Marek Has Taken To Improving Mesa's GLSL Compiler Performance

[pal666]

s/wander/wonder/

[trek]

i wonder which relative impact on fps could have each of faster compiler, background theads compiler and shader cache

instead of an fps increase, I think this will impact more on a smoother experience, with less freeze time when a new shader need to be compiled

[dungeon]

i wander which relative impact on fps could have each of faster compiler, background theads compiler and shader cache

All those features save some CPU time, saving any CPU time anywhere has positive impact everywhere

[pal666]

All those features save some CPU time, saving any CPU time anywhere has positive impact everywhere

sure, but you can save one percent of time or one hundred

[pal666]

instead of an fps increase, I think this will impact more on a smoother experience, with less freeze time when a new shader need to be compiled

what other thing could have most impact on fps then?

[dungeon]

sure, but you can save one percent of time or one hundred

That is how nvidia is fastest driver, they are very fast on the CPU side... so clever CPU render managment is all the magic.

Everybody expected evolution and 10 GHz single core CPUs to appear, so you can say "who cares about that, my next CPU with eat all that"... but as we know that does not happened, so saving CPU time is highest priority if you want render performance

Even with Vulkan, see Dota that is again CPU bound too just less

[pal666]

so it looks like most promising is multithreading, but not only for shader compiler

[pal666]

in other words let marek do it for 17.0

[dungeon]

There was some unsuccessful Intel's patches back from 2012. ... but it is never late to try to do that again, even 5 year later

[microcode]

It is important to note that this will not manifest itself in 25% faster compilation times for radeonsi. This test is done without any backend compiler. It's only the glsl compiler compiling, and then the result is dropped on the floor. For radeonsi it will have to be handed over to LLVM for compiling to bytecode for the graphics card. The benefit will therefore be lower. The article did not really point this out all that well.

I can see how it could have an outsize positive effect on latency for frames with shader compiles. The shader compiler needs to validate the shader synchronously when the application calls glCompileShader; however the driver does not need to do final codegen at that point, and in fact it may need to defer it anyway. Also, having less code invoked through glCompileShader should improve cache locality for a hot loop of shader compiles.

Then again, could turn out not to work that way, who knows.