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

minor SURF_GPU optimization (descriptor calculation, change block size to 6x6)

Browse Source
This commit is contained in:
Vladislav Vinogradov 2011-05-23 18:40:09 +00:00
parent 44e9fdaa2d
commit 7fbcc1ec05
1 changed files with 14 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
modules/gpu/src/cuda/surf.cu
View File

@ -770,18 +770,16 @@ namespace cv { namespace gpu { namespace surf
// since grids are 2D, need to compute xBlock and yBlock indices
const int xBlock = (blockIdx.y & 3); // blockIdx.y % 4
const int yBlock = (blockIdx.y >> 2); // floor(blockIdx.y/4)
const int xIndex = xBlock * blockDim.x + threadIdx.x;
const int yIndex = yBlock * blockDim.y + threadIdx.y;
const int xIndex = xBlock * 5 + threadIdx.x;
const int yIndex = yBlock * 5 + threadIdx.y;
s_PATCH[threadIdx.y][threadIdx.x] = calcPATCH(yIndex, xIndex, centerX, centerY, win_offset, cos_dir, sin_dir, win_size);
if (threadIdx.x == 0)
s_PATCH[threadIdx.y][5] = calcPATCH(yIndex, xBlock * blockDim.x + 5, centerX, centerY, win_offset, cos_dir, sin_dir, win_size);
if (threadIdx.y == 0)
s_PATCH[5][threadIdx.x] = calcPATCH(yBlock * blockDim.y + 5, xIndex, centerX, centerY, win_offset, cos_dir, sin_dir, win_size);
if (threadIdx.x == 0 && threadIdx.y == 0)
s_PATCH[5][5] = calcPATCH(yBlock * blockDim.y + 5, xBlock * blockDim.x + 5, centerX, centerY, win_offset, cos_dir, sin_dir, win_size);
__syncthreads();
if (threadIdx.x < 5 && threadIdx.y < 5)
{
tid = threadIdx.y * 5 + threadIdx.x;
const float dw = c_DW[yIndex * PATCH_SZ + xIndex];
const float vx = (s_PATCH[threadIdx.y ][threadIdx.x + 1] - s_PATCH[threadIdx.y][threadIdx.x] + s_PATCH[threadIdx.y + 1][threadIdx.x + 1] - s_PATCH[threadIdx.y + 1][threadIdx.x ]) * dw;
@ -790,6 +788,7 @@ namespace cv { namespace gpu { namespace surf
s_dx_bin[tid] = vx;
s_dy_bin[tid] = vy;
}
}
__device__ void reduce_sum25(volatile float* sdata1, volatile float* sdata2, volatile float* sdata3, volatile float* sdata4, int tid)
{
@ -986,7 +985,7 @@ namespace cv { namespace gpu { namespace surf
if (descriptors.cols == 64)
{
compute_descriptors64<<<dim3(nFeatures, 16, 1), dim3(5, 5, 1)>>>(descriptors, featureX, featureY, featureSize, featureDir);
compute_descriptors64<<<dim3(nFeatures, 16, 1), dim3(6, 6, 1)>>>(descriptors, featureX, featureY, featureSize, featureDir);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaThreadSynchronize() );
@ -998,7 +997,7 @@ namespace cv { namespace gpu { namespace surf
}
else
{
compute_descriptors128<<<dim3(nFeatures, 16, 1), dim3(5, 5, 1)>>>(descriptors, featureX, featureY, featureSize, featureDir);
compute_descriptors128<<<dim3(nFeatures, 16, 1), dim3(6, 6, 1)>>>(descriptors, featureX, featureY, featureSize, featureDir);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaThreadSynchronize() );