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

removed BEGIN_OPENCV_DEVICE_NAMESPACE macros

Browse Source
This commit is contained in:
Vladislav Vinogradov
2011-11-14 09:02:06 +00:00
parent d926541311
commit 0f53f2993e
73 changed files with 19272 additions and 19504 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1972
modules/gpu/src/cuda/bf_knnmatch.cu
View File

File diff suppressed because it is too large Load Diff

1312
modules/gpu/src/cuda/bf_match.cu
View File

File diff suppressed because it is too large Load Diff

748
modules/gpu/src/cuda/bf_radius_match.cu
View File

@@ -45,423 +45,421 @@
#include "opencv2/gpu/device/vec_distance.hpp"
#include "opencv2/gpu/device/datamov_utils.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace bf_radius_match {
///////////////////////////////////////////////////////////////////////////////
// Match Unrolled
template <int BLOCK_SIZE, int MAX_DESC_LEN, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void matchUnrolled(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
namespace cv { namespace gpu { namespace device
{
#if __CUDA_ARCH__ >= 110
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
#pragma unroll
for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; ++i)
namespace bf_radius_match
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
///////////////////////////////////////////////////////////////////////////////
// Match Unrolled
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
template <int BLOCK_SIZE, int MAX_DESC_LEN, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void matchUnrolled(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
T val;
#if __CUDA_ARCH__ >= 110
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
extern __shared__ int smem[];
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
#pragma unroll
for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T, typename Mask>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, cudaStream_t stream)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
#endif
}
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T, typename Mask>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
template <int BLOCK_SIZE, int MAX_DESC_LEN, typename Dist, typename T>
void matchUnrolled(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
///////////////////////////////////////////////////////////////////////////////
// Match
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
template <int BLOCK_SIZE, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void match(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
#if __CUDA_ARCH__ >= 110
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
extern __shared__ int smem[];
if (masks != 0 && masks[i].data)
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
matchUnrolled<BLOCK_SIZE, MAX_DESC_LEN, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
for (int i = 0, endi = (query.cols + BLOCK_SIZE - 1) / BLOCK_SIZE; i < endi; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
__syncthreads();
///////////////////////////////////////////////////////////////////////////////
// Match
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
template <int BLOCK_SIZE, bool SAVE_IMG_IDX, typename Dist, typename T, typename Mask>
__global__ void match(const DevMem2D_<T> query, int imgIdx, const DevMem2D_<T> train, float maxDistance, const Mask mask,
PtrStepi bestTrainIdx, PtrStepi bestImgIdx, PtrStepf bestDistance, unsigned int* nMatches, int maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
#if __CUDA_ARCH__ >= 110
extern __shared__ int smem[];
const int queryIdx = blockIdx.y * BLOCK_SIZE + threadIdx.y;
const int trainIdx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
typename Dist::value_type* s_query = (typename Dist::value_type*)(smem);
typename Dist::value_type* s_train = (typename Dist::value_type*)(smem + BLOCK_SIZE * BLOCK_SIZE);
Dist dist;
for (int i = 0, endi = (query.cols + BLOCK_SIZE - 1) / BLOCK_SIZE; i < endi; ++i)
{
const int loadX = threadIdx.x + i * BLOCK_SIZE;
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = 0;
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = 0;
if (loadX < query.cols)
{
T val;
ForceGlob<T>::Load(query.ptr(::min(queryIdx, query.rows - 1)), loadX, val);
s_query[threadIdx.y * BLOCK_SIZE + threadIdx.x] = val;
ForceGlob<T>::Load(train.ptr(::min(blockIdx.x * BLOCK_SIZE + threadIdx.y, train.rows - 1)), loadX, val);
s_train[threadIdx.x * BLOCK_SIZE + threadIdx.y] = val;
}
__syncthreads();
#pragma unroll
for (int j = 0; j < BLOCK_SIZE; ++j)
dist.reduceIter(s_query[threadIdx.y * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + threadIdx.x]);
__syncthreads();
}
float distVal = (typename Dist::result_type)dist;
if (queryIdx < query.rows && trainIdx < train.rows && mask(queryIdx, trainIdx) && distVal < maxDistance)
{
unsigned int ind = atomicInc(nMatches + queryIdx, (unsigned int) -1);
if (ind < maxCount)
{
bestTrainIdx.ptr(queryIdx)[ind] = trainIdx;
if (SAVE_IMG_IDX) bestImgIdx.ptr(queryIdx)[ind] = imgIdx;
bestDistance.ptr(queryIdx)[ind] = distVal;
}
}
#endif
}
}
#endif
}
template <int BLOCK_SIZE, typename Dist, typename T, typename Mask>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
match<BLOCK_SIZE, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <int BLOCK_SIZE, typename Dist, typename T>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
template <int BLOCK_SIZE, typename Dist, typename T, typename Mask>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
match<BLOCK_SIZE, false, Dist><<<grid, block, smemSize, stream>>>(query, 0, train, maxDistance, mask,
trainIdx, PtrStepi(), distance, nMatches.data, trainIdx.cols);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
else
template <int BLOCK_SIZE, typename Dist, typename T>
void match(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
cudaStream_t stream)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
const dim3 block(BLOCK_SIZE, BLOCK_SIZE);
const size_t smemSize = (2 * BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
for (int i = 0; i < n; ++i)
{
const DevMem2D_<T> train = trains[i];
const dim3 grid(divUp(train.rows, BLOCK_SIZE), divUp(query.rows, BLOCK_SIZE));
if (masks != 0 && masks[i].data)
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, SingleMask(masks[i]),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
else
{
match<BLOCK_SIZE, true, Dist><<<grid, block, smemSize, stream>>>(query, i, train, maxDistance, WithOutMask(),
trainIdx, imgIdx, distance, nMatches.data, trainIdx.cols);
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
cudaSafeCall( cudaGetLastError() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////////////////////
// Match dispatcher
///////////////////////////////////////////////////////////////////////////////
// Match dispatcher
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
}
template <typename Dist, typename T, typename Mask>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>& train, float maxDistance, const Mask& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, train, maxDistance, mask, trainIdx, distance, nMatches, stream);
}
}
template <typename Dist, typename T>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
}
template <typename Dist, typename T>
void matchDispatcher(const DevMem2D_<T>& query, const DevMem2D_<T>* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (query.cols <= 64)
{
matchUnrolled<16, 64, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 128)
{
matchUnrolled<16, 128, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
/*else if (query.cols <= 256)
{
matchUnrolled<16, 256, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 512)
{
matchUnrolled<16, 512, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
else if (query.cols <= 1024)
{
matchUnrolled<16, 1024, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}*/
else
{
match<16, Dist>(query, trains, n, maxDistance, masks, trainIdx, imgIdx, distance, nMatches, stream);
}
}
///////////////////////////////////////////////////////////////////////////////
// Radius Match caller
///////////////////////////////////////////////////////////////////////////////
// Radius Match caller
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchL1_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
//template void matchL2_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db& train, float maxDistance, const DevMem2Db& mask,
const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
if (mask.data)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, SingleMask(mask),
trainIdx, distance, nMatches,
cc, stream);
}
else
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), static_cast< DevMem2D_<T> >(train), maxDistance, WithOutMask(),
trainIdx, distance, nMatches,
cc, stream);
}
}
template void matchHamming_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<uchar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& queryDescs, const DevMem2Db& trainDescs, float maxDistance, const DevMem2Db& mask, const DevMem2Di& trainIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template <typename T> void matchL1_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher< L1Dist<T> >(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchL1_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL1_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL1_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template <typename T> void matchL2_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<L2Dist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
//template void matchL2_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchL2_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchL2_gpu<float >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template <typename T> void matchHamming_gpu(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks,
const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches,
int cc, cudaStream_t stream)
{
matchDispatcher<HammingDist>(static_cast< DevMem2D_<T> >(query), (const DevMem2D_<T>*)trains, n, maxDistance, masks,
trainIdx, imgIdx, distance, nMatches,
cc, stream);
}
template void matchHamming_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
} // namespace bf_radius_match
END_OPENCV_DEVICE_NAMESPACE
template void matchHamming_gpu<uchar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<schar >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<ushort>(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
//template void matchHamming_gpu<short >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
template void matchHamming_gpu<int >(const DevMem2Db& query, const DevMem2Db* trains, int n, float maxDistance, const DevMem2Db* masks, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance, const DevMem2D_<unsigned int>& nMatches, int cc, cudaStream_t stream);
} // namespace bf_radius_match
}}} // namespace cv { namespace gpu { namespace device

312
modules/gpu/src/cuda/bilateral_filter.cu
View File

@@ -43,186 +43,184 @@
#include "internal_shared.hpp"
#include "opencv2/gpu/device/limits.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace bilateral_filter {
__constant__ float* ctable_color;
__constant__ float* ctable_space;
__constant__ size_t ctable_space_step;
__constant__ int cndisp;
__constant__ int cradius;
__constant__ short cedge_disc;
__constant__ short cmax_disc;
void load_constants(float* table_color, DevMem2Df table_space, int ndisp, int radius, short edge_disc, short max_disc)
namespace cv { namespace gpu { namespace device
{
cudaSafeCall( cudaMemcpyToSymbol(ctable_color, &table_color, sizeof(table_color)) );
cudaSafeCall( cudaMemcpyToSymbol(ctable_space, &table_space.data, sizeof(table_space.data)) );
size_t table_space_step = table_space.step / sizeof(float);
cudaSafeCall( cudaMemcpyToSymbol(ctable_space_step, &table_space_step, sizeof(size_t)) );
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int)) );
cudaSafeCall( cudaMemcpyToSymbol(cradius, &radius, sizeof(int)) );
cudaSafeCall( cudaMemcpyToSymbol(cedge_disc, &edge_disc, sizeof(short)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc, &max_disc, sizeof(short)) );
}
template <int channels>
struct DistRgbMax
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
namespace bilateral_filter
{
uchar x = ::abs(a[0] - b[0]);
uchar y = ::abs(a[1] - b[1]);
uchar z = ::abs(a[2] - b[2]);
return (::max(::max(x, y), z));
}
};
__constant__ float* ctable_color;
__constant__ float* ctable_space;
__constant__ size_t ctable_space_step;
template <>
struct DistRgbMax<1>
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
{
return ::abs(a[0] - b[0]);
}
};
__constant__ int cndisp;
__constant__ int cradius;
template <int channels, typename T>
__global__ void bilateral_filter(int t, T* disp, size_t disp_step, const uchar* img, size_t img_step, int h, int w)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
__constant__ short cedge_disc;
__constant__ short cmax_disc;
T dp[5];
if (y > 0 && y < h - 1 && x > 0 && x < w - 1)
{
dp[0] = *(disp + (y ) * disp_step + x + 0);
dp[1] = *(disp + (y-1) * disp_step + x + 0);
dp[2] = *(disp + (y ) * disp_step + x - 1);
dp[3] = *(disp + (y+1) * disp_step + x + 0);
dp[4] = *(disp + (y ) * disp_step + x + 1);
if(::abs(dp[1] - dp[0]) >= cedge_disc || ::abs(dp[2] - dp[0]) >= cedge_disc || ::abs(dp[3] - dp[0]) >= cedge_disc || ::abs(dp[4] - dp[0]) >= cedge_disc)
void load_constants(float* table_color, DevMem2Df table_space, int ndisp, int radius, short edge_disc, short max_disc)
{
const int ymin = ::max(0, y - cradius);
const int xmin = ::max(0, x - cradius);
const int ymax = ::min(h - 1, y + cradius);
const int xmax = ::min(w - 1, x + cradius);
cudaSafeCall( cudaMemcpyToSymbol(ctable_color, &table_color, sizeof(table_color)) );
cudaSafeCall( cudaMemcpyToSymbol(ctable_space, &table_space.data, sizeof(table_space.data)) );
size_t table_space_step = table_space.step / sizeof(float);
cudaSafeCall( cudaMemcpyToSymbol(ctable_space_step, &table_space_step, sizeof(size_t)) );
float cost[] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int)) );
cudaSafeCall( cudaMemcpyToSymbol(cradius, &radius, sizeof(int)) );
const uchar* ic = img + y * img_step + channels * x;
cudaSafeCall( cudaMemcpyToSymbol(cedge_disc, &edge_disc, sizeof(short)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc, &max_disc, sizeof(short)) );
}
for(int yi = ymin; yi <= ymax; yi++)
template <int channels>
struct DistRgbMax
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
{
const T* disp_y = disp + yi * disp_step;
uchar x = ::abs(a[0] - b[0]);
uchar y = ::abs(a[1] - b[1]);
uchar z = ::abs(a[2] - b[2]);
return (::max(::max(x, y), z));
}
};
for(int xi = xmin; xi <= xmax; xi++)
template <>
struct DistRgbMax<1>
{
static __device__ __forceinline__ uchar calc(const uchar* a, const uchar* b)
{
return ::abs(a[0] - b[0]);
}
};
template <int channels, typename T>
__global__ void bilateral_filter(int t, T* disp, size_t disp_step, const uchar* img, size_t img_step, int h, int w)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
T dp[5];
if (y > 0 && y < h - 1 && x > 0 && x < w - 1)
{
dp[0] = *(disp + (y ) * disp_step + x + 0);
dp[1] = *(disp + (y-1) * disp_step + x + 0);
dp[2] = *(disp + (y ) * disp_step + x - 1);
dp[3] = *(disp + (y+1) * disp_step + x + 0);
dp[4] = *(disp + (y ) * disp_step + x + 1);
if(::abs(dp[1] - dp[0]) >= cedge_disc || ::abs(dp[2] - dp[0]) >= cedge_disc || ::abs(dp[3] - dp[0]) >= cedge_disc || ::abs(dp[4] - dp[0]) >= cedge_disc)
{
const uchar* in = img + yi * img_step + channels * xi;
const int ymin = ::max(0, y - cradius);
const int xmin = ::max(0, x - cradius);
const int ymax = ::min(h - 1, y + cradius);
const int xmax = ::min(w - 1, x + cradius);
uchar dist_rgb = DistRgbMax<channels>::calc(in, ic);
float cost[] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
const float weight = ctable_color[dist_rgb] * (ctable_space + ::abs(y-yi)* ctable_space_step)[::abs(x-xi)];
const uchar* ic = img + y * img_step + channels * x;
const T disp_reg = disp_y[xi];
for(int yi = ymin; yi <= ymax; yi++)
{
const T* disp_y = disp + yi * disp_step;
cost[0] += ::min(cmax_disc, ::abs(disp_reg - dp[0])) * weight;
cost[1] += ::min(cmax_disc, ::abs(disp_reg - dp[1])) * weight;
cost[2] += ::min(cmax_disc, ::abs(disp_reg - dp[2])) * weight;
cost[3] += ::min(cmax_disc, ::abs(disp_reg - dp[3])) * weight;
cost[4] += ::min(cmax_disc, ::abs(disp_reg - dp[4])) * weight;
for(int xi = xmin; xi <= xmax; xi++)
{
const uchar* in = img + yi * img_step + channels * xi;
uchar dist_rgb = DistRgbMax<channels>::calc(in, ic);
const float weight = ctable_color[dist_rgb] * (ctable_space + ::abs(y-yi)* ctable_space_step)[::abs(x-xi)];
const T disp_reg = disp_y[xi];
cost[0] += ::min(cmax_disc, ::abs(disp_reg - dp[0])) * weight;
cost[1] += ::min(cmax_disc, ::abs(disp_reg - dp[1])) * weight;
cost[2] += ::min(cmax_disc, ::abs(disp_reg - dp[2])) * weight;
cost[3] += ::min(cmax_disc, ::abs(disp_reg - dp[3])) * weight;
cost[4] += ::min(cmax_disc, ::abs(disp_reg - dp[4])) * weight;
}
}
float minimum = numeric_limits<float>::max();
int id = 0;
if (cost[0] < minimum)
{
minimum = cost[0];
id = 0;
}
if (cost[1] < minimum)
{
minimum = cost[1];
id = 1;
}
if (cost[2] < minimum)
{
minimum = cost[2];
id = 2;
}
if (cost[3] < minimum)
{
minimum = cost[3];
id = 3;
}
if (cost[4] < minimum)
{
minimum = cost[4];
id = 4;
}
*(disp + y * disp_step + x) = dp[id];
}
}
float minimum = numeric_limits<float>::max();
int id = 0;
if (cost[0] < minimum)
{
minimum = cost[0];
id = 0;
}
if (cost[1] < minimum)
{
minimum = cost[1];
id = 1;
}
if (cost[2] < minimum)
{
minimum = cost[2];
id = 2;
}
if (cost[3] < minimum)
{
minimum = cost[3];
id = 3;
}
if (cost[4] < minimum)
{
minimum = cost[4];
id = 4;
}
*(disp + y * disp_step + x) = dp[id];
}
}
}
template <typename T>
void bilateral_filter_caller(DevMem2D_<T> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(disp.cols, threads.x << 1);
grid.y = divUp(disp.rows, threads.y);
switch (channels)
{
case 1:
for (int i = 0; i < iters; ++i)
template <typename T>
void bilateral_filter_caller(DevMem2D_<T> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter<1><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(disp.cols, threads.x << 1);
grid.y = divUp(disp.rows, threads.y);
bilateral_filter<1><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
switch (channels)
{
case 1:
for (int i = 0; i < iters; ++i)
{
bilateral_filter<1><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter<1><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
}
break;
case 3:
for (int i = 0; i < iters; ++i)
{
bilateral_filter<3><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter<3><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
}
break;
default:
cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
if (stream != 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
break;
case 3:
for (int i = 0; i < iters; ++i)
void bilateral_filter_gpu(DevMem2Db disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter<3><<<grid, threads, 0, stream>>>(0, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter<3><<<grid, threads, 0, stream>>>(1, disp.data, disp.step/sizeof(T), img.data, img.step, disp.rows, disp.cols);
cudaSafeCall( cudaGetLastError() );
bilateral_filter_caller(disp, img, channels, iters, stream);
}
break;
default:
cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
if (stream != 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void bilateral_filter_gpu(DevMem2Db disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter_caller(disp, img, channels, iters, stream);
}
void bilateral_filter_gpu(DevMem2D_<short> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter_caller(disp, img, channels, iters, stream);
}
} // namespace bilateral_filter
END_OPENCV_DEVICE_NAMESPACE
void bilateral_filter_gpu(DevMem2D_<short> disp, DevMem2Db img, int channels, int iters, cudaStream_t stream)
{
bilateral_filter_caller(disp, img, channels, iters, stream);
}
} // namespace bilateral_filter
}}} // namespace cv { namespace gpu { namespace device

126
modules/gpu/src/cuda/blend.cu
View File

@@ -42,77 +42,75 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace blend {
template <typename T>
__global__ void blendLinearKernel(int rows, int cols, int cn, const PtrStep<T> img1, const PtrStep<T> img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStep<T> result)
namespace cv { namespace gpu { namespace device
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y < rows && x < cols)
namespace blend
{
int x_ = x / cn;
float w1 = weights1.ptr(y)[x_];
float w2 = weights2.ptr(y)[x_];
T p1 = img1.ptr(y)[x];
T p2 = img2.ptr(y)[x];
result.ptr(y)[x] = (p1 * w1 + p2 * w2) / (w1 + w2 + 1e-5f);
}
}
template <typename T>
__global__ void blendLinearKernel(int rows, int cols, int cn, const PtrStep<T> img1, const PtrStep<T> img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStep<T> result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols * cn, threads.x), divUp(rows, threads.y));
blendLinearKernel<<<grid, threads, 0, stream>>>(rows, cols * cn, cn, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
if (y < rows && x < cols)
{
int x_ = x / cn;
float w1 = weights1.ptr(y)[x_];
float w2 = weights2.ptr(y)[x_];
T p1 = img1.ptr(y)[x];
T p2 = img2.ptr(y)[x];
result.ptr(y)[x] = (p1 * w1 + p2 * w2) / (w1 + w2 + 1e-5f);
}
}
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
void blendLinearCaller(int rows, int cols, int cn, PtrStep<T> img1, PtrStep<T> img2, PtrStepf weights1, PtrStepf weights2, PtrStep<T> result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols * cn, threads.x), divUp(rows, threads.y));
blendLinearKernel<<<grid, threads, 0, stream>>>(rows, cols * cn, cn, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
template void blendLinearCaller<uchar>(int, int, int, PtrStep<uchar>, PtrStep<uchar>, PtrStepf, PtrStepf, PtrStep<uchar>, cudaStream_t stream);
template void blendLinearCaller<float>(int, int, int, PtrStep<float>, PtrStep<float>, PtrStepf, PtrStepf, PtrStep<float>, cudaStream_t stream);
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template void blendLinearCaller<uchar>(int, int, int, PtrStep<uchar>, PtrStep<uchar>, PtrStepf, PtrStepf, PtrStep<uchar>, cudaStream_t stream);
template void blendLinearCaller<float>(int, int, int, PtrStep<float>, PtrStep<float>, PtrStepf, PtrStepf, PtrStep<float>, cudaStream_t stream);
__global__ void blendLinearKernel8UC4(int rows, int cols, const PtrStepb img1, const PtrStepb img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStepb result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
__global__ void blendLinearKernel8UC4(int rows, int cols, const PtrStepb img1, const PtrStepb img2,
const PtrStepf weights1, const PtrStepf weights2, PtrStepb result)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y < rows && x < cols)
{
float w1 = weights1.ptr(y)[x];
float w2 = weights2.ptr(y)[x];
float sum_inv = 1.f / (w1 + w2 + 1e-5f);
w1 *= sum_inv;
w2 *= sum_inv;
uchar4 p1 = ((const uchar4*)img1.ptr(y))[x];
uchar4 p2 = ((const uchar4*)img2.ptr(y))[x];
((uchar4*)result.ptr(y))[x] = make_uchar4(p1.x * w1 + p2.x * w2, p1.y * w1 + p2.y * w2,
p1.z * w1 + p2.z * w2, p1.w * w1 + p2.w * w2);
}
}
if (y < rows && x < cols)
{
float w1 = weights1.ptr(y)[x];
float w2 = weights2.ptr(y)[x];
float sum_inv = 1.f / (w1 + w2 + 1e-5f);
w1 *= sum_inv;
w2 *= sum_inv;
uchar4 p1 = ((const uchar4*)img1.ptr(y))[x];
uchar4 p2 = ((const uchar4*)img2.ptr(y))[x];
((uchar4*)result.ptr(y))[x] = make_uchar4(p1.x * w1 + p2.x * w2, p1.y * w1 + p2.y * w2,
p1.z * w1 + p2.z * w2, p1.w * w1 + p2.w * w2);
}
}
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
blendLinearKernel8UC4<<<grid, threads, 0, stream>>>(rows, cols, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
void blendLinearCaller8UC4(int rows, int cols, PtrStepb img1, PtrStepb img2, PtrStepf weights1, PtrStepf weights2, PtrStepb result, cudaStream_t stream)
{
dim3 threads(16, 16);
dim3 grid(divUp(cols, threads.x), divUp(rows, threads.y));
blendLinearKernel8UC4<<<grid, threads, 0, stream>>>(rows, cols, img1, img2, weights1, weights2, result);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
} // namespace blend
END_OPENCV_DEVICE_NAMESPACE
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
} // namespace blend
}}} // namespace cv { namespace gpu { namespace device

