com.github.asus4.onnxruntime 0.1.14 → 0.2.0

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64_x86_64-simulator/onnxruntime.framework/Headers/onnxruntime_run_options_config_keys.h

@@ -30,3 +30,22 @@ static const char* const kOrtRunOptionsConfigEnableMemoryArenaShrinkage = "memor
 // Per default it will be set to '0'
 // Taking CUDA EP as an example, it omit triggering cudaStreamSynchronize on the compute stream.
 static const char* const kOrtRunOptionsConfigDisableSynchronizeExecutionProviders = "disable_synchronize_execution_providers";
+
+// Set HTP performance mode for QNN HTP backend before session run.
+// options for HTP performance mode: "burst", "balanced", "default", "high_performance",
+// "high_power_saver", "low_balanced", "extreme_power_saver", "low_power_saver", "power_saver",
+// "sustained_high_performance". Default to "default".
+static const char* const kOrtRunOptionsConfigQnnPerfMode = "qnn.htp_perf_mode";
+
+// Set HTP performance mode for QNN HTP backend post session run.
+static const char* const kOrtRunOptionsConfigQnnPerfModePostRun = "qnn.htp_perf_mode_post_run";
+
+// Set RPC control latency for QNN HTP backend
+static const char* const kOrtRunOptionsConfigQnnRpcControlLatency = "qnn.rpc_control_latency";
+
+// Set graph annotation id for CUDA EP. Use with enable_cuda_graph=true.
+// The value should be an integer. If the value is not set, the default value is 0 and
+// ORT session only captures one cuda graph before another capture is requested.
+// If the value is set to -1, cuda graph capture/replay is disabled in that run.
+// User are not expected to set the value to 0 as it is reserved for internal use.
+static const char* const kOrtRunOptionsConfigCudaGraphAnnotation = "gpu_graph_id";

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64_x86_64-simulator/onnxruntime.framework/Headers/onnxruntime_session_options_config_keys.h

@@ -78,21 +78,35 @@ static const char* const kOrtSessionOptionsEnableGeluApproximation = "optimizati
 static const char* const kOrtSessionOptionsDisableAheadOfTimeFunctionInlining = "session.disable_aot_function_inlining";
 
 #ifdef ENABLE_TRAINING
-// Specifies a list of op types for memory footprint reduction.
-// The value should be a ","-delimited list of pair of
-// <subgraph string: optimization strategy: number of subgraph to apply>.
-// For example, "Gelu+Cast+:1:0,Dropout+:1:1".
-//   A valid "subgraph string" should be one subgraph representation output by ORT graph transformations.
-//   "optimization strategy" currently has valid values: 0 - disabled, 1 - recompute.
-//   "number of subgraph to apply" is used to control how many subgraphs to apply optimization, to avoid "oversaving"
-//   the memory.
-static const char* const kOrtSessionOptionsMemoryOptimizerEnabler = "optimization.memory_optimizer_config";
+// Specifies a path of the file containing a list of memory optimization configurations.
+// The value should be a string indicating the file path of the config file.
+// The content of the config file is a JSON struct like this:
+// [
+//   "Gelu+Cast+:1:0",
+//   "Dropout+:1:1"
+// ]
+// Taking the example of "Gelu+Cast+:1:0",
+// > "Gelu+Cast+" is the subgraph string, a valid "subgraph string" should be one subgraph representation
+//    output by ORT graph transformations.
+// > "1" is "optimization strategy", valid values: 0 - disabled, 1 - recompute.
+// > "0" is "number of subgraph to apply" which is used to control how many subgraphs to apply optimization,
+//    to avoid "oversaving" the memory.
+static const char* const kOrtSessionOptionsMemoryOptimizerApplyConfig = "optimization.memory_optimizer_config";
 
 // Specifies the config for detecting subgraphs for memory footprint reduction.
 // The value should be a string contains int separated using commas. The default value is "0:0".
 static const char* const kOrtSessionOptionsMemoryOptimizerProbeConfig = "optimization.enable_memory_probe_recompute_config";
 #endif
 
