rx7900xt regression on av1 decoding since contiguous vram flag changes

On my VCN 4.0 based decode hardware, AV1 decoding with radv has been broken since

e362b7c8f8c7af00d06f0ab609629101aebae993 Author: Arunpravin Paneer Selvam Arunpravin.PaneerSelvam@amd.com Date: Thu Apr 25 21:05:55 2024 +0530

drm/amdgpu: Modify the contiguous flags behaviour

It just decodes garbage or green bits.

added 7000 dGPU series AMDgpu TTM VCN bisected labels

@arunpravin24 @ckoenig @agd5f any ideas?

Mhm, not of hand. What could be is that the AV1 decoder has some undocumented requirements for linear memory.

Going to discuss that with Leo and the multimedia team today.

It's also strange, vaapi decode still seems to work, but vulkan doesn't, I've played with a bunch of placements trying to figure it out, but to do no avail so far. Also only seems to be on gfx11

Okay I tracked this down to the vcn ib patching code not working, I'll send a patch to fix the regression

Yeah, we figured that out as well. Question is why the IB patching isn't working.

Somebody from Leo's team is already looking into that.

it's because you dropped the CONTIGUOUS flag from non-kernel bos, patch on the list in 5 minutes

I wasn't able to figure out why it would not work with non-contiguous, so my patch is to make sure the video bos used with the IB parsing will be contiguous.

0001-drm-amdgpu-Ensure-contiguous-memory-in-amdgpu_cs_fin.patch

Another option would be to remove the check for kernel bos as @airlied said, but I assume that was actually intended?

0001-amdgpu-allow-setting-contiguous-on-non-kernel-bos-fo.patch

mine just reverted that one liner, but yes if we do it another way I'm also fine

non-contiguous BO's in VRAM can't be ioremapped

No, that is certainly incorrect. We had non-contiguous mappings before and they should have worked.

the code forces the IB to be contiguous before mapping it.

and that used to work, but then you add the kernel bo only check and it fails, since there are user bos

Ah, so the code in amdgpu_cs_find_mapping() doesn't work any more. Ok that makes perfect sense.

Give me a moment to fix that.

yes that breaks, maybe we used to have IBs in GTT before?

since amdgpu_ttm_alloc_gart is called there, that would have worked for objects placed in gart

radeonsi have IBs in GTT, so that's why vaapi isn't affected.

makes sense now. radv has started using VRAM IBs

Yeah, exactly that.

Give me a second to fix that properly. It's just like 4 lines of code changed.

Potential fix for the issue.

0001-drm-amdgpu-fix-contiguous-handling-for-IB-parsing.patch

this patch doesn't seem sufficient, but I'll dig in a bit more once I get out of meetings

needs

(*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;

also since later code checks for both (amdgpu_vram_mgr.c)

Good point. That is probably a separate bug as well. But that can come later on, let's get that issue here fixed for now.

0001-drm-amdgpu-fix-contiguous-handling-for-IB-parsing-v2.patch

mentioned in issue #3510

closed

@ckoenig The patch you mentioned above (merged since 6.10.4) has broken H.264 hardware decoding on GCN 1.0 and 2.0 (AMD Radeon 520 and AMD Radeon R9 390X, as reported by another fellow user).

I have bisected the kernel to the commit f3572db3c049b4d32bb5ba77ad5305616c44c7c1, reverting it has fixed hardware decoding on my Radeon 520.

See #3599 (closed).

mentioned in issue #3599 (closed)

mentioned in commit agd5f/linux@6f4b6d55

mentioned in commit agd5f/linux@12f325bc