Merge pull request #21142 from alalek:dnn_two_inputs_ocl_fp16_3.4

modules/dnn/src/dnn.cpp

@@ -597,29 +597,26 @@ struct DataLayer : public Layer
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
+        // FIXIT: add wrapper without exception suppression
         CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
                    forward_ocl(inputs_arr, outputs_arr, internals_arr))
 
-        if (outputs_arr.depth() == CV_16S)
-        {
-            forward_fallback(inputs_arr, outputs_arr, internals_arr);
-            return;
-        }
+        bool isFP16 = outputs_arr.depth() == CV_16S;
 
         std::vector<Mat> outputs, internals;
         outputs_arr.getMatVector(outputs);
         internals_arr.getMatVector(internals);
 
-        // Supported modes:
-        // | Input type | Output type |
-        // |       fp32 |        fp32 |
-        // |      uint8 |        fp32 |
         for (int i = 0; i < inputsData.size(); ++i)
         {
             double scale = scaleFactors[i];
             Scalar& mean = means[i];
+
             CV_Assert(mean == Scalar() || inputsData[i].size[1] <= 4);
-            CV_CheckTypeEQ(outputs[i].type(), CV_32FC1, "");
+            if (isFP16)
+                CV_CheckTypeEQ(outputs[i].type(), CV_16SC1, "");
+            else
+                CV_CheckTypeEQ(outputs[i].type(), CV_32FC1, "");
 
             bool singleMean = true;
             for (int j = 1; j < std::min(4, inputsData[i].size[1]) && singleMean; ++j)
@@ -629,34 +626,49 @@ struct DataLayer : public Layer
 
             if (singleMean)
             {
-                inputsData[i].convertTo(outputs[i], CV_32F, scale, -mean[0] * scale);
+                if (isFP16)
+                {
+                    Mat input_f32;
+                    inputsData[i].convertTo(input_f32, CV_32F, scale, -mean[0] * scale);
+                    convertFp16(input_f32, outputs[i]);
+                }
+                else
+                {
+                    inputsData[i].convertTo(outputs[i], CV_32F, scale, -mean[0] * scale);
+                }
             }
             else
             {
                 for (int n = 0; n < inputsData[i].size[0]; ++n)
+                {
                     for (int c = 0; c < inputsData[i].size[1]; ++c)
                     {
                         Mat inp = getPlane(inputsData[i], n, c);
                         Mat out = getPlane(outputs[i], n, c);
-                        inp.convertTo(out, CV_32F, scale, -mean[c] * scale);
+                        if (isFP16)
+                        {
+                            Mat input_f32;
+                            inp.convertTo(input_f32, CV_32F, scale, -mean[c] * scale);
+                            convertFp16(input_f32, out);
+                        }
+                        else
+                        {
+                            inp.convertTo(out, CV_32F, scale, -mean[c] * scale);
+                        }
                     }
+                }
             }
         }
     }
 
 #ifdef HAVE_OPENCL
-    std::vector<Mat> tmp_expressions;
     bool forward_ocl(InputArrayOfArrays, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
     {
-        // Supported modes:
-        // | Input type | Output type |
-        // |       fp32 |        fp32 |
-        // |       fp32 |        fp16 |
-        // |      uint8 |        fp32 |
+        bool isFP16 = outputs_.depth() == CV_16S;
+
         std::vector<UMat> outputs;
         outputs_.getUMatVector(outputs);
 
-        tmp_expressions.clear();
         for (int i = 0; i < inputsData.size(); ++i)
         {
             Mat inputData = inputsData[i];
@@ -664,58 +676,55 @@ struct DataLayer : public Layer
             double scale = scaleFactors[i];
             Scalar& mean = means[i];
 
-            CV_Assert(mean == Scalar() || inputsData[i].size[1] <= 4);
+            CV_Assert(mean == Scalar() || inputData.size[1] <= 4);
+            if (isFP16)
+                CV_CheckTypeEQ(outputs[i].type(), CV_16SC1, "");
+            else
+                CV_CheckTypeEQ(outputs[i].type(), CV_32FC1, "");
+
             bool singleMean = true;
-            for (int j = 1; j < std::min(4, inputsData[i].size[1]) && singleMean; ++j)
+            for (int j = 1; j < std::min(4, inputData.size[1]) && singleMean; ++j)
             {
                 singleMean = mean[j] == mean[j - 1];
             }
 
