search

xmos / fwk_voice

Voice Framework
Other
11 stars 19 forks source link

Test Aaron's suggestions regarding bfp division #287

Closed uvvpavel closed 2 years ago

uvvpavel commented 2 years ago

rescale_vector function in noise suppression doesn't use vpu because bfp implementation leads to poor precision. The non-vpu implementation is taking a lot of cycles. Aaron gave me some tips on how to implement long division without consuming many cycles. The discussion with the tips can be found here: https://github.com/xmos/lib_xs3_math/issues/89

mbanth commented 2 years ago

@uvvpavel, please update the Story Points field to state the number of engineering days you think this effort will take. Also, update the Status field to reflect the actual status of the issue. Thanks.

uvvpavel commented 2 years ago

@mbanth I did some tests regarding this issue, and they still fail. I fell like it doesn't make much sense to test it with random data since it exposes bfp weakness and is not quite realistic. I can design a python-based test which will generate noise and test this function. However, it needs to test against something, ideally against the pyhton model. In order to do that python model needs to be up to date with xc/c implementation. I've raised an issue in the lib_noise_supression that covers that. Let me know what you think about it.

uvvpavel commented 2 years ago

After talking to Michael we decided to close this issue until we have a python model we can use for testing. The issue https://github.com/xmos/lib_noise_suppression/issues/176 covers the work to fix the model.