generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Normalize line endings and whitespace

Browse Source
This commit is contained in:
OpenCV Buildbot
2012-10-17 11:12:04 +04:00
committed by Andrey Kamaev
parent 0442bca235
commit 81f826db2b
1511 changed files with 258678 additions and 258624 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
modules/gpu/CMakeLists.txt
View File

@@ -44,7 +44,7 @@ if (HAVE_CUDA)
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler /wd4251)
endif()
endif()
ocv_cuda_compile(cuda_objs ${lib_cuda} ${ncv_cuda})
#CUDA_BUILD_CLEAN_TARGET()

2408
modules/gpu/doc/video.rst
View File

File diff suppressed because it is too large Load Diff

1428
modules/gpu/include/opencv2/gpu/device/border_interpolate.hpp
View File

File diff suppressed because it is too large Load Diff

442
modules/gpu/include/opencv2/gpu/device/color.hpp
View File

@@ -1,221 +1,221 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_COLOR_HPP__
#define __OPENCV_GPU_COLOR_HPP__
#include "detail/color_detail.hpp"
namespace cv { namespace gpu { namespace device
{
// All OPENCV_GPU_IMPLEMENT_*_TRAITS(ColorSpace1_to_ColorSpace2, ...) macros implements
// template <typename T> class ColorSpace1_to_ColorSpace2_traits
// {
// typedef ... functor_type;
// static __host__ __device__ functor_type create_functor();
// };
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgba, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr555, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr565, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr555, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr565, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr555, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr565, 4, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr555, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr565, 4, 2, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgb, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgb, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgr, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgr, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgba, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgba, 4, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgra, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgra, 4, 0, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgr, 3)
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgra, 4)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr555, 5)
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr565, 6)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr555_to_gray, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr565_to_gray, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgb_to_gray, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgr_to_gray, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgba_to_gray, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgra_to_gray, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv4, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv4, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgba, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgba, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgr, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgra, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgr, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgra, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_BORDER_INTERPOLATE_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_COLOR_HPP__
#define __OPENCV_GPU_COLOR_HPP__
#include "detail/color_detail.hpp"
namespace cv { namespace gpu { namespace device
{
// All OPENCV_GPU_IMPLEMENT_*_TRAITS(ColorSpace1_to_ColorSpace2, ...) macros implements
// template <typename T> class ColorSpace1_to_ColorSpace2_traits
// {
// typedef ... functor_type;
// static __host__ __device__ functor_type create_functor();
// };
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgba, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr555, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr565, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr555, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr565, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr555, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr565, 4, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr555, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr565, 4, 2, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgb, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgb, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgr, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgr, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgba, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgba, 4, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgra, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgra, 4, 0, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgr, 3)
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgra, 4)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr555, 5)
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr565, 6)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr555_to_gray, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr565_to_gray, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgb_to_gray, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgr_to_gray, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgba_to_gray, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgra_to_gray, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv4, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv4, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgba, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgba, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgr, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgra, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgr, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgra, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_BORDER_INTERPOLATE_HPP__

228
modules/gpu/include/opencv2/gpu/device/common.hpp
View File

@@ -1,114 +1,114 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_COMMON_HPP__
#define __OPENCV_GPU_COMMON_HPP__
#include <cuda_runtime.h>
#include "opencv2/core/cuda_devptrs.hpp"
#ifndef CV_PI
#define CV_PI 3.1415926535897932384626433832795
#endif
#ifndef CV_PI_F
#ifndef CV_PI
#define CV_PI_F 3.14159265f
#else
#define CV_PI_F ((float)CV_PI)
#endif
#endif
#if defined(__GNUC__)
#define cudaSafeCall(expr) ___cudaSafeCall(expr, __FILE__, __LINE__, __func__)
#else /* defined(__CUDACC__) || defined(__MSVC__) */
#define cudaSafeCall(expr) ___cudaSafeCall(expr, __FILE__, __LINE__)
#endif
namespace cv { namespace gpu
{
void error(const char *error_string, const char *file, const int line, const char *func);
template <typename T> static inline bool isAligned(const T* ptr, size_t size)
{
return reinterpret_cast<size_t>(ptr) % size == 0;
}
static inline bool isAligned(size_t step, size_t size)
{
return step % size == 0;
}
}}
static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int line, const char *func = "")
{
if (cudaSuccess != err)
cv::gpu::error(cudaGetErrorString(err), file, line, func);
}
#ifdef __CUDACC__
namespace cv { namespace gpu
{
__host__ __device__ __forceinline__ int divUp(int total, int grain)
{
return (total + grain - 1) / grain;
}
namespace device
{
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef signed char schar;
typedef unsigned int uint;
template<class T> inline void bindTexture(const textureReference* tex, const PtrStepSz<T>& img)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
}
}
}}
#endif // __CUDACC__
#endif // __OPENCV_GPU_COMMON_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_COMMON_HPP__
#define __OPENCV_GPU_COMMON_HPP__
#include <cuda_runtime.h>
#include "opencv2/core/cuda_devptrs.hpp"
#ifndef CV_PI
#define CV_PI 3.1415926535897932384626433832795
#endif
#ifndef CV_PI_F
#ifndef CV_PI
#define CV_PI_F 3.14159265f
#else
#define CV_PI_F ((float)CV_PI)
#endif
#endif
#if defined(__GNUC__)
#define cudaSafeCall(expr) ___cudaSafeCall(expr, __FILE__, __LINE__, __func__)
#else /* defined(__CUDACC__) || defined(__MSVC__) */
#define cudaSafeCall(expr) ___cudaSafeCall(expr, __FILE__, __LINE__)
#endif
namespace cv { namespace gpu
{
void error(const char *error_string, const char *file, const int line, const char *func);
template <typename T> static inline bool isAligned(const T* ptr, size_t size)
{
return reinterpret_cast<size_t>(ptr) % size == 0;
}
static inline bool isAligned(size_t step, size_t size)
{
return step % size == 0;
}
}}
static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int line, const char *func = "")
{
if (cudaSuccess != err)
cv::gpu::error(cudaGetErrorString(err), file, line, func);
}
#ifdef __CUDACC__
namespace cv { namespace gpu
{
__host__ __device__ __forceinline__ int divUp(int total, int grain)
{
return (total + grain - 1) / grain;
}
namespace device
{
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef signed char schar;
typedef unsigned int uint;
template<class T> inline void bindTexture(const textureReference* tex, const PtrStepSz<T>& img)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
}
}
}}
#endif // __CUDACC__
#endif // __OPENCV_GPU_COMMON_HPP__

210
modules/gpu/include/opencv2/gpu/device/datamov_utils.hpp
View File

@@ -1,105 +1,105 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DATAMOV_UTILS_HPP__
#define __OPENCV_GPU_DATAMOV_UTILS_HPP__
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 200
// for Fermi memory space is detected automatically
template <typename T> struct ForceGlob
{
__device__ __forceinline__ static void Load(const T* ptr, int offset, T& val) { val = ptr[offset]; }
};
#else // __CUDA_ARCH__ >= 200
#if defined(_WIN64) || defined(__LP64__)
// 64-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "l"
#else
// 32-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "r"
#endif
template<class T> struct ForceGlob;
#define OPENCV_GPU_DEFINE_FORCE_GLOB(base_type, ptx_type, reg_mod) \
template <> struct ForceGlob<base_type> \
{ \
__device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
{ \
asm("ld.global."#ptx_type" %0, [%1];" : "="#reg_mod(val) : OPENCV_GPU_ASM_PTR(ptr + offset)); \
} \
};
#define OPENCV_GPU_DEFINE_FORCE_GLOB_B(base_type, ptx_type) \
template <> struct ForceGlob<base_type> \
{ \
__device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
{ \
asm("ld.global."#ptx_type" %0, [%1];" : "=r"(*reinterpret_cast<uint*>(&val)) : OPENCV_GPU_ASM_PTR(ptr + offset)); \
} \
};
OPENCV_GPU_DEFINE_FORCE_GLOB_B(uchar, u8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(schar, s8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(char, b8)
OPENCV_GPU_DEFINE_FORCE_GLOB (ushort, u16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (short, s16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (uint, u32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (int, s32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (float, f32, f)
OPENCV_GPU_DEFINE_FORCE_GLOB (double, f64, d)
#undef OPENCV_GPU_DEFINE_FORCE_GLOB
#undef OPENCV_GPU_DEFINE_FORCE_GLOB_B
#undef OPENCV_GPU_ASM_PTR
#endif // __CUDA_ARCH__ >= 200
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_DATAMOV_UTILS_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DATAMOV_UTILS_HPP__
#define __OPENCV_GPU_DATAMOV_UTILS_HPP__
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 200
// for Fermi memory space is detected automatically
template <typename T> struct ForceGlob
{
__device__ __forceinline__ static void Load(const T* ptr, int offset, T& val) { val = ptr[offset]; }
};
#else // __CUDA_ARCH__ >= 200
#if defined(_WIN64) || defined(__LP64__)
// 64-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "l"
#else
// 32-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "r"
#endif
template<class T> struct ForceGlob;
#define OPENCV_GPU_DEFINE_FORCE_GLOB(base_type, ptx_type, reg_mod) \
template <> struct ForceGlob<base_type> \
{ \
__device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
{ \
asm("ld.global."#ptx_type" %0, [%1];" : "="#reg_mod(val) : OPENCV_GPU_ASM_PTR(ptr + offset)); \
} \
};
#define OPENCV_GPU_DEFINE_FORCE_GLOB_B(base_type, ptx_type) \
template <> struct ForceGlob<base_type> \
{ \
__device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
{ \
asm("ld.global."#ptx_type" %0, [%1];" : "=r"(*reinterpret_cast<uint*>(&val)) : OPENCV_GPU_ASM_PTR(ptr + offset)); \
} \
};
OPENCV_GPU_DEFINE_FORCE_GLOB_B(uchar, u8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(schar, s8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(char, b8)
OPENCV_GPU_DEFINE_FORCE_GLOB (ushort, u16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (short, s16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (uint, u32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (int, s32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (float, f32, f)
OPENCV_GPU_DEFINE_FORCE_GLOB (double, f64, d)
#undef OPENCV_GPU_DEFINE_FORCE_GLOB
#undef OPENCV_GPU_DEFINE_FORCE_GLOB_B
#undef OPENCV_GPU_ASM_PTR
#endif // __CUDA_ARCH__ >= 200
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_DATAMOV_UTILS_HPP__

3084
modules/gpu/include/opencv2/gpu/device/detail/color_detail.hpp
View File

File diff suppressed because it is too large Load Diff

1682
modules/gpu/include/opencv2/gpu/device/detail/reduction_detail.hpp
View File

File diff suppressed because it is too large Load Diff

790
modules/gpu/include/opencv2/gpu/device/detail/transform_detail.hpp
View File

@@ -1,395 +1,395 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TRANSFORM_DETAIL_HPP__
#define __OPENCV_GPU_TRANSFORM_DETAIL_HPP__
#include "../common.hpp"
#include "../vec_traits.hpp"
#include "../functional.hpp"
namespace cv { namespace gpu { namespace device
{
namespace transform_detail
{
//! Read Write Traits
template <typename T, typename D, int shift> struct UnaryReadWriteTraits
{
typedef typename TypeVec<T, shift>::vec_type read_type;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
template <typename T1, typename T2, typename D, int shift> struct BinaryReadWriteTraits
{
typedef typename TypeVec<T1, shift>::vec_type read_type1;
typedef typename TypeVec<T2, shift>::vec_type read_type2;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
//! Transform kernels
template <int shift> struct OpUnroller;
template <> struct OpUnroller<1>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
}
};
template <> struct OpUnroller<2>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
}
};
template <> struct OpUnroller<3>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
}
};
template <> struct OpUnroller<4>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
if (mask(y, x_shifted + 3))
dst.w = op(src.w);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
if (mask(y, x_shifted + 3))
dst.w = op(src1.w, src2.w);
}
};
template <> struct OpUnroller<8>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.a0 = op(src.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src.a7);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.a0 = op(src1.a0, src2.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src1.a1, src2.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src1.a2, src2.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src1.a3, src2.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src1.a4, src2.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src1.a5, src2.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src1.a6, src2.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src1.a7, src2.a7);
}
};
template <typename T, typename D, typename UnOp, typename Mask>
static __global__ void transformSmart(const PtrStepSz<T> src_, PtrStep<D> dst_, const Mask mask, const UnOp op)
{
typedef TransformFunctorTraits<UnOp> ft;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::read_type read_type;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::write_type write_type;
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
if (y < src_.rows)
{
const T* src = src_.ptr(y);
D* dst = dst_.ptr(y);
if (x_shifted + ft::smart_shift - 1 < src_.cols)
{
const read_type src_n_el = ((const read_type*)src)[x];
write_type dst_n_el = ((const write_type*)dst)[x];
OpUnroller<ft::smart_shift>::unroll(src_n_el, dst_n_el, mask, op, x_shifted, y);
((write_type*)dst)[x] = dst_n_el;
}
else
{
for (int real_x = x_shifted; real_x < src_.cols; ++real_x)
{
if (mask(y, real_x))
dst[real_x] = op(src[real_x]);
}
}
}
}
template <typename T, typename D, typename UnOp, typename Mask>
__global__ static void transformSimple(const PtrStepSz<T> src, PtrStep<D> dst, const Mask mask, const UnOp op)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < src.cols && y < src.rows && mask(y, x))
{
dst.ptr(y)[x] = op(src.ptr(y)[x]);
}
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __global__ void transformSmart(const PtrStepSz<T1> src1_, const PtrStep<T2> src2_, PtrStep<D> dst_,
const Mask mask, const BinOp op)
{
typedef TransformFunctorTraits<BinOp> ft;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type1 read_type1;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type2 read_type2;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::write_type write_type;
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
if (y < src1_.rows)
{
const T1* src1 = src1_.ptr(y);
const T2* src2 = src2_.ptr(y);
D* dst = dst_.ptr(y);
if (x_shifted + ft::smart_shift - 1 < src1_.cols)
{
const read_type1 src1_n_el = ((const read_type1*)src1)[x];
const read_type2 src2_n_el = ((const read_type2*)src2)[x];
write_type dst_n_el = ((const write_type*)dst)[x];
OpUnroller<ft::smart_shift>::unroll(src1_n_el, src2_n_el, dst_n_el, mask, op, x_shifted, y);
((write_type*)dst)[x] = dst_n_el;
}
else
{
for (int real_x = x_shifted; real_x < src1_.cols; ++real_x)
{
if (mask(y, real_x))
dst[real_x] = op(src1[real_x], src2[real_x]);
}
}
}
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __global__ void transformSimple(const PtrStepSz<T1> src1, const PtrStep<T2> src2, PtrStep<D> dst,
const Mask mask, const BinOp op)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < src1.cols && y < src1.rows && mask(y, x))
{
const T1 src1_data = src1.ptr(y)[x];
const T2 src2_data = src2.ptr(y)[x];
dst.ptr(y)[x] = op(src1_data, src2_data);
}
}
template <bool UseSmart> struct TransformDispatcher;
template<> struct TransformDispatcher<false>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x), divUp(src.rows, threads.y), 1);
transformSimple<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x), divUp(src1.rows, threads.y), 1);
transformSimple<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template<> struct TransformDispatcher<true>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
StaticAssert<ft::smart_shift != 1>::check();
if (!isAligned(src.data, ft::smart_shift * sizeof(T)) || !isAligned(src.step, ft::smart_shift * sizeof(T)) ||
!isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
{
TransformDispatcher<false>::call(src, dst, op, mask, stream);
return;
}
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x * ft::smart_shift), divUp(src.rows, threads.y), 1);
transformSmart<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
StaticAssert<ft::smart_shift != 1>::check();
if (!isAligned(src1.data, ft::smart_shift * sizeof(T1)) || !isAligned(src1.step, ft::smart_shift * sizeof(T1)) ||
!isAligned(src2.data, ft::smart_shift * sizeof(T2)) || !isAligned(src2.step, ft::smart_shift * sizeof(T2)) ||
!isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
{
TransformDispatcher<false>::call(src1, src2, dst, op, mask, stream);
return;
}
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x * ft::smart_shift), divUp(src1.rows, threads.y), 1);
transformSmart<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
} // namespace transform_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_TRANSFORM_DETAIL_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TRANSFORM_DETAIL_HPP__
#define __OPENCV_GPU_TRANSFORM_DETAIL_HPP__
#include "../common.hpp"
#include "../vec_traits.hpp"
#include "../functional.hpp"
namespace cv { namespace gpu { namespace device
{
namespace transform_detail
{
//! Read Write Traits
template <typename T, typename D, int shift> struct UnaryReadWriteTraits
{
typedef typename TypeVec<T, shift>::vec_type read_type;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
template <typename T1, typename T2, typename D, int shift> struct BinaryReadWriteTraits
{
typedef typename TypeVec<T1, shift>::vec_type read_type1;
typedef typename TypeVec<T2, shift>::vec_type read_type2;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
//! Transform kernels
template <int shift> struct OpUnroller;
template <> struct OpUnroller<1>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
}
};
template <> struct OpUnroller<2>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
}
};
template <> struct OpUnroller<3>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
}
};
template <> struct OpUnroller<4>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
if (mask(y, x_shifted + 3))
dst.w = op(src.w);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
if (mask(y, x_shifted + 3))
dst.w = op(src1.w, src2.w);
}
};
template <> struct OpUnroller<8>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.a0 = op(src.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src.a7);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.a0 = op(src1.a0, src2.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src1.a1, src2.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src1.a2, src2.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src1.a3, src2.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src1.a4, src2.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src1.a5, src2.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src1.a6, src2.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src1.a7, src2.a7);
}
};
template <typename T, typename D, typename UnOp, typename Mask>
static __global__ void transformSmart(const PtrStepSz<T> src_, PtrStep<D> dst_, const Mask mask, const UnOp op)
{
typedef TransformFunctorTraits<UnOp> ft;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::read_type read_type;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::write_type write_type;
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
if (y < src_.rows)
{
const T* src = src_.ptr(y);
D* dst = dst_.ptr(y);
if (x_shifted + ft::smart_shift - 1 < src_.cols)
{
const read_type src_n_el = ((const read_type*)src)[x];
write_type dst_n_el = ((const write_type*)dst)[x];
OpUnroller<ft::smart_shift>::unroll(src_n_el, dst_n_el, mask, op, x_shifted, y);
((write_type*)dst)[x] = dst_n_el;
}
else
{
for (int real_x = x_shifted; real_x < src_.cols; ++real_x)
{
if (mask(y, real_x))
dst[real_x] = op(src[real_x]);
}
}
}
}
template <typename T, typename D, typename UnOp, typename Mask>
__global__ static void transformSimple(const PtrStepSz<T> src, PtrStep<D> dst, const Mask mask, const UnOp op)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < src.cols && y < src.rows && mask(y, x))
{
dst.ptr(y)[x] = op(src.ptr(y)[x]);
}
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __global__ void transformSmart(const PtrStepSz<T1> src1_, const PtrStep<T2> src2_, PtrStep<D> dst_,
const Mask mask, const BinOp op)
{
typedef TransformFunctorTraits<BinOp> ft;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type1 read_type1;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type2 read_type2;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::write_type write_type;
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
if (y < src1_.rows)
{
const T1* src1 = src1_.ptr(y);
const T2* src2 = src2_.ptr(y);
D* dst = dst_.ptr(y);
if (x_shifted + ft::smart_shift - 1 < src1_.cols)
{
const read_type1 src1_n_el = ((const read_type1*)src1)[x];
const read_type2 src2_n_el = ((const read_type2*)src2)[x];
write_type dst_n_el = ((const write_type*)dst)[x];
OpUnroller<ft::smart_shift>::unroll(src1_n_el, src2_n_el, dst_n_el, mask, op, x_shifted, y);
((write_type*)dst)[x] = dst_n_el;
}
else
{
for (int real_x = x_shifted; real_x < src1_.cols; ++real_x)
{
if (mask(y, real_x))
dst[real_x] = op(src1[real_x], src2[real_x]);
}
}
}
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __global__ void transformSimple(const PtrStepSz<T1> src1, const PtrStep<T2> src2, PtrStep<D> dst,
const Mask mask, const BinOp op)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < src1.cols && y < src1.rows && mask(y, x))
{
const T1 src1_data = src1.ptr(y)[x];
const T2 src2_data = src2.ptr(y)[x];
dst.ptr(y)[x] = op(src1_data, src2_data);
}
}
template <bool UseSmart> struct TransformDispatcher;
template<> struct TransformDispatcher<false>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x), divUp(src.rows, threads.y), 1);
transformSimple<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x), divUp(src1.rows, threads.y), 1);
transformSimple<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template<> struct TransformDispatcher<true>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
StaticAssert<ft::smart_shift != 1>::check();
if (!isAligned(src.data, ft::smart_shift * sizeof(T)) || !isAligned(src.step, ft::smart_shift * sizeof(T)) ||
!isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
{
TransformDispatcher<false>::call(src, dst, op, mask, stream);
return;
}
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x * ft::smart_shift), divUp(src.rows, threads.y), 1);
transformSmart<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
StaticAssert<ft::smart_shift != 1>::check();
if (!isAligned(src1.data, ft::smart_shift * sizeof(T1)) || !isAligned(src1.step, ft::smart_shift * sizeof(T1)) ||
!isAligned(src2.data, ft::smart_shift * sizeof(T2)) || !isAligned(src2.step, ft::smart_shift * sizeof(T2)) ||
!isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
{
TransformDispatcher<false>::call(src1, src2, dst, op, mask, stream);
return;
}
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x * ft::smart_shift), divUp(src1.rows, threads.y), 1);
transformSmart<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
} // namespace transform_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_TRANSFORM_DETAIL_HPP__

374
modules/gpu/include/opencv2/gpu/device/detail/type_traits_detail.hpp
View File

@@ -1,187 +1,187 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TYPE_TRAITS_DETAIL_HPP__
#define __OPENCV_GPU_TYPE_TRAITS_DETAIL_HPP__
#include "../common.hpp"
#include "../vec_traits.hpp"
namespace cv { namespace gpu { namespace device
{
namespace type_traits_detail
{
template <bool, typename T1, typename T2> struct Select { typedef T1 type; };
template <typename T1, typename T2> struct Select<false, T1, T2> { typedef T2 type; };
template <typename T> struct IsSignedIntergral { enum {value = 0}; };
template <> struct IsSignedIntergral<schar> { enum {value = 1}; };
template <> struct IsSignedIntergral<char1> { enum {value = 1}; };
template <> struct IsSignedIntergral<short> { enum {value = 1}; };
template <> struct IsSignedIntergral<short1> { enum {value = 1}; };
template <> struct IsSignedIntergral<int> { enum {value = 1}; };
template <> struct IsSignedIntergral<int1> { enum {value = 1}; };
template <typename T> struct IsUnsignedIntegral { enum {value = 0}; };
template <> struct IsUnsignedIntegral<uchar> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uchar1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint1> { enum {value = 1}; };
template <typename T> struct IsIntegral { enum {value = IsSignedIntergral<T>::value || IsUnsignedIntegral<T>::value}; };
template <> struct IsIntegral<char> { enum {value = 1}; };
template <> struct IsIntegral<bool> { enum {value = 1}; };
template <typename T> struct IsFloat { enum {value = 0}; };
template <> struct IsFloat<float> { enum {value = 1}; };
template <> struct IsFloat<double> { enum {value = 1}; };
template <typename T> struct IsVec { enum {value = 0}; };
template <> struct IsVec<uchar1> { enum {value = 1}; };
template <> struct IsVec<uchar2> { enum {value = 1}; };
template <> struct IsVec<uchar3> { enum {value = 1}; };
template <> struct IsVec<uchar4> { enum {value = 1}; };
template <> struct IsVec<uchar8> { enum {value = 1}; };
template <> struct IsVec<char1> { enum {value = 1}; };
template <> struct IsVec<char2> { enum {value = 1}; };
template <> struct IsVec<char3> { enum {value = 1}; };
template <> struct IsVec<char4> { enum {value = 1}; };
template <> struct IsVec<char8> { enum {value = 1}; };
template <> struct IsVec<ushort1> { enum {value = 1}; };
template <> struct IsVec<ushort2> { enum {value = 1}; };
template <> struct IsVec<ushort3> { enum {value = 1}; };
template <> struct IsVec<ushort4> { enum {value = 1}; };
template <> struct IsVec<ushort8> { enum {value = 1}; };
template <> struct IsVec<short1> { enum {value = 1}; };
template <> struct IsVec<short2> { enum {value = 1}; };
template <> struct IsVec<short3> { enum {value = 1}; };
template <> struct IsVec<short4> { enum {value = 1}; };
template <> struct IsVec<short8> { enum {value = 1}; };
template <> struct IsVec<uint1> { enum {value = 1}; };
template <> struct IsVec<uint2> { enum {value = 1}; };
template <> struct IsVec<uint3> { enum {value = 1}; };
template <> struct IsVec<uint4> { enum {value = 1}; };
template <> struct IsVec<uint8> { enum {value = 1}; };
template <> struct IsVec<int1> { enum {value = 1}; };
template <> struct IsVec<int2> { enum {value = 1}; };
template <> struct IsVec<int3> { enum {value = 1}; };
template <> struct IsVec<int4> { enum {value = 1}; };
template <> struct IsVec<int8> { enum {value = 1}; };
template <> struct IsVec<float1> { enum {value = 1}; };
template <> struct IsVec<float2> { enum {value = 1}; };
template <> struct IsVec<float3> { enum {value = 1}; };
template <> struct IsVec<float4> { enum {value = 1}; };
template <> struct IsVec<float8> { enum {value = 1}; };
template <> struct IsVec<double1> { enum {value = 1}; };
template <> struct IsVec<double2> { enum {value = 1}; };
template <> struct IsVec<double3> { enum {value = 1}; };
template <> struct IsVec<double4> { enum {value = 1}; };
template <> struct IsVec<double8> { enum {value = 1}; };
template <class U> struct AddParameterType { typedef const U& type; };
template <class U> struct AddParameterType<U&> { typedef U& type; };
template <> struct AddParameterType<void> { typedef void type; };
template <class U> struct ReferenceTraits
{
enum { value = false };
typedef U type;
};
template <class U> struct ReferenceTraits<U&>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits
{
enum { value = false };
typedef void type;
};
template <class U> struct PointerTraits<U*>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits<U*&>
{
enum { value = true };
typedef U type;
};
template <class U> struct UnConst
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnConst<const U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnConst<const U&>
{
typedef U& type;
enum { value = 1 };
};
template <class U> struct UnVolatile
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnVolatile<volatile U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnVolatile<volatile U&>
{
typedef U& type;
enum { value = 1 };
};
} // namespace type_traits_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_TYPE_TRAITS_DETAIL_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TYPE_TRAITS_DETAIL_HPP__
#define __OPENCV_GPU_TYPE_TRAITS_DETAIL_HPP__
#include "../common.hpp"
#include "../vec_traits.hpp"
namespace cv { namespace gpu { namespace device
{
namespace type_traits_detail
{
template <bool, typename T1, typename T2> struct Select { typedef T1 type; };
template <typename T1, typename T2> struct Select<false, T1, T2> { typedef T2 type; };
template <typename T> struct IsSignedIntergral { enum {value = 0}; };
template <> struct IsSignedIntergral<schar> { enum {value = 1}; };
template <> struct IsSignedIntergral<char1> { enum {value = 1}; };
template <> struct IsSignedIntergral<short> { enum {value = 1}; };
template <> struct IsSignedIntergral<short1> { enum {value = 1}; };
template <> struct IsSignedIntergral<int> { enum {value = 1}; };
template <> struct IsSignedIntergral<int1> { enum {value = 1}; };
template <typename T> struct IsUnsignedIntegral { enum {value = 0}; };
template <> struct IsUnsignedIntegral<uchar> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uchar1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint1> { enum {value = 1}; };
template <typename T> struct IsIntegral { enum {value = IsSignedIntergral<T>::value || IsUnsignedIntegral<T>::value}; };
template <> struct IsIntegral<char> { enum {value = 1}; };
template <> struct IsIntegral<bool> { enum {value = 1}; };
template <typename T> struct IsFloat { enum {value = 0}; };
template <> struct IsFloat<float> { enum {value = 1}; };
template <> struct IsFloat<double> { enum {value = 1}; };
template <typename T> struct IsVec { enum {value = 0}; };
template <> struct IsVec<uchar1> { enum {value = 1}; };
template <> struct IsVec<uchar2> { enum {value = 1}; };
template <> struct IsVec<uchar3> { enum {value = 1}; };
template <> struct IsVec<uchar4> { enum {value = 1}; };
template <> struct IsVec<uchar8> { enum {value = 1}; };
template <> struct IsVec<char1> { enum {value = 1}; };
template <> struct IsVec<char2> { enum {value = 1}; };
template <> struct IsVec<char3> { enum {value = 1}; };
template <> struct IsVec<char4> { enum {value = 1}; };
template <> struct IsVec<char8> { enum {value = 1}; };
template <> struct IsVec<ushort1> { enum {value = 1}; };
template <> struct IsVec<ushort2> { enum {value = 1}; };
template <> struct IsVec<ushort3> { enum {value = 1}; };
template <> struct IsVec<ushort4> { enum {value = 1}; };
template <> struct IsVec<ushort8> { enum {value = 1}; };
template <> struct IsVec<short1> { enum {value = 1}; };
template <> struct IsVec<short2> { enum {value = 1}; };
template <> struct IsVec<short3> { enum {value = 1}; };
template <> struct IsVec<short4> { enum {value = 1}; };
template <> struct IsVec<short8> { enum {value = 1}; };
template <> struct IsVec<uint1> { enum {value = 1}; };
template <> struct IsVec<uint2> { enum {value = 1}; };
template <> struct IsVec<uint3> { enum {value = 1}; };
template <> struct IsVec<uint4> { enum {value = 1}; };
template <> struct IsVec<uint8> { enum {value = 1}; };
template <> struct IsVec<int1> { enum {value = 1}; };
template <> struct IsVec<int2> { enum {value = 1}; };
template <> struct IsVec<int3> { enum {value = 1}; };
template <> struct IsVec<int4> { enum {value = 1}; };
template <> struct IsVec<int8> { enum {value = 1}; };
template <> struct IsVec<float1> { enum {value = 1}; };
template <> struct IsVec<float2> { enum {value = 1}; };
template <> struct IsVec<float3> { enum {value = 1}; };
template <> struct IsVec<float4> { enum {value = 1}; };
template <> struct IsVec<float8> { enum {value = 1}; };
template <> struct IsVec<double1> { enum {value = 1}; };
template <> struct IsVec<double2> { enum {value = 1}; };
template <> struct IsVec<double3> { enum {value = 1}; };
template <> struct IsVec<double4> { enum {value = 1}; };
template <> struct IsVec<double8> { enum {value = 1}; };
template <class U> struct AddParameterType { typedef const U& type; };
template <class U> struct AddParameterType<U&> { typedef U& type; };
template <> struct AddParameterType<void> { typedef void type; };
template <class U> struct ReferenceTraits
{
enum { value = false };
typedef U type;
};
template <class U> struct ReferenceTraits<U&>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits
{
enum { value = false };
typedef void type;
};
template <class U> struct PointerTraits<U*>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits<U*&>
{
enum { value = true };
typedef U type;
};
template <class U> struct UnConst
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnConst<const U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnConst<const U&>
{
typedef U& type;
enum { value = 1 };
};
template <class U> struct UnVolatile
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnVolatile<volatile U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnVolatile<volatile U&>
{
typedef U& type;
enum { value = 1 };
};
} // namespace type_traits_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_TYPE_TRAITS_DETAIL_HPP__

234
modules/gpu/include/opencv2/gpu/device/detail/vec_distance_detail.hpp
View File

@@ -1,117 +1,117 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VEC_DISTANCE_DETAIL_HPP__
#define __OPENCV_GPU_VEC_DISTANCE_DETAIL_HPP__
#include "../datamov_utils.hpp"
namespace cv { namespace gpu { namespace device
{
namespace vec_distance_detail
{
template <int THREAD_DIM, int N> struct UnrollVecDiffCached
{
template <typename Dist, typename T1, typename T2>
static __device__ void calcCheck(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int ind)
{
if (ind < len)
{
T1 val1 = *vecCached++;
T2 val2;
ForceGlob<T2>::Load(vecGlob, ind, val2);
dist.reduceIter(val1, val2);
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcCheck(vecCached, vecGlob, len, dist, ind + THREAD_DIM);
}
}
template <typename Dist, typename T1, typename T2>
static __device__ void calcWithoutCheck(const T1* vecCached, const T2* vecGlob, Dist& dist)
{
T1 val1 = *vecCached++;
T2 val2;
ForceGlob<T2>::Load(vecGlob, 0, val2);
vecGlob += THREAD_DIM;
dist.reduceIter(val1, val2);
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcWithoutCheck(vecCached, vecGlob, dist);
}
};
template <int THREAD_DIM> struct UnrollVecDiffCached<THREAD_DIM, 0>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcCheck(const T1*, const T2*, int, Dist&, int)
{
}
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcWithoutCheck(const T1*, const T2*, Dist&)
{
}
};
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN> struct VecDiffCachedCalculator;
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, false>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcCheck(vecCached, vecGlob, len, dist, tid);
}
};
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, true>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcWithoutCheck(vecCached, vecGlob + tid, dist);
}
};
} // namespace vec_distance_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_DISTANCE_DETAIL_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VEC_DISTANCE_DETAIL_HPP__
#define __OPENCV_GPU_VEC_DISTANCE_DETAIL_HPP__
#include "../datamov_utils.hpp"
namespace cv { namespace gpu { namespace device
{
namespace vec_distance_detail
{
template <int THREAD_DIM, int N> struct UnrollVecDiffCached
{
template <typename Dist, typename T1, typename T2>
static __device__ void calcCheck(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int ind)
{
if (ind < len)
{
T1 val1 = *vecCached++;
T2 val2;
ForceGlob<T2>::Load(vecGlob, ind, val2);
dist.reduceIter(val1, val2);
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcCheck(vecCached, vecGlob, len, dist, ind + THREAD_DIM);
}
}
template <typename Dist, typename T1, typename T2>
static __device__ void calcWithoutCheck(const T1* vecCached, const T2* vecGlob, Dist& dist)
{
T1 val1 = *vecCached++;
T2 val2;
ForceGlob<T2>::Load(vecGlob, 0, val2);
vecGlob += THREAD_DIM;
dist.reduceIter(val1, val2);
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcWithoutCheck(vecCached, vecGlob, dist);
}
};
template <int THREAD_DIM> struct UnrollVecDiffCached<THREAD_DIM, 0>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcCheck(const T1*, const T2*, int, Dist&, int)
{
}
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcWithoutCheck(const T1*, const T2*, Dist&)
{
}
};
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN> struct VecDiffCachedCalculator;
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, false>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcCheck(vecCached, vecGlob, len, dist, tid);
}
};
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, true>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcWithoutCheck(vecCached, vecGlob + tid, dist);
}
};
} // namespace vec_distance_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_DISTANCE_DETAIL_HPP__

160
modules/gpu/include/opencv2/gpu/device/dynamic_smem.hpp
View File

@@ -1,80 +1,80 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DYNAMIC_SMEM_HPP__
#define __OPENCV_GPU_DYNAMIC_SMEM_HPP__
namespace cv { namespace gpu { namespace device
{
template<class T> struct DynamicSharedMem
{
__device__ __forceinline__ operator T*()
{
extern __shared__ int __smem[];
return (T*)__smem;
}
__device__ __forceinline__ operator const T*() const
{
extern __shared__ int __smem[];
return (T*)__smem;
}
};
// specialize for double to avoid unaligned memory access compile errors
template<> struct DynamicSharedMem<double>
{
__device__ __forceinline__ operator double*()
{
extern __shared__ double __smem_d[];
return (double*)__smem_d;
}
__device__ __forceinline__ operator const double*() const
{
extern __shared__ double __smem_d[];
return (double*)__smem_d;
}
};
}}}
#endif // __OPENCV_GPU_DYNAMIC_SMEM_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DYNAMIC_SMEM_HPP__
#define __OPENCV_GPU_DYNAMIC_SMEM_HPP__
namespace cv { namespace gpu { namespace device
{
template<class T> struct DynamicSharedMem
{
__device__ __forceinline__ operator T*()
{
extern __shared__ int __smem[];
return (T*)__smem;
}
__device__ __forceinline__ operator const T*() const
{
extern __shared__ int __smem[];
return (T*)__smem;
}
};
// specialize for double to avoid unaligned memory access compile errors
template<> struct DynamicSharedMem<double>
{
__device__ __forceinline__ operator double*()
{
extern __shared__ double __smem_d[];
return (double*)__smem_d;
}
__device__ __forceinline__ operator const double*() const
{
extern __shared__ double __smem_d[];
return (double*)__smem_d;
}
};
}}}
#endif // __OPENCV_GPU_DYNAMIC_SMEM_HPP__

556
modules/gpu/include/opencv2/gpu/device/filters.hpp
View File

@@ -1,278 +1,278 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_FILTERS_HPP__
#define __OPENCV_GPU_FILTERS_HPP__
#include "saturate_cast.hpp"
#include "vec_traits.hpp"
#include "vec_math.hpp"
#include "type_traits.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename Ptr2D> struct PointFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ PointFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
: src(src_)
{
(void)fx;
(void)fy;
}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
return src(__float2int_rz(y), __float2int_rz(x));
}
const Ptr2D src;
};
template <typename Ptr2D> struct LinearFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ LinearFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
: src(src_)
{
(void)fx;
(void)fy;
}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
work_type out = VecTraits<work_type>::all(0);
const int x1 = __float2int_rd(x);
const int y1 = __float2int_rd(y);
const int x2 = x1 + 1;
const int y2 = y1 + 1;
elem_type src_reg = src(y1, x1);
out = out + src_reg * ((x2 - x) * (y2 - y));
src_reg = src(y1, x2);
out = out + src_reg * ((x - x1) * (y2 - y));
src_reg = src(y2, x1);
out = out + src_reg * ((x2 - x) * (y - y1));
src_reg = src(y2, x2);
out = out + src_reg * ((x - x1) * (y - y1));
return saturate_cast<elem_type>(out);
}
const Ptr2D src;
};
template <typename Ptr2D> struct CubicFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
explicit __host__ __device__ __forceinline__ CubicFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
: src(src_)
{
(void)fx;
(void)fy;
}
static __device__ __forceinline__ float bicubicCoeff(float x_)
{
float x = fabsf(x_);
if (x <= 1.0f)
{
return x * x * (1.5f * x - 2.5f) + 1.0f;
}
else if (x < 2.0f)
{
return x * (x * (-0.5f * x + 2.5f) - 4.0f) + 2.0f;
}
else
{
return 0.0f;
}
}
__device__ elem_type operator ()(float y, float x) const
{
const float xmin = ::ceilf(x - 2.0f);
const float xmax = ::floorf(x + 2.0f);
const float ymin = ::ceilf(y - 2.0f);
const float ymax = ::floorf(y + 2.0f);
work_type sum = VecTraits<work_type>::all(0);
float wsum = 0.0f;
for (float cy = ymin; cy <= ymax; cy += 1.0f)
{
for (float cx = xmin; cx <= xmax; cx += 1.0f)
{
const float w = bicubicCoeff(x - cx) * bicubicCoeff(y - cy);
sum = sum + w * src(__float2int_rd(cy), __float2int_rd(cx));
wsum += w;
}
}
work_type res = (!wsum)? VecTraits<work_type>::all(0) : sum / wsum;
return saturate_cast<elem_type>(res);
}
const Ptr2D src;
};
// for integer scaling
template <typename Ptr2D> struct IntegerAreaFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ IntegerAreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
: src(src_), scale_x(scale_x_), scale_y(scale_y_), scale(1.f / (scale_x * scale_y)) {}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
float fsx1 = x * scale_x;
float fsx2 = fsx1 + scale_x;
int sx1 = __float2int_ru(fsx1);
int sx2 = __float2int_rd(fsx2);
float fsy1 = y * scale_y;
float fsy2 = fsy1 + scale_y;
int sy1 = __float2int_ru(fsy1);
int sy2 = __float2int_rd(fsy2);
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
work_type out = VecTraits<work_type>::all(0.f);
for(int dy = sy1; dy < sy2; ++dy)
for(int dx = sx1; dx < sx2; ++dx)
{
out = out + src(dy, dx) * scale;
}
return saturate_cast<elem_type>(out);
}
const Ptr2D src;
float scale_x, scale_y ,scale;
};
template <typename Ptr2D> struct AreaFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ AreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
: src(src_), scale_x(scale_x_), scale_y(scale_y_){}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
float fsx1 = x * scale_x;
float fsx2 = fsx1 + scale_x;
int sx1 = __float2int_ru(fsx1);
int sx2 = __float2int_rd(fsx2);
float fsy1 = y * scale_y;
float fsy2 = fsy1 + scale_y;
int sy1 = __float2int_ru(fsy1);
int sy2 = __float2int_rd(fsy2);
float scale = 1.f / (fminf(scale_x, src.width - fsx1) * fminf(scale_y, src.height - fsy1));
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
work_type out = VecTraits<work_type>::all(0.f);
for (int dy = sy1; dy < sy2; ++dy)
{
for (int dx = sx1; dx < sx2; ++dx)
out = out + src(dy, dx) * scale;
if (sx1 > fsx1)
out = out + src(dy, (sx1 -1) ) * ((sx1 - fsx1) * scale);
if (sx2 < fsx2)
out = out + src(dy, sx2) * ((fsx2 -sx2) * scale);
}
if (sy1 > fsy1)
for (int dx = sx1; dx < sx2; ++dx)
out = out + src( (sy1 - 1) , dx) * ((sy1 -fsy1) * scale);
if (sy2 < fsy2)
for (int dx = sx1; dx < sx2; ++dx)
out = out + src(sy2, dx) * ((fsy2 -sy2) * scale);
if ((sy1 > fsy1) && (sx1 > fsx1))
out = out + src( (sy1 - 1) , (sx1 - 1)) * ((sy1 -fsy1) * (sx1 -fsx1) * scale);
if ((sy1 > fsy1) && (sx2 < fsx2))
out = out + src( (sy1 - 1) , sx2) * ((sy1 -fsy1) * (fsx2 -sx2) * scale);
if ((sy2 < fsy2) && (sx2 < fsx2))
out = out + src(sy2, sx2) * ((fsy2 -sy2) * (fsx2 -sx2) * scale);
if ((sy2 < fsy2) && (sx1 > fsx1))
out = out + src(sy2, (sx1 - 1)) * ((fsy2 -sy2) * (sx1 -fsx1) * scale);
return saturate_cast<elem_type>(out);
}
const Ptr2D src;
float scale_x, scale_y;
int width, haight;
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_FILTERS_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_FILTERS_HPP__
#define __OPENCV_GPU_FILTERS_HPP__
#include "saturate_cast.hpp"
#include "vec_traits.hpp"
#include "vec_math.hpp"
#include "type_traits.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename Ptr2D> struct PointFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ PointFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
: src(src_)
{
(void)fx;
(void)fy;
}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
return src(__float2int_rz(y), __float2int_rz(x));
}
const Ptr2D src;
};
template <typename Ptr2D> struct LinearFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ LinearFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
: src(src_)
{
(void)fx;
(void)fy;
}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
work_type out = VecTraits<work_type>::all(0);
const int x1 = __float2int_rd(x);
const int y1 = __float2int_rd(y);
const int x2 = x1 + 1;
const int y2 = y1 + 1;
elem_type src_reg = src(y1, x1);
out = out + src_reg * ((x2 - x) * (y2 - y));
src_reg = src(y1, x2);
out = out + src_reg * ((x - x1) * (y2 - y));
src_reg = src(y2, x1);
out = out + src_reg * ((x2 - x) * (y - y1));
src_reg = src(y2, x2);
out = out + src_reg * ((x - x1) * (y - y1));
return saturate_cast<elem_type>(out);
}
const Ptr2D src;
};
template <typename Ptr2D> struct CubicFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
explicit __host__ __device__ __forceinline__ CubicFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
: src(src_)
{
(void)fx;
(void)fy;
}
static __device__ __forceinline__ float bicubicCoeff(float x_)
{
float x = fabsf(x_);
if (x <= 1.0f)
{
return x * x * (1.5f * x - 2.5f) + 1.0f;
}
else if (x < 2.0f)
{
return x * (x * (-0.5f * x + 2.5f) - 4.0f) + 2.0f;
}
else
{
return 0.0f;
}
}
__device__ elem_type operator ()(float y, float x) const
{
const float xmin = ::ceilf(x - 2.0f);
const float xmax = ::floorf(x + 2.0f);
const float ymin = ::ceilf(y - 2.0f);
const float ymax = ::floorf(y + 2.0f);
work_type sum = VecTraits<work_type>::all(0);
float wsum = 0.0f;
for (float cy = ymin; cy <= ymax; cy += 1.0f)
{
for (float cx = xmin; cx <= xmax; cx += 1.0f)
{
const float w = bicubicCoeff(x - cx) * bicubicCoeff(y - cy);
sum = sum + w * src(__float2int_rd(cy), __float2int_rd(cx));
wsum += w;
}
}
work_type res = (!wsum)? VecTraits<work_type>::all(0) : sum / wsum;
return saturate_cast<elem_type>(res);
}
const Ptr2D src;
};
// for integer scaling
template <typename Ptr2D> struct IntegerAreaFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ IntegerAreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
: src(src_), scale_x(scale_x_), scale_y(scale_y_), scale(1.f / (scale_x * scale_y)) {}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
float fsx1 = x * scale_x;
float fsx2 = fsx1 + scale_x;
int sx1 = __float2int_ru(fsx1);
int sx2 = __float2int_rd(fsx2);
float fsy1 = y * scale_y;
float fsy2 = fsy1 + scale_y;
int sy1 = __float2int_ru(fsy1);
int sy2 = __float2int_rd(fsy2);
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
work_type out = VecTraits<work_type>::all(0.f);
for(int dy = sy1; dy < sy2; ++dy)
for(int dx = sx1; dx < sx2; ++dx)
{
out = out + src(dy, dx) * scale;
}
return saturate_cast<elem_type>(out);
}
const Ptr2D src;
float scale_x, scale_y ,scale;
};
template <typename Ptr2D> struct AreaFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ AreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
: src(src_), scale_x(scale_x_), scale_y(scale_y_){}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
float fsx1 = x * scale_x;
float fsx2 = fsx1 + scale_x;
int sx1 = __float2int_ru(fsx1);
int sx2 = __float2int_rd(fsx2);
float fsy1 = y * scale_y;
float fsy2 = fsy1 + scale_y;
int sy1 = __float2int_ru(fsy1);
int sy2 = __float2int_rd(fsy2);
float scale = 1.f / (fminf(scale_x, src.width - fsx1) * fminf(scale_y, src.height - fsy1));
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
work_type out = VecTraits<work_type>::all(0.f);
for (int dy = sy1; dy < sy2; ++dy)
{
for (int dx = sx1; dx < sx2; ++dx)
out = out + src(dy, dx) * scale;
if (sx1 > fsx1)
out = out + src(dy, (sx1 -1) ) * ((sx1 - fsx1) * scale);
if (sx2 < fsx2)
out = out + src(dy, sx2) * ((fsx2 -sx2) * scale);
}
if (sy1 > fsy1)
for (int dx = sx1; dx < sx2; ++dx)
out = out + src( (sy1 - 1) , dx) * ((sy1 -fsy1) * scale);
if (sy2 < fsy2)
for (int dx = sx1; dx < sx2; ++dx)
out = out + src(sy2, dx) * ((fsy2 -sy2) * scale);
if ((sy1 > fsy1) && (sx1 > fsx1))
out = out + src( (sy1 - 1) , (sx1 - 1)) * ((sy1 -fsy1) * (sx1 -fsx1) * scale);
if ((sy1 > fsy1) && (sx2 < fsx2))
out = out + src( (sy1 - 1) , sx2) * ((sy1 -fsy1) * (fsx2 -sx2) * scale);
if ((sy2 < fsy2) && (sx2 < fsx2))
out = out + src(sy2, sx2) * ((fsy2 -sy2) * (fsx2 -sx2) * scale);
if ((sy2 < fsy2) && (sx1 > fsx1))
out = out + src(sy2, (sx1 - 1)) * ((fsy2 -sy2) * (sx1 -fsx1) * scale);
return saturate_cast<elem_type>(out);
}
const Ptr2D src;
float scale_x, scale_y;
int width, haight;
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_FILTERS_HPP__

142
modules/gpu/include/opencv2/gpu/device/funcattrib.hpp
View File

@@ -1,72 +1,72 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DEVICE_FUNCATTRIB_HPP_
#define __OPENCV_GPU_DEVICE_FUNCATTRIB_HPP_
#include <cstdio>
namespace cv { namespace gpu { namespace device
{
template<class Func>
void printFuncAttrib(Func& func)
{
cudaFuncAttributes attrs;
cudaFuncGetAttributes(&attrs, func);
printf("=== Function stats ===\n");
printf("Name: \n");
printf("sharedSizeBytes = %d\n", attrs.sharedSizeBytes);
printf("constSizeBytes = %d\n", attrs.constSizeBytes);
printf("localSizeBytes = %d\n", attrs.localSizeBytes);
printf("maxThreadsPerBlock = %d\n", attrs.maxThreadsPerBlock);
printf("numRegs = %d\n", attrs.numRegs);
printf("ptxVersion = %d\n", attrs.ptxVersion);
printf("binaryVersion = %d\n", attrs.binaryVersion);
printf("\n");
fflush(stdout);
}
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DEVICE_FUNCATTRIB_HPP_
#define __OPENCV_GPU_DEVICE_FUNCATTRIB_HPP_
#include <cstdio>
namespace cv { namespace gpu { namespace device
{
template<class Func>
void printFuncAttrib(Func& func)
{
cudaFuncAttributes attrs;
cudaFuncGetAttributes(&attrs, func);
printf("=== Function stats ===\n");
printf("Name: \n");
printf("sharedSizeBytes = %d\n", attrs.sharedSizeBytes);
printf("constSizeBytes = %d\n", attrs.constSizeBytes);
printf("localSizeBytes = %d\n", attrs.localSizeBytes);
printf("maxThreadsPerBlock = %d\n", attrs.maxThreadsPerBlock);
printf("numRegs = %d\n", attrs.numRegs);
printf("ptxVersion = %d\n", attrs.ptxVersion);
printf("binaryVersion = %d\n", attrs.binaryVersion);
printf("\n");
fflush(stdout);
}
}}} // namespace cv { namespace gpu { namespace device
#endif /* __OPENCV_GPU_DEVICE_FUNCATTRIB_HPP_ */

1372
modules/gpu/include/opencv2/gpu/device/functional.hpp
View File

File diff suppressed because it is too large Load Diff

470
modules/gpu/include/opencv2/gpu/device/limits.hpp
View File

@@ -1,235 +1,235 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_LIMITS_GPU_HPP__
#define __OPENCV_GPU_LIMITS_GPU_HPP__
#include <limits>
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
template<class T> struct numeric_limits
{
typedef T type;
__device__ __forceinline__ static type min() { return type(); };
__device__ __forceinline__ static type max() { return type(); };
__device__ __forceinline__ static type epsilon() { return type(); }
__device__ __forceinline__ static type round_error() { return type(); }
__device__ __forceinline__ static type denorm_min() { return type(); }
__device__ __forceinline__ static type infinity() { return type(); }
__device__ __forceinline__ static type quiet_NaN() { return type(); }
__device__ __forceinline__ static type signaling_NaN() { return T(); }
static const bool is_signed;
};
template<> struct numeric_limits<bool>
{
typedef bool type;
__device__ __forceinline__ static type min() { return false; };
__device__ __forceinline__ static type max() { return true; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<char>
{
typedef char type;
__device__ __forceinline__ static type min() { return CHAR_MIN; };
__device__ __forceinline__ static type max() { return CHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (char)-1 == -1;
};
template<> struct numeric_limits<signed char>
{
typedef char type;
__device__ __forceinline__ static type min() { return SCHAR_MIN; };
__device__ __forceinline__ static type max() { return SCHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (signed char)-1 == -1;
};
template<> struct numeric_limits<unsigned char>
{
typedef unsigned char type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UCHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<short>
{
typedef short type;
__device__ __forceinline__ static type min() { return SHRT_MIN; };
__device__ __forceinline__ static type max() { return SHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned short>
{
typedef unsigned short type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return USHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<int>
{
typedef int type;
__device__ __forceinline__ static type min() { return INT_MIN; };
__device__ __forceinline__ static type max() { return INT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned int>
{
typedef unsigned int type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UINT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<long>
{
typedef long type;
__device__ __forceinline__ static type min() { return LONG_MIN; };
__device__ __forceinline__ static type max() { return LONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned long>
{
typedef unsigned long type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return ULONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<float>
{
typedef float type;
__device__ __forceinline__ static type min() { return 1.175494351e-38f/*FLT_MIN*/; };
__device__ __forceinline__ static type max() { return 3.402823466e+38f/*FLT_MAX*/; };
__device__ __forceinline__ static type epsilon() { return 1.192092896e-07f/*FLT_EPSILON*/; };
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<double>
{
typedef double type;
__device__ __forceinline__ static type min() { return 2.2250738585072014e-308/*DBL_MIN*/; };
__device__ __forceinline__ static type max() { return 1.7976931348623158e+308/*DBL_MAX*/; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
}}} // namespace cv { namespace gpu { namespace device {
#endif // __OPENCV_GPU_LIMITS_GPU_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_LIMITS_GPU_HPP__
#define __OPENCV_GPU_LIMITS_GPU_HPP__
#include <limits>
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
template<class T> struct numeric_limits
{
typedef T type;
__device__ __forceinline__ static type min() { return type(); };
__device__ __forceinline__ static type max() { return type(); };
__device__ __forceinline__ static type epsilon() { return type(); }
__device__ __forceinline__ static type round_error() { return type(); }
__device__ __forceinline__ static type denorm_min() { return type(); }
__device__ __forceinline__ static type infinity() { return type(); }
__device__ __forceinline__ static type quiet_NaN() { return type(); }
__device__ __forceinline__ static type signaling_NaN() { return T(); }
static const bool is_signed;
};
template<> struct numeric_limits<bool>
{
typedef bool type;
__device__ __forceinline__ static type min() { return false; };
__device__ __forceinline__ static type max() { return true; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<char>
{
typedef char type;
__device__ __forceinline__ static type min() { return CHAR_MIN; };
__device__ __forceinline__ static type max() { return CHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (char)-1 == -1;
};
template<> struct numeric_limits<signed char>
{
typedef char type;
__device__ __forceinline__ static type min() { return SCHAR_MIN; };
__device__ __forceinline__ static type max() { return SCHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (signed char)-1 == -1;
};
template<> struct numeric_limits<unsigned char>
{
typedef unsigned char type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UCHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<short>
{
typedef short type;
__device__ __forceinline__ static type min() { return SHRT_MIN; };
__device__ __forceinline__ static type max() { return SHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned short>
{
typedef unsigned short type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return USHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<int>
{
typedef int type;
__device__ __forceinline__ static type min() { return INT_MIN; };
__device__ __forceinline__ static type max() { return INT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned int>
{
typedef unsigned int type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UINT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<long>
{
typedef long type;
__device__ __forceinline__ static type min() { return LONG_MIN; };
__device__ __forceinline__ static type max() { return LONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned long>
{
typedef unsigned long type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return ULONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<float>
{
typedef float type;
__device__ __forceinline__ static type min() { return 1.175494351e-38f/*FLT_MIN*/; };
__device__ __forceinline__ static type max() { return 3.402823466e+38f/*FLT_MAX*/; };
__device__ __forceinline__ static type epsilon() { return 1.192092896e-07f/*FLT_EPSILON*/; };
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<double>
{
typedef double type;
__device__ __forceinline__ static type min() { return 2.2250738585072014e-308/*DBL_MIN*/; };
__device__ __forceinline__ static type max() { return 1.7976931348623158e+308/*DBL_MAX*/; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
}}} // namespace cv { namespace gpu { namespace device {
#endif // __OPENCV_GPU_LIMITS_GPU_HPP__

432
modules/gpu/include/opencv2/gpu/device/saturate_cast.hpp
View File

@@ -1,216 +1,216 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_SATURATE_CAST_HPP__
#define __OPENCV_GPU_SATURATE_CAST_HPP__
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uchar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(schar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(ushort v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(short v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uint v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(int v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(float v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(double v) { return _Tp(v); }
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(schar v)
{
return (uchar) ::max((int)v, 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
{
return (uchar) ::min((uint)v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
{
return (uchar)((uint)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
{
return (uchar) ::min(v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
{
return saturate_cast<uchar>((uint)v);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<uchar>(iv);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<uchar>(iv);
#else
return saturate_cast<uchar>((float)v);
#endif
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uchar v)
{
return (schar) ::min((int)v, SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
{
return (schar) ::min((uint)v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
{
return (schar)((uint)(v-SCHAR_MIN) <= (uint)UCHAR_MAX ? v : v > 0 ? SCHAR_MAX : SCHAR_MIN);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
{
return saturate_cast<schar>((int)v);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
{
return (schar) ::min(v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<schar>(iv);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<schar>(iv);
#else
return saturate_cast<schar>((float)v);
#endif
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
{
return (ushort)((uint)v <= (uint)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
{
return (ushort) ::min(v, (uint)USHRT_MAX);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<ushort>(iv);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<ushort>(iv);
#else
return saturate_cast<ushort>((float)v);
#endif
}
template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
{
return (short) ::min((int)v, SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(int v)
{
return (short)((uint)(v - SHRT_MIN) <= (uint)USHRT_MAX ? v : v > 0 ? SHRT_MAX : SHRT_MIN);
}
template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
{
return (short) ::min(v, (uint)SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<short>(iv);
}
template<> __device__ __forceinline__ short saturate_cast<short>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<short>(iv);
#else
return saturate_cast<short>((float)v);
#endif
}
template<> __device__ __forceinline__ int saturate_cast<int>(float v)
{
return __float2int_rn(v);
}
template<> __device__ __forceinline__ int saturate_cast<int>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
return __double2int_rn(v);
#else
return saturate_cast<int>((float)v);
#endif
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
{
return __float2uint_rn(v);
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
return __double2uint_rn(v);
#else
return saturate_cast<uint>((float)v);
#endif
}
}}}
#endif /* __OPENCV_GPU_SATURATE_CAST_HPP__ */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_SATURATE_CAST_HPP__
#define __OPENCV_GPU_SATURATE_CAST_HPP__
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uchar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(schar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(ushort v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(short v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uint v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(int v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(float v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(double v) { return _Tp(v); }
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(schar v)
{
return (uchar) ::max((int)v, 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
{
return (uchar) ::min((uint)v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
{
return (uchar)((uint)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
{
return (uchar) ::min(v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
{
return saturate_cast<uchar>((uint)v);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<uchar>(iv);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<uchar>(iv);
#else
return saturate_cast<uchar>((float)v);
#endif
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uchar v)
{
return (schar) ::min((int)v, SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
{
return (schar) ::min((uint)v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
{
return (schar)((uint)(v-SCHAR_MIN) <= (uint)UCHAR_MAX ? v : v > 0 ? SCHAR_MAX : SCHAR_MIN);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
{
return saturate_cast<schar>((int)v);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
{
return (schar) ::min(v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<schar>(iv);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<schar>(iv);
#else
return saturate_cast<schar>((float)v);
#endif
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
{
return (ushort)((uint)v <= (uint)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
{
return (ushort) ::min(v, (uint)USHRT_MAX);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<ushort>(iv);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<ushort>(iv);
#else
return saturate_cast<ushort>((float)v);
#endif
}
template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
{
return (short) ::min((int)v, SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(int v)
{
return (short)((uint)(v - SHRT_MIN) <= (uint)USHRT_MAX ? v : v > 0 ? SHRT_MAX : SHRT_MIN);
}
template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
{
return (short) ::min(v, (uint)SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<short>(iv);
}
template<> __device__ __forceinline__ short saturate_cast<short>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<short>(iv);
#else
return saturate_cast<short>((float)v);
#endif
}
template<> __device__ __forceinline__ int saturate_cast<int>(float v)
{
return __float2int_rn(v);
}
template<> __device__ __forceinline__ int saturate_cast<int>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
return __double2int_rn(v);
#else
return saturate_cast<int>((float)v);
#endif
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
{
return __float2uint_rn(v);
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(double v)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
return __double2uint_rn(v);
#else
return saturate_cast<uint>((float)v);
#endif
}
}}}
#endif /* __OPENCV_GPU_SATURATE_CAST_HPP__ */

134
modules/gpu/include/opencv2/gpu/device/static_check.hpp
View File

@@ -1,67 +1,67 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_GPU_DEVICE_STATIC_CHECK_HPP__
#define __OPENCV_GPU_GPU_DEVICE_STATIC_CHECK_HPP__
#if defined(__CUDACC__)
#define __OPENCV_GPU_HOST_DEVICE__ __host__ __device__ __forceinline__
#else
#define __OPENCV_GPU_HOST_DEVICE__
#endif
namespace cv { namespace gpu
{
namespace device
{
template<bool expr> struct Static {};
template<> struct Static<true>
{
__OPENCV_GPU_HOST_DEVICE__ static void check() {};
};
}
}}
#undef __OPENCV_GPU_HOST_DEVICE__
#endif /* __OPENCV_GPU_GPU_DEVICE_STATIC_CHECK_HPP__ */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_GPU_DEVICE_STATIC_CHECK_HPP__
#define __OPENCV_GPU_GPU_DEVICE_STATIC_CHECK_HPP__
#if defined(__CUDACC__)
#define __OPENCV_GPU_HOST_DEVICE__ __host__ __device__ __forceinline__
#else
#define __OPENCV_GPU_HOST_DEVICE__
#endif
namespace cv { namespace gpu
{
namespace device
{
template<bool expr> struct Static {};
template<> struct Static<true>
{
__OPENCV_GPU_HOST_DEVICE__ static void check() {};
};
}
}}
#undef __OPENCV_GPU_HOST_DEVICE__
#endif /* __OPENCV_GPU_GPU_DEVICE_STATIC_CHECK_HPP__ */

134
modules/gpu/include/opencv2/gpu/device/transform.hpp
View File

@@ -1,67 +1,67 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TRANSFORM_HPP__
#define __OPENCV_GPU_TRANSFORM_HPP__
#include "common.hpp"
#include "utility.hpp"
#include "detail/transform_detail.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename T, typename D, typename UnOp, typename Mask>
static inline void transform(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
transform_detail::TransformDispatcher<VecTraits<T>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src, dst, op, mask, stream);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static inline void transform(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
transform_detail::TransformDispatcher<VecTraits<T1>::cn == 1 && VecTraits<T2>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src1, src2, dst, op, mask, stream);
}
}}}
#endif // __OPENCV_GPU_TRANSFORM_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TRANSFORM_HPP__
#define __OPENCV_GPU_TRANSFORM_HPP__
#include "common.hpp"
#include "utility.hpp"
#include "detail/transform_detail.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename T, typename D, typename UnOp, typename Mask>
static inline void transform(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
transform_detail::TransformDispatcher<VecTraits<T>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src, dst, op, mask, stream);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static inline void transform(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
transform_detail::TransformDispatcher<VecTraits<T1>::cn == 1 && VecTraits<T2>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src1, src2, dst, op, mask, stream);
}
}}}
#endif // __OPENCV_GPU_TRANSFORM_HPP__

164
modules/gpu/include/opencv2/gpu/device/type_traits.hpp
View File

@@ -1,82 +1,82 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TYPE_TRAITS_HPP__
#define __OPENCV_GPU_TYPE_TRAITS_HPP__
#include "detail/type_traits_detail.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename T> struct IsSimpleParameter
{
enum {value = type_traits_detail::IsIntegral<T>::value || type_traits_detail::IsFloat<T>::value ||
type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<T>::type>::value};
};
template <typename T> struct TypeTraits
{
typedef typename type_traits_detail::UnConst<T>::type NonConstType;
typedef typename type_traits_detail::UnVolatile<T>::type NonVolatileType;
typedef typename type_traits_detail::UnVolatile<typename type_traits_detail::UnConst<T>::type>::type UnqualifiedType;
typedef typename type_traits_detail::PointerTraits<UnqualifiedType>::type PointeeType;
typedef typename type_traits_detail::ReferenceTraits<T>::type ReferredType;
enum { isConst = type_traits_detail::UnConst<T>::value };
enum { isVolatile = type_traits_detail::UnVolatile<T>::value };
enum { isReference = type_traits_detail::ReferenceTraits<UnqualifiedType>::value };
enum { isPointer = type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<UnqualifiedType>::type>::value };
enum { isUnsignedInt = type_traits_detail::IsUnsignedIntegral<UnqualifiedType>::value };
enum { isSignedInt = type_traits_detail::IsSignedIntergral<UnqualifiedType>::value };
enum { isIntegral = type_traits_detail::IsIntegral<UnqualifiedType>::value };
enum { isFloat = type_traits_detail::IsFloat<UnqualifiedType>::value };
enum { isArith = isIntegral || isFloat };
enum { isVec = type_traits_detail::IsVec<UnqualifiedType>::value };
typedef typename type_traits_detail::Select<IsSimpleParameter<UnqualifiedType>::value,
T, typename type_traits_detail::AddParameterType<T>::type>::type ParameterType;
};
}}}
#endif // __OPENCV_GPU_TYPE_TRAITS_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_TYPE_TRAITS_HPP__
#define __OPENCV_GPU_TYPE_TRAITS_HPP__
#include "detail/type_traits_detail.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename T> struct IsSimpleParameter
{
enum {value = type_traits_detail::IsIntegral<T>::value || type_traits_detail::IsFloat<T>::value ||
type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<T>::type>::value};
};
template <typename T> struct TypeTraits
{
typedef typename type_traits_detail::UnConst<T>::type NonConstType;
typedef typename type_traits_detail::UnVolatile<T>::type NonVolatileType;
typedef typename type_traits_detail::UnVolatile<typename type_traits_detail::UnConst<T>::type>::type UnqualifiedType;
typedef typename type_traits_detail::PointerTraits<UnqualifiedType>::type PointeeType;
typedef typename type_traits_detail::ReferenceTraits<T>::type ReferredType;
enum { isConst = type_traits_detail::UnConst<T>::value };
enum { isVolatile = type_traits_detail::UnVolatile<T>::value };
enum { isReference = type_traits_detail::ReferenceTraits<UnqualifiedType>::value };
enum { isPointer = type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<UnqualifiedType>::type>::value };
enum { isUnsignedInt = type_traits_detail::IsUnsignedIntegral<UnqualifiedType>::value };
enum { isSignedInt = type_traits_detail::IsSignedIntergral<UnqualifiedType>::value };
enum { isIntegral = type_traits_detail::IsIntegral<UnqualifiedType>::value };
enum { isFloat = type_traits_detail::IsFloat<UnqualifiedType>::value };
enum { isArith = isIntegral || isFloat };
enum { isVec = type_traits_detail::IsVec<UnqualifiedType>::value };
typedef typename type_traits_detail::Select<IsSimpleParameter<UnqualifiedType>::value,
T, typename type_traits_detail::AddParameterType<T>::type>::type ParameterType;
};
}}}
#endif // __OPENCV_GPU_TYPE_TRAITS_HPP__

474
modules/gpu/include/opencv2/gpu/device/utility.hpp
View File

@@ -1,237 +1,237 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_UTILITY_HPP__
#define __OPENCV_GPU_UTILITY_HPP__
#include "saturate_cast.hpp"
#include "datamov_utils.hpp"
#include "detail/reduction_detail.hpp"
namespace cv { namespace gpu { namespace device
{
#define OPENCV_GPU_LOG_WARP_SIZE (5)
#define OPENCV_GPU_WARP_SIZE (1 << OPENCV_GPU_LOG_WARP_SIZE)
#define OPENCV_GPU_LOG_MEM_BANKS ((__CUDA_ARCH__ >= 200) ? 5 : 4) // 32 banks on fermi, 16 on tesla
#define OPENCV_GPU_MEM_BANKS (1 << OPENCV_GPU_LOG_MEM_BANKS)
///////////////////////////////////////////////////////////////////////////////
// swap
template <typename T> void __device__ __host__ __forceinline__ swap(T& a, T& b)
{
const T temp = a;
a = b;
b = temp;
}
///////////////////////////////////////////////////////////////////////////////
// Mask Reader
struct SingleMask
{
explicit __host__ __device__ __forceinline__ SingleMask(PtrStepb mask_) : mask(mask_) {}
__host__ __device__ __forceinline__ SingleMask(const SingleMask& mask_): mask(mask_.mask){}
__device__ __forceinline__ bool operator()(int y, int x) const
{
return mask.ptr(y)[x] != 0;
}
PtrStepb mask;
};
struct SingleMaskChannels
{
__host__ __device__ __forceinline__ SingleMaskChannels(PtrStepb mask_, int channels_)
: mask(mask_), channels(channels_) {}
__host__ __device__ __forceinline__ SingleMaskChannels(const SingleMaskChannels& mask_)
:mask(mask_.mask), channels(mask_.channels){}
__device__ __forceinline__ bool operator()(int y, int x) const
{
return mask.ptr(y)[x / channels] != 0;
}
PtrStepb mask;
int channels;
};
struct MaskCollection
{
explicit __host__ __device__ __forceinline__ MaskCollection(PtrStepb* maskCollection_)
: maskCollection(maskCollection_) {}
__device__ __forceinline__ MaskCollection(const MaskCollection& masks_)
: maskCollection(masks_.maskCollection), curMask(masks_.curMask){}
__device__ __forceinline__ void next()
{
curMask = *maskCollection++;
}
__device__ __forceinline__ void setMask(int z)
{
curMask = maskCollection[z];
}
__device__ __forceinline__ bool operator()(int y, int x) const
{
uchar val;
return curMask.data == 0 || (ForceGlob<uchar>::Load(curMask.ptr(y), x, val), (val != 0));
}
const PtrStepb* maskCollection;
PtrStepb curMask;
};
struct WithOutMask
{
__device__ __forceinline__ WithOutMask(){}
__device__ __forceinline__ WithOutMask(const WithOutMask& mask){}
__device__ __forceinline__ void next() const
{
}
__device__ __forceinline__ void setMask(int) const
{
}
__device__ __forceinline__ bool operator()(int, int) const
{
return true;
}
__device__ __forceinline__ bool operator()(int, int, int) const
{
return true;
}
static __device__ __forceinline__ bool check(int, int)
{
return true;
}
static __device__ __forceinline__ bool check(int, int, int, uint offset = 0)
{
return true;
}
};
///////////////////////////////////////////////////////////////////////////////
// Reduction
template <int n, typename T, typename Op> __device__ __forceinline__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::ReductionDispatcher<n <= 64>::reduce<n>(data, partial_reduction, tid, op);
}
template <int n, typename T, typename V, typename Pred>
__device__ __forceinline__ void reducePredVal(volatile T* sdata, T& myData, V* sval, V& myVal, int tid, const Pred& pred)
{
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::PredValReductionDispatcher<n <= 64>::reduce<n>(myData, myVal, sdata, sval, tid, pred);
}
template <int n, typename T, typename V1, typename V2, typename Pred>
__device__ __forceinline__ void reducePredVal2(volatile T* sdata, T& myData, V1* sval1, V1& myVal1, V2* sval2, V2& myVal2, int tid, const Pred& pred)
{
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::PredVal2ReductionDispatcher<n <= 64>::reduce<n>(myData, myVal1, myVal2, sdata, sval1, sval2, tid, pred);
}
///////////////////////////////////////////////////////////////////////////////
// Solve linear system
// solve 2x2 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve2x2(const T A[2][2], const T b[2], T x[2])
{
T det = A[0][0] * A[1][1] - A[1][0] * A[0][1];
if (det != 0)
{
double invdet = 1.0 / det;
x[0] = saturate_cast<T>(invdet * (b[0] * A[1][1] - b[1] * A[0][1]));
x[1] = saturate_cast<T>(invdet * (A[0][0] * b[1] - A[1][0] * b[0]));
return true;
}
return false;
}
// solve 3x3 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve3x3(const T A[3][3], const T b[3], T x[3])
{
T det = A[0][0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1])
- A[0][1] * (A[1][0] * A[2][2] - A[1][2] * A[2][0])
+ A[0][2] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]);
if (det != 0)
{
double invdet = 1.0 / det;
x[0] = saturate_cast<T>(invdet *
(b[0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1]) -
A[0][1] * (b[1] * A[2][2] - A[1][2] * b[2] ) +
A[0][2] * (b[1] * A[2][1] - A[1][1] * b[2] )));
x[1] = saturate_cast<T>(invdet *
(A[0][0] * (b[1] * A[2][2] - A[1][2] * b[2] ) -
b[0] * (A[1][0] * A[2][2] - A[1][2] * A[2][0]) +
A[0][2] * (A[1][0] * b[2] - b[1] * A[2][0])));
x[2] = saturate_cast<T>(invdet *
(A[0][0] * (A[1][1] * b[2] - b[1] * A[2][1]) -
A[0][1] * (A[1][0] * b[2] - b[1] * A[2][0]) +
b[0] * (A[1][0] * A[2][1] - A[1][1] * A[2][0])));
return true;
}
return false;
}
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_UTILITY_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_UTILITY_HPP__
#define __OPENCV_GPU_UTILITY_HPP__
#include "saturate_cast.hpp"
#include "datamov_utils.hpp"
#include "detail/reduction_detail.hpp"
namespace cv { namespace gpu { namespace device
{
#define OPENCV_GPU_LOG_WARP_SIZE (5)
#define OPENCV_GPU_WARP_SIZE (1 << OPENCV_GPU_LOG_WARP_SIZE)
#define OPENCV_GPU_LOG_MEM_BANKS ((__CUDA_ARCH__ >= 200) ? 5 : 4) // 32 banks on fermi, 16 on tesla
#define OPENCV_GPU_MEM_BANKS (1 << OPENCV_GPU_LOG_MEM_BANKS)
///////////////////////////////////////////////////////////////////////////////
// swap
template <typename T> void __device__ __host__ __forceinline__ swap(T& a, T& b)
{
const T temp = a;
a = b;
b = temp;
}
///////////////////////////////////////////////////////////////////////////////
// Mask Reader
struct SingleMask
{
explicit __host__ __device__ __forceinline__ SingleMask(PtrStepb mask_) : mask(mask_) {}
__host__ __device__ __forceinline__ SingleMask(const SingleMask& mask_): mask(mask_.mask){}
__device__ __forceinline__ bool operator()(int y, int x) const
{
return mask.ptr(y)[x] != 0;
}
PtrStepb mask;
};
struct SingleMaskChannels
{
__host__ __device__ __forceinline__ SingleMaskChannels(PtrStepb mask_, int channels_)
: mask(mask_), channels(channels_) {}
__host__ __device__ __forceinline__ SingleMaskChannels(const SingleMaskChannels& mask_)
:mask(mask_.mask), channels(mask_.channels){}
__device__ __forceinline__ bool operator()(int y, int x) const
{
return mask.ptr(y)[x / channels] != 0;
}
PtrStepb mask;
int channels;
};
struct MaskCollection
{
explicit __host__ __device__ __forceinline__ MaskCollection(PtrStepb* maskCollection_)
: maskCollection(maskCollection_) {}
__device__ __forceinline__ MaskCollection(const MaskCollection& masks_)
: maskCollection(masks_.maskCollection), curMask(masks_.curMask){}
__device__ __forceinline__ void next()
{
curMask = *maskCollection++;
}
__device__ __forceinline__ void setMask(int z)
{
curMask = maskCollection[z];
}
__device__ __forceinline__ bool operator()(int y, int x) const
{
uchar val;
return curMask.data == 0 || (ForceGlob<uchar>::Load(curMask.ptr(y), x, val), (val != 0));
}
const PtrStepb* maskCollection;
PtrStepb curMask;
};
struct WithOutMask
{
__device__ __forceinline__ WithOutMask(){}
__device__ __forceinline__ WithOutMask(const WithOutMask& mask){}
__device__ __forceinline__ void next() const
{
}
__device__ __forceinline__ void setMask(int) const
{
}
__device__ __forceinline__ bool operator()(int, int) const
{
return true;
}
__device__ __forceinline__ bool operator()(int, int, int) const
{
return true;
}
static __device__ __forceinline__ bool check(int, int)
{
return true;
}
static __device__ __forceinline__ bool check(int, int, int, uint offset = 0)
{
return true;
}
};
///////////////////////////////////////////////////////////////////////////////
// Reduction
template <int n, typename T, typename Op> __device__ __forceinline__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::ReductionDispatcher<n <= 64>::reduce<n>(data, partial_reduction, tid, op);
}
template <int n, typename T, typename V, typename Pred>
__device__ __forceinline__ void reducePredVal(volatile T* sdata, T& myData, V* sval, V& myVal, int tid, const Pred& pred)
{
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::PredValReductionDispatcher<n <= 64>::reduce<n>(myData, myVal, sdata, sval, tid, pred);
}
template <int n, typename T, typename V1, typename V2, typename Pred>
__device__ __forceinline__ void reducePredVal2(volatile T* sdata, T& myData, V1* sval1, V1& myVal1, V2* sval2, V2& myVal2, int tid, const Pred& pred)
{
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::PredVal2ReductionDispatcher<n <= 64>::reduce<n>(myData, myVal1, myVal2, sdata, sval1, sval2, tid, pred);
}
///////////////////////////////////////////////////////////////////////////////
// Solve linear system
// solve 2x2 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve2x2(const T A[2][2], const T b[2], T x[2])
{
T det = A[0][0] * A[1][1] - A[1][0] * A[0][1];
if (det != 0)
{
double invdet = 1.0 / det;
x[0] = saturate_cast<T>(invdet * (b[0] * A[1][1] - b[1] * A[0][1]));
x[1] = saturate_cast<T>(invdet * (A[0][0] * b[1] - A[1][0] * b[0]));
return true;
}
return false;
}
// solve 3x3 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve3x3(const T A[3][3], const T b[3], T x[3])
{
T det = A[0][0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1])
- A[0][1] * (A[1][0] * A[2][2] - A[1][2] * A[2][0])
+ A[0][2] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]);
if (det != 0)
{
double invdet = 1.0 / det;
x[0] = saturate_cast<T>(invdet *
(b[0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1]) -
A[0][1] * (b[1] * A[2][2] - A[1][2] * b[2] ) +
A[0][2] * (b[1] * A[2][1] - A[1][1] * b[2] )));
x[1] = saturate_cast<T>(invdet *
(A[0][0] * (b[1] * A[2][2] - A[1][2] * b[2] ) -
b[0] * (A[1][0] * A[2][2] - A[1][2] * A[2][0]) +
A[0][2] * (A[1][0] * b[2] - b[1] * A[2][0])));
x[2] = saturate_cast<T>(invdet *
(A[0][0] * (A[1][1] * b[2] - b[1] * A[2][1]) -
A[0][1] * (A[1][0] * b[2] - b[1] * A[2][0]) +
b[0] * (A[1][0] * A[2][1] - A[1][1] * A[2][0])));
return true;
}
return false;
}
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_UTILITY_HPP__

448
modules/gpu/include/opencv2/gpu/device/vec_distance.hpp
View File

@@ -1,224 +1,224 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VEC_DISTANCE_HPP__
#define __OPENCV_GPU_VEC_DISTANCE_HPP__
#include "utility.hpp"
#include "functional.hpp"
#include "detail/vec_distance_detail.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename T> struct L1Dist
{
typedef int value_type;
typedef int result_type;
__device__ __forceinline__ L1Dist() : mySum(0) {}
__device__ __forceinline__ void reduceIter(int val1, int val2)
{
mySum = __sad(val1, val2, mySum);
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
__device__ __forceinline__ operator int() const
{
return mySum;
}
int mySum;
};
template <> struct L1Dist<float>
{
typedef float value_type;
typedef float result_type;
__device__ __forceinline__ L1Dist() : mySum(0.0f) {}
__device__ __forceinline__ void reduceIter(float val1, float val2)
{
mySum += ::fabs(val1 - val2);
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
__device__ __forceinline__ operator float() const
{
return mySum;
}
float mySum;
};
struct L2Dist
{
typedef float value_type;
typedef float result_type;
__device__ __forceinline__ L2Dist() : mySum(0.0f) {}
__device__ __forceinline__ void reduceIter(float val1, float val2)
{
float reg = val1 - val2;
mySum += reg * reg;
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
__device__ __forceinline__ operator float() const
{
return sqrtf(mySum);
}
float mySum;
};
struct HammingDist
{
typedef int value_type;
typedef int result_type;
__device__ __forceinline__ HammingDist() : mySum(0) {}
__device__ __forceinline__ void reduceIter(int val1, int val2)
{
mySum += __popc(val1 ^ val2);
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
__device__ __forceinline__ operator int() const
{
return mySum;
}
int mySum;
};
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename Dist, typename T1, typename T2>
__device__ void calcVecDiffGlobal(const T1* vec1, const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
for (int i = tid; i < len; i += THREAD_DIM)
{
T1 val1;
ForceGlob<T1>::Load(vec1, i, val1);
T2 val2;
ForceGlob<T2>::Load(vec2, i, val2);
dist.reduceIter(val1, val2);
}
dist.reduceAll<THREAD_DIM>(smem, tid);
}
// calc distance between two vectors, first vector is cached in register or shared memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename Dist, typename T1, typename T2>
__device__ __forceinline__ void calcVecDiffCached(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
vec_distance_detail::VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>::calc(vecCached, vecGlob, len, dist, tid);
dist.reduceAll<THREAD_DIM>(smem, tid);
}
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename T1> struct VecDiffGlobal
{
explicit __device__ __forceinline__ VecDiffGlobal(const T1* vec1_, int = 0, void* = 0, int = 0, int = 0)
{
vec1 = vec1_;
}
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
{
calcVecDiffGlobal<THREAD_DIM>(vec1, vec2, len, dist, smem, tid);
}
const T1* vec1;
};
// calc distance between two vectors, first vector is cached in register memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename U> struct VecDiffCachedRegister
{
template <typename T1> __device__ __forceinline__ VecDiffCachedRegister(const T1* vec1, int len, U* smem, int glob_tid, int tid)
{
if (glob_tid < len)
smem[glob_tid] = vec1[glob_tid];
__syncthreads();
U* vec1ValsPtr = vec1Vals;
#pragma unroll
for (int i = tid; i < MAX_LEN; i += THREAD_DIM)
*vec1ValsPtr++ = smem[i];
__syncthreads();
}
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
{
calcVecDiffCached<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>(vec1Vals, vec2, len, dist, smem, tid);
}
U vec1Vals[MAX_LEN / THREAD_DIM];
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_DISTANCE_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VEC_DISTANCE_HPP__
#define __OPENCV_GPU_VEC_DISTANCE_HPP__
#include "utility.hpp"
#include "functional.hpp"
#include "detail/vec_distance_detail.hpp"
namespace cv { namespace gpu { namespace device
{
template <typename T> struct L1Dist
{
typedef int value_type;
typedef int result_type;
__device__ __forceinline__ L1Dist() : mySum(0) {}
__device__ __forceinline__ void reduceIter(int val1, int val2)
{
mySum = __sad(val1, val2, mySum);
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
__device__ __forceinline__ operator int() const
{
return mySum;
}
int mySum;
};
template <> struct L1Dist<float>
{
typedef float value_type;
typedef float result_type;
__device__ __forceinline__ L1Dist() : mySum(0.0f) {}
__device__ __forceinline__ void reduceIter(float val1, float val2)
{
mySum += ::fabs(val1 - val2);
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
__device__ __forceinline__ operator float() const
{
return mySum;
}
float mySum;
};
struct L2Dist
{
typedef float value_type;
typedef float result_type;
__device__ __forceinline__ L2Dist() : mySum(0.0f) {}
__device__ __forceinline__ void reduceIter(float val1, float val2)
{
float reg = val1 - val2;
mySum += reg * reg;
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
__device__ __forceinline__ operator float() const
{
return sqrtf(mySum);
}
float mySum;
};
struct HammingDist
{
typedef int value_type;
typedef int result_type;
__device__ __forceinline__ HammingDist() : mySum(0) {}
__device__ __forceinline__ void reduceIter(int val1, int val2)
{
mySum += __popc(val1 ^ val2);
}
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
__device__ __forceinline__ operator int() const
{
return mySum;
}
int mySum;
};
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename Dist, typename T1, typename T2>
__device__ void calcVecDiffGlobal(const T1* vec1, const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
for (int i = tid; i < len; i += THREAD_DIM)
{
T1 val1;
ForceGlob<T1>::Load(vec1, i, val1);
T2 val2;
ForceGlob<T2>::Load(vec2, i, val2);
dist.reduceIter(val1, val2);
}
dist.reduceAll<THREAD_DIM>(smem, tid);
}
// calc distance between two vectors, first vector is cached in register or shared memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename Dist, typename T1, typename T2>
__device__ __forceinline__ void calcVecDiffCached(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
vec_distance_detail::VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>::calc(vecCached, vecGlob, len, dist, tid);
dist.reduceAll<THREAD_DIM>(smem, tid);
}
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename T1> struct VecDiffGlobal
{
explicit __device__ __forceinline__ VecDiffGlobal(const T1* vec1_, int = 0, void* = 0, int = 0, int = 0)
{
vec1 = vec1_;
}
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
{
calcVecDiffGlobal<THREAD_DIM>(vec1, vec2, len, dist, smem, tid);
}
const T1* vec1;
};
// calc distance between two vectors, first vector is cached in register memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename U> struct VecDiffCachedRegister
{
template <typename T1> __device__ __forceinline__ VecDiffCachedRegister(const T1* vec1, int len, U* smem, int glob_tid, int tid)
{
if (glob_tid < len)
smem[glob_tid] = vec1[glob_tid];
__syncthreads();
U* vec1ValsPtr = vec1Vals;
#pragma unroll
for (int i = tid; i < MAX_LEN; i += THREAD_DIM)
*vec1ValsPtr++ = smem[i];
__syncthreads();
}
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
{
calcVecDiffCached<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>(vec1Vals, vec2, len, dist, smem, tid);
}
U vec1Vals[MAX_LEN / THREAD_DIM];
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_DISTANCE_HPP__

658
modules/gpu/include/opencv2/gpu/device/vec_math.hpp
View File

@@ -1,330 +1,330 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VECMATH_HPP__
#define __OPENCV_GPU_VECMATH_HPP__
#include "saturate_cast.hpp"
#include "vec_traits.hpp"
#include "functional.hpp"
namespace cv { namespace gpu { namespace device
{
namespace vec_math_detail
{
template <int cn, typename VecD> struct SatCastHelper;
template <typename VecD> struct SatCastHelper<1, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x));
}
};
template <typename VecD> struct SatCastHelper<2, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y));
}
};
template <typename VecD> struct SatCastHelper<3, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z));
}
};
template <typename VecD> struct SatCastHelper<4, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z), saturate_cast<D>(v.w));
}
};
template <typename VecD, typename VecS> static __device__ __forceinline__ VecD saturate_cast_caller(const VecS& v)
{
return SatCastHelper<VecTraits<VecD>::cn, VecD>::cast(v);
}
}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
#define OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, op, func) \
__device__ __forceinline__ TypeVec<func<type>::result_type, 1>::vec_type op(const type ## 1 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 1>::vec_type>::make(f(a.x)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 2>::vec_type op(const type ## 2 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 2>::vec_type>::make(f(a.x), f(a.y)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 3>::vec_type op(const type ## 3 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 3>::vec_type>::make(f(a.x), f(a.y), f(a.z)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 4>::vec_type op(const type ## 4 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 4>::vec_type>::make(f(a.x), f(a.y), f(a.z), f(a.w)); \
}
namespace vec_math_detail
{
template <typename T1, typename T2> struct BinOpTraits
{
typedef int argument_type;
};
template <typename T> struct BinOpTraits<T, T>
{
typedef T argument_type;
};
template <typename T> struct BinOpTraits<T, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<double, T>
{
typedef double argument_type;
};
template <> struct BinOpTraits<double, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<T, float>
{
typedef float argument_type;
};
template <typename T> struct BinOpTraits<float, T>
{
typedef float argument_type;
};
template <> struct BinOpTraits<float, float>
{
typedef float argument_type;
};
template <> struct BinOpTraits<double, float>
{
typedef double argument_type;
};
template <> struct BinOpTraits<float, double>
{
typedef double argument_type;
};
}
#define OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, op, func) \
__device__ __forceinline__ TypeVec<func<type>::result_type, 1>::vec_type op(const type ## 1 & a, const type ## 1 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 1>::vec_type>::make(f(a.x, b.x)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(const type ## 1 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(v.x, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(T s, const type ## 1 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(s, v.x)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 2>::vec_type op(const type ## 2 & a, const type ## 2 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 2>::vec_type>::make(f(a.x, b.x), f(a.y, b.y)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(const type ## 2 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(v.x, s), f(v.y, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(T s, const type ## 2 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(s, v.x), f(s, v.y)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 3>::vec_type op(const type ## 3 & a, const type ## 3 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 3>::vec_type>::make(f(a.x, b.x), f(a.y, b.y), f(a.z, b.z)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(const type ## 3 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(T s, const type ## 3 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 4>::vec_type op(const type ## 4 & a, const type ## 4 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 4>::vec_type>::make(f(a.x, b.x), f(a.y, b.y), f(a.z, b.z), f(a.w, b.w)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(const type ## 4 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s), f(v.w, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(T s, const type ## 4 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<T, type>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z), f(s, v.w)); \
}
#define OPENCV_GPU_IMPLEMENT_VEC_OP(type) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator +, plus) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator -, minus) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator *, multiplies) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator /, divides) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP (type, operator -, negate) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator ==, equal_to) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator !=, not_equal_to) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator > , greater) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator < , less) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator >=, greater_equal) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator <=, less_equal) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator &&, logical_and) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator ||, logical_or) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP (type, operator ! , logical_not) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, max, maximum) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, min, minimum) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, fabs, fabs_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, sqrt, sqrt_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, exp, exp_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, exp2, exp2_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, exp10, exp10_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, log, log_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, log2, log2_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, log10, log10_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, sin, sin_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, cos, cos_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, tan, tan_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, asin, asin_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, acos, acos_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, atan, atan_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, sinh, sinh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, cosh, cosh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, tanh, tanh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, asinh, asinh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, acosh, acosh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, atanh, atanh_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, hypot, hypot_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, atan2, atan2_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, pow, pow_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, hypot_sqr, hypot_sqr_func)
#define OPENCV_GPU_IMPLEMENT_VEC_INT_OP(type) \
OPENCV_GPU_IMPLEMENT_VEC_OP(type) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator &, bit_and) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator |, bit_or) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator ^, bit_xor) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP (type, operator ~, bit_not)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uchar)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(char)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(ushort)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(short)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(int)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uint)
OPENCV_GPU_IMPLEMENT_VEC_OP(float)
OPENCV_GPU_IMPLEMENT_VEC_OP(double)
#undef OPENCV_GPU_IMPLEMENT_VEC_UNOP
#undef OPENCV_GPU_IMPLEMENT_VEC_BINOP
#undef OPENCV_GPU_IMPLEMENT_VEC_OP
#undef OPENCV_GPU_IMPLEMENT_VEC_INT_OP
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VECMATH_HPP__
#define __OPENCV_GPU_VECMATH_HPP__
#include "saturate_cast.hpp"
#include "vec_traits.hpp"
#include "functional.hpp"
namespace cv { namespace gpu { namespace device
{
namespace vec_math_detail
{
template <int cn, typename VecD> struct SatCastHelper;
template <typename VecD> struct SatCastHelper<1, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x));
}
};
template <typename VecD> struct SatCastHelper<2, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y));
}
};
template <typename VecD> struct SatCastHelper<3, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z));
}
};
template <typename VecD> struct SatCastHelper<4, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z), saturate_cast<D>(v.w));
}
};
template <typename VecD, typename VecS> static __device__ __forceinline__ VecD saturate_cast_caller(const VecS& v)
{
return SatCastHelper<VecTraits<VecD>::cn, VecD>::cast(v);
}
}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
#define OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, op, func) \
__device__ __forceinline__ TypeVec<func<type>::result_type, 1>::vec_type op(const type ## 1 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 1>::vec_type>::make(f(a.x)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 2>::vec_type op(const type ## 2 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 2>::vec_type>::make(f(a.x), f(a.y)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 3>::vec_type op(const type ## 3 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 3>::vec_type>::make(f(a.x), f(a.y), f(a.z)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 4>::vec_type op(const type ## 4 & a) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 4>::vec_type>::make(f(a.x), f(a.y), f(a.z), f(a.w)); \
}
namespace vec_math_detail
{
template <typename T1, typename T2> struct BinOpTraits
{
typedef int argument_type;
};
template <typename T> struct BinOpTraits<T, T>
{
typedef T argument_type;
};
template <typename T> struct BinOpTraits<T, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<double, T>
{
typedef double argument_type;
};
template <> struct BinOpTraits<double, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<T, float>
{
typedef float argument_type;
};
template <typename T> struct BinOpTraits<float, T>
{
typedef float argument_type;
};
template <> struct BinOpTraits<float, float>
{
typedef float argument_type;
};
template <> struct BinOpTraits<double, float>
{
typedef double argument_type;
};
template <> struct BinOpTraits<float, double>
{
typedef double argument_type;
};
}
#define OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, op, func) \
__device__ __forceinline__ TypeVec<func<type>::result_type, 1>::vec_type op(const type ## 1 & a, const type ## 1 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 1>::vec_type>::make(f(a.x, b.x)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(const type ## 1 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(v.x, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(T s, const type ## 1 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(s, v.x)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 2>::vec_type op(const type ## 2 & a, const type ## 2 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 2>::vec_type>::make(f(a.x, b.x), f(a.y, b.y)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(const type ## 2 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(v.x, s), f(v.y, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(T s, const type ## 2 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(s, v.x), f(s, v.y)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 3>::vec_type op(const type ## 3 & a, const type ## 3 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 3>::vec_type>::make(f(a.x, b.x), f(a.y, b.y), f(a.z, b.z)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(const type ## 3 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(T s, const type ## 3 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 4>::vec_type op(const type ## 4 & a, const type ## 4 & b) \
{ \
func<type> f; \
return VecTraits<TypeVec<func<type>::result_type, 4>::vec_type>::make(f(a.x, b.x), f(a.y, b.y), f(a.z, b.z), f(a.w, b.w)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(const type ## 4 & v, T s) \
{ \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s), f(v.w, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(T s, const type ## 4 & v) \
{ \
func<typename vec_math_detail::BinOpTraits<T, type>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z), f(s, v.w)); \
}
#define OPENCV_GPU_IMPLEMENT_VEC_OP(type) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator +, plus) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator -, minus) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator *, multiplies) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator /, divides) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP (type, operator -, negate) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator ==, equal_to) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator !=, not_equal_to) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator > , greater) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator < , less) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator >=, greater_equal) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator <=, less_equal) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator &&, logical_and) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator ||, logical_or) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP (type, operator ! , logical_not) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, max, maximum) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, min, minimum) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, fabs, fabs_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, sqrt, sqrt_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, exp, exp_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, exp2, exp2_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, exp10, exp10_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, log, log_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, log2, log2_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, log10, log10_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, sin, sin_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, cos, cos_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, tan, tan_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, asin, asin_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, acos, acos_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, atan, atan_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, sinh, sinh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, cosh, cosh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, tanh, tanh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, asinh, asinh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, acosh, acosh_func) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, atanh, atanh_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, hypot, hypot_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, atan2, atan2_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, pow, pow_func) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, hypot_sqr, hypot_sqr_func)
#define OPENCV_GPU_IMPLEMENT_VEC_INT_OP(type) \
OPENCV_GPU_IMPLEMENT_VEC_OP(type) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator &, bit_and) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator |, bit_or) \
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator ^, bit_xor) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP (type, operator ~, bit_not)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uchar)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(char)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(ushort)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(short)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(int)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uint)
OPENCV_GPU_IMPLEMENT_VEC_OP(float)
OPENCV_GPU_IMPLEMENT_VEC_OP(double)
#undef OPENCV_GPU_IMPLEMENT_VEC_UNOP
#undef OPENCV_GPU_IMPLEMENT_VEC_BINOP
#undef OPENCV_GPU_IMPLEMENT_VEC_OP
#undef OPENCV_GPU_IMPLEMENT_VEC_INT_OP
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VECMATH_HPP__

560
modules/gpu/include/opencv2/gpu/device/vec_traits.hpp
View File

@@ -1,280 +1,280 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VEC_TRAITS_HPP__
#define __OPENCV_GPU_VEC_TRAITS_HPP__
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
template<typename T, int N> struct TypeVec;
struct __align__(8) uchar8
{
uchar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uchar8 make_uchar8(uchar a0, uchar a1, uchar a2, uchar a3, uchar a4, uchar a5, uchar a6, uchar a7)
{
uchar8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(8) char8
{
schar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ char8 make_char8(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7)
{
char8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) ushort8
{
ushort a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ ushort8 make_ushort8(ushort a0, ushort a1, ushort a2, ushort a3, ushort a4, ushort a5, ushort a6, ushort a7)
{
ushort8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) short8
{
short a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ short8 make_short8(short a0, short a1, short a2, short a3, short a4, short a5, short a6, short a7)
{
short8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) uint8
{
uint a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uint8 make_uint8(uint a0, uint a1, uint a2, uint a3, uint a4, uint a5, uint a6, uint a7)
{
uint8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) int8
{
int a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ int8 make_int8(int a0, int a1, int a2, int a3, int a4, int a5, int a6, int a7)
{
int8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) float8
{
float a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ float8 make_float8(float a0, float a1, float a2, float a3, float a4, float a5, float a6, float a7)
{
float8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct double8
{
double a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ double8 make_double8(double a0, double a1, double a2, double a3, double a4, double a5, double a6, double a7)
{
double8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
#define OPENCV_GPU_IMPLEMENT_TYPE_VEC(type) \
template<> struct TypeVec<type, 1> { typedef type vec_type; }; \
template<> struct TypeVec<type ## 1, 1> { typedef type ## 1 vec_type; }; \
template<> struct TypeVec<type, 2> { typedef type ## 2 vec_type; }; \
template<> struct TypeVec<type ## 2, 2> { typedef type ## 2 vec_type; }; \
template<> struct TypeVec<type, 3> { typedef type ## 3 vec_type; }; \
template<> struct TypeVec<type ## 3, 3> { typedef type ## 3 vec_type; }; \
template<> struct TypeVec<type, 4> { typedef type ## 4 vec_type; }; \
template<> struct TypeVec<type ## 4, 4> { typedef type ## 4 vec_type; }; \
template<> struct TypeVec<type, 8> { typedef type ## 8 vec_type; }; \
template<> struct TypeVec<type ## 8, 8> { typedef type ## 8 vec_type; };
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uchar)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(char)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(ushort)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(short)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(int)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uint)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(float)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(double)
#undef OPENCV_GPU_IMPLEMENT_TYPE_VEC
template<> struct TypeVec<schar, 1> { typedef schar vec_type; };
template<> struct TypeVec<schar, 2> { typedef char2 vec_type; };
template<> struct TypeVec<schar, 3> { typedef char3 vec_type; };
template<> struct TypeVec<schar, 4> { typedef char4 vec_type; };
template<> struct TypeVec<schar, 8> { typedef char8 vec_type; };
template<> struct TypeVec<bool, 1> { typedef uchar vec_type; };
template<> struct TypeVec<bool, 2> { typedef uchar2 vec_type; };
template<> struct TypeVec<bool, 3> { typedef uchar3 vec_type; };
template<> struct TypeVec<bool, 4> { typedef uchar4 vec_type; };
template<> struct TypeVec<bool, 8> { typedef uchar8 vec_type; };
template<typename T> struct VecTraits;
#define OPENCV_GPU_IMPLEMENT_VEC_TRAITS(type) \
template<> struct VecTraits<type> \
{ \
typedef type elem_type; \
enum {cn=1}; \
static __device__ __host__ __forceinline__ type all(type v) {return v;} \
static __device__ __host__ __forceinline__ type make(type x) {return x;} \
static __device__ __host__ __forceinline__ type make(const type* v) {return *v;} \
}; \
template<> struct VecTraits<type ## 1> \
{ \
typedef type elem_type; \
enum {cn=1}; \
static __device__ __host__ __forceinline__ type ## 1 all(type v) {return make_ ## type ## 1(v);} \
static __device__ __host__ __forceinline__ type ## 1 make(type x) {return make_ ## type ## 1(x);} \
static __device__ __host__ __forceinline__ type ## 1 make(const type* v) {return make_ ## type ## 1(*v);} \
}; \
template<> struct VecTraits<type ## 2> \
{ \
typedef type elem_type; \
enum {cn=2}; \
static __device__ __host__ __forceinline__ type ## 2 all(type v) {return make_ ## type ## 2(v, v);} \
static __device__ __host__ __forceinline__ type ## 2 make(type x, type y) {return make_ ## type ## 2(x, y);} \
static __device__ __host__ __forceinline__ type ## 2 make(const type* v) {return make_ ## type ## 2(v[0], v[1]);} \
}; \
template<> struct VecTraits<type ## 3> \
{ \
typedef type elem_type; \
enum {cn=3}; \
static __device__ __host__ __forceinline__ type ## 3 all(type v) {return make_ ## type ## 3(v, v, v);} \
static __device__ __host__ __forceinline__ type ## 3 make(type x, type y, type z) {return make_ ## type ## 3(x, y, z);} \
static __device__ __host__ __forceinline__ type ## 3 make(const type* v) {return make_ ## type ## 3(v[0], v[1], v[2]);} \
}; \
template<> struct VecTraits<type ## 4> \
{ \
typedef type elem_type; \
enum {cn=4}; \
static __device__ __host__ __forceinline__ type ## 4 all(type v) {return make_ ## type ## 4(v, v, v, v);} \
static __device__ __host__ __forceinline__ type ## 4 make(type x, type y, type z, type w) {return make_ ## type ## 4(x, y, z, w);} \
static __device__ __host__ __forceinline__ type ## 4 make(const type* v) {return make_ ## type ## 4(v[0], v[1], v[2], v[3]);} \
}; \
template<> struct VecTraits<type ## 8> \
{ \
typedef type elem_type; \
enum {cn=8}; \
static __device__ __host__ __forceinline__ type ## 8 all(type v) {return make_ ## type ## 8(v, v, v, v, v, v, v, v);} \
static __device__ __host__ __forceinline__ type ## 8 make(type a0, type a1, type a2, type a3, type a4, type a5, type a6, type a7) {return make_ ## type ## 8(a0, a1, a2, a3, a4, a5, a6, a7);} \
static __device__ __host__ __forceinline__ type ## 8 make(const type* v) {return make_ ## type ## 8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);} \
};
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uchar)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(ushort)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(short)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(int)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uint)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(float)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(double)
#undef OPENCV_GPU_IMPLEMENT_VEC_TRAITS
template<> struct VecTraits<char>
{
typedef char elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char all(char v) {return v;}
static __device__ __host__ __forceinline__ char make(char x) {return x;}
static __device__ __host__ __forceinline__ char make(const char* x) {return *x;}
};
template<> struct VecTraits<schar>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ schar all(schar v) {return v;}
static __device__ __host__ __forceinline__ schar make(schar x) {return x;}
static __device__ __host__ __forceinline__ schar make(const schar* x) {return *x;}
};
template<> struct VecTraits<char1>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char1 all(schar v) {return make_char1(v);}
static __device__ __host__ __forceinline__ char1 make(schar x) {return make_char1(x);}
static __device__ __host__ __forceinline__ char1 make(const schar* v) {return make_char1(v[0]);}
};
template<> struct VecTraits<char2>
{
typedef schar elem_type;
enum {cn=2};
static __device__ __host__ __forceinline__ char2 all(schar v) {return make_char2(v, v);}
static __device__ __host__ __forceinline__ char2 make(schar x, schar y) {return make_char2(x, y);}
static __device__ __host__ __forceinline__ char2 make(const schar* v) {return make_char2(v[0], v[1]);}
};
template<> struct VecTraits<char3>
{
typedef schar elem_type;
enum {cn=3};
static __device__ __host__ __forceinline__ char3 all(schar v) {return make_char3(v, v, v);}
static __device__ __host__ __forceinline__ char3 make(schar x, schar y, schar z) {return make_char3(x, y, z);}
static __device__ __host__ __forceinline__ char3 make(const schar* v) {return make_char3(v[0], v[1], v[2]);}
};
template<> struct VecTraits<char4>
{
typedef schar elem_type;
enum {cn=4};
static __device__ __host__ __forceinline__ char4 all(schar v) {return make_char4(v, v, v, v);}
static __device__ __host__ __forceinline__ char4 make(schar x, schar y, schar z, schar w) {return make_char4(x, y, z, w);}
static __device__ __host__ __forceinline__ char4 make(const schar* v) {return make_char4(v[0], v[1], v[2], v[3]);}
};
template<> struct VecTraits<char8>
{
typedef schar elem_type;
enum {cn=8};
static __device__ __host__ __forceinline__ char8 all(schar v) {return make_char8(v, v, v, v, v, v, v, v);}
static __device__ __host__ __forceinline__ char8 make(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7) {return make_char8(a0, a1, a2, a3, a4, a5, a6, a7);}
static __device__ __host__ __forceinline__ char8 make(const schar* v) {return make_char8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);}
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_TRAITS_HPP__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_VEC_TRAITS_HPP__
#define __OPENCV_GPU_VEC_TRAITS_HPP__
#include "common.hpp"
namespace cv { namespace gpu { namespace device
{
template<typename T, int N> struct TypeVec;
struct __align__(8) uchar8
{
uchar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uchar8 make_uchar8(uchar a0, uchar a1, uchar a2, uchar a3, uchar a4, uchar a5, uchar a6, uchar a7)
{
uchar8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(8) char8
{
schar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ char8 make_char8(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7)
{
char8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) ushort8
{
ushort a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ ushort8 make_ushort8(ushort a0, ushort a1, ushort a2, ushort a3, ushort a4, ushort a5, ushort a6, ushort a7)
{
ushort8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) short8
{
short a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ short8 make_short8(short a0, short a1, short a2, short a3, short a4, short a5, short a6, short a7)
{
short8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) uint8
{
uint a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uint8 make_uint8(uint a0, uint a1, uint a2, uint a3, uint a4, uint a5, uint a6, uint a7)
{
uint8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) int8
{
int a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ int8 make_int8(int a0, int a1, int a2, int a3, int a4, int a5, int a6, int a7)
{
int8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) float8
{
float a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ float8 make_float8(float a0, float a1, float a2, float a3, float a4, float a5, float a6, float a7)
{
float8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct double8
{
double a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ double8 make_double8(double a0, double a1, double a2, double a3, double a4, double a5, double a6, double a7)
{
double8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
#define OPENCV_GPU_IMPLEMENT_TYPE_VEC(type) \
template<> struct TypeVec<type, 1> { typedef type vec_type; }; \
template<> struct TypeVec<type ## 1, 1> { typedef type ## 1 vec_type; }; \
template<> struct TypeVec<type, 2> { typedef type ## 2 vec_type; }; \
template<> struct TypeVec<type ## 2, 2> { typedef type ## 2 vec_type; }; \
template<> struct TypeVec<type, 3> { typedef type ## 3 vec_type; }; \
template<> struct TypeVec<type ## 3, 3> { typedef type ## 3 vec_type; }; \
template<> struct TypeVec<type, 4> { typedef type ## 4 vec_type; }; \
template<> struct TypeVec<type ## 4, 4> { typedef type ## 4 vec_type; }; \
template<> struct TypeVec<type, 8> { typedef type ## 8 vec_type; }; \
template<> struct TypeVec<type ## 8, 8> { typedef type ## 8 vec_type; };
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uchar)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(char)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(ushort)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(short)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(int)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uint)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(float)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(double)
#undef OPENCV_GPU_IMPLEMENT_TYPE_VEC
template<> struct TypeVec<schar, 1> { typedef schar vec_type; };
template<> struct TypeVec<schar, 2> { typedef char2 vec_type; };
template<> struct TypeVec<schar, 3> { typedef char3 vec_type; };
template<> struct TypeVec<schar, 4> { typedef char4 vec_type; };
template<> struct TypeVec<schar, 8> { typedef char8 vec_type; };
template<> struct TypeVec<bool, 1> { typedef uchar vec_type; };
template<> struct TypeVec<bool, 2> { typedef uchar2 vec_type; };
template<> struct TypeVec<bool, 3> { typedef uchar3 vec_type; };
template<> struct TypeVec<bool, 4> { typedef uchar4 vec_type; };
template<> struct TypeVec<bool, 8> { typedef uchar8 vec_type; };
template<typename T> struct VecTraits;
#define OPENCV_GPU_IMPLEMENT_VEC_TRAITS(type) \
template<> struct VecTraits<type> \
{ \
typedef type elem_type; \
enum {cn=1}; \
static __device__ __host__ __forceinline__ type all(type v) {return v;} \
static __device__ __host__ __forceinline__ type make(type x) {return x;} \
static __device__ __host__ __forceinline__ type make(const type* v) {return *v;} \
}; \
template<> struct VecTraits<type ## 1> \
{ \
typedef type elem_type; \
enum {cn=1}; \
static __device__ __host__ __forceinline__ type ## 1 all(type v) {return make_ ## type ## 1(v);} \
static __device__ __host__ __forceinline__ type ## 1 make(type x) {return make_ ## type ## 1(x);} \
static __device__ __host__ __forceinline__ type ## 1 make(const type* v) {return make_ ## type ## 1(*v);} \
}; \
template<> struct VecTraits<type ## 2> \
{ \
typedef type elem_type; \
enum {cn=2}; \
static __device__ __host__ __forceinline__ type ## 2 all(type v) {return make_ ## type ## 2(v, v);} \
static __device__ __host__ __forceinline__ type ## 2 make(type x, type y) {return make_ ## type ## 2(x, y);} \
static __device__ __host__ __forceinline__ type ## 2 make(const type* v) {return make_ ## type ## 2(v[0], v[1]);} \
}; \
template<> struct VecTraits<type ## 3> \
{ \
typedef type elem_type; \
enum {cn=3}; \
static __device__ __host__ __forceinline__ type ## 3 all(type v) {return make_ ## type ## 3(v, v, v);} \
static __device__ __host__ __forceinline__ type ## 3 make(type x, type y, type z) {return make_ ## type ## 3(x, y, z);} \
static __device__ __host__ __forceinline__ type ## 3 make(const type* v) {return make_ ## type ## 3(v[0], v[1], v[2]);} \
}; \
template<> struct VecTraits<type ## 4> \
{ \
typedef type elem_type; \
enum {cn=4}; \
static __device__ __host__ __forceinline__ type ## 4 all(type v) {return make_ ## type ## 4(v, v, v, v);} \
static __device__ __host__ __forceinline__ type ## 4 make(type x, type y, type z, type w) {return make_ ## type ## 4(x, y, z, w);} \
static __device__ __host__ __forceinline__ type ## 4 make(const type* v) {return make_ ## type ## 4(v[0], v[1], v[2], v[3]);} \
}; \
template<> struct VecTraits<type ## 8> \
{ \
typedef type elem_type; \
enum {cn=8}; \
static __device__ __host__ __forceinline__ type ## 8 all(type v) {return make_ ## type ## 8(v, v, v, v, v, v, v, v);} \
static __device__ __host__ __forceinline__ type ## 8 make(type a0, type a1, type a2, type a3, type a4, type a5, type a6, type a7) {return make_ ## type ## 8(a0, a1, a2, a3, a4, a5, a6, a7);} \
static __device__ __host__ __forceinline__ type ## 8 make(const type* v) {return make_ ## type ## 8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);} \
};
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uchar)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(ushort)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(short)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(int)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uint)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(float)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(double)
#undef OPENCV_GPU_IMPLEMENT_VEC_TRAITS
template<> struct VecTraits<char>
{
typedef char elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char all(char v) {return v;}
static __device__ __host__ __forceinline__ char make(char x) {return x;}
static __device__ __host__ __forceinline__ char make(const char* x) {return *x;}
};
template<> struct VecTraits<schar>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ schar all(schar v) {return v;}
static __device__ __host__ __forceinline__ schar make(schar x) {return x;}
static __device__ __host__ __forceinline__ schar make(const schar* x) {return *x;}
};
template<> struct VecTraits<char1>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char1 all(schar v) {return make_char1(v);}
static __device__ __host__ __forceinline__ char1 make(schar x) {return make_char1(x);}
static __device__ __host__ __forceinline__ char1 make(const schar* v) {return make_char1(v[0]);}
};
template<> struct VecTraits<char2>
{
typedef schar elem_type;
enum {cn=2};
static __device__ __host__ __forceinline__ char2 all(schar v) {return make_char2(v, v);}
static __device__ __host__ __forceinline__ char2 make(schar x, schar y) {return make_char2(x, y);}
static __device__ __host__ __forceinline__ char2 make(const schar* v) {return make_char2(v[0], v[1]);}
};
template<> struct VecTraits<char3>
{
typedef schar elem_type;
enum {cn=3};
static __device__ __host__ __forceinline__ char3 all(schar v) {return make_char3(v, v, v);}
static __device__ __host__ __forceinline__ char3 make(schar x, schar y, schar z) {return make_char3(x, y, z);}
static __device__ __host__ __forceinline__ char3 make(const schar* v) {return make_char3(v[0], v[1], v[2]);}
};
template<> struct VecTraits<char4>
{
typedef schar elem_type;
enum {cn=4};
static __device__ __host__ __forceinline__ char4 all(schar v) {return make_char4(v, v, v, v);}
static __device__ __host__ __forceinline__ char4 make(schar x, schar y, schar z, schar w) {return make_char4(x, y, z, w);}
static __device__ __host__ __forceinline__ char4 make(const schar* v) {return make_char4(v[0], v[1], v[2], v[3]);}
};
template<> struct VecTraits<char8>
{
typedef schar elem_type;
enum {cn=8};
static __device__ __host__ __forceinline__ char8 all(schar v) {return make_char8(v, v, v, v, v, v, v, v);}
static __device__ __host__ __forceinline__ char8 make(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7) {return make_char8(a0, a1, a2, a3, a4, a5, a6, a7);}
static __device__ __host__ __forceinline__ char8 make(const schar* v) {return make_char8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);}
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_TRAITS_HPP__

222
modules/gpu/include/opencv2/gpu/device/warp.hpp
View File

@@ -1,112 +1,112 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DEVICE_WARP_HPP__
#define __OPENCV_GPU_DEVICE_WARP_HPP__
namespace cv { namespace gpu { namespace device
{
struct Warp
{
enum
{
LOG_WARP_SIZE = 5,
WARP_SIZE = 1 << LOG_WARP_SIZE,
STRIDE = WARP_SIZE
};
/** \brief Returns the warp lane ID of the calling thread. */
static __device__ __forceinline__ unsigned int laneId()
{
unsigned int ret;
asm("mov.u32 %0, %laneid;" : "=r"(ret) );
return ret;
}
template<typename It, typename T>
static __device__ __forceinline__ void fill(It beg, It end, const T& value)
{
for(It t = beg + laneId(); t < end; t += STRIDE)
*t = value;
}
template<typename InIt, typename OutIt>
static __device__ __forceinline__ OutIt copy(InIt beg, InIt end, OutIt out)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = *t;
return out;
}
template<typename InIt, typename OutIt, class UnOp>
static __device__ __forceinline__ OutIt transform(InIt beg, InIt end, OutIt out, UnOp op)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = op(*t);
return out;
}
template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
static __device__ __forceinline__ OutIt transform(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
{
unsigned int lane = laneId();
InIt1 t1 = beg1 + lane;
InIt2 t2 = beg2 + lane;
for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, out += STRIDE)
*out = op(*t1, *t2);
return out;
}
template<typename OutIt, typename T>
static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
{
unsigned int lane = laneId();
value += lane;
for(OutIt t = beg + lane; t < end; t += STRIDE, value += STRIDE)
*t = value;
}
};
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_DEVICE_WARP_HPP__
#define __OPENCV_GPU_DEVICE_WARP_HPP__
namespace cv { namespace gpu { namespace device
{
struct Warp
{
enum
{
LOG_WARP_SIZE = 5,
WARP_SIZE = 1 << LOG_WARP_SIZE,
STRIDE = WARP_SIZE
};
/** \brief Returns the warp lane ID of the calling thread. */
static __device__ __forceinline__ unsigned int laneId()
{
unsigned int ret;
asm("mov.u32 %0, %laneid;" : "=r"(ret) );
return ret;
}
template<typename It, typename T>
static __device__ __forceinline__ void fill(It beg, It end, const T& value)
{
for(It t = beg + laneId(); t < end; t += STRIDE)
*t = value;
}
template<typename InIt, typename OutIt>
static __device__ __forceinline__ OutIt copy(InIt beg, InIt end, OutIt out)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = *t;
return out;
}
template<typename InIt, typename OutIt, class UnOp>
static __device__ __forceinline__ OutIt transform(InIt beg, InIt end, OutIt out, UnOp op)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = op(*t);
return out;
}
template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
static __device__ __forceinline__ OutIt transform(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
{
unsigned int lane = laneId();
InIt1 t1 = beg1 + lane;
InIt2 t2 = beg2 + lane;
for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, out += STRIDE)
*out = op(*t1, *t2);
return out;
}
template<typename OutIt, typename T>
static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
{
unsigned int lane = laneId();
value += lane;
for(OutIt t = beg + lane; t < end; t += STRIDE, value += STRIDE)
*t = value;
}
};
}}} // namespace cv { namespace gpu { namespace device
#endif /* __OPENCV_GPU_DEVICE_WARP_HPP__ */

136
modules/gpu/include/opencv2/gpu/device/warp_reduce.hpp
View File

@@ -1,69 +1,69 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef OPENCV_GPU_WARP_REDUCE_HPP__
#define OPENCV_GPU_WARP_REDUCE_HPP__
namespace cv { namespace gpu { namespace device
{
template <class T>
__device__ __forceinline__ T warp_reduce(volatile T *ptr , const unsigned int tid = threadIdx.x)
{
const unsigned int lane = tid & 31; // index of thread in warp (0..31)
if (lane < 16)
{
T partial = ptr[tid];
ptr[tid] = partial = partial + ptr[tid + 16];
ptr[tid] = partial = partial + ptr[tid + 8];
ptr[tid] = partial = partial + ptr[tid + 4];
ptr[tid] = partial = partial + ptr[tid + 2];
ptr[tid] = partial = partial + ptr[tid + 1];
}
return ptr[tid - lane];
}
}}} // namespace cv { namespace gpu { namespace device {
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef OPENCV_GPU_WARP_REDUCE_HPP__
#define OPENCV_GPU_WARP_REDUCE_HPP__
namespace cv { namespace gpu { namespace device
{
template <class T>
__device__ __forceinline__ T warp_reduce(volatile T *ptr , const unsigned int tid = threadIdx.x)
{
const unsigned int lane = tid & 31; // index of thread in warp (0..31)
if (lane < 16)
{
T partial = ptr[tid];
ptr[tid] = partial = partial + ptr[tid + 16];
ptr[tid] = partial = partial + ptr[tid + 8];
ptr[tid] = partial = partial + ptr[tid + 4];
ptr[tid] = partial = partial + ptr[tid + 2];
ptr[tid] = partial = partial + ptr[tid + 1];
}
return ptr[tid - lane];
}
}}} // namespace cv { namespace gpu { namespace device {
#endif /* OPENCV_GPU_WARP_REDUCE_HPP__ */

86
modules/gpu/include/opencv2/gpu/devmem2d.hpp
View File

@@ -1,43 +1,43 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "opencv2/core/cuda_devptrs.hpp"
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "opencv2/core/cuda_devptrs.hpp"

4996
modules/gpu/include/opencv2/gpu/gpu.hpp
View File

File diff suppressed because it is too large Load Diff

86
modules/gpu/include/opencv2/gpu/gpumat.hpp
View File

@@ -1,43 +1,43 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "opencv2/core/gpumat.hpp"
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "opencv2/core/gpumat.hpp"

126
modules/gpu/include/opencv2/gpu/stream_accessor.hpp
View File

@@ -1,64 +1,64 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_STREAM_ACCESSOR_HPP__
#define __OPENCV_GPU_STREAM_ACCESSOR_HPP__
#include "opencv2/gpu/gpu.hpp"
#include "cuda_runtime_api.h"
namespace cv
{
namespace gpu
{
// This is only header file that depends on Cuda. All other headers are independent.
// So if you use OpenCV binaries you do noot need to install Cuda Toolkit.
// But of you wanna use GPU by yourself, may get cuda stream instance using the class below.
// In this case you have to install Cuda Toolkit.
struct StreamAccessor
{
CV_EXPORTS static cudaStream_t getStream(const Stream& stream);
};
}
}
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_GPU_STREAM_ACCESSOR_HPP__
#define __OPENCV_GPU_STREAM_ACCESSOR_HPP__
#include "opencv2/gpu/gpu.hpp"
#include "cuda_runtime_api.h"
namespace cv
{
namespace gpu
{
// This is only header file that depends on Cuda. All other headers are independent.
// So if you use OpenCV binaries you do noot need to install Cuda Toolkit.
// But of you wanna use GPU by yourself, may get cuda stream instance using the class below.
// In this case you have to install Cuda Toolkit.
struct StreamAccessor
{
CV_EXPORTS static cudaStream_t getStream(const Stream& stream);
};
}
}
#endif /* __OPENCV_GPU_STREAM_ACCESSOR_HPP__ */

0
modules/gpu/misc/mark_nvidia.py Normal file → Executable file
View File

762
modules/gpu/perf/perf_calib3d.cpp
View File

@@ -1,381 +1,381 @@
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// StereoBM
typedef std::tr1::tuple<string, string> pair_string;
DEF_PARAM_TEST_1(ImagePair, pair_string);
PERF_TEST_P(ImagePair, Calib3D_StereoBM, Values(pair_string("gpu/perf/aloe.png", "gpu/perf/aloeR.png")))
{
declare.time(5.0);
const cv::Mat imgLeft = readImage(GET_PARAM(0), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgLeft.empty());
const cv::Mat imgRight = readImage(GET_PARAM(1), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgRight.empty());
const int preset = 0;
const int ndisp = 256;
if (PERF_RUN_GPU())
{
cv::gpu::StereoBM_GPU d_bm(preset, ndisp);
cv::gpu::GpuMat d_imgLeft(imgLeft);
cv::gpu::GpuMat d_imgRight(imgRight);
cv::gpu::GpuMat d_dst;
d_bm(d_imgLeft, d_imgRight, d_dst);
TEST_CYCLE()
{
d_bm(d_imgLeft, d_imgRight, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::StereoBM bm(preset, ndisp);
cv::Mat dst;
bm(imgLeft, imgRight, dst);
TEST_CYCLE()
{
bm(imgLeft, imgRight, dst);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// StereoBeliefPropagation
PERF_TEST_P(ImagePair, Calib3D_StereoBeliefPropagation, Values(pair_string("gpu/stereobp/aloe-L.png", "gpu/stereobp/aloe-R.png")))
{
declare.time(10.0);
const cv::Mat imgLeft = readImage(GET_PARAM(0));
ASSERT_FALSE(imgLeft.empty());
const cv::Mat imgRight = readImage(GET_PARAM(1));
ASSERT_FALSE(imgRight.empty());
const int ndisp = 64;
if (PERF_RUN_GPU())
{
cv::gpu::StereoBeliefPropagation d_bp(ndisp);
cv::gpu::GpuMat d_imgLeft(imgLeft);
cv::gpu::GpuMat d_imgRight(imgRight);
cv::gpu::GpuMat d_dst;
d_bp(d_imgLeft, d_imgRight, d_dst);
TEST_CYCLE()
{
d_bp(d_imgLeft, d_imgRight, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// StereoConstantSpaceBP
PERF_TEST_P(ImagePair, Calib3D_StereoConstantSpaceBP, Values(pair_string("gpu/stereobm/aloe-L.png", "gpu/stereobm/aloe-R.png")))
{
declare.time(10.0);
const cv::Mat imgLeft = readImage(GET_PARAM(0), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgLeft.empty());
const cv::Mat imgRight = readImage(GET_PARAM(1), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgRight.empty());
const int ndisp = 128;
if (PERF_RUN_GPU())
{
cv::gpu::StereoConstantSpaceBP d_csbp(ndisp);
cv::gpu::GpuMat d_imgLeft(imgLeft);
cv::gpu::GpuMat d_imgRight(imgRight);
cv::gpu::GpuMat d_dst;
d_csbp(d_imgLeft, d_imgRight, d_dst);
TEST_CYCLE()
{
d_csbp(d_imgLeft, d_imgRight, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// DisparityBilateralFilter
PERF_TEST_P(ImagePair, Calib3D_DisparityBilateralFilter, Values(pair_string("gpu/stereobm/aloe-L.png", "gpu/stereobm/aloe-disp.png")))
{
const cv::Mat img = readImage(GET_PARAM(0), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
const cv::Mat disp = readImage(GET_PARAM(1), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(disp.empty());
const int ndisp = 128;
if (PERF_RUN_GPU())
{
cv::gpu::DisparityBilateralFilter d_filter(ndisp);
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_disp(disp);
cv::gpu::GpuMat d_dst;
d_filter(d_disp, d_img, d_dst);
TEST_CYCLE()
{
d_filter(d_disp, d_img, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// TransformPoints
DEF_PARAM_TEST_1(Count, int);
PERF_TEST_P(Count, Calib3D_TransformPoints, Values(5000, 10000, 20000))
{
const int count = GetParam();
cv::Mat src(1, count, CV_32FC3);
fillRandom(src, -100, 100);
const cv::Mat rvec = cv::Mat::ones(1, 3, CV_32FC1);
const cv::Mat tvec = cv::Mat::ones(1, 3, CV_32FC1);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::transformPoints(d_src, rvec, tvec, d_dst);
TEST_CYCLE()
{
cv::gpu::transformPoints(d_src, rvec, tvec, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// ProjectPoints
PERF_TEST_P(Count, Calib3D_ProjectPoints, Values(5000, 10000, 20000))
{
const int count = GetParam();
cv::Mat src(1, count, CV_32FC3);
fillRandom(src, -100, 100);
const cv::Mat rvec = cv::Mat::ones(1, 3, CV_32FC1);
const cv::Mat tvec = cv::Mat::ones(1, 3, CV_32FC1);
const cv::Mat camera_mat = cv::Mat::ones(3, 3, CV_32FC1);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::projectPoints(d_src, rvec, tvec, camera_mat, cv::Mat(), d_dst);
TEST_CYCLE()
{
cv::gpu::projectPoints(d_src, rvec, tvec, camera_mat, cv::Mat(), d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::projectPoints(src, rvec, tvec, camera_mat, cv::noArray(), dst);
TEST_CYCLE()
{
cv::projectPoints(src, rvec, tvec, camera_mat, cv::noArray(), dst);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// SolvePnPRansac
PERF_TEST_P(Count, Calib3D_SolvePnPRansac, Values(5000, 10000, 20000))
{
declare.time(10.0);
const int count = GetParam();
cv::Mat object(1, count, CV_32FC3);
fillRandom(object, -100, 100);
cv::Mat camera_mat(3, 3, CV_32FC1);
fillRandom(camera_mat, 0.5, 1);
camera_mat.at<float>(0, 1) = 0.f;
camera_mat.at<float>(1, 0) = 0.f;
camera_mat.at<float>(2, 0) = 0.f;
camera_mat.at<float>(2, 1) = 0.f;
const cv::Mat dist_coef(1, 8, CV_32F, cv::Scalar::all(0));
std::vector<cv::Point2f> image_vec;
cv::Mat rvec_gold(1, 3, CV_32FC1);
fillRandom(rvec_gold, 0, 1);
cv::Mat tvec_gold(1, 3, CV_32FC1);
fillRandom(tvec_gold, 0, 1);
cv::projectPoints(object, rvec_gold, tvec_gold, camera_mat, dist_coef, image_vec);
cv::Mat image(1, count, CV_32FC2, &image_vec[0]);
cv::Mat rvec;
cv::Mat tvec;
if (PERF_RUN_GPU())
{
cv::gpu::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
TEST_CYCLE()
{
cv::gpu::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
}
}
else
{
cv::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
TEST_CYCLE()
{
cv::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
}
}
CPU_SANITY_CHECK(rvec);
CPU_SANITY_CHECK(tvec);
}
//////////////////////////////////////////////////////////////////////
// ReprojectImageTo3D
PERF_TEST_P(Sz_Depth, Calib3D_ReprojectImageTo3D, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16S)))
{
const cv::Size size = GET_PARAM(0);
const int depth = GET_PARAM(1);
cv::Mat src(size, depth);
fillRandom(src, 5.0, 30.0);
cv::Mat Q(4, 4, CV_32FC1);
fillRandom(Q, 0.1, 1.0);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::reprojectImageTo3D(d_src, d_dst, Q);
TEST_CYCLE()
{
cv::gpu::reprojectImageTo3D(d_src, d_dst, Q);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::reprojectImageTo3D(src, dst, Q);
TEST_CYCLE()
{
cv::reprojectImageTo3D(src, dst, Q);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// DrawColorDisp
PERF_TEST_P(Sz_Depth, Calib3D_DrawColorDisp, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16S)))
{
const cv::Size size = GET_PARAM(0);
const int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src, 0, 255);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::drawColorDisp(d_src, d_dst, 255);
TEST_CYCLE()
{
cv::gpu::drawColorDisp(d_src, d_dst, 255);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
} // namespace
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// StereoBM
typedef std::tr1::tuple<string, string> pair_string;
DEF_PARAM_TEST_1(ImagePair, pair_string);
PERF_TEST_P(ImagePair, Calib3D_StereoBM, Values(pair_string("gpu/perf/aloe.png", "gpu/perf/aloeR.png")))
{
declare.time(5.0);
const cv::Mat imgLeft = readImage(GET_PARAM(0), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgLeft.empty());
const cv::Mat imgRight = readImage(GET_PARAM(1), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgRight.empty());
const int preset = 0;
const int ndisp = 256;
if (PERF_RUN_GPU())
{
cv::gpu::StereoBM_GPU d_bm(preset, ndisp);
cv::gpu::GpuMat d_imgLeft(imgLeft);
cv::gpu::GpuMat d_imgRight(imgRight);
cv::gpu::GpuMat d_dst;
d_bm(d_imgLeft, d_imgRight, d_dst);
TEST_CYCLE()
{
d_bm(d_imgLeft, d_imgRight, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::StereoBM bm(preset, ndisp);
cv::Mat dst;
bm(imgLeft, imgRight, dst);
TEST_CYCLE()
{
bm(imgLeft, imgRight, dst);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// StereoBeliefPropagation
PERF_TEST_P(ImagePair, Calib3D_StereoBeliefPropagation, Values(pair_string("gpu/stereobp/aloe-L.png", "gpu/stereobp/aloe-R.png")))
{
declare.time(10.0);
const cv::Mat imgLeft = readImage(GET_PARAM(0));
ASSERT_FALSE(imgLeft.empty());
const cv::Mat imgRight = readImage(GET_PARAM(1));
ASSERT_FALSE(imgRight.empty());
const int ndisp = 64;
if (PERF_RUN_GPU())
{
cv::gpu::StereoBeliefPropagation d_bp(ndisp);
cv::gpu::GpuMat d_imgLeft(imgLeft);
cv::gpu::GpuMat d_imgRight(imgRight);
cv::gpu::GpuMat d_dst;
d_bp(d_imgLeft, d_imgRight, d_dst);
TEST_CYCLE()
{
d_bp(d_imgLeft, d_imgRight, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// StereoConstantSpaceBP
PERF_TEST_P(ImagePair, Calib3D_StereoConstantSpaceBP, Values(pair_string("gpu/stereobm/aloe-L.png", "gpu/stereobm/aloe-R.png")))
{
declare.time(10.0);
const cv::Mat imgLeft = readImage(GET_PARAM(0), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgLeft.empty());
const cv::Mat imgRight = readImage(GET_PARAM(1), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(imgRight.empty());
const int ndisp = 128;
if (PERF_RUN_GPU())
{
cv::gpu::StereoConstantSpaceBP d_csbp(ndisp);
cv::gpu::GpuMat d_imgLeft(imgLeft);
cv::gpu::GpuMat d_imgRight(imgRight);
cv::gpu::GpuMat d_dst;
d_csbp(d_imgLeft, d_imgRight, d_dst);
TEST_CYCLE()
{
d_csbp(d_imgLeft, d_imgRight, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// DisparityBilateralFilter
PERF_TEST_P(ImagePair, Calib3D_DisparityBilateralFilter, Values(pair_string("gpu/stereobm/aloe-L.png", "gpu/stereobm/aloe-disp.png")))
{
const cv::Mat img = readImage(GET_PARAM(0), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
const cv::Mat disp = readImage(GET_PARAM(1), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(disp.empty());
const int ndisp = 128;
if (PERF_RUN_GPU())
{
cv::gpu::DisparityBilateralFilter d_filter(ndisp);
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_disp(disp);
cv::gpu::GpuMat d_dst;
d_filter(d_disp, d_img, d_dst);
TEST_CYCLE()
{
d_filter(d_disp, d_img, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// TransformPoints
DEF_PARAM_TEST_1(Count, int);
PERF_TEST_P(Count, Calib3D_TransformPoints, Values(5000, 10000, 20000))
{
const int count = GetParam();
cv::Mat src(1, count, CV_32FC3);
fillRandom(src, -100, 100);
const cv::Mat rvec = cv::Mat::ones(1, 3, CV_32FC1);
const cv::Mat tvec = cv::Mat::ones(1, 3, CV_32FC1);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::transformPoints(d_src, rvec, tvec, d_dst);
TEST_CYCLE()
{
cv::gpu::transformPoints(d_src, rvec, tvec, d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
//////////////////////////////////////////////////////////////////////
// ProjectPoints
PERF_TEST_P(Count, Calib3D_ProjectPoints, Values(5000, 10000, 20000))
{
const int count = GetParam();
cv::Mat src(1, count, CV_32FC3);
fillRandom(src, -100, 100);
const cv::Mat rvec = cv::Mat::ones(1, 3, CV_32FC1);
const cv::Mat tvec = cv::Mat::ones(1, 3, CV_32FC1);
const cv::Mat camera_mat = cv::Mat::ones(3, 3, CV_32FC1);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::projectPoints(d_src, rvec, tvec, camera_mat, cv::Mat(), d_dst);
TEST_CYCLE()
{
cv::gpu::projectPoints(d_src, rvec, tvec, camera_mat, cv::Mat(), d_dst);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::projectPoints(src, rvec, tvec, camera_mat, cv::noArray(), dst);
TEST_CYCLE()
{
cv::projectPoints(src, rvec, tvec, camera_mat, cv::noArray(), dst);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// SolvePnPRansac
PERF_TEST_P(Count, Calib3D_SolvePnPRansac, Values(5000, 10000, 20000))
{
declare.time(10.0);
const int count = GetParam();
cv::Mat object(1, count, CV_32FC3);
fillRandom(object, -100, 100);
cv::Mat camera_mat(3, 3, CV_32FC1);
fillRandom(camera_mat, 0.5, 1);
camera_mat.at<float>(0, 1) = 0.f;
camera_mat.at<float>(1, 0) = 0.f;
camera_mat.at<float>(2, 0) = 0.f;
camera_mat.at<float>(2, 1) = 0.f;
const cv::Mat dist_coef(1, 8, CV_32F, cv::Scalar::all(0));
std::vector<cv::Point2f> image_vec;
cv::Mat rvec_gold(1, 3, CV_32FC1);
fillRandom(rvec_gold, 0, 1);
cv::Mat tvec_gold(1, 3, CV_32FC1);
fillRandom(tvec_gold, 0, 1);
cv::projectPoints(object, rvec_gold, tvec_gold, camera_mat, dist_coef, image_vec);
cv::Mat image(1, count, CV_32FC2, &image_vec[0]);
cv::Mat rvec;
cv::Mat tvec;
if (PERF_RUN_GPU())
{
cv::gpu::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
TEST_CYCLE()
{
cv::gpu::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
}
}
else
{
cv::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
TEST_CYCLE()
{
cv::solvePnPRansac(object, image, camera_mat, dist_coef, rvec, tvec);
}
}
CPU_SANITY_CHECK(rvec);
CPU_SANITY_CHECK(tvec);
}
//////////////////////////////////////////////////////////////////////
// ReprojectImageTo3D
PERF_TEST_P(Sz_Depth, Calib3D_ReprojectImageTo3D, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16S)))
{
const cv::Size size = GET_PARAM(0);
const int depth = GET_PARAM(1);
cv::Mat src(size, depth);
fillRandom(src, 5.0, 30.0);
cv::Mat Q(4, 4, CV_32FC1);
fillRandom(Q, 0.1, 1.0);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::reprojectImageTo3D(d_src, d_dst, Q);
TEST_CYCLE()
{
cv::gpu::reprojectImageTo3D(d_src, d_dst, Q);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::reprojectImageTo3D(src, dst, Q);
TEST_CYCLE()
{
cv::reprojectImageTo3D(src, dst, Q);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// DrawColorDisp
PERF_TEST_P(Sz_Depth, Calib3D_DrawColorDisp, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16S)))
{
const cv::Size size = GET_PARAM(0);
const int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src, 0, 255);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::drawColorDisp(d_src, d_dst, 255);
TEST_CYCLE()
{
cv::gpu::drawColorDisp(d_src, d_dst, 255);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
FAIL() << "No such CPU implementation analogy.";
}
}
} // namespace

4910
modules/gpu/perf/perf_core.cpp
View File

File diff suppressed because it is too large Load Diff

618
modules/gpu/perf/perf_features2d.cpp
View File

@@ -1,309 +1,309 @@
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// SURF
DEF_PARAM_TEST_1(Image, string);
PERF_TEST_P(Image, Features2D_SURF, Values<string>("gpu/perf/aloe.png"))
{
declare.time(50.0);
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::SURF_GPU d_surf;
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_keypoints, d_descriptors;
d_surf(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
TEST_CYCLE()
{
d_surf(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
}
GPU_SANITY_CHECK(d_descriptors, 1e-4);
GPU_SANITY_CHECK_KEYPOINTS(SURF, d_keypoints);
}
else
{
cv::SURF surf;
std::vector<cv::KeyPoint> keypoints;
cv::Mat descriptors;
surf(img, cv::noArray(), keypoints, descriptors);
TEST_CYCLE()
{
keypoints.clear();
surf(img, cv::noArray(), keypoints, descriptors);
}
SANITY_CHECK_KEYPOINTS(keypoints);
SANITY_CHECK(descriptors, 1e-4);
}
}
//////////////////////////////////////////////////////////////////////
// FAST
PERF_TEST_P(Image, Features2D_FAST, Values<string>("gpu/perf/aloe.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::FAST_GPU d_fast(20);
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_keypoints;
d_fast(d_img, cv::gpu::GpuMat(), d_keypoints);
TEST_CYCLE()
{
d_fast(d_img, cv::gpu::GpuMat(), d_keypoints);
}
GPU_SANITY_CHECK_RESPONSE(FAST, d_keypoints);
}
else
{
std::vector<cv::KeyPoint> keypoints;
cv::FAST(img, keypoints, 20);
TEST_CYCLE()
{
keypoints.clear();
cv::FAST(img, keypoints, 20);
}
SANITY_CHECK_KEYPOINTS(keypoints);
}
}
//////////////////////////////////////////////////////////////////////
// ORB
PERF_TEST_P(Image, Features2D_ORB, Values<string>("gpu/perf/aloe.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::ORB_GPU d_orb(4000);
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_keypoints, d_descriptors;
d_orb(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
TEST_CYCLE()
{
d_orb(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
}
GPU_SANITY_CHECK_KEYPOINTS(ORB, d_keypoints);
GPU_SANITY_CHECK(d_descriptors);
}
else
{
cv::ORB orb(4000);
std::vector<cv::KeyPoint> keypoints;
cv::Mat descriptors;
orb(img, cv::noArray(), keypoints, descriptors);
TEST_CYCLE()
{
keypoints.clear();
orb(img, cv::noArray(), keypoints, descriptors);
}
SANITY_CHECK_KEYPOINTS(keypoints);
SANITY_CHECK(descriptors);
}
}
//////////////////////////////////////////////////////////////////////
// BFMatch
DEF_PARAM_TEST(DescSize_Norm, int, NormType);
PERF_TEST_P(DescSize_Norm, Features2D_BFMatch, Combine(Values(64, 128, 256), Values(NormType(cv::NORM_L1), NormType(cv::NORM_L2), NormType(cv::NORM_HAMMING))))
{
declare.time(20.0);
int desc_size = GET_PARAM(0);
int normType = GET_PARAM(1);
int type = normType == cv::NORM_HAMMING ? CV_8U : CV_32F;
cv::Mat query(3000, desc_size, type);
fillRandom(query);
cv::Mat train(3000, desc_size, type);
fillRandom(train);
if (PERF_RUN_GPU())
{
cv::gpu::BFMatcher_GPU d_matcher(normType);
cv::gpu::GpuMat d_query(query);
cv::gpu::GpuMat d_train(train);
cv::gpu::GpuMat d_trainIdx, d_distance;
d_matcher.matchSingle(d_query, d_train, d_trainIdx, d_distance);
TEST_CYCLE()
{
d_matcher.matchSingle(d_query, d_train, d_trainIdx, d_distance);
}
GPU_SANITY_CHECK(d_trainIdx);
GPU_SANITY_CHECK(d_distance);
}
else
{
cv::BFMatcher matcher(normType);
std::vector<cv::DMatch> matches;
matcher.match(query, train, matches);
TEST_CYCLE()
{
matcher.match(query, train, matches);
}
SANITY_CHECK(matches);
}
}
//////////////////////////////////////////////////////////////////////
// BFKnnMatch
DEF_PARAM_TEST(DescSize_K_Norm, int, int, NormType);
PERF_TEST_P(DescSize_K_Norm, Features2D_BFKnnMatch, Combine(
Values(64, 128, 256),
Values(2, 3),
Values(NormType(cv::NORM_L1), NormType(cv::NORM_L2), NormType(cv::NORM_HAMMING))))
{
declare.time(30.0);
int desc_size = GET_PARAM(0);
int k = GET_PARAM(1);
int normType = GET_PARAM(2);
int type = normType == cv::NORM_HAMMING ? CV_8U : CV_32F;
cv::Mat query(3000, desc_size, type);
fillRandom(query);
cv::Mat train(3000, desc_size, type);
fillRandom(train);
if (PERF_RUN_GPU())
{
cv::gpu::BFMatcher_GPU d_matcher(normType);
cv::gpu::GpuMat d_query(query);
cv::gpu::GpuMat d_train(train);
cv::gpu::GpuMat d_trainIdx, d_distance, d_allDist;
d_matcher.knnMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_allDist, k);
TEST_CYCLE()
{
d_matcher.knnMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_allDist, k);
}
GPU_SANITY_CHECK(d_trainIdx);
GPU_SANITY_CHECK(d_distance);
}
else
{
cv::BFMatcher matcher(normType);
std::vector< std::vector<cv::DMatch> > matches;
matcher.knnMatch(query, train, matches, k);
TEST_CYCLE()
{
matcher.knnMatch(query, train, matches, k);
}
SANITY_CHECK(matches);
}
}
//////////////////////////////////////////////////////////////////////
// BFRadiusMatch
PERF_TEST_P(DescSize_Norm, Features2D_BFRadiusMatch, Combine(Values(64, 128, 256), Values(NormType(cv::NORM_L1), NormType(cv::NORM_L2), NormType(cv::NORM_HAMMING))))
{
declare.time(30.0);
int desc_size = GET_PARAM(0);
int normType = GET_PARAM(1);
int type = normType == cv::NORM_HAMMING ? CV_8U : CV_32F;
cv::Mat query(3000, desc_size, type);
fillRandom(query, 0.0, 1.0);
cv::Mat train(3000, desc_size, type);
fillRandom(train, 0.0, 1.0);
if (PERF_RUN_GPU())
{
cv::gpu::BFMatcher_GPU d_matcher(normType);
cv::gpu::GpuMat d_query(query);
cv::gpu::GpuMat d_train(train);
cv::gpu::GpuMat d_trainIdx, d_nMatches, d_distance;
d_matcher.radiusMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_nMatches, 2.0);
TEST_CYCLE()
{
d_matcher.radiusMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_nMatches, 2.0);
}
GPU_SANITY_CHECK(d_trainIdx);
GPU_SANITY_CHECK(d_distance);
}
else
{
cv::BFMatcher matcher(normType);
std::vector< std::vector<cv::DMatch> > matches;
matcher.radiusMatch(query, train, matches, 2.0);
TEST_CYCLE()
{
matcher.radiusMatch(query, train, matches, 2.0);
}
SANITY_CHECK(matches);
}
}
} // namespace
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// SURF
DEF_PARAM_TEST_1(Image, string);
PERF_TEST_P(Image, Features2D_SURF, Values<string>("gpu/perf/aloe.png"))
{
declare.time(50.0);
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::SURF_GPU d_surf;
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_keypoints, d_descriptors;
d_surf(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
TEST_CYCLE()
{
d_surf(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
}
GPU_SANITY_CHECK(d_descriptors, 1e-4);
GPU_SANITY_CHECK_KEYPOINTS(SURF, d_keypoints);
}
else
{
cv::SURF surf;
std::vector<cv::KeyPoint> keypoints;
cv::Mat descriptors;
surf(img, cv::noArray(), keypoints, descriptors);
TEST_CYCLE()
{
keypoints.clear();
surf(img, cv::noArray(), keypoints, descriptors);
}
SANITY_CHECK_KEYPOINTS(keypoints);
SANITY_CHECK(descriptors, 1e-4);
}
}
//////////////////////////////////////////////////////////////////////
// FAST
PERF_TEST_P(Image, Features2D_FAST, Values<string>("gpu/perf/aloe.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::FAST_GPU d_fast(20);
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_keypoints;
d_fast(d_img, cv::gpu::GpuMat(), d_keypoints);
TEST_CYCLE()
{
d_fast(d_img, cv::gpu::GpuMat(), d_keypoints);
}
GPU_SANITY_CHECK_RESPONSE(FAST, d_keypoints);
}
else
{
std::vector<cv::KeyPoint> keypoints;
cv::FAST(img, keypoints, 20);
TEST_CYCLE()
{
keypoints.clear();
cv::FAST(img, keypoints, 20);
}
SANITY_CHECK_KEYPOINTS(keypoints);
}
}
//////////////////////////////////////////////////////////////////////
// ORB
PERF_TEST_P(Image, Features2D_ORB, Values<string>("gpu/perf/aloe.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::ORB_GPU d_orb(4000);
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_keypoints, d_descriptors;
d_orb(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
TEST_CYCLE()
{
d_orb(d_img, cv::gpu::GpuMat(), d_keypoints, d_descriptors);
}
GPU_SANITY_CHECK_KEYPOINTS(ORB, d_keypoints);
GPU_SANITY_CHECK(d_descriptors);
}
else
{
cv::ORB orb(4000);
std::vector<cv::KeyPoint> keypoints;
cv::Mat descriptors;
orb(img, cv::noArray(), keypoints, descriptors);
TEST_CYCLE()
{
keypoints.clear();
orb(img, cv::noArray(), keypoints, descriptors);
}
SANITY_CHECK_KEYPOINTS(keypoints);
SANITY_CHECK(descriptors);
}
}
//////////////////////////////////////////////////////////////////////
// BFMatch
DEF_PARAM_TEST(DescSize_Norm, int, NormType);
PERF_TEST_P(DescSize_Norm, Features2D_BFMatch, Combine(Values(64, 128, 256), Values(NormType(cv::NORM_L1), NormType(cv::NORM_L2), NormType(cv::NORM_HAMMING))))
{
declare.time(20.0);
int desc_size = GET_PARAM(0);
int normType = GET_PARAM(1);
int type = normType == cv::NORM_HAMMING ? CV_8U : CV_32F;
cv::Mat query(3000, desc_size, type);
fillRandom(query);
cv::Mat train(3000, desc_size, type);
fillRandom(train);
if (PERF_RUN_GPU())
{
cv::gpu::BFMatcher_GPU d_matcher(normType);
cv::gpu::GpuMat d_query(query);
cv::gpu::GpuMat d_train(train);
cv::gpu::GpuMat d_trainIdx, d_distance;
d_matcher.matchSingle(d_query, d_train, d_trainIdx, d_distance);
TEST_CYCLE()
{
d_matcher.matchSingle(d_query, d_train, d_trainIdx, d_distance);
}
GPU_SANITY_CHECK(d_trainIdx);
GPU_SANITY_CHECK(d_distance);
}
else
{
cv::BFMatcher matcher(normType);
std::vector<cv::DMatch> matches;
matcher.match(query, train, matches);
TEST_CYCLE()
{
matcher.match(query, train, matches);
}
SANITY_CHECK(matches);
}
}
//////////////////////////////////////////////////////////////////////
// BFKnnMatch
DEF_PARAM_TEST(DescSize_K_Norm, int, int, NormType);
PERF_TEST_P(DescSize_K_Norm, Features2D_BFKnnMatch, Combine(
Values(64, 128, 256),
Values(2, 3),
Values(NormType(cv::NORM_L1), NormType(cv::NORM_L2), NormType(cv::NORM_HAMMING))))
{
declare.time(30.0);
int desc_size = GET_PARAM(0);
int k = GET_PARAM(1);
int normType = GET_PARAM(2);
int type = normType == cv::NORM_HAMMING ? CV_8U : CV_32F;
cv::Mat query(3000, desc_size, type);
fillRandom(query);
cv::Mat train(3000, desc_size, type);
fillRandom(train);
if (PERF_RUN_GPU())
{
cv::gpu::BFMatcher_GPU d_matcher(normType);
cv::gpu::GpuMat d_query(query);
cv::gpu::GpuMat d_train(train);
cv::gpu::GpuMat d_trainIdx, d_distance, d_allDist;
d_matcher.knnMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_allDist, k);
TEST_CYCLE()
{
d_matcher.knnMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_allDist, k);
}
GPU_SANITY_CHECK(d_trainIdx);
GPU_SANITY_CHECK(d_distance);
}
else
{
cv::BFMatcher matcher(normType);
std::vector< std::vector<cv::DMatch> > matches;
matcher.knnMatch(query, train, matches, k);
TEST_CYCLE()
{
matcher.knnMatch(query, train, matches, k);
}
SANITY_CHECK(matches);
}
}
//////////////////////////////////////////////////////////////////////
// BFRadiusMatch
PERF_TEST_P(DescSize_Norm, Features2D_BFRadiusMatch, Combine(Values(64, 128, 256), Values(NormType(cv::NORM_L1), NormType(cv::NORM_L2), NormType(cv::NORM_HAMMING))))
{
declare.time(30.0);
int desc_size = GET_PARAM(0);
int normType = GET_PARAM(1);
int type = normType == cv::NORM_HAMMING ? CV_8U : CV_32F;
cv::Mat query(3000, desc_size, type);
fillRandom(query, 0.0, 1.0);
cv::Mat train(3000, desc_size, type);
fillRandom(train, 0.0, 1.0);
if (PERF_RUN_GPU())
{
cv::gpu::BFMatcher_GPU d_matcher(normType);
cv::gpu::GpuMat d_query(query);
cv::gpu::GpuMat d_train(train);
cv::gpu::GpuMat d_trainIdx, d_nMatches, d_distance;
d_matcher.radiusMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_nMatches, 2.0);
TEST_CYCLE()
{
d_matcher.radiusMatchSingle(d_query, d_train, d_trainIdx, d_distance, d_nMatches, 2.0);
}
GPU_SANITY_CHECK(d_trainIdx);
GPU_SANITY_CHECK(d_distance);
}
else
{
cv::BFMatcher matcher(normType);
std::vector< std::vector<cv::DMatch> > matches;
matcher.radiusMatch(query, train, matches, 2.0);
TEST_CYCLE()
{
matcher.radiusMatch(query, train, matches, 2.0);
}
SANITY_CHECK(matches);
}
}
} // namespace

830
modules/gpu/perf/perf_filters.cpp
View File

@@ -1,415 +1,415 @@
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// Blur
DEF_PARAM_TEST(Sz_Type_KernelSz, cv::Size, MatType, int);
PERF_TEST_P(Sz_Type_KernelSz, Filters_Blur, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4), Values(3, 5, 7)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::blur(d_src, d_dst, cv::Size(ksize, ksize));
TEST_CYCLE()
{
cv::gpu::blur(d_src, d_dst, cv::Size(ksize, ksize));
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::blur(src, dst, cv::Size(ksize, ksize));
TEST_CYCLE()
{
cv::blur(src, dst, cv::Size(ksize, ksize));
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Sobel
PERF_TEST_P(Sz_Type_KernelSz, Filters_Sobel, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1), Values(3, 5, 7, 9, 11, 13, 15)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::Sobel(d_src, d_dst, -1, 1, 1, d_buf, ksize);
TEST_CYCLE()
{
cv::gpu::Sobel(d_src, d_dst, -1, 1, 1, d_buf, ksize);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::Sobel(src, dst, -1, 1, 1, ksize);
TEST_CYCLE()
{
cv::Sobel(src, dst, -1, 1, 1, ksize);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Scharr
PERF_TEST_P(Sz_Type, Filters_Scharr, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::Scharr(d_src, d_dst, -1, 1, 0, d_buf);
TEST_CYCLE()
{
cv::gpu::Scharr(d_src, d_dst, -1, 1, 0, d_buf);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::Scharr(src, dst, -1, 1, 0);
TEST_CYCLE()
{
cv::Scharr(src, dst, -1, 1, 0);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// GaussianBlur
PERF_TEST_P(Sz_Type_KernelSz, Filters_GaussianBlur, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1), Values(3, 5, 7, 9, 11, 13, 15)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::GaussianBlur(d_src, d_dst, cv::Size(ksize, ksize), d_buf, 0.5);
TEST_CYCLE()
{
cv::gpu::GaussianBlur(d_src, d_dst, cv::Size(ksize, ksize), d_buf, 0.5);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::GaussianBlur(src, dst, cv::Size(ksize, ksize), 0.5);
TEST_CYCLE()
{
cv::GaussianBlur(src, dst, cv::Size(ksize, ksize), 0.5);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Laplacian
PERF_TEST_P(Sz_Type_KernelSz, Filters_Laplacian, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4), Values(1, 3)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::Laplacian(d_src, d_dst, -1, ksize);
TEST_CYCLE()
{
cv::gpu::Laplacian(d_src, d_dst, -1, ksize);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::Laplacian(src, dst, -1, ksize);
TEST_CYCLE()
{
cv::Laplacian(src, dst, -1, ksize);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Erode
PERF_TEST_P(Sz_Type, Filters_Erode, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat ker = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(3, 3));
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::erode(d_src, d_dst, ker, d_buf);
TEST_CYCLE()
{
cv::gpu::erode(d_src, d_dst, ker, d_buf);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::erode(src, dst, ker);
TEST_CYCLE()
{
cv::erode(src, dst, ker);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Dilate
PERF_TEST_P(Sz_Type, Filters_Dilate, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat ker = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(3, 3));
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::dilate(d_src, d_dst, ker, d_buf);
TEST_CYCLE()
{
cv::gpu::dilate(d_src, d_dst, ker, d_buf);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::dilate(src, dst, ker);
TEST_CYCLE()
{
cv::dilate(src, dst, ker);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// MorphologyEx
CV_ENUM(MorphOp, cv::MORPH_OPEN, cv::MORPH_CLOSE, cv::MORPH_GRADIENT, cv::MORPH_TOPHAT, cv::MORPH_BLACKHAT)
#define ALL_MORPH_OPS ValuesIn(MorphOp::all())
DEF_PARAM_TEST(Sz_Type_Op, cv::Size, MatType, MorphOp);
PERF_TEST_P(Sz_Type_Op, Filters_MorphologyEx, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4), ALL_MORPH_OPS))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int morphOp = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat ker = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(3, 3));
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf1;
cv::gpu::GpuMat d_buf2;
cv::gpu::morphologyEx(d_src, d_dst, morphOp, ker, d_buf1, d_buf2);
TEST_CYCLE()
{
cv::gpu::morphologyEx(d_src, d_dst, morphOp, ker, d_buf1, d_buf2);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::morphologyEx(src, dst, morphOp, ker);
TEST_CYCLE()
{
cv::morphologyEx(src, dst, morphOp, ker);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Filter2D
PERF_TEST_P(Sz_Type_KernelSz, Filters_Filter2D, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4), Values(3, 5, 7, 9, 11, 13, 15)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat kernel(ksize, ksize, CV_32FC1);
fillRandom(kernel, 0.0, 1.0);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::filter2D(d_src, d_dst, -1, kernel);
TEST_CYCLE()
{
cv::gpu::filter2D(d_src, d_dst, -1, kernel);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::filter2D(src, dst, -1, kernel);
TEST_CYCLE()
{
cv::filter2D(src, dst, -1, kernel);
}
CPU_SANITY_CHECK(dst);
}
}
} // namespace
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// Blur
DEF_PARAM_TEST(Sz_Type_KernelSz, cv::Size, MatType, int);
PERF_TEST_P(Sz_Type_KernelSz, Filters_Blur, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4), Values(3, 5, 7)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::blur(d_src, d_dst, cv::Size(ksize, ksize));
TEST_CYCLE()
{
cv::gpu::blur(d_src, d_dst, cv::Size(ksize, ksize));
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::blur(src, dst, cv::Size(ksize, ksize));
TEST_CYCLE()
{
cv::blur(src, dst, cv::Size(ksize, ksize));
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Sobel
PERF_TEST_P(Sz_Type_KernelSz, Filters_Sobel, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1), Values(3, 5, 7, 9, 11, 13, 15)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::Sobel(d_src, d_dst, -1, 1, 1, d_buf, ksize);
TEST_CYCLE()
{
cv::gpu::Sobel(d_src, d_dst, -1, 1, 1, d_buf, ksize);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::Sobel(src, dst, -1, 1, 1, ksize);
TEST_CYCLE()
{
cv::Sobel(src, dst, -1, 1, 1, ksize);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Scharr
PERF_TEST_P(Sz_Type, Filters_Scharr, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::Scharr(d_src, d_dst, -1, 1, 0, d_buf);
TEST_CYCLE()
{
cv::gpu::Scharr(d_src, d_dst, -1, 1, 0, d_buf);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::Scharr(src, dst, -1, 1, 0);
TEST_CYCLE()
{
cv::Scharr(src, dst, -1, 1, 0);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// GaussianBlur
PERF_TEST_P(Sz_Type_KernelSz, Filters_GaussianBlur, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1), Values(3, 5, 7, 9, 11, 13, 15)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::GaussianBlur(d_src, d_dst, cv::Size(ksize, ksize), d_buf, 0.5);
TEST_CYCLE()
{
cv::gpu::GaussianBlur(d_src, d_dst, cv::Size(ksize, ksize), d_buf, 0.5);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::GaussianBlur(src, dst, cv::Size(ksize, ksize), 0.5);
TEST_CYCLE()
{
cv::GaussianBlur(src, dst, cv::Size(ksize, ksize), 0.5);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Laplacian
PERF_TEST_P(Sz_Type_KernelSz, Filters_Laplacian, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4), Values(1, 3)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::Laplacian(d_src, d_dst, -1, ksize);
TEST_CYCLE()
{
cv::gpu::Laplacian(d_src, d_dst, -1, ksize);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::Laplacian(src, dst, -1, ksize);
TEST_CYCLE()
{
cv::Laplacian(src, dst, -1, ksize);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Erode
PERF_TEST_P(Sz_Type, Filters_Erode, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat ker = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(3, 3));
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::erode(d_src, d_dst, ker, d_buf);
TEST_CYCLE()
{
cv::gpu::erode(d_src, d_dst, ker, d_buf);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::erode(src, dst, ker);
TEST_CYCLE()
{
cv::erode(src, dst, ker);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Dilate
PERF_TEST_P(Sz_Type, Filters_Dilate, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat ker = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(3, 3));
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf;
cv::gpu::dilate(d_src, d_dst, ker, d_buf);
TEST_CYCLE()
{
cv::gpu::dilate(d_src, d_dst, ker, d_buf);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::dilate(src, dst, ker);
TEST_CYCLE()
{
cv::dilate(src, dst, ker);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// MorphologyEx
CV_ENUM(MorphOp, cv::MORPH_OPEN, cv::MORPH_CLOSE, cv::MORPH_GRADIENT, cv::MORPH_TOPHAT, cv::MORPH_BLACKHAT)
#define ALL_MORPH_OPS ValuesIn(MorphOp::all())
DEF_PARAM_TEST(Sz_Type_Op, cv::Size, MatType, MorphOp);
PERF_TEST_P(Sz_Type_Op, Filters_MorphologyEx, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4), ALL_MORPH_OPS))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int morphOp = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat ker = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(3, 3));
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::GpuMat d_buf1;
cv::gpu::GpuMat d_buf2;
cv::gpu::morphologyEx(d_src, d_dst, morphOp, ker, d_buf1, d_buf2);
TEST_CYCLE()
{
cv::gpu::morphologyEx(d_src, d_dst, morphOp, ker, d_buf1, d_buf2);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::morphologyEx(src, dst, morphOp, ker);
TEST_CYCLE()
{
cv::morphologyEx(src, dst, morphOp, ker);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// Filter2D
PERF_TEST_P(Sz_Type_KernelSz, Filters_Filter2D, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4), Values(3, 5, 7, 9, 11, 13, 15)))
{
declare.time(20.0);
cv::Size size = GET_PARAM(0);
int type = GET_PARAM(1);
int ksize = GET_PARAM(2);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat kernel(ksize, ksize, CV_32FC1);
fillRandom(kernel, 0.0, 1.0);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
cv::gpu::filter2D(d_src, d_dst, -1, kernel);
TEST_CYCLE()
{
cv::gpu::filter2D(d_src, d_dst, -1, kernel);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
cv::filter2D(src, dst, -1, kernel);
TEST_CYCLE()
{
cv::filter2D(src, dst, -1, kernel);
}
CPU_SANITY_CHECK(dst);
}
}
} // namespace

3846
modules/gpu/perf/perf_imgproc.cpp
View File

File diff suppressed because it is too large Load Diff

4
modules/gpu/perf/perf_labeling.cpp
View File

@@ -32,8 +32,8 @@ struct GreedyLabeling
return lo <= d && d <= hi;
}
private:
InInterval& operator=(const InInterval&);
private:
InInterval& operator=(const InInterval&);
};

370
modules/gpu/perf/perf_matop.cpp
View File

@@ -1,185 +1,185 @@
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// SetTo
PERF_TEST_P(Sz_Depth_Cn, MatOp_SetTo, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), GPU_CHANNELS_1_3_4))
{
cv::Size size = GET_PARAM(0);
int depth = GET_PARAM(1);
int channels = GET_PARAM(2);
int type = CV_MAKE_TYPE(depth, channels);
cv::Scalar val(1, 2, 3, 4);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(size, type);
d_src.setTo(val);
TEST_CYCLE()
{
d_src.setTo(val);
}
GPU_SANITY_CHECK(d_src);
}
else
{
cv::Mat src(size, type);
src.setTo(val);
TEST_CYCLE()
{
src.setTo(val);
}
CPU_SANITY_CHECK(src);
}
}
//////////////////////////////////////////////////////////////////////
// SetToMasked
PERF_TEST_P(Sz_Depth_Cn, MatOp_SetToMasked, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), GPU_CHANNELS_1_3_4))
{
cv::Size size = GET_PARAM(0);
int depth = GET_PARAM(1);
int channels = GET_PARAM(2);
int type = CV_MAKE_TYPE(depth, channels);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat mask(size, CV_8UC1);
fillRandom(mask, 0, 2);
cv::Scalar val(1, 2, 3, 4);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_mask(mask);
d_src.setTo(val, d_mask);
TEST_CYCLE()
{
d_src.setTo(val, d_mask);
}
GPU_SANITY_CHECK(d_src);
}
else
{
src.setTo(val, mask);
TEST_CYCLE()
{
src.setTo(val, mask);
}
CPU_SANITY_CHECK(src);
}
}
//////////////////////////////////////////////////////////////////////
// CopyToMasked
PERF_TEST_P(Sz_Depth_Cn, MatOp_CopyToMasked, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), GPU_CHANNELS_1_3_4))
{
cv::Size size = GET_PARAM(0);
int depth = GET_PARAM(1);
int channels = GET_PARAM(2);
int type = CV_MAKE_TYPE(depth, channels);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat mask(size, CV_8UC1);
fillRandom(mask, 0, 2);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_mask(mask);
cv::gpu::GpuMat d_dst;
d_src.copyTo(d_dst, d_mask);
TEST_CYCLE()
{
d_src.copyTo(d_dst, d_mask);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
src.copyTo(dst, mask);
TEST_CYCLE()
{
src.copyTo(dst, mask);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// ConvertTo
DEF_PARAM_TEST(Sz_2Depth, cv::Size, MatDepth, MatDepth);
PERF_TEST_P(Sz_2Depth, MatOp_ConvertTo, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), Values(CV_8U, CV_16U, CV_32F, CV_64F)))
{
cv::Size size = GET_PARAM(0);
int depth1 = GET_PARAM(1);
int depth2 = GET_PARAM(2);
cv::Mat src(size, depth1);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
d_src.convertTo(d_dst, depth2, 0.5, 1.0);
TEST_CYCLE()
{
d_src.convertTo(d_dst, depth2, 0.5, 1.0);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
src.convertTo(dst, depth2, 0.5, 1.0);
TEST_CYCLE()
{
src.convertTo(dst, depth2, 0.5, 1.0);
}
CPU_SANITY_CHECK(dst);
}
}
} // namespace
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
//////////////////////////////////////////////////////////////////////
// SetTo
PERF_TEST_P(Sz_Depth_Cn, MatOp_SetTo, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), GPU_CHANNELS_1_3_4))
{
cv::Size size = GET_PARAM(0);
int depth = GET_PARAM(1);
int channels = GET_PARAM(2);
int type = CV_MAKE_TYPE(depth, channels);
cv::Scalar val(1, 2, 3, 4);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(size, type);
d_src.setTo(val);
TEST_CYCLE()
{
d_src.setTo(val);
}
GPU_SANITY_CHECK(d_src);
}
else
{
cv::Mat src(size, type);
src.setTo(val);
TEST_CYCLE()
{
src.setTo(val);
}
CPU_SANITY_CHECK(src);
}
}
//////////////////////////////////////////////////////////////////////
// SetToMasked
PERF_TEST_P(Sz_Depth_Cn, MatOp_SetToMasked, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), GPU_CHANNELS_1_3_4))
{
cv::Size size = GET_PARAM(0);
int depth = GET_PARAM(1);
int channels = GET_PARAM(2);
int type = CV_MAKE_TYPE(depth, channels);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat mask(size, CV_8UC1);
fillRandom(mask, 0, 2);
cv::Scalar val(1, 2, 3, 4);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_mask(mask);
d_src.setTo(val, d_mask);
TEST_CYCLE()
{
d_src.setTo(val, d_mask);
}
GPU_SANITY_CHECK(d_src);
}
else
{
src.setTo(val, mask);
TEST_CYCLE()
{
src.setTo(val, mask);
}
CPU_SANITY_CHECK(src);
}
}
//////////////////////////////////////////////////////////////////////
// CopyToMasked
PERF_TEST_P(Sz_Depth_Cn, MatOp_CopyToMasked, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), GPU_CHANNELS_1_3_4))
{
cv::Size size = GET_PARAM(0);
int depth = GET_PARAM(1);
int channels = GET_PARAM(2);
int type = CV_MAKE_TYPE(depth, channels);
cv::Mat src(size, type);
fillRandom(src);
cv::Mat mask(size, CV_8UC1);
fillRandom(mask, 0, 2);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_mask(mask);
cv::gpu::GpuMat d_dst;
d_src.copyTo(d_dst, d_mask);
TEST_CYCLE()
{
d_src.copyTo(d_dst, d_mask);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
src.copyTo(dst, mask);
TEST_CYCLE()
{
src.copyTo(dst, mask);
}
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// ConvertTo
DEF_PARAM_TEST(Sz_2Depth, cv::Size, MatDepth, MatDepth);
PERF_TEST_P(Sz_2Depth, MatOp_ConvertTo, Combine(GPU_TYPICAL_MAT_SIZES, Values(CV_8U, CV_16U, CV_32F, CV_64F), Values(CV_8U, CV_16U, CV_32F, CV_64F)))
{
cv::Size size = GET_PARAM(0);
int depth1 = GET_PARAM(1);
int depth2 = GET_PARAM(2);
cv::Mat src(size, depth1);
fillRandom(src);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_dst;
d_src.convertTo(d_dst, depth2, 0.5, 1.0);
TEST_CYCLE()
{
d_src.convertTo(d_dst, depth2, 0.5, 1.0);
}
GPU_SANITY_CHECK(d_dst);
}
else
{
cv::Mat dst;
src.convertTo(dst, depth2, 0.5, 1.0);
TEST_CYCLE()
{
src.convertTo(dst, depth2, 0.5, 1.0);
}
CPU_SANITY_CHECK(dst);
}
}
} // namespace

368
modules/gpu/perf/perf_objdetect.cpp
View File

@@ -1,184 +1,184 @@
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
///////////////////////////////////////////////////////////////
// HOG
DEF_PARAM_TEST_1(Image, string);
PERF_TEST_P(Image, ObjDetect_HOG, Values<string>("gpu/hog/road.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
std::vector<cv::Rect> found_locations;
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_img(img);
cv::gpu::HOGDescriptor d_hog;
d_hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
d_hog.detectMultiScale(d_img, found_locations);
TEST_CYCLE()
{
d_hog.detectMultiScale(d_img, found_locations);
}
}
else
{
cv::HOGDescriptor hog;
hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
hog.detectMultiScale(img, found_locations);
TEST_CYCLE()
{
hog.detectMultiScale(img, found_locations);
}
}
SANITY_CHECK(found_locations);
}
//===========test for CalTech data =============//
DEF_PARAM_TEST_1(HOG, string);
PERF_TEST_P(HOG, CalTech, Values<string>("gpu/caltech/image_00000009_0.png", "gpu/caltech/image_00000032_0.png",
"gpu/caltech/image_00000165_0.png", "gpu/caltech/image_00000261_0.png", "gpu/caltech/image_00000469_0.png",
"gpu/caltech/image_00000527_0.png", "gpu/caltech/image_00000574_0.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
std::vector<cv::Rect> found_locations;
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_img(img);
cv::gpu::HOGDescriptor d_hog;
d_hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
d_hog.detectMultiScale(d_img, found_locations);
TEST_CYCLE()
{
d_hog.detectMultiScale(d_img, found_locations);
}
}
else
{
cv::HOGDescriptor hog;
hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
hog.detectMultiScale(img, found_locations);
TEST_CYCLE()
{
hog.detectMultiScale(img, found_locations);
}
}
SANITY_CHECK(found_locations);
}
///////////////////////////////////////////////////////////////
// HaarClassifier
typedef pair<string, string> pair_string;
DEF_PARAM_TEST_1(ImageAndCascade, pair_string);
PERF_TEST_P(ImageAndCascade, ObjDetect_HaarClassifier,
Values<pair_string>(make_pair("gpu/haarcascade/group_1_640x480_VGA.pgm", "gpu/perf/haarcascade_frontalface_alt.xml")))
{
cv::Mat img = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::CascadeClassifier_GPU d_cascade;
ASSERT_TRUE(d_cascade.load(perf::TestBase::getDataPath(GetParam().second)));
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_objects_buffer;
d_cascade.detectMultiScale(d_img, d_objects_buffer);
TEST_CYCLE()
{
d_cascade.detectMultiScale(d_img, d_objects_buffer);
}
GPU_SANITY_CHECK(d_objects_buffer);
}
else
{
cv::CascadeClassifier cascade;
ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/perf/haarcascade_frontalface_alt.xml")));
std::vector<cv::Rect> rects;
cascade.detectMultiScale(img, rects);
TEST_CYCLE()
{
cascade.detectMultiScale(img, rects);
}
CPU_SANITY_CHECK(rects);
}
}
///////////////////////////////////////////////////////////////
// LBP cascade
PERF_TEST_P(ImageAndCascade, ObjDetect_LBPClassifier,
Values<pair_string>(make_pair("gpu/haarcascade/group_1_640x480_VGA.pgm", "gpu/lbpcascade/lbpcascade_frontalface.xml")))
{
cv::Mat img = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::CascadeClassifier_GPU d_cascade;
ASSERT_TRUE(d_cascade.load(perf::TestBase::getDataPath(GetParam().second)));
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_gpu_rects;
d_cascade.detectMultiScale(d_img, d_gpu_rects);
TEST_CYCLE()
{
d_cascade.detectMultiScale(d_img, d_gpu_rects);
}
GPU_SANITY_CHECK(d_gpu_rects);
}
else
{
cv::CascadeClassifier cascade;
ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/lbpcascade/lbpcascade_frontalface.xml")));
std::vector<cv::Rect> rects;
cascade.detectMultiScale(img, rects);
TEST_CYCLE()
{
cascade.detectMultiScale(img, rects);
}
CPU_SANITY_CHECK(rects);
}
}
} // namespace
#include "perf_precomp.hpp"
using namespace std;
using namespace testing;
namespace {
///////////////////////////////////////////////////////////////
// HOG
DEF_PARAM_TEST_1(Image, string);
PERF_TEST_P(Image, ObjDetect_HOG, Values<string>("gpu/hog/road.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
std::vector<cv::Rect> found_locations;
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_img(img);
cv::gpu::HOGDescriptor d_hog;
d_hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
d_hog.detectMultiScale(d_img, found_locations);
TEST_CYCLE()
{
d_hog.detectMultiScale(d_img, found_locations);
}
}
else
{
cv::HOGDescriptor hog;
hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
hog.detectMultiScale(img, found_locations);
TEST_CYCLE()
{
hog.detectMultiScale(img, found_locations);
}
}
SANITY_CHECK(found_locations);
}
//===========test for CalTech data =============//
DEF_PARAM_TEST_1(HOG, string);
PERF_TEST_P(HOG, CalTech, Values<string>("gpu/caltech/image_00000009_0.png", "gpu/caltech/image_00000032_0.png",
"gpu/caltech/image_00000165_0.png", "gpu/caltech/image_00000261_0.png", "gpu/caltech/image_00000469_0.png",
"gpu/caltech/image_00000527_0.png", "gpu/caltech/image_00000574_0.png"))
{
cv::Mat img = readImage(GetParam(), cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
std::vector<cv::Rect> found_locations;
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_img(img);
cv::gpu::HOGDescriptor d_hog;
d_hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
d_hog.detectMultiScale(d_img, found_locations);
TEST_CYCLE()
{
d_hog.detectMultiScale(d_img, found_locations);
}
}
else
{
cv::HOGDescriptor hog;
hog.setSVMDetector(cv::gpu::HOGDescriptor::getDefaultPeopleDetector());
hog.detectMultiScale(img, found_locations);
TEST_CYCLE()
{
hog.detectMultiScale(img, found_locations);
}
}
SANITY_CHECK(found_locations);
}
///////////////////////////////////////////////////////////////
// HaarClassifier
typedef pair<string, string> pair_string;
DEF_PARAM_TEST_1(ImageAndCascade, pair_string);
PERF_TEST_P(ImageAndCascade, ObjDetect_HaarClassifier,
Values<pair_string>(make_pair("gpu/haarcascade/group_1_640x480_VGA.pgm", "gpu/perf/haarcascade_frontalface_alt.xml")))
{
cv::Mat img = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::CascadeClassifier_GPU d_cascade;
ASSERT_TRUE(d_cascade.load(perf::TestBase::getDataPath(GetParam().second)));
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_objects_buffer;
d_cascade.detectMultiScale(d_img, d_objects_buffer);
TEST_CYCLE()
{
d_cascade.detectMultiScale(d_img, d_objects_buffer);
}
GPU_SANITY_CHECK(d_objects_buffer);
}
else
{
cv::CascadeClassifier cascade;
ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/perf/haarcascade_frontalface_alt.xml")));
std::vector<cv::Rect> rects;
cascade.detectMultiScale(img, rects);
TEST_CYCLE()
{
cascade.detectMultiScale(img, rects);
}
CPU_SANITY_CHECK(rects);
}
}
///////////////////////////////////////////////////////////////
// LBP cascade
PERF_TEST_P(ImageAndCascade, ObjDetect_LBPClassifier,
Values<pair_string>(make_pair("gpu/haarcascade/group_1_640x480_VGA.pgm", "gpu/lbpcascade/lbpcascade_frontalface.xml")))
{
cv::Mat img = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(img.empty());
if (PERF_RUN_GPU())
{
cv::gpu::CascadeClassifier_GPU d_cascade;
ASSERT_TRUE(d_cascade.load(perf::TestBase::getDataPath(GetParam().second)));
cv::gpu::GpuMat d_img(img);
cv::gpu::GpuMat d_gpu_rects;
d_cascade.detectMultiScale(d_img, d_gpu_rects);
TEST_CYCLE()
{
d_cascade.detectMultiScale(d_img, d_gpu_rects);
}
GPU_SANITY_CHECK(d_gpu_rects);
}
else
{
cv::CascadeClassifier cascade;
ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/lbpcascade/lbpcascade_frontalface.xml")));
std::vector<cv::Rect> rects;
cascade.detectMultiScale(img, rects);
TEST_CYCLE()
{
cascade.detectMultiScale(img, rects);
}
CPU_SANITY_CHECK(rects);
}
}
} // namespace

74
modules/gpu/perf/perf_precomp.hpp
View File

@@ -1,37 +1,37 @@
#ifdef __GNUC__
# pragma GCC diagnostic ignored "-Wmissing-declarations"
# pragma GCC diagnostic ignored "-Wmissing-prototypes" //OSX
#endif
#ifndef __OPENCV_PERF_PRECOMP_HPP__
#define __OPENCV_PERF_PRECOMP_HPP__
#include <cstdio>
#include <iostream>
#include "cvconfig.h"
#ifdef HAVE_CUDA
#include <cuda_runtime.h>
#endif
#include "opencv2/ts/ts.hpp"
#include "opencv2/ts/ts_perf.hpp"
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/gpu/gpu.hpp"
#include "opencv2/calib3d/calib3d.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/video/video.hpp"
#include "opencv2/nonfree/nonfree.hpp"
#include "opencv2/legacy/legacy.hpp"
#include "opencv2/photo/photo.hpp"
#include "utility.hpp"
#ifdef GTEST_CREATE_SHARED_LIBRARY
#error no modules except ts should have GTEST_CREATE_SHARED_LIBRARY defined
#endif
#endif
#ifdef __GNUC__
# pragma GCC diagnostic ignored "-Wmissing-declarations"
# pragma GCC diagnostic ignored "-Wmissing-prototypes" //OSX
#endif
#ifndef __OPENCV_PERF_PRECOMP_HPP__
#define __OPENCV_PERF_PRECOMP_HPP__
#include <cstdio>
#include <iostream>
#include "cvconfig.h"
#ifdef HAVE_CUDA
#include <cuda_runtime.h>
#endif
#include "opencv2/ts/ts.hpp"
#include "opencv2/ts/ts_perf.hpp"
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/gpu/gpu.hpp"
#include "opencv2/calib3d/calib3d.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/video/video.hpp"
#include "opencv2/nonfree/nonfree.hpp"
#include "opencv2/legacy/legacy.hpp"
#include "opencv2/photo/photo.hpp"
#include "utility.hpp"
#ifdef GTEST_CREATE_SHARED_LIBRARY
#error no modules except ts should have GTEST_CREATE_SHARED_LIBRARY defined
#endif
#endif

1962
modules/gpu/perf/perf_video.cpp
View File

File diff suppressed because it is too large Load Diff

380
modules/gpu/perf/utility.cpp
View File

@@ -1,191 +1,191 @@
#include "perf_precomp.hpp"
using namespace std;
using namespace cv;
using namespace cv::gpu;
void fillRandom(Mat& m, double a, double b)
{
RNG rng(123456789);
rng.fill(m, RNG::UNIFORM, Scalar::all(a), Scalar::all(b));
}
Mat readImage(const string& fileName, int flags)
{
return imread(perf::TestBase::getDataPath(fileName), flags);
}
void PrintTo(const CvtColorInfo& info, ostream* os)
{
static const char* str[] =
{
"BGR2BGRA",
"BGRA2BGR",
"BGR2RGBA",
"RGBA2BGR",
"BGR2RGB",
"BGRA2RGBA",
"BGR2GRAY",
"RGB2GRAY",
"GRAY2BGR",
"GRAY2BGRA",
"BGRA2GRAY",
"RGBA2GRAY",
"BGR2BGR565",
"RGB2BGR565",
"BGR5652BGR",
"BGR5652RGB",
"BGRA2BGR565",
"RGBA2BGR565",
"BGR5652BGRA",
"BGR5652RGBA",
"GRAY2BGR565",
"BGR5652GRAY",
"BGR2BGR555",
"RGB2BGR555",
"BGR5552BGR",
"BGR5552RGB",
"BGRA2BGR555",
"RGBA2BGR555",
"BGR5552BGRA",
"BGR5552RGBA",
"GRAY2BGR555",
"BGR5552GRAY",
"BGR2XYZ",
"RGB2XYZ",
"XYZ2BGR",
"XYZ2RGB",
"BGR2YCrCb",
"RGB2YCrCb",
"YCrCb2BGR",
"YCrCb2RGB",
"BGR2HSV",
"RGB2HSV",
"",
"",
"BGR2Lab",
"RGB2Lab",
"BayerBG2BGR",
"BayerGB2BGR",
"BayerRG2BGR",
"BayerGR2BGR",
"BGR2Luv",
"RGB2Luv",
"BGR2HLS",
"RGB2HLS",
"HSV2BGR",
"HSV2RGB",
"Lab2BGR",
"Lab2RGB",
"Luv2BGR",
"Luv2RGB",
"HLS2BGR",
"HLS2RGB",
"BayerBG2BGR_VNG",
"BayerGB2BGR_VNG",
"BayerRG2BGR_VNG",
"BayerGR2BGR_VNG",
"BGR2HSV_FULL",
"RGB2HSV_FULL",
"BGR2HLS_FULL",
"RGB2HLS_FULL",
"HSV2BGR_FULL",
"HSV2RGB_FULL",
"HLS2BGR_FULL",
"HLS2RGB_FULL",
"LBGR2Lab",
"LRGB2Lab",
"LBGR2Luv",
"LRGB2Luv",
"Lab2LBGR",
"Lab2LRGB",
"Luv2LBGR",
"Luv2LRGB",
"BGR2YUV",
"RGB2YUV",
"YUV2BGR",
"YUV2RGB",
"BayerBG2GRAY",
"BayerGB2GRAY",
"BayerRG2GRAY",
"BayerGR2GRAY",
//YUV 4:2:0 formats family
"YUV2RGB_NV12",
"YUV2BGR_NV12",
"YUV2RGB_NV21",
"YUV2BGR_NV21",
"YUV2RGBA_NV12",
"YUV2BGRA_NV12",
"YUV2RGBA_NV21",
"YUV2BGRA_NV21",
"YUV2RGB_YV12",
"YUV2BGR_YV12",
"YUV2RGB_IYUV",
"YUV2BGR_IYUV",
"YUV2RGBA_YV12",
"YUV2BGRA_YV12",
"YUV2RGBA_IYUV",
"YUV2BGRA_IYUV",
"YUV2GRAY_420",
//YUV 4:2:2 formats family
"YUV2RGB_UYVY",
"YUV2BGR_UYVY",
"YUV2RGB_VYUY",
"YUV2BGR_VYUY",
"YUV2RGBA_UYVY",
"YUV2BGRA_UYVY",
"YUV2RGBA_VYUY",
"YUV2BGRA_VYUY",
"YUV2RGB_YUY2",
"YUV2BGR_YUY2",
"YUV2RGB_YVYU",
"YUV2BGR_YVYU",
"YUV2RGBA_YUY2",
"YUV2BGRA_YUY2",
"YUV2RGBA_YVYU",
"YUV2BGRA_YVYU",
"YUV2GRAY_UYVY",
"YUV2GRAY_YUY2",
// alpha premultiplication
"RGBA2mRGBA",
"mRGBA2RGBA",
"COLORCVT_MAX"
};
*os << str[info.code];
#include "perf_precomp.hpp"
using namespace std;
using namespace cv;
using namespace cv::gpu;
void fillRandom(Mat& m, double a, double b)
{
RNG rng(123456789);
rng.fill(m, RNG::UNIFORM, Scalar::all(a), Scalar::all(b));
}
Mat readImage(const string& fileName, int flags)
{
return imread(perf::TestBase::getDataPath(fileName), flags);
}
void PrintTo(const CvtColorInfo& info, ostream* os)
{
static const char* str[] =
{
"BGR2BGRA",
"BGRA2BGR",
"BGR2RGBA",
"RGBA2BGR",
"BGR2RGB",
"BGRA2RGBA",
"BGR2GRAY",
"RGB2GRAY",
"GRAY2BGR",
"GRAY2BGRA",
"BGRA2GRAY",
"RGBA2GRAY",
"BGR2BGR565",
"RGB2BGR565",
"BGR5652BGR",
"BGR5652RGB",
"BGRA2BGR565",
"RGBA2BGR565",
"BGR5652BGRA",
"BGR5652RGBA",
"GRAY2BGR565",
"BGR5652GRAY",
"BGR2BGR555",
"RGB2BGR555",
"BGR5552BGR",
"BGR5552RGB",
"BGRA2BGR555",
"RGBA2BGR555",
"BGR5552BGRA",
"BGR5552RGBA",
"GRAY2BGR555",
"BGR5552GRAY",
"BGR2XYZ",
"RGB2XYZ",
"XYZ2BGR",
"XYZ2RGB",
"BGR2YCrCb",
"RGB2YCrCb",
"YCrCb2BGR",
"YCrCb2RGB",
"BGR2HSV",
"RGB2HSV",
"",
"",
"BGR2Lab",
"RGB2Lab",
"BayerBG2BGR",
"BayerGB2BGR",
"BayerRG2BGR",
"BayerGR2BGR",
"BGR2Luv",
"RGB2Luv",
"BGR2HLS",
"RGB2HLS",
"HSV2BGR",
"HSV2RGB",
"Lab2BGR",
"Lab2RGB",
"Luv2BGR",
"Luv2RGB",
"HLS2BGR",
"HLS2RGB",
"BayerBG2BGR_VNG",
"BayerGB2BGR_VNG",
"BayerRG2BGR_VNG",
"BayerGR2BGR_VNG",
"BGR2HSV_FULL",
"RGB2HSV_FULL",
"BGR2HLS_FULL",
"RGB2HLS_FULL",
"HSV2BGR_FULL",
"HSV2RGB_FULL",
"HLS2BGR_FULL",
"HLS2RGB_FULL",
"LBGR2Lab",
"LRGB2Lab",
"LBGR2Luv",
"LRGB2Luv",
"Lab2LBGR",
"Lab2LRGB",
"Luv2LBGR",
"Luv2LRGB",
"BGR2YUV",
"RGB2YUV",
"YUV2BGR",
"YUV2RGB",
"BayerBG2GRAY",
"BayerGB2GRAY",
"BayerRG2GRAY",
"BayerGR2GRAY",
//YUV 4:2:0 formats family
"YUV2RGB_NV12",
"YUV2BGR_NV12",
"YUV2RGB_NV21",
"YUV2BGR_NV21",
"YUV2RGBA_NV12",
"YUV2BGRA_NV12",
"YUV2RGBA_NV21",
"YUV2BGRA_NV21",
"YUV2RGB_YV12",
"YUV2BGR_YV12",
"YUV2RGB_IYUV",
"YUV2BGR_IYUV",
"YUV2RGBA_YV12",
"YUV2BGRA_YV12",
"YUV2RGBA_IYUV",
"YUV2BGRA_IYUV",
"YUV2GRAY_420",
//YUV 4:2:2 formats family
"YUV2RGB_UYVY",
"YUV2BGR_UYVY",
"YUV2RGB_VYUY",
"YUV2BGR_VYUY",
"YUV2RGBA_UYVY",
"YUV2BGRA_UYVY",
"YUV2RGBA_VYUY",
"YUV2BGRA_VYUY",
"YUV2RGB_YUY2",
"YUV2BGR_YUY2",
"YUV2RGB_YVYU",
"YUV2BGR_YVYU",
"YUV2RGBA_YUY2",
"YUV2BGRA_YUY2",
"YUV2RGBA_YVYU",
"YUV2BGRA_YVYU",
"YUV2GRAY_UYVY",
"YUV2GRAY_YUY2",
// alpha premultiplication
"RGBA2mRGBA",
"mRGBA2RGBA",
"COLORCVT_MAX"
};
*os << str[info.code];
}

168
modules/gpu/perf/utility.hpp
View File

@@ -1,84 +1,84 @@
#ifndef __OPENCV_PERF_GPU_UTILITY_HPP__
#define __OPENCV_PERF_GPU_UTILITY_HPP__
#include "opencv2/core/core.hpp"
#include "opencv2/core/gpumat.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/ts/ts_perf.hpp"
void fillRandom(cv::Mat& m, double a = 0.0, double b = 255.0);
cv::Mat readImage(const std::string& fileName, int flags = cv::IMREAD_COLOR);
using perf::MatType;
using perf::MatDepth;
CV_ENUM(BorderMode, cv::BORDER_REFLECT101, cv::BORDER_REPLICATE, cv::BORDER_CONSTANT, cv::BORDER_REFLECT, cv::BORDER_WRAP)
#define ALL_BORDER_MODES testing::ValuesIn(BorderMode::all())
CV_ENUM(Interpolation, cv::INTER_NEAREST, cv::INTER_LINEAR, cv::INTER_CUBIC, cv::INTER_AREA)
#define ALL_INTERPOLATIONS testing::ValuesIn(Interpolation::all())
CV_ENUM(NormType, cv::NORM_INF, cv::NORM_L1, cv::NORM_L2, cv::NORM_HAMMING)
const int Gray = 1, TwoChannel = 2, BGR = 3, BGRA = 4;
CV_ENUM(MatCn, Gray, TwoChannel, BGR, BGRA)
#define GPU_CHANNELS_1_3_4 testing::Values(Gray, BGR, BGRA)
#define GPU_CHANNELS_1_3 testing::Values(Gray, BGR)
struct CvtColorInfo
{
int scn;
int dcn;
int code;
explicit CvtColorInfo(int scn_=0, int dcn_=0, int code_=0) : scn(scn_), dcn(dcn_), code(code_) {}
};
void PrintTo(const CvtColorInfo& info, std::ostream* os);
#define GET_PARAM(k) std::tr1::get< k >(GetParam())
#define DEF_PARAM_TEST(name, ...) typedef ::perf::TestBaseWithParam< std::tr1::tuple< __VA_ARGS__ > > name
#define DEF_PARAM_TEST_1(name, param_type) typedef ::perf::TestBaseWithParam< param_type > name
DEF_PARAM_TEST_1(Sz, cv::Size);
typedef perf::Size_MatType Sz_Type;
DEF_PARAM_TEST(Sz_Depth, cv::Size, MatDepth);
DEF_PARAM_TEST(Sz_Depth_Cn, cv::Size, MatDepth, MatCn);
#define GPU_TYPICAL_MAT_SIZES testing::Values(perf::sz720p, perf::szSXGA, perf::sz1080p)
#define GPU_SANITY_CHECK(dmat, ...) \
do{ \
cv::Mat d##dmat(dmat); \
SANITY_CHECK(d##dmat, ## __VA_ARGS__); \
} while(0)
#define CPU_SANITY_CHECK(cmat, ...) \
do{ \
SANITY_CHECK(cmat, ## __VA_ARGS__); \
} while(0)
#define GPU_SANITY_CHECK_KEYPOINTS(alg, dmat, ...) \
do{ \
cv::Mat d##dmat(dmat); \
cv::Mat __pt_x = d##dmat.row(cv::gpu::alg##_GPU::X_ROW); \
cv::Mat __pt_y = d##dmat.row(cv::gpu::alg##_GPU::Y_ROW); \
cv::Mat __angle = d##dmat.row(cv::gpu::alg##_GPU::ANGLE_ROW); \
cv::Mat __octave = d##dmat.row(cv::gpu::alg##_GPU::OCTAVE_ROW); \
cv::Mat __size = d##dmat.row(cv::gpu::alg##_GPU::SIZE_ROW); \
::perf::Regression::add(this, std::string(#dmat) + "-pt-x-row", __pt_x, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "-pt-y-row", __pt_y, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "-angle-row", __angle, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "octave-row", __octave, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "-pt-size-row", __size, ## __VA_ARGS__); \
} while(0)
#define GPU_SANITY_CHECK_RESPONSE(alg, dmat, ...) \
do{ \
cv::Mat d##dmat(dmat); \
cv::Mat __response = d##dmat.row(cv::gpu::alg##_GPU::RESPONSE_ROW); \
::perf::Regression::add(this, std::string(#dmat) + "-response-row", __response, ## __VA_ARGS__); \
} while(0)
#define FAIL_NO_CPU() FAIL() << "No such CPU implementation analogy"
#endif // __OPENCV_PERF_GPU_UTILITY_HPP__
#ifndef __OPENCV_PERF_GPU_UTILITY_HPP__
#define __OPENCV_PERF_GPU_UTILITY_HPP__
#include "opencv2/core/core.hpp"
#include "opencv2/core/gpumat.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/ts/ts_perf.hpp"
void fillRandom(cv::Mat& m, double a = 0.0, double b = 255.0);
cv::Mat readImage(const std::string& fileName, int flags = cv::IMREAD_COLOR);
using perf::MatType;
using perf::MatDepth;
CV_ENUM(BorderMode, cv::BORDER_REFLECT101, cv::BORDER_REPLICATE, cv::BORDER_CONSTANT, cv::BORDER_REFLECT, cv::BORDER_WRAP)
#define ALL_BORDER_MODES testing::ValuesIn(BorderMode::all())
CV_ENUM(Interpolation, cv::INTER_NEAREST, cv::INTER_LINEAR, cv::INTER_CUBIC, cv::INTER_AREA)
#define ALL_INTERPOLATIONS testing::ValuesIn(Interpolation::all())
CV_ENUM(NormType, cv::NORM_INF, cv::NORM_L1, cv::NORM_L2, cv::NORM_HAMMING)
const int Gray = 1, TwoChannel = 2, BGR = 3, BGRA = 4;
CV_ENUM(MatCn, Gray, TwoChannel, BGR, BGRA)
#define GPU_CHANNELS_1_3_4 testing::Values(Gray, BGR, BGRA)
#define GPU_CHANNELS_1_3 testing::Values(Gray, BGR)
struct CvtColorInfo
{
int scn;
int dcn;
int code;
explicit CvtColorInfo(int scn_=0, int dcn_=0, int code_=0) : scn(scn_), dcn(dcn_), code(code_) {}
};
void PrintTo(const CvtColorInfo& info, std::ostream* os);
#define GET_PARAM(k) std::tr1::get< k >(GetParam())
#define DEF_PARAM_TEST(name, ...) typedef ::perf::TestBaseWithParam< std::tr1::tuple< __VA_ARGS__ > > name
#define DEF_PARAM_TEST_1(name, param_type) typedef ::perf::TestBaseWithParam< param_type > name
DEF_PARAM_TEST_1(Sz, cv::Size);
typedef perf::Size_MatType Sz_Type;
DEF_PARAM_TEST(Sz_Depth, cv::Size, MatDepth);
DEF_PARAM_TEST(Sz_Depth_Cn, cv::Size, MatDepth, MatCn);
#define GPU_TYPICAL_MAT_SIZES testing::Values(perf::sz720p, perf::szSXGA, perf::sz1080p)
#define GPU_SANITY_CHECK(dmat, ...) \
do{ \
cv::Mat d##dmat(dmat); \
SANITY_CHECK(d##dmat, ## __VA_ARGS__); \
} while(0)
#define CPU_SANITY_CHECK(cmat, ...) \
do{ \
SANITY_CHECK(cmat, ## __VA_ARGS__); \
} while(0)
#define GPU_SANITY_CHECK_KEYPOINTS(alg, dmat, ...) \
do{ \
cv::Mat d##dmat(dmat); \
cv::Mat __pt_x = d##dmat.row(cv::gpu::alg##_GPU::X_ROW); \
cv::Mat __pt_y = d##dmat.row(cv::gpu::alg##_GPU::Y_ROW); \
cv::Mat __angle = d##dmat.row(cv::gpu::alg##_GPU::ANGLE_ROW); \
cv::Mat __octave = d##dmat.row(cv::gpu::alg##_GPU::OCTAVE_ROW); \
cv::Mat __size = d##dmat.row(cv::gpu::alg##_GPU::SIZE_ROW); \
::perf::Regression::add(this, std::string(#dmat) + "-pt-x-row", __pt_x, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "-pt-y-row", __pt_y, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "-angle-row", __angle, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "octave-row", __octave, ## __VA_ARGS__); \
::perf::Regression::add(this, std::string(#dmat) + "-pt-size-row", __size, ## __VA_ARGS__); \
} while(0)
#define GPU_SANITY_CHECK_RESPONSE(alg, dmat, ...) \
do{ \
cv::Mat d##dmat(dmat); \
cv::Mat __response = d##dmat.row(cv::gpu::alg##_GPU::RESPONSE_ROW); \
::perf::Regression::add(this, std::string(#dmat) + "-response-row", __response, ## __VA_ARGS__); \
} while(0)
#define FAIL_NO_CPU() FAIL() << "No such CPU implementation analogy"
#endif // __OPENCV_PERF_GPU_UTILITY_HPP__

