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

optimized gpu::multiply

Browse Source
This commit is contained in:
Vladislav Vinogradov 2011-09-12 08:45:56 +00:00
parent 6763bd6d01
commit acac27d8fe
2 changed files with 43 additions and 53 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

77
modules/gpu/src/cuda/element_operations.cu
View File

@ -607,68 +607,59 @@ namespace cv { namespace gpu { namespace device
//////////////////////////////////////////////////////////////////////////
// multiply
// TODO implement more efficient version
template <typename TSrc1, typename TSrc2, typename TDst, int cn>
void __global__ multiplyKernel(const PtrStep src1, const PtrStep src2, int rows, int cols,
PtrStep dst)
struct multiply_8uc4_32f : binary_function<uint, float, uint>
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < cols && y < rows)
__device__ __forceinline__ uint operator ()(uint a, float b) const
{
((TDst*)dst.ptr(y))[x] = saturate_cast<TDst>(((TSrc1*)src1.ptr(y))[x] * ((TSrc2*)src2.ptr(y))[x / cn]);
uint res = 0;
res |= 0xffu & (saturate_cast<uchar>((0xffu & (a )) * b) );
res |= 0xffu & (saturate_cast<uchar>((0xffu & (a >> 8)) * b) << 8);
res |= 0xffu & (saturate_cast<uchar>((0xffu & (a >> 16)) * b) << 16);
res |= 0xffu & (saturate_cast<uchar>((0xffu & (a >> 24)) * b) << 24);
return res;
}
}
};
template <typename TSrc1, typename TSrc2, typename TDst, int cn>
void multiplyCaller(const PtrStep src1, const PtrStep src2, int rows, int cols, PtrStep dst, cudaStream_t stream)
template <> struct TransformFunctorTraits<multiply_8uc4_32f> : DefaultTransformFunctorTraits<multiply_8uc4_32f>
{
dim3 threads(32, 8);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
enum { smart_block_dim_x = 8 };
enum { smart_block_dim_y = 8 };
enum { smart_shift = 8 };
};
multiplyKernel<TSrc1, TSrc2, TDst, cn><<<grid, threads>>>(src1, src2, rows, cols, dst);
cudaSafeCall(cudaGetLastError());
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
void multiply_gpu(const DevMem2D_<uchar4>& src1, const DevMem2Df& src2, const DevMem2D_<uchar4>& dst, cudaStream_t stream)
{
transform(static_cast< DevMem2D_<uint> >(src1), src2, static_cast< DevMem2D_<uint> >(dst), multiply_8uc4_32f(), stream);
}
template void multiplyCaller<uchar, float, uchar, 4>(const PtrStep src1, const PtrStep src2, int rows, int cols, PtrStep dst, cudaStream_t stream);
//////////////////////////////////////////////////////////////////////////
// multiply (by scalar)
// TODO implement more efficient version
template <typename TSrc, typename TDst>
void __global__ multiplyScalarKernel(const PtrStep src1, float scale, int rows, int cols, PtrStep dst)
template <typename T, typename D, typename S> struct MultiplyScalar : unary_function<T, D>
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
__host__ __device__ __forceinline__ MultiplyScalar(typename TypeTraits<S>::ParameterType scale_) : scale(scale_) {}
if (x < cols && y < rows)
__device__ __forceinline__ D operator ()(typename TypeTraits<T>::ParameterType a) const
{
((TDst*)dst.ptr(y))[x] = saturate_cast<TDst>(((TSrc*)src1.ptr(y))[x] * scale);
return saturate_cast<D>(a * scale);
}
}
const S scale;
};
template <typename TSrc, typename TDst>
void multiplyScalarCaller(const PtrStep src, float scale, int rows, int cols, PtrStep dst, cudaStream_t stream)
template <> struct TransformFunctorTraits< MultiplyScalar<uchar, uchar, float> > : DefaultTransformFunctorTraits< MultiplyScalar<uchar, uchar, float> >
{
dim3 threads(32, 8);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
enum { smart_block_dim_y = 8 };
enum { smart_shift = 8 };
};
multiplyScalarKernel<TSrc, TDst><<<grid, threads>>>(src, scale, rows, cols, dst);
cudaSafeCall(cudaGetLastError());
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
template <typename T, typename D>
void multiplyScalar_gpu(const DevMem2D& src, float scale, const DevMem2D& dst, cudaStream_t stream)
{
transform(static_cast< DevMem2D_<T> >(src), static_cast< DevMem2D_<D> >(dst), MultiplyScalar<T, D, float>(scale), stream);
}
template void multiplyScalarCaller<uchar, uchar>(const PtrStep src, float scale, int rows, int cols, PtrStep dst, cudaStream_t stream);
template void multiplyScalar_gpu<uchar, uchar>(const DevMem2D& src, float scale, const DevMem2D& dst, cudaStream_t stream);
}}}

19
modules/gpu/src/element_operations.cpp
View File

@ -199,22 +199,21 @@ void cv::gpu::subtract(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, Stre
namespace cv { namespace gpu { namespace device
{
template <typename TSrc1, typename TSrc2, typename TDst, int cn>
void multiplyCaller(const PtrStep src1, const PtrStep src2, int rows, int cols, PtrStep dst, cudaStream_t stream);
void multiply_gpu(const DevMem2D_<uchar4>& src1, const DevMem2Df& src2, const DevMem2D_<uchar4>& dst, cudaStream_t stream);
template <typename TSrc, typename TDst>
void multiplyScalarCaller(const PtrStep src, float scalar, int rows, int cols, PtrStep dst, cudaStream_t stream);
template <typename T, typename D>
void multiplyScalar_gpu(const DevMem2D& src, float scale, const DevMem2D& dst, cudaStream_t stream);
}}}
void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, Stream& stream)
{
if (src1.type() == CV_8UC4 && src2.type() == CV_32F)
if (src1.type() == CV_8UC4 && src2.type() == CV_32FC1)
{
CV_Assert(src1.size() == src2.size());
dst.create(src1.size(), src1.type());
device::multiplyCaller<uchar, float, uchar, 4>(static_cast<DevMem2D>(src1), static_cast<DevMem2D>(src2),
src1.rows, src1.cols * 4, static_cast<DevMem2D>(dst),
StreamAccessor::getStream(stream));
device::multiply_gpu(src1, src2, dst, StreamAccessor::getStream(stream));
}
else
nppArithmCaller(src1, src2, dst, nppiMul_8u_C1RSfs, nppiMul_8u_C4RSfs, nppiMul_32s_C1R, nppiMul_32f_C1R, StreamAccessor::getStream(stream));
@ -225,8 +224,8 @@ void cv::gpu::multiply(const GpuMat& src, const Scalar& sc, GpuMat& dst, Stream&
if (src.depth() == CV_8U)
{
dst.create(src.size(), src.type());
device::multiplyScalarCaller<uchar, uchar>(static_cast<DevMem2D>(src), (float)(sc[0]), src.rows, src.cols * src.channels(),
static_cast<DevMem2D>(dst), StreamAccessor::getStream(stream));
device::multiplyScalar_gpu<uchar, uchar>(src.reshape(1), (float)(sc[0]), dst, StreamAccessor::getStream(stream));
}
else
{