def _create_trace(self):
trace = btw.Writer(self._trace_path)
clock = btw.Clock('test_clock')
trace.add_clock(clock)
integer_field_type = btw.IntegerFieldDeclaration(32)
event_class = btw.EventClass('simple_event')
event_class.add_field(integer_field_type, 'int_field')
stream_class = btw.StreamClass('empty_packet_stream')
stream_class.add_event_class(event_class)
stream_class.clock = clock
stream = trace.create_stream(stream_class)
for i in range(self._expected_event_count):
event = btw.Event(event_class)
event.payload('int_field').value = i
stream.append_event(event)
stream.flush()
# The CTF writer will not be able to populate the packet context's
# timestamp_begin and timestamp_end fields if it is asked to flush
# without any queued events.
with self.assertRaises(ValueError):
stream.flush()
packet_context = stream.packet_context
packet_context.field('timestamp_begin').value = 1
packet_context.field('timestamp_end').value = 123456
stream.flush()
def sched_switch(self, tid):
ClockManager().sample()
e = btw.Event(self.schedswitchEventClass)
e.payload("next_tid").value = tid
#ctx = btw.StructureField(self.stream_context_decl)
#ctx.field("pid").value = self.id
#e.stream_context = ctx
self.stream.append_event(e)
def apex_START(self, tid):
ClockManager().sample()
e = btw.Event(self.apexStart)
e.payload("tid").value = tid
#ctx = btw.StructureField(self.stream_context_decl)
#ctx.field("pid").value = self.id
#e.stream_context = ctx
self.stream.append_event(e)
def task_create(self, tid):
ClockManager().sample()
e = btw.Event(self.tcreateEventClass)
e.payload("tid").value = tid
#ctx = btw.StructureField(self.stream_context_decl)
#ctx.field("pid").value = self.id
#e.stream_context = ctx
self.stream.append_event(e)
def _write_trace(self):
trace = btw.Writer(self._trace_path)
clock = btw.Clock('test_clock')
trace.add_clock(clock)
integer_field_type = btw.IntegerFieldDeclaration(32)
event_class = btw.EventClass('simple_event')
event_class.add_field(integer_field_type, 'int_field')
stream_class = btw.StreamClass('test_stream')
stream_class.add_event_class(event_class)
stream_class.clock = clock
stream_class.packet_context_type = None
stream = trace.create_stream(stream_class)
for i in range(self._expected_event_count):
event = btw.Event(event_class)
event.payload('int_field').value = i
stream.append_event(event)
stream.flush()
# It is not valid for a stream to contain more than one packet
# if it does not have content_size/packet_size info in its packet
# context
event = btw.Event(event_class)
event.payload('int_field').value = 42
stream.append_event(event)
flush_raises = False
try:
stream.flush()
except ValueError:
flush_raises = True
self.assertTrue(flush_raises)
trace.flush_metadata()
def _create_trace(self, stream_descriptions):
trace_path = tempfile.mkdtemp()
trace = btw.Writer(trace_path)
clock = btw.Clock('test_clock')
clock.uuid = self._clock_uuid
trace.add_clock(clock)
integer_field_type = btw.IntegerFieldDeclaration(32)
event_class = btw.EventClass('simple_event')
event_class.add_field(integer_field_type, 'int_field')
stream_class = btw.StreamClass('test_stream')
stream_class.add_event_class(event_class)
stream_class.clock = clock
streams = []
stream_entries = []
for stream_id, stream_packets in enumerate(stream_descriptions):
stream = trace.create_stream(stream_class)
streams.append(stream)
for packet in stream_packets:
for timestamp in packet.timestamps:
stream_entries.append(Entry(stream_id, timestamp))
