op_meta_info.h 38.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#pragma once

17
#include <iostream>
18 19
#include <string>
#include <unordered_map>
20
#include <utility>
21 22
#include <vector>

23
#include "paddle/phi/api/ext/exception.h"
24
#include "paddle/phi/api/include/dll_decl.h"
25
#include "paddle/phi/api/include/tensor.h"
26
#include "paddle/utils/any.h"
27 28
#include "paddle/utils/none.h"
#include "paddle/utils/optional.h"
29 30 31 32 33 34 35 36 37 38

/**
 * Op Meta Info Related Define.
 *
 * Used to maintain operator core information.
 *
 */

namespace paddle {

39
class PADDLE_API OpMetaInfoHelper;
40
using Tensor = paddle::Tensor;
41

42 43
///////////////// Util Marco Define ////////////////

44 45 46 47 48 49 50
#define PD_DISABLE_COPY_AND_ASSIGN(classname)      \
 private:                                          \
  classname(const classname&) = delete;            \
  classname(classname&&) = delete;                 \
  classname& operator=(const classname&) = delete; \
  classname& operator=(classname&&) = delete

51 52 53 54 55 56
#define STATIC_ASSERT_GLOBAL_NAMESPACE(uniq_name, msg)                        \
  struct __test_global_namespace_##uniq_name##__ {};                          \
  static_assert(std::is_same<::__test_global_namespace_##uniq_name##__,       \
                             __test_global_namespace_##uniq_name##__>::value, \
                msg)

57 58
///////////////// Util Define and Function ////////////////

59 60
constexpr char kGradTensorSuffix[] = "@GRAD";
constexpr char kTensorVectorSuffix[] = "@VECTOR";
61
constexpr char kDoubleGradNewOutSuffix[] = "@NEW";
62
constexpr char kOptionalSuffix[] = "@OPTIONAL";
63 64 65 66 67 68 69 70 71 72 73 74

// Used for Construct Grad Tensor name
inline std::string Grad(const std::string& t_name) {
  std::string result;
  result.reserve(t_name.size() + 5U);
  result += t_name;
  result += kGradTensorSuffix;
  return result;
}

// Used for Construct std::vector<Tensor> name
inline std::string Vec(const std::string& t_name) {
75
  std::string result;
76 77 78
  result.reserve(t_name.size() + 7U);
  result += t_name;
  result += kTensorVectorSuffix;
79 80 81
  return result;
}

82 83 84 85 86 87 88 89 90
// Used for Construct double grad output name
inline std::string New(const std::string& t_name) {
  std::string result;
  result.reserve(t_name.size() + 4U);
  result += t_name;
  result += kDoubleGradNewOutSuffix;
  return result;
}

91 92 93 94 95 96 97 98 99
// Used for Construct paddle::optional name
inline std::string Optional(const std::string& t_name) {
  std::string result;
  result.reserve(t_name.size() + 9U);
  result += t_name;
  result += kOptionalSuffix;
  return result;
}

100 101 102 103 104 105 106 107 108
PADDLE_API void AssignTensorImpl(const Tensor& src, Tensor* dst);

////////////////////// Kernel Context ////////////////////////

class PADDLE_API CustomOpKernelContext {
 public:
  CustomOpKernelContext() = default;

  void EmplaceBackInput(Tensor&& input);
109
  void EmplaceBackInputs(const std::vector<Tensor>& inputs);
110
  void EmplaceBackOutput(Tensor&& output);
111
  void EmplaceBackOutputs(const std::vector<Tensor>& outputs);
112
  void EmplaceBackAttr(paddle::any attr);
113 114 115
  void EmplaceBackAttrs(const std::vector<paddle::any>& attrs) {
    attrs_ = std::move(attrs);
  }
116 117 118 119 120
  const std::pair<size_t, size_t>& InputRangeAt(size_t idx) const;
  const std::pair<size_t, size_t>& OutputRangeAt(size_t idx) const;

  const Tensor& InputAt(size_t idx) const;
  std::vector<Tensor> InputsBetween(size_t start, size_t end) const;
121
  Tensor& MutableInputAt(size_t idx);
122
  paddle::optional<Tensor> OptionalInputAt(size_t idx);
123 124
  paddle::optional<std::vector<Tensor>> OptionalInputsBetween(size_t start,
                                                              size_t end);
125 126 127 128 129 130 131
  const std::vector<paddle::any>& Attrs() const { return attrs_; }
  const std::vector<std::pair<size_t, size_t>>& InputRange() {
    return input_range_;
  }
  const std::vector<std::pair<size_t, size_t>>& OutputRange() {
    return output_range_;
  }
132 133
  Tensor* MutableOutputAt(size_t idx);
  std::vector<Tensor*> MutableOutputBetweeen(size_t start, size_t end);
134
  std::vector<Tensor> OutputsBetweeen(size_t start, size_t end);
135 136 137 138 139 140 141 142 143 144 145
  std::vector<Tensor>* AllMutableOutput();

  template <typename AttrType>
  AttrType AttrAt(size_t idx) const {
    try {
      return paddle::any_cast<AttrType>(attrs_.at(idx));
    } catch (paddle::bad_any_cast&) {
      PD_THROW("Attribute cast error in Custom Op Kernel Context.");
    }
  }

