custom number types

[osqp]

custom number types

[goulart-paul]

I'm not necessarily against doing this, but note that I haved use bool, float and int types in the factorisation to keep the memory footprint small. If we are to enumerate overall a variety of input types then we should probably replace bool with int to minimize the number of combinations.

[mlubin]

I'd agree that there's no real need to let the bool type vary. int (or something else set at compile-time config) is a reasonable choice.

[goulart-paul]

I would be happy to insist that the boolean type should always be of type bool (or _Bool), but that doesn't exist in C89. The reason for making it user definable was largely to preserve the option to define it as 'int' if when there is no native boolean type.

If it is to be the same type as the other ints (most likely int64), then there is a fairly large memory penalty if you have a really huge matrix with very low density.

[bstellato]

Another possibility is to specify the types as CMakeLists.txt options in a qdldl_configure.h.in file as done for osqp here. The same could be done with the bool/int decision for memory usage. What do you think?

[mlubin]

Given that C89/C99 support is known at compile time, it's quite reasonable to let this be a config option or just switch on int/bool with macros based on the value of __STDC_VERSION__.

[bstellato]

This should be fixed in 2644bc2. CMake flags for doubles and integers + __STDC_VERSION__ check for bools.