Why does indexing take longer?

I wrote a search function that the number of loops vary but can be around 30 million. When the function finds a match, it records the information as follows:

 spur_table_1(b2_index + 1,:) = [harm_0 harm_1 harm_2 order_c band-1 rx1 rx1 harm_0_nz * freq0 harm_0_nz * freq0 harm_1_nz * freq1 harm_1_nz * freq1 harm_2_nz * freq2 harm_2_nz * freq2];

For 30 million loops the routine takes around 50 seconds to complete.

If I change the command to write to the array individually like shown below:

 spur_table_1(b1_index + 1,1) = harm_0;
 spur_table_1(b1_index + 1,2) = harm_1;
 spur_table_1(b1_index + 1,3) = harm_2;
 spur_table_1(b1_index + 1,4) = order_c;
 spur_table_1(b1_index + 1,5) = band - 1;
 spur_table_1(b1_index + 1,6) = rx1;
 spur_table_1(b1_index + 1,7) = rx1;
 spur_table_1(b1_index + 1,8) = harm_0_nz * freq0;
 spur_table_1(b1_index + 1,9) = harm_0_nz * freq0;
 spur_table_1(b1_index + 1,10) = harm_1_nz * freq1;
 spur_table_1(b1_index + 1,11) = harm_1_nz * freq1;
 spur_table_1(b1_index + 1,12) = harm_2_nz * freq2;
 spur_table_1(b1_index + 1,13) = harm_2_nz * freq2;

The routine takes less than two seconds. Why is this?

[Sara]

There's quite a discussion here: http://www.mathworks.com/matlabcentral/answers/54522-why-is-indexing-vectors-matrices-in-matlab-very-inefficient

To make your life easier, if you do:

new_vect = [harm_0 harm_1 harm_2 order_c band-1 rx1 rx1 harm_0_nz * freq0 harm_0_nz * freq0 harm_1_nz * freq1 harm_1_nz * freq1 harm_2_nz * freq2 harm_2_nz * freq2];
for j = 1:numel(new_vect)
    spur_table_1(b2_index + 1,j) = new_vect(j);      
end

it is only two times more consuming than its expanded version on my pc.

Answer 1

[per isakson]

I cannot reproduce your results. There is some crucial information missing in your question

• What is the typical size of the resulting spur_table_1?
• Is spur_table_1 preallocated?

Matlab arrays use column major order (I failed to find a good reference)

    >> M = reshape( (1:12), 3, [] )
    M =
         1     4     7    10
         2     5     8    11
         3     6     9    12

.

I made a test that you may want to study. The fastest code is five hundred times faster than slowest. (R2013a,64bit,Win7)

    cssm( 1e4 )
    ans =
        0.2083    0.2051    0.0084    0.0048    0.0072    0.0069    0.0004

and

    >> cssm( 2e4 )
    ans =
        0.8326    0.8087    0.0166    0.0099    0.0142    0.0137    0.0008

The mfile, cssm, is attached.

.

Maybe you could modify cssm.m to reproduce the difference in execution time that you encounter.

[Cedric]

Very interesting. The most surprising may be that M04 without prealloc beats preallocated cases M05, M06.

[Cedric]

[added to main answer]

[per isakson]

Thanks, I added your case.

[Larry]

I got the following results when I ran the cssm file:

>> elapse = cssm( 1e4 )

elapse =

    0.1240    0.1138    0.0076    0.0040    0.0060    0.0055    0.0003

>> elapse1 = cssm( 1e5 )

elapse1 =

   24.4641   24.1554    0.0712    0.0407    0.0590    0.0588    0.0032

Perhaps there's something in the setup of my MATLAB. I have a networked license but I believe the program and execution is on my computer.

[per isakson]

Our results with cssm(1e4) show that your computer is a bit faster than mine. I see nothing remarkable in the differences between our runs. (I interrupted my execution of cssm(1e5). I didn't want to wait.)

"Perhaps there's something in the setup of my MATLAB" . Why do you believe that?

.

Does the code in your question resembles any of my seven cases? And

• What is the typical size of the resulting spur_table_1?
• Is spur_table_1 preallocated?