Linfeng Zhang authored 8 years ago and Jean-Marc Valin committed 7 years ago

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>

Linfeng Zhang authored 8 years ago and Jean-Marc Valin committed 7 years ago

The optimization is bit exact with C function.

This optimization speeds up fixed-point SILK encoder on NEON about
5% to 8%. (Tested on Acer Chromebook, ARMv7 Processor rev 3 (v7l).)

Change-Id: I582f6f3585b7946149e16a2ad3084ebc0ae79a4f

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

- celt/modes.c:430:14: warning: cast from 'const unsigned char *' to
  'opus_int16 *' increases required alignment from 1 to 2 [-Wcast-align]
- 'C[0][1]' may be used uninitialized [-Wmaybe-uninitialized]
- Unused variable/parameter
- Value stored is never read
- MSVC warnings about "possible loss of data" due to type conversions
- MSVC warning C4146: unary minus operator applied to unsigned type
- silk/NLSF_del_dec_quant.c:137:20: warning: array subscript is above
  array bounds [-Warray-bounds] (gcc -O3 false positive)
- src/mlp_train.h:39:20: warning: function declaration isn't a prototype
  [-Wstrict-prototypes]
- Remove SMALL_FOOTPRINT code from SSE 4.1 FIR implementation, matching
  the C implementation.

The clang -Wcast-align warnings with SSE intrinsics are a known
clang issue: https://llvm.org/bugs/show_bug.cgi?id=20670

Linfeng Zhang authored 8 years ago and Jean-Marc Valin committed 8 years ago

The optimization is bit exact with C function.

Change-Id: Ib3bdc26a5a4ebe02e7f24be85104e8e9a2a9a738

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

It checks that no clearly unstable filter passes the LPC_inv_pred_gain()
test. Also, this will make it possible to check assembly for correctness.

Modified from an original patch from Linfeng Zhang <linfengz@google.com>.

James Zern authored 8 years ago and Jean-Marc Valin committed 8 years ago

and rtcd disabled (CFLAGS=-mfpu=neon)

broken since:
cfdaf365 Optimize silk_NSQ_del_dec() for ARM NEON

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

Linfeng Zhang authored 8 years ago and Jean-Marc Valin committed 8 years ago

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>

Casting to unsigned to avoid shifting negative values left.

Linfeng Zhang authored 8 years ago and Felicia Lim committed 8 years ago

NSQ_LPC_BUF_LENGTH is independent of DECISION_DELAY.

Signed-off-by: Felicia Lim <flim@google.com>

Use of celt_fir remains off by default

Linfeng Zhang authored 8 years ago and Felicia Lim committed 8 years ago

The "mem" in celt_fir_c() either is contained in the head of input "x"
in reverse order already, or can be easily attached to the head of "x"
before calling the function. Removing argument "mem" can eliminate the
redundant buffer copies inside.
Update celt_fir_sse4_1() accordingly.

Ensure that the SILK encoder's log gain is 63, not 64, when encoding
a maximum-value delta gain index of 40.  This matches the decoder
and RFC 6716 4.2.7.4, and prevents an assertion failure in the rare
case that the gain is later independently coded.

This was just enabling commented-out code, setting a fixed
default removes dead code.

Also remove the SILK_DEBUG_STORE_CLOSE_FILES flush call from
opus_demo.

This is debugging code which is no longer used, but defining
the symbols for SILK_DEBUG_STORE_CLOSE_FILES and calling it
from opus_demo causes linking problems on Microsoft Visual Studio
where we have strict controls on public symbols and want to
test the compiled DLL.

Since the code isn't in active use, it's better to remove it
to avoid clutter and address the linking issue.

Reported by Mark Warner.

We now declare that anything that would overflow is not stable enough

Speeds up encoding by another ~1-2%

(found through fuzzing)

Felicia Lim authored 8 years ago and Jean-Marc Valin committed 8 years ago

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

For large values of maxabs_Q20, silk_MUL( maxabs_Q20, ind + 1 ) could
overflow.

Reduces the XX values to avoid an overflow in silk_VQ_WMat_EC(), even in
floating-point.

no dither