# Mark the last inserted entry as the end of packet
stream_entries[len(stream_entries) - 1].end_of_packet = True
# Sort stream entries which will provide us with a time-ordered list of
# events to insert in the streams.
for entry in sorted(stream_entries, key=lambda entry: entry.timestamp):
clock.time = entry.timestamp
event = btw.Event(event_class)
event.payload('int_field').value = entry.stream_id
streams[entry.stream_id].append_event(event)
if entry.end_of_packet is True:
streams[entry.stream_id].flush()
return trace_path
init_time = 0
pool_memory_allocate_events = {}
# OpenCL times to synchronize GPU times with CPU times
opencl_device_events = []
queued_times = []
end_times = []
enqueueNDRangeKernel_times = []
callback_times = []
for r_event in collection.events:
name = r_event.name
event_time = r_event.timestamp
fields = r_event.field_list_with_scope(
babeltrace.common.CTFScope.EVENT_FIELDS)
w_event = btw.Event(event_classes[name])
for f in fields:
# print(name, f, r_event[f])
if r_event[f] == "hsa_init":
init_time = event_time
# if hcTracer:kernel_* : fill the grid and groupworker arrays
if f == "workgroup_size" or f == "grid_size":
for i in range(3):
tmp = w_event.payload(f).field(i)
tmp.value = r_event[f][i]
continue
w_event.payload(f).value = r_event[f]
for event_class in event_classes.values():
main_stream_class.add_event_class(event_class)
main_stream = writer.create_stream(main_stream_class)
# Create events, based on event classes
init_time = None
cntol = 0
events = defaultdict(list)
driver_dic, runtime_dic = correlation_cbid()
for r_event in collection.events:
name = r_event.name
event_time = r_event.timestamp
w_event = btw.Event(event_classes[name])
fields = r_event.field_list_with_scope(
babeltrace.common.CTFScope.EVENT_FIELDS)
w_event = btw.Event(event_classes[name])
for f in fields:
# print(name, f, r_event[f])
w_event.payload(f).value = r_event[f]
if "queue" not in name and ("cudaTracer:kernel" in name
or "cudaTracer:memcpy" in name):
event_time = r_event["timestamp"] * 1000
# organize threads
threadId = r_event.field_with_scope(
event_class.add_field(enumeration_type, "enum_field")
# Create an array type
array_type = btw.ArrayFieldDeclaration(int10_type, 5)
event_class.add_field(array_type, "array_field")
# Create a sequence type
sequence_type = btw.SequenceFieldDeclaration(int32_type, "sequence_len")
event_class.add_field(uint16_type, "sequence_len")
event_class.add_field(sequence_type, "sequence_field")
stream_class.add_event_class(event_class)
stream = writer.create_stream(stream_class)
for i in range(100):
event = btw.Event(event_class)
clock.time = i * 1000
structure_field = event.payload("structure_field")
integer_field = structure_field.field("an_integer")
integer_field.value = i
string_field = structure_field.field("a_string_field")
string_field.value = "Test string."
float_field = event.payload("float_field")
float_field.value = float(i) + (float(i) / 100.0)
array_field = event.payload("array_field")
for j in range(5):
element = array_field.field(j)
def partSwitch(self, pid):
ClockManager().sample()
e = btw.Event(self.partSwitchEventClass)
e.payload("pid").value = pid
self.stream.append_event(e)
def syscall_entry(self, irq):
ClockManager().sample()
e = btw.Event(self.syscallEntryEventClass)
e.payload("id").value = irq
self.stream.append_event(e)
def irq_exit(self, irq):
ClockManager().sample()
e = btw.Event(self.irqExitEventClass)
e.payload("irq").value = irq
self.stream.append_event(e)
def create_ctf(trace):
# temporary directory holding the CTF trace
trace_path = tempfile.mkdtemp()
# our writer
writer = btw.Writer(trace_path)
# create one default clock and register it to the writer
clock = btw.Clock('my_clock')
clock.description = 'this is my clock'
#clock.frequency = 3570000000
writer.add_clock(clock)
# create our single stream
streams = []
event_types = []
for i in range(8):
event_types.append({})
print("iter ", i)
# create one default stream class and assign our clock to it
stream_class = btw.StreamClass('stream{}'.format(i))
stream_class.clock = clock
# create one default event class
event_class_cont = btw.EventClass('switch_infcont')
event_class_prev = btw.EventClass('switch_infprev')
event_class_next = btw.EventClass('switch_infnext')
event_class_hypercall = btw.EventClass('hypercall')
event_class_hypercall2 = btw.EventClass('hypercall2')
event_class_hypercall_version = btw.EventClass('hypercall_version')
event_class_hypercall_version_return = btw.EventClass('hypercall_version_return')
event_class_running_runnable = btw.EventClass('running_to_runnable')
event_class_runnable_running = btw.EventClass('runnable_to_running')
event_class_running_block = btw.EventClass('running_to_blocked')
event_class_blocked_runnable = btw.EventClass('blocked_to_runnable')
# create one 32-bit signed integer field
int32_field_decl = btw.IntegerFieldDeclaration(32)
int32_field_decl.signed = True
# create one 32-bit signed integer field
int64_field_decl = btw.IntegerFieldDeclaration(64)
int64_field_decl.signed = True
