[Backend] Remove deprecated ort api (#2034)

* [Backend] Remove deprecated ort api

`Ort::CustomOpApi` deprecated in ONNXRuntime 1.15

fix https://github.com/PaddlePaddle/FastDeploy/issues/2033

* Improve compatibly with ort older than 1.14

* Update fastdeploy/runtime/backends/ort/ops/multiclass_nms.cc

Co-authored-by: Horror Proton <107091537+horror-proton@users.noreply.github.com>

* Fix double free and wrong attrib type in ort/ops

---------

Co-authored-by: Jason <928090362@qq.com>

fastdeploy/runtime/backends/ort/ops/adaptive_pool2d.cc

@@ -17,13 +17,6 @@
 #include "adaptive_pool2d.h"
 
 namespace fastdeploy {
-struct OrtTensorDimensions : std::vector<int64_t> {
-  OrtTensorDimensions(Ort::CustomOpApi ort, const OrtValue* value) {
-    OrtTensorTypeAndShapeInfo* info = ort.GetTensorTypeAndShape(value);
-    std::vector<int64_t>::operator=(ort.GetTensorShape(info));
-    ort.ReleaseTensorTypeAndShapeInfo(info);
-  }
-};
 
 void AdaptivePool2dKernel::CpuAdaptivePool(
     const std::vector<int64_t>& input_size,
@@ -68,25 +61,38 @@ void AdaptivePool2dKernel::CpuAdaptivePool(
 }
 
 void AdaptivePool2dKernel::Compute(OrtKernelContext* context) {
-  const OrtValue* input = ort_.KernelContext_GetInput(context, 0);
+#if ORT_API_VERSION >= 14
+  Ort::KernelContext ort_context{context};
+  Ort::ConstValue input = ort_context.GetInput(0);
+#else
+  Ort::CustomOpApi api{ort_};
+  Ort::Unowned<const Ort::Value> input{
+      const_cast<OrtValue*>(api.KernelContext_GetInput(context, 0))};
+#endif
+  auto input_data = input.GetTensorData<float>();
+  auto input_dim = input.GetTensorTypeAndShapeInfo().GetShape();
 
-  const float* input_data =
-      reinterpret_cast<const float*>(ort_.GetTensorData<float>(input));
-
-  OrtTensorDimensions input_dim(ort_, input);
   output_size_[0] = input_dim[0];
   std::vector<int64_t> input_size;
   for (auto i : input_dim) {
     input_size.push_back(i);
   }
 
-  OrtValue* output = ort_.KernelContext_GetOutput(
-      context, 0, output_size_.data(), output_size_.size());
-
-  float* output_data = ort_.GetTensorMutableData<float>(output);
+#if ORT_API_VERSION >= 14
+  auto output = ort_context.GetOutput(0, output_size_);
+#else
+  Ort::Unowned<Ort::Value> output{api.KernelContext_GetOutput(
+      context, 0, output_size_.data(), output_size_.size())};
+#endif
+  float* output_data = output.GetTensorMutableData<float>();
   if (!strcmp(this->provider_, "CUDAExecutionProvider")) {
 #ifdef WITH_GPU
-    auto compute_stream = ort_.KernelContext_GetGPUComputeStream(context);
+    auto compute_stream =
+#if ORT_API_VERSION >= 14
+        ort_context.GetGPUComputeStream();
+#else
+        api.KernelContext_GetGPUComputeStream(context);
+#endif
     CudaAdaptivePool(input_size, output_size_, output_data, input_data,
                      compute_stream, pooling_type_);
 #else
@@ -100,14 +106,20 @@ void AdaptivePool2dKernel::Compute(OrtKernelContext* context) {
 }
 
 void AdaptivePool2dKernel::GetAttribute(const OrtKernelInfo* info) {
-  pooling_type_ =
-      ort_.KernelInfoGetAttribute<std::string>(info, "pooling_type");
+#if ORT_API_VERSION >= 14
+  Ort::ConstKernelInfo ort_info{info};
+  pooling_type_ = ort_info.GetAttribute<std::string>("pooling_type");
+  output_size_ = ort_info.GetAttributes<int64_t>("output_size");
+#else
+  Ort::CustomOpApi api{ort_};
+  pooling_type_ = api.KernelInfoGetAttribute<std::string>(info, "pooling_type");
   output_size_ =
-      ort_.KernelInfoGetAttribute<std::vector<int64_t>>(info, "output_size");
-  FDASSERT(
-      output_size_.size() == 4 && output_size_[2] > 0 && output_size_[3] > 0,
-      "The output size of adaptive pool must be positive.");
+      api.KernelInfoGetAttribute<std::vector<int64_t>>(info, "output_size");
+#endif
+  FDASSERT(output_size_.size() == 4 && output_size_[2] > 0 &&
+               output_size_[3] > 0,
+           "The output size of adaptive pool must be positive.");
 }
 }  // namespace fastdeploy
 
