Re: [Fortran, Patch] Memory sync after coarray image control statements and assignment

[Tobias Burnus <tobias.burnus@physik.fu-berlin.de>]

On Mon, Dec 07, 2015 at 02:09:22PM +0000, Matthew Wahab wrote:
> >>I wonder whether using
> >>__asm__ __volatile__ ("":::"memory");
> >>would be sufficient as it has a way lower overhead than
> >>__sync_synchronize().
>
> I don't know anything about Fortran or coarrays and I'm curious
> whether this affects architectures with weak memory models. Is the
> barrier only needed to stop reordering by the compiler or is does it
> also need to stop reordering by the hardware?

Short answer: I think no mfence is needed as either the communication
is local (to the thread/process) - in which case the hardware will act
correctly - or the communication is remote (different thread, process,
communication to different computer via interlink [ethernet, infiniband,
...]); and in the later case, the communication library has to deal with
it.

However, I think that except for SYNC using
  __asm__ __volatile__ ("":::"memory");
is the wrong solution - even though it might work as band aid.


 * * *


To gether some suggestions, here is how coarrays work. They are based on:
everything is a local variable - except it is a coarray. And all communication
is explicit - but is often single sideded.  That scheme works well (also)
with distributed memory and is similar to MPI (Message Passing Interface).


Using

integer :: var[*]

one declares a coarray scalar variable. Accessing it as

  var = 5
  tmp = var

acts on the variable on the current 'image' (process). Using

  var[idx] = 5   ! Set 'var' in image 'idx' to '5'
  tmp = var[idx] ! Obtain value of 'var' on image 'idx'

one accesses that variable on a remote image - except if 'idx' matches
the current image; in that case, it acts on the local variable.


In one segment (which ends at explicit or implicit synchronizations,
e.g. using SYNC ALL): If the value of a is changed - either locally
or by a remote image - then only that image is permitted to 'read' the
value (or to change it as that would be another possibility for race
conditions).


Very often, coarray variables are arrays and heavily accessed as local
varable in hot loops. But on the other hand, the value of the variable
can be changed by external means - up to having hardware support to
write into the memory from another computer.


Thus, there two cases were the local view might fail:

 (a) when the variable has been changed in a previous segment by a
     remote process, e.g.
       var = 5   ! assue 'var' is a coarray
       sync all  ! end of segment 
       !         ! var is changed by a remote image
       sync all  ! end of segment 
       tmp = var
     where "var" might or might not have the value 5. Or more fancy
     without locks and global barriers:
       type(event_type) :: var_is_set[*]
       var = 5
       event post(have_set_var[remote_idx])
         ! Remote waits for event, sets out 'var' to 42 and
         ! then posts an event that it is set
       event wait(have_set_var)
       tmp = var
    where one tells the remote process that "var = 5" has been set
    and waits until it has set the local variable "var".


 (b) when the variable is changed on the same image by two means, e.g.
       var = 5
       var[this_image()] = 42
       tmp = var


The communication maps to function calls like:
  __gfortran_caf_send(var_token, idx, 5)    // var[idx] = 5
  __gfortran_caf_get (tmp, var_token, idx)  // tmp = var[idx]
[Real commands, see
https://gcc.gnu.org/onlinedocs/gfortran/index.html#toc_Coarray-Programming]


It is assumed that the library called takes care of the hardware part,
i.e. locking, mfence etc. - such that in the compiler itself, one only
has to deal with restricting compiler optimizations to the places where
it is permitted.


A coarray comes into existance via:

  var = _gfortran_caf_register (size, &var_token);

similar to malloc - and var_token identifies the coarray; what the
content is, is left to the library. - For the good or worse,
_gfortran_caf_register doesn't tell the compiler that it has clobbered
"var" as in a way the pointer address escaped. Nor is "var" marked as
volatile.


Coming back to item (a): After a segment has ended (SYNC ALL, EVENT WAIT
or similar), the compiler can no longer assume that coarray variables
have the same value. Those variables can be hidden as in
    call foo()
    sync all
    call foo()
where 'foo' doesn't know about the 'sync all' and the caller doesn't know
whether 'foo' has accesses to coarrays (and whether it accesses them). I
think
  __asm__ __volatile__ ("":::"memory");
should be sufficient in that case, assuming that the communication library
takes are of making all changes available (mfence, __sync_synchronize or
whatever).


The other question is item (b): Here, one has an alias problem within the
segment - but the alias only rarely happens. Code of the form
  var[idx] = ...
usually does not access the local memory; it happens only in corner cases
like:
  do i = 1, num_images
    var[i] = 5
  end do
which sets 'var' to 5 on all images - including the local image. This will
produce code like (full example below):

  *var = 5
  _gfortran_caf_send(var_token, idx, 42);
  tmp = *var

where the frontend knows that _gfortran_caf_send might modify 'var' iff idx
matches the local image.  The question is only how to tell this the compiler.

Using __asm__ __volatile__ ("":::"memory"); seems to be appropriate, even
though something more tailored would be nice, something like:
  __asm__ __volatile__ ("": : "+rm" (*var) : "")
but I am not sure whether that's valid and works - additionally, it assumes
that one reads "*var" (i.e. that it is defined before), which is not required.
(No idea whether -Wundefined would trigger.)

I there some better way to express this?

Cheers,

Tobias

PS: Full example:

integer :: tmp
integer :: var[*]

var = 5
var[this_image()] = 42
tmp = var
end

This will generate a static constructor, run at program startup:

_caf_init.1 ()
{                           //  v-- allocate 4 byte
  var = _gfortran_caf_register (4, 0, &caf_token.0, 0B, 0B, 0);
}

and the (main) program code (slightly trimmed):

  static void * restrict caf_token.0;
  static integer(kind=4) * restrict var;
  void _caf_init.1 (void);

  *var = 4;
  
  desc.3.data = 42;
  _gfortran_caf_send (caf_token.0, 0B /* offset */ var,
                      _gfortran_caf_this_image (0), &desc.2, 0B, &desc.3, 4, 4, 0);
  __asm__ __volatile__("":::"memory");  // new
  tmp = *var;

The problem is that in that case the compiler does not know that
"_gfortran_caf_send (caf_token.0," can modify "*var".