mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-06-13 04:52:53 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Some small fixes

Browse Source
This commit is contained in:
ElenaGvozdeva 2014-10-16 10:24:44 +03:00
parent c5a2879ce0
commit 65b8a1cb37
2 changed files with 11 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
modules/core/src/matmul.cpp
View File

@ -788,7 +788,7 @@ static bool ocl_gemm( InputArray matA, InputArray matB, double alpha,
const ocl::Device & dev = ocl::Device::getDefault(); const ocl::Device & dev = ocl::Device::getDefault();
bool doubleSupport = dev.doubleFPConfig() > 0; bool doubleSupport = dev.doubleFPConfig() > 0;
if ((!doubleSupport && depth == CV_64F)) if (!doubleSupport && depth == CV_64F)
return false; return false;
bool haveC = matC.kind() != cv::_InputArray::NONE; bool haveC = matC.kind() != cv::_InputArray::NONE;
@ -804,7 +804,7 @@ static bool ocl_gemm( InputArray matA, InputArray matB, double alpha,
Size sizeD(sizeB.width, sizeA.height); Size sizeD(sizeB.width, sizeA.height);
CV_Assert( matB.type() == type && (!haveC || matC.type() == type) ); CV_Assert( !haveC || matC.type() == type );
CV_Assert( sizeA.width == sizeB.height && (!haveC || sizeC == sizeD) ); CV_Assert( sizeA.width == sizeB.height && (!haveC || sizeC == sizeD) );
int max_wg_size = (int)dev.maxWorkGroupSize(); int max_wg_size = (int)dev.maxWorkGroupSize();
@ -822,11 +822,8 @@ static bool ocl_gemm( InputArray matA, InputArray matB, double alpha,
if (haveC) if (haveC)
ctrans ? transpose(matC, D) : matC.copyTo(D); ctrans ? transpose(matC, D) : matC.copyTo(D);
else
D.setTo(Scalar::all(0));
int vectorWidths[] = { 4, 4, 2, 2, 1, 4, cn, -1 }; int vectorWidths[] = { 4, 4, 2, 2, 1, 4, cn, -1 };
int kercn = ocl::checkOptimalVectorWidth(vectorWidths, B, D); int kercn = ocl::checkOptimalVectorWidth(vectorWidths, B, D);
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", 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",

18
modules/core/src/opencl/gemm.cl
View File

@ -22,13 +22,13 @@
#if cn==2 #if cn==2
#if kercn==2 #if kercn==2
#define MUL(i, a, b)\ #define MUL(a, b)\
{\ {\
sum.x += fma(a.x, b.x, - a.y * b.y);\ sum.x += fma(a.x, b.x, - a.y * b.y);\
sum.y += fma(a.x, b.y, a.y * b.x);\ sum.y += fma(a.x, b.y, a.y * b.x);\
} }
#else #else
#define MUL(i, a, b)\ #define MUL(a, b)\
{\ {\
sum.x += fma(a.x, b.x, - a.y * b.y);\ sum.x += fma(a.x, b.x, - a.y * b.y);\
sum.y += fma(a.x, b.y, a.y * b.x);\ sum.y += fma(a.x, b.y, a.y * b.x);\
@ -37,7 +37,7 @@
} }
#endif #endif
#else #else
#define MUL(i, a, b) sum = fma(a, b, sum); #define MUL(a, b) sum = fma(a, b, sum);
#endif #endif
@ -62,7 +62,7 @@ __kernel void gemm(__global const uchar * A_ptr, int A_step, int A_offset,
if (x < D_cols && y < D_rows) if (x < D_cols && y < D_rows)
{ {
for (int i = 0; i < n; ++i) for (int i = 0; i < n; ++i)
MUL(i, A[i], B[i*STEP_B]); MUL(A[i], B[i*STEP_B]);
#else #else
__local T a_local[LOCAL_SIZE*LOCAL_SIZE]; __local T a_local[LOCAL_SIZE*LOCAL_SIZE];
@ -86,14 +86,14 @@ __kernel void gemm(__global const uchar * A_ptr, int A_step, int A_offset,
if (x < D_cols && y < D_rows) if (x < D_cols && y < D_rows)
{ {
for (int i = 0; i < LOCAL_SIZE
#if NO_MULT #if NO_MULT
&& p * LOCAL_SIZE + i < n int ie = min(LOCAL_SIZE, n - p * LOCAL_SIZE);
for (int i = 0; i < ie; ++i)
#else
for (int i = 0; i < LOCAL_SIZE; ++i)
#endif #endif
; ++i) MUL(a_local[mad24(lidy, LOCAL_SIZE, i)], b_local[mad24(i, LOCAL_SIZE, lidx)]);
MUL(i, a_local[mad24(lidy, LOCAL_SIZE, i)], b_local[mad24(i, LOCAL_SIZE, lidx)]);
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
} }