def rename_argument(kernel, old_name, new_name, existing_ok=False): """ .. versionadded:: 2016.2 """ var_name_gen = kernel.get_var_name_generator() if old_name not in kernel.arg_dict: raise LoopyError("old arg name '%s' does not exist" % old_name) does_exist = var_name_gen.is_name_conflicting(new_name) if does_exist and not existing_ok: raise LoopyError( "argument name '%s' conflicts with an existing identifier" "--cannot rename" % new_name) # {{{ instructions from pymbolic import var subst_dict = {old_name: var(new_name)} from loopy.symbolic import (RuleAwareSubstitutionMapper, SubstitutionRuleMappingContext) from pymbolic.mapper.substitutor import make_subst_func rule_mapping_context = SubstitutionRuleMappingContext( kernel.substitutions, var_name_gen) smap = RuleAwareSubstitutionMapper(rule_mapping_context, make_subst_func(subst_dict), within=lambda kernel, insn, stack: True) kernel = rule_mapping_context.finish_kernel(smap.map_kernel(kernel)) # }}} # {{{ args new_args = [] for arg in kernel.args: if arg.name == old_name: arg = arg.copy(name=new_name) new_args.append(arg) # }}} # {{{ domain new_domains = [] for dom in kernel.domains: dom_var_dict = dom.get_var_dict() if old_name in dom_var_dict: dt, pos = dom_var_dict[old_name] dom = dom.set_dim_name(dt, pos, new_name) new_domains.append(dom) # }}} return kernel.copy(domains=new_domains, args=new_args)
def _apply_renames_in_exprs(kernel, var_renames): from loopy.symbolic import (SubstitutionRuleMappingContext, RuleAwareSubstitutionMapper) from pymbolic.mapper.substitutor import make_subst_func from loopy.match import parse_stack_match srmc = SubstitutionRuleMappingContext(kernel.substitutions, kernel.get_var_name_generator()) subst_map = RuleAwareSubstitutionMapper(srmc, make_subst_func(var_renames), within=parse_stack_match(None)) return subst_map.map_kernel(kernel)
def rename_callable(program, old_name, new_name=None, existing_ok=False): """ :arg program: An instance of :class:`loopy.TranslationUnit` :arg old_name: The callable to be renamed :arg new_name: New name for the callable to be renamed :arg existing_ok: An instance of :class:`bool` """ from loopy.symbolic import (RuleAwareSubstitutionMapper, SubstitutionRuleMappingContext) from pymbolic import var assert isinstance(program, TranslationUnit) assert isinstance(old_name, str) if (new_name in program.callables_table) and not existing_ok: raise LoopyError(f"callables named '{new_name}' already exists") if new_name is None: namegen = UniqueNameGenerator(program.callables_table.keys()) new_name = namegen(old_name) assert isinstance(new_name, str) new_callables_table = {} for name, clbl in program.callables_table.items(): if name == old_name: name = new_name if isinstance(clbl, CallableKernel): knl = clbl.subkernel rule_mapping_context = SubstitutionRuleMappingContext( knl.substitutions, knl.get_var_name_generator()) smap = RuleAwareSubstitutionMapper(rule_mapping_context, {var(old_name): var(new_name)}.get, within=lambda *args: True) knl = rule_mapping_context.finish_kernel(smap.map_kernel(knl)) clbl = clbl.copy(subkernel=knl.copy(name=name)) elif isinstance(clbl, ScalarCallable): pass else: raise NotImplementedError(f"{type(clbl)}") new_callables_table[name] = clbl new_entrypoints = program.entrypoints.copy() if old_name in new_entrypoints: new_entrypoints = ((new_entrypoints | frozenset([new_name])) - frozenset([old_name])) return program.copy(callables_table=new_callables_table, entrypoints=new_entrypoints)
def _apply_renames_in_exprs(kernel, var_renames): from loopy.symbolic import ( SubstitutionRuleMappingContext, RuleAwareSubstitutionMapper) from pymbolic.mapper.substitutor import make_subst_func from loopy.match import parse_stack_match srmc = SubstitutionRuleMappingContext( kernel.substitutions, kernel.get_var_name_generator()) subst_map = RuleAwareSubstitutionMapper( srmc, make_subst_func(var_renames), within=parse_stack_match(None)) return subst_map.map_kernel(kernel)
def rename_argument(kernel, old_name, new_name, existing_ok=False): """ .. versionadded:: 2016.2 """ var_name_gen = kernel.get_var_name_generator() if old_name not in kernel.arg_dict: raise LoopyError("old arg name '%s' does not exist" % old_name) does_exist = var_name_gen.is_name_conflicting(new_name) if does_exist and not existing_ok: raise LoopyError("argument name '%s' conflicts with an existing identifier" "--cannot rename" % new_name) from pymbolic import var subst_dict = {old_name: var(new_name)} from loopy.symbolic import ( RuleAwareSubstitutionMapper, SubstitutionRuleMappingContext) from pymbolic.mapper.substitutor import make_subst_func rule_mapping_context = SubstitutionRuleMappingContext( kernel.substitutions, var_name_gen) smap = RuleAwareSubstitutionMapper(rule_mapping_context, make_subst_func(subst_dict), within=lambda knl, insn, stack: True) kernel = smap.map_kernel(kernel) new_args = [] for arg in kernel.args: if arg.name == old_name: arg = arg.copy(name=new_name) new_args.append(arg) return kernel.copy(args=new_args)
