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

Merge branch from CUDA team

Browse Source
This commit is contained in:
marina.kolpakova 2012-12-24 15:08:33 +04:00
commit 15e7712a26
201 changed files with 25627 additions and 15241 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
CMakeLists.txt
View File

@ -110,14 +110,15 @@ endif()
# Optional 3rd party components
# ===================================================
OCV_OPTION(WITH_1394 "Include IEEE1394 support" ON IF (UNIX AND NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_1394 "Include IEEE1394 support" ON IF (UNIX AND NOT ANDROID AND NOT IOS AND NOT CARMA) )
OCV_OPTION(WITH_AVFOUNDATION "Use AVFoundation for Video I/O" ON IF IOS)
OCV_OPTION(WITH_CARBON "Use Carbon for UI instead of Cocoa" OFF IF APPLE )
OCV_OPTION(WITH_CUBLAS "Include NVidia Cuda Basic Linear Algebra Subprograms (BLAS) library support" OFF IF (CMAKE_VERSION VERSION_GREATER "2.8" AND NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_CUDA "Include NVidia Cuda Runtime support" ON IF (CMAKE_VERSION VERSION_GREATER "2.8" AND NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_CUFFT "Include NVidia Cuda Fast Fourier Transform (FFT) library support" ON IF (CMAKE_VERSION VERSION_GREATER "2.8" AND NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_CUBLAS "Include NVidia Cuda Basic Linear Algebra Subprograms (BLAS) library support" OFF IF (CMAKE_VERSION VERSION_GREATER "2.8" AND NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_NVCUVID "Include NVidia Video Decoding library support" OFF IF (CMAKE_VERSION VERSION_GREATER "2.8" AND NOT ANDROID AND NOT IOS AND NOT APPLE) )
OCV_OPTION(WITH_EIGEN "Include Eigen2/Eigen3 support" ON)
OCV_OPTION(WITH_FFMPEG "Include FFMPEG support" ON IF (NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_FFMPEG "Include FFMPEG support" ON IF (NOT ANDROID AND NOT IOS))
OCV_OPTION(WITH_GSTREAMER "Include Gstreamer support" ON IF (UNIX AND NOT APPLE AND NOT ANDROID) )
OCV_OPTION(WITH_GTK "Include GTK support" ON IF (UNIX AND NOT APPLE AND NOT ANDROID) )
OCV_OPTION(WITH_IPP "Include Intel IPP support" OFF IF (MSVC OR X86 OR X86_64) )
@ -139,9 +140,9 @@ OCV_OPTION(WITH_VIDEOINPUT "Build HighGUI with DirectShow support" ON
OCV_OPTION(WITH_XIMEA "Include XIMEA cameras support" OFF IF (NOT ANDROID AND NOT APPLE) )
OCV_OPTION(WITH_XINE "Include Xine support (GPL)" OFF IF (UNIX AND NOT APPLE AND NOT ANDROID) )
OCV_OPTION(WITH_CLP "Include Clp support (EPL)" OFF)
OCV_OPTION(WITH_OPENCL "Include OpenCL Runtime support" OFF IF (NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_OPENCLAMDFFT "Include AMD OpenCL FFT library support" OFF IF (NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_OPENCLAMDBLAS "Include AMD OpenCL BLAS library support" OFF IF (NOT ANDROID AND NOT IOS) )
OCV_OPTION(WITH_OPENCL "Include OpenCL Runtime support" OFF IF (NOT ANDROID AND NOT IOS AND NOT CARMA) )
OCV_OPTION(WITH_OPENCLAMDFFT "Include AMD OpenCL FFT library support" OFF IF (NOT ANDROID AND NOT IOS AND NOT CARMA) )
OCV_OPTION(WITH_OPENCLAMDBLAS "Include AMD OpenCL BLAS library support" OFF IF (NOT ANDROID AND NOT IOS AND NOT CARMA) )
# OpenCV build components
@ -160,12 +161,12 @@ OCV_OPTION(BUILD_ANDROID_SERVICE "Build OpenCV Manager for Google Play" OFF I
OCV_OPTION(BUILD_ANDROID_PACKAGE "Build platform-specific package for Google Play" OFF IF ANDROID )
# 3rd party libs
OCV_OPTION(BUILD_ZLIB "Build zlib from source" WIN32 OR APPLE )
OCV_OPTION(BUILD_TIFF "Build libtiff from source" WIN32 OR ANDROID OR APPLE )
OCV_OPTION(BUILD_JASPER "Build libjasper from source" WIN32 OR ANDROID OR APPLE )
OCV_OPTION(BUILD_JPEG "Build libjpeg from source" WIN32 OR ANDROID OR APPLE )
OCV_OPTION(BUILD_PNG "Build libpng from source" WIN32 OR ANDROID OR APPLE )
OCV_OPTION(BUILD_OPENEXR "Build openexr from source" WIN32 OR ANDROID OR APPLE )
OCV_OPTION(BUILD_ZLIB "Build zlib from source" WIN32 OR APPLE OR CARMA )
OCV_OPTION(BUILD_TIFF "Build libtiff from source" WIN32 OR ANDROID OR APPLE OR CARMA )
OCV_OPTION(BUILD_JASPER "Build libjasper from source" WIN32 OR ANDROID OR APPLE OR CARMA )
OCV_OPTION(BUILD_JPEG "Build libjpeg from source" WIN32 OR ANDROID OR APPLE OR CARMA )
OCV_OPTION(BUILD_PNG "Build libpng from source" WIN32 OR ANDROID OR APPLE OR CARMA )
OCV_OPTION(BUILD_OPENEXR "Build openexr from source" WIN32 OR ANDROID OR APPLE OR CARMA )
# OpenCV installation options

53
cmake/OpenCVDetectCUDA.cmake
View File

@ -3,17 +3,17 @@ if(${CMAKE_VERSION} VERSION_LESS "2.8.3")
return()
endif()
if (WIN32 AND NOT MSVC)
if(WIN32 AND NOT MSVC)
message(STATUS "CUDA compilation is disabled (due to only Visual Studio compiler suppoted on your platform).")
return()
endif()
if (CMAKE_COMPILER_IS_GNUCXX AND NOT APPLE AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
if(CMAKE_COMPILER_IS_GNUCXX AND NOT APPLE AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
message(STATUS "CUDA compilation is disabled (due to Clang unsuppoted on your platform).")
return()
endif()
find_package(CUDA 4.1)
find_package(CUDA 4.2)
if(CUDA_FOUND)
set(HAVE_CUDA 1)
@ -26,15 +26,20 @@ if(CUDA_FOUND)
set(HAVE_CUBLAS 1)
endif()
message(STATUS "CUDA detected: " ${CUDA_VERSION})
if(${CUDA_VERSION_STRING} VERSION_GREATER "4.1")
set(CUDA_ARCH_BIN "1.1 1.2 1.3 2.0 2.1(2.0) 3.0" CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
else()
set(CUDA_ARCH_BIN "1.1 1.2 1.3 2.0 2.1(2.0)" CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
if(WITH_NVCUVID)
find_cuda_helper_libs(nvcuvid)
set(HAVE_NVCUVID 1)
endif()
set(CUDA_ARCH_PTX "2.0" CACHE STRING "Specify 'virtual' PTX architectures to build PTX intermediate code for")
message(STATUS "CUDA detected: " ${CUDA_VERSION})
if (CARMA)
set(CUDA_ARCH_BIN "3.0" CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
set(CUDA_ARCH_PTX "3.0" CACHE STRING "Specify 'virtual' PTX architectures to build PTX intermediate code for")
else()
set(CUDA_ARCH_BIN "1.1 1.2 1.3 2.0 2.1(2.0) 3.0" CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
set(CUDA_ARCH_PTX "2.0 3.0" CACHE STRING "Specify 'virtual' PTX architectures to build PTX intermediate code for")
endif()
string(REGEX REPLACE "\\." "" ARCH_BIN_NO_POINTS "${CUDA_ARCH_BIN}")
string(REGEX REPLACE "\\." "" ARCH_PTX_NO_POINTS "${CUDA_ARCH_PTX}")
@ -72,11 +77,20 @@ if(CUDA_FOUND)
# Tell NVCC to add PTX intermediate code for the specified architectures
string(REGEX MATCHALL "[0-9]+" ARCH_LIST "${ARCH_PTX_NO_POINTS}")
foreach(ARCH IN LISTS ARCH_LIST)
set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${ARCH},code=compute_${ARCH})
set(OPENCV_CUDA_ARCH_PTX "${OPENCV_CUDA_ARCH_PTX} ${ARCH}")
set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${ARCH}")
endforeach()
foreach(ARCH IN LISTS ARCH_LIST)
set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${ARCH},code=compute_${ARCH})
set(OPENCV_CUDA_ARCH_PTX "${OPENCV_CUDA_ARCH_PTX} ${ARCH}")
set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${ARCH}")
endforeach()
if(CARMA)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --target-cpu-architecture=ARM" )
if (CMAKE_VERSION VERSION_LESS 2.8.10)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -ccbin=${CMAKE_CXX_COMPILER}" )
endif()
endif()
# These vars will be processed in other scripts
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} ${NVCC_FLAGS_EXTRA})
@ -84,7 +98,7 @@ if(CUDA_FOUND)
message(STATUS "CUDA NVCC target flags: ${CUDA_NVCC_FLAGS}")
OCV_OPTION(CUDA_FAST_MATH "Enable --use_fast_math for CUDA compiler " OFF)
OCV_OPTION(CUDA_FAST_MATH "Enable --use_fast_math for CUDA compiler " OFF)
if(CUDA_FAST_MATH)
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} --use_fast_math)
@ -92,7 +106,6 @@ if(CUDA_FOUND)
mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD CUDA_SDK_ROOT_DIR)
unset(CUDA_npp_LIBRARY CACHE)
find_cuda_helper_libs(npp)
macro(ocv_cuda_compile VAR)
@ -106,15 +119,15 @@ if(CUDA_FOUND)
string(REPLACE "-ggdb3" "" ${var} "${${var}}")
endforeach()
if (BUILD_SHARED_LIBS)
if(BUILD_SHARED_LIBS)
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -DCVAPI_EXPORTS)
endif()
if(UNIX OR APPLE)
set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fPIC)
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fPIC)
endif()
if(APPLE)
set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fno-finite-math-only)
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fno-finite-math-only)
endif()
# disabled because of multiple warnings during building nvcc auto generated files

21
cmake/templates/cvconfig.h.cmake
View File

@ -172,21 +172,15 @@
/* NVidia Cuda Runtime API*/
#cmakedefine HAVE_CUDA
/* OpenCL Support */
#cmakedefine HAVE_OPENCL
/* AMD's OpenCL Fast Fourier Transform Library*/
#cmakedefine HAVE_CLAMDFFT
/* AMD's Basic Linear Algebra Subprograms Library*/
#cmakedefine HAVE_CLAMDBLAS
/* NVidia Cuda Fast Fourier Transform (FFT) API*/
#cmakedefine HAVE_CUFFT
/* NVidia Cuda Basic Linear Algebra Subprograms (BLAS) API*/
#cmakedefine HAVE_CUBLAS
/* NVidia Video Decoding API*/
#cmakedefine HAVE_NVCUVID
/* Compile for 'real' NVIDIA GPU architectures */
#define CUDA_ARCH_BIN "${OPENCV_CUDA_ARCH_BIN}"
@ -199,6 +193,15 @@
/* Create PTX or BIN for 1.0 compute capability */
#cmakedefine CUDA_ARCH_BIN_OR_PTX_10
/* OpenCL Support */
#cmakedefine HAVE_OPENCL
/* AMD's OpenCL Fast Fourier Transform Library*/
#cmakedefine HAVE_CLAMDFFT
/* AMD's Basic Linear Algebra Subprograms Library*/
#cmakedefine HAVE_CLAMDBLAS
/* VideoInput library */
#cmakedefine HAVE_VIDEOINPUT

1
modules/core/CMakeLists.txt
View File

@ -10,7 +10,6 @@ if(HAVE_CUDA)
file(GLOB lib_cuda "src/cuda/*.cu")
ocv_cuda_compile(cuda_objs ${lib_cuda})
set(cuda_link_libs ${CUDA_LIBRARIES} ${CUDA_npp_LIBRARY})
else()
set(lib_cuda "")

14
modules/core/include/opencv2/core/core.hpp
View File

@ -91,7 +91,7 @@ class SparseMat;
typedef Mat MatND;
class GlBuffer;
class GlTexture;
class GlTexture2D;
class GlArrays;
class GlCamera;
@ -1306,7 +1306,7 @@ public:
STD_VECTOR_MAT = 5 << KIND_SHIFT,
EXPR = 6 << KIND_SHIFT,
OPENGL_BUFFER = 7 << KIND_SHIFT,
OPENGL_TEXTURE = 8 << KIND_SHIFT,
OPENGL_TEXTURE2D = 8 << KIND_SHIFT,
GPU_MAT = 9 << KIND_SHIFT
};
_InputArray();
@ -1323,13 +1323,13 @@ public:
_InputArray(const Scalar& s);
_InputArray(const double& val);
_InputArray(const GlBuffer& buf);
_InputArray(const GlTexture& tex);
_InputArray(const GlTexture2D& tex);
_InputArray(const gpu::GpuMat& d_mat);
virtual Mat getMat(int i=-1) const;
virtual void getMatVector(vector<Mat>& mv) const;
virtual GlBuffer getGlBuffer() const;
virtual GlTexture getGlTexture() const;
virtual GlTexture2D getGlTexture2D() const;
virtual gpu::GpuMat getGpuMat() const;
virtual int kind() const;
@ -1380,6 +1380,8 @@ public:
template<typename _Tp, int m, int n> _OutputArray(Matx<_Tp, m, n>& matx);
template<typename _Tp> _OutputArray(_Tp* vec, int n);
_OutputArray(gpu::GpuMat& d_mat);
_OutputArray(GlBuffer& buf);
_OutputArray(GlTexture2D& tex);
_OutputArray(const Mat& m);
template<typename _Tp> _OutputArray(const vector<_Tp>& vec);
@ -1390,12 +1392,16 @@ public:
template<typename _Tp, int m, int n> _OutputArray(const Matx<_Tp, m, n>& matx);
template<typename _Tp> _OutputArray(const _Tp* vec, int n);
_OutputArray(const gpu::GpuMat& d_mat);
_OutputArray(const GlBuffer& buf);
_OutputArray(const GlTexture2D& tex);
virtual bool fixedSize() const;
virtual bool fixedType() const;
virtual bool needed() const;
virtual Mat& getMatRef(int i=-1) const;
virtual gpu::GpuMat& getGpuMatRef() const;
virtual GlBuffer& getGlBufferRef() const;
virtual GlTexture2D& getGlTexture2DRef() const;
virtual void create(Size sz, int type, int i=-1, bool allowTransposed=false, int fixedDepthMask=0) const;
virtual void create(int rows, int cols, int type, int i=-1, bool allowTransposed=false, int fixedDepthMask=0) const;
virtual void create(int dims, const int* size, int type, int i=-1, bool allowTransposed=false, int fixedDepthMask=0) const;

14
modules/core/include/opencv2/core/cuda_devptrs.hpp
View File

@ -152,6 +152,20 @@ namespace cv
//#undef __CV_GPU_DEPR_BEFORE__
//#undef __CV_GPU_DEPR_AFTER__
namespace device
{
using cv::gpu::PtrSz;
using cv::gpu::PtrStep;
using cv::gpu::PtrStepSz;
using cv::gpu::PtrStepSzb;
using cv::gpu::PtrStepSzf;
using cv::gpu::PtrStepSzi;
using cv::gpu::PtrStepb;
using cv::gpu::PtrStepf;
using cv::gpu::PtrStepi;
}
}
}

18
modules/core/include/opencv2/core/gpumat.hpp
View File

@ -79,6 +79,8 @@ namespace cv { namespace gpu
WARP_SHUFFLE_FUNCTIONS = FEATURE_SET_COMPUTE_30
};
CV_EXPORTS bool deviceSupports(FeatureSet feature_set);
// Gives information about what GPU archs this OpenCV GPU module was
// compiled for
class CV_EXPORTS TargetArchs
@ -545,22 +547,6 @@ namespace cv { namespace gpu
ensureSizeIsEnough(size.height, size.width, type, m);
}
inline void createContinuous(int rows, int cols, int type, GpuMat& m)
{
int area = rows * cols;
if (!m.isContinuous() || m.type() != type || m.size().area() != area)
ensureSizeIsEnough(1, area, type, m);
m = m.reshape(0, rows);
}
inline void ensureSizeIsEnough(int rows, int cols, int type, GpuMat& m)
{
if (m.type() == type && m.rows >= rows && m.cols >= cols)
m = m(Rect(0, 0, cols, rows));
else
m.create(rows, cols, type);
}
inline GpuMat allocMatFromBuf(int rows, int cols, int type, GpuMat &mat)
{
if (!mat.empty() && mat.type() == type && mat.rows >= rows && mat.cols >= cols)

35
modules/core/include/opencv2/core/internal.hpp
View File

@ -750,39 +750,4 @@ typedef struct CvBigFuncTable
(tab).fn_2d[CV_32F] = (void*)FUNCNAME##_32f##FLAG; \
(tab).fn_2d[CV_64F] = (void*)FUNCNAME##_64f##FLAG
#ifdef __cplusplus
//! OpenGL extension table
class CV_EXPORTS CvOpenGlFuncTab
{
public:
virtual ~CvOpenGlFuncTab();
virtual void genBuffers(int n, unsigned int* buffers) const = 0;
virtual void deleteBuffers(int n, const unsigned int* buffers) const = 0;
virtual void bufferData(unsigned int target, ptrdiff_t size, const void* data, unsigned int usage) const = 0;
virtual void bufferSubData(unsigned int target, ptrdiff_t offset, ptrdiff_t size, const void* data) const = 0;
virtual void bindBuffer(unsigned int target, unsigned int buffer) const = 0;
virtual void* mapBuffer(unsigned int target, unsigned int access) const = 0;
virtual void unmapBuffer(unsigned int target) const = 0;
virtual void generateBitmapFont(const std::string& family, int height, int weight, bool italic, bool underline, int start, int count, int base) const = 0;
virtual bool isGlContextInitialized() const = 0;
};
CV_EXPORTS void icvSetOpenGlFuncTab(const CvOpenGlFuncTab* tab);
CV_EXPORTS bool icvCheckGlError(const char* file, const int line, const char* func = "");
#if defined(__GNUC__)
#define CV_CheckGlError() CV_DbgAssert( (::icvCheckGlError(__FILE__, __LINE__, __func__)) )
#else
#define CV_CheckGlError() CV_DbgAssert( (::icvCheckGlError(__FILE__, __LINE__)) )
#endif
#endif //__cplusplus
#endif // __OPENCV_CORE_INTERNAL_HPP__

297
modules/core/include/opencv2/core/opengl_interop.hpp
View File

@ -47,205 +47,212 @@
#include "opencv2/core/core.hpp"
namespace cv
{
namespace cv {
CV_EXPORTS bool checkGlError(const char* file, const int line, const char* func = "");
#if defined(__GNUC__)
#define CV_CheckGlError() CV_DbgAssert( (cv::checkGlError(__FILE__, __LINE__, __func__)) )
#else
#define CV_CheckGlError() CV_DbgAssert( (cv::checkGlError(__FILE__, __LINE__)) )
#endif
/////////////////// OpenGL Objects ///////////////////
//! Smart pointer for OpenGL buffer memory with reference counting.
class CV_EXPORTS GlBuffer
{
public:
enum Usage
enum Target
{
ARRAY_BUFFER = 0x8892, // buffer will use for OpenGL arrays (vertices, colors, normals, etc)
TEXTURE_BUFFER = 0x88EC // buffer will ise for OpenGL textures
ARRAY_BUFFER = 0x8892, //!< The buffer will be used as a source for vertex data
ELEMENT_ARRAY_BUFFER = 0x8893, //!< The buffer will be used for indices (in glDrawElements, for example)
PIXEL_PACK_BUFFER = 0x88EB, //!< The buffer will be used for reading from OpenGL textures
PIXEL_UNPACK_BUFFER = 0x88EC //!< The buffer will be used for writing to OpenGL textures
};
enum Access
{
READ_ONLY = 0x88B8,
WRITE_ONLY = 0x88B9,
READ_WRITE = 0x88BA
};
//! create empty buffer
explicit GlBuffer(Usage usage);
GlBuffer();
//! create buffer from existed buffer id
GlBuffer(int arows, int acols, int atype, unsigned int abufId, bool autoRelease = false);
GlBuffer(Size asize, int atype, unsigned int abufId, bool autoRelease = false);
//! create buffer
GlBuffer(int rows, int cols, int type, Usage usage);
GlBuffer(Size size, int type, Usage usage);
GlBuffer(int arows, int acols, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false);
GlBuffer(Size asize, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false);
//! copy from host/device memory
GlBuffer(InputArray mat, Usage usage);
explicit GlBuffer(InputArray arr, Target target = ARRAY_BUFFER, bool autoRelease = false);
void create(int rows, int cols, int type, Usage usage);
void create(Size size, int type, Usage usage);
void create(int rows, int cols, int type);
void create(Size size, int type);
//! create buffer
void create(int arows, int acols, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false);
void create(Size asize, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false) { create(asize.height, asize.width, atype, target, autoRelease); }
//! release memory and delete buffer object
void release();
//! copy from host/device memory
void copyFrom(InputArray mat);
//! set auto release mode (if true, release will be called in object's destructor)
void setAutoRelease(bool flag);
void bind() const;
void unbind() const;
//! copy from host/device memory
void copyFrom(InputArray arr, Target target = ARRAY_BUFFER, bool autoRelease = false);
//! copy to host/device memory
void copyTo(OutputArray arr, Target target = ARRAY_BUFFER, bool autoRelease = false) const;
//! create copy of current buffer
GlBuffer clone(Target target = ARRAY_BUFFER, bool autoRelease = false) const;
//! bind buffer for specified target
void bind(Target target) const;
//! unbind any buffers from specified target
static void unbind(Target target);
//! map to host memory
Mat mapHost();
Mat mapHost(Access access);
void unmapHost();
//! map to device memory
gpu::GpuMat mapDevice();
void unmapDevice();
inline int rows() const { return rows_; }
inline int cols() const { return cols_; }
inline Size size() const { return Size(cols_, rows_); }
inline bool empty() const { return rows_ == 0 || cols_ == 0; }
int rows() const { return rows_; }
int cols() const { return cols_; }
Size size() const { return Size(cols_, rows_); }
bool empty() const { return rows_ == 0 || cols_ == 0; }
inline int type() const { return type_; }
inline int depth() const { return CV_MAT_DEPTH(type_); }
inline int channels() const { return CV_MAT_CN(type_); }
inline int elemSize() const { return CV_ELEM_SIZE(type_); }
inline int elemSize1() const { return CV_ELEM_SIZE1(type_); }
int type() const { return type_; }
int depth() const { return CV_MAT_DEPTH(type_); }
int channels() const { return CV_MAT_CN(type_); }
int elemSize() const { return CV_ELEM_SIZE(type_); }
int elemSize1() const { return CV_ELEM_SIZE1(type_); }
inline Usage usage() const { return usage_; }
unsigned int bufId() const;
class Impl;
private:
Ptr<Impl> impl_;
int rows_;
int cols_;
int type_;
Usage usage_;
Ptr<Impl> impl_;
};
template <> CV_EXPORTS void Ptr<GlBuffer::Impl>::delete_obj();
//! Smart pointer for OpenGL 2d texture memory with reference counting.
class CV_EXPORTS GlTexture
//! Smart pointer for OpenGL 2D texture memory with reference counting.
class CV_EXPORTS GlTexture2D
{
public:
enum Format
{
NONE = 0,
DEPTH_COMPONENT = 0x1902, //!< Depth
RGB = 0x1907, //!< Red, Green, Blue
RGBA = 0x1908 //!< Red, Green, Blue, Alpha
};
//! create empty texture
GlTexture();
GlTexture2D();
//! create texture from existed texture id
GlTexture2D(int arows, int acols, Format aformat, unsigned int atexId, bool autoRelease = false);
GlTexture2D(Size asize, Format aformat, unsigned int atexId, bool autoRelease = false);
//! create texture
GlTexture(int rows, int cols, int type);
GlTexture(Size size, int type);
GlTexture2D(int arows, int acols, Format aformat, bool autoRelease = false);
GlTexture2D(Size asize, Format aformat, bool autoRelease = false);
//! copy from host/device memory
explicit GlTexture(InputArray mat, bool bgra = true);
explicit GlTexture2D(InputArray arr, bool autoRelease = false);
void create(int rows, int cols, int type);
void create(Size size, int type);
//! create texture
void create(int arows, int acols, Format aformat, bool autoRelease = false);
void create(Size asize, Format aformat, bool autoRelease = false) { create(asize.height, asize.width, aformat, autoRelease); }
//! release memory and delete texture object
void release();
//! set auto release mode (if true, release will be called in object's destructor)
void setAutoRelease(bool flag);
//! copy from host/device memory
void copyFrom(InputArray mat, bool bgra = true);
void copyFrom(InputArray arr, bool autoRelease = false);
//! copy to host/device memory
void copyTo(OutputArray arr, int ddepth = CV_32F, bool autoRelease = false) const;
//! bind texture to current active texture unit for GL_TEXTURE_2D target
void bind() const;
void unbind() const;
inline int rows() const { return rows_; }
inline int cols() const { return cols_; }
inline Size size() const { return Size(cols_, rows_); }
inline bool empty() const { return rows_ == 0 || cols_ == 0; }
int rows() const { return rows_; }
int cols() const { return cols_; }
Size size() const { return Size(cols_, rows_); }
bool empty() const { return rows_ == 0 || cols_ == 0; }
inline int type() const { return type_; }
inline int depth() const { return CV_MAT_DEPTH(type_); }
inline int channels() const { return CV_MAT_CN(type_); }
inline int elemSize() const { return CV_ELEM_SIZE(type_); }
inline int elemSize1() const { return CV_ELEM_SIZE1(type_); }
Format format() const { return format_; }
unsigned int texId() const;
class Impl;
private:
Ptr<Impl> impl_;
int rows_;
int cols_;
int type_;
Ptr<Impl> impl_;
GlBuffer buf_;
Format format_;
};
template <> CV_EXPORTS void Ptr<GlTexture::Impl>::delete_obj();
template <> CV_EXPORTS void Ptr<GlTexture2D::Impl>::delete_obj();
//! OpenGL Arrays
class CV_EXPORTS GlArrays
{
public:
inline GlArrays()
: vertex_(GlBuffer::ARRAY_BUFFER), color_(GlBuffer::ARRAY_BUFFER), bgra_(true), normal_(GlBuffer::ARRAY_BUFFER), texCoord_(GlBuffer::ARRAY_BUFFER)
{
}
GlArrays();
void setVertexArray(InputArray vertex);
inline void resetVertexArray() { vertex_.release(); }
void resetVertexArray();
void setColorArray(InputArray color, bool bgra = true);
inline void resetColorArray() { color_.release(); }
void setColorArray(InputArray color);
void resetColorArray();
void setNormalArray(InputArray normal);
inline void resetNormalArray() { normal_.release(); }
void resetNormalArray();
void setTexCoordArray(InputArray texCoord);
inline void resetTexCoordArray() { texCoord_.release(); }
void resetTexCoordArray();
void release();
void setAutoRelease(bool flag);
void bind() const;
void unbind() const;
inline int rows() const { return vertex_.rows(); }
inline int cols() const { return vertex_.cols(); }
inline Size size() const { return vertex_.size(); }
inline bool empty() const { return vertex_.empty(); }
int size() const { return size_; }
bool empty() const { return size_ == 0; }
private:
int size_;
GlBuffer vertex_;
GlBuffer color_;
bool bgra_;
GlBuffer normal_;
GlBuffer texCoord_;
};
//! OpenGL Font
class CV_EXPORTS GlFont
{
public:
enum Weight
{
WEIGHT_LIGHT = 300,
WEIGHT_NORMAL = 400,
WEIGHT_SEMIBOLD = 600,
WEIGHT_BOLD = 700,
WEIGHT_BLACK = 900
};
enum Style
{
STYLE_NORMAL = 0,
STYLE_ITALIC = 1,
STYLE_UNDERLINE = 2
};
static Ptr<GlFont> get(const std::string& family, int height = 12, Weight weight = WEIGHT_NORMAL, Style style = STYLE_NORMAL);
void draw(const char* str, size_t len) const;
inline const std::string& family() const { return family_; }
inline int height() const { return height_; }
inline Weight weight() const { return weight_; }
inline Style style() const { return style_; }
private:
GlFont(const std::string& family, int height, Weight weight, Style style);
std::string family_;
int height_;
Weight weight_;
Style style_;
unsigned int base_;
GlFont(const GlFont&);
GlFont& operator =(const GlFont&);
};
//! render functions
/////////////////// Render Functions ///////////////////
//! render texture rectangle in window
CV_EXPORTS void render(const GlTexture& tex,
CV_EXPORTS void render(const GlTexture2D& tex,
Rect_<double> wndRect = Rect_<double>(0.0, 0.0, 1.0, 1.0),
Rect_<double> texRect = Rect_<double>(0.0, 0.0, 1.0, 1.0));
@ -267,67 +274,13 @@ namespace RenderMode {
//! render OpenGL arrays
CV_EXPORTS void render(const GlArrays& arr, int mode = RenderMode::POINTS, Scalar color = Scalar::all(255));
CV_EXPORTS void render(const GlArrays& arr, InputArray indices, int mode = RenderMode::POINTS, Scalar color = Scalar::all(255));
CV_EXPORTS void render(const std::string& str, const Ptr<GlFont>& font, Scalar color, Point2d pos);
//! OpenGL camera
class CV_EXPORTS GlCamera
{
public:
GlCamera();
void lookAt(Point3d eye, Point3d center, Point3d up);
void setCameraPos(Point3d pos, double yaw, double pitch, double roll);
void setScale(Point3d scale);
void setProjectionMatrix(const Mat& projectionMatrix, bool transpose = true);
void setPerspectiveProjection(double fov, double aspect, double zNear, double zFar);
void setOrthoProjection(double left, double right, double bottom, double top, double zNear, double zFar);
void setupProjectionMatrix() const;
void setupModelViewMatrix() const;
private:
Point3d eye_;
Point3d center_;
Point3d up_;
Point3d pos_;
double yaw_;
double pitch_;
double roll_;
bool useLookAtParams_;
Point3d scale_;
Mat projectionMatrix_;
double fov_;
double aspect_;
double left_;
double right_;
double bottom_;
double top_;
double zNear_;
double zFar_;
bool perspectiveProjection_;
};
inline void GlBuffer::create(Size _size, int _type, Usage _usage) { create(_size.height, _size.width, _type, _usage); }
inline void GlBuffer::create(int _rows, int _cols, int _type) { create(_rows, _cols, _type, usage()); }
inline void GlBuffer::create(Size _size, int _type) { create(_size.height, _size.width, _type, usage()); }
inline void GlTexture::create(Size _size, int _type) { create(_size.height, _size.width, _type); }
namespace gpu
{
namespace gpu {
//! set a CUDA device to use OpenGL interoperability
CV_EXPORTS void setGlDevice(int device = 0);
}
} // namespace cv
#endif // __cplusplus

106
modules/core/src/cuda/matrix_operations.cu
View File

@ -44,6 +44,7 @@
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/type_traits.hpp"
namespace cv { namespace gpu { namespace device
{
@ -54,6 +55,7 @@ namespace cv { namespace gpu { namespace device
void writeScalar(const int*);
void writeScalar(const float*);
void writeScalar(const double*);
void copyToWithMask_gpu(PtrStepSzb src, PtrStepSzb dst, size_t elemSize1, int cn, PtrStepSzb mask, bool colorMask, cudaStream_t stream);
void convert_gpu(PtrStepSzb, int, PtrStepSzb, int, double, double, cudaStream_t);
}}}
@ -226,16 +228,16 @@ namespace cv { namespace gpu { namespace device
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template <typename T, typename D> struct Convertor : unary_function<T, D>
template <typename T, typename D, typename S> struct Convertor : unary_function<T, D>
{
Convertor(double alpha_, double beta_) : alpha(alpha_), beta(beta_) {}
Convertor(S alpha_, S beta_) : alpha(alpha_), beta(beta_) {}
__device__ __forceinline__ D operator()(const T& src) const
__device__ __forceinline__ D operator()(typename TypeTraits<T>::ParameterType src) const
{
return saturate_cast<D>(alpha * src + beta);
}
double alpha, beta;
S alpha, beta;
};
namespace detail
@ -282,16 +284,16 @@ namespace cv { namespace gpu { namespace device
};
}
template <typename T, typename D> struct TransformFunctorTraits< Convertor<T, D> > : detail::ConvertTraits< Convertor<T, D> >
template <typename T, typename D, typename S> struct TransformFunctorTraits< Convertor<T, D, S> > : detail::ConvertTraits< Convertor<T, D, S> >
{
};
template<typename T, typename D>
template<typename T, typename D, typename S>
void cvt_(PtrStepSzb src, PtrStepSzb dst, double alpha, double beta, cudaStream_t stream)
{
cudaSafeCall( cudaSetDoubleForDevice(&alpha) );
cudaSafeCall( cudaSetDoubleForDevice(&beta) );
Convertor<T, D> op(alpha, beta);
Convertor<T, D, S> op(static_cast<S>(alpha), static_cast<S>(beta));
cv::gpu::device::transform((PtrStepSz<T>)src, (PtrStepSz<D>)dst, op, WithOutMask(), stream);
}
@ -304,36 +306,74 @@ namespace cv { namespace gpu { namespace device
{
typedef void (*caller_t)(PtrStepSzb src, PtrStepSzb dst, double alpha, double beta, cudaStream_t stream);
static const caller_t tab[8][8] =
static const caller_t tab[7][7] =
{
{cvt_<uchar, uchar>, cvt_<uchar, schar>, cvt_<uchar, ushort>, cvt_<uchar, short>,
cvt_<uchar, int>, cvt_<uchar, float>, cvt_<uchar, double>, 0},
{cvt_<schar, uchar>, cvt_<schar, schar>, cvt_<schar, ushort>, cvt_<schar, short>,
cvt_<schar, int>, cvt_<schar, float>, cvt_<schar, double>, 0},
{cvt_<ushort, uchar>, cvt_<ushort, schar>, cvt_<ushort, ushort>, cvt_<ushort, short>,
cvt_<ushort, int>, cvt_<ushort, float>, cvt_<ushort, double>, 0},
{cvt_<short, uchar>, cvt_<short, schar>, cvt_<short, ushort>, cvt_<short, short>,
cvt_<short, int>, cvt_<short, float>, cvt_<short, double>, 0},
{cvt_<int, uchar>, cvt_<int, schar>, cvt_<int, ushort>,
cvt_<int, short>, cvt_<int, int>, cvt_<int, float>, cvt_<int, double>, 0},
{cvt_<float, uchar>, cvt_<float, schar>, cvt_<float, ushort>,
cvt_<float, short>, cvt_<float, int>, cvt_<float, float>, cvt_<float, double>, 0},
{cvt_<double, uchar>, cvt_<double, schar>, cvt_<double, ushort>,
cvt_<double, short>, cvt_<double, int>, cvt_<double, float>, cvt_<double, double>, 0},
{0,0,0,0,0,0,0,0}
{
cvt_<uchar, uchar, float>,
cvt_<uchar, schar, float>,
cvt_<uchar, ushort, float>,
cvt_<uchar, short, float>,
cvt_<uchar, int, float>,
cvt_<uchar, float, float>,
cvt_<uchar, double, double>
},
{
cvt_<schar, uchar, float>,
cvt_<schar, schar, float>,
cvt_<schar, ushort, float>,
cvt_<schar, short, float>,
cvt_<schar, int, float>,
cvt_<schar, float, float>,
cvt_<schar, double, double>
},
{
cvt_<ushort, uchar, float>,
cvt_<ushort, schar, float>,
cvt_<ushort, ushort, float>,
cvt_<ushort, short, float>,
cvt_<ushort, int, float>,
cvt_<ushort, float, float>,
cvt_<ushort, double, double>
},
{
cvt_<short, uchar, float>,
cvt_<short, schar, float>,
cvt_<short, ushort, float>,
cvt_<short, short, float>,
cvt_<short, int, float>,
cvt_<short, float, float>,
cvt_<short, double, double>
},
{
cvt_<int, uchar, float>,
cvt_<int, schar, float>,
cvt_<int, ushort, float>,
cvt_<int, short, float>,
cvt_<int, int, double>,
cvt_<int, float, double>,
cvt_<int, double, double>
},
{
cvt_<float, uchar, float>,
cvt_<float, schar, float>,
cvt_<float, ushort, float>,
cvt_<float, short, float>,
cvt_<float, int, float>,
cvt_<float, float, float>,
cvt_<float, double, double>
},
{
cvt_<double, uchar, double>,
cvt_<double, schar, double>,
cvt_<double, ushort, double>,
cvt_<double, short, double>,
cvt_<double, int, double>,
cvt_<double, float, double>,
cvt_<double, double, double>
}
};
caller_t func = tab[sdepth][ddepth];
if (!func)
cv::gpu::error("Unsupported convert operation", __FILE__, __LINE__, "convert_gpu");
func(src, dst, alpha, beta, stream);
}

