From 677b9f2aade96306102da3976e5f4db8909821a7 Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Wed, 3 Jan 2018 07:54:26 -0500
Subject: [PATCH] Disable performance bug for non-production work
---
.../perftool/loopy/transformations/disjointgroups.py | 10 ++++++++--
python/dune/perftool/options.py | 1 +
2 files changed, 9 insertions(+), 2 deletions(-)
diff --git a/python/dune/perftool/loopy/transformations/disjointgroups.py b/python/dune/perftool/loopy/transformations/disjointgroups.py
index 2f0a64f0..1f22c67d 100644
--- a/python/dune/perftool/loopy/transformations/disjointgroups.py
+++ b/python/dune/perftool/loopy/transformations/disjointgroups.py
@@ -1,6 +1,12 @@
""" A helper transformation that makes all groups conflicting """
+from dune.perftool.options import get_option
def make_groups_conflicting(knl):
- groups = frozenset().union(*tuple(i.groups for i in knl.instructions))
- return knl.copy(instructions=[i.copy(conflicts_with_groups=groups - i.groups) for i in knl.instructions])
+ # As this transformation introduces a performance bug that basically
+ # kills our CI, we only apply it if really needed - meaning in production.
+ if get_option("assure_statement_ordering"):
+ groups = frozenset().union(*tuple(i.groups for i in knl.instructions))
+ return knl.copy(instructions=[i.copy(conflicts_with_groups=groups - i.groups) for i in knl.instructions])
+ else:
+ return knl
diff --git a/python/dune/perftool/options.py b/python/dune/perftool/options.py
index 789f7b86..9ae717c5 100644
--- a/python/dune/perftool/options.py
+++ b/python/dune/perftool/options.py
@@ -64,6 +64,7 @@ class PerftoolOptionsArray(ImmutableRecord):
architecture = PerftoolOption(default="haswell", helpstr="The architecture to optimize for. Possible values: haswell|knl")
grid_offset = PerftoolOption(default=False, helpstr="Set to true if you want a yasp grid where the lower left corner is not in the origin.")
simplify = PerftoolOption(default=True, helpstr="Whether to simplify expressions using sympy")
+ assure_statement_ordering = PerftoolOption(default=False, helpstr="Whether special care should be taken for a good statement ordering in sumfact kernels, runs into a loopy scheduler performance bug, but is necessary for production.")
# Arguments that are mainly to be set by logic depending on other options
max_vector_width = PerftoolOption(default=256, helpstr=None)
--
GitLab