third_party/opencv
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

cmake: AVX512 -> AVX_512F

Browse Source
This commit is contained in:
Alexander Alekhin
2017-12-28 02:16:42 +00:00
parent 2938860b3f
commit 898ca38257
7 changed files with 34 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/CMakeLists.txt
+1 -1
View File
@@ -13,7 +13,7 @@ endif()
set(the_description "Deep neural network module. It allows to load models from different frameworks and to make forward pass")
ocv_add_dispatched_file("layers/layers_common" AVX AVX2 AVX512)
ocv_add_dispatched_file("layers/layers_common" AVX AVX2 AVX_512F)
ocv_add_module(dnn opencv_core opencv_imgproc WRAP python matlab java js)
ocv_warnings_disable(CMAKE_CXX_FLAGS -Wno-shadow -Wno-parentheses -Wmaybe-uninitialized -Wsign-promo
modules/dnn/src/layers/convolution_layer.cpp
+6 -6
View File
@@ -384,7 +384,7 @@ public:
p.is1x1_ = kernel == Size(0,0) && pad == Size(0, 0);
p.useAVX = checkHardwareSupport(CPU_AVX);
p.useAVX2 = checkHardwareSupport(CPU_AVX2);
p.useAVX512 = checkHardwareSupport(CPU_AVX_512DQ);
p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX_512F;
int ncn = std::min(inpCn, (int)BLK_SIZE_CN);
p.ofstab_.resize(kernel.width*kernel.height*ncn);
@@ -564,10 +564,10 @@ public:
// now compute dot product of the weights
// and im2row-transformed part of the tensor
int bsz = ofs1 - ofs0;
#if CV_TRY_AVX512
#if CV_TRY_AVX_512F
/* AVX512 convolution requires an alignment of 16, and ROI is only there for larger vector sizes */
if(useAVX512)
opt_AVX512::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
opt_AVX_512F::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
else
#endif
@@ -1102,7 +1102,7 @@ public:
nstripes_ = nstripes;
useAVX = checkHardwareSupport(CPU_AVX);
useAVX2 = checkHardwareSupport(CPU_AVX2);
useAVX512 = checkHardwareSupport(CPU_AVX_512DQ);
useAVX512 = CV_CPU_HAS_SUPPORT_AVX_512F;
}
void operator()(const Range& range_) const
@@ -1120,9 +1120,9 @@ public:
size_t bstep = b_->step1();
size_t cstep = c_->step1();
#if CV_TRY_AVX512
#if CV_TRY_AVX_512F
if( useAVX512 )
opt_AVX512::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
opt_AVX_512F::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
else
#endif
#if CV_TRY_AVX2
modules/dnn/src/layers/fully_connected_layer.cpp
+3 -3
View File
@@ -161,7 +161,7 @@ public:
p.activ = activ;
p.useAVX = checkHardwareSupport(CPU_AVX);
p.useAVX2 = checkHardwareSupport(CPU_AVX2);
p.useAVX512 = checkHardwareSupport(CPU_AVX_512DQ);
p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX_512F;
parallel_for_(Range(0, nstripes), p, nstripes);
}
@@ -196,9 +196,9 @@ public:
memcpy(sptr, sptr_, vecsize*sizeof(sptr[0]));
#if CV_TRY_AVX512
#if CV_TRY_AVX_512F
if( useAVX512 )
opt_AVX512::fastGEMM1T( sptr, wptr, wstep, biasptr, dptr, nw, vecsize);
opt_AVX_512F::fastGEMM1T( sptr, wptr, wstep, biasptr, dptr, nw, vecsize);
else
#endif
#if CV_TRY_AVX2
modules/dnn/src/layers/layers_common.simd.hpp
+1 -1
View File
@@ -301,7 +301,7 @@ void fastGEMM( const float* aptr, size_t astep, const float* bptr,
{
int n = 0;
#ifdef CV_AVX512
#if CV_AVX_512F
for( ; n <= nb - 32; n += 32 )
{
for( int m = 0; m < ma; m += 4 )