146
  // handle inplace map
147 148 149 150 151 152 153 154
  void MapPlainOutputs(
      const std::vector<std::string>& inputs,
      const std::vector<std::string>& outputs,
      const std::unordered_map<std::string, std::string>& inplace_map);
  void AssignInplaceOutputs();
  std::vector<Tensor*>* AllMutablePlainOutput();
  std::unordered_map<size_t, size_t> GetInplaceTensorMap();

155 156 157 158 159
 private:
  // TODO(chenweihang): replaced be SmallVector
  std::vector<Tensor> inputs_;
  std::vector<Tensor> outputs_;
  std::vector<paddle::any> attrs_;
160
  // handle inplace map
161 162
  std::vector<Tensor*> plain_outputs_;
  std::unordered_map<size_t, size_t> inplace_tensor_map_;
163 164 165 166 167

  std::vector<std::pair<size_t, size_t>> input_range_;
  std::vector<std::pair<size_t, size_t>> output_range_;
};

168 169 170
////////////////////// Kernel Function (PD_KERNEL) ////////////////////////

// Record Op kernel core function
171 172 173 174 175 176
using KernelFunc = void (*)(CustomOpKernelContext*);

#define PD_SPECIALIZE_ComputeCallHelper(attr_type)                             \
  template <typename... Tail>                                                  \
  struct ComputeCallHelper<attr_type, Tail...> {                               \
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs> \
177
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {  \
178 179 180 181 182 183
      attr_type arg = ctx->AttrAt<attr_type>(attr_idx);                        \
      ComputeCallHelper<                                                       \
          Tail...>::template Compute<in_idx, attr_idx + 1, out_idx>(ctx,       \
                                                                    pargs...,  \
                                                                    arg);      \
    }                                                                          \
184 185
  }

186 187 188 189 190 191 192 193
template <typename T>
struct TypeTag {};

template <typename F, F f>
struct KernelFuncImpl;

template <typename Return, typename... Args, Return (*impl_fn)(Args...)>
struct KernelFuncImpl<Return (*)(Args...), impl_fn> {
194 195
  static void Compute(CustomOpKernelContext* ctx) {
    ComputeCallHelper<Args..., TypeTag<int>>::template Compute<0, 0, 0>(ctx);
196 197 198 199 200 201
  }

 private:
  template <typename... RemainingArgs>
  struct ComputeCallHelper;

202
  // Handle args for general Tensor input case
203 204
  template <typename... Tail>
  struct ComputeCallHelper<const Tensor&, Tail...> {
205
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
206
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
207
      auto& range = ctx->InputRangeAt(in_idx);
208
      auto& arg = ctx->MutableInputAt(range.first);
209 210 211 212
      ComputeCallHelper<
          Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                    pargs...,
                                                                    arg);
213 214 215
    }
  };

216
  // Handle args for optional Tensor input case
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
  template <typename... Tail>
  struct ComputeCallHelper<const paddle::optional<paddle::Tensor>&, Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
      auto& range = ctx->InputRangeAt(in_idx);
      auto& arg = ctx->InputAt(range.first);
      if (!arg.is_initialized()) {
        ComputeCallHelper<Tail...>::
            template Compute<in_idx + 1, attr_idx, out_idx>(
                ctx, pargs..., paddle::none);
      } else {
        ComputeCallHelper<
            Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                      pargs...,
                                                                      arg);
      }
    }
  };

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
  // Handle args for inplace Tensor case
  template <typename... Tail>
  struct ComputeCallHelper<Tensor&, Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
      auto& range = ctx->InputRangeAt(in_idx);
      auto& arg = ctx->MutableInputAt(range.first);
      ComputeCallHelper<
          Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                    pargs...,
                                                                    arg);
    }
  };

  // Handle args for optional inplace Tensor input case
  template <typename... Tail>
  struct ComputeCallHelper<paddle::optional<paddle::Tensor>&, Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
      auto& range = ctx->InputRangeAt(in_idx);
      auto arg = ctx->OptionalInputAt(range.first);
      ComputeCallHelper<
          Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                    pargs...,
                                                                    arg);
    }
  };

264
  // Handle args for general vector<Tensor> input case
265 266
  template <typename... Tail>
  struct ComputeCallHelper<const std::vector<Tensor>&, Tail...> {
267
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
268
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
269 270 271 272 273 274
      auto& range = ctx->InputRangeAt(in_idx);
      auto arg = ctx->InputsBetween(range.first, range.second);
      ComputeCallHelper<
          Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                    pargs...,
                                                                    arg);
275 276 277
    }
  };

278
  // Handle args for optional vector<Tensor> input case
279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298
  template <typename... Tail>
  struct ComputeCallHelper<const paddle::optional<std::vector<paddle::Tensor>>&,
                           Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
      auto& range = ctx->InputRangeAt(in_idx);
      auto arg = ctx->InputsBetween(range.first, range.second);
      if (arg.empty() || !arg[0].is_initialized()) {
        ComputeCallHelper<Tail...>::
            template Compute<in_idx + 1, attr_idx, out_idx>(
                ctx, pargs..., paddle::none);
      } else {
        ComputeCallHelper<
            Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                      pargs...,
                                                                      arg);
      }
    }
  };

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
  // Handle args for inplace vector<Tensor> case
  template <typename... Tail>
  struct ComputeCallHelper<std::vector<Tensor>&, Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
      auto& range = ctx->InputRangeAt(in_idx);
      auto arg = ctx->InputsBetween(range.first, range.second);
      ComputeCallHelper<
          Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                    pargs...,
                                                                    arg);
    }
  };

  // Handle args for optional inplace vector<Tensor> case
  template <typename... Tail>
  struct ComputeCallHelper<paddle::optional<std::vector<Tensor>>&, Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
      auto& range = ctx->InputRangeAt(in_idx);
      auto arg = ctx->OptionalInputsBetween(range.first, range.second);
      ComputeCallHelper<
          Tail...>::template Compute<in_idx + 1, attr_idx, out_idx>(ctx,
                                                                    pargs...,
                                                                    arg);
    }
  };

327 328 329 330
  PD_SPECIALIZE_ComputeCallHelper(bool);
  PD_SPECIALIZE_ComputeCallHelper(int);
  PD_SPECIALIZE_ComputeCallHelper(float);
  PD_SPECIALIZE_ComputeCallHelper(int64_t);
331 332 333 334 335 336 337
  PD_SPECIALIZE_ComputeCallHelper(const std::string&);
  PD_SPECIALIZE_ComputeCallHelper(const std::vector<int>&);
  PD_SPECIALIZE_ComputeCallHelper(const std::vector<float>&);
  PD_SPECIALIZE_ComputeCallHelper(const std::vector<int64_t>&);
  PD_SPECIALIZE_ComputeCallHelper(const std::vector<std::string>&);
  // TODO(chenweihang): support other attribute type if needed.
  // Why not support other attribute type here?
R
Ruibiao Chen 已提交
338
  // - paddle::blank, std::vector<bool> and std::vector<double>
339 340 341 342 343
  //   are not used in op
  // - BlockDesc* and std::vector<BlockDesc*> are used in framework

  // NOTE(chenweihang): Used to be compatible with the 2.0.1 released
  // interface, and will be deprecated in the future
344 345 346 347
  PD_SPECIALIZE_ComputeCallHelper(const bool&);
  PD_SPECIALIZE_ComputeCallHelper(const int&);
  PD_SPECIALIZE_ComputeCallHelper(const float&);
  PD_SPECIALIZE_ComputeCallHelper(const int64_t&);
348

C
Chen Weihang 已提交
349 350 351 352 353 354 355 356
  // NOTE(chenweihang): Used to be compatible with the 2.1 released
  // interface, but not recommended
  PD_SPECIALIZE_ComputeCallHelper(std::string);
  PD_SPECIALIZE_ComputeCallHelper(std::vector<int>);
  PD_SPECIALIZE_ComputeCallHelper(std::vector<float>);
  PD_SPECIALIZE_ComputeCallHelper(std::vector<int64_t>);
  PD_SPECIALIZE_ComputeCallHelper(std::vector<std::string>);

357 358
  // Used to be compatible with 2.3 released internal inplace interface, not
  // recommended
359
  // Handle args for compatible inplace case
360 361 362
  template <typename... Tail>
  struct ComputeCallHelper<Tensor*, Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
363
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
364 365 366 367 368 369 370 371 372
      auto& range = ctx->OutputRangeAt(out_idx);
      auto* arg = ctx->MutableOutputAt(range.first);
      ComputeCallHelper<
          Tail...>::template Compute<in_idx, attr_idx, out_idx + 1>(ctx,
                                                                    pargs...,
                                                                    arg);
    }
  };

373 374
  // Used to be compatible with 2.3 released internal inplace interface, not
  // recommended
375 376
  // TODO(chenweihang): What is the appropriate output form?
  // std::vector<Tensor>*? or std::vector<Tensor*>? or std::vector<Tensor*>*
377
  // Handle args for compatible inplace case
378 379 380
  template <typename... Tail>
  struct ComputeCallHelper<std::vector<Tensor*>, Tail...> {
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
381
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396
      auto& range = ctx->OutputRangeAt(out_idx);
      auto arg = ctx->MutableOutputBetweeen(range.first, range.second);
      ComputeCallHelper<
          Tail...>::template Compute<in_idx, attr_idx, out_idx + 1>(ctx,
                                                                    pargs...,
                                                                    arg);
    }
  };

  template <int out_idx, typename T>
  struct ComputeReturnHelper;

