In e_log.c, there was just a off-by-1 (1 ulp) error in the comment
about the threshold. The precision of the threshold is unimportant,
but the magic numbers in the code are easier to understand when the
threshold is described precisely.
In e_logf.c, mistranslation of the magic numbers gave an off-by-1
(1 * 16 ulps) error in the intended negative bound for the threshold
and an off-by-7 (7 * 16 ulps) error in the intended positive bound for
the threshold, and the intended bounds were not translated from the
double precision bounds so they were unnecessarily small by a factor
of about 2048.
The optimization of using the simple Taylor approximation for args
near a power of 2 is dubious since it only applies to a relatively
small proportion of args, but if it is done then doing it 2048 times
as often _may_ be more efficient. (My benchmarks show unexplained
dependencies on the data that increase with further optimizations
in this area.)
2**-28 as a side effect, by merging with the float precision version
of tanh() and the double precision version of sinh().
For tiny x, tanh(x) ~= x, and we used the expression x*(one+x) to
return this value (x) and set the inexact flag iff x != 0. This
doesn't work on ia64 since gcc -O does the dubious optimization
x*(one+x) = x+x*x so as to use fma, so the sign of -0.0 was lost.
Instead, handle tiny x in the same as sinh(), although this is imperfect:
- return x directly and set the inexact flag in a less efficient way.
- increased the threshold for non-tinyness from 2**-55 to 2**-28 so that
many more cases are optimized than are pessimized.
Updated some comments and fixed bugs in others (ranges for half-open
intervals mostly had the open end backwards, and there were nearby style
bugs).
functions are only for compatibility with obsolete standards. They
shouldn't be used, so they shouldn't be optimized. Use the generic
versions instead.
This fixes scalbf() as a side effect. The optimized asm version left
garbage on the FP stack. I fixed the corresponding bug in the optimized
asm scalb() and scalbn() in 1996. NetBSD fixed it in scalb(), scalbn()
and scalbnf() in 1999 but missed fixing it in scalbf(). Then in 2005
the bug was reimplemented in FreeBSD by importing NetBSD's scalbf().
The generic versions have slightly different error handling:
- the asm versions blindly round the second parameter to a (floating
point) integer and proceed, while the generic versions return NaN
if this rounding changes the value. POSIX permits both behaviours
(these functions are XSI extensions and the behaviour for a bogus
non-integral second parameter is unspecified). Apart from this
and the bug in scalbf(), the behaviour of the generic versions seems
to be identical. (I only exhusatively tested
generic_scalbf(1.0F, anyfloat) == asm_scalb(1.0F, anyfloat). This
covers many representative corner cases involving NaNs and Infs but
doesn't test exception flags. The brokenness of scalbf() showed up
as weird behaviour after testing just 7 integer cases sequentially.)
to scalbf(), but ldexpf() cannot be implemented in that way since the
types of the second parameter differ. ldexpf() can be implemented as
a weak or strong reference to scalbnf() (*) but that was already done
long before rev.1.10 was committed. The old implementation uses a
reference, so rev.1.10 had no effect on applications. The C files for
the scalb() family are not used for amd64 or i386, so rev.1.10 had even
less effect for these arches.
(*) scalbnf() raises the radix to the given exponent, while ldexpf()
raises 2 to the given exponent. Thus the functions are equivalent
except possibly for their error handling iff the radix is 2. Standards
more or less require identical error handling. Under FreeBSD, the
functions are equivalent except for more details being missing in
scalbnf()'s man page.
commit broke the 2**24 cases where |x| > DBL_MAX/2. There are exponent
range problems not just for denormals (underflow) but for large values
(overflow). Doubles have more than enough exponent range to avoid the
problems, but I forgot to convert enough terms to double, so there was
an x+x term which was sometimes evaluated in float precision.
Unfortunately, this is a pessimization with some combinations of systems
and compilers (it makes no difference on Athlon XP's, but on Athlon64's
it gives a 5% pessimization with gcc-3.4 but not with gcc-3.3).
Exlain the problem better in comments.
algorithm for the second step significantly to also get a perfectly
rounded result in round-to-nearest mode. The resulting optimization
is about 25% on Athlon64's and 30% on Athlon XP's (about 25 cycles
out of 100 on the former).
Using extra precision, we don't need to do anything special to avoid
large rounding errors in the third step (Newton's method), so we can
regroup terms to avoid a division, increase clarity, and increase
opportunities for parallelism. Rearrangement for parallelism loses
the increase in clarity. We end up with the same number of operations
but with a division reduced to a multiplication.
