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.

tbe_kernel_store.h 1.6 kB

5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455
  1. /**
  2. * Copyright 2019-2020 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_COMMON_TBE_KERNEL_STORE_H_
  17. #define GE_COMMON_TBE_KERNEL_STORE_H_
  18. #include <cstdint>
  19. #include <memory>
  20. #include <string>
  21. #include <unordered_map>
  22. #include <vector>
  23. #include "framework/common/fmk_types.h"
  24. #include "graph/op_desc.h"
  25. #include "graph/op_kernel_bin.h"
  26. namespace ge {
  27. using TBEKernel = ge::OpKernelBin;
  28. using TBEKernelPtr = std::shared_ptr<ge::OpKernelBin>;
  29. class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY TBEKernelStore {
  30. public:
  31. TBEKernelStore();
  32. ~TBEKernelStore() = default;
  33. void AddTBEKernel(const TBEKernelPtr &kernel);
  34. bool Build();
  35. bool Load(const uint8_t *data, const size_t &len);
  36. TBEKernelPtr FindTBEKernel(const std::string &name) const;
  37. void LoadTBEKernelBinToOpDesc(const std::shared_ptr<ge::OpDesc> &op_desc) const;
  38. const uint8_t *Data() const;
  39. size_t DataSize() const;
  40. private:
  41. std::unordered_map<std::string, TBEKernelPtr> kernels_;
  42. std::vector<uint8_t> buffer_;
  43. };
  44. } // namespace ge
  45. #endif // GE_COMMON_TBE_KERNEL_STORE_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