def _add_kernel_axis(kernel, axis_name, start, stop, base_inames):
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, base_inames)
domain = domch.domain
new_dim_idx = domain.dim(dim_type.set)
domain = (domain
.insert_dims(dim_type.set, new_dim_idx, 1)
.set_dim_name(dim_type.set, new_dim_idx, axis_name))
from loopy.symbolic import get_dependencies
deps = get_dependencies(start) | get_dependencies(stop)
assert deps <= kernel.all_params()
param_names = domain.get_var_names(dim_type.param)
for dep in deps:
if dep not in param_names:
new_dim_idx = domain.dim(dim_type.param)
domain = (domain
.insert_dims(dim_type.param, new_dim_idx, 1)
.set_dim_name(dim_type.param, new_dim_idx, dep))
from loopy.isl_helpers import make_slab
slab = make_slab(domain.get_space(), axis_name, start, stop)
domain = domain & slab
return kernel.copy(domains=domch.get_domains_with(domain))
def _add_kernel_axis(kernel, axis_name, start, stop, base_inames):
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, base_inames)
domain = domch.domain
new_dim_idx = domain.dim(dim_type.set)
domain = (domain
.insert_dims(dim_type.set, new_dim_idx, 1)
.set_dim_name(dim_type.set, new_dim_idx, axis_name))
from loopy.symbolic import get_dependencies
deps = get_dependencies(start) | get_dependencies(stop)
assert deps <= kernel.all_params()
param_names = domain.get_var_names(dim_type.param)
for dep in deps:
if dep not in param_names:
new_dim_idx = domain.dim(dim_type.param)
domain = (domain
.insert_dims(dim_type.param, new_dim_idx, 1)
.set_dim_name(dim_type.param, new_dim_idx, dep))
from loopy.isl_helpers import make_slab
slab = make_slab(domain.get_space(), axis_name, start, stop)
domain = domain & slab
return kernel.copy(domains=domch.get_domains_with(domain))
def remove_unused_inames(knl, inames=None):
"""Delete those among *inames* that are unused, i.e. project them
out of the domain. If these inames pose implicit restrictions on
other inames, these restrictions will persist as existentially
quantified variables.
:arg inames: may be an iterable of inames or a string of comma-separated inames.
"""
# {{{ normalize arguments
if inames is None:
inames = knl.all_inames()
elif isinstance(inames, str):
inames = inames.split(",")
# }}}
# {{{ check which inames are unused
import loopy as lp
exp_knl = lp.expand_subst(knl)
inames = set(inames)
used_inames = set()
for insn in exp_knl.instructions:
used_inames.update(
exp_knl.insn_inames(insn.id)
| insn.reduction_inames())
unused_inames = inames - used_inames
# }}}
# {{{ remove them
from loopy.kernel.tools import DomainChanger
for iname in unused_inames:
domch = DomainChanger(knl, (iname, ))
dom = domch.domain
dt, idx = dom.get_var_dict()[iname]
dom = dom.project_out(dt, idx, 1)
knl = knl.copy(domains=domch.get_domains_with(dom))
# }}}
return knl
def remove_unused_inames(knl, inames=None):
"""Delete those among *inames* that are unused, i.e. project them
out of the domain. If these inames pose implicit restrictions on
other inames, these restrictions will persist as existentially
quantified variables.
:arg inames: may be an iterable of inames or a string of comma-separated inames.
"""
# {{{ normalize arguments
if inames is None:
inames = knl.all_inames()
elif isinstance(inames, str):
inames = inames.split(",")
# }}}
# {{{ check which inames are unused
inames = set(inames)
used_inames = set()
for insn in knl.instructions:
used_inames.update(knl.insn_inames(insn.id))
unused_inames = inames - used_inames
# }}}
# {{{ remove them
from loopy.kernel.tools import DomainChanger
for iname in unused_inames:
domch = DomainChanger(knl, (iname,))
dom = domch.domain
dt, idx = dom.get_var_dict()[iname]
dom = dom.project_out(dt, idx, 1)
knl = knl.copy(domains=domch.get_domains_with(dom))
# }}}
return knl
def make_reduction_inames_unique(kernel, inames=None, within=None):
"""
:arg inames: if not *None*, only apply to these inames
:arg within: a stack match as understood by
:func:`loopy.match.parse_stack_match`.
.. versionadded:: 2016.2
"""
name_gen = kernel.get_var_name_generator()
from loopy.match import parse_stack_match
within = parse_stack_match(within)
# {{{ change kernel
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, name_gen)
r_uniq = _ReductionInameUniquifier(rule_mapping_context,
inames,
within=within)
kernel = rule_mapping_context.finish_kernel(r_uniq.map_kernel(kernel))
# }}}
# {{{ duplicate the inames
for old_iname, new_iname in r_uniq.old_to_new:
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, frozenset([old_iname]))
from loopy.isl_helpers import duplicate_axes
kernel = kernel.copy(domains=domch.get_domains_with(
duplicate_axes(domch.domain, [old_iname], [new_iname])))
# }}}
return kernel
def precompute(
kernel,
subst_use,
sweep_inames=[],
within=None,
storage_axes=None,
temporary_name=None,
precompute_inames=None,
precompute_outer_inames=None,
storage_axis_to_tag={},
# "None" is a valid value here, distinct from the default.
default_tag=_not_provided,
dtype=None,
fetch_bounding_box=False,
temporary_address_space=None,
compute_insn_id=None,
**kwargs):
"""Precompute the expression described in the substitution rule determined by
*subst_use* and store it in a temporary array. A precomputation needs two
things to operate, a list of *sweep_inames* (order irrelevant) and an
ordered list of *storage_axes* (whose order will describe the axis ordering
of the temporary array).
:arg subst_use: Describes what to prefetch.
The following objects may be given for *subst_use*:
* The name of the substitution rule.
* The tagged name ("name$tag") of the substitution rule.
* A list of invocations of the substitution rule.
This list of invocations, when swept across *sweep_inames*, then serves
to define the footprint of the precomputation.
Invocations may be tagged ("name$tag") to filter out a subset of the
usage sites of the substitution rule. (Namely those usage sites that
use the same tagged name.)
Invocations may be given as a string or as a
:class:`pymbolic.primitives.Expression` object.
If only one invocation is to be given, then the only entry of the list
may be given directly.
If the list of invocations generating the footprint is not given,
all (tag-matching, if desired) usage sites of the substitution rule
are used to determine the footprint.
The following cases can arise for each sweep axis:
* The axis is an iname that occurs within arguments specified at
usage sites of the substitution rule. This case is assumed covered
by the storage axes provided for the argument.
* The axis is an iname that occurs within the *value* of the rule, but not
within its arguments. A new, dedicated storage axis is allocated for
such an axis.
:arg sweep_inames: A :class:`list` of inames to be swept.
May also equivalently be a comma-separated string.
:arg within: a stack match as understood by
:func:`loopy.match.parse_stack_match`.
:arg storage_axes: A :class:`list` of inames and/or rule argument
names/indices to be used as storage axes.
May also equivalently be a comma-separated string.
:arg temporary_name:
The temporary variable name to use for storing the precomputed data.
If it does not exist, it will be created. If it does exist, its properties
(such as size, type) are checked (and updated, if possible) to match
its use.
:arg precompute_inames:
A tuple of inames to be used to carry out the precomputation.
If the specified inames do not already exist, they will be
created. If they do already exist, their loop domain is verified
against the one required for this precomputation. This tuple may
be shorter than the (provided or automatically found) *storage_axes*
tuple, in which case names will be automatically created.
May also equivalently be a comma-separated string.
:arg precompute_outer_inames: A :class:`frozenset` of inames within which
the compute instruction is nested. If *None*, make an educated guess.
May also be specified as a comma-separated string.
:arg default_tag: The :ref:`iname tag <iname-tags>` to be applied to the
inames created to perform the precomputation. The current default will
make them local axes and automatically split them to fit the work
group size, but this default will disappear in favor of simply leaving them
untagged in 2019. For 2018, a warning will be issued if no *default_tag* is
specified.
:arg compute_insn_id: The ID of the instruction generated to perform the
precomputation.
If `storage_axes` is not specified, it defaults to the arrangement
`<direct sweep axes><arguments>` with the direct sweep axes being the
slower-varying indices.
Trivial storage axes (i.e. axes of length 1 with respect to the sweep) are
eliminated.
"""
# {{{ unify temporary_address_space / temporary_scope
temporary_scope = kwargs.pop("temporary_scope", None)
from loopy.kernel.data import AddressSpace
if temporary_scope is not None:
from warnings import warn
warn(
"temporary_scope is deprecated. Use temporary_address_space instead",
DeprecationWarning,
stacklevel=2)
if temporary_address_space is not None:
raise LoopyError(
"may not specify both temporary_address_space and "
"temporary_scope")
temporary_address_space = temporary_scope
del temporary_scope
# }}}
if kwargs:
raise TypeError("unrecognized keyword arguments: %s" %
", ".join(kwargs.keys()))
# {{{ check, standardize arguments
if isinstance(sweep_inames, str):
sweep_inames = [iname.strip() for iname in sweep_inames.split(",")]
for iname in sweep_inames:
if iname not in kernel.all_inames():
raise RuntimeError("sweep iname '%s' is not a known iname" % iname)
sweep_inames = list(sweep_inames)
sweep_inames_set = frozenset(sweep_inames)
if isinstance(storage_axes, str):
storage_axes = [ax.strip() for ax in storage_axes.split(",")]
if isinstance(precompute_inames, str):
precompute_inames = [
iname.strip() for iname in precompute_inames.split(",")
]
if isinstance(precompute_outer_inames, str):
precompute_outer_inames = frozenset(
iname.strip() for iname in precompute_outer_inames.split(","))
if isinstance(subst_use, str):
subst_use = [subst_use]
footprint_generators = None
subst_name = None
subst_tag = None
from pymbolic.primitives import Variable, Call
from loopy.symbolic import parse, TaggedVariable
for use in subst_use:
if isinstance(use, str):
use = parse(use)
if isinstance(use, Call):
if footprint_generators is None:
footprint_generators = []
footprint_generators.append(use)
subst_name_as_expr = use.function
else:
subst_name_as_expr = use
if isinstance(subst_name_as_expr, TaggedVariable):
new_subst_name = subst_name_as_expr.name
new_subst_tag = subst_name_as_expr.tag
elif isinstance(subst_name_as_expr, Variable):
new_subst_name = subst_name_as_expr.name
new_subst_tag = None
else:
raise ValueError("unexpected type of subst_name")
if (subst_name, subst_tag) == (None, None):
subst_name, subst_tag = new_subst_name, new_subst_tag
else:
if (subst_name, subst_tag) != (new_subst_name, new_subst_tag):
raise ValueError("not all uses in subst_use agree "
"on rule name and tag")
from loopy.match import parse_stack_match
within = parse_stack_match(within)
try:
subst = kernel.substitutions[subst_name]
except KeyError:
raise LoopyError("substitution rule '%s' not found" % subst_name)
c_subst_name = subst_name.replace(".", "_")
# {{{ handle default_tag
from loopy.transform.data import _not_provided \
as transform_data_not_provided
if default_tag is _not_provided or default_tag is transform_data_not_provided:
# no need to warn for scalar precomputes
if sweep_inames:
from warnings import warn
warn(
"Not specifying default_tag is deprecated, and default_tag "
"will become mandatory in 2019.x. "
"Pass 'default_tag=\"l.auto\" to match the current default, "
"or Pass 'default_tag=None to leave the loops untagged, which "
"is the recommended behavior.",
DeprecationWarning,
stacklevel=(
# In this case, we came here through add_prefetch. Increase
# the stacklevel.
3 if default_tag is transform_data_not_provided else 2))
default_tag = "l.auto"
from loopy.kernel.data import parse_tag
default_tag = parse_tag(default_tag)
# }}}
# }}}
# {{{ process invocations in footprint generators, start access_descriptors
if footprint_generators:
from pymbolic.primitives import Variable, Call
access_descriptors = []
for fpg in footprint_generators:
if isinstance(fpg, Variable):
args = ()
elif isinstance(fpg, Call):
args = fpg.parameters
else:
raise ValueError("footprint generator must "
"be substitution rule invocation")
access_descriptors.append(
RuleAccessDescriptor(identifier=access_descriptor_id(
args, None),
args=args))
# }}}
# {{{ gather up invocations in kernel code, finish access_descriptors
if not footprint_generators:
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
invg = RuleInvocationGatherer(rule_mapping_context, kernel, subst_name,
subst_tag, within)
del rule_mapping_context
import loopy as lp
for insn in kernel.instructions:
if isinstance(insn, lp.MultiAssignmentBase):
for assignee in insn.assignees:
invg(assignee, kernel, insn)
invg(insn.expression, kernel, insn)
access_descriptors = invg.access_descriptors
if not access_descriptors:
raise RuntimeError("no invocations of '%s' found" % subst_name)
# }}}
# {{{ find inames used in arguments
expanding_usage_arg_deps = set()
for accdesc in access_descriptors:
for arg in accdesc.args:
expanding_usage_arg_deps.update(
get_dependencies(arg) & kernel.all_inames())
# }}}
var_name_gen = kernel.get_var_name_generator()
# {{{ use given / find new storage_axes
# extra axes made necessary because they don't occur in the arguments
extra_storage_axes = set(sweep_inames_set - expanding_usage_arg_deps)
from loopy.symbolic import SubstitutionRuleExpander
submap = SubstitutionRuleExpander(kernel.substitutions)
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: " +
", ".join(extra_storage_axes - value_inames -
expanding_usage_arg_deps))
new_iname_to_tag = {}
if storage_axes is None:
storage_axes = []
# Add sweep_inames (in given--rather than arbitrary--order) to
# storage_axes *if* they are part of extra_storage_axes.
for iname in sweep_inames:
if iname in extra_storage_axes:
extra_storage_axes.remove(iname)
storage_axes.append(iname)
if extra_storage_axes:
if (precompute_inames is not None
and len(storage_axes) < len(precompute_inames)):
raise LoopyError(
"must specify a sufficient number of "
"storage_axes to uniquely determine the meaning "
"of the given precompute_inames. (%d storage_axes "
"needed)" % len(precompute_inames))
storage_axes.extend(sorted(extra_storage_axes))
storage_axes.extend(range(len(subst.arguments)))
del extra_storage_axes
prior_storage_axis_name_dict = {}
storage_axis_names = []
storage_axis_sources = [] # number for arg#, or iname
# {{{ check for pre-existing precompute_inames
if precompute_inames is not None:
preexisting_precompute_inames = (set(precompute_inames)
& kernel.all_inames())
else:
preexisting_precompute_inames = set()
# }}}
for i, saxis in enumerate(storage_axes):
tag_lookup_saxis = saxis
if saxis in subst.arguments:
saxis = subst.arguments.index(saxis)
storage_axis_sources.append(saxis)
if isinstance(saxis, int):
# argument index
name = old_name = subst.arguments[saxis]
else:
old_name = saxis
name = "%s_%s" % (c_subst_name, old_name)
if (precompute_inames is not None and i < len(precompute_inames)
and precompute_inames[i]):
name = precompute_inames[i]
tag_lookup_saxis = name
if (name not in preexisting_precompute_inames
and var_name_gen.is_name_conflicting(name)):
raise RuntimeError("new storage axis name '%s' "
"conflicts with existing name" % name)
else:
name = var_name_gen(name)
storage_axis_names.append(name)
if name not in preexisting_precompute_inames:
new_iname_to_tag[name] = storage_axis_to_tag.get(
tag_lookup_saxis, default_tag)
prior_storage_axis_name_dict[name] = old_name
del storage_axis_to_tag
del storage_axes
del precompute_inames
# }}}
# {{{ fill out access_descriptors[...].storage_axis_exprs
access_descriptors = [
accdesc.copy(storage_axis_exprs=storage_axis_exprs(
storage_axis_sources, accdesc.args))
for accdesc in access_descriptors
]
# }}}
expanding_inames = sweep_inames_set | frozenset(expanding_usage_arg_deps)
assert expanding_inames <= kernel.all_inames()
if storage_axis_names:
# {{{ find domain to be changed
change_inames = expanding_inames | preexisting_precompute_inames
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, change_inames)
if domch.leaf_domain_index is not None:
# If the sweep inames are at home in parent domains, then we'll add
# fetches with loops over copies of these parent inames that will end
# up being scheduled *within* loops over these parents.
for iname in sweep_inames_set:
if kernel.get_home_domain_index(
iname) != domch.leaf_domain_index:
raise RuntimeError(
"sweep iname '%s' is not 'at home' in the "
"sweep's leaf domain" % iname)
# }}}
abm = ArrayToBufferMap(kernel, domch.domain, sweep_inames,
access_descriptors, len(storage_axis_names))
non1_storage_axis_names = []
for i, saxis in enumerate(storage_axis_names):
if abm.non1_storage_axis_flags[i]:
non1_storage_axis_names.append(saxis)
else:
del new_iname_to_tag[saxis]
if saxis in preexisting_precompute_inames:
raise LoopyError(
"precompute axis %d (1-based) was "
"eliminated as "
"having length 1 but also mapped to existing "
"iname '%s'" % (i + 1, saxis))
mod_domain = domch.domain
# {{{ modify the domain, taking into account preexisting inames
# inames may already exist in mod_domain, add them primed to start
primed_non1_saxis_names = [
iname + "'" for iname in non1_storage_axis_names
]
mod_domain = abm.augment_domain_with_sweep(
domch.domain,
primed_non1_saxis_names,
boxify_sweep=fetch_bounding_box)
check_domain = mod_domain
for i, saxis in enumerate(non1_storage_axis_names):
var_dict = mod_domain.get_var_dict(isl.dim_type.set)
if saxis in preexisting_precompute_inames:
# add equality constraint between existing and new variable
dt, dim_idx = var_dict[saxis]
saxis_aff = isl.Aff.var_on_domain(mod_domain.space, dt,
dim_idx)
dt, dim_idx = var_dict[primed_non1_saxis_names[i]]
new_var_aff = isl.Aff.var_on_domain(mod_domain.space, dt,
dim_idx)
mod_domain = mod_domain.add_constraint(
isl.Constraint.equality_from_aff(new_var_aff - saxis_aff))
# project out the new one
mod_domain = mod_domain.project_out(dt, dim_idx, 1)
else:
# remove the prime from the new variable
dt, dim_idx = var_dict[primed_non1_saxis_names[i]]
mod_domain = mod_domain.set_dim_name(dt, dim_idx, saxis)
def add_assumptions(d):
assumption_non_param = isl.BasicSet.from_params(kernel.assumptions)
assumptions, domain = isl.align_two(assumption_non_param, d)
return assumptions & domain
# {{{ check that we got the desired domain
check_domain = add_assumptions(
check_domain.project_out_except(primed_non1_saxis_names,
[isl.dim_type.set]))
mod_check_domain = add_assumptions(mod_domain)
# re-add the prime from the new variable
var_dict = mod_check_domain.get_var_dict(isl.dim_type.set)
for saxis in non1_storage_axis_names:
dt, dim_idx = var_dict[saxis]
mod_check_domain = mod_check_domain.set_dim_name(
dt, dim_idx, saxis + "'")
mod_check_domain = mod_check_domain.project_out_except(
primed_non1_saxis_names, [isl.dim_type.set])
mod_check_domain, check_domain = isl.align_two(mod_check_domain,
check_domain)
# The modified domain can't get bigger by adding constraints
assert mod_check_domain <= check_domain
if not check_domain <= mod_check_domain:
print(check_domain)
print(mod_check_domain)
raise LoopyError("domain of preexisting inames does not match "
"domain needed for precompute")
# }}}
# {{{ check that we didn't shrink the original domain
# project out the new names from the modified domain
orig_domain_inames = list(domch.domain.get_var_dict(isl.dim_type.set))
mod_check_domain = add_assumptions(
mod_domain.project_out_except(orig_domain_inames,
[isl.dim_type.set]))
check_domain = add_assumptions(domch.domain)
mod_check_domain, check_domain = isl.align_two(mod_check_domain,
check_domain)
# The modified domain can't get bigger by adding constraints
assert mod_check_domain <= check_domain
if not check_domain <= mod_check_domain:
print(check_domain)
print(mod_check_domain)
raise LoopyError(
"original domain got shrunk by applying the precompute")
# }}}
# }}}
new_kernel_domains = domch.get_domains_with(mod_domain)
else:
# leave kernel domains unchanged
new_kernel_domains = kernel.domains
non1_storage_axis_names = []
abm = NoOpArrayToBufferMap()
kernel = kernel.copy(domains=new_kernel_domains)
# {{{ set up compute insn
if temporary_name is None:
temporary_name = var_name_gen(based_on=c_subst_name)
assignee = var(temporary_name)
if non1_storage_axis_names:
assignee = assignee[tuple(
var(iname) for iname in non1_storage_axis_names)]
# {{{ process substitutions on compute instruction
storage_axis_subst_dict = {}
for arg_name, bi in zip(storage_axis_names, abm.storage_base_indices):
if arg_name in non1_storage_axis_names:
arg = var(arg_name)
else:
arg = 0
storage_axis_subst_dict[prior_storage_axis_name_dict.get(
arg_name, arg_name)] = arg + bi
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
from loopy.match import parse_stack_match
expr_subst_map = RuleAwareSubstitutionMapper(
rule_mapping_context,
make_subst_func(storage_axis_subst_dict),
within=parse_stack_match(None))
compute_expression = expr_subst_map(subst.expression, kernel, None)
# }}}
from loopy.kernel.data import Assignment
if compute_insn_id is None:
compute_insn_id = kernel.make_unique_instruction_id(
based_on=c_subst_name)
compute_insn = Assignment(
id=compute_insn_id,
assignee=assignee,
expression=compute_expression,
# within_inames determined below
)
compute_dep_id = compute_insn_id
added_compute_insns = [compute_insn]
if temporary_address_space == AddressSpace.GLOBAL:
barrier_insn_id = kernel.make_unique_instruction_id(
based_on=c_subst_name + "_barrier")
from loopy.kernel.instruction import BarrierInstruction
barrier_insn = BarrierInstruction(id=barrier_insn_id,
depends_on=frozenset(
[compute_insn_id]),
synchronization_kind="global",
mem_kind="global")
compute_dep_id = barrier_insn_id
added_compute_insns.append(barrier_insn)
# }}}
# {{{ substitute rule into expressions in kernel (if within footprint)
from loopy.symbolic import SubstitutionRuleExpander
expander = SubstitutionRuleExpander(kernel.substitutions)
invr = RuleInvocationReplacer(rule_mapping_context,
subst_name,
subst_tag,
within,
access_descriptors,
abm,
storage_axis_names,
storage_axis_sources,
non1_storage_axis_names,
temporary_name,
compute_insn_id,
compute_dep_id,
compute_read_variables=get_dependencies(
expander(compute_expression)))
kernel = invr.map_kernel(kernel)
kernel = kernel.copy(instructions=added_compute_insns +
kernel.instructions)
kernel = rule_mapping_context.finish_kernel(kernel)
# }}}
# {{{ add dependencies to compute insn
kernel = kernel.copy(instructions=[
insn.copy(depends_on=frozenset(invr.compute_insn_depends_on)) if insn.