Using specifically double precision, there is enough extra precision
for the third step to give enough precision for perfect rounding to
float precision provided the previous steps are accurate to 16 bits.
(They were accurate to 12 bits, which was almost minimal for imperfect
rounding in the old version but would be more than enough for imperfect
rounding in this version (9 bits would be enough now).) I couldn't
find any significant time optimizations from optimizing the previous
steps, so I decided to optimize for accuracy instead. The second step
needed a division although a previous commit optimized it to use a
polynomial approximation for its main detail, and this division dominated
the time for the second step. Use the same Newton's method for the
second step as for the third step since this is insignificantly slower
than the division plus the polynomial (now that Newton's method only
needs 1 division), significantly more accurate, and simpler. Single
precision would be precise enough for the second step, but doesn't
have enough exponent range to handle denormals without the special
grouping of terms (as in previous versions) that requires another
division, so we use double precision for both the second and third
steps.
distributed non-large args, this saves about 14 of 134 cycles for
Athlon64s and about 5 of 199 cycles for AthlonXPs.
Moved the check for x == 0 inside the check for subnormals. With
gcc-3.4 on uniformly distributed non-large args, this saves another
5 cycles on Athlon64s and loses 1 cycle on AthlonXPs.
Use INSERT_WORDS() and not SET_HIGH_WORD() when converting the first
approximation from bits to double. With gcc-3.4 on uniformly distributed
non-large args, this saves another 4 cycles on both Athlon64s and and
AthlonXPs.
Accessing doubles as 2 words may be an optimization on old CPUs, but on
current CPUs it tends to cause extra operations and pipeline stalls,
especially for writes, even when only 1 of the words needs to be accessed.
Removed an unused variable.
function approximation for the second step. The polynomial has degree
2 for cbrtf() and 4 for cbrt(). These degrees are minimal for the final
accuracy to be essentially the same as before (slightly smaller).
Adjust the rounding between steps 2 and 3 to match. Unfortunately,
for cbrt(), this breaks the claimed accuracy slightly although incorrect
rounding doesn't. Claim less accuracy since its not worth pessimizing
the polynomial or relying on exhaustive testing to get insignificantly
more accuracy.
This saves about 30 cycles on Athlons (mainly by avoiding 2 divisions)
so it gives an overall optimization in the 10-25% range (a larger
percentage for float precision, especially in 32-bit mode, since other
overheads are more dominant for double precision, surprisingly more
in 32-bit mode).
- in preparing for the third approximation, actually make t larger in
magnitude than cbrt(x). After chopping, t must be incremented by 2
ulps to make it larger, not 1 ulp since chopping can reduce it by
almost 1 ulp and it might already be up to half a different-sized-ulp
smaller than cbrt(x). I have not found any cases where this is
essential, but the think-time error bound depends on it. The relative
smallness of the different-sized-ulp limited the bug. If there are
cases where this is essential, then the final error bound would be
5/6+epsilon instead of of 4/6+epsilon ulps (still < 1).
- in preparing for the third approximation, round more carefully (but
still sloppily to avoid branches) so that the claimed error bound of
0.667 ulps is satisfied in all cases tested for cbrt() and remains
satisfied in all cases for cbrtf(). There isn't enough spare precision
for very sloppy rounding to work:
- in cbrt(), even with the inadequate increment, the actual error was
0.6685 in some cases, and correcting the increment increased this
a little. The fix uses sloppy rounding to 25 bits instead of very
sloppy rounding to 21 bits, and starts using uint64_t instead of 2
words for bit manipulation so that rounding more bits is not much
costly.
- in cbrtf(), the 0.667 bound was already satisfied even with the
inadequate increment, but change the code to almost match cbrt()
anyway. There is not enough spare precision in the Newton
approximation to double the inadequate increment without exceeding
the 0.667 bound, and no spare precision to avoid this problem as
in cbrt(). The fix is to round using an increment of 2 smaller-ulps
before chopping so that an increment of 1 ulp is enough. In cbrt(),
we essentially do the same, but move the chop point so that the
increment of 1 is not needed.
Fixed comments to match code:
- in cbrt(), the second approximation is good to 25 bits, not quite 26 bits.
- in cbrt(), don't claim that the second approximation may be implemented
in single precision. Single precision cannot handle the full exponent
range without minor but pessimal changes to renormalize, and although
single precision is enough, 25 bit precision is now claimed and used.
Added comments about some of the magic for the error bound 4/6+epsilon.
