def __init__(self, codegen_state, type_inf_mapper=None):
self.kernel = codegen_state.kernel
self.codegen_state = codegen_state
if type_inf_mapper is None:
type_inf_mapper = TypeInferenceMapper(self.kernel)
self.type_inf_mapper = type_inf_mapper
def __init__(self, codegen_state, fortran_abi=False, type_inf_mapper=None):
self.kernel = codegen_state.kernel
self.codegen_state = codegen_state
if type_inf_mapper is None:
type_inf_mapper = TypeInferenceMapper(self.kernel)
self.type_inf_mapper = type_inf_mapper
self.allow_complex = codegen_state.allow_complex
self.fortran_abi = fortran_abi
def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True):
"""Rewrites reductions into their imperative form. With *insn_id_filter*
specified, operate only on the instruction with an instruction id matching
*insn_id_filter*.
If *insn_id_filter* is given, only the outermost level of reductions will be
expanded, inner reductions will be left alone (because they end up in a new
instruction with a different ID, which doesn't match the filter).
If *insn_id_filter* is not given, all reductions in all instructions will
be realized.
"""
logger.debug("%s: realize reduction" % kernel.name)
new_insns = []
new_iname_tags = {}
insn_id_gen = kernel.get_instruction_id_generator()
var_name_gen = kernel.get_var_name_generator()
new_temporary_variables = kernel.temporary_variables.copy()
from loopy.type_inference import TypeInferenceMapper
type_inf_mapper = TypeInferenceMapper(kernel)
# {{{ sequential
def map_reduction_seq(expr, rec, nresults, arg_dtype,
reduction_dtypes):
outer_insn_inames = temp_kernel.insn_inames(insn)
from pymbolic import var
acc_var_names = [
var_name_gen("acc_"+"_".join(expr.inames))
for i in range(nresults)]
acc_vars = tuple(var(n) for n in acc_var_names)
from loopy.kernel.data import TemporaryVariable, temp_var_scope
for name, dtype in zip(acc_var_names, reduction_dtypes):
new_temporary_variables[name] = TemporaryVariable(
name=name,
shape=(),
dtype=dtype,
scope=temp_var_scope.PRIVATE)
init_id = insn_id_gen(
"%s_%s_init" % (insn.id, "_".join(expr.inames)))
init_insn = make_assignment(
id=init_id,
assignees=acc_vars,
within_inames=outer_insn_inames - frozenset(expr.inames),
within_inames_is_final=insn.within_inames_is_final,
depends_on=frozenset(),
expression=expr.operation.neutral_element(arg_dtype, expr.inames))
generated_insns.append(init_insn)
update_id = insn_id_gen(
based_on="%s_%s_update" % (insn.id, "_".join(expr.inames)))
update_insn_iname_deps = temp_kernel.insn_inames(insn) | set(expr.inames)
if insn.within_inames_is_final:
update_insn_iname_deps = insn.within_inames | set(expr.inames)
reduction_insn = make_assignment(
id=update_id,
assignees=acc_vars,
expression=expr.operation(
arg_dtype,
acc_vars if len(acc_vars) > 1 else acc_vars[0],
expr.expr, expr.inames),
depends_on=frozenset([init_insn.id]) | insn.depends_on,
within_inames=update_insn_iname_deps,
within_inames_is_final=insn.within_inames_is_final)
generated_insns.append(reduction_insn)
new_insn_add_depends_on.add(reduction_insn.id)
if nresults == 1:
assert len(acc_vars) == 1
return acc_vars[0]
else:
return acc_vars
# }}}
# {{{ local-parallel
def _get_int_iname_size(iname):
from loopy.isl_helpers import static_max_of_pw_aff
from loopy.symbolic import pw_aff_to_expr
size = pw_aff_to_expr(
static_max_of_pw_aff(
kernel.get_iname_bounds(iname).size,
constants_only=True))
assert isinstance(size, six.integer_types)
return size
def _make_slab_set(iname, size):
v = isl.make_zero_and_vars([iname])
bs, = (
v[0].le_set(v[iname])
&
v[iname].lt_set(v[0] + size)).get_basic_sets()
return bs
def map_reduction_local(expr, rec, nresults, arg_dtype,
reduction_dtypes):
red_iname, = expr.inames
size = _get_int_iname_size(red_iname)
outer_insn_inames = temp_kernel.insn_inames(insn)
from loopy.kernel.data import LocalIndexTagBase
outer_local_inames = tuple(
oiname
for oiname in outer_insn_inames
if isinstance(
kernel.iname_to_tag.get(oiname),
LocalIndexTagBase))
from pymbolic import var
outer_local_iname_vars = tuple(
var(oiname) for oiname in outer_local_inames)
outer_local_iname_sizes = tuple(
_get_int_iname_size(oiname)
for oiname in outer_local_inames)
