vulkan: allow graphics queue only through env var

[0cc4m]

Improve #20551 to fix the reported issues. Only use graphics queue on RADV on larger GPUs.

Fixes #20597

[lemmi]

While graphics queue measurably helps performance on strix halo's 8060s with 40 CUs, desktop performance is unusable now, like with other smaller GPUs, as commented here: #20551 (comment)
I think this either needs to be an option, or a different threshold is necessary.

[0cc4m]

I have updated the PR to require manually setting GGML_VK_ALLOW_GRAPHICS_QUEUE=1 to enable graphics queue use (if a compute queue is available). It seems this is only really safe when headless.

[0cc4m]

@zedbytes That means you should now be using GGML_VK_ALLOW_GRAPHICS_QUEUE=1 instead of RADV_DEBUG=nocompute. This is helpful on headless RADV systems or if you don't mind the graphics interruptions.

[zedbytes]

@0cc4m agreed, building with master and updated flag now !

[zedbytes]

@0cc4m post preliminary testing
I still find ~14% better performance with RADV_DEBUG=nocompute over GGML_VK_ALLOW_GRAPHICS_QUEUE=1
wondering if this gain is restricted only to RDNA 4 as other AMD GPUs aren't getting much performance benefit

[tsterbak]

Quick question: where is this env variable and other available env configuration documented? I cannot find it anywhere :/

[NickM-27]

The findings here seem fairly odd, on linux both my 9060XT and 7900XTX the previous change to use the graphics queue resulted in an increase in TG.