topic Suggestions on how to speed up the following proc IML program - vectorizing? in SAS/IML Software and Matrix Computations

Suggestions on how to speed up the following proc IML program - vectorizing?

Kona — Sun, 20 Oct 2024 00:10:10 GMT

Hi,

I am struggling with how to speed up the following program. I have been trying to vectorize things, however, cannot get things to work. In the example below I have a 5x5 matrix, however, in practice the matrix is much larger, e.g., 5000 x 5000. Thanks so much for any help or suggestions.

-Michael

proc IML;

nP=5;

CM= {0 0.5 0.75 0 0.75 , 1 0 1 0 0.5 , 0.375 0.25 0 0.375 0.25 , 0 0 1 0 0, 1 0.333 0.667 0 0 };

print CM;

D = j(nP,nP,0);

do i1 = 1 to nP;

do i2 = i1+1 to nP; /* I only need to do the calculation on one half of the CM matrix as the resulting D matrix is symmetric */

do k=1 to nP;

if k ^= i1 && k ^= i2 then do; /* I don't use the cases when i1=i2 or i2=i1 */

D[i1,i2] = D[i1,i2] + (CM[i1,k] - CM[i2,k])**2 + (CM[k,i1] - CM[k,i2])**2;

end;

D[i1,i2] = sqrt(D[i1,i2]);

D[i2,i1] = D[i1,i2];

end;

print D;

end;

quit;

run;

Re: Suggestions on how to speed up the following proc IML program - vectorizing?

Ksharp — Tue, 22 Oct 2024 03:24:28 GMT


proc IML;
nP=5;
CM= {0 0.5 0.75 0 0.75 , 1 0 1 0 0.5 , 0.375 0.25 0 0.375 0.25 , 0 0 1 0 0, 1 0.333 0.667 0 0 };
print CM;
 
D = j(nP,nP,0);
  do i1 = 1 to nP;
    do i2 = i1+1 to nP; 
	  temp=(CM[i1,]-CM[i2,])##2 + t((CM[,i1]-CM[,i2])##2);
	  temp[i1||i2]=0;/* I don't use the cases when i1=i2 or i2=i1 */
      D[i1,i2] = sqrt(cusum(temp)[nP]);
  end;
 end;
D=D`+D;
print D;
quit;

Re: Suggestions on how to speed up the following proc IML program - vectorizing?

Kona — Sun, 20 Oct 2024 14:21:15 GMT

Many thanks Ksharp. That was exactly what I was looking for.