Cherry pick PR 3120 and 3131 to rel-1.2.0 (#3137)

* Enable DML Nuget Package for x64 or x86 architectures (#3120)

* add dml gpu pipelines

* add x86 to the gpu dml dev build pipeline

* Enable DML x86 builds

* Fix uint64_t -> size_t warning

* fix warnings

* enable dml on x86 ci builds

* operatorHelper 773 error uint32_t vs uint64_t

* operatorHelper 773 error uint32_t vs uint64_t

* make x86 pipeline use the gpu pool

* more warnings

* fix x86 directml path

* make dml nuget package

* disable tf_pnasnet_large

* disable zfnet512

* make validation use wildcards

* disable x86 dml gpu tests

* add args.

* update gpu.yml

* change nupkg wildcard

* add debug statements

* package x86 dml nupkg

* dont drop managed nuget again from dml pipeline build

* Add DML EULA

* directml license should be renamed to not clobber the existing license

* casing on dml package....

* {} to ()

* fix license name

* disable dml from x86 ci

* typo and cr feedback

* remove featurizers

* ship the dml pdb as well

* Fix CUDA PATH (#3131)

Previously, we put the "bin" folder of all the CUDA verions in the system PATH. And 10.2 is in the front. It's a mess.
So I've removed all of them from the system PATH env. But I need to add one of them back through build scripts.

(The problem only affect the C# test, not the C/C++ tests that forked from build.py).

Co-authored-by: smk2007 <smk2007@gmail.com>
Co-authored-by: Changming Sun <chasun@microsoft.com>

cmake/onnxruntime_common.cmake

@@ -56,7 +56,7 @@ elseif(onnxruntime_target_platform STREQUAL "ARM" OR CMAKE_GENERATOR MATCHES "AR
     set(onnxruntime_target_platform "ARM")
 elseif(onnxruntime_target_platform STREQUAL "x64" OR onnxruntime_target_platform STREQUAL "x86_64" OR onnxruntime_target_platform STREQUAL "AMD64" OR CMAKE_GENERATOR MATCHES "Win64")
     set(onnxruntime_target_platform "x64")
-elseif(onnxruntime_target_platform STREQUAL "x86" OR onnxruntime_target_platform STREQUAL "i386" OR onnxruntime_target_platform STREQUAL "i686")
+elseif(onnxruntime_target_platform STREQUAL "Win32" OR onnxruntime_target_platform STREQUAL "x86" OR onnxruntime_target_platform STREQUAL "i386" OR onnxruntime_target_platform STREQUAL "i686")
     set(onnxruntime_target_platform "x86")
 endif()
 

cmake/onnxruntime_providers.cmake

@@ -435,6 +435,7 @@ if (onnxruntime_USE_DML)
     if(NOT onnxruntime_target_platform STREQUAL "x86" AND NOT onnxruntime_target_platform STREQUAL "x64")
       message(FATAL_ERROR "Target platform ${onnxruntime_target_platform} is not supported by DML")
     endif()
+
     foreach(file "DirectML.dll" "DirectML.pdb" "DirectML.Debug.dll" "DirectML.Debug.pdb")
       add_custom_command(TARGET onnxruntime_providers_dml
         POST_BUILD

csharp/OnnxRuntime.CSharp.proj

@@ -18,8 +18,9 @@ CMake creates a target to this project
     <IsReleaseBuild Condition=" '$(IsReleaseBuild)' == '' ">false</IsReleaseBuild>
 
     <!--internal build related properties-->
-    <OnnxRuntimeSourceDirectory Condition="'$(OnnxRuntimeSourceDirectory)'==''">..</OnnxRuntimeSourceDirectory>    
+    <OnnxRuntimeSourceDirectory Condition="'$(OnnxRuntimeSourceDirectory)'==''">..</OnnxRuntimeSourceDirectory>
     <OnnxRuntimeBuildDirectory Condition="'$(OnnxRuntimeBuildDirectory)'==''">..\build\Windows</OnnxRuntimeBuildDirectory>
+    <OnnxRuntimePackagesDirectory Condition="'$(OnnxRuntimePackagesDirectory)'==''">$(OnnxRuntimeBuildDirectory)\packages</OnnxRuntimePackagesDirectory>
     <NativeBuildOutputDir>$(OnnxRuntimeBuildDirectory)\$(Configuration)\$(Configuration)</NativeBuildOutputDir>
     
   </PropertyGroup>
@@ -87,6 +88,9 @@ CMake creates a target to this project
     <ConvertToAbsolutePath Paths="$(NativeBuildOutputDir)">
       <Output TaskParameter="AbsolutePaths" PropertyName="NativeBuildOutputDirAbs"/>
     </ConvertToAbsolutePath>
+    <ConvertToAbsolutePath Paths="$(OnnxRuntimePackagesDirectory)">
+      <Output TaskParameter="AbsolutePaths" PropertyName="OnnxRuntimePackagesDirectoryAbs"/>
+    </ConvertToAbsolutePath>
     <ConvertToAbsolutePath Paths="$(OnnxRuntimeSourceDirectory)">
       <Output TaskParameter="AbsolutePaths" PropertyName="OnnxRuntimeSourceDirectoryAbs"/>
     </ConvertToAbsolutePath>    
@@ -107,7 +111,7 @@ CMake creates a target to this project
              Properties="NoBuild=true;Platform=AnyCPU;PackageVersion=$(PackageVersion);OrtPackageId=$(OrtPackageId)"/>
     
     <Message Importance="High" Text="Generating nuspec for the native Nuget package ..." />
-    <Exec ContinueOnError="False" Command="python ..\tools\nuget\generate_nuspec_for_native_nuget.py --package_version $(PackageVersion) --package_name $(OrtPackageId) --target_architecture $(TargetArchitecture) --build_config $(Configuration) --native_build_path $(NativeBuildOutputDirAbs) --ort_build_path $(OnnxRuntimeBuildDirectoryAbs) --sources_path $(OnnxRuntimeSourceDirectoryAbs) --commit_id $(GitCommitHash) --is_release_build $(IsReleaseBuild)" ConsoleToMSBuild="true">
+    <Exec ContinueOnError="False" Command="python ..\tools\nuget\generate_nuspec_for_native_nuget.py --package_version $(PackageVersion) --package_name $(OrtPackageId) --target_architecture $(TargetArchitecture) --build_config $(Configuration) --native_build_path $(NativeBuildOutputDirAbs) --packages_path $(OnnxRuntimePackagesDirectoryAbs) --ort_build_path $(OnnxRuntimeBuildDirectoryAbs) --sources_path $(OnnxRuntimeSourceDirectoryAbs) --commit_id $(GitCommitHash) --is_release_build $(IsReleaseBuild)" ConsoleToMSBuild="true">
         <Output TaskParameter="ConsoleOutput" PropertyName="GenerateNuspecOutput" />
     </Exec>
 

csharp/src/Microsoft.ML.OnnxRuntime/Microsoft.ML.OnnxRuntime.csproj

@@ -92,8 +92,8 @@
           CopyToOutputDirectory="Always"
           Visible="false"
     />
-    <None Include="$(NativeBuildOutputDir)\windows.ai.machinelearning.dll"
-          Condition="Exists('$(NativeBuildOutputDir)\windows.ai.machinelearning.dll')
+    <None Include="$(NativeBuildOutputDir)\Windows.AI.MachineLearning.dll"
+          Condition="Exists('$(NativeBuildOutputDir)\Windows.AI.MachineLearning.dll')
                      And ('$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime.DirectML'
                      Or '$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime')"
           PackagePath="\runtimes\win-$(TargetArchitecture)\native"
@@ -101,8 +101,8 @@
           CopyToOutputDirectory="Always"
           Visible="false"
     />
-    <None Include="$(NativeBuildOutputDir)\windows.ai.machinelearning.lib"
-          Condition="Exists('$(NativeBuildOutputDir)\windows.ai.machinelearning.lib')
+    <None Include="$(NativeBuildOutputDir)\Windows.AI.MachineLearning.lib"
+          Condition="Exists('$(NativeBuildOutputDir)\Windows.AI.MachineLearning.lib')
                      And ('$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime.DirectML'
                      Or '$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime')"
           PackagePath="\runtimes\win-$(TargetArchitecture)\native"
@@ -110,8 +110,8 @@
           CopyToOutputDirectory="Never"
           Visible="false"
     />
-    <None Include="$(NativeBuildOutputDir)\windows.ai.machinelearning.pdb"
-          Condition="Exists('$(NativeBuildOutputDir)\windows.ai.machinelearning.pdb')
+    <None Include="$(NativeBuildOutputDir)\Windows.AI.MachineLearning.pdb"
+          Condition="Exists('$(NativeBuildOutputDir)\Windows.AI.MachineLearning.pdb')
                      And ('$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime.DirectML'
                      Or '$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime')"
           PackagePath="\runtimes\win-$(TargetArchitecture)\native"
@@ -119,8 +119,8 @@
           CopyToOutputDirectory="Always"
           Visible="false"
     />
-    <None Include="$(OnnxRuntimeBuildDirectory)\$(Configuration)\windows.ai.machinelearning.winmd"
-          Condition="Exists('$(OnnxRuntimeBuildDirectory)\$(Configuration)\windows.ai.machinelearning.winmd')
+    <None Include="$(OnnxRuntimeBuildDirectory)\$(Configuration)\Windows.AI.MachineLearning.winmd"
+          Condition="Exists('$(OnnxRuntimeBuildDirectory)\$(Configuration)\Windows.AI.MachineLearning.winmd')
                      And ('$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime.DirectML'
                      Or '$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime')"
           PackagePath="\build\native\metadata"
@@ -129,7 +129,7 @@
           Visible="false"
     />
     <None Include="$(OnnxRuntimeBuildDirectory)\$(Configuration)\Windows.AI.MachineLearning.h"
-          Condition="Exists('$(OnnxRuntimeBuildDirectory)\$(Configuration)\windows.ai.machinelearning.h')
+          Condition="Exists('$(OnnxRuntimeBuildDirectory)\$(Configuration)\Windows.AI.MachineLearning.h')
                      And ('$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime.DirectML'
                      Or '$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime')"
           PackagePath="\build\native\include"
@@ -138,7 +138,7 @@
           Visible="false"
     />
     <None Include="$(OnnxRuntimeBuildDirectory)\$(Configuration)\Windows.AI.MachineLearning.Native.h"
-          Condition="Exists('$(OnnxRuntimeBuildDirectory)\$(Configuration)\windows.ai.machinelearning.native.h')
+          Condition="Exists('$(OnnxRuntimeBuildDirectory)\$(Configuration)\Windows.AI.MachineLearning.Native.h')
                      And ('$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime.DirectML'
                      Or '$(OrtPackageId)' == 'Microsoft.ML.OnnxRuntime')"
           PackagePath="\build\native\include"

onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlCommandRecorder.cpp

@@ -62,7 +62,7 @@ void DmlCommandRecorder::InitializeOperator(
 
         // Allocate and immediately free a temporary buffer. The buffer resource will still be
         // alive (managed by the pool); freeing allows the resource to be shared with other operators.
-        void* tempResourceHandle = allocator->Alloc(temporaryResourceSize, AllocatorRoundingMode::Enabled);
+        void* tempResourceHandle = allocator->Alloc(static_cast<size_t>(temporaryResourceSize), AllocatorRoundingMode::Enabled);
         if (!tempResourceHandle)
         {
             THROW_HR(E_OUTOFMEMORY);
@@ -137,7 +137,7 @@ void DmlCommandRecorder::ExecuteOperator(
 
         // Allocate and immediately free a temporary buffer. The buffer resource will still be
         // alive (managed by the pool); freeing allows the resource to be shared with other operators.
-        void* tempResourceHandle = allocator->Alloc(temporaryResourceSize, AllocatorRoundingMode::Enabled);
+        void* tempResourceHandle = allocator->Alloc(static_cast<size_t>(temporaryResourceSize), AllocatorRoundingMode::Enabled);
         if (!tempResourceHandle)
         {
             THROW_HR(E_OUTOFMEMORY);

onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionProvider.cpp

@@ -455,7 +455,7 @@ namespace Dml
     {
         assert(!m_closed);
 
-        m_uploadHeap->BeginUploadToGpu(dstData, 0, D3D12_RESOURCE_STATE_UNORDERED_ACCESS, AsByteSpan(srcData, srcDataSize));
+        m_uploadHeap->BeginUploadToGpu(dstData, 0, D3D12_RESOURCE_STATE_UNORDERED_ACCESS, AsByteSpan(srcData, static_cast<size_t>(srcDataSize)));
 
         return S_OK;
     }

onnxruntime/core/providers/dml/DmlExecutionProvider/src/FusedGraphKernel.cpp

@@ -289,7 +289,7 @@ namespace Dml
             if (persistentResourceSize > 0)
             {
                 THROW_IF_FAILED(m_provider->AllocatePooledResource(
-                    persistentResourceSize,
+                    static_cast<size_t>(persistentResourceSize),
                     AllocatorRoundingMode::Disabled,
                     m_persistentResource.GetAddressOf(),
                     m_persistentResourceAllocatorUnk.GetAddressOf()));
@@ -589,7 +589,7 @@ namespace Dml
                 // which is scheduled up to the point that this method returns has completed.
                 ComPtr<IUnknown> tempAlloc;
                 uint64_t tempAllocId = 0;
-                THROW_IF_FAILED(contextWrapper.AllocateTemporaryData(execBindingProps.TemporaryResourceSize, tempAlloc.GetAddressOf(), &tempAllocId));
+                THROW_IF_FAILED(contextWrapper.AllocateTemporaryData(static_cast<size_t>(execBindingProps.TemporaryResourceSize), tempAlloc.GetAddressOf(), &tempAllocId));
 
                 ComPtr<IUnknown> tempResourceUnk;
                 m_winmlProvider->GetABIDataInterface(false, tempAlloc.Get(), &tempResourceUnk);

onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperator.cpp

@@ -80,7 +80,7 @@ namespace Dml
             if (persistentResourceSize > 0)
             {
                 THROW_IF_FAILED(m_executionProvider->AllocatePooledResource(
-                    persistentResourceSize,
+                    static_cast<size_t>(persistentResourceSize),
                     AllocatorRoundingMode::Enabled,
                     m_persistentResource.GetAddressOf(),
                     m_persistentResourcePoolingUnk.GetAddressOf()));
@@ -117,7 +117,7 @@ namespace Dml
             {
                 m_persistentResource = nullptr;
                 THROW_IF_FAILED(m_executionProvider->AllocatePooledResource(
-                    persistentResourceSize,
+                    static_cast<size_t>(persistentResourceSize),
                     AllocatorRoundingMode::Enabled,
                     m_persistentResource.GetAddressOf(),
                     m_persistentResourcePoolingUnk.GetAddressOf()));

onnxruntime/core/providers/dml/OperatorAuthorHelper/OperatorHelper.cpp

@@ -374,7 +374,7 @@ int64_t ReadAsInt64(MLOperatorTensorDataType tensorDataType, const void* p)
         )
     {
         const uint32_t inputDimCount = gsl::narrow_cast<int32_t>(inputDimensions.size());
-        m_axis = HandleNegativeAxis(operatorAttributes.GetOptionalAttribute<int32_t>(AttrName::Axis, 0), inputDimCount);
+        m_axis = static_cast<int>(HandleNegativeAxis(operatorAttributes.GetOptionalAttribute<int32_t>(AttrName::Axis, 0), inputDimCount));
         m_split = operatorAttributes.GetOptionalAttributeVectorInt32(AttrName::Split);
     }
 
@@ -768,9 +768,9 @@ int64_t ReadAsInt64(MLOperatorTensorDataType tensorDataType, const void* p)
         gsl::span<const DimensionType> inputDimensions
         )
     {
-        int32_t inputDimCount = gsl::narrow_cast<int>(inputDimensions.size());
-        m_axis = HandleNegativeAxis(operatorAttributes.GetOptionalAttribute<int>(AttrName::Axis, -1), inputDimCount);
-        ML_CHECK_VALID_ARGUMENT(m_axis < inputDimensions.size());
+        uint32_t inputDimCount = gsl::narrow_cast<uint32_t>(inputDimensions.size());
+        m_axis = static_cast<int>(HandleNegativeAxis(operatorAttributes.GetOptionalAttribute<int>(AttrName::Axis, -1), inputDimCount));
+        ML_CHECK_VALID_ARGUMENT(m_axis < static_cast<int>(inputDimensions.size()));
     }
 
     std::vector<EdgeShapes> ConcatHelper::GetOutputShapes(const MLShapeInferenceContext& shapeInfo) const
@@ -864,7 +864,7 @@ int64_t ReadAsInt64(MLOperatorTensorDataType tensorDataType, const void* p)
         int32_t inputDimCount = gsl::narrow_cast<int32_t>(inputDimensions.size());
         int32_t axis = operatorAttributes.GetOptionalAttribute<int32_t>(AttrName::Axis, 1);
         // Flatten can accept an axis [-r, r], including one past the last absolute index.
-        m_axis = (axis == inputDimCount) ? axis : HandleNegativeAxis(axis, inputDimCount);
+        m_axis = (axis == inputDimCount) ? axis : static_cast<int>(HandleNegativeAxis(axis, inputDimCount));
     }
 
     std::vector<EdgeShapes> FlattenHelper::GetOutputShapes(const MLShapeInferenceContext& shapeInfo) const

onnxruntime/core/providers/dml/OperatorAuthorHelper/OperatorHelper.h

@@ -607,7 +607,7 @@ public:
             {
                 dimIndex = axes[i];
             }
-            ML_CHECK_VALID_ARGUMENT(dimIndex < inputDimensions.size(), "'axes' must be valid with within actual input dimensions.");
+            ML_CHECK_VALID_ARGUMENT(dimIndex < static_cast<int>(inputDimensions.size()), "'axes' must be valid with within actual input dimensions.");
 
             // Positive values are offsets from 0.
             // Negative values are offsets from the dimension's size.

onnxruntime/test/onnx/main.cc

@@ -396,7 +396,7 @@ int real_main(int argc, char* argv[], Ort::Env& env) {
     static const ORTCHAR_T* cuda_flaky_tests[] = {
         ORT_TSTR("fp16_inception_v1"),
         ORT_TSTR("fp16_shufflenet"), ORT_TSTR("fp16_tiny_yolov2")};
-    static const ORTCHAR_T* dml_disabled_tests[] = {ORT_TSTR("mlperf_ssd_resnet34_1200"), ORT_TSTR("mlperf_ssd_mobilenet_300"), ORT_TSTR("mask_rcnn"), ORT_TSTR("faster_rcnn")};
+    static const ORTCHAR_T* dml_disabled_tests[] = {ORT_TSTR("mlperf_ssd_resnet34_1200"), ORT_TSTR("mlperf_ssd_mobilenet_300"), ORT_TSTR("mask_rcnn"), ORT_TSTR("faster_rcnn"), ORT_TSTR("tf_pnasnet_large"), ORT_TSTR("zfnet512")};
     static const ORTCHAR_T* dnnl_disabled_tests[] = {ORT_TSTR("test_densenet121"), ORT_TSTR("test_resnet18v2"), ORT_TSTR("test_resnet34v2"), ORT_TSTR("test_resnet50v2"), ORT_TSTR("test_resnet101v2"),
                                                      ORT_TSTR("test_resnet101v2"), ORT_TSTR("test_vgg19"), ORT_TSTR("tf_inception_resnet_v2"), ORT_TSTR("tf_inception_v1"), ORT_TSTR("tf_inception_v3"), ORT_TSTR("tf_inception_v4"), ORT_TSTR("tf_mobilenet_v1_1.0_224"),
                                                      ORT_TSTR("tf_mobilenet_v2_1.0_224"), ORT_TSTR("tf_mobilenet_v2_1.4_224"), ORT_TSTR("tf_nasnet_large"), ORT_TSTR("tf_pnasnet_large"), ORT_TSTR("tf_resnet_v1_50"), ORT_TSTR("tf_resnet_v1_101"), ORT_TSTR("tf_resnet_v1_101"),

