def test_omp_region_no_slice_no_const_bounds():
''' Test that we generate the correct code when we apply an OpenMP
PARALLEL region transformation to a list of nodes when the Schedule
has been transformed to use loop-bound look-ups '''
psy, invoke = get_invoke("single_invoke_three_kernels.f90", 0)
schedule = invoke.schedule
ompr = OMPParallelTrans()
cbtrans = GOConstLoopBoundsTrans()
newsched, _ = cbtrans.apply(schedule, const_bounds=False)
omp_schedule, _ = ompr.apply(newsched.children)
# Replace the original loop schedule with the transformed one
invoke.schedule = omp_schedule
# Store the results of applying this code transformation as
# a string
gen = str(psy.gen)
gen = gen.lower()
# Iterate over the lines of generated code
within_omp_region = False
call_count = 0
for line in gen.split('\n'):
if '!$omp parallel default' in line:
within_omp_region = True
if '!$omp end parallel' in line:
within_omp_region = False
if ' call ' in line and within_omp_region:
call_count += 1
assert call_count == 3
def test_const_loop_bounds_invalid_offset():
''' Test that we raise an appropriate error if we attempt to generate
code with constant loop bounds for a kernel that expects an
unsupported grid-offset '''
psy, invoke = get_invoke("test26_const_bounds_invalid_offset.f90", 0)
cbtrans = GOConstLoopBoundsTrans()
schedule = invoke.schedule
newsched, _ = cbtrans.apply(schedule, const_bounds=True)
invoke.schedule = newsched
with pytest.raises(GenerationError):
_ = psy.gen
def test_const_loop_bounds_not_schedule():
''' Check that we raise an error if we attempt to apply the
constant loop-bounds transformation to something that is
not a Schedule '''
_, invoke = get_invoke("test11_different_iterates_over_"
"one_invoke.f90", 0)
schedule = invoke.schedule
cbtrans = GOConstLoopBoundsTrans()
with pytest.raises(TransformationError):
_, _ = cbtrans.apply(schedule.children[0])
def test_omp_region_retains_kernel_order1():
''' Test that applying the OpenMP PARALLEL region transformation
to a sub-set of nodes (last 2 of three) does not change their
ordering '''
psy, invoke = get_invoke("single_invoke_three_kernels.f90", 0)
schedule = invoke.schedule
ompr = OMPParallelTrans()
cbtrans = GOConstLoopBoundsTrans()
omp_schedule, _ = ompr.apply(schedule.children[1:])
# Replace the original loop schedule with the transformed one
invoke.schedule = omp_schedule
# Store the results of applying this code transformation as
# a string
gen = str(psy.gen)
gen = gen.lower()
# Iterate over the lines of generated code
cu_idx = -1
cv_idx = -1
ts_idx = -1
for idx, line in enumerate(gen.split('\n')):
if 'call compute_cu' in line:
cu_idx = idx
if 'call compute_cv' in line:
cv_idx = idx
if 'call time_smooth' in line:
ts_idx = idx
# Kernels should be in order {compute_cu, compute_cv, time_smooth}
assert cu_idx < cv_idx and cv_idx < ts_idx
# Repeat after turning off constant loop bounds
newsched, _ = cbtrans.apply(omp_schedule, const_bounds=False)
invoke.schedule = newsched
gen = str(psy.gen)
gen = gen.lower()
# Iterate over the lines of generated code
cu_idx = -1
cv_idx = -1
ts_idx = -1
for idx, line in enumerate(gen.split('\n')):
if 'call compute_cu' in line:
cu_idx = idx
if 'call compute_cv' in line:
cv_idx = idx
if 'call time_smooth' in line:
ts_idx = idx
# Kernels should be in order {compute_cu, compute_cv, time_smooth}
assert cu_idx < cv_idx and cv_idx < ts_idx
def test_omp_region_with_single_loop():
''' Test that we can pass the OpenMP PARALLEL region transformation
a single node in a schedule '''
psy, invoke = get_invoke("single_invoke_three_kernels.f90", 0)
schedule = invoke.schedule
ompr = OMPParallelTrans()
cbtrans = GOConstLoopBoundsTrans()
omp_schedule, _ = ompr.apply(schedule.children[1])
# Replace the original loop schedule with the transformed one
invoke.schedule = omp_schedule
# Store the results of applying this code transformation as
# a string
gen = str(psy.gen)
gen = gen.lower()
# Iterate over the lines of generated code
within_omp_region = False
