def time_function (width, height, repeat, max_tokens, max_threads, grain_size):
# run_application(0, width, height, repeat)
time0 = run_application(0, width, height, repeat, max_tokens, max_threads, grain_size)
time1 = run_application(1, width, height, repeat, max_tokens, max_threads, grain_size)
timediff = (time1 - time0)
bandwidth = (10 * repeat * width * height * 2 / timediff)
bandwidth_string = op.get_bytes(bandwidth) + '/s'
return bandwidth
def run_the_function(print_result, height, width, repeat, text_file_name, grain_size):
path_name= "./data/KnifeQuadBPos1_2_21_int16.h5"
top_level_data_set_name= "KnifeQuadBPos1/"
host_name = socket.gethostname()
#Program to execute
debug_version = './Debug/cppProcessing2.0 '
profile_version = './Profiling/cppProcessing2.0 '
parallel_debug = './parallel_debug/cppProcessing2.0 '
parallel_profile = './parallel_profile/cppProcessing2.0 '
cmdl_arg = '1 ' + str(width) + ' '+ str(height) + ' ' + str(repeat) + ' ' + text_file_name + ' ' + str(grain_size) + ' 1'
program_to_execute = parallel_profile + cmdl_arg
#events to monitor
#instructions
event1 = op.oprofile_events('CPU_CLK_UNHALTED','0x00',100000000)
event2 = op.oprofile_events('INST_RETIRED','0x00',60000000)
#cache misses
event3 = op.oprofile_events('LLC_MISSES','0x41',60000) #divide by LLC_REFS
# event4 = op.oprofile_events('l2_lines_in','0x07',1000000) #100000
# event5 = op.oprofile_events('br_inst_retired', '0x01', 400000) #total branch instructions retired
# event6 = op.oprofile_events('br_misp_retired', '0x01', 400000) #total mispredicted branches. Divide by br_inst_retired
event6 = op.oprofile_events('mem_trans_retired','0x02',2000000)
# event7 = op.oprofile_events('uops_retired', 'stall_cycles',2000000) #no of stall cycles. Divide by cpu cycles
#
# event8 = op.oprofile_events('dtlb_load_misses', '0x01',2000000)
# # event8 = op.oprofile_events('dtlb_load_misses', '0x81',1000) #Ivy Bridge
#
# event9 = op.oprofile_events('LLC_REFS', '0x4f',6000)
#
# event10 = op.oprofile_events('l1d_pend_miss', 'pending',2000000) #cycles of l1d misses outstanding. Divide by CPU cycles
# event11 = op.oprofile_events('resource_stalls', '0x01',2000000) #no of stall cycles/divide by number of instructions
# event12 = op.oprofile_events('l1d', '0x01',2000000) #cycles of l1d misses outstanding. Divide by CPU cycles
list_of_events = [event1, event2, event3, event6]#, event7, event8, event9, event10, event11, event12]
#variable initialisation
dict_of_attributes = {}
total_time = 0.0
dict_of_function_perc_time = {}
list_of_events_recorded = []
image_size = width * height * 2 #memory size
total_processing_time = 0.0
operf_events = op.get_operf_option(list_of_events)
result_directory = './oprof_reports/'+ host_name + '/'
report_destination = result_directory + 'profile_report.txt'
sample_data_destination = result_directory + 'oprof_data/'
#commandline options
cmd_git = 'git rev-parse HEAD > ' + report_destination
cmd_time = '(/usr/bin/time -v ' + program_to_execute + ') &>> ' + report_destination
cmd_operf = 'operf ' + '-d ' + sample_data_destination + ' ' + operf_events + ' '+ program_to_execute
cmd_opreport = 'opreport --symbols --session-dir=' + sample_data_destination + ' >> ' + report_destination
cmd_mkdir = 'mkdir -p ' + sample_data_destination
cmd_annotate = 'opannotate -s --output-dir=' + result_directory + 'annotated/ ' + '--session-dir=' + sample_data_destination + ' '+ program_to_execute
op.cmd_call(cmd_git)
op.cmd_call(cmd_time)
op.cmd_call(cmd_operf)
op.cmd_call(cmd_opreport)
op.cmd_call(cmd_mkdir)
op.cmd_call(cmd_annotate)
f = open(report_destination, 'r')
