saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-03 03:58:54 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Update cast op to support string <-> numeric (#379)

Browse source 
* Update cast kernel to support to/from string

* Update namespace

* Add support for literal numeric case

* Update to support -INF test

* Update kernel registration for cast

* Update ONNX to 1.4.1

* Update registy api

* Resolve some comments

* Update cast kernel implementation

* Resolve comments

* Fixed test data in onnx

* Update cast kernel implementation

* Resolve PR comments

* Update cast_op.cc

* Update onnx commits info

* Update comments
This commit is contained in:
Raymond Yang 2019-02-12 10:10:56 -08:00 committed by GitHub
parent f72474c24b
commit ec8ac04f30
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
9 changed files with 323 additions and 150 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
cgmanifest.json
View file

@ -49,7 +49,7 @@
"component":{
"type":"git",
"git":{
"commitHash":"8a1319733a5518bd0001842db27e2df53a306eff",
"commitHash":"2896c77cfc628f18b6ca6b28e3a380807fa00f53",
"repositoryUrl":"https://github.com/onnx/onnx.git"
}
}

2
cmake/external/onnx vendored

@ -1 +1 @@
Subproject commit dbf3581835e3a05716e10587511d7ab3b2cdc386
Subproject commit 2896c77cfc628f18b6ca6b28e3a380807fa00f53

1
onnxruntime/contrib_ops/cpu/quantize_linear.cc
View file

@ -3,7 +3,6 @@
#include "contrib_ops/cpu/quantize_linear.h"
#include "core/providers/cpu/math/element_wise_ops.h"
#include "core/providers/cpu/tensor/cast_op.h"
#include "core/providers/common.h"
namespace onnxruntime {

50
onnxruntime/core/providers/cpu/cpu_execution_provider.cc
View file

@ -134,18 +134,18 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain,
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, GRU);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, LSTM);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, RNN);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint8_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint16_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint32_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint64_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int8_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int16_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int32_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int64_t, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, bool, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, float, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, double, Cast);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, MLFloat16, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint8_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint16_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint32_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint64_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int8_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int16_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int32_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int64_t, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, bool, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, float, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, double, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, MLFloat16, Cast);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 4, Concat);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Crop);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Gather);
@ -233,6 +233,7 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, Con
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, MeanVarianceNormalization);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, int32_t, Greater);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, int32_t, Less);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, string, Cast);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, EyeLike);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, float, IsNaN);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, MLFloat16, IsNaN);
@ -384,18 +385,18 @@ void RegisterOnnxOperatorKernels(KernelRegistry& kernel_registry) {
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, GRU)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, LSTM)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, RNN)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint8_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint16_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint32_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, uint64_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int8_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int16_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int32_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, int64_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, bool, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, float, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, double, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, MLFloat16, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint8_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint16_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint32_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, uint64_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int8_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int16_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int32_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, int64_t, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, bool, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, float, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, double, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 6, 9, MLFloat16, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 4, Concat)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Crop)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, Gather)>());
@ -485,6 +486,7 @@ void RegisterOnnxOperatorKernels(KernelRegistry& kernel_registry) {
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, int32_t, Greater)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, int32_t, Less)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, EyeLike)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, string, Cast)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, float, IsNaN)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, MLFloat16, IsNaN)>());
kernel_registry.Register(BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, Sign)>());

