def gen_inout_bin_res(bindir, bmpdt, net, pool_num, downsample, name_list, data_list):
file_list = []
alphalist = []
intxt_list = []
outtxt_list = []
from script.run_steps import get_model_name
name = get_model_name(net)
layer_max, opear_max, block_max = getresnet_max(net, name)
def file_exist(name):
if os.path.exists(f'{bmpdt}/{name}'):
file_list.append(name)
for i in range(opear_max):
inname = f'conv{i}.input.6.txt'
outname = f'quant{i}.output.6.txt'
file_exist(inname)
file_exist(outname)
for layerindex in range(1, layer_max+1):
for blockindex in range(int(block_max[layerindex-1].strip())):
for opearindex in range(1, opear_max+1):
inname = f'layer{layerindex}.{blockindex}.conv{opearindex}.input.6.txt'
outname = f'layer{layerindex}.{blockindex}.quant{opearindex}.output.6.txt'
file_exist(inname)
file_exist(outname)
if layerindex and blockindex == 0:
inname = f'layer{layerindex}.0.downsample.0.input.6.txt'
outname = f'layer{layerindex}.0.quant_shortcut.output.6.txt'
file_exist(inname)
file_exist(outname)
avgout = "quant_avg.output.6.txt"
fcin = "fc.input.6.txt"
fcout = "quant_fc.output.6.txt"
if os.path.exists(f'{bmpdt}/{avgout}'):
file_list.append(f'{avgout}')
if os.path.exists(f'{bmpdt}/{fcin}'):
file_list.append(f'{fcin}')
if os.path.exists(f'{bmpdt}/{fcout}'):
file_list.append(f'{fcout}')
if not os.path.exists(bindir):
print(f'run os.mkdir({bindir} error)')
return -1
for i in range(len(name_list)):
name = name_list[i]
if ".alpha" in name:
alphalist.append(data_list[i].tolist())
for i in range(len(file_list)):
if avgout == file_list[i]:
avgin = file_list[i-1]
if "quant_avg" in file_list[i]:
break
if "input" in file_list[i]:
intxt_list.append(file_list[i])
if "output" in file_list[i+1]:
if "output" in file_list[i+2]:
outname = file_list[i+2]
else:
outname = file_list[i+1]
outtxt_list.append(outname)
intxt_list.append(fcin)
outtxt_list.append(fcout)
j = 0
bit = 7
in_q = 5
for i in range(1, layer_cnts+1):
if i == 1:
C_in, H_in, W_in = get_fstin_CHW(net)
elif i in downsample:
C_in, H_in, W_in = get_out_CHW(i-2)
else:
C_in, H_in, W_in = get_out_CHW(i-1)
C_out, H_out, W_out = get_out_CHW(i)
if i in pool_num:
out_q = in_q
continue
out_q = bit - math.ceil(math.log2(0.5 * alphalist[j]))
input_txt = f'{bmpdt}/{intxt_list[j]}'
output_txt = f'{bmpdt}/{outtxt_list[j]}'
input_bin = f'{bindir}/layer{i}_input_act0_file.bin'
output_bin = f'{bindir}/layer{i}_output_file.bin'
write_in = threading.Thread(target=write_inout_bin,
args=(input_txt, input_bin, H_in, W_in, C_in, in_q))
write_in.start()
write_in.join()
write_out = threading.Thread(target=write_inout_bin,
args=(output_txt, output_bin, H_out, W_out, C_out, out_q))
write_out.start()
write_out.join()
j += 1
in_q = out_q
def gen_inout_bin_vgg(bindir, bmpdt, net, pool_num, downsample, name_list, data_list):
fc_output = []
fc_inputs = []
alphalist = []
conv_inputs = []
conv_output = []
list = os.listdir(bmpdt)
for i in range(len(list)):
if "input" in list[i]:
if "conv" in list[i]:
conv_inputs.append(list[i])
elif "classifier" in list[i]:
fc_inputs.append(list[i])
elif "output" in list[i]:
if "quant." in list[i]:
conv_output.append(list[i])
elif "quant_fc" in list[i]:
fc_output.append(list[i])
fc_inputs = sorted(fc_inputs)
fc_output = sorted(fc_output)
conv_inputs.sort(key=lambda x: int(re.match('\D+(\d+)\.conv.input.6.txt', x).group(1)))
