From ded9342c7b96bf5bf6e28c7b8f0bbe4eb11ac00b Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Sun, 5 Jun 2016 01:43:33 -0500
Subject: [PATCH] Deprecate iname infercence. Do iname inference once up front
and never again, fix transformations to work without it, and complain if
iname inference actually does anything
---
loopy/kernel/__init__.py | 14 +++++++-------
loopy/kernel/creation.py | 27 +++++++++++++++++++++++++++
loopy/kernel/data.py | 18 ++----------------
loopy/kernel/tools.py | 26 ++++++++++++++++++--------
loopy/preprocess.py | 12 ++----------
loopy/transform/batch.py | 10 +++++++++-
loopy/transform/buffer.py | 9 +++++++--
loopy/transform/diff.py | 4 +++-
loopy/transform/iname.py | 25 ++++++++++++++++++++++---
loopy/transform/precompute.py | 17 ++++++++++++-----
loopy/version.py | 2 +-
11 files changed, 110 insertions(+), 54 deletions(-)
diff --git a/loopy/kernel/__init__.py b/loopy/kernel/__init__.py
index f9049ee43..3d64d227e 100644
--- a/loopy/kernel/__init__.py
+++ b/loopy/kernel/__init__.py
@@ -669,9 +669,11 @@ class LoopKernel(RecordWithoutPickling):
"""Return a mapping from instruction ids to inames inside which
they should be run.
"""
+ result = {}
+ for insn in self.instructions:
+ result[insn.id] = insn.forced_iname_deps
- from loopy.kernel.tools import find_all_insn_inames
- return find_all_insn_inames(self)
+ return result
@memoize_method
def all_referenced_inames(self):
@@ -681,11 +683,9 @@ class LoopKernel(RecordWithoutPickling):
return result
def insn_inames(self, insn):
- from loopy.kernel.data import InstructionBase
- if isinstance(insn, InstructionBase):
- return self.all_insn_inames()[insn.id]
- else:
- return self.all_insn_inames()[insn]
+ if isinstance(insn, str):
+ insn = self.id_to_insn[insn]
+ return insn.forced_iname_deps
@memoize_method
def iname_to_insns(self):
diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index 1c988e34d..ec3b8b6ef 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -1084,6 +1084,19 @@ def resolve_wildcard_deps(knl):
# }}}
+# {{{ add inferred iname deps
+
+def add_inferred_inames(knl):
+ from loopy.kernel.tools import find_all_insn_inames
+ insn_inames = find_all_insn_inames(knl)
+
+ return knl.copy(instructions=[
+ insn.copy(forced_iname_deps=insn_inames[insn.id])
+ for insn in knl.instructions])
+
+# }}}
+
+
# {{{ kernel creation top-level
def make_kernel(domains, instructions, kernel_data=["..."], **kwargs):
@@ -1302,6 +1315,20 @@ def make_kernel(domains, instructions, kernel_data=["..."], **kwargs):
check_for_nonexistent_iname_deps(knl)
knl = create_temporaries(knl, default_order)
+ # -------------------------------------------------------------------------
+ # Ordering dependency:
+ # -------------------------------------------------------------------------
+ # Must create temporaries before inferring inames (because those temporaries
+ # mediate dependencies that are then used for iname propagation.)
+ # -------------------------------------------------------------------------
+ # NOTE: add_inferred_inames will be phased out and throws warnings if it
+ # does something.
+ knl = add_inferred_inames(knl)
+ # -------------------------------------------------------------------------
+ # Ordering dependency:
+ # -------------------------------------------------------------------------
+ # Must infer inames before determining shapes.
