#
✕ nick:lina
#asahi-gpu · 2023-07-22 08:30
linaSomehow we have batches holding onto gigabytes of BOs...
#asahi-gpu · 2023-07-22 08:22
linaOK, I think it's not the cache, I think the cache just inherits the bloat for the extra 1-2 seconds it's programmed to
#asahi-gpu · 2023-07-22 08:13
linaOof yeah, I see 10+GB of BO cache. This is bad.
#asahi-gpu · 2023-07-22 07:53
linaMaybe the BO cache is just over-enthusiastic?
#asahi-gpu · 2023-07-22 07:41
linaWe have a Firefox problem... it's using crazy amounts of GPU memory for some stuff, like Google Maps and the PDF renderer.
#asahi-gpu · 2023-07-21 11:05
linaSo maybe it's worth trying ^^
#asahi-gpu · 2023-07-21 11:05
linaI wonder if it might be a good test case... I can't imagine that kind of program doing anything "super" fancy in terms of graphics API corner cases?
#asahi-gpu · 2023-07-21 10:48
lina"gl"scopeclient only has a Vulkan renderer? ^^;;
#asahi-gpu · 2023-07-19 11:34
linaOK ^^
#asahi-gpu · 2023-07-19 11:32
linaalyssa: By the way, did you ever look at YCbCr conversion? We're still not loading the conversion coefficients in the kernel, and it might be a good idea to test this and start doing that before upstreaming.
#asahi-gpu · 2023-07-19 11:18
linaThe compression meta buffers are directly after the main buffers with no page alignment, so that's probably just implied when using this configuration
#asahi-gpu · 2023-07-19 11:16
linaWhat patches?
#asahi-gpu · 2023-07-19 11:16
linaNo?
#asahi-gpu · 2023-07-19 11:16
linaYes, and "Compressed 1" is true
#asahi-gpu · 2023-07-19 11:16
lina<field name="Compression" size="2" start="106" type="hex"/> <!-- 0 for uncompressed -->
#asahi-gpu · 2023-07-19 11:15
linaSo it's just the fun descriptors then
#asahi-gpu · 2023-07-19 11:15
linaOhh OK, so that one is already known
#asahi-gpu · 2023-07-19 11:14
linaalyssa: do you know if compression=0 works? also any idea about the TODO.unkB1?
#asahi-gpu · 2023-07-19 11:12
linaMind you, this is probably orthogonal to the actual compression
#asahi-gpu · 2023-07-19 11:12
linaYeah...
#asahi-gpu · 2023-07-19 11:10
linaI guess this is a whole new PIPE_FORMAT, don't see anything matching it...
#asahi-gpu · 2023-07-19 11:05
linaAlso the store shader uses TODO.unkB1 ?
#asahi-gpu · 2023-07-19 11:05
linaSo then I'm not sure if the compression is the same or not because too many things changed, but at least "Compression" is 0x0 which seems different.
#asahi-gpu · 2023-07-19 11:04
linaAnd when you do that, the accel buffer field is the address of the second plane... and I guess the actual compression metadata offset is implied?
#asahi-gpu · 2023-07-19 11:04
linaIt turns out it *does* have support for X2R10G10B10 + A8 biplanar formats directly
#asahi-gpu · 2023-07-19 11:03
linaalyssa: Sooo about that DCP compression thing... https://gist.github.com/asahilina/59cf953b82dd1675909f1cbb1a650fd4
#asahi-gpu · 2023-07-19 01:57
linaalyssa: I think you can just merge it if you're comfortable without line by line review, unless you want me to just review it today ^^
#asahi-gpu · 2023-07-17 22:52
linaalyssa: Because gettimeofday is a complex function of the raw CPU timer counter ^^
#asahi-gpu · 2023-07-16 12:31
linaNot syscall, ioctl
#asahi-gpu · 2023-07-16 12:30
linarosefromthedead, j`ey: Looks like the timekeeping stuff is pretty deep kernel internals and I don't think it's worth fighting to figure out how to precisely expose the mapping to userspace... so I think a syscall that gets the GPU time and the current CLOCK_* time (not synchronized, so with a small error offset) is probably the least painful way to go for now
#asahi-gpu · 2023-07-16 12:13
linaj`ey: Problem is we can't use any internal APIs here, we need to go through the official entrypoints...
#asahi-gpu · 2023-07-16 12:11
linaThey *can* be thunked for 64bit in FEX. 32bit is always emulated right now. I haven't set up thunks so all my tests to date have been emulated.
#asahi-gpu · 2023-07-16 12:08
linaWhat won't work is directly reading the ARM64 register from Mesa
#asahi-gpu · 2023-07-16 12:08
linaThey should, so that should work
#asahi-gpu · 2023-07-16 12:07
linaSince vDSO should be plumbed in properly for emulation
#asahi-gpu · 2023-07-16 12:07
linaThe problem is that won't work from an emulated x86 environment, we need an ioctl or gettimeofday for that
#asahi-gpu · 2023-07-16 12:07
linaWhich... I guess we could just do anyway? Though I need to check whether Linux guarantees that access is always enabled
#asahi-gpu · 2023-07-16 12:06
linaYes, that goes through vdso and I think that directly reads the counter from userspace outright
#asahi-gpu · 2023-07-16 12:03
linaThere might also be some conversion factor we can provide instead to let userspace know how to convert from CLOCK_MONOTONIC or similar, but I need to investigate the kernel time tracking code to see how that works
#asahi-gpu · 2023-07-16 12:03
lina(the raw ARM timer)
#asahi-gpu · 2023-07-16 12:02
linarosefromthedead: It's the same as the CPU time, I've been meaning to add an ioctl to grab that
#asahi-gpu · 2023-07-16 10:36
linaThose messages are harmless, it's just the lockups were related to that so I wanted to see...
#asahi-gpu · 2023-07-16 10:35
linaYeah, that would be expected then!
#asahi-gpu · 2023-07-16 10:35
linaI deliberately made that a bit noisy to see what the behavior is if someone runs into the issue again...
#asahi-gpu · 2023-07-16 10:34
linaj`ey: That was for debugging, if the lockups are gone I'll revert them ^^
#asahi-gpu · 2023-07-15 03:06
linaThis was a while ago though... I could be totally off
#asahi-gpu · 2023-07-15 03:06
linaBut I think I saw something where one of the dispatch modes ended up with some table provided that I assumed had something to do with workgroups
#asahi-gpu · 2023-07-15 03:06
linaMaybe? I might be misremembering too
#asahi-gpu · 2023-07-15 03:02
linaalyssa: I'm pretty sure I saw that some point, when you do non-uniform sizes there's this big structure/table thing with the sizes I think?
#asahi-gpu · 2023-07-14 02:25
linaI wonder if we want to encode the accel buffer offset in the modifier or just declare that the garlic rules are implied?

plus ›