op_registry.h 15.8 KB
Newer Older
F
fengjiayi 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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. */

15 16
#pragma once

17
#include <algorithm>
18
#include <atomic>
Y
Yu Yang 已提交
19
#include <type_traits>
20 21
#include <unordered_map>
#include <unordered_set>
Q
Qiao Longfei 已提交
22
#include "paddle/framework/attr_checker.h"
F
fengjiayi 已提交
23
#include "paddle/framework/grad_op_creator.h"
24
#include "paddle/framework/op_desc.pb.h"
D
dongzhihong 已提交
25
#include "paddle/framework/scope.h"
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

namespace paddle {
namespace framework {

// helper class to set attribute type
struct AttrTypeHelper {
  template <typename T>
  static void SetAttrType(AttrProto* attr);

  static Attribute GetAttrValue(const AttrDesc& attr_desc) {
    switch (attr_desc.type()) {
      case paddle::framework::AttrType::INT: {
        return attr_desc.i();
      }
      case paddle::framework::AttrType::FLOAT: {
        return attr_desc.f();
      }
      case paddle::framework::AttrType::STRING: {
        return attr_desc.s();
      }
      case paddle::framework::AttrType::INTS: {
        std::vector<int> val(attr_desc.ints_size());
        for (int i = 0; i < attr_desc.ints_size(); ++i) {
          val[i] = attr_desc.ints(i);
        }
        return val;
      }
      case paddle::framework::AttrType::FLOATS: {
        std::vector<float> val(attr_desc.floats_size());
        for (int i = 0; i < attr_desc.floats_size(); ++i) {
          val[i] = attr_desc.floats(i);
        }
        return val;
      }
      case paddle::framework::AttrType::STRINGS: {
        std::vector<std::string> val(attr_desc.strings_size());
        for (int i = 0; i < attr_desc.strings_size(); ++i) {
          val[i] = attr_desc.strings(i);
        }
        return val;
      }
    }
    PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
    return boost::blank();
  }
};

// this class not only make proto but also init attribute checkers.
class OpProtoAndCheckerMaker {
 public:
  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
      : proto_(proto), op_checker_(op_checker) {}

79 80 81 82 83 84 85 86
  ~OpProtoAndCheckerMaker() {
    PADDLE_ENFORCE(validated_, "should call Validate after build");
  }

  void Validate() {
    validated_ = true;
    CheckNoDuplicatedInOutAttrs();
  }
87

88
 protected:
89
  void AddInput(const std::string& name, const std::string& comment,
90
                bool multiple = false, bool ignore_gradient = false) {
91
    auto input = proto_->mutable_inputs()->Add();
92 93
    *input->mutable_name() = name;
    *input->mutable_comment() = comment;
D
dongzhihong 已提交
94
    input->set_ignore_gradient(ignore_gradient);
95 96 97 98 99 100
    input->set_multiple(multiple);
    if (multiple) {
      SetHasMultipleInput();
    }
  }

101 102 103
  void AddInputs(const std::string& name, const std::string& comment,
                 bool ignore_gradient = false) {
    AddInput(name, comment, true, ignore_gradient);
104 105
  }

106
  void AddOutput(const std::string& name, const std::string& comment,
107 108
                 bool temporary = false, bool multiple = false,
                 bool ignore_gradient = false) {
109
    auto output = proto_->mutable_outputs()->Add();
110 111
    *output->mutable_name() = name;
    *output->mutable_comment() = comment;
D
dongzhihong 已提交
112
    output->set_ignore_gradient(ignore_gradient);
113 114 115 116 117 118 119 120 121 122 123
    output->set_multiple(multiple);
    if (multiple) {
      SetHasMultipleOutput();
    }
    output->set_temporary(temporary);
    if (temporary) {
      SetHasTemporaryOutput();
    }
  }

  void AddOutputs(const std::string& name, const std::string& comment,
124 125
                  bool temporary = false, bool ignore_gradient = false) {
    AddOutput(name, comment, temporary, true, ignore_gradient);
126 127 128 129
  }

  template <typename T>
  TypedAttrChecker<T>& AddAttr(const std::string& name,
130 131
                               const std::string& comment,
                               bool generated = false) {
132
    auto attr = proto_->mutable_attrs()->Add();
133 134
    *attr->mutable_name() = name;
    *attr->mutable_comment() = comment;
135
    attr->set_generated(generated);
136 137 138 139 140 141 142 143
    AttrTypeHelper::SetAttrType<T>(attr);
    return op_checker_->AddAttrChecker<T>(name);
  }

  void AddComment(const std::string& comment) {
    *(proto_->mutable_comment()) = comment;
  }

