Skip to content

O
opencv
Commit 72f020a8 authored by Alexey Spizhevoy's avatar Alexey Spizhevoy
Browse files
Options

added gpu::count_non_zero version for CC1.0, refactored gpu module a little

parent 120a3b75
Hide whitespace changes
Inline Side-by-side
Showing with 117 additions and 35 deletions
modules/gpu/include/opencv2/gpu/gpu.hpp
View file @ 72f020a8
......@@ -68,6 +68,9 @@ namespace cv
CV_EXPORTS void getGpuMemInfo(size_t& free, size_t& total);
CV_EXPORTS bool hasNativeDoubleSupport(int device);
CV_EXPORTS bool hasAtomicsSupport(int device);
//////////////////////////////// Error handling ////////////////////////
CV_EXPORTS void error(const char *error_string, const char *file, const int line, const char *func);
......
modules/gpu/src/arithm.cpp
View file @ 72f020a8
......@@ -665,15 +665,33 @@ int cv::gpu::countNonZero(const GpuMat& src, GpuMat& buf)
get_buf_size_required(buf_size.width, buf_size.height);
buf.create(buf_size, CV_8U);
switch (src.type())
int device = getDevice();
if (hasAtomicsSupport(device))
{
switch (src.type())
{
case CV_8U: return count_non_zero_caller<unsigned char>(src, buf);
case CV_8S: return count_non_zero_caller<signed char>(src, buf);
case CV_16U: return count_non_zero_caller<unsigned short>(src, buf);
case CV_16S: return count_non_zero_caller<signed short>(src, buf);
case CV_32S: return count_non_zero_caller<int>(src, buf);
case CV_32F: return count_non_zero_caller<float>(src, buf);
case CV_64F:
if (hasNativeDoubleSupport(device))
return count_non_zero_caller<double>(src, buf);
}
}
else
{
case CV_8U: return count_non_zero_caller<unsigned char>(src, buf);
case CV_8S: return count_non_zero_caller<signed char>(src, buf);
case CV_16U: return count_non_zero_caller<unsigned short>(src, buf);
case CV_16S: return count_non_zero_caller<signed short>(src, buf);
case CV_32S: return count_non_zero_caller<int>(src, buf);
case CV_32F: return count_non_zero_caller<float>(src, buf);
case CV_64F: return count_non_zero_caller<double>(src, buf);
switch (src.type())
{
case CV_8U: return count_non_zero_caller_2steps<unsigned char>(src, buf);
case CV_8S: return count_non_zero_caller_2steps<signed char>(src, buf);
case CV_16U: return count_non_zero_caller_2steps<unsigned short>(src, buf);
case CV_16S: return count_non_zero_caller_2steps<signed short>(src, buf);
case CV_32S: return count_non_zero_caller_2steps<int>(src, buf);
case CV_32F: return count_non_zero_caller_2steps<float>(src, buf);
}
}
CV_Error(CV_StsBadArg, "countNonZero: unsupported type");
......
modules/gpu/src/cuda/mathfunc.cu
View file @ 72f020a8
......@@ -908,6 +908,27 @@ namespace cv { namespace gpu { namespace mathfunc
}
template <int size, typename T>
__device__ void sum_shared_mem(volatile T* data, const unsigned int tid)
{
T sum = data[tid];
if (size >= 512) if (tid < 256) { data[tid] = sum = sum + data[tid + 256]; } __syncthreads();
if (size >= 256) if (tid < 128) { data[tid] = sum = sum + data[tid + 128]; } __syncthreads();
if (size >= 128) if (tid < 64) { data[tid] = sum = sum + data[tid + 64]; } __syncthreads();
if (tid < 32)
{
if (size >= 64) data[tid] = sum = sum + data[tid + 32];
if (size >= 32) data[tid] = sum = sum + data[tid + 16];
if (size >= 16) data[tid] = sum = sum + data[tid + 8];
if (size >= 8) data[tid] = sum = sum + data[tid + 4];
if (size >= 4) data[tid] = sum = sum + data[tid + 2];
if (size >= 2) data[tid] = sum = sum + data[tid + 1];
}
}
template <int nthreads, typename T>
__global__ void count_non_zero_kernel(const DevMem2D src, volatile unsigned int* count)
{
......@@ -928,12 +949,9 @@ namespace cv { namespace gpu { namespace mathfunc
scount[tid] = cnt;
__syncthreads();
for (unsigned int step = nthreads / 2; step > 0; step >>= 1)
{
if (tid < step) scount[tid] += scount[tid + step];
__syncthreads();
}
sum_shared_mem<nthreads, unsigned int>(scount, tid);
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 110
__shared__ bool is_last;
if (tid == 0)
......@@ -950,16 +968,12 @@ namespace cv { namespace gpu { namespace mathfunc
if (is_last)
{
scount[tid] = tid < gridDim.x * gridDim.y ? count[tid] : 0;
for (unsigned int step = nthreads / 2; step > 0; step >>= 1)
{
if (tid < step) scount[tid] += scount[tid + step];
__syncthreads();
}
sum_shared_mem<nthreads, unsigned int>(scount, tid);
if (tid == 0) count[0] = scount[0];
}
#else
