def gen_bmp(layercnt, vgglog):
"""
description:
Generate model files
parameters:
vgglog: Path to vgglog file
return code:
None
"""
net = ""
if "vggnet" in ptpath:
net = "vggnet"
elif "resnet18" in ptpath:
net = "resnet18"
elif "resnet34" in ptpath:
net = "resnet34"
elif "resnet50" in ptpath:
net = "resnet50"
else:
print(f'Unknown {ptpath}')
return 1
cmd_list = [
f'python3 input/inout_print.py {net} imagenet/im6.bmp {ptpath} {bmpdtpath} > .debug/genbmp.log'
]
for i in range(len(cmd_list)):
os.system(cmd_list[i])
if layercnt > 1 and "resnet" not in net:
# Muse-v2 interface, v3 is not used.
deal_fc_k(vgglog)
prints("run gen_bmp successfully")
def gen_fpga(filepath):
"""
description:
Generate bin files for fpga
parameters:
filepath: Run config_gen_file.py directory
return code:
None
"""
os.chdir(filepath)
cmd_list = [
"rm -rf con*txt cfg*txt *bn* *bias* *alpha* *weight* *input* *output* \
*k* data_for_fpga",
f'cp -af {confpath}/* {ptdtpath}/* {bmpdtpath}/* imagenet/',
"python3 ../input/config_gen_file.py -d ../imagenet/ -n imagenet_img6 -f > ../.debug/genfile.log",
f'mv {filepath}/data_for_fpga {outputpath}'
]
for i in range(len(cmd_list)):
if "cp -af " in cmd_list[i] or "mv " in cmd_list[i]:
os.chdir("..")
elif "python" in cmd_list[i]:
os.chdir(filepath)
os.system(cmd_list[i])
prints("generate fpga files successfully")
def binary_addr(instdir, datadir, output):
binsize = 0
binsize = get_binsize(os.listdir(instdir), binsize, output)
binsize = get_binsize(os.listdir(datadir), binsize, output)
fmain.write(f'all binary size:{binsize} bytes\n')
fmain.write(f'DDR ADDRBLOCK size :{binsize} bytes\n')
prints("calculate address of binary successfully")
def gen_ddraddr(name_list, data_list, act_bit, wet_bit):
fmain.write(f'dram_capacity:{dram_capacity}\n')
netinout_addr(netpath)
poollist, downsample = weight_addr(name_list, data_list, act_bit, wet_bit)
if layer_cnts > 1:
global CALCULATE
CALCULATE = ADDRBLOCK
otherinout_addr()
prints("run gen_ddraddr successfully")
return layers_act_addr,layers_wet_addr,datas_locate,poollist,downsample
def binary_addr(instdir, datadir, output):
binsize = 0
binsize = get_binsize(os.listdir(instdir), binsize, output)
binsize = get_binsize(os.listdir(datadir), binsize, output)
fmain.write(f'all binary size:{binsize} bytes\n')
global ADDRBLOCK
ADDRBLOCK += binsize
fmain.write(f'DDR ADDRBLOCK size :{ADDRBLOCK} bytes\n')
prints("run binary_addr successfully")
def test_chip_relocate84(layer_locate, outchl, C):
weight_width = align(outchl, BUS_WIDTH)
end = 0
for x in range(int(C / 2)): # outchannel/2组bn+weight
bn_locate = layer_locate[end + BUS_WIDTH * x:end + (x + 2) * BUS_WIDTH]
for i in range(len(bn_locate)):
if bn_locate[i] != '00':
prints(
f'test_chip_relocate84 bnlocate[{i}] failed value:{bn_locate[i]}'
)
return
# prints(f'test_chip_relocate84 {BUS_WIDTH*2} bn_locate:{len(bn_locate)} layer_locate[{end+BUS_WIDTH*x}:{end+(x+2)*BUS_WIDTH-1}]={bn_locate}')
# weight_locate = layer_locate[end + (x + 2) * BUS_WIDTH:end + weight_width*2 + (x + 2) * BUS_WIDTH]
end += weight_width * 2 + BUS_WIDTH
