Skip to content

O
opencv
Commit 98c8ecf8 authored by Andrey Morozov's avatar Andrey Morozov
Browse files
Options

modified kernel setto(), added double type, code has been improved

parent 3f5dd5f1
Show whitespace changes
Inline Side-by-side
Showing with 98 additions and 190 deletions
modules/gpu/src/cuda/cuda_shared.hpp
View file @ 98c8ecf8
......@@ -61,8 +61,8 @@ namespace cv
{
static inline int divUp(int a, int b) { return (a % b == 0) ? a/b : a/b + 1; }
extern "C" void set_to_without_mask (const DevMem2D& mat, const double * scalar, int depth, int channels);
extern "C" void set_to_with_mask (const DevMem2D& mat, const double * scalar, const DevMem2D& mask, int depth, int channels);
extern "C" void set_to_without_mask (const DevMem2D& mat, int depth, const double * scalar, int channels);
extern "C" void set_to_with_mask (const DevMem2D& mat, int depth, const double * scalar, const DevMem2D& mask, int channels);
extern "C" void convert_to(const DevMem2D& src, int sdepth, DevMem2D dst, int ddepth, size_t width, size_t height, double alpha, double beta);
}
......
modules/gpu/src/cuda/matrix_operations.cu
View file @ 98c8ecf8
......@@ -49,7 +49,7 @@
using namespace cv::gpu;
using namespace cv::gpu::impl;
__constant__ __align__(16) float scalar_d[4];
__constant__ __align__(16) double scalar_d[4];
namespace mat_operators
{
......@@ -57,8 +57,8 @@ namespace mat_operators
// SetTo
//////////////////////////////////////////////////////////
template<typename T, int channels>
__global__ void kernel_set_to_without_mask(T * mat, int cols, int rows, int step)
template<typename T>
__global__ void kernel_set_to_without_mask(T * mat, int cols, int rows, int step, int channels)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
......@@ -70,14 +70,14 @@ namespace mat_operators
}
}
template<typename T, int channels>
__global__ void kernel_set_to_with_mask(T * mat, const unsigned char * mask, int cols, int rows, int step, int step_mask)
template<typename T>
__global__ void kernel_set_to_with_mask(T * mat, const unsigned char * mask, int cols, int rows, int step, int channels, int step_mask)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
if (mask[y * step_mask + x] != 0)
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
{
size_t idx = y * (step / sizeof(T)) + x;
mat[idx] = scalar_d[ x % channels ];
......@@ -319,100 +319,100 @@ namespace mat_operators
} // namespace mat_operators
//////////////////////////////////////////////////////////////
// SetTo
//////////////////////////////////////////////////////////////
extern "C" void cv::gpu::impl::set_to_without_mask(const DevMem2D& mat, const double * scalar, int elemSize1, int channels)
namespace cv
{
float data[4];
data[0] = static_cast<float>(scalar[0]);
data[1] = static_cast<float>(scalar[1]);
data[2] = static_cast<float>(scalar[2]);
data[3] = static_cast<float>(scalar[3]);
cudaSafeCall( cudaMemcpyToSymbol(scalar_d, &data, sizeof(data)));
namespace gpu
{
namespace impl
{
dim3 threadsPerBlock(16, 16, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
//////////////////////////////////////////////////////////////
// SetTo
//////////////////////////////////////////////////////////////
if (channels == 1)
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_without_mask<unsigned char, 1><<<numBlocks,threadsPerBlock>>>(mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_without_mask<unsigned short, 1><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_without_mask<float, 1><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, mat.cols, mat.rows, mat.step);
}
if (channels == 2)
typedef void (*SetToFunc_with_mask)(const DevMem2D& mat, const DevMem2D& mask, int channels);
typedef void (*SetToFunc_without_mask)(const DevMem2D& mat, int channels);
template <typename T>
void set_to_with_mask_run(const DevMem2D& mat, const DevMem2D& mask, int channels)
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_without_mask<unsigned char, 2><<<numBlocks,threadsPerBlock>>>(mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_without_mask<unsigned short, 2><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_without_mask<float, 2><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, mat.cols, mat.rows, mat.step);
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
::mat_operators::kernel_set_to_with_mask<T><<<numBlocks,threadsPerBlock>>>((T*)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, channels, mask.step);
cudaSafeCall ( cudaThreadSynchronize() );
}
if (channels == 3)
template <typename T>
void set_to_without_mask_run(const DevMem2D& mat, int channels)
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_without_mask<unsigned char, 3><<<numBlocks,threadsPerBlock>>>(mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_without_mask<unsigned short, 3><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_without_mask<float, 3><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, mat.cols, mat.rows, mat.step);
