Skip to content

O
opencv_contrib
Commit 04041fd7 authored by Namgoo Lee's avatar Namgoo Lee Committed by Namgoo Lee
Browse files
Options

[moved from opencv] cudalegacy: Use safe block scan function 

original commit: https://github.com/opencv/opencv/commit/21eb60f88bce4e9e4bedffcd28b29c75c404e788
parent 796853e0
Hide whitespace changes
Inline Side-by-side
Showing with 9 additions and 202 deletions
modules/cudalegacy/src/cuda/NCVHaarObjectDetection.cu
View file @ 04041fd7
......@@ -59,8 +59,7 @@
#include <algorithm>
#include <cstdio>
#include "opencv2/core/cuda/warp.hpp"
#include "opencv2/core/cuda/warp_shuffle.hpp"
#include "opencv2/cudev.hpp"
#include "opencv2/opencv_modules.hpp"
......@@ -77,92 +76,6 @@
#include "NCVAlg.hpp"
//==============================================================================
//
// BlockScan file
//
//==============================================================================
NCV_CT_ASSERT(K_WARP_SIZE == 32); //this is required for the manual unroll of the loop in warpScanInclusive
//Almost the same as naive scan1Inclusive, but doesn't need __syncthreads()
//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::cuda::device::Warp::laneId();
// scan on shuffl functions
#pragma unroll
for (int i = 1; i <= (K_WARP_SIZE / 2); i *= 2)
{
const Ncv32u n = cv::cuda::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
}
__device__ __forceinline__ Ncv32u warpScanExclusive(Ncv32u idata, volatile Ncv32u *s_Data)
{
return warpScanInclusive(idata, s_Data) - idata;
}
template <Ncv32u tiNumScanThreads>
__device__ Ncv32u scan1Inclusive(Ncv32u idata, volatile Ncv32u *s_Data)
{
if (tiNumScanThreads > K_WARP_SIZE)
{
//Bottom-level inclusive warp scan
Ncv32u warpResult = warpScanInclusive(idata, s_Data);
//Save top elements of each warp for exclusive warp scan
//sync to wait for warp scans to complete (because s_Data is being overwritten)
__syncthreads();
if( (threadIdx.x & (K_WARP_SIZE - 1)) == (K_WARP_SIZE - 1) )
{
s_Data[threadIdx.x >> K_LOG2_WARP_SIZE] = warpResult;
}
//wait for warp scans to complete
__syncthreads();
if( threadIdx.x < (tiNumScanThreads / K_WARP_SIZE) )
{
//grab top warp elements
Ncv32u val = s_Data[threadIdx.x];
//calculate exclusive scan and write back to shared memory
s_Data[threadIdx.x] = warpScanExclusive(val, s_Data);
}
//return updated warp scans with exclusive scan results
__syncthreads();
return warpResult + s_Data[threadIdx.x >> K_LOG2_WARP_SIZE];
}
else
{
return warpScanInclusive(idata, s_Data);
}
}
//==============================================================================
//
// HaarClassifierCascade file
......@@ -260,11 +173,11 @@ __device__ void compactBlockWriteOutAnchorParallel(Ncv32u threadPassFlag, Ncv32u
{
#if __CUDA_ARCH__ && __CUDA_ARCH__ >= 110
__shared__ Ncv32u shmem[NUM_THREADS_ANCHORSPARALLEL * 2];
__shared__ Ncv32u shmem[NUM_THREADS_ANCHORSPARALLEL];
__shared__ Ncv32u numPassed;
__shared__ Ncv32u outMaskOffset;
Ncv32u incScan = scan1Inclusive<NUM_THREADS_ANCHORSPARALLEL>(threadPassFlag, shmem);
Ncv32u incScan = cv::cudev::blockScanInclusive<NUM_THREADS_ANCHORSPARALLEL>(threadPassFlag, shmem, threadIdx.x);
__syncthreads();
if (threadIdx.x == NUM_THREADS_ANCHORSPARALLEL-1)
......
modules/cudalegacy/src/cuda/NPP_staging.cu
View file @ 04041fd7
......@@ -45,8 +45,7 @@
#include <vector>
#include <cuda_runtime.h>
#include "opencv2/core/cuda/warp.hpp"
#include "opencv2/core/cuda/warp_shuffle.hpp"
#include "opencv2/cudev.hpp"
#include "opencv2/cudalegacy/NPP_staging.hpp"
......@@ -81,111 +80,6 @@ cudaStream_t nppStSetActiveCUDAstream(cudaStream_t cudaStream)
}
