def consume(self, t, src, state):
if src.childtag == "right":
my_sch = self.scheme()
declr_template = self._cgenv.get_template("hash_declaration.cpp")
right_template = self._cgenv.get_template("insert_materialize.cpp")
hashname = self._hashname
keypos = self.right_keypos
keyval = t.get_code(self.right_keypos)
if self.rightCondIsRightAttr:
keytype = self.language().typename(
self.condition.right.typeof(
my_sch,
None))
else:
keytype = self.language().typename(
self.condition.left.typeof(
my_sch,
None))
in_tuple_type = t.getTupleTypename()
in_tuple_name = t.name
# declaration of hash map
hashdeclr = declr_template.render(locals())
state.addDeclarations([hashdeclr])
# materialization point
code = right_template.render(locals())
return code
if src.childtag == "left":
left_template = self._cgenv.get_template("lookup.cpp")
hashname = self._hashname
keyname = t.name
keytype = t.getTupleTypename()
keypos = self.left_keypos
keyval = t.get_code(keypos)
right_tuple_name = gensym()
outTuple = CStagedTupleRef(gensym(), self.scheme())
out_tuple_type_def = outTuple.generateDefinition()
out_tuple_type = outTuple.getTupleTypename()
out_tuple_name = outTuple.name
state.addDeclarations([out_tuple_type_def])
inner_plan_compiled = self.parent().consume(outTuple, self, state)
code = left_template.render(locals())
return code
assert False, "src not equal to left or right"
def consume(self, t, src, state):
access_template = self._cgenv.get_template('hash_insert_lookup.cpp')
hashname = self._hashname
keyname = t.name
side = src.childtag
outTuple = self.outTuple
out_tuple_type = self.outTuple.getTupleTypename()
out_tuple_name = self.outTuple.name
global_syncname = state.getPipelineProperty('global_syncname')
if src.childtag == "right":
left_sch = self.left.scheme()
# save for later
self.right_in_tuple_type = t.getTupleTypename()
state.resolveSymbol(self.rightTypeRef, self.right_in_tuple_type)
inner_plan_compiled = self.parent().consume(outTuple, self, state)
keyval = self.__aggregate_val__(t, self.rightcols)
other_tuple_type = self.leftTypeRef.getPlaceholder()
left_type = other_tuple_type
right_type = self.right_in_tuple_type
left_name = gensym()
right_name = keyname
self.right_name = right_name
valname = left_name
code = access_template.render(locals())
return code
if src.childtag == "left":
right_in_tuple_type = self.right_in_tuple_type
left_in_tuple_type = t.getTupleTypename()
state.resolveSymbol(self.leftTypeRef, left_in_tuple_type)
keyval = self.__aggregate_val__(t, self.leftcols)
inner_plan_compiled = self.parent().consume(outTuple, self, state)
left_type = left_in_tuple_type
right_type = self.right_in_tuple_type
other_tuple_type = self.right_in_tuple_type
left_name = keyname
right_name = gensym()
valname = right_name
code = access_template.render(locals())
return code
assert False, "src not equal to left or right"
def consume(self, t, src, state):
if src.childtag == "right":
declr_template = """std::unordered_map\
<int64_t, std::vector<%(in_tuple_type)s>* > %(hashname)s;
"""
right_template = """insert(%(hashname)s, %(keyname)s, %(keypos)s);
"""
hashname = self._hashname
keyname = t.name
keypos = self.right_keypos
in_tuple_type = t.getTupleTypename()
# declaration of hash map
hashdeclr = declr_template % locals()
state.addDeclarations([hashdeclr])
# materialization point
code = right_template % locals()
return code
if src.childtag == "left":
left_template = """
for (auto %(right_tuple_name)s : \
lookup(%(hashname)s, %(keyname)s.get(%(keypos)s))) {
auto %(out_tuple_name)s = \
combine<%(out_tuple_type)s> (%(keyname)s, %(right_tuple_name)s);
%(inner_plan_compiled)s
}
"""
hashname = self._hashname
keyname = t.name
keytype = t.getTupleTypename()
keypos = self.left_keypos
right_tuple_name = gensym()
outTuple = CStagedTupleRef(gensym(), self.scheme())
out_tuple_type_def = outTuple.generateDefinition()
out_tuple_type = outTuple.getTupleTypename()
out_tuple_name = outTuple.name
state.addDeclarations([out_tuple_type_def])
inner_plan_compiled = self.parent().consume(outTuple, self, state)
code = left_template % locals()
return code
assert False, "src not equal to left or right"
def consume(self, t, src, state):
if src.childtag == "right":
right_template = self._cgenv.get_template('insert_materialize.cpp')
hashname = self._hashname
keyname = t.name
keyval = self.__aggregate_val__(t, self.rightcols)
self.right_syncname = get_pipeline_task_name(state)
self.rightTupleTypename = t.getTupleTypename()
if self.rightTupleTypeRef is not None:
state.resolveSymbol(self.rightTupleTypeRef,
self.rightTupleTypename)
pipeline_sync = state.getPipelineProperty('global_syncname')
# materialization point
code = right_template.render(locals())
return code
if src.childtag == "left":
left_template = self._cgenv.get_template('lookup.cpp')
# add a dependence on the right pipeline
state.addToPipelinePropertySet('dependences', self.right_syncname)
hashname = self._hashname
keyname = t.name
input_tuple_type = t.getTupleTypename()
keyval = self.__aggregate_val__(t, self.leftcols)
pipeline_sync = state.getPipelineProperty('global_syncname')
right_tuple_name = gensym()
right_tuple_type = self.rightTupleTypename
outTuple = GrappaStagedTupleRef(gensym(), self.scheme())
out_tuple_type_def = outTuple.generateDefinition()
out_tuple_type = outTuple.getTupleTypename()
out_tuple_name = outTuple.name
state.addDeclarations([out_tuple_type_def])
inner_plan_compiled = self.parent().consume(outTuple, self, state)
code = left_template.render(locals())
return code
assert False, "src not equal to left or right"
def produce(self, state):
# Common subexpression elimination
# don't scan the same file twice
resultsym = state.lookupExpr(self)
_LOG.debug("lookup %s(h=%s) => %s", self, self.__hash__(), resultsym)
if not resultsym:
# TODO for now this will break
# whatever relies on self.bound like reusescans
# Scan is the only place where a relation is declared
resultsym = gensym()
name = str(self.relation_key).split(':')[2]
fstemplate, fsbindings = self.__compileme__(resultsym, name)
state.saveExpr(self, resultsym)
stagedTuple = self.new_tuple_ref(resultsym, self.scheme())
state.saveTupleDef(resultsym, stagedTuple)
tuple_type_def = stagedTuple.generateDefinition()
tuple_type = stagedTuple.getTupleTypename()
state.addDeclarations([tuple_type_def])
rel_decl_template = self.__get_relation_decl_template__(name)
if rel_decl_template:
state.addDeclarations([rel_decl_template.render(locals())])
# now that we have the type, format this in;
state.setPipelineProperty('type', 'scan')
state.setPipelineProperty('source', self.__class__)
state.addPipeline(
fstemplate.render(fsbindings, result_type=tuple_type))
# no return value used because parent is a new pipeline
self.parent().consume(resultsym, self, state)
def create_pipeline_synchronization(state):
"""
The pipeline_synchronization will sync tasks
within a single pipeline. Adds this new object to
the compiler state.
