Am 11.04.2014 11:54, schrieb Lauri Kasanen:
On Fri, 11 Apr 2014 10:33:08 +0200 Christian König deathsimple@vodafone.de wrote:
Actually direct register access shouldn't be necessary so often. Apart from page flips, write/read pointer updates and irq processing there shouldn't be so many of them. Could you clarify a bit more what issue you are seeing here?
Too much cpu usage for such a simple function. 2% makes it #2 in top-10 radeon.ko functions, right after evergreen_cs_parse. For reference, #3 (radeon_cs_packet_parse) is only 0.5%, one fourth of this function's usage.
I think you misunderstood me here. I do believe your numbers that it makes a noticeable difference.
But I've did a couple of perf tests recently on SI and CIK while hacking on VM support, and IIRC r100_mm_rreg didn't showed up in the top 10 on those systems.
So what puzzles me is who the hack is calling r100_mm_rreg so often that it makes a noticeable difference on evergreen/NI?
The biggest caller is cayman_cp_int_cntl_setup. Before inlining it took 0.0013%, after it takes 1%.
Sounds like somebody is constantly turning interrupts on and off.
This is on a Richland APU, so Aruba/Cayman. Urban Terror is an ioq3 game with a lot of cpu-side vertex submissions.
That will probably be the difference, I only tested lightsmark.
Anyway, I would do like Ilia suggested and only put the else branch into a separate, not inlined function.
BTW: It's probably a good idea to do the same for the write function as well.
Christian.
- Lauri