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

optimizations:

Browse Source 
- new reduce implementation (with kepler optimizations)
- saturate_cast via asm command
- video SIMD instructions in element operations
- float arithmetics instead of double
- new deviceSupports function
This commit is contained in:
Vladislav Vinogradov
2013-01-23 14:43:36 +04:00
parent ae6266e101
commit 281d036fcf
45 changed files with 9379 additions and 8846 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
modules/gpu/src/nvidia/NCVHaarObjectDetection.cu
View File

@@ -65,6 +65,8 @@
#include "NPP_staging/NPP_staging.hpp"
#include "NCVRuntimeTemplates.hpp"
#include "NCVHaarObjectDetection.hpp"
#include "opencv2/gpu/device/warp.hpp"
#include "opencv2/gpu/device/warp_shuffle.hpp"
//==============================================================================
@@ -81,6 +83,20 @@ NCV_CT_ASSERT(K_WARP_SIZE == 32); //this is required for the manual unroll of th
//assuming size <= WARP_SIZE and size is power of 2
__device__ Ncv32u warpScanInclusive(Ncv32u idata, volatile Ncv32u *s_Data)
{
#if __CUDA_ARCH__ >= 300
const unsigned int laneId = cv::gpu::device::Warp::laneId();
// scan on shuffl functions
#pragma unroll
for (int i = 1; i <= (K_WARP_SIZE / 2); i *= 2)
{
const Ncv32u n = cv::gpu::device::shfl_up(idata, i);
if (laneId >= i)
idata += n;
}
return idata;
#else
Ncv32u pos = 2 * threadIdx.x - (threadIdx.x & (K_WARP_SIZE - 1));
s_Data[pos] = 0;
pos += K_WARP_SIZE;
@@ -93,6 +109,7 @@ __device__ Ncv32u warpScanInclusive(Ncv32u idata, volatile Ncv32u *s_Data)
s_Data[pos] += s_Data[pos - 16];
return s_Data[pos];
#endif
}
__device__ __forceinline__ Ncv32u warpScanExclusive(Ncv32u idata, volatile Ncv32u *s_Data)
@@ -2317,4 +2334,4 @@ NCVStatus ncvHaarStoreNVBIN_host(const std::string &filename,
return NCV_SUCCESS;
}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */

34
modules/gpu/src/nvidia/NPP_staging/NPP_staging.cu
View File

@@ -44,6 +44,8 @@
#include <vector>
#include <cuda_runtime.h>
#include "NPP_staging.hpp"
#include "opencv2/gpu/device/warp.hpp"
#include "opencv2/gpu/device/warp_shuffle.hpp"
texture<Ncv8u, 1, cudaReadModeElementType> tex8u;
@@ -90,6 +92,36 @@ NCV_CT_ASSERT(K_WARP_SIZE == 32); //this is required for the manual unroll of th
//assuming size <= WARP_SIZE and size is power of 2
template <class T>
inline __device__ T warpScanInclusive(T idata, volatile T *s_Data)
{
#if __CUDA_ARCH__ >= 300
const unsigned int laneId = cv::gpu::device::Warp::laneId();
// scan on shuffl functions
#pragma unroll
for (int i = 1; i <= (K_WARP_SIZE / 2); i *= 2)
{
const T n = cv::gpu::device::shfl_up(idata, i);
if (laneId >= i)
idata += n;
}
return idata;
#else
Ncv32u pos = 2 * threadIdx.x - (threadIdx.x & (K_WARP_SIZE - 1));
s_Data[pos] = 0;
pos += K_WARP_SIZE;
s_Data[pos] = idata;
s_Data[pos] += s_Data[pos - 1];
s_Data[pos] += s_Data[pos - 2];
s_Data[pos] += s_Data[pos - 4];
s_Data[pos] += s_Data[pos - 8];
s_Data[pos] += s_Data[pos - 16];
return s_Data[pos];
#endif
}
inline __device__ Ncv64u warpScanInclusive(Ncv64u idata, volatile Ncv64u *s_Data)
{
Ncv32u pos = 2 * threadIdx.x - (threadIdx.x & (K_WARP_SIZE - 1));
s_Data[pos] = 0;
@@ -2578,4 +2610,4 @@ NCVStatus nppiStResize_32f_C1R(const Ncv32f *pSrc,
return status;
}
#endif /* CUDA_DISABLER */
#endif /* CUDA_DISABLER */