1056
modules/gpu/src/arithm.cpp
View File

File diff suppressed because it is too large Load Diff

316
modules/gpu/src/bilateral_filter.cpp
View File

@@ -1,158 +1,158 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int, int, int) { throw_nogpu(); }
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int, int, int, float, float, float) { throw_nogpu(); }
void cv::gpu::DisparityBilateralFilter::operator()(const GpuMat&, const GpuMat&, GpuMat&, Stream&) { throw_nogpu(); }
#else /* !defined (HAVE_CUDA) */
namespace cv { namespace gpu { namespace device
{
namespace disp_bilateral_filter
{
void disp_load_constants(float* table_color, PtrStepSzf table_space, int ndisp, int radius, short edge_disc, short max_disc);
template<typename T>
void disp_bilateral_filter(PtrStepSz<T> disp, PtrStepSzb img, int channels, int iters, cudaStream_t stream);
}
}}}
using namespace ::cv::gpu::device::disp_bilateral_filter;
namespace
{
const float DEFAULT_EDGE_THRESHOLD = 0.1f;
const float DEFAULT_MAX_DISC_THRESHOLD = 0.2f;
const float DEFAULT_SIGMA_RANGE = 10.0f;
inline void calc_color_weighted_table(GpuMat& table_color, float sigma_range, int len)
{
Mat cpu_table_color(1, len, CV_32F);
float* line = cpu_table_color.ptr<float>();
for(int i = 0; i < len; i++)
line[i] = static_cast<float>(std::exp(-double(i * i) / (2 * sigma_range * sigma_range)));
table_color.upload(cpu_table_color);
}
inline void calc_space_weighted_filter(GpuMat& table_space, int win_size, float dist_space)
{
int half = (win_size >> 1);
Mat cpu_table_space(half + 1, half + 1, CV_32F);
for (int y = 0; y <= half; ++y)
{
float* row = cpu_table_space.ptr<float>(y);
for (int x = 0; x <= half; ++x)
row[x] = exp(-sqrt(float(y * y) + float(x * x)) / dist_space);
}
table_space.upload(cpu_table_space);
}
template <typename T>
void disp_bilateral_filter_operator(int ndisp, int radius, int iters, float edge_threshold,float max_disc_threshold,
GpuMat& table_color, GpuMat& table_space,
const GpuMat& disp, const GpuMat& img, GpuMat& dst, Stream& stream)
{
short edge_disc = max<short>(short(1), short(ndisp * edge_threshold + 0.5));
short max_disc = short(ndisp * max_disc_threshold + 0.5);
disp_load_constants(table_color.ptr<float>(), table_space, ndisp, radius, edge_disc, max_disc);
if (&dst != &disp)
{
if (stream)
stream.enqueueCopy(disp, dst);
else
disp.copyTo(dst);
}
disp_bilateral_filter<T>(dst, img, img.channels(), iters, StreamAccessor::getStream(stream));
}
typedef void (*bilateral_filter_operator_t)(int ndisp, int radius, int iters, float edge_threshold, float max_disc_threshold,
GpuMat& table_color, GpuMat& table_space,
const GpuMat& disp, const GpuMat& img, GpuMat& dst, Stream& stream);
const bilateral_filter_operator_t operators[] =
{disp_bilateral_filter_operator<unsigned char>, 0, 0, disp_bilateral_filter_operator<short>, 0, 0, 0, 0};
}
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int ndisp_, int radius_, int iters_)
: ndisp(ndisp_), radius(radius_), iters(iters_), edge_threshold(DEFAULT_EDGE_THRESHOLD), max_disc_threshold(DEFAULT_MAX_DISC_THRESHOLD),
sigma_range(DEFAULT_SIGMA_RANGE)
{
calc_color_weighted_table(table_color, sigma_range, 255);
calc_space_weighted_filter(table_space, radius * 2 + 1, radius + 1.0f);
}
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int ndisp_, int radius_, int iters_, float edge_threshold_,
float max_disc_threshold_, float sigma_range_)
: ndisp(ndisp_), radius(radius_), iters(iters_), edge_threshold(edge_threshold_), max_disc_threshold(max_disc_threshold_),
sigma_range(sigma_range_)
{
calc_color_weighted_table(table_color, sigma_range, 255);
calc_space_weighted_filter(table_space, radius * 2 + 1, radius + 1.0f);
}
void cv::gpu::DisparityBilateralFilter::operator()(const GpuMat& disp, const GpuMat& img, GpuMat& dst, Stream& stream)
{
CV_DbgAssert(0 < ndisp && 0 < radius && 0 < iters);
CV_Assert(disp.rows == img.rows && disp.cols == img.cols && (disp.type() == CV_8U || disp.type() == CV_16S) && (img.type() == CV_8UC1 || img.type() == CV_8UC3));
operators[disp.type()](ndisp, radius, iters, edge_threshold, max_disc_threshold, table_color, table_space, disp, img, dst, stream);
}
#endif /* !defined (HAVE_CUDA) */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int, int, int) { throw_nogpu(); }
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int, int, int, float, float, float) { throw_nogpu(); }
void cv::gpu::DisparityBilateralFilter::operator()(const GpuMat&, const GpuMat&, GpuMat&, Stream&) { throw_nogpu(); }
#else /* !defined (HAVE_CUDA) */
namespace cv { namespace gpu { namespace device
{
namespace disp_bilateral_filter
{
void disp_load_constants(float* table_color, PtrStepSzf table_space, int ndisp, int radius, short edge_disc, short max_disc);
template<typename T>
void disp_bilateral_filter(PtrStepSz<T> disp, PtrStepSzb img, int channels, int iters, cudaStream_t stream);
}
}}}
using namespace ::cv::gpu::device::disp_bilateral_filter;
namespace
{
const float DEFAULT_EDGE_THRESHOLD = 0.1f;
const float DEFAULT_MAX_DISC_THRESHOLD = 0.2f;
const float DEFAULT_SIGMA_RANGE = 10.0f;
inline void calc_color_weighted_table(GpuMat& table_color, float sigma_range, int len)
{
Mat cpu_table_color(1, len, CV_32F);
float* line = cpu_table_color.ptr<float>();
for(int i = 0; i < len; i++)
line[i] = static_cast<float>(std::exp(-double(i * i) / (2 * sigma_range * sigma_range)));
table_color.upload(cpu_table_color);
}
inline void calc_space_weighted_filter(GpuMat& table_space, int win_size, float dist_space)
{
int half = (win_size >> 1);
Mat cpu_table_space(half + 1, half + 1, CV_32F);
for (int y = 0; y <= half; ++y)
{
float* row = cpu_table_space.ptr<float>(y);
for (int x = 0; x <= half; ++x)
row[x] = exp(-sqrt(float(y * y) + float(x * x)) / dist_space);
}
table_space.upload(cpu_table_space);
}
template <typename T>
void disp_bilateral_filter_operator(int ndisp, int radius, int iters, float edge_threshold,float max_disc_threshold,
GpuMat& table_color, GpuMat& table_space,
const GpuMat& disp, const GpuMat& img, GpuMat& dst, Stream& stream)
{
short edge_disc = max<short>(short(1), short(ndisp * edge_threshold + 0.5));
short max_disc = short(ndisp * max_disc_threshold + 0.5);
disp_load_constants(table_color.ptr<float>(), table_space, ndisp, radius, edge_disc, max_disc);
if (&dst != &disp)
{
if (stream)
stream.enqueueCopy(disp, dst);
else
disp.copyTo(dst);
}
disp_bilateral_filter<T>(dst, img, img.channels(), iters, StreamAccessor::getStream(stream));
}
typedef void (*bilateral_filter_operator_t)(int ndisp, int radius, int iters, float edge_threshold, float max_disc_threshold,
GpuMat& table_color, GpuMat& table_space,
const GpuMat& disp, const GpuMat& img, GpuMat& dst, Stream& stream);
const bilateral_filter_operator_t operators[] =
{disp_bilateral_filter_operator<unsigned char>, 0, 0, disp_bilateral_filter_operator<short>, 0, 0, 0, 0};
}
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int ndisp_, int radius_, int iters_)
: ndisp(ndisp_), radius(radius_), iters(iters_), edge_threshold(DEFAULT_EDGE_THRESHOLD), max_disc_threshold(DEFAULT_MAX_DISC_THRESHOLD),
sigma_range(DEFAULT_SIGMA_RANGE)
{
calc_color_weighted_table(table_color, sigma_range, 255);
calc_space_weighted_filter(table_space, radius * 2 + 1, radius + 1.0f);
}
cv::gpu::DisparityBilateralFilter::DisparityBilateralFilter(int ndisp_, int radius_, int iters_, float edge_threshold_,
float max_disc_threshold_, float sigma_range_)
: ndisp(ndisp_), radius(radius_), iters(iters_), edge_threshold(edge_threshold_), max_disc_threshold(max_disc_threshold_),
sigma_range(sigma_range_)
{
calc_color_weighted_table(table_color, sigma_range, 255);
calc_space_weighted_filter(table_space, radius * 2 + 1, radius + 1.0f);
}
void cv::gpu::DisparityBilateralFilter::operator()(const GpuMat& disp, const GpuMat& img, GpuMat& dst, Stream& stream)
{
CV_DbgAssert(0 < ndisp && 0 < radius && 0 < iters);
CV_Assert(disp.rows == img.rows && disp.cols == img.cols && (disp.type() == CV_8U || disp.type() == CV_16S) && (img.type() == CV_8UC1 || img.type() == CV_8UC3));
operators[disp.type()](ndisp, radius, iters, edge_threshold, max_disc_threshold, table_color, table_space, disp, img, dst, stream);
}
#endif /* !defined (HAVE_CUDA) */

200
modules/gpu/src/blend.cpp
View File

@@ -1,100 +1,100 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace std;
using namespace cv;
using namespace cv::gpu;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
void cv::gpu::blendLinear(const GpuMat&, const GpuMat&, const GpuMat&, const GpuMat&, GpuMat&, Stream&) { throw_nogpu(); }
#else
namespace cv { namespace gpu { namespace device
{
namespace blend
{
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream);
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream);
}
}}}
using namespace ::cv::gpu::device::blend;
void cv::gpu::blendLinear(const GpuMat& img1, const GpuMat& img2, const GpuMat& weights1, const GpuMat& weights2,
GpuMat& result, Stream& stream)
{
CV_Assert(img1.size() == img2.size());
CV_Assert(img1.type() == img2.type());
CV_Assert(weights1.size() == img1.size());
CV_Assert(weights2.size() == img2.size());
CV_Assert(weights1.type() == CV_32F);
CV_Assert(weights2.type() == CV_32F);
const Size size = img1.size();
const int depth = img1.depth();
const int cn = img1.channels();
result.create(size, CV_MAKE_TYPE(depth, cn));
switch (depth)
{
case CV_8U:
if (cn != 4)
blendLinearCaller<uchar>(size.height, size.width, cn, img1, img2, weights1, weights2, result, StreamAccessor::getStream(stream));
else
blendLinearCaller8UC4(size.height, size.width, img1, img2, weights1, weights2, result, StreamAccessor::getStream(stream));
break;
case CV_32F:
blendLinearCaller<float>(size.height, size.width, cn, img1, img2, weights1, weights2, result, StreamAccessor::getStream(stream));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "bad image depth in linear blending function");
}
}
#endif
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace std;
using namespace cv;
using namespace cv::gpu;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
void cv::gpu::blendLinear(const GpuMat&, const GpuMat&, const GpuMat&, const GpuMat&, GpuMat&, Stream&) { throw_nogpu(); }
#else
namespace cv { namespace gpu { namespace device
{
namespace blend
{
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream);
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream);
}
}}}
using namespace ::cv::gpu::device::blend;
void cv::gpu::blendLinear(const GpuMat& img1, const GpuMat& img2, const GpuMat& weights1, const GpuMat& weights2,
GpuMat& result, Stream& stream)
{
CV_Assert(img1.size() == img2.size());
CV_Assert(img1.type() == img2.type());
CV_Assert(weights1.size() == img1.size());
CV_Assert(weights2.size() == img2.size());
CV_Assert(weights1.type() == CV_32F);
CV_Assert(weights2.type() == CV_32F);
const Size size = img1.size();
const int depth = img1.depth();
const int cn = img1.channels();
result.create(size, CV_MAKE_TYPE(depth, cn));
switch (depth)
{
case CV_8U:
if (cn != 4)
blendLinearCaller<uchar>(size.height, size.width, cn, img1, img2, weights1, weights2, result, StreamAccessor::getStream(stream));
else
blendLinearCaller8UC4(size.height, size.width, img1, img2, weights1, weights2, result, StreamAccessor::getStream(stream));
break;
case CV_32F:
blendLinearCaller<float>(size.height, size.width, cn, img1, img2, weights1, weights2, result, StreamAccessor::getStream(stream));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "bad image depth in linear blending function");
}
}
#endif

2096
modules/gpu/src/brute_force_matcher.cpp
View File

File diff suppressed because it is too large Load Diff

590
modules/gpu/src/calib3d.cpp
View File

@@ -1,295 +1,295 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#if !defined HAVE_CUDA || defined(CUDA_DISABLER)
void cv::gpu::transformPoints(const GpuMat&, const Mat&, const Mat&, GpuMat&, Stream&) { throw_nogpu(); }
void cv::gpu::projectPoints(const GpuMat&, const Mat&, const Mat&, const Mat&, const Mat&, GpuMat&, Stream&) { throw_nogpu(); }
void cv::gpu::solvePnPRansac(const Mat&, const Mat&, const Mat&, const Mat&, Mat&, Mat&, bool, int, float, int, vector<int>*) { throw_nogpu(); }
#else
namespace cv { namespace gpu { namespace device
{
namespace transform_points
{
void call(const PtrStepSz<float3> src, const float* rot, const float* transl, PtrStepSz<float3> dst, cudaStream_t stream);
}
namespace project_points
{
void call(const PtrStepSz<float3> src, const float* rot, const float* transl, const float* proj, PtrStepSz<float2> dst, cudaStream_t stream);
}
namespace solve_pnp_ransac
{
int maxNumIters();
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores);
}
}}}
using namespace ::cv::gpu::device;
namespace
{
void transformPointsCaller(const GpuMat& src, const Mat& rvec, const Mat& tvec, GpuMat& dst, cudaStream_t stream)
{
CV_Assert(src.rows == 1 && src.cols > 0 && src.type() == CV_32FC3);
CV_Assert(rvec.size() == Size(3, 1) && rvec.type() == CV_32F);
CV_Assert(tvec.size() == Size(3, 1) && tvec.type() == CV_32F);
// Convert rotation vector into matrix
Mat rot;
Rodrigues(rvec, rot);
dst.create(src.size(), src.type());
transform_points::call(src, rot.ptr<float>(), tvec.ptr<float>(), dst, stream);
}
}
void cv::gpu::transformPoints(const GpuMat& src, const Mat& rvec, const Mat& tvec, GpuMat& dst, Stream& stream)
{
transformPointsCaller(src, rvec, tvec, dst, StreamAccessor::getStream(stream));
}
namespace
{
void projectPointsCaller(const GpuMat& src, const Mat& rvec, const Mat& tvec, const Mat& camera_mat, const Mat& dist_coef, GpuMat& dst, cudaStream_t stream)
{
CV_Assert(src.rows == 1 && src.cols > 0 && src.type() == CV_32FC3);
CV_Assert(rvec.size() == Size(3, 1) && rvec.type() == CV_32F);
CV_Assert(tvec.size() == Size(3, 1) && tvec.type() == CV_32F);
CV_Assert(camera_mat.size() == Size(3, 3) && camera_mat.type() == CV_32F);
CV_Assert(dist_coef.empty()); // Undistortion isn't supported
// Convert rotation vector into matrix
Mat rot;
Rodrigues(rvec, rot);
dst.create(src.size(), CV_32FC2);
project_points::call(src, rot.ptr<float>(), tvec.ptr<float>(), camera_mat.ptr<float>(), dst,stream);
}
}
void cv::gpu::projectPoints(const GpuMat& src, const Mat& rvec, const Mat& tvec, const Mat& camera_mat, const Mat& dist_coef, GpuMat& dst, Stream& stream)
{
projectPointsCaller(src, rvec, tvec, camera_mat, dist_coef, dst, StreamAccessor::getStream(stream));
}
namespace
{
// Selects subset_size random different points from [0, num_points - 1] range
void selectRandom(int subset_size, int num_points, vector<int>& subset)
{
subset.resize(subset_size);
for (int i = 0; i < subset_size; ++i)
{
bool was;
do
{
subset[i] = rand() % num_points;
was = false;
for (int j = 0; j < i; ++j)
if (subset[j] == subset[i])
{
was = true;
break;
}
} while (was);
}
}
// Computes rotation, translation pair for small subsets if the input data
class TransformHypothesesGenerator
{
public:
TransformHypothesesGenerator(const Mat& object_, const Mat& image_, const Mat& dist_coef_,
const Mat& camera_mat_, int num_points_, int subset_size_,
Mat rot_matrices_, Mat transl_vectors_)
: object(&object_), image(&image_), dist_coef(&dist_coef_), camera_mat(&camera_mat_),
num_points(num_points_), subset_size(subset_size_), rot_matrices(rot_matrices_),
transl_vectors(transl_vectors_) {}
void operator()(const BlockedRange& range) const
{
// Input data for generation of the current hypothesis
vector<int> subset_indices(subset_size);
Mat_<Point3f> object_subset(1, subset_size);
Mat_<Point2f> image_subset(1, subset_size);
// Current hypothesis data
Mat rot_vec(1, 3, CV_64F);
Mat rot_mat(3, 3, CV_64F);
Mat transl_vec(1, 3, CV_64F);
for (int iter = range.begin(); iter < range.end(); ++iter)
{
selectRandom(subset_size, num_points, subset_indices);
for (int i = 0; i < subset_size; ++i)
{
object_subset(0, i) = object->at<Point3f>(subset_indices[i]);
image_subset(0, i) = image->at<Point2f>(subset_indices[i]);
}
solvePnP(object_subset, image_subset, *camera_mat, *dist_coef, rot_vec, transl_vec);
// Remember translation vector
Mat transl_vec_ = transl_vectors.colRange(iter * 3, (iter + 1) * 3);
transl_vec = transl_vec.reshape(0, 1);
transl_vec.convertTo(transl_vec_, CV_32F);
// Remember rotation matrix
Rodrigues(rot_vec, rot_mat);
Mat rot_mat_ = rot_matrices.colRange(iter * 9, (iter + 1) * 9).reshape(0, 3);
rot_mat.convertTo(rot_mat_, CV_32F);
}
}
const Mat* object;
const Mat* image;
const Mat* dist_coef;
const Mat* camera_mat;
int num_points;
int subset_size;
// Hypotheses storage (global)
Mat rot_matrices;
Mat transl_vectors;
};
}
void cv::gpu::solvePnPRansac(const Mat& object, const Mat& image, const Mat& camera_mat,
const Mat& dist_coef, Mat& rvec, Mat& tvec, bool use_extrinsic_guess,
int num_iters, float max_dist, int min_inlier_count,
vector<int>* inliers)
{
(void)min_inlier_count;
CV_Assert(object.rows == 1 && object.cols > 0 && object.type() == CV_32FC3);
CV_Assert(image.rows == 1 && image.cols > 0 && image.type() == CV_32FC2);
CV_Assert(object.cols == image.cols);
CV_Assert(camera_mat.size() == Size(3, 3) && camera_mat.type() == CV_32F);
CV_Assert(!use_extrinsic_guess); // We don't support initial guess for now
CV_Assert(num_iters <= solve_pnp_ransac::maxNumIters());
const int subset_size = 4;
const int num_points = object.cols;
CV_Assert(num_points >= subset_size);
// Unapply distortion and intrinsic camera transformations
Mat eye_camera_mat = Mat::eye(3, 3, CV_32F);
Mat empty_dist_coef;
Mat image_normalized;
undistortPoints(image, image_normalized, camera_mat, dist_coef, Mat(), eye_camera_mat);
// Hypotheses storage (global)
Mat rot_matrices(1, num_iters * 9, CV_32F);
Mat transl_vectors(1, num_iters * 3, CV_32F);
// Generate set of hypotheses using small subsets of the input data
TransformHypothesesGenerator body(object, image_normalized, empty_dist_coef, eye_camera_mat,
num_points, subset_size, rot_matrices, transl_vectors);
parallel_for(BlockedRange(0, num_iters), body);
// Compute scores (i.e. number of inliers) for each hypothesis
GpuMat d_object(object);
GpuMat d_image_normalized(image_normalized);
GpuMat d_hypothesis_scores(1, num_iters, CV_32S);
solve_pnp_ransac::computeHypothesisScores(
num_iters, num_points, rot_matrices.ptr<float>(), transl_vectors.ptr<float3>(),
d_object.ptr<float3>(), d_image_normalized.ptr<float2>(), max_dist * max_dist,
d_hypothesis_scores.ptr<int>());
// Find the best hypothesis index
Point best_idx;
double best_score;
minMaxLoc(d_hypothesis_scores, NULL, &best_score, NULL, &best_idx);
int num_inliers = static_cast<int>(best_score);
// Extract the best hypothesis data
Mat rot_mat = rot_matrices.colRange(best_idx.x * 9, (best_idx.x + 1) * 9).reshape(0, 3);
Rodrigues(rot_mat, rvec);
rvec = rvec.reshape(0, 1);
tvec = transl_vectors.colRange(best_idx.x * 3, (best_idx.x + 1) * 3).clone();
tvec = tvec.reshape(0, 1);
// Build vector of inlier indices
if (inliers != NULL)
{
inliers->clear();
inliers->reserve(num_inliers);
Point3f p, p_transf;
Point2f p_proj;
const float* rot = rot_mat.ptr<float>();
const float* transl = tvec.ptr<float>();
for (int i = 0; i < num_points; ++i)
{
p = object.at<Point3f>(0, i);
p_transf.x = rot[0] * p.x + rot[1] * p.y + rot[2] * p.z + transl[0];
p_transf.y = rot[3] * p.x + rot[4] * p.y + rot[5] * p.z + transl[1];
p_transf.z = rot[6] * p.x + rot[7] * p.y + rot[8] * p.z + transl[2];
p_proj.x = p_transf.x / p_transf.z;
p_proj.y = p_transf.y / p_transf.z;
if (norm(p_proj - image_normalized.at<Point2f>(0, i)) < max_dist)
inliers->push_back(i);
}
}
}
#endif
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#if !defined HAVE_CUDA || defined(CUDA_DISABLER)
void cv::gpu::transformPoints(const GpuMat&, const Mat&, const Mat&, GpuMat&, Stream&) { throw_nogpu(); }
void cv::gpu::projectPoints(const GpuMat&, const Mat&, const Mat&, const Mat&, const Mat&, GpuMat&, Stream&) { throw_nogpu(); }
void cv::gpu::solvePnPRansac(const Mat&, const Mat&, const Mat&, const Mat&, Mat&, Mat&, bool, int, float, int, vector<int>*) { throw_nogpu(); }
#else
namespace cv { namespace gpu { namespace device
{
namespace transform_points
{
void call(const PtrStepSz<float3> src, const float* rot, const float* transl, PtrStepSz<float3> dst, cudaStream_t stream);
}
namespace project_points
{
void call(const PtrStepSz<float3> src, const float* rot, const float* transl, const float* proj, PtrStepSz<float2> dst, cudaStream_t stream);
}
namespace solve_pnp_ransac
{
int maxNumIters();
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores);
}
}}}
using namespace ::cv::gpu::device;
namespace
{
void transformPointsCaller(const GpuMat& src, const Mat& rvec, const Mat& tvec, GpuMat& dst, cudaStream_t stream)
{
CV_Assert(src.rows == 1 && src.cols > 0 && src.type() == CV_32FC3);
CV_Assert(rvec.size() == Size(3, 1) && rvec.type() == CV_32F);
CV_Assert(tvec.size() == Size(3, 1) && tvec.type() == CV_32F);
// Convert rotation vector into matrix
Mat rot;
Rodrigues(rvec, rot);
dst.create(src.size(), src.type());
transform_points::call(src, rot.ptr<float>(), tvec.ptr<float>(), dst, stream);
}
}
void cv::gpu::transformPoints(const GpuMat& src, const Mat& rvec, const Mat& tvec, GpuMat& dst, Stream& stream)
{
transformPointsCaller(src, rvec, tvec, dst, StreamAccessor::getStream(stream));
}
namespace
{
void projectPointsCaller(const GpuMat& src, const Mat& rvec, const Mat& tvec, const Mat& camera_mat, const Mat& dist_coef, GpuMat& dst, cudaStream_t stream)
{
CV_Assert(src.rows == 1 && src.cols > 0 && src.type() == CV_32FC3);
CV_Assert(rvec.size() == Size(3, 1) && rvec.type() == CV_32F);
CV_Assert(tvec.size() == Size(3, 1) && tvec.type() == CV_32F);
CV_Assert(camera_mat.size() == Size(3, 3) && camera_mat.type() == CV_32F);
CV_Assert(dist_coef.empty()); // Undistortion isn't supported
// Convert rotation vector into matrix
Mat rot;
Rodrigues(rvec, rot);
dst.create(src.size(), CV_32FC2);
project_points::call(src, rot.ptr<float>(), tvec.ptr<float>(), camera_mat.ptr<float>(), dst,stream);
}
}
void cv::gpu::projectPoints(const GpuMat& src, const Mat& rvec, const Mat& tvec, const Mat& camera_mat, const Mat& dist_coef, GpuMat& dst, Stream& stream)
{
projectPointsCaller(src, rvec, tvec, camera_mat, dist_coef, dst, StreamAccessor::getStream(stream));
}
namespace
{
// Selects subset_size random different points from [0, num_points - 1] range
void selectRandom(int subset_size, int num_points, vector<int>& subset)
{
subset.resize(subset_size);
for (int i = 0; i < subset_size; ++i)
{
bool was;
do
{
subset[i] = rand() % num_points;
was = false;
for (int j = 0; j < i; ++j)
if (subset[j] == subset[i])
{
was = true;
break;
}
} while (was);
}
}
// Computes rotation, translation pair for small subsets if the input data
class TransformHypothesesGenerator
{
public:
TransformHypothesesGenerator(const Mat& object_, const Mat& image_, const Mat& dist_coef_,
const Mat& camera_mat_, int num_points_, int subset_size_,
Mat rot_matrices_, Mat transl_vectors_)
: object(&object_), image(&image_), dist_coef(&dist_coef_), camera_mat(&camera_mat_),
num_points(num_points_), subset_size(subset_size_), rot_matrices(rot_matrices_),
transl_vectors(transl_vectors_) {}
void operator()(const BlockedRange& range) const
{
// Input data for generation of the current hypothesis
vector<int> subset_indices(subset_size);
Mat_<Point3f> object_subset(1, subset_size);
Mat_<Point2f> image_subset(1, subset_size);
// Current hypothesis data
Mat rot_vec(1, 3, CV_64F);
Mat rot_mat(3, 3, CV_64F);
Mat transl_vec(1, 3, CV_64F);
for (int iter = range.begin(); iter < range.end(); ++iter)
{
selectRandom(subset_size, num_points, subset_indices);
for (int i = 0; i < subset_size; ++i)
{
object_subset(0, i) = object->at<Point3f>(subset_indices[i]);
image_subset(0, i) = image->at<Point2f>(subset_indices[i]);
}
solvePnP(object_subset, image_subset, *camera_mat, *dist_coef, rot_vec, transl_vec);
// Remember translation vector
Mat transl_vec_ = transl_vectors.colRange(iter * 3, (iter + 1) * 3);
transl_vec = transl_vec.reshape(0, 1);
transl_vec.convertTo(transl_vec_, CV_32F);
// Remember rotation matrix
Rodrigues(rot_vec, rot_mat);
Mat rot_mat_ = rot_matrices.colRange(iter * 9, (iter + 1) * 9).reshape(0, 3);
rot_mat.convertTo(rot_mat_, CV_32F);
}
}
const Mat* object;
const Mat* image;
const Mat* dist_coef;
const Mat* camera_mat;
int num_points;
int subset_size;
// Hypotheses storage (global)
Mat rot_matrices;
Mat transl_vectors;
};
}
void cv::gpu::solvePnPRansac(const Mat& object, const Mat& image, const Mat& camera_mat,
const Mat& dist_coef, Mat& rvec, Mat& tvec, bool use_extrinsic_guess,
int num_iters, float max_dist, int min_inlier_count,
vector<int>* inliers)
{
(void)min_inlier_count;
CV_Assert(object.rows == 1 && object.cols > 0 && object.type() == CV_32FC3);
CV_Assert(image.rows == 1 && image.cols > 0 && image.type() == CV_32FC2);
CV_Assert(object.cols == image.cols);
CV_Assert(camera_mat.size() == Size(3, 3) && camera_mat.type() == CV_32F);
CV_Assert(!use_extrinsic_guess); // We don't support initial guess for now
CV_Assert(num_iters <= solve_pnp_ransac::maxNumIters());
const int subset_size = 4;
const int num_points = object.cols;
CV_Assert(num_points >= subset_size);
// Unapply distortion and intrinsic camera transformations
Mat eye_camera_mat = Mat::eye(3, 3, CV_32F);
Mat empty_dist_coef;
Mat image_normalized;
undistortPoints(image, image_normalized, camera_mat, dist_coef, Mat(), eye_camera_mat);
// Hypotheses storage (global)
Mat rot_matrices(1, num_iters * 9, CV_32F);
Mat transl_vectors(1, num_iters * 3, CV_32F);
// Generate set of hypotheses using small subsets of the input data
TransformHypothesesGenerator body(object, image_normalized, empty_dist_coef, eye_camera_mat,
num_points, subset_size, rot_matrices, transl_vectors);
parallel_for(BlockedRange(0, num_iters), body);
// Compute scores (i.e. number of inliers) for each hypothesis
GpuMat d_object(object);
GpuMat d_image_normalized(image_normalized);
GpuMat d_hypothesis_scores(1, num_iters, CV_32S);
solve_pnp_ransac::computeHypothesisScores(
num_iters, num_points, rot_matrices.ptr<float>(), transl_vectors.ptr<float3>(),
d_object.ptr<float3>(), d_image_normalized.ptr<float2>(), max_dist * max_dist,
d_hypothesis_scores.ptr<int>());
// Find the best hypothesis index
Point best_idx;
double best_score;
minMaxLoc(d_hypothesis_scores, NULL, &best_score, NULL, &best_idx);
int num_inliers = static_cast<int>(best_score);
// Extract the best hypothesis data
Mat rot_mat = rot_matrices.colRange(best_idx.x * 9, (best_idx.x + 1) * 9).reshape(0, 3);
Rodrigues(rot_mat, rvec);
rvec = rvec.reshape(0, 1);
tvec = transl_vectors.colRange(best_idx.x * 3, (best_idx.x + 1) * 3).clone();
tvec = tvec.reshape(0, 1);
// Build vector of inlier indices
if (inliers != NULL)
{
inliers->clear();
inliers->reserve(num_inliers);
Point3f p, p_transf;
Point2f p_proj;
const float* rot = rot_mat.ptr<float>();
const float* transl = tvec.ptr<float>();
for (int i = 0; i < num_points; ++i)
{
p = object.at<Point3f>(0, i);
p_transf.x = rot[0] * p.x + rot[1] * p.y + rot[2] * p.z + transl[0];
p_transf.y = rot[3] * p.x + rot[4] * p.y + rot[5] * p.z + transl[1];
p_transf.z = rot[6] * p.x + rot[7] * p.y + rot[8] * p.z + transl[2];
p_proj.x = p_transf.x / p_transf.z;
p_proj.y = p_transf.y / p_transf.z;
if (norm(p_proj - image_normalized.at<Point2f>(0, i)) < max_dist)
inliers->push_back(i);
}
}
}
#endif

3242
modules/gpu/src/color.cpp
View File

File diff suppressed because it is too large Load Diff

278
modules/gpu/src/cu_safe_call.cpp
View File

@@ -1,139 +1,139 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "cu_safe_call.h"
#ifdef HAVE_CUDA
namespace
{
#define error_entry(entry) { entry, #entry }
struct ErrorEntry
{
int code;
std::string str;
};
class ErrorEntryComparer
{
public:
inline ErrorEntryComparer(int code) : code_(code) {}
inline bool operator()(const ErrorEntry& e) const { return e.code == code_; }
private:
int code_;
};
std::string getErrorString(int code, const ErrorEntry* errors, size_t n)
{
size_t idx = std::find_if(errors, errors + n, ErrorEntryComparer(code)) - errors;
const std::string& msg = (idx != n) ? errors[idx].str : std::string("Unknown error code");
std::ostringstream ostr;
ostr << msg << " [Code = " << code << "]";
return ostr.str();
}
const ErrorEntry cu_errors [] =
{
error_entry( CUDA_SUCCESS ),
error_entry( CUDA_ERROR_INVALID_VALUE ),
error_entry( CUDA_ERROR_OUT_OF_MEMORY ),
error_entry( CUDA_ERROR_NOT_INITIALIZED ),
error_entry( CUDA_ERROR_DEINITIALIZED ),
error_entry( CUDA_ERROR_PROFILER_DISABLED ),
error_entry( CUDA_ERROR_PROFILER_NOT_INITIALIZED ),
error_entry( CUDA_ERROR_PROFILER_ALREADY_STARTED ),
error_entry( CUDA_ERROR_PROFILER_ALREADY_STOPPED ),
error_entry( CUDA_ERROR_NO_DEVICE ),
error_entry( CUDA_ERROR_INVALID_DEVICE ),
error_entry( CUDA_ERROR_INVALID_IMAGE ),
error_entry( CUDA_ERROR_INVALID_CONTEXT ),
error_entry( CUDA_ERROR_CONTEXT_ALREADY_CURRENT ),
error_entry( CUDA_ERROR_MAP_FAILED ),
error_entry( CUDA_ERROR_UNMAP_FAILED ),
error_entry( CUDA_ERROR_ARRAY_IS_MAPPED ),
error_entry( CUDA_ERROR_ALREADY_MAPPED ),
error_entry( CUDA_ERROR_NO_BINARY_FOR_GPU ),
error_entry( CUDA_ERROR_ALREADY_ACQUIRED ),
error_entry( CUDA_ERROR_NOT_MAPPED ),
error_entry( CUDA_ERROR_NOT_MAPPED_AS_ARRAY ),
error_entry( CUDA_ERROR_NOT_MAPPED_AS_POINTER ),
error_entry( CUDA_ERROR_ECC_UNCORRECTABLE ),
error_entry( CUDA_ERROR_UNSUPPORTED_LIMIT ),
error_entry( CUDA_ERROR_CONTEXT_ALREADY_IN_USE ),
error_entry( CUDA_ERROR_INVALID_SOURCE ),
error_entry( CUDA_ERROR_FILE_NOT_FOUND ),
error_entry( CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND ),
error_entry( CUDA_ERROR_SHARED_OBJECT_INIT_FAILED ),
error_entry( CUDA_ERROR_OPERATING_SYSTEM ),
error_entry( CUDA_ERROR_INVALID_HANDLE ),
error_entry( CUDA_ERROR_NOT_FOUND ),
error_entry( CUDA_ERROR_NOT_READY ),
error_entry( CUDA_ERROR_LAUNCH_FAILED ),
error_entry( CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES ),
error_entry( CUDA_ERROR_LAUNCH_TIMEOUT ),
error_entry( CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING ),
error_entry( CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED ),
error_entry( CUDA_ERROR_PEER_ACCESS_NOT_ENABLED ),
error_entry( CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE ),
error_entry( CUDA_ERROR_CONTEXT_IS_DESTROYED ),
error_entry( CUDA_ERROR_ASSERT ),
error_entry( CUDA_ERROR_TOO_MANY_PEERS ),
error_entry( CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED ),
error_entry( CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED ),
error_entry( CUDA_ERROR_UNKNOWN )
};
const size_t cu_errors_num = sizeof(cu_errors) / sizeof(cu_errors[0]);
}
std::string cv::gpu::detail::cuGetErrString(CUresult res)
{
return getErrorString(res, cu_errors, cu_errors_num);
}
#endif // HAVE_CUDA
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "cu_safe_call.h"
#ifdef HAVE_CUDA
namespace
{
#define error_entry(entry) { entry, #entry }
struct ErrorEntry
{
int code;
std::string str;
};
class ErrorEntryComparer
{
public:
inline ErrorEntryComparer(int code) : code_(code) {}
inline bool operator()(const ErrorEntry& e) const { return e.code == code_; }
private:
int code_;
};
std::string getErrorString(int code, const ErrorEntry* errors, size_t n)
{
size_t idx = std::find_if(errors, errors + n, ErrorEntryComparer(code)) - errors;
const std::string& msg = (idx != n) ? errors[idx].str : std::string("Unknown error code");
std::ostringstream ostr;
ostr << msg << " [Code = " << code << "]";
return ostr.str();
}
const ErrorEntry cu_errors [] =
{
error_entry( CUDA_SUCCESS ),
error_entry( CUDA_ERROR_INVALID_VALUE ),
error_entry( CUDA_ERROR_OUT_OF_MEMORY ),
error_entry( CUDA_ERROR_NOT_INITIALIZED ),
error_entry( CUDA_ERROR_DEINITIALIZED ),
error_entry( CUDA_ERROR_PROFILER_DISABLED ),
error_entry( CUDA_ERROR_PROFILER_NOT_INITIALIZED ),
error_entry( CUDA_ERROR_PROFILER_ALREADY_STARTED ),
error_entry( CUDA_ERROR_PROFILER_ALREADY_STOPPED ),
error_entry( CUDA_ERROR_NO_DEVICE ),
error_entry( CUDA_ERROR_INVALID_DEVICE ),
error_entry( CUDA_ERROR_INVALID_IMAGE ),
error_entry( CUDA_ERROR_INVALID_CONTEXT ),
error_entry( CUDA_ERROR_CONTEXT_ALREADY_CURRENT ),
error_entry( CUDA_ERROR_MAP_FAILED ),
error_entry( CUDA_ERROR_UNMAP_FAILED ),
error_entry( CUDA_ERROR_ARRAY_IS_MAPPED ),
error_entry( CUDA_ERROR_ALREADY_MAPPED ),
error_entry( CUDA_ERROR_NO_BINARY_FOR_GPU ),
error_entry( CUDA_ERROR_ALREADY_ACQUIRED ),
error_entry( CUDA_ERROR_NOT_MAPPED ),
error_entry( CUDA_ERROR_NOT_MAPPED_AS_ARRAY ),
error_entry( CUDA_ERROR_NOT_MAPPED_AS_POINTER ),
error_entry( CUDA_ERROR_ECC_UNCORRECTABLE ),
error_entry( CUDA_ERROR_UNSUPPORTED_LIMIT ),
error_entry( CUDA_ERROR_CONTEXT_ALREADY_IN_USE ),
error_entry( CUDA_ERROR_INVALID_SOURCE ),
error_entry( CUDA_ERROR_FILE_NOT_FOUND ),
error_entry( CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND ),
error_entry( CUDA_ERROR_SHARED_OBJECT_INIT_FAILED ),
error_entry( CUDA_ERROR_OPERATING_SYSTEM ),
error_entry( CUDA_ERROR_INVALID_HANDLE ),
error_entry( CUDA_ERROR_NOT_FOUND ),
error_entry( CUDA_ERROR_NOT_READY ),
error_entry( CUDA_ERROR_LAUNCH_FAILED ),
error_entry( CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES ),
error_entry( CUDA_ERROR_LAUNCH_TIMEOUT ),
error_entry( CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING ),
error_entry( CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED ),
error_entry( CUDA_ERROR_PEER_ACCESS_NOT_ENABLED ),
error_entry( CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE ),
error_entry( CUDA_ERROR_CONTEXT_IS_DESTROYED ),
error_entry( CUDA_ERROR_ASSERT ),
error_entry( CUDA_ERROR_TOO_MANY_PEERS ),
error_entry( CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED ),
error_entry( CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED ),
error_entry( CUDA_ERROR_UNKNOWN )
};
const size_t cu_errors_num = sizeof(cu_errors) / sizeof(cu_errors[0]);
}
std::string cv::gpu::detail::cuGetErrString(CUresult res)
{
return getErrorString(res, cu_errors, cu_errors_num);
}
#endif // HAVE_CUDA

134
modules/gpu/src/cu_safe_call.h
View File

@@ -1,67 +1,67 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __CU_SAFE_CALL_H__
#define __CU_SAFE_CALL_H__
#include "precomp.hpp"
#ifdef HAVE_CUDA
namespace cv { namespace gpu {
namespace detail
{
std::string cuGetErrString(CUresult res);
inline void cuSafeCall_impl(CUresult res, const char* file, int line)
{
if (res != CUDA_SUCCESS)
cv::error( cv::Exception(CV_GpuApiCallError, cuGetErrString(res), "unknown function", file, line) );
}
}
}}
#define cuSafeCall( op ) cv::gpu::detail::cuSafeCall_impl( (op), __FILE__, __LINE__ )
#endif // HAVE_CUDA
#endif // __CU_SAFE_CALL_H__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __CU_SAFE_CALL_H__
#define __CU_SAFE_CALL_H__
#include "precomp.hpp"
#ifdef HAVE_CUDA
namespace cv { namespace gpu {
namespace detail
{
std::string cuGetErrString(CUresult res);
inline void cuSafeCall_impl(CUresult res, const char* file, int line)
{
if (res != CUDA_SUCCESS)
cv::error( cv::Exception(CV_GpuApiCallError, cuGetErrString(res), "unknown function", file, line) );
}
}
}}
#define cuSafeCall( op ) cv::gpu::detail::cuSafeCall_impl( (op), __FILE__, __LINE__ )
#endif // HAVE_CUDA
#endif // __CU_SAFE_CALL_H__

422
modules/gpu/src/cuda/NV12ToARGB.cu
View File

@@ -1,212 +1,212 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
/*
* Copyright 1993-2010 NVIDIA Corporation. All rights reserved.
*
* Please refer to the NVIDIA end user license agreement (EULA) associated
* with this source code for terms and conditions that govern your use of
* this software. Any use, reproduction, disclosure, or distribution of
* this software and related documentation outside the terms of the EULA
* is strictly prohibited.
*
*/
/*
NV12ToARGB color space conversion CUDA kernel
This sample uses CUDA to perform a simple NV12 (YUV 4:2:0 planar)
source and converts to output in ARGB format
*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
namespace cv { namespace gpu { namespace device {
namespace video_decoding
{
__constant__ uint constAlpha = ((uint)0xff << 24);
__constant__ float constHueColorSpaceMat[9];
void loadHueCSC(float hueCSC[9])
{
cudaSafeCall( cudaMemcpyToSymbol(constHueColorSpaceMat, hueCSC, 9 * sizeof(float)) );
}
__device__ void YUV2RGB(const uint* yuvi, float* red, float* green, float* blue)
{
float luma, chromaCb, chromaCr;
// Prepare for hue adjustment
luma = (float)yuvi[0];
chromaCb = (float)((int)yuvi[1] - 512.0f);
chromaCr = (float)((int)yuvi[2] - 512.0f);
// Convert YUV To RGB with hue adjustment
*red = (luma * constHueColorSpaceMat[0]) +
(chromaCb * constHueColorSpaceMat[1]) +
(chromaCr * constHueColorSpaceMat[2]);
*green = (luma * constHueColorSpaceMat[3]) +
(chromaCb * constHueColorSpaceMat[4]) +
(chromaCr * constHueColorSpaceMat[5]);
*blue = (luma * constHueColorSpaceMat[6]) +
(chromaCb * constHueColorSpaceMat[7]) +
(chromaCr * constHueColorSpaceMat[8]);
}
__device__ uint RGBAPACK_10bit(float red, float green, float blue, uint alpha)
{
uint ARGBpixel = 0;
// Clamp final 10 bit results
red = ::fmin(::fmax(red, 0.0f), 1023.f);
green = ::fmin(::fmax(green, 0.0f), 1023.f);
blue = ::fmin(::fmax(blue, 0.0f), 1023.f);
// Convert to 8 bit unsigned integers per color component
ARGBpixel = (((uint)blue >> 2) |
(((uint)green >> 2) << 8) |
(((uint)red >> 2) << 16) |
(uint)alpha);
return ARGBpixel;
}
// CUDA kernel for outputing the final ARGB output from NV12
#define COLOR_COMPONENT_BIT_SIZE 10
#define COLOR_COMPONENT_MASK 0x3FF
__global__ void NV12ToARGB(uchar* srcImage, size_t nSourcePitch,
uint* dstImage, size_t nDestPitch,
uint width, uint height)
{
// Pad borders with duplicate pixels, and we multiply by 2 because we process 2 pixels per thread
const int x = blockIdx.x * (blockDim.x << 1) + (threadIdx.x << 1);
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x >= width || y >= height)
return;
// Read 2 Luma components at a time, so we don't waste processing since CbCr are decimated this way.
// if we move to texture we could read 4 luminance values
uint yuv101010Pel[2];
yuv101010Pel[0] = (srcImage[y * nSourcePitch + x ]) << 2;
yuv101010Pel[1] = (srcImage[y * nSourcePitch + x + 1]) << 2;
const size_t chromaOffset = nSourcePitch * height;
const int y_chroma = y >> 1;
if (y & 1) // odd scanline ?
{
uint chromaCb = srcImage[chromaOffset + y_chroma * nSourcePitch + x ];
uint chromaCr = srcImage[chromaOffset + y_chroma * nSourcePitch + x + 1];
if (y_chroma < ((height >> 1) - 1)) // interpolate chroma vertically
{
chromaCb = (chromaCb + srcImage[chromaOffset + (y_chroma + 1) * nSourcePitch + x ] + 1) >> 1;
chromaCr = (chromaCr + srcImage[chromaOffset + (y_chroma + 1) * nSourcePitch + x + 1] + 1) >> 1;
}
yuv101010Pel[0] |= (chromaCb << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[0] |= (chromaCr << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
yuv101010Pel[1] |= (chromaCb << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[1] |= (chromaCr << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
}
else
{
yuv101010Pel[0] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x ] << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[0] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x + 1] << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
yuv101010Pel[1] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x ] << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[1] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x + 1] << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
}
// this steps performs the color conversion
uint yuvi[6];
float red[2], green[2], blue[2];
yuvi[0] = (yuv101010Pel[0] & COLOR_COMPONENT_MASK );
yuvi[1] = ((yuv101010Pel[0] >> COLOR_COMPONENT_BIT_SIZE) & COLOR_COMPONENT_MASK);
yuvi[2] = ((yuv101010Pel[0] >> (COLOR_COMPONENT_BIT_SIZE << 1)) & COLOR_COMPONENT_MASK);
yuvi[3] = (yuv101010Pel[1] & COLOR_COMPONENT_MASK );
yuvi[4] = ((yuv101010Pel[1] >> COLOR_COMPONENT_BIT_SIZE) & COLOR_COMPONENT_MASK);
yuvi[5] = ((yuv101010Pel[1] >> (COLOR_COMPONENT_BIT_SIZE << 1)) & COLOR_COMPONENT_MASK);
// YUV to RGB Transformation conversion
YUV2RGB(&yuvi[0], &red[0], &green[0], &blue[0]);
YUV2RGB(&yuvi[3], &red[1], &green[1], &blue[1]);
// Clamp the results to RGBA
const size_t dstImagePitch = nDestPitch >> 2;
dstImage[y * dstImagePitch + x ] = RGBAPACK_10bit(red[0], green[0], blue[0], constAlpha);
dstImage[y * dstImagePitch + x + 1 ] = RGBAPACK_10bit(red[1], green[1], blue[1], constAlpha);
}
void NV12ToARGB_gpu(const PtrStepb decodedFrame, PtrStepSz<uint> interopFrame, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(interopFrame.cols, 2 * block.x), divUp(interopFrame.rows, block.y));
NV12ToARGB<<<grid, block, 0, stream>>>(decodedFrame.data, decodedFrame.step, interopFrame.data, interopFrame.step,
interopFrame.cols, interopFrame.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
}}}
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
/*
* Copyright 1993-2010 NVIDIA Corporation. All rights reserved.
*
* Please refer to the NVIDIA end user license agreement (EULA) associated
* with this source code for terms and conditions that govern your use of
* this software. Any use, reproduction, disclosure, or distribution of
* this software and related documentation outside the terms of the EULA
* is strictly prohibited.
*
*/
/*
NV12ToARGB color space conversion CUDA kernel
This sample uses CUDA to perform a simple NV12 (YUV 4:2:0 planar)
source and converts to output in ARGB format
*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
namespace cv { namespace gpu { namespace device {
namespace video_decoding
{
__constant__ uint constAlpha = ((uint)0xff << 24);
__constant__ float constHueColorSpaceMat[9];
void loadHueCSC(float hueCSC[9])
{
cudaSafeCall( cudaMemcpyToSymbol(constHueColorSpaceMat, hueCSC, 9 * sizeof(float)) );
}
__device__ void YUV2RGB(const uint* yuvi, float* red, float* green, float* blue)
{
float luma, chromaCb, chromaCr;
// Prepare for hue adjustment
luma = (float)yuvi[0];
chromaCb = (float)((int)yuvi[1] - 512.0f);
chromaCr = (float)((int)yuvi[2] - 512.0f);
// Convert YUV To RGB with hue adjustment
*red = (luma * constHueColorSpaceMat[0]) +
(chromaCb * constHueColorSpaceMat[1]) +
(chromaCr * constHueColorSpaceMat[2]);
*green = (luma * constHueColorSpaceMat[3]) +
(chromaCb * constHueColorSpaceMat[4]) +
(chromaCr * constHueColorSpaceMat[5]);
*blue = (luma * constHueColorSpaceMat[6]) +
(chromaCb * constHueColorSpaceMat[7]) +
(chromaCr * constHueColorSpaceMat[8]);
}
__device__ uint RGBAPACK_10bit(float red, float green, float blue, uint alpha)
{
uint ARGBpixel = 0;
// Clamp final 10 bit results
red = ::fmin(::fmax(red, 0.0f), 1023.f);
green = ::fmin(::fmax(green, 0.0f), 1023.f);
blue = ::fmin(::fmax(blue, 0.0f), 1023.f);
// Convert to 8 bit unsigned integers per color component
ARGBpixel = (((uint)blue >> 2) |
(((uint)green >> 2) << 8) |
(((uint)red >> 2) << 16) |
(uint)alpha);
return ARGBpixel;
}
// CUDA kernel for outputing the final ARGB output from NV12
#define COLOR_COMPONENT_BIT_SIZE 10
#define COLOR_COMPONENT_MASK 0x3FF
__global__ void NV12ToARGB(uchar* srcImage, size_t nSourcePitch,
uint* dstImage, size_t nDestPitch,
uint width, uint height)
{
// Pad borders with duplicate pixels, and we multiply by 2 because we process 2 pixels per thread
const int x = blockIdx.x * (blockDim.x << 1) + (threadIdx.x << 1);
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x >= width || y >= height)
return;
// Read 2 Luma components at a time, so we don't waste processing since CbCr are decimated this way.
// if we move to texture we could read 4 luminance values
uint yuv101010Pel[2];
yuv101010Pel[0] = (srcImage[y * nSourcePitch + x ]) << 2;
yuv101010Pel[1] = (srcImage[y * nSourcePitch + x + 1]) << 2;
const size_t chromaOffset = nSourcePitch * height;
const int y_chroma = y >> 1;
if (y & 1) // odd scanline ?
{
uint chromaCb = srcImage[chromaOffset + y_chroma * nSourcePitch + x ];
uint chromaCr = srcImage[chromaOffset + y_chroma * nSourcePitch + x + 1];
if (y_chroma < ((height >> 1) - 1)) // interpolate chroma vertically
{
chromaCb = (chromaCb + srcImage[chromaOffset + (y_chroma + 1) * nSourcePitch + x ] + 1) >> 1;
chromaCr = (chromaCr + srcImage[chromaOffset + (y_chroma + 1) * nSourcePitch + x + 1] + 1) >> 1;
}
yuv101010Pel[0] |= (chromaCb << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[0] |= (chromaCr << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
yuv101010Pel[1] |= (chromaCb << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[1] |= (chromaCr << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
}
else
{
yuv101010Pel[0] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x ] << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[0] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x + 1] << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
yuv101010Pel[1] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x ] << ( COLOR_COMPONENT_BIT_SIZE + 2));
yuv101010Pel[1] |= ((uint)srcImage[chromaOffset + y_chroma * nSourcePitch + x + 1] << ((COLOR_COMPONENT_BIT_SIZE << 1) + 2));
}
// this steps performs the color conversion
uint yuvi[6];
float red[2], green[2], blue[2];
yuvi[0] = (yuv101010Pel[0] & COLOR_COMPONENT_MASK );
yuvi[1] = ((yuv101010Pel[0] >> COLOR_COMPONENT_BIT_SIZE) & COLOR_COMPONENT_MASK);
yuvi[2] = ((yuv101010Pel[0] >> (COLOR_COMPONENT_BIT_SIZE << 1)) & COLOR_COMPONENT_MASK);
yuvi[3] = (yuv101010Pel[1] & COLOR_COMPONENT_MASK );
yuvi[4] = ((yuv101010Pel[1] >> COLOR_COMPONENT_BIT_SIZE) & COLOR_COMPONENT_MASK);
yuvi[5] = ((yuv101010Pel[1] >> (COLOR_COMPONENT_BIT_SIZE << 1)) & COLOR_COMPONENT_MASK);
// YUV to RGB Transformation conversion
YUV2RGB(&yuvi[0], &red[0], &green[0], &blue[0]);
YUV2RGB(&yuvi[3], &red[1], &green[1], &blue[1]);
// Clamp the results to RGBA
const size_t dstImagePitch = nDestPitch >> 2;
dstImage[y * dstImagePitch + x ] = RGBAPACK_10bit(red[0], green[0], blue[0], constAlpha);
dstImage[y * dstImagePitch + x + 1 ] = RGBAPACK_10bit(red[1], green[1], blue[1], constAlpha);
}
void NV12ToARGB_gpu(const PtrStepb decodedFrame, PtrStepSz<uint> interopFrame, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(interopFrame.cols, 2 * block.x), divUp(interopFrame.rows, block.y));
NV12ToARGB<<<grid, block, 0, stream>>>(decodedFrame.data, decodedFrame.step, interopFrame.data, interopFrame.step,
interopFrame.cols, interopFrame.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
}}}
#endif /* CUDA_DISABLER */

2332
modules/gpu/src/cuda/bf_knnmatch.cu
View File

File diff suppressed because it is too large Load Diff

1568
modules/gpu/src/cuda/bf_match.cu
View File

File diff suppressed because it is too large Load Diff

942
modules/gpu/src/cuda/bf_radius_match.cu
View File

@@ -1,472 +1,472 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/vec_distance.hpp"
#include "opencv2/gpu/device/datamov_utils.hpp"
namespace cv { namespace gpu { namespace device
{
namespace bf_radius_match
{
///////////////////////////////////////////////////////////////////////////////
// Match Unrolled
template <int BLOCK_SIZE, int MAX_DESC_LEN, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void matchUnrolled(const PtrStepSz<T> query, int imgIdx, const PtrStepSz<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 110)
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
#pragma unroll
for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T, typename Mask>
void matchUnrolled(const PtrStepSz<T>& query, const PtrStepSz<T>& train, float maxDistance, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T>
void matchUnrolled(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const PtrStepSz<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////////
// Match
template <int BLOCK_SIZE, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void match(const PtrStepSz<T> query, int imgIdx, const PtrStepSz<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 110)
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
for (int i = 0, endi = (query.cols + BLOCK_SIZE - 1) / BLOCK_SIZE; i < endi; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
template <int BLOCK_SIZE, typename Dist, typename T, typename Mask>
void match(const PtrStepSz<T>& query, const PtrStepSz<T>& train, float maxDistance, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
match<BLOCK_SIZE, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, typename Dist, typename T>
void match(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const PtrStepSz<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////////
// Match dispatcher
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>& train, float maxDistance, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
}
template <typename Dist, typename T>
void matchDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
}
///////////////////////////////////////////////////////////////////////////////
// Radius Match caller
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchL1_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
//template void matchL2_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchHamming_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchL1_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
//template void matchL2_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchHamming_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
} // namespace bf_radius_match
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/vec_distance.hpp"
#include "opencv2/gpu/device/datamov_utils.hpp"
namespace cv { namespace gpu { namespace device
{
namespace bf_radius_match
{
///////////////////////////////////////////////////////////////////////////////
// Match Unrolled
template <int BLOCK_SIZE, int MAX_DESC_LEN, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void matchUnrolled(const PtrStepSz<T> query, int imgIdx, const PtrStepSz<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 110)
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
#pragma unroll
for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T, typename Mask>
void matchUnrolled(const PtrStepSz<T>& query, const PtrStepSz<T>& train, float maxDistance, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T>
void matchUnrolled(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const PtrStepSz<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////////
// Match
template <int BLOCK_SIZE, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void match(const PtrStepSz<T> query, int imgIdx, const PtrStepSz<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 110)
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
for (int i = 0, endi = (query.cols + BLOCK_SIZE - 1) / BLOCK_SIZE; i < endi; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
template <int BLOCK_SIZE, typename Dist, typename T, typename Mask>
void match(const PtrStepSz<T>& query, const PtrStepSz<T>& train, float maxDistance, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
match<BLOCK_SIZE, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, typename Dist, typename T>
void match(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const PtrStepSz<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////////
// Match dispatcher
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>& train, float maxDistance, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
}
template <typename Dist, typename T>
void matchDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
}
///////////////////////////////////////////////////////////////////////////////
// Radius Match caller
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchL1_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
//template void matchL2_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchHamming_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, float maxDistance, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchL1_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
//template void matchL2_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchHamming_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches, int cc, cudaStream_t stream);
} // namespace bf_radius_match
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

