RadeonSI Ups Its Compiler Threads To Let Shader-DB Run Faster On Modern Systems

[nanonyme]

Should hopefully translate to faster load times on many core CPUs for games as well

This. I really hope to see benchmarks with empty cache that show whether this helped in games where startup was slowed by shader compilation

[konserw]

Though in practice, I'm not aware of any computers where it wouldn't be. 12, 16, 24, 32, 36, 48, etc. cores all divide evenly. I guess maybe in a virtual environment there might be an odd # of cores assigned.

https://en.wikipedia.org/wiki/AMD_Phenom#Phenom_X3

+ of course virtual environments

[droste]

Though in practice, I'm not aware of any computers where it wouldn't be. 12, 16, 24, 32, 36, 48, etc. cores all divide evenly. I guess maybe in a virtual environment there might be an odd # of cores assigned.

Even with numbers that are in theory divisible by 4 you may get a different result because floating point numbers have errors https://en.wikipedia.org/wiki/Floati...uracy_problems

/edit:
Removed the wrong example because I'm bad at math in my head

[LinAGKar]

Even with numbers that are in theory divisible by 4 you may get a different result because floating point numbers have errors

Not if you have integral numbers all the way, which is the case here.

[droste]

Not if you have integral numbers all the way, which is the case here.

Yes but there's a division there so the result may not be a integral number anymore. Easy example: 1 / 10 is NOT 0.1 in floating point arithmetic.

/edit: And the original post I replied to had "0.75" which isn't an integral number

[LinAGKar]

Yes but there's a division there so the result may not be a integral number anymore. Easy example: 1 / 10 is NOT 0.1 in floating point arithmetic.

/edit: And the original post I replied to had "0.75" which isn't an integral number

In this case though, we're dividing by 4, which is a power of two, so the result is exactly representable by a float. 0.75 is also exactly representable in binary (0.11₂). Also, I was talking about thread counts that are divisible by 4, like those listed there.

[marek]

marek what were the effects of this change in shaderdb execution time?

shader-db was bottlenecked by the 3-thread shader compiler queue, so it only used 5 threads or so on a 16-thread CPU. Now it's using ~16.

[fuzz]

Any idea what effect this will have on multiple NUMA nodes? Is each thread mostly self-contained or will there be memory swapping across nodes?