Behavior of ippsFIR_32f

[robert_inkol]

We have examined the performance of the ippFIR_32f implementation of the FIR filter and would like to confirm our understanding of its implementation.

First, we have noticed that when used as a single rate filter, the computational cost (number of clock cycles needed to process an input data sample)tends toriseas the number of filter coefficients isincreased, but drops significantly each time the number of coefficientsreaches a power of two. Second, when used as a multirate filter, we find that the computational cost is a roughly linear function of the number of coefficients, except for a real filter which shows a jump in computational cost when the number of coefficients is about 3x the decimation factor.

My inference is that the single rate filter is based on a frequency domain filter implementation using FFTs and that the multirate implementation is based on a polyphase filter architecure (i.e., a time domain approach).

Is this correct?

Thanks

Robert

[Vladimir_Dudnik]

Hi Robert,

there is answer from our expert

Yes, this is correct. Single rate filter has direct implementation (time domain) until order < BOUNDARY, whereBOUNDARY depends on platform (IA32 or IA64) and real or complex data. If order > BOUNDARY, FFT algorithm is used. MR is always time domain and has 3 different algorithms in dependence on order, upfactor and downfactor.

Regards,
Vladimir

[poozir]

Hi Vladimir,

is there a way to find out, what size "BOUNDARY" is (on my machine) ?
Like a special method for this?

Thank you!