245
modules/gpu/src/cuda/calib3d.cu
View File

@@ -44,149 +44,148 @@
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/functional.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#define SOLVE_PNP_RANSAC_MAX_NUM_ITERS 200
namespace transform_points
namespace cv { namespace gpu { namespace device
{
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
#define SOLVE_PNP_RANSAC_MAX_NUM_ITERS 200
struct TransformOp : unary_function<float3, float3>
namespace transform_points
{
__device__ __forceinline__ float3 operator()(const float3& p) const
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
struct TransformOp : unary_function<float3, float3>
{
return make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
crot1.x * p.x + crot1.y * p.y + crot1.z * p.z + ctransl.y,
crot2.x * p.x + crot2.y * p.y + crot2.z * p.z + ctransl.z);
__device__ __forceinline__ float3 operator()(const float3& p) const
{
return make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
crot1.x * p.x + crot1.y * p.y + crot1.z * p.z + ctransl.y,
crot2.x * p.x + crot2.y * p.y + crot2.z * p.z + ctransl.z);
}
};
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, DevMem2D_<float3> dst,
cudaStream_t stream)
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
::cv::gpu::device::transform(src, dst, TransformOp(), stream);
}
};
} // namespace transform_points
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, DevMem2D_<float3> dst,
cudaStream_t stream)
namespace project_points
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
OPENCV_DEVICE_NAMESPACE_ transform(src, dst, TransformOp(), stream);
}
} // namespace transform_points
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
__constant__ float3 cproj0;
__constant__ float3 cproj1;
namespace project_points
{
__constant__ float3 crot0;
__constant__ float3 crot1;
__constant__ float3 crot2;
__constant__ float3 ctransl;
__constant__ float3 cproj0;
__constant__ float3 cproj1;
struct ProjectOp : unary_function<float3, float3>
{
__device__ __forceinline__ float2 operator()(const float3& p) const
struct ProjectOp : unary_function<float3, float3>
{
// Rotate and translate in 3D
float3 t = make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
crot1.x * p.x + crot1.y * p.y + crot1.z * p.z + ctransl.y,
crot2.x * p.x + crot2.y * p.y + crot2.z * p.z + ctransl.z);
// Project on 2D plane
return make_float2(
(cproj0.x * t.x + cproj0.y * t.y) / t.z + cproj0.z,
(cproj1.x * t.x + cproj1.y * t.y) / t.z + cproj1.z);
__device__ __forceinline__ float2 operator()(const float3& p) const
{
// Rotate and translate in 3D
float3 t = make_float3(
crot0.x * p.x + crot0.y * p.y + crot0.z * p.z + ctransl.x,
crot1.x * p.x + crot1.y * p.y + crot1.z * p.z + ctransl.y,
crot2.x * p.x + crot2.y * p.y + crot2.z * p.z + ctransl.z);
// Project on 2D plane
return make_float2(
(cproj0.x * t.x + cproj0.y * t.y) / t.z + cproj0.z,
(cproj1.x * t.x + cproj1.y * t.y) / t.z + cproj1.z);
}
};
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, const float* proj, DevMem2D_<float2> dst,
cudaStream_t stream)
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj0, proj, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj1, proj + 3, sizeof(float) * 3));
::cv::gpu::device::transform(src, dst, ProjectOp(), stream);
}
};
} // namespace project_points
void call(const DevMem2D_<float3> src, const float* rot,
const float* transl, const float* proj, DevMem2D_<float2> dst,
cudaStream_t stream)
namespace solve_pnp_ransac
{
cudaSafeCall(cudaMemcpyToSymbol(crot0, rot, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot1, rot + 3, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(crot2, rot + 6, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(ctransl, transl, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj0, proj, sizeof(float) * 3));
cudaSafeCall(cudaMemcpyToSymbol(cproj1, proj + 3, sizeof(float) * 3));
OPENCV_DEVICE_NAMESPACE_ transform(src, dst, ProjectOp(), stream);
}
} // namespace project_points
__constant__ float3 crot_matrices[SOLVE_PNP_RANSAC_MAX_NUM_ITERS * 3];
__constant__ float3 ctransl_vectors[SOLVE_PNP_RANSAC_MAX_NUM_ITERS];
namespace solve_pnp_ransac
{
__constant__ float3 crot_matrices[SOLVE_PNP_RANSAC_MAX_NUM_ITERS * 3];
__constant__ float3 ctransl_vectors[SOLVE_PNP_RANSAC_MAX_NUM_ITERS];
int maxNumIters()
{
return SOLVE_PNP_RANSAC_MAX_NUM_ITERS;
}
__device__ __forceinline__ float sqr(float x)
{
return x * x;
}
__global__ void computeHypothesisScoresKernel(
const int num_points, const float3* object, const float2* image,
const float dist_threshold, int* g_num_inliers)
{
const float3* const &rot_mat = crot_matrices + blockIdx.x * 3;
const float3 &transl_vec = ctransl_vectors[blockIdx.x];
int num_inliers = 0;
for (int i = threadIdx.x; i < num_points; i += blockDim.x)
int maxNumIters()
{
float3 p = object[i];
p = make_float3(
rot_mat[0].x * p.x + rot_mat[0].y * p.y + rot_mat[0].z * p.z + transl_vec.x,
rot_mat[1].x * p.x + rot_mat[1].y * p.y + rot_mat[1].z * p.z + transl_vec.y,
rot_mat[2].x * p.x + rot_mat[2].y * p.y + rot_mat[2].z * p.z + transl_vec.z);
p.x /= p.z;
p.y /= p.z;
float2 image_p = image[i];
if (sqr(p.x - image_p.x) + sqr(p.y - image_p.y) < dist_threshold)
++num_inliers;
return SOLVE_PNP_RANSAC_MAX_NUM_ITERS;
}
extern __shared__ float s_num_inliers[];
s_num_inliers[threadIdx.x] = num_inliers;
__syncthreads();
for (int step = blockDim.x / 2; step > 0; step >>= 1)
__device__ __forceinline__ float sqr(float x)
{
if (threadIdx.x < step)
s_num_inliers[threadIdx.x] += s_num_inliers[threadIdx.x + step];
return x * x;
}
__global__ void computeHypothesisScoresKernel(
const int num_points, const float3* object, const float2* image,
const float dist_threshold, int* g_num_inliers)
{
const float3* const &rot_mat = crot_matrices + blockIdx.x * 3;
const float3 &transl_vec = ctransl_vectors[blockIdx.x];
int num_inliers = 0;
for (int i = threadIdx.x; i < num_points; i += blockDim.x)
{
float3 p = object[i];
p = make_float3(
rot_mat[0].x * p.x + rot_mat[0].y * p.y + rot_mat[0].z * p.z + transl_vec.x,
rot_mat[1].x * p.x + rot_mat[1].y * p.y + rot_mat[1].z * p.z + transl_vec.y,
rot_mat[2].x * p.x + rot_mat[2].y * p.y + rot_mat[2].z * p.z + transl_vec.z);
p.x /= p.z;
p.y /= p.z;
float2 image_p = image[i];
if (sqr(p.x - image_p.x) + sqr(p.y - image_p.y) < dist_threshold)
++num_inliers;
}
extern __shared__ float s_num_inliers[];
s_num_inliers[threadIdx.x] = num_inliers;
__syncthreads();
for (int step = blockDim.x / 2; step > 0; step >>= 1)
{
if (threadIdx.x < step)
s_num_inliers[threadIdx.x] += s_num_inliers[threadIdx.x + step];
__syncthreads();
}
if (threadIdx.x == 0)
g_num_inliers[blockIdx.x] = s_num_inliers[0];
}
if (threadIdx.x == 0)
g_num_inliers[blockIdx.x] = s_num_inliers[0];
}
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores)
{
cudaSafeCall(cudaMemcpyToSymbol(crot_matrices, rot_matrices, num_hypotheses * 3 * sizeof(float3)));
cudaSafeCall(cudaMemcpyToSymbol(ctransl_vectors, transl_vectors, num_hypotheses * sizeof(float3)));
void computeHypothesisScores(
const int num_hypotheses, const int num_points, const float* rot_matrices,
const float3* transl_vectors, const float3* object, const float2* image,
const float dist_threshold, int* hypothesis_scores)
{
cudaSafeCall(cudaMemcpyToSymbol(crot_matrices, rot_matrices, num_hypotheses * 3 * sizeof(float3)));
cudaSafeCall(cudaMemcpyToSymbol(ctransl_vectors, transl_vectors, num_hypotheses * sizeof(float3)));
dim3 threads(256);
dim3 grid(num_hypotheses);
int smem_size = threads.x * sizeof(float);
dim3 threads(256);
dim3 grid(num_hypotheses);
int smem_size = threads.x * sizeof(float);
computeHypothesisScoresKernel<<<grid, threads, smem_size>>>(
num_points, object, image, dist_threshold, hypothesis_scores);
cudaSafeCall( cudaGetLastError() );
computeHypothesisScoresKernel<<<grid, threads, smem_size>>>(
num_points, object, image, dist_threshold, hypothesis_scores);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace solvepnp_ransac
END_OPENCV_DEVICE_NAMESPACE
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace solvepnp_ransac
}}} // namespace cv { namespace gpu { namespace device

820
modules/gpu/src/cuda/canny.cu
View File

@@ -44,450 +44,448 @@
#include <algorithm>
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace canny {
__global__ void calcSobelRowPass(const PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
namespace cv { namespace gpu { namespace device
{
__shared__ int smem[16][18];
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (i < rows)
namespace canny
{
smem[threadIdx.y][threadIdx.x + 1] = src.ptr(i)[j];
if (threadIdx.x == 0)
__global__ void calcSobelRowPass(const PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
{
smem[threadIdx.y][0] = src.ptr(i)[::max(j - 1, 0)];
smem[threadIdx.y][17] = src.ptr(i)[::min(j + 16, cols - 1)];
}
__syncthreads();
__shared__ int smem[16][18];
if (j < cols)
{
dx_buf.ptr(i)[j] = -smem[threadIdx.y][threadIdx.x] + smem[threadIdx.y][threadIdx.x + 2];
dy_buf.ptr(i)[j] = smem[threadIdx.y][threadIdx.x] + 2 * smem[threadIdx.y][threadIdx.x + 1] + smem[threadIdx.y][threadIdx.x + 2];
}
}
}
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
void calcSobelRowPass_gpu(PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
calcSobelRowPass<<<grid, block>>>(src, dx_buf, dy_buf, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
struct L1
{
static __device__ __forceinline__ float calc(int x, int y)
{
return ::abs(x) + ::abs(y);
}
};
struct L2
{
static __device__ __forceinline__ float calc(int x, int y)
{
return ::sqrtf(x * x + y * y);
}
};
template <typename Norm> __global__ void calcMagnitude(const PtrStepi dx_buf, const PtrStepi dy_buf,
PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols)
{
__shared__ int sdx[18][16];
__shared__ int sdy[18][16];
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (j < cols)
{
sdx[threadIdx.y + 1][threadIdx.x] = dx_buf.ptr(i)[j];
sdy[threadIdx.y + 1][threadIdx.x] = dy_buf.ptr(i)[j];
if (threadIdx.y == 0)
{
sdx[0][threadIdx.x] = dx_buf.ptr(::max(i - 1, 0))[j];
sdx[17][threadIdx.x] = dx_buf.ptr(::min(i + 16, rows - 1))[j];
sdy[0][threadIdx.x] = dy_buf.ptr(::max(i - 1, 0))[j];
sdy[17][threadIdx.x] = dy_buf.ptr(::min(i + 16, rows - 1))[j];
}
__syncthreads();
if (i < rows)
{
int x = sdx[threadIdx.y][threadIdx.x] + 2 * sdx[threadIdx.y + 1][threadIdx.x] + sdx[threadIdx.y + 2][threadIdx.x];
int y = -sdy[threadIdx.y][threadIdx.x] + sdy[threadIdx.y + 2][threadIdx.x];
dx.ptr(i)[j] = x;
dy.ptr(i)[j] = y;
mag.ptr(i + 1)[j + 1] = Norm::calc(x, y);
}
}
}
void calcMagnitude_gpu(PtrStepi dx_buf, PtrStepi dy_buf, PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
template <typename Norm> __global__ void calcMagnitude(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols)
{
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (i < rows && j < cols)
mag.ptr(i + 1)[j + 1] = Norm::calc(dx.ptr(i)[j], dy.ptr(i)[j]);
}
void calcMagnitude_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx, dy, mag, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
//////////////////////////////////////////////////////////////////////////////////////////
#define CANNY_SHIFT 15
#define TG22 (int)(0.4142135623730950488016887242097*(1<<CANNY_SHIFT) + 0.5)
__global__ void calcMap(const PtrStepi dx, const PtrStepi dy, const PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
__shared__ float smem[18][18];
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
if (ly < 14)
smem[ly][lx] = mag.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
if (ly < 4 && blockIdx.y * 16 + ly + 14 <= rows && blockIdx.x * 16 + lx <= cols)
smem[ly + 14][lx] = mag.ptr(blockIdx.y * 16 + ly + 14)[blockIdx.x * 16 + lx];
__syncthreads();
if (i < rows && j < cols)
{
int x = dx.ptr(i)[j];
int y = dy.ptr(i)[j];
const int s = (x ^ y) < 0 ? -1 : 1;
const float m = smem[threadIdx.y + 1][threadIdx.x + 1];
x = ::abs(x);
y = ::abs(y);
// 0 - the pixel can not belong to an edge
// 1 - the pixel might belong to an edge
// 2 - the pixel does belong to an edge
int edge_type = 0;
if (m > low_thresh)
{
const int tg22x = x * TG22;
const int tg67x = tg22x + ((x + x) << CANNY_SHIFT);
y <<= CANNY_SHIFT;
if (y < tg22x)
if (i < rows)
{
if (m > smem[threadIdx.y + 1][threadIdx.x] && m >= smem[threadIdx.y + 1][threadIdx.x + 2])
edge_type = 1 + (int)(m > high_thresh);
}
else if( y > tg67x )
{
if (m > smem[threadIdx.y][threadIdx.x + 1] && m >= smem[threadIdx.y + 2][threadIdx.x + 1])
edge_type = 1 + (int)(m > high_thresh);
}
else
{
if (m > smem[threadIdx.y][threadIdx.x + 1 - s] && m > smem[threadIdx.y + 2][threadIdx.x + 1 + s])
edge_type = 1 + (int)(m > high_thresh);
}
}
map.ptr(i + 1)[j + 1] = edge_type;
}
}
#undef CANNY_SHIFT
#undef TG22
void calcMap_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
calcMap<<<grid, block>>>(dx, dy, mag, map, rows, cols, low_thresh, high_thresh);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
//////////////////////////////////////////////////////////////////////////////////////////
__device__ unsigned int counter = 0;
__global__ void edgesHysteresisLocal(PtrStepi map, ushort2* st, int rows, int cols)
{
#if __CUDA_ARCH__ >= 120
__shared__ int smem[18][18];
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
if (ly < 14)
smem[ly][lx] = map.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
if (ly < 4 && blockIdx.y * 16 + ly + 14 <= rows && blockIdx.x * 16 + lx <= cols)
smem[ly + 14][lx] = map.ptr(blockIdx.y * 16 + ly + 14)[blockIdx.x * 16 + lx];
__syncthreads();
if (i < rows && j < cols)
{
int n;
#pragma unroll
for (int k = 0; k < 16; ++k)
{
n = 0;
if (smem[threadIdx.y + 1][threadIdx.x + 1] == 1)
{
n += smem[threadIdx.y ][threadIdx.x ] == 2;
n += smem[threadIdx.y ][threadIdx.x + 1] == 2;
n += smem[threadIdx.y ][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 1][threadIdx.x ] == 2;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 2][threadIdx.x ] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 2;
}
if (n > 0)
smem[threadIdx.y + 1][threadIdx.x + 1] = 2;
}
const int e = smem[threadIdx.y + 1][threadIdx.x + 1];
map.ptr(i + 1)[j + 1] = e;
n = 0;
if (e == 2)
{
n += smem[threadIdx.y ][threadIdx.x ] == 1;
n += smem[threadIdx.y ][threadIdx.x + 1] == 1;
n += smem[threadIdx.y ][threadIdx.x + 2] == 1;
n += smem[threadIdx.y + 1][threadIdx.x ] == 1;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 1;
n += smem[threadIdx.y + 2][threadIdx.x ] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 1;
}
if (n > 0)
{
const unsigned int ind = atomicInc(&counter, (unsigned int)(-1));
st[ind] = make_ushort2(j + 1, i + 1);
}
}
#endif
}
void edgesHysteresisLocal_gpu(PtrStepi map, ushort2* st1, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
edgesHysteresisLocal<<<grid, block>>>(map, st1, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
__constant__ int c_dx[8] = {-1, 0, 1, -1, 1, -1, 0, 1};
__constant__ int c_dy[8] = {-1, -1, -1, 0, 0, 1, 1, 1};
__global__ void edgesHysteresisGlobal(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols, int count)
{
#if __CUDA_ARCH__ >= 120
const int stack_size = 512;
__shared__ unsigned int s_counter;
__shared__ unsigned int s_ind;
__shared__ ushort2 s_st[stack_size];
if (threadIdx.x == 0)
s_counter = 0;
__syncthreads();
int ind = blockIdx.y * gridDim.x + blockIdx.x;
if (ind < count)
{
ushort2 pos = st1[ind];
if (pos.x > 0 && pos.x <= cols && pos.y > 0 && pos.y <= rows)
{
if (threadIdx.x < 8)
{
pos.x += c_dx[threadIdx.x];
pos.y += c_dy[threadIdx.x];
if (map.ptr(pos.y)[pos.x] == 1)
smem[threadIdx.y][threadIdx.x + 1] = src.ptr(i)[j];
if (threadIdx.x == 0)
{
map.ptr(pos.y)[pos.x] = 2;
smem[threadIdx.y][0] = src.ptr(i)[::max(j - 1, 0)];
smem[threadIdx.y][17] = src.ptr(i)[::min(j + 16, cols - 1)];
}
__syncthreads();
ind = atomicInc(&s_counter, (unsigned int)(-1));
s_st[ind] = pos;
if (j < cols)
{
dx_buf.ptr(i)[j] = -smem[threadIdx.y][threadIdx.x] + smem[threadIdx.y][threadIdx.x + 2];
dy_buf.ptr(i)[j] = smem[threadIdx.y][threadIdx.x] + 2 * smem[threadIdx.y][threadIdx.x + 1] + smem[threadIdx.y][threadIdx.x + 2];
}
}
}
void calcSobelRowPass_gpu(PtrStepb src, PtrStepi dx_buf, PtrStepi dy_buf, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
calcSobelRowPass<<<grid, block>>>(src, dx_buf, dy_buf, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
struct L1
{
static __device__ __forceinline__ float calc(int x, int y)
{
return ::abs(x) + ::abs(y);
}
};
struct L2
{
static __device__ __forceinline__ float calc(int x, int y)
{
return ::sqrtf(x * x + y * y);
}
};
template <typename Norm> __global__ void calcMagnitude(const PtrStepi dx_buf, const PtrStepi dy_buf,
PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols)
{
__shared__ int sdx[18][16];
__shared__ int sdy[18][16];
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (j < cols)
{
sdx[threadIdx.y + 1][threadIdx.x] = dx_buf.ptr(i)[j];
sdy[threadIdx.y + 1][threadIdx.x] = dy_buf.ptr(i)[j];
if (threadIdx.y == 0)
{
sdx[0][threadIdx.x] = dx_buf.ptr(::max(i - 1, 0))[j];
sdx[17][threadIdx.x] = dx_buf.ptr(::min(i + 16, rows - 1))[j];
sdy[0][threadIdx.x] = dy_buf.ptr(::max(i - 1, 0))[j];
sdy[17][threadIdx.x] = dy_buf.ptr(::min(i + 16, rows - 1))[j];
}
__syncthreads();
if (i < rows)
{
int x = sdx[threadIdx.y][threadIdx.x] + 2 * sdx[threadIdx.y + 1][threadIdx.x] + sdx[threadIdx.y + 2][threadIdx.x];
int y = -sdy[threadIdx.y][threadIdx.x] + sdy[threadIdx.y + 2][threadIdx.x];
dx.ptr(i)[j] = x;
dy.ptr(i)[j] = y;
mag.ptr(i + 1)[j + 1] = Norm::calc(x, y);
}
}
}
void calcMagnitude_gpu(PtrStepi dx_buf, PtrStepi dy_buf, PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx_buf, dy_buf, dx, dy, mag, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
template <typename Norm> __global__ void calcMagnitude(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols)
{
const int j = blockIdx.x * blockDim.x + threadIdx.x;
const int i = blockIdx.y * blockDim.y + threadIdx.y;
if (i < rows && j < cols)
mag.ptr(i + 1)[j + 1] = Norm::calc(dx.ptr(i)[j], dy.ptr(i)[j]);
}
void calcMagnitude_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, int rows, int cols, bool L2Grad)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
if (L2Grad)
calcMagnitude<L2><<<grid, block>>>(dx, dy, mag, rows, cols);
else
calcMagnitude<L1><<<grid, block>>>(dx, dy, mag, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
//////////////////////////////////////////////////////////////////////////////////////////
#define CANNY_SHIFT 15
#define TG22 (int)(0.4142135623730950488016887242097*(1<<CANNY_SHIFT) + 0.5)
__global__ void calcMap(const PtrStepi dx, const PtrStepi dy, const PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
__shared__ float smem[18][18];
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
if (ly < 14)
smem[ly][lx] = mag.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
if (ly < 4 && blockIdx.y * 16 + ly + 14 <= rows && blockIdx.x * 16 + lx <= cols)
smem[ly + 14][lx] = mag.ptr(blockIdx.y * 16 + ly + 14)[blockIdx.x * 16 + lx];
__syncthreads();
while (s_counter > 0 && s_counter <= stack_size - blockDim.x)
if (i < rows && j < cols)
{
const int subTaskIdx = threadIdx.x >> 3;
const int portion = ::min(s_counter, blockDim.x >> 3);
int x = dx.ptr(i)[j];
int y = dy.ptr(i)[j];
const int s = (x ^ y) < 0 ? -1 : 1;
const float m = smem[threadIdx.y + 1][threadIdx.x + 1];
pos.x = pos.y = 0;
x = ::abs(x);
y = ::abs(y);
if (subTaskIdx < portion)
pos = s_st[s_counter - 1 - subTaskIdx];
__syncthreads();
if (threadIdx.x == 0)
s_counter -= portion;
__syncthreads();
if (pos.x > 0 && pos.x <= cols && pos.y > 0 && pos.y <= rows)
// 0 - the pixel can not belong to an edge
// 1 - the pixel might belong to an edge
// 2 - the pixel does belong to an edge
int edge_type = 0;
if (m > low_thresh)
{
pos.x += c_dx[threadIdx.x & 7];
pos.y += c_dy[threadIdx.x & 7];
const int tg22x = x * TG22;
const int tg67x = tg22x + ((x + x) << CANNY_SHIFT);
if (map.ptr(pos.y)[pos.x] == 1)
y <<= CANNY_SHIFT;
if (y < tg22x)
{
map.ptr(pos.y)[pos.x] = 2;
ind = atomicInc(&s_counter, (unsigned int)(-1));
s_st[ind] = pos;
if (m > smem[threadIdx.y + 1][threadIdx.x] && m >= smem[threadIdx.y + 1][threadIdx.x + 2])
edge_type = 1 + (int)(m > high_thresh);
}
else if( y > tg67x )
{
if (m > smem[threadIdx.y][threadIdx.x + 1] && m >= smem[threadIdx.y + 2][threadIdx.x + 1])
edge_type = 1 + (int)(m > high_thresh);
}
else
{
if (m > smem[threadIdx.y][threadIdx.x + 1 - s] && m > smem[threadIdx.y + 2][threadIdx.x + 1 + s])
edge_type = 1 + (int)(m > high_thresh);
}
}
__syncthreads();
}
if (s_counter > 0)
{
if (threadIdx.x == 0)
{
ind = atomicAdd(&counter, s_counter);
s_ind = ind - s_counter;
}
__syncthreads();
ind = s_ind;
for (int i = threadIdx.x; i < s_counter; i += blockDim.x)
{
st2[ind + i] = s_st[i];
}
map.ptr(i + 1)[j + 1] = edge_type;
}
}
}
#endif
}
#undef CANNY_SHIFT
#undef TG22
void edgesHysteresisGlobal_gpu(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, counter) );
unsigned int count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
void calcMap_gpu(PtrStepi dx, PtrStepi dy, PtrStepf mag, PtrStepi map, int rows, int cols, float low_thresh, float high_thresh)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
while (count > 0)
{
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(unsigned int)) );
calcMap<<<grid, block>>>(dx, dy, mag, map, rows, cols, low_thresh, high_thresh);
cudaSafeCall( cudaGetLastError() );
dim3 block(128, 1, 1);
dim3 grid(std::min(count, 65535u), divUp(count, 65535), 1);
edgesHysteresisGlobal<<<grid, block>>>(map, st1, st2, rows, cols, count);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
cudaSafeCall(cudaThreadSynchronize());
//////////////////////////////////////////////////////////////////////////////////////////
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
__device__ unsigned int counter = 0;
std::swap(st1, st2);
}
}
__global__ void edgesHysteresisLocal(PtrStepi map, ushort2* st, int rows, int cols)
{
#if __CUDA_ARCH__ >= 120
__global__ void getEdges(PtrStepi map, PtrStepb dst, int rows, int cols)
{
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
__shared__ int smem[18][18];
if (i < rows && j < cols)
dst.ptr(i)[j] = (uchar)(-(map.ptr(i + 1)[j + 1] >> 1));
}
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
void getEdges_gpu(PtrStepi map, PtrStepb dst, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
const int tid = threadIdx.y * 16 + threadIdx.x;
const int lx = tid % 18;
const int ly = tid / 18;
getEdges<<<grid, block>>>(map, dst, rows, cols);
cudaSafeCall( cudaGetLastError() );
if (ly < 14)
smem[ly][lx] = map.ptr(blockIdx.y * 16 + ly)[blockIdx.x * 16 + lx];
cudaSafeCall(cudaThreadSynchronize());
}
if (ly < 4 && blockIdx.y * 16 + ly + 14 <= rows && blockIdx.x * 16 + lx <= cols)
smem[ly + 14][lx] = map.ptr(blockIdx.y * 16 + ly + 14)[blockIdx.x * 16 + lx];
} // namespace canny
__syncthreads();
END_OPENCV_DEVICE_NAMESPACE
if (i < rows && j < cols)
{
int n;
#pragma unroll
for (int k = 0; k < 16; ++k)
{
n = 0;
if (smem[threadIdx.y + 1][threadIdx.x + 1] == 1)
{
n += smem[threadIdx.y ][threadIdx.x ] == 2;
n += smem[threadIdx.y ][threadIdx.x + 1] == 2;
n += smem[threadIdx.y ][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 1][threadIdx.x ] == 2;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 2;
n += smem[threadIdx.y + 2][threadIdx.x ] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 2;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 2;
}
if (n > 0)
smem[threadIdx.y + 1][threadIdx.x + 1] = 2;
}
const int e = smem[threadIdx.y + 1][threadIdx.x + 1];
map.ptr(i + 1)[j + 1] = e;
n = 0;
if (e == 2)
{
n += smem[threadIdx.y ][threadIdx.x ] == 1;
n += smem[threadIdx.y ][threadIdx.x + 1] == 1;
n += smem[threadIdx.y ][threadIdx.x + 2] == 1;
n += smem[threadIdx.y + 1][threadIdx.x ] == 1;
n += smem[threadIdx.y + 1][threadIdx.x + 2] == 1;
n += smem[threadIdx.y + 2][threadIdx.x ] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 1] == 1;
n += smem[threadIdx.y + 2][threadIdx.x + 2] == 1;
}
if (n > 0)
{
const unsigned int ind = atomicInc(&counter, (unsigned int)(-1));
st[ind] = make_ushort2(j + 1, i + 1);
}
}
#endif
}
void edgesHysteresisLocal_gpu(PtrStepi map, ushort2* st1, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
edgesHysteresisLocal<<<grid, block>>>(map, st1, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
__constant__ int c_dx[8] = {-1, 0, 1, -1, 1, -1, 0, 1};
__constant__ int c_dy[8] = {-1, -1, -1, 0, 0, 1, 1, 1};
__global__ void edgesHysteresisGlobal(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols, int count)
{
#if __CUDA_ARCH__ >= 120
const int stack_size = 512;
__shared__ unsigned int s_counter;
__shared__ unsigned int s_ind;
__shared__ ushort2 s_st[stack_size];
if (threadIdx.x == 0)
s_counter = 0;
__syncthreads();
int ind = blockIdx.y * gridDim.x + blockIdx.x;
if (ind < count)
{
ushort2 pos = st1[ind];
if (pos.x > 0 && pos.x <= cols && pos.y > 0 && pos.y <= rows)
{
if (threadIdx.x < 8)
{
pos.x += c_dx[threadIdx.x];
pos.y += c_dy[threadIdx.x];
if (map.ptr(pos.y)[pos.x] == 1)
{
map.ptr(pos.y)[pos.x] = 2;
ind = atomicInc(&s_counter, (unsigned int)(-1));
s_st[ind] = pos;
}
}
__syncthreads();
while (s_counter > 0 && s_counter <= stack_size - blockDim.x)
{
const int subTaskIdx = threadIdx.x >> 3;
const int portion = ::min(s_counter, blockDim.x >> 3);
pos.x = pos.y = 0;
if (subTaskIdx < portion)
pos = s_st[s_counter - 1 - subTaskIdx];
__syncthreads();
if (threadIdx.x == 0)
s_counter -= portion;
__syncthreads();
if (pos.x > 0 && pos.x <= cols && pos.y > 0 && pos.y <= rows)
{
pos.x += c_dx[threadIdx.x & 7];
pos.y += c_dy[threadIdx.x & 7];
if (map.ptr(pos.y)[pos.x] == 1)
{
map.ptr(pos.y)[pos.x] = 2;
ind = atomicInc(&s_counter, (unsigned int)(-1));
s_st[ind] = pos;
}
}
__syncthreads();
}
if (s_counter > 0)
{
if (threadIdx.x == 0)
{
ind = atomicAdd(&counter, s_counter);
s_ind = ind - s_counter;
}
__syncthreads();
ind = s_ind;
for (int i = threadIdx.x; i < s_counter; i += blockDim.x)
{
st2[ind + i] = s_st[i];
}
}
}
}
#endif
}
void edgesHysteresisGlobal_gpu(PtrStepi map, ushort2* st1, ushort2* st2, int rows, int cols)
{
void* counter_ptr;
cudaSafeCall( cudaGetSymbolAddress(&counter_ptr, counter) );
unsigned int count;
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
while (count > 0)
{
cudaSafeCall( cudaMemset(counter_ptr, 0, sizeof(unsigned int)) );
dim3 block(128, 1, 1);
dim3 grid(std::min(count, 65535u), divUp(count, 65535), 1);
edgesHysteresisGlobal<<<grid, block>>>(map, st1, st2, rows, cols, count);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
cudaSafeCall( cudaMemcpy(&count, counter_ptr, sizeof(unsigned int), cudaMemcpyDeviceToHost) );
std::swap(st1, st2);
}
}
__global__ void getEdges(PtrStepi map, PtrStepb dst, int rows, int cols)
{
const int j = blockIdx.x * 16 + threadIdx.x;
const int i = blockIdx.y * 16 + threadIdx.y;
if (i < rows && j < cols)
dst.ptr(i)[j] = (uchar)(-(map.ptr(i + 1)[j + 1] >> 1));
}
void getEdges_gpu(PtrStepi map, PtrStepb dst, int rows, int cols)
{
dim3 block(16, 16, 1);
dim3 grid(divUp(cols, block.x), divUp(rows, block.y), 1);
getEdges<<<grid, block>>>(map, dst, rows, cols);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall(cudaThreadSynchronize());
}
} // namespace canny
}}} // namespace cv { namespace gpu { namespace device

