[QNN-EP] Implement file mapped weights feature

onnxruntime/core/providers/qnn/builder/onnx_ctx_model_helper.cc

@@ -94,6 +94,7 @@ Status GetEpContextFromMainNode(const onnxruntime::Node& main_context_node,
     const std::string& context_binary = node_helper.Get(EP_CACHE_CONTEXT, "");
     return qnn_backend_manager->LoadCachedQnnContextFromBuffer(const_cast<char*>(context_binary.c_str()),
                                                                static_cast<uint64_t>(context_binary.length()),
+                                                               "",
                                                                main_context_node.Name(),
                                                                qnn_models,
                                                                max_spill_fill_size);
@@ -127,6 +128,18 @@ Status GetEpContextFromMainNode(const onnxruntime::Node& main_context_node,
     return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_GRAPH, "The file path in ep_cache_context does not exist or is not accessible.");
   }
 
+  std::string context_binary_path_str = context_binary_path.string();
+#ifdef QNN_FILE_MAPPED_WEIGHTS_ENABLED
+  if (qnn_backend_manager->FileMappingIsEnabled()) {
+    return qnn_backend_manager->LoadCachedQnnContextFromBuffer(nullptr,
+                                                               0,
+                                                               context_binary_path_str,
+                                                               main_context_node.Name(),
+                                                               qnn_models,
+                                                               max_spill_fill_size);
+  }
+#endif
+
   size_t buffer_size{0};
   std::ifstream cache_file(context_binary_path.string().c_str(), std::ifstream::binary);
   ORT_RETURN_IF(!cache_file || !cache_file.good(), "Failed to open cache file.");
@@ -144,6 +157,7 @@ Status GetEpContextFromMainNode(const onnxruntime::Node& main_context_node,
   cache_file.close();
   return qnn_backend_manager->LoadCachedQnnContextFromBuffer(buffer.get(),
                                                              static_cast<uint64_t>(buffer_size),
+                                                             context_binary_path_str,
                                                              main_context_node.Name(),
                                                              qnn_models,
                                                              max_spill_fill_size);

onnxruntime/core/providers/qnn/builder/qnn_backend_manager.h

@@ -32,6 +32,10 @@
 #include "core/providers/qnn/builder/qnn_profile_serializer.h"
 #include "core/providers/qnn/builder/qnn_node_group/qnn_node_group.h"
 
+#ifdef QNN_FILE_MAPPED_WEIGHTS_ENABLED
+#include "core/providers/qnn/builder/qnn_file_mapping_callback_interface.h"
+#endif
+
 namespace onnxruntime {
 namespace qnn {
 
@@ -154,6 +158,7 @@ class QnnBackendManager : public std::enable_shared_from_this<QnnBackendManager>
   std::unique_ptr<unsigned char[]> GetContextBinaryBuffer(uint64_t& written_buffer_size);
 
   Status LoadCachedQnnContextFromBuffer(char* buffer, uint64_t buffer_length,
+                                        const std::string& context_bin_filepath,
                                         std::string node_name,
                                         std::unordered_map<std::string, std::unique_ptr<qnn::QnnModel>>& qnn_models,
                                         int64_t max_spill_fill_size);
@@ -163,6 +168,7 @@ class QnnBackendManager : public std::enable_shared_from_this<QnnBackendManager>
   Status SetupBackend(const logging::Logger& logger, bool load_from_cached_context,
                       bool need_load_system_lib, bool share_ep_contexts,
                       bool enable_vtcm_backup_buffer_sharing,
+                      bool enable_file_mapped_weights,
                       std::unordered_map<std::string, std::unique_ptr<std::vector<std::string>>>& context_bin_map);
 
   Status CreateHtpPowerCfgId(uint32_t deviceId, uint32_t coreId, uint32_t& htp_power_config_id);
@@ -246,7 +252,14 @@ class QnnBackendManager : public std::enable_shared_from_this<QnnBackendManager>
   bool ProfilingEnabled() { return profiling_enabled_; }
 #endif
 
+  bool FileMappingIsEnabled() { return file_mapped_weights_enabled_; }
+
  private:
+  typedef struct BufferInfo {
+    std::unique_ptr<char[]> data;
+    size_t size;
+  } BufferInfo_t;
+
   Status LoadBackend();
 
   Status InitializeBackend();
@@ -263,9 +276,23 @@ class QnnBackendManager : public std::enable_shared_from_this<QnnBackendManager>
 
   Status CreateContext(bool enable_htp_weight_sharing);
 
