def gen_all_traces(array_h=4,
array_w=4,
ifmap_h=7,
ifmap_w=7,
filt_h=3,
filt_w=3,
num_channels=3,
strides=1,
num_filt=8,
data_flow='os',
word_size_bytes=1,
filter_sram_size=64,
ifmap_sram_size=64,
ofmap_sram_size=64,
filt_base=1000000,
ifmap_base=0,
ofmap_base=2000000,
sram_read_trace_file="sram_read.csv",
sram_write_trace_file="sram_write.csv",
dram_filter_trace_file="dram_filter_read.csv",
dram_ifmap_trace_file="dram_ifmap_read.csv",
dram_ofmap_trace_file="dram_ofmap_write.csv"):
sram_cycles = 0
util = 0
print("Generating traces and bw numbers")
if data_flow == 'os':
sram_cycles, util = \
sram.sram_traffic(
dimension_rows= array_h,
dimension_cols= array_w,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt,
filt_base=filt_base, ifmap_base=ifmap_base,
ofmap_base = ofmap_base,
sram_read_trace_file=sram_read_trace_file,
sram_write_trace_file=sram_write_trace_file
)
elif data_flow == 'ws':
sram_cycles, util = \
sram_ws.sram_traffic(
dimension_rows = array_h,
dimension_cols = array_w,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
strides = strides, num_filt = num_filt,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file,
sram_write_trace_file = sram_write_trace_file
)
elif data_flow == 'is':
sram_cycles, util = \
sram_is.sram_traffic(
dimension_rows = array_h,
dimension_cols = array_w,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
strides = strides, num_filt = num_filt,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file,
sram_write_trace_file = sram_write_trace_file
)
#print("Generating DRAM traffic")
dram.dram_trace_read_v2(
sram_sz=ifmap_sram_size,
word_sz_bytes=word_size_bytes,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_ifmap_trace_file,
)
#print("IFMAP SRAM Size: \t" + str(ifmap_sram_size))
dram.dram_trace_read_v2(
sram_sz=filter_sram_size,
word_sz_bytes=word_size_bytes,
min_addr=filt_base,
max_addr=ofmap_base,
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_filter_trace_file,
)
dram.dram_trace_write(ofmap_sram_size=ofmap_sram_size,
data_width_bytes=word_size_bytes,
sram_write_trace_file=sram_write_trace_file,
dram_write_trace_file=dram_ofmap_trace_file)
print("Average utilization : \t" + str(util) + " %")
print("Cycles for compute : \t" + str(sram_cycles) + " cycles")
bw_numbers, detailed_log = gen_bw_numbers(dram_ifmap_trace_file,
dram_filter_trace_file,
dram_ofmap_trace_file,
sram_write_trace_file,
sram_read_trace_file)
#array_h, array_w)
return bw_numbers, detailed_log, util, sram_cycles
def gen_all_traces_dramsim_format(
array_h=4,
array_w=4,
ifmap_h=7,
ifmap_w=7,
filt_h=3,
filt_w=3,
num_channels=3,
strides=1,
num_filt=8,
data_flow='os',
# word_size_bytes=1,
dram_read_bw=8,
dram_write_bw=8,
dram_read_wsb=1,
dram_write_wsb=1,
filter_sram_size=64,
ifmap_sram_size=64,
ofmap_sram_size=64,
filt_base=1000000,
ifmap_base=0,
ofmap_base=2000000,
sram_read_trace_file="sram_read.csv",
sram_write_trace_file="sram_write.csv",
dram_filter_trace_file="dram_filter_read.csv",
dram_ifmap_trace_file="dram_ifmap_read.csv",
dram_ofmap_trace_file="dram_ofmap_write.csv"):
sram_cycles = 0
util = 0
print("Generating traces and bw numbers")
if data_flow == 'os':
sram_cycles, util = \
sram.sram_traffic(
dimension_rows=array_h,
dimension_cols=array_w,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt,
filt_base=filt_base, ifmap_base=ifmap_base,
ofmap_base=ofmap_base,
sram_read_trace_file=sram_read_trace_file,
sram_write_trace_file=sram_write_trace_file
)
elif data_flow == 'ws':
sram_cycles, util = \
sram_ws.sram_traffic(
dimension_rows=array_h,
dimension_cols=array_w,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt,
ofmap_base=ofmap_base, filt_base=filt_base, ifmap_base=ifmap_base,
sram_read_trace_file=sram_read_trace_file,
sram_write_trace_file=sram_write_trace_file
)
elif data_flow == 'is':
sram_cycles, util = \
sram_is.sram_traffic(
dimension_rows=array_h,
dimension_cols=array_w,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt,
ofmap_base=ofmap_base, filt_base=filt_base, ifmap_base=ifmap_base,
sram_read_trace_file=sram_read_trace_file,
sram_write_trace_file=sram_write_trace_file
)
