Merge pull request #25817 from lamiayous:ly/extend_onnxrt_gapi_backend_handle_i32_i64_type

Handling I32/I64 data types in G-API ONNX back-end #25817

### 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/src/backends/onnx/gonnxbackend.cpp

@@ -123,7 +123,7 @@ class ONNXCompiled {
     std::vector<cv::Mat> out_data;
 
     void Run(const std::vector<cv::Mat>& ins,
-             const std::vector<cv::Mat>& outs);
+             std::vector<cv::Mat>& outs);
 
     std::vector<std::string> in_names_without_const;
 public:
@@ -322,22 +322,20 @@ inline std::vector<int64_t> toORT(const cv::MatSize &sz) {
 inline void preprocess(const cv::Mat& src,
                        const cv::gimpl::onnx::TensorInfo& ti,
                              cv::Mat& dst) {
-    GAPI_Assert(src.depth() == CV_32F || src.depth() == CV_8U);
     // CNN input type
     const auto type = toCV(ti.type);
-    if (src.depth() == CV_32F) {
+    if (src.depth() != CV_8U) {
         // Just pass the tensor as-is.
         // No layout or dimension transformations done here!
         // TODO: This needs to be aligned across all NN backends.
-        GAPI_Assert(type == CV_32F && "Only 32F model input is supported for 32F input data");
         const auto tensor_dims = toORT(src.size);
         if (tensor_dims.size() == ti.dims.size()) {
             for (size_t i = 0; i < ti.dims.size(); ++i) {
                 GAPI_Assert((ti.dims[i] == -1 || ti.dims[i] == tensor_dims[i]) &&
-                            "32F tensor dimensions should match with all non-dynamic NN input dimensions");
+                            "Non-U8 tensor dimensions should match with all non-dynamic NN input dimensions");
             }
         } else {
-            GAPI_Error("32F tensor size should match with NN input");
+            GAPI_Error("Non-U8 tensor size should match with NN input");
         }
 
         dst = src;
@@ -471,6 +469,25 @@ inline Ort::Value createTensor(const Ort::MemoryInfo& memory_info,
         return createTensor<float>(memory_info, tensor_params, data);
     case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32:
         return createTensor<int32_t>(memory_info, tensor_params, data);
+    case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64:{
+        // cv::Mat does not support int64 data directly.
+        // Following steps are applied to create an ONNX tensor from cv::Mat data:
+        // - First create a new ONNX tensor 'i64_tensor' with data type int64_t using the default allocator
+        // - Next retrieve a pointer to the mutable data buffer of 'i64_tensor'
+        // - Convert the data from int32 (see toCV function) to int64 and deep copy it into 'i64_tensor'
+        auto ort_dims = toORT(data.size);
+
+        Ort::AllocatorWithDefaultOptions allocator;
+        Ort::Value i64_tensor = Ort::Value::CreateTensor<int64_t>(allocator,
+                                                                  ort_dims.data(),
+                                                                  ort_dims.size());
+        int64_t* tensor_data = i64_tensor.GetTensorMutableData<int64_t>();
+
+        cv::gimpl::convertInt32ToInt64(data.ptr<int>(),
+                                       tensor_data,
+                                       data.total());
+        return i64_tensor;
+    }
     default:
         GAPI_Error("ONNX. Unsupported data type");
     }
@@ -747,9 +764,11 @@ ONNXCompiled::ONNXCompiled(const gapi::onnx::detail::ParamDesc &pp)
                             in_tensor_info.end(),
                             [](const cv::gimpl::onnx::TensorInfo &p) {
                                 return p.type == ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT
-                                    || p.type == ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8;
+                                    || p.type == ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8
+                                    || p.type == ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32
+                                    || p.type == ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64;
                             })
-                && "Only FP32 and U8 inputs for NN are supported");
+                && "Only FP32, INT32, INT64 and U8 inputs for NN are supported");
 
     // Put mean and std in appropriate tensor params
     if (!params.mean.empty() || !params.stdev.empty()) {
@@ -864,7 +883,7 @@ cv::Mat ONNXCompiled::allocOutput(int i) const {
 }
 
 void ONNXCompiled::Run(const std::vector<cv::Mat>& ins,
-                       const std::vector<cv::Mat>& outs) {
+                       std::vector<cv::Mat>& outs) {
     std::vector<Ort::Value> in_tensors, out_tensors;
 
     // Layer names order for run
@@ -909,6 +928,17 @@ void ONNXCompiled::Run(const std::vector<cv::Mat>& ins,
                          out_run_names.data(),
                          &out_tensors.front(),
                          params.output_names.size());
+        if (out_tensor_info[0].type == ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64) {
+            // cv::Mat does not support int64 output data.
+            // Conversion from int 64 to int 32 is carried in the copyFromONNX function
+            // The output is written to out_mat
+            for (auto &&iter : ade::util::zip(ade::util::toRange(out_tensors),
+                                              ade::util::toRange(outs))) {
+                auto &out_tensor = std::get<0>(iter);
+                auto &out_mat = std::get<1>(iter);
+                copyFromONNX(out_tensor, out_mat);
+            }
+        }
     } else {
         // Hard path - run session & user-defined post-processing
         // NOTE: use another list of output names here