  // For compatibility with the original custom op form
  template <int out_idx>
  struct ComputeReturnHelper<out_idx, std::vector<Tensor>> {
397
    static void Compute(CustomOpKernelContext* ctx, Args&... args) {
398 399
      static_assert(out_idx == 0,
                    "If return std::vector<Tensor> in Custom OpKernel, "
H
HongyuJia 已提交
400
                    "you cannot pass output by kernel function argument.");
401
      auto outs = impl_fn(args...);
402
      auto* orig_outs = ctx->AllMutablePlainOutput();
403 404 405 406 407 408 409 410
      PD_CHECK(orig_outs->size() == outs.size(),
               "The number of element in custom operator outputs is wrong, "
               "expected contains ",
               orig_outs->size(),
               " Tensors, but actually contains ",
               outs.size(),
               " Tensors.");
      for (size_t i = 0; i < outs.size(); ++i) {
411
        AssignTensorImpl(outs.at(i), orig_outs->at(i));
412 413 414 415 416 417
      }
    }
  };

  template <int out_idx>
  struct ComputeReturnHelper<out_idx, void> {
418
    static void Compute(CustomOpKernelContext* ctx, Args&... args) {
419 420 421 422
      impl_fn(args...);
    }
  };

423 424 425
  // end: base template
  template <typename T>
  struct ComputeCallHelper<TypeTag<T>> {
426
    template <int in_idx, int attr_idx, int out_idx, typename... PreviousArgs>
427
    static void Compute(CustomOpKernelContext* ctx, PreviousArgs&... pargs) {
428
      ComputeReturnHelper<out_idx, Return>::Compute(ctx, pargs...);
429 430 431 432 433 434 435 436 437 438 439
    }
  };
};

#define PD_KERNEL(...) \
  ::paddle::KernelFuncImpl<decltype(&__VA_ARGS__), &__VA_ARGS__>::Compute

/////////////// InferShape Function (PD_INFER_SHAPE) ///////////////

// Record Op infershape core function
using InferShapeFunc = std::vector<std::vector<int64_t>> (*)(
440
    const std::vector<std::vector<int64_t>>& input_shapes,
441
    const std::vector<std::vector<std::vector<int64_t>>>& vec_input_shapes,
442 443
    const std::vector<paddle::any>& attrs);

444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
#define PD_SPECIALIZE_InferShapeCallHelper_FOR_SHAPE(input_type)     \
  template <typename... Tail>                                        \
  struct InferShapeCallHelper<input_type, Tail...> {                 \
    template <int in_idx,                                            \
              int vec_in_idx,                                        \
              int attr_idx,                                          \
              typename... PreviousArgs>                              \
    static Return InferShape(                                        \
        const std::vector<std::vector<int64_t>>& input_shapes,       \
        const std::vector<std::vector<std::vector<int64_t>>>&        \
            vec_input_shapes,                                        \
        const std::vector<paddle::any>& attrs,                       \
        const PreviousArgs&... pargs) {                              \
      input_type arg = input_shapes[in_idx];                         \
      return InferShapeCallHelper<Tail...>::                         \
          template InferShape<in_idx + 1, vec_in_idx, attr_idx>(     \
              input_shapes, vec_input_shapes, attrs, pargs..., arg); \
    }                                                                \
462 463
  }

464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481
#define PD_SPECIALIZE_InferShapeCallHelper_FOR_SHAPES(input_type)    \
  template <typename... Tail>                                        \
  struct InferShapeCallHelper<input_type, Tail...> {                 \
    template <int in_idx,                                            \
              int vec_in_idx,                                        \
              int attr_idx,                                          \
              typename... PreviousArgs>                              \
    static Return InferShape(                                        \
        const std::vector<std::vector<int64_t>>& input_shapes,       \
        const std::vector<std::vector<std::vector<int64_t>>>&        \
            vec_input_shapes,                                        \
        const std::vector<paddle::any>& attrs,                       \
        const PreviousArgs&... pargs) {                              \
      input_type arg = vec_input_shapes[vec_in_idx];                 \
      return InferShapeCallHelper<Tail...>::                         \
          template InferShape<in_idx, vec_in_idx + 1, attr_idx>(     \
              input_shapes, vec_input_shapes, attrs, pargs..., arg); \
    }                                                                \
482 483
  }

