def example_psyir_nary():
'''Utility function that creates a PSyIR tree containing a nary MIN
intrinsic operator and returns the operator.
:returns: PSyIR MIN operator instance.
:rtype: :py:class:`psyclone.psyGen.NaryOperation`
'''
symbol_table = SymbolTable()
arg1 = symbol_table.new_symbol("arg",
symbol_type=DataSymbol,
datatype=REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
arg2 = symbol_table.new_symbol("arg",
symbol_type=DataSymbol,
datatype=REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
arg3 = symbol_table.new_symbol("arg",
symbol_type=DataSymbol,
datatype=REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
arg4 = symbol_table.new_symbol(symbol_type=DataSymbol, datatype=REAL_TYPE)
symbol_table.specify_argument_list([arg1, arg2, arg3])
var1 = Reference(arg1)
var2 = Reference(arg2)
var3 = Reference(arg3)
var4 = Reference(arg4)
oper = NaryOperation.Operator.MIN
operation = NaryOperation.create(oper, [var1, var2, var3])
assign = Assignment.create(var4, operation)
_ = KernelSchedule.create("min_example", symbol_table, [assign])
return operation
def test_argumentinterface_access_values(access):
'''Check that all the ArgumentInterface access values are supported.
'''
argument_interface = ArgumentInterface()
argument_interface.access = access
assert argument_interface.access == access
def example_psyir(create_expression):
'''Utility function that creates a PSyIR tree containing a SIGN
intrinsic operator and returns the operator.
:param function create_expresssion: function used to create the \
content of the first argument of the SIGN operator.
:returns: PSyIR SIGN operator instance.
:rtype: :py:class:`psyclone.psyGen.BinaryOperation`
'''
symbol_table = SymbolTable()
arg1 = symbol_table.new_symbol("arg",
symbol_type=DataSymbol,
datatype=REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
arg2 = symbol_table.new_symbol("arg",
symbol_type=DataSymbol,
datatype=REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
arg3 = symbol_table.new_symbol(symbol_type=DataSymbol, datatype=REAL_TYPE)
symbol_table.specify_argument_list([arg1, arg2])
var1 = Reference(arg1)
var2 = Reference(arg2)
var3 = Reference(arg3)
oper = BinaryOperation.Operator.SIGN
operation = BinaryOperation.create(oper, create_expression(var1), var2)
assign = Assignment.create(var3, operation)
_ = KernelSchedule.create("sign_example", symbol_table, [assign])
return operation
def test_symbol_interface_setter():
'''Test that the Symbol interface setter behaves as expected,
including raising an exception if the input is of the wrong
type. Also use this to test the is_local, is_global and
is_argument and is_unresolved properties.
'''
symbol = Symbol('sym1')
assert symbol.is_local
assert not symbol.is_global
assert not symbol.is_argument
assert not symbol.is_unresolved
symbol.interface = GlobalInterface(ContainerSymbol("my_mod"))
assert not symbol.is_local
assert symbol.is_global
assert not symbol.is_argument
assert not symbol.is_unresolved
symbol.interface = ArgumentInterface()
assert not symbol.is_local
assert not symbol.is_global
assert symbol.is_argument
assert not symbol.is_unresolved
symbol.interface = UnresolvedInterface()
assert not symbol.is_local
assert not symbol.is_global
assert not symbol.is_argument
assert symbol.is_unresolved
with pytest.raises(TypeError) as info:
symbol.interface = "hello"
assert ("The interface to a Symbol must be a SymbolInterface but got "
"'str'" in str(info.value))
def test_oclw_gen_declaration():
'''Check the OpenCLWriter class gen_declaration method produces
the expected declarations.
'''
oclwriter = OpenCLWriter()
# Basic entry - Scalar are passed by value and don't have additional
# qualifiers.
symbol = DataSymbol("dummy1", INTEGER_TYPE)
result = oclwriter.gen_declaration(symbol)
assert result == "int dummy1"
# Array argument has a memory qualifier (only __global for now)
array_type = ArrayType(INTEGER_TYPE, [2, ArrayType.Extent.ATTRIBUTE, 2])
symbol = DataSymbol("dummy2", array_type)
result = oclwriter.gen_declaration(symbol)
assert result == "__global int * restrict dummy2"
# Array with unknown intent
array_type = ArrayType(INTEGER_TYPE, [2, ArrayType.Extent.ATTRIBUTE, 2])
symbol = DataSymbol("dummy2",
array_type,
interface=ArgumentInterface(
ArgumentInterface.Access.UNKNOWN))
result = oclwriter.gen_declaration(symbol)
assert result == "__global int * restrict dummy2"
def test_datasymbol_copy_properties():
'''Test that the DataSymbol copy_properties method works as expected.'''
array_type = ArrayType(REAL_SINGLE_TYPE, [1, 2])
symbol = DataSymbol("myname",
array_type,
constant_value=None,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
# Check an exception is raised if an incorrect argument is passed in
with pytest.raises(TypeError) as excinfo:
symbol.copy_properties(None)
assert ("Argument should be of type 'DataSymbol' but found 'NoneType'."
"") in str(excinfo.value)
new_symbol = DataSymbol("other_name",
INTEGER_SINGLE_TYPE,
constant_value=7)
symbol.copy_properties(new_symbol)
assert symbol.name == "myname"
assert symbol.datatype.intrinsic == ScalarType.Intrinsic.INTEGER
assert symbol.datatype.precision == ScalarType.Precision.SINGLE
assert symbol.is_local
assert isinstance(symbol.constant_value, Literal)
assert symbol.constant_value.value == "7"
assert (
symbol.constant_value.datatype.intrinsic == symbol.datatype.intrinsic)
assert (
symbol.constant_value.datatype.precision == symbol.datatype.precision)
def example_psyir(create_expression):
'''Utility function that creates a PSyIR tree containing an ABS
intrinsic operator and returns the operator.
:param function create_expresssion: function used to create the \
content of the ABS operator.
:returns: PSyIR ABS operator instance.
:rtype: :py:class:`psyclone.psyGen.UnaryOperation`
'''
symbol_table = SymbolTable()
name1 = symbol_table.new_symbol_name("arg")
arg1 = DataSymbol(name1,
REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
symbol_table.add(arg1)
name2 = symbol_table.new_symbol_name()
local = DataSymbol(name2, REAL_TYPE)
symbol_table.add(local)
symbol_table.specify_argument_list([arg1])
var1 = Reference(arg1)
var2 = Reference(local)
oper = UnaryOperation.Operator.ABS
operation = UnaryOperation.create(oper, create_expression(var1))
assign = Assignment.create(var2, operation)
_ = KernelSchedule.create("abs_example", symbol_table, [assign])
return operation
def test_generic_scalars(data_type, symbol, intrinsic, precision):
'''Test the generated generic scalar datatypes and symbols are created
correctly.
'''
# datatype
lfric_datatype = data_type()
assert lfric_datatype.intrinsic == intrinsic
assert lfric_datatype.precision is precision
# precision can be set explicitly
lfric_datatype = data_type(precision=4)
assert lfric_datatype.precision == 4
# symbol
lfric_symbol = symbol("symbol")
assert lfric_symbol.name == "symbol"
assert isinstance(lfric_symbol.interface, LocalInterface)
assert isinstance(lfric_symbol.datatype, data_type)
lfric_symbol = symbol("symbol",
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
assert isinstance(lfric_symbol.interface, ArgumentInterface)
assert lfric_symbol.interface.access == ArgumentInterface.Access.READ
# precision can be set explicitly
lfric_symbol = symbol("symbol", precision=4)
assert lfric_symbol.datatype.precision == 4
def scalar(self, scalar_arg, var_accesses=None):
'''Create an LFRic scalar argument and add it to the symbol table and
argument list.
:param scalar_arg: the scalar to add.
:type scalar_arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: if the datatype of the scalar is \
not supported.
'''
mapping = {
"integer": lfric_psyir.LfricIntegerScalarDataSymbol,
"real": lfric_psyir.LfricRealScalarDataSymbol,
"logical": lfric_psyir.LfricLogicalScalarDataSymbol
}
try:
symbol = self._symbol_table.symbol_from_tag(
scalar_arg.name,
symbol_type=mapping[scalar_arg.intrinsic_type],
interface=ArgumentInterface(INTENT_MAPPING[scalar_arg.intent]))
except KeyError as info:
message = (
"scalar of type '{0}' not implemented in KernelInterface "
"class.".format(scalar_arg.intrinsic_type))
six.raise_from(NotImplementedError(message), info)
self._arglist.append(symbol)
def test_cw_gen_declaration():
'''Check the CWriter class gen_declaration method produces
the expected declarations.
'''
cwriter = CWriter()
# Basic entries
symbol = DataSymbol("dummy1", INTEGER_TYPE)
result = cwriter.gen_declaration(symbol)
assert result == "int dummy1"
symbol = DataSymbol("dummy1", CHARACTER_TYPE)
result = cwriter.gen_declaration(symbol)
assert result == "char dummy1"
symbol = DataSymbol("dummy1", BOOLEAN_TYPE)
result = cwriter.gen_declaration(symbol)
assert result == "bool dummy1"
# Array argument
array_type = ArrayType(REAL_TYPE, [2, ArrayType.Extent.ATTRIBUTE, 2])
symbol = DataSymbol("dummy2",
array_type,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
result = cwriter.gen_declaration(symbol)
assert result == "double * restrict dummy2"
# Array with unknown access
array_type = ArrayType(INTEGER_TYPE, [2, ArrayType.Extent.ATTRIBUTE, 2])
symbol = DataSymbol("dummy2",
array_type,
interface=ArgumentInterface(
ArgumentInterface.Access.UNKNOWN))
result = cwriter.gen_declaration(symbol)
assert result == "int * restrict dummy2"
# Check invalid datatype produces and error
symbol._datatype = "invalid"
with pytest.raises(NotImplementedError) as error:
_ = cwriter.gen_declaration(symbol)
assert "Could not generate the C definition for the variable 'dummy2', " \
"type 'invalid' is currently not supported." in str(error.value)
def test_datasymbol_copy():
'''Test that the DataSymbol copy method produces a faithful separate copy
of the original symbol.