conv_output.sort(key=lambda x: int(re.match('\D+(\d+)\.quant.output.6.txt', x).group(1)))
for i in range(len(name_list)):
name = name_list[i]
if ".alpha" in name:
alphalist.append(data_list[i].tolist())
j = 0
bit = 7
in_q = 5
for i in range(1, layer_cnts - len(fc_inputs) + 1):
if i == 1:
C_in, H_in, W_in = get_fstin_CHW(net)
elif i in downsample:
C_in, H_in, W_in = get_out_CHW(i-2)
else:
C_in, H_in, W_in = get_out_CHW(i-1)
C_out, H_out, W_out = get_out_CHW(i)
if i in pool_num:
out_q = in_q
continue
out_q = bit - math.ceil(math.log2(0.5 * alphalist[j]))
input_txt = f'{bmpdt}/{conv_inputs[j]}'
output_txt = f'{bmpdt}/{conv_output[j]}'
input_bin = f'{bindir}/layer{i}_input_act0_file.bin'
output_bin = f'{bindir}/layer{i}_output_file.bin'
write_in = threading.Thread(target=write_inout_bin,
args=(input_txt, input_bin, H_in, W_in, C_in, in_q))
write_in.start()
write_in.join()
write_out = threading.Thread(target=write_inout_bin,
args=(output_txt, output_bin, H_out, W_out, C_out, out_q))
write_out.start()
write_out.join()
j += 1
in_q = out_q
if layer_cnts <= 1:
# Adapt one layer toolchains
return
for i in range(len(fc_inputs)):
layernum = i + len(conv_inputs) + len(pool_num) + 1
C_in, H_in, W_in = get_out_CHW(layernum-1)
C_out, H_out, W_out = get_out_CHW(layernum)
out_q = bit - math.ceil(math.log2(0.5 * alphalist[j + len(fc_inputs)]))
input_txt = f'{bmpdt}/{fc_inputs[i]}'
output_txt = f'{bmpdt}/{fc_output[i]}'
input_bin = f'{bindir}/layer{len(conv_inputs) + i + len(pool_num)}_input_act0_file.bin'
output_bin = f'{bindir}/layer{len(conv_inputs) + i + len(pool_num)}_output_file.bin'
write_in = threading.Thread(target=write_inout_bin,
args=(input_txt, input_bin, H_in, W_in, C_in, in_q))
write_in.start()
write_in.join()
write_out = threading.Thread(target=write_inout_bin,
args=(output_txt, output_bin, H_out, W_out, C_out, in_q))
write_out.start()
write_out.join()
in_q = out_q
def get_nnbaton(num_type, netpth, logpth, actbit, wetbit):
flag = 1
nnbaton = []
layermax = 0
dec_index = 0 # res_branch index
sizes = get_fstin_CHW(netpth)
net = logpth.rsplit('/', 1)[1].split('.')[0]
if 'vggnet' not in net and 'resnet' not in net \
and "simulator" not in net:
print(f'Muse-v3 toolchains not recognized {net}')
exit(0)
if 'vggnet' in net or net == 'simulator':
flag = 0
elif 'resnet' in net:
resnet = get_single_conv(netpth)
layermax = find_layer_quant("layer", netpth, net)[0]
if net == 'resnet18':
flag = 1
resdec = ['a', 'b'] # csv:res2 & csv:res5 = block cnt
bradec = ['2a', '2b'] # csv:branch description == layer cnt
for rescnt in range(1, int(layermax) + 1): # csv:res2~res5
for j in range(len(resdec)):
if resdec[j] == 'a' and rescnt > 1:
bradec = ['2a', '2b', '1']
else:
bradec = ['2a', '2b']
for k in range(len(bradec)):
layer_type = f'res{rescnt + 1}{resdec[j]}_branch{bradec[k]}'
resnet.append(layer_type)
elif net == 'resnet34':
flag = 2
resdec = ['a', 'b', 'c'] # csv:res2 & csv:res5 = block cnt
bradec = ['2a', '2b'] # csv:branch description == layer cnt
for rescnt in range(1, int(layermax) + 1): # csv:res2~res5
if rescnt == 2: # csv:res3
resdec = ['a', 'b', 'c', 'd']
elif rescnt == 3: # csv:res4
resdec = ['a', 'b', 'c', 'd', 'e', 'f']
elif rescnt == 4: # csv:res5
resdec = ['a', 'b', 'c']
for j in range(len(resdec)):
if resdec[j] == 'a' and rescnt > 1:
bradec = ['2a', '2b', '1']
else:
bradec = ['2a', '2b']
for k in range(len(bradec)):
