codec makefiles larger freedom in what they can do to it.
Use this in libopus to prepend the libopus searchpaths to
CODECFLAGS so that its internal config.h will be picked up before
our global one. This avoids having to do a s/config.h/opus_config.h/
when syncing which will be handy soon.
Change-Id: I018d729aa0c8300fa3149f22a5a8c5668b339dfa
Reviewed-on: http://gerrit.rockbox.org/496
Reviewed-by: Nils Wallménius <nils@rockbox.org>
Distractions make logic fail. It only needs one more loop and should
not trigger further compression cycles after not feeding more data.
Change-Id: Ie0dbb34af92e0ca5718480dd4ab4719a141717ff
For mp3_enc, split encoding duties between COP and CPU.
For wavpack_enc, simply run the encoding on COP (splitting that one
needs more consideration) which keeps the it and the UI from running
on the same core.
As a result, at least they are now useable on PP at "normal" sample
rates.
mp3_enc in all this gets an extensive renovation and some optimizations
for speed, to reduce IRAM requirements and remove unneeded stuff.
Change-Id: I215578dbe36f14e516b05a5ca70880eb01ca0ec2
Basically, just give it a good rewrite.
Software codec recording can be implemented in a more straightforward
and simple manner and made more robust through the better codec
control now available.
Encoded audio buffer uses a packed format instead of fixed-size
chunks and uses smaller data headers leading to more efficient usage.
The greatest benefit is with a VBR format like wavpack which needs
to request a maximum size but only actually ends up committing part
of that request.
No guard buffers are used for either PCM or encoded audio. PCM is
read into the codec's provided buffer and mono conversion done at
that time in the core if required. Any highly-specialized sample
conversion is still done within the codec itself, such as 32-bit
(wavpack) or interleaved mono (mp3).
There is no longer a separate filename array. All metadata goes
onto the main encoded audio buffer, eliminating any predermined
file limit on the buffer as well as not wasting the space for
unused path queue slots.
The core and codec interface is less awkward and a bit more sensible.
Some less useful interface features were removed. Threads are kept
on narrow code paths ie. the audio thread never calls encoding
functions and the codec thread never calls file functions as before.
Codecs no longer call file functions directly. Writes are buffered
in the core and data written to storage in larger chunks to speed up
flushing of data. In fact, codecs are no longer aware of the stream
being a file at all and have no access to the fd.
SPDIF frequency detection no longer requires a restart of recording
or plugging the source before entering the screen. It will poll
for changes and update when stopped or prerecording (which does
discard now-invalid prerecorded data).
I've seen to it that writing a proper header on full disk works
when the format makes it reasonably practical to do so. Other cases
may have incorrect data sizes but sample info will be in tact. File
left that way may play anyway.
mp3_enc.codec acquires the ability to write 'Info' headers with LAME
tags to make it gapless (bonus).
Change-Id: I670685166d5eb32ef58ef317f50b8af766ceb653
Reviewed-on: http://gerrit.rockbox.org/493
Reviewed-by: Michael Sevakis <jethead71@rockbox.org>
Tested-by: Michael Sevakis <jethead71@rockbox.org>
Most SoCs are these days are fast enough for realtime BRR, gaussian
interpolation and echo processing.
Change-Id: I180ce8ad45242c67b5e573a406b9522098a3f12b
Affected BRR cached waveforms but not realtime BRR decode as far as
I could ascertain. BRR cached waves required loop points to be inside
the initial waveform but this change removes that restriction.
Change-Id: I0ef4db720e5c28bd7b2fb9ae255d27c0a7213f79
CPU optimization gets its own files in which to fill-in optimizable
routines.
Some pointless #if 0's for profiling need removal. Those macros are
empty if not profiling.
Force some functions that are undesirable to be force-inlined by the
compiler to be not inlined.
Change-Id: Ia7b7e45380d7efb20c9b1a4d52e05db3ef6bbaab
Use the tlsf malloc and friends instead of the silly
codec_malloc to get actually working free and saner
realloc that doesn't leak memory.
Makes files with moderately sized embedded AA play
on targets with large enough codec buffers and files
with too large AA are now skipped rather than crashing.
Fixes crash when playing example file in FS#12842.
Change-Id: I06562955c4d9a95bd90f55738214fba462092b71
DSP_SWITCH_FREQUENCY has been deprecated and the same enumerated value
as DSP_SET_FREQUENCY since major DSP revisions were committed. This
task should have been performed much earlier but, oh well, do it now.
Change-Id: I3f30d651b894136a07c7e17f78fc16a7d98631ff
Will need it soon enough.
Combine the contents of all the various fixedpoint.h files.
Not moving fixedpoint.c for now since I'm not sure where it
should be and it causes some dependency issues.
Change-Id: Ideacbca2ca78f9158c2b114b113c274f68e908d5
Prevents cutoff of tracks, especially short ones:
* Extend looped tracks by fade length to fade at start of loop repeat.
* No fade occurs for non-repeating track only having an intro.
* Uses id3.tail_trim field to store fade duration.
Use libGME built-in elapsed time reporting instead of custom calculation:
* libGME already reports in milliseconds.
* Don't advance time counter when Repeat == One. It just runs the progress
over the length limit.
Fix a comment about sample rate and set the reported bitrate to be
accurate for 44.1 kHz stereo.
