Improving speed of 1D fft over 3D data

HI. I am trying to make an image reconstruction script for OCT data.

I vectorized all the code, but it is still quite slow and memory consuming.

I tried using gpuArray but the size of the data(raw_3D) exceedings the size of videoram.

So, as second trial, I split the data to multiple of 2D images([N,sx] of sy) and put this 2D images to gpuArray.

But it took much longer than vectorized CPU code. might be overhead of cpu-gpu data transfer..

Also, I have no clue using 'arrayfun' for my scripts. using 'parfor' also doesn't help...

I need to run this script at least 600 times per experiments, so reducing this time really matters.

Please Help! cc

% raw_3D : [N,sx,sy], 1d (2048 points)spectrometer data with 1000x1000 of points 
% shiftMap : [sx,sy], contains shifting value per each data points 
% IMGContrast : constant 
% IMGBrightness : constant 

raw_3D = rand(2048,1000,1000);
shiftMap = zeros(sx,sy);

N = size(row_3D,1); 
sx = size(raw_3D,2); 
sy = size(raw_3D,3); 

%windows
win_tukey =	tukeywin(N);
windows =       repmat(coef_tukey,1,sx,sy);

clear basicImage
%% (1) Original script: vectorized CPU code
tic
basicImage = fft(raw_3D.*windows,[],1);
toc
%when variable 'basicImage' exists : 59sec, 70sec, 50sec..., when variable 'basicImage' doesn't exists : 18sec, 20sec

%%-------------

%% (2) GPU implementation
tic
basicImage=zeros(N,sx,sy);

for index=1:sy
	gpuRaw=gpuArray(raw_3D(:,:,index));
	basicImage(:,:,index) =gather(fft(gpuRaw));	
end

toc
%%------------ 78 seconds

%% (3) Parfor
tic
basicImage=zeros(N,sx,sy);
parfor index=1:sy
	gpuRaw=gpuArray(raw_3D(:,:,index));
	basicImage(:,:,index) =gather(fft(gpuRaw));
end
toc
%%------------
% 225 second
% doesn't help because have only one GPU?