def _fix_parameter(kernel, name, value, within=None): def process_set(s): var_dict = s.get_var_dict() try: dt, idx = var_dict[name] except KeyError: return s value_aff = isl.Aff.zero_on_domain(s.space) + value from loopy.isl_helpers import iname_rel_aff name_equal_value_aff = iname_rel_aff(s.space, name, "==", value_aff) s = (s.add_constraint( isl.Constraint.equality_from_aff( name_equal_value_aff)).project_out(dt, idx, 1)) return s new_domains = [process_set(dom) for dom in kernel.domains] from pymbolic.mapper.substitutor import make_subst_func subst_func = make_subst_func({name: value}) from loopy.symbolic import SubstitutionMapper, PartialEvaluationMapper subst_map = SubstitutionMapper(subst_func) ev_map = PartialEvaluationMapper() def map_expr(expr): return ev_map(subst_map(expr)) from loopy.kernel.array import ArrayBase new_args = [] for arg in kernel.args: if arg.name == name: # remove from argument list continue if not isinstance(arg, ArrayBase): new_args.append(arg) else: new_args.append(arg.map_exprs(map_expr)) new_temp_vars = {} for tv in kernel.temporary_variables.values(): new_temp_vars[tv.name] = tv.map_exprs(map_expr) from loopy.match import parse_stack_match within = parse_stack_match(within) rule_mapping_context = SubstitutionRuleMappingContext( kernel.substitutions, kernel.get_var_name_generator()) esubst_map = RuleAwareSubstitutionMapper(rule_mapping_context, subst_func, within=within) return (rule_mapping_context.finish_kernel( esubst_map.map_kernel( kernel, within=within, # overwritten below, no need to map map_tvs=False, map_args=False)).copy( domains=new_domains, args=new_args, temporary_variables=new_temp_vars, assumptions=process_set(kernel.assumptions), ))
def _inline_call_instruction(caller_knl, callee_knl, call_insn): """ Returns a copy of *caller_knl* with the *call_insn* in the *kernel* replaced by inlining *callee_knl* into it within it. :arg call_insn: An instance of `loopy.CallInstruction` of the call-site. """ import pymbolic.primitives as prim from pymbolic.mapper.substitutor import make_subst_func from loopy.kernel.data import ValueArg # {{{ sanity checks assert call_insn.expression.function.name == callee_knl.name # }}} callee_label = callee_knl.name[:4] + "_" vng = caller_knl.get_var_name_generator() ing = caller_knl.get_instruction_id_generator() # {{{ construct callee->caller name mappings # name_map: Mapping[str, str] # A mapping from variable names in the callee kernel's namespace to # the ones they would be referred by in the caller's namespace post inlining. name_map = {} # only consider temporary variables and inames, arguments would be mapping # according to the invocation in call_insn. for name in (callee_knl.all_inames() | set(callee_knl.temporary_variables.keys())): new_name = vng(callee_label + name) name_map[name] = new_name # }}} # {{{ iname_to_tags # new_inames: caller's inames post inlining new_inames = caller_knl.inames for old_name, callee_iname in callee_knl.inames.items(): new_name = name_map[old_name] new_inames[new_name] = callee_iname.copy(name=new_name) # }}} # {{{ register callee's temps as caller's # new_temps: caller's temps post inlining new_temps = caller_knl.temporary_variables.copy() for name, tv in callee_knl.temporary_variables.items(): new_temps[name_map[name]] = tv.copy(name=name_map[name]) # }}} # {{{ get callee args -> parameters passed to the call arg_map = {} # callee arg name -> caller symbols (e.g. SubArrayRef) assignees = call_insn.assignees # writes parameters = call_insn.expression.parameters # reads from loopy.kernel.function_interface import get_kw_pos_association kw_to_pos, pos_to_kw = get_kw_pos_association(callee_knl) for i, par in enumerate(parameters): arg_map[pos_to_kw[i]] = par for i, assignee in enumerate(assignees): arg_map[pos_to_kw[-i - 1]] = assignee # }}} # {{{ process domains/assumptions # rename inames new_domains = callee_knl.domains.copy() for old_iname in callee_knl.all_inames(): new_domains = [ rename_iname(dom, old_iname, name_map[old_iname]) for dom in new_domains ] # realize domains' dim params in terms of caller's variables new_assumptions = callee_knl.assumptions for callee_arg_name, param_expr in arg_map.items(): if isinstance(callee_knl.arg_dict[callee_arg_name], ValueArg): new_domains = [ substitute_into_domain( dom, callee_arg_name, param_expr, get_valid_domain_param_names(caller_knl)) for dom in new_domains ] new_assumptions = substitute_into_domain( new_assumptions, callee_arg_name, param_expr, get_valid_domain_param_names(caller_knl)) # }}} # {{{ rename inames/temporaries in the program rule_mapping_context = SubstitutionRuleMappingContext( callee_knl.substitutions, vng) subst_func = make_subst_func({ old_name: prim.Variable(new_name) for old_name, new_name in name_map.items() }) inames_temps_renamer = RuleAwareSubstitutionMapper( rule_mapping_context, subst_func, within=lambda *args: True) callee_knl = rule_mapping_context.finish_kernel( inames_temps_renamer.map_kernel(callee_knl)) # }}} # {{{ map callee's expressions to get expressions after inlining rule_mapping_context = SubstitutionRuleMappingContext( callee_knl.substitutions, vng) smap = KernelArgumentSubstitutor(rule_mapping_context, caller_knl, callee_knl, arg_map) callee_knl = rule_mapping_context.finish_kernel( smap.map_kernel(callee_knl)) # }}} # {{{ generate new ids for instructions insn_id_map = {} for insn in callee_knl.instructions: insn_id_map[insn.id] = ing(callee_label + insn.id) # }}} # {{{ use NoOp to mark the start and end of callee kernel from loopy.kernel.instruction import NoOpInstruction noop_start = NoOpInstruction(id=ing(callee_label + "_start"), within_inames=call_insn.within_inames, depends_on=call_insn.depends_on) noop_end = NoOpInstruction(id=call_insn.id, within_inames=call_insn.within_inames, depends_on=frozenset(insn_id_map.values())) # }}} # {{{ map callee's instruction ids inlined_insns = [noop_start] for insn in callee_knl.instructions: new_within_inames = (frozenset(name_map[iname] for iname in insn.within_inames) | call_insn.within_inames) new_depends_on = (frozenset(insn_id_map[dep] for dep in insn.depends_on) | {noop_start.id}) new_no_sync_with = frozenset( (insn_id_map[id], scope) for id, scope in insn.no_sync_with) new_id = insn_id_map[insn.id] if isinstance(insn, Assignment): new_atomicity = tuple( type(atomicity)(name_map[atomicity.var_name]) for atomicity in insn.atomicity) insn = insn.copy(id=insn_id_map[insn.id], within_inames=new_within_inames, depends_on=new_depends_on, tags=insn.tags | call_insn.tags, atomicity=new_atomicity, no_sync_with=new_no_sync_with) else: insn = insn.copy(id=new_id, within_inames=new_within_inames, depends_on=new_depends_on, tags=insn.tags | call_insn.tags, no_sync_with=new_no_sync_with) inlined_insns.append(insn) inlined_insns.append(noop_end) # }}} # {{{ swap out call_insn with inlined_instructions idx = caller_knl.instructions.index(call_insn) new_insns = (caller_knl.instructions[:idx] + inlined_insns + caller_knl.instructions[idx + 1:]) # }}} old_assumptions, new_assumptions = isl.align_two(caller_knl.assumptions, new_assumptions) return caller_knl.copy(instructions=new_insns, temporary_variables=new_temps, domains=caller_knl.domains + new_domains, assumptions=(old_assumptions.params() & new_assumptions.params()), inames=new_inames)
def _fix_parameter(kernel, name, value): def process_set(s): var_dict = s.get_var_dict() try: dt, idx = var_dict[name] except KeyError: return s value_aff = isl.Aff.zero_on_domain(s.space) + value from loopy.isl_helpers import iname_rel_aff name_equal_value_aff = iname_rel_aff(s.space, name, "==", value_aff) s = s.add_constraint(isl.Constraint.equality_from_aff(name_equal_value_aff)).project_out(dt, idx, 1) return s new_domains = [process_set(dom) for dom in kernel.domains] from pymbolic.mapper.substitutor import make_subst_func subst_func = make_subst_func({name: value}) from loopy.symbolic import SubstitutionMapper, PartialEvaluationMapper subst_map = SubstitutionMapper(subst_func) ev_map = PartialEvaluationMapper() def map_expr(expr): return ev_map(subst_map(expr)) from loopy.kernel.array import ArrayBase new_args = [] for arg in kernel.args: if arg.name == name: # remove from argument list continue if not isinstance(arg, ArrayBase): new_args.append(arg) else: new_args.append(arg.map_exprs(map_expr)) new_temp_vars = {} for tv in six.itervalues(kernel.temporary_variables): new_temp_vars[tv.name] = tv.map_exprs(map_expr) from loopy.match import parse_stack_match within = parse_stack_match(None) rule_mapping_context = SubstitutionRuleMappingContext(kernel.substitutions, kernel.get_var_name_generator()) esubst_map = RuleAwareSubstitutionMapper(rule_mapping_context, subst_func, within=within) return rule_mapping_context.finish_kernel(esubst_map.map_kernel(kernel)).copy( domains=new_domains, args=new_args, temporary_variables=new_temp_vars, assumptions=process_set(kernel.assumptions), )
