Am 12.10.2018 um 18:56 schrieb Michel Dänzer:
On 2018-10-11 12:09 a.m., Bhawanpreet Lakha wrote:
From: David Francis <David.Francis@xxxxxxx>
[Why]
A quick-and-dirty way of getting performance data for the amdgpu
driver would make performance improvements easier
[How]
The PERF_TRACE functionality is a tic-toc style debug method.
Put PERF_TRACE calls on either side of the code you want to test.
PERF_TRACE requires access to struct dc_context. PERF_TRACE()
will pick up the CTX macro, and PERF_TRACE_CTX(struct dc_context)
allows you to pass the context explicitly.
The last 20 results can be read through the debugfs entry
amdgpu_perf_trace. Each result contains the time in ns and
number of GPU read/writes since the result before it.
In my experimentation, each PERF_TRACE() call uses at most 700ns
Should this use the Linux tracing infrastructure?
Yeah, agree that looks like reimplementing the tracing infrastructure to
me as well.
Christian.
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
