#
…

2024-08-08

00:05rlittl01 has quit [Read error: Network is unreachable]
00:11anuragrao has quit [Ping timeout: 480 seconds]
00:21anuragrao has joined #asahi-gpu
00:32anuragrao has quit [Ping timeout: 480 seconds]
00:40anuragrao has joined #asahi-gpu
00:48anuragrao has quit [Ping timeout: 480 seconds]
01:00KxCORP58940 has quit [Quit: Bye!]
01:00KxCORP58940 has joined #asahi-gpu
01:03pb17 has quit [Ping timeout: 480 seconds]
01:26anuragrao has joined #asahi-gpu
01:29pb17 has joined #asahi-gpu
01:38anuragrao has quit [Ping timeout: 480 seconds]
02:02ddxtanx has quit [Quit: Konversation terminated!]
02:02ddxtanx has joined #asahi-gpu
02:04jeisom has quit [Quit: Leaving]
02:04ddxtanx_ has joined #asahi-gpu
02:05ddxtanx_ has quit [Remote host closed the connection]
02:06ddxtanx_ has joined #asahi-gpu
02:11ddxtanx has quit [Ping timeout: 480 seconds]
02:31karolherbst_ has joined #asahi-gpu
02:36jn has quit [Remote host closed the connection]
02:37jn has joined #asahi-gpu
02:38karolherbst has quit [Ping timeout: 480 seconds]
03:16pb17 has quit [Ping timeout: 480 seconds]
03:17ddxtanx_ has quit [Remote host closed the connection]
03:19ddxtanx has joined #asahi-gpu
03:23pb17 has joined #asahi-gpu
03:40anuragrao has joined #asahi-gpu
03:57Stary has quit [Quit: ZNC - http://znc.in]
04:00anuragrao has quit [Remote host closed the connection]
04:00anuragrao has joined #asahi-gpu
04:02Stary has joined #asahi-gpu
04:07cylm has quit [Ping timeout: 480 seconds]
04:16anuragrao has quit [Ping timeout: 480 seconds]
04:21chadmed has quit [Quit: Konversation terminated!]
04:21chadmed has joined #asahi-gpu
05:13ddxtanx has quit [Quit: Konversation terminated!]
05:14ddxtanx has joined #asahi-gpu
05:14ddxtanx has quit []
05:15ddxtanx has joined #asahi-gpu
05:34anuragrao has joined #asahi-gpu
05:42anuragrao has quit [Ping timeout: 480 seconds]
06:06pjakobsson has joined #asahi-gpu
06:08john-cabaj has joined #asahi-gpu
06:12karolherbst_ is now known as karolherbst
06:26john-cabaj has quit [Ping timeout: 480 seconds]
06:40cr1901 has quit [Read error: Connection reset by peer]
06:44cr1901 has joined #asahi-gpu
06:47pb17 has quit [Ping timeout: 480 seconds]
06:58pb17 has joined #asahi-gpu
06:59anuragrao has joined #asahi-gpu
07:03yuyichao_ has quit [Ping timeout: 480 seconds]
07:09anuragrao has quit [Ping timeout: 480 seconds]
07:31anuragrao has joined #asahi-gpu
07:36abbe has quit [Remote host closed the connection]
07:37abbe has joined #asahi-gpu
07:39anuragrao has quit [Ping timeout: 480 seconds]
08:18pb17 has quit [Ping timeout: 480 seconds]
08:23chadmed has quit [Quit: Konversation terminated!]
08:23chadmed has joined #asahi-gpu
08:27pb17 has joined #asahi-gpu
08:32anuragrao has joined #asahi-gpu
08:43yuyichao_ has joined #asahi-gpu
08:45anuragrao has quit [Ping timeout: 480 seconds]
09:01anuragrao has joined #asahi-gpu
09:04yuyichao has joined #asahi-gpu
09:07yuyichao_ has quit [Ping timeout: 480 seconds]
09:08yuyichao_ has joined #asahi-gpu
09:11anuragrao has quit [Remote host closed the connection]
09:11anuragrao has joined #asahi-gpu
09:12yuyichao has quit [Ping timeout: 480 seconds]
09:22anuragrao has quit [Ping timeout: 480 seconds]
09:24karolherbst"Testing "/" cl_int Verification failed at element 1749980: 0x2ba97a73 / 0xfa8d9df5 = 0xfffffff8, got 0xffffffef" noooo
09:26anuragrao has joined #asahi-gpu
09:41pb17 has quit [Ping timeout: 480 seconds]
09:41pb17 has joined #asahi-gpu
09:49anuragrao has quit [Ping timeout: 480 seconds]
09:52anuragrao has joined #asahi-gpu
10:00anuragrao has quit [Ping timeout: 480 seconds]
10:02ddxtanx_ has joined #asahi-gpu
10:07ddxtanx has quit [Ping timeout: 480 seconds]
10:14karolherbstalyssa: okay, your block/blit stuff works for fp32, but not for fp16
10:15karolherbstimage format that means
10:15karolherbstbut mhh.. only fails for 1D and 2D
10:15karolherbstahh.. and 1Darray
10:32anuragrao has joined #asahi-gpu
10:32flokli has quit [Ping timeout: 480 seconds]
10:40flokli has joined #asahi-gpu
10:54chadmed has quit [Read error: Connection reset by peer]
10:54chadmed has joined #asahi-gpu
11:06anuragrao has quit [Ping timeout: 480 seconds]
11:11pb17 has quit [Ping timeout: 480 seconds]
11:17anuragrao has joined #asahi-gpu
11:25anuragrao has quit [Ping timeout: 480 seconds]
11:28pb17 has joined #asahi-gpu
11:57karolherbstokay.. I think the int bug above is a rusticl one
11:57karolherbstsomething with block/grid sizes
12:32jeisom has joined #asahi-gpu
13:07karolherbstoh no.. it's a real bug..
13:09karolherbstprobably something wrong in nir_lower_idiv
13:12karolherbstI've asked austriancoder if it also happens on etnaviv...
13:13karolherbstthat's gonna be pain to debug
13:14pb17 has quit [Ping timeout: 480 seconds]
13:17karolherbstat 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...
13:23alyssawhat test?
13:23alyssaand how do I reproduce it?
13:24karolherbstin the CL CTS it's integer_ops/test_integer_ops uint_math int_math
13:25karolherbstthe piglit one helps, because you won't have to run it for a minute until hitting it :D
13:26alyssa.. where is that in piglit?
13:26karolherbstneeds -DPIGLIT_BUILD_CL_TESTS=ON
13:27karolherbstand then just bin/cl-program-tester
13:30alyssa Could NOT find PythonNumpy (missing: PythonNumpy_STATUS) (Required is at
13:30alyssaexcuse me
13:33pb17 has joined #asahi-gpu
13:36* alyssarebuilds piglit
13:59anuragrao has joined #asahi-gpu
14:16* alyssabuilds rusticl
14:26karolherbsthuh.. this blit bug is weird...
14:26john-cabaj has joined #asahi-gpu
14:31* alyssapoking at idiv
14:36alyssakarolherbst: ok, this is interesting
14:36alyssathe problem is the value of `rcp` at the start
14:36alyssaint((1.0 / float(b)) * (4294966784.0))
14:36alyssaagx has an off-by-1 compared to my cpu
14:37alyssaand that off-by-1 propagates to the totally wrong final rsult
14:37alyssaIDK if this is an AGX backend bug or a nir_lower_idiv bug, strictly
14:37karolherbstmaybe something with rounding?
14:39alyssaplausible
14:39alyssaI mean I'm just plumbing through the hardware frcp here
14:40karolherbstor just the hw not being precise enough
14:40alyssaalthough Metal's docs claim that 1.0 / x is correctly rounded
14:40karolherbstand we can safely assume they emit it directly without any funky lowering?
14:42alyssahm no that should be ok
14:44alyssafrcp(0x4C4565C8)
14:45alyssashould be 32A60000
14:45alyssabut AGX gives 32A60001
14:48alyssaoh, jeez, ugh!
14:48alyssakarolherbst: frcp with -no-fast-math has lowering indeed. ughhh.
14:49karolherbstjust some scaling like nir_scale_fdiv or something more complex?
14:50alyssamore complex, gimme a few mins
14:57mrosorensen1 has joined #asahi-gpu
14:57mrosorensen has quit [Ping timeout: 480 seconds]
15:00alyssathis is... subtle
15:00alyssathere are two parts to the lowering
15:00alyssaone part is a scaling like scale_fdiv, to reduce the domain for large denomintaors
15:00alyssa*denominators
15:00alyssathe other part though I don't quite understand yet
15:01alyssaunder a certain condition it'll adjust the output as
15:01alyssarcp := fma(fma(-input, rcp, 1.0), rcp, rcp)
15:01karolherbstI've seen that one
15:01karolherbstthat's uhhh
15:01karolherbsthow is that called
15:01alyssa----wait is that a newton raphson step
15:01alyssagot it in 2
15:02alyssamy aborted math degree came in handy!
