def make(registry, value, type_id, writer=None):
"""
:param registry:
:param value:
:param otf2.Type type_id:
:param writer: If set, the string value will be written. Otherwise just added to the registry
:return:
"""
# values are expected to be int, float, or a Definition subclass
union = _otf2.AttributeValue()
if type_id == Type.UINT8:
union.uint8 = int(value)
elif type_id == Type.UINT16:
union.uint16 = int(value)
elif type_id == Type.UINT32:
union.uint32 = int(value)
elif type_id == Type.UINT64:
union.uint64 = int(value)
elif type_id == Type.INT8:
union.int8 = int(value)
elif type_id == Type.INT16:
union.int16 = int(value)
elif type_id == Type.INT32:
union.int32 = int(value)
elif type_id == Type.INT64:
union.int64 = int(value)
elif type_id == Type.FLOAT:
union.float32 = float(value)
elif type_id == Type.DOUBLE:
union.float64 = float(value)
elif type_id == Type.STRING:
union.stringRef = registry.strings.get_ref(value, writer)
def test_3_random_remove(self):
attr_list = _otf2.AttributeList_New()
self.assertIsNotNone(attr_list)
ids_to_add = [4, 8, 15, 16, 23, 42]
# from the end, from the front, from the middle, as index into ids_to_add
ids_to_remove = [5, 0, 2, 1, 4, 3]
attr = _otf2.AttributeValue()
for (i, id_to_add) in enumerate(ids_to_add):
attr.uint32 = i
_otf2.AttributeList_AddAttribute(attr_list, id_to_add, _otf2.TYPE_UINT32, attr)
res = _otf2.AttributeList_TestAttributeByID(attr_list, id_to_add);
self.assertTrue(res, "failed test id")
number_of_elements = _otf2.AttributeList_GetNumberOfElements(attr_list)
self.assertEqual(number_of_elements, i + 1, "incorrect number of elements")
for (i, id_to_remove) in enumerate(ids_to_remove):
id_to_add = ids_to_add[id_to_remove]
(type_, attr) = _otf2.AttributeList_GetAttributeByID(attr_list, id_to_add)
self.assertEqual(type_, _otf2.TYPE_UINT32)
self.assertEqual(attr.uint32, id_to_remove)
_otf2.AttributeList_RemoveAttribute(attr_list, id_to_add)
ret = _otf2.AttributeList_TestAttributeByID(attr_list, id_to_add)
self.assertFalse(ret, "found removed test id")
number_of_elements = _otf2.AttributeList_GetNumberOfElements(attr_list)
self.assertEqual(number_of_elements, len(ids_to_add) - i - 1, "incorrect number of elements")
_otf2.AttributeList_Delete(attr_list)
def test_2_add_get_pop(self):
attr_list = _otf2.AttributeList_New()
self.assertIsNotNone(attr_list)
attr = _otf2.AttributeValue()
attr.uint32 = 1
_otf2.AttributeList_AddAttribute(attr_list, 42, _otf2.TYPE_UINT32, attr)
res = _otf2.AttributeList_TestAttributeByID(attr_list, 42)
self.assertTrue(res, "test id")
number_of_elements = _otf2.AttributeList_GetNumberOfElements(attr_list)
self.assertEqual(number_of_elements, 1, "incorrect number of elements")
# C-style
type_ = _otf2.Type()
id_ = _otf2.AttributeRef()
_otf2.AttributeList_GetAttributeByID(attr_list, 42, type_, attr)
self.assertEqual(type_, _otf2.TYPE_UINT32, "incorrect type")
self.assertNotEqual(type_, _otf2.TYPE_INT32, "incorrect type matches")
self.assertEqual(attr.uint32, 1, "incorrect value")
_otf2.AttributeList_GetAttributeByIndex(attr_list, 0, id_, type_, attr)
self.assertEqual(id_.value, 42, "incorrect id")
self.assertEqual(type_, _otf2.TYPE_UINT32, "incorrect type")
self.assertEqual(attr.uint32, 1, "incorrect value")
# Python-style
(type_, attr) = _otf2.AttributeList_GetAttributeByID(attr_list, 42)
self.assertEqual(type_, _otf2.TYPE_UINT32, "incorrect type")
self.assertNotEqual(type_, _otf2.TYPE_INT32, "incorrect type matches")
self.assertEqual(attr.uint32, 1, "incorrect value")
(id_, type_, attr) = _otf2.AttributeList_GetAttributeByIndex(attr_list, 0)
self.assertEqual(id_, 42, "incorrect id")