559
modules/gpu/src/cuda/color.cu
View File

@@ -44,181 +44,181 @@
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/color.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_x = 8 };
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_x = 8 };
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr555_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_bgr565_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr555_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_bgra_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgr565_to_rgba_traits::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr555_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr565_traits::functor_type)
{
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr555_traits::functor_type)
{
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(gray_to_bgr565_traits::functor_type)
{
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_yuv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(yuv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_YCrCb4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(YCrCb4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_xyz4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(xyz4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hsv4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hsv4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(bgra_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(rgba_to_hls4_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_bgra_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
OPENCV_GPU_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR(name, traits) \
void name(const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream) \
@@ -226,7 +226,7 @@ DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
traits::functor_type functor = traits::create_functor(); \
typedef typename traits::functor_type::argument_type src_t; \
typedef typename traits::functor_type::result_type dst_t; \
OPENCV_DEVICE_NAMESPACE_ transform((DevMem2D_<src_t>)src, (DevMem2D_<dst_t>)dst, functor, stream); \
::cv::gpu::device::transform((DevMem2D_<src_t>)src, (DevMem2D_<dst_t>)dst, functor, stream); \
}
#define OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(name) \
@@ -243,138 +243,137 @@ DEFINE_TRANSFORM_FUNCTOR_TRAITS(hls4_to_rgba_traits<uchar>::functor_type)
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_8u, name ## _full_traits<uchar>) \
OPENCV_GPU_IMPLEMENT_CVTCOLOR(name ## _full_32f, name ## _full_traits<float>)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgb_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgra_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(rgba_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(gray_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr555)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(gray_to_bgr565)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr555_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE(bgr565_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_gray)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_yuv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(yuv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_YCrCb4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(YCrCb4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgb_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(rgba_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgr_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(bgra_to_xyz4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL(xyz4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hsv4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hsv4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgb_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(rgba_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgr_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(bgra_to_hls4)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgb)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_rgba)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls_to_bgra)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgr)
OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F(hls4_to_bgra)
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F
END_OPENCV_DEVICE_NAMESPACE
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ONE
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_ALL
#undef OPENCV_GPU_IMPLEMENT_CVTCOLOR_8U32F
}}} // namespace cv { namespace gpu { namespace device

366
modules/gpu/src/cuda/column_filter.cu
View File

@@ -47,203 +47,201 @@
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
namespace column_filter {
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float kernel[], int ksize)
namespace cv { namespace gpu { namespace device
{
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearColumnFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
__shared__ T smem[BLOCK_DIM_X][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_Y + 1];
//Offset to the upper halo edge
const int x = blockIdx.x * BLOCK_DIM_X + threadIdx.x;
const int y = (blockIdx.y * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_Y + threadIdx.y;
if (x < src.cols)
namespace column_filter
{
const T* src_col = src.ptr() + x;
__constant__ float c_kernel[MAX_KERNEL_SIZE];
//Main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
//Upper halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_low(y + i * BLOCK_DIM_Y, src_col, src.step);
//Lower halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y]= b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
__syncthreads();
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
void loadKernel(const float kernel[], int ksize)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for(int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y + j - anchor] * c_kernel[j];
int dstY = y + i * BLOCK_DIM_Y;
if (dstY < src.rows)
dst.ptr(dstY)[x] = saturate_cast<D>(sum);
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
}
}
template <int ksize, typename T, typename D, template<typename> class B>
void linearColumnFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, RESULT_STEPS * BLOCK_DIM_Y));
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearColumnFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
B<T> b(src.rows);
__shared__ T smem[BLOCK_DIM_X][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_Y + 1];
linearColumnFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
//Offset to the upper halo edge
const int x = blockIdx.x * BLOCK_DIM_X + threadIdx.x;
const int y = (blockIdx.y * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_Y + threadIdx.y;
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (x < src.cols)
{
const T* src_col = src.ptr() + x;
template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
{
0,
linearColumnFilter_caller<1 , T, D, BrdColReflect101>,
linearColumnFilter_caller<2 , T, D, BrdColReflect101>,
linearColumnFilter_caller<3 , T, D, BrdColReflect101>,
linearColumnFilter_caller<4 , T, D, BrdColReflect101>,
linearColumnFilter_caller<5 , T, D, BrdColReflect101>,
linearColumnFilter_caller<6 , T, D, BrdColReflect101>,
linearColumnFilter_caller<7 , T, D, BrdColReflect101>,
linearColumnFilter_caller<8 , T, D, BrdColReflect101>,
linearColumnFilter_caller<9 , T, D, BrdColReflect101>,
linearColumnFilter_caller<10, T, D, BrdColReflect101>,
linearColumnFilter_caller<11, T, D, BrdColReflect101>,
linearColumnFilter_caller<12, T, D, BrdColReflect101>,
linearColumnFilter_caller<13, T, D, BrdColReflect101>,
linearColumnFilter_caller<14, T, D, BrdColReflect101>,
linearColumnFilter_caller<15, T, D, BrdColReflect101>,
linearColumnFilter_caller<16, T, D, BrdColReflect101>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColReplicate>,
linearColumnFilter_caller<2 , T, D, BrdColReplicate>,
linearColumnFilter_caller<3 , T, D, BrdColReplicate>,
linearColumnFilter_caller<4 , T, D, BrdColReplicate>,
linearColumnFilter_caller<5 , T, D, BrdColReplicate>,
linearColumnFilter_caller<6 , T, D, BrdColReplicate>,
linearColumnFilter_caller<7 , T, D, BrdColReplicate>,
linearColumnFilter_caller<8 , T, D, BrdColReplicate>,
linearColumnFilter_caller<9 , T, D, BrdColReplicate>,
linearColumnFilter_caller<10, T, D, BrdColReplicate>,
linearColumnFilter_caller<11, T, D, BrdColReplicate>,
linearColumnFilter_caller<12, T, D, BrdColReplicate>,
linearColumnFilter_caller<13, T, D, BrdColReplicate>,
linearColumnFilter_caller<14, T, D, BrdColReplicate>,
linearColumnFilter_caller<15, T, D, BrdColReplicate>,
linearColumnFilter_caller<16, T, D, BrdColReplicate>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColConstant>,
linearColumnFilter_caller<2 , T, D, BrdColConstant>,
linearColumnFilter_caller<3 , T, D, BrdColConstant>,
linearColumnFilter_caller<4 , T, D, BrdColConstant>,
linearColumnFilter_caller<5 , T, D, BrdColConstant>,
linearColumnFilter_caller<6 , T, D, BrdColConstant>,
linearColumnFilter_caller<7 , T, D, BrdColConstant>,
linearColumnFilter_caller<8 , T, D, BrdColConstant>,
linearColumnFilter_caller<9 , T, D, BrdColConstant>,
linearColumnFilter_caller<10, T, D, BrdColConstant>,
linearColumnFilter_caller<11, T, D, BrdColConstant>,
linearColumnFilter_caller<12, T, D, BrdColConstant>,
linearColumnFilter_caller<13, T, D, BrdColConstant>,
linearColumnFilter_caller<14, T, D, BrdColConstant>,
linearColumnFilter_caller<15, T, D, BrdColConstant>,
linearColumnFilter_caller<16, T, D, BrdColConstant>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColReflect>,
linearColumnFilter_caller<2 , T, D, BrdColReflect>,
linearColumnFilter_caller<3 , T, D, BrdColReflect>,
linearColumnFilter_caller<4 , T, D, BrdColReflect>,
linearColumnFilter_caller<5 , T, D, BrdColReflect>,
linearColumnFilter_caller<6 , T, D, BrdColReflect>,
linearColumnFilter_caller<7 , T, D, BrdColReflect>,
linearColumnFilter_caller<8 , T, D, BrdColReflect>,
linearColumnFilter_caller<9 , T, D, BrdColReflect>,
linearColumnFilter_caller<10, T, D, BrdColReflect>,
linearColumnFilter_caller<11, T, D, BrdColReflect>,
linearColumnFilter_caller<12, T, D, BrdColReflect>,
linearColumnFilter_caller<13, T, D, BrdColReflect>,
linearColumnFilter_caller<14, T, D, BrdColReflect>,
linearColumnFilter_caller<15, T, D, BrdColReflect>,
linearColumnFilter_caller<16, T, D, BrdColReflect>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColWrap>,
linearColumnFilter_caller<2 , T, D, BrdColWrap>,
linearColumnFilter_caller<3 , T, D, BrdColWrap>,
linearColumnFilter_caller<4 , T, D, BrdColWrap>,
linearColumnFilter_caller<5 , T, D, BrdColWrap>,
linearColumnFilter_caller<6 , T, D, BrdColWrap>,
linearColumnFilter_caller<7 , T, D, BrdColWrap>,
linearColumnFilter_caller<8 , T, D, BrdColWrap>,
linearColumnFilter_caller<9 , T, D, BrdColWrap>,
linearColumnFilter_caller<10, T, D, BrdColWrap>,
linearColumnFilter_caller<11, T, D, BrdColWrap>,
linearColumnFilter_caller<12, T, D, BrdColWrap>,
linearColumnFilter_caller<13, T, D, BrdColWrap>,
linearColumnFilter_caller<14, T, D, BrdColWrap>,
linearColumnFilter_caller<15, T, D, BrdColWrap>,
linearColumnFilter_caller<16, T, D, BrdColWrap>,
//Main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
//Upper halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y] = b.at_low(y + i * BLOCK_DIM_Y, src_col, src.step);
//Lower halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y]= b.at_high(y + i * BLOCK_DIM_Y, src_col, src.step);
__syncthreads();
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for(int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.x][threadIdx.y + i * BLOCK_DIM_Y + j - anchor] * c_kernel[j];
int dstY = y + i * BLOCK_DIM_Y;
if (dstY < src.rows)
dst.ptr(dstY)[x] = saturate_cast<D>(sum);
}
}
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
template <int ksize, typename T, typename D, template<typename> class B>
void linearColumnFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, BLOCK_DIM_X), divUp(src.rows, RESULT_STEPS * BLOCK_DIM_Y));
template void linearColumnFilter_gpu<float , uchar >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float4, uchar4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float , short >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float2, short2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float3, short3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , int >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
B<T> b(src.rows);
} // namespace column_filter
linearColumnFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
END_OPENCV_DEVICE_NAMESPACE
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearColumnFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
{
0,
linearColumnFilter_caller<1 , T, D, BrdColReflect101>,
linearColumnFilter_caller<2 , T, D, BrdColReflect101>,
linearColumnFilter_caller<3 , T, D, BrdColReflect101>,
linearColumnFilter_caller<4 , T, D, BrdColReflect101>,
linearColumnFilter_caller<5 , T, D, BrdColReflect101>,
linearColumnFilter_caller<6 , T, D, BrdColReflect101>,
linearColumnFilter_caller<7 , T, D, BrdColReflect101>,
linearColumnFilter_caller<8 , T, D, BrdColReflect101>,
linearColumnFilter_caller<9 , T, D, BrdColReflect101>,
linearColumnFilter_caller<10, T, D, BrdColReflect101>,
linearColumnFilter_caller<11, T, D, BrdColReflect101>,
linearColumnFilter_caller<12, T, D, BrdColReflect101>,
linearColumnFilter_caller<13, T, D, BrdColReflect101>,
linearColumnFilter_caller<14, T, D, BrdColReflect101>,
linearColumnFilter_caller<15, T, D, BrdColReflect101>,
linearColumnFilter_caller<16, T, D, BrdColReflect101>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColReplicate>,
linearColumnFilter_caller<2 , T, D, BrdColReplicate>,
linearColumnFilter_caller<3 , T, D, BrdColReplicate>,
linearColumnFilter_caller<4 , T, D, BrdColReplicate>,
linearColumnFilter_caller<5 , T, D, BrdColReplicate>,
linearColumnFilter_caller<6 , T, D, BrdColReplicate>,
linearColumnFilter_caller<7 , T, D, BrdColReplicate>,
linearColumnFilter_caller<8 , T, D, BrdColReplicate>,
linearColumnFilter_caller<9 , T, D, BrdColReplicate>,
linearColumnFilter_caller<10, T, D, BrdColReplicate>,
linearColumnFilter_caller<11, T, D, BrdColReplicate>,
linearColumnFilter_caller<12, T, D, BrdColReplicate>,
linearColumnFilter_caller<13, T, D, BrdColReplicate>,
linearColumnFilter_caller<14, T, D, BrdColReplicate>,
linearColumnFilter_caller<15, T, D, BrdColReplicate>,
linearColumnFilter_caller<16, T, D, BrdColReplicate>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColConstant>,
linearColumnFilter_caller<2 , T, D, BrdColConstant>,
linearColumnFilter_caller<3 , T, D, BrdColConstant>,
linearColumnFilter_caller<4 , T, D, BrdColConstant>,
linearColumnFilter_caller<5 , T, D, BrdColConstant>,
linearColumnFilter_caller<6 , T, D, BrdColConstant>,
linearColumnFilter_caller<7 , T, D, BrdColConstant>,
linearColumnFilter_caller<8 , T, D, BrdColConstant>,
linearColumnFilter_caller<9 , T, D, BrdColConstant>,
linearColumnFilter_caller<10, T, D, BrdColConstant>,
linearColumnFilter_caller<11, T, D, BrdColConstant>,
linearColumnFilter_caller<12, T, D, BrdColConstant>,
linearColumnFilter_caller<13, T, D, BrdColConstant>,
linearColumnFilter_caller<14, T, D, BrdColConstant>,
linearColumnFilter_caller<15, T, D, BrdColConstant>,
linearColumnFilter_caller<16, T, D, BrdColConstant>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColReflect>,
linearColumnFilter_caller<2 , T, D, BrdColReflect>,
linearColumnFilter_caller<3 , T, D, BrdColReflect>,
linearColumnFilter_caller<4 , T, D, BrdColReflect>,
linearColumnFilter_caller<5 , T, D, BrdColReflect>,
linearColumnFilter_caller<6 , T, D, BrdColReflect>,
linearColumnFilter_caller<7 , T, D, BrdColReflect>,
linearColumnFilter_caller<8 , T, D, BrdColReflect>,
linearColumnFilter_caller<9 , T, D, BrdColReflect>,
linearColumnFilter_caller<10, T, D, BrdColReflect>,
linearColumnFilter_caller<11, T, D, BrdColReflect>,
linearColumnFilter_caller<12, T, D, BrdColReflect>,
linearColumnFilter_caller<13, T, D, BrdColReflect>,
linearColumnFilter_caller<14, T, D, BrdColReflect>,
linearColumnFilter_caller<15, T, D, BrdColReflect>,
linearColumnFilter_caller<16, T, D, BrdColReflect>
},
{
0,
linearColumnFilter_caller<1 , T, D, BrdColWrap>,
linearColumnFilter_caller<2 , T, D, BrdColWrap>,
linearColumnFilter_caller<3 , T, D, BrdColWrap>,
linearColumnFilter_caller<4 , T, D, BrdColWrap>,
linearColumnFilter_caller<5 , T, D, BrdColWrap>,
linearColumnFilter_caller<6 , T, D, BrdColWrap>,
linearColumnFilter_caller<7 , T, D, BrdColWrap>,
linearColumnFilter_caller<8 , T, D, BrdColWrap>,
linearColumnFilter_caller<9 , T, D, BrdColWrap>,
linearColumnFilter_caller<10, T, D, BrdColWrap>,
linearColumnFilter_caller<11, T, D, BrdColWrap>,
linearColumnFilter_caller<12, T, D, BrdColWrap>,
linearColumnFilter_caller<13, T, D, BrdColWrap>,
linearColumnFilter_caller<14, T, D, BrdColWrap>,
linearColumnFilter_caller<15, T, D, BrdColWrap>,
linearColumnFilter_caller<16, T, D, BrdColWrap>,
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
template void linearColumnFilter_gpu<float , uchar >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float4, uchar4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float , short >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearColumnFilter_gpu<float2, short2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float3, short3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , int >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearColumnFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
} // namespace column_filter
}}} // namespace cv { namespace gpu { namespace device

144
modules/gpu/src/cuda/copy_make_border.cu
View File

@@ -43,87 +43,85 @@
#include "internal_shared.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename Ptr2D, typename T> __global__ void copyMakeBorder(const Ptr2D src, DevMem2D_<T> dst, int top, int left)
namespace cv { namespace gpu { namespace device
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = src(y - top, x - left);
}
template <template <typename> class B, typename T> struct CopyMakeBorderDispatcher
{
static void call(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, int top, int left,
const typename VecTraits<T>::elem_type* borderValue, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<T> brd(src.rows, src.cols, VecTraits<T>::make(borderValue));
BorderReader< PtrStep<T>, B<T> > brdSrc(src, brd);
copyMakeBorder<<<grid, block, 0, stream>>>(brdSrc, dst, top, left);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T, int cn> void copyMakeBorder_gpu(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode,
const T* borderValue, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type vec_type;
typedef void (*caller_t)(const DevMem2D_<vec_type>& src, const DevMem2D_<vec_type>& dst, int top, int left, const T* borderValue, cudaStream_t stream);
static const caller_t callers[5] =
namespace imgproc
{
CopyMakeBorderDispatcher<BrdReflect101, vec_type>::call,
CopyMakeBorderDispatcher<BrdReplicate, vec_type>::call,
CopyMakeBorderDispatcher<BrdConstant, vec_type>::call,
CopyMakeBorderDispatcher<BrdReflect, vec_type>::call,
CopyMakeBorderDispatcher<BrdWrap, vec_type>::call
};
template <typename Ptr2D, typename T> __global__ void copyMakeBorder(const Ptr2D src, DevMem2D_<T> dst, int top, int left)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
callers[borderMode](DevMem2D_<vec_type>(src), DevMem2D_<vec_type>(dst), top, left, borderValue, stream);
}
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = src(y - top, x - left);
}
template void copyMakeBorder_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template <template <typename> class B, typename T> struct CopyMakeBorderDispatcher
{
static void call(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, int top, int left,
const typename VecTraits<T>::elem_type* borderValue, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
//template void copyMakeBorder_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
B<T> brd(src.rows, src.cols, VecTraits<T>::make(borderValue));
BorderReader< PtrStep<T>, B<T> > brdSrc(src, brd);
template void copyMakeBorder_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
copyMakeBorder<<<grid, block, 0, stream>>>(brdSrc, dst, top, left);
cudaSafeCall( cudaGetLastError() );
template void copyMakeBorder_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
};
//template void copyMakeBorder_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
template <typename T, int cn> void copyMakeBorder_gpu(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode,
const T* borderValue, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type vec_type;
template void copyMakeBorder_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
typedef void (*caller_t)(const DevMem2D_<vec_type>& src, const DevMem2D_<vec_type>& dst, int top, int left, const T* borderValue, cudaStream_t stream);
} // namespace imgproc
static const caller_t callers[5] =
{
CopyMakeBorderDispatcher<BrdReflect101, vec_type>::call,
CopyMakeBorderDispatcher<BrdReplicate, vec_type>::call,
CopyMakeBorderDispatcher<BrdConstant, vec_type>::call,
CopyMakeBorderDispatcher<BrdReflect, vec_type>::call,
CopyMakeBorderDispatcher<BrdWrap, vec_type>::call
};
END_OPENCV_DEVICE_NAMESPACE
callers[borderMode](DevMem2D_<vec_type>(src), DevMem2D_<vec_type>(dst), top, left, borderValue, stream);
}
template void copyMakeBorder_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const uchar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const schar* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const ushort* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const short* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const int* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
//template void copyMakeBorder_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
template void copyMakeBorder_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int top, int left, int borderMode, const float* borderValue, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

3870
modules/gpu/src/cuda/element_operations.cu
View File

File diff suppressed because it is too large Load Diff

276
modules/gpu/src/cuda/hist.cu
View File

@@ -45,177 +45,175 @@
#include "opencv2/gpu/device/utility.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace cv { namespace gpu { namespace device
{
#define UINT_BITS 32U
#define UINT_BITS 32U
//Warps == subhistograms per threadblock
#define WARP_COUNT 6
//Warps == subhistograms per threadblock
#define WARP_COUNT 6
//Threadblock size
#define HISTOGRAM256_THREADBLOCK_SIZE (WARP_COUNT * OPENCV_GPU_WARP_SIZE)
#define HISTOGRAM256_BIN_COUNT 256
//Threadblock size
#define HISTOGRAM256_THREADBLOCK_SIZE (WARP_COUNT * OPENCV_GPU_WARP_SIZE)
#define HISTOGRAM256_BIN_COUNT 256
//Shared memory per threadblock
#define HISTOGRAM256_THREADBLOCK_MEMORY (WARP_COUNT * HISTOGRAM256_BIN_COUNT)
//Shared memory per threadblock
#define HISTOGRAM256_THREADBLOCK_MEMORY (WARP_COUNT * HISTOGRAM256_BIN_COUNT)
#define PARTIAL_HISTOGRAM256_COUNT 240
#define PARTIAL_HISTOGRAM256_COUNT 240
#define MERGE_THREADBLOCK_SIZE 256
#define MERGE_THREADBLOCK_SIZE 256
#define USE_SMEM_ATOMICS (__CUDA_ARCH__ >= 120)
#define USE_SMEM_ATOMICS (__CUDA_ARCH__ >= 120)
namespace hist {
#if (!USE_SMEM_ATOMICS)
#define TAG_MASK ( (1U << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE)) - 1U )
__forceinline__ __device__ void addByte(volatile uint* s_WarpHist, uint data, uint threadTag)
namespace hist
{
uint count;
do
#if (!USE_SMEM_ATOMICS)
#define TAG_MASK ( (1U << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE)) - 1U )
__forceinline__ __device__ void addByte(volatile uint* s_WarpHist, uint data, uint threadTag)
{
uint count;
do
{
count = s_WarpHist[data] & TAG_MASK;
count = threadTag | (count + 1);
s_WarpHist[data] = count;
} while (s_WarpHist[data] != count);
}
#else
#define TAG_MASK 0xFFFFFFFFU
__forceinline__ __device__ void addByte(uint* s_WarpHist, uint data, uint threadTag)
{
atomicAdd(s_WarpHist + data, 1);
}
#endif
__forceinline__ __device__ void addWord(uint* s_WarpHist, uint data, uint tag, uint pos_x, uint cols)
{
count = s_WarpHist[data] & TAG_MASK;
count = threadTag | (count + 1);
s_WarpHist[data] = count;
} while (s_WarpHist[data] != count);
}
uint x = pos_x << 2;
#else
if (x + 0 < cols) addByte(s_WarpHist, (data >> 0) & 0xFFU, tag);
if (x + 1 < cols) addByte(s_WarpHist, (data >> 8) & 0xFFU, tag);
if (x + 2 < cols) addByte(s_WarpHist, (data >> 16) & 0xFFU, tag);
if (x + 3 < cols) addByte(s_WarpHist, (data >> 24) & 0xFFU, tag);
}
#define TAG_MASK 0xFFFFFFFFU
__global__ void histogram256(const PtrStep<uint> d_Data, uint* d_PartialHistograms, uint dataCount, uint cols)
{
//Per-warp subhistogram storage
__shared__ uint s_Hist[HISTOGRAM256_THREADBLOCK_MEMORY];
uint* s_WarpHist= s_Hist + (threadIdx.x >> OPENCV_GPU_LOG_WARP_SIZE) * HISTOGRAM256_BIN_COUNT;
__forceinline__ __device__ void addByte(uint* s_WarpHist, uint data, uint threadTag)
{
atomicAdd(s_WarpHist + data, 1);
}
//Clear shared memory storage for current threadblock before processing
#pragma unroll
for (uint i = 0; i < (HISTOGRAM256_THREADBLOCK_MEMORY / HISTOGRAM256_THREADBLOCK_SIZE); i++)
s_Hist[threadIdx.x + i * HISTOGRAM256_THREADBLOCK_SIZE] = 0;
#endif
//Cycle through the entire data set, update subhistograms for each warp
const uint tag = threadIdx.x << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE);
__forceinline__ __device__ void addWord(uint* s_WarpHist, uint data, uint tag, uint pos_x, uint cols)
{
uint x = pos_x << 2;
__syncthreads();
const uint colsui = d_Data.step / sizeof(uint);
for(uint pos = blockIdx.x * blockDim.x + threadIdx.x; pos < dataCount; pos += blockDim.x * gridDim.x)
{
uint pos_y = pos / colsui;
uint pos_x = pos % colsui;
uint data = d_Data.ptr(pos_y)[pos_x];
addWord(s_WarpHist, data, tag, pos_x, cols);
}
if (x + 0 < cols) addByte(s_WarpHist, (data >> 0) & 0xFFU, tag);
if (x + 1 < cols) addByte(s_WarpHist, (data >> 8) & 0xFFU, tag);
if (x + 2 < cols) addByte(s_WarpHist, (data >> 16) & 0xFFU, tag);
if (x + 3 < cols) addByte(s_WarpHist, (data >> 24) & 0xFFU, tag);
}
//Merge per-warp histograms into per-block and write to global memory
__syncthreads();
for(uint bin = threadIdx.x; bin < HISTOGRAM256_BIN_COUNT; bin += HISTOGRAM256_THREADBLOCK_SIZE)
{
uint sum = 0;
__global__ void histogram256(const PtrStep<uint> d_Data, uint* d_PartialHistograms, uint dataCount, uint cols)
{
//Per-warp subhistogram storage
__shared__ uint s_Hist[HISTOGRAM256_THREADBLOCK_MEMORY];
uint* s_WarpHist= s_Hist + (threadIdx.x >> OPENCV_GPU_LOG_WARP_SIZE) * HISTOGRAM256_BIN_COUNT;
for (uint i = 0; i < WARP_COUNT; i++)
sum += s_Hist[bin + i * HISTOGRAM256_BIN_COUNT] & TAG_MASK;
//Clear shared memory storage for current threadblock before processing
#pragma unroll
for (uint i = 0; i < (HISTOGRAM256_THREADBLOCK_MEMORY / HISTOGRAM256_THREADBLOCK_SIZE); i++)
s_Hist[threadIdx.x + i * HISTOGRAM256_THREADBLOCK_SIZE] = 0;
d_PartialHistograms[blockIdx.x * HISTOGRAM256_BIN_COUNT + bin] = sum;
}
}
//Cycle through the entire data set, update subhistograms for each warp
const uint tag = threadIdx.x << (UINT_BITS - OPENCV_GPU_LOG_WARP_SIZE);
////////////////////////////////////////////////////////////////////////////////
// Merge histogram256() output
// Run one threadblock per bin; each threadblock adds up the same bin counter
// from every partial histogram. Reads are uncoalesced, but mergeHistogram256
// takes only a fraction of total processing time
////////////////////////////////////////////////////////////////////////////////
__syncthreads();
const uint colsui = d_Data.step / sizeof(uint);
for(uint pos = blockIdx.x * blockDim.x + threadIdx.x; pos < dataCount; pos += blockDim.x * gridDim.x)
{
uint pos_y = pos / colsui;
uint pos_x = pos % colsui;
uint data = d_Data.ptr(pos_y)[pos_x];
addWord(s_WarpHist, data, tag, pos_x, cols);
}
__global__ void mergeHistogram256(const uint* d_PartialHistograms, int* d_Histogram)
{
uint sum = 0;
//Merge per-warp histograms into per-block and write to global memory
__syncthreads();
for(uint bin = threadIdx.x; bin < HISTOGRAM256_BIN_COUNT; bin += HISTOGRAM256_THREADBLOCK_SIZE)
{
uint sum = 0;
#pragma unroll
for (uint i = threadIdx.x; i < PARTIAL_HISTOGRAM256_COUNT; i += MERGE_THREADBLOCK_SIZE)
sum += d_PartialHistograms[blockIdx.x + i * HISTOGRAM256_BIN_COUNT];
for (uint i = 0; i < WARP_COUNT; i++)
sum += s_Hist[bin + i * HISTOGRAM256_BIN_COUNT] & TAG_MASK;
__shared__ uint data[MERGE_THREADBLOCK_SIZE];
data[threadIdx.x] = sum;
d_PartialHistograms[blockIdx.x * HISTOGRAM256_BIN_COUNT + bin] = sum;
}
}
for (uint stride = MERGE_THREADBLOCK_SIZE / 2; stride > 0; stride >>= 1)
{
__syncthreads();
if(threadIdx.x < stride)
data[threadIdx.x] += data[threadIdx.x + stride];
}
////////////////////////////////////////////////////////////////////////////////
// Merge histogram256() output
// Run one threadblock per bin; each threadblock adds up the same bin counter
// from every partial histogram. Reads are uncoalesced, but mergeHistogram256
// takes only a fraction of total processing time
////////////////////////////////////////////////////////////////////////////////
if(threadIdx.x == 0)
d_Histogram[blockIdx.x] = saturate_cast<int>(data[0]);
}
__global__ void mergeHistogram256(const uint* d_PartialHistograms, int* d_Histogram)
{
uint sum = 0;
void histogram256_gpu(DevMem2Db src, int* hist, uint* buf, cudaStream_t stream)
{
histogram256<<<PARTIAL_HISTOGRAM256_COUNT, HISTOGRAM256_THREADBLOCK_SIZE, 0, stream>>>(
DevMem2D_<uint>(src),
buf,
static_cast<uint>(src.rows * src.step / sizeof(uint)),
src.cols);
#pragma unroll
for (uint i = threadIdx.x; i < PARTIAL_HISTOGRAM256_COUNT; i += MERGE_THREADBLOCK_SIZE)
sum += d_PartialHistograms[blockIdx.x + i * HISTOGRAM256_BIN_COUNT];
cudaSafeCall( cudaGetLastError() );
__shared__ uint data[MERGE_THREADBLOCK_SIZE];
data[threadIdx.x] = sum;
mergeHistogram256<<<HISTOGRAM256_BIN_COUNT, MERGE_THREADBLOCK_SIZE, 0, stream>>>(buf, hist);
for (uint stride = MERGE_THREADBLOCK_SIZE / 2; stride > 0; stride >>= 1)
{
__syncthreads();
if(threadIdx.x < stride)
data[threadIdx.x] += data[threadIdx.x + stride];
}
cudaSafeCall( cudaGetLastError() );
if(threadIdx.x == 0)
d_Histogram[blockIdx.x] = saturate_cast<int>(data[0]);
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void histogram256_gpu(DevMem2Db src, int* hist, uint* buf, cudaStream_t stream)
{
histogram256<<<PARTIAL_HISTOGRAM256_COUNT, HISTOGRAM256_THREADBLOCK_SIZE, 0, stream>>>(
DevMem2D_<uint>(src),
buf,
static_cast<uint>(src.rows * src.step / sizeof(uint)),
src.cols);
__constant__ int c_lut[256];
cudaSafeCall( cudaGetLastError() );
__global__ void equalizeHist(const DevMem2Db src, PtrStepb dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
mergeHistogram256<<<HISTOGRAM256_BIN_COUNT, MERGE_THREADBLOCK_SIZE, 0, stream>>>(buf, hist);
if (x < src.cols && y < src.rows)
{
const uchar val = src.ptr(y)[x];
const int lut = c_lut[val];
dst.ptr(y)[x] = __float2int_rn(255.0f / (src.cols * src.rows) * lut);
}
}
cudaSafeCall( cudaGetLastError() );
void equalizeHist_gpu(DevMem2Db src, DevMem2Db dst, const int* lut, cudaStream_t stream)
{
dim3 block(16, 16);
dim3 grid(divUp(src.cols, block.x), divUp(src.rows, block.y));
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
cudaSafeCall( cudaMemcpyToSymbol(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice) );
__constant__ int c_lut[256];
equalizeHist<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
__global__ void equalizeHist(const DevMem2Db src, PtrStepb dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < src.cols && y < src.rows)
{
const uchar val = src.ptr(y)[x];
const int lut = c_lut[val];
dst.ptr(y)[x] = __float2int_rn(255.0f / (src.cols * src.rows) * lut);
}
}
void equalizeHist_gpu(DevMem2Db src, DevMem2Db dst, const int* lut, cudaStream_t stream)
{
dim3 block(16, 16);
dim3 grid(divUp(src.cols, block.x), divUp(src.rows, block.y));
cudaSafeCall( cudaMemcpyToSymbol(c_lut, lut, 256 * sizeof(int), 0, cudaMemcpyDeviceToDevice) );
equalizeHist<<<grid, block, 0, stream>>>(src, dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace hist
END_OPENCV_DEVICE_NAMESPACE
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace hist
}}} // namespace cv { namespace gpu { namespace device

1298
modules/gpu/src/cuda/hog.cu
View File

File diff suppressed because it is too large Load Diff

1870
modules/gpu/src/cuda/imgproc.cu
View File

File diff suppressed because it is too large Load Diff

111
modules/gpu/src/cuda/internal_shared.hpp
View File

@@ -50,7 +50,7 @@
#include "safe_call.hpp"
#ifndef CV_PI
#define CV_PI 3.1415926535897932384626433832795f
#define CV_PI 3.1415926535897932384626433832795
#endif
#ifndef CV_PI_F
@@ -61,27 +61,21 @@
#endif
#endif
#define BEGIN_OPENCV_DEVICE_NAMESPACE namespace cv { namespace gpu { namespace device {
#define END_OPENCV_DEVICE_NAMESPACE }}}
#define OPENCV_DEVICE_NAMESPACE ::cv::gpu::device
#define OPENCV_DEVICE_NAMESPACE_ ::cv::gpu::device::
#ifdef __CUDACC__
BEGIN_OPENCV_DEVICE_NAMESPACE
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef signed char schar;
typedef unsigned int uint;
template<class T> static inline void bindTexture(const textureReference* tex, const DevMem2D_<T>& img)
namespace cv { namespace gpu { namespace device
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
}
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef signed char schar;
typedef unsigned int uint;
END_OPENCV_DEVICE_NAMESPACE
template<class T> static inline void bindTexture(const textureReference* tex, const DevMem2D_<T>& img)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
}
}}}
#endif
@@ -102,87 +96,6 @@ namespace cv { namespace gpu
static inline int divUp(int total, int grain) { return (total + grain - 1) / grain; }
/*template<class T> static inline void uploadConstant(const char* name, const T& value)
{
cudaSafeCall( cudaMemcpyToSymbol(name, &value, sizeof(T)) );
}
template<class T> static inline void uploadConstant(const char* name, const T& value, cudaStream_t stream)
{
cudaSafeCall( cudaMemcpyToSymbolAsync(name, &value, sizeof(T), 0, cudaMemcpyHostToDevice, stream) );
} */
//template<class T> static inline void bindTexture(const char* name, const DevMem2D_<T>& img)
//{
// //!!!! const_cast is disabled!
// //!!!! Please use constructor of 'class texture' instead.
//
// //textureReference* tex;
// //cudaSafeCall( cudaGetTextureReference((const textureReference**)&tex, name) );
// //tex->normalized = normalized;
// //tex->filterMode = filterMode;
// //tex->addressMode[0] = addrMode;
// //tex->addressMode[1] = addrMode;
//
// const textureReference* tex;
// cudaSafeCall( cudaGetTextureReference(&tex, name) );
//
// cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
// cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
//}
//static inline void unbindTexture(const char *name)
//{
// const textureReference* tex;
// cudaSafeCall( cudaGetTextureReference(&tex, name) );
// cudaSafeCall( cudaUnbindTexture(tex) );
//}
//class TextureBinder
//{
//public:
// TextureBinder() : tex_(0) {}
// template <typename T> TextureBinder(const textureReference* tex, const DevMem2D_<T>& img) : tex_(0)
// {
// bind(tex, img);
// }
// template <typename T> TextureBinder(const char* tex_name, const DevMem2D_<T>& img) : tex_(0)
// {
// bind(tex_name, img);
// }
// ~TextureBinder() { unbind(); }
//
// template <typename T> void bind(const textureReference* tex, const DevMem2D_<T>& img)
// {
// unbind();
//
// cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
// cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
//
// tex_ = tex;
// }
// template <typename T> void bind(const char* tex_name, const DevMem2D_<T>& img)
// {
// const textureReference* tex;
// cudaSafeCall( cudaGetTextureReference(&tex, tex_name) );
// bind(tex, img);
// }
//
// void unbind()
// {
// if (tex_)
// {
// cudaUnbindTexture(tex_);
// tex_ = 0;
// }
// }
//
//private:
// const textureReference* tex_;
//};
class NppStreamHandler
{
public:

1710
modules/gpu/src/cuda/match_template.cu
View File

File diff suppressed because it is too large Load Diff

300
modules/gpu/src/cuda/mathfunc.cu
View File

@@ -42,174 +42,172 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace mathfunc {
//////////////////////////////////////////////////////////////////////////////////////
// Cart <-> Polar
struct Nothing
namespace cv { namespace gpu { namespace device
{
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
namespace mathfunc
{
}
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = ::sqrtf(x_data * x_data + y_data * y_data);
}
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = ::atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
//////////////////////////////////////////////////////////////////////////////////////
// Cart <-> Polar
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
struct NonEmptyMag
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
{
return mag[y * mag_step + x];
}
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
::sincosf(scale * angle_data, &sin_a, &cos_a);
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
}
template <typename Mag, typename Angle>
void cartToPolar_caller(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(x.cols, threads.x);
grid.y = divUp(x.rows, threads.y);
const float scale = angleInDegrees ? (float)(180.0f / CV_PI) : 1.f;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void cartToPolar_gpu(DevMem2Df x, DevMem2Df y, DevMem2Df mag, bool magSqr, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
struct Nothing
{
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
{
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
}
},
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
dst[y * dst_step + x] = ::sqrtf(x_data * x_data + y_data * y_data);
}
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = ::atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
};
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
struct NonEmptyMag
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
{
return mag[y * mag_step + x];
}
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
template <typename Mag>
void polarToCart_caller(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
if (x < width && y < height)
{
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
grid.x = divUp(mag.cols, threads.x);
grid.y = divUp(mag.rows, threads.y);
const float scale = angleInDegrees ? (float)(CV_PI / 180.0f) : 1.0f;
::sincosf(scale * angle_data, &sin_a, &cos_a);
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename Mag, typename Angle>
void cartToPolar_caller(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
void polarToCart_gpu(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
grid.x = divUp(x.cols, threads.x);
grid.y = divUp(x.rows, threads.y);
const float scale = angleInDegrees ? (float)(180.0f / CV_PI) : 1.f;
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
} // namespace mathfunc
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
END_OPENCV_DEVICE_NAMESPACE
void cartToPolar_gpu(DevMem2Df x, DevMem2Df y, DevMem2Df mag, bool magSqr, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
{
{
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
}
},
{
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
}
}
};
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
template <typename Mag>
void polarToCart_caller(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(mag.cols, threads.x);
grid.y = divUp(mag.rows, threads.y);
const float scale = angleInDegrees ? (float)(CV_PI / 180.0f) : 1.0f;
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void polarToCart_gpu(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
} // namespace mathfunc
}}} // namespace cv { namespace gpu { namespace device

529
modules/gpu/src/cuda/matrix_operations.cu
View File

@@ -45,304 +45,303 @@
#include "opencv2/gpu/device/transform.hpp"
#include "opencv2/gpu/device/functional.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
template <typename T> struct shift_and_sizeof;
template <> struct shift_and_sizeof<signed char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<unsigned char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<unsigned short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<int> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<float> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<double> { enum { shift = 3 }; };
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// CopyTo /////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template<typename T>
__global__ void copy_to_with_mask(const T* mat_src, T* mat_dst, const uchar* mask, int cols, int rows, size_t step_mat, size_t step_mask, int channels)
namespace cv { namespace gpu { namespace device
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
template <typename T> struct shift_and_sizeof;
template <> struct shift_and_sizeof<signed char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<unsigned char> { enum { shift = 0 }; };
template <> struct shift_and_sizeof<short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<unsigned short> { enum { shift = 1 }; };
template <> struct shift_and_sizeof<int> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<float> { enum { shift = 2 }; };
template <> struct shift_and_sizeof<double> { enum { shift = 3 }; };
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
{
size_t idx = y * ( step_mat >> shift_and_sizeof<T>::shift ) + x;
mat_dst[idx] = mat_src[idx];
}
}
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// CopyTo /////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template<typename T>
void copy_to_with_mask_run(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
dim3 threadsPerBlock(16,16, 1);
dim3 numBlocks ( divUp(mat_src.cols * channels , threadsPerBlock.x) , divUp(mat_src.rows , threadsPerBlock.y), 1);
copy_to_with_mask<T><<<numBlocks,threadsPerBlock, 0, stream>>>
((T*)mat_src.data, (T*)mat_dst.data, (unsigned char*)mask.data, mat_src.cols, mat_src.rows, mat_src.step, mask.step, channels);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
void copy_to_with_mask(const DevMem2Db& mat_src, DevMem2Db mat_dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
typedef void (*CopyToFunc)(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream);
static CopyToFunc tab[8] =
template<typename T>
__global__ void copy_to_with_mask(const T* mat_src, T* mat_dst, const uchar* mask, int cols, int rows, size_t step_mat, size_t step_mask, int channels)
{
copy_to_with_mask_run<unsigned char>,
copy_to_with_mask_run<signed char>,
copy_to_with_mask_run<unsigned short>,
copy_to_with_mask_run<short>,
copy_to_with_mask_run<int>,
copy_to_with_mask_run<float>,
copy_to_with_mask_run<double>,
0
};
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
CopyToFunc func = tab[depth];
if (func == 0) cv::gpu::error("Unsupported copyTo operation", __FILE__, __LINE__);
func(mat_src, mat_dst, mask, channels, stream);
}
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// SetTo //////////////////////////////////
///////////////////////////////////////////////////////////////////////////
__constant__ uchar scalar_8u[4];
__constant__ schar scalar_8s[4];
__constant__ ushort scalar_16u[4];
__constant__ short scalar_16s[4];
__constant__ int scalar_32s[4];
__constant__ float scalar_32f[4];
__constant__ double scalar_64f[4];
template <typename T> __device__ __forceinline__ T readScalar(int i);
template <> __device__ __forceinline__ uchar readScalar<uchar>(int i) {return scalar_8u[i];}
template <> __device__ __forceinline__ schar readScalar<schar>(int i) {return scalar_8s[i];}
template <> __device__ __forceinline__ ushort readScalar<ushort>(int i) {return scalar_16u[i];}
template <> __device__ __forceinline__ short readScalar<short>(int i) {return scalar_16s[i];}
template <> __device__ __forceinline__ int readScalar<int>(int i) {return scalar_32s[i];}
template <> __device__ __forceinline__ float readScalar<float>(int i) {return scalar_32f[i];}
template <> __device__ __forceinline__ double readScalar<double>(int i) {return scalar_64f[i];}
void writeScalar(const uchar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8u, vals, sizeof(uchar) * 4) );
}
void writeScalar(const schar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8s, vals, sizeof(schar) * 4) );
}
void writeScalar(const ushort* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16u, vals, sizeof(ushort) * 4) );
}
void writeScalar(const short* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16s, vals, sizeof(short) * 4) );
}
void writeScalar(const int* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_32s, vals, sizeof(int) * 4) );
}
void writeScalar(const float* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_32f, vals, sizeof(float) * 4) );
}
void writeScalar(const double* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_64f, vals, sizeof(double) * 4) );
}
template<typename T>
__global__ void set_to_without_mask(T* mat, int cols, int rows, size_t step, int channels)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
if ((x < cols * channels ) && (y < rows))
{
size_t idx = y * ( step >> shift_and_sizeof<T>::shift ) + x;
mat[idx] = readScalar<T>(x % channels);
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
{
size_t idx = y * ( step_mat >> shift_and_sizeof<T>::shift ) + x;
mat_dst[idx] = mat_src[idx];
}
}
}
template<typename T>
__global__ void set_to_with_mask(T* mat, const uchar* mask, int cols, int rows, size_t step, int channels, size_t step_mask)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
template<typename T>
void copy_to_with_mask_run(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
dim3 threadsPerBlock(16,16, 1);
dim3 numBlocks ( divUp(mat_src.cols * channels , threadsPerBlock.x) , divUp(mat_src.rows , threadsPerBlock.y), 1);
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
copy_to_with_mask<T><<<numBlocks,threadsPerBlock, 0, stream>>>
((T*)mat_src.data, (T*)mat_dst.data, (unsigned char*)mask.data, mat_src.cols, mat_src.rows, mat_src.step, mask.step, channels);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
void copy_to_with_mask(const DevMem2Db& mat_src, DevMem2Db mat_dst, int depth, const DevMem2Db& mask, int channels, const cudaStream_t & stream)
{
typedef void (*CopyToFunc)(const DevMem2Db& mat_src, const DevMem2Db& mat_dst, const DevMem2Db& mask, int channels, const cudaStream_t & stream);
static CopyToFunc tab[8] =
{
copy_to_with_mask_run<unsigned char>,
copy_to_with_mask_run<signed char>,
copy_to_with_mask_run<unsigned short>,
copy_to_with_mask_run<short>,
copy_to_with_mask_run<int>,
copy_to_with_mask_run<float>,
copy_to_with_mask_run<double>,
0
};
CopyToFunc func = tab[depth];
if (func == 0) cv::gpu::error("Unsupported copyTo operation", __FILE__, __LINE__);
func(mat_src, mat_dst, mask, channels, stream);
}
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// SetTo //////////////////////////////////
///////////////////////////////////////////////////////////////////////////
__constant__ uchar scalar_8u[4];
__constant__ schar scalar_8s[4];
__constant__ ushort scalar_16u[4];
__constant__ short scalar_16s[4];
__constant__ int scalar_32s[4];
__constant__ float scalar_32f[4];
__constant__ double scalar_64f[4];
template <typename T> __device__ __forceinline__ T readScalar(int i);
template <> __device__ __forceinline__ uchar readScalar<uchar>(int i) {return scalar_8u[i];}
template <> __device__ __forceinline__ schar readScalar<schar>(int i) {return scalar_8s[i];}
template <> __device__ __forceinline__ ushort readScalar<ushort>(int i) {return scalar_16u[i];}
template <> __device__ __forceinline__ short readScalar<short>(int i) {return scalar_16s[i];}
template <> __device__ __forceinline__ int readScalar<int>(int i) {return scalar_32s[i];}
template <> __device__ __forceinline__ float readScalar<float>(int i) {return scalar_32f[i];}
template <> __device__ __forceinline__ double readScalar<double>(int i) {return scalar_64f[i];}
void writeScalar(const uchar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8u, vals, sizeof(uchar) * 4) );
}
void writeScalar(const schar* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_8s, vals, sizeof(schar) * 4) );
}
void writeScalar(const ushort* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16u, vals, sizeof(ushort) * 4) );
}
void writeScalar(const short* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_16s, vals, sizeof(short) * 4) );
}
void writeScalar(const int* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_32s, vals, sizeof(int) * 4) );
}
void writeScalar(const float* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_32f, vals, sizeof(float) * 4) );
}
void writeScalar(const double* vals)
{
cudaSafeCall( cudaMemcpyToSymbol(scalar_64f, vals, sizeof(double) * 4) );
}
template<typename T>
__global__ void set_to_without_mask(T* mat, int cols, int rows, size_t step, int channels)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
if ((x < cols * channels ) && (y < rows))
{
size_t idx = y * ( step >> shift_and_sizeof<T>::shift ) + x;
mat[idx] = readScalar<T>(x % channels);
}
}
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream)
{
writeScalar(scalar);
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
set_to_with_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, (uchar*)mask.data, mat.cols, mat.rows, mat.step, channels, mask.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream)
{
writeScalar(scalar);
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
set_to_without_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, mat.cols, mat.rows, mat.step, channels);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, int channels, cudaStream_t stream);
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template <typename T, typename D> struct Convertor : unary_function<T, D>
{
Convertor(double alpha_, double beta_) : alpha(alpha_), beta(beta_) {}
__device__ __forceinline__ D operator()(const T& src) const
{
return saturate_cast<D>(alpha * src + beta);
}
const double alpha, beta;
};
template<typename T>
__global__ void set_to_with_mask(T* mat, const uchar* mask, int cols, int rows, size_t step, int channels, size_t step_mask)
{
size_t x = blockIdx.x * blockDim.x + threadIdx.x;
size_t y = blockIdx.y * blockDim.y + threadIdx.y;
namespace detail
{
template <size_t src_size, size_t dst_size, typename F> struct ConvertTraitsDispatcher : DefaultTransformFunctorTraits<F>
if ((x < cols * channels ) && (y < rows))
if (mask[y * step_mask + x / channels] != 0)
{
size_t idx = y * ( step >> shift_and_sizeof<T>::shift ) + x;
mat[idx] = readScalar<T>(x % channels);
}
}
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream)
{
};
template <typename F> struct ConvertTraitsDispatcher<1, 1, F> : DefaultTransformFunctorTraits<F>
writeScalar(scalar);
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
set_to_with_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, (uchar*)mask.data, mat.cols, mat.rows, mat.step, channels, mask.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, const DevMem2Db& mask, int channels, cudaStream_t stream);
template <typename T>
void set_to_gpu(const DevMem2Db& mat, const T* scalar, int channels, cudaStream_t stream)
{
enum { smart_shift = 8 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 2, F> : DefaultTransformFunctorTraits<F>
writeScalar(scalar);
dim3 threadsPerBlock(32, 8, 1);
dim3 numBlocks (mat.cols * channels / threadsPerBlock.x + 1, mat.rows / threadsPerBlock.y + 1, 1);
set_to_without_mask<T><<<numBlocks, threadsPerBlock, 0, stream>>>((T*)mat.data, mat.cols, mat.rows, mat.step, channels);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall ( cudaDeviceSynchronize() );
}
template void set_to_gpu<uchar >(const DevMem2Db& mat, const uchar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<schar >(const DevMem2Db& mat, const schar* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<ushort>(const DevMem2Db& mat, const ushort* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<short >(const DevMem2Db& mat, const short* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<int >(const DevMem2Db& mat, const int* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<float >(const DevMem2Db& mat, const float* scalar, int channels, cudaStream_t stream);
template void set_to_gpu<double>(const DevMem2Db& mat, const double* scalar, int channels, cudaStream_t stream);
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
template <typename T, typename D> struct Convertor : unary_function<T, D>
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
Convertor(double alpha_, double beta_) : alpha(alpha_), beta(beta_) {}
__device__ __forceinline__ D operator()(const T& src) const
{
return saturate_cast<D>(alpha * src + beta);
}
const double alpha, beta;
};
template <typename F> struct ConvertTraitsDispatcher<2, 2, F> : DefaultTransformFunctorTraits<F>
namespace detail
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<2, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 2 };
};
template <size_t src_size, size_t dst_size, typename F> struct ConvertTraitsDispatcher : DefaultTransformFunctorTraits<F>
{
};
template <typename F> struct ConvertTraitsDispatcher<1, 1, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 8 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<1, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<4, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<4, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 2 };
};
template <typename F> struct ConvertTraitsDispatcher<2, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<2, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_shift = 2 };
};
template <typename F> struct ConvertTraits : ConvertTraitsDispatcher<sizeof(typename F::argument_type), sizeof(typename F::result_type), F>
template <typename F> struct ConvertTraitsDispatcher<4, 2, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 4 };
};
template <typename F> struct ConvertTraitsDispatcher<4, 4, F> : DefaultTransformFunctorTraits<F>
{
enum { smart_block_dim_y = 8 };
enum { smart_shift = 2 };
};
template <typename F> struct ConvertTraits : ConvertTraitsDispatcher<sizeof(typename F::argument_type), sizeof(typename F::result_type), F>
{
};
}
template <typename T, typename D> struct TransformFunctorTraits< Convertor<T, D> > : detail::ConvertTraits< Convertor<T, D> >
{
};
}
template <typename T, typename D> struct TransformFunctorTraits< Convertor<T, D> > : detail::ConvertTraits< Convertor<T, D> >
{
};
template<typename T, typename D>
void cvt_(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta, cudaStream_t stream)
{
cudaSafeCall( cudaSetDoubleForDevice(&alpha) );
cudaSafeCall( cudaSetDoubleForDevice(&beta) );
Convertor<T, D> op(alpha, beta);
OPENCV_DEVICE_NAMESPACE_ transform((DevMem2D_<T>)src, (DevMem2D_<D>)dst, op, stream);
}
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta,
cudaStream_t stream = 0)
{
typedef void (*caller_t)(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta,
cudaStream_t stream);
static const caller_t tab[8][8] =
template<typename T, typename D>
void cvt_(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta, cudaStream_t stream)
{
{cvt_<uchar, uchar>, cvt_<uchar, schar>, cvt_<uchar, ushort>, cvt_<uchar, short>,
cvt_<uchar, int>, cvt_<uchar, float>, cvt_<uchar, double>, 0},
cudaSafeCall( cudaSetDoubleForDevice(&alpha) );
cudaSafeCall( cudaSetDoubleForDevice(&beta) );
Convertor<T, D> op(alpha, beta);
::cv::gpu::device::transform((DevMem2D_<T>)src, (DevMem2D_<D>)dst, op, stream);
}
{cvt_<schar, uchar>, cvt_<schar, schar>, cvt_<schar, ushort>, cvt_<schar, short>,
cvt_<schar, int>, cvt_<schar, float>, cvt_<schar, double>, 0},
void convert_gpu(const DevMem2Db& src, int sdepth, const DevMem2Db& dst, int ddepth, double alpha, double beta,
cudaStream_t stream = 0)
{
typedef void (*caller_t)(const DevMem2Db& src, const DevMem2Db& dst, double alpha, double beta,
cudaStream_t stream);
{cvt_<ushort, uchar>, cvt_<ushort, schar>, cvt_<ushort, ushort>, cvt_<ushort, short>,
cvt_<ushort, int>, cvt_<ushort, float>, cvt_<ushort, double>, 0},
static const caller_t tab[8][8] =
{
{cvt_<uchar, uchar>, cvt_<uchar, schar>, cvt_<uchar, ushort>, cvt_<uchar, short>,
cvt_<uchar, int>, cvt_<uchar, float>, cvt_<uchar, double>, 0},
{cvt_<short, uchar>, cvt_<short, schar>, cvt_<short, ushort>, cvt_<short, short>,
cvt_<short, int>, cvt_<short, float>, cvt_<short, double>, 0},
{cvt_<schar, uchar>, cvt_<schar, schar>, cvt_<schar, ushort>, cvt_<schar, short>,
cvt_<schar, int>, cvt_<schar, float>, cvt_<schar, double>, 0},
{cvt_<int, uchar>, cvt_<int, schar>, cvt_<int, ushort>,
cvt_<int, short>, cvt_<int, int>, cvt_<int, float>, cvt_<int, double>, 0},
{cvt_<ushort, uchar>, cvt_<ushort, schar>, cvt_<ushort, ushort>, cvt_<ushort, short>,
cvt_<ushort, int>, cvt_<ushort, float>, cvt_<ushort, double>, 0},
{cvt_<float, uchar>, cvt_<float, schar>, cvt_<float, ushort>,
cvt_<float, short>, cvt_<float, int>, cvt_<float, float>, cvt_<float, double>, 0},
{cvt_<short, uchar>, cvt_<short, schar>, cvt_<short, ushort>, cvt_<short, short>,
cvt_<short, int>, cvt_<short, float>, cvt_<short, double>, 0},
{cvt_<double, uchar>, cvt_<double, schar>, cvt_<double, ushort>,
cvt_<double, short>, cvt_<double, int>, cvt_<double, float>, cvt_<double, double>, 0},
{cvt_<int, uchar>, cvt_<int, schar>, cvt_<int, ushort>,
cvt_<int, short>, cvt_<int, int>, cvt_<int, float>, cvt_<int, double>, 0},
{0,0,0,0,0,0,0,0}
};
{cvt_<float, uchar>, cvt_<float, schar>, cvt_<float, ushort>,
cvt_<float, short>, cvt_<float, int>, cvt_<float, float>, cvt_<float, double>, 0},
caller_t func = tab[sdepth][ddepth];
if (!func)
cv::gpu::error("Unsupported convert operation", __FILE__, __LINE__);
{cvt_<double, uchar>, cvt_<double, schar>, cvt_<double, ushort>,
cvt_<double, short>, cvt_<double, int>, cvt_<double, float>, cvt_<double, double>, 0},
func(src, dst, alpha, beta, stream);
}
{0,0,0,0,0,0,0,0}
};
END_OPENCV_DEVICE_NAMESPACE
caller_t func = tab[sdepth][ddepth];
if (!func)
cv::gpu::error("Unsupported convert operation", __FILE__, __LINE__);
func(src, dst, alpha, beta, stream);
}
}}} // namespace cv { namespace gpu { namespace device

3864
modules/gpu/src/cuda/matrix_reductions.cu
View File

File diff suppressed because it is too large Load Diff

218
modules/gpu/src/cuda/pyr_down.cu
View File

@@ -46,142 +46,140 @@
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename T, typename B> __global__ void pyrDown(const PtrStep<T> src, PtrStep<T> dst, const B b, int dst_cols)
namespace cv { namespace gpu { namespace device
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y;
__shared__ value_type smem[256 + 4];
value_type sum;
const int src_y = 2*y;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, x, src.data, src.step);
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);
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;
if (threadIdx.x < 2)
namespace imgproc
{
const int left_x = x - 2 + threadIdx.x;
template <typename T, typename B> __global__ void pyrDown(const PtrStep<T> src, PtrStep<T> dst, const B b, int dst_cols)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, left_x, src.data, src.step);
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);
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);
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y;
smem[threadIdx.x] = sum;
}
__shared__ value_type smem[256 + 4];
if (threadIdx.x > 253)
{
const int right_x = x + threadIdx.x + 2;
value_type sum;
const int src_y = 2*y;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, right_x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, right_x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, right_x, src.data, src.step);
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, x, src.data, src.step);
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);
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[4 + threadIdx.x] = sum;
}
smem[2 + threadIdx.x] = sum;
__syncthreads();
if (threadIdx.x < 2)
{
const int left_x = x - 2 + threadIdx.x;
if (threadIdx.x < 128)
{
const int tid2 = threadIdx.x * 2;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, left_x, src.data, src.step);
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);
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);
sum = VecTraits<value_type>::all(0);
smem[threadIdx.x] = sum;
}
sum = sum + 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
sum = sum + 0.0625f * smem[2 + tid2 + 2];
if (threadIdx.x > 253)
{
const int right_x = x + threadIdx.x + 2;
const int dst_x = (blockIdx.x * blockDim.x + tid2) / 2;
sum = VecTraits<value_type>::all(0);
sum = sum + 0.0625f * b.at(src_y - 2, right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y - 1, right_x, src.data, src.step);
sum = sum + 0.375f * b.at(src_y , right_x, src.data, src.step);
sum = sum + 0.25f * b.at(src_y + 1, right_x, src.data, src.step);
sum = sum + 0.0625f * b.at(src_y + 2, right_x, src.data, src.step);
if (dst_x < dst_cols)
dst.ptr(y)[dst_x] = saturate_cast<T>(sum);
}
}
smem[4 + threadIdx.x] = sum;
}
template <typename T, template <typename> class B> void pyrDown_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
{
const dim3 block(256);
const dim3 grid(divUp(src.cols, block.x), dst.rows);
__syncthreads();
B<T> b(src.rows, src.cols);
if (threadIdx.x < 128)
{
const int tid2 = threadIdx.x * 2;
pyrDown<T><<<grid, block, 0, stream>>>(src, dst, b, dst.cols);
cudaSafeCall( cudaGetLastError() );
sum = VecTraits<value_type>::all(0);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
sum = sum + 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
sum = sum + 0.0625f * smem[2 + tid2 + 2];
template <typename T, int cn> void pyrDown_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
const int dst_x = (blockIdx.x * blockDim.x + tid2) / 2;
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
if (dst_x < dst_cols)
dst.ptr(y)[dst_x] = saturate_cast<T>(sum);
}
}
static const caller_t callers[] =
{
pyrDown_caller<type, BrdReflect101>, pyrDown_caller<type, BrdReplicate>, pyrDown_caller<type, BrdConstant>, pyrDown_caller<type, BrdReflect>, pyrDown_caller<type, BrdWrap>
};
template <typename T, template <typename> class B> void pyrDown_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
{
const dim3 block(256);
const dim3 grid(divUp(src.cols, block.x), dst.rows);
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
B<T> b(src.rows, src.cols);
template void pyrDown_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
pyrDown<T><<<grid, block, 0, stream>>>(src, dst, b, dst.cols);
cudaSafeCall( cudaGetLastError() );
template void pyrDown_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void pyrDown_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template <typename T, int cn> void pyrDown_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
template void pyrDown_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
template void pyrDown_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
static const caller_t callers[] =
{
pyrDown_caller<type, BrdReflect101>, pyrDown_caller<type, BrdReplicate>, pyrDown_caller<type, BrdConstant>, pyrDown_caller<type, BrdReflect>, pyrDown_caller<type, BrdWrap>
};
template void pyrDown_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
} // namespace imgproc
template void pyrDown_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
END_OPENCV_DEVICE_NAMESPACE
template void pyrDown_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrDown_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

212
modules/gpu/src/cuda/pyr_up.cu
View File

@@ -46,137 +46,135 @@
#include "opencv2/gpu/device/vec_math.hpp"
#include "opencv2/gpu/device/saturate_cast.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename T, typename B> __global__ void pyrUp(const PtrStep<T> src, DevMem2D_<T> dst, const B b)
namespace cv { namespace gpu { namespace device
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
__shared__ T smem1[10][10];
__shared__ value_type smem2[20][16];
value_type sum;
if (threadIdx.x < 10 && threadIdx.y < 10)
smem1[threadIdx.y][threadIdx.x] = b.at(blockIdx.y * blockDim.y / 2 + threadIdx.y - 1, blockIdx.x * blockDim.x / 2 + threadIdx.x - 1, src.data, src.step);
__syncthreads();
const int tidx = threadIdx.x;
sum = VecTraits<value_type>::all(0);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[1 + threadIdx.y / 2][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 2) >> 1)];
smem2[2 + threadIdx.y][tidx] = sum;
if (threadIdx.y < 2)
namespace imgproc
{
sum = VecTraits<value_type>::all(0);
template <typename T, typename B> __global__ void pyrUp(const PtrStep<T> src, DevMem2D_<T> dst, const B b)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type value_type;
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[0][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx + 2) >> 1)];
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
smem2[threadIdx.y][tidx] = sum;
}
__shared__ T smem1[10][10];
__shared__ value_type smem2[20][16];
if (threadIdx.y > 13)
{
sum = VecTraits<value_type>::all(0);
value_type sum;
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[9][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx + 2) >> 1)];
if (threadIdx.x < 10 && threadIdx.y < 10)
smem1[threadIdx.y][threadIdx.x] = b.at(blockIdx.y * blockDim.y / 2 + threadIdx.y - 1, blockIdx.x * blockDim.x / 2 + threadIdx.x - 1, src.data, src.step);
smem2[4 + threadIdx.y][tidx] = sum;
}
__syncthreads();
__syncthreads();
const int tidx = threadIdx.x;
sum = VecTraits<value_type>::all(0);
sum = VecTraits<value_type>::all(0);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y - 2][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y - 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.375f * smem2[2 + threadIdx.y ][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y + 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y + 2][tidx];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[1 + threadIdx.y / 2][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[1 + threadIdx.y / 2][1 + ((tidx + 2) >> 1)];
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = saturate_cast<T>(4.0f * sum);
}
smem2[2 + threadIdx.y][tidx] = sum;
template <typename T, template <typename> class B> void pyrUp_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
{
const dim3 block(16, 16);
const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
if (threadIdx.y < 2)
{
sum = VecTraits<value_type>::all(0);
B<T> b(src.rows, src.cols);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[0][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[0][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[0][1 + ((tidx + 2) >> 1)];
pyrUp<T><<<grid, block, 0, stream>>>(src, dst, b);
cudaSafeCall( cudaGetLastError() );
smem2[threadIdx.y][tidx] = sum;
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (threadIdx.y > 13)
{
sum = VecTraits<value_type>::all(0);
template <typename T, int cn> void pyrUp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx - 2) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx - 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.375f * smem1[9][1 + ((tidx ) >> 1)];
sum = sum + (tidx % 2 != 0) * 0.25f * smem1[9][1 + ((tidx + 1) >> 1)];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem1[9][1 + ((tidx + 2) >> 1)];
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
smem2[4 + threadIdx.y][tidx] = sum;
}
static const caller_t callers[] =
{
pyrUp_caller<type, BrdReflect101>, pyrUp_caller<type, BrdReplicate>, pyrUp_caller<type, BrdConstant>, pyrUp_caller<type, BrdReflect>, pyrUp_caller<type, BrdWrap>
};
__syncthreads();
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
sum = VecTraits<value_type>::all(0);
template void pyrUp_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y - 2][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y - 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.375f * smem2[2 + threadIdx.y ][tidx];
sum = sum + (tidx % 2 != 0) * 0.25f * smem2[2 + threadIdx.y + 1][tidx];
sum = sum + (tidx % 2 == 0) * 0.0625f * smem2[2 + threadIdx.y + 2][tidx];
template void pyrUp_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
if (x < dst.cols && y < dst.rows)
dst.ptr(y)[x] = saturate_cast<T>(4.0f * sum);
}
template void pyrUp_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template <typename T, template <typename> class B> void pyrUp_caller(const DevMem2D_<T>& src, const DevMem2D_<T>& dst, cudaStream_t stream)
{
const dim3 block(16, 16);
const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
template void pyrUp_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
B<T> b(src.rows, src.cols);
template void pyrUp_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
pyrUp<T><<<grid, block, 0, stream>>>(src, dst, b);
cudaSafeCall( cudaGetLastError() );
template void pyrUp_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
} // namespace imgproc
template <typename T, int cn> void pyrUp_gpu(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream)
{
typedef typename TypeVec<T, cn>::vec_type type;
END_OPENCV_DEVICE_NAMESPACE
typedef void (*caller_t)(const DevMem2D_<type>& src, const DevMem2D_<type>& dst, cudaStream_t stream);
static const caller_t callers[] =
{
pyrUp_caller<type, BrdReflect101>, pyrUp_caller<type, BrdReplicate>, pyrUp_caller<type, BrdConstant>, pyrUp_caller<type, BrdReflect>, pyrUp_caller<type, BrdWrap>
};
callers[borderType](static_cast< DevMem2D_<type> >(src), static_cast< DevMem2D_<type> >(dst), stream);
}
template void pyrUp_gpu<uchar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<uchar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<schar, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<ushort, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<short, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<int, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 1>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 2>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 3>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
template void pyrUp_gpu<float, 4>(const DevMem2Db& src, const DevMem2Db& dst, int borderType, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

382
modules/gpu/src/cuda/remap.cu
View File

@@ -47,208 +47,206 @@
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename Ptr2D, typename T> __global__ void remap(const Ptr2D src, const PtrStepf mapx, const PtrStepf mapy, DevMem2D_<T> dst)
namespace cv { namespace gpu { namespace device
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void remap(const Ptr2D src, const PtrStepf mapx, const PtrStepf mapy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = mapx.ptr(y)[x];
const float ycoo = mapy.ptr(y)[x];
if (x < dst.cols && y < dst.rows)
{
const float xcoo = mapx.ptr(y)[x];
const float ycoo = mapy.ptr(y)[x];
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherStream
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block, 0, stream>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst, const float* borderValue, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
remap<<<grid, block>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_REMAP_TEX(type) \
texture< type , cudaTextureType2D> tex_remap_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_remap_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_remap_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter, template <typename> class B> struct RemapDispatcherNonStream<Filter, B, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_remap_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <template <typename> class Filter> struct RemapDispatcherNonStream<Filter, BrdReplicate, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
Filter< tex_remap_ ## type ##_reader > filter_src(texSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(short2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(float2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_REMAP_TEX
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcher
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
RemapDispatcherNonStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, cc);
else
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc);
}
};
template <typename T> void remap_gpu(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc);
static const caller_t callers[3][5] =
{
{
RemapDispatcher<PointFilter, BrdReflect101, T>::call,
RemapDispatcher<PointFilter, BrdReplicate, T>::call,
RemapDispatcher<PointFilter, BrdConstant, T>::call,
RemapDispatcher<PointFilter, BrdReflect, T>::call,
RemapDispatcher<PointFilter, BrdWrap, T>::call
},
{
RemapDispatcher<LinearFilter, BrdReflect101, T>::call,
RemapDispatcher<LinearFilter, BrdReplicate, T>::call,
RemapDispatcher<LinearFilter, BrdConstant, T>::call,
RemapDispatcher<LinearFilter, BrdReflect, T>::call,
RemapDispatcher<LinearFilter, BrdWrap, T>::call
},
{
RemapDispatcher<CubicFilter, BrdReflect101, T>::call,
RemapDispatcher<CubicFilter, BrdReplicate, T>::call,
RemapDispatcher<CubicFilter, BrdConstant, T>::call,
RemapDispatcher<CubicFilter, BrdReflect, T>::call,
RemapDispatcher<CubicFilter, BrdWrap, T>::call
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
};
callers[interpolation][borderMode](static_cast< DevMem2D_<T> >(src), xmap, ymap, static_cast< DevMem2D_<T> >(dst), borderValue, stream, cc);
}
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherStream
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
template void remap_gpu<uchar >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<uchar2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
//template void remap_gpu<schar>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
remap<<<grid, block, 0, stream>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template void remap_gpu<ushort >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<ushort2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst, const float* borderValue, int)
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
template void remap_gpu<short >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<short2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue));
BorderReader< PtrStep<T>, B<work_type> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, B<work_type> > > filter_src(brdSrc);
//template void remap_gpu<int >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
remap<<<grid, block>>>(filter_src, mapx, mapy, dst);
cudaSafeCall( cudaGetLastError() );
template void remap_gpu<float >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<float2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
cudaSafeCall( cudaDeviceSynchronize() );
}
};
} // namespace imgproc
#define OPENCV_GPU_IMPLEMENT_REMAP_TEX(type) \
texture< type , cudaTextureType2D> tex_remap_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_remap_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_remap_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter, template <typename> class B> struct RemapDispatcherNonStream<Filter, B, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float* borderValue, int cc) \
{ \
typedef typename TypeVec<float, VecTraits< type >::cn>::vec_type work_type; \
dim3 block(32, cc >= 20 ? 8 : 4); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
B<work_type> brd(src.rows, src.cols, VecTraits<work_type>::make(borderValue)); \
BorderReader< tex_remap_ ## type ##_reader, B<work_type> > brdSrc(texSrc, brd); \
Filter< BorderReader< tex_remap_ ## type ##_reader, B<work_type> > > filter_src(brdSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <template <typename> class Filter> struct RemapDispatcherNonStream<Filter, BrdReplicate, type> \
{ \
static void call(const DevMem2D_< type >& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_< type >& dst, const float*, int) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_remap_ ## type , src); \
tex_remap_ ## type ##_reader texSrc; \
Filter< tex_remap_ ## type ##_reader > filter_src(texSrc); \
remap<<<grid, block>>>(filter_src, mapx, mapy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(uchar4)
END_OPENCV_DEVICE_NAMESPACE
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(schar)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(char4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(short2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(short4)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int2)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(int4)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float)
//OPENCV_GPU_IMPLEMENT_REMAP_TEX(float2)
OPENCV_GPU_IMPLEMENT_REMAP_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_REMAP_TEX
template <template <typename> class Filter, template <typename> class B, typename T> struct RemapDispatcher
{
static void call(const DevMem2D_<T>& src, const DevMem2Df& mapx, const DevMem2Df& mapy, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc)
{
if (stream == 0)
RemapDispatcherNonStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, cc);
else
RemapDispatcherStream<Filter, B, T>::call(src, mapx, mapy, dst, borderValue, stream, cc);
}
};
template <typename T> void remap_gpu(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation,
int borderMode, const float* borderValue, cudaStream_t stream, int cc)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2D_<T>& dst,
const float* borderValue, cudaStream_t stream, int cc);
static const caller_t callers[3][5] =
{
{
RemapDispatcher<PointFilter, BrdReflect101, T>::call,
RemapDispatcher<PointFilter, BrdReplicate, T>::call,
RemapDispatcher<PointFilter, BrdConstant, T>::call,
RemapDispatcher<PointFilter, BrdReflect, T>::call,
RemapDispatcher<PointFilter, BrdWrap, T>::call
},
{
RemapDispatcher<LinearFilter, BrdReflect101, T>::call,
RemapDispatcher<LinearFilter, BrdReplicate, T>::call,
RemapDispatcher<LinearFilter, BrdConstant, T>::call,
RemapDispatcher<LinearFilter, BrdReflect, T>::call,
RemapDispatcher<LinearFilter, BrdWrap, T>::call
},
{
RemapDispatcher<CubicFilter, BrdReflect101, T>::call,
RemapDispatcher<CubicFilter, BrdReplicate, T>::call,
RemapDispatcher<CubicFilter, BrdConstant, T>::call,
RemapDispatcher<CubicFilter, BrdReflect, T>::call,
RemapDispatcher<CubicFilter, BrdWrap, T>::call
}
};
callers[interpolation][borderMode](static_cast< DevMem2D_<T> >(src), xmap, ymap, static_cast< DevMem2D_<T> >(dst), borderValue, stream, cc);
}
template void remap_gpu<uchar >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<uchar2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<uchar4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<schar>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<char4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<ushort2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<ushort4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<short2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<short4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<int4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float >(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
//template void remap_gpu<float2>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float3>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
template void remap_gpu<float4>(const DevMem2Db& src, const DevMem2Df& xmap, const DevMem2Df& ymap, const DevMem2Db& dst, int interpolation, int borderMode, const float* borderValue, cudaStream_t stream, int cc);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

362
modules/gpu/src/cuda/resize.cu
View File

@@ -47,219 +47,217 @@
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/filters.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace imgproc {
template <typename Ptr2D, typename T> __global__ void resize(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
namespace cv { namespace gpu { namespace device
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
namespace imgproc
{
template <typename Ptr2D, typename T> __global__ void resize(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <typename Ptr2D, typename T> __global__ void resizeNN(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
dst.ptr(y)[x] = saturate_cast<T>(src(ycoo, xcoo));
}
}
template <typename Ptr2D, typename T> __global__ void resizeNN(const Ptr2D src, float fx, float fy, DevMem2D_<T> dst)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
if (x < dst.cols && y < dst.rows)
{
const float xcoo = x / fx;
const float ycoo = y / fy;
dst.ptr(y)[x] = src(__float2int_rd(ycoo), __float2int_rd(xcoo));
}
}
dst.ptr(y)[x] = src(__float2int_rd(ycoo), __float2int_rd(xcoo));
}
}
template <template <typename> class Filter, typename T> struct ResizeDispatcherStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
template <template <typename> class Filter, typename T> struct ResizeDispatcherStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
resize<<<grid, block, 0, stream>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <typename T> struct ResizeDispatcherStream<PointFilter, T>
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
resize<<<grid, block, 0, stream>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <typename T> struct ResizeDispatcherStream<PointFilter, T>
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
resizeNN<<<grid, block, 0, stream>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
resizeNN<<<grid, block, 0, stream>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
}
};
template <template <typename> class Filter, typename T> struct ResizeDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
template <template <typename> class Filter, typename T> struct ResizeDispatcherNonStream
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filter_src(brdSrc);
resize<<<grid, block>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
resize<<<grid, block>>>(filter_src, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T> struct ResizeDispatcherNonStream<PointFilter, T>
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
cudaSafeCall( cudaDeviceSynchronize() );
}
};
template <typename T> struct ResizeDispatcherNonStream<PointFilter, T>
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst)
{
dim3 block(32, 8);
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
BrdReplicate<T> brd(src.rows, src.cols);
BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
resizeNN<<<grid, block>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
resizeNN<<<grid, block>>>(brdSrc, fx, fy, dst);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
}
};
cudaSafeCall( cudaDeviceSynchronize() );
}
};
#define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
texture< type , cudaTextureType2D> tex_resize_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_resize_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_resize_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter> struct ResizeDispatcherNonStream<Filter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
Filter< tex_resize_ ## type ##_reader > filter_src(texSrc); \
resize<<<grid, block>>>(filter_src, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <> struct ResizeDispatcherNonStream<PointFilter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
resizeNN<<<grid, block>>>(texSrc, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
#define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
texture< type , cudaTextureType2D> tex_resize_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
struct tex_resize_ ## type ## _reader \
{ \
typedef type elem_type; \
typedef int index_type; \
__device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
{ \
return tex2D(tex_resize_ ## type , x, y); \
} \
}; \
template <template <typename> class Filter> struct ResizeDispatcherNonStream<Filter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
Filter< tex_resize_ ## type ##_reader > filter_src(texSrc); \
resize<<<grid, block>>>(filter_src, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
}; \
template <> struct ResizeDispatcherNonStream<PointFilter, type> \
{ \
static void call(const DevMem2D_< type >& src, float fx, float fy, const DevMem2D_< type >& dst) \
{ \
dim3 block(32, 8); \
dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
bindTexture(&tex_resize_ ## type , src); \
tex_resize_ ## type ##_reader texSrc; \
resizeNN<<<grid, block>>>(texSrc, fx, fy, dst); \
cudaSafeCall( cudaGetLastError() ); \
cudaSafeCall( cudaDeviceSynchronize() ); \
} \
};
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(schar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(schar)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int4)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int2)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
//OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float2)
OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
#undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
#undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
template <template <typename> class Filter, typename T> struct ResizeDispatcher
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
if (stream == 0)
ResizeDispatcherNonStream<Filter, T>::call(src, fx, fy, dst);
else
ResizeDispatcherStream<Filter, T>::call(src, fx, fy, dst, stream);
}
};
template <template <typename> class Filter, typename T> struct ResizeDispatcher
{
static void call(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream)
{
if (stream == 0)
ResizeDispatcherNonStream<Filter, T>::call(src, fx, fy, dst);
else
ResizeDispatcherStream<Filter, T>::call(src, fx, fy, dst, stream);
}
};
template <typename T> void resize_gpu(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream);
template <typename T> void resize_gpu(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, float fx, float fy, const DevMem2D_<T>& dst, cudaStream_t stream);
static const caller_t callers[3] =
{
ResizeDispatcher<PointFilter, T>::call, ResizeDispatcher<LinearFilter, T>::call, ResizeDispatcher<CubicFilter, T>::call
};
static const caller_t callers[3] =
{
ResizeDispatcher<PointFilter, T>::call, ResizeDispatcher<LinearFilter, T>::call, ResizeDispatcher<CubicFilter, T>::call
};
callers[interpolation](static_cast< DevMem2D_<T> >(src), fx, fy, static_cast< DevMem2D_<T> >(dst), stream);
}
callers[interpolation](static_cast< DevMem2D_<T> >(src), fx, fy, static_cast< DevMem2D_<T> >(dst), stream);
}
template void resize_gpu<uchar >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<uchar2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<uchar2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<uchar4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<schar>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<schar>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<char4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort >(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<ushort2>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort3>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort4>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort >(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<ushort2>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort3>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<ushort4>(const DevMem2Db& src,float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<short2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<short2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<short4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<int4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<float2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
} // namespace imgproc
END_OPENCV_DEVICE_NAMESPACE
template void resize_gpu<float >(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
//template void resize_gpu<float2>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float3>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
template void resize_gpu<float4>(const DevMem2Db& src, float fx, float fy, const DevMem2Db& dst, int interpolation, cudaStream_t stream);
} // namespace imgproc
}}} // namespace cv { namespace gpu { namespace device

