Uh oh!
There was an error while loading. Please reload this page.
Scale P and Q with L2 cache size for SVE - #4397
Conversation
The defaults in param.h now reflect an L2 size of 128KB, and that is scaled based on the actual size.
Mousius
commented
Dec 27, 2023
@martin-frbg , this is closer to what I was thinking previously, what do you think? I can see others have done similar in |
martin-frbg
commented
Dec 27, 2023
yes, for a specific cpu TARGET build I think the factor would have to be applied in common_param.h but I have limited brain capacity for that right now |
Mousius
commented
Dec 29, 2023
Thanks @martin-frbg, I'll look into it 😸 ! |
| #define SGEMM_DEFAULT_P 128 | ||
| #define DGEMM_DEFAULT_P 160 | ||
| #define SGEMM_DEFAULT_P 30 |
There was a problem hiding this comment.
How were the default P and Q chosen for 128KB cache?
There was a problem hiding this comment.
#4381 demonstrated values that worked well for a 1MB L2 cache, so I divided that by 8.
If you have a more scientific approach, I'd be happy to hear it 😸
The defaults in param.h now reflect an L2 size of 128KB, and that is scaled based on the actual size.