self.assertEqual(type_, _otf2.TYPE_UINT32, "incorrect type")
self.assertEqual(attr.uint32, 1, "incorrect value")
(id, type, attr) = _otf2.AttributeList_PopAttribute(attr_list)
self.assertEqual(id_, 42, "incorrect id")
self.assertEqual(type_, _otf2.TYPE_UINT32, "incorrect type")
self.assertEqual(attr.uint32, 1, "incorrect value")
number_of_elements = _otf2.AttributeList_GetNumberOfElements(attr_list)
self.assertEqual(number_of_elements, 0, "incorrect number of elements")
_otf2.AttributeList_Delete(attr_list)
def write_paradigm(global_def_writer):
str_id_mpi = write_string(global_def_writer, "MPI")
_otf2.GlobalDefWriter_WriteParadigm(global_def_writer, _otf2.PARADIGM_MPI,
str_id_mpi,
_otf2.PARADIGM_CLASS_PROCESS)
attribute_value = _otf2.AttributeValue()
attribute_value.stringRef = write_string(global_def_writer,
"MPI_COMM_${id}")
_otf2.GlobalDefWriter_WriteParadigmProperty(
global_def_writer, _otf2.PARADIGM_MPI,
_otf2.PARADIGM_PROPERTY_COMM_NAME_TEMPLATE, _otf2.TYPE_STRING,
attribute_value)
attribute_value.stringRef = write_string(global_def_writer,
"MPI_WIN_${id}")
_otf2.GlobalDefWriter_WriteParadigmProperty(
global_def_writer, _otf2.PARADIGM_MPI,
_otf2.PARADIGM_PROPERTY_RMA_WIN_NAME_TEMPLATE, _otf2.TYPE_STRING,
attribute_value)
def make(registry, value, type_id, writer=None):
"""
:param registry:
:param value:
:param otf2.Type type_id:
:param writer: If set, the string value will be written. Otherwise just added to the registry
:return:
"""
# values are expected to be int, float, or a Definition subclass
union = _otf2.AttributeValue()
if type_id == Type.UINT8:
union.uint8 = int(value)
elif type_id == Type.UINT16:
union.uint16 = int(value)
elif type_id == Type.UINT32:
union.uint32 = int(value)
elif type_id == Type.UINT64:
union.uint64 = int(value)
elif type_id == Type.INT8:
union.int8 = int(value)
elif type_id == Type.INT16:
union.int16 = int(value)
elif type_id == Type.INT32:
union.int32 = int(value)
elif type_id == Type.INT64:
union.int64 = int(value)
elif type_id == Type.FLOAT:
union.float32 = float(value)
elif type_id == Type.DOUBLE:
union.float64 = float(value)
elif type_id == Type.STRING:
union.stringRef = registry.strings.get_ref(value, writer)
elif type_id == Type.ATTRIBUTE:
union.attributeRef = value._ref
elif type_id == Type.LOCATION:
union.locationRef = value._ref
elif type_id == Type.REGION:
union.regionRef = value._ref
elif type_id == Type.GROUP:
union.groupRef = value._ref
elif type_id == Type.METRIC:
union.metricRef = value._ref
elif type_id == Type.COMM:
union.commRef = value._ref
elif type_id == Type.PARAMETER:
union.parameterRef = value._ref
elif type_id == Type.RMA_WIN:
union.rmaWinRef = value._ref
elif type_id == Type.SOURCE_CODE_LOCATION:
union.sourceCodeLocationRef = value._ref
elif type_id == Type.CALLING_CONTEXT:
union.callingContextRef = value._ref
elif type_id == Type.INTERRUPT_GENERATOR:
union.interruptGeneratorRef = value._ref
elif type_id == Type.IO_FILE:
union.ioFileRef = value._ref
elif type_id == Type.IO_HANDLE:
union.ioHandleRef = value._ref
else:
raise TypeError("Unexpected type id given")
return union
def main():
archive_name = "TestTrace"
# Array used to store metric values */
metric_values = (_otf2.MetricValue * NUM_OF_METRICS)()
# If you add new metrics please adapt also the array of metric types
# and the routine to generate random metric values. */
# Array of metric types */
metric_types = [
_otf2.TYPE_UINT64, _otf2.TYPE_DOUBLE, _otf2.TYPE_INT64,
_otf2.TYPE_DOUBLE, _otf2.TYPE_INT64
]
# Initialize accumulated metric */
metric_values[0].unsigned_int = 0
metric_values[1].floating_point = 12.5
metric_values[2].signed_int = 1