+// This setting if set should contain a comma separated list of optimizers names that should be disabled.
+// Optimizers may take time to execute and affect model loading time. If you feel that a specific optimizer
+// does not provider runtime benefits, but affects your model loading time you may disable it using this config
+// entry. This option is not enabled in ORT_MINIMAL_BUILD build.
+// A list of optimizes is available in onnxruntime/core/optimizer/graph_transformer_utils.cc
+//
+// Default is an empty string which means no optimizers are disabled.
+static const char* const kOrtSessionOptionsDisableSpecifiedOptimizers = "optimization.disable_specified_optimizers";
+
 // Enable or disable using device allocator for allocating initialized tensor memory. "1": enable; "0": disable. The default is "0".
 // Using device allocators means the memory allocation is made using malloc/new.
 static const char* const kOrtSessionOptionsUseDeviceAllocatorForInitializers = "session.use_device_allocator_for_initializers";
@@ -251,8 +265,17 @@ static const char* const kOrtSessionOptionEpContextFilePath = "ep.context_file_p
 // "1": dump the EP context into the Onnx model. (default).
 static const char* const kOrtSessionOptionEpContextEmbedMode = "ep.context_embed_mode";
 
+// Specify the EPContext node name prefix to make it unique
+// in case user need to merge/connect multiple EPContext nodes in one model
+static const char* const kOrtSessionOptionEpContextNodeNamePrefix = "ep.context_node_name_prefix";
+
 // Gemm fastmath mode provides fp32 gemm acceleration with bfloat16 based matmul.
 // Option values:
 // - "0": Gemm FastMath mode is not enabled. [DEFAULT]
 // - "1": Gemm FastMath mode is enabled.
 static const char* const kOrtSessionOptionsMlasGemmFastMathArm64Bfloat16 = "mlas.enable_gemm_fastmath_arm64_bfloat16";
+
+// When converting DQ + MatMul -> MatMulNBits, the accuracy level of the MatMulNBits is controlled by this option.
+// Refer to MatMulNBits op schema for more details.
+// If not provided, default is 4.
+static const char* const kOrtSessionOptionsQDQMatMulNBitsAccuracyLevel = "session.qdq_matmulnbits_accuracy_level";

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64_x86_64-simulator/onnxruntime.framework/Info.plist

@@ -9,9 +9,9 @@
 	<key>CFBundleIdentifier</key>
 	<string>com.microsoft.onnxruntime</string>
 	<key>CFBundleVersion</key>
-	<string>1.17.3</string>
+	<string>1.19.2</string>
 	<key>CFBundleShortVersionString</key>
-	<string>1.17.3</string>
+	<string>1.19.2</string>
 	<key>CFBundleSignature</key>
 	<string>????</string>
 	<key>CFBundlePackageType</key>

package/Plugins/iOS~/onnxruntime.xcframework/macos-arm64_x86_64/onnxruntime.framework/{Headers → Versions/A/Headers}/coreml_provider_factory.h

@@ -28,9 +28,12 @@ enum COREMLFlags {
   // dynamic shapes. However, the performance may be negatively impacted if inputs have dynamic shapes.
   COREML_FLAG_ONLY_ALLOW_STATIC_INPUT_SHAPES = 0x008,
 
+  // Create an MLProgram. By default it will create a NeuralNetwork model. Requires Core ML 5 or later.
+  COREML_FLAG_CREATE_MLPROGRAM = 0x010,
+
   // Keep COREML_FLAG_LAST at the end of the enum definition
   // And assign the last COREMLFlag to it
-  COREML_FLAG_LAST = COREML_FLAG_ONLY_ALLOW_STATIC_INPUT_SHAPES,
+  COREML_FLAG_LAST = COREML_FLAG_CREATE_MLPROGRAM,
 };
 
 #ifdef __cplusplus

package/Plugins/iOS~/onnxruntime.xcframework/macos-arm64_x86_64/onnxruntime.framework/{Headers → Versions/A/Headers}/onnxruntime_c_api.h

@@ -38,7 +38,7 @@
  *
  * This value is used by some API functions to behave as this version of the header expects.
  */
-#define ORT_API_VERSION 17
+#define ORT_API_VERSION 19
 
 #ifdef __cplusplus
 extern "C" {
@@ -196,7 +196,10 @@ typedef enum ONNXTensorElementDataType {
   ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FN,    // Non-IEEE floating-point format based on IEEE754 single-precision
   ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FNUZ,  // Non-IEEE floating-point format based on IEEE754 single-precision
   ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2,      // Non-IEEE floating-point format based on IEEE754 single-precision
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2FNUZ   // Non-IEEE floating-point format based on IEEE754 single-precision
+  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2FNUZ,  // Non-IEEE floating-point format based on IEEE754 single-precision
+  // Int4 types were introduced in ONNX 1.16. See https://onnx.ai/onnx/technical/int4.html
+  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT4,  // maps to a pair of packed uint4 values (size == 1 byte)
+  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT4    // maps to a pair of packed int4 values (size == 1 byte)
 } ONNXTensorElementDataType;
 
