Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

optimized gpu pyrDown and sepFilter2D

Browse Source
pull/31/head^2
Vladislav Vinogradov 13 years ago
parent 7591ee1688
commit fd77ef7ece
4 changed files with 157 additions and 60 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 2
      modules/gpu/perf/perf_imgproc.cpp
  2. 49
      modules/gpu/src/cuda/column_filter.cu
  3. 117
      modules/gpu/src/cuda/pyr_down.cu
  4. 49
      modules/gpu/src/cuda/row_filter.cu

2
modules/gpu/perf/perf_imgproc.cpp
Unescape View File

@ -1720,7 +1720,7 @@ CV_FLAGS(GHMethod, cv::GHT_POSITION, cv::GHT_SCALE, cv::GHT_ROTATION);
DEF_PARAM_TEST(Method_Sz, GHMethod, cv::Size); DEF_PARAM_TEST(Method_Sz, GHMethod, cv::Size);
PERF_TEST_P(Method_Sz, GeneralizedHough, Combine( PERF_TEST_P(Method_Sz, ImgProc_GeneralizedHough, Combine(
Values(GHMethod(cv::GHT_POSITION), GHMethod(cv::GHT_POSITION | cv::GHT_SCALE), GHMethod(cv::GHT_POSITION | cv::GHT_ROTATION), GHMethod(cv::GHT_POSITION | cv::GHT_SCALE | cv::GHT_ROTATION)), Values(GHMethod(cv::GHT_POSITION), GHMethod(cv::GHT_POSITION | cv::GHT_SCALE), GHMethod(cv::GHT_POSITION | cv::GHT_ROTATION), GHMethod(cv::GHT_POSITION | cv::GHT_SCALE | cv::GHT_ROTATION)),
GPU_TYPICAL_MAT_SIZES)) GPU_TYPICAL_MAT_SIZES))
{ {

49
modules/gpu/src/cuda/column_filter.cu
Unescape View File

@ -89,20 +89,45 @@ namespace cv { namespace gpu { namespace device
const int yStart = blockIdx.y * (BLOCK_DIM_Y * PATCH_PER_BLOCK) + threadIdx.y; const int yStart = blockIdx.y * (BLOCK_DIM_Y * PATCH_PER_BLOCK) + threadIdx.y;
//Upper halo if (blockIdx.y > 0)
#pragma unroll {
for (int j = 0; j < HALO_SIZE; ++j) //Upper halo
smem[threadIdx.y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_low(yStart - (HALO_SIZE - j) * BLOCK_DIM_Y, src_col, src.step)); #pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart - (HALO_SIZE - j) * BLOCK_DIM_Y, x));
}
else
{
//Upper halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_low(yStart - (HALO_SIZE - j) * BLOCK_DIM_Y, src_col, src.step));
}
//Main data if (blockIdx.y + 2 < gridDim.y)
#pragma unroll {
for (int j = 0; j < PATCH_PER_BLOCK; ++j) //Main data
smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + j * BLOCK_DIM_Y, src_col, src.step)); #pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart + j * BLOCK_DIM_Y, x));
//Lower halo //Lower halo
#pragma unroll #pragma unroll
for (int j = 0; j < HALO_SIZE; ++j) for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_Y, src_col, src.step)); smem[threadIdx.y + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(src(yStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_Y, x));
}
else
{
//Main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y + HALO_SIZE * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + j * BLOCK_DIM_Y, src_col, src.step));
//Lower halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_Y + j * BLOCK_DIM_Y][threadIdx.x] = saturate_cast<sum_t>(brd.at_high(yStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_Y, src_col, src.step));
}
__syncthreads(); __syncthreads();

