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

naive implementation

Browse Source
This commit is contained in:
Elena Gvozdeva
2014-08-22 15:05:29 +04:00
committed by ElenaGvozdeva
parent dee56598e9
commit d78bc3c321
2 changed files with 140 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

82
modules/core/src/matmul.cpp
View File

@@ -693,7 +693,7 @@ static void GEMMStore_64fc( const Complexd* c_data, size_t c_step,
#ifdef HAVE_CLAMDBLAS
static bool ocl_gemm( InputArray matA, InputArray matB, double alpha,
static bool ocl_gemm_amdblas( InputArray matA, InputArray matB, double alpha,
InputArray matC, double beta, OutputArray matD, int flags )
{
int type = matA.type(), esz = CV_ELEM_SIZE(type);
@@ -775,6 +775,79 @@ static bool ocl_gemm( InputArray matA, InputArray matB, double alpha,
#endif
#ifdef HAVE_OPENCL
static bool ocl_gemm( InputArray matA, InputArray matB, double alpha,
InputArray matC, double beta, OutputArray matD, int flags )
{
int depth = matA.depth(), cn = matA.channels();
int type = CV_MAKETYPE(depth, cn);
CV_Assert( type == matB.type() && (type == CV_32FC1 || type == CV_64FC1 || type == CV_32FC2 || type == CV_64FC2) );
const ocl::Device & dev = ocl::Device::getDefault();
bool doubleSupport = dev.doubleFPConfig() > 0;
if ((!doubleSupport && depth == CV_64F))
return false;
bool haveC = matC.kind() != cv::_InputArray::NONE;
Size sizeA = matA.size(), sizeB = matB.size(), sizeC = haveC ? matC.size() : Size(0, 0);
bool atrans = (flags & GEMM_1_T) != 0, btrans = (flags & GEMM_2_T) != 0, ctrans = (flags & GEMM_3_T) != 0;
if (atrans)
sizeA = Size(sizeA.height, sizeA.width);
if (btrans)
sizeB = Size(sizeB.height, sizeB.width);
if (haveC && ctrans)
sizeC = Size(sizeC.height, sizeC.width);
Size sizeD(sizeB.width, sizeA.height);
CV_Assert( matB.type() == type && (!haveC || matC.type() == type) );
CV_Assert( sizeA.width == sizeB.height && (!haveC || sizeC == sizeD) );
String opts = format("-D T=%s -D T1=%s -D cn=%d %s %s",
ocl::typeToStr(type), ocl::typeToStr(depth), cn,
haveC ? "-D HAVE_C" : "",
doubleSupport ? " -D DOUBLE_SUPPORT" : "");
ocl::Kernel k("gemm", cv::ocl::core::gemm_oclsrc, opts);
if (k.empty())
return false;
matD.create(sizeD, type);
UMat A = matA.getUMat(), B = matB.getUMat(), D = matD.getUMat();
if (atrans)
A = A.t();
if (btrans)
B = B.t();
if (haveC)
ctrans ? transpose(matC, D) : matC.copyTo(D);
else
D.setTo(Scalar::all(0));
if (depth == CV_64F)
k.args(ocl::KernelArg::ReadOnlyNoSize(A),
ocl::KernelArg::ReadOnlyNoSize(B),
ocl::KernelArg::ReadWrite(D),
sizeA.width, alpha, beta);
else
k.args(ocl::KernelArg::ReadOnlyNoSize(A),
ocl::KernelArg::ReadOnlyNoSize(B),
ocl::KernelArg::ReadWrite(D),
sizeA.width, (float)alpha, (float)beta);
size_t globalsize[2] = { sizeD.width, sizeD.height};
return k.run(2, globalsize, NULL, false);
}
#endif
}
void cv::gemm( InputArray matA, InputArray matB, double alpha,
@@ -783,7 +856,12 @@ void cv::gemm( InputArray matA, InputArray matB, double alpha,
#ifdef HAVE_CLAMDBLAS
CV_OCL_RUN(ocl::haveAmdBlas() && matA.dims() <= 2 && matB.dims() <= 2 && matC.dims() <= 2 && _matD.isUMat() &&
matA.cols() > 20 && matA.rows() > 20 && matB.cols() > 20, // since it works incorrect for small sizes
ocl_gemm(matA, matB, alpha, matC, beta, _matD, flags))
ocl_gemm_amdblas(matA, matB, alpha, matC, beta, _matD, flags))
#endif
#ifdef HAVE_OPENCL
CV_OCL_RUN(_matD.isUMat() && matA.dims() <= 2 && matB.dims() <= 2 && matC.dims() <= 2,
ocl_gemm(matA, matB, alpha, matC, beta, _matD, flags))
#endif
const int block_lin_size = 128;