Convert to 16 bits only at the very end

Compensates for the spectral leakage and the fact that we don't have
an explicit masking curve.

Detecting tones can help us prevent the encoder from getting confused
by them.

Git will now remember it for us forever.

Signed-off-by: Jean-Marc Valin <jeanmarcv@google.com>

Even if it is followed by repeated short extensions with payloads.
We track the total size of the short extension payloads that need
 to be repeated, and remove that from the available space for the
 last long extension.
This means we can no longer use L=0 on a repeat to skip coding a
 frame separator when the extensions to be repeated contain a long
 extension followed by one or more short extensions with payloads
 (and there are no more non-repeated extensions in the current
 frame, but there are extensions in the next frame), but this case
 seems uncommon, and hard to explain.
The savings from always being able to skip coding a length when the
 final extensions are repeated extensions with at least one long
 extension is likely higher.
We can still skip a frame separator if we repeat only short
 extensions.

Also update existing tests and add a test for the case where we do
 not have enough space for the repeated short extensions after the
 last long extension.

Signed-off-by: Jean-Marc Valin <jeanmarcv@google.com>

This avoids the need to code a frame separator in the case that the
 last repeated extension is a short extension, and the next
 non-repeated extension is in the following frame instead of the
 current one (saving one additional byte).
Also add tests for both encoding and decoding this.

Signed-off-by: Jean-Marc Valin <jeanmarcv@google.com>

This gives more confidence that the generation code always produces
 output that the parsing code can parse.

Signed-off-by: Jean-Marc Valin <jeanmarcv@google.com>

Also adds tests which exercise generating repeated extensions as
 well as the count_ext() and parse_ext() API for parsing extensions
 in frame order.

Signed-off-by: Jean-Marc Valin <jeanmarcv@google.com>

Right now, opus_packet_extensions_parse() returns the extensions in
 the order they appear in the packet, which is no longer necessarily
 in frame order.
This adds a new (still private) API that returns parsed extensions
 in frame order, even when repeated extensions are used.
Nothing has been converted to use this new API yet.

Signed-off-by: Jean-Marc Valin <jeanmarcv@google.com>

We were originally saturating the comb filter with enough margin
that the de-emphasis couldn't overflow at low frequency
(1/(1-.85) gain). The problem is that didn't leave enough headroom
for a full-scale HF signal.

EXTEND32() would warn that the input isn't 16-bit

Created test_opus_custom script which does a more heavy evaluation of the various use cases of OpusCustom, testing:
- Mixed float/fixed use cases
- Mixed Opus/OpusCustom use cases
- Wide mixture of run-time configurables
- RMS difference (if RESYNTH) is defined

Signed-off-by: Jean-Marc Valin <jeanmarcv@google.com>

Was due to nbAvailableBytes not matching the real CBR budget.
Also adding same tell+16<=total_bits condition that the decoder
has for the comb filter just to be on the safe side.
This bug doesn't appear to be possible with the regular Opus API.

Otherwise we would shrink the buffer before initializing it.

Seems to have been in 682b6cf1. There's nothing wrong with either the
encoder or the decoder. Just about how we resynthesize anti-collapse
for mono in the encoder.

Makes it possible to just encode or just decode using the same format as
opus_demo

update to osce evaluation script

Address the same issue in the mips code.

IOhannes m zmölnig authored 2 years ago and Ralph Giles committed 6 months ago

The doxygen-generated manpages contain some typos...

Signed-off-by: Ralph Giles <giles@thaumas.net>

Rather than repeating the code to iterate through extensions in
 three different places, each with slight differences, different
 edge cases, different error handling, etc., create an iterator
 that can be used everywhere.

Nam Se Hyun authored 6 months ago and Tristan Matthews committed 6 months ago

Signed-off-by: Tristan Matthews <tmatth@videolan.org>

Test multiple buffer lengths and ensure we do not write past the
 end of the provided buffer.

The function takes an opus_int32 *, but we were passing the address
 of an int, which might not be the same.
This is only likely to affect platforms with a 16-bit int.

Without this check, a DRED extension encoded with an invalid frame
 number would still be used, potentially with a surprisingly large
 dred_frame_offset.

With an 8+ MB packet it is possible to craft an extension length
 that would overflow, bypassing the checks to ensure the extension
 data remains inside the packet.
This patch fixes that and adds a test for it.