call_count = 0
for line in gen.split('\n'):
if '!$omp parallel default' in line:
within_omp_region = True
if '!$omp end parallel' in line:
within_omp_region = False
if ' call ' in line and within_omp_region:
call_count += 1
assert call_count == 1
# Repeat the test after turning off constant loop bounds
newsched, _ = cbtrans.apply(omp_schedule, const_bounds=False)
invoke.schedule = newsched
gen = str(psy.gen)
gen = gen.lower()
within_omp_region = False
call_count = 0
for line in gen.split('\n'):
if '!$omp parallel default' in line:
within_omp_region = True
if '!$omp end parallel' in line:
within_omp_region = False
if ' call ' in line and within_omp_region:
call_count += 1
assert call_count == 1
def test_loop_fuse_different_iterates_over():
''' Test that an appropriate error is raised when we attempt to
fuse two loops that have differing values of ITERATES_OVER '''
_, invoke = get_invoke("test11_different_iterates_over_"
"one_invoke.f90", 0)
schedule = invoke.schedule
lftrans = LoopFuseTrans()
cbtrans = GOConstLoopBoundsTrans()
# Attempt to fuse two loops that are iterating over different
# things
with pytest.raises(TransformationError):
_, _ = lftrans.apply(schedule.children[0], schedule.children[1])
# Turn off constant loop bounds (which should have no effect)
# and repeat
newsched, _ = cbtrans.apply(schedule, const_bounds=False)
with pytest.raises(TransformationError):
_, _ = lftrans.apply(newsched.children[0], newsched.children[1])
def test_const_loop_bounds_toggle():
''' Check that we can toggle constant loop bounds on and off and
that the default behaviour is "on" '''
psy, invoke = get_invoke(
"test11_different_iterates_over_"
"one_invoke.f90", 0)
schedule = invoke.schedule
cbtrans = GOConstLoopBoundsTrans()
# First check that the generated code uses constant loop
# bounds by default
gen = str(psy.gen)
assert "INTEGER istop, jstop" in gen
assert "istop = cv_fld%grid%simulation_domain%xstop" in gen
assert "jstop = cv_fld%grid%simulation_domain%ystop" in gen
assert "DO j=2,jstop-1" in gen
assert "DO i=2,istop" in gen
# Next, check that applying the constant loop-bounds
# transformation has no effect (in this case)
newsched, _ = cbtrans.apply(schedule)
invoke.schedule = newsched
# Store the generated code as a string
gen = str(psy.gen)
assert "INTEGER istop, jstop" in gen
assert "istop = cv_fld%grid%simulation_domain%xstop" in gen
assert "jstop = cv_fld%grid%simulation_domain%ystop" in gen
assert "DO j=2,jstop-1" in gen
assert "DO i=2,istop" in gen
# Finally, test that we can turn-off constant loop bounds
newsched, _ = cbtrans.apply(schedule, const_bounds=False)
invoke.schedule = newsched
# Store the generated code as a string
gen = str(psy.gen)
assert "DO j=cv_fld%internal%ystart,cv_fld%internal%ystop" in gen
assert "DO i=cv_fld%internal%xstart,cv_fld%internal%xstop" in gen
assert "DO j=p_fld%whole%ystart,p_fld%whole%ystop" in gen
assert "DO i=p_fld%whole%xstart,p_fld%whole%xstop" in gen
def test_omp_region_commutes_with_loop_trans_bounds_lookup():
''' Test that the OpenMP PARALLEL region and (orphan) loop
transformations commute after constant bounds have been
switched off - i.e. we get the same result
independent of the order in which they are applied. '''
psy, invoke = get_invoke("single_invoke_two_kernels.f90", 0)
schedule = invoke.schedule
# Turn-off constant loop bounds
cbtrans = GOConstLoopBoundsTrans()
newsched, _ = cbtrans.apply(schedule, const_bounds=False)
# Put an OpenMP do directive around each loop contained
# in the schedule
ompl = GOceanOMPLoopTrans()
for child in newsched.children:
omp_schedule, _ = ompl.apply(child)
# Now put an OpenMP parallel region around that set of
# loops
ompr = OMPParallelTrans()
schedule, _ = ompr.apply(omp_schedule.children)
# Replace the original loop schedule with the transformed one
invoke.schedule = schedule
# Store the results of applying this code transformation as
# a string
loop_before_region_gen = str(psy.gen)
