If this is your first visit, be sure to
check out the FAQ by clicking the
link above. You may have to register
before you can post: click the register link above to proceed. To start viewing messages,
select the forum that you want to visit from the selection below.
Problem is that "2D" is a catch-all term covering a wide range of different things, from line-drawing to video playback to the graphics memory manager's ability to deal with a lot of simultaneous graphics activities (eg compositor plus desktop plus apps). Each of these involves different code, different hardware, and different levels of suckage across the different HW and driver offerings.
My impression was that Catalyst *used* to suck at the kind of 2D things measured here but that was greatly improved over the last couple of years.
The depressing thing is that software rendering into a shadow buffer still seems to be the fastest for *these* 2D operations, although of course that falls down if you try to *scroll* a big screen with software rendering. When you see reports that "switching to EXA" makes 2D operations go faster with radeonSI that's not really HW accelerated EXA they're running AFAIK, it's software rendering into a shadow buffer (which is really fast for small operations, which is what you get on typical 2D benchmarks).
Of course the downside of switching to EXA with GCN hardware is losing 3D acceleration as well and falling back to llvmpipe. The real solution is to improve glamor so that more drawing functions are hardware accelerated rather than SW rendering without shadowfb, and that is happening now.
The real solution is to improve glamor so that more drawing functions are hardware accelerated rather than SW rendering without shadowfb, and that is happening now.
Yeah, the slowest things in gtkperf are line drawing and the circles test, mostly because their fallbacks are unoptimized. There have been recent attempts to improve the worst case scenarios for line drawing (which was making certain workflows unusable), but I haven't seen anything for the circles test yet.
Honestly, I'm not sure how much effort I should put into it, given that it sounds like Eric Anholt has some pretty major optimization work that he's preparing for review/merging in the hopefully-near future which will do much more than just incremental improvements to parts of glamor.
I'm not sure, but it won't stop me from hypothesizing
Michael's system is an APU while Pontostroy's using a dedicated GPU. The APU has a core clock of 720MHz and the RAM is DDR3 clocked at 2133MHz. The APU is also sharing the 5MB of L2 cache. The 7790 has a core clock 1050MHz and uses 2GB of dedicated DDR5 memory clocked at 1500MHz (effective clock of 6400Mhz).
The CPU and GPU cores on the APU are sharing the thermal head room, though I don't think the CPU is using much of it during these benchmarks. More benchmarks would be needed to see if it is a hardware limitation or software regression.