id == compute_insn_id else insn for insn in kernel.instructions
])
# }}}
# {{{ propagate storage iname subst to dependencies of compute instructions
from loopy.kernel.tools import find_recursive_dependencies
compute_deps = find_recursive_dependencies(kernel,
frozenset([compute_insn_id]))
# FIXME: Need to verify that there are no outside dependencies
# on compute_deps
prior_storage_axis_names = frozenset(storage_axis_subst_dict)
new_insns = []
for insn in kernel.instructions:
if (insn.id in compute_deps
and insn.within_inames & prior_storage_axis_names):
insn = (insn.with_transformed_expressions(
lambda expr: expr_subst_map(expr, kernel, insn)).copy(
within_inames=frozenset(
storage_axis_subst_dict.get(iname, var(iname)).name
for iname in insn.within_inames)))
new_insns.append(insn)
else:
new_insns.append(insn)
kernel = kernel.copy(instructions=new_insns)
# }}}
# {{{ determine inames for compute insn
if precompute_outer_inames is None:
from loopy.kernel.tools import guess_iname_deps_based_on_var_use
precompute_outer_inames = (
frozenset(non1_storage_axis_names)
| frozenset((expanding_usage_arg_deps | value_inames) -
sweep_inames_set)
| guess_iname_deps_based_on_var_use(kernel, compute_insn))
else:
if not isinstance(precompute_outer_inames, frozenset):
raise TypeError("precompute_outer_inames must be a frozenset")
precompute_outer_inames = precompute_outer_inames \
| frozenset(non1_storage_axis_names)
kernel = kernel.copy(instructions=[
insn.copy(within_inames=precompute_outer_inames) if insn.id ==
compute_insn_id else insn for insn in kernel.instructions
])
# }}}
# {{{ set up temp variable
import loopy as lp
if dtype is not None:
dtype = np.dtype(dtype)
if temporary_address_space is None:
temporary_address_space = lp.auto
new_temp_shape = tuple(abm.non1_storage_shape)
new_temporary_variables = kernel.temporary_variables.copy()
if temporary_name not in new_temporary_variables:
temp_var = lp.TemporaryVariable(
name=temporary_name,
dtype=dtype,
base_indices=(0, ) * len(new_temp_shape),
shape=tuple(abm.non1_storage_shape),
address_space=temporary_address_space,
dim_names=tuple(non1_storage_axis_names))
else:
temp_var = new_temporary_variables[temporary_name]
# {{{ check and adapt existing temporary
if temp_var.dtype is lp.auto:
pass
elif temp_var.dtype is not lp.auto and dtype is lp.auto:
dtype = temp_var.dtype
elif temp_var.dtype is not lp.auto and dtype is not lp.auto:
if temp_var.dtype != dtype:
raise LoopyError("Existing and new dtype of temporary '%s' "
"do not match (existing: %s, new: %s)" %
(temporary_name, temp_var.dtype, dtype))
temp_var = temp_var.copy(dtype=dtype)
if len(temp_var.shape) != len(new_temp_shape):
raise LoopyError(
"Existing and new temporary '%s' do not "
"have matching number of dimensions ('%d' vs. '%d') " %
(temporary_name, len(temp_var.shape), len(new_temp_shape)))
if temp_var.base_indices != (0, ) * len(new_temp_shape):
raise LoopyError(
"Existing and new temporary '%s' do not "
"have matching number of dimensions ('%d' vs. '%d') " %
(temporary_name, len(temp_var.shape), len(new_temp_shape)))
new_temp_shape = tuple(
max(i, ex_i) for i, ex_i in zip(new_temp_shape, temp_var.shape))
temp_var = temp_var.copy(shape=new_temp_shape)
if temporary_address_space == temp_var.address_space:
pass
elif temporary_address_space is lp.auto:
temporary_address_space = temp_var.address_space
elif temp_var.address_space is lp.auto:
pass
else:
raise LoopyError("Existing and new temporary '%s' do not "
"have matching scopes (existing: %s, new: %s)" %
(temporary_name,
AddressSpace.stringify(temp_var.address_space),
AddressSpace.stringify(temporary_address_space)))
temp_var = temp_var.copy(address_space=temporary_address_space)
# }}}
new_temporary_variables[temporary_name] = temp_var
kernel = kernel.copy(temporary_variables=new_temporary_variables)
# }}}
from loopy import tag_inames
kernel = tag_inames(kernel, new_iname_to_tag)
from loopy.kernel.data import AutoFitLocalIndexTag, filter_iname_tags_by_type
if filter_iname_tags_by_type(new_iname_to_tag.values(),
AutoFitLocalIndexTag):
from loopy.kernel.tools import assign_automatic_axes
kernel = assign_automatic_axes(kernel)
return kernel
def precompute(kernel, subst_use, sweep_inames=[], within=None,
storage_axes=None, temporary_name=None, precompute_inames=None,
storage_axis_to_tag={}, default_tag="l.auto", dtype=None,
fetch_bounding_box=False, temporary_is_local=None,
compute_insn_id=None):
"""Precompute the expression described in the substitution rule determined by
*subst_use* and store it in a temporary array. A precomputation needs two
things to operate, a list of *sweep_inames* (order irrelevant) and an
ordered list of *storage_axes* (whose order will describe the axis ordering
of the temporary array).
:arg subst_use: Describes what to prefetch.
The following objects may be given for *subst_use*:
* The name of the substitution rule.
* The tagged name ("name$tag") of the substitution rule.
* A list of invocations of the substitution rule.
This list of invocations, when swept across *sweep_inames*, then serves
to define the footprint of the precomputation.
Invocations may be tagged ("name$tag") to filter out a subset of the
usage sites of the substitution rule. (Namely those usage sites that
use the same tagged name.)
Invocations may be given as a string or as a
:class:`pymbolic.primitives.Expression` object.
If only one invocation is to be given, then the only entry of the list
may be given directly.
If the list of invocations generating the footprint is not given,
all (tag-matching, if desired) usage sites of the substitution rule
are used to determine the footprint.
The following cases can arise for each sweep axis:
* The axis is an iname that occurs within arguments specified at
usage sites of the substitution rule. This case is assumed covered
by the storage axes provided for the argument.
* The axis is an iname that occurs within the *value* of the rule, but not
within its arguments. A new, dedicated storage axis is allocated for
such an axis.
:arg sweep_inames: A :class:`list` of inames and/or rule argument
names to be swept.
May also equivalently be a comma-separated string.
:arg storage_axes: A :class:`list` of inames and/or rule argument
names/indices to be used as storage axes.
May also equivalently be a comma-separated string.
:arg within: a stack match as understood by
:func:`loopy.context_matching.parse_stack_match`.
:arg temporary_name:
The temporary variable name to use for storing the precomputed data.
If it does not exist, it will be created. If it does exist, its properties
(such as size, type) are checked (and updated, if possible) to match
its use.
:arg precompute_inames:
A tuple of inames to be used to carry out the precomputation.
If the specified inames do not already exist, they will be
created. If they do already exist, their loop domain is verified
against the one required for this precomputation. This tuple may
be shorter than the (provided or automatically found) *storage_axes*
tuple, in which case names will be automatically created.
May also equivalently be a comma-separated string.
:arg compute_insn_id: The ID of the instruction performing the precomputation.
If `storage_axes` is not specified, it defaults to the arrangement
`<direct sweep axes><arguments>` with the direct sweep axes being the
slower-varying indices.
Trivial storage axes (i.e. axes of length 1 with respect to the sweep) are
eliminated.
"""
# {{{ check, standardize arguments
if isinstance(sweep_inames, str):
sweep_inames = [iname.strip() for iname in sweep_inames.split(",")]
for iname in sweep_inames:
if iname not in kernel.all_inames():
raise RuntimeError("sweep iname '%s' is not a known iname"
% iname)
sweep_inames = list(sweep_inames)
sweep_inames_set = frozenset(sweep_inames)
if isinstance(storage_axes, str):
storage_axes = [ax.strip() for ax in storage_axes.split(",")]
if isinstance(precompute_inames, str):
precompute_inames = [iname.strip() for iname in precompute_inames.split(",")]
if isinstance(subst_use, str):
subst_use = [subst_use]
footprint_generators = None
subst_name = None
subst_tag = None
from pymbolic.primitives import Variable, Call
from loopy.symbolic import parse, TaggedVariable
for use in subst_use:
if isinstance(use, str):
use = parse(use)
if isinstance(use, Call):
if footprint_generators is None:
footprint_generators = []
footprint_generators.append(use)
subst_name_as_expr = use.function
else:
subst_name_as_expr = use
if isinstance(subst_name_as_expr, TaggedVariable):
new_subst_name = subst_name_as_expr.name
new_subst_tag = subst_name_as_expr.tag
elif isinstance(subst_name_as_expr, Variable):
new_subst_name = subst_name_as_expr.name
new_subst_tag = None
else:
raise ValueError("unexpected type of subst_name")
if (subst_name, subst_tag) == (None, None):
subst_name, subst_tag = new_subst_name, new_subst_tag
else:
if (subst_name, subst_tag) != (new_subst_name, new_subst_tag):
raise ValueError("not all uses in subst_use agree "
"on rule name and tag")
from loopy.context_matching import parse_stack_match
within = parse_stack_match(within)
from loopy.kernel.data import parse_tag
default_tag = parse_tag(default_tag)
subst = kernel.substitutions[subst_name]
c_subst_name = subst_name.replace(".", "_")
# }}}
# {{{ process invocations in footprint generators, start access_descriptors
if footprint_generators:
from pymbolic.primitives import Variable, Call
access_descriptors = []
for fpg in footprint_generators:
if isinstance(fpg, Variable):
args = ()
elif isinstance(fpg, Call):
args = fpg.parameters
else:
raise ValueError("footprint generator must "
"be substitution rule invocation")
access_descriptors.append(
RuleAccessDescriptor(
identifier=access_descriptor_id(args, None),
args=args
))
# }}}
# {{{ gather up invocations in kernel code, finish access_descriptors
if not footprint_generators:
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
invg = RuleInvocationGatherer(
rule_mapping_context, kernel, subst_name, subst_tag, within)
del rule_mapping_context
import loopy as lp
for insn in kernel.instructions:
if isinstance(insn, lp.Assignment):
invg(insn.assignee, kernel, insn)
invg(insn.expression, kernel, insn)
access_descriptors = invg.access_descriptors
if not access_descriptors:
raise RuntimeError("no invocations of '%s' found" % subst_name)
# }}}
# {{{ find inames used in arguments
expanding_usage_arg_deps = set()
for accdesc in access_descriptors:
for arg in accdesc.args:
expanding_usage_arg_deps.update(
get_dependencies(arg) & kernel.all_inames())
# }}}
var_name_gen = kernel.get_var_name_generator()
# {{{ use given / find new storage_axes
# extra axes made necessary because they don't occur in the arguments
extra_storage_axes = set(sweep_inames_set - expanding_usage_arg_deps)
from loopy.symbolic import SubstitutionRuleExpander
submap = SubstitutionRuleExpander(kernel.substitutions)
value_inames = get_dependencies(
submap(subst.expression)
) & kernel.all_inames()
if value_inames - expanding_usage_arg_deps < extra_storage_axes:
raise RuntimeError("unreferenced sweep inames specified: "
+ ", ".join(extra_storage_axes
- value_inames - expanding_usage_arg_deps))
new_iname_to_tag = {}
if storage_axes is None:
storage_axes = []
# Add sweep_inames (in given--rather than arbitrary--order) to
# storage_axes *if* they are part of extra_storage_axes.
for iname in sweep_inames:
if iname in extra_storage_axes:
extra_storage_axes.remove(iname)
storage_axes.append(iname)
if extra_storage_axes:
if (precompute_inames is not None
and len(storage_axes) < len(precompute_inames)):
raise LoopyError("must specify a sufficient number of "
"storage_axes to uniquely determine the meaning "
"of the given precompute_inames. (%d storage_axes "
"needed)" % len(precompute_inames))
storage_axes.extend(sorted(extra_storage_axes))
storage_axes.extend(range(len(subst.arguments)))
del extra_storage_axes
prior_storage_axis_name_dict = {}
storage_axis_names = []
storage_axis_sources = [] # number for arg#, or iname
# {{{ check for pre-existing precompute_inames
if precompute_inames is not None:
preexisting_precompute_inames = (
set(precompute_inames) & kernel.all_inames())
else:
preexisting_precompute_inames = set()
# }}}
for i, saxis in enumerate(storage_axes):
tag_lookup_saxis = saxis
if saxis in subst.arguments:
saxis = subst.arguments.index(saxis)
storage_axis_sources.append(saxis)
if isinstance(saxis, int):
# argument index
name = old_name = subst.arguments[saxis]
else:
old_name = saxis
name = "%s_%s" % (c_subst_name, old_name)
if (precompute_inames is not None
and i < len(precompute_inames)
and precompute_inames[i]):
name = precompute_inames[i]
tag_lookup_saxis = name
if (name not in preexisting_precompute_inames
and var_name_gen.is_name_conflicting(name)):
raise RuntimeError("new storage axis name '%s' "
"conflicts with existing name" % name)
else:
name = var_name_gen(name)
storage_axis_names.append(name)
if name not in preexisting_precompute_inames:
new_iname_to_tag[name] = storage_axis_to_tag.get(
tag_lookup_saxis, default_tag)
prior_storage_axis_name_dict[name] = old_name
del storage_axis_to_tag
del storage_axes
del precompute_inames
# }}}
# {{{ fill out access_descriptors[...].storage_axis_exprs
access_descriptors = [
accdesc.copy(
storage_axis_exprs=storage_axis_exprs(
storage_axis_sources, accdesc.args))
for accdesc in access_descriptors]
# }}}
expanding_inames = sweep_inames_set | frozenset(expanding_usage_arg_deps)
assert expanding_inames <= kernel.all_inames()
if storage_axis_names:
# {{{ find domain to be changed
change_inames = expanding_inames | preexisting_precompute_inames
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, change_inames)
if domch.leaf_domain_index is not None:
# If the sweep inames are at home in parent domains, then we'll add
# fetches with loops over copies of these parent inames that will end
# up being scheduled *within* loops over these parents.
for iname in sweep_inames_set:
if kernel.get_home_domain_index(iname) != domch.leaf_domain_index:
raise RuntimeError("sweep iname '%s' is not 'at home' in the "
"sweep's leaf domain" % iname)
# }}}
abm = ArrayToBufferMap(kernel, domch.domain, sweep_inames,
access_descriptors, len(storage_axis_names))
non1_storage_axis_names = []
for i, saxis in enumerate(storage_axis_names):
if abm.non1_storage_axis_flags[i]:
non1_storage_axis_names.append(saxis)
else:
del new_iname_to_tag[saxis]
if saxis in preexisting_precompute_inames:
raise LoopyError("precompute axis %d (1-based) was "
"eliminated as "
"having length 1 but also mapped to existing "
"iname '%s'" % (i+1, saxis))
mod_domain = domch.domain
# {{{ modify the domain, taking into account preexisting inames
# inames may already exist in mod_domain, add them primed to start
primed_non1_saxis_names = [
iname+"'" for iname in non1_storage_axis_names]
mod_domain = abm.augment_domain_with_sweep(
domch.domain, primed_non1_saxis_names,
boxify_sweep=fetch_bounding_box)
check_domain = mod_domain
for i, saxis in enumerate(non1_storage_axis_names):
var_dict = mod_domain.get_var_dict(isl.dim_type.set)
if saxis in preexisting_precompute_inames:
# add equality constraint between existing and new variable
dt, dim_idx = var_dict[saxis]
saxis_aff = isl.Aff.var_on_domain(mod_domain.space, dt, dim_idx)
dt, dim_idx = var_dict[primed_non1_saxis_names[i]]
new_var_aff = isl.Aff.var_on_domain(mod_domain.space, dt, dim_idx)
mod_domain = mod_domain.add_constraint(
isl.Constraint.equality_from_aff(new_var_aff - saxis_aff))
# project out the new one
mod_domain = mod_domain.project_out(dt, dim_idx, 1)
else:
# remove the prime from the new variable
dt, dim_idx = var_dict[primed_non1_saxis_names[i]]
mod_domain = mod_domain.set_dim_name(dt, dim_idx, saxis)
# {{{ check that we got the desired domain
check_domain = check_domain.project_out_except(
primed_non1_saxis_names, [isl.dim_type.set])
mod_check_domain = mod_domain
# re-add the prime from the new variable
var_dict = mod_check_domain.get_var_dict(isl.dim_type.set)
for saxis in non1_storage_axis_names:
dt, dim_idx = var_dict[saxis]
mod_check_domain = mod_check_domain.set_dim_name(dt, dim_idx, saxis+"'")
mod_check_domain = mod_check_domain.project_out_except(
primed_non1_saxis_names, [isl.dim_type.set])
mod_check_domain, check_domain = isl.align_two(
mod_check_domain, check_domain)
# The modified domain can't get bigger by adding constraints
assert mod_check_domain <= check_domain
if not check_domain <= mod_check_domain:
print(check_domain)
print(mod_check_domain)
raise LoopyError("domain of preexisting inames does not match "
"domain needed for precompute")
# }}}
# {{{ check that we didn't shrink the original domain
# project out the new names from the modified domain
orig_domain_inames = list(domch.domain.get_var_dict(isl.dim_type.set))
mod_check_domain = mod_domain.project_out_except(
orig_domain_inames, [isl.dim_type.set])
check_domain = domch.domain
mod_check_domain, check_domain = isl.align_two(
mod_check_domain, check_domain)
# The modified domain can't get bigger by adding constraints
assert mod_check_domain <= check_domain
if not check_domain <= mod_check_domain:
print(check_domain)
print(mod_check_domain)
raise LoopyError("original domain got shrunk by applying the precompute")
# }}}
# }}}
new_kernel_domains = domch.get_domains_with(mod_domain)
else:
# leave kernel domains unchanged
new_kernel_domains = kernel.domains
non1_storage_axis_names = []
abm = NoOpArrayToBufferMap()
kernel = kernel.copy(domains=new_kernel_domains)
# {{{ set up compute insn
if temporary_name is None:
temporary_name = var_name_gen(based_on=c_subst_name)
assignee = var(temporary_name)
if non1_storage_axis_names:
assignee = assignee.index(
tuple(var(iname) for iname in non1_storage_axis_names))
# {{{ process substitutions on compute instruction
storage_axis_subst_dict = {}
for arg_name, bi in zip(storage_axis_names, abm.storage_base_indices):
if arg_name in non1_storage_axis_names:
arg = var(arg_name)
else:
arg = 0
storage_axis_subst_dict[
prior_storage_axis_name_dict.get(arg_name, arg_name)] = arg+bi
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
from loopy.context_matching import parse_stack_match
expr_subst_map = RuleAwareSubstitutionMapper(
rule_mapping_context,
make_subst_func(storage_axis_subst_dict),
within=parse_stack_match(None))
compute_expression = expr_subst_map(subst.expression, kernel, None)
# }}}
from loopy.kernel.data import Assignment
if compute_insn_id is None:
compute_insn_id = kernel.make_unique_instruction_id(based_on=c_subst_name)
compute_insn = Assignment(
id=compute_insn_id,
assignee=assignee,
expression=compute_expression)
# }}}
# {{{ substitute rule into expressions in kernel (if within footprint)
invr = RuleInvocationReplacer(rule_mapping_context,
subst_name, subst_tag, within,
access_descriptors, abm,
storage_axis_names, storage_axis_sources,
non1_storage_axis_names,
temporary_name, compute_insn_id)
kernel = invr.map_kernel(kernel)
kernel = kernel.copy(
instructions=[compute_insn] + kernel.instructions)
kernel = rule_mapping_context.finish_kernel(kernel)
# }}}
# {{{ set up temp variable
import loopy as lp
if dtype is None:
dtype = lp.auto
else:
dtype = np.dtype(dtype)
import loopy as lp
if temporary_is_local is None:
temporary_is_local = lp.auto
new_temp_shape = tuple(abm.non1_storage_shape)
new_temporary_variables = kernel.temporary_variables.copy()
if temporary_name not in new_temporary_variables:
temp_var = lp.TemporaryVariable(
name=temporary_name,
dtype=dtype,
base_indices=(0,)*len(new_temp_shape),
shape=tuple(abm.non1_storage_shape),
is_local=temporary_is_local)
else:
temp_var = new_temporary_variables[temporary_name]
# {{{ check and adapt existing temporary
if temp_var.dtype is lp.auto:
pass
elif temp_var.dtype is not lp.auto and dtype is lp.auto:
dtype = temp_var.dtype
elif temp_var.dtype is not lp.auto and dtype is not lp.auto:
if temp_var.dtype != dtype:
raise LoopyError("Existing and new dtype of temporary '%s' "
"do not match (existing: %s, new: %s)"
% (temporary_name, temp_var.dtype, dtype))
temp_var = temp_var.copy(dtype=dtype)
if len(temp_var.shape) != len(new_temp_shape):
raise LoopyError("Existing and new temporary '%s' do not "
"have matching number of dimensions "
% (temporary_name,
len(temp_var.shape), len(new_temp_shape)))
if temp_var.base_indices != (0,) * len(new_temp_shape):
raise LoopyError("Existing and new temporary '%s' do not "
"have matching number of dimensions "
% (temporary_name,
len(temp_var.shape), len(new_temp_shape)))
new_temp_shape = tuple(
max(i, ex_i)
for i, ex_i in zip(new_temp_shape, temp_var.shape))
temp_var = temp_var.copy(shape=new_temp_shape)
if temporary_is_local == temp_var.is_local:
pass
elif temporary_is_local is lp.auto:
temporary_is_local = temp_var.is_local
elif temp_var.is_local is lp.auto:
pass
else:
raise LoopyError("Existing and new temporary '%s' do not "
"have matching values of 'is_local'"
% (temporary_name,
temp_var.is_local, temporary_is_local))
temp_var = temp_var.copy(is_local=temporary_is_local)
# }}}
new_temporary_variables[temporary_name] = temp_var
kernel = kernel.copy(
temporary_variables=new_temporary_variables)
# }}}
from loopy import tag_inames
kernel = tag_inames(kernel, new_iname_to_tag)
from loopy.kernel.data import AutoFitLocalIndexTag
has_automatic_axes = any(
isinstance(tag, AutoFitLocalIndexTag)
for tag in new_iname_to_tag.values())
if has_automatic_axes:
from loopy.kernel.tools import assign_automatic_axes
kernel = assign_automatic_axes(kernel)
return kernel
def buffer_array(kernel, var_name, buffer_inames, init_expression=None,
store_expression=None, within=None, default_tag="l.auto",
temporary_scope=None, temporary_is_local=None,
fetch_bounding_box=False):
"""Replace accesses to *var_name* with ones to a temporary, which is
created and acts as a buffer. To perform this transformation, the access
footprint to *var_name* is determined and a temporary of a suitable
:class:`loopy.AddressSpace` and shape is created.
By default, the value of the buffered cells in *var_name* are read prior to
any (read/write) use, and the modified values are written out after use has
concluded, but for special use cases (e.g. additive accumulation), the
behavior can be modified using *init_expression* and *store_expression*.
:arg buffer_inames: The inames across which the buffer should be usable--i.e.
all possible values of these inames will be covered by the buffer footprint.
A tuple of inames or a comma-separated string.
:arg init_expression: Either *None* (indicating the prior value of the buffered
array should be read) or an expression optionally involving the
variable 'base' (which references the associated location in the array
being buffered).
:arg store_expression: Either *None*, *False*, or an expression involving
variables 'base' and 'buffer' (without array indices).
(*None* indicates that a default storage instruction should be used,
*False* indicates that no storing of the temporary should occur
at all.)
:arg within: If not None, limit the action of the transformation to
matching contexts. See :func:`loopy.match.parse_stack_match`
for syntax.
:arg temporary_scope: If given, override the choice of
:class:`AddressSpace` for the created temporary.
:arg default_tag: The default :ref:`iname-tags` to be assigned to the
inames used for fetching and storing
:arg fetch_bounding_box: If the access footprint is non-convex
(resulting in an error), setting this argument to *True* will force a
rectangular (and hence convex) superset of the footprint to be
fetched.
"""
# {{{ unify temporary_scope / temporary_is_local
from loopy.kernel.data import AddressSpace
if temporary_is_local is not None:
from warnings import warn
warn("temporary_is_local is deprecated. Use temporary_scope instead",
DeprecationWarning, stacklevel=2)
if temporary_scope is not None:
raise LoopyError("may not specify both temporary_is_local and "
"temporary_scope")
if temporary_is_local:
temporary_scope = AddressSpace.LOCAL
else:
temporary_scope = AddressSpace.PRIVATE
del temporary_is_local
# }}}
# {{{ process arguments
if isinstance(init_expression, str):
from loopy.symbolic import parse
init_expression = parse(init_expression)
if isinstance(store_expression, str):
from loopy.symbolic import parse
store_expression = parse(store_expression)
if isinstance(buffer_inames, str):
buffer_inames = [s.strip()
for s in buffer_inames.split(",") if s.strip()]
for iname in buffer_inames:
if iname not in kernel.all_inames():
raise RuntimeError("sweep iname '%s' is not a known iname"
% iname)
buffer_inames = list(buffer_inames)
buffer_inames_set = frozenset(buffer_inames)
from loopy.match import parse_stack_match
within = parse_stack_match(within)
if var_name in kernel.arg_dict:
var_descr = kernel.arg_dict[var_name]
elif var_name in kernel.temporary_variables:
var_descr = kernel.temporary_variables[var_name]
else:
raise ValueError("variable '%s' not found" % var_name)
from loopy.kernel.data import ArrayBase
if isinstance(var_descr, ArrayBase):
var_shape = var_descr.shape
else:
var_shape = ()
if temporary_scope is None:
import loopy as lp
temporary_scope = lp.auto
# }}}
# {{{ caching
from loopy import CACHING_ENABLED
from loopy.preprocess import prepare_for_caching
key_kernel = prepare_for_caching(kernel)
cache_key = (key_kernel, var_name, tuple(buffer_inames),
PymbolicExpressionHashWrapper(init_expression),
PymbolicExpressionHashWrapper(store_expression), within,
default_tag, temporary_scope, fetch_bounding_box)
if CACHING_ENABLED:
try:
result = buffer_array_cache[cache_key]
logger.info("%s: buffer_array cache hit" % kernel.name)
return result
except KeyError:
pass
# }}}
var_name_gen = kernel.get_var_name_generator()
within_inames = set()
access_descriptors = []
for insn in kernel.instructions:
if not within(kernel, insn.id, ()):
continue
from pymbolic.primitives import Variable, Subscript
from loopy.symbolic import LinearSubscript
for assignee in insn.assignees:
if isinstance(assignee, Variable):
assignee_name = assignee.name
index = ()
elif isinstance(assignee, Subscript):
assignee_name = assignee.aggregate.name
index = assignee.index_tuple
elif isinstance(assignee, LinearSubscript):
if assignee.aggregate.name == var_name:
raise LoopyError("buffer_array may not be applied in the "
"presence of linear write indexing into '%s'" % var_name)
else:
raise LoopyError("invalid lvalue '%s'" % assignee)
if assignee_name == var_name:
within_inames.update(
(get_dependencies(index) & kernel.all_inames())
- buffer_inames_set)
access_descriptors.append(
AccessDescriptor(
identifier=insn.id,
storage_axis_exprs=index))
# {{{ find fetch/store inames
init_inames = []
store_inames = []
new_iname_to_tag = {}
for i in range(len(var_shape)):
dim_name = str(i)
if isinstance(var_descr, ArrayBase) and var_descr.dim_names is not None:
dim_name = var_descr.dim_names[i]
init_iname = var_name_gen(f"{var_name}_init_{dim_name}")
store_iname = var_name_gen(f"{var_name}_store_{dim_name}")
new_iname_to_tag[init_iname] = default_tag
new_iname_to_tag[store_iname] = default_tag
init_inames.append(init_iname)
store_inames.append(store_iname)
# }}}
# {{{ modify loop domain
non1_init_inames = []
non1_store_inames = []
if var_shape:
# {{{ find domain to be changed
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, buffer_inames_set | within_inames)
if domch.leaf_domain_index is not None:
# If the sweep inames are at home in parent domains, then we'll add
# fetches with loops over copies of these parent inames that will end
# up being scheduled *within* loops over these parents.
for iname in buffer_inames_set:
if kernel.get_home_domain_index(iname) != domch.leaf_domain_index:
raise RuntimeError("buffer iname '%s' is not 'at home' in the "
"sweep's leaf domain" % iname)
# }}}
abm = ArrayToBufferMap(kernel, domch.domain, buffer_inames,
access_descriptors, len(var_shape))
for i in range(len(var_shape)):
if abm.non1_storage_axis_flags[i]:
non1_init_inames.append(init_inames[i])
non1_store_inames.append(store_inames[i])
else:
del new_iname_to_tag[init_inames[i]]
del new_iname_to_tag[store_inames[i]]
new_domain = domch.domain
new_domain = abm.augment_domain_with_sweep(
new_domain, non1_init_inames,
boxify_sweep=fetch_bounding_box)
new_domain = abm.augment_domain_with_sweep(
new_domain, non1_store_inames,
boxify_sweep=fetch_bounding_box)
new_kernel_domains = domch.get_domains_with(new_domain)
del new_domain
else:
# leave kernel domains unchanged
new_kernel_domains = kernel.domains
abm = NoOpArrayToBufferMap()
# }}}
# {{{ set up temp variable
import loopy as lp
buf_var_name = var_name_gen(based_on=var_name+"_buf")
new_temporary_variables = kernel.temporary_variables.copy()
temp_var = lp.TemporaryVariable(
name=buf_var_name,
dtype=var_descr.dtype,
base_indices=(0,)*len(abm.non1_storage_shape),
shape=tuple(abm.non1_storage_shape),
address_space=temporary_scope)
new_temporary_variables[buf_var_name] = temp_var
# }}}
new_insns = []
buf_var = var(buf_var_name)
# {{{ generate init instruction
buf_var_init = buf_var
if non1_init_inames:
buf_var_init = buf_var_init.index(
tuple(var(iname) for iname in non1_init_inames))
init_base = var(var_name)
init_subscript = []
init_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_init_inames[init_iname_idx])
init_iname_idx += 1
init_subscript.append(ax_subscript)
if init_subscript:
init_base = init_base.index(tuple(init_subscript))
if init_expression is None:
init_expression = init_base
else:
init_expression = init_expression
init_expression = SubstitutionMapper(
make_subst_func({
"base": init_base,
}))(init_expression)
init_insn_id = kernel.make_unique_instruction_id(based_on="init_"+var_name)
from loopy.kernel.data import Assignment
init_instruction = Assignment(id=init_insn_id,
assignee=buf_var_init,
expression=init_expression,
within_inames=(
frozenset(within_inames)
| frozenset(non1_init_inames)),
depends_on=frozenset(),
depends_on_is_final=True)
# }}}
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
aar = ArrayAccessReplacer(rule_mapping_context, var_name,
within, abm, buf_var)
kernel = rule_mapping_context.finish_kernel(aar.map_kernel(kernel))
did_write = False
for insn_id in aar.modified_insn_ids:
insn = kernel.id_to_insn[insn_id]
if buf_var_name in insn.assignee_var_names():
did_write = True
# {{{ add init_insn_id to depends_on
new_insns = []
def none_to_empty_set(s):
if s is None:
return frozenset()
else:
return s
for insn in kernel.instructions:
if insn.id in aar.modified_insn_ids:
new_insns.append(
insn.copy(
depends_on=(
none_to_empty_set(insn.depends_on)
| frozenset([init_insn_id]))))
else:
new_insns.append(insn)
# }}}
# {{{ generate store instruction
buf_var_store = buf_var
if non1_store_inames:
buf_var_store = buf_var_store.index(
tuple(var(iname) for iname in non1_store_inames))
store_subscript = []
store_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_store_inames[store_iname_idx])
store_iname_idx += 1
store_subscript.append(ax_subscript)
store_target = var(var_name)
if store_subscript:
store_target = store_target.index(tuple(store_subscript))
if store_expression is None:
store_expression = buf_var_store
else:
store_expression = SubstitutionMapper(
make_subst_func({
"base": store_target,
"buffer": buf_var_store,
}))(store_expression)
if store_expression is not False:
from loopy.kernel.data import Assignment
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, "any")]),
assignee=store_target,
expression=store_expression,
within_inames=(
frozenset(within_inames)
| frozenset(non1_store_inames)))
else:
did_write = False
# }}}
new_insns.append(init_instruction)
if did_write:
new_insns.append(store_instruction)
else:
for iname in store_inames:
del new_iname_to_tag[iname]
kernel = kernel.copy(
domains=new_kernel_domains,
instructions=new_insns,
temporary_variables=new_temporary_variables)
from loopy import tag_inames
kernel = tag_inames(kernel, new_iname_to_tag)
from loopy.kernel.tools import assign_automatic_axes
kernel = assign_automatic_axes(kernel)
if CACHING_ENABLED:
from loopy.preprocess import prepare_for_caching
buffer_array_cache.store_if_not_present(
cache_key, prepare_for_caching(kernel))
return kernel
def buffer_array(kernel, var_name, buffer_inames, init_expression=None,
store_expression=None, within=None, default_tag="l.auto",
temporary_scope=None, temporary_is_local=None,
fetch_bounding_box=False):
"""
:arg init_expression: Either *None* (indicating the prior value of the buffered
array should be read) or an expression optionally involving the
variable 'base' (which references the associated location in the array
being buffered).
:arg store_expression: Either *None*, *False*, or an expression involving
variables 'base' and 'buffer' (without array indices).
(*None* indicates that a default storage instruction should be used,
*False* indicates that no storing of the temporary should occur
at all.)
"""
# {{{ unify temporary_scope / temporary_is_local
from loopy.kernel.data import temp_var_scope
if temporary_is_local is not None:
from warnings import warn
warn("temporary_is_local is deprecated. Use temporary_scope instead",
DeprecationWarning, stacklevel=2)
if temporary_scope is not None:
raise LoopyError("may not specify both temporary_is_local and "
"temporary_scope")
if temporary_is_local:
temporary_scope = temp_var_scope.LOCAL
else:
temporary_scope = temp_var_scope.PRIVATE
del temporary_is_local
# }}}
# {{{ process arguments
if isinstance(init_expression, str):
from loopy.symbolic import parse
init_expression = parse(init_expression)
if isinstance(store_expression, str):
from loopy.symbolic import parse
store_expression = parse(store_expression)
if isinstance(buffer_inames, str):
buffer_inames = [s.strip()
for s in buffer_inames.split(",") if s.strip()]
for iname in buffer_inames:
if iname not in kernel.all_inames():
raise RuntimeError("sweep iname '%s' is not a known iname"
% iname)
buffer_inames = list(buffer_inames)
buffer_inames_set = frozenset(buffer_inames)
from loopy.match import parse_stack_match
within = parse_stack_match(within)
if var_name in kernel.arg_dict:
var_descr = kernel.arg_dict[var_name]
elif var_name in kernel.temporary_variables:
var_descr = kernel.temporary_variables[var_name]
else:
raise ValueError("variable '%s' not found" % var_name)
from loopy.kernel.data import ArrayBase
if isinstance(var_descr, ArrayBase):
var_shape = var_descr.shape
else:
var_shape = ()
if temporary_scope is None:
import loopy as lp
temporary_scope = lp.auto
# }}}
# {{{ caching
from loopy import CACHING_ENABLED
from loopy.preprocess import prepare_for_caching
key_kernel = prepare_for_caching(kernel)
cache_key = (key_kernel, var_name, tuple(buffer_inames),
PymbolicExpressionHashWrapper(init_expression),
PymbolicExpressionHashWrapper(store_expression), within,
default_tag, temporary_scope, fetch_bounding_box)
if CACHING_ENABLED:
try:
result = buffer_array_cache[cache_key]
logger.info("%s: buffer_array cache hit" % kernel.name)
return result
except KeyError:
pass
# }}}
var_name_gen = kernel.get_var_name_generator()
within_inames = set()
access_descriptors = []
for insn in kernel.instructions:
if not within(kernel, insn.id, ()):
continue
from pymbolic.primitives import Variable, Subscript
from loopy.symbolic import LinearSubscript
for assignee in insn.assignees:
if isinstance(assignee, Variable):
assignee_name = assignee.name
index = ()
elif isinstance(assignee, Subscript):
assignee_name = assignee.aggregate.name
index = assignee.index_tuple
elif isinstance(assignee, LinearSubscript):
if assignee.aggregate.name == var_name:
raise LoopyError("buffer_array may not be applied in the "
"presence of linear write indexing into '%s'" % var_name)
else:
raise LoopyError("invalid lvalue '%s'" % assignee)
if assignee_name == var_name:
within_inames.update(
(get_dependencies(index) & kernel.all_inames())
- buffer_inames_set)
access_descriptors.append(
AccessDescriptor(
identifier=insn.id,
storage_axis_exprs=index))
# {{{ find fetch/store inames
init_inames = []
store_inames = []
new_iname_to_tag = {}
for i in range(len(var_shape)):
dim_name = str(i)
if isinstance(var_descr, ArrayBase) and var_descr.dim_names is not None:
dim_name = var_descr.dim_names[i]
init_iname = var_name_gen("%s_init_%s" % (var_name, dim_name))
store_iname = var_name_gen("%s_store_%s" % (var_name, dim_name))
new_iname_to_tag[init_iname] = default_tag
new_iname_to_tag[store_iname] = default_tag
init_inames.append(init_iname)
store_inames.append(store_iname)
# }}}
# {{{ modify loop domain
non1_init_inames = []
non1_store_inames = []
if var_shape:
# {{{ find domain to be changed
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, buffer_inames_set | within_inames)
if domch.leaf_domain_index is not None:
# If the sweep inames are at home in parent domains, then we'll add
# fetches with loops over copies of these parent inames that will end
# up being scheduled *within* loops over these parents.
for iname in buffer_inames_set:
if kernel.get_home_domain_index(iname) != domch.leaf_domain_index:
raise RuntimeError("buffer iname '%s' is not 'at home' in the "
"sweep's leaf domain" % iname)
# }}}
abm = ArrayToBufferMap(kernel, domch.domain, buffer_inames,
access_descriptors, len(var_shape))
for i in range(len(var_shape)):
if abm.non1_storage_axis_flags[i]:
non1_init_inames.append(init_inames[i])
non1_store_inames.append(store_inames[i])
else:
del new_iname_to_tag[init_inames[i]]
del new_iname_to_tag[store_inames[i]]
new_domain = domch.domain
new_domain = abm.augment_domain_with_sweep(
new_domain, non1_init_inames,
boxify_sweep=fetch_bounding_box)
new_domain = abm.augment_domain_with_sweep(
new_domain, non1_store_inames,
boxify_sweep=fetch_bounding_box)
new_kernel_domains = domch.get_domains_with(new_domain)
del new_domain
else:
# leave kernel domains unchanged
new_kernel_domains = kernel.domains
abm = NoOpArrayToBufferMap()
# }}}
# {{{ set up temp variable
import loopy as lp
buf_var_name = var_name_gen(based_on=var_name+"_buf")
new_temporary_variables = kernel.temporary_variables.copy()
temp_var = lp.TemporaryVariable(
name=buf_var_name,
dtype=var_descr.dtype,
base_indices=(0,)*len(abm.non1_storage_shape),
shape=tuple(abm.non1_storage_shape),
scope=temporary_scope)
new_temporary_variables[buf_var_name] = temp_var
# }}}
new_insns = []
buf_var = var(buf_var_name)
# {{{ generate init instruction
buf_var_init = buf_var
if non1_init_inames:
buf_var_init = buf_var_init.index(
tuple(var(iname) for iname in non1_init_inames))
init_base = var(var_name)
init_subscript = []
init_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_init_inames[init_iname_idx])
init_iname_idx += 1
init_subscript.append(ax_subscript)
if init_subscript:
init_base = init_base.index(tuple(init_subscript))
if init_expression is None:
init_expression = init_base
else:
init_expression = init_expression
init_expression = SubstitutionMapper(
make_subst_func({
"base": init_base,
}))(init_expression)
init_insn_id = kernel.make_unique_instruction_id(based_on="init_"+var_name)
from loopy.kernel.data import Assignment
init_instruction = Assignment(id=init_insn_id,
assignee=buf_var_init,
expression=init_expression,
forced_iname_deps=(
frozenset(within_inames)
| frozenset(non1_init_inames)),
depends_on=frozenset(),
depends_on_is_final=True)
# }}}
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
aar = ArrayAccessReplacer(rule_mapping_context, var_name,
within, abm, buf_var)
kernel = rule_mapping_context.finish_kernel(aar.map_kernel(kernel))
did_write = False
for insn_id in aar.modified_insn_ids:
insn = kernel.id_to_insn[insn_id]
if buf_var_name in insn.assignee_var_names():
did_write = True
# {{{ add init_insn_id to depends_on
new_insns = []
def none_to_empty_set(s):
if s is None:
return frozenset()
else:
return s
for insn in kernel.instructions:
if insn.id in aar.modified_insn_ids:
new_insns.append(
insn.copy(
depends_on=(
none_to_empty_set(insn.depends_on)
| frozenset([init_insn_id]))))
else:
new_insns.append(insn)
# }}}
# {{{ generate store instruction
buf_var_store = buf_var
if non1_store_inames:
buf_var_store = buf_var_store.index(
tuple(var(iname) for iname in non1_store_inames))
store_subscript = []
store_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_store_inames[store_iname_idx])
store_iname_idx += 1
store_subscript.append(ax_subscript)
store_target = var(var_name)
if store_subscript:
store_target = store_target.index(tuple(store_subscript))
if store_expression is None:
store_expression = buf_var_store
else:
store_expression = SubstitutionMapper(
make_subst_func({
"base": store_target,
"buffer": buf_var_store,
}))(store_expression)
if store_expression is not False:
from loopy.kernel.data import Assignment
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)
| frozenset(non1_store_inames)))
else:
did_write = False
# }}}
new_insns.append(init_instruction)
if did_write:
new_insns.append(store_instruction)
else:
for iname in store_inames:
del new_iname_to_tag[iname]
kernel = kernel.copy(
domains=new_kernel_domains,
instructions=new_insns,
temporary_variables=new_temporary_variables)
from loopy import tag_inames
kernel = tag_inames(kernel, new_iname_to_tag)
from loopy.kernel.tools import assign_automatic_axes
kernel = assign_automatic_axes(kernel)
if CACHING_ENABLED:
from loopy.preprocess import prepare_for_caching
buffer_array_cache[cache_key] = prepare_for_caching(kernel)
return kernel
def buffer_array(
kernel,
var_name,
buffer_inames,
init_expression=None,
store_expression=None,
within=None,
default_tag="l.auto",
temporary_is_local=None,
fetch_bounding_box=False,
):
"""
:arg init_expression: Either *None* (indicating the prior value of the buffered
array should be read) or an expression optionally involving the
variable 'base' (which references the associated location in the array
being buffered).
