Hi Oded,
the commit in the subject (5a8888a3119eb5b67f6e90605be9ed2e04807126) somehow causes degraded CPU (!) performance on my Kaveri box. E.g. a piglit run takes about 15% longer, and I suspect the same is true for e.g. compiling the kernel, though I haven't measured that explicitly.
In addition to the measurably worse performance, the commit causes the CPU fan to stay significantly quieter during CPU-bound tasks, so maybe the commit somehow causes the CPU not to achieve turbo clocks or something like that.
I'm attaching my .config. I don't see any difference in dmesg with or without the commit.
On 02/13/2015 12:01 PM, Michel Dänzer wrote:
Hi Oded,
the commit in the subject (5a8888a3119eb5b67f6e90605be9ed2e04807126) somehow causes degraded CPU (!) performance on my Kaveri box. E.g. a piglit run takes about 15% longer, and I suspect the same is true for e.g. compiling the kernel, though I haven't measured that explicitly.
In addition to the measurably worse performance, the commit causes the CPU fan to stay significantly quieter during CPU-bound tasks, so maybe the commit somehow causes the CPU not to achieve turbo clocks or something like that.
I'm attaching my .config. I don't see any difference in dmesg with or without the commit.
Hi Michel, Thanks for catching that. As we talked on IRC, there is indeed a bug in amdkfd. I will publish the fix you tested to be included in 3.20 and mark it as stable. I will also add piglit to my regression tests.
Oded
On 16.02.2015 20:27, Oded Gabbay wrote:
On 02/13/2015 12:01 PM, Michel Dänzer wrote:
Hi Oded,
the commit in the subject (5a8888a3119eb5b67f6e90605be9ed2e04807126) somehow causes degraded CPU (!) performance on my Kaveri box. E.g. a piglit run takes about 15% longer, and I suspect the same is true for e.g. compiling the kernel, though I haven't measured that explicitly.
In addition to the measurably worse performance, the commit causes the CPU fan to stay significantly quieter during CPU-bound tasks, so maybe the commit somehow causes the CPU not to achieve turbo clocks or something like that.
I'm attaching my .config. I don't see any difference in dmesg with or without the commit.
Hi Michel, Thanks for catching that. As we talked on IRC, there is indeed a bug in amdkfd. I will publish the fix you tested to be included in 3.20 and mark it as stable. I will also add piglit to my regression tests.
Sounds good, thanks!
dri-devel@lists.freedesktop.org
-
Michel Dänzer -
Oded Gabbay