# Now we do it again but in the opposite order...
# ...we re-generate the original schedule here rather than
# keeping a (deep) copy of it from earlier as that can
# cause resource problems.
psy, invoke = get_invoke("single_invoke_two_kernels.f90", 0)
schedule = invoke.schedule
# Turn-off constant loop bounds
cbtrans = GOConstLoopBoundsTrans()
schedule, _ = cbtrans.apply(schedule, const_bounds=False)
# Put all of the loops in the schedule within a single
# OpenMP region
ompr = OMPParallelTrans()
omp_schedule, _ = ompr.apply(schedule.children)
# Put an OpenMP do directive around each loop contained
# in the region
ompl = GOceanOMPLoopTrans()
for child in omp_schedule.children[0].children:
schedule, _ = ompl.apply(child)
omp_schedule = schedule
# Replace the original loop schedule with the transformed one
invoke.schedule = omp_schedule
# Store the results of applying this code transformation as
# a string
region_before_loop_gen = str(psy.gen)
assert region_before_loop_gen == loop_before_region_gen
def test_omp_parallel_loop():
'''Test that we can generate an OMP PARALLEL DO correctly,
independent of whether or not we are generating constant loop bounds '''
psy, invoke = get_invoke("single_invoke_three_kernels.f90", 0)
schedule = invoke.schedule
omp = GOceanOMPParallelLoopTrans()
cbtrans = GOConstLoopBoundsTrans()
omp_sched, _ = omp.apply(schedule.children[0])
invoke.schedule = omp_sched
gen = str(psy.gen)
gen = gen.lower()
expected = ("!$omp parallel do default(shared), private(j,i), "
"schedule(static)\n"
" do j=2,jstop\n"
" do i=2,istop+1\n"
" call compute_cu_code(i, j, cu_fld%data, "
"p_fld%data, u_fld%data)\n"
" end do \n"
" end do \n"
" !$omp end parallel do")
assert expected in gen
newsched, _ = cbtrans.apply(omp_sched, const_bounds=False)
invoke.schedule = newsched
gen = str(psy.gen)
gen = gen.lower()
expected = (
" !$omp parallel do default(shared), private(j,i), "
"schedule(static)\n"
" do j=cu_fld%internal%ystart,cu_fld%internal%ystop\n"
" do i=cu_fld%internal%xstart,cu_fld%internal%xstop\n"
" call compute_cu_code(i, j, cu_fld%data, p_fld%data, "
"u_fld%data)\n"
" end do \n"
" end do \n"
" !$omp end parallel do")
assert expected in gen
def test_driver_creation(tmpdir):
'''Test that driver is created correctly for all variable access \
modes (input, input-output, output).
'''
# Use tmpdir so that the driver is created in tmp
tmpdir.chdir()
etrans = GOceanExtractTrans()
ctrans = GOConstLoopBoundsTrans()
psy, invoke = get_invoke("driver_test.f90",
GOCEAN_API,
idx=0,
dist_mem=False)
schedule = invoke.schedule
# This test expects constant loop bounds
ctrans.apply(schedule)
etrans.apply(schedule.children[0], {'create_driver': True})
# We are only interested in the driver, so ignore results.
str(psy.gen)
driver = tmpdir.join("driver-psy_extract_example_with_various_variable_"
"access_patterns-invoke_0_compute_kernel:compute_"
"kernel_code:r0.f90")
assert driver.isfile()
with driver.open("r") as driver_file:
driver_code = driver_file.read()