+ # -------------------------------------------------------------------------
knl = determine_shapes_of_temporaries(knl)
knl = expand_defines_in_shapes(knl, defines)
knl = guess_arg_shape_if_requested(knl, default_order)
diff --git a/loopy/kernel/data.py b/loopy/kernel/data.py
index 3ba3f1918..44a5618e7 100644
--- a/loopy/kernel/data.py
+++ b/loopy/kernel/data.py
@@ -561,15 +561,14 @@ class InstructionBase(Record):
"groups conflicts_with_groups "
"no_sync_with "
"predicates "
- "forced_iname_deps_is_final stop_iname_dep_propagation "
- "stop_iname_dep_propagation "
+ "forced_iname_deps_is_final forced_iname_deps "
"priority boostable boostable_into".split())
def __init__(self, id, depends_on, depends_on_is_final,
groups, conflicts_with_groups,
no_sync_with,
forced_iname_deps_is_final, forced_iname_deps,
- stop_iname_dep_propagation, priority,
+ priority,
boostable, boostable_into, predicates, tags,
insn_deps=None, insn_deps_is_final=None):
@@ -598,9 +597,6 @@ class InstructionBase(Record):
if forced_iname_deps_is_final is None:
forced_iname_deps_is_final = False
- if stop_iname_dep_propagation is None:
- stop_iname_dep_propagation = frozenset()
-
if depends_on_is_final is None:
depends_on_is_final = False
@@ -623,7 +619,6 @@ class InstructionBase(Record):
# assert all(is_interned(pred) for pred in predicates)
assert isinstance(forced_iname_deps, frozenset)
- assert isinstance(stop_iname_dep_propagation, frozenset)
assert isinstance(depends_on, frozenset) or depends_on is None
assert isinstance(groups, frozenset)
assert isinstance(conflicts_with_groups, frozenset)
@@ -636,7 +631,6 @@ class InstructionBase(Record):
groups=groups, conflicts_with_groups=conflicts_with_groups,
forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
- stop_iname_dep_propagation=stop_iname_dep_propagation,
priority=priority,
boostable=boostable,
boostable_into=boostable_into,
@@ -808,8 +802,6 @@ class InstructionBase(Record):
intern_frozenset_of_ids(self.forced_iname_deps))
self.predicates = (
intern_frozenset_of_ids(self.predicates))
- self.stop_iname_dep_propagation = (
- intern_frozenset_of_ids(self.stop_iname_dep_propagation))
# }}}
@@ -1101,7 +1093,6 @@ class Assignment(MultiAssignmentBase):
no_sync_with=None,
forced_iname_deps_is_final=None,
forced_iname_deps=frozenset(),
- stop_iname_dep_propagation=None,
boostable=None, boostable_into=None, tags=None,
temp_var_type=None, atomicity=(),
priority=0, predicates=frozenset(),
@@ -1116,7 +1107,6 @@ class Assignment(MultiAssignmentBase):
no_sync_with=no_sync_with,
forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
- stop_iname_dep_propagation=stop_iname_dep_propagation,
boostable=boostable,
boostable_into=boostable_into,
priority=priority,
@@ -1245,7 +1235,6 @@ class CallInstruction(MultiAssignmentBase):
no_sync_with=None,
forced_iname_deps_is_final=None,
forced_iname_deps=frozenset(),
- stop_iname_dep_propagation=None,
boostable=None, boostable_into=None, tags=None,
temp_var_types=None,
priority=0, predicates=frozenset(),
@@ -1260,7 +1249,6 @@ class CallInstruction(MultiAssignmentBase):
no_sync_with=no_sync_with,
forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
- stop_iname_dep_propagation=stop_iname_dep_propagation,
boostable=boostable,
boostable_into=boostable_into,
priority=priority,
@@ -1421,7 +1409,6 @@ class CInstruction(InstructionBase):
groups=None, conflicts_with_groups=None,
no_sync_with=None,
forced_iname_deps_is_final=None, forced_iname_deps=frozenset(),
- stop_iname_dep_propagation=None,
priority=0, boostable=None, boostable_into=None,
predicates=frozenset(), tags=None,
insn_deps=None, insn_deps_is_final=None):
@@ -1443,7 +1430,6 @@ class CInstruction(InstructionBase):
no_sync_with=no_sync_with,
forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
- stop_iname_dep_propagation=stop_iname_dep_propagation,
boostable=boostable,
boostable_into=boostable_into,
priority=priority, predicates=predicates, tags=tags,
diff --git a/loopy/kernel/tools.py b/loopy/kernel/tools.py
index 5a100d712..1775032cb 100644
--- a/loopy/kernel/tools.py
+++ b/loopy/kernel/tools.py
@@ -30,7 +30,7 @@ from six.moves import intern
import numpy as np
