Slightly different results on differents MCX and environments

[Ilya Kuznetsov]

Hi, Prof. Fang!

Thank your very much for MCX tool!

We are trying to move from MATLAB to more 'open' environments and I am testing my code with Octave and pmcx.
I found slightly different results and I would like to check it with you.

I have: MATLAB + Windows

$Rev::0fa060 $ v2025 $Date::2025-02-13 17:38:58 -05$ by $Author::Qianqian Fang$
- code name: [Jumbo Jolt] compiled by nvcc [7.5] for CUDA-arch [350] on [Feb 13 2025]
- compiled with: RNG [xorshift128+] with Seed Length [4]

Octave + WSL2

$Rev::5332f0 $ v2025 $Date::2025-02-16 00:22:41 -05$ by $Author::Qianqian Fang$
- code name: [Jumbo Jolt] compiled by nvcc [9.0] for CUDA-arch [350] on [Feb 23 2025]
- compiled with: RNG [xorshift128+] with Seed Length [4]

PMCX + Windows 

$Rev::0fa060 $ v2025 $Date::2025-02-13 17:38:58 -05$ by $Author::Qianqian Fang$
- code name: [Fermi MCX] compiled by nvcc [10.2] for CUDA-arch [350] on [Feb 13 2025]
- compiled with: RNG [xorshift128+] with Seed Length [4]

• For Octave and MATLAB I have an identical code, I literraly run the same file.
• For pmcx I save cfg from MATLAB mat file, convert it to python format and just run pmcx.mcxlab(cfg)
• seed everywhere is constant and equal 10000000

For MATLAB and PMCX I see identical amount of detected photons, but different amount of E on the detectors (like 4.5773e-06 for MATLAB and 4.047e-06 for PMCX version).

But for MATLAB and Octave I see different amount of detected photons (0.05% difference), but result of simulation (after mcxdref) are closer between MATLAB and Octave.

Could you give advice or share your ideas why the results are different everywhere? 

I thought pmcx and MATLAB versions should get the same results due the same revision and OS (but the codename is strangely different, maybe a typo).

[Qianqian Fang]

hi Ilya,

exact reproducibility between simulations, even within the same environment, is not guaranteed for MCX, and accumulation operations on the GPU in general.

please see previous discussions on this topic

https://groups.google.com/g/mcx-users/c/DXJ9C70Sajo/m/TVmGaK0NAQAJ

https://groups.google.com/g/mcx-users/c/pFp8wstktw4/m/KXeRmkcyAwAJ

https://groups.google.com/g/mcx-users/c/E9VgPb4q62M/m/d6G_9Dr7AgAJ

if the simulations runs on different GPUs, or different OSes, the chance of getting nearly identical output is unlikely - by default, mcx determines a total number of threads based on the GPUs core count, and different GPUs will use different number of threads - even the seeds are the same, the results won't be the same; similarly, different GPU driver version also changes the behaviors of thread scheduler and causes the order of operations non-repeatible.

> For MATLAB and PMCX I see identical amount of detected photons, but different amount of E on the detectors (like 4.5773e-06 for MATLAB and 4.047e-06 for PMCX version).

what do you mean by different amount of "E"? detected photon weight? how did you compute this number?

Qianqian

--
You received this message because you are subscribed to the Google Groups "mcx-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mcx-users+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/mcx-users/d5e502db-23e7-415b-89eb-fe61bea6a870n%40googlegroups.com.

[Ilya Kuznetsov]

Thanks for your answer,  I understand now.

Regarding amount of E on the detector - I run simulation using cfg.outputtype = 'energy' and then calculate energy on the detector using mcxdref, but I changed mcxdref.m in my project - I don't divide detweightsum./cfg.nphoton by area of the detector. 

This way I am getting amount of energy on each detector, assuming that input energy is 1 joule.

[Qianqian Fang]

if you are trying to get the total detected "energy" at a given detector, you do not need to set outputtype to 'energy', because that flag only specifies the volumetric output (flux.data), and has no effect to detected photon data (detp).

I assume you meant mcxcwdref.m instead of mcxdref.m.

To view this discussion visit https://groups.google.com/d/msgid/mcx-users/14264c26-6f07-4e2c-b1c5-43a701ea7603n%40googlegroups.com.

[Ilya Kuznetsov]

Thanks!

Yes, I meant mcxcwdref.m.