idx = ~isnan(Val);
Val2 = Val;
Val2(idx) = cumsum(Val(idx));
Or:
Val2 = cumsum(Val, 'omitnan');
Val2(isnan(Val)) = NaN;
[EDITED] And with a reset of the sum after each NaN block:
Data = [NaN, NaN, 0.5, 0.1, 0.4, NaN, 0.1, 0.2];
X = Data;
idx = isnan(X);
new = strfind(idx(:).', [true, false]); % (:).' because STRFIND needs a row
Y = cumsum(X, 'omitnan');
X(new) = [0; -diff(Y(new))]; % [EDITED 2] ROUNDING PROBLEMS!!!
R = cumsum(X, 'omitnan');
R(idx) = NaN;
[EDITED 2] I've fixed the code, but it suffers from rounding problems. After a certain number of elements you get e.g. 0.0999999999997975 instead of 0.1 . So let's try a simply loop:
function R = cumsumResetNaN(R)
c = 0;
for k = 1:numel(R)
if isnan(R(k))
c = 0;
else
c = c + R(k);
R(k) = c;
end
end
end
This needs 0.00068 sec for input data with 64700 elements, which is faster than the vectorized method above (which has the severe rounding problem).
[EDITED 3] And for completeness a C-Mex function:
#include "mex.h"
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
// Jan Simon, 2017, License: CC BY-SA 3.0
// Create cumulative sum of a double vector, which is reset to 0
// at every NaN.
double *x, *xf, c;
if (!mxIsDouble(prhs[0]) || mxIsSparse(prhs[0]) || mxIsComplex(prhs[0]) ||
(mxGetM(prhs[0]) != 1 && mxGetM(prhs[0]) != 1)) {
mexErrMsgIdAndTxt("JSimon:cumsumRestartNaN:BadInput",
"Input must be a real full double vector.");
}
plhs[0] = mxDuplicateArray(prhs[0]);
x = mxGetPr(plhs[0]);
xf = x + mxGetNumberOfElements(plhs[0]);
c = 0;
while (x < xf) {
if (*x == *x) { // NaN==NaN is FALSE by defintion
c += *x;
*x++ = c;
} else {
c = 0;
x++;
}
}
}
This needs 0.26 sec compared to 0.91 sec for the M-version [EDITED 2] (input data: [64700] elements, 1000 iterations, MSCV2012, Win7/64, Matlab 2016b)
