dist_attr.cc 22.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
/* Copyright (c) 2022 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. */

#include <algorithm>
#include <iostream>
#include <iterator>

#include "paddle/fluid/distributed/auto_parallel/dist_attr.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/var_desc.h"

namespace paddle {
namespace distributed {
namespace auto_parallel {

std::vector<std::string> TensorDistAttr::fields_{
    "process_mesh", "dims_mapping", "batch_dim", "dynamic_dims"};

32 33 34 35 36 37 38 39 40 41 42
static inline std::vector<int64_t> get_tensor_shape(const VarDesc* tensor) {
  if (tensor == nullptr) return std::vector<int64_t>();
  switch (tensor->GetType()) {
    case framework::proto::VarType::READER:
    case framework::proto::VarType::LOD_TENSOR_ARRAY:
    case framework::proto::VarType::STEP_SCOPES:
    case framework::proto::VarType::FEED_MINIBATCH:
    case framework::proto::VarType::FETCH_LIST:
      return std::vector<int64_t>();
    default:
      return tensor->GetShape();
43 44 45
  }
}

46 47 48 49 50 51 52
TensorDistAttr::TensorDistAttr(const VarDesc& tensor) {
  VLOG(4) << "[TensorDistAttr constructor] tensor name: " << tensor.Name();
  std::vector<int64_t> tensor_shape = get_tensor_shape(&tensor);
  set_default_dims_mapping(tensor_shape);
  set_default_dynamic_dims(tensor_shape);
}

53
TensorDistAttr::TensorDistAttr(const TensorDistAttr& dist_attr) {
54
  copy_from(dist_attr);
55 56 57
}

TensorDistAttr& TensorDistAttr::operator=(const TensorDistAttr& dist_attr) {
58 59 60 61 62 63 64
  if (this == &dist_attr) return *this;
  TensorDistAttr tmp(dist_attr);
  std::swap(this->process_mesh_, tmp.process_mesh_);
  std::swap(this->dims_mapping_, tmp.dims_mapping_);
  std::swap(this->batch_dim_, tmp.batch_dim_);
  std::swap(this->dynamic_dims_, tmp.dynamic_dims_);
  std::swap(this->annotated_, tmp.annotated_);
65 66 67 68
  return *this;
}

void TensorDistAttr::copy_from(const TensorDistAttr& dist_attr) {
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
  set_process_mesh(dist_attr.process_mesh());
  set_dims_mapping(dist_attr.dims_mapping());
  set_batch_dim(dist_attr.batch_dim());
  set_dynamic_dims(dist_attr.dynamic_dims());
  set_annotated(dist_attr.annotated());
}

void TensorDistAttr::set_process_mesh(const ProcessMesh& process_mesh) {
  process_mesh_ = process_mesh;
}

void TensorDistAttr::set_dims_mapping(
    const std::vector<int64_t>& dims_mapping) {
  dims_mapping_ = dims_mapping;
}

void TensorDistAttr::set_batch_dim(int64_t batch_dim) {
86
  batch_dim_ = batch_dim;
87 88 89 90 91 92 93 94 95 96 97
}

void TensorDistAttr::set_dynamic_dims(const std::vector<bool>& dynamic_dims) {
  dynamic_dims_ = dynamic_dims;
}

void TensorDistAttr::set_annotated(
    const std::map<std::string, bool>& annotated) {
  annotated_ = annotated;
}

98 99 100 101 102 103 104 105 106 107 108
void TensorDistAttr::set_default_dims_mapping(
    const std::vector<int64_t>& tensor_shape) {
  if (tensor_shape.size() != 0) {
    dims_mapping_ = std::vector<int64_t>(tensor_shape.size(), -1);
  }
}

void TensorDistAttr::set_default_dynamic_dims(
    const std::vector<int64_t>& tensor_shape) {
  if (tensor_shape.size() != 0) {
    dynamic_dims_ = std::vector<bool>(tensor_shape.size(), false);
109 110 111
  }
}

