onnxruntime-extensions/include/op_def_struct.h

307 строки
11 KiB
C++

// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
//.A very thin wrapper of ONNXRuntime Custom Operator Callback ABI, which
// is only used in the custom-op kernels. For the general ORT C++ invocation, like end-to-end
// testing, the ONNXRuntime public C++ APIs should be used since there is no binary compatible requirement.
#pragma once
#include <cstdint>
#include <cstddef>
#include <array>
#include <memory>
#include <string>
#include <vector>
#include <utility>
#include <type_traits>
#include <optional>
#include <functional>
#include "exceptions.h"
#include "onnxruntime_extensions.h"
#include "custom_op/custom_op_lite.h"
#define MIN_ORT_VERSION_SUPPORTED 11
namespace Ort {
namespace Custom {
template <typename T>
inline OrtStatusPtr ToApiStatus(const T& status) {
return (OrtStatus*)status;
}
template <>
inline OrtStatusPtr ToApiStatus(const OrtStatusPtr& status) {
return status;
}
template <typename RType, typename... Args>
struct FunctionKernel {
using ComputeFn = std::function<RType(Args...)>;
RType Compute(Args... args) const {
return compute_fn_(args...);
}
ComputeFn compute_fn_;
};
// primary template handles types that have no nested ::type member:
template <class, class = void>
struct IsFunctionKernel {
typedef std::false_type type;
};
// specialization recognizes types that do have a nested ::type member:
template <class T>
struct IsFunctionKernel<T, std::void_t<typename T::ComputeFn>> {
typedef std::true_type type;
};
// Helper type
template <typename T>
struct ComputeArgsList;
// Specialization for member function
template <typename RType, typename C, typename... Args>
struct ComputeArgsList<RType (C::*)(Args...) const> {
using FunctionType = RType (*)(Args...);
using MemberFunctionType = RType (C::*)(Args...) const;
using ResultType = RType;
};
template<typename, typename T>
struct HasOnModelAttach {
static_assert(
std::integral_constant<T, false>::value,
"Second template parameter needs to be of function type.");
};
// specialization that does the checking
template<typename C, typename Ret, typename... Args>
struct HasOnModelAttach<C, Ret(Args...)> {
private:
template<typename T>
static constexpr auto check(T*)
-> typename
std::is_same<
decltype( std::declval<T>().OnModelAttach( std::declval<Args>()... ) ),
Ret
>::type; // attempt to call it and see if the return type is correct
template<typename>
static constexpr std::false_type check(...);
typedef decltype(check<C>(0)) type;
public:
static constexpr bool value = type::value;
};
template <typename T, typename = void>
struct CustomOp_defined_getInputMemoryType : std::false_type {};
template <typename T>
struct CustomOp_defined_getInputMemoryType<T, std::void_t<decltype(&T::GetInputMemoryType)>> : std::true_type {};
template <typename T, typename = void>
struct CustomOp_defined_getMayInplace : std::false_type {};
template <typename T>
struct CustomOp_defined_getMayInplace<T, std::void_t<decltype(&T::GetMayInplace)>> : std::true_type {};
template <typename T, typename = void>
struct CustomOp_defined_releaseMayInplace : std::false_type {};
template <typename T>
struct CustomOp_defined_releaseMayInplace<T, std::void_t<decltype(&T::ReleaseMayInplace)>> : std::true_type {};
template <typename CustomOpKernel>
struct OrtLiteCustomStructV2 : public OrtLiteCustomOp {
using ComputeFunction = decltype(&CustomOpKernel::Compute);
using RegularComputeType = typename ComputeArgsList<ComputeFunction>::FunctionType;
using RType = typename ComputeArgsList<ComputeFunction>::ResultType;
template <typename... Args>
using MemberComputeType = RType (CustomOpKernel::*)(Args...) const;
struct KernelEx : public CustomOpKernel {
struct {
std::string ep_{};
std::unique_ptr<OrtW::CustomOpApi> api_;
} extra_;
};
template <typename T>
static OrtStatusPtr InitKernel(KernelEx& kernel,
const OrtApi& api, const OrtKernelInfo& info, RegularComputeType fn, T t) {
if constexpr (HasOnModelAttach<KernelEx, OrtStatusPtr(const OrtApi&, const OrtKernelInfo&)>::value){
auto status = kernel.OnModelAttach(api, info);
return ToApiStatus(status);
}
else {
auto status = kernel.OnModelAttach(OrtAttributeReader(api, info));
return ToApiStatus(status);
}
}
static OrtStatusPtr InitKernel(
KernelEx& kernel,
const OrtApi& api, const OrtKernelInfo& info, RegularComputeType fn, std::true_type) {
kernel.compute_fn_ = fn;
return nullptr;
}
template <typename... Args>
void ParseArgs(MemberComputeType<Args...> fn) {
OrtLiteCustomOp::ParseArgs<Args...>(OrtLiteCustomOp::input_types_, OrtLiteCustomOp::output_types_);
}
// TODO: consider to disable these legacy functions for mobile build to save binary size
template <typename... Args>
void DefineCallbackFunctionsLegacy(MemberComputeType<Args...> fn, RegularComputeType regular_fn) {
OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* ort_api, const OrtKernelInfo* info) {
auto self = static_cast<const OrtLiteCustomStructV2<CustomOpKernel>*>(this_);