 // Synced with onnx TypeProto oneof
@@ -267,7 +270,7 @@ typedef enum OrtOpAttrType {
 //! @}
 #define ORT_RUNTIME_CLASS(X) \
   struct Ort##X;             \
-  typedef struct Ort##X Ort##X;
+  typedef struct Ort##X Ort##X
 
 /** \addtogroup Global
  * ONNX Runtime C API
@@ -319,6 +322,12 @@ typedef struct OrtAllocator {
   void*(ORT_API_CALL* Alloc)(struct OrtAllocator* this_, size_t size);                ///< Returns a pointer to an allocated block of `size` bytes
   void(ORT_API_CALL* Free)(struct OrtAllocator* this_, void* p);                      ///< Free a block of memory previously allocated with OrtAllocator::Alloc
   const struct OrtMemoryInfo*(ORT_API_CALL* Info)(const struct OrtAllocator* this_);  ///< Return a pointer to an ::OrtMemoryInfo that describes this allocator
+  /**
+   * @brief Optional allocation function to use for memory allocations made during session initialization.
+   * Use this function if you want to separate allocations made by ORT during Run() calls from
+   * those made during session initialization. This allows for separate memory management strategies for these allocations.
+   */
+  void*(ORT_API_CALL* Reserve)(struct OrtAllocator* this_, size_t size);  ///< Returns a pointer to an allocated block of `size` bytes
 } OrtAllocator;
 
 typedef void(ORT_API_CALL* OrtLoggingFunction)(
@@ -464,13 +473,13 @@ typedef struct OrtCUDAProviderOptions {
 
   /** \brief Enable TunableOp for using.
    *   Set it to 1/0 to enable/disable TunableOp. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_CUDA_TUNABLE_OP_ENABLE.
+   *   This option can be overridden by environment variable ORT_CUDA_TUNABLE_OP_ENABLE.
    */
   int tunable_op_enable;
 
   /** \brief Enable TunableOp for tuning.
    *   Set it to 1/0 to enable/disable TunableOp tuning. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_CUDA_TUNABLE_OP_TUNING_ENABLE.
+   *   This option can be overridden by environment variable ORT_CUDA_TUNABLE_OP_TUNING_ENABLE.
    */
   int tunable_op_tuning_enable;
 
@@ -496,6 +505,7 @@ typedef struct OrtROCMProviderOptions {
         has_user_compute_stream{},
         user_compute_stream{},
         default_memory_arena_cfg{},
+        enable_hip_graph{false},
         tunable_op_enable{false},
         tunable_op_tuning_enable{false},
         tunable_op_max_tuning_duration_ms{} {}
@@ -548,15 +558,17 @@ typedef struct OrtROCMProviderOptions {
    */
   OrtArenaCfg* default_memory_arena_cfg;
 
+  int enable_hip_graph;
+
   /** \brief Enable TunableOp for using.
    *   Set it to 1/0 to enable/disable TunableOp. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_ROCM_TUNABLE_OP_ENABLE.
+   *   This option can be overridden by environment variable ORT_ROCM_TUNABLE_OP_ENABLE.
    */
   int tunable_op_enable;
 
   /** \brief Enable TunableOp for tuning.
    *   Set it to 1/0 to enable/disable TunableOp tuning. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_ROCM_TUNABLE_OP_TUNING_ENABLE.
+   *   This option can be overridden by environment variable ORT_ROCM_TUNABLE_OP_TUNING_ENABLE.
    */
   int tunable_op_tuning_enable;
 
@@ -605,6 +617,10 @@ typedef struct OrtMIGraphXProviderOptions {
   int migraphx_int8_enable;                          // MIGraphX INT8 precision. Default 0 = false, nonzero = true
   int migraphx_use_native_calibration_table;         // MIGraphx INT8 cal table. Default 0 = false, noznero = true
   const char* migraphx_int8_calibration_table_name;  // MIGraphx INT8 calibration table name
+  int migraphx_save_compiled_model;                  // migraphx save compiled model. Default 0 = false, noznero = true
+  const char* migraphx_save_model_path;              // migraphx model path name
+  int migraphx_load_compiled_model;                  // migraphx int8 cal table. Default 0 = false, noznero = true
+  const char* migraphx_load_model_path;              // migraphx model path name
 } OrtMIGraphXProviderOptions;
 
