1) Enable asm/intrinsics even for floating-point
2) Make sure ARMv8 asimd enables EDSP/MEDIA/Neon
3) Add dotp architecture to rtcd table since AArch *can* have dotp

Used for DNN matrix multiplies

for cmake, force PRESEUME_SSE4_1 on PRESUME_AVX2

Reducing dependency chains

Should never occur on amd64, but it could on 32-bit x86

No RTCD yet

Starting with compute_linear()

Avoids switching to CNG unless we just have a "refresh packet"

The sinc filter offset was incorrectly handled. Since it perfectly compensates
for the analysis offset, nothing has to be done.
Also, the preemphasis memory was never initialized.

Controls whether deep PLC is enabled

No dependency on the data files if no DNN code enabled

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

Those intrinsics don't actually require alignment so we're OK

Jean-Marc Valin authored 1 year ago and Jean-Marc Valin committed 1 year ago

This only affects custom modes (builds with --enable-custom-modes) with frame
sizes 2.5, 5, 10, and 20ms and sampling rates below 40 kHz. The problem does
not affect normal use of Opus (using OpusEncoder/OpusDecoder) even when built
with custom modes enabled, but only special applications that use
OpusCustomEncoder/OpusCustomDecoder.

If configuring CELT for CBR but controlling the bitrate with
OPUS_SET_BITRATE rather than nbCompressedBytes, then the range
coder buffer would never get resized. AFAICT this could never
be triggered in Opus because CBR was also controlled by
nbCompressedBytes.

CELT encoding would just fail when setting CELT to CBR in hybrid mode.
It was never a problem because hybrid CBR was always used with
OPUS_BITRATE_MAX.

Makes it possible to run the pre-emphasis with the correct period/coeffs

Not used by anything yet

Zheng Lv authored 2 years ago and Felicia Lim committed 2 years ago

This makes kiss_twiddle_cpx 4-byte aligned (instead of 2-byte) for
fixed-point builds. Tested with an armv6j+nofp development board, CELT
encoding becomes 1.4x as fast, and decoding over 2x.

Performance gain is mostly attributed to the proper alignment of the
static const array mdct_twiddles960.

Co-authored-by: David Gao <davidgao@google.com>
Signed-off-by: Felicia Lim <flim@google.com>

Reviewed by Mark Harris

MULT16_32_QX() is now implemented using a signed-unsigned multiply,
so the second argument can now have one extra bit compared to the
old signed-signed implementation.

Reviewed by Mark Harris

We need to move the history out of the way before we write to the
shape array X, or else we get corruption of the audio.

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

Marcus Asteborg authored 2 years ago and Mark Harris committed 2 years ago

When building with FLOAT_APPROX.

Signed-off-by: Mark Harris <mark.hsj@gmail.com>