def test_chip_relocate84(layer_locate, outchl, C):
fmain.write(f'test_relocate...Start Rearrange weight...len(layer_locate):{len(layer_locate)}\n')
weight_width = align(outchl, BUS_WIDTH)
end = 0
for x in range(int(C/2)): # outchannel/2组bn+weight
bn_locate = layer_locate[end + BUS_WIDTH * x:end + (x + 2) * BUS_WIDTH]
for i in range(len(bn_locate)):
if bn_locate[i] != '00':
prints(f'test_chip_relocate84 bnlocate[{i}] failed value:{bn_locate[i]}')
return
# prints(f'test_chip_relocate84 {BUS_WIDTH*2} bn_locate:{len(bn_locate)} layer_locate[{end+BUS_WIDTH*x}:{end+(x+2)*BUS_WIDTH-1}]={bn_locate}')
# weight_locate = layer_locate[end + (x + 2) * BUS_WIDTH:end + weight_width*2 + (x + 2) * BUS_WIDTH]
end += weight_width * 2 + BUS_WIDTH
fmain.write(f'test_relocate...Endof Rearrange weight...len(layer_locate):{len(layer_locate)}\n')
fmain.write(f'test_chip_relocate84 fc weight data success\n')
def gen_bmp(layercnt, actbit, vgglog):
"""
description:
Generate model files
parameters:
vgglog: Path to vgglog file
return code:
None
"""
net = logpath.split('/')[1].split('.')[0]
cmd_list = [
f'python3 input/inout_print.py {net} imagenet/im6.bmp {ptpath} {bmpdtpath} {actbit} > .debug/genbmp.log 2>&1'
]
for i in range(len(cmd_list)):
os.system(cmd_list[i])
if layercnt > 1 and "resnet" not in net:
# Muse-v2 interface, v3 is not used.
deal_fc_k(vgglog)
prints("generate bmp files successfully")
def test_relocate(layer_locate, outchl, C, H, W):
weight_width = align(C * H * W, BUS_WIDTH)
if H == 1 and W == 1:
weight_width = align(outchl, BUS_WIDTH)
outchl = C
end = 0
for x in range(outchl): # outchannel组bn+weight
bn_locate = layer_locate[end + BUS_WIDTH * x:end + (x + 1) * BUS_WIDTH]
if H == 1 and W == 1:
for i in range(len(bn_locate)):
if bn_locate[i] != '00':
prints(
f'test_relocatetest bnlocate[{i}] failed value:{bn_locate[i]}'
)
return
# fmain.write(f'test_relocate {BUS_WIDTH} bn_locate:{len(bn_locate)} layer_locate[{end+BUS_WIDTH*x}:{end+(x+1)*BUS_WIDTH-1}]={bn_locate}\n')
# fmain.write(f'test_relocate {weight_width} weight_locate:{len(weight_locate)} layer_locate[{end+(x+1)*BUS_WIDTH}:{end+weight_width+(x+1)*BUS_WIDTH-1}]={weight_locate}\n')
end += weight_width
if H == 1 and W == 1:
fmain.write(f'test_relocate fc weight data success\n')
def weight_addr(name_list, data_list, active_bit, weight_bit):
"""
description:
Load pt file and format output
parameters:
loadpt: The Class of load_pt
return code:
None
"""
bn_bias = []
k = scale = 0
bn_weight = []
fc_flag = False
running_var = []
weight_bn_k = []
weight_bn_b = []
weight_data = []
global ADDRBLOCK
global BUS_WIDTH
word_address = 0
start = ADDRBLOCK
running_mean = []
layer_locate = []
N = C = H = W = 0
global datas_locate
global layers_wet_addr
layer_cnt = fc_cnt = 0
BUS_WIDTH = int(256 / weight_bit)
poollist,downsample = get_layer_num("pool")
fmain.write(f'weight_addr start:{ADDRBLOCK} bus_addr:{bus_address()} chiplet:{2**chiplet_id}\n')
for i in range(len(name_list)):
name = name_list[i]
data = data_list[i]
