3 ms·
The problem is that the different SIMD instruction sets are genuinely... different. The basics of “8-bit unsigned add” and similar are possible to abstract over
by Sesse__ 2y ago
The problem is that the different SIMD instruction sets are genuinely... different. The basics of “8-bit unsigned add” and similar are possible to abstract over, but for a lot of cases, you may have to switch your entire algorithm around between different CPUs to get reasonable performance (or even gain over the scalar code at all). There's no way a compiler or SIMD abstraction library will do that for you.