Enabling Metal argument buffers in 1.2.10+ breaks rendering in Dolphin

[OatmealDome]

It appears that using Metal argument buffers, which is enabled by default in 1.2.10+, breaks rendering in Dolphin. Output is either a black screen or corrupted:

Expected Output

Apple M1

Intel UHD 630

Here's a GPU trace from the M1: LineWidthTest.gputrace.zip

If you would like to run the test for yourself, you can download a build of Dolphin with the latest MoltenVK commit (1d2ec638532268d076a550262df2e5e5e4fd2c9e) here: pr-13205-dolphin-latest-universal.dmg. Switch the video backend in Graphics button -> General tab to Vulkan, and then drag and drop this file on the Dolphin window to start rendering the test image: LineWidthTest.dff.

Turning off Metal argument buffers with MVK_CONFIG_USE_METAL_ARGUMENT_BUFFERS=0 resolves the issue.

I was able to reproduce the issue on:

• M1 Pro / Sequoia 15.2 beta 4 (black screen)
• M1 / Sonoma 14.6.1 (black screen)
• Intel UHD 630 / Sequoia 15.1 (corruption)

I was unable to reproduce the issue on:

• AMD Radeon Pro 5500M / Sequoia 15.1

[billhollings]

Sorry for the delay in getting to this.

It seems to run fine on my M3 Pro. I don't have an M1 to verify your specific condition.

I was not able to start your app from Xcode, so I was not able to either set environment variables, or review the sys err logs to confirm state such as argument buffer use. If you can let me know how to do those, I'll have another shot at it.

[squidbus]

For the Intel corruption specifically, I've been seeing the same in multiple apps on Intel GPUs.

[dboyan]

I did some investigation earlier, this is once again related to combined image-sampler in argument buffer, at least on M1.

I was able to reproduce the issue on my m1, and I recorded the metal API trace with argument buffers enabled and disabled from rendering a single frame. I'm attaching them here.

• Argument buffer enabled: dolphin-argbuf.gputrace.zip
• Argument buffer disabled: dolphin-noargbuf.gputrace.zip (compare call 307 in the former and 328 in the latter)

Note: they are generated when replaying the following vulkan gfxr trace dolphin.gfxr.zip. I cannot guarantee these can be replayed on other platforms, but hopefully it is more or less possible on other Apple Silicons at least.

When argument buffer is enabled, GLSL layout(binding = 0) uniform sampler2DArray samp0; is translated into a single pair of texture2d_array and sampler:

struct spvDescriptorSetBuffer1
{
    texture2d_array<float> samp0 [[id(0)]];
    sampler samp0Smplr [[id(1)]];
};

In the meantime, the application writes an array of 8 image info into the descriptor binding. During this process (there is no bound check):

• element 0 writes texture (image) info to [[id(0)]] and sampler to [[id(1)]]
• element 1 writes texture (image) info to [[id(1)]] (BOOM) and sampler to [[id(2)]]
• ...

In the end, the memory around the argument buffer looks like (only the first row is the argument buffer and you can see the OOB writes and overwrites):

I'm not an enough expert to say if the application or spirv-cross translation has their share of the issue here but I'd say the overwriting between overlapping ids between textures and samplers is pretty evil. (when argument buffer is not on, textures and samplers live in different spaces so they won't have id clash)

I came up with the hacky #2427 that is able to address the symptom. The idea is to use stride of 2 when argument buffer is enabled. So the above process becomes:

• element 0 writes texture (image) info to [[id(0)]] and sampler to [[id(1)]]
• element 1 writes texture (image) info to [[id(2)]] and sampler to [[id(3)]]
• ...

They still write out of the bound, but they no longer overwrite each other. I'm not sure if that's the right way to go, so I left it as a draft there for people with better ideas to figure out.

[squidbus]

For what its worth I don't think the corruption on Intel is related to combined image-sampler in argument buffers, as I've seen it in apps that are not using that.

[dboyan]

Well, yeah, maybe there are other issues involved. Unfortunately I don't have intel macs to reproduce them and understand what's happening there. But combined image-sampler is definitely broken here, and that was the cause on m1.

On a side note, there is a bigger issue with overlapping ids. If we have:

layout(binding = 0) uniform sampler2D samp0;
layout(binding = 1) uniform sampler2D samp1;

this will break when argument buffer is on, and I think that's what #2395 is about.