tools/ci_build/github/azure-pipelines/azure-pipelines-py-packaging.yml

@@ -216,7 +216,7 @@ jobs:
   timeoutInMinutes:  60  
   variables:
     CUDA_VERSION: '10.1'
-    EnvSetupScript: setup_env.bat
+    EnvSetupScript: setup_env_cuda.bat
   strategy:
     matrix:
       Python35:

tools/ci_build/github/azure-pipelines/c-api-packaging-pipelines.yml

@@ -165,7 +165,7 @@ jobs:
   pool: 'Win-GPU-2019'
   timeoutInMinutes:  120
   variables:
-    EnvSetupScript: setup_env.bat
+    EnvSetupScript: setup_env_cuda.bat
     buildArch: x64
     msbuildArch: amd64
     msbuildPlatform: x64

tools/ci_build/github/azure-pipelines/nuget/templates/gpu.yml

@@ -11,7 +11,7 @@ jobs:
     BuildCommand: --build_dir $(Build.BinariesDirectory) --skip_submodule_sync --build_shared_lib --enable_onnx_tests --use_telemetry --cmake_generator "Visual Studio 16 2019" --use_cuda --cuda_version=10.1 --cuda_home="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1" --cudnn_home="C:\local\cudnn-10.1-windows10-x64-v7.6.5.32\cuda"
     BuildArch: 'x64'
     msbuildArchitecture: 'amd64'
-    EnvSetupScript: 'setup_env.bat'    
+    EnvSetupScript: 'setup_env_cuda.bat'    
     sln_platform: 'x64'
     DoDebugBuild: 'false'
     DoNugetPack : 'true'
@@ -26,6 +26,51 @@ jobs:
      mkdir $(Build.ArtifactStagingDirectory)\testdata
      copy $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\custom_op_library.* $(Build.ArtifactStagingDirectory)\testdata
 
+- template: ../../templates/win-ci-2019.yml
+  parameters:
+    AgentPool : 'Win-GPU-2019'
+    ArtifactName: 'drop-nuget-dml'    
+    JobName: 'Windows_CI_GPU_DML_Dev'
+    BuildCommand: --build_dir $(Build.BinariesDirectory) --skip_submodule_sync --build_shared_lib --enable_onnx_tests --use_telemetry --use_dml --use_winml --cmake_generator "Visual Studio 16 2019"
+    BuildArch: 'x64'
+    msbuildArchitecture: 'amd64'
+    EnvSetupScript: 'setup_env_cuda.bat'    
+    sln_platform: 'x64'
+    DoDebugBuild: 'false'
+    DoNugetPack : 'true'
+    DoCompliance: 'false'
+    DoEsrp: ${{ parameters.DoEsrp }}    
+    CudaVersion: '10.0'
+    OrtPackageId: 'Microsoft.ML.OnnxRuntime.DirectML'
+    NuPackScript: |
+     msbuild $(Build.SourcesDirectory)\csharp\OnnxRuntime.CSharp.proj /p:Configuration=RelWithDebInfo /t:CreatePackage /p:OrtPackageId=Microsoft.ML.OnnxRuntime.DirectML
+     copy $(Build.SourcesDirectory)\csharp\src\Microsoft.ML.OnnxRuntime\bin\RelWithDebInfo\*.nupkg $(Build.ArtifactStagingDirectory)
+     copy $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\*.nupkg $(Build.ArtifactStagingDirectory)
+     mkdir $(Build.ArtifactStagingDirectory)\testdata
+     copy $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\custom_op_library.* $(Build.ArtifactStagingDirectory)\testdata
+
+- template: ../../templates/win-ci-2019.yml
+  parameters:
+    AgentPool : 'Win-GPU-2019'
+    ArtifactName: 'drop-win-dml-x86-zip'
+    JobName: 'Windows_CI_GPU_DML_Dev_x86'
+    BuildCommand: --build_dir $(Build.BinariesDirectory) --x86 --skip_submodule_sync --build_shared_lib --enable_onnx_tests --use_telemetry --use_dml --use_winml --cmake_generator "Visual Studio 16 2019"
+    BuildArch: 'x86'
+    EnvSetupScript: 'setup_env_x86.bat'
+    sln_platform: 'Win32'
+    DoDebugBuild: 'false'
+    DoNugetPack : 'true'
+    DoCompliance: ${{ parameters.DoCompliance }}
+    DoEsrp: ${{ parameters.DoEsrp }}
+    RunTests: 'false'
+    NuPackScript: |
+     msbuild $(Build.SourcesDirectory)\csharp\OnnxRuntime.CSharp.proj /p:Configuration=RelWithDebInfo /p:TargetArchitecture=x86 /t:CreatePackage /p:OrtPackageId=Microsoft.ML.OnnxRuntime.DirectML
+     cd $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\
+     ren Microsoft.ML.OnnxRuntime.DirectML.* win-dml-x86.zip
+     copy $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\win-dml-x86.zip $(Build.ArtifactStagingDirectory)
+     mkdir $(Build.ArtifactStagingDirectory)\testdata
+     copy $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\custom_op_library.* $(Build.ArtifactStagingDirectory)\testdata
+
 - job: 'Linux_CI_GPU_Dev'
   workspace:
     clean: all
@@ -68,15 +113,29 @@ jobs:
   pool: 'Win-GPU-2019'
   dependsOn:
   - Windows_CI_GPU_CUDA_Dev
+  - Windows_CI_GPU_DML_Dev
+  - Windows_CI_GPU_DML_Dev_x86
   - Linux_CI_GPU_Dev
   condition: succeeded()
   steps:  
   - task: DownloadPipelineArtifact@0
-    displayName: 'Download Pipeline Artifact - NuGet'
+    displayName: 'Download Pipeline Artifact - NuGet CUDA'
     inputs:
       artifactName: 'drop-nuget'
       targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
 
+  - task: DownloadPipelineArtifact@0
+    displayName: 'Download Pipeline Artifact - NuGet DirectML'
+    inputs:
+      artifactName: 'drop-nuget-dml'
+      targetPath: '$(Build.BinariesDirectory)/nuget-artifact-dml'
+      
+  - task: DownloadPipelineArtifact@0
+    displayName: 'Download Pipeline Artifact - NuGet DirectML x86'
+    inputs:
+      artifactName: 'drop-win-dml-x86-zip'
+      targetPath: '$(Build.BinariesDirectory)/nuget-artifact-dml'
+
   - task: DownloadPipelineArtifact@0
     displayName: 'Download Pipeline Artifact - Linux'
     inputs:
@@ -110,6 +169,40 @@ jobs:
      copy $(Build.BinariesDirectory)\nuget-artifact\*.nupkg $(Build.ArtifactStagingDirectory)
     displayName: 'Bundle Native NuGet and other binaries'
 
+  - script: |
+     pushd $(Build.BinariesDirectory)\nuget-artifact-dml
+     dir
+     powershell -Command "Invoke-WebRequest http://stahlworks.com/dev/unzip.exe -OutFile unzip.exe"
+     powershell -Command "Invoke-WebRequest http://stahlworks.com/dev/zip.exe -OutFile zip.exe"
+     set PATH=%CD%;%PATH%
+     FOR /R %%i IN (*.nupkg) do ( 
+         rename %%~ni.nupkg %%~ni.zip
+         unzip %%~ni.zip -d %%~ni
+         del /Q %%~ni.zip
+
+         unzip win-dml-x86.zip -d win-x86
+         mkdir %%~ni\runtimes\win-x86
+         mkdir %%~ni\runtimes\win-x86\native
+
+         move win-x86\runtimes\win-x86\native\onnxruntime.dll %%~ni\runtimes\win-x86\native\onnxruntime.dll
+         move win-x86\runtimes\win-x86\native\onnxruntime.lib %%~ni\runtimes\win-x86\native\onnxruntime.lib
+         move win-x86\runtimes\win-x86\native\onnxruntime.pdb %%~ni\runtimes\win-x86\native\onnxruntime.pdb
+
+         move win-x86\runtimes\win-x86\native\windows.ai.machinelearning.dll %%~ni\runtimes\win-x86\native\windows.ai.machinelearning.dll
+         move win-x86\runtimes\win-x86\native\windows.ai.machinelearning.lib %%~ni\runtimes\win-x86\native\windows.ai.machinelearning.lib
+         move win-x86\runtimes\win-x86\native\windows.ai.machinelearning.pdb %%~ni\runtimes\win-x86\native\windows.ai.machinelearning.pdb
+
+         move win-x86\runtimes\win-x86\native\directml.dll %%~ni\runtimes\win-x86\native\directml.dll
+
+         pushd %%~ni
+         zip -r ..\%%~ni.zip .
+         popd
+         move %%~ni.zip %%~ni.nupkg
+     )
+     popd
+     copy $(Build.BinariesDirectory)\nuget-artifact-dml\Microsoft.ML.OnnxRuntime.DirectML*nupkg $(Build.ArtifactStagingDirectory)
+    displayName: 'Bundle DML NuGet and other binaries'
+
   - template: ../../templates/esrp_nuget.yml
     parameters:
       DisplayName: 'ESRP - sign NuGet package'
@@ -119,9 +212,17 @@ jobs:
   - template: ../../templates/validate-nuget.yml
     parameters:
       NugetPath: '$(Build.ArtifactStagingDirectory)'
+      NugetPackage: 'Microsoft.ML.OnnxRuntime.Gpu*nupkg'
       PlatformsSupported: 'win-x64,linux-x64'
       VerifyNugetSigning: ${{ parameters.DoEsrp }}
 