def affine_map_inames(kernel, old_inames, new_inames, equations): """Return a new *kernel* where the affine transform specified by *equations* has been applied to the inames. :arg old_inames: A list of inames to be replaced by affine transforms of their values. May also be a string of comma-separated inames. :arg new_inames: A list of new inames that are not yet used in *kernel*, but have their values established in terms of *old_inames* by *equations*. May also be a string of comma-separated inames. :arg equations: A list of equations estabilishing a relationship between *old_inames* and *new_inames*. Each equation may be a tuple ``(lhs, rhs)`` of expressions or a string, with left and right hand side of the equation separated by ``=``. """ # {{{ check and parse arguments if isinstance(new_inames, str): new_inames = new_inames.split(",") new_inames = [iname.strip() for iname in new_inames] if isinstance(old_inames, str): old_inames = old_inames.split(",") old_inames = [iname.strip() for iname in old_inames] if isinstance(equations, str): equations = [equations] import re eqn_re = re.compile(r"^([^=]+)=([^=]+)$") def parse_equation(eqn): if isinstance(eqn, str): eqn_match = eqn_re.match(eqn) if not eqn_match: raise ValueError("invalid equation: %s" % eqn) from loopy.symbolic import parse lhs = parse(eqn_match.group(1)) rhs = parse(eqn_match.group(2)) return (lhs, rhs) elif isinstance(eqn, tuple): if len(eqn) != 2: raise ValueError("unexpected length of equation tuple, " "got %d, should be 2" % len(eqn)) return eqn else: raise ValueError("unexpected type of equation" "got %d, should be string or tuple" % type(eqn).__name__) equations = [parse_equation(eqn) for eqn in equations] all_vars = kernel.all_variable_names() for iname in new_inames: if iname in all_vars: raise LoopyError("new iname '%s' is already used in kernel" % iname) for iname in old_inames: if iname not in kernel.all_inames(): raise LoopyError("old iname '%s' not known" % iname) # }}} # {{{ substitute iname use from pymbolic.algorithm import solve_affine_equations_for old_inames_to_expr = solve_affine_equations_for(old_inames, equations) subst_dict = dict((v.name, expr) for v, expr in old_inames_to_expr.items()) var_name_gen = kernel.get_var_name_generator() from pymbolic.mapper.substitutor import make_subst_func from loopy.match import parse_stack_match rule_mapping_context = SubstitutionRuleMappingContext( kernel.substitutions, var_name_gen) old_to_new = RuleAwareSubstitutionMapper(rule_mapping_context, make_subst_func(subst_dict), within=parse_stack_match(None)) kernel = (rule_mapping_context.finish_kernel( old_to_new.map_kernel(kernel)).copy( applied_iname_rewrites=kernel.applied_iname_rewrites + [subst_dict])) # }}} # {{{ change domains new_inames_set = frozenset(new_inames) old_inames_set = frozenset(old_inames) new_domains = [] for idom, dom in enumerate(kernel.domains): dom_var_dict = dom.get_var_dict() old_iname_overlap = [ iname for iname in old_inames if iname in dom_var_dict ] if not old_iname_overlap: new_domains.append(dom) continue from loopy.symbolic import get_dependencies dom_new_inames = set() dom_old_inames = set() # mapping for new inames to dim_types new_iname_dim_types = {} dom_equations = [] for iname in old_iname_overlap: for ieqn, (lhs, rhs) in enumerate(equations): eqn_deps = get_dependencies(lhs) | get_dependencies(rhs) if iname in eqn_deps: dom_new_inames.update(eqn_deps & new_inames_set) dom_old_inames.update(eqn_deps & old_inames_set) if dom_old_inames: dom_equations.append((lhs, rhs)) this_eqn_old_iname_dim_types = set(dom_var_dict[old_iname][0] for old_iname in eqn_deps & old_inames_set) if this_eqn_old_iname_dim_types: if len(this_eqn_old_iname_dim_types) > 1: raise ValueError( "inames '%s' (from equation %d (0-based)) " "in domain %d (0-based) are not " "of a uniform dim_type" % (", ".join(eqn_deps & old_inames_set), ieqn, idom)) this_eqn_new_iname_dim_type, = this_eqn_old_iname_dim_types for new_iname in eqn_deps & new_inames_set: if new_iname in new_iname_dim_types: if (this_eqn_new_iname_dim_type != new_iname_dim_types[new_iname]): raise ValueError( "dim_type disagreement for " "iname '%s' (from equation %d (0-based)) " "in domain %d (0-based)" % (new_iname, ieqn, idom)) else: new_iname_dim_types[new_iname] = \ this_eqn_new_iname_dim_type if not dom_old_inames <= set(dom_var_dict): raise ValueError( "domain %d (0-based) does not know about " "all old inames (specifically '%s') needed to define new inames" % (idom, ", ".join(dom_old_inames - set(dom_var_dict)))) # add inames to domain with correct dim_types dom_new_inames = list(dom_new_inames) for iname in dom_new_inames: dt = new_iname_dim_types[iname] iname_idx = dom.dim(dt) dom = dom.add_dims(dt, 1) dom = dom.set_dim_name(dt, iname_idx, iname) # add equations from loopy.symbolic import aff_from_expr for lhs, rhs in dom_equations: dom = dom.add_constraint( isl.Constraint.equality_from_aff( aff_from_expr(dom.space, rhs - lhs))) # project out old inames for iname in dom_old_inames: dt, idx = dom.get_var_dict()[iname] dom = dom.project_out(dt, idx, 1) new_domains.append(dom) # }}} # {{{ switch iname refs in instructions def fix_iname_set(insn_id, inames): if old_inames_set <= inames: return (inames - old_inames_set) | new_inames_set elif old_inames_set & inames: raise LoopyError( "instruction '%s' uses only a part (%s), not all, " "of the old inames" % (insn_id, ", ".join(old_inames_set & inames))) else: return inames new_instructions = [ insn.copy(within_inames=fix_iname_set(insn.id, insn.within_inames)) for insn in kernel.instructions ] # }}} return kernel.copy(domains=new_domains, instructions=new_instructions)
def affine_map_inames(kernel, old_inames, new_inames, equations): """Return a new *kernel* where the affine transform specified by *equations* has been applied to the inames. :arg old_inames: A list of inames to be replaced by affine transforms of their values. May also be a string of comma-separated inames. :arg new_inames: A list of new inames that are not yet used in *kernel*, but have their values established in terms of *old_inames* by *equations*. May also be a string of comma-separated inames. :arg equations: A list of equations estabilishing a relationship between *old_inames* and *new_inames*. Each equation may be a tuple ``(lhs, rhs)`` of expressions or a string, with left and right hand side of the equation separated by ``=``. """ # {{{ check and parse arguments if isinstance(new_inames, str): new_inames = new_inames.split(",") new_inames = [iname.strip() for iname in new_inames] if isinstance(old_inames, str): old_inames = old_inames.split(",") old_inames = [iname.strip() for iname in old_inames] if isinstance(equations, str): equations = [equations] import re eqn_re = re.compile(r"^([^=]+)=([^=]+)$") def parse_equation(eqn): if isinstance(eqn, str): eqn_match = eqn_re.match(eqn) if not eqn_match: raise ValueError("invalid equation: %s" % eqn) from loopy.symbolic import parse lhs = parse(eqn_match.group(1)) rhs = parse(eqn_match.group(2)) return (lhs, rhs) elif isinstance(eqn, tuple): if len(eqn) != 2: raise ValueError("unexpected length of equation tuple, " "got %d, should be 2" % len(eqn)) return eqn else: raise ValueError("unexpected type of equation" "got %d, should be string or tuple" % type(eqn).__name__) equations = [parse_equation(eqn) for eqn in equations] all_vars = kernel.all_variable_names() for iname in new_inames: if iname in all_vars: raise LoopyError("new iname '%s' is already used in kernel" % iname) for iname in old_inames: if iname not in kernel.all_inames(): raise LoopyError("old iname '%s' not known" % iname) # }}} # {{{ substitute iname use from pymbolic.algorithm import solve_affine_equations_for old_inames_to_expr = solve_affine_equations_for(old_inames, equations) subst_dict = dict( (v.name, expr) for v, expr in old_inames_to_expr.items()) var_name_gen = kernel.get_var_name_generator() from pymbolic.mapper.substitutor import make_subst_func from loopy.context_matching import parse_stack_match rule_mapping_context = SubstitutionRuleMappingContext( kernel.substitutions, var_name_gen) old_to_new = RuleAwareSubstitutionMapper(rule_mapping_context, make_subst_func(subst_dict), within=parse_stack_match(None)) kernel = ( rule_mapping_context.finish_kernel( old_to_new.map_kernel(kernel)) .copy( applied_iname_rewrites=kernel.applied_iname_rewrites + [subst_dict] )) # }}} # {{{ change domains new_inames_set = set(new_inames) old_inames_set = set(old_inames) new_domains = [] for idom, dom in enumerate(kernel.domains): dom_var_dict = dom.get_var_dict() old_iname_overlap = [ iname for iname in old_inames if iname in dom_var_dict] if not old_iname_overlap: new_domains.append(dom) continue from loopy.symbolic import get_dependencies dom_new_inames = set() dom_old_inames = set() # mapping for new inames to dim_types new_iname_dim_types = {} dom_equations = [] for iname in old_iname_overlap: for ieqn, (lhs, rhs) in enumerate(equations): eqn_deps = get_dependencies(lhs) | get_dependencies(rhs) if iname in eqn_deps: dom_new_inames.update(eqn_deps & new_inames_set) dom_old_inames.update(eqn_deps & old_inames_set) if dom_old_inames: dom_equations.append((lhs, rhs)) this_eqn_old_iname_dim_types = set( dom_var_dict[old_iname][0] for old_iname in eqn_deps & old_inames_set) if this_eqn_old_iname_dim_types: if len(this_eqn_old_iname_dim_types) > 1: raise ValueError("inames '%s' (from equation %d (0-based)) " "in domain %d (0-based) are not " "of a uniform dim_type" % (", ".join(eqn_deps & old_inames_set), ieqn, idom)) this_eqn_new_iname_dim_type, = this_eqn_old_iname_dim_types for new_iname in eqn_deps & new_inames_set: if new_iname in new_iname_dim_types: if (this_eqn_new_iname_dim_type != new_iname_dim_types[new_iname]): raise ValueError("dim_type disagreement for " "iname '%s' (from equation %d (0-based)) " "in domain %d (0-based)" % (new_iname, ieqn, idom)) else: new_iname_dim_types[new_iname] = \ this_eqn_new_iname_dim_type if not dom_old_inames <= set(dom_var_dict): raise ValueError("domain %d (0-based) does not know about " "all old inames (specifically '%s') needed to define new inames" % (idom, ", ".join(dom_old_inames - set(dom_var_dict)))) # add inames to domain with correct dim_types dom_new_inames = list(dom_new_inames) for iname in dom_new_inames: dt = new_iname_dim_types[iname] iname_idx = dom.dim(dt) dom = dom.add_dims(dt, 1) dom = dom.set_dim_name(dt, iname_idx, iname) # add equations from loopy.symbolic import aff_from_expr for lhs, rhs in dom_equations: dom = dom.add_constraint( isl.Constraint.equality_from_aff( aff_from_expr(dom.space, rhs - lhs))) # project out old inames for iname in dom_old_inames: dt, idx = dom.get_var_dict()[iname] dom = dom.project_out(dt, idx, 1) new_domains.append(dom) # }}} return kernel.copy(domains=new_domains)
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 rename_iname(knl, old_iname, new_iname, existing_ok=False, within=None): """ :arg within: a stack match as understood by :func:`loopy.context_matching.parse_stack_match`. :arg existing_ok: execute even if *new_iname* already exists """ var_name_gen = knl.get_var_name_generator() does_exist = var_name_gen.is_name_conflicting(new_iname) if does_exist and not existing_ok: raise ValueError("iname '%s' conflicts with an existing identifier" "--cannot rename" % new_iname) if does_exist: # {{{ check that the domains match up dom = knl.get_inames_domain(frozenset((old_iname, new_iname))) var_dict = dom.get_var_dict() _, old_idx = var_dict[old_iname] _, new_idx = var_dict[new_iname] par_idx = dom.dim(dim_type.param) dom_old = dom.move_dims( dim_type.param, par_idx, dim_type.set, old_idx, 1) dom_old = dom_old.move_dims( dim_type.set, dom_old.dim(dim_type.set), dim_type.param, par_idx, 1) dom_old = dom_old.project_out( dim_type.set, new_idx if new_idx < old_idx else new_idx - 1, 1) par_idx = dom.dim(dim_type.param) dom_new = dom.move_dims( dim_type.param, par_idx, dim_type.set, new_idx, 1) dom_new = dom_new.move_dims( dim_type.set, dom_new.dim(dim_type.set), dim_type.param, par_idx, 1) dom_new = dom_new.project_out( dim_type.set, old_idx if old_idx < new_idx else old_idx - 1, 1) if not (dom_old <= dom_new and dom_new <= dom_old): raise LoopyError( "inames {old} and {new} do not iterate over the same domain" .format(old=old_iname, new=new_iname)) # }}} from pymbolic import var subst_dict = {old_iname: var(new_iname)} 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)) new_instructions = [] for insn in knl.instructions: if (old_iname in insn.forced_iname_deps and within(knl, insn, ())): insn = insn.copy( forced_iname_deps=( (insn.forced_iname_deps - frozenset([old_iname])) | frozenset([new_iname]))) new_instructions.append(insn) knl = knl.copy(instructions=new_instructions) else: knl = duplicate_inames( knl, [old_iname], within=within, new_inames=[new_iname]) knl = remove_unused_inames(knl, [old_iname]) return knl