2718
modules/core/src/gl_core_3_1.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

1331
modules/core/src/gl_core_3_1.hpp Normal file
View File

File diff suppressed because it is too large Load Diff

200
modules/core/src/gpumat.cpp
View File

@ -45,8 +45,7 @@
#include <iostream>
#ifdef HAVE_CUDA
#include <cuda.h>
#include <cuda_runtime_api.h>
#include <cuda_runtime.h>
#include <npp.h>
#define CUDART_MINIMUM_REQUIRED_VERSION 4010
@ -69,33 +68,89 @@ using namespace cv::gpu;
namespace
{
// Compares value to set using the given comparator. Returns true if
// there is at least one element x in the set satisfying to: x cmp value
// predicate.
template <typename Comparer>
bool compareToSet(const std::string& set_as_str, int value, Comparer cmp)
class CudaArch
{
public:
CudaArch();
bool builtWith(FeatureSet feature_set) const;
bool hasPtx(int major, int minor) const;
bool hasBin(int major, int minor) const;
bool hasEqualOrLessPtx(int major, int minor) const;
bool hasEqualOrGreaterPtx(int major, int minor) const;
bool hasEqualOrGreaterBin(int major, int minor) const;
private:
static void fromStr(const string& set_as_str, vector<int>& arr);
vector<int> bin;
vector<int> ptx;
vector<int> features;
};
const CudaArch cudaArch;
CudaArch::CudaArch()
{
#ifdef HAVE_CUDA
fromStr(CUDA_ARCH_BIN, bin);
fromStr(CUDA_ARCH_PTX, ptx);
fromStr(CUDA_ARCH_FEATURES, features);
#endif
}
bool CudaArch::builtWith(FeatureSet feature_set) const
{
return !features.empty() && (features.back() >= feature_set);
}
bool CudaArch::hasPtx(int major, int minor) const
{
return find(ptx.begin(), ptx.end(), major * 10 + minor) != ptx.end();
}
bool CudaArch::hasBin(int major, int minor) const
{
return find(bin.begin(), bin.end(), major * 10 + minor) != bin.end();
}
bool CudaArch::hasEqualOrLessPtx(int major, int minor) const
{
return !ptx.empty() && (ptx.front() <= major * 10 + minor);
}
bool CudaArch::hasEqualOrGreaterPtx(int major, int minor) const
{
return !ptx.empty() && (ptx.back() >= major * 10 + minor);
}
bool CudaArch::hasEqualOrGreaterBin(int major, int minor) const
{
return !bin.empty() && (bin.back() >= major * 10 + minor);
}
void CudaArch::fromStr(const string& set_as_str, vector<int>& arr)
{
if (set_as_str.find_first_not_of(" ") == string::npos)
return false;
return;
std::stringstream stream(set_as_str);
istringstream stream(set_as_str);
int cur_value;
while (!stream.eof())
{
stream >> cur_value;
if (cmp(cur_value, value))
return true;
arr.push_back(cur_value);
}
return false;
sort(arr.begin(), arr.end());
}
}
bool cv::gpu::TargetArchs::builtWith(cv::gpu::FeatureSet feature_set)
{
#if defined (HAVE_CUDA)
return ::compareToSet(CUDA_ARCH_FEATURES, feature_set, std::greater_equal<int>());
return cudaArch.builtWith(feature_set);
#else
(void)feature_set;
return false;
@ -110,7 +165,7 @@ bool cv::gpu::TargetArchs::has(int major, int minor)
bool cv::gpu::TargetArchs::hasPtx(int major, int minor)
{
#if defined (HAVE_CUDA)
return ::compareToSet(CUDA_ARCH_PTX, major * 10 + minor, std::equal_to<int>());
return cudaArch.hasPtx(major, minor);
#else
(void)major;
(void)minor;
@ -121,7 +176,7 @@ bool cv::gpu::TargetArchs::hasPtx(int major, int minor)
bool cv::gpu::TargetArchs::hasBin(int major, int minor)
{
#if defined (HAVE_CUDA)
return ::compareToSet(CUDA_ARCH_BIN, major * 10 + minor, std::equal_to<int>());
return cudaArch.hasBin(major, minor);
#else
(void)major;
(void)minor;
@ -132,8 +187,7 @@ bool cv::gpu::TargetArchs::hasBin(int major, int minor)
bool cv::gpu::TargetArchs::hasEqualOrLessPtx(int major, int minor)
{
#if defined (HAVE_CUDA)
return ::compareToSet(CUDA_ARCH_PTX, major * 10 + minor,
std::less_equal<int>());
return cudaArch.hasEqualOrLessPtx(major, minor);
#else
(void)major;
(void)minor;
@ -143,14 +197,13 @@ bool cv::gpu::TargetArchs::hasEqualOrLessPtx(int major, int minor)
bool cv::gpu::TargetArchs::hasEqualOrGreater(int major, int minor)
{
return hasEqualOrGreaterPtx(major, minor) ||
hasEqualOrGreaterBin(major, minor);
return hasEqualOrGreaterPtx(major, minor) || hasEqualOrGreaterBin(major, minor);
}
bool cv::gpu::TargetArchs::hasEqualOrGreaterPtx(int major, int minor)
{
#if defined (HAVE_CUDA)
return ::compareToSet(CUDA_ARCH_PTX, major * 10 + minor, std::greater_equal<int>());
return cudaArch.hasEqualOrGreaterPtx(major, minor);
#else
(void)major;
(void)minor;
@ -161,8 +214,7 @@ bool cv::gpu::TargetArchs::hasEqualOrGreaterPtx(int major, int minor)
bool cv::gpu::TargetArchs::hasEqualOrGreaterBin(int major, int minor)
{
#if defined (HAVE_CUDA)
return ::compareToSet(CUDA_ARCH_BIN, major * 10 + minor,
std::greater_equal<int>());
return cudaArch.hasEqualOrGreaterBin(major, minor);
#else
(void)major;
(void)minor;
@ -170,6 +222,31 @@ bool cv::gpu::TargetArchs::hasEqualOrGreaterBin(int major, int minor)
#endif
}
bool cv::gpu::deviceSupports(FeatureSet feature_set)
{
static int versions[] =
{
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1
};
static const int cache_size = static_cast<int>(sizeof(versions) / sizeof(versions[0]));
const int devId = getDevice();
int version;
if (devId < cache_size && versions[devId] >= 0)
version = versions[devId];
else
{
DeviceInfo dev(devId);
version = dev.majorVersion() * 10 + dev.minorVersion();
if (devId < cache_size)
versions[devId] = version;
}
return TargetArchs::builtWith(feature_set) && (version >= feature_set);
}
#if !defined (HAVE_CUDA)
#define throw_nogpu CV_Error(CV_GpuNotSupported, "The library is compiled without CUDA support")
@ -316,18 +393,6 @@ void cv::gpu::DeviceInfo::queryMemory(size_t& free_memory, size_t& total_memory)
namespace
{
template <class T> void getCudaAttribute(T *attribute, CUdevice_attribute device_attribute, int device)
{
*attribute = T();
//CUresult error = CUDA_SUCCESS;// = cuDeviceGetAttribute( attribute, device_attribute, device ); why link erros under ubuntu??
CUresult error = cuDeviceGetAttribute( attribute, device_attribute, device );
if( CUDA_SUCCESS == error )
return;
printf("Driver API error = %04d\n", error);
cv::gpu::error("driver API error", __FILE__, __LINE__);
}
int convertSMVer2Cores(int major, int minor)
{
// Defines for GPU Architecture types (using the SM version to determine the # of cores per SM
@ -336,7 +401,7 @@ namespace
int Cores;
} SMtoCores;
SMtoCores gpuArchCoresPerSM[] = { { 0x10, 8 }, { 0x11, 8 }, { 0x12, 8 }, { 0x13, 8 }, { 0x20, 32 }, { 0x21, 48 }, {0x30, 192}, { -1, -1 } };
SMtoCores gpuArchCoresPerSM[] = { { 0x10, 8 }, { 0x11, 8 }, { 0x12, 8 }, { 0x13, 8 }, { 0x20, 32 }, { 0x21, 48 }, {0x30, 192}, {0x35, 192}, { -1, -1 } };
int index = 0;
while (gpuArchCoresPerSM[index].SM != -1)
@ -345,7 +410,7 @@ namespace
return gpuArchCoresPerSM[index].Cores;
index++;
}
printf("MapSMtoCores undefined SMversion %d.%d!\n", major, minor);
return -1;
}
}
@ -383,22 +448,13 @@ void cv::gpu::printCudaDeviceInfo(int device)
printf(" CUDA Driver Version / Runtime Version %d.%d / %d.%d\n", driverVersion/1000, driverVersion%100, runtimeVersion/1000, runtimeVersion%100);
printf(" CUDA Capability Major/Minor version number: %d.%d\n", prop.major, prop.minor);
printf(" Total amount of global memory: %.0f MBytes (%llu bytes)\n", (float)prop.totalGlobalMem/1048576.0f, (unsigned long long) prop.totalGlobalMem);
printf(" (%2d) Multiprocessors x (%2d) CUDA Cores/MP: %d CUDA Cores\n",
prop.multiProcessorCount, convertSMVer2Cores(prop.major, prop.minor),
convertSMVer2Cores(prop.major, prop.minor) * prop.multiProcessorCount);
int cores = convertSMVer2Cores(prop.major, prop.minor);
if (cores > 0)
printf(" (%2d) Multiprocessors x (%2d) CUDA Cores/MP: %d CUDA Cores\n", prop.multiProcessorCount, cores, cores * prop.multiProcessorCount);
printf(" GPU Clock Speed: %.2f GHz\n", prop.clockRate * 1e-6f);
// This is not available in the CUDA Runtime API, so we make the necessary calls the driver API to support this for output
int memoryClock, memBusWidth, L2CacheSize;
getCudaAttribute<int>( &memoryClock, CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE, dev );
getCudaAttribute<int>( &memBusWidth, CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH, dev );
getCudaAttribute<int>( &L2CacheSize, CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE, dev );
printf(" Memory Clock rate: %.2f Mhz\n", memoryClock * 1e-3f);
printf(" Memory Bus Width: %d-bit\n", memBusWidth);
if (L2CacheSize)
printf(" L2 Cache Size: %d bytes\n", L2CacheSize);
printf(" Max Texture Dimension Size (x,y,z) 1D=(%d), 2D=(%d,%d), 3D=(%d,%d,%d)\n",
prop.maxTexture1D, prop.maxTexture2D[0], prop.maxTexture2D[1],
prop.maxTexture3D[0], prop.maxTexture3D[1], prop.maxTexture3D[2]);
@ -458,7 +514,12 @@ void cv::gpu::printShortCudaDeviceInfo(int device)
const char *arch_str = prop.major < 2 ? " (not Fermi)" : "";
printf("Device %d: \"%s\" %.0fMb", dev, prop.name, (float)prop.totalGlobalMem/1048576.0f);
printf(", sm_%d%d%s, %d cores", prop.major, prop.minor, arch_str, convertSMVer2Cores(prop.major, prop.minor) * prop.multiProcessorCount);
printf(", sm_%d%d%s", prop.major, prop.minor, arch_str);
int cores = convertSMVer2Cores(prop.major, prop.minor);
if (cores > 0)
printf(", %d cores", cores * prop.multiProcessorCount);
printf(", Driver/Runtime ver.%d.%d/%d.%d\n", driverVersion/1000, driverVersion%100, runtimeVersion/1000, runtimeVersion%100);
}
fflush(stdout);
@ -704,6 +765,43 @@ cv::Mat::Mat(const GpuMat& m) : flags(0), dims(0), rows(0), cols(0), data(0), re
m.download(*this);
}
void cv::gpu::createContinuous(int rows, int cols, int type, GpuMat& m)
{
int area = rows * cols;
if (m.empty() || m.type() != type || !m.isContinuous() || m.size().area() < area)
m.create(1, area, type);
m.cols = cols;
m.rows = rows;
m.step = m.elemSize() * cols;
m.flags |= Mat::CONTINUOUS_FLAG;
}
void cv::gpu::ensureSizeIsEnough(int rows, int cols, int type, GpuMat& m)
{
if (m.empty() || m.type() != type || m.data != m.datastart)
m.create(rows, cols, type);
else
{
const size_t esz = m.elemSize();
const ptrdiff_t delta2 = m.dataend - m.datastart;
const size_t minstep = m.cols * esz;
Size wholeSize;
wholeSize.height = std::max(static_cast<int>((delta2 - minstep) / m.step + 1), m.rows);
wholeSize.width = std::max(static_cast<int>((delta2 - m.step * (wholeSize.height - 1)) / esz), m.cols);
if (wholeSize.height < rows || wholeSize.width < cols)
m.create(rows, cols, type);
else
{
m.cols = cols;
m.rows = rows;
}
}
}
namespace
{
class GpuFuncTable

65
modules/core/src/matrix.cpp
View File

@ -922,8 +922,8 @@ _InputArray::_InputArray(const Mat& m) : flags(MAT), obj((void*)&m) {}
_InputArray::_InputArray(const vector<Mat>& vec) : flags(STD_VECTOR_MAT), obj((void*)&vec) {}
_InputArray::_InputArray(const double& val) : flags(FIXED_TYPE + FIXED_SIZE + MATX + CV_64F), obj((void*)&val), sz(Size(1,1)) {}
_InputArray::_InputArray(const MatExpr& expr) : flags(FIXED_TYPE + FIXED_SIZE + EXPR), obj((void*)&expr) {}
_InputArray::_InputArray(const GlBuffer& buf) : flags(FIXED_TYPE + FIXED_SIZE + OPENGL_BUFFER), obj((void*)&buf) {}
_InputArray::_InputArray(const GlTexture& tex) : flags(FIXED_TYPE + FIXED_SIZE + OPENGL_TEXTURE), obj((void*)&tex) {}
_InputArray::_InputArray(const GlBuffer& buf) : flags(OPENGL_BUFFER), obj((void*)&buf) {}
_InputArray::_InputArray(const GlTexture2D &tex) : flags(OPENGL_TEXTURE2D), obj((void*)&tex) {}
_InputArray::_InputArray(const gpu::GpuMat& d_mat) : flags(GPU_MAT), obj((void*)&d_mat) {}
Mat _InputArray::getMat(int i) const
@ -1076,14 +1076,14 @@ GlBuffer _InputArray::getGlBuffer() const
}
}
GlTexture _InputArray::getGlTexture() const
GlTexture2D _InputArray::getGlTexture2D() const
{
int k = kind();
CV_Assert(k == OPENGL_TEXTURE);
CV_Assert(k == OPENGL_TEXTURE2D);
//if( k == OPENGL_TEXTURE )
{
const GlTexture* tex = (const GlTexture*)obj;
const GlTexture2D* tex = (const GlTexture2D*)obj;
return *tex;
}
}
@ -1168,10 +1168,10 @@ Size _InputArray::size(int i) const
return buf->size();
}
if( k == OPENGL_TEXTURE )
if( k == OPENGL_TEXTURE2D )
{
CV_Assert( i < 0 );
const GlTexture* tex = (const GlTexture*)obj;
const GlTexture2D* tex = (const GlTexture2D*)obj;
return tex->size();
}
@ -1216,9 +1216,6 @@ int _InputArray::type(int i) const
if( k == OPENGL_BUFFER )
return ((const GlBuffer*)obj)->type();
if( k == OPENGL_TEXTURE )
return ((const GlTexture*)obj)->type();
CV_Assert( k == GPU_MAT );
//if( k == GPU_MAT )
return ((const gpu::GpuMat*)obj)->type();
@ -1271,8 +1268,8 @@ bool _InputArray::empty() const
if( k == OPENGL_BUFFER )
return ((const GlBuffer*)obj)->empty();
if( k == OPENGL_TEXTURE )
return ((const GlTexture*)obj)->empty();
if( k == OPENGL_TEXTURE2D )
return ((const GlTexture2D*)obj)->empty();
CV_Assert( k == GPU_MAT );
//if( k == GPU_MAT )
@ -1285,10 +1282,14 @@ _OutputArray::~_OutputArray() {}
_OutputArray::_OutputArray(Mat& m) : _InputArray(m) {}
_OutputArray::_OutputArray(vector<Mat>& vec) : _InputArray(vec) {}
_OutputArray::_OutputArray(gpu::GpuMat& d_mat) : _InputArray(d_mat) {}
_OutputArray::_OutputArray(GlBuffer& buf) : _InputArray(buf) {}
_OutputArray::_OutputArray(GlTexture2D& tex) : _InputArray(tex) {}
_OutputArray::_OutputArray(const Mat& m) : _InputArray(m) {flags |= FIXED_SIZE|FIXED_TYPE;}
_OutputArray::_OutputArray(const vector<Mat>& vec) : _InputArray(vec) {flags |= FIXED_SIZE;}
_OutputArray::_OutputArray(const gpu::GpuMat& d_mat) : _InputArray(d_mat) {flags |= FIXED_SIZE|FIXED_TYPE;}
_OutputArray::_OutputArray(const GlBuffer& buf) : _InputArray(buf) {flags |= FIXED_SIZE|FIXED_TYPE;}
_OutputArray::_OutputArray(const GlTexture2D& tex) : _InputArray(tex) {flags |= FIXED_SIZE|FIXED_TYPE;}
bool _OutputArray::fixedSize() const
@ -1318,6 +1319,13 @@ void _OutputArray::create(Size _sz, int mtype, int i, bool allowTransposed, int
((gpu::GpuMat*)obj)->create(_sz, mtype);
return;
}
if( k == OPENGL_BUFFER && i < 0 && !allowTransposed && fixedDepthMask == 0 )
{
CV_Assert(!fixedSize() || ((GlBuffer*)obj)->size() == _sz);
CV_Assert(!fixedType() || ((GlBuffer*)obj)->type() == mtype);
((GlBuffer*)obj)->create(_sz, mtype);
return;
}
int sizes[] = {_sz.height, _sz.width};
create(2, sizes, mtype, i, allowTransposed, fixedDepthMask);
}
@ -1339,6 +1347,13 @@ void _OutputArray::create(int rows, int cols, int mtype, int i, bool allowTransp
((gpu::GpuMat*)obj)->create(rows, cols, mtype);
return;
}
if( k == OPENGL_BUFFER && i < 0 && !allowTransposed && fixedDepthMask == 0 )
{
CV_Assert(!fixedSize() || ((GlBuffer*)obj)->size() == Size(cols, rows));
CV_Assert(!fixedType() || ((GlBuffer*)obj)->type() == mtype);
((GlBuffer*)obj)->create(rows, cols, mtype);
return;
}
int sizes[] = {rows, cols};
create(2, sizes, mtype, i, allowTransposed, fixedDepthMask);
}
@ -1558,6 +1573,18 @@ void _OutputArray::release() const
return;
}
if( k == OPENGL_BUFFER )
{
((GlBuffer*)obj)->release();
return;
}
if( k == OPENGL_TEXTURE2D )
{
((GlTexture2D*)obj)->release();
return;
}
if( k == NONE )
return;
@ -1623,6 +1650,20 @@ gpu::GpuMat& _OutputArray::getGpuMatRef() const
return *(gpu::GpuMat*)obj;
}
GlBuffer& _OutputArray::getGlBufferRef() const
{
int k = kind();
CV_Assert( k == OPENGL_BUFFER );
return *(GlBuffer*)obj;
}
GlTexture2D& _OutputArray::getGlTexture2DRef() const
{
int k = kind();
CV_Assert( k == OPENGL_TEXTURE2D );
return *(GlTexture2D*)obj;
}
static _OutputArray _none;
OutputArray noArray() { return _none; }

2001
modules/core/src/opengl_interop.cpp
View File

File diff suppressed because it is too large Load Diff

16
modules/gpu/CMakeLists.txt
View File

@ -22,17 +22,14 @@ source_group("Device" FILES ${lib_device_hdrs})
source_group("Device\\Detail" FILES ${lib_device_hdrs_detail})
if (HAVE_CUDA)
file(GLOB_RECURSE ncv_srcs "src/nvidia/*.cpp")
file(GLOB_RECURSE ncv_srcs "src/nvidia/*.cpp" "src/nvidia/*.h*")
file(GLOB_RECURSE ncv_cuda "src/nvidia/*.cu")
file(GLOB_RECURSE ncv_hdrs "src/nvidia/*.hpp" "src/nvidia/*.h")
set(ncv_files ${ncv_srcs} ${ncv_hdrs} ${ncv_cuda})
set(ncv_files ${ncv_srcs} ${ncv_cuda})
source_group("Src\\NVidia" FILES ${ncv_files})
ocv_include_directories("src/nvidia" "src/nvidia/core" "src/nvidia/NPP_staging" ${CUDA_INCLUDE_DIRS})
ocv_warnings_disable(CMAKE_CXX_FLAGS -Wundef -Wmissing-declarations /wd4211 /wd4201 /wd4100 /wd4505 /wd4408)
string(REPLACE "-Wsign-promo" "" CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
#set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-keep")
#set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler;/EHsc-;")
if(MSVC)
@ -47,23 +44,18 @@ if (HAVE_CUDA)
ocv_cuda_compile(cuda_objs ${lib_cuda} ${ncv_cuda})
#CUDA_BUILD_CLEAN_TARGET()
set(cuda_link_libs ${CUDA_LIBRARIES} ${CUDA_npp_LIBRARY})
if(NOT APPLE)
unset(CUDA_nvcuvid_LIBRARY CACHE)
find_cuda_helper_libs(nvcuvid)
if(WITH_NVCUVID)
set(cuda_link_libs ${cuda_link_libs} ${CUDA_nvcuvid_LIBRARY})
endif()
if(WIN32)
unset(CUDA_nvcuvenc_LIBRARY CACHE)
find_cuda_helper_libs(nvcuvenc)
set(cuda_link_libs ${cuda_link_libs} ${CUDA_nvcuvenc_LIBRARY})
endif()
if(NOT APPLE AND WITH_FFMPEG)
if(WITH_FFMPEG)
set(cuda_link_libs ${cuda_link_libs} ${HIGHGUI_LIBRARIES})
endif()
else()

10
modules/gpu/app/nv_perf_test/CMakeLists.txt Normal file
View File

@ -0,0 +1,10 @@
cmake_minimum_required(VERSION 2.8.3)
project(nv_perf_test)
find_package(OpenCV REQUIRED)
include_directories(${OpenCV_INCLUDE_DIR})
add_executable(${PROJECT_NAME} main.cpp)
target_link_libraries(${PROJECT_NAME} ${OpenCV_LIBS})

BIN
modules/gpu/app/nv_perf_test/im1_1280x800.jpg Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 140 KiB

BIN
modules/gpu/app/nv_perf_test/im2_1280x800.jpg Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 140 KiB

489
modules/gpu/app/nv_perf_test/main.cpp Normal file
View File

@ -0,0 +1,489 @@
#include <cstdio>
#define HAVE_CUDA 1
#include <opencv2/core/core.hpp>
#include <opencv2/gpu/gpu.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/video/video.hpp>
#include <opencv2/legacy/legacy.hpp>
#include <opencv2/ts/ts.hpp>
#include <opencv2/ts/ts_perf.hpp>
static void printOsInfo()
{
#if defined _WIN32
# if defined _WIN64
printf("[----------]\n[ GPU INFO ] \tRun on OS Windows x64.\n[----------]\n"); fflush(stdout);
# else
printf("[----------]\n[ GPU INFO ] \tRun on OS Windows x32.\n[----------]\n"); fflush(stdout);
# endif
#elif defined linux
# if defined _LP64
printf("[----------]\n[ GPU INFO ] \tRun on OS Linux x64.\n[----------]\n"); fflush(stdout);
# else
printf("[----------]\n[ GPU INFO ] \tRun on OS Linux x32.\n[----------]\n"); fflush(stdout);
# endif
#elif defined __APPLE__
# if defined _LP64
printf("[----------]\n[ GPU INFO ] \tRun on OS Apple x64.\n[----------]\n"); fflush(stdout);
# else
printf("[----------]\n[ GPU INFO ] \tRun on OS Apple x32.\n[----------]\n"); fflush(stdout);
# endif
#endif
}
static void printCudaInfo()
{
const int deviceCount = cv::gpu::getCudaEnabledDeviceCount();
printf("[----------]\n"); fflush(stdout);
printf("[ GPU INFO ] \tCUDA device count:: %d.\n", deviceCount); fflush(stdout);
printf("[----------]\n"); fflush(stdout);
for (int i = 0; i < deviceCount; ++i)
{
cv::gpu::DeviceInfo info(i);
printf("[----------]\n"); fflush(stdout);
printf("[ DEVICE ] \t# %d %s.\n", i, info.name().c_str()); fflush(stdout);
printf("[ ] \tCompute capability: %d.%d\n", info.majorVersion(), info.minorVersion()); fflush(stdout);
printf("[ ] \tMulti Processor Count: %d\n", info.multiProcessorCount()); fflush(stdout);
printf("[ ] \tTotal memory: %d Mb\n", static_cast<int>(static_cast<int>(info.totalMemory() / 1024.0) / 1024.0)); fflush(stdout);
printf("[ ] \tFree memory: %d Mb\n", static_cast<int>(static_cast<int>(info.freeMemory() / 1024.0) / 1024.0)); fflush(stdout);
if (!info.isCompatible())
printf("[ GPU INFO ] \tThis device is NOT compatible with current GPU module build\n");
printf("[----------]\n"); fflush(stdout);
}
}
int main(int argc, char* argv[])
{
printOsInfo();
printCudaInfo();
perf::Regression::Init("nv_perf_test");
perf::TestBase::Init(argc, argv);
testing::InitGoogleTest(&argc, argv);
return RUN_ALL_TESTS();
}
#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
//////////////////////////////////////////////////////////
// HoughLinesP
DEF_PARAM_TEST_1(Image, std::string);
PERF_TEST_P(Image, HoughLinesP,
testing::Values(std::string("im1_1280x800.jpg")))
{
declare.time(30.0);
std::string fileName = GetParam();
const double rho = 1.0;
const double theta = 1.0;
const int threshold = 40;
const int minLineLenght = 20;
const int maxLineGap = 5;
cv::Mat image = cv::imread(fileName, cv::IMREAD_GRAYSCALE);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_image(image);
cv::gpu::GpuMat d_lines;
cv::gpu::HoughLinesBuf d_buf;
cv::gpu::HoughLinesP(d_image, d_lines, d_buf, rho, theta, minLineLenght, maxLineGap);
TEST_CYCLE()
{
cv::gpu::HoughLinesP(d_image, d_lines, d_buf, rho, theta, minLineLenght, maxLineGap);
}
}
else
{
cv::Mat mask;
cv::Canny(image, mask, 50, 100);
std::vector<cv::Vec4i> lines;
cv::HoughLinesP(mask, lines, rho, theta, threshold, minLineLenght, maxLineGap);
TEST_CYCLE()
{
cv::HoughLinesP(mask, lines, rho, theta, threshold, minLineLenght, maxLineGap);
}
}
SANITY_CHECK(0);
}
//////////////////////////////////////////////////////////
// GoodFeaturesToTrack
DEF_PARAM_TEST(Image_Depth, std::string, perf::MatDepth);
PERF_TEST_P(Image_Depth, GoodFeaturesToTrack,
testing::Combine(
testing::Values(std::string("im1_1280x800.jpg")),
testing::Values(CV_8U, CV_16U)
))
{
declare.time(60);
const std::string fileName = std::tr1::get<0>(GetParam());
const int depth = std::tr1::get<1>(GetParam());
const int maxCorners = 5000;
const double qualityLevel = 0.05;
const int minDistance = 5;
const int blockSize = 3;
const bool useHarrisDetector = true;
const double k = 0.05;
cv::Mat src = cv::imread(fileName, cv::IMREAD_GRAYSCALE);
if (src.empty())
FAIL() << "Unable to load source image [" << fileName << "]";
if (depth != CV_8U)
src.convertTo(src, depth);
cv::Mat mask(src.size(), CV_8UC1, cv::Scalar::all(1));
mask(cv::Rect(0, 0, 100, 100)).setTo(cv::Scalar::all(0));
if (PERF_RUN_GPU())
{
cv::gpu::GoodFeaturesToTrackDetector_GPU d_detector(maxCorners, qualityLevel, minDistance, blockSize, useHarrisDetector, k);
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_mask(mask);
cv::gpu::GpuMat d_pts;
d_detector(d_src, d_pts, d_mask);
TEST_CYCLE()
{
d_detector(d_src, d_pts, d_mask);
}
}
else
{
if (depth != CV_8U)
FAIL() << "Unsupported depth";
cv::Mat pts;
cv::goodFeaturesToTrack(src, pts, maxCorners, qualityLevel, minDistance, mask, blockSize, useHarrisDetector, k);
TEST_CYCLE()
{
cv::goodFeaturesToTrack(src, pts, maxCorners, qualityLevel, minDistance, mask, blockSize, useHarrisDetector, k);
}
}
SANITY_CHECK(0);
}
//////////////////////////////////////////////////////////
// OpticalFlowPyrLKSparse
typedef std::pair<std::string, std::string> string_pair;
DEF_PARAM_TEST(ImagePair_Depth_GraySource, string_pair, perf::MatDepth, bool);
PERF_TEST_P(ImagePair_Depth_GraySource, OpticalFlowPyrLKSparse,
testing::Combine(
testing::Values(string_pair("im1_1280x800.jpg", "im2_1280x800.jpg")),
testing::Values(CV_8U, CV_16U),
testing::Bool()
))
{
declare.time(60);
const string_pair fileNames = std::tr1::get<0>(GetParam());
const int depth = std::tr1::get<1>(GetParam());
const bool graySource = std::tr1::get<2>(GetParam());
// PyrLK params
const cv::Size winSize(15, 15);
const int maxLevel = 5;
const cv::TermCriteria criteria(cv::TermCriteria::COUNT + cv::TermCriteria::EPS, 30, 0.01);
// GoodFeaturesToTrack params
const int maxCorners = 5000;
const double qualityLevel = 0.05;
const int minDistance = 5;
const int blockSize = 3;
const bool useHarrisDetector = true;
const double k = 0.05;
cv::Mat src1 = cv::imread(fileNames.first, graySource ? cv::IMREAD_GRAYSCALE : cv::IMREAD_COLOR);
if (src1.empty())
FAIL() << "Unable to load source image [" << fileNames.first << "]";
cv::Mat src2 = cv::imread(fileNames.second, graySource ? cv::IMREAD_GRAYSCALE : cv::IMREAD_COLOR);
if (src2.empty())
FAIL() << "Unable to load source image [" << fileNames.second << "]";
cv::Mat gray_src;
if (graySource)
gray_src = src1;
else
cv::cvtColor(src1, gray_src, cv::COLOR_BGR2GRAY);
cv::Mat pts;
cv::goodFeaturesToTrack(gray_src, pts, maxCorners, qualityLevel, minDistance, cv::noArray(), blockSize, useHarrisDetector, k);
if (depth != CV_8U)
{
src1.convertTo(src1, depth);
src2.convertTo(src2, depth);
}
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src1(src1);
cv::gpu::GpuMat d_src2(src2);
cv::gpu::GpuMat d_pts(pts.reshape(2, 1));
cv::gpu::GpuMat d_nextPts;
cv::gpu::GpuMat d_status;
cv::gpu::PyrLKOpticalFlow d_pyrLK;
d_pyrLK.winSize = winSize;
d_pyrLK.maxLevel = maxLevel;
d_pyrLK.iters = criteria.maxCount;
d_pyrLK.useInitialFlow = false;
d_pyrLK.sparse(d_src1, d_src2, d_pts, d_nextPts, d_status);
TEST_CYCLE()
{
d_pyrLK.sparse(d_src1, d_src2, d_pts, d_nextPts, d_status);
}
}
else
{
if (depth != CV_8U)
FAIL() << "Unsupported depth";
cv::Mat nextPts;
cv::Mat status;
cv::calcOpticalFlowPyrLK(src1, src2, pts, nextPts, status, cv::noArray(), winSize, maxLevel, criteria);
TEST_CYCLE()
{
cv::calcOpticalFlowPyrLK(src1, src2, pts, nextPts, status, cv::noArray(), winSize, maxLevel, criteria);
}
}
SANITY_CHECK(0);
}
//////////////////////////////////////////////////////////
// OpticalFlowFarneback
DEF_PARAM_TEST(ImagePair_Depth, string_pair, perf::MatDepth);
PERF_TEST_P(ImagePair_Depth, OpticalFlowFarneback,
testing::Combine(
testing::Values(string_pair("im1_1280x800.jpg", "im2_1280x800.jpg")),
testing::Values(CV_8U, CV_16U)
))
{
declare.time(500);
const string_pair fileNames = std::tr1::get<0>(GetParam());
const int depth = std::tr1::get<1>(GetParam());
const double pyrScale = 0.5;
const int numLevels = 6;
const int winSize = 7;
const int numIters = 15;
const int polyN = 7;
const double polySigma = 1.5;
const int flags = cv::OPTFLOW_USE_INITIAL_FLOW;
cv::Mat src1 = cv::imread(fileNames.first, cv::IMREAD_GRAYSCALE);
if (src1.empty())
FAIL() << "Unable to load source image [" << fileNames.first << "]";
cv::Mat src2 = cv::imread(fileNames.second, cv::IMREAD_GRAYSCALE);
if (src2.empty())
FAIL() << "Unable to load source image [" << fileNames.second << "]";
if (depth != CV_8U)
{
src1.convertTo(src1, depth);
src2.convertTo(src2, depth);
}
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src1(src1);
cv::gpu::GpuMat d_src2(src2);
cv::gpu::GpuMat d_u(src1.size(), CV_32FC1, cv::Scalar::all(0));
cv::gpu::GpuMat d_v(src1.size(), CV_32FC1, cv::Scalar::all(0));
cv::gpu::FarnebackOpticalFlow d_farneback;
d_farneback.pyrScale = pyrScale;
d_farneback.numLevels = numLevels;
d_farneback.winSize = winSize;
d_farneback.numIters = numIters;
d_farneback.polyN = polyN;
d_farneback.polySigma = polySigma;
d_farneback.flags = flags;
d_farneback(d_src1, d_src2, d_u, d_v);
TEST_CYCLE_N(10)
{
d_farneback(d_src1, d_src2, d_u, d_v);
}
}
else
{
if (depth != CV_8U)
FAIL() << "Unsupported depth";
cv::Mat flow(src1.size(), CV_32FC2, cv::Scalar::all(0));
cv::calcOpticalFlowFarneback(src1, src2, flow, pyrScale, numLevels, winSize, numIters, polyN, polySigma, flags);
TEST_CYCLE_N(10)
{
cv::calcOpticalFlowFarneback(src1, src2, flow, pyrScale, numLevels, winSize, numIters, polyN, polySigma, flags);
}
}
SANITY_CHECK(0);
}
//////////////////////////////////////////////////////////
// OpticalFlowBM
void calcOpticalFlowBM(const cv::Mat& prev, const cv::Mat& curr,
cv::Size bSize, cv::Size shiftSize, cv::Size maxRange, int usePrevious,
cv::Mat& velx, cv::Mat& vely)
{
cv::Size sz((curr.cols - bSize.width + shiftSize.width)/shiftSize.width, (curr.rows - bSize.height + shiftSize.height)/shiftSize.height);
velx.create(sz, CV_32FC1);
vely.create(sz, CV_32FC1);
CvMat cvprev = prev;
CvMat cvcurr = curr;
CvMat cvvelx = velx;
CvMat cvvely = vely;
cvCalcOpticalFlowBM(&cvprev, &cvcurr, bSize, shiftSize, maxRange, usePrevious, &cvvelx, &cvvely);
}
DEF_PARAM_TEST(ImagePair_BlockSize_ShiftSize_MaxRange, string_pair, cv::Size, cv::Size, cv::Size);
PERF_TEST_P(ImagePair_BlockSize_ShiftSize_MaxRange, OpticalFlowBM,
testing::Combine(
testing::Values(string_pair("im1_1280x800.jpg", "im2_1280x800.jpg")),
testing::Values(cv::Size(16, 16)),
testing::Values(cv::Size(2, 2)),
testing::Values(cv::Size(16, 16))
))
{
declare.time(1000);
const string_pair fileNames = std::tr1::get<0>(GetParam());
const cv::Size block_size = std::tr1::get<1>(GetParam());
const cv::Size shift_size = std::tr1::get<2>(GetParam());
const cv::Size max_range = std::tr1::get<3>(GetParam());
cv::Mat src1 = cv::imread(fileNames.first, cv::IMREAD_GRAYSCALE);
if (src1.empty())
FAIL() << "Unable to load source image [" << fileNames.first << "]";
cv::Mat src2 = cv::imread(fileNames.second, cv::IMREAD_GRAYSCALE);
if (src2.empty())
FAIL() << "Unable to load source image [" << fileNames.second << "]";
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src1(src1);
cv::gpu::GpuMat d_src2(src2);
cv::gpu::GpuMat d_velx, d_vely, buf;
cv::gpu::calcOpticalFlowBM(d_src1, d_src2, block_size, shift_size, max_range, false, d_velx, d_vely, buf);
TEST_CYCLE_N(10)
{
cv::gpu::calcOpticalFlowBM(d_src1, d_src2, block_size, shift_size, max_range, false, d_velx, d_vely, buf);
}
}
else
{
cv::Mat velx, vely;
calcOpticalFlowBM(src1, src2, block_size, shift_size, max_range, false, velx, vely);
TEST_CYCLE_N(10)
{
calcOpticalFlowBM(src1, src2, block_size, shift_size, max_range, false, velx, vely);
}
}
SANITY_CHECK(0);
}
PERF_TEST_P(ImagePair_BlockSize_ShiftSize_MaxRange, FastOpticalFlowBM,
testing::Combine(
testing::Values(string_pair("im1_1280x800.jpg", "im2_1280x800.jpg")),
testing::Values(cv::Size(16, 16)),
testing::Values(cv::Size(1, 1)),
testing::Values(cv::Size(16, 16))
))
{
declare.time(1000);
const string_pair fileNames = std::tr1::get<0>(GetParam());
const cv::Size block_size = std::tr1::get<1>(GetParam());
const cv::Size shift_size = std::tr1::get<2>(GetParam());
const cv::Size max_range = std::tr1::get<3>(GetParam());
cv::Mat src1 = cv::imread(fileNames.first, cv::IMREAD_GRAYSCALE);
if (src1.empty())
FAIL() << "Unable to load source image [" << fileNames.first << "]";
cv::Mat src2 = cv::imread(fileNames.second, cv::IMREAD_GRAYSCALE);
if (src2.empty())
FAIL() << "Unable to load source image [" << fileNames.second << "]";
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_src1(src1);
cv::gpu::GpuMat d_src2(src2);
cv::gpu::GpuMat d_velx, d_vely;
cv::gpu::FastOpticalFlowBM fastBM;
fastBM(d_src1, d_src2, d_velx, d_vely, max_range.width, block_size.width);
TEST_CYCLE_N(10)
{
fastBM(d_src1, d_src2, d_velx, d_vely, max_range.width, block_size.width);
}
}
else
{
cv::Mat velx, vely;
calcOpticalFlowBM(src1, src2, block_size, shift_size, max_range, false, velx, vely);
TEST_CYCLE_N(10)
{
calcOpticalFlowBM(src1, src2, block_size, shift_size, max_range, false, velx, vely);
}
}
SANITY_CHECK(0);
}

85
modules/gpu/doc/object_detection.rst
View File

@ -199,6 +199,91 @@ Returns block descriptors computed for the whole image.
The function is mainly used to learn the classifier.
Soft Cascade Classifier
==========================
Soft Cascade Classifier for Object Detection
----------------------------------------------------------
Cascade detectors have been shown to operate extremely rapidly, with high accuracy, and have important applications in different spheres. The initial goal for this cascade implementation was the fast and accurate pedestrian detector but it also useful in general. Soft cascade is trained with AdaBoost. But instead of training sequence of stages, the soft cascade is trained as a one long stage of T weak classifiers. Soft cascade is formulated as follows:
.. math::
\texttt{H}(x) = \sum _{\texttt{t}=1..\texttt{T}} {\texttt{s}_t(x)}
where :math:`\texttt{s}_t(x) = \alpha_t\texttt{h}_t(x)` are the set of thresholded weak classifiers selected during AdaBoost training scaled by the associated weights. Let
.. math::
\texttt{H}_t(x) = \sum _{\texttt{i}=1..\texttt{t}} {\texttt{s}_i(x)}
be the partial sum of sample responses before :math:`t`-the weak classifier will be applied. The funtcion :math:`\texttt{H}_t(x)` of :math:`t` for sample :math:`x` named *sample trace*.
After each weak classifier evaluation, the sample trace at the point :math:`t` is compared with the rejection threshold :math:`r_t`. The sequence of :math:`r_t` named *rejection trace*.
The sample has been rejected if it fall rejection threshold. So stageless cascade allows to reject not-object sample as soon as possible. Another meaning of the sample trace is a confidence with that sample recognized as desired object. At each :math:`t` that confidence depend on all previous weak classifier. This feature of soft cascade is resulted in more accurate detection. The original formulation of soft cascade can be found in [BJ05]_.
.. [BJ05] Lubomir Bourdev and Jonathan Brandt. tRobust Object Detection Via Soft Cascade. IEEE CVPR, 2005.
.. [BMTG12] Rodrigo Benenson, Markus Mathias, Radu Timofte and Luc Van Gool. Pedestrian detection at 100 frames per second. IEEE CVPR, 2012.
gpu::SCascade
-----------------------------------------------
.. ocv:class:: gpu::SCascade : public Algorithm
Implementation of soft (stageless) cascaded detector. ::
class CV_EXPORTS SCascade : public Algorithm
{
struct CV_EXPORTS Detection
{
ushort x;
ushort y;
ushort w;
ushort h;
float confidence;
int kind;
enum {PEDESTRIAN = 0};
};
SCascade(const double minScale = 0.4, const double maxScale = 5., const int scales = 55, const int rejfactor = 1);
virtual ~SCascade();
virtual bool load(const FileNode& fn);
virtual void detect(InputArray image, InputArray rois, OutputArray objects, Stream& stream = Stream::Null()) const;
virtual void genRoi(InputArray roi, OutputArray mask, Stream& stream = Stream::Null()) const;
};
gpu::SCascade::~SCascade
---------------------------
Destructor for SCascade.
.. ocv:function:: gpu::SCascade::~SCascade()
gpu::SCascade::load
--------------------------
Load cascade from FileNode.
.. ocv:function:: bool gpu::SCascade::load(const FileNode& fn)
:param fn: File node from which the soft cascade are read.
gpu::SCascade::detect
--------------------------
Apply cascade to an input frame and return the vector of Decection objcts.
.. ocv:function:: void gpu::SCascade::detect(InputArray image, InputArray rois, OutputArray objects, Stream& stream = Stream::Null()) const
:param image: a frame on which detector will be applied.
:param rois: a regions of interests mask generated by genRoi. Only the objects that fall into one of the regions will be returned.
:param objects: an output array of Detections represented as GpuMat of detections (SCascade::Detection). The first element of the matrix is actually a count of detections.
:param stream: a high-level CUDA stream abstraction used for asynchronous execution.
gpu::CascadeClassifier_GPU
--------------------------

0
modules/gpu/src/opencv2/gpu/device/block.hpp → modules/gpu/include/opencv2/gpu/device/block.hpp
View File

12
modules/gpu/include/opencv2/gpu/device/common.hpp
View File

@ -85,8 +85,6 @@ static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int
cv::gpu::error(cudaGetErrorString(err), file, line, func);
}
#ifdef __CUDACC__
namespace cv { namespace gpu
{
__host__ __device__ __forceinline__ int divUp(int total, int grain)
@ -96,19 +94,25 @@ namespace cv { namespace gpu
namespace device
{
using cv::gpu::divUp;
#ifdef __CUDACC__
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef signed char schar;
typedef unsigned int uint;
#ifdef WIN32
typedef unsigned int uint;
#endif
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 // __CUDACC__
#endif // __OPENCV_GPU_COMMON_HPP__

2
modules/gpu/include/opencv2/gpu/device/detail/color_detail.hpp
View File

@ -807,9 +807,9 @@ namespace cv { namespace gpu { namespace device
template <int bidx, typename T, typename D> static __device__ __forceinline__ void RGB2XYZConvert(const T* src, D& dst)
{
dst.z = saturate_cast<T>(CV_DESCALE(src[bidx^2] * c_RGB2XYZ_D65i[6] + src[1] * c_RGB2XYZ_D65i[7] + src[bidx] * c_RGB2XYZ_D65i[8], xyz_shift));
dst.x = saturate_cast<T>(CV_DESCALE(src[bidx^2] * c_RGB2XYZ_D65i[0] + src[1] * c_RGB2XYZ_D65i[1] + src[bidx] * c_RGB2XYZ_D65i[2], xyz_shift));
dst.y = saturate_cast<T>(CV_DESCALE(src[bidx^2] * c_RGB2XYZ_D65i[3] + src[1] * c_RGB2XYZ_D65i[4] + src[bidx] * c_RGB2XYZ_D65i[5], xyz_shift));
dst.z = saturate_cast<T>(CV_DESCALE(src[bidx^2] * c_RGB2XYZ_D65i[6] + src[1] * c_RGB2XYZ_D65i[7] + src[bidx] * c_RGB2XYZ_D65i[8], xyz_shift));
}
template <int bidx> static __device__ __forceinline__ uint RGB2XYZConvert(uint src)

361
modules/gpu/include/opencv2/gpu/device/detail/reduce.hpp Normal file
View File

@ -0,0 +1,361 @@
/*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_REDUCE_DETAIL_HPP__
#define __OPENCV_GPU_REDUCE_DETAIL_HPP__
#include <thrust/tuple.h>
#include "../warp.hpp"
#include "../warp_shuffle.hpp"
namespace cv { namespace gpu { namespace device
{
namespace reduce_detail
{
template <typename T> struct GetType;
template <typename T> struct GetType<T*>
{
typedef T type;
};
template <typename T> struct GetType<volatile T*>
{
typedef T type;
};
template <typename T> struct GetType<T&>
{
typedef T type;
};
template <unsigned int I, unsigned int N>
struct For
{
template <class PointerTuple, class ValTuple>
static __device__ void loadToSmem(const PointerTuple& smem, const ValTuple& val, unsigned int tid)
{
thrust::get<I>(smem)[tid] = thrust::get<I>(val);
For<I + 1, N>::loadToSmem(smem, val, tid);
}
template <class PointerTuple, class ValTuple>
static __device__ void loadFromSmem(const PointerTuple& smem, const ValTuple& val, unsigned int tid)
{
thrust::get<I>(val) = thrust::get<I>(smem)[tid];
For<I + 1, N>::loadFromSmem(smem, val, tid);
}
template <class PointerTuple, class ValTuple, class OpTuple>
static __device__ void merge(const PointerTuple& smem, const ValTuple& val, unsigned int tid, unsigned int delta, const OpTuple& op)
{
typename GetType<typename thrust::tuple_element<I, PointerTuple>::type>::type reg = thrust::get<I>(smem)[tid + delta];
thrust::get<I>(smem)[tid] = thrust::get<I>(val) = thrust::get<I>(op)(thrust::get<I>(val), reg);
For<I + 1, N>::merge(smem, val, tid, delta, op);
}
template <class ValTuple, class OpTuple>
static __device__ void mergeShfl(const ValTuple& val, unsigned int delta, unsigned int width, const OpTuple& op)
{
typename GetType<typename thrust::tuple_element<I, ValTuple>::type>::type reg = shfl_down(thrust::get<I>(val), delta, width);
thrust::get<I>(val) = thrust::get<I>(op)(thrust::get<I>(val), reg);
For<I + 1, N>::mergeShfl(val, delta, width, op);
}
};
template <unsigned int N>
struct For<N, N>
{
template <class PointerTuple, class ValTuple>
static __device__ void loadToSmem(const PointerTuple&, const ValTuple&, unsigned int)
{
}
template <class PointerTuple, class ValTuple>
static __device__ void loadFromSmem(const PointerTuple&, const ValTuple&, unsigned int)
{
}
template <class PointerTuple, class ValTuple, class OpTuple>
static __device__ void merge(const PointerTuple&, const ValTuple&, unsigned int, unsigned int, const OpTuple&)
{
}
template <class ValTuple, class OpTuple>
static __device__ void mergeShfl(const ValTuple&, unsigned int, unsigned int, const OpTuple&)
{
}
};
template <typename T>
__device__ __forceinline__ void loadToSmem(volatile T* smem, T& val, unsigned int tid)
{
smem[tid] = val;
}
template <typename T>
__device__ __forceinline__ void loadFromSmem(volatile T* smem, T& val, unsigned int tid)
{
val = smem[tid];
}
template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9>
__device__ __forceinline__ void loadToSmem(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
unsigned int tid)
{
For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::loadToSmem(smem, val, tid);
}
template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9>
__device__ __forceinline__ void loadFromSmem(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
unsigned int tid)
{
For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::loadFromSmem(smem, val, tid);
}
template <typename T, class Op>
__device__ __forceinline__ void merge(volatile T* smem, T& val, unsigned int tid, unsigned int delta, const Op& op)
{
T reg = smem[tid + delta];
smem[tid] = val = op(val, reg);
}
template <typename T, class Op>
__device__ __forceinline__ void mergeShfl(T& val, unsigned int delta, unsigned int width, const Op& op)
{
T reg = shfl_down(val, delta, width);
val = op(val, reg);
}
template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
__device__ __forceinline__ void merge(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
unsigned int tid,
unsigned int delta,
const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
{
For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::merge(smem, val, tid, delta, op);
}
template <typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
__device__ __forceinline__ void mergeShfl(const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
unsigned int delta,
unsigned int width,
const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
{
For<0, thrust::tuple_size<thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9> >::value>::mergeShfl(val, delta, width, op);
}
template <unsigned int N> struct Generic
{
template <typename Pointer, typename Reference, class Op>
static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
{
loadToSmem(smem, val, tid);
if (N >= 32)
__syncthreads();
if (N >= 2048)
{
if (tid < 1024)
merge(smem, val, tid, 1024, op);
__syncthreads();
}
if (N >= 1024)
{
if (tid < 512)
merge(smem, val, tid, 512, op);
__syncthreads();
}
if (N >= 512)
{
if (tid < 256)
merge(smem, val, tid, 256, op);
__syncthreads();
}
if (N >= 256)
{
if (tid < 128)
merge(smem, val, tid, 128, op);
__syncthreads();
}
if (N >= 128)
{
if (tid < 64)
merge(smem, val, tid, 64, op);
__syncthreads();
}
if (N >= 64)
{
if (tid < 32)
merge(smem, val, tid, 32, op);
}
if (tid < 16)
{
merge(smem, val, tid, 16, op);
merge(smem, val, tid, 8, op);
merge(smem, val, tid, 4, op);
merge(smem, val, tid, 2, op);
merge(smem, val, tid, 1, op);
}
}
};
template <unsigned int I, typename Pointer, typename Reference, class Op>
struct Unroll
{
static __device__ void loopShfl(Reference val, Op op, unsigned int N)
{
mergeShfl(val, I, N, op);
Unroll<I / 2, Pointer, Reference, Op>::loopShfl(val, op, N);
}
static __device__ void loop(Pointer smem, Reference val, unsigned int tid, Op op)
{
merge(smem, val, tid, I, op);
Unroll<I / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
}
};
template <typename Pointer, typename Reference, class Op>
struct Unroll<0, Pointer, Reference, Op>
{
static __device__ void loopShfl(Reference, Op, unsigned int)
{
}
static __device__ void loop(Pointer, Reference, unsigned int, Op)
{
}
};
template <unsigned int N> struct WarpOptimized
{
template <typename Pointer, typename Reference, class Op>
static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
{
#if __CUDA_ARCH__ >= 300
(void) smem;
(void) tid;
Unroll<N / 2, Pointer, Reference, Op>::loopShfl(val, op, N);
#else
loadToSmem(smem, val, tid);
if (tid < N / 2)
Unroll<N / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
#endif
}
};
template <unsigned int N> struct GenericOptimized32
{
enum { M = N / 32 };
template <typename Pointer, typename Reference, class Op>
static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
{
const unsigned int laneId = Warp::laneId();
#if __CUDA_ARCH__ >= 300
Unroll<16, Pointer, Reference, Op>::loopShfl(val, op, warpSize);
if (laneId == 0)
loadToSmem(smem, val, tid / 32);
#else
loadToSmem(smem, val, tid);
if (laneId < 16)
Unroll<16, Pointer, Reference, Op>::loop(smem, val, tid, op);
__syncthreads();
if (laneId == 0)
loadToSmem(smem, val, tid / 32);
#endif
__syncthreads();
loadFromSmem(smem, val, tid);
if (tid < 32)
{
#if __CUDA_ARCH__ >= 300
Unroll<M / 2, Pointer, Reference, Op>::loopShfl(val, op, M);
#else
Unroll<M / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
#endif
}
}
};
template <bool val, class T1, class T2> struct StaticIf;
template <class T1, class T2> struct StaticIf<true, T1, T2>
{
typedef T1 type;
};
template <class T1, class T2> struct StaticIf<false, T1, T2>
{
typedef T2 type;
};
template <unsigned int N> struct IsPowerOf2
{
enum { value = ((N != 0) && !(N & (N - 1))) };
};
template <unsigned int N> struct Dispatcher
{
typedef typename StaticIf<
(N <= 32) && IsPowerOf2<N>::value,
WarpOptimized<N>,
typename StaticIf<
(N <= 1024) && IsPowerOf2<N>::value,
GenericOptimized32<N>,
Generic<N>
>::type
>::type reductor;
};
}
}}}
#endif // __OPENCV_GPU_REDUCE_DETAIL_HPP__

498
modules/gpu/include/opencv2/gpu/device/detail/reduce_key_val.hpp Normal file
View File

@ -0,0 +1,498 @@
/*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_PRED_VAL_REDUCE_DETAIL_HPP__
#define __OPENCV_GPU_PRED_VAL_REDUCE_DETAIL_HPP__
#include <thrust/tuple.h>
#include "../warp.hpp"
#include "../warp_shuffle.hpp"
namespace cv { namespace gpu { namespace device
{
namespace reduce_key_val_detail
{
template <typename T> struct GetType;
template <typename T> struct GetType<T*>
{
typedef T type;
};
template <typename T> struct GetType<volatile T*>
{
typedef T type;
};
template <typename T> struct GetType<T&>
{
typedef T type;
};
template <unsigned int I, unsigned int N>
struct For
{
template <class PointerTuple, class ReferenceTuple>
static __device__ void loadToSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid)
{
thrust::get<I>(smem)[tid] = thrust::get<I>(data);
For<I + 1, N>::loadToSmem(smem, data, tid);
}
template <class PointerTuple, class ReferenceTuple>
static __device__ void loadFromSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid)
{
thrust::get<I>(data) = thrust::get<I>(smem)[tid];
For<I + 1, N>::loadFromSmem(smem, data, tid);
}
template <class ReferenceTuple>
static __device__ void copyShfl(const ReferenceTuple& val, unsigned int delta, int width)
{
thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width);
For<I + 1, N>::copyShfl(val, delta, width);
}
template <class PointerTuple, class ReferenceTuple>
static __device__ void copy(const PointerTuple& svals, const ReferenceTuple& val, unsigned int tid, unsigned int delta)
{
thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta];
For<I + 1, N>::copy(svals, val, tid, delta);
}
template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple>
static __device__ void mergeShfl(const KeyReferenceTuple& key, const ValReferenceTuple& val, const CmpTuple& cmp, unsigned int delta, int width)
{
typename GetType<typename thrust::tuple_element<I, KeyReferenceTuple>::type>::type reg = shfl_down(thrust::get<I>(key), delta, width);
if (thrust::get<I>(cmp)(reg, thrust::get<I>(key)))
{
thrust::get<I>(key) = reg;
thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width);
}
For<I + 1, N>::mergeShfl(key, val, cmp, delta, width);
}
template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple>
static __device__ void merge(const KeyPointerTuple& skeys, const KeyReferenceTuple& key,
const ValPointerTuple& svals, const ValReferenceTuple& val,
const CmpTuple& cmp,
unsigned int tid, unsigned int delta)
{
typename GetType<typename thrust::tuple_element<I, KeyPointerTuple>::type>::type reg = thrust::get<I>(skeys)[tid + delta];
if (thrust::get<I>(cmp)(reg, thrust::get<I>(key)))
{
thrust::get<I>(skeys)[tid] = thrust::get<I>(key) = reg;
thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta];
}
For<I + 1, N>::merge(skeys, key, svals, val, cmp, tid, delta);
}
};
template <unsigned int N>
struct For<N, N>
{
template <class PointerTuple, class ReferenceTuple>
static __device__ void loadToSmem(const PointerTuple&, const ReferenceTuple&, unsigned int)
{
}
template <class PointerTuple, class ReferenceTuple>
static __device__ void loadFromSmem(const PointerTuple&, const ReferenceTuple&, unsigned int)
{
}
template <class ReferenceTuple>
static __device__ void copyShfl(const ReferenceTuple&, unsigned int, int)
{
}
template <class PointerTuple, class ReferenceTuple>
static __device__ void copy(const PointerTuple&, const ReferenceTuple&, unsigned int, unsigned int)
{
}
template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple>
static __device__ void mergeShfl(const KeyReferenceTuple&, const ValReferenceTuple&, const CmpTuple&, unsigned int, int)
{
}
template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple>
static __device__ void merge(const KeyPointerTuple&, const KeyReferenceTuple&,
const ValPointerTuple&, const ValReferenceTuple&,
const CmpTuple&,
unsigned int, unsigned int)
{
}
};
//////////////////////////////////////////////////////
// loadToSmem
template <typename T>
__device__ __forceinline__ void loadToSmem(volatile T* smem, T& data, unsigned int tid)
{
smem[tid] = data;
}
template <typename T>
__device__ __forceinline__ void loadFromSmem(volatile T* smem, T& data, unsigned int tid)
{
data = smem[tid];
}
template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
__device__ __forceinline__ void loadToSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data,
unsigned int tid)
{
For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadToSmem(smem, data, tid);
}
template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
__device__ __forceinline__ void loadFromSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data,
unsigned int tid)
{
For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadFromSmem(smem, data, tid);
}
//////////////////////////////////////////////////////
// copyVals
template <typename V>
__device__ __forceinline__ void copyValsShfl(V& val, unsigned int delta, int width)
{
val = shfl_down(val, delta, width);
}
template <typename V>
__device__ __forceinline__ void copyVals(volatile V* svals, V& val, unsigned int tid, unsigned int delta)
{
svals[tid] = val = svals[tid + delta];
}
template <typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
__device__ __forceinline__ void copyValsShfl(const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
unsigned int delta,
int width)
{
For<0, thrust::tuple_size<thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9> >::value>::copyShfl(val, delta, width);
}
template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
__device__ __forceinline__ void copyVals(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
unsigned int tid, unsigned int delta)
{
For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::copy(svals, val, tid, delta);
}
//////////////////////////////////////////////////////
// merge
template <typename K, typename V, class Cmp>
__device__ __forceinline__ void mergeShfl(K& key, V& val, const Cmp& cmp, unsigned int delta, int width)
{
K reg = shfl_down(key, delta, width);
if (cmp(reg, key))
{
key = reg;
copyValsShfl(val, delta, width);
}
}
template <typename K, typename V, class Cmp>
__device__ __forceinline__ void merge(volatile K* skeys, K& key, volatile V* svals, V& val, const Cmp& cmp, unsigned int tid, unsigned int delta)
{
K reg = skeys[tid + delta];
if (cmp(reg, key))
{
skeys[tid] = key = reg;
copyVals(svals, val, tid, delta);
}
}
template <typename K,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
class Cmp>
__device__ __forceinline__ void mergeShfl(K& key,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
const Cmp& cmp,
unsigned int delta, int width)
{
K reg = shfl_down(key, delta, width);
if (cmp(reg, key))
{
key = reg;
copyValsShfl(val, delta, width);
}
}
template <typename K,
typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
class Cmp>
__device__ __forceinline__ void merge(volatile K* skeys, K& key,
const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
const Cmp& cmp, unsigned int tid, unsigned int delta)
{
K reg = skeys[tid + delta];
if (cmp(reg, key))
{
skeys[tid] = key = reg;
copyVals(svals, val, tid, delta);
}
}
template <typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
__device__ __forceinline__ void mergeShfl(const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp,
unsigned int delta, int width)
{
For<0, thrust::tuple_size<thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9> >::value>::mergeShfl(key, val, cmp, delta, width);
}
template <typename KP0, typename KP1, typename KP2, typename KP3, typename KP4, typename KP5, typename KP6, typename KP7, typename KP8, typename KP9,
typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
__device__ __forceinline__ void merge(const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>& skeys,
const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp,
unsigned int tid, unsigned int delta)
{
For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::merge(skeys, key, svals, val, cmp, tid, delta);
}
//////////////////////////////////////////////////////
// Generic
template <unsigned int N> struct Generic
{
template <class KP, class KR, class VP, class VR, class Cmp>
static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
{
loadToSmem(skeys, key, tid);
loadValsToSmem(svals, val, tid);
if (N >= 32)
__syncthreads();
if (N >= 2048)
{
if (tid < 1024)
merge(skeys, key, svals, val, cmp, tid, 1024);
__syncthreads();
}
if (N >= 1024)
{
if (tid < 512)
merge(skeys, key, svals, val, cmp, tid, 512);
__syncthreads();
}
if (N >= 512)
{
if (tid < 256)
merge(skeys, key, svals, val, cmp, tid, 256);
__syncthreads();
}
if (N >= 256)
{
if (tid < 128)
merge(skeys, key, svals, val, cmp, tid, 128);
__syncthreads();
}
if (N >= 128)
{
if (tid < 64)
merge(skeys, key, svals, val, cmp, tid, 64);
__syncthreads();
}
if (N >= 64)
{
if (tid < 32)
merge(skeys, key, svals, val, cmp, tid, 32);
}
if (tid < 16)
{
merge(skeys, key, svals, val, cmp, tid, 16);
merge(skeys, key, svals, val, cmp, tid, 8);
merge(skeys, key, svals, val, cmp, tid, 4);
merge(skeys, key, svals, val, cmp, tid, 2);
merge(skeys, key, svals, val, cmp, tid, 1);
}
}
};
template <unsigned int I, class KP, class KR, class VP, class VR, class Cmp>
struct Unroll
{
static __device__ void loopShfl(KR key, VR val, Cmp cmp, unsigned int N)
{
mergeShfl(key, val, cmp, I, N);
Unroll<I / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N);
}
static __device__ void loop(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
{
merge(skeys, key, svals, val, cmp, tid, I);
Unroll<I / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
}
};
template <class KP, class KR, class VP, class VR, class Cmp>
struct Unroll<0, KP, KR, VP, VR, Cmp>
{
static __device__ void loopShfl(KR, VR, Cmp, unsigned int)
{
}
static __device__ void loop(KP, KR, VP, VR, unsigned int, Cmp)
{
}
};
template <unsigned int N> struct WarpOptimized
{
template <class KP, class KR, class VP, class VR, class Cmp>
static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
{
#if 0 // __CUDA_ARCH__ >= 300
(void) skeys;
(void) svals;
(void) tid;
Unroll<N / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N);
#else
loadToSmem(skeys, key, tid);
loadToSmem(svals, val, tid);
if (tid < N / 2)
Unroll<N / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
#endif
}
};
template <unsigned int N> struct GenericOptimized32
{
enum { M = N / 32 };
template <class KP, class KR, class VP, class VR, class Cmp>
static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
{
const unsigned int laneId = Warp::laneId();
#if 0 // __CUDA_ARCH__ >= 300
Unroll<16, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, warpSize);
if (laneId == 0)
{
loadToSmem(skeys, key, tid / 32);
loadToSmem(svals, val, tid / 32);
}
#else
loadToSmem(skeys, key, tid);
loadToSmem(svals, val, tid);
if (laneId < 16)
Unroll<16, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
__syncthreads();
if (laneId == 0)
{
loadToSmem(skeys, key, tid / 32);
loadToSmem(svals, val, tid / 32);
}
#endif
__syncthreads();
loadFromSmem(skeys, key, tid);
if (tid < 32)
{
#if 0 // __CUDA_ARCH__ >= 300
loadFromSmem(svals, val, tid);
Unroll<M / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, M);
#else
Unroll<M / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
#endif
}
}
};
template <bool val, class T1, class T2> struct StaticIf;
template <class T1, class T2> struct StaticIf<true, T1, T2>
{
typedef T1 type;
};
template <class T1, class T2> struct StaticIf<false, T1, T2>
{
typedef T2 type;
};
template <unsigned int N> struct IsPowerOf2
{
enum { value = ((N != 0) && !(N & (N - 1))) };
};
template <unsigned int N> struct Dispatcher
{
typedef typename StaticIf<
(N <= 32) && IsPowerOf2<N>::value,
WarpOptimized<N>,
typename StaticIf<
(N <= 1024) && IsPowerOf2<N>::value,
GenericOptimized32<N>,
Generic<N>
>::type
>::type reductor;
};
}
}}}
#endif // __OPENCV_GPU_PRED_VAL_REDUCE_DETAIL_HPP__

841
modules/gpu/include/opencv2/gpu/device/detail/reduction_detail.hpp
View File

@ -1,841 +0,0 @@
/*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_REDUCTION_DETAIL_HPP__
#define __OPENCV_GPU_REDUCTION_DETAIL_HPP__
namespace cv { namespace gpu { namespace device
{
namespace utility_detail
{
///////////////////////////////////////////////////////////////////////////////
// Reductor
template <int n> struct WarpReductor
{
template <typename T, typename Op> static __device__ __forceinline__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
if (tid < n)
data[tid] = partial_reduction;
if (n > 32) __syncthreads();
if (n > 32)
{
if (tid < n - 32)
data[tid] = partial_reduction = op(partial_reduction, data[tid + 32]);
if (tid < 16)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 16]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1]);
}
}
else if (n > 16)
{
if (tid < n - 16)
data[tid] = partial_reduction = op(partial_reduction, data[tid + 16]);
if (tid < 8)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1]);
}
}
else if (n > 8)
{
if (tid < n - 8)
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8]);
if (tid < 4)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1]);
}
}
else if (n > 4)
{
if (tid < n - 4)
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4]);
if (tid < 2)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1]);
}
}
else if (n > 2)
{
if (tid < n - 2)
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2]);
if (tid < 2)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1]);
}
}
}
};
template <> struct WarpReductor<64>
{
template <typename T, typename Op> static __device__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
data[tid] = partial_reduction;
__syncthreads();
if (tid < 32)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 32]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 16]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1 ]);
}
}
};
template <> struct WarpReductor<32>
{
template <typename T, typename Op> static __device__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
data[tid] = partial_reduction;
if (tid < 16)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 16]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1 ]);
}
}
};
template <> struct WarpReductor<16>
{
template <typename T, typename Op> static __device__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
data[tid] = partial_reduction;
if (tid < 8)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1 ]);
}
}
};
template <> struct WarpReductor<8>
{
template <typename T, typename Op> static __device__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
data[tid] = partial_reduction;
if (tid < 4)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1 ]);
}
}
};
template <bool warp> struct ReductionDispatcher;
template <> struct ReductionDispatcher<true>
{
template <int n, typename T, typename Op> static __device__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
WarpReductor<n>::reduce(data, partial_reduction, tid, op);
}
};
template <> struct ReductionDispatcher<false>
{
template <int n, typename T, typename Op> static __device__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
if (tid < n)
data[tid] = partial_reduction;
__syncthreads();
if (n == 512) { if (tid < 256) { data[tid] = partial_reduction = op(partial_reduction, data[tid + 256]); } __syncthreads(); }
if (n >= 256) { if (tid < 128) { data[tid] = partial_reduction = op(partial_reduction, data[tid + 128]); } __syncthreads(); }
if (n >= 128) { if (tid < 64) { data[tid] = partial_reduction = op(partial_reduction, data[tid + 64]); } __syncthreads(); }
if (tid < 32)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 32]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 16]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1]);
}
}
};
///////////////////////////////////////////////////////////////////////////////
// PredValWarpReductor
template <int n> struct PredValWarpReductor;
template <> struct PredValWarpReductor<64>
{
template <typename T, typename V, typename Pred>
static __device__ void reduce(T& myData, V& myVal, volatile T* sdata, V* sval, int tid, const Pred& pred)
{
if (tid < 32)
{
myData = sdata[tid];
myVal = sval[tid];
T reg = sdata[tid + 32];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 32];
}
reg = sdata[tid + 16];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 16];
}
reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 8];
}
reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 1];
}
}
}
};
template <> struct PredValWarpReductor<32>
{
template <typename T, typename V, typename Pred>
static __device__ void reduce(T& myData, V& myVal, volatile T* sdata, V* sval, int tid, const Pred& pred)
{
if (tid < 16)
{
myData = sdata[tid];
myVal = sval[tid];
T reg = sdata[tid + 16];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 16];
}
reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 8];
}
reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 1];
}
}
}
};
template <> struct PredValWarpReductor<16>
{
template <typename T, typename V, typename Pred>
static __device__ void reduce(T& myData, V& myVal, volatile T* sdata, V* sval, int tid, const Pred& pred)
{
if (tid < 8)
{
myData = sdata[tid];
myVal = sval[tid];
T reg = reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 8];
}
reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 1];
}
}
}
};
template <> struct PredValWarpReductor<8>
{
template <typename T, typename V, typename Pred>
static __device__ void reduce(T& myData, V& myVal, volatile T* sdata, V* sval, int tid, const Pred& pred)
{
if (tid < 4)
{
myData = sdata[tid];
myVal = sval[tid];
T reg = reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 1];
}
}
}
};
template <bool warp> struct PredValReductionDispatcher;
template <> struct PredValReductionDispatcher<true>
{
template <int n, typename T, typename V, typename Pred> static __device__ void reduce(T& myData, V& myVal, volatile T* sdata, V* sval, int tid, const Pred& pred)
{
PredValWarpReductor<n>::reduce(myData, myVal, sdata, sval, tid, pred);
}
};
template <> struct PredValReductionDispatcher<false>
{
template <int n, typename T, typename V, typename Pred> static __device__ void reduce(T& myData, V& myVal, volatile T* sdata, V* sval, int tid, const Pred& pred)
{
myData = sdata[tid];
myVal = sval[tid];
if (n >= 512 && tid < 256)
{
T reg = sdata[tid + 256];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 256];
}
__syncthreads();
}
if (n >= 256 && tid < 128)
{
T reg = sdata[tid + 128];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 128];
}
__syncthreads();
}
if (n >= 128 && tid < 64)
{
T reg = sdata[tid + 64];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 64];
}
__syncthreads();
}
if (tid < 32)
{
if (n >= 64)
{
T reg = sdata[tid + 32];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 32];
}
}
if (n >= 32)
{
T reg = sdata[tid + 16];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 16];
}
}
if (n >= 16)
{
T reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 8];
}
}
if (n >= 8)
{
T reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 4];
}
}
if (n >= 4)
{
T reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 2];
}
}
if (n >= 2)
{
T reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval[tid] = myVal = sval[tid + 1];
}
}
}
}
};
///////////////////////////////////////////////////////////////////////////////
// PredVal2WarpReductor
template <int n> struct PredVal2WarpReductor;
template <> struct PredVal2WarpReductor<64>
{
template <typename T, typename V1, typename V2, typename Pred>
static __device__ void reduce(T& myData, V1& myVal1, V2& myVal2, volatile T* sdata, V1* sval1, V2* sval2, int tid, const Pred& pred)
{
if (tid < 32)
{
myData = sdata[tid];
myVal1 = sval1[tid];
myVal2 = sval2[tid];
T reg = sdata[tid + 32];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 32];
sval2[tid] = myVal2 = sval2[tid + 32];
}
reg = sdata[tid + 16];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 16];
sval2[tid] = myVal2 = sval2[tid + 16];
}
reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 8];
sval2[tid] = myVal2 = sval2[tid + 8];
}
reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 4];
sval2[tid] = myVal2 = sval2[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 2];
sval2[tid] = myVal2 = sval2[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 1];
sval2[tid] = myVal2 = sval2[tid + 1];
}
}
}
};
template <> struct PredVal2WarpReductor<32>
{
template <typename T, typename V1, typename V2, typename Pred>
static __device__ void reduce(T& myData, V1& myVal1, V2& myVal2, volatile T* sdata, V1* sval1, V2* sval2, int tid, const Pred& pred)
{
if (tid < 16)
{
myData = sdata[tid];
myVal1 = sval1[tid];
myVal2 = sval2[tid];
T reg = sdata[tid + 16];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 16];
sval2[tid] = myVal2 = sval2[tid + 16];
}
reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 8];
sval2[tid] = myVal2 = sval2[tid + 8];
}
reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 4];
sval2[tid] = myVal2 = sval2[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 2];
sval2[tid] = myVal2 = sval2[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 1];
sval2[tid] = myVal2 = sval2[tid + 1];
}
}
}
};
template <> struct PredVal2WarpReductor<16>
{
template <typename T, typename V1, typename V2, typename Pred>
static __device__ void reduce(T& myData, V1& myVal1, V2& myVal2, volatile T* sdata, V1* sval1, V2* sval2, int tid, const Pred& pred)
{
if (tid < 8)
{
myData = sdata[tid];
myVal1 = sval1[tid];
myVal2 = sval2[tid];
T reg = reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 8];
sval2[tid] = myVal2 = sval2[tid + 8];
}
reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 4];
sval2[tid] = myVal2 = sval2[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 2];
sval2[tid] = myVal2 = sval2[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 1];
sval2[tid] = myVal2 = sval2[tid + 1];
}
}
}
};
template <> struct PredVal2WarpReductor<8>
{
template <typename T, typename V1, typename V2, typename Pred>
static __device__ void reduce(T& myData, V1& myVal1, V2& myVal2, volatile T* sdata, V1* sval1, V2* sval2, int tid, const Pred& pred)
{
if (tid < 4)
{
myData = sdata[tid];
myVal1 = sval1[tid];
myVal2 = sval2[tid];
T reg = reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 4];
sval2[tid] = myVal2 = sval2[tid + 4];
}
reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 2];
sval2[tid] = myVal2 = sval2[tid + 2];
}
reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 1];
sval2[tid] = myVal2 = sval2[tid + 1];
}
}
}
};
template <bool warp> struct PredVal2ReductionDispatcher;
template <> struct PredVal2ReductionDispatcher<true>
{
template <int n, typename T, typename V1, typename V2, typename Pred>
static __device__ void reduce(T& myData, V1& myVal1, V2& myVal2, volatile T* sdata, V1* sval1, V2* sval2, int tid, const Pred& pred)
{
PredVal2WarpReductor<n>::reduce(myData, myVal1, myVal2, sdata, sval1, sval2, tid, pred);
}
};
template <> struct PredVal2ReductionDispatcher<false>
{
template <int n, typename T, typename V1, typename V2, typename Pred>
static __device__ void reduce(T& myData, V1& myVal1, V2& myVal2, volatile T* sdata, V1* sval1, V2* sval2, int tid, const Pred& pred)
{
myData = sdata[tid];
myVal1 = sval1[tid];
myVal2 = sval2[tid];
if (n >= 512 && tid < 256)
{
T reg = sdata[tid + 256];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 256];
sval2[tid] = myVal2 = sval2[tid + 256];
}
__syncthreads();
}
if (n >= 256 && tid < 128)
{
T reg = sdata[tid + 128];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 128];
sval2[tid] = myVal2 = sval2[tid + 128];
}
__syncthreads();
}
if (n >= 128 && tid < 64)
{
T reg = sdata[tid + 64];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 64];
sval2[tid] = myVal2 = sval2[tid + 64];
}
__syncthreads();
}
if (tid < 32)
{
if (n >= 64)
{
T reg = sdata[tid + 32];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 32];
sval2[tid] = myVal2 = sval2[tid + 32];
}
}
if (n >= 32)
{
T reg = sdata[tid + 16];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 16];
sval2[tid] = myVal2 = sval2[tid + 16];
}
}
if (n >= 16)
{
T reg = sdata[tid + 8];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 8];
sval2[tid] = myVal2 = sval2[tid + 8];
}
}
if (n >= 8)
{
T reg = sdata[tid + 4];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 4];
sval2[tid] = myVal2 = sval2[tid + 4];
}
}
if (n >= 4)
{
T reg = sdata[tid + 2];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 2];
sval2[tid] = myVal2 = sval2[tid + 2];
}
}
if (n >= 2)
{
T reg = sdata[tid + 1];
if (pred(reg, myData))
{
sdata[tid] = myData = reg;
sval1[tid] = myVal1 = sval1[tid + 1];
sval2[tid] = myVal2 = sval2[tid + 1];
}
}
}
}
};
} // namespace utility_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_REDUCTION_DETAIL_HPP__

1
modules/gpu/include/opencv2/gpu/device/emulation.hpp
View File

@ -44,7 +44,6 @@
#define OPENCV_GPU_EMULATION_HPP_
#include "warp_reduce.hpp"
#include <stdio.h>
namespace cv { namespace gpu { namespace device
{

123
modules/gpu/include/opencv2/gpu/device/functional.hpp
View File

@ -302,18 +302,18 @@ namespace cv { namespace gpu { namespace device
template <> struct name<type> : binary_function<type, type, type> \
{ \
__device__ __forceinline__ type operator()(type lhs, type rhs) const {return op(lhs, rhs);} \
__device__ __forceinline__ name(const name& other):binary_function<type, type, type>(){}\
__device__ __forceinline__ name():binary_function<type, type, type>(){}\
__device__ __forceinline__ name() {}\
__device__ __forceinline__ name(const name&) {}\
};
template <typename T> struct maximum : binary_function<T, T, T>
{
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
{
return lhs < rhs ? rhs : lhs;
return max(lhs, rhs);
}
__device__ __forceinline__ maximum(const maximum& other):binary_function<T, T, T>(){}
__device__ __forceinline__ maximum():binary_function<T, T, T>(){}
__device__ __forceinline__ maximum() {}
__device__ __forceinline__ maximum(const maximum&) {}
};
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, uchar, ::max)
@ -330,10 +330,10 @@ namespace cv { namespace gpu { namespace device
{
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
{
return lhs < rhs ? lhs : rhs;
return min(lhs, rhs);
}
__device__ __forceinline__ minimum(const minimum& other):binary_function<T, T, T>(){}
__device__ __forceinline__ minimum():binary_function<T, T, T>(){}
__device__ __forceinline__ minimum() {}
__device__ __forceinline__ minimum(const minimum&) {}
};
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, uchar, ::min)
@ -350,6 +350,108 @@ namespace cv { namespace gpu { namespace device
// Math functions
///bound=========================================
template <typename T> struct abs_func : unary_function<T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType x) const
{
return abs(x);
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<unsigned char> : unary_function<unsigned char, unsigned char>
{
__device__ __forceinline__ unsigned char operator ()(unsigned char x) const
{
return x;
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<signed char> : unary_function<signed char, signed char>
{
__device__ __forceinline__ signed char operator ()(signed char x) const
{
return ::abs((int)x);
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<char> : unary_function<char, char>
{
__device__ __forceinline__ char operator ()(char x) const
{
return ::abs((int)x);
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<unsigned short> : unary_function<unsigned short, unsigned short>
{
__device__ __forceinline__ unsigned short operator ()(unsigned short x) const
{
return x;
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<short> : unary_function<short, short>
{
__device__ __forceinline__ short operator ()(short x) const
{
return ::abs((int)x);
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<unsigned int> : unary_function<unsigned int, unsigned int>
{
__device__ __forceinline__ unsigned int operator ()(unsigned int x) const
{
return x;
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<int> : unary_function<int, int>
{
__device__ __forceinline__ int operator ()(int x) const
{
return ::abs(x);
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<float> : unary_function<float, float>
{
__device__ __forceinline__ float operator ()(float x) const
{
return ::fabsf(x);
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
template <> struct abs_func<double> : unary_function<double, double>
{
__device__ __forceinline__ double operator ()(double x) const
{
return ::fabs(x);
}
__device__ __forceinline__ abs_func() {}
__device__ __forceinline__ abs_func(const abs_func&) {}
};
#define OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(name, func) \
template <typename T> struct name ## _func : unary_function<T, float> \
{ \
@ -357,6 +459,8 @@ namespace cv { namespace gpu { namespace device
{ \
return func ## f(v); \
} \
__device__ __forceinline__ name ## _func() {} \
__device__ __forceinline__ name ## _func(const name ## _func&) {} \
}; \
template <> struct name ## _func<double> : unary_function<double, double> \
{ \
@ -364,6 +468,8 @@ namespace cv { namespace gpu { namespace device
{ \
return func(v); \
} \
__device__ __forceinline__ name ## _func() {} \
__device__ __forceinline__ name ## _func(const name ## _func&) {} \
};
#define OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(name, func) \
@ -382,7 +488,6 @@ namespace cv { namespace gpu { namespace device
} \
};
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(fabs, ::fabs)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(sqrt, ::sqrt)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(exp, ::exp)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(exp2, ::exp2)

197
modules/gpu/include/opencv2/gpu/device/reduce.hpp Normal file
View File

@ -0,0 +1,197 @@
/*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_REDUCE_HPP__
#define __OPENCV_GPU_REDUCE_HPP__
#include <thrust/tuple.h>
#include "detail/reduce.hpp"
#include "detail/reduce_key_val.hpp"
namespace cv { namespace gpu { namespace device
{
template <int N, typename T, class Op>
__device__ __forceinline__ void reduce(volatile T* smem, T& val, unsigned int tid, const Op& op)
{
reduce_detail::Dispatcher<N>::reductor::template reduce<volatile T*, T&, const Op&>(smem, val, tid, op);
}
template <int N,
typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
__device__ __forceinline__ void reduce(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
unsigned int tid,
const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
{
reduce_detail::Dispatcher<N>::reductor::template reduce<
const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>&,
const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>&,
const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>&>(smem, val, tid, op);
}
template <unsigned int N, typename K, typename V, class Cmp>
__device__ __forceinline__ void reduceKeyVal(volatile K* skeys, K& key, volatile V* svals, V& val, unsigned int tid, const Cmp& cmp)
{
reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<volatile K*, K&, volatile V*, V&, const Cmp&>(skeys, key, svals, val, tid, cmp);
}
template <unsigned int N,
typename K,
typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
class Cmp>
__device__ __forceinline__ void reduceKeyVal(volatile K* skeys, K& key,
const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
unsigned int tid, const Cmp& cmp)
{
reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<volatile K*, K&,
const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>&,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>&,
const Cmp&>(skeys, key, svals, val, tid, cmp);
}
template <unsigned int N,
typename KP0, typename KP1, typename KP2, typename KP3, typename KP4, typename KP5, typename KP6, typename KP7, typename KP8, typename KP9,
typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
__device__ __forceinline__ void reduceKeyVal(const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>& skeys,
const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
unsigned int tid,
const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp)
{
reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<
const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>&,
const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>&,
const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>&,
const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>&,
const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>&
>(skeys, key, svals, val, tid, cmp);
}
// smem_tuple
template <typename T0>
__device__ __forceinline__
thrust::tuple<volatile T0*>
smem_tuple(T0* t0)
{
return thrust::make_tuple((volatile T0*) t0);
}
template <typename T0, typename T1>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*>
smem_tuple(T0* t0, T1* t1)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1);
}
template <typename T0, typename T1, typename T2>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*>
smem_tuple(T0* t0, T1* t1, T2* t2)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2);
}
template <typename T0, typename T1, typename T2, typename T3>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*>
smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3);
}
template <typename T0, typename T1, typename T2, typename T3, typename T4>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*>
smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4);
}
template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*>
smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5);
}
template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*>
smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6);
}
template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*>
smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7);
}
template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7, typename T8>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*, volatile T8*>
smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7, T8* t8)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7, (volatile T8*) t8);
}
template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7, typename T8, typename T9>
__device__ __forceinline__
thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*, volatile T8*, volatile T9*>
smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7, T8* t8, T9* t9)
{
return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7, (volatile T8*) t8, (volatile T9*) t9);
}
}}}
#endif // __OPENCV_GPU_UTILITY_HPP__

176
modules/gpu/include/opencv2/gpu/device/saturate_cast.hpp
View File

@ -58,35 +58,47 @@ namespace cv { namespace gpu { namespace device
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);
uint res = 0;
int vi = v;
asm("cvt.sat.u8.s8 %0, %1;" : "=r"(res) : "r"(vi));
return res;
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
{
return saturate_cast<uchar>((uint)v);
uint res = 0;
asm("cvt.sat.u8.s16 %0, %1;" : "=r"(res) : "h"(v));
return res;
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
{
uint res = 0;
asm("cvt.sat.u8.u16 %0, %1;" : "=r"(res) : "h"(v));
return res;
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
{
uint res = 0;
asm("cvt.sat.u8.s32 %0, %1;" : "=r"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
{
uint res = 0;
asm("cvt.sat.u8.u32 %0, %1;" : "=r"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<uchar>(iv);
uint res = 0;
asm("cvt.rni.sat.u8.f32 %0, %1;" : "=r"(res) : "f"(v));
return res;
}
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);
#if __CUDA_ARCH__ >= 130
uint res = 0;
asm("cvt.rni.sat.u8.f64 %0, %1;" : "=r"(res) : "d"(v));
return res;
#else
return saturate_cast<uchar>((float)v);
#endif
@ -94,35 +106,47 @@ namespace cv { namespace gpu { namespace device
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);
uint res = 0;
uint vi = v;
asm("cvt.sat.s8.u8 %0, %1;" : "=r"(res) : "r"(vi));
return res;
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
{
return saturate_cast<schar>((int)v);
uint res = 0;
asm("cvt.sat.s8.s16 %0, %1;" : "=r"(res) : "h"(v));
return res;
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
{
uint res = 0;
asm("cvt.sat.s8.u16 %0, %1;" : "=r"(res) : "h"(v));
return res;
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
{
uint res = 0;
asm("cvt.sat.s8.s32 %0, %1;" : "=r"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
{
return (schar) ::min(v, (uint)SCHAR_MAX);
uint res = 0;
asm("cvt.sat.s8.u32 %0, %1;" : "=r"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<schar>(iv);
uint res = 0;
asm("cvt.rni.sat.s8.f32 %0, %1;" : "=r"(res) : "f"(v));
return res;
}
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);
#if __CUDA_ARCH__ >= 130
uint res = 0;
asm("cvt.rni.sat.s8.f64 %0, %1;" : "=r"(res) : "d"(v));
return res;
#else
return saturate_cast<schar>((float)v);
#endif
@ -130,30 +154,41 @@ namespace cv { namespace gpu { namespace device
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
{
return (ushort) ::max((int)v, 0);
ushort res = 0;
int vi = v;
asm("cvt.sat.u16.s8 %0, %1;" : "=h"(res) : "r"(vi));
return res;
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
{
return (ushort) ::max((int)v, 0);
ushort res = 0;
asm("cvt.sat.u16.s16 %0, %1;" : "=h"(res) : "h"(v));
return res;
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
{
return (ushort)((uint)v <= (uint)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0);
ushort res = 0;
asm("cvt.sat.u16.s32 %0, %1;" : "=h"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
{
return (ushort) ::min(v, (uint)USHRT_MAX);
ushort res = 0;
asm("cvt.sat.u16.u32 %0, %1;" : "=h"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<ushort>(iv);
ushort res = 0;
asm("cvt.rni.sat.u16.f32 %0, %1;" : "=h"(res) : "f"(v));
return res;
}
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);
#if __CUDA_ARCH__ >= 130
ushort res = 0;
asm("cvt.rni.sat.u16.f64 %0, %1;" : "=h"(res) : "d"(v));
return res;
#else
return saturate_cast<ushort>((float)v);
#endif
@ -161,31 +196,45 @@ namespace cv { namespace gpu { namespace device
template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
{
return (short) ::min((int)v, SHRT_MAX);
short res = 0;
asm("cvt.sat.s16.u16 %0, %1;" : "=h"(res) : "h"(v));
return res;
}
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);
short res = 0;
asm("cvt.sat.s16.s32 %0, %1;" : "=h"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
{
return (short) ::min(v, (uint)SHRT_MAX);
short res = 0;
asm("cvt.sat.s16.u32 %0, %1;" : "=h"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ short saturate_cast<short>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<short>(iv);
short res = 0;
asm("cvt.rni.sat.s16.f32 %0, %1;" : "=h"(res) : "f"(v));
return res;
}
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);
#if __CUDA_ARCH__ >= 130
short res = 0;
asm("cvt.rni.sat.s16.f64 %0, %1;" : "=h"(res) : "d"(v));
return res;
#else
return saturate_cast<short>((float)v);
#endif
}
template<> __device__ __forceinline__ int saturate_cast<int>(uint v)
{
int res = 0;
asm("cvt.sat.s32.u32 %0, %1;" : "=r"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ int saturate_cast<int>(float v)
{
return __float2int_rn(v);
@ -199,6 +248,25 @@ namespace cv { namespace gpu { namespace device
#endif
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(schar v)
{
uint res = 0;
int vi = v;
asm("cvt.sat.u32.s8 %0, %1;" : "=r"(res) : "r"(vi));
return res;
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(short v)
{
uint res = 0;
asm("cvt.sat.u32.s16 %0, %1;" : "=r"(res) : "h"(v));
return res;
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(int v)
{
uint res = 0;
asm("cvt.sat.u32.s32 %0, %1;" : "=r"(res) : "r"(v));
return res;
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
{
return __float2uint_rn(v);

24
modules/gpu/include/opencv2/gpu/device/utility.hpp
View File

@ -45,7 +45,6 @@
#include "saturate_cast.hpp"
#include "datamov_utils.hpp"
#include "detail/reduction_detail.hpp"
namespace cv { namespace gpu { namespace device
{
@ -156,29 +155,6 @@ namespace cv { namespace gpu { namespace device
}
};
///////////////////////////////////////////////////////////////////////////////
// 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

10
modules/gpu/include/opencv2/gpu/device/vec_distance.hpp
View File

@ -43,7 +43,7 @@
#ifndef __OPENCV_GPU_VEC_DISTANCE_HPP__
#define __OPENCV_GPU_VEC_DISTANCE_HPP__
#include "utility.hpp"
#include "reduce.hpp"
#include "functional.hpp"
#include "detail/vec_distance_detail.hpp"
@ -63,7 +63,7 @@ namespace cv { namespace gpu { namespace device
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
}
__device__ __forceinline__ operator int() const
@ -87,7 +87,7 @@ namespace cv { namespace gpu { namespace device
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
}
__device__ __forceinline__ operator float() const
@ -113,7 +113,7 @@ namespace cv { namespace gpu { namespace device
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
}
__device__ __forceinline__ operator float() const
@ -138,7 +138,7 @@ namespace cv { namespace gpu { namespace device
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
}
__device__ __forceinline__ operator int() const

4
modules/gpu/include/opencv2/gpu/device/vec_math.hpp
View File

@ -280,7 +280,7 @@ namespace cv { namespace gpu { namespace device
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, abs, abs_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) \
@ -327,4 +327,4 @@ namespace cv { namespace gpu { namespace device
#undef OPENCV_GPU_IMPLEMENT_VEC_INT_OP
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VECMATH_HPP__
#endif // __OPENCV_GPU_VECMATH_HPP__

145
modules/gpu/include/opencv2/gpu/device/warp_shuffle.hpp Normal file
View File

@ -0,0 +1,145 @@
/*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_WARP_SHUFFLE_HPP__
#define __OPENCV_GPU_WARP_SHUFFLE_HPP__
namespace cv { namespace gpu { namespace device
{
template <typename T>
__device__ __forceinline__ T shfl(T val, int srcLane, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
return __shfl(val, srcLane, width);
#else
return T();
#endif
}
__device__ __forceinline__ unsigned int shfl(unsigned int val, int srcLane, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
return (unsigned int) __shfl((int) val, srcLane, width);
#else
return 0;
#endif
}
__device__ __forceinline__ double shfl(double val, int srcLane, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
int lo = __double2loint(val);
int hi = __double2hiint(val);
lo = __shfl(lo, srcLane, width);
hi = __shfl(hi, srcLane, width);
return __hiloint2double(hi, lo);
#else
return 0.0;
#endif
}
template <typename T>
__device__ __forceinline__ T shfl_down(T val, unsigned int delta, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
return __shfl_down(val, delta, width);
#else
return T();
#endif
}
__device__ __forceinline__ unsigned int shfl_down(unsigned int val, unsigned int delta, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
return (unsigned int) __shfl_down((int) val, delta, width);
#else
return 0;
#endif
}
__device__ __forceinline__ double shfl_down(double val, unsigned int delta, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
int lo = __double2loint(val);
int hi = __double2hiint(val);
lo = __shfl_down(lo, delta, width);
hi = __shfl_down(hi, delta, width);
return __hiloint2double(hi, lo);
#else
return 0.0;
#endif
}
template <typename T>
__device__ __forceinline__ T shfl_up(T val, unsigned int delta, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
return __shfl_up(val, delta, width);
#else
return T();
#endif
}
__device__ __forceinline__ unsigned int shfl_up(unsigned int val, unsigned int delta, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
return (unsigned int) __shfl_up((int) val, delta, width);
#else
return 0;
#endif
}
__device__ __forceinline__ double shfl_up(double val, unsigned int delta, int width = warpSize)
{
#if __CUDA_ARCH__ >= 300
int lo = __double2loint(val);
int hi = __double2hiint(val);
lo = __shfl_up(lo, delta, width);
hi = __shfl_up(hi, delta, width);
return __hiloint2double(hi, lo);
#else
return 0.0;
#endif
}
}}}
#endif // __OPENCV_GPU_WARP_SHUFFLE_HPP__

236
modules/gpu/include/opencv2/gpu/gpu.hpp
View File

@ -792,31 +792,23 @@ private:
GpuMat lab, l, ab;
};
struct CV_EXPORTS CannyBuf
{
void create(const Size& image_size, int apperture_size = 3);
void release();
struct CV_EXPORTS CannyBuf;
GpuMat dx, dy;
GpuMat mag;
GpuMat map;
GpuMat st1, st2;
Ptr<FilterEngine_GPU> filterDX, filterDY;
};
CV_EXPORTS void Canny(const GpuMat& image, GpuMat& edges, double low_thresh, double high_thresh, int apperture_size = 3, bool L2gradient = false);
CV_EXPORTS void Canny(const GpuMat& image, CannyBuf& buf, GpuMat& edges, double low_thresh, double high_thresh, int apperture_size = 3, bool L2gradient = false);
CV_EXPORTS void Canny(const GpuMat& dx, const GpuMat& dy, GpuMat& edges, double low_thresh, double high_thresh, bool L2gradient = false);
CV_EXPORTS void Canny(const GpuMat& dx, const GpuMat& dy, CannyBuf& buf, GpuMat& edges, double low_thresh, double high_thresh, bool L2gradient = false);
struct CV_EXPORTS CannyBuf
{
CannyBuf() {}
explicit CannyBuf(const Size& image_size, int apperture_size = 3) {create(image_size, apperture_size);}
CannyBuf(const GpuMat& dx_, const GpuMat& dy_);
void create(const Size& image_size, int apperture_size = 3);
void release();
GpuMat dx, dy;
GpuMat dx_buf, dy_buf;
GpuMat edgeBuf;
GpuMat trackBuf1, trackBuf2;
Ptr<FilterEngine_GPU> filterDX, filterDY;
};
class CV_EXPORTS ImagePyramid
{
public:
@ -855,6 +847,11 @@ CV_EXPORTS void HoughLines(const GpuMat& src, GpuMat& lines, float rho, float th
CV_EXPORTS void HoughLines(const GpuMat& src, GpuMat& lines, HoughLinesBuf& buf, float rho, float theta, int threshold, bool doSort = false, int maxLines = 4096);
CV_EXPORTS void HoughLinesDownload(const GpuMat& d_lines, OutputArray h_lines, OutputArray h_votes = noArray());
//! HoughLinesP
//! finds line segments in the black-n-white image using probabalistic Hough transform
CV_EXPORTS void HoughLinesP(const GpuMat& image, GpuMat& lines, HoughLinesBuf& buf, float rho, float theta, int minLineLength, int maxLineGap, int maxLines = 4096);
//! HoughCircles
struct HoughCirclesBuf
@ -1036,11 +1033,9 @@ CV_EXPORTS void histRange(const GpuMat& src, GpuMat hist[4], const GpuMat levels
//! Calculates histogram for 8u one channel image
//! Output hist will have one row, 256 cols and CV32SC1 type.
CV_EXPORTS void calcHist(const GpuMat& src, GpuMat& hist, Stream& stream = Stream::Null());
CV_EXPORTS void calcHist(const GpuMat& src, GpuMat& hist, GpuMat& buf, Stream& stream = Stream::Null());
//! normalizes the grayscale image brightness and contrast by normalizing its histogram
CV_EXPORTS void equalizeHist(const GpuMat& src, GpuMat& dst, Stream& stream = Stream::Null());
CV_EXPORTS void equalizeHist(const GpuMat& src, GpuMat& dst, GpuMat& hist, Stream& stream = Stream::Null());
CV_EXPORTS void equalizeHist(const GpuMat& src, GpuMat& dst, GpuMat& hist, GpuMat& buf, Stream& stream = Stream::Null());
//////////////////////////////// StereoBM_GPU ////////////////////////////////
@ -1532,6 +1527,97 @@ public:
int detectMultiScale(const GpuMat& image, GpuMat& objectsBuf, Size maxObjectSize, Size minSize = Size(), double scaleFactor = 1.1, int minNeighbors = 4);
};
// ======================== GPU version for soft cascade ===================== //
class CV_EXPORTS ChannelsProcessor
{
public:
enum
{
GENERIC = 1 << 4,
SEPARABLE = 2 << 4
};
// Appends specified number of HOG first-order features integrals into given vector.
// Param frame is an input 3-channel bgr image.
// Param channels is a GPU matrix of optionally shrinked channels
// Param stream is stream is a high-level CUDA stream abstraction used for asynchronous execution.
virtual void apply(InputArray frame, OutputArray channels, Stream& stream = Stream::Null()) = 0;
// Creates a specific preprocessor implementation.
// Param shrinkage is a resizing factor. Resize is applied before the computing integral sum
// Param bins is a number of HOG-like channels.
// Param flags is a channel computing extra flags.
static cv::Ptr<ChannelsProcessor> create(const int shrinkage, const int bins, const int flags = GENERIC);
virtual ~ChannelsProcessor();
protected:
ChannelsProcessor();
};
// Implementation of soft (stageless) cascaded detector.
class CV_EXPORTS SCascade : public Algorithm
{
public:
// Representation of detectors result.
struct CV_EXPORTS Detection
{
ushort x;
ushort y;
ushort w;
ushort h;
float confidence;
int kind;
enum {PEDESTRIAN = 0};
};
enum { NO_REJECT = 1, DOLLAR = 2, /*PASCAL = 4,*/ DEFAULT = NO_REJECT, NMS_MASK = 0xF};
// An empty cascade will be created.
// Param minScale is a minimum scale relative to the original size of the image on which cascade will be applyed.
// Param minScale is a maximum scale relative to the original size of the image on which cascade will be applyed.
// Param scales is a number of scales from minScale to maxScale.
// Param flags is an extra tuning flags.
SCascade(const double minScale = 0.4, const double maxScale = 5., const int scales = 55,
const int flags = NO_REJECT || ChannelsProcessor::GENERIC);
virtual ~SCascade();
cv::AlgorithmInfo* info() const;
// Load cascade from FileNode.
// Param fn is a root node for cascade. Should be <cascade>.
virtual bool load(const FileNode& fn);
// Load cascade config.
virtual void read(const FileNode& fn);
// Return the matrix of of detectioned objects.
// Param image is a frame on which detector will be applied.
// Param rois is a regions of interests mask generated by genRoi.
// Only the objects that fall into one of the regions will be returned.
// Param objects is an output array of Detections represented as GpuMat of detections (SCascade::Detection)
// The first element of the matrix is actually a count of detections.
// Param stream is stream is a high-level CUDA stream abstraction used for asynchronous execution
virtual void detect(InputArray image, InputArray rois, OutputArray objects, Stream& stream = Stream::Null()) const;
private:
struct Fields;
Fields* fields;
double minScale;
double maxScale;
int scales;
int flags;
};
CV_EXPORTS bool initModule_gpu(void);
////////////////////////////////// SURF //////////////////////////////////////////
class CV_EXPORTS SURF_GPU
@ -1877,8 +1963,6 @@ private:
GpuMat uPyr_[2];
GpuMat vPyr_[2];
bool isDeviceArch11_;
};
@ -1895,7 +1979,6 @@ public:
polyN = 5;
polySigma = 1.1;
flags = 0;
isDeviceArch11_ = !DeviceInfo().supports(FEATURE_SET_COMPUTE_12);
}
int numLevels;
@ -1943,8 +2026,113 @@ private:
GpuMat frames_[2];
GpuMat pyrLevel_[2], M_, bufM_, R_[2], blurredFrame_[2];
std::vector<GpuMat> pyramid0_, pyramid1_;
};
bool isDeviceArch11_;
// Implementation of the Zach, Pock and Bischof Dual TV-L1 Optical Flow method
//
// see reference:
// [1] C. Zach, T. Pock and H. Bischof, "A Duality Based Approach for Realtime TV-L1 Optical Flow".
// [2] Javier Sanchez, Enric Meinhardt-Llopis and Gabriele Facciolo. "TV-L1 Optical Flow Estimation".
class CV_EXPORTS OpticalFlowDual_TVL1_GPU
{
public:
OpticalFlowDual_TVL1_GPU();
void operator ()(const GpuMat& I0, const GpuMat& I1, GpuMat& flowx, GpuMat& flowy);
void collectGarbage();
/**
* Time step of the numerical scheme.
*/
double tau;
/**
* Weight parameter for the data term, attachment parameter.
* This is the most relevant parameter, which determines the smoothness of the output.
* The smaller this parameter is, the smoother the solutions we obtain.
* It depends on the range of motions of the images, so its value should be adapted to each image sequence.
*/
double lambda;
/**
* Weight parameter for (u - v)^2, tightness parameter.
* It serves as a link between the attachment and the regularization terms.
* In theory, it should have a small value in order to maintain both parts in correspondence.
* The method is stable for a large range of values of this parameter.
*/
double theta;
/**
* Number of scales used to create the pyramid of images.
*/
int nscales;
/**
* Number of warpings per scale.
* Represents the number of times that I1(x+u0) and grad( I1(x+u0) ) are computed per scale.
* This is a parameter that assures the stability of the method.
* It also affects the running time, so it is a compromise between speed and accuracy.
*/
int warps;
/**
* Stopping criterion threshold used in the numerical scheme, which is a trade-off between precision and running time.
* A small value will yield more accurate solutions at the expense of a slower convergence.
*/
double epsilon;
/**
* Stopping criterion iterations number used in the numerical scheme.
*/
int iterations;
bool useInitialFlow;
private:
void procOneScale(const GpuMat& I0, const GpuMat& I1, GpuMat& u1, GpuMat& u2);
std::vector<GpuMat> I0s;
std::vector<GpuMat> I1s;
std::vector<GpuMat> u1s;
std::vector<GpuMat> u2s;
GpuMat I1x_buf;
GpuMat I1y_buf;
GpuMat I1w_buf;
GpuMat I1wx_buf;
GpuMat I1wy_buf;
GpuMat grad_buf;
GpuMat rho_c_buf;
GpuMat p11_buf;
GpuMat p12_buf;
GpuMat p21_buf;
GpuMat p22_buf;
GpuMat diff_buf;
GpuMat norm_buf;
};
//! Calculates optical flow for 2 images using block matching algorithm */
CV_EXPORTS void calcOpticalFlowBM(const GpuMat& prev, const GpuMat& curr,
Size block_size, Size shift_size, Size max_range, bool use_previous,
GpuMat& velx, GpuMat& vely, GpuMat& buf,
Stream& stream = Stream::Null());
class CV_EXPORTS FastOpticalFlowBM
{
public:
void operator ()(const GpuMat& I0, const GpuMat& I1, GpuMat& flowx, GpuMat& flowy, int search_window = 21, int block_window = 7, Stream& s = Stream::Null());
private:
GpuMat buffer;
GpuMat extended_I0;
GpuMat extended_I1;
};

26
modules/gpu/misc/carma.toolchain.cmake Normal file
View File

@ -0,0 +1,26 @@
set(CMAKE_SYSTEM_NAME Linux)
set(CMAKE_SYSTEM_VERSION 1)
set(CMAKE_SYSTEM_PROCESSOR arm)
set(CMAKE_C_COMPILER arm-linux-gnueabi-gcc-4.5)
set(CMAKE_CXX_COMPILER arm-linux-gnueabi-g++-4.5)
#suppress compiller varning
set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-psabi" )
set( CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-psabi" )
# can be any other plases
set(__arm_linux_eabi_root /usr/arm-linux-gnueabi)
set(CMAKE_FIND_ROOT_PATH ${CMAKE_FIND_ROOT_PATH} ${__arm_linux_eabi_root})
if(EXISTS ${CUDA_TOOLKIT_ROOT_DIR})
set(CMAKE_FIND_ROOT_PATH ${CMAKE_FIND_ROOT_PATH} ${CUDA_TOOLKIT_ROOT_DIR})
endif()
set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM ONLY)
set(CARMA 1)
add_definitions(-DCARMA)

881
modules/gpu/perf/perf_core.cpp
View File

File diff suppressed because it is too large Load Diff

111
modules/gpu/perf/perf_imgproc.cpp
View File

@ -581,13 +581,12 @@ PERF_TEST_P(Sz, ImgProc_CalcHist, GPU_TYPICAL_MAT_SIZES)
{
cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat d_hist;
cv::gpu::GpuMat d_buf;
cv::gpu::calcHist(d_src, d_hist, d_buf);
cv::gpu::calcHist(d_src, d_hist);
TEST_CYCLE()
{
cv::gpu::calcHist(d_src, d_hist, d_buf);
cv::gpu::calcHist(d_src, d_hist);
}
GPU_SANITY_CHECK(d_hist);
@ -1706,10 +1705,40 @@ PERF_TEST_P(Sz_Depth_Cn, ImgProc_ImagePyramidGetLayer, Combine(GPU_TYPICAL_MAT_S
}
}
namespace {
struct Vec4iComparator
{
bool operator()(const cv::Vec4i& a, const cv::Vec4i b) const
{
if (a[0] != b[0]) return a[0] < b[0];
else if(a[1] != b[1]) return a[1] < b[1];
else if(a[2] != b[2]) return a[2] < b[2];
else return a[3] < b[3];
}
};
struct Vec3fComparator
{
bool operator()(const cv::Vec3f& a, const cv::Vec3f b) const
{
if(a[0] != b[0]) return a[0] < b[0];
else if(a[1] != b[1]) return a[1] < b[1];
else return a[2] < b[2];
}
};
struct Vec2fComparator
{
bool operator()(const cv::Vec2f& a, const cv::Vec2f b) const
{
if(a[0] != b[0]) return a[0] < b[0];
else return a[1] < b[1];
}
};
}
//////////////////////////////////////////////////////////////////////
// HoughLines
PERF_TEST_P(Sz, DISABLED_ImgProc_HoughLines, GPU_TYPICAL_MAT_SIZES)
PERF_TEST_P(Sz, ImgProc_HoughLines, GPU_TYPICAL_MAT_SIZES)
{
declare.time(30.0);
@ -1744,7 +1773,11 @@ PERF_TEST_P(Sz, DISABLED_ImgProc_HoughLines, GPU_TYPICAL_MAT_SIZES)
cv::gpu::HoughLines(d_src, d_lines, d_buf, rho, theta, threshold);
}
GPU_SANITY_CHECK(d_lines);
cv::Mat h_lines(d_lines);
cv::Vec2f* begin = (cv::Vec2f*)(h_lines.ptr<char>(0));
cv::Vec2f* end = (cv::Vec2f*)(h_lines.ptr<char>(0) + (h_lines.cols) * 2 * sizeof(float));
std::sort(begin, end, Vec2fComparator());
SANITY_CHECK(h_lines);
}
else
{
@ -1756,7 +1789,64 @@ PERF_TEST_P(Sz, DISABLED_ImgProc_HoughLines, GPU_TYPICAL_MAT_SIZES)
cv::HoughLines(src, lines, rho, theta, threshold);
}
CPU_SANITY_CHECK(lines);
std::sort(lines.begin(), lines.end(), Vec2fComparator());
SANITY_CHECK(lines);
}
}
//////////////////////////////////////////////////////////////////////
// HoughLinesP
DEF_PARAM_TEST_1(Image, std::string);
PERF_TEST_P(Image, ImgProc_HoughLinesP, testing::Values("cv/shared/pic5.png", "stitching/a1.png"))
{
declare.time(30.0);
std::string fileName = getDataPath(GetParam());
const double rho = 1.0f;
const double theta = CV_PI / 180.0;
const int threshold = 100;
const int minLineLenght = 50;
const int maxLineGap = 5;
cv::Mat image = cv::imread(fileName, cv::IMREAD_GRAYSCALE);
cv::Mat mask;
cv::Canny(image, mask, 50, 100);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_mask(mask);
cv::gpu::GpuMat d_lines;
cv::gpu::HoughLinesBuf d_buf;
cv::gpu::HoughLinesP(d_mask, d_lines, d_buf, rho, theta, minLineLenght, maxLineGap);
TEST_CYCLE()
{
cv::gpu::HoughLinesP(d_mask, d_lines, d_buf, rho, theta, minLineLenght, maxLineGap);
}
cv::Mat h_lines(d_lines);
cv::Vec4i* begin = h_lines.ptr<cv::Vec4i>();
cv::Vec4i* end = h_lines.ptr<cv::Vec4i>() + h_lines.cols;
std::sort(begin, end, Vec4iComparator());
SANITY_CHECK(h_lines);
}
else
{
std::vector<cv::Vec4i> lines;
cv::HoughLinesP(mask, lines, rho, theta, threshold, minLineLenght, maxLineGap);
TEST_CYCLE()
{
cv::HoughLinesP(mask, lines, rho, theta, threshold, minLineLenght, maxLineGap);
}
std::sort(lines.begin(), lines.end(), Vec4iComparator());
SANITY_CHECK(lines);
}
}
@ -1804,7 +1894,11 @@ PERF_TEST_P(Sz_Dp_MinDist, ImgProc_HoughCircles, Combine(GPU_TYPICAL_MAT_SIZES,
cv::gpu::HoughCircles(d_src, d_circles, d_buf, CV_HOUGH_GRADIENT, dp, minDist, cannyThreshold, votesThreshold, minRadius, maxRadius);
}
GPU_SANITY_CHECK(d_circles);
cv::Mat h_circles(d_circles);
cv::Vec3f* begin = (cv::Vec3f*)(h_circles.ptr<char>(0));
cv::Vec3f* end = (cv::Vec3f*)(h_circles.ptr<char>(0) + (h_circles.cols) * 3 * sizeof(float));
std::sort(begin, end, Vec3fComparator());
SANITY_CHECK(h_circles);
}
else
{
@ -1817,7 +1911,8 @@ PERF_TEST_P(Sz_Dp_MinDist, ImgProc_HoughCircles, Combine(GPU_TYPICAL_MAT_SIZES,
cv::HoughCircles(src, circles, CV_HOUGH_GRADIENT, dp, minDist, cannyThreshold, votesThreshold, minRadius, maxRadius);
}
CPU_SANITY_CHECK(circles);
std::sort(circles.begin(), circles.end(), Vec3fComparator());
SANITY_CHECK(circles);
}
}

