op_registry.h 14.7 KB
Newer Older
1 2
#pragma once

3
#include <algorithm>
4
#include <atomic>
Y
Yu Yang 已提交
5
#include <type_traits>
6 7
#include <unordered_map>
#include <unordered_set>
Q
Qiao Longfei 已提交
8
#include "paddle/framework/attr_checker.h"
9 10
#include "paddle/framework/op_desc.pb.h"
#include "paddle/framework/op_proto.pb.h"
Q
Qiao Longfei 已提交
11
#include "paddle/framework/operator.h"
D
dongzhihong 已提交
12
#include "paddle/framework/scope.h"
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

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) {}

66 67 68 69 70 71 72 73
  ~OpProtoAndCheckerMaker() {
    PADDLE_ENFORCE(validated_, "should call Validate after build");
  }

  void Validate() {
    validated_ = true;
    CheckNoDuplicatedInOutAttrs();
  }
74

75
 protected:
76 77
  void AddInput(const std::string& name, const std::string& comment,
                bool multiple = false) {
78
    auto input = proto_->mutable_inputs()->Add();
79 80
    *input->mutable_name() = name;
    *input->mutable_comment() = comment;
81 82 83 84 85 86 87 88
    input->set_multiple(multiple);
    if (multiple) {
      SetHasMultipleInput();
    }
  }

  void AddInputs(const std::string& name, const std::string& comment) {
    AddInput(name, comment, true);
89 90
  }

91 92
  void AddOutput(const std::string& name, const std::string& comment,
                 bool temporary = false, bool multiple = false) {
93
    auto output = proto_->mutable_outputs()->Add();
94 95
    *output->mutable_name() = name;
    *output->mutable_comment() = comment;
96 97 98 99 100 101 102 103 104 105 106 107 108
    output->set_multiple(multiple);
    if (multiple) {
      SetHasMultipleOutput();
    }
    output->set_temporary(temporary);
    if (temporary) {
      SetHasTemporaryOutput();
    }
  }

  void AddOutputs(const std::string& name, const std::string& comment,
                  bool temporary = false) {
    AddOutput(name, comment, temporary, true);
109 110 111 112
  }

  template <typename T>
  TypedAttrChecker<T>& AddAttr(const std::string& name,
113 114
                               const std::string& comment,
                               bool generated = false) {
115
    auto attr = proto_->mutable_attrs()->Add();
116 117
    *attr->mutable_name() = name;
    *attr->mutable_comment() = comment;
118
    attr->set_generated(generated);
119 120 121 122 123 124 125 126
    AttrTypeHelper::SetAttrType<T>(attr);
    return op_checker_->AddAttrChecker<T>(name);
  }

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

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
 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;
    }
  }

175
  void CheckNoDuplicatedInOutAttrs() {
176
    std::unordered_set<std::string> names;
177 178 179 180
    auto checker = [&](const std::string& name) {
      PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
      names.insert(name);
    };
181
    for (auto& attr : proto_->attrs()) {
182 183 184 185 186 187 188
      checker(attr.name());
    }
    for (auto& input : proto_->inputs()) {
      checker(input.name());
    }
    for (auto& output : proto_->outputs()) {
      checker(output.name());
189 190 191
    }
  }

192 193
  OpProto* proto_;
  OpAttrChecker* op_checker_;
194
  bool validated_{false};
195 196 197
  bool has_multiple_input_{false};
  bool has_multiple_output_{false};
  bool has_temporary_output_{false};
198 199 200
};

class OpRegistry {
Q
Qiao Longfei 已提交
201
  using OpCreator = std::function<OperatorBase*()>;
Y
Yu Yang 已提交
202
  using VarIndexMap = std::unordered_map<std::string, int>;
Y
Yu Yang 已提交
203
  using VarNameList = std::vector<std::string>;
204 205 206 207

 public:
  template <typename OpType, typename ProtoMakerType>
  static void RegisterOp(const std::string& op_type) {
208 209
    creators()[op_type] = [] { return new OpType; };
    OpAttrChecker& op_checker = op_checkers()[op_type];
D
dongzhihong 已提交
210
    OpProto& op_proto = protos()[op_type];
211 212
    auto maker = ProtoMakerType(&op_proto, &op_checker);
    maker.Validate();
Y
Yu Yang 已提交
213 214 215 216 217
    *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
218 219 220 221 222 223 224 225 226 227 228