 /** \brief OpenVINO Provider Options
@@ -1834,14 +1850,30 @@ struct OrtApi {
 
   /** \brief Used for custom operators, get an input of a kernel
    *
-   * \see ::OrtCustomOp
+   * The function attempts fetches the input of the kernel. If the input is optional
+   * and not present, the function returns success and out is set to nullptr.
+   *
+   * \param[in] context ::OrtKernelContext instance
+   * \param[in] index See KernelContext_GetInputCount for boundaries check.
+   * \param[out] out OrtValue if the input is present otherwise is set nullptr
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
    */
   ORT_API2_STATUS(KernelContext_GetInput, _In_ const OrtKernelContext* context, _In_ size_t index,
                   _Out_ const OrtValue** out);
 
   /** \brief Used for custom operators, get an output of a kernel
    *
-   * \see ::OrtCustomOp
+   * The function attempts fetches the output of the kernel. If the output is optional
+   * and not present, the function returns success and out is set to nullptr.
+   *
+   * \param[in] context ::OrtKernelContext instance
+   * \param[in] index See KernelContext_GetOutputCount for boundaries check.
+   * \param[in] dim_values output dimensions
+   * \param[in] dim_count number of dimensions
+   * \param[out] out a ptr to OrtValue to output otherwise set to nullptr
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
    */
   ORT_API2_STATUS(KernelContext_GetOutput, _Inout_ OrtKernelContext* context, _In_ size_t index,
                   _In_ const int64_t* dim_values, size_t dim_count, _Outptr_ OrtValue** out);
@@ -2766,7 +2798,7 @@ struct OrtApi {
    * "initial_growth_chunk_size_bytes": (Possible) Size of the second allocation in the arena.
    *  Only relevant if arena strategy is `kNextPowerOfTwo`. Use -1 to allow ORT to choose the default.
    * "max_power_of_two_extend_bytes": The maximum enxtend size if arena strategy is `kNextPowerOfTwo`.
-   *  It is not an allocation limit, it is only a limit for extention when requested byte is less than the limit.
+   *  It is not an allocation limit, it is only a limit for extension when requested byte is less than the limit.
    *  When requested bytes is more than the limit, allocator will still return as requested.
    *  Use -1 to allow ORT to choose the default 1GB for max_power_of_two_extend_bytes.
    *  Ultimately, the allocation size is determined by the allocation memory request.
@@ -2914,7 +2946,7 @@ struct OrtApi {
    *
    * Please refer to https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#cc
    * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtTensorRTProviderOptionsV2
-   * and value should be its related range.
+   * and value should be its related range. Recreates the options and only sets the supplied values.
    *
    * For example, key="trt_max_workspace_size" and value="2147483648"
    *
@@ -3410,7 +3442,7 @@ struct OrtApi {
    *
    * Please refer to https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#configuration-options
    * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtCUDAProviderOptionsV2
-   * and value should be its related range.
+   * and value should be its related range. Recreates the options and only sets the supplied values.
    *
    * For example, key="device_id" and value="0"
    *
@@ -3483,15 +3515,15 @@ struct OrtApi {
    * \param[in] options
    * \param[in] initializer_names Array of null terminated UTF-8 encoded strings of the initializers names.
    * \param[in] initializers Array of ::OrtValue type
-   * \param[in] initializers_num Number of elements in the initializer_names and initializers
+   * \param[in] num_initializers Number of elements in the initializer_names and initializers
    *
    * \snippet{doc} snippets.dox OrtStatus Return Value
    *
    * \since Version 1.12.
    */
   ORT_API2_STATUS(AddExternalInitializers, _In_ OrtSessionOptions* options,
-                  _In_reads_(input_len) const char* const* initializer_names,
-                  _In_reads_(input_len) const OrtValue* const* initializers, size_t initializers_num);
+                  _In_reads_(num_initializers) const char* const* initializer_names,
+                  _In_reads_(num_initializers) const OrtValue* const* initializers, size_t num_initializers);
 