'''
array_type = ArrayType(REAL_SINGLE_TYPE, [1, 2])
symbol = DataSymbol("myname",
array_type,
constant_value=None,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
new_symbol = symbol.copy()
# Check the new symbol has the same properties as the original
assert symbol.name == new_symbol.name
assert symbol.datatype == new_symbol.datatype
assert symbol.shape == new_symbol.shape
assert symbol.constant_value == new_symbol.constant_value
assert symbol.interface == new_symbol.interface
# Change the properties of the new symbol and check the original
# is not affected. Can't check constant_value yet as we have a
# shape value
new_symbol._name = "new"
new_symbol.datatype = ArrayType(
ScalarType(ScalarType.Intrinsic.INTEGER, ScalarType.Precision.DOUBLE),
[3, 4])
new_symbol._interface = LocalInterface()
assert symbol.name == "myname"
assert symbol.datatype.intrinsic == ScalarType.Intrinsic.REAL
assert symbol.datatype.precision == ScalarType.Precision.SINGLE
assert len(symbol.shape) == 2
assert isinstance(symbol.shape[0], Literal)
assert symbol.shape[0].value == "1"
assert symbol.shape[0].datatype.intrinsic == ScalarType.Intrinsic.INTEGER
assert symbol.shape[0].datatype.precision == ScalarType.Precision.UNDEFINED
assert isinstance(symbol.shape[1], Literal)
assert symbol.shape[1].value == "2"
assert symbol.shape[1].datatype.intrinsic == ScalarType.Intrinsic.INTEGER
assert symbol.shape[1].datatype.precision == ScalarType.Precision.UNDEFINED
assert not symbol.constant_value
# Now check constant_value
new_symbol.constant_value = 3
assert isinstance(symbol.shape[0], Literal)
assert symbol.shape[0].value == "1"
assert symbol.shape[0].datatype.intrinsic == ScalarType.Intrinsic.INTEGER
assert symbol.shape[0].datatype.precision == ScalarType.Precision.UNDEFINED
assert isinstance(symbol.shape[1], Literal)
assert symbol.shape[1].value == "2"
assert symbol.shape[1].datatype.intrinsic == ScalarType.Intrinsic.INTEGER
assert symbol.shape[1].datatype.precision == ScalarType.Precision.UNDEFINED
assert not symbol.constant_value
def test_datasymbol_initialisation():
'''Test that a DataSymbol instance can be created when valid arguments are
given, otherwise raise relevant exceptions.'''
# Test with valid arguments
assert isinstance(DataSymbol('a', REAL_SINGLE_TYPE), DataSymbol)
assert isinstance(DataSymbol('a', REAL_DOUBLE_TYPE), DataSymbol)
assert isinstance(DataSymbol('a', REAL4_TYPE), DataSymbol)
kind = DataSymbol('r_def', INTEGER_SINGLE_TYPE)
real_kind_type = ScalarType(ScalarType.Intrinsic.REAL, kind)
assert isinstance(DataSymbol('a', real_kind_type),
DataSymbol)
# real constants are not currently supported
assert isinstance(DataSymbol('a', INTEGER_SINGLE_TYPE), DataSymbol)
assert isinstance(DataSymbol('a', INTEGER_DOUBLE_TYPE, constant_value=0),
DataSymbol)
assert isinstance(DataSymbol('a', INTEGER4_TYPE),
DataSymbol)
assert isinstance(DataSymbol('a', CHARACTER_TYPE), DataSymbol)
assert isinstance(DataSymbol('a', CHARACTER_TYPE,
constant_value="hello"), DataSymbol)
assert isinstance(DataSymbol('a', BOOLEAN_TYPE), DataSymbol)
assert isinstance(DataSymbol('a', BOOLEAN_TYPE,
constant_value=False),
DataSymbol)
array_type = ArrayType(REAL_SINGLE_TYPE, [ArrayType.Extent.ATTRIBUTE])
assert isinstance(DataSymbol('a', array_type), DataSymbol)
array_type = ArrayType(REAL_SINGLE_TYPE, [3])
assert isinstance(DataSymbol('a', array_type), DataSymbol)
array_type = ArrayType(REAL_SINGLE_TYPE, [3, ArrayType.Extent.ATTRIBUTE])
assert isinstance(DataSymbol('a', array_type), DataSymbol)
assert isinstance(DataSymbol('a', REAL_SINGLE_TYPE), DataSymbol)
assert isinstance(DataSymbol('a', REAL8_TYPE), DataSymbol)
dim = DataSymbol('dim', INTEGER_SINGLE_TYPE,
interface=UnresolvedInterface())
array_type = ArrayType(REAL_SINGLE_TYPE, [Reference(dim)])
assert isinstance(DataSymbol('a', array_type), DataSymbol)
array_type = ArrayType(REAL_SINGLE_TYPE,
[3, Reference(dim), ArrayType.Extent.ATTRIBUTE])
assert isinstance(DataSymbol('a', array_type), DataSymbol)
assert isinstance(
DataSymbol('a', REAL_SINGLE_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE)), DataSymbol)
assert isinstance(
DataSymbol('a', REAL_SINGLE_TYPE,
visibility=Symbol.Visibility.PRIVATE), DataSymbol)
assert isinstance(DataSymbol('field', DataTypeSymbol("field_type",
DeferredType())),
DataSymbol)
def test_datasymbol_constant_value_setter_invalid():
'''Test that a DataSymbol constant value setter raises the appropriate
error if an invalid value and/or datatype are given.'''
# Test with invalid constant values
sym = DataSymbol('a', DeferredType())
with pytest.raises(ValueError) as error:
sym.constant_value = 1.0
assert ("Error setting constant value for symbol 'a'. A DataSymbol with "
"a constant value must be a scalar or an array but found "
"'DeferredType'." in str(error.value))
# Test with invalid constant expressions
ct_expr = Return()
with pytest.raises(ValueError) as error:
_ = DataSymbol('a', INTEGER_SINGLE_TYPE, constant_value=ct_expr)
assert "Error setting constant value for symbol 'a'. PSyIR static " \
"expressions can only contain PSyIR literal, operation or reference" \
" nodes but found:" in str(error.value)
with pytest.raises(ValueError) as error:
DataSymbol('a',
INTEGER_SINGLE_TYPE,
interface=ArgumentInterface(),
constant_value=9)
assert ("Error setting constant value for symbol 'a'. A DataSymbol with "
"an ArgumentInterface can not have a constant value."
in str(error.value))
with pytest.raises(ValueError) as error:
DataSymbol('a', INTEGER_SINGLE_TYPE, constant_value=9.81)
assert ("Error setting constant value for symbol 'a'. This DataSymbol "
"instance datatype is 'Scalar<INTEGER, SINGLE>' which "
"means the constant value is expected to be") in str(error.value)
assert "'int'>' but found " in str(error.value)
assert "'float'>'." in str(error.value)
with pytest.raises(ValueError) as error:
DataSymbol('a', CHARACTER_TYPE, constant_value=42)
assert ("Error setting constant value for symbol 'a'. This DataSymbol "
"instance datatype is 'Scalar<CHARACTER, UNDEFINED>' which "
"means the constant value is expected to be") in str(error.value)
assert "'str'>' but found " in str(error.value)
assert "'int'>'." in str(error.value)
with pytest.raises(ValueError) as error:
DataSymbol('a', BOOLEAN_TYPE, constant_value="hello")
assert ("Error setting constant value for symbol 'a'. This DataSymbol "
"instance datatype is 'Scalar<BOOLEAN, UNDEFINED>' which "
"means the constant value is expected to be") in str(error.value)
assert "'bool'>' but found " in str(error.value)
assert "'str'>'." in str(error.value)
def operator(self, arg, var_accesses=None):
'''Create an LFRic operator argument and an ncells argument and add
them to the symbol table and argument list. Also declare the
associated 'fs_from', 'fs_to' symbols if they have not already
been declared so that they can be used to dimension the
operator symbol (as well as ncells).
:param arg: the operator to add.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: if the datatype of the field is \
not supported.
'''
fs_from_name = arg.function_space_from.orig_name
ndf_symbol_from = self._symbol_table.symbol_from_tag(
"ndf_{0}".format(fs_from_name),
fs=fs_from_name,
symbol_type=lfric_psyir.NumberOfDofsDataSymbol,
interface=self._read_access)
fs_to_name = arg.function_space_to.orig_name
ndf_symbol_to = self._symbol_table.symbol_from_tag(
"ndf_{0}".format(fs_to_name),
fs=fs_to_name,
symbol_type=lfric_psyir.NumberOfDofsDataSymbol,
interface=self._read_access)
ncells = lfric_psyir.NumberOfCellsDataSymbol(
"ncell_3d", interface=self._read_access)
self._symbol_table.add(ncells)
self._arglist.append(ncells)
op_arg_symbol = self._symbol_table.symbol_from_tag(
arg.name,
symbol_type=lfric_psyir.OperatorDataSymbol,
dims=[
Reference(ndf_symbol_from),
Reference(ndf_symbol_to),
Reference(ncells)
],
fs_from=fs_from_name,
fs_to=fs_to_name,
interface=ArgumentInterface(INTENT_MAPPING[arg.intent]))
self._arglist.append(op_arg_symbol)
def test_specific_scalar_symbols(symbol, generic_symbol, attribute_map):
'''Test the generated specific scalar symbols are
created correctly.
'''
args = ["symbol"] + list(attribute_map.values())
lfric_symbol = symbol(*args)
assert isinstance(lfric_symbol, generic_symbol)
assert lfric_symbol.name == "symbol"
assert isinstance(lfric_symbol.interface, LocalInterface)
for attribute in attribute_map:
assert getattr(lfric_symbol, attribute) == attribute_map[attribute]
lfric_symbol = symbol(*args,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
assert isinstance(lfric_symbol.interface, ArgumentInterface)
assert lfric_symbol.interface.access == ArgumentInterface.Access.READ
def test_field_vector(monkeypatch):
'''Test the KernelInterface class field_vector method. We want to
check that the correct symbol is referenced for the dimension of
the vector field symbols so the simplest solution is to use one of
the Fortran test examples. Also check that the expected exception
is raised if the type of the vector field is not supported.
