def __init__(self,
SimConfig_path,
indata=0,
rdata=0,
outprecision=8,
default_inbuf_size=16,
default_outbuf_size=4,
default_inchannel=64,
default_size=9):
# indata: volume of input data (for pooling) (Byte)
# rdata: volume of data from buffer to iReg (Byte)
# default_inbuf_size: the default PE-level input buffer size (unit: KB)
# default_outbuf_size: the default Tile-level output buffer size (unit: KB)
self.pooling = Pooling(SimConfig_path=SimConfig_path)
self.inbuf = buffer(SimConfig_path=SimConfig_path,
buf_level=1,
default_buf_size=default_inbuf_size)
self.inbuf.calculate_buf_write_latency(indata)
self.inbuf_wlatency = self.inbuf.buf_wlatency
# unit: ns
self.inbuf.calculate_buf_read_latency(rdata)
self.inbuf_rlatency = self.inbuf.buf_rlatency
self.pooling.calculate_Pooling_latency(inchannel=default_inchannel,
insize=default_size)
self.digital_latency = self.pooling.Pooling_latency
self.outbuf = buffer(SimConfig_path=SimConfig_path,
buf_level=2,
default_buf_size=default_outbuf_size)
self.outbuf.calculate_buf_write_latency(wdata=(default_inchannel *
outprecision / 8))
self.outbuf_rlatency = 0
self.outbuf_wlatency = self.outbuf.buf_wlatency
self.pooling_latency = self.inbuf_wlatency + self.inbuf_rlatency + self.digital_latency + self.outbuf_rlatency + self.outbuf_wlatency
def calculate_PE_area(self, SimConfig_path=None, default_inbuf_size=16):
# unit: um^2
self.inbuf = buffer(SimConfig_path=SimConfig_path,
buf_level=1,
default_buf_size=default_inbuf_size)
self.inbuf.calculate_buf_area()
self.calculate_xbar_area()
self.calculate_demux_area()
self.calculate_mux_area()
self.calculate_DAC_area()
self.calculate_ADC_area()
self.PE_adder.calculate_adder_area()
self.PE_shiftreg.calculate_shiftreg_area()
self.PE_iReg.calculate_reg_area()
self.PE_oReg.calculate_reg_area()
self.PE_xbar_area = self.PE_xbar_num * self.xbar_area
self.PE_ADC_area = self.ADC_area * self.PE_ADC_num
self.PE_DAC_area = self.DAC_area * self.PE_DAC_num
self.PE_adder_area = self.PE_group_ADC_num * self.PE_adder_num * self.PE_adder.adder_area
self.PE_shiftreg_area = self.PE_ADC_num * self.PE_shiftreg.shiftreg_area
self.PE_iReg_area = self.PE_DAC_num * self.PE_iReg.reg_area
self.PE_oReg_area = self.PE_ADC_num * self.PE_oReg.reg_area
self.PE_input_demux_area = self.input_demux_area * self.PE_DAC_num
self.PE_output_mux_area = self.output_mux_area * self.PE_ADC_num
self.PE_digital_area = self.PE_adder_area + self.PE_shiftreg_area + self.PE_input_demux_area + \
self.PE_output_mux_area + self.PE_iReg_area + self.PE_oReg_area
self.PE_inbuf_area = self.inbuf.buf_area
self.PE_area = self.PE_xbar_area + self.PE_ADC_area + self.PE_DAC_area + self.PE_digital_area + self.PE_inbuf_area
def __init__(self, SimConfig_path, read_row=0, read_column=0, indata=0, rdata=0, inprecision = 8,
PE_num=0, default_inbuf_size = 16, default_outbuf_size =4):
# read_row: activated WL number in crossbar
# read_column: activated BL number in crossbar
# indata: volume of input data (for PE) (Byte)
# rdata: volume of data from buffer to iReg (Byte)
# outdata: volume of output data (for PE) (Byte)
# inprecision: input data precision of each Xbar
# PE_num: used PE_number in one tile
# default_inbuf_size: the default PE-level input buffer size (unit: KB)
# default_outbuf_size: the default Tile-level output buffer size (unit: KB)
PE_latency_analysis.__init__(self, SimConfig_path, read_row=read_row, read_column=read_column,
indata=indata, rdata=rdata, inprecision=inprecision, default_buf_size = default_inbuf_size)
tilel_config = cp.ConfigParser()
tilel_config.read(SimConfig_path, encoding='UTF-8')
self.intra_tile_bandwidth = float(tilel_config.get('Tile level', 'Intra_Tile_Bandwidth'))
merge_time = math.ceil(math.log2(PE_num))
self.tile_PE_num = list(map(int, tilel_config.get('Tile level', 'PE_Num').split(',')))
if self.tile_PE_num[0] == 0:
self.tile_PE_num[0] = 4
self.tile_PE_num[1] = 4
assert self.tile_PE_num[0] > 0, "PE number in one PE < 0"
assert self.tile_PE_num[1] > 0, "PE number in one PE < 0"
self.tile_PE_total_num = self.tile_PE_num[0] * self.tile_PE_num[1]
assert PE_num <= self.tile_PE_total_num, "PE number exceeds the range"
self.outbuf = buffer(SimConfig_path=SimConfig_path, buf_level=2, default_buf_size=default_outbuf_size)
total_level = math.ceil(math.log2(self.tile_PE_total_num))
self.jointmodule_latency = merge_time * self.digital_period
self.transfer_latency = (total_level*(self.PE.ADC_precision+merge_time)-merge_time*(merge_time+1)/2)\
