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.

Starting with compute_linear()

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

Not so much for old machines, as for getting decent performance
when not setting -march= (SSE2 is part of the amd64 ABI).

Should be able to handle all previous GRU variants and more.

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

Depending on what defines are set there is collisions with the ones
in Opus. To avoid these errors we rename the exp functions and
macros.

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

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

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

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

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

AVX without AVX2 should now work again too.

The 4* is now stored in the table to avoid computing it in the loop

Saves on the MDense/softmax computation since we only need to compute
8 values instead of 256.

Using rational function approximation for tanh() and sigmoid.

This isn't necessary since valid exponents can't flip the sign bit

See https://github.com/mozilla/LPCNet/commit/bffdcee95#commitcomment-46372726