search

parallella / pal

An optimized C library for math, parallel processing and data movement
Apache License 2.0
301 stars 110 forks source link

Implemented decimating FIR filter #186

Closed shimshir closed 9 years ago

shimshir commented 9 years ago

I didn't understand what is mean by "This function retains the address of the delay filter memory containing the previous delayed values to allow consecutive processing of blocks."

aolofsson commented 9 years ago

The question is whether the function retains state.

shimshir commented 9 years ago

State of what exactly, the state of the taps?! If so then there are (nh + 1) * nx different states, what should be done with them? I would be happy to implement additional behavior I just need a more detailed explanation of the requirements.

lfochamon commented 9 years ago

@shimshir The idea is to internally retain the last buffer used in the calculations, in this case nh*df last elements from x. However, I think a better option would be to initialize the delay line instead of saving the last state. Keeping an internal state would make it very hard for a user to filter two different buffers sequentially.

I've been thinking that maybe a better solution would be to only return values with a full delay line, ignoring the initial transient. The output of the filter function would be nx-nh+1 (instead of nx) and the initial state would be the first nh elements of x. Since the length of r is smaller than the length of x, in-place filtering is possible (although there would be nh-1 "garbage values" at the end of the output array). Still, I think it would be a good trade off.

@aolofsson What do you think?

shimshir commented 9 years ago

@lchamon Having the option to initialize the delay line would be a good solution, but wouldn't the function signature have to be changed/extended by the initialization vector. As far as I understand it you want something similar to the Python lfilter function of the scipy package, particularly the zi vector, am I right?

lfochamon commented 9 years ago

The idea is that the "initialization" would already be contained in x: the first nh elements of x. In your implementation, it would be equivalent to removing the first loop and adjusting the indexing of r.

And you are right: the idea is that the function keeps zi/zf from scipy internally (I believe it's similar to the *dbuf from the TI DSP library that @aolofsson referenced in #158).