def test_omp_transform():
'''Tests that the profiling transform works correctly with OMP
parallelisation.'''
_, invoke = get_invoke("test27_loop_swap.f90",
"gocean1.0",
name="invoke_loop1")
schedule = invoke.schedule
prt = ProfileRegionTrans()
omp_loop = GOceanOMPLoopTrans()
omp_par = OMPParallelTrans()
# Parallelise the first loop:
sched1, _ = omp_loop.apply(schedule.children[0])
sched2, _ = omp_par.apply(sched1.children[0])
sched3, _ = prt.apply(sched2.children[0])
correct = (
" CALL ProfileStart(\"boundary_conditions_ne_offset_mod\", "
"\"bc_ssh_code\", profile)\n"
" !$omp parallel default(shared), private(i,j)\n"
" !$omp do schedule(static)\n"
" DO j=2,jstop\n"
" DO i=2,istop\n"
" CALL bc_ssh_code(i, j, 1, t%data, t%grid%tmask)\n"
" END DO \n"
" END DO \n"
" !$omp end do\n"
" !$omp end parallel\n"
" CALL ProfileEnd(profile)")
code = str(invoke.gen())
assert correct in code
# Now add another profile node between the omp parallel and omp do
# directives:
sched3, _ = prt.apply(sched3.children[0].children[0].children[0])
code = str(invoke.gen())
correct = ''' CALL ProfileStart("boundary_conditions_ne_offset_mod", \
"bc_ssh_code", profile)
!$omp parallel default(shared), private(i,j)
CALL ProfileStart("boundary_conditions_ne_offset_mod", "bc_ssh_code_1", \
profile_1)
!$omp do schedule(static)
DO j=2,jstop
DO i=2,istop
CALL bc_ssh_code(i, j, 1, t%data, t%grid%tmask)
END DO\x20
END DO\x20
!$omp end do
CALL ProfileEnd(profile_1)
!$omp end parallel
CALL ProfileEnd(profile)'''
assert correct in code
def trans(psy):
'''
Take the supplied psy object, add OpenACC directives and then enclose
the whole schedule within a profiling region.
:param psy: the PSy layer to transform.
:type psy: :py:class:`psyclone.gocean1p0.GOPSy`
:returns: the transformed PSy object.
:rtype: :py:class:`psyclone.gocean1p0.GOPSy`
'''
from psyclone.transformations import ProfileRegionTrans
proftrans = ProfileRegionTrans()
# Use the trans() routine in acc_transform.py to add the OpenACC directives
psy = acc_trans(psy)
invoke = psy.invokes.get('invoke_0_inc_field')
schedule = invoke.schedule
# Enclose everything in a profiling region
newschedule, _ = proftrans.apply(schedule.children)
invoke.schedule = newschedule
newschedule.view()
return psy
def add_profile_nodes(schedule, loop_class):
'''This function inserts all required Profiling Nodes (for invokes
and kernels, as specified on the command line) into a schedule.
:param schedule: The schedule to instrument.
:type schedule: :py::class::`psyclone.psyGen.Schedule` or derived class
:param loop_class: The loop class (e.g. GOLoop, DynLoop) to instrument.
:type loop_class: :py::class::`psyclone.psyGen.Loop` or derived class.
'''
from psyclone.transformations import ProfileRegionTrans
profile_trans = ProfileRegionTrans()
if Profiler.profile_kernels():
for i in schedule.children:
if isinstance(i, loop_class):
profile_trans.apply(i)
if Profiler.profile_invokes():
profile_trans.apply(schedule.children)
def test_profile_basic(capsys):
'''Check basic functionality: node names, schedule view.
'''
from psyclone.psyGen import colored, SCHEDULE_COLOUR_MAP
Profiler.set_options([Profiler.INVOKES])
_, invoke = get_invoke("test11_different_iterates_over_one_invoke.f90",
"gocean1.0",
idx=0)
assert isinstance(invoke.schedule.children[0], ProfileNode)
invoke.schedule.view()
out, _ = capsys.readouterr()
gsched = colored("GOInvokeSchedule", SCHEDULE_COLOUR_MAP["Schedule"])
loop = Loop().coloured_text
profile = invoke.schedule.children[0].coloured_text
# Do one test based on schedule view, to make sure colouring
# and indentation is correct
expected = (gsched + "[invoke='invoke_0', Constant loop bounds=True]\n"
" " + profile + "\n"
" " + loop + "[type='outer', field_space='go_cv', "
"it_space='go_internal_pts']\n")
assert expected in out
prt = ProfileRegionTrans()
