From 66b46d35f7943966cb2f1c9407cf50095c49ba9b Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Tue, 28 Aug 2018 16:54:44 +0200
Subject: [PATCH] Dabble with the cost model
---
python/dune/perftool/sumfact/vectorization.py | 30 +++++++++++--------
1 file changed, 18 insertions(+), 12 deletions(-)
diff --git a/python/dune/perftool/sumfact/vectorization.py b/python/dune/perftool/sumfact/vectorization.py
index bf463446..2b51a853 100644
--- a/python/dune/perftool/sumfact/vectorization.py
+++ b/python/dune/perftool/sumfact/vectorization.py
@@ -54,20 +54,18 @@ def attach_vectorization_info(sf):
def costmodel(sf):
- # Penalize vertical vectorization
- vertical_penalty = 1 + math.log(sf.vertical_width)
+ # Penalize vertical vectorization and scalar execution
+ verticality = sf.vertical_width
+ if isinstance(sf, SumfactKernel):
+ verticality = get_vcl_type_size(dtype_floatingpoint())
+ vertical_penalty = 1 + 0.5 * math.log(verticality, 2)
memory_penalty = 1.0
if isinstance(sf, VectorizedSumfactKernel):
- memory_penalty = 1.0 + math.log(len(set(k.interface for k in sf.kernels)), 2)
-
- # Penalize scalar sum factorization kernels
- scalar_penalty = 1
- if isinstance(sf, SumfactKernel):
- scalar_penalty = get_vcl_type_size(dtype_floatingpoint())
+ memory_penalty = 1.0 + 0.25 * math.log(len(set(k.interface for k in sf.kernels)), 2)
# Return total operations
- return sf.operations * vertical_penalty * memory_penalty * scalar_penalty
+ return sf.operations * vertical_penalty * memory_penalty
def explicit_costfunction(sf):
@@ -265,21 +263,27 @@ def level1_optimal_vectorization_strategy(sumfacts, width):
# Find the minimum cost strategy between all the quadrature point tuples
optimal_strategies = {qp: level2_optimal_vectorization_strategy(sumfacts, width, qp) for qp in quad_points}
- qp = min(optimal_strategies, key=lambda qp: strategy_cost((qp, optimal_strategies[qp])))
# If we are using the 'target' strategy, we might want to log some information.
if get_form_option("vectorization_strategy") == "target":
# Print the achieved cost and the target cost on the screen
set_form_option("vectorization_strategy", "model")
+ qp = min(optimal_strategies, key=lambda qp: strategy_cost((qp, optimal_strategies[qp])))
cost = strategy_cost((qp, optimal_strategies[qp]))
+
print("The target cost was: {}".format(get_form_option("vectorization_target")))
print("The achieved cost was: {}".format(cost))
- print("The optimal cost would be: {}".format(strategy_cost(level1_optimal_vectorization_strategy(sumfacts, width))))
+ optimum = level1_optimal_vectorization_strategy(sumfacts, width)
+ print("The optimal cost would be: {}".format(strategy_cost(optimum)))
set_form_option("vectorization_strategy", "target")
print("The score in 'target' logic was: {}".format(strategy_cost((qp, optimal_strategies[qp]))))
# Print the employed vectorization strategy into a file
- filename = "targetstrat_{}.log".format(int(float(get_form_option("vectorization_target"))))
+ suffix = ""
+ if get_global_context_value("integral_type") == "interior_facet":
+ suffix = "_dir{}_mod{}".format(get_global_context_value("facedir_s"),
+ get_global_context_value("facemod_s"))
+ filename = "targetstrat_{}{}.log".format(int(float(get_form_option("vectorization_target"))), suffix)
with open(filename, 'w') as f:
f.write("\n".join(stringify_vectorization_strategy((qp, optimal_strategies[qp]))))
@@ -291,6 +295,8 @@ def level1_optimal_vectorization_strategy(sumfacts, width):
# TODO: Depending on the number of samples, we might need a file lock here.
with open("mapping.csv", 'a') as f:
f.write(" ".join((identifier, str(cost), short_stringify_vectorization_strategy((qp, optimal_strategies[qp])))) + "\n")
+ else:
+ qp = min(optimal_strategies, key=lambda qp: strategy_cost((qp, optimal_strategies[qp])))
return qp, optimal_strategies[qp]
--
GitLab