[TensorRT RTX EP] Be able to specify aux streams (microsoft#26569)

### Description
Be able to specify auxiliary streams to TensorRT RTX EP. 


### Motivation and Context
In some use cases, we want to have full control over all the streams
used by TRT-RTX, even auxiliary ones.

Author: hadiFute

include/onnxruntime/core/providers/nv_tensorrt_rtx/nv_provider_options.h

@@ -7,8 +7,10 @@
  * - `kDeviceId`: Specifies the GPU device ID to use.
  * - `kHasUserComputeStream`: Indicates whether a user-provided compute stream is used.
  * - `kUserComputeStream`: Specifies the user-provided compute stream.
+ * - `kUserAuxStreamArray`: Specifies the user-provided aux stream.
  * - `kMaxWorkspaceSize`: Sets the maximum workspace size for GPU memory allocation.
  * - 'kMaxSharedMemSize': Sets the maximum amount of shared memory that TensorRT kernels are allowed to use
+ * - `kLengthAuxStreamArray`: Specifies the length/size of the auxiliary streams array (kUserAuxStreamArray). Also sets the maximum number of auxiliary streams for TensorRT execution.
  * - `kDumpSubgraphs`: Enables or disables dumping of subgraphs for debugging.
  * - `kDetailedBuildLog`: Enables or disables detailed build logs for debugging.
  * - `kProfilesMinShapes`: Specifies the minimum shapes for profiling.
@@ -24,8 +26,10 @@ namespace provider_option_names {
 constexpr const char* kDeviceId = "device_id";
 constexpr const char* kHasUserComputeStream = "has_user_compute_stream";
 constexpr const char* kUserComputeStream = "user_compute_stream";
+constexpr const char* kUserAuxStreamArray = "user_aux_stream_array";
 constexpr const char* kMaxWorkspaceSize = "nv_max_workspace_size";
 constexpr const char* kMaxSharedMemSize = "nv_max_shared_mem_size";
+constexpr const char* kLengthAuxStreamArray = "nv_length_aux_stream_array";
 constexpr const char* kDumpSubgraphs = "nv_dump_subgraphs";
 constexpr const char* kDetailedBuildLog = "nv_detailed_build_log";
 constexpr const char* kProfilesMinShapes = "nv_profile_min_shapes";

onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc

@@ -984,6 +984,17 @@ NvExecutionProvider::NvExecutionProvider(const NvExecutionProviderInfo& info)
     stream_ = nullptr;  // Will be created in compute function
   }
 
+  if (info.user_aux_stream_array != nullptr) {
+    if (info.auxiliary_streams <= 0) {
+      ORT_THROW_IF_ERROR(ORT_MAKE_STATUS(ONNXRUNTIME, EP_FAIL, "Auxiliary streams must be greater than 0 when using external auxiliary streams"));
+    }
+    external_aux_streams_ = true;
+    aux_streams_ = reinterpret_cast<cudaStream_t*>(info.user_aux_stream_array);
+  } else {
+    external_aux_streams_ = false;
+    aux_streams_ = nullptr;
+  }
+
   std::string profile_min_shapes, profile_max_shapes, profile_opt_shapes;
 
   // incase the EP context is dumped the engine cache has to be enabled
@@ -3039,6 +3050,11 @@ Status NvExecutionProvider::CreateNodeComputeInfoFromGraph(const GraphViewer& gr
         return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "NvTensorRTRTX EP select an optimization profile for the current context failed");
     }
 
+    // Set auxiliary stream if provided by user
+    if (external_aux_streams_ && aux_streams_ != nullptr) {
+      trt_context->setAuxStreams(aux_streams_, (int32_t)auxiliary_streams_);
+    }
+
     // Check before using trt_engine
     if (trt_engine == nullptr) {
       return ORT_MAKE_STATUS(ONNXRUNTIME, EP_FAIL, "No engine is found.");
@@ -3450,6 +3466,11 @@ Status NvExecutionProvider::CreateNodeComputeInfoFromPrecompiledEngine(const Gra
       }
     }
 
+    // Set auxiliary stream if provided by user
+    if (external_aux_streams_ && aux_streams_ != nullptr) {
+      trt_context->setAuxStreams(aux_streams_, (int32_t)auxiliary_streams_);
+    }
+
     // Start CUDA graph capture with the correct stream
     // Note: We need to set the stream and start capture here because this is where we have access to the actual compute stream
     // Get the graph annotation ID that was stored during OnRunStart

onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.h

@@ -349,6 +349,8 @@ class NvExecutionProvider : public IExecutionProvider {
   mutable NvExecutionProviderInfo info_;
   bool external_stream_ = false;
   cudaStream_t stream_ = nullptr;
+  bool external_aux_streams_ = false;
+  cudaStream_t* aux_streams_ = nullptr;
   int max_partition_iterations_ = 1000;
   size_t min_subgraph_size_ = 1;
   size_t max_workspace_size_ = 0;

onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider_info.cc

@@ -16,6 +16,7 @@ NvExecutionProviderInfo NvExecutionProviderInfo::FromProviderOptions(const Provi
                                                                      const ConfigOptions& session_options) {
   NvExecutionProviderInfo info{};
   void* user_compute_stream = nullptr;
+  void* user_aux_stream_array = nullptr;
   void* onnx_bytestream = nullptr;
   void* external_data_bytestream = nullptr;
   ORT_THROW_IF_ERROR(
@@ -41,8 +42,17 @@ NvExecutionProviderInfo NvExecutionProviderInfo::FromProviderOptions(const Provi
                 user_compute_stream = reinterpret_cast<void*>(address);
                 return Status::OK();
               })
+          .AddValueParser(
+              nv::provider_option_names::kUserAuxStreamArray,
+              [&user_aux_stream_array](const std::string& value_str) -> Status {
+                size_t address;
+                ORT_RETURN_IF_ERROR(ParseStringWithClassicLocale(value_str, address));
+                user_aux_stream_array = reinterpret_cast<void*>(address);
+                return Status::OK();
+              })
           .AddAssignmentToReference(nv::provider_option_names::kMaxWorkspaceSize, info.max_workspace_size)
           .AddAssignmentToReference(nv::provider_option_names::kMaxSharedMemSize, info.max_shared_mem_size)
+          .AddAssignmentToReference(nv::provider_option_names::kLengthAuxStreamArray, info.auxiliary_streams)
           .AddAssignmentToReference(nv::provider_option_names::kDumpSubgraphs, info.dump_subgraphs)
           .AddAssignmentToReference(nv::provider_option_names::kDetailedBuildLog, info.detailed_build_log)
           .AddAssignmentToReference(nv::provider_option_names::kProfilesMinShapes, info.profile_min_shapes)
@@ -56,6 +66,7 @@ NvExecutionProviderInfo NvExecutionProviderInfo::FromProviderOptions(const Provi
 
   info.user_compute_stream = user_compute_stream;
   info.has_user_compute_stream = (user_compute_stream != nullptr);
+  info.user_aux_stream_array = user_aux_stream_array;
   info.onnx_bytestream = onnx_bytestream;
   info.external_data_bytestream = external_data_bytestream;
 
@@ -98,8 +109,10 @@ ProviderOptions NvExecutionProviderInfo::ToProviderOptions(const NvExecutionProv
       {nv::provider_option_names::kDeviceId, MakeStringWithClassicLocale(info.device_id)},
       {nv::provider_option_names::kHasUserComputeStream, MakeStringWithClassicLocale(info.has_user_compute_stream)},
       {nv::provider_option_names::kUserComputeStream, MakeStringWithClassicLocale(reinterpret_cast<size_t>(info.user_compute_stream))},
+      {nv::provider_option_names::kUserAuxStreamArray, MakeStringWithClassicLocale(reinterpret_cast<size_t>(info.user_aux_stream_array))},
       {nv::provider_option_names::kMaxWorkspaceSize, MakeStringWithClassicLocale(info.max_workspace_size)},
       {nv::provider_option_names::kMaxSharedMemSize, MakeStringWithClassicLocale(info.max_shared_mem_size)},
+      {nv::provider_option_names::kLengthAuxStreamArray, MakeStringWithClassicLocale(info.auxiliary_streams)},
       {nv::provider_option_names::kDumpSubgraphs, MakeStringWithClassicLocale(info.dump_subgraphs)},
       {nv::provider_option_names::kDetailedBuildLog, MakeStringWithClassicLocale(info.detailed_build_log)},
       {nv::provider_option_names::kProfilesMinShapes, MakeStringWithClassicLocale(info.profile_min_shapes)},

onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider_info.h

@@ -21,6 +21,7 @@ struct NvExecutionProviderInfo {
   int device_id{0};
   bool has_user_compute_stream{false};
   void* user_compute_stream{nullptr};
+  void* user_aux_stream_array{nullptr};
   int max_partition_iterations{1000};
   int min_subgraph_size{1};
   size_t max_workspace_size{0};