On Fri, Aug 05, 2022 at 03:04:34PM -0700, Andres Freund wrote:
> But mainly I'd expect to find a difference if the SIMD code were optimized a
> further on the basis of not needing to return the offset. E.g. by
> replacing _mm_packs_epi32 with _mm_or_si128, that's cheaper.
I haven't been able to find a significant difference between the two. If
anything, the _mm_packs_epi* approach actually seems to be slightly faster
in some cases. For something marginally more concrete, I compared the two
in perf-top and saw the following for the relevant instructions:
_mm_packs_epi*:
0.19 │ packssdw %xmm1,%xmm0
0.62 │ packssdw %xmm1,%xmm0
7.14 │ packsswb %xmm1,%xmm0
_mm_or_si128:
1.52 │ por %xmm1,%xmm0
2.05 │ por %xmm1,%xmm0
5.66 │ por %xmm1,%xmm0
I also tried a combined approach where I replaced _mm_packs_epi16 with
_mm_or_si128:
1.16 │ packssdw %xmm1,%xmm0
1.47 │ packssdw %xmm1,%xmm0
8.17 │ por %xmm1,%xmm0
Of course, this simplistic analysis leaves out the impact of the
surrounding instructions, but it seems to support the idea that the
_mm_packs_epi* approach might have a slight edge.
--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com
