dist_attr.cc 24.3 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
TensorDistAttr::TensorDistAttr(const VarDesc& tensor) : tensor_(&tensor) {
  VLOG(4) << "[TensorDistAttr constructor] tensor name: " << tensor_->Name();
  if (tensor_->GetType() == framework::proto::VarType::READER) return;
  if (tensor_->GetType() == framework::proto::VarType::LOD_TENSOR_ARRAY) return;
  if (tensor_->GetType() == framework::proto::VarType::STEP_SCOPES) return;
  tensor_shape_ = tensor_->GetShape();
  VLOG(4) << "[TensorDistAttr constructor] tensor shape: "
          << str_join(tensor_shape_);
40
  set_default_dims_mapping();
41
  for (std::size_t i = 0; i < tensor_shape_.size(); ++i) {
42 43 44 45 46 47
    dynamic_dims_.push_back(false);
  }
}

TensorDistAttr::TensorDistAttr(const TensorDistAttr& dist_attr) {
  if (tensor_ == nullptr) {
48 49
    tensor_ = dist_attr.tensor_;
    tensor_shape_ = dist_attr.tensor_shape_;
50
  }
51 52 53 54 55 56 57 58
  if (tensor_ != nullptr) {
    VLOG(4) << "[TensorDistAttr copy constructor] tensor name:  "
            << tensor_->Name() << ", tensro shape: " << str_join(tensor_shape_);
  } else {
    VLOG(4) << "[TensorDistAttr copy constructor] tensor name:  None"
            << ", tensro shape: " << str_join(tensor_shape_);
  }
  copy_from(dist_attr);
59 60 61 62
}

TensorDistAttr& TensorDistAttr::operator=(const TensorDistAttr& dist_attr) {
  if (tensor_ == nullptr) {
63 64
    tensor_ = dist_attr.tensor_;
    tensor_shape_ = dist_attr.tensor_shape_;
65
  }
66 67 68 69 70 71 72 73 74 75 76 77
  if (tensor_ != nullptr) {
    VLOG(4) << "[TensorDistAttr assign constructor] tensor name:  "
            << tensor_->Name() << ", tensro shape: " << str_join(tensor_shape_);
  } else {
    VLOG(4) << "[TensorDistAttr assign constructor] tensor name:  None"
            << ", tensro shape: " << str_join(tensor_shape_);
  }
  copy_from(dist_attr);
  return *this;
}

void TensorDistAttr::copy_from(const TensorDistAttr& dist_attr) {
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
  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) {
  PADDLE_ENFORCE_EQ(verify_process_mesh(process_mesh),
                    true,
                    platform::errors::InvalidArgument(
                        "Wrong process mesh %s.", process_mesh.to_string()));
  process_mesh_ = process_mesh;
}

void TensorDistAttr::set_dims_mapping(
    const std::vector<int64_t>& dims_mapping) {
  PADDLE_ENFORCE_EQ(verify_dims_mapping(dims_mapping),
                    true,
                    platform::errors::InvalidArgument("Wrong dims_mapping %s.",
                                                      str_join(dims_mapping)));
  dims_mapping_ = dims_mapping;
}

void TensorDistAttr::set_batch_dim(int64_t batch_dim) {
  PADDLE_ENFORCE_EQ(
      verify_batch_dim(batch_dim),
      true,
      platform::errors::InvalidArgument(
          "Wrong batch_dim %d in this distributed attribute.", batch_dim));
108 109 110
  if (tensor_ != nullptr && tensor_shape_.size() > 0) {
    int64_t canonical_batch_dim =
        canonical_dim(batch_dim, tensor_shape_.size());
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136
    batch_dim_ = canonical_batch_dim;
  } else {
    batch_dim_ = batch_dim;
  }
}

void TensorDistAttr::set_dynamic_dims(const std::vector<bool>& dynamic_dims) {
  PADDLE_ENFORCE_EQ(
      verify_dynamic_dims(dynamic_dims),
      true,
      platform::errors::InvalidArgument("The dynamic_dims [%s] is wrong.",
                                        str_join(dynamic_dims)));
  dynamic_dims_ = dynamic_dims;
}

void TensorDistAttr::set_annotated(
    const std::map<std::string, bool>& annotated) {
  PADDLE_ENFORCE_EQ(verify_annotated(annotated),
                    true,
                    platform::errors::InvalidArgument(
                        "The annotated [%s] is wrong.", str_join(annotated)));
  annotated_ = annotated;
}

void TensorDistAttr::set_default_dims_mapping() {
  if (tensor_ != nullptr) {
137
    dims_mapping_ = std::vector<int64_t>(tensor_shape_.size(), -1);
138 139 140 141 142 143 144 145 146 147 148 149
  }
}

void TensorDistAttr::annotate(const std::string& name) {
  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 {
150 151
  VLOG(4) << "[TensorDistAttr verify_process_mesh] "
          << process_mesh.to_string();
152 153 154 155 156 157 158 159 160 161 162 163
  if (!process_mesh_.empty()) {
    for (int64_t dim_mapping : dims_mapping_) {
      if (dim_mapping < -1 || dim_mapping >= process_mesh_.ndim()) {
        return false;
      }
    }
  }
  return true;
}

bool TensorDistAttr::verify_dims_mapping(
    const std::vector<int64_t>& dims_mapping) const {
164 165 166
  VLOG(4) << "[TensorDistAttr verify_dims_mapping] " << str_join(dims_mapping);
  if (dims_mapping.size() != tensor_shape_.size()) {
    return false;
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
  }
  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;
}

bool TensorDistAttr::verify_batch_dim(int64_t dim) const {
191 192 193
  VLOG(4) << "[TensorDistAttr verify_batch_dim] " << dim;
  int64_t ndim = tensor_shape_.size();
  if (tensor_ != nullptr && ndim > 0) {
194 195 196 197 198 199 200 201 202 203 204 205
    if (dim < 0) {
      dim = dim + ndim;
    }
    if (dim < 0 || dim >= ndim) {
      return false;
    }
  }
  return true;
}

bool TensorDistAttr::verify_dynamic_dims(
    const std::vector<bool>& dynamic_dims) const {
206 207 208
  VLOG(4) << "[TensorDistAttr verify_dynamic_dims] " << str_join(dynamic_dims);
  if (dynamic_dims.size() != tensor_shape_.size()) {
    return false;
209 210 211 212 213 214
  }
  return true;
}

bool TensorDistAttr::verify_annotated(
    const std::map<std::string, bool>& annotated) const {
215
  VLOG(4) << "[TensorDistAttr verify_annotated] " << str_join(annotated);
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258
  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;
}

bool TensorDistAttr::verify() const {
  if (!verify_process_mesh(process_mesh_)) {
    return false;
  }
  if (!verify_dims_mapping(dims_mapping_)) {
    return false;
  }
  if (!verify_batch_dim(batch_dim_)) {
    return false;
  }
  if (!verify_dynamic_dims(dynamic_dims_)) {
    return false;
  }
  if (!verify_annotated(annotated_)) {
    return false;
  }
  return true;
}

std::string TensorDistAttr::to_string() const {
  std::string dist_str;
  if (tensor_ != nullptr) {
    dist_str = "{tensor_name: " + tensor_->Name() + ", ";
  } else {
    dist_str = "{tensor_name: None, ";
  }
  dist_str += "process_mesh: " + process_mesh_.to_string() + ", ";
  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;
}

259 260 261
void TensorDistAttr::from_proto(const TensorDistAttrProto& proto) {
  process_mesh_ = ProcessMesh::from_proto(proto.process_mesh());
  dims_mapping_.resize(proto.dims_mapping_size());
262
  for (int64_t i = 0; i < proto.dims_mapping_size(); ++i) {
263
    dims_mapping_[i] = proto.dims_mapping(i);
264
  }
265 266
  batch_dim_ = proto.batch_dim();
  dynamic_dims_.resize(proto.dynamic_dims_size());
267
  for (int64_t i = 0; i < proto.dynamic_dims_size(); ++i) {
268
    dynamic_dims_[i] = proto.dynamic_dims(i);
269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
  }
}

