Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

removed BEGIN_OPENCV_DEVICE_NAMESPACE macros

Browse Source
pull/13383/head
Vladislav Vinogradov 13 years ago
parent d926541311
commit 0f53f2993e
73 changed files with 18038 additions and 18270 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 19
      modules/gpu/src/arithm.cpp
  2. 19
      modules/gpu/src/bilateral_filter.cpp
  3. 19
      modules/gpu/src/blend.cpp
  4. 153
      modules/gpu/src/brute_force_matcher.cpp
  5. 39
      modules/gpu/src/calib3d.cpp
  6. 275
      modules/gpu/src/color.cpp
  7. 1816
      modules/gpu/src/cuda/bf_knnmatch.cu
  8. 1190
      modules/gpu/src/cuda/bf_match.cu
  9. 712
      modules/gpu/src/cuda/bf_radius_match.cu
  10. 308
      modules/gpu/src/cuda/bilateral_filter.cu
  11. 124
      modules/gpu/src/cuda/blend.cu
  12. 243
      modules/gpu/src/cuda/calib3d.cu
  13. 690
      modules/gpu/src/cuda/canny.cu
  14. 621
      modules/gpu/src/cuda/color.cu
  15. 352
      modules/gpu/src/cuda/column_filter.cu
  16. 162
      modules/gpu/src/cuda/copy_make_border.cu
  17. 3642
      modules/gpu/src/cuda/element_operations.cu
  18. 284
      modules/gpu/src/cuda/hist.cu
  19. 1274
      modules/gpu/src/cuda/hog.cu
  20. 1568
      modules/gpu/src/cuda/imgproc.cu
  21. 111
      modules/gpu/src/cuda/internal_shared.hpp
  22. 1536
      modules/gpu/src/cuda/match_template.cu
  23. 294
      modules/gpu/src/cuda/mathfunc.cu
  24. 495
      modules/gpu/src/cuda/matrix_operations.cu
  25. 3408
      modules/gpu/src/cuda/matrix_reductions.cu
  26. 260
      modules/gpu/src/cuda/pyr_down.cu
  27. 200
      modules/gpu/src/cuda/pyr_up.cu
  28. 398
      modules/gpu/src/cuda/remap.cu
  29. 428
      modules/gpu/src/cuda/resize.cu
  30. 388
      modules/gpu/src/cuda/row_filter.cu
  31. 63
      modules/gpu/src/cuda/safe_call.hpp
  32. 922
      modules/gpu/src/cuda/split_merge.cu
  33. 800
      modules/gpu/src/cuda/stereobm.cu
  34. 782
      modules/gpu/src/cuda/stereobp.cu
  35. 1334
      modules/gpu/src/cuda/stereocsbp.cu
  36. 1570
      modules/gpu/src/cuda/surf.cu
  37. 21
      modules/gpu/src/cudastream.cpp
  38. 250
      modules/gpu/src/element_operations.cpp
  39. 29
      modules/gpu/src/filtering.cpp
  40. 71
      modules/gpu/src/hog.cpp
  41. 341
      modules/gpu/src/imgproc.cpp
  42. 29
      modules/gpu/src/initialization.cpp
  43. 163
      modules/gpu/src/match_template.cpp
  44. 162
      modules/gpu/src/matrix_reductions.cpp
  45. 1141
      modules/gpu/src/opencv2/gpu/device/border_interpolate.hpp
  46. 345
      modules/gpu/src/opencv2/gpu/device/color.hpp
  47. 46
      modules/gpu/src/opencv2/gpu/device/datamov_utils.hpp
  48. 1787
      modules/gpu/src/opencv2/gpu/device/detail/color_detail.hpp
  49. 577
      modules/gpu/src/opencv2/gpu/device/detail/transform_detail.hpp
  50. 257
      modules/gpu/src/opencv2/gpu/device/detail/type_traits_detail.hpp
  51. 1255
      modules/gpu/src/opencv2/gpu/device/detail/utility_detail.hpp
  52. 103
      modules/gpu/src/opencv2/gpu/device/detail/vec_distance_detail.hpp
  53. 33
      modules/gpu/src/opencv2/gpu/device/emulation.hpp
  54. 131
      modules/gpu/src/opencv2/gpu/device/filters.hpp
  55. 39
      modules/gpu/src/opencv2/gpu/device/funcattrib.hpp
  56. 742
      modules/gpu/src/opencv2/gpu/device/functional.hpp
  57. 370
      modules/gpu/src/opencv2/gpu/device/limits.hpp
  58. 319
      modules/gpu/src/opencv2/gpu/device/saturate_cast.hpp
  59. 29
      modules/gpu/src/opencv2/gpu/device/static_check.hpp
  60. 45
      modules/gpu/src/opencv2/gpu/device/transform.hpp
  61. 55
      modules/gpu/src/opencv2/gpu/device/type_traits.hpp
  62. 255
      modules/gpu/src/opencv2/gpu/device/utility.hpp
  63. 263
      modules/gpu/src/opencv2/gpu/device/vec_distance.hpp
  64. 295
      modules/gpu/src/opencv2/gpu/device/vec_math.hpp
  65. 319
      modules/gpu/src/opencv2/gpu/device/vec_traits.hpp
  66. 111
      modules/gpu/src/opencv2/gpu/device/warp.hpp
  67. 39
      modules/gpu/src/opencv2/gpu/device/warp_reduce.hpp
  68. 21
      modules/gpu/src/split_merge.cpp
  69. 19
      modules/gpu/src/stereobm.cpp
  70. 41
      modules/gpu/src/stereobp.cpp
  71. 63
      modules/gpu/src/stereocsbp.cpp
  72. 41
      modules/gpu/src/surf.cpp
  73. 2
      modules/gpu/test/test_video.cpp

19
modules/gpu/src/arithm.cpp
Unescape View File

@ -425,21 +425,20 @@ void cv::gpu::magnitudeSqr(const GpuMat& src, GpuMat& dst, Stream& stream)
////////////////////////////////////////////////////////////////////////
// Polar <-> Cart
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace mathfunc
namespace cv { namespace gpu { namespace device
{
void cartToPolar_gpu(DevMem2Df x, DevMem2Df y, DevMem2Df mag, bool magSqr, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream);
void polarToCart_gpu(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace mathfunc
{
void cartToPolar_gpu(DevMem2Df x, DevMem2Df y, DevMem2Df mag, bool magSqr, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream);
void polarToCart_gpu(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream);
}
}}}
namespace
{
inline void cartToPolar_caller(const GpuMat& x, const GpuMat& y, GpuMat* mag, bool magSqr, GpuMat* angle, bool angleInDegrees, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ mathfunc;
using namespace ::cv::gpu::device::mathfunc;
CV_DbgAssert(x.size() == y.size() && x.type() == y.type());
CV_Assert(x.depth() == CV_32F);
@ -459,7 +458,7 @@ namespace
inline void polarToCart_caller(const GpuMat& mag, const GpuMat& angle, GpuMat& x, GpuMat& y, bool angleInDegrees, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ mathfunc;
using namespace ::cv::gpu::device::mathfunc;
CV_DbgAssert((mag.empty() || mag.size() == angle.size()) && mag.type() == angle.type());
CV_Assert(mag.depth() == CV_32F);

19
modules/gpu/src/bilateral_filter.cpp
Unescape View File

@ -55,19 +55,18 @@ void cv::gpu::DisparityBilateralFilter::operator()(const GpuMat&, const GpuMat&,
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace bilateral_filter
namespace cv { namespace gpu { namespace device
{
void load_constants(float* table_color, DevMem2Df table_space, int ndisp, int radius, short edge_disc, short max_disc);
void bilateral_filter_gpu(DevMem2Db disp, DevMem2Db img, int channels, int iters, cudaStream_t stream);
void bilateral_filter_gpu(DevMem2D_<short> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream);
}
namespace bilateral_filter
{
void load_constants(float* table_color, DevMem2Df table_space, int ndisp, int radius, short edge_disc, short max_disc);
END_OPENCV_DEVICE_NAMESPACE
void bilateral_filter_gpu(DevMem2Db disp, DevMem2Db img, int channels, int iters, cudaStream_t stream);
void bilateral_filter_gpu(DevMem2D_<short> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream);
}
}}}
using namespace OPENCV_DEVICE_NAMESPACE_ bilateral_filter;
using namespace ::cv::gpu::device::bilateral_filter;
namespace
{

19
modules/gpu/src/blend.cpp
Unescape View File

@ -52,19 +52,18 @@ void cv::gpu::blendLinear(const GpuMat&, const GpuMat&, const GpuMat&, const Gpu
#else
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace blend
namespace cv { namespace gpu { namespace device
{
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream);
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream);
}
namespace blend
{
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream);
}
}}}
using namespace OPENCV_DEVICE_NAMESPACE_ blend;
using namespace ::cv::gpu::device::blend;
void cv::gpu::blendLinear(const GpuMat& img1, const GpuMat& img2, const GpuMat& weights1, const GpuMat& weights2,
GpuMat& result, Stream& stream)

153
modules/gpu/src/brute_force_matcher.cpp
Unescape View File

@ -82,80 +82,79 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatch(const GpuMat&, vector< vec
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace bf_match
{
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
}
namespace bf_knnmatch
namespace cv { namespace gpu { namespace device
{
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, int k, const DevMem2Db& mask,
const DevMem2Db& trainIdx, const DevMem2Db& distance, const DevMem2Df& allDist,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, int k, const DevMem2Db& mask,
const DevMem2Db& trainIdx, const DevMem2Db& distance, const DevMem2Df& allDist,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, int k, const DevMem2Db& mask,
const DevMem2Db& trainIdx, const DevMem2Db& distance, const DevMem2Df& allDist,
int cc, cudaStream_t stream);
template <typename T> void match2L1_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Db& trainIdx, const DevMem2Db& imgIdx, const DevMem2Db& distance,
int cc, cudaStream_t stream);
template <typename T> void match2L2_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Db& trainIdx, const DevMem2Db& imgIdx, const DevMem2Db& distance,
int cc, cudaStream_t stream);
template <typename T> void match2Hamming_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Db& trainIdx, const DevMem2Db& imgIdx, const DevMem2Db& distance,
int cc, cudaStream_t stream);
}
namespace bf_match
{
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance,
int cc, cudaStream_t stream);
}
namespace bf_radius_match
{
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
}
namespace bf_knnmatch
{
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, int k, const DevMem2Db& mask,
const DevMem2Db& trainIdx, const DevMem2Db& distance, const DevMem2Df& allDist,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, int k, const DevMem2Db& mask,
const DevMem2Db& trainIdx, const DevMem2Db& distance, const DevMem2Df& allDist,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, int k, const DevMem2Db& mask,
const DevMem2Db& trainIdx, const DevMem2Db& distance, const DevMem2Df& allDist,
int cc, cudaStream_t stream);
template <typename T> void match2L1_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Db& trainIdx, const DevMem2Db& imgIdx, const DevMem2Db& distance,
int cc, cudaStream_t stream);
template <typename T> void match2L2_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Db& trainIdx, const DevMem2Db& imgIdx, const DevMem2Db& distance,
int cc, cudaStream_t stream);
template <typename T> void match2Hamming_gpu(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Db& trainIdx, const DevMem2Db& imgIdx, const DevMem2Db& distance,
int cc, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace bf_radius_match
{
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream);
}
}}}
////////////////////////////////////////////////////////////////////
// Train collection
@ -199,7 +198,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchSingle(const GpuMat& query, const
if (query.empty() || train.empty())
return;
using namespace OPENCV_DEVICE_NAMESPACE_ bf_match;
using namespace ::cv::gpu::device::bf_match;
typedef void (*caller_t)(const DevMem2Db& query, const DevMem2Db& train, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance,
@ -341,7 +340,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchCollection(const GpuMat& query, c
if (query.empty() || trainCollection.empty())
return;
using namespace OPENCV_DEVICE_NAMESPACE_ bf_match;
using namespace ::cv::gpu::device::bf_match;
typedef void (*caller_t)(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance,
@ -452,7 +451,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatchSingle(const GpuMat& query, co
if (query.empty() || train.empty())
return;
using namespace OPENCV_DEVICE_NAMESPACE_ bf_knnmatch;
using namespace ::cv::gpu::device::bf_knnmatch;
typedef void (*caller_t)(const DevMem2Db& query, const DevMem2Db& train, int k, const DevMem2Db& mask,
const DevMem2Db& trainIdx, const DevMem2Db& distance, const DevMem2Df& allDist,
@ -581,7 +580,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatch2Collection(const GpuMat& quer
if (query.empty() || trainCollection.empty())
return;
using namespace OPENCV_DEVICE_NAMESPACE_ bf_knnmatch;
using namespace ::cv::gpu::device::bf_knnmatch;
typedef void (*caller_t)(const DevMem2Db& query, const DevMem2Db& trains, const DevMem2D_<PtrStepb>& masks,
const DevMem2Db& trainIdx, const DevMem2Db& imgIdx, const DevMem2Db& distance,
@ -762,7 +761,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchSingle(const GpuMat& query,
if (query.empty() || train.empty())
return;
using namespace OPENCV_DEVICE_NAMESPACE_ bf_radius_match;
using namespace ::cv::gpu::device::bf_radius_match;
typedef void (*caller_t)(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
@ -893,7 +892,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchCollection(const GpuMat& qu
if (query.empty() || empty())
return;
using namespace OPENCV_DEVICE_NAMESPACE_ bf_radius_match;
using namespace ::cv::gpu::device::bf_radius_match;
typedef void (*caller_t)(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,

39
modules/gpu/src/calib3d.cpp
Unescape View File

@ -56,31 +56,30 @@ void cv::gpu::solvePnPRansac(const Mat&, const Mat&, const Mat&, const Mat&, Mat
#else
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace transform_points
{
void call(const DevMem2D_<float3> src, const float* rot, const float* transl, DevMem2D_<float3> dst, cudaStream_t stream);
}
namespace project_points
namespace cv { namespace gpu { namespace device
{
void call(const DevMem2D_<float3> src, const float* rot, const float* transl, const float* proj, DevMem2D_<float2> dst, cudaStream_t stream);
}
namespace transform_points
{
void call(const DevMem2D_<float3> src, const float* rot, const float* transl, DevMem2D_<float3> dst, cudaStream_t stream);
}
namespace solve_pnp_ransac
{
int maxNumIters();
namespace project_points
{
void call(const DevMem2D_<float3> src, const float* rot, const float* transl, const float* proj, DevMem2D_<float2> dst, cudaStream_t stream);
}
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores);
}
namespace solve_pnp_ransac
{
int maxNumIters();
END_OPENCV_DEVICE_NAMESPACE
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores);
}
}}}
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
namespace
{

275
modules/gpu/src/color.cpp
Unescape View File

@ -51,8 +51,8 @@ void cv::gpu::cvtColor(const GpuMat&, GpuMat&, int, int, Stream&) { throw_nogpu(
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
#define OPENCV_GPU_DECLARE_CVTCOLOR_ONE(name) \
void name(const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
@ -67,142 +67,141 @@ BEGIN_OPENCV_DEVICE_NAMESPACE
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(name ## _full_8u) \
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(name ## _full_32f)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_bgra)
#undef OPENCV_GPU_DECLARE_CVTCOLOR_ONE
#undef OPENCV_GPU_DECLARE_CVTCOLOR_ALL
#undef OPENCV_GPU_DECLARE_CVTCOLOR_8U32F
END_OPENCV_DEVICE_NAMESPACE
using namespace OPENCV_DEVICE_NAMESPACE;
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_DECLARE_CVTCOLOR_8U32F(hls4_to_bgra)
#undef OPENCV_GPU_DECLARE_CVTCOLOR_ONE
#undef OPENCV_GPU_DECLARE_CVTCOLOR_ALL
#undef OPENCV_GPU_DECLARE_CVTCOLOR_8U32F
}}}
using namespace ::cv::gpu::device;
namespace
{

1816
modules/gpu/src/cuda/bf_knnmatch.cu
View File

File diff suppressed because it is too large Load Diff

1190
modules/gpu/src/cuda/bf_match.cu
View File

File diff suppressed because it is too large Load Diff

712
modules/gpu/src/cuda/bf_radius_match.cu
Unescape View File

@ -45,423 +45,421 @@
#include "opencv2/gpu/device/vec_distance.hpp"
#include "opencv2/gpu/device/datamov_utils.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace bf_radius_match {
///////////////////////////////////////////////////////////////////////////////
// Match Unrolled
template <int BLOCK_SIZE, int MAX_DESC_LEN, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void matchUnrolled(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
namespace cv { namespace gpu { namespace device
{
#if __CUDA_ARCH__ >= 110
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
namespace bf_radius_match
{
///////////////////////////////////////////////////////////////////////////////
// Match Unrolled
Dist dist;
template <int BLOCK_SIZE, int MAX_DESC_LEN, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void matchUnrolled(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if __CUDA_ARCH__ >= 110
#pragma unroll
for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
extern __shared__ int smem[];
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
if (loadX < query.cols)
{
T val;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
Dist dist;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
#pragma unroll
for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
__syncthreads();
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
if (loadX < query.cols)
{
T val;
__syncthreads();
}
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
float distVal = (typename Dist::result_type)dist;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
__syncthreads();
#endif
}
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T, typename Mask>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
__syncthreads();
}
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
float distVal = (typename Dist::result_type)dist;
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
#endif
}
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T, typename Mask>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (masks != 0 && masks[i].data)
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////////
// Match
///////////////////////////////////////////////////////////////////////////////
// Match
template <int BLOCK_SIZE, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void match(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if __CUDA_ARCH__ >= 110
template <int BLOCK_SIZE, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void match(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if __CUDA_ARCH__ >= 110
extern __shared__ int smem[];
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
Dist dist;
for (int i = 0, endi = (query.cols + BLOCK_SIZE - 1) / BLOCK_SIZE; i < endi; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
for (int i = 0, endi = (query.cols + BLOCK_SIZE - 1) / BLOCK_SIZE; i < endi; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
__syncthreads();
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
#endif
}
}
#endif
}
template <int BLOCK_SIZE, typename Dist, typename T, typename Mask>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
template <int BLOCK_SIZE, typename Dist, typename T, typename Mask>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
match<BLOCK_SIZE, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
match<BLOCK_SIZE, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, typename Dist, typename T>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, typename Dist, typename T>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
///////////////////////////////////////////////////////////////////////////////
// Match dispatcher
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
}
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
template <typename Dist, typename T>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
}
///////////////////////////////////////////////////////////////////////////////
// Radius Match caller
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
template void matchL1_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
if (masks != 0 && masks[i].data)
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
if (mask.data)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
else
//template void matchL2_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
if (mask.data)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////////
// Match dispatcher
template void matchHamming_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
}
template <typename Dist, typename T>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
}
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
///////////////////////////////////////////////////////////////////////////////
// Radius Match caller
template void matchL1_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchL1_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
//template void matchL2_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchHamming_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchL1_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
//template void matchL2_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchHamming_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
} // namespace bf_radius_match
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
END_OPENCV_DEVICE_NAMESPACE
template void matchHamming_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
} // namespace bf_radius_match
}}} // namespace cv { namespace gpu { namespace device

308
modules/gpu/src/cuda/bilateral_filter.cu
Unescape View File

@ -43,186 +43,184 @@
#include "internal_shared.hpp"
#include "opencv2/gpu/device/limits.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace bilateral_filter {
__constant__ float* ctable_color;
__constant__ float* ctable_space;
__constant__ size_t ctable_space_step;
__constant__ int cndisp;
__constant__ int cradius;
__constant__ short cedge_disc;
__constant__ short cmax_disc;
void load_constants(float* table_color, DevMem2Df table_space, int ndisp, int radius, short edge_disc, short max_disc)
{
cudaSafeCall( cudaMemcpyToSymbol(ctable_color, &table_color, sizeof(table_color)) );
cudaSafeCall( cudaMemcpyToSymbol(ctable_space, &table_space.data, sizeof(table_space.data)) );
size_t table_space_step = table_space.step / sizeof(float);
cudaSafeCall( cudaMemcpyToSymbol(ctable_space_step, &table_space_step, sizeof(size_t)) );
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int)) );
cudaSafeCall( cudaMemcpyToSymbol(cradius, &radius, sizeof(int)) );
cudaSafeCall( cudaMemcpyToSymbol(cedge_disc, &edge_disc, sizeof(short)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc, &max_disc, sizeof(short)) );
}
template <int channels>
struct DistRgbMax
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
{
uchar x = ::abs(a[0] - b[0]);
uchar y = ::abs(a[1] - b[1]);
uchar z = ::abs(a[2] - b[2]);
return (::max(::max(x, y), z));
}
};
template <>
struct DistRgbMax<1>
namespace cv { namespace gpu { namespace device
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
namespace bilateral_filter
{
return ::abs(a[0] - b[0]);
}
};
__constant__ float* ctable_color;
__constant__ float* ctable_space;
__constant__ size_t ctable_space_step;
template <int channels, typename T>
__global__ void bilateral_filter(int t, T* disp, size_t disp_step, const uchar* img, size_t img_step, int h, int w)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
T dp[5];
__constant__ int cndisp;
__constant__ int cradius;
if (y > 0 && y < h - 1 && x > 0 && x < w - 1)
{
dp[0] = *(disp + (y ) * disp_step + x + 0);
dp[1] = *(disp + (y-1) * disp_step + x + 0);
dp[2] = *(disp + (y ) * disp_step + x - 1);
dp[3] = *(disp + (y+1) * disp_step + x + 0);
dp[4] = *(disp + (y ) * disp_step + x + 1);
__constant__ short cedge_disc;
__constant__ short cmax_disc;
if(::abs(dp[1] - dp[0]) >= cedge_disc || ::abs(dp[2] - dp[0]) >= cedge_disc || ::abs(dp[3] - dp[0]) >= cedge_disc || ::abs(dp[4] - dp[0]) >= cedge_disc)
void load_constants(float* table_color, DevMem2Df table_space, int ndisp, int radius, short edge_disc, short max_disc)
{
const int ymin = ::max(0, y - cradius);
const int xmin = ::max(0, x - cradius);
const int ymax = ::min(h - 1, y + cradius);
const int xmax = ::min(w - 1, x + cradius);
cudaSafeCall( cudaMemcpyToSymbol(ctable_color, &table_color, sizeof(table_color)) );
cudaSafeCall( cudaMemcpyToSymbol(ctable_space, &table_space.data, sizeof(table_space.data)) );
size_t table_space_step = table_space.step / sizeof(float);
cudaSafeCall( cudaMemcpyToSymbol(ctable_space_step, &table_space_step, sizeof(size_t)) );
float cost[] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int)) );
cudaSafeCall( cudaMemcpyToSymbol(cradius, &radius, sizeof(int)) );
const uchar* ic = img + y * img_step + channels * x;
cudaSafeCall( cudaMemcpyToSymbol(cedge_disc, &edge_disc, sizeof(short)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc, &max_disc, sizeof(short)) );
}
for(int yi = ymin; yi <= ymax; yi++)
template <int channels>
struct DistRgbMax
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
{
const T* disp_y = disp + yi * disp_step;
uchar x = ::abs(a[0] - b[0]);
uchar y = ::abs(a[1] - b[1]);
uchar z = ::abs(a[2] - b[2]);
return (::max(::max(x, y), z));
}
};
for(int xi = xmin; xi <= xmax; xi++)
{
const uchar* in = img + yi * img_step + channels * xi;
template <>
struct DistRgbMax<1>
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
{
return ::abs(a[0] - b[0]);
}
};
uchar dist_rgb = DistRgbMax<channels>::calc(in, ic);
template <int channels, typename T>
__global__ void bilateral_filter(int t, T* disp, size_t disp_step, const uchar* img, size_t img_step, int h, int w)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
const float weight = ctable_color[dist_rgb] * (ctable_space + ::abs(y-yi)* ctable_space_step)[::abs(x-xi)];
T dp[5];
const T disp_reg = disp_y[xi];
if (y > 0 && y < h - 1 && x > 0 && x < w - 1)
{
dp[0] = *(disp + (y ) * disp_step + x + 0);
dp[1] = *(disp + (y-1) * disp_step + x + 0);
dp[2] = *(disp + (y ) * disp_step + x - 1);
dp[3] = *(disp + (y+1) * disp_step + x + 0);
dp[4] = *(disp + (y ) * disp_step + x + 1);
cost[0] += ::min(cmax_disc, ::abs(disp_reg - dp[0])) * weight;
cost[1] += ::min(cmax_disc, ::abs(disp_reg - dp[1])) * weight;
cost[2] += ::min(cmax_disc, ::abs(disp_reg - dp[2])) * weight;
cost[3] += ::min(cmax_disc, ::abs(disp_reg - dp[3])) * weight;
cost[4] += ::min(cmax_disc, ::abs(disp_reg - dp[4])) * weight;
if(::abs(dp[1] - dp[0]) >= cedge_disc || ::abs(dp[2] - dp[0]) >= cedge_disc || ::abs(dp[3] - dp[0]) >= cedge_disc || ::abs(dp[4] - dp[0]) >= cedge_disc)
{
const int ymin = ::max(0, y - cradius);
const int xmin = ::max(0, x - cradius);
const int ymax = ::min(h - 1, y + cradius);
const int xmax = ::min(w - 1, x + cradius);
float cost[] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
const uchar* ic = img + y * img_step + channels * x;
for(int yi = ymin; yi <= ymax; yi++)
{
const T* disp_y = disp + yi * disp_step;
for(int xi = xmin; xi <= xmax; xi++)
{
const uchar* in = img + yi * img_step + channels * xi;
uchar dist_rgb = DistRgbMax<channels>::calc(in, ic);
const float weight = ctable_color[dist_rgb] * (ctable_space + ::abs(y-yi)* ctable_space_step)[::abs(x-xi)];
const T disp_reg = disp_y[xi];
cost[0] += ::min(cmax_disc, ::abs(disp_reg - dp[0])) * weight;
cost[1] += ::min(cmax_disc, ::abs(disp_reg - dp[1])) * weight;
cost[2] += ::min(cmax_disc, ::abs(disp_reg - dp[2])) * weight;
cost[3] += ::min(cmax_disc, ::abs(disp_reg - dp[3])) * weight;
cost[4] += ::min(cmax_disc, ::abs(disp_reg - dp[4])) * weight;
}
}
float minimum = numeric_limits<float>::max();
int id = 0;
if (cost[0] < minimum)
{
minimum = cost[0];
id = 0;
}
if (cost[1] < minimum)
{
minimum = cost[1];
id = 1;
}
if (cost[2] < minimum)
{
minimum = cost[2];
id = 2;
}
if (cost[3] < minimum)
{
minimum = cost[3];
id = 3;
}
if (cost[4] < minimum)
{
minimum = cost[4];
id = 4;
}
*(disp + y * disp_step + x) = dp[id];
}
}
}
float minimum = numeric_limits<float>::max();
int id = 0;
template <typename T>
void bilateral_filter_caller(DevMem2D_<T> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(disp.cols, threads.x << 1);
grid.y = divUp(disp.rows, threads.y);
if (cost[0] < minimum)
switch (channels)
{
minimum = cost[0];
id = 0;
}
if (cost[1] < minimum)
{
minimum = cost[1];
id = 1;
}
if (cost[2] < minimum)
{
minimum = cost[2];
id = 2;
}
if (cost[3] < minimum)
{
minimum = cost[3];
id = 3;
}
if (cost[4] < minimum)
{
minimum = cost[4];
id = 4;
case 1:
for (int i = 0; i < iters; ++i)
{
bilateral_filter<1><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter<1><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
}
break;
case 3:
for (int i = 0; i < iters; ++i)
{
bilateral_filter<3><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter<3><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
}
break;
default:
cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
*(disp + y * disp_step + x) = dp[id];
if (stream != 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
}
template <typename T>
void bilateral_filter_caller(DevMem2D_<T> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(disp.cols, threads.x << 1);
grid.y = divUp(disp.rows, threads.y);
switch (channels)
{
case 1:
for (int i = 0; i < iters; ++i)
void bilateral_filter_gpu(DevMem2Db disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter<1><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter<1><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter_caller(disp, img, channels, iters, stream);
}
break;
case 3:
for (int i = 0; i < iters; ++i)
{
bilateral_filter<3><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter<3><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
void bilateral_filter_gpu(DevMem2D_<short> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter_caller(disp, img, channels, iters, stream);
}
break;
default:
cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
if (stream != 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void bilateral_filter_gpu(DevMem2Db disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter_caller(disp, img, channels, iters, stream);
}
void bilateral_filter_gpu(DevMem2D_<short> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter_caller(disp, img, channels, iters, stream);
}
} // namespace bilateral_filter
END_OPENCV_DEVICE_NAMESPACE
} // namespace bilateral_filter
}}} // namespace cv { namespace gpu { namespace device

124
modules/gpu/src/cuda/blend.cu
Unescape View File

@ -42,77 +42,75 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace blend {
template <typename T>
__global__ void blendLinearKernel(int rows, int cols, int cn, const PtrStep<T> img1, const PtrStep<T> img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStep<T> result)
namespace cv { namespace gpu { namespace device
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y < rows && x < cols)
namespace blend
{
int x_ = x / cn;
float w1 = weights1.ptr(y)[x_];
float w2 = weights2.ptr(y)[x_];
T p1 = img1.ptr(y)[x];
T p2 = img2.ptr(y)[x];
result.ptr(y)[x] = (p1 * w1 + p2 * w2) / (w1 + w2 + 1e-5f);
}
}
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols * cn, threads.x), divUp(rows, threads.y));
blendLinearKernel<<<grid, threads, 0, stream>>>(rows, cols * cn, cn, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
template <typename T>
__global__ void blendLinearKernel(int rows, int cols, int cn, const PtrStep<T> img1, const PtrStep<T> img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStep<T> result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
if (y < rows && x < cols)
{
int x_ = x / cn;
float w1 = weights1.ptr(y)[x_];
float w2 = weights2.ptr(y)[x_];
T p1 = img1.ptr(y)[x];
T p2 = img2.ptr(y)[x];
result.ptr(y)[x] = (p1 * w1 + p2 * w2) / (w1 + w2 + 1e-5f);
}
}
template void blendLinearCaller<uchar>(int, int, int, PtrStep<uchar>, PtrStep<uchar>, PtrStepf, PtrStepf, PtrStep<uchar>, cudaStream_t stream);
template void blendLinearCaller<float>(int, int, int, PtrStep<float>, PtrStep<float>, PtrStepf, PtrStepf, PtrStep<float>, cudaStream_t stream);
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols * cn, threads.x), divUp(rows, threads.y));
blendLinearKernel<<<grid, threads, 0, stream>>>(rows, cols * cn, cn, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
__global__ void blendLinearKernel8UC4(int rows, int cols, const PtrStepb img1, const PtrStepb img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStepb result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
template void blendLinearCaller<uchar>(int, int, int, PtrStep<uchar>, PtrStep<uchar>, PtrStepf, PtrStepf, PtrStep<uchar>, cudaStream_t stream);
template void blendLinearCaller<float>(int, int, int, PtrStep<float>, PtrStep<float>, PtrStepf, PtrStepf, PtrStep<float>, cudaStream_t stream);
if (y < rows && x < cols)
{
float w1 = weights1.ptr(y)[x];
float w2 = weights2.ptr(y)[x];
float sum_inv = 1.f / (w1 + w2 + 1e-5f);
w1 *= sum_inv;
w2 *= sum_inv;
uchar4 p1 = ((const uchar4*)img1.ptr(y))[x];
uchar4 p2 = ((const uchar4*)img2.ptr(y))[x];
((uchar4*)result.ptr(y))[x] = make_uchar4(p1.x * w1 + p2.x * w2, p1.y * w1 + p2.y * w2,
p1.z * w1 + p2.z * w2, p1.w * w1 + p2.w * w2);
}
}
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
blendLinearKernel8UC4<<<grid, threads, 0, stream>>>(rows, cols, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
__global__ void blendLinearKernel8UC4(int rows, int cols, const PtrStepb img1, const PtrStepb img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStepb result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
if (y < rows && x < cols)
{
float w1 = weights1.ptr(y)[x];
float w2 = weights2.ptr(y)[x];
float sum_inv = 1.f / (w1 + w2 + 1e-5f);
w1 *= sum_inv;
w2 *= sum_inv;
uchar4 p1 = ((const uchar4*)img1.ptr(y))[x];
uchar4 p2 = ((const uchar4*)img2.ptr(y))[x];
((uchar4*)result.ptr(y))[x] = make_uchar4(p1.x * w1 + p2.x * w2, p1.y * w1 + p2.y * w2,
p1.z * w1 + p2.z * w2, p1.w * w1 + p2.w * w2);
}
}
} // namespace blend
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
blendLinearKernel8UC4<<<grid, threads, 0, stream>>>(rows, cols, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
END_OPENCV_DEVICE_NAMESPACE
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
} // namespace blend
}}} // namespace cv { namespace gpu { namespace device

243
modules/gpu/src/cuda/calib3d.cu
Unescape View File

@ -44,149 +44,148 @@
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/functional.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#define SOLVE_PNP_RANSAC_MAX_NUM_ITERS 200
namespace transform_points
namespace cv { namespace gpu { namespace device
{
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
#define SOLVE_PNP_RANSAC_MAX_NUM_ITERS 200
struct TransformOp : unary_function<float3, float3>
namespace transform_points
{
__device__ __forceinline__ float3 operator()(const float3& p) const
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
struct TransformOp : unary_function<float3, float3>
{
return make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
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__ float3 operator()(const float3& p) const
{
return make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
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);
}
};
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, DevMem2D_<float3> dst,
cudaStream_t stream)
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
::cv::gpu::device::transform(src, dst, TransformOp(), stream);
}
};
} // namespace transform_points
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, DevMem2D_<float3> dst,
cudaStream_t stream)
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
OPENCV_DEVICE_NAMESPACE_ transform(src, dst, TransformOp(), stream);
}
} // namespace transform_points
namespace project_points
{
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
__constant__ float3 cproj0;
__constant__ float3 cproj1;
struct ProjectOp : unary_function<float3, float3>
namespace project_points
{
__device__ __forceinline__ float2 operator()(const float3& p) const
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
__constant__ float3 cproj0;
__constant__ float3 cproj1;
struct ProjectOp : unary_function<float3, float3>
{
// Rotate and translate in 3D
float3 t = make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
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);
// Project on 2D plane
return make_float2(
(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__ float2 operator()(const float3& p) const
{
// Rotate and translate in 3D
float3 t = make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
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);
// Project on 2D plane
return make_float2(
(cproj0.x * t.x + cproj0.y * t.y) / t.z + cproj0.z,
(cproj1.x * t.x + cproj1.y * t.y) / t.z + cproj1.z);
}
};
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, const float* proj, DevMem2D_<float2> dst,
cudaStream_t stream)
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj0, proj, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj1, proj + 3, sizeof(float) * 3));
::cv::gpu::device::transform(src, dst, ProjectOp(), stream);
}
};
} // namespace project_points
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, const float* proj, DevMem2D_<float2> dst,
cudaStream_t stream)
namespace solve_pnp_ransac
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj0, proj, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj1, proj + 3, sizeof(float) * 3));
OPENCV_DEVICE_NAMESPACE_ transform(src, dst, ProjectOp(), stream);
}
} // namespace project_points
namespace solve_pnp_ransac
{
__constant__ float3 crot_matrices[SOLVE_PNP_RANSAC_MAX_NUM_ITERS * 3];
__constant__ float3 ctransl_vectors[SOLVE_PNP_RANSAC_MAX_NUM_ITERS];
__constant__ float3 crot_matrices[SOLVE_PNP_RANSAC_MAX_NUM_ITERS * 3];
__constant__ float3 ctransl_vectors[SOLVE_PNP_RANSAC_MAX_NUM_ITERS];
int maxNumIters()
{
return SOLVE_PNP_RANSAC_MAX_NUM_ITERS;
}
__device__ __forceinline__ float sqr(float x)
{
return x * x;
}
__global__ void computeHypothesisScoresKernel(
const int num_points, const float3* object, const float2* image,
const float dist_threshold, int* g_num_inliers)
{
const float3* const &rot_mat = crot_matrices + blockIdx.x * 3;
const float3 &transl_vec = ctransl_vectors[blockIdx.x];
int num_inliers = 0;
for (int i = threadIdx.x; i < num_points; i += blockDim.x)
int maxNumIters()
{
float3 p = object[i];
p = make_float3(
rot_mat[0].x * p.x + rot_mat[0].y * p.y + rot_mat[0].z * p.z + transl_vec.x,
rot_mat[1].x * p.x + rot_mat[1].y * p.y + rot_mat[1].z * p.z + transl_vec.y,
rot_mat[2].x * p.x + rot_mat[2].y * p.y + rot_mat[2].z * p.z + transl_vec.z);
p.x /= p.z;
p.y /= p.z;
float2 image_p = image[i];
if (sqr(p.x - image_p.x) + sqr(p.y - image_p.y) < dist_threshold)
++num_inliers;
return SOLVE_PNP_RANSAC_MAX_NUM_ITERS;
}
extern __shared__ float s_num_inliers[];
s_num_inliers[threadIdx.x] = num_inliers;
__syncthreads();
__device__ __forceinline__ float sqr(float x)
{
return x * x;
}
for (int step = blockDim.x / 2; step > 0; step >>= 1)
__global__ void computeHypothesisScoresKernel(
const int num_points, const float3* object, const float2* image,
const float dist_threshold, int* g_num_inliers)
{
if (threadIdx.x < step)
s_num_inliers[threadIdx.x] += s_num_inliers[threadIdx.x + step];
const float3* const &rot_mat = crot_matrices + blockIdx.x * 3;
const float3 &transl_vec = ctransl_vectors[blockIdx.x];
int num_inliers = 0;
for (int i = threadIdx.x; i < num_points; i += blockDim.x)
{
float3 p = object[i];
p = make_float3(
rot_mat[0].x * p.x + rot_mat[0].y * p.y + rot_mat[0].z * p.z + transl_vec.x,
rot_mat[1].x * p.x + rot_mat[1].y * p.y + rot_mat[1].z * p.z + transl_vec.y,
rot_mat[2].x * p.x + rot_mat[2].y * p.y + rot_mat[2].z * p.z + transl_vec.z);
p.x /= p.z;
p.y /= p.z;
float2 image_p = image[i];
if (sqr(p.x - image_p.x) + sqr(p.y - image_p.y) < dist_threshold)
++num_inliers;
}
extern __shared__ float s_num_inliers[];
s_num_inliers[threadIdx.x] = num_inliers;
__syncthreads();
}
if (threadIdx.x == 0)
g_num_inliers[blockIdx.x] = s_num_inliers[0];
}
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();
}
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores)
{
cudaSafeCall(cudaMemcpyToSymbol(crot_matrices, rot_matrices, num_hypotheses * 3 * sizeof(float3)));
cudaSafeCall(cudaMemcpyToSymbol(ctransl_vectors, transl_vectors, num_hypotheses * sizeof(float3)));
if (threadIdx.x == 0)
g_num_inliers[blockIdx.x] = s_num_inliers[0];
}
dim3 threads(256);
dim3 grid(num_hypotheses);
int smem_size = threads.x * sizeof(float);
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores)
{
cudaSafeCall(cudaMemcpyToSymbol(crot_matrices, rot_matrices, num_hypotheses * 3 * sizeof(float3)));
cudaSafeCall(cudaMemcpyToSymbol(ctransl_vectors, transl_vectors, num_hypotheses * sizeof(float3)));
computeHypothesisScoresKernel<<<grid, threads, smem_size>>>(
num_points, object, image, dist_threshold, hypothesis_scores);
cudaSafeCall( cudaGetLastError() );
dim3 threads(256);
dim3 grid(num_hypotheses);
int smem_size = threads.x * sizeof(float);
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace solvepnp_ransac
computeHypothesisScoresKernel<<<grid, threads, smem_size>>>(
num_points, object, image, dist_threshold, hypothesis_scores);
cudaSafeCall( cudaGetLastError() );
END_OPENCV_DEVICE_NAMESPACE
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace solvepnp_ransac
}}} // namespace cv { namespace gpu { namespace device

690
modules/gpu/src/cuda/canny.cu
Unescape View File

@ -44,450 +44,448 @@
#include <algorithm>
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace canny {
__global__ void calcSobelRowPass(const PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
namespace cv { namespace gpu { namespace device
{
__shared__ int smem[16][18];
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (i < rows)
namespace canny
{
smem[threadIdx.y][threadIdx.x + 1] = src.ptr(i)[j];
if (threadIdx.x == 0)
__global__ void calcSobelRowPass(const PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
{
smem[threadIdx.y][0] = src.ptr(i)[::max(j - 1, 0)];
smem[threadIdx.y][17] = src.ptr(i)[::min(j + 16, cols - 1)];
__shared__ int smem[16][18];
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = 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 (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];
}
}
}
__syncthreads();
if (j < cols)
void calcSobelRowPass_gpu(PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int 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];
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(cudaThreadSynchronize());
}
}
}
void calcSobelRowPass_gpu(PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
struct L1
{
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);
}
};
calcSobelRowPass<<<grid, block>>>(src, dx_buf, dy_buf, rows, cols);
cudaSafeCall( cudaGetLastError() );
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];
cudaSafeCall(cudaThreadSynchronize());
}
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
struct L1
{
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);
}
};
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];
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];
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();
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
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];
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];
dx.ptr(i)[j] = x;
dy.ptr(i)[j] = y;
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];
mag.ptr(i + 1)[j + 1] = Norm::calc(x, y);
}
}
}
__syncthreads();
if (i < rows)
void calcMagnitude_gpu(PtrStepi dx_buf, PtrStepi dy_buf, PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
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];
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
dx.ptr(i)[j] = x;
dy.ptr(i)[j] = y;
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
mag.ptr(i + 1)[j + 1] = Norm::calc(x, y);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
}
}
void calcMagnitude_gpu(PtrStepi dx_buf, PtrStepi dy_buf, PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
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;
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
if (i < rows && j < cols)
mag.ptr(i + 1)[j + 1] = Norm::calc(dx.ptr(i)[j], dy.ptr(i)[j]);
}
cudaSafeCall( cudaGetLastError() );
void calcMagnitude_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
cudaSafeCall(cudaThreadSynchronize());
}
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx, dy, mag, rows, cols);
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;
cudaSafeCall( cudaGetLastError() );
if (i < rows && j < cols)
mag.ptr(i + 1)[j + 1] = Norm::calc(dx.ptr(i)[j], dy.ptr(i)[j]);
}
cudaSafeCall(cudaThreadSynchronize());
}
void calcMagnitude_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
//////////////////////////////////////////////////////////////////////////////////////////
#define CANNY_SHIFT 15
#define TG22 (int)(0.4142135623730950488016887242097*(1<<CANNY_SHIFT) + 0.5)
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx, dy, mag, rows, cols);
__global__ void calcMap(const PtrStepi dx, const PtrStepi dy, const PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
__shared__ float smem[18][18];
cudaSafeCall( cudaGetLastError() );
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
cudaSafeCall(cudaThreadSynchronize());
}
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
//////////////////////////////////////////////////////////////////////////////////////////
#define CANNY_SHIFT 15
#define TG22 (int)(0.4142135623730950488016887242097*(1<<CANNY_SHIFT) + 0.5)
if (ly < 14)
smem[ly][lx] = mag.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
__global__ void calcMap(const PtrStepi dx, const PtrStepi dy, const PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
__shared__ float smem[18][18];
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];
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
__syncthreads();
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
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];
if (ly < 14)
smem[ly][lx] = mag.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
x = ::abs(x);
y = ::abs(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];
// 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;
__syncthreads();
if (m > low_thresh)
{
const int tg22x = x * TG22;
const int tg67x = tg22x + ((x + x) << CANNY_SHIFT);
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];
y <<= CANNY_SHIFT;
x = ::abs(x);
y = ::abs(y);
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;
}
}
// 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;
#undef CANNY_SHIFT
#undef TG22
if (m > low_thresh)
void calcMap_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
const int tg22x = x * TG22;
const int tg67x = tg22x + ((x + x) << CANNY_SHIFT);
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
y <<= CANNY_SHIFT;
calcMap<<<grid, block>>>(dx, dy, mag, map, rows, cols, low_thresh, high_thresh);
cudaSafeCall( cudaGetLastError() );
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);
}
cudaSafeCall(cudaThreadSynchronize());
}
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() );
__device__ unsigned int counter = 0;
cudaSafeCall(cudaThreadSynchronize());
}
//////////////////////////////////////////////////////////////////////////////////////////
__global__ void edgesHysteresisLocal(PtrStepi map, ushort2* st, int rows, int cols)
{
#if __CUDA_ARCH__ >= 120
__device__ unsigned int counter = 0;
__shared__ int smem[18][18];
__global__ void edgesHysteresisLocal(PtrStepi map, ushort2* st, int rows, int cols)
{
#if __CUDA_ARCH__ >= 120
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
__shared__ int smem[18][18];
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
if (ly < 14)
smem[ly][lx] = map.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
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];
if (ly < 14)
smem[ly][lx] = map.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
__syncthreads();
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];
if (i < rows && j < cols)
{
int n;
__syncthreads();
#pragma unroll
for (int k = 0; k < 16; ++k)
{
n = 0;
if (i < rows && j < cols)
{
int n;
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;
}
#pragma unroll
for (int k = 0; k < 16; ++k)
{
n = 0;
if (n > 0)
smem[threadIdx.y + 1][threadIdx.x + 1] = 2;
}
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;
}
const int e = smem[threadIdx.y + 1][threadIdx.x + 1];
if (n > 0)
smem[threadIdx.y + 1][threadIdx.x + 1] = 2;
}
map.ptr(i + 1)[j + 1] = e;
const int e = smem[threadIdx.y + 1][threadIdx.x + 1];
n = 0;
map.ptr(i + 1)[j + 1] = e;
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;
}
n = 0;
if (n > 0)
{
const unsigned int ind = atomicInc(&counter, (unsigned int)(-1));
st[ind] = make_ushort2(j + 1, i + 1);
}
}
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;
#endif
}
if (n > 0)
void edgesHysteresisLocal_gpu(PtrStepi map, ushort2* st1, int rows, int cols)
{
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)
{
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(cudaThreadSynchronize());
}
__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 __CUDA_ARCH__ >= 120
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
const int stack_size = 512;
__shared__ unsigned int s_counter;
__shared__ unsigned int s_ind;
__shared__ ushort2 s_st[stack_size];
edgesHysteresisLocal<<<grid, block>>>(map, st1, rows, cols);
cudaSafeCall( cudaGetLastError() );
if (threadIdx.x == 0)
s_counter = 0;
__syncthreads();
cudaSafeCall(cudaThreadSynchronize());
}
int ind = blockIdx.y * gridDim.x + blockIdx.x;
__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};
if (ind < count)
{
ushort2 pos = st1[ind];
if (pos.x > 0 && pos.x <= cols && pos.y > 0 && pos.y <= rows)
__global__ void edgesHysteresisGlobal(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols, int count)
{
if (threadIdx.x < 8)
{
pos.x += c_dx[threadIdx.x];
pos.y += c_dy[threadIdx.x];
#if __CUDA_ARCH__ >= 120
if (map.ptr(pos.y)[pos.x] == 1)
{
map.ptr(pos.y)[pos.x] = 2;
ind = atomicInc(&s_counter, (unsigned int)(-1));
const int stack_size = 512;
__shared__ unsigned int s_counter;
__shared__ unsigned int s_ind;
__shared__ ushort2 s_st[stack_size];
s_st[ind] = pos;
}
}
if (threadIdx.x == 0)
s_counter = 0;
__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);
int ind = blockIdx.y * gridDim.x + blockIdx.x;
pos.x = pos.y = 0;
if (ind < count)
{
ushort2 pos = st1[ind];
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)
if (threadIdx.x < 8)
{
map.ptr(pos.y)[pos.x] = 2;
pos.x += c_dx[threadIdx.x];
pos.y += c_dy[threadIdx.x];
ind = atomicInc(&s_counter, (unsigned int)(-1));
if (map.ptr(pos.y)[pos.x] == 1)
{
map.ptr(pos.y)[pos.x] = 2;
s_st[ind] = pos;
}
}
__syncthreads();
}
ind = atomicInc(&s_counter, (unsigned int)(-1));
if (s_counter > 0)
{
if (threadIdx.x == 0)
{
ind = atomicAdd(&counter, s_counter);
s_ind = ind - s_counter;
}
__syncthreads();
s_st[ind] = pos;
}
}
__syncthreads();
ind = s_ind;
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();
}
for (int i = threadIdx.x; i < s_counter; i += blockDim.x)
{
st2[ind + i] = s_st[i];
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];
}
}
}
}
}
}
#endif
}
void edgesHysteresisGlobal_gpu(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, counter) );
unsigned int count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
while (count > 0)
{
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(unsigned int)) );
#endif
}
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() );
void edgesHysteresisGlobal_gpu(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, counter) );
unsigned int count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
cudaSafeCall(cudaThreadSynchronize());
while (count > 0)
{
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(unsigned int)) );
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
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() );
std::swap(st1, st2);
}
}
cudaSafeCall(cudaThreadSynchronize());
__global__ void getEdges(PtrStepi map, PtrStepb dst, int rows, int cols)
{
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
if (i < rows && j < cols)
dst.ptr(i)[j] = (uchar)(-(map.ptr(i + 1)[j + 1] >> 1));
}
std::swap(st1, st2);
}
}
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);
__global__ void getEdges(PtrStepi map, PtrStepb dst, int rows, int cols)
{
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
getEdges<<<grid, block>>>(map, dst, rows, cols);
cudaSafeCall( cudaGetLastError() );
if (i < rows && j < cols)
dst.ptr(i)[j] = (uchar)(-(map.ptr(i + 1)[j + 1] >> 1));
}
cudaSafeCall(cudaThreadSynchronize());
}
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);
} // namespace canny
getEdges<<<grid, block>>>(map, dst, rows, cols);
cudaSafeCall( cudaGetLastError() );
END_OPENCV_DEVICE_NAMESPACE
cudaSafeCall(cudaThreadSynchronize());
}
} // namespace canny
}}} // namespace cv { namespace gpu { namespace device

621
modules/gpu/src/cuda/color.cu
Unescape View File

@ -44,181 +44,181 @@
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/color.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_x = 8 };
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr555_traits::functor_type)
{
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr565_traits::functor_type)
{
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
namespace cv { namespace gpu { namespace device
{
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_x = 8 };
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr555_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr565_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR(name, traits) \
void name(const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream) \
@ -226,7 +226,7 @@ DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
traits::functor_type functor = traits::create_functor(); \
typedef typename traits::functor_type::argument_type src_t; \
typedef typename traits::functor_type::result_type dst_t; \
OPENCV_DEVICE_NAMESPACE_ transform((DevMem2D_<src_t>)src, (DevMem2D_<dst_t>)dst, functor, stream); \
::cv::gpu::device::transform((DevMem2D_<src_t>)src, (DevMem2D_<dst_t>)dst, functor, stream); \
}
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(name) \
@ -243,138 +243,137 @@ DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_8u, name ## _full_traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_32f, name ## _full_traits<float>)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgra)
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F
END_OPENCV_DEVICE_NAMESPACE
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgra)
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F
}}} // namespace cv { namespace gpu { namespace device

352
modules/gpu/src/cuda/column_filter.cu
Unescape View File

@ -47,203 +47,201 @@
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
namespace column_filter {
namespace cv { namespace gpu { namespace device
{
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
__constant__ float c_kernel[MAX_KERNEL_SIZE];
namespace column_filter
{
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float kernel[], int ksize)
{
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
void loadKernel(const float kernel[], int ksize)
{
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearColumnFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearColumnFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
__shared__ T smem[BLOCK_DIM_X][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_Y + 1];
__shared__ T smem[BLOCK_DIM_X][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_Y + 1];
//Offset to the upper halo edge
const int x = blockIdx.x * BLOCK_DIM_X + threadIdx.x;
const int y = (blockIdx.y * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_Y + threadIdx.y;
//Offset to the upper halo edge
const int x = blockIdx.x * BLOCK_DIM_X + threadIdx.x;
const int y = (blockIdx.y * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_Y + threadIdx.y;
if (x < src.cols)
{
const T* src_col = src.ptr() + x;
if (x < src.cols)
{
const T* src_col = src.ptr() + x;
//Main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
//Main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
//Upper halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_low(y + i * BLOCK_DIM_Y, src_col, src.step);
//Upper halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_low(y + i * BLOCK_DIM_Y, src_col, src.step);
//Lower halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y]= b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
//Lower halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y]= b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
__syncthreads();
__syncthreads();
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for(int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y + j - anchor] * c_kernel[j];
#pragma unroll
for(int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y + j - anchor] * c_kernel[j];
int dstY = y + i * BLOCK_DIM_Y;
int dstY = y + i * BLOCK_DIM_Y;
if (dstY < src.rows)
dst.ptr(dstY)[x] = saturate_cast<D>(sum);
if (dstY < src.rows)
dst.ptr(dstY)[x] = saturate_cast<D>(sum);
}
}
}
}
}
template <int ksize, typename T, typename D, template<typename> class B>
void linearColumnFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, RESULT_STEPS * BLOCK_DIM_Y));
template <int ksize, typename T, typename D, template<typename> class B>
void linearColumnFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, RESULT_STEPS * BLOCK_DIM_Y));
B<T> b(src.rows);
B<T> b(src.rows);
linearColumnFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
linearColumnFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
{
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>
},
{
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>
},
template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
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>
},
{
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>
},
{
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>,
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
{
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>
},
{
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>
},
{
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>
},
{
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>
},
{
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>,
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
template void linearColumnFilter_gpu<float , uchar >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float4, uchar4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float , short >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float2, short2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float3, short3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , int >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
} // namespace column_filter
END_OPENCV_DEVICE_NAMESPACE
template void linearColumnFilter_gpu<float , uchar >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float4, uchar4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float , short >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float2, short2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float3, short3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , int >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
} // namespace column_filter
}}} // namespace cv { namespace gpu { namespace device

162
modules/gpu/src/cuda/copy_make_border.cu
Unescape View File

@ -43,87 +43,85 @@
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename Ptr2D, typename T> __global__ void copyMakeBorder(const Ptr2D src, DevMem2D_<T> dst, int top, int left)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = src(y - top, x - left);
}
template <template <typename> class B, typename T> struct CopyMakeBorderDispatcher
{
static void call(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, int top, int left,
const typename VecTraits<T>::elem_type* borderValue, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<T> brd(src.rows, src.cols, VecTraits<T>::make(borderValue));
BorderReader< PtrStep<T>, B<T> > brdSrc(src, brd);
copyMakeBorder<<<grid, block, 0, stream>>>(brdSrc, dst, top, left);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T, int cn> void copyMakeBorder_gpu(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode,
const T* borderValue, cudaStream_t stream)
namespace cv { namespace gpu { namespace device
{
typedef typename TypeVec<T, cn>::vec_type vec_type;
typedef void (*caller_t)(const DevMem2D_<vec_type>& src, const DevMem2D_<vec_type>& dst, int top, int left, const T* borderValue, cudaStream_t stream);
static const caller_t callers[5] =
namespace imgproc
{
CopyMakeBorderDispatcher<BrdReflect101, vec_type>::call,
CopyMakeBorderDispatcher<BrdReplicate, vec_type>::call,
CopyMakeBorderDispatcher<BrdConstant, vec_type>::call,
CopyMakeBorderDispatcher<BrdReflect, vec_type>::call,
CopyMakeBorderDispatcher<BrdWrap, vec_type>::call
};
callers[borderMode](DevMem2D_<vec_type>(src), DevMem2D_<vec_type>(dst), top, left, borderValue, stream);
}
template void copyMakeBorder_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
} // namespace imgproc
END_OPENCV_DEVICE_NAMESPACE
template <typename Ptr2D, typename T> __global__ void copyMakeBorder(const Ptr2D src, DevMem2D_<T> dst, int top, int left)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = src(y - top, x - left);
}
template <template <typename> class B, typename T> struct CopyMakeBorderDispatcher
{
static void call(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, int top, int left,
const typename VecTraits<T>::elem_type* borderValue, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<T> brd(src.rows, src.cols, VecTraits<T>::make(borderValue));
BorderReader< PtrStep<T>, B<T> > brdSrc(src, brd);
copyMakeBorder<<<grid, block, 0, stream>>>(brdSrc, dst, top, left);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T, int cn> void copyMakeBorder_gpu(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode,
const T* borderValue, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type vec_type;
typedef void (*caller_t)(const DevMem2D_<vec_type>& src, const DevMem2D_<vec_type>& dst, int top, int left, const T* borderValue, cudaStream_t stream);
static const caller_t callers[5] =
{
CopyMakeBorderDispatcher<BrdReflect101, vec_type>::call,
CopyMakeBorderDispatcher<BrdReplicate, vec_type>::call,
CopyMakeBorderDispatcher<BrdConstant, vec_type>::call,
CopyMakeBorderDispatcher<BrdReflect, vec_type>::call,
CopyMakeBorderDispatcher<BrdWrap, vec_type>::call
};
callers[borderMode](DevMem2D_<vec_type>(src), DevMem2D_<vec_type>(dst), top, left, borderValue, stream);
}
template void copyMakeBorder_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

3642
modules/gpu/src/cuda/element_operations.cu
View File

File diff suppressed because it is too large Load Diff

284
modules/gpu/src/cuda/hist.cu
Unescape View File

@ -45,177 +45,175 @@
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#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 (__CUDA_ARCH__ >= 120)
namespace cv { namespace gpu { namespace device
{
#define UINT_BITS 32U
namespace hist {
//Warps == subhistograms per threadblock
#define WARP_COUNT 6
#if (!USE_SMEM_ATOMICS)
//Threadblock size
#define HISTOGRAM256_THREADBLOCK_SIZE (WARP_COUNT * OPENCV_GPU_WARP_SIZE)
#define HISTOGRAM256_BIN_COUNT 256
#define TAG_MASK ( (1U << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE)) - 1U )
//Shared memory per threadblock
#define HISTOGRAM256_THREADBLOCK_MEMORY (WARP_COUNT * HISTOGRAM256_BIN_COUNT)
__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);
}
#define PARTIAL_HISTOGRAM256_COUNT 240
#else
#define MERGE_THREADBLOCK_SIZE 256
#define TAG_MASK 0xFFFFFFFFU
#define USE_SMEM_ATOMICS (__CUDA_ARCH__ >= 120)
__forceinline__ __device__ void addByte(uint* s_WarpHist, uint data, uint threadTag)
namespace hist
{
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;
#if (!USE_SMEM_ATOMICS)
//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;
#define TAG_MASK ( (1U << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE)) - 1U )
//Cycle through the entire data set, update subhistograms for each warp
const uint tag = threadIdx.x << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE);
__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);
}
__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;
}
}
#else
////////////////////////////////////////////////////////////////////////////////
// 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
////////////////////////////////////////////////////////////////////////////////
#define TAG_MASK 0xFFFFFFFFU
__global__ void mergeHistogram256(const uint* d_PartialHistograms, int* d_Histogram)
{
uint sum = 0;
__forceinline__ __device__ void addByte(uint* s_WarpHist, uint data, uint threadTag)
{
atomicAdd(s_WarpHist + data, 1);
}
#pragma unroll
for (uint i = threadIdx.x; i < PARTIAL_HISTOGRAM256_COUNT; i += MERGE_THREADBLOCK_SIZE)
sum += d_PartialHistograms[blockIdx.x + i * HISTOGRAM256_BIN_COUNT];
#endif
__shared__ uint data[MERGE_THREADBLOCK_SIZE];
data[threadIdx.x] = sum;
__forceinline__ __device__ void addWord(uint* s_WarpHist, uint data, uint tag, uint pos_x, uint cols)
{
uint x = pos_x << 2;
for (uint stride = MERGE_THREADBLOCK_SIZE / 2; stride > 0; stride >>= 1)
{
__syncthreads();
if(threadIdx.x < stride)
data[threadIdx.x] += data[threadIdx.x + stride];
}
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);
}
if(threadIdx.x == 0)
d_Histogram[blockIdx.x] = saturate_cast<int>(data[0]);
}
__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;
void histogram256_gpu(DevMem2Db src, int* hist, uint* buf, cudaStream_t stream)
{
histogram256<<<PARTIAL_HISTOGRAM256_COUNT, HISTOGRAM256_THREADBLOCK_SIZE, 0, stream>>>(
DevMem2D_<uint>(src),
buf,
static_cast<uint>(src.rows * src.step / sizeof(uint)),
src.cols);
#pragma unroll
for (uint i = threadIdx.x; i < PARTIAL_HISTOGRAM256_COUNT; i += MERGE_THREADBLOCK_SIZE)
sum += d_PartialHistograms[blockIdx.x + i * HISTOGRAM256_BIN_COUNT];
cudaSafeCall( cudaGetLastError() );
__shared__ uint data[MERGE_THREADBLOCK_SIZE];
data[threadIdx.x] = sum;
mergeHistogram256<<<HISTOGRAM256_BIN_COUNT, MERGE_THREADBLOCK_SIZE, 0, stream>>>(buf, hist);
for (uint stride = MERGE_THREADBLOCK_SIZE / 2; stride > 0; stride >>= 1)
{
__syncthreads();
if(threadIdx.x < stride)
data[threadIdx.x] += data[threadIdx.x + stride];
}
cudaSafeCall( cudaGetLastError() );
if(threadIdx.x == 0)
d_Histogram[blockIdx.x] = saturate_cast<int>(data[0]);
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void histogram256_gpu(DevMem2Db src, int* hist, uint* buf, cudaStream_t stream)
{
histogram256<<<PARTIAL_HISTOGRAM256_COUNT, HISTOGRAM256_THREADBLOCK_SIZE, 0, stream>>>(
DevMem2D_<uint>(src),
buf,
static_cast<uint>(src.rows * src.step / sizeof(uint)),
src.cols);
__constant__ int c_lut[256];
cudaSafeCall( cudaGetLastError() );
__global__ void equalizeHist(const DevMem2Db src, PtrStepb dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
mergeHistogram256<<<HISTOGRAM256_BIN_COUNT, MERGE_THREADBLOCK_SIZE, 0, stream>>>(buf, hist);
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);
}
}
cudaSafeCall( cudaGetLastError() );
void equalizeHist_gpu(DevMem2Db src, DevMem2Db dst, const int* lut, cudaStream_t stream)
{
dim3 block(16, 16);
dim3 grid(divUp(src.cols, block.x), divUp(src.rows, block.y));
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
cudaSafeCall( cudaMemcpyToSymbol(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice) );
__constant__ int c_lut[256];
equalizeHist<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
__global__ void equalizeHist(const DevMem2Db 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(DevMem2Db src, DevMem2Db dst, const int* lut, cudaStream_t stream)
{
dim3 block(16, 16);
dim3 grid(divUp(src.cols, block.x), divUp(src.rows, block.y));
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
cudaSafeCall( cudaMemcpyToSymbol(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice) );
} // namespace hist
equalizeHist<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
END_OPENCV_DEVICE_NAMESPACE
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace hist
}}} // namespace cv { namespace gpu { namespace device

1274
modules/gpu/src/cuda/hog.cu
View File

File diff suppressed because it is too large Load Diff

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

File diff suppressed because it is too large Load Diff

111
modules/gpu/src/cuda/internal_shared.hpp
Unescape View File

@ -50,7 +50,7 @@
#include "safe_call.hpp"
#ifndef CV_PI
#define CV_PI 3.1415926535897932384626433832795f
#define CV_PI 3.1415926535897932384626433832795
#endif
#ifndef CV_PI_F
@ -61,27 +61,21 @@
#endif
#endif
#define BEGIN_OPENCV_DEVICE_NAMESPACE namespace cv { namespace gpu { namespace device {
#define END_OPENCV_DEVICE_NAMESPACE }}}
#define OPENCV_DEVICE_NAMESPACE ::cv::gpu::device
#define OPENCV_DEVICE_NAMESPACE_ ::cv::gpu::device::
#ifdef __CUDACC__
BEGIN_OPENCV_DEVICE_NAMESPACE
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef signed char schar;
typedef unsigned int uint;
template<class T> static inline void bindTexture(const textureReference* tex, const DevMem2D_<T>& img)
namespace cv { namespace gpu { namespace device
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
}
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef signed char schar;
typedef unsigned int uint;
END_OPENCV_DEVICE_NAMESPACE
template<class T> static inline void bindTexture(const textureReference* tex, const DevMem2D_<T>& img)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
}
}}}
#endif
@ -102,87 +96,6 @@ namespace cv { namespace gpu
static inline int divUp(int total, int grain) { return (total + grain - 1) / grain; }
/*template<class T> static inline void uploadConstant(const char* name, const T& value)
{
cudaSafeCall( cudaMemcpyToSymbol(name, &value, sizeof(T)) );
}
template<class T> static inline void uploadConstant(const char* name, const T& value, cudaStream_t stream)
{
cudaSafeCall( cudaMemcpyToSymbolAsync(name, &value, sizeof(T), 0, cudaMemcpyHostToDevice, stream) );
} */
//template<class T> static inline void bindTexture(const char* name, const DevMem2D_<T>& img)
//{
// //!!!! const_cast is disabled!
// //!!!! Please use constructor of 'class texture' instead.
//
// //textureReference* tex;
// //cudaSafeCall( cudaGetTextureReference((const textureReference**)&tex, name) );
// //tex->normalized = normalized;
// //tex->filterMode = filterMode;
// //tex->addressMode[0] = addrMode;
// //tex->addressMode[1] = addrMode;
//
// const textureReference* tex;
// cudaSafeCall( cudaGetTextureReference(&tex, name) );
//
// cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
// cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
//}
//static inline void unbindTexture(const char *name)
//{
// const textureReference* tex;
// cudaSafeCall( cudaGetTextureReference(&tex, name) );
// cudaSafeCall( cudaUnbindTexture(tex) );
//}
//class TextureBinder
//{
//public:
// TextureBinder() : tex_(0) {}
// template <typename T> TextureBinder(const textureReference* tex, const DevMem2D_<T>& img) : tex_(0)
// {
// bind(tex, img);
// }
// template <typename T> TextureBinder(const char* tex_name, const DevMem2D_<T>& img) : tex_(0)
// {
// bind(tex_name, img);
// }
// ~TextureBinder() { unbind(); }
//
// template <typename T> void bind(const textureReference* tex, const DevMem2D_<T>& img)
// {
// unbind();
//
// cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
// cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
//
// tex_ = tex;
// }
// template <typename T> void bind(const char* tex_name, const DevMem2D_<T>& img)
// {
// const textureReference* tex;
// cudaSafeCall( cudaGetTextureReference(&tex, tex_name) );
// bind(tex, img);
// }
//
// void unbind()
// {
// if (tex_)
// {
// cudaUnbindTexture(tex_);
// tex_ = 0;
// }
// }
//
//private:
// const textureReference* tex_;
//};
class NppStreamHandler
{
public:

1536
modules/gpu/src/cuda/match_template.cu
View File

File diff suppressed because it is too large Load Diff

294
modules/gpu/src/cuda/mathfunc.cu
Unescape View File

@ -42,174 +42,172 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace mathfunc {
//////////////////////////////////////////////////////////////////////////////////////
// Cart <-> Polar
struct Nothing
{
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
{
}
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = ::sqrtf(x_data * x_data + y_data * y_data);
}
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = ::atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
struct NonEmptyMag
namespace cv { namespace gpu { namespace device
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
namespace mathfunc
{
return mag[y * mag_step + x];
}
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
::sincosf(scale * angle_data, &sin_a, &cos_a);
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
}
template <typename Mag, typename Angle>
void cartToPolar_caller(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
//////////////////////////////////////////////////////////////////////////////////////
// Cart <-> Polar
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;
struct Nothing
{
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
{
}
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = ::sqrtf(x_data * x_data + y_data * y_data);
}
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = ::atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
void cartToPolar_gpu(DevMem2Df x, DevMem2Df y, DevMem2Df mag, bool magSqr, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
struct NonEmptyMag
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
{
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
return mag[y * mag_step + x];
}
},
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
::sincosf(scale * angle_data, &sin_a, &cos_a);
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
}
};
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
template <typename Mag, typename Angle>
void cartToPolar_caller(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
template <typename Mag>
void polarToCart_caller(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(x.cols, threads.x);
grid.y = divUp(x.rows, threads.y);
const float scale = angleInDegrees ? (float)(180.0f / CV_PI) : 1.f;
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;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void cartToPolar_gpu(DevMem2Df x, DevMem2Df y, DevMem2Df mag, bool magSqr, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
{
{
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
}
},
{
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
}
}
};
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
void polarToCart_gpu(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
template <typename Mag>
void polarToCart_caller(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(mag.cols, threads.x);
grid.y = divUp(mag.rows, threads.y);
const float scale = angleInDegrees ? (float)(CV_PI / 180.0f) : 1.0f;
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace mathfunc
void polarToCart_gpu(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
END_OPENCV_DEVICE_NAMESPACE
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
} // namespace mathfunc
}}} // namespace cv { namespace gpu { namespace device

495
modules/gpu/src/cuda/matrix_operations.cu
Unescape View File

@ -45,304 +45,303 @@
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/functional.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T> struct shift_and_sizeof;
template <> struct shift_and_sizeof<signed char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<unsigned char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<unsigned short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<int> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<float> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<double> { enum { shift = 3 }; };
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// CopyTo /////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template<typename T>
__global__ void copy_to_with_mask(const T* mat_src, T* mat_dst, const uchar* mask, int cols, int rows, size_t step_mat, size_t step_mask, int channels)
namespace cv { namespace gpu { namespace device
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
{
size_t idx = y * ( step_mat >> shift_and_sizeof<T>::shift ) + x;
mat_dst[idx] = mat_src[idx];
}
}
template<typename T>
void copy_to_with_mask_run(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
dim3 threadsPerBlock(16,16, 1);
dim3 numBlocks ( divUp(mat_src.cols * channels , threadsPerBlock.x) , divUp(mat_src.rows , threadsPerBlock.y), 1);
copy_to_with_mask<T><<<numBlocks,threadsPerBlock, 0, stream>>>
((T*)mat_src.data, (T*)mat_dst.data, (unsigned char*)mask.data, mat_src.cols, mat_src.rows, mat_src.step, mask.step, channels);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
void copy_to_with_mask(const DevMem2Db& mat_src, DevMem2Db mat_dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
typedef void (*CopyToFunc)(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream);
static CopyToFunc tab[8] =
template <typename T> struct shift_and_sizeof;
template <> struct shift_and_sizeof<signed char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<unsigned char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<unsigned short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<int> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<float> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<double> { enum { shift = 3 }; };
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// CopyTo /////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template<typename T>
__global__ void copy_to_with_mask(const T* mat_src, T* mat_dst, const uchar* mask, int cols, int rows, size_t step_mat, size_t step_mask, int channels)
{
copy_to_with_mask_run<unsigned char>,
copy_to_with_mask_run<signed char>,
copy_to_with_mask_run<unsigned short>,
copy_to_with_mask_run<short>,
copy_to_with_mask_run<int>,
copy_to_with_mask_run<float>,
copy_to_with_mask_run<double>,
0
};
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
{
size_t idx = y * ( step_mat >> shift_and_sizeof<T>::shift ) + x;
mat_dst[idx] = mat_src[idx];
}
}
CopyToFunc func = tab[depth];
template<typename T>
void copy_to_with_mask_run(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
dim3 threadsPerBlock(16,16, 1);
dim3 numBlocks ( divUp(mat_src.cols * channels , threadsPerBlock.x) , divUp(mat_src.rows , threadsPerBlock.y), 1);
if (func == 0) cv::gpu::error("Unsupported copyTo operation", __FILE__, __LINE__);
copy_to_with_mask<T><<<numBlocks,threadsPerBlock, 0, stream>>>
((T*)mat_src.data, (T*)mat_dst.data, (unsigned char*)mask.data, mat_src.cols, mat_src.rows, mat_src.step, mask.step, channels);
cudaSafeCall( cudaGetLastError() );
func(mat_src, mat_dst, mask, channels, stream);
}
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// SetTo //////////////////////////////////
///////////////////////////////////////////////////////////////////////////
void copy_to_with_mask(const DevMem2Db& mat_src, DevMem2Db mat_dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
typedef void (*CopyToFunc)(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream);
__constant__ uchar scalar_8u[4];
__constant__ schar scalar_8s[4];
__constant__ ushort scalar_16u[4];
__constant__ short scalar_16s[4];
__constant__ int scalar_32s[4];
__constant__ float scalar_32f[4];
__constant__ double scalar_64f[4];
static CopyToFunc tab[8] =
{
copy_to_with_mask_run<unsigned char>,
copy_to_with_mask_run<signed char>,
copy_to_with_mask_run<unsigned short>,
copy_to_with_mask_run<short>,
copy_to_with_mask_run<int>,
copy_to_with_mask_run<float>,
copy_to_with_mask_run<double>,
0
};
template <typename T> __device__ __forceinline__ T readScalar(int i);
template <> __device__ __forceinline__ uchar readScalar<uchar>(int i) {return scalar_8u[i];}
template <> __device__ __forceinline__ schar readScalar<schar>(int i) {return scalar_8s[i];}
template <> __device__ __forceinline__ ushort readScalar<ushort>(int i) {return scalar_16u[i];}
template <> __device__ __forceinline__ short readScalar<short>(int i) {return scalar_16s[i];}
template <> __device__ __forceinline__ int readScalar<int>(int i) {return scalar_32s[i];}
template <> __device__ __forceinline__ float readScalar<float>(int i) {return scalar_32f[i];}
template <> __device__ __forceinline__ double readScalar<double>(int i) {return scalar_64f[i];}
CopyToFunc func = tab[depth];
void writeScalar(const uchar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8u, vals, sizeof(uchar) * 4) );
}
void writeScalar(const schar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8s, vals, sizeof(schar) * 4) );
}
void writeScalar(const ushort* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16u, vals, sizeof(ushort) * 4) );
}
void writeScalar(const short* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16s, vals, sizeof(short) * 4) );
}
void writeScalar(const int* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_32s, vals, sizeof(int) * 4) );
}
void writeScalar(const float* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_32f, vals, sizeof(float) * 4) );
}
void writeScalar(const double* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_64f, vals, sizeof(double) * 4) );
}
if (func == 0) cv::gpu::error("Unsupported copyTo operation", __FILE__, __LINE__);
template<typename T>
__global__ void set_to_without_mask(T* mat, int cols, int rows, size_t step, int channels)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
func(mat_src, mat_dst, mask, channels, stream);
}
if ((x < cols * channels ) && (y < rows))
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// SetTo //////////////////////////////////
///////////////////////////////////////////////////////////////////////////
__constant__ uchar scalar_8u[4];
__constant__ schar scalar_8s[4];
__constant__ ushort scalar_16u[4];
__constant__ short scalar_16s[4];
__constant__ int scalar_32s[4];
__constant__ float scalar_32f[4];
__constant__ double scalar_64f[4];
template <typename T> __device__ __forceinline__ T readScalar(int i);
template <> __device__ __forceinline__ uchar readScalar<uchar>(int i) {return scalar_8u[i];}
template <> __device__ __forceinline__ schar readScalar<schar>(int i) {return scalar_8s[i];}
template <> __device__ __forceinline__ ushort readScalar<ushort>(int i) {return scalar_16u[i];}
template <> __device__ __forceinline__ short readScalar<short>(int i) {return scalar_16s[i];}
template <> __device__ __forceinline__ int readScalar<int>(int i) {return scalar_32s[i];}
template <> __device__ __forceinline__ float readScalar<float>(int i) {return scalar_32f[i];}
template <> __device__ __forceinline__ double readScalar<double>(int i) {return scalar_64f[i];}
void writeScalar(const uchar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8u, vals, sizeof(uchar) * 4) );
}
void writeScalar(const schar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8s, vals, sizeof(schar) * 4) );
}
void writeScalar(const ushort* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16u, vals, sizeof(ushort) * 4) );
}
void writeScalar(const short* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16s, vals, sizeof(short) * 4) );
}
void writeScalar(const int* vals)
{
size_t idx = y * ( step >> shift_and_sizeof<T>::shift ) + x;
mat[idx] = readScalar<T>(x % channels);
cudaSafeCall( cudaMemcpyToSymbol(scalar_32s, vals, sizeof(int) * 4) );
}
void writeScalar(const float* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_32f, vals, sizeof(float) * 4) );
}
void writeScalar(const double* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_64f, vals, sizeof(double) * 4) );
}
}
template<typename T>
__global__ void set_to_with_mask(T* mat, const uchar* mask, int cols, int rows, size_t step, int channels, size_t step_mask)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
template<typename T>
__global__ void set_to_without_mask(T* mat, int cols, int rows, size_t step, int channels)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
if ((x < cols * channels ) && (y < rows))
{
size_t idx = y * ( step >> shift_and_sizeof<T>::shift ) + x;
mat[idx] = readScalar<T>(x % channels);
}
}
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream)
{
writeScalar(scalar);
}
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
template<typename T>
__global__ void set_to_with_mask(T* mat, const uchar* mask, int cols, int rows, size_t step, int channels, size_t step_mask)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
{
size_t idx = y * ( step >> shift_and_sizeof<T>::shift ) + x;
mat[idx] = readScalar<T>(x % channels);
}
}
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream)
{
writeScalar(scalar);
set_to_with_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, (uchar*)mask.data, mat.cols, mat.rows, mat.step, channels, mask.step);
cudaSafeCall( cudaGetLastError() );
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
set_to_with_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, (uchar*)mask.data, mat.cols, mat.rows, mat.step, channels, mask.step);
cudaSafeCall( cudaGetLastError() );
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream)
{
writeScalar(scalar);
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream)
{
writeScalar(scalar);
set_to_without_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, mat.cols, mat.rows, mat.step, channels);
cudaSafeCall( cudaGetLastError() );
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
set_to_without_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, mat.cols, mat.rows, mat.step, channels);
cudaSafeCall( cudaGetLastError() );
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, int channels, cudaStream_t stream);
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, int channels, cudaStream_t stream);
template <typename T, typename D> struct Convertor : unary_function<T, D>
{
Convertor(double alpha_, double beta_) : alpha(alpha_), beta(beta_) {}
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
__device__ __forceinline__ D operator()(const T& src) const
template <typename T, typename D> struct Convertor : unary_function<T, D>
{
return saturate_cast<D>(alpha * src + beta);
}
Convertor(double alpha_, double beta_) : alpha(alpha_), beta(beta_) {}
const double alpha, beta;
};
__device__ __forceinline__ D operator()(const T& src) const
{
return saturate_cast<D>(alpha * src + beta);
}
namespace detail
{
template <size_t src_size, size_t dst_size, typename F> struct ConvertTraitsDispatcher : DefaultTransformFunctorTraits<F>
{
};
template <typename F> struct ConvertTraitsDispatcher<1, 1, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 8 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
const double alpha, beta;
};
template <typename F> struct ConvertTraitsDispatcher<2, 2, F> : DefaultTransformFunctorTraits<F>
namespace detail
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<2, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 2 };
};
template <size_t src_size, size_t dst_size, typename F> struct ConvertTraitsDispatcher : DefaultTransformFunctorTraits<F>
{
};
template <typename F> struct ConvertTraitsDispatcher<1, 1, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 8 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<4, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<4, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 2 };
};
template <typename F> struct ConvertTraitsDispatcher<2, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<2, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 2 };
};
template <typename F> struct ConvertTraitsDispatcher<4, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<4, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 2 };
};
template <typename F> struct ConvertTraits : ConvertTraitsDispatcher<sizeof(typename F::argument_type), sizeof(typename F::result_type), F>
{
};
}
template <typename F> struct ConvertTraits : ConvertTraitsDispatcher<sizeof(typename F::argument_type), sizeof(typename F::result_type), F>
template <typename T, typename D> struct TransformFunctorTraits< Convertor<T, D> > : detail::ConvertTraits< Convertor<T, D> >
{
};
}
template <typename T, typename D> struct TransformFunctorTraits< Convertor<T, D> > : detail::ConvertTraits< Convertor<T, D> >
{
};
template<typename T, typename D>
void cvt_(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta, cudaStream_t stream)
{
cudaSafeCall( cudaSetDoubleForDevice(&alpha) );
cudaSafeCall( cudaSetDoubleForDevice(&beta) );
Convertor<T, D> op(alpha, beta);
OPENCV_DEVICE_NAMESPACE_ transform((DevMem2D_<T>)src, (DevMem2D_<D>)dst, op, stream);
}
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta,
cudaStream_t stream = 0)
{
typedef void (*caller_t)(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta,
cudaStream_t stream);
template<typename T, typename D>
void cvt_(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta, cudaStream_t stream)
{
cudaSafeCall( cudaSetDoubleForDevice(&alpha) );
cudaSafeCall( cudaSetDoubleForDevice(&beta) );
Convertor<T, D> op(alpha, beta);
::cv::gpu::device::transform((DevMem2D_<T>)src, (DevMem2D_<D>)dst, op, stream);
}
static const caller_t tab[8][8] =
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta,
cudaStream_t stream = 0)
{
{cvt_<uchar, uchar>, cvt_<uchar, schar>, cvt_<uchar, ushort>, cvt_<uchar, short>,
cvt_<uchar, int>, cvt_<uchar, float>, cvt_<uchar, double>, 0},
typedef void (*caller_t)(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta,
cudaStream_t stream);
{cvt_<schar, uchar>, cvt_<schar, schar>, cvt_<schar, ushort>, cvt_<schar, short>,
cvt_<schar, int>, cvt_<schar, float>, cvt_<schar, double>, 0},
static const caller_t tab[8][8] =
{
{cvt_<uchar, uchar>, cvt_<uchar, schar>, cvt_<uchar, ushort>, cvt_<uchar, short>,
cvt_<uchar, int>, cvt_<uchar, float>, cvt_<uchar, 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_<schar, uchar>, cvt_<schar, schar>, cvt_<schar, ushort>, cvt_<schar, short>,
cvt_<schar, int>, cvt_<schar, float>, cvt_<schar, 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_<ushort, uchar>, cvt_<ushort, schar>, cvt_<ushort, ushort>, cvt_<ushort, short>,
cvt_<ushort, int>, cvt_<ushort, float>, cvt_<ushort, 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_<short, uchar>, cvt_<short, schar>, cvt_<short, ushort>, cvt_<short, short>,
cvt_<short, int>, cvt_<short, float>, cvt_<short, 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_<int, uchar>, cvt_<int, schar>, cvt_<int, ushort>,
cvt_<int, short>, cvt_<int, int>, cvt_<int, float>, cvt_<int, double>, 0},
{cvt_<double, uchar>, cvt_<double, schar>, cvt_<double, ushort>,
cvt_<double, short>, cvt_<double, int>, cvt_<double, float>, cvt_<double, double>, 0},
{cvt_<float, uchar>, cvt_<float, schar>, cvt_<float, ushort>,
cvt_<float, short>, cvt_<float, int>, cvt_<float, float>, cvt_<float, double>, 0},
{0,0,0,0,0,0,0,0}
};
{cvt_<double, uchar>, cvt_<double, schar>, cvt_<double, ushort>,
cvt_<double, short>, cvt_<double, int>, cvt_<double, float>, cvt_<double, double>, 0},
caller_t func = tab[sdepth][ddepth];
if (!func)
cv::gpu::error("Unsupported convert operation", __FILE__, __LINE__);
{0,0,0,0,0,0,0,0}
};
func(src, dst, alpha, beta, stream);
}
caller_t func = tab[sdepth][ddepth];
if (!func)
cv::gpu::error("Unsupported convert operation", __FILE__, __LINE__);
END_OPENCV_DEVICE_NAMESPACE
func(src, dst, alpha, beta, stream);
}
}}} // namespace cv { namespace gpu { namespace device

3408
modules/gpu/src/cuda/matrix_reductions.cu
View File

File diff suppressed because it is too large Load Diff

260
modules/gpu/src/cuda/pyr_down.cu
Unescape View File

@ -46,142 +46,140 @@
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename T, typename B> __global__ void pyrDown(const PtrStep<T> src, PtrStep<T> dst, const B b, int dst_cols)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y;
__shared__ value_type smem[256 + 4];
value_type sum;
const int src_y = 2*y;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, x, src.data, src.step);
smem[2 + threadIdx.x] = sum;
if (threadIdx.x < 2)
{
const int left_x = x - 2 + threadIdx.x;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, left_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, left_x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , left_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, left_x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, left_x, src.data, src.step);
smem[threadIdx.x] = sum;
}
if (threadIdx.x > 253)
{
const int right_x = x + threadIdx.x + 2;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, right_x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, right_x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, right_x, src.data, src.step);
smem[4 + threadIdx.x] = sum;
}
__syncthreads();
if (threadIdx.x < 128)
{
const int tid2 = threadIdx.x * 2;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
sum = sum + 0.0625f * smem[2 + tid2 + 2];
const int dst_x = (blockIdx.x * blockDim.x + tid2) / 2;
if (dst_x < dst_cols)
dst.ptr(y)[dst_x] = saturate_cast<T>(sum);
}
}
template <typename T, template <typename> class B> void pyrDown_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
namespace cv { namespace gpu { namespace device
{
const dim3 block(256);
const dim3 grid(divUp(src.cols, block.x), dst.rows);
B<T> b(src.rows, src.cols);
pyrDown<T><<<grid, block, 0, stream>>>(src, dst, b, dst.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, int cn> void pyrDown_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
static const caller_t callers[] =
namespace imgproc
{
pyrDown_caller<type, BrdReflect101>, pyrDown_caller<type, BrdReplicate>, pyrDown_caller<type, BrdConstant>, pyrDown_caller<type, BrdReflect>, pyrDown_caller<type, BrdWrap>
};
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
template void pyrDown_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template <typename T, typename B> __global__ void pyrDown(const PtrStep<T> src, PtrStep<T> dst, const B b, int dst_cols)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
template void pyrDown_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y;
template void pyrDown_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
__shared__ value_type smem[256 + 4];
template void pyrDown_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
value_type sum;
const int src_y = 2*y;
template void pyrDown_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, x, src.data, src.step);
smem[2 + threadIdx.x] = sum;
} // namespace imgproc
if (threadIdx.x < 2)
{
const int left_x = x - 2 + threadIdx.x;
END_OPENCV_DEVICE_NAMESPACE
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, left_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, left_x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , left_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, left_x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, left_x, src.data, src.step);
smem[threadIdx.x] = sum;
}
if (threadIdx.x > 253)
{
const int right_x = x + threadIdx.x + 2;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, right_x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, right_x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, right_x, src.data, src.step);
smem[4 + threadIdx.x] = sum;
}
__syncthreads();
if (threadIdx.x < 128)
{
const int tid2 = threadIdx.x * 2;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
sum = sum + 0.0625f * smem[2 + tid2 + 2];
const int dst_x = (blockIdx.x * blockDim.x + tid2) / 2;
if (dst_x < dst_cols)
dst.ptr(y)[dst_x] = saturate_cast<T>(sum);
}
}
template <typename T, template <typename> class B> void pyrDown_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
{
const dim3 block(256);
const dim3 grid(divUp(src.cols, block.x), dst.rows);
B<T> b(src.rows, src.cols);
pyrDown<T><<<grid, block, 0, stream>>>(src, dst, b, dst.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, int cn> void pyrDown_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
static const caller_t callers[] =
{
pyrDown_caller<type, BrdReflect101>, pyrDown_caller<type, BrdReplicate>, pyrDown_caller<type, BrdConstant>, pyrDown_caller<type, BrdReflect>, pyrDown_caller<type, BrdWrap>
};
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
template void pyrDown_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

200
modules/gpu/src/cuda/pyr_up.cu
Unescape View File

@ -46,137 +46,135 @@
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename T, typename B> __global__ void pyrUp(const PtrStep<T> src, DevMem2D_<T> dst, const B b)
namespace cv { namespace gpu { namespace device
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
namespace imgproc
{
template <typename T, typename B> __global__ void pyrUp(const PtrStep<T> src, DevMem2D_<T> dst, const B b)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
__shared__ T smem1[10][10];
__shared__ value_type smem2[20][16];
__shared__ T smem1[10][10];
__shared__ value_type smem2[20][16];
value_type sum;
value_type sum;
if (threadIdx.x < 10 && threadIdx.y < 10)
smem1[threadIdx.y][threadIdx.x] = b.at(blockIdx.y * blockDim.y / 2 + threadIdx.y - 1, blockIdx.x * blockDim.x / 2 + threadIdx.x - 1, src.data, src.step);
if (threadIdx.x < 10 && threadIdx.y < 10)
smem1[threadIdx.y][threadIdx.x] = b.at(blockIdx.y * blockDim.y / 2 + threadIdx.y - 1, blockIdx.x * blockDim.x / 2 + threadIdx.x - 1, src.data, src.step);
__syncthreads();
__syncthreads();
const int tidx = threadIdx.x;
const int tidx = threadIdx.x;
sum = VecTraits<value_type>::all(0);
sum = VecTraits<value_type>::all(0);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[1 + threadIdx.y / 2][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 2) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[1 + threadIdx.y / 2][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 2) >> 1)];
smem2[2 + threadIdx.y][tidx] = sum;
smem2[2 + threadIdx.y][tidx] = sum;
if (threadIdx.y < 2)
{
sum = VecTraits<value_type>::all(0);
if (threadIdx.y < 2)
{
sum = VecTraits<value_type>::all(0);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[0][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx + 2) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[0][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx + 2) >> 1)];
smem2[threadIdx.y][tidx] = sum;
}
smem2[threadIdx.y][tidx] = sum;
}
if (threadIdx.y > 13)
{
sum = VecTraits<value_type>::all(0);
if (threadIdx.y > 13)
{
sum = VecTraits<value_type>::all(0);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[9][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx + 2) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[9][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx + 2) >> 1)];
smem2[4 + threadIdx.y][tidx] = sum;
}
smem2[4 + threadIdx.y][tidx] = sum;
}
__syncthreads();
__syncthreads();
sum = VecTraits<value_type>::all(0);
sum = VecTraits<value_type>::all(0);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y - 2][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y - 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.375f * smem2[2 + threadIdx.y ][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y + 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y + 2][tidx];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y - 2][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y - 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.375f * smem2[2 + threadIdx.y ][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y + 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y + 2][tidx];
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = saturate_cast<T>(4.0f * sum);
}
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = saturate_cast<T>(4.0f * sum);
}
template <typename T, template <typename> class B> void pyrUp_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
{
const dim3 block(16, 16);
const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
template <typename T, template <typename> class B> void pyrUp_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
{
const dim3 block(16, 16);
const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<T> b(src.rows, src.cols);
B<T> b(src.rows, src.cols);
pyrUp<T><<<grid, block, 0, stream>>>(src, dst, b);
cudaSafeCall( cudaGetLastError() );
pyrUp<T><<<grid, block, 0, stream>>>(src, dst, b);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, int cn> void pyrUp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
static const caller_t callers[] =
{
pyrUp_caller<type, BrdReflect101>, pyrUp_caller<type, BrdReplicate>, pyrUp_caller<type, BrdConstant>, pyrUp_caller<type, BrdReflect>, pyrUp_caller<type, BrdWrap>
};
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
template <typename T, int cn> void pyrUp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
template void pyrUp_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
template void pyrUp_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
static const caller_t callers[] =
{
pyrUp_caller<type, BrdReflect101>, pyrUp_caller<type, BrdReplicate>, pyrUp_caller<type, BrdConstant>, pyrUp_caller<type, BrdReflect>, pyrUp_caller<type, BrdWrap>
};
template void pyrUp_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
template void pyrUp_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
} // namespace imgproc
template void pyrUp_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template void pyrUp_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