auto kernel = std::make_unique<KernelEx>();
typedef typename IsFunctionKernel<CustomOpKernel>::type type_flag;
auto status = InitKernel(*kernel, *ort_api, *info, self->regular_fn_, type_flag());
OrtW::ThrowOnError(*ort_api, status);
kernel->extra_.ep_ = self->execution_provider_;
kernel->extra_.api_ = std::make_unique<OrtW::CustomOpApi>(*ort_api);
return reinterpret_cast<void*>(kernel.release());
};
OrtCustomOp::KernelCompute = [](void* op_kernel, OrtKernelContext* context) {
auto kernel = reinterpret_cast<KernelEx*>(op_kernel);
std::vector<TensorPtr> tensors;
auto t = CreateTuple<0, 0, Args...>(kernel->extra_.api_.get(),
context,
tensors,
kernel->extra_.api_->KernelContext_GetInputCount(context),
kernel->extra_.api_->KernelContext_GetOutputCount(context),
kernel->extra_.ep_);
std::apply([kernel](Args const&... t_args) {
auto status = kernel->Compute(t_args...); OrtW::API::ThrowOnError(ToApiStatus(status)); }, t);
};
OrtCustomOp::KernelDestroy = [](void* op_kernel) {
std::unique_ptr<KernelEx>(reinterpret_cast<KernelEx*>(op_kernel)).reset();
};
}
#if ORT_API_VERSION >= 16
template <typename... Args>
void DefineCallbackFunctions(MemberComputeType<Args...> fn, RegularComputeType regular_fn) {
OrtCustomOp::CreateKernel = nullptr;
OrtCustomOp::KernelCompute = nullptr;
if constexpr (CustomOp_defined_getInputMemoryType<CustomOpKernel>::value) {
OrtCustomOp::GetInputMemoryType = [](const OrtCustomOp* /*this_*/, size_t index) -> OrtMemType {
return CustomOpKernel::GetInputMemoryType(index);
};
}
#if ORT_API_VERSION >= 18
if constexpr (CustomOp_defined_getMayInplace<CustomOpKernel>::value) {
OrtCustomOp::GetMayInplace = [](int** input_index, int** output_index) -> size_t {
return CustomOpKernel::GetMayInplace(input_index, output_index);
};
}
if constexpr (CustomOp_defined_releaseMayInplace<CustomOpKernel>::value) {
OrtCustomOp::ReleaseMayInplace = [](int* input_index, int* output_index) -> void {
CustomOpKernel::ReleaseMayInplace(input_index, output_index);
};
}
#endif
OrtCustomOp::CreateKernelV2 = [](const OrtCustomOp* this_,
const OrtApi* api, const OrtKernelInfo* info, void** op_kernel) -> OrtStatusPtr {
if (api == nullptr) {
assert(false && "Got a null pointer for ORT api on calling CreateKernelV2");
// should never happened, what we can do?
return nullptr;
}
if (this_ == nullptr || info == nullptr || op_kernel == nullptr) {
return api->CreateStatus(ORT_INVALID_ARGUMENT, "OrtCustomOp::CreateKernelV2: received a null pointer");
}
auto self = static_cast<const OrtLiteCustomStructV2<CustomOpKernel>*>(this_);
auto kernel = std::make_unique<KernelEx>();
if (kernel == nullptr) {
return api->CreateStatus(ORT_FAIL, "OrtCustomOp::CreateKernelV2: failed to new a kernel, OOM?");
}
typedef typename IsFunctionKernel<CustomOpKernel>::type flag_type;
auto status = InitKernel(*kernel, *api, *info, self->regular_fn_, flag_type());
if (status == nullptr) {
kernel->extra_.ep_ = self->execution_provider_;
kernel->extra_.api_ = std::make_unique<OrtW::CustomOpApi>(*api);
*op_kernel = reinterpret_cast<void*>(kernel.release());
}
return status;
};
OrtCustomOp::KernelComputeV2 = [](void* op_kernel, OrtKernelContext* context) -> OrtStatusPtr {
auto kernel = reinterpret_cast<KernelEx*>(op_kernel);
std::vector<TensorPtr> tensors;
auto t = CreateTuple<0, 0, Args...>(kernel->extra_.api_.get(),
context,
tensors,
kernel->extra_.api_->KernelContext_GetInputCount(context),
kernel->extra_.api_->KernelContext_GetOutputCount(context),
kernel->extra_.ep_);
return std::apply([kernel](Args const&... t_args) {
auto status = kernel->Compute(t_args...);
return ToApiStatus(status); }, t);
};
OrtCustomOp::KernelDestroy = [](void* op_kernel) {
std::unique_ptr<KernelEx>(reinterpret_cast<KernelEx*>(op_kernel)).reset();
};
}
#endif // ORT_API_VERSION >= 16
OrtLiteCustomStructV2(const char* op_name,
const char* execution_provider,
RegularComputeType fn_compute = nullptr)
: OrtLiteCustomOp(op_name, execution_provider), regular_fn_(fn_compute) {
ParseArgs(&CustomOpKernel::Compute);
#if ORT_API_VERSION >= 16
if (OrtCustomOp::version >= 16) {
DefineCallbackFunctions(&CustomOpKernel::Compute, fn_compute);
} else
#endif // ORT_API_VERSION >= 16
{
DefineCallbackFunctionsLegacy(&CustomOpKernel::Compute, fn_compute);
}
}
RegularComputeType regular_fn_{};
};
template <typename RType, typename... Args>
OrtLiteCustomOp* CreateLiteCustomOpV2(const char* op_name,
const char* execution_provider,
RType (*custom_compute_fn)(Args...)) {
using LiteOp = OrtLiteCustomStructV2<FunctionKernel<RType, Args...>>;
return std::make_unique<LiteOp>(op_name, execution_provider, custom_compute_fn).release();
}
template <typename OpKernel>
OrtLiteCustomOp* CreateLiteCustomOpV2(const char* op_name,
const char* execution_provider) {
using LiteOp = OrtLiteCustomStructV2<OpKernel>;
return std::make_unique<LiteOp>(op_name, execution_provider).release();
}
} // namespace Custom
} // namespace Ort
namespace ortc = Ort::Custom;