+  Status ReadContextBinIfValid(const std::string& context_bin_filepath,
+                               BufferInfo_t& buffer_info,
+                               bool read_file_contents);
+
   Status CreateContextVtcmBackupBufferSharingEnabled(std::unordered_map<std::string,
                                                                         std::unique_ptr<std::vector<std::string>>>& context_bin_map);
 
+  Status CreateContextFromListAsyncV1(const QnnContext_Config_t** configs,
+                                      std::unordered_map<std::string,
+                                                         std::unique_ptr<std::vector<std::string>>>& context_bin_map);
+
+#ifdef QNN_FILE_MAPPED_WEIGHTS_ENABLED
+  Status CreateContextFromListAsyncV2(const QnnContext_Config_t** configs,
+                                      std::unordered_map<std::string,
+                                                         std::unique_ptr<std::vector<std::string>>>& context_bin_map);
+#endif
+
   Status ReleaseContext();
 
   // Sets the ORT logger and creates a corresponding QNN logger with the same log level.
@@ -451,6 +478,15 @@ class QnnBackendManager : public std::enable_shared_from_this<QnnBackendManager>
   bool context_created_ = false;
   bool backend_setup_completed_ = false;
   bool vtcm_backup_buffer_sharing_enabled_ = false;
+  bool file_mapped_weights_enabled_ = false;
+
+#ifdef QNN_FILE_MAPPED_WEIGHTS_ENABLED
+  std::shared_ptr<FileMappingCallbackInterface> file_mapper_ = nullptr;
+  // Notify params for file mapping must persist throughout lifetime of
+  // QnnBackendManager for release of DMA data callback on destruction
+  std::vector<std::pair<FileMappingCallbackInterface*, void*>> file_mapping_notify_params_;
+#endif
+
   // NPU backend requires quantized model
   QnnBackendType qnn_backend_type_ = QnnBackendType::CPU;
   Qnn_ProfileHandle_t profile_backend_handle_ = nullptr;

onnxruntime/core/providers/qnn/builder/qnn_def.h

@@ -19,6 +19,12 @@ namespace qnn {
 #define QNN_SYSTEM_PROFILE_API_ENABLED
 #endif
 
+#if defined(_WIN32) && (defined(__aarch64__) || defined(_M_ARM64))
+#if QNN_API_VERSION_MAJOR > 2 || ((QNN_API_VERSION_MAJOR) == 2 && (QNN_API_VERSION_MINOR >= 32))
+#define QNN_FILE_MAPPED_WEIGHTS_ENABLED
+#endif
+#endif
+
 // QNN only support subset of POSIX of dlopen/dlsym/dladdr/dlerror/dlclose
 // except the following flags for dlopen, others should be done only
 // when we really need them

onnxruntime/core/providers/qnn/builder/qnn_file_mapping_callback_interface.h

@@ -0,0 +1,45 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+
+#include "core/providers/qnn/builder/qnn_def.h"
+#ifdef QNN_FILE_MAPPED_WEIGHTS_ENABLED
+
+#include <string>
+
+#include <QnnContext.h>
+
+#include "core/providers/qnn/ort_api.h"
+
+namespace onnxruntime {
+namespace qnn {
+
+class FileMappingCallbackInterface {
+ public:
+  virtual ~FileMappingCallbackInterface() = default;
+  virtual Status MapContextBin(const std::string& bin_filepath,
+                               void** notify_param) = 0;
+  virtual Status ReleaseContextBin(const std::string& model_name) = 0;
+
+  virtual Status GetContextBinMappingPointer(const std::string& bin_filepath, void** mapping_ptr) = 0;
+
+  virtual Status FreeContextBinMappingPointer(LPVOID bin_mapping_pointer) = 0;
+
+  virtual Qnn_ErrorHandle_t MapDmaData(Qnn_ContextBinaryDataRequest_t request,
+                                       Qnn_ContextBinaryDmaDataResponse_t* response,
+                                       void* notify_param) = 0;
+  virtual Qnn_ErrorHandle_t ReleaseDmaData(Qnn_ContextBinaryDmaDataMem_t data_mem,
+                                           void* notify_param) = 0;
+
+  virtual Qnn_ErrorHandle_t MapRawData(Qnn_ContextBinaryDataRequest_t request,
+                                       Qnn_ContextBinaryRawDataResponse_t* response,
+                                       void* notify_param) = 0;
+  virtual Qnn_ErrorHandle_t ReleaseRawData(Qnn_ContextBinaryRawDataMem_t data_mem,
+                                           void* notify_param) = 0;
+};
+
+}  // namespace qnn
+}  // namespace onnxruntime
+
+#endif  // QNN_FILE_MAPPED_WEIGHTS_ENABLED