Update grid_sample_test.cc to run on all execution providers & change GridSample NHWC version to 16.

Author: mtavenrath

onnxruntime/contrib_ops/cuda/cuda_contrib_kernels.cc

@@ -204,7 +204,7 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1
 #endif
 
 #ifdef ENABLE_CUDA_NHWC_OPS
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSInternalNHWCDomain, 20, float, GridSample);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSInternalNHWCDomain, 16, float, GridSample);
 #endif
 
 template <>
@@ -413,7 +413,7 @@ Status RegisterCudaContribKernels(KernelRegistry& kernel_registry) {
 #endif
 
 #ifdef ENABLE_CUDA_NHWC_OPS
-    BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSInternalNHWCDomain, 20, float, GridSample)>,
+    BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSInternalNHWCDomain, 16, float, GridSample)>,
 #endif
   };
 

onnxruntime/contrib_ops/cuda/grid_sample.cc

@@ -19,10 +19,10 @@ namespace cuda {
       (*KernelDefBuilder::Create())                                \
           .TypeConstraint("T1", DataTypeImpl::GetTensorType<T>())  \
           .TypeConstraint("T2", DataTypeImpl::GetTensorType<T>()), \
-      GridSample<T, LAYOUT>);
+      onnxruntime::contrib::cuda::GridSample<T, LAYOUT>);
 
 REGISTER_KERNEL_TYPED(float, 1, LAYOUT_NCHW, kMSDomain)
-REGISTER_KERNEL_TYPED(float, 20, LAYOUT_NHWC, kMSInternalNHWCDomain)
+REGISTER_KERNEL_TYPED(float, 16, LAYOUT_NHWC, kMSInternalNHWCDomain)
 
 template <typename T, bool IsNHWC>
 GridSample<T, IsNHWC>::GridSample(const OpKernelInfo& info) : CudaKernel(info) {
@@ -92,4 +92,8 @@ Status GridSample<T, IsNHWC>::ComputeInternal(OpKernelContext* context) const {
 }
 }  // namespace cuda
 }  // namespace contrib
+
+namespace cuda {
+  REGISTER_KERNEL_TYPED(float, 16, LAYOUT_NCHW, kOnnxDomain)
+}  // namespace cuda
 }  // namespace onnxruntime

onnxruntime/core/providers/cuda/cuda_execution_provider.cc

@@ -1256,6 +1256,7 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, double, LessOrEqual);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, MLFloat16, LessOrEqual);
 class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, 17, ScatterElements);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, float, GridSample);
 
 // Opset 17
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 17, float, LayerNormalization);
@@ -2143,6 +2144,7 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
     BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, double, LessOrEqual)>,
     BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, MLFloat16, LessOrEqual)>,
     BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, 17, ScatterElements)>,
+    BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 16, float, GridSample)>,
 
     // Opset 17
     BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 17, float, LayerNormalization)>,