Skip last 1by1 layer and residual connection

[fps]

This saves between 0.5 and 1% of cpu (relative to without this change). In my own implementation I have it as a template parameter [1][2]. But I wasn't sure how to add it to yours. I guess the 20 comparisons (one for each layer) might pale in comparison to the matrix multiplication and addition that is saved. But possibly making it a template parameter might save another small fraction of a percent :)

[1] https://github.com/fps/anna/blob/master/examples/nam_wavenet.cpp
[2] https://github.com/fps/anna/blob/master/include/anna/nam.hpp

[mikeoliphant]

Yeah, the outputs from the last layer aren't used. Not a big performance impact, but every bit helps.

I'm inclined to try to do this with a template parameter (although it is unlikely to make much performance impact).

Should be able to get rid of the "arrayOutputs" matrix as well, since it is just there to catch the unused output from the last layer.

[fps]

Here's another small optimization opportunity:

NeuralAudio/NeuralAudio/WaveNet.h

Line 399 in 51ba25a

headArray.setZero();

sets the headArray to zero. The first WaveNetLayerT could instead use = instead of +=. This would save the zeroing. Again this could be done with a template parameter and an if constexpr(first). I have done so in my implementation:

https://github.com/fps/anna/blob/cdef14fe0b9e3bf96aef1b0b492e771071710348/include/anna/nam.hpp#L68

I'm still not beating yours though. Probably because of my (more costly) buffer handling ;)

[mikeoliphant]

👍