op_registry.h 18.6 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
  void AddInput(const std::string& name, const std::string& comment,
77
                bool multiple = false, bool ignore_gradient = false) {
78
    auto input = proto_->mutable_inputs()->Add();
79 80
    *input->mutable_name() = name;
    *input->mutable_comment() = comment;
D
dongzhihong 已提交
81
    input->set_ignore_gradient(ignore_gradient);
82 83 84 85 86 87
    input->set_multiple(multiple);
    if (multiple) {
      SetHasMultipleInput();
    }
  }

88 89 90
  void AddInputs(const std::string& name, const std::string& comment,
                 bool ignore_gradient = false) {
    AddInput(name, comment, true, ignore_gradient);
91 92
  }

93
  void AddOutput(const std::string& name, const std::string& comment,
94 95
                 bool temporary = false, bool multiple = false,
                 bool ignore_gradient = false) {
96
    auto output = proto_->mutable_outputs()->Add();
97 98
    *output->mutable_name() = name;
    *output->mutable_comment() = comment;
D
dongzhihong 已提交
99
    output->set_ignore_gradient(ignore_gradient);
100 101 102 103 104 105 106 107 108 109 110
    output->set_multiple(multiple);
    if (multiple) {
      SetHasMultipleOutput();
    }
    output->set_temporary(temporary);
    if (temporary) {
      SetHasTemporaryOutput();
    }
  }

  void AddOutputs(const std::string& name, const std::string& comment,
111 112
                  bool temporary = false, bool ignore_gradient = false) {
    AddOutput(name, comment, temporary, true, ignore_gradient);
113 114 115 116
  }

  template <typename T>
  TypedAttrChecker<T>& AddAttr(const std::string& name,
117 118
                               const std::string& comment,
                               bool generated = false) {
119
    auto attr = proto_->mutable_attrs()->Add();
120 121
    *attr->mutable_name() = name;
    *attr->mutable_comment() = comment;
122
    attr->set_generated(generated);
123 124 125 126 127 128 129 130
    AttrTypeHelper::SetAttrType<T>(attr);
    return op_checker_->AddAttrChecker<T>(name);
  }

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

131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
 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;
    }
  }

179
  void CheckNoDuplicatedInOutAttrs() {
180
    std::unordered_set<std::string> names;
181 182 183 184
    auto checker = [&](const std::string& name) {
      PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
      names.insert(name);
    };
185
    for (auto& attr : proto_->attrs()) {
186 187 188 189 190 191 192
      checker(attr.name());
    }
    for (auto& input : proto_->inputs()) {
      checker(input.name());
    }
    for (auto& output : proto_->outputs()) {
      checker(output.name());
193 194 195
    }
  }

196 197
  OpProto* proto_;
  OpAttrChecker* op_checker_;
198
  bool validated_{false};
199 200 201
  bool has_multiple_input_{false};
  bool has_multiple_output_{false};
  bool has_temporary_output_{false};
202 203 204
};

class OpRegistry {
Q
Qiao Longfei 已提交
205
  using OpCreator = std::function<OperatorBase*()>;
Y
Yu Yang 已提交
206
  using VarIndexMap = std::unordered_map<std::string, int>;
Y
Yu Yang 已提交
207
  using VarNameList = std::vector<std::string>;
208 209 210 211

 public:
  template <typename OpType, typename ProtoMakerType>
  static void RegisterOp(const std::string& op_type) {
212 213
    creators()[op_type] = [] { return new OpType; };
    OpAttrChecker& op_checker = op_checkers()[op_type];
D
dongzhihong 已提交
214
    OpProto& op_proto = protos()[op_type];
215 216
    auto maker = ProtoMakerType(&op_proto, &op_checker);
    maker.Validate();
Y
Yu Yang 已提交
217 218 219 220 221
    *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
222 223 224 225 226 227 228 229 230 231 232

    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++;
    }
233 234
  }

F
fengjiayi 已提交
235 236 237 238 239
  template <typename OpType>
  static void RegisterGradOp(const std::string& op_type) {
    grad_creators()[op_type] = [] { return new OpType; };
  }

Y
Yu Yang 已提交
240 241 242 243 244 245
  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
246
                   "Operator %s cannot be found.", type);
247

Y
Yu Yang 已提交
248 249 250 251
    auto op = op_create_it->second();
    op->type_ = type;
    op->inputs_ = inputs;
    op->outputs_ = outputs;
F
fengjiayi 已提交
252

Y
Yu Yang 已提交
253 254
    op->attrs_ = attrs;
    op_checkers().at(type).Check(op->attrs_);
255

Y
Yu Yang 已提交
256
    GenerateTempVariableName(op);
257

Y
Yu Yang 已提交
258
    {
Y
Yu Yang 已提交
259
      auto var_index_it = VarIndexMaps().find(type);
Y
Yu Yang 已提交
260 261 262 263
      if (var_index_it != VarIndexMaps().end()) {
        op->in_out_idxs_ = var_index_it->second;
      }
    }
Y
Yu Yang 已提交
264

Q
Qiao Longfei 已提交
265
    op->Init();
Y
Yu Yang 已提交
266
    return OperatorPtr(op);
267 268
  }