list_of_functions = op.get_function_list(f)
f.close()
for function_name in list_of_functions:
dict_of_function_perc_time[function_name] = 0
dict_of_attributes[function_name] = [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
f = open(report_destination, 'r')
s = f.readline()
commit_key = s;
while s != '':
if 'Counted' in s:
for event in list_of_events:
if (event.event_name + ' events') in s:
list_of_events_recorded.append(event)
if 'Elapsed (wall clock) time ' in s:
total_time = op.parse_time(s)
for function_name in list_of_functions:
if function_name in s:
delimited = s.split(' ')
parsed = [item for item in delimited if item != '']
attributes = []
dict_of_function_perc_time[function_name] = float(parsed[1]) + dict_of_function_perc_time[function_name]
for index in xrange(len(list_of_events_recorded)): # manually add the percentage clock cycles
attributes.append( float(parsed[index * 2]) + dict_of_attributes[function_name][index] )
dict_of_attributes[function_name] = attributes
s = f.readline()
llc_misses_per_instruction = op.get_llc_misses(list_of_functions, dict_of_attributes, list_of_events_recorded)
CPI = op.get_CPI(list_of_functions, dict_of_attributes, list_of_events_recorded)
bandwidth = op.get_bandwidth(list_of_functions, dict_of_function_perc_time, total_time, image_size, repeat)
function_time = op.get_time(list_of_functions, dict_of_function_perc_time, total_time, repeat)
l1d_miss_rate = op.get_L1D_miss_rate(list_of_functions, dict_of_attributes, list_of_events_recorded)
DTLB_miss_rate = op.get_DTLB_miss_rate(list_of_functions, dict_of_attributes, list_of_events_recorded)
LLC_miss_rate = op.get_LLC_miss_rate(list_of_functions, dict_of_attributes, list_of_events_recorded)
br_misspre_rate = op.get_br_mispre_rate(list_of_functions, dict_of_attributes, list_of_events_recorded)
resource_stall = op.get_resource_stall_rate(list_of_functions, dict_of_attributes, list_of_events_recorded)
l2_miss_rate = op.get_L2_miss_rate(list_of_functions, dict_of_attributes, list_of_events_recorded)
l1d_repl_rate = op.get_L1D_repl_rate(list_of_functions, dict_of_attributes, list_of_events_recorded)
memory_bandwidth = op.get_memory_bandwidth(list_of_functions, dict_of_attributes, list_of_events_recorded)
for name, perc_time in dict_of_function_perc_time.iteritems():
total_processing_time = total_processing_time + perc_time * total_time /1000 /100 #in seconds
#printing reports
if print_result == True:
print
print datetime.datetime.now().date(),
print datetime.datetime.now().time(),
print 'Report:'
print cmd_time
print cmd_operf
print cmd_opreport
print 'oprofile sample directory: ' + sample_data_destination
print cmd_annotate
print '=' * 100
print commit_key,
print 'operf ' + '-d ' + sample_data_destination + ' ' + operf_events + ' '+ program_to_execute
print 'The program took {0:.4} ms in total. {1:.4} ms per sample/reset pair.'.format(total_time, total_time/repeat)
print 'Of which the processing functions took in total {0:.4} ms to run.'.format(total_processing_time)
print 'Image size {0:d} ({1:d} * {2:d}) pixels.'.format(width * height, height, width),
print op.get_bytes(image_size)
print 'Statistics collected for '+str(repeat)+' iterations'
print 'Bandwidth:'
print '\t' + 'Total: {0:.4} MB/s'.format(op.get_bytes(image_size * 2 * repeat/total_processing_time) ) #2 because of sample and reset
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0}'.format(op.get_bytes((bandwidth[index]))) + '/s'
print 'LLC misses per instruction:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2%}'.format(llc_misses_per_instruction[index])
