#asahi-gpu on 2024-08-08 — irc logs at oftc.irclog.whitequark.org

2024-07-16 04:52 ChanServ changed the topic of #asahi-gpu to: Asahi Linux GPU development (no user support, NO binary reversing) | Keep things on topic | GitHub: https://alx.sh/g | Wiki: https://alx.sh/w | Logs: https://alx.sh/l/asahi-gpu

00:05 rlittl01 has quit [Read error: Network is unreachable]

00:11 anuragrao has quit [Ping timeout: 480 seconds]

00:21 anuragrao has joined #asahi-gpu

00:32 anuragrao has quit [Ping timeout: 480 seconds]

00:40 anuragrao has joined #asahi-gpu

00:48 anuragrao has quit [Ping timeout: 480 seconds]

01:00 KxCORP58940 has quit [Quit: Bye!]

01:00 KxCORP58940 has joined #asahi-gpu

01:03 pb17 has quit [Ping timeout: 480 seconds]

01:26 anuragrao has joined #asahi-gpu

01:29 pb17 has joined #asahi-gpu

01:38 anuragrao has quit [Ping timeout: 480 seconds]

02:02 ddxtanx has quit [Quit: Konversation terminated!]

02:02 ddxtanx has joined #asahi-gpu

02:04 jeisom has quit [Quit: Leaving]

02:04 ddxtanx_ has joined #asahi-gpu

02:05 ddxtanx_ has quit [Remote host closed the connection]

02:06 ddxtanx_ has joined #asahi-gpu

02:11 ddxtanx has quit [Ping timeout: 480 seconds]

02:31 karolherbst_ has joined #asahi-gpu

02:36 jn has quit [Remote host closed the connection]

02:37 jn has joined #asahi-gpu

02:38 karolherbst has quit [Ping timeout: 480 seconds]

03:16 pb17 has quit [Ping timeout: 480 seconds]

03:17 ddxtanx_ has quit [Remote host closed the connection]

03:19 ddxtanx has joined #asahi-gpu

03:23 pb17 has joined #asahi-gpu

03:40 anuragrao has joined #asahi-gpu

03:57 Stary has quit [Quit: ZNC - http://znc.in]

04:00 anuragrao has quit [Remote host closed the connection]

04:00 anuragrao has joined #asahi-gpu

04:02 Stary has joined #asahi-gpu

04:07 cylm has quit [Ping timeout: 480 seconds]

04:16 anuragrao has quit [Ping timeout: 480 seconds]

04:21 chadmed has quit [Quit: Konversation terminated!]

04:21 chadmed has joined #asahi-gpu

05:13 ddxtanx has quit [Quit: Konversation terminated!]

05:14 ddxtanx has joined #asahi-gpu

05:14 ddxtanx has quit []

05:15 ddxtanx has joined #asahi-gpu

05:34 anuragrao has joined #asahi-gpu

05:42 anuragrao has quit [Ping timeout: 480 seconds]

06:06 pjakobsson has joined #asahi-gpu

06:08 john-cabaj has joined #asahi-gpu

06:12 karolherbst_ is now known as karolherbst

06:26 john-cabaj has quit [Ping timeout: 480 seconds]

06:40 cr1901 has quit [Read error: Connection reset by peer]

06:44 cr1901 has joined #asahi-gpu

06:47 pb17 has quit [Ping timeout: 480 seconds]

06:58 pb17 has joined #asahi-gpu

06:59 anuragrao has joined #asahi-gpu

07:03 yuyichao_ has quit [Ping timeout: 480 seconds]

07:09 anuragrao has quit [Ping timeout: 480 seconds]

07:31 anuragrao has joined #asahi-gpu

07:36 abbe has quit [Remote host closed the connection]

07:37 abbe has joined #asahi-gpu

07:39 anuragrao has quit [Ping timeout: 480 seconds]

08:18 pb17 has quit [Ping timeout: 480 seconds]

08:23 chadmed has quit [Quit: Konversation terminated!]

