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

moved integral to gpuarithm module

Browse Source
This commit is contained in:
Vladislav Vinogradov 2013-04-09 14:18:18 +04:00
parent 10ac854358
commit c56bdbc1c5
11 changed files with 236 additions and 219 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -168,18 +168,6 @@ CV_EXPORTS void buildWarpSphericalMaps(Size src_size, Rect dst_roi, const Mat &K
CV_EXPORTS void rotate(const GpuMat& src, GpuMat& dst, Size dsize, double angle, double xShift = 0, double yShift = 0,
int interpolation = INTER_LINEAR, Stream& stream = Stream::Null());
//! computes the integral image
//! sum will have CV_32S type, but will contain unsigned int values
//! supports only CV_8UC1 source type
CV_EXPORTS void integral(const GpuMat& src, GpuMat& sum, Stream& stream = Stream::Null());
//! buffered version
CV_EXPORTS void integralBuffered(const GpuMat& src, GpuMat& sum, GpuMat& buffer, Stream& stream = Stream::Null());
//! computes squared integral image
//! result matrix will have 64F type, but will contain 64U values
//! supports source images of 8UC1 type only
CV_EXPORTS void sqrIntegral(const GpuMat& src, GpuMat& sqsum, Stream& stream = Stream::Null());
//! computes Harris cornerness criteria at each image pixel
CV_EXPORTS void cornerHarris(const GpuMat& src, GpuMat& dst, int blockSize, int ksize, double k, int borderType = BORDER_REFLECT101);
CV_EXPORTS void cornerHarris(const GpuMat& src, GpuMat& dst, GpuMat& Dx, GpuMat& Dy, int blockSize, int ksize, double k, int borderType = BORDER_REFLECT101);

57
modules/gpu/perf/perf_imgproc.cpp
View File

@ -363,63 +363,6 @@ PERF_TEST_P(Sz_Depth_Op, ImgProc_Threshold,
}
}
//////////////////////////////////////////////////////////////////////
// Integral
PERF_TEST_P(Sz, ImgProc_Integral,
GPU_TYPICAL_MAT_SIZES)
{
const cv::Size size = GetParam();
cv::Mat src(size, CV_8UC1);
declare.in(src, WARMUP_RNG);
if (PERF_RUN_GPU())
{
const cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat dst;
cv::gpu::GpuMat d_buf;
TEST_CYCLE() cv::gpu::integralBuffered(d_src, dst, d_buf);
GPU_SANITY_CHECK(dst);
}
else
{
cv::Mat dst;
TEST_CYCLE() cv::integral(src, dst);
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// IntegralSqr
PERF_TEST_P(Sz, ImgProc_IntegralSqr,
GPU_TYPICAL_MAT_SIZES)
{
const cv::Size size = GetParam();
cv::Mat src(size, CV_8UC1);
declare.in(src, WARMUP_RNG);
if (PERF_RUN_GPU())
{
const cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat dst;
TEST_CYCLE() cv::gpu::sqrIntegral(d_src, dst);
GPU_SANITY_CHECK(dst);
}
else
{
FAIL_NO_CPU();
}
}
//////////////////////////////////////////////////////////////////////
// HistEvenC1