print 'Cycle per instruction (CPI):'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2}'.format(CPI[index])
print 'Processing time (ms) for each call:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0}'.format(int(function_time[index]))
if len(l1d_miss_rate) != 0:
print 'L1D misses percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2%}'.format(l1d_miss_rate[index])
if len(l1d_repl_rate) != 0:
print 'L1D replacement rate:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2%}'.format(l1d_repl_rate[index])
if len(l2_miss_rate) != 0:
print 'L2 misses percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2%}'.format(l2_miss_rate[index])
if len(DTLB_miss_rate) != 0:
print 'DTLB miss per instruction:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2}'.format(DTLB_miss_rate[index])
if len(LLC_miss_rate) != 0:
print 'LLC miss rate:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2%}'.format(LLC_miss_rate[index])
if len(br_misspre_rate) != 0:
print 'Branch misprediction percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2%}'.format(br_misspre_rate[index])
if len(resource_stall) != 0:
print 'Resource stall cycle percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + '{0:.2%}'.format(resource_stall[index])
if len(memory_bandwidth) != 0:
print 'Memory Bandwidth'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function +':' + op.get_bytes(memory_bandwidth[index])
print '=' * 100
return image_size * 2 * repeat/total_processing_time
print grain_size_file_name
cmd_rm_file = "rm -f " + grain_size_file_name
cmd_create_file = "touch "+ grain_size_file_name
op.cmd_call(cmd_rm_file)
op.cmd_call(cmd_create_file)
# for max_tokens in range(10,50,5):
#for grain_size in (3528/7,3528/6, 3528/4, 3528/3,3528/2,3528,3528*3, 3528*7,3528*9, 3528*21, 3528*59, 3528*63, 3528*177,3528*413,3528*531,3528*1239):
# for multiple in range(20, 100, 5):
# grain_size = multiple * 3528
for max_threads in range(1,25,1):
a = op.accumulator(grain_size_file_name, [max_threads, max_tokens, grain_size], 10)
for repeats in xrange(1,3):
bandwidth = time_function(width, height, repeat, max_tokens, max_threads, grain_size)
a.add_number(bandwidth)
print grain_size, op.get_bytes(bandwidth)
a.average()
# a = op.accumulator(grain_size_file_name, 1, 10)
# for repeats in xrange(1,10):
# bandwidth = time_function(width, height, repeat, max_tokens, max_threads, grain_size)
# a.add_number(bandwidth)
# print grain_size, op.get_bytes(bandwidth)
# a.average()
print bandwidth
###############################################################################################################
def run_the_function(print_result, height, width, repeat, text_file_name,
grain_size):
path_name = "./data/KnifeQuadBPos1_2_21_int16.h5"
top_level_data_set_name = "KnifeQuadBPos1/"
host_name = socket.gethostname()
#Program to execute
debug_version = './Debug/cppProcessing2.0 '
profile_version = './Profiling/cppProcessing2.0 '
parallel_debug = './parallel_debug/cppProcessing2.0 '
parallel_profile = './parallel_profile/cppProcessing2.0 '
cmdl_arg = '1 ' + str(width) + ' ' + str(height) + ' ' + str(
repeat) + ' ' + text_file_name + ' ' + str(grain_size) + ' 1'
program_to_execute = parallel_profile + cmdl_arg
#events to monitor
#instructions
event1 = op.oprofile_events('CPU_CLK_UNHALTED', '0x00', 100000000)
event2 = op.oprofile_events('INST_RETIRED', '0x00', 60000000)
#cache misses
event3 = op.oprofile_events('LLC_MISSES', '0x41',
60000) #divide by LLC_REFS
# event4 = op.oprofile_events('l2_lines_in','0x07',1000000) #100000
# event5 = op.oprofile_events('br_inst_retired', '0x01', 400000) #total branch instructions retired
# event6 = op.oprofile_events('br_misp_retired', '0x01', 400000) #total mispredicted branches. Divide by br_inst_retired
event6 = op.oprofile_events('mem_trans_retired', '0x02', 2000000)