I still don't understand why it is 4/6+ and not 6/6+ ulps.
Indent comments at the right of code more consistently.
the second step of approximating cbrt(x). It turns out to be neither
very magic not nor very good. It is just the (2,2) Pade approximation
to 1/cbrt(r) at r = 1, arranged in a strange way to use fewer operations
at a cost of replacing 4 multiplications by 1 division, which is an
especially bad tradeoff on machines where some of the multiplications
can be done in parallel. A Remez rational approximation would give
at least 2 more bits of accuracy, but the (2,2) Pade approximation
already gives 6 more bits than needed. (Changed the comment which
essentially says that it gives 3 more bits.)
Lower order Pade approximations are not quite accurate enough for
double precision but are plenty for float precision. A lower order
Remez rational approximation might be enough for double precision too.
However, rational approximations inherently require an extra division,
and polynomial approximations work well for 1/cbrt(r) at r = 1, so I
plan to switch to using the latter. There are some technical
complications that tend to cost a division in another way.
This gives an optimization of between 9 and 22% on Athlons (largest
for cbrt() on amd64 -- from 205 to 159 cycles).
We extracted the sign bit and worked with |x|, and restored the sign
bit as the last step. We avoided branches to a fault by using accesses
to FP values as bits to clear and restore the sign bit. Avoiding
branches is usually good, but the bit access macros are not so good
(especially for setting FP values), and here they always caused pipeline
stalls on Athlons. Even using branches would be faster except on args
that give perfect branch misprediction, since only mispredicted branches
cause stalls, but it possible to avoid touching the sign bit in FP
values at all (except to preserve it in conversions from bits to FP
not related to the sign bit). Do this. The results are identical
except in 2 of the 3 unsupported rounding modes, since all the
approximations use odd rational functions so they work right on strictly
negative values, and the special case of -0 doesn't use an approximation.
<cbrt(x) in bits> ~= <x in bits>/3 + BIAS.
Keep the large comments only in the double version as usual.
Fixed some style bugs (mainly grammar and spelling errors in comments).
It was because I forgot to translate the part of the double precision
algorithm that chops t so that t*t is exact. Now the maximum error
is the same as for double precision (almost exactly 2.0/3 ulps).
The maximum error was 3.56 ulps.
The bug was another translation error. The double precision version
has a comment saying "new cbrt to 23 bits, may be implemented in
precision". This means exactly what it says -- that the 23 bit second
approximation for the double precision cbrt() may be implemented in
single (i.e., float) precision. It doesn't mean what the translation
assumed -- that this approximation, when implemented in float precision,
is good enough for the the final approximation in float precision.
First, float precision needs a 24 bit approximation. The "23 bit"
approximation is actually good to 24 bits on float precision args, but
only if it is evaluated in double precision. Second, the algorithm
requires a cleanup step to ensure its error bound.
In float precision, any reasonable algorithm works for the cleanup
step. Use the same algorithm as for double precision, although this
is much more than enough and is a significant pessimization, and don't
optimize or simplify anything using double precision to implement the
float case, so that the whole double precision algorithm can be verified
in float precision. A maximum error of 0.667 ulps is claimed for cbrt()
and the max for cbrtf() using the same algorithm shouldn't be different,
but the actual max for cbrtf() on amd64 is now 0.9834 ulps. (On i386
-O1 the max is 0.5006 (down from < 0.7) due to extra precision.)
The threshold for not being tiny was too small. Use the usual 2**-12
threshold. As for sinhf, use a different method (now the same as for
sinhf) to set the inexact flag for tiny nonzero x so that the larger
threshold works, although this method is imperfect. As for sinhf,
this change is not just an optimization, since the general code that
we fell into has accuracy problems even for tiny x. On amd64, avoiding
it fixes tanhf on 2*13495596 args with errors of between 1 and 1.3
ulps and thus reduces the total number of args with errors of >= 1 ulp
from 37533748 to 5271278; the maximum error is unchanged at 2.2 ulps.
The magic number 22 is log(DBL_MAX)/2 plus slop. This is bogus for
float precision. Use 9 (log(FLT_MAX)/2 plus less slop than for
double precision). Unlike for coshf and tanhf, this is just an
optimization, and MAX isn't misspelled EPSILON in the commit log.
I started testing with nonstandard rounding modes, and verified that
the chosen thresholds work for all modes modulo problems not related
to thresholds. The best thresholds are not very dependent on the mode,
at least for tanhf.
shares its low half with pio2_hi. pio2_hi is rounded down although
rounding to nearest would be a tiny bit better, so pio4_hi must be
rounded down too. It was rounded to nearest, which happens to be
different in float precision but the same in double precision.
