saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-10 17:37:14 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Register OpSet13 CUDA Kernels for BERT/UniLMv2 (#4856)

Browse source 
* opset13 cuda kernels for BERT.

* add opset13 SoftmaxCrossEntropyLoss.

* opset13 size.

* fix argmax/min for ut.

* fix ut failure for argmax/min.

* OrtMemTypeCPUInput

Co-authored-by: Vincent Wang <weicwang@microsoft.com>
This commit is contained in:
Vincent Wang 2020-09-05 08:09:52 +08:00 committed by GitHub
parent 370d194db7
commit 84de14a833
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
31 changed files with 1213 additions and 423 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

17
include/onnxruntime/core/framework/op_kernel.h
View file

@ -432,6 +432,23 @@ using BuildKernelCreateInfoFn = KernelCreateInfo (*)();
static_cast<KernelCreatePtrFn>([](const OpKernelInfo& info) -> OpKernel* { return new __VA_ARGS__(info); })); \
}
#define ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type1, type2, name) \
provider##_##name##_##domain##_ver##startver##_##endver##_##type1##_##type2
#define ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_EX(name, domain, startver, endver, type1, type2, provider, builder, ...) \
class ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type1, type2, name); \
template <> \
KernelCreateInfo \
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type1, type2, name)>() { \
return KernelCreateInfo( \
builder.SetName(#name) \
.SetDomain(domain) \
.SinceVersion(startver, endver) \
.Provider(provider) \
.Build(), \
static_cast<KernelCreatePtrFn>([](const OpKernelInfo& info) -> OpKernel* { return new __VA_ARGS__(info); })); \
}
// Use within macro definitions to create a custom vector of constraints.
// Example: #define REG_KERNEL(OP, VERSION, KERNEL_CLASS, Type, ...)
// .TypeConstraint("T", BuildKernelDefConstraints<Type, __VA_ARGS_>())

6
onnxruntime/core/providers/cpu/cpu_execution_provider.cc
View file

@ -187,7 +187,7 @@ class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDoma
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 4, Reshape_1);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 5, Reshape);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Shape);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Size);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 12, Size);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 9, Slice);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, SpaceToDepth);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 10, DepthToSpace);
@ -456,6 +456,7 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 12, double_double, Dropout);
// opset 13
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, Size);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, Cast);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, Sign);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, uint8_t, DequantizeLinear);
@ -710,7 +711,7 @@ Status RegisterOnnxOperatorKernels(KernelRegistry& kernel_registry) {
Reshape_1)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 5, Reshape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Shape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Size)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 12, Size)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 9,
Slice)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, SpaceToDepth)>,
@ -1159,6 +1160,7 @@ Status RegisterOnnxOperatorKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 12, double_double, Dropout)>,
// opset 13
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, Size)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, Cast)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, Sign)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, uint8_t,

21
onnxruntime/core/providers/cpu/tensor/size.cc
View file

@ -26,9 +26,28 @@ Status Size::Compute(OpKernelContext* ctx) const {
// omit this.
// TODO: Both onnxruntime and ONNX lists of types seem somewhat incomplete and incomparable.
ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
Size,
1, 12,
KernelDefBuilder().TypeConstraint("T",
std::vector<MLDataType>({DataTypeImpl::GetTensorType<float>(),
DataTypeImpl::GetTensorType<double>(),
DataTypeImpl::GetTensorType<int8_t>(),
DataTypeImpl::GetTensorType<int16_t>(),
DataTypeImpl::GetTensorType<int32_t>(),
DataTypeImpl::GetTensorType<int64_t>(),
DataTypeImpl::GetTensorType<uint8_t>(),
DataTypeImpl::GetTensorType<uint16_t>(),
DataTypeImpl::GetTensorType<uint32_t>(),
DataTypeImpl::GetTensorType<uint64_t>(),
DataTypeImpl::GetTensorType<std::string>(),
DataTypeImpl::GetTensorType<bool>()}))
.TypeConstraint("T1", DataTypeImpl::GetTensorType<int64_t>()),
Size);
ONNX_CPU_OPERATOR_KERNEL(
Size,
1,
13,
KernelDefBuilder().TypeConstraint("T",
std::vector<MLDataType>({DataTypeImpl::GetTensorType<float>(),
DataTypeImpl::GetTensorType<double>(),

