Re:Beating NumPy matrix multiplication in 150 lines of C
by dzaima on Friday July 05, 2024 @01:59PM
(#40884734)
Oh yeah, I did forget that, despite being separated into uops, sign-extend-mem is still more efficient than literally being separated as such (some other similar things include memory insert/extract, and, perhaps most significantly, broadcast; with various results across intel & AMD). I imagine the memory subsystem is able to simultaneously supply ≤128-bit results to both 128-bit ALU halves, thus avoiding the need for cross-128-bit transfers.
The _mm256_setr_epi32 by itself would indeed be very inefficient, but clang manages to vectorize it[1] to vpaddd+vpsllvd, which require some constant loads (also it generates some weird blends, idk).
[1]: https://godbolt.org/z/7jq4z39GT - L833-847 or so in the assembly, or on L67 in the source, right click → "Reveal linked code"
Indeed, but it reveals something else interesting. On Zen2 and Zen3 processors, the throughput of vpmovsxbd ymm, [...] is more than twice as efficient compared to sign extension from another vector register i.e. vpmovsxbd ymm, xmm
> the original shifting version also has multiple loads
I believe _mm256_setr_epi32 like that is typically compiled into a sequence of vmovd / vpinsrd / vinserti128 instructions. These involve no loads, just multiple instructions assembling the vector from int32 pieces produced in scalar registers.