400
modules/gpu/src/cuda/bilateral_filter.cu
View File

@@ -1,201 +1,201 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
using namespace cv::gpu;
typedef unsigned char uchar;
typedef unsigned short ushort;
//////////////////////////////////////////////////////////////////////////////////
/// Bilateral filtering
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
__device__ __forceinline__ float norm_l1(const float& a) { return ::fabs(a); }
__device__ __forceinline__ float norm_l1(const float2& a) { return ::fabs(a.x) + ::fabs(a.y); }
__device__ __forceinline__ float norm_l1(const float3& a) { return ::fabs(a.x) + ::fabs(a.y) + ::fabs(a.z); }
__device__ __forceinline__ float norm_l1(const float4& a) { return ::fabs(a.x) + ::fabs(a.y) + ::fabs(a.z) + ::fabs(a.w); }
__device__ __forceinline__ float sqr(const float& a) { return a * a; }
template<typename T, typename B>
__global__ void bilateral_kernel(const PtrStepSz<T> src, PtrStep<T> dst, const B b, const int ksz, const float sigma_spatial2_inv_half, const float sigma_color2_inv_half)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
int x = threadIdx.x + blockIdx.x * blockDim.x;
int y = threadIdx.y + blockIdx.y * blockDim.y;
if (x >= src.cols || y >= src.rows)
return;
value_type center = saturate_cast<value_type>(src(y, x));
value_type sum1 = VecTraits<value_type>::all(0);
float sum2 = 0;
int r = ksz / 2;
float r2 = (float)(r * r);
int tx = x - r + ksz;
int ty = y - r + ksz;
if (x - ksz/2 >=0 && y - ksz/2 >=0 && tx < src.cols && ty < src.rows)
{
for (int cy = y - r; cy < ty; ++cy)
for (int cx = x - r; cx < tx; ++cx)
{
float space2 = (x - cx) * (x - cx) + (y - cy) * (y - cy);
if (space2 > r2)
continue;
value_type value = saturate_cast<value_type>(src(cy, cx));
float weight = ::exp(space2 * sigma_spatial2_inv_half + sqr(norm_l1(value - center)) * sigma_color2_inv_half);
sum1 = sum1 + weight * value;
sum2 = sum2 + weight;
}
}
else
{
for (int cy = y - r; cy < ty; ++cy)
for (int cx = x - r; cx < tx; ++cx)
{
float space2 = (x - cx) * (x - cx) + (y - cy) * (y - cy);
if (space2 > r2)
continue;
value_type value = saturate_cast<value_type>(b.at(cy, cx, src.data, src.step));
float weight = ::exp(space2 * sigma_spatial2_inv_half + sqr(norm_l1(value - center)) * sigma_color2_inv_half);
sum1 = sum1 + weight * value;
sum2 = sum2 + weight;
}
}
dst(y, x) = saturate_cast<T>(sum1 / sum2);
}
template<typename T, template <typename> class B>
void bilateral_caller(const PtrStepSzb& src, PtrStepSzb dst, int kernel_size, float sigma_spatial, float sigma_color, cudaStream_t stream)
{
dim3 block (32, 8);
dim3 grid (divUp (src.cols, block.x), divUp (src.rows, block.y));
B<T> b(src.rows, src.cols);
float sigma_spatial2_inv_half = -0.5f/(sigma_spatial * sigma_spatial);
float sigma_color2_inv_half = -0.5f/(sigma_color * sigma_color);
cudaSafeCall( cudaFuncSetCacheConfig (bilateral_kernel<T, B<T> >, cudaFuncCachePreferL1) );
bilateral_kernel<<<grid, block>>>((PtrStepSz<T>)src, (PtrStepSz<T>)dst, b, kernel_size, sigma_spatial2_inv_half, sigma_color2_inv_half);
cudaSafeCall ( cudaGetLastError () );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template<typename T>
void bilateral_filter_gpu(const PtrStepSzb& src, PtrStepSzb dst, int kernel_size, float gauss_spatial_coeff, float gauss_color_coeff, int borderMode, cudaStream_t stream)
{
typedef void (*caller_t)(const PtrStepSzb& src, PtrStepSzb dst, int kernel_size, float sigma_spatial, float sigma_color, cudaStream_t stream);
static caller_t funcs[] =
{
bilateral_caller<T, BrdReflect101>,
bilateral_caller<T, BrdReplicate>,
bilateral_caller<T, BrdConstant>,
bilateral_caller<T, BrdReflect>,
bilateral_caller<T, BrdWrap>,
};
funcs[borderMode](src, dst, kernel_size, gauss_spatial_coeff, gauss_color_coeff, stream);
}
}
}}}
#define OCV_INSTANTIATE_BILATERAL_FILTER(T) \
template void cv::gpu::device::imgproc::bilateral_filter_gpu<T>(const PtrStepSzb&, PtrStepSzb, int, float, float, int, cudaStream_t);
OCV_INSTANTIATE_BILATERAL_FILTER(uchar)
//OCV_INSTANTIATE_BILATERAL_FILTER(uchar2)
OCV_INSTANTIATE_BILATERAL_FILTER(uchar3)
OCV_INSTANTIATE_BILATERAL_FILTER(uchar4)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar2)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar3)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar4)
OCV_INSTANTIATE_BILATERAL_FILTER(short)
//OCV_INSTANTIATE_BILATERAL_FILTER(short2)
OCV_INSTANTIATE_BILATERAL_FILTER(short3)
OCV_INSTANTIATE_BILATERAL_FILTER(short4)
OCV_INSTANTIATE_BILATERAL_FILTER(ushort)
//OCV_INSTANTIATE_BILATERAL_FILTER(ushort2)
OCV_INSTANTIATE_BILATERAL_FILTER(ushort3)
OCV_INSTANTIATE_BILATERAL_FILTER(ushort4)
//OCV_INSTANTIATE_BILATERAL_FILTER(int)
//OCV_INSTANTIATE_BILATERAL_FILTER(int2)
//OCV_INSTANTIATE_BILATERAL_FILTER(int3)
//OCV_INSTANTIATE_BILATERAL_FILTER(int4)
OCV_INSTANTIATE_BILATERAL_FILTER(float)
//OCV_INSTANTIATE_BILATERAL_FILTER(float2)
OCV_INSTANTIATE_BILATERAL_FILTER(float3)
OCV_INSTANTIATE_BILATERAL_FILTER(float4)
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
using namespace cv::gpu;
typedef unsigned char uchar;
typedef unsigned short ushort;
//////////////////////////////////////////////////////////////////////////////////
/// Bilateral filtering
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
__device__ __forceinline__ float norm_l1(const float& a) { return ::fabs(a); }
__device__ __forceinline__ float norm_l1(const float2& a) { return ::fabs(a.x) + ::fabs(a.y); }
__device__ __forceinline__ float norm_l1(const float3& a) { return ::fabs(a.x) + ::fabs(a.y) + ::fabs(a.z); }
__device__ __forceinline__ float norm_l1(const float4& a) { return ::fabs(a.x) + ::fabs(a.y) + ::fabs(a.z) + ::fabs(a.w); }
__device__ __forceinline__ float sqr(const float& a) { return a * a; }
template<typename T, typename B>
__global__ void bilateral_kernel(const PtrStepSz<T> src, PtrStep<T> dst, const B b, const int ksz, const float sigma_spatial2_inv_half, const float sigma_color2_inv_half)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
int x = threadIdx.x + blockIdx.x * blockDim.x;
int y = threadIdx.y + blockIdx.y * blockDim.y;
if (x >= src.cols || y >= src.rows)
return;
value_type center = saturate_cast<value_type>(src(y, x));
value_type sum1 = VecTraits<value_type>::all(0);
float sum2 = 0;
int r = ksz / 2;
float r2 = (float)(r * r);
int tx = x - r + ksz;
int ty = y - r + ksz;
if (x - ksz/2 >=0 && y - ksz/2 >=0 && tx < src.cols && ty < src.rows)
{
for (int cy = y - r; cy < ty; ++cy)
for (int cx = x - r; cx < tx; ++cx)
{
float space2 = (x - cx) * (x - cx) + (y - cy) * (y - cy);
if (space2 > r2)
continue;
value_type value = saturate_cast<value_type>(src(cy, cx));
float weight = ::exp(space2 * sigma_spatial2_inv_half + sqr(norm_l1(value - center)) * sigma_color2_inv_half);
sum1 = sum1 + weight * value;
sum2 = sum2 + weight;
}
}
else
{
for (int cy = y - r; cy < ty; ++cy)
for (int cx = x - r; cx < tx; ++cx)
{
float space2 = (x - cx) * (x - cx) + (y - cy) * (y - cy);
if (space2 > r2)
continue;
value_type value = saturate_cast<value_type>(b.at(cy, cx, src.data, src.step));
float weight = ::exp(space2 * sigma_spatial2_inv_half + sqr(norm_l1(value - center)) * sigma_color2_inv_half);
sum1 = sum1 + weight * value;
sum2 = sum2 + weight;
}
}
dst(y, x) = saturate_cast<T>(sum1 / sum2);
}
template<typename T, template <typename> class B>
void bilateral_caller(const PtrStepSzb& src, PtrStepSzb dst, int kernel_size, float sigma_spatial, float sigma_color, cudaStream_t stream)
{
dim3 block (32, 8);
dim3 grid (divUp (src.cols, block.x), divUp (src.rows, block.y));
B<T> b(src.rows, src.cols);
float sigma_spatial2_inv_half = -0.5f/(sigma_spatial * sigma_spatial);
float sigma_color2_inv_half = -0.5f/(sigma_color * sigma_color);
cudaSafeCall( cudaFuncSetCacheConfig (bilateral_kernel<T, B<T> >, cudaFuncCachePreferL1) );
bilateral_kernel<<<grid, block>>>((PtrStepSz<T>)src, (PtrStepSz<T>)dst, b, kernel_size, sigma_spatial2_inv_half, sigma_color2_inv_half);
cudaSafeCall ( cudaGetLastError () );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template<typename T>
void bilateral_filter_gpu(const PtrStepSzb& src, PtrStepSzb dst, int kernel_size, float gauss_spatial_coeff, float gauss_color_coeff, int borderMode, cudaStream_t stream)
{
typedef void (*caller_t)(const PtrStepSzb& src, PtrStepSzb dst, int kernel_size, float sigma_spatial, float sigma_color, cudaStream_t stream);
static caller_t funcs[] =
{
bilateral_caller<T, BrdReflect101>,
bilateral_caller<T, BrdReplicate>,
bilateral_caller<T, BrdConstant>,
bilateral_caller<T, BrdReflect>,
bilateral_caller<T, BrdWrap>,
};
funcs[borderMode](src, dst, kernel_size, gauss_spatial_coeff, gauss_color_coeff, stream);
}
}
}}}
#define OCV_INSTANTIATE_BILATERAL_FILTER(T) \
template void cv::gpu::device::imgproc::bilateral_filter_gpu<T>(const PtrStepSzb&, PtrStepSzb, int, float, float, int, cudaStream_t);
OCV_INSTANTIATE_BILATERAL_FILTER(uchar)
//OCV_INSTANTIATE_BILATERAL_FILTER(uchar2)
OCV_INSTANTIATE_BILATERAL_FILTER(uchar3)
OCV_INSTANTIATE_BILATERAL_FILTER(uchar4)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar2)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar3)
//OCV_INSTANTIATE_BILATERAL_FILTER(schar4)
OCV_INSTANTIATE_BILATERAL_FILTER(short)
//OCV_INSTANTIATE_BILATERAL_FILTER(short2)
OCV_INSTANTIATE_BILATERAL_FILTER(short3)
OCV_INSTANTIATE_BILATERAL_FILTER(short4)
OCV_INSTANTIATE_BILATERAL_FILTER(ushort)
//OCV_INSTANTIATE_BILATERAL_FILTER(ushort2)
OCV_INSTANTIATE_BILATERAL_FILTER(ushort3)
OCV_INSTANTIATE_BILATERAL_FILTER(ushort4)
//OCV_INSTANTIATE_BILATERAL_FILTER(int)
//OCV_INSTANTIATE_BILATERAL_FILTER(int2)
//OCV_INSTANTIATE_BILATERAL_FILTER(int3)
//OCV_INSTANTIATE_BILATERAL_FILTER(int4)
OCV_INSTANTIATE_BILATERAL_FILTER(float)
//OCV_INSTANTIATE_BILATERAL_FILTER(float2)
OCV_INSTANTIATE_BILATERAL_FILTER(float3)
OCV_INSTANTIATE_BILATERAL_FILTER(float4)
#endif /* CUDA_DISABLER */

240
modules/gpu/src/cuda/blend.cu
View File

@@ -1,121 +1,121 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
namespace cv { namespace gpu { namespace device
{
namespace blend
{
template <typename T>
__global__ void blendLinearKernel(int rows, int cols, int cn, const PtrStep<T> img1, const PtrStep<T> img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStep<T> result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y < rows && x < cols)
{
int x_ = x / cn;
float w1 = weights1.ptr(y)[x_];
float w2 = weights2.ptr(y)[x_];
T p1 = img1.ptr(y)[x];
T p2 = img2.ptr(y)[x];
result.ptr(y)[x] = (p1 * w1 + p2 * w2) / (w1 + w2 + 1e-5f);
}
}
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols * cn, threads.x), divUp(rows, threads.y));
blendLinearKernel<<<grid, threads, 0, stream>>>(rows, cols * cn, cn, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template void blendLinearCaller<uchar>(int, int, int, PtrStep<uchar>, PtrStep<uchar>, PtrStepf, PtrStepf, PtrStep<uchar>, cudaStream_t stream);
template void blendLinearCaller<float>(int, int, int, PtrStep<float>, PtrStep<float>, PtrStepf, PtrStepf, PtrStep<float>, cudaStream_t stream);
__global__ void blendLinearKernel8UC4(int rows, int cols, const PtrStepb img1, const PtrStepb img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStepb result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y < rows && x < cols)
{
float w1 = weights1.ptr(y)[x];
float w2 = weights2.ptr(y)[x];
float sum_inv = 1.f / (w1 + w2 + 1e-5f);
w1 *= sum_inv;
w2 *= sum_inv;
uchar4 p1 = ((const uchar4*)img1.ptr(y))[x];
uchar4 p2 = ((const uchar4*)img2.ptr(y))[x];
((uchar4*)result.ptr(y))[x] = make_uchar4(p1.x * w1 + p2.x * w2, p1.y * w1 + p2.y * w2,
p1.z * w1 + p2.z * w2, p1.w * w1 + p2.w * w2);
}
}
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
blendLinearKernel8UC4<<<grid, threads, 0, stream>>>(rows, cols, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
} // namespace blend
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
namespace cv { namespace gpu { namespace device
{
namespace blend
{
template <typename T>
__global__ void blendLinearKernel(int rows, int cols, int cn, const PtrStep<T> img1, const PtrStep<T> img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStep<T> result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y < rows && x < cols)
{
int x_ = x / cn;
float w1 = weights1.ptr(y)[x_];
float w2 = weights2.ptr(y)[x_];
T p1 = img1.ptr(y)[x];
T p2 = img2.ptr(y)[x];
result.ptr(y)[x] = (p1 * w1 + p2 * w2) / (w1 + w2 + 1e-5f);
}
}
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols * cn, threads.x), divUp(rows, threads.y));
blendLinearKernel<<<grid, threads, 0, stream>>>(rows, cols * cn, cn, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template void blendLinearCaller<uchar>(int, int, int, PtrStep<uchar>, PtrStep<uchar>, PtrStepf, PtrStepf, PtrStep<uchar>, cudaStream_t stream);
template void blendLinearCaller<float>(int, int, int, PtrStep<float>, PtrStep<float>, PtrStepf, PtrStepf, PtrStep<float>, cudaStream_t stream);
__global__ void blendLinearKernel8UC4(int rows, int cols, const PtrStepb img1, const PtrStepb img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStepb result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y < rows && x < cols)
{
float w1 = weights1.ptr(y)[x];
float w2 = weights2.ptr(y)[x];
float sum_inv = 1.f / (w1 + w2 + 1e-5f);
w1 *= sum_inv;
w2 *= sum_inv;
uchar4 p1 = ((const uchar4*)img1.ptr(y))[x];
uchar4 p2 = ((const uchar4*)img2.ptr(y))[x];
((uchar4*)result.ptr(y))[x] = make_uchar4(p1.x * w1 + p2.x * w2, p1.y * w1 + p2.y * w2,
p1.z * w1 + p2.z * w2, p1.w * w1 + p2.w * w2);
}
}
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
blendLinearKernel8UC4<<<grid, threads, 0, stream>>>(rows, cols, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
} // namespace blend
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

1000
modules/gpu/src/cuda/canny.cu
View File

File diff suppressed because it is too large Load Diff

766
modules/gpu/src/cuda/color.cu
View File

@@ -1,384 +1,384 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include <internal_shared.hpp>
#include <opencv2/gpu/device/transform.hpp>
#include <opencv2/gpu/device/color.hpp>
#include <cvt_colot_internal.h>
namespace cv { namespace gpu { namespace device
{
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_x = 8 };
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr555_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr565_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR(name, traits) \
void name(const PtrStepSzb& src, const PtrStepSzb& dst, cudaStream_t stream) \
{ \
traits::functor_type functor = traits::create_functor(); \
typedef typename traits::functor_type::argument_type src_t; \
typedef typename traits::functor_type::result_type dst_t; \
cv::gpu::device::transform((PtrStepSz<src_t>)src, (PtrStepSz<dst_t>)dst, functor, WithOutMask(), stream); \
}
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(name) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name, name ## _traits)
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(name) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _8u, name ## _traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _16u, name ## _traits<ushort>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _32f, name ## _traits<float>)
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(name) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _8u, name ## _traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _32f, name ## _traits<float>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_8u, name ## _full_traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_32f, name ## _full_traits<float>)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgra)
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include <internal_shared.hpp>
#include <opencv2/gpu/device/transform.hpp>
#include <opencv2/gpu/device/color.hpp>
#include <cvt_colot_internal.h>
namespace cv { namespace gpu { namespace device
{
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_x = 8 };
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr555_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr565_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR(name, traits) \
void name(const PtrStepSzb& src, const PtrStepSzb& dst, cudaStream_t stream) \
{ \
traits::functor_type functor = traits::create_functor(); \
typedef typename traits::functor_type::argument_type src_t; \
typedef typename traits::functor_type::result_type dst_t; \
cv::gpu::device::transform((PtrStepSz<src_t>)src, (PtrStepSz<dst_t>)dst, functor, WithOutMask(), stream); \
}
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(name) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name, name ## _traits)
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(name) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _8u, name ## _traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _16u, name ## _traits<ushort>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _32f, name ## _traits<float>)
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(name) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _8u, name ## _traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _32f, name ## _traits<float>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_8u, name ## _full_traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_32f, name ## _full_traits<float>)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgra)
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

776
modules/gpu/src/cuda/column_filter.cu
View File

@@ -1,388 +1,388 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/static_check.hpp"
namespace cv { namespace gpu { namespace device
{
namespace column_filter
{
#define MAX_KERNEL_SIZE 32
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float* kernel, int ksize, cudaStream_t stream)
{
if (stream == 0)
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice) );
else
cudaSafeCall( cudaMemcpyToSymbolAsync(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice, stream) );
}
template <int KSIZE, typename T, typename D, typename B>
__global__ void linearColumnFilter(const PtrStepSz<T> src, PtrStep<D> dst, const int anchor, const B brd)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
const int BLOCK_DIM_X = 16;
const int BLOCK_DIM_Y = 16;
const int PATCH_PER_BLOCK = 4;
const int HALO_SIZE = KSIZE <= 16 ? 1 : 2;
#else
const int BLOCK_DIM_X = 16;
const int BLOCK_DIM_Y = 8;
const int PATCH_PER_BLOCK = 2;
const int HALO_SIZE = 2;
#endif
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
__shared__ sum_t smem[(PATCH_PER_BLOCK + 2 * HALO_SIZE) * BLOCK_DIM_Y][BLOCK_DIM_X];
const int x = blockIdx.x * BLOCK_DIM_X + threadIdx.x;
if (x >= src.cols)
return;
const T* src_col = src.ptr() + x;
const int yStart = blockIdx.y * (BLOCK_DIM_Y * PATCH_PER_BLOCK) + threadIdx.y;
if (blockIdx.y > 0)
{
//Upper halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart - (HALO_SIZE - j) * BLOCK_DIM_Y, x));
}
else
{
//Upper halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_low(yStart - (HALO_SIZE - j) * BLOCK_DIM_Y, src_col, src.step));
}
if (blockIdx.y + 2 < gridDim.y)
{
//Main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart + j * BLOCK_DIM_Y, x));
//Lower halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_Y, x));
}
else
{
//Main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + j * BLOCK_DIM_Y, src_col, src.step));
//Lower halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_Y, src_col, src.step));
}
__syncthreads();
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
{
const int y = yStart + j * BLOCK_DIM_Y;
if (y < src.rows)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for (int k = 0; k < KSIZE; ++k)
sum = sum + smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y - anchor + k][threadIdx.x] * c_kernel[k];
dst(y, x) = saturate_cast<D>(sum);
}
}
}
template <int KSIZE, typename T, typename D, template<typename> class B>
void linearColumnFilter_caller(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream)
{
int BLOCK_DIM_X;
int BLOCK_DIM_Y;
int PATCH_PER_BLOCK;
if (cc >= 20)
{
BLOCK_DIM_X = 16;
BLOCK_DIM_Y = 16;
PATCH_PER_BLOCK = 4;
}
else
{
BLOCK_DIM_X = 16;
BLOCK_DIM_Y = 8;
PATCH_PER_BLOCK = 2;
}
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y * PATCH_PER_BLOCK));
B<T> brd(src.rows);
linearColumnFilter<KSIZE, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, brd);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearColumnFilter_gpu(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream);
static const caller_t callers[5][33] =
{
{
0,
linearColumnFilter_caller< 1, T, D, BrdColReflect101>,
linearColumnFilter_caller< 2, T, D, BrdColReflect101>,
linearColumnFilter_caller< 3, T, D, BrdColReflect101>,
linearColumnFilter_caller< 4, T, D, BrdColReflect101>,
linearColumnFilter_caller< 5, T, D, BrdColReflect101>,
linearColumnFilter_caller< 6, T, D, BrdColReflect101>,
linearColumnFilter_caller< 7, T, D, BrdColReflect101>,
linearColumnFilter_caller< 8, T, D, BrdColReflect101>,
linearColumnFilter_caller< 9, T, D, BrdColReflect101>,
linearColumnFilter_caller<10, T, D, BrdColReflect101>,
linearColumnFilter_caller<11, T, D, BrdColReflect101>,
linearColumnFilter_caller<12, T, D, BrdColReflect101>,
linearColumnFilter_caller<13, T, D, BrdColReflect101>,
linearColumnFilter_caller<14, T, D, BrdColReflect101>,
linearColumnFilter_caller<15, T, D, BrdColReflect101>,
linearColumnFilter_caller<16, T, D, BrdColReflect101>,
linearColumnFilter_caller<17, T, D, BrdColReflect101>,
linearColumnFilter_caller<18, T, D, BrdColReflect101>,
linearColumnFilter_caller<19, T, D, BrdColReflect101>,
linearColumnFilter_caller<20, T, D, BrdColReflect101>,
linearColumnFilter_caller<21, T, D, BrdColReflect101>,
linearColumnFilter_caller<22, T, D, BrdColReflect101>,
linearColumnFilter_caller<23, T, D, BrdColReflect101>,
linearColumnFilter_caller<24, T, D, BrdColReflect101>,
linearColumnFilter_caller<25, T, D, BrdColReflect101>,
linearColumnFilter_caller<26, T, D, BrdColReflect101>,
linearColumnFilter_caller<27, T, D, BrdColReflect101>,
linearColumnFilter_caller<28, T, D, BrdColReflect101>,
linearColumnFilter_caller<29, T, D, BrdColReflect101>,
linearColumnFilter_caller<30, T, D, BrdColReflect101>,
linearColumnFilter_caller<31, T, D, BrdColReflect101>,
linearColumnFilter_caller<32, T, D, BrdColReflect101>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColReplicate>,
linearColumnFilter_caller< 2, T, D, BrdColReplicate>,
linearColumnFilter_caller< 3, T, D, BrdColReplicate>,
linearColumnFilter_caller< 4, T, D, BrdColReplicate>,
linearColumnFilter_caller< 5, T, D, BrdColReplicate>,
linearColumnFilter_caller< 6, T, D, BrdColReplicate>,
linearColumnFilter_caller< 7, T, D, BrdColReplicate>,
linearColumnFilter_caller< 8, T, D, BrdColReplicate>,
linearColumnFilter_caller< 9, T, D, BrdColReplicate>,
linearColumnFilter_caller<10, T, D, BrdColReplicate>,
linearColumnFilter_caller<11, T, D, BrdColReplicate>,
linearColumnFilter_caller<12, T, D, BrdColReplicate>,
linearColumnFilter_caller<13, T, D, BrdColReplicate>,
linearColumnFilter_caller<14, T, D, BrdColReplicate>,
linearColumnFilter_caller<15, T, D, BrdColReplicate>,
linearColumnFilter_caller<16, T, D, BrdColReplicate>,
linearColumnFilter_caller<17, T, D, BrdColReplicate>,
linearColumnFilter_caller<18, T, D, BrdColReplicate>,
linearColumnFilter_caller<19, T, D, BrdColReplicate>,
linearColumnFilter_caller<20, T, D, BrdColReplicate>,
linearColumnFilter_caller<21, T, D, BrdColReplicate>,
linearColumnFilter_caller<22, T, D, BrdColReplicate>,
linearColumnFilter_caller<23, T, D, BrdColReplicate>,
linearColumnFilter_caller<24, T, D, BrdColReplicate>,
linearColumnFilter_caller<25, T, D, BrdColReplicate>,
linearColumnFilter_caller<26, T, D, BrdColReplicate>,
linearColumnFilter_caller<27, T, D, BrdColReplicate>,
linearColumnFilter_caller<28, T, D, BrdColReplicate>,
linearColumnFilter_caller<29, T, D, BrdColReplicate>,
linearColumnFilter_caller<30, T, D, BrdColReplicate>,
linearColumnFilter_caller<31, T, D, BrdColReplicate>,
linearColumnFilter_caller<32, T, D, BrdColReplicate>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColConstant>,
linearColumnFilter_caller< 2, T, D, BrdColConstant>,
linearColumnFilter_caller< 3, T, D, BrdColConstant>,
linearColumnFilter_caller< 4, T, D, BrdColConstant>,
linearColumnFilter_caller< 5, T, D, BrdColConstant>,
linearColumnFilter_caller< 6, T, D, BrdColConstant>,
linearColumnFilter_caller< 7, T, D, BrdColConstant>,
linearColumnFilter_caller< 8, T, D, BrdColConstant>,
linearColumnFilter_caller< 9, T, D, BrdColConstant>,
linearColumnFilter_caller<10, T, D, BrdColConstant>,
linearColumnFilter_caller<11, T, D, BrdColConstant>,
linearColumnFilter_caller<12, T, D, BrdColConstant>,
linearColumnFilter_caller<13, T, D, BrdColConstant>,
linearColumnFilter_caller<14, T, D, BrdColConstant>,
linearColumnFilter_caller<15, T, D, BrdColConstant>,
linearColumnFilter_caller<16, T, D, BrdColConstant>,
linearColumnFilter_caller<17, T, D, BrdColConstant>,
linearColumnFilter_caller<18, T, D, BrdColConstant>,
linearColumnFilter_caller<19, T, D, BrdColConstant>,
linearColumnFilter_caller<20, T, D, BrdColConstant>,
linearColumnFilter_caller<21, T, D, BrdColConstant>,
linearColumnFilter_caller<22, T, D, BrdColConstant>,
linearColumnFilter_caller<23, T, D, BrdColConstant>,
linearColumnFilter_caller<24, T, D, BrdColConstant>,
linearColumnFilter_caller<25, T, D, BrdColConstant>,
linearColumnFilter_caller<26, T, D, BrdColConstant>,
linearColumnFilter_caller<27, T, D, BrdColConstant>,
linearColumnFilter_caller<28, T, D, BrdColConstant>,
linearColumnFilter_caller<29, T, D, BrdColConstant>,
linearColumnFilter_caller<30, T, D, BrdColConstant>,
linearColumnFilter_caller<31, T, D, BrdColConstant>,
linearColumnFilter_caller<32, T, D, BrdColConstant>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColReflect>,
linearColumnFilter_caller< 2, T, D, BrdColReflect>,
linearColumnFilter_caller< 3, T, D, BrdColReflect>,
linearColumnFilter_caller< 4, T, D, BrdColReflect>,
linearColumnFilter_caller< 5, T, D, BrdColReflect>,
linearColumnFilter_caller< 6, T, D, BrdColReflect>,
linearColumnFilter_caller< 7, T, D, BrdColReflect>,
linearColumnFilter_caller< 8, T, D, BrdColReflect>,
linearColumnFilter_caller< 9, T, D, BrdColReflect>,
linearColumnFilter_caller<10, T, D, BrdColReflect>,
linearColumnFilter_caller<11, T, D, BrdColReflect>,
linearColumnFilter_caller<12, T, D, BrdColReflect>,
linearColumnFilter_caller<13, T, D, BrdColReflect>,
linearColumnFilter_caller<14, T, D, BrdColReflect>,
linearColumnFilter_caller<15, T, D, BrdColReflect>,
linearColumnFilter_caller<16, T, D, BrdColReflect>,
linearColumnFilter_caller<17, T, D, BrdColReflect>,
linearColumnFilter_caller<18, T, D, BrdColReflect>,
linearColumnFilter_caller<19, T, D, BrdColReflect>,
linearColumnFilter_caller<20, T, D, BrdColReflect>,
linearColumnFilter_caller<21, T, D, BrdColReflect>,
linearColumnFilter_caller<22, T, D, BrdColReflect>,
linearColumnFilter_caller<23, T, D, BrdColReflect>,
linearColumnFilter_caller<24, T, D, BrdColReflect>,
linearColumnFilter_caller<25, T, D, BrdColReflect>,
linearColumnFilter_caller<26, T, D, BrdColReflect>,
linearColumnFilter_caller<27, T, D, BrdColReflect>,
linearColumnFilter_caller<28, T, D, BrdColReflect>,
linearColumnFilter_caller<29, T, D, BrdColReflect>,
linearColumnFilter_caller<30, T, D, BrdColReflect>,
linearColumnFilter_caller<31, T, D, BrdColReflect>,
linearColumnFilter_caller<32, T, D, BrdColReflect>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColWrap>,
linearColumnFilter_caller< 2, T, D, BrdColWrap>,
linearColumnFilter_caller< 3, T, D, BrdColWrap>,
linearColumnFilter_caller< 4, T, D, BrdColWrap>,
linearColumnFilter_caller< 5, T, D, BrdColWrap>,
linearColumnFilter_caller< 6, T, D, BrdColWrap>,
linearColumnFilter_caller< 7, T, D, BrdColWrap>,
linearColumnFilter_caller< 8, T, D, BrdColWrap>,
linearColumnFilter_caller< 9, T, D, BrdColWrap>,
linearColumnFilter_caller<10, T, D, BrdColWrap>,
linearColumnFilter_caller<11, T, D, BrdColWrap>,
linearColumnFilter_caller<12, T, D, BrdColWrap>,
linearColumnFilter_caller<13, T, D, BrdColWrap>,
linearColumnFilter_caller<14, T, D, BrdColWrap>,
linearColumnFilter_caller<15, T, D, BrdColWrap>,
linearColumnFilter_caller<16, T, D, BrdColWrap>,
linearColumnFilter_caller<17, T, D, BrdColWrap>,
linearColumnFilter_caller<18, T, D, BrdColWrap>,
linearColumnFilter_caller<19, T, D, BrdColWrap>,
linearColumnFilter_caller<20, T, D, BrdColWrap>,
linearColumnFilter_caller<21, T, D, BrdColWrap>,
linearColumnFilter_caller<22, T, D, BrdColWrap>,
linearColumnFilter_caller<23, T, D, BrdColWrap>,
linearColumnFilter_caller<24, T, D, BrdColWrap>,
linearColumnFilter_caller<25, T, D, BrdColWrap>,
linearColumnFilter_caller<26, T, D, BrdColWrap>,
linearColumnFilter_caller<27, T, D, BrdColWrap>,
linearColumnFilter_caller<28, T, D, BrdColWrap>,
linearColumnFilter_caller<29, T, D, BrdColWrap>,
linearColumnFilter_caller<30, T, D, BrdColWrap>,
linearColumnFilter_caller<31, T, D, BrdColWrap>,
linearColumnFilter_caller<32, T, D, BrdColWrap>
}
};
loadKernel(kernel, ksize, stream);
callers[brd_type][ksize]((PtrStepSz<T>)src, (PtrStepSz<D>)dst, anchor, cc, stream);
}
template void linearColumnFilter_gpu<float , uchar >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float4, uchar4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float3, short3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float , int >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
} // namespace column_filter
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/static_check.hpp"
namespace cv { namespace gpu { namespace device
{
namespace column_filter
{
#define MAX_KERNEL_SIZE 32
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float* kernel, int ksize, cudaStream_t stream)
{
if (stream == 0)
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice) );
else
cudaSafeCall( cudaMemcpyToSymbolAsync(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice, stream) );
}
template <int KSIZE, typename T, typename D, typename B>
__global__ void linearColumnFilter(const PtrStepSz<T> src, PtrStep<D> dst, const int anchor, const B brd)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
const int BLOCK_DIM_X = 16;
const int BLOCK_DIM_Y = 16;
const int PATCH_PER_BLOCK = 4;
const int HALO_SIZE = KSIZE <= 16 ? 1 : 2;
#else
const int BLOCK_DIM_X = 16;
const int BLOCK_DIM_Y = 8;
const int PATCH_PER_BLOCK = 2;
const int HALO_SIZE = 2;
#endif
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
__shared__ sum_t smem[(PATCH_PER_BLOCK + 2 * HALO_SIZE) * BLOCK_DIM_Y][BLOCK_DIM_X];
const int x = blockIdx.x * BLOCK_DIM_X + threadIdx.x;
if (x >= src.cols)
return;
const T* src_col = src.ptr() + x;
const int yStart = blockIdx.y * (BLOCK_DIM_Y * PATCH_PER_BLOCK) + threadIdx.y;
if (blockIdx.y > 0)
{
//Upper halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart - (HALO_SIZE - j) * BLOCK_DIM_Y, x));
}
else
{
//Upper halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_low(yStart - (HALO_SIZE - j) * BLOCK_DIM_Y, src_col, src.step));
}
if (blockIdx.y + 2 < gridDim.y)
{
//Main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart + j * BLOCK_DIM_Y, x));
//Lower halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_Y, x));
}
else
{
//Main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + j * BLOCK_DIM_Y, src_col, src.step));
//Lower halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_Y, src_col, src.step));
}
__syncthreads();
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
{
const int y = yStart + j * BLOCK_DIM_Y;
if (y < src.rows)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for (int k = 0; k < KSIZE; ++k)
sum = sum + smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y - anchor + k][threadIdx.x] * c_kernel[k];
dst(y, x) = saturate_cast<D>(sum);
}
}
}
template <int KSIZE, typename T, typename D, template<typename> class B>
void linearColumnFilter_caller(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream)
{
int BLOCK_DIM_X;
int BLOCK_DIM_Y;
int PATCH_PER_BLOCK;
if (cc >= 20)
{
BLOCK_DIM_X = 16;
BLOCK_DIM_Y = 16;
PATCH_PER_BLOCK = 4;
}
else
{
BLOCK_DIM_X = 16;
BLOCK_DIM_Y = 8;
PATCH_PER_BLOCK = 2;
}
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y * PATCH_PER_BLOCK));
B<T> brd(src.rows);
linearColumnFilter<KSIZE, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, brd);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearColumnFilter_gpu(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream);
static const caller_t callers[5][33] =
{
{
0,
linearColumnFilter_caller< 1, T, D, BrdColReflect101>,
linearColumnFilter_caller< 2, T, D, BrdColReflect101>,
linearColumnFilter_caller< 3, T, D, BrdColReflect101>,
linearColumnFilter_caller< 4, T, D, BrdColReflect101>,
linearColumnFilter_caller< 5, T, D, BrdColReflect101>,
linearColumnFilter_caller< 6, T, D, BrdColReflect101>,
linearColumnFilter_caller< 7, T, D, BrdColReflect101>,
linearColumnFilter_caller< 8, T, D, BrdColReflect101>,
linearColumnFilter_caller< 9, T, D, BrdColReflect101>,
linearColumnFilter_caller<10, T, D, BrdColReflect101>,
linearColumnFilter_caller<11, T, D, BrdColReflect101>,
linearColumnFilter_caller<12, T, D, BrdColReflect101>,
linearColumnFilter_caller<13, T, D, BrdColReflect101>,
linearColumnFilter_caller<14, T, D, BrdColReflect101>,
linearColumnFilter_caller<15, T, D, BrdColReflect101>,
linearColumnFilter_caller<16, T, D, BrdColReflect101>,
linearColumnFilter_caller<17, T, D, BrdColReflect101>,
linearColumnFilter_caller<18, T, D, BrdColReflect101>,
linearColumnFilter_caller<19, T, D, BrdColReflect101>,
linearColumnFilter_caller<20, T, D, BrdColReflect101>,
linearColumnFilter_caller<21, T, D, BrdColReflect101>,
linearColumnFilter_caller<22, T, D, BrdColReflect101>,
linearColumnFilter_caller<23, T, D, BrdColReflect101>,
linearColumnFilter_caller<24, T, D, BrdColReflect101>,
linearColumnFilter_caller<25, T, D, BrdColReflect101>,
linearColumnFilter_caller<26, T, D, BrdColReflect101>,
linearColumnFilter_caller<27, T, D, BrdColReflect101>,
linearColumnFilter_caller<28, T, D, BrdColReflect101>,
linearColumnFilter_caller<29, T, D, BrdColReflect101>,
linearColumnFilter_caller<30, T, D, BrdColReflect101>,
linearColumnFilter_caller<31, T, D, BrdColReflect101>,
linearColumnFilter_caller<32, T, D, BrdColReflect101>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColReplicate>,
linearColumnFilter_caller< 2, T, D, BrdColReplicate>,
linearColumnFilter_caller< 3, T, D, BrdColReplicate>,
linearColumnFilter_caller< 4, T, D, BrdColReplicate>,
linearColumnFilter_caller< 5, T, D, BrdColReplicate>,
linearColumnFilter_caller< 6, T, D, BrdColReplicate>,
linearColumnFilter_caller< 7, T, D, BrdColReplicate>,
linearColumnFilter_caller< 8, T, D, BrdColReplicate>,
linearColumnFilter_caller< 9, T, D, BrdColReplicate>,
linearColumnFilter_caller<10, T, D, BrdColReplicate>,
linearColumnFilter_caller<11, T, D, BrdColReplicate>,
linearColumnFilter_caller<12, T, D, BrdColReplicate>,
linearColumnFilter_caller<13, T, D, BrdColReplicate>,
linearColumnFilter_caller<14, T, D, BrdColReplicate>,
linearColumnFilter_caller<15, T, D, BrdColReplicate>,
linearColumnFilter_caller<16, T, D, BrdColReplicate>,
linearColumnFilter_caller<17, T, D, BrdColReplicate>,
linearColumnFilter_caller<18, T, D, BrdColReplicate>,
linearColumnFilter_caller<19, T, D, BrdColReplicate>,
linearColumnFilter_caller<20, T, D, BrdColReplicate>,
linearColumnFilter_caller<21, T, D, BrdColReplicate>,
linearColumnFilter_caller<22, T, D, BrdColReplicate>,
linearColumnFilter_caller<23, T, D, BrdColReplicate>,
linearColumnFilter_caller<24, T, D, BrdColReplicate>,
linearColumnFilter_caller<25, T, D, BrdColReplicate>,
linearColumnFilter_caller<26, T, D, BrdColReplicate>,
linearColumnFilter_caller<27, T, D, BrdColReplicate>,
linearColumnFilter_caller<28, T, D, BrdColReplicate>,
linearColumnFilter_caller<29, T, D, BrdColReplicate>,
linearColumnFilter_caller<30, T, D, BrdColReplicate>,
linearColumnFilter_caller<31, T, D, BrdColReplicate>,
linearColumnFilter_caller<32, T, D, BrdColReplicate>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColConstant>,
linearColumnFilter_caller< 2, T, D, BrdColConstant>,
linearColumnFilter_caller< 3, T, D, BrdColConstant>,
linearColumnFilter_caller< 4, T, D, BrdColConstant>,
linearColumnFilter_caller< 5, T, D, BrdColConstant>,
linearColumnFilter_caller< 6, T, D, BrdColConstant>,
linearColumnFilter_caller< 7, T, D, BrdColConstant>,
linearColumnFilter_caller< 8, T, D, BrdColConstant>,
linearColumnFilter_caller< 9, T, D, BrdColConstant>,
linearColumnFilter_caller<10, T, D, BrdColConstant>,
linearColumnFilter_caller<11, T, D, BrdColConstant>,
linearColumnFilter_caller<12, T, D, BrdColConstant>,
linearColumnFilter_caller<13, T, D, BrdColConstant>,
linearColumnFilter_caller<14, T, D, BrdColConstant>,
linearColumnFilter_caller<15, T, D, BrdColConstant>,
linearColumnFilter_caller<16, T, D, BrdColConstant>,
linearColumnFilter_caller<17, T, D, BrdColConstant>,
linearColumnFilter_caller<18, T, D, BrdColConstant>,
linearColumnFilter_caller<19, T, D, BrdColConstant>,
linearColumnFilter_caller<20, T, D, BrdColConstant>,
linearColumnFilter_caller<21, T, D, BrdColConstant>,
linearColumnFilter_caller<22, T, D, BrdColConstant>,
linearColumnFilter_caller<23, T, D, BrdColConstant>,
linearColumnFilter_caller<24, T, D, BrdColConstant>,
linearColumnFilter_caller<25, T, D, BrdColConstant>,
linearColumnFilter_caller<26, T, D, BrdColConstant>,
linearColumnFilter_caller<27, T, D, BrdColConstant>,
linearColumnFilter_caller<28, T, D, BrdColConstant>,
linearColumnFilter_caller<29, T, D, BrdColConstant>,
linearColumnFilter_caller<30, T, D, BrdColConstant>,
linearColumnFilter_caller<31, T, D, BrdColConstant>,
linearColumnFilter_caller<32, T, D, BrdColConstant>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColReflect>,
linearColumnFilter_caller< 2, T, D, BrdColReflect>,
linearColumnFilter_caller< 3, T, D, BrdColReflect>,
linearColumnFilter_caller< 4, T, D, BrdColReflect>,
linearColumnFilter_caller< 5, T, D, BrdColReflect>,
linearColumnFilter_caller< 6, T, D, BrdColReflect>,
linearColumnFilter_caller< 7, T, D, BrdColReflect>,
linearColumnFilter_caller< 8, T, D, BrdColReflect>,
linearColumnFilter_caller< 9, T, D, BrdColReflect>,
linearColumnFilter_caller<10, T, D, BrdColReflect>,
linearColumnFilter_caller<11, T, D, BrdColReflect>,
linearColumnFilter_caller<12, T, D, BrdColReflect>,
linearColumnFilter_caller<13, T, D, BrdColReflect>,
linearColumnFilter_caller<14, T, D, BrdColReflect>,
linearColumnFilter_caller<15, T, D, BrdColReflect>,
linearColumnFilter_caller<16, T, D, BrdColReflect>,
linearColumnFilter_caller<17, T, D, BrdColReflect>,
linearColumnFilter_caller<18, T, D, BrdColReflect>,
linearColumnFilter_caller<19, T, D, BrdColReflect>,
linearColumnFilter_caller<20, T, D, BrdColReflect>,
linearColumnFilter_caller<21, T, D, BrdColReflect>,
linearColumnFilter_caller<22, T, D, BrdColReflect>,
linearColumnFilter_caller<23, T, D, BrdColReflect>,
linearColumnFilter_caller<24, T, D, BrdColReflect>,
linearColumnFilter_caller<25, T, D, BrdColReflect>,
linearColumnFilter_caller<26, T, D, BrdColReflect>,
linearColumnFilter_caller<27, T, D, BrdColReflect>,
linearColumnFilter_caller<28, T, D, BrdColReflect>,
linearColumnFilter_caller<29, T, D, BrdColReflect>,
linearColumnFilter_caller<30, T, D, BrdColReflect>,
linearColumnFilter_caller<31, T, D, BrdColReflect>,
linearColumnFilter_caller<32, T, D, BrdColReflect>
},
{
0,
linearColumnFilter_caller< 1, T, D, BrdColWrap>,
linearColumnFilter_caller< 2, T, D, BrdColWrap>,
linearColumnFilter_caller< 3, T, D, BrdColWrap>,
linearColumnFilter_caller< 4, T, D, BrdColWrap>,
linearColumnFilter_caller< 5, T, D, BrdColWrap>,
linearColumnFilter_caller< 6, T, D, BrdColWrap>,
linearColumnFilter_caller< 7, T, D, BrdColWrap>,
linearColumnFilter_caller< 8, T, D, BrdColWrap>,
linearColumnFilter_caller< 9, T, D, BrdColWrap>,
linearColumnFilter_caller<10, T, D, BrdColWrap>,
linearColumnFilter_caller<11, T, D, BrdColWrap>,
linearColumnFilter_caller<12, T, D, BrdColWrap>,
linearColumnFilter_caller<13, T, D, BrdColWrap>,
linearColumnFilter_caller<14, T, D, BrdColWrap>,
linearColumnFilter_caller<15, T, D, BrdColWrap>,
linearColumnFilter_caller<16, T, D, BrdColWrap>,
linearColumnFilter_caller<17, T, D, BrdColWrap>,
linearColumnFilter_caller<18, T, D, BrdColWrap>,
linearColumnFilter_caller<19, T, D, BrdColWrap>,
linearColumnFilter_caller<20, T, D, BrdColWrap>,
linearColumnFilter_caller<21, T, D, BrdColWrap>,
linearColumnFilter_caller<22, T, D, BrdColWrap>,
linearColumnFilter_caller<23, T, D, BrdColWrap>,
linearColumnFilter_caller<24, T, D, BrdColWrap>,
linearColumnFilter_caller<25, T, D, BrdColWrap>,
linearColumnFilter_caller<26, T, D, BrdColWrap>,
linearColumnFilter_caller<27, T, D, BrdColWrap>,
linearColumnFilter_caller<28, T, D, BrdColWrap>,
linearColumnFilter_caller<29, T, D, BrdColWrap>,
linearColumnFilter_caller<30, T, D, BrdColWrap>,
linearColumnFilter_caller<31, T, D, BrdColWrap>,
linearColumnFilter_caller<32, T, D, BrdColWrap>
}
};
loadKernel(kernel, ksize, stream);
callers[brd_type][ksize]((PtrStepSz<T>)src, (PtrStepSz<D>)dst, anchor, cc, stream);
}
template void linearColumnFilter_gpu<float , uchar >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float4, uchar4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float3, short3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float , int >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
} // namespace column_filter
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

260
modules/gpu/src/cuda/copy_make_border.cu
View File

@@ -1,131 +1,131 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void copyMakeBorder(const Ptr2D src, PtrStepSz<T> dst, int top, int left)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = src(y - top, x - left);
}
template <template <typename> class B, typename T> struct CopyMakeBorderDispatcher
{
static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& dst, int top, int left,
const typename VecTraits<T>::elem_type* borderValue, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<T> brd(src.rows, src.cols, VecTraits<T>::make(borderValue));
BorderReader< PtrStep<T>, B<T> > brdSrc(src, brd);
copyMakeBorder<<<grid, block, 0, stream>>>(brdSrc, dst, top, left);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T, int cn> void copyMakeBorder_gpu(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode,
const T* borderValue, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type vec_type;
typedef void (*caller_t)(const PtrStepSz<vec_type>& src, const PtrStepSz<vec_type>& dst, int top, int left, const T* borderValue, cudaStream_t stream);
static const caller_t callers[5] =
{
CopyMakeBorderDispatcher<BrdReflect101, vec_type>::call,
CopyMakeBorderDispatcher<BrdReplicate, vec_type>::call,
CopyMakeBorderDispatcher<BrdConstant, vec_type>::call,
CopyMakeBorderDispatcher<BrdReflect, vec_type>::call,
CopyMakeBorderDispatcher<BrdWrap, vec_type>::call
};
callers[borderMode](PtrStepSz<vec_type>(src), PtrStepSz<vec_type>(dst), top, left, borderValue, stream);
}
template void copyMakeBorder_gpu<uchar, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<ushort, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<short, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<float, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void copyMakeBorder(const Ptr2D src, PtrStepSz<T> dst, int top, int left)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = src(y - top, x - left);
}
template <template <typename> class B, typename T> struct CopyMakeBorderDispatcher
{
static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& dst, int top, int left,
const typename VecTraits<T>::elem_type* borderValue, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<T> brd(src.rows, src.cols, VecTraits<T>::make(borderValue));
BorderReader< PtrStep<T>, B<T> > brdSrc(src, brd);
copyMakeBorder<<<grid, block, 0, stream>>>(brdSrc, dst, top, left);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T, int cn> void copyMakeBorder_gpu(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode,
const T* borderValue, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type vec_type;
typedef void (*caller_t)(const PtrStepSz<vec_type>& src, const PtrStepSz<vec_type>& dst, int top, int left, const T* borderValue, cudaStream_t stream);
static const caller_t callers[5] =
{
CopyMakeBorderDispatcher<BrdReflect101, vec_type>::call,
CopyMakeBorderDispatcher<BrdReplicate, vec_type>::call,
CopyMakeBorderDispatcher<BrdConstant, vec_type>::call,
CopyMakeBorderDispatcher<BrdReflect, vec_type>::call,
CopyMakeBorderDispatcher<BrdWrap, vec_type>::call
};
callers[borderMode](PtrStepSz<vec_type>(src), PtrStepSz<vec_type>(dst), top, left, borderValue, stream);
}
template void copyMakeBorder_gpu<uchar, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<ushort, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<short, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 1>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<float, 2>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 3>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 4>(const PtrStepSzb& src, const PtrStepSzb& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

4744
modules/gpu/src/cuda/element_operations.cu
View File

File diff suppressed because it is too large Load Diff

778
modules/gpu/src/cuda/fast.cu
View File

File diff suppressed because one or more lines are too long

300
modules/gpu/src/cuda/gftt.cu
View File

@@ -1,151 +1,151 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
// Copyright (c) 2010, Paul Furgale, Chi Hay Tong
//
// The original code was written by Paul Furgale and Chi Hay Tong
// and later optimized and prepared for integration into OpenCV by Itseez.
//
//M*/
#if !defined CUDA_DISABLER
#include <thrust/sort.h>
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
namespace cv { namespace gpu { namespace device
{
namespace gfft
{
texture<float, cudaTextureType2D, cudaReadModeElementType> eigTex(0, cudaFilterModePoint, cudaAddressModeClamp);
__device__ uint g_counter = 0;
template <class Mask> __global__ void findCorners(float threshold, const Mask mask, float2* corners, uint max_count, int rows, int cols)
{
#if __CUDA_ARCH__ >= 110
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (i > 0 && i < rows - 1 && j > 0 && j < cols - 1 && mask(i, j))
{
float val = tex2D(eigTex, j, i);
if (val > threshold)
{
float maxVal = val;
maxVal = ::fmax(tex2D(eigTex, j - 1, i - 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j , i - 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j + 1, i - 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j - 1, i), maxVal);
maxVal = ::fmax(tex2D(eigTex, j + 1, i), maxVal);
maxVal = ::fmax(tex2D(eigTex, j - 1, i + 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j , i + 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j + 1, i + 1), maxVal);
if (val == maxVal)
{
const uint ind = atomicInc(&g_counter, (uint)(-1));
if (ind < max_count)
corners[ind] = make_float2(j, i);
}
}
}
#endif // __CUDA_ARCH__ >= 110
}
int findCorners_gpu(PtrStepSzf eig, float threshold, PtrStepSzb mask, float2* corners, int max_count)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, g_counter) );
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(uint)) );
bindTexture(&eigTex, eig);
dim3 block(16, 16);
dim3 grid(divUp(eig.cols, block.x), divUp(eig.rows, block.y));
if (mask.data)
findCorners<<<grid, block>>>(threshold, SingleMask(mask), corners, max_count, eig.rows, eig.cols);
else
findCorners<<<grid, block>>>(threshold, WithOutMask(), corners, max_count, eig.rows, eig.cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
uint count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(uint), cudaMemcpyDeviceToHost) );
return min(count, max_count);
}
class EigGreater
{
public:
__device__ __forceinline__ bool operator()(float2 a, float2 b) const
{
return tex2D(eigTex, a.x, a.y) > tex2D(eigTex, b.x, b.y);
}
};
void sortCorners_gpu(PtrStepSzf eig, float2* corners, int count)
{
bindTexture(&eigTex, eig);
thrust::device_ptr<float2> ptr(corners);
thrust::sort(ptr, ptr + count, EigGreater());
}
} // namespace optical_flow
}}}
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
// Copyright (c) 2010, Paul Furgale, Chi Hay Tong
//
// The original code was written by Paul Furgale and Chi Hay Tong
// and later optimized and prepared for integration into OpenCV by Itseez.
//
//M*/
#if !defined CUDA_DISABLER
#include <thrust/sort.h>
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
namespace cv { namespace gpu { namespace device
{
namespace gfft
{
texture<float, cudaTextureType2D, cudaReadModeElementType> eigTex(0, cudaFilterModePoint, cudaAddressModeClamp);
__device__ uint g_counter = 0;
template <class Mask> __global__ void findCorners(float threshold, const Mask mask, float2* corners, uint max_count, int rows, int cols)
{
#if __CUDA_ARCH__ >= 110
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (i > 0 && i < rows - 1 && j > 0 && j < cols - 1 && mask(i, j))
{
float val = tex2D(eigTex, j, i);
if (val > threshold)
{
float maxVal = val;
maxVal = ::fmax(tex2D(eigTex, j - 1, i - 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j , i - 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j + 1, i - 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j - 1, i), maxVal);
maxVal = ::fmax(tex2D(eigTex, j + 1, i), maxVal);
maxVal = ::fmax(tex2D(eigTex, j - 1, i + 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j , i + 1), maxVal);
maxVal = ::fmax(tex2D(eigTex, j + 1, i + 1), maxVal);
if (val == maxVal)
{
const uint ind = atomicInc(&g_counter, (uint)(-1));
if (ind < max_count)
corners[ind] = make_float2(j, i);
}
}
}
#endif // __CUDA_ARCH__ >= 110
}
int findCorners_gpu(PtrStepSzf eig, float threshold, PtrStepSzb mask, float2* corners, int max_count)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, g_counter) );
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(uint)) );
bindTexture(&eigTex, eig);
dim3 block(16, 16);
dim3 grid(divUp(eig.cols, block.x), divUp(eig.rows, block.y));
if (mask.data)
findCorners<<<grid, block>>>(threshold, SingleMask(mask), corners, max_count, eig.rows, eig.cols);
else
findCorners<<<grid, block>>>(threshold, WithOutMask(), corners, max_count, eig.rows, eig.cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
uint count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(uint), cudaMemcpyDeviceToHost) );
return min(count, max_count);
}
class EigGreater
{
public:
__device__ __forceinline__ bool operator()(float2 a, float2 b) const
{
return tex2D(eigTex, a.x, a.y) > tex2D(eigTex, b.x, b.y);
}
};
void sortCorners_gpu(PtrStepSzf eig, float2* corners, int count)
{
bindTexture(&eigTex, eig);
thrust::device_ptr<float2> ptr(corners);
thrust::sort(ptr, ptr + count, EigGreater());
}
} // namespace optical_flow
}}}
#endif /* CUDA_DISABLER */

446
modules/gpu/src/cuda/hist.cu
View File

@@ -1,224 +1,224 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
namespace cv { namespace gpu { namespace device
{
#define UINT_BITS 32U
//Warps == subhistograms per threadblock
#define WARP_COUNT 6
//Threadblock size
#define HISTOGRAM256_THREADBLOCK_SIZE (WARP_COUNT * OPENCV_GPU_WARP_SIZE)
#define HISTOGRAM256_BIN_COUNT 256
//Shared memory per threadblock
#define HISTOGRAM256_THREADBLOCK_MEMORY (WARP_COUNT * HISTOGRAM256_BIN_COUNT)
#define PARTIAL_HISTOGRAM256_COUNT 240
#define MERGE_THREADBLOCK_SIZE 256
#define USE_SMEM_ATOMICS (defined (__CUDA_ARCH__) && (__CUDA_ARCH__ >= 120))
namespace hist
{
#if (!USE_SMEM_ATOMICS)
#define TAG_MASK ( (1U << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE)) - 1U )
__forceinline__ __device__ void addByte(volatile uint* s_WarpHist, uint data, uint threadTag)
{
uint count;
do
{
count = s_WarpHist[data] & TAG_MASK;
count = threadTag | (count + 1);
s_WarpHist[data] = count;
} while (s_WarpHist[data] != count);
}
#else
#define TAG_MASK 0xFFFFFFFFU
__forceinline__ __device__ void addByte(uint* s_WarpHist, uint data, uint threadTag)
{
atomicAdd(s_WarpHist + data, 1);
}
#endif
__forceinline__ __device__ void addWord(uint* s_WarpHist, uint data, uint tag, uint pos_x, uint cols)
{
uint x = pos_x << 2;
if (x + 0 < cols) addByte(s_WarpHist, (data >> 0) & 0xFFU, tag);
if (x + 1 < cols) addByte(s_WarpHist, (data >> 8) & 0xFFU, tag);
if (x + 2 < cols) addByte(s_WarpHist, (data >> 16) & 0xFFU, tag);
if (x + 3 < cols) addByte(s_WarpHist, (data >> 24) & 0xFFU, tag);
}
__global__ void histogram256(const PtrStep<uint> d_Data, uint* d_PartialHistograms, uint dataCount, uint cols)
{
//Per-warp subhistogram storage
__shared__ uint s_Hist[HISTOGRAM256_THREADBLOCK_MEMORY];
uint* s_WarpHist= s_Hist + (threadIdx.x >> OPENCV_GPU_LOG_WARP_SIZE) * HISTOGRAM256_BIN_COUNT;
//Clear shared memory storage for current threadblock before processing
#pragma unroll
for (uint i = 0; i < (HISTOGRAM256_THREADBLOCK_MEMORY / HISTOGRAM256_THREADBLOCK_SIZE); i++)
s_Hist[threadIdx.x + i * HISTOGRAM256_THREADBLOCK_SIZE] = 0;
//Cycle through the entire data set, update subhistograms for each warp
const uint tag = threadIdx.x << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE);
__syncthreads();
const uint colsui = d_Data.step / sizeof(uint);
for(uint pos = blockIdx.x * blockDim.x + threadIdx.x; pos < dataCount; pos += blockDim.x * gridDim.x)
{
uint pos_y = pos / colsui;
uint pos_x = pos % colsui;
uint data = d_Data.ptr(pos_y)[pos_x];
addWord(s_WarpHist, data, tag, pos_x, cols);
}
//Merge per-warp histograms into per-block and write to global memory
__syncthreads();
for(uint bin = threadIdx.x; bin < HISTOGRAM256_BIN_COUNT; bin += HISTOGRAM256_THREADBLOCK_SIZE)
{
uint sum = 0;
for (uint i = 0; i < WARP_COUNT; i++)
sum += s_Hist[bin + i * HISTOGRAM256_BIN_COUNT] & TAG_MASK;
d_PartialHistograms[blockIdx.x * HISTOGRAM256_BIN_COUNT + bin] = sum;
}
}
////////////////////////////////////////////////////////////////////////////////
// Merge histogram256() output
// Run one threadblock per bin; each threadblock adds up the same bin counter
// from every partial histogram. Reads are uncoalesced, but mergeHistogram256
// takes only a fraction of total processing time
////////////////////////////////////////////////////////////////////////////////
__global__ void mergeHistogram256(const uint* d_PartialHistograms, int* d_Histogram)
{
uint sum = 0;
#pragma unroll
for (uint i = threadIdx.x; i < PARTIAL_HISTOGRAM256_COUNT; i += MERGE_THREADBLOCK_SIZE)
sum += d_PartialHistograms[blockIdx.x + i * HISTOGRAM256_BIN_COUNT];
__shared__ uint data[MERGE_THREADBLOCK_SIZE];
data[threadIdx.x] = sum;
for (uint stride = MERGE_THREADBLOCK_SIZE / 2; stride > 0; stride >>= 1)
{
__syncthreads();
if(threadIdx.x < stride)
data[threadIdx.x] += data[threadIdx.x + stride];
}
if(threadIdx.x == 0)
d_Histogram[blockIdx.x] = saturate_cast<int>(data[0]);
}
void histogram256_gpu(PtrStepSzb src, int* hist, uint* buf, cudaStream_t stream)
{
histogram256<<<PARTIAL_HISTOGRAM256_COUNT, HISTOGRAM256_THREADBLOCK_SIZE, 0, stream>>>(
PtrStepSz<uint>(src),
buf,
static_cast<uint>(src.rows * src.step / sizeof(uint)),
src.cols);
cudaSafeCall( cudaGetLastError() );
mergeHistogram256<<<HISTOGRAM256_BIN_COUNT, MERGE_THREADBLOCK_SIZE, 0, stream>>>(buf, hist);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
__constant__ int c_lut[256];
__global__ void equalizeHist(const PtrStepSzb src, PtrStepb dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < src.cols && y < src.rows)
{
const uchar val = src.ptr(y)[x];
const int lut = c_lut[val];
dst.ptr(y)[x] = __float2int_rn(255.0f / (src.cols * src.rows) * lut);
}
}
void equalizeHist_gpu(PtrStepSzb src, PtrStepSzb dst, const int* lut, cudaStream_t stream)
{
dim3 block(16, 16);
dim3 grid(divUp(src.cols, block.x), divUp(src.rows, block.y));
cudaSafeCall( cudaMemcpyToSymbol(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice) );
equalizeHist<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace hist
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
namespace cv { namespace gpu { namespace device
{
#define UINT_BITS 32U
//Warps == subhistograms per threadblock
#define WARP_COUNT 6
//Threadblock size
#define HISTOGRAM256_THREADBLOCK_SIZE (WARP_COUNT * OPENCV_GPU_WARP_SIZE)
#define HISTOGRAM256_BIN_COUNT 256
//Shared memory per threadblock
#define HISTOGRAM256_THREADBLOCK_MEMORY (WARP_COUNT * HISTOGRAM256_BIN_COUNT)
#define PARTIAL_HISTOGRAM256_COUNT 240
#define MERGE_THREADBLOCK_SIZE 256
#define USE_SMEM_ATOMICS (defined (__CUDA_ARCH__) && (__CUDA_ARCH__ >= 120))
namespace hist
{
#if (!USE_SMEM_ATOMICS)
#define TAG_MASK ( (1U << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE)) - 1U )
__forceinline__ __device__ void addByte(volatile uint* s_WarpHist, uint data, uint threadTag)
{
uint count;
do
{
count = s_WarpHist[data] & TAG_MASK;
count = threadTag | (count + 1);
s_WarpHist[data] = count;
} while (s_WarpHist[data] != count);
}
#else
#define TAG_MASK 0xFFFFFFFFU
__forceinline__ __device__ void addByte(uint* s_WarpHist, uint data, uint threadTag)
{
atomicAdd(s_WarpHist + data, 1);
}
#endif
__forceinline__ __device__ void addWord(uint* s_WarpHist, uint data, uint tag, uint pos_x, uint cols)
{
uint x = pos_x << 2;
if (x + 0 < cols) addByte(s_WarpHist, (data >> 0) & 0xFFU, tag);
if (x + 1 < cols) addByte(s_WarpHist, (data >> 8) & 0xFFU, tag);
if (x + 2 < cols) addByte(s_WarpHist, (data >> 16) & 0xFFU, tag);
if (x + 3 < cols) addByte(s_WarpHist, (data >> 24) & 0xFFU, tag);
}
__global__ void histogram256(const PtrStep<uint> d_Data, uint* d_PartialHistograms, uint dataCount, uint cols)
{
//Per-warp subhistogram storage
__shared__ uint s_Hist[HISTOGRAM256_THREADBLOCK_MEMORY];
uint* s_WarpHist= s_Hist + (threadIdx.x >> OPENCV_GPU_LOG_WARP_SIZE) * HISTOGRAM256_BIN_COUNT;
//Clear shared memory storage for current threadblock before processing
#pragma unroll
for (uint i = 0; i < (HISTOGRAM256_THREADBLOCK_MEMORY / HISTOGRAM256_THREADBLOCK_SIZE); i++)
s_Hist[threadIdx.x + i * HISTOGRAM256_THREADBLOCK_SIZE] = 0;
//Cycle through the entire data set, update subhistograms for each warp
const uint tag = threadIdx.x << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE);
__syncthreads();
const uint colsui = d_Data.step / sizeof(uint);
for(uint pos = blockIdx.x * blockDim.x + threadIdx.x; pos < dataCount; pos += blockDim.x * gridDim.x)
{
uint pos_y = pos / colsui;
uint pos_x = pos % colsui;
uint data = d_Data.ptr(pos_y)[pos_x];
addWord(s_WarpHist, data, tag, pos_x, cols);
}
//Merge per-warp histograms into per-block and write to global memory
__syncthreads();
for(uint bin = threadIdx.x; bin < HISTOGRAM256_BIN_COUNT; bin += HISTOGRAM256_THREADBLOCK_SIZE)
{
uint sum = 0;
for (uint i = 0; i < WARP_COUNT; i++)
sum += s_Hist[bin + i * HISTOGRAM256_BIN_COUNT] & TAG_MASK;
d_PartialHistograms[blockIdx.x * HISTOGRAM256_BIN_COUNT + bin] = sum;
}
}
////////////////////////////////////////////////////////////////////////////////
// Merge histogram256() output
// Run one threadblock per bin; each threadblock adds up the same bin counter
// from every partial histogram. Reads are uncoalesced, but mergeHistogram256
// takes only a fraction of total processing time
////////////////////////////////////////////////////////////////////////////////
__global__ void mergeHistogram256(const uint* d_PartialHistograms, int* d_Histogram)
{
uint sum = 0;
#pragma unroll
for (uint i = threadIdx.x; i < PARTIAL_HISTOGRAM256_COUNT; i += MERGE_THREADBLOCK_SIZE)
sum += d_PartialHistograms[blockIdx.x + i * HISTOGRAM256_BIN_COUNT];
__shared__ uint data[MERGE_THREADBLOCK_SIZE];
data[threadIdx.x] = sum;
for (uint stride = MERGE_THREADBLOCK_SIZE / 2; stride > 0; stride >>= 1)
{
__syncthreads();
if(threadIdx.x < stride)
data[threadIdx.x] += data[threadIdx.x + stride];
}
if(threadIdx.x == 0)
d_Histogram[blockIdx.x] = saturate_cast<int>(data[0]);
}
void histogram256_gpu(PtrStepSzb src, int* hist, uint* buf, cudaStream_t stream)
{
histogram256<<<PARTIAL_HISTOGRAM256_COUNT, HISTOGRAM256_THREADBLOCK_SIZE, 0, stream>>>(
PtrStepSz<uint>(src),
buf,
static_cast<uint>(src.rows * src.step / sizeof(uint)),
src.cols);
cudaSafeCall( cudaGetLastError() );
mergeHistogram256<<<HISTOGRAM256_BIN_COUNT, MERGE_THREADBLOCK_SIZE, 0, stream>>>(buf, hist);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
__constant__ int c_lut[256];
__global__ void equalizeHist(const PtrStepSzb src, PtrStepb dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < src.cols && y < src.rows)
{
const uchar val = src.ptr(y)[x];
const int lut = c_lut[val];
dst.ptr(y)[x] = __float2int_rn(255.0f / (src.cols * src.rows) * lut);
}
}
void equalizeHist_gpu(PtrStepSzb src, PtrStepSzb dst, const int* lut, cudaStream_t stream)
{
dim3 block(16, 16);
dim3 grid(divUp(src.cols, block.x), divUp(src.rows, block.y));
cudaSafeCall( cudaMemcpyToSymbol(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice) );
equalizeHist<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace hist
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

1740
modules/gpu/src/cuda/hog.cu
View File

File diff suppressed because it is too large Load Diff

2014
modules/gpu/src/cuda/imgproc.cu
View File

File diff suppressed because it is too large Load Diff

200
modules/gpu/src/cuda/internal_shared.hpp
View File

@@ -1,100 +1,100 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_internal_shared_HPP__
#define __OPENCV_internal_shared_HPP__
#include <cuda_runtime.h>
#include <npp.h>
#include "NPP_staging.hpp"
#include "opencv2/gpu/devmem2d.hpp"
#include "safe_call.hpp"
#include "opencv2/gpu/device/common.hpp"
namespace cv { namespace gpu
{
enum
{
BORDER_REFLECT101_GPU = 0,
BORDER_REPLICATE_GPU,
BORDER_CONSTANT_GPU,
BORDER_REFLECT_GPU,
BORDER_WRAP_GPU
};
class NppStreamHandler
{
public:
inline explicit NppStreamHandler(cudaStream_t newStream = 0)
{
oldStream = nppGetStream();
nppSetStream(newStream);
}
inline ~NppStreamHandler()
{
nppSetStream(oldStream);
}
private:
cudaStream_t oldStream;
};
class NppStStreamHandler
{
public:
inline explicit NppStStreamHandler(cudaStream_t newStream = 0)
{
oldStream = nppStSetActiveCUDAstream(newStream);
}
inline ~NppStStreamHandler()
{
nppStSetActiveCUDAstream(oldStream);
}
private:
cudaStream_t oldStream;
};
}}
#endif /* __OPENCV_internal_shared_HPP__ */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_internal_shared_HPP__
#define __OPENCV_internal_shared_HPP__
#include <cuda_runtime.h>
#include <npp.h>
#include "NPP_staging.hpp"
#include "opencv2/gpu/devmem2d.hpp"
#include "safe_call.hpp"
#include "opencv2/gpu/device/common.hpp"
namespace cv { namespace gpu
{
enum
{
BORDER_REFLECT101_GPU = 0,
BORDER_REPLICATE_GPU,
BORDER_CONSTANT_GPU,
BORDER_REFLECT_GPU,
BORDER_WRAP_GPU
};
class NppStreamHandler
{
public:
inline explicit NppStreamHandler(cudaStream_t newStream = 0)
{
oldStream = nppGetStream();
nppSetStream(newStream);
}
inline ~NppStreamHandler()
{
nppSetStream(oldStream);
}
private:
cudaStream_t oldStream;
};
class NppStStreamHandler
{
public:
inline explicit NppStStreamHandler(cudaStream_t newStream = 0)
{
oldStream = nppStSetActiveCUDAstream(newStream);
}
inline ~NppStStreamHandler()
{
nppStSetActiveCUDAstream(oldStream);
}
private:
cudaStream_t oldStream;
};
}}
#endif /* __OPENCV_internal_shared_HPP__ */

