[committed, amdgcn] Update OpenACC testcases for amdgcn

Thomas Schwinge thomas@codesourcery.com
Tue Jun 8 09:51:24 GMT 2021


Hi!

On 2020-01-20T16:53:49+0000, Andrew Stubbs <ams@codesourcery.com> wrote:
> I've committed this testsuite-only patch to fix some test cases that
> need GCN-specific settings in order to pass.

>       * testsuite/libgomp.oacc-c-c++-common/parallel-dims.c
>       (acc_gang): Recognise acc_device_radeon.
>       (acc_worker): Likewise.
>       (acc_vector): Likewise.

> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/parallel-dims.c
> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/parallel-dims.c
> @@ -14,7 +14,8 @@ static unsigned int __attribute__ ((optimize ("O2"))) acc_gang ()
>  {
>    if (acc_on_device ((int) acc_device_host))
>      return 0;
> -  else if (acc_on_device ((int) acc_device_nvidia))
> +  else if (acc_on_device ((int) acc_device_nvidia)
> +        || acc_on_device ((int) acc_device_radeon))
>      return __builtin_goacc_parlevel_id (GOMP_DIM_GANG);
>    else
>      __builtin_abort ();
> @@ -25,7 +26,8 @@ static unsigned int __attribute__ ((optimize ("O2"))) acc_worker ()
>  {
>    if (acc_on_device ((int) acc_device_host))
>      return 0;
> -  else if (acc_on_device ((int) acc_device_nvidia))
> +  else if (acc_on_device ((int) acc_device_nvidia)
> +        || acc_on_device ((int) acc_device_radeon))
>      return __builtin_goacc_parlevel_id (GOMP_DIM_WORKER);
>    else
>      __builtin_abort ();
> @@ -36,7 +38,8 @@ static unsigned int __attribute__ ((optimize ("O2"))) acc_vector ()
>  {
>    if (acc_on_device ((int) acc_device_host))
>      return 0;
> -  else if (acc_on_device ((int) acc_device_nvidia))
> +  else if (acc_on_device ((int) acc_device_nvidia)
> +        || acc_on_device ((int) acc_device_radeon))
>      return __builtin_goacc_parlevel_id (GOMP_DIM_VECTOR);
>    else
>      __builtin_abort ();

Similar changes are necessary for
'libgomp.oacc-fortran/parallel-dims.f90', too -- and actually all that
can be simplified; pushed "Fix 'libgomp.oacc-fortran/parallel-dims.f90'
for 'acc_device_radeon'" to master branch in commit
32099c0d24adb93a031e0301ffd77b065b6f5472, see attached.

>       (main): Set expectations for amdgcn.

> @@ -282,6 +285,12 @@ int main ()
>         /* The GCC nvptx back end enforces num_workers (32).  */
>         workers_actual = 32;
>       }
> +      else if (acc_on_device (acc_device_radeon))
> +     {
> +       /* The GCC GCN back end is limited to num_workers (16).
> +          Temporarily set this to 1 until multiple workers are permitted. */
> +       workers_actual = 1; // 16;
> +     }

ACK; working on that.

> @@ -328,6 +337,11 @@ int main ()
>         /* We're actually executing with num_workers (32).  */
>         /* workers_actual = 32; */
>       }
> +      else if (acc_on_device (acc_device_radeon))
> +     {
> +       /* The GCC GCN back end is limited to num_workers (16).  */
> +       workers_actual = 16;
> +     }

That's surprising however, that here we're not similarly having to set
'workers_actual = 1' -- find the explanation in a forthcoming email.  ;-)

> [...]


Grüße
 Thomas


-----------------
Mentor Graphics (Deutschland) GmbH, Arnulfstrasse 201, 80634 München Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Frank Thürauf
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Fix-libgomp.oacc-fortran-parallel-dims.f90-for-acc_d.patch
Type: text/x-diff
Size: 4438 bytes
Desc: not available
URL: <https://gcc.gnu.org/pipermail/gcc-patches/attachments/20210608/ea8620f5/attachment-0001.bin>


More information about the Gcc-patches mailing list