112
void TensorDistAttr::mark_annotated(const std::string& name) {
113 114 115 116 117 118 119 120
  auto result = std::find(std::begin(fields_), std::end(fields_), name);
  if (result != std::end(fields_)) {
    annotated_[name] = true;
  }
}

bool TensorDistAttr::verify_process_mesh(
    const ProcessMesh& process_mesh) const {
121 122
  VLOG(4) << "[TensorDistAttr verify_process_mesh] "
          << process_mesh.to_string();
123 124
  if (!process_mesh_.empty()) {
    for (int64_t dim_mapping : dims_mapping_) {
125
      if (dim_mapping >= process_mesh_.ndim()) {
126 127 128 129 130 131 132 133
        return false;
      }
    }
  }
  return true;
}

bool TensorDistAttr::verify_dims_mapping(
134 135
    const std::vector<int64_t>& dims_mapping,
    const std::vector<int64_t>& tensor_shape) const {
136
  VLOG(4) << "[TensorDistAttr verify_dims_mapping] " << str_join(dims_mapping);
137
  if (dims_mapping.size() != tensor_shape.size()) {
138
    return false;
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
  }
  std::unordered_map<int64_t, int64_t> map;
  if (!process_mesh_.empty()) {
    for (int64_t i : dims_mapping) {
      if (i < -1 || i >= process_mesh_.ndim()) {
        return false;
      }
      ++map[i];
      if (i != -1 && map[i] > 1) {
        return false;
      }
    }
  } else {
    for (int64_t i : dims_mapping) {
      ++map[i];
      if (i != -1 && map[i] > 1) {
        return false;
      }
    }
  }
  return true;
}

162 163
bool TensorDistAttr::verify_batch_dim(
    int64_t dim, const std::vector<int64_t>& tensor_shape) const {
164
  VLOG(4) << "[TensorDistAttr verify_batch_dim] " << dim;
165 166
  int64_t ndim = tensor_shape.size();
  if (ndim > 0) {
167 168 169 170 171 172 173 174 175 176 177
    if (dim < 0) {
      dim = dim + ndim;
    }
    if (dim < 0 || dim >= ndim) {
      return false;
    }
  }
  return true;
}

bool TensorDistAttr::verify_dynamic_dims(
178 179
    const std::vector<bool>& dynamic_dims,
    const std::vector<int64_t>& tensor_shape) const {
180
  VLOG(4) << "[TensorDistAttr verify_dynamic_dims] " << str_join(dynamic_dims);
181
  if (dynamic_dims.size() > 0 && dynamic_dims.size() != tensor_shape.size()) {
182
    return false;
183 184 185 186 187 188
  }
  return true;
}

bool TensorDistAttr::verify_annotated(
    const std::map<std::string, bool>& annotated) const {
189
  VLOG(4) << "[TensorDistAttr verify_annotated] " << str_join(annotated);
190 191 192 193 194 195 196 197 198
  for (const auto& item : annotated) {
    auto result = std::find(std::begin(fields_), std::end(fields_), item.first);
    if (result == std::end(fields_)) {
      return false;
    }
  }
  return true;
}

199 200
bool TensorDistAttr::verify(const VarDesc* tensor) const {
  auto tensor_shape = get_tensor_shape(tensor);
201 202 203
  if (!verify_process_mesh(process_mesh_)) {
    return false;
  }
204
  if (!verify_dims_mapping(dims_mapping_, tensor_shape)) {
205 206
    return false;
  }
207
  if (!verify_batch_dim(batch_dim_, tensor_shape)) {
208 209
    return false;
  }
210
  if (!verify_dynamic_dims(dynamic_dims_, tensor_shape)) {
211 212 213 214 215 216 217 218 219 220
    return false;
  }
  if (!verify_annotated(annotated_)) {
    return false;
  }
  return true;
}

std::string TensorDistAttr::to_string() const {
  std::string dist_str;
221
  dist_str += "{process_mesh: " + process_mesh_.to_string() + ", ";
222 223 224 225 226 227 228
  dist_str += "dims_mappings: [" + str_join(dims_mapping_) + "], ";
  dist_str += "batch_dim: " + std::to_string(batch_dim_) + ", ";
  dist_str += "dynamic_dims: [" + str_join(dynamic_dims_) + "], ";
  dist_str += "annotated: [" + str_join(annotated_) + "]}";
  return dist_str;
}

229 230 231
void TensorDistAttr::from_proto(const TensorDistAttrProto& proto) {
  process_mesh_ = ProcessMesh::from_proto(proto.process_mesh());
  dims_mapping_.resize(proto.dims_mapping_size());
232
  for (int64_t i = 0; i < proto.dims_mapping_size(); ++i) {
233
    dims_mapping_[i] = proto.dims_mapping(i);
234
  }
235 236
  batch_dim_ = proto.batch_dim();
  dynamic_dims_.resize(proto.dynamic_dims_size());
237
  for (int64_t i = 0; i < proto.dynamic_dims_size(); ++i) {
238
    dynamic_dims_[i] = proto.dynamic_dims(i);
239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
  }
}

TensorDistAttrProto TensorDistAttr::to_proto() const {
  TensorDistAttrProto proto;
  proto.mutable_process_mesh()->CopyFrom(process_mesh_.to_proto());
  for (const auto& i : dims_mapping_) {
    proto.add_dims_mapping(i);
  }
  proto.set_batch_dim(batch_dim_);
  for (const auto& i : dynamic_dims_) {
    proto.add_dynamic_dims(i);
  }
  return proto;
}

255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274
std::string TensorDistAttr::serialize_to_string() {
  std::string data;
  auto proto = to_proto();
  proto.SerializeToString(&data);
  PADDLE_ENFORCE_EQ(to_proto().SerializeToString(&data),
                    true,
                    platform::errors::InvalidArgument(
                        "Failed to serialize tensor dist attr to string."));
  return data;
}

void TensorDistAttr::parse_from_string(const std::string& data) {
  TensorDistAttrProto proto;
  PADDLE_ENFORCE_EQ(proto.ParseFromString(data),
                    true,
                    platform::errors::InvalidArgument(
                        "Failed to parse tensor dist attr from string."));
  from_proto(proto);
}

275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290
bool operator==(const TensorDistAttr& lhs, const TensorDistAttr& rhs) {
  if (lhs.process_mesh() != rhs.process_mesh()) {
    return false;
  }
  if (lhs.dims_mapping() != rhs.dims_mapping()) {
    return false;
  }
  if (lhs.batch_dim() != rhs.batch_dim()) {
    return false;
  }
  if (lhs.dynamic_dims() != rhs.dynamic_dims()) {
    return false;
  }
  return true;
}

291 292 293
std::vector<std::string> OperatorDistAttr::fields_{"process_mesh",
                                                   "impl_type",
                                                   "impl_idx",
294
                                                   "is_recompute",
295 296
                                                   "execution_stream",
                                                   "scheduling_priority"};
297

298 299 300
OperatorDistAttr::OperatorDistAttr(const OpDesc& op) {
  VLOG(4) << "[OperatorDistAttr constructor] op type: " << op.Type();
  initialize(&op);
301 302 303
}

OperatorDistAttr::OperatorDistAttr(const OperatorDistAttr& dist_attr) {
304
  VLOG(4) << "[OperatorDistAttr copy constructor]";
305 306 307 308 309
  copy_from(dist_attr);
}

OperatorDistAttr& OperatorDistAttr::operator=(
    const OperatorDistAttr& dist_attr) {
310 311 312 313 314 315 316 317 318 319 320 321 322 323
  VLOG(4) << "[OperatorDistAttr assign constructor]";
  if (this == &dist_attr) return *this;
  OperatorDistAttr tmp(dist_attr);
  std::swap(this->input_dist_attrs_, tmp.input_dist_attrs_);
  std::swap(this->output_dist_attrs_, tmp.output_dist_attrs_);
  std::swap(this->process_mesh_, tmp.process_mesh_);
  std::swap(this->op_type_, tmp.op_type_);
  std::swap(this->impl_type_, tmp.impl_type_);
  std::swap(this->impl_idx_, tmp.impl_idx_);
  std::swap(this->is_recompute_, tmp.is_recompute_);
  std::swap(this->execution_stream_, tmp.execution_stream_);
  std::swap(this->annotated_, tmp.annotated_);
  // Note: Make sure all tensor dist attr has the same process_mesh
  set_process_mesh(this->process_mesh_);
324 325 326
  return *this;
}