484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511
#define PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(attr_type)               \
  template <typename... Tail>                                                \
  struct InferShapeCallHelper<attr_type, Tail...> {                          \
    template <int in_idx,                                                    \
              int vec_in_idx,                                                \
              int attr_idx,                                                  \
              typename... PreviousArgs>                                      \
    static Return InferShape(                                                \
        const std::vector<std::vector<int64_t>>& input_shapes,               \
        const std::vector<std::vector<std::vector<int64_t>>>&                \
            vec_input_shapes,                                                \
        const std::vector<paddle::any>& attrs,                               \
        const PreviousArgs&... pargs) {                                      \
      try {                                                                  \
        attr_type arg = paddle::any_cast<attr_type>(attrs[attr_idx]);        \
        return InferShapeCallHelper<Tail...>::                               \
            template InferShape<in_idx, vec_in_idx, attr_idx + 1>(           \
                input_shapes, vec_input_shapes, attrs, pargs..., arg);       \
      } catch (paddle::bad_any_cast&) {                                      \
        PD_THROW(                                                            \
            "Attribute cast error in custom operator InferShapeFn. "         \
            "Expected " #attr_type                                           \
            " value. InferShapeFn's attribute list must be exactly same as " \
            "Forward "                                                       \
            "KernelFn's attribute list except std::vector<int64_t> "         \
            "attribute.");                                                   \
      }                                                                      \
    }                                                                        \
512
  }
513 514 515 516 517 518

template <typename F, F f>
struct InferShapeFuncImpl;

template <typename Return, typename... Args, Return (*impl_fn)(Args...)>
struct InferShapeFuncImpl<Return (*)(Args...), impl_fn> {
519
  static Return InferShape(
520
      const std::vector<std::vector<int64_t>>& input_shapes,
521
      const std::vector<std::vector<std::vector<int64_t>>>& vec_input_shapes,
522
      const std::vector<paddle::any>& attrs) {
523 524
    return InferShapeCallHelper<Args..., TypeTag<int>>::
        template InferShape<0, 0, 0>(input_shapes, vec_input_shapes, attrs);
525 526 527 528 529 530
  }

 private:
  template <typename... RemainingArgs>
  struct InferShapeCallHelper;

531 532 533
  PD_SPECIALIZE_InferShapeCallHelper_FOR_SHAPE(const std::vector<int64_t>&);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_SHAPES(
      const std::vector<std::vector<int64_t>>&);
534

535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559
  template <typename... Tail>
  struct InferShapeCallHelper<const paddle::optional<std::vector<int64_t>>&,
                              Tail...> {
    template <int in_idx,
              int vec_in_idx,
              int attr_idx,
              typename... PreviousArgs>
    static Return InferShape(
        const std::vector<std::vector<int64_t>>& input_shapes,
        const std::vector<std::vector<std::vector<int64_t>>>& vec_input_shapes,
        const std::vector<paddle::any>& attrs,
        const PreviousArgs&... pargs) {
      const std::vector<int64_t>& arg = input_shapes[in_idx];
      if (arg.empty()) {
        return InferShapeCallHelper<Tail...>::
            template InferShape<in_idx + 1, vec_in_idx, attr_idx>(
                input_shapes, vec_input_shapes, attrs, pargs..., paddle::none);
      } else {
        return InferShapeCallHelper<Tail...>::
            template InferShape<in_idx + 1, vec_in_idx, attr_idx>(
                input_shapes, vec_input_shapes, attrs, pargs..., arg);
      }
    }
  };

560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586
  template <typename... Tail>
  struct InferShapeCallHelper<
      const paddle::optional<std::vector<std::vector<int64_t>>>&,
      Tail...> {
    template <int in_idx,
              int vec_in_idx,
              int attr_idx,
              typename... PreviousArgs>
    static Return InferShape(
        const std::vector<std::vector<int64_t>>& input_shapes,
        const std::vector<std::vector<std::vector<int64_t>>>& vec_input_shapes,
        const std::vector<paddle::any>& attrs,
        const PreviousArgs&... pargs) {
      const std::vector<std::vector<int64_t>>& arg =
          vec_input_shapes[vec_in_idx];
      if (arg.empty()) {
        return InferShapeCallHelper<Tail...>::
            template InferShape<in_idx, vec_in_idx + 1, attr_idx>(
                input_shapes, vec_input_shapes, attrs, pargs..., paddle::none);
      } else {
        return InferShapeCallHelper<Tail...>::
            template InferShape<in_idx, vec_in_idx + 1, attr_idx>(
                input_shapes, vec_input_shapes, attrs, pargs..., arg);
      }
    }
  };

587 588 589 590 591
  // NOTE(chenweihang): Used to be compatible with the 2.0.1 released
  // interface, and will be deprecated in the future
  PD_SPECIALIZE_InferShapeCallHelper_FOR_SHAPE(std::vector<int64_t>);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_SHAPES(
      std::vector<std::vector<int64_t>>);
592