if pt_skip(name):
continue
# 从pt中获取running_mean running_var bn_bias scale和weight
elif "scale" in name:
scale = data.tolist()
elif "weight" in name:
if "classifier" in name or "fc" in name:
fc_flag = True
elif "bn" in name or "downsample.1" in name:
bn_weight = data
continue
weight_data = data
layer_cnt += 1
if layer_cnt in poollist:
layer_cnt += 1
elif "bias" in name:
bn_bias = data
elif "running_mean" in name:
running_mean = data
elif "running_var" in name:
running_var = data
# 通过running_mean running_var bn_bias bn_weight计算bn_k和bn_b
if len(running_mean) and len(running_var) and len(bn_bias) and len(bn_weight):
for i in range(len(bn_bias)):
bn_k = bn_weight[i] / (running_var[i].sqrt() + 1e-6)
bn_b = -bn_weight[i] * running_mean[i] / \
(running_var[i].sqrt() + 1e-6) + bn_bias[i]
weight_bn_k.append(bn_k.half().float())
weight_bn_b.append(bn_b.half().float())
running_mean = []
running_var = []
bn_weight = []
bn_bias = []
# activation地址排布和weight数据保存
if scale and (len(weight_bn_k) and len(weight_bn_b) and len(weight_data) or fc_flag):
if "classifier" in name or "fc" in name:
fc_cnt += 1
layer_cnt = get_layer_num(f'fc{fc_cnt}')
layer_name = f'layer{str(layer_cnt)}'
dim_list = list(weight_data.size())
dim_lens = len(dim_list)
if dim_lens == 4:
N = dim_list[0]
C = dim_list[1]
H = dim_list[2]
W = dim_list[3]
elif dim_lens == 2:
N = C = 1
H = dim_list[0]
W = dim_list[1]
for n in range(N):
if len(weight_bn_k) and len(weight_bn_b): # pt中的bnkb存入layer_locate
hexdata_k = '%X' % st.unpack('I', st.pack('f', weight_bn_k[n]))[0]
hexdata_b = '%X' % st.unpack('I', st.pack('f', weight_bn_b[n]))[0]
for x in range(len(str(hexdata_k))):
if (x % 2) == 0:
layer_locate.append(hexdata_k[x] + hexdata_k[x + 1])
for y in range(len(str(hexdata_b))):
if (y % 2) == 0:
layer_locate.append(hexdata_b[y] + hexdata_b[y + 1])
layer_locate = add_padding(layer_locate)
for c in range(C):
for h in range(H):
if dim_lens == 2:
for i in range(BUS_WIDTH):
layer_locate.append("00")
for w in range(W):
if dim_lens == 4:
word_addr = w + h * W + c * H * W
rounds = round(weight_data[n][c][h][w].tolist() / scale)
elif dim_lens == 2:
word_addr = w + h * W
rounds = round(weight_data[h][w].tolist() / scale)
else:
prints(f'Unknown weight:{name} length:{dim_lens} dim:{dim_list}')
continue
hexdata_w = HexProcess(rounds)
layer_locate.append(hexdata_w)
word_address += BUS_WIDTH # bn预留(256bit)
word_address += align(word_addr, BUS_WIDTH)
layer_locate = add_padding(layer_locate)
end = ADDRBLOCK + word_address
layers_wet_addr.append([f'{layer_name} weight start and end:', [start, end]])
start = end
if active_bit == 16 or active_bit == 4: # weight 16*8 4*4不重排
pass
elif fc_flag == False: # active_bit=8 conv weight重排
if chiplet_id:
fmain.write(f'{layer_name} relocate conv\'s weight...\n')
layer_locate = relocate(layer_locate, N, C, H, W)
# test_relocate(layer_locate, N, C, H, W)
else: # fc weight重排
if weight_bit == 8: # fc weight 8*8重排
layer_locate = fc_relocate88(layer_locate, W, H)
elif weight_bit == 4: # fc weight 8*4重排
