doc/tutorial.rst

@@ -1059,7 +1059,6 @@ earlier:
         acc_k = 0.0f;
       if (-1 + -16 * gid(0) + -1 * lid(0) + n >= 0)
         a_fetch[lid(0)] = a[16 * gid(0) + lid(0)];
-      barrier(CLK_LOCAL_MEM_FENCE) /* for a_fetch (insn_k_update depends on a_fetch_rule) */;
       if (-1 + -16 * gid(0) + -1 * lid(0) + n >= 0)
       {
         for (int k = 0; k <= 15; ++k)

loopy/schedule/tools.py

@@ -335,11 +335,15 @@ def get_return_from_kernel_mapping(kernel):
 
 # {{{ check for write races in accesses
 
-def _check_for_access_races(map_a, insn_a, map_b, insn_b, knl, callables_table):
+def _check_for_access_races(map_a, insn_a, map_b, insn_b, knl, callables_table,
+                            address_space):
     """
     Returns *True* if the execution instances of *insn_a* and *insn_b*, accessing
     the same variable via access maps *map_a* and *map_b*, result in an access race.
 
+    :arg address_space: An instance of :class:`loopy.kernel.data.AddressSpace`
+        of the variable whose accesses are being checked for a race.
+
     .. note::
 
         The accesses ``map_a``, ``map_b`` lead to write races iff there exists 2
@@ -348,9 +352,12 @@ def _check_for_access_races(map_a, insn_a, map_b, insn_b, knl, callables_table):
     import pymbolic.primitives as p
     from loopy.symbolic import isl_set_from_expr, aff_from_expr, aff_to_expr
     from loopy.kernel.data import (filter_iname_tags_by_type,
-                                   HardwareConcurrentTag)
+                                   HardwareConcurrentTag,
+                                   AddressSpace)
     from loopy.kernel.tools import get_hw_axis_base_for_codegen
 
+    assert address_space in [AddressSpace.LOCAL, AddressSpace.GLOBAL]
+
     gsize, lsize = knl.get_grid_size_upper_bounds(callables_table,
                                                   return_dict=True)
 
@@ -470,25 +477,40 @@ def _check_for_access_races(map_a, insn_a, map_b, insn_b, knl, callables_table):
     # {{{ Step 5: create the set any(l.i.A != l.i.B) OR any(g.i.A != g.i.B)
 
     space = set_a.space
-    unequal_global_id_set = isl.Set.empty(set_a.get_space())
+    unequal_local_id_set = isl.Set.empty(set_a.get_space())
+    unequal_group_id_set = isl.Set.empty(set_a.get_space())
+    equal_group_id_set = isl.BasicSet.universe(set_a.get_space())
 
     for i_l in lsize:
         lid_a = p.Variable(f"l.{i_l}.A")
         lid_b = p.Variable(f"l.{i_l}.B")
-        unequal_global_id_set |= (isl_set_from_expr(space,
-                                                    p.Comparison(lid_a, "!=", lid_b))
-                                  )
+        unequal_local_id_set |= (isl_set_from_expr(space,
+                                                   p.Comparison(lid_a, "!=", lid_b))
+                                 )
 
     for i_g in gsize:
         gid_a = p.Variable(f"g.{i_g}.A")
         gid_b = p.Variable(f"g.{i_g}.B")
-        unequal_global_id_set |= (isl_set_from_expr(space,
-                                                    p.Comparison(gid_a, "!=", gid_b))
-                                  )
+        unequal_group_id_set |= (isl_set_from_expr(space,
+                                                   p.Comparison(gid_a, "!=", gid_b))
+                                 )
+        equal_group_id_set &= (isl_set_from_expr(space,
+                                                 p.Comparison(gid_a, "==", gid_b))
+                               )
 
     # }}}
 
-    return not (set_a & set_b & unequal_global_id_set).is_empty()
+    if address_space == AddressSpace.GLOBAL:
+        return not (set_a
+                    & set_b
+                    & (unequal_local_id_set
+                       | unequal_group_id_set)
+                    ).is_empty()
+    else:
+        return not (set_a
+                    & set_b
+                    & unequal_local_id_set
+                    & equal_group_id_set).is_empty()
 
 
 class AccessMapDescriptor(enum.Enum):
@@ -582,7 +604,10 @@ class WriteRaceChecker:
 
         return _check_for_access_races(insn1_amap, self.kernel.id_to_insn[insn1],
                                        insn2_amap, self.kernel.id_to_insn[insn2],
-                                       self.kernel, self.callables_table)
+                                       self.kernel, self.callables_table,
+                                       (self.kernel
+                                        .get_var_descriptor(var_name)
+                                        .address_space))
 
 # }}}
 

test/test_loopy.py

@@ -3644,6 +3644,30 @@ def test_barrier_non_zero_hw_lbound():
     assert barrier_between(knl, "w_a", "w_b")
 
 
+def test_no_unnecessary_lbarrier(ctx_factory):
+    # This regression would fail on loopy.git <= 268a7f4
+    # (Issue reported by @thilinarmtb)
+
+    t_unit = lp.make_kernel(
+        "{[i_outer, i_inner]: 0 <= i_outer < n and 0 <= i_inner < 16}",
+        """
+        <> s_a[i_inner] = ai[i_outer * 16 + i_inner] {id=write_s_a}
+        ao[i_outer * 16 + i_inner] = 2.0 * s_a[i_inner] {id=write_ao, dep=write_s_a}
+        """,
+        assumptions="n>=0")
+
+    t_unit = lp.add_dtypes(t_unit, dict(ai=np.float32))
+    t_unit = lp.tag_inames(t_unit, dict(i_inner="l.0", i_outer="g.0"))
+    t_unit = lp.set_temporary_address_space(t_unit, "s_a", "local")
+    t_unit = lp.prioritize_loops(t_unit, "i_outer,i_inner")
+
+    t_unit = lp.preprocess_kernel(t_unit)
+    knl = lp.get_one_linearized_kernel(t_unit.default_entrypoint,
+                                       t_unit.callables_table)
+
+    assert not barrier_between(knl, "write_s_a", "write_ao")
+
+
 if __name__ == "__main__":
     if len(sys.argv) > 1:
         exec(sys.argv[1])