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
::mat_operators::kernel_set_to_without_mask<T><<<numBlocks,threadsPerBlock>>>((T*)mat.ptr, mat.cols, mat.rows, mat.step, channels);
cudaSafeCall ( cudaThreadSynchronize() );
}
if (channels == 4)
extern "C" void set_to_without_mask(const DevMem2D& mat, int depth, const double * scalar, int channels)
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_without_mask<unsigned char, 4><<<numBlocks,threadsPerBlock>>>(mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_without_mask<unsigned short, 4><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, mat.cols, mat.rows, mat.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_without_mask<float, 4><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, mat.cols, mat.rows, mat.step);
}
double data[4];
data[0] = scalar[0];
data[1] = scalar[1];
data[2] = scalar[2];
data[3] = scalar[3];
cudaSafeCall( cudaMemcpyToSymbol(scalar_d, &data, sizeof(data)));
cudaSafeCall ( cudaThreadSynchronize() );
}
static SetToFunc_without_mask tab[8] =
{
set_to_without_mask_run<unsigned char>,
set_to_without_mask_run<char>,
set_to_without_mask_run<unsigned short>,
set_to_without_mask_run<short>,
set_to_without_mask_run<int>,
set_to_without_mask_run<float>,
set_to_without_mask_run<double>,
0
};
extern "C" void cv::gpu::impl::set_to_with_mask(const DevMem2D& mat, const double * scalar, const DevMem2D& mask, int elemSize1, int channels)
{
float data[4];
data[0] = static_cast<float>(scalar[0]);
data[1] = static_cast<float>(scalar[1]);
data[2] = static_cast<float>(scalar[2]);
data[3] = static_cast<float>(scalar[3]);
cudaSafeCall( cudaMemcpyToSymbol(scalar_d, &data, sizeof(data)));
SetToFunc_without_mask func = tab[depth];
dim3 threadsPerBlock(16, 16, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
if (func == 0) error("Operation \'ConvertTo\' doesn't supported on your GPU model", __FILE__, __LINE__);
if (channels == 1)
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_with_mask<unsigned char, 1><<<numBlocks,threadsPerBlock>>>(mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_with_mask<unsigned short, 1><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_with_mask<float, 1><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
}
if (channels == 2)
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_with_mask<unsigned char, 2><<<numBlocks,threadsPerBlock>>>(mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_with_mask<unsigned short, 2><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_with_mask<float, 2><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
func(mat, channels);
}
if (channels == 3)
extern "C" void set_to_with_mask(const DevMem2D& mat, int depth, const double * scalar, const DevMem2D& mask, int channels)
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_with_mask<unsigned char, 3><<<numBlocks,threadsPerBlock>>>(mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_with_mask<unsigned short, 3><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_with_mask<float, 3><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
}
if (channels == 4)
double data[4];
data[0] = scalar[0];
data[1] = scalar[1];
data[2] = scalar[2];
data[3] = scalar[3];
cudaSafeCall( cudaMemcpyToSymbol(scalar_d, &data, sizeof(data)));
static SetToFunc_with_mask tab[8] =
{
if (elemSize1 == 1) ::mat_operators::kernel_set_to_with_mask<unsigned char, 4><<<numBlocks,threadsPerBlock>>>(mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 2) ::mat_operators::kernel_set_to_with_mask<unsigned short, 4><<<numBlocks,threadsPerBlock>>>((unsigned short *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
if (elemSize1 == 4) ::mat_operators::kernel_set_to_with_mask<float, 4><<<numBlocks,threadsPerBlock>>>((float *)mat.ptr, (unsigned char *)mask.ptr, mat.cols, mat.rows, mat.step, mask.step);
set_to_with_mask_run<unsigned char>,
set_to_with_mask_run<char>,
set_to_with_mask_run<unsigned short>,
set_to_with_mask_run<short>,
set_to_with_mask_run<int>,
set_to_with_mask_run<float>,
set_to_with_mask_run<double>,
0
};
SetToFunc_with_mask func = tab[depth];
if (func == 0) error("Operation \'ConvertTo\' doesn't supported on your GPU model", __FILE__, __LINE__);
func(mat, mask, channels);
}
cudaSafeCall ( cudaThreadSynchronize() );
}
//////////////////////////////////////////////////////////////
// ConvertTo
//////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////
// ConvertTo
//////////////////////////////////////////////////////////////
namespace cv
{
namespace gpu
{
namespace impl
{
typedef void (*CvtFunc)(const DevMem2D& src, DevMem2D& dst, size_t width, size_t height, double alpha, double beta);
......
