[patch,gomp-4_0-branch] openacc parallel reduction part 1

[Thomas Schwinge]

Hi Cesar!

On Sun, 6 Jul 2014 16:10:56 -0700, Cesar Philippidis <cesar_philippidis@mentor.com> wrote:
> This patch is the first step to enabling parallel reductions in openacc.

Thanks!

> As mentioned earlier, this patch isn't complete yet. For starters, parts
> of it depends on our internal ptx backend. I've temporarily remapped the
> ptx dependencies to their openmp equivalent, but without a proper
> openacc runtime this infrastructure won't do much.

For the curious: we're working on preparing our implementation of the
OpenACC Runtime Library for upstream submission; if only the weeks had
more days...

> Thomas, is this patch OK for gomp-4_0-branch?

I still :-( haven't managed to allocate the time for a proper review, but
given this doesn't regress any existing test cases, it's fine to commit,
and then we can take it from there.

A few minor comments:

> 2014-07-06  Cesar Philippidis  <cesar@codesourcery.com>
> 	    Thomas Schwinge  <thomas@codesourcery.com>

By the way, on gomp-4_0-branch, ChangeLog snippets go into the respective
ChangeLog.gomp files.

> --- a/gcc/c/c-parser.c
> +++ b/gcc/c/c-parser.c
> @@ -11706,7 +11710,8 @@ c_parser_oacc_kernels (location_t loc, c_parser *parser, char *p_name)
>  */
>  
>  #define OACC_LOOP_CLAUSE_MASK						\
> -	(OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_NONE)
> +	( (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_COLLAPSE)		\

Not yet.  ;-)

> +	| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_REDUCTION))

> --- a/gcc/fortran/types.def
> +++ b/gcc/fortran/types.def
> @@ -86,6 +86,7 @@ DEF_FUNCTION_TYPE_1 (BT_FN_UINT_UINT, BT_UINT, BT_UINT)
>  DEF_FUNCTION_TYPE_1 (BT_FN_PTR_PTR, BT_PTR, BT_PTR)
>  DEF_FUNCTION_TYPE_1 (BT_FN_VOID_INT, BT_VOID, BT_INT)
>  DEF_FUNCTION_TYPE_1 (BT_FN_BOOL_INT, BT_BOOL, BT_INT)
> +DEF_FUNCTION_TYPE_1 (BT_FN_INT_INT, BT_INT, BT_INT)

That one's not actually needed, because...

> --- a/gcc/omp-builtins.def
> +++ b/gcc/omp-builtins.def
> @@ -236,3 +236,6 @@ DEF_GOMP_BUILTIN (BUILT_IN_GOMP_TARGET_UPDATE, "GOMP_target_update",
>  		  BT_FN_VOID_INT_PTR_SIZE_PTR_PTR_PTR, ATTR_NOTHROW_LIST)
>  DEF_GOMP_BUILTIN (BUILT_IN_GOMP_TEAMS, "GOMP_teams",
>  		  BT_FN_VOID_UINT_UINT, ATTR_NOTHROW_LIST)
> +
> +DEF_GOMP_BUILTIN (BUILT_IN_OMP_SET_NUM_THREADS, "omp_set_num_threads",
> +		  BT_FN_INT_INT, ATTR_CONST_NOTHROW_LEAF_LIST)

... it's actually »void omp_set_num_threads (int)«, so BT_FN_VOID_INT.
As this is only temporary code, please add a FIXME comment here.  Hmm,
and I wonder, given this is using DEF_*GOMP*_BUILTIN, does this actually
do the right thing if -openmp is not specified?


Grüße,
 Thomas
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 472 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20140707/b3588978/attachment.sig>