Re-arrange matrix based on index on GPU using arrayfun

2 次查看(过去 30 天)
Hi,
I want to re-arrange a two dimensions frequency domain x,y to p=sqrt(x^2+y^2) on GPU. I can do so with the code below. However, it it is pretty slow for my application (2 seconds). I have read about arrayfun but could not make it work for my purpose. Any help is much appreciated.
nx=200;
Nx=floor(nx/2);
ny=100;
Ny=floor(ny/2);
t=200;
Nt=floor(t/2);
A=gpuArray(single(rand(nx,ny,t)));
A=fft2(A);
A=A(1:Ny,1:Nx,:);
freqx=0:1/nx:1/2-(1/nx);
freqy=0:1/ny:1/2-(1/ny);
A=fft(A,nz,3);
G=gpuArray(single(zeros(Ny*Nx,Nt+1)));
for ii=1:Nx
for jj=1:Ny
G((ii-1)*ny+jj,1)=sqrt(freqx(1,ii)^2+freqy(1,jj)^2);
G((ii-1)*ny+jj,2:end)=abs(squeeze(A(jj,ii,1:Nt)));
end
end
Ma=max(G(:,1));
bin=Ma/100;
Y=gpuArray(single(zeros(100,Nt+1)));
for ii=1:100
loik= (G(:,1)>=(bin*(ii-1)) & G(:,1)<(bin*(ii)));
Y(ii,:)=mean(G(loik,:));
end
Any help is appreciated.
Thanks
  4 个评论
Walter Roberson
Walter Roberson 2019-7-30
dsolve cannot be vectorized and cannot use gpu. It can be run in parallel on the cpu using the Parallel Computing toolbox.
The ability to use character vectors to define ode is going away soon. You should be rewriting in terms of symbolic equations.
Sometimes you can define the boundary conditions as symbolic and dsolve that, and then afterwards just subs() the numeric boundaries into the result. This does not always work even in cases where dsolve seems to find solutions for all initial conditions that you try.
You should also look at using odeFunction to transform your ode into something that can be processed numerically with ode45 or similar.
Tutu
Tutu 2019-8-2
Thanks Walter Roberson for your answers. I will accept your repply as the answer to this question.

请先登录,再进行评论。

回答(0 个)

类别

Help CenterFile Exchange 中查找有关 Special Functions 的更多信息

产品


版本

R2019a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by