def override_log(key, old, new):
logging.info(
'Replacing {} for test type: {}. Old value:'
' ({}), New value: ({})'.format(
key,
stringify_args(
[ttype, test['eval-type'], jtype, stype],
joiner='.'), stringify_args(listify(old)),
stringify_args(listify(new))))
def _raise(desc, inp, nameref, shape):
logger = logging.getLogger(__name__)
logger.debug('{} array for driver kernel {} does not '
'match expected shape (from array {}). '
'Expected: ({}), got: ({})'.format(
desc, inp.name, nameref,
stringify_args(inp.shape),
stringify_args(shape))
)
raise InvalidInputSpecificationException(inp.name)
def __init__(self, bad_inputs):
from pyjac.utils import stringify_args, listify
self.message = (
'Inputs: ({}) were incorrectly, or conflictingly specified. '
'See debug output for more information'.format(
stringify_args(listify(bad_inputs))))
super(InvalidInputSpecificationException, self).__init__(self.message)
def __internal_validator(self,
field,
valuelist,
valid,
message,
necessary=True):
valuelist = listify(valuelist)
if six.callable(valid):
badvals = [x for x in valuelist if not valid(x)]
else:
badvals = [x for x in valuelist if x not in valid]
if badvals and necessary:
args = (badvals, )
if not six.callable(valid):
args = (badvals, valid)
self._error(
field, message.format(*tuple(stringify_args(x) for x in args)))
def test_lockstep_driver(self):
# get rate info
rate_info = determine_jac_inds(self.store.reacs, self.store.specs,
RateSpecialization.fixed)
mod_test = get_run_source()
for kind, loopy_opts in OptionLoopWrapper.from_get_oploop(
self, do_ratespec=False, langs=get_test_langs(),
do_vector=True, yield_index=True):
# make namestore
namestore = arc.NameStore(loopy_opts, rate_info)
# kernel 1 - need the jacobian reset kernel
reset = reset_arrays(loopy_opts, namestore)
# kernel 2 - incrementer
# make mapstore, arrays and kernel info
mapstore = arc.MapStore(loopy_opts, namestore.phi_inds, None)
# use arrays of 2 & 3 dimensions to test the driver's copying
base_phi_shape = namestore.n_arr.shape
P_lp, P_str = mapstore.apply_maps(namestore.P_arr,
arc.global_ind)
phi_lp, phi_str = mapstore.apply_maps(namestore.n_arr,
arc.global_ind,
arc.var_name)
inputs = [P_lp.name, phi_lp.name]
base_jac_shape = namestore.jac.shape
jac_lp, jac_str = mapstore.apply_maps(namestore.jac,
arc.global_ind,
arc.var_name,
arc.var_name)
outputs = [jac_lp.name]
kernel_data = [P_lp, phi_lp, jac_lp]
kernel_data.extend(arc.initial_condition_dimension_vars(
loopy_opts, None))
instructions = Template("""
${phi_str} = ${phi_str} + ${P_str} {id=0, dep=*}
${jac_str} = ${jac_str} + ${phi_str} {id=1, dep=0, nosync=0}
""").safe_substitute(**locals())
# handle atomicity
can_vec, vec_spec = ic.get_deep_specializer(
loopy_opts, atomic_ids=['1'])
barriers = []
if loopy_opts.depth:
# need a barrier between the reset & the kernel
barriers = [(0, 1, 'global')]
inner_kernel = k_gen.knl_info(
name='inner',
instructions=instructions,
mapstore=mapstore,
var_name=arc.var_name,
kernel_data=kernel_data,
silenced_warnings=['write_race(0)', 'write_race(1)'],
can_vectorize=can_vec,
vectorization_specializer=vec_spec)
# put it in a generator
generator = k_gen.make_kernel_generator(
loopy_opts, kernel_type=KernelType.dummy,
name='inner_kernel', kernels=[reset, inner_kernel],
namestore=namestore,
input_arrays=inputs[:],
output_arrays=outputs[:],
is_validation=True,
driver_type=DriverType.lockstep,
barriers=barriers)
# use a "weird" (non-evenly divisibly by vector width) test-size to
# properly test the copy-in / copy-out
test_size = self.store.test_size - 37
if test_size <= 0:
test_size = self.store.test_size - 1
assert test_size > 0
# and make
with temporary_build_dirs() as (build, obj, lib):
numpy_arrays = []
def __save(shape, name, zero=False):
data = np.zeros(shape)
if not zero:
# make it a simple range
data.flat[:] = np.arange(np.prod(shape))
# save
myname = pjoin(lib, name + '.npy')
# need to split inputs / answer
np.save(myname, data.flatten('K'))
numpy_arrays.append(data.flatten('K'))
