[Tracking Issue] Improve CPU backend support without CUDA

hidet-org / hidet

An open-source efficient deep learning framework/compiler, written in python.

https://hidet.org

Apache License 2.0

648 stars 52 forks source link

[Tracking Issue] Improve CPU backend support without CUDA #187

Open LDY1998 opened 1 year ago

LDY1998 commented 1 year ago

This is a tracking issue for adding CPU compilation without CUDA:

[x] Refactor codegen to separate GPU/CPU code generation
176
[x] Refactor Storage api to use libc for memory management instead of CUDA Host
189
[ ] Add test workflow in cpu-only enivironment

tonyzhao-jt commented 1 year ago

Hi, when will the CPU version be available？ Looking forward to it.

yaoyaoding commented 1 year ago

Hi @SpringWave1, you can try it even now.

a = hidet.randn([3, 4])
b = hidet.randn([4, 3])
c = a @ b

Hidet will generate the matrix multiplication kernel and run it with CPU. However, now our auto-scheduler only generates kernel that uses a single thread, and there is not efficient matmul schedule template. We are still working on these. The good support might come in the end of this year.

hidet-org / hidet

[Tracking Issue] Improve CPU backend support without CUDA #187

176

189