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;
}

285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304
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);
}

305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
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;
}

std::vector<std::string> OperatorDistAttr::fields_{
322
    "process_mesh", "impl_type", "impl_idx", "execution_stream"};
323 324

OperatorDistAttr::OperatorDistAttr(const OpDesc& op) : op_(&op) {
325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
  VLOG(4) << "[OperatorDistAttr constructor] op type: " << op_->Type();
  initialize();
}

OperatorDistAttr::OperatorDistAttr(const OperatorDistAttr& dist_attr) {
  if (op_ == nullptr) {
    op_ = dist_attr.op();
  }
  if (op_ != nullptr) {
    VLOG(4) << "[OperatorDistAttr copy constructor] op type: " << op_->Type();
  } else {
    VLOG(4) << "[OperatorDistAttr copy constructor] op type: None";
  }
  initialize();
  copy_from(dist_attr);
}

OperatorDistAttr& OperatorDistAttr::operator=(
    const OperatorDistAttr& dist_attr) {
  if (op_ == nullptr) {
    op_ = dist_attr.op();
  }
  if (op_ != nullptr) {
    VLOG(4) << "[OperatorDistAttr assign constructor] op type: " << op_->Type();
  } else {
    VLOG(4) << "[OperatorDistAttr assign constructor] op type: None";
  }
  initialize();
  copy_from(dist_attr);
  return *this;
}

void OperatorDistAttr::initialize() {
  if (op_ == nullptr) return;
359 360
  for (std::string name : op_->InputArgumentNames()) {
    VarDesc* input = op_->Block()->FindVarRecursive(name);
361
    VLOG(4) << "[OperatorDistAttr create input dist attr] " << name;
362
    inputs_[name] = input;
363 364 365 366 367
    if (input == nullptr || op_->Type() == "create_py_reader") {
      input_dist_attrs_[name] = TensorDistAttr();
    } else {
      input_dist_attrs_[name] = TensorDistAttr(*input);
    }
368 369 370
  }
  for (std::string name : op_->OutputArgumentNames()) {
    VarDesc* output = op_->Block()->FindVarRecursive(name);
371
    VLOG(4) << "[OperatorDistAttr create output dist attr] " << name;
372
    outputs_[name] = output;
373 374 375 376 377
    if (output == nullptr) {
      output_dist_attrs_[name] = TensorDistAttr();
    } else {
      output_dist_attrs_[name] = TensorDistAttr(*output);
    }
378
  }
379
  impl_type_ = kDefault;
380
  impl_idx_ = 0;
381
  execution_stream_ = kDefault;
382 383
}

384 385 386
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());
387 388 389
  set_process_mesh(dist_attr.process_mesh());
  set_impl_type(dist_attr.impl_type());
  set_impl_idx(dist_attr.impl_idx());
390
  set_execution_stream(dist_attr.execution_stream());
391 392 393
  set_annotated(dist_attr.annotated());
}

394 395
void OperatorDistAttr::set_input_dist_attrs(
    const std::map<std::string, TensorDistAttr>& dist_attrs) {
396
  if (op_ == nullptr) {
397 398 399 400 401 402 403 404 405
    for (const auto& item : dist_attrs) {
      set_input_dist_attr(item.first, item.second);
    }
  } else {
    for (const auto& item : input_dist_attrs_) {
      if (dist_attrs.count(item.first) == 1) {
        set_input_dist_attr(item.first, dist_attrs.at(item.first));
      }
    }
406
  }
407 408 409 410 411 412 413 414 415 416 417 418 419 420
}

void OperatorDistAttr::set_output_dist_attrs(
    const std::map<std::string, TensorDistAttr>& dist_attrs) {
  if (op_ == nullptr) {
    for (const auto& item : dist_attrs) {
      set_output_dist_attr(item.first, item.second);
    }
  } else {
    for (const auto& item : output_dist_attrs_) {
      if (dist_attrs.count(item.first) == 1) {
        set_output_dist_attr(item.first, dist_attrs.at(item.first));
      }
    }
421 422 423 424 425 426 427 428
  }
}

