The printf() writes the text into stdout, which may be undesirable
in some use cases. Moreover, the printed information was redundant.
So, let's drop such calls.
Since this change, the libosmogapk uses the Osmocom logging
framework. By default, logging is disabled and could be enabled
by the external applications calling the osmo_gapk_log_init()
with a desired log target as an argument.
Since GAPK package contains a library and the representative
osmo-gapk application, the 'main.c' looks a bit confusing. Let's
use the common naming scheme.
The usage of linuxlist is more flexible than having a limited
array of pointers. This approach allows to have as much items
in a processing queue as required.
To simplify the benchrarking process via the library API, this
change introduces two new functions, which are intended to
provide total cycle and frame count.
Having statically allocated memory for benchmark data of every
codec causes high memory usage, especially if actual benchmarking
is not required for a particular use case. Instead of that, let's
provide an optional opportunity to enable benchmarking for a
particular codec by calling the osmo_gapk_bench_enable(). The
required amount of memory would be allocated, and then can be
freed by calling the osmo_gapk_bench_free() or manually.
The usage of a 'static inline' function definition in the 'bench.h'
is resulting in separate independent function definitions in each
translation unit from which the header is included. This is
increasing the size of compiled code unnecessarily.
There is no need to expose the implementation details of both
BENCHMARK_START and BENCHMARK_STOP macros via public header.
This change moves them to a separate private header 'bench.h'.
The benchmark_dump() is only used by the osmo-gapk binary, and
is intended to prepare and print benchmarking results to stderr,
what is most likely unusable for the library users.
To avoid a naming conflict between libosmogapk and other projects
during linkage, all the exposed symbols should have an unique
prefix. Let's use 'osmo_gapk' for that.
The most compilers today do support the '#pragma once', which is
designed to cause the current source file to be included only once
in a single compilation. One has several advantages, including:
less code, avoidance of name clashes, and sometimes improvement
in compilation speed.
See: https://en.wikipedia.org/wiki/Pragma_once for details.
To be able to use the library, external applications need to know,
which symbols are exposed. This information is provided by header
files, which are being installed to a system's ${includedir}
since this change.
The previous GAPK implementation was represented as a single
executable. So, all audio transcoding operations were available
only by calling the 'gapk' binary. This approach didn't allow
external applications to benefit from using GAPK API directly.
Since there are some projects (such as GR-GSM and OsmocomBB),
which are potential users of GAPK code base, it would be better
to have all transcoding functions within a shared library.
So, this change separates the common code into a shared library,
named 'libosmogapk', and links the 'gapk' binary against one.
Currently there are no shared headers, pkg-config manifest and
the export map, but they will be done latter.
On some systems the ALSA output buffer is pretty big, and
if the audio samples are not being passed into the buffer
quickly enough, it becomes starved for data, resulting
in an error called underrun.
Previously, when it happenned, GAPK used to stop processing
with the following message (where X is a random number):
[+] PQ: Adding ALSA output (dev='default', blk_len=320)
[!] pq_execute(): abort, item returned -1
[+] Processed X frames
According to the ALSA documentation, the pcm_handle
changes its state when the problem happens, and should
be recovered using the snd_pcm_prepare() call. This change
actually does that.
This testsuite takes a PCM auidio file and encodes from it every
supported format, and compares that output with a known sample (from an
earlier, trusted version of gakp, shipped as part of this project).
I then re-decodes this file to PCM and also compares that with a shipped
reference re-decode.
While EFR has a canonical format of 31 bytes, the codec_efr.c *does not*
use that canonical format as input. Rather, it uses the format of .amr
files with a 0x3C header as first byte. So the resulting encode/decode
functions should not assume 31 bytes, but 32 bytes.
I noticed that ti-hr format doesn't pass an encode-decode-playback test,
and discussion with tnt resulted in the following conclusion:
19:29 <@tnt> looking at fr and efr, it's always msb_xxx
19:30 <@tnt> and if I ever used it, then most likely it was for decoding
meaning ti_hr_to_canon would have been used and not the
other way around.
The RTP EFR payload is a bit like the FR payload: one nibble magic
marker, then followed by the actual codec bits. So we need to
add/remove that magic marker and shift the remainder by one nibble.
The ETSI reference codec actually uses an array of 20/22 16bit values
rather than a "canonical" format. The conversion is what fmt_hr_ref.c
is doing. However, codec_hr.c must then subsequently not check for the
canonical input/output sizes, but those specific to it.
After merging this change, there is support for the AMR codec (by means
of libopencore-amr, which is already used for EFR).
In terms of gapk formats, we introdude
* the "amr-opencore" format, which serves both as the canonical format,
and as the input format to opencore-amrnb itself.
* the "rtp-amr" format, which is the payload of RFC4867 octet-aligned mode
You can use the following command for a real-time RTP playback for AMR
frames:
./gapk -I 0.0.0.0/30000 -f rtp-amr -A default -g rawpcm-s16le
In fact, it should probably be better to silently ignore all those
errors as opposed to aborting the entire processing queue? But that's
for another patch...
Vadim Yanitskiy
Harald Welte
Sylvain Munaut