800k parameters, 600 MFLOPS, with a receptive field of 3 feature vectors

Saves ~270 kB of weights in the decoder

Only code non-degenerate symbols, which makes the encoder faster

Removing one of the 2d conv layers for pitch estimation reduces
complexity without noticeable degradation. FARGAN model has more
adversarial training.
Also, no need for the double precision in the low-pass filter.

Using a max lambda of 0.04

Remove old pitch estimator and retrain all models

Michael Klingbeil authored 1 year ago and Jean-Marc Valin committed 1 year ago

Add unit tests for opus_packet_extensions_count, opus_packet_extensions_parse,
and opus_packet_extensions_generate. Add various assertions for NULL buffers or
negative length arguments. Add explicit check for invalid length in
opus_packet_extensions_count. Check for extension id >127 in
opus_packet_extensions_generate. Check for invalid length input in
opus_packet_extensions_generate.

Signed-off-by: Jean-Marc Valin <jmvalin@amazon.com>

autoreconf looks for ACLOCAL_*FLAGS in Makefile.am and will
pass that on when it invokes aclocal. Furthermore, setting
a local like this doesn't make it available to subcommands,
so the line has no effect.

Needed by commit 972a34ec.

Use autoreconf in autogen.sh instead of the handwritten version,
it's simpler, and also updates things that we weren't handling.

Drop the hand-written INSTALL file.  Its information content was
~zero, and autotools wants to overwrite it with its own version,
so don't fight that, just .gitignore it.

Original patch by Aurélien Zanelli <aurelien.zanelli@parrot.com>:
 http://lists.xiph.org/pipermail/opus/2013-May/002078.html

Revised version:
- Add autconf detection (ported from libtheora).
- Rename ARM5E to ARMv5E (an ARM5 is not the same thing as ARMv5!).
- Use actual macros so they can still be selectively overridden.
- Split out ARMv4 parts and add a few more ARMv4 macros.
- Label blocks to make them easy to find in generated assembly.
- Fix MULT16_32_Q15() so we can pass make check.
  The MDCT test passes in values larger than 2**30 for b.
  The new version should be just as fast (or faster, since it's
   easier to merge the shift with following instructions), and
   there's no appreciable impact on accuracy (FFT/MDCT SNR actually
   goes up in most cases).
- Fix register constraints.
  We were using early-clobber flags in a bunch of places that
   didn't need them, and commutative-pair flags in a bunch of
   places that weren't actually commutative.
  This was Jean-Marc's fault (the original code came from Speex).
- Simplify silk_CLZ16().
- Port over iFFT C_MULC asm by Andree Buschmann
   <AndreeBuschmann@t-online.de> from Rockbox.
- Speed up the C_MULC asm by using LDRD, allowing more flexible
   addressing, re-ordering instructions to avoid some stalls,
   allowing more flexible register allocation, and getting things
   out of the inline asm block so the compiler can schedule them
   better.
- Add C_MUL and C_MUL4 asm for the FFT to the encoder based, on the
   new C_MULC.

In total, this patch gives a 22.3% speed-up on test_opus_encoder on
 a 600 MHz Cortex A8 using gcc 4.2.1,
When restricted to ARMv4 optimizations, it gives a 9.6% speed-up
 on the same processor/compiler.
On the conformance test vectors:
 Average mono quality is 97.0583 %
 Average stereo quality is 97.775 %

Previously the autotools build litered whatever directory
'make' was invoked from with intermediate files. Adding
'subdir-objects' to AUTOMAKE_OPTIONS tells it to put them
in subdirectories according to the relative paths to the
source files, .i.e. silk/schur.o for $(srcdir)/silk/schur.c

The change to autogen.sh is necessary to handle the
(apparently new) case where an AUTOMAKE_OPTIONS line is
given without a minimum version number.