418
modules/gpu/src/cuda/row_filter.cu
View File

@@ -47,226 +47,224 @@
#include "opencv2/gpu/device/limits.hpp"
#include "opencv2/gpu/device/border_interpolate.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
namespace row_filter {
__constant__ float c_kernel[MAX_KERNEL_SIZE];
void loadKernel(const float kernel[], int ksize)
namespace cv { namespace gpu { namespace device
{
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
#define MAX_KERNEL_SIZE 16
#define BLOCK_DIM_X 16
#define BLOCK_DIM_Y 4
#define RESULT_STEPS 8
#define HALO_STEPS 1
namespace detail
{
template <typename T, size_t size> struct SmemType
namespace row_filter
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type smem_t;
};
__constant__ float c_kernel[MAX_KERNEL_SIZE];
template <typename T> struct SmemType<T, 4>
{
typedef T smem_t;
};
}
template <typename T> struct SmemType
{
typedef typename detail::SmemType<T, sizeof(T)>::smem_t smem_t;
};
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearRowFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename SmemType<T>::smem_t smem_t;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
__shared__ smem_t smem[BLOCK_DIM_Y][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_X];
//Offset to the left halo edge
const int x = (blockIdx.x * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_X + threadIdx.x;
const int y = blockIdx.y * BLOCK_DIM_Y + threadIdx.y;
if (y < src.rows)
{
const T* src_row = src.ptr(y);
//Load main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
//Load left halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_low(i * BLOCK_DIM_X + x, src_row);
//Load right halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
__syncthreads();
D* dst_row = dst.ptr(y);
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
void loadKernel(const float kernel[], int ksize)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for (int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X + j - anchor] * c_kernel[j];
int dstX = x + i * BLOCK_DIM_X;
if (dstX < src.cols)
dst_row[dstX] = saturate_cast<D>(sum);
cudaSafeCall( cudaMemcpyToSymbol(c_kernel, kernel, ksize * sizeof(float)) );
}
}
}
template <int ksize, typename T, typename D, template<typename> class B>
void linearRowFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
typedef typename SmemType<T>::smem_t smem_t;
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, RESULT_STEPS * BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y));
B<smem_t> b(src.cols);
linearRowFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
namespace detail
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect101>,
linearRowFilter_caller<2 , T, D, BrdRowReflect101>,
linearRowFilter_caller<3 , T, D, BrdRowReflect101>,
linearRowFilter_caller<4 , T, D, BrdRowReflect101>,
linearRowFilter_caller<5 , T, D, BrdRowReflect101>,
linearRowFilter_caller<6 , T, D, BrdRowReflect101>,
linearRowFilter_caller<7 , T, D, BrdRowReflect101>,
linearRowFilter_caller<8 , T, D, BrdRowReflect101>,
linearRowFilter_caller<9 , T, D, BrdRowReflect101>,
linearRowFilter_caller<10, T, D, BrdRowReflect101>,
linearRowFilter_caller<11, T, D, BrdRowReflect101>,
linearRowFilter_caller<12, T, D, BrdRowReflect101>,
linearRowFilter_caller<13, T, D, BrdRowReflect101>,
linearRowFilter_caller<14, T, D, BrdRowReflect101>,
linearRowFilter_caller<15, T, D, BrdRowReflect101>,
linearRowFilter_caller<16, T, D, BrdRowReflect101>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReplicate>,
linearRowFilter_caller<2 , T, D, BrdRowReplicate>,
linearRowFilter_caller<3 , T, D, BrdRowReplicate>,
linearRowFilter_caller<4 , T, D, BrdRowReplicate>,
linearRowFilter_caller<5 , T, D, BrdRowReplicate>,
linearRowFilter_caller<6 , T, D, BrdRowReplicate>,
linearRowFilter_caller<7 , T, D, BrdRowReplicate>,
linearRowFilter_caller<8 , T, D, BrdRowReplicate>,
linearRowFilter_caller<9 , T, D, BrdRowReplicate>,
linearRowFilter_caller<10, T, D, BrdRowReplicate>,
linearRowFilter_caller<11, T, D, BrdRowReplicate>,
linearRowFilter_caller<12, T, D, BrdRowReplicate>,
linearRowFilter_caller<13, T, D, BrdRowReplicate>,
linearRowFilter_caller<14, T, D, BrdRowReplicate>,
linearRowFilter_caller<15, T, D, BrdRowReplicate>,
linearRowFilter_caller<16, T, D, BrdRowReplicate>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowConstant>,
linearRowFilter_caller<2 , T, D, BrdRowConstant>,
linearRowFilter_caller<3 , T, D, BrdRowConstant>,
linearRowFilter_caller<4 , T, D, BrdRowConstant>,
linearRowFilter_caller<5 , T, D, BrdRowConstant>,
linearRowFilter_caller<6 , T, D, BrdRowConstant>,
linearRowFilter_caller<7 , T, D, BrdRowConstant>,
linearRowFilter_caller<8 , T, D, BrdRowConstant>,
linearRowFilter_caller<9 , T, D, BrdRowConstant>,
linearRowFilter_caller<10, T, D, BrdRowConstant>,
linearRowFilter_caller<11, T, D, BrdRowConstant>,
linearRowFilter_caller<12, T, D, BrdRowConstant>,
linearRowFilter_caller<13, T, D, BrdRowConstant>,
linearRowFilter_caller<14, T, D, BrdRowConstant>,
linearRowFilter_caller<15, T, D, BrdRowConstant>,
linearRowFilter_caller<16, T, D, BrdRowConstant>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect>,
linearRowFilter_caller<2 , T, D, BrdRowReflect>,
linearRowFilter_caller<3 , T, D, BrdRowReflect>,
linearRowFilter_caller<4 , T, D, BrdRowReflect>,
linearRowFilter_caller<5 , T, D, BrdRowReflect>,
linearRowFilter_caller<6 , T, D, BrdRowReflect>,
linearRowFilter_caller<7 , T, D, BrdRowReflect>,
linearRowFilter_caller<8 , T, D, BrdRowReflect>,
linearRowFilter_caller<9 , T, D, BrdRowReflect>,
linearRowFilter_caller<10, T, D, BrdRowReflect>,
linearRowFilter_caller<11, T, D, BrdRowReflect>,
linearRowFilter_caller<12, T, D, BrdRowReflect>,
linearRowFilter_caller<13, T, D, BrdRowReflect>,
linearRowFilter_caller<14, T, D, BrdRowReflect>,
linearRowFilter_caller<15, T, D, BrdRowReflect>,
linearRowFilter_caller<16, T, D, BrdRowReflect>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowWrap>,
linearRowFilter_caller<2 , T, D, BrdRowWrap>,
linearRowFilter_caller<3 , T, D, BrdRowWrap>,
linearRowFilter_caller<4 , T, D, BrdRowWrap>,
linearRowFilter_caller<5 , T, D, BrdRowWrap>,
linearRowFilter_caller<6 , T, D, BrdRowWrap>,
linearRowFilter_caller<7 , T, D, BrdRowWrap>,
linearRowFilter_caller<8 , T, D, BrdRowWrap>,
linearRowFilter_caller<9 , T, D, BrdRowWrap>,
linearRowFilter_caller<10, T, D, BrdRowWrap>,
linearRowFilter_caller<11, T, D, BrdRowWrap>,
linearRowFilter_caller<12, T, D, BrdRowWrap>,
linearRowFilter_caller<13, T, D, BrdRowWrap>,
linearRowFilter_caller<14, T, D, BrdRowWrap>,
linearRowFilter_caller<15, T, D, BrdRowWrap>,
linearRowFilter_caller<16, T, D, BrdRowWrap>
template <typename T, size_t size> struct SmemType
{
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type smem_t;
};
template <typename T> struct SmemType<T, 4>
{
typedef T smem_t;
};
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
template <typename T> struct SmemType
{
typedef typename detail::SmemType<T, sizeof(T)>::smem_t smem_t;
};
template void linearRowFilter_gpu<uchar , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<uchar4, float4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short2, float2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<short3, float3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<int , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template <int KERNEL_SIZE, typename T, typename D, typename B>
__global__ void linearRowFilter(const DevMem2D_<T> src, PtrStep<D> dst, int anchor, const B b)
{
typedef typename SmemType<T>::smem_t smem_t;
typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type sum_t;
} // namespace row_filter
__shared__ smem_t smem[BLOCK_DIM_Y][(RESULT_STEPS + 2 * HALO_STEPS) * BLOCK_DIM_X];
END_OPENCV_DEVICE_NAMESPACE
//Offset to the left halo edge
const int x = (blockIdx.x * RESULT_STEPS - HALO_STEPS) * BLOCK_DIM_X + threadIdx.x;
const int y = blockIdx.y * BLOCK_DIM_Y + threadIdx.y;
if (y < src.rows)
{
const T* src_row = src.ptr(y);
//Load main data
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
//Load left halo
#pragma unroll
for(int i = 0; i < HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_low(i * BLOCK_DIM_X + x, src_row);
//Load right halo
#pragma unroll
for(int i = HALO_STEPS + RESULT_STEPS; i < HALO_STEPS + RESULT_STEPS + HALO_STEPS; ++i)
smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X] = b.at_high(i * BLOCK_DIM_X + x, src_row);
__syncthreads();
D* dst_row = dst.ptr(y);
#pragma unroll
for(int i = HALO_STEPS; i < HALO_STEPS + RESULT_STEPS; ++i)
{
sum_t sum = VecTraits<sum_t>::all(0);
#pragma unroll
for (int j = 0; j < KERNEL_SIZE; ++j)
sum = sum + smem[threadIdx.y][threadIdx.x + i * BLOCK_DIM_X + j - anchor] * c_kernel[j];
int dstX = x + i * BLOCK_DIM_X;
if (dstX < src.cols)
dst_row[dstX] = saturate_cast<D>(sum);
}
}
}
template <int ksize, typename T, typename D, template<typename> class B>
void linearRowFilter_caller(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream)
{
typedef typename SmemType<T>::smem_t smem_t;
const dim3 block(BLOCK_DIM_X, BLOCK_DIM_Y);
const dim3 grid(divUp(src.cols, RESULT_STEPS * BLOCK_DIM_X), divUp(src.rows, BLOCK_DIM_Y));
B<smem_t> b(src.cols);
linearRowFilter<ksize, T, D><<<grid, block, 0, stream>>>(src, dst, anchor, b);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <typename T, typename D>
void linearRowFilter_gpu(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2D_<T>& src, const DevMem2D_<D>& dst, int anchor, cudaStream_t stream);
static const caller_t callers[5][17] =
{
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect101>,
linearRowFilter_caller<2 , T, D, BrdRowReflect101>,
linearRowFilter_caller<3 , T, D, BrdRowReflect101>,
linearRowFilter_caller<4 , T, D, BrdRowReflect101>,
linearRowFilter_caller<5 , T, D, BrdRowReflect101>,
linearRowFilter_caller<6 , T, D, BrdRowReflect101>,
linearRowFilter_caller<7 , T, D, BrdRowReflect101>,
linearRowFilter_caller<8 , T, D, BrdRowReflect101>,
linearRowFilter_caller<9 , T, D, BrdRowReflect101>,
linearRowFilter_caller<10, T, D, BrdRowReflect101>,
linearRowFilter_caller<11, T, D, BrdRowReflect101>,
linearRowFilter_caller<12, T, D, BrdRowReflect101>,
linearRowFilter_caller<13, T, D, BrdRowReflect101>,
linearRowFilter_caller<14, T, D, BrdRowReflect101>,
linearRowFilter_caller<15, T, D, BrdRowReflect101>,
linearRowFilter_caller<16, T, D, BrdRowReflect101>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReplicate>,
linearRowFilter_caller<2 , T, D, BrdRowReplicate>,
linearRowFilter_caller<3 , T, D, BrdRowReplicate>,
linearRowFilter_caller<4 , T, D, BrdRowReplicate>,
linearRowFilter_caller<5 , T, D, BrdRowReplicate>,
linearRowFilter_caller<6 , T, D, BrdRowReplicate>,
linearRowFilter_caller<7 , T, D, BrdRowReplicate>,
linearRowFilter_caller<8 , T, D, BrdRowReplicate>,
linearRowFilter_caller<9 , T, D, BrdRowReplicate>,
linearRowFilter_caller<10, T, D, BrdRowReplicate>,
linearRowFilter_caller<11, T, D, BrdRowReplicate>,
linearRowFilter_caller<12, T, D, BrdRowReplicate>,
linearRowFilter_caller<13, T, D, BrdRowReplicate>,
linearRowFilter_caller<14, T, D, BrdRowReplicate>,
linearRowFilter_caller<15, T, D, BrdRowReplicate>,
linearRowFilter_caller<16, T, D, BrdRowReplicate>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowConstant>,
linearRowFilter_caller<2 , T, D, BrdRowConstant>,
linearRowFilter_caller<3 , T, D, BrdRowConstant>,
linearRowFilter_caller<4 , T, D, BrdRowConstant>,
linearRowFilter_caller<5 , T, D, BrdRowConstant>,
linearRowFilter_caller<6 , T, D, BrdRowConstant>,
linearRowFilter_caller<7 , T, D, BrdRowConstant>,
linearRowFilter_caller<8 , T, D, BrdRowConstant>,
linearRowFilter_caller<9 , T, D, BrdRowConstant>,
linearRowFilter_caller<10, T, D, BrdRowConstant>,
linearRowFilter_caller<11, T, D, BrdRowConstant>,
linearRowFilter_caller<12, T, D, BrdRowConstant>,
linearRowFilter_caller<13, T, D, BrdRowConstant>,
linearRowFilter_caller<14, T, D, BrdRowConstant>,
linearRowFilter_caller<15, T, D, BrdRowConstant>,
linearRowFilter_caller<16, T, D, BrdRowConstant>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowReflect>,
linearRowFilter_caller<2 , T, D, BrdRowReflect>,
linearRowFilter_caller<3 , T, D, BrdRowReflect>,
linearRowFilter_caller<4 , T, D, BrdRowReflect>,
linearRowFilter_caller<5 , T, D, BrdRowReflect>,
linearRowFilter_caller<6 , T, D, BrdRowReflect>,
linearRowFilter_caller<7 , T, D, BrdRowReflect>,
linearRowFilter_caller<8 , T, D, BrdRowReflect>,
linearRowFilter_caller<9 , T, D, BrdRowReflect>,
linearRowFilter_caller<10, T, D, BrdRowReflect>,
linearRowFilter_caller<11, T, D, BrdRowReflect>,
linearRowFilter_caller<12, T, D, BrdRowReflect>,
linearRowFilter_caller<13, T, D, BrdRowReflect>,
linearRowFilter_caller<14, T, D, BrdRowReflect>,
linearRowFilter_caller<15, T, D, BrdRowReflect>,
linearRowFilter_caller<16, T, D, BrdRowReflect>
},
{
0,
linearRowFilter_caller<1 , T, D, BrdRowWrap>,
linearRowFilter_caller<2 , T, D, BrdRowWrap>,
linearRowFilter_caller<3 , T, D, BrdRowWrap>,
linearRowFilter_caller<4 , T, D, BrdRowWrap>,
linearRowFilter_caller<5 , T, D, BrdRowWrap>,
linearRowFilter_caller<6 , T, D, BrdRowWrap>,
linearRowFilter_caller<7 , T, D, BrdRowWrap>,
linearRowFilter_caller<8 , T, D, BrdRowWrap>,
linearRowFilter_caller<9 , T, D, BrdRowWrap>,
linearRowFilter_caller<10, T, D, BrdRowWrap>,
linearRowFilter_caller<11, T, D, BrdRowWrap>,
linearRowFilter_caller<12, T, D, BrdRowWrap>,
linearRowFilter_caller<13, T, D, BrdRowWrap>,
linearRowFilter_caller<14, T, D, BrdRowWrap>,
linearRowFilter_caller<15, T, D, BrdRowWrap>,
linearRowFilter_caller<16, T, D, BrdRowWrap>
}
};
loadKernel(kernel, ksize);
callers[brd_type][ksize]((DevMem2D_<T>)src, (DevMem2D_<D>)dst, anchor, stream);
}
template void linearRowFilter_gpu<uchar , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<uchar4, float4>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
//template void linearRowFilter_gpu<short2, float2>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<short3, float3>(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<int , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
template void linearRowFilter_gpu<float , float >(const DevMem2Db& src, const DevMem2Db& dst, const float kernel[], int ksize, int anchor, int brd_type, cudaStream_t stream);
} // namespace row_filter
}}} // namespace cv { namespace gpu { namespace device

63
modules/gpu/src/cuda/safe_call.hpp
View File

@@ -62,44 +62,43 @@
#define cublasSafeCall(expr) ___cublasSafeCall(expr, __FILE__, __LINE__)
#endif
namespace cv { namespace gpu {
void error(const char *error_string, const char *file, const int line, const char *func = "");
void nppError(int err, const char *file, const int line, const char *func = "");
void ncvError(int err, const char *file, const int line, const char *func = "");
void cufftError(int err, const char *file, const int line, const char *func = "");
void cublasError(int err, const char *file, const int line, const char *func = "");
static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int line, const char *func = "")
namespace cv { namespace gpu
{
if (cudaSuccess != err)
cv::gpu::error(cudaGetErrorString(err), file, line, func);
}
void error(const char *error_string, const char *file, const int line, const char *func = "");
void nppError(int err, const char *file, const int line, const char *func = "");
void ncvError(int err, const char *file, const int line, const char *func = "");
void cufftError(int err, const char *file, const int line, const char *func = "");
void cublasError(int err, const char *file, const int line, const char *func = "");
static inline void ___nppSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (err < 0)
cv::gpu::nppError(err, file, line, func);
}
static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int line, const char *func = "")
{
if (cudaSuccess != err)
cv::gpu::error(cudaGetErrorString(err), file, line, func);
}
static inline void ___ncvSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (NCV_SUCCESS != err)
cv::gpu::ncvError(err, file, line, func);
}
static inline void ___nppSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (err < 0)
cv::gpu::nppError(err, file, line, func);
}
static inline void ___cufftSafeCall(cufftResult_t err, const char *file, const int line, const char *func = "")
{
if (CUFFT_SUCCESS != err)
cv::gpu::cufftError(err, file, line, func);
}
static inline void ___ncvSafeCall(int err, const char *file, const int line, const char *func = "")
{
if (NCV_SUCCESS != err)
cv::gpu::ncvError(err, file, line, func);
}
static inline void ___cublasSafeCall(cublasStatus_t err, const char *file, const int line, const char *func = "")
{
if (CUBLAS_STATUS_SUCCESS != err)
cv::gpu::cublasError(err, file, line, func);
}
static inline void ___cufftSafeCall(cufftResult_t err, const char *file, const int line, const char *func = "")
{
if (CUFFT_SUCCESS != err)
cv::gpu::cufftError(err, file, line, func);
}
static inline void ___cublasSafeCall(cublasStatus_t err, const char *file, const int line, const char *func = "")
{
if (CUBLAS_STATUS_SUCCESS != err)
cv::gpu::cublasError(err, file, line, func);
}
}}
#endif /* __OPENCV_CUDA_SAFE_CALL_HPP__ */

838
modules/gpu/src/cuda/split_merge.cu
View File

@@ -42,467 +42,465 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace split_merge {
template <typename T, size_t elem_size = sizeof(T)>
struct TypeTraits
namespace cv { namespace gpu { namespace device
{
typedef T type;
typedef T type2;
typedef T type3;
typedef T type4;
};
template <typename T>
struct TypeTraits<T, 1>
{
typedef char type;
typedef char2 type2;
typedef char3 type3;
typedef char4 type4;
};
template <typename T>
struct TypeTraits<T, 2>
{
typedef short type;
typedef short2 type2;
typedef short3 type3;
typedef short4 type4;
};
template <typename T>
struct TypeTraits<T, 4>
{
typedef int type;
typedef int2 type2;
typedef int3 type3;
typedef int4 type4;
};
template <typename T>
struct TypeTraits<T, 8>
{
typedef double type;
typedef double2 type2;
//typedef double3 type3;
//typedef double4 type3;
};
typedef void (*MergeFunction)(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream);
typedef void (*SplitFunction)(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream);
//------------------------------------------------------------
// Merge
template <typename T>
__global__ void mergeC2_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type2 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_y[x] = dst_elem;
}
}
template <typename T>
__global__ void mergeC3_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type3 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC3_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
double* dst_y = (double*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[3 * x] = src0_y[x];
dst_y[3 * x + 1] = src1_y[x];
dst_y[3 * x + 2] = src2_y[x];
}
}
template <typename T>
__global__ void mergeC4_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type4 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
const T* src3_y = (const T*)(src3 + y * src3_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_elem.w = src3_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC4_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
const double* src3_y = (const double*)(src3 + y * src3_step);
double2* dst_y = (double2*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[2 * x] = make_double2(src0_y[x], src1_y[x]);
dst_y[2 * x + 1] = make_double2(src2_y[x], src3_y[x]);
}
}
template <typename T>
static void mergeC2_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC2_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC3_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC3_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC4_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC4_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
src[3].data, src[3].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
void merge_caller(const DevMem2Db* src, DevMem2Db& dst,
int total_channels, size_t elem_size,
const cudaStream_t& stream)
{
static MergeFunction merge_func_tbl[] =
namespace split_merge
{
mergeC2_<char>, mergeC2_<short>, mergeC2_<int>, 0, mergeC2_<double>,
mergeC3_<char>, mergeC3_<short>, mergeC3_<int>, 0, mergeC3_<double>,
mergeC4_<char>, mergeC4_<short>, mergeC4_<int>, 0, mergeC4_<double>,
};
template <typename T, size_t elem_size = sizeof(T)>
struct TypeTraits
{
typedef T type;
typedef T type2;
typedef T type3;
typedef T type4;
};
size_t merge_func_id = (total_channels - 2) * 5 + (elem_size >> 1);
MergeFunction merge_func = merge_func_tbl[merge_func_id];
template <typename T>
struct TypeTraits<T, 1>
{
typedef char type;
typedef char2 type2;
typedef char3 type3;
typedef char4 type4;
};
if (merge_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
template <typename T>
struct TypeTraits<T, 2>
{
typedef short type;
typedef short2 type2;
typedef short3 type3;
typedef short4 type4;
};
merge_func(src, dst, stream);
}
template <typename T>
struct TypeTraits<T, 4>
{
typedef int type;
typedef int2 type2;
typedef int3 type3;
typedef int4 type4;
};
template <typename T>
struct TypeTraits<T, 8>
{
typedef double type;
typedef double2 type2;
//typedef double3 type3;
//typedef double4 type3;
};
typedef void (*MergeFunction)(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream);
typedef void (*SplitFunction)(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream);
//------------------------------------------------------------
// Merge
template <typename T>
__global__ void mergeC2_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type2 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_y[x] = dst_elem;
}
}
template <typename T>
__global__ void mergeC3_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type3 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC3_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
double* dst_y = (double*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[3 * x] = src0_y[x];
dst_y[3 * x + 1] = src1_y[x];
dst_y[3 * x + 2] = src2_y[x];
}
}
template <typename T>
__global__ void mergeC4_(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
typedef typename TypeTraits<T>::type4 dst_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const T* src0_y = (const T*)(src0 + y * src0_step);
const T* src1_y = (const T*)(src1 + y * src1_step);
const T* src2_y = (const T*)(src2 + y * src2_step);
const T* src3_y = (const T*)(src3 + y * src3_step);
dst_type* dst_y = (dst_type*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_type dst_elem;
dst_elem.x = src0_y[x];
dst_elem.y = src1_y[x];
dst_elem.z = src2_y[x];
dst_elem.w = src3_y[x];
dst_y[x] = dst_elem;
}
}
template <>
__global__ void mergeC4_<double>(const uchar* src0, size_t src0_step,
const uchar* src1, size_t src1_step,
const uchar* src2, size_t src2_step,
const uchar* src3, size_t src3_step,
int rows, int cols, uchar* dst, size_t dst_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src0_y = (const double*)(src0 + y * src0_step);
const double* src1_y = (const double*)(src1 + y * src1_step);
const double* src2_y = (const double*)(src2 + y * src2_step);
const double* src3_y = (const double*)(src3 + y * src3_step);
double2* dst_y = (double2*)(dst + y * dst_step);
if (x < cols && y < rows)
{
dst_y[2 * x] = make_double2(src0_y[x], src1_y[x]);
dst_y[2 * x + 1] = make_double2(src2_y[x], src3_y[x]);
}
}
template <typename T>
static void mergeC2_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC2_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC3_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC3_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void mergeC4_(const DevMem2Db* src, DevMem2Db& dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(dst.cols, blockDim.x), divUp(dst.rows, blockDim.y));
mergeC4_<T><<<gridDim, blockDim, 0, stream>>>(
src[0].data, src[0].step,
src[1].data, src[1].step,
src[2].data, src[2].step,
src[3].data, src[3].step,
dst.rows, dst.cols, dst.data, dst.step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
void merge_caller(const DevMem2Db* src, DevMem2Db& dst,
int total_channels, size_t elem_size,
const cudaStream_t& stream)
{
static MergeFunction merge_func_tbl[] =
{
mergeC2_<char>, mergeC2_<short>, mergeC2_<int>, 0, mergeC2_<double>,
mergeC3_<char>, mergeC3_<short>, mergeC3_<int>, 0, mergeC3_<double>,
mergeC4_<char>, mergeC4_<short>, mergeC4_<int>, 0, mergeC4_<double>,
};
size_t merge_func_id = (total_channels - 2) * 5 + (elem_size >> 1);
MergeFunction merge_func = merge_func_tbl[merge_func_id];
if (merge_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
merge_func(src, dst, stream);
}
//------------------------------------------------------------
// Split
//------------------------------------------------------------
// Split
template <typename T>
__global__ void splitC2_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step)
{
typedef typename TypeTraits<T>::type2 src_type;
template <typename T>
__global__ void splitC2_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step)
{
typedef typename TypeTraits<T>::type2 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
}
}
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
}
}
template <typename T>
__global__ void splitC3_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
{
typedef typename TypeTraits<T>::type3 src_type;
template <typename T>
__global__ void splitC3_(const uchar* src, size_t src_step,
int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
{
typedef typename TypeTraits<T>::type3 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
}
}
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
}
}
template <>
__global__ void splitC3_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
template <>
__global__ void splitC3_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double* src_y = (const double*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
const double* src_y = (const double*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
if (x < cols && y < rows)
{
dst0_y[x] = src_y[3 * x];
dst1_y[x] = src_y[3 * x + 1];
dst2_y[x] = src_y[3 * x + 2];
}
}
if (x < cols && y < rows)
{
dst0_y[x] = src_y[3 * x];
dst1_y[x] = src_y[3 * x + 1];
dst2_y[x] = src_y[3 * x + 2];
}
}
template <typename T>
__global__ void splitC4_(const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
typedef typename TypeTraits<T>::type4 src_type;
template <typename T>
__global__ void splitC4_(const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
typedef typename TypeTraits<T>::type4 src_type;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
T* dst3_y = (T*)(dst3 + y * dst3_step);
const src_type* src_y = (const src_type*)(src + y * src_step);
T* dst0_y = (T*)(dst0 + y * dst0_step);
T* dst1_y = (T*)(dst1 + y * dst1_step);
T* dst2_y = (T*)(dst2 + y * dst2_step);
T* dst3_y = (T*)(dst3 + y * dst3_step);
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
dst3_y[x] = src_elem.w;
}
}
if (x < cols && y < rows)
{
src_type src_elem = src_y[x];
dst0_y[x] = src_elem.x;
dst1_y[x] = src_elem.y;
dst2_y[x] = src_elem.z;
dst3_y[x] = src_elem.w;
}
}
template <>
__global__ void splitC4_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
template <>
__global__ void splitC4_<double>(
const uchar* src, size_t src_step, int rows, int cols,
uchar* dst0, size_t dst0_step,
uchar* dst1, size_t dst1_step,
uchar* dst2, size_t dst2_step,
uchar* dst3, size_t dst3_step)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const double2* src_y = (const double2*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
double* dst3_y = (double*)(dst3 + y * dst3_step);
const double2* src_y = (const double2*)(src + y * src_step);
double* dst0_y = (double*)(dst0 + y * dst0_step);
double* dst1_y = (double*)(dst1 + y * dst1_step);
double* dst2_y = (double*)(dst2 + y * dst2_step);
double* dst3_y = (double*)(dst3 + y * dst3_step);
if (x < cols && y < rows)
{
double2 src_elem1 = src_y[2 * x];
double2 src_elem2 = src_y[2 * x + 1];
dst0_y[x] = src_elem1.x;
dst1_y[x] = src_elem1.y;
dst2_y[x] = src_elem2.x;
dst3_y[x] = src_elem2.y;
}
}
if (x < cols && y < rows)
{
double2 src_elem1 = src_y[2 * x];
double2 src_elem2 = src_y[2 * x + 1];
dst0_y[x] = src_elem1.x;
dst1_y[x] = src_elem1.y;
dst2_y[x] = src_elem2.x;
dst3_y[x] = src_elem2.y;
}
}
template <typename T>
static void splitC2_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC2_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step);
cudaSafeCall( cudaGetLastError() );
template <typename T>
static void splitC2_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC2_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void splitC3_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC3_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step);
cudaSafeCall( cudaGetLastError() );
template <typename T>
static void splitC3_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC3_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
template <typename T>
static void splitC4_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC4_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step,
dst[3].data, dst[3].step);
cudaSafeCall( cudaGetLastError() );
template <typename T>
static void splitC4_(const DevMem2Db& src, DevMem2Db* dst, const cudaStream_t& stream)
{
dim3 blockDim(32, 8);
dim3 gridDim(divUp(src.cols, blockDim.x), divUp(src.rows, blockDim.y));
splitC4_<T><<<gridDim, blockDim, 0, stream>>>(
src.data, src.step, src.rows, src.cols,
dst[0].data, dst[0].step,
dst[1].data, dst[1].step,
dst[2].data, dst[2].step,
dst[3].data, dst[3].step);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
if (stream == 0)
cudaSafeCall(cudaDeviceSynchronize());
}
void split_caller(const DevMem2Db& src, DevMem2Db* dst, int num_channels, size_t elem_size1, const cudaStream_t& stream)
{
static SplitFunction split_func_tbl[] =
{
splitC2_<char>, splitC2_<short>, splitC2_<int>, 0, splitC2_<double>,
splitC3_<char>, splitC3_<short>, splitC3_<int>, 0, splitC3_<double>,
splitC4_<char>, splitC4_<short>, splitC4_<int>, 0, splitC4_<double>,
};
void split_caller(const DevMem2Db& src, DevMem2Db* dst, int num_channels, size_t elem_size1, const cudaStream_t& stream)
{
static SplitFunction split_func_tbl[] =
{
splitC2_<char>, splitC2_<short>, splitC2_<int>, 0, splitC2_<double>,
splitC3_<char>, splitC3_<short>, splitC3_<int>, 0, splitC3_<double>,
splitC4_<char>, splitC4_<short>, splitC4_<int>, 0, splitC4_<double>,
};
size_t split_func_id = (num_channels - 2) * 5 + (elem_size1 >> 1);
SplitFunction split_func = split_func_tbl[split_func_id];
size_t split_func_id = (num_channels - 2) * 5 + (elem_size1 >> 1);
SplitFunction split_func = split_func_tbl[split_func_id];
if (split_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
if (split_func == 0)
cv::gpu::error("Unsupported channel count or data type", __FILE__, __LINE__);
split_func(src, dst, stream);
}
} // namespace split_merge
END_OPENCV_DEVICE_NAMESPACE
split_func(src, dst, stream);
}
} // namespace split_merge
}}} // namespace cv { namespace gpu { namespace device

