Merge pull request #10513 from pengli:dnn

modules/dnn/src/layers/batch_norm_layer.cpp

@@ -22,6 +22,7 @@ class BatchNormLayerImpl : public BatchNormLayer
 {
 public:
     Mat weights_, bias_;
+    Mat weightMat, biasMat;
 
     BatchNormLayerImpl(const LayerParams& params)
     {
@@ -96,17 +97,81 @@ public:
         return true;
     }
 
+    void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs)
+    {
+        if (inputs[0]->dims == 4)
+        {
+            int groups = inputs[0]->size[0];
+            int channels = inputs[0]->size[1];
+            int rows = inputs[0]->size[2];
+            int cols = inputs[0]->size[3];
+            MatShape s = shape(groups * channels, rows * cols);
+            weightMat = Mat(s[0], s[1], CV_32FC1);
+            biasMat = Mat(s[0], s[1], CV_32FC1);
+            for (int n = 0; n < s[0]; n++)
+            {
+                weightMat.row(n).setTo(weights_.at<float>(n % channels));
+                biasMat.row(n).setTo(bias_.at<float>(n % channels));
+            }
+        }
+    }
+
     virtual bool supportBackend(int backendId)
     {
         return backendId == DNN_BACKEND_DEFAULT ||
                backendId == DNN_BACKEND_HALIDE && haveHalide();
     }
 
+#ifdef HAVE_OPENCL
+    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
+    {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inputs_.getUMatVector(inputs);
+        outputs_.getUMatVector(outputs);
+
+        CV_Assert(blobs.size() >= 2);
+        CV_Assert(inputs.size() == 1);
+
+        UMat &inpBlob = inputs[0];
+        CV_Assert(inpBlob.dims == 2 || inpBlob.dims == 4);
+        int groups = inpBlob.size[0];
+        int channels = inpBlob.size[1];
+        int rows = inpBlob.dims > 2 ? inpBlob.size[2] : 1;
+        int cols = inpBlob.dims > 2 ? inpBlob.size[3] : 1;
+
+        for (size_t ii = 0; ii < outputs.size(); ii++)
+        {
+            if (inpBlob.dims == 2)
+            {
+                UMat& src = inputs[ii];
+                UMat& dst = outputs[ii];
+                multiply(src, weights_, dst);
+                add(dst, bias_, dst);
+            }
+            else
+            {
+                MatShape s = shape(groups * channels, rows * cols);
+                UMat src = inputs[ii].reshape(1, s.size(), &s[0]);
+                UMat dst = outputs[ii].reshape(1, s.size(), &s[0]);
+                multiply(src, weightMat, dst);
+                add(dst, biasMat, dst);
+            }
+        }
+        return true;
+    }
+#endif
+
     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
+        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
+                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
         Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
     }
 

modules/dnn/src/layers/blank_layer.cpp

@@ -63,8 +63,22 @@ public:
     }
 
 #ifdef HAVE_OPENCL
-    bool forward_ocl(InputArrayOfArrays inputs, OutputArrayOfArrays outputs, OutputArrayOfArrays internals)
+    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
     {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inputs_.getUMatVector(inputs);
+        outputs_.getUMatVector(outputs);
+
+        for (int i = 0, n = outputs.size(); i < n; ++i)
+        {
+            void *src_handle = inputs[i].handle(ACCESS_READ);
+            void *dst_handle = outputs[i].handle(ACCESS_WRITE);
+            if (src_handle != dst_handle)
+                inputs[i].copyTo(outputs[i]);
+        }
+
         return true;
     }
 #endif

modules/dnn/src/layers/eltwise_layer.cpp

@@ -259,11 +259,63 @@ public:
         }
     };
 
+#ifdef HAVE_OPENCL
+    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
+    {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inputs_.getUMatVector(inputs);
+        outputs_.getUMatVector(outputs);
+
+        switch (op)
+        {
+            case SUM:
+                if (coeffs.empty())
+                {
+                    add(inputs[0], inputs[1], outputs[0]);
+                    for (int i = 2; i < inputs.size(); ++i)
+                        add(outputs[0], inputs[i], outputs[0]);
+                }
+                else
+                {
+                    UMat mul0, mul1;
+                    multiply(coeffs[0], inputs[0], mul0);
+                    multiply(coeffs[1], inputs[1], mul1);
+                    add(mul0, mul1, outputs[0]);
+                    for (int i = 2; i < inputs.size(); ++i)
+                    {
+                        multiply(coeffs[i], inputs[i], mul0);
+                        add(mul0, outputs[0], outputs[0]);
+                    }
+                }
+                break;
+            case PROD:
+                multiply(inputs[0], inputs[1], outputs[0]);
+                for (int i = 2; i < inputs.size(); ++i)
+                    multiply(inputs[i], outputs[0], outputs[0]);
+                break;
+            case MAX:
+                max(inputs[0], inputs[1], outputs[0]);
+                for (int i = 2; i < inputs.size(); ++i)
+                    max(inputs[i], outputs[0], outputs[0]);
+                break;
+            default:
+                return false;
+        }
+        return true;
+    }
+#endif
+
     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
+        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
+                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
         Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
     }
 