'''
kernel_interface = KernelInterface(None)
_, invoke_info = parse(os.path.join(BASE_PATH, "8_vector_field.f90"),
api="dynamo0.3")
psy = PSyFactory("dynamo0.3", distributed_memory=False).create(invoke_info)
schedule = psy.invokes.invoke_list[0].schedule
kernel = schedule[0].loop_body[0]
vector_arg = kernel.args[1]
kernel_interface.field_vector(vector_arg)
# undf symbol declared
undf_tag = "undf_{0}".format(vector_arg.function_space.orig_name)
undf_symbol = kernel_interface._symbol_table.lookup(undf_tag)
assert isinstance(undf_symbol, lfric_psyir.NumberOfUniqueDofsDataSymbol)
assert isinstance(undf_symbol.interface, ArgumentInterface)
assert (
undf_symbol.interface.access == kernel_interface._read_access.access)
# vector fields declared, added to argument list, correct function
# space specified and dimensioned correctly
for idx in range(vector_arg.vector_size):
tag = "{0}_v{1}".format(vector_arg.name, idx)
symbol = kernel_interface._symbol_table.lookup(tag)
assert isinstance(symbol, lfric_psyir.RealVectorFieldDataDataSymbol)
assert isinstance(symbol.interface, ArgumentInterface)
assert (symbol.interface.access == ArgumentInterface(
INTENT_MAPPING[vector_arg.intent]).access)
assert kernel_interface._arglist[idx - 3] is symbol
assert symbol.fs == vector_arg.function_space.orig_name
assert len(symbol.shape) == 1
assert isinstance(symbol.shape[0], Reference)
assert symbol.shape[0].symbol is undf_symbol
# Force an exception by setting the intrinsic type to an unsupported value
monkeypatch.setattr(vector_arg, "_intrinsic_type", "unsupported")
with pytest.raises(NotImplementedError) as info:
kernel_interface.field_vector(vector_arg)
assert ("kernel interface does not support a vector field of type "
"'unsupported'." in str(info.value))
def field_vector(self, argvect, var_accesses=None):
'''Create LFRic field vector arguments and add them to the symbol
table and argument list. Also declare the associated "undf"
symbol if it has not already been declared so that it can be
used to dimension the field vector arguments.
:param argvect: the field vector to add.
:type argvect: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: if the datatype of the vector \
field is not supported.
'''
fs_name = argvect.function_space.orig_name
undf_symbol = self._symbol_table.symbol_from_tag(
"undf_{0}".format(fs_name),
fs=fs_name,
symbol_type=lfric_psyir.NumberOfUniqueDofsDataSymbol,
interface=self._read_access)
interface = ArgumentInterface(INTENT_MAPPING[argvect.intent])
try:
field_class = self.vector_field_mapping[argvect.intrinsic_type]
except KeyError as info:
message = ("kernel interface does not support a vector field of "
"type '{0}'.".format(argvect.intrinsic_type))
six.raise_from(NotImplementedError(message), info)
for idx in range(argvect.vector_size):
tag = "{0}_v{1}".format(argvect.name, idx)
field_data_symbol = self._symbol_table.symbol_from_tag(
tag,
symbol_type=field_class,
dims=[Reference(undf_symbol)],
fs=fs_name,
interface=interface)
self._arglist.append(field_data_symbol)
def test_arrays(data_type, symbol, scalar_type, dims, attribute_map):
'''Test the generated array datatypes and datasymbols are created
correctly. This includes field datatypes and symbols which are
kept as a separate list in psyir.py
'''
# Datatype creation
lfric_datatype = data_type(dims)
assert isinstance(lfric_datatype, ArrayType)
assert isinstance(lfric_datatype._datatype, scalar_type)
for idx, dim in enumerate(lfric_datatype.shape):
if isinstance(dim, Literal):
assert dim.value == str(dims[idx])
elif isinstance(dim, Reference):
assert dim is dims[idx]
assert dim.symbol is dims[idx].symbol
else:
assert False, "unexpected type of dimension found"
# Wrong number of dims
with pytest.raises(TypeError) as info:
_ = data_type([])
assert ("{0} expected the number of supplied dimensions to be {1} but "
"found 0.".format(type(lfric_datatype).__name__, len(dims))
in str(info.value))
# Datasymbol creation
args = list(attribute_map.values())
lfric_symbol = symbol("symbol", dims, *args)
assert isinstance(lfric_symbol, DataSymbol)
assert lfric_symbol.name == "symbol"
assert isinstance(lfric_symbol.interface, LocalInterface)
assert isinstance(lfric_symbol.datatype, data_type)
lfric_symbol = symbol("symbol",
dims,
*args,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
assert isinstance(lfric_symbol.interface, ArgumentInterface)
assert lfric_symbol.interface.access == ArgumentInterface.Access.READ
def test_argumentinterface_init():
'''Check that the ArgumentInterface can be created successfully and
has the expected values. Also checks the access property and that
an exception is raised if the supplied access value is the wrong
type.
'''
argument_interface = ArgumentInterface()
assert argument_interface._access == ArgumentInterface.Access.UNKNOWN
assert argument_interface.access == argument_interface._access
assert argument_interface._pass_by_value is False
argument_interface = ArgumentInterface(ArgumentInterface.Access.READ)
assert argument_interface._access == ArgumentInterface.Access.READ
assert argument_interface.access == argument_interface._access
with pytest.raises(TypeError) as info:
_ = ArgumentInterface("hello")
assert ("SymbolInterface.access must be an 'ArgumentInterface.Access' but "
"got 'str'." in str(info.value))
with pytest.raises(TypeError) as info:
argument_interface.access = "hello"
assert ("SymbolInterface.access must be an 'ArgumentInterface.Access' but "
"got 'str'." in str(info.value))
Once you have psyclone installed, this script may be run by doing:
>>> python create.py
This should output a Fortran representation of the LFRic-PSyIR.
'''
# pylint: disable=no-member
from __future__ import print_function
from psyclone.psyir.nodes import Call, Reference, Container, KernelSchedule
from psyclone.psyir.symbols import RoutineSymbol, SymbolTable, \
ArgumentInterface
from psyclone.domain.lfric import psyir as lfric_psyir
from psyclone.psyir.backend.fortran import FortranWriter
READ_ARG = ArgumentInterface(ArgumentInterface.Access.READ)
# Add LFRic precision symbols and the module in which they are
# contained to the symbol table
SYMBOL_TABLE = SymbolTable()
for symbol in [
lfric_psyir.I_DEF, lfric_psyir.R_DEF, lfric_psyir.CONSTANTS_MOD
]:
SYMBOL_TABLE.add(symbol)
# Create LFRic ndf and undf symbols and add them to the symbol table
NDF_W3 = lfric_psyir.NumberOfDofsDataSymbol("ndf_w3", "w3", interface=READ_ARG)
UNDF_W3 = lfric_psyir.NumberOfUniqueDofsDataSymbol("undf_w3",
"w3",
interface=READ_ARG)
for symbol in [NDF_W3, UNDF_W3]:
def test_symtab_implementation_for_opencl():
''' Tests that the GOcean specialised Symbol Table implements the
abstract properties needed to generate OpenCL.
'''
kschedule = GOKernelSchedule('test')
# Test symbol table without any kernel argument
with pytest.raises(GenerationError) as err:
_ = kschedule.symbol_table.iteration_indices
assert ("GOcean 1.0 API kernels should always have at least two "
"arguments representing the iteration indices but the Symbol "
"Table for kernel 'test' has only 0 argument(s)."
in str(err.value))
# Test symbol table with 1 kernel argument
arg1 = DataSymbol("arg1",
INTEGER_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
kschedule.symbol_table.add(arg1)
kschedule.symbol_table.specify_argument_list([arg1])
with pytest.raises(GenerationError) as err:
_ = kschedule.symbol_table.iteration_indices
assert ("GOcean 1.0 API kernels should always have at least two "
"arguments representing the iteration indices but the Symbol "
"Table for kernel 'test' has only 1 argument(s)."
in str(err.value))
# Test symbol table with 2 kernel argument
arg2 = DataSymbol("arg2",
INTEGER_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
kschedule.symbol_table.add(arg2)
kschedule.symbol_table.specify_argument_list([arg1, arg2])
iteration_indices = kschedule.symbol_table.iteration_indices
assert iteration_indices[0] is arg1
assert iteration_indices[1] is arg2
# Test symbol table with 3 kernel argument
array_type = ArrayType(REAL_TYPE, [10, 10])
arg3 = DataSymbol("buffer1",
array_type,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
kschedule.symbol_table.add(arg3)
kschedule.symbol_table.specify_argument_list([arg1, arg2, arg3])
iteration_indices = kschedule.symbol_table.iteration_indices
data_args = kschedule.symbol_table.data_arguments
assert iteration_indices[0] is arg1
assert iteration_indices[1] is arg2
assert data_args[0] is arg3
# Test gen_ocl with wrong iteration indices types and shapes.
arg1._datatype._intrinsic = ScalarType.Intrinsic.REAL
with pytest.raises(GenerationError) as err:
_ = kschedule.symbol_table.iteration_indices
assert ("GOcean 1.0 API kernels first argument should be a scalar "
"integer but got 'Scalar<REAL, UNDEFINED>' for kernel 'test'."
in str(err.value))
arg1._datatype._intrinsic = ScalarType.Intrinsic.INTEGER # restore
arg2._datatype = ArrayType(INTEGER_TYPE, [10])
with pytest.raises(GenerationError) as err:
_ = kschedule.symbol_table.iteration_indices
assert ("GOcean 1.0 API kernels second argument should be a scalar integer"
" but got 'Array<Scalar<INTEGER, UNDEFINED>, shape=[10]>' for "
"kernel 'test'." in str(err.value))
def test_scoping_node_copy_hierarchy():
''' Test that the ScopingNode copy() method creates a new symbol table
with copied symbols and updates the children references.