# event7 = op.oprofile_events('uops_retired', 'stall_cycles',2000000) #no of stall cycles. Divide by cpu cycles
#
# event8 = op.oprofile_events('dtlb_load_misses', '0x01',2000000)
# # event8 = op.oprofile_events('dtlb_load_misses', '0x81',1000) #Ivy Bridge
#
# event9 = op.oprofile_events('LLC_REFS', '0x4f',6000)
#
# event10 = op.oprofile_events('l1d_pend_miss', 'pending',2000000) #cycles of l1d misses outstanding. Divide by CPU cycles
# event11 = op.oprofile_events('resource_stalls', '0x01',2000000) #no of stall cycles/divide by number of instructions
# event12 = op.oprofile_events('l1d', '0x01',2000000) #cycles of l1d misses outstanding. Divide by CPU cycles
list_of_events = [event1, event2, event3, event6
] #, event7, event8, event9, event10, event11, event12]
#variable initialisation
dict_of_attributes = {}
total_time = 0.0
dict_of_function_perc_time = {}
list_of_events_recorded = []
image_size = width * height * 2 #memory size
total_processing_time = 0.0
operf_events = op.get_operf_option(list_of_events)
result_directory = './oprof_reports/' + host_name + '/'
report_destination = result_directory + 'profile_report.txt'
sample_data_destination = result_directory + 'oprof_data/'
#commandline options
cmd_git = 'git rev-parse HEAD > ' + report_destination
cmd_time = '(/usr/bin/time -v ' + program_to_execute + ') &>> ' + report_destination
cmd_operf = 'operf ' + '-d ' + sample_data_destination + ' ' + operf_events + ' ' + program_to_execute
cmd_opreport = 'opreport --symbols --session-dir=' + sample_data_destination + ' >> ' + report_destination
cmd_mkdir = 'mkdir -p ' + sample_data_destination
cmd_annotate = 'opannotate -s --output-dir=' + result_directory + 'annotated/ ' + '--session-dir=' + sample_data_destination + ' ' + program_to_execute
op.cmd_call(cmd_git)
op.cmd_call(cmd_time)
op.cmd_call(cmd_operf)
op.cmd_call(cmd_opreport)
op.cmd_call(cmd_mkdir)
op.cmd_call(cmd_annotate)
f = open(report_destination, 'r')
list_of_functions = op.get_function_list(f)
f.close()
for function_name in list_of_functions:
dict_of_function_perc_time[function_name] = 0
dict_of_attributes[function_name] = [
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0
]
f = open(report_destination, 'r')
s = f.readline()
commit_key = s
while s != '':
if 'Counted' in s:
for event in list_of_events:
if (event.event_name + ' events') in s:
list_of_events_recorded.append(event)
if 'Elapsed (wall clock) time ' in s:
total_time = op.parse_time(s)
for function_name in list_of_functions:
if function_name in s:
delimited = s.split(' ')
parsed = [item for item in delimited if item != '']
attributes = []
dict_of_function_perc_time[function_name] = float(
parsed[1]) + dict_of_function_perc_time[function_name]
for index in xrange(
len(list_of_events_recorded
)): # manually add the percentage clock cycles
attributes.append(
float(parsed[index * 2]) +
dict_of_attributes[function_name][index])
dict_of_attributes[function_name] = attributes
s = f.readline()
llc_misses_per_instruction = op.get_llc_misses(list_of_functions,
dict_of_attributes,
list_of_events_recorded)
CPI = op.get_CPI(list_of_functions, dict_of_attributes,
list_of_events_recorded)
bandwidth = op.get_bandwidth(list_of_functions, dict_of_function_perc_time,
total_time, image_size, repeat)
function_time = op.get_time(list_of_functions, dict_of_function_perc_time,
total_time, repeat)
l1d_miss_rate = op.get_L1D_miss_rate(list_of_functions, dict_of_attributes,
list_of_events_recorded)
