GitLab

Not everyone who includes speexdsp_config_types.h will have a test
which defines those, and if we've chosen to use the stdint types at
configure time then we know exactly which header(s) are available, so
just choose the best one then and generate the header to use it.

This patch, including the above text, is copied from a commit in the
speex repository[1]. The original commit for speex was made by Ron
<ron@debian.org>.

[1] https://git.xiph.org/?p=speex.git;a=commitdiff;h=774c87d6cb7dd8dabdd17677fc6da753ecf4aa87Signed-off-by: Tanu Kaskinen <tanu.kaskinen@linux.intel.com>

Reported-by: Fabian Henze <flyser42@gmx.de>

_USE_SSE2 was only being defined for win32 builds on x86-64

Signed-off-by: Tristan Matthews <le.businessman@gmail.com>

Signed-off-by: Tristan Matthews <le.businessman@gmail.com>

Signed-off-by: Tristan Matthews <le.businessman@gmail.com>

Patch by chutz.
https://trac.xiph.org/ticket/1825

Patch adapted from https://trac.xiph.org/ticket/1298Signed-off-by: Tristan Matthews <le.businessman@gmail.com>

Signed-off-by: Tristan Matthews <le.businessman@gmail.com>

This is a little more consistent than spx_int32_t with other lengths.
Signed-off-by: Tristan Matthews <le.businessman@gmail.com>

The only difference between speex_alloc(size) and speex_realloc(NULL, size)
is that speex_alloc zero-initializes the memory, but every byte of memory
allocated is subsequently set anyway.

This also means that st->sinc_table_length is initialized even on the first
allocation of sinc_table.
Signed-off-by: Tristan Matthews <le.businessman@gmail.com>

Also adds inline asm implementations of WORD2INT(x) macro for fixed
and floating point.

Semantics of inner_product_single have also been changed to contain
the final right shift and saturation so it can also be implemented in
the optimal way for the used platform. This change affects fixed point
calculations only.

I also added a new fixed point macro SATURATE32PSHR(x, shift, a). It
does pretty much the same thing as SATURATE32(PSHR32(x, shift), a),
but it avoids over flowing in rounding up phase in the rare occasion
where x has already been saturated. It should also be slightly faster.

Use --enable-neon to force NEON optimization on. The auto detection
should also work if your CFLAGS supports NEON.

On many systems taking couple of kilo-bytes for full sinc table is not
a problem. With a straight filter table the number of MAC operations
needed drops to one quarter. Also the inner loop of the filter becomes
more simple and easier to optimize. So the full sinc table version
runs much faster in most cases. The trade off is some kilo-bytes
bigger memory usage and more CPU needed when initializing the
resampler.

configure.ac: Add --enable-resample-full-sinc-table conf flag

Use this flag to enable resampling with full sinc table.

Since the math.h one is what most systems will use anyway.

The $pkgincludedir for this package is now $includedir/speexdsp, which
isn't where established users are looking for these headers. Thanks to
the vlc folks for noticing this.

sample_frac_num is the array most likely to be non-zero after resampling a
previous stream, but the other arrays can also be left at non-zero values in
other situations.

Using a previous sample_frac_num can mean that the first input sample is
ignored.  This may usually not be significant, but may just be detectable when
only one channel of a stereo signal is affected.

Reviewed by Jean-Marc Valin <jmvalin@mozilla.com> at
https://bugzilla.mozilla.org/show_bug.cgi?id=973374Signed-off-by: Tristan Matthews <le.businessman@gmail.com>