"""
global_syncname = gensym()
# true = tracked by gce user metrics
global_sync_decl_template = ct("""
GlobalCompletionEvent %(global_syncname)s(true);
""")
global_sync_decl = global_sync_decl_template % locals()
gce_metric_template = """
GRAPPA_DEFINE_METRIC(CallbackMetric<int64_t>, \
app_%(pipeline_id)s_gce_incomplete, []{
return %(global_syncname)s.incomplete();
});
"""
pipeline_id = state.getCurrentPipelineId()
gce_metric_def = gce_metric_template % locals()
state.addDeclarations([global_sync_decl, gce_metric_def])
state.setPipelineProperty('global_syncname', global_syncname)
return global_syncname
def expression_combine(cls, args, operator="&&"):
codes, decls, inits = cls._extract_code_decl_init(args)
# special case for integer divide. C doesn't have this syntax
# Rely on automatic conversion from float to int
this_decls = []
this_inits = []
if operator == "//":
operator = "/"
# special case for string LIKE, use overloaded mod operator
elif operator == "like":
# NOTE: LIKE probably shouldn't be implemented as
# a "binop" because the input type != output type
assert len(args) == 2, "LIKE only combines 2 arguments"
operator = "%"
# hoist pattern compilation out of the loop processing
# Unchecked precondition: codes[1] is independent of the tuple
varname = gensym()
this_decls.append("std::regex {var};\n".format(var=varname))
this_inits.append(cls.on_all(
"""{var} = compile_like_pattern({str});
""".format(var=varname, str=codes[1])))
# replace the string literal with the regex
codes[1] = varname
opstr = " %s " % operator
conjunc = opstr.join(["(%s)" % c for c in codes])
_LOG.debug("conjunc: %s", conjunc)
return "( %s )" % conjunc, \
decls + this_decls, \
inits + this_inits
def expression_combine(cls, args, operator="&&"):
codes, decls, inits = cls._extract_code_decl_init(args)
# special case for integer divide. C doesn't have this syntax
# Rely on automatic conversion from float to int
this_decls = []
this_inits = []
if operator == "//":
operator = "/"
# special case for string LIKE, use overloaded mod operator
elif operator == "like":
# NOTE: LIKE probably shouldn't be implemented as
# a "binop" because the input type != output type
assert len(args) == 2, "LIKE only combines 2 arguments"
operator = "%"
# hoist pattern compilation out of the loop processing
# Unchecked precondition: codes[1] is independent of the tuple
varname = gensym()
this_decls.append("std::regex {var};\n".format(var=varname))
this_inits.append(cls.on_all(
"""{var} = compile_like_pattern({str});
""".format(var=varname, str=codes[1])))
# replace the string literal with the regex
codes[1] = varname
opstr = " %s " % operator
conjunc = opstr.join(["(%s)" % c for c in codes])
_LOG.debug("conjunc: %s", conjunc)
return "( %s )" % conjunc, \
decls + this_decls, \
inits + this_inits
def produce(self, state):
# declare a single new type for project
# TODO: instead do mark used-columns?
# always does an assignment to new tuple
self.newtuple = self.new_tuple_ref(gensym(), self.scheme())
state.addDeclarations([self.newtuple.generateDefinition()])
self.input.produce(state)
def get_append(out_tuple_type, type1, type1numfields,
type2, type2numfields):
append_func_name = "create_" + gensym()
result_type = out_tuple_type
combine_function_def = _cgenv.get_template(
"materialized_tuple_create_two.cpp").render(locals())
return append_func_name, combine_function_def
def produce(self, state):
