Re: [Patch] Fortran: %re/%im fixes for OpenMP/OpenACC + gfc_is_simplify_contiguous

[Julian Brown <julian@codesourcery.com>]

On Tue, 9 Feb 2021 13:45:36 +0100
Tobias Burnus <tobias@codesourcery.com> wrote:

> On 09.02.21 12:58, Thomas Schwinge wrote:
> >> Granted. The array(:)%re access might update too much, but that's
> >> not different to array with strides or with contiguous arrays
> >> sections which contain component reference (and more than one
> >> component).  
> > (Is that indeed allowed to "update too much"?)  
> 
> Yes - that's the general problem with strides or bit sets;
> copying only a subset – and doing so atomically – is not
> always possible or feasible.
> 
> *OpenACC* 3.1 has for "2.14.4  Update Directive" the restriction:
> 
> "Noncontiguous subarrays may appear. It is implementation-specific
>   whether noncontiguous regions are updated by using one transfer
>   for each contiguous subregion, or whether the noncontiguous data
>   is packed, transferred once, and unpacked, or whether one or more
>   larger subarrays (no larger than the smallest contiguous region
>   that contains the specified subarray) are updated."
> 
> For map, I saw that that's the case – but I think Julian's
> patch does not handle this correctly for:
> 
> type t
>    integer :: i, j, k
> end type t
> type(t) :: A(100)
>    ... host(A(:)%j)
> 
> I think instead of transferring A(1)%j to A(100)%j, it transfers
> all of A(:), i.e. also A(1)%i and A(100)%k :-(
> 
> ^– Julian?

Yes it will -- but given that A(2)%i and A(99)%k (and all the in-between
values) can legitimately be transferred according to the spec, how much
of a problem is that? In particular, are there situations where this
"over-updating" can lead to incorrect results in a conforming program?

Perhaps the question is, can a user legitimately expect the host and
offloaded versions of some given memory block to hold different data,
like maintaining different data in a cache than the storage backing
that cache? One use-case for that might be double buffering a "single
array" (i.e. the host and device versions of that array). I don't think
that's something we'd want to encourage, though.

I think, rather, that partial updates are an optimisation the user can
use when they know that only part of an array has been updated, so
slight over-copying is harmless.

Thanks,

Julian