Merge pull request #19632 from l-bat:lb/ie_arm_target

Added OpenVINO ARM target

* Added IE ARM target

* Added OpenVINO ARM target

* Delete ARM target

* Detect ARM platform

* Changed device name in ArmPlugin

* Change ARM detection

modules/dnn/include/opencv2/dnn/utils/inference_engine.hpp

@@ -49,6 +49,8 @@ CV_EXPORTS_W void resetMyriadDevice();
 #define CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_2 "Myriad2"
 /// Intel(R) Neural Compute Stick 2, NCS2 (USB 03e7:2485), MyriadX (https://software.intel.com/ru-ru/neural-compute-stick)
 #define CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X "MyriadX"
+#define CV_DNN_INFERENCE_ENGINE_CPU_TYPE_ARM_COMPUTE "ARM_COMPUTE"
+#define CV_DNN_INFERENCE_ENGINE_CPU_TYPE_X86         "X86"
 
 
 /** @brief Returns Inference Engine VPU type.
@@ -57,6 +59,11 @@ CV_EXPORTS_W void resetMyriadDevice();
  */
 CV_EXPORTS_W cv::String getInferenceEngineVPUType();
 
+/** @brief Returns Inference Engine CPU type.
+ *
+ * Specify OpenVINO plugin: CPU or ARM.
+ */
+CV_EXPORTS_W cv::String getInferenceEngineCPUType();
 
 CV__DNN_EXPERIMENTAL_NS_END
 }} // namespace

modules/dnn/src/dnn.cpp

@@ -1286,17 +1286,19 @@ struct Net::Impl : public detail::NetImplBase
         CV_Assert(preferableBackend != DNN_BACKEND_HALIDE ||
                   preferableTarget == DNN_TARGET_CPU ||
                   preferableTarget == DNN_TARGET_OPENCL);
+#ifdef HAVE_INF_ENGINE
         if (preferableBackend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
             preferableBackend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
         {
             CV_Assert(
-                  preferableTarget == DNN_TARGET_CPU ||
+                  (preferableTarget == DNN_TARGET_CPU && (!isArmComputePlugin() || preferableBackend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)) ||
                   preferableTarget == DNN_TARGET_OPENCL ||
                   preferableTarget == DNN_TARGET_OPENCL_FP16 ||
                   preferableTarget == DNN_TARGET_MYRIAD ||
                   preferableTarget == DNN_TARGET_FPGA
             );
         }
+#endif
         if (!netWasAllocated || this->blobsToKeep != blobsToKeep_)
         {
             if (preferableBackend == DNN_BACKEND_OPENCV && IS_DNN_OPENCL_TARGET(preferableTarget))
@@ -1972,8 +1974,8 @@ struct Net::Impl : public detail::NetImplBase
             return;
         }
 
-        bool supportsCPUFallback = preferableTarget == DNN_TARGET_CPU ||
-                                   BackendRegistry::checkIETarget(DNN_TARGET_CPU);
+        bool supportsCPUFallback = !isArmComputePlugin() && (preferableTarget == DNN_TARGET_CPU ||
+                                   BackendRegistry::checkIETarget(DNN_TARGET_CPU));
 
         // Build Inference Engine networks from sets of layers that support this
         // backend. Split a whole model on several Inference Engine networks if

modules/dnn/src/layers/batch_norm_layer.cpp

@@ -382,7 +382,11 @@ public:
         shape[1] = weights_.total();
         auto weight = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), weights_.data);
         auto bias = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), bias_.data);
+#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2021_2)
+        auto scale_node = std::make_shared<ngraph::op::v1::Multiply>(ieInpNode, weight, ngraph::op::AutoBroadcastType::NUMPY);
+#else
         auto scale_node = std::make_shared<ngraph::op::v0::Multiply>(ieInpNode, weight, ngraph::op::AutoBroadcastType::NUMPY);
+#endif
         auto scale_shift = std::make_shared<ngraph::op::v1::Add>(scale_node, bias, ngraph::op::AutoBroadcastType::NUMPY);
         return Ptr<BackendNode>(new InfEngineNgraphNode(scale_shift));
     }

modules/dnn/src/layers/convolution_layer.cpp

@@ -273,10 +273,13 @@ public:
 #ifdef HAVE_INF_ENGINE
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
         {
-            if (ksize == 1)
+            bool isArmTarget = preferableTarget == DNN_TARGET_CPU && isArmComputePlugin();
+            if (isArmTarget && blobs.empty())
                 return false;
+            if (ksize == 1)
+                return isArmTarget;
             if (ksize == 3)
-                return preferableTarget == DNN_TARGET_CPU;
+                return preferableTarget != DNN_TARGET_MYRIAD && !isArmTarget;
             if ((backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || preferableTarget != DNN_TARGET_MYRIAD) && blobs.empty())
                 return false;
             return (preferableTarget != DNN_TARGET_MYRIAD || dilation.width == dilation.height);
@@ -578,7 +581,7 @@ public:
         CV_Assert_N(inputs.size() >= 1, nodes.size() >= 1);
         auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
         std::vector<size_t> dims = ieInpNode->get_shape();
-        CV_Assert(dims.size() == 4 || dims.size() == 5);
+        CV_Check(dims.size(), dims.size() >= 3 && dims.size() <= 5, "");
         std::shared_ptr<ngraph::Node> ieWeights = nodes.size() > 1 ? nodes[1].dynamicCast<InfEngineNgraphNode>()->node : nullptr;
         if (nodes.size() > 1)
             CV_Assert(ieWeights);  // dynamic_cast should not fail
@@ -616,7 +619,7 @@ public:
         else
         {
             auto shape = std::make_shared<ngraph::op::Constant>(ngraph::element::i64,
-                             ngraph::Shape{kernel_shape.size()}, kernel_shape.data());
+                             ngraph::Shape{kernel_shape.size()}, std::vector<int64_t>(kernel_shape.begin(), kernel_shape.end()));
             ieWeights  = std::make_shared<ngraph::op::v1::Reshape>(ieWeights, shape, true);
         }
 
@@ -651,7 +654,7 @@ public:
             if (nodes.size() == 3)
             {
                 auto bias_shape = std::make_shared<ngraph::op::Constant>(ngraph::element::i64,
-                                    ngraph::Shape{shape.size()}, shape.data());
+                                    ngraph::Shape{shape.size()}, std::vector<int64_t>(shape.begin(), shape.end()));
                 bias = std::make_shared<ngraph::op::v1::Reshape>(nodes[2].dynamicCast<InfEngineNgraphNode>()->node, bias_shape, true);
             }
             else

modules/dnn/src/layers/elementwise_layers.cpp

@@ -1164,11 +1164,15 @@ struct PowerFunctor : public BaseFunctor
                                                                  ngraph::Shape{1}, &scale);
         auto shift_node = std::make_shared<ngraph::op::Constant>(ngraph::element::f32,
                                                                  ngraph::Shape{1}, &shift);
-        auto power_node = std::make_shared<ngraph::op::Constant>(ngraph::element::f32,
-                                                                 ngraph::Shape{1}, &power);
 
         auto mul = std::make_shared<ngraph::op::v1::Multiply>(scale_node, node, ngraph::op::AutoBroadcastType::NUMPY);
         auto scale_shift = std::make_shared<ngraph::op::v1::Add>(mul, shift_node, ngraph::op::AutoBroadcastType::NUMPY);
+
+        if (power == 1)
+            return scale_shift;
+
+        auto power_node = std::make_shared<ngraph::op::Constant>(ngraph::element::f32,
+                                                                 ngraph::Shape{1}, &power);
         return std::make_shared<ngraph::op::v1::Power>(scale_shift, power_node, ngraph::op::AutoBroadcastType::NUMPY);
     }
 #endif  // HAVE_DNN_NGRAPH

modules/dnn/src/layers/normalize_bbox_layer.cpp

@@ -324,8 +324,8 @@ public:
         if (!acrossSpatial) {
             axes_data.push_back(1);
         } else {
-            axes_data.resize(ieInpNode->get_shape().size());
-            std::iota(axes_data.begin(), axes_data.end(), 0);
+            axes_data.resize(ieInpNode->get_shape().size() - 1);
+            std::iota(axes_data.begin(), axes_data.end(), 1);
         }
         auto axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes_data.size()}, axes_data);
         auto norm = std::make_shared<ngraph::op::NormalizeL2>(ieInpNode, axes, epsilon, ngraph::op::EpsMode::ADD);
@@ -334,19 +334,18 @@ public:
         std::vector<size_t> shape(ieInpNode->get_shape().size(), 1);
         shape[0] = blobs.empty() ? 1 : batch;
         shape[1] = numChannels;
-        std::shared_ptr<ngraph::op::Constant> weight;
-        if (blobs.empty())
+        if (!blobs.empty())
         {
-            std::vector<float> ones(numChannels, 1);
-            weight = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), ones.data());
-        }
-        else
-        {
-            weight = std::make_shared<ngraph::op::Constant>(
+            auto weight = std::make_shared<ngraph::op::Constant>(
                                       ngraph::element::f32, ngraph::Shape(shape), blobs[0].data);
+#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2021_2)
+            auto mul = std::make_shared<ngraph::op::v1::Multiply>(norm, weight, ngraph::op::AutoBroadcastType::NUMPY);
+#else
+            auto mul = std::make_shared<ngraph::op::v0::Multiply>(norm, weight, ngraph::op::AutoBroadcastType::NUMPY);
+#endif
+            return Ptr<BackendNode>(new InfEngineNgraphNode(mul));
         }
-        auto mul = std::make_shared<ngraph::op::v0::Multiply>(norm, weight, ngraph::op::AutoBroadcastType::NUMPY);
-        return Ptr<BackendNode>(new InfEngineNgraphNode(mul));
+        return Ptr<BackendNode>(new InfEngineNgraphNode(norm));
     }
 #endif  // HAVE_DNN_NGRAPH
 

