def analyze_iterations(nodes):
"""
Attach :class:`IterationProperty` to :class:`Iteration` objects within
``nodes``. The recognized IterationProperty decorators are listed in
``nodes.IterationProperty._KNOWN``.
"""
sections = FindSections().visit(nodes)
# Local analysis: detect Iteration properties, inspecting trees in isolation
mapper = OrderedDict()
for tree, exprs in sections.items():
deps_graph = compute_dependency_graph(exprs)
mapper = detect_fully_parallel(tree, deps_graph, mapper)
mapper = detect_outermost_parallel(tree, deps_graph, mapper)
mapper = detect_outermost_sequential_inner_parallel(
tree, deps_graph, mapper)
mapper = detect_innermost_unitstride(tree, deps_graph, mapper)
mapper = detect_wrappable_iterations(tree, deps_graph, mapper)
# Global analysis
for k, v in list(mapper.items()):
args = k.args
# SEQUENTIAL kills PARALLEL
properties = [i for i in v if i != PARALLEL] if SEQUENTIAL in v else v
properties = as_tuple(args.pop('properties')) + as_tuple(properties)
mapper[k] = Iteration(properties=properties, **args)
# Store the discovered properties in the Iteration/Expression tree
processed = NestedTransformer(mapper).visit(nodes)
return processed
def retrieve_iteration_tree(node, mode='normal'):
"""Return a list of all :class:`Iteration` sub-trees rooted in ``node``.
For example, given the Iteration tree:
.. code-block:: c
Iteration i
expr0
Iteration j
Iteraion k
expr1
Iteration p
expr2
Return the list: ::
[(Iteration i, Iteration j, Iteration k), (Iteration i, Iteration p)]
:param node: The searched Iteration/Expression tree.
:param mode: Accepted values are 'normal' (default) and 'superset', in which
case iteration trees that are subset of larger iteration trees
are dropped.
"""
assert mode in ('normal', 'superset')
trees = [i for i in FindSections().visit(node) if i]
if mode == 'normal':
return trees
else:
match = []
for i in trees:
if any(set(i).issubset(set(j)) for j in trees if i != j):
continue
match.append(i)
return match
def test_find_sections(exprs, block1, block2, block3):
finder = FindSections()
sections = finder.visit(block1)
assert len(sections) == 1
sections = finder.visit(block2)
assert len(sections) == 2
found = list(sections.values())
assert len(found[0]) == 1
assert len(found[1]) == 1
sections = finder.visit(block3)
assert len(sections) == 3
found = list(sections.values())
assert len(found[0]) == 1
assert len(found[1]) == 2
assert len(found[2]) == 1
def retrieve_iteration_tree(node, mode='normal'):
"""
A list with all :class:`Iteration` sub-trees within an IET.
Examples
--------
Given the Iteration tree:
.. code-block:: c
Iteration i
expr0
Iteration j
Iteraion k
expr1
Iteration p
expr2
Return the list: ::
[(Iteration i, Iteration j, Iteration k), (Iteration i, Iteration p)]
Parameters
----------
iet : Node
The searched Iteration/Expression tree.
mode : str, optional
- ``normal``
- ``superset``: Iteration trees that are subset of larger iteration trees
are dropped.
"""
assert mode in ('normal', 'superset')
trees = [IterationTree(i) for i in FindSections().visit(node) if i]
if mode == 'normal':
return trees
else:
match = []
for i in trees:
if any(set(i).issubset(set(j)) for j in trees if i != j):
continue
match.append(i)
return IterationTree(match)
def create_profile(node):
"""
Create a :class:`Profiler` for the Iteration/Expression tree ``node``.
The following code sections are profiled: ::
* The whole ``node``;
* A sequence of perfectly nested loops that have common :class:`Iteration`
dimensions, but possibly different extent. For example: ::
for x = 0 to N
..
for x = 1 to N-1
..
Both Iterations have dimension ``x``, and will be profiled as a single
section, though their extent is different.
* Any perfectly nested loops.
"""
profiler = Profiler()
# Group by root Iteration
mapper = OrderedDict()
for itspace in FindSections().visit(node):
mapper.setdefault(itspace[0], []).append(itspace)
# Group sections if their iteration spaces overlap
key = lambda itspace: set([i.dim for i in itspace])
found = []
for v in mapper.values():
queue = list(v)
handle = []
while queue:
item = queue.pop(0)
if not handle or key(item) == key(handle[0]):
handle.append(item)
else:
# Found a timing section
found.append(tuple(handle))
handle = [item]
if handle:
found.append(tuple(handle))
# Create and track C-level timers
mapper = OrderedDict()
for i, group in enumerate(found):
name = 'section_%d' % i
# We time at the single timestep level
for i in zip(*group):
root = i[0]
remainder = tuple(j for j in i if j is not root)
if not (root.dim.is_Time or root.dim.is_Stepping):
break
# Prepare to transform the Iteration/Expression tree
body = (root, ) + remainder
mapper[root] = TimedList(gname=profiler.varname, lname=name, body=body)
mapper.update(OrderedDict([(j, None) for j in remainder]))
# Estimate computational properties of the profiled section
expressions = FindNodes(Expression).visit(body)
ops = estimate_cost([e.expr for e in expressions])
memory = estimate_memory([e.expr for e in expressions])
# Keep track of the new profiled section
profiler.add(name, group[0], ops, memory)
# Transform the Iteration/Expression tree introducing the C-level timers
processed = Transformer(mapper).visit(node)
return processed, profiler