   /** \brief: Create attribute of onnxruntime operator
    *
@@ -3595,6 +3627,7 @@ struct OrtApi {
    * QNN supported keys:
    *   "backend_path": file path to QNN backend library.
    *   "profiling_level": QNN profiling level, options: "off", "basic", "detailed". Default to off.
+   *   "profiling_file_path": QNN profiling file path if ETW not enabled.
    *   "rpc_control_latency": QNN RPC control latency.
    *   "vtcm_mb": QNN VTCM size in MB. default to 0(not set).
    *   "htp_performance_mode": QNN performance mode, options: "burst", "balanced", "default", "high_performance",
@@ -3616,6 +3649,10 @@ struct OrtApi {
    *     - "73"
    *     - "75"
    *   "device_id": The ID of the device to use when setting 'htp_arch'. Defaults to "0" (for single device).
+       "enable_htp_fp16_precision": Only used for float32 model.
+       Enable the float32 model to be inferenced with fp16 precision. Otherwise, it will be fp32 precision.
+         - "0": Default. With fp32 precision.
+         - "1": With fp16 precision.
    *
    * SNPE supported keys:
    *   "runtime": SNPE runtime engine, options: "CPU", "CPU_FLOAT32", "GPU", "GPU_FLOAT32_16_HYBRID", "GPU_FLOAT16",
@@ -4430,13 +4467,14 @@ struct OrtApi {
    * E.g. a cuda stream or a cublas handle
    *
    * \param context - Kernel context
-   * \param resouce_version - Version of the resource
+   * \param resource_version - Version of the resource
    * \param resource_id - Type of resource
    * \param resource - A pointer to returned resource
    *
    * \since Version 1.16.
    */
-  ORT_API2_STATUS(KernelContext_GetResource, _In_ const OrtKernelContext* context, _In_ int resouce_version, _In_ int resource_id, _Outptr_ void** resource);
+  ORT_API2_STATUS(KernelContext_GetResource, _In_ const OrtKernelContext* context, _In_ int resource_version,
+                  _In_ int resource_id, _Outptr_ void** resource);
 
   /** \brief Set user logging function
    *
@@ -4491,10 +4529,10 @@ struct OrtApi {
   ORT_API2_STATUS(ShapeInferContext_GetAttribute, _In_ const OrtShapeInferContext* context, _In_ const char* attr_name, _Outptr_ const OrtOpAttr** attr);
 
   /**
-   * Set type and shape info of an ouput
+   * Set type and shape info of an output
    *
    * \param[in] context
-   * \param[in] index The index of the ouput
+   * \param[in] index The index of the output
    * \param[out] info Type shape info of the output
    *
    * \since Version 1.17.
@@ -4566,6 +4604,68 @@ struct OrtApi {
                   _In_reads_(num_keys) const char* const* provider_options_keys,
                   _In_reads_(num_keys) const char* const* provider_options_values,
                   _In_ size_t num_keys);
+
+  /** \brief Append VitisAI provider to session options
+   *
+   * If VitisAI is not available (due to a non VitisAI enabled build, or if VitisAI is not installed on the system), this function will return failure.
+   *
+   * \param[in] options
+   * \param[in] provider_options_keys
+   * \param[in] provider_options_values
+   * \param[in] num_keys
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_VitisAI,
+                  _In_ OrtSessionOptions* options,
+                  _In_reads_(num_keys) const char* const* provider_options_keys,
+                  _In_reads_(num_keys) const char* const* provider_options_values,
+                  _In_ size_t num_keys);
+
+  /** \brief Get scratch buffer from the corresponding allocator under the sepcific OrtMemoryInfo object.
+   *         NOTE: callers are responsible to release this scratch buffer from the corresponding allocator
+   *  \param[in] context OrtKernelContext instance
+   *  \param[in] mem_info OrtMemoryInfo instance
+   *  \param[in] count_or_bytes How many bytes is this scratch buffer
+   *  \param[out] out A pointer to the scrach buffer
+   *  \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(KernelContext_GetScratchBuffer, _In_ const OrtKernelContext* context, _In_ const OrtMemoryInfo* mem_info, _In_ size_t count_or_bytes, _Outptr_ void** out);
+
+  /** \brief Get allocator from KernelInfo for a specific memory type. Please use C API ReleaseAllocator to release out object
+   *
+   * \param[in] info OrtKernelInfo instance
+   * \param[in] mem_type OrtMemType object
+   * \param[out] out A pointer to OrtAllocator
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(KernelInfoGetAllocator, _In_ const OrtKernelInfo* info, _In_ OrtMemType mem_type, _Outptr_ OrtAllocator** out);
+
+  /** \brief Replace initialized Tensors with external data with the provided files in memory
+   *
+   * The function will find the initialized TensorProtos with external data in the graph with the provided
+   * external file names and the file content in memory. The API gets the external file name, offset, data length
+   * from TensorProto, and locate the tensor data from the file in memory buffer.
+   * It creates a Tensor to replace the existing Tensor in graph. The replacement
+   * will occur before any of the optimizations take place. The data will be copied into the graph
+   * since TensorProto can't refer to the user provided buffers.
+   *
+   * \param[in] options
+   * \param[in] external_initializer_file_names Array of null terminated UTF-8 encoded strings of the file names
+   *            which holds the external initializers.
+   * \param[in] external_initializer_file_buffer_array Array of pointers to the buffer of the file content.
+   *            The buffer can be freed after session creation.
+   * \param[in] external_initializer_file_lengths Array of size_t to indicate the length of file content
+   * \param[in] num_external_initializer_files Number of external files
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(AddExternalInitializersFromFilesInMemory, _In_ OrtSessionOptions* options,
+                  _In_reads_(num_external_initializer_files) const ORTCHAR_T* const* external_initializer_file_names,
+                  _In_reads_(num_external_initializer_files) char* const* external_initializer_file_buffer_array,
+                  _In_reads_(num_external_initializer_files) const size_t* external_initializer_file_lengths,
+                  size_t num_external_initializer_files);
 };
 