# {{{ add separate iname to carry out the reduction
# Doing this sheds any odd conditionals that may be active
# on our red_iname.
base_exec_iname = var_name_gen("red_"+red_iname)
domains.append(_make_slab_set(base_exec_iname, size))
new_iname_tags[base_exec_iname] = kernel.iname_to_tag[red_iname]
# }}}
neutral_var_names = [
var_name_gen("neutral_"+red_iname)
for i in range(nresults)]
acc_var_names = [
var_name_gen("acc_"+red_iname)
for i in range(nresults)]
acc_vars = tuple(var(n) for n in acc_var_names)
from loopy.kernel.data import TemporaryVariable, temp_var_scope
for name, dtype in zip(acc_var_names, reduction_dtypes):
new_temporary_variables[name] = TemporaryVariable(
name=name,
shape=outer_local_iname_sizes + (size,),
dtype=dtype,
scope=temp_var_scope.LOCAL)
for name, dtype in zip(neutral_var_names, reduction_dtypes):
new_temporary_variables[name] = TemporaryVariable(
name=name,
shape=(),
dtype=dtype,
scope=temp_var_scope.PRIVATE)
base_iname_deps = outer_insn_inames - frozenset(expr.inames)
neutral = expr.operation.neutral_element(arg_dtype, expr.inames)
init_id = insn_id_gen("%s_%s_init" % (insn.id, red_iname))
init_insn = make_assignment(
id=init_id,
assignees=tuple(
acc_var[outer_local_iname_vars + (var(base_exec_iname),)]
for acc_var in acc_vars),
expression=neutral,
within_inames=base_iname_deps | frozenset([base_exec_iname]),
within_inames_is_final=insn.within_inames_is_final,
depends_on=frozenset())
generated_insns.append(init_insn)
def _strip_if_scalar(c):
if len(acc_vars) == 1:
return c[0]
else:
return c
init_neutral_id = insn_id_gen("%s_%s_init_neutral" % (insn.id, red_iname))
init_neutral_insn = make_assignment(
id=init_neutral_id,
assignees=tuple(var(nvn) for nvn in neutral_var_names),
expression=neutral,
within_inames=base_iname_deps | frozenset([base_exec_iname]),
within_inames_is_final=insn.within_inames_is_final,
depends_on=frozenset())
generated_insns.append(init_neutral_insn)
transfer_id = insn_id_gen("%s_%s_transfer" % (insn.id, red_iname))
transfer_insn = make_assignment(
id=transfer_id,
assignees=tuple(
acc_var[outer_local_iname_vars + (var(red_iname),)]
for acc_var in acc_vars),
expression=expr.operation(
arg_dtype,
_strip_if_scalar(tuple(var(nvn) for nvn in neutral_var_names)),
expr.expr, expr.inames),
within_inames=(
(outer_insn_inames - frozenset(expr.inames))
| frozenset([red_iname])),
within_inames_is_final=insn.within_inames_is_final,
depends_on=frozenset([init_id, init_neutral_id]) | insn.depends_on,
no_sync_with=frozenset([(init_id, "any")]))
generated_insns.append(transfer_insn)
cur_size = 1
while cur_size < size:
cur_size *= 2
prev_id = transfer_id
bound = size
istage = 0
while cur_size > 1:
new_size = cur_size // 2
assert new_size * 2 == cur_size
stage_exec_iname = var_name_gen("red_%s_s%d" % (red_iname, istage))
domains.append(_make_slab_set(stage_exec_iname, bound-new_size))
