def test_no_parent_accdirective():
''' Test that applying Extract Transformation on an orphaned
ACCLoopDirective without its ancestor ACCParallelDirective
when optimisations are applied raises a TransformationError. '''
etrans = GOceanExtractTrans()
acclpt = ACCLoopTrans()
accpara = ACCParallelTrans()
accdata = ACCEnterDataTrans()
_, invoke = get_invoke("single_invoke_three_kernels.f90",
GOCEAN_API,
idx=0,
dist_mem=False)
schedule = invoke.schedule
# Apply the OpenACC Loop transformation to every loop in the Schedule
for child in schedule.children:
if isinstance(child, Loop):
acclpt.apply(child)
# Enclose all of these loops within a single ACC Parallel region
accpara.apply(schedule.children)
# Add a mandatory ACC enter-data directive
accdata.apply(schedule)
orphaned_directive = schedule.children[1].children[0]
with pytest.raises(TransformationError) as excinfo:
_, _ = etrans.apply(orphaned_directive)
assert "Error in GOceanExtractTrans: Application to Nodes enclosed " \
"within a thread-parallel region is not allowed." \
in str(excinfo.value)
def trans(psy):
''' Take the supplied psy object, apply OpenACC transformations
to the schedule of invoke_0 and return the new psy object '''
ptrans = ACCParallelTrans()
ltrans = ACCLoopTrans()
dtrans = ACCEnterDataTrans()
ktrans = ACCRoutineTrans()
invoke = psy.invokes.get('invoke_0_inc_field')
schedule = invoke.schedule
schedule.view()
# Apply the OpenACC Loop transformation to *every* loop
# nest in the schedule
for child in schedule.children:
if isinstance(child, Loop):
ltrans.apply(child, {"collapse": 2})
# Put all of the loops in a single parallel region
ptrans.apply(schedule.children)
# Add an enter-data directive
dtrans.apply(schedule)
# Put an 'acc routine' directive inside each kernel
for kern in schedule.coded_kernels():
ktrans.apply(kern)
# Ideally we would module-inline the kernel here (to save having to
# rely on the compiler to do it) but this does not currently work
# for the fparser2 AST (issue #229).
# _, _ = itrans.apply(kern)
schedule.view()
return psy
def test_lfric_stencil():
'''Check variable usage detection when OpenACC is used with a
kernel that uses a stencil.
'''
# Use the OpenACC transforms to create the required kernels
from psyclone.transformations import ACCParallelTrans, ACCEnterDataTrans
from psyclone.psyGen import CodedKern
acc_par_trans = ACCParallelTrans()
acc_enter_trans = ACCEnterDataTrans()
_, invoke = get_invoke("14.4_halo_vector.f90", "dynamo0.3",
idx=0, dist_mem=False)
sched = invoke.schedule
_ = acc_par_trans.apply(sched.children)
_ = acc_enter_trans.apply(sched)
# Find the first kernel:
kern = invoke.schedule.walk(CodedKern)[0]
create_acc_arg_list = KernCallAccArgList(kern)
var_accesses = VariablesAccessInfo()
create_acc_arg_list.generate(var_accesses=var_accesses)
var_info = str(var_accesses)
assert "f1: READ+WRITE" in var_info
assert "f2: READ" in var_info
assert "f2_stencil_dofmap: READ" in var_info
def test_lfric_acc_operator():
'''Check variable usage detection when OpenACC is used with
a kernel that uses an operator.
'''
# Use the OpenACC transforms to enclose the kernels
# with OpenACC directives.
from psyclone.transformations import ACCParallelTrans, ACCEnterDataTrans
from psyclone.psyGen import CodedKern
acc_par_trans = ACCParallelTrans()
acc_enter_trans = ACCEnterDataTrans()
_, invoke = get_invoke("20.0_cma_assembly.f90", "dynamo0.3",
idx=0, dist_mem=False)
sched = invoke.schedule
_ = acc_par_trans.apply(sched.children)
_ = acc_enter_trans.apply(sched)
# Find the first kernel:
kern = invoke.schedule.walk(CodedKern)[0]
create_acc_arg_list = KernCallAccArgList(kern)
var_accesses = VariablesAccessInfo()
create_acc_arg_list.generate(var_accesses=var_accesses)
var_info = str(var_accesses)
assert "lma_op1_proxy%ncell_3d: READ" in var_info
assert "lma_op1_proxy%local_stencil: WRITE" in var_info
def test_lfric_acc():
'''Check variable usage detection when OpenACC is used.
'''
# Use the OpenACC transforms to enclose the kernels
# with OpenACC directives.
from psyclone.transformations import ACCParallelTrans, ACCEnterDataTrans
from psyclone.psyGen import CodedKern
acc_par_trans = ACCParallelTrans()
acc_enter_trans = ACCEnterDataTrans()
_, invoke = get_invoke("1_single_invoke.f90", "dynamo0.3",
name="invoke_0_testkern_type", dist_mem=False)
sched = invoke.schedule
_ = acc_par_trans.apply(sched.children)
_ = acc_enter_trans.apply(sched)
# Find the first kernel:
kern = invoke.schedule.walk(CodedKern)[0]
create_acc_arg_list = KernCallAccArgList(kern)
var_accesses = VariablesAccessInfo()
create_acc_arg_list.generate(var_accesses=var_accesses)
var_info = str(var_accesses)
assert "f1: READ+WRITE" in var_info
assert "f2: READ" in var_info
assert "m1: READ" in var_info
assert "m2: READ" in var_info
assert "undf_w1: READ" in var_info
assert "map_w1: READ" in var_info
assert "undf_w2: READ" in var_info
assert "map_w2: READ" in var_info
assert "undf_w3: READ" in var_info
assert "map_w3: READ" in var_info
def test_no_parent_accdirective():
''' Test that applying Extract Transformation on an orphaned
ACCLoopDirective without its ancestor ACCParallelDirective
when optimisations are applied raises a TransformationError. '''
from psyclone.transformations import ACCParallelTrans, ACCEnterDataTrans, \
ACCLoopTrans
etrans = GOceanExtractRegionTrans()
acclpt = ACCLoopTrans()
accpara = ACCParallelTrans()
accdata = ACCEnterDataTrans()
_, invoke_info = parse(os.path.join(GOCEAN_BASE_PATH,
"single_invoke_three_kernels.f90"),
api=GOCEAN_API)
psy = PSyFactory(GOCEAN_API, distributed_memory=False).create(invoke_info)
invoke = psy.invokes.invoke_list[0]
schedule = invoke.schedule
# Apply the OpenACC Loop transformation to every loop in the Schedule
for child in schedule.children:
if isinstance(child, Loop):
schedule, _ = acclpt.apply(child)
# Enclose all of these loops within a single ACC Parallel region
schedule, _ = accpara.apply(schedule.children)
# Add a mandatory ACC enter-data directive
schedule, _ = accdata.apply(schedule)
orphaned_directive = schedule.children[1].children[0]
with pytest.raises(TransformationError) as excinfo:
_, _ = etrans.apply(orphaned_directive)
assert ("Extraction of Nodes enclosed within a thread parallel "
"region is not allowed.") in str(excinfo)
def trans(psy):
'''PSyclone transformation script for the dynamo0p3 api to apply
OpenACC loop, parallel and enter data directives generically.
:param psy: a PSyclone PSy object which captures the algorithm and \
kernel information required by PSyclone.
:type psy: subclass of :py:class:`psyclone.psyGen.PSy`
'''
kernels_trans = ACCKernelsTrans()
routine_trans = ACCRoutineTrans()
ctrans = Dynamo0p3ColourTrans()
loop_trans = ACCLoopTrans()
enter_trans = ACCEnterDataTrans()
const = LFRicConstants()
# Loop over all of the Invokes in the PSy object
for invoke in psy.invokes.invoke_list:
schedule = invoke.schedule
# Colour loops as required
for loop in schedule.loops():
if loop.field_space.orig_name \
not in const.VALID_DISCONTINUOUS_NAMES \
and loop.iteration_space == "cell_column":
ctrans.apply(loop)
# Add Kernels and Loop directives
for loop in schedule.loops():
if loop.loop_type != "colours":
kernels_trans.apply([loop])
loop_trans.apply(loop)
# Add Routine directive to kernels
for kernel in schedule.coded_kernels():
routine_trans.apply(kernel)
