Further speedups possible by getting rid of exp2f...
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Faster methods possible; since exponent is always a multiple of 1/8.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Much faster generation possible; but array is small so don't want to bloat
the binary.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
This exploits an approach based on the sieve of Eratosthenes, a popular
method for generating prime numbers.
Tables are identical to previous ones.
Tested with FATE with/without --enable-hardcoded-tables.
Sample benchmark (Haswell, GNU/Linux+gcc):
prev:
7860100 decicycles in cbrt_tableinit, 1 runs, 0 skips
7777490 decicycles in cbrt_tableinit, 2 runs, 0 skips
[...]
7582339 decicycles in cbrt_tableinit, 256 runs, 0 skips
7563556 decicycles in cbrt_tableinit, 512 runs, 0 skips
new:
2099480 decicycles in cbrt_tableinit, 1 runs, 0 skips
2044470 decicycles in cbrt_tableinit, 2 runs, 0 skips
[...]
1796544 decicycles in cbrt_tableinit, 256 runs, 0 skips
1791631 decicycles in cbrt_tableinit, 512 runs, 0 skips
Both small and large run count given as this is called once so small run
count may give a better picture, small numbers are fairly consistent,
and there is a consistent downward trend from small to large runs,
at which point it stabilizes to a new value.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
In the spirit of commit a956840cbc. Simple method to reproduce:
pass -vstats_file /dev/full to ffmpeg.
All raw fclose usages in ffmpeg.c taken care of here.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
sqrt is faster, and is sometimes more accurate depending on the libm.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
avio_closep is not guaranteed to succeed, and its return value can
contain information regarding failure of preceding writes and silent
loss of data (man 2 close, man fclose). Users should know when the
progress was not successfully logged, and so a diagnostic is printed
here.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
ff_clz is faster, and uses an intrinsic (at the moment on GCC). exp2 is
a wasteful function for a simple integer exponentiation.
Untested.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Commit b272c3a5aa has sped up dsd_tablegen, and now table generation takes
~ 40k cycles. Thus, these tables can always be generated at runtime.
Tested with/without --enable-hardcoded-tables.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Tables are bit identical.
Sample benchmark (Haswell, GNU/Linux+gcc):
old:
814485 decicycles in dsd_ctables_tableinit, 512 runs, 0 skips
new:
356808 decicycles in dsd_ctable_tableinit, 512 runs, 0 skips
Binary size should essentially be identical, and is in fact identical on
the configuration I tested on.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
These are just for prefixes and may be hardcoded easily; see lavu/eval
for this approach.
Reviewed-by: Stefano Sabatini <stefasab@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Replaces by real arithmetic. Tested the validity of these transformations separately.
Numerical differences are ~1e-15, and should not matter: it is not even
clear which is more precise mathematically.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Should fix the regression, and also speeds up table generation.
Tables tested on GNU/Linux+clang: they are identical to the ones prior
to 5495c7f. ff_exp10 caused one slight change in one entry, 50000 became
50001 due to somewhat incorrect rounding.
Untested on ICC; passes FATE on GNU/Linux+gcc.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
avoids the float to integer cast, and is slightly superior in terms of
rounding ("Dutch/Gauss rounding").
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Fast, reasonably accurate 10^x. Alternative of detection of libm exp10 at configure
time is not worth the trouble, since it is anyway not POSIX or ISO C,
and currently only the GNU libm has it. Furthermore, GNU libm's variant
is ~ 2x slower, and is ironically not correctly rounded (2 ulp off) to justify all
that slowdown.
Reviewed-by: James Almer <jamrial@gmail.com>
Reviewed-by: Hendrik Leppkes <h.leppkes@gmail.com>
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Addition of comments marking the end of ifdef blocks, correction of an
incorrect (at double precision) M_LN2, removal of an unnecessary undef.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
exp10 is a function available in GNU libm. Looks like no other common
libm has it. This adds support for it to FFmpeg.
There are essentially 2 ways of handling the fallback:
1. Using pow(10, x)
2. Using exp2(M_LOG2_10 * x).
First one represents a Pareto improvement, with no speed or accuracy
regression anywhere, but speed improvement limited to GNU libm.
