AVX runtime float/double swizzle small improvement #1189
Conversation
AntoinePrv
changed the title
|
Thanks! nice fine tuning \o/ |
|
Looks like AVX tests are good now, but I'm unsure what the remaining failure is. |
|
LGTM, I'll fix the emulated part, not something you should worry on. Would you mind squashing the last two commits? |
|
@AntoinePrv when you squash, you can also rebase on |
| __m256 swapped = _mm256_permute2f128_ps(self, self, 0x01); // [high | low] | ||
|
|
||
| // normalize mask | ||
| batch<uint32_t, A> half_mask = mask % 4; |
There was a problem hiding this comment.
does this generate different asm actually?
PS I am fine either way. Might be worth having a normalize<value> API so that we can use everywhere that converts value to mask if it is a pow2
There was a problem hiding this comment.
I'm unsure. For a regular integer operation I would say so but with intrinsics I had a doubt.
|
I like this PR! Nice that you found an ulterior way to optimize this! I would suggest that generating the blend mask and the normalization to be pure method or class method that we can use elsewhere. I think we might need them here and there. This also allow to unit test them individually making debugging and development easier. Again, this is just a suggestion. Feel free to ignore it. |
|
Merged as 9d41ad9 (once squashed) |
3 participants
I have a number of of swizzle improvements to suggest, but I am starting small to get better accustomed to xsimd.
What do you make of the following change? My motivation was that
_mm256_permute2f128_psis the most expensive operation (though not sure if that's a problem in a CPU pipeline) so this PR suggest using it only once.It also replaces modulo with a select mask to make sure this is properly optimized.