dnn(opencl): bypass unsupported fusion cases 2

diff --git a/modules/dnn/src/dnn.cpp b/modules/dnn/src/dnn.cpp
index 9ee688f497e5ec5c01b3d45e79c64b6f4e34ac23..c789638793a40cf8070d8bc06f3ee7689ded5c1e 100644 (file)
--- a/modules/dnn/src/dnn.cpp
+++ b/modules/dnn/src/dnn.cpp
@@ -2413,14 +2413,42 @@ struct Net::Impl : public detail::NetImplBase
                 }
 
                 // fuse convolution layer followed by eltwise + relu
-                if ( IS_DNN_OPENCL_TARGET(preferableTarget) && ld.layerInstance->type == "Convolution" )
+                while (nextData && IS_DNN_OPENCL_TARGET(preferableTarget) && ld.layerInstance->type == "Convolution")  // semantic of 'if'
                 {
-                    Ptr<EltwiseLayer> nextEltwiseLayer;
-                    if( nextData )
-                        nextEltwiseLayer = nextData->layerInstance.dynamicCast<EltwiseLayer>();
+                    Ptr<EltwiseLayer> nextEltwiseLayer = nextData->layerInstance.dynamicCast<EltwiseLayer>();
+                    if (nextEltwiseLayer.empty())
+                        break;
+
+                    if (pinsToKeep.count(lpNext) != 0)
+                        break;
+                    if (nextData->inputBlobsId.size() != 2)
+                        break;
+
+                    if (!nextData->params.has("operation") || nextData->params.get<String>("operation").toLowerCase() == "sum")
+                    {
+                        if (nextData->params.has("coeff"))
+                        {
+                            DictValue paramCoeff = nextData->params.get("coeff");
+                            int n = paramCoeff.size();
+                            bool isCoeffOneOne = (n == 2);
+                            for (int i = 0; isCoeffOneOne && i < n; i++)
+                            {
+                                float c = paramCoeff.get<float>(i);
+                                isCoeffOneOne &= (c == 1.0f);
+                            }
+                            if (!isCoeffOneOne)
+                            {
+                                CV_LOG_DEBUG(NULL, "DNN/OpenCL: fusion of 'Sum' without coeffs (or {1.0, 1.0}) is supported only");
+                                break;
+                            }
+                        }
+                    }
+                    else
+                    {
+                        CV_LOG_DEBUG(NULL, "DNN/OpenCL: fusion with eltwise operation is not supported: " << nextData->params.get<String>("operation"));
+                        break;
+                    }
 
-                    if( !nextEltwiseLayer.empty() && pinsToKeep.count(lpNext) == 0 &&
-                        nextData && nextData->inputBlobsId.size() == 2 )
                     {
                         LayerData *eltwiseData = nextData;
 
@@ -2517,6 +2545,8 @@ struct Net::Impl : public detail::NetImplBase
                             }
                         }
                     }
+
+                    break;
                 }
             }
 
@@ -2698,11 +2728,11 @@ struct Net::Impl : public detail::NetImplBase
 
         Ptr<Layer> layer = ld.layerInstance;
 
-        TickMeter tm;
-        tm.start();
-
         if( !ld.skip )
         {
+            TickMeter tm;
+            tm.start();
+
             std::map<int, Ptr<BackendNode> >::iterator it = ld.backendNodes.find(preferableBackend);
             if (preferableBackend == DNN_BACKEND_OPENCV || it == ld.backendNodes.end() || it->second.empty())
             {
@@ -2881,12 +2911,15 @@ struct Net::Impl : public detail::NetImplBase
                     CV_Error(Error::StsNotImplemented, "Unknown backend identifier");
                 }
             }
+
+            tm.stop();
+            int64 t = tm.getTimeTicks();
+            layersTimings[ld.id] = (t > 0) ? t : t + 1;  // zero for skipped layers only
         }
         else
-            tm.reset();
-
-        tm.stop();
-        layersTimings[ld.id] = tm.getTimeTicks();
+        {
+            layersTimings[ld.id] = 0;
+        }
 
         ld.flag = 1;
     }

diff --git a/modules/dnn/src/layers/convolution_layer.cpp b/modules/dnn/src/layers/convolution_layer.cpp
index 206ce72fa0b3c0ecd84b32b6df55902e7a4e9ee4..473c07b755033e208e430b591cb73e3f4b8259b6 100644 (file)
--- a/modules/dnn/src/layers/convolution_layer.cpp
+++ b/modules/dnn/src/layers/convolution_layer.cpp
@@ -376,7 +376,7 @@ public:
                 if (activ_power->scale != 1.0f)  // not supported well by implementation, #17964
                 {
                     // FIXIT no way to check number of blobs (like, eltwise input)
-                    CV_LOG_INFO(NULL, "DNN/OpenCL: can't configure Power activation (scale != 1.0f)");
+                    CV_LOG_DEBUG(NULL, "DNN/OpenCL: can't configure Power activation (scale != 1.0f)");
                     activ.release();
                     newActiv = false;
                     return false;

diff --git a/modules/dnn/test/test_layers.cpp b/modules/dnn/test/test_layers.cpp
index 3872f562efbf663962540df53239a640b2d58e99..327f3e9abdfbec13a5deea472a87dd98ecbc44a9 100644 (file)
--- a/modules/dnn/test/test_layers.cpp
+++ b/modules/dnn/test/test_layers.cpp
@@ -2341,10 +2341,6 @@ TEST_P(ConvolutionEltwiseActivationFusion, Accuracy)
     Backend backendId = get<0>(get<4>(GetParam()));
     Target targetId = get<1>(get<4>(GetParam()));
 
-    // bug: https://github.com/opencv/opencv/issues/17945
-    if ((eltwiseOp != "sum" || weightedEltwise) && backendId == DNN_BACKEND_OPENCV && (targetId == DNN_TARGET_OPENCL || targetId == DNN_TARGET_OPENCL_FP16))
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL);
-
     Net net;
     int convId = net.addLayer(convParams.name, convParams.type, convParams);
     int eltwiseId = net.addLayer(eltwiseParams.name, eltwiseParams.type, eltwiseParams);
@@ -2361,7 +2357,9 @@ TEST_P(ConvolutionEltwiseActivationFusion, Accuracy)
             expectedFusedLayers.push_back(activId); // activation is fused with eltwise layer
         else if (targetId == DNN_TARGET_OPENCL || targetId == DNN_TARGET_OPENCL_FP16)
         {
-            if (actType == "ReLU" || actType == "ChannelsPReLU" /*|| actType == "Power"*/)
+            if (eltwiseOp == "sum" && !weightedEltwise &&
+                (actType == "ReLU" || actType == "ChannelsPReLU" /*|| actType == "Power"*/)
+            )
             {
                 expectedFusedLayers.push_back(eltwiseId);
                 expectedFusedLayers.push_back(activId);