Speed Testing CoDiPack

[bradbell]

I am working on a tool for comparing AD packages and recently added CoDiPack gradient calculations; see
https://cmpad.readthedocs.io/codi_gradient.hpp.html
The following comparison is for toy sized problems:
https://cmpad.readthedocs.io/xam_main.py.html#xam-main-py-xam-main-csv

The following output indicates that for larger problems CoDiPack is relatively fast (when the time to create the AD tape is included for each function evaluation):

rate,min_time,package,algorithm,n_arg,n_other,time_setup,date,compiler,debug,language,special
1.5e+00,0.500,adept,det_by_minor,100,0,true,2024-7-21,Clang-17.0.6,false,c++,false
1.6e+00,0.500,adolc,det_by_minor,100,0,true,2024-7-21,Clang-17.0.6,false,c++,false
3.8e-01,0.500,autodiff,det_by_minor,100,0,true,2024-7-21,Clang-17.0.6,false,c++,false
1.2e+01,0.500,codi,det_by_minor,100,0,true,2024-7-21,Clang-17.0.6,false,c++,false
3.9e+00,0.500,cppad,det_by_minor,100,0,true,2024-7-21,Clang-17.0.6,false,c++,false
9.6e-01,0.500,sacado,det_by_minor,100,0,true,2024-7-21,Clang-17.0.6,false,c++,false

[MaxSagebaum]

Thanks for the info. Yes these results look quite good.

I had a look at the value on https://cmpad.readthedocs.io/xam_main.py.html#xam-main-py-xam-main-csv. Here, the problem sizes seem to be smaller and CoDiPack is most of the time slower. (I have to plot a graph for this.)

I will have a look in the next weeks. Maybe I can tweak some things. ;-)

[bradbell]

The following instructions:
https://cmpad.readthedocs.io/get_started.html
should enable you to run just the codi speed tests on your own machine and with parameters of you choosing.

[bradbell]

There was a problem with how cmpad was detecting that CoDiPack was available (it was looking for codi.h instead of codi.hpp). This created a problem following the get_started instructions above. It has been fixed.