Suggestions on how to optimize this code (avoid for-loop)?

Hello,

My code organizes data (USD) from a table (tableA) into a 3D matrix (matrixA) using a for-loop. The resulting matrix (matrixA) is as follows:

1st dimension: country 2nd dimension: time 3rd dimension: product

Here's an example:

% CREATE TABLE
product = {'LON';'LON';'LON';'GUA';'GUA'};
country = {'US';'AU';'CA';'US';'CA'};
dt = datetime({'06/30/2016';'03/31/2016';'12/31/2016';'06/30/2016';'03/31/2016'});
USD = [150;200;100;50;75];

tableA = table(product,country,dt,USD);
clear product country dt USD

% VECTORS
products = {'LON';'GUA'};
countries = {'US';'CA';'AU'};
t = datetime('12/31/2015') + calquarters(0:1:4);

% CREATE MATRIX A
matrixA = zeros(3,4,2);

for i = 1:3
    for j = 1:2
        for k = 1:4
            A = ismember(tableA.country,countries(i)) & ...
                ismember(tableA.product,products(j)) & ...
                gt(tableA.dt,t(k)) & le(tableA.dt,t(k+1));
            B = tableA.USD(A);
            matrixA(i,k,j) = sum(B);
        end
    end
end

clear A B i j k

I would appreciate any help on how to do this without a for-loop. Many thanks!!!

[Adam]

Have you run the profiler on it to check that the for loops are the bottleneck?

I had a function recently in which I was clearing some variables mid-function (it was code I wrote 8 years ago!) which seemed harmless and helpful, but when I ran the profiler on the function the one line clearing variables was taking over 50% of the time of the whole function.

Answer 1

[Sean de Wolski]

This does what you want but does not order based on your country/product order that you have hardwired but rather based on uvp, uvc, the first output of unique. The corresponding rows/columns are correct.

[uvp,~,idxp] = unique(tableA.product)
[uvc,~,idxc] = unique(tableA.country)
idxdt = discretize(datenum(tableA.dt),datenum(t))
matrixB = accumarray([idxc idxdt idxp],tableA.USD)

[Alexandre C Borges]

That's great! Exactly what I was trying to do. Thanks!