# Insert a profile call between outer and inner loop.
# This tests that we find the subroutine node even
# if it is not the immediate parent.
new_sched, _ = prt.apply(
invoke.schedule.children[0].children[0].children[0])
new_sched_str = str(new_sched)
correct = ("""GOInvokeSchedule(Constant loop bounds=True):
ProfileStart[var=profile]
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'2']
Literal[value:'jstop-1']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'2']
Literal[value:'istop']
Literal[value:'1']
Schedule:
kern call: compute_cv_code
End Schedule
End GOLoop
End Schedule
End GOLoop
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'1']
Literal[value:'jstop+1']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'1']
Literal[value:'istop+1']
Literal[value:'1']
Schedule:
kern call: bc_ssh_code
End Schedule
End GOLoop
End Schedule
End GOLoop
ProfileEnd
End Schedule""")
assert correct in new_sched_str
Profiler.set_options(None)
def test_transform_errors(capsys):
'''Tests error handling of the profile region transformation.'''
# This has been imported and tested before, so we can assume
# here that this all works as expected/
_, invoke = get_invoke("test27_loop_swap.f90",
"gocean1.0",
name="invoke_loop1")
schedule = invoke.schedule
prt = ProfileRegionTrans()
with pytest.raises(TransformationError) as excinfo:
prt.apply([schedule.children[0].children[0], schedule.children[1]])
assert "supplied nodes are not children of the same parent." \
in str(excinfo)
# Supply not a node object:
with pytest.raises(TransformationError) as excinfo:
prt.apply(5)
assert "Argument must be a single Node in a schedule or a list of Nodes " \
"in a schedule but have been passed an object of type: " \
in str(excinfo)
# Python 3 reports 'class', python 2 'type' - so just check for both
assert "<type 'int'>" in str(excinfo) or "<class 'int'>" in str(excinfo)
# Test that it will only allow correctly ordered nodes:
with pytest.raises(TransformationError) as excinfo:
sched1, _ = prt.apply([schedule.children[1], schedule.children[0]])
assert "Children are not consecutive children of one parent:" \
in str(excinfo)
with pytest.raises(TransformationError) as excinfo:
sched1, _ = prt.apply([schedule.children[0], schedule.children[2]])
assert "Children are not consecutive children of one parent:" \
in str(excinfo)
# Test 3 element lists: first various incorrect ordering:
with pytest.raises(TransformationError) as excinfo:
sched1, _ = prt.apply(
[schedule.children[0], schedule.children[2], schedule.children[1]])
assert "Children are not consecutive children of one parent:" \
in str(excinfo)
with pytest.raises(TransformationError) as excinfo:
sched1, _ = prt.apply(
[schedule.children[1], schedule.children[0], schedule.children[2]])
assert "Children are not consecutive children of one parent:" \
in str(excinfo)
# Just to be sure: also check that the right order does indeed work!
sched1, _ = prt.apply(
[schedule.children[0], schedule.children[1], schedule.children[2]])
sched1.view()
out, _ = capsys.readouterr()
# out is unicode, and has no replace function, so convert to string first
out = str(out).replace("\n", "")
correct_re = (".*GOInvokeSchedule.*"
r" .*Profile.*"
r" .*Loop.*\[type='outer'.*"
r" .*Loop.*\[type='outer'.*"
r" .*Loop.*\[type='outer'.*")
assert re.search(correct_re, out)
# Test that we don't add a profile node inside a OMP do loop (which
# would be invalid syntax):
_, invoke = get_invoke("test27_loop_swap.f90",
"gocean1.0",
name="invoke_loop1")
schedule = invoke.schedule
prt = ProfileRegionTrans()
omp_loop = GOceanOMPLoopTrans()
# Parallelise the first loop:
sched1, _ = omp_loop.apply(schedule.children[0])
# Inserting a ProfileRegion inside a omp do loop is syntactically
# incorrect, the inner part must be a do loop only:
with pytest.raises(TransformationError) as excinfo:
prt.apply(sched1.children[0].children[0])
assert "A ProfileNode cannot be inserted between an OpenMP/ACC directive "\
"and the loop(s) to which it applies!" in str(excinfo)
def test_transform(capsys):
'''Tests normal behaviour of profile region transformation.'''
