#asahi-gpu · 2023-08-16 12:27
linaThere is indeed a magic VA range for dynamic allocation and the messages you'd expect back&forth plus some management structures I didn't look at yet#asahi-gpu · 2023-08-16 12:27
linaI also looked at mesh shaders, it's what you said - no magic CDM in VDM thing, just discrete compute...#asahi-gpu · 2023-08-16 12:26
linaSeemed related to the helper program for compute but unclear#asahi-gpu · 2023-08-16 12:26
linaAnd some mystery "no preemption" (maybe?) flag that would be nice to test#asahi-gpu · 2023-08-16 12:26
linaAlso threw in the bindless sampler thing#asahi-gpu · 2023-08-16 12:26
linaAll the helper programs are exposed in the UAPI now too, you just need to implement it ^^#asahi-gpu · 2023-08-16 12:25
linaLayered rendering is fixed, needed a kernel bugfix (missing one magic buffer) + all the strides in the UAPI for Z/S#asahi-gpu · 2023-08-16 12:25
linaKernel update in the usual branch, agx/uapi-updates-202308 on top of your layered branch but it's just one commit you can cherry-pick#asahi-gpu · 2023-08-16 12:25
linaI did see it when bringing up G14X but I didn't make the connection that it was the same helper as always...#asahi-gpu · 2023-08-16 12:24
linaalyssa: The helper program. It's for all three.#asahi-gpu · 2023-08-16 09:06
linaThat suddenly makes a lot of sense though...#asahi-gpu · 2023-08-16 09:06
linaBad news: It's actually setting up the stacks, so we get to implement all that if we want spilling ^^;;#asahi-gpu · 2023-08-16 09:06
linaalyssa: Good news! I figured out what the helper program is for.#asahi-gpu · 2023-08-12 14:57
linaalyssa: How do we do that?#asahi-gpu · 2023-08-12 06:04
linaalyssa: I think we want to no16 shame firefox... they haven't replied to my bug report (with patch) about the Sheets bug yet, but there's also all the broken WebGL games/apps...#asahi-gpu · 2023-08-11 13:26
linaIt could also be randomly "fixed" though, given that it's a heisenbug..#asahi-gpu · 2023-08-11 13:26
linaThat one's been broken in firefox for ages...#asahi-gpu · 2023-08-11 13:25
linaPlease...#asahi-gpu · 2023-08-11 13:25
linaOh.#asahi-gpu · 2023-08-11 13:25
linaYeah, I didn't see anything obvious but I didn't trace every single line...#asahi-gpu · 2023-08-11 13:24
linaSo I think it has to do something with that#asahi-gpu · 2023-08-11 13:24
linaBut one side effect of that is the order of the basic blocks, and the sense of the condition variable, get flipped in the asm#asahi-gpu · 2023-08-11 13:24
linaSo one thing that fixes it is changing the condition in fetch_clip_area() to `if (index >= CLIP_TASK_EMPTY && index != 0xdead) {` (which should be a no-op as long as the index isn't that weird value)#asahi-gpu · 2023-08-11 13:22
linaI mean what about cmpsel type stuff?#asahi-gpu · 2023-08-11 13:22
linaDo we know how NaN propagation works in general on AGX?#asahi-gpu · 2023-08-11 13:22
linaThe square is supposed to be a map thing, not glitched out#asahi-gpu · 2023-08-11 13:21
linaand --loop=300 to see the last frame more#asahi-gpu · 2023-08-11 13:21
linaUmm try --sb#asahi-gpu · 2023-08-11 13:20
linaThe FS side manifests in do_clip() and happens even if all paths there return good near-1.0 alpha values (but not if all return 1.0 because that causes everything to be optimized out)#asahi-gpu · 2023-08-11 13:19
linaFrom which I deduce that cmpsel stuff propagates NaNs? If not, something else is even more wrong...#asahi-gpu · 2023-08-11 13:19
linaThere whole thing seems to spill into the FS, my guess is NaN propagation? That's the only thing that makes sense since the FS breaks even when all code paths must return good values.#asahi-gpu · 2023-08-11 13:18
linaIn the VS, messing with fetch_clip_area causes it to work or not, seemingly depending on the order/polarity of the condition there#asahi-gpu · 2023-08-11 13:18
linaThe bad draw is 6918#asahi-gpu · 2023-08-11 13:17
linaalyssa: https://cloud.asahilina.net/s/omCgkcc9LAHKjte#asahi-gpu · 2023-08-11 10:23
linaIt's what's causing one of the firefox issues#asahi-gpu · 2023-08-11 10:23
linaI have this really weird shader bug which goes away if I add dummy code#asahi-gpu · 2023-08-11 10:23
linaalyssa: I need your help ;;#asahi-gpu · 2023-08-10 00:26
linaWe do have a dummy framebuffer but it's not probing, I think something is messed up with the power domains since these devices don't actually have disp0#asahi-gpu · 2023-08-10 00:25
linaalyssa: The problem with kmsro is that it hardcodes what display controllers go with what drivers#asahi-gpu · 2023-08-09 11:49
linaCool! ^^#asahi-gpu · 2023-08-09 11:44
linaOnly problem is I don't know how to run the "real" CTS without a video output for X/Wayland ^^;;#asahi-gpu · 2023-08-09 11:40
linaalyssa: G14D passes the surfaceless dEQP/KHR tests up to GLES31 with the MR I sent, except KHR-GLES31.core.texture_buffer.texture_buffer_atomic_functions is flaky. This is on asahi/mesa with the deqp debug flag.#asahi-gpu · 2023-08-02 05:49
linaSo it just needs a m1n1 fix to ignore it in this case instead of fail#asahi-gpu · 2023-08-02 05:48
linajannau: Ohh ok, looks like the Mac Studio doesn't actually use any of this for the GPU, I guess iBoot just throws that info in but the driver doesn't use it.#asahi-gpu · 2023-08-02 01:39
linaNot just that one I guess, probably everything t600x?#asahi-gpu · 2023-08-02 01:39
linajannau: Oof, that means I need to work out the fuses for that one too ;;#asahi-gpu · 2023-07-30 06:13
linajannau: Pushed a fixup directly to main#asahi-gpu · 2023-07-30 05:49
linaI guess I can just change the condition to check for agx_linear_allowed, and if it's not, shrug and let it be compressed/tiled#asahi-gpu · 2023-07-30 05:48
lina... why are we trying to allocate mipmapped or cubemaps as shared? That looks wrong...#asahi-gpu · 2023-07-30 03:52
linahttps://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21685 was my attempt but...