Merge pull request opencv#9007 from alalek:issue_9001

mshabunin · mshabunin · commit 1c4fb418cfa3 · 2017-06-27T14:38:43.000Z
diff --git a/cmake/OpenCVCompilerOptimizations.cmake b/cmake/OpenCVCompilerOptimizations.cmake
@@ -647,12 +647,15 @@ macro(ocv_compiler_optimization_fill_cpu_config)
     if(NOT DEFINED CPU_${OPT}_FEATURE_ALIAS OR NOT "x${CPU_${OPT}_FEATURE_ALIAS}" STREQUAL "x")
       set(OPENCV_CPU_CONTROL_DEFINITIONS_CONFIGMAKE "${OPENCV_CPU_CONTROL_DEFINITIONS_CONFIGMAKE}
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_${OPT}
+#  define CV_TRY_${OPT} 1
 #  define CV_CPU_HAS_SUPPORT_${OPT} 1
 #  define CV_CPU_CALL_${OPT}(fn, args) return (opt_${OPT}::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_${OPT}
+#  define CV_TRY_${OPT} 1
 #  define CV_CPU_HAS_SUPPORT_${OPT} (cv::checkHardwareSupport(CV_CPU_${OPT}))
 #  define CV_CPU_CALL_${OPT}(fn, args) if (CV_CPU_HAS_SUPPORT_${OPT}) return (opt_${OPT}::fn args)
 #else
+#  define CV_TRY_${OPT} 0
 #  define CV_CPU_HAS_SUPPORT_${OPT} 0
 #  define CV_CPU_CALL_${OPT}(fn, args)
 #endif
diff --git a/modules/core/include/opencv2/core/cv_cpu_helper.h b/modules/core/include/opencv2/core/cv_cpu_helper.h
@@ -1,144 +1,180 @@
 // AUTOGENERATED, DO NOT EDIT
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE
+#  define CV_TRY_SSE 1
 #  define CV_CPU_HAS_SUPPORT_SSE 1
 #  define CV_CPU_CALL_SSE(fn, args) return (opt_SSE::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE
+#  define CV_TRY_SSE 1
 #  define CV_CPU_HAS_SUPPORT_SSE (cv::checkHardwareSupport(CV_CPU_SSE))
 #  define CV_CPU_CALL_SSE(fn, args) if (CV_CPU_HAS_SUPPORT_SSE) return (opt_SSE::fn args)
 #else
+#  define CV_TRY_SSE 0
 #  define CV_CPU_HAS_SUPPORT_SSE 0
 #  define CV_CPU_CALL_SSE(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_SSE(fn, args, mode, ...)  CV_CPU_CALL_SSE(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE2
+#  define CV_TRY_SSE2 1
 #  define CV_CPU_HAS_SUPPORT_SSE2 1
 #  define CV_CPU_CALL_SSE2(fn, args) return (opt_SSE2::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE2
+#  define CV_TRY_SSE2 1
 #  define CV_CPU_HAS_SUPPORT_SSE2 (cv::checkHardwareSupport(CV_CPU_SSE2))
 #  define CV_CPU_CALL_SSE2(fn, args) if (CV_CPU_HAS_SUPPORT_SSE2) return (opt_SSE2::fn args)
 #else
+#  define CV_TRY_SSE2 0
 #  define CV_CPU_HAS_SUPPORT_SSE2 0
 #  define CV_CPU_CALL_SSE2(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_SSE2(fn, args, mode, ...)  CV_CPU_CALL_SSE2(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE3
+#  define CV_TRY_SSE3 1
 #  define CV_CPU_HAS_SUPPORT_SSE3 1
 #  define CV_CPU_CALL_SSE3(fn, args) return (opt_SSE3::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE3
+#  define CV_TRY_SSE3 1
 #  define CV_CPU_HAS_SUPPORT_SSE3 (cv::checkHardwareSupport(CV_CPU_SSE3))
 #  define CV_CPU_CALL_SSE3(fn, args) if (CV_CPU_HAS_SUPPORT_SSE3) return (opt_SSE3::fn args)
 #else
+#  define CV_TRY_SSE3 0
 #  define CV_CPU_HAS_SUPPORT_SSE3 0
 #  define CV_CPU_CALL_SSE3(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_SSE3(fn, args, mode, ...)  CV_CPU_CALL_SSE3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSSE3
+#  define CV_TRY_SSSE3 1