928
modules/gpu/src/cuda/stereobm.cu
View File

@@ -42,496 +42,494 @@
#include "internal_shared.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace stereobm {
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Stereo BM ////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
#define ROWSperTHREAD 21 // the number of rows a thread will process
#define BLOCK_W 128 // the thread block width (464)
#define N_DISPARITIES 8
#define STEREO_MIND 0 // The minimum d range to check
#define STEREO_DISP_STEP N_DISPARITIES // the d step, must be <= 1 to avoid aliasing
__constant__ unsigned int* cminSSDImage;
__constant__ size_t cminSSD_step;
__constant__ int cwidth;
__constant__ int cheight;
__device__ __forceinline__ int SQ(int a)
namespace cv { namespace gpu { namespace device
{
return a * a;
}
template<int RADIUS>
__device__ unsigned int CalcSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int cache = 0;
unsigned int cache2 = 0;
for(int i = 1; i <= RADIUS; i++)
cache += col_ssd[i];
col_ssd_cache[0] = cache;
__syncthreads();
if (threadIdx.x < BLOCK_W - RADIUS)
cache2 = col_ssd_cache[RADIUS];
else
for(int i = RADIUS + 1; i < (2 * RADIUS + 1); i++)
cache2 += col_ssd[i];
return col_ssd[0] + cache + cache2;
}
template<int RADIUS>
__device__ uint2 MinSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int ssd[N_DISPARITIES];
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
ssd[0] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 0 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[1] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 1 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[2] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 2 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[3] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 3 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[4] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 4 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[5] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 5 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[6] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 6 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[7] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 7 * (BLOCK_W + 2 * RADIUS));
int mssd = ::min(::min(::min(ssd[0], ssd[1]), ::min(ssd[4], ssd[5])), ::min(::min(ssd[2], ssd[3]), ::min(ssd[6], ssd[7])));
int bestIdx = 0;
for (int i = 0; i < N_DISPARITIES; i++)
namespace stereobm
{
if (mssd == ssd[i])
bestIdx = i;
}
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Stereo BM ////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
return make_uint2(mssd, bestIdx);
}
#define ROWSperTHREAD 21 // the number of rows a thread will process
template<int RADIUS>
__device__ void StepDown(int idx1, int idx2, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
unsigned char leftPixel1;
unsigned char leftPixel2;
unsigned char rightPixel1[8];
unsigned char rightPixel2[8];
unsigned int diff1, diff2;
#define BLOCK_W 128 // the thread block width (464)
#define N_DISPARITIES 8
leftPixel1 = imageL[idx1];
leftPixel2 = imageL[idx2];
#define STEREO_MIND 0 // The minimum d range to check
#define STEREO_DISP_STEP N_DISPARITIES // the d step, must be <= 1 to avoid aliasing
idx1 = idx1 - d;
idx2 = idx2 - d;
__constant__ unsigned int* cminSSDImage;
__constant__ size_t cminSSD_step;
__constant__ int cwidth;
__constant__ int cheight;
rightPixel1[7] = imageR[idx1 - 7];
rightPixel1[0] = imageR[idx1 - 0];
rightPixel1[1] = imageR[idx1 - 1];
rightPixel1[2] = imageR[idx1 - 2];
rightPixel1[3] = imageR[idx1 - 3];
rightPixel1[4] = imageR[idx1 - 4];
rightPixel1[5] = imageR[idx1 - 5];
rightPixel1[6] = imageR[idx1 - 6];
rightPixel2[7] = imageR[idx2 - 7];
rightPixel2[0] = imageR[idx2 - 0];
rightPixel2[1] = imageR[idx2 - 1];
rightPixel2[2] = imageR[idx2 - 2];
rightPixel2[3] = imageR[idx2 - 3];
rightPixel2[4] = imageR[idx2 - 4];
rightPixel2[5] = imageR[idx2 - 5];
rightPixel2[6] = imageR[idx2 - 6];
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
diff1 = leftPixel1 - rightPixel1[0];
diff2 = leftPixel2 - rightPixel2[0];
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[1];
diff2 = leftPixel2 - rightPixel2[1];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[2];
diff2 = leftPixel2 - rightPixel2[2];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[3];
diff2 = leftPixel2 - rightPixel2[3];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[4];
diff2 = leftPixel2 - rightPixel2[4];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[5];
diff2 = leftPixel2 - rightPixel2[5];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[6];
diff2 = leftPixel2 - rightPixel2[6];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[7];
diff2 = leftPixel2 - rightPixel2[7];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
}
template<int RADIUS>
__device__ void InitColSSD(int x_tex, int y_tex, int im_pitch, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
unsigned char leftPixel1;
int idx;
unsigned int diffa[] = {0, 0, 0, 0, 0, 0, 0, 0};
for(int i = 0; i < (2 * RADIUS + 1); i++)
{
idx = y_tex * im_pitch + x_tex;
leftPixel1 = imageL[idx];
idx = idx - d;
diffa[0] += SQ(leftPixel1 - imageR[idx - 0]);
diffa[1] += SQ(leftPixel1 - imageR[idx - 1]);
diffa[2] += SQ(leftPixel1 - imageR[idx - 2]);
diffa[3] += SQ(leftPixel1 - imageR[idx - 3]);
diffa[4] += SQ(leftPixel1 - imageR[idx - 4]);
diffa[5] += SQ(leftPixel1 - imageR[idx - 5]);
diffa[6] += SQ(leftPixel1 - imageR[idx - 6]);
diffa[7] += SQ(leftPixel1 - imageR[idx - 7]);
y_tex += 1;
}
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] = diffa[0];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] = diffa[1];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] = diffa[2];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] = diffa[3];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] = diffa[4];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] = diffa[5];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] = diffa[6];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] = diffa[7];
}
template<int RADIUS>
__global__ void stereoKernel(unsigned char *left, unsigned char *right, size_t img_step, PtrStepb disp, int maxdisp)
{
extern __shared__ unsigned int col_ssd_cache[];
volatile unsigned int *col_ssd = col_ssd_cache + BLOCK_W + threadIdx.x;
volatile unsigned int *col_ssd_extra = threadIdx.x < (2 * RADIUS) ? col_ssd + BLOCK_W : 0; //#define N_DIRTY_PIXELS (2 * RADIUS)
//#define X (blockIdx.x * BLOCK_W + threadIdx.x + STEREO_MAXD)
int X = (blockIdx.x * BLOCK_W + threadIdx.x + maxdisp + RADIUS);
//#define Y (__mul24(blockIdx.y, ROWSperTHREAD) + RADIUS)
#define Y (blockIdx.y * ROWSperTHREAD + RADIUS)
//int Y = blockIdx.y * ROWSperTHREAD + RADIUS;
unsigned int* minSSDImage = cminSSDImage + X + Y * cminSSD_step;
unsigned char* disparImage = disp.data + X + Y * disp.step;
/* if (X < cwidth)
{
unsigned int *minSSDImage_end = minSSDImage + min(ROWSperTHREAD, cheight - Y) * minssd_step;
for(uint *ptr = minSSDImage; ptr != minSSDImage_end; ptr += minssd_step )
*ptr = 0xFFFFFFFF;
}*/
int end_row = ::min(ROWSperTHREAD, cheight - Y - RADIUS);
int y_tex;
int x_tex = X - RADIUS;
if (x_tex >= cwidth)
return;
for(int d = STEREO_MIND; d < maxdisp; d += STEREO_DISP_STEP)
{
y_tex = Y - RADIUS;
InitColSSD<RADIUS>(x_tex, y_tex, img_step, left, right, d, col_ssd);
if (col_ssd_extra > 0)
if (x_tex + BLOCK_W < cwidth)
InitColSSD<RADIUS>(x_tex + BLOCK_W, y_tex, img_step, left, right, d, col_ssd_extra);
__syncthreads(); //before MinSSD function
if (X < cwidth - RADIUS && Y < cheight - RADIUS)
__device__ __forceinline__ int SQ(int a)
{
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[0])
{
disparImage[0] = (unsigned char)(d + minSSD.y);
minSSDImage[0] = minSSD.x;
}
return a * a;
}
for(int row = 1; row < end_row; row++)
{
int idx1 = y_tex * img_step + x_tex;
int idx2 = (y_tex + (2 * RADIUS + 1)) * img_step + x_tex;
template<int RADIUS>
__device__ unsigned int CalcSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int cache = 0;
unsigned int cache2 = 0;
for(int i = 1; i <= RADIUS; i++)
cache += col_ssd[i];
col_ssd_cache[0] = cache;
__syncthreads();
StepDown<RADIUS>(idx1, idx2, left, right, d, col_ssd);
if (threadIdx.x < BLOCK_W - RADIUS)
cache2 = col_ssd_cache[RADIUS];
else
for(int i = RADIUS + 1; i < (2 * RADIUS + 1); i++)
cache2 += col_ssd[i];
if (col_ssd_extra)
if (x_tex + BLOCK_W < cwidth)
StepDown<RADIUS>(idx1, idx2, left + BLOCK_W, right + BLOCK_W, d, col_ssd_extra);
return col_ssd[0] + cache + cache2;
}
y_tex += 1;
template<int RADIUS>
__device__ uint2 MinSSD(volatile unsigned int *col_ssd_cache, volatile unsigned int *col_ssd)
{
unsigned int ssd[N_DISPARITIES];
__syncthreads(); //before MinSSD function
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
ssd[0] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 0 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[1] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 1 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[2] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 2 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[3] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 3 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[4] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 4 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[5] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 5 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[6] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 6 * (BLOCK_W + 2 * RADIUS));
__syncthreads();
ssd[7] = CalcSSD<RADIUS>(col_ssd_cache, col_ssd + 7 * (BLOCK_W + 2 * RADIUS));
if (X < cwidth - RADIUS && row < cheight - RADIUS - Y)
int mssd = ::min(::min(::min(ssd[0], ssd[1]), ::min(ssd[4], ssd[5])), ::min(::min(ssd[2], ssd[3]), ::min(ssd[6], ssd[7])));
int bestIdx = 0;
for (int i = 0; i < N_DISPARITIES; i++)
{
int idx = row * cminSSD_step;
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[idx])
if (mssd == ssd[i])
bestIdx = i;
}
return make_uint2(mssd, bestIdx);
}
template<int RADIUS>
__device__ void StepDown(int idx1, int idx2, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
unsigned char leftPixel1;
unsigned char leftPixel2;
unsigned char rightPixel1[8];
unsigned char rightPixel2[8];
unsigned int diff1, diff2;
leftPixel1 = imageL[idx1];
leftPixel2 = imageL[idx2];
idx1 = idx1 - d;
idx2 = idx2 - d;
rightPixel1[7] = imageR[idx1 - 7];
rightPixel1[0] = imageR[idx1 - 0];
rightPixel1[1] = imageR[idx1 - 1];
rightPixel1[2] = imageR[idx1 - 2];
rightPixel1[3] = imageR[idx1 - 3];
rightPixel1[4] = imageR[idx1 - 4];
rightPixel1[5] = imageR[idx1 - 5];
rightPixel1[6] = imageR[idx1 - 6];
rightPixel2[7] = imageR[idx2 - 7];
rightPixel2[0] = imageR[idx2 - 0];
rightPixel2[1] = imageR[idx2 - 1];
rightPixel2[2] = imageR[idx2 - 2];
rightPixel2[3] = imageR[idx2 - 3];
rightPixel2[4] = imageR[idx2 - 4];
rightPixel2[5] = imageR[idx2 - 5];
rightPixel2[6] = imageR[idx2 - 6];
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
diff1 = leftPixel1 - rightPixel1[0];
diff2 = leftPixel2 - rightPixel2[0];
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[1];
diff2 = leftPixel2 - rightPixel2[1];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[2];
diff2 = leftPixel2 - rightPixel2[2];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[3];
diff2 = leftPixel2 - rightPixel2[3];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[4];
diff2 = leftPixel2 - rightPixel2[4];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[5];
diff2 = leftPixel2 - rightPixel2[5];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[6];
diff2 = leftPixel2 - rightPixel2[6];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
diff1 = leftPixel1 - rightPixel1[7];
diff2 = leftPixel2 - rightPixel2[7];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] += SQ(diff2) - SQ(diff1);
}
template<int RADIUS>
__device__ void InitColSSD(int x_tex, int y_tex, int im_pitch, unsigned char* imageL, unsigned char* imageR, int d, volatile unsigned int *col_ssd)
{
unsigned char leftPixel1;
int idx;
unsigned int diffa[] = {0, 0, 0, 0, 0, 0, 0, 0};
for(int i = 0; i < (2 * RADIUS + 1); i++)
{
idx = y_tex * im_pitch + x_tex;
leftPixel1 = imageL[idx];
idx = idx - d;
diffa[0] += SQ(leftPixel1 - imageR[idx - 0]);
diffa[1] += SQ(leftPixel1 - imageR[idx - 1]);
diffa[2] += SQ(leftPixel1 - imageR[idx - 2]);
diffa[3] += SQ(leftPixel1 - imageR[idx - 3]);
diffa[4] += SQ(leftPixel1 - imageR[idx - 4]);
diffa[5] += SQ(leftPixel1 - imageR[idx - 5]);
diffa[6] += SQ(leftPixel1 - imageR[idx - 6]);
diffa[7] += SQ(leftPixel1 - imageR[idx - 7]);
y_tex += 1;
}
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
col_ssd[0 * (BLOCK_W + 2 * RADIUS)] = diffa[0];
col_ssd[1 * (BLOCK_W + 2 * RADIUS)] = diffa[1];
col_ssd[2 * (BLOCK_W + 2 * RADIUS)] = diffa[2];
col_ssd[3 * (BLOCK_W + 2 * RADIUS)] = diffa[3];
col_ssd[4 * (BLOCK_W + 2 * RADIUS)] = diffa[4];
col_ssd[5 * (BLOCK_W + 2 * RADIUS)] = diffa[5];
col_ssd[6 * (BLOCK_W + 2 * RADIUS)] = diffa[6];
col_ssd[7 * (BLOCK_W + 2 * RADIUS)] = diffa[7];
}
template<int RADIUS>
__global__ void stereoKernel(unsigned char *left, unsigned char *right, size_t img_step, PtrStepb disp, int maxdisp)
{
extern __shared__ unsigned int col_ssd_cache[];
volatile unsigned int *col_ssd = col_ssd_cache + BLOCK_W + threadIdx.x;
volatile unsigned int *col_ssd_extra = threadIdx.x < (2 * RADIUS) ? col_ssd + BLOCK_W : 0; //#define N_DIRTY_PIXELS (2 * RADIUS)
//#define X (blockIdx.x * BLOCK_W + threadIdx.x + STEREO_MAXD)
int X = (blockIdx.x * BLOCK_W + threadIdx.x + maxdisp + RADIUS);
//#define Y (__mul24(blockIdx.y, ROWSperTHREAD) + RADIUS)
#define Y (blockIdx.y * ROWSperTHREAD + RADIUS)
//int Y = blockIdx.y * ROWSperTHREAD + RADIUS;
unsigned int* minSSDImage = cminSSDImage + X + Y * cminSSD_step;
unsigned char* disparImage = disp.data + X + Y * disp.step;
/* if (X < cwidth)
{
unsigned int *minSSDImage_end = minSSDImage + min(ROWSperTHREAD, cheight - Y) * minssd_step;
for(uint *ptr = minSSDImage; ptr != minSSDImage_end; ptr += minssd_step )
*ptr = 0xFFFFFFFF;
}*/
int end_row = ::min(ROWSperTHREAD, cheight - Y - RADIUS);
int y_tex;
int x_tex = X - RADIUS;
if (x_tex >= cwidth)
return;
for(int d = STEREO_MIND; d < maxdisp; d += STEREO_DISP_STEP)
{
y_tex = Y - RADIUS;
InitColSSD<RADIUS>(x_tex, y_tex, img_step, left, right, d, col_ssd);
if (col_ssd_extra > 0)
if (x_tex + BLOCK_W < cwidth)
InitColSSD<RADIUS>(x_tex + BLOCK_W, y_tex, img_step, left, right, d, col_ssd_extra);
__syncthreads(); //before MinSSD function
if (X < cwidth - RADIUS && Y < cheight - RADIUS)
{
disparImage[disp.step * row] = (unsigned char)(d + minSSD.y);
minSSDImage[idx] = minSSD.x;
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[0])
{
disparImage[0] = (unsigned char)(d + minSSD.y);
minSSDImage[0] = minSSD.x;
}
}
for(int row = 1; row < end_row; row++)
{
int idx1 = y_tex * img_step + x_tex;
int idx2 = (y_tex + (2 * RADIUS + 1)) * img_step + x_tex;
__syncthreads();
StepDown<RADIUS>(idx1, idx2, left, right, d, col_ssd);
if (col_ssd_extra)
if (x_tex + BLOCK_W < cwidth)
StepDown<RADIUS>(idx1, idx2, left + BLOCK_W, right + BLOCK_W, d, col_ssd_extra);
y_tex += 1;
__syncthreads(); //before MinSSD function
if (X < cwidth - RADIUS && row < cheight - RADIUS - Y)
{
int idx = row * cminSSD_step;
uint2 minSSD = MinSSD<RADIUS>(col_ssd_cache + threadIdx.x, col_ssd);
if (minSSD.x < minSSDImage[idx])
{
disparImage[disp.step * row] = (unsigned char)(d + minSSD.y);
minSSDImage[idx] = minSSD.x;
}
}
} // for row loop
} // for d loop
}
template<int RADIUS> void kernel_caller(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream)
{
dim3 grid(1,1,1);
dim3 threads(BLOCK_W, 1, 1);
grid.x = divUp(left.cols - maxdisp - 2 * RADIUS, BLOCK_W);
grid.y = divUp(left.rows - 2 * RADIUS, ROWSperTHREAD);
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
size_t smem_size = (BLOCK_W + N_DISPARITIES * (BLOCK_W + 2 * RADIUS)) * sizeof(unsigned int);
stereoKernel<RADIUS><<<grid, threads, smem_size, stream>>>(left.data, right.data, left.step, disp, maxdisp);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
};
typedef void (*kernel_caller_t)(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream);
const static kernel_caller_t callers[] =
{
0,
kernel_caller< 1>, kernel_caller< 2>, kernel_caller< 3>, kernel_caller< 4>, kernel_caller< 5>,
kernel_caller< 6>, kernel_caller< 7>, kernel_caller< 8>, kernel_caller< 9>, kernel_caller<10>,
kernel_caller<11>, kernel_caller<12>, kernel_caller<13>, kernel_caller<15>, kernel_caller<15>,
kernel_caller<16>, kernel_caller<17>, kernel_caller<18>, kernel_caller<19>, kernel_caller<20>,
kernel_caller<21>, kernel_caller<22>, kernel_caller<23>, kernel_caller<24>, kernel_caller<25>
//0,0,0, 0,0,0, 0,0,kernel_caller<9>
};
const int calles_num = sizeof(callers)/sizeof(callers[0]);
void stereoBM_GPU(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, int winsz, const DevMem2D_<unsigned int>& minSSD_buf, cudaStream_t& stream)
{
int winsz2 = winsz >> 1;
if (winsz2 == 0 || winsz2 >= calles_num)
cv::gpu::error("Unsupported window size", __FILE__, __LINE__);
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferL1) );
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferShared) );
cudaSafeCall( cudaMemset2D(disp.data, disp.step, 0, disp.cols, disp.rows) );
cudaSafeCall( cudaMemset2D(minSSD_buf.data, minSSD_buf.step, 0xFF, minSSD_buf.cols * minSSD_buf.elemSize(), disp.rows) );
cudaSafeCall( cudaMemcpyToSymbol( cwidth, &left.cols, sizeof(left.cols) ) );
cudaSafeCall( cudaMemcpyToSymbol( cheight, &left.rows, sizeof(left.rows) ) );
cudaSafeCall( cudaMemcpyToSymbol( cminSSDImage, &minSSD_buf.data, sizeof(minSSD_buf.data) ) );
size_t minssd_step = minSSD_buf.step/minSSD_buf.elemSize();
cudaSafeCall( cudaMemcpyToSymbol( cminSSD_step, &minssd_step, sizeof(minssd_step) ) );
callers[winsz2](left, right, disp, maxdisp, stream);
}
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Sobel Prefiler ///////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
texture<unsigned char, 2, cudaReadModeElementType> texForSobel;
__global__ void prefilter_kernel(DevMem2Db output, int prefilterCap)
{
int x = blockDim.x * blockIdx.x + threadIdx.x;
int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < output.cols && y < output.rows)
{
int conv = (int)tex2D(texForSobel, x - 1, y - 1) * (-1) + (int)tex2D(texForSobel, x + 1, y - 1) * (1) +
(int)tex2D(texForSobel, x - 1, y ) * (-2) + (int)tex2D(texForSobel, x + 1, y ) * (2) +
(int)tex2D(texForSobel, x - 1, y + 1) * (-1) + (int)tex2D(texForSobel, x + 1, y + 1) * (1);
conv = ::min(::min(::max(-prefilterCap, conv), prefilterCap) + prefilterCap, 255);
output.ptr(y)[x] = conv & 0xFF;
}
}
void prefilter_xsobel(const DevMem2Db& input, const DevMem2Db& output, int prefilterCap, cudaStream_t & stream)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForSobel, input.data, desc, input.cols, input.rows, input.step ) );
dim3 threads(16, 16, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, threads.y);
prefilter_kernel<<<grid, threads, 0, stream>>>(output, prefilterCap);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
cudaSafeCall( cudaUnbindTexture (texForSobel ) );
}
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////// Textureness filtering ////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
texture<unsigned char, 2, cudaReadModeNormalizedFloat> texForTF;
__device__ __forceinline__ float sobel(int x, int y)
{
float conv = tex2D(texForTF, x - 1, y - 1) * (-1) + tex2D(texForTF, x + 1, y - 1) * (1) +
tex2D(texForTF, x - 1, y ) * (-2) + tex2D(texForTF, x + 1, y ) * (2) +
tex2D(texForTF, x - 1, y + 1) * (-1) + tex2D(texForTF, x + 1, y + 1) * (1);
return fabs(conv);
}
__device__ float CalcSums(float *cols, float *cols_cache, int winsz)
{
float cache = 0;
float cache2 = 0;
int winsz2 = winsz/2;
for(int i = 1; i <= winsz2; i++)
cache += cols[i];
cols_cache[0] = cache;
__syncthreads();
if (threadIdx.x < blockDim.x - winsz2)
cache2 = cols_cache[winsz2];
else
for(int i = winsz2 + 1; i < winsz; i++)
cache2 += cols[i];
return cols[0] + cache + cache2;
}
#define RpT (2 * ROWSperTHREAD) // got experimentally
__global__ void textureness_kernel(DevMem2Db disp, int winsz, float threshold)
{
int winsz2 = winsz/2;
int n_dirty_pixels = (winsz2) * 2;
extern __shared__ float cols_cache[];
float *cols = cols_cache + blockDim.x + threadIdx.x;
float *cols_extra = threadIdx.x < n_dirty_pixels ? cols + blockDim.x : 0;
int x = blockIdx.x * blockDim.x + threadIdx.x;
int beg_row = blockIdx.y * RpT;
int end_row = ::min(beg_row + RpT, disp.rows);
if (x < disp.cols)
{
int y = beg_row;
float sum = 0;
float sum_extra = 0;
for(int i = y - winsz2; i <= y + winsz2; ++i)
{
sum += sobel(x - winsz2, i);
if (cols_extra)
sum_extra += sobel(x + blockDim.x - winsz2, i);
}
*cols = sum;
if (cols_extra)
*cols_extra = sum_extra;
__syncthreads();
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
__syncthreads();
for(int y = beg_row + 1; y < end_row; ++y)
{
sum = sum - sobel(x - winsz2, y - winsz2 - 1) + sobel(x - winsz2, y + winsz2);
*cols = sum;
if (cols_extra)
{
sum_extra = sum_extra - sobel(x + blockDim.x - winsz2, y - winsz2 - 1) + sobel(x + blockDim.x - winsz2, y + winsz2);
*cols_extra = sum_extra;
}
__syncthreads();
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
__syncthreads();
}
}
} // for row loop
} // for d loop
}
template<int RADIUS> void kernel_caller(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream)
{
dim3 grid(1,1,1);
dim3 threads(BLOCK_W, 1, 1);
grid.x = divUp(left.cols - maxdisp - 2 * RADIUS, BLOCK_W);
grid.y = divUp(left.rows - 2 * RADIUS, ROWSperTHREAD);
//See above: #define COL_SSD_SIZE (BLOCK_W + 2 * RADIUS)
size_t smem_size = (BLOCK_W + N_DISPARITIES * (BLOCK_W + 2 * RADIUS)) * sizeof(unsigned int);
stereoKernel<RADIUS><<<grid, threads, smem_size, stream>>>(left.data, right.data, left.step, disp, maxdisp);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
};
typedef void (*kernel_caller_t)(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, cudaStream_t & stream);
const static kernel_caller_t callers[] =
{
0,
kernel_caller< 1>, kernel_caller< 2>, kernel_caller< 3>, kernel_caller< 4>, kernel_caller< 5>,
kernel_caller< 6>, kernel_caller< 7>, kernel_caller< 8>, kernel_caller< 9>, kernel_caller<10>,
kernel_caller<11>, kernel_caller<12>, kernel_caller<13>, kernel_caller<15>, kernel_caller<15>,
kernel_caller<16>, kernel_caller<17>, kernel_caller<18>, kernel_caller<19>, kernel_caller<20>,
kernel_caller<21>, kernel_caller<22>, kernel_caller<23>, kernel_caller<24>, kernel_caller<25>
//0,0,0, 0,0,0, 0,0,kernel_caller<9>
};
const int calles_num = sizeof(callers)/sizeof(callers[0]);
void stereoBM_GPU(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& disp, int maxdisp, int winsz, const DevMem2D_<unsigned int>& minSSD_buf, cudaStream_t& stream)
{
int winsz2 = winsz >> 1;
if (winsz2 == 0 || winsz2 >= calles_num)
cv::gpu::error("Unsupported window size", __FILE__, __LINE__);
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferL1) );
//cudaSafeCall( cudaFuncSetCacheConfig(&stereoKernel, cudaFuncCachePreferShared) );
cudaSafeCall( cudaMemset2D(disp.data, disp.step, 0, disp.cols, disp.rows) );
cudaSafeCall( cudaMemset2D(minSSD_buf.data, minSSD_buf.step, 0xFF, minSSD_buf.cols * minSSD_buf.elemSize(), disp.rows) );
cudaSafeCall( cudaMemcpyToSymbol( cwidth, &left.cols, sizeof(left.cols) ) );
cudaSafeCall( cudaMemcpyToSymbol( cheight, &left.rows, sizeof(left.rows) ) );
cudaSafeCall( cudaMemcpyToSymbol( cminSSDImage, &minSSD_buf.data, sizeof(minSSD_buf.data) ) );
size_t minssd_step = minSSD_buf.step/minSSD_buf.elemSize();
cudaSafeCall( cudaMemcpyToSymbol( cminSSD_step, &minssd_step, sizeof(minssd_step) ) );
callers[winsz2](left, right, disp, maxdisp, stream);
}
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Sobel Prefiler ///////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
texture<unsigned char, 2, cudaReadModeElementType> texForSobel;
__global__ void prefilter_kernel(DevMem2Db output, int prefilterCap)
{
int x = blockDim.x * blockIdx.x + threadIdx.x;
int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < output.cols && y < output.rows)
{
int conv = (int)tex2D(texForSobel, x - 1, y - 1) * (-1) + (int)tex2D(texForSobel, x + 1, y - 1) * (1) +
(int)tex2D(texForSobel, x - 1, y ) * (-2) + (int)tex2D(texForSobel, x + 1, y ) * (2) +
(int)tex2D(texForSobel, x - 1, y + 1) * (-1) + (int)tex2D(texForSobel, x + 1, y + 1) * (1);
conv = ::min(::min(::max(-prefilterCap, conv), prefilterCap) + prefilterCap, 255);
output.ptr(y)[x] = conv & 0xFF;
}
}
void prefilter_xsobel(const DevMem2Db& input, const DevMem2Db& output, int prefilterCap, cudaStream_t & stream)
{
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForSobel, input.data, desc, input.cols, input.rows, input.step ) );
dim3 threads(16, 16, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, threads.y);
prefilter_kernel<<<grid, threads, 0, stream>>>(output, prefilterCap);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
cudaSafeCall( cudaUnbindTexture (texForSobel ) );
}
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////// Textureness filtering ////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
texture<unsigned char, 2, cudaReadModeNormalizedFloat> texForTF;
__device__ __forceinline__ float sobel(int x, int y)
{
float conv = tex2D(texForTF, x - 1, y - 1) * (-1) + tex2D(texForTF, x + 1, y - 1) * (1) +
tex2D(texForTF, x - 1, y ) * (-2) + tex2D(texForTF, x + 1, y ) * (2) +
tex2D(texForTF, x - 1, y + 1) * (-1) + tex2D(texForTF, x + 1, y + 1) * (1);
return fabs(conv);
}
__device__ float CalcSums(float *cols, float *cols_cache, int winsz)
{
float cache = 0;
float cache2 = 0;
int winsz2 = winsz/2;
for(int i = 1; i <= winsz2; i++)
cache += cols[i];
cols_cache[0] = cache;
__syncthreads();
if (threadIdx.x < blockDim.x - winsz2)
cache2 = cols_cache[winsz2];
else
for(int i = winsz2 + 1; i < winsz; i++)
cache2 += cols[i];
return cols[0] + cache + cache2;
}
#define RpT (2 * ROWSperTHREAD) // got experimentally
__global__ void textureness_kernel(DevMem2Db disp, int winsz, float threshold)
{
int winsz2 = winsz/2;
int n_dirty_pixels = (winsz2) * 2;
extern __shared__ float cols_cache[];
float *cols = cols_cache + blockDim.x + threadIdx.x;
float *cols_extra = threadIdx.x < n_dirty_pixels ? cols + blockDim.x : 0;
int x = blockIdx.x * blockDim.x + threadIdx.x;
int beg_row = blockIdx.y * RpT;
int end_row = ::min(beg_row + RpT, disp.rows);
if (x < disp.cols)
{
int y = beg_row;
float sum = 0;
float sum_extra = 0;
for(int i = y - winsz2; i <= y + winsz2; ++i)
{
sum += sobel(x - winsz2, i);
if (cols_extra)
sum_extra += sobel(x + blockDim.x - winsz2, i);
}
*cols = sum;
if (cols_extra)
*cols_extra = sum_extra;
__syncthreads();
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
__syncthreads();
for(int y = beg_row + 1; y < end_row; ++y)
void postfilter_textureness(const DevMem2Db& input, int winsz, float avgTexturenessThreshold, const DevMem2Db& disp, cudaStream_t & stream)
{
sum = sum - sobel(x - winsz2, y - winsz2 - 1) + sobel(x - winsz2, y + winsz2);
*cols = sum;
avgTexturenessThreshold *= winsz * winsz;
if (cols_extra)
{
sum_extra = sum_extra - sobel(x + blockDim.x - winsz2, y - winsz2 - 1) + sobel(x + blockDim.x - winsz2, y + winsz2);
*cols_extra = sum_extra;
}
texForTF.filterMode = cudaFilterModeLinear;
texForTF.addressMode[0] = cudaAddressModeWrap;
texForTF.addressMode[1] = cudaAddressModeWrap;
__syncthreads();
float sum_win = CalcSums(cols, cols_cache + threadIdx.x, winsz) * 255;
if (sum_win < threshold)
disp.data[y * disp.step + x] = 0;
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForTF, input.data, desc, input.cols, input.rows, input.step ) );
__syncthreads();
dim3 threads(128, 1, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, RpT);
size_t smem_size = (threads.x + threads.x + (winsz/2) * 2 ) * sizeof(float);
textureness_kernel<<<grid, threads, smem_size, stream>>>(disp, winsz, avgTexturenessThreshold);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
cudaSafeCall( cudaUnbindTexture (texForTF) );
}
}
}
void postfilter_textureness(const DevMem2Db& input, int winsz, float avgTexturenessThreshold, const DevMem2Db& disp, cudaStream_t & stream)
{
avgTexturenessThreshold *= winsz * winsz;
texForTF.filterMode = cudaFilterModeLinear;
texForTF.addressMode[0] = cudaAddressModeWrap;
texForTF.addressMode[1] = cudaAddressModeWrap;
cudaChannelFormatDesc desc = cudaCreateChannelDesc<unsigned char>();
cudaSafeCall( cudaBindTexture2D( 0, texForTF, input.data, desc, input.cols, input.rows, input.step ) );
dim3 threads(128, 1, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(input.cols, threads.x);
grid.y = divUp(input.rows, RpT);
size_t smem_size = (threads.x + threads.x + (winsz/2) * 2 ) * sizeof(float);
textureness_kernel<<<grid, threads, smem_size, stream>>>(disp, winsz, avgTexturenessThreshold);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
cudaSafeCall( cudaUnbindTexture (texForTF) );
}
} // namespace stereobm
END_OPENCV_DEVICE_NAMESPACE
} // namespace stereobm
}}} // namespace cv { namespace gpu { namespace device

