Added Halide backend support for deep learning layers

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

@@ -55,6 +55,23 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
 
     typedef std::vector<int> MatShape;
 
+    /**
+     * @brief Enum of computation backends supported by layers.
+     */
+    enum Backend
+    {
+        DNN_BACKEND_DEFAULT,
+        DNN_BACKEND_HALIDE
+    };
+
+    /**
+     * @brief Enum of target devices for computations.
+     */
+    enum Target
+    {
+        DNN_TARGET_CPU
+    };
+
     /** @brief Initialize dnn module and built-in layers.
      *
      * This function automatically called on most of OpenCV builds,
@@ -77,6 +94,54 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
         String type; //!< Type name which was used for creating layer by layer factory (optional).
     };
 
+   /**
+    * @brief Derivatives of this class encapsulates functions of certain backends.
+    */
+    class BackendNode
+    {
+    public:
+        BackendNode(int backendId);
+
+        virtual ~BackendNode(); //!< Virtual destructor to make polymorphism.
+
+        int backendId; //!< Backend identifier.
+    };
+
+    /**
+     * @brief Derivatives of this class wraps cv::Mat for different backends and targets.
+     */
+    class BackendWrapper
+    {
+    public:
+        BackendWrapper(int backendId, int targetId);
+
+        /**
+         * @brief Wrap cv::Mat for specific backend and target.
+         * @param[in] targetId Target identifier.
+         * @param[in] m cv::Mat for wrapping.
+         *
+         * Make CPU->GPU data transfer if it's require for the target.
+         */
+        BackendWrapper(int targetId, const cv::Mat& m);
+
+        /**
+         * @brief Make wrapper for reused cv::Mat.
+         * @param[in] base Wrapper of cv::Mat that will be reused.
+         * @param[in] shape Specific shape.
+         *
+         * Initialize wrapper from another one. It'll wrap the same host CPU
+         * memory and mustn't allocate memory on device(i.e. GPU). It might
+         * has different shape. Use in case of CPU memory reusing for reuse
+         * associented memory on device too.
+         */
+        BackendWrapper(const Ptr<BackendWrapper>& base, const MatShape& shape);
+
+        virtual ~BackendWrapper(); //!< Virtual destructor to make polymorphism.
+
+        int backendId;  //!< Backend identifier.
+        int targetId;   //!< Target identifier.
+    };
+
     /** @brief This interface class allows to build new Layers - are building blocks of networks.
      *
      * Each class, derived from Layer, must implement allocate() methods to declare own outputs and forward() to compute outputs.
@@ -131,6 +196,50 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
          */
         virtual int outputNameToIndex(String outputName);
 
+        /**
+         * @brief Ask layer if it support specific backend for doing computations.
+         * @param[in] backendId computation backend identifier.
+         * @see Backend
+         */
+        virtual bool supportBackend(int backendId);
+
+        /**
+         * @brief Returns Halide backend node.
+         * @param[in] inputs Input Halide buffers.
+         * @see BackendNode, BackendWrapper
+         *
+         * Input buffers should be exactly the same that will be used in forward invocations.
+         * Despite we can use Halide::ImageParam based on input shape only,
+         * it helps prevent some memory management issues (if something wrong,
+         * Halide tests will be failed).
+         */
+        virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs);
+
+       /**
+        * @brief Automatic Halide scheduling based on layer hyper-parameters.
+        * @param[in] node Backend node with Halide functions.
+        * @param[in] inputs Blobs that will be used in forward invocations.
+        * @param[in] outputs Blobs that will be used in forward invocations.
+        * @see BackendNode
+        *
+        * Layer don't use own Halide::Func members because we can have applied
+        * layers fusing. In this way the fused function should be scheduled.
+        */
+        virtual void applyHalideScheduler(Ptr<BackendNode>& node,
+                                          const std::vector<Mat*> &inputs,
+                                          const std::vector<Mat> &outputs) const;
+
+        /**
+         * @brief Implement layers fusing.
+         * @param[in] node Backend node of bottom layer.
+         * @see BackendNode
+         *
+         * Actual for graph-based backends. If layer attached successfully,
+         * returns non-empty cv::Ptr to node of the same backend.
+         * Fuse only over the last function.
+         */
+        virtual Ptr<BackendNode> tryAttach(const Ptr<BackendNode>& node);
+
         virtual bool getMemoryShapes(const std::vector<MatShape> &inputs,
                                      const int requiredOutputs,
                                      std::vector<MatShape> &outputs,
@@ -251,6 +360,24 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
         /** @overload */
         void forwardOpt(const std::vector<LayerId> &toLayers);
 
+        /**
+         * @brief Compile Halide layers.
+         * @param[in] scheduler Path to YAML file with scheduling directives.
+         * @see setPreferableBackend
+         *
+         * Schedule layers that support Halide backend. Then compile them for
+         * specific target. For layers that not represented in scheduling file
+         * or if no manual scheduling used at all, automatic scheduling will be applied.
+         */
+        void compileHalide(const std::string& scheduler = "");
+
+        /**
+         * @brief Ask network to use specific computation backend where it supported.
+         * @param[in] backendId backend identifier.
+         * @see Backend
+         */
+        void setPreferableBackend(int backendId);
+
         /** @brief Sets the new value for the layer output blob
          *  @param outputName descriptor of the updating layer output blob.
          *  @param blob new blob.

