00:05 <qzed> hmm, it doesn't even try to init the ucode/sqe yet... so gmu is initialized before ucode, is that correct?

00:46 <HdkR> ,,,,,,,,.

00:46 <HdkR> oops.

02:13 <robclark> qzed: so for newer a660+ fw, the afuc disasm actually needs to use emulation mode enough to run thru the bootstrap code (which loads the pm4 packet jumptables).. if it is stuck in a loop, it could be because the initial sqe register values are not correct, so I'd have to look at that.. but gmu is loaded first, it is what powers up the gpu, so you can ignore sqe for a bit

02:14 <robclark> zero length block might be zero-init (thinking in terms of elf binaries.. although that might not be exactly how they represent it embedded in the driver)

02:16 <qzed> thanks! I'll upload the files in a bit (after I'm done cleaning up some other patches...) so you can have a look if you want

02:16 <qzed> I'll also try to have another look at the windows driver gmu loader thing, I guess that could make sense with zero-init

02:54 <robclark> no hurry on the sqe, I won't really have time to look until at least the weekend (but I guess figuring out the couple bits of initial state needed to not get stuck in an infinite loop in the afuc bootstrap emulation isn't really necessary for bringing up the kernel driver)

03:10 <qzed> I've pushed them to https://github.com/linux-surface/aarch64-firmware/tree/main/qcom

03:11 <qzed> no worries, thought it might be an issue with afuc so I thought this might interest you

03:12 <qzed> I guess I'm busy with the gmu for the time being and if I'm lucky the sqe/ucode works out of the box

14:35 <CosmicPenguin> qzed: I assume you've seen https://git.codelinaro.org/clo/la/kernel/msm-5.10/-/blob/KERNEL.PLATFORM.1.0.r1-09000-WAIPIO.0/drivers/gpu/msm/adreno_gen7_gmu.c#L408

15:12 <qzed> CosmicPenguin: I hadn't, thanks!

15:13 <qzed> looks mostly similar to the a6xx_gmu_fw_load() as far as I can tell at a quick glance

15:13 <CosmicPenguin> That mgiht help you make a little bit more sense of what you are seeing

15:14 <qzed> yeah, so I think the zero-length entries are used in combination with another (memory map? table to allocate and map memory regions

15:14 <CosmicPenguin> Yeah. Its kind of elf like except its not elf

15:15 <qzed> and there's a comment that seems to agree with that: https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/msm/adreno/a6xx_gmu.c#L1512

15:16 <qzed> the windows code looks actually quite a bit like the link you've posted

15:19 <CosmicPenguin> I'm not up to date on what the new post 5.10 world looks like on codelinaro after Qualcomm moved to GKI

15:21 <qzed> I'm now trying to validate that everything gets allocated that should be allocated

15:21 <qzed> after that I'll have to try and understand a bit better how the linux driver works and how everything fits together

15:22 <qzed> otherwise I'm flying a bit blind... which isn't really ideal

15:25 <CosmicPenguin> the downstream kernel driver is your best guide to how GMU is loaded at the price of wading through some truly bad code

15:26 <qzed> I guess, yeah. Thanks again for that link!