 #  define CV_CPU_HAS_SUPPORT_SSSE3 1
 #  define CV_CPU_CALL_SSSE3(fn, args) return (opt_SSSE3::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSSE3
+#  define CV_TRY_SSSE3 1
 #  define CV_CPU_HAS_SUPPORT_SSSE3 (cv::checkHardwareSupport(CV_CPU_SSSE3))
 #  define CV_CPU_CALL_SSSE3(fn, args) if (CV_CPU_HAS_SUPPORT_SSSE3) return (opt_SSSE3::fn args)
 #else
+#  define CV_TRY_SSSE3 0
 #  define CV_CPU_HAS_SUPPORT_SSSE3 0
 #  define CV_CPU_CALL_SSSE3(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_SSSE3(fn, args, mode, ...)  CV_CPU_CALL_SSSE3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE4_1
+#  define CV_TRY_SSE4_1 1
 #  define CV_CPU_HAS_SUPPORT_SSE4_1 1
 #  define CV_CPU_CALL_SSE4_1(fn, args) return (opt_SSE4_1::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE4_1
+#  define CV_TRY_SSE4_1 1
 #  define CV_CPU_HAS_SUPPORT_SSE4_1 (cv::checkHardwareSupport(CV_CPU_SSE4_1))
 #  define CV_CPU_CALL_SSE4_1(fn, args) if (CV_CPU_HAS_SUPPORT_SSE4_1) return (opt_SSE4_1::fn args)
 #else
+#  define CV_TRY_SSE4_1 0
 #  define CV_CPU_HAS_SUPPORT_SSE4_1 0
 #  define CV_CPU_CALL_SSE4_1(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_SSE4_1(fn, args, mode, ...)  CV_CPU_CALL_SSE4_1(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE4_2
+#  define CV_TRY_SSE4_2 1
 #  define CV_CPU_HAS_SUPPORT_SSE4_2 1
 #  define CV_CPU_CALL_SSE4_2(fn, args) return (opt_SSE4_2::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE4_2
+#  define CV_TRY_SSE4_2 1
 #  define CV_CPU_HAS_SUPPORT_SSE4_2 (cv::checkHardwareSupport(CV_CPU_SSE4_2))
 #  define CV_CPU_CALL_SSE4_2(fn, args) if (CV_CPU_HAS_SUPPORT_SSE4_2) return (opt_SSE4_2::fn args)
 #else
+#  define CV_TRY_SSE4_2 0
 #  define CV_CPU_HAS_SUPPORT_SSE4_2 0
 #  define CV_CPU_CALL_SSE4_2(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_SSE4_2(fn, args, mode, ...)  CV_CPU_CALL_SSE4_2(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_POPCNT
+#  define CV_TRY_POPCNT 1
 #  define CV_CPU_HAS_SUPPORT_POPCNT 1
 #  define CV_CPU_CALL_POPCNT(fn, args) return (opt_POPCNT::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_POPCNT
+#  define CV_TRY_POPCNT 1
 #  define CV_CPU_HAS_SUPPORT_POPCNT (cv::checkHardwareSupport(CV_CPU_POPCNT))
 #  define CV_CPU_CALL_POPCNT(fn, args) if (CV_CPU_HAS_SUPPORT_POPCNT) return (opt_POPCNT::fn args)
 #else
+#  define CV_TRY_POPCNT 0
 #  define CV_CPU_HAS_SUPPORT_POPCNT 0
 #  define CV_CPU_CALL_POPCNT(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_POPCNT(fn, args, mode, ...)  CV_CPU_CALL_POPCNT(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX
+#  define CV_TRY_AVX 1
 #  define CV_CPU_HAS_SUPPORT_AVX 1
 #  define CV_CPU_CALL_AVX(fn, args) return (opt_AVX::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX
+#  define CV_TRY_AVX 1
 #  define CV_CPU_HAS_SUPPORT_AVX (cv::checkHardwareSupport(CV_CPU_AVX))
 #  define CV_CPU_CALL_AVX(fn, args) if (CV_CPU_HAS_SUPPORT_AVX) return (opt_AVX::fn args)
 #else
+#  define CV_TRY_AVX 0