3
modules/gpu/perf/perf_objdetect.cpp
View File

@ -89,7 +89,6 @@ PERF_TEST_P(HOG, CalTech, Values<string>("gpu/caltech/image_00000009_0.png", "gp
SANITY_CHECK(found_locations);
}
///////////////////////////////////////////////////////////////
// HaarClassifier
@ -181,4 +180,4 @@ PERF_TEST_P(ImageAndCascade, ObjDetect_LBPClassifier,
}
}
} // namespace
} // namespace

279
modules/gpu/perf/perf_softcascade.cpp Normal file
View File

@ -0,0 +1,279 @@
#include "perf_precomp.hpp"
#define GPU_PERF_TEST_P(fixture, name, params) \
class fixture##_##name : public fixture {\
public:\
fixture##_##name() {}\
protected:\
virtual void __cpu();\
virtual void __gpu();\
virtual void PerfTestBody();\
};\
TEST_P(fixture##_##name, name /*perf*/){ RunPerfTestBody(); }\
INSTANTIATE_TEST_CASE_P(/*none*/, fixture##_##name, params);\
void fixture##_##name::PerfTestBody() { if (PERF_RUN_GPU()) __gpu(); else __cpu(); }
#define RUN_CPU(fixture, name)\
void fixture##_##name::__cpu()
#define RUN_GPU(fixture, name)\
void fixture##_##name::__gpu()
#define NO_CPU(fixture, name)\
void fixture##_##name::__cpu() { FAIL() << "No such CPU implementation analogy";}
namespace {
struct DetectionLess
{
bool operator()(const cv::gpu::SCascade::Detection& a,
const cv::gpu::SCascade::Detection& b) const
{
if (a.x != b.x) return a.x < b.x;
else if (a.y != b.y) return a.y < b.y;
else if (a.w != b.w) return a.w < b.w;
else return a.h < b.h;
}
};
cv::Mat sortDetections(cv::gpu::GpuMat& objects)
{
cv::Mat detections(objects);
typedef cv::gpu::SCascade::Detection Detection;
Detection* begin = (Detection*)(detections.ptr<char>(0));
Detection* end = (Detection*)(detections.ptr<char>(0) + detections.cols);
std::sort(begin, end, DetectionLess());
return detections;
}
}
typedef std::tr1::tuple<std::string, std::string> fixture_t;
typedef perf::TestBaseWithParam<fixture_t> SCascadeTest;
GPU_PERF_TEST_P(SCascadeTest, detect,
testing::Combine(
testing::Values(std::string("cv/cascadeandhog/sc_cvpr_2012_to_opencv.xml")),
testing::Values(std::string("cv/cascadeandhog/bahnhof/image_00000000_0.png"))))
RUN_GPU(SCascadeTest, detect)
{
cv::Mat cpu = readImage (GET_PARAM(1));
ASSERT_FALSE(cpu.empty());
cv::gpu::GpuMat colored(cpu);
cv::gpu::SCascade cascade;
cv::FileStorage fs(perf::TestBase::getDataPath(GET_PARAM(0)), cv::FileStorage::READ);
ASSERT_TRUE(fs.isOpened());
ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1);
rois.setTo(1);
cascade.detect(colored, rois, objectBoxes);
TEST_CYCLE()
{
cascade.detect(colored, rois, objectBoxes);
}
SANITY_CHECK(sortDetections(objectBoxes));
}
NO_CPU(SCascadeTest, detect)
static cv::Rect getFromTable(int idx)
{
static const cv::Rect rois[] =
{
cv::Rect( 65 * 4, 20 * 4, 35 * 4, 80 * 4),
cv::Rect( 95 * 4, 35 * 4, 45 * 4, 40 * 4),
cv::Rect( 45 * 4, 35 * 4, 45 * 4, 40 * 4),
cv::Rect( 25 * 4, 27 * 4, 50 * 4, 45 * 4),
cv::Rect(100 * 4, 50 * 4, 45 * 4, 40 * 4),
cv::Rect( 60 * 4, 30 * 4, 45 * 4, 40 * 4),
cv::Rect( 40 * 4, 55 * 4, 50 * 4, 40 * 4),
cv::Rect( 48 * 4, 37 * 4, 72 * 4, 80 * 4),
cv::Rect( 48 * 4, 32 * 4, 85 * 4, 58 * 4),
cv::Rect( 48 * 4, 0 * 4, 32 * 4, 27 * 4)
};
return rois[idx];
}
typedef std::tr1::tuple<std::string, std::string, int> roi_fixture_t;
typedef perf::TestBaseWithParam<roi_fixture_t> SCascadeTestRoi;
GPU_PERF_TEST_P(SCascadeTestRoi, detectInRoi,
testing::Combine(
testing::Values(std::string("cv/cascadeandhog/sc_cvpr_2012_to_opencv.xml")),
testing::Values(std::string("cv/cascadeandhog/bahnhof/image_00000000_0.png")),
testing::Range(0, 5)))
RUN_GPU(SCascadeTestRoi, detectInRoi)
{
cv::Mat cpu = readImage (GET_PARAM(1));
ASSERT_FALSE(cpu.empty());
cv::gpu::GpuMat colored(cpu);
cv::gpu::SCascade cascade;
cv::FileStorage fs(perf::TestBase::getDataPath(GET_PARAM(0)), cv::FileStorage::READ);
ASSERT_TRUE(fs.isOpened());
ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
cv::gpu::GpuMat objectBoxes(1, 16384 * 20, CV_8UC1), rois(colored.size(), CV_8UC1);
rois.setTo(0);
int nroi = GET_PARAM(2);
cv::RNG rng;
for (int i = 0; i < nroi; ++i)
{
cv::Rect r = getFromTable(rng(10));
cv::gpu::GpuMat sub(rois, r);
sub.setTo(1);
}
cascade.detect(colored, rois, objectBoxes);
TEST_CYCLE()
{
cascade.detect(colored, rois, objectBoxes);
}
SANITY_CHECK(sortDetections(objectBoxes));
}
NO_CPU(SCascadeTestRoi, detectInRoi)
GPU_PERF_TEST_P(SCascadeTestRoi, detectEachRoi,
testing::Combine(
testing::Values(std::string("cv/cascadeandhog/sc_cvpr_2012_to_opencv.xml")),
testing::Values(std::string("cv/cascadeandhog/bahnhof/image_00000000_0.png")),
testing::Range(0, 10)))
RUN_GPU(SCascadeTestRoi, detectEachRoi)
{
cv::Mat cpu = readImage (GET_PARAM(1));
ASSERT_FALSE(cpu.empty());
cv::gpu::GpuMat colored(cpu);
cv::gpu::SCascade cascade;
cv::FileStorage fs(perf::TestBase::getDataPath(GET_PARAM(0)), cv::FileStorage::READ);
ASSERT_TRUE(fs.isOpened());
ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
cv::gpu::GpuMat objectBoxes(1, 16384 * 20, CV_8UC1), rois(colored.size(), CV_8UC1);
rois.setTo(0);
int idx = GET_PARAM(2);
cv::Rect r = getFromTable(idx);
cv::gpu::GpuMat sub(rois, r);
sub.setTo(1);
cascade.detect(colored, rois, objectBoxes);
TEST_CYCLE()
{
cascade.detect(colored, rois, objectBoxes);
}
SANITY_CHECK(sortDetections(objectBoxes));
}
NO_CPU(SCascadeTestRoi, detectEachRoi)
GPU_PERF_TEST_P(SCascadeTest, detectOnIntegral,
testing::Combine(
testing::Values(std::string("cv/cascadeandhog/sc_cvpr_2012_to_opencv.xml")),
testing::Values(std::string("cv/cascadeandhog/integrals.xml"))))
static std::string itoa(long i)
{
static char s[65];
sprintf(s, "%ld", i);
return std::string(s);
}
RUN_GPU(SCascadeTest, detectOnIntegral)
{
cv::FileStorage fsi(perf::TestBase::getDataPath(GET_PARAM(1)), cv::FileStorage::READ);
ASSERT_TRUE(fsi.isOpened());
cv::gpu::GpuMat hogluv(121 * 10, 161, CV_32SC1);
for (int i = 0; i < 10; ++i)
{
cv::Mat channel;
fsi[std::string("channel") + itoa(i)] >> channel;
cv::gpu::GpuMat gchannel(hogluv, cv::Rect(0, 121 * i, 161, 121));
gchannel.upload(channel);
}
cv::gpu::SCascade cascade;
cv::FileStorage fs(perf::TestBase::getDataPath(GET_PARAM(0)), cv::FileStorage::READ);
ASSERT_TRUE(fs.isOpened());
ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(cv::Size(640, 480), CV_8UC1);
rois.setTo(1);
cascade.detect(hogluv, rois, objectBoxes);
TEST_CYCLE()
{
cascade.detect(hogluv, rois, objectBoxes);
}
SANITY_CHECK(sortDetections(objectBoxes));
}
NO_CPU(SCascadeTest, detectOnIntegral)
GPU_PERF_TEST_P(SCascadeTest, detectStream,
testing::Combine(
testing::Values(std::string("cv/cascadeandhog/sc_cvpr_2012_to_opencv.xml")),
testing::Values(std::string("cv/cascadeandhog/bahnhof/image_00000000_0.png"))))
RUN_GPU(SCascadeTest, detectStream)
{
cv::Mat cpu = readImage (GET_PARAM(1));
ASSERT_FALSE(cpu.empty());
cv::gpu::GpuMat colored(cpu);
cv::gpu::SCascade cascade;
cv::FileStorage fs(perf::TestBase::getDataPath(GET_PARAM(0)), cv::FileStorage::READ);
ASSERT_TRUE(fs.isOpened());
ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1);
rois.setTo(1);
cv::gpu::Stream s;
cascade.detect(colored, rois, objectBoxes, s);
TEST_CYCLE()
{
cascade.detect(colored, rois, objectBoxes, s);
}
#ifdef HAVE_CUDA
cudaDeviceSynchronize();
#endif
SANITY_CHECK(sortDetections(objectBoxes));
}
NO_CPU(SCascadeTest, detectStream)

167
modules/gpu/perf/perf_video.cpp
View File

@ -394,6 +394,173 @@ PERF_TEST_P(ImagePair, Video_FarnebackOpticalFlow,
}
}
//////////////////////////////////////////////////////
// OpticalFlowDual_TVL1
PERF_TEST_P(ImagePair, Video_OpticalFlowDual_TVL1,
Values<pair_string>(make_pair("gpu/opticalflow/frame0.png", "gpu/opticalflow/frame1.png")))
{
declare.time(20);
cv::Mat frame0 = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(frame0.empty());
cv::Mat frame1 = readImage(GetParam().second, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(frame1.empty());
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_frame0(frame0);
cv::gpu::GpuMat d_frame1(frame1);
cv::gpu::GpuMat d_flowx;
cv::gpu::GpuMat d_flowy;
cv::gpu::OpticalFlowDual_TVL1_GPU d_alg;
d_alg(d_frame0, d_frame1, d_flowx, d_flowy);
TEST_CYCLE()
{
d_alg(d_frame0, d_frame1, d_flowx, d_flowy);
}
GPU_SANITY_CHECK(d_flowx);
GPU_SANITY_CHECK(d_flowy);
}
else
{
cv::Mat flow;
cv::OpticalFlowDual_TVL1 alg;
alg(frame0, frame1, flow);
TEST_CYCLE()
{
alg(frame0, frame1, flow);
}
CPU_SANITY_CHECK(flow);
}
}
//////////////////////////////////////////////////////
// OpticalFlowBM
void calcOpticalFlowBM(const cv::Mat& prev, const cv::Mat& curr,
cv::Size bSize, cv::Size shiftSize, cv::Size maxRange, int usePrevious,
cv::Mat& velx, cv::Mat& vely)
{
cv::Size sz((curr.cols - bSize.width + shiftSize.width)/shiftSize.width, (curr.rows - bSize.height + shiftSize.height)/shiftSize.height);
velx.create(sz, CV_32FC1);
vely.create(sz, CV_32FC1);
CvMat cvprev = prev;
CvMat cvcurr = curr;
CvMat cvvelx = velx;
CvMat cvvely = vely;
cvCalcOpticalFlowBM(&cvprev, &cvcurr, bSize, shiftSize, maxRange, usePrevious, &cvvelx, &cvvely);
}
PERF_TEST_P(ImagePair, Video_OpticalFlowBM,
Values<pair_string>(make_pair("gpu/opticalflow/frame0.png", "gpu/opticalflow/frame1.png")))
{
declare.time(400);
cv::Mat frame0 = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(frame0.empty());
cv::Mat frame1 = readImage(GetParam().second, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(frame1.empty());
cv::Size block_size(16, 16);
cv::Size shift_size(1, 1);
cv::Size max_range(16, 16);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_frame0(frame0);
cv::gpu::GpuMat d_frame1(frame1);
cv::gpu::GpuMat d_velx, d_vely, buf;
cv::gpu::calcOpticalFlowBM(d_frame0, d_frame1, block_size, shift_size, max_range, false, d_velx, d_vely, buf);
TEST_CYCLE()
{
cv::gpu::calcOpticalFlowBM(d_frame0, d_frame1, block_size, shift_size, max_range, false, d_velx, d_vely, buf);
}
GPU_SANITY_CHECK(d_velx);
GPU_SANITY_CHECK(d_vely);
}
else
{
cv::Mat velx, vely;
calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
TEST_CYCLE()
{
calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
}
CPU_SANITY_CHECK(velx);
CPU_SANITY_CHECK(vely);
}
}
PERF_TEST_P(ImagePair, Video_FastOpticalFlowBM,
Values<pair_string>(make_pair("gpu/opticalflow/frame0.png", "gpu/opticalflow/frame1.png")))
{
declare.time(400);
cv::Mat frame0 = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(frame0.empty());
cv::Mat frame1 = readImage(GetParam().second, cv::IMREAD_GRAYSCALE);
ASSERT_FALSE(frame1.empty());
cv::Size block_size(16, 16);
cv::Size shift_size(1, 1);
cv::Size max_range(16, 16);
if (PERF_RUN_GPU())
{
cv::gpu::GpuMat d_frame0(frame0);
cv::gpu::GpuMat d_frame1(frame1);
cv::gpu::GpuMat d_velx, d_vely;
cv::gpu::FastOpticalFlowBM fastBM;
fastBM(d_frame0, d_frame1, d_velx, d_vely, max_range.width, block_size.width);
TEST_CYCLE()
{
fastBM(d_frame0, d_frame1, d_velx, d_vely, max_range.width, block_size.width);
}
GPU_SANITY_CHECK(d_velx);
GPU_SANITY_CHECK(d_vely);
}
else
{
cv::Mat velx, vely;
calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
TEST_CYCLE()
{
calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
}
CPU_SANITY_CHECK(velx);
CPU_SANITY_CHECK(vely);
}
}
//////////////////////////////////////////////////////
// FGDStatModel

16
modules/gpu/src/arithm.cpp
View File

@ -68,11 +68,16 @@ void cv::gpu::polarToCart(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, bool,
void cv::gpu::gemm(const GpuMat& src1, const GpuMat& src2, double alpha, const GpuMat& src3, double beta, GpuMat& dst, int flags, Stream& stream)
{
#ifndef HAVE_CUBLAS
(void)src1; (void)src2; (void)alpha; (void)src3; (void)beta; (void)dst; (void)flags; (void)stream;
(void)src1;
(void)src2;
(void)alpha;
(void)src3;
(void)beta;
(void)dst;
(void)flags;
(void)stream;
CV_Error(CV_StsNotImplemented, "The library was build without CUBLAS");
#else
// CUBLAS works with column-major matrices
CV_Assert(src1.type() == CV_32FC1 || src1.type() == CV_32FC2 || src1.type() == CV_64FC1 || src1.type() == CV_64FC2);
@ -80,7 +85,7 @@ void cv::gpu::gemm(const GpuMat& src1, const GpuMat& src2, double alpha, const G
if (src1.depth() == CV_64F)
{
if (!TargetArchs::builtWith(NATIVE_DOUBLE) || !DeviceInfo().supports(NATIVE_DOUBLE))
if (!deviceSupports(NATIVE_DOUBLE))
CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
}
@ -188,7 +193,6 @@ void cv::gpu::gemm(const GpuMat& src1, const GpuMat& src2, double alpha, const G
}
cublasSafeCall( cublasDestroy_v2(handle) );
#endif
}
@ -227,7 +231,7 @@ void cv::gpu::transpose(const GpuMat& src, GpuMat& dst, Stream& s)
}
else // if (src.elemSize() == 8)
{
if (!TargetArchs::builtWith(NATIVE_DOUBLE) || !DeviceInfo().supports(NATIVE_DOUBLE))
if (!deviceSupports(NATIVE_DOUBLE))
CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
NppStStreamHandler h(stream);

84
modules/gpu/src/brute_force_matcher.cpp
View File

@ -88,71 +88,71 @@ namespace cv { namespace gpu { namespace device
{
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
}
namespace bf_knnmatch
{
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void match2L1_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void match2L2_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void match2Hamming_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
}
namespace bf_radius_match
{
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);
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);
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);
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);
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);
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);
cudaStream_t stream);
}
}}}
@ -202,7 +202,7 @@ void cv::gpu::BFMatcher_GPU::matchSingle(const GpuMat& query, const GpuMat& trai
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callersL1[] =
{
@ -238,10 +238,7 @@ void cv::gpu::BFMatcher_GPU::matchSingle(const GpuMat& query, const GpuMat& trai
caller_t func = callers[query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, train, mask, trainIdx, distance, cc, StreamAccessor::getStream(stream));
func(query, train, mask, trainIdx, distance, StreamAccessor::getStream(stream));
}
void cv::gpu::BFMatcher_GPU::matchDownload(const GpuMat& trainIdx, const GpuMat& distance, vector<DMatch>& matches)
@ -348,7 +345,7 @@ void cv::gpu::BFMatcher_GPU::matchCollection(const GpuMat& query, const GpuMat&
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callersL1[] =
{
@ -383,10 +380,7 @@ void cv::gpu::BFMatcher_GPU::matchCollection(const GpuMat& query, const GpuMat&
caller_t func = callers[query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, trainCollection, masks, trainIdx, imgIdx, distance, cc, StreamAccessor::getStream(stream));
func(query, trainCollection, masks, trainIdx, imgIdx, distance, StreamAccessor::getStream(stream));
}
void cv::gpu::BFMatcher_GPU::matchDownload(const GpuMat& trainIdx, const GpuMat& imgIdx, const GpuMat& distance, vector<DMatch>& matches)
@ -462,7 +456,7 @@ void cv::gpu::BFMatcher_GPU::knnMatchSingle(const GpuMat& query, const GpuMat& t
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callersL1[] =
{
@ -512,10 +506,7 @@ void cv::gpu::BFMatcher_GPU::knnMatchSingle(const GpuMat& query, const GpuMat& t
caller_t func = callers[query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, train, k, mask, trainIdx, distance, allDist, cc, StreamAccessor::getStream(stream));
func(query, train, k, mask, trainIdx, distance, allDist, StreamAccessor::getStream(stream));
}
void cv::gpu::BFMatcher_GPU::knnMatchDownload(const GpuMat& trainIdx, const GpuMat& distance,
@ -594,7 +585,7 @@ void cv::gpu::BFMatcher_GPU::knnMatch2Collection(const GpuMat& query, const GpuM
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callersL1[] =
{
@ -634,10 +625,7 @@ void cv::gpu::BFMatcher_GPU::knnMatch2Collection(const GpuMat& query, const GpuM
caller_t func = callers[query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, trainCollection, maskCollection, trainIdx, imgIdx, distance, cc, StreamAccessor::getStream(stream));
func(query, trainCollection, maskCollection, trainIdx, imgIdx, distance, StreamAccessor::getStream(stream));
}
void cv::gpu::BFMatcher_GPU::knnMatch2Download(const GpuMat& trainIdx, const GpuMat& imgIdx, const GpuMat& distance,
@ -778,7 +766,7 @@ void cv::gpu::BFMatcher_GPU::radiusMatchSingle(const GpuMat& query, const GpuMat
typedef void (*caller_t)(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);
cudaStream_t stream);
static const caller_t callersL1[] =
{
@ -799,12 +787,6 @@ void cv::gpu::BFMatcher_GPU::radiusMatchSingle(const GpuMat& query, const GpuMat
matchHamming_gpu<int>, 0/*matchHamming_gpu<float>*/
};
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
const int nQuery = query.rows;
const int nTrain = train.rows;
@ -830,7 +812,7 @@ void cv::gpu::BFMatcher_GPU::radiusMatchSingle(const GpuMat& query, const GpuMat
caller_t func = callers[query.depth()];
CV_Assert(func != 0);
func(query, train, maxDistance, mask, trainIdx, distance, nMatches, cc, StreamAccessor::getStream(stream));
func(query, train, maxDistance, mask, trainIdx, distance, nMatches, StreamAccessor::getStream(stream));
}
void cv::gpu::BFMatcher_GPU::radiusMatchDownload(const GpuMat& trainIdx, const GpuMat& distance, const GpuMat& nMatches,
@ -913,7 +895,7 @@ void cv::gpu::BFMatcher_GPU::radiusMatchCollection(const GpuMat& query, GpuMat&
typedef void (*caller_t)(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);
cudaStream_t stream);
static const caller_t callersL1[] =
{
@ -934,12 +916,6 @@ void cv::gpu::BFMatcher_GPU::radiusMatchCollection(const GpuMat& query, GpuMat&
matchHamming_gpu<int>, 0/*matchHamming_gpu<float>*/
};
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
const int nQuery = query.rows;
CV_Assert(query.channels() == 1 && query.depth() < CV_64F);
@ -968,7 +944,7 @@ void cv::gpu::BFMatcher_GPU::radiusMatchCollection(const GpuMat& query, GpuMat&
vector<PtrStepSzb> masks_(masks.begin(), masks.end());
func(query, &trains_[0], static_cast<int>(trains_.size()), maxDistance, masks_.size() == 0 ? 0 : &masks_[0],
trainIdx, imgIdx, distance, nMatches, cc, StreamAccessor::getStream(stream));
trainIdx, imgIdx, distance, nMatches, StreamAccessor::getStream(stream));
}
void cv::gpu::BFMatcher_GPU::radiusMatchDownload(const GpuMat& trainIdx, const GpuMat& imgIdx, const GpuMat& distance, const GpuMat& nMatches,

2
modules/gpu/src/cascadeclassifier.cpp
View File

@ -623,7 +623,7 @@ private:
}
// copy data structures on gpu
stage_mat.upload(cv::Mat(1, stages.size() * sizeof(Stage), CV_8UC1, (uchar*)&(stages[0]) ));
stage_mat.upload(cv::Mat(1, (int) (stages.size() * sizeof(Stage)), CV_8UC1, (uchar*)&(stages[0]) ));
trees_mat.upload(cv::Mat(cl_trees).reshape(1,1));
nodes_mat.upload(cv::Mat(cl_nodes).reshape(1,1));
leaves_mat.upload(cv::Mat(cl_leaves).reshape(1,1));

220
modules/gpu/src/cuda/bf_knnmatch.cu
View File

@ -42,10 +42,13 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/reduce.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/vec_distance.hpp"
#include "opencv2/gpu/device/datamov_utils.hpp"
#include "opencv2/gpu/device/warp_shuffle.hpp"
namespace cv { namespace gpu { namespace device
{
@ -59,6 +62,45 @@ namespace cv { namespace gpu { namespace device
int& bestTrainIdx1, int& bestTrainIdx2,
float* s_distance, int* s_trainIdx)
{
#if __CUDA_ARCH__ >= 300
(void) s_distance;
(void) s_trainIdx;
float d1, d2;
int i1, i2;
#pragma unroll
for (int i = BLOCK_SIZE / 2; i >= 1; i /= 2)
{
d1 = shfl_down(bestDistance1, i, BLOCK_SIZE);
d2 = shfl_down(bestDistance2, i, BLOCK_SIZE);
i1 = shfl_down(bestTrainIdx1, i, BLOCK_SIZE);
i2 = shfl_down(bestTrainIdx2, i, BLOCK_SIZE);
if (bestDistance1 < d1)
{
if (d1 < bestDistance2)
{
bestDistance2 = d1;
bestTrainIdx2 = i1;
}
}
else
{
bestDistance2 = bestDistance1;
bestTrainIdx2 = bestTrainIdx1;
bestDistance1 = d1;
bestTrainIdx1 = i1;
if (d2 < bestDistance2)
{
bestDistance2 = d2;
bestTrainIdx2 = i2;
}
}
}
#else
float myBestDistance1 = numeric_limits<float>::max();
float myBestDistance2 = numeric_limits<float>::max();
int myBestTrainIdx1 = -1;
@ -122,6 +164,7 @@ namespace cv { namespace gpu { namespace device
bestTrainIdx1 = myBestTrainIdx1;
bestTrainIdx2 = myBestTrainIdx2;
#endif
}
template <int BLOCK_SIZE>
@ -130,6 +173,53 @@ namespace cv { namespace gpu { namespace device
int& bestImgIdx1, int& bestImgIdx2,
float* s_distance, int* s_trainIdx, int* s_imgIdx)
{
#if __CUDA_ARCH__ >= 300
(void) s_distance;
(void) s_trainIdx;
(void) s_imgIdx;
float d1, d2;
int i1, i2;
int j1, j2;
#pragma unroll
for (int i = BLOCK_SIZE / 2; i >= 1; i /= 2)
{
d1 = shfl_down(bestDistance1, i, BLOCK_SIZE);
d2 = shfl_down(bestDistance2, i, BLOCK_SIZE);
i1 = shfl_down(bestTrainIdx1, i, BLOCK_SIZE);
i2 = shfl_down(bestTrainIdx2, i, BLOCK_SIZE);
j1 = shfl_down(bestImgIdx1, i, BLOCK_SIZE);
j2 = shfl_down(bestImgIdx2, i, BLOCK_SIZE);
if (bestDistance1 < d1)
{
if (d1 < bestDistance2)
{
bestDistance2 = d1;
bestTrainIdx2 = i1;
bestImgIdx2 = j1;
}
}
else
{
bestDistance2 = bestDistance1;
bestTrainIdx2 = bestTrainIdx1;
bestImgIdx2 = bestImgIdx1;
bestDistance1 = d1;
bestTrainIdx1 = i1;
bestImgIdx1 = j1;
if (d2 < bestDistance2)
{
bestDistance2 = d2;
bestTrainIdx2 = i2;
bestImgIdx2 = j2;
}
}
}
#else
float myBestDistance1 = numeric_limits<float>::max();
float myBestDistance2 = numeric_limits<float>::max();
int myBestTrainIdx1 = -1;
@ -205,6 +295,7 @@ namespace cv { namespace gpu { namespace device
bestImgIdx1 = myBestImgIdx1;
bestImgIdx2 = myBestImgIdx2;
#endif
}
///////////////////////////////////////////////////////////////////////////////
@ -748,9 +839,8 @@ namespace cv { namespace gpu { namespace device
template <typename Dist, typename T, typename Mask>
void match2Dispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>& train, const Mask& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolledCached<16, 64, Dist>(query, train, mask, static_cast< PtrStepSz<int2> >(trainIdx), static_cast< PtrStepSz<float2> > (distance), stream);
@ -780,9 +870,8 @@ namespace cv { namespace gpu { namespace device
template <typename Dist, typename T, typename Mask>
void match2Dispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, const Mask& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolledCached<16, 64, Dist>(query, trains, n, mask, static_cast< PtrStepSz<int2> >(trainIdx), static_cast< PtrStepSz<int2> >(imgIdx), static_cast< PtrStepSz<float2> > (distance), stream);
@ -945,9 +1034,8 @@ namespace cv { namespace gpu { namespace device
template <typename Dist, typename T, typename Mask>
void calcDistanceDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>& train, const Mask& mask,
const PtrStepSzf& allDist,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
calcDistanceUnrolled<16, 64, Dist>(query, train, mask, allDist, stream);
@ -1005,7 +1093,7 @@ namespace cv { namespace gpu { namespace device
s_trainIdx[threadIdx.x] = bestIdx;
__syncthreads();
reducePredVal<BLOCK_SIZE>(s_dist, dist, s_trainIdx, bestIdx, threadIdx.x, less<volatile float>());
reduceKeyVal<BLOCK_SIZE>(s_dist, dist, s_trainIdx, bestIdx, threadIdx.x, less<float>());
if (threadIdx.x == 0)
{
@ -1034,7 +1122,7 @@ namespace cv { namespace gpu { namespace device
cudaSafeCall( cudaDeviceSynchronize() );
}
void findKnnMatchDispatcher(int k, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream)
void findKnnMatchDispatcher(int k, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream)
{
findKnnMatch<256>(k, static_cast<PtrStepSzi>(trainIdx), static_cast<PtrStepSzf>(distance), allDist, stream);
}
@ -1045,16 +1133,16 @@ namespace cv { namespace gpu { namespace device
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>& train, int k, const Mask& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (k == 2)
{
match2Dispatcher<Dist>(query, train, mask, trainIdx, distance, cc, stream);
match2Dispatcher<Dist>(query, train, mask, trainIdx, distance, stream);
}
else
{
calcDistanceDispatcher<Dist>(query, train, mask, allDist, cc, stream);
findKnnMatchDispatcher(k, trainIdx, distance, allDist, cc, stream);
calcDistanceDispatcher<Dist>(query, train, mask, allDist, stream);
findKnnMatchDispatcher(k, trainIdx, distance, allDist, stream);
}
}
@ -1063,105 +1151,105 @@ namespace cv { namespace gpu { namespace device
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (mask.data)
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, SingleMask(mask), trainIdx, distance, allDist, cc, stream);
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, SingleMask(mask), trainIdx, distance, allDist, stream);
else
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, WithOutMask(), trainIdx, distance, allDist, cc, stream);
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, WithOutMask(), trainIdx, distance, allDist, stream);
}
template void matchL1_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchL1_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (mask.data)
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, SingleMask(mask), trainIdx, distance, allDist, cc, stream);
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, SingleMask(mask), trainIdx, distance, allDist, stream);
else
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, WithOutMask(), trainIdx, distance, allDist, cc, stream);
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, WithOutMask(), trainIdx, distance, allDist, stream);
}
//template void matchL2_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchL2_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (mask.data)
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, SingleMask(mask), trainIdx, distance, allDist, cc, stream);
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, SingleMask(mask), trainIdx, distance, allDist, stream);
else
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, WithOutMask(), trainIdx, distance, allDist, cc, stream);
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), k, WithOutMask(), trainIdx, distance, allDist, stream);
}
template void matchHamming_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, int cc, cudaStream_t stream);
template void matchHamming_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, int k, const PtrStepSzb& mask, const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist, cudaStream_t stream);
template <typename T> void match2L1_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (masks.data)
match2Dispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data), trainIdx, imgIdx, distance, cc, stream);
match2Dispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data), trainIdx, imgIdx, distance, stream);
else
match2Dispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(), trainIdx, imgIdx, distance, cc, stream);
match2Dispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(), trainIdx, imgIdx, distance, stream);
}
template void match2L1_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2L1_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2L1_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2L1_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2L1_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2L1_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2L1_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
//template void match2L1_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template void match2L1_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template void match2L1_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template void match2L1_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template void match2L1_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template <typename T> void match2L2_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (masks.data)
match2Dispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data), trainIdx, imgIdx, distance, cc, stream);
match2Dispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data), trainIdx, imgIdx, distance, stream);
else
match2Dispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(), trainIdx, imgIdx, distance, cc, stream);
match2Dispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(), trainIdx, imgIdx, distance, stream);
}
//template void match2L2_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2L2_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2L2_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2L2_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2L2_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2L2_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2L2_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
//template void match2L2_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
//template void match2L2_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
//template void match2L2_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
//template void match2L2_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template void match2L2_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template <typename T> void match2Hamming_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (masks.data)
match2Dispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data), trainIdx, imgIdx, distance, cc, stream);
match2Dispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data), trainIdx, imgIdx, distance, stream);
else
match2Dispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(), trainIdx, imgIdx, distance, cc, stream);
match2Dispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(), trainIdx, imgIdx, distance, stream);
}
template void match2Hamming_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2Hamming_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2Hamming_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
//template void match2Hamming_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2Hamming_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, int cc, cudaStream_t stream);
template void match2Hamming_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
//template void match2Hamming_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template void match2Hamming_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
//template void match2Hamming_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
template void match2Hamming_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance, cudaStream_t stream);
} // namespace bf_knnmatch
}}} // namespace cv { namespace gpu { namespace device {
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

131
modules/gpu/src/cuda/bf_match.cu
View File

@ -42,7 +42,9 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/reduce.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/vec_distance.hpp"
#include "opencv2/gpu/device/datamov_utils.hpp"
@ -60,12 +62,7 @@ namespace cv { namespace gpu { namespace device
s_distance += threadIdx.y * BLOCK_SIZE;
s_trainIdx += threadIdx.y * BLOCK_SIZE;
s_distance[threadIdx.x] = bestDistance;
s_trainIdx[threadIdx.x] = bestTrainIdx;
__syncthreads();
reducePredVal<BLOCK_SIZE>(s_distance, bestDistance, s_trainIdx, bestTrainIdx, threadIdx.x, less<volatile float>());
reduceKeyVal<BLOCK_SIZE>(s_distance, bestDistance, s_trainIdx, bestTrainIdx, threadIdx.x, less<float>());
}
template <int BLOCK_SIZE>
@ -75,13 +72,7 @@ namespace cv { namespace gpu { namespace device
s_trainIdx += threadIdx.y * BLOCK_SIZE;
s_imgIdx += threadIdx.y * BLOCK_SIZE;
s_distance[threadIdx.x] = bestDistance;
s_trainIdx[threadIdx.x] = bestTrainIdx;
s_imgIdx [threadIdx.x] = bestImgIdx;
__syncthreads();
reducePredVal2<BLOCK_SIZE>(s_distance, bestDistance, s_trainIdx, bestTrainIdx, s_imgIdx, bestImgIdx, threadIdx.x, less<volatile float>());
reduceKeyVal<BLOCK_SIZE>(s_distance, bestDistance, smem_tuple(s_trainIdx, s_imgIdx), thrust::tie(bestTrainIdx, bestImgIdx), threadIdx.x, less<float>());
}
///////////////////////////////////////////////////////////////////////////////
@ -567,9 +558,8 @@ namespace cv { namespace gpu { namespace device
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>& train, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolledCached<16, 64, Dist>(query, train, mask, trainIdx, distance, stream);
@ -599,9 +589,8 @@ namespace cv { namespace gpu { namespace device
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const PtrStepSz<T>& query, const PtrStepSz<T>* trains, int n, const Mask& mask,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolledCached<16, 64, Dist>(query, trains, n, mask, trainIdx, imgIdx, distance, stream);
@ -633,153 +622,153 @@ namespace cv { namespace gpu { namespace device
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), SingleMask(mask),
trainIdx, distance,
cc, stream);
stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), WithOutMask(),
trainIdx, distance,
cc, stream);
stream);
}
}
template void matchL1_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), SingleMask(mask),
trainIdx, distance,
cc, stream);
stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), WithOutMask(),
trainIdx, distance,
cc, stream);
stream);
}
}
//template void matchL2_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), SingleMask(mask),
trainIdx, distance,
cc, stream);
stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), WithOutMask(),
trainIdx, distance,
cc, stream);
stream);
}
}
template void matchHamming_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchHamming_gpu<uchar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& queryDescs, const PtrStepSzb& trainDescs, const PtrStepSzb& mask, const PtrStepSzi& trainIdx, const PtrStepSzf& distance, cudaStream_t stream);
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (masks.data)
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data),
trainIdx, imgIdx, distance,
cc, stream);
stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(),
trainIdx, imgIdx, distance,
cc, stream);
stream);
}
}
template void matchL1_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL1_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL1_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL1_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (masks.data)
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data),
trainIdx, imgIdx, distance,
cc, stream);
stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(),
trainIdx, imgIdx, distance,
cc, stream);
stream);
}
}
//template void matchL2_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& maskCollection, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchL2_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchL2_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchL2_gpu<float >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& maskCollection, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream)
cudaStream_t stream)
{
if (masks.data)
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, MaskCollection(masks.data),
trainIdx, imgIdx, distance,
cc, stream);
stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains.ptr(), trains.cols, WithOutMask(),
trainIdx, imgIdx, distance,
cc, stream);
stream);
}
}
template void matchHamming_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, int cc, cudaStream_t stream);
template void matchHamming_gpu<uchar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
//template void matchHamming_gpu<short >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
template void matchHamming_gpu<int >(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks, const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, cudaStream_t stream);
} // namespace bf_match
}}} // namespace cv { namespace gpu { namespace device {
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

117
modules/gpu/src/cuda/bf_radius_match.cu
View File

@ -42,7 +42,8 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/vec_distance.hpp"
#include "opencv2/gpu/device/datamov_utils.hpp"
@ -58,8 +59,6 @@ namespace cv { namespace gpu { namespace device
__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;
@ -110,8 +109,6 @@ namespace cv { namespace gpu { namespace device
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T, typename Mask>
@ -170,8 +167,6 @@ namespace cv { namespace gpu { namespace device
__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;
@ -221,8 +216,6 @@ namespace cv { namespace gpu { namespace device
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
template <int BLOCK_SIZE, typename Dist, typename T, typename Mask>
@ -281,9 +274,8 @@ namespace cv { namespace gpu { namespace device
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)
cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
@ -313,9 +305,8 @@ namespace cv { namespace gpu { namespace device
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)
cudaStream_t stream)
{
(void)cc;
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
@ -347,126 +338,126 @@ namespace cv { namespace gpu { namespace device
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)
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);
stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
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 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, 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, 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, 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, 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, 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, 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)
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);
stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
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 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, 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, 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, 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, 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, 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, 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)
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);
stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), static_cast< PtrStepSz<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
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 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, 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, 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, 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, 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, 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)
cudaStream_t stream)
{
matchDispatcher< L1Dist<T> >(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
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 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, 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, 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, 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, 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, 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, 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)
cudaStream_t stream)
{
matchDispatcher<L2Dist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
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 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, 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, 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, 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, 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, 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, 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)
cudaStream_t stream)
{
matchDispatcher<HammingDist>(static_cast< PtrStepSz<T> >(query), (const PtrStepSz<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
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);
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, 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, 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, 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, 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, cudaStream_t stream);
} // namespace bf_radius_match
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

27
modules/gpu/src/cuda/calib3d.cu
View File

@ -42,9 +42,10 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/reduce.hpp"
namespace cv { namespace gpu { namespace device
{
@ -66,6 +67,8 @@ namespace cv { namespace gpu { namespace device
crot1.x * p.x + crot1.y * p.y + crot1.z * p.z + ctransl.y,
crot2.x * p.x + crot2.y * p.y + crot2.z * p.z + ctransl.z);
}
__device__ __forceinline__ TransformOp() {}
__device__ __forceinline__ TransformOp(const TransformOp&) {}
};
void call(const PtrStepSz<float3> src, const float* rot,
@ -103,6 +106,8 @@ namespace cv { namespace gpu { namespace device
(cproj0.x * t.x + cproj0.y * t.y) / t.z + cproj0.z,
(cproj1.x * t.x + cproj1.y * t.y) / t.z + cproj1.z);
}
__device__ __forceinline__ ProjectOp() {}
__device__ __forceinline__ ProjectOp(const ProjectOp&) {}
};
void call(const PtrStepSz<float3> src, const float* rot,
@ -134,6 +139,7 @@ namespace cv { namespace gpu { namespace device
return x * x;
}
template <int BLOCK_SIZE>
__global__ void computeHypothesisScoresKernel(
const int num_points, const float3* object, const float2* image,
const float dist_threshold, int* g_num_inliers)
@ -156,19 +162,11 @@ namespace cv { namespace gpu { namespace device
++num_inliers;
}
extern __shared__ float s_num_inliers[];
s_num_inliers[threadIdx.x] = num_inliers;
__syncthreads();
for (int step = blockDim.x / 2; step > 0; step >>= 1)
{
if (threadIdx.x < step)
s_num_inliers[threadIdx.x] += s_num_inliers[threadIdx.x + step];
__syncthreads();
}
__shared__ int s_num_inliers[BLOCK_SIZE];
reduce<BLOCK_SIZE>(s_num_inliers, num_inliers, threadIdx.x, plus<int>());
if (threadIdx.x == 0)
g_num_inliers[blockIdx.x] = s_num_inliers[0];
g_num_inliers[blockIdx.x] = num_inliers;
}
void computeHypothesisScores(
@ -181,9 +179,8 @@ namespace cv { namespace gpu { namespace device
dim3 threads(256);
dim3 grid(num_hypotheses);
int smem_size = threads.x * sizeof(float);
computeHypothesisScoresKernel<<<grid, threads, smem_size>>>(
computeHypothesisScoresKernel<256><<<grid, threads>>>(
num_points, object, image, dist_threshold, hypothesis_scores);
cudaSafeCall( cudaGetLastError() );
@ -193,4 +190,4 @@ namespace cv { namespace gpu { namespace device
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

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

@ -43,459 +43,451 @@
#if !defined CUDA_DISABLER
#include <utility>
#include <algorithm>
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/emulation.hpp"
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/utility.hpp"
using namespace cv::gpu;
using namespace cv::gpu::device;
namespace canny
{
struct L1 : binary_function<int, int, float>
{
__device__ __forceinline__ float operator ()(int x, int y) const
{
return ::abs(x) + ::abs(y);
}
__device__ __forceinline__ L1() {}
__device__ __forceinline__ L1(const L1&) {}
};
struct L2 : binary_function<int, int, float>
{
__device__ __forceinline__ float operator ()(int x, int y) const
{
return ::sqrtf(x * x + y * y);
}
__device__ __forceinline__ L2() {}
__device__ __forceinline__ L2(const L2&) {}
};
}
namespace cv { namespace gpu { namespace device
{
namespace canny
template <> struct TransformFunctorTraits<canny::L1> : DefaultTransformFunctorTraits<canny::L1>
{
__global__ void calcSobelRowPass(const PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
enum { smart_shift = 4 };
};
template <> struct TransformFunctorTraits<canny::L2> : DefaultTransformFunctorTraits<canny::L2>
{
enum { smart_shift = 4 };
};
}}}
namespace canny
{
texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_src(false, cudaFilterModePoint, cudaAddressModeClamp);
struct SrcTex
{
const int xoff;
const int yoff;
__host__ SrcTex(int _xoff, int _yoff) : xoff(_xoff), yoff(_yoff) {}
__device__ __forceinline__ int operator ()(int y, int x) const
{
__shared__ int smem[16][18];
return tex2D(tex_src, x + xoff, y + yoff);
}
};
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
template <class Norm> __global__
void calcMagnitudeKernel(const SrcTex src, PtrStepi dx, PtrStepi dy, PtrStepSzf mag, const Norm norm)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (i < rows)
{
smem[threadIdx.y][threadIdx.x + 1] = src.ptr(i)[j];
if (threadIdx.x == 0)
{
smem[threadIdx.y][0] = src.ptr(i)[::max(j - 1, 0)];
smem[threadIdx.y][17] = src.ptr(i)[::min(j + 16, cols - 1)];
}
__syncthreads();
if (y >= mag.rows || x >= mag.cols)
return;
if (j < cols)
{
dx_buf.ptr(i)[j] = -smem[threadIdx.y][threadIdx.x] + smem[threadIdx.y][threadIdx.x + 2];
dy_buf.ptr(i)[j] = smem[threadIdx.y][threadIdx.x] + 2 * smem[threadIdx.y][threadIdx.x + 1] + smem[threadIdx.y][threadIdx.x + 2];
}
}
int dxVal = (src(y - 1, x + 1) + 2 * src(y, x + 1) + src(y + 1, x + 1)) - (src(y - 1, x - 1) + 2 * src(y, x - 1) + src(y + 1, x - 1));
int dyVal = (src(y + 1, x - 1) + 2 * src(y + 1, x) + src(y + 1, x + 1)) - (src(y - 1, x - 1) + 2 * src(y - 1, x) + src(y - 1, x + 1));
dx(y, x) = dxVal;
dy(y, x) = dyVal;
mag(y, x) = norm(dxVal, dyVal);
}
void calcMagnitude(PtrStepSzb srcWhole, int xoff, int yoff, PtrStepSzi dx, PtrStepSzi dy, PtrStepSzf mag, bool L2Grad)
{
const dim3 block(16, 16);
const dim3 grid(divUp(mag.cols, block.x), divUp(mag.rows, block.y));
bindTexture(&tex_src, srcWhole);
SrcTex src(xoff, yoff);
if (L2Grad)
{
L2 norm;
calcMagnitudeKernel<<<grid, block>>>(src, dx, dy, mag, norm);
}
else
{
L1 norm;
calcMagnitudeKernel<<<grid, block>>>(src, dx, dy, mag, norm);
}
void calcSobelRowPass_gpu(PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
void calcMagnitude(PtrStepSzi dx, PtrStepSzi dy, PtrStepSzf mag, bool L2Grad)
{
if (L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
calcSobelRowPass<<<grid, block>>>(src, dx_buf, dy_buf, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
L2 norm;
transform(dx, dy, mag, norm, WithOutMask(), 0);
}
struct L1
else
{
static __device__ __forceinline__ float calc(int x, int y)
{
return ::abs(x) + ::abs(y);
}
};
struct L2
{
static __device__ __forceinline__ float calc(int x, int y)
{
return ::sqrtf(x * x + y * y);
}
};
template <typename Norm> __global__ void calcMagnitude(const PtrStepi dx_buf, const PtrStepi dy_buf,
PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols)
{
__shared__ int sdx[18][16];
__shared__ int sdy[18][16];
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (j < cols)
{
sdx[threadIdx.y + 1][threadIdx.x] = dx_buf.ptr(i)[j];
sdy[threadIdx.y + 1][threadIdx.x] = dy_buf.ptr(i)[j];
if (threadIdx.y == 0)
{
sdx[0][threadIdx.x] = dx_buf.ptr(::max(i - 1, 0))[j];
sdx[17][threadIdx.x] = dx_buf.ptr(::min(i + 16, rows - 1))[j];
sdy[0][threadIdx.x] = dy_buf.ptr(::max(i - 1, 0))[j];
sdy[17][threadIdx.x] = dy_buf.ptr(::min(i + 16, rows - 1))[j];
}
__syncthreads();
if (i < rows)
{
int x = sdx[threadIdx.y][threadIdx.x] + 2 * sdx[threadIdx.y + 1][threadIdx.x] + sdx[threadIdx.y + 2][threadIdx.x];
int y = -sdy[threadIdx.y][threadIdx.x] + sdy[threadIdx.y + 2][threadIdx.x];
dx.ptr(i)[j] = x;
dy.ptr(i)[j] = y;
mag.ptr(i + 1)[j + 1] = Norm::calc(x, y);
}
}
L1 norm;
transform(dx, dy, mag, norm, WithOutMask(), 0);
}
}
}
void calcMagnitude_gpu(PtrStepi dx_buf, PtrStepi dy_buf, PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
//////////////////////////////////////////////////////////////////////////////////////////
namespace canny
{
texture<float, cudaTextureType2D, cudaReadModeElementType> tex_mag(false, cudaFilterModePoint, cudaAddressModeClamp);
__global__ void calcMapKernel(const PtrStepSzi dx, const PtrStepi dy, PtrStepi map, const float low_thresh, const float high_thresh)
{
const int CANNY_SHIFT = 15;
const int TG22 = (int)(0.4142135623730950488016887242097*(1<<CANNY_SHIFT) + 0.5);
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x == 0 || x >= dx.cols - 1 || y == 0 || y >= dx.rows - 1)
return;
int dxVal = dx(y, x);
int dyVal = dy(y, x);
const int s = (dxVal ^ dyVal) < 0 ? -1 : 1;
const float m = tex2D(tex_mag, x, y);
dxVal = ::abs(dxVal);
dyVal = ::abs(dyVal);
// 0 - the pixel can not belong to an edge
// 1 - the pixel might belong to an edge
// 2 - the pixel does belong to an edge
int edge_type = 0;
if (m > low_thresh)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
const int tg22x = dxVal * TG22;
const int tg67x = tg22x + ((dxVal + dxVal) << CANNY_SHIFT);
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
dyVal <<= CANNY_SHIFT;
if (dyVal < tg22x)
{
if (m > tex2D(tex_mag, x - 1, y) && m >= tex2D(tex_mag, x + 1, y))
edge_type = 1 + (int)(m > high_thresh);
}
else if(dyVal > tg67x)
{
if (m > tex2D(tex_mag, x, y - 1) && m >= tex2D(tex_mag, x, y + 1))
edge_type = 1 + (int)(m > high_thresh);
}
else
calcMagnitude<L1><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
{
if (m > tex2D(tex_mag, x - s, y - 1) && m >= tex2D(tex_mag, x + s, y + 1))
edge_type = 1 + (int)(m > high_thresh);
}
}
template <typename Norm> __global__ void calcMagnitude(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols)
{
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
map(y, x) = edge_type;
}
if (i < rows && j < cols)
mag.ptr(i + 1)[j + 1] = Norm::calc(dx.ptr(i)[j], dy.ptr(i)[j]);
void calcMap(PtrStepSzi dx, PtrStepSzi dy, PtrStepSzf mag, PtrStepSzi map, float low_thresh, float high_thresh)
{
const dim3 block(16, 16);
const dim3 grid(divUp(dx.cols, block.x), divUp(dx.rows, block.y));
bindTexture(&tex_mag, mag);
calcMapKernel<<<grid, block>>>(dx, dy, map, low_thresh, high_thresh);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
}
//////////////////////////////////////////////////////////////////////////////////////////
namespace canny
{
__device__ int counter = 0;
__global__ void edgesHysteresisLocalKernel(PtrStepSzi map, ushort2* st)
{
__shared__ volatile int smem[18][18];
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
smem[threadIdx.y + 1][threadIdx.x + 1] = x < map.cols && y < map.rows ? map(y, x) : 0;
if (threadIdx.y == 0)
smem[0][threadIdx.x + 1] = y > 0 ? map(y - 1, x) : 0;
if (threadIdx.y == blockDim.y - 1)
smem[blockDim.y + 1][threadIdx.x + 1] = y + 1 < map.rows ? map(y + 1, x) : 0;
if (threadIdx.x == 0)
smem[threadIdx.y + 1][0] = x > 0 ? map(y, x - 1) : 0;
if (threadIdx.x == blockDim.x - 1)
smem[threadIdx.y + 1][blockDim.x + 1] = x + 1 < map.cols ? map(y, x + 1) : 0;
if (threadIdx.x == 0 && threadIdx.y == 0)
smem[0][0] = y > 0 && x > 0 ? map(y - 1, x - 1) : 0;
if (threadIdx.x == blockDim.x - 1 && threadIdx.y == 0)
smem[0][blockDim.x + 1] = y > 0 && x + 1 < map.cols ? map(y - 1, x + 1) : 0;
if (threadIdx.x == 0 && threadIdx.y == blockDim.y - 1)
smem[blockDim.y + 1][0] = y + 1 < map.rows && x > 0 ? map(y + 1, x - 1) : 0;
if (threadIdx.x == blockDim.x - 1 && threadIdx.y == blockDim.y - 1)
smem[blockDim.y + 1][blockDim.x + 1] = y + 1 < map.rows && x + 1 < map.cols ? map(y + 1, x + 1) : 0;
__syncthreads();
if (x >= map.cols || y >= map.rows)
return;
int n;
#pragma unroll
for (int k = 0; k < 16; ++k)
{
n = 0;
if (smem[threadIdx.y + 1][threadIdx.x + 1] == 1)
{
n += smem[threadIdx.y ][threadIdx.x ] == 2;
n += smem[threadIdx.y ][threadIdx.x + 1] == 2;
n += smem[threadIdx.y ][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 1][threadIdx.x ] == 2;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 2][threadIdx.x ] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 2;
}
if (n > 0)
smem[threadIdx.y + 1][threadIdx.x + 1] = 2;
}
void calcMagnitude_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
const int e = smem[threadIdx.y + 1][threadIdx.x + 1];
map(y, x) = e;
n = 0;
if (e == 2)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
n += smem[threadIdx.y ][threadIdx.x ] == 1;
n += smem[threadIdx.y ][threadIdx.x + 1] == 1;
n += smem[threadIdx.y ][threadIdx.x + 2] == 1;
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx, dy, mag, rows, cols);
n += smem[threadIdx.y + 1][threadIdx.x ] == 1;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 1;
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
n += smem[threadIdx.y + 2][threadIdx.x ] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 1;
}
//////////////////////////////////////////////////////////////////////////////////////////
#define CANNY_SHIFT 15
#define TG22 (int)(0.4142135623730950488016887242097*(1<<CANNY_SHIFT) + 0.5)
__global__ void calcMap(const PtrStepi dx, const PtrStepi dy, const PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
if (n > 0)
{
__shared__ float smem[18][18];
const int ind = ::atomicAdd(&counter, 1);
st[ind] = make_ushort2(x, y);
}
}
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
void edgesHysteresisLocal(PtrStepSzi map, ushort2* st1)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, counter) );
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(int)) );
if (ly < 14)
smem[ly][lx] = mag.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
const dim3 block(16, 16);
const dim3 grid(divUp(map.cols, block.x), divUp(map.rows, block.y));
if (ly < 4 && blockIdx.y * 16 + ly + 14 <= rows && blockIdx.x * 16 + lx <= cols)
smem[ly + 14][lx] = mag.ptr(blockIdx.y * 16 + ly + 14)[blockIdx.x * 16 + lx];
edgesHysteresisLocalKernel<<<grid, block>>>(map, st1);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
}
//////////////////////////////////////////////////////////////////////////////////////////
namespace canny
{
__constant__ int c_dx[8] = {-1, 0, 1, -1, 1, -1, 0, 1};
__constant__ int c_dy[8] = {-1, -1, -1, 0, 0, 1, 1, 1};
__global__ void edgesHysteresisGlobalKernel(PtrStepSzi map, ushort2* st1, ushort2* st2, const int count)
{
const int stack_size = 512;
__shared__ int s_counter;
__shared__ int s_ind;
__shared__ ushort2 s_st[stack_size];
if (threadIdx.x == 0)
s_counter = 0;
__syncthreads();
int ind = blockIdx.y * gridDim.x + blockIdx.x;
if (ind >= count)
return;
ushort2 pos = st1[ind];
if (threadIdx.x < 8)
{
pos.x += c_dx[threadIdx.x];
pos.y += c_dy[threadIdx.x];
if (pos.x > 0 && pos.x < map.cols && pos.y > 0 && pos.y < map.rows && map(pos.y, pos.x) == 1)
{
map(pos.y, pos.x) = 2;
ind = Emulation::smem::atomicAdd(&s_counter, 1);
s_st[ind] = pos;
}
}
__syncthreads();
while (s_counter > 0 && s_counter <= stack_size - blockDim.x)
{
const int subTaskIdx = threadIdx.x >> 3;
const int portion = ::min(s_counter, blockDim.x >> 3);
if (subTaskIdx < portion)
pos = s_st[s_counter - 1 - subTaskIdx];
__syncthreads();
if (i < rows && j < cols)
{
int x = dx.ptr(i)[j];
int y = dy.ptr(i)[j];
const int s = (x ^ y) < 0 ? -1 : 1;
const float m = smem[threadIdx.y + 1][threadIdx.x + 1];
x = ::abs(x);
y = ::abs(y);
// 0 - the pixel can not belong to an edge
// 1 - the pixel might belong to an edge
// 2 - the pixel does belong to an edge
int edge_type = 0;
if (m > low_thresh)
{
const int tg22x = x * TG22;
const int tg67x = tg22x + ((x + x) << CANNY_SHIFT);
y <<= CANNY_SHIFT;
if (y < tg22x)
{
if (m > smem[threadIdx.y + 1][threadIdx.x] && m >= smem[threadIdx.y + 1][threadIdx.x + 2])
edge_type = 1 + (int)(m > high_thresh);
}
else if( y > tg67x )
{
if (m > smem[threadIdx.y][threadIdx.x + 1] && m >= smem[threadIdx.y + 2][threadIdx.x + 1])
edge_type = 1 + (int)(m > high_thresh);
}
else
{
if (m > smem[threadIdx.y][threadIdx.x + 1 - s] && m > smem[threadIdx.y + 2][threadIdx.x + 1 + s])
edge_type = 1 + (int)(m > high_thresh);
}
}
map.ptr(i + 1)[j + 1] = edge_type;
}
}
#undef CANNY_SHIFT
#undef TG22
void calcMap_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
calcMap<<<grid, block>>>(dx, dy, mag, map, rows, cols, low_thresh, high_thresh);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
//////////////////////////////////////////////////////////////////////////////////////////
__device__ unsigned int counter = 0;
__global__ void edgesHysteresisLocal(PtrStepi map, ushort2* st, int rows, int cols)
{
#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ >= 120)
__shared__ int smem[18][18];
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
if (ly < 14)
smem[ly][lx] = map.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
if (ly < 4 && blockIdx.y * 16 + ly + 14 <= rows && blockIdx.x * 16 + lx <= cols)
smem[ly + 14][lx] = map.ptr(blockIdx.y * 16 + ly + 14)[blockIdx.x * 16 + lx];
__syncthreads();
if (i < rows && j < cols)
{
int n;
#pragma unroll
for (int k = 0; k < 16; ++k)
{
n = 0;
if (smem[threadIdx.y + 1][threadIdx.x + 1] == 1)
{
n += smem[threadIdx.y ][threadIdx.x ] == 2;
n += smem[threadIdx.y ][threadIdx.x + 1] == 2;
n += smem[threadIdx.y ][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 1][threadIdx.x ] == 2;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 2][threadIdx.x ] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 2;
}
if (n > 0)
smem[threadIdx.y + 1][threadIdx.x + 1] = 2;
}
const int e = smem[threadIdx.y + 1][threadIdx.x + 1];
map.ptr(i + 1)[j + 1] = e;
n = 0;
if (e == 2)
{
n += smem[threadIdx.y ][threadIdx.x ] == 1;
n += smem[threadIdx.y ][threadIdx.x + 1] == 1;
n += smem[threadIdx.y ][threadIdx.x + 2] == 1;
n += smem[threadIdx.y + 1][threadIdx.x ] == 1;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 1;
n += smem[threadIdx.y + 2][threadIdx.x ] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 1;
}
if (n > 0)
{
const unsigned int ind = atomicInc(&counter, (unsigned int)(-1));
st[ind] = make_ushort2(j + 1, i + 1);
}
}
#endif
}
void edgesHysteresisLocal_gpu(PtrStepi map, ushort2* st1, int rows, int cols)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, counter) );
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(unsigned int)) );
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
edgesHysteresisLocal<<<grid, block>>>(map, st1, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
__constant__ int c_dx[8] = {-1, 0, 1, -1, 1, -1, 0, 1};
__constant__ int c_dy[8] = {-1, -1, -1, 0, 0, 1, 1, 1};
__global__ void edgesHysteresisGlobal(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols, int count)
{
#if defined (__CUDA_ARCH__) && __CUDA_ARCH__ >= 120
const int stack_size = 512;
__shared__ unsigned int s_counter;
__shared__ unsigned int s_ind;
__shared__ ushort2 s_st[stack_size];
if (threadIdx.x == 0)
s_counter = 0;
s_counter -= portion;
__syncthreads();
int ind = blockIdx.y * gridDim.x + blockIdx.x;
if (ind < count)
if (subTaskIdx < portion)
{
ushort2 pos = st1[ind];
pos.x += c_dx[threadIdx.x & 7];
pos.y += c_dy[threadIdx.x & 7];
if (pos.x > 0 && pos.x <= cols && pos.y > 0 && pos.y <= rows)
if (pos.x > 0 && pos.x < map.cols && pos.y > 0 && pos.y < map.rows && map(pos.y, pos.x) == 1)
{
if (threadIdx.x < 8)
{
pos.x += c_dx[threadIdx.x];
pos.y += c_dy[threadIdx.x];
map(pos.y, pos.x) = 2;
if (map.ptr(pos.y)[pos.x] == 1)
{
map.ptr(pos.y)[pos.x] = 2;
ind = Emulation::smem::atomicAdd(&s_counter, 1);
ind = atomicInc(&s_counter, (unsigned int)(-1));
s_st[ind] = pos;
}
}
__syncthreads();
while (s_counter > 0 && s_counter <= stack_size - blockDim.x)
{
const int subTaskIdx = threadIdx.x >> 3;
const int portion = ::min(s_counter, blockDim.x >> 3);
pos.x = pos.y = 0;
if (subTaskIdx < portion)
pos = s_st[s_counter - 1 - subTaskIdx];
__syncthreads();
if (threadIdx.x == 0)
s_counter -= portion;
__syncthreads();
if (pos.x > 0 && pos.x <= cols && pos.y > 0 && pos.y <= rows)
{
pos.x += c_dx[threadIdx.x & 7];
pos.y += c_dy[threadIdx.x & 7];
if (map.ptr(pos.y)[pos.x] == 1)
{
map.ptr(pos.y)[pos.x] = 2;
ind = atomicInc(&s_counter, (unsigned int)(-1));
s_st[ind] = pos;
}
}
__syncthreads();
}
if (s_counter > 0)
{
if (threadIdx.x == 0)
{
ind = atomicAdd(&counter, s_counter);
s_ind = ind - s_counter;
}
__syncthreads();
ind = s_ind;
for (int i = threadIdx.x; i < s_counter; i += blockDim.x)
{
st2[ind + i] = s_st[i];
}
}
s_st[ind] = pos;
}
}
#endif
__syncthreads();
}
void edgesHysteresisGlobal_gpu(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols)
if (s_counter > 0)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, counter) );
unsigned int count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
while (count > 0)
if (threadIdx.x == 0)
{
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(unsigned int)) );
dim3 block(128, 1, 1);
dim3 grid(std::min(count, 65535u), divUp(count, 65535), 1);
edgesHysteresisGlobal<<<grid, block>>>(map, st1, st2, rows, cols, count);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
std::swap(st1, st2);
ind = ::atomicAdd(&counter, s_counter);
s_ind = ind - s_counter;
}
__syncthreads();
ind = s_ind;
for (int i = threadIdx.x; i < s_counter; i += blockDim.x)
st2[ind + i] = s_st[i];
}
}
__global__ void getEdges(PtrStepi map, PtrStepb dst, int rows, int cols)
void edgesHysteresisGlobal(PtrStepSzi map, ushort2* st1, ushort2* st2)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, canny::counter) );
int count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(int), cudaMemcpyDeviceToHost) );
while (count > 0)
{
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(int)) );
if (i < rows && j < cols)
dst.ptr(i)[j] = (uchar)(-(map.ptr(i + 1)[j + 1] >> 1));
}
const dim3 block(128);
const dim3 grid(::min(count, 65535u), divUp(count, 65535), 1);
void getEdges_gpu(PtrStepi map, PtrStepb dst, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
getEdges<<<grid, block>>>(map, dst, rows, cols);
edgesHysteresisGlobalKernel<<<grid, block>>>(map, st1, st2, count);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(int), cudaMemcpyDeviceToHost) );
std::swap(st1, st2);
}
} // namespace canny
}}} // namespace cv { namespace gpu { namespace device
}
}
//////////////////////////////////////////////////////////////////////////////////////////
#endif /* CUDA_DISABLER */
namespace canny
{
struct GetEdges : unary_function<int, uchar>
{
__device__ __forceinline__ uchar operator ()(int e) const
{
return (uchar)(-(e >> 1));
}
__device__ __forceinline__ GetEdges() {}
__device__ __forceinline__ GetEdges(const GetEdges&) {}
};
}
namespace cv { namespace gpu { namespace device
{
template <> struct TransformFunctorTraits<canny::GetEdges> : DefaultTransformFunctorTraits<canny::GetEdges>
{
enum { smart_shift = 4 };
};
}}}
namespace canny
{
void getEdges(PtrStepSzi map, PtrStepSzb dst)
{
transform(map, dst, GetEdges(), WithOutMask(), 0);
}
}
#endif /* CUDA_DISABLER */

3
modules/gpu/src/cuda/ccomponetns.cu
View File

@ -497,6 +497,7 @@ namespace cv { namespace gpu { namespace device
void labelComponents(const PtrStepSzb& edges, PtrStepSzi comps, int flags, cudaStream_t stream)
{
(void) flags;
dim3 block(CTA_SIZE_X, CTA_SIZE_Y);
dim3 grid(divUp(edges.cols, TILE_COLS), divUp(edges.rows, TILE_ROWS));
@ -529,4 +530,4 @@ namespace cv { namespace gpu { namespace device
}
} } }
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.0.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float, uchar>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.1.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float3, uchar3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.10.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float, unsigned short>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.11.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float3, ushort3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.12.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float4, ushort4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.13.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float3, int3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.14.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float4, int4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.2.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float4, uchar4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.3.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float3, short3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.4.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float, int>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.5.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float, float>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.6.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.7.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float4, float4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.8.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float, short>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/column_filter.9.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "column_filter.h"
namespace filter
{
template void linearColumn<float4, short4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

391
modules/gpu/src/cuda/column_filter.cu
View File

@ -1,391 +0,0 @@
/*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<float3, uchar3>(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);
template void linearColumnFilter_gpu<float3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
template void linearColumnFilter_gpu<float4, float4>(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 */

