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

Only code non-degenerate symbols, which makes the encoder faster

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

Update silk/x86/NSQ_del_dec_sse4_1.c to match the remaining
silk/NSQ_del_dec.c changes made in
c913dc38



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

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

Co-authored-by: James Zern <jzern@google.com>

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

Including removing useless code

Removing one of the 2d conv layers for pitch estimation reduces
complexity without noticeable degradation. FARGAN model has more
adversarial training.
Also, no need for the double precision in the low-pass filter.