if (tid == 0) count[blockIdx.y * gridDim.x + blockIdx.x] = scount[0];
#endif
}
......@@ -990,6 +1004,47 @@ namespace cv { namespace gpu { namespace mathfunc
template int count_non_zero_caller<float>(const DevMem2D, PtrStep);
template int count_non_zero_caller<double>(const DevMem2D, PtrStep);
template <int nthreads, typename T>
__global__ void count_non_zero_kernel_2ndstep(unsigned int* count, int size)
{
__shared__ unsigned int scount[nthreads];
unsigned int tid = threadIdx.y * blockDim.x + threadIdx.x;
scount[tid] = tid < size ? count[tid] : 0;
sum_shared_mem<nthreads, unsigned int>(scount, tid);
if (tid == 0) count[0] = scount[0];
}
template <typename T>
int count_non_zero_caller_2steps(const DevMem2D src, PtrStep buf)
{
dim3 threads, grid;
estimate_thread_cfg(threads, grid);
estimate_kernel_consts(src.cols, src.rows, threads, grid);
unsigned int* count_buf = (unsigned int*)buf.ptr(0);
cudaSafeCall(cudaMemcpyToSymbol(blocks_finished, &czero, sizeof(blocks_finished)));
count_non_zero_kernel<256, T><<<grid, threads>>>(src, count_buf);
count_non_zero_kernel_2ndstep<256, T><<<1, 256>>>(count_buf, grid.x * grid.y);
cudaSafeCall(cudaThreadSynchronize());
unsigned int count;
cudaSafeCall(cudaMemcpy(&count, count_buf, sizeof(int), cudaMemcpyDeviceToHost));
return count;
}
template int count_non_zero_caller_2steps<unsigned char>(const DevMem2D, PtrStep);
template int count_non_zero_caller_2steps<signed char>(const DevMem2D, PtrStep);
template int count_non_zero_caller_2steps<unsigned short>(const DevMem2D, PtrStep);
template int count_non_zero_caller_2steps<signed short>(const DevMem2D, PtrStep);
template int count_non_zero_caller_2steps<int>(const DevMem2D, PtrStep);
template int count_non_zero_caller_2steps<float>(const DevMem2D, PtrStep);
} // namespace countnonzero
}}}
modules/gpu/src/initialization.cpp
View file @ 72f020a8
......@@ -55,6 +55,8 @@ CV_EXPORTS int cv::gpu::getDevice() { throw_nogpu(); return 0; }
CV_EXPORTS void cv::gpu::getComputeCapability(int /*device*/, int& /*major*/, int& /*minor*/) { throw_nogpu(); }
CV_EXPORTS int cv::gpu::getNumberOfSMs(int /*device*/) { throw_nogpu(); return 0; }
CV_EXPORTS void cv::gpu::getGpuMemInfo(size_t& /*free*/, size_t& /*total*/) { throw_nogpu(); }
CV_EXPORTS bool cv::gpu::hasNativeDoubleSupport(int /*device*/) { throw_nogpu(); return false; }
CV_EXPORTS bool cv::gpu::hasAtomicsSupport(int /*device*/) { throw_nogpu(); return false; }
#else /* !defined (HAVE_CUDA) */
......@@ -106,5 +108,19 @@ CV_EXPORTS void cv::gpu::getGpuMemInfo(size_t& free, size_t& total)
cudaSafeCall( cudaMemGetInfo( &free, &total ) );
}
CV_EXPORTS bool cv::gpu::hasNativeDoubleSupport(int device)
{
int major, minor;
getComputeCapability(device, major, minor);
return major > 1 || (major == 1 && minor >= 3);
}
CV_EXPORTS bool cv::gpu::hasAtomicsSupport(int device)
{
int major, minor;
getComputeCapability(device, major, minor);
return major > 1 || (major == 1 && minor >= 1);
}
#endif
tests/gpu/src/arithm.cpp
View file @ 72f020a8
......@@ -681,11 +681,7 @@ struct CV_GpuMinMaxTest: public CvTest
void run(int)
{
int depth_end;
int major, minor;
cv::gpu::getComputeCapability(getDevice(), major, minor);
if (minor >= 1) depth_end = CV_64F; else depth_end = CV_32F;
if (cv::gpu::hasNativeDoubleSupport(cv::gpu::getDevice())) depth_end = CV_64F; else depth_end = CV_32F;
for (int cn = 1; cn <= 4; ++cn)
for (int depth = CV_8U; depth <= depth_end; ++depth)
{
......@@ -760,10 +756,7 @@ struct CV_GpuMinMaxLocTest: public CvTest
void run(int)
{
int depth_end;
int major, minor;
cv::gpu::getComputeCapability(getDevice(), major, minor);
if (minor >= 1) depth_end = CV_64F; else depth_end = CV_32F;
if (cv::gpu::hasNativeDoubleSupport(cv::gpu::getDevice())) depth_end = CV_64F; else depth_end = CV_32F;
for (int depth = CV_8U; depth <= depth_end; ++depth)
{
int rows = 1, cols = 3;
......@@ -829,11 +822,8 @@ struct CV_GpuCountNonZeroTest: CvTest
{
srand(0);
int depth_end;
int major, minor;
cv::gpu::getComputeCapability(getDevice(), major, minor);
if (minor >= 1) depth_end = CV_64F; else depth_end = CV_32F;
for (int depth = CV_8U; depth <= depth_end; ++depth)
if (cv::gpu::hasNativeDoubleSupport(cv::gpu::getDevice())) depth_end = CV_64F; else depth_end = CV_32F;
for (int depth = CV_8U; depth <= CV_32F; ++depth)
{
for (int i = 0; i < 4; ++i)
{
......
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