File size: 22,661 Bytes
c206440 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 | // Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
#pragma once
#include "boost/mp11.hpp"
// It is safe to include the below header even if SHARED_PROVIDER macro is enabled
// as it doesn't include any pb headers.
#include "core/framework/prepacked_weights_container.h"
#ifndef SHARED_PROVIDER
#include <functional>
#include "core/common/exceptions.h"
#include "core/common/logging/logging.h"
#include "core/common/status.h"
#include "core/framework/execution_provider.h"
#include "core/framework/kernel_def_builder.h"
#include "core/framework/ort_value.h"
#include "core/framework/op_kernel_info.h"
#include "core/framework/op_node_proto_helper.h"
#include "core/framework/tensor.h"
#include "core/framework/sparse_tensor.h"
#include "core/graph/constants.h"
#include "core/graph/graph_viewer.h"
#if !defined(ORT_MINIMAL_BUILD)
#include "onnx/defs/schema.h"
#else
#include "onnx/defs/data_type_utils.h"
#endif
#include "onnx/onnx_pb.h"
#include "onnx/onnx-operators_pb.h"
#include "core/common/gsl.h"
namespace onnxruntime {
class OpKernelContext;
}
#endif
namespace onnxruntime {
std::unique_ptr<OpKernelInfo> CopyOpKernelInfo(const OpKernelInfo& info);
class OpKernel {
public:
using DoneCallback = std::function<void()>;
explicit OpKernel(const OpKernelInfo& info) : op_kernel_info_(CopyOpKernelInfo(info)) {}
virtual ~OpKernel() = default;
const onnxruntime::Node& Node() const;
const onnxruntime::KernelDef& KernelDef() const;
[[nodiscard]] virtual Status Compute(_Inout_ OpKernelContext* context) const = 0;
[[nodiscard]] virtual bool IsAsync() const {
// by default all kernels are sync version.
return false;
}
[[nodiscard]] virtual Status ComputeAsync(_Inout_ OpKernelContext*, DoneCallback) const {
ORT_NOT_IMPLEMENTED(__FUNCTION__, " is not implemented");
}
// Override this function to PrePack initialized constant tensor to the format as needed.
// For example, MatMul kernel can pack the input B if it is constant like code below.
// Status PrePack(const Tensor& tensor, int input_idx, /*out*/ bool& is_packed,
// /*out*/ PrePackedWeights* prepacked_weight_for_caching,
// AllocatorPtr alloc) override {
// is_packed = false;
// if (input_idx == 1) {
// is_packed = true;
// this.Pack(tensor, this.buffer_, alloc);
// if (prepacked_weight_for_caching) {
// // LOGIC TO CACHE `this.buffer_` SINCE THE KERNEL DOESN"T OWN THE PACKED WEIGHT
// }
// }
// return Status::OK();
// }
// Please refer to MatMulIntegerToFloatBase for a complete example
// @param tensor: The initialized constant tensor
// @param input_idx: The input index of the tensor in this kernel
// @param alloc: The kernel's PrePack() method MUST use this allocator for allocating the pre-packed
// weights' buffers. The alloc that the PrePack() method will receive will be either
// the allocator tied to the session if the kernel owns the pre-packed buffer or an
// allocator shared between sessions if the pre-packed buffer is to be shared across sessions
// (i.e.) the kernel does not own the buffer.
// @param is_packed: Set it to true if the kernel packed the tensor or to false
// The kernel is responsible for keeping the packed data and related metadata if is_packed is true,
// and the original initialized constant tensor will be released and not accessible anymore in
// the Compute function.
// @param prepacked_weights: A PrePackedWeights instance will be provided to the kernel IF the pre-packed weights
// are meant to be stored in a shared container.
virtual Status
PrePack(const Tensor& /*tensor*/, int /*input_idx*/, AllocatorPtr /*alloc*/,
/*out*/ bool& is_packed, /*out*/ PrePackedWeights* /*prepacked_weights*/) {
is_packed = false;
return Status::OK();
}
// Override this function to use provided pre-packed weight.
// Status UseSharedPrePackedBuffers(std::vector<BufferUniquePtr>& prepacked_buffers,
// int input_idx,
// /*out*/ bool& used_shared_buffers) {
// used_shared_buffers = true;
// this.buffer_ = std::move(prepacked_buffers[0]);
// return Status::OK();
// }
// Please refer to MatMulIntegerToFloatBase for a complete example
// @param prepacked_buffers: The pre-packed buffers to be used by this kernel for the provided input index
// (Sometimes a single constant initializer may have multiple pre-packed buffers associated
// with it and it upto the kernel developer to store it in any order of their choice in PrePack()
// and must use the same order for retrieval in UseSharedPrePackedBuffers().
// @param input_idx: The input index of the tensor in this kernel
// @param used_shared_buffers: Boolean flag set by the kernel implementation indicating
// that the provided weight has been used by the kernel.
virtual Status UseSharedPrePackedBuffers(std::vector<BufferUniquePtr>& /*prepacked_buffers*/,
int /*input_idx*/,
/*out*/ bool& used_shared_buffers) {
used_shared_buffers = false;
return Status::OK();
}
const OrtMemoryInfo& Allocator(OrtMemType mem_type) const;
const OpKernelInfo& Info() const {
return *op_kernel_info_;
}
private:
ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(OpKernel);
std::unique_ptr<OpKernelInfo> op_kernel_info_;
};
class FuncManager;
using KernelCreateFn = std::function<Status(FuncManager& func_mgr, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out)>;
using KernelCreatePtrFn = std::add_pointer<Status(FuncManager& func_mgr, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out)>::type;
struct KernelCreateInfo {
std::unique_ptr<KernelDef> kernel_def; // Owned and stored in the global kernel registry.
KernelCreateFn kernel_create_func;
Status status;
KernelCreateInfo(std::unique_ptr<KernelDef> definition,
KernelCreateFn create_func)
: kernel_def(std::move(definition)),
kernel_create_func(create_func) {}
KernelCreateInfo(KernelCreateInfo&& other) noexcept
: kernel_def(std::move(other.kernel_def)),
kernel_create_func(std::move(other.kernel_create_func)) {}
KernelCreateInfo() = default;
};
// Forward declarations for the non-specialized BuildKernelCreateInfo method.
template <typename T>
KernelCreateInfo BuildKernelCreateInfo();
namespace ml {
template <typename T>
KernelCreateInfo BuildKernelCreateInfo();
} // namespace ml
namespace contrib {
template <typename T>
KernelCreateInfo BuildKernelCreateInfo();
} // namespace contrib
namespace contrib {
namespace cuda {
template <typename T>
KernelCreateInfo BuildKernelCreateInfo();
} // namespace cuda
} // namespace contrib
namespace contrib {
namespace rocm {
template <typename T>
KernelCreateInfo BuildKernelCreateInfo();
} // namespace rocm
} // namespace contrib
namespace contrib {
namespace snpe {
template <typename T>
KernelCreateInfo BuildKernelCreateInfo();
} // namespace snpe
} // namespace contrib
using BuildKernelCreateInfoFn = KernelCreateInfo (*)();
// Naming convention for operator kernel classes
#define ONNX_OPERATOR_KERNEL_CLASS_NAME(provider, domain, ver, name) \
provider##_##name##_##domain##_ver##ver
#define ONNX_CPU_OPERATOR_KERNEL(name, ver, builder, ...) \
ONNX_OPERATOR_KERNEL_EX(name, kOnnxDomain, ver, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_CPU_OPERATOR_ML_KERNEL(name, ver, builder, ...) \
ONNX_OPERATOR_KERNEL_EX(name, kMLDomain, ver, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_CPU_OPERATOR_MS_KERNEL(name, ver, builder, ...) \
ONNX_OPERATOR_KERNEL_EX(name, kMSDomain, ver, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_OPERATOR_KERNEL_EX(name, domain, ver, provider, builder, ...) \
class ONNX_OPERATOR_KERNEL_CLASS_NAME(provider, domain, ver, name); \
template <> \
KernelCreateInfo \
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(provider, domain, ver, name)>() { \
return KernelCreateInfo( \
builder.SetName(#name) \
.SetDomain(domain) \
.SinceVersion(ver) \
.Provider(provider) \
.Build(), \
static_cast<KernelCreatePtrFn>( \
[](FuncManager&, \
const OpKernelInfo& info, \
std::unique_ptr<OpKernel>& out) -> Status { \
out = std::make_unique<__VA_ARGS__>(info); \
return Status::OK(); \
})); \
}
#define ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(provider, domain, startver, endver, name) \
provider##_##name##_##domain##_ver##startver##_##endver
#define ONNX_CPU_OPERATOR_VERSIONED_KERNEL(name, startver, endver, builder, ...) \
ONNX_OPERATOR_VERSIONED_KERNEL_EX(name, kOnnxDomain, startver, endver, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_CPU_OPERATOR_VERSIONED_ML_KERNEL(name, startver, endver, builder, ...) \
ONNX_OPERATOR_VERSIONED_KERNEL_EX(name, kMLDomain, startver, endver, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_OPERATOR_VERSIONED_KERNEL_EX(name, domain, startver, endver, provider, builder, ...) \
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(provider, domain, startver, endver, name); \
template <> \
KernelCreateInfo \
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(provider, domain, startver, endver, name)>() { \
return KernelCreateInfo( \
builder.SetName(#name) \
.SetDomain(domain) \
.SinceVersion(startver, endver) \
.Provider(provider) \
.Build(), \
static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
