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

Not hooked up

Reducing dependency chains

Reducing the dependency chain between tmp1 and tmp2 at the
cost of an extra multiply.

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

No RTCD yet

800k parameters, 600 MFLOPS, with a receptive field of 3 feature vectors

Starting with compute_linear()

Saves ~270 kB of weights in the decoder

Removing some indirections