+  - template: ../../templates/validate-nuget.yml
+    parameters:
+      NugetPath: '$(Build.ArtifactStagingDirectory)'
+      NugetPackage: 'Microsoft.ML.OnnxRuntime.DirectML*nupkg'
+      PlatformsSupported: 'win-x64,win-x86'
+      VerifyNugetSigning: ${{ parameters.DoEsrp }}
+
   - task: PublishPipelineArtifact@0
     displayName: 'Publish Pipeline NuGet Artifact'
     inputs:

tools/ci_build/github/azure-pipelines/nuget/templates/test_win.yml

@@ -32,7 +32,7 @@ jobs:
   - task: BatchScript@1
     displayName: 'setup env'
     inputs:
-      filename: '$(Build.SourcesDirectory)\tools\ci_build\github\windows\setup_env.bat'
+      filename: '$(Build.SourcesDirectory)\tools\ci_build\github\windows\setup_env_cuda.bat'
       modifyEnvironment: true
       workingFolder: '$(Build.BinariesDirectory)'
 

tools/ci_build/github/azure-pipelines/templates/validate-nuget.yml

@@ -2,6 +2,7 @@ parameters:
   NugetPath: ''
   PlatformsSupported: ''
   VerifyNugetSigning: ''
+  NugetPackage: '*'
 
 steps:
     - task: UsePythonVersion@0
@@ -13,5 +14,5 @@ steps:
       displayName: 'Validate Nuget'
       inputs:
         scriptPath: '$(Build.SourcesDirectory)\tools\nuget\validate_nuget.py'
-        arguments: '--nuget_path ${{parameters.NugetPath}} --platforms_supported ${{parameters.PlatformsSupported}} --verify_nuget_signing ${{parameters.VerifyNugetSigning}}'
+        arguments: '--nuget_package ${{parameters.NugetPackage}} --nuget_path ${{parameters.NugetPath}} --platforms_supported ${{parameters.PlatformsSupported}} --verify_nuget_signing ${{parameters.VerifyNugetSigning}}'
         workingDirectory: "$(Build.BinariesDirectory)"

tools/ci_build/github/azure-pipelines/templates/win-ci-2019.yml

@@ -15,6 +15,7 @@ parameters:
   AgentPool: 'Win-CPU-2019'
   AgentDemands: []
   OrtPackageId: Microsoft.ML.OnnxRuntime
+  RunTests : 'true'
 jobs:
 - job: ${{ parameters.JobName }}
   timeoutInMinutes: 120
@@ -133,18 +134,20 @@ jobs:
           arguments: '--configuration RelWithDebInfo -p:Platform="Any CPU" -p:OnnxRuntimeBuildDirectory="$(Build.BinariesDirectory)" -p:OrtPackageId=${{ parameters.OrtPackageId }}'
           workingDirectory: '$(Build.SourcesDirectory)\csharp'
 
-      - script: |
-         mklink  /D /J $(Build.BinariesDirectory)\RelWithDebInfo\models $(Build.BinariesDirectory)\models
-         DIR dist\ /S /B > wheel_filename_file
-         set /p WHEEL_FILENAME=<wheel_filename_file
-         del wheel_filename_file
-         python.exe -m pip install -q --upgrade %WHEEL_FILENAME%
-         set PATH=%PATH%;$(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo
-         @echo %PATH%
-         python $(Build.SourcesDirectory)\tools\ci_build\build.py $(BuildCommand) --test --config RelWithDebInfo --enable_lto
-        workingDirectory: '$(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo'
-        displayName: 'Run tests'
+      - ${{ if eq(parameters.RunTests, true) }}:
+        - script: |
+           mklink  /D /J $(Build.BinariesDirectory)\RelWithDebInfo\models $(Build.BinariesDirectory)\models
+           DIR dist\ /S /B > wheel_filename_file
+           set /p WHEEL_FILENAME=<wheel_filename_file
+           del wheel_filename_file
+           python.exe -m pip install -q --upgrade %WHEEL_FILENAME%
+           set PATH=%PATH%;$(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo
+           @echo %PATH%
+           python $(Build.SourcesDirectory)\tools\ci_build\build.py $(BuildCommand) --test --config RelWithDebInfo --enable_lto
+          workingDirectory: '$(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo'
+          displayName: 'Run tests'
 
+    - ${{ if eq(parameters.RunTests, true) }}:
       - task: PublishTestResults@2
         displayName: 'Publish unit test results'
         inputs:
@@ -180,7 +183,7 @@ jobs:
         displayName: 'Create NuGet Package'
 