398
modules/gpu/src/cuda/remap.cu
Unescape View File

@ -47,208 +47,206 @@
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename Ptr2D, typename T> __global__ void remap(const Ptr2D src, const PtrStepf mapx, const PtrStepf mapy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = mapx.ptr(y)[x];
const float ycoo = mapy.ptr(y)[x];
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherStream
namespace cv { namespace gpu { namespace device
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block, 0, stream>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst, const float* borderValue, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_REMAP_TEX(type) \
texture< type , cudaTextureType2D> tex_remap_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_remap_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_remap_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter, template <typename> class B> struct RemapDispatcherNonStream<Filter, B, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_remap_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <template <typename> class Filter> struct RemapDispatcherNonStream<Filter, BrdReplicate, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
Filter< tex_remap_ ## type ##_reader > filter_src(texSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(short2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(float2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_REMAP_TEX
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcher
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
RemapDispatcherNonStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, cc);
else
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc);
}
};
template <typename T> void remap_gpu(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc);
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void remap(const Ptr2D src, const PtrStepf mapx, const PtrStepf mapy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = mapx.ptr(y)[x];
const float ycoo = mapy.ptr(y)[x];
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
static const caller_t callers[3][5] =
{
{
RemapDispatcher<PointFilter, BrdReflect101, T>::call,
RemapDispatcher<PointFilter, BrdReplicate, T>::call,
RemapDispatcher<PointFilter, BrdConstant, T>::call,
RemapDispatcher<PointFilter, BrdReflect, T>::call,
RemapDispatcher<PointFilter, BrdWrap, T>::call
},
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherStream
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block, 0, stream>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst, const float* borderValue, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_REMAP_TEX(type) \
texture< type , cudaTextureType2D> tex_remap_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_remap_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_remap_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter, template <typename> class B> struct RemapDispatcherNonStream<Filter, B, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_remap_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <template <typename> class Filter> struct RemapDispatcherNonStream<Filter, BrdReplicate, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
Filter< tex_remap_ ## type ##_reader > filter_src(texSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(short2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(float2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_REMAP_TEX
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcher
{
RemapDispatcher<LinearFilter, BrdReflect101, T>::call,
RemapDispatcher<LinearFilter, BrdReplicate, T>::call,
RemapDispatcher<LinearFilter, BrdConstant, T>::call,
RemapDispatcher<LinearFilter, BrdReflect, T>::call,
RemapDispatcher<LinearFilter, BrdWrap, T>::call
},
{
RemapDispatcher<CubicFilter, BrdReflect101, T>::call,
RemapDispatcher<CubicFilter, BrdReplicate, T>::call,
RemapDispatcher<CubicFilter, BrdConstant, T>::call,
RemapDispatcher<CubicFilter, BrdReflect, T>::call,
RemapDispatcher<CubicFilter, BrdWrap, T>::call
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
RemapDispatcherNonStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, cc);
else
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc);
}
};
template <typename T> void remap_gpu(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc);
static const caller_t callers[3][5] =
{
{
RemapDispatcher<PointFilter, BrdReflect101, T>::call,
RemapDispatcher<PointFilter, BrdReplicate, T>::call,
RemapDispatcher<PointFilter, BrdConstant, T>::call,
RemapDispatcher<PointFilter, BrdReflect, T>::call,
RemapDispatcher<PointFilter, BrdWrap, T>::call
},
{
RemapDispatcher<LinearFilter, BrdReflect101, T>::call,
RemapDispatcher<LinearFilter, BrdReplicate, T>::call,
RemapDispatcher<LinearFilter, BrdConstant, T>::call,
RemapDispatcher<LinearFilter, BrdReflect, T>::call,
RemapDispatcher<LinearFilter, BrdWrap, T>::call
},
{
RemapDispatcher<CubicFilter, BrdReflect101, T>::call,
RemapDispatcher<CubicFilter, BrdReplicate, T>::call,
RemapDispatcher<CubicFilter, BrdConstant, T>::call,
RemapDispatcher<CubicFilter, BrdReflect, T>::call,
RemapDispatcher<CubicFilter, BrdWrap, T>::call
}
};
callers[interpolation][borderMode](static_cast< DevMem2D_<T> >(src), xmap, ymap, static_cast< DevMem2D_<T> >(dst), borderValue, stream, cc);
}
};
callers[interpolation][borderMode](static_cast< DevMem2D_<T> >(src), xmap, ymap, static_cast< DevMem2D_<T> >(dst), borderValue, stream, cc);
}
template void remap_gpu<uchar >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<uchar2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<schar>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<ushort2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<short2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<float2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
} // namespace imgproc
END_OPENCV_DEVICE_NAMESPACE
template void remap_gpu<uchar >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<uchar2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<schar>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<ushort2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<short2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<float2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

428
modules/gpu/src/cuda/resize.cu
Unescape View File

@ -47,219 +47,217 @@
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename Ptr2D, typename T> __global__ void resize(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <typename Ptr2D, typename T> __global__ void resizeNN(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
dst.ptr(y)[x] = src(__float2int_rd(ycoo), __float2int_rd(xcoo));
}
}
template <template <typename> class Filter, typename T> struct ResizeDispatcherStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
resize<<<grid, block, 0, stream>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <typename T> struct ResizeDispatcherStream<PointFilter, T>
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
resizeNN<<<grid, block, 0, stream>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, typename T> struct ResizeDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
resize<<<grid, block>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T> struct ResizeDispatcherNonStream<PointFilter, T>
namespace cv { namespace gpu { namespace device
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
resizeNN<<<grid, block>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
texture< type , cudaTextureType2D> tex_resize_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_resize_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_resize_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter> struct ResizeDispatcherNonStream<Filter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
Filter< tex_resize_ ## type ##_reader > filter_src(texSrc); \
resize<<<grid, block>>>(filter_src, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <> struct ResizeDispatcherNonStream<PointFilter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
resizeNN<<<grid, block>>>(texSrc, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(schar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
template <template <typename> class Filter, typename T> struct ResizeDispatcher
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
if (stream == 0)
ResizeDispatcherNonStream<Filter, T>::call(src, fx, fy, dst);
else
ResizeDispatcherStream<Filter, T>::call(src, fx, fy, dst, stream);
}
};
template <typename T> void resize_gpu(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream);
static const caller_t callers[3] =
{
ResizeDispatcher<PointFilter, T>::call, ResizeDispatcher<LinearFilter, T>::call, ResizeDispatcher<CubicFilter, T>::call
};
callers[interpolation](static_cast< DevMem2D_<T> >(src), fx, fy, static_cast< DevMem2D_<T> >(dst), stream);
}
template void resize_gpu<uchar >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<uchar2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<schar>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort >(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<ushort2>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort3>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort4>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<short2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<float2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
} // namespace imgproc
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void resize(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <typename Ptr2D, typename T> __global__ void resizeNN(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
dst.ptr(y)[x] = src(__float2int_rd(ycoo), __float2int_rd(xcoo));
}
}
template <template <typename> class Filter, typename T> struct ResizeDispatcherStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
resize<<<grid, block, 0, stream>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <typename T> struct ResizeDispatcherStream<PointFilter, T>
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
resizeNN<<<grid, block, 0, stream>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, typename T> struct ResizeDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
resize<<<grid, block>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T> struct ResizeDispatcherNonStream<PointFilter, T>
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
resizeNN<<<grid, block>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
texture< type , cudaTextureType2D> tex_resize_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_resize_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_resize_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter> struct ResizeDispatcherNonStream<Filter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
Filter< tex_resize_ ## type ##_reader > filter_src(texSrc); \
resize<<<grid, block>>>(filter_src, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <> struct ResizeDispatcherNonStream<PointFilter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
resizeNN<<<grid, block>>>(texSrc, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(schar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
template <template <typename> class Filter, typename T> struct ResizeDispatcher
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
if (stream == 0)
ResizeDispatcherNonStream<Filter, T>::call(src, fx, fy, dst);
else
ResizeDispatcherStream<Filter, T>::call(src, fx, fy, dst, stream);
}
};
template <typename T> void resize_gpu(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream);
static const caller_t callers[3] =
{
ResizeDispatcher<PointFilter, T>::call, ResizeDispatcher<LinearFilter, T>::call, ResizeDispatcher<CubicFilter, T>::call
};
callers[interpolation](static_cast< DevMem2D_<T> >(src), fx, fy, static_cast< DevMem2D_<T> >(dst), stream);
}
template void resize_gpu<uchar >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<uchar2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<schar>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort >(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<ushort2>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort3>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort4>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<short2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<float2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

388
modules/gpu/src/cuda/row_filter.cu
Unescape View File

@ -47,226 +47,224 @@
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
namespace row_filter {
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float kernel[], int ksize)
namespace cv { namespace gpu { namespace device
{
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
namespace detail
{
template <typename T, size_t size> struct SmemType
namespace row_filter
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type smem_t;
};
__constant__ float c_kernel[MAX_KERNEL_SIZE];
template <typename T> struct SmemType<T, 4>
{
typedef T smem_t;
};
}
void loadKernel(const float kernel[], int ksize)
{
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
template <typename T> struct SmemType
{
typedef typename detail::SmemType<T, sizeof(T)>::smem_t smem_t;
};
namespace detail
{
template <typename T, size_t size> struct SmemType
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type smem_t;
};
template <typename T> struct SmemType<T, 4>
{
typedef T smem_t;
};
}
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearRowFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename SmemType<T>::smem_t smem_t;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
template <typename T> struct SmemType
{
typedef typename detail::SmemType<T, sizeof(T)>::smem_t smem_t;
};
__shared__ smem_t smem[BLOCK_DIM_Y][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_X];
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearRowFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename SmemType<T>::smem_t smem_t;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
//Offset to the left halo edge
const int x = (blockIdx.x * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_X + threadIdx.x;
const int y = blockIdx.y * BLOCK_DIM_Y + threadIdx.y;
__shared__ smem_t smem[BLOCK_DIM_Y][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_X];
if (y < src.rows)
{
const T* src_row = src.ptr(y);
//Offset to the left halo edge
const int x = (blockIdx.x * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_X + threadIdx.x;
const int y = blockIdx.y * BLOCK_DIM_Y + threadIdx.y;
//Load main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
if (y < src.rows)
{
const T* src_row = src.ptr(y);
//Load left halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_low(i * BLOCK_DIM_X + x, src_row);
//Load main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
//Load right halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
//Load left halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_low(i * BLOCK_DIM_X + x, src_row);
__syncthreads();
//Load right halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
D* dst_row = dst.ptr(y);
__syncthreads();
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
{
sum_t sum = VecTraits<sum_t>::all(0);
D* dst_row = dst.ptr(y);
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for (int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X + j - anchor] * c_kernel[j];
#pragma unroll
for (int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X + j - anchor] * c_kernel[j];
int dstX = x + i * BLOCK_DIM_X;
int dstX = x + i * BLOCK_DIM_X;
if (dstX < src.cols)
dst_row[dstX] = saturate_cast<D>(sum);
if (dstX < src.cols)
dst_row[dstX] = saturate_cast<D>(sum);
}
}
}
}
}
template <int ksize, typename T, typename D, template<typename> class B>
void linearRowFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
typedef typename SmemType<T>::smem_t smem_t;
template <int ksize, typename T, typename D, template<typename> class B>
void linearRowFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
typedef typename SmemType<T>::smem_t smem_t;
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, RESULT_STEPS * BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y));
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, RESULT_STEPS * BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y));
B<smem_t> b(src.cols);
B<smem_t> b(src.cols);
linearRowFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
linearRowFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect101>,
linearRowFilter_caller<2 , T, D, BrdRowReflect101>,
linearRowFilter_caller<3 , T, D, BrdRowReflect101>,
linearRowFilter_caller<4 , T, D, BrdRowReflect101>,
linearRowFilter_caller<5 , T, D, BrdRowReflect101>,
linearRowFilter_caller<6 , T, D, BrdRowReflect101>,
linearRowFilter_caller<7 , T, D, BrdRowReflect101>,
linearRowFilter_caller<8 , T, D, BrdRowReflect101>,
linearRowFilter_caller<9 , T, D, BrdRowReflect101>,
linearRowFilter_caller<10, T, D, BrdRowReflect101>,
linearRowFilter_caller<11, T, D, BrdRowReflect101>,
linearRowFilter_caller<12, T, D, BrdRowReflect101>,
linearRowFilter_caller<13, T, D, BrdRowReflect101>,
linearRowFilter_caller<14, T, D, BrdRowReflect101>,
linearRowFilter_caller<15, T, D, BrdRowReflect101>,
linearRowFilter_caller<16, T, D, BrdRowReflect101>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReplicate>,
linearRowFilter_caller<2 , T, D, BrdRowReplicate>,
linearRowFilter_caller<3 , T, D, BrdRowReplicate>,
linearRowFilter_caller<4 , T, D, BrdRowReplicate>,
linearRowFilter_caller<5 , T, D, BrdRowReplicate>,
linearRowFilter_caller<6 , T, D, BrdRowReplicate>,
linearRowFilter_caller<7 , T, D, BrdRowReplicate>,
linearRowFilter_caller<8 , T, D, BrdRowReplicate>,
linearRowFilter_caller<9 , T, D, BrdRowReplicate>,
linearRowFilter_caller<10, T, D, BrdRowReplicate>,
linearRowFilter_caller<11, T, D, BrdRowReplicate>,
linearRowFilter_caller<12, T, D, BrdRowReplicate>,
linearRowFilter_caller<13, T, D, BrdRowReplicate>,
linearRowFilter_caller<14, T, D, BrdRowReplicate>,
linearRowFilter_caller<15, T, D, BrdRowReplicate>,
linearRowFilter_caller<16, T, D, BrdRowReplicate>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowConstant>,
linearRowFilter_caller<2 , T, D, BrdRowConstant>,
linearRowFilter_caller<3 , T, D, BrdRowConstant>,
linearRowFilter_caller<4 , T, D, BrdRowConstant>,
linearRowFilter_caller<5 , T, D, BrdRowConstant>,
linearRowFilter_caller<6 , T, D, BrdRowConstant>,
linearRowFilter_caller<7 , T, D, BrdRowConstant>,
linearRowFilter_caller<8 , T, D, BrdRowConstant>,
linearRowFilter_caller<9 , T, D, BrdRowConstant>,
linearRowFilter_caller<10, T, D, BrdRowConstant>,
linearRowFilter_caller<11, T, D, BrdRowConstant>,
linearRowFilter_caller<12, T, D, BrdRowConstant>,
linearRowFilter_caller<13, T, D, BrdRowConstant>,
linearRowFilter_caller<14, T, D, BrdRowConstant>,
linearRowFilter_caller<15, T, D, BrdRowConstant>,
linearRowFilter_caller<16, T, D, BrdRowConstant>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect>,
linearRowFilter_caller<2 , T, D, BrdRowReflect>,
linearRowFilter_caller<3 , T, D, BrdRowReflect>,
linearRowFilter_caller<4 , T, D, BrdRowReflect>,
linearRowFilter_caller<5 , T, D, BrdRowReflect>,
linearRowFilter_caller<6 , T, D, BrdRowReflect>,
linearRowFilter_caller<7 , T, D, BrdRowReflect>,
linearRowFilter_caller<8 , T, D, BrdRowReflect>,
linearRowFilter_caller<9 , T, D, BrdRowReflect>,
linearRowFilter_caller<10, T, D, BrdRowReflect>,
linearRowFilter_caller<11, T, D, BrdRowReflect>,
linearRowFilter_caller<12, T, D, BrdRowReflect>,
linearRowFilter_caller<13, T, D, BrdRowReflect>,
linearRowFilter_caller<14, T, D, BrdRowReflect>,
linearRowFilter_caller<15, T, D, BrdRowReflect>,
linearRowFilter_caller<16, T, D, BrdRowReflect>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowWrap>,
linearRowFilter_caller<2 , T, D, BrdRowWrap>,
linearRowFilter_caller<3 , T, D, BrdRowWrap>,
linearRowFilter_caller<4 , T, D, BrdRowWrap>,
linearRowFilter_caller<5 , T, D, BrdRowWrap>,
linearRowFilter_caller<6 , T, D, BrdRowWrap>,
linearRowFilter_caller<7 , T, D, BrdRowWrap>,
linearRowFilter_caller<8 , T, D, BrdRowWrap>,
linearRowFilter_caller<9 , T, D, BrdRowWrap>,
linearRowFilter_caller<10, T, D, BrdRowWrap>,
linearRowFilter_caller<11, T, D, BrdRowWrap>,
linearRowFilter_caller<12, T, D, BrdRowWrap>,
linearRowFilter_caller<13, T, D, BrdRowWrap>,
linearRowFilter_caller<14, T, D, BrdRowWrap>,
linearRowFilter_caller<15, T, D, BrdRowWrap>,
linearRowFilter_caller<16, T, D, BrdRowWrap>
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect101>,
linearRowFilter_caller<2 , T, D, BrdRowReflect101>,
linearRowFilter_caller<3 , T, D, BrdRowReflect101>,
linearRowFilter_caller<4 , T, D, BrdRowReflect101>,
linearRowFilter_caller<5 , T, D, BrdRowReflect101>,
linearRowFilter_caller<6 , T, D, BrdRowReflect101>,
linearRowFilter_caller<7 , T, D, BrdRowReflect101>,
linearRowFilter_caller<8 , T, D, BrdRowReflect101>,
linearRowFilter_caller<9 , T, D, BrdRowReflect101>,
linearRowFilter_caller<10, T, D, BrdRowReflect101>,
linearRowFilter_caller<11, T, D, BrdRowReflect101>,
linearRowFilter_caller<12, T, D, BrdRowReflect101>,
linearRowFilter_caller<13, T, D, BrdRowReflect101>,
linearRowFilter_caller<14, T, D, BrdRowReflect101>,
linearRowFilter_caller<15, T, D, BrdRowReflect101>,
linearRowFilter_caller<16, T, D, BrdRowReflect101>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReplicate>,
linearRowFilter_caller<2 , T, D, BrdRowReplicate>,
linearRowFilter_caller<3 , T, D, BrdRowReplicate>,
linearRowFilter_caller<4 , T, D, BrdRowReplicate>,
linearRowFilter_caller<5 , T, D, BrdRowReplicate>,
linearRowFilter_caller<6 , T, D, BrdRowReplicate>,
linearRowFilter_caller<7 , T, D, BrdRowReplicate>,
linearRowFilter_caller<8 , T, D, BrdRowReplicate>,
linearRowFilter_caller<9 , T, D, BrdRowReplicate>,
linearRowFilter_caller<10, T, D, BrdRowReplicate>,
linearRowFilter_caller<11, T, D, BrdRowReplicate>,
linearRowFilter_caller<12, T, D, BrdRowReplicate>,
linearRowFilter_caller<13, T, D, BrdRowReplicate>,
linearRowFilter_caller<14, T, D, BrdRowReplicate>,
linearRowFilter_caller<15, T, D, BrdRowReplicate>,
linearRowFilter_caller<16, T, D, BrdRowReplicate>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowConstant>,
linearRowFilter_caller<2 , T, D, BrdRowConstant>,
linearRowFilter_caller<3 , T, D, BrdRowConstant>,
linearRowFilter_caller<4 , T, D, BrdRowConstant>,
linearRowFilter_caller<5 , T, D, BrdRowConstant>,
linearRowFilter_caller<6 , T, D, BrdRowConstant>,
linearRowFilter_caller<7 , T, D, BrdRowConstant>,
linearRowFilter_caller<8 , T, D, BrdRowConstant>,
linearRowFilter_caller<9 , T, D, BrdRowConstant>,
linearRowFilter_caller<10, T, D, BrdRowConstant>,
linearRowFilter_caller<11, T, D, BrdRowConstant>,
linearRowFilter_caller<12, T, D, BrdRowConstant>,
linearRowFilter_caller<13, T, D, BrdRowConstant>,
linearRowFilter_caller<14, T, D, BrdRowConstant>,
linearRowFilter_caller<15, T, D, BrdRowConstant>,
linearRowFilter_caller<16, T, D, BrdRowConstant>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect>,
linearRowFilter_caller<2 , T, D, BrdRowReflect>,
linearRowFilter_caller<3 , T, D, BrdRowReflect>,
linearRowFilter_caller<4 , T, D, BrdRowReflect>,
linearRowFilter_caller<5 , T, D, BrdRowReflect>,
linearRowFilter_caller<6 , T, D, BrdRowReflect>,
linearRowFilter_caller<7 , T, D, BrdRowReflect>,
linearRowFilter_caller<8 , T, D, BrdRowReflect>,
linearRowFilter_caller<9 , T, D, BrdRowReflect>,
linearRowFilter_caller<10, T, D, BrdRowReflect>,
linearRowFilter_caller<11, T, D, BrdRowReflect>,
linearRowFilter_caller<12, T, D, BrdRowReflect>,
linearRowFilter_caller<13, T, D, BrdRowReflect>,
linearRowFilter_caller<14, T, D, BrdRowReflect>,
linearRowFilter_caller<15, T, D, BrdRowReflect>,
linearRowFilter_caller<16, T, D, BrdRowReflect>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowWrap>,
linearRowFilter_caller<2 , T, D, BrdRowWrap>,
linearRowFilter_caller<3 , T, D, BrdRowWrap>,
linearRowFilter_caller<4 , T, D, BrdRowWrap>,
linearRowFilter_caller<5 , T, D, BrdRowWrap>,
linearRowFilter_caller<6 , T, D, BrdRowWrap>,
linearRowFilter_caller<7 , T, D, BrdRowWrap>,
linearRowFilter_caller<8 , T, D, BrdRowWrap>,
linearRowFilter_caller<9 , T, D, BrdRowWrap>,
linearRowFilter_caller<10, T, D, BrdRowWrap>,
linearRowFilter_caller<11, T, D, BrdRowWrap>,
linearRowFilter_caller<12, T, D, BrdRowWrap>,
linearRowFilter_caller<13, T, D, BrdRowWrap>,
linearRowFilter_caller<14, T, D, BrdRowWrap>,
linearRowFilter_caller<15, T, D, BrdRowWrap>,
linearRowFilter_caller<16, T, D, BrdRowWrap>
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
template void linearRowFilter_gpu<uchar , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<uchar4, float4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short2, float2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<short3, float3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<int , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
} // namespace row_filter
END_OPENCV_DEVICE_NAMESPACE
template void linearRowFilter_gpu<uchar , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<uchar4, float4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short2, float2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<short3, float3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<int , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
} // namespace row_filter
}}} // namespace cv { namespace gpu { namespace device

63
modules/gpu/src/cuda/safe_call.hpp
Unescape View File

@ -62,44 +62,43 @@
#define cublasSafeCall(expr) ___cublasSafeCall(expr, __FILE__, __LINE__)
#endif
namespace cv { namespace gpu {
void error(const char *error_string, const char *file, const int line, const char *func = "");
void nppError(int err, const char *file, const int line, const char *func = "");
void ncvError(int err, const char *file, const int line, const char *func = "");
void cufftError(int err, const char *file, const int line, const char *func = "");
void cublasError(int err, const char *file, const int line, const char *func = "");
static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int line, const char *func = "")
namespace cv { namespace gpu
{
if (cudaSuccess != err)
cv::gpu::error(cudaGetErrorString(err), file, line, func);
}
void error(const char *error_string, const char *file, const int line, const char *func = "");
void nppError(int err, const char *file, const int line, const char *func = "");
void ncvError(int err, const char *file, const int line, const char *func = "");
void cufftError(int err, const char *file, const int line, const char *func = "");
void cublasError(int err, const char *file, const int line, const char *func = "");
static inline void ___nppSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (err < 0)
cv::gpu::nppError(err, file, line, func);
}
static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int line, const char *func = "")
{
if (cudaSuccess != err)
cv::gpu::error(cudaGetErrorString(err), file, line, func);
}
static inline void ___ncvSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (NCV_SUCCESS != err)
cv::gpu::ncvError(err, file, line, func);
}
static inline void ___nppSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (err < 0)
cv::gpu::nppError(err, file, line, func);
}
static inline void ___cufftSafeCall(cufftResult_t err, const char *file, const int line, const char *func = "")
{
if (CUFFT_SUCCESS != err)
cv::gpu::cufftError(err, file, line, func);
}
static inline void ___ncvSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (NCV_SUCCESS != err)
cv::gpu::ncvError(err, file, line, func);
}
static inline void ___cublasSafeCall(cublasStatus_t err, const char *file, const int line, const char *func = "")
{
if (CUBLAS_STATUS_SUCCESS != err)
cv::gpu::cublasError(err, file, line, func);
}
static inline void ___cufftSafeCall(cufftResult_t err, const char *file, const int line, const char *func = "")
{
if (CUFFT_SUCCESS != err)
cv::gpu::cufftError(err, file, line, func);
}
static inline void ___cublasSafeCall(cublasStatus_t err, const char *file, const int line, const char *func = "")
{
if (CUBLAS_STATUS_SUCCESS != err)
cv::gpu::cublasError(err, file, line, func);
}
}}
#endif /* __OPENCV_CUDA_SAFE_CALL_HPP__ */

922
modules/gpu/src/cuda/split_merge.cu
Unescape View File

@ -42,467 +42,465 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace split_merge {
template <typename T, size_t elem_size = sizeof(T)>
struct TypeTraits
{
typedef T type;
typedef T type2;
typedef T type3;
typedef T type4;
};
template <typename T>
struct TypeTraits<T, 1>
{
typedef char type;
typedef char2 type2;
typedef char3 type3;
typedef char4 type4;
};
template <typename T>
struct TypeTraits<T, 2>
{
typedef short type;
typedef short2 type2;
typedef short3 type3;
typedef short4 type4;
};
template <typename T>
struct TypeTraits<T, 4>
{
typedef int type;
typedef int2 type2;
typedef int3 type3;
typedef int4 type4;
};
template <typename T>
struct TypeTraits<T, 8>
{
typedef double type;
typedef double2 type2;
//typedef double3 type3;
//typedef double4 type3;
};
typedef void (*MergeFunction)(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream);
typedef void (*SplitFunction)(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream);
//------------------------------------------------------------
// Merge
template <typename T>
__global__ void mergeC2_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type2 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_y[x] = dst_elem;
}
}
template <typename T>
__global__ void mergeC3_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type3 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC3_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
double* dst_y = (double*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[3 * x] = src0_y[x];
dst_y[3 * x + 1] = src1_y[x];
dst_y[3 * x + 2] = src2_y[x];
}
}
template <typename T>
__global__ void mergeC4_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type4 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
const T* src3_y = (const T*)(src3 + y * src3_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_elem.w = src3_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC4_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
const double* src3_y = (const double*)(src3 + y * src3_step);
double2* dst_y = (double2*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[2 * x] = make_double2(src0_y[x], src1_y[x]);
dst_y[2 * x + 1] = make_double2(src2_y[x], src3_y[x]);
}
}
template <typename T>
static void mergeC2_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC2_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC3_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC3_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC4_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC4_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
src[3].data, src[3].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
void merge_caller(const DevMem2Db* src, DevMem2Db& dst,
int total_channels, size_t elem_size,
const cudaStream_t& stream)
{
static MergeFunction merge_func_tbl[] =
{
mergeC2_<char>, mergeC2_<short>, mergeC2_<int>, 0, mergeC2_<double>,
mergeC3_<char>, mergeC3_<short>, mergeC3_<int>, 0, mergeC3_<double>,
mergeC4_<char>, mergeC4_<short>, mergeC4_<int>, 0, mergeC4_<double>,
};
size_t merge_func_id = (total_channels - 2) * 5 + (elem_size >> 1);
MergeFunction merge_func = merge_func_tbl[merge_func_id];
if (merge_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
merge_func(src, dst, stream);
}
//------------------------------------------------------------
// Split
template <typename T>
__global__ void splitC2_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step)
{
typedef typename TypeTraits<T>::type2 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
}
}
template <typename T>
__global__ void splitC3_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
namespace cv { namespace gpu { namespace device
{
typedef typename TypeTraits<T>::type3 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
if (x < cols && y < rows)
namespace split_merge
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
}
}
template <>
__global__ void splitC3_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src_y = (const double*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
if (x < cols && y < rows)
{
dst0_y[x] = src_y[3 * x];
dst1_y[x] = src_y[3 * x + 1];
dst2_y[x] = src_y[3 * x + 2];
}
}
template <typename T>
__global__ void splitC4_(const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
typedef typename TypeTraits<T>::type4 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
T* dst3_y = (T*)(dst3 + y * dst3_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
dst3_y[x] = src_elem.w;
}
}
template <>
__global__ void splitC4_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double2* src_y = (const double2*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
double* dst3_y = (double*)(dst3 + y * dst3_step);
if (x < cols && y < rows)
{
double2 src_elem1 = src_y[2 * x];
double2 src_elem2 = src_y[2 * x + 1];
dst0_y[x] = src_elem1.x;
dst1_y[x] = src_elem1.y;
dst2_y[x] = src_elem2.x;
dst3_y[x] = src_elem2.y;
}
}
template <typename T>
static void splitC2_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC2_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void splitC3_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC3_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void splitC4_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC4_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step,
dst[3].data, dst[3].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
void split_caller(const DevMem2Db& src, DevMem2Db* dst, int num_channels, size_t elem_size1, const cudaStream_t& stream)
{
static SplitFunction split_func_tbl[] =
{
splitC2_<char>, splitC2_<short>, splitC2_<int>, 0, splitC2_<double>,
splitC3_<char>, splitC3_<short>, splitC3_<int>, 0, splitC3_<double>,
splitC4_<char>, splitC4_<short>, splitC4_<int>, 0, splitC4_<double>,
};
size_t split_func_id = (num_channels - 2) * 5 + (elem_size1 >> 1);
SplitFunction split_func = split_func_tbl[split_func_id];
if (split_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
split_func(src, dst, stream);
}
} // namespace split_merge
END_OPENCV_DEVICE_NAMESPACE
template <typename T, size_t elem_size = sizeof(T)>
struct TypeTraits
{
typedef T type;
typedef T type2;
typedef T type3;
typedef T type4;
};
template <typename T>
struct TypeTraits<T, 1>
{
typedef char type;
typedef char2 type2;
typedef char3 type3;
typedef char4 type4;
};
template <typename T>
struct TypeTraits<T, 2>
{
typedef short type;
typedef short2 type2;
typedef short3 type3;
typedef short4 type4;
};
template <typename T>
struct TypeTraits<T, 4>
{
typedef int type;
typedef int2 type2;
typedef int3 type3;
typedef int4 type4;
};
template <typename T>
struct TypeTraits<T, 8>
{
typedef double type;
typedef double2 type2;
//typedef double3 type3;
//typedef double4 type3;
};
typedef void (*MergeFunction)(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream);
typedef void (*SplitFunction)(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream);
//------------------------------------------------------------
// Merge
template <typename T>
__global__ void mergeC2_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type2 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_y[x] = dst_elem;
}
}
template <typename T>
__global__ void mergeC3_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type3 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC3_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
double* dst_y = (double*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[3 * x] = src0_y[x];
dst_y[3 * x + 1] = src1_y[x];
dst_y[3 * x + 2] = src2_y[x];
}
}
template <typename T>
__global__ void mergeC4_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type4 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
const T* src3_y = (const T*)(src3 + y * src3_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_elem.w = src3_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC4_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
const double* src3_y = (const double*)(src3 + y * src3_step);
double2* dst_y = (double2*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[2 * x] = make_double2(src0_y[x], src1_y[x]);
dst_y[2 * x + 1] = make_double2(src2_y[x], src3_y[x]);
}
}
template <typename T>
static void mergeC2_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC2_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC3_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC3_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC4_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC4_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
src[3].data, src[3].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
void merge_caller(const DevMem2Db* src, DevMem2Db& dst,
int total_channels, size_t elem_size,
const cudaStream_t& stream)
{
static MergeFunction merge_func_tbl[] =
{
mergeC2_<char>, mergeC2_<short>, mergeC2_<int>, 0, mergeC2_<double>,
mergeC3_<char>, mergeC3_<short>, mergeC3_<int>, 0, mergeC3_<double>,
mergeC4_<char>, mergeC4_<short>, mergeC4_<int>, 0, mergeC4_<double>,
};
size_t merge_func_id = (total_channels - 2) * 5 + (elem_size >> 1);
MergeFunction merge_func = merge_func_tbl[merge_func_id];
if (merge_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
merge_func(src, dst, stream);
}
//------------------------------------------------------------
// Split
template <typename T>
__global__ void splitC2_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step)
{
typedef typename TypeTraits<T>::type2 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
}
}
template <typename T>
__global__ void splitC3_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
{
typedef typename TypeTraits<T>::type3 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
}
}
template <>
__global__ void splitC3_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src_y = (const double*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
if (x < cols && y < rows)
{
dst0_y[x] = src_y[3 * x];
dst1_y[x] = src_y[3 * x + 1];
dst2_y[x] = src_y[3 * x + 2];
}
}
template <typename T>
__global__ void splitC4_(const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
typedef typename TypeTraits<T>::type4 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
T* dst3_y = (T*)(dst3 + y * dst3_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
dst3_y[x] = src_elem.w;
}
}
template <>
__global__ void splitC4_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double2* src_y = (const double2*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
double* dst3_y = (double*)(dst3 + y * dst3_step);
if (x < cols && y < rows)
{
double2 src_elem1 = src_y[2 * x];
double2 src_elem2 = src_y[2 * x + 1];
dst0_y[x] = src_elem1.x;
dst1_y[x] = src_elem1.y;
dst2_y[x] = src_elem2.x;
dst3_y[x] = src_elem2.y;
}
}
template <typename T>
static void splitC2_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC2_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void splitC3_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC3_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void splitC4_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC4_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step,
dst[3].data, dst[3].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
void split_caller(const DevMem2Db& src, DevMem2Db* dst, int num_channels, size_t elem_size1, const cudaStream_t& stream)
{
static SplitFunction split_func_tbl[] =
{
splitC2_<char>, splitC2_<short>, splitC2_<int>, 0, splitC2_<double>,
splitC3_<char>, splitC3_<short>, splitC3_<int>, 0, splitC3_<double>,
splitC4_<char>, splitC4_<short>, splitC4_<int>, 0, splitC4_<double>,
};
size_t split_func_id = (num_channels - 2) * 5 + (elem_size1 >> 1);
SplitFunction split_func = split_func_tbl[split_func_id];
if (split_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
split_func(src, dst, stream);
}
} // namespace split_merge
}}} // namespace cv { namespace gpu { namespace device

