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.

resource_handle.proto 1.3 kB

12345678910111213141516171819202122232425262728293031323334353637
  1. /**
  2. * This file is part of Open Source Software TensorFlow, version 1.15.0 https://github.com/tensorflow/tensorflow
  3. *
  4. * This file is included by GraphEngine so as to support model format conversion from tensorflow model to GraphEngine model.
  5. * This file in this distribution may have been modified by Huawei Technologies Co., Ltd ("Huawei Modifications").
  6. * All Huawei Modifications are Copyright 2019-2020 Huawei Technologies Co., Ltd.
  7. */
  8. syntax = "proto3";
  9. package domi.tensorflow;
  10. option cc_enable_arenas = true;
  11. option java_outer_classname = "ResourceHandle";
  12. option java_multiple_files = true;
  13. option java_package = "org.tensorflow.framework";
  14. // Protocol buffer representing a handle to a tensorflow resource. Handles are
  15. // not valid across executions, but can be serialized back and forth from within
  16. // a single run.
  17. message ResourceHandleProto {
  18. // Unique name for the device containing the resource.
  19. string device = 1;
  20. // Container in which this resource is placed.
  21. string container = 2;
  22. // Unique name of this resource.
  23. string name = 3;
  24. // Hash code for the type of the resource. Is only valid in the same device
  25. // and in the same execution.
  26. uint64 hash_code = 4;
  27. // For debug-only, the name of the type pointed to by this handle, if
  28. // available.
  29. string maybe_type_name = 5;
  30. };

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