[CPU] Limit the use of [8, 32, 16] gemm vector sizes to CPUs w/ avx512f feature

iree-org / iree

A retargetable MLIR-based machine learning compiler and runtime toolkit.

http://iree.dev/

Apache License 2.0

2.47k stars 551 forks source link

[CPU] Limit the use of [8, 32, 16] gemm vector sizes to CPUs w/ avx512f feature #17727

Closed hanhanW closed 6 days ago

hanhanW commented 1 week ago

The tile sizes were tuned for targets that have avx512f features. The revision update the default vector sizes to [1, 1, vector_size] for targets w/o avx512f feature, which avoids large vector sizes.

Fixes https://github.com/iree-org/iree/issues/17683