potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

add support for CUDA streams

Browse Source
This commit is contained in:
marina.kolpakova
2012-11-10 03:59:09 +04:00
parent 40600fa504
commit 0cbf9eb22a
7 changed files with 140 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/gpu/src/cuda/integral_image.cu
-1
View File
@@ -444,7 +444,6 @@ namespace cv { namespace gpu { namespace device
}
// used for frame preprocessing before Soft Cascade evaluation: no synchronization needed
// ToDo: partial dy
void shfl_integral_gpu_buffered(PtrStepSzb img, PtrStepSz<uint4> buffer, PtrStepSz<unsigned int> integral,
int blockStep, cudaStream_t stream)
{
modules/gpu/src/cuda/isf-sc.cu
+17 -11
View File
@@ -71,7 +71,7 @@ namespace icf {
}
void fillBins(cv::gpu::PtrStepSzb hogluv, const cv::gpu::PtrStepSzf& nangle,
const int fw, const int fh, const int bins)
const int fw, const int fh, const int bins, cudaStream_t stream )
{
const uchar* mag = (const uchar*)hogluv.ptr(fh * bins);
uchar* hog = (uchar*)hogluv.ptr();
@@ -80,9 +80,12 @@ namespace icf {
dim3 block(32, 8);
dim3 grid(fw / 32, fh / 8);
magToHist<<<grid, block>>>(mag, angle, nangle.step / sizeof(float), hog, hogluv.step, fh);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
magToHist<<<grid, block, 0, stream>>>(mag, angle, nangle.step / sizeof(float), hog, hogluv.step, fh);
if (!stream)
{
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
}
texture<int, cudaTextureType2D, cudaReadModeElementType> thogluv;
@@ -305,7 +308,7 @@ namespace icf {
template<>
void CascadeInvoker<CascadePolicy>::operator()(const PtrStepSzb& roi, const PtrStepSzi& hogluv,
PtrStepSz<uchar4> objects, PtrStepSzi counter, const int downscales, const int scale) const
PtrStepSz<uchar4> objects, PtrStepSzi counter, const int downscales, const int scale, const cudaStream_t& stream) const
{
int fw = 160;
int fh = 120;
@@ -325,22 +328,25 @@ namespace icf {
if (scale == -1)
{
test_kernel_warp<false><<<grid, block>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, 0);
test_kernel_warp<false><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, 0);
cudaSafeCall( cudaGetLastError());
grid = dim3(fw, fh / 8, 47 - downscales);
test_kernel_warp<true><<<grid, block>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, downscales);
test_kernel_warp<true><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, downscales);
}
else
{
if (scale >= downscales)
test_kernel_warp<true><<<grid, block>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, scale);
test_kernel_warp<true><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, scale);
else
test_kernel_warp<false><<<grid, block>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, scale);
test_kernel_warp<false><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, scale);
}
cudaSafeCall( cudaGetLastError());
cudaSafeCall( cudaDeviceSynchronize());
if (!stream)
{
cudaSafeCall( cudaGetLastError());
cudaSafeCall( cudaDeviceSynchronize());
}
}
}
}}}