op_registry.h 16.1 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_builder.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:
Y
Yu Yang 已提交
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
  struct VariableBuilder {
    VarProto* var_;
    std::function<void()> on_multiple_;
    std::function<void()> on_temporary_;

    VariableBuilder& SetMultiple() {
      var_->set_multiple(true);
      on_multiple_();
      return *this;
    }

    VariableBuilder& SetTemporary() {
      PADDLE_ENFORCE(bool(on_temporary_), "Cannot set temporary");
      var_->set_temporary(true);
      on_temporary_();
      return *this;
    }

    VariableBuilder& IgnoreGradient() {
      var_->set_ignore_gradient(true);
      return *this;
    }
  };

  VariableBuilder AddInput(const std::string& name,
                           const std::string& comment) {
115
    auto input = proto_->mutable_inputs()->Add();
116 117
    *input->mutable_name() = name;
    *input->mutable_comment() = comment;
Y
Yu Yang 已提交
118 119
    return VariableBuilder{input, [=] { this->SetHasMultipleInput(); },
                           nullptr};
120 121
  }

Y
Yu Yang 已提交
122 123
  VariableBuilder AddOutput(const std::string& name,
                            const std::string& comment) {
124
    auto output = proto_->mutable_outputs()->Add();
125 126
    *output->mutable_name() = name;
    *output->mutable_comment() = comment;
Y
Yu Yang 已提交
127 128
    return VariableBuilder{output, [=] { this->SetHasMultipleOutput(); },
                           [=] { this->SetHasTemporaryOutput(); }};
129 130 131 132
  }

  template <typename T>
  TypedAttrChecker<T>& AddAttr(const std::string& name,
133 134
                               const std::string& comment,
                               bool generated = false) {
135
    auto attr = proto_->mutable_attrs()->Add();
136 137
    *attr->mutable_name() = name;
    *attr->mutable_comment() = comment;
138
    attr->set_generated(generated);
139 140 141 142 143 144 145 146
    AttrTypeHelper::SetAttrType<T>(attr);
    return op_checker_->AddAttrChecker<T>(name);
  }

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

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
 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;
    }
  }

195
  void CheckNoDuplicatedInOutAttrs() {
196
    std::unordered_set<std::string> names;
197 198 199 200
    auto checker = [&](const std::string& name) {
      PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
      names.insert(name);
    };
201
    for (auto& attr : proto_->attrs()) {
202 203 204 205 206 207 208
      checker(attr.name());
    }
    for (auto& input : proto_->inputs()) {
      checker(input.name());
    }
    for (auto& output : proto_->outputs()) {
      checker(output.name());
209 210 211
    }
  }

212 213
  OpProto* proto_;
  OpAttrChecker* op_checker_;
214
  bool validated_{false};
215 216 217
  bool has_multiple_input_{false};
  bool has_multiple_output_{false};
  bool has_temporary_output_{false};
218 219 220
};

class OpRegistry {
Q
Qiao Longfei 已提交
221
  using OpCreator = std::function<OperatorBase*()>;
Y
Yu Yang 已提交
222
  using VarIndexMap = std::unordered_map<std::string, int>;
Y
Yu Yang 已提交
223
  using VarNameList = std::vector<std::string>;
224 225 226 227

 public:
  template <typename OpType, typename ProtoMakerType>
  static void RegisterOp(const std::string& op_type) {
228
    op_creators()[op_type] = [] { return new OpType; };
229
    OpAttrChecker& op_checker = op_checkers()[op_type];
D
dongzhihong 已提交
230
    OpProto& op_proto = protos()[op_type];
231 232
    auto maker = ProtoMakerType(&op_proto, &op_checker);
    maker.Validate();
Y
Yu Yang 已提交
233 234 235 236 237
    *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
238 239 240 241 242 243 244 245 246 247 248

    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++;
    }
249 250
  }

251 252 253 254 255
  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
256 257
  }

Y
Yu Yang 已提交
258 259 260 261
  static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
                                                const VarNameList& inputs,
                                                const VarNameList& outputs,
                                                const AttributeMap& attrs) {
262 263
    auto op_create_it = op_creators().find(type);
    PADDLE_ENFORCE(op_create_it != op_creators().end(),
F
fengjiayi 已提交
264
                   "Operator %s cannot be found.", type);
265

Y
Yu Yang 已提交
266 267 268 269
    auto op = op_create_it->second();
    op->type_ = type;
    op->inputs_ = inputs;
    op->outputs_ = outputs;
F
fengjiayi 已提交
270

Y
Yu Yang 已提交
271 272
    op->attrs_ = attrs;
    op_checkers().at(type).Check(op->attrs_);
273

Y
Yu Yang 已提交
274
    GenerateTempVariableName(op);
275

Y
Yu Yang 已提交
276
    {
Y
Yu Yang 已提交
277
      auto var_index_it = VarIndexMaps().find(type);
Y
Yu Yang 已提交
278 279 280 281
      if (var_index_it != VarIndexMaps().end()) {
        op->in_out_idxs_ = var_index_it->second;
      }
    }
Y
Yu Yang 已提交
282

Q
Qiao Longfei 已提交
283
    op->Init();
Y
Yu Yang 已提交
284
    return std::shared_ptr<OperatorBase>(op);
285 286
  }

