Fix for fixed point overflow in synthio #10288
Conversation
|
I updated the tests as this change breaks most of them. I looked and as far as I could tell it is small rounding differences due to how I fixed this issue. I also listened but would appreciate it if someone else can confirm that things sound right. The rounding changes were about 1 (for ints) or less then 0.0001 for floats so not large for audio. |
Not that it's necessary for this PR, but would |
| @@ -212,7 +213,7 @@ void synthio_biquad_filter_samples(mp_obj_t self_in, biquad_filter_state *st, in | |||
|
|
|||
| for (size_t n = n_samples; n; --n, ++buffer) { | |||
| int32_t input = *buffer; | |||
| int32_t output = (b0 * input + b1 * x0 + b2 * x1 - a1 * y0 - a2 * y1 + (1 << (BIQUAD_SHIFT - 1))) >> BIQUAD_SHIFT; | |||
| int32_t output = synthio_sat16((b0 * input + b1 * x0 + b2 * x1 - a1 * y0 - a2 * y1 + (1 << (BIQUAD_SHIFT - 1))), 15); | |||
There was a problem hiding this comment.
This should be BIQUAD_SHIFT instead of 15.
|
Just finished running tests. No more distortion across the board! Now that we can push Q up to much higher values, I'm a bit disappointed that it doesn't appear to want to self-oscillate like you'd expect from an analog filter. After a bit of research, I was able to find some anecdotes about biquads not being able to self-oscillate without the help of additional processing (using an envelope follower controlling gain). For our purposes, I think we can live without it. |
Fix for #10286 and #10200.
There may still be more locations where fixed point could be handled better in
synthiobut this fix hits two big ones that I discovered so worth fixing those now.Also refactored
sat16that will shift a fixed point result back into 16 bits out ofaudiofreeverband intosynthioto be used by any audio module that needs it. Potentially worth moving this to a generic "audiohelper" type module in the future in case someone builds without synthio but still wants audio effects.