327 328 329 330
void OperatorDistAttr::initialize(const OpDesc* op) {
  if (op == nullptr) return;
  for (std::string name : op->InputArgumentNames()) {
    VarDesc* input = op->Block()->FindVarRecursive(name);
331
    VLOG(4) << "[OperatorDistAttr create input dist attr] " << name;
332
    if (input == nullptr || op->Type() == "create_py_reader") {
333 334 335 336
      input_dist_attrs_[name] = TensorDistAttr();
    } else {
      input_dist_attrs_[name] = TensorDistAttr(*input);
    }
337
  }
338 339
  for (std::string name : op->OutputArgumentNames()) {
    VarDesc* output = op->Block()->FindVarRecursive(name);
340 341 342 343 344 345
    VLOG(4) << "[OperatorDistAttr create output dist attr] " << name;
    if (output == nullptr) {
      output_dist_attrs_[name] = TensorDistAttr();
    } else {
      output_dist_attrs_[name] = TensorDistAttr(*output);
    }
346
  }
347
  op_type_ = op->Type();
348
  impl_type_ = kDefault;
349
  impl_idx_ = 0;
350
  is_recompute_ = false;
351
  execution_stream_ = kDefault;
352
  scheduling_priority_ = 0;
353 354
}

355 356 357
void OperatorDistAttr::copy_from(const OperatorDistAttr& dist_attr) {
  set_input_dist_attrs(dist_attr.input_dist_attrs());
  set_output_dist_attrs(dist_attr.output_dist_attrs());
358
  set_process_mesh(dist_attr.process_mesh());
359
  set_op_type(dist_attr.op_type());
360 361
  set_impl_type(dist_attr.impl_type());
  set_impl_idx(dist_attr.impl_idx());
362
  set_is_recompute(dist_attr.is_recompute());
363
  set_execution_stream(dist_attr.execution_stream());
364
  set_scheduling_priority(dist_attr.scheduling_priority());
365 366 367
  set_annotated(dist_attr.annotated());
}

368 369
void OperatorDistAttr::set_input_dist_attrs(
    const std::map<std::string, TensorDistAttr>& dist_attrs) {
370 371
  for (const auto& item : dist_attrs) {
    set_input_dist_attr(item.first, item.second);
372
  }
373 374 375 376
}

void OperatorDistAttr::set_output_dist_attrs(
    const std::map<std::string, TensorDistAttr>& dist_attrs) {
377 378
  for (const auto& item : dist_attrs) {
    set_output_dist_attr(item.first, item.second);
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405
  }
}

void OperatorDistAttr::set_input_dist_attr(const std::string& name,
                                           const TensorDistAttr& dist_attr) {
  input_dist_attrs_[name] = dist_attr;
  // Make sure the process mesh of input be same as that of the op
  input_dist_attrs_[name].set_process_mesh(process_mesh_);
}

void OperatorDistAttr::set_output_dist_attr(const std::string& name,
                                            const TensorDistAttr& dist_attr) {
  output_dist_attrs_[name] = dist_attr;
  // Make sure the process mesh of output be same as that of the op
  output_dist_attrs_[name].set_process_mesh(process_mesh_);
}

void OperatorDistAttr::set_process_mesh(const ProcessMesh& process_mesh) {
  for (auto& item : input_dist_attrs_) {
    item.second.set_process_mesh(process_mesh);
  }
  for (auto& item : output_dist_attrs_) {
    item.second.set_process_mesh(process_mesh);
  }
  process_mesh_ = process_mesh;
}

406 407 408 409 410 411
void OperatorDistAttr::set_annotated(
    const std::map<std::string, bool>& annotated) {
  annotated_ = annotated;
}

void OperatorDistAttr::mark_annotated(const std::string& name) {
412 413 414 415 416 417
  auto result = std::find(std::begin(fields_), std::end(fields_), name);
  if (result != std::end(fields_)) {
    annotated_[name] = true;
  }
  if (name == "process_mesh") {
    for (auto& item : input_dist_attrs_) {
418
      item.second.mark_annotated(name);
419 420
    }
    for (auto& item : output_dist_attrs_) {
421
      item.second.mark_annotated(name);
422 423 424 425
    }
  }
}

426 427 428 429 430 431 432 433
void OperatorDistAttr::clear_annotated() {
  annotated_.clear();
  for (auto& item : input_dist_attrs_) {
    item.second.clear_annotated();
  }
  for (auto& item : output_dist_attrs_) {
    item.second.clear_annotated();
  }
434 435
}

436 437 438 439 440 441 442
const std::vector<int64_t>& OperatorDistAttr::input_dims_mapping(
    const std::string& name) const {
  return input_dist_attr(name).dims_mapping();
}

void OperatorDistAttr::set_input_dims_mapping(
    const std::string& name, const std::vector<int64_t>& dims_mapping) {
443 444
  input_dist_attrs_[name].set_dims_mapping(dims_mapping);
  input_dist_attrs_[name].set_process_mesh(process_mesh_);
445 446 447 448 449 450 451 452 453
}

const std::vector<int64_t>& OperatorDistAttr::output_dims_mapping(
    const std::string& name) {
  return output_dist_attr(name).dims_mapping();
}

void OperatorDistAttr::set_output_dims_mapping(
    const std::string& name, const std::vector<int64_t>& dims_mapping) {
454 455
  output_dist_attrs_[name].set_dims_mapping(dims_mapping);
  output_dist_attrs_[name].set_process_mesh(process_mesh_);
456 457
}

458 459 460
bool OperatorDistAttr::verify_input_dist_attr(const std::string& name,
                                              const TensorDistAttr& dist_attr,
                                              const VarDesc* tensor) const {
461 462
  VLOG(4) << "[OperatorDistAttr verify_input_dist_attr] " << name << " "
          << dist_attr.to_string();
463
  if (!dist_attr.verify(tensor)) {
464 465
    return false;
  }
466 467
  if (tensor != nullptr) {
    if (name != tensor->Name()) {
468 469 470
      return false;
    }
  }
471 472 473
  if (input_dist_attrs_.count(name) == 0) {
    return false;
  }
474 475 476
  return true;
}

477 478 479
bool OperatorDistAttr::verify_output_dist_attr(const std::string& name,
                                               const TensorDistAttr& dist_attr,
                                               const VarDesc* tensor) const {
480 481
  VLOG(4) << "[OperatorDistAttr verify_output_dist_attr] " << name << " "
          << dist_attr.to_string();
482
  if (!dist_attr.verify(tensor)) {
483 484
    return false;
  }
485 486
  if (tensor != nullptr) {
    if (name != tensor->Name()) {
487 488 489
      return false;
    }
  }
490 491 492
  if (output_dist_attrs_.count(name) == 0) {
    return false;
  }
493 494 495 496 497
  return true;
}

bool OperatorDistAttr::verify_process_mesh(
    const ProcessMesh& process_mesh) const {
498 499
  VLOG(4) << "[OperatorDistAttr verify_process_mesh] "
          << process_mesh.to_string();
500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517
  if (process_mesh != process_mesh_) {
    return false;
  }
  for (auto& item : input_dist_attrs_) {
    if (item.second.process_mesh() != process_mesh) {
      return false;
    }
  }
  for (auto& item : output_dist_attrs_) {
    if (item.second.process_mesh() != process_mesh) {
      return false;
    }
  }
  return true;
}

bool OperatorDistAttr::verify_annotated(
    const std::map<std::string, bool>& annotated) const {
518
  VLOG(4) << "[OperatorDistAttr verify_annotated] " << str_join(annotated);
519 520 521 522 523 524 525
  for (const auto& item : annotated) {
    auto result = std::find(std::begin(fields_), std::end(fields_), item.first);
    if (result == std::end(fields_)) {
      return false;
    }
  }
  for (auto& item : input_dist_attrs_) {
526 527
    VLOG(4) << "[OperatorDistAttr verify_annotated input] "
            << str_join(item.second.annotated());
528 529 530 531 532
    if (!item.second.verify_annotated(item.second.annotated())) {
      return false;
    }
  }
  for (auto& item : output_dist_attrs_) {
533 534
    VLOG(4) << "[OperatorDistAttr verify_annotated output] "
            << str_join(item.second.annotated());
535 536 537 538 539 540 541
    if (!item.second.verify_annotated(item.second.annotated())) {
      return false;
    }
  }
  return true;
}

542
bool OperatorDistAttr::verify(const OpDesc* op) const {
543 544 545 546
  if (!verify_process_mesh(process_mesh_)) {
    return false;
  }
  for (auto const& item : input_dist_attrs_) {
547
    auto input_names = op->InputArgumentNames();
548 549 550 551 552
    auto found =
        std::find(std::begin(input_names), std::end(input_names), item.first);
    if (found == std::end(input_names)) {
      return false;
    }
553 554
    auto tensor = op->Block()->FindVarRecursive(item.first);
    if (!verify_input_dist_attr(item.first, item.second, tensor)) {
555 556 557 558
      return false;
    }
  }
  for (auto const& item : output_dist_attrs_) {
559
    auto output_names = op->OutputArgumentNames();
560 561 562 563 564
    auto found =
        std::find(std::begin(output_names), std::end(output_names), item.first);
    if (found == std::end(output_names)) {
      return false;
    }
565 566
    auto tensor = op->Block()->FindVarRecursive(item.first);
    if (!verify_output_dist_attr(item.first, item.second, tensor)) {
567 568 569 570 571 572
      return false;
    }
  }
  return true;
}

573 574
void OperatorDistAttr::rename_input(const std::string& old_name,
                                    const std::string& new_name) {
575
  if (old_name == new_name) return;
576 577
  for (auto& item : input_dist_attrs_) {
    if (item.first == old_name) {
578
      input_dist_attrs_[new_name].copy_from(input_dist_attrs_[old_name]);
579 580 581 582 583 584 585 586
      input_dist_attrs_.erase(old_name);
      break;
    }
  }
}

void OperatorDistAttr::rename_output(const std::string& old_name,
                                     const std::string& new_name) {
587
  if (old_name == new_name) return;
588 589
  for (auto& item : output_dist_attrs_) {
    if (item.first == old_name) {
590
      output_dist_attrs_[new_name].copy_from(output_dist_attrs_[old_name]);
591 592 593 594 595 596
      output_dist_attrs_.erase(old_name);
      break;
    }
  }
}

597 598
std::string OperatorDistAttr::to_string() const {
  std::string str;
599
  str += "{impl_type: " + impl_type_ + ", ";
600
  str += "impl_idx: " + std::to_string(impl_idx_) + ", ";
601
  str += "execution_stream: " + execution_stream_ + ", ";
602
  str += "scheduling_priority: " + std::to_string(scheduling_priority_) + ", ";
603 604 605 606
  str += "annotated: [" + str_join(annotated_) + "], ";
  str += "\nprocess_mesh: " + process_mesh_.to_string() + ", ";
  str += "\ninput_dist_attrs: [\n";
  for (auto const& item : input_dist_attrs_) {
607
    str += "  " + item.first + ": " + item.second.to_string() + ",\n";
608 609 610 611
  }
  str.replace(str.size() - 2, 2, "]");
  str += "\noutput_dist_attrs: [\n";
  for (auto const& item : output_dist_attrs_) {
612
    str += "  " + item.first + ": " + item.second.to_string() + ",\n";
613 614 615 616 617
  }
  str.replace(str.size() - 2, 2, "]}");
  return str;
}