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
 private:
  void SetHasMultiple(const std::string& in_out, bool* flag) {
    if (!*flag) {
      AddAttr<std::vector<int>>(in_out + "_format",
                                "The multiple index of " + in_out +
                                    "\n"
                                    R"DOC(
This attribute is used by Paddle core framework. Paddle's Op support each input
or output could be a list of variable. This attribute is used to show how that
list organized.

e.g.
  input = ["a", "b", "c", "d", "e", "f"]
  input_format = [0, 4, 5, 6]

means
  The number of all input variables this op is six, and they are segmented into
  three inputs.

  The first input is input[0:4], second is input[4:5], third is input[5:6].
)DOC",
                                /*generated*/ true);
      *flag = true;
    }
  }

  void SetHasMultipleInput() { SetHasMultiple("input", &has_multiple_input_); }
  void SetHasMultipleOutput() {
    SetHasMultiple("output", &has_multiple_output_);
  }

  void SetHasTemporaryOutput() {
    if (!has_temporary_output_) {
      AddAttr<std::vector<int>>("temporary_index",
                                R"DOC(The temporary index of output.

Not all output of Paddle Op is used by user. For faster computation, each op
could output some its internal state to other op, other op could take that
output to make compute faster.

Add a mark to which output is temporary is helpful for future optimization.
)DOC",
                                /*generated*/ true)
          .SetDefault(std::vector<int>());
      has_temporary_output_ = true;
    }
  }

192
  void CheckNoDuplicatedInOutAttrs() {
193
    std::unordered_set<std::string> names;
194 195 196 197
    auto checker = [&](const std::string& name) {
      PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
      names.insert(name);
    };
198
    for (auto& attr : proto_->attrs()) {
199 200 201 202 203 204 205
      checker(attr.name());
    }
    for (auto& input : proto_->inputs()) {
      checker(input.name());
    }
    for (auto& output : proto_->outputs()) {
      checker(output.name());
206 207 208
    }
  }

209 210
  OpProto* proto_;
  OpAttrChecker* op_checker_;
211
  bool validated_{false};
212 213 214
  bool has_multiple_input_{false};
  bool has_multiple_output_{false};
  bool has_temporary_output_{false};
215 216 217
};

class OpRegistry {
Q
Qiao Longfei 已提交
218
  using OpCreator = std::function<OperatorBase*()>;
Y
Yu Yang 已提交
219
  using VarIndexMap = std::unordered_map<std::string, int>;
Y
Yu Yang 已提交
220
  using VarNameList = std::vector<std::string>;
221 222 223 224

 public:
  template <typename OpType, typename ProtoMakerType>
  static void RegisterOp(const std::string& op_type) {
225
    op_creators()[op_type] = [] { return new OpType; };
226
    OpAttrChecker& op_checker = op_checkers()[op_type];
D
dongzhihong 已提交
227
    OpProto& op_proto = protos()[op_type];
228 229
    auto maker = ProtoMakerType(&op_proto, &op_checker);
    maker.Validate();
Y
Yu Yang 已提交
230 231 232 233 234
    *op_proto.mutable_type() = op_type;
    PADDLE_ENFORCE(
        op_proto.IsInitialized(),
        "Fail to initialize %s's OpProto, because %s is not initialized",
        op_type, op_proto.InitializationErrorString());
Y
Yu Yang 已提交
235 236 237 238 239 240 241 242 243 244 245

    VarIndexMaps()[op_type].reset(new VarIndexMap());
    auto& varmap = *VarIndexMaps()[op_type];
    int idx = 0;
    for (auto& var : op_proto.inputs()) {
      varmap[var.name()] = idx++;
    }
    idx = 0;
    for (auto& var : op_proto.outputs()) {
      varmap[var.name()] = idx++;
    }
246 247
  }

248 249 250 251 252
  template <typename GradOpType>
  static void RegisterGradOp(const std::string& op_type,
                             const std::string& grad_op_type) {
    op_creators()[grad_op_type] = [] { return new GradOpType; };
    grad_ops()[op_type] = grad_op_type;
F
fengjiayi 已提交
253 254
  }