layer_type = f'res{rescnt + 1}{resdec[j]}_branch{bradec[k]}'
resnet.append(layer_type)
elif net == 'resnet50':
flag = 3
resdec = ['a', 'b', 'c'] # csv:res2 & csv:res5 = block cnt
bradec = ['1', '2a', '2b',
'2c'] # csv:branch description == layer cnt
for rescnt in range(1, int(layermax) + 1): # csv:res2~res5
if rescnt == 2: # csv:res3
resdec = ['a', 'b', 'c', 'd']
elif rescnt == 3: # csv:res4
resdec = ['a', 'b', 'c', 'd', 'e', 'f']
elif rescnt == 4: # csv:res5
resdec = ['a', 'b', 'c']
for j in range(len(resdec)):
if resdec[j] == 'a' and rescnt == 1:
bradec = ['2a', '2b', '2c', '1']
elif resdec[j] == 'a' and rescnt > 1:
bradec = ['1', '2a', '2b', '2c']
else:
bradec = ['2a', '2b', '2c']
for k in range(len(bradec)):
layer_type = f'res{rescnt + 1}{resdec[j]}_branch{bradec[k]}'
resnet.append(layer_type)
else:
print(f'Muse-v3 toolchains unknown model {net}')
exit(0)
worklad = f'nnbaton/workload/{net}-{sizes[1]}.csv'
batonlg = f'../.debug/nnbaton_{net}-{sizes[1]}.log'
csvname = f'nnbaton/csv/sort/reuslt_{net}-{sizes[1]}.csv'
csvpath = ['sort', 'raw4/output', 'cooked4', 'workload']
data = [
'cin', 'in_size', 'cout', 'out_size', 'kernel_size', 'stride', 'note'
]
run_list = [
f'python3 main.py {net} {sizes[1]} {actbit} {wetbit} >{batonlg} 2>&1',
f'python3 view.py >> {batonlg} 2>&1'
]
for i in range(len(csvpath)):
csvdir = f'nnbaton/csv/{csvpath[i]}'
if 'workload' == csvpath[i]:
csvdir = f'nnbaton/workload'
if os.path.exists(csvdir):
shutil.rmtree(csvdir)
os.makedirs(csvdir)
with open(worklad, 'w', newline='') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=data)
writer.writeheader()
for i in range(1, len(num_type)):
layercnt = int(num_type[i][0])
layerdec = num_type[i][1].split(':')[1]
if flag == 0:
layerdec = layerdec.split(' ')[0]
if "conv" in layerdec:
if layercnt == 1:
CHW_in = sizes
elif "downsample" in layerdec:
CHW_in = get_out_CHW(i - 2)
else:
CHW_in = get_out_CHW(i - 1)
CHW_out = get_out_CHW(i)
cout = CHW_out[0]
stride = num_type[i][4]
out_size = num_type[i][3]
ker_size = num_type[i][2]
cin = align(CHW_in[0], 8)
in_size = out_size * stride + ker_size - stride
if flag and 'conv' in layerdec:
layerdec = resnet[dec_index]
dec_index += 1
data = [
cin, in_size, cout, out_size, ker_size, stride, layerdec
]
writer = csv.DictWriter(csvfile, fieldnames=data)
writer.writeheader()
os.chdir("nnbaton")
for i in range(len(run_list)):
os.system(run_list[i])
os.chdir("..")
X1_index = Y1_index = K1_index = 0
X2_index = Y2_index = K2_index = 0
Kc_index = Yc_index = Xc_index = 0
Kp_index = Yp_index = 0
C1_index = C0_index = 0
X0_index = Y0_index = 0
op_index = divindex = 0
with open(csvname, 'r') as my_csv:
reader = csv.reader(my_csv)
for row in reader:
for i in range(len(row)):
if row[i] == "note":
op_index = i
elif row[i] == "X1":
X1_index = i
elif row[i] == "Y1":
Y1_index = i
elif row[i] == "K1":
K1_index = i
elif row[i] == "X2":
X2_index = i
elif row[i] == "Y2":
Y2_index = i
elif row[i] == "K2":
K2_index = i
elif row[i] == "Kp":
Kp_index = i
elif row[i] == "Yp":
Yp_index = i
elif row[i] == "Xc":
Xc_index = i
elif row[i] == "Yc":
Yc_index = i
elif row[i] == "Kc":
Kc_index = i
elif row[i] == "C1":
C1_index = i
elif row[i] == "C0":
C0_index = i
elif row[i] == "X0":
X0_index = i
elif row[i] == "Y0":
Y0_index = i
elif row[i] == "reorder_case":
divindex = i
break
for i in range(1, len(num_type)):
layer_type = num_type[i][1].split(':')[1].split(' ', 1)[0]
with open(csvname, 'r') as my_csv:
reader = csv.reader(my_csv)
for row in reader:
if row[op_index] == layer_type:
nnbaton.append([
int(row[X1_index]),
int(row[Y1_index]),
int(row[K1_index]),
int(row[X2_index]),
int(row[Y2_index]),
int(row[K2_index]),
int(row[Kp_index]),
int(row[Yp_index]),
int(row[Kc_index]),
int(row[Yc_index]),
int(row[Xc_index]),
int(row[C1_index]),
int(row[C0_index]),
int(row[X0_index]),
int(row[Y0_index]),
int(row[divindex])
])
if flag == 1:
for rescnt in range(1, int(layermax) + 1): # csv:res2~res5
for j in range(len(resdec)):
if resdec[j] == 'a' and rescnt > 1:
bradec = ['2a', '2b', '1']
else:
bradec = ['2a', '2b']
for k in range(len(bradec)):
layer_type = f'res{rescnt + 1}{resdec[j]}_branch{bradec[k]}'
with open(csvname, 'r') as my_csv:
reader = csv.reader(my_csv)
for row in reader:
if row[op_index] == layer_type:
nnbaton.append([
int(row[X1_index]),
int(row[Y1_index]),
int(row[K1_index]),
int(row[X2_index]),
int(row[Y2_index]),
int(row[K2_index]),
int(row[Kp_index]),
int(row[Yp_index]),
int(row[Kc_index]),
int(row[Yc_index]),
int(row[Xc_index]),
int(row[C1_index]),
int(row[C0_index]),
int(row[X0_index]),
int(row[Y0_index]),
int(row[divindex])
])
elif flag == 2:
for rescnt in range(1, int(layermax) + 1): # csv:res2~res5
if rescnt == 2: # csv:res3
resdec = ['a', 'b', 'c', 'd']
elif rescnt == 3: # csv:res4
resdec = ['a', 'b', 'c', 'd', 'e', 'f']
elif rescnt == 4: # csv:res5
resdec = ['a', 'b', 'c']
for j in range(len(resdec)):
if resdec[j] == 'a' and rescnt > 1:
bradec = ['2a', '2b', '1']
else:
bradec = ['2a', '2b']
for k in range(len(bradec)):
layer_type = f'res{rescnt + 1}{resdec[j]}_branch{bradec[k]}'
with open(csvname, 'r') as my_csv:
reader = csv.reader(my_csv)
for row in reader:
if row[op_index] == layer_type:
nnbaton.append([
int(row[X1_index]),
int(row[Y1_index]),
int(row[K1_index]),
int(row[X2_index]),
int(row[Y2_index]),
int(row[K2_index]),
int(row[Kp_index]),
int(row[Yp_index]),
int(row[Kc_index]),
int(row[Yc_index]),
int(row[Xc_index]),
int(row[C1_index]),
int(row[C0_index]),
int(row[X0_index]),
int(row[Y0_index]),
int(row[divindex])
])
elif flag == 3:
for rescnt in range(1, int(layermax) + 1): # csv:res2~res5
if rescnt == 2: # csv:res3
resdec = ['a', 'b', 'c', 'd']
elif rescnt == 3: # csv:res4
resdec = ['a', 'b', 'c', 'd', 'e', 'f']
elif rescnt == 4: # csv:res5
resdec = ['a', 'b', 'c']
for j in range(len(resdec)):
if resdec[j] == 'a' and rescnt == 1:
bradec = ["2a", "2b", "2c", '1']
elif resdec[j] == 'a' and rescnt > 1:
bradec = ['1', "2a", "2b", "2c"]
else:
bradec = ["2a", "2b", "2c"]
for k in range(len(bradec)):
layer_type = f'res{rescnt + 1}{resdec[j]}_branch{bradec[k]}'
with open(csvname, 'r') as my_csv:
reader = csv.reader(my_csv)
for row in reader:
if row[op_index] == layer_type:
nnbaton.append([
int(row[X1_index]),
int(row[Y1_index]),
int(row[K1_index]),
int(row[X2_index]),
int(row[Y2_index]),
int(row[K2_index]),
int(row[Kp_index]),
int(row[Yp_index]),
int(row[Kc_index]),
int(row[Yc_index]),
int(row[Xc_index]),
int(row[C1_index]),
int(row[C0_index]),
int(row[X0_index]),
int(row[Y0_index]),
int(row[divindex])
])
for i in range(1, len(num_type)):
layer_type = num_type[i][1].split(':')[1].split(' ', 1)[0]
if "conv" not in layer_type:
nnbaton.insert(i - 1,
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0])
return nnbaton