metric_values[3].floating_point = 10.0
metric_values[4].unsigned_int = 1
# Distribute metric members to its classes */
metric_members_of_class_1 = [METRIC_A, METRIC_B]
metric_members_of_class_2 = [METRIC_C]
metric_members_of_class_3 = [METRIC_D]
metric_members_of_class_4 = [METRIC_E]
# Array of offsets needed to get correct type or values of a metric in corresponding arrays. */
metric_classes_offsets = [0, 2, 3, 4, 4]
# The collection of metric classes of instances for which we write metric events */
class MetricWriteParameter:
def __init__(self, id_, offset, count, record_for_location):
self.id = id_
self.metric_types = metric_types[offset:offset + count]
self.metric_values = metric_values[offset:offset + count]
self.record_for_location = record_for_location
# FIXME missing class 4
metric_write_parameters = [
MetricWriteParameter(METRIC_CLASS_1, metric_classes_offsets[0],
len(metric_members_of_class_1),
record_for_all_locations),
MetricWriteParameter(METRIC_CLASS_2, metric_classes_offsets[1],
len(metric_members_of_class_2),
record_for_all_locations),
MetricWriteParameter(METRIC_CLASS_3, metric_classes_offsets[2],
len(metric_members_of_class_3),
record_for_all_locations),
MetricWriteParameter(METRIC_INSTANCE_1, metric_classes_offsets[3],
len(metric_members_of_class_4),
record_for_location_0)
]
def create_new_metric_values():
metric_values[0].unsigned_int += 1
metric_values[1].floating_point += 12.5
metric_values[2].unsigned_int = -metric_values[2].unsigned_int
metric_values[3].floating_point += 10.0
metric_values[4].unsigned_int += 1
def write_metrics(new, metric_time, metric_rank, metric_thread):
if new:
create_new_metric_values()
for metric_write_parameter in metric_write_parameters:
if metric_write_parameter.record_for_location(
metric_rank, metric_thread):
_otf2.EvtWriter_Metric(evt_writer, None, metric_time,
metric_write_parameter.id,
metric_write_parameter.metric_types,
metric_write_parameter.metric_values)
# Calculate number of MPI ranks and OpenMP threads. */
num_ranks = 1
num_threads = 1
if otf2_MPI:
num_ranks = otf2_LOCATIONS
archive_path = "otf2_trace_gen_trace_mpi_{}_{}{}".format(
num_ranks, otf2_EVENTS, "_attr" if otf2_ATTRIBUTES else "")
elif otf2_OPENMP:
num_threads = otf2_LOCATIONS
archive_path = "otf2_trace_gen_trace_omp_{}_{}{}".format(
num_threads, otf2_EVENTS, "_attr" if otf2_ATTRIBUTES else "")
elif otf2_HYBRID:
if (otf2_LOCATIONS % 4) == 0:
num_threads = 4
elif (otf2_LOCATIONS % 3) == 0:
num_threads = 3
elif (otf2_LOCATIONS % 2) == 0:
num_threads = 2
else:
# FIXME shadow
for num_threads in range(5, otf2_LOCATIONS):
if (otf2_LOCATIONS % num_threads) == 0:
break
num_ranks = otf2_LOCATIONS // num_threads
archive_path = "otf2_trace_gen_trace_mpi_omp_{}x{}_{}{}".format(
num_ranks, num_threads, otf2_EVENTS,
"_attr" if otf2_ATTRIBUTES else "")
else:
raise RuntimeError("No paradigm set")
print("\nGenerating an OTF2 test archive")
print("with {} rank(s) and {} thread(s) each ({} locations total)".format(
num_ranks, num_threads, otf2_LOCATIONS))
print("to {}/{}.otf2\n".format(archive_path, archive_name))
if os.path.isdir(archive_path):
for i in itertools.count():
archive_path_backup = "{}_bk_{}".format(archive_path, i)
if not os.path.exists(archive_path_backup):
os.rename(archive_path, archive_path_backup)
break
# Create new archive handle. */
archive = _otf2.Archive_Open(archive_path, archive_name,
_otf2.FILEMODE_WRITE, 1024 * 1024,
4 * 1024 * 1024, _otf2.SUBSTRATE_POSIX,
_otf2.COMPRESSION_NONE)
flush_callbacks = _otf2.FlushCallbacks(pre_flush=pre_flush,
post_flush=post_flush)
_otf2.Archive_SetFlushCallbacks(archive, flush_callbacks, None)