DTLB_miss_rate = op.get_DTLB_miss_rate(list_of_functions,
dict_of_attributes,
list_of_events_recorded)
LLC_miss_rate = op.get_LLC_miss_rate(list_of_functions, dict_of_attributes,
list_of_events_recorded)
br_misspre_rate = op.get_br_mispre_rate(list_of_functions,
dict_of_attributes,
list_of_events_recorded)
resource_stall = op.get_resource_stall_rate(list_of_functions,
dict_of_attributes,
list_of_events_recorded)
l2_miss_rate = op.get_L2_miss_rate(list_of_functions, dict_of_attributes,
list_of_events_recorded)
l1d_repl_rate = op.get_L1D_repl_rate(list_of_functions, dict_of_attributes,
list_of_events_recorded)
memory_bandwidth = op.get_memory_bandwidth(list_of_functions,
dict_of_attributes,
list_of_events_recorded)
for name, perc_time in dict_of_function_perc_time.iteritems():
total_processing_time = total_processing_time + perc_time * total_time / 1000 / 100 #in seconds
#printing reports
if print_result == True:
print
print datetime.datetime.now().date(),
print datetime.datetime.now().time(),
print 'Report:'
print cmd_time
print cmd_operf
print cmd_opreport
print 'oprofile sample directory: ' + sample_data_destination
print cmd_annotate
print '=' * 100
print commit_key,
print 'operf ' + '-d ' + sample_data_destination + ' ' + operf_events + ' ' + program_to_execute
print 'The program took {0:.4} ms in total. {1:.4} ms per sample/reset pair.'.format(
total_time, total_time / repeat)
print 'Of which the processing functions took in total {0:.4} ms to run.'.format(
total_processing_time)
print 'Image size {0:d} ({1:d} * {2:d}) pixels.'.format(
width * height, height, width),
print op.get_bytes(image_size)
print 'Statistics collected for ' + str(repeat) + ' iterations'
print 'Bandwidth:'
print '\t' + 'Total: {0:.4} MB/s'.format(
op.get_bytes(
image_size * 2 * repeat /
total_processing_time)) #2 because of sample and reset
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0}'.format(
op.get_bytes((bandwidth[index]))) + '/s'
print 'LLC misses per instruction:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2%}'.format(
llc_misses_per_instruction[index])
print 'Cycle per instruction (CPI):'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2}'.format(CPI[index])
print 'Processing time (ms) for each call:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0}'.format(
int(function_time[index]))
if len(l1d_miss_rate) != 0:
print 'L1D misses percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2%}'.format(
l1d_miss_rate[index])
if len(l1d_repl_rate) != 0:
print 'L1D replacement rate:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2%}'.format(
l1d_repl_rate[index])
if len(l2_miss_rate) != 0:
print 'L2 misses percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2%}'.format(
l2_miss_rate[index])
if len(DTLB_miss_rate) != 0:
print 'DTLB miss per instruction:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2}'.format(
DTLB_miss_rate[index])
if len(LLC_miss_rate) != 0:
print 'LLC miss rate:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2%}'.format(
LLC_miss_rate[index])
if len(br_misspre_rate) != 0:
print 'Branch misprediction percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2%}'.format(
br_misspre_rate[index])
if len(resource_stall) != 0:
print 'Resource stall cycle percentage:'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + '{0:.2%}'.format(
resource_stall[index])
if len(memory_bandwidth) != 0:
print 'Memory Bandwidth'
for function in list_of_functions:
index = list_of_functions.index(function)
print '\t' + function + ':' + op.get_bytes(
memory_bandwidth[index])
print '=' * 100
return image_size * 2 * repeat / total_processing_time