add enableWinograd API for Net.

modules/dnn/include/opencv2/dnn/all_layers.hpp

@@ -259,6 +259,7 @@ CV__DNN_INLINE_NS_BEGIN
         bool fusedActivation = false;
         bool fusedAdd = false;
         bool isConv2D = false; // Should be deleted after fastconv branch support Conv1D and Conv3D.
+        bool useWinograd = false; // Flag whether to use Winograd to speed up 3x3 convolution.
     };
 
     class CV_EXPORTS ConvolutionLayerInt8 : public BaseConvolutionLayer
@@ -270,6 +271,7 @@ CV__DNN_INLINE_NS_BEGIN
         // quantization type flag. The perChannel default is true, that means it contains the parameters
         // of per-Channel quantization. Otherwise, that means this layer contains per-Tensor quantized parameters.
         bool per_channel;
+        bool useWinograd = true; // Flag whether to use Winograd to speed up 3x3 convolution.
         static Ptr<BaseConvolutionLayer> create(const LayerParams& params);
     };
 

modules/dnn/include/opencv2/dnn/dnn.hpp

@@ -837,6 +837,12 @@ CV__DNN_INLINE_NS_BEGIN
          */
         CV_WRAP void enableFusion(bool fusion);
 
+        /** @brief Enables or disables the Winograd compute branch. The Winograd compute branch can speed up
+         * 3x3 Convolution at a small loss of accuracy.
+        * @param useWinograd true to enable the Winograd compute branch. The default is true.
+        */
+        CV_WRAP void enableWinograd(bool useWinograd);
+
         /** @brief Returns overall time for inference and timings (in ticks) for layers.
          *
          * Indexes in returned vector correspond to layers ids. Some layers can be fused with others,

modules/dnn/src/int8layers/convolution_layer.cpp

@@ -41,7 +41,7 @@ public:
     BaseConvolutionLayerInt8Impl(const LayerParams &params)
     {
         setParamsFrom(params);
-        getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode, adjust_pads);
+        getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode, adjust_pads, useWinograd);
 
         numOutput = params.get<int>("num_output");
         int ngroups = params.get<int>("group", 1);

modules/dnn/src/int8layers/layers_common.hpp

@@ -23,7 +23,7 @@ namespace dnn
 {
 void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin,
                                 std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations,
-                                cv::String &padMode, std::vector<size_t>& adjust_pads);
+                                cv::String &padMode, std::vector<size_t>& adjust_pads, bool& useWinograd);
 
 void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<bool>& globalPooling,
                             std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end, std::vector<size_t>& strides, cv::String &padMode);

modules/dnn/src/layers/convolution_layer.cpp

@@ -89,7 +89,8 @@ public:
     BaseConvolutionLayerImpl(const LayerParams &params)
     {
         setParamsFrom(params);
-        getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode, adjust_pads);
+        getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations,
+                                   padMode, adjust_pads, useWinograd);
 
         numOutput = params.get<int>("num_output");
         int ngroups = params.get<int>("group", 1);
@@ -2112,7 +2113,7 @@ public:
                 int dilation_w = dilations.back();
 
                 fastConv2dImpl = initFastConv2d(ngroups, K, C, Hk, Wk, stride_w, stride_h, dilation_w,
-                                                dilation_h, pads_begin, pads_end, weightsMat, &biasvec[0]);
+                                                dilation_h, pads_begin, pads_end, weightsMat, &biasvec[0], useWinograd);
             }
 
             if (fastConv2dImpl)

modules/dnn/src/layers/fast_convolution/fast_convolution.cpp

@@ -23,7 +23,8 @@ Ptr<FastConv2d> initFastConv2d(
         const std::vector<size_t>& pads_begin,
         const std::vector<size_t>& pads_end,
         InputArray _weightsMat,
-        float* srcBias)
+        float* srcBias,
+        bool useWinograd)
 {
     Ptr<FastConv2d> conv = makePtr<FastConv2d>();
 
@@ -48,11 +49,11 @@ Ptr<FastConv2d> initFastConv2d(
     const size_t wstep = weightsMat.step1();
 