# write 'data'
import loopy as lp
for arr in kernel_data:
if not isinstance(arr, lp.ValueArg):
__save((test_size,) + arr.shape[1:], arr.name,
arr.name in outputs)
# and a parameter
param = np.zeros((test_size,))
param[:] = np.arange(test_size)
# build code
generator.generate(build,
data_order=loopy_opts.order,
data_filename='data.bin',
for_validation=True)
# write header
write_aux(build, loopy_opts, self.store.specs, self.store.reacs)
# generate wrapper
pywrap(loopy_opts.lang, build,
obj_dir=obj, out_dir=lib,
ktype=KernelType.dummy,
file_base=generator.name,
additional_inputs=inputs[:],
additional_outputs=outputs[:])
# and calling script
test = pjoin(lib, 'test.py')
inputs = utils.stringify_args(
[pjoin(lib, inp + '.npy') for inp in inputs], use_quotes=True)
str_outputs = utils.stringify_args(
[pjoin(lib, inp + '.npy') for inp in outputs], use_quotes=True)
num_threads = _get_test_input(
'num_threads', psutil.cpu_count(logical=False))
with open(test, 'w') as file:
file.write(mod_test.safe_substitute(
package='pyjac_{lang}'.format(
lang=utils.package_lang[loopy_opts.lang]),
input_args=inputs,
test_arrays=str_outputs,
output_files=str_outputs,
looser_tols='[]',
loose_rtol=0,
loose_atol=0,
rtol=0,
atol=0,
non_array_args='{}, {}'.format(
test_size, num_threads),
kernel_name=generator.name.title(),))
try:
utils.run_with_our_python([test])
except subprocess.CalledProcessError:
logger = logging.getLogger(__name__)
logger.debug(utils.stringify_args(vars(loopy_opts), kwd=True))
assert False, 'lockstep_driver error'
# calculate answers
ns = base_jac_shape[1]
# pressure is added to phi
phi = numpy_arrays[1].reshape((test_size, ns),
order=loopy_opts.order)
p_arr = numpy_arrays[0]
phi = phi + p_arr[:, np.newaxis]
jac = numpy_arrays[2].reshape((test_size, ns, ns),
order=loopy_opts.order)
# and the diagonal of the jacobian has the updated pressure added
jac[:, range(ns), range(ns)] += phi[:, range(ns)]
# and read in outputs
test = np.load(pjoin(lib, outputs[0] + '.npy')).reshape(
jac.shape, order=loopy_opts.order)
assert np.array_equal(test, jac)
def __init__(self, otype, value, allowed):
from pyjac.utils import stringify_args
self.message = ('Value "{}" for override type "{}" is not allowed. '
'Allowed values are: {}'.format(
otype, value, stringify_args(allowed)))
super(InvalidOverrideException, self).__init__(self.message)
def get_test_matrix(work_dir,
test_type,
test_matrix,
for_validation,
raise_on_missing=True,
langs=get_test_langs()):
"""Runs a set of mechanisms and an ordered dictionary for
performance and functional testing
Parameters
----------
work_dir : str
Working directory with mechanisms and for data
test_type: :class:`build_type.jacobian`
Controls some testing options (e.g., whether to do a sparse matrix or not)
test_matrix: str
The test matrix file to load
for_validation: bool
If determines which test type to load from the test matrix,
validation or performance
raise_on_missing: bool
Raise an exception of the specified :param:`test_matrix` file is not found
langs: list of str
The allowed languages, modifiable by the :envvar:`TEST_LANGS` or test_langs
in :file:`test_setup.py`
Returns
-------
mechanisms : dict
A dictionary indicating which mechanism are available for testing,
The structure is as follows:
mech_name : {'mech' : file path to the Cantera mechanism
'ns' : number of species in the mechanism
'limits' : {'full': XXX, 'sparse': XXX}}: a dictionary of
limits on the number of conditions that can be evaluated
for this mechanism (full & sparse jacobian respectively)
due to memory constraints
params : OrderedDict
The parameters to put in an oploop
max_vec_width : int
The maximum vector width to test
"""
work_dir = abspath(work_dir)
# validate the test matrix
matrix_name = test_matrix
test_matrix = build_and_validate('test_matrix_schema.yaml', test_matrix)
# check that we have the working directory
if not exists(work_dir):
raise Exception('Work directory {} for '.format(work_dir) +
'testing not found, exiting...')
