Contiguous/non-contiguous decorator

[EmilyBourne]

@dominikbell noticed that the compilation time for Pyccel became significantly slower when we released version 1.8.0.
One reason for this is that the wrapper is now more complicated as it also handles slices. This is not useful for every application.

Additionally knowing that an array is contiguous allows some compilers to provide additional optimisations (Fortran provides a keyword for this).

I wonder if it would be useful to add an @contiguous decorator? This would act similarly to the @allow_negative_indexes decorator and would lead to the mentioned array being marked as contiguous in Fortran. Additionally it would remove the need for the step parameter in the wrapper. In the example provided by @dominikbell this change reduces the time spent compiling the wrapper by half.

[yguclu]

Can you please share a minimal working example?

[said-hadjout]

By the way, I see a potential runtime benefit to specifying the memory contiguous-ness. Have you observed anything like that?

For the runtime benefit, I think we can get some performance because we are doing the memory allocation outside of Fortran using Python, but I don't have any example

[EmilyBourne]

I think we can get some performance because we are doing the memory allocation outside of Fortran using Python

This is true for both the contiguous and the non-contiguous case

[saidctb]

yes, but in the contiguous case, we are telling the compiler that the memory is contiguous, otherwise it will not know, when the array is allocated in Python and it will assume the general case

[EmilyBourne]

This is something I need to test at some point. Does assuming the general case actually lead to slower performance?

[saidctb]

probably it will not allow some optimizations, you have a discussion here https://fortran-lang.discourse.group/t/when-is-contiguous-useful/3082

[saidctb]

If you add the contiguous keyword in a function argument and pass a non-contiguous array, Fortran will create a copy of the array