This fixes about 13.5 million errors of more than 1 ulp in asinf().
The largest error was 2.81 ulps on amd64 and 2.57 ulps on i386 -O1.
Now the largest error is 0.93 ulps on amd65 and 0.67 ulps on i386 -O1.
sqrt(2)/2-1. For log1p(), fixed the approximation to sqrt(2)/2-1.
The end result is to fix an error of 1.293 ulps in
log1pf(0.41421395540 (hex 0x3ed413da))
and an error of 1.783 ulps in
log1p(-0.292893409729003961761) (hex 0x12bec4 00000001)).
The former was the only error of > 1 ulp for log1pf() and the latter
is the only such error that I know of for log1p().
The approximations don't need to be very accurate, but the last 2 need
to be related to the first and be rounded up a little (even more than
1 ulp for sqrt(2)/2-1) for the following implementation-detail reason:
when the arg (x) is not between (the approximations to) sqrt(2)/2-1
and sqrt(2)-1, we commit to using a correction term, but we only
actually use it if 1+x is between sqrt(2)/2 and sqrt(2) according to
the first approximation. Thus we must ensure that
!(sqrt(2)/2-1 < x < sqrt(2)-1) implies !(sqrt(2)/2 < x+1 < sqrt(2)),
where all the sqrt(2)'s are really slightly different approximations
to sqrt(2) and some of the "<"'s are really "<="'s. This was not done.
In log1pf(), the last 2 approximations were rounded up by about 6 ulps
more than needed relative to a good approximation to sqrt(2), but the
actual approximation to sqrt(2) was off by 3 ulps. The approximation
to sqrt(2)-1 ended up being 4 ulps too small, so the algoritm was
broken in 4 cases. The result happened to be broken in 1 case. This
is fixed by using a natural approximation to sqrt(2) and derived
approximations for the others.
In logf(), all the approximations made sense, but the approximation
to sqrt(2)/2-1 was 2 ulps too small (a tiny amount, since we compare
with a granularity of 2**32 ulps), so the algorithm was broken in 2
cases. The result was broken in 1 case. This is fixed by rounding
up the approximation to sqrt(2)/2-1 by 2**32 ulps, so 2**32 cases are
now handled a little differently (still correctly according to my
assertion that the approximations don't need to be very accurate, but
this has not been checked).
on assignment.
Extra precision on i386's broke hi+lo decomposition in the usual way.
It caused all except 1 of the 62343 errors of more than 1 ulp for
log1pf() on i386's with gcc -O [-fno-float-store].
according to the highest nonzero bit in a denormal was missing.
fdlibm ilogbf() and ilogb() have always had the adjustment, but only
use a small part of their method for handling denormals; use the
normalization method in log[f]() for the main part.
It was lost in rev.1.9. The log message for rev.1.9 says that the
special case of +-0 is handled twice, but it was only handled once,
so it became unhandled, and this happened to break half of the cases
that return +-0:
- round-towards-minus-infinity: 0 < x < 1: result was -0 not 0
- round-to-nearest: -0.5 <= x < 0: result was 0 not -0
- round-towards-plus-infinity: -1 < x < 0: result was 0 not -0
- round-towards-zero: -1 < x < 0: result was 0 not -0
TWO52[sx] to trick gcc into correctly converting TWO52[sx]+x to double
on assignment to "double w", force a correct assignment by assigning
to *(double *)&w. This is cleaner and avoids the double rounding
problem on machines that evaluate double expressions in double
precision. It is not necessary to convert w-TWO52[sx] to double
precision on return as implied in the comment in rev.1.3, since
the difference is exact.
(1) In round-to-nearest mode, on all machines, fdlibm rint() never
worked for |x| = n+0.75 where n is an even integer between 262144
and 524286 inclusive (2*131072 cases). To avoid double rounding
on some machines, we begin by adjusting x to a value with the 0.25
bit not set, essentially by moving the 0.25 bit to a lower bit
where it works well enough as a guard, but we botched the adjustment
when log2(|x|) == 18 (2*2**52 cases) and ended up just clearing
the 0.25 bit then. Most subcases still worked accidentally since
another lower bit serves as a guard. The case of odd n worked
accidentally because the rounding goes the right way then. However,
for even n, after mangling n+0.75 to 0.5, rounding gives n but the
correct result is n+1.
