Merge pull request #1061 from bitwangyaoyao:2.4_fix2

This commit is contained in:
Roman Donchenko
2013-07-01 13:21:12 +04:00
committed by OpenCV Buildbot
10 changed files with 186 additions and 38 deletions

View File

@@ -1816,8 +1816,14 @@ void cv::ocl::device::hog::normalize_hists(int nbins,
openCLExecuteKernel(clCxt, &objdetect_hog, kernelName, globalThreads,
localThreads, args, -1, -1, "-D CPU");
else
{
cl_kernel kernel = openCLGetKernelFromSource(clCxt, &objdetect_hog, kernelName);
int wave_size = queryDeviceInfo<WAVEFRONT_SIZE, int>(kernel);
char opt[32] = {0};
sprintf(opt, "-D WAVE_SIZE=%d", wave_size);
openCLExecuteKernel(clCxt, &objdetect_hog, kernelName, globalThreads,
localThreads, args, -1, -1);
localThreads, args, -1, -1, opt);
}
}
void cv::ocl::device::hog::classify_hists(int win_height, int win_width,
@@ -1879,8 +1885,14 @@ void cv::ocl::device::hog::classify_hists(int win_height, int win_width,
openCLExecuteKernel(clCxt, &objdetect_hog, kernelName, globalThreads,
localThreads, args, -1, -1, "-D CPU");
else
{
cl_kernel kernel = openCLGetKernelFromSource(clCxt, &objdetect_hog, kernelName);
int wave_size = queryDeviceInfo<WAVEFRONT_SIZE, int>(kernel);
char opt[32] = {0};
sprintf(opt, "-D WAVE_SIZE=%d", wave_size);
openCLExecuteKernel(clCxt, &objdetect_hog, kernelName, globalThreads,
localThreads, args, -1, -1);
localThreads, args, -1, -1, opt);
}
}
void cv::ocl::device::hog::extract_descrs_by_rows(int win_height, int win_width,

View File

@@ -133,7 +133,9 @@ __kernel void compute_hists_lut_kernel(
final_hist[(cell_x * 2 + cell_y) * cnbins + bin_id] =
hist_[0] + hist_[1] + hist_[2];
}
#ifdef CPU
barrier(CLK_LOCAL_MEM_FENCE);
#endif
int tid = (cell_y * CELLS_PER_BLOCK_Y + cell_x) * 12 + cell_thread_x;
if ((tid < cblock_hist_size) && (gid < blocks_total))
@@ -225,8 +227,9 @@ __kernel void compute_hists_kernel(
final_hist[(cell_x * 2 + cell_y) * cnbins + bin_id] =
hist_[0] + hist_[1] + hist_[2];
}
#ifdef CPU
barrier(CLK_LOCAL_MEM_FENCE);
#endif
int tid = (cell_y * CELLS_PER_BLOCK_Y + cell_x) * 12 + cell_thread_x;
if ((tid < cblock_hist_size) && (gid < blocks_total))
{
@@ -318,6 +321,10 @@ float reduce_smem(volatile __local float* smem, int size)
if (tid < 32)
{
if (size >= 64) smem[tid] = sum = sum + smem[tid + 32];
#if WAVE_SIZE < 32
} barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16) {
#endif
if (size >= 32) smem[tid] = sum = sum + smem[tid + 16];
if (size >= 16) smem[tid] = sum = sum + smem[tid + 8];
if (size >= 8) smem[tid] = sum = sum + smem[tid + 4];
@@ -418,6 +425,9 @@ __kernel void classify_hists_180_kernel(
{
smem[tid] = product = product + smem[tid + 32];
}
#if WAVE_SIZE < 32
barrier(CLK_LOCAL_MEM_FENCE);
#endif
if (tid < 16)
{
smem[tid] = product = product + smem[tid + 16];
@@ -487,6 +497,10 @@ __kernel void classify_hists_252_kernel(
if (tid < 32)
{
smem[tid] = product = product + smem[tid + 32];
#if WAVE_SIZE < 32
} barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16) {
#endif
smem[tid] = product = product + smem[tid + 16];
smem[tid] = product = product + smem[tid + 8];
smem[tid] = product = product + smem[tid + 4];
@@ -553,6 +567,10 @@ __kernel void classify_hists_kernel(
if (tid < 32)
{
smem[tid] = product = product + smem[tid + 32];
#if WAVE_SIZE < 32
} barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16) {
#endif
smem[tid] = product = product + smem[tid + 16];
smem[tid] = product = product + smem[tid + 8];
smem[tid] = product = product + smem[tid + 4];

View File

@@ -258,27 +258,13 @@ float sobel(__global unsigned char *input, int x, int y, int rows, int cols)
float CalcSums(__local float *cols, __local float *cols_cache, int winsz)
{
float cache = 0;
float cache2 = 0;
int winsz2 = winsz/2;
unsigned int cache = cols[0];
int x = get_local_id(0);
int group_size_x = get_local_size(0);
for(int i = 1; i <= winsz2; i++)
#pragma unroll
for(int i = 1; i <= winsz; i++)
cache += cols[i];
cols_cache[0] = cache;
barrier(CLK_LOCAL_MEM_FENCE);
if (x < group_size_x - winsz2)
cache2 = cols_cache[winsz2];
else
for(int i = winsz2 + 1; i < winsz; i++)
cache2 += cols[i];
return cols[0] + cache + cache2;
return cache;
}
#define RpT (2 * ROWSperTHREAD) // got experimentally
@@ -301,8 +287,7 @@ __kernel void textureness_kernel(__global unsigned char *disp, int disp_rows, in
int beg_row = group_id_y * RpT;
int end_row = min(beg_row + RpT, disp_rows);
// if (x < disp_cols)
// {
int y = beg_row;
float sum = 0;
@@ -340,11 +325,15 @@ __kernel void textureness_kernel(__global unsigned char *disp, int disp_rows, in
}
barrier(CLK_LOCAL_MEM_FENCE);
float sum_win = CalcSums(cols, cols_cache + local_id_x, winsz) * 255;
if (sum_win < threshold)
disp[y * disp_step + x] = 0;
if (x < disp_cols)
{
float sum_win = CalcSums(cols, cols_cache + local_id_x, winsz) * 255;
if (sum_win < threshold)
disp[y * disp_step + x] = 0;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
// }
}

View File

@@ -118,6 +118,8 @@ int main(int argc, char **argv)
setDevice(oclinfo[pid], device);
setBinaryDiskCache(CACHE_UPDATE);
cout << "Device type:" << type << endl << "Device name:" << oclinfo[pid].DeviceName[device] << endl;
return RUN_ALL_TESTS();
}

View File

@@ -146,17 +146,17 @@ TEST_P(HOG, Detect)
if (winSize.width == 48 && winSize.height == 96)
{
// daimler's base
ocl_hog.setSVMDetector(ocl_hog.getPeopleDetector48x96());
ocl_hog.setSVMDetector(hog.getDaimlerPeopleDetector());
hog.setSVMDetector(hog.getDaimlerPeopleDetector());
}
else if (winSize.width == 64 && winSize.height == 128)
{
ocl_hog.setSVMDetector(ocl_hog.getPeopleDetector64x128());
ocl_hog.setSVMDetector(hog.getDefaultPeopleDetector());
hog.setSVMDetector(hog.getDefaultPeopleDetector());
}
else
{
ocl_hog.setSVMDetector(ocl_hog.getDefaultPeopleDetector());
ocl_hog.setSVMDetector(hog.getDefaultPeopleDetector());
hog.setSVMDetector(hog.getDefaultPeopleDetector());
}