373
modules/gpu/src/cuda/column_filter.h Normal file
View File

@ -0,0 +1,373 @@
/*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*/
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
using namespace cv::gpu;
using namespace cv::gpu::device;
namespace column_filter
{
#define MAX_KERNEL_SIZE 32
__constant__ float c_kernel[MAX_KERNEL_SIZE];
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 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() );
}
}
namespace filter
{
template <typename T, typename D>
void linearColumn(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,
column_filter::caller< 1, T, D, BrdColReflect101>,
column_filter::caller< 2, T, D, BrdColReflect101>,
column_filter::caller< 3, T, D, BrdColReflect101>,
column_filter::caller< 4, T, D, BrdColReflect101>,
column_filter::caller< 5, T, D, BrdColReflect101>,
column_filter::caller< 6, T, D, BrdColReflect101>,
column_filter::caller< 7, T, D, BrdColReflect101>,
column_filter::caller< 8, T, D, BrdColReflect101>,
column_filter::caller< 9, T, D, BrdColReflect101>,
column_filter::caller<10, T, D, BrdColReflect101>,
column_filter::caller<11, T, D, BrdColReflect101>,
column_filter::caller<12, T, D, BrdColReflect101>,
column_filter::caller<13, T, D, BrdColReflect101>,
column_filter::caller<14, T, D, BrdColReflect101>,
column_filter::caller<15, T, D, BrdColReflect101>,
column_filter::caller<16, T, D, BrdColReflect101>,
column_filter::caller<17, T, D, BrdColReflect101>,
column_filter::caller<18, T, D, BrdColReflect101>,
column_filter::caller<19, T, D, BrdColReflect101>,
column_filter::caller<20, T, D, BrdColReflect101>,
column_filter::caller<21, T, D, BrdColReflect101>,
column_filter::caller<22, T, D, BrdColReflect101>,
column_filter::caller<23, T, D, BrdColReflect101>,
column_filter::caller<24, T, D, BrdColReflect101>,
column_filter::caller<25, T, D, BrdColReflect101>,
column_filter::caller<26, T, D, BrdColReflect101>,
column_filter::caller<27, T, D, BrdColReflect101>,
column_filter::caller<28, T, D, BrdColReflect101>,
column_filter::caller<29, T, D, BrdColReflect101>,
column_filter::caller<30, T, D, BrdColReflect101>,
column_filter::caller<31, T, D, BrdColReflect101>,
column_filter::caller<32, T, D, BrdColReflect101>
},
{
0,
column_filter::caller< 1, T, D, BrdColReplicate>,
column_filter::caller< 2, T, D, BrdColReplicate>,
column_filter::caller< 3, T, D, BrdColReplicate>,
column_filter::caller< 4, T, D, BrdColReplicate>,
column_filter::caller< 5, T, D, BrdColReplicate>,
column_filter::caller< 6, T, D, BrdColReplicate>,
column_filter::caller< 7, T, D, BrdColReplicate>,
column_filter::caller< 8, T, D, BrdColReplicate>,
column_filter::caller< 9, T, D, BrdColReplicate>,
column_filter::caller<10, T, D, BrdColReplicate>,
column_filter::caller<11, T, D, BrdColReplicate>,
column_filter::caller<12, T, D, BrdColReplicate>,
column_filter::caller<13, T, D, BrdColReplicate>,
column_filter::caller<14, T, D, BrdColReplicate>,
column_filter::caller<15, T, D, BrdColReplicate>,
column_filter::caller<16, T, D, BrdColReplicate>,
column_filter::caller<17, T, D, BrdColReplicate>,
column_filter::caller<18, T, D, BrdColReplicate>,
column_filter::caller<19, T, D, BrdColReplicate>,
column_filter::caller<20, T, D, BrdColReplicate>,
column_filter::caller<21, T, D, BrdColReplicate>,
column_filter::caller<22, T, D, BrdColReplicate>,
column_filter::caller<23, T, D, BrdColReplicate>,
column_filter::caller<24, T, D, BrdColReplicate>,
column_filter::caller<25, T, D, BrdColReplicate>,
column_filter::caller<26, T, D, BrdColReplicate>,
column_filter::caller<27, T, D, BrdColReplicate>,
column_filter::caller<28, T, D, BrdColReplicate>,
column_filter::caller<29, T, D, BrdColReplicate>,
column_filter::caller<30, T, D, BrdColReplicate>,
column_filter::caller<31, T, D, BrdColReplicate>,
column_filter::caller<32, T, D, BrdColReplicate>
},
{
0,
column_filter::caller< 1, T, D, BrdColConstant>,
column_filter::caller< 2, T, D, BrdColConstant>,
column_filter::caller< 3, T, D, BrdColConstant>,
column_filter::caller< 4, T, D, BrdColConstant>,
column_filter::caller< 5, T, D, BrdColConstant>,
column_filter::caller< 6, T, D, BrdColConstant>,
column_filter::caller< 7, T, D, BrdColConstant>,
column_filter::caller< 8, T, D, BrdColConstant>,
column_filter::caller< 9, T, D, BrdColConstant>,
column_filter::caller<10, T, D, BrdColConstant>,
column_filter::caller<11, T, D, BrdColConstant>,
column_filter::caller<12, T, D, BrdColConstant>,
column_filter::caller<13, T, D, BrdColConstant>,
column_filter::caller<14, T, D, BrdColConstant>,
column_filter::caller<15, T, D, BrdColConstant>,
column_filter::caller<16, T, D, BrdColConstant>,
column_filter::caller<17, T, D, BrdColConstant>,
column_filter::caller<18, T, D, BrdColConstant>,
column_filter::caller<19, T, D, BrdColConstant>,
column_filter::caller<20, T, D, BrdColConstant>,
column_filter::caller<21, T, D, BrdColConstant>,
column_filter::caller<22, T, D, BrdColConstant>,
column_filter::caller<23, T, D, BrdColConstant>,
column_filter::caller<24, T, D, BrdColConstant>,
column_filter::caller<25, T, D, BrdColConstant>,
column_filter::caller<26, T, D, BrdColConstant>,
column_filter::caller<27, T, D, BrdColConstant>,
column_filter::caller<28, T, D, BrdColConstant>,
column_filter::caller<29, T, D, BrdColConstant>,
column_filter::caller<30, T, D, BrdColConstant>,
column_filter::caller<31, T, D, BrdColConstant>,
column_filter::caller<32, T, D, BrdColConstant>
},
{
0,
column_filter::caller< 1, T, D, BrdColReflect>,
column_filter::caller< 2, T, D, BrdColReflect>,
column_filter::caller< 3, T, D, BrdColReflect>,
column_filter::caller< 4, T, D, BrdColReflect>,
column_filter::caller< 5, T, D, BrdColReflect>,
column_filter::caller< 6, T, D, BrdColReflect>,
column_filter::caller< 7, T, D, BrdColReflect>,
column_filter::caller< 8, T, D, BrdColReflect>,
column_filter::caller< 9, T, D, BrdColReflect>,
column_filter::caller<10, T, D, BrdColReflect>,
column_filter::caller<11, T, D, BrdColReflect>,
column_filter::caller<12, T, D, BrdColReflect>,
column_filter::caller<13, T, D, BrdColReflect>,
column_filter::caller<14, T, D, BrdColReflect>,
column_filter::caller<15, T, D, BrdColReflect>,
column_filter::caller<16, T, D, BrdColReflect>,
column_filter::caller<17, T, D, BrdColReflect>,
column_filter::caller<18, T, D, BrdColReflect>,
column_filter::caller<19, T, D, BrdColReflect>,
column_filter::caller<20, T, D, BrdColReflect>,
column_filter::caller<21, T, D, BrdColReflect>,
column_filter::caller<22, T, D, BrdColReflect>,
column_filter::caller<23, T, D, BrdColReflect>,
column_filter::caller<24, T, D, BrdColReflect>,
column_filter::caller<25, T, D, BrdColReflect>,
column_filter::caller<26, T, D, BrdColReflect>,
column_filter::caller<27, T, D, BrdColReflect>,
column_filter::caller<28, T, D, BrdColReflect>,
column_filter::caller<29, T, D, BrdColReflect>,
column_filter::caller<30, T, D, BrdColReflect>,
column_filter::caller<31, T, D, BrdColReflect>,
column_filter::caller<32, T, D, BrdColReflect>
},
{
0,
column_filter::caller< 1, T, D, BrdColWrap>,
column_filter::caller< 2, T, D, BrdColWrap>,
column_filter::caller< 3, T, D, BrdColWrap>,
column_filter::caller< 4, T, D, BrdColWrap>,
column_filter::caller< 5, T, D, BrdColWrap>,
column_filter::caller< 6, T, D, BrdColWrap>,
column_filter::caller< 7, T, D, BrdColWrap>,
column_filter::caller< 8, T, D, BrdColWrap>,
column_filter::caller< 9, T, D, BrdColWrap>,
column_filter::caller<10, T, D, BrdColWrap>,
column_filter::caller<11, T, D, BrdColWrap>,
column_filter::caller<12, T, D, BrdColWrap>,
column_filter::caller<13, T, D, BrdColWrap>,
column_filter::caller<14, T, D, BrdColWrap>,
column_filter::caller<15, T, D, BrdColWrap>,
column_filter::caller<16, T, D, BrdColWrap>,
column_filter::caller<17, T, D, BrdColWrap>,
column_filter::caller<18, T, D, BrdColWrap>,
column_filter::caller<19, T, D, BrdColWrap>,
column_filter::caller<20, T, D, BrdColWrap>,
column_filter::caller<21, T, D, BrdColWrap>,
column_filter::caller<22, T, D, BrdColWrap>,
column_filter::caller<23, T, D, BrdColWrap>,
column_filter::caller<24, T, D, BrdColWrap>,
column_filter::caller<25, T, D, BrdColWrap>,
column_filter::caller<26, T, D, BrdColWrap>,
column_filter::caller<27, T, D, BrdColWrap>,
column_filter::caller<28, T, D, BrdColWrap>,
column_filter::caller<29, T, D, BrdColWrap>,
column_filter::caller<30, T, D, BrdColWrap>,
column_filter::caller<31, T, D, BrdColWrap>,
column_filter::caller<32, T, D, BrdColWrap>
}
};
if (stream == 0)
cudaSafeCall( cudaMemcpyToSymbol(column_filter::c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice) );
else
cudaSafeCall( cudaMemcpyToSymbolAsync(column_filter::c_kernel, kernel, ksize * sizeof(float), 0, cudaMemcpyDeviceToDevice, stream) );
callers[brd_type][ksize]((PtrStepSz<T>)src, (PtrStepSz<D>)dst, anchor, cc, stream);
}
}

4282
modules/gpu/src/cuda/element_operations.cu
View File

File diff suppressed because it is too large Load Diff

69
modules/gpu/src/cuda/fgd_bgfg.cu
View File

@ -46,6 +46,8 @@
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/reduce.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "fgd_bgfg_common.hpp"
using namespace cv::gpu;
@ -181,57 +183,8 @@ namespace bgfg
__shared__ unsigned int data1[MERGE_THREADBLOCK_SIZE];
__shared__ unsigned int data2[MERGE_THREADBLOCK_SIZE];
data0[threadIdx.x] = sum0;
data1[threadIdx.x] = sum1;
data2[threadIdx.x] = sum2;
__syncthreads();
if (threadIdx.x < 128)
{
data0[threadIdx.x] = sum0 += data0[threadIdx.x + 128];
data1[threadIdx.x] = sum1 += data1[threadIdx.x + 128];
data2[threadIdx.x] = sum2 += data2[threadIdx.x + 128];
}
__syncthreads();
if (threadIdx.x < 64)
{
data0[threadIdx.x] = sum0 += data0[threadIdx.x + 64];
data1[threadIdx.x] = sum1 += data1[threadIdx.x + 64];
data2[threadIdx.x] = sum2 += data2[threadIdx.x + 64];
}
__syncthreads();
if (threadIdx.x < 32)
{
volatile unsigned int* vdata0 = data0;
volatile unsigned int* vdata1 = data1;
volatile unsigned int* vdata2 = data2;
vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 32];
vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 32];
vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 32];
vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 16];
vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 16];
vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 16];
vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 8];
vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 8];
vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 8];
vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 4];
vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 4];
vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 4];
vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 2];
vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 2];
vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 2];
vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 1];
vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 1];
vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 1];
}
plus<unsigned int> op;
reduce<MERGE_THREADBLOCK_SIZE>(smem_tuple(data0, data1, data2), thrust::tie(sum0, sum1, sum2), threadIdx.x, thrust::make_tuple(op, op, op));
if(threadIdx.x == 0)
{
@ -245,9 +198,9 @@ namespace bgfg
void calcDiffHistogram_gpu(PtrStepSzb prevFrame, PtrStepSzb curFrame,
unsigned int* hist0, unsigned int* hist1, unsigned int* hist2,
unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2,
int cc, cudaStream_t stream)
bool cc20, cudaStream_t stream)
{
const int HISTOGRAM_WARP_COUNT = cc < 20 ? 4 : 6;
const int HISTOGRAM_WARP_COUNT = cc20 ? 6 : 4;
const int HISTOGRAM_THREADBLOCK_SIZE = HISTOGRAM_WARP_COUNT * WARP_SIZE;
calcPartialHistogram<PT, CT><<<PARTIAL_HISTOGRAM_COUNT, HISTOGRAM_THREADBLOCK_SIZE, 0, stream>>>(
@ -261,10 +214,10 @@ namespace bgfg
cudaSafeCall( cudaDeviceSynchronize() );
}
template void calcDiffHistogram_gpu<uchar3, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
template void calcDiffHistogram_gpu<uchar3, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
template void calcDiffHistogram_gpu<uchar4, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
template void calcDiffHistogram_gpu<uchar4, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
template void calcDiffHistogram_gpu<uchar3, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
template void calcDiffHistogram_gpu<uchar3, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
template void calcDiffHistogram_gpu<uchar4, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
template void calcDiffHistogram_gpu<uchar4, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
/////////////////////////////////////////////////////////////////////////
// calcDiffThreshMask
@ -845,4 +798,4 @@ namespace bgfg
template void updateBackgroundModel_gpu<uchar4, uchar4, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, PtrStepSzb Ftd, PtrStepSzb Fbd, PtrStepSzb foreground, PtrStepSzb background, int deltaC, int deltaCC, float alpha1, float alpha2, float alpha3, int N1c, int N1cc, int N2c, int N2cc, float T, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

2
modules/gpu/src/cuda/fgd_bgfg_common.hpp
View File

@ -125,7 +125,7 @@ namespace bgfg
void calcDiffHistogram_gpu(cv::gpu::PtrStepSzb prevFrame, cv::gpu::PtrStepSzb curFrame,
unsigned int* hist0, unsigned int* hist1, unsigned int* hist2,
unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2,
int cc, cudaStream_t stream);
bool cc20, cudaStream_t stream);
template <typename PT, typename CT>
void calcDiffThreshMask_gpu(cv::gpu::PtrStepSzb prevFrame, cv::gpu::PtrStepSzb curFrame, uchar3 bestThres, cv::gpu::PtrStepSzb changeMask, cudaStream_t stream);

3
modules/gpu/src/cuda/gftt.cu
View File

@ -47,6 +47,7 @@
#if !defined CUDA_DISABLER
#include <thrust/device_ptr.h>
#include <thrust/sort.h>
#include "opencv2/gpu/device/common.hpp"
@ -148,4 +149,4 @@ namespace cv { namespace gpu { namespace device
}}}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

13
modules/gpu/src/cuda/global_motion.cu
View File

@ -43,12 +43,11 @@
#if !defined CUDA_DISABLER
#include "thrust/device_ptr.h"
#include "thrust/remove.h"
#include "thrust/functional.h"
#include "internal_shared.hpp"
#include <thrust/device_ptr.h>
#include <thrust/remove.h>
#include <thrust/functional.h>
using namespace thrust;
#include "internal_shared.hpp"
namespace cv { namespace gpu { namespace device { namespace globmotion {
@ -64,7 +63,7 @@ int compactPoints(int N, float *points0, float *points1, const uchar *mask)
return thrust::remove_if(thrust::make_zip_iterator(thrust::make_tuple(dpoints0, dpoints1)),
thrust::make_zip_iterator(thrust::make_tuple(dpoints0 + N, dpoints1 + N)),
dmask, thrust::not1(thrust::identity<uchar>()))
- make_zip_iterator(make_tuple(dpoints0, dpoints1));
- thrust::make_zip_iterator(make_tuple(dpoints0, dpoints1));
}
@ -117,4 +116,4 @@ void calcWobbleSuppressionMaps(
}}}}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

270
modules/gpu/src/cuda/hist.cu
View File

@ -43,182 +43,112 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/emulation.hpp"
#include "opencv2/gpu/device/transform.hpp"
using namespace cv::gpu;
using namespace cv::gpu::device;
namespace hist
{
__global__ void histogram256Kernel(const uchar* src, int cols, int rows, size_t step, int* hist)
{
__shared__ int shist[256];
const int y = blockIdx.x * blockDim.y + threadIdx.y;
const int tid = threadIdx.y * blockDim.x + threadIdx.x;
shist[tid] = 0;
__syncthreads();
if (y < rows)
{
const unsigned int* rowPtr = (const unsigned int*) (src + y * step);
const int cols_4 = cols / 4;
for (int x = threadIdx.x; x < cols_4; x += blockDim.x)
{
unsigned int data = rowPtr[x];
Emulation::smem::atomicAdd(&shist[(data >> 0) & 0xFFU], 1);
Emulation::smem::atomicAdd(&shist[(data >> 8) & 0xFFU], 1);
Emulation::smem::atomicAdd(&shist[(data >> 16) & 0xFFU], 1);
Emulation::smem::atomicAdd(&shist[(data >> 24) & 0xFFU], 1);
}
if (cols % 4 != 0 && threadIdx.x == 0)
{
for (int x = cols_4 * 4; x < cols; ++x)
{
unsigned int data = ((const uchar*)rowPtr)[x];
Emulation::smem::atomicAdd(&shist[data], 1);
}
}
}
__syncthreads();
const int histVal = shist[tid];
if (histVal > 0)
::atomicAdd(hist + tid, histVal);
}
void histogram256(PtrStepSzb src, int* hist, cudaStream_t stream)
{
const dim3 block(32, 8);
const dim3 grid(divUp(src.rows, block.y));
histogram256Kernel<<<grid, block, 0, stream>>>(src.data, src.cols, src.rows, src.step, hist);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
/////////////////////////////////////////////////////////////////////////
namespace hist
{
__constant__ int c_lut[256];
struct EqualizeHist : unary_function<uchar, uchar>
{
float scale;
__host__ EqualizeHist(float _scale) : scale(_scale) {}
__device__ __forceinline__ uchar operator ()(uchar val) const
{
const int lut = c_lut[val];
return __float2int_rn(scale * lut);
}
};
}
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
template <> struct TransformFunctorTraits<hist::EqualizeHist> : DefaultTransformFunctorTraits<hist::EqualizeHist>
{
#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));
enum { smart_shift = 4 };
};
}}}
namespace hist
{
void equalizeHist(PtrStepSzb src, PtrStepSzb dst, const int* lut, cudaStream_t stream)
{
if (stream == 0)
cudaSafeCall( cudaMemcpyToSymbol(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice) );
else
cudaSafeCall( cudaMemcpyToSymbolAsync(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice, stream) );
equalizeHist<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
const float scale = 255.0f / (src.cols * src.rows);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace hist
}}} // namespace cv { namespace gpu { namespace device
transform(src, dst, EqualizeHist(scale), WithOutMask(), stream);
}
}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

181
modules/gpu/src/cuda/hog.cu
View File

@ -42,7 +42,10 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/reduce.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/warp_shuffle.hpp"
namespace cv { namespace gpu { namespace device
{
@ -226,29 +229,32 @@ namespace cv { namespace gpu { namespace device
template<int size>
__device__ float reduce_smem(volatile float* smem)
__device__ float reduce_smem(float* smem, float val)
{
unsigned int tid = threadIdx.x;
float sum = smem[tid];
float sum = val;
if (size >= 512) { if (tid < 256) smem[tid] = sum = sum + smem[tid + 256]; __syncthreads(); }
if (size >= 256) { if (tid < 128) smem[tid] = sum = sum + smem[tid + 128]; __syncthreads(); }
if (size >= 128) { if (tid < 64) smem[tid] = sum = sum + smem[tid + 64]; __syncthreads(); }
reduce<size>(smem, sum, tid, plus<float>());
if (tid < 32)
if (size == 32)
{
if (size >= 64) smem[tid] = sum = sum + smem[tid + 32];
if (size >= 32) smem[tid] = sum = sum + smem[tid + 16];
if (size >= 16) smem[tid] = sum = sum + smem[tid + 8];
if (size >= 8) smem[tid] = sum = sum + smem[tid + 4];
if (size >= 4) smem[tid] = sum = sum + smem[tid + 2];
if (size >= 2) smem[tid] = sum = sum + smem[tid + 1];
#if __CUDA_ARCH__ >= 300
return shfl(sum, 0);
#else
return smem[0];
#endif
}
else
{
#if __CUDA_ARCH__ >= 300
if (threadIdx.x == 0)
smem[0] = sum;
#endif
__syncthreads();
sum = smem[0];
__syncthreads();
return sum;
return smem[0];
}
}
@ -272,19 +278,13 @@ namespace cv { namespace gpu { namespace device
if (threadIdx.x < block_hist_size)
elem = hist[0];
squares[threadIdx.x] = elem * elem;
__syncthreads();
float sum = reduce_smem<nthreads>(squares);
float sum = reduce_smem<nthreads>(squares, elem * elem);
float scale = 1.0f / (::sqrtf(sum) + 0.1f * block_hist_size);
elem = ::min(elem * scale, threshold);
__syncthreads();
squares[threadIdx.x] = elem * elem;
sum = reduce_smem<nthreads>(squares, elem * elem);
__syncthreads();
sum = reduce_smem<nthreads>(squares);
scale = 1.0f / (::sqrtf(sum) + 1e-3f);
if (threadIdx.x < block_hist_size)
@ -330,65 +330,36 @@ namespace cv { namespace gpu { namespace device
// return confidence values not just positive location
template <int nthreads, // Number of threads per one histogram block
int nblocks> // Number of histogram block processed by single GPU thread block
int nblocks> // Number of histogram block processed by single GPU thread block
__global__ void compute_confidence_hists_kernel_many_blocks(const int img_win_width, const int img_block_width,
const int win_block_stride_x, const int win_block_stride_y,
const float* block_hists, const float* coefs,
float free_coef, float threshold, float* confidences)
{
const int win_x = threadIdx.z;
if (blockIdx.x * blockDim.z + win_x >= img_win_width)
return;
const int win_x = threadIdx.z;
if (blockIdx.x * blockDim.z + win_x >= img_win_width)
return;
const float* hist = block_hists + (blockIdx.y * win_block_stride_y * img_block_width +
blockIdx.x * win_block_stride_x * blockDim.z + win_x) *
cblock_hist_size;
const float* hist = block_hists + (blockIdx.y * win_block_stride_y * img_block_width +
blockIdx.x * win_block_stride_x * blockDim.z + win_x) *
cblock_hist_size;
float product = 0.f;
for (int i = threadIdx.x; i < cdescr_size; i += nthreads)
{
int offset_y = i / cdescr_width;
int offset_x = i - offset_y * cdescr_width;
product += coefs[i] * hist[offset_y * img_block_width * cblock_hist_size + offset_x];
}
float product = 0.f;
for (int i = threadIdx.x; i < cdescr_size; i += nthreads)
{
int offset_y = i / cdescr_width;
int offset_x = i - offset_y * cdescr_width;
product += coefs[i] * hist[offset_y * img_block_width * cblock_hist_size + offset_x];
}
__shared__ float products[nthreads * nblocks];
__shared__ float products[nthreads * nblocks];
const int tid = threadIdx.z * nthreads + threadIdx.x;
products[tid] = product;
const int tid = threadIdx.z * nthreads + threadIdx.x;
__syncthreads();
reduce<nthreads>(products, product, tid, plus<float>());
if (nthreads >= 512)
{
if (threadIdx.x < 256) products[tid] = product = product + products[tid + 256];
__syncthreads();
}
if (nthreads >= 256)
{
if (threadIdx.x < 128) products[tid] = product = product + products[tid + 128];
__syncthreads();
}
if (nthreads >= 128)
{
if (threadIdx.x < 64) products[tid] = product = product + products[tid + 64];
__syncthreads();
}
if (threadIdx.x < 32)
{
volatile float* smem = products;
if (nthreads >= 64) smem[tid] = product = product + smem[tid + 32];
if (nthreads >= 32) smem[tid] = product = product + smem[tid + 16];
if (nthreads >= 16) smem[tid] = product = product + smem[tid + 8];
if (nthreads >= 8) smem[tid] = product = product + smem[tid + 4];
if (nthreads >= 4) smem[tid] = product = product + smem[tid + 2];
if (nthreads >= 2) smem[tid] = product = product + smem[tid + 1];
}
if (threadIdx.x == 0)
confidences[blockIdx.y * img_win_width + blockIdx.x * blockDim.z + win_x]
= (float)(product + free_coef);
if (threadIdx.x == 0)
confidences[blockIdx.y * img_win_width + blockIdx.x * blockDim.z + win_x] = product + free_coef;
}
@ -396,32 +367,32 @@ namespace cv { namespace gpu { namespace device
int win_stride_y, int win_stride_x, int height, int width, float* block_hists,
float* coefs, float free_coef, float threshold, float *confidences)
{
const int nthreads = 256;
const int nblocks = 1;
const int nthreads = 256;
const int nblocks = 1;
int win_block_stride_x = win_stride_x / block_stride_x;
int win_block_stride_y = win_stride_y / block_stride_y;
int img_win_width = (width - win_width + win_stride_x) / win_stride_x;
int img_win_height = (height - win_height + win_stride_y) / win_stride_y;
int win_block_stride_x = win_stride_x / block_stride_x;
int win_block_stride_y = win_stride_y / block_stride_y;
int img_win_width = (width - win_width + win_stride_x) / win_stride_x;
int img_win_height = (height - win_height + win_stride_y) / win_stride_y;
dim3 threads(nthreads, 1, nblocks);
dim3 grid(divUp(img_win_width, nblocks), img_win_height);
dim3 threads(nthreads, 1, nblocks);
dim3 grid(divUp(img_win_width, nblocks), img_win_height);
cudaSafeCall(cudaFuncSetCacheConfig(compute_confidence_hists_kernel_many_blocks<nthreads, nblocks>,
cudaFuncCachePreferL1));
cudaSafeCall(cudaFuncSetCacheConfig(compute_confidence_hists_kernel_many_blocks<nthreads, nblocks>,
cudaFuncCachePreferL1));
int img_block_width = (width - CELLS_PER_BLOCK_X * CELL_WIDTH + block_stride_x) /
block_stride_x;
compute_confidence_hists_kernel_many_blocks<nthreads, nblocks><<<grid, threads>>>(
img_win_width, img_block_width, win_block_stride_x, win_block_stride_y,
block_hists, coefs, free_coef, threshold, confidences);
cudaSafeCall(cudaThreadSynchronize());
int img_block_width = (width - CELLS_PER_BLOCK_X * CELL_WIDTH + block_stride_x) /
block_stride_x;
compute_confidence_hists_kernel_many_blocks<nthreads, nblocks><<<grid, threads>>>(
img_win_width, img_block_width, win_block_stride_x, win_block_stride_y,
block_hists, coefs, free_coef, threshold, confidences);
cudaSafeCall(cudaThreadSynchronize());
}
template <int nthreads, // Number of threads per one histogram block
int nblocks> // Number of histogram block processed by single GPU thread block
int nblocks> // Number of histogram block processed by single GPU thread block
__global__ void classify_hists_kernel_many_blocks(const int img_win_width, const int img_block_width,
const int win_block_stride_x, const int win_block_stride_y,
const float* block_hists, const float* coefs,
@ -446,36 +417,8 @@ namespace cv { namespace gpu { namespace device
__shared__ float products[nthreads * nblocks];
const int tid = threadIdx.z * nthreads + threadIdx.x;
products[tid] = product;
__syncthreads();
if (nthreads >= 512)
{
if (threadIdx.x < 256) products[tid] = product = product + products[tid + 256];
__syncthreads();
}
if (nthreads >= 256)
{
if (threadIdx.x < 128) products[tid] = product = product + products[tid + 128];
__syncthreads();
}
if (nthreads >= 128)
{
if (threadIdx.x < 64) products[tid] = product = product + products[tid + 64];
__syncthreads();
}
if (threadIdx.x < 32)
{
volatile float* smem = products;
if (nthreads >= 64) smem[tid] = product = product + smem[tid + 32];
if (nthreads >= 32) smem[tid] = product = product + smem[tid + 16];
if (nthreads >= 16) smem[tid] = product = product + smem[tid + 8];
if (nthreads >= 8) smem[tid] = product = product + smem[tid + 4];
if (nthreads >= 4) smem[tid] = product = product + smem[tid + 2];
if (nthreads >= 2) smem[tid] = product = product + smem[tid + 1];
}
reduce<nthreads>(products, product, tid, plus<float>());
if (threadIdx.x == 0)
labels[blockIdx.y * img_win_width + blockIdx.x * blockDim.z + win_x] = (product + free_coef >= threshold);
@ -868,4 +811,4 @@ namespace cv { namespace gpu { namespace device
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

199
modules/gpu/src/cuda/hough.cu
View File

@ -42,7 +42,9 @@
#if !defined CUDA_DISABLER
#include <thrust/device_ptr.h>
#include <thrust/sort.h>
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/emulation.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
@ -291,6 +293,201 @@ namespace cv { namespace gpu { namespace device
return totalCount;
}
////////////////////////////////////////////////////////////////////////
// houghLinesProbabilistic
texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_mask(false, cudaFilterModePoint, cudaAddressModeClamp);
__global__ void houghLinesProbabilistic(const PtrStepSzi accum,
int4* out, const int maxSize,
const float rho, const float theta,
const int lineGap, const int lineLength,
const int rows, const int cols)
{
const int r = blockIdx.x * blockDim.x + threadIdx.x;
const int n = blockIdx.y * blockDim.y + threadIdx.y;
if (r >= accum.cols - 2 || n >= accum.rows - 2)
return;
const int curVotes = accum(n + 1, r + 1);
if (curVotes >= lineLength &&
curVotes > accum(n, r) &&
curVotes > accum(n, r + 1) &&
curVotes > accum(n, r + 2) &&
curVotes > accum(n + 1, r) &&
curVotes > accum(n + 1, r + 2) &&
curVotes > accum(n + 2, r) &&
curVotes > accum(n + 2, r + 1) &&
curVotes > accum(n + 2, r + 2))
{
const float radius = (r - (accum.cols - 2 - 1) * 0.5f) * rho;
const float angle = n * theta;
float cosa;
float sina;
sincosf(angle, &sina, &cosa);
float2 p0 = make_float2(cosa * radius, sina * radius);
float2 dir = make_float2(-sina, cosa);
float2 pb[4] = {make_float2(-1, -1), make_float2(-1, -1), make_float2(-1, -1), make_float2(-1, -1)};
float a;
if (dir.x != 0)
{
a = -p0.x / dir.x;
pb[0].x = 0;
pb[0].y = p0.y + a * dir.y;
a = (cols - 1 - p0.x) / dir.x;
pb[1].x = cols - 1;
pb[1].y = p0.y + a * dir.y;
}
if (dir.y != 0)
{
a = -p0.y / dir.y;
pb[2].x = p0.x + a * dir.x;
pb[2].y = 0;
a = (rows - 1 - p0.y) / dir.y;
pb[3].x = p0.x + a * dir.x;
pb[3].y = rows - 1;
}
if (pb[0].x == 0 && (pb[0].y >= 0 && pb[0].y < rows))
{
p0 = pb[0];
if (dir.x < 0)
dir = -dir;
}
else if (pb[1].x == cols - 1 && (pb[0].y >= 0 && pb[0].y < rows))
{
p0 = pb[1];
if (dir.x > 0)
dir = -dir;
}
else if (pb[2].y == 0 && (pb[2].x >= 0 && pb[2].x < cols))
{
p0 = pb[2];
if (dir.y < 0)
dir = -dir;
}
else if (pb[3].y == rows - 1 && (pb[3].x >= 0 && pb[3].x < cols))
{
p0 = pb[3];
if (dir.y > 0)
dir = -dir;
}
float2 d;
if (::fabsf(dir.x) > ::fabsf(dir.y))
{
d.x = dir.x > 0 ? 1 : -1;
d.y = dir.y / ::fabsf(dir.x);
}
else
{
d.x = dir.x / ::fabsf(dir.y);
d.y = dir.y > 0 ? 1 : -1;
}
float2 line_end[2];
int gap;
bool inLine = false;
float2 p1 = p0;
if (p1.x < 0 || p1.x >= cols || p1.y < 0 || p1.y >= rows)
return;
for (;;)
{
if (tex2D(tex_mask, p1.x, p1.y))
{
gap = 0;
if (!inLine)
{
line_end[0] = p1;
line_end[1] = p1;
inLine = true;
}
else
{
line_end[1] = p1;
}
}
else if (inLine)
{
if (++gap > lineGap)
{
bool good_line = ::abs(line_end[1].x - line_end[0].x) >= lineLength ||
::abs(line_end[1].y - line_end[0].y) >= lineLength;
if (good_line)
{
const int ind = ::atomicAdd(&g_counter, 1);
if (ind < maxSize)
out[ind] = make_int4(line_end[0].x, line_end[0].y, line_end[1].x, line_end[1].y);
}
gap = 0;
inLine = false;
}
}
p1 = p1 + d;
if (p1.x < 0 || p1.x >= cols || p1.y < 0 || p1.y >= rows)
{
if (inLine)
{
bool good_line = ::abs(line_end[1].x - line_end[0].x) >= lineLength ||
::abs(line_end[1].y - line_end[0].y) >= lineLength;
if (good_line)
{
const int ind = ::atomicAdd(&g_counter, 1);
if (ind < maxSize)
out[ind] = make_int4(line_end[0].x, line_end[0].y, line_end[1].x, line_end[1].y);
}
}
break;
}
}
}
}
int houghLinesProbabilistic_gpu(PtrStepSzb mask, PtrStepSzi accum, int4* out, int maxSize, float rho, float theta, int lineGap, int lineLength)
{
void* counterPtr;
cudaSafeCall( cudaGetSymbolAddress(&counterPtr, g_counter) );
cudaSafeCall( cudaMemset(counterPtr, 0, sizeof(int)) );
const dim3 block(32, 8);
const dim3 grid(divUp(accum.cols - 2, block.x), divUp(accum.rows - 2, block.y));
bindTexture(&tex_mask, mask);
houghLinesProbabilistic<<<grid, block>>>(accum,
out, maxSize,
rho, theta,
lineGap, lineLength,
mask.rows, mask.cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
int totalCount;
cudaSafeCall( cudaMemcpy(&totalCount, counterPtr, sizeof(int), cudaMemcpyDeviceToHost) );
totalCount = ::min(totalCount, maxSize);
return totalCount;
}
////////////////////////////////////////////////////////////////////////
// circlesAccumCenters
@ -1509,4 +1706,4 @@ namespace cv { namespace gpu { namespace device
}}}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