593 594 595 596
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(bool);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(int);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(float);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(int64_t);
597 598 599 600 601 602 603 604
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const std::string&);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const std::vector<int>&);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const std::vector<float>&);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const std::vector<std::string>&);
  // NOTE(chenweihang): InferShape can't support std::vector<int64_t> attr type,
  // because the input type is std::vector<int64_t>, only can use one rule to
  // parse std::vector<int64_t> parameter

605 606 607 608 609 610 611
  // NOTE(chenweihang): Used to be compatible with the 2.0.1 released
  // interface, and will be deprecated in the future
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const bool&);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const int&);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const float&);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(const int64_t&);

C
Chen Weihang 已提交
612 613 614 615 616 617 618
  // NOTE(chenweihang): Used to be compatible with the 2.1 released
  // interface, but not recommended
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(std::string);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(std::vector<int>);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(std::vector<float>);
  PD_SPECIALIZE_InferShapeCallHelper_FOR_ATTR(std::vector<std::string>);

619 620 621
  // end: base template
  template <typename T>
  struct InferShapeCallHelper<TypeTag<T>> {
622
    template <int in_idx, int vec_in_idx, int attr_idx>
623
    static Return InferShape(
624 625
        const std::vector<std::vector<int64_t>>& input_shapes,
        const std::vector<std::vector<std::vector<int64_t>>>& vec_input_shapes,
626 627
        const std::vector<paddle::any>& attrs,
        const Args&... args) {
628 629 630 631 632 633 634 635 636 637 638
      return impl_fn(args...);
    }
  };
};

#define PD_INFER_SHAPE(...) \
  ::paddle::InferShapeFuncImpl<decltype(&__VA_ARGS__), &__VA_ARGS__>::InferShape

/////////////// InferDataType Function (PD_INFER_DTYPE) ///////////////

// Record Op Infer dtype core function
639
using InferDtypeFunc = std::vector<DataType> (*)(
640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657
    const std::vector<DataType>& input_dtypes,
    const std::vector<std::vector<DataType>>& vec_input_dtypes);

#define PD_SPECIALIZE_InferDtypeCallHelper_TO_DTYPE(input_type)              \
  template <typename... Tail>                                                \
  struct InferDtypeCallHelper<input_type, Tail...> {                         \
    template <int in_idx, int vec_in_idx, typename... PreviousArgs>          \
    static Return InferDtype(                                                \
        const std::vector<DataType>& input_dtypes,                           \
        const std::vector<std::vector<DataType>>& vec_input_dtypes,          \
        const PreviousArgs&... pargs) {                                      \
      input_type arg = input_dtypes[in_idx];                                 \
      return InferDtypeCallHelper<Tail...>::template InferDtype<in_idx + 1,  \
                                                                vec_in_idx>( \
          input_dtypes, vec_input_dtypes, pargs..., arg);                    \
    }                                                                        \
  }

658 659 660 661 662 663 664 665 666 667 668 669 670
#define PD_SPECIALIZE_InferDtypeCallHelper_FOR_DTYPES(input_type)   \
  template <typename... Tail>                                       \
  struct InferDtypeCallHelper<input_type, Tail...> {                \
    template <int in_idx, int vec_in_idx, typename... PreviousArgs> \
    static Return InferDtype(                                       \
        const std::vector<DataType>& input_dtypes,                  \
        const std::vector<std::vector<DataType>>& vec_input_dtypes, \
        const PreviousArgs&... pargs) {                             \
      input_type arg = vec_input_dtypes[vec_in_idx];                \
      return InferDtypeCallHelper<Tail...>::                        \
          template InferDtype<in_idx, vec_in_idx + 1>(              \
              input_dtypes, vec_input_dtypes, pargs..., arg);       \
    }                                                               \
671
  }
672 673 674 675 676 677

template <typename F, F f>
struct InferDtypeFuncImpl;

template <typename Return, typename... Args, Return (*impl_fn)(Args...)>
struct InferDtypeFuncImpl<Return (*)(Args...), impl_fn> {
678
  static Return InferDtype(
679 680
      const std::vector<DataType>& input_dtypes,
      const std::vector<std::vector<DataType>>& vec_input_dtypes) {
681 682 683
    return InferDtypeCallHelper<Args..., TypeTag<int>>::template InferDtype<0,
                                                                            0>(
        input_dtypes, vec_input_dtypes);
684 685 686 687 688 689
  }

 private:
  template <typename... RemainingArgs>
  struct InferDtypeCallHelper;

690 691
  PD_SPECIALIZE_InferDtypeCallHelper_TO_DTYPE(const DataType&);
  PD_SPECIALIZE_InferDtypeCallHelper_FOR_DTYPES(const std::vector<DataType>&);
692