# declare a single new type for project
# TODO: instead do mark used-columns?
# always does an assignment to new tuple
self.newtuple = self.new_tuple_ref(gensym(), self.scheme())
state.addDeclarations([self.newtuple.generateDefinition()])
self.input.produce(state)
def get_append(out_tuple_type, type1, type1numfields,
type2, type2numfields):
append_func_name = "create_" + gensym()
result_type = out_tuple_type
combine_function_def = _cgenv.get_template(
"materialized_tuple_create_two.cpp").render(locals())
return append_func_name, combine_function_def
def produce(self, state):
self.symBase = self.__genBaseName__()
if not isinstance(self.condition, expression.EQ):
msg = "The C compiler can only handle equi-join conditions\
of a single attribute: %s" % self.condition
raise ValueError(msg)
init_template = ct("""%(hashname)s.init_global_DHT( &%(hashname)s, \
cores()*16*1024 );
""")
declr_template = ct("""typedef DoubleDHT<int64_t, \
%(left_in_tuple_type)s, \
%(right_in_tuple_type)s,
std_hash> \
DHT_%(left_in_tuple_type)s_%(right_in_tuple_type)s;
DHT_%(left_in_tuple_type)s_%(right_in_tuple_type)s %(hashname)s;
""")
my_sch = self.scheme()
left_sch = self.left.scheme()
# declaration of hash map
self._hashname = self.__getHashName__()
hashname = self._hashname
self.leftTypeRef = state.createUnresolvedSymbol()
left_in_tuple_type = self.leftTypeRef.getPlaceholder()
self.rightTypeRef = state.createUnresolvedSymbol()
right_in_tuple_type = self.rightTypeRef.getPlaceholder()
hashdeclr = declr_template % locals()
state.addDeclarationsUnresolved([hashdeclr])
self.outTuple = GrappaStagedTupleRef(gensym(), my_sch)
out_tuple_type_def = self.outTuple.generateDefinition()
state.addDeclarations([out_tuple_type_def])
# find the attribute that corresponds to the right child
self.rightCondIsRightAttr = \
self.condition.right.position >= len(left_sch)
self.leftCondIsRightAttr = \
self.condition.left.position >= len(left_sch)
assert self.rightCondIsRightAttr ^ self.leftCondIsRightAttr
self.right.childtag = "right"
state.addInitializers([init_template % locals()])
self.right.produce(state)
self.left.childtag = "left"
self.left.produce(state)
def produce(self, state):
# Common subexpression elimination
# don't scan the same file twice
resultsym = state.lookupExpr(self)
_LOG.debug("lookup %s(h=%s) => %s", self, self.__hash__(),
resultsym)
if not resultsym:
# TODO for now this will break
# whatever relies on self.bound like reusescans
# Scan is the only place where a relation is declared
resultsym = gensym()
name = str(self.relation_key).split(':')[2]
fstemplate, fsbindings = self.__compileme__(resultsym, name)
state.saveExpr(self, resultsym)
stagedTuple = self.new_tuple_ref_for_filescan(
resultsym,
self.scheme())
state.saveTupleDef(resultsym, stagedTuple)
tuple_type_def = stagedTuple.generateDefinition()
tuple_type = stagedTuple.getTupleTypename()
state.addDeclarations([tuple_type_def])
colnames = self.scheme().get_names()
rel_decl_template = self.__get_relation_decl_template__(name)
if rel_decl_template:
state.addDeclarations([rel_decl_template.render(locals())])
rel_aux_decl_template = self._get_input_aux_decls_template()
if rel_aux_decl_template:
state.addDeclarations([rel_aux_decl_template.render(locals())])
# now that we have the type, format this in;
state.setPipelineProperty('type', 'scan')
state.setPipelineProperty('source', self.__class__)
state.addPipeline(
fstemplate.render(fsbindings, result_type=tuple_type))
# no return value used because parent is a new pipeline
self.parent().consume(resultsym, self, state)
def createTupleTypeConversion(lang, state, input_tuple, result_tuple):
# add declaration for function to convert from one type to the other
type1 = input_tuple.getTupleTypename()
type1numfields = len(input_tuple.scheme)
convert_func_name = "create_" + gensym()
result_type = result_tuple.getTupleTypename()
result_name = result_tuple.name
input_tuple_name = input_tuple.name
convert_func = lang._cgenv.get_template(
'materialized_tuple_create_one.cpp').render(locals())
state.addDeclarations([convert_func])
return lang._cgenv.get_template('tuple_type_convert.cpp').render(
result_type=result_type,
result_name=result_name,
convert_func_name=convert_func_name,
input_tuple_name=input_tuple_name
)
def consume(self, t, src, state):
union_template = _cgenv.get_template('union.cpp')
unified_tuple_typename = self.unifiedTupleType.getTupleTypename()
unified_tuple_name = self.unifiedTupleType.name
src_tuple_name = t.name
# add declaration for function to convert from one type to the other
type1 = t.getTupleTypename()
type1numfields = len(t.scheme)
convert_func_name = "create_" + gensym()
result_type = unified_tuple_typename
convert_func = _cgenv.get_template(
'materialized_tuple_create_one.cpp').render(locals())
state.addDeclarations([convert_func])
inner_plan_compiled = \
self.parent().consume(self.unifiedTupleType, self, state)
return union_template.render(locals())
def produce(self, state):
self.symBase = self.__genBaseName__()
init_template = self._cgenv.get_template('hash_init.cpp')
declr_template = self._cgenv.get_template('hash_declaration.cpp')
my_sch = self.scheme()
left_sch = self.left.scheme()
right_sch = self.right.scheme()
self.leftcols, self.rightcols = \
algebra.convertcondition(self.condition,
len(left_sch),
left_sch + right_sch)
# declaration of hash map
self._hashname = self.__getHashName__()
keytype = self.__aggregate_type__(my_sch, self.rightcols)
hashname = self._hashname
self.leftTypeRef = state.createUnresolvedSymbol()
left_in_tuple_type = self.leftTypeRef.getPlaceholder()
self.rightTypeRef = state.createUnresolvedSymbol()
right_in_tuple_type = self.rightTypeRef.getPlaceholder()
hashdeclr = declr_template.render(locals())
state.addDeclarationsUnresolved([hashdeclr])
self.outTuple = GrappaStagedTupleRef(gensym(), my_sch)
out_tuple_type_def = self.outTuple.generateDefinition()
state.addDeclarations([out_tuple_type_def])
self.right.childtag = "right"
state.addInitializers([init_template.render(locals())])
self.right.produce(state)
self.left.childtag = "left"
self.left.produce(state)
def create_pipeline_synchronization(state):
"""
The pipeline_synchronization will sync tasks
within a single pipeline. Adds this new object to
the compiler state.