117
modules/gpu/src/cuda/pyr_down.cu
Unescape View File

@ -40,7 +40,7 @@
// //
//M*/ //M*/
#include "internal_shared.hpp" #include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp" #include "opencv2/gpu/device/border_interpolate.hpp"
#include "opencv2/gpu/device/vec_traits.hpp" #include "opencv2/gpu/device/vec_traits.hpp"
#include "opencv2/gpu/device/vec_math.hpp" #include "opencv2/gpu/device/vec_math.hpp"
@ -50,57 +50,104 @@ namespace cv { namespace gpu { namespace device
{ {
namespace imgproc namespace imgproc
{ {
template <typename T, typename B> __global__ void pyrDown(const PtrStep<T> src, PtrStep<T> dst, const B b, int dst_cols) template <typename T, typename B> __global__ void pyrDown(const PtrStepSz<T> src, PtrStep<T> dst, const B b, int dst_cols)
{ {
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type; typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_t;
__shared__ work_t smem[256 + 4];
const int x = blockIdx.x * blockDim.x + threadIdx.x; const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y; const int y = blockIdx.y;
__shared__ value_type smem[256 + 4]; const int src_y = 2 * y;
value_type sum; if (src_y >= 2 && src_y < src.rows - 2 && x >= 2 && x < src.cols - 2)
{
{
work_t sum;
const int src_y = 2*y; sum = 0.0625f * src(src_y - 2, x);
sum = sum + 0.25f * src(src_y - 1, x);
sum = sum + 0.375f * src(src_y , x);
sum = sum + 0.25f * src(src_y + 1, x);
sum = sum + 0.0625f * src(src_y + 2, x);
sum = VecTraits<value_type>::all(0); smem[2 + threadIdx.x] = sum;
}
sum = sum + 0.0625f * b.at(src_y - 2, x, src.data, src.step); if (threadIdx.x < 2)
sum = sum + 0.25f * b.at(src_y - 1, x, src.data, src.step); {
sum = sum + 0.375f * b.at(src_y , x, src.data, src.step); const int left_x = x - 2;
sum = sum + 0.25f * b.at(src_y + 1, x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, x, src.data, src.step);
smem[2 + threadIdx.x] = sum; work_t sum;
if (threadIdx.x < 2) sum = 0.0625f * src(src_y - 2, left_x);
{ sum = sum + 0.25f * src(src_y - 1, left_x);
const int left_x = x - 2; sum = sum + 0.375f * src(src_y , left_x);
sum = sum + 0.25f * src(src_y + 1, left_x);
sum = sum + 0.0625f * src(src_y + 2, left_x);
sum = VecTraits<value_type>::all(0); smem[threadIdx.x] = sum;
}
sum = sum + 0.0625f * b.at(src_y - 2, left_x, src.data, src.step); if (threadIdx.x > 253)
sum = sum + 0.25f * b.at(src_y - 1, left_x, src.data, src.step); {
sum = sum + 0.375f * b.at(src_y , left_x, src.data, src.step); const int right_x = x + 2;
sum = sum + 0.25f * b.at(src_y + 1, left_x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, left_x, src.data, src.step);
smem[threadIdx.x] = sum; work_t sum;
}
if (threadIdx.x > 253) sum = 0.0625f * src(src_y - 2, right_x);
sum = sum + 0.25f * src(src_y - 1, right_x);
sum = sum + 0.375f * src(src_y , right_x);
sum = sum + 0.25f * src(src_y + 1, right_x);
sum = sum + 0.0625f * src(src_y + 2, right_x);
smem[4 + threadIdx.x] = sum;
}
}
else
{ {
const int right_x = x + 2; {
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col_high(x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col_high(x));
sum = sum + 0.375f * src(src_y , b.idx_col_high(x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col_high(x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col_high(x));
smem[2 + threadIdx.x] = sum;
}
if (threadIdx.x < 2)
{
const int left_x = x - 2;
work_t sum;
sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col(left_x));
sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col(left_x));
sum = sum + 0.375f * src(src_y , b.idx_col(left_x));
sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col(left_x));
sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col(left_x));
smem[threadIdx.x] = sum;
}
if (threadIdx.x > 253)
{
const int right_x = x + 2;
sum = VecTraits<value_type>::all(0); work_t sum;
sum = sum + 0.0625f * b.at(src_y - 2, right_x, src.data, src.step); sum = 0.0625f * src(b.idx_row_low (src_y - 2), b.idx_col_high(right_x));
sum = sum + 0.25f * b.at(src_y - 1, right_x, src.data, src.step); sum = sum + 0.25f * src(b.idx_row_low (src_y - 1), b.idx_col_high(right_x));
sum = sum + 0.375f * b.at(src_y , right_x, src.data, src.step); sum = sum + 0.375f * src(src_y , b.idx_col_high(right_x));
sum = sum + 0.25f * b.at(src_y + 1, right_x, src.data, src.step); sum = sum + 0.25f * src(b.idx_row_high(src_y + 1), b.idx_col_high(right_x));
sum = sum + 0.0625f * b.at(src_y + 2, right_x, src.data, src.step); sum = sum + 0.0625f * src(b.idx_row_high(src_y + 2), b.idx_col_high(right_x));
smem[4 + threadIdx.x] = sum; smem[4 + threadIdx.x] = sum;
}
} }
__syncthreads(); __syncthreads();
@ -109,9 +156,9 @@ namespace cv { namespace gpu { namespace device
{ {
const int tid2 = threadIdx.x * 2; const int tid2 = threadIdx.x * 2;
sum = VecTraits<value_type>::all(0); work_t sum;
sum = sum + 0.0625f * smem[2 + tid2 - 2]; sum = 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1]; sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ]; sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1]; sum = sum + 0.25f * smem[2 + tid2 + 1];

