On Fri, 7 Feb 2025 15:04:23 +0000, Michael S wrote:Years ago I solved this problem for an optimized Ogg Vorbis decoder:res += _mm256_extract_epi32(res0, 0);Simple question:: how would you port this code to a machine
res += _mm256_extract_epi32(res0, 4);
return res;
with a different SIMD instruction set ??
I wrote a set of #defines which wrapped MMX/SSE intrinsics on the x86 side and Motorola's more capable Altivec instructions on the Apple side.
I had to limit myself a tiny bit in a couple of places, as well as expanding a Motorola operation into a pair of SSE instrinsics, but the resulting code still ran faster than all commercialy available libraries on both platforms.
If/when those instrinsics diverge more, then the problem would be significantly harder, but back then both Altivec and SSE used 128-bit registers.
Terje
--
- <Terje.Mathisen at tmsw.no>
"almost all programming can be viewed as an exercise in caching"
Les messages affichés proviennent d'usenet.