[gomp4] partially enable GOMP_MAP_FIRSTPRIVATE_POINTER in gfortran

Thomas Schwinge thomas@codesourcery.com
Mon Jan 30 15:26:00 GMT 2017


Hi Cesar!  (It's me, again!)  ;-)

On Fri, 27 Jan 2017 09:13:06 -0800, Cesar Philippidis <cesar@codesourcery.com> wrote:
> This patch partially enables GOMP_MAP_FIRSTPRIVATE_POINTER in gfortran.
> gfortran still falls back to GOMP_MAP_POINTER for arrays with
> descriptors and derived types. The limitation on derived types is there
> because we don't have much test coverage for it, and this patch series
> was more exploratory for performance enhancements.

Now that you still freshly remember it, please file an issue so that
we'll take care of that later.

> With that in mind,
> there are a couple of shortcomings with this patch.
> 
>  1) Dummy reduction variables fallback to GOMP_MAP_POINTER because of a
>     pointer dereferencing bug.

Please also file an issue for that.


>     The state of debugging such problems on
>     PTX targets leaves something to be desired, especially since print
>     isn't working on nvptx targets currently.

If the following is what you mean, then that's working for me:

    $ cat < ../printf.c
    int main(int argc, char *argv[])
    {
    #pragma acc parallel copyin(argv[0][0:__builtin_strlen(argv[0]) + 1])
      {
        __builtin_printf("Offloaded from %s.\n", argv[0]);
      }
    
      return 0;
    }
    $ build-gcc/gcc/xgcc [...] -Wall -Wextra -g ../printf.c -fopenacc -O2
    $ GOMP_DEBUG=1 ./a.out
    [...]
      nvptx_exec: kernel main$_omp_fn$0: launch gangs=1, workers=1, vectors=32
    Offloaded from ./a.out.
      nvptx_exec: kernel main$_omp_fn$0: finished
    GOMP_offload_unregister_ver (1, 0x400c20, 5, 0x401560)
    GOMP_offload_unregister_ver (0, 0x400c20, 6, 0x602050)

Again, please file an issue as appropriate.  ;-)


>  2) Apparently, firstprivate pointers negatively affects the alias
>     analysis used by ACC KERNELS and parloops, so a couple of more
>     execution tests fail to generate offloaded code.
> 
> I plan to resolve issue 1) in a follow up patch later on (but maybe not
> in the immediate future). Regarding 2), ACC KERNELS are eventually going
> to need a significant rework, but that's not going to happen in the near
> future either. I've been pushing to get the performance of ACC PARALLEL
> regions on par to other OpenACC compilers first, and hopefully that
> won't be too far way.

Hmm, hmm.


> With this patch, I'm observing an approximate 0.6s reduction in
> CloverLeaf's original 0.9s execution time (it takes approximate 0.9s
> after the GOMP_MAP_FIRSTPRIVATE_INT and GOMP_MAP_TO_PSET patches), to
> yield a final execution time somewhere in the neighborhood of 0.3s.
> That's about a one second savings from the unpatched version of GCC.

Yay!  \o/


> This patch has been committed to gomp-4_0-branch.

(Not reviewed in detail.)

> --- a/gcc/fortran/trans-openmp.c
> +++ b/gcc/fortran/trans-openmp.c
> @@ -2005,9 +2005,12 @@ gfc_trans_omp_clauses_1 (stmtblock_t *block, gfc_omp_clauses *clauses,
>  					(TREE_TYPE (TREE_TYPE (decl)))))
>  		    {
>  		      tree orig_decl = decl;
> +		      enum gomp_map_kind gmk = GOMP_MAP_FIRSTPRIVATE_POINTER;
> +		      if (n->u.map_op == OMP_MAP_FORCE_DEVICEPTR)
> +			gmk = GOMP_MAP_POINTER;

Curious, why is "deviceptr" different?

>  		      node4 = build_omp_clause (input_location,
>  						OMP_CLAUSE_MAP);
> -		      OMP_CLAUSE_SET_MAP_KIND (node4, GOMP_MAP_POINTER);
> +		      OMP_CLAUSE_SET_MAP_KIND (node4, gmk);
>  		      OMP_CLAUSE_DECL (node4) = decl;
>  		      OMP_CLAUSE_SIZE (node4) = size_int (0);
>  		      decl = build_fold_indirect_ref (decl);

> --- a/gcc/gimplify.c
> +++ b/gcc/gimplify.c

> @@ -6605,11 +6636,10 @@ gimplify_scan_omp_clauses (tree *list_p, gimple_seq *pre_p,
>    ctx = new_omp_context (region_type);
>    ctx->clauses = *list_p;
>    outer_ctx = ctx->outer_context;
> -  if (code == OMP_TARGET && !lang_GNU_Fortran ())
> +  if (code == OMP_TARGET && !(lang_GNU_Fortran () && !(region_type & ORT_ACC)))
>      {
> -      ctx->target_map_pointers_as_0len_arrays = true;
> -      /* FIXME: For Fortran we want to set this too, when
> -	 the Fortran FE is updated to OpenMP 4.5.  */
> +      if (!lang_GNU_Fortran () || region_type & ORT_ACC)
> +	ctx->target_map_pointers_as_0len_arrays = true;
>        ctx->target_map_scalars_firstprivate = true;
>      }

I guess the Fortran OpenMP comment should stay?  And, isn't that logic a
bit complicated; could simplify this as follows, unless I'm confused?