This test has 2 ScopingNodes, and the copy will only be applied to the
inner one. This means that the References to the symbols on the outer
scope should not be duplicated. Also it contains argument symbols and
a reference inside another reference to make sure all these are copied
appropriately.
'''
parent_node = Container("module")
symbol_b = parent_node.symbol_table.new_symbol("b",
symbol_type=DataSymbol,
datatype=ArrayType(
INTEGER_TYPE, [5]))
schedule = Routine("routine")
parent_node.addchild(schedule)
symbol_a = schedule.symbol_table.new_symbol(
"a",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE,
interface=ArgumentInterface(ArgumentInterface.Access.READWRITE))
schedule.symbol_table.specify_argument_list([symbol_a])
symbol_i = schedule.symbol_table.new_symbol("i",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE)
schedule.addchild(
Assignment.create(
Reference(symbol_a),
ArrayReference.create(symbol_b, [Reference(symbol_i)])))
new_schedule = schedule.copy()
# Check that the symbol_table has been deep copied
assert new_schedule.symbol_table is not schedule.symbol_table
assert new_schedule.symbol_table.lookup("i") is not \
schedule.symbol_table.lookup("i")
assert new_schedule.symbol_table.lookup("a") is not \
schedule.symbol_table.lookup("a")
# Check that 'a' and 'i' have been copied to the new symbol table.
assert new_schedule[0].lhs.symbol not in schedule.symbol_table.symbols
assert new_schedule[0].lhs.symbol in new_schedule.symbol_table.symbols
assert new_schedule[0].rhs.children[0].symbol not in \
schedule.symbol_table.symbols
assert new_schedule[0].rhs.children[0].symbol in \
new_schedule.symbol_table.symbols
# Add the "_new" suffix to all symbol in the copied schedule
for symbol in new_schedule.symbol_table.symbols:
new_schedule.symbol_table.rename_symbol(symbol, symbol.name + "_new")
# An update to a symbol in the outer scope must affect both copies of the
# inner schedule.
parent_node.symbol_table.rename_symbol(symbol_b, symbol_b.name + "_global")
# Insert the schedule back to the original container
parent_node.addchild(new_schedule)
# Check that the expected code is generated
# TODO #1200: the new 'routine' RoutineSymbol also needs to change.
expected = '''\
module module
implicit none
integer, dimension(5) :: b_global
contains
subroutine routine(a)
integer, intent(inout) :: a
integer :: i
a = b_global(i)
end subroutine routine
subroutine routine(a_new)
integer, intent(inout) :: a_new
integer :: i_new
a_new = b_global(i_new)
end subroutine routine
end module module
'''
writer = FortranWriter()
output = writer(parent_node)
assert expected == output
def create_psyir_tree():
''' Create an example PSyIR Tree.
:returns: an example PSyIR tree.
:rtype: :py:class:`psyclone.psyir.nodes.Container`
'''
# Symbol table, symbols and scalar datatypes
symbol_table = SymbolTable()
arg1 = symbol_table.new_symbol(symbol_type=DataSymbol,
datatype=REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
symbol_table.specify_argument_list([arg1])
tmp_symbol = symbol_table.new_symbol(symbol_type=DataSymbol,
datatype=REAL_DOUBLE_TYPE)
index_symbol = symbol_table.new_symbol(root_name="i",
symbol_type=DataSymbol,
datatype=INTEGER4_TYPE)
real_kind = symbol_table.new_symbol(root_name="RKIND",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE,
constant_value=8)
routine_symbol = RoutineSymbol("my_sub")
# Array using precision defined by another symbol
scalar_type = ScalarType(ScalarType.Intrinsic.REAL, real_kind)
array = symbol_table.new_symbol(root_name="a",
symbol_type=DataSymbol,
datatype=ArrayType(scalar_type, [10]))
# Make generators for nodes which do not have other Nodes as children,
# with some predefined scalar datatypes
def zero():
return Literal("0.0", REAL_TYPE)
def one():
return Literal("1.0", REAL4_TYPE)
def two():
return Literal("2.0", scalar_type)
def int_zero():
return Literal("0", INTEGER_SINGLE_TYPE)
def int_one():
return Literal("1", INTEGER8_TYPE)
def tmp1():
return Reference(arg1)
def tmp2():
return Reference(tmp_symbol)
# Unary Operation
oper = UnaryOperation.Operator.SIN
unaryoperation = UnaryOperation.create(oper, tmp2())
# Binary Operation
oper = BinaryOperation.Operator.ADD
binaryoperation = BinaryOperation.create(oper, one(), unaryoperation)
# Nary Operation
oper = NaryOperation.Operator.MAX
naryoperation = NaryOperation.create(oper, [tmp1(), tmp2(), one()])
# Array reference using a range
lbound = BinaryOperation.create(BinaryOperation.Operator.LBOUND,
Reference(array), int_one())
ubound = BinaryOperation.create(BinaryOperation.Operator.UBOUND,
Reference(array), int_one())
my_range = Range.create(lbound, ubound)
tmparray = ArrayReference.create(array, [my_range])
# Assignments
assign1 = Assignment.create(tmp1(), zero())
assign2 = Assignment.create(tmp2(), zero())
assign3 = Assignment.create(tmp2(), binaryoperation)
assign4 = Assignment.create(tmp1(), tmp2())
assign5 = Assignment.create(tmp1(), naryoperation)
assign6 = Assignment.create(tmparray, two())
# Call
call = Call.create(routine_symbol, [tmp1(), binaryoperation.copy()])
# If statement
if_condition = BinaryOperation.create(BinaryOperation.Operator.GT, tmp1(),
zero())
ifblock = IfBlock.create(if_condition, [assign3, assign4])
# Loop
loop = Loop.create(index_symbol, int_zero(), int_one(), int_one(),
[ifblock])
# KernelSchedule
kernel_schedule = KernelSchedule.create(
"work", symbol_table, [assign1, call, assign2, loop, assign5, assign6])
# Container
container_symbol_table = SymbolTable()
container = Container.create("CONTAINER", container_symbol_table,
[kernel_schedule])
# Import data from another container
external_container = ContainerSymbol("some_mod")
container_symbol_table.add(external_container)
external_var = DataSymbol("some_var",
INTEGER_TYPE,
interface=GlobalInterface(external_container))
container_symbol_table.add(external_var)
routine_symbol.interface = GlobalInterface(external_container)
container_symbol_table.add(routine_symbol)
return container
def apply(self, node, options=None):
'''Apply this transformation to the supplied node.
:param node: the node to transform.
:type node: :py:class:`psyclone.gocean1p0.GOKern`
:param options: a dictionary with options for transformations.
:type options: dict of string:values or None
:returns: 2-tuple of new schedule and memento of transform.
:rtype: (:py:class:`psyclone.gocean1p0.GOInvokeSchedule`, \
:py:class:`psyclone.undoredo.Memento`)
'''
self.validate(node, options)
# Get useful references
invoke_st = node.ancestor(InvokeSchedule).symbol_table
inner_loop = node.ancestor(Loop)
outer_loop = inner_loop.ancestor(Loop)
cursor = outer_loop.position
# Make sure the boundary symbols in the PSylayer exist
inv_xstart = invoke_st.symbol_from_tag("xstart_" + node.name,
root_name="xstart",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE)
inv_xstop = invoke_st.symbol_from_tag("xstop_" + node.name,
root_name="xstop",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE)
inv_ystart = invoke_st.symbol_from_tag("ystart_" + node.name,
root_name="ystart",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE)
inv_ystop = invoke_st.symbol_from_tag("ystop_" + node.name,
root_name="ystop",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE)
# If the kernel acts on the whole iteration space, the boundary values
# are not needed. This also avoids adding duplicated arguments if this
# transformation is applied more than once to the same kernel. But the
# declaration and initialisation above still needs to exist because the
# boundary variables are expected to exist by the generation code.
if (inner_loop.field_space == "go_every"
and outer_loop.field_space == "go_every"
and inner_loop.iteration_space == "go_all_pts"
and outer_loop.iteration_space == "go_all_pts"):
return node.root, None
# Initialise the boundary values provided by the Loop construct
assign1 = Assignment.create(Reference(inv_xstart),
inner_loop.lower_bound().copy())
outer_loop.parent.children.insert(cursor, assign1)
cursor = cursor + 1
assign2 = Assignment.create(Reference(inv_xstop),
inner_loop.upper_bound().copy())
outer_loop.parent.children.insert(cursor, assign2)
cursor = cursor + 1
assign3 = Assignment.create(Reference(inv_ystart),
outer_loop.lower_bound().copy())
outer_loop.parent.children.insert(cursor, assign3)
cursor = cursor + 1
assign4 = Assignment.create(Reference(inv_ystop),
outer_loop.upper_bound().copy())
outer_loop.parent.children.insert(cursor, assign4)
# Update Kernel Call argument list
for symbol in [inv_xstart, inv_xstop, inv_ystart, inv_ystop]:
node.arguments.append(symbol.name, "go_i_scalar")
# Now that the boundaries are inside the kernel, the looping should go
# through all the field points
inner_loop.field_space = "go_every"
outer_loop.field_space = "go_every"
inner_loop.iteration_space = "go_all_pts"
outer_loop.iteration_space = "go_all_pts"
# Update Kernel
kschedule = node.get_kernel_schedule()
kernel_st = kschedule.symbol_table
iteration_indices = kernel_st.iteration_indices
data_arguments = kernel_st.data_arguments
# Create new symbols and insert them as kernel arguments at the end of
# the kernel argument list
xstart_symbol = kernel_st.new_symbol(
"xstart",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE,
interface=ArgumentInterface(ArgumentInterface.Access.READ))
xstop_symbol = kernel_st.new_symbol("xstop",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
ystart_symbol = kernel_st.new_symbol(
"ystart",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE,
interface=ArgumentInterface(ArgumentInterface.Access.READ))
ystop_symbol = kernel_st.new_symbol("ystop",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READ))
kernel_st.specify_argument_list(
iteration_indices + data_arguments +
[xstart_symbol, xstop_symbol, ystart_symbol, ystop_symbol])
# Create boundary masking conditions
condition1 = BinaryOperation.create(BinaryOperation.Operator.LT,
Reference(iteration_indices[0]),
Reference(xstart_symbol))
condition2 = BinaryOperation.create(BinaryOperation.Operator.GT,
Reference(iteration_indices[0]),
Reference(xstop_symbol))
condition3 = BinaryOperation.create(BinaryOperation.Operator.LT,
Reference(iteration_indices[1]),
Reference(ystart_symbol))
condition4 = BinaryOperation.create(BinaryOperation.Operator.GT,
Reference(iteration_indices[1]),
Reference(ystop_symbol))
condition = BinaryOperation.create(
BinaryOperation.Operator.OR,
BinaryOperation.create(BinaryOperation.Operator.OR, condition1,
condition2),
BinaryOperation.create(BinaryOperation.Operator.OR, condition3,
condition4))
# Insert the conditional mask as the first statement of the kernel
if_statement = IfBlock.create(condition, [Return()])
kschedule.children.insert(0, if_statement)
return node.root, None
def test_oclw_kernelschedule():
'''Check the OpenCLWriter class kernelschedule_node visitor produces
the expected OpenCL code.
