Introduction
Previously, we identified several packages that do not support SVE2 codes yet. We ultimately decided that Opus Audio Codec is the best candidate. In this post, we will explore the package in detail and see how we can implement SVE2 into it.
Before We Start...
When we clone the package, we can see the following files:
$ ls
AUTHORS cmake LICENSE_PLEASE_READ.txt meson.build opus_sources.mk silk update_version
autogen.sh CMakeLists.txt m4 meson_options.txt opus-uninstalled.pc.in silk_headers.mk win32
celt configure.ac Makefile.am NEWS README silk_sources.mk
celt_headers.mk COPYING Makefile.mips opus_headers.mk README.draft src
celt_sources.mk doc Makefile.unix opus.m4 releases.sha2 tests
ChangeLog include meson opus.pc.in scripts training
As we can see, the packages contains several Makefile and configure template files. These gives us an idea that this package may use the GNU Autotools to generate Makefile and configure scripts. To have a clear understanding of how we can install this package, we can read the README file.
# README
# ...
1) Clone the repository:
% git clone https://gitlab.xiph.org/xiph/opus.git
% cd opus
2) Compiling the source
% ./autogen.sh
% ./configure
% make
3) Install the codec libraries (optional)
% sudo make install
Once you have compiled the codec, there will be a opus_demo executable
in the top directory.
Usage: opus_demo [-e] <application> <sampling rate (Hz)> <channels (1/2)>
<bits per second> [options] <input> <output>
opus_demo -d <sampling rate (Hz)> <channels (1/2)> [options]
<input> <output>
# ...
Now, let's follow the instruction. Once we run the ./autogen.sh, we get the following list of files.
$ ./autogen.sh
$ ls
aclocal.m4 CMakeLists.txt doc Makefile.mips opus.pc.in silk_headers.mk
AUTHORS compile include Makefile.unix opus_sources.mk silk_sources.mk
autogen.sh config.guess INSTALL meson opus-uninstalled.pc.in src
autom4te.cache config.h.in install-sh meson.build package_version test-driver
celt config.sub LICENSE_PLEASE_READ.txt meson_options.txt README tests
celt_headers.mk configure ltmain.sh missing README.draft training
celt_sources.mk configure.ac m4 NEWS releases.sha2 update_version
ChangeLog COPYING Makefile.am opus_headers.mk scripts win32
cmake depcomp Makefile.in opus.m4 silk
We have more files now. The notable files are Makefile.in and configure script file. Makefile.in is generated from Makefile.am file but still is missing some values that are going to be filled with the configure script. Now, let's run the configure script.
$ ./configure
$ ls
aclocal.m4 config.guess include Makefile.unix opus-uninstalled.pc stamp-h1
AUTHORS config.h INSTALL meson opus-uninstalled.pc.in test-driver
autogen.sh config.h.in install-sh meson.build package_version tests
autom4te.cache config.log libtool meson_options.txt README training
celt config.status LICENSE_PLEASE_READ.txt missing README.draft update_version
celt_headers.mk config.sub ltmain.sh NEWS releases.sha2 win32
celt_sources.mk configure m4 opus_headers.mk scripts
ChangeLog configure.ac Makefile opus.m4 silk
cmake COPYING Makefile.am opus.pc silk_headers.mk
CMakeLists.txt depcomp Makefile.in opus.pc.in silk_sources.mk
compile doc Makefile.mips opus_sources.mk src
Not surprisingly, we get a Makefile amongst the newly generated files. If we inspect the Makefile, we can see what CFLAG it uses to compile the package.
# Makefile
# ...
CFLAGS = -g -O2 -fvisibility=hidden -D_FORTIFY_SOURCE=2 -W -Wall -Wextra -Wcast-align -Wnested-externs -Wshadow -Wstrict-prototypes
# ...
From these flags, we can know that the package does not utilize auto-vectorization. This means we can also implement SVE2 codes by utilizing auto-vectorization in this package.