108
modules/gpu/src/imgproc.cpp
View File

@ -55,9 +55,6 @@ void cv::gpu::buildWarpPlaneMaps(Size, Rect, const Mat&, const Mat&, const Mat&,
void cv::gpu::buildWarpCylindricalMaps(Size, Rect, const Mat&, const Mat&, float, GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::gpu::buildWarpSphericalMaps(Size, Rect, const Mat&, const Mat&, float, GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::gpu::rotate(const GpuMat&, GpuMat&, Size, double, double, double, int, Stream&) { throw_no_cuda(); }
void cv::gpu::integral(const GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::gpu::integralBuffered(const GpuMat&, GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::gpu::sqrIntegral(const GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::gpu::evenLevels(GpuMat&, int, int, int) { throw_no_cuda(); }
void cv::gpu::histEven(const GpuMat&, GpuMat&, int, int, int, Stream&) { throw_no_cuda(); }
void cv::gpu::histEven(const GpuMat&, GpuMat&, GpuMat&, int, int, int, Stream&) { throw_no_cuda(); }
@ -412,111 +409,6 @@ void cv::gpu::rotate(const GpuMat& src, GpuMat& dst, Size dsize, double angle, d
funcs[src.depth()][src.channels() - 1](src, dst, dsize, angle, xShift, yShift, interpolation, StreamAccessor::getStream(stream));
}
////////////////////////////////////////////////////////////////////////
// integral
void cv::gpu::integral(const GpuMat& src, GpuMat& sum, Stream& s)
{
GpuMat buffer;
integralBuffered(src, sum, buffer, s);
}
namespace cv { namespace gpu { namespace cudev
{
namespace imgproc
{
void shfl_integral_gpu(const PtrStepSzb& img, PtrStepSz<unsigned int> integral, cudaStream_t stream);
}
}}}
void cv::gpu::integralBuffered(const GpuMat& src, GpuMat& sum, GpuMat& buffer, Stream& s)
{
CV_Assert(src.type() == CV_8UC1);
cudaStream_t stream = StreamAccessor::getStream(s);
cv::Size whole;
cv::Point offset;
src.locateROI(whole, offset);
if (deviceSupports(WARP_SHUFFLE_FUNCTIONS) && src.cols <= 2048
&& offset.x % 16 == 0 && ((src.cols + 63) / 64) * 64 <= (static_cast<int>(src.step) - offset.x))
{
ensureSizeIsEnough(((src.rows + 7) / 8) * 8, ((src.cols + 63) / 64) * 64, CV_32SC1, buffer);
cv::gpu::cudev::imgproc::shfl_integral_gpu(src, buffer, stream);
sum.create(src.rows + 1, src.cols + 1, CV_32SC1);
if (s)
s.enqueueMemSet(sum, Scalar::all(0));
else
sum.setTo(Scalar::all(0));
GpuMat inner = sum(Rect(1, 1, src.cols, src.rows));
GpuMat res = buffer(Rect(0, 0, src.cols, src.rows));
if (s)
s.enqueueCopy(res, inner);
else
res.copyTo(inner);
}
else
{
sum.create(src.rows + 1, src.cols + 1, CV_32SC1);
NcvSize32u roiSize;
roiSize.width = src.cols;
roiSize.height = src.rows;
cudaDeviceProp prop;
cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) );
Ncv32u bufSize;
ncvSafeCall( nppiStIntegralGetSize_8u32u(roiSize, &bufSize, prop) );
ensureSizeIsEnough(1, bufSize, CV_8UC1, buffer);
NppStStreamHandler h(stream);
ncvSafeCall( nppiStIntegral_8u32u_C1R(const_cast<Ncv8u*>(src.ptr<Ncv8u>()), static_cast<int>(src.step),
sum.ptr<Ncv32u>(), static_cast<int>(sum.step), roiSize, buffer.ptr<Ncv8u>(), bufSize, prop) );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
//////////////////////////////////////////////////////////////////////////////
// sqrIntegral
void cv::gpu::sqrIntegral(const GpuMat& src, GpuMat& sqsum, Stream& s)
{
CV_Assert(src.type() == CV_8U);
NcvSize32u roiSize;
roiSize.width = src.cols;
roiSize.height = src.rows;
cudaDeviceProp prop;
cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) );
Ncv32u bufSize;
ncvSafeCall(nppiStSqrIntegralGetSize_8u64u(roiSize, &bufSize, prop));
GpuMat buf(1, bufSize, CV_8U);
cudaStream_t stream = StreamAccessor::getStream(s);
NppStStreamHandler h(stream);
sqsum.create(src.rows + 1, src.cols + 1, CV_64F);
ncvSafeCall(nppiStSqrIntegral_8u64u_C1R(const_cast<Ncv8u*>(src.ptr<Ncv8u>(0)), static_cast<int>(src.step),
sqsum.ptr<Ncv64u>(0), static_cast<int>(sqsum.step), roiSize, buf.ptr<Ncv8u>(0), bufSize, prop));
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////
// Histogram

37
modules/gpu/test/test_imgproc.cpp
View File

@ -46,43 +46,6 @@
using namespace cvtest;
///////////////////////////////////////////////////////////////////////////////////////////////////////
// Integral
PARAM_TEST_CASE(Integral, cv::gpu::DeviceInfo, cv::Size, UseRoi)
{
cv::gpu::DeviceInfo devInfo;
cv::Size size;
bool useRoi;
virtual void SetUp()
{
devInfo = GET_PARAM(0);
size = GET_PARAM(1);
useRoi = GET_PARAM(2);
cv::gpu::setDevice(devInfo.deviceID());
}
};
GPU_TEST_P(Integral, Accuracy)
{
cv::Mat src = randomMat(size, CV_8UC1);
cv::gpu::GpuMat dst = createMat(cv::Size(src.cols + 1, src.rows + 1), CV_32SC1, useRoi);
cv::gpu::integral(loadMat(src, useRoi), dst);
cv::Mat dst_gold;
cv::integral(src, dst_gold, CV_32S);
EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
}
INSTANTIATE_TEST_CASE_P(GPU_ImgProc, Integral, testing::Combine(
ALL_DEVICES,
DIFFERENT_SIZES,
WHOLE_SUBMAT));
///////////////////////////////////////////////////////////////////////////////////////////////////////
// HistEven

2
modules/gpuarithm/CMakeLists.txt
View File

@ -6,7 +6,7 @@ set(the_description "GPU-accelerated Operations on Matrices")
ocv_warnings_disable(CMAKE_CXX_FLAGS -Wundef -Wmissing-declarations)
ocv_define_module(gpuarithm opencv_core OPTIONAL opencv_imgproc)
ocv_define_module(gpuarithm opencv_core OPTIONAL opencv_gpunvidia opencv_imgproc)
if(HAVE_CUBLAS)
CUDA_ADD_CUBLAS_TO_TARGET(${the_module})

12
modules/gpuarithm/include/opencv2/gpuarithm.hpp
View File

@ -283,6 +283,18 @@ CV_EXPORTS void rectStdDev(const GpuMat& src, const GpuMat& sqr, GpuMat& dst, co
CV_EXPORTS void copyMakeBorder(const GpuMat& src, GpuMat& dst, int top, int bottom, int left, int right, int borderType,
const Scalar& value = Scalar(), Stream& stream = Stream::Null());
//! computes the integral image
//! sum will have CV_32S type, but will contain unsigned int values
//! supports only CV_8UC1 source type
CV_EXPORTS void integral(const GpuMat& src, GpuMat& sum, Stream& stream = Stream::Null());
//! buffered version
CV_EXPORTS void integralBuffered(const GpuMat& src, GpuMat& sum, GpuMat& buffer, Stream& stream = Stream::Null());
//! computes squared integral image
//! result matrix will have 64F type, but will contain 64U values
//! supports source images of 8UC1 type only
CV_EXPORTS void sqrIntegral(const GpuMat& src, GpuMat& sqsum, Stream& stream = Stream::Null());
}} // namespace cv { namespace gpu {
#endif /* __OPENCV_GPUARITHM_HPP__ */

61
modules/gpuarithm/perf/perf_core.cpp
View File

@ -2156,11 +2156,11 @@ PERF_TEST_P(Sz_Depth_NormType, Core_Normalize,
}
}
#ifdef HAVE_OPENCV_IMGPROC
//////////////////////////////////////////////////////////////////////
// CopyMakeBorder
#ifdef HAVE_OPENCV_IMGPROC
DEF_PARAM_TEST(Sz_Depth_Cn_Border, cv::Size, MatDepth, MatCn, BorderMode);
PERF_TEST_P(Sz_Depth_Cn_Border, ImgProc_CopyMakeBorder,
@ -2198,4 +2198,61 @@ PERF_TEST_P(Sz_Depth_Cn_Border, ImgProc_CopyMakeBorder,
}
}
//////////////////////////////////////////////////////////////////////
// Integral
PERF_TEST_P(Sz, ImgProc_Integral,
GPU_TYPICAL_MAT_SIZES)
{
const cv::Size size = GetParam();
cv::Mat src(size, CV_8UC1);
declare.in(src, WARMUP_RNG);
if (PERF_RUN_GPU())
{
const cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat dst;
cv::gpu::GpuMat d_buf;
TEST_CYCLE() cv::gpu::integralBuffered(d_src, dst, d_buf);
GPU_SANITY_CHECK(dst);
}
else
{
cv::Mat dst;
TEST_CYCLE() cv::integral(src, dst);
CPU_SANITY_CHECK(dst);
}
}
//////////////////////////////////////////////////////////////////////
// IntegralSqr
PERF_TEST_P(Sz, ImgProc_IntegralSqr,
GPU_TYPICAL_MAT_SIZES)
{
const cv::Size size = GetParam();
cv::Mat src(size, CV_8UC1);
declare.in(src, WARMUP_RNG);
if (PERF_RUN_GPU())
{
const cv::gpu::GpuMat d_src(src);
cv::gpu::GpuMat dst;
TEST_CYCLE() cv::gpu::sqrIntegral(d_src, dst);
GPU_SANITY_CHECK(dst);
}
else
{
FAIL_NO_CPU();
}
}
#endif