--- gcc/gimplify.c
+++ gcc/gimplify.c
@@ -6636,10 +6636,11 @@ gimplify_scan_omp_clauses (tree *list_p, gimple_seq *pre_p,
   ctx = new_omp_context (region_type);
   ctx->clauses = *list_p;
   outer_ctx = ctx->outer_context;
-  if (code == OMP_TARGET && !(lang_GNU_Fortran () && !(region_type & ORT_ACC)))
+  /* FIXME: For Fortran OpenMP we want to set this too, when
+     the Fortran FE is updated to OpenMP 4.5.  */
+  if (code == OMP_TARGET && (!lang_GNU_Fortran () || (region_type & ORT_ACC)))
     {
-      if (!lang_GNU_Fortran () || region_type & ORT_ACC)
-	ctx->target_map_pointers_as_0len_arrays = true;
+      ctx->target_map_pointers_as_0len_arrays = true;
       ctx->target_map_scalars_firstprivate = true;
     }
   if (!lang_GNU_Fortran ())

> --- a/gcc/testsuite/gfortran.dg/goacc/kernels-loop-n.f95
> +++ b/gcc/testsuite/gfortran.dg/goacc/kernels-loop-n.f95
> @@ -37,4 +37,6 @@ end module test
>  ! Check that the loop has been split off into a function.
>  ! { dg-final { scan-tree-dump-times "(?n);; Function __test_MOD_foo._omp_fn.0 " 1 "optimized" } }
>  
> -! { dg-final { scan-tree-dump-times "(?n)oacc function \\(0," 1 "parloops1" } }
> +! This failure was introduced with the GOMP_MAP_POINTER ->
> +! GOMP_MAP_FIRSTPRIVATE_POINTER conversion.
> +! { dg-final { scan-tree-dump-times "(?n)oacc function \\(0," 1 "parloops1" { xfail *-*-* } } }

Hmm, hmm.

> --- a/libgomp/testsuite/libgomp.oacc-fortran/deviceptr-1.f90
> +++ b/libgomp/testsuite/libgomp.oacc-fortran/deviceptr-1.f90
> @@ -3,6 +3,7 @@
>  ! the deviceptr variable is implied.
>  
>  ! { dg-do run }
> +! { dg-additional-options "-foffload-force" }
>  
>  subroutine subr1 (a, b)
>    implicit none

This is also an OpenACC kernels issue.

> --- a/libgomp/testsuite/libgomp.oacc-fortran/non-scalar-data.f90
> +++ b/libgomp/testsuite/libgomp.oacc-fortran/non-scalar-data.f90
> @@ -2,6 +2,7 @@
>  ! offloaded regions are properly mapped using present_or_copy.
>  
>  ! { dg-do run }
> +! { dg-additional-options "-foffload-force" }
>  
>  program main
>    implicit none

Likweise.

I do agree that our OpenACC kernels implementation leaves a lot to be
desired, but that we're now also regressing such very simple cases, is a
bit unfortunate.  Have you already made an attempt at figuring out what's
going wrong?


Another OpenMP regression:

    PASS: libgomp.fortran/target2.f90   -O0  (test for excess errors)
    [-PASS:-]{+FAIL:+} libgomp.fortran/target2.f90   -O0  execution test
    PASS: libgomp.fortran/target2.f90   -O1  (test for excess errors)
    [-PASS:-]{+FAIL:+} libgomp.fortran/target2.f90   -O1  execution test
    PASS: libgomp.fortran/target2.f90   -O2  (test for excess errors)
    [-PASS:-]{+FAIL:+} libgomp.fortran/target2.f90   -O2  execution test
    PASS: libgomp.fortran/target2.f90   -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions  (test for excess errors)
    [-PASS:-]{+FAIL:+} libgomp.fortran/target2.f90   -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions  execution test
    PASS: libgomp.fortran/target2.f90   -O3 -g  (test for excess errors)
    [-PASS:-]{+FAIL:+} libgomp.fortran/target2.f90   -O3 -g  execution test
    PASS: libgomp.fortran/target2.f90   -Os  (test for excess errors)
    [-PASS:-]{+FAIL:+} libgomp.fortran/target2.f90   -Os  execution test

That is:

    offload error: process on the device 0 unexpectedly exited with code 0

..., which, as far as I remember, basically means "SIGSEGV" in the Intel
MIC (emulated) offloaded code.

Porting this gomp-4_0-branch r244987 "Partially enable
GOMP_MAP_FIRSTPRIVATE_POINTER in gfortran." to trunk (see attached, if
you want to experiment with that), I can reproduce some (maybe even the
same?) issue with OpenMP nvptx offloading: "libgomp: cuCtxSynchronize
error: an illegal memory access was encountered".  Do you have an idea
which of your changes might cause that?


Grüße
 Thomas


-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Partially-enable-GOMP_MAP_FIRSTPRIVATE_POINTER-in-gf.patch
Type: text/x-diff
Size: 6929 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20170130/f0cccd6c/attachment.bin>


More information about the Fortran mailing list