Chiebot-Mirror
/
opencv
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fixed bug in gpu filter engine (incorrect buffer type) and in vector's saturate_cast.

Browse Source 
changed buffer type in linear filters to float.
added support of 1 channel image to linear filters.
added support of BORDER_REFLECT101, BORDER_REPLICATE and BORDER_CONSTANT border type to gpu linear filters.
minor fix in tests.
update comments in gpu.hpp.
pull/13383/head
Vladislav Vinogradov 15 years ago
parent 108ab94023
commit 49ec8ba742
9 changed files with 676 additions and 309 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 50
      modules/gpu/include/opencv2/gpu/gpu.hpp
  2. 4
      modules/gpu/src/arithm.cpp
  3. 503
      modules/gpu/src/cuda/filters.cu
  4. 3
      modules/gpu/src/cuda/internal_shared.hpp
  5. 240
      modules/gpu/src/filtering.cpp
  6. 6
      modules/gpu/src/imgproc_gpu.cpp
  7. 149
      modules/gpu/src/opencv2/gpu/device/vecmath.hpp
  8. 6
      tests/gpu/src/brute_force_matcher.cpp
  9. 14
      tests/gpu/src/filters.cpp

50
modules/gpu/include/opencv2/gpu/gpu.hpp
Unescape View File

@ -388,7 +388,7 @@ namespace cv
CV_EXPORTS void divide(const GpuMat& a, const Scalar& sc, GpuMat& c); CV_EXPORTS void divide(const GpuMat& a, const Scalar& sc, GpuMat& c);
//! transposes the matrix //! transposes the matrix
//! supports only CV_8UC1 type //! supports CV_8UC1, CV_8SC1, CV_8UC4, CV_8SC4, CV_16UC2, CV_16SC2, CV_32SC1, CV_32FC1 type
CV_EXPORTS void transpose(const GpuMat& src1, GpuMat& dst); CV_EXPORTS void transpose(const GpuMat& src1, GpuMat& dst);
//! computes element-wise absolute difference of two arrays (c = abs(a - b)) //! computes element-wise absolute difference of two arrays (c = abs(a - b))
@ -725,11 +725,11 @@ namespace cv
}; };
//! returns the non-separable filter engine with the specified filter //! returns the non-separable filter engine with the specified filter
CV_EXPORTS Ptr<FilterEngine_GPU> createFilter2D_GPU(const Ptr<BaseFilter_GPU> filter2D); CV_EXPORTS Ptr<FilterEngine_GPU> createFilter2D_GPU(const Ptr<BaseFilter_GPU> filter2D, int srcType, int dstType);
//! returns the separable filter engine with the specified filters //! returns the separable filter engine with the specified filters
CV_EXPORTS Ptr<FilterEngine_GPU> createSeparableFilter_GPU(const Ptr<BaseRowFilter_GPU>& rowFilter, CV_EXPORTS Ptr<FilterEngine_GPU> createSeparableFilter_GPU(const Ptr<BaseRowFilter_GPU>& rowFilter,
const Ptr<BaseColumnFilter_GPU>& columnFilter); const Ptr<BaseColumnFilter_GPU>& columnFilter, int srcType, int bufType, int dstType);
//! returns horizontal 1D box filter //! returns horizontal 1D box filter
//! supports only CV_8UC1 source type and CV_32FC1 sum type //! supports only CV_8UC1 source type and CV_32FC1 sum type
@ -767,23 +767,40 @@ namespace cv
CV_EXPORTS Ptr<FilterEngine_GPU> createLinearFilter_GPU(int srcType, int dstType, const Mat& kernel, CV_EXPORTS Ptr<FilterEngine_GPU> createLinearFilter_GPU(int srcType, int dstType, const Mat& kernel,
const Point& anchor = Point(-1,-1)); const Point& anchor = Point(-1,-1));
//! returns the primitive row filter with the specified kernel //! returns the primitive row filter with the specified kernel.
//! supports only CV_8UC1, CV_8UC4, CV_16SC1, CV_16SC2, CV_32SC1, CV_32FC1 source type.
//! there are two version of algorithm: NPP and OpenCV.
//! NPP calls when srcType == CV_8UC1 or srcType == CV_8UC4 and bufType == srcType,
//! otherwise calls OpenCV version.
//! NPP supports only BORDER_CONSTANT border type.
//! OpenCV version supports only CV_32F as buffer depth and
//! BORDER_REFLECT101, BORDER_REPLICATE and BORDER_CONSTANT border types.
CV_EXPORTS Ptr<BaseRowFilter_GPU> getLinearRowFilter_GPU(int srcType, int bufType, const Mat& rowKernel, CV_EXPORTS Ptr<BaseRowFilter_GPU> getLinearRowFilter_GPU(int srcType, int bufType, const Mat& rowKernel,
int anchor = -1); int anchor = -1, int borderType = BORDER_CONSTANT);
//! returns the primitive column filter with the specified kernel //! returns the primitive column filter with the specified kernel.
//! supports only CV_8UC1, CV_8UC4, CV_16SC1, CV_16SC2, CV_32SC1, CV_32FC1 dst type.
//! there are two version of algorithm: NPP and OpenCV.
//! NPP calls when dstType == CV_8UC1 or dstType == CV_8UC4 and bufType == dstType,
//! otherwise calls OpenCV version.
//! NPP supports only BORDER_CONSTANT border type.
//! OpenCV version supports only CV_32F as buffer depth and
//! BORDER_REFLECT101, BORDER_REPLICATE and BORDER_CONSTANT border types.
CV_EXPORTS Ptr<BaseColumnFilter_GPU> getLinearColumnFilter_GPU(int bufType, int dstType, const Mat& columnKernel, CV_EXPORTS Ptr<BaseColumnFilter_GPU> getLinearColumnFilter_GPU(int bufType, int dstType, const Mat& columnKernel,
int anchor = -1); int anchor = -1, int borderType = BORDER_CONSTANT);
//! returns the separable linear filter engine //! returns the separable linear filter engine
CV_EXPORTS Ptr<FilterEngine_GPU> createSeparableLinearFilter_GPU(int srcType, int dstType, const Mat& rowKernel, CV_EXPORTS Ptr<FilterEngine_GPU> createSeparableLinearFilter_GPU(int srcType, int dstType, const Mat& rowKernel,
const Mat& columnKernel, const Point& anchor = Point(-1,-1)); const Mat& columnKernel, const Point& anchor = Point(-1,-1), int rowBorderType = BORDER_DEFAULT,
int columnBorderType = -1);
//! returns filter engine for the generalized Sobel operator //! returns filter engine for the generalized Sobel operator
CV_EXPORTS Ptr<FilterEngine_GPU> createDerivFilter_GPU(int srcType, int dstType, int dx, int dy, int ksize); CV_EXPORTS Ptr<FilterEngine_GPU> createDerivFilter_GPU(int srcType, int dstType, int dx, int dy, int ksize,
int rowBorderType = BORDER_DEFAULT, int columnBorderType = -1);
//! returns the Gaussian filter engine //! returns the Gaussian filter engine
CV_EXPORTS Ptr<FilterEngine_GPU> createGaussianFilter_GPU(int type, Size ksize, double sigma1, double sigma2 = 0); CV_EXPORTS Ptr<FilterEngine_GPU> createGaussianFilter_GPU(int type, Size ksize, double sigma1, double sigma2 = 0,
int rowBorderType = BORDER_DEFAULT, int columnBorderType = -1);
//! returns maximum filter //! returns maximum filter
CV_EXPORTS Ptr<BaseFilter_GPU> getMaxFilter_GPU(int srcType, int dstType, const Size& ksize, Point anchor = Point(-1,-1)); CV_EXPORTS Ptr<BaseFilter_GPU> getMaxFilter_GPU(int srcType, int dstType, const Size& ksize, Point anchor = Point(-1,-1));
@ -812,16 +829,19 @@ namespace cv
//! applies separable 2D linear filter to the image //! applies separable 2D linear filter to the image
CV_EXPORTS void sepFilter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& kernelX, const Mat& kernelY, CV_EXPORTS void sepFilter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& kernelX, const Mat& kernelY,
Point anchor = Point(-1,-1)); Point anchor = Point(-1,-1), int rowBorderType = BORDER_DEFAULT, int columnBorderType = -1);
//! applies generalized Sobel operator to the image //! applies generalized Sobel operator to the image
CV_EXPORTS void Sobel(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, int ksize = 3, double scale = 1); CV_EXPORTS void Sobel(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, int ksize = 3, double scale = 1,
int rowBorderType = BORDER_DEFAULT, int columnBorderType = -1);
//! applies the vertical or horizontal Scharr operator to the image //! applies the vertical or horizontal Scharr operator to the image
CV_EXPORTS void Scharr(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, double scale = 1); CV_EXPORTS void Scharr(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, double scale = 1,
int rowBorderType = BORDER_DEFAULT, int columnBorderType = -1);
//! smooths the image using Gaussian filter. //! smooths the image using Gaussian filter.
CV_EXPORTS void GaussianBlur(const GpuMat& src, GpuMat& dst, Size ksize, double sigma1, double sigma2 = 0); CV_EXPORTS void GaussianBlur(const GpuMat& src, GpuMat& dst, Size ksize, double sigma1, double sigma2 = 0,
int rowBorderType = BORDER_DEFAULT, int columnBorderType = -1);
//! applies Laplacian operator to the image //! applies Laplacian operator to the image
//! supports only ksize = 1 and ksize = 3 //! supports only ksize = 1 and ksize = 3

4
modules/gpu/src/arithm.cpp
Unescape View File

@ -277,12 +277,12 @@ namespace cv { namespace gpu { namespace mathfunc
void cv::gpu::transpose(const GpuMat& src, GpuMat& dst) void cv::gpu::transpose(const GpuMat& src, GpuMat& dst)
{ {
CV_Assert(src.type() == CV_8UC1 || src.type() == CV_8UC4 || src.type() == CV_8SC4 CV_Assert(src.type() == CV_8UC1 || src.type() == CV_8SC1 || src.type() == CV_8UC4 || src.type() == CV_8SC4
|| src.type() == CV_16UC2 || src.type() == CV_16SC2 || src.type() == CV_32SC1 || src.type() == CV_32FC1); || src.type() == CV_16UC2 || src.type() == CV_16SC2 || src.type() == CV_32SC1 || src.type() == CV_32FC1);
dst.create( src.cols, src.rows, src.type() ); dst.create( src.cols, src.rows, src.type() );
if (src.type() == CV_8UC1) if (src.type() == CV_8UC1 || src.type() == CV_8SC1)
{ {
NppiSize sz; NppiSize sz;
sz.width = src.cols; sz.width = src.cols;

