mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-07-27 20:02:15 +00:00
Enable more ROCM ops that are sharing CUDA code. Some are needed for Turing NLG models.
This commit is contained in:
Jesse Benson
Jesse Benson
parent
dbe31361bc
commit
d18aa45b46
4 changed files with 30 additions and 47 deletions
|
|
@ -570,8 +570,8 @@ class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kO
|
|||
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 2, 10, float, Pad);
|
||||
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 2, 10, double, Pad);
|
||||
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 2, 10, MLFloat16, Pad);
|
||||
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 4, Reshape);
|
||||
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 5, 12, Reshape);
|
||||
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 4, Reshape_1);
|
||||
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 12, Shape);
|
||||
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 12, Size);
|
||||
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 6, 12, Tile);
|
||||
|
|
@ -1255,12 +1255,12 @@ static Status RegisterRocmKernels(KernelRegistry& kernel_registry) {
|
|||
// BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 2, 10, float, Pad)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 2, 10, double, Pad)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 2, 10, MLFloat16, Pad)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 4, Reshape)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 5, 12, Reshape)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 4, Reshape_1)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 12, Shape)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 12, Size)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 6, 12, Tile)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Tile)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 12, Size)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 6, 12, Tile)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Tile)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 1, 12, Transpose)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 6, float, InstanceNormalization)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 6, double, InstanceNormalization)>,
|
||||
|
|
@ -1512,13 +1512,13 @@ static Status RegisterRocmKernels(KernelRegistry& kernel_registry) {
|
|||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, float, Mul)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, double, Mul)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, MLFloat16, Mul)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int32_t, Div)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int64_t, Div)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint32_t, Div)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint64_t, Div)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, float, Div)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, double, Div)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, MLFloat16, Div)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int32_t, Div)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int64_t, Div)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint32_t, Div)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint64_t, Div)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, float, Div)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, double, Div)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, MLFloat16, Div)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int8_t, Abs)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int16_t, Abs)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int32_t, Abs)>,
|
||||
|
|
@ -1584,21 +1584,21 @@ static Status RegisterRocmKernels(KernelRegistry& kernel_registry) {
|
|||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int32_t, NonZero)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int64_t, NonZero)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, float, NonZero)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, float, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, double, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, MLFloat16, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int8_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int16_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int32_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int64_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint8_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint16_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint32_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint64_t, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, bool, Cast)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Reshape)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Shape)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Size)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, float, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, double, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, MLFloat16, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int8_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int16_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int32_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int64_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint8_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint16_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint32_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, uint64_t, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, bool, Cast)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Reshape)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Shape)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Size)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, Transpose)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, ScatterElements)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kOnnxDomain, 13, int32_t, Slice)>,
|
||||
|
|
|
|||
|
|
@ -235,9 +235,9 @@ Status RegisterRocmTrainingKernels(KernelRegistry& kernel_registry) {
|
|||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, MLFloat16_float, InvertibleLayerNormalizationGrad)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, SliceGrad)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, GatherElementsGrad)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, MLFloat16, Scale)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, float, Scale)>,
|
||||
// BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, double, Scale)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, MLFloat16, Scale)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, float, Scale)>,
|
||||
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kRocmExecutionProvider, kMSDomain, 1, double, Scale)>,
|
||||
|
||||
// P2P communication operators.
|
||||
#if defined(ORT_USE_NCCL) || defined(USE_MPI)
|
||||
|
|
|
|||
|
|
@ -178,11 +178,6 @@ provider_excluded_files = [
|
|||
'tensor/reverse_sequence.h',
|
||||
'tensor/reverse_sequence_impl.cu',
|
||||
'tensor/reverse_sequence_impl.h',
|
||||
'tensor/size.cc',
|
||||
'tensor/tile.cc',
|
||||
'tensor/tile.h',
|
||||
'tensor/tile_impl.cu',
|
||||
'tensor/tile_impl.h',
|
||||
'tensor/transpose.cc',
|
||||
'tensor/transpose.h',
|
||||
'tensor/upsample.cc',
|
||||
|
|
@ -238,9 +233,6 @@ training_ops_excluded_files = [
|
|||
'math/div_grad.h',
|
||||
'math/div_grad_impl.cu',
|
||||
'math/div_grad_impl.h',
|
||||
'math/scale.cc',
|
||||
'math/scale.cu',
|
||||
'math/scale.h',
|
||||
'math/softmax_grad_impl.cu',
|
||||
'math/softmax_grad.cc',
|
||||
'nn/batch_norm_grad.cc',
|
||||
|
|
|
|||
|
|
@ -20,13 +20,6 @@ CudaKernelTest.SparseSoftmaxCrossEntropy_LargeSizeTensor
|
|||
CudaKernelTest.NegativeLogLikelihoodLoss_TinySizeTensor
|
||||
CudaKernelTest.NegativeLogLikelihoodLoss_SmallSizeTensor
|
||||
CudaKernelTest.NegativeLogLikelihoodLoss_MediumSizeTensor
|
||||
CudaKernelTest.ReduceSum_SmallTensorTrailingAxes
|
||||
CudaKernelTest.ReduceSum_MidTensorTrailingAxes
|
||||
CudaKernelTest.ReduceSum_LargeTensorTrailingAxes
|
||||
CudaKernelTest.ScaleHalfHalfScaleUp
|
||||
CudaKernelTest.ScaleHalfInt64ScaleUp
|
||||
CudaKernelTest.ScaleHalfHalfScaleDown
|
||||
CudaKernelTest.ScaleHalfInt64ScaleDown
|
||||
ReductionOpTest.ReductionVariationTest
|
||||
ReductionOpTest.ReduceL1_default_axes_keepdims
|
||||
ReductionOpTest.ReduceL1_do_not_keep_dims
|
||||
|
|
@ -106,8 +99,6 @@ GradientCheckerTest.SubGrad
|
|||
GradientCheckerTest.MulGrad
|
||||
GradientCheckerTest.MatMulGrad
|
||||
GradientCheckerTest.ReduceMeanGrad
|
||||
GradientCheckerTest.ReduceSumGrad
|
||||
GradientCheckerTest.ReduceLogSumExpGrad
|
||||
GradientCheckerTest.ReduceL2Grad
|
||||
GradientCheckerTest.SoftmaxCrossEntropyGrad
|
||||
GradientCheckerTest.ExpandGrad
|
||||
|
|
|
|||
Loading…
Reference in a new issue