"""
global_syncname = gensym()
# true = tracked by gce user metrics
global_sync_decl = GrappaLanguage.cgenv().get_template(
'sync_declaration.cpp').render(locals())
gce_metric_template = GrappaLanguage.cgenv().get_template(
'gce_app_metric.cpp')
pipeline_id = state.getCurrentPipelineId()
gce_metric_def = gce_metric_template.render(locals())
state.addDeclarations([global_sync_decl, gce_metric_def])
state.setPipelineProperty('global_syncname', global_syncname)
return global_syncname
def produce(self, state):
assert len(self.grouping_list) <= 2, \
"""%s does not currently support \
"groupings of more than 2 attributes"""\
% self.__class__.__name__
assert len(self.aggregate_list) == 1, \
"%s currently only supports aggregates of 1 attribute"\
% self.__class__.__name__
for agg_term in self.aggregate_list:
assert isinstance(agg_term,
expression.BuiltinAggregateExpression), \
"""%s only supports simple aggregate expressions.
A rule should create Apply[GroupBy]""" \
% self.__class__.__name__
self.useKey = len(self.grouping_list) > 0
_LOG.debug("groupby uses keys? %s" % self.useKey)
declr_template = None
if self.useKey:
if len(self.grouping_list) == 1:
declr_template = ct("""typedef DHT_symmetric<int64_t, \
int64_t, std_hash> \
DHT_int64;
""")
elif len(self.grouping_list) == 2:
declr_template = ct("""typedef DHT_symmetric<\
std::pair<int64_t,int64_t>, \
int64_t, pair_hash> \
DHT_pair_int64;
""")
self._hashname = self.__genHashName__()
_LOG.debug("generate hashname %s for %s", self._hashname, self)
hashname = self._hashname
if declr_template is not None:
hashdeclr = declr_template % locals()
state.addDeclarationsUnresolved([hashdeclr])
if self.useKey:
if len(self.grouping_list) == 1:
init_template = ct("""auto %(hashname)s = \
DHT_int64::create_DHT_symmetric( );""")
elif len(self.grouping_list) == 2:
init_template = ct("""auto %(hashname)s = \
DHT_pair_int64::create_DHT_symmetric( );""")
else:
init_template = ct("""auto %(hashname)s = counter::create();
""")
state.addInitializers([init_template % locals()])
self.input.produce(state)
# now that everything is aggregated, produce the tuples
assert len(self.column_list()) == 1 \
or isinstance(self.column_list()[0],
expression.AttributeRef), \
"""assumes first column is the key and second is aggregate result
column_list: %s""" % self.column_list()
if self.useKey:
mapping_var_name = gensym()
if len(self.grouping_list) == 1:
produce_template = ct("""%(hashname)s->\
forall_entries<&%(pipeline_sync)s>\
([=](std::pair<const int64_t,int64_t>& %(mapping_var_name)s) {
%(output_tuple_type)s %(output_tuple_name)s(\
{%(mapping_var_name)s.first, %(mapping_var_name)s.second});
%(inner_code)s
});
""")
elif len(self.grouping_list) == 2:
produce_template = ct("""%(hashname)s->\
forall_entries<&%(pipeline_sync)s>\
([=](std::pair<const std::pair<int64_t,int64_t>,int64_t>& \
%(mapping_var_name)s) {
%(output_tuple_type)s %(output_tuple_name)s(\
{%(mapping_var_name)s.first.first,\
%(mapping_var_name)s.first.second,\
%(mapping_var_name)s.second});
%(inner_code)s
});
""")
else:
op = self.aggregate_list[0].__class__.__name__
# translations for Grappa::reduce predefined ops
coll_op = {'COUNT': 'COLL_ADD',
'SUM': 'COLL_ADD',
'MAX': 'COLL_MAX',
'MIN': 'COLL_MIN'}[op]
produce_template = ct("""auto %(output_tuple_name)s_tmp = \
reduce<int64_t, \
counter, \
%(coll_op)s, \
&get_count>\
(%(hashname)s);
%(output_tuple_type)s %(output_tuple_name)s;
%(output_tuple_name)s.set(0, %(output_tuple_name)s_tmp);
%(inner_code)s
""")
pipeline_sync = create_pipeline_synchronization(state)
get_pipeline_task_name(state)
# add a dependence on the input aggregation pipeline
state.addToPipelinePropertySet('dependences', self.input_syncname)
output_tuple = GrappaStagedTupleRef(gensym(), self.scheme())
output_tuple_name = output_tuple.name
output_tuple_type = output_tuple.getTupleTypename()
state.addDeclarations([output_tuple.generateDefinition()])
inner_code = self.parent().consume(output_tuple, self, state)
code = produce_template % locals()
state.setPipelineProperty("type", "in_memory")
state.addPipeline(code)
def produce(self, state):
self.unifiedTupleType = self.new_tuple_ref(gensym(), self.scheme())
state.addDeclarations([self.unifiedTupleType.generateDefinition()])
self.right.produce(state)
self.left.produce(state)
def consume(self, t, src, state):
if src.childtag == "right":
right_template = ct("""
%(hashname)s.insert_async<&%(pipeline_sync)s>(\
%(keyname)s.get(%(keypos)s), %(keyname)s);
""")
hashname = self._hashname
keyname = t.name
keypos = self.right_keypos
self.right_syncname = get_pipeline_task_name(state)
self.rightTupleTypename = t.getTupleTypename()
if self.rightTupleTypeRef is not None:
state.resolveSymbol(self.rightTupleTypeRef,
self.rightTupleTypename)
pipeline_sync = state.getPipelineProperty('global_syncname')
# materialization point
code = right_template % locals()
return code
if src.childtag == "left":
left_template = ct("""
%(hashname)s.lookup_iter<&%(pipeline_sync)s>( \
%(keyname)s.get(%(keypos)s), \
[=](%(right_tuple_type)s& %(right_tuple_name)s) {
join_coarse_result_count++;
%(out_tuple_type)s %(out_tuple_name)s = \
combine<%(out_tuple_type)s, \
%(keytype)s, \
%(right_tuple_type)s> \
(%(keyname)s, %(right_tuple_name)s);
%(inner_plan_compiled)s
});
""")
# add a dependence on the right pipeline
state.addToPipelinePropertySet('dependences', self.right_syncname)
hashname = self._hashname
keyname = t.name
keytype = t.getTupleTypename()
pipeline_sync = state.getPipelineProperty('global_syncname')
