4 ms·
> up to 16 load instructions separately That's not the only alternative. Here's AVX2 code that copies 32 bytes with 2 instructions: vmovdqu ymm0, ymmword
by Const-me 6y ago
> up to 16 load instructions separately
That's not the only alternative. Here's AVX2 code that copies 32 bytes with 2 instructions:
vmovdqu ymm0, ymmword ptr[rsi]
vmovdqu ymmword ptr[rdi], ymm0
Pretty sure ARM NEON can do something similar.
- monocasa 6y agoThe point isn't the number of bytes transferred, but being able to dump the integer register file to/from main memory quickly (and issue an indirect jump at the same time!) Also, the gate count where these instructions really shine don't tend to have vector units.
- usr1106 6y agoBut NEON is optional, not available in many "truly" embedded implementations.