Why are exprnd and binornd so slow?

A couple of times I've run into the issue that the function exprnd() for generating exponential random variables is excruciatingly slow. My own function super simple function below is much faster.

function [ ret ] = myexprnd( rate )
  ret = -log(rand(size(rate)))./rate;
end

The same goes for binornd. The following is much much faster:

function ret = mybinornd( n, p )
  ret = sum(rand(1,n)<p);
end

Does matlab have a different set of functions for random number generation that are optimized? Why doesn't mathworks replace exprnd() and binornd() with the code above?

[David]

Hmm, so it seems exprnd( ones(1,10000) ) is much faster than myexprnd( ones( 1, 10000) ), so I guess exprnd is optimized for generating many numbers at once. Fair enough.

[H. Birkan]

Consider the following custom code for generating matrix of binomial rnd numbers

function [ res ] = TEST_SUB_mybinornd( N, p )

[row_cnt, col_cnt] = size(N);

res = zeros(row_cnt, col_cnt);

for ii=1:row_cnt
   for jj=1:col_cnt
      res(ii, jj) = sum(rand(1,N(ii,jj))<p(ii,jj));
   end
end

end

Even if it loops over the matrix entries when I run the following script

ns = 1000*ones(1,100000);
randprobs = rand(1,100000);

fprintf('\nStarting built-in function\n')
t=tic;
myrand = binornd(ns, randprobs);
toc(t)

fprintf('\n\nStarting TEST FUNCTION\n')
t=tic;
myrand = TEST_SUB_mybinornd( ns, randprobs );
toc(t)

I see a huge difference !!!

Starting built-in function Elapsed time is 6.246857 seconds.

Starting TEST FUNCTION Elapsed time is 2.491758 seconds.

Answer 1

[the cyclist]

You can directly see what MATLAB's exprnd() and binornd() functions are doing:

>> edit exprnd
>> edit binornd

The primary difference from yours is that MATLAB is doing various error-checking tasks, which could cause noticeable overhead when generating just one or a few values.