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>
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>
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>
This should be useful for the sofalizer filter.
Reviewed-by: Kieran Kunhya <kierank@ob-encoder.com>
Reviewed-by: Clément Bœsch <u@pkh.me>
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>
Fix a dead lock under certain conditions. Let's assume we have a queue of 1
message max, 2 senders, and 1 receiver.
Scenario (real record obtained with debug added):
[...]
SENDER #0: acquired lock
SENDER #0: queue is full, wait
SENDER #1: acquired lock
SENDER #1: queue is full, wait
RECEIVER: acquired lock
RECEIVER: reading a msg from the queue
RECEIVER: signal the cond
RECEIVER: acquired lock
RECEIVER: queue is empty, wait
SENDER #0: writing a msg the queue
SENDER #0: signal the cond
SENDER #0: acquired lock
SENDER #0: queue is full, wait
SENDER #1: queue is full, wait
Translated:
- initially the queue contains 1/1 message with 2 senders blocking on
it, waiting to push another message.
- Meanwhile the receiver is obtaining the lock, read the message,
signal & release the lock. For some reason it is able to acquire the
lock again before the signal wakes up one of the sender. Since it
just emptied the queue, the reader waits for the queue to fill up
again.
- The signal finally reaches one of the sender, which writes a message
and then signal the condition. Unfortunately, instead of waking up
the reader, it actually wakes up the other worker (signal = notify
the condition just for 1 waiter), who can't push another message in
the queue because it's full.
- Meanwhile, the receiver is still waiting. Deadlock.
This scenario can be triggered with for example:
tests/api/api-threadmessage-test 1 2 100 100 1 1000 1000
One working solution is to make av_thread_message_queue_{send,recv}()
call pthread_cond_broadcast() instead of pthread_cond_signal() so both
senders and receivers are unlocked when work is done (be it reading or
writing).
This second solution replaces the condition with two: one to notify the
senders, and one to notify the receivers. This prevents senders from
notifying other senders instead of a reader, and the other way around.
It also avoid broadcasting to everyone like the first solution, and is,
as a result in theory more optimized.
Pretty standard macros, these should help libav*
users avoid repeating ver.si.on parsing code,
which aids in compatibility-checking tasks like
identifying FFmpeg from Libav (_MICRO >= 100 check).
Something many are doing since we are not
intercompatible anymore.
Signed-off-by: Reynaldo H. Verdejo Pinochet <reynaldo@osg.samsung.com>
The fps variable is explicitly set to -1 in case of some errors, the check must
thus be signed or the code setting it needs to use 0 as error code
the type of the field could be changed as well but its in an installed header
Fixes: integer overflow
Fixes: 9982cc157b1ea90429435640a989122f/asan_generic_3ad004a_3799_22cf198d9cd09928e2d9ad250474fa58.mov
Found-by: Mateusz "j00ru" Jurczyk and Gynvael Coldwind
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
There was no reason AFAIK for making AV_CRC_24_IEEE 12. This simply
resulted in wasted space under --enable-hardcoded-tables:
dynamic: 1318672 libavutil/libavutil.so.55
old : 1330680 libavutil/libavutil.so.55
new : 1326488 libavutil/libavutil.so.55
Minor version number is bumped, with ifdefry due to API breakage.
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
This is useful for build-time table generation (--enable-hardcoded-tables),
by providing compat shims for hosts that have broken libms.
This file is deliberately kept minimal; functions can always be added on
an as-needed basis.
Reviewed-by: Clément Bœsch <u@pkh.me>
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
The code expects actual positive numbers and gives completely wrong
results if INT64_MIN is treated as positive
Instead clip it into the valid range that is add 1 and treat it as
negative
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This is a trivial rewrite of the loops that results in better
prefetching and associated cache efficiency. Essentially, the problem is
that modern prefetching logic is based on finite state Markov memory, a reasonable
assumption that is used elsewhere in CPU's in for instance branch
predictors.
Surrounding loops all iterate forward through the array, making the
predictor think of prefetching in the forward direction, but the
intermediate loop is unnecessarily in the backward direction.
Speedup is nontrivial. Benchmarks obtained by 10^6 iterations within
solve_lls, with START/STOP_TIMER. File is tests/data/fate/flac-16-lpc-cholesky.err.
