Source
- URL: https://x.com/NOTimothyLottes/status/1738744404414759381
- Author: NOTimothyLottes (@NOTimothyLottes)
- Posted: 2023-12-24 02:12:24
Thread
1/ @NOTimothyLottes
Two great graphs from https://chipsandcheese.com/2023/12/23/nintendo-switchs-igpu-maxwell-nerfed-edition/ - Why I always manually vectorize 16-bit code, and also why you always need to manually check disassembly.
Branches: 2023-12-24-BrettRidel-do-you-know-other-equivalent-graphs-or-ressources
2/ @AgileJebrim
@NOTimothyLottes
Why are float16s so much faster than int16s?
3/ @NOTimothyLottes
@AgileJebrim
Same bit width, FMA needs smaller multipler than IMAD. Thus HW mul24 fast paths that never get exposed. But 16-bit is small enough already. I thought on AMD, packed 16-bit int or float are the same perf. Some vendors though try area savings by de-throughputing integer/etc ALUs
4/ @AgileJebrim
@NOTimothyLottes
Oh are you only comparing FMA? If it was just additions, would they be on par?