Merge pull request #24045 from TolyaTalamanov:at/add-onnx-directml-execution-provider

G-API: Support DirectML Execution Provider for ONNXRT Backend #24045

### Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [ ] I agree to contribute to the project under Apache 2 License.
- [ ] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [ ] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake

cmake/FindONNX.cmake

@@ -16,7 +16,22 @@ if(ONNXRT_ROOT_DIR)
     CMAKE_FIND_ROOT_PATH_BOTH)
 endif()
 
+macro(detect_onxxrt_ep filename dir have_ep_var)
+    find_path(ORT_EP_INCLUDE ${filename} ${dir} CMAKE_FIND_ROOT_PATH_BOTH)
+    if(ORT_EP_INCLUDE)
+       set(${have_ep_var} TRUE)
+    endif()
+endmacro()
+
 if(ORT_LIB AND ORT_INCLUDE)
+  # Check DirectML Execution Provider availability
+  get_filename_component(dml_dir ${ONNXRT_ROOT_DIR}/include/onnxruntime/core/providers/dml ABSOLUTE)
+  detect_onxxrt_ep(
+      dml_provider_factory.h
+      ${dml_dir}
+      HAVE_ONNX_DML
+  )
+
   set(HAVE_ONNX TRUE)
   # For CMake output only
   set(ONNX_LIBRARIES "${ORT_LIB}" CACHE STRING "ONNX Runtime libraries")

modules/gapi/CMakeLists.txt

@@ -162,6 +162,7 @@ set(gapi_srcs
 
     # ONNX backend
     src/backends/onnx/gonnxbackend.cpp
+    src/backends/onnx/dml_ep.cpp
 
     # Render backend
     src/backends/render/grenderocv.cpp
@@ -366,6 +367,9 @@ endif()
 if(HAVE_ONNX)
   ocv_target_link_libraries(${the_module} PRIVATE ${ONNX_LIBRARY})
   ocv_target_compile_definitions(${the_module} PRIVATE HAVE_ONNX=1)
+  if(HAVE_ONNX_DML)
+    ocv_target_compile_definitions(${the_module} PRIVATE HAVE_ONNX_DML=1)
+  endif()
   if(TARGET opencv_test_gapi)
     ocv_target_compile_definitions(opencv_test_gapi PRIVATE HAVE_ONNX=1)
     ocv_target_link_libraries(opencv_test_gapi PRIVATE ${ONNX_LIBRARY})

modules/gapi/include/opencv2/gapi/infer/bindings_onnx.hpp

@@ -34,7 +34,10 @@ public:
     PyParams& cfgNormalize(const std::string &layer_name, bool flag);
 
     GAPI_WRAP
-    PyParams& cfgExecutionProvider(ep::OpenVINO ov_ep);
+    PyParams& cfgAddExecutionProvider(ep::OpenVINO ep);
+
+    GAPI_WRAP
+    PyParams& cfgAddExecutionProvider(ep::DirectML ep);
 
     GAPI_WRAP
     PyParams& cfgDisableMemPattern();

modules/gapi/include/opencv2/gapi/infer/onnx.hpp

@@ -45,28 +45,13 @@ struct GAPI_EXPORTS_W_SIMPLE OpenVINO {
 
     /** @brief Class constructor.
 
-    Constructs OpenVINO parameters based on device information.
+    Constructs OpenVINO parameters based on device type information.
 
-    @param device Target device to use.
+    @param dev_type Target device type to use. ("CPU_FP32", "GPU_FP16", etc)
     */
     GAPI_WRAP
-    OpenVINO(const std::string &device)
-        : device_id(device) {
-    }
-
-    /** @brief Specifies OpenVINO Execution Provider device type.
-
-    This function is used to override the accelerator hardware type
-    and precision at runtime. If this option is not explicitly configured, default
-    hardware and precision specified during onnxruntime build time is used.
-
-    @param type Device type ("CPU_FP32", "GPU_FP16", etc)
-    @return reference to this parameter structure.
-    */
-    GAPI_WRAP
-    OpenVINO& cfgDeviceType(const std::string &type) {
-        device_type = cv::util::make_optional(type);
-        return *this;
+    explicit OpenVINO(const std::string &dev_type)
+        : device_type(dev_type) {
     }
 
     /** @brief Specifies OpenVINO Execution Provider cache dir.
@@ -86,15 +71,14 @@ struct GAPI_EXPORTS_W_SIMPLE OpenVINO {
     /** @brief Specifies OpenVINO Execution Provider number of threads.
 
