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

refactor cudaarithm reductions:

Browse Source 
* remove overloads with explicit buffer, now BufferPool is used
* added async versions for all reduce functions
This commit is contained in:
Vladislav Vinogradov
2014-12-24 13:40:33 +03:00
parent a4e598f474
commit cd0e95de16
16 changed files with 1075 additions and 519 deletions
Download Patch File Download Diff File Expand all files Collapse all files

200
modules/cudaarithm/src/reductions.cpp
View File

@@ -47,110 +47,106 @@ using namespace cv::cuda;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
double cv::cuda::norm(InputArray, int, InputArray, GpuMat&) { throw_no_cuda(); return 0.0; }
double cv::cuda::norm(InputArray, InputArray, GpuMat&, int) { throw_no_cuda(); return 0.0; }
double cv::cuda::norm(InputArray, int, InputArray) { throw_no_cuda(); return 0.0; }
void cv::cuda::calcNorm(InputArray, OutputArray, int, InputArray, Stream&) { throw_no_cuda(); }
double cv::cuda::norm(InputArray, InputArray, int) { throw_no_cuda(); return 0.0; }
void cv::cuda::calcNormDiff(InputArray, InputArray, OutputArray, int, Stream&) { throw_no_cuda(); }
Scalar cv::cuda::sum(InputArray, InputArray, GpuMat&) { throw_no_cuda(); return Scalar(); }
Scalar cv::cuda::absSum(InputArray, InputArray, GpuMat&) { throw_no_cuda(); return Scalar(); }
Scalar cv::cuda::sqrSum(InputArray, InputArray, GpuMat&) { throw_no_cuda(); return Scalar(); }
Scalar cv::cuda::sum(InputArray, InputArray) { throw_no_cuda(); return Scalar(); }
void cv::cuda::calcSum(InputArray, OutputArray, InputArray, Stream&) { throw_no_cuda(); }
Scalar cv::cuda::absSum(InputArray, InputArray) { throw_no_cuda(); return Scalar(); }
void cv::cuda::calcAbsSum(InputArray, OutputArray, InputArray, Stream&) { throw_no_cuda(); }
Scalar cv::cuda::sqrSum(InputArray, InputArray) { throw_no_cuda(); return Scalar(); }
void cv::cuda::calcSqrSum(InputArray, OutputArray, InputArray, Stream&) { throw_no_cuda(); }
void cv::cuda::minMax(InputArray, double*, double*, InputArray, GpuMat&) { throw_no_cuda(); }
void cv::cuda::minMaxLoc(InputArray, double*, double*, Point*, Point*, InputArray, GpuMat&, GpuMat&) { throw_no_cuda(); }
void cv::cuda::minMax(InputArray, double*, double*, InputArray) { throw_no_cuda(); }
void cv::cuda::findMinMax(InputArray, OutputArray, InputArray, Stream&) { throw_no_cuda(); }
void cv::cuda::minMaxLoc(InputArray, double*, double*, Point*, Point*, InputArray) { throw_no_cuda(); }
void cv::cuda::findMinMaxLoc(InputArray, OutputArray, OutputArray, InputArray, Stream&) { throw_no_cuda(); }
int cv::cuda::countNonZero(InputArray, GpuMat&) { throw_no_cuda(); return 0; }
int cv::cuda::countNonZero(InputArray) { throw_no_cuda(); return 0; }
void cv::cuda::countNonZero(InputArray, OutputArray, Stream&) { throw_no_cuda(); }
void cv::cuda::reduce(InputArray, OutputArray, int, int, int, Stream&) { throw_no_cuda(); }
void cv::cuda::meanStdDev(InputArray, Scalar&, Scalar&, GpuMat&) { throw_no_cuda(); }
void cv::cuda::meanStdDev(InputArray, Scalar&, Scalar&) { throw_no_cuda(); }
void cv::cuda::meanStdDev(InputArray, OutputArray, Stream&) { throw_no_cuda(); }
void cv::cuda::rectStdDev(InputArray, InputArray, OutputArray, Rect, Stream&) { throw_no_cuda(); }
void cv::cuda::normalize(InputArray, OutputArray, double, double, int, int, InputArray, GpuMat&, GpuMat&) { throw_no_cuda(); }
