Alignment pragma

[sheino]

Added alignment pragma, which replaces ivdep pragma if running with OpenMP

[sheino]

Probably set_aligned_pragma needs to go somewhere else...
I just put it inside compiler as i found self.pragma_aligned var there and wanted to be consistent with having all pragma declarations inside compiler.

Thoughs @mlange05 ?

[navjotk]

I think this belongs in Propagator. Also, I think the method could be made more generic. I would hold on to accepting this PR until the test comparing the save=True and save=False case as pointed out in #62 is ready and passes here.

[pvelesko]

but all the pragmas are set up and inside compiler.py and this is a pragma as well.

[navjotk]

It needs to be aware of the concepts of factors, loop counters and time steppers. All implementation choices in Propagator which might be changed very soon. compiler.py deals with the compiler interface, something we expect to not change frequently.

Why introduce this unnecessary coupling between the two?

Why does someone working on the compiler interface need to know what time steppers are?

If I change how my factors are put into the stencil, why should the compiler interface need to be changed?

If I change my loop variables, why should my compiler interface need to know about it?

[mlange05]

Ok, so the compiler is supposed to provide the "wording" of pragmas (due to toolchain-dependent variants of ivdep originally). So at the most it should provide the string omp simd aligned, but since the whole thing needs to be parameterised, how about providing it as a function rather than a property? This way the propagator does all the sanity checking and only inserts the pragma_str, something like compiler.aligned_pragma(pragma_str).

[FabioLuporini]

is this equivalent to

for item in flatten(stencil.free_symbols for stencil in stencils):
    if str(item) not in factorized and item not in loop_counters + time_steppers:
        array_names.add(item)

[sheino]

@navjotk @mlange05 waiting for comments or a go if it looks alright

[FabioLuporini]

I pushed a small change to make everything more pythonic (also fixes the docstring)

[mlange05]

Was the pragma string not supposed to come from the Compiler?

[mlange05]

Should the alignment size (:64) really be hard-coded, or should this come from the respective Compiler ?

[ggorman]

Good catch. I think we talked about this before. Such constants should be determined programmatically using sysctlbyname: http://www.manpagez.com/man/3/sysctlbyname/
In the case of cache line size you are looking for hw.cachelinesize

However, and I believe @FabioLuporini raised this previously as well, we might instead want to align to page size in which case we should use hw.pagesize so we can reduce TBL misses. The amount of wasted memory is still small.

[mlange05]

The use of %s without replacement value looks wrong. Can we move the (%s:%s) part of the string to the caller?

[mlange05]

Looks good now. Merging...