remove 'sh' parameter from FFT.fft

[amigalemming]

The 'sh' parameter is not needed for planning the FFT, thus I propose to remove it. This would allow to re-use the same plan for arrays with arbitrary global shape.

[Haskell-mouse]

I have tried to do this just now, but without any success.
I did it for native and ptx cases (you can write any "sh" value and it will be ignored).

With interpreter problem is that i cant restore innermost dimensions from "Exp sh" value(if i dont know range of array - DIM1 or DIM2 etc...).
Maybe it's a lack of my knowledge of haskell (
Maybe there is solution through Template Haskell.

Another way is writing 3 functions: fft1, fft2, fft3, but i think that it is ugly....

[tmcdonell]

Yes, the reason the sh parameter is still required as a real Haskell value is explained here.

That is a very naive implementation though, so if we can replace it with a real algorithm then perhaps we can avoid this problem. I think this package might have such an implementation, so maybe we should use that instead?

On the other hand, there is little reason to be using the naive implementation at all (at least in the current form; a better algorithm might perform well enough to be useful), so maybe it should be removed and have a requirement that the CUDA or FFTW foreign libraries are used instead?

[Haskell-mouse]

I think that there is should be implementation for interpreter - it is very useful sometimes. For example i can't properly build llvm-4 in my win10(i tried for 2 days, lol....), but i still can use accelerate-fft library and test my code with interpreter on my laptop.

I will try to use accelerate-fourier to make implementation for interpreter without sh param.

[amigalemming]

On Thu, 17 Aug 2017, Rinat Striungis wrote: I think, that i can try to use accelerate-fourier to make implementation for interpreter without sh param.

The accelerate-fourier library provides a full implementation of Fast-Fourier transformations for all data set sizes in plain Accelerate code.

[Haskell-mouse]

That is nice :-)
So - i ll try to use it to replace "interpreter fft" function.

[tmcdonell]

@Haskell-mouse by the way, @yuhangwang has had some progress installing LLVM on windows, and has written up some instructions here; maybe that has some useful points for you too.

[amigalemming]

On Wed, 23 Aug 2017, Trevor L. McDonell wrote: @Haskell-mouse by the way, @yuhangwang has had some progress installing LLVM on windows, and has written up some instructions here; maybe that has some useful points for you too.

Last time, I tried, I succeeded building LLVM-3.8 on MSYS2.

[Haskell-mouse]

@tmcdonell , thank you !
I ll try it later due to very busy right now.

I tried to build llvm-4.0, but it has been built successfully only without recommended by llvm-hs flags, and llvm-hs didnt work.
So... thank you again =)