*read_column/self.intra_tile_bandwidth
self.outbuf.calculate_buf_write_latency(wdata=((self.PE.ADC_precision + merge_time)*read_column*PE_num/8))
self.tile_buf_rlatency = 0
self.tile_buf_wlatency = self.outbuf.buf_wlatency
# do not consider
self.tile_latency = self.PE_latency + self.jointmodule_latency + self.transfer_latency + self.tile_buf_wlatency
def calculate_PE_read_power_fast(self,
max_column=0,
max_row=0,
max_group=0,
SimConfig_path=None,
default_inbuf_size=16):
# unit: W
# coarse but fast estimation
# max_column: maximum used column in one crossbar in this tile
# max_row: maximum used row in one crossbar in this tile
# max_group: maximum used groups in one PE
self.inbuf = buffer(SimConfig_path=SimConfig_path,
buf_level=1,
default_buf_size=default_inbuf_size)
self.inbuf.calculate_buf_read_power()
self.inbuf.calculate_buf_write_power()
self.calculate_DAC_power()
self.calculate_ADC_power()
self.calculate_demux_power()
self.calculate_mux_power()
self.PE_shiftreg.calculate_shiftreg_power()
self.PE_iReg.calculate_reg_power()
self.PE_oReg.calculate_reg_power()
self.PE_adder.calculate_adder_power()
self.PE_read_power = 0
self.PE_xbar_read_power = 0
self.PE_ADC_read_power = 0
self.PE_DAC_read_power = 0
self.PE_adder_read_power = 0
self.PE_shiftreg_read_power = 0
self.PE_iReg_read_power = 0
self.PE_oReg_read_power = 0
self.input_demux_read_power = 0
self.output_mux_read_power = 0
self.PE_digital_read_power = 0
self.xbar_read_config(read_row=max_row, read_column=max_column)
self.calculate_xbar_read_power()
self.PE_xbar_read_power = self.PE_multiplex_xbar_num[
1] * max_group * self.xbar_read_power / self.input_demux / self.output_mux
self.PE_DAC_read_power = max_group * math.ceil(
max_row / self.input_demux) * self.DAC_power
self.PE_ADC_read_power = max_group * math.ceil(
max_column / self.output_mux) * self.ADC_power
self.input_demux_read_power = max_group * math.ceil(
max_row / self.input_demux) * self.input_demux_power
self.output_mux_read_power = max_group * math.ceil(
max_column / self.output_mux) * self.output_mux_power
self.PE_adder_read_power = (max_group - 1) * math.ceil(
max_column / self.output_mux) * self.PE_adder.adder_power
self.PE_shiftreg_read_power = max_group * math.ceil(
max_column / self.output_mux) * self.PE_shiftreg.shiftreg_power
self.PE_iReg_read_power = max_group * math.ceil(
max_row / self.input_demux) * self.PE_iReg.reg_power
self.PE_oReg_read_power = max_group * math.ceil(
max_column / self.output_mux) * self.PE_oReg.reg_power
self.PE_digital_read_power = self.input_demux_read_power + self.output_mux_read_power + self.PE_adder_read_power + self.PE_shiftreg_read_power + self.PE_iReg_read_power + self.PE_oReg_read_power
self.PE_inbuf_read_rpower = self.inbuf.buf_rpower * 1e-3
self.PE_inbuf_read_wpower = self.inbuf.buf_wpower * 1e-3
self.PE_inbuf_read_power = self.PE_inbuf_read_rpower + self.PE_inbuf_read_wpower
self.PE_read_power = self.PE_xbar_read_power + self.PE_DAC_read_power + self.PE_ADC_read_power + self.PE_digital_read_power + self.PE_inbuf_read_power
def calculate_model_area(self): #Todo: Noc area
self.graph.tile.calculate_tile_area(
SimConfig_path=self.SimConfig_path,
default_inbuf_size=self.graph.max_inbuf_size,
default_outbuf_size=self.graph.max_outbuf_size)
self.global_buf = buffer(SimConfig_path=self.SimConfig_path,
buf_level=1,
default_buf_size=self.graph.global_buf_size)
self.global_buf.calculate_buf_area()
self.global_add = adder(SimConfig_path=self.SimConfig_path,
bitwidth=self.graph.global_adder_bitwidth)
self.global_add.calculate_adder_area()
for i in range(self.total_layer_num):
tile_num = self.graph.layer_tileinfo[i]['tilenum']
self.arch_area[i] = self.graph.tile.tile_area * tile_num
self.arch_xbar_area[i] = self.graph.tile.tile_xbar_area * tile_num
self.arch_ADC_area[i] = self.graph.tile.tile_ADC_area * tile_num
self.arch_DAC_area[i] = self.graph.tile.tile_DAC_area * tile_num
self.arch_digital_area[
i] = self.graph.tile.tile_digital_area * tile_num
self.arch_adder_area[
i] = self.graph.tile.tile_adder_area * tile_num
self.arch_shiftreg_area[
i] = self.graph.tile.tile_shiftreg_area * tile_num
self.arch_iReg_area[i] = self.graph.tile.tile_iReg_area * tile_num
self.arch_oReg_area[i] = self.graph.tile.tile_oReg_area * tile_num
self.arch_input_demux_area[
i] = self.graph.tile.tile_input_demux_area * tile_num
self.arch_output_mux_area[
i] = self.graph.tile.tile_output_mux_area * tile_num
self.arch_jointmodule_area[