# Set description, and creator. */
_otf2.Archive_SetSerialCollectiveCallbacks(archive)
_otf2.Archive_SetDescription(
archive, "Simple OTF2 trace archive to valid reader components")
_otf2.Archive_SetCreator(archive, "otf2_trace_gen")
# Generate location IDs. Just to have non-consecutive location IDs. */
locations = []
mpi_ranks = []
master_threads = []
for rank in range(num_ranks):
for thread in range(num_threads):
locations.append((rank << 16) + thread)
mpi_ranks.append(rank)
master_threads.append(rank << 16)
_otf2.Archive_OpenEvtFiles(archive)
# Create local event writer. */
# Create an attribute list. */
attributes = None
if otf2_ATTRIBUTES:
attributes = _otf2.AttributeList_New()
t = get_time()
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(archive, master_threads[rank])
# Enter main. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t, REGION_MAIN)
for event in range(otf2_EVENTS):
t = get_time()
# FIXME - this makes no sense
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
# Write enter MPI_Send. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_SEND)
_otf2.EvtWriter_MpiSend(evt_writer, None, t,
(rank + 1) % num_ranks, rank, 0, 1024)
t = get_time()
# Write leave MPI_Send. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_SEND)
# Immediately enter the MPI_Recv */
t = get_time()
# Write enter MPI_Recv. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_RECV)
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
_otf2.EvtWriter_MpiRecv(evt_writer, None, t,
(rank + num_ranks - 1) % num_ranks,
rank, 0, 1024)
# Write enter MPI_Recv. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_RECV)
# MPI Barrier. */
t = get_time()
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_BARRIER)
_otf2.EvtWriter_MpiCollectiveBegin(evt_writer, None, t)
# End all started MPI_Barrier */
t = get_time()
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
_otf2.EvtWriter_MpiCollectiveEnd(evt_writer, None, t,
_otf2.COLLECTIVE_OP_BARRIER,
rank, _otf2.UNDEFINED_UINT32,
0, 0)
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_BARRIER)
# If in OpenMP or Hybrid mode. */
if not otf2_MPI:
t_enter = get_time()
t_leave = get_time()
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
# At this point the master creates a team of several parallel threads */
_otf2.EvtWriter_OmpFork(evt_writer, attributes, t_enter,
num_threads)
for rank in range(num_ranks):
for thread in range(num_threads):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, locations[num_threads * rank + thread])
# Enter OpenMP parallel loop (outer region). */
write_metrics(True, t_enter, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t_enter,
REGION_OPENMP_PARALLEL)
# Enter OpenMP parallel loop (inner region). */
write_metrics(False, t_enter, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t_enter,
REGION_OPENMP_LOOP)
# Enter OpenMP implicit barrier. */
write_metrics(False, t_enter, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t_enter,
REGION_OPENMP_IMPLICIT_BARRIER)
for rank in range(num_ranks):
for thread in range(num_threads):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, locations[num_threads * rank + thread])
# Leave OpenMP implicit barrier. */
write_metrics(True, t_leave, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t_leave,
REGION_OPENMP_IMPLICIT_BARRIER)
# Leave OpenMP parallel loop (inner region). */
write_metrics(False, t_leave, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t_leave,
REGION_OPENMP_LOOP)
# Leave OpenMP parallel loop (outer region). */
write_metrics(False, t_leave, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t_leave,
REGION_OPENMP_PARALLEL)