Q
Qiao Longfei 已提交
269
  static OperatorPtr CreateOp(const OpDesc& op_desc) {
Y
Yu Yang 已提交
270 271
    std::vector<std::string> inputs;
    inputs.reserve((size_t)op_desc.inputs_size());
272
    std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
Y
Yu Yang 已提交
273 274 275 276
              std::back_inserter(inputs));

    std::vector<std::string> outputs;
    outputs.reserve((size_t)op_desc.outputs_size());
277
    std::copy(op_desc.outputs().begin(), op_desc.outputs().end(),
Y
Yu Yang 已提交
278 279 280
              std::back_inserter(outputs));

    AttributeMap attrs;
281
    for (auto& attr : op_desc.attrs()) {
Y
Yu Yang 已提交
282
      attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
283
    }
Y
Yu Yang 已提交
284 285

    return CreateOp(op_desc.type(), inputs, outputs, attrs);
286 287
  }

F
fengjiayi 已提交
288 289 290 291 292 293 294 295 296 297
  static OperatorPtr CreateGradOp(OperatorPtr op) {
    OperatorPtr grad_op(grad_creators().at(op->type_)());
    grad_op->type_ = op->type_;

    AssembleGradInOut(op, grad_op);
    GenerateGradArgOffset(op, grad_op);
    GenerateGradAttr(op, grad_op);

    grad_op->Init();
    return grad_op;
D
dongzhihong 已提交
298 299
  }

Y
Yu Yang 已提交
300 301 302 303 304
  static std::unordered_map<std::string, OpProto>& protos() {
    static std::unordered_map<std::string, OpProto> protos_;
    return protos_;
  };

305
 private:
Y
Yu Yang 已提交
306 307 308 309 310 311
  static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>>&
  VarIndexMaps() {
    static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>> maps_;
    return maps_;
  }

F
fengjiayi 已提交
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326
  static std::unordered_map<std::string, OpCreator>& creators() {
    static std::unordered_map<std::string, OpCreator> creators_;
    return creators_;
  }

  static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
    static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
    return op_checkers_;
  };

  static std::unordered_map<std::string, OpCreator>& grad_creators() {
    static std::unordered_map<std::string, OpCreator> grad_creators_;
    return grad_creators_;
  }

327
  static void GenerateTempVariableName(OperatorBase* op) {
328 329 330
    static std::atomic<size_t> gUniqId(0UL);
    for (auto& outname : op->outputs_) {
      if (outname == OperatorBase::TMP_VAR_NAME()) {
331
        outname += op->type_;
332 333 334 335 336 337
        outname += "@";
        outname += std::to_string(gUniqId.fetch_add(1));
      }
    }
  }

F
fengjiayi 已提交
338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
  static void AssembleGradInOut(OperatorPtr op, OperatorPtr grad_op) {
    size_t in_sz = op->inputs_.size() + op->outputs_.size() * 2;
    grad_op->inputs_.reserve(in_sz);
    size_t out_sz = op->inputs_.size();
    grad_op->outputs_.reserve(out_sz);
    // copy op->inputs_ to grad_op->inputs_
    std::copy(op->inputs_.begin(), op->inputs_.end(),
              std::back_inserter(grad_op->inputs_));
    // copy op->outputs_ to grad_op->inputs_
    std::copy(op->outputs_.begin(), op->outputs_.end(),
              std::back_inserter(grad_op->inputs_));
    // add gradients of op->outputs_ to grad_op->inputs_
    for (const std::string& name : op->outputs_) {
      grad_op->inputs_.emplace_back(name + OperatorBase::GRAD_VAR_SUFFIX());
    }
    // add gradients of op->inputs_ to grad_op->outputs_
    for (const std::string& name : op->inputs_) {
      grad_op->outputs_.emplace_back(name + OperatorBase::GRAD_VAR_SUFFIX());
    }
357
  }
358

F
fengjiayi 已提交
359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381
  static void GenerateGradArgOffset(OperatorPtr op, OperatorPtr grad_op) {
    VarIndexMap* grad_varmap = new VarIndexMap();
    const OpProto& op_proto = protos()[op->type_];
    int idx = 0;
    // offset of op's inputs
    for (const auto& var : op_proto.inputs()) {
      (*grad_varmap)[var.name()] = idx++;
    }
    // offset of op's outputs
    for (const auto& var : op_proto.outputs()) {
      (*grad_varmap)[var.name()] = idx++;
    }
    // offset of gradients of op's output
    for (const auto& var : op_proto.outputs()) {
      (*grad_varmap)[var.name() + OperatorBase::GRAD_VAR_SUFFIX()] = idx++;
    }
    idx = 0;
    // offset of gradients of op's input
    for (const auto& var : op_proto.inputs()) {
      (*grad_varmap)[var.name() + OperatorBase::GRAD_VAR_SUFFIX()] = idx++;
    }
    grad_op->in_out_idxs_.reset(grad_varmap);
  }
D
dongzhihong 已提交
382

F
fengjiayi 已提交
383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429
  static void GenerateGradAttr(OperatorPtr op, OperatorPtr grad_op) {
    const OpProto& op_proto = protos()[op->type_];
    grad_op->attrs_ = op->attrs_;
    grad_op->attrs_.erase("input_format");
    grad_op->attrs_.erase("output_format");
    bool has_in_format = op->attrs_.count("input_format");
    bool has_out_format = op->attrs_.count("output_format");
    // grad_op's inputs_ contains op's inputs_, outputs_ and gradients of
    // outpus_. So grad_op's input_format is necessary when op has
    // either input_format or output_format.
    if (has_in_format || has_out_format) {
      std::vector<int> old_in_format;
      std::vector<int> old_out_format;
      has_in_format
          ? old_in_format = op->GetAttr<std::vector<int>>("input_format")
          : old_in_format = std::vector<int>(op_proto.inputs_size()),
            std::iota(old_in_format.begin(), old_in_format.end(), 0);
      has_out_format
          ? old_out_format = op->GetAttr<std::vector<int>>("output_format")
          : old_out_format = std::vector<int>(op_proto.outputs_size()),
            std::iota(old_out_format.begin(), old_out_format.end(), 0);

      std::vector<int> in_format;
      in_format.reserve(old_in_format.size() + old_out_format.size() * 2);
      int base = 0;
      for (const int& idx : old_in_format) {
        in_format.emplace_back(idx + base);
      }
      base += op->inputs_.size();
      for (const int& idx : old_out_format) {
        in_format.emplace_back(idx + base);
      }
      base += op->outputs_.size();
      for (const int& idx : old_in_format) {
        in_format.emplace_back(idx + base);
      }
      grad_op->attrs_["input_format"] = in_format;
      // grad_op's outputs_ contains gradients of op's inputs_. So grad_op's
      // output_format is necessary only when op has input_format.
      if (has_in_format) {
        std::vector<int> out_format;
        out_format.reserve(op_proto.inputs_size());
        std::copy(old_in_format.begin(), old_in_format.end(),
                  std::back_inserter(out_format));
        grad_op->attrs_["output_format"] = out_format;
      }
    }
D
dongzhihong 已提交
430
  }
431
};
432 433 434 435

template <typename OpType, typename ProtoMakerType>
class OpRegisterHelper {
 public:
Y
Yu Yang 已提交
436
  OpRegisterHelper(const char* op_type) {
437 438 439 440
    OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
  }
};

D
dongzhihong 已提交
441 442 443 444 445 446 447 448
template <typename OpType>
class GradOpRegisterHelper {
 public:
  GradOpRegisterHelper(const char* op_type) {
    OpRegistry::RegisterGradOp<OpType>(op_type);
  }
};

449 450 451
/**
 * check if MACRO is used in GLOBAL NAMESPACE.
 */
Y
Yu Yang 已提交
452 453 454 455 456 457
#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)

458 459 460
/**
 * Macro to Register Operator.
 */
Y
Yu Yang 已提交
461 462 463 464 465 466 467
#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
468
/**
F
fengjiayi 已提交
469
 * Macro to Register Gradient Operator.
D
dongzhihong 已提交
470 471 472
 */
#define REGISTER_GRADIENT_OP(__op_type, __op_class)            \
  STATIC_ASSERT_GLOBAL_NAMESPACE(                              \
F
fengjiayi 已提交
473
      __reg_gradient_op_##__reg_op__##__op_type,               \
D
dongzhihong 已提交
474 475 476 477 478
      "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; }

479 480 481
/**
 * Macro to Register OperatorKernel.
 */
482
#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...)             \
Y
Yu Yang 已提交
483
  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
484
      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
Y
Yu Yang 已提交
485 486 487 488 489 490
      "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] \
491
          .reset(new __VA_ARGS__());                                      \
Y
Yu Yang 已提交
492 493 494
    }                                                                     \
  };                                                                      \
  static __op_kernel_register__##type##__ __reg_kernel_##type##__;        \
495
  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
Y
Yu Yang 已提交
496

497 498 499
// (type, KernelType)
#define REGISTER_OP_GPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
500

501 502 503
// (type, KernelType)
#define REGISTER_OP_CPU_KERNEL(type, ...) \
  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
Y
Yu Yang 已提交
504

505 506 507 508
/**
 * 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
509 510 511 512 513 514 515 516
#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
517 518 519 520 521 522 523 524
#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
525

526 527
// use Operator with only cpu kernel.
#define USE_OP_CPU(op_type)       \
Y
Yu Yang 已提交
528
  USE_OP_WITHOUT_KERNEL(op_type); \
529
  USE_OP_KERNEL(op_type, CPU)
Y
Yu Yang 已提交
530

531 532
#ifdef PADDLE_ONLY_CPU
#define USE_OP(op_type) USE_OP_CPU(op_type)
Y
Yu Yang 已提交
533
#else
534 535
#define USE_OP(op_type) \
  USE_OP_CPU(op_type);  \
Y
Yu Yang 已提交
536 537
  USE_OP_KERNEL(op_type, GPU)
#endif
538 539 540

}  // namespace framework
}  // namespace paddle