The chunk size for SIMD loops should be right for the current device; too big
allocates too much memory, too small is inefficient. Getting it wrong doesn't
actually break anything though.
This patch attempts to choose the optimal setting based on the context. Both
host-fallback and device will get the same chunk size, but device performance
is the most important in this case.
gcc/ChangeLog:
* omp-expand.cc (is_in_offload_region): New function.
(omp_adjust_chunk_size): Add pass-through "offload" parameter.
(get_ws_args_for): Likewise.
(determine_parallel_type): Use is_in_offload_region to adjust call to
get_ws_args_for.
(expand_omp_for_generic): Likewise.
(expand_omp_for_static_chunk): Likewise.
---
gcc/omp-expand.cc | 36 ++++++++++++++++++++++++++++--------
1 file changed, 28 insertions(+), 8 deletions(-)
diff --git a/gcc/omp-expand.cc b/gcc/omp-expand.cc
index 907fd46a5b2..b0f9d375b6c 100644
--- a/gcc/omp-expand.cc
+++ b/gcc/omp-expand.cc
@@ -127,6 +127,23 @@ is_combined_parallel (struct omp_region *region)
return region->is_combined_parallel;
}
+/* Return true is REGION is or is contained within an offload region. */
+
+static bool
+is_in_offload_region (struct omp_region *region)
+{
+ gimple *entry_stmt = last_nondebug_stmt (region->entry);
+ if (is_gimple_omp (entry_stmt)
+ && is_gimple_omp_offloaded (entry_stmt))
+ return true;
+ else if (region->outer)
+ return is_in_offload_region (region->outer);
+ else
+ return (lookup_attribute ("omp declare target",
+ DECL_ATTRIBUTES (current_function_decl))
+ != NULL);
+}
+
/* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB
is the immediate dominator of PAR_ENTRY_BB, return true if there
are no data dependencies that would prevent expanding the parallel
@@ -207,12 +224,12 @@ workshare_safe_to_combine_p (basic_block ws_entry_bb)
presence (SIMD_SCHEDULE). */
static tree
-omp_adjust_chunk_size (tree chunk_size, bool simd_schedule)
+omp_adjust_chunk_size (tree chunk_size, bool simd_schedule, bool offload)
{
if (!simd_schedule || integer_zerop (chunk_size))
return chunk_size;
- poly_uint64 vf = omp_max_vf (false);
+ poly_uint64 vf = omp_max_vf (offload);
if (known_eq (vf, 1U))
return chunk_size;
@@ -228,7 +245,7 @@ omp_adjust_chunk_size (tree chunk_size, bool simd_schedule)
expanded. */
static vec<tree, va_gc> *
-get_ws_args_for (gimple *par_stmt, gimple *ws_stmt)
+get_ws_args_for (gimple *par_stmt, gimple *ws_stmt, bool offload)
{
tree t;
location_t loc = gimple_location (ws_stmt);
@@ -270,7 +287,7 @@ get_ws_args_for (gimple *par_stmt, gimple *ws_stmt)
if (fd.chunk_size)
{
t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size);
- t = omp_adjust_chunk_size (t, fd.simd_schedule);
+ t = omp_adjust_chunk_size (t, fd.simd_schedule, offload);
ws_args->quick_push (t);
}
@@ -366,7 +383,8 @@ determine_parallel_type (struct omp_region *region)
region->is_combined_parallel = true;
region->inner->is_combined_parallel = true;
- region->ws_args = get_ws_args_for (par_stmt, ws_stmt);
+ region->ws_args = get_ws_args_for (par_stmt, ws_stmt,
+ is_in_offload_region (region));
}
}
@@ -3929,6 +3947,7 @@ expand_omp_for_generic (struct omp_region *region,
tree *counts = NULL;
int i;
bool ordered_lastprivate = false;
+ bool offload = is_in_offload_region (region);
gcc_assert (!broken_loop || !in_combined_parallel);
gcc_assert (fd->iter_type == long_integer_type_node
@@ -4196,7 +4215,7 @@ expand_omp_for_generic (struct omp_region *region,
if (fd->chunk_size)
{
t = fold_convert (fd->iter_type, fd->chunk_size);
- t = omp_adjust_chunk_size (t, fd->simd_schedule);
+ t = omp_adjust_chunk_size (t, fd->simd_schedule, offload);
if (sched_arg)
{
if (fd->ordered)
@@ -4240,7 +4259,7 @@ expand_omp_for_generic (struct omp_region *region,
{
tree bfn_decl = builtin_decl_explicit (start_fn);
t = fold_convert (fd->iter_type, fd->chunk_size);
- t = omp_adjust_chunk_size (t, fd->simd_schedule);
+ t = omp_adjust_chunk_size (t, fd->simd_schedule, offload);
if (sched_arg)
t = build_call_expr (bfn_decl, 10, t5, t0, t1, t2, sched_arg,
t, t3, t4, reductions, mem);
@@ -5937,7 +5956,8 @@ expand_omp_for_static_chunk (struct omp_region *region,
step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step),
true, NULL_TREE, true, GSI_SAME_STMT);
tree chunk_size = fold_convert (itype, fd->chunk_size);
- chunk_size = omp_adjust_chunk_size (chunk_size, fd->simd_schedule);
+ chunk_size = omp_adjust_chunk_size (chunk_size, fd->simd_schedule,
+ is_in_offload_region (region));
chunk_size
= force_gimple_operand_gsi (&gsi, chunk_size, true, NULL_TREE, true,
GSI_SAME_STMT);
--
2.46.0