503
modules/gpu/src/cuda/filters.cu
Unescape View File

@ -43,6 +43,7 @@
#include "opencv2/gpu/devmem2d.hpp" #include "opencv2/gpu/devmem2d.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp" #include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/vecmath.hpp" #include "opencv2/gpu/device/vecmath.hpp"
#include "opencv2/gpu/device/limits_gpu.hpp"
#include "safe_call.hpp" #include "safe_call.hpp"
#include "internal_shared.hpp" #include "internal_shared.hpp"
@ -50,14 +51,198 @@
using namespace cv::gpu; using namespace cv::gpu;
using namespace cv::gpu::device; using namespace cv::gpu::device;
#ifndef FLT_MAX namespace cv
#define FLT_MAX 3.402823466e+30F {
#endif namespace gpu
{
namespace device
{
struct BrdReflect101
{
explicit BrdReflect101(int len): last(len - 1) {}
__device__ int idx_low(int i) const
{
return abs(i);
}
__device__ int idx_high(int i) const
{
return last - abs(last - i);
}
__device__ int idx(int i) const
{
return abs(idx_high(i));
}
bool is_range_safe(int mini, int maxi) const
{
return -last <= mini && maxi <= 2 * last;
}
int last;
};
template <typename D>
struct BrdRowReflect101: BrdReflect101
{
explicit BrdRowReflect101(int len): BrdReflect101(len) {}
template <typename T>
__device__ D at_low(int i, const T* data) const
{
return saturate_cast<D>(data[idx_low(i)]);
}
template <typename T>
__device__ D at_high(int i, const T* data) const
{
return saturate_cast<D>(data[idx_high(i)]);
}
};
template <typename D>
struct BrdColReflect101: BrdReflect101
{
BrdColReflect101(int len, int step): BrdReflect101(len), step(step) {}
template <typename T>
__device__ D at_low(int i, const T* data) const
{
return saturate_cast<D>(data[idx_low(i) * step]);
}
template <typename T>
__device__ D at_high(int i, const T* data) const
{
return saturate_cast<D>(data[idx_high(i) * step]);
}
int step;
};
struct BrdReplicate
{
explicit BrdReplicate(int len): last(len - 1) {}
__device__ int idx_low(int i) const
{
return max(i, 0);
}
__device__ int idx_high(int i) const
{
return min(i, last);
}
__device__ int idx(int i) const
{
return max(min(i, last), 0);
}
bool is_range_safe(int mini, int maxi) const
{
return true;
}
int last;
};
template <typename D>
struct BrdRowReplicate: BrdReplicate
{
explicit BrdRowReplicate(int len): BrdReplicate(len) {}
template <typename T>
__device__ D at_low(int i, const T* data) const
{
return saturate_cast<D>(data[idx_low(i)]);
}
template <typename T>
__device__ D at_high(int i, const T* data) const
{
return saturate_cast<D>(data[idx_high(i)]);
}
};
template <typename D>
struct BrdColReplicate: BrdReplicate
{
BrdColReplicate(int len, int step): BrdReplicate(len), step(step) {}
template <typename T>
__device__ D at_low(int i, const T* data) const
{
return saturate_cast<D>(data[idx_low(i) * step]);
}
template <typename T>
__device__ D at_high(int i, const T* data) const
{
return saturate_cast<D>(data[idx_high(i) * step]);
}
int step;
};
template <typename D>
struct BrdRowConstant
{
explicit BrdRowConstant(int len_, const D& val_ = VecTraits<D>::all(0)): len(len_), val(val_) {}
template <typename T>
__device__ D at_low(int i, const T* data) const
{
return i >= 0 ? saturate_cast<D>(data[i]) : val;
}
template <typename T>
__device__ D at_high(int i, const T* data) const
{
return i < len ? saturate_cast<D>(data[i]) : val;
}
bool is_range_safe(int mini, int maxi) const
{
return true;
}
int len;
D val;
};
template <typename D>
struct BrdColConstant
{
BrdColConstant(int len_, int step_, const D& val_ = VecTraits<D>::all(0)): len(len_), step(step_), val(val_) {}
template <typename T>
__device__ D at_low(int i, const T* data) const
{
return i >= 0 ? saturate_cast<D>(data[i * step]) : val;
}
template <typename T>
__device__ D at_high(int i, const T* data) const
{
return i < len ? saturate_cast<D>(data[i * step]) : val;
}
bool is_range_safe(int mini, int maxi) const
{
return true;
}
int len;
int step;
D val;
};
}
}
}
///////////////////////////////////////////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////////////////////////////////////////
// Linear filters // Linear filters
#define MAX_KERNEL_SIZE 16 #define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 16
namespace filter_krnls namespace filter_krnls
{ {
@ -74,46 +259,53 @@ namespace cv { namespace gpu { namespace filters
namespace filter_krnls namespace filter_krnls
{ {
template <int BLOCK_DIM_X, int BLOCK_DIM_Y, int KERNEL_SIZE, typename T, typename D> template <typename T, size_t size> struct SmemType_
__global__ void linearRowFilter(const T* src, size_t src_step, D* dst, size_t dst_step, int anchor, int width, int height)
{ {
__shared__ T smem[BLOCK_DIM_Y * BLOCK_DIM_X * 3]; typedef typename TypeVec<float, VecTraits<T>::cn>::vec_t smem_t;
};
template <typename T> struct SmemType_<T, 4>
{
typedef T smem_t;
};
template <typename T> struct SmemType
{
typedef typename SmemType_<T, sizeof(T)>::smem_t smem_t;
};
const int blockStartX = blockDim.x * blockIdx.x; template <int ksize, typename T, typename D, typename B>
const int blockStartY = blockDim.y * blockIdx.y; __global__ void linearRowFilter(const DevMem2D_<T> src, PtrStep_<D> dst, int anchor, const B b)
{
typedef typename SmemType<T>::smem_t smem_t;
const int threadX = blockStartX + threadIdx.x; __shared__ smem_t smem[BLOCK_DIM_Y * BLOCK_DIM_X * 3];
const int prevThreadX = threadX - blockDim.x;
const int nextThreadX = threadX + blockDim.x;
const int threadY = blockStartY + threadIdx.y; const int x = BLOCK_DIM_X * blockIdx.x + threadIdx.x;
const int y = BLOCK_DIM_Y * blockIdx.y + threadIdx.y;
T* sDataRow = smem + threadIdx.y * blockDim.x * 3; smem_t* sDataRow = smem + threadIdx.y * BLOCK_DIM_X * 3;
if (threadY < height) if (y < src.rows)
{ {
const T* rowSrc = src + threadY * src_step; const T* rowSrc = src.ptr(y);
sDataRow[threadIdx.x + blockDim.x] = threadX < width ? rowSrc[threadX] : VecTraits<T>::all(0); sDataRow[threadIdx.x ] = b.at_low(x - BLOCK_DIM_X, rowSrc);
sDataRow[threadIdx.x + BLOCK_DIM_X ] = b.at_high(x, rowSrc);
sDataRow[threadIdx.x] = prevThreadX >= 0 ? rowSrc[prevThreadX] : VecTraits<T>::all(0); sDataRow[threadIdx.x + BLOCK_DIM_X * 2] = b.at_high(x + BLOCK_DIM_X, rowSrc);
sDataRow[(blockDim.x << 1) + threadIdx.x] = nextThreadX < width ? rowSrc[nextThreadX] : VecTraits<T>::all(0);
__syncthreads(); __syncthreads();
if (threadX < width) if (x < src.cols)
{ {
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_t sum_t; typedef typename TypeVec<float, VecTraits<T>::cn>::vec_t sum_t;
sum_t sum = VecTraits<sum_t>::all(0); sum_t sum = VecTraits<sum_t>::all(0);
sDataRow += threadIdx.x + blockDim.x - anchor; sDataRow += threadIdx.x + BLOCK_DIM_X - anchor;
#pragma unroll #pragma unroll
for(int i = 0; i < KERNEL_SIZE; ++i) for(int i = 0; i < ksize; ++i)
sum = sum + sDataRow[i] * cLinearKernel[i]; sum = sum + sDataRow[i] * cLinearKernel[i];
dst[threadY * dst_step + threadX] = saturate_cast<D>(sum); dst.ptr(y)[x] = saturate_cast<D>(sum);
} }
} }
} }
@ -121,100 +313,138 @@ namespace filter_krnls
namespace cv { namespace gpu { namespace filters namespace cv { namespace gpu { namespace filters
{ {
template <int KERNEL_SIZE, typename T, typename D> template <int ksize, typename T, typename D, template<typename> class B>
void linearRowFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor) void linearRowFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor)
{ {
const int BLOCK_DIM_X = 16;
const int BLOCK_DIM_Y = 16;
dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y); dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y);
dim3 blocks(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y)); dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y));
typedef typename filter_krnls::SmemType<T>::smem_t smem_t;
B<smem_t> b(src.cols);
filter_krnls::linearRowFilter<BLOCK_DIM_X, BLOCK_DIM_Y, KERNEL_SIZE><<<blocks, threads>>>(src.data, src.step/src.elemSize(), if (!b.is_range_safe(-BLOCK_DIM_X, (grid.x + 1) * BLOCK_DIM_X - 1))
dst.data, dst.step/dst.elemSize(), anchor, src.cols, src.rows); {
cv::gpu::error("linearRowFilter: can't use specified border extrapolation, image is too small, "
"try bigger image or another border extrapolation mode", __FILE__, __LINE__);
}
filter_krnls::linearRowFilter<ksize, T, D><<<grid, threads>>>(src, dst, anchor, b);
cudaSafeCall( cudaThreadSynchronize() ); cudaSafeCall( cudaThreadSynchronize() );
} }
template <typename T, typename D> template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor) void linearRowFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type)
{ {
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor); typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor);
static const caller_t callers[] = static const caller_t callers[3][17] =
{linearRowFilter_caller<0 , T, D>, linearRowFilter_caller<1 , T, D>, {
linearRowFilter_caller<2 , T, D>, linearRowFilter_caller<3 , T, D>, {
linearRowFilter_caller<4 , T, D>, linearRowFilter_caller<5 , T, D>, 0,
linearRowFilter_caller<6 , T, D>, linearRowFilter_caller<7 , T, D>, linearRowFilter_caller<1 , T, D, BrdRowReflect101>,
linearRowFilter_caller<8 , T, D>, linearRowFilter_caller<9 , T, D>, linearRowFilter_caller<2 , T, D, BrdRowReflect101>,
linearRowFilter_caller<10, T, D>, linearRowFilter_caller<11, T, D>, linearRowFilter_caller<3 , T, D, BrdRowReflect101>,
