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.

rts_kernel_task_builder.cc 1.8 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445
  1. /**
  2. * Copyright 2021 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 "single_op/task/rts_kernel_task_builder.h"
  17. #include "single_op/task/build_task_utils.h"
  18. namespace ge {
  19. namespace {
  20. const size_t kNumAddresses = 2;
  21. } // namespace
  22. Status RtsKernelTaskBuilder::BuildMemcpyAsyncTask(const OpDescPtr &op_desc,
  23. const domi::MemcpyAsyncDef &kernel_def,
  24. const SingleOpModelParam &param,
  25. std::unique_ptr<MemcpyAsyncTask> &task) {
  26. task.reset(new(std::nothrow)MemcpyAsyncTask());
  27. GE_CHECK_NOTNULL(task);
  28. task->SetOpDesc(op_desc);
  29. task->dst_max_ = kernel_def.dst_max();
  30. task->count_ = kernel_def.count();
  31. task->kind_ = static_cast<rtMemcpyKind_t>(kernel_def.kind());
  32. auto addresses = BuildTaskUtils::JoinAddresses(BuildTaskUtils::GetAddresses(op_desc, param, false));
  33. if (addresses.size() != kNumAddresses) {
  34. GELOGE(INTERNAL_ERROR, "[Build][MemcpyAsyncTask] Invalid address count: %zu", addresses.size());
  35. return INTERNAL_ERROR;
  36. }
  37. task->addresses_[0] = reinterpret_cast<uintptr_t>(addresses[0]);
  38. task->addresses_[1] = reinterpret_cast<uintptr_t>(addresses[1]);
  39. return SUCCESS;
  40. }
  41. } // 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