(2) In round-towards-minus-infinity mode, on all machines, fdlibm rint()
never for x = n+0.25 where n is any integer between -524287 and
-262144 inclusive (262144 cases). In these cases, after mangling
n+0.25 to n, rounding gives n but the correct result is n-1.
(3) In round-towards-plus-infinity mode, on all machines, fdlibm rint()
never for x = n+0.25 where n is any integer between 262144 and
524287 inclusive (262144 cases). In these cases, after mangling
n+0.25 to n, rounding gives n but the correct result is n+1.
A variant of this bug was fixed for the float case in rev.1.9 of s_rintf.c,
but the analysis there is incomplete (it only mentions (1)) and the fix
is buggy.
Example of the problem with double rounding: rint(1.375) on a machine
which evaluates double expressions with just 1 bit of extra precision
and is in round-to-nearest mode. We evaluate the result using
(double)(2**52 + 1.375) - 2**52. Evaluating 2**52 + 1.375 in (53+1) bit
prcision gives 2**52 + 1.5 (first rounding). (Second) rounding of this
to double gives 2**52 + 2.0. Subtracting 2**52 from this gives 2.0 but
we want 1.0. Evaluating 2**52 + 1.375 in double precision would have
given the desired intermediate result of 2**52 + 1.0.
The double rounding problem is relatively rare, so the botched adjustment
can be fixed for most machines by removing the entire adjustment. This
would be a wrong fix (using it is 1 of the bugs in rev.1.9 of s_rintf.c)
since fdlibm is supposed to be generic, but it works in the following cases:
- on all machines that evaluate double expressions in double precision,
provided either long double has the same precision as double (alpha,
and i386's with precision forced to double) or my earlier fix to use
a long double 2**52 is modified to avoid using long double precision.
- on all machines that evaluate double expressions in many more than 11
bits of extra precision. The 1 bit of extra precision in the example
is the worst case. With N bits of extra precision, it sufices to
adjust the bit N bits below the 0.5 bit. For N >= about 52 there is
no such bit so the adjustment is both impossible and unnecessary. The
fix in rev.1.9 of s_rintf.c apparently depends on corresponding magic
in float precision: on all supported machines N is either 0 or >= 24,
so double rounding doesn't occur in practice.
- on all machines that don't use fdlibm rint*() (i386's).
So under FreeBSD, the double rounding problem only affects amd64 now, but
should only affect i386 in future (when double expressions are evaluated
in long double precision).
- in round-towards-minus-infinity mode, on all machines, roundf(x) never
worked for 0 < |x| < 0.5 (2*0x3effffff cases in all, or almost half of
float space). It was -0 for 0 < x < 0.5 and 0 for -0.5 < x < 0, but
should be 0 and -0, respectively. This is because t = ceilf(|x|) = 1
for these args, and when we adjust t from 1 to 0 by subtracting 1, we
get -0 in this rounding mode, but we want and expected to get 0.
- in round-towards-minus-infinity, round towards zero and round-to-nearest
modes, on machines that evaluate float expressions in float precision
(most machines except i386's), roundf(x) never worked for |x| =
<float value immediately below 0.5> (2 cases in all). It was +-1 but
should have been +-0. This is because t = ceilf(|x|) = 1 for these
args, and when we try to classify |x| by subtracting it from 1 we
get an unexpected rounding error -- the result is 0.5 after rounding
to float in all 3 rounding modes, so we we have forgotten the
difference between |x| and 0.5 and end up returning the same value
as for +-0.5.
The fix is to use floorf() instead of ceilf() and to add 1 instead of
-1 in the adjustment. With floorf() all the expressions used are
always evaluated exactly so there are no rounding problems, and with
adjustments of +1 we don't go near -0 when adjusting.
Attempted to fix round() and roundl() by cloning the fix for roundf().
This has only been tested for round(), only on args representable as
floats. Double expressions are evaluated in double precision even on
i386's, so round(0.5-epsilon) was broken even on i386's. roundl()
must be completely broken on i386's since long double precision is not
really supported. There seem to be no other dependencies on the
precision.
k_tanf.c but with different details.
The polynomial is odd with degree 13 for tanf() and odd with degree
9 for sinf(), so the details are not very different for sinf() -- the
term with the x**11 and x**13 coefficients goes awaym and (mysteriously)
it helps to do the evaluation of w = z*z early although moving it later
was a key optimization for tanf(). The details are different but simpler
for cosf() because the polynomial is even and of lower degree.
