Merge pull request #24059 from TolyaTalamanov:at/add-onnx-cuda-execution-provider

G-API: Support CUDA & TensoRT Execution Providers for ONNXRT Backend #24059

### 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

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

@@ -39,6 +39,12 @@ public:
     GAPI_WRAP
     PyParams& cfgAddExecutionProvider(ep::DirectML ep);
 
+    GAPI_WRAP
+    PyParams& cfgAddExecutionProvider(ep::CUDA ep);
+
+    GAPI_WRAP
+    PyParams& cfgAddExecutionProvider(ep::TensorRT ep);
+
     GAPI_WRAP
     PyParams& cfgDisableMemPattern();
 

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

@@ -32,6 +32,56 @@ namespace onnx {
  */
 namespace ep {
 
+/**
+ * @brief This structure provides functions
+ * that fill inference options for CUDA Execution Provider.
+ * Please follow https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#cuda-execution-provider
+ */
+struct GAPI_EXPORTS_W_SIMPLE CUDA {
+    // NB: Used from python.
+    /// @private -- Exclude this constructor from OpenCV documentation
+    GAPI_WRAP
+    CUDA() = default;
+
+    /** @brief Class constructor.
+
+    Constructs CUDA parameters based on device type information.
+
+    @param dev_id Target device id to use.
+    */
+    GAPI_WRAP
+    explicit CUDA(const int dev_id)
+        : device_id(dev_id) {
+    }
+
+    int device_id;
+};
+
+/**
+ * @brief This structure provides functions
+ * that fill inference options for TensorRT Execution Provider.
+ * Please follow https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#tensorrt-execution-provider
+ */
+struct GAPI_EXPORTS_W_SIMPLE TensorRT {
+    // NB: Used from python.
+    /// @private -- Exclude this constructor from OpenCV documentation
+    GAPI_WRAP
+    TensorRT() = default;
+
+    /** @brief Class constructor.
+
+    Constructs TensorRT parameters based on device type information.
+
+    @param dev_id Target device id to use.
+    */
+    GAPI_WRAP
+    explicit TensorRT(const int dev_id)
+        : device_id(dev_id) {
+    }
+
+    int device_id;
+};
+
 /**
  * @brief This structure provides functions
  * that fill inference options for ONNX OpenVINO Execution Provider.
@@ -143,7 +193,11 @@ public:
     DeviceDesc ddesc;
 };
 
-using EP = cv::util::variant<cv::util::monostate, OpenVINO, DirectML>;
+using EP = cv::util::variant< cv::util::monostate
+                            , OpenVINO
+                            , DirectML
+                            , CUDA
+                            , TensorRT>;
 
 } // namespace ep
 
@@ -431,6 +485,34 @@ public:
         return *this;
     }
 
+    /** @brief Adds execution provider for runtime.
+
+    The function is used to add ONNX Runtime CUDA Execution Provider options.
+
+    @param ep CUDA Execution Provider options.
+    @see cv::gapi::onnx::ep::CUDA.
+
+    @return the reference on modified object.
+    */
+    Params<Net>& cfgAddExecutionProvider(ep::CUDA&& 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 TensorRT Execution Provider options.
+
+    @param ep TensorRT Execution Provider options.
+    @see cv::gapi::onnx::ep::TensorRT.
+
+    @return the reference on modified object.
+    */
+    Params<Net>& cfgAddExecutionProvider(ep::TensorRT&& ep) {
+        desc.execution_providers.emplace_back(std::move(ep));
+        return *this;
+    }
+
     /** @brief Disables the memory pattern optimization.
 
