Support DML provider on Windows

CMakeLists.txt

@@ -82,6 +82,14 @@ else()
   list(REMOVE_ITEM generator_srcs ${generator_cuda_srcs})
 endif()
 
+if(USE_DML)
+  if(WIN32)
+    add_compile_definitions(USE_DML=1)
+  else()
+    message(FATAL_ERROR "USE_DML is ON but this isn't windows.")
+  endif()
+endif()
+
 if(ENABLE_TESTS AND TEST_PHI2)
   add_compile_definitions(TEST_PHI2=1)
 else()

build.py

@@ -96,6 +96,7 @@ def validate_cuda_home(cuda_home: str | bytes | os.PathLike | None):
 def build(
     skip_wheel: bool = False,
     use_cuda: bool | None = None,
+    use_dml: bool | None = None,
     cuda_home: str | bytes | os.PathLike | None = None,
     cmake_generator: str | None = None,
     ort_home: str | bytes | os.PathLike | None = None,
@@ -141,6 +142,7 @@ def build(
         "-DCMAKE_POSITION_INDEPENDENT_CODE=ON",
         "-DUSE_CXX17=ON",
         "-DUSE_CUDA=ON" if cuda_home else "-DUSE_CUDA=OFF",
+        "-DUSE_DML=ON" if use_dml else "-DUSE_DML=OFF",
         f"-DBUILD_WHEEL={build_wheel}",
     ]
 
@@ -218,6 +220,7 @@ def build(
     parser.add_argument("--skip_csharp", action="store_true", help="Skip building the C# API.")
     parser.add_argument("--build_dir", default=None, help="Path to output directory.")
     parser.add_argument("--use_cuda", action="store_true", help="Whether to use CUDA. Default is to not use cuda.")
+    parser.add_argument("--use_dml", action="store_true", help="Whether to use DML. Default is to not use DML.")
     parser.add_argument("--parallel", action="store_true", help="Enable parallel build.")
     parser.add_argument(
         "--config",
@@ -231,6 +234,7 @@ def build(
     build(
         skip_wheel=args.skip_wheel,
         use_cuda=args.use_cuda,
+        use_dml=args.use_dml,
         cuda_home=args.cuda_home,
         cmake_generator=args.cmake_generator,
         ort_home=args.ort_home,

cmake/options.cmake

@@ -1,6 +1,7 @@
 include(CMakeDependentOption)
 
 option(USE_CUDA "Build with CUDA support" ON)
+option(USE_DML "Build with DML support" OFF)
 option(NO_TOKENIZER "Don't include the Tokenizer" OFF)
 option(ENABLE_PYTHON "Build the Python API." ON)
 option(ENABLE_TESTS "Enable tests" ON)

src/models/model.cpp

@@ -9,6 +9,11 @@
 #include "decoder_only.h"
 #include "whisper.h"
 #include "kernels.h"
+#ifdef USE_DML
+//  Because dml_provider_factory includes windows headers that #define min and max, this next line will prevent this from happening
+#define NOMINMAX
+#include "dml_provider_factory.h"
+#endif
 
 namespace Generators {
 
@@ -291,7 +296,14 @@ void Model::CreateSessionOptions() {
 
       Ort::ThrowOnError(Ort::api->UpdateROCMProviderOptions(&ort_provider_options, keys.data(), values.data(), keys.size()));
       ort_options.AppendExecutionProvider_ROCM(ort_provider_options);
-      device_type_ = DeviceType::CPU;  // Scoring uses CPU, even though the model uses ROCM
+#ifdef USE_DML
+    } else if (provider_options.name == "dml") {
+      const OrtDmlApi* p_dml_api{};
+      Ort::ThrowOnError(Ort::api->GetExecutionProviderApi("DML", ORT_API_VERSION, reinterpret_cast<const void**>(&p_dml_api)));
+      if (!p_dml_api)
+        throw std::runtime_error("Unexpected nullptr getting OrtDmlApi");
+      p_dml_api->SessionOptionsAppendExecutionProvider_DML(&ort_options, 0);
+#endif
     } else
       throw std::runtime_error("Unknown provider type: " + provider_options.name);
   }