On Athlons, for uniformly distributed args in [-2pi, 2pi], this gives
an optimization of about 4 cycles (10%) in most cases (13% for sinf()
on AXP, but 0% for cosf() with gcc-3.3 -O1 on AXP). The best case
(sinf() with gcc-3.4 -O1 -fcaller-saves on A64) now takes 33-39 cycles
(was 37-45 cycles). Hardware sinf takes 74-129 cycles. Despite
being fine tuned for Athlons, the optimization is even larger on
some other arches (about 15% on ia64 (pluto2) and 20% on alpha (beast)
with gcc -O2 -fomit-frame-pointer).
cosf(x) is supposed to return something like x when x is a NaN, and
we actually fairly consistently return x-x which is normally very like
x (on i386 and and it is x if x is a quiet NaN and x with the quiet bit
set if x is a signaling NaN. Rev.1.10 broke this by normalising x to
fabsf(x). It's not clear if fabsf(x) is should preserve x if x is a NaN,
but it actually clears the sign bit, and other parts of the code depended
on this.
The bugs can be fixed by saving x before normalizing it, and using the
saved x only for NaNs, and using uint32_t instead of int32_t for ix
so that negative NaNs are not misclassified even if fabsf() doesn't
clear their sign bit, but gcc pessimizes the saving very well, especially
on Athlon XPs (it generates extra loads and stores, and mixes use of
the SSE and i387, and this somehow messes up pipelines). Normalizing
x is not a very good optimization anyway, so stop doing it. (It adds
latency to the FPU pipelines, but in previous versions it helped except
for |x| <= 3pi/4 by simplifying the integer pipelines.) Use the same
organization as in s_sinf.c and s_tanf.c with some branches reordered.
These changes combined recover most of the performance of the unfixed
version on A64 but still lose 10% on AXP with gcc-3.4 -O1 but not with
gcc-3.3 -O1.
that was used doesn't work normally here, since we want to be able to
multiply `hi' by the exponent of x _exactly_, and the exponent of x has
more than 7 significant bits for most denormal x's, so the multiplication
was not always exact despite a cloned comment claiming that it was. (The
comment is correct in the double precision case -- with the normal 33+53
bit decomposition the exponent can have 20 significant bits and the extra
bit for denormals is only the 11th.)
Fixing this had little or no effect for denormals (I think because
more precision is inherently lost for denormals than is lost by roundoff
errors in the multiplication).
The fix is to reduce the precision of the decomposition to 16+24 bits.
Due to 2 bugs in the old deomposition and numerical accidents, reducing
the precision actually increased the precision of hi+lo. The old hi+lo
had about 39 bits instead of at least 41 like it should have had.
There were off-by-1-bit errors in each of hi and lo, apparently due
to mistranslation from the double precision hi and lo. The correct
16 bit hi happens to give about 19 bits of precision, so the correct
hi+lo gives about 43 bits instead of at least 40. The end result is
that expf() is now perfectly rounded (to nearest) except in 52561 cases
instead of except in 67027 cases, and the maximum error is 0.5013 ulps
instead of 0.5023 ulps.
Instead of echoing the code in a comment, try to describe why we split
up the evaluation in a special way.
The new optimization is mostly to move the evaluation of w = z*z later
so that everything else (except z = x*x) doesn't have to wait for w.
On Athlons, FP multiplication has a latency of 4 cycles so this
optimization saves 4 cycles per call provided no new dependencies are
introduced. Tweaking the other terms in to reduce dependencies saves
a couple more cycles in some cases (more on AXP than on A64; up to 8
cycles out of 56 altogether in some cases). The previous version had
a similar optimization for s = z*x. Special optimizations like these
probably have a larger effect than the simple 2-way vectorization
permitted (but not activated by gcc) in the old version, since 2-way
vectorization is not enough and the polynomial's degree is so small
in the float case that non-vectorizable dependencies dominate.
On an AXP, tanf() on uniformly distributed args in [-2pi, 2pi] now
takes 34-55 cycles (was 39-59 cycles).
of between 1.0 and 1.8509 ulps for lgammaf(x) with x between -2**-21 and
-2**-70.
As usual, the cutoff for tiny args was not correctly translated to
float precision. It was 2**-70 but 2**-21 works. Not as usual, having
a too-small threshold was worse than a pessimization. It was just a
pessimization for (positive) args between 2**-70 and 2**-21, but for
the first ~50 million (negative) args below -2**-70, the general code
overflowed and gave a result of infinity instead of correct (finite)
results near 70*log(2). For the remaining ~361 million negative args
above -2**21, the general code gave almost-acceptable errors (lgamma[f]()
is not very accurate in general) but the pessimization was larger than
for misclassified tiny positive args.