'''
# The kernelschedule OpenCL Backend relies on abstract methods that
# need to be implemented by the APIs. A generic kernelschedule will
# produce a NotImplementedError.
oclwriter = OpenCLWriter()
kschedule = KernelSchedule("kname")
with pytest.raises(NotImplementedError) as error:
_ = oclwriter(kschedule)
assert "Abstract property. Which symbols are data arguments is " \
"API-specific." in str(error.value)
# Mock abstract properties. (pytest monkeypatch does not work
# with properties, used sub-class instead)
class MockSymbolTable(SymbolTable):
''' Mock needed abstract methods of the Symbol Table '''
@property
def iteration_indices(self):
return self.argument_list[:2]
@property
def data_arguments(self):
return self.argument_list[2:]
kschedule.symbol_table.__class__ = MockSymbolTable
# Create a sample symbol table and kernel schedule
interface = ArgumentInterface(ArgumentInterface.Access.UNKNOWN)
i = DataSymbol('i', INTEGER_TYPE, interface=interface)
j = DataSymbol('j', INTEGER_TYPE, interface=interface)
array_type = ArrayType(REAL_TYPE, [10, 10])
data1 = DataSymbol('data1', array_type, interface=interface)
data2 = DataSymbol('data2', array_type, interface=interface)
kschedule.symbol_table.add(i)
kschedule.symbol_table.add(j)
kschedule.symbol_table.add(data1)
kschedule.symbol_table.add(data2)
kschedule.symbol_table.specify_argument_list([i, j, data1, data2])
kschedule.addchild(Return(parent=kschedule))
result = oclwriter(kschedule)
assert result == "" \
"__kernel void kname(\n" \
" __global double * restrict data1,\n" \
" __global double * restrict data2\n" \
" ){\n" \
" int data1LEN1 = get_global_size(0);\n" \
" int data1LEN2 = get_global_size(1);\n" \
" int data2LEN1 = get_global_size(0);\n" \
" int data2LEN2 = get_global_size(1);\n" \
" int i = get_global_id(0);\n" \
" int j = get_global_id(1);\n" \
" return;\n" \
"}\n\n"
# Set a local_size value different to 1 into the KernelSchedule
oclwriter = OpenCLWriter(kernels_local_size=4)
result = oclwriter(kschedule)
assert result == "" \
"__attribute__((reqd_work_group_size(4, 1, 1)))\n" \
"__kernel void kname(\n" \
" __global double * restrict data1,\n" \
" __global double * restrict data2\n" \
" ){\n" \
" int data1LEN1 = get_global_size(0);\n" \
" int data1LEN2 = get_global_size(1);\n" \
" int data2LEN1 = get_global_size(0);\n" \
" int data2LEN2 = get_global_size(1);\n" \
" int i = get_global_id(0);\n" \
" int j = get_global_id(1);\n" \
" return;\n" \
"}\n\n"
# Add a symbol with a deferred interface and check that this raises the
# expected error
array_type = ArrayType(REAL_TYPE, [10, 10])
kschedule.symbol_table.add(
DataSymbol('broken', array_type, interface=UnresolvedInterface()))
with pytest.raises(VisitorError) as err:
_ = oclwriter(kschedule)
assert ("symbol table contains unresolved data entries (i.e. that have no "
"defined Interface) which are not used purely to define the "
"precision of other symbols: 'broken'" in str(err.value))
def test_argumentinterface_str():
'''Test that an ArgumentInterface instance can be stringified'''
argument_interface = ArgumentInterface()
assert str(argument_interface) == "Argument(pass-by-value=False)"
def test_operator():
'''Test the KernelInterface class operator method. We want to check
that the correct symbol is referenced for the dimension of the
operator symbol so the simplest solution is to use one of the Fortran
test examples.
'''
kernel_interface = KernelInterface(None)
_, invoke_info = parse(os.path.join(BASE_PATH, "10_operator.f90"),
api="dynamo0.3")
psy = PSyFactory("dynamo0.3", distributed_memory=False).create(invoke_info)
schedule = psy.invokes.invoke_list[0].schedule
kernel = schedule[0].loop_body[0]
operator_arg = kernel.args[0]
kernel_interface.operator(operator_arg)
# fs_from symbol declared
fs_from_name = operator_arg.function_space_from.orig_name
fs_from_tag = "ndf_{0}".format(fs_from_name)
fs_from_symbol = kernel_interface._symbol_table.lookup(fs_from_tag)
assert isinstance(fs_from_symbol, lfric_psyir.NumberOfDofsDataSymbol)
assert fs_from_symbol.fs == fs_from_name
assert isinstance(fs_from_symbol.interface, ArgumentInterface)
assert (fs_from_symbol.interface.access ==
kernel_interface._read_access.access)
# fs_to symbol declared
fs_to_name = operator_arg.function_space_from.orig_name
fs_to_tag = "ndf_{0}".format(fs_to_name)
fs_to_symbol = kernel_interface._symbol_table.lookup(fs_to_tag)
assert isinstance(fs_to_symbol, lfric_psyir.NumberOfDofsDataSymbol)
assert fs_to_symbol.fs == fs_to_name
assert isinstance(fs_to_symbol.interface, ArgumentInterface)
assert (
fs_to_symbol.interface.access == kernel_interface._read_access.access)
# ncells symbol declared
ncells_symbol = kernel_interface._symbol_table.lookup("ncell_3d")
assert isinstance(ncells_symbol, lfric_psyir.NumberOfCellsDataSymbol)
assert isinstance(ncells_symbol.interface, ArgumentInterface)
assert (
ncells_symbol.interface.access == kernel_interface._read_access.access)
# operator declared, added to argument list, correct function
# spaces specified and dimensioned correctly
tag = operator_arg.name
symbol = kernel_interface._symbol_table.lookup(tag)
assert isinstance(symbol, lfric_psyir.OperatorDataSymbol)
assert isinstance(symbol.interface, ArgumentInterface)
assert (symbol.interface.access == ArgumentInterface(
INTENT_MAPPING[operator_arg.intent]).access)
assert kernel_interface._arglist[-1] is symbol
assert symbol.fs_from == operator_arg.function_space_from.orig_name
assert symbol.fs_to == operator_arg.function_space_to.orig_name
assert len(symbol.shape) == 3
assert isinstance(symbol.shape[0], Reference)
assert symbol.shape[0].symbol is fs_from_symbol
assert isinstance(symbol.shape[1], Reference)
assert symbol.shape[1].symbol is fs_to_symbol
assert isinstance(symbol.shape[2], Reference)
assert symbol.shape[2].symbol is ncells_symbol
def create_psyir_tree():
''' Create an example PSyIR Tree.
:returns: an example PSyIR tree.
