yao.qi at linaro.org
Mon Aug 9 11:42:17 BST 2010
Michael Hope wrote:
> # Future focus
> * Emphasise performance, neutral on correctness
Usually, the starting point of performance tuning cycle is to identify
some scenarios that performance is very poor. Then, profile it to
identify bottlenecks, and figure out how to generate more efficient code
from compiler. Some one call it as "Workload Optimization".
Do we have some typical workloads so far? We had a discussion some
weeks ago on benchmarks, and personally, I prefer scenarios/workload
than benchmarks, because it is more useful to improve the performance of
a real scenario/workload, from the performance tuning's point of view.
In one word, we should get workload/scenario/benchmark ready before we
start to tune the performance.
Any performance profilers available so far? Oprofile is widely used on
Linux, but I haven't ever tried to use Oprofile on ARM.
> * Topics to investigate
> o Vectoriser
> o Pipeline
> o PGO
> o Upstream research topics
I agree with Michael that we'd better talk with someone who is using
linaro or building software on top of ARM. Their feedback/comment is
very useful for us to determine which should be done first, or to
prioritize these items.
Once the typical workloads/scenarios are got, it is not hard to figure
out which items we should do first.
More information about the linaro-toolchain