clear;
clear all;
clc;

for i=500:500:9000
A=rand(i,i);
B=rand(i,1);
i

tic;
C=A\B;
time1=toc;
fprintf('CPU time = %3.3f sec\n', time1);

Ag=gpuArray(A);
Bg=gpuArray(B);
tic;
Cg=Ag\Bg;
time2=toc;
C=gather(Cg);
fprintf('GPU time for only solving = %3.3f sec\n', time2);
clear A B C Ag Bg Cg;

A=rand(i,i);
B=rand(i,1);
Ag=gpuArray(A);
tic;
Bg=gpuArray(B);
Cg=Ag\Bg;
time2=toc;
C=gather(Cg);
fprintf('GPU time for transfer of F plus solving = %3.3f sec\n', time2);
clear A B C Ag Bg Cg;

A=rand(i,i);
B=rand(i,1);
Ag=gpuArray(A);
Bg=gpuArray(B);
tic;
Cg=Ag\Bg;
C=gather(Cg);
time2=toc;
fprintf('GPU Time for solving plus transfer of result = %3.3f sec\n', time2);
clear A B C Ag Bg Cg;

A=rand(i,i);
B=rand(i,1);
Ag=gpuArray(A);
tic;
Bg=gpuArray(B);
Cg=Ag\Bg;
C=gather(Cg);
time2=toc;
fprintf('GPU Time for transfer of F plus solving plus transfer of result = %3.3f sec\n', time2);
clear A B C Ag Bg Cg;

A=rand(i,i);
B=rand(i,1);
tic;
Ag=gpuArray(A);
Bg=gpuArray(B);
Cg=Ag\Bg;
C=gather(Cg);
time2=toc;
fprintf('GPU Time for solving and all data transfers including K = %3.3f sec\n', time2);
clear A B C Ag Bg Cg;
end