800
modules/gpu/src/cuda/stereobm.cu
Unescape View File

@ -42,496 +42,494 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace stereobm {
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Stereo BM ////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
namespace cv { namespace gpu { namespace device
{
namespace stereobm
{
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Stereo BM ////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
#define ROWSperTHREAD 21 // the number of rows a thread will process
#define ROWSperTHREAD 21 // the number of rows a thread will process
#define BLOCK_W 128 // the thread block width (464)
#define N_DISPARITIES 8
#define BLOCK_W 128 // the thread block width (464)
#define N_DISPARITIES 8
#define STEREO_MIND 0 // The minimum d range to check
#define STEREO_DISP_STEP N_DISPARITIES // the d step, must be <= 1 to avoid aliasing
#define STEREO_MIND 0 // The minimum d range to check
#define STEREO_DISP_STEP N_DISPARITIES // the d step, must be <= 1 to avoid aliasing
__constant__ unsigned int* cminSSDImage;
__constant__ size_t cminSSD_step;
__constant__ int cwidth;
__constant__ int cheight;
__constant__ unsigned int* cminSSDImage;
__constant__ size_t cminSSD_step;
__constant__ int cwidth;
__constant__ int cheight;
__device__ __forceinline__ int SQ(int a)
{
return a * a;
}
__device__ __forceinline__ int SQ(int a)
{
return a * a;
}
template<int RADIUS>
__device__ unsigned int CalcSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int cache = 0;
unsigned int cache2 = 0;
template<int RADIUS>
__device__ unsigned int CalcSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int cache = 0;
unsigned int cache2 = 0;
for(int i = 1; i <= RADIUS; i++)
cache += col_ssd[i];
for(int i = 1; i <= RADIUS; i++)
cache += col_ssd[i];
col_ssd_cache[0] = cache;
col_ssd_cache[0] = cache;
__syncthreads();
__syncthreads();
if (threadIdx.x < BLOCK_W - RADIUS)
cache2 = col_ssd_cache[RADIUS];
else
for(int i = RADIUS + 1; i < (2 * RADIUS + 1); i++)
cache2 += col_ssd[i];
if (threadIdx.x < BLOCK_W - RADIUS)
cache2 = col_ssd_cache[RADIUS];
else
for(int i = RADIUS + 1; i < (2 * RADIUS + 1); i++)
cache2 += col_ssd[i];
return col_ssd[0] + cache + cache2;
}
return col_ssd[0] + cache + cache2;
}
template<int RADIUS>
__device__ uint2 MinSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int ssd[N_DISPARITIES];
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
ssd[0] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 0 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[1] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 1 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[2] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 2 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[3] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 3 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[4] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 4 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[5] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 5 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[6] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 6 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[7] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 7 * (BLOCK_W + 2 * RADIUS));
int mssd = ::min(::min(::min(ssd[0], ssd[1]), ::min(ssd[4], ssd[5])), ::min(::min(ssd[2], ssd[3]), ::min(ssd[6], ssd[7])));
int bestIdx = 0;
for (int i = 0; i < N_DISPARITIES; i++)
{
if (mssd == ssd[i])
bestIdx = i;
}
template<int RADIUS>
__device__ uint2 MinSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int ssd[N_DISPARITIES];
return make_uint2(mssd, bestIdx);
}
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
ssd[0] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 0 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[1] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 1 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[2] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 2 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[3] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 3 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[4] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 4 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[5] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 5 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[6] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 6 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[7] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 7 * (BLOCK_W + 2 * RADIUS));
template<int RADIUS>
__device__ void StepDown(int idx1, int idx2, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
unsigned char leftPixel1;
unsigned char leftPixel2;
unsigned char rightPixel1[8];
unsigned char rightPixel2[8];
unsigned int diff1, diff2;
leftPixel1 = imageL[idx1];
leftPixel2 = imageL[idx2];
idx1 = idx1 - d;
idx2 = idx2 - d;
rightPixel1[7] = imageR[idx1 - 7];
rightPixel1[0] = imageR[idx1 - 0];
rightPixel1[1] = imageR[idx1 - 1];
rightPixel1[2] = imageR[idx1 - 2];
rightPixel1[3] = imageR[idx1 - 3];
rightPixel1[4] = imageR[idx1 - 4];
rightPixel1[5] = imageR[idx1 - 5];
rightPixel1[6] = imageR[idx1 - 6];
rightPixel2[7] = imageR[idx2 - 7];
rightPixel2[0] = imageR[idx2 - 0];
rightPixel2[1] = imageR[idx2 - 1];
rightPixel2[2] = imageR[idx2 - 2];
rightPixel2[3] = imageR[idx2 - 3];
rightPixel2[4] = imageR[idx2 - 4];
rightPixel2[5] = imageR[idx2 - 5];
rightPixel2[6] = imageR[idx2 - 6];
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
diff1 = leftPixel1 - rightPixel1[0];
diff2 = leftPixel2 - rightPixel2[0];
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[1];
diff2 = leftPixel2 - rightPixel2[1];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[2];
diff2 = leftPixel2 - rightPixel2[2];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[3];
diff2 = leftPixel2 - rightPixel2[3];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[4];
diff2 = leftPixel2 - rightPixel2[4];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[5];
diff2 = leftPixel2 - rightPixel2[5];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[6];
diff2 = leftPixel2 - rightPixel2[6];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[7];
diff2 = leftPixel2 - rightPixel2[7];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
}
template<int RADIUS>
__device__ void InitColSSD(int x_tex, int y_tex, int im_pitch, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
unsigned char leftPixel1;
int idx;
unsigned int diffa[] = {0, 0, 0, 0, 0, 0, 0, 0};
int mssd = ::min(::min(::min(ssd[0], ssd[1]), ::min(ssd[4], ssd[5])), ::min(::min(ssd[2], ssd[3]), ::min(ssd[6], ssd[7])));
for(int i = 0; i < (2 * RADIUS + 1); i++)
{
idx = y_tex * im_pitch + x_tex;
leftPixel1 = imageL[idx];
idx = idx - d;
diffa[0] += SQ(leftPixel1 - imageR[idx - 0]);
diffa[1] += SQ(leftPixel1 - imageR[idx - 1]);
diffa[2] += SQ(leftPixel1 - imageR[idx - 2]);
diffa[3] += SQ(leftPixel1 - imageR[idx - 3]);
diffa[4] += SQ(leftPixel1 - imageR[idx - 4]);
diffa[5] += SQ(leftPixel1 - imageR[idx - 5]);
diffa[6] += SQ(leftPixel1 - imageR[idx - 6]);
diffa[7] += SQ(leftPixel1 - imageR[idx - 7]);
y_tex += 1;
}
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] = diffa[0];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] = diffa[1];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] = diffa[2];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] = diffa[3];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] = diffa[4];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] = diffa[5];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] = diffa[6];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] = diffa[7];
}
template<int RADIUS>
__global__ void stereoKernel(unsigned char *left, unsigned char *right, size_t img_step, PtrStepb disp, int maxdisp)
{
extern __shared__ unsigned int col_ssd_cache[];
volatile unsigned int *col_ssd = col_ssd_cache + BLOCK_W + threadIdx.x;
volatile unsigned int *col_ssd_extra = threadIdx.x < (2 * RADIUS) ? col_ssd + BLOCK_W : 0; //#define N_DIRTY_PIXELS (2 * RADIUS)
//#define X (blockIdx.x * BLOCK_W + threadIdx.x + STEREO_MAXD)
int X = (blockIdx.x * BLOCK_W + threadIdx.x + maxdisp + RADIUS);
//#define Y (__mul24(blockIdx.y, ROWSperTHREAD) + RADIUS)
#define Y (blockIdx.y * ROWSperTHREAD + RADIUS)
//int Y = blockIdx.y * ROWSperTHREAD + RADIUS;
unsigned int* minSSDImage = cminSSDImage + X + Y * cminSSD_step;
unsigned char* disparImage = disp.data + X + Y * disp.step;
/* if (X < cwidth)
{
unsigned int *minSSDImage_end = minSSDImage + min(ROWSperTHREAD, cheight - Y) * minssd_step;
for(uint *ptr = minSSDImage; ptr != minSSDImage_end; ptr += minssd_step )
*ptr = 0xFFFFFFFF;
}*/
int end_row = ::min(ROWSperTHREAD, cheight - Y - RADIUS);
int y_tex;
int x_tex = X - RADIUS;
if (x_tex >= cwidth)
return;
for(int d = STEREO_MIND; d < maxdisp; d += STEREO_DISP_STEP)
{
y_tex = Y - RADIUS;
InitColSSD<RADIUS>(x_tex, y_tex, img_step, left, right, d, col_ssd);
int bestIdx = 0;
for (int i = 0; i < N_DISPARITIES; i++)
{
if (mssd == ssd[i])
bestIdx = i;
}
if (col_ssd_extra > 0)
if (x_tex + BLOCK_W < cwidth)
InitColSSD<RADIUS>(x_tex + BLOCK_W, y_tex, img_step, left, right, d, col_ssd_extra);
return make_uint2(mssd, bestIdx);
}
__syncthreads(); //before MinSSD function
template<int RADIUS>
__device__ void StepDown(int idx1, int idx2, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
unsigned char leftPixel1;
unsigned char leftPixel2;
unsigned char rightPixel1[8];
unsigned char rightPixel2[8];
unsigned int diff1, diff2;
leftPixel1 = imageL[idx1];
leftPixel2 = imageL[idx2];
idx1 = idx1 - d;
idx2 = idx2 - d;
rightPixel1[7] = imageR[idx1 - 7];
rightPixel1[0] = imageR[idx1 - 0];
rightPixel1[1] = imageR[idx1 - 1];
rightPixel1[2] = imageR[idx1 - 2];
rightPixel1[3] = imageR[idx1 - 3];
rightPixel1[4] = imageR[idx1 - 4];
rightPixel1[5] = imageR[idx1 - 5];
rightPixel1[6] = imageR[idx1 - 6];
rightPixel2[7] = imageR[idx2 - 7];
rightPixel2[0] = imageR[idx2 - 0];
rightPixel2[1] = imageR[idx2 - 1];
rightPixel2[2] = imageR[idx2 - 2];
rightPixel2[3] = imageR[idx2 - 3];
rightPixel2[4] = imageR[idx2 - 4];
rightPixel2[5] = imageR[idx2 - 5];
rightPixel2[6] = imageR[idx2 - 6];
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
diff1 = leftPixel1 - rightPixel1[0];
diff2 = leftPixel2 - rightPixel2[0];
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[1];
diff2 = leftPixel2 - rightPixel2[1];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[2];
diff2 = leftPixel2 - rightPixel2[2];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[3];
diff2 = leftPixel2 - rightPixel2[3];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[4];
diff2 = leftPixel2 - rightPixel2[4];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[5];
diff2 = leftPixel2 - rightPixel2[5];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[6];
diff2 = leftPixel2 - rightPixel2[6];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[7];
diff2 = leftPixel2 - rightPixel2[7];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
}
if (X < cwidth - RADIUS && Y < cheight - RADIUS)
template<int RADIUS>
__device__ void InitColSSD(int x_tex, int y_tex, int im_pitch, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[0])
unsigned char leftPixel1;
int idx;
unsigned int diffa[] = {0, 0, 0, 0, 0, 0, 0, 0};
for(int i = 0; i < (2 * RADIUS + 1); i++)
{
disparImage[0] = (unsigned char)(d + minSSD.y);
minSSDImage[0] = minSSD.x;
idx = y_tex * im_pitch + x_tex;
leftPixel1 = imageL[idx];
idx = idx - d;
diffa[0] += SQ(leftPixel1 - imageR[idx - 0]);
diffa[1] += SQ(leftPixel1 - imageR[idx - 1]);
diffa[2] += SQ(leftPixel1 - imageR[idx - 2]);
diffa[3] += SQ(leftPixel1 - imageR[idx - 3]);
diffa[4] += SQ(leftPixel1 - imageR[idx - 4]);
diffa[5] += SQ(leftPixel1 - imageR[idx - 5]);
diffa[6] += SQ(leftPixel1 - imageR[idx - 6]);
diffa[7] += SQ(leftPixel1 - imageR[idx - 7]);
y_tex += 1;
}
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] = diffa[0];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] = diffa[1];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] = diffa[2];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] = diffa[3];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] = diffa[4];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] = diffa[5];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] = diffa[6];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] = diffa[7];
}
for(int row = 1; row < end_row; row++)
template<int RADIUS>
__global__ void stereoKernel(unsigned char *left, unsigned char *right, size_t img_step, PtrStepb disp, int maxdisp)
{
int idx1 = y_tex * img_step + x_tex;
int idx2 = (y_tex + (2 * RADIUS + 1)) * img_step + x_tex;
__syncthreads();
StepDown<RADIUS>(idx1, idx2, left, right, d, col_ssd);
extern __shared__ unsigned int col_ssd_cache[];
volatile unsigned int *col_ssd = col_ssd_cache + BLOCK_W + threadIdx.x;
volatile unsigned int *col_ssd_extra = threadIdx.x < (2 * RADIUS) ? col_ssd + BLOCK_W : 0; //#define N_DIRTY_PIXELS (2 * RADIUS)
//#define X (blockIdx.x * BLOCK_W + threadIdx.x + STEREO_MAXD)
int X = (blockIdx.x * BLOCK_W + threadIdx.x + maxdisp + RADIUS);
//#define Y (__mul24(blockIdx.y, ROWSperTHREAD) + RADIUS)
#define Y (blockIdx.y * ROWSperTHREAD + RADIUS)
//int Y = blockIdx.y * ROWSperTHREAD + RADIUS;
unsigned int* minSSDImage = cminSSDImage + X + Y * cminSSD_step;
unsigned char* disparImage = disp.data + X + Y * disp.step;
/* if (X < cwidth)
{
unsigned int *minSSDImage_end = minSSDImage + min(ROWSperTHREAD, cheight - Y) * minssd_step;
for(uint *ptr = minSSDImage; ptr != minSSDImage_end; ptr += minssd_step )
*ptr = 0xFFFFFFFF;
}*/
int end_row = ::min(ROWSperTHREAD, cheight - Y - RADIUS);
int y_tex;
int x_tex = X - RADIUS;
if (x_tex >= cwidth)
return;
for(int d = STEREO_MIND; d < maxdisp; d += STEREO_DISP_STEP)
{
y_tex = Y - RADIUS;
if (col_ssd_extra)
if (x_tex + BLOCK_W < cwidth)
StepDown<RADIUS>(idx1, idx2, left + BLOCK_W, right + BLOCK_W, d, col_ssd_extra);
InitColSSD<RADIUS>(x_tex, y_tex, img_step, left, right, d, col_ssd);
y_tex += 1;
if (col_ssd_extra > 0)
if (x_tex + BLOCK_W < cwidth)
InitColSSD<RADIUS>(x_tex + BLOCK_W, y_tex, img_step, left, right, d, col_ssd_extra);
__syncthreads(); //before MinSSD function
__syncthreads(); //before MinSSD function
if (X < cwidth - RADIUS && row < cheight - RADIUS - Y)
{
int idx = row * cminSSD_step;
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[idx])
if (X < cwidth - RADIUS && Y < cheight - RADIUS)
{
disparImage[disp.step * row] = (unsigned char)(d + minSSD.y);
minSSDImage[idx] = minSSD.x;
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[0])
{
disparImage[0] = (unsigned char)(d + minSSD.y);
minSSDImage[0] = minSSD.x;
}
}
}
} // for row loop
} // for d loop
}
for(int row = 1; row < end_row; row++)
{
int idx1 = y_tex * img_step + x_tex;
int idx2 = (y_tex + (2 * RADIUS + 1)) * img_step + x_tex;
template<int RADIUS> void kernel_caller(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream)
{
dim3 grid(1,1,1);
dim3 threads(BLOCK_W, 1, 1);
__syncthreads();
grid.x = divUp(left.cols - maxdisp - 2 * RADIUS, BLOCK_W);
grid.y = divUp(left.rows - 2 * RADIUS, ROWSperTHREAD);
StepDown<RADIUS>(idx1, idx2, left, right, d, col_ssd);
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
size_t smem_size = (BLOCK_W + N_DISPARITIES * (BLOCK_W + 2 * RADIUS)) * sizeof(unsigned int);
if (col_ssd_extra)
if (x_tex + BLOCK_W < cwidth)
StepDown<RADIUS>(idx1, idx2, left + BLOCK_W, right + BLOCK_W, d, col_ssd_extra);
stereoKernel<RADIUS><<<grid, threads, smem_size, stream>>>(left.data, right.data, left.step, disp, maxdisp);
cudaSafeCall( cudaGetLastError() );
y_tex += 1;
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
};
__syncthreads(); //before MinSSD function
typedef void (*kernel_caller_t)(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream);
if (X < cwidth - RADIUS && row < cheight - RADIUS - Y)
{
int idx = row * cminSSD_step;
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[idx])
{
disparImage[disp.step * row] = (unsigned char)(d + minSSD.y);
minSSDImage[idx] = minSSD.x;
}
}
} // for row loop
} // for d loop
}
const static kernel_caller_t callers[] =
{
0,
kernel_caller< 1>, kernel_caller< 2>, kernel_caller< 3>, kernel_caller< 4>, kernel_caller< 5>,
kernel_caller< 6>, kernel_caller< 7>, kernel_caller< 8>, kernel_caller< 9>, kernel_caller<10>,
kernel_caller<11>, kernel_caller<12>, kernel_caller<13>, kernel_caller<15>, kernel_caller<15>,
kernel_caller<16>, kernel_caller<17>, kernel_caller<18>, kernel_caller<19>, kernel_caller<20>,
kernel_caller<21>, kernel_caller<22>, kernel_caller<23>, kernel_caller<24>, kernel_caller<25>
//0,0,0, 0,0,0, 0,0,kernel_caller<9>
};
const int calles_num = sizeof(callers)/sizeof(callers[0]);
void stereoBM_GPU(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, int winsz, const DevMem2D_<unsigned int>& minSSD_buf, cudaStream_t& stream)
{
int winsz2 = winsz >> 1;
if (winsz2 == 0 || winsz2 >= calles_num)
cv::gpu::error("Unsupported window size", __FILE__, __LINE__);
template<int RADIUS> void kernel_caller(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream)
{
dim3 grid(1,1,1);
dim3 threads(BLOCK_W, 1, 1);
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferL1) );
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferShared) );
grid.x = divUp(left.cols - maxdisp - 2 * RADIUS, BLOCK_W);
grid.y = divUp(left.rows - 2 * RADIUS, ROWSperTHREAD);
cudaSafeCall( cudaMemset2D(disp.data, disp.step, 0, disp.cols, disp.rows) );
cudaSafeCall( cudaMemset2D(minSSD_buf.data, minSSD_buf.step, 0xFF, minSSD_buf.cols * minSSD_buf.elemSize(), disp.rows) );
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
size_t smem_size = (BLOCK_W + N_DISPARITIES * (BLOCK_W + 2 * RADIUS)) * sizeof(unsigned int);
cudaSafeCall( cudaMemcpyToSymbol( cwidth, &left.cols, sizeof(left.cols) ) );
cudaSafeCall( cudaMemcpyToSymbol( cheight, &left.rows, sizeof(left.rows) ) );
cudaSafeCall( cudaMemcpyToSymbol( cminSSDImage, &minSSD_buf.data, sizeof(minSSD_buf.data) ) );
stereoKernel<RADIUS><<<grid, threads, smem_size, stream>>>(left.data, right.data, left.step, disp, maxdisp);
cudaSafeCall( cudaGetLastError() );
size_t minssd_step = minSSD_buf.step/minSSD_buf.elemSize();
cudaSafeCall( cudaMemcpyToSymbol( cminSSD_step, &minssd_step, sizeof(minssd_step) ) );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
};
callers[winsz2](left, right, disp, maxdisp, stream);
}
typedef void (*kernel_caller_t)(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream);
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Sobel Prefiler ///////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
const static kernel_caller_t callers[] =
{
0,
kernel_caller< 1>, kernel_caller< 2>, kernel_caller< 3>, kernel_caller< 4>, kernel_caller< 5>,
kernel_caller< 6>, kernel_caller< 7>, kernel_caller< 8>, kernel_caller< 9>, kernel_caller<10>,
kernel_caller<11>, kernel_caller<12>, kernel_caller<13>, kernel_caller<15>, kernel_caller<15>,
kernel_caller<16>, kernel_caller<17>, kernel_caller<18>, kernel_caller<19>, kernel_caller<20>,
kernel_caller<21>, kernel_caller<22>, kernel_caller<23>, kernel_caller<24>, kernel_caller<25>
//0,0,0, 0,0,0, 0,0,kernel_caller<9>
};
const int calles_num = sizeof(callers)/sizeof(callers[0]);
void stereoBM_GPU(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, int winsz, const DevMem2D_<unsigned int>& minSSD_buf, cudaStream_t& stream)
{
int winsz2 = winsz >> 1;
texture<unsigned char, 2, cudaReadModeElementType> texForSobel;
if (winsz2 == 0 || winsz2 >= calles_num)
cv::gpu::error("Unsupported window size", __FILE__, __LINE__);
__global__ void prefilter_kernel(DevMem2Db output, int prefilterCap)
{
int x = blockDim.x * blockIdx.x + threadIdx.x;
int y = blockDim.y * blockIdx.y + threadIdx.y;
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferL1) );
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferShared) );
if (x < output.cols && y < output.rows)
{
int conv = (int)tex2D(texForSobel, x - 1, y - 1) * (-1) + (int)tex2D(texForSobel, x + 1, y - 1) * (1) +
(int)tex2D(texForSobel, x - 1, y ) * (-2) + (int)tex2D(texForSobel, x + 1, y ) * (2) +
(int)tex2D(texForSobel, x - 1, y + 1) * (-1) + (int)tex2D(texForSobel, x + 1, y + 1) * (1);
cudaSafeCall( cudaMemset2D(disp.data, disp.step, 0, disp.cols, disp.rows) );
cudaSafeCall( cudaMemset2D(minSSD_buf.data, minSSD_buf.step, 0xFF, minSSD_buf.cols * minSSD_buf.elemSize(), disp.rows) );
cudaSafeCall( cudaMemcpyToSymbol( cwidth, &left.cols, sizeof(left.cols) ) );
cudaSafeCall( cudaMemcpyToSymbol( cheight, &left.rows, sizeof(left.rows) ) );
cudaSafeCall( cudaMemcpyToSymbol( cminSSDImage, &minSSD_buf.data, sizeof(minSSD_buf.data) ) );
conv = ::min(::min(::max(-prefilterCap, conv), prefilterCap) + prefilterCap, 255);
output.ptr(y)[x] = conv & 0xFF;
}
}
size_t minssd_step = minSSD_buf.step/minSSD_buf.elemSize();
cudaSafeCall( cudaMemcpyToSymbol( cminSSD_step, &minssd_step, sizeof(minssd_step) ) );
void prefilter_xsobel(const DevMem2Db& input, const DevMem2Db& output, int prefilterCap, cudaStream_t & stream)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForSobel, input.data, desc, input.cols, input.rows, input.step ) );
callers[winsz2](left, right, disp, maxdisp, stream);
}
dim3 threads(16, 16, 1);
dim3 grid(1, 1, 1);
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Sobel Prefiler ///////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, threads.y);
texture<unsigned char, 2, cudaReadModeElementType> texForSobel;
prefilter_kernel<<<grid, threads, 0, stream>>>(output, prefilterCap);
cudaSafeCall( cudaGetLastError() );
__global__ void prefilter_kernel(DevMem2Db output, int prefilterCap)
{
int x = blockDim.x * blockIdx.x + threadIdx.x;
int y = blockDim.y * blockIdx.y + threadIdx.y;
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
if (x < output.cols && y < output.rows)
{
int conv = (int)tex2D(texForSobel, x - 1, y - 1) * (-1) + (int)tex2D(texForSobel, x + 1, y - 1) * (1) +
(int)tex2D(texForSobel, x - 1, y ) * (-2) + (int)tex2D(texForSobel, x + 1, y ) * (2) +
(int)tex2D(texForSobel, x - 1, y + 1) * (-1) + (int)tex2D(texForSobel, x + 1, y + 1) * (1);
cudaSafeCall( cudaUnbindTexture (texForSobel ) );
}
conv = ::min(::min(::max(-prefilterCap, conv), prefilterCap) + prefilterCap, 255);
output.ptr(y)[x] = conv & 0xFF;
}
}
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////// Textureness filtering ////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
void prefilter_xsobel(const DevMem2Db& input, const DevMem2Db& output, int prefilterCap, cudaStream_t & stream)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForSobel, input.data, desc, input.cols, input.rows, input.step ) );
texture<unsigned char, 2, cudaReadModeNormalizedFloat> texForTF;
dim3 threads(16, 16, 1);
dim3 grid(1, 1, 1);
__device__ __forceinline__ float sobel(int x, int y)
{
float conv = tex2D(texForTF, x - 1, y - 1) * (-1) + tex2D(texForTF, x + 1, y - 1) * (1) +
tex2D(texForTF, x - 1, y ) * (-2) + tex2D(texForTF, x + 1, y ) * (2) +
tex2D(texForTF, x - 1, y + 1) * (-1) + tex2D(texForTF, x + 1, y + 1) * (1);
return fabs(conv);
}
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, threads.y);
__device__ float CalcSums(float *cols, float *cols_cache, int winsz)
{
float cache = 0;
float cache2 = 0;
int winsz2 = winsz/2;
prefilter_kernel<<<grid, threads, 0, stream>>>(output, prefilterCap);
cudaSafeCall( cudaGetLastError() );
for(int i = 1; i <= winsz2; i++)
cache += cols[i];
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
cols_cache[0] = cache;
cudaSafeCall( cudaUnbindTexture (texForSobel ) );
}
__syncthreads();
if (threadIdx.x < blockDim.x - winsz2)
cache2 = cols_cache[winsz2];
else
for(int i = winsz2 + 1; i < winsz; i++)
cache2 += cols[i];
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////// Textureness filtering ////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
return cols[0] + cache + cache2;
}
texture<unsigned char, 2, cudaReadModeNormalizedFloat> texForTF;
#define RpT (2 * ROWSperTHREAD) // got experimentally
__device__ __forceinline__ float sobel(int x, int y)
{
float conv = tex2D(texForTF, x - 1, y - 1) * (-1) + tex2D(texForTF, x + 1, y - 1) * (1) +
tex2D(texForTF, x - 1, y ) * (-2) + tex2D(texForTF, x + 1, y ) * (2) +
tex2D(texForTF, x - 1, y + 1) * (-1) + tex2D(texForTF, x + 1, y + 1) * (1);
return fabs(conv);
}
__global__ void textureness_kernel(DevMem2Db disp, int winsz, float threshold)
{
int winsz2 = winsz/2;
int n_dirty_pixels = (winsz2) * 2;
__device__ float CalcSums(float *cols, float *cols_cache, int winsz)
{
float cache = 0;
float cache2 = 0;
int winsz2 = winsz/2;
extern __shared__ float cols_cache[];
float *cols = cols_cache + blockDim.x + threadIdx.x;
float *cols_extra = threadIdx.x < n_dirty_pixels ? cols + blockDim.x : 0;
for(int i = 1; i <= winsz2; i++)
cache += cols[i];
int x = blockIdx.x * blockDim.x + threadIdx.x;
int beg_row = blockIdx.y * RpT;
int end_row = ::min(beg_row + RpT, disp.rows);
cols_cache[0] = cache;
if (x < disp.cols)
{
int y = beg_row;
__syncthreads();
float sum = 0;
float sum_extra = 0;
if (threadIdx.x < blockDim.x - winsz2)
cache2 = cols_cache[winsz2];
else
for(int i = winsz2 + 1; i < winsz; i++)
cache2 += cols[i];
for(int i = y - winsz2; i <= y + winsz2; ++i)
{
sum += sobel(x - winsz2, i);
if (cols_extra)
sum_extra += sobel(x + blockDim.x - winsz2, i);
return cols[0] + cache + cache2;
}
*cols = sum;
if (cols_extra)
*cols_extra = sum_extra;
__syncthreads();
#define RpT (2 * ROWSperTHREAD) // got experimentally
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
__global__ void textureness_kernel(DevMem2Db disp, int winsz, float threshold)
{
int winsz2 = winsz/2;
int n_dirty_pixels = (winsz2) * 2;
__syncthreads();
extern __shared__ float cols_cache[];
float *cols = cols_cache + blockDim.x + threadIdx.x;
float *cols_extra = threadIdx.x < n_dirty_pixels ? cols + blockDim.x : 0;
for(int y = beg_row + 1; y < end_row; ++y)
{
sum = sum - sobel(x - winsz2, y - winsz2 - 1) + sobel(x - winsz2, y + winsz2);
*cols = sum;
int x = blockIdx.x * blockDim.x + threadIdx.x;
int beg_row = blockIdx.y * RpT;
int end_row = ::min(beg_row + RpT, disp.rows);
if (cols_extra)
if (x < disp.cols)
{
sum_extra = sum_extra - sobel(x + blockDim.x - winsz2, y - winsz2 - 1) + sobel(x + blockDim.x - winsz2, y + winsz2);
*cols_extra = sum_extra;
}
int y = beg_row;
__syncthreads();
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
float sum = 0;
float sum_extra = 0;
__syncthreads();
}
}
}
for(int i = y - winsz2; i <= y + winsz2; ++i)
{
sum += sobel(x - winsz2, i);
if (cols_extra)
sum_extra += sobel(x + blockDim.x - winsz2, i);
}
*cols = sum;
if (cols_extra)
*cols_extra = sum_extra;
void postfilter_textureness(const DevMem2Db& input, int winsz, float avgTexturenessThreshold, const DevMem2Db& disp, cudaStream_t & stream)
{
avgTexturenessThreshold *= winsz * winsz;
__syncthreads();
texForTF.filterMode = cudaFilterModeLinear;
texForTF.addressMode[0] = cudaAddressModeWrap;
texForTF.addressMode[1] = cudaAddressModeWrap;
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForTF, input.data, desc, input.cols, input.rows, input.step ) );
__syncthreads();
dim3 threads(128, 1, 1);
dim3 grid(1, 1, 1);
for(int y = beg_row + 1; y < end_row; ++y)
{
sum = sum - sobel(x - winsz2, y - winsz2 - 1) + sobel(x - winsz2, y + winsz2);
*cols = sum;
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, RpT);
if (cols_extra)
{
sum_extra = sum_extra - sobel(x + blockDim.x - winsz2, y - winsz2 - 1) + sobel(x + blockDim.x - winsz2, y + winsz2);
*cols_extra = sum_extra;
}
size_t smem_size = (threads.x + threads.x + (winsz/2) * 2 ) * sizeof(float);
textureness_kernel<<<grid, threads, smem_size, stream>>>(disp, winsz, avgTexturenessThreshold);
cudaSafeCall( cudaGetLastError() );
__syncthreads();
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
__syncthreads();
}
}
}
void postfilter_textureness(const DevMem2Db& input, int winsz, float avgTexturenessThreshold, const DevMem2Db& disp, cudaStream_t & stream)
{
avgTexturenessThreshold *= winsz * winsz;
cudaSafeCall( cudaUnbindTexture (texForTF) );
}
texForTF.filterMode = cudaFilterModeLinear;
texForTF.addressMode[0] = cudaAddressModeWrap;
texForTF.addressMode[1] = cudaAddressModeWrap;
} // namespace stereobm
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForTF, input.data, desc, input.cols, input.rows, input.step ) );
END_OPENCV_DEVICE_NAMESPACE
dim3 threads(128, 1, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, RpT);
size_t smem_size = (threads.x + threads.x + (winsz/2) * 2 ) * sizeof(float);
textureness_kernel<<<grid, threads, smem_size, stream>>>(disp, winsz, avgTexturenessThreshold);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
cudaSafeCall( cudaUnbindTexture (texForTF) );
}
} // namespace stereobm
}}} // namespace cv { namespace gpu { namespace device

782
modules/gpu/src/cuda/stereobp.cu
Unescape View File

@ -44,489 +44,487 @@
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/limits.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace stereobp {
///////////////////////////////////////////////////////////////
/////////////////////// load constants ////////////////////////
///////////////////////////////////////////////////////////////
__constant__ int cndisp;
__constant__ float cmax_data_term;
__constant__ float cdata_weight;
__constant__ float cmax_disc_term;
__constant__ float cdisc_single_jump;
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump)
{
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int )) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_data_term, &max_data_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdata_weight, &data_weight, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc_term, &max_disc_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdisc_single_jump, &disc_single_jump, sizeof(float)) );
}
///////////////////////////////////////////////////////////////
////////////////////////// comp data //////////////////////////
///////////////////////////////////////////////////////////////
template <int cn> struct PixDiff;
template <> struct PixDiff<1>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *ls;
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
return ::abs((int)l - *rs);
}
uchar l;
};
template <> struct PixDiff<3>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *((uchar3*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
float val = tb * ::abs((int)l.x - rs[0]);
val += tg * ::abs((int)l.y - rs[1]);
val += tr * ::abs((int)l.z - rs[2]);
return val;
}
uchar3 l;
};
template <> struct PixDiff<4>
namespace cv { namespace gpu { namespace device
{
__device__ __forceinline__ PixDiff(const uchar* ls)
namespace stereobp
{
l = *((uchar4*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
uchar4 r = *((uchar4*)rs);
///////////////////////////////////////////////////////////////
/////////////////////// load constants ////////////////////////
///////////////////////////////////////////////////////////////
float val = tb * ::abs((int)l.x - r.x);
val += tg * ::abs((int)l.y - r.y);
val += tr * ::abs((int)l.z - r.z);
__constant__ int cndisp;
__constant__ float cmax_data_term;
__constant__ float cdata_weight;
__constant__ float cmax_disc_term;
__constant__ float cdisc_single_jump;
return val;
}
uchar4 l;
};
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump)
{
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int )) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_data_term, &max_data_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdata_weight, &data_weight, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc_term, &max_disc_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdisc_single_jump, &disc_single_jump, sizeof(float)) );
}
template <int cn, typename D>
__global__ void comp_data(const DevMem2Db left, const PtrStepb right, PtrElemStep_<D> data)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
///////////////////////////////////////////////////////////////
////////////////////////// comp data //////////////////////////
///////////////////////////////////////////////////////////////
if (y > 0 && y < left.rows - 1 && x > 0 && x < left.cols - 1)
{
const uchar* ls = left.ptr(y) + x * cn;
const PixDiff<cn> pixDiff(ls);
const uchar* rs = right.ptr(y) + x * cn;
template <int cn> struct PixDiff;
template <> struct PixDiff<1>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *ls;
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
return ::abs((int)l - *rs);
}
uchar l;
};
template <> struct PixDiff<3>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *((uchar3*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
D* ds = data.ptr(y) + x;
const size_t disp_step = data.step * left.rows;
float val = tb * ::abs((int)l.x - rs[0]);
val += tg * ::abs((int)l.y - rs[1]);
val += tr * ::abs((int)l.z - rs[2]);
for (int disp = 0; disp < cndisp; disp++)
return val;
}
uchar3 l;
};
template <> struct PixDiff<4>
{
if (x - disp >= 1)
__device__ __forceinline__ PixDiff(const uchar* ls)
{
float val = pixDiff(rs - disp * cn);
l = *((uchar4*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
uchar4 r = *((uchar4*)rs);
ds[disp * disp_step] = saturate_cast<D>(fmin(cdata_weight * val, cdata_weight * cmax_data_term));
float val = tb * ::abs((int)l.x - r.x);
val += tg * ::abs((int)l.y - r.y);
val += tr * ::abs((int)l.z - r.z);
return val;
}
else
uchar4 l;
};
template <int cn, typename D>
__global__ void comp_data(const DevMem2Db left, const PtrStepb right, PtrElemStep_<D> data)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y > 0 && y < left.rows - 1 && x > 0 && x < left.cols - 1)
{
ds[disp * disp_step] = saturate_cast<D>(cdata_weight * cmax_data_term);
const uchar* ls = left.ptr(y) + x * cn;
const PixDiff<cn> pixDiff(ls);
const uchar* rs = right.ptr(y) + x * cn;
D* ds = data.ptr(y) + x;
const size_t disp_step = data.step * left.rows;
for (int disp = 0; disp < cndisp; disp++)
{
if (x - disp >= 1)
{
float val = pixDiff(rs - disp * cn);
ds[disp * disp_step] = saturate_cast<D>(fmin(cdata_weight * val, cdata_weight * cmax_data_term));
}
else
{
ds[disp * disp_step] = saturate_cast<D>(cdata_weight * cmax_data_term);
}
}
}
}
}
}
template<typename T, typename D>
void comp_data_gpu(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream);
template <> void comp_data_gpu<uchar, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
template<typename T, typename D>
void comp_data_gpu(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
template <> void comp_data_gpu<uchar, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<1, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<1, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<1, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
comp_data<1, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
template <> void comp_data_gpu<uchar3, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
template <> void comp_data_gpu<uchar3, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<3, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar3, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<3, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar3, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<3, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
comp_data<3, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
template <> void comp_data_gpu<uchar4, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
template <> void comp_data_gpu<uchar4, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<4, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar4, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<4, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar4, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
comp_data<4, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
comp_data<4, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
///////////////////////////////////////////////////////////////
//////////////////////// data step down ///////////////////////
///////////////////////////////////////////////////////////////
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T>
__global__ void data_step_down(int dst_cols, int dst_rows, int src_rows, const PtrStep<T> src, PtrStep<T> dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
///////////////////////////////////////////////////////////////
//////////////////////// data step down ///////////////////////
///////////////////////////////////////////////////////////////
if (x < dst_cols && y < dst_rows)
{
for (int d = 0; d < cndisp; ++d)
template <typename T>
__global__ void data_step_down(int dst_cols, int dst_rows, int src_rows, const PtrStep<T> src, PtrStep<T> dst)
{
float dst_reg = src.ptr(d * src_rows + (2*y+0))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+0))[(2*x+1)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+1)];
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
dst.ptr(d * dst_rows + y)[x] = saturate_cast<T>(dst_reg);
if (x < dst_cols && y < dst_rows)
{
for (int d = 0; d < cndisp; ++d)
{
float dst_reg = src.ptr(d * src_rows + (2*y+0))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+0))[(2*x+1)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+1)];
dst.ptr(d * dst_rows + y)[x] = saturate_cast<T>(dst_reg);
}
}
}
}
}
template<typename T>
void data_step_down_gpu(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
template<typename T>
void data_step_down_gpu(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
data_step_down<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)src, (DevMem2D_<T>)dst);
cudaSafeCall( cudaGetLastError() );
data_step_down<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)src, (DevMem2D_<T>)dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void data_step_down_gpu<short>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
template void data_step_down_gpu<float>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
template void data_step_down_gpu<short>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
template void data_step_down_gpu<float>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
///////////////////////////////////////////////////////////////
/////////////////// level up messages ////////////////////////
///////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////
/////////////////// level up messages ////////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__global__ void level_up_message(int dst_cols, int dst_rows, int src_rows, const PtrElemStep_<T> src, PtrElemStep_<T> dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
template <typename T>
__global__ void level_up_message(int dst_cols, int dst_rows, int src_rows, const PtrElemStep_<T> src, PtrElemStep_<T> dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < dst_cols && y < dst_rows)
{
const size_t dst_disp_step = dst.step * dst_rows;
const size_t src_disp_step = src.step * src_rows;
if (x < dst_cols && y < dst_rows)
{
const size_t dst_disp_step = dst.step * dst_rows;
const size_t src_disp_step = src.step * src_rows;
T* dstr = dst.ptr(y ) + x;
const T* srcr = src.ptr(y/2) + x/2;
T* dstr = dst.ptr(y ) + x;
const T* srcr = src.ptr(y/2) + x/2;
for (int d = 0; d < cndisp; ++d)
dstr[d * dst_disp_step] = srcr[d * src_disp_step];
}
}
for (int d = 0; d < cndisp; ++d)
dstr[d * dst_disp_step] = srcr[d * src_disp_step];
}
}
template <typename T>
void level_up_messages_gpu(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
template <typename T>
void level_up_messages_gpu(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
int src_idx = (dst_idx + 1) & 1;
int src_idx = (dst_idx + 1) & 1;
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mus[src_idx], (DevMem2D_<T>)mus[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mus[src_idx], (DevMem2D_<T>)mus[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mds[src_idx], (DevMem2D_<T>)mds[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mds[src_idx], (DevMem2D_<T>)mds[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mls[src_idx], (DevMem2D_<T>)mls[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mls[src_idx], (DevMem2D_<T>)mls[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mrs[src_idx], (DevMem2D_<T>)mrs[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mrs[src_idx], (DevMem2D_<T>)mrs[dst_idx]);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void level_up_messages_gpu<short>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
template void level_up_messages_gpu<float>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
template void level_up_messages_gpu<short>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
template void level_up_messages_gpu<float>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
///////////////////////////////////////////////////////////////
//////////////////// calc all iterations /////////////////////
///////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////
//////////////////// calc all iterations /////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__device__ void calc_min_linear_penalty(T* dst, size_t step)
{
float prev = dst[0];
float cur;
for (int disp = 1; disp < cndisp; ++disp)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
template <typename T>
__device__ void calc_min_linear_penalty(T* dst, size_t step)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
float prev = dst[0];
float cur;
for (int disp = 1; disp < cndisp; ++disp)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
}
prev = cur;
}
prev = dst[(cndisp - 1) * step];
for (int disp = cndisp - 2; disp >= 0; disp--)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
}
prev = cur;
}
}
prev = cur;
}
prev = dst[(cndisp - 1) * step];
for (int disp = cndisp - 2; disp >= 0; disp--)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
template <typename T>
__device__ void message(const T* msg1, const T* msg2, const T* msg3, const T* data, T* dst, size_t msg_disp_step, size_t data_disp_step)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
}
prev = cur;
}
}
float minimum = device::numeric_limits<float>::max();
template <typename T>
__device__ void message(const T* msg1, const T* msg2, const T* msg3, const T* data, T* dst, size_t msg_disp_step, size_t data_disp_step)
{
float minimum = device::numeric_limits<float>::max();
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = msg1[msg_disp_step * i];
dst_reg += msg2[msg_disp_step * i];
dst_reg += msg3[msg_disp_step * i];
dst_reg += data[data_disp_step * i];
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = msg1[msg_disp_step * i];
dst_reg += msg2[msg_disp_step * i];
dst_reg += msg3[msg_disp_step * i];
dst_reg += data[data_disp_step * i];
if (dst_reg < minimum)
minimum = dst_reg;
if (dst_reg < minimum)
minimum = dst_reg;
dst[msg_disp_step * i] = saturate_cast<T>(dst_reg);
}
dst[msg_disp_step * i] = saturate_cast<T>(dst_reg);
}
calc_min_linear_penalty(dst, msg_disp_step);
calc_min_linear_penalty(dst, msg_disp_step);
minimum += cmax_disc_term;
minimum += cmax_disc_term;
float sum = 0;
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = dst[msg_disp_step * i];
if (dst_reg > minimum)
{
dst_reg = minimum;
dst[msg_disp_step * i] = saturate_cast<T>(minimum);
}
sum += dst_reg;
}
sum /= cndisp;
float sum = 0;
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = dst[msg_disp_step * i];
if (dst_reg > minimum)
{
dst_reg = minimum;
dst[msg_disp_step * i] = saturate_cast<T>(minimum);
for(int i = 0; i < cndisp; ++i)
dst[msg_disp_step * i] -= sum;
}
sum += dst_reg;
}
sum /= cndisp;
for(int i = 0; i < cndisp; ++i)
dst[msg_disp_step * i] -= sum;
}
template <typename T>
__global__ void one_iteration(int t, PtrElemStep_<T> u, T* d, T* l, T* r, const PtrElemStep_<T> data, int cols, int rows)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
if ((y > 0) && (y < rows - 1) && (x > 0) && (x < cols - 1))
{
T* us = u.ptr(y) + x;
T* ds = d + y * u.step + x;
T* ls = l + y * u.step + x;
T* rs = r + y * u.step + x;
const T* dt = data.ptr(y) + x;
size_t msg_disp_step = u.step * rows;
size_t data_disp_step = data.step * rows;
message(us + u.step, ls + 1, rs - 1, dt, us, msg_disp_step, data_disp_step);
message(ds - u.step, ls + 1, rs - 1, dt, ds, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, rs - 1, dt, rs, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, ls + 1, dt, ls, msg_disp_step, data_disp_step);
}
}
template <typename T>
void calc_all_iterations_gpu(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d,
const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(cols, threads.x << 1);
grid.y = divUp(rows, threads.y);
template <typename T>
__global__ void one_iteration(int t, PtrElemStep_<T> u, T* d, T* l, T* r, const PtrElemStep_<T> data, int cols, int rows)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
for(int t = 0; t < iters; ++t)
{
one_iteration<T><<<grid, threads, 0, stream>>>(t, (DevMem2D_<T>)u, (T*)d.data, (T*)l.data, (T*)r.data, (DevMem2D_<T>)data, cols, rows);
cudaSafeCall( cudaGetLastError() );
if ((y > 0) && (y < rows - 1) && (x > 0) && (x < cols - 1))
{
T* us = u.ptr(y) + x;
T* ds = d + y * u.step + x;
T* ls = l + y * u.step + x;
T* rs = r + y * u.step + x;
const T* dt = data.ptr(y) + x;
size_t msg_disp_step = u.step * rows;
size_t data_disp_step = data.step * rows;
message(us + u.step, ls + 1, rs - 1, dt, us, msg_disp_step, data_disp_step);
message(ds - u.step, ls + 1, rs - 1, dt, ds, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, rs - 1, dt, rs, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, ls + 1, dt, ls, msg_disp_step, data_disp_step);
}
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
template <typename T>
void calc_all_iterations_gpu(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d,
const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
template void calc_all_iterations_gpu<short>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
template void calc_all_iterations_gpu<float>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
grid.x = divUp(cols, threads.x << 1);
grid.y = divUp(rows, threads.y);
///////////////////////////////////////////////////////////////
/////////////////////////// output ////////////////////////////
///////////////////////////////////////////////////////////////
for(int t = 0; t < iters; ++t)
{
one_iteration<T><<<grid, threads, 0, stream>>>(t, (DevMem2D_<T>)u, (T*)d.data, (T*)l.data, (T*)r.data, (DevMem2D_<T>)data, cols, rows);
cudaSafeCall( cudaGetLastError() );
template <typename T>
__global__ void output(const PtrElemStep_<T> u, const T* d, const T* l, const T* r, const T* data,
DevMem2D_<short> disp)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
if (y > 0 && y < disp.rows - 1 && x > 0 && x < disp.cols - 1)
{
const T* us = u.ptr(y + 1) + x;
const T* ds = d + (y - 1) * u.step + x;
const T* ls = l + y * u.step + (x + 1);
const T* rs = r + y * u.step + (x - 1);
const T* dt = data + y * u.step + x;
template void calc_all_iterations_gpu<short>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
template void calc_all_iterations_gpu<float>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
size_t disp_step = disp.rows * u.step;
///////////////////////////////////////////////////////////////
/////////////////////////// output ////////////////////////////
///////////////////////////////////////////////////////////////
int best = 0;
float best_val = numeric_limits<float>::max();
for (int d = 0; d < cndisp; ++d)
template <typename T>
__global__ void output(const PtrElemStep_<T> u, const T* d, const T* l, const T* r, const T* data,
DevMem2D_<short> disp)
{
float val = us[d * disp_step];
val += ds[d * disp_step];
val += ls[d * disp_step];
val += rs[d * disp_step];
val += dt[d * disp_step];
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (val < best_val)
if (y > 0 && y < disp.rows - 1 && x > 0 && x < disp.cols - 1)
{
best_val = val;
best = d;
const T* us = u.ptr(y + 1) + x;
const T* ds = d + (y - 1) * u.step + x;
const T* ls = l + y * u.step + (x + 1);
const T* rs = r + y * u.step + (x - 1);
const T* dt = data + y * u.step + x;
size_t disp_step = disp.rows * u.step;
int best = 0;
float best_val = numeric_limits<float>::max();
for (int d = 0; d < cndisp; ++d)
{
float val = us[d * disp_step];
val += ds[d * disp_step];
val += ls[d * disp_step];
val += rs[d * disp_step];
val += dt[d * disp_step];
if (val < best_val)
{
best_val = val;
best = d;
}
}
disp.ptr(y)[x] = saturate_cast<short>(best);
}
}
disp.ptr(y)[x] = saturate_cast<short>(best);
}
}
template <typename T>
void output_gpu(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data,
const DevMem2D_<short>& disp, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(disp.cols, threads.x);
grid.y = divUp(disp.rows, threads.y);
output<T><<<grid, threads, 0, stream>>>((DevMem2D_<T>)u, (const T*)d.data, (const T*)l.data, (const T*)r.data, (const T*)data.data, disp);
cudaSafeCall( cudaGetLastError() );
template <typename T>
void output_gpu(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data,
const DevMem2D_<short>& disp, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
grid.x = divUp(disp.cols, threads.x);
grid.y = divUp(disp.rows, threads.y);
template void output_gpu<short>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
template void output_gpu<float>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
output<T><<<grid, threads, 0, stream>>>((DevMem2D_<T>)u, (const T*)d.data, (const T*)l.data, (const T*)r.data, (const T*)data.data, disp);
cudaSafeCall( cudaGetLastError() );
} // namespace stereobp
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
END_OPENCV_DEVICE_NAMESPACE
template void output_gpu<short>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
template void output_gpu<float>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
} // namespace stereobp
}}} // namespace cv { namespace gpu { namespace device

1334
modules/gpu/src/cuda/stereocsbp.cu
View File

File diff suppressed because it is too large Load Diff

1570
modules/gpu/src/cuda/surf.cu
View File

File diff suppressed because it is too large Load Diff

21
modules/gpu/src/cudastream.cpp
Unescape View File

@ -71,20 +71,19 @@ cv::gpu::Stream::operator bool() const { throw_nogpu(); return false; }
#include "opencv2/gpu/stream_accessor.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
void copy_to_with_mask(const DevMem2Db& src, DevMem2Db dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t & stream = 0);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
void copy_to_with_mask(const DevMem2Db& src, DevMem2Db dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t & stream = 0);
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta, cudaStream_t stream = 0);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta, cudaStream_t stream = 0);
}}}
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
struct Stream::Impl
{

