Why is spdiags() much slower now than in previous versions

I have used spdiags() over many years.

Lately MathWorks have updated the script and the response is much slower.

I have compared spdiags() from R2019a with R2021a on a mac.

The older spdiags() runs in my present test 15 to 40 times faster than the latest spdiags().

I wonder what benefit I get from waiting that much longer, for something that involves plugging numbers into a sparse matrix.

[James Tursa]

Please show the exact code (with arguments) you are using for the comparison.

[Bruno Luong]

Not sure how you get to this conclusion

version
runtest

function runtest
% code that specifies the diagonals and the matrix size.
N = 100000;
D = rand(N,4);
M = 10;
tic
for n = 1:M
    matrix     = spdiags(D,[-2,-1,1,2],N,N); % R2021a spdiags
end
toc
end

Result on 2018b

>> testspdiags

ans =

    '9.5.0.1298439 (R2018b) Update 7'

Elapsed time is 0.498342 seconds.

on R2021a, same PC

>> testspdiags

ans =

    '9.10.0.1649659 (R2021a) Update 1'

Elapsed time is 0.188962 seconds.

[Matt J]

Mads' comment moved here:

When I run your code, I also see no problem. I confirm your results.

I can't send the entire code of mine or the data.

However:

P = 4096

Dm2_... are complex, but with zeros in some places.

Profiler showed quite some time spent in makeSparseGeneral(B, d, A)

    version
    disp('old')
    tic
    matrix = spdiags_old([Dm2_,Dm1_,Dp1_,Dp2_],[-2,-1,1,2],P*3,P*3);
    toc
    disp('new')
    tic
    matrix = spdiags([Dm2_,Dm1_,Dp1_,Dp2_],[-2,-1,1,2],P*3,P*3);
    toc

Two outputs from R2020b for reference:

ans =
    '9.9.0.1592791 (R2020b) Update 5'
old
Elapsed time is 0.007126 seconds.
new
Elapsed time is 0.261705 seconds.
ans =
    '9.9.0.1592791 (R2020b) Update 5'
old
Elapsed time is 0.006587 seconds.
new
Elapsed time is 0.270015 seconds.

MANY outsputs from R2021a, which by the way runs very unstable.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.011837 seconds.
new
Elapsed time is 0.250930 seconds.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.006751 seconds.
new
Elapsed time is 0.249982 seconds.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.006735 seconds.
new
Elapsed time is 0.250596 seconds.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.006552 seconds.
new
Elapsed time is 0.254115 seconds.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.006326 seconds.
new
Elapsed time is 0.249205 seconds.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.007143 seconds.
new
Elapsed time is 0.256610 seconds.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.006193 seconds.
new
Elapsed time is 0.254189 seconds.

ans =

    '9.10.0.1602886 (R2021a)'

old
Elapsed time is 0.006665 seconds.
new
Elapsed time is 0.267214 seconds.

[Brandon Caasenbrood]

I can confirm the same. spdiags has poor performance compared to previous matlab version (e.g. 2021a outperforms 2023a signficantly). Any clue, @MathWorks Support Team?

[Steven Lord]

Please send a small example with which you can reproduce this slowdown to Technical Support directly using the Contact Support link under the Get Support heading at the end of this page.

[Bruno Luong]

@Brandon Caasenbrood "I can confirm the same. spdiags has poor performance compared to previous matlab version (e.g. 2021a outperforms 2023a signficantly)"

I just test the same code I posted above with R2023a, and it is faster than R2021a.

>> testspdiags
ans =
'9.14.0.2206163 (R2023a)'
Elapsed time is 0.153604 seconds. % REMINDER: Elapsed time is 0.188962 seconds. for R2021a

[Brandon Caasenbrood]

I might have shouted wolf too early. But I'm getting subpar performance with spdaig in comparison to speye when make diagonal matrices. I've recently moved from 2021a to 2023a, and only now noticed the performance dip; which made me believe the "6 May 2021" issue above was related. Consider the following code (a segment from the MMA solver by Krister Svanberg - https://gitlab.gbar.dtu.dk/s144426/s144426-TopOpt/-/blob/9e34bb5bb19b2e0f5aebbb4640f4d296627140d6/subsolv.m)

GG and GG_ are identical, but speye outperforms spdiags here. Nevertheless, running the below code, gives indeed similar performance. What am I missing here?

clc;
M = 3e4;
N = 1e3;

tic;
disp('test: spdiags');
a = rand(M,1);
for ii = 1:N
    A = spdiags(a,0,M,M) - spdiags(a/10,0,M,M);
end
toc

tic;
disp('test: speye');
a = rand(M,1);
for ii = 1:N
    A = a.*speye(M,M) - (a/10).*speye(M,M);
end
toc

% test: spdiags
% Elapsed time is 0.814878 seconds.
% test: speye
% Elapsed time is 0.847202 seconds.

The profiler says the bottleneck is in spdiags>makeSparseGeneral.Slightly odd, since I, J, and V are preallocated above, and I also never seen a profiler highlighting the iteration counter before.