From e4467b6f0c78b52acf3c0e01eb9372667101be8b Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Fri, 31 Mar 2017 11:56:06 +0200
Subject: [PATCH] fixup
---
python/dune/perftool/sumfact/realization.py | 48 +++++++++++----------
1 file changed, 26 insertions(+), 22 deletions(-)
diff --git a/python/dune/perftool/sumfact/realization.py b/python/dune/perftool/sumfact/realization.py
index 9032ad1b..6d99175e 100644
--- a/python/dune/perftool/sumfact/realization.py
+++ b/python/dune/perftool/sumfact/realization.py
@@ -33,30 +33,14 @@ import numpy as np
import pymbolic.primitives as prim
-@generator_factory(item_tags=("sumfactkernel",),
- context_tags=("kernel",),
- cache_key_generator=lambda s, **kw: s.cache_key)
-def realize_sum_factorization_kernel(sf, insn_dep=frozenset(), outshape=None, direct_input=None, direct_output=None):
- # Unify the insn_dep parameter to be a frozenset
- if isinstance(insn_dep, str):
- insn_dep = frozenset({insn_dep})
- assert isinstance(insn_dep, frozenset)
-
- # Get the vectorization info. During dry run, this is a now op
-# sf = attach_vectorization_info(sf)
- if get_global_context_value("dry_run", False):
- # During the dry run, we just return the kernel as passed into this
- # function. After the dry run, it can be used to attach information
- # about vectorization.
- return sf, insn_dep
-
- # Get the instruction dependencies of the sumfact kernel. This variable will be
- # updated throughout this function.
- insn_dep = insn_dep.union(sf.insn_dep)
+def realize_sum_factorization_kernel(sf, **kwargs):
+ insn_dep = kwargs.pop('insn_dep', frozenset())
+ if not get_global_context_value("dry_run", False):
+ insn_dep = _realize_input(sf, insn_dep)
+ return _realize_sum_factorization_kernel(sf, insn_dep=insn_dep, **kwargs)
- # Define some helper constructs that make it easier to write generic code later
- vecindex = () if sf.index is None else (sf.index,)
+def _realize_input(sf, insn_dep):
# Set up the input for stage 1
if sf.stage == 1 and not get_option("fastdg"):
assert sf.coeff_func
@@ -71,10 +55,30 @@ def realize_sum_factorization_kernel(sf, insn_dep=frozenset(), outshape=None, di
basisiname = sumfact_iname(name_lfs_bound(lfs), "basis")
container = sf.coeff_func(sf.restriction)
coeff = pymbolic_coefficient(container, lfs, basisiname)
+ vecindex = () if sf.index is None else (sf.index,)
assignee = prim.Subscript(prim.Variable(input_setup), (prim.Variable(basisiname),) + vecindex)
insn_dep = instruction(assignee=assignee,
expression=coeff,
+ depends_on = insn_dep,
)
+ return insn_dep
+
+
+@generator_factory(item_tags=("sumfactkernel",),
+ context_tags=("kernel",),
+ cache_key_generator=lambda s, **kw: s.cache_key)
+def _realize_sum_factorization_kernel(sf, insn_dep=frozenset(), outshape=None, direct_input=None, direct_output=None):
+ # Get the vectorization info. During dry run, this is a now op
+# sf = attach_vectorization_info(sf)
+ if get_global_context_value("dry_run", False):
+ # During the dry run, we just return the kernel as passed into this
+ # function. After the dry run, it can be used to attach information
+ # about vectorization.
+ return sf, insn_dep
+
+ # Get the instruction dependencies of the sumfact kernel. This variable will be
+ # updated throughout this function.
+ insn_dep = insn_dep.union(sf.insn_dep)
# Prepare some dim_tags/shapes for later use
ftags = ",".join(["f"] * sf.length)
--
GitLab