GLSL Copy Propagation Optimizations For Mesa

[phoronix]

Phoronix: GLSL Copy Propagation Optimizations For Mesa

A developer has published a set of 14 patches providing copy propagation optimizations for Mesa's GLSL/Nir code...

http://www.phoronix.com/scan.php?page=news_item&px=Mesa-GLSL-Copy-Propagation-Opt

[rene]

With some luck, and hard work, maybe 2017 will be the year were Mesa's GL performance will overtake the binary only blobs in more benchmarks ;-)

[slacka]

According to the bug report Nvidia proprietary takes 1.2 seconds ( with shader cache disabled), Intel proprietary takes 7 seconds, but Intel Mesa takes 262 seconds! Are the Intel Linux guys not allowed to talk to the WIndows guys? Is there cooperation or an information firewall for some reason?

[helland]

According to the bug report Nvidia proprietary takes 1.2 seconds ( with shader cache disabled), Intel proprietary takes 7 seconds, but Intel Mesa takes 262 seconds! Are the Intel Linux guys not allowed to talk to the WIndows guys? Is there cooperation or an information firewall for some reason?

Since I wrote the patches, I feel like I can give a short and concise answer to this: Lack of time and prioritization. There really is no need to speak to the Windows guys at all about this. It is well known compiler theory, and nothing really fancy. The fact that we are dealing with glsl, instead of say, C, changes things slightly, but not in a way that makes it impossible to do this better. The shader in question is also a torture test, and is in no way indicative of an actual real life workload. Since this is not a big issue with normal workloads, time has been spent on other things.

Following from the above: This is not the patch series that will dramatically improve shader compile times. In fact, it barely shows any changes on real life workloads. The patch series will also likely be respun to a different approach that should prove even more benefitial. There are however other work going on in the parser-lexer area that will prove much more benefitial for actual real-life workloads. But that work is not mine, so I'll let the author speak for himself.