RADV Vulkan Squeezes A Bit More Performance Out Of Mad Max

[phoronix]

Phoronix: RADV Vulkan Squeezes A Bit More Performance Out Of Mad Max

David Airlie has managed to squeeze a bit more performance out of Feral Interactive's Mad Max Vulkan port with one line of code change to RADV...

http://www.phoronix.com/scan.php?page=news_item&px=RADV-FMULADD

[amehaye]

10% improvement for such a small change is nothing to sneer at.
Also remember that even very small changes sometimes require a lot of work in tracking down the exact problem and solution.

[jaxxed]

yeah that level of interaction must be great.

Perhaps is should be reinforced how small that patch was too, so it's not like a lot of custom coding was required, just a different operation (?) used.

[haagch]

I'm somehow reminded of a compiler lecture where the professor mentioned that for some compilers the instruction selection is actually bruteforced. They take a target CPU and for commom mathematical operations they just generate random instruction sequences and hope to find one that produces the same results but runs faster.

[R41N3R]

Now Max Max runs on Vulkan even better :-) Thanks

[elldekaa]

I am curious about how he is tracking such performance improvement... Is there a specific protocol with different GPUs? Are they using phoronix tools?

[Michael]

I am curious about how he is tracking such performance improvement... Is there a specific protocol with different GPUs? Are they using phoronix tools?

In the commit he says he is using the PTS (Phoronix Test Suite) test for comparison.

[orome]

I'm somehow reminded of a compiler lecture where the professor mentioned that for some compilers the instruction selection is actually bruteforced. They take a target CPU and for commom mathematical operations they just generate random instruction sequences and hope to find one that produces the same results but runs faster.

the approach is called 'superoptimization', modern approach is to generate replacement directories (inst seq -> superoptimized inst sequence).

just a different operation (?) used.

it's not a different operation, just slightly different semantics. FMA traditionally require 'fused' semantics, i.e. the multiply and add operations happen without rounding the intermediate result. AMD GPUs implement this by using double precision and the instruction runs at DP rate (usually 1/8 or 1/16 of SP rate).
The alternative is MAD (multiply-add) which gives the same rounding/precision as mul+add but it's faster (usually single cycle)

[elldekaa]

In the commit he says he is using the PTS (Phoronix Test Suite) test for comparison.

Very nice!

[stqn]

Reaching 60 FPS at high settings at 4K is very good, but I wonder what the hardware is…