#
…

2021-04-16

00:18dougallbloom: great solve! that all properly makes sense now :)
00:23snalty has quit [Read error: Connection reset by peer]
00:23snalty has joined #asahi-gpu
00:26nickiminjaj has quit [Ping timeout: 246 seconds]
01:00DragoonAethis has quit [Quit: hej-hej!]
01:00DragoonAethis has joined #asahi-gpu
01:56opticron has quit [Ping timeout: 265 seconds]
01:57opticron has joined #asahi-gpu
02:14odmir has quit [Remote host closed the connection]
02:15odmir has joined #asahi-gpu
02:20odmir has quit [Ping timeout: 240 seconds]
02:46odmir has joined #asahi-gpu
02:51phiologe has quit [Ping timeout: 250 seconds]
02:52phiologe has joined #asahi-gpu
03:09bloomand implemented in the compiler :)
03:09bloomso now all the known float ops are handled
03:35odmir has quit [Remote host closed the connection]
03:56bloomdougall: "# okay, this is very lazy
03:56bloom" wrote a proper C encoder
03:57bloomwill test tomorrow 😴
03:58dougallyeah, that sounds much better :) g'night!
06:57nickiminjaj has joined #asahi-gpu
07:58nickiminjaj has quit [Ping timeout: 240 seconds]
08:00nickiminjaj has joined #asahi-gpu
08:15jix has quit [Quit: WeeChat 3.0]
08:21jix has joined #asahi-gpu
08:35nickiminjaj has quit [Ping timeout: 265 seconds]
08:42nickiminjaj has joined #asahi-gpu
10:57nickiminjaj has quit [Quit: leaving]
11:23qyousef_ has quit [Ping timeout: 265 seconds]
11:25qyousef_ has joined #asahi-gpu
13:10Necrosporus has quit [Ping timeout: 265 seconds]
13:12odmir has joined #asahi-gpu
13:15linkmauve has quit [Ping timeout: 250 seconds]
14:13Necrosporus has joined #asahi-gpu
17:08odmir has quit [Remote host closed the connection]
17:09odmir has joined #asahi-gpu
17:14odmir has quit [Ping timeout: 240 seconds]
17:34linkmauve has joined #asahi-gpu
17:40odmir has joined #asahi-gpu
17:43odmir has quit [Remote host closed the connection]
17:43odmir has joined #asahi-gpu
19:27odmir has quit []
19:39bloomInteresting... Metal implements unary fneg/fabs with bitop (twiddling the sign bit) instead of fadd.
19:41odmir has joined #asahi-gpu
19:44bloom has quit [Ping timeout: 246 seconds]
19:44bloom has joined #asahi-gpu
19:57bloomnote to self: `convert` is only for data type converts
19:57bloomto change sizes, iadd/fadd are used.
19:57bloomalso, modifiers are unsupported (both float and int mods)
20:10snalty has quit [Quit: ZNC 1.8.2 - https://znc.in]
20:12bloomnote: Metal compiler likes bitop_mov more than xor
20:12bloomDunno if that's deliberate.
20:13bloom(was trying to find out if there's an xchg/swap instruction)
21:18bloomThe assembly prduced for precise::sin(x) is horrifying
21:19bloom(~90 instructions)
21:27bloomdefault in Metal is fast::sin(x) which is super cheap
21:49glibcprecise sin is expensive, there's no getting around that (either instruction, or LUT)
22:45bloomthat it is..
22:46bloomwhat's more interesting to me is that they don't use the fast sine anywhere in it
22:46bloomI would expect if you have fast imprecise transcendentals available, you'd use them and feed the estimate into Newton's method or something
23:03chrisfwhat does the precise sin code look like?
…