278
onnxruntime/core/providers/cpu/tensor/cast_op.cc
View file

@ -1,13 +1,202 @@
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
#include "core/providers/cpu/tensor/cast_op.h"
#include <sstream>
#include "core/common/common.h"
#include "core/framework/op_kernel.h"
#include "core/util/math.h"
#include "core/util/math_cpuonly.h"
#include "Eigen/src/Core/arch/CUDA/Half.h"
#include "core/common/common.h"
#if defined(USE_MLAS) && defined(_M_AMD64)
#include "core/mlas/inc/mlas.h"
#endif
using namespace ONNX_NAMESPACE;
namespace onnxruntime {
template <typename SrcType,
typename DstType>
inline void CastData(const Tensor* in, Tensor* out, const TensorShape& shape) {
auto shape_size = shape.Size();
auto in_vector = ConstEigenVectorMap<SrcType>(in->template Data<SrcType>(), shape_size);
auto output_vector = EigenVectorMap<DstType>(out->template MutableData<DstType>(), shape_size);
output_vector = in_vector.template cast<DstType>();
}
template <>
inline void CastData<float, MLFloat16>(const Tensor* in, Tensor* out, const TensorShape& shape) {
auto out_data = out->template MutableData<MLFloat16>();
auto shape_size = shape.Size();
auto in_vector = ConstEigenVectorMap<float>(in->template Data<float>(), shape_size);
auto output_vector = EigenVectorMap<Eigen::half>(static_cast<Eigen::half*>(static_cast<void*>(out_data)), shape_size);
output_vector = in_vector.template cast<Eigen::half>();
}
template <>
inline void CastData<MLFloat16, float>(const Tensor* in, Tensor* out, const TensorShape& shape) {
auto out_data = out->template MutableData<float>();
auto in_data = in->template Data<MLFloat16>();
auto shape_size = shape.Size();
#if defined(USE_MLAS) && defined(_M_AMD64)
MlasConvertHalfToFloatBuffer(&in_data[0].val, out_data, shape_size);
#else
auto in_vector = ConstEigenVectorMap<Eigen::half>(static_cast<const Eigen::half*>(static_cast<const void*>(in_data)), shape_size);
auto output_vector = EigenVectorMap<float>(out_data, shape_size);
output_vector = in_vector.template cast<float>();
#endif
}
template <typename SrcType,
typename DstType>
inline void CastFloat16Data(const Tensor* in, Tensor* out, const TensorShape& shape, const AllocatorPtr& allocator) {
ORT_ENFORCE(allocator != nullptr);
const int64_t len = shape.Size();
ORT_ENFORCE(len > 0);
void* buffer = allocator->AllocArray(sizeof(float), len);
ORT_ENFORCE(buffer);
Tensor tmp_tensor(DataTypeImpl::GetType<float>(), shape, buffer, allocator->Info(), nullptr);
if (std::is_same<SrcType, MLFloat16>::value) {
CastData<MLFloat16, float>(in, &tmp_tensor, shape); // first cast to float
CastData<float, DstType>(&tmp_tensor, out, shape); // then cast to the destination type.
} else if (std::is_same<DstType, MLFloat16>::value) {
CastData<SrcType, float>(in, &tmp_tensor, shape);
CastData<float, MLFloat16>(&tmp_tensor, out, shape);
}
allocator->Free(buffer);
}
template <typename SrcType>
inline void CastToStringData(const Tensor* in, Tensor* out, const TensorShape& shape) {
const int64_t len = shape.Size();
ORT_ENFORCE(len > 0);
for (int i = 0; i < len; ++i) {
if (std::is_floating_point<SrcType>::value && std::isnan(in->Data<SrcType>()[i])) {