# add this field declaration to our event class
event_class_blocked_runnable.add_field(int32_field_decl, 'dom')
event_class_blocked_runnable.add_field(int32_field_decl, 'vcpu')
event_class_running_block.add_field(int32_field_decl, 'dom')
event_class_running_block.add_field(int32_field_decl, 'vcpu')
event_class_runnable_running.add_field(int32_field_decl, 'dom')
event_class_runnable_running.add_field(int32_field_decl, 'vcpu')
event_class_running_runnable.add_field(int32_field_decl, 'dom')
event_class_running_runnable.add_field(int32_field_decl, 'vcpu')
event_class_cont.add_field(int32_field_decl, 'dom')
event_class_cont.add_field(int32_field_decl, 'vcpu')
event_class_prev.add_field(int32_field_decl, 'dom')
event_class_prev.add_field(int32_field_decl, 'vcpu')
event_class_next.add_field(int32_field_decl, 'dom')
event_class_next.add_field(int32_field_decl, 'vcpu')
event_class_hypercall.add_field(int32_field_decl, 'op')
event_class_hypercall2.add_field(int32_field_decl, 'op')
event_class_hypercall_version.add_field(int32_field_decl, 'id')
event_class_hypercall_version_return.add_field(int32_field_decl, 'id')
event_types[i]['switch_infcont'] = event_class_cont
event_types[i]['switch_infprev'] = event_class_prev
event_types[i]['switch_infnext'] = event_class_next
event_types[i]['hypercall'] = event_class_hypercall
event_types[i]['hypercall2'] = event_class_hypercall2
event_types[i]['hypercall_version'] = event_class_hypercall_version
event_types[i]['hypercall_version_return'] = event_class_hypercall_version_return
event_types[i]['blocked_to_runnable'] = event_class_blocked_runnable
event_types[i]['running_to_blocked'] = event_class_running_block
event_types[i]['runnable_to_running'] = event_class_runnable_running
event_types[i]['running_to_runnable'] = event_class_running_runnable
# register our event class to our stream class
stream_class.add_event_class(event_class_cont)
stream_class.add_event_class(event_class_prev)
stream_class.add_event_class(event_class_next)
stream_class.add_event_class(event_class_hypercall)
stream_class.add_event_class(event_class_hypercall2)
stream_class.add_event_class(event_class_hypercall_version)
stream_class.add_event_class(event_class_hypercall_version_return)
stream_class.add_event_class(event_class_running_block)
stream_class.add_event_class(event_class_blocked_runnable)
stream_class.add_event_class(event_class_running_runnable)
stream_class.add_event_class(event_class_runnable_running)
stream_class.packet_context_type.add_field(int32_field_decl, 'cpu_id')
streams.append(writer.create_stream(stream_class))
streams[-1].packet_context.field('cpu_id').value = i
used_streams = set()
done = False
while not done:
time = 2 ** 64
e = None
selected_s = None
for s in trace:
if len(s) > 0:
temp = s[0]
if temp["timestamp"] < time:
e = temp
time = e["timestamp"]
selected_s = s
if e is None:
done = True
break
else:
e = selected_s.pop(0)
event = None
cpu = e["cpu"]
if e["name"] == "switch_infcont":
event = btw.Event(event_types[cpu]["switch_infcont"])
event.payload('dom').value = e["dom"]
event.payload('vcpu').value = e["vcpu"]
elif e["name"] == "switch_infnext":
event = btw.Event(event_types[cpu]["switch_infnext"])
event.payload('dom').value = e["dom"]
event.payload('vcpu').value = e["vcpu"]
elif e["name"] == "switch_infprev":
event = btw.Event(event_types[cpu]["switch_infprev"])
event.payload('dom').value = e["dom"]
event.payload('vcpu').value = e["vcpu"]
elif e["name"] == "blocked_to_runnable":
event = btw.Event(event_types[cpu]["blocked_to_runnable"])
event.payload('dom').value = e["dom"]
event.payload('vcpu').value = e["vcpu"]
elif e["name"] == "running_to_blocked":
event = btw.Event(event_types[cpu]["running_to_blocked"])
event.payload('dom').value = e["dom"]
event.payload('vcpu').value = e["vcpu"]
elif e["name"] == "runnable_to_running":
event = btw.Event(event_types[cpu]["runnable_to_running"])
event.payload('dom').value = e["dom"]
event.payload('vcpu').value = e["vcpu"]
elif e["name"] == "running_to_runnable":
event = btw.Event(event_types[cpu]["running_to_runnable"])
event.payload('dom').value = e["dom"]
event.payload('vcpu').value = e["vcpu"]
elif e["name"] == "hypercall_op":
class_ = event_types[cpu]["hypercall"]
event = btw.Event(event_types[cpu]["hypercall"])
#print("cpu={}' class={} ev={}".format(cpu, class_, event))
event.payload('op').value = e["op"]
elif e["name"] == "hypercall_op2":
event = btw.Event(event_types[cpu]["hypercall2"])
event.payload('op').value = e["op"]
elif e["name"] == "hypercall_version":
event = btw.Event(event_types[cpu]["hypercall_version"])
event.payload('id').value = e["id"]
used_streams.add(cpu)
stream = streams[cpu]
if event != None:
clock.time = to_ns(e["timestamp"])
stream.append_event(event)
if e["name"] == "hypercall_version":
event = btw.Event(event_types[cpu]["hypercall_version_return"])
event.payload('id').value = e["id"]
stream.append_event(event)
# flush the stream
print('trace path: {}'.format(trace_path))
for s in used_streams:
streams[s].flush()