Gregory Maxwell
a65db56e54
Fix declaration after statement in fixed point.
2014-01-08 11:48:38 -08:00
Jean-Marc Valin
e775090427
pseudostack instrumentation (off by default)
2014-01-07 21:32:41 -05:00
Jean-Marc Valin
9134e96cb2
Fixes SMALL_FOOTPRINT for float
2014-01-07 17:50:46 -05:00
Jean-Marc Valin
e17ca25617
Don't allocate pulses on the stack when calling the SILK PLC.
...
Also minor C89 fix for the previous commit
2014-01-07 15:27:02 -05:00
Jean-Marc Valin
b63e7110cb
Moves CELT PLC pitch search to a separate function to reduce peak stack
2014-01-07 15:02:43 -05:00
Jean-Marc Valin
5f807c176f
Adds SMALL_FOOTPRINT hack to the CELT PLC too
2014-01-07 04:48:42 -05:00
Jean-Marc Valin
9d1b6fef2a
Moves deemphasis() call out of celt_decode_lost() to reduce peak stack
2014-01-07 04:32:41 -05:00
Jean-Marc Valin
ad8371d172
Cleaning up leftovers of "freq" in celt_decode_with_ec()
2014-01-06 17:45:57 -05:00
Jean-Marc Valin
4d07b1357e
Reduces the decoder stack use by removing the pcm_silk buffer in fixed-point
...
We only keep when concealing less than 10ms with SILK.
2014-01-06 17:43:20 -05:00
Jean-Marc Valin
14ca4ed682
Moves the remains of compute_inv_mdcts() to celt_synthesis()
2014-01-06 09:31:09 -05:00
Jean-Marc Valin
32454dcadc
Hack that makes the SMALL_FOOTPRINT CELT decoder use only 4.25 kB of stack.
2014-01-06 09:11:52 -05:00
Jean-Marc Valin
bdc7b93358
Reduces decoder stack usage by only storing one channel of denormalized MDCT
2014-01-06 08:58:38 -05:00
Jean-Marc Valin
4a6744a446
Some cleaning up of the synthesis code.
2014-01-05 21:40:02 -05:00
Jean-Marc Valin
ed01a596dc
Making exp_rotation1() use MAC16_16(), which saves a few cycles on ARM
2014-01-04 21:06:24 -05:00
Jean-Marc Valin
ccec752a38
Silences unused parameter warning
2014-01-04 00:49:46 -05:00
Jean-Marc Valin
ef0eac497f
Moving the radix-2 to expose trivial twiddle factors
2014-01-03 23:55:52 -05:00
Jean-Marc Valin
c8f62a4aed
Improving the accuracy of the fixed-point radix-3 and radix-5
2013-12-31 00:00:37 -05:00
Jean-Marc Valin
e1f846208e
Minor cleanup -- nothing to see here
2013-12-29 18:45:49 -05:00
Jean-Marc Valin
05291fd6a6
Fixed-point: slight accuracy improvement in the comb filter
2013-12-29 18:31:17 -05:00
Jean-Marc Valin
30f52cbe2d
Remove a SAVE_STACK that was pasted accidentally in the previous commit
2013-12-29 16:21:06 -05:00
Jean-Marc Valin
e1dc1e2238
Unifying scaling of fixed-point and float FFT
2013-12-29 13:34:17 -05:00
Jean-Marc Valin
dbb96ab5cc
Fixes C89 issue
2013-12-29 00:09:06 -05:00
Jean-Marc Valin
4c1a90a847
Getting rid of some negations
...
Since we're doing two rotations, we can invert the sign on both.
Also adding a few comments for optimizing the FFT.
2013-12-28 23:14:26 -05:00
Jean-Marc Valin
cc344fb8ff
Slightly improving the accuracy of the fixed-point MDCT downscale
...
Also simplifying the code
2013-12-28 19:10:44 -05:00
Jean-Marc Valin
e0c00e27d8
Commit 99968ab was causing us to allocate too much stack in the MDCT
2013-12-27 03:16:34 -05:00
Jean-Marc Valin
e43a0abe0a
Removes the separate 1/8N rotation in the (I)MDCT and unmerges the MDCT sizes
...
Undoes commits f7547a4e72513f3c
2013-12-27 00:10:54 -05:00
Jean-Marc Valin
a5e3c8a6a6
Inverse MDCT no longer requires any scratch space
2013-12-23 02:26:03 -05:00
Jean-Marc Valin
e2bcb3fe9b
Reverse the ordering of the FFT stage to optimize a degenerate radix-4 case.
