I'm convinced that this is a good idea already:
A = [A(2:end, 2:end), B; B', b];
You save some micro-seconds by avoiding end, because the explicit definition if the size is faster:
[s1, s2] = size(A);
A = [A(2:s1, 2:s2), B; B', b];
In my measurements this is remarkably slower:
A(1:s1-1, 1:s2-1) = A(2:s1, 2:s2);
A(s1, 1:s2-1) = B';
A(1:s1-1, s2) = B;
A(s1, s2) = b;
This might save some temporary memory, but the indexing needs more time.
[EDITED]
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
double *A, *B, *Bp, b, *R;
mwSize i, n, n_m1;
% Get inputs:
A = mxGetPr(prhs[0]);
B = mxGetPr(prhs[1]);
b = mxGetScalar(prhs[2]);
n = mxGetM(prhs[0]);
% Create outputs:
plhs[0] = mxCreateDoubleMatrix(n, n, mxReal);
R = mxGetPr(plhs[0]);
Bp = B;
n_m1 = n - 1;
for (i = 0; i < n_m1; i++) {
memcpy(R, A+1, n_m1 * sizeof(double));
R[n_m1] = *Bp++;
R += n;
A += n;
}
memcpy(R, B, n_m1 * sizeof(double));
R[n_m1] = b;
}
Use this with great care and after exhaustive testing only! This is written in the forum's editor and neither compiled nor tested. Do you know how to compile a C-Mex file?
