From 0d78c5844023a765aae67a585a60682e78d433a2 Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Thu, 23 Aug 2018 15:53:07 +0200
Subject: [PATCH] Beautify the implementation a small bit
---
python/dune/perftool/sumfact/vectorization.py | 47 ++++++++++---------
1 file changed, 24 insertions(+), 23 deletions(-)
diff --git a/python/dune/perftool/sumfact/vectorization.py b/python/dune/perftool/sumfact/vectorization.py
index cd0f7ffa..d98ba5e7 100644
--- a/python/dune/perftool/sumfact/vectorization.py
+++ b/python/dune/perftool/sumfact/vectorization.py
@@ -85,13 +85,15 @@ def explicit_costfunction(sf):
return 1000000000000
-_global_kernel_amount = 0
-
-
def target_costfunction(sf):
+ # The cost of a kernel is given by the difference to the desired target cost.
+ # Pitfall: The target cost needs to be weighed to account for this being called
+ # on subsets and not on a full vectorization strategy!
+ all_sf, _ = filter_active_inactive_sumfacts()
+ total = len(all_sf)
target = float(get_form_option("vectorization_target"))
realcost = costmodel(sf)
- ratio = sf.horizontal_width / _global_kernel_amount
+ ratio = sf.horizontal_width / total
return abs(realcost - ratio * target)
@@ -110,8 +112,7 @@ def strategy_cost(strat_tuple):
raise NotImplementedError("Vectorization strategy '{}' unknown!".format(s))
keys = set(sf.cache_key for sf in strategy.values())
- if qp is not None:
- set_quadrature_points(qp)
+ set_quadrature_points(qp)
# Sum over all the sum factorization kernels in the realization
score = 0.0
@@ -154,13 +155,7 @@ def stringify_vectorization_strategy(strategy):
return result
-def decide_vectorization_strategy():
- """ Decide how to vectorize!
- Note that the vectorization of the quadrature loop is independent of this,
- as it is implemented through a post-processing (== loopy transformation) step.
- """
- logger = logging.getLogger(__name__)
-
+def filter_active_inactive_sumfacts():
# Retrieve all sum factorization kernels for stage 1 and 3
from dune.perftool.generation import retrieve_cache_items
all_sumfacts = [i for i in retrieve_cache_items("kernel_default and sumfactnodes")]
@@ -174,6 +169,18 @@ def decide_vectorization_strategy():
# All sum factorization kernels that get used
active_sumfacts = [i for i in all_sumfacts if i.stage == 3 or i in basis_sumfacts]
+ return active_sumfacts, inactive_sumfacts
+
+
+def decide_vectorization_strategy():
+ """ Decide how to vectorize!
+ Note that the vectorization of the quadrature loop is independent of this,
+ as it is implemented through a post-processing (== loopy transformation) step.
+ """
+ logger = logging.getLogger(__name__)
+
+ active_sumfacts, inactive_sumfacts = filter_active_inactive_sumfacts()
+
# If no vectorization is needed, abort now
if get_form_option("vectorization_strategy") == "none" or (get_global_context_value("form_type") == "jacobian" and not get_form_option("vectorization_jacobians")):
for sf in all_sumfacts:
@@ -208,18 +215,12 @@ def decide_vectorization_strategy():
sfdict = add_to_frozendict(sfdict, {sf: 0 for sf in inactive_sumfacts})
# Register the results
- for sf in all_sumfacts:
+ from itertools import chain
+ for sf in chain(active_sumfacts, inactive_sumfacts):
_cache_vectorization_info(sf, sfdict[sf])
def level1_optimal_vectorization_strategy(sumfacts, width):
- # If this uses the 'target' cost model, we need to store information on how many
- # sum factorization kernels need to be implemented. This will be used to correctly
- # weight the cost target in the cost function.
- if get_form_option("vectorization_strategy") == "target":
- global _global_kernel_amount
- _global_kernel_amount = len(sumfacts)
-
# Gather a list of possible quadrature point tuples
quad_points = [quadrature_points_per_direction()]
if get_form_option("vectorization_allow_quadrature_changes"):
@@ -255,11 +256,11 @@ def level1_optimal_vectorization_strategy(sumfacts, width):
# Write an entry into a csvfile which connects the given measuring identifier with a cost
from dune.testtools.parametertree.parser import parse_ini_file
- inifile = parse_ini_file(get_option("inifile"))
+ inifile = parse_ini_file(get_option("ini_file"))
identifier = inifile["identifier"]
filename = join(get_option("project_basedir"), "costmodel-verification", "mapping.csv")
- #TODO: Depending on the number of samples, we might need a file lock here.
+ # TODO: Depending on the number of samples, we might need a file lock here.
with open(filename, 'a') as f:
f.write("{} {}".format(identifier, cost))
--
GitLab