From 2a56e7bae429a014c285c1275b621e61392c9ef6 Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Mon, 8 Oct 2018 13:54:36 +0200
Subject: [PATCH] Add a penalty for quadrature point increase
This penalty accounts for the increased work in the quadrature loop.
Still a bit experimental.
---
python/dune/perftool/sumfact/vectorization.py | 39 +++++++++++++++----
1 file changed, 31 insertions(+), 8 deletions(-)
diff --git a/python/dune/perftool/sumfact/vectorization.py b/python/dune/perftool/sumfact/vectorization.py
index 58df262f..cb6e8793 100644
--- a/python/dune/perftool/sumfact/vectorization.py
+++ b/python/dune/perftool/sumfact/vectorization.py
@@ -99,6 +99,20 @@ def target_costfunction(sf):
return abs(realcost - ratio * target)
+def accumulate_for_strategy(strategy, func, reduce=lambda x, y: x + y, start=0):
+ """ Accumulate a quantity over a given vectorization strategy """
+ accum = start
+
+ # Make sure to count each implementation exactly once
+ keys = set(sf.cache_key for sf in strategy.values())
+ for sf in strategy.values():
+ if sf.cache_key in keys:
+ accum = reduce(accum, func(sf))
+ keys.discard(sf.cache_key)
+
+ return accum
+
+
def strategy_cost(strat_tuple):
qp, strategy = strat_tuple
@@ -119,14 +133,23 @@ def strategy_cost(strat_tuple):
keys = set(sf.cache_key for sf in strategy.values())
set_quadrature_points(qp)
- # Sum over all the sum factorization kernels in the realization
- score = 0.0
- for sf in strategy.values():
- if sf.cache_key in keys:
- score = score + float(func(sf))
- keys.discard(sf.cache_key)
+ return accumulate_for_strategy(strategy, lambda sf: float(func(sf)))
+
+
+def quadrature_penalized_strategy_cost(strat_tuple):
+ cost = strategy_cost(strat_tuple)
+
+ qp, strategy = strat_tuple
+ num_qp_new = product(qp)
+ sf_flops = accumulate_for_strategy(strategy, lambda sf: sf.operations)
+
+ set_quadrature_points(None)
+ num_qp_old = product(quadrature_points_per_direction())
+ set_quadrature_points(qp)
- return score
+ # TODO determine this
+ ops_per_qp = 100
+ return float((sf_flops + ops_per_qp * num_qp_new) / (sf_flops + ops_per_qp * num_qp_old)) * cost
def fixedqp_strategy_costfunction(qp):
@@ -300,7 +323,7 @@ def level1_optimal_vectorization_strategy(sumfacts, width):
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])))
+ qp = min(optimal_strategies, key=lambda qp: quadrature_penalized_strategy_cost((qp, optimal_strategies[qp])))
return qp, optimal_strategies[qp]
--
GitLab