This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [RFC][PATCH 0/2] aarch64: Add optimized ASIMD versions of sinf/cosf
- From: Siddhesh Poyarekar <siddhesh at gotplt dot org>
- To: Wilco Dijkstra <Wilco dot Dijkstra at arm dot com>, "Ashwin dot Sekhar at cavium dot com" <Ashwin dot Sekhar at cavium dot com>, "libc-alpha at sourceware dot org" <libc-alpha at sourceware dot org>
- Cc: nd <nd at arm dot com>
- Date: Tue, 13 Jun 2017 16:36:15 +0530
- Subject: Re: [RFC][PATCH 0/2] aarch64: Add optimized ASIMD versions of sinf/cosf
- Authentication-results: sourceware.org; auth=none
- References: <VI1PR0802MB2621642BCA3BB3FF95A405C483C20@VI1PR0802MB2621.eurprd08.prod.outlook.com>
On Tuesday 13 June 2017 04:27 PM, Wilco Dijkstra wrote:
> That is fine for benchmarking the individual code paths of a specific implementation.
>
> However a good benchmark would run through a representative subset of calls from
> actual code rather than repeating the same input many times. This avoids focusing too
> much on special cases that never occur in the real world or failing to take the cost of
> branch mispredictions into account due to varying inputs.
The problem with libm benchmarks is that real world workloads are not
easy to find. For sincos tonto from SPEC2006 may provide some input,
but otherwise there isn't much to talk about. Other more popular
benchmarks (like UnixBench) are again synthetic.
All of the math microbenchmarks in glibc are currently synthetic and aim
to test all code paths. We had once talked about a system benchmarking
project (in 2014 or 2015 I think) that could come up with more relevant
inputs for benchtests but not much has happened since then.
Siddhesh