This applies to every library where performance is not critical.

modelled after aarch64 code

on Cortex-A8, s16 and s32 code is about 2x faster,
float code about 7x faster

Signed-off-by: Peter Meerwald <pmeerw@pmeerw.net>
Signed-off-by: Martin Storsjö <martin@martin.st>

Signed-off-by: Peter Meerwald <pmeerw@pmeerw.net>
Signed-off-by: Martin Storsjö <martin@martin.st>

CC: libav-stable@libav.org
Bug-Id: CID 732225

CC: libav-stable@libav.org
Bug-Id: CID 1231986

Leftover from fbc0b865.

Signed-off-by: Luca Barbato <lu_zero@gentoo.org>

This treats mono as planar internally within libavresample rather
than changing the sample format.

Signed-off-by: Luca Barbato <lu_zero@gentoo.org>

Utility function to get the upper bound on the number of samples the
resampler would output.

Instead of setting FFLIBS in each library Makefile configure
exports FFLIBS-$library in config.mak.

Optimized for the default filter length 16.

30% faster opus silk decoding.

Required for arch optimized resampling.

This fixes cases where very few input samples (fewer than needed for one
output sample) are passed to lavr at the beginning.
CC:libav-stable@libav.org

This simplifies the code, since we do not have to deal with a possibly
negative source index anymore.

negative sample_index is handled in the block above.

The correct "next" input sample is not the first sample of the
resampling buffer, but the center sample of the filter_length-sized
block at the beginning.

CC:libav-stable@libav.org

Dead code elimination is enough to avoid undefined references in these cases.

When mixing 3 front channels into 2, the center channel is mixed into left and right with the center mix level, not -3dB.

Don't rely on the fact that an unprefixed label currently exists.

Signed-off-by: Martin Storsjö <martin@martin.st>

Signed-off-by: Martin Storsjö <martin@martin.st>

Specifically, when the corresponding input channel exists and its matrix
column is all-zero (which is necessary for zeroing the output), the
matrix column must be removed from the matrix.

This is not done currently, so the mixing code would end up using
uninitialized pointers from stack.

Found-by: Mateusz "j00ru" Jurczyk and Gynvael Coldwind

This should make it easier to catch problems where some of those
pointers are used uninitialized, since reading from NULL should always
crash, while random numbers from stack can turn out to be valid
pointers, so random memory may be silently overwritten.

This is a more sensible place for these macros.

Fixes a segfault during resampling when compiled with -DDEBUG.
Fixes all fate-lavr-resample tests with -DDEBUG.

CC:libav-stable@libav.org

This is so we can sync to x264's version of FMA4 support.

This partialy reverts commit 79687079.

Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>

With the forthcoming VEX instruction emulation, mulps
must have only the third operand point to memory, as
this is what vmulps expects.

Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>