# print("Generating DRAM traffic")
dramsim_format.dram_trace_read_dramsim(
sram_sz=ifmap_sram_size,
word_sz_bytes=dram_read_wsb,
default_read_bw=dram_read_bw,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_ifmap_trace_file,
)
dramsim_format.dram_trace_read_dramsim(
sram_sz=filter_sram_size,
word_sz_bytes=dram_read_wsb,
default_read_bw=dram_read_bw,
min_addr=filt_base,
max_addr=ofmap_base,
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_filter_trace_file,
)
dramsim_format.dram_trace_write_dramsim(
ofmap_sram_size=ofmap_sram_size,
data_width_bytes=dram_write_wsb,
default_write_bw=dram_write_bw,
sram_write_trace_file=sram_write_trace_file,
dram_write_trace_file=dram_ofmap_trace_file)
print("Average utilization : \t" + str(util) + " %")
print("Cycles for compute : \t" + str(sram_cycles) + " cycles")
print(
"Fred [WARNING]: The bw stats part are not taken care, yet to be correct by parsing dramsim format dram "
"traces")
# bw_numbers, detailed_log = gen_bw_numbers(dram_ifmap_trace_file, dram_filter_trace_file,
# dram_ofmap_trace_file, sram_write_trace_file,
# sram_read_trace_file)
# array_h, array_w)
return util, sram_cycles
def gen_all_traces(array_h=4,
array_w=4,
ifmap_h=7,
ifmap_w=7,
filt_h=3,
filt_w=3,
num_channels=3,
strides=1,
num_filt=8,
data_flow='os',
layer_tag="Conv1",
word_size_bytes=1,
filter_sram_size=64,
ifmap_sram_size=64,
ofmap_sram_size=64,
filt_base=1000000,
ifmap_base=0,
ofmap_base=2000000,
sram_read_trace_file="sram_read.csv",
sram_write_trace_file="sram_write.csv",
dram_filter_trace_file="dram_filter_read.csv",
dram_ifmap_trace_file="dram_ifmap_read.csv",
dram_ofmap_trace_file="dram_ofmap_write.csv"):
sram_cycles = 0
util = 0
tag = layer_tag + "_" + str(array_h) + "_" + str(array_w)
# Anand: Removing the following for SILENT mode
#print("Generating traces and bw numbers")
if data_flow == 'os':
tag_match = False
if lookup_flag:
sram_tag = tag + "_os"
tag_match, dir_name = lut.lookup(sram_tag)
if tag_match:
#this_dir = lut.to_string(subprocess.check_output(['pwd']))
this_dir = os.getcwd()
sram_cycles, util = lut.get_sram_stats(dir_name, dest_dir=this_dir)
else:
sram_cycles, util = \
sram.sram_traffic(
dimension_rows= array_h,
dimension_cols= array_w,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt,
filt_base=filt_base, ifmap_base=ifmap_base,
ofmap_base = ofmap_base,
sram_read_trace_file=sram_read_trace_file,
sram_write_trace_file=sram_write_trace_file
)
# If the lookup_flag is set make a new data base entry
#if lookup_flag:
# #this_dir = lut.to_string(subprocess.check_output(['pwd']))
# this_dir = os.getcwd()
# lut.create_sram_entry(sram_tag, this_dir, sram_cycles, util)
elif data_flow == 'ws':
tag_match = False
if lookup_flag:
sram_tag = tag + "_ws"
tag_match, dir_name = lut.lookup(sram_tag)
if tag_match:
#this_dir = lut.to_string(subprocess.check_output(['pwd']))
this_dir = os.getcwd()
sram_cycles, util = lut.get_sram_stats(dir_name, dest_dir=this_dir)
else:
sram_cycles, util = \
sram_ws.sram_traffic(
dimension_rows = array_h,
dimension_cols = array_w,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
strides = strides, num_filt = num_filt,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file,
sram_write_trace_file = sram_write_trace_file
)
# If the lookup_flag is set make a new data base entry
#if lookup_flag:
# #this_dir = lut.to_string(subprocess.check_output(['pwd']))
# this_dir = os.getcwd()
# lut.create_sram_entry(sram_tag, this_dir, sram_cycles, util)
elif data_flow == 'is':
tag_match = False
if lookup_flag:
sram_tag = tag + "_is"
tag_match, dir_name = lut.lookup(sram_tag)
if tag_match:
#this_dir = lut.to_string(subprocess.check_output(['pwd']))
this_dir = os.getcwd()
sram_cycles, util = lut.get_sram_stats(dir_name, dest_dir=this_dir)
else:
sram_cycles, util = \
sram_is.sram_traffic(
dimension_rows = array_h,
dimension_cols = array_w,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
strides = strides, num_filt = num_filt,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file,
sram_write_trace_file = sram_write_trace_file
)
# If the lookup_flag is set make a new data base entry
#if lookup_flag:
# #this_dir = lut.to_string(subprocess.check_output(['pwd']))
# this_dir = os.getcwd()
# lut.create_sram_entry(sram_tag, this_dir, sram_cycles, util)
tag_match = False
if lookup_flag:
dram_tag = tag + "_" + data_flow + "_" + str(
ifmap_sram_size) + "_ifmap"
tag_match, dir_name = lut.lookup(dram_tag)
if tag_match:
#this_dir = lut.to_string(subprocess.check_output(['pwd']))
this_dir = os.getcwd()
lut.get_dram_trace(dir_name, this_dir, cat='ifmap')
else:
#print("Generating DRAM traffic")
dram.dram_trace_read_v2(
sram_sz=ifmap_sram_size,
word_sz_bytes=word_size_bytes,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_ifmap_trace_file,
)
#if lookup_flag:
# #this_dir = lut.to_string(subprocess.check_output(['pwd']))
# this_dir = os.getcwd()
# lut.create_dram_entry(dram_tag, this_dir, cat='ifmap')
tag_match = False
if lookup_flag:
dram_tag = tag + "_" + data_flow + "_" + str(
filter_sram_size) + "_filter"
#dram_tag += "_" + str(filter_sram_size) + "_filter"
tag_match, dir_name = lut.lookup(dram_tag)
if tag_match:
#this_dir = lut.to_string(subprocess.check_output(['pwd']))
this_dir = os.getcwd()
lut.get_dram_trace(dir_name, this_dir, cat='filter')
else:
dram.dram_trace_read_v2(
sram_sz=filter_sram_size,
word_sz_bytes=word_size_bytes,
min_addr=filt_base,
max_addr=(filt_base * 10000),
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_filter_trace_file,
)
#if lookup_flag:
# #this_dir = lut.to_string(subprocess.check_output(['pwd']))
# this_dir = os.getcwd()
# lut.create_dram_entry(dram_tag, this_dir, cat='filter')
tag_match = False
if lookup_flag:
dram_tag = tag + "_" + data_flow + "_" + str(
ofmap_sram_size) + "_ofmap"
#dram_tag += "_" + str(filter_sram_size) + "_filter"
#dram_tag += "_" + str(ofmap_sram_size) + "_ofmap"
tag_match, dir_name = lut.lookup(dram_tag)
if tag_match:
#this_dir = lut.to_string(subprocess.check_output(['pwd']))
this_dir = os.getcwd()
lut.get_dram_trace(dir_name, this_dir, cat='ofmap')
else:
dram.dram_trace_write(ofmap_sram_size=ofmap_sram_size,
data_width_bytes=word_size_bytes,
sram_write_trace_file=sram_write_trace_file,
dram_write_trace_file=dram_ofmap_trace_file)
#if lookup_flag:
# #this_dir = lut.to_string(subprocess.check_output(['pwd']))
# this_dir = os.getcwd()
# lut.create_dram_entry(dram_tag, this_dir, cat='ofmap')
tag_match = False
if lookup_flag:
dram_tag = tag + "_" + data_flow + "_" + str(
ifmap_sram_size) + "_ifmap"
dram_tag += "_" + str(filter_sram_size) + "_filter"
dram_tag += "_" + str(ofmap_sram_size) + "_ofmap"
tag_match, dir_name = lut.lookup(dram_tag)
if tag_match:
#this_dir = lut.to_string(subprocess.check_output(['pwd']))
this_dir = os.getcwd()
bw_numbers, detailed_log =\
lut.get_log_entries(dir_name, this_dir)
else:
# Anand: Removing the following for SILENT mode
#print("Average utilization : \t" + str(util) + " %")
#print("Cycles for compute : \t" + str(sram_cycles) + " cycles")
bw_numbers, detailed_log = gen_bw_numbers(dram_ifmap_trace_file,
dram_filter_trace_file,
dram_ofmap_trace_file,
sram_write_trace_file,
sram_read_trace_file)
#array_h, array_w)
#if lookup_flag:
# #this_dir = lut.to_string(subprocess.check_output(['pwd']))
# this_dir = os.getcwd()
# lut.create_log_entry(dram_tag, bw_numbers, detailed_log)
return bw_numbers, detailed_log, util, sram_cycles
def gen_all_traces(array_h_first=4,
array_w_first=4,
array_h_second=4,
array_w_second=4,
single_array=1,
ifmap_h=7,
ifmap_w=7,
filt_h=3,
filt_w=3,
num_channels=3,
strides=1,
num_filt=8,
data_flow='os',
word_size_bytes=1,
filter_sram_size_first=64,
ifmap_sram_size_first=64,
ofmap_sram_size_first=64,
filter_sram_size_second=64,
ifmap_sram_size_second=64,
ofmap_sram_size_second=64,
filt_base=1000000,
ifmap_base=0,
ofmap_base=2000000,
sram_read_trace_file_first="sram0_read.csv",
sram_read_trace_file_second="sram1_read.csv",
sram_write_trace_file_first="sram0_write.csv",
sram_write_trace_file_second="sram1_write.csv",
dram_filter_trace_file="dram_filter_read.csv",
dram_ifmap_trace_file="dram_ifmap_read.csv",
dram_ofmap_trace_file="dram_ofmap_write.csv"):
sram_cycles = 0
sram_cycles_first = 0
sram_cycles_second = 0
array_one_used = 0
array_two_used = 0