void OperatorDistAttr::set_input_dist_attr(const std::string& name,
                                           const TensorDistAttr& dist_attr) {
  PADDLE_ENFORCE_EQ(
      verify_input_dist_attr(name, dist_attr),
      true,
429 430 431 432
      platform::errors::InvalidArgument("Wrong dist_attr %s for %s. %s",
                                        dist_attr.to_string(),
                                        name,
                                        to_string()));
433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483
  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) {
  PADDLE_ENFORCE_EQ(
      verify_output_dist_attr(name, dist_attr),
      true,
      platform::errors::InvalidArgument(
          "Wrong dist_attr %s for %s.", dist_attr.to_string(), name));
  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;
}

void OperatorDistAttr::annotate(const std::string& name) {
  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_) {
      item.second.annotate(name);
    }
    for (auto& item : output_dist_attrs_) {
      item.second.annotate(name);
    }
  }
}

void OperatorDistAttr::set_annotated(
    const std::map<std::string, bool>& annotated) {
  PADDLE_ENFORCE_EQ(verify_annotated(annotated),
                    true,
                    platform::errors::InvalidArgument(
                        "The annotated [%s] is wrong.", str_join(annotated)));
  annotated_ = annotated;
}

484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503
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) {
  input_dist_attr(name).set_dims_mapping(dims_mapping);
}

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) {
  output_dist_attr(name).set_dims_mapping(dims_mapping);
}

504 505
bool OperatorDistAttr::verify_input_dist_attr(
    const std::string& name, const TensorDistAttr& dist_attr) const {
506 507
  VLOG(4) << "[OperatorDistAttr verify_input_dist_attr] " << name << " "
          << dist_attr.to_string();
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525
  if (!dist_attr.verify()) {
    return false;
  }
  if (op_ != nullptr) {
    if (dist_attr.tensor() != nullptr) {
      if (name != dist_attr.tensor()->Name()) {
        return false;
      }
    }
    if (input_dist_attrs_.count(name) == 0) {
      return false;
    }
  }
  return true;
}

bool OperatorDistAttr::verify_output_dist_attr(
    const std::string& name, const TensorDistAttr& dist_attr) const {
526 527
  VLOG(4) << "[OperatorDistAttr verify_output_dist_attr] " << name << " "
          << dist_attr.to_string();
528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545
  if (!dist_attr.verify()) {
    return false;
  }
  if (op_ != nullptr) {
    if (dist_attr.tensor() != nullptr) {
      if (name != dist_attr.tensor()->Name()) {
        return false;
      }
    }
    if (output_dist_attrs_.count(name) == 0) {
      return false;
    }
  }
  return true;
}

bool OperatorDistAttr::verify_process_mesh(
    const ProcessMesh& process_mesh) const {
546 547
  VLOG(4) << "[OperatorDistAttr verify_process_mesh] "
          << process_mesh.to_string();
548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565
  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 {
566
  VLOG(4) << "[OperatorDistAttr verify_annotated] " << str_join(annotated);
567 568 569 570 571 572 573
  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_) {
574 575
    VLOG(4) << "[OperatorDistAttr verify_annotated input] "
            << str_join(item.second.annotated());
576 577 578 579 580
    if (!item.second.verify_annotated(item.second.annotated())) {
      return false;
    }
  }
  for (auto& item : output_dist_attrs_) {
581 582
    VLOG(4) << "[OperatorDistAttr verify_annotated output] "
            << str_join(item.second.annotated());
583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621
    if (!item.second.verify_annotated(item.second.annotated())) {
      return false;
    }
  }
  return true;
}