Second one represents a slight accuracy loss (relative error ~ 1e-13)
for non GNU libm. Speedup of > 2x is obtained on non GNU libm platforms,
~30% on GNU libm. These are "average case numbers", another benefit is
the lack of triggering of the well-known terrible worst case paths
through pow.
Based on reviews, second one chosen. Comment added accordingly.
Reviewed-by: Hendrik Leppkes <h.leppkes@gmail.com>
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
complex is not available on all platforms. Furthermore, it is trivial to
rewrite complex number expressions to real arithmetic, and in fact
sometimes advantageous for performance reasons: by wrapping as a complex,
one forces a particular Cartesian representation that is not necessarily optimal for the purpose.
Configure dependencies also removed, and aemphasis is now available across
all platforms.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Source code is from Boost:
http://www.boost.org/doc/libs/1_46_1/boost/math/special_functions/erf.hpp
with appropriate modifications for FFmpeg.
Tested on interval -6 to 6 (beyond which it saturates), +/-NAN, +/-INFINITY
under -fsanitize=undefined on clang to test for possible undefined behavior.
This function turns out to actually be essentially as accurate and faster than the
libm (GNU/BSD's/Mac OS X), and I can think of 3 reasons why upstream
does not use this:
1. They are not aware of it.
2. They are concerned about licensing - this applies especially to GNU
libm.
3. They do not know and/or appreciate the benefits of rational
approximations over polynomial approximations. Boost uses them to great
effect, see e.g swr/resample for bessel derived from them, which is also
similarly superior to libm variants.
First, performance.
sample benchmark (clang -O3, Haswell, GNU/Linux):
3e8 values evenly spaced from 0 to 6
time (libm):
./test 13.39s user 0.00s system 100% cpu 13.376 total
time (boost based):
./test 9.20s user 0.00s system 100% cpu 9.190 total
Second, accuracy.
1e8 eval pts from 0 to 6
maxdiff (absolute): 2.2204460492503131e-16
occuring at point where libm erf is correctly rounded, this is not.
Illustration of superior rounding of this function:
arg : 0.83999999999999997
erf : 0.76514271145499457
boost : 0.76514271145499446
real : 0.76514271145499446
i.e libm is actually incorrectly rounded. Note that this is clear from:
https://github.com/JuliaLang/openlibm/blob/master/src/s_erf.c (the Sun
implementation used by both BSD and GNU libm's), where only 1 ulp is
guaranteed.
Reasons it is not easy/worthwhile to create a "correctly rounded"
variant of this function (i.e 0.5ulp):
1. Upstream libm's don't do it anyway, so we can't guarantee this unless
we force this implementation on all platforms. This is not easy, as the
linker would complain unless measures are taken.
2. Nothing in FFmpeg cares or can care about such things, due to the
above and FFmpeg's nature.
3. Creating a correctly rounded function will in practice need some use of long
double/fma. long double, although C89/C90, unfortunately has problems on
ppc. This needs fixing of toolchain flags/configure. In any case this
will be slower for miniscule gain.
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Here it is mostly a cosmetic change, but there might be benefits in that
there are no compat hacks for lround, while there are for lrint.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
lrint avoids an implicit cast, and is not slower on non-broken libm's. Thus this
represents a Pareto improvement.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
lrint is faster here on -ftree-vectorize with GCC. This is likely simply
an artifact of GCC's rather terrible auto-vectorizer, since as per the
instruction set manuals cvtsd2si and cvttsd2si (or their vector equivalents)
have identical cycle timings.
Anyway, regardless of above, lrint is superior to round accuracy wise.
Safety guaranteed as long int has at least 32 bits.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
For systems with broken libms.
Tested with NAN, -NAN, INFINITY, -INFINITY, +/-x for regular double x and
combinations of these.
Old versions of MSVC need some UINT64_C hackery.
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Reviewed-by: Hendrik Leppkes <h.leppkes@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
lrint is faster and conveys the intent better here. It is safe as long int has
at least 32 bits.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
exp2 is faster.