:rtype: :py:class:`psyclone.psyir.nodes.Container`
'''
# Symbol table, symbols and scalar datatypes
symbol_table = SymbolTable()
arg1 = symbol_table.new_symbol(symbol_type=DataSymbol,
datatype=REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
symbol_table.specify_argument_list([arg1])
tmp_symbol = symbol_table.new_symbol(symbol_type=DataSymbol,
datatype=REAL_DOUBLE_TYPE)
index_symbol = symbol_table.new_symbol(root_name="i",
symbol_type=DataSymbol,
datatype=INTEGER4_TYPE)
real_kind = symbol_table.new_symbol(root_name="RKIND",
symbol_type=DataSymbol,
datatype=INTEGER_TYPE,
constant_value=8)
routine_symbol = RoutineSymbol("my_sub")
# Array using precision defined by another symbol
scalar_type = ScalarType(ScalarType.Intrinsic.REAL, real_kind)
array = symbol_table.new_symbol(root_name="a",
symbol_type=DataSymbol,
datatype=ArrayType(scalar_type, [10]))
# Nodes which do not have Nodes as children and (some) predefined
# scalar datatypes
# TODO: Issue #1136 looks at how to avoid all of the _x versions
zero_1 = Literal("0.0", REAL_TYPE)
zero_2 = Literal("0.0", REAL_TYPE)
zero_3 = Literal("0.0", REAL_TYPE)
one_1 = Literal("1.0", REAL4_TYPE)
one_2 = Literal("1.0", REAL4_TYPE)
one_3 = Literal("1.0", REAL4_TYPE)
two = Literal("2.0", scalar_type)
int_zero = Literal("0", INTEGER_SINGLE_TYPE)
int_one_1 = Literal("1", INTEGER8_TYPE)
int_one_2 = Literal("1", INTEGER8_TYPE)
int_one_3 = Literal("1", INTEGER8_TYPE)
int_one_4 = Literal("1", INTEGER8_TYPE)
tmp1_1 = Reference(arg1)
tmp1_2 = Reference(arg1)
tmp1_3 = Reference(arg1)
tmp1_4 = Reference(arg1)
tmp1_5 = Reference(arg1)
tmp1_6 = Reference(arg1)
tmp2_1 = Reference(tmp_symbol)
tmp2_2 = Reference(tmp_symbol)
tmp2_3 = Reference(tmp_symbol)
tmp2_4 = Reference(tmp_symbol)
tmp2_5 = Reference(tmp_symbol)
tmp2_6 = Reference(tmp_symbol)
# Unary Operation
oper = UnaryOperation.Operator.SIN
unaryoperation_1 = UnaryOperation.create(oper, tmp2_1)
unaryoperation_2 = UnaryOperation.create(oper, tmp2_2)
# Binary Operation
oper = BinaryOperation.Operator.ADD
binaryoperation_1 = BinaryOperation.create(oper, one_1, unaryoperation_1)
binaryoperation_2 = BinaryOperation.create(oper, one_2, unaryoperation_2)
# Nary Operation
oper = NaryOperation.Operator.MAX
naryoperation = NaryOperation.create(oper, [tmp1_1, tmp2_3, one_3])
# Array reference using a range
lbound = BinaryOperation.create(BinaryOperation.Operator.LBOUND,
Reference(array), int_one_1)
ubound = BinaryOperation.create(BinaryOperation.Operator.UBOUND,
Reference(array), int_one_2)
my_range = Range.create(lbound, ubound)
tmparray = ArrayReference.create(array, [my_range])
# Assignments
assign1 = Assignment.create(tmp1_2, zero_1)
assign2 = Assignment.create(tmp2_4, zero_2)
assign3 = Assignment.create(tmp2_5, binaryoperation_1)
assign4 = Assignment.create(tmp1_3, tmp2_6)
assign5 = Assignment.create(tmp1_4, naryoperation)
assign6 = Assignment.create(tmparray, two)
# Call
call = Call.create(routine_symbol, [tmp1_5, binaryoperation_2])
# If statement
if_condition = BinaryOperation.create(BinaryOperation.Operator.GT, tmp1_6,
zero_3)
ifblock = IfBlock.create(if_condition, [assign3, assign4])
# Loop
loop = Loop.create(index_symbol, int_zero, int_one_3, int_one_4, [ifblock])
# KernelSchedule
kernel_schedule = KernelSchedule.create(
"work", symbol_table, [assign1, call, assign2, loop, assign5, assign6])
# Container
container_symbol_table = SymbolTable()
container = Container.create("CONTAINER", container_symbol_table,
[kernel_schedule])
# Import data from another container
external_container = ContainerSymbol("some_mod")
container_symbol_table.add(external_container)
external_var = DataSymbol("some_var",
INTEGER_TYPE,
interface=GlobalInterface(external_container))
container_symbol_table.add(external_var)
routine_symbol.interface = GlobalInterface(external_container)
container_symbol_table.add(routine_symbol)
return container
def test_validate_kernel_code_arg(monkeypatch):
'''Test that this method returns successfully if its two arguments
have identical content, otherwise test that the expected
exceptions are raised.
'''
kernel = DynKern()
# Kernel name needs to be set when testing exceptions.
kernel._name = "dummy"
read_access = ArgumentInterface(ArgumentInterface.Access.READ)
real_scalar_symbol = DataSymbol("generic_real_scalar",
REAL_TYPE,
interface=read_access)
int_scalar_symbol = DataSymbol("generic_int_scalar",
INTEGER_TYPE,
interface=read_access)
real_scalar_rw_symbol = DataSymbol("generic_scalar_rw",
REAL_TYPE,
interface=ArgumentInterface(
ArgumentInterface.Access.READWRITE))
lfric_real_scalar_symbol = LfricRealScalarDataSymbol("scalar",
interface=read_access)
lfric_int_scalar_symbol = LfricIntegerScalarDataSymbol(
"scalar", interface=read_access)
lfric_real_field_symbol = RealFieldDataDataSymbol("field",
dims=[1],
fs="w0",
interface=read_access)
kernel._validate_kernel_code_arg(lfric_real_scalar_symbol,
lfric_real_scalar_symbol)
with pytest.raises(GenerationError) as info:
kernel._validate_kernel_code_arg(lfric_real_scalar_symbol,
lfric_int_scalar_symbol)
assert ("Kernel argument 'scalar' has datatype 'Intrinsic.REAL' in kernel "
"'dummy' but the LFRic API expects 'Intrinsic.INTEGER'"
in str(info.value))
with pytest.raises(GenerationError) as info:
kernel._validate_kernel_code_arg(real_scalar_symbol,
lfric_real_scalar_symbol)
assert ("Kernel argument 'generic_real_scalar' has precision 'UNDEFINED' "
"in kernel 'dummy' but the LFRic API expects 'r_def'."
in str(info.value))
with pytest.raises(GenerationError) as info:
kernel._validate_kernel_code_arg(real_scalar_symbol,
real_scalar_rw_symbol)
assert ("Kernel argument 'generic_real_scalar' has intent 'READ' in "
"kernel 'dummy' but the LFRic API expects intent "
"'READWRITE'." in str(info.value))
with pytest.raises(GenerationError) as info:
kernel._validate_kernel_code_arg(lfric_real_field_symbol,
lfric_real_scalar_symbol)
assert ("Argument 'field' to kernel 'dummy' should be a scalar "
"according to the LFRic API, but it is not." in str(info.value))
with pytest.raises(GenerationError) as info:
kernel._validate_kernel_code_arg(lfric_real_scalar_symbol,
lfric_real_field_symbol)
assert ("Argument 'scalar' to kernel 'dummy' should be an array "
"according to the LFRic API, but it is not." in str(info.value))
undf = NumberOfUniqueDofsDataSymbol("undf", fs="w0", interface=read_access)
lfric_real_field_symbol2 = RealFieldDataDataSymbol("field",
dims=[Reference(undf)],
fs="w0",
interface=read_access)
# if one of the dimensions is not a datasymbol then the arguments
# are not checked.
kernel._validate_kernel_code_arg(lfric_real_field_symbol,
lfric_real_field_symbol2)
kernel._validate_kernel_code_arg(lfric_real_field_symbol2,
lfric_real_field_symbol)
lfric_real_field_symbol3 = RealFieldDataDataSymbol("field",
dims=[Reference(undf)],
fs="w0",
interface=read_access)
monkeypatch.setattr(lfric_real_field_symbol3.datatype, "_shape",
[Reference(undf), Reference(undf)])
with pytest.raises(GenerationError) as info:
kernel._validate_kernel_code_arg(lfric_real_field_symbol2,
lfric_real_field_symbol3)
assert ("Argument 'field' to kernel 'dummy' should be an array with 2 "
"dimension(s) according to the LFRic API, but found 1."
in str(info.value))
lfric_real_field_symbol4 = RealFieldDataDataSymbol(
"field",
dims=[Reference(int_scalar_symbol)],
fs="w0",
interface=read_access)
with pytest.raises(GenerationError) as info:
kernel._validate_kernel_code_arg(lfric_real_field_symbol4,
lfric_real_field_symbol2)
assert ("For dimension 1 in array argument 'field' to kernel 'dummy' the "
"following error was found: Kernel argument 'generic_int_scalar' "
"has precision 'UNDEFINED' in kernel 'dummy' but the LFRic API "
"expects 'i_def'" in str(info.value))
# monkeypatch lfric_real_scalar_symbol to return that it is not a
# scalar in order to force the required exception. We do this by
# changing the ScalarType as it is used when determining whether
# the symbol is a scalar.
monkeypatch.setattr(psyclone.psyir.symbols, "ScalarType", str)
with pytest.raises(InternalError) as info:
kernel._validate_kernel_code_arg(lfric_real_scalar_symbol,
lfric_real_scalar_symbol)
assert ("unexpected argument type found for 'scalar' in kernel 'dummy'. "
"Expecting a scalar or an array." in str(info.value))
class KernelInterface(ArgOrdering):
'''Create the kernel arguments for the supplied kernel as specified by
the associated kernel metadata and the kernel ordering rules
encoded in the ArgOrdering base class as method callbacks.
A PSyIR symbol table is created containing appropriate LFRic PSyIR
symbols to specify the arguments. If an argument is an array with
one or more dimension sizes specified by another argument, then
the associated array symbol is created so that it references the
appropriate symbol.
Related arguments - e.g. a field has an associated dofmap - are
not directly connected, they must be inferred from the function
space names. It is not yet clear whether this would be useful or
not.
TBD: This class should replace the current kernel stub generation
code when all of its methods are implemented, see issue #928.
:param kern: the kernel for which to create arguments.
:type kern: :py:class:`psyclone.dynamo0p3.DynKern`
'''
#: Mapping from a generic PSyIR datatype to the equivalent
#: LFRic-specific field datasymbol.
field_mapping = {
"integer": lfric_psyir.IntegerFieldDataDataSymbol,
"real": lfric_psyir.RealFieldDataDataSymbol,
"logical": lfric_psyir.LogicalFieldDataDataSymbol
}
#: Mapping from a generic PSyIR datatype to the equivalent
#: LFRic-specific vector field datasymbol.
vector_field_mapping = {
"integer": lfric_psyir.IntegerVectorFieldDataDataSymbol,
"real": lfric_psyir.RealVectorFieldDataDataSymbol,
"logical": lfric_psyir.LogicalVectorFieldDataDataSymbol
}
#: Mapping from the LFRic metadata description of quadrature to the
#: associated LFRic-specific basis function datasymbol.
basis_mapping = {
"gh_quadrature_xyoz": lfric_psyir.BasisFunctionQrXyozDataSymbol,
"gh_quadrature_face": lfric_psyir.BasisFunctionQrFaceDataSymbol,
"gh_quadrature_edge": lfric_psyir.BasisFunctionQrEdgeDataSymbol
}
#: Mapping from the LFRic metadata description of quadrature to the
#: associated LFRic-specific differential basis function datasymbol.
diff_basis_mapping = {
"gh_quadrature_xyoz": lfric_psyir.DiffBasisFunctionQrXyozDataSymbol,
"gh_quadrature_face": lfric_psyir.DiffBasisFunctionQrFaceDataSymbol,
"gh_quadrature_edge": lfric_psyir.DiffBasisFunctionQrEdgeDataSymbol
}
_read_access = ArgumentInterface(ArgumentInterface.Access.READ)
def __init__(self, kern):
super(KernelInterface, self).__init__(kern)
self._symbol_table = SymbolTable()
self._arglist = []
def generate(self, var_accesses=None):
'''Call the generate base class then add the argument list as it can't
be appended as we go along.