618
void OperatorDistAttr::from_proto(const OperatorDistAttrProto& proto) {
619
  for (int64_t i = 0; i < proto.input_dist_attrs_size(); ++i) {
620 621 622 623
    TensorDistAttr dist_attr;
    std::string name = proto.input_dist_attrs(i).name();
    dist_attr.from_proto(proto.input_dist_attrs(i).tensor_dist_attr());
    input_dist_attrs_[name] = dist_attr;
624 625
  }
  for (int64_t i = 0; i < proto.output_dist_attrs_size(); ++i) {
626 627 628 629
    TensorDistAttr dist_attr;
    std::string name = proto.output_dist_attrs(i).name();
    dist_attr.from_proto(proto.output_dist_attrs(i).tensor_dist_attr());
    output_dist_attrs_[name] = dist_attr;
630
  }
631 632 633
  process_mesh_ = ProcessMesh::from_proto(proto.process_mesh());
  impl_type_ = proto.impl_type();
  impl_idx_ = proto.impl_idx();
634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653
}

OperatorDistAttrProto OperatorDistAttr::to_proto() const {
  OperatorDistAttrProto proto;
  for (const auto& item : input_dist_attrs_) {
    auto proto_item = proto.mutable_input_dist_attrs()->Add();
    proto_item->set_name(item.first);
    proto_item->mutable_tensor_dist_attr()->CopyFrom(item.second.to_proto());
  }
  for (const auto& item : output_dist_attrs_) {
    auto proto_item = proto.mutable_output_dist_attrs()->Add();
    proto_item->set_name(item.first);
    proto_item->mutable_tensor_dist_attr()->CopyFrom(item.second.to_proto());
  }
  proto.mutable_process_mesh()->CopyFrom(process_mesh_.to_proto());
  proto.set_impl_type(impl_type_);
  proto.set_impl_idx(impl_idx_);
  return proto;
}

654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673
std::string OperatorDistAttr::serialize_to_string() {
  std::string data;
  auto proto = to_proto();
  proto.SerializeToString(&data);
  PADDLE_ENFORCE_EQ(to_proto().SerializeToString(&data),
                    true,
                    platform::errors::InvalidArgument(
                        "Failed to serialize op dist attr to string."));
  return data;
}

void OperatorDistAttr::parse_from_string(const std::string& data) {
  OperatorDistAttrProto proto;
  PADDLE_ENFORCE_EQ(proto.ParseFromString(data),
                    true,
                    platform::errors::InvalidArgument(
                        "Failed to parse op dist attr from string."));
  from_proto(proto);
}

674 675 676 677 678 679 680 681 682 683
bool operator==(const OperatorDistAttr& lhs, const OperatorDistAttr& rhs) {
  if (lhs.process_mesh() != rhs.process_mesh()) {
    return false;
  }
  if (lhs.impl_type() != rhs.impl_type()) {
    return false;
  }
  if (lhs.impl_idx() != rhs.impl_idx()) {
    return false;
  }
684 685 686
  if (lhs.execution_stream() != rhs.execution_stream()) {
    return false;
  }
687 688 689
  if (lhs.scheduling_priority() != rhs.scheduling_priority()) {
    return false;
  }
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713
  for (auto const& item : lhs.input_dist_attrs()) {
    if (rhs.input_dist_attrs().count(item.first) != 1) {
      return false;
    }
    if (rhs.input_dist_attrs().at(item.first) !=
        lhs.input_dist_attrs().at(item.first)) {
      return false;
    }
  }
  for (auto const& item : lhs.output_dist_attrs()) {
    if (rhs.output_dist_attrs().count(item.first) != 1) {
      return false;
    }
    if (rhs.output_dist_attrs().at(item.first) !=
        lhs.output_dist_attrs().at(item.first)) {
      return false;
    }
  }
  return true;
}

}  // namespace auto_parallel
}  // namespace distributed
}  // namespace paddle