# load the models
models = load_models(work_dir, test_matrix)
assert isinstance(test_type, build_type)
# load tests
tests = load_tests(test_matrix, matrix_name)
# filter those that match the test type
valid_str = 'validation' if for_validation else 'performance'
tests = [test for test in tests if test['test-type'] == valid_str]
tests = [
test for test in tests
if test['eval-type'] == enum_to_string(test_type)
or test['eval-type'] == 'both'
]
# and dictify
tests = [OrderedDict(test) for test in tests]
if not tests:
raise Exception('No tests found in matrix {} for {} test of {}, '
'exiting...'.format(matrix_name, valid_str,
enum_to_string(test_type)))
# get defaults we haven't migrated to schema yet
rate_spec = ['fixed', 'hybrid'] if test_type != build_type.jacobian \
else ['fixed']
sparse = ([
enum_to_string(JacobianFormat.sparse),
enum_to_string(JacobianFormat.full)
] if test_type == build_type.jacobian else
[enum_to_string(JacobianFormat.full)])
jac_types = [
enum_to_string(JacobianType.exact),
enum_to_string(JacobianType.finite_difference)
] if (test_type == build_type.jacobian
and not for_validation) else [enum_to_string(JacobianType.exact)]
split_kernels = [False]
# and default # of cores, this may be overriden
default_num_cores, can_override_cores = num_cores_default()
# load platforms
platforms = load_platforms(test_matrix,
langs=langs,
raise_on_empty=raise_on_missing)
platforms = [OrderedDict(platform) for platform in platforms]
out_params = []
logger = logging.getLogger(__name__)
for test in tests:
# filter platforms
plats = [p.copy() for p in platforms]
if 'platforms' in test:
plats = [
plat for plat in plats if plat['platform'] in test['platforms']
]
if len(plats) < len(platforms):
logger.info(
'Platforms ({}) filtered out for test type: {}'.format(
', '.join([
p['platform'] for p in platforms if p not in plats
]), ' - '.join([test['test-type'],
test['eval-type']])))
if not len(plats):
logger.warn('No platforms found for test {}, skipping...'.format(
' - '.join([test['test-type'], test['eval-type']])))
continue
for plookup in plats:
clean = plookup.copy()
# get default number of cores
cores = default_num_cores[:]
# get default vector widths
widths = plookup['width']
is_wide = widths is not None
depths = plookup['depth']
is_deep = depths is not None
if is_deep and not is_wide:
widths = depths[:]
# sanity check
if is_wide or is_deep:
assert widths is not None
# special gpu handling for cores
is_gpu = False
# test platform type
if platform_is_gpu(plookup['platform']):
# set cores to 1
is_gpu = True
cores = [1]
def apply_vectypes(lookup,
widths,
is_wide=is_wide,
is_deep=is_deep):
if is_wide or is_deep:
# set vec widths
use_par = None in widths or (is_wide and is_deep)
lookup['vecsize'] = [x for x in widths[:] if x is not None]
base = [True] if not use_par else [True, False]
if is_wide:
lookup['wide'] = base[:]
base.pop()
if is_deep:
lookup['deep'] = base[:]
else:
lookup['vecsize'] = [None]
lookup['wide'] = [False]
lookup['deep'] = [False]
del lookup['width']
del lookup['depth']
apply_vectypes(plookup, widths)
# default is both conp / conv
conp = [True, False]
order = ['C', 'F']
# loop over possible overrides
oploop = OptionLoop(
OrderedDict([('ttype', [enum_to_string(test_type)]),
('jtype', jac_types), ('stype', sparse)]))
for i, state in enumerate(oploop):
ttype = state['ttype']
jtype = state['jtype']
stype = state['stype']
def override_log(key, old, new):
logging.info(
'Replacing {} for test type: {}. Old value:'
' ({}), New value: ({})'.format(
key,
stringify_args(
[ttype, test['eval-type'], jtype, stype],
joiner='.'), stringify_args(listify(old)),
stringify_args(listify(new))))
# copy defaults
icores = cores[:]
iorder = order[:]
iconp = conp[:]
ivecsizes = widths[:] if widths is not None else [None]
imodels = tuple(models.keys())
# load overides
overrides = get_overrides(test, ttype, jtype, stype)
# check that we can apply
if 'num_cores' in overrides and not can_override_cores:
raise InvalidTestEnivironmentException(
ttype, 'num_cores', matrix_name, 'num_threads')
elif 'num_cores' in overrides and is_gpu:
logger = logging.getLogger(__name__)
logger.info(
'Discarding unused "num_cores" override for GPU '
'platform {}'.format(plookup['platform']))
del overrides['num_cores']
# 'num_cores', 'order', 'conp', 'vecsize', 'vectype'
# now apply overrides
outplat = plookup.copy()
for current in overrides:
ivectypes_override = None
for override in overrides:
if override == 'num_cores':
override_log('num_cores', icores,
overrides[override])
icores = overrides[override]
elif override == 'order' and not is_gpu:
override_log('order', iorder, overrides[override])
iorder = overrides[override]
elif override == 'gpuorder' and is_gpu:
override_log('order', iorder, overrides[override])
iorder = overrides[override]
elif override == 'conp':
iconp_save = iconp[:]
iconp = []
if 'conp' in overrides[override]:
iconp.append(True)
if 'conv' in overrides[override]:
iconp.append(False)
override_log('conp', iconp_save, iconp)
elif override == 'vecsize' and not is_gpu:
override_log('vecsize', ivecsizes,
overrides[override])
outplat['vecsize'] = listify(overrides[override])
elif override == 'gpuvecsize' and is_gpu:
override_log('gpuvecsize', ivecsizes,
overrides[override])
outplat['vecsize'] = listify(overrides[override])
elif override == 'vectype' and not is_gpu:
# we have to do this at the end
ivectypes_override = overrides[override]
elif override == 'gpuvectype' and is_gpu:
ivectypes_override = overrides[override]
elif override == 'models':
# check that all models are valid
for model in overrides[override]:
if model not in imodels:
raise InvalidOverrideException(
override, model, imodels)
# and replace
override_log('models', stringify_args(imodels),
stringify_args(overrides[override]))
imodels = tuple(overrides[override])
if ivectypes_override is not None:
c = clean.copy()
apply_vectypes(c,
outplat['vecsize'],
is_wide='wide' in ivectypes_override,
is_deep='deep' in ivectypes_override)
# and copy into working
outplat['wide'] = c['wide'] if 'wide' in c else [False]
outplat['deep'] = c['deep'] if 'deep' in c else [False]
outplat['vecsize'] = c['vecsize']
old = ['']
if is_wide:
old += ['wide']
if is_deep:
old += ['deep']
elif not is_wide:
old += ['par']
override_log('vecsize', old, ivectypes_override)
# and finally, convert back to an option loop format
out_params.append(
[('num_cores', icores), ('order',
iorder), ('rate_spec', rate_spec),
('split_kernels', split_kernels), ('conp', iconp),
('sparse', [stype]), ('jac_type',
[jtype]), ('models', [imodels])] +
[(key, value) for key, value in six.iteritems(outplat)])
max_vec_width = 1
vector_params = [
dict(p)['vecsize'] for p in out_params
if 'vecsize' in dict(p) and dict(p)['vecsize'] != [None]
]
if vector_params:
max_vec_width = max(max_vec_width,
max([max(x) for x in vector_params]))
from . import reduce_oploop
loop = reduce_oploop(out_params)
return models, loop, max_vec_width
def get_driver(loopy_opts, namestore, inputs, outputs, driven,
test_size=None):
"""
Implements a driver function for kernel evaluation.