250
modules/gpu/src/element_operations.cpp
Unescape View File

@ -123,19 +123,18 @@ namespace
////////////////////////////////////////////////////////////////////////
// add
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T, typename D>
void add_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
template <typename T, typename D>
void add_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
template <typename T, typename D>
void add_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T, typename D>
void add_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
}}}
void cv::gpu::add(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const GpuMat& mask, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
@ -174,7 +173,7 @@ void cv::gpu::add(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const Gpu
void cv::gpu::add(const GpuMat& src, const Scalar& sc, GpuMat& dst, const GpuMat& mask, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, double val, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
@ -236,19 +235,18 @@ void cv::gpu::add(const GpuMat& src, const Scalar& sc, GpuMat& dst, const GpuMat
////////////////////////////////////////////////////////////////////////
// subtract
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T, typename D>
void subtract_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
template <typename T, typename D>
void subtract_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
template <typename T, typename D>
void subtract_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T, typename D>
void subtract_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
}}}
void cv::gpu::subtract(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const GpuMat& mask, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
@ -287,7 +285,7 @@ void cv::gpu::subtract(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, cons
void cv::gpu::subtract(const GpuMat& src, const Scalar& sc, GpuMat& dst, const GpuMat& mask, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, double val, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream);
@ -349,22 +347,21 @@ void cv::gpu::subtract(const GpuMat& src, const Scalar& sc, GpuMat& dst, const G
////////////////////////////////////////////////////////////////////////
// multiply
BEGIN_OPENCV_DEVICE_NAMESPACE
void multiply_gpu(const DevMem2D_<uchar4>& src1, const DevMem2Df& src2, const DevMem2D_<uchar4>& dst, cudaStream_t stream);
void multiply_gpu(const DevMem2D_<short4>& src1, const DevMem2Df& src2, const DevMem2D_<short4>& dst, cudaStream_t stream);
template <typename T, typename D>
void multiply_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
void multiply_gpu(const DevMem2D_<uchar4>& src1, const DevMem2Df& src2, const DevMem2D_<uchar4>& dst, cudaStream_t stream);
void multiply_gpu(const DevMem2D_<short4>& src1, const DevMem2Df& src2, const DevMem2D_<short4>& dst, cudaStream_t stream);
template <typename T, typename D>
void multiply_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, double scale, cudaStream_t stream);
template <typename T, typename D>
void multiply_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T, typename D>
void multiply_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, double scale, cudaStream_t stream);
}}}
void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, double scale, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream);
@ -422,7 +419,7 @@ void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, doub
void cv::gpu::multiply(const GpuMat& src, const Scalar& sc, GpuMat& dst, double scale, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, double val, const DevMem2Db& dst, double scale, cudaStream_t stream);
@ -472,25 +469,24 @@ void cv::gpu::multiply(const GpuMat& src, const Scalar& sc, GpuMat& dst, double
////////////////////////////////////////////////////////////////////////
// divide
BEGIN_OPENCV_DEVICE_NAMESPACE
void divide_gpu(const DevMem2D_<uchar4>& src1, const DevMem2Df& src2, const DevMem2D_<uchar4>& dst, cudaStream_t stream);
void divide_gpu(const DevMem2D_<short4>& src1, const DevMem2Df& src2, const DevMem2D_<short4>& dst, cudaStream_t stream);
template <typename T, typename D>
void divide_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
void divide_gpu(const DevMem2D_<uchar4>& src1, const DevMem2Df& src2, const DevMem2D_<uchar4>& dst, cudaStream_t stream);
void divide_gpu(const DevMem2D_<short4>& src1, const DevMem2Df& src2, const DevMem2D_<short4>& dst, cudaStream_t stream);
template <typename T, typename D>
void divide_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, double scale, cudaStream_t stream);
template <typename T, typename D>
void divide_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream);
template <typename T, typename D>
void divide_gpu(double scalar, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T, typename D>
void divide_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, double scale, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T, typename D>
void divide_gpu(double scalar, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
}}}
void cv::gpu::divide(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, double scale, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream);
@ -548,7 +544,7 @@ void cv::gpu::divide(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, double
void cv::gpu::divide(const GpuMat& src, const Scalar& sc, GpuMat& dst, double scale, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, double val, const DevMem2Db& dst, double scale, cudaStream_t stream);
@ -597,7 +593,7 @@ void cv::gpu::divide(const GpuMat& src, const Scalar& sc, GpuMat& dst, double sc
void cv::gpu::divide(double scale, const GpuMat& src, GpuMat& dst, int dtype, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(double scalar, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
@ -630,19 +626,18 @@ void cv::gpu::divide(double scale, const GpuMat& src, GpuMat& dst, int dtype, St
//////////////////////////////////////////////////////////////////////////////
// absdiff
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T>
void absdiff_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T>
void absdiff_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
template <typename T>
void absdiff_gpu(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T>
void absdiff_gpu(const DevMem2Db& src1, double val, const DevMem2Db& dst, cudaStream_t stream);
}}}
void cv::gpu::absdiff(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
@ -714,7 +709,7 @@ void cv::gpu::absdiff(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, Strea
void cv::gpu::absdiff(const GpuMat& src1, const Scalar& src2, GpuMat& dst, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, double val, const DevMem2Db& dst, cudaStream_t stream);
@ -758,18 +753,17 @@ void cv::gpu::absdiff(const GpuMat& src1, const Scalar& src2, GpuMat& dst, Strea
//////////////////////////////////////////////////////////////////////////////
// Comparison of two matrixes
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T> void compare_eq(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T> void compare_ne(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T> void compare_lt(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T> void compare_le(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
template <typename T> void compare_eq(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T> void compare_ne(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T> void compare_lt(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
template <typename T> void compare_le(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
}}}
void cv::gpu::compare(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, int cmpop, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*func_t)(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, cudaStream_t stream);
@ -835,14 +829,13 @@ void cv::gpu::compare(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, int c
//////////////////////////////////////////////////////////////////////////////
// Unary bitwise logical operations
BEGIN_OPENCV_DEVICE_NAMESPACE
void bitwiseNotCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src, PtrStepb dst, cudaStream_t stream);
template <typename T>
void bitwiseMaskNotCaller(int rows, int cols, int cn, const PtrStepb src, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
void bitwiseNotCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src, PtrStepb dst, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T>
void bitwiseMaskNotCaller(int rows, int cols, int cn, const PtrStepb src, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
}}}
namespace
{
@ -850,13 +843,13 @@ namespace
{
dst.create(src.size(), src.type());
OPENCV_DEVICE_NAMESPACE_ bitwiseNotCaller(src.rows, src.cols, src.elemSize1(), dst.channels(), src, dst, stream);
::cv::gpu::device::bitwiseNotCaller(src.rows, src.cols, src.elemSize1(), dst.channels(), src, dst, stream);
}
void bitwiseNotCaller(const GpuMat& src, GpuMat& dst, const GpuMat& mask, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*Caller)(int, int, int, const PtrStepb, const PtrStepb, PtrStepb, cudaStream_t);
@ -893,24 +886,23 @@ void cv::gpu::bitwise_not(const GpuMat& src, GpuMat& dst, const GpuMat& mask, St
//////////////////////////////////////////////////////////////////////////////
// Binary bitwise logical operations
BEGIN_OPENCV_DEVICE_NAMESPACE
void bitwiseOrCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src1, const PtrStepb src2, PtrStepb dst, cudaStream_t stream);
template <typename T>
void bitwiseMaskOrCaller(int rows, int cols, int cn, const PtrStepb src1, const PtrStepb src2, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
void bitwiseOrCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src1, const PtrStepb src2, PtrStepb dst, cudaStream_t stream);
void bitwiseAndCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src1, const PtrStepb src2, PtrStepb dst, cudaStream_t stream);
template <typename T>
void bitwiseMaskOrCaller(int rows, int cols, int cn, const PtrStepb src1, const PtrStepb src2, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
template <typename T>
void bitwiseMaskAndCaller(int rows, int cols, int cn, const PtrStepb src1, const PtrStepb src2, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
void bitwiseAndCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src1, const PtrStepb src2, PtrStepb dst, cudaStream_t stream);
void bitwiseXorCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src1, const PtrStepb src2, PtrStepb dst, cudaStream_t stream);
template <typename T>
void bitwiseMaskAndCaller(int rows, int cols, int cn, const PtrStepb src1, const PtrStepb src2, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
template <typename T>
void bitwiseMaskXorCaller(int rows, int cols, int cn, const PtrStepb src1, const PtrStepb src2, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
void bitwiseXorCaller(int rows, int cols, size_t elem_size1, int cn, const PtrStepb src1, const PtrStepb src2, PtrStepb dst, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T>
void bitwiseMaskXorCaller(int rows, int cols, int cn, const PtrStepb src1, const PtrStepb src2, const PtrStepb mask, PtrStepb dst, cudaStream_t stream);
}}}
namespace
{
@ -919,12 +911,12 @@ namespace
CV_Assert(src1.size() == src2.size() && src1.type() == src2.type());
dst.create(src1.size(), src1.type());
OPENCV_DEVICE_NAMESPACE_ bitwiseOrCaller(dst.rows, dst.cols, dst.elemSize1(), dst.channels(), src1, src2, dst, stream);
::cv::gpu::device::bitwiseOrCaller(dst.rows, dst.cols, dst.elemSize1(), dst.channels(), src1, src2, dst, stream);
}
void bitwiseOrCaller(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const GpuMat& mask, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*Caller)(int, int, int, const PtrStepb, const PtrStepb, const PtrStepb, PtrStepb, cudaStream_t);
@ -952,13 +944,13 @@ namespace
CV_Assert(src1.size() == src2.size() && src1.type() == src2.type());
dst.create(src1.size(), src1.type());
OPENCV_DEVICE_NAMESPACE_ bitwiseAndCaller(dst.rows, dst.cols, dst.elemSize1(), dst.channels(), src1, src2, dst, stream);
::cv::gpu::device::bitwiseAndCaller(dst.rows, dst.cols, dst.elemSize1(), dst.channels(), src1, src2, dst, stream);
}
void bitwiseAndCaller(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const GpuMat& mask, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*Caller)(int, int, int, const PtrStepb, const PtrStepb, const PtrStepb, PtrStepb, cudaStream_t);
@ -986,13 +978,13 @@ namespace
CV_Assert(src1.size() == src2.size() && src1.type() == src2.type());
dst.create(src1.size(), src1.type());
OPENCV_DEVICE_NAMESPACE_ bitwiseXorCaller(dst.rows, dst.cols, dst.elemSize1(), dst.channels(), src1, src2, dst, stream);
::cv::gpu::device::bitwiseXorCaller(dst.rows, dst.cols, dst.elemSize1(), dst.channels(), src1, src2, dst, stream);
}
void bitwiseXorCaller(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const GpuMat& mask, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
typedef void (*Caller)(int, int, int, const PtrStepb, const PtrStepb, const PtrStepb, PtrStepb, cudaStream_t);
@ -1046,21 +1038,20 @@ void cv::gpu::bitwise_xor(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, c
//////////////////////////////////////////////////////////////////////////////
// Minimum and maximum operations
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T>
void min_gpu(const DevMem2D_<T>& src1, const DevMem2D_<T>& src2, const DevMem2D_<T>& dst, cudaStream_t stream);
template <typename T>
void max_gpu(const DevMem2D_<T>& src1, const DevMem2D_<T>& src2, const DevMem2D_<T>& dst, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
template <typename T>
void min_gpu(const DevMem2D_<T>& src1, const DevMem2D_<T>& src2, const DevMem2D_<T>& dst, cudaStream_t stream);
template <typename T>
void min_gpu(const DevMem2D_<T>& src1, T src2, const DevMem2D_<T>& dst, cudaStream_t stream);
template <typename T>
void max_gpu(const DevMem2D_<T>& src1, const DevMem2D_<T>& src2, const DevMem2D_<T>& dst, cudaStream_t stream);
template <typename T>
void max_gpu(const DevMem2D_<T>& src1, T src2, const DevMem2D_<T>& dst, cudaStream_t stream);
template <typename T>
void min_gpu(const DevMem2D_<T>& src1, T src2, const DevMem2D_<T>& dst, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template <typename T>
void max_gpu(const DevMem2D_<T>& src1, T src2, const DevMem2D_<T>& dst, cudaStream_t stream);
}}}
namespace
{
@ -1069,14 +1060,14 @@ namespace
{
CV_Assert(src1.size() == src2.size() && src1.type() == src2.type());
dst.create(src1.size(), src1.type());
OPENCV_DEVICE_NAMESPACE_ min_gpu<T>(src1.reshape(1), src2.reshape(1), dst.reshape(1), stream);
::cv::gpu::device::min_gpu<T>(src1.reshape(1), src2.reshape(1), dst.reshape(1), stream);
}
template <typename T>
void min_caller(const GpuMat& src1, double src2, GpuMat& dst, cudaStream_t stream)
{
dst.create(src1.size(), src1.type());
OPENCV_DEVICE_NAMESPACE_ min_gpu<T>(src1.reshape(1), saturate_cast<T>(src2), dst.reshape(1), stream);
::cv::gpu::device::min_gpu<T>(src1.reshape(1), saturate_cast<T>(src2), dst.reshape(1), stream);
}
template <typename T>
@ -1084,14 +1075,14 @@ namespace
{
CV_Assert(src1.size() == src2.size() && src1.type() == src2.type());
dst.create(src1.size(), src1.type());
OPENCV_DEVICE_NAMESPACE_ max_gpu<T>(src1.reshape(1), src2.reshape(1), dst.reshape(1), stream);
::cv::gpu::device::max_gpu<T>(src1.reshape(1), src2.reshape(1), dst.reshape(1), stream);
}
template <typename T>
void max_caller(const GpuMat& src1, double src2, GpuMat& dst, cudaStream_t stream)
{
dst.create(src1.size(), src1.type());
OPENCV_DEVICE_NAMESPACE_ max_gpu<T>(src1.reshape(1), saturate_cast<T>(src2), dst.reshape(1), stream);
::cv::gpu::device::max_gpu<T>(src1.reshape(1), saturate_cast<T>(src2), dst.reshape(1), stream);
}
}
@ -1155,18 +1146,17 @@ void cv::gpu::max(const GpuMat& src1, double src2, GpuMat& dst, Stream& stream)
////////////////////////////////////////////////////////////////////////
// threshold
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T>
void threshold_gpu(const DevMem2Db& src, const DevMem2Db& dst, T thresh, T maxVal, int type, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
template <typename T>
void threshold_gpu(const DevMem2Db& src, const DevMem2Db& dst, T thresh, T maxVal, int type, cudaStream_t stream);
}}}
namespace
{
template <typename T> void threshold_caller(const GpuMat& src, GpuMat& dst, double thresh, double maxVal, int type, cudaStream_t stream)
{
OPENCV_DEVICE_NAMESPACE_ threshold_gpu<T>(src, dst, saturate_cast<T>(thresh), saturate_cast<T>(maxVal), type, stream);
::cv::gpu::device::threshold_gpu<T>(src, dst, saturate_cast<T>(thresh), saturate_cast<T>(maxVal), type, stream);
}
}
@ -1223,16 +1213,15 @@ double cv::gpu::threshold(const GpuMat& src, GpuMat& dst, double thresh, double
////////////////////////////////////////////////////////////////////////
// pow
BEGIN_OPENCV_DEVICE_NAMESPACE
template<typename T>
void pow_caller(const DevMem2Db& src, float power, DevMem2Db dst, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
template<typename T>
void pow_caller(const DevMem2Db& src, float power, DevMem2Db dst, cudaStream_t stream);
}}}
void cv::gpu::pow(const GpuMat& src, double power, GpuMat& dst, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
CV_Assert(src.depth() != CV_64F);
dst.create(src.size(), src.type());
@ -1252,16 +1241,15 @@ void cv::gpu::pow(const GpuMat& src, double power, GpuMat& dst, Stream& stream)
////////////////////////////////////////////////////////////////////////
// addWeighted
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T1, typename T2, typename D>
void addWeighted_gpu(const DevMem2Db& src1, double alpha, const DevMem2Db& src2, double beta, double gamma, const DevMem2Db& dst, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
template <typename T1, typename T2, typename D>
void addWeighted_gpu(const DevMem2Db& src1, double alpha, const DevMem2Db& src2, double beta, double gamma, const DevMem2Db& dst, cudaStream_t stream);
}}}
void cv::gpu::addWeighted(const GpuMat& src1, double alpha, const GpuMat& src2, double beta, double gamma, GpuMat& dst, int dtype, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
CV_Assert(src1.size() == src2.size());
CV_Assert(src1.type() == src2.type() || (dtype >= 0 && src1.channels() == src2.channels()));

29
modules/gpu/src/filtering.cpp
Unescape View File

@ -735,21 +735,20 @@ void cv::gpu::filter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& ke
////////////////////////////////////////////////////////////////////////////////////////////////////
// Separable Linear Filter
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace row_filter
namespace cv { namespace gpu { namespace device
{
template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
}
namespace column_filter
{
template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
}
namespace row_filter
{
template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace column_filter
{
template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
}
}}}
namespace
{
@ -803,7 +802,7 @@ namespace
Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType, const Mat& rowKernel, int anchor, int borderType)
{
using namespace OPENCV_DEVICE_NAMESPACE_ row_filter;
using namespace ::cv::gpu::device::row_filter;
static const nppFilter1D_t nppFilter1D_callers[] = {0, nppiFilterRow_8u_C1R, 0, 0, nppiFilterRow_8u_C4R};
@ -918,7 +917,7 @@ namespace
Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int dstType, const Mat& columnKernel, int anchor, int borderType)
{
using namespace OPENCV_DEVICE_NAMESPACE_ column_filter;
using namespace ::cv::gpu::device::column_filter;
static const nppFilter1D_t nppFilter1D_callers[] = {0, nppiFilterColumn_8u_C1R, 0, 0, nppiFilterColumn_8u_C4R};

71
modules/gpu/src/hog.cpp
Unescape View File

@ -60,44 +60,43 @@ std::vector<float> cv::gpu::HOGDescriptor::getPeopleDetector64x128() { throw_nog
#else
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace hog
namespace cv { namespace gpu { namespace device
{
void set_up_constants(int nbins, int block_stride_x, int block_stride_y,
int nblocks_win_x, int nblocks_win_y);
void compute_hists(int nbins, int block_stride_x, int blovck_stride_y,
int height, int width, const cv::gpu::DevMem2Df& grad,
const cv::gpu::DevMem2Db& qangle, float sigma, float* block_hists);
void normalize_hists(int nbins, int block_stride_x, int block_stride_y,
int height, int width, float* block_hists, float threshold);
void classify_hists(int win_height, int win_width, int block_stride_y,
int block_stride_x, int win_stride_y, int win_stride_x, int height,
int width, float* block_hists, float* coefs, float free_coef,
float threshold, unsigned char* labels);
void extract_descrs_by_rows(int win_height, int win_width, int block_stride_y, int block_stride_x,
int win_stride_y, int win_stride_x, int height, int width, float* block_hists,
cv::gpu::DevMem2Df descriptors);
void extract_descrs_by_cols(int win_height, int win_width, int block_stride_y, int block_stride_x,
int win_stride_y, int win_stride_x, int height, int width, float* block_hists,
cv::gpu::DevMem2Df descriptors);
void compute_gradients_8UC1(int nbins, int height, int width, const cv::gpu::DevMem2Db& img,
float angle_scale, cv::gpu::DevMem2Df grad, cv::gpu::DevMem2Db qangle, bool correct_gamma);
void compute_gradients_8UC4(int nbins, int height, int width, const cv::gpu::DevMem2Db& img,
float angle_scale, cv::gpu::DevMem2Df grad, cv::gpu::DevMem2Db qangle, bool correct_gamma);
void resize_8UC1(const cv::gpu::DevMem2Db& src, cv::gpu::DevMem2Db dst);
void resize_8UC4(const cv::gpu::DevMem2Db& src, cv::gpu::DevMem2Db dst);
}
END_OPENCV_DEVICE_NAMESPACE
namespace hog
{
void set_up_constants(int nbins, int block_stride_x, int block_stride_y,
int nblocks_win_x, int nblocks_win_y);
void compute_hists(int nbins, int block_stride_x, int blovck_stride_y,
int height, int width, const cv::gpu::DevMem2Df& grad,
const cv::gpu::DevMem2Db& qangle, float sigma, float* block_hists);
void normalize_hists(int nbins, int block_stride_x, int block_stride_y,
int height, int width, float* block_hists, float threshold);
void classify_hists(int win_height, int win_width, int block_stride_y,
int block_stride_x, int win_stride_y, int win_stride_x, int height,
int width, float* block_hists, float* coefs, float free_coef,
float threshold, unsigned char* labels);
void extract_descrs_by_rows(int win_height, int win_width, int block_stride_y, int block_stride_x,
int win_stride_y, int win_stride_x, int height, int width, float* block_hists,
cv::gpu::DevMem2Df descriptors);
void extract_descrs_by_cols(int win_height, int win_width, int block_stride_y, int block_stride_x,
int win_stride_y, int win_stride_x, int height, int width, float* block_hists,
cv::gpu::DevMem2Df descriptors);
void compute_gradients_8UC1(int nbins, int height, int width, const cv::gpu::DevMem2Db& img,
float angle_scale, cv::gpu::DevMem2Df grad, cv::gpu::DevMem2Db qangle, bool correct_gamma);
void compute_gradients_8UC4(int nbins, int height, int width, const cv::gpu::DevMem2Db& img,
float angle_scale, cv::gpu::DevMem2Df grad, cv::gpu::DevMem2Db qangle, bool correct_gamma);
void resize_8UC1(const cv::gpu::DevMem2Db& src, cv::gpu::DevMem2Db dst);
void resize_8UC4(const cv::gpu::DevMem2Db& src, cv::gpu::DevMem2Db dst);
}
}}}
using namespace OPENCV_DEVICE_NAMESPACE;
using namespace ::cv::gpu::device;
cv::gpu::HOGDescriptor::HOGDescriptor(Size win_size, Size block_size, Size block_stride, Size cell_size,
int nbins, double win_sigma, double threshold_L2hys, bool gamma_correction, int nlevels)

341
modules/gpu/src/imgproc.cpp
Unescape View File

@ -107,20 +107,19 @@ void cv::gpu::CannyBuf::release() { throw_nogpu(); }
////////////////////////////////////////////////////////////////////////
// remap
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
template <typename T>
void remap_gpu(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst,
int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
template <typename T>
void remap_gpu(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst,
int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
}
}}}
void cv::gpu::remap(const GpuMat& src, GpuMat& dst, const GpuMat& xmap, const GpuMat& ymap, int interpolation, int borderMode, const Scalar& borderValue, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
typedef void (*caller_t)(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc);
@ -160,18 +159,17 @@ void cv::gpu::remap(const GpuMat& src, GpuMat& dst, const GpuMat& xmap, const Gp
////////////////////////////////////////////////////////////////////////
// meanShiftFiltering_GPU
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void meanShiftFiltering_gpu(const DevMem2Db& src, DevMem2Db dst, int sp, int sr, int maxIter, float eps, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void meanShiftFiltering_gpu(const DevMem2Db& src, DevMem2Db dst, int sp, int sr, int maxIter, float eps, cudaStream_t stream);
}
}}}
void cv::gpu::meanShiftFiltering(const GpuMat& src, GpuMat& dst, int sp, int sr, TermCriteria criteria, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
if( src.empty() )
CV_Error( CV_StsBadArg, "The input image is empty" );
@ -197,18 +195,17 @@ void cv::gpu::meanShiftFiltering(const GpuMat& src, GpuMat& dst, int sp, int sr,
////////////////////////////////////////////////////////////////////////
// meanShiftProc_GPU
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void meanShiftProc_gpu(const DevMem2Db& src, DevMem2Db dstr, DevMem2Db dstsp, int sp, int sr, int maxIter, float eps, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void meanShiftProc_gpu(const DevMem2Db& src, DevMem2Db dstr, DevMem2Db dstsp, int sp, int sr, int maxIter, float eps, cudaStream_t stream);
}
}}}
void cv::gpu::meanShiftProc(const GpuMat& src, GpuMat& dstr, GpuMat& dstsp, int sp, int sr, TermCriteria criteria, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
if( src.empty() )
CV_Error( CV_StsBadArg, "The input image is empty" );
@ -235,22 +232,21 @@ void cv::gpu::meanShiftProc(const GpuMat& src, GpuMat& dstr, GpuMat& dstsp, int
////////////////////////////////////////////////////////////////////////
// drawColorDisp
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void drawColorDisp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int ndisp, const cudaStream_t& stream);
void drawColorDisp_gpu(const DevMem2D_<short>& src, const DevMem2Db& dst, int ndisp, const cudaStream_t& stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void drawColorDisp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int ndisp, const cudaStream_t& stream);
void drawColorDisp_gpu(const DevMem2D_<short>& src, const DevMem2Db& dst, int ndisp, const cudaStream_t& stream);
}
}}}
namespace
{
template <typename T>
void drawColorDisp_caller(const GpuMat& src, GpuMat& dst, int ndisp, const cudaStream_t& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
dst.create(src.size(), CV_8UC4);
@ -272,22 +268,21 @@ void cv::gpu::drawColorDisp(const GpuMat& src, GpuMat& dst, int ndisp, Stream& s
////////////////////////////////////////////////////////////////////////
// reprojectImageTo3D
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void reprojectImageTo3D_gpu(const DevMem2Db& disp, const DevMem2Df& xyzw, const float* q, const cudaStream_t& stream);
void reprojectImageTo3D_gpu(const DevMem2D_<short>& disp, const DevMem2Df& xyzw, const float* q, const cudaStream_t& stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void reprojectImageTo3D_gpu(const DevMem2Db& disp, const DevMem2Df& xyzw, const float* q, const cudaStream_t& stream);
void reprojectImageTo3D_gpu(const DevMem2D_<short>& disp, const DevMem2Df& xyzw, const float* q, const cudaStream_t& stream);
}
}}}
namespace
{
template <typename T>
void reprojectImageTo3D_caller(const GpuMat& disp, GpuMat& xyzw, const Mat& Q, const cudaStream_t& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
xyzw.create(disp.rows, disp.cols, CV_32FC4);
@ -309,14 +304,13 @@ void cv::gpu::reprojectImageTo3D(const GpuMat& disp, GpuMat& xyzw, const Mat& Q,
////////////////////////////////////////////////////////////////////////
// resize
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
template <typename T> void resize_gpu(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
template <typename T> void resize_gpu(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
}
}}}
void cv::gpu::resize(const GpuMat& src, GpuMat& dst, Size dsize, double fx, double fy, int interpolation, Stream& s)
{
@ -380,7 +374,7 @@ void cv::gpu::resize(const GpuMat& src, GpuMat& dst, Size dsize, double fx, doub
}
else
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
typedef void (*caller_t)(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
static const caller_t callers[6][4] =
@ -400,20 +394,19 @@ void cv::gpu::resize(const GpuMat& src, GpuMat& dst, Size dsize, double fx, doub
////////////////////////////////////////////////////////////////////////
// copyMakeBorder
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
template <typename T, int cn> void copyMakeBorder_gpu(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const T* borderValue, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
template <typename T, int cn> void copyMakeBorder_gpu(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const T* borderValue, cudaStream_t stream);
}
}}}
namespace
{
template <typename T, int cn> void copyMakeBorder_caller(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderType, const Scalar& value, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
Scalar_<T> val(saturate_cast<T>(value[0]), saturate_cast<T>(value[1]), saturate_cast<T>(value[2]), saturate_cast<T>(value[3]));
@ -666,21 +659,20 @@ void cv::gpu::warpPerspective(const GpuMat& src, GpuMat& dst, const Mat& M, Size
//////////////////////////////////////////////////////////////////////////////
// buildWarpPlaneMaps
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void buildWarpPlaneMaps(int tl_u, int tl_v, DevMem2Df map_x, DevMem2Df map_y,
const float k_rinv[9], const float r_kinv[9], const float t[3], float scale,
cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void buildWarpPlaneMaps(int tl_u, int tl_v, DevMem2Df map_x, DevMem2Df map_y,
const float k_rinv[9], const float r_kinv[9], const float t[3], float scale,
cudaStream_t stream);
}
}}}
void cv::gpu::buildWarpPlaneMaps(Size src_size, Rect dst_roi, const Mat &K, const Mat& R, const Mat &T,
float scale, GpuMat& map_x, GpuMat& map_y, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
CV_Assert(K.size() == Size(3,3) && K.type() == CV_32F);
CV_Assert(R.size() == Size(3,3) && R.type() == CV_32F);
@ -700,21 +692,20 @@ void cv::gpu::buildWarpPlaneMaps(Size src_size, Rect dst_roi, const Mat &K, cons
//////////////////////////////////////////////////////////////////////////////
// buildWarpCylyndricalMaps
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void buildWarpCylindricalMaps(int tl_u, int tl_v, DevMem2Df map_x, DevMem2Df map_y,
const float k_rinv[9], const float r_kinv[9], float scale,
cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void buildWarpCylindricalMaps(int tl_u, int tl_v, DevMem2Df map_x, DevMem2Df map_y,
const float k_rinv[9], const float r_kinv[9], float scale,
cudaStream_t stream);
}
}}}
void cv::gpu::buildWarpCylindricalMaps(Size src_size, Rect dst_roi, const Mat &K, const Mat& R, float scale,
GpuMat& map_x, GpuMat& map_y, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
CV_Assert(K.size() == Size(3,3) && K.type() == CV_32F);
CV_Assert(R.size() == Size(3,3) && R.type() == CV_32F);
@ -733,21 +724,20 @@ void cv::gpu::buildWarpCylindricalMaps(Size src_size, Rect dst_roi, const Mat &K
//////////////////////////////////////////////////////////////////////////////
// buildWarpSphericalMaps
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void buildWarpSphericalMaps(int tl_u, int tl_v, DevMem2Df map_x, DevMem2Df map_y,
const float k_rinv[9], const float r_kinv[9], float scale,
cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void buildWarpSphericalMaps(int tl_u, int tl_v, DevMem2Df map_x, DevMem2Df map_y,
const float k_rinv[9], const float r_kinv[9], float scale,
cudaStream_t stream);
}
}}}
void cv::gpu::buildWarpSphericalMaps(Size src_size, Rect dst_roi, const Mat &K, const Mat& R, float scale,
GpuMat& map_x, GpuMat& map_y, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
CV_Assert(K.size() == Size(3,3) && K.type() == CV_32F);
CV_Assert(R.size() == Size(3,3) && R.type() == CV_32F);
@ -899,18 +889,17 @@ void cv::gpu::sqrIntegral(const GpuMat& src, GpuMat& sqsum, Stream& s)
//////////////////////////////////////////////////////////////////////////////
// columnSum
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void columnSum_32F(const DevMem2Db src, const DevMem2Db dst);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void columnSum_32F(const DevMem2Db src, const DevMem2Db dst);
}
}}}
void cv::gpu::columnSum(const GpuMat& src, GpuMat& dst)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
CV_Assert(src.type() == CV_32F);
@ -1245,19 +1234,18 @@ void cv::gpu::histRange(const GpuMat& src, GpuMat hist[4], const GpuMat levels[4
hist_callers[src.depth()](src, hist, levels, buf, StreamAccessor::getStream(stream));
}
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace hist
namespace cv { namespace gpu { namespace device
{
void histogram256_gpu(DevMem2Db src, int* hist, unsigned int* buf, cudaStream_t stream);
const int PARTIAL_HISTOGRAM256_COUNT = 240;
const int HISTOGRAM256_BIN_COUNT = 256;
namespace hist
{
void histogram256_gpu(DevMem2Db src, int* hist, unsigned int* buf, cudaStream_t stream);
void equalizeHist_gpu(DevMem2Db src, DevMem2Db dst, const int* lut, cudaStream_t stream);
}
const int PARTIAL_HISTOGRAM256_COUNT = 240;
const int HISTOGRAM256_BIN_COUNT = 256;
END_OPENCV_DEVICE_NAMESPACE
void equalizeHist_gpu(DevMem2Db src, DevMem2Db dst, const int* lut, cudaStream_t stream);
}
}}}
void cv::gpu::calcHist(const GpuMat& src, GpuMat& hist, Stream& stream)
{
@ -1267,7 +1255,7 @@ void cv::gpu::calcHist(const GpuMat& src, GpuMat& hist, Stream& stream)
void cv::gpu::calcHist(const GpuMat& src, GpuMat& hist, GpuMat& buf, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ hist;
using namespace ::cv::gpu::device::hist;
CV_Assert(src.type() == CV_8UC1);
@ -1293,7 +1281,7 @@ void cv::gpu::equalizeHist(const GpuMat& src, GpuMat& dst, GpuMat& hist, Stream&
void cv::gpu::equalizeHist(const GpuMat& src, GpuMat& dst, GpuMat& hist, GpuMat& buf, Stream& s)
{
using namespace OPENCV_DEVICE_NAMESPACE_ hist;
using namespace ::cv::gpu::device::hist;
CV_Assert(src.type() == CV_8UC1);
@ -1327,16 +1315,15 @@ void cv::gpu::equalizeHist(const GpuMat& src, GpuMat& dst, GpuMat& hist, GpuMat&
////////////////////////////////////////////////////////////////////////
// cornerHarris & minEgenVal
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void extractCovData_caller(const DevMem2Df Dx, const DevMem2Df Dy, PtrStepf dst, cudaStream_t stream);
void cornerHarris_caller(const int block_size, const float k, const DevMem2Db Dx, const DevMem2Db Dy, DevMem2Db dst, int border_type, cudaStream_t stream);
void cornerMinEigenVal_caller(const int block_size, const DevMem2Db Dx, const DevMem2Db Dy, DevMem2Db dst, int border_type, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void extractCovData_caller(const DevMem2Df Dx, const DevMem2Df Dy, PtrStepf dst, cudaStream_t stream);
void cornerHarris_caller(const int block_size, const float k, const DevMem2Db Dx, const DevMem2Db Dy, DevMem2Db dst, int border_type, cudaStream_t stream);
void cornerMinEigenVal_caller(const int block_size, const DevMem2Db Dx, const DevMem2Db Dy, DevMem2Db dst, int border_type, cudaStream_t stream);
}
}}}
namespace
{
@ -1421,7 +1408,7 @@ void cv::gpu::cornerHarris(const GpuMat& src, GpuMat& dst, GpuMat& Dx, GpuMat& D
void cv::gpu::cornerHarris(const GpuMat& src, GpuMat& dst, GpuMat& Dx, GpuMat& Dy, GpuMat& buf, int blockSize, int ksize, double k, int borderType, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
CV_Assert(borderType == cv::BORDER_REFLECT101 ||
borderType == cv::BORDER_REPLICATE);
@ -1448,7 +1435,7 @@ void cv::gpu::cornerMinEigenVal(const GpuMat& src, GpuMat& dst, GpuMat& Dx, GpuM
void cv::gpu::cornerMinEigenVal(const GpuMat& src, GpuMat& dst, GpuMat& Dx, GpuMat& Dy, GpuMat& buf, int blockSize, int ksize, int borderType, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
CV_Assert(borderType == cv::BORDER_REFLECT101 ||
borderType == cv::BORDER_REPLICATE);
@ -1464,20 +1451,19 @@ void cv::gpu::cornerMinEigenVal(const GpuMat& src, GpuMat& dst, GpuMat& Dx, GpuM
//////////////////////////////////////////////////////////////////////////////
// mulSpectrums
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void mulSpectrums(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, DevMem2D_<cufftComplex> c, cudaStream_t stream);
void mulSpectrums_CONJ(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, DevMem2D_<cufftComplex> c, cudaStream_t stream);
}
namespace imgproc
{
void mulSpectrums(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, DevMem2D_<cufftComplex> c, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
void mulSpectrums_CONJ(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, DevMem2D_<cufftComplex> c, cudaStream_t stream);
}
}}}
void cv::gpu::mulSpectrums(const GpuMat& a, const GpuMat& b, GpuMat& c, int flags, bool conjB, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
typedef void (*Caller)(const PtrStep<cufftComplex>, const PtrStep<cufftComplex>, DevMem2D_<cufftComplex>, cudaStream_t stream);
@ -1495,20 +1481,19 @@ void cv::gpu::mulSpectrums(const GpuMat& a, const GpuMat& b, GpuMat& c, int flag
//////////////////////////////////////////////////////////////////////////////
// mulAndScaleSpectrums
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void mulAndScaleSpectrums(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, float scale, DevMem2D_<cufftComplex> c, cudaStream_t stream);
void mulAndScaleSpectrums_CONJ(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, float scale, DevMem2D_<cufftComplex> c, cudaStream_t stream);
}
namespace imgproc
{
void mulAndScaleSpectrums(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, float scale, DevMem2D_<cufftComplex> c, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
void mulAndScaleSpectrums_CONJ(const PtrStep<cufftComplex> a, const PtrStep<cufftComplex> b, float scale, DevMem2D_<cufftComplex> c, cudaStream_t stream);
}
}}}
void cv::gpu::mulAndScaleSpectrums(const GpuMat& a, const GpuMat& b, GpuMat& c, int flags, float scale, bool conjB, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
typedef void (*Caller)(const PtrStep<cufftComplex>, const PtrStep<cufftComplex>, float scale, DevMem2D_<cufftComplex>, cudaStream_t stream);
static Caller callers[] = { mulAndScaleSpectrums, mulAndScaleSpectrums_CONJ };
@ -1673,18 +1658,17 @@ void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
convolve(image, templ, result, ccorr, buf);
}
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
void convolve_gpu(const DevMem2Df& src, const PtrStepf& dst, int kWidth, int kHeight, float* kernel, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
void convolve_gpu(const DevMem2Df& src, const PtrStepf& dst, int kWidth, int kHeight, float* kernel, cudaStream_t stream);
}
}}}
void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result, bool ccorr, ConvolveBuf& buf, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
#ifndef HAVE_CUFFT
@ -1811,18 +1795,17 @@ void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
//////////////////////////////////////////////////////////////////////////////
// pyrDown
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
template <typename T, int cn> void pyrDown_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
template <typename T, int cn> void pyrDown_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
}
}}}
void cv::gpu::pyrDown(const GpuMat& src, GpuMat& dst, int borderType, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
typedef void (*func_t)(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
@ -1851,18 +1834,17 @@ void cv::gpu::pyrDown(const GpuMat& src, GpuMat& dst, int borderType, Stream& st
//////////////////////////////////////////////////////////////////////////////
// pyrUp
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc
namespace cv { namespace gpu { namespace device
{
template <typename T, int cn> void pyrUp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace imgproc
{
template <typename T, int cn> void pyrUp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
}
}}}
void cv::gpu::pyrUp(const GpuMat& src, GpuMat& dst, int borderType, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ imgproc;
using namespace ::cv::gpu::device::imgproc;
typedef void (*func_t)(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
@ -1933,31 +1915,30 @@ void cv::gpu::CannyBuf::release()
trackBuf2.release();
}
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace canny
namespace cv { namespace gpu { namespace device
{
void calcSobelRowPass_gpu(PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols);
void calcMagnitude_gpu(PtrStepi dx_buf, PtrStepi dy_buf, PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad);
void calcMagnitude_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad);
namespace canny
{
void calcSobelRowPass_gpu(PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols);
void calcMap_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh);
void edgesHysteresisLocal_gpu(PtrStepi map, ushort2* st1, int rows, int cols);
void calcMagnitude_gpu(PtrStepi dx_buf, PtrStepi dy_buf, PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad);
void calcMagnitude_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad);
void edgesHysteresisGlobal_gpu(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols);
void calcMap_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh);
void edgesHysteresisLocal_gpu(PtrStepi map, ushort2* st1, int rows, int cols);
void getEdges_gpu(PtrStepi map, PtrStepb dst, int rows, int cols);
}
void edgesHysteresisGlobal_gpu(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols);
END_OPENCV_DEVICE_NAMESPACE
void getEdges_gpu(PtrStepi map, PtrStepb dst, int rows, int cols);
}
}}}
namespace
{
void CannyCaller(CannyBuf& buf, GpuMat& dst, float low_thresh, float high_thresh)
{
using namespace OPENCV_DEVICE_NAMESPACE_ canny;
using namespace ::cv::gpu::device::canny;
calcMap_gpu(buf.dx, buf.dy, buf.edgeBuf, buf.edgeBuf, dst.rows, dst.cols, low_thresh, high_thresh);
@ -1977,7 +1958,7 @@ void cv::gpu::Canny(const GpuMat& src, GpuMat& dst, double low_thresh, double hi
void cv::gpu::Canny(const GpuMat& src, CannyBuf& buf, GpuMat& dst, double low_thresh, double high_thresh, int apperture_size, bool L2gradient)
{
using namespace OPENCV_DEVICE_NAMESPACE_ canny;
using namespace ::cv::gpu::device::canny;
CV_Assert(TargetArchs::builtWith(SHARED_ATOMICS) && DeviceInfo().supports(SHARED_ATOMICS));
CV_Assert(src.type() == CV_8UC1);
@ -2016,7 +1997,7 @@ void cv::gpu::Canny(const GpuMat& dx, const GpuMat& dy, GpuMat& dst, double low_
void cv::gpu::Canny(const GpuMat& dx, const GpuMat& dy, CannyBuf& buf, GpuMat& dst, double low_thresh, double high_thresh, bool L2gradient)
{
using namespace OPENCV_DEVICE_NAMESPACE_ canny;
using namespace ::cv::gpu::device::canny;
CV_Assert(TargetArchs::builtWith(SHARED_ATOMICS) && DeviceInfo().supports(SHARED_ATOMICS));
CV_Assert(dx.type() == CV_32SC1 && dy.type() == CV_32SC1 && dx.size() == dy.size());

29
modules/gpu/src/initialization.cpp
Unescape View File

@ -274,18 +274,17 @@ void cv::gpu::DeviceInfo::queryMemory(size_t& free_memory, size_t& total_memory)
////////////////////////////////////////////////////////////////////
// GpuFuncTable
BEGIN_OPENCV_DEVICE_NAMESPACE
void copy_to_with_mask(const DevMem2Db& src, DevMem2Db dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t& stream = 0);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
namespace cv { namespace gpu { namespace device
{
void copy_to_with_mask(const DevMem2Db& src, DevMem2Db dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t& stream = 0);
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta, cudaStream_t stream = 0);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta, cudaStream_t stream = 0);
}}}
namespace
{
@ -345,7 +344,7 @@ namespace
void convertToKernelCaller(const GpuMat& src, GpuMat& dst)
{
OPENCV_DEVICE_NAMESPACE_ convert_gpu(src.reshape(1), src.depth(), dst.reshape(1), dst.depth(), 1.0, 0.0);
::cv::gpu::device::convert_gpu(src.reshape(1), src.depth(), dst.reshape(1), dst.depth(), 1.0, 0.0);
}
//////////////////////////////////////////////////////////////////////////
@ -403,7 +402,7 @@ namespace
void kernelSet(GpuMat& src, Scalar s)
{
Scalar_<T> sf = s;
OPENCV_DEVICE_NAMESPACE_ set_to_gpu(src, sf.val, src.channels(), 0);
::cv::gpu::device::set_to_gpu(src, sf.val, src.channels(), 0);
}
template<int SDEPTH, int SCN> struct NppSetMaskFunc
@ -458,7 +457,7 @@ namespace
void kernelSetMask(GpuMat& src, Scalar s, const GpuMat& mask)
{
Scalar_<T> sf = s;
OPENCV_DEVICE_NAMESPACE_ set_to_gpu(src, sf.val, mask, src.channels(), 0);
::cv::gpu::device::set_to_gpu(src, sf.val, mask, src.channels(), 0);
}
class CudaFuncTable : public GpuFuncTable
@ -479,7 +478,7 @@ namespace
void copyWithMask(const GpuMat& src, GpuMat& dst, const GpuMat& mask) const
{
OPENCV_DEVICE_NAMESPACE_ copy_to_with_mask(src, dst, src.depth(), mask, src.channels());
::cv::gpu::device::copy_to_with_mask(src, dst, src.depth(), mask, src.channels());
}
void convert(const GpuMat& src, GpuMat& dst) const
@ -560,7 +559,7 @@ namespace
void convert(const GpuMat& src, GpuMat& dst, double alpha, double beta) const
{
device::convert_gpu(src.reshape(1), src.depth(), dst.reshape(1), dst.depth(), alpha, beta);
::cv::gpu::device::convert_gpu(src.reshape(1), src.depth(), dst.reshape(1), dst.depth(), alpha, beta);
}
void setTo(GpuMat& m, Scalar s, const GpuMat& mask) const

163
modules/gpu/src/match_template.cpp
Unescape View File

@ -52,94 +52,93 @@ void cv::gpu::matchTemplate(const GpuMat&, const GpuMat&, GpuMat&, int, Stream&)
#else
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace match_template
namespace cv { namespace gpu { namespace device
{
void matchTemplateNaive_CCORR_8U(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplateNaive_CCORR_32F(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplateNaive_SQDIFF_8U(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplateNaive_SQDIFF_32F(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result,
int cn, cudaStream_t stream);
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result,
int cn, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8U(int w, int h, const DevMem2D_<unsigned int> image_sum, unsigned int templ_sum, DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8UC2(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r,
const DevMem2D_<unsigned int> image_sum_g,
unsigned int templ_sum_r,
unsigned int templ_sum_g,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8UC3(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r,
const DevMem2D_<unsigned int> image_sum_g,
const DevMem2D_<unsigned int> image_sum_b,
unsigned int templ_sum_r,
unsigned int templ_sum_g,
unsigned int templ_sum_b,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8UC4(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r,
const DevMem2D_<unsigned int> image_sum_g,
const DevMem2D_<unsigned int> image_sum_b,
const DevMem2D_<unsigned int> image_sum_a,
unsigned int templ_sum_r,
unsigned int templ_sum_g,
unsigned int templ_sum_b,
unsigned int templ_sum_a,
DevMem2Df result, cudaStream_t stream);
namespace match_template
{
void matchTemplateNaive_CCORR_8U(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplateNaive_CCORR_32F(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8U(
int w, int h, const DevMem2D_<unsigned int> image_sum,
const DevMem2D_<unsigned long long> image_sqsum,
unsigned int templ_sum, unsigned int templ_sqsum,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8UC2(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r, const DevMem2D_<unsigned long long> image_sqsum_r,
const DevMem2D_<unsigned int> image_sum_g, const DevMem2D_<unsigned long long> image_sqsum_g,
unsigned int templ_sum_r, unsigned int templ_sqsum_r,
unsigned int templ_sum_g, unsigned int templ_sqsum_g,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8UC3(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r, const DevMem2D_<unsigned long long> image_sqsum_r,
const DevMem2D_<unsigned int> image_sum_g, const DevMem2D_<unsigned long long> image_sqsum_g,
const DevMem2D_<unsigned int> image_sum_b, const DevMem2D_<unsigned long long> image_sqsum_b,
unsigned int templ_sum_r, unsigned int templ_sqsum_r,
unsigned int templ_sum_g, unsigned int templ_sqsum_g,
unsigned int templ_sum_b, unsigned int templ_sqsum_b,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8UC4(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r, const DevMem2D_<unsigned long long> image_sqsum_r,
const DevMem2D_<unsigned int> image_sum_g, const DevMem2D_<unsigned long long> image_sqsum_g,
const DevMem2D_<unsigned int> image_sum_b, const DevMem2D_<unsigned long long> image_sqsum_b,
const DevMem2D_<unsigned int> image_sum_a, const DevMem2D_<unsigned long long> image_sqsum_a,
unsigned int templ_sum_r, unsigned int templ_sqsum_r,
unsigned int templ_sum_g, unsigned int templ_sqsum_g,
unsigned int templ_sum_b, unsigned int templ_sqsum_b,
unsigned int templ_sum_a, unsigned int templ_sqsum_a,
DevMem2Df result, cudaStream_t stream);
void matchTemplateNaive_SQDIFF_8U(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplateNaive_SQDIFF_32F(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void normalize_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum,
unsigned int templ_sqsum, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result,
int cn, cudaStream_t stream);
void extractFirstChannel_32F(const DevMem2Db image, DevMem2Df result, int cn, cudaStream_t stream);
}
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result,
int cn, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
void matchTemplatePrepared_CCOFF_8U(int w, int h, const DevMem2D_<unsigned int> image_sum, unsigned int templ_sum, DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8UC2(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r,
const DevMem2D_<unsigned int> image_sum_g,
unsigned int templ_sum_r,
unsigned int templ_sum_g,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8UC3(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r,
const DevMem2D_<unsigned int> image_sum_g,
const DevMem2D_<unsigned int> image_sum_b,
unsigned int templ_sum_r,
unsigned int templ_sum_g,
unsigned int templ_sum_b,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8UC4(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r,
const DevMem2D_<unsigned int> image_sum_g,
const DevMem2D_<unsigned int> image_sum_b,
const DevMem2D_<unsigned int> image_sum_a,
unsigned int templ_sum_r,
unsigned int templ_sum_g,
unsigned int templ_sum_b,
unsigned int templ_sum_a,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8U(
int w, int h, const DevMem2D_<unsigned int> image_sum,
const DevMem2D_<unsigned long long> image_sqsum,
unsigned int templ_sum, unsigned int templ_sqsum,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8UC2(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r, const DevMem2D_<unsigned long long> image_sqsum_r,
const DevMem2D_<unsigned int> image_sum_g, const DevMem2D_<unsigned long long> image_sqsum_g,
unsigned int templ_sum_r, unsigned int templ_sqsum_r,
unsigned int templ_sum_g, unsigned int templ_sqsum_g,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8UC3(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r, const DevMem2D_<unsigned long long> image_sqsum_r,
const DevMem2D_<unsigned int> image_sum_g, const DevMem2D_<unsigned long long> image_sqsum_g,
const DevMem2D_<unsigned int> image_sum_b, const DevMem2D_<unsigned long long> image_sqsum_b,
unsigned int templ_sum_r, unsigned int templ_sqsum_r,
unsigned int templ_sum_g, unsigned int templ_sqsum_g,
unsigned int templ_sum_b, unsigned int templ_sqsum_b,
DevMem2Df result, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_NORMED_8UC4(
int w, int h,
const DevMem2D_<unsigned int> image_sum_r, const DevMem2D_<unsigned long long> image_sqsum_r,
const DevMem2D_<unsigned int> image_sum_g, const DevMem2D_<unsigned long long> image_sqsum_g,
const DevMem2D_<unsigned int> image_sum_b, const DevMem2D_<unsigned long long> image_sqsum_b,
const DevMem2D_<unsigned int> image_sum_a, const DevMem2D_<unsigned long long> image_sqsum_a,
unsigned int templ_sum_r, unsigned int templ_sqsum_r,
unsigned int templ_sum_g, unsigned int templ_sqsum_g,
unsigned int templ_sum_b, unsigned int templ_sqsum_b,
unsigned int templ_sum_a, unsigned int templ_sqsum_a,
DevMem2Df result, cudaStream_t stream);
void normalize_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum,
unsigned int templ_sqsum, DevMem2Df result, int cn, cudaStream_t stream);
void extractFirstChannel_32F(const DevMem2Db image, DevMem2Df result, int cn, cudaStream_t stream);
}
}}}
using namespace OPENCV_DEVICE_NAMESPACE_ match_template;
using namespace ::cv::gpu::device::match_template;
namespace
{