import islpy as isl
from islpy import dim_type
-from loopy.diagnostic import LoopyError
+from loopy.diagnostic import LoopyError, warn_with_kernel
import logging
logger = logging.getLogger(__name__)
@@ -164,9 +164,7 @@ def find_all_insn_inames(kernel):
))
insn_id_to_inames[insn.id] = iname_deps
- insn_assignee_inames[insn.id] = (
- write_deps & kernel.all_inames()
- | insn.stop_iname_dep_propagation)
+ insn_assignee_inames[insn.id] = write_deps & kernel.all_inames()
written_vars = kernel.get_written_variables()
@@ -216,8 +214,14 @@ def find_all_insn_inames(kernel):
if inames_new != inames_old:
did_something = True
- logger.debug("%s: find_all_insn_inames: %s -> %s (dep-based)" % (
- kernel.name, insn.id, ", ".join(sorted(inames_new))))
+
+ warn_with_kernel(kernel, "inferred_iname",
+ "The iname(s) '%s' on instruction '%s' was "
+ "automatically added. "
+ "This is deprecated. Please add the iname "
+ "to the instruction "
+ "implicitly, e.g. by adding '{inames=...}"
+ % (inames_new-inames_old, insn.id))
# }}}
@@ -247,8 +251,14 @@ def find_all_insn_inames(kernel):
if inames_new != inames_old:
did_something = True
insn_id_to_inames[insn.id] = frozenset(inames_new)
- logger.debug("%s: find_all_insn_inames: %s -> %s (domain-based)" % (
- kernel.name, insn.id, ", ".join(sorted(inames_new))))
+
+ warn_with_kernel(kernel, "inferred_iname",
+ "The iname(s) '%s' on instruction '%s' was "
+ "automatically added. "
+ "This is deprecated. Please add the iname "
+ "to the instruction "
+ "implicitly, e.g. by adding '{inames=...}"
+ % (inames_new-inames_old, insn.id))
# }}}
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 5690218e6..d333cdf18 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -544,8 +544,7 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True):
expr.expr, expr.inames),
depends_on=frozenset([init_insn.id]) | insn.depends_on,
forced_iname_deps=update_insn_iname_deps,
- forced_iname_deps_is_final=insn.forced_iname_deps_is_final,
- stop_iname_dep_propagation=frozenset(expr.inames))
+ forced_iname_deps_is_final=insn.forced_iname_deps_is_final)
generated_insns.append(reduction_insn)
@@ -716,8 +715,6 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True):
new_insn_add_depends_on.add(prev_id)
new_insn_add_no_sync_with.add(prev_id)
new_insn_add_forced_iname_deps.add(stage_exec_iname or base_exec_iname)
- new_insn_add_stop_iname_dep_propagation.add(
- stage_exec_iname or base_exec_iname)
if nresults == 1:
assert len(acc_vars) == 1
@@ -828,7 +825,6 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True):
new_insn_add_depends_on = set()
new_insn_add_no_sync_with = set()
new_insn_add_forced_iname_deps = set()
- new_insn_add_stop_iname_dep_propagation = set()
generated_insns = []
@@ -859,11 +855,7 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True):
| frozenset(new_insn_add_no_sync_with),
forced_iname_deps=(
temp_kernel.insn_inames(insn)
- | new_insn_add_forced_iname_deps),
- stop_iname_dep_propagation=(
- insn.stop_iname_dep_propagation
- | new_insn_add_stop_iname_dep_propagation),
- )
+ | new_insn_add_forced_iname_deps))
kwargs.pop("id")
kwargs.pop("expression")
diff --git a/loopy/transform/batch.py b/loopy/transform/batch.py
index 967e14de6..97c15c6f2 100644
--- a/loopy/transform/batch.py
+++ b/loopy/transform/batch.py
@@ -143,9 +143,17 @@ def to_batched(knl, nbatches, batch_varying_args, batch_iname_prefix="ibatch",
bvc = _BatchVariableChanger(rule_mapping_context,
knl, batch_varying_args, batch_iname_expr,
sequential=sequential)
- return rule_mapping_context.finish_kernel(
+ kernel = rule_mapping_context.finish_kernel(
bvc.map_kernel(knl))
+ batch_iname_set = frozenset([batch_iname])
+ kernel = kernel.copy(
+ instructions=[
+ insn.copy(forced_iname_deps=insn.forced_iname_deps | batch_iname_set)
+ for insn in kernel.instructions])
+
+ return kernel
+
# }}}
# vim: foldmethod=marker
diff --git a/loopy/transform/buffer.py b/loopy/transform/buffer.py
index fb32b3ce8..0c25b6393 100644
--- a/loopy/transform/buffer.py
+++ b/loopy/transform/buffer.py
@@ -400,7 +400,9 @@ def buffer_array(kernel, var_name, buffer_inames, init_expression=None,
