Intel Still Making Source Engine Performance Optimizations

[phoronix]

Phoronix: Intel Still Making Source Engine Performance Optimizations

For those not actively following the Mesa Git repository, there continues to be new performance-optimizing patches flowing in from Intel's developers for their open-source Linux graphics driver...

http://www.phoronix.com/vr.php?view=MTQzNjQ

[dh04000]

That looks to my untrained eye to be a big improvement in reduction of instructions. Someone will have to actually benchmark to see what the real world effects will be though.

[leif81]

I still have to pinch myself everytime I see a title like that.

[smitty3268]

Phoronix: Intel Still Making Source Engine Performance Optimizations

For those not actively following the Mesa Git repository, there continues to be new performance-optimizing patches flowing in from Intel's developers for their open-source Linux graphics driver...

http://www.phoronix.com/vr.php?view=MTQzNjQ

I'm pretty sure this is one of the optimizations (but not the only) enabled for radeon by setting R600_DEBUG=sb on r600g.

So, judging by that, it's unlikely to make much difference on the benchmarks michael runs, but can be very big on real applications like the Source games or WINE.

[vadimg]

I'm pretty sure this is one of the optimizations (but not the only) enabled for radeon by setting R600_DEBUG=sb on r600g.

Yes, r600g/sb has an if-conversion pass that produces the same result for the IF-MOV-ELSE-MOV-ENDIF pattern, though it's a bit more generic and handles almost any suitable conditional code, e.g. with more than one MOV in each branch, other instructions, nested IFs, etc.

So, judging by that, it's unlikely to make much difference on the benchmarks michael runs, but can be very big on real applications like the Source games or WINE.

Many apps in Michael's benchmarks don't use shaders with conditional control flow at all, in such cases these optimizations have nothing to work on.

[curaga]

This sounds like something the GLSL compiler should do, not the drivers. Why not? Does the IR not have a SEL equivalent?

[Ancurio]

This sounds like something the GLSL compiler should do, not the drivers. Why not? Does the IR not have a SEL equivalent?

Where do you think the compiler lives?

[curaga]

Where do you think the compiler lives?

It lives in src/glsl, while the drivers live under src/mesa/drivers and src/gallium/drivers. Having this kind of simplification in the GLSL compiler would be in one place, instead of duplicated in every driver.

[shirish]

does those optimizations in anyway relate to older cpus as well

Hi all,
Do these source engine perf. optimizations have any advantage for older generations of Intel CPU's as well ?

[dh04000]

Anyone benchmark this yet?