Hi Martin!

Thanks for you suggestion!
Could you tell me what dataflow diagram you are refering to?
Using 2d FFT or doing the convolution in the spatial domain?


Martin Brown wrote:
> eugenemi@gmail.com wrote:
> > Rune - thanks for your answer!
> >
> > I understand that I can use 2D FFT to do both
> > conv2(X,K) and conv2(X,K').
> > But I'm trying to use the fact that I'm convolving with
> > K and its transpose to generate results even faster.
> > Do you know of any technique that utilizes
> > the fact that convolution is done with K and K'
> > to achieve faster runtime?
>
> Yes. Provided that K is a given function and you want both
> conv2(X,K) and conv2(X,K') most of the time
>
> Look carefully at redundant operations in the dataflow diagram.
> It should be obvious where you can save time.
> 
> Regards,
> Martin Brown

eugenemi@gmail.com wrote:
> Rune - thanks for your answer!
> 
> I understand that I can use 2D FFT to do both
> conv2(X,K) and conv2(X,K').
> But I'm trying to use the fact that I'm convolving with
> K and its transpose to generate results even faster.
> Do you know of any technique that utilizes
> the fact that convolution is done with K and K'
> to achieve faster runtime?

Yes. Provided that K is a given function and you want both
conv2(X,K) and conv2(X,K') most of the time

Look carefully at redundant operations in the dataflow diagram.
It should be obvious where you can save time.

Regards,
Martin Brown

Rune - thanks for your answer!

I understand that I can use 2D FFT to do both
conv2(X,K) and conv2(X,K').
But I'm trying to use the fact that I'm convolving with
K and its transpose to generate results even faster.
Do you know of any technique that utilizes
the fact that convolution is done with K and K'
to achieve faster runtime?

Rune Allnor wrote:
> eugenemi@gmail.com wrote:
> > Hi - I'm interested in computing 2d convolutions of matrix X
> > with arbitrary kernel K and its non-conjugate transpose K'.
> > Is there a way to compute conv2(X,K) and conv2(X,K') faster than
> doing
> > 2 generic convolutions?
> > Also, if kernel KK has radial symmetry, can conv2(X,KK) be computed
> > faster
> > than doing a generic convolution?
> > Thanks!
> 
> What about a 2D FFT?
> 
> Rune

eugenemi@gmail.com wrote:
> Hi - I'm interested in computing 2d convolutions of matrix X
> with arbitrary kernel K and its non-conjugate transpose K'.
> Is there a way to compute conv2(X,K) and conv2(X,K') faster than
doing
> 2 generic convolutions?
> Also, if kernel KK has radial symmetry, can conv2(X,KK) be computed
> faster
> than doing a generic convolution?
> Thanks!

What about a 2D FFT?

Rune

Hi - I'm interested in computing 2d convolutions of matrix X
with arbitrary kernel K and its non-conjugate transpose K'.
Is there a way to compute conv2(X,K) and conv2(X,K') faster than doing
2 generic convolutions?
Also, if kernel KK has radial symmetry, can conv2(X,KK) be computed
faster
than doing a generic convolution?
Thanks!