 /*
@@ -4663,6 +4763,21 @@ struct OrtCustomOp {
   // Get start range
   int(ORT_API_CALL* GetStartVersion)(_In_ const struct OrtCustomOp* op);
   int(ORT_API_CALL* GetEndVersion)(_In_ const struct OrtCustomOp* op);
+
+  // Get the inplace_map that defines which output can reuse which input
+  // Callers will provide 2 raw int* and pass in their address, this function will fill these 2 arrays
+  // when return, output (*output_index)[i] may reuse the input (*input_index[i]).
+  // The return value is the size of these 2 arrays.
+  // Callers are responsible to delete these 2 arrays after use by calling OrtCustomOp::ReleaseMayInplace().
+  size_t(ORT_API_CALL* GetMayInplace)(_Out_ int** input_index, _Out_ int** output_index);
+
+  // Release the pointer input_index and output_index allocated from GetMayInplace() function.
+  // If GetMayInplace() is defined, this function MUST be defined as well.
+  void(ORT_API_CALL* ReleaseMayInplace)(_Frees_ptr_opt_ int* input_index, _Frees_ptr_opt_ int* output_index);
+
+  // Same as GetMayInplace() and ReleaseMayInplace()
+  size_t(ORT_API_CALL* GetAliasMap)(_Out_ int** input_index, _Out_ int** output_index);
+  void(ORT_API_CALL* ReleaseAliasMap)(_Frees_ptr_opt_ int* input_index, _Frees_ptr_opt_ int* output_index);
 };
 
 /*

package/Plugins/iOS~/onnxruntime.xcframework/macos-arm64_x86_64/onnxruntime.framework/{Headers → Versions/A/Headers}/onnxruntime_cxx_api.h

@@ -873,6 +873,9 @@ struct SessionOptionsImpl : ConstSessionOptionsImpl<T> {
 
   SessionOptionsImpl& AddInitializer(const char* name, const OrtValue* ort_val);                                             ///< Wraps OrtApi::AddInitializer
   SessionOptionsImpl& AddExternalInitializers(const std::vector<std::string>& names, const std::vector<Value>& ort_values);  ///< Wraps OrtApi::AddExternalInitializers
+  SessionOptionsImpl& AddExternalInitializersFromFilesInMemory(const std::vector<std::basic_string<ORTCHAR_T>>& external_initializer_file_names,
+                                                               const std::vector<char*>& external_initializer_file_buffer_array,
+                                                               const std::vector<size_t>& external_initializer_file_lengths);  ///< Wraps OrtApi::AddExternalInitializersFromFilesInMemory
 
   SessionOptionsImpl& AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options);          ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA
   SessionOptionsImpl& AppendExecutionProvider_CUDA_V2(const OrtCUDAProviderOptionsV2& provider_options);     ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA_V2
@@ -901,6 +904,9 @@ struct SessionOptionsImpl : ConstSessionOptionsImpl<T> {
   SessionOptionsImpl& RegisterCustomOpsLibrary(const ORTCHAR_T* library_name, const CustomOpConfigs& custom_op_configs = {});
 
   SessionOptionsImpl& RegisterCustomOpsUsingFunction(const char* function_name);  ///< Wraps OrtApi::RegisterCustomOpsUsingFunction
+
+  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_VitisAI
+  SessionOptionsImpl& AppendExecutionProvider_VitisAI(const std::unordered_map<std::string, std::string>& provider_options = {});
 };
 }  // namespace detail
 