08:23 chadmed has joined #asahi-gpu

08:27 pb17 has joined #asahi-gpu

08:32 anuragrao has joined #asahi-gpu

08:43 yuyichao_ has joined #asahi-gpu

08:45 anuragrao has quit [Ping timeout: 480 seconds]

09:01 anuragrao has joined #asahi-gpu

09:04 yuyichao has joined #asahi-gpu

09:07 yuyichao_ has quit [Ping timeout: 480 seconds]

09:08 yuyichao_ has joined #asahi-gpu

09:11 anuragrao has quit [Remote host closed the connection]

09:11 anuragrao has joined #asahi-gpu

09:12 yuyichao has quit [Ping timeout: 480 seconds]

09:22 anuragrao has quit [Ping timeout: 480 seconds]

09:24 <karolherbst> "Testing "/" cl_int Verification failed at element 1749980: 0x2ba97a73 / 0xfa8d9df5 = 0xfffffff8, got 0xffffffef" noooo

09:26 anuragrao has joined #asahi-gpu

09:41 pb17 has quit [Ping timeout: 480 seconds]

09:41 pb17 has joined #asahi-gpu

09:49 anuragrao has quit [Ping timeout: 480 seconds]

09:52 anuragrao has joined #asahi-gpu

10:00 anuragrao has quit [Ping timeout: 480 seconds]

10:02 ddxtanx_ has joined #asahi-gpu

10:07 ddxtanx has quit [Ping timeout: 480 seconds]

10:14 <karolherbst> alyssa: okay, your block/blit stuff works for fp32, but not for fp16

10:15 <karolherbst> image format that means

10:15 <karolherbst> but mhh.. only fails for 1D and 2D

10:15 <karolherbst> ahh.. and 1Darray

10:32 anuragrao has joined #asahi-gpu

10:32 flokli has quit [Ping timeout: 480 seconds]

10:40 flokli has joined #asahi-gpu

10:54 chadmed has quit [Read error: Connection reset by peer]

10:54 chadmed has joined #asahi-gpu

11:06 anuragrao has quit [Ping timeout: 480 seconds]

11:11 pb17 has quit [Ping timeout: 480 seconds]

11:17 anuragrao has joined #asahi-gpu

11:25 anuragrao has quit [Ping timeout: 480 seconds]

11:28 pb17 has joined #asahi-gpu

11:57 <karolherbst> okay.. I think the int bug above is a rusticl one

11:57 <karolherbst> something with block/grid sizes

12:32 jeisom has joined #asahi-gpu

13:07 <karolherbst> oh no.. it's a real bug..

13:09 <karolherbst> probably something wrong in nir_lower_idiv

13:12 <alyssa> uhoh

13:12 <karolherbst> I've asked austriancoder if it also happens on etnaviv...

13:13 <karolherbst> that's gonna be pain to debug

13:14 pb17 has quit [Ping timeout: 480 seconds]

13:17 <karolherbst> at least reverting 65e431e61a3bace7e50c11d699880ae860f76133 doesn't help.. but maybe the other changes might? But I'm not in the mood of doing reverts across the big nir rename...

13:23 <alyssa> what test?

13:23 <alyssa> and how do I reproduce it?

13:24 <karolherbst> with piglits cl-program-tester: https://gist.githubusercontent.com/karolherbst/f49e8f5091887d60bd88da58f54d5b7e/raw/c55ef2d681c3eb6481a0645cf7beb71556181157/test.cl

13:24 <karolherbst> in the CL CTS it's integer_ops/test_integer_ops uint_math int_math

13:25 <alyssa> uhhh ok

13:25 <karolherbst> the piglit one helps, because you won't have to run it for a minute until hitting it :D

13:26 <alyssa> .. where is that in piglit?

13:26 <karolherbst> needs -DPIGLIT_BUILD_CL_TESTS=ON

13:27 <karolherbst> and then just bin/cl-program-tester

13:28 <alyssa> ah

13:30 <alyssa> Could NOT find PythonNumpy (missing: PythonNumpy_STATUS) (Required is at

13:30 <alyssa> excuse me

13:30 <karolherbst> how rude

13:33 pb17 has joined #asahi-gpu

13:36 * alyssa rebuilds piglit

13:59 anuragrao has joined #asahi-gpu

14:16 * alyssa builds rusticl

14:26 <karolherbst> huh.. this blit bug is weird...

14:26 john-cabaj has joined #asahi-gpu

14:31 * alyssa poking at idiv

14:36 <alyssa> karolherbst: ok, this is interesting

14:36 <alyssa> the problem is the value of `rcp` at the start

14:36 <alyssa> int((1.0 / float(b)) * (4294966784.0))

14:36 <alyssa> agx has an off-by-1 compared to my cpu

14:37 <alyssa> and that off-by-1 propagates to the totally wrong final rsult

14:37 <karolherbst> funky

14:37 <alyssa> IDK if this is an AGX backend bug or a nir_lower_idiv bug, strictly

14:37 <karolherbst> maybe something with rounding?

14:39 <alyssa> plausible

14:39 <alyssa> I mean I'm just plumbing through the hardware frcp here

14:40 <karolherbst> or just the hw not being precise enough

14:40 <alyssa> although Metal's docs claim that 1.0 / x is correctly rounded

14:40 <karolherbst> and we can safely assume they emit it directly without any funky lowering?

14:40 <alyssa> yes

14:41 <alyssa> er wait

14:42 <alyssa> hm no that should be ok

14:44 <alyssa> yeah..

14:44 <alyssa> frcp(0x4C4565C8)

14:45 <alyssa> should be 32A60000

14:45 <alyssa> but AGX gives 32A60001

14:48 <alyssa> oh, jeez, ugh!

14:48 <alyssa> karolherbst: frcp with -no-fast-math has lowering indeed. ughhh.

14:48 <alyssa> cute.

14:48 <karolherbst> :3

14:49 <karolherbst> just some scaling like nir_scale_fdiv or something more complex?

14:50 <alyssa> more complex, gimme a few mins

14:57 mrosorensen1 has joined #asahi-gpu

14:57 mrosorensen has quit [Ping timeout: 480 seconds]

15:00 <alyssa> this is... subtle

15:00 <alyssa> there are two parts to the lowering

15:00 <alyssa> one part is a scaling like scale_fdiv, to reduce the domain for large denomintaors

15:00 <alyssa> *denominators

15:00 <alyssa> the other part though I don't quite understand yet

15:01 <alyssa> under a certain condition it'll adjust the output as

15:01 <alyssa> rcp := fma(fma(-input, rcp, 1.0), rcp, rcp)

15:01 <karolherbst> ohh wait

15:01 <karolherbst> I've seen that one

15:01 <karolherbst> that's uhhh

15:01 <karolherbst> how is that called

15:01 <alyssa> ----wait is that a newton raphson step

15:01 <karolherbst> yep

15:01 <karolherbst> that one

15:01 <alyssa> lolz

15:01 <alyssa> ok

15:01 <alyssa> got it in 2

15:02 <alyssa> :P

15:02 <alyssa> my aborted math degree came in handy!

15:03 <karolherbst> there is a good thing about this.. I think libclc already has the code for the lowering... but...

15:03 <karolherbst> somewhere

15:03 <karolherbst> maybe not

15:04 <alyssa> needs to be in-tree for idiv lowering anyhow

15:04 <karolherbst> yeah...

15:04 <alyssa> I'm also wondering if there's a way to make lower idiv tolerate the error, for faster overall

15:05 <alyssa> maybe not

15:05 mrosorensen14 has joined #asahi-gpu

15:06 mrosorensen1 has quit [Ping timeout: 480 seconds]

15:07 <karolherbst> libclc be like: https://github.com/llvm/llvm-project/blob/main/libclc/generic/lib/math/acos.cl#L128

15:09 <karolherbst> (which gets real funny real quick when you also need fma lowering)

15:17 cylm has joined #asahi-gpu

15:36 <alyssa> karolherbst: ook. so what I realized is real funny is that CL CTS is going to hit libclc for frcp

15:36 <alyssa> but you need exact frcp! in the backend *anyway* for correct idiv

15:37 <alyssa> so that seems.. bad

15:38 <alyssa> although i dont actually see where we call libclc frcp

15:40 <alyssa> I guess libclc doesn't do divide?

15:40 <karolherbst> libclc doesn't have a frcp opcode

15:40 <karolherbst> ehh clc

15:41 <karolherbst> the spirv env just defines precision for "1.0 / x" explicitly

15:41 <alyssa> ah. but does the CL CTS test it.