118
modules/gpuarithm/src/arithm.cpp
View File

@ -61,6 +61,9 @@ void cv::gpu::polarToCart(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, bool,
void cv::gpu::normalize(const GpuMat&, GpuMat&, double, double, int, int, const GpuMat&) { throw_no_cuda(); }
void cv::gpu::normalize(const GpuMat&, GpuMat&, double, double, int, int, const GpuMat&, GpuMat&, GpuMat&) { throw_no_cuda(); }
void cv::gpu::copyMakeBorder(const GpuMat&, GpuMat&, int, int, int, int, int, const Scalar&, Stream&) { throw_no_cuda(); }
void cv::gpu::integral(const GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::gpu::integralBuffered(const GpuMat&, GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::gpu::sqrIntegral(const GpuMat&, GpuMat&, Stream&) { throw_no_cuda(); }
#else /* !defined (HAVE_CUDA) */
@ -718,4 +721,119 @@ void cv::gpu::copyMakeBorder(const GpuMat& src, GpuMat& dst, int top, int bottom
}
}
////////////////////////////////////////////////////////////////////////
// integral
void cv::gpu::integral(const GpuMat& src, GpuMat& sum, Stream& s)
{
GpuMat buffer;
integralBuffered(src, sum, buffer, s);
}
namespace cv { namespace gpu { namespace cudev
{
namespace imgproc
{
void shfl_integral_gpu(const PtrStepSzb& img, PtrStepSz<unsigned int> integral, cudaStream_t stream);
}
}}}
void cv::gpu::integralBuffered(const GpuMat& src, GpuMat& sum, GpuMat& buffer, Stream& s)
{
CV_Assert(src.type() == CV_8UC1);
cudaStream_t stream = StreamAccessor::getStream(s);
cv::Size whole;
cv::Point offset;
src.locateROI(whole, offset);
if (deviceSupports(WARP_SHUFFLE_FUNCTIONS) && src.cols <= 2048
&& offset.x % 16 == 0 && ((src.cols + 63) / 64) * 64 <= (static_cast<int>(src.step) - offset.x))
{
ensureSizeIsEnough(((src.rows + 7) / 8) * 8, ((src.cols + 63) / 64) * 64, CV_32SC1, buffer);
cv::gpu::cudev::imgproc::shfl_integral_gpu(src, buffer, stream);
sum.create(src.rows + 1, src.cols + 1, CV_32SC1);
if (s)
s.enqueueMemSet(sum, Scalar::all(0));
else
sum.setTo(Scalar::all(0));
GpuMat inner = sum(Rect(1, 1, src.cols, src.rows));
GpuMat res = buffer(Rect(0, 0, src.cols, src.rows));
if (s)
s.enqueueCopy(res, inner);
else
res.copyTo(inner);
}
else
{
#ifndef HAVE_OPENCV_GPUNVIDIA
throw_no_cuda();
#else
sum.create(src.rows + 1, src.cols + 1, CV_32SC1);
NcvSize32u roiSize;
roiSize.width = src.cols;
roiSize.height = src.rows;
cudaDeviceProp prop;
cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) );
Ncv32u bufSize;
ncvSafeCall( nppiStIntegralGetSize_8u32u(roiSize, &bufSize, prop) );
ensureSizeIsEnough(1, bufSize, CV_8UC1, buffer);
NppStStreamHandler h(stream);
ncvSafeCall( nppiStIntegral_8u32u_C1R(const_cast<Ncv8u*>(src.ptr<Ncv8u>()), static_cast<int>(src.step),
sum.ptr<Ncv32u>(), static_cast<int>(sum.step), roiSize, buffer.ptr<Ncv8u>(), bufSize, prop) );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
#endif
}
}
//////////////////////////////////////////////////////////////////////////////
// sqrIntegral
void cv::gpu::sqrIntegral(const GpuMat& src, GpuMat& sqsum, Stream& s)
{
#ifndef HAVE_OPENCV_GPUNVIDIA
(void) src;
(void) sqsum;
(void) s;
throw_no_cuda();
#else
CV_Assert(src.type() == CV_8U);
NcvSize32u roiSize;
roiSize.width = src.cols;
roiSize.height = src.rows;
cudaDeviceProp prop;
cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) );
Ncv32u bufSize;
ncvSafeCall(nppiStSqrIntegralGetSize_8u64u(roiSize, &bufSize, prop));
GpuMat buf(1, bufSize, CV_8U);
cudaStream_t stream = StreamAccessor::getStream(s);
NppStStreamHandler h(stream);
sqsum.create(src.rows + 1, src.cols + 1, CV_64F);
ncvSafeCall(nppiStSqrIntegral_8u64u_C1R(const_cast<Ncv8u*>(src.ptr<Ncv8u>(0)), static_cast<int>(src.step),
sqsum.ptr<Ncv64u>(0), static_cast<int>(sqsum.step), roiSize, buf.ptr<Ncv8u>(0), bufSize, prop));
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
#endif
}
#endif /* !defined (HAVE_CUDA) */