keypos = self.left_keypos
right_tuple_name = gensym()
right_tuple_type = self.rightTupleTypename
outTuple = GrappaStagedTupleRef(gensym(), self.scheme())
out_tuple_type_def = outTuple.generateDefinition()
out_tuple_type = outTuple.getTupleTypename()
out_tuple_name = outTuple.name
state.addDeclarations([out_tuple_type_def])
inner_plan_compiled = self.parent().consume(outTuple, self, state)
code = left_template % locals()
return code
assert False, "src not equal to left or right"
def produce(self, state):
assert len(self.grouping_list) <= 2, \
"%s does not currently support groupings of \
more than 2 attributes" % self.__class__.__name__
assert len(self.aggregate_list) == 1, \
"""%s currently only supports aggregates of 1 attribute
(aggregate_list=%s)""" \
% (self.__class__.__name__, self.aggregate_list)
for agg_term in self.aggregate_list:
assert isinstance(agg_term,
expression.BuiltinAggregateExpression), \
"""%s only supports simple aggregate expressions.
A rule should create Apply[GroupBy]""" \
% self.__class__.__name__
inp_sch = self.input.scheme()
self.useMap = len(self.grouping_list) > 0
if self.useMap:
if len(self.grouping_list) == 1:
declr_template = self._cgenv.get_template(
'1key_declaration.cpp')
keytype = self.language().typename(
self.grouping_list[0].typeof(
inp_sch,
None))
elif len(self.grouping_list) == 2:
declr_template = self._cgenv.get_template(
'2key_declaration.cpp')
keytypes = ','.join(
[self.language().typename(g.typeof(inp_sch, None))
for g in self.grouping_list])
else:
initial_value = self.__get_initial_value__(
0,
cached_inp_sch=inp_sch)
declr_template = self._cgenv.get_template('0key_declaration.cpp')
valtype = self.language().typename(
self.aggregate_list[0].typeof(
inp_sch,
None))
self.hashname = CGroupBy.__genHashName__()
hashname = self.hashname
hash_declr = declr_template.render(locals())
state.addDeclarations([hash_declr])
my_sch = self.scheme()
_LOG.debug("aggregates: %s", self.aggregate_list)
_LOG.debug("columns: %s", self.column_list())
_LOG.debug("groupings: %s", self.grouping_list)
_LOG.debug("groupby scheme: %s", my_sch)
_LOG.debug("groupby scheme[0] type: %s", type(my_sch[0]))
self.input.produce(state)
# now that everything is aggregated, produce the tuples
assert (not self.useMap) \
or isinstance(self.column_list()[0],
expression.AttributeRef), \
"assumes first column is the key and " \
"second is aggregate result: %s" % (self.column_list()[0])
if self.useMap:
if len(self.grouping_list) == 1:
produce_template = self._cgenv.get_template('1key_scan.cpp')
elif len(self.grouping_list) == 2:
produce_template = self._cgenv.get_template('2key_scan.cpp')
else:
produce_template = self._cgenv.get_template('0key_scan.cpp')
output_tuple = CStagedTupleRef(gensym(), my_sch)
output_tuple_name = output_tuple.name
output_tuple_type = output_tuple.getTupleTypename()
state.addDeclarations([output_tuple.generateDefinition()])
inner_code = self.parent().consume(output_tuple, self, state)
code = produce_template.render(locals())
state.setPipelineProperty("type", "in_memory")
state.addPipeline(code)
def consume(self, t, src, state):
access_template = ct("""
%(hashname)s.insert_lookup_iter_%(side)s<&%(global_syncname)s>(\
%(keyname)s.get(%(keypos)s), %(keyname)s, \
[=](%(other_tuple_type)s %(valname)s) {
join_coarse_result_count++;
%(out_tuple_type)s %(out_tuple_name)s = \
combine<%(out_tuple_type)s, \
%(left_type)s, \
%(right_type)s> (%(left_name)s, \
%(right_name)s);
%(inner_plan_compiled)s
});
""")
hashname = self._hashname
keyname = t.name
side = src.childtag
outTuple = self.outTuple
out_tuple_type = self.outTuple.getTupleTypename()
out_tuple_name = self.outTuple.name
global_syncname = state.getPipelineProperty('global_syncname')
if src.childtag == "right":
left_sch = self.left.scheme()
# save for later
self.right_in_tuple_type = t.getTupleTypename()
state.resolveSymbol(self.rightTypeRef, self.right_in_tuple_type)
if self.rightCondIsRightAttr:
keypos = self.condition.right.position \
- len(left_sch)
else:
keypos = self.condition.left.position \
- len(left_sch)
inner_plan_compiled = self.parent().consume(outTuple, self, state)
other_tuple_type = self.leftTypeRef.getPlaceholder()
left_type = other_tuple_type
right_type = self.right_in_tuple_type
left_name = gensym()
right_name = keyname
self.right_name = right_name
valname = left_name
code = access_template % locals()
return code
if src.childtag == "left":
right_in_tuple_type = self.right_in_tuple_type
left_in_tuple_type = t.getTupleTypename()
state.resolveSymbol(self.leftTypeRef, left_in_tuple_type)
if self.rightCondIsRightAttr:
keypos = self.condition.left.position
else:
keypos = self.condition.right.position
inner_plan_compiled = self.parent().consume(outTuple, self, state)
left_type = left_in_tuple_type
right_type = self.right_in_tuple_type
other_tuple_type = self.right_in_tuple_type
left_name = keyname
right_name = gensym()
valname = right_name
code = access_template % locals()
return code
assert False, "src not equal to left or right"
def produce(self, state):
assert len(self.grouping_list) <= 2, \
"%s does not currently support groupings of \
more than 2 attributes" % self.__class__.__name__
assert len(self.aggregate_list) == 1, \
"""%s currently only supports aggregates of 1 attribute
(aggregate_list=%s)""" \
% (self.__class__.__name__, self.aggregate_list)
for agg_term in self.aggregate_list:
assert isinstance(agg_term,
expression.BuiltinAggregateExpression), \
"""%s only supports simple aggregate expressions.
