ggml-hexagon: Does ggml-hexagon not use the HMX unit? #17655
Unanswered
kimminsu38oo
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Hello, I was reviewing the Hexagon backend source code in matmul-ops.c and noticed that the matrix multiplication operators are implemented using HVX intrinsics.
However, I could not find any usage of the HMX (Hexagon Matrix Extensions) unit. I understand that HMX usage is possible, as demonstrated in this paper and repository which implement FlashAttention using HMX.
I have a few questions regarding this:
Is it correct that the current implementation does not utilize HMX?
Is programming with HMX currently not feasible? (e.g., due to lack of public documentation or support from Qualcomm?)
Or is avoiding HMX a deliberate design choice?
I would appreciate your insights on why HMX is not used in this backend.
Thank you.
Beta Was this translation helpful? Give feedback.
All reactions