From 432c1060159eac7536a43f9594196a6c894bd22e Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Fri, 28 Jul 2017 17:09:31 -0500
Subject: [PATCH] Make a way to spell literal local barriers
---
loopy/codegen/control.py | 42 +++++++++++++++++++++++++++----------
loopy/kernel/creation.py | 3 +++
loopy/kernel/instruction.py | 4 ++--
test/test_loopy.py | 18 ++++++++++++++++
4 files changed, 54 insertions(+), 13 deletions(-)
diff --git a/loopy/codegen/control.py b/loopy/codegen/control.py
index d206faad5..789c00d33 100644
--- a/loopy/codegen/control.py
+++ b/loopy/codegen/control.py
@@ -1,8 +1,6 @@
"""Loop nest build top-level control/hoisting."""
-from __future__ import division
-from __future__ import absolute_import
-import six
+from __future__ import division, absolute_import
__copyright__ = "Copyright (C) 2012 Andreas Kloeckner"
@@ -26,12 +24,13 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
"""
-
+import six
from loopy.codegen.result import merge_codegen_results, wrap_in_if
import islpy as isl
from loopy.schedule import (
EnterLoop, LeaveLoop, RunInstruction, Barrier, CallKernel,
gather_schedule_block, generate_sub_sched_items)
+from loopy.diagnostic import LoopyError
def get_admissible_conditional_inames_for(codegen_state, sched_index):
@@ -150,15 +149,36 @@ def generate_code_for_sched_index(codegen_state, sched_index):
return func(codegen_state, sched_index)
elif isinstance(sched_item, Barrier):
+ # {{{ emit barrier code
+
+ from loopy.codegen.result import CodeGenerationResult
+
if codegen_state.is_generating_device_code:
- return codegen_state.ast_builder.emit_barrier(
+ barrier_ast = codegen_state.ast_builder.emit_barrier(
sched_item.kind, sched_item.comment)
- from loopy.codegen.result import CodeGenerationResult
- return CodeGenerationResult(
- host_program=None,
- device_programs=[],
- implemented_domains={},
- implemented_data_info=codegen_state.implemented_data_info)
+ if sched_item.originating_insn_id:
+ return CodeGenerationResult.new(
+ codegen_state,
+ sched_item.originating_insn_id,
+ barrier_ast,
+ codegen_state.implemented_domain)
+ else:
+ return barrier_ast
+ else:
+ # host code
+ if sched_item.kind in ["global", "local"]:
+ # host code is assumed globally and locally synchronous
+ return CodeGenerationResult(
+ host_program=None,
+ device_programs=[],
+ implemented_domains={},
+ implemented_data_info=codegen_state.implemented_data_info)
+
+ else:
+ raise LoopyError("do not know how to emit code for barrier kind '%s'"
+ "in host code" % sched_item.kind)
+
+ # }}}
elif isinstance(sched_item, RunInstruction):
insn = kernel.id_to_insn[sched_item.insn_id]
diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index 3bffc8546..89cb5f26a 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -582,6 +582,9 @@ def parse_special_insn(groups, insn_options):
if special_insn_kind == "gbarrier":
cls = BarrierInstruction
kwargs["kind"] = "global"
+ elif special_insn_kind == "lbarrier":
+ cls = BarrierInstruction
+ kwargs["kind"] = "local"
elif special_insn_kind == "nop":
cls = NoOpInstruction
else:
diff --git a/loopy/kernel/instruction.py b/loopy/kernel/instruction.py
index 08268ca9f..d5c388af6 100644
--- a/loopy/kernel/instruction.py
+++ b/loopy/kernel/instruction.py
@@ -1312,11 +1312,12 @@ class BarrierInstruction(_DataObliviousInstruction):
.. attribute:: kind
- A string, currently only ``"global"``.
+ A string, ``"global"`` or ``"local"``.
The textual syntax in a :mod:`loopy` kernel is::
... gbarrier
+ ... lbarrier
"""
fields = _DataObliviousInstruction.fields | set(["kind"])
@@ -1328,7 +1329,6 @@ class BarrierInstruction(_DataObliviousInstruction):
priority=None,
boostable=None, boostable_into=None,
predicates=None, tags=None, kind="global"):
- assert kind == "global"
if predicates:
raise LoopyError("conditional barriers are not supported")
diff --git a/test/test_loopy.py b/test/test_loopy.py
index 78633abbd..49679ce5b 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -1046,6 +1046,24 @@ def test_within_inames_and_reduction():
print(k.stringify(with_dependencies=True))
+def test_literal_local_barrier(ctx_factory):
+ ctx = ctx_factory()
+
+ knl = lp.make_kernel(
+ "{ [i]: 0<=i<n }",
+ """
+ for i
+ ... lbarrier
+ end
+ """, seq_dependencies=True)
+
+ knl = lp.fix_parameters(knl, n=128)
+
+ ref_knl = knl
+
+ lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
+
+
def test_kernel_splitting(ctx_factory):
ctx = ctx_factory()
--
GitLab