Hardware: x86-64, Haswell, GNU/Linux.
new:
17291 decicycles in solve_lls, 2096706 runs, 446 skips
17255 decicycles in solve_lls, 4193657 runs, 647 skips
17231 decicycles in solve_lls, 8384997 runs, 3611 skips
17189 decicycles in solve_lls,16771010 runs, 6206 skips
17132 decicycles in solve_lls,33544757 runs, 9675 skips
17092 decicycles in solve_lls,67092404 runs, 16460 skips
17058 decicycles in solve_lls,134188213 runs, 29515 skips
old:
18009 decicycles in solve_lls, 2096665 runs, 487 skips
17805 decicycles in solve_lls, 4193320 runs, 984 skips
17779 decicycles in solve_lls, 8386855 runs, 1753 skips
18289 decicycles in solve_lls,16774280 runs, 2936 skips
18158 decicycles in solve_lls,33548104 runs, 6328 skips
18420 decicycles in solve_lls,67091793 runs, 17071 skips
18310 decicycles in solve_lls,134187219 runs, 30509 skips
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
Commit 14ea4151d7 had a bug in that the
conversion of the uint64_t result to an int (the return signature) would
lead to implementation defined behavior, and in this case simply
returned 0 for NAN. A fix via AND'ing the result with 1 does the trick,
simply by ensuring a 0 or 1 return value.
Patch tested with FATE on x86-64, GNU/Linux by forcing the compatibility
code via an ifdef hack suggested by Michael.
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
* commit '588b6215b4c74945994eb9636b0699028c069ed2':
rtmpcrypt: Do the xtea decryption in little endian mode
xtea: Add functions for little endian mode
Conflicts:
libavutil/xtea.c
Merged-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
This improves the mathematical behavior of hypotenuse computation.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
It is known that the naive sqrt(x*x + y*y) approach for computing the
hypotenuse suffers from overflow and accuracy issues, see e.g
http://www.johndcook.com/blog/2010/06/02/whats-so-hard-about-finding-a-hypotenuse/.
This adds hypot support to FFmpeg, a C99 function.
On platforms without hypot, this patch does a reaonable workaround, that
although not as accurate as GNU libm, is readable and does not suffer
from the overflow issue. Improvements can be made separately.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
isnan and isinf are actually macros as per the standard. In particular,
the existing implementation has incorrect signature. Furthermore, this
results in undefined behavior for e.g double values outside float range
as per the standard.
This patch corrects the undefined behavior for all usage within FFmpeg.
Note that long double is not handled as it is not used in FFmpeg.
Furthermore, even if at some point long double gets used, it is likely
not needed to modify the macro in practice for usage in FFmpeg. See
below for analysis.
Getting long double to work strictly per the spec is significantly harder
since a long double may be an IEEE 128 bit quad (very rare), 80 bit
extended precision value (on GCC/Clang), or simply double (on recent Microsoft).
On the other hand, any potential future usage of long double is likely
for precision (when a platform offers extra precision) and not for range, since
the range anyway varies and is not as portable as IEEE 754 single/double
precision. In such cases, the implicit cast to a double is well defined
and isinf and isnan should work as intended.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
* commit '1fc94724f1fd52944bb5ae571475c621da4b77a0':
xtea: Clarify that the current API works in big endian mode
Merged-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
The function is renamed to ff_rint64_clip()
This should avoid build failures on VS2012
Feel free to changes this to a different solution
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The rationale for this function is reflected in the documentation for
it, and is copied here:
Clip a double value into the long long amin-amax range.
This function is needed because conversion of floating point to integers when
it does not fit in the integer's representation does not necessarily saturate
correctly (usually converted to a cvttsd2si on x86) which saturates numbers
> INT64_MAX to INT64_MIN. The standard marks such conversions as undefined
behavior, allowing this sort of mathematically bogus conversions. This provides
a safe alternative that is slower obviously but assures safety and better
mathematical behavior.
API:
@param a value to clip
@param amin minimum value of the clip range
@param amax maximum value of the clip range
@return clipped value
Note that a priori if one can guarantee from the calling side that the
double is in range, it is safe to simply do an explicit/implicit cast,
and that will be far faster. However, otherwise this function should be
used.
avutil minor version is bumped.
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>
This adds msvc optimisations as well as fixing an error in icl whereby it will generate invalid code otherwise.
Signed-off-by: Matt Oliver <protogonoi@gmail.com>
ICC versions older than atleast 12.1.6 dont have the tzcnt intrinsics.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Matt Oliver <protogonoi@gmail.com>
Fixes compilation of host tool aacps_fixed_tablegen.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: James Almer <jamrial@gmail.com>