162
modules/gpu/src/matrix_reductions.cpp
Unescape View File

@ -190,35 +190,34 @@ double cv::gpu::norm(const GpuMat& src1, const GpuMat& src2, int normType)
////////////////////////////////////////////////////////////////////////
// Sum
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace matrix_reductions
namespace cv { namespace gpu { namespace device
{
namespace sum
namespace matrix_reductions
{
template <typename T>
void sumCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
namespace sum
{
template <typename T>
void sumCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void sumMultipassCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void sumMultipassCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void absSumCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void absSumCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void absSumMultipassCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void absSumMultipassCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void sqrSumCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void sqrSumCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void sqrSumMultipassCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
template <typename T>
void sqrSumMultipassCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn);
void getBufSizeRequired(int cols, int rows, int cn, int& bufcols, int& bufrows);
void getBufSizeRequired(int cols, int rows, int cn, int& bufcols, int& bufrows);
}
}
}
END_OPENCV_DEVICE_NAMESPACE
}}}
Scalar cv::gpu::sum(const GpuMat& src)
{
@ -229,7 +228,7 @@ Scalar cv::gpu::sum(const GpuMat& src)
Scalar cv::gpu::sum(const GpuMat& src, GpuMat& buf)
{
using namespace OPENCV_DEVICE_NAMESPACE_ matrix_reductions::sum;
using namespace ::cv::gpu::device::matrix_reductions::sum;
typedef void (*Caller)(const DevMem2Db, PtrStepb, double*, int);
@ -272,7 +271,7 @@ Scalar cv::gpu::absSum(const GpuMat& src)
Scalar cv::gpu::absSum(const GpuMat& src, GpuMat& buf)
{
using namespace OPENCV_DEVICE_NAMESPACE_ matrix_reductions::sum;
using namespace ::cv::gpu::device::matrix_reductions::sum;
typedef void (*Caller)(const DevMem2Db, PtrStepb, double*, int);
@ -316,7 +315,7 @@ Scalar cv::gpu::sqrSum(const GpuMat& src)
Scalar cv::gpu::sqrSum(const GpuMat& src, GpuMat& buf)
{
using namespace OPENCV_DEVICE_NAMESPACE_ matrix_reductions::sum;
using namespace ::cv::gpu::device::matrix_reductions::sum;
typedef void (*Caller)(const DevMem2Db, PtrStepb, double*, int);
@ -353,29 +352,28 @@ Scalar cv::gpu::sqrSum(const GpuMat& src, GpuMat& buf)
////////////////////////////////////////////////////////////////////////
// Find min or max
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace matrix_reductions
namespace cv { namespace gpu { namespace device
{
namespace minmax
namespace matrix_reductions
{
void getBufSizeRequired(int cols, int rows, int elem_size, int& bufcols, int& bufrows);
template <typename T>
void minMaxCaller(const DevMem2Db src, double* minval, double* maxval, PtrStepb buf);
namespace minmax
{
void getBufSizeRequired(int cols, int rows, int elem_size, int& bufcols, int& bufrows);
template <typename T>
void minMaxCaller(const DevMem2Db src, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMaskCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMaskCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMultipassCaller(const DevMem2Db src, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMultipassCaller(const DevMem2Db src, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMaskMultipassCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMaskMultipassCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
}
}
}
END_OPENCV_DEVICE_NAMESPACE
}}}
void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const GpuMat& mask)
@ -387,7 +385,7 @@ void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const Gp
void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const GpuMat& mask, GpuMat& buf)
{
using namespace OPENCV_DEVICE_NAMESPACE_ matrix_reductions::minmax;
using namespace ::cv::gpu::device::matrix_reductions::minmax;
typedef void (*Caller)(const DevMem2Db, double*, double*, PtrStepb);
typedef void (*MaskedCaller)(const DevMem2Db, const PtrStepb, double*, double*, PtrStepb);
@ -457,34 +455,33 @@ void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const Gp
////////////////////////////////////////////////////////////////////////
// Locate min and max
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace matrix_reductions
namespace cv { namespace gpu { namespace device
{
namespace minmaxloc
namespace matrix_reductions
{
void getBufSizeRequired(int cols, int rows, int elem_size, int& b1cols,
int& b1rows, int& b2cols, int& b2rows);
template <typename T>
void minMaxLocCaller(const DevMem2Db src, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
namespace minmaxloc
{
void getBufSizeRequired(int cols, int rows, int elem_size, int& b1cols,
int& b1rows, int& b2cols, int& b2rows);
template <typename T>
void minMaxLocMaskCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval,
template <typename T>
void minMaxLocCaller(const DevMem2Db src, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
template <typename T>
void minMaxLocMultipassCaller(const DevMem2Db src, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
template <typename T>
void minMaxLocMaskCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
template <typename T>
void minMaxLocMaskMultipassCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval,
template <typename T>
void minMaxLocMultipassCaller(const DevMem2Db src, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
}
}
END_OPENCV_DEVICE_NAMESPACE
template <typename T>
void minMaxLocMaskMultipassCaller(const DevMem2Db src, const PtrStepb mask, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
}
}
}}}
void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point* minLoc, Point* maxLoc, const GpuMat& mask)
{
@ -495,7 +492,7 @@ void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point
void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point* minLoc, Point* maxLoc,
const GpuMat& mask, GpuMat& valBuf, GpuMat& locBuf)
{
using namespace OPENCV_DEVICE_NAMESPACE_ matrix_reductions::minmaxloc;
using namespace ::cv::gpu::device::matrix_reductions::minmaxloc;
typedef void (*Caller)(const DevMem2Db, double*, double*, int[2], int[2], PtrStepb, PtrStepb);
typedef void (*MaskedCaller)(const DevMem2Db, const PtrStepb, double*, double*, int[2], int[2], PtrStepb, PtrStepb);
@ -571,23 +568,22 @@ void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point
//////////////////////////////////////////////////////////////////////////////
// Count non-zero elements
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace matrix_reductions
namespace cv { namespace gpu { namespace device
{
namespace countnonzero
namespace matrix_reductions
{
void getBufSizeRequired(int cols, int rows, int& bufcols, int& bufrows);
namespace countnonzero
{
void getBufSizeRequired(int cols, int rows, int& bufcols, int& bufrows);
template <typename T>
int countNonZeroCaller(const DevMem2Db src, PtrStepb buf);
template <typename T>
int countNonZeroCaller(const DevMem2Db src, PtrStepb buf);
template <typename T>
int countNonZeroMultipassCaller(const DevMem2Db src, PtrStepb buf);
template <typename T>
int countNonZeroMultipassCaller(const DevMem2Db src, PtrStepb buf);
}
}
}
END_OPENCV_DEVICE_NAMESPACE
}}}
int cv::gpu::countNonZero(const GpuMat& src)
{
@ -598,7 +594,7 @@ int cv::gpu::countNonZero(const GpuMat& src)
int cv::gpu::countNonZero(const GpuMat& src, GpuMat& buf)
{
using namespace OPENCV_DEVICE_NAMESPACE_ matrix_reductions::countnonzero;
using namespace ::cv::gpu::device::matrix_reductions::countnonzero;
typedef int (*Caller)(const DevMem2Db src, PtrStepb buf);
@ -632,19 +628,19 @@ int cv::gpu::countNonZero(const GpuMat& src, GpuMat& buf)
//////////////////////////////////////////////////////////////////////////////
// reduce
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace matrix_reductions
namespace cv { namespace gpu { namespace device
{
template <typename T, typename S, typename D> void reduceRows_gpu(const DevMem2Db& src, const DevMem2Db& dst, int reduceOp, cudaStream_t stream);
template <typename T, typename S, typename D> void reduceCols_gpu(const DevMem2Db& src, int cn, const DevMem2Db& dst, int reduceOp, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace matrix_reductions
{
template <typename T, typename S, typename D> void reduceRows_gpu(const DevMem2Db& src, const DevMem2Db& dst, int reduceOp, cudaStream_t stream);
template <typename T, typename S, typename D> void reduceCols_gpu(const DevMem2Db& src, int cn, const DevMem2Db& dst, int reduceOp, cudaStream_t stream);
}
}}}
void cv::gpu::reduce(const GpuMat& src, GpuMat& dst, int dim, int reduceOp, int dtype, Stream& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ matrix_reductions;
using namespace ::cv::gpu::device::matrix_reductions;
CV_Assert(src.depth() <= CV_32F && src.channels() <= 4 && dtype <= CV_32F);
CV_Assert(dim == 0 || dim == 1);

1141
modules/gpu/src/opencv2/gpu/device/border_interpolate.hpp
View File

File diff suppressed because it is too large Load Diff

345
modules/gpu/src/opencv2/gpu/device/color.hpp
Unescape View File

@ -46,178 +46,177 @@
#include "internal_shared.hpp"
#include "detail/color_detail.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
// All OPENCV_GPU_IMPLEMENT_*_TRAITS(ColorSpace1_to_ColorSpace2, ...) macros implements
// template <typename T> class ColorSpace1_to_ColorSpace2_traits
// {
// typedef ... functor_type;
// static __host__ __device__ functor_type create_functor();
// };
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgba, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr555, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr565, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr555, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr565, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr555, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr565, 4, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr555, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr565, 4, 2, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgb, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgb, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgr, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgr, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgba, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgba, 4, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgra, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgra, 4, 0, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgr, 3)
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgra, 4)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr555, 5)
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr565, 6)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr555_to_gray, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr565_to_gray, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgb_to_gray, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgr_to_gray, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgba_to_gray, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgra_to_gray, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv4, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv4, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgba, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgba, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgr, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgra, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgr, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgra, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS
END_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
// All OPENCV_GPU_IMPLEMENT_*_TRAITS(ColorSpace1_to_ColorSpace2, ...) macros implements
// template <typename T> class ColorSpace1_to_ColorSpace2_traits
// {
// typedef ... functor_type;
// static __host__ __device__ functor_type create_functor();
// };
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgba, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr555, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr565, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr555, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr565, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr555, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr565, 4, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr555, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr565, 4, 2, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgb, 3, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgb, 3, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgr, 3, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgr, 3, 0, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgba, 4, 2, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgba, 4, 2, 6)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgra, 4, 0, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgra, 4, 0, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgr, 3)
OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgra, 4)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr555, 5)
OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr565, 6)
#undef OPENCV_GPU_IMPLEMENT_GRAY2RGB5x5_TRAITS
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr555_to_gray, 5)
OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr565_to_gray, 6)
#undef OPENCV_GPU_IMPLEMENT_RGB5x52GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgb_to_gray, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgr_to_gray, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(rgba_to_gray, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS(bgra_to_gray, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2GRAY_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv4, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv4, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_RGB2YUV_TRAITS
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgba, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgba, 4, 4, 0)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgr, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgra, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgr, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgra, 4, 4, 2)
#undef OPENCV_GPU_IMPLEMENT_YUV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2YCrCb_TRAITS
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_YCrCb2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2XYZ_TRAITS
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_XYZ2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HSV_TRAITS
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HSV2RGB_TRAITS
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls4, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls4, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls4, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls4, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_RGB2HLS_TRAITS
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgb, 3, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgba, 3, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgb, 4, 3, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgba, 4, 4, 2)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgr, 3, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgra, 3, 4, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgr, 4, 3, 0)
OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgra, 4, 4, 0)
#undef OPENCV_GPU_IMPLEMENT_HLS2RGB_TRAITS
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_BORDER_INTERPOLATE_HPP__

46
modules/gpu/src/opencv2/gpu/device/datamov_utils.hpp
Unescape View File

@ -45,16 +45,8 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#if defined(_WIN64) || defined(__LP64__)
// 64-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "l"
#else
// 32-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "r"
#endif
namespace cv { namespace gpu { namespace device
{
#if __CUDA_ARCH__ >= 200
// for Fermi memory space is detected automatically
@ -65,6 +57,14 @@ BEGIN_OPENCV_DEVICE_NAMESPACE
#else // __CUDA_ARCH__ >= 200
#if defined(_WIN64) || defined(__LP64__)
// 64-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "l"
#else
// 32-bit register modifier for inlined asm
#define OPENCV_GPU_ASM_PTR "r"
#endif
template<class T> struct ForceGlob;
#define OPENCV_GPU_DEFINE_FORCE_GLOB(base_type, ptx_type, reg_mod) \
@ -85,21 +85,21 @@ BEGIN_OPENCV_DEVICE_NAMESPACE
} \
};
OPENCV_GPU_DEFINE_FORCE_GLOB_B(uchar, u8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(schar, s8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(char, b8)
OPENCV_GPU_DEFINE_FORCE_GLOB (ushort, u16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (short, s16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (uint, u32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (int, s32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (float, f32, f)
OPENCV_GPU_DEFINE_FORCE_GLOB (double, f64, d)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(uchar, u8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(schar, s8)
OPENCV_GPU_DEFINE_FORCE_GLOB_B(char, b8)
OPENCV_GPU_DEFINE_FORCE_GLOB (ushort, u16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (short, s16, h)
OPENCV_GPU_DEFINE_FORCE_GLOB (uint, u32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (int, s32, r)
OPENCV_GPU_DEFINE_FORCE_GLOB (float, f32, f)
OPENCV_GPU_DEFINE_FORCE_GLOB (double, f64, d)
#undef OPENCV_GPU_DEFINE_FORCE_GLOB
#undef OPENCV_GPU_DEFINE_FORCE_GLOB_B
#undef OPENCV_GPU_DEFINE_FORCE_GLOB
#undef OPENCV_GPU_DEFINE_FORCE_GLOB_B
#undef OPENCV_GPU_ASM_PTR
#endif // __CUDA_ARCH__ >= 200
END_OPENCV_DEVICE_NAMESPACE
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_DATAMOV_UTILS_HPP__

1787
modules/gpu/src/opencv2/gpu/device/detail/color_detail.hpp
View File

File diff suppressed because it is too large Load Diff

577
modules/gpu/src/opencv2/gpu/device/detail/transform_detail.hpp
Unescape View File

@ -47,365 +47,348 @@
#include "../vec_traits.hpp"
#include "../functional.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace detail
namespace cv { namespace gpu { namespace device
{
//! Mask accessor
struct MaskReader
{
explicit MaskReader(const PtrStepb& mask_): mask(mask_) {}
__device__ __forceinline__ bool operator()(int y, int x) const { return mask.ptr(y)[x]; }
const PtrStepb mask;
};
struct NoMask
namespace transform_detail
{
__device__ __forceinline__ bool operator()(int y, int x) const { return true; }
};
//! Read Write Traits
//! Read Write Traits
template <typename T, typename D, int shift> struct UnaryReadWriteTraits
{
typedef typename TypeVec<T, shift>::vec_type read_type;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
template <typename T, typename D, int shift> struct UnaryReadWriteTraits
{
typedef typename TypeVec<T, shift>::vec_type read_type;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
template <typename T1, typename T2, typename D, int shift> struct BinaryReadWriteTraits
{
typedef typename TypeVec<T1, shift>::vec_type read_type1;
typedef typename TypeVec<T2, shift>::vec_type read_type2;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
template <typename T1, typename T2, typename D, int shift> struct BinaryReadWriteTraits
{
typedef typename TypeVec<T1, shift>::vec_type read_type1;
typedef typename TypeVec<T2, shift>::vec_type read_type2;
typedef typename TypeVec<D, shift>::vec_type write_type;
};
//! Transform kernels
//! Transform kernels
template <int shift> struct OpUnroller;
template <> struct OpUnroller<1>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
template <int shift> struct OpUnroller;
template <> struct OpUnroller<1>
{
if (mask(y, x_shifted))
dst.x = op(src.x);
}
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
}
};
template <> struct OpUnroller<2>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
}
};
template <> struct OpUnroller<2>
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
}
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
}
};
template <> struct OpUnroller<3>
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
}
};
template <> struct OpUnroller<3>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
}
};
template <> struct OpUnroller<4>
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
}
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
if (mask(y, x_shifted + 3))
dst.w = op(src.w);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
if (mask(y, x_shifted + 3))
dst.w = op(src1.w, src2.w);
}
};
template <> struct OpUnroller<8>
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
}
};
template <> struct OpUnroller<4>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.a0 = op(src.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src.a7);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.a0 = op(src1.a0, src2.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src1.a1, src2.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src1.a2, src2.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src1.a3, src2.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src1.a4, src2.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src1.a5, src2.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src1.a6, src2.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src1.a7, src2.a7);
}
};
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
__global__ static void transformSmart(const DevMem2D_<T> src_, PtrStep<D> dst_, const Mask mask, const UnOp op)
{
if (mask(y, x_shifted))
dst.x = op(src.x);
if (mask(y, x_shifted + 1))
dst.y = op(src.y);
if (mask(y, x_shifted + 2))
dst.z = op(src.z);
if (mask(y, x_shifted + 3))
dst.w = op(src.w);
}
typedef TransformFunctorTraits<UnOp> ft;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::read_type read_type;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::write_type write_type;
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
{
if (mask(y, x_shifted))
dst.x = op(src1.x, src2.x);
if (mask(y, x_shifted + 1))
dst.y = op(src1.y, src2.y);
if (mask(y, x_shifted + 2))
dst.z = op(src1.z, src2.z);
if (mask(y, x_shifted + 3))
dst.w = op(src1.w, src2.w);
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
if (y < src_.rows)
{
const T* src = src_.ptr(y);
D* dst = dst_.ptr(y);
if (x_shifted + ft::smart_shift - 1 < src_.cols)
{
const read_type src_n_el = ((const read_type*)src)[x];
write_type dst_n_el;
OpUnroller<ft::smart_shift>::unroll(src_n_el, dst_n_el, mask, op, x_shifted, y);
((write_type*)dst)[x] = dst_n_el;
}
else
{
for (int real_x = x_shifted; real_x < src_.cols; ++real_x)
{
if (mask(y, real_x))
dst[real_x] = op(src[real_x]);
}
}
}
}
};
template <> struct OpUnroller<8>
{
template <typename T, typename D, typename UnOp, typename Mask>
static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
static __global__ void transformSimple(const DevMem2D_<T> src, PtrStep<D> dst, const Mask mask, const UnOp op)
{
if (mask(y, x_shifted))
dst.a0 = op(src.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src.a7);
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < src.cols && y < src.rows && mask(y, x))
{
dst.ptr(y)[x] = op(src.ptr(y)[x]);
}
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
__global__ static void transformSmart(const DevMem2D_<T1> src1_, const PtrStep<T2> src2_, PtrStep<D> dst_,
const Mask mask, const BinOp op)
{
if (mask(y, x_shifted))
dst.a0 = op(src1.a0, src2.a0);
if (mask(y, x_shifted + 1))
dst.a1 = op(src1.a1, src2.a1);
if (mask(y, x_shifted + 2))
dst.a2 = op(src1.a2, src2.a2);
if (mask(y, x_shifted + 3))
dst.a3 = op(src1.a3, src2.a3);
if (mask(y, x_shifted + 4))
dst.a4 = op(src1.a4, src2.a4);
if (mask(y, x_shifted + 5))
dst.a5 = op(src1.a5, src2.a5);
if (mask(y, x_shifted + 6))
dst.a6 = op(src1.a6, src2.a6);
if (mask(y, x_shifted + 7))
dst.a7 = op(src1.a7, src2.a7);
}
};
template <typename T, typename D, typename UnOp, typename Mask>
__global__ static void transformSmart(const DevMem2D_<T> src_, PtrStep<D> dst_, const Mask mask, const UnOp op)
{
typedef TransformFunctorTraits<UnOp> ft;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::read_type read_type;
typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::write_type write_type;
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
typedef TransformFunctorTraits<BinOp> ft;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type1 read_type1;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type2 read_type2;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::write_type write_type;
if (y < src_.rows)
{
const T* src = src_.ptr(y);
D* dst = dst_.ptr(y);
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
if (x_shifted + ft::smart_shift - 1 < src_.cols)
if (y < src1_.rows)
{
const read_type src_n_el = ((const read_type*)src)[x];
write_type dst_n_el;
const T1* src1 = src1_.ptr(y);
const T2* src2 = src2_.ptr(y);
D* dst = dst_.ptr(y);
OpUnroller<ft::smart_shift>::unroll(src_n_el, dst_n_el, mask, op, x_shifted, y);
if (x_shifted + ft::smart_shift - 1 < src1_.cols)
{
const read_type1 src1_n_el = ((const read_type1*)src1)[x];
const read_type2 src2_n_el = ((const read_type2*)src2)[x];
write_type dst_n_el;
OpUnroller<ft::smart_shift>::unroll(src1_n_el, src2_n_el, dst_n_el, mask, op, x_shifted, y);
((write_type*)dst)[x] = dst_n_el;
}
else
{
for (int real_x = x_shifted; real_x < src_.cols; ++real_x)
((write_type*)dst)[x] = dst_n_el;
}
else
{
if (mask(y, real_x))
dst[real_x] = op(src[real_x]);
for (int real_x = x_shifted; real_x < src1_.cols; ++real_x)
{
if (mask(y, real_x))
dst[real_x] = op(src1[real_x], src2[real_x]);
}
}
}
}
}
template <typename T, typename D, typename UnOp, typename Mask>
static __global__ void transformSimple(const DevMem2D_<T> src, PtrStep<D> dst, const Mask mask, const UnOp op)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < src.cols && y < src.rows && mask(y, x))
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __global__ void transformSimple(const DevMem2D_<T1> src1, const PtrStep<T2> src2, PtrStep<D> dst,
const Mask mask, const BinOp op)
{
dst.ptr(y)[x] = op(src.ptr(y)[x]);
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < src1.cols && y < src1.rows && mask(y, x))
{
const T1 src1_data = src1.ptr(y)[x];
const T2 src2_data = src2.ptr(y)[x];
dst.ptr(y)[x] = op(src1_data, src2_data);
}
}
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
__global__ static void transformSmart(const DevMem2D_<T1> src1_, const PtrStep<T2> src2_, PtrStep<D> dst_,
const Mask mask, const BinOp op)
{
typedef TransformFunctorTraits<BinOp> ft;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type1 read_type1;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type2 read_type2;
typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::write_type write_type;
template <bool UseSmart> struct TransformDispatcher;
template<> struct TransformDispatcher<false>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
const int x = threadIdx.x + blockIdx.x * blockDim.x;
const int y = threadIdx.y + blockIdx.y * blockDim.y;
const int x_shifted = x * ft::smart_shift;
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x), divUp(src.rows, threads.y), 1);
if (y < src1_.rows)
{
const T1* src1 = src1_.ptr(y);
const T2* src2 = src2_.ptr(y);
D* dst = dst_.ptr(y);
transformSimple<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (x_shifted + ft::smart_shift - 1 < src1_.cols)
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, const Mask& mask, cudaStream_t stream)
{
const read_type1 src1_n_el = ((const read_type1*)src1)[x];
const read_type2 src2_n_el = ((const read_type2*)src2)[x];
write_type dst_n_el;
OpUnroller<ft::smart_shift>::unroll(src1_n_el, src2_n_el, dst_n_el, mask, op, x_shifted, y);
typedef TransformFunctorTraits<BinOp> ft;
((write_type*)dst)[x] = dst_n_el;
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x), divUp(src1.rows, threads.y), 1);
transformSimple<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
else
};
template<> struct TransformDispatcher<true>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, const Mask& mask, cudaStream_t stream)
{
for (int real_x = x_shifted; real_x < src1_.cols; ++real_x)
{
if (mask(y, real_x))
dst[real_x] = op(src1[real_x], src2[real_x]);
}
}
}
}
typedef TransformFunctorTraits<UnOp> ft;
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static __global__ void transformSimple(const DevMem2D_<T1> src1, const PtrStep<T2> src2, PtrStep<D> dst,
const Mask mask, const BinOp op)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
StaticAssert<ft::smart_shift != 1>::check();
if (x < src1.cols && y < src1.rows && mask(y, x))
{
const T1 src1_data = src1.ptr(y)[x];
const T2 src2_data = src2.ptr(y)[x];
dst.ptr(y)[x] = op(src1_data, src2_data);
}
}
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x * ft::smart_shift), divUp(src.rows, threads.y), 1);
template <bool UseSmart> struct TransformDispatcher;
template<> struct TransformDispatcher<false>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
transformSmart<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x), divUp(src.rows, threads.y), 1);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
transformSimple<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
StaticAssert<ft::smart_shift != 1>::check();
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x * ft::smart_shift), divUp(src1.rows, threads.y), 1);
const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x), divUp(src1.rows, threads.y), 1);
transformSmart<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
transformSimple<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template<> struct TransformDispatcher<true>
{
template <typename T, typename D, typename UnOp, typename Mask>
static void call(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, const Mask& mask, cudaStream_t stream)
static void transform_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
StaticAssert<ft::smart_shift != 1>::check();
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src.cols, threads.x * ft::smart_shift), divUp(src.rows, threads.y), 1);
transformSmart<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
TransformDispatcher<VecTraits<T>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src, dst, op, mask, stream);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void call(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, const Mask& mask, cudaStream_t stream)
static void transform_caller(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
StaticAssert<ft::smart_shift != 1>::check();
const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
const dim3 grid(divUp(src1.cols, threads.x * ft::smart_shift), divUp(src1.rows, threads.y), 1);
transformSmart<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
TransformDispatcher<VecTraits<T1>::cn == 1 && VecTraits<T2>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src1, src2, dst, op, mask, stream);
}
};
template <typename T, typename D, typename UnOp, typename Mask>
static void transform_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<UnOp> ft;
TransformDispatcher<VecTraits<T>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src, dst, op, mask, stream);
}
template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
static void transform_caller(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, const Mask& mask, cudaStream_t stream)
{
typedef TransformFunctorTraits<BinOp> ft;
TransformDispatcher<VecTraits<T1>::cn == 1 && VecTraits<T2>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src1, src2, dst, op, mask, stream);
}
}
END_OPENCV_DEVICE_NAMESPACE
} // namespace transform_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_TRANSFORM_DETAIL_HPP__

257
modules/gpu/src/opencv2/gpu/device/detail/type_traits_detail.hpp
Unescape View File

@ -46,143 +46,142 @@
#include "internal_shared.hpp"
#include "../vec_traits.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace detail
namespace cv { namespace gpu { namespace device
{
template <bool, typename T1, typename T2> struct Select { typedef T1 type; };
template <typename T1, typename T2> struct Select<false, T1, T2> { typedef T2 type; };
template <typename T> struct IsSignedIntergral { enum {value = 0}; };
template <> struct IsSignedIntergral<schar> { enum {value = 1}; };
template <> struct IsSignedIntergral<char1> { enum {value = 1}; };
template <> struct IsSignedIntergral<short> { enum {value = 1}; };
template <> struct IsSignedIntergral<short1> { enum {value = 1}; };
template <> struct IsSignedIntergral<int> { enum {value = 1}; };
template <> struct IsSignedIntergral<int1> { enum {value = 1}; };
namespace type_traits_detail
{
template <bool, typename T1, typename T2> struct Select { typedef T1 type; };
template <typename T1, typename T2> struct Select<false, T1, T2> { typedef T2 type; };
template <typename T> struct IsUnsignedIntegral { enum {value = 0}; };
template <> struct IsUnsignedIntegral<uchar> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uchar1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint1> { enum {value = 1}; };
template <typename T> struct IsSignedIntergral { enum {value = 0}; };
template <> struct IsSignedIntergral<schar> { enum {value = 1}; };
template <> struct IsSignedIntergral<char1> { enum {value = 1}; };
template <> struct IsSignedIntergral<short> { enum {value = 1}; };
template <> struct IsSignedIntergral<short1> { enum {value = 1}; };
template <> struct IsSignedIntergral<int> { enum {value = 1}; };
template <> struct IsSignedIntergral<int1> { enum {value = 1}; };
template <typename T> struct IsIntegral { enum {value = IsSignedIntergral<T>::value || IsUnsignedIntegral<T>::value}; };
template <> struct IsIntegral<char> { enum {value = 1}; };
template <> struct IsIntegral<bool> { enum {value = 1}; };
template <typename T> struct IsUnsignedIntegral { enum {value = 0}; };
template <> struct IsUnsignedIntegral<uchar> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uchar1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<ushort1> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint> { enum {value = 1}; };
template <> struct IsUnsignedIntegral<uint1> { enum {value = 1}; };
template <typename T> struct IsFloat { enum {value = 0}; };
template <> struct IsFloat<float> { enum {value = 1}; };
template <> struct IsFloat<double> { enum {value = 1}; };
template <typename T> struct IsIntegral { enum {value = IsSignedIntergral<T>::value || IsUnsignedIntegral<T>::value}; };
template <> struct IsIntegral<char> { enum {value = 1}; };
template <> struct IsIntegral<bool> { enum {value = 1}; };
template <typename T> struct IsVec { enum {value = 0}; };
template <> struct IsVec<uchar1> { enum {value = 1}; };
template <> struct IsVec<uchar2> { enum {value = 1}; };
template <> struct IsVec<uchar3> { enum {value = 1}; };
template <> struct IsVec<uchar4> { enum {value = 1}; };
template <> struct IsVec<uchar8> { enum {value = 1}; };
template <> struct IsVec<char1> { enum {value = 1}; };
template <> struct IsVec<char2> { enum {value = 1}; };
template <> struct IsVec<char3> { enum {value = 1}; };
template <> struct IsVec<char4> { enum {value = 1}; };
template <> struct IsVec<char8> { enum {value = 1}; };
template <> struct IsVec<ushort1> { enum {value = 1}; };
template <> struct IsVec<ushort2> { enum {value = 1}; };
template <> struct IsVec<ushort3> { enum {value = 1}; };
template <> struct IsVec<ushort4> { enum {value = 1}; };
template <> struct IsVec<ushort8> { enum {value = 1}; };
template <> struct IsVec<short1> { enum {value = 1}; };
template <> struct IsVec<short2> { enum {value = 1}; };
template <> struct IsVec<short3> { enum {value = 1}; };
template <> struct IsVec<short4> { enum {value = 1}; };
template <> struct IsVec<short8> { enum {value = 1}; };
template <> struct IsVec<uint1> { enum {value = 1}; };
template <> struct IsVec<uint2> { enum {value = 1}; };
template <> struct IsVec<uint3> { enum {value = 1}; };
template <> struct IsVec<uint4> { enum {value = 1}; };
template <> struct IsVec<uint8> { enum {value = 1}; };
template <> struct IsVec<int1> { enum {value = 1}; };
template <> struct IsVec<int2> { enum {value = 1}; };
template <> struct IsVec<int3> { enum {value = 1}; };
template <> struct IsVec<int4> { enum {value = 1}; };
template <> struct IsVec<int8> { enum {value = 1}; };
template <> struct IsVec<float1> { enum {value = 1}; };
template <> struct IsVec<float2> { enum {value = 1}; };
template <> struct IsVec<float3> { enum {value = 1}; };
template <> struct IsVec<float4> { enum {value = 1}; };
template <> struct IsVec<float8> { enum {value = 1}; };
template <> struct IsVec<double1> { enum {value = 1}; };
template <> struct IsVec<double2> { enum {value = 1}; };
template <> struct IsVec<double3> { enum {value = 1}; };
template <> struct IsVec<double4> { enum {value = 1}; };
template <> struct IsVec<double8> { enum {value = 1}; };
template <typename T> struct IsFloat { enum {value = 0}; };
template <> struct IsFloat<float> { enum {value = 1}; };
template <> struct IsFloat<double> { enum {value = 1}; };
template <class U> struct AddParameterType { typedef const U& type; };
template <class U> struct AddParameterType<U&> { typedef U& type; };
template <> struct AddParameterType<void> { typedef void type; };
template <typename T> struct IsVec { enum {value = 0}; };
template <> struct IsVec<uchar1> { enum {value = 1}; };
template <> struct IsVec<uchar2> { enum {value = 1}; };
template <> struct IsVec<uchar3> { enum {value = 1}; };
template <> struct IsVec<uchar4> { enum {value = 1}; };
template <> struct IsVec<uchar8> { enum {value = 1}; };
template <> struct IsVec<char1> { enum {value = 1}; };
template <> struct IsVec<char2> { enum {value = 1}; };
template <> struct IsVec<char3> { enum {value = 1}; };
template <> struct IsVec<char4> { enum {value = 1}; };
template <> struct IsVec<char8> { enum {value = 1}; };
template <> struct IsVec<ushort1> { enum {value = 1}; };
template <> struct IsVec<ushort2> { enum {value = 1}; };
template <> struct IsVec<ushort3> { enum {value = 1}; };
template <> struct IsVec<ushort4> { enum {value = 1}; };
template <> struct IsVec<ushort8> { enum {value = 1}; };
template <> struct IsVec<short1> { enum {value = 1}; };
template <> struct IsVec<short2> { enum {value = 1}; };
template <> struct IsVec<short3> { enum {value = 1}; };
template <> struct IsVec<short4> { enum {value = 1}; };
template <> struct IsVec<short8> { enum {value = 1}; };
template <> struct IsVec<uint1> { enum {value = 1}; };
template <> struct IsVec<uint2> { enum {value = 1}; };
template <> struct IsVec<uint3> { enum {value = 1}; };
template <> struct IsVec<uint4> { enum {value = 1}; };
template <> struct IsVec<uint8> { enum {value = 1}; };
template <> struct IsVec<int1> { enum {value = 1}; };
template <> struct IsVec<int2> { enum {value = 1}; };
template <> struct IsVec<int3> { enum {value = 1}; };
template <> struct IsVec<int4> { enum {value = 1}; };
template <> struct IsVec<int8> { enum {value = 1}; };
template <> struct IsVec<float1> { enum {value = 1}; };
template <> struct IsVec<float2> { enum {value = 1}; };
template <> struct IsVec<float3> { enum {value = 1}; };
template <> struct IsVec<float4> { enum {value = 1}; };
template <> struct IsVec<float8> { enum {value = 1}; };
template <> struct IsVec<double1> { enum {value = 1}; };
template <> struct IsVec<double2> { enum {value = 1}; };
template <> struct IsVec<double3> { enum {value = 1}; };
template <> struct IsVec<double4> { enum {value = 1}; };
template <> struct IsVec<double8> { enum {value = 1}; };
template <class U> struct ReferenceTraits
{
enum { value = false };
typedef U type;
};
template <class U> struct ReferenceTraits<U&>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits
{
enum { value = false };
typedef void type;
};
template <class U> struct PointerTraits<U*>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits<U*&>
{
enum { value = true };
typedef U type;
};
template <class U> struct UnConst
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnConst<const U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnConst<const U&>
{
typedef U& type;
enum { value = 1 };
};
template <class U> struct AddParameterType { typedef const U& type; };
template <class U> struct AddParameterType<U&> { typedef U& type; };
template <> struct AddParameterType<void> { typedef void type; };
template <class U> struct UnVolatile
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnVolatile<volatile U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnVolatile<volatile U&>
{
typedef U& type;
enum { value = 1 };
};
}
template <class U> struct ReferenceTraits
{
enum { value = false };
typedef U type;
};
template <class U> struct ReferenceTraits<U&>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits
{
enum { value = false };
typedef void type;
};
template <class U> struct PointerTraits<U*>
{
enum { value = true };
typedef U type;
};
template <class U> struct PointerTraits<U*&>
{
enum { value = true };
typedef U type;
};
template <class U> struct UnConst
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnConst<const U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnConst<const U&>
{
typedef U& type;
enum { value = 1 };
};
END_OPENCV_DEVICE_NAMESPACE
template <class U> struct UnVolatile
{
typedef U type;
enum { value = 0 };
};
template <class U> struct UnVolatile<volatile U>
{
typedef U type;
enum { value = 1 };
};
template <class U> struct UnVolatile<volatile U&>
{
typedef U& type;
enum { value = 1 };
};
} // namespace type_traits_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_TYPE_TRAITS_DETAIL_HPP__

1255
modules/gpu/src/opencv2/gpu/device/detail/utility_detail.hpp
View File

File diff suppressed because it is too large Load Diff

103
modules/gpu/src/opencv2/gpu/device/detail/vec_distance_detail.hpp
Unescape View File

@ -46,74 +46,73 @@
#include "internal_shared.hpp"
#include "../datamov_utils.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace detail
namespace cv { namespace gpu { namespace device
{
template <int THREAD_DIM, int N> struct UnrollVecDiffCached
namespace vec_distance_detail
{
template <typename Dist, typename T1, typename T2>
static __device__ void calcCheck(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int ind)
template <int THREAD_DIM, int N> struct UnrollVecDiffCached
{
if (ind < len)
template <typename Dist, typename T1, typename T2>
static __device__ void calcCheck(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int ind)
{
T1 val1 = *vecCached++;
if (ind < len)
{
T1 val1 = *vecCached++;
T2 val2;
ForceGlob<T2>::Load(vecGlob, ind, val2);
T2 val2;
ForceGlob<T2>::Load(vecGlob, ind, val2);
dist.reduceIter(val1, val2);
dist.reduceIter(val1, val2);
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcCheck(vecCached, vecGlob, len, dist, ind + THREAD_DIM);
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcCheck(vecCached, vecGlob, len, dist, ind + THREAD_DIM);
}
}
}
template <typename Dist, typename T1, typename T2>
static __device__ void calcWithoutCheck(const T1* vecCached, const T2* vecGlob, Dist& dist)
{
T1 val1 = *vecCached++;
template <typename Dist, typename T1, typename T2>
static __device__ void calcWithoutCheck(const T1* vecCached, const T2* vecGlob, Dist& dist)
{
T1 val1 = *vecCached++;
T2 val2;
ForceGlob<T2>::Load(vecGlob, 0, val2);
vecGlob += THREAD_DIM;
T2 val2;
ForceGlob<T2>::Load(vecGlob, 0, val2);
vecGlob += THREAD_DIM;
dist.reduceIter(val1, val2);
dist.reduceIter(val1, val2);
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcWithoutCheck(vecCached, vecGlob, dist);
}
};
template <int THREAD_DIM> struct UnrollVecDiffCached<THREAD_DIM, 0>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcCheck(const T1*, const T2*, int, Dist&, int)
UnrollVecDiffCached<THREAD_DIM, N - 1>::calcWithoutCheck(vecCached, vecGlob, dist);
}
};
template <int THREAD_DIM> struct UnrollVecDiffCached<THREAD_DIM, 0>
{
}
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcCheck(const T1*, const T2*, int, Dist&, int)
{
}
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcWithoutCheck(const T1*, const T2*, Dist&)
{
}
};
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calcWithoutCheck(const T1*, const T2*, Dist&)
{
}
};
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN> struct VecDiffCachedCalculator;
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, false>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN> struct VecDiffCachedCalculator;
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, false>
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcCheck(vecCached, vecGlob, len, dist, tid);
}
};
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, true>
{
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcCheck(vecCached, vecGlob, len, dist, tid);
}
};
template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, true>
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcWithoutCheck(vecCached, vecGlob + tid, dist);
}
};
}
END_OPENCV_DEVICE_NAMESPACE
template <typename Dist, typename T1, typename T2>
static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
{
UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcWithoutCheck(vecCached, vecGlob + tid, dist);
}
};
} // namespace vec_distance_detail
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_DISTANCE_DETAIL_HPP__

33
modules/gpu/src/opencv2/gpu/device/emulation.hpp
Unescape View File

@ -46,23 +46,22 @@
#include "internal_shared.hpp"
#include "warp_reduce.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
struct Emulation
namespace cv { namespace gpu { namespace device
{
static __forceinline__ __device__ int Ballot(int predicate, volatile int* cta_buffer)
{
#if __CUDA_ARCH__ >= 200
(void)cta_buffer;
return __ballot(predicate);
#else
int tid = threadIdx.x;
cta_buffer[tid] = predicate ? (1 << (tid & 31)) : 0;
return warp_reduce(cta_buffer);
#endif
}
};
END_OPENCV_DEVICE_NAMESPACE
struct Emulation
{
static __forceinline__ __device__ int Ballot(int predicate, volatile int* cta_buffer)
{
#if __CUDA_ARCH__ >= 200
(void)cta_buffer;
return __ballot(predicate);
#else
int tid = threadIdx.x;
cta_buffer[tid] = predicate ? (1 << (tid & 31)) : 0;
return warp_reduce(cta_buffer);
#endif
}
};
}}} // namespace cv { namespace gpu { namespace device
#endif /* OPENCV_GPU_EMULATION_HPP_ */

131
modules/gpu/src/opencv2/gpu/device/filters.hpp
Unescape View File

@ -48,90 +48,89 @@
#include "vec_traits.hpp"
#include "vec_math.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename Ptr2D> struct PointFilter
namespace cv { namespace gpu { namespace device
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ PointFilter(const Ptr2D& src_) : src(src_) {}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
template <typename Ptr2D> struct PointFilter
{
return src(__float2int_rn(y), __float2int_rn(x));
}
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
const Ptr2D src;
};
template <typename Ptr2D> struct LinearFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
explicit __host__ __device__ __forceinline__ PointFilter(const Ptr2D& src_) : src(src_) {}
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
return src(__float2int_rn(y), __float2int_rn(x));
}
explicit __host__ __device__ __forceinline__ LinearFilter(const Ptr2D& src_) : src(src_) {}
const Ptr2D src;
};
__device__ __forceinline__ elem_type operator ()(float y, float x) const
template <typename Ptr2D> struct LinearFilter
{
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
work_type out = VecTraits<work_type>::all(0);
explicit __host__ __device__ __forceinline__ LinearFilter(const Ptr2D& src_) : src(src_) {}
const int x1 = __float2int_rd(x);
const int y1 = __float2int_rd(y);
const int x2 = x1 + 1;
const int y2 = y1 + 1;
__device__ __forceinline__ elem_type operator ()(float y, float x) const
{
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
elem_type src_reg = src(y1, x1);
out = out + src_reg * ((x2 - x) * (y2 - y));
work_type out = VecTraits<work_type>::all(0);
src_reg = src(y1, x2);
out = out + src_reg * ((x - x1) * (y2 - y));
const int x1 = __float2int_rd(x);
const int y1 = __float2int_rd(y);
const int x2 = x1 + 1;
const int y2 = y1 + 1;
src_reg = src(y2, x1);
out = out + src_reg * ((x2 - x) * (y - y1));
elem_type src_reg = src(y1, x1);
out = out + src_reg * ((x2 - x) * (y2 - y));
src_reg = src(y2, x2);
out = out + src_reg * ((x - x1) * (y - y1));
src_reg = src(y1, x2);
out = out + src_reg * ((x - x1) * (y2 - y));
return saturate_cast<elem_type>(out);
}
src_reg = src(y2, x1);
out = out + src_reg * ((x2 - x) * (y - y1));
const Ptr2D src;
};
src_reg = src(y2, x2);
out = out + src_reg * ((x - x1) * (y - y1));
template <typename Ptr2D> struct CubicFilter
{
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
return saturate_cast<elem_type>(out);
}
explicit __host__ __device__ __forceinline__ CubicFilter(const Ptr2D& src_) : src(src_) {}
static __device__ __forceinline__ work_type cubicInterpolate(const work_type& p0, const work_type& p1, const work_type& p2, const work_type& p3, float x)
{
return p1 + 0.5f * x * (p2 - p0 + x * (2.0f * p0 - 5.0f * p1 + 4.0f * p2 - p3 + x * (3.0f * (p1 - p2) + p3 - p0)));
}
const Ptr2D src;
};
__device__ elem_type operator ()(float y, float x) const
template <typename Ptr2D> struct CubicFilter
{
const int xi = __float2int_rn(x);
const int yi = __float2int_rn(y);
work_type arr[4];
arr[0] = cubicInterpolate(saturate_cast<work_type>(src(yi - 1, xi - 1)), saturate_cast<work_type>(src(yi - 1, xi)), saturate_cast<work_type>(src(yi - 1, xi + 1)), saturate_cast<work_type>(src(yi - 1, xi + 2)), x - xi);
arr[1] = cubicInterpolate(saturate_cast<work_type>(src(yi , xi - 1)), saturate_cast<work_type>(src(yi , xi)), saturate_cast<work_type>(src(yi , xi + 1)), saturate_cast<work_type>(src(yi , xi + 2)), x - xi);
arr[2] = cubicInterpolate(saturate_cast<work_type>(src(yi + 1, xi - 1)), saturate_cast<work_type>(src(yi + 1, xi)), saturate_cast<work_type>(src(yi + 1, xi + 1)), saturate_cast<work_type>(src(yi + 1, xi + 2)), x - xi);
arr[3] = cubicInterpolate(saturate_cast<work_type>(src(yi + 2, xi - 1)), saturate_cast<work_type>(src(yi + 2, xi)), saturate_cast<work_type>(src(yi + 2, xi + 1)), saturate_cast<work_type>(src(yi + 2, xi + 2)), x - xi);
return saturate_cast<elem_type>(cubicInterpolate(arr[0], arr[1], arr[2], arr[3], y - yi));
}
const Ptr2D src;
};
typedef typename Ptr2D::elem_type elem_type;
typedef float index_type;
typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
END_OPENCV_DEVICE_NAMESPACE
explicit __host__ __device__ __forceinline__ CubicFilter(const Ptr2D& src_) : src(src_) {}
static __device__ __forceinline__ work_type cubicInterpolate(const work_type& p0, const work_type& p1, const work_type& p2, const work_type& p3, float x)
{
return p1 + 0.5f * x * (p2 - p0 + x * (2.0f * p0 - 5.0f * p1 + 4.0f * p2 - p3 + x * (3.0f * (p1 - p2) + p3 - p0)));
}
__device__ elem_type operator ()(float y, float x) const
{
const int xi = __float2int_rn(x);
const int yi = __float2int_rn(y);
work_type arr[4];
arr[0] = cubicInterpolate(saturate_cast<work_type>(src(yi - 1, xi - 1)), saturate_cast<work_type>(src(yi - 1, xi)), saturate_cast<work_type>(src(yi - 1, xi + 1)), saturate_cast<work_type>(src(yi - 1, xi + 2)), x - xi);
arr[1] = cubicInterpolate(saturate_cast<work_type>(src(yi , xi - 1)), saturate_cast<work_type>(src(yi , xi)), saturate_cast<work_type>(src(yi , xi + 1)), saturate_cast<work_type>(src(yi , xi + 2)), x - xi);
arr[2] = cubicInterpolate(saturate_cast<work_type>(src(yi + 1, xi - 1)), saturate_cast<work_type>(src(yi + 1, xi)), saturate_cast<work_type>(src(yi + 1, xi + 1)), saturate_cast<work_type>(src(yi + 1, xi + 2)), x - xi);
arr[3] = cubicInterpolate(saturate_cast<work_type>(src(yi + 2, xi - 1)), saturate_cast<work_type>(src(yi + 2, xi)), saturate_cast<work_type>(src(yi + 2, xi + 1)), saturate_cast<work_type>(src(yi + 2, xi + 2)), x - xi);
return saturate_cast<elem_type>(cubicInterpolate(arr[0], arr[1], arr[2], arr[3], y - yi));
}
const Ptr2D src;
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_FILTERS_HPP__

