Plans for Enhanced CPU Backend Support

[xlinsist]

I'm seeking clarification on two points:

• Does FlagGems plan to enhance CPU runtime support, especially with operator-specific tuning?
• Following the ARM backend support PR, would contributions for backend support of other CPUs—particularly for RISC-V with Vector extensions(RVV)—be welcomed?

Background

Tiling-based optimization is crucial in efficient AI kernel programming, and Triton’s popularity stems from its ease of use in this space. FlagGems is one of the earliest open-source libraries exploring this approach and already supports platforms beyond NVIDIA GPUs, according to backend code.

Given the relevance of tiling on CPUs, Triton has begun developing a triton-cpu branch. However, CPU backend support in FlagGems remains limited. Aside from a recent ARM backend support PR, support for other CPUs is minimal. Since tuning configurations optimized for ARM architectures may not generalize well to other CPU platforms such as x86 or RISC-V. Is that a way we can do something about this? Thanks!

[tongxin]

Thanks for reaching out and offering to volunteer. As you've noticed we've been teaming with community on porting flaggems kernels to ARM CPUs. We do like to extend the tuning support to other CPUs as well, particularly targeting riscv with rvv but we are not yet able to invest resources now. It would be great that you collaborate with us and work out a plan on CPU side tuning. Please contact Zhixin @StrongSpoon and Ruilin @Galaxy1458 for further information.

[xlinsist]

Thanks for sharing the details, we will look into this and circle back when we have explored further. Appreciate the connection!