Avoids valgrind complaining about use of uninitialized memory

Removes float debug weights, as well as useless spaces

Use OPUS_DRED instead of NEURAL_FEC

Also default to disabling dnn float debugging

Still don't quite know what I'm doing

Compilers sometimes replace vmlaq*() with fmul+fadd instead of fmla.
Trying to use vfmaq*() instead when possible.

Finished adversarial training on 800k model. Also, move weights to a new
location.

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

Adds RTCD tables for compute_activation() and compute_conv2d()

Avoids having to write intrinsics for simple loops

Enabling only on platforms that have been tested just in case we
run into a non-IEEE754 platform where they would break.

Still missing some intrinsics

Used for DNN matrix multiplies

Code moved to compute_frame_features()

for cmake, force PRESEUME_SSE4_1 on PRESUME_AVX2

aka banging on it until it builds on my machine.
Further improvements welcome

Not yet with rtcd

DRED will absorb the bitrate variation

Fixes warnings, undefined behaviour, and check-asm failure

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

The optimization is bit-exact with C function.

This optimization speeds up SILK encoder (floating point) as following:

AMD Zen:
Complexity 0-5 :      0%
Complexity 6-7 : 3 -  7%
Complexity 8-10: 8 - 15%

Intel Skylake:
Complexity 0-5 :       0%
Complexity 6-7 : 14 - 18%
Complexity 8-10: 17 - 22%

Adapted by Jean-Marc Valin