i] = self.graph.tile.tile_jointmodule_area * tile_num
self.arch_buf_area[i] = self.graph.tile.tile_buffer_area * tile_num
self.arch_pooling_area[
i] = self.graph.tile.tile_pooling_area * tile_num
self.arch_total_area = sum(self.arch_area)
self.arch_total_xbar_area = sum(self.arch_xbar_area)
self.arch_total_ADC_area = sum(self.arch_ADC_area)
self.arch_total_DAC_area = sum(self.arch_DAC_area)
self.arch_total_digital_area = sum(
self.arch_digital_area
) + self.global_add.adder_area * self.graph.global_adder_num
self.arch_total_adder_area = sum(
self.arch_adder_area
) + self.global_add.adder_area * self.graph.global_adder_num
self.arch_total_shiftreg_area = sum(self.arch_shiftreg_area)
self.arch_total_iReg_area = sum(self.arch_iReg_area)
self.arch_total_oReg_area = sum(self.arch_oReg_area)
self.arch_total_input_demux_area = sum(self.arch_input_demux_area)
self.arch_total_output_mux_area = sum(self.arch_output_mux_area)
self.arch_total_jointmodule_area = sum(self.arch_jointmodule_area)
self.arch_total_buf_area = sum(
self.arch_buf_area) + self.global_buf.buf_area
self.arch_total_pooling_area = sum(self.arch_pooling_area)
def calculate_tile_area(self,
SimConfig_path=None,
default_inbuf_size=16,
default_outbuf_size=4):
# unit: um^2
self.tile_area = 0
self.tile_xbar_area = 0
self.tile_ADC_area = 0
self.tile_DAC_area = 0
self.tile_input_demux_area = 0
self.tile_output_mux_area = 0
self.tile_shiftreg_area = 0
self.tile_iReg_area = 0
self.tile_oReg_area = 0
self.tile_adder_area = 0
self.tile_buffer_area = 0
self.tile_digital_area = 0
self.tile_adder.calculate_adder_area()
self.tile_shiftreg.calculate_shiftreg_area()
self.tile_iReg.calculate_reg_area()
self.tile_oReg.calculate_reg_area()
self.tile_jointmodule.calculate_jointmodule_area()
self.tile_buffer = buffer(SimConfig_path=SimConfig_path,
buf_level=2,
default_buf_size=default_outbuf_size)
self.tile_buffer.calculate_buf_area()
self.tile_pooling.calculate_Pooling_area()
for i in range(self.tile_PE_num[0]):
for j in range(self.tile_PE_num[1]):
self.tile_PE_list[i][j].calculate_PE_area(
SimConfig_path=SimConfig_path,
default_inbuf_size=default_inbuf_size)
self.tile_xbar_area += self.tile_PE_list[i][j].PE_xbar_area
self.tile_ADC_area += self.tile_PE_list[i][j].PE_ADC_area
self.tile_DAC_area += self.tile_PE_list[i][j].PE_DAC_area
# self.tile_digital_area += self.tile_PE_list[i][j].PE_digital_area
self.tile_input_demux_area += self.tile_PE_list[i][
j].PE_input_demux_area
self.tile_output_mux_area += self.tile_PE_list[i][
j].PE_output_mux_area
self.tile_shiftreg_area += self.tile_PE_list[i][
j].PE_shiftreg_area
self.tile_iReg_area += self.tile_PE_list[i][j].PE_iReg_area
self.tile_oReg_area += self.tile_PE_list[i][j].PE_oReg_area
self.tile_adder_area += self.tile_PE_list[i][j].PE_adder_area
self.tile_buffer_area += self.tile_PE_list[i][j].PE_inbuf_area
# self.tile_adder_area += self.tile_adder_num * self.tile_adder.adder_area
# self.tile_shiftreg_area += self.tile_shiftreg_num * self.tile_shiftreg.shiftreg_area
self.tile_jointmodule_area = self.tile_jointmodule_num * self.tile_jointmodule.jointmodule_area
self.tile_digital_area = self.tile_input_demux_area + self.tile_output_mux_area + self.tile_adder_area \
+ self.tile_shiftreg_area + self.tile_jointmodule_area + self.tile_iReg_area + self.tile_oReg_area
self.tile_pooling_area = self.tile_pooling.Pooling_area
self.tile_buffer_area += self.tile_buffer.buf_area
self.tile_area = self.tile_xbar_area + self.tile_ADC_area + self.tile_DAC_area + self.tile_digital_area + self.tile_buffer_area + self.tile_pooling_area
def calculate_model_energy(self):
#print(self.model_latency.total_buffer_r_latency)
self.global_buf = buffer(SimConfig_path=self.SimConfig_path,
buf_level=1,
default_buf_size=self.graph.global_buf_size)
self.global_buf.calculate_buf_read_power()
self.global_buf.calculate_buf_write_power()
self.global_add = adder(SimConfig_path=self.SimConfig_path,
bitwidth=self.graph.global_adder_bitwidth)
self.global_add.calculate_adder_power()
for i in range(self.total_layer_num):
tile_num = self.graph.layer_tileinfo[i]['tilenum']
self.arch_xbar_energy[i] = self.model_power.arch_xbar_power[
i] * self.model_latency.total_xbar_latency[i]
self.arch_ADC_energy[i] = self.model_power.arch_ADC_power[
i] * self.model_latency.total_ADC_latency[i]
self.arch_DAC_energy[i] = self.model_power.arch_DAC_power[
i] * self.model_latency.total_DAC_latency[i]
self.arch_adder_energy[i] = self.model_power.arch_adder_power[
i] * self.model_latency.total_adder_latency[i]