39
modules/gpu/src/opencv2/gpu/device/funcattrib.hpp
Unescape View File

@ -47,28 +47,27 @@
#include <cstdio>
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template<class Func>
void printFuncAttrib(Func& func)
namespace cv { namespace gpu { namespace device
{
template<class Func>
void printFuncAttrib(Func& func)
{
cudaFuncAttributes attrs;
cudaFuncGetAttributes(&attrs, func);
printf("=== Function stats ===\n");
printf("Name: \n");
printf("sharedSizeBytes = %d\n", attrs.sharedSizeBytes);
printf("constSizeBytes = %d\n", attrs.constSizeBytes);
printf("localSizeBytes = %d\n", attrs.localSizeBytes);
printf("maxThreadsPerBlock = %d\n", attrs.maxThreadsPerBlock);
printf("numRegs = %d\n", attrs.numRegs);
printf("ptxVersion = %d\n", attrs.ptxVersion);
printf("binaryVersion = %d\n", attrs.binaryVersion);
printf("\n");
fflush(stdout);
}
cudaFuncAttributes attrs;
cudaFuncGetAttributes(&attrs, func);
END_OPENCV_DEVICE_NAMESPACE
printf("=== Function stats ===\n");
printf("Name: \n");
printf("sharedSizeBytes = %d\n", attrs.sharedSizeBytes);
printf("constSizeBytes = %d\n", attrs.constSizeBytes);
printf("localSizeBytes = %d\n", attrs.localSizeBytes);
printf("maxThreadsPerBlock = %d\n", attrs.maxThreadsPerBlock);
printf("numRegs = %d\n", attrs.numRegs);
printf("ptxVersion = %d\n", attrs.ptxVersion);
printf("binaryVersion = %d\n", attrs.binaryVersion);
printf("\n");
fflush(stdout);
}
}}} // namespace cv { namespace gpu { namespace device
#endif /* __OPENCV_GPU_DEVICE_FUNCATTRIB_HPP_ */

742
modules/gpu/src/opencv2/gpu/device/functional.hpp
Unescape View File

@ -49,182 +49,182 @@
#include "vec_traits.hpp"
#include "type_traits.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
// Function Objects
namespace cv { namespace gpu { namespace device
{
// Function Objects
using thrust::unary_function;
using thrust::binary_function;
using thrust::unary_function;
using thrust::binary_function;
// Arithmetic Operations
// Arithmetic Operations
template <typename T> struct plus : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
template <typename T> struct plus : binary_function<T, T, T>
{
return a + b;
}
};
template <typename T> struct minus : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a + b;
}
};
template <typename T> struct minus : binary_function<T, T, T>
{
return a - b;
}
};
template <typename T> struct multiplies : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a - b;
}
};
template <typename T> struct multiplies : binary_function<T, T, T>
{
return a * b;
}
};
template <typename T> struct divides : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a * b;
}
};
template <typename T> struct divides : binary_function<T, T, T>
{
return a / b;
}
};
template <typename T> struct modulus : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a / b;
}
};
template <typename T> struct modulus : binary_function<T, T, T>
{
return a % b;
}
};
template <typename T> struct negate : unary_function<T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a % b;
}
};
template <typename T> struct negate : unary_function<T, T>
{
return -a;
}
};
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a) const
{
return -a;
}
};
// Comparison Operations
// Comparison Operations
template <typename T> struct equal_to : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
template <typename T> struct equal_to : binary_function<T, T, bool>
{
return a == b;
}
};
template <typename T> struct not_equal_to : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a == b;
}
};
template <typename T> struct not_equal_to : binary_function<T, T, bool>
{
return a != b;
}
};
template <typename T> struct greater : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a != b;
}
};
template <typename T> struct greater : binary_function<T, T, bool>
{
return a > b;
}
};
template <typename T> struct less : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a > b;
}
};
template <typename T> struct less : binary_function<T, T, bool>
{
return a < b;
}
};
template <typename T> struct greater_equal : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a < b;
}
};
template <typename T> struct greater_equal : binary_function<T, T, bool>
{
return a >= b;
}
};
template <typename T> struct less_equal : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a >= b;
}
};
template <typename T> struct less_equal : binary_function<T, T, bool>
{
return a <= b;
}
};
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a <= b;
}
};
// Logical Operations
// Logical Operations
template <typename T> struct logical_and : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
template <typename T> struct logical_and : binary_function<T, T, bool>
{
return a && b;
}
};
template <typename T> struct logical_or : binary_function<T, T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a && b;
}
};
template <typename T> struct logical_or : binary_function<T, T, bool>
{
return a || b;
}
};
template <typename T> struct logical_not : unary_function<T, bool>
{
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a) const
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a || b;
}
};
template <typename T> struct logical_not : unary_function<T, bool>
{
return !a;
}
};
__device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a) const
{
return !a;
}
};
// Bitwise Operations
// Bitwise Operations
template <typename T> struct bit_and : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
template <typename T> struct bit_and : binary_function<T, T, T>
{
return a & b;
}
};
template <typename T> struct bit_or : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a & b;
}
};
template <typename T> struct bit_or : binary_function<T, T, T>
{
return a | b;
}
};
template <typename T> struct bit_xor : binary_function<T, T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a | b;
}
};
template <typename T> struct bit_xor : binary_function<T, T, T>
{
return a ^ b;
}
};
template <typename T> struct bit_not : unary_function<T, T>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType v) const
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a, typename TypeTraits<T>::ParameterType b) const
{
return a ^ b;
}
};
template <typename T> struct bit_not : unary_function<T, T>
{
return ~v;
}
};
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType v) const
{
return ~v;
}
};
// Generalized Identity Operations
// Generalized Identity Operations
template <typename T> struct identity : unary_function<T, T>
{
__device__ __forceinline__ typename TypeTraits<T>::ParameterType operator()(typename TypeTraits<T>::ParameterType x) const
template <typename T> struct identity : unary_function<T, T>
{
return x;
}
};
__device__ __forceinline__ typename TypeTraits<T>::ParameterType operator()(typename TypeTraits<T>::ParameterType x) const
{
return x;
}
};
template <typename T1, typename T2> struct project1st : binary_function<T1, T2, T1>
{
__device__ __forceinline__ typename TypeTraits<T1>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
template <typename T1, typename T2> struct project1st : binary_function<T1, T2, T1>
{
return lhs;
}
};
template <typename T1, typename T2> struct project2nd : binary_function<T1, T2, T2>
{
__device__ __forceinline__ typename TypeTraits<T2>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
__device__ __forceinline__ typename TypeTraits<T1>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
{
return lhs;
}
};
template <typename T1, typename T2> struct project2nd : binary_function<T1, T2, T2>
{
return rhs;
}
};
__device__ __forceinline__ typename TypeTraits<T2>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
{
return rhs;
}
};
// Min/Max Operations
@ -234,39 +234,41 @@ template <typename T1, typename T2> struct project2nd : binary_function<T1, T2,
__device__ __forceinline__ type operator()(type lhs, type rhs) const {return op(lhs, rhs);} \
};
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
template <typename T> struct maximum : binary_function<T, T, T>
{
return lhs < rhs ? rhs : lhs;
}
};
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, uchar, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, schar, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, char, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, ushort, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, short, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, int, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, uint, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, float, ::fmax)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, double, ::fmax)
template <typename T> struct minimum : binary_function<T, T, T>
{
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
{
return lhs < rhs ? lhs : rhs;
}
};
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, uchar, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, schar, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, char, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, ushort, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, short, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, int, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, uint, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, float, ::fmin)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, double, ::fmin)
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
{
return lhs < rhs ? rhs : lhs;
}
};
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, uchar, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, schar, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, char, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, ushort, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, short, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, int, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, uint, ::max)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, float, ::fmax)
OPENCV_GPU_IMPLEMENT_MINMAX(maximum, double, ::fmax)
template <typename T> struct minimum : binary_function<T, T, T>
{
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
{
return lhs < rhs ? lhs : rhs;
}
};
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, uchar, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, schar, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, char, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, ushort, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, short, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, int, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, uint, ::min)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, float, ::fmin)
OPENCV_GPU_IMPLEMENT_MINMAX(minimum, double, ::fmin)
#undef OPENCV_GPU_IMPLEMENT_MINMAX
@ -287,6 +289,7 @@ OPENCV_GPU_IMPLEMENT_MINMAX(minimum, double, ::fmin)
return func(v); \
} \
};
#define OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(name, func) \
template <typename T> struct name ## _func : binary_function<T, T, float> \
{ \
@ -303,259 +306,258 @@ OPENCV_GPU_IMPLEMENT_MINMAX(minimum, double, ::fmin)
} \
};
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)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(exp10, ::exp10)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(log, ::log)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(log2, ::log2)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(log10, ::log10)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(sin, ::sin)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(cos, ::cos)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(tan, ::tan)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(asin, ::asin)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(acos, ::acos)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(atan, ::atan)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(sinh, ::sinh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(cosh, ::cosh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(tanh, ::tanh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(asinh, ::asinh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(acosh, ::acosh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(atanh, ::atanh)
OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(hypot, ::hypot)
OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(atan2, ::atan2)
OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(pow, ::pow)
#undef OPENCV_GPU_IMPLEMENT_UN_FUNCTOR
#undef OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR
template<typename T> struct hypot_sqr_func : binary_function<T, T, float>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType src1, typename TypeTraits<T>::ParameterType src2) const
{
return src1 * src1 + src2 * src2;
}
};
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)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(exp10, ::exp10)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(log, ::log)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(log2, ::log2)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(log10, ::log10)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(sin, ::sin)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(cos, ::cos)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(tan, ::tan)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(asin, ::asin)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(acos, ::acos)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(atan, ::atan)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(sinh, ::sinh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(cosh, ::cosh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(tanh, ::tanh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(asinh, ::asinh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(acosh, ::acosh)
OPENCV_GPU_IMPLEMENT_UN_FUNCTOR(atanh, ::atanh)
OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(hypot, ::hypot)
OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(atan2, ::atan2)
OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR(pow, ::pow)
#undef OPENCV_GPU_IMPLEMENT_UN_FUNCTOR
#undef OPENCV_GPU_IMPLEMENT_BIN_FUNCTOR
template<typename T> struct hypot_sqr_func : binary_function<T, T, float>
{
__device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType src1, typename TypeTraits<T>::ParameterType src2) const
{
return src1 * src1 + src2 * src2;
}
};
// Saturate Cast Functor
// Saturate Cast Functor
template <typename T, typename D> struct saturate_cast_func : unary_function<T, D>
{
__device__ __forceinline__ D operator ()(typename TypeTraits<T>::ParameterType v) const
template <typename T, typename D> struct saturate_cast_func : unary_function<T, D>
{
return saturate_cast<D>(v);
}
};
// Threshold Functors
__device__ __forceinline__ D operator ()(typename TypeTraits<T>::ParameterType v) const
{
return saturate_cast<D>(v);
}
};
template <typename T> struct thresh_binary_func : unary_function<T, T>
{
__host__ __device__ __forceinline__ thresh_binary_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
// Threshold Functors
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
template <typename T> struct thresh_binary_func : unary_function<T, T>
{
return (src > thresh) * maxVal;
}
__host__ __device__ __forceinline__ thresh_binary_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
const T thresh;
const T maxVal;
};
template <typename T> struct thresh_binary_inv_func : unary_function<T, T>
{
__host__ __device__ __forceinline__ thresh_binary_inv_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
{
return (src > thresh) * maxVal;
}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
const T thresh;
const T maxVal;
};
template <typename T> struct thresh_binary_inv_func : unary_function<T, T>
{
return (src <= thresh) * maxVal;
}
__host__ __device__ __forceinline__ thresh_binary_inv_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
const T thresh;
const T maxVal;
};
template <typename T> struct thresh_trunc_func : unary_function<T, T>
{
explicit __host__ __device__ __forceinline__ thresh_trunc_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
{
return (src <= thresh) * maxVal;
}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
const T thresh;
const T maxVal;
};
template <typename T> struct thresh_trunc_func : unary_function<T, T>
{
return minimum<T>()(src, thresh);
}
explicit __host__ __device__ __forceinline__ thresh_trunc_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {}
const T thresh;
};
template <typename T> struct thresh_to_zero_func : unary_function<T, T>
{
explicit __host__ __device__ __forceinline__ thresh_to_zero_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
{
return minimum<T>()(src, thresh);
}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
const T thresh;
};
template <typename T> struct thresh_to_zero_func : unary_function<T, T>
{
return (src > thresh) * src;
}
explicit __host__ __device__ __forceinline__ thresh_to_zero_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {}
const T thresh;
};
template <typename T> struct thresh_to_zero_inv_func : unary_function<T, T>
{
explicit __host__ __device__ __forceinline__ thresh_to_zero_inv_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
{
return (src > thresh) * src;
}
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
const T thresh;
};
template <typename T> struct thresh_to_zero_inv_func : unary_function<T, T>
{
return (src <= thresh) * src;
}
explicit __host__ __device__ __forceinline__ thresh_to_zero_inv_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {}
const T thresh;
};
__device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
{
return (src <= thresh) * src;
}
// Function Object Adaptors
template <typename Predicate> struct unary_negate : unary_function<typename Predicate::argument_type, bool>
{
explicit __host__ __device__ __forceinline__ unary_negate(const Predicate& p) : pred(p) {}
const T thresh;
};
__device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::argument_type>::ParameterType x) const
{
return !pred(x);
}
// Function Object Adaptors
const Predicate pred;
};
template <typename Predicate> __host__ __device__ __forceinline__ unary_negate<Predicate> not1(const Predicate& pred)
{
return unary_negate<Predicate>(pred);
}
template <typename Predicate> struct unary_negate : unary_function<typename Predicate::argument_type, bool>
{
explicit __host__ __device__ __forceinline__ unary_negate(const Predicate& p) : pred(p) {}
template <typename Predicate> struct binary_negate : binary_function<typename Predicate::first_argument_type, typename Predicate::second_argument_type, bool>
{
explicit __host__ __device__ __forceinline__ binary_negate(const Predicate& p) : pred(p) {}
__device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::argument_type>::ParameterType x) const
{
return !pred(x);
}
__device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::first_argument_type>::ParameterType x, typename TypeTraits<typename Predicate::second_argument_type>::ParameterType y) const
{
return !pred(x,y);
const Predicate pred;
};
template <typename Predicate> __host__ __device__ __forceinline__ unary_negate<Predicate> not1(const Predicate& pred)
{
return unary_negate<Predicate>(pred);
}
const Predicate pred;
};
template <typename BinaryPredicate> __host__ __device__ __forceinline__ binary_negate<BinaryPredicate> not2(const BinaryPredicate& pred)
{
return binary_negate<BinaryPredicate>(pred);
}
template <typename Predicate> struct binary_negate : binary_function<typename Predicate::first_argument_type, typename Predicate::second_argument_type, bool>
{
explicit __host__ __device__ __forceinline__ binary_negate(const Predicate& p) : pred(p) {}
template <typename Op> struct binder1st : unary_function<typename Op::second_argument_type, typename Op::result_type>
{
__host__ __device__ __forceinline__ binder1st(const Op& op_, const typename Op::first_argument_type& arg1_) : op(op_), arg1(arg1_) {}
__device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::first_argument_type>::ParameterType x, typename TypeTraits<typename Predicate::second_argument_type>::ParameterType y) const
{
return !pred(x,y);
}
__device__ __forceinline__ typename Op::result_type operator ()(typename TypeTraits<typename Op::second_argument_type>::ParameterType a) const
const Predicate pred;
};
template <typename BinaryPredicate> __host__ __device__ __forceinline__ binary_negate<BinaryPredicate> not2(const BinaryPredicate& pred)
{
return op(arg1, a);
return binary_negate<BinaryPredicate>(pred);
}
const Op op;
const typename Op::first_argument_type arg1;
};
template <typename Op, typename T> __host__ __device__ __forceinline__ binder1st<Op> bind1st(const Op& op, const T& x)
{
return binder1st<Op>(op, typename Op::first_argument_type(x));
}
template <typename Op> struct binder1st : unary_function<typename Op::second_argument_type, typename Op::result_type>
{
__host__ __device__ __forceinline__ binder1st(const Op& op_, const typename Op::first_argument_type& arg1_) : op(op_), arg1(arg1_) {}
template <typename Op> struct binder2nd : unary_function<typename Op::first_argument_type, typename Op::result_type>
{
__host__ __device__ __forceinline__ binder2nd(const Op& op_, const typename Op::second_argument_type& arg2_) : op(op_), arg2(arg2_) {}
__device__ __forceinline__ typename Op::result_type operator ()(typename TypeTraits<typename Op::second_argument_type>::ParameterType a) const
{
return op(arg1, a);
}
__forceinline__ __device__ typename Op::result_type operator ()(typename TypeTraits<typename Op::first_argument_type>::ParameterType a) const
const Op op;
const typename Op::first_argument_type arg1;
};
template <typename Op, typename T> __host__ __device__ __forceinline__ binder1st<Op> bind1st(const Op& op, const T& x)
{
return op(a, arg2);
return binder1st<Op>(op, typename Op::first_argument_type(x));
}
const Op op;
const typename Op::second_argument_type arg2;
};
template <typename Op, typename T> __host__ __device__ __forceinline__ binder2nd<Op> bind2nd(const Op& op, const T& x)
{
return binder2nd<Op>(op, typename Op::second_argument_type(x));
}
template <typename Op> struct binder2nd : unary_function<typename Op::first_argument_type, typename Op::result_type>
{
__host__ __device__ __forceinline__ binder2nd(const Op& op_, const typename Op::second_argument_type& arg2_) : op(op_), arg2(arg2_) {}
// Functor Traits
__forceinline__ __device__ typename Op::result_type operator ()(typename TypeTraits<typename Op::first_argument_type>::ParameterType a) const
{
return op(a, arg2);
}
template <typename F> struct IsUnaryFunction
{
typedef char Yes;
struct No {Yes a[2];};
const Op op;
const typename Op::second_argument_type arg2;
};
template <typename Op, typename T> __host__ __device__ __forceinline__ binder2nd<Op> bind2nd(const Op& op, const T& x)
{
return binder2nd<Op>(op, typename Op::second_argument_type(x));
}
template <typename T, typename D> static Yes check(unary_function<T, D>);
static No check(...);
// Functor Traits
static F makeF();
template <typename F> struct IsUnaryFunction
{
typedef char Yes;
struct No {Yes a[2];};
enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
};
template <typename T, typename D> static Yes check(unary_function<T, D>);
static No check(...);
template <typename F> struct IsBinaryFunction
{
typedef char Yes;
struct No {Yes a[2];};
static F makeF();
template <typename T1, typename T2, typename D> static Yes check(binary_function<T1, T2, D>);
static No check(...);
enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
};
static F makeF();
template <typename F> struct IsBinaryFunction
{
typedef char Yes;
struct No {Yes a[2];};
enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
};
template <typename T1, typename T2, typename D> static Yes check(binary_function<T1, T2, D>);
static No check(...);
namespace detail
{
template <size_t src_elem_size, size_t dst_elem_size> struct UnOpShift { enum { shift = 1 }; };
template <size_t src_elem_size> struct UnOpShift<src_elem_size, 1> { enum { shift = 4 }; };
template <size_t src_elem_size> struct UnOpShift<src_elem_size, 2> { enum { shift = 2 }; };
static F makeF();
template <typename T, typename D> struct DefaultUnaryShift
{
enum { shift = detail::UnOpShift<sizeof(T), sizeof(D)>::shift };
enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
};
template <size_t src_elem_size1, size_t src_elem_size2, size_t dst_elem_size> struct BinOpShift { enum { shift = 1 }; };
template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 1> { enum { shift = 4 }; };
template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 2> { enum { shift = 2 }; };
template <typename T1, typename T2, typename D> struct DefaultBinaryShift
namespace functional_detail
{
enum { shift = detail::BinOpShift<sizeof(T1), sizeof(T2), sizeof(D)>::shift };
};
template <size_t src_elem_size, size_t dst_elem_size> struct UnOpShift { enum { shift = 1 }; };
template <size_t src_elem_size> struct UnOpShift<src_elem_size, 1> { enum { shift = 4 }; };
template <size_t src_elem_size> struct UnOpShift<src_elem_size, 2> { enum { shift = 2 }; };
template <typename Func, bool unary = IsUnaryFunction<Func>::value> struct ShiftDispatcher;
template <typename Func> struct ShiftDispatcher<Func, true>
{
enum { shift = DefaultUnaryShift<typename Func::argument_type, typename Func::result_type>::shift };
};
template <typename Func> struct ShiftDispatcher<Func, false>
template <typename T, typename D> struct DefaultUnaryShift
{
enum { shift = UnOpShift<sizeof(T), sizeof(D)>::shift };
};
template <size_t src_elem_size1, size_t src_elem_size2, size_t dst_elem_size> struct BinOpShift { enum { shift = 1 }; };
template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 1> { enum { shift = 4 }; };
template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 2> { enum { shift = 2 }; };
template <typename T1, typename T2, typename D> struct DefaultBinaryShift
{
enum { shift = BinOpShift<sizeof(T1), sizeof(T2), sizeof(D)>::shift };
};
template <typename Func, bool unary = IsUnaryFunction<Func>::value> struct ShiftDispatcher;
template <typename Func> struct ShiftDispatcher<Func, true>
{
enum { shift = DefaultUnaryShift<typename Func::argument_type, typename Func::result_type>::shift };
};
template <typename Func> struct ShiftDispatcher<Func, false>
{
enum { shift = DefaultBinaryShift<typename Func::first_argument_type, typename Func::second_argument_type, typename Func::result_type>::shift };
};
}
template <typename Func> struct DefaultTransformShift
{
enum { shift = DefaultBinaryShift<typename Func::first_argument_type, typename Func::second_argument_type, typename Func::result_type>::shift };
enum { shift = functional_detail::ShiftDispatcher<Func>::shift };
};
}
template <typename Func> struct DefaultTransformShift
{
enum { shift = detail::ShiftDispatcher<Func>::shift };
};
template <typename Func> struct DefaultTransformFunctorTraits
{
enum { simple_block_dim_x = 16 };
enum { simple_block_dim_y = 16 };
template <typename Func> struct DefaultTransformFunctorTraits
{
enum { simple_block_dim_x = 16 };
enum { simple_block_dim_y = 16 };
enum { smart_block_dim_x = 16 };
enum { smart_block_dim_y = 16 };
enum { smart_shift = DefaultTransformShift<Func>::shift };
};
enum { smart_block_dim_x = 16 };
enum { smart_block_dim_y = 16 };
enum { smart_shift = DefaultTransformShift<Func>::shift };
};
template <typename Func> struct TransformFunctorTraits : DefaultTransformFunctorTraits<Func> {};
template <typename Func> struct TransformFunctorTraits : DefaultTransformFunctorTraits<Func> {};
#define DEFINE_TRANSFORM_FUNCTOR_TRAITS(type) \
#define OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(type) \
template <> struct TransformFunctorTraits< type > : DefaultTransformFunctorTraits< type >
END_OPENCV_DEVICE_NAMESPACE
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_FUNCTIONAL_HPP__

370
modules/gpu/src/opencv2/gpu/device/limits.hpp
Unescape View File

@ -43,193 +43,193 @@
#ifndef __OPENCV_GPU_LIMITS_GPU_HPP__
#define __OPENCV_GPU_LIMITS_GPU_HPP__
#include <limits>
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template<class T> struct numeric_limits
{
typedef T type;
__device__ __forceinline__ static type min() { return type(); };
__device__ __forceinline__ static type max() { return type(); };
__device__ __forceinline__ static type epsilon() { return type(); }
__device__ __forceinline__ static type round_error() { return type(); }
__device__ __forceinline__ static type denorm_min() { return type(); }
__device__ __forceinline__ static type infinity() { return type(); }
__device__ __forceinline__ static type quiet_NaN() { return type(); }
__device__ __forceinline__ static type signaling_NaN() { return T(); }
static const bool is_signed;
};
template<> struct numeric_limits<bool>
{
typedef bool type;
__device__ __forceinline__ static type min() { return false; };
__device__ __forceinline__ static type max() { return true; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<char>
{
typedef char type;
__device__ __forceinline__ static type min() { return CHAR_MIN; };
__device__ __forceinline__ static type max() { return CHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (char)-1 == -1;
};
template<> struct numeric_limits<signed char>
{
typedef char type;
__device__ __forceinline__ static type min() { return CHAR_MIN; };
__device__ __forceinline__ static type max() { return CHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (signed char)-1 == -1;
};
template<> struct numeric_limits<unsigned char>
{
typedef unsigned char type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UCHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<short>
{
typedef short type;
__device__ __forceinline__ static type min() { return SHRT_MIN; };
__device__ __forceinline__ static type max() { return SHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned short>
{
typedef unsigned short type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return USHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<int>
{
typedef int type;
__device__ __forceinline__ static type min() { return INT_MIN; };
__device__ __forceinline__ static type max() { return INT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned int>
{
typedef unsigned int type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UINT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<long>
{
typedef long type;
__device__ __forceinline__ static type min() { return LONG_MIN; };
__device__ __forceinline__ static type max() { return LONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned long>
{
typedef unsigned long type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return ULONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<float>
{
typedef float type;
__device__ __forceinline__ static type min() { return 1.175494351e-38f/*FLT_MIN*/; };
__device__ __forceinline__ static type max() { return 3.402823466e+38f/*FLT_MAX*/; };
__device__ __forceinline__ static type epsilon() { return 1.192092896e-07f/*FLT_EPSILON*/; };
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<double>
namespace cv { namespace gpu { namespace device
{
typedef double type;
__device__ __forceinline__ static type min() { return 2.2250738585072014e-308/*DBL_MIN*/; };
__device__ __forceinline__ static type max() { return 1.7976931348623158e+308/*DBL_MAX*/; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
END_OPENCV_DEVICE_NAMESPACE
template<class T> struct numeric_limits
{
typedef T type;
__device__ __forceinline__ static type min() { return type(); };
__device__ __forceinline__ static type max() { return type(); };
__device__ __forceinline__ static type epsilon() { return type(); }
__device__ __forceinline__ static type round_error() { return type(); }
__device__ __forceinline__ static type denorm_min() { return type(); }
__device__ __forceinline__ static type infinity() { return type(); }
__device__ __forceinline__ static type quiet_NaN() { return type(); }
__device__ __forceinline__ static type signaling_NaN() { return T(); }
static const bool is_signed;
};
template<> struct numeric_limits<bool>
{
typedef bool type;
__device__ __forceinline__ static type min() { return false; };
__device__ __forceinline__ static type max() { return true; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<char>
{
typedef char type;
__device__ __forceinline__ static type min() { return CHAR_MIN; };
__device__ __forceinline__ static type max() { return CHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (char)-1 == -1;
};
template<> struct numeric_limits<signed char>
{
typedef char type;
__device__ __forceinline__ static type min() { return SCHAR_MIN; };
__device__ __forceinline__ static type max() { return SCHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = (signed char)-1 == -1;
};
template<> struct numeric_limits<unsigned char>
{
typedef unsigned char type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UCHAR_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<short>
{
typedef short type;
__device__ __forceinline__ static type min() { return SHRT_MIN; };
__device__ __forceinline__ static type max() { return SHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned short>
{
typedef unsigned short type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return USHRT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<int>
{
typedef int type;
__device__ __forceinline__ static type min() { return INT_MIN; };
__device__ __forceinline__ static type max() { return INT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned int>
{
typedef unsigned int type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return UINT_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<long>
{
typedef long type;
__device__ __forceinline__ static type min() { return LONG_MIN; };
__device__ __forceinline__ static type max() { return LONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<unsigned long>
{
typedef unsigned long type;
__device__ __forceinline__ static type min() { return 0; };
__device__ __forceinline__ static type max() { return ULONG_MAX; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = false;
};
template<> struct numeric_limits<float>
{
typedef float type;
__device__ __forceinline__ static type min() { return 1.175494351e-38f/*FLT_MIN*/; };
__device__ __forceinline__ static type max() { return 3.402823466e+38f/*FLT_MAX*/; };
__device__ __forceinline__ static type epsilon() { return 1.192092896e-07f/*FLT_EPSILON*/; };
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
template<> struct numeric_limits<double>
{
typedef double type;
__device__ __forceinline__ static type min() { return 2.2250738585072014e-308/*DBL_MIN*/; };
__device__ __forceinline__ static type max() { return 1.7976931348623158e+308/*DBL_MAX*/; };
__device__ __forceinline__ static type epsilon();
__device__ __forceinline__ static type round_error();
__device__ __forceinline__ static type denorm_min();
__device__ __forceinline__ static type infinity();
__device__ __forceinline__ static type quiet_NaN();
__device__ __forceinline__ static type signaling_NaN();
static const bool is_signed = true;
};
}}} // namespace cv { namespace gpu { namespace device {
#endif // __OPENCV_GPU_LIMITS_GPU_HPP__

319
modules/gpu/src/opencv2/gpu/device/saturate_cast.hpp
Unescape View File

@ -45,173 +45,172 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uchar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(schar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(ushort v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(short v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uint v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(int v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(float v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(double v) { return _Tp(v); }
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(schar v)
{
return (uchar) ::max((int)v, 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
{
return (uchar) ::min((uint)v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
{
return (uchar)((uint)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
{
return (uchar) ::min(v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
{
return saturate_cast<uchar>((uint)v);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<uchar>(iv);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(double v)
namespace cv { namespace gpu { namespace device
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<uchar>(iv);
#else
return saturate_cast<uchar>((float)v);
#endif
}
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uchar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(schar v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(ushort v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(short v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uint v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(int v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(float v) { return _Tp(v); }
template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(double v) { return _Tp(v); }
template<> __device__ __forceinline__ schar saturate_cast<schar>(uchar v)
{
return (schar) ::min((int)v, SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
{
return (schar) ::min((uint)v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
{
return (schar)((uint)(v-SCHAR_MIN) <= (uint)UCHAR_MAX ? v : v > 0 ? SCHAR_MAX : SCHAR_MIN);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
{
return saturate_cast<schar>((int)v);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
{
return (schar) ::min(v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(schar v)
{
return (uchar) ::max((int)v, 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
{
return (uchar) ::min((uint)v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
{
return (uchar)((uint)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
{
return (uchar) ::min(v, (uint)UCHAR_MAX);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
{
return saturate_cast<uchar>((uint)v);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<schar>(iv);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(double v)
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<schar>(iv);
#else
return saturate_cast<schar>((float)v);
#endif
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<uchar>(iv);
}
template<> __device__ __forceinline__ uchar saturate_cast<uchar>(double v)
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<uchar>(iv);
#else
return saturate_cast<uchar>((float)v);
#endif
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
{
return (ushort)((uint)v <= (uint)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
{
return (ushort) ::min(v, (uint)USHRT_MAX);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<ushort>(iv);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(double v)
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<ushort>(iv);
#else
return saturate_cast<ushort>((float)v);
#endif
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uchar v)
{
return (schar) ::min((int)v, SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
{
return (schar) ::min((uint)v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
{
return (schar)((uint)(v-SCHAR_MIN) <= (uint)UCHAR_MAX ? v : v > 0 ? SCHAR_MAX : SCHAR_MIN);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
{
return saturate_cast<schar>((int)v);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
{
return (schar) ::min(v, (uint)SCHAR_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
{
return (short) ::min((int)v, SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(int v)
{
return (short)((uint)(v - SHRT_MIN) <= (uint)USHRT_MAX ? v : v > 0 ? SHRT_MAX : SHRT_MIN);
}
template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
{
return (short) ::min(v, (uint)SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<short>(iv);
}
template<> __device__ __forceinline__ short saturate_cast<short>(double v)
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<short>(iv);
#else
return saturate_cast<short>((float)v);
#endif
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<schar>(iv);
}
template<> __device__ __forceinline__ schar saturate_cast<schar>(double v)
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<schar>(iv);
#else
return saturate_cast<schar>((float)v);
#endif
}
template<> __device__ __forceinline__ int saturate_cast<int>(float v)
{
return __float2int_rn(v);
}
template<> __device__ __forceinline__ int saturate_cast<int>(double v)
{
#if __CUDA_ARCH__ >= 130
return __double2int_rn(v);
#else
return saturate_cast<int>((float)v);
#endif
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
{
return (ushort) ::max((int)v, 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
{
return (ushort)((uint)v <= (uint)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
{
return (ushort) ::min(v, (uint)USHRT_MAX);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<ushort>(iv);
}
template<> __device__ __forceinline__ ushort saturate_cast<ushort>(double v)
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<ushort>(iv);
#else
return saturate_cast<ushort>((float)v);
#endif
}
template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
{
return (short) ::min((int)v, SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(int v)
{
return (short)((uint)(v - SHRT_MIN) <= (uint)USHRT_MAX ? v : v > 0 ? SHRT_MAX : SHRT_MIN);
}
template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
{
return (short) ::min(v, (uint)SHRT_MAX);
}
template<> __device__ __forceinline__ short saturate_cast<short>(float v)
{
int iv = __float2int_rn(v);
return saturate_cast<short>(iv);
}
template<> __device__ __forceinline__ short saturate_cast<short>(double v)
{
#if __CUDA_ARCH__ >= 130
int iv = __double2int_rn(v);
return saturate_cast<short>(iv);
#else
return saturate_cast<short>((float)v);
#endif
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
{
return __float2uint_rn(v);
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(double v)
{
#if __CUDA_ARCH__ >= 130
return __double2uint_rn(v);
#else
return saturate_cast<uint>((float)v);
#endif
}
template<> __device__ __forceinline__ int saturate_cast<int>(float v)
{
return __float2int_rn(v);
}
template<> __device__ __forceinline__ int saturate_cast<int>(double v)
{
#if __CUDA_ARCH__ >= 130
return __double2int_rn(v);
#else
return saturate_cast<int>((float)v);
#endif
}
END_OPENCV_DEVICE_NAMESPACE
template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
{
return __float2uint_rn(v);
}
template<> __device__ __forceinline__ uint saturate_cast<uint>(double v)
{
#if __CUDA_ARCH__ >= 130
return __double2uint_rn(v);
#else
return saturate_cast<uint>((float)v);
#endif
}
}}}
#endif /* __OPENCV_GPU_SATURATE_CAST_HPP__ */

29
modules/gpu/src/opencv2/gpu/device/static_check.hpp
Unescape View File

@ -49,24 +49,21 @@
#define __OPENCV_GPU_HOST_DEVICE__
#endif
namespace cv
{
namespace gpu
namespace cv { namespace gpu
{
namespace device
{
namespace device
{
template<bool expr> struct Static {};
template<> struct Static<true>
{
__OPENCV_GPU_HOST_DEVICE__ static void check() {};
};
}
template<bool expr> struct Static {};
template<> struct Static<true>
{
__OPENCV_GPU_HOST_DEVICE__ static void check() {};
};
}
using cv::gpu::device::Static;
}
}
using ::cv::gpu::device::Static;
}}
#undef __PCL_GPU_HOST_DEVICE__
#undef __OPENCV_GPU_HOST_DEVICE__
#endif /* __OPENCV_GPU_GPU_DEVICE_STATIC_CHECK_HPP__ */

45
modules/gpu/src/opencv2/gpu/device/transform.hpp
Unescape View File

@ -47,30 +47,31 @@
#include "utility.hpp"
#include "detail/transform_detail.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T, typename D, typename UnOp>
void transform(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, cudaStream_t stream = 0)
{
detail::transform_caller(src, dst, op, WithOutMask(), stream);
}
template <typename T, typename D, typename UnOp>
void transform(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const PtrStepb& mask, const UnOp& op, cudaStream_t stream = 0)
namespace cv { namespace gpu { namespace device
{
detail::transform_caller(src, dst, op, SingleMask(mask), stream);
}
template <typename T, typename D, typename UnOp>
void transform(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const UnOp& op, cudaStream_t stream = 0)
{
transform_detail::transform_caller(src, dst, op, WithOutMask(), stream);
}
template <typename T1, typename T2, typename D, typename BinOp>
void transform(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, cudaStream_t stream = 0)
{
detail::transform_caller(src1, src2, dst, op, WithOutMask(), stream);
}
template <typename T1, typename T2, typename D, typename BinOp>
void transform(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const PtrStepb& mask, const BinOp& op, cudaStream_t stream = 0)
{
detail::transform_caller(src1, src2, dst, op, SingleMask(mask), stream);
}
template <typename T, typename D, typename UnOp>
void transform(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, const PtrStepb& mask, const UnOp& op, cudaStream_t stream = 0)
{
transform_detail::transform_caller(src, dst, op, SingleMask(mask), stream);
}
template <typename T1, typename T2, typename D, typename BinOp>
void transform(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const BinOp& op, cudaStream_t stream = 0)
{
transform_detail::transform_caller(src1, src2, dst, op, WithOutMask(), stream);
}
END_OPENCV_DEVICE_NAMESPACE
template <typename T1, typename T2, typename D, typename BinOp>
void transform(const DevMem2D_<T1>& src1, const DevMem2D_<T2>& src2, const DevMem2D_<D>& dst, const PtrStepb& mask, const BinOp& op, cudaStream_t stream = 0)
{
transform_detail::transform_caller(src1, src2, dst, op, SingleMask(mask), stream);
}
}}}
#endif // __OPENCV_GPU_TRANSFORM_HPP__

55
modules/gpu/src/opencv2/gpu/device/type_traits.hpp
Unescape View File

@ -46,37 +46,38 @@
#include "internal_shared.hpp"
#include "detail/type_traits_detail.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T> struct IsSimpleParameter
{
enum {value = detail::IsIntegral<T>::value || detail::IsFloat<T>::value || detail::PointerTraits<typename detail::ReferenceTraits<T>::type>::value};
};
template <typename T> struct TypeTraits
namespace cv { namespace gpu { namespace device
{
typedef typename detail::UnConst<T>::type NonConstType;
typedef typename detail::UnVolatile<T>::type NonVolatileType;
typedef typename detail::UnVolatile<typename detail::UnConst<T>::type>::type UnqualifiedType;
typedef typename detail::PointerTraits<UnqualifiedType>::type PointeeType;
typedef typename detail::ReferenceTraits<T>::type ReferredType;
template <typename T> struct IsSimpleParameter
{
enum {value = type_traits_detail::IsIntegral<T>::value || type_traits_detail::IsFloat<T>::value ||
type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<T>::type>::value};
};
enum { isConst = detail::UnConst<T>::value };
enum { isVolatile = detail::UnVolatile<T>::value };
template <typename T> struct TypeTraits
{
typedef typename type_traits_detail::UnConst<T>::type NonConstType;
typedef typename type_traits_detail::UnVolatile<T>::type NonVolatileType;
typedef typename type_traits_detail::UnVolatile<typename type_traits_detail::UnConst<T>::type>::type UnqualifiedType;
typedef typename type_traits_detail::PointerTraits<UnqualifiedType>::type PointeeType;
typedef typename type_traits_detail::ReferenceTraits<T>::type ReferredType;
enum { isReference = detail::ReferenceTraits<UnqualifiedType>::value };
enum { isPointer = detail::PointerTraits<typename detail::ReferenceTraits<UnqualifiedType>::type>::value };
enum { isConst = type_traits_detail::UnConst<T>::value };
enum { isVolatile = type_traits_detail::UnVolatile<T>::value };
enum { isUnsignedInt = detail::IsUnsignedIntegral<UnqualifiedType>::value };
enum { isSignedInt = detail::IsSignedIntergral<UnqualifiedType>::value };
enum { isIntegral = detail::IsIntegral<UnqualifiedType>::value };
enum { isFloat = detail::IsFloat<UnqualifiedType>::value };
enum { isArith = isIntegral || isFloat };
enum { isVec = detail::IsVec<UnqualifiedType>::value };
typedef typename detail::Select<IsSimpleParameter<UnqualifiedType>::value, T, typename detail::AddParameterType<T>::type>::type ParameterType;
};
enum { isReference = type_traits_detail::ReferenceTraits<UnqualifiedType>::value };
enum { isPointer = type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<UnqualifiedType>::type>::value };
END_OPENCV_DEVICE_NAMESPACE
enum { isUnsignedInt = type_traits_detail::IsUnsignedIntegral<UnqualifiedType>::value };
enum { isSignedInt = type_traits_detail::IsSignedIntergral<UnqualifiedType>::value };
enum { isIntegral = type_traits_detail::IsIntegral<UnqualifiedType>::value };
enum { isFloat = type_traits_detail::IsFloat<UnqualifiedType>::value };
enum { isArith = isIntegral || isFloat };
enum { isVec = type_traits_detail::IsVec<UnqualifiedType>::value };
typedef typename type_traits_detail::Select<IsSimpleParameter<UnqualifiedType>::value,
T, typename type_traits_detail::AddParameterType<T>::type>::type ParameterType;
};
}}}
#endif // __OPENCV_GPU_TYPE_TRAITS_HPP__

255
modules/gpu/src/opencv2/gpu/device/utility.hpp
Unescape View File

@ -48,168 +48,167 @@
#include "datamov_utils.hpp"
#include "detail/utility_detail.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#define OPENCV_GPU_LOG_WARP_SIZE (5)
#define OPENCV_GPU_WARP_SIZE (1 << OPENCV_GPU_LOG_WARP_SIZE)
#define OPENCV_GPU_LOG_MEM_BANKS ((__CUDA_ARCH__ >= 200) ? 5 : 4) // 32 banks on fermi, 16 on tesla
#define OPENCV_GPU_MEM_BANKS (1 << OPENCV_GPU_LOG_MEM_BANKS)
///////////////////////////////////////////////////////////////////////////////
// swap
template <typename T> void __device__ __host__ __forceinline__ swap(T& a, T& b)
namespace cv { namespace gpu { namespace device
{
const T temp = a;
a = b;
b = temp;
}
#define OPENCV_GPU_LOG_WARP_SIZE (5)
#define OPENCV_GPU_WARP_SIZE (1 << OPENCV_GPU_LOG_WARP_SIZE)
#define OPENCV_GPU_LOG_MEM_BANKS ((__CUDA_ARCH__ >= 200) ? 5 : 4) // 32 banks on fermi, 16 on tesla
#define OPENCV_GPU_MEM_BANKS (1 << OPENCV_GPU_LOG_MEM_BANKS)
///////////////////////////////////////////////////////////////////////////////
// Mask Reader
///////////////////////////////////////////////////////////////////////////////
// swap
struct SingleMask
{
explicit __host__ __device__ __forceinline__ SingleMask(const PtrStepb& mask_) : mask(mask_) {}
__device__ __forceinline__ bool operator()(int y, int x) const
{
return mask.ptr(y)[x] != 0;
template <typename T> void __device__ __host__ __forceinline__ swap(T& a, T& b)
{
const T temp = a;
a = b;
b = temp;
}
const PtrStepb mask;
};
struct MaskCollection
{
explicit __host__ __device__ __forceinline__ MaskCollection(PtrStepb* maskCollection_) : maskCollection(maskCollection_) {}
///////////////////////////////////////////////////////////////////////////////
// Mask Reader
__device__ __forceinline__ void next()
{
curMask = *maskCollection++;
}
__device__ __forceinline__ void setMask(int z)
struct SingleMask
{
curMask = maskCollection[z];
}
__device__ __forceinline__ bool operator()(int y, int x) const
{
uchar val;
return curMask.data == 0 || (ForceGlob<uchar>::Load(curMask.ptr(y), x, val), (val != 0));
}
explicit __host__ __device__ __forceinline__ SingleMask(const PtrStepb& mask_) : mask(mask_) {}
__device__ __forceinline__ bool operator()(int y, int x) const
{
return mask.ptr(y)[x] != 0;
}
const PtrStepb* maskCollection;
PtrStepb curMask;
};
const PtrStepb mask;
};
struct WithOutMask
{
__device__ __forceinline__ void next() const
struct MaskCollection
{
}
__device__ __forceinline__ void setMask(int) const
explicit __host__ __device__ __forceinline__ MaskCollection(PtrStepb* maskCollection_) : maskCollection(maskCollection_) {}
__device__ __forceinline__ void next()
{
curMask = *maskCollection++;
}
__device__ __forceinline__ void setMask(int z)
{
curMask = maskCollection[z];
}
__device__ __forceinline__ bool operator()(int y, int x) const
{
uchar val;
return curMask.data == 0 || (ForceGlob<uchar>::Load(curMask.ptr(y), x, val), (val != 0));
}
const PtrStepb* maskCollection;
PtrStepb curMask;
};
struct WithOutMask
{
}
__device__ __forceinline__ bool operator()(int, int) const
__device__ __forceinline__ void next() const
{
}
__device__ __forceinline__ void setMask(int) const
{
}
__device__ __forceinline__ bool operator()(int, int) const
{
return true;
}
__device__ __forceinline__ bool operator()(int, int, int) const
{
return true;
}
static __device__ __forceinline__ bool check(int, int)
{
return true;
}
static __device__ __forceinline__ bool check(int, int, int)
{
return true;
}
};
///////////////////////////////////////////////////////////////////////////////
// Reduction
template <int n, typename T, typename Op> __device__ __forceinline__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
return true;
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::ReductionDispatcher<n <= 64>::reduce<n>(data, partial_reduction, tid, op);
}
__device__ __forceinline__ bool operator()(int, int, int) const
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)
{
return true;
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::PredValReductionDispatcher<n <= 64>::reduce<n>(myData, myVal, sdata, sval, tid, pred);
}
static __device__ __forceinline__ bool check(int, int)
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)
{
return true;
StaticAssert<n >= 8 && n <= 512>::check();
utility_detail::PredVal2ReductionDispatcher<n <= 64>::reduce<n>(myData, myVal1, myVal2, sdata, sval1, sval2, tid, pred);
}
static __device__ __forceinline__ bool check(int, int, int)
{
return true;
}
};
///////////////////////////////////////////////////////////////////////////////
// Reduction
template <int n, typename T, typename Op> __device__ __forceinline__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
{
StaticAssert<n >= 8 && n <= 512>::check();
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();
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();
detail::PredVal2ReductionDispatcher<n <= 64>::reduce<n>(myData, myVal1, myVal2, sdata, sval1, sval2, tid, pred);
}
///////////////////////////////////////////////////////////////////////////////
// Solve linear system
// solve 2x2 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve2x2(const T A[2][2], const T b[2], T x[2])
{
T det = A[0][0] * A[1][1] - A[1][0] * A[0][1];
///////////////////////////////////////////////////////////////////////////////
// Solve linear system
if (det != 0)
// solve 2x2 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve2x2(const T A[2][2], const T b[2], T x[2])
{
double invdet = 1.0 / det;
T det = A[0][0] * A[1][1] - A[1][0] * A[0][1];
x[0] = saturate_cast<T>(invdet * (b[0] * A[1][1] - b[1] * A[0][1]));
if (det != 0)
{
double invdet = 1.0 / det;
x[1] = saturate_cast<T>(invdet * (A[0][0] * b[1] - A[1][0] * b[0]));
x[0] = saturate_cast<T>(invdet * (b[0] * A[1][1] - b[1] * A[0][1]));
return true;
}
x[1] = saturate_cast<T>(invdet * (A[0][0] * b[1] - A[1][0] * b[0]));
return false;
}
return true;
}
// solve 3x3 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve3x3(const T A[3][3], const T b[3], T x[3])
{
T det = A[0][0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1])
- A[0][1] * (A[1][0] * A[2][2] - A[1][2] * A[2][0])
+ A[0][2] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]);
return false;
}
if (det != 0)
// solve 3x3 linear system Ax=b
template <typename T> __device__ __forceinline__ bool solve3x3(const T A[3][3], const T b[3], T x[3])
{
double invdet = 1.0 / det;
T det = A[0][0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1])
- A[0][1] * (A[1][0] * A[2][2] - A[1][2] * A[2][0])
+ A[0][2] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]);
x[0] = saturate_cast<T>(invdet *
(b[0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1]) -
A[0][1] * (b[1] * A[2][2] - A[1][2] * b[2] ) +
A[0][2] * (b[1] * A[2][1] - A[1][1] * b[2] )));
if (det != 0)
{
double invdet = 1.0 / det;
x[1] = saturate_cast<T>(invdet *
(A[0][0] * (b[1] * A[2][2] - A[1][2] * b[2] ) -
b[0] * (A[1][0] * A[2][2] - A[1][2] * A[2][0]) +
A[0][2] * (A[1][0] * b[2] - b[1] * A[2][0])));
x[0] = saturate_cast<T>(invdet *
(b[0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1]) -
A[0][1] * (b[1] * A[2][2] - A[1][2] * b[2] ) +
A[0][2] * (b[1] * A[2][1] - A[1][1] * b[2] )));
x[2] = saturate_cast<T>(invdet *
(A[0][0] * (A[1][1] * b[2] - b[1] * A[2][1]) -
A[0][1] * (A[1][0] * b[2] - b[1] * A[2][0]) +
b[0] * (A[1][0] * A[2][1] - A[1][1] * A[2][0])));
x[1] = saturate_cast<T>(invdet *
(A[0][0] * (b[1] * A[2][2] - A[1][2] * b[2] ) -
b[0] * (A[1][0] * A[2][2] - A[1][2] * A[2][0]) +
A[0][2] * (A[1][0] * b[2] - b[1] * A[2][0])));
return true;
}
x[2] = saturate_cast<T>(invdet *
(A[0][0] * (A[1][1] * b[2] - b[1] * A[2][1]) -
A[0][1] * (A[1][0] * b[2] - b[1] * A[2][0]) +
b[0] * (A[1][0] * A[2][1] - A[1][1] * A[2][0])));
return false;
}
return true;
}
END_OPENCV_DEVICE_NAMESPACE
return false;
}
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_UTILITY_HPP__