     This function is used to override the accelerator default value
-    of number of threads with this value at runtime. If this option
-    is not explicitly set, default value of 8 is used during build time.
+    of number of threads with this value at runtime.
 
     @param nthreads Number of threads.
     @return reference to this parameter structure.
     */
     GAPI_WRAP
     OpenVINO& cfgNumThreads(size_t nthreads) {
-        num_of_threads = cv::util::make_optional(nthreads);
+        num_of_threads = nthreads;
         return *this;
     }
 
@@ -127,15 +111,39 @@ struct GAPI_EXPORTS_W_SIMPLE OpenVINO {
         return *this;
     }
 
-    std::string device_id;
+    std::string device_type;
     std::string cache_dir;
-    cv::optional<std::string> device_type;
-    cv::optional<size_t> num_of_threads;
+    size_t num_of_threads = 0;
     bool enable_opencl_throttling = false;
     bool enable_dynamic_shapes = false;
 };
 
-using EP = cv::util::variant<cv::util::monostate, OpenVINO>;
+/**
+ * @brief This structure provides functions
+ * that fill inference options for ONNX DirectML Execution Provider.
+ * Please follow https://onnxruntime.ai/docs/execution-providers/DirectML-ExecutionProvider.html#directml-execution-provider
+ */
+class GAPI_EXPORTS_W_SIMPLE DirectML {
+public:
+    // NB: Used from python.
+    /// @private -- Exclude this constructor from OpenCV documentation
+    GAPI_WRAP
+    DirectML() = default;
+
+    /** @brief Class constructor.
+
+    Constructs DirectML parameters based on device id.
+
+    @param device_id Target device id to use. ("0", "1", etc)
+    */
+    GAPI_WRAP
+    explicit DirectML(const int device_id) : ddesc(device_id) { };
+
+    using DeviceDesc = cv::util::variant<int>;
+    DeviceDesc ddesc;
+};
+
+using EP = cv::util::variant<cv::util::monostate, OpenVINO, DirectML>;
 
 } // namespace ep
 
@@ -191,7 +199,7 @@ struct ParamDesc {
     std::unordered_map<std::string, std::pair<cv::Scalar, cv::Scalar> > generic_mstd;
     std::unordered_map<std::string, bool> generic_norm;
 
-    cv::gapi::onnx::ep::EP execution_provider;
+    std::vector<cv::gapi::onnx::ep::EP> execution_providers;
     bool disable_mem_pattern;
 };
 } // namespace detail
@@ -395,17 +403,31 @@ public:
         return *this;
     }
 
-    /** @brief Specifies execution provider for runtime.
+    /** @brief Adds execution provider for runtime.
 
-    The function is used to set ONNX Runtime OpenVINO Execution Provider options.
+    The function is used to add ONNX Runtime OpenVINO Execution Provider options.
 
-    @param ovep OpenVINO Execution Provider options.
+    @param ep OpenVINO Execution Provider options.
     @see cv::gapi::onnx::ep::OpenVINO.
 
     @return the reference on modified object.
     */
-    Params<Net>& cfgExecutionProvider(ep::OpenVINO&& ovep) {
-        desc.execution_provider = std::move(ovep);
+    Params<Net>& cfgAddExecutionProvider(ep::OpenVINO&& ep) {
+        desc.execution_providers.emplace_back(std::move(ep));
+        return *this;
+    }
+
+    /** @brief Adds execution provider for runtime.
+
+    The function is used to add ONNX Runtime DirectML Execution Provider options.
+
+    @param ep DirectML Execution Provider options.
+    @see cv::gapi::onnx::ep::DirectML.
+
+    @return the reference on modified object.
+    */
+    Params<Net>& cfgAddExecutionProvider(ep::DirectML&& ep) {
+        desc.execution_providers.emplace_back(std::move(ep));
         return *this;
     }
 
@@ -447,20 +469,29 @@ public:
     Params(const std::string& tag, const std::string& model_path)
         : desc{model_path, 0u, 0u, {}, {}, {}, {}, {}, {}, {}, {}, {}, true, {}, {}, {}, false }, m_tag(tag) {}
 
+    /** @see onnx::Params::cfgMeanStdDev. */
     void cfgMeanStdDev(const std::string &layer,
                        const cv::Scalar &m,
                        const cv::Scalar &s) {
         desc.generic_mstd[layer] = std::make_pair(m, s);
     }
 
+    /** @see onnx::Params::cfgNormalize. */
     void cfgNormalize(const std::string &layer, bool flag) {
         desc.generic_norm[layer] = flag;
     }
 