A rule should create Apply[GroupBy]""" \
% self.__class__.__name__
self.useMap = len(self.grouping_list) > 0
if self.useMap:
if len(self.grouping_list) == 1:
declr_template = """std::unordered_map<int64_t, int64_t> \
%(hashname)s;
"""
elif len(self.grouping_list) == 2:
declr_template = """std::unordered_map<\
std::pair<int64_t, int64_t>, int64_t, pairhash> \
%(hashname)s;
"""
else:
declr_template = """int64_t %(hashname)s;
"""
self.hashname = self.__genHashName__()
hashname = self.hashname
hash_declr = declr_template % locals()
state.addDeclarations([hash_declr])
my_sch = self.scheme()
_LOG.debug("aggregates: %s", self.aggregate_list)
_LOG.debug("columns: %s", self.column_list())
_LOG.debug("groupings: %s", self.grouping_list)
_LOG.debug("groupby scheme: %s", my_sch)
_LOG.debug("groupby scheme[0] type: %s", type(my_sch[0]))
self.input.produce(state)
# now that everything is aggregated, produce the tuples
assert (not self.useMap) \
or isinstance(self.column_list()[0],
expression.AttributeRef), \
"assumes first column is the key and " \
"second is aggregate result: %s" % (self.column_list()[0])
if self.useMap:
if len(self.grouping_list) == 1:
produce_template = """for (auto it=%(hashname)s.begin(); \
it!=%(hashname)s.end(); it++) {
%(output_tuple_type)s %(output_tuple_name)s(\
{it->first, it->second});
%(inner_code)s
}
"""
elif len(self.grouping_list) == 2:
produce_template = """for (auto it=%(hashname)s.begin(); \
it!=%(hashname)s.end(); it++) {
%(output_tuple_type)s %(output_tuple_name)s(\
{it->first.first, it->first.second, it->second});
%(inner_code)s
}
"""
else:
produce_template = """{
%(output_tuple_type)s %(output_tuple_name)s({ %(hashname)s });
%(inner_code)s
}
"""
output_tuple = CStagedTupleRef(gensym(), my_sch)
output_tuple_name = output_tuple.name
output_tuple_type = output_tuple.getTupleTypename()
state.addDeclarations([output_tuple.generateDefinition()])
inner_code = self.parent().consume(output_tuple, self, state)
code = produce_template % locals()
state.setPipelineProperty("type", "in_memory")
state.addPipeline(code)
def produce(self, state):
self._agg_mode = None
if len(self.aggregate_list) == 1 \
and isinstance(self.aggregate_list[0],
expression.BuiltinAggregateExpression):
self._agg_mode = self._ONE_BUILT_IN
elif all([isinstance(a, expression.UdaAggregateExpression)
for a in self.aggregate_list]):
self._agg_mode = self._MULTI_UDA
assert self._agg_mode is not None, \
"unsupported aggregates {0}".format(self.aggregate_list)
_LOG.debug("%s _agg_mode was set to %s", self, self._agg_mode)
self.useKey = len(self.grouping_list) > 0
_LOG.debug("groupby uses keys? %s" % self.useKey)
inp_sch = self.input.scheme()
if self._agg_mode == self._ONE_BUILT_IN:
state_type = self.language().typename(
self.aggregate_list[0].input.typeof(inp_sch, None))
op = self.aggregate_list[0].__class__.__name__
self.update_func = "Aggregates::{op}<{type}, {type}>".format(
op=op, type=state_type)
elif self._agg_mode == self._MULTI_UDA:
# for now just name the aggregate after the first state variable
self.func_name = self.updaters[0][0]
self.state_tuple = GrappaStagedTupleRef(gensym(),
self.state_scheme)
state.addDeclarations([self.state_tuple.generateDefinition()])
state_type = self.state_tuple.getTupleTypename()
self.update_func = "{name}_update".format(name=self.func_name)
update_func = self.update_func
if self.useKey:
numkeys = len(self.grouping_list)
keytype = "std::tuple<{types}>".format(
types=','.join([self.language().typename(
g.typeof(inp_sch, None)) for g in self.grouping_list]))
self._hashname = self.__genHashName__()
_LOG.debug("generate hashname %s for %s", self._hashname, self)
hashname = self._hashname
if self.useKey:
init_template = self._cgenv.get_template('withkey_init.cpp')
valtype = state_type
else:
if self._agg_mode == self._ONE_BUILT_IN:
initial_value = \
self.__get_initial_value__(0, cached_inp_sch=inp_sch)
no_key_state_initializer = \
"counter<{state_type}>::create({valinit})".format(
state_type=state_type, valinit=initial_value)
elif self._agg_mode == self._MULTI_UDA:
no_key_state_initializer = \
"symmetric_global_alloc<{state_tuple_type}>()".format(
state_tuple_type=self.state_tuple.getTupleTypename())
init_template = self._cgenv.get_template('withoutkey_init.cpp')
initializer = no_key_state_initializer
state.addInitializers([init_template.render(locals())])
self.input.produce(state)
# now that everything is aggregated, produce the tuples
# assert len(self.column_list()) == 1 \
# or isinstance(self.column_list()[0],
# expression.AttributeRef), \
# """assumes first column is the key and second is aggregate result
# column_list: %s""" % self.column_list()
if self.useKey:
mapping_var_name = gensym()
if self._agg_mode == self._ONE_BUILT_IN:
emit_type = self.language().typename(
self.aggregate_list[0].input.typeof(
self.input.scheme(), None))
elif self._agg_mode == self._MULTI_UDA:
emit_type = self.state_tuple.getTupleTypename()
if self._agg_mode == self._ONE_BUILT_IN:
# need to force type in make_tuple
produce_template = self._cgenv.get_template(
'one_built_in_scan.cpp')
elif self._agg_mode == self._MULTI_UDA:
# pass in attribute values individually
produce_template = self._cgenv.get_template(
'multi_uda_scan.cpp')
else:
if self._agg_mode == self._ONE_BUILT_IN:
produce_template = self._cgenv.get_template(
'one_built_in_0key_output.cpp')
elif self._agg_mode == self._MULTI_UDA:
produce_template = self._cgenv.get_template(
'multi_uda_0key_output.cpp')
pipeline_sync = create_pipeline_synchronization(state)
get_pipeline_task_name(state)
# add a dependence on the input aggregation pipeline
state.addToPipelinePropertySet('dependences', self.input_syncname)
output_tuple = GrappaStagedTupleRef(gensym(), self.scheme())
output_tuple_name = output_tuple.name
output_tuple_type = output_tuple.getTupleTypename()
output_tuple_set_func = output_tuple.set_func_code(0)
state.addDeclarations([output_tuple.generateDefinition()])
inner_code = self.parent().consume(output_tuple, self, state)
code = produce_template.render(locals())
state.setPipelineProperty("type", "in_memory")
state.addPipeline(code)
def produce(self, state):
assert len(self.grouping_list) <= 2, \
"%s does not currently support groupings of \
more than 2 attributes" % self.__class__.__name__
assert len(self.aggregate_list) == 1, \
"""%s currently only supports aggregates of 1 attribute
(aggregate_list=%s)""" \
% (self.__class__.__name__, self.aggregate_list)
for agg_term in self.aggregate_list:
assert isinstance(agg_term,
expression.BuiltinAggregateExpression), \
"""%s only supports simple aggregate expressions.
A rule should create Apply[GroupBy]""" \
% self.__class__.__name__
inp_sch = self.input.scheme()
self.useMap = len(self.grouping_list) > 0
if self.useMap:
if len(self.grouping_list) == 1:
declr_template = self._cgenv.get_template(
'1key_declaration.cpp')
keytype = self.language().typename(
self.grouping_list[0].typeof(inp_sch, None))
elif len(self.grouping_list) == 2:
declr_template = self._cgenv.get_template(
'2key_declaration.cpp')
keytypes = ','.join([
self.language().typename(g.typeof(inp_sch, None))
for g in self.grouping_list
])
else:
initial_value = self.__get_initial_value__(0,
cached_inp_sch=inp_sch)
declr_template = self._cgenv.get_template('0key_declaration.cpp')
valtype = self.language().typename(self.aggregate_list[0].typeof(
inp_sch, None))
self.hashname = CGroupBy.__genHashName__()
hashname = self.hashname
hash_declr = declr_template.render(locals())
state.addDeclarations([hash_declr])
my_sch = self.scheme()
_LOG.debug("aggregates: %s", self.aggregate_list)
_LOG.debug("columns: %s", self.column_list())
_LOG.debug("groupings: %s", self.grouping_list)
_LOG.debug("groupby scheme: %s", my_sch)
_LOG.debug("groupby scheme[0] type: %s", type(my_sch[0]))
self.input.produce(state)
# now that everything is aggregated, produce the tuples
assert (not self.useMap) \
or isinstance(self.column_list()[0],
expression.AttributeRef), \
"assumes first column is the key and " \
"second is aggregate result: %s" % (self.column_list()[0])
if self.useMap:
if len(self.grouping_list) == 1:
produce_template = self._cgenv.get_template('1key_scan.cpp')
elif len(self.grouping_list) == 2:
produce_template = self._cgenv.get_template('2key_scan.cpp')
else:
produce_template = self._cgenv.get_template('0key_scan.cpp')
output_tuple = CStagedTupleRef(gensym(), my_sch)
output_tuple_name = output_tuple.name
output_tuple_type = output_tuple.getTupleTypename()
state.addDeclarations([output_tuple.generateDefinition()])
inner_code = self.parent().consume(output_tuple, self, state)
code = produce_template.render(locals())
state.setPipelineProperty("type", "in_memory")
state.addPipeline(code)
def produce(self, state):
self.unifiedTupleType = self.new_tuple_ref(gensym(), self.scheme())
state.addDeclarations([self.unifiedTupleType.generateDefinition()])
for arg in self.args:
arg.produce(state)
def consume(self, t, src, state):
if src.childtag == "right":
my_sch = self.scheme()
declr_template = self._cgenv.get_template("hash_declaration.cpp")
right_template = self._cgenv.get_template("insert_materialize.cpp")
hashname = self._hashname
keypos = self.right_keypos
keyval = t.get_code(self.right_keypos)
if self.rightCondIsRightAttr:
keytype = self.language().typename(
self.condition.right.typeof(my_sch, None))
else:
keytype = self.language().typename(
self.condition.left.typeof(my_sch, None))
in_tuple_type = t.getTupleTypename()
in_tuple_name = t.name