_, invoke = get_invoke("test27_loop_swap.f90",
"gocean1.0",
name="invoke_loop1")
schedule = invoke.schedule
prt = ProfileRegionTrans()
assert str(prt) == "Insert a profile start and end call."
assert prt.name == "ProfileRegionTrans"
# Try applying it to a list
sched1, _ = prt.apply(schedule.children)
correct = ("""GOInvokeSchedule(Constant loop bounds=True):
ProfileStart[var=profile]
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'2']
Literal[value:'jstop']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'2']
Literal[value:'istop']
Literal[value:'1']
Schedule:
kern call: bc_ssh_code
End Schedule
End GOLoop
End Schedule
End GOLoop
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'1']
Literal[value:'jstop+1']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'1']
Literal[value:'istop']
Literal[value:'1']
Schedule:
kern call: bc_solid_u_code
End Schedule
End GOLoop
End Schedule
End GOLoop
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'1']
Literal[value:'jstop']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'1']
Literal[value:'istop+1']
Literal[value:'1']
Schedule:
kern call: bc_solid_v_code
End Schedule
End GOLoop
End Schedule
End GOLoop
ProfileEnd
End Schedule""")
assert correct in str(sched1)
# Now only wrap a single node - the middle loop:
sched2, _ = prt.apply(schedule.children[0].children[1])
correct = ("""GOInvokeSchedule(Constant loop bounds=True):
ProfileStart[var=profile]
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'2']
Literal[value:'jstop']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'2']
Literal[value:'istop']
Literal[value:'1']
Schedule:
kern call: bc_ssh_code
End Schedule
End GOLoop
End Schedule
End GOLoop
ProfileStart[var=profile_1]
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'1']
Literal[value:'jstop+1']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'1']
Literal[value:'istop']
Literal[value:'1']
Schedule:
kern call: bc_solid_u_code
End Schedule
End GOLoop
End Schedule
End GOLoop
ProfileEnd
GOLoop[id:'', variable:'j', loop_type:'outer']
Literal[value:'1']
Literal[value:'jstop']
Literal[value:'1']
Schedule:
GOLoop[id:'', variable:'i', loop_type:'inner']
Literal[value:'1']
Literal[value:'istop+1']
Literal[value:'1']
Schedule:
kern call: bc_solid_v_code
End Schedule
End GOLoop
End Schedule
End GOLoop
ProfileEnd
End Schedule""")
assert correct in str(sched2)
# Check that an sublist created from individual elements
# can be wrapped
sched3, _ = prt.apply(
[sched2.children[0].children[0], sched2.children[0].children[1]])
sched3.view()
out, _ = capsys.readouterr() # .replace("\n", "")
# out is unicode, and has no replace function, so convert to string first
out = str(out).replace("\n", "")
correct_re = (".*GOInvokeSchedule.*"
r" .*Profile.*"
r" .*Profile.*"
r" .*Loop.*\[type='outer'.*"
r" .*Profile.*"
r" .*Loop.*\[type='outer'.*"
r" .*Loop.*\[type='outer'.*")
assert re.search(correct_re, out)
def test_profile_basic(capsys):
'''Check basic functionality: node names, schedule view.
'''
Profiler.set_options([Profiler.INVOKES])
_, invoke = get_invoke("test11_different_iterates_over_one_invoke.f90",
"gocean1.0", idx=0)
assert isinstance(invoke.schedule.children[0], ProfileNode)
invoke.schedule.view()
out, _ = capsys.readouterr()
coloured_schedule = GOSchedule([]).coloured_text
coloured_loop = Loop().coloured_text
coloured_kern = GOKern().coloured_text
coloured_profile = invoke.schedule.children[0].coloured_text
# Do one test based on schedule view, to make sure colouring
# and indentation is correct
correct = (
'''{0}[invoke='invoke_0',Constant loop bounds=True]
{3}
{1}[type='outer',field_space='go_cv',it_space='go_internal_pts']
{1}[type='inner',field_space='go_cv',it_space='go_internal_pts']
{2} compute_cv_code(cv_fld,p_fld,v_fld) '''
'''[module_inline=False]
{1}[type='outer',field_space='go_ct',it_space='go_all_pts']
{1}[type='inner',field_space='go_ct',it_space='go_all_pts']
{2} bc_ssh_code(ncycle,p_fld,tmask) '''
'''[module_inline=False]'''.format(coloured_schedule, coloured_loop,
coloured_kern, coloured_profile)
)
assert correct in out
prt = ProfileRegionTrans()
# Insert a profile call between outer and inner loop.
# This tests that we find the subroutine node even
# if it is not the immediate parent.
new_sched, _ = prt.apply(invoke.schedule.children[0]
.children[0].children[0])
new_sched_str = str(new_sched)
correct = ("""GOSchedule(Constant loop bounds=True):
ProfileStart[var=profile]
Loop[]: j= lower=2,jstop-1,1
ProfileStart[var=profile_1]
Loop[]: i= lower=2,istop,1
kern call: compute_cv_code
EndLoop
ProfileEnd
EndLoop
Loop[]: j= lower=1,jstop+1,1
Loop[]: i= lower=1,istop+1,1
kern call: bc_ssh_code
EndLoop
EndLoop
ProfileEnd
End Schedule""")
assert correct in new_sched_str
Profiler.set_options(None)