This allows pyJac to utilize a smaller working-buffer (sized to the
global work size), and implements a static(like) scheduling algorithm
Notes
-----
Currently Loopy doesn't have the machinery to enable native calling of other
loopy kernels, so we have to fudge this a bit (and this can't be used for
unit-tests). Future versions will allow us to natively wrap test functions
(i.e., once the new function calling interface is in place in Loopy)
:see:`driver-function` for more information
Parameters
----------
loopy_opts: :class:`loopy_options`
The loopy options specifying how to create this kernel
namestore: :class:`NameStore`
The namestore class that owns our arrays
inputs: list of :class:`lp.KernelArgument`
The arrays that should be copied into internal working buffers
before calling subfunctions
outputs: list of :class:`lp.KernelArgument`
The arrays should be copied back into global memory after calling
subfunctions
driven: :class:`kernel_generator`
The kernel generator to wrap in the driver
Returns
-------
knl_list : list of :class:`knl_info`
The generated infos for feeding into the kernel generator
"""
# we have to do some shennanigains here to get this to work in loopy:
#
# 1. Loopy currently doesn't allow you to alter the for-loop increment size,
# so for OpenCL where we must increment by the global work size, we have to
# put a dummy for-loop in, and teach the kernel generator to work around it
#
# 2. Additionally, the OpenMP target in Loopy is Coming Soon (TM), hence we need
# our own dynamic scheduling preamble for the driver loop (
# if we're operating in queue-form)
#
# 3. Finally, Loopy is just now supporting the ability to natively call other
# kernels, so for the moment we still need to utilize the dummy function
# calling we have set-up for the finite difference Jacobian
# first, get our input / output arrays
arrays = {}
to_find = set(listify(inputs)) | set(listify(outputs))
# create mapping of array names
array_names = {v.name: v for k, v in six.iteritems(vars(namestore))
if isinstance(v, arc.creator) and not (
v.fixed_indicies or v.affine)}
for arr in to_find:
arr_creator = next((array_names[x] for x in array_names if x == arr), None)
if arr_creator is None:
continue
arrays[arr] = arr_creator
if len(arrays) != len(to_find):
missing = to_find - set(arrays.keys())
logger = logging.getLogger(__name__)
logger.debug('Input/output arrays for queue_driver kernel {} not found.'
.format(stringify_args(missing)))
raise InvalidInputSpecificationException(missing)
def arr_non_ic(array_input):
return len(array_input.shape) > 1
# ensure the inputs and output are all identically sized (among those that have)
# a non-initial condition dimension
def __check(check_input):
shape = ()
def _raise(desc, inp, nameref, shape):
logger = logging.getLogger(__name__)
logger.debug('{} array for driver kernel {} does not '
'match expected shape (from array {}). '
'Expected: ({}), got: ({})'.format(
desc, inp.name, nameref,
stringify_args(inp.shape),
stringify_args(shape))
)
raise InvalidInputSpecificationException(inp.name)
nameref = None
desc = 'Input' if check_input else 'Output'
for inp in [arrays[x] for x in (inputs if check_input else outputs)]:
if not arr_non_ic(inp):
# only the initial condition dimension, fine
continue
if shape:
if inp.shape != shape and len(inp.shape) == len(shape):
# allow different shapes in the last index
if not all(x == y for x, y in zip(*(
inp.shape[:-1], shape[:-1]))):
_raise(desc, inp, nameref, shape)
# otherwise, take the maximum of the shape entry
shape = shape[:-1] + (max(shape[-1], inp.shape[-1]),)
elif inp.shape != shape:
_raise(desc, inp, nameref, shape)
else:
nameref = inp.name
shape = inp.shape[:]
if not shape:
logger = logging.getLogger(__name__)
logger.debug('No {} arrays supplied to driver that require '
'copying to working buffer!'.format(desc))
raise InvalidInputSpecificationException('Driver ' + desc + ' arrays')
return shape
def create_interior_kernel(for_input):
shape = __check(for_input)
name = 'copy_{}'.format('in' if for_input else 'out')
# get arrays
arrs = [arrays[x] for x in (inputs if for_input else outputs)]
# create a dummy map and store
map_shape = np.arange(shape[1], dtype=arc.kint_type)
mapper = arc.creator(name, arc.kint_type, map_shape.shape, 'C',
