The optimization is bit exact with C function.
Change-Id: Ifb8f04b19f2d576e79ce5dcfa7e0fc374d71d6c8
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
Split to silk_biquad_alt_stride1() and silk_biquad_alt_stride2(),
so that it can be optimized more efficiently when stride is 2.
This change in C code is bit exact with the origin.
Change-Id: Idaefe670397016ace2a489e3435ac61b7dbe79d5
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
The optimization is bit exact with C function.
Change-Id: Ib3bdc26a5a4ebe02e7f24be85104e8e9a2a9a738
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
The optimization is bit exact with C function.
This optimization speeds up SILK encoder on NEON as following.
Fixed-point:
Complexity 0-5: 0%
Complexity 6-7: 6%
Complexity 8-9: 10%
Complexity 10: 8%
Got similar results on floating-point.
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
1. Only for fixed point on x86 platform (32bit and 64bit, uses SIMD
intrinsics up to SSE4.2)
2. Use "configure --enable-fixed-point --enable-intrinsics" to enable
optimization, default is disabled.
3. Official test cases are verified and passed.
Signed-off-by: Timothy B. Terriberry <tterribe@xiph.org>
Newer versions of MSVC are unhappy with the strategy of the build
environment redefining "inline" (even though they don't support the
actual keyword). Instead we define OPUS_INLINE to the right thing
in opus_defines.h.
This is the same approach we use for restrict.
Original patch by Aurélien Zanelli <aurelien.zanelli@parrot.com>:
http://lists.xiph.org/pipermail/opus/2013-May/002078.html
Revised version:
- Add autconf detection (ported from libtheora).
- Rename ARM5E to ARMv5E (an ARM5 is not the same thing as ARMv5!).
- Use actual macros so they can still be selectively overridden.
- Split out ARMv4 parts and add a few more ARMv4 macros.
- Label blocks to make them easy to find in generated assembly.
- Fix MULT16_32_Q15() so we can pass make check.
The MDCT test passes in values larger than 2**30 for b.
The new version should be just as fast (or faster, since it's
easier to merge the shift with following instructions), and
there's no appreciable impact on accuracy (FFT/MDCT SNR actually
goes up in most cases).
- Fix register constraints.
We were using early-clobber flags in a bunch of places that
didn't need them, and commutative-pair flags in a bunch of
places that weren't actually commutative.
This was Jean-Marc's fault (the original code came from Speex).
- Simplify silk_CLZ16().
- Port over iFFT C_MULC asm by Andree Buschmann
<AndreeBuschmann@t-online.de> from Rockbox.
- Speed up the C_MULC asm by using LDRD, allowing more flexible
addressing, re-ordering instructions to avoid some stalls,
allowing more flexible register allocation, and getting things
out of the inline asm block so the compiler can schedule them
better.
- Add C_MUL and C_MUL4 asm for the FFT to the encoder based, on the
new C_MULC.
In total, this patch gives a 22.3% speed-up on test_opus_encoder on
a 600 MHz Cortex A8 using gcc 4.2.1,
When restricted to ARMv4 optimizations, it gives a 9.6% speed-up
on the same processor/compiler.
On the conformance test vectors:
Average mono quality is 97.0583 %
Average stereo quality is 97.775 %
On MacOS, stdlib.h ends up including sys/signal.h, generating
warnings about the local variables called 'signal' shadowing
the global symbol signal(3).
This was originally done in 86476906 but it missed some use
of 'signal' in prototypes in headers where it didn't cause
warnings. Later the prototypes were moved around and the
warnings came back.
This also cleans up some cases in where stdlib.h was used
but shouldn't be required.
decoder:
- fixed incorrect scaling of filter states for the smallest quantization
step sizes
- NLSF2A now limits the prediction gain of LPC filters
encoder:
- increased damping of LTP coefficients in LTP analysis
- increased white noise fraction in noise shaping LPC analysis
- introduced maximum total prediction gain. Used by Burg's method to
exit early if prediction gain is exceeded. This improves packet
loss robustness and numerical robustness in Burg's method
- Prefiltered signal is now in int32 Q10 domain, from int16 Q0
- Increased max number of iterations in CBR gain control loop from 5 to 6
- Removed useless code from LTP scaling control
- Optimization: smarter LPC loop unrolling
- Switched default win32 compile mode to be floating-point
resampler:
- made resampler have constant delay of 0.75 ms; removed delay
compensation from silk code.
- removed obsolete table entries (~850 Bytes)
- increased downsampling filter order from 16 to 18/24/36 (depending on
frequency ratio)
- reoptimized filter coefficients
C reserves identifiers of the from _[A-Z]+ and we have a number of
those in the code. This patch renames the various function arguments,
MACROS and preprocessor symbols to avoid the reserved form.
It also removes the CHANNELS() macro altogether. This was a
minor optimization for TI DSP to force a mono-only build,
as were the associated local 'const' versions. Since stereo
support is manditory, it wasn't worth keeping.
Thanks to John Ridges for raising the issue, and Jean-Marc Valin
and Greg Maxwell for reviewing the changes.
This is achieved by running the encoding process in a loop and
padding when we don't reach the exact rate. It also implements
VBR-with-cap, which means we no longer need to artificially decrease
the SILK bandwidth when it's close to the cap.