:arg store_expression: Either *None* or an expression involving
variables 'base' and 'buffer' (without array indices).
"""
# {{{ process arguments
if isinstance(init_expression, str):
from loopy.symbolic import parse
init_expression = parse(init_expression)
if isinstance(store_expression, str):
from loopy.symbolic import parse
store_expression = parse(store_expression)
if isinstance(buffer_inames, str):
buffer_inames = [s.strip() for s in buffer_inames.split(",") if s.strip()]
for iname in buffer_inames:
if iname not in kernel.all_inames():
raise RuntimeError("sweep iname '%s' is not a known iname" % iname)
buffer_inames = list(buffer_inames)
buffer_inames_set = frozenset(buffer_inames)
from loopy.context_matching import parse_stack_match
within = parse_stack_match(within)
if var_name in kernel.arg_dict:
var_descr = kernel.arg_dict[var_name]
elif var_name in kernel.temporary_variables:
var_descr = kernel.temporary_variables[var_name]
else:
raise ValueError("variable '%s' not found" % var_name)
from loopy.kernel.data import ArrayBase
if isinstance(var_descr, ArrayBase):
var_shape = var_descr.shape
else:
var_shape = ()
if temporary_is_local is None:
import loopy as lp
temporary_is_local = lp.auto
# }}}
var_name_gen = kernel.get_var_name_generator()
within_inames = set()
access_descriptors = []
for insn in kernel.instructions:
if not within(kernel, insn.id, ()):
continue
for assignee, index in insn.assignees_and_indices():
if assignee == var_name:
within_inames.update((get_dependencies(index) & kernel.all_inames()) - buffer_inames_set)
access_descriptors.append(AccessDescriptor(identifier=insn.id, storage_axis_exprs=index))
# {{{ find fetch/store inames
init_inames = []
store_inames = []
new_iname_to_tag = {}
for i in range(len(var_shape)):
init_iname = var_name_gen("%s_init_%d" % (var_name, i))
store_iname = var_name_gen("%s_store_%d" % (var_name, i))
new_iname_to_tag[init_iname] = default_tag
new_iname_to_tag[store_iname] = default_tag
init_inames.append(init_iname)
store_inames.append(store_iname)
# }}}
# {{{ modify loop domain
non1_init_inames = []
non1_store_inames = []
if var_shape:
# {{{ find domain to be changed
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, buffer_inames_set | within_inames)
if domch.leaf_domain_index is not None:
# If the sweep inames are at home in parent domains, then we'll add
# fetches with loops over copies of these parent inames that will end
# up being scheduled *within* loops over these parents.
for iname in buffer_inames_set:
if kernel.get_home_domain_index(iname) != domch.leaf_domain_index:
raise RuntimeError("buffer iname '%s' is not 'at home' in the " "sweep's leaf domain" % iname)
# }}}
abm = ArrayToBufferMap(kernel, domch.domain, buffer_inames, access_descriptors, len(var_shape))
for i in range(len(var_shape)):
if abm.non1_storage_axis_flags[i]:
non1_init_inames.append(init_inames[i])
non1_store_inames.append(store_inames[i])
else:
del new_iname_to_tag[init_inames[i]]
del new_iname_to_tag[store_inames[i]]
new_domain = domch.domain
new_domain = abm.augment_domain_with_sweep(new_domain, non1_init_inames, boxify_sweep=fetch_bounding_box)
new_domain = abm.augment_domain_with_sweep(new_domain, non1_store_inames, boxify_sweep=fetch_bounding_box)
new_kernel_domains = domch.get_domains_with(new_domain)
del new_domain
else:
# leave kernel domains unchanged
new_kernel_domains = kernel.domains
abm = NoOpArrayToBufferMap()
# }}}
# {{{ set up temp variable
import loopy as lp
buf_var_name = var_name_gen(based_on=var_name + "_buf")
new_temporary_variables = kernel.temporary_variables.copy()
temp_var = lp.TemporaryVariable(
name=buf_var_name,
dtype=var_descr.dtype,
base_indices=(0,) * len(abm.non1_storage_shape),
shape=tuple(abm.non1_storage_shape),
is_local=temporary_is_local,
)
new_temporary_variables[buf_var_name] = temp_var
# }}}
new_insns = []
buf_var = var(buf_var_name)
# {{{ generate init instruction
buf_var_init = buf_var
if non1_init_inames:
buf_var_init = buf_var_init.index(tuple(var(iname) for iname in non1_init_inames))
init_base = var(var_name)
init_subscript = []
init_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_init_inames[init_iname_idx])
init_iname_idx += 1
init_subscript.append(ax_subscript)
if init_subscript:
init_base = init_base.index(tuple(init_subscript))
if init_expression is None:
init_expression = init_base
else:
init_expression = init_expression
init_expression = SubstitutionMapper(make_subst_func({"base": init_base}))(init_expression)
init_insn_id = kernel.make_unique_instruction_id(based_on="init_" + var_name)
from loopy.kernel.data import ExpressionInstruction
init_instruction = ExpressionInstruction(
id=init_insn_id,
assignee=buf_var_init,
expression=init_expression,
forced_iname_deps=frozenset(within_inames),
insn_deps=frozenset(),
insn_deps_is_final=True,
)
# }}}
rule_mapping_context = SubstitutionRuleMappingContext(kernel.substitutions, kernel.get_var_name_generator())
aar = ArrayAccessReplacer(rule_mapping_context, var_name, within, abm, buf_var)
kernel = rule_mapping_context.finish_kernel(aar.map_kernel(kernel))
did_write = False
for insn_id in aar.modified_insn_ids:
insn = kernel.id_to_insn[insn_id]
if any(assignee_name == buf_var_name for assignee_name, _ in insn.assignees_and_indices()):
did_write = True
# {{{ add init_insn_id to insn_deps
new_insns = []
def none_to_empty_set(s):
if s is None:
return frozenset()
else:
return s
for insn in kernel.instructions:
if insn.id in aar.modified_insn_ids:
new_insns.append(insn.copy(insn_deps=(none_to_empty_set(insn.insn_deps) | frozenset([init_insn_id]))))
else:
new_insns.append(insn)
# }}}
# {{{ generate store instruction
buf_var_store = buf_var
if non1_store_inames:
buf_var_store = buf_var_store.index(tuple(var(iname) for iname in non1_store_inames))
store_subscript = []
store_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_store_inames[store_iname_idx])
store_iname_idx += 1
store_subscript.append(ax_subscript)
store_target = var(var_name)
if store_subscript:
store_target = store_target.index(tuple(store_subscript))
if store_expression is None:
store_expression = buf_var_store
else:
store_expression = SubstitutionMapper(make_subst_func({"base": store_target, "buffer": buf_var_store}))(
store_expression
)
from loopy.kernel.data import ExpressionInstruction
store_instruction = ExpressionInstruction(
id=kernel.make_unique_instruction_id(based_on="store_" + var_name),
insn_deps=frozenset(aar.modified_insn_ids),
assignee=store_target,
expression=store_expression,
forced_iname_deps=frozenset(within_inames),
)
# }}}
new_insns.append(init_instruction)
if did_write:
new_insns.append(store_instruction)
kernel = kernel.copy(
domains=new_kernel_domains, instructions=new_insns, temporary_variables=new_temporary_variables
)
from loopy import tag_inames
kernel = tag_inames(kernel, new_iname_to_tag)
return kernel
def link_inames(knl, inames, new_iname, within=None, tag=None):
# {{{ normalize arguments
if isinstance(inames, str):
inames = inames.split(",")
var_name_gen = knl.get_var_name_generator()
new_iname = var_name_gen(new_iname)
# }}}
# {{{ ensure that each iname is used at most once in each instruction
inames_set = set(inames)
if 0:
# FIXME!
for insn in knl.instructions:
insn_inames = knl.insn_inames(insn.id) | insn.reduction_inames()
if len(insn_inames & inames_set) > 1:
raise LoopyError("To-be-linked inames '%s' are used in "
"instruction '%s'. No more than one such iname can "
"be used in one instruction."
% (", ".join(insn_inames & inames_set), insn.id))
# }}}
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(knl, tuple(inames))
# {{{ ensure that projections are identical
unrelated_dom_inames = list(
set(domch.domain.get_var_names(dim_type.set))
- inames_set)
domain = domch.domain
# move all inames to be linked to end to prevent shuffly confusion
for iname in inames:
dt, index = domain.get_var_dict()[iname]
assert dt == dim_type.set
# move to tail of param dim_type
domain = domain.move_dims(
dim_type.param, domain.dim(dim_type.param),
dt, index, 1)
# move to tail of set dim_type
domain = domain.move_dims(
dim_type.set, domain.dim(dim_type.set),
dim_type.param, domain.dim(dim_type.param)-1, 1)
projections = [
domch.domain.project_out_except(
unrelated_dom_inames + [iname], [dim_type.set])
for iname in inames]
all_equal = True
first_proj = projections[0]
for proj in projections[1:]:
all_equal = all_equal and (proj <= first_proj and first_proj <= proj)
if not all_equal:
raise LoopyError("Inames cannot be linked because their domain "
"constraints are not the same.")
del domain # messed up for testing, do not use
# }}}
# change the domain
from loopy.isl_helpers import duplicate_axes
knl = knl.copy(
domains=domch.get_domains_with(
duplicate_axes(domch.domain, [inames[0]], [new_iname])))
# {{{ change the code
from pymbolic import var
subst_dict = dict((iname, var(new_iname)) for iname in inames)
from loopy.context_matching import parse_stack_match
within = parse_stack_match(within)
from pymbolic.mapper.substitutor import make_subst_func
rule_mapping_context = SubstitutionRuleMappingContext(
knl.substitutions, var_name_gen)
ijoin = RuleAwareSubstitutionMapper(rule_mapping_context,
make_subst_func(subst_dict), within)
knl = rule_mapping_context.finish_kernel(
ijoin.map_kernel(knl))
# }}}
knl = remove_unused_inames(knl, inames)
if tag is not None:
knl = tag_inames(knl, {new_iname: tag})
return knl
def duplicate_inames(knl, inames, within, new_inames=None, suffix=None,
tags={}):
"""
:arg within: a stack match as understood by
:func:`loopy.context_matching.parse_stack_match`.
"""
# {{{ normalize arguments, find unique new_inames
if isinstance(inames, str):
inames = [iname.strip() for iname in inames.split(",")]
if isinstance(new_inames, str):
new_inames = [iname.strip() for iname in new_inames.split(",")]
from loopy.context_matching import parse_stack_match
within = parse_stack_match(within)
if new_inames is None:
new_inames = [None] * len(inames)
if len(new_inames) != len(inames):
raise ValueError("new_inames must have the same number of entries as inames")
name_gen = knl.get_var_name_generator()
for i, iname in enumerate(inames):
new_iname = new_inames[i]
if new_iname is None:
new_iname = iname
if suffix is not None:
new_iname += suffix
new_iname = name_gen(new_iname)
else:
if name_gen.is_name_conflicting(new_iname):
raise ValueError("new iname '%s' conflicts with existing names"
% new_iname)
name_gen.add_name(new_iname)
new_inames[i] = new_iname
# }}}
# {{{ duplicate the inames
for old_iname, new_iname in zip(inames, new_inames):
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(knl, frozenset([old_iname]))
from loopy.isl_helpers import duplicate_axes
knl = knl.copy(
domains=domch.get_domains_with(
duplicate_axes(domch.domain, [old_iname], [new_iname])))
# }}}
# {{{ change the inames in the code
rule_mapping_context = SubstitutionRuleMappingContext(
knl.substitutions, name_gen)
indup = _InameDuplicator(rule_mapping_context,
old_to_new=dict(list(zip(inames, new_inames))),
within=within)
knl = rule_mapping_context.finish_kernel(
indup.map_kernel(knl))
# }}}
# {{{ realize tags
for old_iname, new_iname in zip(inames, new_inames):
new_tag = tags.get(old_iname)
if new_tag is not None:
knl = tag_inames(knl, {new_iname: new_tag})
# }}}
return knl
def join_inames(kernel, inames, new_iname=None, tag=None, within=None):
"""
:arg inames: fastest varying last
:arg within: a stack match as understood by
:func:`loopy.context_matching.parse_stack_match`.
"""
# now fastest varying first
inames = inames[::-1]
if new_iname is None:
new_iname = kernel.get_var_name_generator()("_and_".join(inames))
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, frozenset(inames))
for iname in inames:
if kernel.get_home_domain_index(iname) != domch.leaf_domain_index:
raise LoopyError("iname '%s' is not 'at home' in the "
"join's leaf domain" % iname)
new_domain = domch.domain
new_dim_idx = new_domain.dim(dim_type.set)
new_domain = new_domain.add_dims(dim_type.set, 1)
new_domain = new_domain.set_dim_name(dim_type.set, new_dim_idx, new_iname)
joint_aff = zero = isl.Aff.zero_on_domain(new_domain.space)
subst_dict = {}
base_divisor = 1
from pymbolic import var
for i, iname in enumerate(inames):
iname_dt, iname_idx = zero.get_space().get_var_dict()[iname]
iname_aff = zero.add_coefficient_val(iname_dt, iname_idx, 1)
joint_aff = joint_aff + base_divisor*iname_aff
bounds = kernel.get_iname_bounds(iname, constants_only=True)
from loopy.isl_helpers import (
static_max_of_pw_aff, static_value_of_pw_aff)
from loopy.symbolic import pw_aff_to_expr
length = int(pw_aff_to_expr(
static_max_of_pw_aff(bounds.size, constants_only=True)))
try:
lower_bound_aff = static_value_of_pw_aff(
bounds.lower_bound_pw_aff.coalesce(),
constants_only=False)
except Exception as e:
raise type(e)("while finding lower bound of '%s': " % iname)
my_val = var(new_iname) // base_divisor
if i+1 < len(inames):
my_val %= length
my_val += pw_aff_to_expr(lower_bound_aff)
subst_dict[iname] = my_val
base_divisor *= length
from loopy.isl_helpers import iname_rel_aff
new_domain = new_domain.add_constraint(
isl.Constraint.equality_from_aff(
iname_rel_aff(new_domain.get_space(), new_iname, "==", joint_aff)))
for i, iname in enumerate(inames):
iname_to_dim = new_domain.get_space().get_var_dict()
iname_dt, iname_idx = iname_to_dim[iname]
if within is None:
new_domain = new_domain.project_out(iname_dt, iname_idx, 1)
def subst_forced_iname_deps(fid):
result = set()
for iname in fid:
if iname in inames:
result.add(new_iname)
else:
result.add(iname)
return frozenset(result)
new_insns = [
insn.copy(
forced_iname_deps=subst_forced_iname_deps(insn.forced_iname_deps))
for insn in kernel.instructions]
kernel = (kernel
.copy(
instructions=new_insns,
domains=domch.get_domains_with(new_domain),
applied_iname_rewrites=kernel.applied_iname_rewrites + [subst_dict]
))
from loopy.context_matching import parse_stack_match
within = parse_stack_match(within)
from pymbolic.mapper.substitutor import make_subst_func
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
ijoin = _InameJoiner(rule_mapping_context, within,
make_subst_func(subst_dict),
inames, new_iname)
kernel = rule_mapping_context.finish_kernel(
ijoin.map_kernel(kernel))
if tag is not None:
kernel = tag_inames(kernel, {new_iname: tag})
return kernel
def duplicate_inames(knl,
inames,
within,
new_inames=None,
suffix=None,
tags={}):
"""
:arg within: a stack match as understood by
:func:`loopy.match.parse_stack_match`.
"""
# {{{ normalize arguments, find unique new_inames
if isinstance(inames, str):
inames = [iname.strip() for iname in inames.split(",")]
if isinstance(new_inames, str):
new_inames = [iname.strip() for iname in new_inames.split(",")]
from loopy.match import parse_stack_match
within = parse_stack_match(within)
if new_inames is None:
new_inames = [None] * len(inames)
if len(new_inames) != len(inames):
raise ValueError(
"new_inames must have the same number of entries as inames")
name_gen = knl.get_var_name_generator()
for i, iname in enumerate(inames):
new_iname = new_inames[i]
if new_iname is None:
new_iname = iname
if suffix is not None:
new_iname += suffix
new_iname = name_gen(new_iname)
else:
if name_gen.is_name_conflicting(new_iname):
raise ValueError(
"new iname '%s' conflicts with existing names" % new_iname)
name_gen.add_name(new_iname)
new_inames[i] = new_iname
# }}}
# {{{ duplicate the inames
for old_iname, new_iname in zip(inames, new_inames):
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(knl, frozenset([old_iname]))
from loopy.isl_helpers import duplicate_axes
knl = knl.copy(domains=domch.get_domains_with(
duplicate_axes(domch.domain, [old_iname], [new_iname])))
# }}}
# {{{ change the inames in the code
rule_mapping_context = SubstitutionRuleMappingContext(
knl.substitutions, name_gen)
indup = _InameDuplicator(rule_mapping_context,
old_to_new=dict(list(zip(inames, new_inames))),
within=within)
knl = rule_mapping_context.finish_kernel(indup.map_kernel(knl))
# }}}
# {{{ realize tags
for old_iname, new_iname in zip(inames, new_inames):
new_tag = tags.get(old_iname)
if new_tag is not None:
knl = tag_inames(knl, {new_iname: new_tag})
# }}}
return knl
def join_inames(kernel, inames, new_iname=None, tag=None, within=None):
"""
:arg inames: fastest varying last
:arg within: a stack match as understood by
:func:`loopy.match.parse_stack_match`.
"""
# now fastest varying first
inames = inames[::-1]
if new_iname is None:
new_iname = kernel.get_var_name_generator()("_and_".join(inames))
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, frozenset(inames))
for iname in inames:
if kernel.get_home_domain_index(iname) != domch.leaf_domain_index:
raise LoopyError("iname '%s' is not 'at home' in the "
"join's leaf domain" % iname)
new_domain = domch.domain
new_dim_idx = new_domain.dim(dim_type.set)
new_domain = new_domain.add_dims(dim_type.set, 1)
new_domain = new_domain.set_dim_name(dim_type.set, new_dim_idx, new_iname)
joint_aff = zero = isl.Aff.zero_on_domain(new_domain.space)
subst_dict = {}
base_divisor = 1
from pymbolic import var
for i, iname in enumerate(inames):
iname_dt, iname_idx = zero.get_space().get_var_dict()[iname]
iname_aff = zero.add_coefficient_val(iname_dt, iname_idx, 1)
joint_aff = joint_aff + base_divisor * iname_aff
bounds = kernel.get_iname_bounds(iname, constants_only=True)
from loopy.isl_helpers import (static_max_of_pw_aff,
static_value_of_pw_aff)
from loopy.symbolic import pw_aff_to_expr
length = int(
pw_aff_to_expr(
static_max_of_pw_aff(bounds.size, constants_only=True)))
try:
lower_bound_aff = static_value_of_pw_aff(
bounds.lower_bound_pw_aff.coalesce(), constants_only=False)
except Exception as e:
raise type(e)("while finding lower bound of '%s': " % iname)
my_val = var(new_iname) // base_divisor
if i + 1 < len(inames):
my_val %= length
my_val += pw_aff_to_expr(lower_bound_aff)
subst_dict[iname] = my_val
base_divisor *= length
from loopy.isl_helpers import iname_rel_aff
new_domain = new_domain.add_constraint(
isl.Constraint.equality_from_aff(
iname_rel_aff(new_domain.get_space(), new_iname, "==", joint_aff)))
for i, iname in enumerate(inames):
iname_to_dim = new_domain.get_space().get_var_dict()
iname_dt, iname_idx = iname_to_dim[iname]
if within is None:
new_domain = new_domain.project_out(iname_dt, iname_idx, 1)
def subst_within_inames(fid):
result = set()
for iname in fid:
if iname in inames:
result.add(new_iname)
else:
result.add(iname)
return frozenset(result)
new_insns = [
insn.copy(within_inames=subst_within_inames(insn.within_inames))
for insn in kernel.instructions
]
kernel = (kernel.copy(
instructions=new_insns,
domains=domch.get_domains_with(new_domain),
applied_iname_rewrites=kernel.applied_iname_rewrites + [subst_dict]))
from loopy.match import parse_stack_match
within = parse_stack_match(within)
from pymbolic.mapper.substitutor import make_subst_func
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
ijoin = _InameJoiner(rule_mapping_context, within,
make_subst_func(subst_dict), inames, new_iname)
kernel = rule_mapping_context.finish_kernel(ijoin.map_kernel(kernel))
if tag is not None:
kernel = tag_inames(kernel, {new_iname: tag})
return kernel