how do I use filtfilt on a GPU?

Hello,

I am trying to get filtfilt to run on a GPU and I'm unable to get the formatting right (I think).

Eventually, I'd like to calculate phase-amplitude coupling over a wide range of frequencies with a zero-phase filter.This ends up taking a very long time and the slowest step so far seems to be the bandpass filtering of my signals, so I'd like to run those on my GPU instead of the CPU. I'm basing my filter design on eegfilt.m from the EEGLab toolbox.

I'm trying to get my code to run with a small EEG snippet first. I've replaced the real signal with random numbers for this question, but I can share some real signal later if that helps.

I have tried two different ways of calling filtfilt with arrayfun and am receiving two different error messages:

Attempt1: I tried to use arrayfun in the way I'd normally use it (on a non-gpuArray). This gives 'Use of functional workspace is not supported.'

In attempt 2: when I try to use arrayfun in a way that more closely resembles the GPU-related documentation, I get 'Indexing is not supported'.

Is it possible to use filtfilt with a GPU array at all? Can someone please tell me what stupid thing I'm doing wrong?

srate = 1000;  
locutoff = 10;  
hicutoff = 12;  
testbit = rand(12*srate,1) %testbit = EEG(1:12*srate); 

%filter settings 
minfac = 3;    % this many (lo)cutoff-freq cycles in filter 
min_filtorder = 15;   % minimum filter length 
filtorder = minfac*fix(srate/hicutoff); 
if filtorder < min_filtorder 
    filtorder = min_filtorder; 
end 
filtwts = fir1(filtorder, [locutoff, hicutoff]./(srate/2)); 

% without GPU 
filtsig1 = filtfilt(filtwts,1,testbit');  

%attempt 1 with GPU (does not work)
testbit = gpuArray(testbit');
filtsig2 = arrayfun(@(x) filtfilt(filtwts,1,x),testbit);  

%attempt 2 with GPU (also does not work)
testbit = gpuArray(testbit');
filtsig2 = arrayfun(@filtfilt, filtwts,1,testbit);  

Answer 1

[Infinite_king]

Hi Susan Leemburg,

I understand that you are attempting to utilize the 'filtfilt' function and wish to perform the computation on the GPU.

The 'filtfilt' function fully supports GPU Arrays. Once the input is passed as a 'gpuArray,' the computations are automatically performed on the GPU, eliminating the need for the 'arrayfun' function

You can use the 'filtfilt' as follows to shift the computational load to GPU.

srate = 1000;  
locutoff = 10;  
hicutoff = 12;
testbit = rand(12*srate,1) %testbit = EEG(1:12*srate); 
%filter settings 
minfac = 3;    % this many (lo)cutoff-freq cycles in filter 
min_filtorder = 15;   % minimum filter length 
filtorder = minfac*fix(srate/hicutoff); 
if filtorder < min_filtorder 
    filtorder = min_filtorder; 
end 
filtwts = fir1(filtorder, [locutoff, hicutoff]./(srate/2)); 
% without GPU 
filtsig1 = filtfilt(filtwts,1,testbit');

% Transfer the data to GPU
gpu_testbit = gpuArray(testbit');

% Call the 'filtfilt', with the gpuArray as input
filtsig1 = filtfilt(filtwts,1,gpu_testbit);
% Now the computations are shifted to GPU.

For more information, refer the following MATLAB documentations,

'filtfilt' function - https://www.mathworks.com/help/signal/ref/filtfilt.html#refsect-extended-capabilities

Run MATLAB Functions on a GPU - https://www.mathworks.com/help/parallel-computing/run-matlab-functions-on-a-gpu.html

Hope this is helpful.

[Susan Leemburg]

Thanks! This actually solved the earlier errors I've been getting.

I'm now running into another error: "Error using filtfilt (line 81) Invalid data type. Arguments must be double matrices."

It looks like filtfilt cannot use gpuArray double as a double (my guess: because it's not a regular double). I also have no idea how to fix this, because the documentation insists that filtfilt is fully GPU compatible.

[Infinite_king]

Can you share your modified code, so that i can reproduce the error ?

[Guy Ye]

Strangely, I used the above code to compare the time taken by the filtfilt function in CPU mode and GPU mode, and the result showed that the time taken by the filtfilt function was the same, that is, there was no acceleration in GPU mode. Why is this phenomenon？