693
  template <typename... Tail>
694
  struct InferDtypeCallHelper<const paddle::optional<DataType>&, Tail...> {
695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712
    template <int in_idx, int vec_in_idx, typename... PreviousArgs>
    static Return InferDtype(
        const std::vector<DataType>& input_dtypes,
        const std::vector<std::vector<DataType>>& vec_input_dtypes,
        const PreviousArgs&... pargs) {
      const DataType& arg = input_dtypes[in_idx];
      if (arg == DataType::UNDEFINED) {
        return InferDtypeCallHelper<Tail...>::template InferDtype<in_idx + 1,
                                                                  vec_in_idx>(
            input_dtypes, vec_input_dtypes, pargs..., paddle::none);
      } else {
        return InferDtypeCallHelper<Tail...>::template InferDtype<in_idx + 1,
                                                                  vec_in_idx>(
            input_dtypes, vec_input_dtypes, pargs..., arg);
      }
    }
  };

713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733
  template <typename... Tail>
  struct InferDtypeCallHelper<const paddle::optional<std::vector<DataType>>&,
                              Tail...> {
    template <int in_idx, int vec_in_idx, typename... PreviousArgs>
    static Return InferDtype(
        const std::vector<DataType>& input_dtypes,
        const std::vector<std::vector<DataType>>& vec_input_dtypes,
        const PreviousArgs&... pargs) {
      const std::vector<DataType>& arg = vec_input_dtypes[vec_in_idx];
      if (arg.empty()) {
        return InferDtypeCallHelper<Tail...>::
            template InferDtype<in_idx, vec_in_idx + 1>(
                input_dtypes, vec_input_dtypes, pargs..., paddle::none);
      } else {
        return InferDtypeCallHelper<Tail...>::
            template InferDtype<in_idx, vec_in_idx + 1>(
                input_dtypes, vec_input_dtypes, pargs..., arg);
      }
    }
  };

734 735 736 737
  // NOTE(chenweihang): Used to be compatible with the 2.0.1 released
  // interface, and will be deprecated in the future
  PD_SPECIALIZE_InferDtypeCallHelper_TO_DTYPE(DataType);
  PD_SPECIALIZE_InferDtypeCallHelper_FOR_DTYPES(std::vector<DataType>);
738 739 740 741

  // end: base template
  template <typename T>
  struct InferDtypeCallHelper<TypeTag<T>> {
742 743
    template <int in_idx, int vec_in_idx>
    static Return InferDtype(
744 745
        const std::vector<DataType>& input_dtypes,
        const std::vector<std::vector<DataType>>& vec_input_dtypes,
746
        const Args&... args) {
747 748 749 750 751 752 753 754 755 756
      return impl_fn(args...);
    }
  };
};

#define PD_INFER_DTYPE(...) \
  ::paddle::InferDtypeFuncImpl<decltype(&__VA_ARGS__), &__VA_ARGS__>::InferDtype

////////////////////// Op Meta Info //////////////////////

757
class PADDLE_API OpMetaInfo {
758 759
 public:
  explicit OpMetaInfo(const std::string& op_name) : name_(op_name) {}
760 761

  // format: {"<name1>", "<name2>", ...}
762
  OpMetaInfo& Inputs(std::vector<std::string>&& inputs);
763 764

  // format: {"<name1>", "<name2>", ...}
765
  OpMetaInfo& Outputs(std::vector<std::string>&& outputs);
766

767
  // format: {"<name1>:<type1>", "<name2>:<type2>", ...}
768 769
  OpMetaInfo& Attrs(std::vector<std::string>&& attrs);

770 771
  // format: {"<input_name1>:<output_name1>",
  // "<input_name2>:<output_name2>",...}
772
  OpMetaInfo& SetInplaceMap(
773 774
      std::unordered_map<std::string, std::string>&& inplace_map);

775
  // format: PD_KERNEL(...)
776
  OpMetaInfo& SetKernelFn(KernelFunc&& func);
777 778

  // format: PD_INFER_SHAPE(...)
779
  OpMetaInfo& SetInferShapeFn(InferShapeFunc&& func);
780 781

  // format: PD_INFER_DTYPE(...)
782 783 784
  OpMetaInfo& SetInferDtypeFn(InferDtypeFunc&& func);

 private:
785
  friend class OpMetaInfoHelper;
786 787 788 789 790 791

  // 1. desc info
  std::string name_;
  std::vector<std::string> inputs_;
  std::vector<std::string> outputs_;
  std::vector<std::string> attrs_;
792
  std::unordered_map<std::string, std::string> inplace_map_;
793
  std::unordered_map<std::string, std::string> inplace_reverse_map_;
794
  // 2. func info
795 796 797
  KernelFunc kernel_fn_{nullptr};
  InferShapeFunc infer_shape_fn_{nullptr};
  InferDtypeFunc infer_dtype_fn_{nullptr};
798 799
};

