diff --git a/MEMO b/MEMO
index 148999e6ae10f8030b1585287d67648161473f70..dd00893d8c7085ccd78cdd6adb103eed194f0059 100644
--- a/MEMO
+++ b/MEMO
@@ -42,9 +42,6 @@ Things to consider
To-do
^^^^^
-- Just touching a variable written to by a non-idempotent
- instruction makes that instruction also not idempotent
-
- user interface for dim length prescription
- Deal with equality constraints.
@@ -89,6 +86,11 @@ Future ideas
Dealt with
^^^^^^^^^^
+- Just touching a variable written to by a non-idempotent
+ instruction makes that instruction also not idempotent
+ -> Idempotent renamed to boostable.
+ -> Done.
+
- Give the user control over which reduction inames are
duplicated.
diff --git a/loopy/kernel.py b/loopy/kernel.py
index 9af083c2e3dadfafaa174e492ff31ad145468b39..2eff255384969f11d7b37818e7f383da36c3c6c1 100644
--- a/loopy/kernel.py
+++ b/loopy/kernel.py
@@ -223,8 +223,9 @@ class Instruction(Record):
:ivar insn_deps: a list of ids of :class:`Instruction` instances that
*must* be executed before this one. Note that loop scheduling augments this
by adding dependencies on any writes to temporaries read by this instruction.
- :ivar idempotent: Whether the instruction may be executed repeatedly (while obeying
- dependencies) without changing the meaning of the program.
+ :ivar boostable: Whether the instruction may safely be executed
+ inside more loops than advertised without changing the meaning
+ of the program. Allowed values are *None* (for unknwon), *True*, and *False*.
The following two instance variables are only used until :func:`loopy.kernel.make_kernel` is
finished:
@@ -236,13 +237,13 @@ class Instruction(Record):
"""
def __init__(self,
id, assignee, expression,
- forced_iname_deps=[], insn_deps=[], idempotent=None,
+ forced_iname_deps=[], insn_deps=[], boostable=None,
temp_var_type=None, duplicate_inames_and_tags=[]):
Record.__init__(self,
id=id, assignee=assignee, expression=expression,
forced_iname_deps=forced_iname_deps,
- insn_deps=insn_deps, idempotent=idempotent,
+ insn_deps=insn_deps, boostable=boostable,
temp_var_type=temp_var_type, duplicate_inames_and_tags=duplicate_inames_and_tags)
@memoize_method
@@ -268,14 +269,14 @@ class Instruction(Record):
result = "%s: %s <- %s\n [%s]" % (self.id,
self.assignee, self.expression, ", ".join(sorted(self.all_inames())))
- if self.idempotent == True:
- result += " (idempotent)"
- elif self.idempotent == False:
- result += " (not idempotent)"
- elif self.idempotent is None:
- result += " (idempotence unknown)"
+ if self.boostable == True:
+ result += " (boostable)"
+ elif self.boostable == False:
+ result += " (not boostable)"
+ elif self.boostable is None:
+ result += " (boostability unknown)"
else:
- raise RuntimeError("unexpected value for Instruction.idempotent")
+ raise RuntimeError("unexpected value for Instruction.boostable")
if self.insn_deps:
result += "\n : " + ", ".join(self.insn_deps)
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 6f02fd2af542d379522ac12f91f5f967ebb19201..a06e28b4eff24e6451281e3b70c3a86f7697e59a 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -98,7 +98,7 @@ def realize_reduction(kernel):
# }}}
-# {{{ automatic dependencies, find idempotent instructions
+# {{{ automatic dependencies, find boostability of instructions
def find_accessors(kernel, readers):
"""
@@ -129,7 +129,7 @@ def find_accessors(kernel, readers):
-def add_idempotence_and_automatic_dependencies(kernel):
+def add_boostability_and_automatic_dependencies(kernel):
writer_map = find_accessors(kernel, readers=False)
arg_names = set(arg.name for arg in kernel.args)
@@ -145,11 +145,14 @@ def add_idempotence_and_automatic_dependencies(kernel):
set(var.name for var in dm(insn.expression))
& var_names)
+ non_boostable_vars = set()
+
new_insns = []
for insn in kernel.instructions:
auto_deps = []
# {{{ add automatic dependencies
+
all_my_var_writers = set()
for var in dep_map[insn.id]:
var_writers = writer_map.get(var, set())
@@ -171,7 +174,7 @@ def add_idempotence_and_automatic_dependencies(kernel):
# }}}
- # {{{ find dependency loops, flag idempotence
+ # {{{ find dependency loops, flag boostability
while True:
last_all_my_var_writers = all_my_var_writers
@@ -185,12 +188,30 @@ def add_idempotence_and_automatic_dependencies(kernel):
# }}}
+ boostable = insn.id not in all_my_var_writers
+
+ if not boostable:
+ non_boostable_vars.add(insn.get_assignee_var_name())
+
new_insns.append(
insn.copy(
insn_deps=insn.insn_deps + auto_deps,
- idempotent=insn.id not in all_my_var_writers))
+ boostable=boostable))
+
+ # {{{ remove boostability from isns that access non-boostable vars
+
+ new2_insns = []
+ for insn in new_insns:
+ accessed_vars = (
+ set([insn.get_assignee_var_name()])
+ | insn.get_read_var_names())
+
+ boostable = insn.boostable and not bool(non_boostable_vars & accessed_vars)
+ new2_insns.append(insn.copy(boostable=boostable))
+
+ # }}}
- return kernel.copy(instructions=new_insns)
+ return kernel.copy(instructions=new2_insns)
# }}}
@@ -480,7 +501,7 @@ def preprocess_kernel(kernel):
# }}}
kernel = assign_automatic_axes(kernel)
- kernel = add_idempotence_and_automatic_dependencies(kernel)
+ kernel = add_boostability_and_automatic_dependencies(kernel)
kernel = adjust_local_temp_var_storage(kernel)
import loopy.check as chk
diff --git a/loopy/schedule.py b/loopy/schedule.py
index d9e52dfa8aedacf003f3b3190dc57ff5dacb2297..d6129567666f0b5464e0db9e1272dfecd9b22c98 100644
--- a/loopy/schedule.py
+++ b/loopy/schedule.py
@@ -191,8 +191,8 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[]):
schedule_now = set(insn.insn_deps) <= scheduled_insn_ids
- if insn.idempotent == True:
- # If insn is idempotent, it may be placed inside a more deeply
+ if insn.boostable == True:
+ # If insn is boostable, it may be placed inside a more deeply
# nested loop without harm.
# But if it can be scheduled on the way *out* of the currently
@@ -212,8 +212,8 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[]):
if schedulable_at_loop_levels != [len(active_inames)]:
schedule_now = False
- elif insn.idempotent == False:
- # If insn is not idempotent, we must insist that it is placed inside
+ elif insn.boostable == False:
+ # If insn is not boostable, we must insist that it is placed inside
# the exactly correct set of loops.
schedule_now = schedule_now and (
@@ -222,7 +222,7 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[]):
active_inames_set - parallel_inames)
else:
- raise RuntimeError("instruction '%s' has undetermined idempotence"
+ raise RuntimeError("instruction '%s' has undetermined boostability"
% insn.id)
if schedule_now: