raster has quit [Remote host closed the connection]
ckeepax has joined #panfrost
_whitelogger has joined #panfrost
<tomeu>
robher: I'm starting to implement job submission in the drm backend, based on a guessed up ioctl
<robher>
tomeu: that boils down to code in mali_kbase_jm_hw.c. Which then got me looking at feature regs that tell how many cores and coherent clusters.
<tomeu>
oh, i saw that passing by today
<tomeu>
have you found it?
<tomeu>
robher: ^
<robher>
tomeu: there's only one cluster and we don't need jobslot 2 at least for what we care about for now.
<tomeu>
robher: so you have pushed everything you had to push already?
<robher>
tomeu: All I have is dumping some feature registers.
<tomeu>
ah, thought you had some fixes for the MMU
<tomeu>
robher: so the SUBMIT ioctl would basically contain an array of kbase_jd_atom?
<robher>
tomeu: wait, I do need to push the gpu scheduler skeleton.
<tomeu>
robher: is pantest running successively for you without errors?
<tomeu>
the one you pushed is oopsing after I run it for 3 or 4 times
<robher>
Just run it 2 times then. ;)
<tomeu>
yeah well, the backtraces look very similar to when I try to allocate more than a few thousand bytes :)
<tomeu>
robher: have you started moving pantest into igt? I could do that once I finish this
<robher>
tomeu: nope. And I need to go work on DT for a few days...
<tomeu>
ok, I prefer to not invest much time on the MMU, as you seem to have gone quite deep there already
<tomeu>
so I can do igt today and next week will check if you have made progress on the MMU issues
afaerber has quit [Quit: Leaving]
afaerber has joined #panfrost
<tomeu>
robher: do you have a tree in kernelci? we are already running igt on a few machines with midgard there, so it could be quite easy to run these tests there automatically
<robher>
tomeu: no, I don't.
<tomeu>
ok, maybe we'll think later of a good way of doing it
<robher>
tomeu: I figured out the mapping issue. The io-pgtable lib expects everything aligned. Testing a fix now.
<tomeu>
just in time, I'm debugging the igt tests now :)
<alyssa>
robher: "Don't need 3rd one until we have compute support"
<alyssa>
FWIW, compute support is really simple (at least for simple stuff, ES3.1 style compute... OpenCL obviously makes things a lot more complicated), so it'll probably be implemented at least a little bit sooner than later, for debug purposes if nothing else
<alyssa>
(Simple since vertex shaders are just compute jobs in the hw)
<alyssa>
It's true that the raw values of GPU feature detection registers are passed through, but they're _also_ processed in kernel space (at least partially)
<alyssa>
See mali_kbase_gpuprops.c for that in action
<alyssa>
Not that we have to worry about that :)
<HdkR>
alyssa: But what about OpenCL embedded profile? :P
<alyssa>
HdkR: I don't know much about CL to answer that :P