08:01 <ella-0> alyssa: Got no problems with clang-formatting agxv :3

15:03 <alyssa> ella-0: sprakles

15:03 <alyssa> sparkles

17:00 <alyssa> ella-0: by the way, jekstrand complained about the following code pattern I have in Asahi (and Panfrost):

17:01 <alyssa> - lots of high level system values in the NIR e.g. nir_load_ubo_base_agx

17:01 <alyssa> - compiler translates those to moves of magic uniforms

17:01 <alyssa> - compiler passes back a list of uniforms that the driver needs to upload in a certain layout

17:01 <alyssa> - driver needs to upload those uniforms in the layout the compiler decided

17:02 <alyssa> Apparently, this works well enough for GL, but is bad for Vulkan aparently

17:03 <alyssa> The issue being the layout is compiler-driven, even though the dettails are all driver-specific

17:03 <alyssa> so in effect I'm hardcoding GLisms into the compiler and then VK would have to pretend to be GL

17:03 <alyssa> whereas VK would want to do certain things in a different way maybe, e.g. descriptor sets instead of open coding binding tables

17:04 <alyssa> It also turns out that it isn't great for GL either but I digress

17:04 <alyssa> He argues that what should actually happen is that the driver preprocesses the NIR before the compiler sees it, with a little NIR pass to lower system values to load_uniform/load_preamble instructions, and the driver internally keeps track of what it wants to upload

17:05 <alyssa> then the compiler has no knowledge of any special sysvals. It just translates load_preamble instructions 1:1 to hardware uniform registers

17:06 <alyssa> and the GL and VK drivers get to (independently and completely) control the layouts

17:06 <alyssa> It's not clear to me why this is better for VK seeing as I don't work on VK yet

17:06 <alyssa> but. I did write the patches last night to convert the compiler + GL over to the new model. giving you a heads up that you'll need to add some code to VK when this lands and you rebase

17:15 <alyssa> I'd ask jekstrand to chime in but he's on vacation

17:51 <TellowKrinkle> BTW how do uniform registers get initialized? I noticed the Metal compiler seems to reference a lot of constants from them without loading them in a preamble shader, including things from constant address space buffers (with arbitrary constant offsets too), which can be bound to arbitrary GPU buffers. How does that work?

17:52 <alyssa> TellowKrinkle: USC load uniform commands

17:52 <alyssa> see "USC Uniform" in mesa/src/asahi/lib/cmdbuf.xml

17:52 <alyssa> if you use libwrap.dylib, the pointers will be dumped

17:52 <alyssa> but yeah

17:53 <alyssa> can push arbitrary ranges of GPU memory

18:40 <ella-0> alyssa: yeah, in vulkan we ideally want to tell the compiler where everything is not the otherway round; Converting VK to that from what I have now should be fairly trivial :3

18:51 <alyssa> awesome

18:52 <alyssa> I still don't understand *why* Vulkan wants that so bad

18:52 <alyssa> but I have the patches to make it so :p

19:03 <ella-0> yah something something VkPipelineLayout

19:06 <alyssa> Hmmm

19:06 <alyssa> OK

23:07 <alyssa> Mm, 2.5fps with exclusive ubershaders running at 4K :p

23:54 <jannau> any ideas how dcp runtime power management could crash the agx coproc?

23:58 <jannau> might be power domain, iirc the dependency on ps_pmp was missing for dcp as well