Fortran ports

[jeffhammond]

This is a new implementation of BabelStream using Fortran.

The code uses a Fortran driver that is largely equivalent to the C++ one, with a few exceptions. First, it does not use a C++ class for the stream object, since that doesn't seem like a useful way to do things in Fortran. Instead, I use a module that contains the same methods, and which has alloc and dealloc that act like CTOR and DTOR.

The current implementations are:

• DO CONCURRENT
• Fortran array notation
• Sequential DO loops
• OpenACC parallel loop
• OpenACC kernels on Fortran array notation
• OpenMP parallel do
• OpenMP taskloop
• OpenMP target teams distribute parallel do simd
• OpenMP target teams loop
• CUDA Fortran (handwritten CUDA Fortran kernels, except DOT)
• CUDA Fortran kernels (!$cuf kernel do <<<*,*>>>)

I have tested with GCC, Intel (ifort and ifx), and NVHPC compilers on AArch64, x86_64 and NVIDIA GPU targets, although not exhaustively.

The current build system is GNU Make, and requires the user to manually specify the compiler and implementation. I have not, and will not, do anything related to CMake.

~The only thing missing now is CSV output.~ CSV printing is supported.

[jefflarkin]

Comment to the reviewers: For the array version, would you consider adding !$acc kernels around the sections of array syntax? In theory, it would also be valid to put an !$omp workshare around them for OpenMP compilers. I'm not sure whether any compilers will auto-parallelize, much less auto-offload, the array syntax, but a variety of compilers will happily support one or both of those parallelization hints.

[tomdeakin]

@jefflarkin - definitely worth exploring what happens in both of those cases for sure!

[jeffhammond]

@jefflarkin See OpenACCArrayStream.F90

[jeffhammond]

OpenMPWorkshareStream.F90 is there too now, but it should not offload, because OpenMP does not have omp target workshare yet, as far as I know. It should, but someone has to argue with the workshare haters on the committee.

[jeffhammond]

I removed the IEEE NaN check from this branch, because I did it wrong. It is now fixed, and there are other improvements in other branches, which I'll merge after the paper is accepted. I don't want to break the branch linked in the paper for obvious reasons.

I've now implemented CSV printing, while REAL32 and INT32 options both seem to be working now too.

[jeffhammond]

This is ready for review. I merged in all the good stuff done post-submission.