linearRowFilter_caller<12, T, D>, linearRowFilter_caller<13, T, D>, linearRowFilter_caller<4 , T, D, BrdRowReflect101>,
linearRowFilter_caller<14, T, D>, linearRowFilter_caller<15, T, D>}; 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>,
}
};
loadLinearKernel(kernel, ksize); loadLinearKernel(kernel, ksize);
callers[ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor); callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor);
} }
template void linearRowFilter_gpu<uchar4, uchar4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearRowFilter_gpu<uchar , float >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearRowFilter_gpu<uchar4, char4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearRowFilter_gpu<uchar4, float4>(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearRowFilter_gpu<char4, uchar4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearRowFilter_gpu<short , float >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);;
template void linearRowFilter_gpu<char4, char4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearRowFilter_gpu<short2, float2>(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearRowFilter_gpu<int , float >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearRowFilter_gpu<ushort2, ushort2>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearRowFilter_gpu<float , float >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearRowFilter_gpu<ushort2, short2>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearRowFilter_gpu<short2, ushort2>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearRowFilter_gpu<short2, short2>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearRowFilter_gpu<int, int>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearRowFilter_gpu<int, float>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearRowFilter_gpu<float, int>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearRowFilter_gpu<float, float>(const DevMem2D&, const DevMem2D&, const float[], int , int);
}}} }}}
namespace filter_krnls namespace filter_krnls
{ {
template <int BLOCK_DIM_X, int BLOCK_DIM_Y, int KERNEL_SIZE, typename T, typename D> template <int ksize, typename T, typename D, typename B>
__global__ void linearColumnFilter(const T* src, size_t src_step, D* dst, size_t dst_step, int anchor, int width, int height) __global__ void linearColumnFilter(const DevMem2D_<T> src, PtrStep_<D> dst, int anchor, const B b)
{ {
__shared__ T smem[BLOCK_DIM_Y * BLOCK_DIM_X * 3]; __shared__ T smem[BLOCK_DIM_Y * BLOCK_DIM_X * 3];
const int blockStartX = blockDim.x * blockIdx.x; const int x = BLOCK_DIM_X * blockIdx.x + threadIdx.x;
const int blockStartY = blockDim.y * blockIdx.y; const int y = BLOCK_DIM_Y * blockIdx.y + threadIdx.y;
const int threadX = blockStartX + threadIdx.x;
const int threadY = blockStartY + threadIdx.y;
const int prevThreadY = threadY - blockDim.y;
const int nextThreadY = threadY + blockDim.y;
const int smem_step = blockDim.x;
T* sDataColumn = smem + threadIdx.x; T* sDataColumn = smem + threadIdx.x;
if (threadX < width) if (x < src.cols)
{ {
const T* colSrc = src + threadX; const T* srcCol = src.ptr() + x;
sDataColumn[(threadIdx.y + blockDim.y) * smem_step] = threadY < height ? colSrc[threadY * src_step] : VecTraits<T>::all(0);
sDataColumn[threadIdx.y * smem_step] = prevThreadY >= 0 ? colSrc[prevThreadY * src_step] : VecTraits<T>::all(0); sDataColumn[ threadIdx.y * BLOCK_DIM_X] = b.at_low(y - BLOCK_DIM_Y, srcCol);
sDataColumn[(threadIdx.y + BLOCK_DIM_Y) * BLOCK_DIM_X] = b.at_high(y, srcCol);
sDataColumn[(threadIdx.y + (blockDim.y << 1)) * smem_step] = nextThreadY < height ? colSrc[nextThreadY * src_step] : VecTraits<T>::all(0); sDataColumn[(threadIdx.y + BLOCK_DIM_Y * 2) * BLOCK_DIM_X] = b.at_high(y + BLOCK_DIM_Y, srcCol);
__syncthreads(); __syncthreads();
if (threadY < height) if (y < src.rows)
{ {
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_t sum_t; typedef typename TypeVec<float, VecTraits<T>::cn>::vec_t sum_t;
sum_t sum = VecTraits<sum_t>::all(0); sum_t sum = VecTraits<sum_t>::all(0);
sDataColumn += (threadIdx.y + blockDim.y - anchor)* smem_step; sDataColumn += (threadIdx.y + BLOCK_DIM_Y - anchor) * BLOCK_DIM_X;
#pragma unroll #pragma unroll
for(int i = 0; i < KERNEL_SIZE; ++i) for(int i = 0; i < ksize; ++i)
sum = sum + sDataColumn[i * smem_step] * cLinearKernel[i]; sum = sum + sDataColumn[i * BLOCK_DIM_X] * cLinearKernel[i];
dst[threadY * dst_step + threadX] = saturate_cast<D>(sum); dst.ptr(y)[x] = saturate_cast<D>(sum);
} }
} }
} }
@ -222,54 +452,101 @@ namespace filter_krnls
namespace cv { namespace gpu { namespace filters namespace cv { namespace gpu { namespace filters
{ {
template <int KERNEL_SIZE, typename T, typename D> template <int ksize, typename T, typename D, template<typename> class B>
void linearColumnFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor) void linearColumnFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor)
{ {
const int BLOCK_DIM_X = 16;
const int BLOCK_DIM_Y = 16;
dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y); dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y);
dim3 blocks(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y)); dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y));
filter_krnls::linearColumnFilter<BLOCK_DIM_X, BLOCK_DIM_Y, KERNEL_SIZE><<<blocks, threads>>>(src.data, src.step/src.elemSize(), B<T> b(src.rows, src.step / src.elemSize());
dst.data, dst.step/dst.elemSize(), anchor, src.cols, src.rows);
if (!b.is_range_safe(-BLOCK_DIM_Y, (grid.y + 1) * BLOCK_DIM_Y - 1))
{
cv::gpu::error("linearColumnFilter: can't use specified border extrapolation, image is too small, "
"try bigger image or another border extrapolation mode", __FILE__, __LINE__);
}
filter_krnls::linearColumnFilter<ksize, T, D><<<grid, threads>>>(src, dst, anchor, b);
cudaSafeCall( cudaThreadSynchronize() ); cudaSafeCall( cudaThreadSynchronize() );
} }
template <typename T, typename D> template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor) void linearColumnFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type)
{ {
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor); typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor);
static const caller_t callers[] = static const caller_t callers[3][17] =
{linearColumnFilter_caller<0 , T, D>, linearColumnFilter_caller<1 , T, D>, {
linearColumnFilter_caller<2 , T, D>, linearColumnFilter_caller<3 , T, D>, {
linearColumnFilter_caller<4 , T, D>, linearColumnFilter_caller<5 , T, D>, 0,
linearColumnFilter_caller<6 , T, D>, linearColumnFilter_caller<7 , T, D>, linearColumnFilter_caller<1 , T, D, BrdColReflect101>,
linearColumnFilter_caller<8 , T, D>, linearColumnFilter_caller<9 , T, D>, linearColumnFilter_caller<2 , T, D, BrdColReflect101>,
linearColumnFilter_caller<10, T, D>, linearColumnFilter_caller<11, T, D>, linearColumnFilter_caller<3 , T, D, BrdColReflect101>,
linearColumnFilter_caller<12, T, D>, linearColumnFilter_caller<13, T, D>, linearColumnFilter_caller<4 , T, D, BrdColReflect101>,
linearColumnFilter_caller<14, T, D>, linearColumnFilter_caller<15, T, D>}; 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>,
}
};
loadLinearKernel(kernel, ksize); loadLinearKernel(kernel, ksize);
callers[ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor); callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor);
} }
template void linearColumnFilter_gpu<uchar4, uchar4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearColumnFilter_gpu<float , uchar >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearColumnFilter_gpu<uchar4, char4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearColumnFilter_gpu<float4, uchar4>(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearColumnFilter_gpu<char4, uchar4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearColumnFilter_gpu<float , short >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearColumnFilter_gpu<char4, char4>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearColumnFilter_gpu<float2, short2>(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearColumnFilter_gpu<float , int >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearColumnFilter_gpu<ushort2, ushort2>(const DevMem2D&, const DevMem2D&, const float[], int , int); template void linearColumnFilter_gpu<float , float >(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template void linearColumnFilter_gpu<ushort2, short2>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearColumnFilter_gpu<short2, ushort2>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearColumnFilter_gpu<short2, short2>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearColumnFilter_gpu<int, int>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearColumnFilter_gpu<int, float>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearColumnFilter_gpu<float, int>(const DevMem2D&, const DevMem2D&, const float[], int , int);
template void linearColumnFilter_gpu<float, float>(const DevMem2D&, const DevMem2D&, const float[], int , int);
}}} }}}
///////////////////////////////////////////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////////////////////////////////////////
@ -377,7 +654,7 @@ namespace bf_krnls
} }
} }
float minimum = FLT_MAX; float minimum = numeric_limits_gpu<float>::max();
int id = 0; int id = 0;
if (cost[0] < minimum) if (cost[0] < minimum)

