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

optimizations:

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

92
modules/gpu/src/brute_force_matcher.cpp
View File

@@ -88,71 +88,71 @@ namespace cv { namespace gpu { namespace device
{
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
}
namespace bf_knnmatch
{
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void match2L1_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void match2L2_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void match2Hamming_gpu(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
}
namespace bf_radius_match
{
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL1_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchL2_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
template <typename T> void matchHamming_gpu(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
}
}}}
@@ -198,11 +198,11 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchSingle(const GpuMat& query, const
if (query.empty() || train.empty())
return;
using namespace ::cv::gpu::device::bf_match;
using namespace cv::gpu::device::bf_match;
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& train, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callers[3][6] =
{
@@ -234,10 +234,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchSingle(const GpuMat& query, const
caller_t func = callers[distType][query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, train, mask, trainIdx, distance, cc, StreamAccessor::getStream(stream));
func(query, train, mask, trainIdx, distance, StreamAccessor::getStream(stream));
}
void cv::gpu::BruteForceMatcher_GPU_base::matchDownload(const GpuMat& trainIdx, const GpuMat& distance, vector<DMatch>& matches)
@@ -340,11 +337,11 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchCollection(const GpuMat& query, c
if (query.empty() || trainCollection.empty())
return;
using namespace ::cv::gpu::device::bf_match;
using namespace cv::gpu::device::bf_match;
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callers[3][6] =
{
@@ -376,10 +373,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchCollection(const GpuMat& query, c
caller_t func = callers[distType][query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, trainCollection, masks, trainIdx, imgIdx, distance, cc, StreamAccessor::getStream(stream));
func(query, trainCollection, masks, trainIdx, imgIdx, distance, StreamAccessor::getStream(stream));
}
void cv::gpu::BruteForceMatcher_GPU_base::matchDownload(const GpuMat& trainIdx, const GpuMat& imgIdx, const GpuMat& distance, vector<DMatch>& matches)
@@ -451,11 +445,11 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatchSingle(const GpuMat& query, co
if (query.empty() || train.empty())
return;
using namespace ::cv::gpu::device::bf_knnmatch;
using namespace cv::gpu::device::bf_knnmatch;
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& train, int k, const PtrStepSzb& mask,
const PtrStepSzb& trainIdx, const PtrStepSzb& distance, const PtrStepSzf& allDist,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callers[3][6] =
{
@@ -502,10 +496,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatchSingle(const GpuMat& query, co
caller_t func = callers[distType][query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, train, k, mask, trainIdx, distance, allDist, cc, StreamAccessor::getStream(stream));
func(query, train, k, mask, trainIdx, distance, allDist, StreamAccessor::getStream(stream));
}
void cv::gpu::BruteForceMatcher_GPU_base::knnMatchDownload(const GpuMat& trainIdx, const GpuMat& distance,
@@ -580,11 +571,11 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatch2Collection(const GpuMat& quer
if (query.empty() || trainCollection.empty())
return;
using namespace ::cv::gpu::device::bf_knnmatch;
using namespace cv::gpu::device::bf_knnmatch;
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& trains, const PtrStepSz<PtrStepb>& masks,
const PtrStepSzb& trainIdx, const PtrStepSzb& imgIdx, const PtrStepSzb& distance,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callers[3][6] =
{
@@ -621,10 +612,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatch2Collection(const GpuMat& quer
caller_t func = callers[distType][query.depth()];
CV_Assert(func != 0);
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
func(query, trainCollection, maskCollection, trainIdx, imgIdx, distance, cc, StreamAccessor::getStream(stream));
func(query, trainCollection, maskCollection, trainIdx, imgIdx, distance, StreamAccessor::getStream(stream));
}
void cv::gpu::BruteForceMatcher_GPU_base::knnMatch2Download(const GpuMat& trainIdx, const GpuMat& imgIdx, const GpuMat& distance,
@@ -765,7 +753,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchSingle(const GpuMat& query,
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb& train, float maxDistance, const PtrStepSzb& mask,
const PtrStepSzi& trainIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callers[3][6] =
{
@@ -786,12 +774,6 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchSingle(const GpuMat& query,
}
};
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
const int nQuery = query.rows;
const int nTrain = train.rows;
@@ -814,7 +796,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchSingle(const GpuMat& query,
caller_t func = callers[distType][query.depth()];
CV_Assert(func != 0);
func(query, train, maxDistance, mask, trainIdx, distance, nMatches, cc, StreamAccessor::getStream(stream));
func(query, train, maxDistance, mask, trainIdx, distance, nMatches, StreamAccessor::getStream(stream));
}
void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchDownload(const GpuMat& trainIdx, const GpuMat& distance, const GpuMat& nMatches,
@@ -897,7 +879,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchCollection(const GpuMat& qu
typedef void (*caller_t)(const PtrStepSzb& query, const PtrStepSzb* trains, int n, float maxDistance, const PtrStepSzb* masks,
const PtrStepSzi& trainIdx, const PtrStepSzi& imgIdx, const PtrStepSzf& distance, const PtrStepSz<unsigned int>& nMatches,
int cc, cudaStream_t stream);
cudaStream_t stream);
static const caller_t callers[3][6] =
{
@@ -918,12 +900,6 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchCollection(const GpuMat& qu
}
};
DeviceInfo info;
int cc = info.majorVersion() * 10 + info.minorVersion();
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
const int nQuery = query.rows;
CV_Assert(query.channels() == 1 && query.depth() < CV_64F);
@@ -949,7 +925,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchCollection(const GpuMat& qu
vector<PtrStepSzb> masks_(masks.begin(), masks.end());
func(query, &trains_[0], static_cast<int>(trains_.size()), maxDistance, masks_.size() == 0 ? 0 : &masks_[0],
trainIdx, imgIdx, distance, nMatches, cc, StreamAccessor::getStream(stream));
trainIdx, imgIdx, distance, nMatches, StreamAccessor::getStream(stream));
}
void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchDownload(const GpuMat& trainIdx, const GpuMat& imgIdx, const GpuMat& distance, const GpuMat& nMatches,