void cv::cuda::normalize(InputArray, OutputArray, double, double, int, int, InputArray, Stream&) { throw_no_cuda(); }
void cv::cuda::integral(InputArray, OutputArray, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::cuda::sqrIntegral(InputArray, OutputArray, GpuMat&, Stream&) { throw_no_cuda(); }
void cv::cuda::integral(InputArray, OutputArray, Stream&) { throw_no_cuda(); }
void cv::cuda::sqrIntegral(InputArray, OutputArray, Stream&) { throw_no_cuda(); }
#else
namespace
{
class DeviceBuffer
{
public:
explicit DeviceBuffer(int count_ = 1) : count(count_)
{
cudaSafeCall( cudaMalloc(&pdev, count * sizeof(double)) );
}
~DeviceBuffer()
{
cudaSafeCall( cudaFree(pdev) );
}
operator double*() {return pdev;}
void download(double* hptr)
{
double hbuf;
cudaSafeCall( cudaMemcpy(&hbuf, pdev, sizeof(double), cudaMemcpyDeviceToHost) );
*hptr = hbuf;
}
void download(double** hptrs)
{
AutoBuffer<double, 2 * sizeof(double)> hbuf(count);
cudaSafeCall( cudaMemcpy((void*)hbuf, pdev, count * sizeof(double), cudaMemcpyDeviceToHost) );
for (int i = 0; i < count; ++i)
*hptrs[i] = hbuf[i];
}
private:
double* pdev;
int count;
};
}
////////////////////////////////////////////////////////////////////////
// norm
double cv::cuda::norm(InputArray _src, int normType, InputArray _mask, GpuMat& buf)
{
GpuMat src = _src.getGpuMat();
GpuMat mask = _mask.getGpuMat();
namespace cv { namespace cuda { namespace internal {
void normL2(cv::InputArray _src, cv::OutputArray _dst, cv::InputArray _mask, Stream& stream);
void findMaxAbs(cv::InputArray _src, cv::OutputArray _dst, cv::InputArray _mask, Stream& stream);
}}}
void cv::cuda::calcNorm(InputArray _src, OutputArray dst, int normType, InputArray mask, Stream& stream)
{
CV_Assert( normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 );
CV_Assert( mask.empty() || (mask.type() == CV_8UC1 && mask.size() == src.size() && src.channels() == 1) );
GpuMat src = getInputMat(_src, stream);
GpuMat src_single_channel = src.reshape(1);
if (normType == NORM_L1)
return cuda::absSum(src_single_channel, mask, buf)[0];
{
calcAbsSum(src_single_channel, dst, mask, stream);
}
else if (normType == NORM_L2)
{
internal::normL2(src_single_channel, dst, mask, stream);
}
else // NORM_INF
{
internal::findMaxAbs(src_single_channel, dst, mask, stream);
}
}
if (normType == NORM_L2)
return std::sqrt(cuda::sqrSum(src_single_channel, mask, buf)[0]);
double cv::cuda::norm(InputArray _src, int normType, InputArray _mask)
{
Stream& stream = Stream::Null();
// NORM_INF
double min_val, max_val;
cuda::minMax(src_single_channel, &min_val, &max_val, mask, buf);
return std::max(std::abs(min_val), std::abs(max_val));
HostMem dst;
calcNorm(_src, dst, normType, _mask, stream);
stream.waitForCompletion();
double val;
dst.createMatHeader().convertTo(Mat(1, 1, CV_64FC1, &val), CV_64F);
return val;
}
////////////////////////////////////////////////////////////////////////
// meanStdDev
void cv::cuda::meanStdDev(InputArray _src, Scalar& mean, Scalar& stddev, GpuMat& buf)
void cv::cuda::meanStdDev(InputArray _src, OutputArray _dst, Stream& stream)
{
GpuMat src = _src.getGpuMat();
if (!deviceSupports(FEATURE_SET_COMPUTE_13))
CV_Error(cv::Error::StsNotImplemented, "Not sufficient compute capebility");
const GpuMat src = getInputMat(_src, stream);
CV_Assert( src.type() == CV_8UC1 );
