SIMD performance comparison

Open chasingegg opened this issue 3 years ago • 4 comments

Hi, it is really useful to see how SIMD works at 4-bit PQ by simulate_kernels_PQ4, but I'm wondering why the first attempt is worse than the scann implementation since it seems that the first attempt has fewer loops. And furthermore, do we have a real performance comparison instead of this simulation?

Nov 10 '22 12:11 chasingegg

I did the comparison (in C of course) and it was much slower than the code layout used in scann.

Nov 11 '22 11:11 mdouze

Another question... why we have such code layout? 0, 8, 1, 9.... will the sequence affect the efficiency? Why don't we set 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 in the first 4bit of each byte in one 128 bit register.

Dec 03 '22 15:12 chasingegg

@mdouze Do you know the reason...?

Dec 15 '22 06:12 chasingegg

I also got the same result that it is more slower with simd

Aug 04 '23 09:08 ryankeke