new_iname_tags[stage_exec_iname] = kernel.iname_to_tag[red_iname]
stage_id = insn_id_gen("red_%s_stage_%d" % (red_iname, istage))
stage_insn = make_assignment(
id=stage_id,
assignees=tuple(
acc_var[outer_local_iname_vars + (var(stage_exec_iname),)]
for acc_var in acc_vars),
expression=expr.operation(
arg_dtype,
_strip_if_scalar(tuple(
acc_var[
outer_local_iname_vars + (var(stage_exec_iname),)]
for acc_var in acc_vars)),
_strip_if_scalar(tuple(
acc_var[
outer_local_iname_vars + (
var(stage_exec_iname) + new_size,)]
for acc_var in acc_vars)),
expr.inames),
within_inames=(
base_iname_deps | frozenset([stage_exec_iname])),
within_inames_is_final=insn.within_inames_is_final,
depends_on=frozenset([prev_id]),
)
generated_insns.append(stage_insn)
prev_id = stage_id
cur_size = new_size
bound = cur_size
istage += 1
new_insn_add_depends_on.add(prev_id)
new_insn_add_no_sync_with.add((prev_id, "any"))
new_insn_add_within_inames.add(base_exec_iname or stage_exec_iname)
if nresults == 1:
assert len(acc_vars) == 1
return acc_vars[0][outer_local_iname_vars + (0,)]
else:
return [acc_var[outer_local_iname_vars + (0,)] for acc_var in acc_vars]
# }}}
# {{{ seq/par dispatch
def map_reduction(expr, rec, nresults=1):
# Only expand one level of reduction at a time, going from outermost to
# innermost. Otherwise we get the (iname + insn) dependencies wrong.
try:
arg_dtype = type_inf_mapper(expr.expr)
except DependencyTypeInferenceFailure:
if unknown_types_ok:
arg_dtype = lp.auto
reduction_dtypes = (lp.auto,)*nresults
else:
raise LoopyError("failed to determine type of accumulator for "
"reduction '%s'" % expr)
else:
arg_dtype = arg_dtype.with_target(kernel.target)
reduction_dtypes = expr.operation.result_dtypes(
kernel, arg_dtype, expr.inames)
reduction_dtypes = tuple(
dt.with_target(kernel.target) for dt in reduction_dtypes)
outer_insn_inames = temp_kernel.insn_inames(insn)
bad_inames = frozenset(expr.inames) & outer_insn_inames
if bad_inames:
raise LoopyError("reduction used within loop(s) that it was "
"supposed to reduce over: " + ", ".join(bad_inames))
n_sequential = 0
n_local_par = 0
from loopy.kernel.data import (
LocalIndexTagBase, UnrolledIlpTag, UnrollTag, VectorizeTag,
ParallelTag)
for iname in expr.inames:
iname_tag = kernel.iname_to_tag.get(iname)
if isinstance(iname_tag, (UnrollTag, UnrolledIlpTag)):
# These are nominally parallel, but we can live with
# them as sequential.
n_sequential += 1
elif isinstance(iname_tag, LocalIndexTagBase):
n_local_par += 1
elif isinstance(iname_tag, (ParallelTag, VectorizeTag)):
raise LoopyError("the only form of parallelism supported "
"by reductions is 'local'--found iname '%s' "
"tagged '%s'"
% (iname, type(iname_tag).__name__))
else:
n_sequential += 1
if n_local_par and n_sequential:
raise LoopyError("Reduction over '%s' contains both parallel and "
"sequential inames. It must be split "
"(using split_reduction_{in,out}ward) "
"before code generation."
