From a8eadc594f95f683dab3238d0a2b536f3179ab8d Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Wed, 5 Apr 2017 14:40:24 +0200
Subject: [PATCH] Use padding on theta large matrices and get trid of explicit
4s
---
python/dune/perftool/sumfact/accumulation.py | 2 +-
python/dune/perftool/sumfact/realization.py | 2 +-
python/dune/perftool/sumfact/symbolic.py | 9 +++++----
python/dune/perftool/sumfact/tabulation.py | 20 ++++++++++++++-----
python/dune/perftool/sumfact/vectorization.py | 14 ++++++++++---
5 files changed, 33 insertions(+), 14 deletions(-)
diff --git a/python/dune/perftool/sumfact/accumulation.py b/python/dune/perftool/sumfact/accumulation.py
index d77f8145..d09a91b7 100644
--- a/python/dune/perftool/sumfact/accumulation.py
+++ b/python/dune/perftool/sumfact/accumulation.py
@@ -203,7 +203,7 @@ def generate_accumulation_instruction(visitor, accterm, measure, subdomain_id):
vecinames = ()
# TODO: evaluate whether the following line would be okay with vsf.vectorized
if vsf.vec_index(sf) is not None:
- iname = accum_iname((accterm.argument.restriction, restriction), 4, "vec")
+ iname = accum_iname((accterm.argument.restriction, restriction), vsf.vector_width, "vec")
vecinames = (iname,)
transform(lp.tag_inames, [(iname, "vec")])
from dune.perftool.tools import maybe_wrap_subscript
diff --git a/python/dune/perftool/sumfact/realization.py b/python/dune/perftool/sumfact/realization.py
index e2e375a2..583f8f4d 100644
--- a/python/dune/perftool/sumfact/realization.py
+++ b/python/dune/perftool/sumfact/realization.py
@@ -95,7 +95,7 @@ def _realize_sum_factorization_kernel(sf):
if sf.vectorized:
ftags = ftags + ",vec"
ctags = ctags + ",vec"
- vec_shape = (4,)
+ vec_shape = (sf.vector_width,)
# Measure times and count operations in c++ code
if get_option("instrumentation_level") >= 4:
diff --git a/python/dune/perftool/sumfact/symbolic.py b/python/dune/perftool/sumfact/symbolic.py
index a9a0dd4c..37fa5e3f 100644
--- a/python/dune/perftool/sumfact/symbolic.py
+++ b/python/dune/perftool/sumfact/symbolic.py
@@ -344,6 +344,7 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
transpose=self.kernels[0].matrix_sequence[i].transpose,
face=self.kernels[0].matrix_sequence[i].face,
derivative=tuple(k.matrix_sequence[i].derivative for k in self.kernels),
+ width=self.vector_width,
)
for i in range(self.length))
@@ -419,14 +420,14 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
@property
def flat_input_shape(self):
- return (product(mat.cols for mat in self.matrix_sequence), 4)
+ return (product(mat.cols for mat in self.matrix_sequence), self.vector_width)
@property
def quadrature_shape(self):
if self.transposed:
- return tuple(mat.cols for mat in self.matrix_sequence if mat.face is None) + (4,)
+ return tuple(mat.cols for mat in self.matrix_sequence if mat.face is None) + (self.vector_width,)
else:
- return tuple(mat.rows for mat in self.matrix_sequence if mat.face is None) + (4,)
+ return tuple(mat.rows for mat in self.matrix_sequence if mat.face is None) + (self.vector_width,)
@property
def quadrature_dimtags(self):
@@ -436,7 +437,7 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
@property
def dof_shape(self):
- return tuple(mat.rows for mat in self.matrix_sequence) + (4,)
+ return tuple(mat.rows for mat in self.matrix_sequence) + (self.vector_width,)
@property
def dof_dimtags(self):
diff --git a/python/dune/perftool/sumfact/tabulation.py b/python/dune/perftool/sumfact/tabulation.py
index c6fa99e3..e6465985 100644
--- a/python/dune/perftool/sumfact/tabulation.py
+++ b/python/dune/perftool/sumfact/tabulation.py
@@ -59,7 +59,7 @@ class BasisTabulationMatrix(BasisTabulationMatrixBase):
class BasisTabulationMatrixArray(BasisTabulationMatrixBase):
- def __init__(self, rows, cols, transpose, derivative, face):
+ def __init__(self, rows, cols, transpose, derivative, face, width):
assert isinstance(derivative, tuple)
BasisTabulationMatrixBase.__init__(self,
rows=rows,
@@ -67,6 +67,7 @@ class BasisTabulationMatrixArray(BasisTabulationMatrixBase):
transpose=transpose,
derivative=derivative,
face=face,
+ width=width,
)
@property
@@ -77,13 +78,22 @@ class BasisTabulationMatrixArray(BasisTabulationMatrixBase):
quadrature_points_per_direction(),
)
for i, d in enumerate(self.derivative):
- define_theta(name, (self.rows, self.cols), self.transpose, d, face=self.face, additional_indices=(i,))
+ define_theta(name, (self.rows, self.cols), self.transpose, d, face=self.face, additional_indices=(i,), width=self.width)
+
+ # Apply padding to those fields not used. This is necessary because you may get memory
+ # initialized with NaN and those NaNs will screw the horizontal_add.