self.arch_shiftreg_energy[
i] = self.model_power.arch_shiftreg_power[
i] * self.model_latency.total_shiftreg_latency[i]
self.arch_iReg_energy[i] = self.model_power.arch_iReg_power[
i] * self.model_latency.total_iReg_latency[i]
self.arch_oReg_energy[i] = self.model_power.arch_oReg_power[
i] * self.model_latency.total_oReg_latency[i]
self.arch_input_demux_energy[
i] = self.model_power.arch_input_demux_power[
i] * self.model_latency.total_input_demux_latency[i]
self.arch_output_mux_energy[
i] = self.model_power.arch_output_mux_power[
i] * self.model_latency.total_output_mux_latency[i]
self.arch_jointmodule_energy[
i] = self.model_power.arch_jointmodule_power[
i] * self.model_latency.total_jointmodule_latency[i]
self.arch_buf_r_energy[i] = self.model_power.arch_buf_r_power[
i] * self.model_latency.total_buffer_r_latency[i]
self.arch_buf_w_energy[i] = self.model_power.arch_buf_w_power[
i] * self.model_latency.total_buffer_w_latency[i]
self.arch_buf_energy[
i] = self.arch_buf_r_energy[i] + self.arch_buf_w_energy[i]
self.arch_pooling_energy[i] = self.model_power.arch_pooling_power[
i] * self.model_latency.total_pooling_latency[i]
self.arch_digital_energy[i] = self.arch_shiftreg_energy[i]+self.arch_iReg_energy[i]+self.arch_oReg_energy[i]+\
self.arch_input_demux_energy[i]+self.arch_output_mux_energy[i]+self.arch_jointmodule_energy[i]
self.arch_energy[i] = self.arch_xbar_energy[i]+self.arch_ADC_energy[i]+self.arch_DAC_energy[i]+\
self.arch_digital_energy[i]+self.arch_buf_energy[i]+self.arch_pooling_energy[i]
self.arch_total_energy = sum(self.arch_energy) + self.arch_Noc_energy
self.arch_total_xbar_energy = sum(self.arch_xbar_energy)
self.arch_total_ADC_energy = sum(self.arch_ADC_energy)
self.arch_total_DAC_energy = sum(self.arch_DAC_energy)
self.arch_total_digital_energy = sum(self.arch_digital_energy)+\
self.global_add.adder_power*self.graph.global_adder_num*self.global_add.adder_latency
self.arch_total_adder_energy = sum(self.arch_adder_energy)+\
self.global_add.adder_power*self.graph.global_adder_num*self.global_add.adder_latency
self.arch_total_shiftreg_energy = sum(self.arch_shiftreg_energy)
self.arch_total_iReg_energy = sum(self.arch_iReg_energy)
self.arch_total_input_demux_energy = sum(self.arch_input_demux_energy)
self.arch_total_output_mux_energy = sum(self.arch_output_mux_energy)
self.arch_total_jointmodule_energy = sum(self.arch_jointmodule_energy)
self.arch_total_buf_energy = sum(self.arch_buf_energy) + self.global_buf.buf_rpower*1e-3*self.global_buf.buf_rlatency \
+ self.global_buf.buf_wpower*1e-3*self.global_buf.buf_wlatency
self.arch_total_buf_r_energy = sum(
self.arch_buf_r_energy
) + self.global_buf.buf_rpower * 1e-3 * self.global_buf.buf_rlatency
self.arch_total_buf_w_energy = sum(
self.arch_buf_w_energy
) + self.global_buf.buf_wpower * 1e-3 * self.global_buf.buf_wlatency
self.arch_total_pooling_energy = sum(self.arch_pooling_energy)
def calculate_model_area(self): #Todo: Noc area
self.graph.tile.calculate_tile_area(
SimConfig_path=self.SimConfig_path,
default_inbuf_size=self.graph.max_inbuf_size,
default_outbuf_size=self.graph.max_outbuf_size)
self.global_buf = buffer(SimConfig_path=self.SimConfig_path,
buf_level=1,
default_buf_size=self.graph.global_buf_size)
self.global_buf.calculate_buf_area()
self.global_add = adder(SimConfig_path=self.SimConfig_path,
bitwidth=self.graph.global_adder_bitwidth)
self.global_add.calculate_adder_area()
self.tile = tile(SimConfig_path=self.SimConfig_path)
self.tile_xbar_num = self.tile.tile_PE_total_num * self.tile.group_num * self.tile.xbar_column * self.tile.xbar_row
self.tile_DAC_num = self.tile.tile_PE_total_num * self.tile.group_num * self.tile.xbar_row
self.tile_ADC_num = self.tile.tile_PE_total_num * self.tile.group_num * self.tile.xbar_column
total_tile_num = 0
used_total_xbar_num = 0
used_total_DAC_num = 0
used_total_ADC_num = 0
# not the real DAC/ADC num, but it reflects the DAC/ADC num
for i in range(self.total_layer_num):
layer_dict = self.NetStruct[i][0][0]
tile_num = self.graph.layer_tileinfo[i]['tilenum']
self.arch_area[i] = self.graph.tile.tile_area * tile_num
self.arch_xbar_area[i] = self.graph.tile.tile_xbar_area * tile_num
self.arch_ADC_area[i] = self.graph.tile.tile_ADC_area * tile_num
self.arch_DAC_area[i] = self.graph.tile.tile_DAC_area * tile_num
self.arch_digital_area[
i] = self.graph.tile.tile_digital_area * tile_num
self.arch_adder_area[
i] = self.graph.tile.tile_adder_area * tile_num
self.arch_shiftreg_area[
i] = self.graph.tile.tile_shiftreg_area * tile_num
self.arch_iReg_area[i] = self.graph.tile.tile_iReg_area * tile_num
self.arch_oReg_area[i] = self.graph.tile.tile_oReg_area * tile_num
self.arch_input_demux_area[
i] = self.graph.tile.tile_input_demux_area * tile_num
self.arch_output_mux_area[
i] = self.graph.tile.tile_output_mux_area * tile_num
self.arch_jointmodule_area[
i] = self.graph.tile.tile_jointmodule_area * tile_num
self.arch_buf_area[i] = self.graph.tile.tile_buffer_area * tile_num
self.arch_pooling_area[
i] = self.graph.tile.tile_pooling_area * tile_num
if self.graph.layer_tileinfo[i]['type'] == 'conv':
# only consider the utilization rate of conv layer and fc layer
total_tile_num += tile_num
used_xbar_num = self.graph.layer_tileinfo[i][
'x_width'] * self.graph.layer_tileinfo[i]['y_height']
used_DAC_num = self.graph.layer_tileinfo[i][
'y_height'] * self.graph.layer_tileinfo[i][
'weight_precision'] * math.ceil(
int(layer_dict['Outputchannel']) /
self.tile.xbar_column)
used_ADC_num = self.graph.layer_tileinfo[i][
'x_width'] * self.graph.layer_tileinfo[i]['my']
self.arch_xbar_utilization[i] = used_xbar_num / (
tile_num * self.tile_xbar_num)
self.arch_DAC_utilization[i] = used_DAC_num / (
tile_num * self.tile_DAC_num)
self.arch_ADC_utilization[i] = used_ADC_num / (
tile_num * self.tile_ADC_num)
used_total_xbar_num += used_xbar_num
used_total_DAC_num += used_DAC_num
used_total_ADC_num += used_ADC_num
if self.graph.layer_tileinfo[i]['type'] == 'fc':
# only consider the utilization rate of conv layer and fc layer
total_tile_num += tile_num
used_xbar_num = self.graph.layer_tileinfo[i][
'x_width'] * self.graph.layer_tileinfo[i]['y_height']
used_DAC_num = self.graph.layer_tileinfo[i][
'y_height'] * self.graph.layer_tileinfo[i][
'weight_precision'] * math.ceil(
int(layer_dict['Outfeature']) /
self.tile.xbar_column)
used_ADC_num = self.graph.layer_tileinfo[i][
'x_width'] * self.graph.layer_tileinfo[i]['my']
self.arch_xbar_utilization[i] = used_xbar_num / (
tile_num * self.tile_xbar_num)
self.arch_DAC_utilization[i] = used_DAC_num / (
tile_num * self.tile_DAC_num)
self.arch_ADC_utilization[i] = used_ADC_num / (
tile_num * self.tile_ADC_num)
used_total_xbar_num += used_xbar_num
used_total_DAC_num += used_DAC_num
used_total_ADC_num += used_ADC_num
self.arch_total_area = sum(self.arch_area)
self.arch_total_xbar_area = sum(self.arch_xbar_area)
self.arch_total_ADC_area = sum(self.arch_ADC_area)
self.arch_total_DAC_area = sum(self.arch_DAC_area)
self.arch_total_digital_area = sum(
self.arch_digital_area
) + self.global_add.adder_area * self.graph.global_adder_num
self.arch_total_adder_area = sum(
self.arch_adder_area
) + self.global_add.adder_area * self.graph.global_adder_num
self.arch_total_shiftreg_area = sum(self.arch_shiftreg_area)
self.arch_total_iReg_area = sum(self.arch_iReg_area)
self.arch_total_oReg_area = sum(self.arch_oReg_area)
self.arch_total_input_demux_area = sum(self.arch_input_demux_area)
self.arch_total_output_mux_area = sum(self.arch_output_mux_area)
self.arch_total_jointmodule_area = sum(self.arch_jointmodule_area)
self.arch_total_buf_area = sum(
self.arch_buf_area) + self.global_buf.buf_area
self.arch_total_pooling_area = sum(self.arch_pooling_area)
self.arch_total_xbar_utilization = used_total_xbar_num / (
total_tile_num * self.tile_xbar_num)
self.arch_total_DAC_utilization = used_total_DAC_num / (
total_tile_num * self.tile_DAC_num)
self.arch_total_ADC_utilization = used_total_ADC_num / (
total_tile_num * self.tile_ADC_num)
def calculate_model_power(self):
self.global_buf = buffer(SimConfig_path=self.SimConfig_path,
buf_level=1,
default_buf_size=self.graph.global_buf_size)
self.global_buf.calculate_buf_read_power()
self.global_buf.calculate_buf_write_power()
self.global_add = adder(SimConfig_path=self.SimConfig_path,
bitwidth=self.graph.global_adder_bitwidth)
self.global_add.calculate_adder_power()
for i in range(self.total_layer_num):
tile_num = self.graph.layer_tileinfo[i]['tilenum']
max_column = self.graph.layer_tileinfo[i]['max_column']
max_row = self.graph.layer_tileinfo[i]['max_row']
max_PE = self.graph.layer_tileinfo[i]['max_PE']
max_group = self.graph.layer_tileinfo[i]['max_group']
layer_type = self.graph.net[i][0][0]['type']
self.graph.tile.calculate_tile_read_power_fast(
max_column=max_column,
max_row=max_row,
max_PE=max_PE,
max_group=max_group,
layer_type=layer_type,
SimConfig_path=self.SimConfig_path,
default_inbuf_size=self.graph.max_inbuf_size,
default_outbuf_size=self.graph.max_outbuf_size)
self.arch_power[i] = self.graph.tile.tile_read_power * tile_num
self.arch_xbar_power[
i] = self.graph.tile.tile_xbar_read_power * tile_num
self.arch_ADC_power[
i] = self.graph.tile.tile_ADC_read_power * tile_num
self.arch_DAC_power[
i] = self.graph.tile.tile_DAC_read_power * tile_num
self.arch_digital_power[
i] = self.graph.tile.tile_digital_read_power * tile_num
self.arch_adder_power[
i] = self.graph.tile.tile_adder_read_power * tile_num
self.arch_shiftreg_power[
i] = self.graph.tile.tile_shiftreg_read_power * tile_num
self.arch_iReg_power[
i] = self.graph.tile.tile_iReg_read_power * tile_num
self.arch_oReg_power[
i] = self.graph.tile.tile_oReg_read_power * tile_num
self.arch_input_demux_power[
i] = self.graph.tile.tile_input_demux_read_power * tile_num
self.arch_output_mux_power[
i] = self.graph.tile.tile_output_mux_read_power * tile_num
self.arch_jointmodule_power[
i] = self.graph.tile.tile_jointmodule_read_power * tile_num
self.arch_buf_power[
i] = self.graph.tile.tile_buffer_read_power * tile_num
self.arch_buf_r_power[
i] = self.graph.tile.tile_buffer_r_read_power * tile_num
self.arch_buf_w_power[
i] = self.graph.tile.tile_buffer_w_read_power * tile_num
self.arch_pooling_power[
i] = self.graph.tile.tile_pooling_read_power * tile_num
self.arch_total_power = sum(self.arch_power)
self.arch_total_xbar_power = sum(self.arch_xbar_power)
self.arch_total_ADC_power = sum(self.arch_ADC_power)
self.arch_total_DAC_power = sum(self.arch_DAC_power)
self.arch_total_digital_power = sum(
self.arch_digital_power
) + self.global_add.adder_power * self.graph.global_adder_num
self.arch_total_adder_power = sum(
self.arch_adder_power
) + self.global_add.adder_power * self.graph.global_adder_num
self.arch_total_shiftreg_power = sum(self.arch_shiftreg_power)
self.arch_total_iReg_power = sum(self.arch_iReg_power)
self.arch_total_oReg_power = sum(self.arch_oReg_power)
self.arch_total_input_demux_power = sum(self.arch_input_demux_power)
self.arch_total_output_mux_power = sum(self.arch_output_mux_power)
self.arch_total_jointmodule_power = sum(self.arch_jointmodule_power)
self.arch_total_buf_power = sum(self.arch_buf_power) + (
self.global_buf.buf_wpower + self.global_buf.buf_rpower) * 1e-3
self.arch_total_buf_r_power = sum(
self.arch_buf_r_power) + self.global_buf.buf_rpower * 1e-3
self.arch_total_buf_w_power = sum(
self.arch_buf_w_power) + self.global_buf.buf_wpower * 1e-3
self.arch_total_pooling_power = sum(self.arch_pooling_power)
def calculate_tile_read_power_fast(self,
max_column=0,
max_row=0,
max_PE=0,
max_group=0,
layer_type=None,
SimConfig_path=None,
default_inbuf_size=16,
default_outbuf_size=4):
# max_column: maximum used column in one crossbar in this tile
# max_row: maximum used row in one crossbar in this tile
# max_PE: maximum used PE in this tile
# max_group: maximum used groups in one PE
# unit: W
# coarse but fast estimation
self.tile_read_power = 0
self.tile_xbar_read_power = 0
self.tile_ADC_read_power = 0
self.tile_DAC_read_power = 0
self.tile_digital_read_power = 0
self.tile_adder_read_power = 0
self.tile_shiftreg_read_power = 0
self.tile_iReg_read_power = 0
self.tile_oReg_read_power = 0
self.tile_input_demux_read_power = 0
self.tile_output_mux_read_power = 0
self.tile_jointmodule_read_power = 0
self.tile_pooling_read_power = 0
self.tile_buffer_read_power = 0
self.tile_buffer_r_read_power = 0
self.tile_buffer_w_read_power = 0
self.tile_buffer = buffer(SimConfig_path=SimConfig_path,
buf_level=2,
default_buf_size=default_outbuf_size)
if layer_type == 'pooling':
self.tile_pooling.calculate_Pooling_power()
self.tile_pooling_read_power = self.tile_pooling.Pooling_power
elif layer_type == 'conv' or layer_type == 'fc':
self.calculate_PE_read_power_fast(
max_column=max_column,
max_row=max_row,
max_group=max_group,
SimConfig_path=SimConfig_path,
default_inbuf_size=default_inbuf_size)
self.tile_xbar_read_power = max_PE * self.PE_xbar_read_power
self.tile_ADC_read_power = max_PE * self.PE_ADC_read_power
self.tile_DAC_read_power = max_PE * self.PE_DAC_read_power
self.tile_adder_read_power = max_PE * self.PE_adder_read_power
self.tile_shiftreg_read_power = max_PE * self.PE_shiftreg_read_power
self.tile_iReg_read_power = max_PE * self.PE_iReg_read_power
self.tile_oReg_read_power = max_PE * self.PE_oReg_read_power
self.tile_input_demux_read_power = max_PE * self.input_demux_read_power
self.tile_output_mux_read_power = max_PE * self.output_mux_read_power
self.tile_jointmodule_read_power = (max_PE - 1) * math.ceil(
max_column /
self.output_mux) * self.tile_jointmodule.jointmodule_power
self.tile_digital_read_power = self.tile_adder_read_power+self.tile_shiftreg_read_power+\
self.tile_input_demux_read_power+self.tile_output_mux_read_power+self.tile_jointmodule_read_power
self.tile_buffer_r_read_power = max_PE * self.PE_inbuf_read_rpower
self.tile_buffer_w_read_power = max_PE * self.PE_inbuf_read_wpower
self.tile_buffer.calculate_buf_read_power()
self.tile_buffer.calculate_buf_write_power()
self.tile_buffer_r_read_power += self.tile_buffer.buf_rpower * 1e-3
self.tile_buffer_w_read_power += self.tile_buffer.buf_wpower * 1e-3
self.tile_buffer_read_power = self.tile_buffer_r_read_power + self.tile_buffer_w_read_power
self.tile_digital_read_power = self.tile_adder_read_power+self.tile_shiftreg_read_power+self.tile_iReg_read_power+self.tile_oReg_read_power+\
self.tile_input_demux_read_power+self.tile_output_mux_read_power+self.tile_jointmodule_read_power
self.tile_read_power = self.tile_xbar_read_power+self.tile_ADC_read_power+self.tile_DAC_read_power+\
self.tile_digital_read_power+self.tile_pooling_read_power+self.tile_buffer_read_power
def __init__(self, SimConfig_path):
# layer_num is a list with the size of 1xPE_num
ProcessElement.__init__(self, SimConfig_path)
tile_config = cp.ConfigParser()
tile_config.read(SimConfig_path, encoding='UTF-8')
self.tile_PE_num = list(
map(int,
tile_config.get('Tile level', 'PE_Num').split(',')))
if self.tile_PE_num[0] == 0:
self.tile_PE_num[0] = 4
self.tile_PE_num[1] = 4
assert self.tile_PE_num[0] > 0, "PE number in one PE < 0"
assert self.tile_PE_num[1] > 0, "PE number in one PE < 0"
self.tile_PE_total_num = self.tile_PE_num[0] * self.tile_PE_num[1]
self.tile_simulation_level = int(
tile_config.get('Algorithm Configuration', 'Simulation_Level'))
self.tile_PE_list = []
self.tile_PE_enable = []
for i in range(self.tile_PE_num[0]):
self.tile_PE_list.append([])
self.tile_PE_enable.append([])
for j in range(self.tile_PE_num[1]):
__PE = ProcessElement(SimConfig_path)
self.tile_PE_list[i].append(__PE)
self.tile_PE_enable[i].append(0)
self.layer_type = 'conv'
self.tile_layer_num = 0
self.tile_activation_precision = 0
self.tile_sliding_times = 0
self.tile_adder_num = 0
self.tile_shiftreg_num = 0
self.tile_jointmodule_num = 0
self.tile_adder = adder(SimConfig_path)
self.tile_shiftreg = shiftreg(SimConfig_path)
self.tile_iReg = reg(SimConfig_path)
self.tile_oReg = reg(SimConfig_path)
self.tile_jointmodule = JointModule(SimConfig_path)
self.tile_buffer = buffer(SimConfig_path)
self.tile_pooling = Pooling(SimConfig_path)
self.tile_utilization = 0
self.num_occupied_PE = 0
self.tile_area = 0
self.tile_xbar_area = 0
self.tile_ADC_area = 0
self.tile_DAC_area = 0
self.tile_digital_area = 0
self.tile_adder_area = 0
self.tile_shiftreg_area = 0
self.tile_iReg_area = 0
self.tile_oReg_area = 0
self.tile_input_demux_area = 0
self.tile_output_mux_area = 0
self.tile_jointmodule_area = 0
self.tile_pooling_area = 0
self.tile_buffer_area = 0
self.tile_read_power = 0
self.tile_xbar_read_power = 0
self.tile_ADC_read_power = 0
self.tile_DAC_read_power = 0
self.tile_digital_read_power = 0
self.tile_adder_read_power = 0
self.tile_shiftreg_read_power = 0
self.tile_iReg_read_power = 0
self.tile_oReg_read_power = 0
self.tile_input_demux_read_power = 0
self.tile_output_mux_read_power = 0
self.tile_jointmodule_read_power = 0
self.tile_pooling_read_power = 0
self.tile_buffer_read_power = 0
self.tile_buffer_r_read_power = 0
self.tile_buffer_w_read_power = 0
self.tile_write_power = 0
self.tile_xbar_write_power = 0
self.tile_ADC_write_power = 0
self.tile_DAC_write_power = 0
self.tile_digital_write_power = 0
self.tile_adder_write_power = 0
self.tile_shiftreg_write_power = 0
self.tile_iReg_write_power = 0
self.tile_input_demux_write_power = 0
self.tile_output_mux_write_power = 0
self.tile_jointmodule_write_power = 0
self.tile_read_latency = 0
self.tile_xbar_read_latency = 0
self.tile_ADC_read_latency = 0
self.tile_DAC_read_latency = 0
self.tile_digital_read_latency = 0
self.tile_adder_read_latency = 0
self.tile_shiftreg_read_latency = 0
self.tile_iReg_read_latency = 0
self.tile_input_demux_read_latency = 0
self.tile_output_mux_read_latency = 0
self.tile_jointmodule_read_latency = 0
# self.tile_layer_read_latency = {0:0}
self.tile_write_latency = 0
self.tile_xbar_write_latency = 0
self.tile_ADC_write_latency = 0
