Split cell array rows by delimiter (2016b)

I have a vertical cell array of char vectors that I want to split into smaller vertical cell arrays based on rows in the array that serve as delimiters. For example,

x = ...
    {'LINE1'; ...
    '* THIS IS A COMMENT LINE'; ...
    '* THERE CAN BE MORE THAN ONE COMMENT LINE'; ...
    'LINE2'; ...
    'LINE3'};

should be split into

x_split = ...
    {{{'LINE1'}}; ...
    {'LINE2';'LINE3'}};

where lines starting with '* ' are comment identifiers.

I would like the operation to be as fast as possible so I would like a vectorized approach, perhaps involving cellfun/arrayfun. I can get the indices of the comment lines easily enough using cellfun and strncmp, but I'm not sure how to proceed with the splitting.

[Jan]

You forgot to mention, why the first line is stored as a scalar cell array, while the other 2 are a cell vector. Do you want to join the char vectors by using all blocks of comments as separators?

[Hau Kit Yong]

Thank you for the spot. I have edited the desired output. I want every element in the output cell to be cell vectors.

Answer 1

[Jan]

Let's start with a loop approach to clarify at first, what you exactly want:

C = {'LINE1'; ...
    '* THIS IS A COMMENT LINE'; ...
    '* THERE CAN BE MORE THAN ONE COMMENT LINE'; ...
    'LINE2'; ...
    'LINE3'};
limit  = [true, strncmp(C, '*', 1).', true];  % no need for the slow cellfun here!
ini    = strfind(limit, [true, false]);
fin    = strfind(limit, [false, true]) - 1;
n      = numel(ini);
Result = cell(n, 1);
for k = 1:n
    Result{k} = C(ini(k):fin(k));
end

Now you hope that a vectorized approach or cellfun is faster? I do not think so.

Maybe find(diff()) this is faster than calling strfind twice:

limit  = [true, strncmp(C, '*', 1).', true];  % no need for the slow cellfun here!
index  = find(diff(limit))
n      = numel(index) / 2;
Result = cell(n, 1);
for k = 1:n
    Result{k} = C(index(2*k-1):index(2*k)-1);
end

Well, let's try splitapply:

isComment = strncmp(C, '*', 1);
index     = zeros(size(C));
index(strfind([true, isComment], [true, false])) = 1;
index     = cumsum(index);
index(isComment) = NaN;

Result = splitapply(@(x) {x}, C, index);

This seems to be too complex. mat2cell is more direct:

isCmt  = strncmp(C, '*', 1);
limit  = [true, isCmt.', true];
ini    = strfind(limit, [true, false]);
fin    = strfind(limit, [false, true]) - 1;
Rexult = mat2cell(C(~isCmt), (fin - ini + 1).')

Some timings:

C = repmat(C, 10000, 1);  % A larger input
% With tic/toc, Matlab 2019a ONLINE:
% STRFIND:    0.084 sec
% FIND(DIFF): 0.091 sec
% SPLITAPPLY: 0.235 sec
% MAT2CELL:   0.046 sec

The timings in the ONLINE machine need not be accurate, so test it locally again.

[Hau Kit Yong]

I was asking for a vectorized approach as I thought it was generally faster than for loops, but your method is speedy enough! It took about 0.05s to split 150k lines. Thanks for the reminder that strncmp works for cells without cellfun as well.

[Jan]

I've edited the answer and added a splitapply and mat2cell appraoch, which might be considered as "vectorized".