diff mbox

[gomp4] auto partitioning tweak

Message ID e8b338bc-f008-f5a7-5140-c55a1c37872f@acm.org
State New
Headers show

Commit Message

Nathan Sidwell Oct. 4, 2016, 1:14 p.m. UTC 
We were determining in the host compiler as to whethe a particular loop had no 
explicit partitioning.  That'll break down if device_type comes into play.  My 
tiling patch exposed this difficulty, as tiling itself ends up essentially being 
two nested loops, one of which could be explicitly partitioned and the other 
not.  So addressing this problem now.

Hopefully this is the last tweak before the tile patch itself.

nathan
diff mbox

Patch

2016-10-04  Nathan Sidwell  <nathan@codesourcery.com>

	* omp-low.c (lower_oacc_head_mark): Don't determine default auto
	here ...
	(oacc_loop_fixed_partitions): ... do it here instead.

Index: omp-low.c
===================================================================
--- omp-low.c	(revision 240731)
+++ omp-low.c	(working copy)
@@ -6394,14 +6394,13 @@  lower_oacc_head_mark (location_t loc, tr
   if (!tgt || is_oacc_parallel (tgt))
     tag |= OLF_INDEPENDENT;
 
-  /* A loop lacking SEQ, GANG, WORKER and/or VECTOR is implicitly AUTO.  */
-  if (!(tag & (((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1) << OLF_DIM_BASE)
-	       | OLF_SEQ)))
-      tag |= OLF_AUTO;
-
-  /* Ensure at least one level, or 2 for AUTO partitioning  */
-  if (levels < 1 + ((tag & OLF_AUTO) != 0))
-    levels = 1 + ((tag & OLF_AUTO) != 0);
+  /* A loop lacking SEQ, GANG, WORKER and/or VECTOR could be AUTO  */
+  bool maybe_auto = !(tag & (((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1)
+			      << OLF_DIM_BASE) | OLF_SEQ));
+
+  /* Ensure at least one level, or 2 for possible auto partitioning  */
+  if (levels < 1u + maybe_auto)
+    levels = 1u + maybe_auto;
 
   args.quick_push (build_int_cst (integer_type_node, levels));
   args.quick_push (build_int_cst (integer_type_node, tag));
@@ -19759,6 +19758,8 @@  oacc_loop_fixed_partitions (oacc_loop *l
       this_mask = ((loop->flags >> OLF_DIM_BASE)
 		   & (GOMP_DIM_MASK (GOMP_DIM_MAX) - 1));
 
+      bool maybe_auto = !seq_par && !this_mask;
+
       if ((this_mask != 0) + auto_par + seq_par > 1)
 	{
 	  if (noisy)
@@ -19766,7 +19767,7 @@  oacc_loop_fixed_partitions (oacc_loop *l
 		      seq_par
 		      ? "%<seq%> overrides other OpenACC loop specifiers"
 		      : "%<auto%> conflicts with other OpenACC loop specifiers");
-	  auto_par = false;
+	  maybe_auto = false;
 	  loop->flags &= ~OLF_AUTO;
 	  if (seq_par)
 	    {
@@ -19775,8 +19776,11 @@  oacc_loop_fixed_partitions (oacc_loop *l
 	      this_mask = 0;
 	    }
 	}
-      if (auto_par && (loop->flags & OLF_INDEPENDENT))
-	mask_all |= GOMP_DIM_MASK (GOMP_DIM_MAX);
+      if (maybe_auto && (loop->flags & OLF_INDEPENDENT))
+	{
+	  loop->flags |= OLF_AUTO;
+	  mask_all |= GOMP_DIM_MASK (GOMP_DIM_MAX);
+	}
     }
 
   if (this_mask & outer_mask)