diff --git a/loopy/check.py b/loopy/check.py
index e72f9e3e6c4db797220729a5f282d4944b31d6ac..8d78e684bc0d793d6c17bbfebefc428d71fbb0b4 100644
--- a/loopy/check.py
+++ b/loopy/check.py
@@ -493,6 +493,7 @@ def check_that_atomic_ops_are_used_exactly_on_atomic_arrays(kernel):
accessed_atomic_vars = insn.dependency_names() & atomicity_candidates
if not accessed_atomic_vars <= atomic_accesses:
+ missed = atomic_accesses - accessed_atomic_vars
raise LoopyError("atomic variable(s) '%s' in instruction '%s' "
"used in non-atomic access"
% (
diff --git a/loopy/kernel/instruction.py b/loopy/kernel/instruction.py
index d5c388af60a39987c09092fc93325f067a8f4cf7..1d20ff71c3f13035d2f01b56a69e2d983c7660af 100644
--- a/loopy/kernel/instruction.py
+++ b/loopy/kernel/instruction.py
@@ -603,7 +603,7 @@ class AtomicInit(VarAtomicity):
memory_scope.to_string(self.scope))
-class AtomicUpdate(VarAtomicity):
+class OrderedAtomic(VarAtomicity):
"""Properties of an atomic operation. A subclass of :class:`VarAtomicity`.
.. attribute:: ordering
@@ -624,7 +624,7 @@ class AtomicUpdate(VarAtomicity):
"""
super(AtomicUpdate, self).update_persistent_hash(key_hash, key_builder)
- key_builder.rec(key_hash, "AtomicUpdate")
+ key_builder.rec(key_hash, str(self.__class__.__name__))
key_builder.rec(key_hash, self.ordering)
key_builder.rec(key_hash, self.scope)
@@ -634,11 +634,40 @@ class AtomicUpdate(VarAtomicity):
and self.scope == other.scope)
def __str__(self):
- return "update[%s]%s/%s" % (
+ return "%s[%s]%s/%s" % (
+ self.op_name,
self.var_name,
memory_ordering.to_string(self.ordering),
memory_scope.to_string(self.scope))
+
+class AtomicUpdate(VarAtomicity):
+ """Properties of an atomic update. A subclass of :class:`VarAtomicity`.
+
+ .. attribute:: ordering
+
+ One of the values from :class:`memory_ordering`
+
+ .. attribute:: scope
+
+ One of the values from :class:`memory_scope`
+ """
+ op_name = 'update'
+
+
+class AtomicLoad(VarAtomicity):
+ """Properties of an atomic load. A subclass of :class:`VarAtomicity`.
+
+ .. attribute:: ordering
+
+ One of the values from :class:`memory_ordering`
+
+ .. attribute:: scope
+
+ One of the values from :class:`memory_scope`
+ """
+ op_name = 'load'
+
# }}}
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index ced1aaaa13ed8275c1e3a376d1c24895287b3239..554bf6b24bfcd8e3d3db8e40dd6535d19984ae8c 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -2020,6 +2020,48 @@ def limit_boostability(kernel):
# }}}
+# {{{ check for loads of atomic variables
+
+def check_atomic_loads(kernel):
+ """Find instances of AtomicInit or AtomicUpdate with use of other atomic
+ variables to update the atomicity
+ """
+
+ logger.debug("%s: check atomic loads" % kernel.name)
+ from loopy.types import AtomicType
+ from loopy.kernel.array import ArrayBase
+ from loopy.kernel.instruction import Assignment, AtomicLoad
+
+ # find atomic variables
+ atomicity_candidates = (
+ set(v.name for v in six.itervalues(kernel.temporary_variables)
+ if isinstance(v.dtype, AtomicType))
+ |
+ set(v.name for v in kernel.args
+ if isinstance(v, ArrayBase)
+ and isinstance(v.dtype, AtomicType)))
+
+ new_insns = []
+ for insn in kernel.instructions:
+ if isinstance(insn, Assignment):
+ # look for atomic variables
+ atomic_accesses = set(a.var_name for a in insn.atomicity)
+ accessed_atomic_vars = insn.dependency_names() & atomicity_candidates
+ if not accessed_atomic_vars <= atomic_accesses:
+ #if we're missing some
+ missed = accessed_atomic_vars - atomic_accesses
+ for x in missed:
+ if set([x]) & atomicity_candidates:
+ insn = insn.copy(
+ atomicity=insn.atomicity + (AtomicLoad(x),))
+
+ new_insns.append(insn)
+
+ return kernel.copy(instructions=new_insns)
+
+# }}}
+
+
preprocess_cache = PersistentDict("loopy-preprocess-cache-v2-"+DATA_MODEL_VERSION,
key_builder=LoopyKeyBuilder())
@@ -2104,6 +2146,10 @@ def preprocess_kernel(kernel, device=None):
kernel = find_idempotence(kernel)
kernel = limit_boostability(kernel)
+ # check for atomic loads, much easier to do here now that the dependencies
+ # have been established
+ kernel = check_atomic_loads(kernel)
+
kernel = kernel.target.preprocess(kernel)
logger.info("%s: preprocess done" % kernel.name)
diff --git a/test/test_loopy.py b/test/test_loopy.py
index 3593019ad2ca7e41f7db4c95616184e1e8972125..f767197b09e2cddebf5e7b1eb9fc25da23e17746 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -1011,6 +1011,40 @@ def test_atomic(ctx_factory, dtype):
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=10000))
+def test_atomic_load(ctx_factory):
+ dtype = np.int32
+ ctx = ctx_factory()
+
+ if (
+ np.dtype(dtype).itemsize == 8
+ and "cl_khr_int64_base_atomics" not in ctx.devices[0].extensions):
+ pytest.skip("64-bit atomics not supported on device")
+
+ import pyopencl.version # noqa
+ if (
+ cl.version.VERSION < (2015, 2)
+ and dtype == np.int64):
+ pytest.skip("int64 RNG not supported in PyOpenCL < 2015.2")
+
+ knl = lp.make_kernel(
+ "{ [i]: 0<=i<n }",
+ """
+ temp[0] = 5 {id=init, atomic}
+ out[i%20] = out[i%20] + temp[0] {dep=init, nosync=init, atomic}
+ """,
+ [
+ lp.GlobalArg("out", dtype, shape=lp.auto, for_atomic=True),
+ lp.GlobalArg('temp', dtype, shape=lp.auto, for_atomic=True),
+ "..."
+ ],
+ assumptions="n>0")
+
+ ref_knl = knl
+ knl = lp.split_iname(knl, "i", 512)
+ knl = lp.split_iname(knl, "i_inner", 128, outer_tag="unr", inner_tag="g.0")
+ lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=10000))
+
+
def test_within_inames_and_reduction():
# See https://github.com/inducer/loopy/issues/24