+ i = theta_iname("i", self.rows)
+ j = theta_iname("j", self.cols)
+ for index in set(range(self.width)) - set(range(len(self.derivative))):
+ instruction(assignee=prim.Subscript(prim.Variable(name), (prim.Variable(i), prim.Variable(j), index)),
+ expression=0,
+ kernel="operator")
return loopy_class_member(name,
classtag="operator",
dtype=numpy.float64,
dim_tags="f,f,vec",
- shape=(self.rows, self.cols, 4),
+ shape=(self.rows, self.cols, self.width),
potentially_vectorized=True,
managed=True,
)
@@ -214,7 +224,7 @@ def polynomial_lookup_mangler(target, func, dtypes):
return CallMangleInfo(func.name, (NumpyType(numpy.float64),), (NumpyType(numpy.int32), NumpyType(numpy.float64)))
-def define_theta(name, shape, transpose, derivative, face=None, additional_indices=()):
+def define_theta(name, shape, transpose, derivative, face=None, additional_indices=(), width=None):
qp = name_oned_quadrature_points()
qw = name_oned_quadrature_weights()
sort_quadrature_points_weights(qp, qw)
@@ -223,7 +233,7 @@ def define_theta(name, shape, transpose, derivative, face=None, additional_indic
dim_tags = "f,f"
if additional_indices:
dim_tags = dim_tags + ",c"
- shape = shape + (4,)
+ shape = shape + (width,)
loopy_class_member(name,
dtype=numpy.float64,
diff --git a/python/dune/perftool/sumfact/vectorization.py b/python/dune/perftool/sumfact/vectorization.py
index 84c7e615..460ae1e9 100644
--- a/python/dune/perftool/sumfact/vectorization.py
+++ b/python/dune/perftool/sumfact/vectorization.py
@@ -1,5 +1,6 @@
""" Sum factorization vectorization """
+from dune.perftool.loopy.vcl import get_vcl_type_size
from dune.perftool.loopy.symbolic import SumfactKernel, VectorizedSumfactKernel
from dune.perftool.generation import (generator_factory,
get_counted_variable,
@@ -13,6 +14,7 @@ from dune.perftool.error import PerftoolError
from dune.perftool.options import get_option
import loopy as lp
+import numpy as np
@generator_factory(item_tags=("vecinfo", "dryrundata"), cache_key_generator=lambda o, n: o)
@@ -40,10 +42,16 @@ def no_vectorization(sumfacts):
def horizontal_vectorization_strategy(sumfacts):
- if len(sumfacts) in (3, 4):
+ width = get_vcl_type_size(np.float64)
+ # We currently heuristically allow horizontal vectorization if the number
+ # of kernels matches the maximum vector width or is one below that. There
+ # is a lot of TODOs with that procedure:
+ # * You might have *more* kernels than vector width
+ # * You might want to combine horizontal and vertical strategies
+ if len(sumfacts) in (width - 1, width):
# Map the sum factorization to their position in the joint kernel
position_mapping = {}
- available = set(range(4))
+ available = set(range(width))
for sf in sumfacts:
if sf.preferred_position is not None:
# This asserts that no two kernels want to take the same position
@@ -69,7 +77,7 @@ def horizontal_vectorization_strategy(sumfacts):
for sumf in sumfacts:
_cache_vectorization_info(sumf,
VectorizedSumfactKernel(kernels=kernels,
- vector_width=4,
+ vector_width=width,
buffer=buffer,
input=input,
)
--
GitLab