FFT is suddenly slow down for repeated operations.

Hello. I am conducting a simulation with a 3-dimensional matrix using Matlab. While performing FFT about one dimension of the 3-dimensional matrix using GPU for repeated operations, I noticed that the computation speed suddenly slows down. Below is a brief summary of the problematic code.

matlab

clear all; close all; clc;
k = 501;
A = zeros(k,k,k,'gpuArray');
B = zeros(k,k,k,'gpuArray');
for n = 1:k
    tic
    B(:,:,:) = fftshift(fft(ifftshift(A,1),[],1),1);  
    disp(num2str([n, toc]))
end

The result of running the code is as follows:

Result

....

236 7.52e-05

237 7.03e-05

238 7.11e-05

239 7.34e-05

240 7.29e-05

241 7.02e-05

242 0.0235287

243 0.0472623

244 0.0312132

245 0.0444045

246 0.0305904

247 0.0463637

248 0.0309244

....

As the result, from the 242nd iteration, the FFT operation significantly slows down.

I am curious about the cause of this issue.

The GPU used is 4090.

I would appreciate it if you could let me know what the problem might be.

Answer 1

[Edric Ellis]

This is due to the asynchronous nature of gpuArray operations. Where possible, operations on the GPU are queued up and run in the background. At some point, we run out of queue, and the operations have to actually run - that's the point at which you see the change in time.

The best way to time operations on the GPU is to use gputimeit which takes into account this asynchronous behaviour. The other option is to call wait on the gpuDevice before calling toc, like this:

d = gpuDevice();
t = tic();
% do stuff on the GPU
wait(d);
time = toc(t)

[Hosung Jeon]

Thank you for your response. I understand now!