Re: Enhance 'libgomp.c-c++-common/requires-4.c', 'libgomp.c-c++-common/requires-5.c' testing (was: [Patch][v4] OpenMP: Move omp requires checks to libgomp)

[Tobias Burnus <tobias@codesourcery.com>]

On 07.07.22 10:42, Thomas Schwinge wrote:
> In preparation for other changes:
...
> On 2022-06-29T16:33:02+0200, Tobias Burnus<tobias@codesourcery.com>  wrote:
>> +/* { dg-output "devices present but 'omp requires unified_address, unified_shared_memory, reverse_offload' cannot be fulfilled" } */
> (The latter diagnostic later got conditionalized by 'GOMP_DEBUG=1'.)
> OK to push the attached "Enhance 'libgomp.c-c++-common/requires-4.c',
> 'libgomp.c-c++-common/requires-5.c' testing"?
...
>       libgomp/
>       * testsuite/libgomp.c-c++-common/requires-4.c: Enhance testing.
>       * testsuite/libgomp.c-c++-common/requires-5.c: Likewise.
...
> --- a/libgomp/testsuite/libgomp.c-c++-common/requires-4.c
> +++ b/libgomp/testsuite/libgomp.c-c++-common/requires-4.c
> @@ -1,22 +1,29 @@
> -/* { dg-do link { target offloading_enabled } } */
>   /* { dg-additional-options "-flto" } */
>   /* { dg-additional-sources requires-4-aux.c } */
>
> -/* Check diagnostic by device-compiler's or host compiler's lto1.
> +/* Check no diagnostic by device-compiler's or host compiler's lto1.

I note that without ENABLE_OFFLOADING that there is never any lto1
diagnostic.

However, given that no diagnostic is expected, it also works for "!
offloading_enabled".

Thus, the change fine.

>      Other file uses: 'requires reverse_offload', but that's inactive as
>      there are no declare target directives, device constructs nor device routines  */
>
> +/* For actual offload execution, prints the following (only) if GOMP_DEBUG=1:
> +   "devices present but 'omp requires unified_address, unified_shared_memory, reverse_offload' cannot be fulfilled"
> +   and does host-fallback execution.  */

The latter is only true when also device code is produced – and a device
is available for that/those device types. I think that's what you imply
by "For actual offload execution", but it is a bit hidden.

Maybe s/For actual offload execution, prints/It may print/ is clearer?

In principle, it would be nice if we could test for the output, but
currently setting an env var for remote execution does not work, yet.
Cf. https://gcc.gnu.org/pipermail/gcc-patches/2022-July/597773.html -
When set, we could use offload_target_nvptx etc. (..._amdgcn, ..._any)
to test – as this guarantees that it is compiled for that device + the
device is available.
> +
>   #pragma omp requires unified_address,unified_shared_memory
>
> -int a[10];
> +int a[10] = { 0 };
>   extern void foo (void);
>
>   int
>   main (void)
>   {
> -  #pragma omp target
> +  #pragma omp target map(to: a)

Hmm, I wonder whether I like it or not. Without, there is an implicit
"map(tofrom:a)". On the other hand, OpenMP permits that – even with
unified-shared memory – the implementation my copy the data to the
device. (For instance, to permit faster access to "a".)

Thus, ...

> +  for (int i = 0; i < 10; i++)
> +    a[i] = i;
> +
>     for (int i = 0; i < 10; i++)
> -    a[i] = 0;
> +    if (a[i] != i)
> +      __builtin_abort ();
... this condition (back on the host) could also fail with USM. However,
given that to my knowledge no USM implementation actually copies the
data, I believe it is fine. (Disclaimer: I have not checked what OG12,
but I guess it also does not copy it.)
> --- a/libgomp/testsuite/libgomp.c-c++-common/requires-5.c
> +++ b/libgomp/testsuite/libgomp.c-c++-common/requires-5.c
> @@ -1,21 +1,25 @@
> -/* { dg-do run { target { offload_target_nvptx || offload_target_amdgcn } } } */
>   /* { dg-additional-sources requires-5-aux.c } */
>
> +/* For actual offload execution, prints the following (only) if GOMP_DEBUG=1:
> +   "devices present but 'omp requires unified_address, unified_shared_memory, reverse_offload' cannot be fulfilled"
> +   and does host-fallback execution.  */
> +
This wording is correct with the now-removed check – but if you remove
the offload_target..., it only "might" print it, depending, well, on the
conditions set by offload_target...
>   #pragma omp requires unified_shared_memory, unified_address, reverse_offload
>
> -int a[10];
> +int a[10] = { 0 };
>   extern void foo (void);
>
>   int
>   main (void)
>   {
> -  #pragma omp target
> +  #pragma omp target map(to: a)
> +  for (int i = 0; i < 10; i++)
> +    a[i] = i;
> +
>     for (int i = 0; i < 10; i++)
> -    a[i] = 0;
> +    if (a[i] != i)
> +      __builtin_abort ();
>
>     foo ();
>     return 0;
>   }

Thus: LGTM, if you update the GOMP_DEBUG=... wording, either using
"might" (etc.) or by being more explicit.

Once we have remote setenv, we probably want to add another testcase to
check for the GOMP_DEBUG=1, copying an existing one, adding dg-output
and restricting it to target offload_target_...

Tobias

-----------------
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