Tile Shading sample — Area Dispatch path status and experience

[kenchen3000]

Hi ,

I am looking at the samples/tile_shading Area Dispatch path, it seems to be currently disabled with m_IsAreaDispatchActive = False. Could i ask a few questions?

1. Was Area Dispatch intended as a performance optimization or a different programming model exploration? The shading-rate approach seems to trade fine-grained workgroup parallelism (one WG per cluster, with subgroup AABB reduction) for coarse-grained per-tile invocation with per-pixel loops. What was the hypothesis?

2. Did you measure Area Dispatch vs regular tile shading dispatch on Adreno hardware? If so, what were the results? Was it slower, or did it hit correctness issues?

3. Is the missing "PerTileDispatch": true in the JSON intentional or an oversight? Without it, the framework would call vkCmdDispatch instead of vkCmdDispatchTileQCOM, which seems wrong for a shading-rate shader.

Thank you very much!

[RodrigoHolztrattner-QuIC]

Hi @kenchen3000 ,

Area dispatch was actually hardcoded off on purpose. The technique we're showing here (tile based light clustering) doesn't really fit the area dispatch model that well. Clusters don't decompose cleanly into independent screen space regions the way area dispatch assumes, so you don't get the benefits you'd hope for (it's just not the best example to demonstrate area-based dispatch). I just added a few optimizations (subgroup AABB reduction) and re-enabled it (while keeping it mostly 1:1 comparable) but it ends up performing a bit worse than the regular tile shading path on Adreno.

On measurements, we didn't do power profiling on this sample specifically since it's not representative enough of a real game workload to draw useful conclusions from. The meaningful numbers came from a few titles that implemented tile shading and those showed decent power savings when using area dispatch, but to be clear, you probably want to ensure that your technique fits well on area dispatch first, otherwise you could reach a similar situation as this sample and not get any savings and potentially the opposite since you end up discarding a lot of invocations, from personal experience I find it easier to map to ordinary tile shading first and only then consider area based dispatch.

On the PerTileDispatch: true question, yeah that'd be a bug. Without it the framework would just call vkCmdDispatch instead of vkCmdDispatchTileQCOM and the tile attachment reads would break entirely. Fixed with my last push.