3
modules/gpu/src/cuda/internal_shared.hpp
Unescape View File

@ -59,7 +59,8 @@ namespace cv
enum enum
{ {
BORDER_REFLECT101_GPU = 0, BORDER_REFLECT101_GPU = 0,
BORDER_REPLICATE_GPU BORDER_REPLICATE_GPU,
BORDER_CONSTANT_GPU
}; };
// Converts CPU border extrapolation mode into GPU internal analogue. // Converts CPU border extrapolation mode into GPU internal analogue.

240
modules/gpu/src/filtering.cpp
Unescape View File

@ -48,8 +48,8 @@ using namespace cv::gpu;
#if !defined (HAVE_CUDA) #if !defined (HAVE_CUDA)
Ptr<FilterEngine_GPU> cv::gpu::createFilter2D_GPU(const Ptr<BaseFilter_GPU>) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); } Ptr<FilterEngine_GPU> cv::gpu::createFilter2D_GPU(const Ptr<BaseFilter_GPU>, int, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); }
Ptr<FilterEngine_GPU> cv::gpu::createSeparableFilter_GPU(const Ptr<BaseRowFilter_GPU>&, const Ptr<BaseColumnFilter_GPU>&) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); } Ptr<FilterEngine_GPU> cv::gpu::createSeparableFilter_GPU(const Ptr<BaseRowFilter_GPU>&, const Ptr<BaseColumnFilter_GPU>&, int, int, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); }
Ptr<BaseRowFilter_GPU> cv::gpu::getRowSumFilter_GPU(int, int, int, int) { throw_nogpu(); return Ptr<BaseRowFilter_GPU>(0); } Ptr<BaseRowFilter_GPU> cv::gpu::getRowSumFilter_GPU(int, int, int, int) { throw_nogpu(); return Ptr<BaseRowFilter_GPU>(0); }
Ptr<BaseColumnFilter_GPU> cv::gpu::getColumnSumFilter_GPU(int, int, int, int) { throw_nogpu(); return Ptr<BaseColumnFilter_GPU>(0); } Ptr<BaseColumnFilter_GPU> cv::gpu::getColumnSumFilter_GPU(int, int, int, int) { throw_nogpu(); return Ptr<BaseColumnFilter_GPU>(0); }
Ptr<BaseFilter_GPU> cv::gpu::getBoxFilter_GPU(int, int, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); } Ptr<BaseFilter_GPU> cv::gpu::getBoxFilter_GPU(int, int, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); }
@ -58,11 +58,11 @@ Ptr<BaseFilter_GPU> cv::gpu::getMorphologyFilter_GPU(int, int, const Mat&, const
Ptr<FilterEngine_GPU> cv::gpu::createMorphologyFilter_GPU(int, int, const Mat&, const Point&, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); } Ptr<FilterEngine_GPU> cv::gpu::createMorphologyFilter_GPU(int, int, const Mat&, const Point&, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); }
Ptr<BaseFilter_GPU> cv::gpu::getLinearFilter_GPU(int, int, const Mat&, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); } Ptr<BaseFilter_GPU> cv::gpu::getLinearFilter_GPU(int, int, const Mat&, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); }
Ptr<FilterEngine_GPU> cv::gpu::createLinearFilter_GPU(int, int, const Mat&, const Point&) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); } Ptr<FilterEngine_GPU> cv::gpu::createLinearFilter_GPU(int, int, const Mat&, const Point&) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); }
Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int, int, const Mat&, int) { throw_nogpu(); return Ptr<BaseRowFilter_GPU>(0); } Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int, int, const Mat&, int, int) { throw_nogpu(); return Ptr<BaseRowFilter_GPU>(0); }
Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int, int, const Mat&, int) { throw_nogpu(); return Ptr<BaseColumnFilter_GPU>(0); } Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int, int, const Mat&, int, int) { throw_nogpu(); return Ptr<BaseColumnFilter_GPU>(0); }
Ptr<FilterEngine_GPU> cv::gpu::createSeparableLinearFilter_GPU(int, int, const Mat&, const Mat&, const Point&) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); } Ptr<FilterEngine_GPU> cv::gpu::createSeparableLinearFilter_GPU(int, int, const Mat&, const Mat&, const Point&, int, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); }
Ptr<FilterEngine_GPU> cv::gpu::createDerivFilter_GPU(int, int, int, int, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); } Ptr<FilterEngine_GPU> cv::gpu::createDerivFilter_GPU(int, int, int, int, int, int, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); }
Ptr<FilterEngine_GPU> cv::gpu::createGaussianFilter_GPU(int, Size, double, double) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); } Ptr<FilterEngine_GPU> cv::gpu::createGaussianFilter_GPU(int, Size, double, double, int, int) { throw_nogpu(); return Ptr<FilterEngine_GPU>(0); }
Ptr<BaseFilter_GPU> cv::gpu::getMaxFilter_GPU(int, int, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); } Ptr<BaseFilter_GPU> cv::gpu::getMaxFilter_GPU(int, int, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); }
Ptr<BaseFilter_GPU> cv::gpu::getMinFilter_GPU(int, int, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); } Ptr<BaseFilter_GPU> cv::gpu::getMinFilter_GPU(int, int, const Size&, Point) { throw_nogpu(); return Ptr<BaseFilter_GPU>(0); }
@ -71,10 +71,10 @@ void cv::gpu::erode( const GpuMat&, GpuMat&, const Mat&, Point, int) { throw_nog
void cv::gpu::dilate( const GpuMat&, GpuMat&, const Mat&, Point, int) { throw_nogpu(); } void cv::gpu::dilate( const GpuMat&, GpuMat&, const Mat&, Point, int) { throw_nogpu(); }
void cv::gpu::morphologyEx( const GpuMat&, GpuMat&, int, const Mat&, Point, int) { throw_nogpu(); } void cv::gpu::morphologyEx( const GpuMat&, GpuMat&, int, const Mat&, Point, int) { throw_nogpu(); }
void cv::gpu::filter2D(const GpuMat&, GpuMat&, int, const Mat&, Point) { throw_nogpu(); } void cv::gpu::filter2D(const GpuMat&, GpuMat&, int, const Mat&, Point) { throw_nogpu(); }
void cv::gpu::sepFilter2D(const GpuMat&, GpuMat&, int, const Mat&, const Mat&, Point) { throw_nogpu(); } void cv::gpu::sepFilter2D(const GpuMat&, GpuMat&, int, const Mat&, const Mat&, Point, int, int) { throw_nogpu(); }
void cv::gpu::Sobel(const GpuMat&, GpuMat&, int, int, int, int, double) { throw_nogpu(); } void cv::gpu::Sobel(const GpuMat&, GpuMat&, int, int, int, int, double, int, int) { throw_nogpu(); }
void cv::gpu::Scharr(const GpuMat&, GpuMat&, int, int, int, double) { throw_nogpu(); } void cv::gpu::Scharr(const GpuMat&, GpuMat&, int, int, int, double, int, int) { throw_nogpu(); }
void cv::gpu::GaussianBlur(const GpuMat&, GpuMat&, Size, double, double) { throw_nogpu(); } void cv::gpu::GaussianBlur(const GpuMat&, GpuMat&, Size, double, double, int, int) { throw_nogpu(); }
void cv::gpu::Laplacian(const GpuMat&, GpuMat&, int, int, double) { throw_nogpu(); } void cv::gpu::Laplacian(const GpuMat&, GpuMat&, int, int, double) { throw_nogpu(); }
#else #else
@ -133,13 +133,17 @@ namespace
class Filter2DEngine_GPU : public FilterEngine_GPU class Filter2DEngine_GPU : public FilterEngine_GPU
{ {
public: public:
Filter2DEngine_GPU(const Ptr<BaseFilter_GPU>& filter2D_) : filter2D(filter2D_) {} Filter2DEngine_GPU(const Ptr<BaseFilter_GPU>& filter2D_, int srcType_, int dstType_) :
filter2D(filter2D_), srcType(srcType_), dstType(dstType_)
{}
virtual void apply(const GpuMat& src, GpuMat& dst, Rect roi = Rect(0,0,-1,-1)) virtual void apply(const GpuMat& src, GpuMat& dst, Rect roi = Rect(0,0,-1,-1))
{ {
CV_Assert(src.type() == srcType);
Size src_size = src.size(); Size src_size = src.size();
dst.create(src_size, src.type()); dst.create(src_size, dstType);
dst = Scalar(0.0); dst = Scalar(0.0);
normalizeROI(roi, filter2D->ksize, filter2D->anchor, src_size); normalizeROI(roi, filter2D->ksize, filter2D->anchor, src_size);
@ -151,12 +155,13 @@ namespace
} }
Ptr<BaseFilter_GPU> filter2D; Ptr<BaseFilter_GPU> filter2D;
int srcType, dstType;
}; };
} }
Ptr<FilterEngine_GPU> cv::gpu::createFilter2D_GPU(const Ptr<BaseFilter_GPU> filter2D) Ptr<FilterEngine_GPU> cv::gpu::createFilter2D_GPU(const Ptr<BaseFilter_GPU> filter2D, int srcType, int dstType)
{ {
return Ptr<FilterEngine_GPU>(new Filter2DEngine_GPU(filter2D)); return Ptr<FilterEngine_GPU>(new Filter2DEngine_GPU(filter2D, srcType, dstType));
} }
//////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////
@ -168,8 +173,9 @@ namespace
{ {
public: public:
SeparableFilterEngine_GPU(const Ptr<BaseRowFilter_GPU>& rowFilter_, SeparableFilterEngine_GPU(const Ptr<BaseRowFilter_GPU>& rowFilter_,
const Ptr<BaseColumnFilter_GPU>& columnFilter_) : const Ptr<BaseColumnFilter_GPU>& columnFilter_, int srcType_, int bufType_, int dstType_) :
rowFilter(rowFilter_), columnFilter(columnFilter_) rowFilter(rowFilter_), columnFilter(columnFilter_),
srcType(srcType_), bufType(bufType_), dstType(dstType_)
{ {
ksize = Size(rowFilter->ksize, columnFilter->ksize); ksize = Size(rowFilter->ksize, columnFilter->ksize);
anchor = Point(rowFilter->anchor, columnFilter->anchor); anchor = Point(rowFilter->anchor, columnFilter->anchor);
@ -177,19 +183,20 @@ namespace
virtual void apply(const GpuMat& src, GpuMat& dst, Rect roi = Rect(0,0,-1,-1)) virtual void apply(const GpuMat& src, GpuMat& dst, Rect roi = Rect(0,0,-1,-1))
{ {
CV_Assert(src.type() == srcType);
Size src_size = src.size(); Size src_size = src.size();
int src_type = src.type();
dst.create(src_size, src_type); dst.create(src_size, dstType);
dst = Scalar(0.0); dst = Scalar(0.0);
dstBuf.create(src_size, src_type); dstBuf.create(src_size, bufType);
dstBuf = Scalar(0.0); dstBuf = Scalar(0.0);
normalizeROI(roi, ksize, anchor, src_size); normalizeROI(roi, ksize, anchor, src_size);
srcROI = src(roi); GpuMat srcROI = src(roi);
dstROI = dst(roi); GpuMat dstROI = dst(roi);
dstBufROI = dstBuf(roi); GpuMat dstBufROI = dstBuf(roi);
(*rowFilter)(srcROI, dstBufROI); (*rowFilter)(srcROI, dstBufROI);
(*columnFilter)(dstBufROI, dstROI); (*columnFilter)(dstBufROI, dstROI);
@ -197,19 +204,19 @@ namespace
Ptr<BaseRowFilter_GPU> rowFilter; Ptr<BaseRowFilter_GPU> rowFilter;
Ptr<BaseColumnFilter_GPU> columnFilter; Ptr<BaseColumnFilter_GPU> columnFilter;
int srcType, bufType, dstType;
Size ksize; Size ksize;