modules/gpu/src/matrix_operations.cpp
View file @ 98c8ecf8
......@@ -133,7 +133,7 @@ void cv::gpu::GpuMat::convertTo( GpuMat& dst, int rtype, double alpha, double be
GpuMat& GpuMat::operator = (const Scalar& s)
{
cv::gpu::impl::set_to_without_mask(*this, s.val, this->elemSize1(), this->channels());
cv::gpu::impl::set_to_without_mask( *this, this->depth(), s.val, this->channels());
return *this;
}
......@@ -145,11 +145,11 @@ GpuMat& GpuMat::setTo(const Scalar& s, const GpuMat& mask)
if (mask.empty())
{
cv::gpu::impl::set_to_without_mask(*this, s.val, this->elemSize1(), this->channels());
cv::gpu::impl::set_to_without_mask( *this, this->depth(), s.val, this->channels());
}
else
{
cv::gpu::impl::set_to_with_mask(*this, s.val, mask, this->elemSize1(), this->channels());
cv::gpu::impl::set_to_with_mask( *this, this->depth(), s.val, mask, this->channels());
}
return *this;
......
tests/gpu/src/operator_set_to.cpp
View file @ 98c8ecf8
......@@ -24,19 +24,6 @@ class CV_GpuMatOpSetTo : public CvTest
bool compare_matrix(cv::Mat & cpumat, gpu::GpuMat & gpumat);
bool test_cv_8u_c1();
bool test_cv_8u_c2();
bool test_cv_8u_c3();
bool test_cv_8u_c4();
bool test_cv_16u_c4();
bool test_cv_32f_c1();
bool test_cv_32f_c2();
bool test_cv_32f_c3();
bool test_cv_32f_c4();
private:
int rows;
int cols;
......@@ -45,13 +32,13 @@ class CV_GpuMatOpSetTo : public CvTest
CV_GpuMatOpSetTo::CV_GpuMatOpSetTo(): CvTest( "GpuMatOperatorSetTo", "setTo" )
{
rows = 129;
cols = 127;
rows = 256;
cols = 124;
s.val[0] = 128.0;
s.val[1] = 128.0;
s.val[2] = 128.0;
s.val[3] = 128.0;
s.val[0] = 127.0;
s.val[1] = 127.0;
s.val[2] = 127.0;
s.val[3] = 127.0;
//#define PRINT_MATRIX
}
......@@ -99,95 +86,16 @@ bool CV_GpuMatOpSetTo::compare_matrix(cv::Mat & cpumat, gpu::GpuMat & gpumat)
}
}
bool CV_GpuMatOpSetTo::test_cv_8u_c1()
{
Mat cpumat(rows, cols, CV_8U, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_8u_c2()
{
Mat cpumat(rows, cols, CV_8UC2, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_8u_c3()
{
Mat cpumat(rows, cols, CV_8UC3, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_8u_c4()
{
Mat cpumat(rows, cols, CV_8UC4, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_16u_c4()
{
Mat cpumat(rows, cols, CV_16UC4, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_32f_c1()
{
Mat cpumat(rows, cols, CV_32F, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_32f_c2()
{
Mat cpumat(rows, cols, CV_32FC2, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_32f_c3()
{
Mat cpumat(rows, cols, CV_32FC3, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
bool CV_GpuMatOpSetTo::test_cv_32f_c4()
{
Mat cpumat(rows, cols, CV_32FC4, Scalar::all(0));
GpuMat gpumat(cpumat);
return compare_matrix(cpumat, gpumat);
}
void CV_GpuMatOpSetTo::run( int /* start_from */)
{
bool is_test_good = true;
is_test_good &= test_cv_8u_c1();
is_test_good &= test_cv_8u_c2();
is_test_good &= test_cv_8u_c3();
is_test_good &= test_cv_8u_c4();
is_test_good &= test_cv_16u_c4();
is_test_good &= test_cv_32f_c1();
is_test_good &= test_cv_32f_c2();
is_test_good &= test_cv_32f_c3();
is_test_good &= test_cv_32f_c4();
for (int i = 0; i < 7; i++)
{
Mat cpumat(rows, cols, i, Scalar::all(0));
GpuMat gpumat(cpumat);
is_test_good &= compare_matrix(cpumat, gpumat);
}
if (is_test_good == true)
ts->set_failed_test_info(CvTS::OK);
......
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