From f6b4eae7546db76c808eb37d88b242b36f0b161d Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Wed, 23 Nov 2016 16:40:07 +0100
Subject: [PATCH] Use kernel tag in localoperator code generation
---
python/dune/perftool/loopy/buffer.py | 2 +-
python/dune/perftool/pdelab/localoperator.py | 31 +++++++++++++-------
python/dune/perftool/pdelab/spaces.py | 4 +--
3 files changed, 23 insertions(+), 14 deletions(-)
diff --git a/python/dune/perftool/loopy/buffer.py b/python/dune/perftool/loopy/buffer.py
index bd94d995..e7207c2b 100644
--- a/python/dune/perftool/loopy/buffer.py
+++ b/python/dune/perftool/loopy/buffer.py
@@ -44,7 +44,7 @@ class FlipFlopBuffer(object):
return name
-@generator_factory(item_tags=("kernel", "buffer"), cache_key_generator=lambda i, **kw: i)
+@generator_factory(item_tags=("buffer"), cache_key_generator=lambda i, **kw: i, context_tags=("kernel",))
def initialize_buffer(identifier, base_storage_size=None, num=2):
if base_storage_size is None:
raise PerftoolLoopyError("The buffer for identifier {} has not been initialized.".format(identifier))
diff --git a/python/dune/perftool/pdelab/localoperator.py b/python/dune/perftool/pdelab/localoperator.py
index 1c261d94..d9838de3 100644
--- a/python/dune/perftool/pdelab/localoperator.py
+++ b/python/dune/perftool/pdelab/localoperator.py
@@ -10,6 +10,7 @@ from dune.perftool.generation import (backend,
domain,
dump_accumulate_timer,
get_backend,
+ get_global_context_value,
global_context,
iname,
include_file,
@@ -474,16 +475,21 @@ def generate_kernel(integrals):
visitor = UFL2LoopyVisitor(interface, measure, indexmap)
get_backend(interface="accum_insn")(visitor, term, measure, subdomain_id)
+ tag = get_global_context_value("kernel")
+ return extract_kernel_from_cache(tag)
+
+
+def extract_kernel_from_cache(tag):
# Extract the information, which is needed to create a loopy kernel.
# First extracting it, might be useful to alter it before kernel generation.
from dune.perftool.generation import retrieve_cache_functions, retrieve_cache_items
from dune.perftool.loopy.target import DuneTarget
- domains = [i for i in retrieve_cache_items("domain")]
- instructions = [i for i in retrieve_cache_items("instruction")]
- temporaries = {i.name: i for i in retrieve_cache_items("temporary")}
- arguments = [i for i in retrieve_cache_items("argument")]
- silenced = [l for l in retrieve_cache_items("silenced_warning")]
- transformations = [t for t in retrieve_cache_items("transformation")]
+ domains = [i for i in retrieve_cache_items("{} and domain".format(tag))]
+ instructions = [i for i in retrieve_cache_items("{} and instruction".format(tag))]
+ temporaries = {i.name: i for i in retrieve_cache_items("{} and temporary".format(tag))}
+ arguments = [i for i in retrieve_cache_items("{} and argument".format(tag))]
+ silenced = [l for l in retrieve_cache_items("{} and silenced_warning".format(tag))]
+ transformations = [t for t in retrieve_cache_items("{} and transformation".format(tag))]
# Construct an options object
from loopy import Options
@@ -524,7 +530,7 @@ def generate_kernel(integrals):
raise NotImplementedError("Only vectorizing sumfactoized code right now!")
# Now add the preambles to the kernel
- preambles = [(i, p) for i, p in enumerate(retrieve_cache_items("preamble"))]
+ preambles = [(i, p) for i, p in enumerate(retrieve_cache_items("{} and preamble".format(tag)))]
kernel = kernel.copy(preambles=preambles)
# Do the loopy preprocessing!
@@ -532,7 +538,7 @@ def generate_kernel(integrals):
# All items with the kernel tags can be destroyed once a kernel has been generated
from dune.perftool.generation import delete_cache_items
- delete_cache_items("(not file) and (not clazz)")
+ delete_cache_items(tag)
return kernel
@@ -682,7 +688,8 @@ def generate_localoperator_kernels(formdata, data):
enum_pattern()
pattern_baseclass()
enum_alpha()
- kernel = generate_kernel(form.integrals_by_type(measure))
+ with global_context(kernel=assembler_routine_name()):
+ kernel = generate_kernel(form.integrals_by_type(measure))
# Maybe add numerical differentiation
if get_option("numerical_jacobian"):
@@ -737,7 +744,8 @@ def generate_localoperator_kernels(formdata, data):
with global_context(form_type="jacobian"):
for measure in set(i.integral_type() for i in jacform.integrals()):
with global_context(integral_type=measure):
- kernel = generate_kernel(jacform.integrals_by_type(measure))
+ with global_context(kernel=assembler_routine_name()):
+ kernel = generate_kernel(jacform.integrals_by_type(measure))
operator_kernels[(measure, 'jacobian')] = kernel
# Generate dummy functions for those kernels, that vanished in the differentiation process
@@ -762,7 +770,8 @@ def generate_localoperator_kernels(formdata, data):
with global_context(form_type="jacobian_apply"):
for measure in set(i.integral_type() for i in jac_apply_form.integrals()):
with global_context(integral_type=measure):
- kernel = generate_kernel(jac_apply_form.integrals_by_type(measure))
+ with global_context(kernel=assembler_routine_name()):
+ kernel = generate_kernel(jac_apply_form.integrals_by_type(measure))
operator_kernels[(measure, 'jacobian_apply')] = kernel
# Generate dummy functions for those kernels, that vanished in the differentiation process
diff --git a/python/dune/perftool/pdelab/spaces.py b/python/dune/perftool/pdelab/spaces.py
index 18fe134f..710f86b0 100644
--- a/python/dune/perftool/pdelab/spaces.py
+++ b/python/dune/perftool/pdelab/spaces.py
@@ -99,7 +99,7 @@ def name_leaf_lfs(leaf_element, restriction, val=None):
return val
-@generator_factory(cache_key_generator=lambda e, r, c, **kw: (e, r, c))
+@generator_factory(cache_key_generator=lambda e, r, c, **kw: (e, r, c), context_tags=("kernel",))
def name_lfs(element, restriction, component, prefix=None):
# Omitting the prefix is only valid upon a second call, which will
# result in a cache hit.
@@ -178,7 +178,7 @@ def traverse_lfs_tree(arg):
type_gfs(arg.argexpr.ufl_element(), basetype=gfs_basename, index_stack=())
-@generator_factory(item_tags=("iname",), cache_key_generator=lambda e, r, c: (e, c))
+@generator_factory(item_tags=("iname",), cache_key_generator=lambda e, r, c: (e, c), context_tags=("kernel",))
def _lfs_iname(element, restriction, context):
lfs = name_leaf_lfs(element, restriction)
bound = name_lfs_bound(lfs)
--
GitLab