[DML EP] Update DirectML version to 1.12.0 (microsoft#16011)

edgchen1 · May 19, 2023 · 310b22a · 310b22a
1 parent d78bbf5
commit 310b22a
Show file tree

Hide file tree

Showing 9 changed files with 13 additions and 18 deletions.
diff --git a/.pipelines/nuget_config/x64/packages.config b/.pipelines/nuget_config/x64/packages.config
@@ -1,6 +1,6 @@
 <?xml version="1.0" encoding="utf-8"?>
 <packages>
   <package id="python" version="3.9.7" targetFramework="native" />
-  <package id="Microsoft.AI.DirectML" version="1.11.0" targetFramework="native" />
+  <package id="Microsoft.AI.DirectML" version="1.12.0" targetFramework="native" />
   <package id="Microsoft.Windows.CppWinRT" version="2.0.201201.7" targetFramework="native" />
 </packages>
diff --git a/.pipelines/nuget_config/x86/packages.config b/.pipelines/nuget_config/x86/packages.config
@@ -1,6 +1,6 @@
 <?xml version="1.0" encoding="utf-8"?>
 <packages>
   <package id="pythonx86" version="3.9.7" targetFramework="native" />
-  <package id="Microsoft.AI.DirectML" version="1.11.0" targetFramework="native" />
+  <package id="Microsoft.AI.DirectML" version="1.12.0" targetFramework="native" />
   <package id="Microsoft.Windows.CppWinRT" version="2.0.201201.7" targetFramework="native" />
 </packages>
diff --git a/cmake/external/dml.cmake b/cmake/external/dml.cmake
@@ -41,7 +41,7 @@ if (NOT onnxruntime_USE_CUSTOM_DIRECTML)
   set(NUGET_CONFIG ${PROJECT_SOURCE_DIR}/../NuGet.config)
   set(PACKAGES_CONFIG ${PROJECT_SOURCE_DIR}/../packages.config)
   get_filename_component(PACKAGES_DIR ${CMAKE_CURRENT_BINARY_DIR}/../packages ABSOLUTE)
-  set(DML_PACKAGE_DIR ${PACKAGES_DIR}/Microsoft.AI.DirectML.1.11.0)
+  set(DML_PACKAGE_DIR ${PACKAGES_DIR}/Microsoft.AI.DirectML.1.12.0)
 
   # Restore nuget packages, which will pull down the DirectML redist package.
   add_custom_command(

diff --git a/docs/OperatorKernels.md b/docs/OperatorKernels.md
@@ -1204,6 +1204,7 @@ Do not modify directly.*
 |DequantizeLinear|*in* x:**T1**<br> *in* x_scale:**T2**<br> *in* x_zero_point:**T1**<br> *out* y:**T2**|1+|**T1** = tensor(float)<br/> **T2** = tensor(uint8)|
 |EmbedLayerNormalization|*in* input_ids:**T1**<br> *in* segment_ids:**T1**<br> *in* word_embedding:**T**<br> *in* position_embedding:**T**<br> *in* segment_embedding:**T**<br> *in* gamma:**T**<br> *in* beta:**T**<br> *in* mask:**T1**<br> *in* position_ids:**T1**<br> *out* output:**T**<br> *out* mask_index:**T1**<br> *out* embedding_sum:**T**|1+|**T** = tensor(float), tensor(float16)|
 |FusedMatMul|*in* A:**T**<br> *in* B:**T**<br> *out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|
+|FusedMatMulActivation|*in* A:**T**<br> *in* B:**T**<br> *out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|
 |Gelu|*in* X:**T**<br> *out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|
 |GroupNorm|*in* X:**T**<br> *in* gamma:**M**<br> *in* beta:**M**<br> *out* Y:**T**|1+|**M** = tensor(float), tensor(float16)<br/> **T** = tensor(float), tensor(float16)|
 |NhwcConv|*in* X:**T**<br> *in* W:**T**<br> *in* B:**T**<br> *out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|

diff --git a/onnxruntime/core/optimizer/graph_transformer_utils.cc b/onnxruntime/core/optimizer/graph_transformer_utils.cc
@@ -182,6 +182,7 @@ InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
 #ifndef DISABLE_CONTRIB_OPS
   const InlinedHashSet<std::string_view> cpu_ep = {onnxruntime::kCpuExecutionProvider};
 #endif
+  const InlinedHashSet<std::string_view> dml_ep = {onnxruntime::kDmlExecutionProvider};
   switch (level) {
     case TransformerLevel::Level1: {
       // RewriteRule optimizations are the simplest (they generally remove unnecessary nodes and are cheap to run)
@@ -308,6 +309,7 @@ InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
       transformers.emplace_back(std::make_unique<QuickGeluFusion>(cpu_cuda_dml_rocm_eps));
 
       transformers.emplace_back(std::make_unique<MatMulScaleFusion>(cpu_cuda_dml_rocm_eps));
+      transformers.emplace_back(std::make_unique<MatMulActivationFusion>(dml_ep));
 