1830
modules/gpu/src/cuda/match_template.cu
View File

File diff suppressed because it is too large Load Diff

432
modules/gpu/src/cuda/mathfunc.cu
View File

@@ -1,217 +1,217 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
namespace cv { namespace gpu { namespace device
{
namespace mathfunc
{
//////////////////////////////////////////////////////////////////////////////////////
// Cart <-> Polar
struct Nothing
{
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
{
}
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = ::sqrtf(x_data * x_data + y_data * y_data);
}
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = ::atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
struct NonEmptyMag
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
{
return mag[y * mag_step + x];
}
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
::sincosf(scale * angle_data, &sin_a, &cos_a);
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
}
template <typename Mag, typename Angle>
void cartToPolar_caller(PtrStepSzf x, PtrStepSzf y, PtrStepSzf mag, PtrStepSzf angle, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(x.cols, threads.x);
grid.y = divUp(x.rows, threads.y);
const float scale = angleInDegrees ? (float)(180.0f / CV_PI) : 1.f;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void cartToPolar_gpu(PtrStepSzf x, PtrStepSzf y, PtrStepSzf mag, bool magSqr, PtrStepSzf angle, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSzf x, PtrStepSzf y, PtrStepSzf mag, PtrStepSzf angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
{
{
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
}
},
{
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
}
}
};
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
template <typename Mag>
void polarToCart_caller(PtrStepSzf mag, PtrStepSzf angle, PtrStepSzf x, PtrStepSzf y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(mag.cols, threads.x);
grid.y = divUp(mag.rows, threads.y);
const float scale = angleInDegrees ? (float)(CV_PI / 180.0f) : 1.0f;
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void polarToCart_gpu(PtrStepSzf mag, PtrStepSzf angle, PtrStepSzf x, PtrStepSzf y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSzf mag, PtrStepSzf angle, PtrStepSzf x, PtrStepSzf y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
} // namespace mathfunc
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
namespace cv { namespace gpu { namespace device
{
namespace mathfunc
{
//////////////////////////////////////////////////////////////////////////////////////
// Cart <-> Polar
struct Nothing
{
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
{
}
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = ::sqrtf(x_data * x_data + y_data * y_data);
}
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = ::atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
struct NonEmptyMag
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
{
return mag[y * mag_step + x];
}
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
::sincosf(scale * angle_data, &sin_a, &cos_a);
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
}
template <typename Mag, typename Angle>
void cartToPolar_caller(PtrStepSzf x, PtrStepSzf y, PtrStepSzf mag, PtrStepSzf angle, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(x.cols, threads.x);
grid.y = divUp(x.rows, threads.y);
const float scale = angleInDegrees ? (float)(180.0f / CV_PI) : 1.f;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void cartToPolar_gpu(PtrStepSzf x, PtrStepSzf y, PtrStepSzf mag, bool magSqr, PtrStepSzf angle, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSzf x, PtrStepSzf y, PtrStepSzf mag, PtrStepSzf angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
{
{
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
}
},
{
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
}
}
};
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
template <typename Mag>
void polarToCart_caller(PtrStepSzf mag, PtrStepSzf angle, PtrStepSzf x, PtrStepSzf y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(mag.cols, threads.x);
grid.y = divUp(mag.rows, threads.y);
const float scale = angleInDegrees ? (float)(CV_PI / 180.0f) : 1.0f;
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void polarToCart_gpu(PtrStepSzf mag, PtrStepSzf angle, PtrStepSzf x, PtrStepSzf y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSzf mag, PtrStepSzf angle, PtrStepSzf x, PtrStepSzf y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
} // namespace mathfunc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

4204
modules/gpu/src/cuda/matrix_reductions.cu
View File

File diff suppressed because it is too large Load Diff

72
modules/gpu/src/cuda/nlm.cu
View File

@@ -74,12 +74,12 @@ namespace cv { namespace gpu { namespace device
const int i = blockDim.y * blockIdx.y + threadIdx.y;
const int j = blockDim.x * blockIdx.x + threadIdx.x;
if (j >= dst.cols || i >= dst.rows)
return;
int bsize = search_radius + block_radius;
int search_window = 2 * search_radius + 1;
int search_window = 2 * search_radius + 1;
float minus_search_window2_inv = -1.f/(search_window * search_window);
value_type sum1 = VecTraits<value_type>::all(0);
@@ -87,16 +87,16 @@ namespace cv { namespace gpu { namespace device
if (j - bsize >= 0 && j + bsize < dst.cols && i - bsize >= 0 && i + bsize < dst.rows)
{
for(float y = -search_radius; y <= search_radius; ++y)
for(float y = -search_radius; y <= search_radius; ++y)
for(float x = -search_radius; x <= search_radius; ++x)
{
{
float dist2 = 0;
for(float ty = -block_radius; ty <= block_radius; ++ty)
for(float tx = -block_radius; tx <= block_radius; ++tx)
{
value_type bv = saturate_cast<value_type>(src(i + y + ty, j + x + tx));
value_type av = saturate_cast<value_type>(src(i + ty, j + tx));
dist2 += norm2(av - bv);
}
@@ -119,10 +119,10 @@ namespace cv { namespace gpu { namespace device
for(float tx = -block_radius; tx <= block_radius; ++tx)
{
value_type bv = saturate_cast<value_type>(b.at(i + y + ty, j + x + tx, src));
value_type av = saturate_cast<value_type>(b.at(i + ty, j + tx, src));
value_type av = saturate_cast<value_type>(b.at(i + ty, j + tx, src));
dist2 += norm2(av - bv);
}
float w = __expf(dist2 * noise_mult + (x * x + y * y) * minus_search_window2_inv);
sum1 = sum1 + w * saturate_cast<value_type>(b.at(i + y, j + x, src));
@@ -144,9 +144,9 @@ namespace cv { namespace gpu { namespace device
B<T> b(src.rows, src.cols);
int block_window = 2 * block_radius + 1;
float minus_h2_inv = -1.f/(h * h * VecTraits<T>::cn);
float minus_h2_inv = -1.f/(h * h * VecTraits<T>::cn);
float noise_mult = minus_h2_inv/(block_window * block_window);
cudaSafeCall( cudaFuncSetCacheConfig (nlm_kernel<T, B<T> >, cudaFuncCachePreferL1) );
nlm_kernel<<<grid, block>>>((PtrStepSz<T>)src, (PtrStepSz<T>)dst, b, search_radius, block_radius, noise_mult);
cudaSafeCall ( cudaGetLastError () );
@@ -160,7 +160,7 @@ namespace cv { namespace gpu { namespace device
{
typedef void (*func_t)(const PtrStepSzb src, PtrStepSzb dst, int search_radius, int block_radius, float h, cudaStream_t stream);
static func_t funcs[] =
static func_t funcs[] =
{
nlm_caller<T, BrdReflect101>,
nlm_caller<T, BrdReplicate>,
@@ -183,7 +183,7 @@ namespace cv { namespace gpu { namespace device
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
{
__device__ __forceinline__ int calcDist(const uchar& a, const uchar& b) { return (a-b)*(a-b); }
__device__ __forceinline__ int calcDist(const uchar2& a, const uchar2& b) { return (a.x-b.x)*(a.x-b.x) + (a.y-b.y)*(a.y-b.y); }
__device__ __forceinline__ int calcDist(const uchar3& a, const uchar3& b) { return (a.x-b.x)*(a.x-b.x) + (a.y-b.y)*(a.y-b.y) + (a.z-b.z)*(a.z-b.z); }
@@ -193,7 +193,7 @@ namespace cv { namespace gpu { namespace device
enum
{
CTA_SIZE = 128,
TILE_COLS = 128,
TILE_ROWS = 32,
@@ -217,7 +217,7 @@ namespace cv { namespace gpu { namespace device
PtrStep<T> src;
mutable PtrStepi buffer;
__device__ __forceinline__ void initSums_BruteForce(int i, int j, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
{
for(int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
@@ -256,7 +256,7 @@ namespace cv { namespace gpu { namespace device
int dist = calcDist(src(ay + ty, ax + tx), src(by + ty, bx + tx));
dist_sums[index] += dist;
col_sums(tx + block_radius, index) += dist;
col_sums(tx + block_radius, index) += dist;
}
#endif
@@ -265,9 +265,9 @@ namespace cv { namespace gpu { namespace device
}
__device__ __forceinline__ void shiftRight_FirstRow(int i, int j, int first, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
{
{
for(int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
{
{
int y = index / search_window;
int x = index - y * search_window;
@@ -280,8 +280,8 @@ namespace cv { namespace gpu { namespace device
int col_sum = 0;
for (int ty = -block_radius; ty <= block_radius; ++ty)
col_sum += calcDist(src(ay + ty, ax), src(by + ty, bx));
col_sum += calcDist(src(ay + ty, ax), src(by + ty, bx));
dist_sums[index] += col_sum - col_sums(first, index);
col_sums(first, index) = col_sum;
@@ -298,7 +298,7 @@ namespace cv { namespace gpu { namespace device
T a_down = src(ay + block_radius, ax);
for(int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
{
{
int y = index / search_window;
int x = index - y * search_window;
@@ -309,9 +309,9 @@ namespace cv { namespace gpu { namespace device
T b_down = src(by + block_radius, bx);
int col_sum = up_col_sums(j, index) + calcDist(a_down, b_down) - calcDist(a_up, b_up);
dist_sums[index] += col_sum - col_sums(first, index);
col_sums(first, index) = col_sum;
col_sums(first, index) = col_sum;
up_col_sums(j, index) = col_sum;
}
}
@@ -339,14 +339,14 @@ namespace cv { namespace gpu { namespace device
sum = sum + weight * saturate_cast<sum_type>(src(sy + y, sx + x));
}
volatile __shared__ float cta_buffer[CTA_SIZE];
int tid = threadIdx.x;
cta_buffer[tid] = weights_sum;
__syncthreads();
Block::reduce<CTA_SIZE>(cta_buffer, plus());
Block::reduce<CTA_SIZE>(cta_buffer, plus());
weights_sum = cta_buffer[0];
__syncthreads();
@@ -356,7 +356,7 @@ namespace cv { namespace gpu { namespace device
{
cta_buffer[tid] = reinterpret_cast<float*>(&sum)[n];
__syncthreads();
Block::reduce<CTA_SIZE>(cta_buffer, plus());
Block::reduce<CTA_SIZE>(cta_buffer, plus());
reinterpret_cast<float*>(&sum)[n] = cta_buffer[0];
__syncthreads();
@@ -388,7 +388,7 @@ namespace cv { namespace gpu { namespace device
for (int i = tby; i < tey; ++i)
for (int j = tbx; j < tex; ++j)
{
{
__syncthreads();
if (j == tbx)
@@ -397,7 +397,7 @@ namespace cv { namespace gpu { namespace device
first = 0;
}
else
{
{
if (i == tby)
shiftRight_FirstRow(i, j, first, dist_sums, col_sums, up_col_sums);
else
@@ -407,7 +407,7 @@ namespace cv { namespace gpu { namespace device
}
__syncthreads();
convolve_window(i, j, dist_sums, col_sums, up_col_sums, dst(i, j));
}
}
@@ -418,7 +418,7 @@ namespace cv { namespace gpu { namespace device
__global__ void fast_nlm_kernel(const FastNonLocalMenas<T> fnlm, PtrStepSz<T> dst) { fnlm(dst); }
void nln_fast_get_buffer_size(const PtrStepSzb& src, int search_window, int block_window, int& buffer_cols, int& buffer_rows)
{
{
typedef FastNonLocalMenas<uchar> FNLM;
dim3 grid(divUp(src.cols, FNLM::TILE_COLS), divUp(src.rows, FNLM::TILE_ROWS));
@@ -434,13 +434,13 @@ namespace cv { namespace gpu { namespace device
FNLM fnlm(search_window, block_window, h);
fnlm.src = (PtrStepSz<T>)src;
fnlm.buffer = buffer;
fnlm.buffer = buffer;
dim3 block(FNLM::CTA_SIZE, 1);
dim3 grid(divUp(src.cols, FNLM::TILE_COLS), divUp(src.rows, FNLM::TILE_ROWS));
int smem = search_window * search_window * sizeof(int);
fast_nlm_kernel<<<grid, block, smem>>>(fnlm, (PtrStepSz<T>)dst);
cudaSafeCall ( cudaGetLastError () );
if (stream == 0)
@@ -451,7 +451,7 @@ namespace cv { namespace gpu { namespace device
template void nlm_fast_gpu<uchar2>(const PtrStepSzb&, PtrStepSzb, PtrStepi, int, int, float, cudaStream_t);
template void nlm_fast_gpu<uchar3>(const PtrStepSzb&, PtrStepSzb, PtrStepi, int, int, float, cudaStream_t);
__global__ void fnlm_split_kernel(const PtrStepSz<uchar3> lab, PtrStepb l, PtrStep<uchar2> ab)
{
@@ -462,10 +462,10 @@ namespace cv { namespace gpu { namespace device
{
uchar3 p = lab(y, x);
ab(y,x) = make_uchar2(p.y, p.z);
l(y,x) = p.x;
l(y,x) = p.x;
}
}
void fnlm_split_channels(const PtrStepSz<uchar3>& lab, PtrStepb l, PtrStep<uchar2> ab, cudaStream_t stream)
{
dim3 b(32, 8);
@@ -485,7 +485,7 @@ namespace cv { namespace gpu { namespace device
if (x < lab.cols && y < lab.rows)
{
uchar2 p = ab(y, x);
lab(y, x) = make_uchar3(l(y, x), p.x, p.y);
lab(y, x) = make_uchar3(l(y, x), p.x, p.y);
}
}

438
modules/gpu/src/cuda/optical_flow.cu
View File

@@ -1,220 +1,220 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
namespace cv { namespace gpu { namespace device
{
namespace optical_flow
{
#define NEEDLE_MAP_SCALE 16
#define NUM_VERTS_PER_ARROW 6
__global__ void NeedleMapAverageKernel(const PtrStepSzf u, const PtrStepf v, PtrStepf u_avg, PtrStepf v_avg)
{
__shared__ float smem[2 * NEEDLE_MAP_SCALE];
volatile float* u_col_sum = smem;
volatile float* v_col_sum = u_col_sum + NEEDLE_MAP_SCALE;
const int x = blockIdx.x * NEEDLE_MAP_SCALE + threadIdx.x;
const int y = blockIdx.y * NEEDLE_MAP_SCALE;
u_col_sum[threadIdx.x] = 0;
v_col_sum[threadIdx.x] = 0;
#pragma unroll
for(int i = 0; i < NEEDLE_MAP_SCALE; ++i)
{
u_col_sum[threadIdx.x] += u(::min(y + i, u.rows - 1), x);
v_col_sum[threadIdx.x] += v(::min(y + i, u.rows - 1), x);
}
if (threadIdx.x < 8)
{
// now add the column sums
const uint X = threadIdx.x;
if (X | 0xfe == 0xfe) // bit 0 is 0
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 1];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 1];
}
if (X | 0xfe == 0xfc) // bits 0 & 1 == 0
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 2];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 2];
}
if (X | 0xf8 == 0xf8)
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 4];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 4];
}
if (X == 0)
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 8];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 8];
}
}
if (threadIdx.x == 0)
{
const float coeff = 1.0f / (NEEDLE_MAP_SCALE * NEEDLE_MAP_SCALE);
u_col_sum[0] *= coeff;
v_col_sum[0] *= coeff;
u_avg(blockIdx.y, blockIdx.x) = u_col_sum[0];
v_avg(blockIdx.y, blockIdx.x) = v_col_sum[0];
}
}
void NeedleMapAverage_gpu(PtrStepSzf u, PtrStepSzf v, PtrStepSzf u_avg, PtrStepSzf v_avg)
{
const dim3 block(NEEDLE_MAP_SCALE);
const dim3 grid(u_avg.cols, u_avg.rows);
NeedleMapAverageKernel<<<grid, block>>>(u, v, u_avg, v_avg);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
__global__ void NeedleMapVertexKernel(const PtrStepSzf u_avg, const PtrStepf v_avg, float* vertex_data, float* color_data, float max_flow, float xscale, float yscale)
{
// test - just draw a triangle at each pixel
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const float arrow_x = x * NEEDLE_MAP_SCALE + NEEDLE_MAP_SCALE / 2.0f;
const float arrow_y = y * NEEDLE_MAP_SCALE + NEEDLE_MAP_SCALE / 2.0f;
float3 v[NUM_VERTS_PER_ARROW];
if (x < u_avg.cols && y < u_avg.rows)
{
const float u_avg_val = u_avg(y, x);
const float v_avg_val = v_avg(y, x);
const float theta = ::atan2f(v_avg_val, u_avg_val);// + CV_PI;
float r = ::sqrtf(v_avg_val * v_avg_val + u_avg_val * u_avg_val);
r = fmin(14.0f * (r / max_flow), 14.0f);
v[0].z = 1.0f;
v[1].z = 0.7f;
v[2].z = 0.7f;
v[3].z = 0.7f;
v[4].z = 0.7f;
v[5].z = 1.0f;
v[0].x = arrow_x;
v[0].y = arrow_y;
v[5].x = arrow_x;
v[5].y = arrow_y;
v[2].x = arrow_x + r * ::cosf(theta);
v[2].y = arrow_y + r * ::sinf(theta);
v[3].x = v[2].x;
v[3].y = v[2].y;
r = ::fmin(r, 2.5f);
v[1].x = arrow_x + r * ::cosf(theta - CV_PI / 2.0f);
v[1].y = arrow_y + r * ::sinf(theta - CV_PI / 2.0f);
v[4].x = arrow_x + r * ::cosf(theta + CV_PI / 2.0f);
v[4].y = arrow_y + r * ::sinf(theta + CV_PI / 2.0f);
int indx = (y * u_avg.cols + x) * NUM_VERTS_PER_ARROW * 3;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[0].x * xscale;
vertex_data[indx++] = v[0].y * yscale;
vertex_data[indx++] = v[0].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[1].x * xscale;
vertex_data[indx++] = v[1].y * yscale;
vertex_data[indx++] = v[1].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[2].x * xscale;
vertex_data[indx++] = v[2].y * yscale;
vertex_data[indx++] = v[2].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[3].x * xscale;
vertex_data[indx++] = v[3].y * yscale;
vertex_data[indx++] = v[3].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[4].x * xscale;
vertex_data[indx++] = v[4].y * yscale;
vertex_data[indx++] = v[4].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[5].x * xscale;
vertex_data[indx++] = v[5].y * yscale;
vertex_data[indx++] = v[5].z;
}
}
void CreateOpticalFlowNeedleMap_gpu(PtrStepSzf u_avg, PtrStepSzf v_avg, float* vertex_buffer, float* color_data, float max_flow, float xscale, float yscale)
{
const dim3 block(16);
const dim3 grid(divUp(u_avg.cols, block.x), divUp(u_avg.rows, block.y));
NeedleMapVertexKernel<<<grid, block>>>(u_avg, v_avg, vertex_buffer, color_data, max_flow, xscale, yscale);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
}
}}}
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
namespace cv { namespace gpu { namespace device
{
namespace optical_flow
{
#define NEEDLE_MAP_SCALE 16
#define NUM_VERTS_PER_ARROW 6
__global__ void NeedleMapAverageKernel(const PtrStepSzf u, const PtrStepf v, PtrStepf u_avg, PtrStepf v_avg)
{
__shared__ float smem[2 * NEEDLE_MAP_SCALE];
volatile float* u_col_sum = smem;
volatile float* v_col_sum = u_col_sum + NEEDLE_MAP_SCALE;
const int x = blockIdx.x * NEEDLE_MAP_SCALE + threadIdx.x;
const int y = blockIdx.y * NEEDLE_MAP_SCALE;
u_col_sum[threadIdx.x] = 0;
v_col_sum[threadIdx.x] = 0;
#pragma unroll
for(int i = 0; i < NEEDLE_MAP_SCALE; ++i)
{
u_col_sum[threadIdx.x] += u(::min(y + i, u.rows - 1), x);
v_col_sum[threadIdx.x] += v(::min(y + i, u.rows - 1), x);
}
if (threadIdx.x < 8)
{
// now add the column sums
const uint X = threadIdx.x;
if (X | 0xfe == 0xfe) // bit 0 is 0
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 1];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 1];
}
if (X | 0xfe == 0xfc) // bits 0 & 1 == 0
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 2];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 2];
}
if (X | 0xf8 == 0xf8)
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 4];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 4];
}
if (X == 0)
{
u_col_sum[threadIdx.x] += u_col_sum[threadIdx.x + 8];
v_col_sum[threadIdx.x] += v_col_sum[threadIdx.x + 8];
}
}
if (threadIdx.x == 0)
{
const float coeff = 1.0f / (NEEDLE_MAP_SCALE * NEEDLE_MAP_SCALE);
u_col_sum[0] *= coeff;
v_col_sum[0] *= coeff;
u_avg(blockIdx.y, blockIdx.x) = u_col_sum[0];
v_avg(blockIdx.y, blockIdx.x) = v_col_sum[0];
}
}
void NeedleMapAverage_gpu(PtrStepSzf u, PtrStepSzf v, PtrStepSzf u_avg, PtrStepSzf v_avg)
{
const dim3 block(NEEDLE_MAP_SCALE);
const dim3 grid(u_avg.cols, u_avg.rows);
NeedleMapAverageKernel<<<grid, block>>>(u, v, u_avg, v_avg);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
__global__ void NeedleMapVertexKernel(const PtrStepSzf u_avg, const PtrStepf v_avg, float* vertex_data, float* color_data, float max_flow, float xscale, float yscale)
{
// test - just draw a triangle at each pixel
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const float arrow_x = x * NEEDLE_MAP_SCALE + NEEDLE_MAP_SCALE / 2.0f;
const float arrow_y = y * NEEDLE_MAP_SCALE + NEEDLE_MAP_SCALE / 2.0f;
float3 v[NUM_VERTS_PER_ARROW];
if (x < u_avg.cols && y < u_avg.rows)
{
const float u_avg_val = u_avg(y, x);
const float v_avg_val = v_avg(y, x);
const float theta = ::atan2f(v_avg_val, u_avg_val);// + CV_PI;
float r = ::sqrtf(v_avg_val * v_avg_val + u_avg_val * u_avg_val);
r = fmin(14.0f * (r / max_flow), 14.0f);
v[0].z = 1.0f;
v[1].z = 0.7f;
v[2].z = 0.7f;
v[3].z = 0.7f;
v[4].z = 0.7f;
v[5].z = 1.0f;
v[0].x = arrow_x;
v[0].y = arrow_y;
v[5].x = arrow_x;
v[5].y = arrow_y;
v[2].x = arrow_x + r * ::cosf(theta);
v[2].y = arrow_y + r * ::sinf(theta);
v[3].x = v[2].x;
v[3].y = v[2].y;
r = ::fmin(r, 2.5f);
v[1].x = arrow_x + r * ::cosf(theta - CV_PI / 2.0f);
v[1].y = arrow_y + r * ::sinf(theta - CV_PI / 2.0f);
v[4].x = arrow_x + r * ::cosf(theta + CV_PI / 2.0f);
v[4].y = arrow_y + r * ::sinf(theta + CV_PI / 2.0f);
int indx = (y * u_avg.cols + x) * NUM_VERTS_PER_ARROW * 3;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[0].x * xscale;
vertex_data[indx++] = v[0].y * yscale;
vertex_data[indx++] = v[0].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[1].x * xscale;
vertex_data[indx++] = v[1].y * yscale;
vertex_data[indx++] = v[1].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[2].x * xscale;
vertex_data[indx++] = v[2].y * yscale;
vertex_data[indx++] = v[2].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[3].x * xscale;
vertex_data[indx++] = v[3].y * yscale;
vertex_data[indx++] = v[3].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[4].x * xscale;
vertex_data[indx++] = v[4].y * yscale;
vertex_data[indx++] = v[4].z;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
vertex_data[indx++] = v[5].x * xscale;
vertex_data[indx++] = v[5].y * yscale;
vertex_data[indx++] = v[5].z;
}
}
void CreateOpticalFlowNeedleMap_gpu(PtrStepSzf u_avg, PtrStepSzf v_avg, float* vertex_buffer, float* color_data, float max_flow, float xscale, float yscale)
{
const dim3 block(16);
const dim3 grid(divUp(u_avg.cols, block.x), divUp(u_avg.rows, block.y));
NeedleMapVertexKernel<<<grid, block>>>(u_avg, v_avg, vertex_buffer, color_data, max_flow, xscale, yscale);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
}
}}}
#endif /* CUDA_DISABLER */

842
modules/gpu/src/cuda/orb.cu
View File

@@ -1,422 +1,422 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
// Copyright (c) 2010, Paul Furgale, Chi Hay Tong
//
// The original code was written by Paul Furgale and Chi Hay Tong
// and later optimized and prepared for integration into OpenCV by Itseez.
//
//M*/
#if !defined CUDA_DISABLER
#include <thrust/sort.h>
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/functional.hpp"
namespace cv { namespace gpu { namespace device
{
namespace orb
{
////////////////////////////////////////////////////////////////////////////////////////////////////////
// cull
int cull_gpu(int* loc, float* response, int size, int n_points)
{
thrust::device_ptr<int> loc_ptr(loc);
thrust::device_ptr<float> response_ptr(response);
thrust::sort_by_key(response_ptr, response_ptr + size, loc_ptr, thrust::greater<float>());
return n_points;
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// HarrisResponses
__global__ void HarrisResponses(const PtrStepb img, const short2* loc_, float* response, const int npoints, const int blockSize, const float harris_k)
{
__shared__ int smem[8 * 32];
volatile int* srow = smem + threadIdx.y * blockDim.x;
const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
if (ptidx < npoints)
{
const short2 loc = loc_[ptidx];
const int r = blockSize / 2;
const int x0 = loc.x - r;
const int y0 = loc.y - r;
int a = 0, b = 0, c = 0;
for (int ind = threadIdx.x; ind < blockSize * blockSize; ind += blockDim.x)
{
const int i = ind / blockSize;
const int j = ind % blockSize;
int Ix = (img(y0 + i, x0 + j + 1) - img(y0 + i, x0 + j - 1)) * 2 +
(img(y0 + i - 1, x0 + j + 1) - img(y0 + i - 1, x0 + j - 1)) +
(img(y0 + i + 1, x0 + j + 1) - img(y0 + i + 1, x0 + j - 1));
int Iy = (img(y0 + i + 1, x0 + j) - img(y0 + i - 1, x0 + j)) * 2 +
(img(y0 + i + 1, x0 + j - 1) - img(y0 + i - 1, x0 + j - 1)) +
(img(y0 + i + 1, x0 + j + 1) - img(y0 + i - 1, x0 + j + 1));
a += Ix * Ix;
b += Iy * Iy;
c += Ix * Iy;
}
reduce<32>(srow, a, threadIdx.x, plus<volatile int>());
reduce<32>(srow, b, threadIdx.x, plus<volatile int>());
reduce<32>(srow, c, threadIdx.x, plus<volatile int>());
if (threadIdx.x == 0)
{
float scale = (1 << 2) * blockSize * 255.0f;
scale = 1.0f / scale;
const float scale_sq_sq = scale * scale * scale * scale;
response[ptidx] = ((float)a * b - (float)c * c - harris_k * ((float)a + b) * ((float)a + b)) * scale_sq_sq;
}
}
}
void HarrisResponses_gpu(PtrStepSzb img, const short2* loc, float* response, const int npoints, int blockSize, float harris_k, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid;
grid.x = divUp(npoints, block.y);
HarrisResponses<<<grid, block, 0, stream>>>(img, loc, response, npoints, blockSize, harris_k);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// IC_Angle
__constant__ int c_u_max[32];
void loadUMax(const int* u_max, int count)
{
cudaSafeCall( cudaMemcpyToSymbol(c_u_max, u_max, count * sizeof(int)) );
}
__global__ void IC_Angle(const PtrStepb image, const short2* loc_, float* angle, const int npoints, const int half_k)
{
__shared__ int smem[8 * 32];
volatile int* srow = smem + threadIdx.y * blockDim.x;
const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
if (ptidx < npoints)
{
int m_01 = 0, m_10 = 0;
const short2 loc = loc_[ptidx];
// Treat the center line differently, v=0
for (int u = threadIdx.x - half_k; u <= half_k; u += blockDim.x)
m_10 += u * image(loc.y, loc.x + u);
reduce<32>(srow, m_10, threadIdx.x, plus<volatile int>());
for (int v = 1; v <= half_k; ++v)
{
// Proceed over the two lines
int v_sum = 0;
int m_sum = 0;
const int d = c_u_max[v];
for (int u = threadIdx.x - d; u <= d; u += blockDim.x)
{
int val_plus = image(loc.y + v, loc.x + u);
int val_minus = image(loc.y - v, loc.x + u);
v_sum += (val_plus - val_minus);
m_sum += u * (val_plus + val_minus);
}
reduce<32>(srow, v_sum, threadIdx.x, plus<volatile int>());
reduce<32>(srow, m_sum, threadIdx.x, plus<volatile int>());
m_10 += m_sum;
m_01 += v * v_sum;
}
if (threadIdx.x == 0)
{
float kp_dir = ::atan2f((float)m_01, (float)m_10);
kp_dir += (kp_dir < 0) * (2.0f * CV_PI);
kp_dir *= 180.0f / CV_PI;
angle[ptidx] = kp_dir;
}
}
}
void IC_Angle_gpu(PtrStepSzb image, const short2* loc, float* angle, int npoints, int half_k, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid;
grid.x = divUp(npoints, block.y);
IC_Angle<<<grid, block, 0, stream>>>(image, loc, angle, npoints, half_k);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// computeOrbDescriptor
template <int WTA_K> struct OrbDescriptor;
#define GET_VALUE(idx) \
img(loc.y + __float2int_rn(pattern_x[idx] * sina + pattern_y[idx] * cosa), \
loc.x + __float2int_rn(pattern_x[idx] * cosa - pattern_y[idx] * sina))
template <> struct OrbDescriptor<2>
{
__device__ static int calc(const PtrStepb& img, short2 loc, const int* pattern_x, const int* pattern_y, float sina, float cosa, int i)
{
pattern_x += 16 * i;
pattern_y += 16 * i;
int t0, t1, val;
t0 = GET_VALUE(0); t1 = GET_VALUE(1);
val = t0 < t1;
t0 = GET_VALUE(2); t1 = GET_VALUE(3);
val |= (t0 < t1) << 1;
t0 = GET_VALUE(4); t1 = GET_VALUE(5);
val |= (t0 < t1) << 2;
t0 = GET_VALUE(6); t1 = GET_VALUE(7);
val |= (t0 < t1) << 3;
t0 = GET_VALUE(8); t1 = GET_VALUE(9);
val |= (t0 < t1) << 4;
t0 = GET_VALUE(10); t1 = GET_VALUE(11);
val |= (t0 < t1) << 5;
t0 = GET_VALUE(12); t1 = GET_VALUE(13);
val |= (t0 < t1) << 6;
t0 = GET_VALUE(14); t1 = GET_VALUE(15);
val |= (t0 < t1) << 7;
return val;
}
};
template <> struct OrbDescriptor<3>
{
__device__ static int calc(const PtrStepb& img, short2 loc, const int* pattern_x, const int* pattern_y, float sina, float cosa, int i)
{
pattern_x += 12 * i;
pattern_y += 12 * i;
int t0, t1, t2, val;
t0 = GET_VALUE(0); t1 = GET_VALUE(1); t2 = GET_VALUE(2);
val = t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0);
t0 = GET_VALUE(3); t1 = GET_VALUE(4); t2 = GET_VALUE(5);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 2;
t0 = GET_VALUE(6); t1 = GET_VALUE(7); t2 = GET_VALUE(8);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 4;
t0 = GET_VALUE(9); t1 = GET_VALUE(10); t2 = GET_VALUE(11);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 6;
return val;
}
};
template <> struct OrbDescriptor<4>
{
__device__ static int calc(const PtrStepb& img, short2 loc, const int* pattern_x, const int* pattern_y, float sina, float cosa, int i)
{
pattern_x += 16 * i;
pattern_y += 16 * i;
int t0, t1, t2, t3, k, val;
int a, b;
t0 = GET_VALUE(0); t1 = GET_VALUE(1);
t2 = GET_VALUE(2); t3 = GET_VALUE(3);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val = k;
t0 = GET_VALUE(4); t1 = GET_VALUE(5);
t2 = GET_VALUE(6); t3 = GET_VALUE(7);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val |= k << 2;
t0 = GET_VALUE(8); t1 = GET_VALUE(9);
t2 = GET_VALUE(10); t3 = GET_VALUE(11);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val |= k << 4;
t0 = GET_VALUE(12); t1 = GET_VALUE(13);
t2 = GET_VALUE(14); t3 = GET_VALUE(15);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val |= k << 6;
return val;
}
};
#undef GET_VALUE
template <int WTA_K>
__global__ void computeOrbDescriptor(const PtrStepb img, const short2* loc, const float* angle_, const int npoints,
const int* pattern_x, const int* pattern_y, PtrStepb desc, int dsize)
{
const int descidx = blockIdx.x * blockDim.x + threadIdx.x;
const int ptidx = blockIdx.y * blockDim.y + threadIdx.y;
if (ptidx < npoints && descidx < dsize)
{
float angle = angle_[ptidx];
angle *= (float)(CV_PI / 180.f);
float sina, cosa;
::sincosf(angle, &sina, &cosa);
desc.ptr(ptidx)[descidx] = OrbDescriptor<WTA_K>::calc(img, loc[ptidx], pattern_x, pattern_y, sina, cosa, descidx);
}
}
void computeOrbDescriptor_gpu(PtrStepb img, const short2* loc, const float* angle, const int npoints,
const int* pattern_x, const int* pattern_y, PtrStepb desc, int dsize, int WTA_K, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid;
grid.x = divUp(dsize, block.x);
grid.y = divUp(npoints, block.y);
switch (WTA_K)
{
case 2:
computeOrbDescriptor<2><<<grid, block, 0, stream>>>(img, loc, angle, npoints, pattern_x, pattern_y, desc, dsize);
break;
case 3:
computeOrbDescriptor<3><<<grid, block, 0, stream>>>(img, loc, angle, npoints, pattern_x, pattern_y, desc, dsize);
break;
case 4:
computeOrbDescriptor<4><<<grid, block, 0, stream>>>(img, loc, angle, npoints, pattern_x, pattern_y, desc, dsize);
break;
}
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// mergeLocation
__global__ void mergeLocation(const short2* loc_, float* x, float* y, const int npoints, float scale)
{
const int ptidx = blockIdx.x * blockDim.x + threadIdx.x;
if (ptidx < npoints)
{
short2 loc = loc_[ptidx];
x[ptidx] = loc.x * scale;
y[ptidx] = loc.y * scale;
}
}
void mergeLocation_gpu(const short2* loc, float* x, float* y, int npoints, float scale, cudaStream_t stream)
{
dim3 block(256);
dim3 grid;
grid.x = divUp(npoints, block.x);
mergeLocation<<<grid, block, 0, stream>>>(loc, x, y, npoints, scale);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
}}}
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
// Copyright (c) 2010, Paul Furgale, Chi Hay Tong
//
// The original code was written by Paul Furgale and Chi Hay Tong
// and later optimized and prepared for integration into OpenCV by Itseez.
//
//M*/
#if !defined CUDA_DISABLER
#include <thrust/sort.h>
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/functional.hpp"
namespace cv { namespace gpu { namespace device
{
namespace orb
{
////////////////////////////////////////////////////////////////////////////////////////////////////////
// cull
int cull_gpu(int* loc, float* response, int size, int n_points)
{
thrust::device_ptr<int> loc_ptr(loc);
thrust::device_ptr<float> response_ptr(response);
thrust::sort_by_key(response_ptr, response_ptr + size, loc_ptr, thrust::greater<float>());
return n_points;
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// HarrisResponses
__global__ void HarrisResponses(const PtrStepb img, const short2* loc_, float* response, const int npoints, const int blockSize, const float harris_k)
{
__shared__ int smem[8 * 32];
volatile int* srow = smem + threadIdx.y * blockDim.x;
const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
if (ptidx < npoints)
{
const short2 loc = loc_[ptidx];
const int r = blockSize / 2;
const int x0 = loc.x - r;
const int y0 = loc.y - r;
int a = 0, b = 0, c = 0;
for (int ind = threadIdx.x; ind < blockSize * blockSize; ind += blockDim.x)
{
const int i = ind / blockSize;
const int j = ind % blockSize;
int Ix = (img(y0 + i, x0 + j + 1) - img(y0 + i, x0 + j - 1)) * 2 +
(img(y0 + i - 1, x0 + j + 1) - img(y0 + i - 1, x0 + j - 1)) +
(img(y0 + i + 1, x0 + j + 1) - img(y0 + i + 1, x0 + j - 1));
int Iy = (img(y0 + i + 1, x0 + j) - img(y0 + i - 1, x0 + j)) * 2 +
(img(y0 + i + 1, x0 + j - 1) - img(y0 + i - 1, x0 + j - 1)) +
(img(y0 + i + 1, x0 + j + 1) - img(y0 + i - 1, x0 + j + 1));
a += Ix * Ix;
b += Iy * Iy;
c += Ix * Iy;
}
reduce<32>(srow, a, threadIdx.x, plus<volatile int>());
reduce<32>(srow, b, threadIdx.x, plus<volatile int>());
reduce<32>(srow, c, threadIdx.x, plus<volatile int>());
if (threadIdx.x == 0)
{
float scale = (1 << 2) * blockSize * 255.0f;
scale = 1.0f / scale;
const float scale_sq_sq = scale * scale * scale * scale;
response[ptidx] = ((float)a * b - (float)c * c - harris_k * ((float)a + b) * ((float)a + b)) * scale_sq_sq;
}
}
}
void HarrisResponses_gpu(PtrStepSzb img, const short2* loc, float* response, const int npoints, int blockSize, float harris_k, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid;
grid.x = divUp(npoints, block.y);
HarrisResponses<<<grid, block, 0, stream>>>(img, loc, response, npoints, blockSize, harris_k);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// IC_Angle
__constant__ int c_u_max[32];
void loadUMax(const int* u_max, int count)
{
cudaSafeCall( cudaMemcpyToSymbol(c_u_max, u_max, count * sizeof(int)) );
}
__global__ void IC_Angle(const PtrStepb image, const short2* loc_, float* angle, const int npoints, const int half_k)
{
__shared__ int smem[8 * 32];
volatile int* srow = smem + threadIdx.y * blockDim.x;
const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
if (ptidx < npoints)
{
int m_01 = 0, m_10 = 0;
const short2 loc = loc_[ptidx];
// Treat the center line differently, v=0
for (int u = threadIdx.x - half_k; u <= half_k; u += blockDim.x)
m_10 += u * image(loc.y, loc.x + u);
reduce<32>(srow, m_10, threadIdx.x, plus<volatile int>());
for (int v = 1; v <= half_k; ++v)
{
// Proceed over the two lines
int v_sum = 0;
int m_sum = 0;
const int d = c_u_max[v];
for (int u = threadIdx.x - d; u <= d; u += blockDim.x)
{
int val_plus = image(loc.y + v, loc.x + u);
int val_minus = image(loc.y - v, loc.x + u);
v_sum += (val_plus - val_minus);
m_sum += u * (val_plus + val_minus);
}
reduce<32>(srow, v_sum, threadIdx.x, plus<volatile int>());
reduce<32>(srow, m_sum, threadIdx.x, plus<volatile int>());
m_10 += m_sum;
m_01 += v * v_sum;
}
if (threadIdx.x == 0)
{
float kp_dir = ::atan2f((float)m_01, (float)m_10);
kp_dir += (kp_dir < 0) * (2.0f * CV_PI);
kp_dir *= 180.0f / CV_PI;
angle[ptidx] = kp_dir;
}
}
}
void IC_Angle_gpu(PtrStepSzb image, const short2* loc, float* angle, int npoints, int half_k, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid;
grid.x = divUp(npoints, block.y);
IC_Angle<<<grid, block, 0, stream>>>(image, loc, angle, npoints, half_k);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// computeOrbDescriptor
template <int WTA_K> struct OrbDescriptor;
#define GET_VALUE(idx) \
img(loc.y + __float2int_rn(pattern_x[idx] * sina + pattern_y[idx] * cosa), \
loc.x + __float2int_rn(pattern_x[idx] * cosa - pattern_y[idx] * sina))
template <> struct OrbDescriptor<2>
{
__device__ static int calc(const PtrStepb& img, short2 loc, const int* pattern_x, const int* pattern_y, float sina, float cosa, int i)
{
pattern_x += 16 * i;
pattern_y += 16 * i;
int t0, t1, val;
t0 = GET_VALUE(0); t1 = GET_VALUE(1);
val = t0 < t1;
t0 = GET_VALUE(2); t1 = GET_VALUE(3);
val |= (t0 < t1) << 1;
t0 = GET_VALUE(4); t1 = GET_VALUE(5);
val |= (t0 < t1) << 2;
t0 = GET_VALUE(6); t1 = GET_VALUE(7);
val |= (t0 < t1) << 3;
t0 = GET_VALUE(8); t1 = GET_VALUE(9);
val |= (t0 < t1) << 4;
t0 = GET_VALUE(10); t1 = GET_VALUE(11);
val |= (t0 < t1) << 5;
t0 = GET_VALUE(12); t1 = GET_VALUE(13);
val |= (t0 < t1) << 6;
t0 = GET_VALUE(14); t1 = GET_VALUE(15);
val |= (t0 < t1) << 7;
return val;
}
};
template <> struct OrbDescriptor<3>
{
__device__ static int calc(const PtrStepb& img, short2 loc, const int* pattern_x, const int* pattern_y, float sina, float cosa, int i)
{
pattern_x += 12 * i;
pattern_y += 12 * i;
int t0, t1, t2, val;
t0 = GET_VALUE(0); t1 = GET_VALUE(1); t2 = GET_VALUE(2);
val = t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0);
t0 = GET_VALUE(3); t1 = GET_VALUE(4); t2 = GET_VALUE(5);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 2;
t0 = GET_VALUE(6); t1 = GET_VALUE(7); t2 = GET_VALUE(8);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 4;
t0 = GET_VALUE(9); t1 = GET_VALUE(10); t2 = GET_VALUE(11);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 6;
return val;
}
};
template <> struct OrbDescriptor<4>
{
__device__ static int calc(const PtrStepb& img, short2 loc, const int* pattern_x, const int* pattern_y, float sina, float cosa, int i)
{
pattern_x += 16 * i;
pattern_y += 16 * i;
int t0, t1, t2, t3, k, val;
int a, b;
t0 = GET_VALUE(0); t1 = GET_VALUE(1);
t2 = GET_VALUE(2); t3 = GET_VALUE(3);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val = k;
t0 = GET_VALUE(4); t1 = GET_VALUE(5);
t2 = GET_VALUE(6); t3 = GET_VALUE(7);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val |= k << 2;
t0 = GET_VALUE(8); t1 = GET_VALUE(9);
t2 = GET_VALUE(10); t3 = GET_VALUE(11);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val |= k << 4;
t0 = GET_VALUE(12); t1 = GET_VALUE(13);
t2 = GET_VALUE(14); t3 = GET_VALUE(15);
a = 0, b = 2;
if( t1 > t0 ) t0 = t1, a = 1;
if( t3 > t2 ) t2 = t3, b = 3;
k = t0 > t2 ? a : b;
val |= k << 6;
return val;
}
};
#undef GET_VALUE
template <int WTA_K>
__global__ void computeOrbDescriptor(const PtrStepb img, const short2* loc, const float* angle_, const int npoints,
const int* pattern_x, const int* pattern_y, PtrStepb desc, int dsize)
{
const int descidx = blockIdx.x * blockDim.x + threadIdx.x;
const int ptidx = blockIdx.y * blockDim.y + threadIdx.y;
if (ptidx < npoints && descidx < dsize)
{
float angle = angle_[ptidx];
angle *= (float)(CV_PI / 180.f);
float sina, cosa;
::sincosf(angle, &sina, &cosa);
desc.ptr(ptidx)[descidx] = OrbDescriptor<WTA_K>::calc(img, loc[ptidx], pattern_x, pattern_y, sina, cosa, descidx);
}
}
void computeOrbDescriptor_gpu(PtrStepb img, const short2* loc, const float* angle, const int npoints,
const int* pattern_x, const int* pattern_y, PtrStepb desc, int dsize, int WTA_K, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid;
grid.x = divUp(dsize, block.x);
grid.y = divUp(npoints, block.y);
switch (WTA_K)
{
case 2:
computeOrbDescriptor<2><<<grid, block, 0, stream>>>(img, loc, angle, npoints, pattern_x, pattern_y, desc, dsize);
break;
case 3:
computeOrbDescriptor<3><<<grid, block, 0, stream>>>(img, loc, angle, npoints, pattern_x, pattern_y, desc, dsize);
break;
case 4:
computeOrbDescriptor<4><<<grid, block, 0, stream>>>(img, loc, angle, npoints, pattern_x, pattern_y, desc, dsize);
break;
}
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// mergeLocation
__global__ void mergeLocation(const short2* loc_, float* x, float* y, const int npoints, float scale)
{
const int ptidx = blockIdx.x * blockDim.x + threadIdx.x;
if (ptidx < npoints)
{
short2 loc = loc_[ptidx];
x[ptidx] = loc.x * scale;
y[ptidx] = loc.y * scale;
}
}
void mergeLocation_gpu(const short2* loc, float* x, float* y, int npoints, float scale, cudaStream_t stream)
{
dim3 block(256);
dim3 grid;
grid.x = divUp(npoints, block.x);
mergeLocation<<<grid, block, 0, stream>>>(loc, x, y, npoints, scale);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
}}}
#endif /* CUDA_DISABLER */

454
modules/gpu/src/cuda/pyr_down.cu
View File

@@ -1,228 +1,228 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename T, typename B> __global__ void pyrDown(const PtrStepSz<T> src, PtrStep<T> dst, const B b, int dst_cols)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_t;
__shared__ work_t smem[256 + 4];
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y;
const int src_y = 2 * y;
if (src_y >= 2 && src_y < src.rows - 2 && x >= 2 && x < src.cols - 2)
{
{
work_t sum;
sum = 0.0625f * src(src_y - 2, x);
sum = sum + 0.25f * src(src_y - 1, x);
sum = sum + 0.375f * src(src_y , x);
sum = sum + 0.25f * src(src_y + 1, x);
sum = sum + 0.0625f * src(src_y + 2, x);
smem[2 + threadIdx.x] = sum;
}
if (threadIdx.x < 2)
{
const int left_x = x - 2;
work_t sum;
sum = 0.0625f * src(src_y - 2, left_x);
sum = sum + 0.25f * src(src_y - 1, left_x);
sum = sum + 0.375f * src(src_y , left_x);
sum = sum + 0.25f * src(src_y + 1, left_x);
sum = sum + 0.0625f * src(src_y + 2, left_x);
smem[threadIdx.x] = sum;
}
if (threadIdx.x > 253)
{
const int right_x = x + 2;
work_t sum;
sum = 0.0625f * src(src_y - 2, right_x);
sum = sum + 0.25f * src(src_y - 1, right_x);
sum = sum + 0.375f * src(src_y , right_x);
sum = sum + 0.25f * src(src_y + 1, right_x);
sum = sum + 0.0625f * src(src_y + 2, right_x);
smem[4 + threadIdx.x] = sum;
}
}
else
{
{
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col_high(x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col_high(x));
sum = sum + 0.375f * src(src_y , b.idx_col_high(x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col_high(x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col_high(x));
smem[2 + threadIdx.x] = sum;
}
if (threadIdx.x < 2)
{
const int left_x = x - 2;
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col(left_x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col(left_x));
sum = sum + 0.375f * src(src_y , b.idx_col(left_x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col(left_x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col(left_x));
smem[threadIdx.x] = sum;
}
if (threadIdx.x > 253)
{
const int right_x = x + 2;
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col_high(right_x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col_high(right_x));
sum = sum + 0.375f * src(src_y , b.idx_col_high(right_x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col_high(right_x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col_high(right_x));
smem[4 + threadIdx.x] = sum;
}
}
__syncthreads();
if (threadIdx.x < 128)
{
const int tid2 = threadIdx.x * 2;
work_t sum;
sum = 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
sum = sum + 0.0625f * smem[2 + tid2 + 2];
const int dst_x = (blockIdx.x * blockDim.x + tid2) / 2;
if (dst_x < dst_cols)
dst.ptr(y)[dst_x] = saturate_cast<T>(sum);
}
}
template <typename T, template <typename> class B> void pyrDown_caller(PtrStepSz<T> src, PtrStepSz<T> dst, cudaStream_t stream)
{
const dim3 block(256);
const dim3 grid(divUp(src.cols, block.x), dst.rows);
B<T> b(src.rows, src.cols);
pyrDown<T><<<grid, block, 0, stream>>>(src, dst, b, dst.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T> void pyrDown_gpu(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream)
{
pyrDown_caller<T, BrdReflect101>(static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(dst), stream);
}
template void pyrDown_gpu<uchar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<uchar2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<uchar3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<uchar4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<schar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<char2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<char3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<char4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<ushort>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<ushort2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<ushort3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<ushort4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<short>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<short2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<short3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<short4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<float>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<float2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<float3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<float4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename T, typename B> __global__ void pyrDown(const PtrStepSz<T> src, PtrStep<T> dst, const B b, int dst_cols)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_t;
__shared__ work_t smem[256 + 4];
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y;
const int src_y = 2 * y;
if (src_y >= 2 && src_y < src.rows - 2 && x >= 2 && x < src.cols - 2)
{
{
work_t sum;
sum = 0.0625f * src(src_y - 2, x);
sum = sum + 0.25f * src(src_y - 1, x);
sum = sum + 0.375f * src(src_y , x);
sum = sum + 0.25f * src(src_y + 1, x);
sum = sum + 0.0625f * src(src_y + 2, x);
smem[2 + threadIdx.x] = sum;
}
if (threadIdx.x < 2)
{
const int left_x = x - 2;
work_t sum;
sum = 0.0625f * src(src_y - 2, left_x);
sum = sum + 0.25f * src(src_y - 1, left_x);
sum = sum + 0.375f * src(src_y , left_x);
sum = sum + 0.25f * src(src_y + 1, left_x);
sum = sum + 0.0625f * src(src_y + 2, left_x);
smem[threadIdx.x] = sum;
}
if (threadIdx.x > 253)
{
const int right_x = x + 2;
work_t sum;
sum = 0.0625f * src(src_y - 2, right_x);
sum = sum + 0.25f * src(src_y - 1, right_x);
sum = sum + 0.375f * src(src_y , right_x);
sum = sum + 0.25f * src(src_y + 1, right_x);
sum = sum + 0.0625f * src(src_y + 2, right_x);
smem[4 + threadIdx.x] = sum;
}
}
else
{
{
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col_high(x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col_high(x));
sum = sum + 0.375f * src(src_y , b.idx_col_high(x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col_high(x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col_high(x));
smem[2 + threadIdx.x] = sum;
}
if (threadIdx.x < 2)
{
const int left_x = x - 2;
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col(left_x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col(left_x));
sum = sum + 0.375f * src(src_y , b.idx_col(left_x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col(left_x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col(left_x));
smem[threadIdx.x] = sum;
}
if (threadIdx.x > 253)
{
const int right_x = x + 2;
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col_high(right_x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col_high(right_x));
sum = sum + 0.375f * src(src_y , b.idx_col_high(right_x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col_high(right_x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col_high(right_x));
smem[4 + threadIdx.x] = sum;
}
}
__syncthreads();
if (threadIdx.x < 128)
{
const int tid2 = threadIdx.x * 2;
work_t sum;
sum = 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
sum = sum + 0.0625f * smem[2 + tid2 + 2];
const int dst_x = (blockIdx.x * blockDim.x + tid2) / 2;
if (dst_x < dst_cols)
dst.ptr(y)[dst_x] = saturate_cast<T>(sum);
}
}
template <typename T, template <typename> class B> void pyrDown_caller(PtrStepSz<T> src, PtrStepSz<T> dst, cudaStream_t stream)
{
const dim3 block(256);
const dim3 grid(divUp(src.cols, block.x), dst.rows);
B<T> b(src.rows, src.cols);
pyrDown<T><<<grid, block, 0, stream>>>(src, dst, b, dst.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T> void pyrDown_gpu(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream)
{
pyrDown_caller<T, BrdReflect101>(static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(dst), stream);
}
template void pyrDown_gpu<uchar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<uchar2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<uchar3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<uchar4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<schar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<char2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<char3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<char4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<ushort>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<ushort2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<ushort3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<ushort4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<short>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<short2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<short3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<short4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<int4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<float>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrDown_gpu<float2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<float3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrDown_gpu<float4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

390
modules/gpu/src/cuda/pyr_up.cu
View File

@@ -1,196 +1,196 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename T> __global__ void pyrUp(const PtrStepSz<T> src, PtrStepSz<T> dst)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
__shared__ sum_t s_srcPatch[10][10];
__shared__ sum_t s_dstPatch[20][16];
if (threadIdx.x < 10 && threadIdx.y < 10)
{
int srcx = static_cast<int>((blockIdx.x * blockDim.x) / 2 + threadIdx.x) - 1;
int srcy = static_cast<int>((blockIdx.y * blockDim.y) / 2 + threadIdx.y) - 1;
srcx = ::abs(srcx);
srcx = ::min(src.cols - 1, srcx);
srcy = ::abs(srcy);
srcy = ::min(src.rows - 1, srcy);
s_srcPatch[threadIdx.y][threadIdx.x] = saturate_cast<sum_t>(src(srcy, srcx));
}
__syncthreads();
sum_t sum = VecTraits<sum_t>::all(0);
const int evenFlag = static_cast<int>((threadIdx.x & 1) == 0);
const int oddFlag = static_cast<int>((threadIdx.x & 1) != 0);
const bool eveny = ((threadIdx.y & 1) == 0);
const int tidx = threadIdx.x;
if (eveny)
{
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * 0.375f ) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[2 + threadIdx.y][threadIdx.x] = sum;
if (threadIdx.y < 2)
{
sum = VecTraits<sum_t>::all(0);
if (eveny)
{
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[0][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[0][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * 0.375f ) * s_srcPatch[0][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[0][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[0][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[threadIdx.y][threadIdx.x] = sum;
}
if (threadIdx.y > 13)
{
sum = VecTraits<sum_t>::all(0);
if (eveny)
{
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[9][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[9][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * 0.375f ) * s_srcPatch[9][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[9][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[9][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[4 + threadIdx.y][threadIdx.x] = sum;
}
__syncthreads();
sum = VecTraits<sum_t>::all(0);
const int tidy = threadIdx.y;
sum = sum + 0.0625f * s_dstPatch[2 + tidy - 2][threadIdx.x];
sum = sum + 0.25f * s_dstPatch[2 + tidy - 1][threadIdx.x];
sum = sum + 0.375f * s_dstPatch[2 + tidy ][threadIdx.x];
sum = sum + 0.25f * s_dstPatch[2 + tidy + 1][threadIdx.x];
sum = sum + 0.0625f * s_dstPatch[2 + tidy + 2][threadIdx.x];
if (x < dst.cols && y < dst.rows)
dst(y, x) = saturate_cast<T>(4.0f * sum);
}
template <typename T> void pyrUp_caller(PtrStepSz<T> src, PtrStepSz<T> dst, cudaStream_t stream)
{
const dim3 block(16, 16);
const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
pyrUp<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T> void pyrUp_gpu(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream)
{
pyrUp_caller<T>(static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(dst), stream);
}
template void pyrUp_gpu<uchar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<uchar2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<uchar3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<uchar4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<schar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<char2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<char3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<char4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<ushort>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<ushort2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<ushort3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<ushort4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<short>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<short2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<short3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<short4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<float>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<float2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<float3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<float4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename T> __global__ void pyrUp(const PtrStepSz<T> src, PtrStepSz<T> dst)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
__shared__ sum_t s_srcPatch[10][10];
__shared__ sum_t s_dstPatch[20][16];
if (threadIdx.x < 10 && threadIdx.y < 10)
{
int srcx = static_cast<int>((blockIdx.x * blockDim.x) / 2 + threadIdx.x) - 1;
int srcy = static_cast<int>((blockIdx.y * blockDim.y) / 2 + threadIdx.y) - 1;
srcx = ::abs(srcx);
srcx = ::min(src.cols - 1, srcx);
srcy = ::abs(srcy);
srcy = ::min(src.rows - 1, srcy);
s_srcPatch[threadIdx.y][threadIdx.x] = saturate_cast<sum_t>(src(srcy, srcx));
}
__syncthreads();
sum_t sum = VecTraits<sum_t>::all(0);
const int evenFlag = static_cast<int>((threadIdx.x & 1) == 0);
const int oddFlag = static_cast<int>((threadIdx.x & 1) != 0);
const bool eveny = ((threadIdx.y & 1) == 0);
const int tidx = threadIdx.x;
if (eveny)
{
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * 0.375f ) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[1 + (threadIdx.y >> 1)][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[2 + threadIdx.y][threadIdx.x] = sum;
if (threadIdx.y < 2)
{
sum = VecTraits<sum_t>::all(0);
if (eveny)
{
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[0][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[0][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * 0.375f ) * s_srcPatch[0][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[0][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[0][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[threadIdx.y][threadIdx.x] = sum;
}
if (threadIdx.y > 13)
{
sum = VecTraits<sum_t>::all(0);
if (eveny)
{
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[9][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[9][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * 0.375f ) * s_srcPatch[9][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * 0.25f ) * s_srcPatch[9][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * 0.0625f) * s_srcPatch[9][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[4 + threadIdx.y][threadIdx.x] = sum;
}
__syncthreads();
sum = VecTraits<sum_t>::all(0);
const int tidy = threadIdx.y;
sum = sum + 0.0625f * s_dstPatch[2 + tidy - 2][threadIdx.x];
sum = sum + 0.25f * s_dstPatch[2 + tidy - 1][threadIdx.x];
sum = sum + 0.375f * s_dstPatch[2 + tidy ][threadIdx.x];
sum = sum + 0.25f * s_dstPatch[2 + tidy + 1][threadIdx.x];
sum = sum + 0.0625f * s_dstPatch[2 + tidy + 2][threadIdx.x];
if (x < dst.cols && y < dst.rows)
dst(y, x) = saturate_cast<T>(4.0f * sum);
}
template <typename T> void pyrUp_caller(PtrStepSz<T> src, PtrStepSz<T> dst, cudaStream_t stream)
{
const dim3 block(16, 16);
const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
pyrUp<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T> void pyrUp_gpu(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream)
{
pyrUp_caller<T>(static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(dst), stream);
}
template void pyrUp_gpu<uchar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<uchar2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<uchar3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<uchar4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<schar>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<char2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<char3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<char4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<ushort>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<ushort2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<ushort3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<ushort4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<short>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<short2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<short3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<short4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<int4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<float>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
//template void pyrUp_gpu<float2>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<float3>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
template void pyrUp_gpu<float4>(PtrStepSzb src, PtrStepSzb dst, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

1366
modules/gpu/src/cuda/pyrlk.cu
View File

File diff suppressed because it is too large Load Diff

546
modules/gpu/src/cuda/remap.cu
View File

@@ -1,274 +1,274 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void remap(const Ptr2D src, const PtrStepf mapx, const PtrStepf mapy, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = mapx.ptr(y)[x];
const float ycoo = mapy.ptr(y)[x];
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherStream
{
static void call(PtrStepSz<T> src, PtrStepSzf mapx, PtrStepSzf mapy, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block, 0, stream>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherNonStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy, PtrStepSz<T> dst, const float* borderValue, int)
{
(void)srcWhole;
(void)xoff;
(void)yoff;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_REMAP_TEX(type) \
texture< type , cudaTextureType2D> tex_remap_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_remap_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
int xoff, yoff; \
tex_remap_ ## type ## _reader (int xoff_, int yoff_) : xoff(xoff_), yoff(yoff_) {} \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_remap_ ## type , x + xoff, y + yoff); \
} \
}; \
template <template <typename> class Filter, template <typename> class B> struct RemapDispatcherNonStream<Filter, B, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy, \
PtrStepSz< type > dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , srcWhole); \
tex_remap_ ## type ##_reader texSrc(xoff, yoff); \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_remap_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <template <typename> class Filter> struct RemapDispatcherNonStream<Filter, BrdReplicate, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy, \
PtrStepSz< type > dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , srcWhole); \
tex_remap_ ## type ##_reader texSrc(xoff, yoff); \
if (srcWhole.cols == src.cols && srcWhole.rows == src.rows) \
{ \
Filter< tex_remap_ ## type ##_reader > filter_src(texSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
} \
else \
{ \
BrdReplicate<type> brd(src.rows, src.cols); \
BorderReader< tex_remap_ ## type ##_reader, BrdReplicate<type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, BrdReplicate<type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
} \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(short2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(float2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_REMAP_TEX
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcher
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy,
PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
RemapDispatcherNonStream<Filter, B, T>::call(src, srcWhole, xoff, yoff, mapx, mapy, dst, borderValue, cc);
else
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc);
}
};
template <typename T> void remap_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap,
PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap,
PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc);
static const caller_t callers[3][5] =
{
{
RemapDispatcher<PointFilter, BrdReflect101, T>::call,
RemapDispatcher<PointFilter, BrdReplicate, T>::call,
RemapDispatcher<PointFilter, BrdConstant, T>::call,
RemapDispatcher<PointFilter, BrdReflect, T>::call,
RemapDispatcher<PointFilter, BrdWrap, T>::call
},
{
RemapDispatcher<LinearFilter, BrdReflect101, T>::call,
RemapDispatcher<LinearFilter, BrdReplicate, T>::call,
RemapDispatcher<LinearFilter, BrdConstant, T>::call,
RemapDispatcher<LinearFilter, BrdReflect, T>::call,
RemapDispatcher<LinearFilter, BrdWrap, T>::call
},
{
RemapDispatcher<CubicFilter, BrdReflect101, T>::call,
RemapDispatcher<CubicFilter, BrdReplicate, T>::call,
RemapDispatcher<CubicFilter, BrdConstant, T>::call,
RemapDispatcher<CubicFilter, BrdReflect, T>::call,
RemapDispatcher<CubicFilter, BrdWrap, T>::call
}
};
callers[interpolation][borderMode](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), xoff, yoff, xmap, ymap,
static_cast< PtrStepSz<T> >(dst), borderValue, stream, cc);
}
template void remap_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void remap(const Ptr2D src, const PtrStepf mapx, const PtrStepf mapy, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = mapx.ptr(y)[x];
const float ycoo = mapy.ptr(y)[x];
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherStream
{
static void call(PtrStepSz<T> src, PtrStepSzf mapx, PtrStepSzf mapy, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block, 0, stream>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherNonStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy, PtrStepSz<T> dst, const float* borderValue, int)
{
(void)srcWhole;
(void)xoff;
(void)yoff;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_REMAP_TEX(type) \
texture< type , cudaTextureType2D> tex_remap_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_remap_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
int xoff, yoff; \
tex_remap_ ## type ## _reader (int xoff_, int yoff_) : xoff(xoff_), yoff(yoff_) {} \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_remap_ ## type , x + xoff, y + yoff); \
} \
}; \
template <template <typename> class Filter, template <typename> class B> struct RemapDispatcherNonStream<Filter, B, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy, \
PtrStepSz< type > dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , srcWhole); \
tex_remap_ ## type ##_reader texSrc(xoff, yoff); \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_remap_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <template <typename> class Filter> struct RemapDispatcherNonStream<Filter, BrdReplicate, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy, \
PtrStepSz< type > dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , srcWhole); \
tex_remap_ ## type ##_reader texSrc(xoff, yoff); \
if (srcWhole.cols == src.cols && srcWhole.rows == src.rows) \
{ \
Filter< tex_remap_ ## type ##_reader > filter_src(texSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
} \
else \
{ \
BrdReplicate<type> brd(src.rows, src.cols); \
BorderReader< tex_remap_ ## type ##_reader, BrdReplicate<type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, BrdReplicate<type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
} \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(short2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(float2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_REMAP_TEX
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcher
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy,
PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
RemapDispatcherNonStream<Filter, B, T>::call(src, srcWhole, xoff, yoff, mapx, mapy, dst, borderValue, cc);
else
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc);
}
};
template <typename T> void remap_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap,
PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap,
PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc);
static const caller_t callers[3][5] =
{
{
RemapDispatcher<PointFilter, BrdReflect101, T>::call,
RemapDispatcher<PointFilter, BrdReplicate, T>::call,
RemapDispatcher<PointFilter, BrdConstant, T>::call,
RemapDispatcher<PointFilter, BrdReflect, T>::call,
RemapDispatcher<PointFilter, BrdWrap, T>::call
},
{
RemapDispatcher<LinearFilter, BrdReflect101, T>::call,
RemapDispatcher<LinearFilter, BrdReplicate, T>::call,
RemapDispatcher<LinearFilter, BrdConstant, T>::call,
RemapDispatcher<LinearFilter, BrdReflect, T>::call,
RemapDispatcher<LinearFilter, BrdWrap, T>::call
},
{
RemapDispatcher<CubicFilter, BrdReflect101, T>::call,
RemapDispatcher<CubicFilter, BrdReplicate, T>::call,
RemapDispatcher<CubicFilter, BrdConstant, T>::call,
RemapDispatcher<CubicFilter, BrdReflect, T>::call,
RemapDispatcher<CubicFilter, BrdWrap, T>::call
}
};
callers[interpolation][borderMode](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), xoff, yoff, xmap, ymap,
static_cast< PtrStepSz<T> >(dst), borderValue, stream, cc);
}
template void remap_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzf xmap, PtrStepSzf ymap, PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