out->MutableData<std::string>()[i] = "NaN";
} else if (std::is_floating_point<SrcType>::value && std::isinf(in->Data<SrcType>()[i])) {
if (in->Data<SrcType>()[i] < std::numeric_limits<SrcType>::lowest()) {
out->MutableData<std::string>()[i] = "-INF";
} else {
out->MutableData<std::string>()[i] = "INF";
}
} else {
std::ostringstream convert;
convert << in->Data<SrcType>()[i];
out->MutableData<std::string>()[i] = convert.str();
}
}
}
template <typename DstType>
inline void CastFromStringData(const Tensor* in, Tensor* out, const TensorShape& shape) {
if (std::is_same<DstType, std::string>::value) return;
const int64_t len = shape.Size();
ORT_ENFORCE(len > 0);
if (std::is_same<DstType, float>::value) {
float* mutable_data = out->MutableData<float>();
for (int i = 0; i < len; ++i) {
mutable_data[i] = std::stof(in->Data<std::string>()[i]);
}
} else if (std::is_same<DstType, double>::value) {
double* mutable_data = out->MutableData<double>();
for (int i = 0; i < len; ++i) {
mutable_data[i] = std::stod(in->Data<std::string>()[i]);
}
} else if (std::is_same<DstType, int8_t>::value) {
int8_t* mutable_data = out->MutableData<int8_t>();
for (int i = 0; i < len; ++i) {
int temp_i = std::stoi(in->Data<std::string>()[i]);
mutable_data[i] = static_cast<int8_t>(temp_i);
}
} else if (std::is_same<DstType, uint8_t>::value) {
uint8_t* mutable_data = out->MutableData<uint8_t>();
for (int i = 0; i < len; ++i) {
unsigned long temp_ui = std::stoul(in->Data<std::string>()[i]);
mutable_data[i] = static_cast<uint8_t>(temp_ui);
}
} else if (std::is_same<DstType, int16_t>::value) {
int16_t* mutable_data = out->MutableData<int16_t>();
for (int i = 0; i < len; ++i) {
int temp_i = std::stoi(in->Data<std::string>()[i]);
mutable_data[i] = static_cast<int16_t>(temp_i);
}
} else if (std::is_same<DstType, uint16_t>::value) {
uint16_t* mutable_data = out->MutableData<uint16_t>();
for (int i = 0; i < len; ++i) {
unsigned long temp_ui = std::stoul(in->Data<std::string>()[i]);
mutable_data[i] = static_cast<uint16_t>(temp_ui);
}
} else if (std::is_same<DstType, int32_t>::value) {
int32_t* mutable_data = out->MutableData<int32_t>();
for (int i = 0; i < len; ++i) {
mutable_data[i] = std::stol(in->Data<std::string>()[i]);
}
} else if (std::is_same<DstType, uint32_t>::value) {
uint32_t* mutable_data = out->MutableData<uint32_t>();
for (int i = 0; i < len; ++i) {
mutable_data[i] = std::stoul(in->Data<std::string>()[i]);
}
} else if (std::is_same<DstType, int64_t>::value) {
int64_t* mutable_data = out->MutableData<int64_t>();
for (int i = 0; i < len; ++i) {
mutable_data[i] = std::stoll(in->Data<std::string>()[i]);
}
} else if (std::is_same<DstType, uint64_t>::value) {
uint64_t* mutable_data = out->MutableData<uint64_t>();
for (int i = 0; i < len; ++i) {
mutable_data[i] = std::stoull(in->Data<std::string>()[i]);
}
} else {
ORT_THROW("Unsupported type in cast op: from String to ", typeid(DstType).name());
}
} // namespace onnxruntime
template <typename T>
class Cast final : public OpKernel {
public:
Cast(const OpKernelInfo& info) : OpKernel(info) {
int64_t to;