       // GeluApproximation has side effects which may change results. It needs to be manually enabled,
       // or alternatively the model can be updated offline using a model conversion script

diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp
@@ -934,6 +934,7 @@ constexpr static OperatorRegistrationInformation operatorRegistrationInformation
     {REG_INFO_MS(   1,  Gelu,                               typeNameListDefault,            supportedTypeListFloat16to32,           DmlGraphSupport::Supported)},
     {REG_INFO_MS(   1,  BiasGelu,                           typeNameListDefault,            supportedTypeListFloat16to32,           DmlGraphSupport::Supported)},
     {REG_INFO_MS(   1,  FusedMatMul,                        typeNameListDefault,            supportedTypeListFloat16to32,           DmlGraphSupport::Supported)},
+    {REG_INFO_MS(   1,  FusedMatMulActivation,              typeNameListDefault,            supportedTypeListFloat16to32,           DmlGraphSupport::Supported)},
     {REG_INFO_MS(   1,  QLinearSigmoid,                     typeNameListDefault,            supportedTypeListQLinearSigmoid,        DmlGraphSupport::Supported, requiredConstantCpuInputs(), std::nullopt, QueryQLinearSigmoid)},
     {REG_INFO_MS(   1,  Attention,                          typeNameListAttention,          supportedTypeListAttention,             DmlGraphSupport::Supported, requiredConstantCpuInputs(), std::nullopt, QueryAttention)},
 

diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorUtility.cpp b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorUtility.cpp
@@ -364,19 +364,10 @@ namespace Dml
                 int onnxAxis = HandleNegativeAxis(kernelInfo.GetOptionalAttribute<int>(AttrName::GraphFusedAxis, -1), onnxDimCount);
 
                 auto dmlAdjustedAxis = GetDmlAdjustedAxis(onnxAxis, onnxDimCount, kernelInfo.GetTensorShapeDescription().GetInputTensorDimensionCount(0));
-
-                // If the axis is supported by Softmax, use this version instead since it's more likely to be supported by metacommands
-                if (dmlAdjustedAxis == onnxDimCount - 1)
-                {
-                    activation.desc.activationType = DML_OPERATOR_ACTIVATION_SOFTMAX;
-                }
-                else
-                {
-                    activation.desc.activationType = DML_OPERATOR_ACTIVATION_SOFTMAX1;
-                    activation.dmlAxes.push_back(dmlAdjustedAxis);
-                    activation.desc.params.softmax1.Axes = activation.dmlAxes.data();
-                    activation.desc.params.softmax1.AxisCount = gsl::narrow_cast<uint32_t>(activation.dmlAxes.size());
-                }
+                activation.desc.activationType = DML_OPERATOR_ACTIVATION_SOFTMAX1;
+                activation.dmlAxes.push_back(dmlAdjustedAxis);
+                activation.desc.params.softmax1.Axes = activation.dmlAxes.data();
+                activation.desc.params.softmax1.AxisCount = gsl::narrow_cast<uint32_t>(activation.dmlAxes.size());
             }
             else
             {

diff --git a/packages.config b/packages.config
@@ -1,6 +1,6 @@
 <?xml version="1.0" encoding="utf-8"?>
 <packages>
-  <package id="Microsoft.AI.DirectML" version="1.11.0" targetFramework="native" />
+  <package id="Microsoft.AI.DirectML" version="1.12.0" targetFramework="native" />
   <package id="Microsoft.Windows.CppWinRT" version="2.0.201201.7" targetFramework="native" />
   <package id="google.protobuf.tools" version="3.21.12" targetFramework="native" />
 </packages>
diff --git a/tools/nuget/generate_nuspec_for_native_nuget.py b/tools/nuget/generate_nuspec_for_native_nuget.py
@@ -190,7 +190,7 @@ def generate_repo_url(line_list, repo_url, commit_id):
 
 
 def generate_dependencies(xml_text, package_name, version):
-    dml_dependency = '<dependency id="Microsoft.AI.DirectML" version="1.11.0"/>'
+    dml_dependency = '<dependency id="Microsoft.AI.DirectML" version="1.12.0"/>'
 
     if package_name == "Microsoft.AI.MachineLearning":
         xml_text.append("<dependencies>")