-    void cfgExecutionProvider(ep::OpenVINO&& ov_ep) {
-        desc.execution_provider = std::move(ov_ep);
+    /** @see onnx::Params::cfgAddExecutionProvider. */
+    void cfgAddExecutionProvider(ep::OpenVINO&& ep) {
+        desc.execution_providers.emplace_back(std::move(ep));
     }
 
+    /** @see onnx::Params::cfgAddExecutionProvider. */
+    void cfgAddExecutionProvider(ep::DirectML&& ep) {
+        desc.execution_providers.emplace_back(std::move(ep));
+    }
+
+    /** @see onnx::Params::cfgDisableMemPattern. */
     void cfgDisableMemPattern() {
         desc.disable_mem_pattern = true;
     }

modules/gapi/misc/python/pyopencv_gapi.hpp

@@ -30,6 +30,7 @@ using map_string_and_vector_size_t  = std::map<std::string, std::vector<size_t>>
 using map_string_and_vector_float   = std::map<std::string, std::vector<float>>;
 using map_int_and_double            = std::map<int, double>;
 using ep_OpenVINO                   = cv::gapi::onnx::ep::OpenVINO;
+using ep_DirectML                   = cv::gapi::onnx::ep::DirectML;
 
 // NB: Python wrapper generate T_U for T<U>
 // This behavior is only observed for inputs

modules/gapi/src/backends/onnx/bindings_onnx.cpp

@@ -22,8 +22,14 @@ cv::gapi::onnx::PyParams& cv::gapi::onnx::PyParams::cfgNormalize(const std::stri
 }
 
 cv::gapi::onnx::PyParams&
-cv::gapi::onnx::PyParams::cfgExecutionProvider(cv::gapi::onnx::ep::OpenVINO ov_ep) {
-    m_priv->cfgExecutionProvider(std::move(ov_ep));
+cv::gapi::onnx::PyParams::cfgAddExecutionProvider(cv::gapi::onnx::ep::OpenVINO ep) {
+    m_priv->cfgAddExecutionProvider(std::move(ep));
+    return *this;
+}
+
+cv::gapi::onnx::PyParams&
+cv::gapi::onnx::PyParams::cfgAddExecutionProvider(cv::gapi::onnx::ep::DirectML ep) {
+    m_priv->cfgAddExecutionProvider(std::move(ep));
     return *this;
 }
 

modules/gapi/src/backends/onnx/dml_ep.cpp

@@ -0,0 +1,40 @@
+// 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) 2023 Intel Corporation
+
+#include "backends/onnx/dml_ep.hpp"
+#include "logger.hpp"
+
+#ifdef HAVE_ONNX
+#include <onnxruntime_cxx_api.h>
+
+#ifdef HAVE_ONNX_DML
+#include "../providers/dml/dml_provider_factory.h"
+
+void cv::gimpl::onnx::addDMLExecutionProvider(Ort::SessionOptions *session_options,
+                                              const cv::gapi::onnx::ep::DirectML &dml_ep) {
+    namespace ep = cv::gapi::onnx::ep;
+    GAPI_Assert(cv::util::holds_alternative<int>(dml_ep.ddesc));
+    const int device_id = cv::util::get<int>(dml_ep.ddesc);
+    try {
+        OrtSessionOptionsAppendExecutionProvider_DML(*session_options, device_id);
+    } catch (const std::exception &e) {
+        std::stringstream ss;
+        ss << "ONNX Backend: Failed to enable DirectML"
+           << " Execution Provider: " << e.what();
+        cv::util::throw_error(std::runtime_error(ss.str()));
+    }
+}
+
+#else  // HAVE_ONNX_DML
+
+void cv::gimpl::onnx::addDMLExecutionProvider(Ort::SessionOptions*,
+                                              const cv::gapi::onnx::ep::DirectML&) {
+     util::throw_error(std::runtime_error("G-API has been compiled with ONNXRT"
+                                          " without DirectML support"));
+}
+
+#endif  // HAVE_ONNX_DML
+#endif  // HAVE_ONNX

modules/gapi/src/backends/onnx/dml_ep.hpp

@@ -0,0 +1,23 @@
+// 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) 2023 Intel Corporation
+
+#ifndef OPENCV_GAPI_DML_EP_HPP
+#define OPENCV_GAPI_DML_EP_HPP
+
+#include "opencv2/gapi/infer/onnx.hpp"
+#ifdef HAVE_ONNX
+
+#include <onnxruntime_cxx_api.h>
+
+namespace cv {
+namespace gimpl {
+namespace onnx {
+void addDMLExecutionProvider(Ort::SessionOptions *session_options,
+                             const cv::gapi::onnx::ep::DirectML &dml_ep);
+}}}
+
+#endif  // HAVE_ONNX
+#endif  // OPENCV_GAPI_DML_EP_HPP

modules/gapi/src/backends/onnx/gonnxbackend.cpp

@@ -9,6 +9,8 @@
 
 #ifdef HAVE_ONNX
 
+#include "backends/onnx/dml_ep.hpp"
+
 #include <ade/util/algorithm.hpp> // any_of
 #include <ade/util/zip_range.hpp>
 #include <opencv2/gapi/infer.hpp>
@@ -143,37 +145,44 @@ public:
     void run();
 };
 
