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

hog

Browse Source
This commit is contained in:
Vladislav Vinogradov 2012-11-12 13:19:48 +04:00
parent 0ddd16cf78
commit 0e339dd137
1 changed files with 60 additions and 119 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

109
modules/gpu/src/cuda/hog.cu
View File

@ -42,7 +42,10 @@
#if !defined CUDA_DISABLER
#include "internal_shared.hpp"
#include "opencv2/gpu/device/common.hpp"
#include "opencv2/gpu/device/reduce.hpp"
#include "opencv2/gpu/device/functional.hpp"
#include "opencv2/gpu/device/warp_shuffle.hpp"
namespace cv { namespace gpu { namespace device
{
@ -226,29 +229,30 @@ namespace cv { namespace gpu { namespace device
template<int size>
__device__ float reduce_smem(volatile float* smem)
__device__ float reduce_smem(float* smem, float val)
{
unsigned int tid = threadIdx.x;
float sum = smem[tid];
float sum = val;
if (size >= 512) { if (tid < 256) smem[tid] = sum = sum + smem[tid + 256]; __syncthreads(); }
if (size >= 256) { if (tid < 128) smem[tid] = sum = sum + smem[tid + 128]; __syncthreads(); }
if (size >= 128) { if (tid < 64) smem[tid] = sum = sum + smem[tid + 64]; __syncthreads(); }
reduce<size>(smem, sum, tid, plus<float>());
if (tid < 32)
if (size == 32)
{
if (size >= 64) smem[tid] = sum = sum + smem[tid + 32];
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];
if (size >= 4) smem[tid] = sum = sum + smem[tid + 2];
if (size >= 2) smem[tid] = sum = sum + smem[tid + 1];
#if __CUDA_ARCH__ >= 300
return shfl(sum, 0);
#else
return smem[0];
#endif
}
__syncthreads();
sum = smem[0];
#if __CUDA_ARCH__ >= 300
if (threadIdx.x == 0)
smem[0] = sum;
#endif
return sum;
__syncthreads();
return smem[0];
}
@ -272,19 +276,13 @@ namespace cv { namespace gpu { namespace device
if (threadIdx.x < block_hist_size)
elem = hist[0];
squares[threadIdx.x] = elem * elem;
__syncthreads();
float sum = reduce_smem<nthreads>(squares);
float sum = reduce_smem<nthreads>(squares, elem * elem);
float scale = 1.0f / (::sqrtf(sum) + 0.1f * block_hist_size);
elem = ::min(elem * scale, threshold);
__syncthreads();
squares[threadIdx.x] = elem * elem;
sum = reduce_smem<nthreads>(squares, elem * elem);
__syncthreads();
sum = reduce_smem<nthreads>(squares);
scale = 1.0f / (::sqrtf(sum) + 1e-3f);
if (threadIdx.x < block_hist_size)
@ -355,40 +353,11 @@ namespace cv { namespace gpu { namespace device
__shared__ float products[nthreads * nblocks];
const int tid = threadIdx.z * nthreads + threadIdx.x;
products[tid] = product;
__syncthreads();
if (nthreads >= 512)
{
if (threadIdx.x < 256) products[tid] = product = product + products[tid + 256];
__syncthreads();
}
if (nthreads >= 256)
{
if (threadIdx.x < 128) products[tid] = product = product + products[tid + 128];
__syncthreads();
}
if (nthreads >= 128)
{
if (threadIdx.x < 64) products[tid] = product = product + products[tid + 64];
__syncthreads();
}
if (threadIdx.x < 32)
{
volatile float* smem = products;
if (nthreads >= 64) smem[tid] = product = product + smem[tid + 32];
if (nthreads >= 32) smem[tid] = product = product + smem[tid + 16];
if (nthreads >= 16) smem[tid] = product = product + smem[tid + 8];
if (nthreads >= 8) smem[tid] = product = product + smem[tid + 4];
if (nthreads >= 4) smem[tid] = product = product + smem[tid + 2];
if (nthreads >= 2) smem[tid] = product = product + smem[tid + 1];
}
reduce<nthreads>(products, product, tid, plus<float>());
if (threadIdx.x == 0)
confidences[blockIdx.y * img_win_width + blockIdx.x * blockDim.z + win_x]
= (float)(product + free_coef);
confidences[blockIdx.y * img_win_width + blockIdx.x * blockDim.z + win_x] = product + free_coef;
}
@ -446,36 +415,8 @@ namespace cv { namespace gpu { namespace device
__shared__ float products[nthreads * nblocks];
const int tid = threadIdx.z * nthreads + threadIdx.x;
products[tid] = product;
__syncthreads();
if (nthreads >= 512)
{
if (threadIdx.x < 256) products[tid] = product = product + products[tid + 256];
__syncthreads();
}
if (nthreads >= 256)
{
if (threadIdx.x < 128) products[tid] = product = product + products[tid + 128];
__syncthreads();
}
if (nthreads >= 128)
{
if (threadIdx.x < 64) products[tid] = product = product + products[tid + 64];
__syncthreads();
}
if (threadIdx.x < 32)
{
volatile float* smem = products;
if (nthreads >= 64) smem[tid] = product = product + smem[tid + 32];
if (nthreads >= 32) smem[tid] = product = product + smem[tid + 16];
if (nthreads >= 16) smem[tid] = product = product + smem[tid + 8];
if (nthreads >= 8) smem[tid] = product = product + smem[tid + 4];
if (nthreads >= 4) smem[tid] = product = product + smem[tid + 2];
if (nthreads >= 2) smem[tid] = product = product + smem[tid + 1];
}
reduce<nthreads>(products, product, tid, plus<float>());
if (threadIdx.x == 0)
labels[blockIdx.y * img_win_width + blockIdx.x * blockDim.z + win_x] = (product + free_coef >= threshold);