49
modules/gpu/src/cuda/row_filter.cu
Unescape View File

@ -89,20 +89,45 @@ namespace cv { namespace gpu { namespace device
const int xStart = blockIdx.x * (PATCH_PER_BLOCK * BLOCK_DIM_X) + threadIdx.x; const int xStart = blockIdx.x * (PATCH_PER_BLOCK * BLOCK_DIM_X) + threadIdx.x;
//Load left halo if (blockIdx.x > 0)
#pragma unroll {
for (int j = 0; j < HALO_SIZE; ++j) //Load left halo
smem[threadIdx.y][threadIdx.x + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_low(xStart - (HALO_SIZE - j) * BLOCK_DIM_X, src_row)); #pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart - (HALO_SIZE - j) * BLOCK_DIM_X]);
}
else
{
//Load left halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_low(xStart - (HALO_SIZE - j) * BLOCK_DIM_X, src_row));
}
//Load main data if (blockIdx.x + 2 < gridDim.x)
#pragma unroll {
for (int j = 0; j < PATCH_PER_BLOCK; ++j) //Load main data
smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + j * BLOCK_DIM_X, src_row)); #pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart + j * BLOCK_DIM_X]);
//Load right halo //Load right halo
#pragma unroll #pragma unroll
for (int j = 0; j < HALO_SIZE; ++j) for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_X, src_row)); smem[threadIdx.y][threadIdx.x + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(src_row[xStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_X]);
}
else
{
//Load main data
#pragma unroll
for (int j = 0; j < PATCH_PER_BLOCK; ++j)
smem[threadIdx.y][threadIdx.x + HALO_SIZE * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + j * BLOCK_DIM_X, src_row));
//Load right halo
#pragma unroll
for (int j = 0; j < HALO_SIZE; ++j)
smem[threadIdx.y][threadIdx.x + (PATCH_PER_BLOCK + HALO_SIZE) * BLOCK_DIM_X + j * BLOCK_DIM_X] = saturate_cast<sum_t>(brd.at_high(xStart + (PATCH_PER_BLOCK + j) * BLOCK_DIM_X, src_row));
}
__syncthreads(); __syncthreads();

Write Preview
Loading…
Cancel
Save