Re: [PATCH] Add sinf and cosf traces

[Siddhesh Poyarekar <siddhesh at gotplt dot org>]

On 15/12/18 4:02 AM, Wilco Dijkstra wrote:
> Hi Carlos,
>
> I'll see whether the description could be improved. However the specific details
> don't matter all that much - the most important thing is to ensure a real trace is
> used when optimizing code since repeating the same input in a loop does not give
> useful results at all. Periodic functions using range reduction must have inputs
> which exercise the various code paths.
>
> When replaying the traces I noticed there are very distinct usage patterns across
> the range which showed large variations in throughput/latency. So that's why
> there are 3 separate scenarios representing those cases. These traces can't
> capture the full behaviour of real workloads of course, however they are
> representative enough that faster execution of the traces also speeds up any
> application which uses sin/cos/sincos. And that's what matters.

I think the point of a clearer description is for us to know why we 
added that trace in the first place.  In that sense I would argue that a 
clear description for the data set is as important as the data set 
itself since without the former we're very likely to forget in a few 
years (days/weeks in my case!) why the data set is important.

I'm happy to approve the data set with a clearer description of what it 
is meant to measure.

Siddhesh