15:42 <alyssa> reciprocal is suspiciously absent from math_bruteforce

15:42 <karolherbst> yeah.. I don't think it's tested besides through divide

15:42 <alyssa> :=/

15:43 <alyssa> I would like math_bruteforce to test the full 2^32 input space because otherwise I have no confidence in writing the lowering correctly

15:43 <karolherbst> reference_recip

15:43 <karolherbst> wait..

15:43 <karolherbst> is that even used

15:44 <karolherbst> ohhh

15:44 <karolherbst> reciprocal

15:44 <karolherbst> //ENTRY(reciprocal, 1.0f, 1.0f, FTZ_OFF, unaryF),

15:44 <karolherbst> 🙃

15:44 <alyssa> real

15:45 <karolherbst> only disabled test besides tgamma

15:46 <karolherbst> maybe enable it and see if it works?

15:47 <alyssa> was slightly more complicated than that but yeah, added the test and indeed we're failing for large inputs

15:47 <alyssa> i can work with this :+1:

15:47 <alyssa> though not right now I have a big todo list this week

15:47 <karolherbst> yeah, fair

15:47 <alyssa> i'll get to it soon

15:48 <karolherbst> I'll just be busy figuring out what's wrong with the blit, because it's not the actual copy being all wrong, so it's a bit confusing...

15:49 <alyssa> yeehaw

15:59 <karolherbst> alyssa: mhhh.. okay, maybe it's the blit.. but what seems to happen is, that the image gets initialized with 0xffff over it's entire size, then an API copy happens at offset srcOffset = 79, destOffset = 121 and then the destination becomes 0x7e00 starting at offset 96.... but 121+ is as expected copied from the source

16:00 <alyssa> ok

16:01 <alyssa> which test case is that?

16:01 <alyssa> I think I know what's happening but I'd like to confirm

16:01 <karolherbst> images/clCopyImage/test_cl_copy_images 1D CL_HALF_FLOAT

16:01 <karolherbst> you need to skip the first blit

16:01 <karolherbst> the second hits the bug

16:01 <karolherbst> (the first is a full copy)

16:02 <alyssa> or I can just give a patch in a bit

16:02 <alyssa> i was busy reading french wikipedia's page on newton's method :P

16:02 <karolherbst> :D

16:02 <karolherbst> fair

16:06 <alyssa> yeah i need a break too many things happening right now

16:07 <karolherbst> have a good rest

16:25 anuragrao has quit [Ping timeout: 480 seconds]

16:29 anuragrao has joined #asahi-gpu

16:37 anuragrao has quit [Ping timeout: 480 seconds]

16:37 pb17 has quit [Ping timeout: 480 seconds]

16:38 anuragrao has joined #asahi-gpu

16:41 pb17 has joined #asahi-gpu

17:41 chadmed has quit [Quit: Konversation terminated!]

17:41 chadmed has joined #asahi-gpu

17:46 DarkShadow4444 has joined #asahi-gpu

17:48 DarkShadow44 has quit [Ping timeout: 480 seconds]

17:54 anuragrao has quit [Ping timeout: 480 seconds]

17:56 anuragrao has joined #asahi-gpu

18:04 anuragrao has quit [Ping timeout: 480 seconds]

18:05 cyrinux has quit []

18:07 cyrinux has joined #asahi-gpu

18:22 alih has quit []

19:16 flokli has quit [Ping timeout: 480 seconds]

19:19 john-cabaj has quit [Remote host closed the connection]

19:19 john-cabaj has joined #asahi-gpu

19:25 flokli has joined #asahi-gpu

20:24 alyssa has quit [Quit: alyssa]

20:56 john-cabaj has quit [Ping timeout: 480 seconds]

21:51 anuragrao has joined #asahi-gpu

21:55 pb17 has quit [Ping timeout: 480 seconds]

21:59 anuragrao has quit [Ping timeout: 480 seconds]

22:02 ddxtanx_ has quit [Quit: Konversation terminated!]

22:02 ddxtanx_ has joined #asahi-gpu

22:09 pb17 has joined #asahi-gpu

23:27 pb17 has quit [Ping timeout: 480 seconds]

23:33 pb17 has joined #asahi-gpu