diff mbox

OpenACC subarray data alignment in fortran

Message ID 56017292.9050605@codesourcery.com
State New
Headers show

Commit Message

Cesar Philippidis Sept. 22, 2015, 3:24 p.m. UTC 
In both OpenACC and OpenMP, each subarray has at least two data mappings
associated with them, one for the pointer and another for the data in
the array section (fortan also has a pset mapping). One problem I
observed in fortran is that array section data is casted to char *.
Consequently, when lower_omp_target assigns alignment for the subarray
data, it does so incorrectly. This is a problem on nvptx if you have a
data clause such as

  integer foo
  real*8 bar (100)

  !$acc data copy (foo, bar(1:100))

Here, the data associated with bar could get aligned on a 4 byte
boundary instead of 8 byte. That causes problems on nvptx targets.

My fix for this is to prevent the fortran front end from casting the
data pointers to char *. I only prevented casting on the code which
handles OMP_CLAUSE_MAP. The subarrays associated with OMP_CLAUSE_SHARED
also get casted to char *, but I left those as-is because I'm not that
familiar with how non-OpenMP target regions get lowered.

Is this patch OK for trunk?

Thanks,
Cesar

Comments

Cesar Philippidis Sept. 29, 2015, 1:53 p.m. UTC | #1 
Ping.

In the meantime, I'll apply this patch to gomp-4_0-branch.

Cesar

On 09/22/2015 08:24 AM, Cesar Philippidis wrote:
> In both OpenACC and OpenMP, each subarray has at least two data mappings
> associated with them, one for the pointer and another for the data in
> the array section (fortan also has a pset mapping). One problem I
> observed in fortran is that array section data is casted to char *.
> Consequently, when lower_omp_target assigns alignment for the subarray
> data, it does so incorrectly. This is a problem on nvptx if you have a
> data clause such as
> 
>   integer foo
>   real*8 bar (100)
> 
>   !$acc data copy (foo, bar(1:100))
> 
> Here, the data associated with bar could get aligned on a 4 byte
> boundary instead of 8 byte. That causes problems on nvptx targets.
> 
> My fix for this is to prevent the fortran front end from casting the
> data pointers to char *. I only prevented casting on the code which
> handles OMP_CLAUSE_MAP. The subarrays associated with OMP_CLAUSE_SHARED
> also get casted to char *, but I left those as-is because I'm not that
> familiar with how non-OpenMP target regions get lowered.
> 
> Is this patch OK for trunk?
> 
> Thanks,
> Cesar
>
diff mbox

Patch

2015-09-22  Cesar Philippidis  <cesar@codesourcery.com>

	gcc/
	* fortran/trans-openmp.c (gfc_omp_finish_clause): Don't cast ptr
	into a character pointer.
	(gfc_trans_omp_clauses_1): Likewise.

	libgomp/
	* testsuite/libgomp.oacc-fortran/data-alignment.f90: New test.

diff --git a/gcc/fortran/trans-openmp.c b/gcc/fortran/trans-openmp.c
index cd76f2a..8c1e897 100644
--- a/gcc/fortran/trans-openmp.c
+++ b/gcc/fortran/trans-openmp.c
@@ -1065,7 +1065,6 @@  gfc_omp_finish_clause (tree c, gimple_seq *pre_p)
       gfc_start_block (&block);
       tree type = TREE_TYPE (decl);
       tree ptr = gfc_conv_descriptor_data_get (decl);
-      ptr = fold_convert (build_pointer_type (char_type_node), ptr);
       ptr = build_fold_indirect_ref (ptr);
       OMP_CLAUSE_DECL (c) = ptr;
       c2 = build_omp_clause (input_location, OMP_CLAUSE_MAP);
@@ -1972,8 +1971,6 @@  gfc_trans_omp_clauses_1 (stmtblock_t *block, gfc_omp_clauses *clauses,
 		    {
 		      tree type = TREE_TYPE (decl);
 		      tree ptr = gfc_conv_descriptor_data_get (decl);
-		      ptr = fold_convert (build_pointer_type (char_type_node),
-					  ptr);
 		      ptr = build_fold_indirect_ref (ptr);
 		      OMP_CLAUSE_DECL (node) = ptr;
 		      node2 = build_omp_clause (input_location,
@@ -2066,8 +2063,6 @@  gfc_trans_omp_clauses_1 (stmtblock_t *block, gfc_omp_clauses *clauses,
 				       OMP_CLAUSE_SIZE (node), elemsz);
 		    }
 		  gfc_add_block_to_block (block, &se.post);
-		  ptr = fold_convert (build_pointer_type (char_type_node),
-				      ptr);
 		  OMP_CLAUSE_DECL (node) = build_fold_indirect_ref (ptr);
 
 		  if (POINTER_TYPE_P (TREE_TYPE (decl))
diff --git a/libgomp/testsuite/libgomp.oacc-fortran/data-alignment.f90 b/libgomp/testsuite/libgomp.oacc-fortran/data-alignment.f90
new file mode 100644
index 0000000..3c309c0
--- /dev/null
+++ b/libgomp/testsuite/libgomp.oacc-fortran/data-alignment.f90
@@ -0,0 +1,35 @@ 
+! Test if the array data associated with c is properly aligned
+! on the accelerator.  If it is not, this program will crash.
+
+! { dg-do run }
+
+integer function routine_align()
+  implicit none
+  integer, parameter :: n = 10000
+  real*8, dimension(:), allocatable :: c
+  integer :: i, idx
+
+  allocate (c(n))
+  routine_align = 0
+  c = 0.0
+
+  !$acc data copyin(idx) copy(c(1:n))
+
+  !$acc parallel vector_length(32)
+  !$acc loop vector
+  do i=1, n
+     c(i) = i
+  enddo
+  !$acc end parallel
+
+  !$acc end data
+end function routine_align
+
+
+! main driver
+program routine_align_main
+  implicit none
+  integer :: success
+  integer routine_align
+  success = routine_align()
+end program routine_align_main