From 48b2991da2b5b15ca85700a70f3ad9556b4884bb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ren=C3=A9=20He=C3=9F?= <rene.hess@iwr.uni-heidelberg.de>
Date: Thu, 26 Jan 2017 10:04:39 +0100
Subject: [PATCH] Add documentation
---
python/dune/perftool/sumfact/sumfact.py | 27 +++++++++++++++----------
1 file changed, 16 insertions(+), 11 deletions(-)
diff --git a/python/dune/perftool/sumfact/sumfact.py b/python/dune/perftool/sumfact/sumfact.py
index 65f26ece..7e2594df 100644
--- a/python/dune/perftool/sumfact/sumfact.py
+++ b/python/dune/perftool/sumfact/sumfact.py
@@ -242,7 +242,9 @@ def generate_accumulation_instruction(visitor, accterm, measure, subdomain_id):
# Construct the expression representing "{r,jac}.accumulate(..)"
accum = name_accumulation_variable(test_lfs.get_restriction() + ansatz_lfs.get_restriction())
-
+ # We can directly accumulate the solution in the last step of
+ # the sumfactorization. Currently this is only implemented for
+ # FastDGGridOperator.
direct_output = None
if get_option('fastdg'):
ft = get_global_context_value("form_type")
@@ -281,19 +283,23 @@ def generate_accumulation_instruction(visitor, accterm, measure, subdomain_id):
)
# In the case of FastDGGridOperator we can write directly into the resiudal/jacobi
- #
- # TODO: At the moment this only works if we do not vectorize
- # (over gradients) because loopy tries to acces a vectorclass
- # variable.
if get_option('fastdg'):
- ft = get_global_context_value("form_type")
- if ft == 'residual' or ft == 'jacobian_apply':
- if get_global_context_value("dry_run", False):
+ # TODO: We accumulate the result directly in the last step
+ # of the sumfactorization. This part of code is only here
+ # because the pymbolic sumfactorization node seems to
+ # vanish if there is no assignement afterwards. In the dry
+ # run we force the node to be present by doing an
+ # assignement here. After the dry run everything works as
+ # expected.
+ #
+ # (Note: This code does the right thing if you do not use direct_output.)
+ if get_global_context_value("dry_run", False):
+ ft = get_global_context_value("form_type")
+ if ft == 'residual' or ft == 'jacobian_apply':
shape = (basis_functions_per_direction(),) * world_dimension()
ftags = ",".join(["f"]*len(shape))
globalarg(accum, dtype=np.float64, shape=shape, dim_tags=ftags)
assignee = prim.Subscript(prim.Variable(accum), tuple(prim.Variable(i) for i in inames))
-
expression = prim.Sum((assignee, result))
instruction(assignee=assignee,
expression=expression,
@@ -301,8 +307,7 @@ def generate_accumulation_instruction(visitor, accterm, measure, subdomain_id):
forced_iname_deps_is_final=True,
depends_on=insn_dep,
)
- else:
- if get_global_context_value("dry_run", False):
+ else:
assert ft == 'jacobian'
shape = (basis_functions_per_direction(),) * (world_dimension() * 2)
ftags = ",".join(["f"] * len(shape))
--
GitLab