bool OperatorDistAttr::verify() const {
  if (op_ == nullptr) {
    return false;
  }
  if (!verify_process_mesh(process_mesh_)) {
    return false;
  }
  for (auto const& item : input_dist_attrs_) {
    auto input_names = op_->InputArgumentNames();
    auto found =
        std::find(std::begin(input_names), std::end(input_names), item.first);
    if (found == std::end(input_names)) {
      return false;
    }
    if (!verify_input_dist_attr(item.first, item.second)) {
      return false;
    }
  }
  for (auto const& item : output_dist_attrs_) {
    auto output_names = op_->OutputArgumentNames();
    auto found =
        std::find(std::begin(output_names), std::end(output_names), item.first);
    if (found == std::end(output_names)) {
      return false;
    }
    if (!verify_output_dist_attr(item.first, item.second)) {
      return false;
    }
  }
  return true;
}

622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659
void OperatorDistAttr::rename_input(const std::string& old_name,
                                    const std::string& new_name) {
  for (auto& item : input_dist_attrs_) {
    if (item.first == old_name) {
      VarDesc* new_input = op_->Block()->FindVarRecursive(new_name);
      inputs_[new_name] = new_input;
      if (new_input == nullptr) {
        input_dist_attrs_[new_name] = TensorDistAttr();
      } else {
        input_dist_attrs_[new_name] = TensorDistAttr(*new_input);
        input_dist_attrs_[new_name].copy_from(input_dist_attrs_[old_name]);
      }
      inputs_.erase(old_name);
      input_dist_attrs_.erase(old_name);
      break;
    }
  }
}

void OperatorDistAttr::rename_output(const std::string& old_name,
                                     const std::string& new_name) {
  for (auto& item : output_dist_attrs_) {
    if (item.first == old_name) {
      VarDesc* new_output = op_->Block()->FindVarRecursive(new_name);
      outputs_[new_name] = new_output;
      if (new_output == nullptr) {
        output_dist_attrs_[new_name] = TensorDistAttr();
      } else {
        output_dist_attrs_[new_name] = TensorDistAttr(*new_output);
        output_dist_attrs_[new_name].copy_from(output_dist_attrs_[old_name]);
      }
      outputs_.erase(old_name);
      output_dist_attrs_.erase(old_name);
      break;
    }
  }
}

660 661 662 663 664 665 666 667 668
std::string OperatorDistAttr::to_string() const {
  std::string str;
  if (op_ != nullptr) {
    str += "{op_type: " + op_->Type() + ", ";
  } else {
    str += "{op_type: None, ";
  }
  str += "impl_type: " + impl_type_ + ", ";
  str += "impl_idx: " + std::to_string(impl_idx_) + ", ";
669
  str += "execution_stream: " + execution_stream_ + ", ";
670 671 672 673 674 675 676 677 678 679 680 681 682 683 684
  str += "annotated: [" + str_join(annotated_) + "], ";
  str += "\nprocess_mesh: " + process_mesh_.to_string() + ", ";
  str += "\ninput_dist_attrs: [\n";
  for (auto const& item : input_dist_attrs_) {
    str += "  " + item.second.to_string() + ",\n";
  }
  str.replace(str.size() - 2, 2, "]");
  str += "\noutput_dist_attrs: [\n";
  for (auto const& item : output_dist_attrs_) {
    str += "  " + item.second.to_string() + ",\n";
  }
  str.replace(str.size() - 2, 2, "]}");
  return str;
}

685
void OperatorDistAttr::from_proto(const OperatorDistAttrProto& proto) {
686
  for (int64_t i = 0; i < proto.input_dist_attrs_size(); ++i) {
687 688 689 690
    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;
691 692
  }
  for (int64_t i = 0; i < proto.output_dist_attrs_size(); ++i) {
693 694 695 696
    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;
697
  }
698 699 700
  process_mesh_ = ProcessMesh::from_proto(proto.process_mesh());
  impl_type_ = proto.impl_type();
  impl_idx_ = proto.impl_idx();
701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720
}

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;
}

721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740
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);
}

741 742 743 744 745 746 747 748 749 750
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;
  }
751 752 753
  if (lhs.execution_stream() != rhs.execution_stream()) {
    return false;
  }
754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777
  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