-#endif
+#endif

fastdeploy/runtime/backends/ort/ops/adaptive_pool2d.h

@@ -33,12 +33,12 @@ struct AdaptivePool2dKernel {
  protected:
   std::string pooling_type_ = "avg";
   std::vector<int64_t> output_size_ = {};
-  Ort::CustomOpApi ort_;
+  OrtApi ort_;
   void* compute_stream_;
   const char* provider_;
 
  public:
-  AdaptivePool2dKernel(Ort::CustomOpApi ort, const OrtKernelInfo* info,
+  AdaptivePool2dKernel(OrtApi ort, const OrtKernelInfo* info,
                        const char* provider)
       : ort_(ort) {
     GetAttribute(info);
@@ -57,7 +57,7 @@ struct AdaptivePool2dKernel {
 struct AdaptivePool2dOp
     : Ort::CustomOpBase<AdaptivePool2dOp, AdaptivePool2dKernel> {
   explicit AdaptivePool2dOp(const char* provider) : provider_(provider) {}
-  void* CreateKernel(Ort::CustomOpApi api, const OrtKernelInfo* info) const {
+  void* CreateKernel(OrtApi api, const OrtKernelInfo* info) const {
     return new AdaptivePool2dKernel(api, info, provider_);
   }
 

fastdeploy/runtime/backends/ort/ops/multiclass_nms.cc

@@ -23,14 +23,6 @@
 
 namespace fastdeploy {
 
-struct OrtTensorDimensions : std::vector<int64_t> {
-  OrtTensorDimensions(Ort::CustomOpApi ort, const OrtValue* value) {
-    OrtTensorTypeAndShapeInfo* info = ort.GetTensorTypeAndShape(value);
-    std::vector<int64_t>::operator=(ort.GetTensorShape(info));
-    ort.ReleaseTensorTypeAndShapeInfo(info);
-  }
-};
-
 template <class T>
 bool SortScorePairDescend(const std::pair<float, T>& pair1,
                           const std::pair<float, T>& pair2) {
@@ -165,14 +157,24 @@ int MultiClassNmsKernel::NMSForEachSample(
 }
 
 void MultiClassNmsKernel::Compute(OrtKernelContext* context) {
-  const OrtValue* boxes = ort_.KernelContext_GetInput(context, 0);
-  const OrtValue* scores = ort_.KernelContext_GetInput(context, 1);
-  const float* boxes_data =
-      reinterpret_cast<const float*>(ort_.GetTensorData<float>(boxes));
-  const float* scores_data =
-      reinterpret_cast<const float*>(ort_.GetTensorData<float>(scores));
-  OrtTensorDimensions boxes_dim(ort_, boxes);
-  OrtTensorDimensions scores_dim(ort_, scores);
+#if ORT_API_VERSION >= 14
+  Ort::KernelContext ort_context{context};
+  Ort::ConstValue boxes = ort_context.GetInput(0);
+  Ort::ConstValue scores = ort_context.GetInput(1);
+#else
+  Ort::CustomOpApi api{ort_};
+  Ort::Unowned<const Ort::Value> boxes{
+      const_cast<OrtValue*>(api.KernelContext_GetInput(context, 0))};
+  Ort::Unowned<const Ort::Value> scores{
+      const_cast<OrtValue*>(api.KernelContext_GetInput(context, 1))};
+#endif
+
+  auto boxes_data = boxes.GetTensorData<float>();
+  auto scores_data = scores.GetTensorData<float>();
+
+  auto boxes_dim = boxes.GetTensorTypeAndShapeInfo().GetShape();
+  auto scores_dim = scores.GetTensorTypeAndShapeInfo().GetShape();
+
   int score_size = scores_dim.size();
 
   int64_t batch_size = scores_dim[0];
@@ -183,12 +185,16 @@ void MultiClassNmsKernel::Compute(OrtKernelContext* context) {
   FDASSERT(score_size == 3,
            "Require rank of input scores be 3, but now it's %d.", score_size);
   FDASSERT(boxes_dim[2] == 4,
-           "Require the 3-dimension of input boxes be 4, but now it's %lld.",
+           "Require the 3-dimension of input boxes be 4, but now it's %ld.",
            box_dim);
   std::vector<int64_t> out_num_rois_dims = {batch_size};
-  OrtValue* out_num_rois = ort_.KernelContext_GetOutput(
-      context, 2, out_num_rois_dims.data(), out_num_rois_dims.size());
-  int32_t* out_num_rois_data = ort_.GetTensorMutableData<int32_t>(out_num_rois);
+#if ORT_API_VERSION >= 14
+  auto out_num_rois = ort_context.GetOutput(2, out_num_rois_dims);
+#else
+  Ort::Unowned<Ort::Value> out_num_rois{api.KernelContext_GetOutput(
+      context, 2, out_num_rois_dims.data(), out_num_rois_dims.size())};
+#endif
+  int32_t* out_num_rois_data = out_num_rois.GetTensorMutableData<int32_t>();
 
   std::vector<std::map<int, std::vector<int>>> all_indices;
   for (size_t i = 0; i < batch_size; ++i) {
@@ -205,20 +211,26 @@ void MultiClassNmsKernel::Compute(OrtKernelContext* context) {
   }
   std::vector<int64_t> out_box_dims = {num_nmsed_out, 6};
   std::vector<int64_t> out_index_dims = {num_nmsed_out, 1};
-  OrtValue* out_box = ort_.KernelContext_GetOutput(
-      context, 0, out_box_dims.data(), out_box_dims.size());
-  OrtValue* out_index = ort_.KernelContext_GetOutput(
-      context, 1, out_index_dims.data(), out_index_dims.size());
+
+#if ORT_API_VERSION >= 14
+  auto out_box = ort_context.GetOutput(0, out_box_dims);
+  auto out_index = ort_context.GetOutput(1, out_index_dims);
+#else
+  Ort::Unowned<Ort::Value> out_box{api.KernelContext_GetOutput(
+      context, 0, out_box_dims.data(), out_box_dims.size())};
+  Ort::Unowned<Ort::Value> out_index{api.KernelContext_GetOutput(
+      context, 1, out_index_dims.data(), out_index_dims.size())};
+#endif
+
   if (num_nmsed_out == 0) {
-    int32_t* out_num_rois_data =
-        ort_.GetTensorMutableData<int32_t>(out_num_rois);
+    int32_t* out_num_rois_data = out_num_rois.GetTensorMutableData<int32_t>();
     for (size_t i = 0; i < batch_size; ++i) {
       out_num_rois_data[i] = 0;
     }
     return;
   }
-  float* out_box_data = ort_.GetTensorMutableData<float>(out_box);
-  int32_t* out_index_data = ort_.GetTensorMutableData<int32_t>(out_index);
+  float* out_box_data = out_box.GetTensorMutableData<float>();
+  int32_t* out_index_data = out_index.GetTensorMutableData<int32_t>();
 
   int count = 0;
   for (size_t i = 0; i < batch_size; ++i) {
@@ -249,15 +261,27 @@ void MultiClassNmsKernel::Compute(OrtKernelContext* context) {
 }
 
 void MultiClassNmsKernel::GetAttribute(const OrtKernelInfo* info) {
+#if ORT_API_VERSION >= 14
+  Ort::ConstKernelInfo ort_info{info};
+  background_label = ort_info.GetAttribute<int64_t>("background_label");
+  keep_top_k = ort_info.GetAttribute<int64_t>("keep_top_k");
+  nms_eta = ort_info.GetAttribute<float>("nms_eta");
+  nms_threshold = ort_info.GetAttribute<float>("nms_threshold");
+  nms_top_k = ort_info.GetAttribute<int64_t>("nms_top_k");
+  normalized = ort_info.GetAttribute<int64_t>("normalized");
+  score_threshold = ort_info.GetAttribute<float>("score_threshold");
+#else
+  Ort::CustomOpApi api{ort_};
   background_label =
-      ort_.KernelInfoGetAttribute<int64_t>(info, "background_label");
-  keep_top_k = ort_.KernelInfoGetAttribute<int64_t>(info, "keep_top_k");
-  nms_eta = ort_.KernelInfoGetAttribute<float>(info, "nms_eta");
-  nms_threshold = ort_.KernelInfoGetAttribute<float>(info, "nms_threshold");
-  nms_top_k = ort_.KernelInfoGetAttribute<int64_t>(info, "nms_top_k");
-  normalized = ort_.KernelInfoGetAttribute<int64_t>(info, "normalized");
-  score_threshold = ort_.KernelInfoGetAttribute<float>(info, "score_threshold");
+      api.KernelInfoGetAttribute<int64_t>(info, "background_label");
+  keep_top_k = api.KernelInfoGetAttribute<int64_t>(info, "keep_top_k");
+  nms_eta = api.KernelInfoGetAttribute<float>(info, "nms_eta");
+  nms_threshold = api.KernelInfoGetAttribute<float>(info, "nms_threshold");
+  nms_top_k = api.KernelInfoGetAttribute<int64_t>(info, "nms_top_k");
+  normalized = api.KernelInfoGetAttribute<int64_t>(info, "normalized");
+  score_threshold = api.KernelInfoGetAttribute<float>(info, "score_threshold");
+#endif
 }
 }  // namespace fastdeploy
 
-#endif
+#endif

fastdeploy/runtime/backends/ort/ops/multiclass_nms.h

@@ -30,11 +30,10 @@ struct MultiClassNmsKernel {
   int64_t nms_top_k;
   bool normalized;
   float score_threshold;
-  Ort::CustomOpApi ort_;
+  OrtApi ort_;
 
  public:
-  MultiClassNmsKernel(Ort::CustomOpApi ort, const OrtKernelInfo* info)
-      : ort_(ort) {
+  MultiClassNmsKernel(OrtApi ort, const OrtKernelInfo* info) : ort_(ort) {
     GetAttribute(info);
   }
 
@@ -50,7 +49,7 @@ struct MultiClassNmsKernel {
 
 struct MultiClassNmsOp
     : Ort::CustomOpBase<MultiClassNmsOp, MultiClassNmsKernel> {
-  void* CreateKernel(Ort::CustomOpApi api, const OrtKernelInfo* info) const {
+  void* CreateKernel(OrtApi api, const OrtKernelInfo* info) const {
     return new MultiClassNmsKernel(api, info);
   }