self.right_type = in_tuple_type
state.saveExpr((self.right, self.right_keypos),
(self._hashname, self.right_type))
# declaration of hash map
hashdeclr = declr_template.render(locals())
state.addDeclarations([hashdeclr])
# materialization point
code = right_template.render(locals())
return code
if src.childtag == "left":
left_template = self._cgenv.get_template("lookup.cpp")
hashname = self._hashname
keyname = t.name
keytype = t.getTupleTypename()
keypos = self.left_keypos
keyval = t.get_code(keypos)
right_tuple_name = gensym()
outTuple = CStagedTupleRef(gensym(), self.scheme())
out_tuple_type_def = outTuple.generateDefinition()
out_tuple_type = outTuple.getTupleTypename()
out_tuple_name = outTuple.name
type1 = keytype
type1numfields = len(t.scheme)
type2 = self.right_type
type2numfields = len(self.right.scheme())
append_func_name, combine_function_def = \
CStagedTupleRef.get_append(
out_tuple_type,
type1, type1numfields,
type2, type2numfields)
state.addDeclarations([out_tuple_type_def, combine_function_def])
inner_plan_compiled = self.parent().consume(outTuple, self, state)
code = left_template.render(locals())
return code
assert False, "src not equal to left or right"
def produce(self, state):
left_sch = self.left.scheme()
self.syncnames = []
self.symBase = self.__genBaseName__()
self.right.childtag = "right"
self.rightTupleTypeRef = None # may remain None if CSE succeeds
self.leftTupleTypeRef = None # may remain None if CSE succeeds
# find the attribute that corresponds to the right child
self.rightCondIsRightAttr = \
self.condition.right.position >= len(left_sch)
self.leftCondIsRightAttr = \
self.condition.left.position >= len(left_sch)
assert self.rightCondIsRightAttr ^ self.leftCondIsRightAttr
# find right key position
if self.rightCondIsRightAttr:
self.right_keypos = self.condition.right.position \
- len(left_sch)
else:
self.right_keypos = self.condition.left.position \
- len(left_sch)
# find left key position
if self.rightCondIsRightAttr:
self.left_keypos = self.condition.left.position
else:
self.left_keypos = self.condition.right.position
# define output tuple
outTuple = GrappaStagedTupleRef(gensym(), self.scheme())
out_tuple_type_def = outTuple.generateDefinition()
out_tuple_type = outTuple.getTupleTypename()
out_tuple_name = outTuple.name
# common index is defined by same right side and same key
# TODO: probably want also left side
hashtableInfo = state.lookupExpr((self.right, self.right_keypos))
if not hashtableInfo:
# if right child never bound then store hashtable symbol and
# call right child produce
self._hashname = self.__getHashName__()
_LOG.debug("generate hashname %s for %s", self._hashname, self)
hashname = self._hashname
# declaration of hash map
self.rightTupleTypeRef = state.createUnresolvedSymbol()
self.leftTupleTypeRef = state.createUnresolvedSymbol()
self.outTupleTypeRef = state.createUnresolvedSymbol()
right_type = self.rightTupleTypeRef.getPlaceholder()
left_type = self.leftTupleTypeRef.getPlaceholder()
# TODO: really want this addInitializers to be addPreCode
# TODO: *for all pipelines that use this hashname*
init_template = self._cgenv.get_template('hash_init.cpp')
state.addInitializers([init_template.render(locals())])
self.right.produce(state)
self.left.childtag = "left"
self.left.produce(state)
state.saveExpr((self.right, self.right_keypos),
(self._hashname, right_type, left_type,
self.right_syncname, self.left_syncname))
else:
# if found a common subexpression on right child then
# use the same hashtable
self._hashname, right_type, left_type,\
self.right_syncname, self.left_syncname = hashtableInfo
_LOG.debug("reuse hash %s for %s", self._hashname, self)
# now that Relation is produced, produce its contents by iterating over
# the join result
iterate_template = self._cgenv.get_template('result_scan.cpp')
hashname = self._hashname
state.addDeclarations([out_tuple_type_def])
pipeline_sync = create_pipeline_synchronization(state)
get_pipeline_task_name(state)
# add dependences on left and right inputs
state.addToPipelinePropertySet('dependences', self.right_syncname)
state.addToPipelinePropertySet('dependences', self.left_syncname)
# reduce is a single self contained pipeline.
# future hashjoin implementations may pipeline out of it
# by passing a continuation to reduceExecute
reduce_template = self._cgenv.get_template('reduce.cpp')
state.addPreCode(reduce_template.render(locals()))
delete_template = self._cgenv.get_template('delete.cpp')
state.addPostCode(delete_template.render(locals()))
inner_code_compiled = self.parent().consume(outTuple, self, state)
code = iterate_template % locals()
state.setPipelineProperty('type', 'in_memory')
state.setPipelineProperty('source', self.__class__)
state.addPipeline(code)