init_instruction = Assignment(id=init_insn_id,
assignee=buf_var_init,
expression=init_expression,
- forced_iname_deps=frozenset(within_inames),
+ forced_iname_deps=(
+ frozenset(within_inames)
+ | frozenset(non1_init_inames)),
depends_on=frozenset(),
depends_on_is_final=True)
@@ -475,9 +477,12 @@ def buffer_array(kernel, var_name, buffer_inames, init_expression=None,
store_instruction = Assignment(
id=kernel.make_unique_instruction_id(based_on="store_"+var_name),
depends_on=frozenset(aar.modified_insn_ids),
+ no_sync_with=frozenset([init_insn_id]),
assignee=store_target,
expression=store_expression,
- forced_iname_deps=frozenset(within_inames))
+ forced_iname_deps=(
+ frozenset(within_inames)
+ | frozenset(non1_store_inames)))
else:
did_write = False
diff --git a/loopy/transform/diff.py b/loopy/transform/diff.py
index 1f75a60b8..a8c3b5849 100644
--- a/loopy/transform/diff.py
+++ b/loopy/transform/diff.py
@@ -310,7 +310,9 @@ class DifferentiationContext(object):
id=new_insn_id,
assignee=var(new_var_name)[
lhs_ind + diff_iname_exprs],
- expression=diff_expr)
+ expression=diff_expr,
+ forced_iname_deps=(
+ orig_writer_insn.forced_iname_deps | frozenset(diff_inames)))
self.new_instructions.append(insn)
diff --git a/loopy/transform/iname.py b/loopy/transform/iname.py
index 12749a303..4c3cd0a69 100644
--- a/loopy/transform/iname.py
+++ b/loopy/transform/iname.py
@@ -1142,8 +1142,8 @@ def affine_map_inames(kernel, old_inames, new_inames, equations):
# {{{ change domains
- new_inames_set = set(new_inames)
- old_inames_set = set(old_inames)
+ new_inames_set = frozenset(new_inames)
+ old_inames_set = frozenset(old_inames)
new_domains = []
for idom, dom in enumerate(kernel.domains):
@@ -1229,7 +1229,26 @@ def affine_map_inames(kernel, old_inames, new_inames, equations):
# }}}
- return kernel.copy(domains=new_domains)
+ # {{{ switch iname refs in instructions
+
+ def fix_iname_set(insn_id, inames):
+ if old_inames_set <= inames:
+ return (inames - old_inames_set) | new_inames_set
+ elif old_inames_set & inames:
+ raise LoopyError("instruction '%s' uses only a part (%s), not all, "
+ "of the old inames"
+ % (insn_id, ", ".join(old_inames_set & inames)))
+ else:
+ return inames
+
+ new_instructions = [
+ insn.copy(forced_iname_deps=fix_iname_set(
+ insn.id, insn.forced_iname_deps))
+ for insn in kernel.instructions]
+
+ # }}}
+
+ return kernel.copy(domains=new_domains, instructions=new_instructions)
# }}}
diff --git a/loopy/transform/precompute.py b/loopy/transform/precompute.py
index fd6f33efc..f9d71b9f1 100644
--- a/loopy/transform/precompute.py
+++ b/loopy/transform/precompute.py
@@ -480,8 +480,9 @@ def precompute(kernel, subst_use, sweep_inames=[], within=None,
from loopy.symbolic import SubstitutionRuleExpander
submap = SubstitutionRuleExpander(kernel.substitutions)
- value_inames = get_dependencies(
- submap(subst.expression)
+ value_inames = (
+ get_dependencies(submap(subst.expression))
+ - frozenset(subst.arguments)
) & kernel.all_inames()
if value_inames - expanding_usage_arg_deps < extra_storage_axes:
raise RuntimeError("unreferenced sweep inames specified: "
@@ -728,8 +729,8 @@ def precompute(kernel, subst_use, sweep_inames=[], within=None,
assignee = var(temporary_name)
if non1_storage_axis_names:
- assignee = assignee.index(
- tuple(var(iname) for iname in non1_storage_axis_names))
+ assignee = assignee[
+ tuple(var(iname) for iname in non1_storage_axis_names)]
# {{{ process substitutions on compute instruction
@@ -764,7 +765,13 @@ def precompute(kernel, subst_use, sweep_inames=[], within=None,
compute_insn = Assignment(
id=compute_insn_id,
assignee=assignee,
- expression=compute_expression)
+ expression=compute_expression,
+ forced_iname_deps=(
+ frozenset(non1_storage_axis_names)
+ | frozenset(
+ (expanding_usage_arg_deps | value_inames)
+ - sweep_inames_set))
+ )
# }}}
diff --git a/loopy/version.py b/loopy/version.py
index bb1b123cb..0fcb9c04d 100644
--- a/loopy/version.py
+++ b/loopy/version.py
@@ -32,4 +32,4 @@ except ImportError:
else:
_islpy_version = islpy.version.VERSION_TEXT
-DATA_MODEL_VERSION = "v35-islpy%s" % _islpy_version
+DATA_MODEL_VERSION = "v36-islpy%s" % _islpy_version
--
GitLab