Help Wanted Dithering Algorithms Lag on Metal; Not Others

Hi all! A bit of a strange issue. I have been working on a Unity project in both a Windows environment and an OSX environment, and both are pretty capable machines. Obviously the Windows has the better GPU but the OSX puts up a good fight and can, for example, handle pretty expensive blurs. However I run into the weirdest example - dithering blurs to save on GPU cost actually seems to 100% my OSX (and ONLY on OSX) GPU for seemingly no reason.

I've tried two algorithms:
- return frac(dot(float3(Pos.xy, FrameIndexMod4), uint3(2, 7, 23) / 17.0f)); (https://developer.oculus.com/blog/tech-note-shader-snippets-for-efficient-2d-dithering/)
- https://docs.unity3d.com/Packages/com.unity.shadergraph@6.9/manual/Dither-Node.html (4x4 and 8x8)

I genuinely have no clue why dither of all things would perform so poorly on Metal, but I really would like to find a fix for it! I have narrowed it down to these pieces of code so I am 100% sure there is something with them.

Thanks all

 

Maybe mac GPU bad at uint's? Did you try to change uint to float?

 

Yeah, I thought that might be the case - I tried replacing all uint's with both normal integers and floats, and even tried to pre-calculate some of the math (although less accurate) to see if that was causing the issue. No luck on any of it.

 

Have you tried using a dither texture to see if that's any better? My theory is the noise function is a red herring itself, and the cost is coming from the change in memory access patterns from using a noise presumably to jitter the sample pattern that's the performance issue. Mobile GPUs are generally very bandwidth limited, and Apple's GPUs are no different. There might be some fancy prediction Apple's GPUs are doing that a noisy sample pattern is tripping up.

 

It seems like using a dithering map works! I tried both a 16x16 and 32x32 map, and didn't run into the same issue. Nice catch, and thanks a lot for the suggestion. Just for my own knowledge and future reference, is there any way to "bypass" this prediction model on Apple's GPUs? I'm not too concerned about mobile, but I feel like modern generation Apple Laptops/Desktops are more than capable of handling psuedo-random dithering. For reference, I primarily want to get this working (from Oculus's dev blog on optimized dithering in VR):

Code (CSharp):

1. frac(dot(seed, uint2(2, 7) / 17.0));

Much thanks either way!

 

Well, if the noise texture worked, then my prediction for the cause was wrong ... and I have no idea what the issue actually is.

 

My bad, let me rephrase: I used a dithering texture, not a noise. Like an obra-dithering or similar, it was patterned to look like a dither from the start. I didn't use a psuedo-random texture. My bad for the confusion

 

Dither or noise, it doesn't really matter. The fact using a texture removed the perf problem vs using a very simple bit of math is very confusing.