Status status = info.GetAttr("to", &to);
ORT_ENFORCE(status.IsOK(), "Attribute to is not set.");
to_ = gsl::narrow_cast<ONNX_NAMESPACE::TensorProto_DataType>(to);
}
Status Compute(OpKernelContext* context) const override;
private:
template <typename SrcType,
typename DstType>
void CastData(const Tensor* in, Tensor* out, const TensorShape& shape) const {
::onnxruntime::CastData<SrcType, DstType>(in, out, shape);
}
template <typename SrcType,
typename DstType>
Status CastFloat16Data(const Tensor* in, Tensor* out, const TensorShape& shape, OpKernelContext* context) const {
AllocatorPtr allocator;
ORT_RETURN_IF_ERROR(context->GetTempSpaceAllocator(&allocator));
::onnxruntime::CastFloat16Data<SrcType, DstType>(in, out, shape, allocator);
return Status::OK();
}
template <typename SrcType>
Status CastToStringData(const Tensor* in, Tensor* out, const TensorShape& shape) const {
::onnxruntime::CastToStringData<SrcType>(in, out, shape);
return Status::OK();
}
template <typename DstType>
Status CastFromStringData(const Tensor* in, Tensor* out, const TensorShape& shape) const {
::onnxruntime::CastFromStringData<DstType>(in, out, shape);
return Status::OK();
}
ONNX_NAMESPACE::TensorProto_DataType to_;
};
const std::vector<MLDataType> castOpTypeConstraints{
DataTypeImpl::GetTensorType<bool>(),
DataTypeImpl::GetTensorType<float>(),
@ -20,12 +209,14 @@ const std::vector<MLDataType> castOpTypeConstraints{
DataTypeImpl::GetTensorType<int16_t>(),
DataTypeImpl::GetTensorType<int32_t>(),
DataTypeImpl::GetTensorType<int64_t>(),
DataTypeImpl::GetTensorType<MLFloat16>()};
DataTypeImpl::GetTensorType<MLFloat16>(),
DataTypeImpl::GetTensorType<std::string>()};
#define ADD_FROM_CAST_OP(in_type) \
ONNX_CPU_OPERATOR_TYPED_KERNEL( \
ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL( \
Cast, \
6, \
9, \
in_type, \
KernelDefBuilder().TypeConstraint("T1", DataTypeImpl::GetTensorType<in_type>()).TypeConstraint("T2", castOpTypeConstraints), \
Cast<in_type>); \
@ -33,7 +224,6 @@ const std::vector<MLDataType> castOpTypeConstraints{
template <> \
Status Cast<in_type>::Compute(OpKernelContext* context) const { \
const Tensor* X = context->Input<Tensor>(0); \
if (X == nullptr) return Status(common::ONNXRUNTIME, common::FAIL, "input count mismatch"); \
const TensorShape& shape = X->Shape(); \
Tensor* Y = context->Output(0, TensorShape(shape)); \
\
@ -80,11 +270,12 @@ const std::vector<MLDataType> castOpTypeConstraints{
} \
break; \
case TensorProto_DataType_STRING: \
ORT_THROW("Casting to and from strings is not supported yet."); /*break;*/ \
CastToStringData<in_type>(X, Y, shape); \
break; \
case TensorProto_DataType_UNDEFINED: \
ORT_THROW("Cast op must have 'to' argument of type DataType"); /*break;*/ \
ORT_THROW("Cast op must have 'to' argument of type DataType"); /*break;*/ \
default: \
ORT_THROW("Unexpected 'to' argument value: ", to_); \
ORT_THROW("Unexpected 'to' argument value: ", to_); \
} \
return Status::OK(); \
}
@ -101,9 +292,10 @@ ADD_FROM_CAST_OP(bool);
ADD_FROM_CAST_OP(float);
ADD_FROM_CAST_OP(double);
ONNX_CPU_OPERATOR_TYPED_KERNEL(
ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