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
# All threads have finished execution of statements in parallel region.
# Now they synchronize and terminates. Only the master continues execution. */
t_leave = get_time()
_otf2.EvtWriter_OmpJoin(evt_writer, attributes, t_leave)
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
t = get_time()
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_BARRIER)
_otf2.EvtWriter_MpiCollectiveBegin(evt_writer, None, t)
# End all started MPI_Barrier */
t = get_time()
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
_otf2.EvtWriter_MpiCollectiveEnd(evt_writer, None, t,
_otf2.COLLECTIVE_OP_BARRIER,
rank, _otf2.UNDEFINED_UINT32,
0, 0)
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_BARRIER)
t = get_time()
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(archive, master_threads[rank])
# Leave main. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t, REGION_MAIN)
# last timestamp */
t = get_time()
number_of_events_per_location = []
for rank in range(num_ranks):
for thread in range(num_threads):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, locations[num_threads * rank + thread])
n = _otf2.EvtWriter_GetNumberOfEvents(evt_writer)
number_of_events_per_location.append(n)
_otf2.Archive_CloseEvtWriter(archive, evt_writer)
_otf2.Archive_CloseEvtFiles(archive)
# Write the global and local definitions. */
glob_def_writer = _otf2.Archive_GetGlobalDefWriter(archive)
# Write the timerange of the trace */
_otf2.GlobalDefWriter_WriteClockProperties(glob_def_writer, 1, 0, t)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, STRING_EMPTY, "")
string = STRING_EMPTY + 1
# Write paradigm definitions */
if otf2_MPI or otf2_HYBRID:
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI")
_otf2.GlobalDefWriter_WriteParadigm(glob_def_writer,
_otf2.PARADIGM_MPI, string,
_otf2.PARADIGM_CLASS_PROCESS)
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"MPI_COMM_${id}")
value = _otf2.AttributeValue()
value.stringRef = string
string += 1
_otf2.GlobalDefWriter_WriteParadigmProperty(
glob_def_writer, _otf2.PARADIGM_MPI,
_otf2.PARADIGM_PROPERTY_COMM_NAME_TEMPLATE, _otf2.TYPE_STRING,
value)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"MPI_WIN_${id}")
value.stringRef = string
string += 1
_otf2.GlobalDefWriter_WriteParadigmProperty(
glob_def_writer, _otf2.PARADIGM_MPI,
_otf2.PARADIGM_PROPERTY_RMA_WIN_NAME_TEMPLATE, _otf2.TYPE_STRING,
value)
(type_, value) = _otf2.AttributeValue_SetBoolean(_otf2.FALSE)
_otf2.GlobalDefWriter_WriteParadigmProperty(
glob_def_writer, _otf2.PARADIGM_OPENMP,
_otf2.PARADIGM_PROPERTY_RMA_ONLY, type_, value)
if otf2_OPENMP or otf2_HYBRID:
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "OpenMP")
_otf2.GlobalDefWriter_WriteParadigm(
glob_def_writer, _otf2.PARADIGM_OPENMP, string,
_otf2.PARADIGM_CLASS_THREAD_FORK_JOIN)
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"OMP Thread Team ${id}")
value = _otf2.AttributeValue()
value.stringRef = string
string += 1
# FIXME seems like a bad idea to write a hardcoded number 4 here
_otf2.GlobalDefWriter_WriteParadigmProperty(glob_def_writer,
_otf2.PARADIGM_OPENMP, 4,
_otf2.TYPE_STRING, value)
# Write system tree definitions. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Rack 0")
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string + 1, "Rack")
_otf2.GlobalDefWriter_WriteSystemTreeNode(glob_def_writer, 0, string,
string + 1,
_otf2.UNDEFINED_SYSTEM_TREE_NODE)
string += 2
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Blade")
string_blade = string
string += 1
for rank in range(num_ranks // 2):
name_buffer = "Blade #{}".format(rank)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, name_buffer)
_otf2.GlobalDefWriter_WriteSystemTreeNode(glob_def_writer, rank + 1,
string, string_blade, 0)
string += 1
# Write location group and location definitions. */
for rank in range(num_ranks):
name_buffer = "Process {}".format(rank)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, name_buffer)
_otf2.GlobalDefWriter_WriteLocationGroup(
glob_def_writer, rank, string, _otf2.LOCATION_GROUP_TYPE_PROCESS,
(rank // 2) + 1)
string += 1
for thread in range(num_threads):
name_buffer = "Thread {}.{}".format(rank, thread)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
name_buffer)
_otf2.GlobalDefWriter_WriteLocation(
glob_def_writer, locations[rank * num_threads + thread],
string, _otf2.LOCATION_TYPE_CPU_THREAD,
number_of_events_per_location[rank * num_threads + thread],
rank)
string += 1
# Write region definition for "main". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"main (mangled name)")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"main (demangled name)")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MAIN, string,
string - 1, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for an all-threads function. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"!$omp parallelfor")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_OPENMP_PARALLEL,