602
modules/gpu/src/cuda/resize.cu
View File

@@ -1,302 +1,302 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
#include <cfloat>
#include <opencv2/gpu/device/scan.hpp>
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void resize(const Ptr2D src, float fx, float fy, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x * fx;
const float ycoo = y * fy;
dst(y, x) = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <typename Ptr2D, typename T> __global__ void resize_area(const Ptr2D src, float fx, float fy, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
dst(y, x) = saturate_cast<T>(src(y, x));
}
}
template <template <typename> class Filter, typename T> struct ResizeDispatcherStream
{
static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filteredSrc(brdSrc, fx, fy);
resize<<<grid, block, 0, stream>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <typename T> struct ResizeDispatcherStream<AreaFilter, T>
{
static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdConstant<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
AreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
resize_area<<<grid, block, 0, stream>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T> struct ResizeDispatcherStream<IntegerAreaFilter, T>
{
static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdConstant<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
IntegerAreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
resize_area<<<grid, block, 0, stream>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <template <typename> class Filter, typename T> struct ResizeDispatcherNonStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst)
{
(void)srcWhole;
(void)xoff;
(void)yoff;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filteredSrc(brdSrc);
resize<<<grid, block>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
texture< type , cudaTextureType2D> tex_resize_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_resize_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
const int xoff; \
const int yoff; \
__host__ tex_resize_ ## type ## _reader(int xoff_, int yoff_) : xoff(xoff_), yoff(yoff_) {} \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_resize_ ## type, x + xoff, y + yoff); \
} \
}; \
template <template <typename> class Filter> struct ResizeDispatcherNonStream<Filter, type > \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz< type > dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type, srcWhole); \
tex_resize_ ## type ## _reader texSrc(xoff, yoff); \
if (srcWhole.cols == src.cols && srcWhole.rows == src.rows) \
{ \
Filter<tex_resize_ ## type ## _reader> filteredSrc(texSrc); \
resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
} \
else \
{ \
BrdReplicate< type > brd(src.rows, src.cols); \
BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > brdSrc(texSrc, brd); \
Filter< BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > > filteredSrc(brdSrc); \
resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
} \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(schar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
template <template <typename> class Filter, typename T> struct ResizeDispatcher
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
if (stream == 0)
ResizeDispatcherNonStream<Filter, T>::call(src, srcWhole, xoff, yoff, fx, fy, dst);
else
ResizeDispatcherStream<Filter, T>::call(src, fx, fy, dst, stream);
}
};
template <typename T> struct ResizeDispatcher<AreaFilter, T>
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
(void)srcWhole;
(void)xoff;
(void)yoff;
int iscale_x = (int)round(fx);
int iscale_y = (int)round(fy);
if( std::abs(fx - iscale_x) < FLT_MIN && std::abs(fy - iscale_y) < FLT_MIN)
ResizeDispatcherStream<IntegerAreaFilter, T>::call(src, fx, fy, dst, stream);
else
ResizeDispatcherStream<AreaFilter, T>::call(src, fx, fy, dst, stream);
}
};
template <typename T> void resize_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy,
PtrStepSzb dst, int interpolation, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream);
static const caller_t callers[4] =
{
ResizeDispatcher<PointFilter, T>::call,
ResizeDispatcher<LinearFilter, T>::call,
ResizeDispatcher<CubicFilter, T>::call,
ResizeDispatcher<AreaFilter, T>::call
};
// chenge to linear if area interpolation upscaling
if (interpolation == 3 && (fx <= 1.f || fy <= 1.f))
interpolation = 1;
callers[interpolation](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), xoff, yoff, fx, fy,
static_cast< PtrStepSz<T> >(dst), stream);
}
template void resize_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template<typename T> struct scan_traits{};
template<> struct scan_traits<uchar>
{
typedef float scan_line_type;
};
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
#include <cfloat>
#include <opencv2/gpu/device/scan.hpp>
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void resize(const Ptr2D src, float fx, float fy, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x * fx;
const float ycoo = y * fy;
dst(y, x) = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <typename Ptr2D, typename T> __global__ void resize_area(const Ptr2D src, float fx, float fy, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
dst(y, x) = saturate_cast<T>(src(y, x));
}
}
template <template <typename> class Filter, typename T> struct ResizeDispatcherStream
{
static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filteredSrc(brdSrc, fx, fy);
resize<<<grid, block, 0, stream>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <typename T> struct ResizeDispatcherStream<AreaFilter, T>
{
static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdConstant<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
AreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
resize_area<<<grid, block, 0, stream>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T> struct ResizeDispatcherStream<IntegerAreaFilter, T>
{
static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdConstant<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
IntegerAreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
resize_area<<<grid, block, 0, stream>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <template <typename> class Filter, typename T> struct ResizeDispatcherNonStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst)
{
(void)srcWhole;
(void)xoff;
(void)yoff;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filteredSrc(brdSrc);
resize<<<grid, block>>>(filteredSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
texture< type , cudaTextureType2D> tex_resize_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_resize_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
const int xoff; \
const int yoff; \
__host__ tex_resize_ ## type ## _reader(int xoff_, int yoff_) : xoff(xoff_), yoff(yoff_) {} \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_resize_ ## type, x + xoff, y + yoff); \
} \
}; \
template <template <typename> class Filter> struct ResizeDispatcherNonStream<Filter, type > \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz< type > dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type, srcWhole); \
tex_resize_ ## type ## _reader texSrc(xoff, yoff); \
if (srcWhole.cols == src.cols && srcWhole.rows == src.rows) \
{ \
Filter<tex_resize_ ## type ## _reader> filteredSrc(texSrc); \
resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
} \
else \
{ \
BrdReplicate< type > brd(src.rows, src.cols); \
BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > brdSrc(texSrc, brd); \
Filter< BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > > filteredSrc(brdSrc); \
resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
} \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(schar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
template <template <typename> class Filter, typename T> struct ResizeDispatcher
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
if (stream == 0)
ResizeDispatcherNonStream<Filter, T>::call(src, srcWhole, xoff, yoff, fx, fy, dst);
else
ResizeDispatcherStream<Filter, T>::call(src, fx, fy, dst, stream);
}
};
template <typename T> struct ResizeDispatcher<AreaFilter, T>
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
{
(void)srcWhole;
(void)xoff;
(void)yoff;
int iscale_x = (int)round(fx);
int iscale_y = (int)round(fy);
if( std::abs(fx - iscale_x) < FLT_MIN && std::abs(fy - iscale_y) < FLT_MIN)
ResizeDispatcherStream<IntegerAreaFilter, T>::call(src, fx, fy, dst, stream);
else
ResizeDispatcherStream<AreaFilter, T>::call(src, fx, fy, dst, stream);
}
};
template <typename T> void resize_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy,
PtrStepSzb dst, int interpolation, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream);
static const caller_t callers[4] =
{
ResizeDispatcher<PointFilter, T>::call,
ResizeDispatcher<LinearFilter, T>::call,
ResizeDispatcher<CubicFilter, T>::call,
ResizeDispatcher<AreaFilter, T>::call
};
// chenge to linear if area interpolation upscaling
if (interpolation == 3 && (fx <= 1.f || fy <= 1.f))
interpolation = 1;
callers[interpolation](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), xoff, yoff, fx, fy,
static_cast< PtrStepSz<T> >(dst), stream);
}
template void resize_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
template<typename T> struct scan_traits{};
template<> struct scan_traits<uchar>
{
typedef float scan_line_type;
};
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

348
modules/gpu/src/cuda/rgb_to_yv12.cu
View File

@@ -1,175 +1,175 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
namespace cv { namespace gpu { namespace device
{
namespace video_encoding
{
__device__ __forceinline__ void rgbtoy(const uchar b, const uchar g, const uchar r, uchar& y)
{
y = static_cast<uchar>(((int)(30 * r) + (int)(59 * g) + (int)(11 * b)) / 100);
}
__device__ __forceinline__ void rgbtoyuv(const uchar b, const uchar g, const uchar r, uchar& y, uchar& u, uchar& v)
{
rgbtoy(b, g, r, y);
u = static_cast<uchar>(((int)(-17 * r) - (int)(33 * g) + (int)(50 * b) + 12800) / 100);
v = static_cast<uchar>(((int)(50 * r) - (int)(42 * g) - (int)(8 * b) + 12800) / 100);
}
__global__ void Gray_to_YV12(const PtrStepSzb src, PtrStepb dst)
{
const int x = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const int y = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
if (x + 1 >= src.cols || y + 1 >= src.rows)
return;
// get pointers to the data
const size_t planeSize = src.rows * dst.step;
PtrStepb y_plane(dst.data, dst.step);
PtrStepb u_plane(y_plane.data + planeSize, dst.step / 2);
PtrStepb v_plane(u_plane.data + (planeSize / 4), dst.step / 2);
uchar pix;
uchar y_val, u_val, v_val;
pix = src(y, x);
rgbtoy(pix, pix, pix, y_val);
y_plane(y, x) = y_val;
pix = src(y, x + 1);
rgbtoy(pix, pix, pix, y_val);
y_plane(y, x + 1) = y_val;
pix = src(y + 1, x);
rgbtoy(pix, pix, pix, y_val);
y_plane(y + 1, x) = y_val;
pix = src(y + 1, x + 1);
rgbtoyuv(pix, pix, pix, y_val, u_val, v_val);
y_plane(y + 1, x + 1) = y_val;
u_plane(y / 2, x / 2) = u_val;
v_plane(y / 2, x / 2) = v_val;
}
template <typename T>
__global__ void BGR_to_YV12(const PtrStepSz<T> src, PtrStepb dst)
{
const int x = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const int y = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
if (x + 1 >= src.cols || y + 1 >= src.rows)
return;
// get pointers to the data
const size_t planeSize = src.rows * dst.step;
PtrStepb y_plane(dst.data, dst.step);
PtrStepb u_plane(y_plane.data + planeSize, dst.step / 2);
PtrStepb v_plane(u_plane.data + (planeSize / 4), dst.step / 2);
T pix;
uchar y_val, u_val, v_val;
pix = src(y, x);
rgbtoy(pix.z, pix.y, pix.x, y_val);
y_plane(y, x) = y_val;
pix = src(y, x + 1);
rgbtoy(pix.z, pix.y, pix.x, y_val);
y_plane(y, x + 1) = y_val;
pix = src(y + 1, x);
rgbtoy(pix.z, pix.y, pix.x, y_val);
y_plane(y + 1, x) = y_val;
pix = src(y + 1, x + 1);
rgbtoyuv(pix.z, pix.y, pix.x, y_val, u_val, v_val);
y_plane(y + 1, x + 1) = y_val;
u_plane(y / 2, x / 2) = u_val;
v_plane(y / 2, x / 2) = v_val;
}
void Gray_to_YV12_caller(const PtrStepSzb src, PtrStepb dst)
{
dim3 block(32, 8);
dim3 grid(divUp(src.cols, block.x * 2), divUp(src.rows, block.y * 2));
Gray_to_YV12<<<grid, block>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int cn>
void BGR_to_YV12_caller(const PtrStepSzb src, PtrStepb dst)
{
typedef typename TypeVec<uchar, cn>::vec_type src_t;
dim3 block(32, 8);
dim3 grid(divUp(src.cols, block.x * 2), divUp(src.rows, block.y * 2));
BGR_to_YV12<<<grid, block>>>(static_cast< PtrStepSz<src_t> >(src), dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
void YV12_gpu(const PtrStepSzb src, int cn, PtrStepSzb dst)
{
typedef void (*func_t)(const PtrStepSzb src, PtrStepb dst);
static const func_t funcs[] =
{
0, Gray_to_YV12_caller, 0, BGR_to_YV12_caller<3>, BGR_to_YV12_caller<4>
};
funcs[cn](src, dst);
}
}
}}}
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
namespace cv { namespace gpu { namespace device
{
namespace video_encoding
{
__device__ __forceinline__ void rgbtoy(const uchar b, const uchar g, const uchar r, uchar& y)
{
y = static_cast<uchar>(((int)(30 * r) + (int)(59 * g) + (int)(11 * b)) / 100);
}
__device__ __forceinline__ void rgbtoyuv(const uchar b, const uchar g, const uchar r, uchar& y, uchar& u, uchar& v)
{
rgbtoy(b, g, r, y);
u = static_cast<uchar>(((int)(-17 * r) - (int)(33 * g) + (int)(50 * b) + 12800) / 100);
v = static_cast<uchar>(((int)(50 * r) - (int)(42 * g) - (int)(8 * b) + 12800) / 100);
}
__global__ void Gray_to_YV12(const PtrStepSzb src, PtrStepb dst)
{
const int x = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const int y = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
if (x + 1 >= src.cols || y + 1 >= src.rows)
return;
// get pointers to the data
const size_t planeSize = src.rows * dst.step;
PtrStepb y_plane(dst.data, dst.step);
PtrStepb u_plane(y_plane.data + planeSize, dst.step / 2);
PtrStepb v_plane(u_plane.data + (planeSize / 4), dst.step / 2);
uchar pix;
uchar y_val, u_val, v_val;
pix = src(y, x);
rgbtoy(pix, pix, pix, y_val);
y_plane(y, x) = y_val;
pix = src(y, x + 1);
rgbtoy(pix, pix, pix, y_val);
y_plane(y, x + 1) = y_val;
pix = src(y + 1, x);
rgbtoy(pix, pix, pix, y_val);
y_plane(y + 1, x) = y_val;
pix = src(y + 1, x + 1);
rgbtoyuv(pix, pix, pix, y_val, u_val, v_val);
y_plane(y + 1, x + 1) = y_val;
u_plane(y / 2, x / 2) = u_val;
v_plane(y / 2, x / 2) = v_val;
}
template <typename T>
__global__ void BGR_to_YV12(const PtrStepSz<T> src, PtrStepb dst)
{
const int x = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const int y = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
if (x + 1 >= src.cols || y + 1 >= src.rows)
return;
// get pointers to the data
const size_t planeSize = src.rows * dst.step;
PtrStepb y_plane(dst.data, dst.step);
PtrStepb u_plane(y_plane.data + planeSize, dst.step / 2);
PtrStepb v_plane(u_plane.data + (planeSize / 4), dst.step / 2);
T pix;
uchar y_val, u_val, v_val;
pix = src(y, x);
rgbtoy(pix.z, pix.y, pix.x, y_val);
y_plane(y, x) = y_val;
pix = src(y, x + 1);
rgbtoy(pix.z, pix.y, pix.x, y_val);
y_plane(y, x + 1) = y_val;
pix = src(y + 1, x);
rgbtoy(pix.z, pix.y, pix.x, y_val);
y_plane(y + 1, x) = y_val;
pix = src(y + 1, x + 1);
rgbtoyuv(pix.z, pix.y, pix.x, y_val, u_val, v_val);
y_plane(y + 1, x + 1) = y_val;
u_plane(y / 2, x / 2) = u_val;
v_plane(y / 2, x / 2) = v_val;
}
void Gray_to_YV12_caller(const PtrStepSzb src, PtrStepb dst)
{
dim3 block(32, 8);
dim3 grid(divUp(src.cols, block.x * 2), divUp(src.rows, block.y * 2));
Gray_to_YV12<<<grid, block>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int cn>
void BGR_to_YV12_caller(const PtrStepSzb src, PtrStepb dst)
{
typedef typename TypeVec<uchar, cn>::vec_type src_t;
dim3 block(32, 8);
dim3 grid(divUp(src.cols, block.x * 2), divUp(src.rows, block.y * 2));
BGR_to_YV12<<<grid, block>>>(static_cast< PtrStepSz<src_t> >(src), dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
void YV12_gpu(const PtrStepSzb src, int cn, PtrStepSzb dst)
{
typedef void (*func_t)(const PtrStepSzb src, PtrStepb dst);
static const func_t funcs[] =
{
0, Gray_to_YV12_caller, 0, BGR_to_YV12_caller<3>, BGR_to_YV12_caller<4>
};
funcs[cn](src, dst);
}
}
}}}
#endif /* CUDA_DISABLER */

774
modules/gpu/src/cuda/row_filter.cu
View File

@@ -1,387 +1,387 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/static_check.hpp"
namespace cv { namespace gpu { namespace device
{
namespace row_filter
{
#define MAX_KERNEL_SIZE 32
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float* kernel, int ksize, cudaStream_t stream)
{
if (stream == 0)
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice) );
else
cudaSafeCall( cudaMemcpyToSymbolAsync(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice, stream) );
}
template <int KSIZE, typename T, typename D, typename B>
__global__ void linearRowFilter(const PtrStepSz<T> src, PtrStep<D> dst, const int anchor, const B brd)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
const int BLOCK_DIM_X = 32;
const int BLOCK_DIM_Y = 8;
const int PATCH_PER_BLOCK = 4;
const int HALO_SIZE = 1;
#else
const int BLOCK_DIM_X = 32;
const int BLOCK_DIM_Y = 4;
const int PATCH_PER_BLOCK = 4;
const int HALO_SIZE = 1;
#endif
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
__shared__ sum_t smem[BLOCK_DIM_Y][(PATCH_PER_BLOCK + 2 * HALO_SIZE) * BLOCK_DIM_X];
const int y = blockIdx.y * BLOCK_DIM_Y + threadIdx.y;
if (y >= src.rows)
return;
const T* src_row = src.ptr(y);
const int xStart = blockIdx.x * (PATCH_PER_BLOCK * BLOCK_DIM_X) + threadIdx.x;
if (blockIdx.x > 0)
{
//Load left halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart - (HALO_SIZE - j) * BLOCK_DIM_X]);
}
else
{
//Load left halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_low(xStart - (HALO_SIZE - j) * BLOCK_DIM_X, src_row));
}
if (blockIdx.x + 2 < gridDim.x)
{
//Load main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart + j * BLOCK_DIM_X]);
//Load right halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_X]);
}
else
{
//Load main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + j * BLOCK_DIM_X, src_row));
//Load right halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_X, src_row));
}
__syncthreads();
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
{
const int x = xStart + j * BLOCK_DIM_X;
if (x < src.cols)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for (int k = 0; k < KSIZE; ++k)
sum = sum + smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X - anchor + k] * c_kernel[k];
dst(y, x) = saturate_cast<D>(sum);
}
}
}
template <int KSIZE, typename T, typename D, template<typename> class B>
void linearRowFilter_caller(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream)
{
int BLOCK_DIM_X;
int BLOCK_DIM_Y;
int PATCH_PER_BLOCK;
if (cc >= 20)
{
BLOCK_DIM_X = 32;
BLOCK_DIM_Y = 8;
PATCH_PER_BLOCK = 4;
}
else
{
BLOCK_DIM_X = 32;
BLOCK_DIM_Y = 4;
PATCH_PER_BLOCK = 4;
}
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X * PATCH_PER_BLOCK), divUp(src.rows, BLOCK_DIM_Y));
B<T> brd(src.cols);
linearRowFilter<KSIZE, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, brd);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearRowFilter_gpu(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream);
static const caller_t callers[5][33] =
{
{
0,
linearRowFilter_caller< 1, T, D, BrdRowReflect101>,
linearRowFilter_caller< 2, T, D, BrdRowReflect101>,
linearRowFilter_caller< 3, T, D, BrdRowReflect101>,
linearRowFilter_caller< 4, T, D, BrdRowReflect101>,
linearRowFilter_caller< 5, T, D, BrdRowReflect101>,
linearRowFilter_caller< 6, T, D, BrdRowReflect101>,
linearRowFilter_caller< 7, T, D, BrdRowReflect101>,
linearRowFilter_caller< 8, T, D, BrdRowReflect101>,
linearRowFilter_caller< 9, T, D, BrdRowReflect101>,
linearRowFilter_caller<10, T, D, BrdRowReflect101>,
linearRowFilter_caller<11, T, D, BrdRowReflect101>,
linearRowFilter_caller<12, T, D, BrdRowReflect101>,
linearRowFilter_caller<13, T, D, BrdRowReflect101>,
linearRowFilter_caller<14, T, D, BrdRowReflect101>,
linearRowFilter_caller<15, T, D, BrdRowReflect101>,
linearRowFilter_caller<16, T, D, BrdRowReflect101>,
linearRowFilter_caller<17, T, D, BrdRowReflect101>,
linearRowFilter_caller<18, T, D, BrdRowReflect101>,
linearRowFilter_caller<19, T, D, BrdRowReflect101>,
linearRowFilter_caller<20, T, D, BrdRowReflect101>,
linearRowFilter_caller<21, T, D, BrdRowReflect101>,
linearRowFilter_caller<22, T, D, BrdRowReflect101>,
linearRowFilter_caller<23, T, D, BrdRowReflect101>,
linearRowFilter_caller<24, T, D, BrdRowReflect101>,
linearRowFilter_caller<25, T, D, BrdRowReflect101>,
linearRowFilter_caller<26, T, D, BrdRowReflect101>,
linearRowFilter_caller<27, T, D, BrdRowReflect101>,
linearRowFilter_caller<28, T, D, BrdRowReflect101>,
linearRowFilter_caller<29, T, D, BrdRowReflect101>,
linearRowFilter_caller<30, T, D, BrdRowReflect101>,
linearRowFilter_caller<31, T, D, BrdRowReflect101>,
linearRowFilter_caller<32, T, D, BrdRowReflect101>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowReplicate>,
linearRowFilter_caller< 2, T, D, BrdRowReplicate>,
linearRowFilter_caller< 3, T, D, BrdRowReplicate>,
linearRowFilter_caller< 4, T, D, BrdRowReplicate>,
linearRowFilter_caller< 5, T, D, BrdRowReplicate>,
linearRowFilter_caller< 6, T, D, BrdRowReplicate>,
linearRowFilter_caller< 7, T, D, BrdRowReplicate>,
linearRowFilter_caller< 8, T, D, BrdRowReplicate>,
linearRowFilter_caller< 9, T, D, BrdRowReplicate>,
linearRowFilter_caller<10, T, D, BrdRowReplicate>,
linearRowFilter_caller<11, T, D, BrdRowReplicate>,
linearRowFilter_caller<12, T, D, BrdRowReplicate>,
linearRowFilter_caller<13, T, D, BrdRowReplicate>,
linearRowFilter_caller<14, T, D, BrdRowReplicate>,
linearRowFilter_caller<15, T, D, BrdRowReplicate>,
linearRowFilter_caller<16, T, D, BrdRowReplicate>,
linearRowFilter_caller<17, T, D, BrdRowReplicate>,
linearRowFilter_caller<18, T, D, BrdRowReplicate>,
linearRowFilter_caller<19, T, D, BrdRowReplicate>,
linearRowFilter_caller<20, T, D, BrdRowReplicate>,
linearRowFilter_caller<21, T, D, BrdRowReplicate>,
linearRowFilter_caller<22, T, D, BrdRowReplicate>,
linearRowFilter_caller<23, T, D, BrdRowReplicate>,
linearRowFilter_caller<24, T, D, BrdRowReplicate>,
linearRowFilter_caller<25, T, D, BrdRowReplicate>,
linearRowFilter_caller<26, T, D, BrdRowReplicate>,
linearRowFilter_caller<27, T, D, BrdRowReplicate>,
linearRowFilter_caller<28, T, D, BrdRowReplicate>,
linearRowFilter_caller<29, T, D, BrdRowReplicate>,
linearRowFilter_caller<30, T, D, BrdRowReplicate>,
linearRowFilter_caller<31, T, D, BrdRowReplicate>,
linearRowFilter_caller<32, T, D, BrdRowReplicate>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowConstant>,
linearRowFilter_caller< 2, T, D, BrdRowConstant>,
linearRowFilter_caller< 3, T, D, BrdRowConstant>,
linearRowFilter_caller< 4, T, D, BrdRowConstant>,
linearRowFilter_caller< 5, T, D, BrdRowConstant>,
linearRowFilter_caller< 6, T, D, BrdRowConstant>,
linearRowFilter_caller< 7, T, D, BrdRowConstant>,
linearRowFilter_caller< 8, T, D, BrdRowConstant>,
linearRowFilter_caller< 9, T, D, BrdRowConstant>,
linearRowFilter_caller<10, T, D, BrdRowConstant>,
linearRowFilter_caller<11, T, D, BrdRowConstant>,
linearRowFilter_caller<12, T, D, BrdRowConstant>,
linearRowFilter_caller<13, T, D, BrdRowConstant>,
linearRowFilter_caller<14, T, D, BrdRowConstant>,
linearRowFilter_caller<15, T, D, BrdRowConstant>,
linearRowFilter_caller<16, T, D, BrdRowConstant>,
linearRowFilter_caller<17, T, D, BrdRowConstant>,
linearRowFilter_caller<18, T, D, BrdRowConstant>,
linearRowFilter_caller<19, T, D, BrdRowConstant>,
linearRowFilter_caller<20, T, D, BrdRowConstant>,
linearRowFilter_caller<21, T, D, BrdRowConstant>,
linearRowFilter_caller<22, T, D, BrdRowConstant>,
linearRowFilter_caller<23, T, D, BrdRowConstant>,
linearRowFilter_caller<24, T, D, BrdRowConstant>,
linearRowFilter_caller<25, T, D, BrdRowConstant>,
linearRowFilter_caller<26, T, D, BrdRowConstant>,
linearRowFilter_caller<27, T, D, BrdRowConstant>,
linearRowFilter_caller<28, T, D, BrdRowConstant>,
linearRowFilter_caller<29, T, D, BrdRowConstant>,
linearRowFilter_caller<30, T, D, BrdRowConstant>,
linearRowFilter_caller<31, T, D, BrdRowConstant>,
linearRowFilter_caller<32, T, D, BrdRowConstant>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowReflect>,
linearRowFilter_caller< 2, T, D, BrdRowReflect>,
linearRowFilter_caller< 3, T, D, BrdRowReflect>,
linearRowFilter_caller< 4, T, D, BrdRowReflect>,
linearRowFilter_caller< 5, T, D, BrdRowReflect>,
linearRowFilter_caller< 6, T, D, BrdRowReflect>,
linearRowFilter_caller< 7, T, D, BrdRowReflect>,
linearRowFilter_caller< 8, T, D, BrdRowReflect>,
linearRowFilter_caller< 9, T, D, BrdRowReflect>,
linearRowFilter_caller<10, T, D, BrdRowReflect>,
linearRowFilter_caller<11, T, D, BrdRowReflect>,
linearRowFilter_caller<12, T, D, BrdRowReflect>,
linearRowFilter_caller<13, T, D, BrdRowReflect>,
linearRowFilter_caller<14, T, D, BrdRowReflect>,
linearRowFilter_caller<15, T, D, BrdRowReflect>,
linearRowFilter_caller<16, T, D, BrdRowReflect>,
linearRowFilter_caller<17, T, D, BrdRowReflect>,
linearRowFilter_caller<18, T, D, BrdRowReflect>,
linearRowFilter_caller<19, T, D, BrdRowReflect>,
linearRowFilter_caller<20, T, D, BrdRowReflect>,
linearRowFilter_caller<21, T, D, BrdRowReflect>,
linearRowFilter_caller<22, T, D, BrdRowReflect>,
linearRowFilter_caller<23, T, D, BrdRowReflect>,
linearRowFilter_caller<24, T, D, BrdRowReflect>,
linearRowFilter_caller<25, T, D, BrdRowReflect>,
linearRowFilter_caller<26, T, D, BrdRowReflect>,
linearRowFilter_caller<27, T, D, BrdRowReflect>,
linearRowFilter_caller<28, T, D, BrdRowReflect>,
linearRowFilter_caller<29, T, D, BrdRowReflect>,
linearRowFilter_caller<30, T, D, BrdRowReflect>,
linearRowFilter_caller<31, T, D, BrdRowReflect>,
linearRowFilter_caller<32, T, D, BrdRowReflect>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowWrap>,
linearRowFilter_caller< 2, T, D, BrdRowWrap>,
linearRowFilter_caller< 3, T, D, BrdRowWrap>,
linearRowFilter_caller< 4, T, D, BrdRowWrap>,
linearRowFilter_caller< 5, T, D, BrdRowWrap>,
linearRowFilter_caller< 6, T, D, BrdRowWrap>,
linearRowFilter_caller< 7, T, D, BrdRowWrap>,
linearRowFilter_caller< 8, T, D, BrdRowWrap>,
linearRowFilter_caller< 9, T, D, BrdRowWrap>,
linearRowFilter_caller<10, T, D, BrdRowWrap>,
linearRowFilter_caller<11, T, D, BrdRowWrap>,
linearRowFilter_caller<12, T, D, BrdRowWrap>,
linearRowFilter_caller<13, T, D, BrdRowWrap>,
linearRowFilter_caller<14, T, D, BrdRowWrap>,
linearRowFilter_caller<15, T, D, BrdRowWrap>,
linearRowFilter_caller<16, T, D, BrdRowWrap>,
linearRowFilter_caller<17, T, D, BrdRowWrap>,
linearRowFilter_caller<18, T, D, BrdRowWrap>,
linearRowFilter_caller<19, T, D, BrdRowWrap>,
linearRowFilter_caller<20, T, D, BrdRowWrap>,
linearRowFilter_caller<21, T, D, BrdRowWrap>,
linearRowFilter_caller<22, T, D, BrdRowWrap>,
linearRowFilter_caller<23, T, D, BrdRowWrap>,
linearRowFilter_caller<24, T, D, BrdRowWrap>,
linearRowFilter_caller<25, T, D, BrdRowWrap>,
linearRowFilter_caller<26, T, D, BrdRowWrap>,
linearRowFilter_caller<27, T, D, BrdRowWrap>,
linearRowFilter_caller<28, T, D, BrdRowWrap>,
linearRowFilter_caller<29, T, D, BrdRowWrap>,
linearRowFilter_caller<30, T, D, BrdRowWrap>,
linearRowFilter_caller<31, T, D, BrdRowWrap>,
linearRowFilter_caller<32, T, D, BrdRowWrap>
}
};
loadKernel(kernel, ksize, stream);
callers[brd_type][ksize]((PtrStepSz<T>)src, (PtrStepSz<D>)dst, anchor, cc, stream);
}
template void linearRowFilter_gpu<uchar , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<uchar4, float4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<short3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<int , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<float , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
} // namespace row_filter
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 1993-2011, NVIDIA Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/static_check.hpp"
namespace cv { namespace gpu { namespace device
{
namespace row_filter
{
#define MAX_KERNEL_SIZE 32
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float* kernel, int ksize, cudaStream_t stream)
{
if (stream == 0)
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice) );
else
cudaSafeCall( cudaMemcpyToSymbolAsync(c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice, stream) );
}
template <int KSIZE, typename T, typename D, typename B>
__global__ void linearRowFilter(const PtrStepSz<T> src, PtrStep<D> dst, const int anchor, const B brd)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
const int BLOCK_DIM_X = 32;
const int BLOCK_DIM_Y = 8;
const int PATCH_PER_BLOCK = 4;
const int HALO_SIZE = 1;
#else
const int BLOCK_DIM_X = 32;
const int BLOCK_DIM_Y = 4;
const int PATCH_PER_BLOCK = 4;
const int HALO_SIZE = 1;
#endif
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
__shared__ sum_t smem[BLOCK_DIM_Y][(PATCH_PER_BLOCK + 2 * HALO_SIZE) * BLOCK_DIM_X];
const int y = blockIdx.y * BLOCK_DIM_Y + threadIdx.y;
if (y >= src.rows)
return;
const T* src_row = src.ptr(y);
const int xStart = blockIdx.x * (PATCH_PER_BLOCK * BLOCK_DIM_X) + threadIdx.x;
if (blockIdx.x > 0)
{
//Load left halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart - (HALO_SIZE - j) * BLOCK_DIM_X]);
}
else
{
//Load left halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_low(xStart - (HALO_SIZE - j) * BLOCK_DIM_X, src_row));
}
if (blockIdx.x + 2 < gridDim.x)
{
//Load main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart + j * BLOCK_DIM_X]);
//Load right halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_X]);
}
else
{
//Load main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + j * BLOCK_DIM_X, src_row));
//Load right halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_X, src_row));
}
__syncthreads();
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
{
const int x = xStart + j * BLOCK_DIM_X;
if (x < src.cols)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for (int k = 0; k < KSIZE; ++k)
sum = sum + smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X - anchor + k] * c_kernel[k];
dst(y, x) = saturate_cast<D>(sum);
}
}
}
template <int KSIZE, typename T, typename D, template<typename> class B>
void linearRowFilter_caller(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream)
{
int BLOCK_DIM_X;
int BLOCK_DIM_Y;
int PATCH_PER_BLOCK;
if (cc >= 20)
{
BLOCK_DIM_X = 32;
BLOCK_DIM_Y = 8;
PATCH_PER_BLOCK = 4;
}
else
{
BLOCK_DIM_X = 32;
BLOCK_DIM_Y = 4;
PATCH_PER_BLOCK = 4;
}
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X * PATCH_PER_BLOCK), divUp(src.rows, BLOCK_DIM_Y));
B<T> brd(src.cols);
linearRowFilter<KSIZE, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, brd);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearRowFilter_gpu(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream)
{
typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<D> dst, int anchor, int cc, cudaStream_t stream);
static const caller_t callers[5][33] =
{
{
0,
linearRowFilter_caller< 1, T, D, BrdRowReflect101>,
linearRowFilter_caller< 2, T, D, BrdRowReflect101>,
linearRowFilter_caller< 3, T, D, BrdRowReflect101>,
linearRowFilter_caller< 4, T, D, BrdRowReflect101>,
linearRowFilter_caller< 5, T, D, BrdRowReflect101>,
linearRowFilter_caller< 6, T, D, BrdRowReflect101>,
linearRowFilter_caller< 7, T, D, BrdRowReflect101>,
linearRowFilter_caller< 8, T, D, BrdRowReflect101>,
linearRowFilter_caller< 9, T, D, BrdRowReflect101>,
linearRowFilter_caller<10, T, D, BrdRowReflect101>,
linearRowFilter_caller<11, T, D, BrdRowReflect101>,
linearRowFilter_caller<12, T, D, BrdRowReflect101>,
linearRowFilter_caller<13, T, D, BrdRowReflect101>,
linearRowFilter_caller<14, T, D, BrdRowReflect101>,
linearRowFilter_caller<15, T, D, BrdRowReflect101>,
linearRowFilter_caller<16, T, D, BrdRowReflect101>,
linearRowFilter_caller<17, T, D, BrdRowReflect101>,
linearRowFilter_caller<18, T, D, BrdRowReflect101>,
linearRowFilter_caller<19, T, D, BrdRowReflect101>,
linearRowFilter_caller<20, T, D, BrdRowReflect101>,
linearRowFilter_caller<21, T, D, BrdRowReflect101>,
linearRowFilter_caller<22, T, D, BrdRowReflect101>,
linearRowFilter_caller<23, T, D, BrdRowReflect101>,
linearRowFilter_caller<24, T, D, BrdRowReflect101>,
linearRowFilter_caller<25, T, D, BrdRowReflect101>,
linearRowFilter_caller<26, T, D, BrdRowReflect101>,
linearRowFilter_caller<27, T, D, BrdRowReflect101>,
linearRowFilter_caller<28, T, D, BrdRowReflect101>,
linearRowFilter_caller<29, T, D, BrdRowReflect101>,
linearRowFilter_caller<30, T, D, BrdRowReflect101>,
linearRowFilter_caller<31, T, D, BrdRowReflect101>,
linearRowFilter_caller<32, T, D, BrdRowReflect101>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowReplicate>,
linearRowFilter_caller< 2, T, D, BrdRowReplicate>,
linearRowFilter_caller< 3, T, D, BrdRowReplicate>,
linearRowFilter_caller< 4, T, D, BrdRowReplicate>,
linearRowFilter_caller< 5, T, D, BrdRowReplicate>,
linearRowFilter_caller< 6, T, D, BrdRowReplicate>,
linearRowFilter_caller< 7, T, D, BrdRowReplicate>,
linearRowFilter_caller< 8, T, D, BrdRowReplicate>,
linearRowFilter_caller< 9, T, D, BrdRowReplicate>,
linearRowFilter_caller<10, T, D, BrdRowReplicate>,
linearRowFilter_caller<11, T, D, BrdRowReplicate>,
linearRowFilter_caller<12, T, D, BrdRowReplicate>,
linearRowFilter_caller<13, T, D, BrdRowReplicate>,
linearRowFilter_caller<14, T, D, BrdRowReplicate>,
linearRowFilter_caller<15, T, D, BrdRowReplicate>,
linearRowFilter_caller<16, T, D, BrdRowReplicate>,
linearRowFilter_caller<17, T, D, BrdRowReplicate>,
linearRowFilter_caller<18, T, D, BrdRowReplicate>,
linearRowFilter_caller<19, T, D, BrdRowReplicate>,
linearRowFilter_caller<20, T, D, BrdRowReplicate>,
linearRowFilter_caller<21, T, D, BrdRowReplicate>,
linearRowFilter_caller<22, T, D, BrdRowReplicate>,
linearRowFilter_caller<23, T, D, BrdRowReplicate>,
linearRowFilter_caller<24, T, D, BrdRowReplicate>,
linearRowFilter_caller<25, T, D, BrdRowReplicate>,
linearRowFilter_caller<26, T, D, BrdRowReplicate>,
linearRowFilter_caller<27, T, D, BrdRowReplicate>,
linearRowFilter_caller<28, T, D, BrdRowReplicate>,
linearRowFilter_caller<29, T, D, BrdRowReplicate>,
linearRowFilter_caller<30, T, D, BrdRowReplicate>,
linearRowFilter_caller<31, T, D, BrdRowReplicate>,
linearRowFilter_caller<32, T, D, BrdRowReplicate>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowConstant>,
linearRowFilter_caller< 2, T, D, BrdRowConstant>,
linearRowFilter_caller< 3, T, D, BrdRowConstant>,
linearRowFilter_caller< 4, T, D, BrdRowConstant>,
linearRowFilter_caller< 5, T, D, BrdRowConstant>,
linearRowFilter_caller< 6, T, D, BrdRowConstant>,
linearRowFilter_caller< 7, T, D, BrdRowConstant>,
linearRowFilter_caller< 8, T, D, BrdRowConstant>,
linearRowFilter_caller< 9, T, D, BrdRowConstant>,
linearRowFilter_caller<10, T, D, BrdRowConstant>,
linearRowFilter_caller<11, T, D, BrdRowConstant>,
linearRowFilter_caller<12, T, D, BrdRowConstant>,
linearRowFilter_caller<13, T, D, BrdRowConstant>,
linearRowFilter_caller<14, T, D, BrdRowConstant>,
linearRowFilter_caller<15, T, D, BrdRowConstant>,
linearRowFilter_caller<16, T, D, BrdRowConstant>,
linearRowFilter_caller<17, T, D, BrdRowConstant>,
linearRowFilter_caller<18, T, D, BrdRowConstant>,
linearRowFilter_caller<19, T, D, BrdRowConstant>,
linearRowFilter_caller<20, T, D, BrdRowConstant>,
linearRowFilter_caller<21, T, D, BrdRowConstant>,
linearRowFilter_caller<22, T, D, BrdRowConstant>,
linearRowFilter_caller<23, T, D, BrdRowConstant>,
linearRowFilter_caller<24, T, D, BrdRowConstant>,
linearRowFilter_caller<25, T, D, BrdRowConstant>,
linearRowFilter_caller<26, T, D, BrdRowConstant>,
linearRowFilter_caller<27, T, D, BrdRowConstant>,
linearRowFilter_caller<28, T, D, BrdRowConstant>,
linearRowFilter_caller<29, T, D, BrdRowConstant>,
linearRowFilter_caller<30, T, D, BrdRowConstant>,
linearRowFilter_caller<31, T, D, BrdRowConstant>,
linearRowFilter_caller<32, T, D, BrdRowConstant>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowReflect>,
linearRowFilter_caller< 2, T, D, BrdRowReflect>,
linearRowFilter_caller< 3, T, D, BrdRowReflect>,
linearRowFilter_caller< 4, T, D, BrdRowReflect>,
linearRowFilter_caller< 5, T, D, BrdRowReflect>,
linearRowFilter_caller< 6, T, D, BrdRowReflect>,
linearRowFilter_caller< 7, T, D, BrdRowReflect>,
linearRowFilter_caller< 8, T, D, BrdRowReflect>,
linearRowFilter_caller< 9, T, D, BrdRowReflect>,
linearRowFilter_caller<10, T, D, BrdRowReflect>,
linearRowFilter_caller<11, T, D, BrdRowReflect>,
linearRowFilter_caller<12, T, D, BrdRowReflect>,
linearRowFilter_caller<13, T, D, BrdRowReflect>,
linearRowFilter_caller<14, T, D, BrdRowReflect>,
linearRowFilter_caller<15, T, D, BrdRowReflect>,
linearRowFilter_caller<16, T, D, BrdRowReflect>,
linearRowFilter_caller<17, T, D, BrdRowReflect>,
linearRowFilter_caller<18, T, D, BrdRowReflect>,
linearRowFilter_caller<19, T, D, BrdRowReflect>,
linearRowFilter_caller<20, T, D, BrdRowReflect>,
linearRowFilter_caller<21, T, D, BrdRowReflect>,
linearRowFilter_caller<22, T, D, BrdRowReflect>,
linearRowFilter_caller<23, T, D, BrdRowReflect>,
linearRowFilter_caller<24, T, D, BrdRowReflect>,
linearRowFilter_caller<25, T, D, BrdRowReflect>,
linearRowFilter_caller<26, T, D, BrdRowReflect>,
linearRowFilter_caller<27, T, D, BrdRowReflect>,
linearRowFilter_caller<28, T, D, BrdRowReflect>,
linearRowFilter_caller<29, T, D, BrdRowReflect>,
linearRowFilter_caller<30, T, D, BrdRowReflect>,
linearRowFilter_caller<31, T, D, BrdRowReflect>,
linearRowFilter_caller<32, T, D, BrdRowReflect>
},
{
0,
linearRowFilter_caller< 1, T, D, BrdRowWrap>,
linearRowFilter_caller< 2, T, D, BrdRowWrap>,
linearRowFilter_caller< 3, T, D, BrdRowWrap>,
linearRowFilter_caller< 4, T, D, BrdRowWrap>,
linearRowFilter_caller< 5, T, D, BrdRowWrap>,
linearRowFilter_caller< 6, T, D, BrdRowWrap>,
linearRowFilter_caller< 7, T, D, BrdRowWrap>,
linearRowFilter_caller< 8, T, D, BrdRowWrap>,
linearRowFilter_caller< 9, T, D, BrdRowWrap>,
linearRowFilter_caller<10, T, D, BrdRowWrap>,
linearRowFilter_caller<11, T, D, BrdRowWrap>,
linearRowFilter_caller<12, T, D, BrdRowWrap>,
linearRowFilter_caller<13, T, D, BrdRowWrap>,
linearRowFilter_caller<14, T, D, BrdRowWrap>,
linearRowFilter_caller<15, T, D, BrdRowWrap>,
linearRowFilter_caller<16, T, D, BrdRowWrap>,
linearRowFilter_caller<17, T, D, BrdRowWrap>,
linearRowFilter_caller<18, T, D, BrdRowWrap>,
linearRowFilter_caller<19, T, D, BrdRowWrap>,
linearRowFilter_caller<20, T, D, BrdRowWrap>,
linearRowFilter_caller<21, T, D, BrdRowWrap>,
linearRowFilter_caller<22, T, D, BrdRowWrap>,
linearRowFilter_caller<23, T, D, BrdRowWrap>,
linearRowFilter_caller<24, T, D, BrdRowWrap>,
linearRowFilter_caller<25, T, D, BrdRowWrap>,
linearRowFilter_caller<26, T, D, BrdRowWrap>,
linearRowFilter_caller<27, T, D, BrdRowWrap>,
linearRowFilter_caller<28, T, D, BrdRowWrap>,
linearRowFilter_caller<29, T, D, BrdRowWrap>,
linearRowFilter_caller<30, T, D, BrdRowWrap>,
linearRowFilter_caller<31, T, D, BrdRowWrap>,
linearRowFilter_caller<32, T, D, BrdRowWrap>
}
};
loadKernel(kernel, ksize, stream);
callers[brd_type][ksize]((PtrStepSz<T>)src, (PtrStepSz<D>)dst, anchor, cc, stream);
}
template void linearRowFilter_gpu<uchar , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<uchar4, float4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<short3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<int , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearRowFilter_gpu<float , float >(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
} // namespace row_filter
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

190
modules/gpu/src/cuda/safe_call.hpp
View File

@@ -1,95 +1,95 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_CUDA_SAFE_CALL_HPP__
#define __OPENCV_CUDA_SAFE_CALL_HPP__
#include <cuda_runtime_api.h>
#include <cufft.h>
#include <cublas.h>
#include "NCV.hpp"
#if defined(__GNUC__)
#define nppSafeCall(expr) ___nppSafeCall(expr, __FILE__, __LINE__, __func__)
#define ncvSafeCall(expr) ___ncvSafeCall(expr, __FILE__, __LINE__, __func__)
#define cufftSafeCall(expr) ___cufftSafeCall(expr, __FILE__, __LINE__, __func__)
#define cublasSafeCall(expr) ___cublasSafeCall(expr, __FILE__, __LINE__, __func__)
#else /* defined(__CUDACC__) || defined(__MSVC__) */
#define nppSafeCall(expr) ___nppSafeCall(expr, __FILE__, __LINE__)
#define ncvSafeCall(expr) ___ncvSafeCall(expr, __FILE__, __LINE__)
#define cufftSafeCall(expr) ___cufftSafeCall(expr, __FILE__, __LINE__)
#define cublasSafeCall(expr) ___cublasSafeCall(expr, __FILE__, __LINE__)
#endif
namespace cv { namespace gpu
{
void nppError(int err, const char *file, const int line, const char *func = "");
void ncvError(int err, const char *file, const int line, const char *func = "");
void cufftError(int err, const char *file, const int line, const char *func = "");
void cublasError(int err, const char *file, const int line, const char *func = "");
}}
static inline void ___nppSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (err < 0)
cv::gpu::nppError(err, file, line, func);
}
static inline void ___ncvSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (NCV_SUCCESS != err)
cv::gpu::ncvError(err, file, line, func);
}
static inline void ___cufftSafeCall(cufftResult_t err, const char *file, const int line, const char *func = "")
{
if (CUFFT_SUCCESS != err)
cv::gpu::cufftError(err, file, line, func);
}
static inline void ___cublasSafeCall(cublasStatus_t err, const char *file, const int line, const char *func = "")
{
if (CUBLAS_STATUS_SUCCESS != err)
cv::gpu::cublasError(err, file, line, func);
}
#endif /* __OPENCV_CUDA_SAFE_CALL_HPP__ */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_CUDA_SAFE_CALL_HPP__
#define __OPENCV_CUDA_SAFE_CALL_HPP__
#include <cuda_runtime_api.h>
#include <cufft.h>
#include <cublas.h>
#include "NCV.hpp"
#if defined(__GNUC__)
#define nppSafeCall(expr) ___nppSafeCall(expr, __FILE__, __LINE__, __func__)
#define ncvSafeCall(expr) ___ncvSafeCall(expr, __FILE__, __LINE__, __func__)
#define cufftSafeCall(expr) ___cufftSafeCall(expr, __FILE__, __LINE__, __func__)
#define cublasSafeCall(expr) ___cublasSafeCall(expr, __FILE__, __LINE__, __func__)
#else /* defined(__CUDACC__) || defined(__MSVC__) */
#define nppSafeCall(expr) ___nppSafeCall(expr, __FILE__, __LINE__)
#define ncvSafeCall(expr) ___ncvSafeCall(expr, __FILE__, __LINE__)
#define cufftSafeCall(expr) ___cufftSafeCall(expr, __FILE__, __LINE__)
#define cublasSafeCall(expr) ___cublasSafeCall(expr, __FILE__, __LINE__)
#endif
namespace cv { namespace gpu
{
void nppError(int err, const char *file, const int line, const char *func = "");
void ncvError(int err, const char *file, const int line, const char *func = "");
void cufftError(int err, const char *file, const int line, const char *func = "");
void cublasError(int err, const char *file, const int line, const char *func = "");
}}
static inline void ___nppSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (err < 0)
cv::gpu::nppError(err, file, line, func);
}
static inline void ___ncvSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (NCV_SUCCESS != err)
cv::gpu::ncvError(err, file, line, func);
}
static inline void ___cufftSafeCall(cufftResult_t err, const char *file, const int line, const char *func = "")
{
if (CUFFT_SUCCESS != err)
cv::gpu::cufftError(err, file, line, func);
}
static inline void ___cublasSafeCall(cublasStatus_t err, const char *file, const int line, const char *func = "")
{
if (CUBLAS_STATUS_SUCCESS != err)
cv::gpu::cublasError(err, file, line, func);
}
#endif /* __OPENCV_CUDA_SAFE_CALL_HPP__ */

1020
modules/gpu/src/cuda/split_merge.cu
View File

File diff suppressed because it is too large Load Diff

1078
modules/gpu/src/cuda/stereobm.cu
View File

File diff suppressed because it is too large Load Diff

1074
modules/gpu/src/cuda/stereobp.cu
View File

File diff suppressed because it is too large Load Diff

1782
modules/gpu/src/cuda/stereocsbp.cu
View File

File diff suppressed because it is too large Load Diff

2010
modules/gpu/src/cuda/surf.cu
View File

File diff suppressed because it is too large Load Diff

6
modules/gpu/src/cuda/texture_binder.hpp
View File

@@ -56,14 +56,14 @@ namespace cv
template<class T, enum cudaTextureReadMode readMode>
TextureBinder(const PtrStepSz<T>& arr, const struct texture<T, 2, readMode>& tex) : texref(&tex)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture2D(0, tex, arr.data, desc, arr.cols, arr.rows, arr.step) );
}
template<class T, enum cudaTextureReadMode readMode>
TextureBinder(const PtrSz<T>& arr, const struct texture<T, 1, readMode> &tex) : texref(&tex)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture(0, tex, arr.data, desc, arr.size * arr.elemSize()) );
}

776
modules/gpu/src/cuda/warp.cu
View File

@@ -1,389 +1,389 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
__constant__ float c_warpMat[3 * 3];
struct AffineTransform
{
static __device__ __forceinline__ float2 calcCoord(int x, int y)
{
const float xcoo = c_warpMat[0] * x + c_warpMat[1] * y + c_warpMat[2];
const float ycoo = c_warpMat[3] * x + c_warpMat[4] * y + c_warpMat[5];
return make_float2(xcoo, ycoo);
}
};
struct PerspectiveTransform
{
static __device__ __forceinline__ float2 calcCoord(int x, int y)
{
const float coeff = 1.0f / (c_warpMat[6] * x + c_warpMat[7] * y + c_warpMat[8]);
const float xcoo = coeff * (c_warpMat[0] * x + c_warpMat[1] * y + c_warpMat[2]);
const float ycoo = coeff * (c_warpMat[3] * x + c_warpMat[4] * y + c_warpMat[5]);
return make_float2(xcoo, ycoo);
}
};
///////////////////////////////////////////////////////////////////
// Build Maps
template <class Transform> __global__ void buildWarpMaps(PtrStepSzf xmap, PtrStepf ymap)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < xmap.cols && y < xmap.rows)
{
const float2 coord = Transform::calcCoord(x, y);
xmap(y, x) = coord.x;
ymap(y, x) = coord.y;
}
}
template <class Transform> void buildWarpMaps_caller(PtrStepSzf xmap, PtrStepSzf ymap, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(xmap.cols, block.x), divUp(xmap.rows, block.y));
buildWarpMaps<Transform><<<grid, block, 0, stream>>>(xmap, ymap);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void buildWarpAffineMaps_gpu(float coeffs[2 * 3], PtrStepSzf xmap, PtrStepSzf ymap, cudaStream_t stream)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 2 * 3 * sizeof(float)) );
buildWarpMaps_caller<AffineTransform>(xmap, ymap, stream);
}
void buildWarpPerspectiveMaps_gpu(float coeffs[3 * 3], PtrStepSzf xmap, PtrStepSzf ymap, cudaStream_t stream)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 3 * 3 * sizeof(float)) );
buildWarpMaps_caller<PerspectiveTransform>(xmap, ymap, stream);
}
///////////////////////////////////////////////////////////////////
// Warp
template <class Transform, class Ptr2D, typename T> __global__ void warp(const Ptr2D src, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float2 coord = Transform::calcCoord(x, y);
dst.ptr(y)[x] = saturate_cast<T>(src(coord.y, coord.x));
}
}
template <class Transform, template <typename> class Filter, template <typename> class B, typename T> struct WarpDispatcherStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
warp<Transform><<<grid, block, 0, stream>>>(filter_src, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <class Transform, template <typename> class Filter, template <typename> class B, typename T> struct WarpDispatcherNonStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<T> dst, const float* borderValue, int)
{
(void)xoff;
(void)yoff;
(void)srcWhole;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
warp<Transform><<<grid, block>>>(filter_src, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_WARP_TEX(type) \
texture< type , cudaTextureType2D > tex_warp_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_warp_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
int xoff, yoff; \
tex_warp_ ## type ## _reader (int xoff_, int yoff_) : xoff(xoff_), yoff(yoff_) {} \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_warp_ ## type , x + xoff, y + yoff); \
} \
}; \
template <class Transform, template <typename> class Filter, template <typename> class B> struct WarpDispatcherNonStream<Transform, Filter, B, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSz< type > dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_warp_ ## type , srcWhole); \
tex_warp_ ## type ##_reader texSrc(xoff, yoff); \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_warp_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_warp_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
warp<Transform><<<grid, block>>>(filter_src, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <class Transform, template <typename> class Filter> struct WarpDispatcherNonStream<Transform, Filter, BrdReplicate, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSz< type > dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_warp_ ## type , srcWhole); \
tex_warp_ ## type ##_reader texSrc(xoff, yoff); \
if (srcWhole.cols == src.cols && srcWhole.rows == src.rows) \
{ \
Filter< tex_warp_ ## type ##_reader > filter_src(texSrc); \
warp<Transform><<<grid, block>>>(filter_src, dst); \
} \
else \
{ \
BrdReplicate<type> brd(src.rows, src.cols); \
BorderReader< tex_warp_ ## type ##_reader, BrdReplicate<type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_warp_ ## type ##_reader, BrdReplicate<type> > > filter_src(brdSrc); \
warp<Transform><<<grid, block>>>(filter_src, dst); \
} \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_WARP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(char4)
OPENCV_GPU_IMPLEMENT_WARP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_WARP_TEX(short)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(short2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(int)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(int4)
OPENCV_GPU_IMPLEMENT_WARP_TEX(float)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(float2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_WARP_TEX
template <class Transform, template <typename> class Filter, template <typename> class B, typename T> struct WarpDispatcher
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
WarpDispatcherNonStream<Transform, Filter, B, T>::call(src, srcWhole, xoff, yoff, dst, borderValue, cc);
else
WarpDispatcherStream<Transform, Filter, B, T>::call(src, dst, borderValue, stream, cc);
}
};
template <class Transform, typename T>
void warp_caller(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzb dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*func_t)(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc);
static const func_t funcs[3][5] =
{
{
WarpDispatcher<Transform, PointFilter, BrdReflect101, T>::call,
WarpDispatcher<Transform, PointFilter, BrdReplicate, T>::call,
WarpDispatcher<Transform, PointFilter, BrdConstant, T>::call,
WarpDispatcher<Transform, PointFilter, BrdReflect, T>::call,
WarpDispatcher<Transform, PointFilter, BrdWrap, T>::call
},
{
WarpDispatcher<Transform, LinearFilter, BrdReflect101, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdReplicate, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdConstant, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdReflect, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdWrap, T>::call
},
{
WarpDispatcher<Transform, CubicFilter, BrdReflect101, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdReplicate, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdConstant, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdReflect, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdWrap, T>::call
}
};
funcs[interpolation][borderMode](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), xoff, yoff,
static_cast< PtrStepSz<T> >(dst), borderValue, stream, cc);
}
template <typename T> void warpAffine_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 2 * 3 * sizeof(float)) );
warp_caller<AffineTransform, T>(src, srcWhole, xoff, yoff, dst, interpolation, borderMode, borderValue, stream, cc);
}
template void warpAffine_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template <typename T> void warpPerspective_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 3 * 3 * sizeof(float)) );
warp_caller<PerspectiveTransform, T>(src, srcWhole, xoff, yoff, dst, interpolation, borderMode, borderValue, stream, cc);
}
template void warpPerspective_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
__constant__ float c_warpMat[3 * 3];
struct AffineTransform
{
static __device__ __forceinline__ float2 calcCoord(int x, int y)
{
const float xcoo = c_warpMat[0] * x + c_warpMat[1] * y + c_warpMat[2];
const float ycoo = c_warpMat[3] * x + c_warpMat[4] * y + c_warpMat[5];
return make_float2(xcoo, ycoo);
}
};
struct PerspectiveTransform
{
static __device__ __forceinline__ float2 calcCoord(int x, int y)
{
const float coeff = 1.0f / (c_warpMat[6] * x + c_warpMat[7] * y + c_warpMat[8]);
const float xcoo = coeff * (c_warpMat[0] * x + c_warpMat[1] * y + c_warpMat[2]);
const float ycoo = coeff * (c_warpMat[3] * x + c_warpMat[4] * y + c_warpMat[5]);
return make_float2(xcoo, ycoo);
}
};
///////////////////////////////////////////////////////////////////
// Build Maps
template <class Transform> __global__ void buildWarpMaps(PtrStepSzf xmap, PtrStepf ymap)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < xmap.cols && y < xmap.rows)
{
const float2 coord = Transform::calcCoord(x, y);
xmap(y, x) = coord.x;
ymap(y, x) = coord.y;
}
}
template <class Transform> void buildWarpMaps_caller(PtrStepSzf xmap, PtrStepSzf ymap, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(xmap.cols, block.x), divUp(xmap.rows, block.y));
buildWarpMaps<Transform><<<grid, block, 0, stream>>>(xmap, ymap);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void buildWarpAffineMaps_gpu(float coeffs[2 * 3], PtrStepSzf xmap, PtrStepSzf ymap, cudaStream_t stream)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 2 * 3 * sizeof(float)) );
buildWarpMaps_caller<AffineTransform>(xmap, ymap, stream);
}
void buildWarpPerspectiveMaps_gpu(float coeffs[3 * 3], PtrStepSzf xmap, PtrStepSzf ymap, cudaStream_t stream)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 3 * 3 * sizeof(float)) );
buildWarpMaps_caller<PerspectiveTransform>(xmap, ymap, stream);
}
///////////////////////////////////////////////////////////////////
// Warp
template <class Transform, class Ptr2D, typename T> __global__ void warp(const Ptr2D src, PtrStepSz<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float2 coord = Transform::calcCoord(x, y);
dst.ptr(y)[x] = saturate_cast<T>(src(coord.y, coord.x));
}
}
template <class Transform, template <typename> class Filter, template <typename> class B, typename T> struct WarpDispatcherStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
warp<Transform><<<grid, block, 0, stream>>>(filter_src, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <class Transform, template <typename> class Filter, template <typename> class B, typename T> struct WarpDispatcherNonStream
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<T> dst, const float* borderValue, int)
{
(void)xoff;
(void)yoff;
(void)srcWhole;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
warp<Transform><<<grid, block>>>(filter_src, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_WARP_TEX(type) \
texture< type , cudaTextureType2D > tex_warp_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_warp_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
int xoff, yoff; \
tex_warp_ ## type ## _reader (int xoff_, int yoff_) : xoff(xoff_), yoff(yoff_) {} \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_warp_ ## type , x + xoff, y + yoff); \
} \
}; \
template <class Transform, template <typename> class Filter, template <typename> class B> struct WarpDispatcherNonStream<Transform, Filter, B, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSz< type > dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_warp_ ## type , srcWhole); \
tex_warp_ ## type ##_reader texSrc(xoff, yoff); \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_warp_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_warp_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
warp<Transform><<<grid, block>>>(filter_src, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <class Transform, template <typename> class Filter> struct WarpDispatcherNonStream<Transform, Filter, BrdReplicate, type> \
{ \
static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, PtrStepSz< type > dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_warp_ ## type , srcWhole); \
tex_warp_ ## type ##_reader texSrc(xoff, yoff); \
if (srcWhole.cols == src.cols && srcWhole.rows == src.rows) \
{ \
Filter< tex_warp_ ## type ##_reader > filter_src(texSrc); \
warp<Transform><<<grid, block>>>(filter_src, dst); \
} \
else \
{ \
BrdReplicate<type> brd(src.rows, src.cols); \
BorderReader< tex_warp_ ## type ##_reader, BrdReplicate<type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_warp_ ## type ##_reader, BrdReplicate<type> > > filter_src(brdSrc); \
warp<Transform><<<grid, block>>>(filter_src, dst); \
} \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_WARP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(char4)
OPENCV_GPU_IMPLEMENT_WARP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_WARP_TEX(short)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(short2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(int)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(int4)
OPENCV_GPU_IMPLEMENT_WARP_TEX(float)
//OPENCV_GPU_IMPLEMENT_WARP_TEX(float2)
OPENCV_GPU_IMPLEMENT_WARP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_WARP_TEX
template <class Transform, template <typename> class Filter, template <typename> class B, typename T> struct WarpDispatcher
{
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
WarpDispatcherNonStream<Transform, Filter, B, T>::call(src, srcWhole, xoff, yoff, dst, borderValue, cc);
else
WarpDispatcherStream<Transform, Filter, B, T>::call(src, dst, borderValue, stream, cc);
}
};
template <class Transform, typename T>
void warp_caller(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzb dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*func_t)(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, int cc);
static const func_t funcs[3][5] =
{
{
WarpDispatcher<Transform, PointFilter, BrdReflect101, T>::call,
WarpDispatcher<Transform, PointFilter, BrdReplicate, T>::call,
WarpDispatcher<Transform, PointFilter, BrdConstant, T>::call,
WarpDispatcher<Transform, PointFilter, BrdReflect, T>::call,
WarpDispatcher<Transform, PointFilter, BrdWrap, T>::call
},
{
WarpDispatcher<Transform, LinearFilter, BrdReflect101, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdReplicate, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdConstant, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdReflect, T>::call,
WarpDispatcher<Transform, LinearFilter, BrdWrap, T>::call
},
{
WarpDispatcher<Transform, CubicFilter, BrdReflect101, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdReplicate, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdConstant, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdReflect, T>::call,
WarpDispatcher<Transform, CubicFilter, BrdWrap, T>::call
}
};
funcs[interpolation][borderMode](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), xoff, yoff,
static_cast< PtrStepSz<T> >(dst), borderValue, stream, cc);
}
template <typename T> void warpAffine_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 2 * 3 * sizeof(float)) );
warp_caller<AffineTransform, T>(src, srcWhole, xoff, yoff, dst, interpolation, borderMode, borderValue, stream, cc);
}
template void warpAffine_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpAffine_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpAffine_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template <typename T> void warpPerspective_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
cudaSafeCall( cudaMemcpyToSymbol(c_warpMat, coeffs, 3 * 3 * sizeof(float)) );
warp_caller<PerspectiveTransform, T>(src, srcWhole, xoff, yoff, dst, interpolation, borderMode, borderValue, stream, cc);
}
template void warpPerspective_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void warpPerspective_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void warpPerspective_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[3 * 3], PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */

506
modules/gpu/src/cudastream.cpp
View File

@@ -1,253 +1,253 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
#if defined HAVE_CUDA
struct Stream::Impl
{
static cudaStream_t getStream(const Impl* impl) { return impl ? impl->stream : 0; }
cudaStream_t stream;
int ref_counter;
};
#include "opencv2/gpu/stream_accessor.hpp"
CV_EXPORTS cudaStream_t cv::gpu::StreamAccessor::getStream(const Stream& stream)
{
return Stream::Impl::getStream(stream.impl);
};
#endif /* !defined (HAVE_CUDA) */
#if !defined (HAVE_CUDA)
void cv::gpu::Stream::create() { throw_nogpu(); }
void cv::gpu::Stream::release() { throw_nogpu(); }
cv::gpu::Stream::Stream() : impl(0) { throw_nogpu(); }
cv::gpu::Stream::~Stream() { throw_nogpu(); }
cv::gpu::Stream::Stream(const Stream& /*stream*/) { throw_nogpu(); }
Stream& cv::gpu::Stream::operator=(const Stream& /*stream*/) { throw_nogpu(); return *this; }
bool cv::gpu::Stream::queryIfComplete() { throw_nogpu(); return true; }
void cv::gpu::Stream::waitForCompletion() { throw_nogpu(); }
void cv::gpu::Stream::enqueueDownload(const GpuMat& /*src*/, Mat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueDownload(const GpuMat& /*src*/, CudaMem& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueUpload(const CudaMem& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueUpload(const Mat& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueCopy(const GpuMat& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueMemSet(GpuMat& /*src*/, Scalar /*val*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueMemSet(GpuMat& /*src*/, Scalar /*val*/, const GpuMat& /*mask*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueConvert(const GpuMat& /*src*/, GpuMat& /*dst*/, int /*type*/, double /*a*/, double /*b*/) { throw_nogpu(); }
Stream& cv::gpu::Stream::Null() { throw_nogpu(); static Stream s; return s; }
cv::gpu::Stream::operator bool() const { throw_nogpu(); return false; }
#else /* !defined (HAVE_CUDA) */
namespace cv { namespace gpu
{
void copyWithMask(const GpuMat& src, GpuMat& dst, const GpuMat& mask, cudaStream_t stream);
void convertTo(const GpuMat& src, GpuMat& dst, double alpha, double beta, cudaStream_t stream);
void setTo(GpuMat& src, Scalar s, cudaStream_t stream);
void setTo(GpuMat& src, Scalar s, const GpuMat& mask, cudaStream_t stream);
}}
namespace
{
template<class S, class D> void devcopy(const S& src, D& dst, cudaStream_t s, cudaMemcpyKind k)
{
dst.create(src.size(), src.type());
size_t bwidth = src.cols * src.elemSize();
cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, k, s) );
};
}
void cv::gpu::Stream::create()
{
if (impl)
release();
cudaStream_t stream;
cudaSafeCall( cudaStreamCreate( &stream ) );
impl = (Stream::Impl*)fastMalloc(sizeof(Stream::Impl));
impl->stream = stream;
impl->ref_counter = 1;
}
void cv::gpu::Stream::release()
{
if( impl && CV_XADD(&impl->ref_counter, -1) == 1 )
{
cudaSafeCall( cudaStreamDestroy( impl->stream ) );
cv::fastFree( impl );
}
}
cv::gpu::Stream::Stream() : impl(0) { create(); }
cv::gpu::Stream::~Stream() { release(); }
cv::gpu::Stream::Stream(const Stream& stream) : impl(stream.impl)
{
if( impl )
CV_XADD(&impl->ref_counter, 1);
}
Stream& cv::gpu::Stream::operator=(const Stream& stream)
{
if( this != &stream )
{
if( stream.impl )
CV_XADD(&stream.impl->ref_counter, 1);
release();
impl = stream.impl;
}
return *this;
}
bool cv::gpu::Stream::queryIfComplete()
{
cudaError_t err = cudaStreamQuery( Impl::getStream(impl) );
if (err == cudaErrorNotReady || err == cudaSuccess)
return err == cudaSuccess;
cudaSafeCall(err);
return false;
}
void cv::gpu::Stream::waitForCompletion() { cudaSafeCall( cudaStreamSynchronize( Impl::getStream(impl) ) ); }
void cv::gpu::Stream::enqueueDownload(const GpuMat& src, Mat& dst)
{
// if not -> allocation will be done, but after that dst will not point to page locked memory
CV_Assert(src.cols == dst.cols && src.rows == dst.rows && src.type() == dst.type() );
devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToHost);
}
void cv::gpu::Stream::enqueueDownload(const GpuMat& src, CudaMem& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToHost); }
void cv::gpu::Stream::enqueueUpload(const CudaMem& src, GpuMat& dst){ devcopy(src, dst, Impl::getStream(impl), cudaMemcpyHostToDevice); }
void cv::gpu::Stream::enqueueUpload(const Mat& src, GpuMat& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyHostToDevice); }
void cv::gpu::Stream::enqueueCopy(const GpuMat& src, GpuMat& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToDevice); }
void cv::gpu::Stream::enqueueMemSet(GpuMat& src, Scalar s)
{
CV_Assert((src.depth() != CV_64F) ||
(TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
if (s[0] == 0.0 && s[1] == 0.0 && s[2] == 0.0 && s[3] == 0.0)
{
cudaSafeCall( cudaMemset2DAsync(src.data, src.step, 0, src.cols * src.elemSize(), src.rows, Impl::getStream(impl)) );
return;
}
if (src.depth() == CV_8U)
{
int cn = src.channels();
if (cn == 1 || (cn == 2 && s[0] == s[1]) || (cn == 3 && s[0] == s[1] && s[0] == s[2]) || (cn == 4 && s[0] == s[1] && s[0] == s[2] && s[0] == s[3]))
{
int val = saturate_cast<uchar>(s[0]);
cudaSafeCall( cudaMemset2DAsync(src.data, src.step, val, src.cols * src.elemSize(), src.rows, Impl::getStream(impl)) );
return;
}
}
setTo(src, s, Impl::getStream(impl));
}
void cv::gpu::Stream::enqueueMemSet(GpuMat& src, Scalar val, const GpuMat& mask)
{
CV_Assert((src.depth() != CV_64F) ||
(TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
CV_Assert(mask.type() == CV_8UC1);
setTo(src, val, mask, Impl::getStream(impl));
}
void cv::gpu::Stream::enqueueConvert(const GpuMat& src, GpuMat& dst, int rtype, double alpha, double beta)
{
CV_Assert((src.depth() != CV_64F && CV_MAT_DEPTH(rtype) != CV_64F) ||
(TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
bool noScale = fabs(alpha-1) < std::numeric_limits<double>::epsilon() && fabs(beta) < std::numeric_limits<double>::epsilon();
if( rtype < 0 )
rtype = src.type();
else
rtype = CV_MAKETYPE(CV_MAT_DEPTH(rtype), src.channels());
int sdepth = src.depth(), ddepth = CV_MAT_DEPTH(rtype);
if( sdepth == ddepth && noScale )
{
src.copyTo(dst);
return;
}
GpuMat temp;
const GpuMat* psrc = &src;
if( sdepth != ddepth && psrc == &dst )
psrc = &(temp = src);
dst.create( src.size(), rtype );
convertTo(src, dst, alpha, beta, Impl::getStream(impl));
}
cv::gpu::Stream::operator bool() const
{
return impl && impl->stream;
}
cv::gpu::Stream::Stream(Impl* impl_) : impl(impl_) {}
cv::gpu::Stream& cv::gpu::Stream::Null()
{
static Stream s((Impl*)0);
return s;
}
#endif /* !defined (HAVE_CUDA) */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
#if defined HAVE_CUDA
struct Stream::Impl
{
static cudaStream_t getStream(const Impl* impl) { return impl ? impl->stream : 0; }
cudaStream_t stream;
int ref_counter;
};
#include "opencv2/gpu/stream_accessor.hpp"
CV_EXPORTS cudaStream_t cv::gpu::StreamAccessor::getStream(const Stream& stream)
{
return Stream::Impl::getStream(stream.impl);
};
#endif /* !defined (HAVE_CUDA) */
#if !defined (HAVE_CUDA)
void cv::gpu::Stream::create() { throw_nogpu(); }
void cv::gpu::Stream::release() { throw_nogpu(); }
cv::gpu::Stream::Stream() : impl(0) { throw_nogpu(); }
cv::gpu::Stream::~Stream() { throw_nogpu(); }
cv::gpu::Stream::Stream(const Stream& /*stream*/) { throw_nogpu(); }
Stream& cv::gpu::Stream::operator=(const Stream& /*stream*/) { throw_nogpu(); return *this; }
bool cv::gpu::Stream::queryIfComplete() { throw_nogpu(); return true; }
void cv::gpu::Stream::waitForCompletion() { throw_nogpu(); }
void cv::gpu::Stream::enqueueDownload(const GpuMat& /*src*/, Mat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueDownload(const GpuMat& /*src*/, CudaMem& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueUpload(const CudaMem& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueUpload(const Mat& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueCopy(const GpuMat& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueMemSet(GpuMat& /*src*/, Scalar /*val*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueMemSet(GpuMat& /*src*/, Scalar /*val*/, const GpuMat& /*mask*/) { throw_nogpu(); }
void cv::gpu::Stream::enqueueConvert(const GpuMat& /*src*/, GpuMat& /*dst*/, int /*type*/, double /*a*/, double /*b*/) { throw_nogpu(); }
Stream& cv::gpu::Stream::Null() { throw_nogpu(); static Stream s; return s; }
cv::gpu::Stream::operator bool() const { throw_nogpu(); return false; }
#else /* !defined (HAVE_CUDA) */
namespace cv { namespace gpu
{
void copyWithMask(const GpuMat& src, GpuMat& dst, const GpuMat& mask, cudaStream_t stream);
void convertTo(const GpuMat& src, GpuMat& dst, double alpha, double beta, cudaStream_t stream);
void setTo(GpuMat& src, Scalar s, cudaStream_t stream);
void setTo(GpuMat& src, Scalar s, const GpuMat& mask, cudaStream_t stream);
}}
namespace
{
template<class S, class D> void devcopy(const S& src, D& dst, cudaStream_t s, cudaMemcpyKind k)
{
dst.create(src.size(), src.type());
size_t bwidth = src.cols * src.elemSize();
cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, k, s) );
};
}
void cv::gpu::Stream::create()
{
if (impl)
release();
cudaStream_t stream;
cudaSafeCall( cudaStreamCreate( &stream ) );
impl = (Stream::Impl*)fastMalloc(sizeof(Stream::Impl));
impl->stream = stream;
impl->ref_counter = 1;
}
void cv::gpu::Stream::release()
{
if( impl && CV_XADD(&impl->ref_counter, -1) == 1 )
{
cudaSafeCall( cudaStreamDestroy( impl->stream ) );
cv::fastFree( impl );
}
}
cv::gpu::Stream::Stream() : impl(0) { create(); }
cv::gpu::Stream::~Stream() { release(); }
cv::gpu::Stream::Stream(const Stream& stream) : impl(stream.impl)
{
if( impl )
CV_XADD(&impl->ref_counter, 1);
}
Stream& cv::gpu::Stream::operator=(const Stream& stream)
{
if( this != &stream )
{
if( stream.impl )
CV_XADD(&stream.impl->ref_counter, 1);
release();
impl = stream.impl;
}
return *this;
}
bool cv::gpu::Stream::queryIfComplete()
{
cudaError_t err = cudaStreamQuery( Impl::getStream(impl) );
if (err == cudaErrorNotReady || err == cudaSuccess)
return err == cudaSuccess;
cudaSafeCall(err);
return false;
}
void cv::gpu::Stream::waitForCompletion() { cudaSafeCall( cudaStreamSynchronize( Impl::getStream(impl) ) ); }
void cv::gpu::Stream::enqueueDownload(const GpuMat& src, Mat& dst)
{
// if not -> allocation will be done, but after that dst will not point to page locked memory
CV_Assert(src.cols == dst.cols && src.rows == dst.rows && src.type() == dst.type() );
devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToHost);
}
void cv::gpu::Stream::enqueueDownload(const GpuMat& src, CudaMem& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToHost); }
void cv::gpu::Stream::enqueueUpload(const CudaMem& src, GpuMat& dst){ devcopy(src, dst, Impl::getStream(impl), cudaMemcpyHostToDevice); }
void cv::gpu::Stream::enqueueUpload(const Mat& src, GpuMat& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyHostToDevice); }
void cv::gpu::Stream::enqueueCopy(const GpuMat& src, GpuMat& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToDevice); }
void cv::gpu::Stream::enqueueMemSet(GpuMat& src, Scalar s)
{
CV_Assert((src.depth() != CV_64F) ||
(TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
if (s[0] == 0.0 && s[1] == 0.0 && s[2] == 0.0 && s[3] == 0.0)
{
cudaSafeCall( cudaMemset2DAsync(src.data, src.step, 0, src.cols * src.elemSize(), src.rows, Impl::getStream(impl)) );
return;
}
if (src.depth() == CV_8U)
{
int cn = src.channels();
if (cn == 1 || (cn == 2 && s[0] == s[1]) || (cn == 3 && s[0] == s[1] && s[0] == s[2]) || (cn == 4 && s[0] == s[1] && s[0] == s[2] && s[0] == s[3]))
{
int val = saturate_cast<uchar>(s[0]);
cudaSafeCall( cudaMemset2DAsync(src.data, src.step, val, src.cols * src.elemSize(), src.rows, Impl::getStream(impl)) );
return;
}
}
setTo(src, s, Impl::getStream(impl));
}
void cv::gpu::Stream::enqueueMemSet(GpuMat& src, Scalar val, const GpuMat& mask)
{
CV_Assert((src.depth() != CV_64F) ||
(TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
CV_Assert(mask.type() == CV_8UC1);
setTo(src, val, mask, Impl::getStream(impl));
}
void cv::gpu::Stream::enqueueConvert(const GpuMat& src, GpuMat& dst, int rtype, double alpha, double beta)
{
CV_Assert((src.depth() != CV_64F && CV_MAT_DEPTH(rtype) != CV_64F) ||
(TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
bool noScale = fabs(alpha-1) < std::numeric_limits<double>::epsilon() && fabs(beta) < std::numeric_limits<double>::epsilon();
if( rtype < 0 )
rtype = src.type();
else
rtype = CV_MAKETYPE(CV_MAT_DEPTH(rtype), src.channels());
int sdepth = src.depth(), ddepth = CV_MAT_DEPTH(rtype);
if( sdepth == ddepth && noScale )
{
src.copyTo(dst);
return;
}
GpuMat temp;
const GpuMat* psrc = &src;
if( sdepth != ddepth && psrc == &dst )
psrc = &(temp = src);
dst.create( src.size(), rtype );
convertTo(src, dst, alpha, beta, Impl::getStream(impl));
}
cv::gpu::Stream::operator bool() const
{
return impl && impl->stream;
}
cv::gpu::Stream::Stream(Impl* impl_) : impl(impl_) {}
cv::gpu::Stream& cv::gpu::Stream::Null()
{
static Stream s((Impl*)0);
return s;
}
#endif /* !defined (HAVE_CUDA) */

126
modules/gpu/src/cuvid_video_source.cpp
View File

@@ -1,63 +1,63 @@
#include "cuvid_video_source.h"
#include "cu_safe_call.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
cv::gpu::detail::CuvidVideoSource::CuvidVideoSource(const std::string& fname)
{
CUVIDSOURCEPARAMS params;
std::memset(&params, 0, sizeof(CUVIDSOURCEPARAMS));
// Fill parameter struct
params.pUserData = this; // will be passed to data handlers
params.pfnVideoDataHandler = HandleVideoData; // our local video-handler callback
params.pfnAudioDataHandler = 0;
// now create the actual source
CUresult res = cuvidCreateVideoSource(&videoSource_, fname.c_str(), &params);
if (res == CUDA_ERROR_INVALID_SOURCE)
throw std::runtime_error("Unsupported video source");
cuSafeCall( res );
CUVIDEOFORMAT vidfmt;
cuSafeCall( cuvidGetSourceVideoFormat(videoSource_, &vidfmt, 0) );
format_.codec = static_cast<VideoReader_GPU::Codec>(vidfmt.codec);
format_.chromaFormat = static_cast<VideoReader_GPU::ChromaFormat>(vidfmt.chroma_format);
format_.width = vidfmt.coded_width;
format_.height = vidfmt.coded_height;
}
cv::gpu::VideoReader_GPU::FormatInfo cv::gpu::detail::CuvidVideoSource::format() const
{
return format_;
}
void cv::gpu::detail::CuvidVideoSource::start()
{
cuSafeCall( cuvidSetVideoSourceState(videoSource_, cudaVideoState_Started) );
}
void cv::gpu::detail::CuvidVideoSource::stop()
{
cuSafeCall( cuvidSetVideoSourceState(videoSource_, cudaVideoState_Stopped) );
}
bool cv::gpu::detail::CuvidVideoSource::isStarted() const
{
return (cuvidGetVideoSourceState(videoSource_) == cudaVideoState_Started);
}
bool cv::gpu::detail::CuvidVideoSource::hasError() const
{
return (cuvidGetVideoSourceState(videoSource_) == cudaVideoState_Error);
}
int CUDAAPI cv::gpu::detail::CuvidVideoSource::HandleVideoData(void* userData, CUVIDSOURCEDATAPACKET* packet)
{
CuvidVideoSource* thiz = static_cast<CuvidVideoSource*>(userData);
return thiz->parseVideoData(packet->payload, packet->payload_size, (packet->flags & CUVID_PKT_ENDOFSTREAM) != 0);
}
#endif // defined(HAVE_CUDA) && !defined(__APPLE__)
#include "cuvid_video_source.h"
#include "cu_safe_call.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
cv::gpu::detail::CuvidVideoSource::CuvidVideoSource(const std::string& fname)
{
CUVIDSOURCEPARAMS params;
std::memset(&params, 0, sizeof(CUVIDSOURCEPARAMS));
// Fill parameter struct
params.pUserData = this; // will be passed to data handlers
params.pfnVideoDataHandler = HandleVideoData; // our local video-handler callback
params.pfnAudioDataHandler = 0;
// now create the actual source
CUresult res = cuvidCreateVideoSource(&videoSource_, fname.c_str(), &params);
if (res == CUDA_ERROR_INVALID_SOURCE)
throw std::runtime_error("Unsupported video source");
cuSafeCall( res );
CUVIDEOFORMAT vidfmt;
cuSafeCall( cuvidGetSourceVideoFormat(videoSource_, &vidfmt, 0) );
format_.codec = static_cast<VideoReader_GPU::Codec>(vidfmt.codec);
format_.chromaFormat = static_cast<VideoReader_GPU::ChromaFormat>(vidfmt.chroma_format);
format_.width = vidfmt.coded_width;
format_.height = vidfmt.coded_height;
}
cv::gpu::VideoReader_GPU::FormatInfo cv::gpu::detail::CuvidVideoSource::format() const
{
return format_;
}
void cv::gpu::detail::CuvidVideoSource::start()
{
cuSafeCall( cuvidSetVideoSourceState(videoSource_, cudaVideoState_Started) );
}
void cv::gpu::detail::CuvidVideoSource::stop()
{
cuSafeCall( cuvidSetVideoSourceState(videoSource_, cudaVideoState_Stopped) );
}
bool cv::gpu::detail::CuvidVideoSource::isStarted() const
{
return (cuvidGetVideoSourceState(videoSource_) == cudaVideoState_Started);
}
bool cv::gpu::detail::CuvidVideoSource::hasError() const
{
return (cuvidGetVideoSourceState(videoSource_) == cudaVideoState_Error);
}
int CUDAAPI cv::gpu::detail::CuvidVideoSource::HandleVideoData(void* userData, CUVIDSOURCEDATAPACKET* packet)
{
CuvidVideoSource* thiz = static_cast<CuvidVideoSource*>(userData);
return thiz->parseVideoData(packet->payload, packet->payload_size, (packet->flags & CUVID_PKT_ENDOFSTREAM) != 0);
}
#endif // defined(HAVE_CUDA) && !defined(__APPLE__)

180
modules/gpu/src/cuvid_video_source.h
View File

@@ -1,90 +1,90 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __CUVUD_VIDEO_SOURCE_H__
#define __CUVUD_VIDEO_SOURCE_H__
#include "precomp.hpp"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
namespace cv { namespace gpu
{
namespace detail
{
class CuvidVideoSource : public VideoReader_GPU::VideoSource
{
public:
explicit CuvidVideoSource(const std::string& fname);
~CuvidVideoSource() { cuvidDestroyVideoSource(videoSource_); }
VideoReader_GPU::FormatInfo format() const;
void start();
void stop();
bool isStarted() const;
bool hasError() const;
private:
CuvidVideoSource(const CuvidVideoSource&);
CuvidVideoSource& operator =(const CuvidVideoSource&);
// Callback for handling packages of demuxed video data.
//
// Parameters:
// pUserData - Pointer to user data. We must pass a pointer to a
// VideoSourceData struct here, that contains a valid CUvideoparser
// and FrameQueue.
// pPacket - video-source data packet.
//
// NOTE: called from a different thread that doesn't not have a cuda context
//
static int CUDAAPI HandleVideoData(void* pUserData, CUVIDSOURCEDATAPACKET* pPacket);
CUvideosource videoSource_;
VideoReader_GPU::FormatInfo format_;
};
}
}}
#endif // defined(HAVE_CUDA) && !defined(__APPLE__)
#endif // __CUVUD_VIDEO_SOURCE_H__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __CUVUD_VIDEO_SOURCE_H__
#define __CUVUD_VIDEO_SOURCE_H__
#include "precomp.hpp"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
namespace cv { namespace gpu
{
namespace detail
{
class CuvidVideoSource : public VideoReader_GPU::VideoSource
{
public:
explicit CuvidVideoSource(const std::string& fname);
~CuvidVideoSource() { cuvidDestroyVideoSource(videoSource_); }
VideoReader_GPU::FormatInfo format() const;
void start();
void stop();
bool isStarted() const;
bool hasError() const;
private:
CuvidVideoSource(const CuvidVideoSource&);
CuvidVideoSource& operator =(const CuvidVideoSource&);
// Callback for handling packages of demuxed video data.
//
// Parameters:
// pUserData - Pointer to user data. We must pass a pointer to a
// VideoSourceData struct here, that contains a valid CUvideoparser
// and FrameQueue.
// pPacket - video-source data packet.
//
// NOTE: called from a different thread that doesn't not have a cuda context
//
static int CUDAAPI HandleVideoData(void* pUserData, CUVIDSOURCEDATAPACKET* pPacket);
CUvideosource videoSource_;
VideoReader_GPU::FormatInfo format_;
};
}
}}
#endif // defined(HAVE_CUDA) && !defined(__APPLE__)
#endif // __CUVUD_VIDEO_SOURCE_H__

26
modules/gpu/src/denoising.cpp
View File

@@ -124,7 +124,7 @@ void cv::gpu::nonLocalMeans(const GpuMat& src, GpuMat& dst, float h, int search_
int gpuBorderType;
CV_Assert(tryConvertToGpuBorderType(borderMode, gpuBorderType));
dst.create(src.size(), src.type());
func(src, dst, search_window/2, block_window/2, h, gpuBorderType, StreamAccessor::getStream(s));
}
@@ -143,7 +143,7 @@ namespace cv { namespace gpu { namespace device
template<typename T>
void nlm_fast_gpu(const PtrStepSzb& src, PtrStepSzb dst, PtrStepi buffer,
int search_window, int block_window, float h, cudaStream_t stream);
void fnlm_split_channels(const PtrStepSz<uchar3>& lab, PtrStepb l, PtrStep<uchar2> ab, cudaStream_t stream);
void fnlm_merge_channels(const PtrStepb& l, const PtrStep<uchar2>& ab, PtrStepSz<uchar3> lab, cudaStream_t stream);
}
@@ -152,30 +152,30 @@ namespace cv { namespace gpu { namespace device
void cv::gpu::FastNonLocalMeansDenoising::simpleMethod(const GpuMat& src, GpuMat& dst, float h, int search_window, int block_window, Stream& s)
{
CV_Assert(src.depth() == CV_8U && src.channels() < 4);
int border_size = search_window/2 + block_window/2;
int border_size = search_window/2 + block_window/2;
Size esize = src.size() + Size(border_size, border_size) * 2;
cv::gpu::ensureSizeIsEnough(esize, CV_8UC3, extended_src_buffer);
cv::gpu::ensureSizeIsEnough(esize, CV_8UC3, extended_src_buffer);
GpuMat extended_src(esize, src.type(), extended_src_buffer.ptr(), extended_src_buffer.step);
cv::gpu::copyMakeBorder(src, extended_src, border_size, border_size, border_size, border_size, cv::BORDER_DEFAULT, Scalar(), s);
GpuMat src_hdr = extended_src(Rect(Point2i(border_size, border_size), src.size()));
int bcols, brows;
device::imgproc::nln_fast_get_buffer_size(src_hdr, search_window, block_window, bcols, brows);
buffer.create(brows, bcols, CV_32S);
using namespace cv::gpu::device::imgproc;
typedef void (*nlm_fast_t)(const PtrStepSzb&, PtrStepSzb, PtrStepi, int, int, float, cudaStream_t);
static const nlm_fast_t funcs[] = { nlm_fast_gpu<uchar>, nlm_fast_gpu<uchar2>, nlm_fast_gpu<uchar3>, 0};
static const nlm_fast_t funcs[] = { nlm_fast_gpu<uchar>, nlm_fast_gpu<uchar2>, nlm_fast_gpu<uchar3>, 0};
dst.create(src.size(), src.type());
funcs[src.channels()-1](src_hdr, dst, buffer, search_window, block_window, h, StreamAccessor::getStream(s));
}
void cv::gpu::FastNonLocalMeansDenoising::labMethod( const GpuMat& src, GpuMat& dst, float h_luminance, float h_color, int search_window, int block_window, Stream& s)
{
{
#if (CUDA_VERSION < 5000)
(void)src;
(void)dst;
@@ -183,14 +183,14 @@ void cv::gpu::FastNonLocalMeansDenoising::labMethod( const GpuMat& src, GpuMat&
(void)h_color;
(void)search_window;
(void)block_window;
(void)s;
(void)s;
CV_Error( CV_GpuApiCallError, "Lab method required CUDA 5.0 and higher" );
#else
CV_Assert(src.type() == CV_8UC3);
lab.create(src.size(), src.type());
cv::gpu::cvtColor(src, lab, CV_BGR2Lab, 0, s);
@@ -201,7 +201,7 @@ void cv::gpu::FastNonLocalMeansDenoising::labMethod( const GpuMat& src, GpuMat&
l.create(src.size(), CV_8U);
ab.create(src.size(), CV_8UC2);
device::imgproc::fnlm_split_channels(lab, l, ab, StreamAccessor::getStream(s));
simpleMethod(l, l, h_luminance, search_window, block_window, s);
simpleMethod(ab, ab, h_color, search_window, block_window, s);

5438
modules/gpu/src/element_operations.cpp
View File

File diff suppressed because it is too large Load Diff

498
modules/gpu/src/error.cpp
View File

@@ -1,249 +1,249 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#ifdef HAVE_CUDA
namespace
{
#define error_entry(entry) { entry, #entry }
struct ErrorEntry
{
int code;
string str;
};
struct ErrorEntryComparer
{
int code;
ErrorEntryComparer(int code_) : code(code_) {};
bool operator()(const ErrorEntry& e) const { return e.code == code; }
};
string getErrorString(int code, const ErrorEntry* errors, size_t n)
{
size_t idx = find_if(errors, errors + n, ErrorEntryComparer(code)) - errors;
const string& msg = (idx != n) ? errors[idx].str : string("Unknown error code");
ostringstream ostr;
ostr << msg << " [Code = " << code << "]";
return ostr.str();
}
//////////////////////////////////////////////////////////////////////////
// NPP errors
const ErrorEntry npp_errors [] =
{
error_entry( NPP_NOT_SUPPORTED_MODE_ERROR ),
error_entry( NPP_ROUND_MODE_NOT_SUPPORTED_ERROR ),
error_entry( NPP_RESIZE_NO_OPERATION_ERROR ),
#if defined (_MSC_VER)
error_entry( NPP_NOT_SUFFICIENT_COMPUTE_CAPABILITY ),
#endif
error_entry( NPP_BAD_ARG_ERROR ),
error_entry( NPP_LUT_NUMBER_OF_LEVELS_ERROR ),
error_entry( NPP_TEXTURE_BIND_ERROR ),
error_entry( NPP_COEFF_ERROR ),
error_entry( NPP_RECT_ERROR ),
error_entry( NPP_QUAD_ERROR ),
error_entry( NPP_WRONG_INTERSECTION_ROI_ERROR ),
error_entry( NPP_NOT_EVEN_STEP_ERROR ),
error_entry( NPP_INTERPOLATION_ERROR ),
error_entry( NPP_RESIZE_FACTOR_ERROR ),
error_entry( NPP_HAAR_CLASSIFIER_PIXEL_MATCH_ERROR ),
error_entry( NPP_MEMFREE_ERR ),
error_entry( NPP_MEMSET_ERR ),
error_entry( NPP_MEMCPY_ERROR ),
error_entry( NPP_MEM_ALLOC_ERR ),
error_entry( NPP_HISTO_NUMBER_OF_LEVELS_ERROR ),
error_entry( NPP_MIRROR_FLIP_ERR ),
error_entry( NPP_INVALID_INPUT ),
error_entry( NPP_ALIGNMENT_ERROR ),
error_entry( NPP_STEP_ERROR ),
error_entry( NPP_SIZE_ERROR ),
error_entry( NPP_POINTER_ERROR ),
error_entry( NPP_NULL_POINTER_ERROR ),
error_entry( NPP_CUDA_KERNEL_EXECUTION_ERROR ),
error_entry( NPP_NOT_IMPLEMENTED_ERROR ),
error_entry( NPP_ERROR ),
error_entry( NPP_NO_ERROR ),
error_entry( NPP_SUCCESS ),
error_entry( NPP_WARNING ),
error_entry( NPP_WRONG_INTERSECTION_QUAD_WARNING ),
error_entry( NPP_MISALIGNED_DST_ROI_WARNING ),
error_entry( NPP_AFFINE_QUAD_INCORRECT_WARNING ),
error_entry( NPP_DOUBLE_SIZE_WARNING ),
error_entry( NPP_ODD_ROI_WARNING )
};
const size_t npp_error_num = sizeof(npp_errors) / sizeof(npp_errors[0]);
//////////////////////////////////////////////////////////////////////////
// NCV errors
const ErrorEntry ncv_errors [] =
{
error_entry( NCV_SUCCESS ),
error_entry( NCV_UNKNOWN_ERROR ),
error_entry( NCV_CUDA_ERROR ),
error_entry( NCV_NPP_ERROR ),
error_entry( NCV_FILE_ERROR ),
error_entry( NCV_NULL_PTR ),
error_entry( NCV_INCONSISTENT_INPUT ),
error_entry( NCV_TEXTURE_BIND_ERROR ),
error_entry( NCV_DIMENSIONS_INVALID ),
error_entry( NCV_INVALID_ROI ),
error_entry( NCV_INVALID_STEP ),
error_entry( NCV_INVALID_SCALE ),
error_entry( NCV_INVALID_SCALE ),
error_entry( NCV_ALLOCATOR_NOT_INITIALIZED ),
error_entry( NCV_ALLOCATOR_BAD_ALLOC ),
error_entry( NCV_ALLOCATOR_BAD_DEALLOC ),
error_entry( NCV_ALLOCATOR_INSUFFICIENT_CAPACITY ),
error_entry( NCV_ALLOCATOR_DEALLOC_ORDER ),
error_entry( NCV_ALLOCATOR_BAD_REUSE ),
error_entry( NCV_MEM_COPY_ERROR ),
error_entry( NCV_MEM_RESIDENCE_ERROR ),
error_entry( NCV_MEM_INSUFFICIENT_CAPACITY ),
error_entry( NCV_HAAR_INVALID_PIXEL_STEP ),
error_entry( NCV_HAAR_TOO_MANY_FEATURES_IN_CLASSIFIER ),
error_entry( NCV_HAAR_TOO_MANY_FEATURES_IN_CASCADE ),
error_entry( NCV_HAAR_TOO_LARGE_FEATURES ),
error_entry( NCV_HAAR_XML_LOADING_EXCEPTION ),
error_entry( NCV_NOIMPL_HAAR_TILTED_FEATURES ),
error_entry( NCV_WARNING_HAAR_DETECTIONS_VECTOR_OVERFLOW ),
error_entry( NPPST_SUCCESS ),
error_entry( NPPST_ERROR ),
error_entry( NPPST_CUDA_KERNEL_EXECUTION_ERROR ),
error_entry( NPPST_NULL_POINTER_ERROR ),
error_entry( NPPST_TEXTURE_BIND_ERROR ),
error_entry( NPPST_MEMCPY_ERROR ),
error_entry( NPPST_MEM_ALLOC_ERR ),
error_entry( NPPST_MEMFREE_ERR ),
error_entry( NPPST_INVALID_ROI ),
error_entry( NPPST_INVALID_STEP ),
error_entry( NPPST_INVALID_SCALE ),
error_entry( NPPST_MEM_INSUFFICIENT_BUFFER ),
error_entry( NPPST_MEM_RESIDENCE_ERROR ),
error_entry( NPPST_MEM_INTERNAL_ERROR )
};
const size_t ncv_error_num = sizeof(ncv_errors) / sizeof(ncv_errors[0]);
//////////////////////////////////////////////////////////////////////////
// CUFFT errors
const ErrorEntry cufft_errors[] =
{
error_entry( CUFFT_INVALID_PLAN ),
error_entry( CUFFT_ALLOC_FAILED ),
error_entry( CUFFT_INVALID_TYPE ),
error_entry( CUFFT_INVALID_VALUE ),
error_entry( CUFFT_INTERNAL_ERROR ),
error_entry( CUFFT_EXEC_FAILED ),
error_entry( CUFFT_SETUP_FAILED ),
error_entry( CUFFT_INVALID_SIZE ),
error_entry( CUFFT_UNALIGNED_DATA )
};
const int cufft_error_num = sizeof(cufft_errors) / sizeof(cufft_errors[0]);
//////////////////////////////////////////////////////////////////////////
// CUBLAS errors
const ErrorEntry cublas_errors[] =
{
error_entry( CUBLAS_STATUS_SUCCESS ),
error_entry( CUBLAS_STATUS_NOT_INITIALIZED ),
error_entry( CUBLAS_STATUS_ALLOC_FAILED ),
error_entry( CUBLAS_STATUS_INVALID_VALUE ),
error_entry( CUBLAS_STATUS_ARCH_MISMATCH ),
error_entry( CUBLAS_STATUS_MAPPING_ERROR ),
error_entry( CUBLAS_STATUS_EXECUTION_FAILED ),
error_entry( CUBLAS_STATUS_INTERNAL_ERROR )
};
const int cublas_error_num = sizeof(cublas_errors) / sizeof(cublas_errors[0]);
}
namespace cv
{
namespace gpu
{
void nppError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, npp_errors, npp_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
void ncvError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, ncv_errors, ncv_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
void cufftError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, cufft_errors, cufft_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
void cublasError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, cublas_errors, cublas_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
}
}
#endif
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#ifdef HAVE_CUDA
namespace
{
#define error_entry(entry) { entry, #entry }
struct ErrorEntry
{
int code;
string str;
};
struct ErrorEntryComparer
{
int code;
ErrorEntryComparer(int code_) : code(code_) {};
bool operator()(const ErrorEntry& e) const { return e.code == code; }
};
string getErrorString(int code, const ErrorEntry* errors, size_t n)
{
size_t idx = find_if(errors, errors + n, ErrorEntryComparer(code)) - errors;
const string& msg = (idx != n) ? errors[idx].str : string("Unknown error code");
ostringstream ostr;
ostr << msg << " [Code = " << code << "]";
return ostr.str();
}
//////////////////////////////////////////////////////////////////////////
// NPP errors
const ErrorEntry npp_errors [] =
{
error_entry( NPP_NOT_SUPPORTED_MODE_ERROR ),
error_entry( NPP_ROUND_MODE_NOT_SUPPORTED_ERROR ),
error_entry( NPP_RESIZE_NO_OPERATION_ERROR ),
#if defined (_MSC_VER)
error_entry( NPP_NOT_SUFFICIENT_COMPUTE_CAPABILITY ),
#endif
error_entry( NPP_BAD_ARG_ERROR ),
error_entry( NPP_LUT_NUMBER_OF_LEVELS_ERROR ),
error_entry( NPP_TEXTURE_BIND_ERROR ),
error_entry( NPP_COEFF_ERROR ),
error_entry( NPP_RECT_ERROR ),
error_entry( NPP_QUAD_ERROR ),
error_entry( NPP_WRONG_INTERSECTION_ROI_ERROR ),
error_entry( NPP_NOT_EVEN_STEP_ERROR ),
error_entry( NPP_INTERPOLATION_ERROR ),
error_entry( NPP_RESIZE_FACTOR_ERROR ),
error_entry( NPP_HAAR_CLASSIFIER_PIXEL_MATCH_ERROR ),
error_entry( NPP_MEMFREE_ERR ),
error_entry( NPP_MEMSET_ERR ),
error_entry( NPP_MEMCPY_ERROR ),
error_entry( NPP_MEM_ALLOC_ERR ),
error_entry( NPP_HISTO_NUMBER_OF_LEVELS_ERROR ),
error_entry( NPP_MIRROR_FLIP_ERR ),
error_entry( NPP_INVALID_INPUT ),
error_entry( NPP_ALIGNMENT_ERROR ),
error_entry( NPP_STEP_ERROR ),
error_entry( NPP_SIZE_ERROR ),
error_entry( NPP_POINTER_ERROR ),
error_entry( NPP_NULL_POINTER_ERROR ),
error_entry( NPP_CUDA_KERNEL_EXECUTION_ERROR ),
error_entry( NPP_NOT_IMPLEMENTED_ERROR ),
error_entry( NPP_ERROR ),
error_entry( NPP_NO_ERROR ),
error_entry( NPP_SUCCESS ),
error_entry( NPP_WARNING ),
error_entry( NPP_WRONG_INTERSECTION_QUAD_WARNING ),
error_entry( NPP_MISALIGNED_DST_ROI_WARNING ),
error_entry( NPP_AFFINE_QUAD_INCORRECT_WARNING ),
error_entry( NPP_DOUBLE_SIZE_WARNING ),
error_entry( NPP_ODD_ROI_WARNING )
};
const size_t npp_error_num = sizeof(npp_errors) / sizeof(npp_errors[0]);
//////////////////////////////////////////////////////////////////////////
// NCV errors
const ErrorEntry ncv_errors [] =
{
error_entry( NCV_SUCCESS ),
error_entry( NCV_UNKNOWN_ERROR ),
error_entry( NCV_CUDA_ERROR ),
error_entry( NCV_NPP_ERROR ),
error_entry( NCV_FILE_ERROR ),
error_entry( NCV_NULL_PTR ),
error_entry( NCV_INCONSISTENT_INPUT ),
error_entry( NCV_TEXTURE_BIND_ERROR ),
error_entry( NCV_DIMENSIONS_INVALID ),
error_entry( NCV_INVALID_ROI ),
error_entry( NCV_INVALID_STEP ),
error_entry( NCV_INVALID_SCALE ),
error_entry( NCV_INVALID_SCALE ),
error_entry( NCV_ALLOCATOR_NOT_INITIALIZED ),
error_entry( NCV_ALLOCATOR_BAD_ALLOC ),
error_entry( NCV_ALLOCATOR_BAD_DEALLOC ),
error_entry( NCV_ALLOCATOR_INSUFFICIENT_CAPACITY ),
error_entry( NCV_ALLOCATOR_DEALLOC_ORDER ),
error_entry( NCV_ALLOCATOR_BAD_REUSE ),
error_entry( NCV_MEM_COPY_ERROR ),
error_entry( NCV_MEM_RESIDENCE_ERROR ),
error_entry( NCV_MEM_INSUFFICIENT_CAPACITY ),
error_entry( NCV_HAAR_INVALID_PIXEL_STEP ),
error_entry( NCV_HAAR_TOO_MANY_FEATURES_IN_CLASSIFIER ),
error_entry( NCV_HAAR_TOO_MANY_FEATURES_IN_CASCADE ),
error_entry( NCV_HAAR_TOO_LARGE_FEATURES ),
error_entry( NCV_HAAR_XML_LOADING_EXCEPTION ),
error_entry( NCV_NOIMPL_HAAR_TILTED_FEATURES ),
error_entry( NCV_WARNING_HAAR_DETECTIONS_VECTOR_OVERFLOW ),
error_entry( NPPST_SUCCESS ),
error_entry( NPPST_ERROR ),
error_entry( NPPST_CUDA_KERNEL_EXECUTION_ERROR ),
error_entry( NPPST_NULL_POINTER_ERROR ),
error_entry( NPPST_TEXTURE_BIND_ERROR ),
error_entry( NPPST_MEMCPY_ERROR ),
error_entry( NPPST_MEM_ALLOC_ERR ),
error_entry( NPPST_MEMFREE_ERR ),
error_entry( NPPST_INVALID_ROI ),
error_entry( NPPST_INVALID_STEP ),
error_entry( NPPST_INVALID_SCALE ),
error_entry( NPPST_MEM_INSUFFICIENT_BUFFER ),
error_entry( NPPST_MEM_RESIDENCE_ERROR ),
error_entry( NPPST_MEM_INTERNAL_ERROR )
};
const size_t ncv_error_num = sizeof(ncv_errors) / sizeof(ncv_errors[0]);
//////////////////////////////////////////////////////////////////////////
// CUFFT errors
const ErrorEntry cufft_errors[] =
{
error_entry( CUFFT_INVALID_PLAN ),
error_entry( CUFFT_ALLOC_FAILED ),
error_entry( CUFFT_INVALID_TYPE ),
error_entry( CUFFT_INVALID_VALUE ),
error_entry( CUFFT_INTERNAL_ERROR ),
error_entry( CUFFT_EXEC_FAILED ),
error_entry( CUFFT_SETUP_FAILED ),
error_entry( CUFFT_INVALID_SIZE ),
error_entry( CUFFT_UNALIGNED_DATA )
};
const int cufft_error_num = sizeof(cufft_errors) / sizeof(cufft_errors[0]);
//////////////////////////////////////////////////////////////////////////
// CUBLAS errors
const ErrorEntry cublas_errors[] =
{
error_entry( CUBLAS_STATUS_SUCCESS ),
error_entry( CUBLAS_STATUS_NOT_INITIALIZED ),
error_entry( CUBLAS_STATUS_ALLOC_FAILED ),
error_entry( CUBLAS_STATUS_INVALID_VALUE ),
error_entry( CUBLAS_STATUS_ARCH_MISMATCH ),
error_entry( CUBLAS_STATUS_MAPPING_ERROR ),
error_entry( CUBLAS_STATUS_EXECUTION_FAILED ),
error_entry( CUBLAS_STATUS_INTERNAL_ERROR )
};
const int cublas_error_num = sizeof(cublas_errors) / sizeof(cublas_errors[0]);
}
namespace cv
{
namespace gpu
{
void nppError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, npp_errors, npp_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
void ncvError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, ncv_errors, ncv_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
void cufftError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, cufft_errors, cufft_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
void cublasError(int code, const char *file, const int line, const char *func)
{
string msg = getErrorString(code, cublas_errors, cublas_error_num);
cv::gpu::error(msg.c_str(), file, line, func);
}
}
}
#endif

354
modules/gpu/src/fast.cpp
View File

@@ -1,177 +1,177 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
cv::gpu::FAST_GPU::FAST_GPU(int, bool, double) { throw_nogpu(); }
void cv::gpu::FAST_GPU::operator ()(const GpuMat&, const GpuMat&, GpuMat&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::operator ()(const GpuMat&, const GpuMat&, std::vector<KeyPoint>&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::downloadKeypoints(const GpuMat&, std::vector<KeyPoint>&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::convertKeypoints(const Mat&, std::vector<KeyPoint>&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::release() { throw_nogpu(); }
int cv::gpu::FAST_GPU::calcKeyPointsLocation(const GpuMat&, const GpuMat&) { throw_nogpu(); return 0; }
int cv::gpu::FAST_GPU::getKeyPoints(GpuMat&) { throw_nogpu(); return 0; }
#else /* !defined (HAVE_CUDA) */
cv::gpu::FAST_GPU::FAST_GPU(int _threshold, bool _nonmaxSupression, double _keypointsRatio) :
nonmaxSupression(_nonmaxSupression), threshold(_threshold), keypointsRatio(_keypointsRatio), count_(0)
{
}
void cv::gpu::FAST_GPU::operator ()(const GpuMat& image, const GpuMat& mask, std::vector<KeyPoint>& keypoints)
{
if (image.empty())
return;
(*this)(image, mask, d_keypoints_);
downloadKeypoints(d_keypoints_, keypoints);
}
void cv::gpu::FAST_GPU::downloadKeypoints(const GpuMat& d_keypoints, std::vector<KeyPoint>& keypoints)
{
if (d_keypoints.empty())
return;
Mat h_keypoints(d_keypoints);
convertKeypoints(h_keypoints, keypoints);
}
void cv::gpu::FAST_GPU::convertKeypoints(const Mat& h_keypoints, std::vector<KeyPoint>& keypoints)
{
if (h_keypoints.empty())
return;
CV_Assert(h_keypoints.rows == ROWS_COUNT && h_keypoints.elemSize() == 4);
int npoints = h_keypoints.cols;
keypoints.resize(npoints);
const short2* loc_row = h_keypoints.ptr<short2>(LOCATION_ROW);
const float* response_row = h_keypoints.ptr<float>(RESPONSE_ROW);
for (int i = 0; i < npoints; ++i)
{
KeyPoint kp(loc_row[i].x, loc_row[i].y, static_cast<float>(FEATURE_SIZE), -1, response_row[i]);
keypoints[i] = kp;
}
}
void cv::gpu::FAST_GPU::operator ()(const GpuMat& img, const GpuMat& mask, GpuMat& keypoints)
{
calcKeyPointsLocation(img, mask);
keypoints.cols = getKeyPoints(keypoints);
}
namespace cv { namespace gpu { namespace device
{
namespace fast
{
int calcKeypoints_gpu(PtrStepSzb img, PtrStepSzb mask, short2* kpLoc, int maxKeypoints, PtrStepSzi score, int threshold);
int nonmaxSupression_gpu(const short2* kpLoc, int count, PtrStepSzi score, short2* loc, float* response);
}
}}}
int cv::gpu::FAST_GPU::calcKeyPointsLocation(const GpuMat& img, const GpuMat& mask)
{
using namespace cv::gpu::device::fast;
CV_Assert(img.type() == CV_8UC1);
CV_Assert(mask.empty() || (mask.type() == CV_8UC1 && mask.size() == img.size()));
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
int maxKeypoints = static_cast<int>(keypointsRatio * img.size().area());
ensureSizeIsEnough(1, maxKeypoints, CV_16SC2, kpLoc_);
if (nonmaxSupression)
{
ensureSizeIsEnough(img.size(), CV_32SC1, score_);
score_.setTo(Scalar::all(0));
}
count_ = calcKeypoints_gpu(img, mask, kpLoc_.ptr<short2>(), maxKeypoints, nonmaxSupression ? score_ : PtrStepSzi(), threshold);
count_ = std::min(count_, maxKeypoints);
return count_;
}
int cv::gpu::FAST_GPU::getKeyPoints(GpuMat& keypoints)
{
using namespace cv::gpu::device::fast;
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
if (count_ == 0)
return 0;
ensureSizeIsEnough(ROWS_COUNT, count_, CV_32FC1, keypoints);
if (nonmaxSupression)
return nonmaxSupression_gpu(kpLoc_.ptr<short2>(), count_, score_, keypoints.ptr<short2>(LOCATION_ROW), keypoints.ptr<float>(RESPONSE_ROW));
GpuMat locRow(1, count_, kpLoc_.type(), keypoints.ptr(0));
kpLoc_.colRange(0, count_).copyTo(locRow);
keypoints.row(1).setTo(Scalar::all(0));
return count_;
}
void cv::gpu::FAST_GPU::release()
{
kpLoc_.release();
score_.release();
d_keypoints_.release();
}
#endif /* !defined (HAVE_CUDA) */
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other GpuMaterials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or bpied warranties, including, but not limited to, the bpied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
cv::gpu::FAST_GPU::FAST_GPU(int, bool, double) { throw_nogpu(); }
void cv::gpu::FAST_GPU::operator ()(const GpuMat&, const GpuMat&, GpuMat&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::operator ()(const GpuMat&, const GpuMat&, std::vector<KeyPoint>&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::downloadKeypoints(const GpuMat&, std::vector<KeyPoint>&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::convertKeypoints(const Mat&, std::vector<KeyPoint>&) { throw_nogpu(); }
void cv::gpu::FAST_GPU::release() { throw_nogpu(); }
int cv::gpu::FAST_GPU::calcKeyPointsLocation(const GpuMat&, const GpuMat&) { throw_nogpu(); return 0; }
int cv::gpu::FAST_GPU::getKeyPoints(GpuMat&) { throw_nogpu(); return 0; }
#else /* !defined (HAVE_CUDA) */
cv::gpu::FAST_GPU::FAST_GPU(int _threshold, bool _nonmaxSupression, double _keypointsRatio) :
nonmaxSupression(_nonmaxSupression), threshold(_threshold), keypointsRatio(_keypointsRatio), count_(0)
{
}
void cv::gpu::FAST_GPU::operator ()(const GpuMat& image, const GpuMat& mask, std::vector<KeyPoint>& keypoints)
{
if (image.empty())
return;
(*this)(image, mask, d_keypoints_);
downloadKeypoints(d_keypoints_, keypoints);
}
void cv::gpu::FAST_GPU::downloadKeypoints(const GpuMat& d_keypoints, std::vector<KeyPoint>& keypoints)
{
if (d_keypoints.empty())
return;
Mat h_keypoints(d_keypoints);
convertKeypoints(h_keypoints, keypoints);
}
void cv::gpu::FAST_GPU::convertKeypoints(const Mat& h_keypoints, std::vector<KeyPoint>& keypoints)
{
if (h_keypoints.empty())
return;
CV_Assert(h_keypoints.rows == ROWS_COUNT && h_keypoints.elemSize() == 4);
int npoints = h_keypoints.cols;
keypoints.resize(npoints);
const short2* loc_row = h_keypoints.ptr<short2>(LOCATION_ROW);
const float* response_row = h_keypoints.ptr<float>(RESPONSE_ROW);
for (int i = 0; i < npoints; ++i)
{
KeyPoint kp(loc_row[i].x, loc_row[i].y, static_cast<float>(FEATURE_SIZE), -1, response_row[i]);
keypoints[i] = kp;
}
}
void cv::gpu::FAST_GPU::operator ()(const GpuMat& img, const GpuMat& mask, GpuMat& keypoints)
{
calcKeyPointsLocation(img, mask);
keypoints.cols = getKeyPoints(keypoints);
}
namespace cv { namespace gpu { namespace device
{
namespace fast
{
int calcKeypoints_gpu(PtrStepSzb img, PtrStepSzb mask, short2* kpLoc, int maxKeypoints, PtrStepSzi score, int threshold);
int nonmaxSupression_gpu(const short2* kpLoc, int count, PtrStepSzi score, short2* loc, float* response);
}
}}}
int cv::gpu::FAST_GPU::calcKeyPointsLocation(const GpuMat& img, const GpuMat& mask)
{
using namespace cv::gpu::device::fast;
CV_Assert(img.type() == CV_8UC1);
CV_Assert(mask.empty() || (mask.type() == CV_8UC1 && mask.size() == img.size()));
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
int maxKeypoints = static_cast<int>(keypointsRatio * img.size().area());
ensureSizeIsEnough(1, maxKeypoints, CV_16SC2, kpLoc_);
if (nonmaxSupression)
{
ensureSizeIsEnough(img.size(), CV_32SC1, score_);
score_.setTo(Scalar::all(0));
}
count_ = calcKeypoints_gpu(img, mask, kpLoc_.ptr<short2>(), maxKeypoints, nonmaxSupression ? score_ : PtrStepSzi(), threshold);
count_ = std::min(count_, maxKeypoints);
return count_;
}
int cv::gpu::FAST_GPU::getKeyPoints(GpuMat& keypoints)
{
using namespace cv::gpu::device::fast;
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
if (count_ == 0)
return 0;
ensureSizeIsEnough(ROWS_COUNT, count_, CV_32FC1, keypoints);
if (nonmaxSupression)
return nonmaxSupression_gpu(kpLoc_.ptr<short2>(), count_, score_, keypoints.ptr<short2>(LOCATION_ROW), keypoints.ptr<float>(RESPONSE_ROW));
GpuMat locRow(1, count_, kpLoc_.type(), keypoints.ptr(0));
kpLoc_.colRange(0, count_).copyTo(locRow);
keypoints.row(1).setTo(Scalar::all(0));
return count_;
}
void cv::gpu::FAST_GPU::release()
{
kpLoc_.release();
score_.release();
d_keypoints_.release();
}
#endif /* !defined (HAVE_CUDA) */

364
modules/gpu/src/ffmpeg_video_source.cpp
View File

@@ -1,182 +1,182 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "ffmpeg_video_source.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
#ifdef HAVE_FFMPEG
#include "cap_ffmpeg_impl.hpp"
#else
#include "cap_ffmpeg_api.hpp"
#endif
namespace
{
Create_InputMediaStream_FFMPEG_Plugin create_InputMediaStream_FFMPEG_p = 0;
Release_InputMediaStream_FFMPEG_Plugin release_InputMediaStream_FFMPEG_p = 0;
Read_InputMediaStream_FFMPEG_Plugin read_InputMediaStream_FFMPEG_p = 0;
bool init_MediaStream_FFMPEG()
{
static bool initialized = 0;
if (!initialized)
{
#if defined WIN32 || defined _WIN32
const char* module_name = "opencv_ffmpeg"
CVAUX_STR(CV_MAJOR_VERSION) CVAUX_STR(CV_MINOR_VERSION) CVAUX_STR(CV_SUBMINOR_VERSION)
#if (defined _MSC_VER && defined _M_X64) || (defined __GNUC__ && defined __x86_64__)
"_64"
#endif
".dll";
static HMODULE cvFFOpenCV = LoadLibrary(module_name);
if (cvFFOpenCV)
{
create_InputMediaStream_FFMPEG_p =
(Create_InputMediaStream_FFMPEG_Plugin)GetProcAddress(cvFFOpenCV, "create_InputMediaStream_FFMPEG");
release_InputMediaStream_FFMPEG_p =
(Release_InputMediaStream_FFMPEG_Plugin)GetProcAddress(cvFFOpenCV, "release_InputMediaStream_FFMPEG");
read_InputMediaStream_FFMPEG_p =
(Read_InputMediaStream_FFMPEG_Plugin)GetProcAddress(cvFFOpenCV, "read_InputMediaStream_FFMPEG");
initialized = create_InputMediaStream_FFMPEG_p != 0 && release_InputMediaStream_FFMPEG_p != 0 && read_InputMediaStream_FFMPEG_p != 0;
}
#elif defined HAVE_FFMPEG
create_InputMediaStream_FFMPEG_p = create_InputMediaStream_FFMPEG;
release_InputMediaStream_FFMPEG_p = release_InputMediaStream_FFMPEG;
read_InputMediaStream_FFMPEG_p = read_InputMediaStream_FFMPEG;
initialized = true;
#endif
}
return initialized;
}
}
cv::gpu::detail::FFmpegVideoSource::FFmpegVideoSource(const std::string& fname) :
stream_(0)
{
CV_Assert( init_MediaStream_FFMPEG() );
int codec;
int chroma_format;
int width;
int height;
stream_ = create_InputMediaStream_FFMPEG_p(fname.c_str(), &codec, &chroma_format, &width, &height);
if (!stream_)
CV_Error(CV_StsUnsupportedFormat, "Unsupported video source");
format_.codec = static_cast<VideoReader_GPU::Codec>(codec);
format_.chromaFormat = static_cast<VideoReader_GPU::ChromaFormat>(chroma_format);
format_.width = width;
format_.height = height;
}
cv::gpu::detail::FFmpegVideoSource::~FFmpegVideoSource()
{
release_InputMediaStream_FFMPEG_p(stream_);
}
cv::gpu::VideoReader_GPU::FormatInfo cv::gpu::detail::FFmpegVideoSource::format() const
{
return format_;
}
void cv::gpu::detail::FFmpegVideoSource::start()
{
stop_ = false;
hasError_ = false;
thread_.reset(new Thread(readLoop, this));
}
void cv::gpu::detail::FFmpegVideoSource::stop()
{
stop_ = true;
thread_->wait();
thread_.reset();
}
bool cv::gpu::detail::FFmpegVideoSource::isStarted() const
{
return !stop_;
}
bool cv::gpu::detail::FFmpegVideoSource::hasError() const
{
return hasError_;
}
void cv::gpu::detail::FFmpegVideoSource::readLoop(void* userData)
{
FFmpegVideoSource* thiz = static_cast<FFmpegVideoSource*>(userData);
for (;;)
{
unsigned char* data;
int size;
int endOfFile;
if (!read_InputMediaStream_FFMPEG_p(thiz->stream_, &data, &size, &endOfFile))
{
thiz->hasError_ = !endOfFile;
break;
}
if (!thiz->parseVideoData(data, size))
{
thiz->hasError_ = true;
break;
}
if (thiz->stop_)
break;
}
thiz->parseVideoData(0, 0, true);
}
#endif // HAVE_CUDA
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "ffmpeg_video_source.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
#ifdef HAVE_FFMPEG
#include "cap_ffmpeg_impl.hpp"
#else
#include "cap_ffmpeg_api.hpp"
#endif
namespace
{
Create_InputMediaStream_FFMPEG_Plugin create_InputMediaStream_FFMPEG_p = 0;
Release_InputMediaStream_FFMPEG_Plugin release_InputMediaStream_FFMPEG_p = 0;
Read_InputMediaStream_FFMPEG_Plugin read_InputMediaStream_FFMPEG_p = 0;
bool init_MediaStream_FFMPEG()
{
static bool initialized = 0;
if (!initialized)
{
#if defined WIN32 || defined _WIN32
const char* module_name = "opencv_ffmpeg"
CVAUX_STR(CV_MAJOR_VERSION) CVAUX_STR(CV_MINOR_VERSION) CVAUX_STR(CV_SUBMINOR_VERSION)
#if (defined _MSC_VER && defined _M_X64) || (defined __GNUC__ && defined __x86_64__)
"_64"
#endif
".dll";
static HMODULE cvFFOpenCV = LoadLibrary(module_name);
if (cvFFOpenCV)
{
create_InputMediaStream_FFMPEG_p =
(Create_InputMediaStream_FFMPEG_Plugin)GetProcAddress(cvFFOpenCV, "create_InputMediaStream_FFMPEG");
release_InputMediaStream_FFMPEG_p =
(Release_InputMediaStream_FFMPEG_Plugin)GetProcAddress(cvFFOpenCV, "release_InputMediaStream_FFMPEG");
read_InputMediaStream_FFMPEG_p =
(Read_InputMediaStream_FFMPEG_Plugin)GetProcAddress(cvFFOpenCV, "read_InputMediaStream_FFMPEG");
initialized = create_InputMediaStream_FFMPEG_p != 0 && release_InputMediaStream_FFMPEG_p != 0 && read_InputMediaStream_FFMPEG_p != 0;
}
#elif defined HAVE_FFMPEG
create_InputMediaStream_FFMPEG_p = create_InputMediaStream_FFMPEG;
release_InputMediaStream_FFMPEG_p = release_InputMediaStream_FFMPEG;
read_InputMediaStream_FFMPEG_p = read_InputMediaStream_FFMPEG;
initialized = true;
#endif
}
return initialized;
}
}
cv::gpu::detail::FFmpegVideoSource::FFmpegVideoSource(const std::string& fname) :
stream_(0)
{
CV_Assert( init_MediaStream_FFMPEG() );
int codec;
int chroma_format;
int width;
int height;
stream_ = create_InputMediaStream_FFMPEG_p(fname.c_str(), &codec, &chroma_format, &width, &height);
if (!stream_)
CV_Error(CV_StsUnsupportedFormat, "Unsupported video source");
format_.codec = static_cast<VideoReader_GPU::Codec>(codec);
format_.chromaFormat = static_cast<VideoReader_GPU::ChromaFormat>(chroma_format);
format_.width = width;
format_.height = height;
}
cv::gpu::detail::FFmpegVideoSource::~FFmpegVideoSource()
{
release_InputMediaStream_FFMPEG_p(stream_);
}
cv::gpu::VideoReader_GPU::FormatInfo cv::gpu::detail::FFmpegVideoSource::format() const
{
return format_;
}
void cv::gpu::detail::FFmpegVideoSource::start()
{
stop_ = false;
hasError_ = false;
thread_.reset(new Thread(readLoop, this));
}
void cv::gpu::detail::FFmpegVideoSource::stop()
{
stop_ = true;
thread_->wait();
thread_.reset();
}
bool cv::gpu::detail::FFmpegVideoSource::isStarted() const
{
return !stop_;
}
bool cv::gpu::detail::FFmpegVideoSource::hasError() const
{
return hasError_;
}
void cv::gpu::detail::FFmpegVideoSource::readLoop(void* userData)
{
FFmpegVideoSource* thiz = static_cast<FFmpegVideoSource*>(userData);
for (;;)
{
unsigned char* data;
int size;
int endOfFile;
if (!read_InputMediaStream_FFMPEG_p(thiz->stream_, &data, &size, &endOfFile))
{
thiz->hasError_ = !endOfFile;
break;
}
if (!thiz->parseVideoData(data, size))
{
thiz->hasError_ = true;
break;
}
if (thiz->stop_)
break;
}
thiz->parseVideoData(0, 0, true);
}
#endif // HAVE_CUDA

176
modules/gpu/src/ffmpeg_video_source.h
View File

@@ -1,88 +1,88 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __FFMPEG_VIDEO_SOURCE_H__
#define __FFMPEG_VIDEO_SOURCE_H__
#include "precomp.hpp"
#include "thread_wrappers.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
struct InputMediaStream_FFMPEG;
namespace cv { namespace gpu
{
namespace detail
{
class FFmpegVideoSource : public VideoReader_GPU::VideoSource
{
public:
FFmpegVideoSource(const std::string& fname);
~FFmpegVideoSource();
VideoReader_GPU::FormatInfo format() const;
void start();
void stop();
bool isStarted() const;
bool hasError() const;
private:
FFmpegVideoSource(const FFmpegVideoSource&);
FFmpegVideoSource& operator =(const FFmpegVideoSource&);
VideoReader_GPU::FormatInfo format_;
InputMediaStream_FFMPEG* stream_;
std::auto_ptr<Thread> thread_;
volatile bool stop_;
volatile bool hasError_;
static void readLoop(void* userData);
};
}
}}
#endif // HAVE_CUDA
#endif // __CUVUD_VIDEO_SOURCE_H__
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __FFMPEG_VIDEO_SOURCE_H__
#define __FFMPEG_VIDEO_SOURCE_H__
#include "precomp.hpp"
#include "thread_wrappers.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
struct InputMediaStream_FFMPEG;
namespace cv { namespace gpu
{
namespace detail
{
class FFmpegVideoSource : public VideoReader_GPU::VideoSource
{
public:
FFmpegVideoSource(const std::string& fname);
~FFmpegVideoSource();
VideoReader_GPU::FormatInfo format() const;
void start();
void stop();
bool isStarted() const;
bool hasError() const;
private:
FFmpegVideoSource(const FFmpegVideoSource&);
FFmpegVideoSource& operator =(const FFmpegVideoSource&);
VideoReader_GPU::FormatInfo format_;
InputMediaStream_FFMPEG* stream_;
std::auto_ptr<Thread> thread_;
volatile bool stop_;
volatile bool hasError_;
static void readLoop(void* userData);
};
}
}}
#endif // HAVE_CUDA
#endif // __CUVUD_VIDEO_SOURCE_H__

2718
modules/gpu/src/filtering.cpp
View File

File diff suppressed because it is too large Load Diff

234
modules/gpu/src/frame_queue.cpp
View File

@@ -1,117 +1,117 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "frame_queue.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
cv::gpu::detail::FrameQueue::FrameQueue() :
endOfDecode_(0),
framesInQueue_(0),
readPosition_(0)
{
std::memset(displayQueue_, 0, sizeof(displayQueue_));
std::memset((void*)isFrameInUse_, 0, sizeof(isFrameInUse_));
}
bool cv::gpu::detail::FrameQueue::waitUntilFrameAvailable(int pictureIndex)
{
while (isInUse(pictureIndex))
{
// Decoder is getting too far ahead from display
Thread::sleep(1);
if (isEndOfDecode())
return false;
}
return true;
}
void cv::gpu::detail::FrameQueue::enqueue(const CUVIDPARSERDISPINFO* picParams)
{
// Mark the frame as 'in-use' so we don't re-use it for decoding until it is no longer needed
// for display
isFrameInUse_[picParams->picture_index] = true;
// Wait until we have a free entry in the display queue (should never block if we have enough entries)
do
{
bool isFramePlaced = false;
{
CriticalSection::AutoLock autoLock(criticalSection_);
if (framesInQueue_ < MaximumSize)
{
int writePosition = (readPosition_ + framesInQueue_) % MaximumSize;
displayQueue_[writePosition] = *picParams;
framesInQueue_++;
isFramePlaced = true;
}
}
if (isFramePlaced) // Done
break;
// Wait a bit
Thread::sleep(1);
} while (!isEndOfDecode());
}
bool cv::gpu::detail::FrameQueue::dequeue(CUVIDPARSERDISPINFO& displayInfo)
{
CriticalSection::AutoLock autoLock(criticalSection_);
if (framesInQueue_ > 0)
{
int entry = readPosition_;
displayInfo = displayQueue_[entry];
readPosition_ = (entry + 1) % MaximumSize;
framesInQueue_--;
return true;
}
return false;
}
#endif // HAVE_CUDA
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "frame_queue.h"
#if defined(HAVE_CUDA) && !defined(__APPLE__)
cv::gpu::detail::FrameQueue::FrameQueue() :
endOfDecode_(0),
framesInQueue_(0),
readPosition_(0)
{
std::memset(displayQueue_, 0, sizeof(displayQueue_));
std::memset((void*)isFrameInUse_, 0, sizeof(isFrameInUse_));
}
bool cv::gpu::detail::FrameQueue::waitUntilFrameAvailable(int pictureIndex)
{
while (isInUse(pictureIndex))
{
// Decoder is getting too far ahead from display
Thread::sleep(1);
if (isEndOfDecode())
return false;
}
return true;
}
void cv::gpu::detail::FrameQueue::enqueue(const CUVIDPARSERDISPINFO* picParams)
{
// Mark the frame as 'in-use' so we don't re-use it for decoding until it is no longer needed
// for display
isFrameInUse_[picParams->picture_index] = true;
// Wait until we have a free entry in the display queue (should never block if we have enough entries)
do
{
bool isFramePlaced = false;
{
CriticalSection::AutoLock autoLock(criticalSection_);
if (framesInQueue_ < MaximumSize)
{
int writePosition = (readPosition_ + framesInQueue_) % MaximumSize;
displayQueue_[writePosition] = *picParams;
framesInQueue_++;
isFramePlaced = true;
}
}
if (isFramePlaced) // Done
break;
// Wait a bit
Thread::sleep(1);
} while (!isEndOfDecode());
}
bool cv::gpu::detail::FrameQueue::dequeue(CUVIDPARSERDISPINFO& displayInfo)
{
CriticalSection::AutoLock autoLock(criticalSection_);
if (framesInQueue_ > 0)
{
int entry = readPosition_;
displayInfo = displayQueue_[entry];
readPosition_ = (entry + 1) % MaximumSize;
framesInQueue_--;
return true;
}
return false;
}
#endif // HAVE_CUDA

Some files were not shown because too many files have changed in this diff Show More