util_first = 0
util_second = 0
dram_filter_trace_file_first = "dram_sram0_filter_read.csv"
dram_ifmap_trace_file_first = "dram_sram0_ifmap_read.csv"
dram_ofmap_trace_file_first = "dram_sram0_ofmap_write.csv"
dram_filter_trace_file_second = "dram_sram1_filter_read.csv"
dram_ifmap_trace_file_second = "dram_sram1_ifmap_read.csv"
dram_ofmap_trace_file_second = "dram_sram1_ofmap_write.csv"
print("Generating traces and bw numbers")
if data_flow == 'os':
num_filt_first = 0
num_filt_second = 0
i = 1
no_of_filt_px = filt_h * filt_w * num_channels
max_parallel_window_first = 1 #Since OS can't have more than one filter in a column given er vertical fold
max_parallel_window_second = 1
avail_filt_per_fold = (array_w_first * max_parallel_window_first) + (
array_w_second * max_parallel_window_second)
while True:
filt_processing = i * avail_filt_per_fold
if num_filt <= filt_processing:
filt_pend = num_filt - ((i - 1) * avail_filt_per_fold)
if filt_pend <= (array_w_first *
max_parallel_window_first) and filt_pend > (
array_w_second *
max_parallel_window_second
): ##Accomodating the last fold in systolic 1
num_filt_first = num_filt_first + filt_pend
elif filt_pend > (
array_w_first *
max_parallel_window_first) and filt_pend <= (
array_w_second * max_parallel_window_second
): ## Accomodating the last fold in systolic 2
num_filt_second = num_filt_second + filt_pend
elif filt_pend == avail_filt_per_fold:
num_filt_first = num_filt_first + (
array_w_first * max_parallel_window_first)
num_filt_second = num_filt_second + (
array_w_second * max_parallel_window_second)
elif filt_pend <= (
array_w_first *
max_parallel_window_first) and filt_pend <= (
array_w_second * max_parallel_window_second):
col_ratio_first = float(
filt_pend /
(array_w_first * max_parallel_window_first))
col_ratio_second = float(
filt_pend /
(array_w_second * max_parallel_window_second))
if (col_ratio_first >= col_ratio_second):
num_filt_first = num_filt_first + filt_pend
else:
num_filt_second = num_filt_second + filt_pend
else:
col_ratio_first = float(
(filt_pend -
(array_w_second * max_parallel_window_second)) /
(array_w_first * max_parallel_window_first))
col_ratio_second = float(
(filt_pend -
(array_w_first * max_parallel_window_first)) /
(array_w_second * max_parallel_window_second))
if (col_ratio_first > col_ratio_second):
num_filt_second = num_filt_second + (
array_w_second * max_parallel_window_second)
num_filt_first = num_filt_first + (
filt_pend -
(array_w_second * max_parallel_window_second))
else:
num_filt_first = num_filt_first + (
array_w_first * max_parallel_window_first)
num_filt_second = num_filt_second + (
filt_pend -
(array_w_first * max_parallel_window_first))
col_idx_base = num_filt_first ##Starting from systolic 1 and taking the systolic 1 filter count as the beginning for the next systolic as base addresss
break
else:
num_filt_first = num_filt_first + (array_w_first *
max_parallel_window_first)
num_filt_second = num_filt_second + (
array_w_second * max_parallel_window_second)
i = i + 1
if single_array == 1:
num_filt_first = num_filt
num_filt_second = 0
if num_filt_first > 0:
array_one_used = 1
sram_cycles_first, util_first = \
sram.sram_traffic(
dimension_rows= array_h_first,
dimension_cols= array_w_first,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt_first, total_num_filt = num_filt,
filt_base=filt_base, ifmap_base=ifmap_base, col_idx_base = 0,
ofmap_base = ofmap_base,
sram_read_trace_file=sram_read_trace_file_first,
sram_write_trace_file=sram_write_trace_file_first
)
else:
sram_cycles_first = 0
if num_filt_second > 0:
array_two_used = 1
sram_cycles_second, util_second = \
sram.sram_traffic(
dimension_rows= array_h_second,
dimension_cols= array_w_second,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt_second, total_num_filt = num_filt,
filt_base=filt_base, ifmap_base=ifmap_base, col_idx_base = col_idx_base,
ofmap_base = ofmap_base,
sram_read_trace_file=sram_read_trace_file_second,
sram_write_trace_file=sram_write_trace_file_second
)
else:
sram_cycles_second = 0
elif data_flow == 'ws':
num_filt_first = 0
num_filt_second = 0
i = 1
