From 81aa44584776ff8f7f3e41e5fdc5246eb476091c Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Thu, 30 Mar 2017 16:42:38 +0200
Subject: [PATCH] fixup
---
python/dune/perftool/loopy/symbolic.py | 4 ++--
python/dune/perftool/sumfact/realization.py | 16 ++++++++--------
python/dune/perftool/sumfact/sumfact.py | 2 +-
python/dune/perftool/ufl/visitor.py | 10 +++++++---
4 files changed, 18 insertions(+), 14 deletions(-)
diff --git a/python/dune/perftool/loopy/symbolic.py b/python/dune/perftool/loopy/symbolic.py
index 91dadf16..cb877c7e 100644
--- a/python/dune/perftool/loopy/symbolic.py
+++ b/python/dune/perftool/loopy/symbolic.py
@@ -23,7 +23,7 @@ class SumfactKernel(ImmutableRecord, prim.Variable):
stage=1,
preferred_position=None,
restriction=0,
- within_inames=frozenset(),
+ within_inames=(),
input=None,
padding=frozenset(),
index=None,
@@ -41,7 +41,7 @@ class SumfactKernel(ImmutableRecord, prim.Variable):
if not isinstance(restriction, tuple):
restriction = (restriction, 0)
- assert isinstance(within_inames, frozenset)
+ assert isinstance(within_inames, tuple)
ImmutableRecord.__init__(self,
a_matrices=a_matrices,
diff --git a/python/dune/perftool/sumfact/realization.py b/python/dune/perftool/sumfact/realization.py
index 7b84f4f6..739514e8 100644
--- a/python/dune/perftool/sumfact/realization.py
+++ b/python/dune/perftool/sumfact/realization.py
@@ -16,6 +16,7 @@ from dune.perftool.generation import (barrier,
from dune.perftool.loopy.buffer import (get_buffer_temporary,
switch_base_storage,
)
+from dune.perftool.pdelab.geometry import world_dimension
from dune.perftool.options import get_option
from dune.perftool.pdelab.signatures import assembler_routine_name
from dune.perftool.sumfact.permutation import (_sf_permutation_strategy,
@@ -66,11 +67,11 @@ def realize_sum_factorization_kernel(sf, insn_dep=frozenset(), outshape=None, di
dump_accumulate_timer(timer_name)
insn_dep = frozenset({instruction(code="HP_TIMER_START({});".format(timer_name),
depends_on=insn_dep,
- within_inames=sf.within_inames)})
+ within_inames=frozenset(sf.within_inames))})
# Put a barrier before the sumfactorization kernel
insn_dep = frozenset({barrier(depends_on=insn_dep,
- within_inames=sf.within_inames,
+ within_inames=frozenset(sf.within_inames),
)})
# Decide in which order we want to process directions in the
@@ -195,10 +196,9 @@ def realize_sum_factorization_kernel(sf, insn_dep=frozenset(), outshape=None, di
dtype=np.float64,
shape=output_shape + output_shape,
dim_tags=novec_ftags + "," + novec_ftags)
- # TODO the next line should get its inames from
- # elsewhere. This is *NOT* robust (but works right
- # now)
- _ansatz_inames = tuple(Variable(sf.within_inames[i]) for i in range(world_dimension()))
+ # TODO: It is at least questionnable, whether using the *order* of the inames in here
+ # for indexing is a good idea. Then again, it is hard to find an alternative.
+ _ansatz_inames = tuple(prim.Variable(i) for i in sf.within_inames)
assignee = prim.Subscript(prim.Variable(direct_output), _ansatz_inames + output_inames)
# In case of vectorization we need to apply a horizontal add
@@ -215,7 +215,7 @@ def realize_sum_factorization_kernel(sf, insn_dep=frozenset(), outshape=None, di
# at the same time store the instruction ID for the next instruction to depend on
insn_dep = frozenset({instruction(assignee=assignee,
expression=matprod,
- forced_iname_deps=frozenset([iname for iname in out_inames]).union(sf.within_inames),
+ forced_iname_deps=frozenset([iname for iname in out_inames]).union(frozenset(sf.within_inames)),
forced_iname_deps_is_final=True,
depends_on=insn_dep,
)
@@ -225,7 +225,7 @@ def realize_sum_factorization_kernel(sf, insn_dep=frozenset(), outshape=None, di
if get_option("instrumentation_level") >= 4:
insn_dep = frozenset({instruction(code="HP_TIMER_STOP({});".format(timer_name),
depends_on=insn_dep,
- within_inames=sf.within_inames)})
+ within_inames=frozenset(sf.within_inames))})
if sf.stage == 1:
qp_timer_name = assembler_routine_name() + '_kernel' + '_quadratureloop'
post_include('HP_DECLARE_TIMER({});'.format(timer_name), filetag='operatorfile')
diff --git a/python/dune/perftool/sumfact/sumfact.py b/python/dune/perftool/sumfact/sumfact.py
index d46bf6b9..cc05067b 100644
--- a/python/dune/perftool/sumfact/sumfact.py
+++ b/python/dune/perftool/sumfact/sumfact.py
@@ -152,7 +152,7 @@ def generate_accumulation_instruction(visitor, accterm, measure, subdomain_id):
restriction=(accterm.argument.restriction, restriction),
stage=3,
preferred_position=i if accterm.new_indices else None,
- within_inames=frozenset(visitor.inames),
+ within_inames=visitor.inames,
)
# TODO: Move this away!
diff --git a/python/dune/perftool/ufl/visitor.py b/python/dune/perftool/ufl/visitor.py
index 96f1cb85..38b7eee6 100644
--- a/python/dune/perftool/ufl/visitor.py
+++ b/python/dune/perftool/ufl/visitor.py
@@ -98,7 +98,8 @@ class UFL2LoopyVisitor(ModifiedTerminalTracker):
shape = len(element.value_shape())
indices = indices[shape:]
for i in range(len(element.value_shape())):
- self.inames = self.inames + (self.interface.dimension_iname(context='arg', count=i),)
+ if self.interface.dimension_iname(context='arg', count=i) not in self.inames:
+ self.inames = self.inames + (self.interface.dimension_iname(context='arg', count=i),)
# For the purpose of basis evaluation, we need to take the leaf element
leaf_element = element.sub_elements()[0]
@@ -108,7 +109,9 @@ class UFL2LoopyVisitor(ModifiedTerminalTracker):
# Have the issued instruction depend on the iname for this localfunction space
inames = self.interface.lfs_inames(leaf_element, restriction, o.number())
- self.inames = self.inames + inames
+ for iname in inames:
+ if iname not in self.inames:
+ self.inames = self.inames + (iname,)
if self.reference_grad:
return maybe_wrap_subscript(self.interface.pymbolic_reference_gradient(leaf_element, restriction, o.number()), indices)
@@ -215,7 +218,8 @@ class UFL2LoopyVisitor(ModifiedTerminalTracker):
return index._value
else:
if index in self.dimension_indices:
- self.inames = self.inames + (self.dimension_indices[index],)
+ if self.dimension_indices[index] not in self.inames:
+ self.inames = self.inames + (self.dimension_indices[index],)
return Variable(self.dimension_indices[index])
else:
return Variable(self.interface.name_index(index))
--
GitLab