def fix_parameters(kernel, within=None, **value_dict): """Fix the values of the arguments to specific constants. *value_dict* consists of *name*/*value* pairs, where *name* will be fixed to be *value*. *name* may refer to :ref:`domain-parameters` or :ref:`arguments`. """ if not value_dict: return kernel def process_set_one_param(s, name, value): var_dict = s.get_var_dict() try: dt, idx = var_dict[name] except KeyError: return s value_aff = isl.Aff.zero_on_domain(s.space) + value from loopy.isl_helpers import iname_rel_aff name_equal_value_aff = iname_rel_aff(s.space, name, "==", value_aff) s = (s.add_constraint( isl.Constraint.equality_from_aff( name_equal_value_aff)).project_out(dt, idx, 1)) return s def process_set(s): for name, value in value_dict.items(): s = process_set_one_param(s, name, value) return s new_domains = [process_set(dom) for dom in kernel.domains] from pymbolic.mapper.substitutor import make_subst_func subst_func = make_subst_func(value_dict) from loopy.symbolic import SubstitutionMapper, PartialEvaluationMapper subst_map = SubstitutionMapper(subst_func) ev_map = PartialEvaluationMapper() def map_expr(expr): return ev_map(subst_map(expr)) from loopy.kernel.array import ArrayBase new_args = [] for arg in kernel.args: if arg.name in value_dict.keys(): # remove from argument list continue if not isinstance(arg, ArrayBase): new_args.append(arg) else: new_args.append(arg.map_exprs(map_expr)) new_temp_vars = {} for tv in kernel.temporary_variables.values(): new_temp_vars[tv.name] = tv.map_exprs(map_expr) from loopy.match import parse_stack_match within = parse_stack_match(within) rule_mapping_context = SubstitutionRuleMappingContext( kernel.substitutions, kernel.get_var_name_generator()) esubst_map = RuleAwareSubstitutionMapper(rule_mapping_context, subst_func, within=within) return (rule_mapping_context.finish_kernel( esubst_map.map_kernel(kernel, within=within)).copy( domains=new_domains, args=new_args, temporary_variables=new_temp_vars, assumptions=process_set(kernel.assumptions), ))
def rename_iname(knl, old_iname, new_iname, existing_ok=False, within=None): """ :arg within: a stack match as understood by :func:`loopy.match.parse_stack_match`. :arg existing_ok: execute even if *new_iname* already exists """ var_name_gen = knl.get_var_name_generator() does_exist = var_name_gen.is_name_conflicting(new_iname) if old_iname not in knl.all_inames(): raise LoopyError("old iname '%s' does not exist" % old_iname) if does_exist and not existing_ok: raise LoopyError("iname '%s' conflicts with an existing identifier" "--cannot rename" % new_iname) if does_exist: # {{{ check that the domains match up dom = knl.get_inames_domain(frozenset((old_iname, new_iname))) var_dict = dom.get_var_dict() _, old_idx = var_dict[old_iname] _, new_idx = var_dict[new_iname] par_idx = dom.dim(dim_type.param) dom_old = dom.move_dims(dim_type.param, par_idx, dim_type.set, old_idx, 1) dom_old = dom_old.move_dims(dim_type.set, dom_old.dim(dim_type.set), dim_type.param, par_idx, 1) dom_old = dom_old.project_out( dim_type.set, new_idx if new_idx < old_idx else new_idx - 1, 1) par_idx = dom.dim(dim_type.param) dom_new = dom.move_dims(dim_type.param, par_idx, dim_type.set, new_idx, 1) dom_new = dom_new.move_dims(dim_type.set, dom_new.dim(dim_type.set), dim_type.param, par_idx, 1) dom_new = dom_new.project_out( dim_type.set, old_idx if old_idx < new_idx else old_idx - 1, 1) if not (dom_old <= dom_new and dom_new <= dom_old): raise LoopyError( "inames {old} and {new} do not iterate over the same domain". format(old=old_iname, new=new_iname)) # }}} from pymbolic import var subst_dict = {old_iname: var(new_iname)} from loopy.match 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) smap = RuleAwareSubstitutionMapper(rule_mapping_context, make_subst_func(subst_dict), within) knl = rule_mapping_context.finish_kernel(smap.map_kernel(knl)) new_instructions = [] for insn in knl.instructions: if (old_iname in insn.within_inames and within(knl, insn, ())): insn = insn.copy(within_inames=( (insn.within_inames - frozenset([old_iname])) | frozenset([new_iname]))) new_instructions.append(insn) knl = knl.copy(instructions=new_instructions) else: knl = duplicate_inames(knl, [old_iname], within=within, new_inames=[new_iname]) knl = remove_unused_inames(knl, [old_iname]) return knl
