Big VMEM stall in FXAA

Hi,

Sorry to disturb you guys. I'm new in rendering field. I tried to migrate your optimization for FXAA to my engine recently. But it turned out that I got a really big VMEM stall when dispatching the first step of FXAA CS. And it caused by texture sampling.

At first, I thought it could be because I was using color texture as input to calculate luminance, which needed to be sampled twice per thread. And the sampling addresses are not so coherent, which may cost a lot of time in TA. So I tried with Gather a prepared luma texture like you did. But the issue was still there. You can refer to the screenshot of my profiling result.

Have you ever met this kind of issue? What could be the reason? Any sugguestions are really appreciated.

Thanks a lot.

Yan

Start by reviewing the FXAA compute-shader first-step dispatch and the texture-sampling behavior described in the issue. Use the attached profiling screenshots to investigate the reported VMEM stall, then compare the color-texture sampling and prepared-luma Gather approaches. Done means identifying a likely cause and documenting actionable guidance.