Now the max error for lgammaf(x) with |x| < 2**-21 is 0.7885 ulps, and
speed and accuracy are almost the same for positive and negative args
in this range. The maximum error overall is still infinity ulps.
A cutoff of 2**-70 is probably wastefully small for the double precision
case. Smaller cutoffs can be used to reduce the max error to nearly
0.5 ulps for tiny args, but this is useless since the general algrorithm
for nearly-tiny args is not nearly that accurate -- it has a max error of
about 1 ulp.
gives a tiny but hopefully always free optimization in the 2 quadrants
to which it applies. On Athlons, it reduces maximum latency by 4 cycles
in these quadrants but has usually has a smaller effect on total time
(typically ~2 cycles (~5%), but sometimes 8 cycles when the compiler
generates poor code).
of the function name.
Added my (non-)copyright.
In k_tanf.c, added the first set of redundant parentheses to control
grouping which was claimed to be added in the previous commit.
returning float). The functions are renamed from __kernel_{cos,sin}f()
to __kernel_{cos,sin}df() so that misuses of them will cause link errors
and not crashes.
This version is an almost-routine translation with no special optimizations
for accuracy or efficiency. The not-quite-routine part is that in
__kernel_cosf(), regenerating the minimax polynomial with double
precision coefficients gives a coefficient for the x**2 term that is
not quite -0.5, so the literal 0.5 in the code and the related `hz'
variable need to be modified; also, the special code for reducing the
error in 1.0-x**2*0.5 is no longer needed, so it is convenient to
adjust all the logic for the x**2 term a little. Note that without
extra precision, it would be very bad to use a coefficient of other
than -0.5 for the x**2 term -- the old version depends on multiplication
by -0.5 being infinitely precise so as not to need even more special
code for reducing the error in 1-x**2*0.5.
This gives an unimportant increase in accuracy, from ~0.8 to ~0.501
ulps. Almost all of the error is from the final rounding step, since
the choice of the minimax polynomials so that their contribution to the
error is a bit less than 0.5 ulps just happens to give contributions that
are significantly less (~.001 ulps).
An Athlons, for uniformly distributed args in [-2pi, 2pi], this gives
overall speed increases in the 10-20% range, despite giving a speed
decrease of typically 19% (from 31 cycles up to 37) for sinf() on args
in [-pi/4, pi/4].
- Remove dead code that I forgot to remove in the previous commit.
- Calculate the sum of the lower terms of the polynomial (divided by
x**5) in a single expression (sum of odd terms) + (sum of even terms)
with parentheses to control grouping. This is clearer and happens to
give better instruction scheduling for a tiny optimization (an
average of about ~0.5 cycles/call on Athlons).
- Calculate the final sum in a single expression with parentheses to
control grouping too. Change the grouping from
first_term + (second_term + sum_of_lower_terms) to
(first_term + second_term) + sum_of_lower_terms. Normally the first
grouping must be used for accuracy, but extra precision makes any
grouping give a correct result so we can group for efficiency. This
is a larger optimization (average 3-4 cycles/call or 5%).
- Use parentheses to indicate that the C order of left to right evaluation
is what is wanted (for efficiency) in a multiplication too.
The old fdlibm code has several optimizations related to these. 2
involve doing an extra operation that can be done almost in parallel
on some superscalar machines but are pessimizations on sequential
machines. Others involve statement ordering or expression grouping.
All of these except the ordering for the combining the sums of the odd
and even terms seem to be ideal for Athlons, but parallelism is still
limited so all of these optimizations combined together with the ones
in this commit save only ~6-8 cycles (~10%).
On an AXP, tanf() on uniformly distributed args in [-2pi, 2pi] now
takes 39-59 cycles. I don't know of any more optimizations for tanf()
short of writing it all in asm with very MD instruction scheduling.
Hardware fsin takes 122-138 cycles. Most of the optimizations for
tanf() don't work very well for tan[l](). fdlibm tan() now takes
145-365 cycles.
A single polynomial approximation for tan(x) works in infinite precision
up to |x| < pi/2, but in finite precision, to restrict the accumulated
roundoff error to < 1 ulp, |x| must be restricted to less than about
sqrt(0.5/((1.5+1.5)/3)) ~= 0.707. We restricted it a bit more to
give a safety margin including some slop for optimizations. Now that
we use double precision for the calculations, the accumulated roundoff
error is in double-precision ulps so it can easily be made almost 2**29
times smaller than a single-precision ulp. Near x = pi/4 its maximum
is about 0.5+(1.5+1.5)*x**2/3 ~= 1.117 double-precision ulps.
