#asahi-gpu · 2023-02-15 16:36
linaBTW, I just set the governor to "performance" and don't mess with frequencies#asahi-gpu · 2023-02-15 16:36
linaNice!#asahi-gpu · 2023-02-15 16:35
linaOkay, so we're actually doing reasonably well for a tiler ^^#asahi-gpu · 2023-02-15 16:33
linaOkay what does panfrost get?#asahi-gpu · 2023-02-15 16:33
linawwwww#asahi-gpu · 2023-02-15 16:33
linaWhat numbers are you expecting for terrain vs. other GPUs?#asahi-gpu · 2023-02-15 16:32
linaAh, I'm using gbm which probably gains a bit#asahi-gpu · 2023-02-15 16:31
linaI get 169 ^^#asahi-gpu · 2023-02-15 16:31
linataskset to 4-7?#asahi-gpu · 2023-02-15 16:29
linaBut also I have my suspicions about mailbox hurting with that too...#asahi-gpu · 2023-02-15 16:29
linaWasn't chadmed working on that?#asahi-gpu · 2023-02-15 16:28
linaThe CPU scheduler is our enemy...#asahi-gpu · 2023-02-15 16:28
linaOh, did you set your CPUs to performance and pin to the p-cores?#asahi-gpu · 2023-02-15 16:27
linaWhat resolution?#asahi-gpu · 2023-02-15 16:27
linaterrain is like 160 now I think?#asahi-gpu · 2023-02-15 16:26
linawwwwww#asahi-gpu · 2023-02-15 16:25
linaWhat was the FPS before?#asahi-gpu · 2023-02-15 16:25
linaNice!#asahi-gpu · 2023-02-15 16:19
lina^^;;#asahi-gpu · 2023-02-15 16:13
linaAlso right now it doesn't try to track batch-to-batch dependencies so it always does a full barrier between batches (so no vert/frag concurrency). That can clearly be fixed but then we need to be more careful about when we need to insert the barrier and not.#asahi-gpu · 2023-02-15 16:12
linaAny comments on the general approach? I just hacked on it until it worked, I want to know what you think about the submitted bitfield and all that#asahi-gpu · 2023-02-15 16:10
linaYeah ^^#asahi-gpu · 2023-02-15 16:09
lina*touch#asahi-gpu · 2023-02-15 16:09
linaI didn't even though the UAPI or its wrapper#asahi-gpu · 2023-02-15 16:09
linaMost of what I did today is not UAPI-specific, it's all the batch tracking stuff to make explicit sync work#asahi-gpu · 2023-02-15 15:53
linaI'll be back in a bit, I'm going to get some conbini dinner or something#asahi-gpu · 2023-02-15 15:51
linaIt wasn't actually breaking anything, just warning#asahi-gpu · 2023-02-15 15:51
linaI think a compiler upgrade started complaining about that one#asahi-gpu · 2023-02-15 15:51
lina^^;;#asahi-gpu · 2023-02-15 15:49
linaI couldn't figure out how to get real-time traces, so I don't know how much of the overhead is context switching, mailbox stuff, etc.#asahi-gpu · 2023-02-15 15:48
linaalyssa: Okay ^^#asahi-gpu · 2023-02-15 15:48
linaAt this point I think the next biggest submission overhead thing is just memory allocs, so I probably want to add some general mechanism to pool together random unrelated allocs for a single submission. But that can probably wait a bit.#asahi-gpu · 2023-02-15 15:47
linaalyssa: I also did some perf opt on the kernel side, there was some easy low-hanging CPU fruit with array initialization that gave me an extra 10% or so on glmark (for 4000+ FPS workloads, won't do as much for things that are more GPU-limited)#asahi-gpu · 2023-02-15 15:45
linajannau: Oh, I thought it was more general. So basically just Apple then?#asahi-gpu · 2023-02-15 15:44
linaAny ideas on how to land this? Most of the changes are in general code, but then end up talking to syncobjs of course. At this point we probably want to drop macOS first?#asahi-gpu · 2023-02-15 15:42
linaI'm regularly hitting 200+ FPS on xonotic now at least ^^#asahi-gpu · 2023-02-15 15:41
linaAlso perf works again. It was an upstream regression (on all ARM!), jannau found it while I was busy with mesa ^^#asahi-gpu · 2023-02-15 15:40
linaI'm still not sure if the firmware stuff is really 100% robust yet, we'll find out when I turn on timestamps again since that made it really obvious in the past.#asahi-gpu · 2023-02-15 15:32
linaOne little thing: sometimes you get GPU faults when killing things. Since everything is async now and BOs are tracked by userspace, if userspace dies in-progress GPU operations lose their BOs. The driver itself keeps all the critical structures alive until the jobs complete/fail, so nothing breaks, you just get GPU fault logs.#asahi-gpu · 2023-02-15 15:26
linaPushed all the explicit sync stuff to gpu/explicit-sync (kernel) and agx/explicit-sync (mesa), seems to work well ^^#asahi-gpu · 2023-02-15 15:18
lina800x600 surfaceless glmark score: 4496#asahi-gpu · 2023-02-15 13:33
linaI did fix some more threading bugs ^^#asahi-gpu · 2023-02-15 13:33
linaalyssa: With that commented out, we pass all the EGL tests except robust_gl_3* context creation stuff which sounds not applicable... ^^#asahi-gpu · 2023-02-15 12:11
linadaniels: x11/egl (I think that's what the dEQP build picked by default?)#asahi-gpu · 2023-02-15 10:07
linadEQP-EGL.functional.swap_buffers_with_damage.resize_after_swap.buffer_age_render#asahi-gpu · 2023-02-15 10:07
linaalyssa: This is triggering on the EGL testsuite: Assertion `maxx > minx && maxy > miny' failed#asahi-gpu · 2023-02-12 13:15
linaThe problem with 4K kernels is that I have no way to tell the GEM shmem backend to allocate in contiguous 16K pages...#asahi-gpu · 2023-02-04 02:38
linaCool ^^#asahi-gpu · 2023-02-04 02:28
linaLike I'd expect something like every 2 layers to have identical averaged out mip level 1s, every 4 layers for level 2s, etc. (or something like that)