Merge remote-tracking branch 'origin/main' into snnn/replace_pool

.github/workflows/windows.yml

@@ -37,7 +37,7 @@ jobs:
       - name: Delete build folder
         run: |
           if (Test-Path D:\b) { Remove-Item -Recurse -Force D:\b }
-          &tools\ci_build\github\windows\install_third_party_deps.ps1 -cpu_arch x64 -install_prefix D:\b\Debug\installed -build_config Debug
+
 
       # The build machine doesn't have a GPU. So the value of CMAKE_CUDA_ARCHITECTURES doesn't matter.
       - name: Build code

cgmanifests/generated/cgmanifest.json

@@ -346,7 +346,7 @@
       "component": {
         "type": "git",
         "git": {
-          "commitHash": "12a3b24c456cebd9fd11f23ac0164f78129b00c6",
+          "commitHash": "b9b4a37041dec3dd62ac92014a6cc1aece48d9f3",
           "repositoryUrl": "https://github.com/google/dawn.git"
         },
         "comments": "dawn"

cmake/CMakeLists.txt

@@ -93,6 +93,7 @@ option(onnxruntime_USE_OPENVINO "Build with OpenVINO support" OFF)
 option(onnxruntime_USE_COREML "Build with CoreML support" OFF)
 option(onnxruntime_USE_NNAPI_BUILTIN "Build with builtin NNAPI lib for Android NNAPI support" OFF)
 option(onnxruntime_USE_QNN "Build with QNN support" OFF)
+option(onnxruntime_BUILD_QNN_EP_STATIC_LIB "Build with QNN EP as a static library" OFF)
 option(onnxruntime_USE_SNPE "Build with SNPE support" OFF)
 option(onnxruntime_USE_RKNPU "Build with RKNPU support" OFF)
 option(onnxruntime_USE_DNNL "Build with DNNL support" OFF)
@@ -240,6 +241,9 @@ option(onnxruntime_ENABLE_CPUINFO "Enable cpuinfo" ON)
 # ATen fallback support
 option(onnxruntime_ENABLE_ATEN "Enable ATen fallback" OFF)
 
+# dlpack support
+cmake_dependent_option(onnxruntime_ENABLE_DLPACK "Enable dlpack" ON "onnxruntime_ENABLE_TRAINING OR onnxruntime_ENABLE_ATEN OR onnxruntime_ENABLE_PYTHON" OFF)
+
 # Triton support
 option(onnxruntime_ENABLE_TRITON "Enable Triton" OFF)
 
@@ -258,6 +262,12 @@ option(onnxruntime_USE_AZURE "Build with azure inferencing support" OFF)
 option(onnxruntime_USE_LOCK_FREE_QUEUE "Build with lock-free task queue for threadpool." OFF)
 option(onnxruntime_FORCE_GENERIC_ALGORITHMS "Disable optimized arch-specific algorithms. Use only for testing and debugging generic algorithms." OFF)
 
+option(onnxruntime_USE_TENSORRT_INTERFACE "Build ONNXRuntime shared lib which is compatible with TensorRT EP interface" OFF)
+option(onnxruntime_USE_CUDA_INTERFACE "Build ONNXRuntime shared lib which is compatible with Cuda EP interface" OFF)
+option(onnxruntime_USE_OPENVINO_INTERFACE "Build ONNXRuntime shared lib which is compatible with OpenVINO EP interface" OFF)
+option(onnxruntime_USE_VITISAI_INTERFACE "Build ONNXRuntime shared lib which is compatible with Vitis-AI EP interface" OFF)
+option(onnxruntime_USE_QNN_INTERFACE "Build ONNXRuntime shared lib which is compatible with QNN EP interface" OFF)
+
 # ENABLE_TRAINING includes all training functionality
 # The following 2 entry points
 # 1. ORTModule
@@ -702,7 +712,7 @@ if (WIN32)
     # structure was padded due to __declspec(align())
     list(APPEND ORT_WARNING_FLAGS "/wd4324")
     # warning C4800: Implicit conversion from 'X' to bool. Possible information loss
-    if (onnxruntime_USE_OPENVINO)
+    if (onnxruntime_USE_OPENVINO OR onnxruntime_USE_OPENVINO_INTERFACE)
        list(APPEND ORT_WARNING_FLAGS "/wd4800")
     endif()
     # operator 'operator-name': deprecated between enumerations of different types
@@ -863,7 +873,7 @@ else()
   set(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION OFF)
 endif()
 
-if (onnxruntime_USE_CUDA)
+if (onnxruntime_USE_CUDA OR onnxruntime_USE_CUDA_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_CUDA=1)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_CUDA=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES cuda)
@@ -887,7 +897,7 @@ if (onnxruntime_USE_CUDA)
     endif()
 endif()
 
-if (onnxruntime_USE_VITISAI)
+if (onnxruntime_USE_VITISAI OR onnxruntime_USE_VITISAI_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_VITISAI=1)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_VITISAI=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES vitisai)
@@ -897,12 +907,12 @@ if (onnxruntime_USE_DNNL)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES dnnl)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_DNNL=1)
 endif()
-if (onnxruntime_USE_OPENVINO)
+if (onnxruntime_USE_OPENVINO OR onnxruntime_USE_OPENVINO_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_OPENVINO=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES openvino)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_OPENVINO=1)
 endif()
-if (onnxruntime_USE_TENSORRT)
+if (onnxruntime_USE_TENSORRT OR onnxruntime_USE_TENSORRT_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_TENSORRT=1)
     #TODO: remove the following line and change the test code in onnxruntime_shared_lib_test to use the new EP API.
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES tensorrt)
@@ -928,7 +938,7 @@ if (onnxruntime_USE_JSEP)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_JSEP=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES js)
 endif()
-if (onnxruntime_USE_QNN)
+if (onnxruntime_USE_QNN OR onnxruntime_USE_QNN_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_QNN=1)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_QNN=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES qnn)
@@ -956,7 +966,7 @@ if (onnxruntime_USE_QNN)
       endif()
     endif()
 
-    if (MSVC OR ${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
+    if ((NOT onnxruntime_USE_QNN_INTERFACE) AND (MSVC OR ${CMAKE_SYSTEM_NAME} STREQUAL "Linux"))
       file(GLOB QNN_LIB_FILES LIST_DIRECTORIES false "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/libQnn*.so"
 	                                     "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/Qnn*.dll"
                                              "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/libHtpPrepare.so"
@@ -1415,7 +1425,7 @@ if (onnxruntime_ENABLE_TRAINING_APIS)
   )
 endif()
 
-if (onnxruntime_USE_OPENVINO)
+if (onnxruntime_USE_OPENVINO OR onnxruntime_USE_OPENVINO_INTERFACE)
 
   add_definitions(-DUSE_OPENVINO=1)
 
@@ -1428,7 +1438,7 @@ if (onnxruntime_USE_OPENVINO)
     add_definitions(-DOPENVINO_CONFIG_GPU=1)
   endif()
 
-  if (onnxruntime_USE_OPENVINO_CPU)
+  if (onnxruntime_USE_OPENVINO_CPU OR onnxruntime_USE_OPENVINO_INTERFACE) # OpenVino CPU interface is default built.
     add_definitions(-DOPENVINO_CONFIG_CPU=1)
   endif()
 
@@ -1596,6 +1606,10 @@ if (onnxruntime_ENABLE_TRAINING)
   list(APPEND onnxruntime_EXTERNAL_LIBRARIES tensorboard)
 endif()
 
+if (onnxruntime_ENABLE_DLPACK)
+  add_compile_definitions(ENABLE_DLPACK)
+endif()
+
 if (UNIX OR onnxruntime_USE_NCCL)
   # MPI is INDEPENDENT of NCCL for now. You can build NCLL without MPI and launch multi-GPU with your own launcher.
   if (onnxruntime_USE_MPI)

cmake/deps.txt

@@ -58,5 +58,5 @@ extensions;https://github.com/microsoft/onnxruntime-extensions/archive/f3f6caa6e
 composable_kernel;https://github.com/ROCmSoftwarePlatform/composable_kernel/archive/204da9c522cebec5220bba52cd3542ebcaf99e7a.zip;1827348efd47831c13074245274d41b7cae8a557
 directx_headers;https://github.com/microsoft/DirectX-Headers/archive/refs/tags/v1.613.1.zip;47653509a3371eabb156360f42faf582f314bf2e
 cudnn_frontend;https://github.com/NVIDIA/cudnn-frontend/archive/refs/tags/v1.7.0.zip;d0753d8d5b39947ca0729d7773cb84653a129eb1
-dawn;https://github.com/google/dawn/archive/12a3b24c456cebd9fd11f23ac0164f78129b00c6.zip;ad428f6dc16f1336d584f7bad5714e1097dafc43
+dawn;https://github.com/google/dawn/archive/b9b4a37041dec3dd62ac92014a6cc1aece48d9f3.zip;e8b8c2ebabdedb7c57d931fc4a19ae22146d31e1
 kleidiai;https://gitlab.arm.com/kleidi/kleidiai/-/archive/d15722976120710080ca098fe8ddabf4556cb40f/kleidiai-d15722976120710080ca098fe8ddabf4556cb40f.zip;d6c840d00c3b05aedf06e957ddaece1013d1f40b

cmake/external/abseil-cpp.cmake

@@ -32,7 +32,13 @@ FetchContent_Declare(
 
 onnxruntime_fetchcontent_makeavailable(abseil_cpp)
 FetchContent_GetProperties(abseil_cpp)
-set(ABSEIL_SOURCE_DIR ${abseil_cpp_SOURCE_DIR})
+if(abseil_cpp_SOURCE_DIR)
+  set(ABSEIL_SOURCE_DIR ${abseil_cpp_SOURCE_DIR})
+  if(onnxruntime_USE_WEBGPU)
+    set(DAWN_ABSEIL_DIR ${abseil_cpp_SOURCE_DIR})
+  endif()
+endif()
+
 # abseil_cpp_SOURCE_DIR is non-empty if we build it from source
 message(STATUS "Abseil source dir:" ${ABSEIL_SOURCE_DIR})
 # abseil_cpp_VERSION  is non-empty if we find a preinstalled ABSL

cmake/external/onnxruntime_external_deps.cmake

@@ -222,6 +222,11 @@ onnxruntime_fetchcontent_makeavailable(Protobuf)
 if(Protobuf_FOUND)
   message(STATUS "Protobuf version: ${Protobuf_VERSION}")
 else()
+  if(protobuf_SOURCE_DIR)
+    if(onnxruntime_USE_WEBGPU)
+      set(DAWN_PROTOBUF_DIR ${protobuf_SOURCE_DIR})
+    endif()
+  endif()
   # Adjust warning flags
   if (TARGET libprotoc)
     if (NOT MSVC)
@@ -571,8 +576,8 @@ if (onnxruntime_RUN_ONNX_TESTS)
 endif()
 
 
-if(onnxruntime_ENABLE_ATEN)
-  message(STATUS "Aten fallback is enabled.")
+if(onnxruntime_ENABLE_DLPACK)
+  message(STATUS "dlpack is enabled.")
   FetchContent_Declare(
     dlpack
     URL ${DEP_URL_dlpack}
@@ -626,7 +631,7 @@ if (onnxruntime_USE_WEBGPU)
       URL_HASH SHA1=${DEP_SHA1_dawn}
       # All previous patches are merged into the upstream dawn project. We don't need to apply any patches right now.
       # if we need to apply patches in the future, we can uncomment the following line.
-      PATCH_COMMAND ${Patch_EXECUTABLE} --binary --ignore-whitespace -p1 < ${PROJECT_SOURCE_DIR}/patches/dawn/dawn.patch
+      # PATCH_COMMAND ${Patch_EXECUTABLE} --binary --ignore-whitespace -p1 < ${PROJECT_SOURCE_DIR}/patches/dawn/dawn.patch
     )
   endif()