[XPU][TritonGPUToLLVM] Avoid bank conflicts in sub-group transposes #2769
Conversation
victor-eds
requested review from
a team,
chengjunlu,
etiotto and
whitneywhtsang
|
In my mind, the bank number of SLM is 65 on PVC. I think we need to make sure the bank number of the SLM. |
- Store the whole matrix using SIMD block stores for each row leaving a single garbage item at the end of the row so each row has `sub_group_size + 1` elements - Load each row with vector loads By introducing this garbage item at the end of each row, we ensure matrix loading avoid bank conflicts as the offset between the position loaded by work-item `i` and `i+j` is `N * (sub_group_size + 1)` (assuming `sub_group_size` banks). Signed-off-by: victor-eds <[email protected]>
victor-eds
force-pushed
the
reduction-opt-avoid-bank-conflicts
branch
from
a1c5725 to
b6cd04c
Compare
PVC has 64 8B banks. This PR indeed helps avoid bank conflicts. It's also true we could make this more optimal, but I'd go with this for now and fine-tune in a followup PR. WDYT? |
LGTM. |
Good impact on the |
|
@chengjunlu I created #2797 |
4 participants
a single garbage item at the end of the row so each row has
sub_group_size + 1elementsBy introducing this garbage item at the end of each row, we ensure matrix
loading avoid bank conflicts as the offset between the position loaded by
work-item
iandi+jisN * (sub_group_size + 1)(assumingsub_group_sizebanks).
Closes #2751