15:03karolherbstthere is a good thing about this.. I think libclc already has the code for the lowering... but...
15:04alyssaneeds to be in-tree for idiv lowering anyhow
15:04alyssaI'm also wondering if there's a way to make lower idiv tolerate the error, for faster overall
15:05alyssamaybe not
15:05mrosorensen14 has joined #asahi-gpu
15:06mrosorensen1 has quit [Ping timeout: 480 seconds]
15:09karolherbst(which gets real funny real quick when you also need fma lowering)
15:17cylm has joined #asahi-gpu
15:36alyssakarolherbst: ook. so what I realized is real funny is that CL CTS is going to hit libclc for frcp
15:36alyssabut you need exact frcp! in the backend *anyway* for correct idiv
15:37alyssaso that seems.. bad
15:38alyssaalthough i dont actually see where we call libclc frcp
15:40alyssaI guess libclc doesn't do divide?
15:40karolherbstlibclc doesn't have a frcp opcode
15:41karolherbstthe spirv env just defines precision for "1.0 / x" explicitly
15:41alyssaah. but does the CL CTS test it.
15:42alyssareciprocal is suspiciously absent from math_bruteforce
15:42karolherbstyeah.. I don't think it's tested besides through divide
15:43alyssaI would like math_bruteforce to test the full 2^32 input space because otherwise I have no confidence in writing the lowering correctly
15:43karolherbstreference_recip
15:43karolherbstis that even used
15:44karolherbst //ENTRY(reciprocal, 1.0f, 1.0f, FTZ_OFF, unaryF),
15:45karolherbstonly disabled test besides tgamma
15:46karolherbstmaybe enable it and see if it works?
15:47alyssawas slightly more complicated than that but yeah, added the test and indeed we're failing for large inputs
15:47alyssai can work with this :+1:
15:47alyssathough not right now I have a big todo list this week
15:47alyssai'll get to it soon
15:48karolherbstI'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...
15:59karolherbstalyssa: 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
16:01alyssawhich test case is that?
16:01alyssaI think I know what's happening but I'd like to confirm
16:01karolherbstimages/clCopyImage/test_cl_copy_images 1D CL_HALF_FLOAT
16:01karolherbstyou need to skip the first blit
16:01karolherbstthe second hits the bug
16:01karolherbst(the first is a full copy)
16:02alyssaor I can just give a patch in a bit
16:02alyssai was busy reading french wikipedia's page on newton's method :P
16:06alyssayeah i need a break too many things happening right now
16:07karolherbsthave a good rest
16:25anuragrao has quit [Ping timeout: 480 seconds]
16:29anuragrao has joined #asahi-gpu
16:37anuragrao has quit [Ping timeout: 480 seconds]
16:37pb17 has quit [Ping timeout: 480 seconds]
16:38anuragrao has joined #asahi-gpu
16:41pb17 has joined #asahi-gpu
17:41chadmed has quit [Quit: Konversation terminated!]
17:41chadmed has joined #asahi-gpu
17:46DarkShadow4444 has joined #asahi-gpu
17:48DarkShadow44 has quit [Ping timeout: 480 seconds]
17:54anuragrao has quit [Ping timeout: 480 seconds]
17:56anuragrao has joined #asahi-gpu
18:04anuragrao has quit [Ping timeout: 480 seconds]
18:05cyrinux has quit []
18:07cyrinux has joined #asahi-gpu
18:22alih has quit []
19:16flokli has quit [Ping timeout: 480 seconds]
19:19john-cabaj has quit [Remote host closed the connection]
19:19john-cabaj has joined #asahi-gpu
19:25flokli has joined #asahi-gpu
20:24alyssa has quit [Quit: alyssa]
20:56john-cabaj has quit [Ping timeout: 480 seconds]
21:51anuragrao has joined #asahi-gpu
21:55pb17 has quit [Ping timeout: 480 seconds]
21:59anuragrao has quit [Ping timeout: 480 seconds]
22:02ddxtanx_ has quit [Quit: Konversation terminated!]
22:02ddxtanx_ has joined #asahi-gpu
22:09pb17 has joined #asahi-gpu
23:27pb17 has quit [Ping timeout: 480 seconds]
23:33pb17 has joined #asahi-gpu
…