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.

aicpu_task_builder.h 1.5 kB

4 years ago
4 years ago
4 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344
  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. #ifndef GE_SINGLE_OP_TASK_AICPU_TASK_BUILDER_H_
  17. #define GE_SINGLE_OP_TASK_AICPU_TASK_BUILDER_H_
  18. #include <vector>
  19. #include "graph/op_desc.h"
  20. #include "single_op/single_op.h"
  21. #include "single_op/single_op_model.h"
  22. #include "cce/aicpu_engine_struct.h"
  23. namespace ge {
  24. class AiCpuTaskBuilder {
  25. public:
  26. AiCpuTaskBuilder(const OpDescPtr &op_desc, const domi::KernelExDef &kernel_def);
  27. ~AiCpuTaskBuilder() = default;
  28. Status BuildTask(AiCpuTask &task, const SingleOpModelParam &param, uint64_t kernel_id);
  29. private:
  30. static Status SetKernelArgs(void **args, STR_FWK_OP_KERNEL &kernel);
  31. Status SetFmkOpKernel(void *io_addr, void *ws_addr, STR_FWK_OP_KERNEL &kernel);
  32. Status InitWorkspaceAndIO(AiCpuTask &task, const SingleOpModelParam &param);
  33. const OpDescPtr op_desc_;
  34. const domi::KernelExDef &kernel_def_;
  35. };
  36. } // namespace ge
  37. #endif // GE_SINGLE_OP_TASK_AICPU_TASK_BUILDER_H_

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