Change-Id: I3ede22bda0f9a941a3fef751f4d678eb0027344c
Flush decoder state and frame out buffer upon a forced stop to prevent
a short burst of stale audio from the previously decoding track from
playing when skipping from one WMA track to another.
Change-Id: I24c910c5dbd83caed2510db68d9e39a474332a79
Reviewed-on: http://gerrit.rockbox.org/406
Reviewed-by: Michael Sevakis <jethead71@rockbox.org>
Tested-by: Michael Sevakis <jethead71@rockbox.org>
Speeds up decoding of 128k opus files by 1.2MHz on AMSv2. Rounding
error is 1 bit due to KissFFT using a 15 bit shift instead of a 16 bit shift.
Also, change an LDMIA in the armv4 code to LDM as the pointer should not
increment.
Change-Id: I626a207c6a056a1984e33cfe89415c35d0caed93
Reviewed-on: http://gerrit.rockbox.org/377
Reviewed-by: Michael Giacomelli <giac2000@hotmail.com>
Tested-by: Michael Giacomelli <giac2000@hotmail.com>
I'm not 100% sure that the rounding of denormals is correct. As compared to foobar2000,
some samples are off by +1 LSB. However, since I can't output 24 bit PCM easily with
rockbox, I'm not sure if this is due to a bug or just how rockbox rounds. In practice
I don't think it matters so I'm just going to commit this for now.
Change-Id: Ic0792fcb172e4369a5512d202121c2b918b36079
avoids complicated index calculations in the loops.
saves 0.3MHz decoding a 64kbps test file on h300 (cf) and
0.2MHz on c200 (pp)
Change-Id: I1918912d9a4502f89980c6bb270ec2ef10a07010
Signed-off-by: Nils Wallménius <nils@rockbox.org>
on a target with a disk.
Change-Id: I37c875c9cd014eb61fe5232dab0f4b8f15f057dd
Reviewed-on: http://gerrit.rockbox.org/319
Tested-by: Thiago Okada <thiago.mast3r@gmail.com>
Reviewed-by: Frederik Vestre <freqmod@gmail.com>
Tested-by: Frederik Vestre <freqmod@gmail.com>
speeds up decoding of a 64kbps test_file by 1.5MHz on c200 (pp)
and 1.9MHz on fuzev1 (amsv1)
Change-Id: I1db460b634eba608c3e00541d96fc93d5a05710b
Signed-off-by: Nils Wallménius <nils@rockbox.org>
speeds up decoding of a 64kbps test file by 0.5Hz on h300 (cf)
0.9MHz on c200 (pp) and 0.2MHz on fuzev1 (amsv1)
Change-Id: Ib537c2393fa6dca0b61e4e9f80eef5e688c2c2bd
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Unroll overlap add loop by four and use memcpy for copying
instead of loops.
Change-Id: I17114626a395d5972130251d892f851bc86e3a6a
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Replace complicated macro doing three 16*16 muls and add an inline
asm implementation for arm, speeds up decoding a 64kbps test file
by 0.5MHz on c200 (pp) and gives slightly better precision.
Change-Id: I6fc5b83c210f01bffdc38aec54cc5a8b646d8169
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Hoist load of coefficients out of the loop.
Speeds up decoding of a 64kbps test file by 0.6MHz on h300 (cf)
0.2MHz on c200 (pp) and 0.1MHz on fuzev1 (amsv1)
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Change-Id: I4be0059fc2a77748575f5fc9378f7f348d64f1c4
Skip expensive multiply-accumulate loop when gains are 0 and
just copy using memcpy if soure and destination are not the same
Speeds up decoding of a 64kbps test file by 6MHz on h300 (cf)
7MHz on c200 (pp) and 6MHz on fuzev1 (amsv1)
Change-Id: Ibbc9ddfd45a9ac661467b1327b8c67761924fb8b
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Reorder operands to take advantage of the early termination of
multiplications. Saves 2.5MHz decoding a 64kbps opus test file
on c200 (pp).
Change-Id: I470266dc870ab183ece3b23426d41e2a64342a71
Speeds up decoding of 64kbps test file by 6.3MHz on h300 (cf)
and 1.2MHz on c200 (pp).
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Change-Id: I08c2c332153abcbef9447c81986777fd2fcc73fe
speeds up decoding of 64kbps test file by 19MHz on h300 (cf)
and 2.5MHz on c200 (pp)
Change-Id: Idacd2f8962c20c518055d586daeec6b932b7ded2
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Speeds up decoding of a 64kbps test file 26MHz on H300 (cf) and
2MHz on c200 (pp)
Change-Id: I2fb4fe6c0a29321087e02fbd17fd1b1eb84e7b57
Signed-off-by: Nils Wallménius <nils@rockbox.org>
Remove downsampling code from deemphasis loop as we don't use
it and remove multiplications that are not relevant when
not using custom modes. Saves 1.4MHz on h300 (cf), 4.3MHz on
c200 (pp) and 4.6 on fuzev1 (amsv1).
Change-Id: Iab3f1d737a656a563aaa351d50db987a9cff2287
Saves about 30MHz on h300 (cf) and 1.5MHz on c200 (pp) decoding a
64kbps test file. Stack usage is still below 70%.
Change-Id: Ib13df9011adb4eef4bb91a52e5a32741c8bf8988
