DCT-II and DCT-III

[ZacDiggum]

Hi,
I know the scope of this project is FFTs. And I know I can substitute DCTs with FFTs of twice the size of each dimension through mirror reflection of the array. For large 2D arrays this leads to huge memory and speed demands.
Given the similarity of DFT and DCT would it be possible for you to implement kind of a 'fast cosine transform' based on your existing work? Leaving the sine terms out?

[bragadeesh]

Thanks for your interest. Unfortunately, we don't have the resources to implement DCT/DST at this point. We will record this feature request and see if we can implement this in the future. This clFFT library is an open-source project, and we do invite collaborators to extend our library with new features.

[ZacDiggum]

Okay, I understand. Thanks for considering. I'll close this issue now.

[pavanky]

@ZacDiggum I think the issue can stay open.

@bragadeesh The issue originally came through as an arrayfire request. We need to implement it in arrayfire anyway. I'll send it upstream if the API is acceptable on your end.

[bragadeesh]

@pavanky I can take a look after you have the feature implemented. We can evaluate how best to pull it in.

[axeldavy]

I want to add my voice: This would be a great feature !
It is sad to have to use the CPU for the DCT when the DFT speed on GPU is that much improved compared to CPU.

My understanding of:
https://dsp.stackexchange.com/questions/2807/fast-cosine-transform-via-fft/10606

Is that basically the input should be read in a slightly different order, and a final multiplication added (+ take the real part).