Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[MFMA] Implement MFMA 4x64 v3 #550

Draft
wants to merge 8 commits into
base: triton-mlir
Choose a base branch
from

Conversation

binarman
Copy link

@binarman binarman commented Apr 1, 2024

This implements version 3 mfma 4x64 layout with swizzled operands

This PR:
- adds an `matrix_instr_nonkdim` options to force MFMA 64x4 and 4x64 layout: 464 corresponds 4(M)x64(N), 644 corresponds 64(M)x4(N)
- adds tests for this option
- fixes swizzling patter in some cases

MFMA size heuristic now looks like this:

1. If kernel specific option is set, pick it
2. If the result tile shape is larger than 32x32, pick mfma32
3. If the tile shape is smaller than 32x32 but larger than 16x16, pick mfma16
4. if the tile shape is smaller than 4x64 or 64x4, pick mfma4x4
5. Otherwise, pick mfma4x64 or mfma64x4, depending on what tile fits into matrices
Extend K dimension of mfma4x64 and mfma64x4 dot operand layout from 4 to 64.
tuning command: `python3 tune_gemm.py --compare -m 16 -n 64 -k 64 -dtype_a "fp16" -dtype_b "fp16" -dtype_c "fp32" --num_threads=16 --keep`
This implements version 3 mfma 4x64 layout with swizzled operands
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant