Same code taking 15x longer to execute on a faster machine

I recently upgraded to a faster computer (from a 16Gb i7 processor to a 32Gb RYZEN processor with a 16Gb NIVIDIA RTX GPU (CUDA compatible)).

I used to execute the following code in a matter of 5-10 seconds on my older machine (MATLAB 2021a). Now it is taking >2 minutes on a way faster machine (MATLAB 2022b).

I think it is related to a certain memory allocation setting in MATLAB which I might have applied on my previous machine, but I can't remember it anymore.

Is there someone who could help me in resolving this issue, please? For example, which setting I could modulate to execute this code faster?

Or is it is related to the MATLAB version?

dE00=zeros(6,7362497,"single"); %dE00 refers to a numerical value of color difference
% lab_NF contains six CIELAB color values (L*, a* and b*) while labDB is a
% database of 7362497 CIELAB color values.
for nClrs=1:7362497
    dE00(1:6,nClrs)=deltaE2000(lab_NF(1:6,1:3),labDB(nClrs,1:3));
end

Thanks in advance!

[DGM]

What function are you using? It's not either of these

https://github.com/debuglevel/MatLab-Image-Analyses/blob/master/deltaE2000.m

https://github.com/scienstanford/iqmetrics/blob/master/SCIELAB_FR/deltaE2000.m

because those doesn't support the way you're using it.

It's not this one

https://www.mathworks.com/matlabcentral/fileexchange/65536-camera-color-accuracy-analysis

because that doesn't either, and it doesn't return the same size output.

So what is it?

[Aiman Raza]

Indeed, it is neither of the functions provided by you, I wrote my own DeltaE2000 code as per the original article

https://doi.org/10.1002/col.20070. I have attached the function though I am still wondering if it is the code that is poorly written, it should work even badly on a slow machine, not the opposite. My code works fine and it can take the inputs in a versatile manner. It is just very slow as compared to my old machine which is a big disadvantage for me, since I will be running this snippet multiple times.

Thanks !

[Bruno Luong]

What is the spec of the older computer? I guess it's intel base? Some people complain the slowness of AMD proessor possibly due to the library.

Have you run

bench

command?

I believe TMW claims many operations of small matrices is faster on R2022b (see the release note).

There is memory setting for java heap space, but I don't think it affects the computation speed.

[Bruno Luong]

BTW I run your code with random input

dE00=zeros(6,7362497,"single"); %dE00 refers to a numerical value of color difference
% lab_NF contains six CIELAB color values (L*, a* and b*) while labDB is a
% database of 7362497 CIELAB color values.
lab_NF = rand(6,3,'single');
labDB = rand(7362497,3,'single');
tic
for nClrs=1:7362497
    dE00(1:6,nClrs)=deltaE2000(lab_NF(1:6,1:3),labDB(nClrs,1:3));
end
toc

It takes 90 seconds on my computer (Intel + Windows 11 + R2022b)

[Aiman Raza]

Thanks for your comment. Now that you say that, I do think the AMD processor could be the issue. Earlier I had an intel xeon i7 processor with half the RAM and no GPU and the same command worked better.

I ran the bench command and it shows clear superiority for sparse/2D and 3D operations for a simimar intel i7 processor. Only if I knew this before buying this expensive computer.

JAVA heap space would also reduce available memory for storing arrays, so I am a bit not sure about it.

Any other workaround? Should I try updating to Win 11 or degrading to MATLAB 2021a?

[Bruno Luong]

I don't believe the MATLAB version has anything to do with it. I see constant improvement of speed over time, but nothing that can explain x 15. Also your code does seem to use any gpu so Nvidia card is irrelevant.

[Aiman Raza]

Exactly, the GPU isn't involved so I think it is clearly an issue with Intel vs AMD. Though 90s is an improvement over 2.5 minutes, but slower than 5s-10s.

Do you think I could write this snippet better? I mean just this loop to perform faster or without loop with vectorisation?

[Bruno Luong]

I don't want to answer speeding up your code here. It's off topic.

[Aiman Raza]

Thanks for your help though, it makes more sense to why my code suddenly runs slower. Have a great day!

[Star Strider]

This has been a problem for a while.

See a relevant discussion here: How can I use the BLAS and LAPACK implementations included in AMD Optimizing CPU Libraries (AOCL) with MATLAB R2022a? - .

The last posted Comment (posted about a month ago) expresses my frustration on this.

All of my computers have AMD processors because I prefer them.

.

[Bruno Luong]

@Star Strider can you please run the code (the one I posted) and post the result?

[Aiman Raza]

Thank you for your suggestion. However for me using the BLAS/LAPACK did not change the timing. Unfortunately it still takes the same amount of time. PS: I double checked to make sure that the AOC libraries were correctly installed.

[Bruno Luong]

What you should double check is this:

"I used to execute the following code in a matter of 5-10 seconds on my older machine "

[Aiman Raza]

I ran the code on both the machines before posting my question, and it still holds true. I think I just have a poor processor for MATLAB. In python my code goes way faster than before.

[Bruno Luong]

My laptop intel is quite recent, and it is no where 15 s, as I mention 90 seconds.

To be sure I'll download R2021a and do the same tic/toc.

[Bruno Luong]

I run gain the code on my Intel laptop

R2021a: 85 seconds

R2022b: 87 seconds

[Aiman Raza]

So it clearly is my machine which is just not great. I will try to rewrite the code more smartly to improve the processing duration. Thanks a lot.

Answer 1

[Jan]

n = 32000; % 7362497;

dE00   = zeros(6, n,"single");
lab_NF = rand(6, 3, 'single');
labDB  = rand(n, 3, 'single');
tic
for nClrs = 1:n
    dE00(1:6, nClrs) = deltaE2000(lab_NF(1:6, 1:3), labDB(nClrs, 1:3));
end
toc

Elapsed time is 1.013390 seconds.

% Faster without explicit indexing:
tic
for nClrs = 1:n
    dE00(:, nClrs) = deltaE2000(lab_NF, labDB(nClrs, :));
end
toc

Elapsed time is 0.722065 seconds.

[Aiman Raza]

Thanks a lot! Removing the indexing did reduce the time to 106s. It is still better than before. I will try to code without a loop. That might further improve the timing.