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

clang exposes this intrinsic even in 32-bit mode, if targeting >= armv8,
whereas gcc does not, see:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95399


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

silk_burg_analysis() could return a slightly negative value on zero
input, which would cause a negative, which the log didn't like.

-fargan-synthesis not -fargan_synthesis

Timothy B. Terriberry authored 1 year ago and Jean-Marc Valin committed 1 year ago

The existing code in vec_avx.h produced
  warning: dereferencing type-punned pointer will break
   strict-aliasing rules
 with gcc 6.4.0.
We already had a macro to work around this within the rules of the
 C standard, but trying to use that here does not get optimized
 into a single MOVD like we were hoping.
Replacing it with memcpy() instead does get optimized correctly,
 but requires switching from a macro to an inline function in order
 to be able to declare a local variable and return a value.
We already have such an inline function in NSQ_del_dec_avx2.c, so
 hoist that out and use it everywhere, and then convert vec_avx.h
 to use it also.

Also skip the first loss values being generated since they're
biased towards "not lost" due to the initialization.

Fails ubsan because memcpy declares args as non-null

No longer needed now that PLC is trained with PyTorch stack

Should handle the history in a more consistent way.
Slightly increase the model size and re-enable biased band loss in
training.

mostly untested

Should match the TF2 code, but mostly untested