Hello,
I posted a relevant question about 2d convolution a
while ago but no one answered :(… so here’s another one!
Better typed I hope…
In the convolutionFFT2d example, how do we reconstruct the
convolved output after the ifft?
I mean in terms of indexing…
What I m doing now to put the results in the correct order
is something like this (this loop was already in the example
by the way) :
[codebox]for(y = 0; y < DATA_H; y++){
for(x = 0; x < DATA_W; x++){
conv_GPU << h_ResultGPU[y * FFT_W + x].x << “,” << endl;
}
}[/codebox]
[i]where h_‘ResultGPU’ is the complex array that
contains the results and ‘conv_GPU’ is an ofstream
that writes the values to a text file
which i plot in matlab.
DATA_W * DATA_H is the length of the largest
of the two inputs so i think it reads all the components
I’m interested in from the output.
For both the signals I’m using sizes like 256x256, 512x256, 1024x512 etc.
Quite large…[/i]
The result is not correct though…
I m also doing the comparison with the CPU convolution
and it prints out “TEST PASSED” !!??
I really can’t get my head around this
and I ve tried so many things…
It basically looks as if some blocks of samples in the output,
have a different offset rather than zero…which produces a really
bad distortion.
[u]Do i need to do any more normalization or special indexing
to obtain the correct output values in the correct order
(apart from times 1/N which is in the example anyway…)?[/u]
Is this convolution taking care of all elements (rows and columns)?
Hope there’s someone that can
help a bit with these stuff please?
Any old signal processing wizard? 
Sorries for the number questions…
P.S.
I ve attached a plot of a part of the 2d convolution result
and another of a part of the 1d convolution result
(which is working fine), in case you notice
something in it visually…any common errors.
Both convolutions use the same inputs and same
kernels but they don’t look the same at all.
Thanks for looking,
Filippos