string, string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_OPENMP_LOOP,
string, string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for OpenMP implicit barrier. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"!$omp implicit barrier")
_otf2.GlobalDefWriter_WriteRegion(
glob_def_writer, REGION_OPENMP_IMPLICIT_BARRIER, string, string,
STRING_EMPTY, _otf2.REGION_ROLE_UNKNOWN, _otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0, 0)
string += 1
# Write region definition for "MPI_Send". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_Send")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MPI_SEND, string,
string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for "MPI_Recv". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_Recv")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MPI_RECV, string,
string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for "MPI_Barrier". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_Barrier")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MPI_BARRIER,
string, string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Define the list of locations which are mpi ranks */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_LOCATIONS,
STRING_EMPTY,
_otf2.GROUP_TYPE_COMM_LOCATIONS,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
master_threads)
# Define the MPI group for MPI_COMM_WORLD */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_COMM_WORLD,
STRING_EMPTY, _otf2.GROUP_TYPE_COMM_GROUP,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
mpi_ranks)
# Define the MPI group for COMM_SELF like communicators */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_COMM_SELF,
STRING_EMPTY, _otf2.GROUP_TYPE_COMM_SELF,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
[])
# Define a group that includes all locations. It will be used by metric instance. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "All locations")
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_ALL_LOCATIONS,
string, _otf2.GROUP_TYPE_LOCATIONS,
_otf2.PARADIGM_UNKNOWN,
_otf2.GROUP_FLAG_NONE, locations)
string += 1
# Define MPI_COMM_WORD */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"My MPI_COMM_WORLD communicator")
_otf2.GlobalDefWriter_WriteComm(glob_def_writer, MPI_COMM_MPI_COMM_WORLD,
string, GROUP_MPI_COMM_WORLD,
_otf2.UNDEFINED_COMM)
string += 1
# Define MPI_COMM_SELF */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"My MPI_COMM_SELF communicator")
_otf2.GlobalDefWriter_WriteComm(glob_def_writer, MPI_COMM_MPI_COMM_SELF,
string, GROUP_MPI_COMM_SELF,
_otf2.UNDEFINED_COMM)
string += 1
# Write member definition of first metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_A")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "FLOPS")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_A,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_PAPI,
_otf2.METRIC_ACCUMULATED_START,
metric_types[0],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Write member definition of second metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_B")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "s")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_B,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_ABSOLUTE_POINT,
metric_types[1],
_otf2.BASE_DECIMAL, -6, string)
string += 1
# Write member definition of third metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_C")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "apples")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_C,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_RELATIVE_LAST,
metric_types[2],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Write member definition of fourth metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_D")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "pears")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_D,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_ABSOLUTE_POINT,
metric_types[3],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Write member definition of fifth metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_E")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "cherries")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_E,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_ABSOLUTE_POINT,
metric_types[4],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Define metric classes and instances */
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_1,
metric_members_of_class_1,
_otf2.METRIC_SYNCHRONOUS_STRICT,
_otf2.RECORDER_KIND_CPU)
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_2,
metric_members_of_class_2,
_otf2.METRIC_SYNCHRONOUS_STRICT,
_otf2.RECORDER_KIND_CPU)
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_3,
metric_members_of_class_3,
_otf2.METRIC_SYNCHRONOUS_STRICT,
_otf2.RECORDER_KIND_CPU)
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_4,
metric_members_of_class_4,
_otf2.METRIC_ASYNCHRONOUS,
_otf2.RECORDER_KIND_ABSTRACT)
