Skip to content

[CUDA EP] Add pad op version from 19 to 23 support for CUDA #27416

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ShirasawaSama wants to merge 1 commit into
base: main
Choose a base branch
from
+181 −20
Open

[CUDA EP] Add pad op version from 19 to 23 support for CUDA #27416

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 32 additions & 8 deletions onnxruntime/core/providers/cuda/cuda_execution_provider.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1439,17 +1439,21 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int64_t, ReduceMax);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, ScatterElements);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, ScatterND);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, float, Pad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, double, Pad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, MLFloat16, Pad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, bool, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, float, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, double, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, MLFloat16, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, bool, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, float, Resize);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, double, Resize);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, MLFloat16, Resize);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, int32_t, Resize);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, uint8_t, Resize);

// Opset 19
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, float, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, double, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, MLFloat16, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, bool, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 21, float, AveragePool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 21, double, AveragePool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 21, MLFloat16, AveragePool);
Expand Down Expand Up @@ -1570,6 +1574,10 @@ class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDom
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Transpose);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Squeeze);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Unsqueeze);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, float, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, double, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, MLFloat16, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, bool, Pad);

// Opset 22.
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, float, AveragePool);
Expand Down Expand Up @@ -1619,6 +1627,10 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Float8E4M3FN, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Float8E5M2, Cast);
#endif
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, float, Pad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, double, Pad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, MLFloat16, Pad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, bool, Pad);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Shape);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Reshape);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Transpose);
Expand Down Expand Up @@ -2517,17 +2529,21 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int64_t, ReduceMax)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, ScatterElements)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, ScatterND)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, float, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, double, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, MLFloat16, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, bool, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, float, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, double, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, MLFloat16, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, bool, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, float, Resize)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, double, Resize)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, MLFloat16, Resize)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, int32_t, Resize)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, 18, uint8_t, Resize)>,

// Opset 19-20
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, float, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, double, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, MLFloat16, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 20, bool, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 21, float, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 21, double, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 19, 21, MLFloat16, AveragePool)>,
Expand Down Expand Up @@ -2597,6 +2613,10 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 20, IsNaN)>,

// Opset 21
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, float, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, double, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, MLFloat16, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, bool, Pad)>,
// TODO(fajin): support other quantized types
BuildKernelCreateInfo<ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, uint8_t, float, DequantizeLinear)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, int8_t, float, DequantizeLinear)>,
Expand Down Expand Up @@ -2669,6 +2689,10 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, BFloat16, HardSwish)>,

// Opset 23
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, float, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, double, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, MLFloat16, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, bool, Pad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, float, Attention)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, MLFloat16, Attention)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, BFloat16, Attention)>,
Expand Down
50 changes: 49 additions & 1 deletion onnxruntime/core/providers/cuda/tensor/pad.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,10 +40,46 @@ namespace cuda {
.InputMemoryType(OrtMemTypeCPUInput, 2) \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Pad<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Pad, \
kOnnxDomain, \
18, 18, \
T, \
kCudaExecutionProvider, \
(*KernelDefBuilder::Create()) \
.InputMemoryType(OrtMemTypeCPUInput, 1) \
.InputMemoryType(OrtMemTypeCPUInput, 2) \
.InputMemoryType(OrtMemTypeCPUInput, 3) \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Pad<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Pad, \
kOnnxDomain, \
19, 20, \
T, \
kCudaExecutionProvider, \
(*KernelDefBuilder::Create()) \
.InputMemoryType(OrtMemTypeCPUInput, 1) \
.InputMemoryType(OrtMemTypeCPUInput, 2) \
.InputMemoryType(OrtMemTypeCPUInput, 3) \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Pad<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Pad, \
kOnnxDomain, \
21, 22, \
T, \
kCudaExecutionProvider, \
(*KernelDefBuilder::Create()) \
.InputMemoryType(OrtMemTypeCPUInput, 1) \
.InputMemoryType(OrtMemTypeCPUInput, 2) \
.InputMemoryType(OrtMemTypeCPUInput, 3) \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Pad<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
Pad, \
kOnnxDomain, \
18, \
23, \
T, \
kCudaExecutionProvider, \
(*KernelDefBuilder::Create()) \
Expand Down Expand Up @@ -154,6 +190,12 @@ Status Pad<T>::ComputeInternal(OpKernelContext* ctx) const {
effective_input_extents.push_back(extent);
}

TArray<int64_t> slice_starts(dimension_count);
for (size_t i = 0; i < dimension_count; ++i) {
slice_starts[i] = (*p_slices)[i];
}
TArray<int64_t> effective_input_dims(effective_input_extents);

TensorShape output_shape(output_dims);
auto& output_tensor = *ctx->Output(0, output_shape);

Expand Down Expand Up @@ -248,6 +290,10 @@ Status Pad<T>::ComputeInternal(OpKernelContext* ctx) const {
output_dims[width_dim],
lower_pads[height_dim],
lower_pads[width_dim],
slice_starts[height_dim],
slice_starts[width_dim],
effective_input_dims[height_dim],
effective_input_dims[width_dim],
value,
static_cast<int>(mode_),
reinterpret_cast<const typename ToCudaType<T>::MappedType*>(input_tensor.Data<T>()),
Expand All @@ -269,6 +315,8 @@ Status Pad<T>::ComputeInternal(OpKernelContext* ctx) const {
input_dims,
input_strides,
lower_pads,
slice_starts,
effective_input_dims,
value,
static_cast<int>(mode_),
reinterpret_cast<const typename ToCudaType<T>::MappedType*>(input_tensor.Data<T>()),
Expand Down
Loading