Cast,
6,
9,
MLFloat16,
KernelDefBuilder().TypeConstraint("T1", DataTypeImpl::GetTensorType<MLFloat16>()).TypeConstraint("T2", castOpTypeConstraints),
Cast<MLFloat16>);
@ -111,7 +303,6 @@ ONNX_CPU_OPERATOR_TYPED_KERNEL(
template <>
Status Cast<MLFloat16>::Compute(OpKernelContext* context) const {
const Tensor* X = context->Input<Tensor>(0);
if (X == nullptr) return Status(common::ONNXRUNTIME, common::FAIL, "input count mismatch");
const TensorShape& shape = X->Shape();
Tensor* Y = context->Output(0, TensorShape(shape));
Status st;
@ -144,14 +335,14 @@ Status Cast<MLFloat16>::Compute(OpKernelContext* context) const {
CastData<MLFloat16, float>(X, Y, shape);
break;
case TensorProto_DataType_FLOAT16: {
auto X_type = X->DataType();
const void* source = X->DataRaw(X_type);
void* target = Y->MutableDataRaw(X_type);
// if source and target pointers are not equal, we need to copy the data.
if (target != source) {
memcpy(target, source, shape.Size() * X_type->Size());
}
st = Status::OK();
auto X_type = X->DataType();
const void* source = X->DataRaw(X_type);
void* target = Y->MutableDataRaw(X_type);
// if source and target pointers are not equal, we need to copy the data.
if (target != source) {
memcpy(target, source, shape.Size() * X_type->Size());
}
st = Status::OK();
break;
}
case TensorProto_DataType_DOUBLE:
@ -170,4 +361,57 @@ Status Cast<MLFloat16>::Compute(OpKernelContext* context) const {
return st;
}
ONNX_CPU_OPERATOR_TYPED_KERNEL(
Cast,
9,
string,
KernelDefBuilder().TypeConstraint("T1", DataTypeImpl::GetTensorType<std::string>()).TypeConstraint("T2", castOpTypeConstraints),
Cast<std::string>);
template <>
Status Cast<std::string>::Compute(OpKernelContext* context) const {
const Tensor* X = context->Input<Tensor>(0);
if (X == nullptr) return Status(common::ONNXRUNTIME, common::FAIL,
"Input is missing. The operator Cast expects one and only one input");
const TensorShape& shape = X->Shape();
Tensor* Y = context->Output(0, TensorShape(shape));
Status st;
switch (to_) {
case TensorProto_DataType_INT16:
st = CastFromStringData<int16_t>(X, Y, shape);
break;
case TensorProto_DataType_INT32:
st = CastFromStringData<int32_t>(X, Y, shape);
break;
case TensorProto_DataType_INT64:
st = CastFromStringData<int64_t>(X, Y, shape);
break;
case TensorProto_DataType_UINT8:
st = CastFromStringData<uint8_t>(X, Y, shape);
break;
case TensorProto_DataType_UINT16:
st = CastFromStringData<uint16_t>(X, Y, shape);
break;
case TensorProto_DataType_UINT32:
st = CastFromStringData<uint32_t>(X, Y, shape);
break;
case TensorProto_DataType_UINT64:
st = CastFromStringData<uint64_t>(X, Y, shape);
break;
case TensorProto_DataType_FLOAT:
st = CastFromStringData<float>(X, Y, shape);
break;
case TensorProto_DataType_DOUBLE:
st = CastFromStringData<double>(X, Y, shape);
break;
case TensorProto_DataType_INT8:
st = CastFromStringData<int8_t>(X, Y, shape);
break;
case TensorProto_DataType_UNDEFINED:
ORT_THROW("Cast op must have 'to' argument of type DataType");
default:
ORT_THROW("Unexpected 'to' argument value: ", to_);
}
return st;
}
} //namespace onnxruntime