if (!deviceSupports(FEATURE_SET_COMPUTE_13))
CV_Error(cv::Error::StsNotImplemented, "Not sufficient compute capebility");
GpuMat dst = getOutputMat(_dst, 1, 2, CV_64FC1, stream);
NppiSize sz;
sz.width = src.cols;
sz.height = src.rows;
DeviceBuffer dbuf(2);
int bufSize;
#if (CUDA_VERSION <= 4020)
nppSafeCall( nppiMeanStdDev8uC1RGetBufferHostSize(sz, &bufSize) );
@@ -158,14 +154,30 @@ void cv::cuda::meanStdDev(InputArray _src, Scalar& mean, Scalar& stddev, GpuMat&
nppSafeCall( nppiMeanStdDevGetBufferHostSize_8u_C1R(sz, &bufSize) );
#endif
ensureSizeIsEnough(1, bufSize, CV_8UC1, buf);
BufferPool pool(stream);
GpuMat buf = pool.getBuffer(1, bufSize, CV_8UC1);
nppSafeCall( nppiMean_StdDev_8u_C1R(src.ptr<Npp8u>(), static_cast<int>(src.step), sz, buf.ptr<Npp8u>(), dbuf, (double*)dbuf + 1) );
NppStreamHandler h(StreamAccessor::getStream(stream));
cudaSafeCall( cudaDeviceSynchronize() );
nppSafeCall( nppiMean_StdDev_8u_C1R(src.ptr<Npp8u>(), static_cast<int>(src.step), sz, buf.ptr<Npp8u>(), dst.ptr<Npp64f>(), dst.ptr<Npp64f>() + 1) );
double* ptrs[2] = {mean.val, stddev.val};
dbuf.download(ptrs);
syncOutput(dst, _dst, stream);
}
void cv::cuda::meanStdDev(InputArray _src, Scalar& mean, Scalar& stddev)
{
Stream& stream = Stream::Null();
HostMem dst;
meanStdDev(_src, dst, stream);
stream.waitForCompletion();
double vals[2];
dst.createMatHeader().copyTo(Mat(1, 2, CV_64FC1, &vals[0]));
mean = Scalar(vals[0]);
stddev = Scalar(vals[1]);
}
//////////////////////////////////////////////////////////////////////////////
@@ -173,13 +185,12 @@ void cv::cuda::meanStdDev(InputArray _src, Scalar& mean, Scalar& stddev, GpuMat&
void cv::cuda::rectStdDev(InputArray _src, InputArray _sqr, OutputArray _dst, Rect rect, Stream& _stream)
{
GpuMat src = _src.getGpuMat();
GpuMat sqr = _sqr.getGpuMat();
GpuMat src = getInputMat(_src, _stream);
GpuMat sqr = getInputMat(_sqr, _stream);
CV_Assert( src.type() == CV_32SC1 && sqr.type() == CV_64FC1 );
_dst.create(src.size(), CV_32FC1);
GpuMat dst = _dst.getGpuMat();
GpuMat dst = getOutputMat(_dst, src.size(), CV_32FC1, _stream);
NppiSize sz;
sz.width = src.cols;
@@ -200,45 +211,8 @@ void cv::cuda::rectStdDev(InputArray _src, InputArray _sqr, OutputArray _dst, Re
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
////////////////////////////////////////////////////////////////////////
// normalize
void cv::cuda::normalize(InputArray _src, OutputArray dst, double a, double b, int norm_type, int dtype, InputArray mask, GpuMat& norm_buf, GpuMat& cvt_buf)
{
GpuMat src = _src.getGpuMat();
double scale = 1, shift = 0;
if (norm_type == NORM_MINMAX)
{
double smin = 0, smax = 0;
double dmin = std::min(a, b), dmax = std::max(a, b);
cuda::minMax(src, &smin, &smax, mask, norm_buf);
scale = (dmax - dmin) * (smax - smin > std::numeric_limits<double>::epsilon() ? 1.0 / (smax - smin) : 0.0);
shift = dmin - smin * scale;
}
else if (norm_type == NORM_L2 || norm_type == NORM_L1 || norm_type == NORM_INF)
{
scale = cuda::norm(src, norm_type, mask, norm_buf);
scale = scale > std::numeric_limits<double>::epsilon() ? a / scale : 0.0;
shift = 0;
}
else
{
CV_Error(cv::Error::StsBadArg, "Unknown/unsupported norm type");
}
if (mask.empty())
{
src.convertTo(dst, dtype, scale, shift);
}
else
{
src.convertTo(cvt_buf, dtype, scale, shift);
cvt_buf.copyTo(dst, mask);
}
syncOutput(dst, _dst, _stream);
}
#endif