This is the mail archive of the
mailing list for the glibc project.
Re: [RFC][PATCH 0/2] aarch64: Add optimized ASIMD versions of sinf/cosf
On Friday 23 June 2017 03:20 PM, Sekhar, Ashwin wrote:
> Were the weights for different input ranges based on some benchmark?
No, they were not based on any benchmarks, the lower weighted ones were
basically all of the fast return paths. All other branches (i.e. the
different slow paths) are equally weighted.
> And does the chosen values within an input range have some significance
> or were they random values within that different input range ?
Random values, generated using systemtap probes within parts of code IIRC.
> For sinf/cosf, I am planning to use the same weights as sin/cos and
> random values within different input ranges. Thats why the question.
Yeah the guideline should probably be the same, i.e. don't put much
weightage on the faster paths (i.e. the ones that return 0 or NaN or
similar) for the general set of inputs. In addition to that, if the
functions figure in a known benchmark (like SPEC2006) then add those in
a separate ##name section like Wilco did for powf; going forward those
will get higher weight when comparing performance over the general
inputs since they will be more relevant.