% ", ".join(expr.inames))
if n_local_par > 1:
raise LoopyError("Reduction over '%s' contains more than"
"one parallel iname. It must be split "
"(using split_reduction_{in,out}ward) "
"before code generation."
% ", ".join(expr.inames))
if n_sequential:
assert n_local_par == 0
return map_reduction_seq(expr, rec, nresults, arg_dtype,
reduction_dtypes)
elif n_local_par:
return map_reduction_local(expr, rec, nresults, arg_dtype,
reduction_dtypes)
else:
from loopy.diagnostic import warn_with_kernel
warn_with_kernel(kernel, "empty_reduction",
"Empty reduction found (no inames to reduce over). "
"Eliminating.")
return expr.expr
# }}}
from loopy.symbolic import ReductionCallbackMapper
cb_mapper = ReductionCallbackMapper(map_reduction)
insn_queue = kernel.instructions[:]
insn_id_replacements = {}
domains = kernel.domains[:]
temp_kernel = kernel
import loopy as lp
while insn_queue:
new_insn_add_depends_on = set()
new_insn_add_no_sync_with = set()
new_insn_add_within_inames = set()
generated_insns = []
insn = insn_queue.pop(0)
if insn_id_filter is not None and insn.id != insn_id_filter \
or not isinstance(insn, lp.MultiAssignmentBase):
new_insns.append(insn)
continue
nresults = len(insn.assignees)
# Run reduction expansion.
from loopy.symbolic import Reduction
if isinstance(insn.expression, Reduction) and nresults > 1:
new_expressions = cb_mapper(insn.expression, nresults=nresults)
else:
new_expressions = (cb_mapper(insn.expression),)
if generated_insns:
# An expansion happened, so insert the generated stuff plus
# ourselves back into the queue.
kwargs = insn.get_copy_kwargs(
depends_on=insn.depends_on
| frozenset(new_insn_add_depends_on),
no_sync_with=insn.no_sync_with
| frozenset(new_insn_add_no_sync_with),
within_inames=(
temp_kernel.insn_inames(insn)
| new_insn_add_within_inames))
kwargs.pop("id")
kwargs.pop("expression")
kwargs.pop("assignee", None)
kwargs.pop("assignees", None)
kwargs.pop("temp_var_type", None)
kwargs.pop("temp_var_types", None)
if isinstance(insn.expression, Reduction) and nresults > 1:
replacement_insns = [
lp.Assignment(
id=insn_id_gen(insn.id),
assignee=assignee,
expression=new_expr,
**kwargs)
for assignee, new_expr in zip(
insn.assignees, new_expressions)]
else:
new_expr, = new_expressions
replacement_insns = [
make_assignment(
id=insn_id_gen(insn.id),
assignees=insn.assignees,
expression=new_expr,
**kwargs)
]
insn_id_replacements[insn.id] = [
rinsn.id for rinsn in replacement_insns]
insn_queue = generated_insns + replacement_insns + insn_queue
# The reduction expander needs an up-to-date kernel
# object to find dependencies. Keep temp_kernel up-to-date.
temp_kernel = kernel.copy(
instructions=new_insns + insn_queue,
temporary_variables=new_temporary_variables,
domains=domains)
temp_kernel = lp.replace_instruction_ids(
temp_kernel, insn_id_replacements)
else:
# nothing happened, we're done with insn
assert not new_insn_add_depends_on
new_insns.append(insn)
kernel = kernel.copy(
instructions=new_insns,
temporary_variables=new_temporary_variables,
domains=domains)
kernel = lp.replace_instruction_ids(kernel, insn_id_replacements)
kernel = lp.tag_inames(kernel, new_iname_tags)
return kernel