00:43 <macc24> how to know if it's video game being slow, panfrost being slow or g31 being g31?

00:43 <macc24> i got re3 to run on my odroid go super and it barely can do 30fps

00:49 <HdkR> Likely all of the above

00:50 <macc24> ugh

00:55 <alyssa> macc24: i landed some commits yesterday that should make bifrost a bit faster

00:55 <alyssa> how much faster and will it help? TBD.

00:55 <macc24> ok

00:55 <macc24> will try rebuilding

00:59 <alyssa> I don't expect it to be radically better fwiw

00:59 <alyssa> but they were low hanging fruit

00:59 <macc24> ,_,

00:59 <macc24> bit by bit it's going to reach 60fpx

00:59 <alyssa> and there were a bunch of them so shrug

00:59 <macc24> i think i also have like 10% free performance from overclocking dmc

01:01 <macc24> but yea, panfrost is a lot faster than like, 10 months ago? when it struggled to render retroarch menu

01:02 <alyssa> on g31?

01:02 <macc24> yep

01:02 <alyssa> nod

01:02 <alyssa> bifrost compiler was a toy back then

01:02 <alyssa> admittedly it's still not an adult compiler but yknow

01:02 <alyssa> mid adolescence

01:03 <macc24> >name of display: :0 \n Segmentation fault

09:50 <icecream95> urja: Although openscad does segfault for me on exit, it happens whether Panfrost is used or not

09:54 <urja> oh. can confirm (i thought i had confirmed that it didnt happen w/ llvmpipe way back)

09:57 <urja> okay i'll ignore that for now then (i'll check it on x86 when I get back on an x86 box w/ display ... too hot here to have the main PC running to make the room even hotter lol)

09:58 <urja> might be the newer version of openscad too (last testing was w/ 2019.05 ... based on what was in my aur directory :P)

10:01 <urja> wow

10:01 <icecream95> It's the sort of bug that I bet Mesa would keep hitting if it was written in C++

10:07 <HdkR> Shaming all my C++ projects :P

10:27 <icecream95> For the "ctrl-c causes a pile of GPU faults" bug, at least on Midgard the kernel patch someone posted a link to isn't enough, because the fancy new pan_blitter stuff doesn't add all the BOs it uses to the list passed to the kernel, so the kernel doesn't know not to delete those BOs while the job is still running

10:29 <icecream95> (This would be a problem even in normal usage if those BOs didn't last for the entire screen lifetime)

10:32 <urja> hmmm...

11:54 <macc24> hmm

11:54 <macc24> what if something is wrong with my cross compiling setup

11:54 <macc24> *proceeds to compile panfrost on quadcore cortex-a35*

23:53 <icecream95> alyssa: Using a staging buffer for pool uploads and memcpying to the BO before submit increases drawoverhead scores by as much as 15%

23:55 <icecream95> Maybe that means there are still some reads from pool-allocated memory, or it could just be that a linear memcpy is faster than uploading bit by bit

23:55 <icecream95> (not literally "bit by bit")