#asahi-gpu · 2023-06-29 14:28
alyssalina: ^#asahi-gpu · 2023-06-29 14:28
alyssaand see if the shaders diff (other than adding an atomic, obviously) and/or if the structures differ#asahi-gpu · 2023-06-29 14:28
alyssathe way to go about checking this is to diff the agxdecode on t600x for a compute kernel without atomic + compute kernel with#asahi-gpu · 2023-06-29 14:28
alyssa3. pass a flag to the kernel telling the kernel to pass a flag to the hardware to force things together?#asahi-gpu · 2023-06-29 14:28
alyssa2. set a flag in the CDM dispatch structs forcing things to stay together?#asahi-gpu · 2023-06-29 14:27
alyssa1. add more barriers?#asahi-gpu · 2023-06-29 14:27
alyssaand so the compiler needs to report whether atomics are used, and if so either#asahi-gpu · 2023-06-29 14:27
alyssabecause the atomics aren't coherent across clusters#asahi-gpu · 2023-06-29 14:27
alyssaor need to have barriers on multi-cluster parts that aren't needed on single-cluster parts#asahi-gpu · 2023-06-29 14:27
alyssaI wonder if compute kernels with atomics need to be pinned to a single cluster for correct results#asahi-gpu · 2023-06-29 14:26
alyssaOr, geez, I wonder#asahi-gpu · 2023-06-29 14:26
alyssathinking either the kernel is configuring the CDM badly on t600x, or Mesa is doing something silly, or..#asahi-gpu · 2023-06-29 14:23
alyssaflush_finish.* are the stupid tests that we skip in CI#asahi-gpu · 2023-06-29 14:22
alyssaYes.#asahi-gpu · 2023-06-29 14:22
jannauis it expected that dEQP-GLES3.functional.flush_finish.flush_wait takes a long time (several seconds) and makes desktop use slow?#asahi-gpu · 2023-06-29 14:22
alyssathis really could be either Mesa or kernel, tbh#asahi-gpu · 2023-06-29 14:15
jannauerr, the other compute shader tests in the same group pass#asahi-gpu · 2023-06-29 14:14
jannauyes, the other tests in the same group fail, also only ~60% of the atomic_counter tests with multiple threads fail#asahi-gpu · 2023-06-29 14:10
alyssaAll of these tests use atomics.#asahi-gpu · 2023-06-29 14:09
alyssathere are also single invocation/group versions of all those, which I assume are passing#asahi-gpu · 2023-06-29 14:09
alyssaThey're all compute shader tests, but tests with multiple invocations/groups#asahi-gpu · 2023-06-29 14:09
alyssajannau: Those tests at the end are interesting#asahi-gpu · 2023-06-29 14:07
jannauat least this ones: https://paste.debian.net/1284465/ - rerunning the cts atm due to mesa config misshap on my side#asahi-gpu · 2023-06-29 13:57
alyssaspicy.#asahi-gpu · 2023-06-29 13:56
jannauresults on m1 max looked very similar or identical#asahi-gpu · 2023-06-29 13:55
jannaulina: quite a few atomic_counter tests#asahi-gpu · 2023-06-29 13:41
linaWhat is failing?#asahi-gpu · 2023-06-29 13:05
_jannau__that is with asahi.initial_tvb_size=64 and without unknown event messages#asahi-gpu · 2023-06-29 13:04
_jannau__cts results are worse on t6002 compared to t8103, testing now t6001#asahi-gpu · 2023-06-29 13:03
_jannau__lina: mac studio m1 ultra, 48 gpu cores#asahi-gpu · 2023-06-29 13:00
linajannau: Okay, so it's the TVB size thing... what machine is this?#asahi-gpu · 2023-06-29 08:42
jannautests pass with asahi.initial_tvb_size=64#asahi-gpu · 2023-06-29 08:35
jannauall precision/matrixcompmult/highp_fragment/mat3* tests fail but mat2* / mat4* don't, I doubt the tests use MSAA#asahi-gpu · 2023-06-29 08:12
linaAlso I know the initial TVB size is different with clustering, so it wouldn't surprise me if this breaks on Ultra only (or on the whole Pro/Max/Ultra series)#asahi-gpu · 2023-06-29 08:09
linaI think Ultra really is mostly the same execution path when it comes to rendering, just initialization is probably different.#asahi-gpu · 2023-06-29 08:09
linaRe CTS runs, I think it's largely the same execution paths for all of a series, plus it's not like we're submitting conformance separately for each firmware version either...#asahi-gpu · 2023-06-29 08:07
linaDo you know if the hanging test uses MSAA? I wouldn't be surprised if I need a *sample_count factor in there or something#asahi-gpu · 2023-06-29 08:06
linaI think the logic to compute the minimum initial TVB size is not complete#asahi-gpu · 2023-06-29 08:05
linaIs it reproducible if you increase asahi.initial_tvb_size=64 or something like that?#asahi-gpu · 2023-06-29 08:04
linajannau: "Unknown event" is something we can't handle, usually dynamic TVB allocation or similar, so it's no surprise it would hang after that.#asahi-gpu · 2023-06-29 07:49
jannauso it could be either a mesa driver bug, a wayland wsi bug or a VK-GL-CTS bug#asahi-gpu · 2023-06-29 07:48
jannautest passes in the cts run with alyssa's es31 branch and VK-GL-CTS opengl-es-cts-3.2.9.3 with x11_egl#asahi-gpu · 2023-06-29 07:25
jannau'deqp-gles3/functional/shaders/builtin_functions/precision/matrixcompmult/highp_fragment/mat3' seems to be the hanging test. I do not want to disturb the cts run so I will try to confirm that later#asahi-gpu · 2023-06-29 06:12
jannauare "Unknown event message" interesting? There is one (reproducible) ~100 seconds prior#asahi-gpu · 2023-06-29 06:11
jannauseems related to "asahi 406400000.gpu: QueueJob 651041: Job timed out on the DRM scheduler, things will probably break (ran: true)"#asahi-gpu · 2023-06-29 06:11
jannaukernel includes yesterday's drm sched fix#asahi-gpu · 2023-06-29 06:10
jannaulina: piglit run deqp_gles3 seems to hang at test 11377, using VK-GL-CTS opengl-es-cts-3.2.9 DEQP_TARGET=wayland on m1 ultra#asahi-gpu · 2023-06-29 05:54
jannauI'd be surprised if "there are twice as many clusters" is the only difference on the ultra#asahi-gpu · 2023-06-29 05:53
jannauI was under the impression that we wanted to test t6000/t6001/t6002 separately under the assumption that we effectively test Apple's firmware as well and probably hit different execution paths #asahi-gpu · 2023-06-29 03:29
linaThe actual driver codepath there is just a GFP_KERNEL alloc with lots of memory free/cached is failing, which is definitely not a problem with the driver itself, but I wouldn't be surprised if the 4K/16K shmem thing is breaking memory management...