928
modules/gpu/src/cuda/stereobp.cu
View File

@@ -44,489 +44,487 @@
#include "opencv2/gpu/device/saturate_cast.hpp"
#include "opencv2/gpu/device/limits.hpp"
BEGIN_OPENCV_DEVICE_NAMESPACE
namespace stereobp {
///////////////////////////////////////////////////////////////
/////////////////////// load constants ////////////////////////
///////////////////////////////////////////////////////////////
__constant__ int cndisp;
__constant__ float cmax_data_term;
__constant__ float cdata_weight;
__constant__ float cmax_disc_term;
__constant__ float cdisc_single_jump;
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump)
namespace cv { namespace gpu { namespace device
{
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int )) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_data_term, &max_data_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdata_weight, &data_weight, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc_term, &max_disc_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdisc_single_jump, &disc_single_jump, sizeof(float)) );
}
///////////////////////////////////////////////////////////////
////////////////////////// comp data //////////////////////////
///////////////////////////////////////////////////////////////
template <int cn> struct PixDiff;
template <> struct PixDiff<1>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
namespace stereobp
{
l = *ls;
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
return ::abs((int)l - *rs);
}
uchar l;
};
template <> struct PixDiff<3>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *((uchar3*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
///////////////////////////////////////////////////////////////
/////////////////////// load constants ////////////////////////
///////////////////////////////////////////////////////////////
float val = tb * ::abs((int)l.x - rs[0]);
val += tg * ::abs((int)l.y - rs[1]);
val += tr * ::abs((int)l.z - rs[2]);
__constant__ int cndisp;
__constant__ float cmax_data_term;
__constant__ float cdata_weight;
__constant__ float cmax_disc_term;
__constant__ float cdisc_single_jump;
return val;
}
uchar3 l;
};
template <> struct PixDiff<4>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *((uchar4*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
uchar4 r = *((uchar4*)rs);
float val = tb * ::abs((int)l.x - r.x);
val += tg * ::abs((int)l.y - r.y);
val += tr * ::abs((int)l.z - r.z);
return val;
}
uchar4 l;
};
template <int cn, typename D>
__global__ void comp_data(const DevMem2Db left, const PtrStepb right, PtrElemStep_<D> data)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y > 0 && y < left.rows - 1 && x > 0 && x < left.cols - 1)
{
const uchar* ls = left.ptr(y) + x * cn;
const PixDiff<cn> pixDiff(ls);
const uchar* rs = right.ptr(y) + x * cn;
D* ds = data.ptr(y) + x;
const size_t disp_step = data.step * left.rows;
for (int disp = 0; disp < cndisp; disp++)
void load_constants(int ndisp, float max_data_term, float data_weight, float max_disc_term, float disc_single_jump)
{
if (x - disp >= 1)
{
float val = pixDiff(rs - disp * cn);
cudaSafeCall( cudaMemcpyToSymbol(cndisp, &ndisp, sizeof(int )) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_data_term, &max_data_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdata_weight, &data_weight, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cmax_disc_term, &max_disc_term, sizeof(float)) );
cudaSafeCall( cudaMemcpyToSymbol(cdisc_single_jump, &disc_single_jump, sizeof(float)) );
}
ds[disp * disp_step] = saturate_cast<D>(fmin(cdata_weight * val, cdata_weight * cmax_data_term));
///////////////////////////////////////////////////////////////
////////////////////////// comp data //////////////////////////
///////////////////////////////////////////////////////////////
template <int cn> struct PixDiff;
template <> struct PixDiff<1>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *ls;
}
else
__device__ __forceinline__ float operator()(const uchar* rs) const
{
ds[disp * disp_step] = saturate_cast<D>(cdata_weight * cmax_data_term);
return ::abs((int)l - *rs);
}
}
}
}
template<typename T, typename D>
void comp_data_gpu(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream);
template <> void comp_data_gpu<uchar, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<1, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<1, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar3, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<3, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar3, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<3, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar4, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<4, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar4, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<4, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////
//////////////////////// data step down ///////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__global__ void data_step_down(int dst_cols, int dst_rows, int src_rows, const PtrStep<T> src, PtrStep<T> dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < dst_cols && y < dst_rows)
{
for (int d = 0; d < cndisp; ++d)
uchar l;
};
template <> struct PixDiff<3>
{
float dst_reg = src.ptr(d * src_rows + (2*y+0))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+0))[(2*x+1)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+1)];
dst.ptr(d * dst_rows + y)[x] = saturate_cast<T>(dst_reg);
}
}
}
template<typename T>
void data_step_down_gpu(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
data_step_down<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)src, (DevMem2D_<T>)dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void data_step_down_gpu<short>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
template void data_step_down_gpu<float>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
///////////////////////////////////////////////////////////////
/////////////////// level up messages ////////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__global__ void level_up_message(int dst_cols, int dst_rows, int src_rows, const PtrElemStep_<T> src, PtrElemStep_<T> dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < dst_cols && y < dst_rows)
{
const size_t dst_disp_step = dst.step * dst_rows;
const size_t src_disp_step = src.step * src_rows;
T* dstr = dst.ptr(y ) + x;
const T* srcr = src.ptr(y/2) + x/2;
for (int d = 0; d < cndisp; ++d)
dstr[d * dst_disp_step] = srcr[d * src_disp_step];
}
}
template <typename T>
void level_up_messages_gpu(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
int src_idx = (dst_idx + 1) & 1;
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mus[src_idx], (DevMem2D_<T>)mus[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mds[src_idx], (DevMem2D_<T>)mds[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mls[src_idx], (DevMem2D_<T>)mls[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mrs[src_idx], (DevMem2D_<T>)mrs[dst_idx]);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void level_up_messages_gpu<short>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
template void level_up_messages_gpu<float>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
///////////////////////////////////////////////////////////////
//////////////////// calc all iterations /////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__device__ void calc_min_linear_penalty(T* dst, size_t step)
{
float prev = dst[0];
float cur;
for (int disp = 1; disp < cndisp; ++disp)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
}
prev = cur;
}
prev = dst[(cndisp - 1) * step];
for (int disp = cndisp - 2; disp >= 0; disp--)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
}
prev = cur;
}
}
template <typename T>
__device__ void message(const T* msg1, const T* msg2, const T* msg3, const T* data, T* dst, size_t msg_disp_step, size_t data_disp_step)
{
float minimum = device::numeric_limits<float>::max();
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = msg1[msg_disp_step * i];
dst_reg += msg2[msg_disp_step * i];
dst_reg += msg3[msg_disp_step * i];
dst_reg += data[data_disp_step * i];
if (dst_reg < minimum)
minimum = dst_reg;
dst[msg_disp_step * i] = saturate_cast<T>(dst_reg);
}
calc_min_linear_penalty(dst, msg_disp_step);
minimum += cmax_disc_term;
float sum = 0;
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = dst[msg_disp_step * i];
if (dst_reg > minimum)
{
dst_reg = minimum;
dst[msg_disp_step * i] = saturate_cast<T>(minimum);
}
sum += dst_reg;
}
sum /= cndisp;
for(int i = 0; i < cndisp; ++i)
dst[msg_disp_step * i] -= sum;
}
template <typename T>
__global__ void one_iteration(int t, PtrElemStep_<T> u, T* d, T* l, T* r, const PtrElemStep_<T> data, int cols, int rows)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
if ((y > 0) && (y < rows - 1) && (x > 0) && (x < cols - 1))
{
T* us = u.ptr(y) + x;
T* ds = d + y * u.step + x;
T* ls = l + y * u.step + x;
T* rs = r + y * u.step + x;
const T* dt = data.ptr(y) + x;
size_t msg_disp_step = u.step * rows;
size_t data_disp_step = data.step * rows;
message(us + u.step, ls + 1, rs - 1, dt, us, msg_disp_step, data_disp_step);
message(ds - u.step, ls + 1, rs - 1, dt, ds, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, rs - 1, dt, rs, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, ls + 1, dt, ls, msg_disp_step, data_disp_step);
}
}
template <typename T>
void calc_all_iterations_gpu(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d,
const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(cols, threads.x << 1);
grid.y = divUp(rows, threads.y);
for(int t = 0; t < iters; ++t)
{
one_iteration<T><<<grid, threads, 0, stream>>>(t, (DevMem2D_<T>)u, (T*)d.data, (T*)l.data, (T*)r.data, (DevMem2D_<T>)data, cols, rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
template void calc_all_iterations_gpu<short>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
template void calc_all_iterations_gpu<float>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
///////////////////////////////////////////////////////////////
/////////////////////////// output ////////////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__global__ void output(const PtrElemStep_<T> u, const T* d, const T* l, const T* r, const T* data,
DevMem2D_<short> disp)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y > 0 && y < disp.rows - 1 && x > 0 && x < disp.cols - 1)
{
const T* us = u.ptr(y + 1) + x;
const T* ds = d + (y - 1) * u.step + x;
const T* ls = l + y * u.step + (x + 1);
const T* rs = r + y * u.step + (x - 1);
const T* dt = data + y * u.step + x;
size_t disp_step = disp.rows * u.step;
int best = 0;
float best_val = numeric_limits<float>::max();
for (int d = 0; d < cndisp; ++d)
{
float val = us[d * disp_step];
val += ds[d * disp_step];
val += ls[d * disp_step];
val += rs[d * disp_step];
val += dt[d * disp_step];
if (val < best_val)
__device__ __forceinline__ PixDiff(const uchar* ls)
{
best_val = val;
best = d;
l = *((uchar3*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
float val = tb * ::abs((int)l.x - rs[0]);
val += tg * ::abs((int)l.y - rs[1]);
val += tr * ::abs((int)l.z - rs[2]);
return val;
}
uchar3 l;
};
template <> struct PixDiff<4>
{
__device__ __forceinline__ PixDiff(const uchar* ls)
{
l = *((uchar4*)ls);
}
__device__ __forceinline__ float operator()(const uchar* rs) const
{
const float tr = 0.299f;
const float tg = 0.587f;
const float tb = 0.114f;
uchar4 r = *((uchar4*)rs);
float val = tb * ::abs((int)l.x - r.x);
val += tg * ::abs((int)l.y - r.y);
val += tr * ::abs((int)l.z - r.z);
return val;
}
uchar4 l;
};
template <int cn, typename D>
__global__ void comp_data(const DevMem2Db left, const PtrStepb right, PtrElemStep_<D> data)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y > 0 && y < left.rows - 1 && x > 0 && x < left.cols - 1)
{
const uchar* ls = left.ptr(y) + x * cn;
const PixDiff<cn> pixDiff(ls);
const uchar* rs = right.ptr(y) + x * cn;
D* ds = data.ptr(y) + x;
const size_t disp_step = data.step * left.rows;
for (int disp = 0; disp < cndisp; disp++)
{
if (x - disp >= 1)
{
float val = pixDiff(rs - disp * cn);
ds[disp * disp_step] = saturate_cast<D>(fmin(cdata_weight * val, cdata_weight * cmax_data_term));
}
else
{
ds[disp * disp_step] = saturate_cast<D>(cdata_weight * cmax_data_term);
}
}
}
}
disp.ptr(y)[x] = saturate_cast<short>(best);
}
}
template<typename T, typename D>
void comp_data_gpu(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream);
template <typename T>
void output_gpu(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data,
const DevMem2D_<short>& disp, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
template <> void comp_data_gpu<uchar, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(disp.cols, threads.x);
grid.y = divUp(disp.rows, threads.y);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
output<T><<<grid, threads, 0, stream>>>((DevMem2D_<T>)u, (const T*)d.data, (const T*)l.data, (const T*)r.data, (const T*)data.data, disp);
cudaSafeCall( cudaGetLastError() );
comp_data<1, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
template void output_gpu<short>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
template void output_gpu<float>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
} // namespace stereobp
comp_data<1, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
END_OPENCV_DEVICE_NAMESPACE
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar3, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<3, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar3, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<3, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar4, short>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<4, short><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<short>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template <> void comp_data_gpu<uchar4, float>(const DevMem2Db& left, const DevMem2Db& right, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(left.cols, threads.x);
grid.y = divUp(left.rows, threads.y);
comp_data<4, float><<<grid, threads, 0, stream>>>(left, right, (DevMem2D_<float>)data);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
///////////////////////////////////////////////////////////////
//////////////////////// data step down ///////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__global__ void data_step_down(int dst_cols, int dst_rows, int src_rows, const PtrStep<T> src, PtrStep<T> dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < dst_cols && y < dst_rows)
{
for (int d = 0; d < cndisp; ++d)
{
float dst_reg = src.ptr(d * src_rows + (2*y+0))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+0)];
dst_reg += src.ptr(d * src_rows + (2*y+0))[(2*x+1)];
dst_reg += src.ptr(d * src_rows + (2*y+1))[(2*x+1)];
dst.ptr(d * dst_rows + y)[x] = saturate_cast<T>(dst_reg);
}
}
}
template<typename T>
void data_step_down_gpu(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
data_step_down<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)src, (DevMem2D_<T>)dst);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void data_step_down_gpu<short>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
template void data_step_down_gpu<float>(int dst_cols, int dst_rows, int src_rows, const DevMem2Db& src, const DevMem2Db& dst, cudaStream_t stream);
///////////////////////////////////////////////////////////////
/////////////////// level up messages ////////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__global__ void level_up_message(int dst_cols, int dst_rows, int src_rows, const PtrElemStep_<T> src, PtrElemStep_<T> dst)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < dst_cols && y < dst_rows)
{
const size_t dst_disp_step = dst.step * dst_rows;
const size_t src_disp_step = src.step * src_rows;
T* dstr = dst.ptr(y ) + x;
const T* srcr = src.ptr(y/2) + x/2;
for (int d = 0; d < cndisp; ++d)
dstr[d * dst_disp_step] = srcr[d * src_disp_step];
}
}
template <typename T>
void level_up_messages_gpu(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(dst_cols, threads.x);
grid.y = divUp(dst_rows, threads.y);
int src_idx = (dst_idx + 1) & 1;
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mus[src_idx], (DevMem2D_<T>)mus[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mds[src_idx], (DevMem2D_<T>)mds[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mls[src_idx], (DevMem2D_<T>)mls[dst_idx]);
cudaSafeCall( cudaGetLastError() );
level_up_message<T><<<grid, threads, 0, stream>>>(dst_cols, dst_rows, src_rows, (DevMem2D_<T>)mrs[src_idx], (DevMem2D_<T>)mrs[dst_idx]);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void level_up_messages_gpu<short>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
template void level_up_messages_gpu<float>(int dst_idx, int dst_cols, int dst_rows, int src_rows, DevMem2Db* mus, DevMem2Db* mds, DevMem2Db* mls, DevMem2Db* mrs, cudaStream_t stream);
///////////////////////////////////////////////////////////////
//////////////////// calc all iterations /////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__device__ void calc_min_linear_penalty(T* dst, size_t step)
{
float prev = dst[0];
float cur;
for (int disp = 1; disp < cndisp; ++disp)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
}
prev = cur;
}
prev = dst[(cndisp - 1) * step];
for (int disp = cndisp - 2; disp >= 0; disp--)
{
prev += cdisc_single_jump;
cur = dst[step * disp];
if (prev < cur)
{
cur = prev;
dst[step * disp] = saturate_cast<T>(prev);
}
prev = cur;
}
}
template <typename T>
__device__ void message(const T* msg1, const T* msg2, const T* msg3, const T* data, T* dst, size_t msg_disp_step, size_t data_disp_step)
{
float minimum = device::numeric_limits<float>::max();
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = msg1[msg_disp_step * i];
dst_reg += msg2[msg_disp_step * i];
dst_reg += msg3[msg_disp_step * i];
dst_reg += data[data_disp_step * i];
if (dst_reg < minimum)
minimum = dst_reg;
dst[msg_disp_step * i] = saturate_cast<T>(dst_reg);
}
calc_min_linear_penalty(dst, msg_disp_step);
minimum += cmax_disc_term;
float sum = 0;
for(int i = 0; i < cndisp; ++i)
{
float dst_reg = dst[msg_disp_step * i];
if (dst_reg > minimum)
{
dst_reg = minimum;
dst[msg_disp_step * i] = saturate_cast<T>(minimum);
}
sum += dst_reg;
}
sum /= cndisp;
for(int i = 0; i < cndisp; ++i)
dst[msg_disp_step * i] -= sum;
}
template <typename T>
__global__ void one_iteration(int t, PtrElemStep_<T> u, T* d, T* l, T* r, const PtrElemStep_<T> data, int cols, int rows)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = ((blockIdx.x * blockDim.x + threadIdx.x) << 1) + ((y + t) & 1);
if ((y > 0) && (y < rows - 1) && (x > 0) && (x < cols - 1))
{
T* us = u.ptr(y) + x;
T* ds = d + y * u.step + x;
T* ls = l + y * u.step + x;
T* rs = r + y * u.step + x;
const T* dt = data.ptr(y) + x;
size_t msg_disp_step = u.step * rows;
size_t data_disp_step = data.step * rows;
message(us + u.step, ls + 1, rs - 1, dt, us, msg_disp_step, data_disp_step);
message(ds - u.step, ls + 1, rs - 1, dt, ds, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, rs - 1, dt, rs, msg_disp_step, data_disp_step);
message(us + u.step, ds - u.step, ls + 1, dt, ls, msg_disp_step, data_disp_step);
}
}
template <typename T>
void calc_all_iterations_gpu(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d,
const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(cols, threads.x << 1);
grid.y = divUp(rows, threads.y);
for(int t = 0; t < iters; ++t)
{
one_iteration<T><<<grid, threads, 0, stream>>>(t, (DevMem2D_<T>)u, (T*)d.data, (T*)l.data, (T*)r.data, (DevMem2D_<T>)data, cols, rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
}
template void calc_all_iterations_gpu<short>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
template void calc_all_iterations_gpu<float>(int cols, int rows, int iters, const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, cudaStream_t stream);
///////////////////////////////////////////////////////////////
/////////////////////////// output ////////////////////////////
///////////////////////////////////////////////////////////////
template <typename T>
__global__ void output(const PtrElemStep_<T> u, const T* d, const T* l, const T* r, const T* data,
DevMem2D_<short> disp)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
if (y > 0 && y < disp.rows - 1 && x > 0 && x < disp.cols - 1)
{
const T* us = u.ptr(y + 1) + x;
const T* ds = d + (y - 1) * u.step + x;
const T* ls = l + y * u.step + (x + 1);
const T* rs = r + y * u.step + (x - 1);
const T* dt = data + y * u.step + x;
size_t disp_step = disp.rows * u.step;
int best = 0;
float best_val = numeric_limits<float>::max();
for (int d = 0; d < cndisp; ++d)
{
float val = us[d * disp_step];
val += ds[d * disp_step];
val += ls[d * disp_step];
val += rs[d * disp_step];
val += dt[d * disp_step];
if (val < best_val)
{
best_val = val;
best = d;
}
}
disp.ptr(y)[x] = saturate_cast<short>(best);
}
}
template <typename T>
void output_gpu(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data,
const DevMem2D_<short>& disp, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(disp.cols, threads.x);
grid.y = divUp(disp.rows, threads.y);
output<T><<<grid, threads, 0, stream>>>((DevMem2D_<T>)u, (const T*)d.data, (const T*)l.data, (const T*)r.data, (const T*)data.data, disp);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
template void output_gpu<short>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
template void output_gpu<float>(const DevMem2Db& u, const DevMem2Db& d, const DevMem2Db& l, const DevMem2Db& r, const DevMem2Db& data, const DevMem2D_<short>& disp, cudaStream_t stream);
} // namespace stereobp
}}} // namespace cv { namespace gpu { namespace device

1608
modules/gpu/src/cuda/stereocsbp.cu
View File

File diff suppressed because it is too large Load Diff

1804
modules/gpu/src/cuda/surf.cu
View File

File diff suppressed because it is too large Load Diff