layer_locate = fc_relocate84(layer_locate, W, H)
test_chip_relocate84(layer_locate, W, H)
if chiplet_id:
fmain.write(f'{layer_name} relocate fc\'s weight...\n')
if weight_bit == 8: # chiplet mode fc weight 8*8重排
layer_locate = relocate(layer_locate, W, H, 1, 1)
test_relocate(layer_locate, W, H, 1, 1)
elif weight_bit == 4: # chiplet mode fc weight 8*4重排
layer_locate = chip_relocate84(layer_locate, W, H)
test_chip_relocate84(layer_locate, W, H)
fmain.write("%7s bn_k+bn_b+conv_weight data save data success\n" % layer_name)
datas_locate[k] = [f'layer {str(layer_cnt)} data', [layer_locate]]
datas_locate.append(datas_locate[k])
layer_locate = []
weight_bn_k = []
weight_bn_b = []
weight_data = []
fc_flag = 0
scale = 0
k += 1
if layer_cnts <= k:
break
ADDRBLOCK += word_address
return poollist, downsample
def gen_txt(self):
"""
description:
Load pt file and format output
parameters:
loadpt: The Class of load_pt
return code:
None
"""
counts = 0
name_list = []
data_list = []
quant_list = []
onelayer_cnt = []
self.get_tensorinfo(self.netpath)
with open(self.ptpath, 'rb') as f:
ptdata = torch.load(self.ptpath, map_location=torch.device('cpu'))
chip_id = 0
if "vggnet16" in self.logpath or "resnet34" in self.logpath:
act_bit = ptdata['hyper_parameters']['act_quant_bit']
wet_bit = ptdata['hyper_parameters']['weight_quant_bit']
else:
act_bit = 8
wet_bit = 8
for k, v in ptdata['state_dict'].items():
k = k.split(".", 1)[1].strip()
name_list.append(k)
data_list.append(v)
if "resnet" not in self.netpath.lower():
if "quant_" in k or "classifier." in k:
quant_list.append(k)
quant_list.append(v)
if "resnet" in self.netpath.lower():
return name_list, data_list, act_bit, wet_bit, chip_id
else:
with open(f'{self.ptdtpath}/img.input.q.txt', 'w') as fq:
fq.write('{}{}'.format(self.in_q, '\n'))
self.fw.write(f'{self.ptdtpath}/img.input.q.txt write success\n')
for i in range(self.layer_cnts):
layer = f'layers.{i}.'
for j in range(len(name_list)):
if layer in name_list[j]:
self.layers.append([name_list[j], data_list[j]])
counts += 1
onelayer_cnt.append(str(counts))
counts = 0
del (self.layers[0])
for i in range(self.layer_cnts):
layername = f'layer_num:{str(i + 1)}'
self.layermsg = self.get_layer_info(self.logpath, layername)
self.splicing_output(int(onelayer_cnt[i]), counts, quant_list)
counts += int(onelayer_cnt[i])
scale = fcname = weight = ""
for i in range(len(quant_list)):
tmpstr = str(quant_list[i])
if ".scale" in tmpstr or ".weight" in tmpstr:
if ".scale" in tmpstr:
scale = quant_list[i + 1]
else:
fcname = quant_list[i]
weight = quant_list[i + 1]
if len(fcname) and len(str(scale)) and len(str(weight)):
write_data = threading.Thread(target=self.write_pt_data,
args=(fcname, weight, scale))
write_data.start()
write_data.join()
continue
elif "quant_" in tmpstr or "classifier" in tmpstr:
write_data = threading.Thread(target=self.write_pt_data,
args=(quant_list[i], quant_list[i + 1], scale))
write_data.start()
write_data.join()
prints("run gen_txt successfully")
return name_list, data_list, act_bit, wet_bit, chip_id