#
…

2021-03-27

00:00chrisfDarkShadow44, where's the packed format field?
00:04chrisfthe `u` fields may just be `unknown field X`
00:04DarkShadow44chrisf: Packed format field? You mean to specify what format the data has? AFAIK that's all Fx:F
00:06chrisfi had mapped out the bottom 2 bits of F as `size` but it makes sense if the packed formats use the rest
00:06chrisfyeah, i agree that's what it is then :)
00:09DarkShadow44I assume it's split into Fx:D because of the different instruction length, since all bytes that are cut off are assumed 0
00:09DarkShadow44although, it raises a question: Since registers are specified as Rx:R, maybe bit 31 is for another "cut off"?
00:09DarkShadow44just a theory
00:10chrisfit's an interesting theory
00:11chrisfive never seen the metal compiler do that though
00:11DarkShadow44that doesn't mean we can't manually generate bytecode like that >:D
00:11chrisfwhat is this `L` bit?
00:11DarkShadow44shoul cut off the last two bytes
00:11bpye has quit [Ping timeout: 240 seconds]
00:12chrisfactually i dont think ive seen a non-8-byte load or store
00:13DarkShadow44neither have I, but maybe they planned it into the hardware
00:13DarkShadow44most instructions support such a cut off
00:13chrisfah, `L` is always that, i see the note up the top now.
00:13bpye has joined #asahi-gpu
00:14chrisfi would be careful about things that you never see the metal compiler do -- they may well not actually work in silicon
00:15chrisfthere's always stuff that doesnt work
00:15DarkShadow44Sure thing, but that's where tests come in, no?
00:16chrisfif you can show it does work then great
00:16DarkShadow44I mean, I wouldn't necessarily use it, but it'd be good to know, IMHO
00:16DarkShadow44no unknown bits is best
00:16DarkShadow44although I gotta admit, I don't understand in having the benefit of having variable length instructions like that
00:17DarkShadow44I thought an advantage of, for example, ARM was easier decoding due to fix lengths
00:17chrisftradeoffs
00:18Yuzuvariable length ops = higher density (typically, if done right)
00:18chrisfapple does appear willing to spend area making things not suck
00:19DarkShadow44yeah, easier on the cache
00:21odmir_ has joined #asahi-gpu
00:21chrisfoh, i guess the reason i might not have seen short load/store is i dont have examples of the packed cases
00:22chrisfand so `mask` is always in play in the examples i have
00:22DarkShadow44mh, it's a bit odd that the mask is one of the parts that are cut off
00:22DarkShadow44need to make tests for that
00:24odmir has quit [Ping timeout: 240 seconds]
00:41bloomthat especially matters since G13 is a pure scalar arch, yet it's designed for graphics (vector) workloads
00:41bloom're not careful about code density, you'll blow through your i-cache budget
00:42bloom(Since everything is 4x worse over in graphics land, since you repeat the same instruction a bunch of times.)
00:42bloomDifferent vendors cope with this in different ways.
00:42bloomOlder GPU arches were genuinely vector, some newer ones (GCN, Bifrost) have a fp16vec2 thing going on, special oddball handling goes to Adreno which has a special "repeat N times" modifier :-p
00:43DarkShadow44huh, interesting
00:43DarkShadow44what exactly is that "fp16vec2" thing?
00:43bloomIt's.. cute
00:44bloomSome AMD and Mali GPUs are scalar for 32-bit instructions, but allow 2 channel 16-bit vectors (what AMD markets as rapid packed math or sth)
00:45bloomFor Mali, even have 4 channel 8-bit vectors targeting ML workloads
00:45bloomIt follows naturally from 32-bit arithmetic. I.e. packed 2x16-bit add is the same as 32-bit add up to handling of carry bits.
00:46bloom^ integer
00:46bloomand fp16 is so much cheaper in hw than fp32 that it again works out
00:46bloomSo in theoretical benchmarks, it means a 2x reduction in cycle count for 16-bit vs 32-bit workloads
00:46bloom...in theory. In practice it's a pain for compilers, sometimes more so than plain old vec4 hardware.
00:47bloomCode like `a.xy + b.xz` can't be vectorized (exercise: why not?)
00:48bloomAnyway, G13 does *not* work this way.
00:48DarkShadow44heh, should make things easier
00:48bloomIndeed.
00:49bloomI really hate how inactive I've been
00:49bloomDealing with personal stuff but still :<
00:51DarkShadow44I know that feeling
00:54bloomWhat are your relevant interests ?
01:05DarkShadow44What do you mean?
01:42odmir_ has quit [Remote host closed the connection]
01:42odmir has joined #asahi-gpu
01:47odmir has quit [Ping timeout: 260 seconds]
02:00JusticeEX has joined #asahi-gpu
02:20Emantor has quit [Quit: ZNC - http://znc.in]
02:20Emantor has joined #asahi-gpu
03:26phiologe has quit [Ping timeout: 250 seconds]
03:26phiologe has joined #asahi-gpu
04:03odmir has joined #asahi-gpu
04:08odmir has quit [Ping timeout: 265 seconds]
06:06rwhitby has quit [Ping timeout: 258 seconds]
06:22TheJollyRoger has quit [Remote host closed the connection]
06:22TheJollyRoger has joined #asahi-gpu
06:56TheJollyRoger has quit [Remote host closed the connection]
06:59TheJollyRoger has joined #asahi-gpu
08:21rwhitby has joined #asahi-gpu
10:37zkrx has quit [Ping timeout: 265 seconds]
10:46zkrx has joined #asahi-gpu
10:55rwhitby has quit [Ping timeout: 258 seconds]
11:52JusticeEX has quit [Ping timeout: 240 seconds]
16:06Necrosporus has quit [Read error: Connection reset by peer]
16:06Necrosporus has joined #asahi-gpu
17:07Baughn has quit [Read error: Connection reset by peer]
17:22Necrosporus is now known as Guest59740
17:22Guest59740 has quit [Killed (barjavel.freenode.net (Nickname regained by services))]
17:22Necrosporus has joined #asahi-gpu
17:23Baughn has joined #asahi-gpu
17:27odmir has joined #asahi-gpu
17:37JusticeEX has joined #asahi-gpu
18:34odmir has quit [Remote host closed the connection]
18:35odmir has joined #asahi-gpu
18:42odmir has quit [Ping timeout: 268 seconds]
18:47phiologe has quit [Ping timeout: 250 seconds]
18:47phiologe has joined #asahi-gpu
19:07odmir has joined #asahi-gpu
19:21Baughn has quit [Ping timeout: 260 seconds]
19:39odmir has quit [Ping timeout: 246 seconds]
19:40Baughn has joined #asahi-gpu
20:00odmir has joined #asahi-gpu
20:34odmir has quit [Ping timeout: 260 seconds]
20:38JusticeEX has quit [Ping timeout: 268 seconds]
20:48odmir has joined #asahi-gpu
21:20odmir has quit [Ping timeout: 240 seconds]
21:34odmir has joined #asahi-gpu
22:08odmir has quit [Ping timeout: 268 seconds]
22:20odmir has joined #asahi-gpu
22:34solarkraft has quit [Quit: Bye!]
22:38JusticeEX has joined #asahi-gpu
22:53odmir has quit [Ping timeout: 246 seconds]
23:01odmir has joined #asahi-gpu
…