modules/dnn/src/layers/normalize_bbox_layer.cpp

@@ -69,11 +69,74 @@ public:
         return true;
     }
 
+#ifdef HAVE_OPENCL
+    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
+    {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+        std::vector<UMat> internals;
+
+        inputs_.getUMatVector(inputs);
+        outputs_.getUMatVector(outputs);
+        internals_.getUMatVector(internals);
+
+        CV_Assert(inputs.size() == 1 && outputs.size() == 1);
+        CV_Assert(inputs[0].total() == outputs[0].total());
+
+        const UMat& inp0 = inputs[0];
+        UMat& buffer = internals[0];
+        size_t num = inp0.size[0];
+        size_t channels = inp0.size[1];
+        size_t channelSize = inp0.total() / (num * channels);
+        for (size_t i = 0; i < num; ++i)
+        {
+            MatShape s = shape(channels, channelSize);
+            UMat src = inputs[i].reshape(1, s.size(), &s[0]);
+            UMat dst = outputs[i].reshape(1, s.size(), &s[0]);
+
+            UMat abs_mat;
+            absdiff(src, cv::Scalar::all(0), abs_mat);
+            pow(abs_mat, pnorm, buffer);
+
+            if (acrossSpatial)
+            {
+                // add eps to avoid overflow
+                float absSum = sum(buffer)[0] + epsilon;
+                float norm = pow(absSum, 1.0f / pnorm);
+                multiply(src, 1.0f / norm, dst);
+            }
+
+            if (!blobs.empty())
+            {
+                // scale the output
+                Mat scale = blobs[0];
+                if (scale.total() == 1)
+                {
+                    // _scale: 1 x 1
+                    multiply(dst, scale.at<float>(0, 0), dst);
+                }
+                else
+                {
+                    // _scale: _channels x 1
+                    CV_Assert(scale.total() == channels);
+                    repeat(scale, 1, dst.cols, buffer);
+                    multiply(dst, buffer, dst);
+                }
+            }
+        }
+        return true;
+    }
+#endif
+
     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
+        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
+                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
         Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
     }
 

modules/dnn/test/test_layers.cpp

@@ -320,6 +320,11 @@ TEST(Layer_Test_Eltwise, Accuracy)
     testLayerUsingCaffeModels("layer_eltwise");
 }
 
+OCL_TEST(Layer_Test_Eltwise, Accuracy)
+{
+    testLayerUsingCaffeModels("layer_eltwise", DNN_TARGET_OPENCL);
+}
+
 TEST(Layer_Test_PReLU, Accuracy)
 {
     testLayerUsingCaffeModels("layer_prelu", DNN_TARGET_CPU, true);

modules/dnn/test/test_tf_importer.cpp

@@ -76,7 +76,7 @@ static std::string path(const std::string& file)
     return findDataFile("dnn/tensorflow/" + file, false);
 }
 
-static void runTensorFlowNet(const std::string& prefix, bool hasText = false,
+static void runTensorFlowNet(const std::string& prefix, int targetId = DNN_TARGET_CPU, bool hasText = false,
                              double l1 = 1e-5, double lInf = 1e-4,
                              bool memoryLoad = false)
 {
@@ -104,6 +104,9 @@ static void runTensorFlowNet(const std::string& prefix, bool hasText = false,
 
     ASSERT_FALSE(net.empty());
 
+    net.setPreferableBackend(DNN_BACKEND_DEFAULT);
+    net.setPreferableTarget(targetId);
+
     cv::Mat input = blobFromNPY(inpPath);
     cv::Mat target = blobFromNPY(outPath);
 
@@ -132,6 +135,11 @@ TEST(Test_TensorFlow, eltwise_add_mul)
     runTensorFlowNet("eltwise_add_mul");
 }
 
+OCL_TEST(Test_TensorFlow, eltwise_add_mul)
+{
+    runTensorFlowNet("eltwise_add_mul", DNN_TARGET_OPENCL);
+}
+
 TEST(Test_TensorFlow, pad_and_concat)
 {
     runTensorFlowNet("pad_and_concat");
@@ -141,7 +149,14 @@ TEST(Test_TensorFlow, batch_norm)
 {
     runTensorFlowNet("batch_norm");
     runTensorFlowNet("fused_batch_norm");
-    runTensorFlowNet("batch_norm_text", true);
+    runTensorFlowNet("batch_norm_text", DNN_TARGET_CPU, true);
+}
+
+OCL_TEST(Test_TensorFlow, batch_norm)
+{
+    runTensorFlowNet("batch_norm", DNN_TARGET_OPENCL);
+    runTensorFlowNet("fused_batch_norm", DNN_TARGET_OPENCL);
+    runTensorFlowNet("batch_norm_text", DNN_TARGET_OPENCL, true);
 }
 
