Age | Commit message (Collapse) | Author |
|
The (c) is not strictly required, but it was only missing from one file.
|
|
For both vector and scalar routines we reduce the order from 6 to
5. For vector routines, this requires reducing RangeVal as for large
values the tan polynomial is not quite accurate enough. However the
cotan polynomial is used in the inaccurate region in the scalar
routine, so this does not need to change.
Accuracy of scalar routine is unchanged. Accuracy in both vector
routines is now 3.45 ULP, with the same worst-case.
|
|
New routine uses a similar technique to the single-precision Neon
routine, but with an extra reduction to pi/8 using the double-angle
formula. It is accurate to 3.5 ULP.
|
|
Variant was wrongly set in structures used to benchmark SVE functions.
Before this change only half of the lanes were set as expected.
Also reformat for ease of reading.
|
|
All files in pl/math updated to 2023.
|
|
These were technically undefined behaviour - they have been rewritten
without the shift so that their type is unsigned int by default.
|
|
New routine is based on a vector implementation from log1p, which has
been reused (with some modification for improved accuracy close to 0)
from Neon atanh. Accurate to 3.5 ULP.
|
|
New routine uses inlined log1pf helper, and is accurate to 3.1 ULP
(2.8 ULP if fp exceptions are enabled).
|
|
New routines use the same algorithm, reliant on a modified version of
expm1, and are accurate to 3 ULP.
|
|
The new SVE implementation is a direct port of Neon log2, and is
accurate to 2.58 ULPs.
Update error threshold and comments for Neon log2 too, new
approximate argmax but same threshold.
|
|
New SVE routine is an SVE port of the Neon algorithm
and is accurate to 2.48 ULPs.
|
|
New routines are both based on existing log1p routines. Scalar is
accurate to 3 ULP, Neon to 3.5 ULP. Both set fp exceptions correctly
regardless of build config.
|
|
The simplest way to set fenv in Neon atan is by using a scalar
fallback for under/overflow cases, however this routine did not have a
scalar counterpart so we add a new one, based on the same algorithm
and polynomial as the vector variants, and accurate to 2.5 ULP. This
is now used as the fallback for all lanes, when any lane of the Neon
input is special.
|
|
Both routines previously relied on the vector expm1(f) routine exposed
by the library, which depended on WANT_SIMD_EXCEPT for its fenv
behaviour, however both routines were expected to always trigger fp
exceptions correctly. To remedy this, both routines now use an inlined
helper for expm1 (reused from vector tanhf in the case of sinhf), and
special-case small input as well as large when WANT_SIMD_EXCEPT is
enabled.
|
|
The pipe prevented FAILs and PASSs being counted properly - the while
read loop has been rewritten without a pipe, as it was prior to the
changes here.
fenv checking is temporarily disabled in Neon sinh and sinhf, as they
do not get it right. This will be re-enabled once they have been
fixed.
|
|
Updated comment and test threshold.
|
|
The simplest way to set fenv in Neon atanf is by using a scalar
fallback to under/overflow cases, however this routine did not have a
scalar counterpart so we add a new one, based on the same algorithm
and polynomial as the vector variants, and accurate to 2.9 ULP. This
is now used as the fallback for all lanes, when any lane of the Neon
input is special.
|
|
New routines use the same algorithm, with simplified argument
reduction and recombination in the vector variant. Both are accurate
to 2 ULP.
|
|
New max observed - updated filenames, comments and runulp threshold.
|
|
We were previously misusing the WANT_ERRNO build flag. This is now
replaced everywhere appropriate with WANT_SIMD_EXCEPT. A small number
of vector routines get fp exceptions right with no modification - the
tests have been updated to track this.
|
|
A new implementation based on the same approach as
Neon logf, that is accurate to 2.48 ULPs.
Flags set correctly regardless of WANT_ERRNO.
|
|
To conclude the work on simplifying the runulp.sh script, a new macro
has been introduced to specify the intervals in which a routine should
be tested in the routine source. This is eventually consumed by
runulp.sh.
|
|
Introduces a new macro, similar to how ULP thresholds are now
handled, that emits a list of routines which are expected to
correctly trigger fenv exceptions, to be consumed by runulp.sh.
All scalar routines are expected to do so. A small number of Neon
routines are also expected to, dependent on WANT_ERRNO.
|
|
Introduces a new set of macros and Make rules for mechanically
generating a list of ULP limits for each routine, to be consumed
by runulp.sh. This removes the need to maintain long lists of
thresholds in runulp.sh.
|
|
Instead of maintaining three separate lists of routines, which
are cumbersome and prone to merge conflicts, we provide a new
macro, PL_SIG, which by some preprocessor machinery outputs the
lists in the required format (macro formats have been changed
very slightly to make the generation simpler). Only routines with
simple signatures are handled - binary functions still need
mathbench wrappers defined manually. As well, routines with
non-standard references (i.e. powi/powk) still need entries and
wrappers manually defined.
|
|
New behaviour is hidden behind WANT_ERRNO config option.
|
|
New behaviour is hidden behind WANT_ERRNO config option.
|
|
Flags set correctly regardless of WANT_ERRNO.
|
|
Test threshold fixed.
|
|
New behaviour is hidden behind WANT_ERRNO config option.
|
|
Add macros for simplifying polynomial evaluation using either Horner,
pairwise Horner or Estrin. Several routines have been modified to use
the new helpers. Readability is improved slightly, and we expect that
this will make prototyping new routines simpler.
|
|
Small simplification - pl routines do not support different rounding
modes, so there is no need to support them in runulp.sh. As a result
we can also remove Ldir.
|
|
Special lanes were not being properly masked when a lane was
tiny. This is now fixed.
|
|
Fixing a bug that resulted in potentially random results
in boring domain by saturating index at an appropriate value.
|
|
Routine no longer relies on vector log1pf, as this has to become more
complex to deal with fenv itself. Instead we re-use a log1pf helper
from Neon atanhf which does no special-case handling, instead leaving
it all up to the main routine. We now just fall back to the scalar
routine for special-case handling. This uncovered a mistake in
asinhf's handling of NaNs, which has been fixed.
|
|
The ldexp shortcut was left-shifting a signed value. We now bias the
exponent first, will allows the shift to be done on an unsigned value.
|
|
Both routines use simplified inline versions of expm1f, and are
accurate to 2.6 ULP.
|
|
New routine uses two separate algorithms for input greater and less
than 1 (similar to the scalar routine). It is accurate to 2.5 ULP.
|
|
New max observed for both Neon and SVE.
|
|
Both routines use the same algorithm - one Newton iteration with the
initial guess obtained by a low-order polynomial. Scalar is used as a
fallback for subnormal and special cases for the vector routine, which
allows vastly simplified argument reduction and reassembly. Both
routines accurate to 1.5 ULP.
|
|
Both routines are based on a simplified version of log1pf, and are
accurate to 3.1 ULP. Also enabled -c flag from runulp.sh - we need
this for atanhf so that we can set the control lane to something other
than 1, since atanh(1) is infinite.
|
|
New routines are based on double-precision exp, both accurate to 2 ULP.
|
|
New routines are based on the single-precision versions and are
accurate to 3 ULP.
|
|
Reduce the order of the polynomial used in Neon log2 by one (from 7 to
6). In order to calculate the new coefficients required we rescale the
coefficients from log_data.c by log2(e) in extended precision and
round back.
The maximum observed error is unchanged (2.59 ULPs) but the point at
which it is observed has changed slightly.
|
|
There is collision for math-tests and math-rtests between math/ and
pl/math, which can lead to failures if running both concurrently. We
rename the pl-specific lists to avoid this.
|
|
Extra special-case check.
|
|
These were broken in the previous patch, now fixed.
|
|
Reduces the amount of boilerplate developers need to write for new
routines.
|
|
New routine is a vector port of the scalar algorithm, with fallback to
the scalar variant for large and special input. This enables us to
simplify elements of the algorithm which were necessary for large
input. It also means that, as long as we fall back to the scalar for
tiny input as well (dependent on the value of WANT_ERRNO), the routine
sets fenv flags correctly.
|
|
New routine uses the same algorithm as the single-precision routine,
and is accurate to 2.5 ULP.
|