The minimax polynomial needs to be different to work for the larger
interval. I didn't increase its degree the old degree is just large
enough to keep the final error less than 1 ulp and increasing the
degree would be a pessimization. The maximum error is now ~0.80
ulps instead of ~0.53 ulps.
The speedup from this optimization for uniformly distributed args in
[-2pi, 2pi] is 28-43% on athlons, depending on how badly gcc selected
and scheduled the instructions in the old version. The old version
has some int-to-float conversions that are apparently difficult to schedule
well, but gcc-3.3 somehow did everything ~10 cycles or ~10% faster than
gcc-3.4, with the difference especially large on AXPs. On A64s, the
problem seems to be related to documented penalties for moving single
precision data to undead xmm registers. With this version, the speed
is cycles is almost independent of the athlon and gcc version despite
the large differences in instruction selection to use the FPU on AXPs
and SSE on A64s.
This is a minor interface change. The function is renamed from
__kernel_tanf() to __kernel_tandf() so that misues of it will cause
link errors and not crashes.
This version is a routine translation with no special optimizations
for accuracy or efficiency. It gives an unimportant increase in
accuracy, from ~0.9 ulps to 0.5285 ulps. Almost all of the error is
from the minimax polynomial (~0.03 ulps and the final rounding step
(< 0.5 ulps). It gives strange differences in efficiency in the -5
to +10% range, with -O1 fairly consistently becoming faster and -O2
slower on AXP and A64 with gcc-3.3 and gcc-3.4.
arg to __kernel_rem_pio2() gives 53-bit (double) precision, not single
precision and/or the array dimension like I thought. prec == 2 is
used in e_rem_pio2.c for double precision although it is documented
to be for 64-bit (extended) precision, and I just reduced it by 1
thinking that this would give the value suitable for 24-bit (float)
precision. Reducing it 1 more to the documented value for float
precision doesn't actually work (it gives errors of ~0.75 ulps in the
reduced arg, but errors of much less than 0.5 ulps are needed; the bug
seems to be in kernel_rem_pio2.c). Keep using a value 1 larger than
the documented value but supply an array large enough hold the extra
unused result from this.
The bug can also be fixed quickly by increasing init_jk[0] in
k_rem_pio2.c from 2 to 3. This gives behaviour identical to using
prec == 1 except it doesn't create the extra result. It isn't clear
how the precision bug affects higher precisions. 113-bit (quad) is
the largest precision, so there is no way to use a large precision
to fix it.
they can be #included in other .c files to give inline functions, and
use them to inline the functions in most callers (not in e_lgammaf_r.c).
__kernel_tanf() is too large and complicated for gcc to inline very well.
An athlons, this gives a speed increase under favourable pipeline
conditions of about 10% overall (larger for AXP, smaller for A64).
E.g., on AXP, sinf() on uniformly distributed args in [-2Pi, 2Pi]
now takes 30-56 cycles; it used to take 45-61 cycles; hardware fsin
takes 65-129.
On athlons, this gives a speedup of 10-20% for tanf() on uniformly
distributed args in [-2Pi, 2Pi]. (It only directly applies for 43%
of the args and gives a 16-20% speedup for these (more for AXP than
A64) and this gives an overall speedup of 10-12% which is all that it
should; however, it gives an overall speedup of 17-20% with gcc-3.3
on AXP-A64 by mysteriously effected cases where it isn't executed.)
I originally intended to use double precision for all internals of
float trig functions and will probably still do this, but benchmarking
showed that converting to double precision and back is a pessimization
in cases where a simple float precision calculation works, so it may
be optimal to switch precisions only when using extra precision is
much simpler.
__ieee754_rem_pio2f() to its 3 callers and manually inline them.
On Athlons, with favourable compiler flags and optimizations and
favourable pipeline conditions, this gives a speedup of 30-40 cycles
for cosf(), sinf() and tanf() on the range pi/4 < |x| <= 9pi/4, so
thes functions are now signifcantly faster than the hardware trig
functions in many cases. E.g., in a benchmark with uniformly distributed
x in [-2pi, 2pi], A64 hardware fcos took 72-129 cycles and cosf() took
37-55 cycles. Out-of-order execution is needed to get both of these
times. The optimizations in this commit apparently work more by
removing 1 serialization point than by reducing latency.