@@ -2052,7 +2058,11 @@ struct KernelContext {
   explicit KernelContext(OrtKernelContext* context);
   size_t GetInputCount() const;
   size_t GetOutputCount() const;
+  // If input is optional and is not present, the method returns en empty ConstValue
+  // which can be compared to nullptr.
   ConstValue GetInput(size_t index) const;
+  // If outout is optional and is not present, the method returns en empty UnownedValue
+  // which can be compared to nullptr.
   UnownedValue GetOutput(size_t index, const int64_t* dim_values, size_t dim_count) const;
   UnownedValue GetOutput(size_t index, const std::vector<int64_t>& dims) const;
   void* GetGPUComputeStream() const;
@@ -2165,8 +2175,8 @@ struct Op : detail::Base<OrtOp> {
 /// </summary>
 struct ShapeInferContext {
   struct SymbolicInteger {
-    SymbolicInteger(int64_t i) : i_(i), is_int_(true){};
-    SymbolicInteger(const char* s) : s_(s), is_int_(false){};
+    SymbolicInteger(int64_t i) : i_(i), is_int_(true) {};
+    SymbolicInteger(const char* s) : s_(s), is_int_(false) {};
     SymbolicInteger(const SymbolicInteger&) = default;
     SymbolicInteger(SymbolicInteger&&) = default;
 
@@ -2206,7 +2216,7 @@ struct ShapeInferContext {
 
   size_t GetInputCount() const { return input_shapes_.size(); }
 
-  Status SetOutputShape(size_t indice, const Shape& shape);
+  Status SetOutputShape(size_t indice, const Shape& shape, ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT);
 
   int64_t GetAttrInt(const char* attr_name);
 
@@ -2294,6 +2304,11 @@ struct CustomOpBase : OrtCustomOp {
     OrtCustomOp::GetEndVersion = [](const OrtCustomOp* this_) {
       return static_cast<const TOp*>(this_)->end_ver_;
     };
+
+    OrtCustomOp::GetMayInplace = nullptr;
+    OrtCustomOp::ReleaseMayInplace = nullptr;
+    OrtCustomOp::GetAliasMap = nullptr;
+    OrtCustomOp::ReleaseAliasMap = nullptr;
   }
 
   // Default implementation of GetExecutionProviderType that returns nullptr to default to the CPU provider

package/Plugins/iOS~/onnxruntime.xcframework/macos-arm64_x86_64/onnxruntime.framework/{Headers → Versions/A/Headers}/onnxruntime_cxx_inline.h

@@ -7,17 +7,27 @@
 // These are the inline implementations of the C++ header APIs. They're in this separate file as to not clutter
 // the main C++ file with implementation details.
 
-#include <cstring>
+#include <algorithm>
 #include <functional>
-
-#define RETURN_ON_API_FAIL(expression) \
-  {                                    \
-    auto err = (expression);           \
-    if (err) {                         \
-      return Status(err);              \
-    }                                  \
+#include <iterator>
+#include <type_traits>
+
+// Convert OrtStatus to Ort::Status and return
+// instead of throwing
+#define ORT_CXX_RETURN_ON_API_FAIL(expression) \
+  {                                            \
+    auto ort_status = (expression);            \
+    if (ort_status) {                          \
+      return Ort::Status(ort_status);          \
+    }                                          \
   }
 
+#ifdef __cpp_if_constexpr
+#define ORT_CXX_IF_CONSTEXPR if constexpr
+#else
+#define ORT_CXX_IF_CONSTEXPR if
+#endif
+
 namespace Ort {
 
 namespace detail {
@@ -771,6 +781,27 @@ inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddExternalInitializers(con
   return *this;
 }
 
+template <typename T>
+inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddExternalInitializersFromFilesInMemory(const std::vector<std::basic_string<ORTCHAR_T>>& file_names,
+                                                                                              const std::vector<char*>& buffer_array,
+                                                                                              const std::vector<size_t>& file_lengths) {
+  const size_t inputs_num = file_names.size();
+  if (inputs_num != buffer_array.size()) {
+    ORT_CXX_API_THROW("Expecting names and buffer_array to have the same length", ORT_INVALID_ARGUMENT);
+  }
+  if (inputs_num != file_lengths.size()) {
+    ORT_CXX_API_THROW("Expecting names and file_lengths to have the same length", ORT_INVALID_ARGUMENT);
+  }
+  std::vector<const ORTCHAR_T*> names_ptr;
+  names_ptr.reserve(inputs_num);
+  for (size_t i = 0; i < inputs_num; ++i) {
+    names_ptr.push_back(file_names[i].c_str());
+  }
+  ThrowOnError(GetApi().AddExternalInitializersFromFilesInMemory(this->p_, names_ptr.data(), buffer_array.data(),
+                                                                 file_lengths.data(), inputs_num));
+  return *this;
+}
+
 template <typename T>
 inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options) {
   ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CUDA(this->p_, &provider_options));
@@ -885,6 +916,25 @@ inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_Ope
   return *this;
 }
 
+template <typename T>
+inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_VitisAI(const std::unordered_map<std::string, std::string>& provider_options) {
+  auto num_entries = provider_options.size();
+  std::vector<const char*> keys, values;
+  if (num_entries > 0) {
+    keys.reserve(num_entries);
+    values.reserve(num_entries);
+
+    for (const auto& entry : provider_options) {
+      keys.push_back(entry.first.c_str());
+      values.push_back(entry.second.c_str());
+    }
+  }
+
+  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_VitisAI(this->p_, keys.data(), values.data(), num_entries));
+
+  return *this;
+}
+
 template <typename T>
 inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::RegisterCustomOpsLibrary(const ORTCHAR_T* library_name,
                                                                               const CustomOpConfigs& custom_op_configs) {
@@ -1932,7 +1982,9 @@ inline ShapeInferContext::ShapeInferContext(const OrtApi* ort_api,
     TensorTypeAndShapeInfo type_shape_info(info);
     auto integer_shape = type_shape_info.GetShape();
     std::vector<const char*> symbolic_shape(integer_shape.size(), {});
-    type_shape_info.GetSymbolicDimensions(&symbolic_shape[0], integer_shape.size());
+    if (!integer_shape.empty()) {
+      type_shape_info.GetSymbolicDimensions(&symbolic_shape[0], integer_shape.size());
+    }
     Shape shape;
     for (size_t ith = 0; ith < integer_shape.size(); ++ith) {
       if (symbolic_shape[ith] && std::string{symbolic_shape[ith]}.size() > 0) {
@@ -1946,9 +1998,10 @@ inline ShapeInferContext::ShapeInferContext(const OrtApi* ort_api,
   }
 }
 
-inline Status ShapeInferContext::SetOutputShape(size_t indice, const Shape& shape) {
+inline Status ShapeInferContext::SetOutputShape(size_t indice, const Shape& shape, ONNXTensorElementDataType type) {
   OrtTensorTypeAndShapeInfo* info = {};
-  RETURN_ON_API_FAIL(ort_api_->CreateTensorTypeAndShapeInfo(&info));
+  ORT_CXX_RETURN_ON_API_FAIL(ort_api_->CreateTensorTypeAndShapeInfo(&info));
+  ORT_CXX_RETURN_ON_API_FAIL(ort_api_->SetTensorElementType(info, type));
 
   using InfoPtr = std::unique_ptr<OrtTensorTypeAndShapeInfo, std::function<void(OrtTensorTypeAndShapeInfo*)>>;
 
@@ -1961,7 +2014,7 @@ inline Status ShapeInferContext::SetOutputShape(size_t indice, const Shape& shap
 
   for (const auto dim : shape) {
     if (dim.IsInt()) {
-      integer_dims.push_back(dim.IsInt());
+      integer_dims.push_back(dim.AsInt());
       symbolic_dims.push_back("");
     } else {
       if (!dim.AsSym() || std::string{dim.AsSym()}.empty()) {
@@ -1972,9 +2025,9 @@ inline Status ShapeInferContext::SetOutputShape(size_t indice, const Shape& shap
     }
   }
 
-  RETURN_ON_API_FAIL(ort_api_->SetDimensions(info, integer_dims.data(), integer_dims.size()));
-  RETURN_ON_API_FAIL(ort_api_->SetSymbolicDimensions(info, symbolic_dims.data(), symbolic_dims.size()));
-  RETURN_ON_API_FAIL(ort_api_->ShapeInferContext_SetOutputTypeShape(ctx_, indice, info));
+  ORT_CXX_RETURN_ON_API_FAIL(ort_api_->SetDimensions(info, integer_dims.data(), integer_dims.size()));
+  ORT_CXX_RETURN_ON_API_FAIL(ort_api_->SetSymbolicDimensions(info, symbolic_dims.data(), symbolic_dims.size()));
+  ORT_CXX_RETURN_ON_API_FAIL(ort_api_->ShapeInferContext_SetOutputTypeShape(ctx_, indice, info));
   return Status{nullptr};
 }