263
modules/gpu/src/opencv2/gpu/device/vec_distance.hpp
Unescape View File

@ -48,179 +48,178 @@
#include "functional.hpp"
#include "detail/vec_distance_detail.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T> struct L1Dist
namespace cv { namespace gpu { namespace device
{
typedef int value_type;
typedef int result_type;
__device__ __forceinline__ L1Dist() : mySum(0) {}
__device__ __forceinline__ void reduceIter(int val1, int val2)
template <typename T> struct L1Dist
{
mySum = __sad(val1, val2, mySum);
}
typedef int value_type;
typedef int result_type;
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
__device__ __forceinline__ L1Dist() : mySum(0) {}
__device__ __forceinline__ operator int() const
{
return mySum;
}
__device__ __forceinline__ void reduceIter(int val1, int val2)
{
mySum = __sad(val1, val2, mySum);
}
int mySum;
};
template <> struct L1Dist<float>
{
typedef float value_type;
typedef float result_type;
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
__device__ __forceinline__ L1Dist() : mySum(0.0f) {}
__device__ __forceinline__ operator int() const
{
return mySum;
}
__device__ __forceinline__ void reduceIter(float val1, float val2)
int mySum;
};
template <> struct L1Dist<float>
{
mySum += ::fabs(val1 - val2);
}
typedef float value_type;
typedef float result_type;
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
__device__ __forceinline__ L1Dist() : mySum(0.0f) {}
__device__ __forceinline__ operator float() const
{
return mySum;
}
__device__ __forceinline__ void reduceIter(float val1, float val2)
{
mySum += ::fabs(val1 - val2);
}
float mySum;
};
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
struct L2Dist
{
typedef float value_type;
typedef float result_type;
__device__ __forceinline__ operator float() const
{
return mySum;
}
__device__ __forceinline__ L2Dist() : mySum(0.0f) {}
float mySum;
};
__device__ __forceinline__ void reduceIter(float val1, float val2)
struct L2Dist
{
float reg = val1 - val2;
mySum += reg * reg;
}
typedef float value_type;
typedef float result_type;
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
__device__ __forceinline__ L2Dist() : mySum(0.0f) {}
__device__ __forceinline__ operator float() const
{
return sqrtf(mySum);
}
__device__ __forceinline__ void reduceIter(float val1, float val2)
{
float reg = val1 - val2;
mySum += reg * reg;
}
float mySum;
};
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
}
struct HammingDist
{
typedef int value_type;
typedef int result_type;
__device__ __forceinline__ operator float() const
{
return sqrtf(mySum);
}
__device__ __forceinline__ HammingDist() : mySum(0) {}
float mySum;
};
__device__ __forceinline__ void reduceIter(int val1, int val2)
struct HammingDist
{
mySum += __popc(val1 ^ val2);
}
typedef int value_type;
typedef int result_type;
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
__device__ __forceinline__ HammingDist() : mySum(0) {}
__device__ __forceinline__ operator int() const
{
return mySum;
}
__device__ __forceinline__ void reduceIter(int val1, int val2)
{
mySum += __popc(val1 ^ val2);
}
int mySum;
};
template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
{
reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
}
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename Dist, typename T1, typename T2>
__device__ void calcVecDiffGlobal(const T1* vec1, const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
for (int i = tid; i < len; i += THREAD_DIM)
{
T1 val1;
ForceGlob<T1>::Load(vec1, i, val1);
__device__ __forceinline__ operator int() const
{
return mySum;
}
T2 val2;
ForceGlob<T2>::Load(vec2, i, val2);
int mySum;
};
dist.reduceIter(val1, val2);
}
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename Dist, typename T1, typename T2>
__device__ void calcVecDiffGlobal(const T1* vec1, const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
for (int i = tid; i < len; i += THREAD_DIM)
{
T1 val1;
ForceGlob<T1>::Load(vec1, i, val1);
dist.reduceAll<THREAD_DIM>(smem, tid);
}
T2 val2;
ForceGlob<T2>::Load(vec2, i, val2);
// calc distance between two vectors, first vector is cached in register or shared memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename Dist, typename T1, typename T2>
__device__ __forceinline__ void calcVecDiffCached(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
detail::VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>::calc(vecCached, vecGlob, len, dist, tid);
dist.reduceAll<THREAD_DIM>(smem, tid);
}
dist.reduceIter(val1, val2);
}
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename T1> struct VecDiffGlobal
{
explicit __device__ __forceinline__ VecDiffGlobal(const T1* vec1_, int = 0, void* = 0, int = 0, int = 0)
{
vec1 = vec1_;
dist.reduceAll<THREAD_DIM>(smem, tid);
}
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
{
calcVecDiffGlobal<THREAD_DIM>(vec1, vec2, len, dist, smem, tid);
// calc distance between two vectors, first vector is cached in register or shared memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename Dist, typename T1, typename T2>
__device__ __forceinline__ void calcVecDiffCached(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, typename Dist::result_type* smem, int tid)
{
vec_distance_detail::VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>::calc(vecCached, vecGlob, len, dist, tid);
dist.reduceAll<THREAD_DIM>(smem, tid);
}
const T1* vec1;
};
// calc distance between two vectors, first vector is cached in register memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename U> struct VecDiffCachedRegister
{
template <typename T1> __device__ __forceinline__ VecDiffCachedRegister(const T1* vec1, int len, U* smem, int glob_tid, int tid)
// calc distance between two vectors in global memory
template <int THREAD_DIM, typename T1> struct VecDiffGlobal
{
if (glob_tid < len)
smem[glob_tid] = vec1[glob_tid];
__syncthreads();
U* vec1ValsPtr = vec1Vals;
explicit __device__ __forceinline__ VecDiffGlobal(const T1* vec1_, int = 0, void* = 0, int = 0, int = 0)
{
vec1 = vec1_;
}
#pragma unroll
for (int i = tid; i < MAX_LEN; i += THREAD_DIM)
*vec1ValsPtr++ = smem[i];
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
{
calcVecDiffGlobal<THREAD_DIM>(vec1, vec2, len, dist, smem, tid);
}
__syncthreads();
}
const T1* vec1;
};
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
// calc distance between two vectors, first vector is cached in register memory, second vector is in global memory
template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename U> struct VecDiffCachedRegister
{
calcVecDiffCached<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>(vec1Vals, vec2, len, dist, smem, tid);
}
template <typename T1> __device__ __forceinline__ VecDiffCachedRegister(const T1* vec1, int len, U* smem, int glob_tid, int tid)
{
if (glob_tid < len)
smem[glob_tid] = vec1[glob_tid];
__syncthreads();
U* vec1ValsPtr = vec1Vals;
#pragma unroll
for (int i = tid; i < MAX_LEN; i += THREAD_DIM)
*vec1ValsPtr++ = smem[i];
__syncthreads();
}
U vec1Vals[MAX_LEN / THREAD_DIM];
};
template <typename T2, typename Dist>
__device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
{
calcVecDiffCached<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>(vec1Vals, vec2, len, dist, smem, tid);
}
END_OPENCV_DEVICE_NAMESPACE
U vec1Vals[MAX_LEN / THREAD_DIM];
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_DISTANCE_HPP__

295
modules/gpu/src/opencv2/gpu/device/vec_math.hpp
Unescape View File

@ -48,85 +48,85 @@
#include "vec_traits.hpp"
#include "functional.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace detail
namespace cv { namespace gpu { namespace device
{
template <int cn, typename VecD> struct SatCastHelper;
template <typename VecD> struct SatCastHelper<1, VecD>
namespace vec_math_detail
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
template <int cn, typename VecD> struct SatCastHelper;
template <typename VecD> struct SatCastHelper<1, VecD>
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x));
}
};
template <typename VecD> struct SatCastHelper<2, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x));
}
};
template <typename VecD> struct SatCastHelper<2, VecD>
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y));
}
};
template <typename VecD> struct SatCastHelper<3, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y));
}
};
template <typename VecD> struct SatCastHelper<3, VecD>
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z));
}
};
template <typename VecD> struct SatCastHelper<4, VecD>
{
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z));
}
};
template <typename VecD> struct SatCastHelper<4, VecD>
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z), saturate_cast<D>(v.w));
}
};
template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
{
typedef typename VecTraits<VecD>::elem_type D;
return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z), saturate_cast<D>(v.w));
}
};
template <typename VecD, typename VecS> static __device__ __forceinline__ VecD saturate_cast_caller(const VecS& v)
{
return SatCastHelper<VecTraits<VecD>::cn, VecD>::cast(v);
template <typename VecD, typename VecS> static __device__ __forceinline__ VecD saturate_cast_caller(const VecS& v)
{
return SatCastHelper<VecTraits<VecD>::cn, VecD>::cast(v);
}
}
}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double1& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double1& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double2& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double2& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double3& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double3& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double4& v) {return detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uchar4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const char4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const ushort4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const short4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const uint4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const int4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const float4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const double4& v) {return vec_math_detail::saturate_cast_caller<_Tp>(v);}
#define OPENCV_GPU_IMPLEMENT_VEC_UNOP(type, op, func) \
__device__ __forceinline__ TypeVec<func<type>::result_type, 1>::vec_type op(const type ## 1 & a) \
@ -150,49 +150,49 @@ template<typename _Tp> static __device__ __forceinline__ _Tp saturate_cast(const
return VecTraits<TypeVec<func<type>::result_type, 4>::vec_type>::make(f(a.x), f(a.y), f(a.z), f(a.w)); \
}
namespace detail
{
template <typename T1, typename T2> struct BinOpTraits
{
typedef int argument_type;
};
template <typename T> struct BinOpTraits<T, T>
{
typedef T argument_type;
};
template <typename T> struct BinOpTraits<T, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<double, T>
{
typedef double argument_type;
};
template <> struct BinOpTraits<double, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<T, float>
{
typedef float argument_type;
};
template <typename T> struct BinOpTraits<float, T>
{
typedef float argument_type;
};
template <> struct BinOpTraits<float, float>
{
typedef float argument_type;
};
template <> struct BinOpTraits<double, float>
{
typedef double argument_type;
};
template <> struct BinOpTraits<float, double>
{
typedef double argument_type;
};
}
namespace vec_math_detail
{
template <typename T1, typename T2> struct BinOpTraits
{
typedef int argument_type;
};
template <typename T> struct BinOpTraits<T, T>
{
typedef T argument_type;
};
template <typename T> struct BinOpTraits<T, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<double, T>
{
typedef double argument_type;
};
template <> struct BinOpTraits<double, double>
{
typedef double argument_type;
};
template <typename T> struct BinOpTraits<T, float>
{
typedef float argument_type;
};
template <typename T> struct BinOpTraits<float, T>
{
typedef float argument_type;
};
template <> struct BinOpTraits<float, float>
{
typedef float argument_type;
};
template <> struct BinOpTraits<double, float>
{
typedef double argument_type;
};
template <> struct BinOpTraits<float, double>
{
typedef double argument_type;
};
}
#define OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, op, func) \
__device__ __forceinline__ TypeVec<func<type>::result_type, 1>::vec_type op(const type ## 1 & a, const type ## 1 & b) \
@ -201,16 +201,16 @@ namespace detail
return VecTraits<TypeVec<func<type>::result_type, 1>::vec_type>::make(f(a.x, b.x)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(const type ## 1 & v, T s) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(const type ## 1 & v, T s) \
{ \
func<typename detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(v.x, s)); \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(v.x, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(T s, const type ## 1 & v) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type op(T s, const type ## 1 & v) \
{ \
func<typename detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(s, v.x)); \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 1>::vec_type>::make(f(s, v.x)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 2>::vec_type op(const type ## 2 & a, const type ## 2 & b) \
{ \
@ -218,16 +218,16 @@ namespace detail
return VecTraits<TypeVec<func<type>::result_type, 2>::vec_type>::make(f(a.x, b.x), f(a.y, b.y)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(const type ## 2 & v, T s) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(const type ## 2 & v, T s) \
{ \
func<typename detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(v.x, s), f(v.y, s)); \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(v.x, s), f(v.y, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(T s, const type ## 2 & v) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type op(T s, const type ## 2 & v) \
{ \
func<typename detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(s, v.x), f(s, v.y)); \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 2>::vec_type>::make(f(s, v.x), f(s, v.y)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 3>::vec_type op(const type ## 3 & a, const type ## 3 & b) \
{ \
@ -235,16 +235,16 @@ namespace detail
return VecTraits<TypeVec<func<type>::result_type, 3>::vec_type>::make(f(a.x, b.x), f(a.y, b.y), f(a.z, b.z)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(const type ## 3 & v, T s) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(const type ## 3 & v, T s) \
{ \
func<typename detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s)); \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(T s, const type ## 3 & v) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type op(T s, const type ## 3 & v) \
{ \
func<typename detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z)); \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 3>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z)); \
} \
__device__ __forceinline__ TypeVec<func<type>::result_type, 4>::vec_type op(const type ## 4 & a, const type ## 4 & b) \
{ \
@ -252,16 +252,16 @@ namespace detail
return VecTraits<TypeVec<func<type>::result_type, 4>::vec_type>::make(f(a.x, b.x), f(a.y, b.y), f(a.z, b.z), f(a.w, b.w)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(const type ## 4 & v, T s) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(const type ## 4 & v, T s) \
{ \
func<typename detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s), f(v.w, s)); \
func<typename vec_math_detail::BinOpTraits<type, T>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(v.x, s), f(v.y, s), f(v.z, s), f(v.w, s)); \
} \
template <typename T> \
__device__ __forceinline__ typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(T s, const type ## 4 & v) \
__device__ __forceinline__ typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type op(T s, const type ## 4 & v) \
{ \
func<typename detail::BinOpTraits<T, type>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z), f(s, v.w)); \
func<typename vec_math_detail::BinOpTraits<T, type>::argument_type> f; \
return VecTraits<typename TypeVec<typename func<typename vec_math_detail::BinOpTraits<type, T>::argument_type>::result_type, 4>::vec_type>::make(f(s, v.x), f(s, v.y), f(s, v.z), f(s, v.w)); \
}
#define OPENCV_GPU_IMPLEMENT_VEC_OP(type) \
@ -313,20 +313,19 @@ namespace detail
OPENCV_GPU_IMPLEMENT_VEC_BINOP(type, operator ^, bit_xor) \
OPENCV_GPU_IMPLEMENT_VEC_UNOP (type, operator ~, bit_not)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uchar)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(char)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(ushort)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(short)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(int)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uint)
OPENCV_GPU_IMPLEMENT_VEC_OP(float)
OPENCV_GPU_IMPLEMENT_VEC_OP(double)
#undef OPENCV_GPU_IMPLEMENT_VEC_UNOP
#undef OPENCV_GPU_IMPLEMENT_VEC_BINOP
#undef OPENCV_GPU_IMPLEMENT_VEC_OP
#undef OPENCV_GPU_IMPLEMENT_VEC_INT_OP
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uchar)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(char)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(ushort)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(short)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(int)
OPENCV_GPU_IMPLEMENT_VEC_INT_OP(uint)
OPENCV_GPU_IMPLEMENT_VEC_OP(float)
OPENCV_GPU_IMPLEMENT_VEC_OP(double)
END_OPENCV_DEVICE_NAMESPACE
#undef OPENCV_GPU_IMPLEMENT_VEC_UNOP
#undef OPENCV_GPU_IMPLEMENT_VEC_BINOP
#undef OPENCV_GPU_IMPLEMENT_VEC_OP
#undef OPENCV_GPU_IMPLEMENT_VEC_INT_OP
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VECMATH_HPP__

319
modules/gpu/src/opencv2/gpu/device/vec_traits.hpp
Unescape View File

@ -45,82 +45,82 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template<typename T, int N> struct TypeVec;
struct __align__(8) uchar8
{
uchar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uchar8 make_uchar8(uchar a0, uchar a1, uchar a2, uchar a3, uchar a4, uchar a5, uchar a6, uchar a7)
{
uchar8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(8) char8
{
schar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ char8 make_char8(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7)
{
char8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) ushort8
{
ushort a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ ushort8 make_ushort8(ushort a0, ushort a1, ushort a2, ushort a3, ushort a4, ushort a5, ushort a6, ushort a7)
{
ushort8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) short8
{
short a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ short8 make_short8(short a0, short a1, short a2, short a3, short a4, short a5, short a6, short a7)
{
short8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) uint8
namespace cv { namespace gpu { namespace device
{
uint a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uint8 make_uint8(uint a0, uint a1, uint a2, uint a3, uint a4, uint a5, uint a6, uint a7)
{
uint8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) int8
{
int a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ int8 make_int8(int a0, int a1, int a2, int a3, int a4, int a5, int a6, int a7)
{
int8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) float8
{
float a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ float8 make_float8(float a0, float a1, float a2, float a3, float a4, float a5, float a6, float a7)
{
float8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct double8
{
double a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ double8 make_double8(double a0, double a1, double a2, double a3, double a4, double a5, double a6, double a7)
{
double8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
template<typename T, int N> struct TypeVec;
struct __align__(8) uchar8
{
uchar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uchar8 make_uchar8(uchar a0, uchar a1, uchar a2, uchar a3, uchar a4, uchar a5, uchar a6, uchar a7)
{
uchar8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(8) char8
{
schar a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ char8 make_char8(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7)
{
char8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) ushort8
{
ushort a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ ushort8 make_ushort8(ushort a0, ushort a1, ushort a2, ushort a3, ushort a4, ushort a5, ushort a6, ushort a7)
{
ushort8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(16) short8
{
short a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ short8 make_short8(short a0, short a1, short a2, short a3, short a4, short a5, short a6, short a7)
{
short8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) uint8
{
uint a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ uint8 make_uint8(uint a0, uint a1, uint a2, uint a3, uint a4, uint a5, uint a6, uint a7)
{
uint8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) int8
{
int a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ int8 make_int8(int a0, int a1, int a2, int a3, int a4, int a5, int a6, int a7)
{
int8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct __align__(32) float8
{
float a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ float8 make_float8(float a0, float a1, float a2, float a3, float a4, float a5, float a6, float a7)
{
float8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
struct double8
{
double a0, a1, a2, a3, a4, a5, a6, a7;
};
static __host__ __device__ __forceinline__ double8 make_double8(double a0, double a1, double a2, double a3, double a4, double a5, double a6, double a7)
{
double8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
return val;
}
#define OPENCV_GPU_IMPLEMENT_TYPE_VEC(type) \
template<> struct TypeVec<type, 1> { typedef type vec_type; }; \
@ -134,28 +134,28 @@ static __host__ __device__ __forceinline__ double8 make_double8(double a0, doubl
template<> struct TypeVec<type, 8> { typedef type ## 8 vec_type; }; \
template<> struct TypeVec<type ## 8, 8> { typedef type ## 8 vec_type; };
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uchar)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(char)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(ushort)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(short)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(int)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uint)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(float)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(double)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uchar)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(char)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(ushort)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(short)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(int)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(uint)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(float)
OPENCV_GPU_IMPLEMENT_TYPE_VEC(double)
#undef OPENCV_GPU_IMPLEMENT_TYPE_VEC
#undef OPENCV_GPU_IMPLEMENT_TYPE_VEC
template<> struct TypeVec<schar, 1> { typedef schar vec_type; };
template<> struct TypeVec<schar, 2> { typedef char2 vec_type; };
template<> struct TypeVec<schar, 3> { typedef char3 vec_type; };
template<> struct TypeVec<schar, 4> { typedef char4 vec_type; };
template<> struct TypeVec<schar, 8> { typedef char8 vec_type; };
template<> struct TypeVec<schar, 1> { typedef schar vec_type; };
template<> struct TypeVec<schar, 2> { typedef char2 vec_type; };
template<> struct TypeVec<schar, 3> { typedef char3 vec_type; };
template<> struct TypeVec<schar, 4> { typedef char4 vec_type; };
template<> struct TypeVec<schar, 8> { typedef char8 vec_type; };
template<> struct TypeVec<bool, 1> { typedef uchar vec_type; };
template<> struct TypeVec<bool, 2> { typedef uchar2 vec_type; };
template<> struct TypeVec<bool, 3> { typedef uchar3 vec_type; };
template<> struct TypeVec<bool, 4> { typedef uchar4 vec_type; };
template<> struct TypeVec<bool, 8> { typedef uchar8 vec_type; };
template<> struct TypeVec<bool, 1> { typedef uchar vec_type; };
template<> struct TypeVec<bool, 2> { typedef uchar2 vec_type; };
template<> struct TypeVec<bool, 3> { typedef uchar3 vec_type; };
template<> struct TypeVec<bool, 4> { typedef uchar4 vec_type; };
template<> struct TypeVec<bool, 8> { typedef uchar8 vec_type; };
template<typename T> struct VecTraits;
@ -209,73 +209,72 @@ template<> struct TypeVec<bool, 8> { typedef uchar8 vec_type; };
static __device__ __host__ __forceinline__ type ## 8 make(const type* v) {return make_ ## type ## 8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);} \
};
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uchar)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(ushort)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(short)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(int)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uint)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(float)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(double)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uchar)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(ushort)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(short)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(int)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(uint)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(float)
OPENCV_GPU_IMPLEMENT_VEC_TRAITS(double)
#undef OPENCV_GPU_IMPLEMENT_VEC_TRAITS
#undef OPENCV_GPU_IMPLEMENT_VEC_TRAITS
template<> struct VecTraits<char>
{
typedef char elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char all(char v) {return v;}
static __device__ __host__ __forceinline__ char make(char x) {return x;}
static __device__ __host__ __forceinline__ char make(const char* x) {return *x;}
};
template<> struct VecTraits<schar>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ schar all(schar v) {return v;}
static __device__ __host__ __forceinline__ schar make(schar x) {return x;}
static __device__ __host__ __forceinline__ schar make(const schar* x) {return *x;}
};
template<> struct VecTraits<char1>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char1 all(schar v) {return make_char1(v);}
static __device__ __host__ __forceinline__ char1 make(schar x) {return make_char1(x);}
static __device__ __host__ __forceinline__ char1 make(const schar* v) {return make_char1(v[0]);}
};
template<> struct VecTraits<char2>
{
typedef schar elem_type;
enum {cn=2};
static __device__ __host__ __forceinline__ char2 all(schar v) {return make_char2(v, v);}
static __device__ __host__ __forceinline__ char2 make(schar x, schar y) {return make_char2(x, y);}
static __device__ __host__ __forceinline__ char2 make(const schar* v) {return make_char2(v[0], v[1]);}
};
template<> struct VecTraits<char3>
{
typedef schar elem_type;
enum {cn=3};
static __device__ __host__ __forceinline__ char3 all(schar v) {return make_char3(v, v, v);}
static __device__ __host__ __forceinline__ char3 make(schar x, schar y, schar z) {return make_char3(x, y, z);}
static __device__ __host__ __forceinline__ char3 make(const schar* v) {return make_char3(v[0], v[1], v[2]);}
};
template<> struct VecTraits<char4>
{
typedef schar elem_type;
enum {cn=4};
static __device__ __host__ __forceinline__ char4 all(schar v) {return make_char4(v, v, v, v);}
static __device__ __host__ __forceinline__ char4 make(schar x, schar y, schar z, schar w) {return make_char4(x, y, z, w);}
static __device__ __host__ __forceinline__ char4 make(const schar* v) {return make_char4(v[0], v[1], v[2], v[3]);}
};
template<> struct VecTraits<char8>
{
typedef schar elem_type;
enum {cn=8};
static __device__ __host__ __forceinline__ char8 all(schar v) {return make_char8(v, v, v, v, v, v, v, v);}
static __device__ __host__ __forceinline__ char8 make(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7) {return make_char8(a0, a1, a2, a3, a4, a5, a6, a7);}
static __device__ __host__ __forceinline__ char8 make(const schar* v) {return make_char8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);}
};
END_OPENCV_DEVICE_NAMESPACE
template<> struct VecTraits<char>
{
typedef char elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char all(char v) {return v;}
static __device__ __host__ __forceinline__ char make(char x) {return x;}
static __device__ __host__ __forceinline__ char make(const char* x) {return *x;}
};
template<> struct VecTraits<schar>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ schar all(schar v) {return v;}
static __device__ __host__ __forceinline__ schar make(schar x) {return x;}
static __device__ __host__ __forceinline__ schar make(const schar* x) {return *x;}
};
template<> struct VecTraits<char1>
{
typedef schar elem_type;
enum {cn=1};
static __device__ __host__ __forceinline__ char1 all(schar v) {return make_char1(v);}
static __device__ __host__ __forceinline__ char1 make(schar x) {return make_char1(x);}
static __device__ __host__ __forceinline__ char1 make(const schar* v) {return make_char1(v[0]);}
};
template<> struct VecTraits<char2>
{
typedef schar elem_type;
enum {cn=2};
static __device__ __host__ __forceinline__ char2 all(schar v) {return make_char2(v, v);}
static __device__ __host__ __forceinline__ char2 make(schar x, schar y) {return make_char2(x, y);}
static __device__ __host__ __forceinline__ char2 make(const schar* v) {return make_char2(v[0], v[1]);}
};
template<> struct VecTraits<char3>
{
typedef schar elem_type;
enum {cn=3};
static __device__ __host__ __forceinline__ char3 all(schar v) {return make_char3(v, v, v);}
static __device__ __host__ __forceinline__ char3 make(schar x, schar y, schar z) {return make_char3(x, y, z);}
static __device__ __host__ __forceinline__ char3 make(const schar* v) {return make_char3(v[0], v[1], v[2]);}
};
template<> struct VecTraits<char4>
{
typedef schar elem_type;
enum {cn=4};
static __device__ __host__ __forceinline__ char4 all(schar v) {return make_char4(v, v, v, v);}
static __device__ __host__ __forceinline__ char4 make(schar x, schar y, schar z, schar w) {return make_char4(x, y, z, w);}
static __device__ __host__ __forceinline__ char4 make(const schar* v) {return make_char4(v[0], v[1], v[2], v[3]);}
};
template<> struct VecTraits<char8>
{
typedef schar elem_type;
enum {cn=8};
static __device__ __host__ __forceinline__ char8 all(schar v) {return make_char8(v, v, v, v, v, v, v, v);}
static __device__ __host__ __forceinline__ char8 make(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7) {return make_char8(a0, a1, a2, a3, a4, a5, a6, a7);}
static __device__ __host__ __forceinline__ char8 make(const schar* v) {return make_char8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);}
};
}}} // namespace cv { namespace gpu { namespace device
#endif // __OPENCV_GPU_VEC_TRAITS_HPP__

111
modules/gpu/src/opencv2/gpu/device/warp.hpp
Unescape View File

@ -45,71 +45,70 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
struct Warp
namespace cv { namespace gpu { namespace device
{
enum
{
LOG_WARP_SIZE = 5,
WARP_SIZE = 1 << LOG_WARP_SIZE,
STRIDE = WARP_SIZE
};
/** \brief Returns the warp lane ID of the calling thread. */
static __device__ __forceinline__ unsigned int laneId()
struct Warp
{
unsigned int ret;
asm("mov.u32 %0, %laneid;" : "=r"(ret) );
return ret;
}
enum
{
LOG_WARP_SIZE = 5,
WARP_SIZE = 1 << LOG_WARP_SIZE,
STRIDE = WARP_SIZE
};
template<typename It, typename T>
static __device__ __forceinline__ void fill(It beg, It end, const T& value)
{
for(It t = beg + laneId(); t < end; t += STRIDE)
*t = value;
}
/** \brief Returns the warp lane ID of the calling thread. */
static __device__ __forceinline__ unsigned int laneId()
{
unsigned int ret;
asm("mov.u32 %0, %laneid;" : "=r"(ret) );
return ret;
}
template<typename InIt, typename OutIt>
static __device__ __forceinline__ OutIt copy(InIt beg, InIt end, OutIt out)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = *t;
return out;
}
template<typename It, typename T>
static __device__ __forceinline__ void fill(It beg, It end, const T& value)
{
for(It t = beg + laneId(); t < end; t += STRIDE)
*t = value;
}
template<typename InIt, typename OutIt, class UnOp>
static __device__ __forceinline__ OutIt transform(InIt beg, InIt end, OutIt out, UnOp op)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = op(*t);
return out;
}
template<typename InIt, typename OutIt>
static __device__ __forceinline__ OutIt copy(InIt beg, InIt end, OutIt out)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = *t;
return out;
}
template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
static __device__ __forceinline__ OutIt transform(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
{
unsigned int lane = laneId();
template<typename InIt, typename OutIt, class UnOp>
static __device__ __forceinline__ OutIt transform(InIt beg, InIt end, OutIt out, UnOp op)
{
for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
*out = op(*t);
return out;
}
InIt1 t1 = beg1 + lane;
InIt2 t2 = beg2 + lane;
for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, out += STRIDE)
*out = op(*t1, *t2);
return out;
}
template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
static __device__ __forceinline__ OutIt transform(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
{
unsigned int lane = laneId();
template<typename OutIt, typename T>
static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
{
unsigned int lane = laneId();
value += lane;
InIt1 t1 = beg1 + lane;
InIt2 t2 = beg2 + lane;
for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, out += STRIDE)
*out = op(*t1, *t2);
return out;
}
for(OutIt t = beg + lane; t < end; t += STRIDE, value += STRIDE)
*t = value;
}
};
template<typename OutIt, typename T>
static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
{
unsigned int lane = laneId();
value += lane;
END_OPENCV_DEVICE_NAMESPACE
for(OutIt t = beg + lane; t < end; t += STRIDE, value += STRIDE)
*t = value;
}
};
}}} // namespace cv { namespace gpu { namespace device
#endif /* __OPENCV_GPU_DEVICE_WARP_HPP__ */

39
modules/gpu/src/opencv2/gpu/device/warp_reduce.hpp
Unescape View File

@ -46,27 +46,26 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template <class T>
__device__ __forceinline__ T warp_reduce ( volatile T *ptr , const unsigned int tid = threadIdx.x )
{
const unsigned int lane = tid & 31; // index of thread in warp (0..31)
if (lane < 16)
{
T partial = ptr[tid];
namespace cv { namespace gpu { namespace device
{
template <class T>
__device__ __forceinline__ T warp_reduce(volatile T *ptr , const unsigned int tid = threadIdx.x)
{
const unsigned int lane = tid & 31; // index of thread in warp (0..31)
if (lane < 16)
{
T partial = ptr[tid];
ptr[tid] = partial = partial + ptr[tid + 16];
ptr[tid] = partial = partial + ptr[tid + 8];
ptr[tid] = partial = partial + ptr[tid + 4];
ptr[tid] = partial = partial + ptr[tid + 2];
ptr[tid] = partial = partial + ptr[tid + 1];
}
ptr[tid] = partial = partial + ptr[tid + 16];
ptr[tid] = partial = partial + ptr[tid + 8];
ptr[tid] = partial = partial + ptr[tid + 4];
ptr[tid] = partial = partial + ptr[tid + 2];
ptr[tid] = partial = partial + ptr[tid + 1];
}
return ptr[tid - lane];
}
END_OPENCV_DEVICE_NAMESPACE
return ptr[tid - lane];
}
}}} // namespace cv { namespace gpu { namespace device {
#endif /* OPENCV_GPU_WARP_REDUCE_HPP__ */

21
modules/gpu/src/split_merge.cpp
Unescape View File

@ -55,21 +55,20 @@ void cv::gpu::split(const GpuMat& /*src*/, vector<GpuMat>& /*dst*/, Stream& /*st
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace split_merge
{
void merge_caller(const DevMem2Db* src, DevMem2Db& dst, int total_channels, size_t elem_size, const cudaStream_t& stream);
void split_caller(const DevMem2Db& src, DevMem2Db* dst, int num_channels, size_t elem_size1, const cudaStream_t& stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
namespace split_merge
{
void merge_caller(const DevMem2Db* src, DevMem2Db& dst, int total_channels, size_t elem_size, const cudaStream_t& stream);
void split_caller(const DevMem2Db& src, DevMem2Db* dst, int num_channels, size_t elem_size1, const cudaStream_t& stream);
}
}}}
namespace
{
void merge(const GpuMat* src, size_t n, GpuMat& dst, const cudaStream_t& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ split_merge;
using namespace ::cv::gpu::device::split_merge;
CV_Assert(src);
CV_Assert(n > 0);
@ -108,7 +107,7 @@ namespace
void split(const GpuMat& src, GpuMat* dst, const cudaStream_t& stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ split_merge;
using namespace ::cv::gpu::device::split_merge;
CV_Assert(dst);

19
modules/gpu/src/stereobm.cpp
Unescape View File

@ -55,16 +55,15 @@ void cv::gpu::StereoBM_GPU::operator() ( const GpuMat&, const GpuMat&, GpuMat&,
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace stereobm
namespace cv { namespace gpu { namespace device
{
void stereoBM_GPU(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int ndisp, int winsz, const DevMem2D_<unsigned int>& minSSD_buf, cudaStream_t & stream);
void prefilter_xsobel(const DevMem2Db& input, const DevMem2Db& output, int prefilterCap /*= 31*/, cudaStream_t & stream);
void postfilter_textureness(const DevMem2Db& input, int winsz, float avgTexturenessThreshold, const DevMem2Db& disp, cudaStream_t & stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace stereobm
{
void stereoBM_GPU(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int ndisp, int winsz, const DevMem2D_<unsigned int>& minSSD_buf, cudaStream_t & stream);
void prefilter_xsobel(const DevMem2Db& input, const DevMem2Db& output, int prefilterCap /*= 31*/, cudaStream_t & stream);
void postfilter_textureness(const DevMem2Db& input, int winsz, float avgTexturenessThreshold, const DevMem2Db& disp, cudaStream_t & stream);
}
}}}
const float defaultAvgTexThreshold = 3;
@ -99,7 +98,7 @@ namespace
{
void stereo_bm_gpu_operator( GpuMat& minSSD, GpuMat& leBuf, GpuMat& riBuf, int preset, int ndisp, int winSize, float avergeTexThreshold, const GpuMat& left, const GpuMat& right, GpuMat& disparity, cudaStream_t stream)
{
using namespace OPENCV_DEVICE_NAMESPACE_ stereobm;
using namespace ::cv::gpu::device::stereobm;
CV_DbgAssert(left.rows == right.rows && left.cols == right.cols);
CV_DbgAssert(left.type() == CV_8UC1);

41
modules/gpu/src/stereobp.cpp
Unescape View File

@ -59,28 +59,27 @@ void cv::gpu::StereoBeliefPropagation::operator()(const GpuMat&, GpuMat&, Stream
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace stereobp
namespace cv { namespace gpu { namespace device
{
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump);
template<typename T, typename D>
void comp_data_gpu(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream);
template<typename T>
void data_step_down_gpu(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
template <typename T>
void level_up_messages_gpu(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
template <typename T>
void calc_all_iterations_gpu(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d,
const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
template <typename T>
void output_gpu(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data,
const DevMem2D_<short>& disp, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
using namespace OPENCV_DEVICE_NAMESPACE_ stereobp;
namespace stereobp
{
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump);
template<typename T, typename D>
void comp_data_gpu(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream);
template<typename T>
void data_step_down_gpu(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
template <typename T>
void level_up_messages_gpu(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
template <typename T>
void calc_all_iterations_gpu(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d,
const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
template <typename T>
void output_gpu(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data,
const DevMem2D_<short>& disp, cudaStream_t stream);
}
}}}
using namespace ::cv::gpu::device::stereobp;
namespace
{

63
modules/gpu/src/stereocsbp.cpp
Unescape View File

@ -57,40 +57,39 @@ void cv::gpu::StereoConstantSpaceBP::operator()(const GpuMat&, const GpuMat&, Gp
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace stereocsbp
namespace cv { namespace gpu { namespace device
{
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump, int min_disp_th,
const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& temp);
template<class T>
void init_data_cost(int rows, int cols, T* disp_selected_pyr, T* data_cost_selected, size_t msg_step,
int h, int w, int level, int nr_plane, int ndisp, int channels, bool use_local_init_data_cost, cudaStream_t stream);
template<class T>
void compute_data_cost(const T* disp_selected_pyr, T* data_cost, size_t msg_step1, size_t msg_step2,
int rows, int cols, int h, int w, int h2, int level, int nr_plane, int channels, cudaStream_t stream);
template<class T>
void init_message(T* u_new, T* d_new, T* l_new, T* r_new,
const T* u_cur, const T* d_cur, const T* l_cur, const T* r_cur,
T* selected_disp_pyr_new, const T* selected_disp_pyr_cur,
T* data_cost_selected, const T* data_cost, size_t msg_step1, size_t msg_step2,
int h, int w, int nr_plane, int h2, int w2, int nr_plane2, cudaStream_t stream);
template<class T>
void calc_all_iterations(T* u, T* d, T* l, T* r, const T* data_cost_selected,
const T* selected_disp_pyr_cur, size_t msg_step, int h, int w, int nr_plane, int iters, cudaStream_t stream);
template<class T>
void compute_disp(const T* u, const T* d, const T* l, const T* r, const T* data_cost_selected, const T* disp_selected, size_t msg_step,
const DevMem2D_<short>& disp, int nr_plane, cudaStream_t stream);
}
END_OPENCV_DEVICE_NAMESPACE
namespace stereocsbp
{
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump, int min_disp_th,
const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& temp);
template<class T>
void init_data_cost(int rows, int cols, T* disp_selected_pyr, T* data_cost_selected, size_t msg_step,
int h, int w, int level, int nr_plane, int ndisp, int channels, bool use_local_init_data_cost, cudaStream_t stream);
template<class T>
void compute_data_cost(const T* disp_selected_pyr, T* data_cost, size_t msg_step1, size_t msg_step2,
int rows, int cols, int h, int w, int h2, int level, int nr_plane, int channels, cudaStream_t stream);
template<class T>
void init_message(T* u_new, T* d_new, T* l_new, T* r_new,
const T* u_cur, const T* d_cur, const T* l_cur, const T* r_cur,
T* selected_disp_pyr_new, const T* selected_disp_pyr_cur,
T* data_cost_selected, const T* data_cost, size_t msg_step1, size_t msg_step2,
int h, int w, int nr_plane, int h2, int w2, int nr_plane2, cudaStream_t stream);
template<class T>
void calc_all_iterations(T* u, T* d, T* l, T* r, const T* data_cost_selected,
const T* selected_disp_pyr_cur, size_t msg_step, int h, int w, int nr_plane, int iters, cudaStream_t stream);
template<class T>
void compute_disp(const T* u, const T* d, const T* l, const T* r, const T* data_cost_selected, const T* disp_selected, size_t msg_step,
const DevMem2D_<short>& disp, int nr_plane, cudaStream_t stream);
}
}}}
using namespace OPENCV_DEVICE_NAMESPACE_ stereocsbp;
using namespace ::cv::gpu::device::stereocsbp;
namespace
{

41
modules/gpu/src/surf.cpp
Unescape View File

@ -63,35 +63,34 @@ void cv::gpu::SURF_GPU::releaseMemory() { throw_nogpu(); }
#else /* !defined (HAVE_CUDA) */
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace surf
namespace cv { namespace gpu { namespace device
{
void loadGlobalConstants(int maxCandidates, int maxFeatures, int img_rows, int img_cols, int nOctaveLayers, float hessianThreshold);
void loadOctaveConstants(int octave, int layer_rows, int layer_cols);
namespace surf
{
void loadGlobalConstants(int maxCandidates, int maxFeatures, int img_rows, int img_cols, int nOctaveLayers, float hessianThreshold);
void loadOctaveConstants(int octave, int layer_rows, int layer_cols);
void bindImgTex(DevMem2Db img);
void bindSumTex(DevMem2D_<unsigned int> sum);
void bindMaskSumTex(DevMem2D_<unsigned int> maskSum);
void bindImgTex(DevMem2Db img);
void bindSumTex(DevMem2D_<unsigned int> sum);
void bindMaskSumTex(DevMem2D_<unsigned int> maskSum);
void icvCalcLayerDetAndTrace_gpu(const PtrStepf& det, const PtrStepf& trace, int img_rows, int img_cols, int octave, int nOctaveLayers);
void icvCalcLayerDetAndTrace_gpu(const PtrStepf& det, const PtrStepf& trace, int img_rows, int img_cols, int octave, int nOctaveLayers);
void icvFindMaximaInLayer_gpu(const PtrStepf& det, const PtrStepf& trace, int4* maxPosBuffer, unsigned int* maxCounter,
int img_rows, int img_cols, int octave, bool use_mask, int nLayers);
void icvFindMaximaInLayer_gpu(const PtrStepf& det, const PtrStepf& trace, int4* maxPosBuffer, unsigned int* maxCounter,
int img_rows, int img_cols, int octave, bool use_mask, int nLayers);
void icvInterpolateKeypoint_gpu(const PtrStepf& det, const int4* maxPosBuffer, unsigned int maxCounter,
float* featureX, float* featureY, int* featureLaplacian, float* featureSize, float* featureHessian,
unsigned int* featureCounter);
void icvInterpolateKeypoint_gpu(const PtrStepf& det, const int4* maxPosBuffer, unsigned int maxCounter,
float* featureX, float* featureY, int* featureLaplacian, float* featureSize, float* featureHessian,
unsigned int* featureCounter);
void icvCalcOrientation_gpu(const float* featureX, const float* featureY, const float* featureSize, float* featureDir, int nFeatures);
void icvCalcOrientation_gpu(const float* featureX, const float* featureY, const float* featureSize, float* featureDir, int nFeatures);
void compute_descriptors_gpu(const DevMem2Df& descriptors,
const float* featureX, const float* featureY, const float* featureSize, const float* featureDir, int nFeatures);
}
END_OPENCV_DEVICE_NAMESPACE
void compute_descriptors_gpu(const DevMem2Df& descriptors,
const float* featureX, const float* featureY, const float* featureSize, const float* featureDir, int nFeatures);
}
}}}
using namespace OPENCV_DEVICE_NAMESPACE_ surf;
using namespace ::cv::gpu::device::surf;
namespace
{

2
modules/gpu/test/test_video.cpp
Unescape View File

@ -225,7 +225,7 @@ TEST_P(InterpolateFrames, Regression)
#ifndef DUMP
EXPECT_MAT_NEAR(newFrame_gold, newFrame, 1e-4);
EXPECT_MAT_NEAR(newFrame_gold, newFrame, 1e-3);
#else

Write Preview
Loading…
Cancel
Save