}
#define ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type, name) \
provider##_##name##_##domain##_ver##ver##_##type
#define ONNX_CPU_OPERATOR_TYPED_KERNEL(name, ver, type, builder, ...) \
ONNX_OPERATOR_TYPED_KERNEL_EX(name, kOnnxDomain, ver, type, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_CPU_OPERATOR_TYPED_ML_KERNEL(name, ver, type, builder, ...) \
ONNX_OPERATOR_TYPED_KERNEL_EX(name, kMLDomain, ver, type, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_CPU_OPERATOR_TYPED_MS_KERNEL(name, ver, type, builder, ...) \
ONNX_OPERATOR_TYPED_KERNEL_EX(name, kMSDomain, ver, type, kCpuExecutionProvider, builder, __VA_ARGS__)
#define ONNX_OPERATOR_TYPED_KERNEL_EX(name, domain, ver, type, provider, builder, ...) \
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type, name); \
template <> \
KernelCreateInfo \
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type, name)>() { \
return KernelCreateInfo( \
builder.SetName(#name) \
.SetDomain(domain) \
.SinceVersion(ver) \
.Provider(provider) \
.Build(), \
static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
}
#define ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type1, type2, name) \
provider##_##name##_##domain##_ver##ver##_##type1##_##type2
#define ONNX_OPERATOR_TWO_TYPED_KERNEL_EX(name, domain, ver, type1, type2, provider, builder, ...) \
class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type1, type2, name); \
template <> \
KernelCreateInfo \
BuildKernelCreateInfo<ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type1, type2, name)>() { \
return KernelCreateInfo( \
builder.SetName(#name) \
.SetDomain(domain) \
.SinceVersion(ver) \
.Provider(provider) \
.Build(), \
static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
}
#define ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type, name) \
provider##_##name##_##domain##_ver##startver##_##endver##_##type
#define ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(name, startver, endver, type, builder, ...) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, kOnnxDomain, startver, endver, type, kCpuExecutionProvider, builder, \
__VA_ARGS__)
#define ONNX_CPU_OPERATOR_VERSIONED_TYPED_ML_KERNEL(name, startver, endver, type, builder, ...) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, kMLDomain, startver, endver, type, kCpuExecutionProvider, builder, \
__VA_ARGS__)
#define ONNX_CPU_OPERATOR_VERSIONED_TYPED_MS_KERNEL(name, startver, endver, type, builder, ...) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, kMSDomain, startver, endver, type, kCpuExecutionProvider, builder, \
__VA_ARGS__)
#define ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, domain, startver, endver, type, provider, builder, ...) \
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type, name); \
template <> \
KernelCreateInfo \
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, \
type, name)>() { \
return KernelCreateInfo( \
builder.SetName(#name) \
.SetDomain(domain) \
.SinceVersion(startver, endver) \
.Provider(provider) \
.Build(), \
static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
}
#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>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
}
template <typename... Types>
struct BuildKernelDefConstraintsImpl {
std::vector<MLDataType> operator()() const {
return {DataTypeImpl::GetTensorType<Types>()...};
}
};
#if !defined(DISABLE_SPARSE_TENSORS)
template <typename... Types>
struct BuildKernelDefSparseConstraintsImpl {
std::vector<MLDataType> operator()() const {
return {DataTypeImpl::GetSparseTensorType<Types>()...};
}
};
#endif
// 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_>())
template <typename... Types>
inline std::vector<MLDataType> BuildKernelDefConstraints() {
return BuildKernelDefConstraintsImpl<Types...>{}();
}
#if !defined(DISABLE_SPARSE_TENSORS)
template <typename... Types>
inline std::vector<MLDataType> BuildKernelDefSparseConstraints() {
return BuildKernelDefSparseConstraintsImpl<Types...>{}();
}
#endif
// version of BuildKernelDefConstraints() which takes a type list
template <typename L>
inline std::vector<MLDataType> BuildKernelDefConstraintsFromTypeList() {
return boost::mp11::mp_apply<BuildKernelDefConstraintsImpl, L>{}();
}
#if !defined(DISABLE_SPARSE_TENSORS)
template <typename L>
inline std::vector<MLDataType> BuildKernelDefSparseConstraintsFromTypeList() {
return boost::mp11::mp_apply<BuildKernelDefSparseConstraintsImpl, L>{}();
}
#endif
} // namespace onnxruntime
#ifndef SHARED_PROVIDER
#include "core/framework/op_kernel_context.h"
#endif
|