Merge pull request #20287 from hanliutong:dev-rvv-0.10

hanliutong · web-flow · commit aaca4987c9ff · 2021-08-11T01:16:03.000+03:00
Optimization of DNN using native RISC-V vector intrinsics.

* Use RVV to optimize fastGEMM (FP32) in DNN.

* Use RVV to optimize fastGEMM1T in DNN.

* Use RVV to optimize fastConv in DNN.

* Use RVV to optimize fastDepthwiseConv in DNN.

* Vectorize tails using vl.

* Use "vl" instead of scalar to handle small block in fastConv.

* Fix memory access out of bound in "fastGEMM1T".

* Remove setvl.

* Remove useless initialization.

* Use loop unrolling to handle tail part instead of switch.
diff --git a/modules/dnn/CMakeLists.txt b/modules/dnn/CMakeLists.txt
@@ -8,7 +8,7 @@ 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_force_all("layers/layers_common" AVX AVX2 AVX512_SKX)
+ocv_add_dispatched_file_force_all("layers/layers_common" AVX AVX2 AVX512_SKX RVV)
 
 ocv_add_module(dnn opencv_core opencv_imgproc WRAP python java objc js)
 
diff --git a/modules/dnn/src/layers/convolution_layer.cpp b/modules/dnn/src/layers/convolution_layer.cpp
@@ -914,11 +914,12 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
         bool useAVX;
         bool useAVX2;
         bool useAVX512;
+        bool useRVV;
         int blk_size_cn;
 
         ParallelConv()
             : input_(0), weights_(0), output_(0), ngroups_(0), nstripes_(0),
-              biasvec_(0), reluslope_(0), activ_(0), is1x1_(false), useAVX(false), useAVX2(false), useAVX512(false)
+              biasvec_(0), reluslope_(0), activ_(0), is1x1_(false), useAVX(false), useAVX2(false), useAVX512(false), useRVV(false)
             , blk_size_cn(0)
         {}
 
@@ -976,6 +977,7 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
             p.useAVX    = checkHardwareSupport(CPU_AVX)  && isConv2D;
             p.useAVX2   = checkHardwareSupport(CPU_AVX2) && isConv2D;
             p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX  && isConv2D;
+            p.useRVV   = checkHardwareSupport(CPU_RVV) && isConv2D;
 
             int kernel_d = isConv3D? kernel_size[0] : 1;
             int kernel_h = isConv1D? 1 : kernel_size[kernel_size.size() - 2];
@@ -1176,6 +1178,13 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
                                     stride_h, stride_w, dilation_h, dilation_w, pad_t, pad_l,
                                     biasptr, relu, inptr_, height, width, outptr_, out_d, outH, outW);
                             else
+                        #endif
+                        #if CV_TRY_RVV
+                            if(useRVV)
+                                opt_RVV::fastDepthwiseConv(wptr, kernel_h, kernel_w,
+                                    stride_h, stride_w, dilation_h, dilation_w, pad_t, pad_l,
+                                    biasptr, relu, inptr_, height, width, outptr_, out_d, outH, outW);
+                            else
                         #endif
                             {
                                 const float w00_ = wptr[0], w01_ = wptr[1], w02_ = wptr[2],
@@ -1546,6 +1555,12 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
                             opt_AVX::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
                                          outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
                         else
+                    #endif
+                    #if CV_TRY_RVV
+                        if(useRVV)
+                            opt_RVV::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
+                                         outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
+                        else
                     #endif
                         for( int i = 0; i < outCn; i += 2 )
                         {
@@ -2297,6 +2312,7 @@ class DeConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
             useAVX = checkHardwareSupport(CPU_AVX);
             useAVX2 = checkHardwareSupport(CPU_AVX2);
             useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX;
+            useRVV = checkHardwareSupport(CPU_RVV);
         }
 
         void operator()(const Range& range_) const CV_OVERRIDE
@@ -2328,6 +2344,12 @@ class DeConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
             if( useAVX )
                 opt_AVX::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
             else
+        #endif
+        #if CV_TRY_RVV
+            if( useRVV ) {
+                opt_RVV::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
+            }
+            else
         #endif
             for( m = 0; m < mmax; m += 2 )
             {
@@ -2427,6 +2449,7 @@ class DeConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
         bool useAVX;
         bool useAVX2;
         bool useAVX512;
+        bool useRVV;
     };
 
     class Col2ImInvoker : public cv::ParallelLoopBody
diff --git a/modules/dnn/src/layers/fully_connected_layer.cpp b/modules/dnn/src/layers/fully_connected_layer.cpp
@@ -168,7 +168,7 @@ class FullyConnectedLayerImpl CV_FINAL : public InnerProductLayer
     class FullyConnected : public ParallelLoopBody
     {
     public:
-        FullyConnected() : srcMat(0), weights(0), biasMat(0), activ(0), dstMat(0), nstripes(0), useAVX(false), useAVX2(false), useAVX512(false) {}
+        FullyConnected() : srcMat(0), weights(0), biasMat(0), activ(0), dstMat(0), nstripes(0), useAVX(false), useAVX2(false), useAVX512(false), useRVV(false) {}
 
         static void run(const Mat& srcMat, const Mat& weights, const Mat& biasMat,
                         Mat& dstMat, const ActivationLayer* activ, int nstripes)
@@ -191,6 +191,7 @@ class FullyConnectedLayerImpl CV_FINAL : public InnerProductLayer
             p.useAVX = checkHardwareSupport(CPU_AVX);
             p.useAVX2 = checkHardwareSupport(CPU_AVX2);
             p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX;
+            p.useRVV = checkHardwareSupport(CPU_RVV);
 
             parallel_for_(Range(0, nstripes), p, nstripes);
         }
@@ -239,6 +240,11 @@ class FullyConnectedLayerImpl CV_FINAL : public InnerProductLayer
                 if( useAVX )
                     opt_AVX::fastGEMM1T( sptr, wptr, wstep, biasptr, dptr, nw, vecsize);
                 else
+            #endif
+            #if CV_TRY_RVV
+                if( useRVV )
+                    opt_RVV::fastGEMM1T( sptr, wptr, wstep, biasptr, dptr, nw, vecsize);
+                else
             #endif
                 {
                     int i = 0;
@@ -293,6 +299,7 @@ class FullyConnectedLayerImpl CV_FINAL : public InnerProductLayer
         bool useAVX;
         bool useAVX2;
         bool useAVX512;
+        bool useRVV;
     };
 
 #ifdef HAVE_OPENCL
diff --git a/modules/dnn/src/layers/layers_common.simd.hpp b/modules/dnn/src/layers/layers_common.simd.hpp