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 _fuse_two_kernels(knla, knlb): from loopy.kernel import kernel_state if knla.state != kernel_state.INITIAL or knlb.state != kernel_state.INITIAL: raise LoopyError("can only fuse kernels in INITIAL state") # {{{ fuse domains new_domains = knla.domains[:] for dom_b in knlb.domains: i_fuse = _find_fusable_loop_domain_index(dom_b, new_domains) if i_fuse is None: new_domains.append(dom_b) else: dom_a = new_domains[i_fuse] dom_a, dom_b = isl.align_two(dom_a, dom_b) shared_inames = list( set(dom_a.get_var_dict(dim_type.set)) & set(dom_b.get_var_dict(dim_type.set))) dom_a_s = dom_a.project_out_except(shared_inames, [dim_type.set]) dom_b_s = dom_a.project_out_except(shared_inames, [dim_type.set]) if not (dom_a_s <= dom_b_s and dom_b_s <= dom_a_s): raise LoopyError("kernels do not agree on domain of " "inames '%s'" % (",".join(shared_inames))) new_domain = dom_a & dom_b new_domains[i_fuse] = new_domain # }}} vng = knla.get_var_name_generator() b_var_renames = {} # {{{ fuse args new_args = knla.args[:] for b_arg in knlb.args: if b_arg.name not in knla.arg_dict: new_arg_name = vng(b_arg.name) if new_arg_name != b_arg.name: b_var_renames[b_arg.name] = var(new_arg_name) new_args.append(b_arg.copy(name=new_arg_name)) else: if b_arg != knla.arg_dict[b_arg.name]: raise LoopyError( "argument '%s' has inconsistent definition between " "the two kernels being merged" % b_arg.name) # }}} # {{{ fuse temporaries new_temporaries = knla.temporary_variables.copy() for b_name, b_tv in six.iteritems(knlb.temporary_variables): assert b_name == b_tv.name new_tv_name = vng(b_name) if new_tv_name != b_name: b_var_renames[b_name] = var(new_tv_name) assert new_tv_name not in new_temporaries new_temporaries[new_tv_name] = b_tv.copy(name=new_tv_name) # }}} # {{{ apply renames in kernel b from loopy.symbolic import ( SubstitutionRuleMappingContext, RuleAwareSubstitutionMapper) from pymbolic.mapper.substitutor import make_subst_func from loopy.context_matching import parse_stack_match srmc = SubstitutionRuleMappingContext( knlb.substitutions, knlb.get_var_name_generator()) subst_map = RuleAwareSubstitutionMapper( srmc, make_subst_func(b_var_renames), within=parse_stack_match(None)) knlb = subst_map.map_kernel(knlb) # }}} # {{{ fuse instructions new_instructions = knla.instructions[:] from pytools import UniqueNameGenerator insn_id_gen = UniqueNameGenerator( set([insna.id for insna in new_instructions])) knl_b_instructions = [] old_b_id_to_new_b_id = {} for insnb in knlb.instructions: old_id = insnb.id new_id = insn_id_gen(old_id) old_b_id_to_new_b_id[old_id] = new_id knl_b_instructions.append( insnb.copy(id=new_id)) for insnb in knl_b_instructions: new_instructions.append( insnb.copy( insn_deps=frozenset( old_b_id_to_new_b_id[dep_id] for dep_id in insnb.insn_deps))) # }}} # {{{ fuse assumptions assump_a = knla.assumptions assump_b = knlb.assumptions assump_a, assump_b = isl.align_two(assump_a, assump_b) shared_param_names = list( set(dom_a.get_var_dict(dim_type.set)) & set(dom_b.get_var_dict(dim_type.set))) assump_a_s = assump_a.project_out_except(shared_param_names, [dim_type.param]) assump_b_s = assump_a.project_out_except(shared_param_names, [dim_type.param]) if not (assump_a_s <= assump_b_s and assump_b_s <= assump_a_s): raise LoopyError("assumptions do not agree on kernels to be merged") new_assumptions = (assump_a & assump_b).params() # }}} from loopy.kernel import LoopKernel return LoopKernel( domains=new_domains, instructions=new_instructions, args=new_args, name="%s_and_%s" % (knla.name, knlb.name), preambles=_ordered_merge_lists(knla.preambles, knlb.preambles), preamble_generators=_ordered_merge_lists( knla.preamble_generators, knlb.preamble_generators), assumptions=new_assumptions, local_sizes=_merge_dicts( "local size", knla.local_sizes, knlb.local_sizes), temporary_variables=new_temporaries, iname_to_tag=_merge_dicts( "iname-to-tag mapping", knla.iname_to_tag, knlb.iname_to_tag), substitutions=_merge_dicts( "substitution", knla.substitutions, knlb.substitutions), function_manglers=_ordered_merge_lists( knla.function_manglers, knlb.function_manglers), symbol_manglers=_ordered_merge_lists( knla.symbol_manglers, knlb.symbol_manglers), iname_slab_increments=_merge_dicts( "iname slab increment", knla.iname_slab_increments, knlb.iname_slab_increments), loop_priority=_ordered_merge_lists( knla.loop_priority, knlb.loop_priority), silenced_warnings=_ordered_merge_lists( knla.silenced_warnings, knlb.silenced_warnings), applied_iname_rewrites=_ordered_merge_lists( knla.applied_iname_rewrites, knlb.applied_iname_rewrites), index_dtype=_merge_values( "index dtype", knla.index_dtype, knlb.index_dtype), target=_merge_values( "target", knla.target, knlb.target), options=knla.options)