It may be possible to optimize further; e.g the exponents seem to be
multiples of 0.25. This requires study though.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
exp2 suffices here. Some trivial speedup is done in addition here by
reusing results.
This retains accuracy, and in particular results in identical values
with GNU libm + gcc/clang.
sample benchmark (Haswell, GNU/Linux):
proposed : 424160 decicycles in pow_table, 512 runs, 0 skips
exp2 only: 1262093 decicycles in pow_table, 512 runs, 0 skips
old : 2849085 decicycles in pow_table, 512 runs, 0 skips
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
arc4random() was designed as a superior interface for system random
number generation, designed for OpenBSD and subsequently incorporated by
other BSD's, Mac OS X, and some non-standard libc's. It is thus an improvement to
use it whenever available.
As a side note, this may or may not get included in glibc, and there is
a proposal to create a posix_random family based on these ideas:
http://austingroupbugs.net/view.php?id=859.
Tested on Mac OS X.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
These postfixes can be computed statically, and there is no need to
waste runtime resources.
Tested with FATE.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
This is a somewhat subtle failure that can occur when the realloc_array
fails in FORMATS_REF.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
pow is a ridiculous function for computing a simple Gaussian.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Recent commits 6aaac24d72 and
3835554bf8 made progress towards cleaning
up usage of the formats API, and in particular fixed possible NULL pointer
dereferences.
This commit addresses the issue of possible resource leaks when some intermediate
call fails.
Tested with valgrind --leak-check=full --show-leak-kinds=all, and manual simulation
of malloc/realloc failures.
Fixes: CID 1250334.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Recent commits 6aaac24d72 and
3835554bf8 made progress towards cleaning
up usage of the formats API, and in particular fixed possible NULL pointer
dereferences.
This commit addresses the issue of possible resource leaks when some intermediate
call fails.
Tested with valgrind --leak-check=full --show-leak-kinds=all, and manual simulation
of malloc/realloc failures.
Fixes: CID 1338330.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Recent commits 6aaac24d72 and
3835554bf8 made progress towards cleaning
up usage of the formats API, and in particular fixed possible NULL pointer
dereferences.
This commit addresses the issue of possible resource leaks when some intermediate
call fails.
Tested with valgrind --leak-check=full --show-leak-kinds=all, and manual simulation
of malloc/realloc failures.
Fixes: CID 1338326, 1338329.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Recent commits 6aaac24d72 and
3835554bf8 made progress towards cleaning
up usage of the formats API, and in particular fixed possible NULL pointer
dereferences.
This commit addresses the issue of possible resource leaks when some intermediate
call fails.
Tested with valgrind --leak-check=full --show-leak-kinds=all, and manual simulation
of malloc/realloc failures.
Fixes: CID 1338327.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Recent commits 6aaac24d72 and
3835554bf8 made progress towards cleaning
up usage of the formats API, and in particular fixed possible NULL pointer
dereferences.
This commit addresses the issue of possible resource leaks when some intermediate
call fails. Unfortunately, even leaving aside this subtle intermediate
failure aspect, commit 8087632027 was only
partially successful in addressing memleaks. Hopefully, this commit
fixes the issue completely.
Tested with valgrind --leak-check=full --show-leak-kinds=all, and manual simulation
of malloc/realloc failures.
Fixes: CID 1270818.
Reviewed-by: Clément Bœsch <u@pkh.me>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
This gets rid of virtually useless hardcoded tables hackery. The reason
it is useless is that a 320 element lut is anyway placed regardless of
--enable-hardcoded-tables, from which all necessary tables are trivially
derived at runtime at very low cost:
sample benchmark (x86-64, Haswell, GNU/Linux, single run is really
what is relevant here since looping drastically changes the bench). Fluctuations
are on the order of 10% for the single run test:
39400 decicycles in aacsbr_tableinit, 1 runs, 0 skips
25325 decicycles in aacsbr_tableinit, 2 runs, 0 skips
18475 decicycles in aacsbr_tableinit, 4 runs, 0 skips
15008 decicycles in aacsbr_tableinit, 8 runs, 0 skips
13016 decicycles in aacsbr_tableinit, 16 runs, 0 skips
12005 decicycles in aacsbr_tableinit, 32 runs, 0 skips
11546 decicycles in aacsbr_tableinit, 64 runs, 0 skips
11506 decicycles in aacsbr_tableinit, 128 runs, 0 skips
11500 decicycles in aacsbr_tableinit, 256 runs, 0 skips
11183 decicycles in aacsbr_tableinit, 509 runs, 3 skips
Tested with FATE with/without --enable-hardcoded-tables.