563
modules/gpu/src/cuda/icf-sc.cu Normal file
View File

@ -0,0 +1,563 @@
/*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) 2008-2012, 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 <opencv2/gpu/device/common.hpp>
#include <opencv2/gpu/device/saturate_cast.hpp>
#include <icf.hpp>
#include <float.h>
#include <stdio.h>
namespace cv { namespace gpu { namespace device {
namespace icf {
template <int FACTOR>
__device__ __forceinline__ uchar shrink(const uchar* ptr, const int pitch, const int y, const int x)
{
int out = 0;
#pragma unroll
for(int dy = 0; dy < FACTOR; ++dy)
#pragma unroll
for(int dx = 0; dx < FACTOR; ++dx)
{
out += ptr[dy * pitch + dx];
}
return static_cast<uchar>(out / (FACTOR * FACTOR));
}
template<int FACTOR>
__global__ void shrink(const uchar* __restrict__ hogluv, const int inPitch,
uchar* __restrict__ shrank, const int outPitch )
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const uchar* ptr = hogluv + (FACTOR * y) * inPitch + (FACTOR * x);
shrank[ y * outPitch + x] = shrink<FACTOR>(ptr, inPitch, y, x);
}
void shrink(const cv::gpu::PtrStepSzb& channels, cv::gpu::PtrStepSzb shrunk)
{
dim3 block(32, 8);
dim3 grid(shrunk.cols / 32, shrunk.rows / 8);
shrink<4><<<grid, block>>>((uchar*)channels.ptr(), channels.step, (uchar*)shrunk.ptr(), shrunk.step);
cudaSafeCall(cudaDeviceSynchronize());
}
__device__ __forceinline__ void luv(const float& b, const float& g, const float& r, uchar& __l, uchar& __u, uchar& __v)
{
// rgb -> XYZ
float x = 0.412453f * r + 0.357580f * g + 0.180423f * b;
float y = 0.212671f * r + 0.715160f * g + 0.072169f * b;
float z = 0.019334f * r + 0.119193f * g + 0.950227f * b;
// computed for D65
const float _ur = 0.19783303699678276f;
const float _vr = 0.46833047435252234f;
const float divisor = fmax((x + 15.f * y + 3.f * z), FLT_EPSILON);
const float _u = __fdividef(4.f * x, divisor);
const float _v = __fdividef(9.f * y, divisor);
float hack = static_cast<float>(__float2int_rn(y * 2047)) / 2047;
const float L = fmax(0.f, ((116.f * cbrtf(hack)) - 16.f));
const float U = 13.f * L * (_u - _ur);
const float V = 13.f * L * (_v - _vr);
// L in [0, 100], u in [-134, 220], v in [-140, 122]
__l = static_cast<uchar>( L * (255.f / 100.f));
__u = static_cast<uchar>((U + 134.f) * (255.f / (220.f + 134.f )));
__v = static_cast<uchar>((V + 140.f) * (255.f / (122.f + 140.f )));
}
__global__ void bgr2Luv_d(const uchar* rgb, const int rgbPitch, uchar* luvg, const int luvgPitch)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
uchar3 color = ((uchar3*)(rgb + rgbPitch * y))[x];
uchar l, u, v;
luv(color.x / 255.f, color.y / 255.f, color.z / 255.f, l, u, v);
luvg[luvgPitch * y + x] = l;
luvg[luvgPitch * (y + 480) + x] = u;
luvg[luvgPitch * (y + 2 * 480) + x] = v;
}
void bgr2Luv(const PtrStepSzb& bgr, PtrStepSzb luv)
{
dim3 block(32, 8);
dim3 grid(bgr.cols / 32, bgr.rows / 8);
bgr2Luv_d<<<grid, block>>>((const uchar*)bgr.ptr(0), bgr.step, (uchar*)luv.ptr(0), luv.step);
cudaSafeCall(cudaDeviceSynchronize());
}
template<bool isDefaultNum>
__device__ __forceinline__ int fast_angle_bin(const float& dx, const float& dy)
{
const float angle_quantum = CV_PI / 6.f;
float angle = atan2(dx, dy) + (angle_quantum / 2.f);
if (angle < 0) angle += CV_PI;
const float angle_scaling = 1.f / angle_quantum;
return static_cast<int>(angle * angle_scaling) % 6;
}
template<>
__device__ __forceinline__ int fast_angle_bin<true>(const float& dy, const float& dx)
{
int index = 0;
float max_dot = fabs(dx);
{
const float dot_product = fabs(dx * 0.8660254037844386f + dy * 0.5f);
if(dot_product > max_dot)
{
max_dot = dot_product;
index = 1;
}
}
{
const float dot_product = fabs(dy * 0.8660254037844386f + dx * 0.5f);
if(dot_product > max_dot)
{
max_dot = dot_product;
index = 2;
}
}
{
int i = 3;
float2 bin_vector_i;
bin_vector_i.x = ::cos(i * (CV_PI / 6.f));
bin_vector_i.y = ::sin(i * (CV_PI / 6.f));
const float dot_product = fabs(dx * bin_vector_i.x + dy * bin_vector_i.y);
if(dot_product > max_dot)
{
max_dot = dot_product;
index = i;
}
}
{
const float dot_product = fabs(dx * (-0.4999999999999998f) + dy * 0.8660254037844387f);
if(dot_product > max_dot)
{
max_dot = dot_product;
index = 4;
}
}
{
const float dot_product = fabs(dx * (-0.8660254037844387f) + dy * 0.49999999999999994f);
if(dot_product > max_dot)
{
max_dot = dot_product;
index = 5;
}
}
return index;
}
texture<uchar, cudaTextureType2D, cudaReadModeElementType> tgray;
template<bool isDefaultNum>
__global__ void gray2hog(PtrStepSzb mag)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const float dx = tex2D(tgray, x + 1, y + 0) - tex2D(tgray, x - 1, y - 0);
const float dy = tex2D(tgray, x + 0, y + 1) - tex2D(tgray, x - 0, y - 1);
const float magnitude = sqrtf((dx * dx) + (dy * dy)) * (1.0f / sqrtf(2));
const uchar cmag = static_cast<uchar>(magnitude);
mag( 480 * 6 + y, x) = cmag;
mag( 480 * fast_angle_bin<isDefaultNum>(dy, dx) + y, x) = cmag;
}
void gray2hog(const PtrStepSzb& gray, PtrStepSzb mag, const int bins)
{
dim3 block(32, 8);
dim3 grid(gray.cols / 32, gray.rows / 8);
cudaChannelFormatDesc desc = cudaCreateChannelDesc<uchar>();
cudaSafeCall( cudaBindTexture2D(0, tgray, gray.data, desc, gray.cols, gray.rows, gray.step) );
if (bins == 6)
gray2hog<true><<<grid, block>>>(mag);
else
gray2hog<false><<<grid, block>>>(mag);
cudaSafeCall(cudaDeviceSynchronize());
}
// ToDo: use textures or uncached load instruction.
__global__ void magToHist(const uchar* __restrict__ mag,
const float* __restrict__ angle, const int angPitch,
uchar* __restrict__ hog, const int hogPitch, const int fh)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int bin = (int)(angle[y * angPitch + x]);
const uchar val = mag[y * hogPitch + x];
hog[((fh * bin) + y) * hogPitch + x] = val;
}
void fillBins(cv::gpu::PtrStepSzb hogluv, const cv::gpu::PtrStepSzf& nangle,
const int fw, const int fh, const int bins, cudaStream_t stream )
{
const uchar* mag = (const uchar*)hogluv.ptr(fh * bins);
uchar* hog = (uchar*)hogluv.ptr();
const float* angle = (const float*)nangle.ptr();
dim3 block(32, 8);
dim3 grid(fw / 32, fh / 8);
magToHist<<<grid, block, 0, stream>>>(mag, angle, nangle.step / sizeof(float), hog, hogluv.step, fh);
if (!stream)
{
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
}
__device__ __forceinline__ float overlapArea(const Detection &a, const Detection &b)
{
int w = ::min(a.x + a.w, b.x + b.w) - ::max(a.x, b.x);
int h = ::min(a.y + a.h, b.y + b.h) - ::max(a.y, b.y);
return (w < 0 || h < 0)? 0.f : (float)(w * h);
}
texture<uint4, cudaTextureType2D, cudaReadModeElementType> tdetections;
__global__ void overlap(const uint* n, uchar* overlaps)
{
const int idx = threadIdx.x;
const int total = *n;
for (int i = idx + 1; i < total; i += 192)
{
const uint4 _a = tex2D(tdetections, i, 0);
const Detection& a = *((Detection*)(&_a));
bool excluded = false;
for (int j = i + 1; j < total; ++j)
{
const uint4 _b = tex2D(tdetections, j, 0);
const Detection& b = *((Detection*)(&_b));
float ovl = overlapArea(a, b) / ::min(a.w * a.h, b.w * b.h);
if (ovl > 0.65f)
{
int suppessed = (a.confidence > b.confidence)? j : i;
overlaps[suppessed] = 1;
excluded = excluded || (suppessed == i);
}
#if __CUDA_ARCH__ >= 120
if (__all(excluded)) break;
#endif
}
}
}
__global__ void collect(const uint* n, uchar* overlaps, uint* ctr, uint4* suppressed)
{
const int idx = threadIdx.x;
const int total = *n;
for (int i = idx; i < total; i += 192)
{
if (!overlaps[i])
{
int oidx = atomicInc(ctr, 50);
suppressed[oidx] = tex2D(tdetections, i + 1, 0);
}
}
}
void suppress(const PtrStepSzb& objects, PtrStepSzb overlaps, PtrStepSzi ndetections,
PtrStepSzb suppressed, cudaStream_t stream)
{
int block = 192;
int grid = 1;
cudaChannelFormatDesc desc = cudaCreateChannelDesc<uint4>();
size_t offset;
cudaSafeCall( cudaBindTexture2D(&offset, tdetections, objects.data, desc, objects.cols / sizeof(uint4), objects.rows, objects.step));
overlap<<<grid, block>>>((uint*)ndetections.ptr(0), (uchar*)overlaps.ptr(0));
collect<<<grid, block>>>((uint*)ndetections.ptr(0), (uchar*)overlaps.ptr(0), (uint*)suppressed.ptr(0), ((uint4*)suppressed.ptr(0)) + 1);
if (!stream)
{
cudaSafeCall( cudaGetLastError());
cudaSafeCall( cudaDeviceSynchronize());
}
}
template<typename Policy>
struct PrefixSum
{
__device static void apply(float& impact)
{
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
#pragma unroll
// scan on shuffl functions
for (int i = 1; i < Policy::WARP; i *= 2)
{
const float n = __shfl_up(impact, i, Policy::WARP);
if (threadIdx.x >= i)
impact += n;
}
#else
__shared__ volatile float ptr[Policy::STA_X * Policy::STA_Y];
const int idx = threadIdx.y * Policy::STA_X + threadIdx.x;
ptr[idx] = impact;
if ( threadIdx.x >= 1) ptr [idx ] = (ptr [idx - 1] + ptr [idx]);
if ( threadIdx.x >= 2) ptr [idx ] = (ptr [idx - 2] + ptr [idx]);
if ( threadIdx.x >= 4) ptr [idx ] = (ptr [idx - 4] + ptr [idx]);
if ( threadIdx.x >= 8) ptr [idx ] = (ptr [idx - 8] + ptr [idx]);
if ( threadIdx.x >= 16) ptr [idx ] = (ptr [idx - 16] + ptr [idx]);
impact = ptr[idx];
#endif
}
};
texture<int, cudaTextureType2D, cudaReadModeElementType> thogluv;
template<bool isUp>
__device__ __forceinline__ float rescale(const Level& level, Node& node)
{
uchar4& scaledRect = node.rect;
float relScale = level.relScale;
float farea = (scaledRect.z - scaledRect.x) * (scaledRect.w - scaledRect.y);
// rescale
scaledRect.x = __float2int_rn(relScale * scaledRect.x);
scaledRect.y = __float2int_rn(relScale * scaledRect.y);
scaledRect.z = __float2int_rn(relScale * scaledRect.z);
scaledRect.w = __float2int_rn(relScale * scaledRect.w);
float sarea = (scaledRect.z - scaledRect.x) * (scaledRect.w - scaledRect.y);
const float expected_new_area = farea * relScale * relScale;
float approx = (sarea == 0)? 1: __fdividef(sarea, expected_new_area);
float rootThreshold = (node.threshold & 0x0FFFFFFFU) * approx * level.scaling[(node.threshold >> 28) > 6];
return rootThreshold;
}
template<>
__device__ __forceinline__ float rescale<true>(const Level& level, Node& node)
{
uchar4& scaledRect = node.rect;
float relScale = level.relScale;
float farea = scaledRect.z * scaledRect.w;
// rescale
scaledRect.x = __float2int_rn(relScale * scaledRect.x);
scaledRect.y = __float2int_rn(relScale * scaledRect.y);
scaledRect.z = __float2int_rn(relScale * scaledRect.z);
scaledRect.w = __float2int_rn(relScale * scaledRect.w);
float sarea = scaledRect.z * scaledRect.w;
const float expected_new_area = farea * relScale * relScale;
float approx = __fdividef(sarea, expected_new_area);
float rootThreshold = (node.threshold & 0x0FFFFFFFU) * approx * level.scaling[(node.threshold >> 28) > 6];
return rootThreshold;
}
template<bool isUp>
__device__ __forceinline__ int get(int x, int y, uchar4 area)
{
int a = tex2D(thogluv, x + area.x, y + area.y);
int b = tex2D(thogluv, x + area.z, y + area.y);
int c = tex2D(thogluv, x + area.z, y + area.w);
int d = tex2D(thogluv, x + area.x, y + area.w);
return (a - b + c - d);
}
template<>
__device__ __forceinline__ int get<true>(int x, int y, uchar4 area)
{
x += area.x;
y += area.y;
int a = tex2D(thogluv, x, y);
int b = tex2D(thogluv, x + area.z, y);
int c = tex2D(thogluv, x + area.z, y + area.w);
int d = tex2D(thogluv, x, y + area.w);
return (a - b + c - d);
}
texture<float2, cudaTextureType2D, cudaReadModeElementType> troi;
template<typename Policy>
template<bool isUp>
__device void CascadeInvoker<Policy>::detect(Detection* objects, const uint ndetections, uint* ctr, const int downscales) const
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x;
// load Lavel
__shared__ Level level;
// check POI
__shared__ volatile char roiCache[Policy::STA_Y];
if (!threadIdx.y && !threadIdx.x)
((float2*)roiCache)[threadIdx.x] = tex2D(troi, blockIdx.y, x);
__syncthreads();
if (!roiCache[threadIdx.y]) return;
if (!threadIdx.x)
level = levels[downscales + blockIdx.z];
if(x >= level.workRect.x || y >= level.workRect.y) return;
int st = level.octave * level.step;
const int stEnd = st + level.step;
const int hogluvStep = gridDim.y * Policy::STA_Y;
float confidence = 0.f;
for(; st < stEnd; st += Policy::WARP)
{
const int nId = (st + threadIdx.x) * 3;
Node node = nodes[nId];
float threshold = rescale<isUp>(level, node);
int sum = get<isUp>(x, y + (node.threshold >> 28) * hogluvStep, node.rect);
int next = 1 + (int)(sum >= threshold);
node = nodes[nId + next];
threshold = rescale<isUp>(level, node);
sum = get<isUp>(x, y + (node.threshold >> 28) * hogluvStep, node.rect);
const int lShift = (next - 1) * 2 + (int)(sum >= threshold);
float impact = leaves[(st + threadIdx.x) * 4 + lShift];
PrefixSum<Policy>::apply(impact);
confidence += impact;
#if __CUDA_ARCH__ >= 120
if(__any((confidence <= stages[(st + threadIdx.x)]))) st += 2048;
#endif
}
if(!threadIdx.x && st == stEnd && ((confidence - FLT_EPSILON) >= 0))
{
int idx = atomicInc(ctr, ndetections);
objects[idx] = Detection(__float2int_rn(x * Policy::SHRINKAGE),
__float2int_rn(y * Policy::SHRINKAGE), level.objSize.x, level.objSize.y, confidence);
}
}
template<typename Policy, bool isUp>
__global__ void soft_cascade(const CascadeInvoker<Policy> invoker, Detection* objects, const uint n, uint* ctr, const int downs)
{
invoker.template detect<isUp>(objects, n, ctr, downs);
}
template<typename Policy>
void CascadeInvoker<Policy>::operator()(const PtrStepSzb& roi, const PtrStepSzi& hogluv,
PtrStepSz<uchar4> objects, const int downscales, const cudaStream_t& stream) const
{
int fw = roi.rows;
int fh = roi.cols;
dim3 grid(fw, fh / Policy::STA_Y, downscales);
uint* ctr = (uint*)(objects.ptr(0));
Detection* det = ((Detection*)objects.ptr(0)) + 1;
uint max_det = objects.cols / sizeof(Detection);
cudaChannelFormatDesc desc = cudaCreateChannelDesc<int>();
cudaSafeCall( cudaBindTexture2D(0, thogluv, hogluv.data, desc, hogluv.cols, hogluv.rows, hogluv.step));
cudaChannelFormatDesc desc_roi = cudaCreateChannelDesc<typename Policy::roi_type>();
cudaSafeCall( cudaBindTexture2D(0, troi, roi.data, desc_roi, roi.cols / Policy::STA_Y, roi.rows, roi.step));
const CascadeInvoker<Policy> inv = *this;
soft_cascade<Policy, false><<<grid, Policy::block(), 0, stream>>>(inv, det, max_det, ctr, 0);
cudaSafeCall( cudaGetLastError());
grid = dim3(fw, fh / Policy::STA_Y, scales - downscales);
soft_cascade<Policy, true><<<grid, Policy::block(), 0, stream>>>(inv, det, max_det, ctr, downscales);
if (!stream)
{
cudaSafeCall( cudaGetLastError());
cudaSafeCall( cudaDeviceSynchronize());
}
}
template void CascadeInvoker<GK107PolicyX4>::operator()(const PtrStepSzb& roi, const PtrStepSzi& hogluv,
PtrStepSz<uchar4> objects, const int downscales, const cudaStream_t& stream) const;
}
}}}

82
modules/gpu/src/cuda/integral_image.cu
View File

@ -384,6 +384,88 @@ namespace cv { namespace gpu { namespace device
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
__global__ void shfl_integral_vertical(PtrStepSz<unsigned int> buffer, PtrStepSz<unsigned int> integral)
{
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 300)
__shared__ unsigned int sums[32][9];
const int tidx = blockIdx.x * blockDim.x + threadIdx.x;
const int lane_id = tidx % 8;
if (tidx >= integral.cols)
return;
sums[threadIdx.x][threadIdx.y] = 0;
__syncthreads();
unsigned int stepSum = 0;
for (int y = threadIdx.y; y < integral.rows; y += blockDim.y)
{
unsigned int* p = buffer.ptr(y) + tidx;
unsigned int* dst = integral.ptr(y + 1) + tidx + 1;
unsigned int sum = *p;
sums[threadIdx.x][threadIdx.y] = sum;
__syncthreads();
// place into SMEM
// shfl scan reduce the SMEM, reformating so the column
// sums are computed in a warp
// then read out properly
const int j = threadIdx.x % 8;
const int k = threadIdx.x / 8 + threadIdx.y * 4;
int partial_sum = sums[k][j];
for (int i = 1; i <= 8; i *= 2)
{
int n = __shfl_up(partial_sum, i, 32);
if (lane_id >= i)
partial_sum += n;
}
sums[k][j] = partial_sum;
__syncthreads();
if (threadIdx.y > 0)
sum += sums[threadIdx.x][threadIdx.y - 1];
sum += stepSum;
stepSum += sums[threadIdx.x][blockDim.y - 1];
__syncthreads();
*dst = sum;
}
#endif
}
// used for frame preprocessing before Soft Cascade evaluation: no synchronization needed
void shfl_integral_gpu_buffered(PtrStepSzb img, PtrStepSz<uint4> buffer, PtrStepSz<unsigned int> integral,
int blockStep, cudaStream_t stream)
{
{
const int block = blockStep;
const int grid = img.rows;
cudaSafeCall( cudaFuncSetCacheConfig(shfl_integral_horizontal, cudaFuncCachePreferL1) );
shfl_integral_horizontal<<<grid, block, 0, stream>>>((PtrStepSz<uint4>) img, buffer);
cudaSafeCall( cudaGetLastError() );
}
{
const dim3 block(32, 8);
const dim3 grid(divUp(integral.cols, block.x), 1);
shfl_integral_vertical<<<grid, block, 0, stream>>>((PtrStepSz<uint>)buffer, integral);
cudaSafeCall( cudaGetLastError() );
}
}
}
}}}

8
modules/gpu/src/cuda/mathfunc.cu
View File

@ -76,7 +76,7 @@ namespace cv { namespace gpu { namespace device
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;
angle += (angle < 0) * 2.0f * CV_PI_F;
dst[y * dst_step + x] = scale * angle;
}
};
@ -140,7 +140,7 @@ namespace cv { namespace gpu { namespace device
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;
const float scale = angleInDegrees ? (180.0f / CV_PI_F) : 1.f;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
@ -190,7 +190,7 @@ namespace cv { namespace gpu { namespace device
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;
const float scale = angleInDegrees ? (CV_PI_F / 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);
@ -214,4 +214,4 @@ namespace cv { namespace gpu { namespace device
} // namespace mathfunc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

3221
modules/gpu/src/cuda/matrix_reductions.cu
View File

File diff suppressed because it is too large Load Diff

116
modules/gpu/src/cuda/nlm.cu
View File

@ -43,11 +43,11 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/block.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/reduce.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
using namespace cv::gpu;
@ -184,6 +184,85 @@ namespace cv { namespace gpu { namespace device
{
namespace imgproc
{
template <int cn> struct Unroll;
template <> struct Unroll<1>
{
template <int BLOCK_SIZE>
static __device__ __forceinline__ thrust::tuple<volatile float*, volatile float*> smem_tuple(float* smem)
{
return cv::gpu::device::smem_tuple(smem, smem + BLOCK_SIZE);
}
static __device__ __forceinline__ thrust::tuple<float&, float&> tie(float& val1, float& val2)
{
return thrust::tie(val1, val2);
}
static __device__ __forceinline__ const thrust::tuple<plus<float>, plus<float> > op()
{
plus<float> op;
return thrust::make_tuple(op, op);
}
};
template <> struct Unroll<2>
{
template <int BLOCK_SIZE>
static __device__ __forceinline__ thrust::tuple<volatile float*, volatile float*, volatile float*> smem_tuple(float* smem)
{
return cv::gpu::device::smem_tuple(smem, smem + BLOCK_SIZE, smem + 2 * BLOCK_SIZE);
}
static __device__ __forceinline__ thrust::tuple<float&, float&, float&> tie(float& val1, float2& val2)
{
return thrust::tie(val1, val2.x, val2.y);
}
static __device__ __forceinline__ const thrust::tuple<plus<float>, plus<float>, plus<float> > op()
{
plus<float> op;
return thrust::make_tuple(op, op, op);
}
};
template <> struct Unroll<3>
{
template <int BLOCK_SIZE>
static __device__ __forceinline__ thrust::tuple<volatile float*, volatile float*, volatile float*, volatile float*> smem_tuple(float* smem)
{
return cv::gpu::device::smem_tuple(smem, smem + BLOCK_SIZE, smem + 2 * BLOCK_SIZE, smem + 3 * BLOCK_SIZE);
}
static __device__ __forceinline__ thrust::tuple<float&, float&, float&, float&> tie(float& val1, float3& val2)
{
return thrust::tie(val1, val2.x, val2.y, val2.z);
}
static __device__ __forceinline__ const thrust::tuple<plus<float>, plus<float>, plus<float>, plus<float> > op()
{
plus<float> op;
return thrust::make_tuple(op, op, op, op);
}
};
template <> struct Unroll<4>
{
template <int BLOCK_SIZE>
static __device__ __forceinline__ thrust::tuple<volatile float*, volatile float*, volatile float*, volatile float*, volatile float*> smem_tuple(float* smem)
{
return cv::gpu::device::smem_tuple(smem, smem + BLOCK_SIZE, smem + 2 * BLOCK_SIZE, smem + 3 * BLOCK_SIZE, smem + 4 * BLOCK_SIZE);
}
static __device__ __forceinline__ thrust::tuple<float&, float&, float&, float&, float&> tie(float& val1, float4& val2)
{
return thrust::tie(val1, val2.x, val2.y, val2.z, val2.w);
}
static __device__ __forceinline__ const thrust::tuple<plus<float>, plus<float>, plus<float>, plus<float>, plus<float> > op()
{
plus<float> op;
return thrust::make_tuple(op, op, op, op, op);
}
};
__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); }
@ -340,30 +419,15 @@ 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];
__shared__ float cta_buffer[CTA_SIZE * (VecTraits<T>::cn + 1)];
int tid = threadIdx.x;
reduce<CTA_SIZE>(Unroll<VecTraits<T>::cn>::template smem_tuple<CTA_SIZE>(cta_buffer),
Unroll<VecTraits<T>::cn>::tie(weights_sum, sum),
threadIdx.x,
Unroll<VecTraits<T>::cn>::op());
cta_buffer[tid] = weights_sum;
__syncthreads();
Block::reduce<CTA_SIZE>(cta_buffer, plus());
weights_sum = cta_buffer[0];
__syncthreads();
for(int n = 0; n < VecTraits<T>::cn; ++n)
{
cta_buffer[tid] = reinterpret_cast<float*>(&sum)[n];
__syncthreads();
Block::reduce<CTA_SIZE>(cta_buffer, plus());
reinterpret_cast<float*>(&sum)[n] = cta_buffer[0];
__syncthreads();
}
if (tid == 0)
dst = saturate_cast<T>(sum/weights_sum);
if (threadIdx.x == 0)
dst = saturate_cast<T>(sum / weights_sum);
}
__device__ __forceinline__ void operator()(PtrStepSz<T>& dst) const
@ -503,4 +567,4 @@ namespace cv { namespace gpu { namespace device
}}}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

414
modules/gpu/src/cuda/optflowbm.cu Normal file
View File

@ -0,0 +1,414 @@
/*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/limits.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/reduce.hpp"
using namespace cv::gpu;
using namespace cv::gpu::device;
namespace optflowbm
{
texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_prev(false, cudaFilterModePoint, cudaAddressModeClamp);
texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_curr(false, cudaFilterModePoint, cudaAddressModeClamp);
__device__ int cmpBlocks(int X1, int Y1, int X2, int Y2, int2 blockSize)
{
int s = 0;
for (int y = 0; y < blockSize.y; ++y)
{
for (int x = 0; x < blockSize.x; ++x)
s += ::abs(tex2D(tex_prev, X1 + x, Y1 + y) - tex2D(tex_curr, X2 + x, Y2 + y));
}
return s;
}
__global__ void calcOptFlowBM(PtrStepSzf velx, PtrStepf vely, const int2 blockSize, const int2 shiftSize, const bool usePrevious,
const int maxX, const int maxY, const int acceptLevel, const int escapeLevel,
const short2* ss, const int ssCount)
{
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (i >= velx.rows || j >= velx.cols)
return;
const int X1 = j * shiftSize.x;
const int Y1 = i * shiftSize.y;
const int offX = usePrevious ? __float2int_rn(velx(i, j)) : 0;
const int offY = usePrevious ? __float2int_rn(vely(i, j)) : 0;
int X2 = X1 + offX;
int Y2 = Y1 + offY;
int dist = numeric_limits<int>::max();
if (0 <= X2 && X2 <= maxX && 0 <= Y2 && Y2 <= maxY)
dist = cmpBlocks(X1, Y1, X2, Y2, blockSize);
int countMin = 1;
int sumx = offX;
int sumy = offY;
if (dist > acceptLevel)
{
// do brute-force search
for (int k = 0; k < ssCount; ++k)
{
const short2 ssVal = ss[k];
const int dx = offX + ssVal.x;
const int dy = offY + ssVal.y;
X2 = X1 + dx;
Y2 = Y1 + dy;
if (0 <= X2 && X2 <= maxX && 0 <= Y2 && Y2 <= maxY)
{
const int tmpDist = cmpBlocks(X1, Y1, X2, Y2, blockSize);
if (tmpDist < acceptLevel)
{
sumx = dx;
sumy = dy;
countMin = 1;
break;
}
if (tmpDist < dist)
{
dist = tmpDist;
sumx = dx;
sumy = dy;
countMin = 1;
}
else if (tmpDist == dist)
{
sumx += dx;
sumy += dy;
countMin++;
}
}
}
if (dist > escapeLevel)
{
sumx = offX;
sumy = offY;
countMin = 1;
}
}
velx(i, j) = static_cast<float>(sumx) / countMin;
vely(i, j) = static_cast<float>(sumy) / countMin;
}
void calc(PtrStepSzb prev, PtrStepSzb curr, PtrStepSzf velx, PtrStepSzf vely, int2 blockSize, int2 shiftSize, bool usePrevious,
int maxX, int maxY, int acceptLevel, int escapeLevel, const short2* ss, int ssCount, cudaStream_t stream)
{
bindTexture(&tex_prev, prev);
bindTexture(&tex_curr, curr);
const dim3 block(32, 8);
const dim3 grid(divUp(velx.cols, block.x), divUp(vely.rows, block.y));
calcOptFlowBM<<<grid, block, 0, stream>>>(velx, vely, blockSize, shiftSize, usePrevious,
maxX, maxY, acceptLevel, escapeLevel, ss, ssCount);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
/////////////////////////////////////////////////////////
// Fast approximate version
namespace optflowbm_fast
{
enum
{
CTA_SIZE = 128,
TILE_COLS = 128,
TILE_ROWS = 32,
STRIDE = CTA_SIZE
};
template <typename T> __device__ __forceinline__ int calcDist(T a, T b)
{
return ::abs(a - b);
}
template <class T> struct FastOptFlowBM
{
int search_radius;
int block_radius;
int search_window;
int block_window;
PtrStepSz<T> I0;
PtrStep<T> I1;
mutable PtrStepi buffer;
FastOptFlowBM(int search_window_, int block_window_,
PtrStepSz<T> I0_, PtrStepSz<T> I1_,
PtrStepi buffer_) :
search_radius(search_window_ / 2), block_radius(block_window_ / 2),
search_window(search_window_), block_window(block_window_),
I0(I0_), I1(I1_),
buffer(buffer_)
{
}
__device__ 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)
{
dist_sums[index] = 0;
for (int tx = 0; tx < block_window; ++tx)
col_sums(tx, index) = 0;
int y = index / search_window;
int x = index - y * search_window;
int ay = i;
int ax = j;
int by = i + y - search_radius;
int bx = j + x - search_radius;
for (int tx = -block_radius; tx <= block_radius; ++tx)
{
int col_sum = 0;
for (int ty = -block_radius; ty <= block_radius; ++ty)
{
int dist = calcDist(I0(ay + ty, ax + tx), I1(by + ty, bx + tx));
dist_sums[index] += dist;
col_sum += dist;
}
col_sums(tx + block_radius, index) = col_sum;
}
up_col_sums(j, index) = col_sums(block_window - 1, index);
}
}
__device__ 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;
int ay = i;
int ax = j + block_radius;
int by = i + y - search_radius;
int bx = j + x - search_radius + block_radius;
int col_sum = 0;
for (int ty = -block_radius; ty <= block_radius; ++ty)
col_sum += calcDist(I0(ay + ty, ax), I1(by + ty, bx));
dist_sums[index] += col_sum - col_sums(first, index);
col_sums(first, index) = col_sum;
up_col_sums(j, index) = col_sum;
}
}
__device__ void shiftRight_UpSums(int i, int j, int first, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
{
int ay = i;
int ax = j + block_radius;
T a_up = I0(ay - block_radius - 1, ax);
T a_down = I0(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;
int by = i + y - search_radius;
int bx = j + x - search_radius + block_radius;
T b_up = I1(by - block_radius - 1, bx);
T b_down = I1(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;
up_col_sums(j, index) = col_sum;
}
}
__device__ void convolve_window(int i, int j, const int* dist_sums, float& velx, float& vely) const
{
int bestDist = numeric_limits<int>::max();
int bestInd = -1;
for (int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
{
int curDist = dist_sums[index];
if (curDist < bestDist)
{
bestDist = curDist;
bestInd = index;
}
}
__shared__ int cta_dist_buffer[CTA_SIZE];
__shared__ int cta_ind_buffer[CTA_SIZE];
reduceKeyVal<CTA_SIZE>(cta_dist_buffer, bestDist, cta_ind_buffer, bestInd, threadIdx.x, less<int>());
if (threadIdx.x == 0)
{
int y = bestInd / search_window;
int x = bestInd - y * search_window;
velx = x - search_radius;
vely = y - search_radius;
}
}
__device__ void operator()(PtrStepf velx, PtrStepf vely) const
{
int tbx = blockIdx.x * TILE_COLS;
int tby = blockIdx.y * TILE_ROWS;
int tex = ::min(tbx + TILE_COLS, I0.cols);
int tey = ::min(tby + TILE_ROWS, I0.rows);
PtrStepi col_sums;
col_sums.data = buffer.ptr(I0.cols + blockIdx.x * block_window) + blockIdx.y * search_window * search_window;
col_sums.step = buffer.step;
PtrStepi up_col_sums;
up_col_sums.data = buffer.data + blockIdx.y * search_window * search_window;
up_col_sums.step = buffer.step;
extern __shared__ int dist_sums[]; //search_window * search_window
int first = 0;
for (int i = tby; i < tey; ++i)
{
for (int j = tbx; j < tex; ++j)
{
__syncthreads();
if (j == tbx)
{
initSums_BruteForce(i, j, dist_sums, col_sums, up_col_sums);
first = 0;
}
else
{
if (i == tby)
shiftRight_FirstRow(i, j, first, dist_sums, col_sums, up_col_sums);
else
shiftRight_UpSums(i, j, first, dist_sums, col_sums, up_col_sums);
first = (first + 1) % block_window;
}
__syncthreads();
convolve_window(i, j, dist_sums, velx(i, j), vely(i, j));
}
}
}
};
template<typename T> __global__ void optflowbm_fast_kernel(const FastOptFlowBM<T> fbm, PtrStepf velx, PtrStepf vely)
{
fbm(velx, vely);
}
void get_buffer_size(int src_cols, int src_rows, int search_window, int block_window, int& buffer_cols, int& buffer_rows)
{
dim3 grid(divUp(src_cols, TILE_COLS), divUp(src_rows, TILE_ROWS));
buffer_cols = search_window * search_window * grid.y;
buffer_rows = src_cols + block_window * grid.x;
}
template <typename T>
void calc(PtrStepSzb I0, PtrStepSzb I1, PtrStepSzf velx, PtrStepSzf vely, PtrStepi buffer, int search_window, int block_window, cudaStream_t stream)
{
FastOptFlowBM<T> fbm(search_window, block_window, I0, I1, buffer);
dim3 block(CTA_SIZE, 1);
dim3 grid(divUp(I0.cols, TILE_COLS), divUp(I0.rows, TILE_ROWS));
size_t smem = search_window * search_window * sizeof(int);
optflowbm_fast_kernel<<<grid, block, smem, stream>>>(fbm, velx, vely);
cudaSafeCall ( cudaGetLastError () );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void calc<uchar>(PtrStepSzb I0, PtrStepSzb I1, PtrStepSzf velx, PtrStepSzf vely, PtrStepi buffer, int search_window, int block_window, cudaStream_t stream);
}
#endif // !defined CUDA_DISABLER

22
modules/gpu/src/cuda/optical_flow.cu
View File

@ -164,40 +164,40 @@ namespace cv { namespace gpu { namespace device
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[1].x = arrow_x + r * ::cosf(theta - CV_PI_F / 2.0f);
v[1].y = arrow_y + r * ::sinf(theta - CV_PI_F / 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);
v[4].x = arrow_x + r * ::cosf(theta + CV_PI_F / 2.0f);
v[4].y = arrow_y + r * ::sinf(theta + CV_PI_F / 2.0f);
int indx = (y * u_avg.cols + x) * NUM_VERTS_PER_ARROW * 3;
color_data[indx] = (theta - CV_PI) / CV_PI * 180.0f;
color_data[indx] = (theta - CV_PI_F) / CV_PI_F * 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;
color_data[indx] = (theta - CV_PI_F) / CV_PI_F * 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;
color_data[indx] = (theta - CV_PI_F) / CV_PI_F * 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;
color_data[indx] = (theta - CV_PI_F) / CV_PI_F * 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;
color_data[indx] = (theta - CV_PI_F) / CV_PI_F * 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;
color_data[indx] = (theta - CV_PI_F) / CV_PI_F * 180.0f;
vertex_data[indx++] = v[5].x * xscale;
vertex_data[indx++] = v[5].y * yscale;
vertex_data[indx++] = v[5].z;
@ -217,4 +217,4 @@ namespace cv { namespace gpu { namespace device
}
}}}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

3
modules/gpu/src/cuda/optical_flow_farneback.cu
View File

@ -42,7 +42,6 @@
#if !defined CUDA_DISABLER
#include <stdio.h>
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
@ -57,8 +56,6 @@
#define BORDER_SIZE 5
#define MAX_KSIZE_HALF 100
using namespace std;
namespace cv { namespace gpu { namespace device { namespace optflow_farneback
{
__constant__ float c_g[8];

33
modules/gpu/src/cuda/orb.cu
View File

@ -47,10 +47,11 @@
#if !defined CUDA_DISABLER
#include <thrust/device_ptr.h>
#include <thrust/sort.h>
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/reduce.hpp"
#include "opencv2/gpu/device/functional.hpp"
namespace cv { namespace gpu { namespace device
@ -75,9 +76,9 @@ namespace cv { namespace gpu { namespace device
__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;
__shared__ int smem0[8 * 32];
__shared__ int smem1[8 * 32];
__shared__ int smem2[8 * 32];
const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
@ -109,9 +110,12 @@ namespace cv { namespace gpu { namespace device
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>());
int* srow0 = smem0 + threadIdx.y * blockDim.x;
int* srow1 = smem1 + threadIdx.y * blockDim.x;
int* srow2 = smem2 + threadIdx.y * blockDim.x;
plus<int> op;
reduce<32>(smem_tuple(srow0, srow1, srow2), thrust::tie(a, b, c), threadIdx.x, thrust::make_tuple(op, op, op));
if (threadIdx.x == 0)
{
@ -151,9 +155,13 @@ namespace cv { namespace gpu { namespace device
__global__ void IC_Angle(const PtrStepb image, const short2* loc_, float* angle, const int npoints, const int half_k)
{
__shared__ int smem[8 * 32];
__shared__ int smem0[8 * 32];
__shared__ int smem1[8 * 32];
volatile int* srow = smem + threadIdx.y * blockDim.x;
int* srow0 = smem0 + threadIdx.y * blockDim.x;
int* srow1 = smem1 + threadIdx.y * blockDim.x;
plus<int> op;
const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
@ -167,7 +175,7 @@ namespace cv { namespace gpu { namespace device
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>());
reduce<32>(srow0, m_10, threadIdx.x, op);
for (int v = 1; v <= half_k; ++v)
{
@ -185,8 +193,7 @@ namespace cv { namespace gpu { namespace device
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>());
reduce<32>(smem_tuple(srow0, srow1), thrust::tie(v_sum, m_sum), threadIdx.x, thrust::make_tuple(op, op));
m_10 += m_sum;
m_01 += v * v_sum;
@ -419,4 +426,4 @@ namespace cv { namespace gpu { namespace device
}
}}}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

967
modules/gpu/src/cuda/pyrlk.cu
View File

File diff suppressed because it is too large Load Diff

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

@ -69,7 +69,7 @@ namespace cv { namespace gpu { namespace device
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)
static void call(PtrStepSz<T> src, PtrStepSzf mapx, PtrStepSzf mapy, PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, bool)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
@ -87,7 +87,7 @@ namespace cv { namespace gpu { namespace device
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)
static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSzf mapx, PtrStepSzf mapy, PtrStepSz<T> dst, const float* borderValue, bool)
{
(void)srcWhole;
(void)xoff;
@ -124,10 +124,10 @@ namespace cv { namespace gpu { namespace device
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) \
PtrStepSz< type > dst, const float* borderValue, bool cc20) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 block(32, cc20 ? 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); \
@ -142,7 +142,7 @@ namespace cv { namespace gpu { namespace device
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) \
PtrStepSz< type > dst, const float*, bool) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
@ -194,20 +194,20 @@ namespace cv { namespace gpu { namespace device
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)
PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, bool cc20)
{
if (stream == 0)
RemapDispatcherNonStream<Filter, B, T>::call(src, srcWhole, xoff, yoff, mapx, mapy, dst, borderValue, cc);
RemapDispatcherNonStream<Filter, B, T>::call(src, srcWhole, xoff, yoff, mapx, mapy, dst, borderValue, cc20);
else
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc);
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc20);
}
};
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)
PtrStepSzb dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, bool cc20)
{
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);
PtrStepSz<T> dst, const float* borderValue, cudaStream_t stream, bool cc20);
static const caller_t callers[3][5] =
{
@ -235,40 +235,40 @@ namespace cv { namespace gpu { namespace device
};
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);
static_cast< PtrStepSz<T> >(dst), borderValue, stream, cc20);
}
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<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, bool cc20);
//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, bool cc20);
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, bool cc20);
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, bool cc20);
//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<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, bool cc20);
//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, bool cc20);
//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, bool cc20);
//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, bool cc20);
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<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, bool cc20);
//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, bool cc20);
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, bool cc20);
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, bool cc20);
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<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, bool cc20);
//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, bool cc20);
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, bool cc20);
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, bool cc20);
//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<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, bool cc20);
//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, bool cc20);
//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, bool cc20);
//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, bool cc20);
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);
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, bool cc20);
//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, bool cc20);
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, bool cc20);
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, bool cc20);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.0.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<uchar, float>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.1.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<uchar3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.10.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<unsigned short, float>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.11.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<ushort3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.12.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<ushort4, float4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.13.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<int3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.14.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<int4, float4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.2.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<uchar4, float4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.3.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<short3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.4.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<int, float>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.5.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<float, float>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.6.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<float3, float3>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

53
modules/gpu/src/cuda/row_filter.7.cu Normal file
View File

@ -0,0 +1,53 @@
/*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 "row_filter.h"
namespace filter
{
template void linearRow<float4, float4>(PtrStepSzb src, PtrStepSzb dst, const float* kernel, int ksize, int anchor, int brd_type, int cc, cudaStream_t stream);
}
#endif /* CUDA_DISABLER */

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