Date: Sun, 31 Oct 2021 17:47:49 -0700 From: Steve Kargl <sgk@troutmask.apl.washington.edu> To: freebsd-current@freebsd.org Subject: Re: access to ld128 system Message-ID: <20211101004749.GA84142@troutmask.apl.washington.edu> In-Reply-To: <20211028222851.GA24101@troutmask.apl.washington.edu> References: <20211028222851.GA24101@troutmask.apl.washington.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thu, Oct 28, 2021 at 03:28:51PM -0700, Steve Kargl wrote: > kib@ recently committed my implementations of cospi[fl], > sinpi[fl], and tanpi[fl]. These functions have been > extensively tested for float, double, and long double > where long double is the Intel 80-bit long double (e.g., > msun/ld80/s_sinpil.c). The 128-bit versions of these > routines have not been tested (e.g., msun/ld128/s_sinpil.c) > > kib pointed me to a system in the FreeBSD, which I have > access to. Unfortunately, that system has double == > long double, and can only test the support for the > weak references. > > Is there a system with 128-bit long double that I can > have access for some testing? > FYI. An individual has provided access to an aarch64 system. A patch has been submitted to fix the ld128 sinpi, cospi, and tanpi. For the record, I did limited testing to not overwhelm the system. The observed max ULP for sinpi and cospi were less than 1.1 ULP, which is slight worse that the desired less than 1 ULP target. This has been traced to the kernel for computing cosl() in the interval [0,pi/4]. I suspect the minmax polynomial coefficients need refinement. -- Steve
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20211101004749.GA84142>