 #if CV_NEON // For now, winograd is ARM platform only.
-    if (ngroups == 1 && Hk ==3 && Wk == 3 && stride_x == 1 && stride_y == 1 &&
+    if (useWinograd && ngroups == 1 && Hk ==3 && Wk == 3 && stride_x == 1 && stride_y == 1 &&
         dilation_x == 1 && dilation_y ==1 && K >= 16 && C >= 16)
-        conv->ifWinograd63 = true;
+        conv->useWinograd63 = true;
 #else
-    conv->ifWinograd63 = false;
+    conv->useWinograd63 = false;
 #endif
 
     float *srcWeights = (float *)weightsMat.data;
@@ -115,7 +116,7 @@ Ptr<FastConv2d> initFastConv2d(
         }});
 
         // Prepare Weight for Winograd F(6x6, 3x3)
-        if (conv->ifWinograd63)
+        if (conv->useWinograd63)
         {
             initWinograd63(conv, weightsMat, K, C);
         }
@@ -191,10 +192,7 @@ void runFastConv2d(InputArray _input, OutputArray _output, const Ptr<FastConv2d>
     }
 
 #if CV_NEON
-    if (conv->ifWinograd63
+    if (conv->useWinograd63 && inputShape[2] > 12 && inputShape[3] > 12)
-         && inputShape[2] > 12 && inputShape[3] > 12
-         && inputShape[2] < 120 && inputShape[3] < 120
-         )
     {
         if (runWinograd63(input, fusedAddMat, output, conv, ntasks, minval, maxval, activ, ifMinMaxAct))
             return;

modules/dnn/src/layers/fast_convolution/fast_convolution.hpp

@@ -44,7 +44,7 @@ struct FastConv2d
     std::vector<float> weightsBuf;        // For generic Conv 2D
     std::vector<float> weightsWino63Buf;  // For Winograd F(6x6, 3x3).
     std::vector<float> biasBuf;
-    bool ifWinograd63 = false;
+    bool useWinograd63 = false;
     bool useAVX2 = checkHardwareSupport(CPU_AVX2);
     bool useNEON = checkHardwareSupport(CPU_NEON);
 };
@@ -58,7 +58,7 @@ Ptr<FastConv2d> initFastConv2d(
         const std::vector<size_t>& pads_begin,
         const std::vector<size_t>& pads_end,
         InputArray weightsMat,
-        float* srcBias);
+        float* srcBias, bool useWinograd);
 
 // It contains different computing branches, like winograd, 1x1 conv.
 void runFastConv2d(InputArray _input, OutputArray _output, const Ptr<FastConv2d>& conv, int ntasks,

modules/dnn/src/layers/fast_convolution/winograd_3x3s1_f63.cpp

@@ -1689,7 +1689,7 @@ int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _outpu
 
 void initWinograd63(Ptr<FastConv2d>& conv, InputArray _weightsMat, int K, int C)
 {
-    conv->ifWinograd63 = false;
+    conv->useWinograd63 = false;
 }
 
 int runWinograd63(InputArray _input, OutputArray _output, const Ptr<FastConv2d>& conv, int ntasks, float minval, float maxval, ActivationLayer* activ, bool ifMinMaxAct)

modules/dnn/src/layers/layers_common.cpp

@@ -187,12 +187,14 @@ void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kern
 
 void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin,
                                 std::vector<size_t>& pads_end, std::vector<size_t>& strides,
-                                std::vector<size_t>& dilations, cv::String &padMode, std::vector<size_t>& adjust_pads)
+                                std::vector<size_t>& dilations, cv::String &padMode, std::vector<size_t>& adjust_pads,
+                                bool& useWinograd)
 {
     util::getKernelSize(params, kernel);
     util::getStrideAndPadding(params, pads_begin, pads_end, strides, padMode, kernel.size());
     util::getParameter(params, "dilation", "dilation", dilations, true, std::vector<size_t>(kernel.size(), 1));
     util::getParameter(params, "adj", "adj", adjust_pads, true, std::vector<size_t>(kernel.size(), 0));
+    useWinograd = params.get<bool>("use_winograd", true);
 
     for (int i = 0; i < dilations.size(); i++)
         CV_Assert(dilations[i] > 0);

modules/dnn/src/layers/layers_common.hpp

@@ -61,7 +61,7 @@ namespace dnn
 {
 void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin,
                                 std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations,
-                                cv::String &padMode, std::vector<size_t>& adjust_pads);
+                                cv::String &padMode, std::vector<size_t>& adjust_pads, bool& useWinograd);
 
 void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<bool>& globalPooling,
                             std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end, std::vector<size_t>& strides, cv::String &padMode);

modules/dnn/src/net.cpp

@@ -395,6 +395,13 @@ void Net::enableFusion(bool fusion)
     return impl->enableFusion(fusion);
 }
 
+void Net::enableWinograd(bool useWinograd)
+{
+    CV_TRACE_FUNCTION();
+    CV_Assert(impl);
+    return impl->enableWinograd(useWinograd);
+}
+
 void Net::setHalideScheduler(const String& scheduler)
 {
     CV_TRACE_FUNCTION();

modules/dnn/src/net_impl.cpp

@@ -55,6 +55,7 @@ Net::Impl::Impl()
     preferableBackend = (Backend)getParam_DNN_BACKEND_DEFAULT();
     preferableTarget = DNN_TARGET_CPU;
     hasDynamicShapes = false;
+    useWinograd = true;
 }
 
 
@@ -2038,6 +2039,37 @@ void Net::Impl::getMemoryConsumption(
     }
 }
 
+void Net::Impl::enableWinograd(bool useWinograd_)
+{
+    if (useWinograd != useWinograd_)
+    {
+        useWinograd = useWinograd_;
+
+        for (MapIdToLayerData::const_iterator it = layers.begin(); it != layers.end(); it++)
+        {
+            int lid = it->first;
+            LayerData &ld = layers[lid];
+            Ptr<Layer>& currLayer = ld.layerInstance;
+
+            if (ld.type == "Convolution")
+            {
+                ld.params.set("use_winograd", useWinograd_);
+                Ptr<ConvolutionLayer> convLayer = ld.layerInstance.dynamicCast<ConvolutionLayer>();
+                if (!convLayer.empty())
+                    convLayer->useWinograd = useWinograd_;
+            }
+
+            if (ld.type == "ConvolutionInt8")
+            {
+                Ptr<ConvolutionLayerInt8> convLayer = currLayer.dynamicCast<ConvolutionLayerInt8>();
+                ld.params.set("use_winograd", useWinograd_);
+                if (!convLayer.empty())
+                    convLayer->useWinograd = useWinograd_;
+            }
+        }
+    }
+}
+
 
 // TODO drop?
 void Net::Impl::getLayerTypes(std::vector<String>& layersTypes) const

modules/dnn/src/net_impl.hpp

@@ -64,6 +64,7 @@ struct Net::Impl : public detail::NetImplBase
     bool netWasQuantized;
     bool fusion;
     bool isAsync;  // FIXIT: drop
+    bool useWinograd;
     std::vector<int64> layersTimings;
 
 
@@ -211,6 +212,7 @@ struct Net::Impl : public detail::NetImplBase
     void enableFusion(bool fusion_);
 
     virtual void fuseLayers(const std::vector<LayerPin>& blobsToKeep_);
+    void enableWinograd(bool useWinograd_);
 
     void allocateLayers(const std::vector<LayerPin>& blobsToKeep_);
 

modules/dnn/src/net_quantization.cpp

@@ -51,6 +51,7 @@ Net Net::Impl::quantize(Net& net, InputArrayOfArrays calibData, int inputsDtype,
     setPreferableBackend(net, DNN_BACKEND_OPENCV);
     setPreferableTarget(DNN_TARGET_CPU);
     enableFusion(false);
+    enableWinograd(false);
 
     if (calibData.isMat())
     {

modules/dnn/test/test_torch_importer.cpp

@@ -476,6 +476,7 @@ TEST_P(Test_Torch_nets, ENet_accuracy)
         ASSERT_TRUE(!net.empty());
     }
 
+    net.enableWinograd(false);
     net.setPreferableBackend(backend);
     net.setPreferableTarget(target);