def trace_unhandled(event_name, context, event_fields_dict, perf_sample_dict):
global unhandled_enter_secs, unhandled_enter_nsecs
global unhandled_infos
if event_name not in unhandled_enter_nsecs:
unhandled_enter_nsecs[event_name] = {}
unhandled_enter_secs[event_name] = {}
unhandled_infos[event_name] = {}
common_pid = event_fields_dict["common_pid"]
common_secs = event_fields_dict["common_s"]
common_nsecs = event_fields_dict["common_ns"]
if "exit__return" not in event_name:
unhandled_enter_nsecs[event_name][common_pid] = common_nsecs
unhandled_enter_secs[event_name][common_pid] = common_secs
else:
func_name = event_name.replace("_exit__return", "")
start_ns = Duration.nanoseconds(
unhandled_enter_secs[func_name][common_pid],
unhandled_enter_nsecs[func_name][common_pid])
end_ns = Duration.nanoseconds(common_secs, common_nsecs)
if common_pid not in unhandled_infos[func_name]:
unhandled_infos[func_name][common_pid] = []
unhandled_infos[func_name][common_pid].append(
TransferInfo(start_ns, end_ns, 0, 0, func_name))
def probe_libdpu__libdpu_dpu_copy_to_wram_for_dpu_exit__return(
event_name, context, common_cpu, common_secs, common_nsecs, common_pid,
common_comm, common_callchain, __probe_func, __probe_ret_ip,
perf_sample_dict):
global wram_enter_nsecs, wram_enter_secs, wram_pid
size = wram_pid[common_pid][0]
path = rank_path[wram_pid[common_pid][1]]
start_ns = Duration.nanoseconds(wram_enter_secs[common_pid],
wram_enter_nsecs[common_pid])
end_ns = Duration.nanoseconds(common_secs, common_nsecs)
wram_write_infos.setdefault(common_pid, []).append(
TransferInfo(start_ns, end_ns, size, path, "dpu_copy_to_wram_for_dpu"))
def probe_libdpu__libdpu_dpu_copy_from_wram_for_matrix_exit__return(
event_name, context, common_cpu, common_secs, common_nsecs, common_pid,
common_comm, common_callchain, __probe_func, __probe_ret_ip,
perf_sample_dict):
global wram_enter_nsecs, wram_enter_secs, wram_pid, wram_xfer_size
start_ns = Duration.nanoseconds(wram_enter_secs[common_pid],
wram_enter_nsecs[common_pid])
end_ns = Duration.nanoseconds(common_secs, common_nsecs)
path = rank_path[wram_pid[common_pid][1]]
nr_dpus = len(
wram_xfer_size[common_pid]) if common_pid in wram_xfer_size else 64
size = wram_pid[common_pid][0] * nr_dpus
wram_read_infos.setdefault(common_pid, []).append(
TransferInfo(start_ns, end_ns, size, path,
"dpu_copy_from_wram_{}dpus".format(nr_dpus)))
def probe_libdpu__libdpu_dpu_copy_from_mrams_exit__return(
event_name, context, common_cpu, common_secs, common_nsecs, common_pid,
common_comm, common_callchain, __probe_func, __probe_ret_ip,
perf_sample_dict):
global mram_enter_secs, mram_enter_nsecs, mram_pid, mram_xfer_size
global mram_read_infos
global rank_path
transfer_matrix = mram_pid[common_pid][0]
path = rank_path[mram_pid[common_pid][1]]
xfer_size = mram_pid[common_pid][2]
start_ns = Duration.nanoseconds(
mram_enter_secs[common_pid][transfer_matrix],
mram_enter_nsecs[common_pid][transfer_matrix])
end_ns = Duration.nanoseconds(common_secs, common_nsecs)
nb_dpus = len(mram_xfer_size[transfer_matrix])
xfer_size *= nb_dpus
mram_read_infos.setdefault(common_pid, []).append(
TransferInfo(start_ns, end_ns, xfer_size, path,
"dpu_copy_from_mrams_{}dpus".format(nb_dpus)))
def print_transfer_info(size,
func_name,
total_duration,
nb_size,
iram_bandwidth=False):
average_duration = Duration.seconds(total_duration) / nb_size
print_duration = PrintResult.format_duration(total_duration / nb_size)
print_size = PrintResult.format_size(size, iram_bandwidth)
print_bandwidth = PrintResult.format_size(size / average_duration,
iram_bandwidth)
print("{}\t{}\t{}/s\t{}\t{}".format(print_size.center(10),
str(nb_size).center(10),
print_bandwidth.center(10),
print_duration.center(10),
" ".join(func_name)))