3 ms·
That's not necessarily a good idea. Small vector uses rely on having nearly no overhead to be faster, so they can't use a generic system.
by astrange 1y ago
That's not necessarily a good idea. Small vector uses rely on having nearly no overhead to be faster, so they can't use a generic system.
- dzaima 1y agoIf the scalable VLEN is the same as the ALU width, which should generally be the target, small vectors would still perform optimally. Of course if you need less than a VLEN-sized vector you're wasting throughput, but that applies just as much when using 128-bit vectors on AVX-capable hardware, and even worse so on AVX-512-capable (which, while double-pumped or equivalent to some extent on most impls, still has 512-bit-exclusive throughput on most).