100
onnxruntime/core/providers/cpu/tensor/cast_op.h
View file

@ -1,100 +0,0 @@
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
#pragma once
#include "core/common/common.h"
#include "core/framework/op_kernel.h"
#include "core/util/math.h"
#include "core/util/math_cpuonly.h"
#include "Eigen/src/Core/arch/CUDA/Half.h"
#if defined(USE_MLAS) && defined(_M_AMD64)
#include "core/mlas/inc/mlas.h"
#endif
namespace onnxruntime {
template <typename SrcType,
typename DstType>
inline void CastData(const Tensor* in, Tensor* out, const TensorShape& shape) {
auto shape_size = shape.Size();
auto in_vector = ConstEigenVectorMap<SrcType>(in->template Data<SrcType>(), shape_size);
auto output_vector = EigenVectorMap<DstType>(out->template MutableData<DstType>(), shape_size);
output_vector = in_vector.template cast<DstType>();
}
template <>
inline void CastData<float, MLFloat16>(const Tensor* in, Tensor* out, const TensorShape& shape) {
auto out_data = out->template MutableData<MLFloat16>();
auto shape_size = shape.Size();
auto in_vector = ConstEigenVectorMap<float>(in->template Data<float>(), shape_size);
auto output_vector = EigenVectorMap<Eigen::half>(static_cast<Eigen::half*>(static_cast<void*>(out_data)), shape_size);
output_vector = in_vector.template cast<Eigen::half>();
}
template <>
inline void CastData<MLFloat16, float>(const Tensor* in, Tensor* out, const TensorShape& shape) {
auto out_data = out->template MutableData<float>();
auto in_data = in->template Data<MLFloat16>();
auto shape_size = shape.Size();
#if defined(USE_MLAS) && defined(_M_AMD64)
MlasConvertHalfToFloatBuffer(&in_data[0].val, out_data, shape_size);
#else
auto in_vector = ConstEigenVectorMap<Eigen::half>(static_cast<const Eigen::half*>(static_cast<const void*>(in_data)), shape_size);
auto output_vector = EigenVectorMap<float>(out_data, shape_size);
output_vector = in_vector.template cast<float>();
#endif
}
template <typename SrcType,
typename DstType>
inline void CastFloat16Data(const Tensor* in, Tensor* out, const TensorShape& shape, const AllocatorPtr& allocator) {
ORT_ENFORCE(allocator != nullptr);
const int64_t len = shape.Size();
ORT_ENFORCE(len > 0);
void* buffer = allocator->AllocArray(sizeof(float), len);
ORT_ENFORCE(buffer);
Tensor tmp_tensor(DataTypeImpl::GetType<float>(), shape, buffer, allocator->Info(), nullptr);
if (std::is_same<SrcType, MLFloat16>::value) {
CastData<MLFloat16, float>(in, &tmp_tensor, shape); // first cast to float
CastData<float, DstType>(&tmp_tensor, out, shape); // then cast to the destination type.
} else if (std::is_same<DstType, MLFloat16>::value) {
CastData<SrcType, float>(in, &tmp_tensor, shape);
CastData<float, MLFloat16>(&tmp_tensor, out, shape);
}
allocator->Free(buffer);
}
template <typename T>
class Cast final : public OpKernel {
public:
Cast(const OpKernelInfo& info) : OpKernel(info) {
int64_t to;
Status status = info.GetAttr("to", &to);
ORT_ENFORCE(status.IsOK(), "Attribute to is not set.");
to_ = gsl::narrow_cast<ONNX_NAMESPACE::TensorProto_DataType>(to);
}
Status Compute(OpKernelContext* context) const override;
private:
template <typename SrcType,
typename DstType>
void CastData(const Tensor* in, Tensor* out, const TensorShape& shape) const {
::onnxruntime::CastData<SrcType, DstType>(in, out, shape);
}
template <typename SrcType,
typename DstType>
Status CastFloat16Data(const Tensor* in, Tensor* out, const TensorShape& shape, OpKernelContext* context) const {
AllocatorPtr allocator;
ORT_RETURN_IF_ERROR(context->GetTempSpaceAllocator(&allocator));
::onnxruntime::CastFloat16Data<SrcType, DstType>(in, out, shape, allocator);
return Status::OK();
}
ONNX_NAMESPACE::TensorProto_DataType to_;
};
} //namespace onnxruntime

29
onnxruntime/test/providers/cpu/tensor/tensor_op_test.cc
View file

@ -3,7 +3,6 @@
#include "gtest/gtest.h"
#include "test/providers/provider_test_utils.h"
#include "core/providers/cpu/tensor/cast_op.h"
#include "core/providers/cpu/tensor/crop.h"
#include "core/util/math.h"
@ -80,7 +79,7 @@ void TestCastOp(const std::initializer_list<SrcType>& input,
int64_t toType,
ExpectResult expect_result = ExpectResult::kExpectSuccess,
const std::string& expected_failure_string = "") {
OpTester test("Cast");
OpTester test("Cast", 9);
test.AddAttribute("to", toType);
test.AddInput<SrcType>("input", dimensions, input);
test.AddOutput<DstType>("output", dimensions, output);
@ -277,6 +276,32 @@ TEST(TensorOpTest, CastFromFloat16) {
TestCastOp(input, int64_t_data, shape, TensorProto::INT64);
}
TEST(TensorOpTest, CastFromString) {
const std::vector<int64_t> shape{2, 2, 2};
std::initializer_list<std::string> string_data = {"-inf", "+INF", "2.0f", "3.0f", "4.0f", "5.0f", "NaN", "nan"};
const std::initializer_list<float> float_output = {-(std::numeric_limits<float>::infinity()), std::numeric_limits<float>::infinity(), 2.0f, 3.0f, 4.0f, 5.0f, NAN, NAN};
TestCastOp(string_data, float_output, shape, TensorProto::FLOAT);
std::initializer_list<std::string> int_16_string_data = {"0", "1", "2", "3", "4", "5", "-32768", "32767"};
const std::initializer_list<int16_t> int_16_output = {0, 1, 2, 3, 4, 5, SHRT_MIN, SHRT_MAX};
TestCastOp(int_16_string_data, int_16_output, shape, TensorProto::INT16);
std::initializer_list<std::string> int_64_string_data = {"0", "1", "2", "3", "4", "5", "-9223372036854775808", "9223372036854775807"};
const std::initializer_list<int64_t> int_64_output = {0, 1, 2, 3, 4, 5, LLONG_MIN, LLONG_MAX};
TestCastOp(int_64_string_data, int_64_output, shape, TensorProto::INT64);
}
TEST(TensorOpTest, CastToString) {
const std::vector<int64_t> shape{2, 2, 2};
const std::initializer_list<float> float_input = {NAN, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()};
std::initializer_list<std::string> string_output = {"NaN", "1", "2", "3", "4", "5", "-INF", "INF"};
TestCastOp(float_input, string_output, shape, TensorProto::STRING);
std::initializer_list<std::string> int_string_data = {"0", "1", "2", "3", "4", "5", "6", "7"};
const std::initializer_list<int16_t> int_16_input = {0, 1, 2, 3, 4, 5, 6, 7};
TestCastOp(int_16_input, int_string_data, shape, TensorProto::STRING);
}
TEST(TensorOpTest, CropBorderOnly) {
const int N = 2, C = 1, H = 3, W = 4;
std::vector<float> X = {1.0f, 2.0f, 3.0f, 4.0f,

6
onnxruntime/test/providers/provider_test_utils.cc
View file

@ -50,9 +50,11 @@ void Check<float>(const OpTester::Data& expected_data, const Tensor& output_tens
#endif
for (int i = 0; i < size; ++i) {
if (std::isinf(expected[i])) // Test infinity for equality
if (std::isinf(expected[i])){ // Test infinity for equality
EXPECT_EQ(expected[i], output[i]);
else {
} else if (std::isnan(expected[i])) {
EXPECT_TRUE(std::isnan(output[i])) << "Expected output " << i << " to be NaN";
} else {
if (!has_abs_err && !has_rel_err) {
// the default for existing tests
EXPECT_NEAR(expected[i], output[i], threshold) << "provider_type: " << provider_type;

5
tools/ci_build/github/linux/docker/scripts/install_deps.sh
View file

@ -37,8 +37,9 @@ else
#Install ONNX
#5af210ca8a1c73aa6bae8754c9346ec54d0a756e is v1.2.3
#bae6333e149a59a3faa9c4d9c44974373dcf5256 is v1.3.0
#dbf3581835e3a05716e10587511d7ab3b2cdc386 is v1.3.0 latest
for onnx_version in "5af210ca8a1c73aa6bae8754c9346ec54d0a756e" "bae6333e149a59a3faa9c4d9c44974373dcf5256" "dbf3581835e3a05716e10587511d7ab3b2cdc386"; do
#9e55ace55aad1ada27516038dfbdc66a8a0763db is v1.4.1
#2896c77cfc628f18b6ca6b28e3a380807fa00f53 is v1.4.1 latest
for onnx_version in "5af210ca8a1c73aa6bae8754c9346ec54d0a756e" "bae6333e149a59a3faa9c4d9c44974373dcf5256" "9e55ace55aad1ada27516038dfbdc66a8a0763db" "2896c77cfc628f18b6ca6b28e3a380807fa00f53"; do
if [ -z ${lastest_onnx_version+x} ]; then
echo "first pass";
else