05:13 <bschiett> hi guys, new to the channel, i'm working on a rk3288 project and was wondering what the current state is of panfrost and if it offers everything the arm drivers offer? i'm mostly interested in openGLES/openGL

05:16 <HdkR> I'd say it offers more for Midgard

05:20 <bschiett> @HdkR OK, so it's definately stable/feature complete at this point?

05:21 <HdkR> It's open source, work is never complete :P

05:21 <bschiett> good point :-)

05:21 <bschiett> I guess as long as it's reasonably stable i'm OK.

05:24 <HdkR> If you find a bug then it is still worth reporting. People are usually pretty responsive if an issue gets tagged with panfrost :P

05:28 <bschiett> cool, thx

09:37 <daniels> robclark: sorry, I can't tell you why that happened, given there was no work scheduled :\

10:38 <justin3> Hi, first, thank you for your hard work! I'm curently looking at MALI drivers and I'm a bit lost. I can map memory in GPU and in kernel fine, but I'd like to issue commands to the GPU to read/write a GPU memory region. I looked at adreno and I managed to do it, but it seems a bit more complicated with mali GPUs. Maybe someone has a pointer to help to do that?

10:41 <HdkR> justin3: Using what API?

10:42 <justin3> well for now I use the kernel api with ioctls

10:43 <HdkR> So talking with the panfrost kernel driver through drm?

10:43 <HdkR> Or are you using kbase in your environment?

10:44 <justin3> hum not through drm, I use the ioctl of the mali drivers, basically like panloader I think

10:44 <justin3> ah yeah kbase

10:48 <HdkR> pretty sure it has some allocate ioctl or something. I try to not look at kbase as much as possible

10:49 <justin3> oh it has, that how I managed to allocate memory. But I'n not sure how to write something I can pass to the GPU to write in the allocated memory

10:50 <justin3> basically, I was working with adreno, and they have a allocate memory ioctl very similar, and I was able to pass "write mem" commands to the GPU. I was curious how to do that with mali

10:50 <HdkR> what is a write memory command? DMA?

10:51 <justin3> well in adreno, it's a specific opcode executed by the GPU

10:52 <HdkR> You'd probably have to setup a full compute shader that does a memcpy of some sort

10:52 <justin3> huum there is no easier way?

10:53 <justin3> also, would you have any pointer on how to do that?

10:53 <HdkR> Use a GL or Vulkan API instead of talking to it through ioctls instead :P

10:54 <HdkR> Since it sounds like you'd be reimplementing a bunch of this anyway for creating an async dma engine effectively

10:54 <justin3> ok, I don't know very well higher level api, will I be able to control at which address I can read/write?

10:55 <HdkR> With ARM's proprietary blob you can't pass arbitrary pointers. Mesa/AMD has some extension to map arbitrary user pointers in to GPU space though

10:57 <justin3> hum I'm very not faniliar with thoses API

10:58 <justin3> seems complicated :p

10:58 <justin3> well, I'll try to look at some doc about mesa

10:59 <justin3> if anyone has insights on how issue a "write/read" memory job from ioctl, please let me know

11:00 <justin3> thx for your help HdkR

13:10 <alyssa> justin3: You mostly can't, the closest thing is a write value job, but on Midgard that's just used for dumping cycle counters and writing a fixed 0

13:10 <alyssa> Anyway, XY problem -- what are you trying to solve?

13:12 <HdkR> Time to bitbang a compute shader task for async memcpy through the GPU :P

13:13 <justin3> Hi alyssa, do you think I could use a write value job to write in a specific region of the job address space?

13:13 <alyssa> Again, why?

13:13 <alyssa> (Not trying to be annoying, just trying to understand the goal so I can help)

13:14 <justin3> I'm mostly interested in how memory management works for mobile gpus

13:14 <justin3> so I try to get the GPU read/write in regions mapped in cpu userland

13:16 <alyssa> Well, if it's purely academic, you could fill memory in CPU with e.g all 1's, and then use a write value job to write a zero to some address, and that will be visible on the CPU if you read it back and get all the coherency right :)

13:17 <alyssa> (On Bifrost, you can write any constant you wish. But not many use cases.)

13:17 <alyssa> https://gitlab.freedesktop.org/mesa/mesa/-/blob/master/src/panfrost/lib/midgard.xml#L1461-1480

13:17 <alyssa> https://gitlab.freedesktop.org/mesa/mesa/-/blob/master/src/panfrost/lib/midgard.xml#L1427-1443

13:18 <alyssa> ^^ The format the write value job needs to be, you place that structure in GPU memory and then tell the GPU to execute it (set jc to its address in GPU memory)

13:18 <alyssa> This is easier with the Panfrost kernel than with kbase, I recommend you use ours.

13:19 <justin3> woow many thanks that's exactly what I want!

13:19 <alyssa> 👍

13:19 <alyssa> Have fun :)

13:19 <justin3> so hum, as I didn't read all panfrost, is tehre an example code that submit a job

13:20 <justin3> or should I look at mesa, or maybe it's easy enough to gigure out how to do it from the xml?

13:21 <alyssa> panfrost_batch_submit_ioctl

13:21 <justin3> ok thank you, I have a last question, is it possible to write a "read value job" too?

13:56 <alyssa> No.

14:13 <macc24> bbrezillon: did suspend work on https://github.com/bbrezillon/linux/tree/panfrost/duet ?

14:15 <bbrezillon> macc24: yes

14:15 <macc24> fuck

14:16 <bbrezillon> what's the problem?

14:17 <macc24> it doesn't poweron dsi/panel something correctly

14:19 <bbrezillon> alyssa: I'm debugging dEQP-GLES3.functional.fbo.msaa.4_samples.rgba16f failure on Midgard, and I notices that samples 0 and 1 are written correctly, but not samples 2 and 3. Any thoughts on what could cause that?

14:19 <bbrezillon> macc24: that used to work

14:20 <bbrezillon> I mean, you had it working too, right?

14:20 <macc24> i have to try it

14:20 <bbrezillon> I don't get what the problem is

14:21 <bbrezillon> suspend/resume?

14:21 <macc24> resuming

14:21 <macc24> it doesn't power the dsi correctly at least on my kernel

14:21 <bbrezillon> ok, that part has not been tested, indeed

14:22 <bbrezillon> macc24: I recommend you look at the chomeos kernel tree

14:22 <macc24> oh i forgot it exists

14:25 <macc24> it looks like in chromeos kernel mtk_dsi.c has no suspend/resume parts

14:26 <bbrezillon> what's the error exactly?

14:26 <macc24> i don't have access to console without screen

14:30 <justin3> thank you very much for your help alyssa, that was very helpful

14:46 <urjaman> how do you know it's the display that's preventing the resume then?

14:47 <urjaman> if it comes up with everything but display you should be able to ssh in or something tbh

14:56 <macc24> hmmmmmmmmmmmmmm

16:27 <macc24> it might be panel driver