self.tile_DAC_write_latency = 0
self.tile_digital_write_latency = 0
self.tile_adder_write_latency = 0
self.tile_shiftreg_write_latency = 0
self.tile_iReg_write_latency = 0
self.tile_input_demux_write_latency = 0
self.tile_output_mux_write_latency = 0
self.tile_jointmodule_write_latency = 0
# self.tile_layer_write_latency = {0:0}
self.tile_read_energy = 0
self.tile_xbar_read_energy = 0
self.tile_ADC_read_energy = 0
self.tile_DAC_read_energy = 0
self.tile_digital_read_energy = 0
self.tile_adder_read_energy = 0
self.tile_shiftreg_read_energy = 0
self.tile_iReg_read_energy = 0
self.tile_input_demux_read_energy = 0
self.tile_output_mux_read_energy = 0
self.tile_jointmodule_read_energy = 0
self.tile_write_energy = 0
self.tile_xbar_write_energy = 0
self.tile_ADC_write_energy = 0
self.tile_DAC_write_energy = 0
self.tile_digital_write_energy = 0
self.tile_adder_write_energy = 0
self.tile_shiftreg_write_energy = 0
self.tile_iReg_write_energy = 0
self.tile_input_demux_write_energy = 0
self.tile_output_mux_write_energy = 0
self.tile_jointmodule_write_energy = 0
# print("tile configuration is loaded")
self.calculate_intra_PE_connection()
def update_tile_buf_size(self, SimConfig_path, default_buf_size=16):
self.tile_buffer = buffer(SimConfig_path=SimConfig_path,
default_buf_size=default_buf_size)
def __init__(self,
SimConfig_path,
read_row=0,
read_column=0,
indata=0,
rdata=0,
inprecision=8,
default_buf_size=16):
# read_row: activated WL number in crossbar
# read_column: activated BL number in crossbar
# indata: volume of input data (for PE) (Byte)
# rdata: volume of data from buffer to iReg (Byte)
# outdata: volume of output data (for PE) (Byte)
# inprecision: input data precision of each Xbar
# default_buf_size: default input buffer size (KB)
PEl_config = cp.ConfigParser()
PEl_config.read(SimConfig_path, encoding='UTF-8')
self.inbuf = buffer(SimConfig_path=SimConfig_path,
buf_level=1,
default_buf_size=default_buf_size)
self.PE = ProcessElement(SimConfig_path)
self.inbuf.calculate_buf_write_latency(indata)
self.PE_buf_wlatency = self.inbuf.buf_wlatency
# unit: ns
self.digital_period = 1 / float(
PEl_config.get('Digital module', 'Digital_Frequency')) * 1e3
self.inbuf.calculate_buf_read_latency(rdata)
self.PE_buf_rlatency = self.inbuf.buf_rlatency
multiple_time = math.ceil(inprecision/self.PE.DAC_precision) * math.ceil(read_row/self.PE.PE_group_DAC_num) *\
math.ceil(read_column/self.PE.PE_group_ADC_num)
self.PE.calculate_xbar_read_latency()
Transistor_Tech = int(
PEl_config.get('Crossbar level', 'Transistor_Tech'))
XBar_size = list(
map(float,
PEl_config.get('Crossbar level', 'Xbar_Size').split(',')))
DAC_num = int(PEl_config.get('Process element level', 'DAC_Num'))
ADC_num = int(PEl_config.get('Process element level', 'ADC_Num'))
Row = XBar_size[0]
Column = XBar_size[1]
# ns (using NVSim)
decoderLatency_dict = {
1: 0.27933 # 1:8, technology 65nm
}
decoder1_8 = decoderLatency_dict[1]
Row_per_DAC = math.ceil(Row / DAC_num)
m = 1
while Row_per_DAC > 0:
Row_per_DAC = Row_per_DAC // 8
m += 1
self.decoderLatency = m * decoder1_8
# ns
muxLatency_dict = {1: 32.744 / 1000}
mux8_1 = muxLatency_dict[1]
m = 1
Column_per_ADC = math.ceil(Column / ADC_num)
while Column_per_ADC > 0:
Column_per_ADC = Column_per_ADC // 8
m += 1
self.muxLatency = m * mux8_1
self.xbar_latency = multiple_time * self.PE.xbar_read_latency
self.PE.calculate_DAC_latency()
self.DAC_latency = multiple_time * self.PE.DAC_latency
self.PE.calculate_ADC_latency()
self.ADC_latency = multiple_time * self.PE.ADC_latency
self.iReg_latency = math.ceil(read_row/self.PE.PE_group_DAC_num)*math.ceil(read_column/self.PE.PE_group_ADC_num)*self.digital_period+\
multiple_time*self.digital_period
# write and read
self.shiftreg_latency = multiple_time * self.digital_period
self.input_demux_latency = multiple_time * self.decoderLatency
self.adder_latency = math.ceil(
read_column / self.PE.PE_group_ADC_num) * math.ceil(
math.log2(self.PE.group_num)) * self.digital_period
self.output_mux_latency = multiple_time * self.muxLatency
self.computing_latency = self.DAC_latency + self.xbar_latency + self.ADC_latency
self.oreg_latency = math.ceil(
read_column / self.PE.PE_group_ADC_num) * self.digital_period
self.PE_digital_latency = self.iReg_latency + self.shiftreg_latency + self.input_demux_latency + \
self.adder_latency + self.output_mux_latency + self.oreg_latency
self.PE_latency = self.PE_buf_wlatency + self.PE_buf_rlatency + self.computing_latency + self.PE_digital_latency