modules/dnn/src/layers/padding_layer.cpp

@@ -97,9 +97,12 @@ public:
     {
 #ifdef HAVE_INF_ENGINE
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
-            return INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R1) &&
-                   (preferableTarget != DNN_TARGET_MYRIAD ||
-                    (dstRanges.size() == 4 && paddings[0].first == 0 && paddings[0].second == 0));
+        {
+            if (INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R1) && preferableTarget == DNN_TARGET_MYRIAD)
+                return dstRanges.size() == 4 && paddings[0].first == 0 && paddings[0].second == 0;
+
+            return (dstRanges.size() <= 4 || !isArmComputePlugin());
+        }
 #endif
         return backendId == DNN_BACKEND_OPENCV ||
                (backendId == DNN_BACKEND_HALIDE && haveHalide() && dstRanges.size() == 4);

modules/dnn/src/layers/permute_layer.cpp

@@ -105,6 +105,10 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
+#ifdef HAVE_INF_ENGINE
+        if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && preferableTarget == DNN_TARGET_CPU)
+            return _order.size() <= 4 || !isArmComputePlugin();
+#endif
         return backendId == DNN_BACKEND_OPENCV ||
                ((backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && haveInfEngine());
     }

modules/dnn/src/layers/pooling_layer.cpp

@@ -205,7 +205,9 @@ public:
 #endif
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
         {
-            return !computeMaxIdx && type != STOCHASTIC && kernel_size.size() > 1;
+#ifdef HAVE_DNN_NGRAPH
+            return !computeMaxIdx && type != STOCHASTIC && kernel_size.size() > 1 && (kernel_size.size() != 3 || !isArmComputePlugin());
+#endif
         }
         else if (backendId == DNN_BACKEND_OPENCV)
         {

modules/dnn/src/layers/region_layer.cpp

@@ -393,8 +393,10 @@ public:
             std::vector<int64_t> mask(anchors, 1);
             region = std::make_shared<ngraph::op::RegionYolo>(tr_input, coords, classes, anchors, useSoftmax, mask, 1, 3, anchors_vec);
 
+            auto tr_shape = tr_input->get_shape();
             auto shape_as_inp = std::make_shared<ngraph::op::Constant>(ngraph::element::i64,
-                                                                       ngraph::Shape{tr_input->get_shape().size()}, tr_input->get_shape().data());
+                                                                       ngraph::Shape{tr_shape.size()},
+                                                                       std::vector<int64_t>(tr_shape.begin(), tr_shape.end()));
 
             region = std::make_shared<ngraph::op::v1::Reshape>(region, shape_as_inp, true);
             new_axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{4}, std::vector<int64_t>{0, 2, 3, 1});
@@ -540,7 +542,7 @@ public:
         result = std::make_shared<ngraph::op::Transpose>(result, tr_axes);
         if (b > 1)
         {
-            std::vector<size_t> sizes = {(size_t)b, result->get_shape()[0] / b, result->get_shape()[1]};
+            std::vector<int64_t> sizes{b, static_cast<int64_t>(result->get_shape()[0]) / b, static_cast<int64_t>(result->get_shape()[1])};
             auto shape_node = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{sizes.size()}, sizes.data());
             result = std::make_shared<ngraph::op::v1::Reshape>(result, shape_node, true);
         }

modules/dnn/src/layers/scale_layer.cpp

@@ -249,7 +249,11 @@ public:
             auto weight = blobs.empty() ? ieInpNode1 :
                           std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), blobs[0].data);
 
-            node = std::make_shared<ngraph::op::v0::Multiply>(node, weight, ngraph::op::AutoBroadcastType::NUMPY);
+#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2021_2)
+        node = std::make_shared<ngraph::op::v1::Multiply>(node, weight, ngraph::op::AutoBroadcastType::NUMPY);
+#else
+        node = std::make_shared<ngraph::op::v0::Multiply>(node, weight, ngraph::op::AutoBroadcastType::NUMPY);
+#endif
         }
         if (hasBias || !hasWeights)
         {

modules/dnn/src/op_inf_engine.cpp

@@ -651,6 +651,22 @@ InferenceEngine::Core& getCore(const std::string& id)
 }
 #endif
 
