From 7c27277d23dee19f117be8883b77730043dc961a Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Wed, 6 Dec 2017 11:36:14 +0100
Subject: [PATCH] Implement a printing facility for vectorization strategies
---
python/dune/perftool/loopy/symbolic.py | 2 +-
python/dune/perftool/sumfact/accumulation.py | 3 +++
python/dune/perftool/sumfact/basis.py | 3 +++
python/dune/perftool/sumfact/symbolic.py | 12 +++++++++++
python/dune/perftool/sumfact/tabulation.py | 16 +++++++++++++++
python/dune/perftool/sumfact/vectorization.py | 20 +++++++++++++++++++
6 files changed, 55 insertions(+), 1 deletion(-)
diff --git a/python/dune/perftool/loopy/symbolic.py b/python/dune/perftool/loopy/symbolic.py
index a592c461..01a99225 100644
--- a/python/dune/perftool/loopy/symbolic.py
+++ b/python/dune/perftool/loopy/symbolic.py
@@ -49,7 +49,7 @@ def walk_map_sumfact_kernel(self, expr, *args):
def stringify_map_sumfact_kernel(self, expr, *args):
- return "SUMFACT"
+ return str(expr)
def dependency_map_sumfact_kernel(self, expr):
diff --git a/python/dune/perftool/sumfact/accumulation.py b/python/dune/perftool/sumfact/accumulation.py
index 53d949c9..3c935161 100644
--- a/python/dune/perftool/sumfact/accumulation.py
+++ b/python/dune/perftool/sumfact/accumulation.py
@@ -92,6 +92,9 @@ class AlreadyAssembledInput(SumfactKernelInputBase):
def __hash__(self):
return hash(self.index)
+ def __str__(self):
+ return "Input{}".format(self.index[0])
+
class SumfactAccumulationInfo(ImmutableRecord):
def __init__(self,
diff --git a/python/dune/perftool/sumfact/basis.py b/python/dune/perftool/sumfact/basis.py
index 1c1823df..889f464f 100644
--- a/python/dune/perftool/sumfact/basis.py
+++ b/python/dune/perftool/sumfact/basis.py
@@ -64,6 +64,9 @@ class LFSSumfactKernelInput(SumfactKernelInputBase, ImmutableRecord):
restriction=restriction,
)
+ def __str__(self):
+ return "{}".format(self.coeff_func(self.restriction))
+
def realize(self, sf, index, insn_dep):
lfs = name_lfs(self.element, self.restriction, self.element_index)
basisiname = sumfact_iname(name_lfs_bound(lfs), "basis")
diff --git a/python/dune/perftool/sumfact/symbolic.py b/python/dune/perftool/sumfact/symbolic.py
index 44ae9f79..857a1d97 100644
--- a/python/dune/perftool/sumfact/symbolic.py
+++ b/python/dune/perftool/sumfact/symbolic.py
@@ -146,6 +146,12 @@ class SumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable):
def stringifier(self):
return lp.symbolic.StringifyMapper
+ def __str__(self):
+ # Above stringifier just calls back into this
+ return "SF{}:[{}]->[{}]".format(self.stage,
+ str(self.input),
+ ", ".join(str(m) for m in self.matrix_sequence))
+
mapper_method = "map_sumfact_kernel"
#
@@ -385,6 +391,12 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
def stringifier(self):
return lp.symbolic.StringifyMapper
+ def __str__(self):
+ # Above stringifier just calls back into this
+ return "VSF{}:[{}]->[{}]".format(self.stage,
+ ", ".join(str(k.input) for k in self.kernels),
+ ", ".join(str(mat) for mat in self.matrix_sequence))
+
mapper_method = "map_vectorized_sumfact_kernel"
init_arg_names = ("kernels", "horizontal_width", "vertical_width", "buffer", "insn_dep")
diff --git a/python/dune/perftool/sumfact/tabulation.py b/python/dune/perftool/sumfact/tabulation.py
index 177c476b..c4ce095b 100644
--- a/python/dune/perftool/sumfact/tabulation.py
+++ b/python/dune/perftool/sumfact/tabulation.py
@@ -63,6 +63,15 @@ class BasisTabulationMatrix(BasisTabulationMatrixBase, ImmutableRecord):
slice_index=slice_index,
)
+ def __str__(self):
+ return "{}{}A{}{}{}" \
+ .format("face{}_".format(self.face) if self.face is not None else "",
+ "d" if self.derivative else "",
+ self.basis_size,
+ "T" if self.transpose else "",
+ "_slice{}".format(self.slice_index) if self.slice_size is not None else "",
+ )
+
@property
def rows(self):
if self.transpose:
@@ -131,6 +140,13 @@ class BasisTabulationMatrixArray(BasisTabulationMatrixBase):
width = len(tabs)
self.width = width
+ def __str__(self):
+ abbrevs = tuple("{}A{}{}".format("d" if t.derivative else "",
+ self.basis_size,
+ "s{}".format(t.slice_index) if t.slice_size is not None else "")
+ for t in self.tabs)
+ return "_".join(abbrevs)
+
@property
def quadrature_size(self):
return self.tabs[0].quadrature_size
diff --git a/python/dune/perftool/sumfact/vectorization.py b/python/dune/perftool/sumfact/vectorization.py
index 6261a380..53a5c278 100644
--- a/python/dune/perftool/sumfact/vectorization.py
+++ b/python/dune/perftool/sumfact/vectorization.py
@@ -203,6 +203,26 @@ def greedy_vectorization_strategy(sumfacts, width):
return result
+def print_vectorization_strategy(strategy):
+ qp, strategy = strategy
+ print "\nPrinting potential vectorization strategy:"
+ print "Quadrature point tuple: {}".format(qp)
+
+ # Look for all realizations in the strategy and iterate over them
+ cache_keys = frozenset(v.cache_key for v in strategy.values())
+ for ck in cache_keys:
+ # Filter all the kernels that are realized by this and print
+ for key in strategy:
+ if strategy[key].cache_key == ck:
+ print "{}:".format(key)
+
+ # Find one representative to print
+ for val in strategy.values():
+ if val.cache_key == ck:
+ print " {}".format(val)
+ break
+
+
def decide_vectorization_strategy():
""" Decide how to vectorize!
Note that the vectorization of the quadrature loop is independent of this,
--
GitLab