#asahi-gpu · 2022-11-08 13:51
linaI think there's like a dozen buffers that get allocated from shmem/mapped/freed for every render w#asahi-gpu · 2022-11-08 13:51
linaBut I think the driver doesn't do anything silly as far as stability so I hope the kernel side works well for you!#asahi-gpu · 2022-11-08 13:50
linaI still need to fix the allocator to not be totally silly, I'll do that next week...#asahi-gpu · 2022-11-08 13:49
lina^^;;#asahi-gpu · 2022-11-08 13:47
linaYay!!!! ^^#asahi-gpu · 2022-11-08 01:20
linaIt's pushed now...#asahi-gpu · 2022-11-08 01:20
linaalyssa: Aaaa sorry I forgot to push m1n1!!! T_T#asahi-gpu · 2022-11-04 09:39
linait should render with scrolling ffs sometimes, and sometimes all-ff, but it might be different on 24c#asahi-gpu · 2022-11-04 09:39
linaCBW. WCTA. STAC.STAS2.[ 68. 8] unk_68 = 0xff000000000000#asahi-gpu · 2022-11-04 09:39
linajannau: If you're around, can you run the regular AGX tracer for a bit (until it renders stuff) and look for this field?#asahi-gpu · 2022-11-04 07:31
linaphire, jannau: Any chance you could dump 0x404d04000 .. 20 on your machines? ^^ (might need to pmgr enable the GPU first)#asahi-gpu · 2022-11-01 01:23
linaI already know f-i is linear, so I left it overnight grabbing v-p curves @ two frequencies to get the dynamic v-i curves (core & sram). If we can solve that one, that's everything.#asahi-gpu · 2022-11-01 01:22
linaYeah, exactly#asahi-gpu · 2022-11-01 01:19
linaProbably at the very least power state decisions#asahi-gpu · 2022-11-01 01:19
linachadmed: There's like 10 copies of the max-power number in initdata, so I'm pretty sure it's used for lots of things...#asahi-gpu · 2022-10-31 12:50
linaIf I just extrapolate the first piecewise component, that's already intersecting the discontinuity (closer to the top side)... maybe we should just go with that?#asahi-gpu · 2022-10-31 12:49
linaThe voltage range that matters is about 400-920mV, especially the high side... which conveniently is precisely dancing on the broken discontinuity.#asahi-gpu · 2022-10-31 12:48
linaI don't know what to do about this one. Obviously copying Apple's curve there is wrong and silly.#asahi-gpu · 2022-10-31 12:48
linaIt turns out their math is demonstrably wrong by several percentage points...#asahi-gpu · 2022-10-31 12:48
linachadmed: Meanwhile Apple: https://twitter.com/LinaAsahi/status/1587058352885882880#asahi-gpu · 2022-10-31 08:34
linachadmed: New (hopefully less cursed) data fitting challenge: https://gist.github.com/asahilina/729b99c214ffcceadb396c5f4c4bd7ec ^^#asahi-gpu · 2022-10-29 14:23
linaAnd the f=0 data is proportional (to within the available significant digits) to the leakage table values.#asahi-gpu · 2022-10-29 14:22
linaSo that means there is no per-core shenanigans for dynamic power, and the static and dynamic power are just added together#asahi-gpu · 2022-10-29 14:22
linaThe core leakage/static power data is additive. If I subtract f=0 data from f=1000 data, the total power becomes constant#asahi-gpu · 2022-10-29 14:21
linaOK, so at least some things are adding up...#asahi-gpu · 2022-10-29 14:14
linajannau: Sure! More data is great ^^#asahi-gpu · 2022-10-29 14:14
linaSo then both would have to line up, which is a 1/16 chance#asahi-gpu · 2022-10-29 14:14
linaThough they get to pick both dies, and presumably they would pick two bad ones if they had enough of them#asahi-gpu · 2022-10-29 14:13
linaTrue...#asahi-gpu · 2022-10-29 14:03
linaphire: If you get a chance, can you do the same trace? initdata with that pwr script + hv/trace_all_more.py#asahi-gpu · 2022-10-29 14:02
lina... and yours are completely regular, so I'm pretty sure Apple is turning cores off for marketing reasons and not yield at this point, at least for you#asahi-gpu · 2022-10-29 14:02
linajannau: Heh, looking at the SGX registers it reads at the beginning, yours has what looks like a mask of active cores...#asahi-gpu · 2022-10-29 12:41
linaI just realized trace_all was untracing pmgr/error_handler which definitely have reg ranges related to fuses in them...#asahi-gpu · 2022-10-29 12:40
linajannau: Actually actually, use hv/trace_all_more.py which I just pushed...#asahi-gpu · 2022-10-29 12:34
linaIt'll be a huge log but it might have something useful...#asahi-gpu · 2022-10-29 12:34
linajannau: Actually, use trace_all (on top of trace_agx) too if you can#asahi-gpu · 2022-10-29 12:27
linaI want to figure out if any of the few SGX registers read at startup have like a variant/model ID, if I can convince my macOS that I have a lower-core model I can find out if that's really what matters or not#asahi-gpu · 2022-10-29 12:27
linajannau: Can you take another dump of the M1 Ultra with that agx trace script, but save the HV log (-l log.txt) and send me that?#asahi-gpu · 2022-10-29 12:00
linaAh yeah, my weirdo old mesa for this still had the "force linear" patch in it...#asahi-gpu · 2022-10-29 06:00
linaI don't think I broke anything but the render command though?#asahi-gpu · 2022-10-29 06:00
linaOh also, the Python side does not implement the up to date UABI yet... I'm running on some old mesa branch still for that#asahi-gpu · 2022-10-29 05:04
linaMake sure to check render.py, I think I put in some dumb condition for what to blit to the FB#asahi-gpu · 2022-10-29 03:22
linaThere's no mechanism for actually doing WSI right now#asahi-gpu · 2022-10-29 03:22
linaYeah, it can't do GLX, you need to use surfaceless EGL and it'll just blit some surfaces into the framebuffer...#asahi-gpu · 2022-10-28 05:40
linaAt the fragment level though, right? Not the vertex level...#asahi-gpu · 2022-10-28 05:40
linaAnyway, I should get ready for the stream (and have a sandwich or something) ^^#asahi-gpu · 2022-10-28 05:39
linaYeah, but if they have to do some kind of flush around alpha draws...#asahi-gpu · 2022-10-28 05:38
linawwwww#asahi-gpu · 2022-10-28 05:37
linaThat seems really wasteful...#asahi-gpu · 2022-10-28 05:36
lina*that