//==============================================================================
//
// BlockScan.cuh
//
//==============================================================================
NCV_CT_ASSERT(K_WARP_SIZE == 32); //this is required for the manual unroll of the loop in warpScanInclusive
//Almost the same as naive scan1Inclusive, but doesn't need __syncthreads()
//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::cuda::device::Warp::laneId();
// scan on shuffl functions
#pragma unroll
for (int i = 1; i <= (K_WARP_SIZE / 2); i *= 2)
{
const T n = cv::cuda::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;
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];
}
template <class T>
inline __device__ T warpScanExclusive(T idata, volatile T *s_Data)
{
return warpScanInclusive(idata, s_Data) - idata;
}
template <class T, Ncv32u tiNumScanThreads>
inline __device__ T blockScanInclusive(T idata, volatile T *s_Data)
{
if (tiNumScanThreads > K_WARP_SIZE)
{
//Bottom-level inclusive warp scan
T warpResult = warpScanInclusive(idata, s_Data);
//Save top elements of each warp for exclusive warp scan
//sync to wait for warp scans to complete (because s_Data is being overwritten)
__syncthreads();
if( (threadIdx.x & (K_WARP_SIZE - 1)) == (K_WARP_SIZE - 1) )
{
s_Data[threadIdx.x >> K_LOG2_WARP_SIZE] = warpResult;
}
//wait for warp scans to complete
__syncthreads();
if( threadIdx.x < (tiNumScanThreads / K_WARP_SIZE) )
{
//grab top warp elements
T val = s_Data[threadIdx.x];
//calculate exclusive scan and write back to shared memory
s_Data[threadIdx.x] = warpScanExclusive(val, s_Data);
}
//return updated warp scans with exclusive scan results
__syncthreads();
return warpResult + s_Data[threadIdx.x >> K_LOG2_WARP_SIZE];
}
else
{
return warpScanInclusive(idata, s_Data);
}
}
//==============================================================================
//
// IntegralImage.cu
......@@ -280,7 +174,7 @@ __global__ void scanRows(T_in *d_src, Ncv32u texOffs, Ncv32u srcWidth, Ncv32u sr
Ncv32u numBuckets = (srcWidth + NUM_SCAN_THREADS - 1) >> LOG2_NUM_SCAN_THREADS;
Ncv32u offsetX = 0;
__shared__ T_out shmem[NUM_SCAN_THREADS * 2];
__shared__ T_out shmem[NUM_SCAN_THREADS];
__shared__ T_out carryElem;
carryElem = 0;
__syncthreads();
......@@ -301,7 +195,7 @@ __global__ void scanRows(T_in *d_src, Ncv32u texOffs, Ncv32u srcWidth, Ncv32u sr
curElemMod = _scanElemOp<T_in, T_out>::scanElemOp<tbDoSqr>(curElem);
//inclusive scan
curScanElem = blockScanInclusive<T_out, NUM_SCAN_THREADS>(curElemMod, shmem);
curScanElem = cv::cudev::blockScanInclusive<NUM_SCAN_THREADS>(curElemMod, shmem, threadIdx.x);
if (curElemOffs <= srcWidth)
{
......@@ -1290,7 +1184,7 @@ __global__ void removePass1Scan(Ncv32u *d_src, Ncv32u srcLen,
return;
}
__shared__ Ncv32u shmem[NUM_REMOVE_THREADS * 2];
__shared__ Ncv32u shmem[NUM_REMOVE_THREADS];
Ncv32u scanElem = 0;
if (elemAddrIn < srcLen)
......@@ -1305,7 +1199,7 @@ __global__ void removePass1Scan(Ncv32u *d_src, Ncv32u srcLen,
}
}
Ncv32u localScanInc = blockScanInclusive<Ncv32u, NUM_REMOVE_THREADS>(scanElem, shmem);
Ncv32u localScanInc = cv::cudev::blockScanInclusive<NUM_REMOVE_THREADS>(scanElem, shmem, threadIdx.x);
__syncthreads();
if (elemAddrIn < srcLen)
......
modules/cudev/include/opencv2/cudev/warp/scan.hpp
View file @ 04041fd7
......@@ -98,7 +98,7 @@ __device__ T warpScanInclusive(T data, volatile T* smem, uint tid)
#pragma unroll
for (int i = 1; i <= (WARP_SIZE / 2); i *= 2)
{
const T val = shfl_up(data, i);
const T val = __shfl_up(data, i, WARP_SIZE);
if (laneId >= i)
data += val;
}
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment