Update patch with latest commits (#19706)

### Description  ### Motivation and Context  --------- Co-authored-by: Sheil Kumar <[email protected]> Co-authored-by: Sheil Kumar <[email protected]> Co-authored-by: Scott McKay <[email protected]> Co-authored-by: Changming Sun <[email protected]> Co-authored-by: Yulong Wang <[email protected]> Co-authored-by: Yi Zhang <[email protected]> Co-authored-by: Your Name <[email protected]> Co-authored-by: kunal-vaishnavi <[email protected]> Co-authored-by: Yufeng Li <[email protected]> Co-authored-by: Sumit Agarwal <[email protected]> Co-authored-by: kailums <[email protected]> Co-authored-by: Rachel Guo <[email protected]> Co-authored-by: rachguo <[email protected]> Co-authored-by: duanshengliu <[email protected]> Co-authored-by: cloudhan <[email protected]> Co-authored-by: Adrian Lizarraga <[email protected]> Co-authored-by: zesongw <[email protected]> Co-authored-by: Maximilian Müller <[email protected]> Co-authored-by: pengwa <[email protected]>
microsoft · Feb 28, 2024 · 07957b0 · 07957b0
1 parent 96bd1fe
commit 07957b0
Show file tree

Hide file tree

Showing 99 changed files with 7,367 additions and 845 deletions.
diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt
@@ -324,15 +324,27 @@ if (onnxruntime_USE_ROCM)
   endif()
 
   # replicate strategy used by pytorch to get ROCM_VERSION
-  # https://github.com/pytorch/pytorch/blob/8eb21488fdcdb8b0e6fa2e46179b5fa6c42e75af/cmake/public/LoadHIP.cmake#L153-L173
-  file(READ "${onnxruntime_ROCM_HOME}/.info/version-dev" ROCM_VERSION_DEV_RAW)
-  string(REGEX MATCH "^([0-9]+)\.([0-9]+)\.([0-9]+)-.*$" ROCM_VERSION_DEV_MATCH ${ROCM_VERSION_DEV_RAW})
-  if (ROCM_VERSION_DEV_MATCH)
+  # https://github.com/pytorch/pytorch/blob/5c5b71b6eebae76d744261715231093e62f0d090/cmake/public/LoadHIP.cmake
+  # with modification
+  if (EXISTS "${onnxruntime_ROCM_HOME}/.info/version-dev")
+    file(READ "${onnxruntime_ROCM_HOME}/.info/version-dev" ROCM_VERSION_DEV_RAW)
+    string(REGEX MATCH "^([0-9]+)\.([0-9]+)\.([0-9]+)-.*$" ROCM_VERSION_MATCH ${ROCM_VERSION_DEV_RAW})
+  elseif (EXISTS "${onnxruntime_ROCM_HOME}/include/rocm_version.h")
+    file(READ "${onnxruntime_ROCM_HOME}/include/rocm_version.h" ROCM_VERSION_H_RAW)
+    string(REGEX MATCH "\"([0-9]+)\.([0-9]+)\.([0-9]+).*\"" ROCM_VERSION_MATCH ${ROCM_VERSION_H_RAW})
+  elseif (EXISTS "${onnxruntime_ROCM_HOME}/include/rocm-core/rocm_version.h")
+    file(READ "${onnxruntime_ROCM_HOME}/include/rocm-core/rocm_version.h" ROCM_VERSION_H_RAW)
+    string(REGEX MATCH "\"([0-9]+)\.([0-9]+)\.([0-9]+).*\"" ROCM_VERSION_MATCH ${ROCM_VERSION_H_RAW})
+  endif()
+
+  if (ROCM_VERSION_MATCH)
     set(ROCM_VERSION_DEV_MAJOR ${CMAKE_MATCH_1})
     set(ROCM_VERSION_DEV_MINOR ${CMAKE_MATCH_2})
     set(ROCM_VERSION_DEV_PATCH ${CMAKE_MATCH_3})
     set(ROCM_VERSION_DEV "${ROCM_VERSION_DEV_MAJOR}.${ROCM_VERSION_DEV_MINOR}.${ROCM_VERSION_DEV_PATCH}")
     math(EXPR ROCM_VERSION_DEV_INT "(${ROCM_VERSION_DEV_MAJOR}*10000) + (${ROCM_VERSION_DEV_MINOR}*100) + ${ROCM_VERSION_DEV_PATCH}")
+  else()
+    message(FATAL_ERROR "Cannot determine ROCm version string")
   endif()
   message("\n***** ROCm version from ${onnxruntime_ROCM_HOME}/.info/version-dev ****\n")
   message("ROCM_VERSION_DEV: ${ROCM_VERSION_DEV}")
@@ -1400,6 +1412,10 @@ endif()
 if (onnxruntime_USE_CUDA)
   set(CMAKE_CUDA_RUNTIME_LIBRARY Shared)
   set(CMAKE_CUDA_STANDARD 17)
+  if(onnxruntime_CUDA_HOME)
+    file(TO_CMAKE_PATH CUDAToolkit_ROOT ${onnxruntime_CUDA_HOME})
+  endif()
+  find_package(CUDAToolkit REQUIRED)
   if(onnxruntime_CUDNN_HOME)
     file(TO_CMAKE_PATH ${onnxruntime_CUDNN_HOME} onnxruntime_CUDNN_HOME)
   endif()
@@ -1729,14 +1745,12 @@ if(onnxruntime_BUILD_KERNEL_EXPLORER)
 endif()
 
 # When GDK_PLATFORM is set then WINAPI_FAMILY is defined in gdk_toolchain.cmake (along with other relevant flags/definitions).
-if (WIN32 AND NOT GDK_PLATFORM)
+if (WIN32 AND NOT GDK_PLATFORM AND NOT CMAKE_CROSSCOMPILING)
   if (NOT CMAKE_CXX_STANDARD_LIBRARIES MATCHES kernel32.lib)
     # On onecore, link to the onecore build of the MSVC runtime
     get_filename_component(msvc_path "${CMAKE_C_COMPILER}/../../../.." ABSOLUTE)
     link_directories(BEFORE "${msvc_path}/lib/onecore/${onnxruntime_target_platform}")
-    # The .lib files in the MSVC runtime have a DEFAULITLIB entry for onecore.lib, which in turn links to reverse forwarders.
-    # We ignore that entry and use onecore_apiset.lib instead, since system components must not rely on reverse forwarders.
-    add_link_options("/NODEFAULTLIB:onecore.lib")
+    # The .lib files in the MSVC runtime have a DEFAULITLIB entry for onecore.lib, but it shold not cause any conflict with onecoreuap.lib
   endif()
 endif()
 

diff --git a/cmake/adjust_global_compile_flags.cmake b/cmake/adjust_global_compile_flags.cmake
@@ -205,7 +205,7 @@ endif()
 
 
 macro(check_nvcc_compiler_flag _FLAG _RESULT)
-    execute_process(COMMAND ${onnxruntime_CUDA_HOME}/bin/nvcc "${_FLAG}" RESULT_VARIABLE NVCC_OUT ERROR_VARIABLE NVCC_ERROR)
+    execute_process(COMMAND ${CUDAToolkit_BIN_DIR}/nvcc "${_FLAG}" RESULT_VARIABLE NVCC_OUT ERROR_VARIABLE NVCC_ERROR)
     message("NVCC_ERROR = ${NVCC_ERROR}")
     message("NVCC_OUT = ${NVCC_OUT}")
     if ("${NVCC_OUT}" MATCHES "0")

diff --git a/cmake/external/onnxruntime_external_deps.cmake b/cmake/external/onnxruntime_external_deps.cmake
@@ -556,16 +556,15 @@ message("Finished fetching external dependencies")
 set(onnxruntime_LINK_DIRS )
 if (onnxruntime_USE_CUDA)
       #TODO: combine onnxruntime_CUDNN_HOME and onnxruntime_CUDA_HOME, assume they are the same
+      find_package(CUDAToolkit REQUIRED)
       if (WIN32)
         if(onnxruntime_CUDNN_HOME)
           list(APPEND onnxruntime_LINK_DIRS ${onnxruntime_CUDNN_HOME}/lib ${onnxruntime_CUDNN_HOME}/lib/x64)
         endif()
-        list(APPEND onnxruntime_LINK_DIRS ${onnxruntime_CUDA_HOME}/x64/lib64)
       else()
         if(onnxruntime_CUDNN_HOME)
           list(APPEND onnxruntime_LINK_DIRS  ${onnxruntime_CUDNN_HOME}/lib ${onnxruntime_CUDNN_HOME}/lib64)
         endif()
-        list(APPEND onnxruntime_LINK_DIRS ${onnxruntime_CUDA_HOME}/lib64)
       endif()
 endif()
 

diff --git a/cmake/onnxruntime_providers_cuda.cmake b/cmake/onnxruntime_providers_cuda.cmake
@@ -178,15 +178,16 @@
     add_dependencies(${target} onnxruntime_providers_shared ${onnxruntime_EXTERNAL_DEPENDENCIES})
     if(onnxruntime_CUDA_MINIMAL)
       target_compile_definitions(${target} PRIVATE USE_CUDA_MINIMAL)
-      target_link_libraries(${target} PRIVATE ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface)
+      target_link_libraries(${target} PRIVATE ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface CUDA::cudart)
     else()
-      target_link_libraries(${target} PRIVATE cublasLt cublas cudnn curand cufft ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface)
+      target_link_libraries(${target} PRIVATE CUDA::cublasLt CUDA::cublas cudnn CUDA::curand CUDA::cufft CUDA::cudart
+              ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface)
       if(onnxruntime_CUDNN_HOME)
           target_include_directories(${target} PRIVATE ${onnxruntime_CUDNN_HOME}/include)
           target_link_directories(${target} PRIVATE ${onnxruntime_CUDNN_HOME}/lib)
       endif()
     endif()
-
+    
     if (onnxruntime_USE_TRITON_KERNEL)
       # compile triton kernel, generate .a and .h files
       include(onnxruntime_compile_triton_kernel.cmake)
@@ -196,25 +197,24 @@
       target_include_directories(${target} PRIVATE ${triton_kernel_header_dir})
       target_link_libraries(${target} PUBLIC -Wl,--whole-archive ${triton_kernel_obj_file} -Wl,--no-whole-archive)
       # lib cuda needed by cuLaunchKernel
-      target_link_libraries(${target} PRIVATE cuda)
+      target_link_libraries(${target} PRIVATE CUDA::cuda_driver)
     endif()
 
     include(cutlass)
     target_include_directories(${target} PRIVATE ${cutlass_SOURCE_DIR}/include ${cutlass_SOURCE_DIR}/examples)
 
-    target_include_directories(${target} PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR}  ${eigen_INCLUDE_DIRS} ${TVM_INCLUDES} PUBLIC ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+    target_include_directories(${target} PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR}  ${eigen_INCLUDE_DIRS} ${TVM_INCLUDES}
+     PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
     # ${CMAKE_CURRENT_BINARY_DIR} is so that #include "onnxruntime_config.h" inside tensor_shape.h is found
     set_target_properties(${target} PROPERTIES LINKER_LANGUAGE CUDA)
     set_target_properties(${target} PROPERTIES FOLDER "ONNXRuntime")
 
     if (onnxruntime_ENABLE_CUDA_PROFILING) # configure cupti for cuda profiling
-      target_include_directories(${target} PRIVATE ${onnxruntime_CUDA_HOME}/extras/CUPTI/include)
-      target_link_directories(${target} PRIVATE ${onnxruntime_CUDA_HOME}/extras/CUPTI/lib64)
-      target_link_libraries(${target} PRIVATE cupti)
+      target_link_libraries(${target} PRIVATE CUDA::cupti)
     endif()
 
-    if (onnxruntime_ENABLE_NVTX_PROFILE AND NOT WIN32)
-      target_link_libraries(${target} PRIVATE nvToolsExt)
+    if (onnxruntime_ENABLE_NVTX_PROFILE)
+      target_link_libraries(${target} PRIVATE CUDA::nvtx3)
     endif()
 
     if (onnxruntime_ENABLE_TRAINING_OPS)

diff --git a/cmake/onnxruntime_providers_tensorrt.cmake b/cmake/onnxruntime_providers_tensorrt.cmake
@@ -8,7 +8,7 @@
   set(BUILD_LIBRARY_ONLY 1)
   add_definitions("-DONNX_ML=1")
   add_definitions("-DONNX_NAMESPACE=onnx")
-  set(CUDA_INCLUDE_DIRS ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+  set(CUDA_INCLUDE_DIRS ${CUDAToolkit_INCLUDE_DIRS})
   set(TENSORRT_ROOT ${onnxruntime_TENSORRT_HOME})
   set(OLD_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
   set(PROTOBUF_LIBRARY ${PROTOBUF_LIB})
@@ -58,7 +58,7 @@
       URL_HASH SHA1=${DEP_SHA1_onnx_tensorrt}
     )
     if (NOT CUDA_INCLUDE_DIR)
-      set(CUDA_INCLUDE_DIR ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}) # onnx-tensorrt repo needs this variable to build
+      set(CUDA_INCLUDE_DIR ${CUDAToolkit_INCLUDE_DIRS}) # onnx-tensorrt repo needs this variable to build
     endif()
     # The onnx_tensorrt repo contains a test program, getSupportedAPITest, which doesn't support Windows. It uses
     # unistd.h. So we must exclude it from our build. onnxruntime_fetchcontent_makeavailable is for the purpose.
@@ -102,11 +102,12 @@
   onnxruntime_add_include_to_target(onnxruntime_providers_tensorrt onnxruntime_common onnx flatbuffers::flatbuffers Boost::mp11 safeint_interface)
   add_dependencies(onnxruntime_providers_tensorrt onnxruntime_providers_shared ${onnxruntime_EXTERNAL_DEPENDENCIES})
   if (onnxruntime_USE_TENSORRT_BUILTIN_PARSER)
-    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${trt_link_libs} cudart ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers Boost::mp11 safeint_interface ${ABSEIL_LIBS})
+    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${trt_link_libs} ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers Boost::mp11 safeint_interface ${ABSEIL_LIBS} PUBLIC CUDA::cudart)
   else()
-    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${onnxparser_link_libs} ${trt_link_libs} cudart ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers ${ABSEIL_LIBS})
+    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${onnxparser_link_libs} ${trt_link_libs} ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers ${ABSEIL_LIBS} PUBLIC CUDA::cudart)
   endif()
-  target_include_directories(onnxruntime_providers_tensorrt PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR} ${eigen_INCLUDE_DIRS} PUBLIC ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+  target_include_directories(onnxruntime_providers_tensorrt PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR} ${eigen_INCLUDE_DIRS}
+    PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
   if(onnxruntime_CUDNN_HOME)
     target_include_directories(onnxruntime_providers_tensorrt PRIVATE ${onnxruntime_CUDNN_HOME}/include)
   endif()

diff --git a/cmake/onnxruntime_python.cmake b/cmake/onnxruntime_python.cmake
@@ -282,10 +282,7 @@ if (WIN32)
     get_filename_component(CUDNN_DLL_NAME ${CUDNN_DLL_PATH} NAME_WE)
     string(REPLACE "cudnn64_" "" CUDNN_VERSION "${CUDNN_DLL_NAME}")
     if(NOT onnxruntime_CUDA_VERSION)
-      message("Reading json file ${onnxruntime_CUDA_HOME}/version.json")
-      set(CUDA_SDK_JSON_FILE_PATH "${onnxruntime_CUDA_HOME}/version.json")
-      file(READ ${CUDA_SDK_JSON_FILE_PATH} CUDA_SDK_JSON_CONTENT)
-      string(JSON onnxruntime_CUDA_VERSION GET ${CUDA_SDK_JSON_CONTENT} "cuda" "version")
+      set(onnxruntime_CUDA_VERSION ${CUDAToolkit_VERSION})
       message("onnxruntime_CUDA_VERSION=${onnxruntime_CUDA_VERSION}")
     endif()
     file(APPEND "${VERSION_INFO_FILE}"

diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake
@@ -67,7 +67,7 @@ function(AddTest)
     if(onnxruntime_USE_CUDA)
       #XXX: we should not need to do this. onnxruntime_test_all.exe should not have direct dependency on CUDA DLLs,
       # otherwise it will impact when CUDA DLLs can be unloaded.
-      target_link_libraries(${_UT_TARGET} PRIVATE cudart)
+      target_link_libraries(${_UT_TARGET} PRIVATE CUDA::cudart)
     endif()
     target_link_libraries(${_UT_TARGET} PRIVATE ${_UT_LIBS} GTest::gtest GTest::gmock ${onnxruntime_EXTERNAL_LIBRARIES})
   endif()
@@ -1268,7 +1268,7 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
       list(APPEND onnxruntime_shared_lib_test_LIBS cpuinfo)
     endif()
     if (onnxruntime_USE_CUDA)
-      list(APPEND onnxruntime_shared_lib_test_LIBS cudart)
+      list(APPEND onnxruntime_shared_lib_test_LIBS CUDA::cudart)
     endif()
     if (onnxruntime_USE_ROCM)
       list(APPEND onnxruntime_shared_lib_test_LIBS hip::host)

diff --git a/cmake/wcos_rules_override.cmake b/cmake/wcos_rules_override.cmake
@@ -1,2 +1,2 @@
-set(CMAKE_C_STANDARD_LIBRARIES_INIT onecoreuap_apiset.lib)
-set(CMAKE_CXX_STANDARD_LIBRARIES_INIT onecoreuap_apiset.lib)
+set(CMAKE_C_STANDARD_LIBRARIES_INIT onecoreuap.lib)
+set(CMAKE_CXX_STANDARD_LIBRARIES_INIT onecoreuap.lib)
diff --git a/js/common/lib/env.ts b/js/common/lib/env.ts
@@ -36,6 +36,7 @@ export declare namespace Env {
     /**
      * set or get a boolean value indicating whether to enable trace.
      *
+     * @deprecated Use `env.trace` instead. If `env.trace` is set, this property will be ignored.
      * @defaultValue `false`
      */
     trace?: boolean;
@@ -167,13 +168,21 @@ export interface Env {
    * @defaultValue `'warning'`
    */
   logLevel?: 'verbose'|'info'|'warning'|'error'|'fatal';
+
   /**
    * Indicate whether run in debug mode.
    *
    * @defaultValue `false`
    */
   debug?: boolean;
 
+  /**
+   * set or get a boolean value indicating whether to enable trace.
+   *
+   * @defaultValue `false`
+   */
+  trace?: boolean;
+
   /**
    * Get version of the current package.
    */

diff --git a/js/common/lib/trace.ts b/js/common/lib/trace.ts
@@ -4,7 +4,7 @@
 import {env} from './env-impl.js';
 
 export const TRACE = (deviceType: string, label: string) => {
-  if (!env.wasm.trace) {
+  if (typeof env.trace === 'undefined' ? !env.wasm.trace : !env.trace) {
     return;
   }
   // eslint-disable-next-line no-console
@@ -30,14 +30,14 @@ const TRACE_FUNC = (msg: string, extraMsg?: string) => {
 };
 
 export const TRACE_FUNC_BEGIN = (extraMsg?: string) => {
-  if (!env.wasm.trace) {
+  if (typeof env.trace === 'undefined' ? !env.wasm.trace : !env.trace) {
     return;
   }
   TRACE_FUNC('BEGIN', extraMsg);
 };
 
 export const TRACE_FUNC_END = (extraMsg?: string) => {
-  if (!env.wasm.trace) {
+  if (typeof env.trace === 'undefined' ? !env.wasm.trace : !env.trace) {
     return;
   }
   TRACE_FUNC('END', extraMsg);

diff --git a/js/web/lib/wasm/jsep/backend-webgpu.ts b/js/web/lib/wasm/jsep/backend-webgpu.ts
@@ -710,7 +710,8 @@ export class WebGpuBackend {
   }
   setQueryType(): void {
     this.queryType = 'none';
-    if (this.env.webgpu.profiling?.mode === 'default' || this.env.wasm.trace) {
+    if (this.env.webgpu.profiling?.mode === 'default' ||
+        (typeof this.env.trace === 'undefined' ? this.env.wasm.trace : this.env.trace)) {
       if (this.device.features.has('chromium-experimental-timestamp-query-inside-passes')) {
         this.queryType = 'inside-passes';
       } else if (this.device.features.has('timestamp-query')) {

diff --git a/js/web/test/data/ops/add.jsonc b/js/web/test/data/ops/add.jsonc
@@ -157,28 +157,6 @@
             "type": "float32"
           }
         ]
-      },
-      {
-        "name": "T[2,0] T[2,1]",
-        "inputs": [
-          {
-            "data": [],
-            "dims": [2, 0],
-            "type": "float32"
-          },
-          {
-            "data": [1, 2],
-            "dims": [2, 1],
-            "type": "float32"
-          }
-        ],
-        "outputs": [
-          {
-            "data": [],
-            "dims": [2, 0],
-            "type": "float32"
-          }
-        ]
       }
     ]
   }

diff --git a/js/web/test/data/ops/add_zero-sized.jsonc b/js/web/test/data/ops/add_zero-sized.jsonc
@@ -0,0 +1,31 @@
+[
+  {
+    "name": "Add with no attributes",
+    "operator": "Add",
+    "attributes": [],
+    "cases": [
+      {
+        "name": "T[2,0] T[2,1]",
+        "inputs": [
+          {
+            "data": [],
+            "dims": [2, 0],
+            "type": "float32"
+          },
+          {
+            "data": [1, 2],
+            "dims": [2, 1],
+            "type": "float32"
+          }
+        ],
+        "outputs": [
+          {
+            "data": [],
+            "dims": [2, 0],
+            "type": "float32"
+          }
+        ]
+      }
+    ]
+  }
+]