no_of_filt_px = filt_h * filt_w * num_channels
if array_h_first < no_of_filt_px:
max_parallel_window_first = 1
else:
max_parallel_window_first = math.floor(array_h_first /
no_of_filt_px)
if array_h_second < no_of_filt_px:
max_parallel_window_second = 1
else:
max_parallel_window_second = math.floor(array_h_second /
no_of_filt_px)
avail_filt_per_fold = (array_w_first * max_parallel_window_first) + (
array_w_second * max_parallel_window_second)
while True:
filt_processing = i * avail_filt_per_fold
if num_filt <= filt_processing:
filt_pend = num_filt - ((i - 1) * avail_filt_per_fold)
if filt_pend <= (array_w_first *
max_parallel_window_first) and filt_pend > (
array_w_second *
max_parallel_window_second
): ##Accomodating the last fold in systolic 1
num_filt_first = num_filt_first + filt_pend
elif filt_pend > (
array_w_first *
max_parallel_window_first) and filt_pend <= (
array_w_second * max_parallel_window_second
): ## Accomodating the last fold in systolic 2
num_filt_second = num_filt_second + filt_pend
elif filt_pend == avail_filt_per_fold:
num_filt_first = num_filt_first + (
array_w_first * max_parallel_window_first)
num_filt_second = num_filt_second + (
array_w_second * max_parallel_window_second)
elif filt_pend <= (
array_w_first *
max_parallel_window_first) and filt_pend <= (
array_w_second * max_parallel_window_second):
col_ratio_first = float(
filt_pend /
(array_w_first * max_parallel_window_first))
col_ratio_second = float(
filt_pend /
(array_w_second * max_parallel_window_second))
if (col_ratio_first >= col_ratio_second):
num_filt_first = num_filt_first + filt_pend
else:
num_filt_second = num_filt_second + filt_pend
else:
col_ratio_first = float(
(filt_pend -
(array_w_second * max_parallel_window_second)) /
(array_w_first * max_parallel_window_first))
col_ratio_second = float(
(filt_pend -
(array_w_first * max_parallel_window_first)) /
(array_w_second * max_parallel_window_second))
if (col_ratio_first > col_ratio_second):
num_filt_second = num_filt_second + (
array_w_second * max_parallel_window_second)
num_filt_first = num_filt_first + (
filt_pend -
(array_w_second * max_parallel_window_second))
else:
num_filt_first = num_filt_first + (
array_w_first * max_parallel_window_first)
num_filt_second = num_filt_second + (
filt_pend -
(array_w_first * max_parallel_window_first))
col_idx_base = num_filt_first ##Starting from systolic 1 and taking the systolic 1 filter count as the beginning for the next systolic as base addresss
break
else:
num_filt_first = num_filt_first + (array_w_first *
max_parallel_window_first)
num_filt_second = num_filt_second + (
array_w_second * max_parallel_window_second)
i = i + 1
if single_array == 1:
num_filt_first = num_filt
num_filt_second = 0
if num_filt_first > 0:
array_one_used = 1
sram_cycles_first, util_first = \
sram_ws.sram_traffic(
dimension_rows = array_h_first,
dimension_cols = array_w_first,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
col_idx_base = 0, total_num_filt = num_filt,
strides = strides, num_filt = num_filt_first,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file_first,
sram_write_trace_file = sram_write_trace_file_first
)
else:
sram_cycles_first = 0
if num_filt_second > 0:
array_two_used = 1
sram_cycles_second, util_second = \
sram_ws.sram_traffic(
dimension_rows = array_h_second,
dimension_cols = array_w_second,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
col_idx_base = col_idx_base, total_num_filt = num_filt,
strides = strides, num_filt = num_filt_second,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file_second,
sram_write_trace_file = sram_write_trace_file_second
)
else:
sram_cycles_second = 0
elif data_flow == 'is':
ofmap_h = (ifmap_h - filt_h) / strides + 1
ofmap_w = (ifmap_w - filt_w) / strides + 1
num_ofmap = ofmap_h * ofmap_w
num_ofmap_first = 0
num_ofmap_second = 0
i = 1
no_of_filt_px = filt_h * filt_w * num_channels
if array_h_first < no_of_filt_px:
max_parallel_window_first = 1
else:
max_parallel_window_first = math.floor(array_h_first /
no_of_filt_px)
if array_h_second < no_of_filt_px:
max_parallel_window_second = 1
else:
max_parallel_window_second = math.floor(array_h_second /
