Hello from the ROCm Software Platform

[Aswin Mathews]

Hello from AMD, 

I was asked to help on-board CuPy into full support with ROCm/HIP and just started building/testing on MI100 yesterday. 

New fork at https://github.com/ROCmSoftwarePlatform/cupy, mainly for internal testing, at the moment. 

I joined https://gitter.im/cupy/community and am checking up on 'cat:hip' issues and disabled tests on ROCm/HIP. 

Looking forward to working with the community. 

Thanks,

Aswin. 

[Leo Fang]

Hello Aswin,

Welcome! So glad to hear that AMD is willing to offer support to us. I'm sure with your help we'll be able to improve the quality of our HIP builds better and move forward faster to get out of the experimental phase.

I am from Brookhaven National Lab and I am one of the main drivers behind CuPy's support for ROCm 3.5+. I will work alongside the CuPy team and provide you feedbacks and guidance whenever needed. Looking forward to working with you. 

Best,

Leo

aswin....@amd.com 在 2021年4月7日 星期三下午4:39:55 [UTC-4] 的信中寫道：

[maeh...@preferred.jp]

Hello Aswin,

I'm Kenichi from the CuPy team. We're happy to welcome you to the CuPy community!

Looking forward to working with you.

> checking up on 'cat:hip' issues and disabled tests on ROCm/HIP. 

Thank you! Currently, some unit tests are marked as an "expected failure" on ROCm (grep xfail in tests directory). These XFAILs indicate the difference between ROCm and CUDA. Most of them are empricially known to fail on ROCm and we haven't investigated much in the detailed condition to reproduce it. After supporting ROCm 4.1 build, we would like to remove these XFAILs as much as possible. Some may need a fix in ROCm, CuPy, or both; and some may only happen on the specific GPU.

Thanks,

Kenichi

2021年4月9日金曜日 9:14:32 UTC+9 leo8...@gmail.com: