Wav2letter: mobile backend

[bill-kalog]

Question

Hi, is it still in your plans to release a backend that will work on mobile devices?
If not, could you share any pointers as to how to do it? Do you have to create something similar
to fbgemm but for QNNPACK or did you follow some other approach?

[vineelpratap]

Hi, we are not working on mobile backend at the moment and it is not in our near future plans.

For extending to mobile, one would have to remove the dependency on fbgemm and use a different backend which supports fp16 operations like xnnpack or Accelerate.

INT8 inference with QNNPACK might be more efficient then fp16 but that would require non trivial changes to the code to support scale, zero_point machinery required for INT8 inference.

[trangtv57]

hi @vineelpratap,
I have same question with bill, so as you say, if we need use wav2letter in mobille. We need remove the dependency on fbgemm and use difference backend like xnnpackk or accelrerate.
Can you give more detail about how we remove this dependency. I understand we just need build xnnpack instead of fbgemm, and wav2letter will automatic pick new backend, or we need more changes in code and build script.
Thankyou.

[vineelpratap]

Hi,
Apart changing the build script, you would alsso have to remove the fbgemm function calls that we use in the code to perform matrix multiplication and replace with equivalent from xnnpack (or whatever library being used).

[trangtv57]

thanks youu @vineelpratap