Point anchor; Point anchor;
GpuMat dstBuf; GpuMat dstBuf;
GpuMat srcROI;
GpuMat dstROI;
GpuMat dstBufROI;
}; };
} }
Ptr<FilterEngine_GPU> cv::gpu::createSeparableFilter_GPU(const Ptr<BaseRowFilter_GPU>& rowFilter, Ptr<FilterEngine_GPU> cv::gpu::createSeparableFilter_GPU(const Ptr<BaseRowFilter_GPU>& rowFilter,
const Ptr<BaseColumnFilter_GPU>& columnFilter) const Ptr<BaseColumnFilter_GPU>& columnFilter, int srcType, int bufType, int dstType)
{ {
return Ptr<FilterEngine_GPU>(new SeparableFilterEngine_GPU(rowFilter, columnFilter)); return Ptr<FilterEngine_GPU>(new SeparableFilterEngine_GPU(rowFilter, columnFilter, srcType, bufType, dstType));
} }
//////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////
@ -315,7 +322,7 @@ Ptr<BaseFilter_GPU> cv::gpu::getBoxFilter_GPU(int srcType, int dstType, const Si
Ptr<FilterEngine_GPU> cv::gpu::createBoxFilter_GPU(int srcType, int dstType, const Size& ksize, const Point& anchor) Ptr<FilterEngine_GPU> cv::gpu::createBoxFilter_GPU(int srcType, int dstType, const Size& ksize, const Point& anchor)
{ {
Ptr<BaseFilter_GPU> boxFilter = getBoxFilter_GPU(srcType, dstType, ksize, anchor); Ptr<BaseFilter_GPU> boxFilter = getBoxFilter_GPU(srcType, dstType, ksize, anchor);
return createFilter2D_GPU(boxFilter); return createFilter2D_GPU(boxFilter, srcType, dstType);
} }
void cv::gpu::boxFilter(const GpuMat& src, GpuMat& dst, int ddepth, Size ksize, Point anchor) void cv::gpu::boxFilter(const GpuMat& src, GpuMat& dst, int ddepth, Size ksize, Point anchor)
@ -386,8 +393,8 @@ namespace
class MorphologyFilterEngine_GPU : public Filter2DEngine_GPU class MorphologyFilterEngine_GPU : public Filter2DEngine_GPU
{ {
public: public:
MorphologyFilterEngine_GPU(const Ptr<BaseFilter_GPU>& filter2D_, int iters_) : MorphologyFilterEngine_GPU(const Ptr<BaseFilter_GPU>& filter2D_, int type, int iters_) :
Filter2DEngine_GPU(filter2D_), iters(iters_) {} Filter2DEngine_GPU(filter2D_, type, type), iters(iters_) {}
virtual void apply(const GpuMat& src, GpuMat& dst, Rect roi = Rect(0,0,-1,-1)) virtual void apply(const GpuMat& src, GpuMat& dst, Rect roi = Rect(0,0,-1,-1))
{ {
@ -415,7 +422,7 @@ Ptr<FilterEngine_GPU> cv::gpu::createMorphologyFilter_GPU(int op, int type, cons
Ptr<BaseFilter_GPU> filter2D = getMorphologyFilter_GPU(op, type, kernel, ksize, anchor); Ptr<BaseFilter_GPU> filter2D = getMorphologyFilter_GPU(op, type, kernel, ksize, anchor);
return Ptr<FilterEngine_GPU>(new MorphologyFilterEngine_GPU(filter2D, iterations)); return Ptr<FilterEngine_GPU>(new MorphologyFilterEngine_GPU(filter2D, type, iterations));
} }
namespace namespace
@ -558,7 +565,7 @@ Ptr<FilterEngine_GPU> cv::gpu::createLinearFilter_GPU(int srcType, int dstType,
Ptr<BaseFilter_GPU> linearFilter = getLinearFilter_GPU(srcType, dstType, kernel, ksize, anchor); Ptr<BaseFilter_GPU> linearFilter = getLinearFilter_GPU(srcType, dstType, kernel, ksize, anchor);
return createFilter2D_GPU(linearFilter); return createFilter2D_GPU(linearFilter, srcType, dstType);
} }
void cv::gpu::filter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& kernel, Point anchor) void cv::gpu::filter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& kernel, Point anchor)
@ -578,10 +585,10 @@ void cv::gpu::filter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& ke
namespace cv { namespace gpu { namespace filters namespace cv { namespace gpu { namespace filters
{ {
template <typename T, typename D> template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor); void linearRowFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
template <typename T, typename D> template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor); void linearColumnFilter_gpu(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
}}} }}}
namespace namespace
@ -589,7 +596,7 @@ namespace
typedef NppStatus (*nppFilter1D_t)(const Npp8u * pSrc, Npp32s nSrcStep, Npp8u * pDst, Npp32s nDstStep, NppiSize oROI, typedef NppStatus (*nppFilter1D_t)(const Npp8u * pSrc, Npp32s nSrcStep, Npp8u * pDst, Npp32s nDstStep, NppiSize oROI,
const Npp32s * pKernel, Npp32s nMaskSize, Npp32s nAnchor, Npp32s nDivisor); const Npp32s * pKernel, Npp32s nMaskSize, Npp32s nAnchor, Npp32s nDivisor);
typedef void (*gpuFilter1D_t)(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor); typedef void (*gpuFilter1D_t)(const DevMem2D& src, const DevMem2D& dst, const float kernel[], int ksize, int anchor, int brd_type);
class NppLinearRowFilter : public BaseRowFilter_GPU class NppLinearRowFilter : public BaseRowFilter_GPU
{ {
@ -614,35 +621,28 @@ namespace
class GpuLinearRowFilter : public BaseRowFilter_GPU class GpuLinearRowFilter : public BaseRowFilter_GPU
{ {
public: public:
GpuLinearRowFilter(int ksize_, int anchor_, const Mat& kernel_, gpuFilter1D_t func_) : GpuLinearRowFilter(int ksize_, int anchor_, const Mat& kernel_, gpuFilter1D_t func_, int brd_type_) :
BaseRowFilter_GPU(ksize_, anchor_), kernel(kernel_), func(func_) {} BaseRowFilter_GPU(ksize_, anchor_), kernel(kernel_), func(func_), brd_type(brd_type_) {}
virtual void operator()(const GpuMat& src, GpuMat& dst) virtual void operator()(const GpuMat& src, GpuMat& dst)
{ {
func(src, dst, kernel.ptr<float>(), ksize, anchor); func(src, dst, kernel.ptr<float>(), ksize, anchor, brd_type);
} }
Mat kernel; Mat kernel;
gpuFilter1D_t func; gpuFilter1D_t func;
int brd_type;
}; };
} }
Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType, const Mat& rowKernel, int anchor) Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType, const Mat& rowKernel, int anchor, int borderType)
{ {
using namespace cv::gpu::filters;
static const nppFilter1D_t nppFilter1D_callers[] = {0, nppiFilterRow_8u_C1R, 0, 0, nppiFilterRow_8u_C4R}; static const nppFilter1D_t nppFilter1D_callers[] = {0, nppiFilterRow_8u_C1R, 0, 0, nppiFilterRow_8u_C4R};
static const gpuFilter1D_t gpuFilter1D_callers[6][6] =
{
{linearRowFilter_gpu<uchar4, uchar4>,linearRowFilter_gpu<uchar4, char4>,0,0,0,0},
{linearRowFilter_gpu<char4, uchar4>,linearRowFilter_gpu<char4, char4>,0,0,0,0},
{0,0,linearRowFilter_gpu<ushort2, ushort2>,linearRowFilter_gpu<ushort2, short2>,0,0},
{0,0,linearRowFilter_gpu<short2, ushort2>,linearRowFilter_gpu<short2, short2>,0,0},
{0,0,0,0,linearRowFilter_gpu<int, int>, linearRowFilter_gpu<int, float>},
{0,0,0,0,linearRowFilter_gpu<float, int>, linearRowFilter_gpu<float, float>}
};
if ((bufType == srcType) && (srcType == CV_8UC1 || srcType == CV_8UC4)) if ((bufType == srcType) && (srcType == CV_8UC1 || srcType == CV_8UC4))
{ {
CV_Assert(borderType == BORDER_CONSTANT);
GpuMat gpu_row_krnl; GpuMat gpu_row_krnl;
int nDivisor; int nDivisor;
normalizeKernel(rowKernel, gpu_row_krnl, CV_32S, &nDivisor, true); normalizeKernel(rowKernel, gpu_row_krnl, CV_32S, &nDivisor, true);
@ -654,8 +654,14 @@ Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType,
nppFilter1D_callers[CV_MAT_CN(srcType)])); nppFilter1D_callers[CV_MAT_CN(srcType)]));
} }
CV_Assert(srcType == CV_8UC4 || srcType == CV_8SC4 || srcType == CV_16UC2 || srcType == CV_16SC2 || srcType == CV_32SC1 || srcType == CV_32FC1); CV_Assert(borderType == BORDER_REFLECT101 || borderType == BORDER_REPLICATE || borderType == BORDER_CONSTANT);
CV_Assert(bufType == CV_8UC4 || bufType == CV_8SC4 || bufType == CV_16UC2 || bufType == CV_16SC2 || bufType == CV_32SC1 || bufType == CV_32FC1); int gpuBorderType;
CV_Assert(tryConvertToGpuBorderType(borderType, gpuBorderType));
CV_Assert(srcType == CV_8UC1 || srcType == CV_8UC4 || srcType == CV_16SC1 || srcType == CV_16SC2
|| srcType == CV_32SC1 || srcType == CV_32FC1);
CV_Assert(CV_MAT_DEPTH(bufType) == CV_32F && CV_MAT_CN(srcType) == CV_MAT_CN(bufType));
Mat temp(rowKernel.size(), CV_32FC1); Mat temp(rowKernel.size(), CV_32FC1);
rowKernel.convertTo(temp, CV_32FC1); rowKernel.convertTo(temp, CV_32FC1);
@ -663,12 +669,35 @@ Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType,
int ksize = cont_krnl.cols; int ksize = cont_krnl.cols;
CV_Assert(ksize < 16); CV_Assert(ksize > 0 && ksize <= 16);
normalizeAnchor(anchor, ksize); normalizeAnchor(anchor, ksize);
return Ptr<BaseRowFilter_GPU>(new GpuLinearRowFilter(ksize, anchor, cont_krnl, gpuFilter1D_t func = 0;
gpuFilter1D_callers[CV_MAT_DEPTH(srcType)][CV_MAT_DEPTH(bufType)]));
switch (srcType)
{
case CV_8UC1:
func = filters::linearRowFilter_gpu<uchar, float>;
break;
case CV_8UC4:
func = filters::linearRowFilter_gpu<uchar4, float4>;
break;
case CV_16SC1:
func = filters::linearRowFilter_gpu<short, float>;
break;
case CV_16SC2:
func = filters::linearRowFilter_gpu<short2, float2>;
break;
case CV_32SC1:
func = filters::linearRowFilter_gpu<int, float>;
break;
case CV_32FC1:
func = filters::linearRowFilter_gpu<float, float>;
break;
}
return Ptr<BaseRowFilter_GPU>(new GpuLinearRowFilter(ksize, anchor, cont_krnl, func, gpuBorderType));
} }
namespace namespace
@ -696,35 +725,28 @@ namespace
class GpuLinearColumnFilter : public BaseColumnFilter_GPU class GpuLinearColumnFilter : public BaseColumnFilter_GPU
{ {
public: public:
GpuLinearColumnFilter(int ksize_, int anchor_, const Mat& kernel_, gpuFilter1D_t func_) : GpuLinearColumnFilter(int ksize_, int anchor_, const Mat& kernel_, gpuFilter1D_t func_, int brd_type_) :
BaseColumnFilter_GPU(ksize_, anchor_), kernel(kernel_), func(func_) {} BaseColumnFilter_GPU(ksize_, anchor_), kernel(kernel_), func(func_), brd_type(brd_type_) {}
virtual void operator()(const GpuMat& src, GpuMat& dst) virtual void operator()(const GpuMat& src, GpuMat& dst)
{ {
func(src, dst, kernel.ptr<float>(), ksize, anchor); func(src, dst, kernel.ptr<float>(), ksize, anchor, brd_type);
} }
Mat kernel; Mat kernel;
gpuFilter1D_t func; gpuFilter1D_t func;
int brd_type;
}; };
} }
Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int dstType, const Mat& columnKernel, int anchor) Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int dstType, const Mat& columnKernel, int anchor, int borderType)
{ {
using namespace cv::gpu::filters;
static const nppFilter1D_t nppFilter1D_callers[] = {0, nppiFilterColumn_8u_C1R, 0, 0, nppiFilterColumn_8u_C4R}; static const nppFilter1D_t nppFilter1D_callers[] = {0, nppiFilterColumn_8u_C1R, 0, 0, nppiFilterColumn_8u_C4R};
static const gpuFilter1D_t gpuFilter1D_callers[6][6] =
{
{linearColumnFilter_gpu<uchar4, uchar4>,linearColumnFilter_gpu<uchar4, char4>,0,0,0,0},
{linearColumnFilter_gpu<char4, uchar4>,linearColumnFilter_gpu<char4, char4>,0,0,0,0},
{0,0,linearColumnFilter_gpu<ushort2, ushort2>,linearColumnFilter_gpu<ushort2, short2>,0,0},
{0,0,linearColumnFilter_gpu<short2, ushort2>,linearColumnFilter_gpu<short2, short2>,0,0},
{0,0,0,0,linearColumnFilter_gpu<int, int>, linearColumnFilter_gpu<int, float>},
{0,0,0,0,linearColumnFilter_gpu<float, int>, linearColumnFilter_gpu<float, float>}
};
if ((bufType == dstType) && (bufType == CV_8UC1 || bufType == CV_8UC4)) if ((bufType == dstType) && (bufType == CV_8UC1 || bufType == CV_8UC4))
{ {
CV_Assert(borderType == BORDER_CONSTANT);
GpuMat gpu_col_krnl; GpuMat gpu_col_krnl;
int nDivisor; int nDivisor;
normalizeKernel(columnKernel, gpu_col_krnl, CV_32S, &nDivisor, true); normalizeKernel(columnKernel, gpu_col_krnl, CV_32S, &nDivisor, true);
@ -736,8 +758,14 @@ Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int ds
nppFilter1D_callers[CV_MAT_CN(bufType)])); nppFilter1D_callers[CV_MAT_CN(bufType)]));
} }
CV_Assert(dstType == CV_8UC4 || dstType == CV_8SC4 || dstType == CV_16UC2 || dstType == CV_16SC2 || dstType == CV_32SC1 || dstType == CV_32FC1); CV_Assert(borderType == BORDER_REFLECT101 || borderType == BORDER_REPLICATE || borderType == BORDER_CONSTANT);
CV_Assert(bufType == CV_8UC4 || bufType == CV_8SC4 || bufType == CV_16UC2 || bufType == CV_16SC2 || bufType == CV_32SC1 || bufType == CV_32FC1); int gpuBorderType;
CV_Assert(tryConvertToGpuBorderType(borderType, gpuBorderType));
CV_Assert(dstType == CV_8UC1 || dstType == CV_8UC4 || dstType == CV_16SC1 || dstType == CV_16SC2
|| dstType == CV_32SC1 || dstType == CV_32FC1);
CV_Assert(CV_MAT_DEPTH(bufType) == CV_32F && CV_MAT_CN(dstType) == CV_MAT_CN(bufType));
Mat temp(columnKernel.size(), CV_32FC1); Mat temp(columnKernel.size(), CV_32FC1);
columnKernel.convertTo(temp, CV_32FC1); columnKernel.convertTo(temp, CV_32FC1);
@ -745,50 +773,76 @@ Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int ds
int ksize = cont_krnl.cols; int ksize = cont_krnl.cols;
CV_Assert(ksize < 16); CV_Assert(ksize > 0 && ksize <= 16);
normalizeAnchor(anchor, ksize); normalizeAnchor(anchor, ksize);
return Ptr<BaseColumnFilter_GPU>(new GpuLinearColumnFilter(ksize, anchor, cont_krnl, gpuFilter1D_t func = 0;
gpuFilter1D_callers[CV_MAT_DEPTH(bufType)][CV_MAT_DEPTH(dstType)]));
switch (dstType)
{
case CV_8UC1:
func = filters::linearColumnFilter_gpu<float, uchar>;
break;
case CV_8UC4:
func = filters::linearColumnFilter_gpu<float4, uchar4>;
break;
case CV_16SC1:
func = filters::linearColumnFilter_gpu<float, short>;
break;
case CV_16SC2:
func = filters::linearColumnFilter_gpu<float2, short2>;
break;
case CV_32SC1:
func = filters::linearColumnFilter_gpu<float, int>;
break;
case CV_32FC1:
func = filters::linearColumnFilter_gpu<float, float>;
break;
}
return Ptr<BaseColumnFilter_GPU>(new GpuLinearColumnFilter(ksize, anchor, cont_krnl, func, gpuBorderType));
} }
Ptr<FilterEngine_GPU> cv::gpu::createSeparableLinearFilter_GPU(int srcType, int dstType, const Mat& rowKernel, const Mat& columnKernel, Ptr<FilterEngine_GPU> cv::gpu::createSeparableLinearFilter_GPU(int srcType, int dstType, const Mat& rowKernel, const Mat& columnKernel,
const Point& anchor) const Point& anchor, int rowBorderType, int columnBorderType)
{ {
if (columnBorderType < 0)
columnBorderType = rowBorderType;
int sdepth = CV_MAT_DEPTH(srcType), ddepth = CV_MAT_DEPTH(dstType); int sdepth = CV_MAT_DEPTH(srcType), ddepth = CV_MAT_DEPTH(dstType);
int cn = CV_MAT_CN(srcType); int cn = CV_MAT_CN(srcType);
int bdepth = std::max(sdepth, ddepth); int bdepth = CV_32F;
int bufType = CV_MAKETYPE(bdepth, cn); int bufType = CV_MAKETYPE(bdepth, cn);
Ptr<BaseRowFilter_GPU> rowFilter = getLinearRowFilter_GPU(srcType, bufType, rowKernel, anchor.x); Ptr<BaseRowFilter_GPU> rowFilter = getLinearRowFilter_GPU(srcType, bufType, rowKernel, anchor.x, rowBorderType);
Ptr<BaseColumnFilter_GPU> columnFilter = getLinearColumnFilter_GPU(bufType, dstType, columnKernel, anchor.y); Ptr<BaseColumnFilter_GPU> columnFilter = getLinearColumnFilter_GPU(bufType, dstType, columnKernel, anchor.y, columnBorderType);
return createSeparableFilter_GPU(rowFilter, columnFilter); return createSeparableFilter_GPU(rowFilter, columnFilter, srcType, bufType, dstType);
} }
void cv::gpu::sepFilter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& kernelX, const Mat& kernelY, Point anchor) void cv::gpu::sepFilter2D(const GpuMat& src, GpuMat& dst, int ddepth, const Mat& kernelX, const Mat& kernelY, Point anchor, int rowBorderType, int columnBorderType)
{ {
if( ddepth < 0 ) if( ddepth < 0 )
ddepth = src.depth(); ddepth = src.depth();
dst.create(src.size(), CV_MAKETYPE(ddepth, src.channels())); dst.create(src.size(), CV_MAKETYPE(ddepth, src.channels()));
Ptr<FilterEngine_GPU> f = createSeparableLinearFilter_GPU(src.type(), dst.type(), kernelX, kernelY, anchor); Ptr<FilterEngine_GPU> f = createSeparableLinearFilter_GPU(src.type(), dst.type(), kernelX, kernelY, anchor, rowBorderType, columnBorderType);
f->apply(src, dst); f->apply(src, dst, Rect(0, 0, src.cols, src.rows));
} }
//////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////
// Deriv Filter // Deriv Filter
Ptr<FilterEngine_GPU> cv::gpu::createDerivFilter_GPU(int srcType, int dstType, int dx, int dy, int ksize) Ptr<FilterEngine_GPU> cv::gpu::createDerivFilter_GPU(int srcType, int dstType, int dx, int dy, int ksize, int rowBorderType, int columnBorderType)
{ {
Mat kx, ky; Mat kx, ky;
getDerivKernels(kx, ky, dx, dy, ksize, false, CV_32F); getDerivKernels(kx, ky, dx, dy, ksize, false, CV_32F);
return createSeparableLinearFilter_GPU(srcType, dstType, kx, ky, Point(-1,-1)); return createSeparableLinearFilter_GPU(srcType, dstType, kx, ky, Point(-1,-1), rowBorderType, columnBorderType);
} }
void cv::gpu::Sobel(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, int ksize, double scale) void cv::gpu::Sobel(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, int ksize, double scale, int rowBorderType, int columnBorderType)
{ {
Mat kx, ky; Mat kx, ky;
getDerivKernels(kx, ky, dx, dy, ksize, false, CV_32F); getDerivKernels(kx, ky, dx, dy, ksize, false, CV_32F);
@ -803,10 +857,10 @@ void cv::gpu::Sobel(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy,
ky *= scale; ky *= scale;
} }
sepFilter2D(src, dst, ddepth, kx, ky, Point(-1,-1)); sepFilter2D(src, dst, ddepth, kx, ky, Point(-1,-1), rowBorderType, columnBorderType);
} }
void cv::gpu::Scharr(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, double scale) void cv::gpu::Scharr(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy, double scale, int rowBorderType, int columnBorderType)
{ {
Mat kx, ky; Mat kx, ky;
getDerivKernels(kx, ky, dx, dy, -1, false, CV_32F); getDerivKernels(kx, ky, dx, dy, -1, false, CV_32F);
@ -821,7 +875,7 @@ void cv::gpu::Scharr(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy,
ky *= scale; ky *= scale;
} }
sepFilter2D(src, dst, ddepth, kx, ky, Point(-1,-1)); sepFilter2D(src, dst, ddepth, kx, ky, Point(-1,-1), rowBorderType, columnBorderType);
} }
void cv::gpu::Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize, double scale) void cv::gpu::Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize, double scale)
@ -843,7 +897,7 @@ void cv::gpu::Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize, d
//////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////
// Gaussian Filter // Gaussian Filter
Ptr<FilterEngine_GPU> cv::gpu::createGaussianFilter_GPU(int type, Size ksize, double sigma1, double sigma2) Ptr<FilterEngine_GPU> cv::gpu::createGaussianFilter_GPU(int type, Size ksize, double sigma1, double sigma2, int rowBorderType, int columnBorderType)
{ {
int depth = CV_MAT_DEPTH(type); int depth = CV_MAT_DEPTH(type);
@ -868,10 +922,10 @@ Ptr<FilterEngine_GPU> cv::gpu::createGaussianFilter_GPU(int type, Size ksize, do
else else
ky = getGaussianKernel( ksize.height, sigma2, std::max(depth, CV_32F) ); ky = getGaussianKernel( ksize.height, sigma2, std::max(depth, CV_32F) );
return createSeparableLinearFilter_GPU(type, type, kx, ky); return createSeparableLinearFilter_GPU(type, type, kx, ky, Point(-1,-1), rowBorderType, columnBorderType);
} }
void cv::gpu::GaussianBlur(const GpuMat& src, GpuMat& dst, Size ksize, double sigma1, double sigma2) void cv::gpu::GaussianBlur(const GpuMat& src, GpuMat& dst, Size ksize, double sigma1, double sigma2, int rowBorderType, int columnBorderType)
{ {
if (ksize.width == 1 && ksize.height == 1) if (ksize.width == 1 && ksize.height == 1)
{ {
@ -881,8 +935,8 @@ void cv::gpu::GaussianBlur(const GpuMat& src, GpuMat& dst, Size ksize, double si
dst.create(src.size(), src.type()); dst.create(src.size(), src.type());
Ptr<FilterEngine_GPU> f = createGaussianFilter_GPU(src.type(), ksize, sigma1, sigma2); Ptr<FilterEngine_GPU> f = createGaussianFilter_GPU(src.type(), ksize, sigma1, sigma2, rowBorderType, columnBorderType);
f->apply(src, dst); f->apply(src, dst, Rect(0, 0, src.cols, src.rows));
} }
//////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////

