From 171b1962cec9563fad310f868624d610599f123b Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Wed, 24 Jan 2018 13:46:51 +0100
Subject: [PATCH] First draft of a cost function that picks a specific strategy
by index
---
python/dune/perftool/options.py | 4 +-
python/dune/perftool/sumfact/vectorization.py | 37 +++++++++++++------
python/setup.py | 1 +
3 files changed, 30 insertions(+), 12 deletions(-)
diff --git a/python/dune/perftool/options.py b/python/dune/perftool/options.py
index 23733a7a..85efbb66 100644
--- a/python/dune/perftool/options.py
+++ b/python/dune/perftool/options.py
@@ -55,11 +55,13 @@ class PerftoolOptionsArray(ImmutableRecord):
fastdg = PerftoolOption(default=False, helpstr="Use FastDGGridOperator from PDELab.")
sumfact = PerftoolOption(default=False, helpstr="Use sumfactorization")
vectorization_quadloop = PerftoolOption(default=False, helpstr="whether to generate code with explicit vectorization")
- vectorization_strategy = PerftoolOption(default="none", helpstr="The identifier of the vectorization cost model. Possible values: none|explicit|model")
+ vectorization_strategy = PerftoolOption(default="none", helpstr="The identifier of the vectorization cost model. Possible values: none|explicit|model|fromlist")
vectorization_horizontal = PerftoolOption(default=None, helpstr="an explicit value for horizontal vectorization read by the 'explicit' strategy")
vectorization_vertical = PerftoolOption(default=None, helpstr="an explicit value for vertical vectorization read by the 'explicit' strategy")
vectorization_padding = PerftoolOption(default=None, helpstr="an explicit value for the allowed padding in vectorization")
vectorization_allow_quadrature_changes = PerftoolOption(default=False, helpstr="whether the vectorization strategy is allowed to alter quadrature point numbers")
+ vectorization_list_index = PerftoolOption(default=None, helpstr="Which vectorization to pick from a list (only valid with vectorization_strategy=fromlist).")
+ vectorization_divide_and_conquer = PerftoolOption(default=True, helpstr="Whether to find cost model optima in subsets, only turn off when you know exactly what you are doing")
turn_off_diagonal_jacobian = PerftoolOption(default=False, helpstr="Do not use diagonal_jacobian transformation on the ufl tree and cast result of jacobianInverseTransposed into a FieldMatrix.")
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.")
diff --git a/python/dune/perftool/sumfact/vectorization.py b/python/dune/perftool/sumfact/vectorization.py
index f3fb96b9..6f5049aa 100644
--- a/python/dune/perftool/sumfact/vectorization.py
+++ b/python/dune/perftool/sumfact/vectorization.py
@@ -89,6 +89,16 @@ def explicit_costfunction(sf):
return 1000000000000
+@backend(interface="vectorization_strategy", name="fromlist")
+def fromlist_costfunction(sf):
+ from dune.perftool.pdelab.signatures import kernel_name
+ count = get_counter(kernel_name())
+ if count == get_option("vectorization_list_index"):
+ return 0.0
+ else:
+ return 1.0
+
+
def strategy_cost(strategy):
func = get_backend(interface="vectorization_strategy",
selector=lambda: get_option("vectorization_strategy"))
@@ -139,7 +149,7 @@ def decide_vectorization_strategy():
from dune.perftool.generation import retrieve_cache_items
all_sumfacts = [i for i in retrieve_cache_items("kernel_default and sumfactnodes")]
- # Stage 1 sumfactorizations that were actually used
+ # Stage 1 sum factorizations that were actually used
basis_sumfacts = [i for i in retrieve_cache_items('kernel_default and basis_sf_kernels')]
# This means we can have sum factorizations that will not get used
@@ -202,18 +212,23 @@ def fixed_quadrature_optimal_vectorization(sumfacts, width, qp):
"""
set_quadrature_points(qp)
- # Find the sets of simultaneously realizable kernels (thats an equivalence relation)
- keys = frozenset(sf.input_key for sf in sumfacts)
+ if get_option("vectorization_divide_and_conquer"):
+ # Find the sets of simultaneously realizable kernels (thats an equivalence relation)
+ keys = frozenset(sf.input_key for sf in sumfacts)
- # Find minimums for each of these sets
- sfdict = frozendict()
- for key in keys:
- key_sumfacts = frozenset(sf for sf in sumfacts if sf.input_key == key)
- minimum = min(fixed_quad_vectorization_opportunity_generator(key_sumfacts, width, qp),
- key=strategy_cost)
- sfdict = add_to_frozendict(sfdict, minimum)
+ # Find minimums for each of these sets
+ sfdict = frozendict()
- return sfdict
+ for key in keys:
+ key_sumfacts = frozenset(sf for sf in sumfacts if sf.input_key == key)
+ minimum = min(fixed_quad_vectorization_opportunity_generator(key_sumfacts, width, qp),
+ key=strategy_cost)
+ sfdict = add_to_frozendict(sfdict, minimum)
+
+ return sfdict
+ else:
+ return min(fixed_quad_vectorization_opportunity_generator(sumfacts, width, qp),
+ key=strategy_cost)
def fixed_quad_vectorization_opportunity_generator(sumfacts, width, qp, already=frozendict()):
diff --git a/python/setup.py b/python/setup.py
index ccbe9e56..c7e80ce1 100644
--- a/python/setup.py
+++ b/python/setup.py
@@ -45,5 +45,6 @@ setup(name='dune.perftool',
entry_points = {
"console_scripts": [
"ufl2pdelab = dune.perftool.compile:compile_form",
+ "picklevecstrats = dune.perftool.sumfact.vectorization:pickle_vectorization_strategies",
]
})
--
GitLab