...
This also happens to increase the accuracy since it appears that the new
ordering is optimal (at least for 20 ms frames), whereas the previous ordering
was pessimal.
2013-12-22 02:17:24 -05:00
Jean-Marc Valin
c8f4e1608a
Merges the FFT scaling with the MDCT pre-rotate
2013-12-21 16:30:49 -05:00
Jean-Marc Valin
153def2884
Getting rid of the inverse FFT entirely
...
IMDCT now uses the forward FFT.
2013-12-21 15:45:17 -05:00
Jean-Marc Valin
99968abba8
Moving bitrev step to forward MDCT too
2013-12-21 14:29:41 -05:00
Jean-Marc Valin
bc13bbaad7
Applying the forward FFT gain up-front for fixed-point too
...
This makes us lose a bit of precision in the first steps, but our gain is more
precise because it's only rounded once. Overall, SNR is slightly improved.
2013-12-21 02:33:22 -05:00
Jean-Marc Valin
2e26b82ec2
Moves the bitrev step to the IMDCT pre-rotation
2013-12-20 23:13:29 -05:00
Jean-Marc Valin
306d7f5a30
fixed-point: slight (but free) accuracy improvement in compute_band_energies()
...
Also moves the VSHR32() condition outside the loop just in case some compilers
don't optimize that properly.
2013-12-16 01:08:21 -05:00
Jean-Marc Valin
e0f26246b0
fixed-point: adds rounding to some shifts to eliminate bias
...
This reduces the peak decoding error by removing small (inaudible) spikes in
the error at the frame boundaries. These were due to the frequency-domain bias
ending up as a small pulse in the middle of the IMDCT overlap. None of this
was ever audible, but fixing it is still cleaner.
2013-12-14 11:07:13 -05:00
Jean-Marc Valin
4a168eb343
Remove useless code in alloc_trim_analysis()
2013-12-11 01:34:06 -05:00
Jean-Marc Valin
5752d659fd
Minor fixed-point accuracy improvements that were completely free
2013-12-11 00:21:38 -05:00
Jean-Marc Valin
91f8010108
Removing indirections
2013-12-10 22:09:33 -05:00
Jean-Marc Valin
5607d5d1c8
Annotating pointer arguments with OPUS_RESTRICT and const
2013-12-10 22:09:29 -05:00
Jean-Marc Valin
122971b8cc
More NaN hardening in the analysis code
2013-12-10 13:56:38 -05:00
Jean-Marc Valin
d5553e8aca
Using OPUS_COPY()/OPUS_CLEAR() in the decoder too
2013-12-10 02:32:26 -05:00
Jean-Marc Valin
15edb78b3e
Making NaN detection more robust to -ffast-math.
2013-12-09 21:56:21 -05:00
Jean-Marc Valin
4fda6b0142
Using celt_inner_prod() in compute_band_energies()
2013-12-09 18:06:34 -05:00
Pedro Becerra
a9b7def9f5
s/MAX16/MAX32/ in transient_analysis()
...
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
2013-12-09 16:08:29 -05:00
Jean-Marc Valin
57cd849cf7
Defining celt_inner_prod() and using it instead of explicit loops.
...
Also adds an SSE-optimized celt_inner_prod().
2013-12-09 15:26:58 -05:00
Jean-Marc Valin
ff072009fe
Replaces inline copies and initialization with OPUS_*() macros.
...
This is a bit faster at -O2 because memcpy()/memmove()/memset() are
vectorized. The code is also cleaner.
2013-12-09 15:26:52 -05:00
Jean-Marc Valin
0f869cba0f
Changes ABS16() and ABS32() to use fabs() in the float build
...
gcc is better at optimizing it than the ?: version
2013-12-09 15:26:43 -05:00
Jean-Marc Valin
c94e4bb103
Optimizes encoder NaN detection and clipping by only running them when needed
...
NaN detection should now be able to catch values that would create NaNs
further down.
2013-12-09 15:26:03 -05:00
Jean-Marc Valin
5626908ec3
Fixed-point fast-path for normal 48 kHz encoding in celt_preemphasis()
2013-12-05 16:40:59 -05:00
Jean-Marc Valin
aed1009df9
Turns a 16x32 multiply into a 16x16 one in celt_preemphasis().
2013-12-05 13:36:48 -05:00
