Reducing The CPU Usage In Mesa To Improve Performance

[dungeon]

@tarceri

It is Athlon 5350 APU so Kabini, has sse4.1... It is not gaming machine, but OK for example CPU usage decreased by 3% and fps dropped by 3% in extremetuxracer.

[tarceri]

@tarceri

It is Athlon 5350 APU so Kabini, has sse4.1... It is not gaming machine, but OK for example CPU usage decreased by 3% and fps dropped by 3% in extremetuxracer.

Hmm, it could be as leonmaxx said the frequency is dropping due to a lesser load, but I would have thought that would caused a bigger drop in framerates. I'll take a look at extremetuxracer later on today, I might not be able to reproduce your results but it would be interesting to see what its doing.

[dungeon]

No, what leonmaxx said is not valid for me, because i don't use cpufreq governor... My kernels even does not have that compiled, can't recommend the same for beefier AMD CPU/APUs and laptops of course but my low power desktop Kabini can live perfectly fine without CPU frequency scaling .

[tarceri]

No, what leonmaxx said is not valid for me, because i don't use cpufreq governor... My kernels even does not have that compiled, can't recommend the same for beefier AMD CPU/APUs and laptops of course but my low power desktop Kabini can live perfectly fine without CPU frequency scaling .

hmm ok, how many times did you run the benchmark? It can be hard to get consitent results sometime. Just looking for an easy answer

[Michael]

Michael: How about getting inspired and making it easier to run profilers with pts? According to the blog it at least used to be a tad cumbersome

The issues expressed in the blog can actually be really easily addressed, but Arceri hadn't explained in full until now his issues (IIRC) but in general without a pressing need / use-case (though would have gladly made the changes back when he was working on the module) or enterprise work item, too busy working on other customer work to just do random improvements without cause.

[dungeon]

hmm ok, how many times did you run the benchmark? It can be hard to get consitent results sometime. Just looking for an easy answer

Few times, so doblechecked . That is how it is for me, CPU usage is slightly lower and fps numbers are slightly lower . So i asking myself - what is actually improvment there

I use kernel 3.18-rc2 right now... 3.16, 3.17, 3.18 all plays (slightly) differently, so easy answer can't be so easy , because of memory managment and changes related to that, for example in mesa PIPE_USAGE_DYNAMIC use RADEON_FLAG_CPU_ACCESS as of kernel 3.17, so some change in CPU sector i guess practicaly can affect fps and actually can be in some cases also be negative if optimized or CPU used is lower... huh it is memory managment pretty tough stuff to understand why happens, but it happens .

http://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/radeon/r600_buffer_common.c

In other cases it is similar or slightly lower fps numbers, but never same/better... but OK, someone else i guess might see improvment on different setup, if it is at least the same for me i will not complain

[cbxbiker61]

Support for CPUS that aren't SSE4.1 capable patch

I threw together a patch which adds an SSE2 version for those CPU's that aren't SSE4.1 capable after reading this page.

http://fossies.org/dox/FRAMEWAVE_1.3.1_SRC/SSE2Intrinsic_8h_source.html

Original patch:

http://www.xilka.com/xilka/source/tmp/X_Mesa-add-SSE-opt-for-glDrawElements.patch

Extension patch that also supports SSE2.

http://www.xilka.com/xilka/source/tmp/X_Mesa-add-SSE-opt-for-glDrawElements-2.patch

[tarceri]

The issues expressed in the blog can actually be really easily addressed, but Arceri hadn't explained in full until now his issues (IIRC) but in general without a pressing need / use-case (though would have gladly made the changes back when he was working on the module) or enterprise work item, too busy working on other customer work to just do random improvements without cause.

Thats correct I hadn't explained the issues I was having, and moved on to working on other things. So much to work on so little time :-(

Out of interest what do you sugget as an easy fix? I may pick up work on the module again at some point if I get a chance. Or someone else might be interested in having a go at finishing it off.

[tarceri]

I threw together a patch which adds an SSE2 version for those CPU's that aren't SSE4.1 capable after reading this page.

http://fossies.org/dox/FRAMEWAVE_1.3.1_SRC/SSE2Intrinsic_8h_source.html

Original patch:

http://www.xilka.com/xilka/source/tmp/X_Mesa-add-SSE-opt-for-glDrawElements.patch

Extension patch that also supports SSE2.

http://www.xilka.com/xilka/source/tm...ements-2.patch

There has also been an alternate SSE2 version posted to the mailing list: http://lists.freedesktop.org/archive...er/069886.html

[tarceri]

Few times, so doblechecked . That is how it is for me, CPU usage is slightly lower and fps numbers are slightly lower . So i asking myself - what is actually improvment there

This is a long shot but does changing the first if statment from:

if (aligned_count >= 4) {

to

if (aligned_count >= 8) {

help at all? If the array size was 4 through to 7 then the patch actually made things worse. Changing this made a very slight improvement to cpu in OpenArena so it does some to make a difference. I'll be able to check extremetuxracer myself later tonight.