# This is an excerpt of the code that should get created.
# It is tested line by line since there is other code in between
# which is not important, and the order might also change. It also
# tests if unique variable names are created in the driver: the user
# program contains a local variable 'dx', which clashes with the grid
# property dx. The grid property will be renamed to 'dx_1':
expected = '''USE extract_psy_data_mod, ONLY: extract_PSyDataType
IMPLICIT NONE
REAL(KIND=8), allocatable, dimension(:,:) :: gphiu
REAL(KIND=8), allocatable, dimension(:,:) :: out_fld
REAL(KIND=8), allocatable, dimension(:,:) :: out_fld_post
REAL(KIND=8), allocatable, dimension(:,:) :: in_fld
REAL(KIND=8), allocatable, dimension(:,:) :: dx
REAL(KIND=8), allocatable, dimension(:,:) :: dx_1
REAL(KIND=8), allocatable, dimension(:,:) :: in_out_fld
REAL(KIND=8), allocatable, dimension(:,:) :: in_out_fld_post
TYPE(extract_PSyDataType) extract_psy_data
CALL extract_psy_data%OpenRead("psy_extract_example_with_various_variable''' \
'''_access_patterns", "invoke_0_compute_kernel:compute_kernel_code:r0")
CALL extract_psy_data%ReadVariable("out_fld_post", out_fld_post)
ALLOCATE (out_fld, mold=out_fld_post)
out_fld = 0.0
CALL extract_psy_data%ReadVariable("in_fld", in_fld)
CALL extract_psy_data%ReadVariable("in_out_fld_post", in_out_fld_post)
CALL extract_psy_data%ReadVariable("dx", dx)
CALL extract_psy_data%ReadVariable("in_fld%grid%dx", dx_1)
CALL extract_psy_data%ReadVariable("in_fld%grid%gphiu", gphiu)
! RegionStart
DO j=2,jstop
DO i=2,istop+1
CALL compute_kernel_code(i, j, out_fld, in_out_fld, in_fld, ''' \
'''dx, dx_1, gphiu)
END DO
END DO
! RegionEnd
!
! Check out_fld
! Check i
! Check j
! Check in_out_fld'''
expected_lines = expected.split("\n")
for line in expected_lines:
assert line in driver_code
def test_node_list_ompparallel_gocean1p0():
''' Test that applying Extract Transformation on a list of Nodes
enclosed within an OMP Parallel Region produces the correct result
in GOcean1.0 API. '''
etrans = GOceanExtractTrans()
ltrans = GOceanOMPLoopTrans()
otrans = OMPParallelTrans()
ctrans = GOConstLoopBoundsTrans()
# Test a Loop nested within the OMP Parallel DO Directive
psy, invoke = get_invoke("single_invoke_three_kernels.f90",
GOCEAN_API,
idx=0,
dist_mem=False)
schedule = invoke.schedule
# Apply GOConstLoopBoundsTrans
ctrans.apply(schedule)
# Apply GOceanOMPParallelLoopTrans to the first two Loops
ltrans.apply(schedule.children[0])
ltrans.apply(schedule.children[1])
# and enclose them within a parallel region
otrans.apply(schedule.children[0:2])
# Now enclose the parallel region within an ExtractNode (inserted
# at the previous location of the OMPParallelDirective
etrans.apply(schedule.children[0])
code = str(psy.gen)
output = """
! ExtractStart
!
CALL extract_psy_data%PreStart("psy_single_invoke_three_kernels", """ \
""""invoke_0:r0", 3, 4)
CALL extract_psy_data%PreDeclareVariable("p_fld", p_fld)
CALL extract_psy_data%PreDeclareVariable("u_fld", u_fld)
CALL extract_psy_data%PreDeclareVariable("v_fld", v_fld)
CALL extract_psy_data%PreDeclareVariable("cu_fld_post", cu_fld)
CALL extract_psy_data%PreDeclareVariable("cv_fld_post", cv_fld)
CALL extract_psy_data%PreDeclareVariable("i_post", i)
CALL extract_psy_data%PreDeclareVariable("j_post", j)
CALL extract_psy_data%PreEndDeclaration
CALL extract_psy_data%ProvideVariable("p_fld", p_fld)
CALL extract_psy_data%ProvideVariable("u_fld", u_fld)
CALL extract_psy_data%ProvideVariable("v_fld", v_fld)
CALL extract_psy_data%PreEnd
!$omp parallel default(shared), private(i,j)
!$omp do schedule(static)
DO j=2,jstop
DO i=2,istop+1
CALL compute_cu_code(i, j, cu_fld%data, p_fld%data, u_fld%data)
END DO
END DO
!$omp end do
!$omp do schedule(static)
DO j=2,jstop+1
DO i=2,istop
CALL compute_cv_code(i, j, cv_fld%data, p_fld%data, v_fld%data)
END DO
END DO
!$omp end do
!$omp end parallel
CALL extract_psy_data%PostStart
CALL extract_psy_data%ProvideVariable("cu_fld_post", cu_fld)
CALL extract_psy_data%ProvideVariable("cv_fld_post", cv_fld)
CALL extract_psy_data%ProvideVariable("i_post", i)
CALL extract_psy_data%ProvideVariable("j_post", j)
CALL extract_psy_data%PostEnd
!
! ExtractEnd"""
assert output in code
def test_single_node_ompparalleldo_gocean1p0_failing_const_loop():
''' Test that applying Extract Transformation on a Node enclosed
within an OMP Parallel DO Directive produces the correct result
in GOcean1.0 API. This test is mostly identical to the previous one,
but uses const loop bounds. At this stage, the dependency analysis
still reports `cv_fld%internal%xstart` etc as loop boundaries, but
the code created will be using istop and jstop.
'''
etrans = GOceanExtractTrans()
otrans = GOceanOMPParallelLoopTrans()
ctrans = GOConstLoopBoundsTrans()
# Test a Loop nested within the OMP Parallel DO Directive
psy, invoke = get_invoke("single_invoke_three_kernels.f90",
GOCEAN_API,
idx=0,
dist_mem=False)
schedule = invoke.schedule
ctrans.apply(schedule)
# Required for #969
schedule.view()
# Apply GOceanOMPParallelLoopTrans to the second Loop
otrans.apply(schedule.children[1])
# Now enclose the parallel region within an ExtractNode (inserted
# at the previous location of the OMPParallelDoDirective
etrans.apply(schedule.children[1])
code = str(psy.gen)
output = """ ! ExtractStart
!
CALL extract_psy_data%PreStart("psy_single_invoke_three_kernels", """ \
""""invoke_0:compute_cv_code:r0", 6, 3)
CALL extract_psy_data%PreDeclareVariable("istop", istop)
CALL extract_psy_data%PreDeclareVariable("jstop", jstop)
CALL extract_psy_data%PreDeclareVariable("p_fld", p_fld)
CALL extract_psy_data%PreDeclareVariable("v_fld", v_fld)
CALL extract_psy_data%PreDeclareVariable("cv_fld_post", cv_fld)
CALL extract_psy_data%PreDeclareVariable("i_post", i)
CALL extract_psy_data%PreDeclareVariable("j_post", j)
CALL extract_psy_data%PreEndDeclaration
CALL extract_psy_data%ProvideVariable("istop", istop)
CALL extract_psy_data%ProvideVariable("jstop", jstop)
CALL extract_psy_data%ProvideVariable("p_fld", p_fld)
CALL extract_psy_data%ProvideVariable("v_fld", v_fld)
CALL extract_psy_data%PreEnd
!$omp parallel do default(shared), private(i,j), schedule(static)
DO j=2,jstop+1
DO i=2,istop
CALL compute_cv_code(i, j, cv_fld%data, p_fld%data, v_fld%data)
END DO
END DO
!$omp end parallel do
CALL extract_psy_data%PostStart
CALL extract_psy_data%ProvideVariable("cv_fld_post", cv_fld)
CALL extract_psy_data%ProvideVariable("i_post", i)
CALL extract_psy_data%ProvideVariable("j_post", j)
CALL extract_psy_data%PostEnd
!
! ExtractEnd"""
assert output in code
