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.

RELEASE.md 1.2 kB

5 years ago
5 years ago
5 years ago
5 years ago
1234567891011121314151617181920
  1. # Release 0.2.0-alpha
  2. ## Major Features and Improvements
  3. - Provides a common graph-level option, and multiple requirements can also share this mechanism in the future.
  4. - Improve graph compilation performance.
  5. - Optimize memory allocation.
  6. - Optimize serveral operators e.g., Slice, StridedSlice, ScatterMax etc.
  7. ## Bugfixes
  8. - Delete redudant codes.([#I1EU2Z](https://gitee.com/mindspore/graphengine/issues/I1EU2Z))
  9. - Fix HCCL initilization bugs under train and eval scenarios.([#I1DIBJ](https://gitee.com/mindspore/graphengine/issues/I1DIBJ))
  10. - Optimize compilation and linking process, enhancing efficiency and performance of concurrent compilation of GraphEngine and MindSpore. ([#I1DFIY](https://gitee.com/mindspore/mindspore/issues/I1DFIY))
  11. - Fix the bug that GE checkpoint cannot save variable names correctly.([#I1DIBJ](https://gitee.com/mindspore/graphengine/issues/I1DIBJ))
  12. - Save dump files on every iteration instead of every execution.([#I1DIBJ](https://gitee.com/mindspore/graphengine/issues/I1DIBJ))
  13. ## Thanks to our Contributors
  14. Thanks goes to these wonderful people: Wang Cong, Tianchun Xu, Haoran Yang.
  15. Contributions of any kind are welcome!

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