30
onnxruntime/core/providers/cuda/activation/activations.cc
View file

@ -6,6 +6,19 @@
namespace onnxruntime {
namespace cuda {
#define REGISTER_ACTIVATION_VERSIONED_KERNEL(x, startver, endver, T) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
x, \
kOnnxDomain, \
startver, \
endver, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder() \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()) \
.MayInplace(0, 0), \
x<T>);
#define REGISTER_ACTIVATION_KERNEL(x, ver, T) \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
x, \
@ -32,6 +45,14 @@ namespace cuda {
return Status::OK(); \
}
#define UNARY_ACTIVATION_OP_VERSIONED_TYPED(name, startver, endver, T) \
REGISTER_ACTIVATION_VERSIONED_KERNEL(name, startver, endver, T)
#define UNARY_ACTIVATION_OP_VERSIONED_HFD(name, startver, endver) \
UNARY_ACTIVATION_OP_VERSIONED_TYPED(name, startver, endver, MLFloat16) \
UNARY_ACTIVATION_OP_VERSIONED_TYPED(name, startver, endver, float) \
UNARY_ACTIVATION_OP_VERSIONED_TYPED(name, startver, endver, double)
#define UNARY_ACTIVATION_OP_TYPED(name, ver, T) \
REGISTER_ACTIVATION_KERNEL(name, ver, T) \
UNARY_ACTIVATION_COMPUTE(name, T)
@ -44,12 +65,15 @@ namespace cuda {
UNARY_ACTIVATION_OP_HFD(Elu, 6);
UNARY_ACTIVATION_OP_HFD(HardSigmoid, 6);
UNARY_ACTIVATION_OP_HFD(LeakyRelu, 6);
UNARY_ACTIVATION_OP_HFD(Relu, 6);
UNARY_ACTIVATION_OP_HFD(Relu, 13);
UNARY_ACTIVATION_OP_VERSIONED_HFD(Relu, 6, 12);
UNARY_ACTIVATION_OP_HFD(Selu, 6);
UNARY_ACTIVATION_OP_HFD(Sigmoid, 6);
UNARY_ACTIVATION_OP_HFD(Sigmoid, 13);
UNARY_ACTIVATION_OP_VERSIONED_HFD(Sigmoid, 6, 12);
UNARY_ACTIVATION_OP_HFD(Softplus, 1);
UNARY_ACTIVATION_OP_HFD(Softsign, 1);
UNARY_ACTIVATION_OP_HFD(Tanh, 6);
UNARY_ACTIVATION_OP_HFD(Tanh, 13);
UNARY_ACTIVATION_OP_VERSIONED_HFD(Tanh, 6, 12);
UNARY_ACTIVATION_OP_HFD(ThresholdedRelu, 10);
} // namespace cuda

1090
onnxruntime/core/providers/cuda/cuda_execution_provider.cc
View file

File diff suppressed because it is too large Load diff

50
onnxruntime/core/providers/cuda/math/binary_elementwise_ops.cc
View file

@ -158,6 +158,9 @@ Status BinaryElementwise<ShouldBroadcast>::Prepare(OpKernelContext* context, Bin
return Status::OK(); \
}
#define BINARY_OP_VERSIONED_TYPED(name, startver, endver, T) \
BINARY_ELEMENTWISE_REGISTER_KERNEL_VERSIONED_TYPED(name, startver, endver, T)
#define BINARY_OP_TYPED(name, ver, T) \
BINARY_ELEMENTWISE_REGISTER_KERNEL_TYPED(name, ver, T) \
BINARY_ELEMENTWISE_COMPUTE(name, T)
@ -185,6 +188,18 @@ Status BinaryElementwise<ShouldBroadcast>::Prepare(OpKernelContext* context, Bin
// D: double
// O: bool
#define BINARY_OP_VERSIONED_HFD(name, startver, endver) \
BINARY_OP_VERSIONED_TYPED(name, startver, endver, MLFloat16) \
BINARY_OP_VERSIONED_TYPED(name, startver, endver, float) \
BINARY_OP_VERSIONED_TYPED(name, startver, endver, double)
#define BINARY_OP_VERSIONED_UZILHFD(name, startver, endver) \
BINARY_OP_VERSIONED_TYPED(name, startver, endver, uint32_t) \
BINARY_OP_VERSIONED_TYPED(name, startver, endver, uint64_t) \
BINARY_OP_VERSIONED_TYPED(name, startver, endver, int32_t) \
BINARY_OP_VERSIONED_TYPED(name, startver, endver, int64_t) \
BINARY_OP_VERSIONED_HFD(name, startver, endver)
#define BINARY_OP_HFD(name, ver) \
BINARY_OP_TYPED(name, ver, MLFloat16) \
BINARY_OP_TYPED(name, ver, float) \
@ -250,21 +265,35 @@ Status BinaryElementwise<ShouldBroadcast>::Prepare(OpKernelContext* context, Bin
BINARY_ELEMENTWISE_REGISTER_KERNEL_VERSIONED_TYPED(name, startver, endver, int64_t) \
BINARY_OP_REGISTER_VERSIONED_HFD(name, startver, endver)
BINARY_OP_UZILHFD(Add, 7)
BINARY_OP_UZILHFD(Sub, 7)
BINARY_OP_UZILHFD(Mul, 7)
BINARY_OP_UZILHFD(Div, 7)
BINARY_OP_VERSIONED_UZILHFD(Add, 7, 12)
BINARY_OP_VERSIONED_UZILHFD(Sub, 7, 12)
BINARY_OP_VERSIONED_UZILHFD(Mul, 7, 12)
BINARY_OP_VERSIONED_UZILHFD(Div, 7, 12)
BINARY_OP_UZILHFD(Add, 13)
BINARY_OP_UZILHFD(Sub, 13)
BINARY_OP_UZILHFD(Mul, 13)
BINARY_OP_UZILHFD(Div, 13)
BINARY_OP_REGISTER_VERSIONED_CLASS_HFD(Pow, Pow_7, 7, 11)
BINARY_LOGICALOP_TYPED(And, 7, bool)
BINARY_LOGICALOP_TYPED(Or, 7, bool)
BINARY_LOGICALOP_TYPED(Xor, 7, bool)
BINARY_OP_HFD(PRelu, 7)
// Pow version 12
// Pow since version 12
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Pow,
kOnnxDomain,
12, 12,
kCudaExecutionProvider,
KernelDefBuilder().TypeConstraint("T", BuildKernelDefConstraints<int32_t, int64_t, float, double>()).TypeConstraint("T1", BuildKernelDefConstraints<int32_t, int64_t, float, double>()),
Pow);
ONNX_OPERATOR_KERNEL_EX(
Pow,
kOnnxDomain,
12,
13,
kCudaExecutionProvider,
KernelDefBuilder().TypeConstraint("T", BuildKernelDefConstraints<int32_t, int64_t, float, double>()).TypeConstraint("T1", BuildKernelDefConstraints<int32_t, int64_t, float, double>()),
Pow);
@ -410,11 +439,14 @@ Status Less<T>::ComputeInternal(OpKernelContext* context) const {
return Status::OK();
}
BINARY_LOGICALOP_REGISTER_UZILHFD(Greater, 9)
BINARY_OP_REGISTER_OIL(Equal, 13)
BINARY_OP_REGISTER_VERSIONED_OIL(Equal, 11, 12)
BINARY_OP_REGISTER_VERSIONED_OIL(Equal, 7, 10)
BINARY_OP_REGISTER_OIL(Equal, 11)
BINARY_LOGICALOP_REGISTER_UZILHFD(Greater, 13)
BINARY_OP_REGISTER_VERSIONED_UZILHFD(Greater, 9, 12)
BINARY_OP_REGISTER_VERSIONED_HFD(Greater, 7, 8)
BINARY_LOGICALOP_REGISTER_UZILHFD(Less, 9)
BINARY_LOGICALOP_REGISTER_UZILHFD(Less, 13)
BINARY_OP_REGISTER_VERSIONED_UZILHFD(Less, 9, 12)
BINARY_OP_REGISTER_VERSIONED_HFD(Less, 7, 8)
} // namespace cuda

12
onnxruntime/core/providers/cuda/math/gemm.cc
View file

@ -30,10 +30,20 @@ namespace cuda {
KernelDefBuilder() \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Gemm<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Gemm, \
kOnnxDomain, \
11, \
12, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder() \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Gemm<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
Gemm, \
kOnnxDomain, \
13, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder() \

11
onnxruntime/core/providers/cuda/math/matmul.cc
View file

@ -19,10 +19,19 @@ namespace cuda {
KernelDefBuilder() \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
MatMul<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
MatMul, \
kOnnxDomain, \
9, 12, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder() \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
MatMul<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
MatMul, \
kOnnxDomain, \
9, \
13, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder() \

20
onnxruntime/core/providers/cuda/math/softmax.cc
View file

@ -65,10 +65,18 @@ SPECIALIZED_SOFTMAX_HELPER_IMPL(MLFloat16)
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Softmax<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Softmax, \
kOnnxDomain, \
11, 12, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Softmax<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
Softmax, \
kOnnxDomain, \
11, \
13, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
@ -81,10 +89,18 @@ SPECIALIZED_SOFTMAX_HELPER_IMPL(MLFloat16)
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Softmax<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
LogSoftmax, \
kOnnxDomain, \
11, 12, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Softmax<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
LogSoftmax, \
kOnnxDomain, \
11, \
13, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \

62
onnxruntime/core/providers/cuda/math/unary_elementwise_ops.cc
View file

@ -13,6 +13,17 @@ Status UnaryElementwise::Prepare(OpKernelContext* context, UnaryElementwisePrepa
return Status::OK();
}
#define UNARY_ELEMENTWISE_REGISTER_VERSIONED_KERNEL(x, startver, endver, T) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
x, \
kOnnxDomain, \
startver, \
endver, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
x<T>);
#define UNARY_ELEMENTWISE_REGISTER_KERNEL(x, ver, T) \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
x, \
@ -46,6 +57,9 @@ Status UnaryElementwise::Prepare(OpKernelContext* context, UnaryElementwisePrepa
return Status::OK(); \
}
#define UNARY_OP_VERSIONED_TYPED(name, startver, endver, T) \
UNARY_ELEMENTWISE_REGISTER_VERSIONED_KERNEL(name, startver, endver, T)
#define UNARY_OP_TYPED(name, ver, T) \
UNARY_ELEMENTWISE_REGISTER_KERNEL(name, ver, T) \
UNARY_ELEMENTWISE_COMPUTE(name, T)
@ -68,6 +82,25 @@ Status UnaryElementwise::Prepare(OpKernelContext* context, UnaryElementwisePrepa
// D: double
// O: bool
#define UNARY_OP_VERSIONED_HFD(name, startver, endver) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, MLFloat16) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, float) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, double)
#define UNARY_OP_VERSIONED_CSILHFD(name, startver, endver) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, int8_t) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, int16_t) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, int32_t) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, int64_t) \
UNARY_OP_VERSIONED_HFD(name, startver, endver)
#define UNARY_OP_VERSIONED_BWUZCSILHFD(name, startver, endver) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, uint8_t) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, uint16_t) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, uint32_t) \
UNARY_OP_VERSIONED_TYPED(name, startver, endver, uint64_t) \
UNARY_OP_VERSIONED_CSILHFD(name, startver, endver)
#define UNARY_OP_HFD(name, ver) \
UNARY_OP_TYPED(name, ver, MLFloat16) \
UNARY_OP_TYPED(name, ver, float) \
@ -87,15 +120,26 @@ Status UnaryElementwise::Prepare(OpKernelContext* context, UnaryElementwisePrepa
UNARY_OP_TYPED(name, ver, uint64_t) \
UNARY_OP_CSILHFD(name, ver)
UNARY_OP_BWUZCSILHFD(Abs, 6)
UNARY_OP_CSILHFD(Neg, 6)
UNARY_OP_HFD(Floor, 6)
UNARY_OP_HFD(Ceil, 6)
UNARY_OP_HFD(Reciprocal, 6)
UNARY_OP_HFD(Sqrt, 6)
UNARY_OP_HFD(Log, 6)
UNARY_OP_HFD(Exp, 6)
UNARY_OP_HFD(Erf, 9)
UNARY_OP_VERSIONED_BWUZCSILHFD(Abs, 6, 12)
UNARY_OP_VERSIONED_CSILHFD(Neg, 6, 12)
UNARY_OP_VERSIONED_HFD(Floor, 6, 12)
UNARY_OP_VERSIONED_HFD(Ceil, 6, 12)
UNARY_OP_VERSIONED_HFD(Reciprocal, 6, 12)
UNARY_OP_VERSIONED_HFD(Sqrt, 6, 12)
UNARY_OP_VERSIONED_HFD(Log, 6, 12)
UNARY_OP_VERSIONED_HFD(Exp, 6, 12)
UNARY_OP_VERSIONED_HFD(Erf, 9, 12)
UNARY_OP_BWUZCSILHFD(Abs, 13)
UNARY_OP_CSILHFD(Neg, 13)
UNARY_OP_HFD(Floor, 13)
UNARY_OP_HFD(Ceil, 13)
UNARY_OP_HFD(Reciprocal, 13)
UNARY_OP_HFD(Sqrt, 13)
UNARY_OP_HFD(Log, 13)
UNARY_OP_HFD(Exp, 13)
UNARY_OP_HFD(Erf, 13)
UNARY_LOGICALOP_TYPED(Not, 1, bool)
UNARY_OP_HFD(Round, 11)

