search

xmos / fwk_voice

Voice Framework
Other
11 stars 19 forks source link

Investigate different performance of avona_ic and py_ic #276

Open uvvpavel opened 2 years ago

uvvpavel commented 2 years ago

While doing #270 Allan and I saw that avona_ic and py_ic can have the same behaviour when using different configurations. We want to investigate that by adding some test criteria to the existing tests.

uvvpavel commented 2 years ago

This turns into comparing the filter coefficient between frames to understand where 2 models diverge from each other

uvvpavel commented 2 years ago

I've extended the py_c_frame_compare test to check the filter coefficients in the first ten frames. The wav file from ic example has passed all the tests. Then I gave it a noisy input from the other test and it started failing due to the issue with a 32-bit shift. I have now implemented a workaround in #266 and the test is now passing. @Allan-xmos and I are unsure if it resolves this issue or if there are some other things that we need to find. I need to wait for Jenkins to verify that

uvvpavel commented 2 years ago

The 32-bit shift bug did not resolve this issue. This investigation will be pending for now

andrewdewhurst commented 2 years ago
  1. This is parked, this issues refers to a specific problem when using different config files, and might be a block floating point “feature”. I suggest we move to AV5, and relook at it once we have the refactored python functions.
mbanth commented 2 years ago

A review by @andrewdewhurst and myself has concluded that we will leave this issue parked for the time being. @andrewdewhurst has requested that the Story Points go to zero during this time.