initializer=map_shape)
mapstore = arc.MapStore(loopy_opts, mapper, test_size)
# determine what other inames we need, if any
namer = UniqueNameGenerator(set([mapstore.iname]))
extra_inames = []
for i in six.moves.range(2, len(shape)):
iname = namer(mapstore.iname)
extra_inames.append((iname, '0 <= {} < {}'.format(
iname, shape[i])))
indicies = [arc.global_ind, mapstore.iname] + [
ex[0] for ex in extra_inames]
global_indicies = indicies[:]
global_indicies[0] += ' + ' + driver_offset.name
# bake in SIMD pre-split
vec_spec = None
split_spec = None
conditional_index = get_problem_index(loopy_opts)
def __build(arr, local, **kwargs):
inds = global_indicies if not local else indicies
if isinstance(arr, arc.jac_creator) and arr.is_sparse:
# this is a sparse Jacobian, hence we have to override the default
# indexing (as we're doing a straight copy)
kwargs['ignore_lookups'] = True
if arr_non_ic(arr):
return mapstore.apply_maps(arr, *inds, **kwargs)
else:
return mapstore.apply_maps(arr, inds[0], **kwargs)
# create working buffer version of arrays
working_buffers = []
working_strs = []
for arr in arrs:
arr_lp, arr_str = __build(arr, True, use_local_name=True)
working_buffers.append(arr_lp)
working_strs.append(arr_str)
# create global versions of arrays
buffers = []
strs = []
for arr in arrs:
arr_lp, arr_str = __build(arr, False, reshape_to_working_buffer=False)
buffers.append(arr_lp)
strs.append(arr_str)
# now create the instructions
instruction_template = Template("""
if ${ind} < ${problem_size} ${shape_check}
${local_buffer} = ${global_buffer} {id=copy_${name}}
end
""") if for_input else Template("""
if ${ind} < ${problem_size} ${shape_check}
${global_buffer} = ${local_buffer} {id=copy_${name}}
end
""")
warnings = []
instructions = []
for i, arr in enumerate(arrs):
# get shape check
shape_check = ''
if arr.shape[-1] != shape[-1] and len(arr.shape) == len(shape):
shape_check = ' and {} < {}'.format(
indicies[-1], arr.shape[-1])
instructions.append(instruction_template.substitute(
local_buffer=working_strs[i],
global_buffer=strs[i],
ind=conditional_index,
problem_size=arc.problem_size.name,
name=arr.name,
shape_check=shape_check))
warnings.append('write_race(copy_{})'.format(arr.name))
if loopy_opts.is_simd:
warnings.append('vectorize_failed')
warnings.append('unrolled_vector_iname_conditional')
instructions = '\n'.join(instructions)
kwargs = {}
if loopy_opts.lang == 'c':
# override the number of copies in this function to 1
# (i.e., 1 per-thread)
kwargs['iname_domain_override'] = [(arc.global_ind, '0 <= {} < 1'.format(
arc.global_ind))]
priorities = ([arc.global_ind + '_outer'] if loopy_opts.pre_split else [
arc.global_ind]) + [arc.var_name]
# and return the kernel info
return knl_info(name=name,
instructions=instructions,
mapstore=mapstore,
var_name=arc.var_name,
extra_inames=extra_inames,
kernel_data=buffers + working_buffers + [
arc.work_size, arc.problem_size, driver_offset],
silenced_warnings=warnings,
vectorization_specializer=vec_spec,
split_specializer=split_spec,
unrolled_vector=True,
loop_priority=set([tuple(priorities + [
iname[0] for iname in extra_inames])]),
**kwargs)
copy_in = create_interior_kernel(True)
# create a dummy kernel info that simply calls our internal function
instructions = driven.name + '()'
# create mapstore
call_name = driven.name
repeats = 1
if loopy_opts.depth:
# we need 'var_name' to have a non-unity size
repeats = loopy_opts.vector_width
map_shape = np.arange(repeats, dtype=arc.kint_type)
mapper = arc.creator(call_name, arc.kint_type, map_shape.shape, 'C',
initializer=map_shape)
mapstore = arc.MapStore(loopy_opts, mapper, test_size)
mangler = lp_pregen.MangleGen(call_name, tuple(), tuple())
kwargs = {}
if loopy_opts.lang == 'c':
# override the number of calls to the driven function in the driver, this
# is currently fixed to 1 (i.e., 1 per-thread)
kwargs['iname_domain_override'] = [(arc.global_ind, '0 <= {} < 1'.format(
arc.global_ind))]
func_call = knl_info(name='driver',
instructions=instructions,
mapstore=mapstore,
kernel_data=[arc.work_size, arc.problem_size],
var_name=arc.var_name,
extra_inames=copy_in.extra_inames[:],
manglers=[mangler],
**kwargs)
copy_out = create_interior_kernel(False)
# and return
return [copy_in, func_call, copy_out]