       - task: PublishPipelineArtifact@0
-        displayName: 'Publish Pipeline Artifact: drop-nuget'
+        displayName: 'Publish Pipeline Artifact: ${{ parameters.ArtifactName }}'
         inputs:
           artifactName: ${{ parameters.ArtifactName }}
           targetPath: '$(Build.ArtifactStagingDirectory)'

tools/ci_build/github/azure-pipelines/win-gpu-ci-pipeline.yml

@@ -13,7 +13,7 @@ jobs:
     MsbuildArguments: '-detailedsummary -maxcpucount -consoleloggerparameters:PerformanceSummary'
     OnnxRuntimeBuildDirectory: '$(Build.BinariesDirectory)'
     DOTNET_SKIP_FIRST_TIME_EXPERIENCE: true
-    EnvSetupScript: setup_env.bat
+    EnvSetupScript: setup_env_cuda.bat
     buildArch: x64
     setVcvars: true    
   timeoutInMinutes: 120

tools/ci_build/github/azure-pipelines/win-gpu-tensorrt-ci-pipeline.yml

@@ -13,7 +13,7 @@ jobs:
     MsbuildArguments: '-detailedsummary -maxcpucount -consoleloggerparameters:PerformanceSummary'
     OnnxRuntimeBuildDirectory: '$(Build.BinariesDirectory)'
     DOTNET_SKIP_FIRST_TIME_EXPERIENCE: true
-    EnvSetupScript: setup_env.bat
+    EnvSetupScript: setup_env_trt.bat
     buildArch: x64
     setVcvars: true
   timeoutInMinutes: 120

tools/ci_build/github/windows/setup_env_cuda.bat

@@ -0,0 +1,2 @@
+set PATH=C:\azcopy;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\bin;C:\local\cudnn-10.1-windows10-x64-v7.6.5.32\cuda\bin;%PATH%
+set GRADLE_OPTS=-Dorg.gradle.daemon=false

tools/ci_build/github/windows/setup_env_trt.bat

@@ -0,0 +1,2 @@
+set PATH=C:\azcopy;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\bin;C:\local\cudnn-10.2-windows10-x64-v7.6.5.32\cuda\bin;%PATH%
+set GRADLE_OPTS=-Dorg.gradle.daemon=false

tools/nuget/generate_nuspec_for_native_nuget.py

@@ -14,7 +14,8 @@ def parse_arguments():
     parser.add_argument("--target_architecture", required=True, help="Eg: x64")
     parser.add_argument("--build_config", required=True, help="Eg: RelWithDebInfo")
     parser.add_argument("--ort_build_path", required=True, help="ORT build directory.")
-    parser.add_argument("--native_build_path", required=True, help="Native build output directory..")
+    parser.add_argument("--native_build_path", required=True, help="Native build output directory.")
+    parser.add_argument("--packages_path", required=True, help="Nuget packages output directory.")
     parser.add_argument("--sources_path", required=True, help="OnnxRuntime source code root.")
     parser.add_argument("--commit_id", required=True, help="The last commit id included in this package.")
     parser.add_argument("--is_release_build", required=False, default=None, type=str, help="Flag indicating if the build is a release build. Accepted values: true/false.")        
@@ -125,7 +126,8 @@ def generate_files(list, args):
         # Process DirectML dll
         if os.path.exists(os.path.join(args.native_build_path, 'DirectML.dll')):
             files_list.append('<file src=' + '"' + os.path.join(args.native_build_path, 'DirectML.dll') + '" target="runtimes\\win-' + args.target_architecture + '\\native" />')
-        
+            files_list.append('<file src=' + '"' + os.path.join(args.native_build_path, 'DirectML.pdb') + '" target="runtimes\\win-' + args.target_architecture + '\\native" />')
+            files_list.append('<file src=' + '"' + os.path.join(args.packages_path, 'DirectML.0.0.1\\LICENSE.txt') + '" target="DirectML_LICENSE.txt" />')
         
     # Process onnxruntime import lib, dll, and pdb
     files_list.append('<file src=' + '"' + os.path.join(args.native_build_path, 'onnxruntime.lib') + '" target="runtimes\\win-' + args.target_architecture + '\\native" />')

tools/nuget/validate_nuget.py

@@ -11,6 +11,7 @@ def parse_arguments():
     parser = argparse.ArgumentParser(description="Validate ONNX Runtime native nuget containing native shared library artifacts spec script",
                                      usage='')
     # Main arguments
+    parser.add_argument("--nuget_package", required=True, help="Nuget package name to be validated.")
     parser.add_argument("--nuget_path", required=True, help="Path containing the Nuget to be validated. Must only contain only one Nuget within this.")
     parser.add_argument("--platforms_supported", required=True, help="Comma separated list (no space). Ex: linux-x64,win-x86,osx-x64")
     parser.add_argument("--verify_nuget_signing", required=True, help="Flag inidicating if Nuget package signing is to be verified. Only accepets 'true' or 'false'")
@@ -69,13 +70,12 @@ def check_if_nuget_is_signed(nuget_path):
 def main():   
     args = parse_arguments()
 
-    files = os.listdir(args.nuget_path)
+    files = glob.glob(os.path.join(args.nuget_path, args.nuget_package))
     nuget_packages_found_in_path = [i for i in files if i.endswith('.nupkg') and "Managed" not in i]
     if (len(nuget_packages_found_in_path) != 1):
         print('Nuget packages found in path: ')
         print(nuget_packages_found_in_path)
         raise Exception('No Nuget packages / more than one Nuget packages found in the given path.')
-    
     nuget_file_name = nuget_packages_found_in_path[0]
     full_nuget_path = os.path.join(args.nuget_path, nuget_file_name)