Y
Yu Yang 已提交
287
  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
Y
Yu Yang 已提交
288 289
    std::vector<std::string> inputs;
    inputs.reserve((size_t)op_desc.inputs_size());
290
    std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
Y
Yu Yang 已提交
291 292 293 294
              std::back_inserter(inputs));

    std::vector<std::string> outputs;
    outputs.reserve((size_t)op_desc.outputs_size());
295
    std::copy(op_desc.outputs().begin(), op_desc.outputs().end(),
Y
Yu Yang 已提交
296 297 298
              std::back_inserter(outputs));

    AttributeMap attrs;
299
    for (auto& attr : op_desc.attrs()) {
Y
Yu Yang 已提交
300
      attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
301
    }
Y
Yu Yang 已提交
302 303

    return CreateOp(op_desc.type(), inputs, outputs, attrs);
304 305
  }

Y
Yu Yang 已提交
306 307
  static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
    PADDLE_ENFORCE(!op.IsNetOp(),
Y
Yu Yang 已提交
308
                   "Use framework::Backward to get backward ops");
309
    std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
F
fengjiayi 已提交
310 311
    grad_op->Init();
    return grad_op;
D
dongzhihong 已提交
312 313
  }

Y
Yu Yang 已提交
314 315 316 317 318
  static std::unordered_map<std::string, OpProto>& protos() {
    static std::unordered_map<std::string, OpProto> protos_;
    return protos_;
  };

319 320 321
  static std::unordered_map<std::string, std::string>& grad_ops() {
    static std::unordered_map<std::string, std::string> grad_ops_;
    return grad_ops_;
322 323
  }

Y
Yu Yang 已提交
324 325 326 327 328 329
  static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>>&
  VarIndexMaps() {
    static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>> maps_;
    return maps_;
  }

330 331 332
  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
333 334
  }

335
 private:
F
fengjiayi 已提交
336 337 338 339 340
  static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
    static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
    return op_checkers_;
  };

341
  static void GenerateTempVariableName(OperatorBase* op) {
342 343 344
    static std::atomic<size_t> gUniqId(0UL);
    for (auto& outname : op->outputs_) {
      if (outname == OperatorBase::TMP_VAR_NAME()) {
345
        outname += op->type_;
346 347 348 349 350
        outname += "@";
        outname += std::to_string(gUniqId.fetch_add(1));
      }
    }
  }
351
};
352 353 354 355

template <typename OpType, typename ProtoMakerType>
class OpRegisterHelper {
 public:
Y
Yu Yang 已提交
356
  OpRegisterHelper(const char* op_type) {
357 358 359 360
    OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
  }
};

361
template <typename GradOpType>
D
dongzhihong 已提交
362 363
class GradOpRegisterHelper {
 public:
364 365
  GradOpRegisterHelper(const char* op_type, const char* grad_op_type) {
    OpRegistry::RegisterGradOp<GradOpType>(op_type, grad_op_type);
D
dongzhihong 已提交
366 367 368
  }
};

369 370 371
/**
 * check if MACRO is used in GLOBAL NAMESPACE.
 */
Y
Yu Yang 已提交
372 373 374 375 376 377
#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)

378 379 380
/**
 * Macro to Register Operator.
 */
Y
Yu Yang 已提交
381 382 383 384 385 386 387
#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
388
/**
F
fengjiayi 已提交
389
 * Macro to Register Gradient Operator.
D
dongzhihong 已提交
390
 */
391 392 393 394 395 396 397 398 399 400
#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
401

D
dongzhihong 已提交
402 403 404 405 406 407 408 409
/**
 * Macro to Forbid user register Gradient Operator.
 */
#define NO_GRADIENT(__op_type)                          \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                       \
      __reg_gradient_op__##__op_type##__op_type##_grad, \
      "NO_GRADIENT must be in global namespace")

410 411 412
/**
 * Macro to Register OperatorKernel.
 */
413
#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...)             \
Y
Yu Yang 已提交
414
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
415
      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
Y
Yu Yang 已提交
416
      "REGISTER_OP_KERNEL must be in global namespace");                  \
Q
qijun 已提交
417 418
  struct __op_kernel_register__##type##__##DEVICE_TYPE##__ {              \
    __op_kernel_register__##type##__##DEVICE_TYPE##__() {                 \
Y
Yu Yang 已提交
419 420 421
      ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
      key.place_ = PlaceType();                                           \
      ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
422
          .reset(new __VA_ARGS__());                                      \
Y
Yu Yang 已提交
423 424
    }                                                                     \
  };                                                                      \
Q
qijun 已提交
425 426
  static __op_kernel_register__##type##__##DEVICE_TYPE##__                \
      __reg_kernel_##type##__##DEVICE_TYPE##__;                           \
427
  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
Y
Yu Yang 已提交
428

429 430 431
// (type, KernelType)
#define REGISTER_OP_GPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
432

433 434 435
// (type, KernelType)
#define REGISTER_OP_CPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
436

437 438 439 440
/**
 * 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
441 442 443 444 445 446 447 448
#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
449 450 451 452 453 454 455 456
#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
457

458 459
// use Operator with only cpu kernel.
#define USE_OP_CPU(op_type)       \
Y
Yu Yang 已提交
460
  USE_OP_WITHOUT_KERNEL(op_type); \
461
  USE_OP_KERNEL(op_type, CPU)
Y
Yu Yang 已提交
462

463 464
#ifdef PADDLE_ONLY_CPU
#define USE_OP(op_type) USE_OP_CPU(op_type)
Y
Yu Yang 已提交
465
#else
466 467
#define USE_OP(op_type) \
  USE_OP_CPU(op_type);  \
Y
Yu Yang 已提交
468 469
  USE_OP_KERNEL(op_type, GPU)
#endif
470 471 472

}  // namespace framework
}  // namespace paddle