-static void appendExecutionProvider(Ort::SessionOptions          *session_options,
-                                    const cv::gapi::onnx::ep::EP &execution_provider) {
+static void addOpenVINOExecutionProvider(Ort::SessionOptions *session_options,
+                                         const cv::gapi::onnx::ep::OpenVINO &ov_ep) {
+     OrtOpenVINOProviderOptions options;
+     options.device_type = ov_ep.device_type.c_str();
+     options.cache_dir = ov_ep.cache_dir.c_str();
+     options.num_of_threads = ov_ep.num_of_threads;
+     options.enable_opencl_throttling = ov_ep.enable_opencl_throttling;
+     options.enable_dynamic_shapes = ov_ep.enable_dynamic_shapes;
+     options.context = nullptr;
+
+     try {
+        session_options->AppendExecutionProvider_OpenVINO(options);
+     } catch (const std::exception &e) {
+         std::stringstream ss;
+         ss << "ONNX Backend: Failed to enable OpenVINO"
+            << " Execution Provider: " << e.what();
+         cv::util::throw_error(std::runtime_error(ss.str()));
+     }
+}
+
+static void addExecutionProvider(Ort::SessionOptions          *session_options,
+                                 const cv::gapi::onnx::ep::EP &execution_provider) {
     namespace ep = cv::gapi::onnx::ep;
     switch (execution_provider.index()) {
         case ep::EP::index_of<ep::OpenVINO>(): {
-             GAPI_LOG_INFO(NULL, "OpenVINO Execution Provider is selected.");
-             const auto &ovep = cv::util::get<ep::OpenVINO>(execution_provider);
-             OrtOpenVINOProviderOptions options;
-             options.device_id = ovep.device_id.c_str();
-             options.cache_dir = ovep.cache_dir.c_str();
-             options.enable_opencl_throttling = ovep.enable_opencl_throttling;
-             options.enable_dynamic_shapes = ovep.enable_dynamic_shapes;
-             // NB: If are not specified, will be taken from onnxruntime build.
-             if (ovep.device_type) {
-                options.device_type = ovep.device_type->c_str();
-             }
-             if (ovep.num_of_threads) {
-                options.num_of_threads = *ovep.num_of_threads;
-             }
-             try {
-                session_options->AppendExecutionProvider_OpenVINO(options);
-             } catch (const std::exception &e) {
-                 std::stringstream ss;
-                 ss << "ONNX Backend: Failed to enable OpenVINO Execution Provider: "
-                    << e.what() << "\nMake sure that onnxruntime has"
-                                   " been compiled with OpenVINO support.";
-                 cv::util::throw_error(std::runtime_error(ss.str()));
-             }
+             GAPI_LOG_INFO(NULL, "OpenVINO Execution Provider is added.");
+             const auto &ov_ep = cv::util::get<ep::OpenVINO>(execution_provider);
+             addOpenVINOExecutionProvider(session_options, ov_ep);
              break;
         }
+        case ep::EP::index_of<ep::DirectML>(): {
+            GAPI_LOG_INFO(NULL, "DirectML Execution Provider is added.");
+            const auto &dml_ep = cv::util::get<ep::DirectML>(execution_provider);
+            addDMLExecutionProvider(session_options, dml_ep);
+            break;
+        }
         default:
+            GAPI_LOG_INFO(NULL, "CPU Execution Provider is added.");
             break;
     }
 }
@@ -629,7 +638,10 @@ ONNXCompiled::ONNXCompiled(const gapi::onnx::detail::ParamDesc &pp)
     }
     // Create and initialize the ONNX session
     Ort::SessionOptions session_options;
-    cv::gimpl::onnx::appendExecutionProvider(&session_options, pp.execution_provider);
+    GAPI_LOG_INFO(NULL, "Adding Execution Providers for \"" << pp.model_path << "\"");
+    for (const auto &ep : pp.execution_providers) {
+        cv::gimpl::onnx::addExecutionProvider(&session_options, ep);
+    }
 
     if (pp.disable_mem_pattern) {
         session_options.DisableMemPattern();