Now, let's compile the package. For this, we can assign more jobs in parallel at a time when we execute the Makefile to increase the speed of compilation. In general, we can calculate the number by doubling the core number plus one. In our case, we can use a value of 24 (16 cores * 2 + 1) to keep every core busy with jobs.
$ make -j 24
$ ls
aclocal.m4 config.guess include Makefile.mips opus.pc silk
AUTHORS config.h INSTALL Makefile.unix opus.pc.in silk_headers.mk
autogen.sh config.h.in install-sh meson opus_sources.mk silk_sources.mk
autom4te.cache config.log libopus.la meson.build opus-uninstalled.pc src
celt config.status libtool meson_options.txt opus-uninstalled.pc.in stamp-h1
celt_headers.mk config.sub LICENSE_PLEASE_READ.txt missing package_version test-driver
celt_sources.mk configure ltmain.sh NEWS README tests
ChangeLog configure.ac m4 opus_compare README.draft training
cmake COPYING Makefile opus_demo releases.sha2 trivial_example
CMakeLists.txt depcomp Makefile.am opus_headers.mk repacketizer_demo update_version
compile doc Makefile.in opus.m4 scripts win32
As the README file mentioned, we have a executable file called opus_demo. When we run it, we can see the package is successfully compiled.
$ ./opus_demo
Usage: /home/swji1/opus/.libs/opus_demo [-e] <application> <sampling rate (Hz)> <channels (1/2)> <bits per second> [options] <input> <output>
/home/swji1/opus/.libs/opus_demo -d <sampling rate (Hz)> <channels (1/2)> [options] <input> <output>
application: voip | audio | restricted-lowdelay
options:
-e : only runs the encoder (output the bit-stream)
-d : only runs the decoder (reads the bit-stream as input)
-cbr : enable constant bitrate; default: variable bitrate
-cvbr : enable constrained variable bitrate; default: unconstrained
-delayed-decision : use look-ahead for speech/music detection (experts only); default: disabled
-bandwidth <NB|MB|WB|SWB|FB> : audio bandwidth (from narrowband to fullband); default: sampling rate
-framesize <2.5|5|10|20|40|60|80|100|120> : frame size in ms; default: 20
-max_payload <bytes> : maximum payload size in bytes, default: 1024
-complexity <comp> : complexity, 0 (lowest) ... 10 (highest); default: 10
-inbandfec : enable SILK inband FEC
-forcemono : force mono encoding, even for stereo input
-dtx : enable SILK DTX
-loss <perc> : simulate packet loss, in percent (0-100); default: 0
Testing
But, how we validate if the binary works as intended? For this, we can refer to the README again.
# README
# ...
== Testing ==
This package includes a collection of automated unit and system tests
which SHOULD be run after compiling the package especially the first
time it is run on a new platform.
To run the integrated tests:
% make check
# ...
Thankfully, the authors provide a set of unit tests to validate the integrity of the executable file. Using this, we can check if the package is compiled correctly.
$ make check
PASS: celt/tests/test_unit_cwrs32
PASS: celt/tests/test_unit_dft
PASS: celt/tests/test_unit_entropy
PASS: celt/tests/test_unit_laplace
PASS: celt/tests/test_unit_mathops
PASS: celt/tests/test_unit_mdct
PASS: celt/tests/test_unit_rotation
PASS: celt/tests/test_unit_types
PASS: silk/tests/test_unit_LPC_inv_pred_gain
PASS: tests/test_opus_api
PASS: tests/test_opus_decode
PASS: tests/test_opus_encode
PASS: tests/test_opus_padding
PASS: tests/test_opus_projection
============================================================================
Testsuite summary for opus 1.3.1-107-gccaaffa9
============================================================================
# TOTAL: 14
# PASS: 14
# SKIP: 0
# XFAIL: 0
# FAIL: 0
# XPASS: 0
# ERROR: 0
============================================================================
Conclusion
In this post, we explored the package and learned how to compile it to generate binary files to execute. We also confirmed that the package does not utilize the auto-vectorization. So, we may try implementing the vectorization in two ways: compiler intrinsics or auto-vectorization. In the next post, we will see how we can add SVE2 codes by using intrinsics.
Top comments (0)