 #  define CV_CPU_HAS_SUPPORT_AVX 0
 #  define CV_CPU_CALL_AVX(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_AVX(fn, args, mode, ...)  CV_CPU_CALL_AVX(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_FP16
+#  define CV_TRY_FP16 1
 #  define CV_CPU_HAS_SUPPORT_FP16 1
 #  define CV_CPU_CALL_FP16(fn, args) return (opt_FP16::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_FP16
+#  define CV_TRY_FP16 1
 #  define CV_CPU_HAS_SUPPORT_FP16 (cv::checkHardwareSupport(CV_CPU_FP16))
 #  define CV_CPU_CALL_FP16(fn, args) if (CV_CPU_HAS_SUPPORT_FP16) return (opt_FP16::fn args)
 #else
+#  define CV_TRY_FP16 0
 #  define CV_CPU_HAS_SUPPORT_FP16 0
 #  define CV_CPU_CALL_FP16(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_FP16(fn, args, mode, ...)  CV_CPU_CALL_FP16(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX2
+#  define CV_TRY_AVX2 1
 #  define CV_CPU_HAS_SUPPORT_AVX2 1
 #  define CV_CPU_CALL_AVX2(fn, args) return (opt_AVX2::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX2
+#  define CV_TRY_AVX2 1
 #  define CV_CPU_HAS_SUPPORT_AVX2 (cv::checkHardwareSupport(CV_CPU_AVX2))
 #  define CV_CPU_CALL_AVX2(fn, args) if (CV_CPU_HAS_SUPPORT_AVX2) return (opt_AVX2::fn args)
 #else
+#  define CV_TRY_AVX2 0
 #  define CV_CPU_HAS_SUPPORT_AVX2 0
 #  define CV_CPU_CALL_AVX2(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_AVX2(fn, args, mode, ...)  CV_CPU_CALL_AVX2(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_FMA3
+#  define CV_TRY_FMA3 1
 #  define CV_CPU_HAS_SUPPORT_FMA3 1
 #  define CV_CPU_CALL_FMA3(fn, args) return (opt_FMA3::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_FMA3
+#  define CV_TRY_FMA3 1
 #  define CV_CPU_HAS_SUPPORT_FMA3 (cv::checkHardwareSupport(CV_CPU_FMA3))
 #  define CV_CPU_CALL_FMA3(fn, args) if (CV_CPU_HAS_SUPPORT_FMA3) return (opt_FMA3::fn args)
 #else
+#  define CV_TRY_FMA3 0
 #  define CV_CPU_HAS_SUPPORT_FMA3 0
 #  define CV_CPU_CALL_FMA3(fn, args)
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_FMA3(fn, args, mode, ...)  CV_CPU_CALL_FMA3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_NEON
+#  define CV_TRY_NEON 1
 #  define CV_CPU_HAS_SUPPORT_NEON 1
 #  define CV_CPU_CALL_NEON(fn, args) return (opt_NEON::fn args)
 #elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_NEON
+#  define CV_TRY_NEON 1
 #  define CV_CPU_HAS_SUPPORT_NEON (cv::checkHardwareSupport(CV_CPU_NEON))
 #  define CV_CPU_CALL_NEON(fn, args) if (CV_CPU_HAS_SUPPORT_NEON) return (opt_NEON::fn args)
 #else
+#  define CV_TRY_NEON 0
 #  define CV_CPU_HAS_SUPPORT_NEON 0
 #  define CV_CPU_CALL_NEON(fn, args)
 #endif
diff --git a/modules/dnn/src/layers/convolution_layer.cpp b/modules/dnn/src/layers/convolution_layer.cpp
@@ -315,7 +315,7 @@ class ConvolutionLayerImpl : public BaseConvolutionLayerImpl
             int inpCnAll = input.size[1], width = input.size[3], height = input.size[2];
             int inpCn = inpCnAll / ngroups;
             p.is1x1_ = kernel == Size(0,0) && pad == Size(0, 0);
-            p.useAVX2 = checkHardwareSupport(CPU_AVX2);
+            p.useAVX2 = CV_CPU_HAS_SUPPORT_AVX2;
 
