3 ms·
SIMD instructions are a good way to avoid this, even if you only use one column of the operands the fact that it's SIMD means operations have to happen in locks
by pjc50 2y ago
SIMD instructions are a good way to avoid this, even if you only use one column of the operands the fact that it's SIMD means operations have to happen in lockstep across the execution units.
- gpderetta 2y agoDo they? An arch with bultin lane predication (like AVX512) could easily implement wide SIMD on top of narrower ALU and then skip the masked out lanes. Actual runtime would depend on the number of non masked lanes. I'm not up to date on GPU architectures, bit I wouldn't be surprised of they do this sort of stuff.