qr decomposition run-time performance

Hi,

I'm comparing the run-time of the "qr" function:

Option1: R = triu(qr(A))

Option2: [~, R]=qr(A) (or [Q, R]=qr(A)),

Only the "R" output is required.

I've tested them on a large number of possible choices of A, a full (not-sparse) matrix with m >> n.

The result was a significant run-time advantage for triu(qr(A)).

To the best of my knowlegde from Matlab documention, option1 use LAPACK xGEQRF Householder Reflectors, but I couldn't find how option2 was implemented.

Can anyone explain what is (or may be) the cause for such performance difference?

(Using Matlab R2019b)

Thank you in advance!

Answer 1

[Bruno Luong]

I'm pretty sure both Q-less-qr and qr use the same algorithm.

What you should try is

Option 3

[~, R] = qr(A,0);

that will return (n x n) R matrix without the bottom part padded with 0.

It's about 2 times slower than q-less qr, but it can be simply explained by Q (m x n) that is built and cleared.

If you do the Option2

[~, R] = qr(A);

the invisible cleared Q is (m x m), totally useless orthogonal much bigger matrix that has been built and then cleared.

[Niv Shapira]

Thank you very much Bruno, i'm looking into it!

[Niv Shapira]

Hi Bruno,

I've done a brief test for the three options (not sure it's done properly):

timeAcc = 0;
for ii = 10000:-1:1
    A = magic(500);
    A = A(:,1:7);
    tStart = tic;
    R = triu(qr(A));
    timeAcc = timeAcc + toc(tStart); 
    disp(timeAcc/(10001-ii));
end
triuQR = timeAcc/10000;

clearvars A tStart R timeAcc
clc

timeAcc = 0;
for ii = 10000:-1:1
    A = magic(500);
    A = A(:,1:7);
    tStart = tic;
    [~,R] = qr(A);
    timeAcc = timeAcc + toc(tStart); 
    disp(timeAcc/(10001-ii));
end
qrTime = timeAcc/10000;

clearvars A tStart R timeAcc
clc

timeAcc = 0;
for ii = 10000:-1:1
    A = magic(500);
    A = A(:,1:7);
    tStart = tic;
    [~,R] = qr(A,0);
    timeAcc = timeAcc + toc(tStart); 
    disp(timeAcc/(10001-ii));
end
compactQR = timeAcc/10000;

The results were:

triuQR = 1.118231647999998e-04 [sec]

qrTime = 0.001107686688700 [sec]

compactQR = 2.532942389999996e-04 [sec]

[Bruno Luong]

This timing results seem to be inline with what I have done earlier today.

[Niv Shapira]

I'll add here a note to whom it may concern:

The test i've performed above on the "magic" matrix "A" was not sufficient. The qr functionality and run-time may vary dramatically depending on the input matrix properties. It seems that for some matrices the run-time difference between triu(qr(A)) to qr(A,0) (or qr(A)) might be even larger.

Answer 2

[Christine Tobler]

When a "~" is used for a return value, this is only for code clarity. MATLAB still needs to compute that value, so the second case is computing the Q, even though you don't intend to use it.