Release 1.2.0
Major Features and Improvements
- [STABLE] Rebuild the AKG repository for providing a new way to support ascend backend by linking a static library contained all the ascend passes. (Ascend)
- [STABLE] Optimize the reduction add operation in ascend backend. (Ascend)
- [STABLE] Add support for tuning elemwise&&reduction operators. (GPU)
Bug fixes
- Fixed a problem that data prefetch cannot be enabled by attributes in DSL.
- Fixed bugs of autotiling algorithms (tiling too small, cannot adapted matmul+bias, etc.) in Ascend platform.
- Fixed local memory promotion for large thread (2980!)
- Fixed reduce binding dimension issue on gpu platform (ff38!)
Contributors
Thanks goes to these wonderful people:
yangsijia, xxxxxxw, polyhedral, zhangrenwei, yiyanzhi, xixixian, hujiahui8, zhengzuohe, lishanni, zhangzhaochuang, xuhui, liuchao, gengzhen, xiaruijie,
chenlei_autodiff, lingyunli63, wYann, lvwenyuan, peiwenfang, hanhuifeng, gaoxiong, chengyun
Contributions of any kind are welcome!
Initial Version
- Upload the initial framework
- Basic support for Ascend910 platform and gpu v100
- Integration with GraphKernel fusion of MindSpore.