6
modules/gpu/src/imgproc_gpu.cpp
Unescape View File

@ -973,6 +973,12 @@ bool cv::gpu::tryConvertToGpuBorderType(int cpuBorderType, int& gpuBorderType)
return true; return true;
} }
if (cpuBorderType == cv::BORDER_CONSTANT)
{
gpuBorderType = cv::gpu::BORDER_CONSTANT_GPU;
return true;
}
return false; return false;
} }

149
modules/gpu/src/opencv2/gpu/device/vecmath.hpp
Unescape View File

@ -123,278 +123,295 @@ namespace cv
{ {
typedef uchar elem_t; typedef uchar elem_t;
enum {cn=1}; enum {cn=1};
static __device__ uchar all(uchar v) {return v;} static __device__ __host__ uchar all(uchar v) {return v;}
static __device__ __host__ uchar make(uchar x) {return x;}
}; };
template<> struct VecTraits<uchar1> template<> struct VecTraits<uchar1>
{ {
typedef uchar elem_t; typedef uchar elem_t;
enum {cn=1}; enum {cn=1};
static __device__ uchar1 all(uchar v) {return make_uchar1(v);} static __device__ __host__ uchar1 all(uchar v) {return make_uchar1(v);}
static __device__ __host__ uchar1 make(uchar x) {return make_uchar1(x);}
}; };
template<> struct VecTraits<uchar2> template<> struct VecTraits<uchar2>
{ {
typedef uchar elem_t; typedef uchar elem_t;
enum {cn=2}; enum {cn=2};
static __device__ uchar2 all(uchar v) {return make_uchar2(v, v);} static __device__ __host__ uchar2 all(uchar v) {return make_uchar2(v, v);}
static __device__ __host__ uchar2 make(uchar x, uchar y) {return make_uchar2(x, y);}
}; };
template<> struct VecTraits<uchar3> template<> struct VecTraits<uchar3>
{ {
typedef uchar elem_t; typedef uchar elem_t;
enum {cn=3}; enum {cn=3};
static __device__ uchar3 all(uchar v) {return make_uchar3(v, v, v);} static __device__ __host__ uchar3 all(uchar v) {return make_uchar3(v, v, v);}
static __device__ __host__ uchar3 make(uchar x, uchar y, uchar z) {return make_uchar3(x, y, z);}
}; };
template<> struct VecTraits<uchar4> template<> struct VecTraits<uchar4>
{ {
typedef uchar elem_t; typedef uchar elem_t;
enum {cn=4}; enum {cn=4};
static __device__ uchar4 all(uchar v) {return make_uchar4(v, v, v, v);} static __device__ __host__ uchar4 all(uchar v) {return make_uchar4(v, v, v, v);}
static __device__ __host__ uchar4 make(uchar x, uchar y, uchar z, uchar w) {return make_uchar4(x, y, z, w);}
}; };
template<> struct VecTraits<char> template<> struct VecTraits<char>
{ {
typedef char elem_t; typedef char elem_t;
enum {cn=1}; enum {cn=1};
static __device__ char all(char v) {return v;} static __device__ __host__ char all(char v) {return v;}
static __device__ __host__ char make(char x) {return x;}
}; };
template<> struct VecTraits<char1> template<> struct VecTraits<char1>
{ {
typedef char elem_t; typedef char elem_t;
enum {cn=1}; enum {cn=1};
static __device__ char1 all(char v) {return make_char1(v);} static __device__ __host__ char1 all(char v) {return make_char1(v);}
static __device__ __host__ char1 make(char x) {return make_char1(x);}
}; };
template<> struct VecTraits<char2> template<> struct VecTraits<char2>
{ {
typedef char elem_t; typedef char elem_t;
enum {cn=2}; enum {cn=2};
static __device__ char2 all(char v) {return make_char2(v, v);} static __device__ __host__ char2 all(char v) {return make_char2(v, v);}
static __device__ __host__ char2 make(char x, char y) {return make_char2(x, y);}
}; };
template<> struct VecTraits<char3> template<> struct VecTraits<char3>
{ {
typedef char elem_t; typedef char elem_t;
enum {cn=3}; enum {cn=3};
static __device__ char3 all(char v) {return make_char3(v, v, v);} static __device__ __host__ char3 all(char v) {return make_char3(v, v, v);}
static __device__ __host__ char3 make(char x, char y, char z) {return make_char3(x, y, z);}
}; };
template<> struct VecTraits<char4> template<> struct VecTraits<char4>
{ {
typedef char elem_t; typedef char elem_t;
enum {cn=4}; enum {cn=4};
static __device__ char4 all(char v) {return make_char4(v, v, v, v);} static __device__ __host__ char4 all(char v) {return make_char4(v, v, v, v);}
static __device__ __host__ char4 make(char x, char y, char z, char w) {return make_char4(x, y, z, w);}
}; };
template<> struct VecTraits<ushort> template<> struct VecTraits<ushort>
{ {
typedef ushort elem_t; typedef ushort elem_t;
enum {cn=1}; enum {cn=1};
static __device__ ushort all(ushort v) {return v;} static __device__ __host__ ushort all(ushort v) {return v;}
static __device__ __host__ ushort make(ushort x) {return x;}
}; };
template<> struct VecTraits<ushort1> template<> struct VecTraits<ushort1>
{ {
typedef ushort elem_t; typedef ushort elem_t;
enum {cn=1}; enum {cn=1};
static __device__ ushort1 all(ushort v) {return make_ushort1(v);} static __device__ __host__ ushort1 all(ushort v) {return make_ushort1(v);}
static __device__ __host__ ushort1 make(ushort x) {return make_ushort1(x);}
}; };
template<> struct VecTraits<ushort2> template<> struct VecTraits<ushort2>
{ {
typedef ushort elem_t; typedef ushort elem_t;
enum {cn=2}; enum {cn=2};
static __device__ ushort2 all(ushort v) {return make_ushort2(v, v);} static __device__ __host__ ushort2 all(ushort v) {return make_ushort2(v, v);}
static __device__ __host__ ushort2 make(ushort x, ushort y) {return make_ushort2(x, y);}
}; };
template<> struct VecTraits<ushort3> template<> struct VecTraits<ushort3>
{ {
typedef ushort elem_t; typedef ushort elem_t;
enum {cn=3}; enum {cn=3};
static __device__ ushort3 all(ushort v) {return make_ushort3(v, v, v);} static __device__ __host__ ushort3 all(ushort v) {return make_ushort3(v, v, v);}
static __device__ __host__ ushort3 make(ushort x, ushort y, ushort z) {return make_ushort3(x, y, z);}
}; };
template<> struct VecTraits<ushort4> template<> struct VecTraits<ushort4>
{ {
typedef ushort elem_t; typedef ushort elem_t;
enum {cn=4}; enum {cn=4};
static __device__ ushort4 all(ushort v) {return make_ushort4(v, v, v, v);} static __device__ __host__ ushort4 all(ushort v) {return make_ushort4(v, v, v, v);}
static __device__ __host__ ushort4 make(ushort x, ushort y, ushort z, ushort w) {return make_ushort4(x, y, z, w);}
}; };
template<> struct VecTraits<short> template<> struct VecTraits<short>
{ {
typedef short elem_t; typedef short elem_t;
enum {cn=1}; enum {cn=1};
static __device__ short all(short v) {return v;} static __device__ __host__ short all(short v) {return v;}
static __device__ __host__ short make(short x) {return x;}
}; };
template<> struct VecTraits<short1> template<> struct VecTraits<short1>
{ {
typedef short elem_t; typedef short elem_t;
enum {cn=1}; enum {cn=1};
static __device__ short1 all(short v) {return make_short1(v);} static __device__ __host__ short1 all(short v) {return make_short1(v);}
static __device__ __host__ short1 make(short x) {return make_short1(x);}
}; };
template<> struct VecTraits<short2> template<> struct VecTraits<short2>
{ {
typedef short elem_t; typedef short elem_t;
enum {cn=2}; enum {cn=2};
static __device__ short2 all(short v) {return make_short2(v, v);} static __device__ __host__ short2 all(short v) {return make_short2(v, v);}
static __device__ __host__ short2 make(short x, short y) {return make_short2(x, y);}
}; };
template<> struct VecTraits<short3> template<> struct VecTraits<short3>
{ {
typedef short elem_t; typedef short elem_t;
enum {cn=3}; enum {cn=3};
static __device__ short3 all(short v) {return make_short3(v, v, v);} static __device__ __host__ short3 all(short v) {return make_short3(v, v, v);}
static __device__ __host__ short3 make(short x, short y, short z) {return make_short3(x, y, z);}
}; };
template<> struct VecTraits<short4> template<> struct VecTraits<short4>
{ {
typedef short elem_t; typedef short elem_t;
enum {cn=4}; enum {cn=4};
static __device__ short4 all(short v) {return make_short4(v, v, v, v);} static __device__ __host__ short4 all(short v) {return make_short4(v, v, v, v);}
static __device__ __host__ short4 make(short x, short y, short z, short w) {return make_short4(x, y, z, w);}
}; };
template<> struct VecTraits<uint> template<> struct VecTraits<uint>
{ {
typedef uint elem_t; typedef uint elem_t;
enum {cn=1}; enum {cn=1};
static __device__ uint all(uint v) {return v;} static __device__ __host__ uint all(uint v) {return v;}
static __device__ __host__ uint make(uint x) {return x;}
}; };
template<> struct VecTraits<uint1> template<> struct VecTraits<uint1>
{ {
typedef uint elem_t; typedef uint elem_t;
enum {cn=1}; enum {cn=1};
static __device__ uint1 all(uint v) {return make_uint1(v);} static __device__ __host__ uint1 all(uint v) {return make_uint1(v);}
static __device__ __host__ uint1 make(uint x) {return make_uint1(x);}
}; };
template<> struct VecTraits<uint2> template<> struct VecTraits<uint2>
{ {
typedef uint elem_t; typedef uint elem_t;
enum {cn=2}; enum {cn=2};
static __device__ uint2 all(uint v) {return make_uint2(v, v);} static __device__ __host__ uint2 all(uint v) {return make_uint2(v, v);}
static __device__ __host__ uint2 make(uint x, uint y) {return make_uint2(x, y);}
}; };
template<> struct VecTraits<uint3> template<> struct VecTraits<uint3>
{ {
typedef uint elem_t; typedef uint elem_t;
enum {cn=3}; enum {cn=3};
static __device__ uint3 all(uint v) {return make_uint3(v, v, v);} static __device__ __host__ uint3 all(uint v) {return make_uint3(v, v, v);}
static __device__ __host__ uint3 make(uint x, uint y, uint z) {return make_uint3(x, y, z);}
}; };
template<> struct VecTraits<uint4> template<> struct VecTraits<uint4>
{ {
typedef uint elem_t; typedef uint elem_t;
enum {cn=4}; enum {cn=4};
static __device__ uint4 all(uint v) {return make_uint4(v, v, v, v);} static __device__ __host__ uint4 all(uint v) {return make_uint4(v, v, v, v);}
static __device__ __host__ uint4 make(uint x, uint y, uint z, uint w) {return make_uint4(x, y, z, w);}
}; };
template<> struct VecTraits<int> template<> struct VecTraits<int>
{ {
typedef int elem_t; typedef int elem_t;
enum {cn=1}; enum {cn=1};
static __device__ int all(int v) {return v;} static __device__ __host__ int all(int v) {return v;}
static __device__ __host__ int make(int x) {return x;}
}; };
template<> struct VecTraits<int1> template<> struct VecTraits<int1>
{ {
typedef int elem_t; typedef int elem_t;
enum {cn=1}; enum {cn=1};
static __device__ int1 all(int v) {return make_int1(v);} static __device__ __host__ int1 all(int v) {return make_int1(v);}
static __device__ __host__ int1 make(int x) {return make_int1(x);}
}; };
template<> struct VecTraits<int2> template<> struct VecTraits<int2>
{ {
typedef int elem_t; typedef int elem_t;
enum {cn=2}; enum {cn=2};
static __device__ int2 all(int v) {return make_int2(v, v);} static __device__ __host__ int2 all(int v) {return make_int2(v, v);}
static __device__ __host__ int2 make(int x, int y) {return make_int2(x, y);}
}; };
template<> struct VecTraits<int3> template<> struct VecTraits<int3>
{ {
typedef int elem_t; typedef int elem_t;
enum {cn=3}; enum {cn=3};
static __device__ int3 all(int v) {return make_int3(v, v, v);} static __device__ __host__ int3 all(int v) {return make_int3(v, v, v);}
static __device__ __host__ int3 make(int x, int y, int z) {return make_int3(x, y, z);}
}; };
template<> struct VecTraits<int4> template<> struct VecTraits<int4>
{ {
typedef int elem_t; typedef int elem_t;
enum {cn=4}; enum {cn=4};
static __device__ int4 all(int v) {return make_int4(v, v, v, v);} static __device__ __host__ int4 all(int v) {return make_int4(v, v, v, v);}
static __device__ __host__ int4 make(int x, int y, int z, int w) {return make_int4(x, y, z, w);}
}; };
template<> struct VecTraits<float> template<> struct VecTraits<float>
{ {
typedef float elem_t; typedef float elem_t;
enum {cn=1}; enum {cn=1};
static __device__ float all(float v) {return v;} static __device__ __host__ float all(float v) {return v;}
static __device__ __host__ float make(float x) {return x;}
}; };
template<> struct VecTraits<float1> template<> struct VecTraits<float1>
{ {
typedef float elem_t; typedef float elem_t;
enum {cn=1}; enum {cn=1};
static __device__ float1 all(float v) {return make_float1(v);} static __device__ __host__ float1 all(float v) {return make_float1(v);}
static __device__ __host__ float1 make(float x) {return make_float1(x);}
}; };
template<> struct VecTraits<float2> template<> struct VecTraits<float2>
{ {
typedef float elem_t; typedef float elem_t;
enum {cn=2}; enum {cn=2};
static __device__ float2 all(float v) {return make_float2(v, v);} static __device__ __host__ float2 all(float v) {return make_float2(v, v);}
static __device__ __host__ float2 make(float x, float y) {return make_float2(x, y);}
}; };
template<> struct VecTraits<float3> template<> struct VecTraits<float3>
{ {
typedef float elem_t; typedef float elem_t;
enum {cn=3}; enum {cn=3};
static __device__ float3 all(float v) {return make_float3(v, v, v);} static __device__ __host__ float3 all(float v) {return make_float3(v, v, v);}
static __device__ __host__ float3 make(float x, float y, float z) {return make_float3(x, y, z);}
}; };
template<> struct VecTraits<float4> template<> struct VecTraits<float4>
{ {
typedef float elem_t; typedef float elem_t;
enum {cn=4}; enum {cn=4};
static __device__ float4 all(float v) {return make_float4(v, v, v, v);} static __device__ __host__ float4 all(float v) {return make_float4(v, v, v, v);}
static __device__ __host__ float4 make(float x, float y, float z, float w) {return make_float4(x, y, z, w);}
}; };
template <int cn, typename VecD> struct SatCast; template <int cn, typename VecD> struct SatCast;
template <typename VecD> struct SatCast<1, VecD> template <typename VecD> struct SatCast<1, VecD>
{ {
template <typename VecS> template <typename VecS>
__device__ VecD operator()(const VecS& v) static __device__ VecD cast(const VecS& v)
{ {
VecD res; typedef typename VecTraits<VecD>::elem_t D;
res.x = saturate_cast< VecTraits<VecD>::elem_t >(v.x); return VecTraits<VecD>::make(saturate_cast<D>(v.x));
return res;
} }
}; };
template <typename VecD> struct SatCast<2, VecD> template <typename VecD> struct SatCast<2, VecD>
{ {
template <typename VecS> template <typename VecS>
__device__ VecD operator()(const VecS& v) static __device__ VecD cast(const VecS& v)
{ {
VecD res; typedef typename VecTraits<VecD>::elem_t D;
res.x = saturate_cast< VecTraits<VecD>::elem_t >(v.x); return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y));
res.y = saturate_cast< VecTraits<VecD>::elem_t >(v.y);
return res;
} }
}; };
template <typename VecD> struct SatCast<3, VecD> template <typename VecD> struct SatCast<3, VecD>
{ {
template <typename VecS> template <typename VecS>
__device__ VecD operator()(const VecS& v) static __device__ VecD cast(const VecS& v)
{ {
VecD res; typedef typename VecTraits<VecD>::elem_t D;
res.x = saturate_cast< VecTraits<VecD>::elem_t >(v.x); return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z));
res.y = saturate_cast< VecTraits<VecD>::elem_t >(v.y);
res.y = saturate_cast< VecTraits<VecD>::elem_t >(v.z);
return res;
} }
}; };
template <typename VecD> struct SatCast<4, VecD> template <typename VecD> struct SatCast<4, VecD>
{ {
template <typename VecS> template <typename VecS>
__device__ VecD operator()(const VecS& v) static __device__ VecD cast(const VecS& v)
{ {
VecD res; typedef typename VecTraits<VecD>::elem_t D;
res.x = saturate_cast< VecTraits<VecD>::elem_t >(v.x); return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z), saturate_cast<D>(v.w));
res.y = saturate_cast< VecTraits<VecD>::elem_t >(v.y);
res.y = saturate_cast< VecTraits<VecD>::elem_t >(v.z);
res.w = saturate_cast< VecTraits<VecD>::elem_t >(v.w);
return res;
} }
}; };
template <typename VecD, typename VecS> static __device__ VecD saturate_cast_caller(const VecS& v) template <typename VecD, typename VecS> static __device__ VecD saturate_cast_caller(const VecS& v)
{ {
SatCast< return SatCast<VecTraits<VecD>::cn, VecD>::cast(v);
VecTraits<VecD>::cn,
VecD
>
cast;
return cast(v);
} }
template<typename _Tp> static __device__ _Tp saturate_cast(const uchar1& v) {return saturate_cast_caller<_Tp>(v);} template<typename _Tp> static __device__ _Tp saturate_cast(const uchar1& v) {return saturate_cast_caller<_Tp>(v);}

6
tests/gpu/src/brute_force_matcher.cpp
Unescape View File

@ -107,7 +107,7 @@ protected:
if (!compareMatches(matchesCPU, matchesGPU)) if (!compareMatches(matchesCPU, matchesGPU))
{ {
ts->printf(CvTS::LOG, "Match FAIL"); ts->printf(CvTS::LOG, "Match FAIL\n");
ts->set_failed_test_info(CvTS::FAIL_MISMATCH); ts->set_failed_test_info(CvTS::FAIL_MISMATCH);
return; return;
} }
@ -119,7 +119,7 @@ protected:
if (!compareMatches(knnMatchesCPU, knnMatchesGPU)) if (!compareMatches(knnMatchesCPU, knnMatchesGPU))
{ {
ts->printf(CvTS::LOG, "KNN Match FAIL"); ts->printf(CvTS::LOG, "KNN Match FAIL\n");
ts->set_failed_test_info(CvTS::FAIL_MISMATCH); ts->set_failed_test_info(CvTS::FAIL_MISMATCH);
return; return;
} }
@ -131,7 +131,7 @@ protected:
if (!compareMatches(radiusMatchesCPU, radiusMatchesGPU)) if (!compareMatches(radiusMatchesCPU, radiusMatchesGPU))
{ {
ts->printf(CvTS::LOG, "Radius Match FAIL"); ts->printf(CvTS::LOG, "Radius Match FAIL\n");
ts->set_failed_test_info(CvTS::FAIL_MISMATCH); ts->set_failed_test_info(CvTS::FAIL_MISMATCH);
return; return;
} }

14
tests/gpu/src/filters.cpp
Unescape View File

@ -80,7 +80,8 @@ protected:
double res = norm(m1ROI, m2ROI, NORM_INF); double res = norm(m1ROI, m2ROI, NORM_INF);
if (res <= 1) // Max difference (2.0) in GaussianBlur
if (res <= 2)
return CvTS::OK; return CvTS::OK;
ts->printf(CvTS::LOG, "Norm: %f\n", res); ts->printf(CvTS::LOG, "Norm: %f\n", res);
@ -166,8 +167,6 @@ struct CV_GpuNppImageSobelTest : public CV_GpuNppFilterTest
int test(const Mat& img) int test(const Mat& img)
{ {
if (img.type() != CV_8UC1)
return CvTS::OK;
int ksizes[] = {3, 5, 7}; int ksizes[] = {3, 5, 7};
int ksizes_num = sizeof(ksizes) / sizeof(int); int ksizes_num = sizeof(ksizes) / sizeof(int);
@ -183,10 +182,8 @@ struct CV_GpuNppImageSobelTest : public CV_GpuNppFilterTest
cv::Sobel(img, cpudst, -1, dx, dy, ksizes[i]); cv::Sobel(img, cpudst, -1, dx, dy, ksizes[i]);
GpuMat gpu1(img); GpuMat gpu1(img);
gpu1.convertTo(gpu1, CV_32S);
GpuMat gpudst; GpuMat gpudst;
cv::gpu::Sobel(gpu1, gpudst, -1, dx, dy, ksizes[i]); cv::gpu::Sobel(gpu1, gpudst, -1, dx, dy, ksizes[i]);
gpudst.convertTo(gpudst, CV_8U);
if (CheckNorm(cpudst, gpudst, Size(ksizes[i], ksizes[i])) != CvTS::OK) if (CheckNorm(cpudst, gpudst, Size(ksizes[i], ksizes[i])) != CvTS::OK)
test_res = CvTS::FAIL_GENERIC; test_res = CvTS::FAIL_GENERIC;
@ -204,19 +201,14 @@ struct CV_GpuNppImageScharrTest : public CV_GpuNppFilterTest
int test(const Mat& img) int test(const Mat& img)
{ {
if (img.type() != CV_8UC1)
return CvTS::OK;
int dx = 1, dy = 0; int dx = 1, dy = 0;
Mat cpudst; Mat cpudst;
cv::Scharr(img, cpudst, -1, dx, dy); cv::Scharr(img, cpudst, -1, dx, dy);
GpuMat gpu1(img); GpuMat gpu1(img);
gpu1.convertTo(gpu1, CV_32S);
GpuMat gpudst; GpuMat gpudst;
cv::gpu::Scharr(gpu1, gpudst, -1, dx, dy); cv::gpu::Scharr(gpu1, gpudst, -1, dx, dy);
gpudst.convertTo(gpudst, CV_8U);
return CheckNorm(cpudst, gpudst, Size(3, 3)); return CheckNorm(cpudst, gpudst, Size(3, 3));
} }
@ -244,7 +236,7 @@ struct CV_GpuNppImageGaussianBlurTest : public CV_GpuNppFilterTest
{ {
cv::Size ksize(ksizes[i], ksizes[j]); cv::Size ksize(ksizes[i], ksizes[j]);
ts->printf(CvTS::LOG, "ksize = (%dx%d)\t", ksizes[i], ksizes[j]); ts->printf(CvTS::LOG, "ksize = (%dx%d)\t\n", ksizes[i], ksizes[j]);
Mat cpudst; Mat cpudst;
cv::GaussianBlur(img, cpudst, ksize, sigma1); cv::GaussianBlur(img, cpudst, ksize, sigma1);

Write Preview
Loading…
Cancel
Save