800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821
//////////////// Op Meta Info Helper /////////////////
class OpMetaInfoHelper {
 public:
  static const std::string& GetOpName(const paddle::OpMetaInfo& info) {
    return info.name_;
  }
  static const std::vector<std::string>& GetInputs(
      const paddle::OpMetaInfo& info) {
    return info.inputs_;
  }
  static const std::vector<std::string>& GetOutputs(
      const paddle::OpMetaInfo& info) {
    return info.outputs_;
  }
  static const std::vector<std::string>& GetAttrs(
      const paddle::OpMetaInfo& info) {
    return info.attrs_;
  }
  static const std::unordered_map<std::string, std::string>& GetInplaceMap(
      const paddle::OpMetaInfo& info) {
    return info.inplace_map_;
  }
822 823 824 825
  static const std::unordered_map<std::string, std::string>&
  GetInplaceReverseMap(const paddle::OpMetaInfo& info) {
    return info.inplace_reverse_map_;
  }
826 827 828 829 830 831 832 833 834 835 836
  static const KernelFunc& GetKernelFn(const paddle::OpMetaInfo& info) {
    return info.kernel_fn_;
  }
  static const InferShapeFunc& GetInferShapeFn(const paddle::OpMetaInfo& info) {
    return info.infer_shape_fn_;
  }
  static const InferDtypeFunc& GetInferDtypeFn(const paddle::OpMetaInfo& info) {
    return info.infer_dtype_fn_;
  }
};

837 838
//////////////// Op Meta Info Map /////////////////

839
class PADDLE_API OpMetaInfoMap {
840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862
 public:
  // this function's impl should keep in header file.
  // if move to cc file, meta info can not be added
  // into map
  static OpMetaInfoMap& Instance() {
    static OpMetaInfoMap g_custom_op_meta_info_map;
    return g_custom_op_meta_info_map;
  }

  std::vector<OpMetaInfo>& operator[](const std::string& name);

  const std::unordered_map<std::string, std::vector<OpMetaInfo>>& GetMap()
      const;

 private:
  OpMetaInfoMap() = default;
  std::unordered_map<std::string, std::vector<OpMetaInfo>> map_;

  PD_DISABLE_COPY_AND_ASSIGN(OpMetaInfoMap);
};

//////////////// Op Meta Info Builder /////////////////

863
class PADDLE_API OpMetaInfoBuilder {
864
 public:
865
  explicit OpMetaInfoBuilder(std::string&& name, size_t index);
866 867
  OpMetaInfoBuilder& Inputs(std::vector<std::string>&& inputs);
  OpMetaInfoBuilder& Outputs(std::vector<std::string>&& outputs);
868
  OpMetaInfoBuilder& Attrs(std::vector<std::string>&& attrs);
869
  OpMetaInfoBuilder& SetInplaceMap(
870
      std::unordered_map<std::string, std::string>&& inplace_map);
871 872 873
  OpMetaInfoBuilder& SetKernelFn(KernelFunc func);
  OpMetaInfoBuilder& SetInferShapeFn(InferShapeFunc func);
  OpMetaInfoBuilder& SetInferDtypeFn(InferDtypeFunc func);
874 875 876 877

 private:
  // Forward Op name
  std::string name_;
878
  // ref current info ptr
879
  OpMetaInfo* info_ptr_;
880 881 882
  // The current op meta info index in vector
  // - 0: op, 1: grad_op, 2: grad_grad_op
  size_t index_;
883 884 885 886
};

/////////////////////// Op register Macro /////////////////////////

887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905
#define PD_BUILD_OP(op_name)                                                   \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                              \
      __reg_op__##op_name, "PD_BUILD_OP must be called in global namespace."); \
  static ::paddle::OpMetaInfoBuilder __op_meta_info_##op_name##__ =            \
      ::paddle::OpMetaInfoBuilder(#op_name, 0)

#define PD_BUILD_GRAD_OP(op_name)                                        \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                        \
      __reg_grad_op__##op_name,                                          \
      "PD_BUILD_GRAD_OP must be called in global namespace.");           \
  static ::paddle::OpMetaInfoBuilder __grad_op_meta_info_##op_name##__ = \
      ::paddle::OpMetaInfoBuilder(#op_name, 1)

#define PD_BUILD_DOUBLE_GRAD_OP(op_name)                                      \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
      __reg_grad_grad_op__##op_name,                                          \
      "PD_BUILD_DOUBLE_GRAD_OP must be called in global namespace.");         \
  static ::paddle::OpMetaInfoBuilder __grad_grad_op_meta_info_##op_name##__ = \
      ::paddle::OpMetaInfoBuilder(#op_name, 2)
906

907 908 909 910 911 912 913 914
}  // namespace paddle

///////////////////// C API ///////////////////

#ifdef __cplusplus
extern "C" {
#endif

915
#if defined(_WIN32)
916
// C-API to get global OpMetaInfoMap.
917 918 919 920
__declspec(dllexport) inline paddle::OpMetaInfoMap& PD_GetOpMetaInfoMap() {
  return paddle::OpMetaInfoMap::Instance();
}
#endif  // _WIN32
921 922 923 924

#ifdef __cplusplus
}
#endif