Add thread throttling profile for SGEMM on NEOVERSEV1 #5071
Conversation
|
@martin-frbg Could I please get a review for this patch ? |
|
As soon as I can concentrate on it. Would you happen to have any benchmark numbers available ? The idea of using varying thread counts has been floated a few times in the past, but the added logic seemed to outweigh gains on x86_64 at least |
|
@martin-frbg Thank you for your comment. As mentioned in the description, the geo. mean speedup is ~1.1 for M=N=K up to 1024. Here is a figure that shows the data on NEOVERSEV1 with 64 cores. |
|
Hmm, I don't really like the added complexity with yet another "hidden" table of parameters (I assume the V2 would need different thresholds), but what's not to like about the added performance... |
|
Yes, V2 would need another set of parameters. I think the complexity is inevitable if we want optimal performance. We could try to somehow organize these tuning parameters, e.g. into files/folders. I am open to suggestions. |
|
Merging to give it more exposure, but I think ideally the table part should be in (or closer to) param.h |
This PR introduces a new thread throttling profile for SGEMM on NEOVERSEV1 CPU core.
Benchmarking for M=N=K=[2, 1024] shows a geometric mean speedup of ~1.1x.