no_of_filt_px)
avail_ofmap_per_fold = (array_w_first * max_parallel_window_first) + (
array_w_second * max_parallel_window_second)
while True:
ofmap_processing = i * avail_ofmap_per_fold
if num_ofmap <= ofmap_processing:
ofmap_pend = num_ofmap - ((i - 1) * avail_ofmap_per_fold)
if ofmap_pend <= (
array_w_first *
max_parallel_window_first) and ofmap_pend > (
array_w_second * max_parallel_window_second
): ##Accomodating the last fold in systolic 1
num_ofmap_first = num_ofmap_first + ofmap_pend
elif ofmap_pend > (
array_w_first *
max_parallel_window_first) and ofmap_pend <= (
array_w_second * max_parallel_window_second
): ## Accomodating the last fold in systolic 2
num_ofmap_second = num_ofmap_second + ofmap_pend
elif ofmap_pend == avail_ofmap_per_fold:
num_ofmap_first = num_ofmap_first + (
array_w_first * max_parallel_window_first)
num_ofmap_second = num_ofmap_second + (
array_w_second * max_parallel_window_second)
elif ofmap_pend <= (
array_w_first *
max_parallel_window_first) and ofmap_pend <= (
array_w_second * max_parallel_window_second):
col_ratio_first = float(
ofmap_pend /
(array_w_first * max_parallel_window_first))
col_ratio_second = float(
ofmap_pend /
(array_w_second * max_parallel_window_second))
if (col_ratio_first >= col_ratio_second):
num_ofmap_first = num_ofmap_first + ofmap_pend
else:
num_ofmap_second = num_ofmap_second + ofmap_pend
else:
col_ratio_first = float(
(ofmap_pend -
(array_w_second * max_parallel_window_second)) /
(array_w_first * max_parallel_window_first))
col_ratio_second = float(
(ofmap_pend -
(array_w_first * max_parallel_window_first)) /
(array_w_second * max_parallel_window_second))
if (col_ratio_first > col_ratio_second):
num_ofmap_second = num_ofmap_second + (
array_w_second * max_parallel_window_second)
num_ofmap_first = num_ofmap_first + (
ofmap_pend -
(array_w_second * max_parallel_window_second))
else:
num_ofmap_first = num_ofmap_first + (
array_w_first * max_parallel_window_first)
num_ofmap_second = num_ofmap_second + (
ofmap_pend -
(array_w_first * max_parallel_window_first))
col_idx_base = num_ofmap_first ##Starting from systolic 1 and taking the systolic 1 filter count as the beginning for the next systolic as base addresss
break
else:
num_ofmap_first = num_ofmap_first + (array_w_first *
max_parallel_window_first)
num_ofmap_second = num_ofmap_second + (
array_w_second * max_parallel_window_second)
i = i + 1
if single_array == 1:
num_ofmap_first = num_ofmap
num_ofmap_second = 0
if num_ofmap_first > 0:
array_one_used = 1
sram_cycles_first, util_first = \
sram_is.sram_traffic(
dimension_rows = array_h_first,
dimension_cols = array_w_first,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels, num_ofmap = num_ofmap_first,
strides = strides, num_filt = num_filt,
col_idx_base = 0,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file_first,
sram_write_trace_file = sram_write_trace_file_first
)
else:
sram_cycles_first = 0
if num_ofmap_second > 0:
array_two_used = 1
sram_cycles_second, util_second = \
sram_is.sram_traffic(
dimension_rows = array_h_second,
dimension_cols = array_w_second,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels, num_ofmap = num_ofmap_second,
strides = strides, num_filt = num_filt,
col_idx_base = col_idx_base,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file_second,
sram_write_trace_file = sram_write_trace_file_second
)
else:
sram_cycles_second = 0
#print("Generating DRAM traffic")
if array_one_used == 1:
if single_array == 1 or array_two_used == 0:
dram.dram_trace_read_v2(sram_sz=ifmap_sram_size_first,
word_sz_bytes=word_size_bytes,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file_first,
dram_trace_file=dram_ifmap_trace_file)
dram.dram_trace_read_v2(sram_sz=filter_sram_size_first,
word_sz_bytes=word_size_bytes,
min_addr=filt_base,
max_addr=ofmap_base,
sram_trace_file=sram_read_trace_file_first,
dram_trace_file=dram_filter_trace_file)
dram.dram_trace_write(
ofmap_sram_size=ofmap_sram_size_first,
data_width_bytes=word_size_bytes,
sram_write_trace_file=sram_write_trace_file_first,
dram_write_trace_file=dram_ofmap_trace_file)
else:
dram.dram_trace_read_v2(
sram_sz=ifmap_sram_size_first,
word_sz_bytes=word_size_bytes,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file_first,
dram_trace_file=dram_ifmap_trace_file_first)
dram.dram_trace_read_v2(
sram_sz=filter_sram_size_first,
word_sz_bytes=word_size_bytes,
min_addr=filt_base,
max_addr=ofmap_base,
sram_trace_file=sram_read_trace_file_first,
dram_trace_file=dram_filter_trace_file_first)
dram.dram_trace_write(
ofmap_sram_size=ofmap_sram_size_first,
data_width_bytes=word_size_bytes,
sram_write_trace_file=sram_write_trace_file_first,
dram_write_trace_file=dram_ofmap_trace_file_first)
if array_two_used == 1:
if array_one_used == 0:
dram.dram_trace_read_v2(
sram_sz=ifmap_sram_size_second,
word_sz_bytes=word_size_bytes,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file_second,
dram_trace_file=dram_ifmap_trace_file)
dram.dram_trace_read_v2(
sram_sz=filter_sram_size_second,
word_sz_bytes=word_size_bytes,
min_addr=filt_base,
max_addr=ofmap_base,
sram_trace_file=sram_read_trace_file_second,
dram_trace_file=dram_filter_trace_file)
dram.dram_trace_write(
ofmap_sram_size=ofmap_sram_size_second,
data_width_bytes=word_size_bytes,
sram_write_trace_file=sram_write_trace_file_second,
dram_write_trace_file=dram_ofmap_trace_file)
else:
dram.dram_trace_read_v2(
sram_sz=ifmap_sram_size_second,
word_sz_bytes=word_size_bytes,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file_second,
dram_trace_file=dram_ifmap_trace_file_second)
dram.dram_trace_read_v2(
sram_sz=filter_sram_size_second,
word_sz_bytes=word_size_bytes,
min_addr=filt_base,
max_addr=ofmap_base,
sram_trace_file=sram_read_trace_file_second,
dram_trace_file=dram_filter_trace_file_second)
dram.dram_trace_write(
ofmap_sram_size=ofmap_sram_size_second,
data_width_bytes=word_size_bytes,
sram_write_trace_file=sram_write_trace_file_second,
dram_write_trace_file=dram_ofmap_trace_file_second)
# Selvaraj: Merge both DRAM traffic CSV's for BW calculations
if array_one_used == 1 and array_two_used == 1:
sram_controller(dram_ifmap_trace_file_first,
dram_ifmap_trace_file_second, dram_ifmap_trace_file)
sram_controller(dram_filter_trace_file_first,
dram_filter_trace_file_second, dram_filter_trace_file)
sram_controller(dram_ofmap_trace_file_first,
dram_ofmap_trace_file_second, dram_ofmap_trace_file)
if array_one_used == 1 and array_two_used == 0:
util = util_first
power_metric = (int(sram_cycles_first) * int(array_h_first) *
int(array_w_first)) / 1000000
elif array_one_used == 0 and array_two_used == 1:
util = util_second
power_metric = (int(sram_cycles_second) * int(array_h_second) *
int(array_w_second)) / 1000000
else:
util = (util_first +
util_second) / 2 #Equally weighted from two systolic arrays
power_metric = ((int(sram_cycles_first) * int(array_h_first) *
int(array_w_first)) +
(int(sram_cycles_second) * int(array_h_second) *
int(array_w_second))) / 1000000
sram_cycles = max(int(sram_cycles_first), int(sram_cycles_second))
print("Average utilization : \t" + str(util) + " %")
print("Cycles for compute : \t" + str(sram_cycles) + " cycles")
print("Power consumed : \t" + str(power_metric) + " Mega-units")
if single_array == 1: # SCALE-Sim used as a single compute array simulator
bw_numbers, detailed_log = gen_bw_numbers(
both_array_used=0,
array_one_idle=0,
array_two_idle=0,
dram_ifmap_trace_file=dram_ifmap_trace_file,
dram_filter_trace_file=
dram_filter_trace_file, #Selvaraj: Add support for two SRAM based BW generation after DRAM merge
dram_ofmap_trace_file=dram_ofmap_trace_file,
sram_write_trace_file_first=sram_write_trace_file_first,
sram_read_trace_file_first=sram_read_trace_file_first)
#array_h, array_w)
elif (array_one_used == 1
and array_two_used == 0): ## Second array not powered on at all
bw_numbers, detailed_log = gen_bw_numbers(
both_array_used=1,
array_one_idle=0,
array_two_idle=1,
dram_ifmap_trace_file=dram_ifmap_trace_file,
dram_filter_trace_file=dram_filter_trace_file,
dram_ofmap_trace_file=dram_ofmap_trace_file,
sram_write_trace_file_first=sram_write_trace_file_first,
sram_read_trace_file_first=sram_read_trace_file_first)
elif (array_one_used == 0
and array_two_used == 1): ## First array not powered on at all
bw_numbers, detailed_log = gen_bw_numbers(
both_array_used=1,
array_one_idle=1,
array_two_idle=0,
dram_ifmap_trace_file=dram_ifmap_trace_file,
dram_filter_trace_file=dram_filter_trace_file,
dram_ofmap_trace_file=dram_ofmap_trace_file,
sram_write_trace_file_second=sram_write_trace_file_second,
sram_read_trace_file_second=sram_read_trace_file_second)
elif array_one_used == 1 and array_two_used == 1:
bw_numbers, detailed_log = gen_bw_numbers(
both_array_used=1,
array_one_idle=0,
array_two_idle=0,
dram_ifmap_trace_file=dram_ifmap_trace_file,
dram_filter_trace_file=dram_filter_trace_file,
dram_ofmap_trace_file=dram_ofmap_trace_file,
sram_write_trace_file_first=sram_write_trace_file_first,
sram_read_trace_file_first=sram_read_trace_file_first,
sram_write_trace_file_second=sram_write_trace_file_second,
sram_read_trace_file_second=sram_read_trace_file_second)
#array_h, array_w)
return bw_numbers, detailed_log, util, str(
sram_cycles), array_one_used, array_two_used, power_metric
def gen_all_traces(array_h=4,
array_w=4,
ifmap_h=7,
ifmap_w=7,
filt_h=3,
filt_w=3,
num_channels=3,
strides=1,
num_filt=8,
data_flow='os',
word_size_bytes=1,
filter_sram_size=64,
ifmap_sram_size=64,
ofmap_sram_size=64,
filt_base=1000000,
ifmap_base=0,
ofmap_base=2000000,
sram_read_trace_file="sram_read.csv",
sram_write_trace_file="sram_write.csv",
dram_filter_trace_file="dram_filter_read.csv",
dram_ifmap_trace_file="dram_ifmap_read.csv",
dram_ofmap_trace_file="dram_ofmap_write.csv"):
sram_cycles = 0
util = 0
ifmap_dram_acc_times = 0
ofmap_dram_acc_times = 0
filt_dram_acc_times = 0
print("Generating traces and bw numbers")
if data_flow == 'os':
sram_cycles, util = \
sram.sram_traffic(
dimension_rows= array_h,
dimension_cols= array_w,
ifmap_h=ifmap_h, ifmap_w=ifmap_w,
filt_h=filt_h, filt_w=filt_w,
num_channels=num_channels,
strides=strides, num_filt=num_filt,
filt_base=filt_base, ifmap_base=ifmap_base,
ofmap_base = ofmap_base,
sram_read_trace_file=sram_read_trace_file,
sram_write_trace_file=sram_write_trace_file
)
elif data_flow == 'ws':
sram_cycles, util = \
sram_ws.sram_traffic(
dimension_rows = array_h,
dimension_cols = array_w,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
strides = strides, num_filt = num_filt,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file,
sram_write_trace_file = sram_write_trace_file
)
elif data_flow == 'is':
sram_cycles, util = \
sram_is.sram_traffic(
dimension_rows = array_h,
dimension_cols = array_w,
ifmap_h = ifmap_h, ifmap_w = ifmap_w,
filt_h = filt_h, filt_w = filt_w,
num_channels = num_channels,
strides = strides, num_filt = num_filt,
ofmap_base = ofmap_base, filt_base = filt_base, ifmap_base = ifmap_base,
sram_read_trace_file = sram_read_trace_file,
sram_write_trace_file = sram_write_trace_file
)
#print("Generating DRAM traffic")
# 利用Sram的时间反推DRAMD执行时间
# 这里还是一个一个调用的，也就是说这三者可以同时发生，但是实际上我们的带宽并没有这么多，
# 而且应该如何cover计算和访存之间的延时
ifmap_dram_acc_times = dram.dram_trace_read_v2(
sram_sz=ifmap_sram_size,
word_sz_bytes=word_size_bytes,
min_addr=ifmap_base,
max_addr=filt_base,
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_ifmap_trace_file,
)
filt_dram_acc_times = dram.dram_trace_read_v2(
sram_sz=filter_sram_size,
word_sz_bytes=word_size_bytes,
min_addr=filt_base,
max_addr=ofmap_base,
sram_trace_file=sram_read_trace_file,
dram_trace_file=dram_filter_trace_file,
)
ofmap_dram_acc_times = dram.dram_trace_write(
ofmap_sram_size=ofmap_sram_size,
data_width_bytes=word_size_bytes,
sram_write_trace_file=sram_write_trace_file,
dram_write_trace_file=dram_ofmap_trace_file)
print("Average utilization : \t" + str(util) + " %")
print("Cycles for compute : \t" + str(sram_cycles) + " cycles")
bw_numbers, detailed_log = gen_bw_numbers(dram_ifmap_trace_file,
dram_filter_trace_file,
dram_ofmap_trace_file,
sram_write_trace_file,
sram_read_trace_file)
#array_h, array_w)
detailed_log += str(ifmap_dram_acc_times) + ",\t" + str(
filt_dram_acc_times) + ",\t" + str(ofmap_dram_acc_times) + ","
print((ifmap_dram_acc_times, filt_dram_acc_times, ofmap_dram_acc_times))
return bw_numbers, detailed_log, util, sram_cycles