From b508835e7a250d32e3cab79c67cf01b99a9ab356 Mon Sep 17 00:00:00 2001
From: KeDengMS <kedeng@microsoft.com>
Date: Sat, 29 Dec 2018 19:57:45 -0800
Subject: [PATCH] Update CUDA Gemm to opset 9 (#266)

---
 .../core/providers/cuda/cuda_execution_provider.cc   | 12 ++++++------
 onnxruntime/core/providers/cuda/math/gemm.cc         |  3 ++-
 2 files changed, 8 insertions(+), 7 deletions(-)

diff --git a/onnxruntime/core/providers/cuda/cuda_execution_provider.cc b/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
index 40a60c0909..96b7d7d3f1 100644
--- a/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
+++ b/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
@@ -250,9 +250,9 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Sq
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Identity);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, Dropout);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Gather);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, float, Gemm);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, double, Gemm);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, MLFloat16, Gemm);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, float, Gemm);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, double, Gemm);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, MLFloat16, Gemm);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, MatMul);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, MatMul);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, MatMul);
@@ -510,9 +510,9 @@ static void RegisterCudaKernels(std::function<void(KernelCreateInfo&&)> fn) {
   fn(BuildKernel<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Identity)>());
   fn(BuildKernel<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, Dropout)>());
   fn(BuildKernel<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Gather)>());
-  fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, float, Gemm)>());
-  fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, double, Gemm)>());
-  fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, MLFloat16, Gemm)>());
+  fn(BuildKernel<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, float, Gemm)>());
+  fn(BuildKernel<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, double, Gemm)>());
+  fn(BuildKernel<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, MLFloat16, Gemm)>());
   fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, MatMul)>());
   fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, MatMul)>());
   fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, MatMul)>());
diff --git a/onnxruntime/core/providers/cuda/math/gemm.cc b/onnxruntime/core/providers/cuda/math/gemm.cc
index 9b4ef08678..0bfb1f2454 100644
--- a/onnxruntime/core/providers/cuda/math/gemm.cc
+++ b/onnxruntime/core/providers/cuda/math/gemm.cc
@@ -10,10 +10,11 @@ namespace onnxruntime {
 namespace cuda {
 
 #define REGISTER_KERNEL_TYPED(T)                                  \
-  ONNX_OPERATOR_TYPED_KERNEL_EX(                                  \
+  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(                        \
       Gemm,                                                       \
       kOnnxDomain,                                                \
       7,                                                          \
+      9,                                                          \
       T,                                                          \
       kCudaExecutionProvider,                                     \
       KernelDefBuilder()                                          \