Re: Fix up 'libgomp.oacc-fortran/print-1.f90' GCN offloading compilation [PR104717] (was: [PATCH] fortran: Fix up gfc_trans_oacc_construct [PR104717])

[Thomas Schwinge <thomas@codesourcery.com>]

Hi!

On 2022-04-26T19:25:31+0200, I wrote:
> On 2022-04-25T23:19:26+0200, I wrote:
>> On 2022-04-20T19:06:17+0200, Jakub Jelinek <jakub@redhat.com> wrote:
>>> So that move_sese_region_to_fn works properly, OpenMP/OpenACC constructs
>>> for which that function is invoked need an extra artificial BIND_EXPR
>>> around their body so that we move all variables of the bodies.
>>>
>>> The C/C++ FEs do that both for OpenMP constructs like OMP_PARALLEL, OMP_TASK
>>> or OMP_TARGET and for OpenACC constructs that behave similarly to
>>> OMP_TARGET, but the Fortran FE only does that for OpenMP constructs.
>>>
>>> The following patch does that for OpenACC constructs too.
>>> This fixes ICE on the attached testcase.
>>
>> ACK, thanks.
>
>>> Unfortunately, it also regresses
>>> FAIL: gfortran.dg/goacc/privatization-1-compute-loop.f90   -O  (test for excess errors)
>>> FAIL: libgomp.oacc-fortran/privatized-ref-2.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O0  (test for excess errors)
>>> FAIL: libgomp.oacc-fortran/privatized-ref-2.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O1  (test for excess errors)
>>> FAIL: libgomp.oacc-fortran/privatized-ref-2.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O2  (test for excess errors)
>>> FAIL: libgomp.oacc-fortran/privatized-ref-2.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions  (test for excess errors)
>>> FAIL: libgomp.oacc-fortran/privatized-ref-2.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O3 -g  (test for excess errors)
>>> FAIL: libgomp.oacc-fortran/privatized-ref-2.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -Os  (test for excess errors)
>>> Those emits emit tons of various messages and now there are some extra ones,
>>
>> I've fixed these up.
>
> One more issue became apparent, where the code changes pushed actually do
> lead to a GCN offloading compilation failure:
>
>     [...]/libgomp.oacc-fortran/print-1.f90: In function ‘MAIN__._omp_fn.0’:
>     [...]/libgomp.oacc-fortran/print-1.f90:13:14: error: 512 bytes of gang-private data-share memory exhausted (increase with ‘-mgang-private-size=560’, for example)
>        13 | !$acc parallel
>           |              ^
>
> In my configuration, I may indeed fix GCN offloading compilation with
> '-foffload-options=amdgcn-amdhsa=-mgang-private-size=560', but I don't
> think that's generally correct/sufficient, so in the the attached
> "Fix up 'libgomp.oacc-fortran/print-1.f90' GCN offloading compilation
> [PR104717]", I instead "raise '-mgang-private-size' to an arbitrary high
> value".  This avoids having to route the actual 'sizeof' from GCC build
> down to the test suite harness (which ought to be doable, but
> non-trivial).  OK to push that:
>
>     +! For GCN offloading compilation, when gang-privatizing 'dt_parm.N'
>     +! (see below), we run into an 'gang-private data-share memory exhausted'
>     +! error: the default '-mgang-private-size' is too small.  Per
>     +! 'gcc/fortran/trans-io.cc'/'libgfortran/io/io.h', that one is
>     +! 'struct st_parameter_dt', which indeed is rather big.  Instead of
>     +! working out its exact size (which may vary per GCC configuration),
>     +! raise '-mgang-private-size' to an arbitrary high value.
>     +! { dg-additional-options "-foffload-options=amdgcn-amdhsa=-mgang-private-size=13579" { target openacc_radeon_accel_selected } }
>
> ... to master branch? (This doubles the use/testing of the
> '-mgang-private-size' option!)  ;-)

Eh.  That only works with the default GCN multilib '-march=fiji', testing
on gfx803 amdfury2 system.  For all of '-march=gfx900' (amdnano2),
'-march=gfx906' (amd_ryzen3), '-march=gfx908' (amd-instinct1), I get:

    libgomp: GCN fatal error: Asynchronous queue error
    Runtime message: HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION: The agent attempted to access memory beyond the largest legal address.

..., and I still get that if lowering the allocation to the minimum,
'-foffload-options=amdgcn-amdhsa=-mgang-private-size=560'.

This is a really simple OpenACC 'parallel' construct:

    !$acc parallel
      write (0, '("The answer is ", I2)') var
    !$acc end parallel

..., which ought to launch a 1-gang x 1-worker x 1-vector GPU kernel, so
I'd assume '-mgang-private-size=560' (or '-mgang-private-size=13579' in
fact) is not a problem?

Help?


Grüße
 Thomas


> We've currently not been doing OpenACC privatization scanning in
> 'libgomp.oacc-fortran/print-1.f90', which I've now added, to help
> document the issue; no need to review that.
>
> Of course, the issue could alternatively be fixed by adding more logic to
> the GCN back end to auto-scale the allocation, or be fixed by adding more
> logic to the compiler to avoid gang-privatizing varibales such as
> 'dt_parm.N' in such cases, but that's not something I'm going to look
> into at this point.
>
> Or, of course, be avoided by re-writing the test case to not require
> gang-privatizing 'dt_parm.N', but the test case is correct as it is.
>
>
> Grüße
>  Thomas
>
>
>>      PR fortran/104717
>>      gcc/fortran/
>>      * trans-openmp.cc (gfc_trans_oacc_construct): Wrap construct body
>>      in an extra BIND_EXPR.
>
>> --- a/gcc/fortran/trans-openmp.cc
>> +++ b/gcc/fortran/trans-openmp.cc
>> @@ -4444,7 +4444,9 @@ gfc_trans_oacc_construct (gfc_code *code)
>>    gfc_start_block (&block);
>>    oacc_clauses = gfc_trans_omp_clauses (&block, code->ext.omp_clauses,
>>                                      code->loc, false, true);
>> +  pushlevel ();
>>    stmt = gfc_trans_omp_code (code->block->next, true);
>> +  stmt = build3_v (BIND_EXPR, NULL, stmt, poplevel (1, 0));
>>    stmt = build2_loc (gfc_get_location (&code->loc), construct_code,
>>                   void_type_node, stmt, oacc_clauses);
>>    gfc_add_expr_to_block (&block, stmt);
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955