 TEST(Test_TensorFlow, pooling)
@@ -179,15 +194,15 @@ TEST(Test_TensorFlow, fp16)
 {
     const float l1 = 1e-3;
     const float lInf = 1e-2;
-    runTensorFlowNet("fp16_single_conv", false, l1, lInf);
+    runTensorFlowNet("fp16_single_conv", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_deconvolution", false, l1, lInf);
+    runTensorFlowNet("fp16_deconvolution", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_max_pool_odd_same", false, l1, lInf);
+    runTensorFlowNet("fp16_max_pool_odd_same", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_padding_valid", false, l1, lInf);
+    runTensorFlowNet("fp16_padding_valid", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_eltwise_add_mul", false, l1, lInf);
+    runTensorFlowNet("fp16_eltwise_add_mul", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_max_pool_odd_valid", false, l1, lInf);
+    runTensorFlowNet("fp16_max_pool_odd_valid", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_pad_and_concat", false, l1, lInf);
+    runTensorFlowNet("fp16_pad_and_concat", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_max_pool_even", false, l1, lInf);
+    runTensorFlowNet("fp16_max_pool_even", DNN_TARGET_CPU, false, l1, lInf);
-    runTensorFlowNet("fp16_padding_same", false, l1, lInf);
+    runTensorFlowNet("fp16_padding_same", DNN_TARGET_CPU, false, l1, lInf);
 }
 
 TEST(Test_TensorFlow, quantized)
@@ -267,7 +282,7 @@ OCL_TEST(Test_TensorFlow, MobileNet_SSD)
 
 TEST(Test_TensorFlow, lstm)
 {
-    runTensorFlowNet("lstm", true);
+    runTensorFlowNet("lstm", DNN_TARGET_CPU, true);
 }
 
 TEST(Test_TensorFlow, split)
@@ -284,11 +299,11 @@ TEST(Test_TensorFlow, memory_read)
 {
     double l1 = 1e-5;
     double lInf = 1e-4;
-    runTensorFlowNet("lstm", true, l1, lInf, true);
+    runTensorFlowNet("lstm", DNN_TARGET_CPU, true, l1, lInf, true);
 
-    runTensorFlowNet("batch_norm", false, l1, lInf, true);
+    runTensorFlowNet("batch_norm", DNN_TARGET_CPU, false, l1, lInf, true);
-    runTensorFlowNet("fused_batch_norm", false, l1, lInf, true);
+    runTensorFlowNet("fused_batch_norm", DNN_TARGET_CPU, false, l1, lInf, true);
-    runTensorFlowNet("batch_norm_text", true, l1, lInf, true);
+    runTensorFlowNet("batch_norm_text", DNN_TARGET_CPU, true, l1, lInf, true);
 }
 
 }

modules/dnn/test/test_torch_importer.cpp

@@ -170,6 +170,11 @@ TEST(Torch_Importer, run_batch_norm)
     runTorchNet("net_batch_norm", DNN_TARGET_CPU, "", false, true);
 }
 
+OCL_TEST(Torch_Importer, run_batch_norm)
+{
+    runTorchNet("net_batch_norm", DNN_TARGET_OPENCL, "", false, true);
+}
+
 TEST(Torch_Importer, net_prelu)
 {
     runTorchNet("net_prelu");
@@ -225,6 +230,11 @@ TEST(Torch_Importer, net_normalize)
     runTorchNet("net_normalize", DNN_TARGET_CPU, "", false, true);
 }
 
+OCL_TEST(Torch_Importer, net_normalize)
+{
+    runTorchNet("net_normalize", DNN_TARGET_OPENCL, "", false, true);
+}
+
 TEST(Torch_Importer, net_padding)
 {
     runTorchNet("net_padding", DNN_TARGET_CPU, "", false, true);
@@ -237,6 +247,11 @@ TEST(Torch_Importer, net_non_spatial)
     runTorchNet("net_non_spatial", DNN_TARGET_CPU, "", false, true);
 }
 
+OCL_TEST(Torch_Importer, net_non_spatial)
+{
+    runTorchNet("net_non_spatial", DNN_TARGET_OPENCL, "", false, true);
+}
+
 TEST(Torch_Importer, ENet_accuracy)
 {
     Net net;