potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #8104 from insoow:master

Browse Source 
Gemm kernels for Intel GPU (#8104)

* Fix an issue with Kernel object reset release when consecutive Kernel::run calls

Kernel::run launch OCL gpu kernels and set a event callback function
to decreate the ref count of UMat or remove UMat when the lauched workloads
are completed. However, for some OCL kernels requires multiple call of
Kernel::run function with some kernel parameter changes (e.g., input
and output buffer offset) to get the final computation result.
In the case, the current implementation requires unnecessary
synchronization and cleanupMat.

This fix requires the user to specify whether there will be more work or not.
If there is no remaining computation, the Kernel::run will reset the
kernel object

Signed-off-by: Woo, Insoo <insoo.woo@intel.com>

* GEMM kernel optimization for Intel GEN

The optimized kernels uses cl_intel_subgroups extension for better
performance.

Note: This optimized kernels will be part of ISAAC in a code generation
way under MIT license.

Signed-off-by: Woo, Insoo <insoo.woo@intel.com>

* Fix API compatibility error

This patch fixes a OCV API compatibility error. The error was reported
due to the interface changes of Kernel::run. To resolve the issue,
An overloaded function of Kernel::run is added. It take a flag indicating
whether there are more work to be done with the kernel object without
releasing resources related to it.

Signed-off-by: Woo, Insoo <insoo.woo@intel.com>

* Renaming intel_gpu_gemm.cpp to intel_gpu_gemm.inl.hpp

Signed-off-by: Woo, Insoo <insoo.woo@intel.com>

* Revert "Fix API compatibility error"

This reverts commit 2ef427db91.

Conflicts:
	modules/core/src/intel_gpu_gemm.inl.hpp

* Revert "Fix an issue with Kernel object reset release when consecutive Kernel::run calls"

This reverts commit cc7f9f5469.

* Fix the case of uninitialization D

When C is null and beta is non-zero, D is used without initialization.
This resloves the issue

Signed-off-by: Woo, Insoo <insoo.woo@intel.com>

* fix potential output error due to 0 * nan

Signed-off-by: Woo, Insoo <insoo.woo@intel.com>

* whitespace fix, eliminate non-ASCII symbols

* fix build warning
This commit is contained in:
insoow
2017-04-19 02:57:54 -07:00
committed by Alexander Alekhin
parent cea0e94376
commit 2922738b6d
5 changed files with 1315 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/matmul.cpp
+66 -38
View File
@@ -41,9 +41,12 @@
//
//M*/
#include <sstream>
#include "precomp.hpp"
#include "opencl_kernels_core.hpp"
#include "opencv2/core/opencl/runtime/opencl_clamdblas.hpp"
#include "opencv2/core/opencl/runtime/opencl_core.hpp"
#include "intel_gpu_gemm.inl.hpp"
namespace cv
{
@@ -787,7 +790,6 @@ static bool ocl_gemm_amdblas( 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 )
{
@@ -806,62 +808,88 @@ static bool ocl_gemm( InputArray matA, InputArray matB, double alpha,
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( !haveC || matC.type() == type );
CV_Assert( sizeA.width == sizeB.height && (!haveC || sizeC == sizeD) );
int max_wg_size = (int)dev.maxWorkGroupSize();
int block_size = (max_wg_size / (32*cn) < 32) ? (max_wg_size / (16*cn) < 16) ? (max_wg_size / (8*cn) < 8) ? 1 : 8 : 16 : 32;
Size sizeD(((btrans)? sizeB.height : sizeB.width),
((atrans)? sizeA.width : sizeA.height));
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 (!dev.intelSubgroupsSupport() || (depth == CV_64F) || cn != 1)
{
String opts;
if (haveC)
ctrans ? transpose(matC, D) : matC.copyTo(D);
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);
int vectorWidths[] = { 4, 4, 2, 2, 1, 4, cn, -1 };
int kercn = ocl::checkOptimalVectorWidth(vectorWidths, B, D);
CV_Assert( sizeA.width == sizeB.height && (!haveC || sizeC == sizeD) );
String opts = format("-D T=%s -D T1=%s -D WT=%s -D cn=%d -D kercn=%d -D LOCAL_SIZE=%d %s %s %s",
int max_wg_size = (int)dev.maxWorkGroupSize();
int block_size = (max_wg_size / (32*cn) < 32) ? (max_wg_size / (16*cn) < 16) ? (max_wg_size / (8*cn) < 8) ? 1 : 8 : 16 : 32;
if (atrans)
A = A.t();
if (btrans)
B = B.t();
if (haveC)
ctrans ? transpose(matC, D) : matC.copyTo(D);
int vectorWidths[] = { 4, 4, 2, 2, 1, 4, cn, -1 };
int kercn = ocl::checkOptimalVectorWidth(vectorWidths, B, D);
opts += format(" -D T=%s -D T1=%s -D WT=%s -D cn=%d -D kercn=%d -D LOCAL_SIZE=%d %s %s %s",
ocl::typeToStr(type), ocl::typeToStr(depth), ocl::typeToStr(CV_MAKETYPE(depth, kercn)),
cn, kercn, block_size,
(sizeA.width % block_size !=0) ? "-D NO_MULT" : "",
haveC ? "-D HAVE_C" : "",
doubleSupport ? " -D DOUBLE_SUPPORT" : "");
ocl::Kernel k("gemm", cv::ocl::core::gemm_oclsrc, opts);
if (k.empty())
return false;
ocl::Kernel k("gemm", cv::ocl::core::gemm_oclsrc, opts);
if (k.empty())
return false;
if (depth == CV_64F)
k.args(ocl::KernelArg::ReadOnlyNoSize(A),
ocl::KernelArg::ReadOnlyNoSize(B, cn, kercn),
ocl::KernelArg::ReadWrite(D, cn, kercn),
sizeA.width, alpha, beta);
if (depth == CV_64F)
k.args(ocl::KernelArg::ReadOnlyNoSize(A),
ocl::KernelArg::ReadOnlyNoSize(B, cn, kercn),
ocl::KernelArg::ReadWrite(D, cn, kercn),
sizeA.width, alpha, beta);
else
k.args(ocl::KernelArg::ReadOnlyNoSize(A),
ocl::KernelArg::ReadOnlyNoSize(B, cn, kercn),
ocl::KernelArg::ReadWrite(D, cn, kercn),
sizeA.width, (float)alpha, (float)beta);
size_t globalsize[2] = { (size_t)sizeD.width * cn / kercn, (size_t)sizeD.height};
size_t localsize[2] = { (size_t)block_size, (size_t)block_size};
return k.run(2, globalsize, block_size!=1 ? localsize : NULL, false);
}
else
k.args(ocl::KernelArg::ReadOnlyNoSize(A),
ocl::KernelArg::ReadOnlyNoSize(B, cn, kercn),
ocl::KernelArg::ReadWrite(D, cn, kercn),
sizeA.width, (float)alpha, (float)beta);
{
if (haveC && beta != 0.0)
{
ctrans ? transpose(matC, D) : matC.copyTo(D);
}
else
{
beta = 0.0;
}
size_t globalsize[2] = { (size_t)sizeD.width * cn / kercn, (size_t)sizeD.height};
size_t localsize[2] = { (size_t)block_size, (size_t)block_size};
return k.run(2, globalsize, block_size!=1 ? localsize : NULL, false);
return intel_gpu_gemm(A, sizeA,
B, sizeB,
D, sizeD,
alpha,
beta,
atrans, btrans);
}
}
#endif