# We can define an instance with differing recorder and scope.
# In this example metric values are valid for all locationa
# but are recorded on first location. */
_otf2.GlobalDefWriter_WriteMetricInstance(glob_def_writer,
METRIC_INSTANCE_1,
METRIC_CLASS_4, locations[0],
_otf2.SCOPE_GROUP,
GROUP_ALL_LOCATIONS)
# write attribute defs */
if otf2_ATTRIBUTES:
for a, attribute_def in enumerate(attribute_defs):
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
attribute_def.name)
_otf2.GlobalDefWriter_WriteAttribute(glob_def_writer, a, string,
STRING_EMPTY,
attribute_def.type)
string += 1
# create a sub communicator of MPI_COMM_WORLD and embed an hyper cube. */
num_cube_ranks = gpot(num_ranks)
cube_ranks = list(range(num_cube_ranks))
# Define the MPI group for the hyper cube */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_COMM_CUBE,
STRING_EMPTY, _otf2.GROUP_TYPE_COMM_GROUP,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
cube_ranks)
# Define MPI_COMM_MPI_COMM_CUBE */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_COMM_CUBE")
_otf2.GlobalDefWriter_WriteComm(glob_def_writer, MPI_COMM_MPI_COMM_CUBE,
string, GROUP_MPI_COMM_WORLD,
MPI_COMM_MPI_COMM_WORLD)
string += 1
num_cube_dims = 0
cube_dim = 1
while cube_dim < num_cube_ranks:
name = "Cube {}".format(cube_dim)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, name)
_otf2.GlobalDefWriter_WriteCartDimension(glob_def_writer,
num_cube_dims, string, 2,
_otf2.CART_PERIODIC_FALSE)
string += 1
num_cube_dims += 1
cube_dim <<= 1
cube_dimensions = [
_otf2.CartDimensionRef(cube_dim) for cube_dim in range(num_cube_dims)
]
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Hyper Cube")
_otf2.GlobalDefWriter_WriteCartTopology(glob_def_writer, 0, string,
MPI_COMM_MPI_COMM_CUBE,
cube_dimensions)
string += 1
for rank in range(num_cube_ranks):
cube_coordinates = [
1 if rank & (2**cube_dim) else 0
for cube_dim in range(num_cube_dims)
]
_otf2.GlobalDefWriter_WriteCartCoordinate(glob_def_writer, 0, rank,
cube_coordinates)
_otf2.Archive_CloseGlobalDefWriter(archive, glob_def_writer)
# write local mappings for MPI communicators and metrics */
# write local mappings for metrics, this is an identity map, just to write
# something out
_otf2.Archive_OpenDefFiles(archive)
metric_map = _otf2.IdMap_Create(_otf2.ID_MAP_DENSE, NUM_OF_CLASSES)
for c in range(NUM_OF_CLASSES):
_otf2.IdMap_AddIdPair(metric_map, c, c)
for rank in range(num_ranks):
mpi_comm_map = _otf2.IdMap_Create(_otf2.ID_MAP_SPARSE, 2)
# Each location uses its rank as the communicator id which maps to the global 0 */
_otf2.IdMap_AddIdPair(mpi_comm_map, rank, MPI_COMM_MPI_COMM_WORLD)
_otf2.IdMap_AddIdPair(mpi_comm_map, rank + num_ranks,
MPI_COMM_MPI_COMM_SELF)
for thread in range(num_threads):
# Open a definition writer, so the file is created. */
def_writer = _otf2.Archive_GetDefWriter(
archive, locations[num_threads * rank + thread])
# Write metric mappings to local definitions. */
_otf2.DefWriter_WriteMappingTable(def_writer, _otf2.MAPPING_METRIC,
metric_map)
if otf2_MPI or otf2_HYBRID:
# Write MPI Comm mappings to local definitions. */
_otf2.DefWriter_WriteMappingTable(def_writer,
_otf2.MAPPING_COMM,
mpi_comm_map)
# Write clock offsets to local definitions. */
_otf2.DefWriter_WriteClockOffset(def_writer, 0, 0, 0.0)
_otf2.DefWriter_WriteClockOffset(def_writer, t, 0, 0.0)
_otf2.Archive_CloseDefWriter(archive, def_writer)
_otf2.IdMap_Free(mpi_comm_map)
_otf2.IdMap_Free(metric_map)
_otf2.Archive_CloseDefFiles(archive)
# Finalize. */
_otf2.Archive_Close(archive)
_otf2.AttributeList_Delete(attributes)
print("Done\n")
# definition IDs for metric classes and instances */
METRIC_CLASS_1 = 0
METRIC_CLASS_2 = 1
METRIC_CLASS_3 = 2
METRIC_CLASS_4 = 3
METRIC_INSTANCE_1 = 4
NUM_OF_CLASSES = 5
def next_value_attr_1():
next_value_attr_1.value.uint64 += 1
return next_value_attr_1.value
next_value_attr_1.value = _otf2.AttributeValue()
next_value_attr_1.value.uint64 = 12345
def next_value_attr_2():
next_value_attr_2.value.int16 += 1
return next_value_attr_2.value
next_value_attr_2.value = _otf2.AttributeValue()
next_value_attr_2.value.int16 = 23456
# attributes */
class AttributeDef:
def __init__(self, name, otf2_type, next_value):