     @return the reference on modified object.
@@ -491,6 +573,16 @@ public:
         desc.execution_providers.emplace_back(std::move(ep));
     }
 
+    /** @see onnx::Params::cfgAddExecutionProvider. */
+    void cfgAddExecutionProvider(ep::CUDA&& ep) {
+        desc.execution_providers.emplace_back(std::move(ep));
+    }
+
+    /** @see onnx::Params::cfgAddExecutionProvider. */
+    void cfgAddExecutionProvider(ep::TensorRT&& 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

@@ -31,6 +31,8 @@ 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;
+using ep_CUDA                       = cv::gapi::onnx::ep::CUDA;
+using ep_TensorRT                   = cv::gapi::onnx::ep::TensorRT;
 
 // NB: Python wrapper generate T_U for T<U>
 // This behavior is only observed for inputs

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

@@ -33,6 +33,18 @@ cv::gapi::onnx::PyParams::cfgAddExecutionProvider(cv::gapi::onnx::ep::DirectML e
     return *this;
 }
 
+cv::gapi::onnx::PyParams&
+cv::gapi::onnx::PyParams::cfgAddExecutionProvider(cv::gapi::onnx::ep::CUDA ep) {
+    m_priv->cfgAddExecutionProvider(std::move(ep));
+    return *this;
+}
+
+cv::gapi::onnx::PyParams&
+cv::gapi::onnx::PyParams::cfgAddExecutionProvider(cv::gapi::onnx::ep::TensorRT ep) {
+    m_priv->cfgAddExecutionProvider(std::move(ep));
+    return *this;
+}
+
 cv::gapi::onnx::PyParams&
 cv::gapi::onnx::PyParams::cfgDisableMemPattern() {
     m_priv->cfgDisableMemPattern();

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

@@ -145,9 +145,39 @@ public:
     void run();
 };
 
+static void addCUDAExecutionProvider(Ort::SessionOptions *session_options,
+                                     const cv::gapi::onnx::ep::CUDA &cuda_ep) {
+     OrtCUDAProviderOptions options{};
+     options.device_id = cuda_ep.device_id;
+
+     try {
+        session_options->AppendExecutionProvider_CUDA(options);
+     } catch (const std::exception &e) {
+         std::stringstream ss;
+         ss << "ONNX Backend: Failed to enable CUDA"
+            << " Execution Provider: " << e.what();
+         cv::util::throw_error(std::runtime_error(ss.str()));
+     }
+}
+
+static void addTensorRTExecutionProvider(Ort::SessionOptions *session_options,
+                                         const cv::gapi::onnx::ep::TensorRT &trt_ep) {
+     OrtTensorRTProviderOptions options{};
+     options.device_id = trt_ep.device_id;
+
+     try {
+        session_options->AppendExecutionProvider_TensorRT(options);
+     } catch (const std::exception &e) {
+         std::stringstream ss;
+         ss << "ONNX Backend: Failed to enable TensorRT"
+            << " Execution Provider: " << e.what();
+         cv::util::throw_error(std::runtime_error(ss.str()));
+     }
+}
+
 static void addOpenVINOExecutionProvider(Ort::SessionOptions *session_options,
                                          const cv::gapi::onnx::ep::OpenVINO &ov_ep) {
-     OrtOpenVINOProviderOptions options;
+     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;
@@ -181,6 +211,18 @@ static void addExecutionProvider(Ort::SessionOptions          *session_options,
             addDMLExecutionProvider(session_options, dml_ep);
             break;
         }
+        case ep::EP::index_of<ep::CUDA>(): {
+            GAPI_LOG_INFO(NULL, "CUDA Execution Provider is added.");
+            const auto &cuda_ep = cv::util::get<ep::CUDA>(execution_provider);
+            addCUDAExecutionProvider(session_options, cuda_ep);
+            break;
+        }
+        case ep::EP::index_of<ep::TensorRT>(): {
+            GAPI_LOG_INFO(NULL, "TensorRT Execution Provider is added.");
+            const auto &trt_ep = cv::util::get<ep::TensorRT>(execution_provider);
+            addTensorRTExecutionProvider(session_options, trt_ep);
+            break;
+        }
         default:
             GAPI_LOG_INFO(NULL, "CPU Execution Provider is added.");
             break;