0
modules/gpu/src/cuda/integral_image.cu → modules/gpuarithm/src/cuda/integral.cu
View File

7
modules/gpuarithm/src/precomp.hpp
View File

@ -51,6 +51,13 @@
#include "opencv2/core/gpu_private.hpp"
#include "opencv2/opencv_modules.hpp"
#ifdef HAVE_OPENCV_GPUNVIDIA
# include "opencv2/gpunvidia.hpp"
# include "opencv2/gpunvidia/private.hpp"
#endif
#ifdef HAVE_CUBLAS
#include <cublas.h>
#endif

41
modules/gpuarithm/test/test_core.cpp
View File

@ -3607,11 +3607,11 @@ INSTANTIATE_TEST_CASE_P(GPU_Core, Normalize, testing::Combine(
testing::Values(NormCode(cv::NORM_L1), NormCode(cv::NORM_L2), NormCode(cv::NORM_INF), NormCode(cv::NORM_MINMAX)),
WHOLE_SUBMAT));
#ifdef HAVE_OPENCV_IMGPROC
//////////////////////////////////////////////////////////////////////////////
// CopyMakeBorder
#ifdef HAVE_OPENCV_IMGPROC
namespace
{
IMPLEMENT_PARAM_CLASS(Border, int)
@ -3669,6 +3669,43 @@ INSTANTIATE_TEST_CASE_P(GPU_ImgProc, CopyMakeBorder, testing::Combine(
ALL_BORDER_TYPES,
WHOLE_SUBMAT));
///////////////////////////////////////////////////////////////////////////////////////////////////////
// Integral
PARAM_TEST_CASE(Integral, cv::gpu::DeviceInfo, cv::Size, UseRoi)
{
cv::gpu::DeviceInfo devInfo;
cv::Size size;
bool useRoi;
virtual void SetUp()
{
devInfo = GET_PARAM(0);
size = GET_PARAM(1);
useRoi = GET_PARAM(2);
cv::gpu::setDevice(devInfo.deviceID());
}
};
GPU_TEST_P(Integral, Accuracy)
{
cv::Mat src = randomMat(size, CV_8UC1);
cv::gpu::GpuMat dst = createMat(cv::Size(src.cols + 1, src.rows + 1), CV_32SC1, useRoi);
cv::gpu::integral(loadMat(src, useRoi), dst);
cv::Mat dst_gold;
cv::integral(src, dst_gold, CV_32S);
EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
}
INSTANTIATE_TEST_CASE_P(GPU_ImgProc, Integral, testing::Combine(
ALL_DEVICES,
DIFFERENT_SIZES,
WHOLE_SUBMAT));
#endif
#endif // HAVE_CUDA