You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

hybrid_mem_assigner.cc 3.1 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/build/memory/hybrid_mem_assigner.h"
  17. #include <utility>
  18. #include <vector>
  19. #include "framework/common/debug/ge_log.h"
  20. #include "graph/build/memory/binary_block_mem_assigner.h"
  21. #include "graph/build/memory/max_block_mem_assigner.h"
  22. namespace ge {
  23. HybridMemAssigner::HybridMemAssigner(ge::ComputeGraphPtr compute_graph)
  24. : compute_graph_(std::move(compute_graph)), priority_assigner_(nullptr) {}
  25. Status HybridMemAssigner::AssignMemory(std::unique_ptr<BlockMemAssigner> &block_assigner, size_t &mem_size) {
  26. vector<int64_t> ranges;
  27. GE_CHECK_NOTNULL(block_assigner);
  28. if (block_assigner->GetMemoryRanges(ranges) != SUCCESS) {
  29. GELOGE(FAILED, "[Get][MemoryRanges] Fail!");
  30. return FAILED;
  31. }
  32. GE_IF_BOOL_EXEC(ranges.empty(), return SUCCESS);
  33. block_assigner->AssignMemoryWithReuse(ranges);
  34. // total size
  35. for (auto it : block_assigner->GetMemOffsets()) {
  36. mem_size += it.second;
  37. }
  38. return SUCCESS;
  39. }
  40. Status HybridMemAssigner::Assign() {
  41. if (GraphUtils::GetRefMapping(compute_graph_, symbol_to_anchors_, anchor_to_symbol_) != GRAPH_SUCCESS) {
  42. REPORT_CALL_ERROR("E19999", "Get ref-mapping for graph %s failed", compute_graph_->GetName().c_str());
  43. GELOGE(FAILED, "[Get][RefMapping] for graph %s failed.", compute_graph_->GetName().c_str());
  44. return FAILED;
  45. }
  46. std::unique_ptr<BlockMemAssigner> binary_assigner(new (std::nothrow) BinaryBlockMemAssigner(
  47. compute_graph_, anchor_to_symbol_, symbol_to_anchors_));
  48. GE_CHECK_NOTNULL(binary_assigner);
  49. std::unique_ptr<BlockMemAssigner> max_assigner(new (std::nothrow) MaxBlockMemAssigner(
  50. compute_graph_, anchor_to_symbol_, symbol_to_anchors_));
  51. GE_CHECK_NOTNULL(max_assigner);
  52. size_t bin_mem_size = 0;
  53. size_t max_mem_size = 0;
  54. GE_CHK_STATUS_RET(AssignMemory(binary_assigner, bin_mem_size), "[Assign][Memory] Fail!");
  55. GE_CHK_STATUS_RET(AssignMemory(max_assigner, max_mem_size), "[Assign][Memory] Fail!");
  56. std::unique_ptr<BlockMemAssigner> priority_assigner;
  57. GELOGD("Binary-block memory size:%zu, max-block memory size:%zu", bin_mem_size, max_mem_size);
  58. if (bin_mem_size <= max_mem_size) {
  59. GELOGD("Use binary-block memory assigner method");
  60. priority_assigner = std::move(binary_assigner);
  61. } else {
  62. GELOGI("Use max-block memory assigner method");
  63. priority_assigner = std::move(max_assigner);
  64. }
  65. priority_assigner->SetOpMemOffset(false);
  66. mem_offsets_ = priority_assigner->GetMemOffsets();
  67. priority_assigner_ = std::move(priority_assigner);
  68. return SUCCESS;
  69. }
  70. } // namespace ge

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示