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
super(KernelInterface, self).generate(var_accesses=var_accesses)
# Set the argument list for the symbol table. This is done at
# the end after incrementally adding symbols to the _args
# list, as it is not possible to incrementally add symbols to
# the symbol table argument list.
self._symbol_table.specify_argument_list(self._arglist)
# While data dependence analysis does not use the symbol
# table, see #845, we have to provide variable information
# separately. This is done by using the base class append()
# method. However, this method also adds the variable names to the
# internal _arglist list which we do not want as we have
# already added our symbols there. Therefore we need to remove
# them afterwards.
# Map from symbol table accesses to dependence analysis accesses.
mapping = {
ArgumentInterface.Access.READ: AccessType.READ,
ArgumentInterface.Access.READWRITE: AccessType.READWRITE,
ArgumentInterface.Access.WRITE: AccessType.WRITE
}
len_arglist = len(self._arglist)
for symbol in self._symbol_table.symbols:
self.append(symbol.name,
var_accesses,
mode=mapping[symbol.interface.access])
self._arglist = self._arglist[:len_arglist]
def cell_position(self, var_accesses=None):
'''Create an LFRic cell-position object and add it to the symbol table
and argument list.
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
symbol = self._symbol_table.symbol_from_tag(
"cell",
symbol_type=lfric_psyir.CellPositionDataSymbol,
interface=self._read_access)
self._arglist.append(symbol)
def mesh_height(self, var_accesses=None):
'''Create an LFRic mesh height object and add it to the symbol table
and argument list.
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
symbol = self._symbol_table.symbol_from_tag(
"nlayers",
symbol_type=lfric_psyir.MeshHeightDataSymbol,
interface=self._read_access)
self._arglist.append(symbol)
def mesh_ncell2d(self, var_accesses=None):
'''Not implemented.
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("mesh_ncell2d not implemented")
def cell_map(self, var_accesses=None):
'''Not implemented.
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("cell_map not implemented")
def field_vector(self, argvect, var_accesses=None):
'''Create LFRic field vector arguments and add them to the symbol
table and argument list. Also declare the associated "undf"
symbol if it has not already been declared so that it can be
used to dimension the field vector arguments.
:param argvect: the field vector to add.
:type argvect: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: if the datatype of the vector \
field is not supported.
'''
fs_name = argvect.function_space.orig_name
undf_symbol = self._symbol_table.symbol_from_tag(
"undf_{0}".format(fs_name),
fs=fs_name,
symbol_type=lfric_psyir.NumberOfUniqueDofsDataSymbol,
interface=self._read_access)
interface = ArgumentInterface(INTENT_MAPPING[argvect.intent])
try:
field_class = self.vector_field_mapping[argvect.intrinsic_type]
except KeyError as info:
message = ("kernel interface does not support a vector field of "
"type '{0}'.".format(argvect.intrinsic_type))
six.raise_from(NotImplementedError(message), info)
for idx in range(argvect.vector_size):
tag = "{0}_v{1}".format(argvect.name, idx)
field_data_symbol = self._symbol_table.symbol_from_tag(
tag,
symbol_type=field_class,
dims=[Reference(undf_symbol)],
fs=fs_name,
interface=interface)
self._arglist.append(field_data_symbol)
def field(self, arg, var_accesses=None):
'''Create an LFRic field argument and add it to the symbol table and
argument list. Also declare the associated "undf" symbol if it
has not already been declared so that it can be used to
dimension the field argument.
:param arg: the field to add.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: if the datatype of the field is \
not supported.
'''
fs_name = arg.function_space.orig_name
undf_symbol = self._symbol_table.symbol_from_tag(
"undf_{0}".format(fs_name),
symbol_type=lfric_psyir.NumberOfUniqueDofsDataSymbol,
fs=fs_name,
interface=self._read_access)
try:
field_class = self.field_mapping[arg.intrinsic_type]
except KeyError as info:
message = ("kernel interface does not support a field of type "
"'{0}'.".format(arg.intrinsic_type))
six.raise_from(NotImplementedError(message), info)
field_data_symbol = self._symbol_table.symbol_from_tag(
arg.name,
interface=ArgumentInterface(INTENT_MAPPING[arg.intent]),
symbol_type=field_class,
dims=[Reference(undf_symbol)],
fs=fs_name)
self._arglist.append(field_data_symbol)
def stencil_unknown_extent(self, arg, var_accesses=None):
'''Not implemented.
:param arg: the kernel argument with which the stencil is associated.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("stencil_unknown_extent not implemented")
def stencil_unknown_direction(self, arg, var_accesses=None):
'''Not implemented.
:param arg: the kernel argument with which the stencil is associated.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("stencil_unknown_direction not implemented")
def stencil(self, arg, var_accesses=None):
'''Not implemented.
:param arg: the kernel argument with which the stencil is associated.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("stencil not implemented")
def operator(self, arg, var_accesses=None):
'''Create an LFRic operator argument and an ncells argument and add
them to the symbol table and argument list. Also declare the
associated 'fs_from', 'fs_to' symbols if they have not already
been declared so that they can be used to dimension the
operator symbol (as well as ncells).
:param arg: the operator to add.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: if the datatype of the field is \
not supported.
'''
fs_from_name = arg.function_space_from.orig_name
ndf_symbol_from = self._symbol_table.symbol_from_tag(
"ndf_{0}".format(fs_from_name),
fs=fs_from_name,
symbol_type=lfric_psyir.NumberOfDofsDataSymbol,
interface=self._read_access)
fs_to_name = arg.function_space_to.orig_name
ndf_symbol_to = self._symbol_table.symbol_from_tag(
"ndf_{0}".format(fs_to_name),
fs=fs_to_name,
symbol_type=lfric_psyir.NumberOfDofsDataSymbol,
interface=self._read_access)
ncells = lfric_psyir.NumberOfCellsDataSymbol(
"ncell_3d", interface=self._read_access)
self._symbol_table.add(ncells)
self._arglist.append(ncells)
op_arg_symbol = self._symbol_table.symbol_from_tag(
arg.name,
symbol_type=lfric_psyir.OperatorDataSymbol,
dims=[
Reference(ndf_symbol_from),
Reference(ndf_symbol_to),
Reference(ncells)
],
fs_from=fs_from_name,
fs_to=fs_to_name,
interface=ArgumentInterface(INTENT_MAPPING[arg.intent]))
self._arglist.append(op_arg_symbol)
def cma_operator(self, arg, var_accesses=None):
'''Not implemented.
:param arg: the CMA operator argument.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("cma_operator not implemented")
def scalar(self, scalar_arg, var_accesses=None):
'''Create an LFRic scalar argument and add it to the symbol table and
argument list.
:param scalar_arg: the scalar to add.
:type scalar_arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: if the datatype of the scalar is \
not supported.
'''
mapping = {
"integer": lfric_psyir.LfricIntegerScalarDataSymbol,
"real": lfric_psyir.LfricRealScalarDataSymbol,
"logical": lfric_psyir.LfricLogicalScalarDataSymbol
}
try:
symbol = self._symbol_table.symbol_from_tag(
scalar_arg.name,
symbol_type=mapping[scalar_arg.intrinsic_type],
interface=ArgumentInterface(INTENT_MAPPING[scalar_arg.intent]))
except KeyError as info:
message = (
"scalar of type '{0}' not implemented in KernelInterface "
"class.".format(scalar_arg.intrinsic_type))
six.raise_from(NotImplementedError(message), info)
self._arglist.append(symbol)
def fs_common(self, function_space, var_accesses=None):
'''Create any arguments that are common to a particular function
space. At this time the only common argument is the number of
degrees of freedom. Create the associated LFRic symbol, and
add it to the symbol table and argument list.
:param function_space: the function space for any common arguments.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
fs_name = function_space.orig_name
ndf_symbol = self._symbol_table.symbol_from_tag(
"ndf_{0}".format(fs_name),
fs=fs_name,
symbol_type=lfric_psyir.NumberOfDofsDataSymbol,
interface=self._read_access)
self._arglist.append(ndf_symbol)
def fs_intergrid(self, function_space, var_accesses=None):
'''Not implemented.
:param arg: the CMA operator argument.
:type arg: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("fs_intergrid not implemented")
def fs_compulsory_field(self, function_space, var_accesses=None):
'''Create any arguments that are compulsory for a field on a
particular function space. At this time the compulsory
arguments are the unique number of degrees of freedom and the
dofmap. Create the associated LFRic symbol, and add it to the
symbol table and argument list. Also declare the number of
degrees of freedom and add to the symbol table if one has not
yet been added.
:param function_space: the function space for any compulsory arguments.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
fs_name = function_space.orig_name
undf_symbol = self._symbol_table.symbol_from_tag(
"undf_{0}".format(fs_name),
fs=fs_name,
symbol_type=lfric_psyir.NumberOfUniqueDofsDataSymbol,
interface=self._read_access)
self._arglist.append(undf_symbol)
fs_name = function_space.orig_name
ndf_symbol = self._symbol_table.symbol_from_tag(
"ndf_{0}".format(fs_name),
fs=fs_name,
symbol_type=lfric_psyir.NumberOfDofsDataSymbol,
interface=self._read_access)
dofmap_symbol = self._symbol_table.symbol_from_tag(
"dofmap_{0}".format(fs_name),
fs=fs_name,
symbol_type=lfric_psyir.DofMapDataSymbol,
dims=[Reference(ndf_symbol)],
interface=self._read_access)
self._arglist.append(dofmap_symbol)
def banded_dofmap(self, function_space, var_accesses=None):
'''Not implemented.