    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++;
    }
229 230
  }

Y
Yu Yang 已提交
231 232 233 234 235 236
  static OperatorPtr CreateOp(const std::string& type,
                              const VarNameList& inputs,
                              const VarNameList& outputs,
                              const AttributeMap& attrs) {
    auto op_create_it = creators().find(type);
    PADDLE_ENFORCE(op_create_it != creators().end(),
F
fengjiayi 已提交
237
                   "Operator %s cannot be found.", type);
238

Y
Yu Yang 已提交
239 240 241 242 243 244
    auto op = op_create_it->second();
    op->type_ = type;
    op->inputs_ = inputs;
    op->outputs_ = outputs;
    op->attrs_ = attrs;
    op_checkers().at(type).Check(op->attrs_);
245

Y
Yu Yang 已提交
246
    GenerateTempVariableName(op);
247

Y
Yu Yang 已提交
248
    {
Y
Yu Yang 已提交
249
      auto var_index_it = VarIndexMaps().find(type);
Y
Yu Yang 已提交
250 251 252 253
      if (var_index_it != VarIndexMaps().end()) {
        op->in_out_idxs_ = var_index_it->second;
      }
    }
Y
Yu Yang 已提交
254

Q
Qiao Longfei 已提交
255
    op->Init();
Y
Yu Yang 已提交
256
    return OperatorPtr(op);
257 258
  }

D
dongzhihong 已提交
259 260 261 262 263
  template <typename OpType>
  static void RegisterGradOp(const std::string& op_type) {
    grad_creators()[op_type] = [] { return new OpType; };
  }

Q
Qiao Longfei 已提交
264
  static OperatorPtr CreateOp(const OpDesc& op_desc) {
Y
Yu Yang 已提交
265 266
    std::vector<std::string> inputs;
    inputs.reserve((size_t)op_desc.inputs_size());
267
    std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
Y
Yu Yang 已提交
268 269 270 271
              std::back_inserter(inputs));

    std::vector<std::string> outputs;
    outputs.reserve((size_t)op_desc.outputs_size());
272
    std::copy(op_desc.outputs().begin(), op_desc.outputs().end(),
Y
Yu Yang 已提交
273 274 275
              std::back_inserter(outputs));

    AttributeMap attrs;
276
    for (auto& attr : op_desc.attrs()) {
Y
Yu Yang 已提交
277
      attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
278
    }
Y
Yu Yang 已提交
279 280

    return CreateOp(op_desc.type(), inputs, outputs, attrs);
281 282
  }

D
dongzhihong 已提交
283 284 285 286 287 288 289 290 291 292 293 294 295 296 297
  static OperatorPtr CreateGradOp(std::shared_ptr<OperatorBase> op) {
    OperatorPtr op_grad(grad_creators().at(op->type_)());
    op_grad->type_ = op->type_;
    op_grad->inputs_.reserve(op->inputs_.size());
    for (auto& input : op->inputs_) {
      op_grad->inputs_.emplace_back(input);
      op_grad->outputs_.emplace_back(input + "@grad");
    }
    for (auto& output : op->outputs_) {
      op_grad->inputs_.emplace_back(output);
      op_grad->inputs_.emplace_back(output + "@grad");
    }
    return op_grad;
  }

Y
Yu Yang 已提交
298 299 300 301 302
  static std::unordered_map<std::string, OpProto>& protos() {
    static std::unordered_map<std::string, OpProto> protos_;
    return protos_;
  };

303
 private:
Y
Yu Yang 已提交
304 305 306 307 308 309
  static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>>&
  VarIndexMaps() {
    static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>> maps_;
    return maps_;
  }

310
  static void GenerateTempVariableName(OperatorBase* op) {
311 312 313
    static std::atomic<size_t> gUniqId(0UL);
    for (auto& outname : op->outputs_) {
      if (outname == OperatorBase::TMP_VAR_NAME()) {
314
        outname += op->type_;
315 316 317 318 319 320
        outname += "@";
        outname += std::to_string(gUniqId.fetch_add(1));
      }
    }
  }

321 322 323 324
  static std::unordered_map<std::string, OpCreator>& creators() {
    static std::unordered_map<std::string, OpCreator> creators_;
    return creators_;
  }
325

326 327 328 329
  static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
    static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
    return op_checkers_;
  };
D
dongzhihong 已提交
330 331 332 333 334

  static std::unordered_map<std::string, OpCreator>& grad_creators() {
    static std::unordered_map<std::string, OpCreator> grad_creators_;
    return grad_creators_;
  }
335
};
336 337 338 339

template <typename OpType, typename ProtoMakerType>
class OpRegisterHelper {
 public:
Y
Yu Yang 已提交
340
  OpRegisterHelper(const char* op_type) {
341 342 343 344
    OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
  }
};

D
dongzhihong 已提交
345 346 347 348 349 350 351 352
template <typename OpType>
class GradOpRegisterHelper {
 public:
  GradOpRegisterHelper(const char* op_type) {
    OpRegistry::RegisterGradOp<OpType>(op_type);
  }
};

353 354 355
/**
 * check if MACRO is used in GLOBAL NAMESPACE.
 */
Y
Yu Yang 已提交
356 357 358 359 360 361
#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)

362 363 364
/**
 * Macro to Register Operator.
 */
Y
Yu Yang 已提交
365 366 367 368 369 370 371
#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
372 373 374 375 376 377 378 379 380 381 382
/**
 * Macro to Register Operator.
 */
#define REGISTER_GRADIENT_OP(__op_type, __op_class)            \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                              \
      __reg_op__##__op_type,                                   \
      "REGISTER_GRADIENT_OP must be in global namespace");     \
  static ::paddle::framework::GradOpRegisterHelper<__op_class> \
      __op_register_##__op_type##__(#__op_type);               \
  int __op_register_##__op_type##_handle__() { return 0; }

383 384 385
/**
 * Macro to Register OperatorKernel.
 */
386
#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...)             \
Y
Yu Yang 已提交
387
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
388
      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
Y
Yu Yang 已提交
389 390 391 392 393 394
      "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] \
395
          .reset(new __VA_ARGS__());                                      \
Y
Yu Yang 已提交
396 397 398
    }                                                                     \
  };                                                                      \
  static __op_kernel_register__##type##__ __reg_kernel_##type##__;        \
399
  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
Y
Yu Yang 已提交
400

401 402 403
// (type, KernelType)
#define REGISTER_OP_GPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
404

405 406 407
// (type, KernelType)
#define REGISTER_OP_CPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
408

409 410 411 412
/**
 * 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
413 414 415 416 417 418 419 420
#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
421 422 423 424 425 426 427 428
#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
429

430 431
// use Operator with only cpu kernel.
#define USE_OP_CPU(op_type)       \
Y
Yu Yang 已提交
432
  USE_OP_WITHOUT_KERNEL(op_type); \
433
  USE_OP_KERNEL(op_type, CPU)
Y
Yu Yang 已提交
434

435 436
#ifdef PADDLE_ONLY_CPU
#define USE_OP(op_type) USE_OP_CPU(op_type)
Y
Yu Yang 已提交
437
#else
438 439
#define USE_OP(op_type) \
  USE_OP_CPU(op_type);  \
Y
Yu Yang 已提交
440 441
  USE_OP_KERNEL(op_type, GPU)
#endif
442 443 444

}  // namespace framework
}  // namespace paddle