#asahi · 2021-02-01 16:45
segherso your choice often is "extra fast code, or correct code"#asahi · 2021-02-01 16:45
segherthe nasty thing is that many algos very much *depend* on operation order#asahi · 2021-02-01 16:44
segherthe very nasty thing about such accelerators (like MMA on Power) is that they implicitly change operation order, so they only can work with -ffast-math or similar#asahi · 2021-02-01 16:24
segherjix, arnd: so AMX is for training neural nets, and ANE is for running them, it seems#asahi · 2021-02-01 16:13
segherso ah, one TLA and another#asahi · 2021-02-01 16:11
segherjix: no, i cannot keep up with the backscroll here... i didn't see arnd even! hi arnd!#asahi · 2021-02-01 15:39
segher(needs about six times as much hardware, and has double the latency)#asahi · 2021-02-01 15:38
seghermaybe it does though... that would be quite surprising#asahi · 2021-02-01 15:38
segherdavidrysk: i do not know. i simply say that afaik the apple thing does not do DP float#asahi · 2021-02-01 15:29
segheralso brainfloat, and maybe some HP#asahi · 2021-02-01 15:23
segherjix: it is SP float only, so useless for almost all BLAS stuff#asahi · 2021-02-01 14:57
segher"for ai"#asahi · 2021-02-01 14:57
segherjix: ah, their convolution stuff#asahi · 2021-02-01 14:48
segherjix: i don't really know what AMX is :-)#asahi · 2021-02-01 14:48
segheri suppose there *is* a way to trap and emulate those on any arm implementation, it is so very useful, but who knows#asahi · 2021-02-01 14:47
segheranything that uses any new instruction will not run on older hardware#asahi · 2021-02-01 14:46
segher"what kind of stuff would require recompiling all userland?"#asahi · 2021-02-01 14:46
segheri was answering the general question :-)#asahi · 2021-02-01 14:39
seghermarcan: and "deeply" sounds lessmlikely than it is... if you simply have some new insn that compilers can use usefully, you already get close to 100% of programs using it#asahi · 2021-02-01 14:37
seghernot using x86 anymore