03:41 <alyssa----> I feel like my school spirit is at an all time high!

03:41 <alyssa----> ~RB

15:56 <bbrezillon> alyssa----: Hi

15:57 <bbrezillon> are you using https://gitlab.freedesktop.org/panfrost/midgard-performance-analyze to dump the counter vals?

16:01 <bbrezillon> are you sure of the block layout used here https://gitlab.freedesktop.org/panfrost/midgard-performance-analyze/blob/master/analyze.c#L45 ?

16:35 <alyssa----> bbrezillon: I am using that to dump it, and no, I'm not sure of the block layout; I'm not convinced it's right

16:35 <alyssa----> bbrezillon: Look into the source code for gator https://github.com/ARM-software/gator (I think?)

16:42 <bbrezillon> alyssa----: ok

16:43 <bbrezillon> looks like we only have 1 MMU block

16:43 <bbrezillon> and 4 shader blocks

16:48 <bbrezillon> alyssa----: was trying to dump the counters as they were dumped with the non-drm driver

16:48 <bbrezillon> and that layout didn't match my version

16:51 <alyssa----> bbrezillon: It's different for every rev of the hardware and it's not exactly discoverable either. The gator source code handles it with some hard-to-follow heuristics. Maybe we can normalize it a bit, but that's still probably a userspace task (don't want to lug around all that complexity in the kernel for maintenance reasons)

17:10 <bbrezillon> alyssa----: the layout can be extracted from the GPU features/caps regs

17:10 <bbrezillon> I just followed what mali kbase does

17:10 <alyssa----> bbrezillon: *can*, it's just inconvenient :P

17:11 <bbrezillon> not sure what you mean

17:12 <bbrezillon> you'll anyway need to do that from userspace based on the same set of info, right?

17:14 <alyssa----> Yeah

17:23 <bbrezillon> alyssa----, tomeu, robher: I'm still debugging it, but given what alyssa---- said, maybe it's worth having a look at the code early on https://github.com/bbrezillon/linux/commits/panfrost/perfcnt

17:24 <bbrezillon> I did abstract quite a few things to make userspace's life easier

17:28 <bbrezillon> figuring out counter names is still left to userspace, but the driver provides a map of valid counters and block instances (see the GET_PERFCNT_LAYOUT ioctl)

17:53 <robher> bbrezillon: LGTM other than my few nits. Though I am wondering why GPU perf counters aren't wired into existing kernel perf counter infrastructure (perf).

17:56 <bbrezillon> robher: good question

17:56 <bbrezillon> I don't know

17:58 <bbrezillon> maybe because it's then hard to enable them on a per-GPU-job basis

18:01 <robher> bbrezillon: right, that's what I suspect too.

19:35 <alyssa----> bbrezillon: "+/* Copyright 2019 Panfrost Team */"

19:36 <alyssa----> Panfrost Team is not a legal entity; this should just be "Collabora" or "Boris Brezillon" (depending if this was a work project or not)

19:36 <alyssa----> As I think robher has commented before

19:37 <alyssa----> "panfrost_perfcnt_job_ctx_eq" memcmp?

19:38 <alyssa----> Only did a preliminary readthrough but looks great! :)

19:41 <Lyude> alyssa----: are you sure about that?

19:41 <Lyude> I'm pretty sure that should be fine, nouveau writes their copyrights the same way

19:42 <Lyude> oops, sorry, I missed the robher part

19:45 <Lyude> alyssa----: yeah, I do spot some older nouveau files that just have (C) The Nouveau Project

19:45 <Lyude> in mesa's src

20:13 <alyssa----> I like giving credit where credit is due anywya

20:14 <Lyude> alyssa----: that's why I usually include an Authors: list below

20:14 <Lyude> but this is all nitpick :P

20:14 <Lyude> whatever works

22:11 <daniels> yeah, the legal entity for our work is Collabora Ltd.

23:53 <alyssa----> bbrezillon: ^ There you go then