Y
Yu Yang 已提交
255 256 257 258
  static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
                                                const VarNameList& inputs,
                                                const VarNameList& outputs,
                                                const AttributeMap& attrs) {
259 260
    auto op_create_it = op_creators().find(type);
    PADDLE_ENFORCE(op_create_it != op_creators().end(),
F
fengjiayi 已提交
261
                   "Operator %s cannot be found.", type);
262

Y
Yu Yang 已提交
263 264 265 266
    auto op = op_create_it->second();
    op->type_ = type;
    op->inputs_ = inputs;
    op->outputs_ = outputs;
F
fengjiayi 已提交
267

Y
Yu Yang 已提交
268 269
    op->attrs_ = attrs;
    op_checkers().at(type).Check(op->attrs_);
270

Y
Yu Yang 已提交
271
    GenerateTempVariableName(op);
272

Y
Yu Yang 已提交
273
    {
Y
Yu Yang 已提交
274
      auto var_index_it = VarIndexMaps().find(type);
Y
Yu Yang 已提交
275 276 277 278
      if (var_index_it != VarIndexMaps().end()) {
        op->in_out_idxs_ = var_index_it->second;
      }
    }
Y
Yu Yang 已提交
279

Q
Qiao Longfei 已提交
280
    op->Init();
Y
Yu Yang 已提交
281
    return std::shared_ptr<OperatorBase>(op);
282 283
  }

Y
Yu Yang 已提交
284
  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
Y
Yu Yang 已提交
285 286
    std::vector<std::string> inputs;
    inputs.reserve((size_t)op_desc.inputs_size());
287
    std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
Y
Yu Yang 已提交
288 289 290 291
              std::back_inserter(inputs));

    std::vector<std::string> outputs;
    outputs.reserve((size_t)op_desc.outputs_size());
292
    std::copy(op_desc.outputs().begin(), op_desc.outputs().end(),
Y
Yu Yang 已提交
293 294 295
              std::back_inserter(outputs));

    AttributeMap attrs;
296
    for (auto& attr : op_desc.attrs()) {
Y
Yu Yang 已提交
297
      attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
298
    }
Y
Yu Yang 已提交
299 300

    return CreateOp(op_desc.type(), inputs, outputs, attrs);
301 302
  }

F
fengjiayi 已提交
303 304
  static std::shared_ptr<OperatorBase> CreateGradOp(
      std::shared_ptr<OperatorBase> op) {
305 306
    GradOpBuilder builder(op.get());
    std::shared_ptr<OperatorBase> grad_op(builder.Build());
F
fengjiayi 已提交
307 308
    grad_op->Init();
    return grad_op;
D
dongzhihong 已提交
309 310
  }

Y
Yu Yang 已提交
311 312 313 314 315
  static std::unordered_map<std::string, OpProto>& protos() {
    static std::unordered_map<std::string, OpProto> protos_;
    return protos_;
  };

316 317 318
  static std::unordered_map<std::string, std::string>& grad_ops() {
    static std::unordered_map<std::string, std::string> grad_ops_;
    return grad_ops_;
319 320
  }

Y
Yu Yang 已提交
321 322 323 324 325 326
  static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>>&
  VarIndexMaps() {
    static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>> maps_;
    return maps_;
  }

327 328 329
  static std::unordered_map<std::string, OpCreator>& op_creators() {
    static std::unordered_map<std::string, OpCreator> op_creators_;
    return op_creators_;
F
fengjiayi 已提交
330 331
  }

332
 private:
F
fengjiayi 已提交
333 334 335 336 337
  static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
    static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
    return op_checkers_;
  };

338
  static void GenerateTempVariableName(OperatorBase* op) {
339 340 341
    static std::atomic<size_t> gUniqId(0UL);
    for (auto& outname : op->outputs_) {
      if (outname == OperatorBase::TMP_VAR_NAME()) {
342
        outname += op->type_;
343 344 345 346 347
        outname += "@";
        outname += std::to_string(gUniqId.fetch_add(1));
      }
    }
  }
348
};
349 350 351 352

template <typename OpType, typename ProtoMakerType>
class OpRegisterHelper {
 public:
Y
Yu Yang 已提交
353
  OpRegisterHelper(const char* op_type) {
354 355 356 357
    OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
  }
};

358
template <typename GradOpType>
D
dongzhihong 已提交
359 360
class GradOpRegisterHelper {
 public:
361 362
  GradOpRegisterHelper(const char* op_type, const char* grad_op_type) {
    OpRegistry::RegisterGradOp<GradOpType>(op_type, grad_op_type);
D
dongzhihong 已提交
363 364 365
  }
};

366 367 368
/**
 * check if MACRO is used in GLOBAL NAMESPACE.
 */
Y
Yu Yang 已提交
369 370 371 372 373 374
#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)

375 376 377
/**
 * Macro to Register Operator.
 */
Y
Yu Yang 已提交
378 379 380 381 382 383 384
#define REGISTER_OP(__op_type, __op_class, __op_maker_class)                 \
  STATIC_ASSERT_GLOBAL_NAMESPACE(__reg_op__##__op_type,                      \
                                 "REGISTER_OP must be in global namespace"); \
  static ::paddle::framework::OpRegisterHelper<__op_class, __op_maker_class> \
      __op_register_##__op_type##__(#__op_type);                             \
  int __op_register_##__op_type##_handle__() { return 0; }

D
dongzhihong 已提交
385
/**
F
fengjiayi 已提交
386
 * Macro to Register Gradient Operator.
D
dongzhihong 已提交
387
 */
388 389 390 391 392 393 394 395 396 397
#define REGISTER_GRADIENT_OP(__op_type, __grad_op_type, __grad_op_class)       \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                              \
      __reg_gradient_op__##__op_type##__grad_op_type,                          \
      "REGISTER_GRADIENT_OP must be in global namespace");                     \
  static ::paddle::framework::GradOpRegisterHelper<__grad_op_class>            \
      __op_gradient_register_##__op_type##__grad_op_type##__(#__op_type,       \
                                                             #__grad_op_type); \
  int __op_gradient_register_##__op_type##__grad_op_type##_handle__() {        \
    return 0;                                                                  \
  }
D
dongzhihong 已提交
398

399 400 401
/**
 * Macro to Register OperatorKernel.
 */
402
#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...)             \
Y
Yu Yang 已提交
403
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
404
      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
Y
Yu Yang 已提交
405 406 407 408 409 410
      "REGISTER_OP_KERNEL must be in global namespace");                  \
  struct __op_kernel_register__##type##__ {                               \
    __op_kernel_register__##type##__() {                                  \
      ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
      key.place_ = PlaceType();                                           \
      ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
411
          .reset(new __VA_ARGS__());                                      \
Y
Yu Yang 已提交
412 413 414
    }                                                                     \
  };                                                                      \
  static __op_kernel_register__##type##__ __reg_kernel_##type##__;        \
415
  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
Y
Yu Yang 已提交
416

417 418 419
// (type, KernelType)
#define REGISTER_OP_GPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
420

421 422 423
// (type, KernelType)
#define REGISTER_OP_CPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
424

425 426 427 428
/**
 * Macro to mark what Operator and Kernel we will use and tell the compiler to
 * link them into target.
 */
Y
Yu Yang 已提交
429 430 431 432 433 434 435 436
#define USE_OP_WITHOUT_KERNEL(op_type)                      \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                           \
      __use_op_without_kernel_##op_type,                    \
      "USE_OP_WITHOUT_KERNEL must be in global namespace"); \
  extern int __op_register_##op_type##_handle__();          \
  static int __use_op_ptr_##op_type##_without_kernel__      \
      __attribute__((unused)) = __op_register_##op_type##_handle__()

Y
Yu Yang 已提交
437 438 439 440 441 442 443 444
#define USE_OP_KERNEL(op_type, DEVICE_TYPE)                               \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
      __use_op_kernel_##op_type##_##DEVICE_TYPE##__,                      \
      "USE_OP_KERNEL must be in global namespace");                       \
  extern int __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__(); \
  static int __use_op_ptr_##op_type##_##DEVICE_TYPE##_kernel__            \
      __attribute__((unused)) =                                           \
          __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
Y
Yu Yang 已提交
445

446 447
// use Operator with only cpu kernel.
#define USE_OP_CPU(op_type)       \
Y
Yu Yang 已提交
448
  USE_OP_WITHOUT_KERNEL(op_type); \
449
  USE_OP_KERNEL(op_type, CPU)
Y
Yu Yang 已提交
450

451 452
#ifdef PADDLE_ONLY_CPU
#define USE_OP(op_type) USE_OP_CPU(op_type)
Y
Yu Yang 已提交
453
#else
454 455
#define USE_OP(op_type) \
  USE_OP_CPU(op_type);  \
Y
Yu Yang 已提交
456 457
  USE_OP_KERNEL(op_type, GPU)
#endif
458 459 460

}  // namespace framework
}  // namespace paddle