-            if (outputs_.depth() == CV_16S)
+            if (singleMean)
             {
-                if (singleMean)
+                if (isFP16)
                 {
-                    tmp_expressions.push_back(Mat(scale * (inputsData[i] - mean[0])));
-                    convertFp16(tmp_expressions.back(), outputs[i]);
+                    UMat input_i;
+                    inputData.convertTo(input_i, CV_32F, scale, -mean[0] * scale);
+                    convertFp16(input_i, outputs[i]);
                 }
                 else
                 {
-                    for (int n = 0; n < inputsData[i].size[0]; ++n)
-                        for (int c = 0; c < inputsData[i].size[1]; ++c)
-                        {
-                            Mat inp = getPlane(inputsData[i], n, c);
-
-                            std::vector<cv::Range> plane(4, Range::all());
-                            plane[0] = Range(n, n + 1);
-                            plane[1] = Range(c, c + 1);
-                            UMat out = outputs[i](plane).reshape(1, inp.dims, inp.size);
-
-                            tmp_expressions.push_back(scale * (inp - mean[c]));
-                            convertFp16(tmp_expressions.back(), out);
-                        }
+                    inputData.convertTo(outputs[i], CV_32F, scale, -mean[0] * scale);
                 }
             }
             else
             {
-                CV_Assert(outputs_.depth() == CV_32F);
-                if (singleMean)
-                {
-                    inputsData[i].convertTo(outputs[i], CV_32F, scale, -mean[0] * scale);
-                }
-                else
+                for (int n = 0; n < inputData.size[0]; ++n)
                 {
-                    for (int n = 0; n < inputsData[i].size[0]; ++n)
-                        for (int c = 0; c < inputsData[i].size[1]; ++c)
-                        {
-                            Mat inp = getPlane(inputsData[i], n, c);
+                    for (int c = 0; c < inputData.size[1]; ++c)
+                    {
+                        Mat inp = getPlane(inputData, n, c);
 
-                            std::vector<cv::Range> plane(4, Range::all());
-                            plane[0] = Range(n, n + 1);
-                            plane[1] = Range(c, c + 1);
-                            UMat out = outputs[i](plane).reshape(1, inp.dims, inp.size);
+                        std::vector<cv::Range> plane(4, Range::all());
+                        plane[0] = Range(n, n + 1);
+                        plane[1] = Range(c, c + 1);
+                        UMat out = outputs[i](plane).reshape(1, inp.dims, inp.size);
 
+                        if (isFP16)
+                        {
+                            UMat input_i;
+                            inp.convertTo(input_i, CV_32F, scale, -mean[c] * scale);
+                            convertFp16(input_i, out);
+                        }
+                        else
+                        {
                             inp.convertTo(out, CV_32F, scale, -mean[c] * scale);
                         }
+                    }
                 }
             }
         }

modules/dnn/test/test_layers.cpp

@@ -1380,57 +1380,6 @@ INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_DLDT_two_inputs_3dim, Combine(
   testing::ValuesIn(list_sizes)
 ));
 
-typedef testing::TestWithParam<tuple<int, int, tuple<Backend, Target> > > Test_DLDT_two_inputs;
-TEST_P(Test_DLDT_two_inputs, as_backend)
-{
-    static const float kScale = 0.5f;
-    static const float kScaleInv = 1.0f / kScale;
-
-    Backend backendId = get<0>(get<2>(GetParam()));
-    Target targetId = get<1>(get<2>(GetParam()));
-
-    Net net;
-    LayerParams lp;
-    lp.type = "Eltwise";
-    lp.name = "testLayer";
-    lp.set("operation", "sum");
-    int eltwiseId = net.addLayerToPrev(lp.name, lp.type, lp);  // connect to a first input
-    net.connect(0, 1, eltwiseId, 1);  // connect to a second input
-
-    int inpSize[] = {1, 2, 3, 4};
-    Mat firstInp(4, &inpSize[0], get<0>(GetParam()));
-    Mat secondInp(4, &inpSize[0], get<1>(GetParam()));
-    randu(firstInp, 0, 255);
-    randu(secondInp, 0, 255);
-
-    net.setInputsNames({"data", "second_input"});
-    net.setInput(firstInp, "data", kScale);
-    net.setInput(secondInp, "second_input", kScaleInv);
-    net.setPreferableBackend(backendId);
-    net.setPreferableTarget(targetId);
-    Mat out = net.forward();
-
-    Mat ref;
-    addWeighted(firstInp, kScale, secondInp, kScaleInv, 0, ref, CV_32F);
-    // Output values are in range [0, 637.5].
-    double l1 = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.06 : 1e-6;
-    double lInf = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.3 : 1e-5;
-    normAssert(out, ref, "", l1, lInf);
-    if (cvtest::debugLevel > 0 || HasFailure())
-    {
-        std::cout << "input1 scale=" << kScale << " input2 scale=" << kScaleInv << std::endl;
-        std::cout << "input1: " << firstInp.size << " " << firstInp.reshape(1, 1) << std::endl;
-        std::cout << "input2: " << secondInp.size << " " << secondInp.reshape(1, 1) << std::endl;
-        std::cout << "ref: " << ref.reshape(1, 1) << std::endl;
-        std::cout << "out: " << out.reshape(1, 1) << std::endl;
-    }
-}
-
-INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_DLDT_two_inputs, Combine(
-  Values(CV_8U, CV_32F), Values(CV_8U, CV_32F),
-  dnnBackendsAndTargets()
-));
-
 class UnsupportedLayer : public Layer
 {
 public:

modules/dnn/test/test_misc.cpp

@@ -828,4 +828,64 @@ INSTANTIATE_TEST_CASE_P(/**/, Test_Model_Optimizer,
 
 #endif  // HAVE_INF_ENGINE
 
+typedef testing::TestWithParam<tuple<MatDepth, MatDepth, tuple<Backend, Target> > > Test_two_inputs;
+TEST_P(Test_two_inputs, basic)
+{
+    static const float kScale = 0.5f;
+    static const float kScaleInv = 1.0f / kScale;
+
+    Backend backendId = get<0>(get<2>(GetParam()));
+    Target targetId = get<1>(get<2>(GetParam()));
+
+    Net net;
+    LayerParams lp;
+    lp.type = "Eltwise";
+    lp.name = "testLayer";
+    lp.set("operation", "sum");
+    int eltwiseId = net.addLayerToPrev(lp.name, lp.type, lp);  // connect to a first input
+    net.connect(0, 1, eltwiseId, 1);  // connect to a second input
+
+    int inpSize[] = {1, 2, 3, 4};
+    Mat firstInp(4, &inpSize[0], get<0>(GetParam()));
+    Mat secondInp(4, &inpSize[0], get<1>(GetParam()));
+    randu(firstInp, 0, 100);
+    randu(secondInp, 0, 100);
+
+#ifndef CV_CXX11
+    std::vector<String> input_names;
+    input_names.push_back("data");
+    input_names.push_back("second_input");
+    net.setInputsNames(input_names);
+#else
+    net.setInputsNames({"data", "second_input"});
+#endif
+    net.setInput(firstInp, "data", kScale);
+    net.setInput(secondInp, "second_input", kScaleInv);
+    net.setPreferableBackend(backendId);
+    net.setPreferableTarget(targetId);
+    Mat out = net.forward();
+
+    Mat ref;
+    addWeighted(firstInp, kScale, secondInp, kScaleInv, 0, ref, CV_32F);
+
+    double l1 = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.06 : 1e-6;
+    double lInf = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.3 : 1e-5;
+    normAssert(out, ref, "", l1, lInf);
+
+    if (cvtest::debugLevel > 0 || HasFailure())
+    {
+        std::cout << "input1 scale=" << kScale << " input2 scale=" << kScaleInv << std::endl;
+        std::cout << "input1: " << firstInp.size << " " << firstInp.reshape(1, 1) << std::endl;
+        std::cout << "input2: " << secondInp.size << " " << secondInp.reshape(1, 1) << std::endl;
+        std::cout << "ref: " << ref.reshape(1, 1) << std::endl;
+        std::cout << "out: " << out.reshape(1, 1) << std::endl;
+    }
+}
+
+INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_two_inputs, Combine(
+    Values(CV_32F, CV_8U),
+    Values(CV_32F, CV_8U),
+    dnnBackendsAndTargets()
+));
+
 }} // namespace