             int ncn = std::min(inpCn, (int)BLK_SIZE_CN);
             p.ofstab_.resize(kernel.width*kernel.height*ncn);
@@ -486,7 +486,7 @@ class ConvolutionLayerImpl : public BaseConvolutionLayerImpl
                         // now compute dot product of the weights
                         // and im2row-transformed part of the tensor
                         int bsz = ofs1 - ofs0;
-                    #if CV_DNN_TRY_AVX2
+                    #if CV_TRY_AVX2
                         if(useAVX2)
                             fastConv_avx2(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
                                           outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
@@ -776,7 +776,7 @@ class DeConvolutionLayerImpl : public BaseConvolutionLayerImpl
             b_ = &b;
             c_ = &c;
             nstripes_ = nstripes;
-            useAVX2 = checkHardwareSupport(CPU_AVX2);
+            useAVX2 = CV_CPU_HAS_SUPPORT_AVX2;
         }
 
         void operator()(const Range& range_) const
@@ -794,7 +794,7 @@ class DeConvolutionLayerImpl : public BaseConvolutionLayerImpl
             size_t bstep = b_->step1();
             size_t cstep = c_->step1();
 
-        #if CV_DNN_TRY_AVX2
+        #if CV_TRY_AVX2
             if( useAVX2 )
                 fastGEMM_avx2( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
             else
diff --git a/modules/dnn/src/layers/fully_connected_layer.cpp b/modules/dnn/src/layers/fully_connected_layer.cpp
@@ -127,7 +127,7 @@ class FullyConnectedLayerImpl : public InnerProductLayer
             biasMat_ = &biasMat;
             dstMat_ = &dstMat;
             nstripes_ = nstripes;
-            useAVX2_ = checkHardwareSupport(CPU_AVX2);
+            useAVX2_ = CV_CPU_HAS_SUPPORT_AVX2;
         }
 
         void operator()(const Range& r) const
@@ -161,7 +161,7 @@ class FullyConnectedLayerImpl : public InnerProductLayer
 
                 memcpy(sptr, sptr_, vecsize*sizeof(sptr[0]));
 
-            #if CV_DNN_TRY_AVX2
+            #if CV_TRY_AVX2
                 if( useAVX2_ )
                     fastGEMM1T_avx2( sptr, wptr, wstep, biasptr, dptr, nw, vecsize);
                 else
diff --git a/modules/dnn/src/layers/layers_common.avx2.cpp b/modules/dnn/src/layers/layers_common.avx2.cpp
@@ -43,10 +43,6 @@
 #include "layers_common.hpp"
 #include "opencv2/core/hal/intrin.hpp"
 
-#if CV_DNN_TRY_AVX2
-
-#include <immintrin.h>
-
 namespace cv {
 namespace dnn {
 
@@ -334,7 +330,6 @@ void fastGEMM_avx2( const float* aptr, size_t astep, const float* bptr,
             _mm256_storeu_ps(cptr3 + n + 8, d31);
         }
     }
-    _mm256_zeroupper();
 
     for( ; n < nb; n++ )
     {
@@ -350,9 +345,8 @@ void fastGEMM_avx2( const float* aptr, size_t astep, const float* bptr,
             cptr0[n] = d0;
         }
     }
+    _mm256_zeroupper();
 }
 
 }
 }
-
-#endif
diff --git a/modules/dnn/src/layers/layers_common.hpp b/modules/dnn/src/layers/layers_common.hpp
@@ -63,9 +63,7 @@ void getConvPoolPaddings(const Size& inp, const Size& out,
                          const Size &kernel, const Size &stride,
                          const String &padMode, Size &pad);
 
-#if CV_SSE2
-#define CV_DNN_TRY_AVX2 1
-
+#if CV_TRY_AVX2
 void fastConv_avx2(const float* weights, size_t wstep, const float* bias,
                    const float* rowbuf, float* output, const int* outShape,
                    int blockSize, int vecsize, int vecsize_aligned,
@@ -76,9 +74,6 @@ void fastGEMM1T_avx2( const float* vec, const float* weights,
 void fastGEMM_avx2( const float* aptr, size_t astep, const float* bptr0,
                    size_t bstep, float* cptr, size_t cstep,
                    int ma, int na, int nb );
-
-#else
-#define CV_DNN_TRY_AVX2 0
 #endif
 
 }

Original file line number	Diff line number	Diff line change
`@@ -43,10 +43,6 @@`
`43`	`43`	`#include "layers_common.hpp"`
`44`	`44`	`#include "opencv2/core/hal/intrin.hpp"`
`45`	`45`
`46`		`-#if CV_DNN_TRY_AVX2`
`47`		`-`
`48`		`-#include <immintrin.h>`
`49`		`-`
`50`	`46`	`namespace cv {`
`51`	`47`	`namespace dnn {`
`52`	`48`
`@@ -334,7 +330,6 @@ void fastGEMM_avx2( const float* aptr, size_t astep, const float* bptr,`
`334`	`330`	`_mm256_storeu_ps(cptr3 + n + 8, d31);`
`335`	`331`	`}`
`336`	`332`	`}`
`337`		`- _mm256_zeroupper();`
`338`	`333`
`339`	`334`	`for( ; n < nb; n++ )`
`340`	`335`	`{`
`@@ -350,9 +345,8 @@ void fastGEMM_avx2( const float* aptr, size_t astep, const float* bptr,`
`350`	`345`	`cptr0[n] = d0;`
`351`	`346`	`}`
`352`	`347`	`}`
	`348`	`+ _mm256_zeroupper();`
`353`	`349`	`}`
`354`	`350`
`355`	`351`	`}`
`356`	`352`	`}`
`357`		`-`
`358`		`-#endif`