modules/dnn/perf/perf_halide_net.cpp

@@ -0,0 +1,141 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+//
+// Copyright (C) 2017, Intel Corporation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+namespace cvtest
+{
+
+#ifdef HAVE_HALIDE
+using namespace cv;
+using namespace dnn;
+
+static void loadNet(const std::string& weights, const std::string& proto,
+                    const std::string& scheduler, int inWidth, int inHeight,
+                    const std::string& outputLayer, const std::string& framework,
+                    int targetId, Net* net, int* outputLayerId)
+{
+    Mat input(inHeight, inWidth, CV_32FC3);
+    randu(input, 0.0f, 1.0f);
+
+    if (framework == "caffe")
+    {
+        *net = cv::dnn::readNetFromCaffe(proto, weights);
+    }
+    else if (framework == "torch")
+    {
+        *net = cv::dnn::readNetFromTorch(weights);
+    }
+    else if (framework == "tensorflow")
+    {
+        *net = cv::dnn::readNetFromTensorflow(weights);
+    }
+    else
+        CV_Error(Error::StsNotImplemented, "Unknown framework " + framework);
+
+    net->setBlob("", cv::dnn::blobFromImage(input, 1.0, false));
+    net->setPreferableBackend(DNN_BACKEND_HALIDE);
+    net->compileHalide(scheduler);
+    *outputLayerId = net->getLayerId(outputLayer);
+    net->forward(*outputLayerId);
+}
+
+PERF_TEST(GoogLeNet, HalidePerfTest)
+{
+    Net net;
+    int outputLayerId;
+    loadNet(findDataFile("dnn/bvlc_googlenet.caffemodel"),
+            findDataFile("dnn/bvlc_googlenet.prototxt"),
+            "", 227, 227, "prob", "caffe", DNN_TARGET_CPU, &net, &outputLayerId);
+
+    TEST_CYCLE_N(10)
+    {
+        net.forward(outputLayerId);
+    }
+    SANITY_CHECK_NOTHING();
+}
+
+PERF_TEST(AlexNet, HalidePerfTest)
+{
+    Net net;
+    int outputLayerId;
+    loadNet(findDataFile("dnn/bvlc_alexnet.caffemodel"),
+            findDataFile("dnn/bvlc_alexnet.prototxt"),
+            findDataFile("dnn/halide_scheduler_alexnet.yml"),
+            227, 227, "prob", "caffe", DNN_TARGET_CPU, &net, &outputLayerId);
+
+    TEST_CYCLE_N(10)
+    {
+        net.forward(outputLayerId);
+    }
+    SANITY_CHECK_NOTHING();
+}
+
+// PERF_TEST(ResNet50, HalidePerfTest)
+// {
+//     Net net;
+//     int outputLayerId;
+//     loadNet(findDataFile("dnn/ResNet-50-model.caffemodel"),
+//             findDataFile("dnn/ResNet-50-deploy.prototxt"),
+//             findDataFile("dnn/halide_scheduler_resnet_50.yml"),
+//             224, 224, "prob", "caffe", DNN_TARGET_CPU, &net, &outputLayerId);
+//
+//     TEST_CYCLE_N(10)
+//     {
+//         net.forward(outputLayerId);
+//     }
+//     SANITY_CHECK_NOTHING();
+// }
+
+// PERF_TEST(SqueezeNet_v1_1, HalidePerfTest)
+// {
+//     Net net;
+//     int outputLayerId;
+//     loadNet(findDataFile("dnn/squeezenet_v1_1.caffemodel"),
+//             findDataFile("dnn/squeezenet_v1_1.prototxt"),
+//             findDataFile("dnn/halide_scheduler_squeezenet_v1_1.yml"),
+//             227, 227, "prob", "caffe", DNN_TARGET_CPU, &net, &outputLayerId);
+//
+//     TEST_CYCLE_N(10)
+//     {
+//         net.forward(outputLayerId);
+//     }
+//     SANITY_CHECK_NOTHING();
+// }
+
+PERF_TEST(Inception_5h, HalidePerfTest)
+{
+    Net net;
+    int outputLayerId;
+    loadNet(findDataFile("dnn/tensorflow_inception_graph.pb"), "",
+            findDataFile("dnn/halide_scheduler_inception_5h.yml"),
+            224, 224, "softmax2", "tensorflow", DNN_TARGET_CPU,
+            &net, &outputLayerId);
+
+    TEST_CYCLE_N(10)
+    {
+        net.forward(outputLayerId);
+    }
+    SANITY_CHECK_NOTHING();
+}
+
+PERF_TEST(ENet, HalidePerfTest)
+{
+    Net net;
+    int outputLayerId;
+    loadNet(findDataFile("dnn/Enet-model-best.net"), "",
+            findDataFile("dnn/halide_scheduler_enet.yml"),
+            512, 256, "l367_Deconvolution", "torch", DNN_TARGET_CPU,
+            &net, &outputLayerId);
+
+    TEST_CYCLE_N(10)
+    {
+        net.forward(outputLayerId);
+    }
+    SANITY_CHECK_NOTHING();
+}
+#endif  // HAVE_HALIDE
+
+}  // namespace cvtest

modules/dnn/perf/perf_main.cpp

@@ -1,3 +1,12 @@
 #include "perf_precomp.hpp"
 
-CV_PERF_TEST_MAIN(dnn)
+static const char* extraTestDataPath =
+#ifdef WINRT
+        NULL;
+#else
+        getenv("OPENCV_DNN_TEST_DATA_PATH");
+#endif
+
+CV_PERF_TEST_MAIN(dnn,
+    extraTestDataPath ? (void)cvtest::addDataSearchPath(extraTestDataPath) : (void)0
+)