Reviewed-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
pow is a very wasteful function for this purpose. A low hanging fruit
would be simply to replace with exp2f, and that does yield some speedup.
However, there are 2 drawbacks of this:
1. It does not exploit the integer nature of the argument.
2. (minor) Some platforms lack a proper exp2f routine, making benefits available
only to non broken libm.
3. exp2f does not solve the same issue that plagues pow, namely terrible
worst case performance. This is a fundamental issue known as the
"table-maker's dilemma" recognized by Prof. Kahan himself and
subsequently elaborated and researched by many others. All this is clear from benchmarks below.
This exploits the IEEE-754 format to get very good performance even in
the worst case for integer powers of 2. This solves all the issues noted
above. Function tested with clang usan over [-1000, 1000] (beyond range of
relevance for this, which is [-255, 255]), patch itself with FATE.
Benchmarks obtained on x86-64, Haswell, GNU-Linux via 10^5 iterations of
the pow call, START/STOP, and command ffplay ~/samples/jpeg2000/chiens_dcinema2K.mxf.
Low number of runs also given to prove the point about worst case:
pow:
216270 decicycles in pow, 1 runs, 0 skips
110175 decicycles in pow, 2 runs, 0 skips
56085 decicycles in pow, 4 runs, 0 skips
29013 decicycles in pow, 8 runs, 0 skips
15472 decicycles in pow, 16 runs, 0 skips
8689 decicycles in pow, 32 runs, 0 skips
5295 decicycles in pow, 64 runs, 0 skips
3599 decicycles in pow, 128 runs, 0 skips
2748 decicycles in pow, 256 runs, 0 skips
2304 decicycles in pow, 511 runs, 1 skips
2072 decicycles in pow, 1022 runs, 2 skips
1963 decicycles in pow, 2044 runs, 4 skips
1894 decicycles in pow, 4091 runs, 5 skips
1860 decicycles in pow, 8184 runs, 8 skips
exp2f:
134140 decicycles in pow, 1 runs, 0 skips
68110 decicycles in pow, 2 runs, 0 skips
34530 decicycles in pow, 4 runs, 0 skips
17677 decicycles in pow, 8 runs, 0 skips
9175 decicycles in pow, 16 runs, 0 skips
4931 decicycles in pow, 32 runs, 0 skips
2808 decicycles in pow, 64 runs, 0 skips
1747 decicycles in pow, 128 runs, 0 skips
1208 decicycles in pow, 256 runs, 0 skips
952 decicycles in pow, 512 runs, 0 skips
822 decicycles in pow, 1024 runs, 0 skips
765 decicycles in pow, 2047 runs, 1 skips
722 decicycles in pow, 4094 runs, 2 skips
693 decicycles in pow, 8190 runs, 2 skips
exp2fi:
2740 decicycles in pow, 1 runs, 0 skips
1530 decicycles in pow, 2 runs, 0 skips
955 decicycles in pow, 4 runs, 0 skips
622 decicycles in pow, 8 runs, 0 skips
477 decicycles in pow, 16 runs, 0 skips
368 decicycles in pow, 32 runs, 0 skips
317 decicycles in pow, 64 runs, 0 skips
291 decicycles in pow, 128 runs, 0 skips
277 decicycles in pow, 256 runs, 0 skips
268 decicycles in pow, 512 runs, 0 skips
265 decicycles in pow, 1024 runs, 0 skips
263 decicycles in pow, 2048 runs, 0 skips
263 decicycles in pow, 4095 runs, 1 skips
260 decicycles in pow, 8191 runs, 1 skips
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
It is required to call va_end for each invocation of va_start within the
same function.
Fixes: CID 1341583.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Ganesh AjjanagaddeandVittorio Giovara