9
onnxruntime/core/providers/cuda/math/variadic_elementwise_ops.cc
View file

@ -233,13 +233,16 @@ const auto k_hfd_datatypes =
KernelDefBuilder().TypeConstraint("T", datatypes), \
impl_class)
REGISTER_KERNEL(Sum, SumOp, 8, k_hfd_datatypes)
REGISTER_KERNEL(Sum, SumOp, 13, k_hfd_datatypes)
REGISTER_VERSIONED_KERNEL(Sum, SumOp, 8, 12, k_hfd_datatypes)
REGISTER_VERSIONED_KERNEL(Sum, SumOp, 6, 7, k_hfd_datatypes)
REGISTER_KERNEL(Min, MinOp, 12, k_uzilhfd_datatypes)
REGISTER_KERNEL(Min, MinOp, 13, k_uzilhfd_datatypes)
REGISTER_VERSIONED_KERNEL(Min, MinOp, 12, 12, k_uzilhfd_datatypes)
REGISTER_VERSIONED_KERNEL(Min, MinOp, 6, 11, k_hfd_datatypes)
REGISTER_KERNEL(Max, MaxOp, 12, k_uzilhfd_datatypes)
REGISTER_KERNEL(Max, MaxOp, 13, k_uzilhfd_datatypes)
REGISTER_VERSIONED_KERNEL(Max, MaxOp, 12, 12, k_uzilhfd_datatypes)
REGISTER_VERSIONED_KERNEL(Max, MaxOp, 6, 11, k_hfd_datatypes)
#undef REGISTER_VERSIONED_KERNEL

15
onnxruntime/core/providers/cuda/nn/dropout.cc
View file

@ -6,10 +6,23 @@
namespace onnxruntime {
namespace cuda {
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Dropout,
kOnnxDomain,
12, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllIEEEFloatTensorTypes())
.TypeConstraint("T1", DataTypeImpl::AllIEEEFloatTensorTypes())
.TypeConstraint("T2", DataTypeImpl::GetTensorType<bool>())
.InputMemoryType<OrtMemTypeCPUInput>(1)
.InputMemoryType<OrtMemTypeCPUInput>(2),
Dropout<false>);
ONNX_OPERATOR_KERNEL_EX(
Dropout,
kOnnxDomain,
12,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllIEEEFloatTensorTypes())

50
onnxruntime/core/providers/cuda/reduction/reduction_ops.cc
View file

@ -24,16 +24,24 @@ namespace cuda {
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
name<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
name, \
kOnnxDomain, \
11, 12, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
name<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
name, \
kOnnxDomain, \
11, \
13, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
name<T>);
// Register with the latest version 12
// Register those with changes in OpSet12.
#define REGISTER_KERNEL_TYPED_12(name, T) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
name, \
@ -51,10 +59,37 @@ namespace cuda {
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
name<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
name, \
kOnnxDomain, \
12, 12, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
name<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
name, \
kOnnxDomain, \
12, \
13, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
name<T>);
// CUDA ArgMax/ArgMin doesn't have OpSet12 implementation (with select_last_index attr), keep it in OpSet11 for now.
#define REGISTER_KERNEL_TYPED_11(name, T) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
name, \
kOnnxDomain, \
1, 10, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
name<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
name, \
kOnnxDomain, \
11, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
@ -890,8 +925,13 @@ template Tensor ReduceCompute<double, CUDNN_REDUCE_TENSOR_NO_INDICES>(
REGISTER_KERNEL_TYPED(name, float) \
REGISTER_KERNEL_TYPED(name, double)
REGISTER_KERNEL_HFD(ArgMax)
REGISTER_KERNEL_HFD(ArgMin)
#define REGISTER_KERNEL_HFD_11(name) \
REGISTER_KERNEL_TYPED_11(name, MLFloat16) \
REGISTER_KERNEL_TYPED_11(name, float) \
REGISTER_KERNEL_TYPED_11(name, double)
REGISTER_KERNEL_HFD_11(ArgMax)
REGISTER_KERNEL_HFD_11(ArgMin)
REGISTER_KERNEL_HFD(ReduceL1)
REGISTER_KERNEL_HFD(ReduceL2)

12
onnxruntime/core/providers/cuda/tensor/cast_op.cc
View file

@ -34,10 +34,20 @@ const std::vector<MLDataType> castOpTypeConstraints{
.TypeConstraint("T1", DataTypeImpl::GetTensorType<T>()) \
.TypeConstraint("T2", castOpTypeConstraints), \
Cast<T>); \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Cast, \
kOnnxDomain, \
9, 12, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder() \
.TypeConstraint("T1", DataTypeImpl::GetTensorType<T>()) \
.TypeConstraint("T2", castOpTypeConstraints), \
Cast<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
Cast, \
kOnnxDomain, \
9, \
13, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder() \

10
onnxruntime/core/providers/cuda/tensor/concat.cc
View file

@ -15,9 +15,17 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(Concat,
Concat);
// opset 11 explicitly support negative axis
ONNX_OPERATOR_VERSIONED_KERNEL_EX(Concat,
kOnnxDomain,
11, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
Concat);
ONNX_OPERATOR_KERNEL_EX(Concat,
kOnnxDomain,
11,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),

12
onnxruntime/core/providers/cuda/tensor/expand.cc
View file

@ -108,10 +108,20 @@ Status Expand::ComputeInternal(OpKernelContext* ctx) const {
}
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Expand,
kOnnxDomain,
8, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
.InputMemoryType<OrtMemTypeCPUInput>(1),
Expand);
ONNX_OPERATOR_KERNEL_EX(
Expand,
kOnnxDomain,
8,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())

14
onnxruntime/core/providers/cuda/tensor/gather.cc
View file

@ -20,11 +20,23 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
DataTypeImpl::GetTensorType<int64_t>()}),
Gather);
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Gather,
kOnnxDomain,
11, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
.TypeConstraint("Tind", std::vector<MLDataType>{
DataTypeImpl::GetTensorType<int32_t>(),
DataTypeImpl::GetTensorType<int64_t>()}),
Gather);
// explicit negative axis support
ONNX_OPERATOR_KERNEL_EX(
Gather,
kOnnxDomain,
11,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())

14
onnxruntime/core/providers/cuda/tensor/gather_elements.cc
View file

@ -12,7 +12,19 @@ namespace cuda {
ONNX_OPERATOR_KERNEL_EX(
GatherElements,
kOnnxDomain,
11,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
.TypeConstraint("Tind", std::vector<MLDataType>{
DataTypeImpl::GetTensorType<int32_t>(),
DataTypeImpl::GetTensorType<int64_t>()}),
GatherElements);
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
GatherElements,
kOnnxDomain,
11, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())

16
onnxruntime/core/providers/cuda/tensor/gather_nd.cc
View file

@ -90,6 +90,19 @@ Status GatherNDBase::PrepareCompute(
return Status::OK();
}
#define REGISTER_KERNEL_VERSIONED_TYPED_GATHER_ND(TIndex, startver, endver) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
GatherND, \
kOnnxDomain, \
startver, \
endver, \
TIndex, \
kCudaExecutionProvider, \
KernelDefBuilder() \
.TypeConstraint("T", DataTypeImpl::AllIEEEFloatTensorTypes()) \
.TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIndex>()), \
GatherND<TIndex>);
#define REGISTER_KERNEL_TYPED_GATHER_ND(TIndex, ver) \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
GatherND, \
@ -112,7 +125,8 @@ Status GatherNDBase::PrepareCompute(
#ifdef ENABLE_TRAINING
REGISTER_KERNEL_TYPED_GATHER_ND(int64_t, 1)
#endif
REGISTER_KERNEL_TYPED_GATHER_ND(int64_t, 12)
REGISTER_KERNEL_TYPED_GATHER_ND(int64_t, 13)
REGISTER_KERNEL_VERSIONED_TYPED_GATHER_ND(int64_t, 12, 12)
template <typename T>
struct GatherNDComputeImpl {

10
onnxruntime/core/providers/cuda/tensor/nonzero_op.cc
View file

@ -11,10 +11,18 @@ namespace cuda {
// kernel builder functions
#define NONZERO_TYPED_KERNEL_WITH_TYPE_NAME(type, type_name) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
NonZero, \
kOnnxDomain, \
9, 12, \
type_name, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<type>()), \
NonZero<type>) \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
NonZero, \
kOnnxDomain, \
9, \
13, \
type_name, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<type>()), \

14
onnxruntime/core/providers/cuda/tensor/reshape.cc
View file

@ -9,7 +9,19 @@ namespace cuda {
ONNX_OPERATOR_KERNEL_EX(
Reshape,
kOnnxDomain,
5,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
.TypeConstraint("shape", DataTypeImpl::GetTensorType<int64_t>())
.Alias(0, 0)
.InputMemoryType<OrtMemTypeCPUInput>(1),
Reshape);
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Reshape,
kOnnxDomain,
5, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())

15
onnxruntime/core/providers/cuda/tensor/scatter_elements.cc
View file

@ -22,10 +22,23 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
DataTypeImpl::GetTensorType<int64_t>()}),
ScatterElements);
ONNX_OPERATOR_KERNEL_EX(
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
ScatterElements,
kOnnxDomain,
11,
12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
.TypeConstraint("Tind", std::vector<MLDataType>{
DataTypeImpl::GetTensorType<int32_t>(),
DataTypeImpl::GetTensorType<int64_t>()}),
ScatterElements);
ONNX_OPERATOR_KERNEL_EX(
ScatterElements,
kOnnxDomain,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())

13
onnxruntime/core/providers/cuda/tensor/shape_op.cc
View file

@ -7,10 +7,21 @@
namespace onnxruntime {
namespace cuda {
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Shape,
kOnnxDomain,
1, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.OutputMemoryType<OrtMemTypeCPUOutput>(0)
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
.TypeConstraint("T1", DataTypeImpl::GetTensorType<int64_t>()),
Shape);
ONNX_OPERATOR_KERNEL_EX(
Shape,
kOnnxDomain,
1,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.OutputMemoryType<OrtMemTypeCPUOutput>(0)

13
onnxruntime/core/providers/cuda/tensor/size.cc
View file

@ -7,10 +7,21 @@
namespace onnxruntime {
namespace cuda {
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Size,
kOnnxDomain,
1, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.OutputMemoryType<OrtMemTypeCPUInput>(0)
.TypeConstraint("T", DataTypeImpl::AllTensorTypes())
.TypeConstraint("T1", DataTypeImpl::GetTensorType<int64_t>()),
Size);
ONNX_OPERATOR_KERNEL_EX(
Size,
kOnnxDomain,
1,
13,
kCudaExecutionProvider,
KernelDefBuilder()
// properly force CPU/GPU synch inside the kernel

32
onnxruntime/core/providers/cuda/tensor/slice.cc
View file

@ -43,11 +43,11 @@ REGISTER_V10_TYPED_SLICE(int32_t)
REGISTER_V10_TYPED_SLICE(int64_t)
REGISTER_V10_TYPED_SLICE(float)
#define REGISTER_V11_TYPED_SLICE(TIND) \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
#define REGISTER_V12_TYPED_SLICE(TIND) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Slice, \
kOnnxDomain, \
11, \
11, 12, \
TIND, \
kCudaExecutionProvider, \
KernelDefBuilder() \
@ -59,9 +59,29 @@ REGISTER_V10_TYPED_SLICE(float)
.TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIND>()), \
Slice<true>);
REGISTER_V11_TYPED_SLICE(int32_t)
REGISTER_V11_TYPED_SLICE(int64_t)
REGISTER_V11_TYPED_SLICE(float)
REGISTER_V12_TYPED_SLICE(int32_t)
REGISTER_V12_TYPED_SLICE(int64_t)
REGISTER_V12_TYPED_SLICE(float)
#define REGISTER_V13_TYPED_SLICE(TIND) \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
Slice, \
kOnnxDomain, \
13, \
TIND, \
kCudaExecutionProvider, \
KernelDefBuilder() \
.InputMemoryType<OrtMemTypeCPUInput>(1) \
.InputMemoryType<OrtMemTypeCPUInput>(2) \
.InputMemoryType<OrtMemTypeCPUInput>(3) \
.InputMemoryType<OrtMemTypeCPUInput>(4) \
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()) \
.TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIND>()), \
Slice<true>);
REGISTER_V13_TYPED_SLICE(int32_t)
REGISTER_V13_TYPED_SLICE(int64_t)
REGISTER_V13_TYPED_SLICE(float)
static Status SliceImpCore(const void* input_data, void* output_data,
size_t element_size, size_t dimension_count,

9
onnxruntime/core/providers/cuda/tensor/split.cc
View file

@ -16,9 +16,16 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(Split,
Split);
// explicitly supports negative axis
ONNX_OPERATOR_VERSIONED_KERNEL_EX(Split,
kOnnxDomain,
11, 12,
kCudaExecutionProvider,
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
Split);
ONNX_OPERATOR_KERNEL_EX(Split,
kOnnxDomain,
11,
13,
kCudaExecutionProvider,
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
Split);

12
onnxruntime/core/providers/cuda/tensor/squeeze.cc
View file

@ -17,10 +17,20 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Squeeze);
// explicit support for negative axis.
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Squeeze,
kOnnxDomain,
11, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.Alias(0, 0)
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
Squeeze);
ONNX_OPERATOR_KERNEL_EX(
Squeeze,
kOnnxDomain,
11,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.Alias(0, 0)

10
onnxruntime/core/providers/cuda/tensor/transpose.cc
View file

@ -9,9 +9,17 @@
namespace onnxruntime {
namespace cuda {
ONNX_OPERATOR_VERSIONED_KERNEL_EX(Transpose,
kOnnxDomain,
1, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
Transpose);
ONNX_OPERATOR_KERNEL_EX(Transpose,
kOnnxDomain,
1,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),

13
onnxruntime/core/providers/cuda/tensor/unsqueeze.cc
View file

@ -17,10 +17,21 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Unsqueeze);
// explicitly support negative axis
ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Unsqueeze,
kOnnxDomain,
11, 12,
kCudaExecutionProvider,
KernelDefBuilder()
.Alias(0, 0)
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
Unsqueeze);
// support bfloat16
ONNX_OPERATOR_KERNEL_EX(
Unsqueeze,
kOnnxDomain,
11,
13,
kCudaExecutionProvider,
KernelDefBuilder()
.Alias(0, 0)

6
orttraining/orttraining/training_ops/cuda/cuda_training_kernels.cc
View file

@ -51,7 +51,8 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1
class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, float, int64_t, SparseSoftmaxCrossEntropy);
// class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, float, int32_t, SparseSoftmaxCrossEntropyGrad);
class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, float, int64_t, SparseSoftmaxCrossEntropyGrad);
class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, float, int64_t, SoftmaxCrossEntropyLoss);
class ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 12, float, int64_t, SoftmaxCrossEntropyLoss);
class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 13, float, int64_t, SoftmaxCrossEntropyLoss);
class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, float, int64_t, SoftmaxCrossEntropyLossGrad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, float, SoftmaxGrad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, double, SoftmaxGrad);
@ -193,7 +194,8 @@ Status RegisterCudaTrainingKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, float, LogSoftmaxGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, double, LogSoftmaxGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, MLFloat16, LogSoftmaxGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, float, int64_t, SoftmaxCrossEntropyLoss)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 12, float, int64_t, SoftmaxCrossEntropyLoss)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 13, float, int64_t, SoftmaxCrossEntropyLoss)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, float, int64_t, SoftmaxCrossEntropyLossGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, float, BatchNormalizationGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMSDomain, 1, double, BatchNormalizationGrad)>,

18
orttraining/orttraining/training_ops/cuda/loss/softmax_cross_entropy_loss_impl.cc
View file

@ -11,6 +11,18 @@
namespace onnxruntime {
namespace cuda {
#define REGISTER_KERNEL_VERSIONED_TYPED_TWO_TYPES(Class, T, Tin, domain, startver, endver) \
ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_EX( \
Class, \
domain, \
startver, endver, \
T, Tin, \
kCudaExecutionProvider, \
KernelDefBuilder() \
.TypeConstraint("T", DataTypeImpl::GetTensorType<T>()) \
.TypeConstraint("Tin", DataTypeImpl::GetTensorType<Tin>()), \
Class<T, Tin>);
#define REGISTER_KERNEL_TYPED_TWO_TYPES(Class, T, Tin, domain, version) \
ONNX_OPERATOR_TWO_TYPED_KERNEL_EX( \
Class, \
@ -243,11 +255,15 @@ Status SoftmaxCrossEntropyLossGrad<T, Tin>::ComputeInternal(OpKernelContext* ctx
return Status::OK();
}
#define SPECIALIZED_VERSIONED_COMPUTE_SPARSE(Class, T, Tin, domain, startver, endvar) \
REGISTER_KERNEL_VERSIONED_TYPED_TWO_TYPES(Class, T, Tin, domain, startver, endvar)
#define SPECIALIZED_COMPUTE_SPARSE(Class, T, Tin, domain, version) \
REGISTER_KERNEL_TYPED_TWO_TYPES(Class, T, Tin, domain, version) \
template Status Class<T, Tin>::ComputeInternal(OpKernelContext* ctx) const;
SPECIALIZED_COMPUTE_SPARSE(SoftmaxCrossEntropyLoss, float, int64_t, kOnnxDomain, 12)
SPECIALIZED_VERSIONED_COMPUTE_SPARSE(SoftmaxCrossEntropyLoss, float, int64_t, kOnnxDomain, 12, 12)
SPECIALIZED_COMPUTE_SPARSE(SoftmaxCrossEntropyLoss, float, int64_t, kOnnxDomain, 13)
SPECIALIZED_COMPUTE_SPARSE(SoftmaxCrossEntropyLossGrad, float, int64_t, kMSDomain, 1)
} // namespace cuda