# Add Enter Data directive covering all of the PSy layer.
enter_trans.apply(schedule)
schedule.view()
return psy
def test_accenterdata_internalerr(monkeypatch):
''' Check that the ACCEnterDataTrans.apply() method raises an internal
error if the validate method fails to throw out an invalid type of
Schedule. '''
acct = ACCEnterDataTrans()
monkeypatch.setattr(acct, "validate", lambda sched, options: None)
with pytest.raises(InternalError) as err:
_, _ = acct.apply("Not a schedule")
assert ("validate() has not rejected an (unsupported) schedule"
in str(err.value))
def test_accenterdata_internalerr(monkeypatch):
''' Check that the ACCEnterDataTrans.apply() method raises an internal
error if the _validate method fails to throw out an invalid type of
Schedule. '''
from psyclone.transformations import ACCEnterDataTrans
from psyclone.psyGen import InternalError
acct = ACCEnterDataTrans()
monkeypatch.setattr(acct, "_validate", lambda sched: None)
with pytest.raises(InternalError) as err:
_, _ = acct.apply("Not a schedule")
assert "validate() has not rejected an (unsupported) schedule" in str(err)
def trans(psy):
''' Take the supplied psy object, apply OpenACC transformations
to the schedule of the first invoke and return the new psy object '''
ptrans = ACCParallelTrans()
ltrans = ACCLoopTrans()
dtrans = ACCEnterDataTrans()
ktrans = ACCRoutineTrans()
itrans = KernelModuleInlineTrans()
g2localtrans = KernelGlobalsToArguments()
invoke = psy.invokes.invoke_list[0]
schedule = invoke.schedule
schedule.view()
# Apply the OpenACC Loop transformation to *every* loop
# nest in the schedule
for child in schedule.children:
if isinstance(child, Loop):
ltrans.apply(child, {"collapse": 2})
# Put all of the loops in a single parallel region
ptrans.apply(schedule.children)
# Add an enter-data directive
dtrans.apply(schedule)
# Convert any accesses to global data into kernel arguments, put an
# 'acc routine' directive inside, and module-inline each kernel
for kern in schedule.coded_kernels():
if kern.name == "kern_use_var_code":
g2localtrans.apply(kern)
ktrans.apply(kern)
itrans.apply(kern)
schedule.view()
return psy
def trans(psy):
''' Take the supplied psy object, apply OpenACC transformations
to the schedule of the first invoke and return the new psy object '''
from psyclone.transformations import ACCParallelTrans, \
ACCEnterDataTrans, ACCLoopTrans, ACCRoutineTrans, \
KernelGlobalsToArguments
ptrans = ACCParallelTrans()
ltrans = ACCLoopTrans()
dtrans = ACCEnterDataTrans()
ktrans = ACCRoutineTrans()
g2localtrans = KernelGlobalsToArguments()
invoke = psy.invokes.invoke_list[0]
schedule = invoke.schedule
schedule.view()
# Apply the OpenACC Loop transformation to *every* loop
# nest in the schedule
from psyclone.psyir.nodes import Loop
for child in schedule.children:
if isinstance(child, Loop):
newschedule, _ = ltrans.apply(child, {"collapse": 2})
schedule = newschedule
# Put all of the loops in a single parallel region
newschedule, _ = ptrans.apply(schedule.children)
# Add an enter-data directive
newschedule, _ = dtrans.apply(schedule)
# Convert any accesses to global data into kernel arguments and then
# put an 'acc routine' directive inside each kernel
for kern in schedule.coded_kernels():
if kern.name == "kern_use_var_code":
# TODO #490 and #663. This currently won't work because the
# KernelGlobalsToArguments transformation works on the PSyIR but
# the subsequent ACCRoutineTrans works on the fparser2 parse tree.
g2localtrans.apply(kern)
_, _ = ktrans.apply(kern)
# Ideally we would module-inline the kernel here (to save having to
# rely on the compiler to do it) but this does not currently work
# for the fparser2 AST (issue #229).
# _, _ = itrans.apply(kern)
invoke.schedule = newschedule
newschedule.view()
return psy
def trans(psy):
'''PSyclone transformation script for the dynamo0p3 api to apply
OpenACC loop, parallel and enter data directives generically.
'''
loop_trans = ACCLoopTrans()
parallel_trans = ACCParallelTrans()
enter_data_trans = ACCEnterDataTrans()
# Loop over all of the Invokes in the PSy object
for invoke in psy.invokes.invoke_list:
print("Transforming invoke '" + invoke.name + "'...")
schedule = invoke.schedule
for loop in schedule.loops():
_, _ = loop_trans.apply(loop)
_, _ = parallel_trans.apply(loop.parent)
_, _ = enter_data_trans.apply(schedule)
return psy