+static bool detectArmPlugin_()
+{
+    InferenceEngine::Core& ie = getCore("CPU");
+    const std::vector<std::string> devices = ie.GetAvailableDevices();
+    for (std::vector<std::string>::const_iterator i = devices.begin(); i != devices.end(); ++i)
+    {
+        if (i->find("CPU") != std::string::npos)
+        {
+            const std::string name = ie.GetMetric(*i, METRIC_KEY(FULL_DEVICE_NAME)).as<std::string>();
+            CV_LOG_INFO(NULL, "CPU plugin: " << name);
+            return name.find("arm_compute::NEON") != std::string::npos;
+        }
+    }
+    return false;
+}
+
 #if !defined(OPENCV_DNN_IE_VPU_TYPE_DEFAULT)
 static bool detectMyriadX_()
 {
@@ -1162,6 +1178,12 @@ bool isMyriadX()
      return myriadX;
 }
 
+bool isArmComputePlugin()
+{
+    static bool armPlugin = getInferenceEngineCPUType() == CV_DNN_INFERENCE_ENGINE_CPU_TYPE_ARM_COMPUTE;
+    return armPlugin;
+}
+
 static std::string getInferenceEngineVPUType_()
 {
     static std::string param_vpu_type = utils::getConfigurationParameterString("OPENCV_DNN_IE_VPU_TYPE", "");
@@ -1199,6 +1221,14 @@ cv::String getInferenceEngineVPUType()
     return vpu_type;
 }
 
+cv::String getInferenceEngineCPUType()
+{
+    static cv::String cpu_type = detectArmPlugin_() ?
+                                 CV_DNN_INFERENCE_ENGINE_CPU_TYPE_ARM_COMPUTE :
+                                 CV_DNN_INFERENCE_ENGINE_CPU_TYPE_X86;
+    return cpu_type;
+}
+
 #else  // HAVE_INF_ENGINE
 
 cv::String getInferenceEngineBackendType()
@@ -1214,6 +1244,11 @@ cv::String getInferenceEngineVPUType()
 {
     CV_Error(Error::StsNotImplemented, "This OpenCV build doesn't include InferenceEngine support");
 }
+
+cv::String getInferenceEngineCPUType()
+{
+    CV_Error(Error::StsNotImplemented, "This OpenCV build doesn't include InferenceEngine support");
+}
 #endif  // HAVE_INF_ENGINE
 
 

modules/dnn/src/op_inf_engine.hpp

@@ -254,6 +254,8 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
 
 bool isMyriadX();
 
+bool isArmComputePlugin();
+
 CV__DNN_EXPERIMENTAL_NS_END
 
 InferenceEngine::Core& getCore(const std::string& id);

modules/dnn/test/test_common.hpp

@@ -35,6 +35,7 @@
 #define CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_2         "dnn_skip_ie_myriad2"
 #define CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X         "dnn_skip_ie_myriadx"
 #define CV_TEST_TAG_DNN_SKIP_IE_MYRIAD           CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_2, CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X
+#define CV_TEST_TAG_DNN_SKIP_IE_ARM_CPU          "dnn_skip_ie_arm_cpu"
 
 
 #ifdef HAVE_INF_ENGINE

modules/dnn/test/test_onnx_importer.cpp

@@ -144,6 +144,10 @@ TEST_P(Test_ONNX_layers, Convolution_variable_weight_bias)
          backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019) && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
 
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_CPU &&
+        getInferenceEngineCPUType() == CV_DNN_INFERENCE_ENGINE_CPU_TYPE_ARM_COMPUTE)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_ARM_CPU, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+
     String basename = "conv_variable_wb";
     Net net = readNetFromONNX(_tf("models/" + basename + ".onnx"));
     ASSERT_FALSE(net.empty());
@@ -717,6 +721,8 @@ TEST_P(Test_ONNX_layers, Conv1d_variable_weight_bias)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
     {
         if (target == DNN_TARGET_MYRIAD) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+        if (target == DNN_TARGET_CPU && getInferenceEngineCPUType() == CV_DNN_INFERENCE_ENGINE_CPU_TYPE_ARM_COMPUTE)
+            applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_ARM_CPU, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
     }
     String basename = "conv1d_variable_wb";
     Net net = readNetFromONNX(_tf("models/" + basename + ".onnx"));