:param function_space: the function space for this dofmap.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("banded_dofmap not implemented")
def indirection_dofmap(self,
function_space,
operator=None,
var_accesses=None):
'''Not implemented.
:param function_space: the function space for this dofmap.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param operator: the CMA operator.
:type operator: :py:class:`psyclone.dynamo0p3.DynKernelArgument`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("indirection_dofmap not implemented")
def basis(self, function_space, var_accesses=None):
'''Create an LFRic basis function argument and add it to the symbol
table and argument list.
:param function_space: the function space for this basis function.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
basis_name_func = function_space.get_basis_name
# This import must be placed here to avoid circular dependencies
# pylint: disable=import-outside-toplevel
from psyclone.dynamo0p3 import DynBasisFunctions
first_dim_value_func = DynBasisFunctions.basis_first_dim_value
self._create_basis(function_space, self.basis_mapping, basis_name_func,
first_dim_value_func)
def diff_basis(self, function_space, var_accesses=None):
'''Create an LFRic differential basis function argument and add it to
the symbol table and argument list.
:param function_space: the function space for this \
differential basis function.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
basis_name_func = function_space.get_diff_basis_name
# This import must be placed here to avoid circular dependencies
# pylint: disable=import-outside-toplevel
from psyclone.dynamo0p3 import DynBasisFunctions
first_dim_value_func = DynBasisFunctions.diff_basis_first_dim_value
self._create_basis(function_space, self.diff_basis_mapping,
basis_name_func, first_dim_value_func)
def field_bcs_kernel(self, function_space, var_accesses=None):
'''Not implemented.
:param function_space: the function space for this boundary condition.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("field_bcs_kernel not implemented")
def operator_bcs_kernel(self, function_space, var_accesses=None):
'''Not implemented.
:param function_space: the function space for this bcs kernel
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises NotImplementedError: as this method is not implemented.
'''
raise NotImplementedError("operator_bcs_kernel not implemented")
def ref_element_properties(self, var_accesses=None):
''' Properties associated with the reference element
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
# This callback does not contribute any kernel arguments
def mesh_properties(self, var_accesses=None):
''' Properties associated with the mesh
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
'''
# This callback does not contribute any kernel arguments
def quad_rule(self, var_accesses=None):
'''Create LFRic arguments associated with the required quadrature, if
they do not already exist, and add them to the symbol table
and argument list. The arguments depend on the type of
quadrature requested.
:param var_accesses: an unused optional argument that stores \
information about variable accesses.
:type var_accesses: :\
py:class:`psyclone.core.access_info.VariablesAccessInfo`
:raises InternalError: if an unsupported quadrature shape is \
found.
'''
# The kernel captures all the required quadrature shapes
for shape in self._kern.qr_rules:
if shape == "gh_quadrature_xyoz":
nqp_xy = self._symbol_table.symbol_from_tag(
"nqp_xy",
symbol_type=lfric_psyir.NumberOfQrPointsInXyDataSymbol,
interface=self._read_access)
nqp_z = self._symbol_table.symbol_from_tag(
"nqp_z",
symbol_type=lfric_psyir.NumberOfQrPointsInZDataSymbol,
interface=self._read_access)
weights_xy = self._symbol_table.symbol_from_tag(
"weights_xy",
symbol_type=lfric_psyir.QrWeightsInXyDataSymbol,
dims=[Reference(nqp_xy)],
interface=self._read_access)
weights_z = self._symbol_table.symbol_from_tag(
"weights_z",
symbol_type=lfric_psyir.QrWeightsInZDataSymbol,
dims=[Reference(nqp_z)],
interface=self._read_access)
self._arglist.extend([nqp_xy, nqp_z, weights_xy, weights_z])
elif shape == "gh_quadrature_face":
nfaces = self._symbol_table.symbol_from_tag(
"nfaces",
symbol_type=lfric_psyir.NumberOfFacesDataSymbol,
interface=self._read_access)
nqp = self._symbol_table.symbol_from_tag(
"nqp_faces",
symbol_type=lfric_psyir.NumberOfQrPointsInFacesDataSymbol,
interface=self._read_access)
weights = self._symbol_table.symbol_from_tag(
"weights_faces",
symbol_type=lfric_psyir.QrWeightsInFacesDataSymbol,
dims=[Reference(nqp)],
interface=self._read_access)
self._arglist.extend([nfaces, nqp, weights])
elif shape == "gh_quadrature_edge":
nedges = self._symbol_table.symbol_from_tag(
"nedges",
symbol_type=lfric_psyir.NumberOfEdgesDataSymbol,
interface=self._read_access)
nqp = self._symbol_table.symbol_from_tag(
"nqp_edges",
symbol_type=lfric_psyir.NumberOfQrPointsInEdgesDataSymbol,
interface=self._read_access)
weights = self._symbol_table.symbol_from_tag(
"weights_edges",
symbol_type=lfric_psyir.QrWeightsInEdgesDataSymbol,
dims=[Reference(nqp)],
interface=self._read_access)
self._arglist.extend([nedges, nqp, weights])
else:
raise InternalError("Unsupported quadrature shape '{0}' "
"found in kernel_interface.".format(shape))
def _create_basis(self, function_space, mapping, basis_name_func,
first_dim_value_func):
'''Internal utility to create an LFRic basis or differential basis
function argument specific to the particular quadrature that
is being used and add it to the symbol table and argument
list. Also declare the associated "ndf" symbol and any
quadrature-specific symbols if they have not already been
declared so that they can be used to dimension the basis or
differential basis symbol.
This utility function is used to avoid code replication as the
structure of a basis function is very similar to the structure
of a differential basis function.
:param function_space: the function space that this basis or \
differential basis function is on.
:type function_space: :py:class:`psyclone.domain.lfric.FunctionSpace`
:param dict mapping: a mapping from quadrature type to basis \
or differential basis class name.
:param method basis_name_func: a method that returns the name \
of the basis or differential basis function for the \
current function space.
:param function first_dim_value_func: a function that returns \
the size of the first dimension of the basis or \
differential basis function for the current function \
space.
:raises NotImplementedError: if an evaluator shape is found \
that is not a quadrature shape (currently just \
'gh_evaluator').
:raises InternalError: if the supplied evaluator shape is not \
recognised.
'''
# pylint: disable=too-many-locals
# This import must be placed here to avoid circular dependencies
# pylint: disable=import-outside-toplevel
from psyclone.dynamo0p3 import VALID_EVALUATOR_SHAPES
for shape in self._kern.eval_shapes:
fs_name = function_space.orig_name
ndf_symbol = self._symbol_table.symbol_from_tag(
"ndf_{0}".format(fs_name),
symbol_type=lfric_psyir.NumberOfDofsDataSymbol,
fs=fs_name,
interface=self._read_access)
# Create the qr tag by appending the last part of the shape
# name to "qr_".
quad_name = shape.split("_")[-1]
basis_tag = basis_name_func(qr_var="qr_" + quad_name)
if shape == "gh_quadrature_xyoz":
nqp_xy = self._symbol_table.symbol_from_tag(
"nqp_xy",
symbol_type=lfric_psyir.NumberOfQrPointsInXyDataSymbol,
interface=self._read_access)
nqp_z = self._symbol_table.symbol_from_tag(
"nqp_z",
symbol_type=lfric_psyir.NumberOfQrPointsInZDataSymbol,
interface=self._read_access)
arg = mapping["gh_quadrature_xyoz"](
basis_tag, [
int(first_dim_value_func(function_space)),
Reference(ndf_symbol),
Reference(nqp_xy),
Reference(nqp_z)
],
fs_name,
interface=self._read_access)
elif shape == "gh_quadrature_face":
nfaces = self._symbol_table.symbol_from_tag(
"nfaces",
symbol_type=lfric_psyir.NumberOfFacesDataSymbol,
interface=self._read_access)
nqp = self._symbol_table.symbol_from_tag(
"nqp_faces",
symbol_type=lfric_psyir.NumberOfQrPointsInFacesDataSymbol,
interface=self._read_access)
arg = mapping["gh_quadrature_face"](
basis_tag, [
int(first_dim_value_func(function_space)),
Reference(ndf_symbol),
Reference(nqp),
Reference(nfaces)
],
fs_name,
interface=self._read_access)
elif shape == "gh_quadrature_edge":
nedges = self._symbol_table.symbol_from_tag(
"nedges",
symbol_type=lfric_psyir.NumberOfEdgesDataSymbol,
interface=self._read_access)
nqp = self._symbol_table.symbol_from_tag(
"nqp_edges",
symbol_type=lfric_psyir.NumberOfQrPointsInEdgesDataSymbol,
interface=self._read_access)
arg = mapping["gh_quadrature_edge"](
basis_tag, [
int(first_dim_value_func(function_space)),
Reference(ndf_symbol),
Reference(nqp),
Reference(nedges)
],
fs_name,
interface=self._read_access)
elif shape in VALID_EVALUATOR_SHAPES:
# Need a (diff) basis array for each target space upon
# which the basis functions have been
# evaluated. _kern.eval_targets is a dict where the
# values are 2-tuples of (FunctionSpace, argument).
for _, _ in self._kern.eval_targets.items():
raise NotImplementedError(
"Evaluator shapes not implemented in kernel_interface "
"class.")
else:
raise InternalError(
"Unrecognised quadrature or evaluator shape '{0}'. "
"Expected one of: {1}.".format(shape,
VALID_EVALUATOR_SHAPES))
self._symbol_table.add(arg)
self._arglist.append(arg)
