Skip to content

O
opencv
Commit 5af529c1 authored by Vladislav Vinogradov's avatar Vladislav Vinogradov
Browse files
Options

fixed and updated gpu implementation of separable liner filters 

now it supports kernel's size up to 32
parent b96a556f
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 13 additions and 27 deletions
modules/gpu/perf/perf_filters.cpp
View file @ 5af529c1
......@@ -139,6 +139,6 @@ INSTANTIATE_TEST_CASE_P(Filter, SeparableLinearFilter, testing::Combine(
ALL_DEVICES,
GPU_TYPICAL_MAT_SIZES,
testing::Values(CV_8UC1, CV_8UC4, CV_32FC1),
testing::Values(3, 5)));
testing::Values(3, 5, 7, 9, 11, 13, 15)));
#endif
modules/gpu/src/cuda/column_filter.cu
View file @ 5af529c1
This diff is collapsed.
modules/gpu/src/cuda/row_filter.cu
View file @ 5af529c1
This diff is collapsed.
modules/gpu/src/filtering.cpp
View file @ 5af529c1
......@@ -740,13 +740,13 @@ namespace cv { namespace gpu { namespace device
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);
void linearRowFilter_gpu(DevMem2Db src, 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);
void linearColumnFilter_gpu(DevMem2Db src, DevMem2Db dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
}
}}}
......@@ -755,7 +755,7 @@ namespace
typedef NppStatus (*nppFilter1D_t)(const Npp8u * pSrc, Npp32s nSrcStep, Npp8u * pDst, Npp32s nDstStep, NppiSize oROI,
const Npp32s * pKernel, Npp32s nMaskSize, Npp32s nAnchor, Npp32s nDivisor);
typedef void (*gpuFilter1D_t)(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
typedef void (*gpuFilter1D_t)(DevMem2Db src, DevMem2Db dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
struct NppLinearRowFilter : public BaseRowFilter_GPU
{
......@@ -825,8 +825,7 @@ Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType,
int gpuBorderType;
CV_Assert(tryConvertToGpuBorderType(borderType, gpuBorderType));
CV_Assert(srcType == CV_8UC1 || srcType == CV_8UC4 /*|| srcType == CV_16SC1*/ /*|| srcType == CV_16SC2*/
|| srcType == CV_16SC3 || srcType == CV_32SC1 || srcType == CV_32FC1);
CV_Assert(srcType == CV_8UC1 || srcType == CV_8UC4 || srcType == CV_16SC3 || srcType == CV_32SC1 || srcType == CV_32FC1);
CV_Assert(CV_MAT_DEPTH(bufType) == CV_32F && CV_MAT_CN(srcType) == CV_MAT_CN(bufType));
......@@ -836,7 +835,7 @@ Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType,
int ksize = cont_krnl.cols;
CV_Assert(ksize > 0 && ksize <= 16);
CV_Assert(ksize > 0 && ksize <= 32);
normalizeAnchor(anchor, ksize);
......@@ -850,12 +849,6 @@ Ptr<BaseRowFilter_GPU> cv::gpu::getLinearRowFilter_GPU(int srcType, int bufType,
case CV_8UC4:
func = linearRowFilter_gpu<uchar4, float4>;
break;
/*case CV_16SC1:
func = linearRowFilter_gpu<short, float>;
break;*/
/*case CV_16SC2:
func = linearRowFilter_gpu<short2, float2>;
break;*/
case CV_16SC3:
func = linearRowFilter_gpu<short3, float3>;
break;
......@@ -940,8 +933,7 @@ Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int ds
int gpuBorderType;
CV_Assert(tryConvertToGpuBorderType(borderType, gpuBorderType));
CV_Assert(dstType == CV_8UC1 || dstType == CV_8UC4 /*|| dstType == CV_16SC1*/ /*|| dstType == CV_16SC2*/
|| dstType == CV_16SC3 || dstType == CV_32SC1 || dstType == CV_32FC1);
CV_Assert(dstType == CV_8UC1 || dstType == CV_8UC4 || dstType == CV_16SC3 || dstType == CV_32SC1 || dstType == CV_32FC1);
CV_Assert(CV_MAT_DEPTH(bufType) == CV_32F && CV_MAT_CN(dstType) == CV_MAT_CN(bufType));
......@@ -951,7 +943,7 @@ Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int ds
int ksize = cont_krnl.cols;
CV_Assert(ksize > 0 && ksize <= 16);
CV_Assert(ksize > 0 && ksize <= 32);
normalizeAnchor(anchor, ksize);
......@@ -965,12 +957,6 @@ Ptr<BaseColumnFilter_GPU> cv::gpu::getLinearColumnFilter_GPU(int bufType, int ds
case CV_8UC4:
func = linearColumnFilter_gpu<float4, uchar4>;
break;
/*case CV_16SC1:
func = linearColumnFilter_gpu<float, short>;
break;*/
/*case CV_16SC2:
func = linearColumnFilter_gpu<float2, short2>;
break;*/
case CV_16SC3:
func = linearColumnFilter_gpu<float3, short3>;
break;
......
modules/gpu/test/test_filters.cpp
View file @ 5af529c1
......@@ -188,7 +188,7 @@ TEST_P(Sobel, Rgba)
dev_dst_rgba.download(dst_rgba);
EXPECT_MAT_NEAR_KSIZE(dst_gold_rgba, dst_rgba, ksize, 0.0);
EXPECT_MAT_NEAR(dst_gold_rgba, dst_rgba, 0.0);
}
TEST_P(Sobel, Gray)
......@@ -204,7 +204,7 @@ TEST_P(Sobel, Gray)
dev_dst_gray.download(dst_gray);
EXPECT_MAT_NEAR_KSIZE(dst_gold_gray, dst_gray, ksize, 0.0);
EXPECT_MAT_NEAR(dst_gold_gray, dst_gray, 0.0);
}
INSTANTIATE_TEST_CASE_P(Filter, Sobel, Combine(
......@@ -342,7 +342,7 @@ TEST_P(GaussianBlur, Rgba)
dev_dst_rgba.download(dst_rgba);
EXPECT_MAT_NEAR_KSIZE(dst_gold_rgba, dst_rgba, ksize, 3.0);
EXPECT_MAT_NEAR(dst_gold_rgba, dst_rgba, 4.0);
}
TEST_P(GaussianBlur, Gray)
......@@ -355,12 +355,12 @@ TEST_P(GaussianBlur, Gray)
dev_dst_gray.download(dst_gray);
EXPECT_MAT_NEAR_KSIZE(dst_gold_gray, dst_gray, ksize, 3.0);
EXPECT_MAT_NEAR(dst_gold_gray, dst_gray, 4.0);
}
INSTANTIATE_TEST_CASE_P(Filter, GaussianBlur, Combine(
ALL_DEVICES,
Values(cv::Size(3, 3), cv::Size(5, 5), cv::Size(7, 7)),
Values(cv::Size(3, 3), cv::Size(5, 5), cv::Size(7, 7), cv::Size(9, 9), cv::Size(11, 11), cv::Size(13, 13), cv::Size(15, 15), cv::Size(17, 17), cv::Size(19, 19), cv::Size(21, 21), cv::Size(23, 23), cv::Size(25, 25), cv::Size(27, 27), cv::Size(29, 29), cv::Size(31, 31)),
USE_ROI));
/////////////////////////////////////////////////////////////////////////////////////////////////
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment