def instantiate(self, setup):
self.class_name = 'PE'
self.row = setup['row']
self.col = setup['col']
self.debug = 'PE[' + str(self.row) + ']' + '[' + str(self.col) + ']'
# ================================================================
# Stats Related Setup
# ================================================================
self.component_class_specification_stats = 'hide'
self.component_specification_stats = 'show'
self.access_counts_stats = 'show'
self.recorder = nnsimRecorder()\
if self.traces_stats == 'show'\
else None
# =================================================================
# IO Channels
# =================================================================
self.weights_data_in_chn = ModuleList(setup['weights_data_in_chn'])
self.ifmap_data_in_chn = ModuleList(setup['ifmap_data_in_chn'])
self.psum_data_in_chn = ModuleList(setup['psum_data_in_chn'])
self.psum_data_out_chn = ModuleList(setup['psum_data_out_chn'])
# =================================================================
# Hardware components
# =================================================================
# >>>> weights scratchpad (reg)
self.weights_reader = ModuleList(Ch())
weights_sp_setup = {'fill_data_ichns': self.weights_data_in_chn,\
'drain_data_ochns': self.weights_reader,\
'num_logical_managers': 1,\
'SRAM': {'depth': setup['wsp_depth'],\
'width': setup['wsp_width'],\
'data_width': setup['wsp_data_width'],\
'nports': setup['wsp_nports'],\
'nbanks': setup['wsp_nbanks'], \
'port_type': setup['wsp_port_type']},\
'debug': self.debug + '_weights_sp'}
self.weight_sp = WeightsSP(weights_sp_setup)
# >>>> mac unit
mac_setup = {'opa_chn': self.ifmap_data_in_chn[0],\
'opb_chn': self.weights_reader[0], \
'opc_chn': self.psum_data_in_chn[0],\
'result_chn': self.psum_data_out_chn[0],\
'latency': setup['mac_latency'],\
'debug': self.debug}
self.mac = mac(mac_setup)
def instantiate(self, setup):
# nnsimTestBench class sets trace generator if there is one
nnsimTestBench.instantiate(self, setup)
self.trace_generator = ws_chip_trace_generator()
self.traces_stats = 'show'
self.recorder = nnsimRecorder() if self.traces_stats == 'show' else None
self.generated_trace = True
# bookkeeping variables
self.result = []
# ---------------------------------------------------------------------
# IO channels for onchip and offchip communications
# ---------------------------------------------------------------------
# -> input channels for sending the input trace data to the GLBs
# smartbuffer needs channels specified as list format
self.weights_in_chn = ModuleList(Channel())
self.ifmap_chn = ModuleList(Channel())
self.psum_in_chn = ModuleList(Channel())
# -> output channel for reciving calculated psum from the chip
self.psum_out_chn = Channel()
# ---------------------------------------------------------------------
# Setup and instantiation of the design under test
# ---------------------------------------------------------------------
chip_setup = {}
chip_setup.update(self.arch)
# attributes that can be automatically derived
chip_setup['weights_seri_ratio'] = self.arch['width']['WeightsGLB']
chip_setup['ifmap_seri_ratio'] = self.arch['width']['IFmapGLB']
chip_setup['psum_seri_ratio'] = self.arch['width']['PsumGLB']
chip_setup['pe_array'] = self.arch['PE_array']
chip_setup['io_chns'] = {'weights': self.weights_in_chn,\
'ifmap': self.ifmap_chn,\
'psum_in': self.psum_in_chn,\
'psum_out': self.psum_out_chn}
chip_setup['PE'] = self.arch['PE']
self.dut = chip(chip_setup)
# for the use of dispatching data inside the testbench
self.WGLB_width = self.arch['width']['WeightsGLB']
self.IGLB_width = self.arch['width']['IFmapGLB']
self.PGLB_width = self.arch['width']['PsumGLB']
def instantiate(self, arr_x, arr_y,
input_chn, output_chn,
chn_per_word,
ifmap_glb_depth):
# PE static configuration (immutable)
self.name = 'chip'
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.post_tr_x = arr_x # num output channels = 8
self.post_tr_y = 4 # num tiles = 4
self.pre_tr_ifmap_x = arr_y # num input channels = 4
self.pre_tr_ifmap_y = 4 # num tiles = 4
self.pre_tr_weights_x = arr_y # num input channels = 4
self.pre_tr_weights_y = arr_x # num output channels = 8
self.stat_type = 'show'
# Instantiate DRAM IO channels
self.input_chn = input_chn
self.output_chn = output_chn
# Instantiate input deserializer and output serializer
self.ifmap_wr_chn = Channel()
self.weights_wr_chn = Channel()
self.bias_wr_chn = Channel()
self.deserializer = InputDeserializer(self.input_chn, self.ifmap_wr_chn,
self.weights_wr_chn, self.bias_wr_chn, arr_x, arr_y,
chn_per_word)
self.psum_output_chn = Channel()
self.serializer = OutputSerializer(self.output_chn, self.psum_output_chn)
# Instantiate GLB and GLB channels
self.ifmap_glb_wr_chn = Channel(3)
self.ifmap_rd_chn = Channel(3)
self.ifmap_glb = IFMapGLB(self.ifmap_glb_wr_chn, self.ifmap_rd_chn,
ifmap_glb_depth, chn_per_word)
self.psum_rd_chn = Channel(3)
self.psum_noc_wr_chn = Channel()
# self.psum_glb = PSumGLB(self.psum_wr_chn, self.psum_noc_wr_chn, self.psum_rd_chn,
# psum_glb_depth, chn_per_word)
self.weights_glb_wr_chn = Channel(3)
self.weights_rd_chn = Channel()
self.weights_glb = WeightsGLB(self.weights_glb_wr_chn, self.weights_rd_chn)
self.bias_rd_chn = Channel()
self.bias_glb = BiasGLB(self.bias_wr_chn, self.bias_rd_chn)
# PE Array and local channel declaration
self.pe_array = ModuleList()
self.pe_ifmap_chns = ModuleList()
self.pe_filter_chns = ModuleList()
self.pe_psum_chns = ModuleList()
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_psum_chns[0].append(Channel(32))
# Actual PE array instantiation
for y in range(self.arr_y):
self.pe_array.append(ModuleList())
self.pe_ifmap_chns.append(ModuleList())
self.pe_filter_chns.append(ModuleList())
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_ifmap_chns[y].append(Channel(32))
self.pe_filter_chns[y].append(Channel(32))
self.pe_psum_chns[y+1].append(Channel(32))
self.pe_array[y].append(
PE(x, y,
self.pe_ifmap_chns[y][x],
self.pe_filter_chns[y][x],
self.pe_psum_chns[y][x],
self.pe_psum_chns[y+1][x]
)
)
# Pre Transform IFMap array and local channel declaration
self.pre_tr_ifmap_array = ModuleList()
self.pre_tr_ifmap_in_chns = ModuleList()
self.pre_tr_ifmap_out_chns = ModuleList()
# Actual pre transform IFMap array instantiation
for y in range(self.pre_tr_ifmap_y):
self.pre_tr_ifmap_array.append(ModuleList())
self.pre_tr_ifmap_in_chns.append(ModuleList())
self.pre_tr_ifmap_out_chns.append(ModuleList())
for x in range(self.pre_tr_ifmap_x):
self.pre_tr_ifmap_in_chns[y].append(Channel(32))
self.pre_tr_ifmap_out_chns[y].append(Channel(32))
self.pre_tr_ifmap_array[y].append(
PreTransformIFMap(x, y,
self.pre_tr_ifmap_in_chns[y][x],
self.pre_tr_ifmap_out_chns[y][x]
)
)
# Pre Transform Weight array and local channel declaration
self.pre_tr_weights_array = ModuleList()
self.pre_tr_weights_in_chns = ModuleList()
self.pre_tr_weights_out_chns = ModuleList()
# Actual pre transform Weight array instantiation
for y in range(self.pre_tr_weights_y):
self.pre_tr_weights_array.append(ModuleList())
self.pre_tr_weights_in_chns.append(ModuleList())
self.pre_tr_weights_out_chns.append(ModuleList())
for x in range(self.pre_tr_weights_x):
self.pre_tr_weights_in_chns[y].append(Channel(32))
self.pre_tr_weights_out_chns[y].append(Channel(32))
self.pre_tr_weights_array[y].append(
PreTransformWeights(x, y,
self.pre_tr_weights_in_chns[y][x],
self.pre_tr_weights_out_chns[y][x]
)
)
# Post Transform Array and local channel declaration
self.post_tr_array = ModuleList()
self.post_tr_bias_chns = ModuleList()
self.post_tr_ofmap_in_chns = ModuleList()
self.post_tr_ofmap_out_chns = ModuleList()
# Actual post transform array instantiation
for y in range(self.post_tr_y):
self.post_tr_array.append(ModuleList())
self.post_tr_bias_chns.append(ModuleList())
self.post_tr_ofmap_in_chns.append(ModuleList())
self.post_tr_ofmap_out_chns.append(ModuleList())
for x in range(self.post_tr_x):
self.post_tr_bias_chns[y].append(Channel(32))
self.post_tr_ofmap_in_chns[y].append(Channel(32))
self.post_tr_ofmap_out_chns[y].append(Channel(32))
self.post_tr_array[y].append(
PostTransform(x, y,
self.post_tr_bias_chns[y][x],
self.post_tr_ofmap_in_chns[y][x],
self.post_tr_ofmap_out_chns[y][x]
)
)
# Setup NoC to deliver weights, ifmaps and psums
self.filter_noc = WeightsNoC(self.weights_rd_chn, self.pe_filter_chns, self.chn_per_word)
self.ifmap_noc = IFMapNoC(self.ifmap_rd_chn, self.pe_ifmap_chns, self.arr_x, self.chn_per_word)
self.psum_rd_noc = PSumRdNoC(self.pe_psum_chns[0], self.chn_per_word)
#self.psum_wr_noc = PSumWrNoC(self.pe_psum_chns[-1], self.psum_output_chn, self.chn_per_word)
self.bias_noc = BiasNoC(self.bias_rd_chn, self.post_tr_bias_chns, self.chn_per_word)
# Setup NoC for post transform blocks
self.post_tr_wr_noc = PostTrWrNoC(self.pe_psum_chns[-1], self.post_tr_ofmap_in_chns, self.chn_per_word)
self.post_tr_rd_noc = PostTrRdNoC(self.post_tr_ofmap_out_chns, self.psum_output_chn, self.chn_per_word)
# Instantiate tiler for ifmaps
self.ifmap_tiler = IFMapTiler(self.ifmap_wr_chn, self.pre_tr_ifmap_in_chns, self.chn_per_word)
# Setup NoC for pre transform blocks
#self.pre_tr_ifmap_wr_noc = PreTrIFMapWrNoC(self.ifmap_wr_chn, self.pre_tr_ifmap_in_chns, self.chn_per_word)
self.pre_tr_ifmap_rd_noc = PreTrIFMapRdNoC(self.pre_tr_ifmap_out_chns, self.ifmap_glb_wr_chn, self.chn_per_word)
self.pre_tr_weights_wr_noc = PreTrWeightsWrNoC(self.weights_wr_chn, self.pre_tr_weights_in_chns, self.chn_per_word)
self.pre_tr_weights_rd_noc = PreTrWeightsRdNoC(self.pre_tr_weights_out_chns, self.weights_glb_wr_chn, self.chn_per_word)
class WSArch(Module):
def instantiate(self, arr_x, arr_y,
input_chn, output_chn,
chn_per_word,
ifmap_glb_depth):
# PE static configuration (immutable)
self.name = 'chip'
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.post_tr_x = arr_x # num output channels = 8
self.post_tr_y = 4 # num tiles = 4
self.stat_type = 'show'
# Instantiate DRAM IO channels
self.input_chn = input_chn
self.output_chn = output_chn
# Instantiate input deserializer and output serializer
self.ifmap_wr_chn = Channel()
self.weights_wr_chn = Channel()
self.bias_wr_chn = Channel()
self.deserializer = InputDeserializer(self.input_chn, self.ifmap_wr_chn,
self.weights_wr_chn, self.bias_wr_chn, arr_x, arr_y,
chn_per_word)
self.psum_output_chn = Channel()
self.serializer = OutputSerializer(self.output_chn, self.psum_output_chn)
# Instantiate GLB and GLB channels
self.ifmap_rd_chn = Channel(3)
self.ifmap_glb = IFMapGLB(self.ifmap_wr_chn, self.ifmap_rd_chn,
ifmap_glb_depth, chn_per_word)
self.psum_rd_chn = Channel(3)
self.psum_noc_wr_chn = Channel()
# self.psum_glb = PSumGLB(self.psum_wr_chn, self.psum_noc_wr_chn, self.psum_rd_chn,
# psum_glb_depth, chn_per_word)
self.weights_rd_chn = Channel()
self.weights_glb = WeightsGLB(self.weights_wr_chn, self.weights_rd_chn)
self.bias_rd_chn = Channel()
self.bias_glb = BiasGLB(self.bias_wr_chn, self.bias_rd_chn)
# PE Array and local channel declaration
self.pe_array = ModuleList()
self.pe_ifmap_chns = ModuleList()
self.pe_filter_chns = ModuleList()
self.pe_psum_chns = ModuleList()
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_psum_chns[0].append(Channel(32))
# Actual PE array instantiation
for y in range(self.arr_y):
self.pe_array.append(ModuleList())
self.pe_ifmap_chns.append(ModuleList())
self.pe_filter_chns.append(ModuleList())
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_ifmap_chns[y].append(Channel(32))
self.pe_filter_chns[y].append(Channel(32))
self.pe_psum_chns[y+1].append(Channel(32))
self.pe_array[y].append(
PE(x, y,
self.pe_ifmap_chns[y][x],
self.pe_filter_chns[y][x],
self.pe_psum_chns[y][x],
self.pe_psum_chns[y+1][x]
)
)
# Post Transform Array and local channel declaration
self.post_tr_array = ModuleList()
self.post_tr_bias_chns = ModuleList()
self.post_tr_ofmap_in_chns = ModuleList()
self.post_tr_ofmap_out_chns = ModuleList()
# Actual post transform array instantiation
for y in range(self.post_tr_y):
self.post_tr_array.append(ModuleList())
self.post_tr_bias_chns.append(ModuleList())
self.post_tr_ofmap_in_chns.append(ModuleList())
self.post_tr_ofmap_out_chns.append(ModuleList())
for x in range(self.post_tr_x):
self.post_tr_bias_chns[y].append(Channel(32))
self.post_tr_ofmap_in_chns[y].append(Channel(32))
self.post_tr_ofmap_out_chns[y].append(Channel(32))
self.post_tr_array[y].append(
PostTransform(x, y,
self.post_tr_bias_chns[y][x],
self.post_tr_ofmap_in_chns[y][x],
self.post_tr_ofmap_out_chns[y][x],
)
)
# Setup NoC to deliver weights, ifmaps and psums
self.filter_noc = WeightsNoC(self.weights_rd_chn, self.pe_filter_chns, self.chn_per_word)
self.ifmap_noc = IFMapNoC(self.ifmap_rd_chn, self.pe_ifmap_chns, self.arr_x, self.chn_per_word)
self.psum_rd_noc = PSumRdNoC(self.pe_psum_chns[0], self.chn_per_word)
#self.psum_wr_noc = PSumWrNoC(self.pe_psum_chns[-1], self.psum_output_chn, self.chn_per_word)
self.bias_noc = BiasNoC(self.bias_rd_chn, self.post_tr_bias_chns, self.chn_per_word)
self.post_tr_wr_noc = PostTrWrNoC(self.pe_psum_chns[-1], self.post_tr_ofmap_in_chns, self.chn_per_word)
self.post_tr_rd_noc = PostTrRdNoC(self.post_tr_ofmap_out_chns, self.psum_output_chn, self.chn_per_word)
def configure(self, image_size, filter_size, in_chn, out_chn):
in_sets = self.arr_y//self.chn_per_word
out_sets = self.arr_x//self.chn_per_word
fmap_per_iteration = image_size[0]*image_size[1]
num_iteration = image_size[0]*image_size[1]
self.deserializer.configure(image_size, filter_size)
self.ifmap_glb.configure(image_size, filter_size, in_sets, fmap_per_iteration)
# self.psum_glb.configure(filter_size, out_sets, fmap_per_iteration)
self.filter_noc.configure(in_sets, self.arr_x)
self.ifmap_noc.configure(in_sets)
self.bias_noc.configure(self.post_tr_x, self.post_tr_y)
self.post_tr_wr_noc.configure(self.post_tr_x, self.post_tr_y)
self.post_tr_rd_noc.configure()
self.psum_rd_noc.configure(self.arr_x)
#self.psum_wr_noc.configure(num_iteration, fmap_per_iteration, out_sets)
for y in range(self.arr_y):
for x in range(self.arr_x):
self.pe_array[y][x].configure(fmap_per_iteration, num_iteration)
for y in range(self.post_tr_y):
for x in range(self.post_tr_x):
self.post_tr_array[y][x].configure()
print("Num PEs: ",self.arr_x*self.arr_y)
print("Num post transform blocks: ", self.post_tr_x*self.post_tr_y)
print("image size: ",image_size)
print("filter size: ",filter_size)
print("input channels: ",in_chn)
print("output channels: ",out_chn)
def instantiate(self, setup):
self.class_name = 'nnsimSmartBuffer'
self.debug = setup['debug']
# ----------------------------------------------------------------
# input channels
# ----------------------------------------------------------------
self.fill_data_ichns = setup['fill_data_ichns']
self.update_data_ichns = setup['update_data_ichns']\
if 'update_data_ichns' in setup\
else None
self.drain_enable_ichns = setup['drain_enable_ichns']\
if 'drain_enable_ichns' in setup\
else None
# ----------------------------------------------------------------
# output channels
# ----------------------------------------------------------------
self.drain_data_ochns = setup['drain_data_ochns']
# ----------------------------------------------------------------
# subcomponents
# ----------------------------------------------------------------
self.logical_managers = ModuleL()
self.memory = SRAM(setup['SRAM'])
self.memory_width = setup['SRAM']['width']
# ----------------------------------------------------------------
# internal channels
# ----------------------------------------------------------------
self.ack_internal_chns = ModuleL()
self.lm_fill_addr_internal_chns = ModuleL()
self.lm_drain_addr_internal_chns = ModuleL()
self.lm_update_addr_internal_chns = ModuleL()
# ----------------------------------------------------------------
# setup logical managers
# ----------------------------------------------------------------
self.num_logical_managers = setup['num_logical_managers']
for logical_unit_idx in range(self.num_logical_managers):
lm_setup = {
'debug': self.debug + ' (LM:' + str(logical_unit_idx) + ')'
}
lm_setup['memory_width'] = self.memory_width
lm_setup.update(setup['LMs'][logical_unit_idx].copy())
# ------ >> ack channel
self.ack_internal_chns.append(Ch())
lm_setup['ack_ichn'] = self.ack_internal_chns[logical_unit_idx]
# ------ >> fill address channel
self.lm_fill_addr_internal_chns.append(Ch())
lm_setup['fill_addr_ochn'] = self.lm_fill_addr_internal_chns[
logical_unit_idx]
# ------ >> drain address channel
self.lm_drain_addr_internal_chns.append(Ch())
lm_setup['drain_addr_ochn'] = self.lm_drain_addr_internal_chns[
logical_unit_idx]
# ------ >> update address channel
if 'update' in setup['LMs'][logical_unit_idx]['AGs']:
self.lm_update_addr_internal_chns.append(Ch())
lm_setup[
'update_addr_ochn'] = self.lm_update_addr_internal_chns[
logical_unit_idx]
else:
self.lm_update_addr_internal_chns.append(None)
# ------ >> instantiate logical manager
self.logical_managers.append(LM(lm_setup))
# ------ >> assign unique class name for each logical manager
self.logical_managers[logical_unit_idx].class_name = \
'LM'+ str(logical_unit_idx) + '_' + self.logical_managers[logical_unit_idx].class_name
# ----------------------------------------------------------------
# internal records containers
# ----------------------------------------------------------------
# >> record for processed reads
# used for retrieving read data information
self.last_read = ModuleL()
for sram_port_idx in range(self.memory.nports):
self.last_read.append(NLCh(depth=1))
# used for detecting drains that can be forwarded
self.approved_write_data_in_cycle = []
for lm_idx in range(self.num_logical_managers):
self.approved_write_data_in_cycle.append({
'fill': None,
'update': None
})
# -----------------------------------------------------------------
# Static and Runtime information
# -----------------------------------------------------------------
# definitin of compound actions
self.setup_access_info()
attrs_dict = setup['SRAM'].copy()
attrs_dict.pop('port_type')
self.attrs.update(attrs_dict)
self.component_with_action = True
class nnsimSmartBuffer(Module):
def instantiate(self, setup):
self.class_name = 'nnsimSmartBuffer'
self.debug = setup['debug']
# ----------------------------------------------------------------
# input channels
# ----------------------------------------------------------------
self.fill_data_ichns = setup['fill_data_ichns']
self.update_data_ichns = setup['update_data_ichns']\
if 'update_data_ichns' in setup\
else None
self.drain_enable_ichns = setup['drain_enable_ichns']\
if 'drain_enable_ichns' in setup\
else None
# ----------------------------------------------------------------
# output channels
# ----------------------------------------------------------------
self.drain_data_ochns = setup['drain_data_ochns']
# ----------------------------------------------------------------
# subcomponents
# ----------------------------------------------------------------
self.logical_managers = ModuleL()
self.memory = SRAM(setup['SRAM'])
self.memory_width = setup['SRAM']['width']
# ----------------------------------------------------------------
# internal channels
# ----------------------------------------------------------------
self.ack_internal_chns = ModuleL()
self.lm_fill_addr_internal_chns = ModuleL()
self.lm_drain_addr_internal_chns = ModuleL()
self.lm_update_addr_internal_chns = ModuleL()
# ----------------------------------------------------------------
# setup logical managers
# ----------------------------------------------------------------
self.num_logical_managers = setup['num_logical_managers']
for logical_unit_idx in range(self.num_logical_managers):
lm_setup = {
'debug': self.debug + ' (LM:' + str(logical_unit_idx) + ')'
}
lm_setup['memory_width'] = self.memory_width
lm_setup.update(setup['LMs'][logical_unit_idx].copy())
# ------ >> ack channel
self.ack_internal_chns.append(Ch())
lm_setup['ack_ichn'] = self.ack_internal_chns[logical_unit_idx]
# ------ >> fill address channel
self.lm_fill_addr_internal_chns.append(Ch())
lm_setup['fill_addr_ochn'] = self.lm_fill_addr_internal_chns[
logical_unit_idx]
# ------ >> drain address channel
self.lm_drain_addr_internal_chns.append(Ch())
lm_setup['drain_addr_ochn'] = self.lm_drain_addr_internal_chns[
logical_unit_idx]
# ------ >> update address channel
if 'update' in setup['LMs'][logical_unit_idx]['AGs']:
self.lm_update_addr_internal_chns.append(Ch())
lm_setup[
'update_addr_ochn'] = self.lm_update_addr_internal_chns[
logical_unit_idx]
else:
self.lm_update_addr_internal_chns.append(None)
# ------ >> instantiate logical manager
self.logical_managers.append(LM(lm_setup))
# ------ >> assign unique class name for each logical manager
self.logical_managers[logical_unit_idx].class_name = \
'LM'+ str(logical_unit_idx) + '_' + self.logical_managers[logical_unit_idx].class_name
# ----------------------------------------------------------------
# internal records containers
# ----------------------------------------------------------------
# >> record for processed reads
# used for retrieving read data information
self.last_read = ModuleL()
for sram_port_idx in range(self.memory.nports):
self.last_read.append(NLCh(depth=1))
# used for detecting drains that can be forwarded
self.approved_write_data_in_cycle = []
for lm_idx in range(self.num_logical_managers):
self.approved_write_data_in_cycle.append({
'fill': None,
'update': None
})
# -----------------------------------------------------------------
# Static and Runtime information
# -----------------------------------------------------------------
# definitin of compound actions
self.setup_access_info()
attrs_dict = setup['SRAM'].copy()
attrs_dict.pop('port_type')
self.attrs.update(attrs_dict)
self.component_with_action = True
def configure(self, config):
for logical_unit_idx in range(self.num_logical_managers):
self.logical_managers[logical_unit_idx].configure(
config['LM'][logical_unit_idx])
# ----------------------------------------------------------------
# internal records configuration
# -------------- --------------------------------------------------
# >> record for logical manager and sram ports correspondance
self.lm_sram_map = config['lm_sram_map']
def tick(self):
self.ack_packet = []
for i in range(self.num_logical_managers):
self.ack_packet.append([])
# ------------------------------------------------------------------
# Check for unprocessed reads
# ------------------------------------------------------------------
for sram_port_idx in range(self.memory.nports):
if self.last_read[sram_port_idx].valid():
read_request_info = self.last_read[sram_port_idx].peek()
# if 'IFmapGLB' in self.debug:
# print(self.debug, read_request_info, 'ochn vacancy:', self.drain_data_ochns[read_request_info['lm_idx']].vacancy())
if self.drain_data_ochns[
read_request_info['lm_idx']].vacancy():
if read_request_info['enabled']:
if not read_request_info['forwarded']:
read_data = [
d for d in self.memory.response(
port=sram_port_idx)
]
# if 'PsumGLB' in self.debug:
# print('DDDDD ' ,self.debug, 'read from memory:', read_data, 'from ', read_request_info['addr'])
else:
read_data = read_request_info['forwarded_data']
# if 'PsumGLB' in self.debug:
# print('DDDDD ' ,self.debug, 'forwared read:', read_data, 'from ', read_request_info['addr'])
else:
read_data = [0] * self.memory.width
# if 'PsumGLB' in self.debug:
# print('DDDDD ' ,self.debug, 'gated data:', read_data, 'from ', read_request_info['addr'])
self.last_read[sram_port_idx].pop()
# if 'PsumGLB' in self.debug:
# print(self.debug, 'port ', sram_port_idx, 'pop')
self.drain_data_ochns[read_request_info['lm_idx']].push(
read_data)
# ------------------------------------------------------------------
# Check for ready requests
# ------------------------------------------------------------------
for lm_idx in range(self.num_logical_managers):
self.check_for_update_request(lm_idx)
self.check_for_fill_request(lm_idx)
self.check_for_drain_request(lm_idx)
if self.logical_managers[lm_idx].fill_round_done.rd()\
and self.logical_managers[lm_idx].drain_round_done.rd():
self.logical_managers[lm_idx].reset_book_keeping_pointers()
# print('round done')
def check_for_drain_request(self, lm_idx):
# if not self.lm_drain_addr_internal_chns[lm_idx].valid():
# if 'IFmapGLB' in self.debug:
# print('!!!!!', self.debug,' no drain request')
# else:
# if 'IFmapGLB' in self.debug:
# print('!!!!!', self.debug,'wait to process request',self.lm_drain_addr_internal_chns[lm_idx].peek() )
if self.lm_drain_addr_internal_chns[lm_idx].valid():
request_info = self.logical_managers[lm_idx].check_request('drain')
# print(request_info)
# if 'IFmapGLB' in self.debug:
# print(self.debug, request_info)
if request_info['addr'] is not None:
enable_signal_ready = True
enabled = True
enabled_buffer = False
if self.drain_enable_ichns is not None and \
self.drain_enable_ichns[lm_idx] is not None:
enable_signal_ready = self.drain_enable_ichns[
lm_idx].valid()
enabled_buffer = True
if enable_signal_ready:
enabled = self.drain_enable_ichns[lm_idx].peek()[0]
# if 'IFmapGLB' in self.debug:
# print(self.debug, request_info, 'enable signal:', enable_signal_ready)
if enable_signal_ready:
sram_rd_port = self.lm_sram_map[lm_idx]['drain']
if self.last_read[sram_rd_port].vacancy(
) and not self.memory.port_in_use(sram_rd_port):
# if 'PsumGLB' in self.debug:
# print(self.debug, 'there is space and available prot for issuing read request')
if enabled_buffer:
self.drain_enable_ichns[lm_idx].pop()
address = request_info['addr']
self.lm_drain_addr_internal_chns[lm_idx].pop()
if enabled and not request_info['forwarded']:
self.memory.request(RD, address, port=sram_rd_port)
if enabled and request_info['forwarded']:
# print('forwarding data')
if request_info['prereq'] == 'fill':
forwarded_data = self.approved_write_data_in_cycle[
lm_idx]['fill']
if forwarded_data is None:
print(
self.debug,
'there is no fill data available for forwarding'
)
sys.exit(0)
elif request_info['prereq'] == 'update':
forwarded_data = self.approved_write_data_in_cycle[
lm_idx]['update']
if forwarded_data is None:
print(
self.debug,
'there is no update data available for forwarding'
)
sys.exit(0)
else:
print(self.debug,
'nowhere to find data being forwarded')
sys.exit(0)
else:
forwarded_data = None
self.last_read[sram_rd_port].push({'lm_idx': lm_idx,\
'enabled': enabled,\
'addr': address,\
'forwarded': request_info['forwarded'],\
'forwarded_data': forwarded_data})
ack_packet = {'type': 'drain', 'addr': address, 'shrink': request_info['shrink'],\
'prereq': request_info['prereq'], 'forwarded': request_info['forwarded'],\
'reset_phy_head': request_info['reset_phy_head'] }
# if 'IFmapGLB' in self.debug:
# print('########## DRAIN #################', self.debug, request_info)
self.logical_managers[lm_idx].update_book_keeping(
ack_packet)
# -------------------------------------------------
# record access
# -------------------------------------------------
if enabled:
last_addr = self.last_read_addr[lm_idx]
if request_info['forwarded']:
arg_name = 'lm' + str(
lm_idx) + '_nforwarded_drain'
self.cycle_access[arg_name] += 1
elif last_addr is not None and last_addr == address:
arg_name = 'lm' + str(
lm_idx) + '_nrepeated_drain'
self.cycle_access[arg_name] += 1
else:
arg_name = 'lm' + str(lm_idx) + '_ndrain'
self.last_read_addr[lm_idx] = address
self.cycle_access[arg_name] += 1
else:
arg_name = 'lm' + str(lm_idx) + '_ngated_drain'
self.cycle_access[arg_name] += 1
# else:
# if 'PsumGLB' in self.debug:
# print(self.debug, 'rd_port:', sram_rd_port, 'last_read_vacant:', self.last_read[sram_rd_port].vacancy(), \
# 'port_in_use:', self.memory.port_in_use(sram_rd_port))
def check_for_update_request(self, lm_idx):
if self.logical_managers[lm_idx].update_AG is None:
return
# if 'PsumGLB' in self.debug:
# stop = 1
if (self.update_data_ichns[lm_idx].valid() and self.lm_update_addr_internal_chns[lm_idx].valid()) or \
(self.lm_update_addr_internal_chns[lm_idx].valid() and self.lm_update_addr_internal_chns[lm_idx].peek()['addr'] == 'reset_phead' ):
request_info = self.logical_managers[lm_idx].check_request(
'update')
# if 'PsumGLB' in self.debug:
# print(self.debug, request_info)
if request_info['addr'] is not None:
sram_wr_port = self.lm_sram_map[lm_idx]['update']
if not self.memory.port_in_use(sram_wr_port):
address = request_info['addr']
data = self.update_data_ichns[lm_idx].pop()
# if 'PsumGLB' in self.debug:
# print(self.debug, request_info, 'data for update:', data)
self.approved_write_data_in_cycle[lm_idx]['update'] = data
self.lm_update_addr_internal_chns[lm_idx].pop()
self.memory.request(WR, address, data, port=sram_wr_port)
ack_packet = {
'type': 'update',
'addr': address,
'reset_phy_head': request_info['reset_phy_head']
}
self.logical_managers[lm_idx].update_book_keeping(
ack_packet)
# ---------------------------------------
# record access
# ---------------------------------------
if not data == self.last_write_data[lm_idx]:
arg_name = 'lm' + str(lm_idx) + '_nupdate'
else:
arg_name = 'lm' + str(
lm_idx) + '_nrepeated_data_update'
self.cycle_access[arg_name] += 1
# self.curr_write_addr[lm_idx] = address
self.last_write_data[lm_idx] = data
if self.traces_stats == 'show':
self.recorder.record(self.debug + '_update.txt',
data[0])
else:
# if the write port is not avaiable, overwrite the logical manager's information
self.logical_managers[lm_idx].approved_write_addr_in_cycle[
'update'] = None
# else:
# if 'PsumGLB' in self.debug:
# print('update_data:', self.update_data_ichns[lm_idx].valid())
def check_for_fill_request(self, lm_idx):
if (self.fill_data_ichns[lm_idx].valid() and self.lm_fill_addr_internal_chns[lm_idx].valid()) or \
(self.lm_fill_addr_internal_chns[lm_idx].valid() and self.lm_fill_addr_internal_chns[lm_idx].peek()['addr'] == 'reset_phead' ):
request_info = self.logical_managers[lm_idx].check_request('fill')
if request_info['addr'] is not None:
sram_wr_port = self.lm_sram_map[lm_idx]['fill']
if not self.memory.port_in_use(sram_wr_port):
address = request_info['addr']
data = self.fill_data_ichns[lm_idx].pop()
self.approved_write_data_in_cycle[lm_idx]['fill'] = data
self.lm_fill_addr_internal_chns[lm_idx].pop()
# print(self.debug, data)
self.memory.request(WR, address, data, port=sram_wr_port)
ack_packet = {
'type': 'fill',
'addr': address,
'reset_phy_head': request_info['reset_phy_head']
}
self.logical_managers[lm_idx].update_book_keeping(
ack_packet)
# if 'IFmapGLB' in self.debug:
# print(self.debug, request_info)
# ---------------------------------------
# record access
# ---------------------------------------
if not data == self.last_write_data[lm_idx]:
arg_name = 'lm' + str(lm_idx) + '_nfill'
else:
arg_name = 'lm' + str(lm_idx) + '_nrepeated_data_fill'
self.cycle_access[arg_name] += 1
# self.curr_write_addr[lm_idx] = address
self.last_write_data[lm_idx] = data
# -------------------------------------------
# record filled data
# -------------------------------------------
if self.traces_stats == 'show':
self.recorder.record(self.debug + '_fill.txt', data[0])
else:
# if the write port is not avaiable, overwrite the logical manager's information
self.logical_managers[lm_idx].approved_write_addr_in_cycle[
'fill'] = None
def setup_access_info(self):
# =====================================================================
# construct action description of the compound action: buffer access
# =====================================================================
self.customized_access = True
self.arg_lst = [] # top level action attributes
total_nread = [] # all drain related counts: aggregated for memory
total_nwrite = [] # all fill related counts: aggregated for memory
total_nrepeated_read = [
] # all repeated drain related counts: aggregated for memory
total_nrepeated_data_write = [] # all repeated data write counts
subcomp_class_actions = {} # all related subcomponent class actions
# ---------------------------------------------------------------------
# Collect related information from each logical manager
# ---------------------------------------------------------------------
for lm_idx in range(self.num_logical_managers):
lm_ndrain = 'lm' + str(lm_idx) + '_ndrain'
lm_nfill = 'lm' + str(lm_idx) + '_nfill'
lm_nupdate = 'lm' + str(lm_idx) + '_nupdate'
lm_nrepeated_data_fill = 'lm' + str(
lm_idx) + '_nrepeated_data_fill'
lm_nrepeated_data_update = 'lm' + str(
lm_idx) + '_nrepeated_data_update'
lm_nrepeated_drain = 'lm' + str(lm_idx) + '_nrepeated_drain'
lm_nforwarded_drain = 'lm' + str(lm_idx) + '_nforwarded_drain'
lm_ngated_drain = 'lm' + str(lm_idx) + '_ngated_drain'
self.arg_lst.append(lm_ndrain)
self.arg_lst.append(lm_nfill)
self.arg_lst.append(lm_nupdate)
self.arg_lst.append(lm_nrepeated_data_fill)
self.arg_lst.append(lm_nrepeated_data_update)
self.arg_lst.append(lm_nrepeated_drain)
self.arg_lst.append(lm_nforwarded_drain)
self.arg_lst.append(lm_ngated_drain)
# sum up the access that will point into RAM
# gated and forwarded drain does not invovle actual RAM accesses
total_nread.append(lm_ndrain)
total_nwrite.append(lm_nfill)
total_nwrite.append(lm_nupdate)
total_nrepeated_read.append(lm_nrepeated_drain)
total_nrepeated_data_write.append(lm_nrepeated_data_fill)
total_nrepeated_data_write.append(lm_nrepeated_data_update)
# ---------------------------------------------------------------------------
# Collect related information for the address generators, embedded in LMs
# ---------------------------------------------------------------------------
subcomp_class_actions.update({self.logical_managers[lm_idx].class_name:\
{'action_name': 'generate',\
'arguments': [lm_ndrain, lm_ngated_drain, lm_nfill, lm_nupdate, lm_nrepeated_data_update, lm_nrepeated_data_fill ],\
'repeat': {'sum':[lm_ndrain, lm_ngated_drain, lm_nfill, lm_nupdate, lm_nrepeated_data_update, lm_nrepeated_data_fill]}}\
})
# ---------------------------------------------------------------------
# Collect related information for the channel subcomponent (move to top level/serializer/deserializer)
# ---------------------------------------------------------------------
self.fill_data_ichns[lm_idx].base_class_name = 'channel'
self.fill_data_ichns[lm_idx].component_class_as_subclass = 'show'
self.drain_data_ochns[lm_idx].base_class_name = 'channel'
self.drain_data_ochns[lm_idx].component_class_as_subclass = 'show'
self.fill_data_ichns[lm_idx].class_name = 'lm_' + str(
lm_idx) + '_fill_chn'
self.drain_data_ochns[lm_idx].class_name = 'lm_' + str(
lm_idx) + '_drain_chn'
subcomp_class_actions.update({self.fill_data_ichns[lm_idx].class_name: {'action_name': 'access',\
'repeat': lm_nfill}})
subcomp_class_actions.update({self.drain_data_ochns[lm_idx].class_name: {'action_name': 'access',\
'repeat': {'sum':[lm_ndrain, lm_nrepeated_drain]}}})
if self.update_data_ichns is not None and self.update_data_ichns[
lm_idx] is not None:
self.update_data_ichns[lm_idx].base_class_name = 'channel'
self.update_data_ichns[
lm_idx].component_class_as_subclass = 'show'
self.update_data_ichns[lm_idx].class_name = 'lm_' + str(
lm_idx) + '_update_chn'
subcomp_class_actions.update({self.update_data_ichns[lm_idx].class_name: {'action_name': 'access',\
'repeat': lm_nupdate}})
# ---------------------------------------------------------------------
# Collect related information for the memory subcomponent
# ---------------------------------------------------------------------
subcomp_class_actions.update({self.memory.class_name: {'action_name': 'RAM_access',\
'arguments': [{'sum': total_nread}, \
{'sum': total_nwrite}, \
{'sum': total_nrepeated_read},\
{'sum': total_nrepeated_data_write}],\
'repeat': 1}})
# ---------------------------------------------------------------------
# Define all of the actions using collected info
# ---------------------------------------------------------------------
smartbuffer_access_action_def = {'arguments': self.arg_lst,\
'subcomponent_class_actions': subcomp_class_actions}
idle_access_action_def = {'subcomponent_class_actions': \
{self.memory.class_name: {'action_name': 'idle','repeat': 1}}}
self.actions = {'idle': idle_access_action_def,\
'buffer_access': smartbuffer_access_action_def}
# =====================================================================
# construct containers for recording access counts
# =====================================================================
self.last_read_addr = [None] * self.num_logical_managers
self.last_write_addr = [None] * self.num_logical_managers
self.last_write_data = [None] * self.num_logical_managers
self.curr_write_addr = [None] * self.num_logical_managers
self.access_stats = {'buffer_access': [], 'idle': {'count': 0}}
self.cycle_access = {}
self.raw_access_stats = {'buffer_access': {}}
self.reset_cycle_access()
def reset_cycle_access(self):
for arg in self.arg_lst:
self.cycle_access.update({arg: 0})
def __ntick__(self):
Module.__ntick__(self)
idle = True
for arg, value in self.cycle_access.items():
if value > 0:
idle = False
break
if idle:
self.access_stats['idle']['count'] += 1
else:
cycle_access_lst = []
for arg in self.arg_lst:
cycle_access_lst.append(self.cycle_access[arg])
cycle_access_tuple = tuple(cycle_access_lst)
if cycle_access_tuple[6] == 1:
self.access_stats['idle']['count'] += 1
if cycle_access_tuple in self.raw_access_stats['buffer_access']:
self.raw_access_stats['buffer_access'][cycle_access_tuple] += 1
else:
self.raw_access_stats['buffer_access'][cycle_access_tuple] = 1
self.reset_cycle_access()
for lm_idx in range(self.num_logical_managers):
self.last_write_addr[lm_idx] = self.curr_write_addr[lm_idx]
self.logical_managers[lm_idx].approved_write_addr_in_cycle[
'fill'] = None
self.logical_managers[lm_idx].approved_write_addr_in_cycle[
'update'] = None
self.approved_write_data_in_cycle[lm_idx]['fill'] = None
self.approved_write_data_in_cycle[lm_idx]['update'] = None
def summerize_access_stats(self):
for access_info_tuple, count in self.raw_access_stats[
'buffer_access'].items():
arg_dict = {}
arg_idx = 0
for arg_name in self.arg_lst:
arg_dict[arg_name] = access_info_tuple[arg_idx]
arg_idx += 1
access_info_dict = {'arguments': arg_dict, 'count': count}
self.access_stats['buffer_access'].append(access_info_dict.copy())
def instantiate(self, arr_x, arr_y, input_chn, output_chn, chn_per_word,
ifmap_glb_depth, psum_glb_depth):
# PE static configuration (immutable)
self.name = 'chip'
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.stat_type = 'show'
# Instantiate DRAM IO channels
self.input_chn = input_chn
self.output_chn = output_chn
# Instantiate input deserializer and output serializer
self.ifmap_wr_chn = Channel(name="ifmap_wr_chn")
self.psum_wr_chn = Channel(name="psum_wr_chn")
self.weights_wr_chn = Channel(name="weights_wr_chn")
self.deserializer = InputDeserializer(self.input_chn,
self.ifmap_wr_chn,
self.weights_wr_chn,
self.psum_wr_chn, arr_x, arr_y,
chn_per_word)
self.psum_output_chn = Channel(name="psum_output_chn")
self.serializer = OutputSerializer(self.output_chn,
self.psum_output_chn)
# Instantiate GLB and GLB channels
self.ifmap_rd_chn = Channel(3, name='ifmap_rd_chn')
self.ifmap_glb = IFMapGLB(self.ifmap_wr_chn, self.ifmap_rd_chn,
ifmap_glb_depth, chn_per_word)
self.psum_rd_chn = Channel(3, name='psum_rd_chn')
self.psum_noc_wr_chn = Channel(name='psum_noc_wr_chn')
self.psum_glb = PSumGLB(self.psum_wr_chn, self.psum_noc_wr_chn,
self.psum_rd_chn, psum_glb_depth, chn_per_word)
self.weights_rd_chn = Channel(name='weights_rd_chn')
self.weights_glb = WeightsGLB(self.weights_wr_chn, self.weights_rd_chn)
# PE Array and local channel declaration
self.pe_array = ModuleList()
self.pe_ifmap_chns = ModuleList()
self.pe_filter_chns = ModuleList()
self.pe_psum_chns = ModuleList()
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_psum_chns[0].append(
Channel(32, name='pe_psum_chns_{}_{}'.format(x, 0)))
# Actual array instantiation
for y in range(self.arr_y):
self.pe_array.append(ModuleList())
self.pe_ifmap_chns.append(ModuleList())
self.pe_filter_chns.append(ModuleList())
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_ifmap_chns[y].append(
Channel(32, name='pe_ifmap_chns_{}_{}'.format(x, y)))
self.pe_filter_chns[y].append(
Channel(32, name='pe_filter_chns_{}_{}'.format(x, y)))
self.pe_psum_chns[y + 1].append(
Channel(32, name='pe_psum_chns_{}_{}'.format(x, y)))
self.pe_array[y].append(
PE(x, y, self.pe_ifmap_chns[y][x],
self.pe_filter_chns[y][x], self.pe_psum_chns[y][x],
self.pe_psum_chns[y + 1][x]))
# Setup NoC to deliver weights, ifmaps and psums
self.filter_noc = WeightsNoC(self.weights_rd_chn, self.pe_filter_chns,
self.chn_per_word)
self.ifmap_noc = IFMapNoC(self.ifmap_rd_chn, self.pe_ifmap_chns,
self.arr_x, self.chn_per_word)
self.psum_rd_noc = PSumRdNoC(self.psum_rd_chn, self.pe_psum_chns[0],
self.chn_per_word)
self.psum_wr_noc = PSumWrNoC(self.pe_psum_chns[-1],
self.psum_noc_wr_chn,
self.psum_output_chn, self.chn_per_word)
class WSArch(Module):
def instantiate(self, arr_x, arr_y, input_chn, output_chn, chn_per_word,
ifmap_glb_depth, psum_glb_depth, weight_glb_depth):
# PE static configuration (immutable)
self.name = 'chip'
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.stat_type = 'show'
# Instantiate DRAM IO channels
self.input_chn = input_chn
self.output_chn = output_chn
# Instantiate input deserializer and output serializer
self.ifmap_wr_chn = Channel(name='ifmap_wr_chn')
self.psum_wr_chn = Channel(name='psum_wr_chn')
self.weights_wr_chn = Channel(name='weights_wr_chn')
self.psum_chn = Channel(32, name='psum_chn')
self.deserializer = InputDeserializer(self.input_chn,
self.ifmap_wr_chn,
self.weights_wr_chn,
self.psum_wr_chn, self.psum_chn,
arr_x, arr_y, chn_per_word)
self.psum_output_chn = Channel(name='psum_output_chn')
self.serializer = OutputSerializer(self.output_chn,
self.psum_output_chn, self.psum_chn,
self.arr_x, self.arr_y,
self.chn_per_word)
# Instantiate GLB and GLB channels
self.ifmap_rd_chn = Channel(3, name='ifmap_rd_chn')
self.ifmap_glb = IFMapGLB(self.ifmap_wr_chn, self.ifmap_rd_chn,
ifmap_glb_depth, chn_per_word)
self.psum_rd_chn = Channel(3, name='psum_rd_chn')
self.psum_noc_wr_chn = Channel(name='psum_noc_wr_chn')
self.psum_glb = PSumGLB(self.psum_wr_chn, self.psum_noc_wr_chn,
self.psum_rd_chn, psum_glb_depth, chn_per_word)
self.weights_rd_chn = Channel(3, name='weights_rd_chn')
self.weights_glb = WeightsGLB(self.weights_wr_chn, self.weights_rd_chn,
weight_glb_depth, chn_per_word)
# PE Array and local channel declaration
self.pe_array = ModuleList()
self.pe_ifmap_chns = ModuleList()
self.pe_filter_chns = ModuleList()
self.pe_psum_chns = ModuleList()
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_psum_chns[0].append(
Channel(32, name='pe_psum_chns_{}_0'.format(x)))
# Actual array instantiation
for y in range(self.arr_y):
self.pe_array.append(ModuleList())
self.pe_ifmap_chns.append(ModuleList())
self.pe_filter_chns.append(ModuleList())
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_ifmap_chns[y].append(
Channel(32, name='pe_ifmap_chns_{}_{}'.format(x, y)))
self.pe_filter_chns[y].append(
Channel(32, name='pe_filter_chns_{}_{}'.format(x, y)))
self.pe_psum_chns[y + 1].append(
Channel(32, name='pe_psum_chns_{}_{}'.format(x, y)))
self.pe_array[y].append(
PE(x, y, self.pe_ifmap_chns[y][x],
self.pe_filter_chns[y][x], self.pe_psum_chns[y][x],
self.pe_psum_chns[y + 1][x]))
# Setup NoC to deliver weights, ifmaps and psums
self.filter_noc = WeightsNoC(self.weights_rd_chn, self.pe_filter_chns,
self.chn_per_word)
self.ifmap_noc = IFMapNoC(self.ifmap_rd_chn, self.pe_ifmap_chns,
self.arr_x, self.chn_per_word)
self.psum_rd_noc = PSumRdNoC(self.psum_rd_chn, self.pe_psum_chns[0],
self.chn_per_word)
self.psum_wr_noc = PSumWrNoC(self.pe_psum_chns[-1],
self.psum_noc_wr_chn,
self.psum_output_chn, self.chn_per_word)
def configure(self, image_size, filter_size, in_chn, out_chn, full_in_chn,
full_out_chn):
full_in_sets = full_in_chn // self.chn_per_word
full_out_sets = full_out_chn // self.chn_per_word
in_sets = self.arr_y // self.chn_per_word
out_sets = self.arr_x // self.chn_per_word
tile_ins = full_in_sets // in_sets
tile_outs = full_out_sets // out_sets
fmap_per_iteration = image_size[0] * image_size[1]
num_iteration = filter_size[0] * filter_size[1]
self.deserializer.configure(image_size, filter_size, full_in_sets,
tile_ins, tile_outs)
self.serializer.configure(image_size, tile_ins, tile_outs)
self.ifmap_glb.configure(image_size, filter_size, in_sets,
full_in_sets, tile_outs, fmap_per_iteration)
self.psum_glb.configure(filter_size, out_sets, fmap_per_iteration)
self.weights_glb.configure(filter_size, in_sets, out_sets)
self.filter_noc.configure(in_sets, self.arr_x)
self.ifmap_noc.configure(in_sets)
self.psum_rd_noc.configure(out_sets)
self.psum_wr_noc.configure(num_iteration, fmap_per_iteration, out_sets)
for y in range(self.arr_y):
for x in range(self.arr_x):
self.pe_array[y][x].configure(fmap_per_iteration,
num_iteration)
def instantiate (self, setup):
self.class_name = "nnsimBuffer"
self.debug = setup['debug']
self.enabled_buffer = setup['enabled_buffer'] if 'enabled_buffer' in setup.keys() else False
self.reg_insert_opt = setup['reg_insert_opt'] if 'reg_insert_opt' in setup.keys() else True
self.nbanks = setup['nbanks']
# physical channels from outside world to write data into the buffer
self.physical_chns = []
# address generators for each memory bank
self.addr_generator_type = []
#======================================================================
# Virtual Data Transfer Channels
#======================================================================
# - instantiated from outside
# - responsible for receiving data from outside
# - responsible for pushing data to outside
self.fill_data_chns = setup['fill_chns']\
if 'fill_chns' in setup.keys() else None
self.update_data_chns = setup['update_chns']\
if 'update_chns' in setup.keys() else None
self.update_src_chns = setup['update_src_chns']\
if 'update_src_chns' in setup.keys() else None
self.drain_rsp_chns = setup['drain_chns']\
if 'drain_chns' in setup.keys() else None
# channels determine whether to perform the read or not, used for zero gating
self.drain_enable_chns = setup['drain_enable_chns']\
if 'drain_enable_chns' in setup.keys() else None
# number of virtual ports for each memory bank
self.nvports = setup['nvports']
# constrcut dictionary to store all the data channels for ease of access
self.vdata_chns = {}
self.vdata_chns['fill'] = self.fill_data_chns if self.fill_data_chns is not None else None
self.vdata_chns['update'] = self.update_data_chns if self.update_data_chns is not None else None
#======================================================================
# Address Generators
#======================================================================
self.addr_generator_types = setup['addr_generators']
#======================================================================
# Hardware Memory Properties
#======================================================================
self.depth = setup['depth'] # depth of each memory bank
self.width = setup['width'] # # of data in each memory bank
self.nports = setup['nports'] # number of physical ports for each memory bank
self.pvmapping = setup['pvmapping']
#======================================================================
# Internal Channels and Data Structures
#======================================================================
# addr generators
self.addr_generators = []
self.faddr_generator = ModuleList()
self.uaddr_generator = ModuleList()
self.daddr_generator = ModuleList()
# addr chns
self.fill_addr_chn = ModuleList()
self.update_addr_chn = ModuleList()
self.drain_addr_chn = ModuleList()
self.fill_local_addr_chn = ModuleList()
self.update_local_addr_chn = ModuleList()
self.drain_local_addr_chn = ModuleList()
#------------------------------------------------------------------
# Address Generators
#------------------------------------------------------------------
self.addr_generators = {'fill': self.faddr_generator, \
'update': self.uaddr_generator, \
'drain': self.daddr_generator}
#------------------------------------------------------------------
# Address Channels for Addr Generators to Push to Sbuffer
#------------------------------------------------------------------
self.vaddr_chns = {'fill': self.fill_addr_chn,\
'update': self.update_addr_chn,\
'drain': self.drain_addr_chn}
self.local_vaddr_chns = {'fill': self.fill_local_addr_chn,\
'update': self.update_local_addr_chn,\
'drain': self.drain_local_addr_chn}
#------------------------------------------------------------------
# Setup Virtual Ports
#------------------------------------------------------------------
# each virtual port has:
# a channel for addr seq
# a channel for input data
# an addr generator
# optionally a destination generator
# optionally a channel for output data
for vport, num in self.nvports.items():
for i in range(num):
# ---------------------------------------------------------
# Address Generator
# ---------------------------------------------------------
# 1. setup single address generator for the virtual port
if setup['multi_addr_generators'][vport] == 1:
self.vaddr_chns[vport].append(NoLatencyChannel())
self.local_vaddr_chns[vport].append(None)
addr_gen_setup = {'addr_chn': self.vaddr_chns[vport][i],\
'width': self.width, \
'type': vport,\
'id': i,
'depth': self.depth,\
'debug': self.debug }
addr_gen_obj = self.addr_generator_types[vport][i](addr_gen_setup)
self.addr_generators[vport].append(addr_gen_obj)
# 2. update address comes from multiple instances of udpate address generators
else:
self.addr_generators[vport].append(ModuleList())
# this channel pushes the chosen address from all the local address generators to the smartbuffer
self.vaddr_chns[vport].append(NoLatencyChannel())
self.local_vaddr_chns[vport].append(ModuleList())
for idx in range(setup['multi_addr_generators'][vport]):
self.local_vaddr_chns[vport][i].append(NoLatencyChannel())
addr_gen_setup = {'addr_chn': self.local_vaddr_chns[vport][i][idx],\
'width': self.width, \
'type': vport,\
'id': i,
'depth': self.depth,\
'debug': self.debug + '[' + str(idx) + ']'}
# assuming these generators are of the same type, just the offset (config) will be different
self.addr_generators[vport][i].append(self.addr_generator_types[vport][i](addr_gen_setup))
#------------------------------------------------------------------
# Setup Smartbuffer (deals with dependency and confict)
#------------------------------------------------------------------
# destination channels are for the destination calculator to push in destinations
# destination_chns = self.destination_chns if self.mode == 'glb' else None
# update_src_chns are channels for sending in external update address sequence
update_src_chn = self.update_src_chns if self.update_src_chns is not None else None
# drain enable channels are channels that are used to send enable signeal to the buffer for read operations
drain_enable_chn = self.drain_enable_chns if self.drain_enable_chns is not None else None
# a smartbuffer is a unit of memory bank
smartbuffer_setup = { 'depth': self.depth,\
'width': self.width, \
'nports': self.nports, \
'pvmapping': self.pvmapping,\
'nvports': self.nvports,\
'vdata_chns': self.vdata_chns, \
'vaddr_chns': self.vaddr_chns,\
'drain_rsp_chns': self.drain_rsp_chns,\
'drain_enable_chns': drain_enable_chn,\
'update_src_chns': update_src_chn,\
'enabled_buffer': self.enabled_buffer,\
'reg_insert_opt': self.reg_insert_opt,\
'debug': self.debug }
self.smartbuffer_type = setup['smartbuffer_type']
self.sbuffer = self.smartbuffer_type(smartbuffer_setup)
#======================================================================
# Stats Collection Info
#======================================================================
self.attrs = { 'dpeth': setup['depth'],\
'width': setup['width'],\
'data_width': setup['data_width'],\
'nbanks': setup['nbanks'],\
'nports': setup['nports']}
class OSArch(Module):
def instantiate(self, arr_y,
input_chn, output_chn,
block_size, num_nonzero,
ifmap_glb_depth, psum_glb_depth, weight_glb_depth):
# PE static configuration (immutable)
self.name = 'chip'
#self.arr_x = arr_x
self.arr_y = arr_y
self.block_size = block_size
self.num_nonzero = num_nonzero
self.stat_type = 'show'
# Instantiate DRAM IO channels
self.input_chn = input_chn
self.output_chn = output_chn
# Instantiate input deserializer and output serializer
self.ifmap_wr_chn = Channel()
self.psum_wr_chn = Channel()
self.weights_wr_chn = Channel()
self.deserializer = InputDeserializer(self.input_chn, self.ifmap_wr_chn,
self.weights_wr_chn, self.psum_wr_chn, arr_y,
block_size, num_nonzero)
self.psum_output_chn = Channel()
self.serializer = OutputSerializer(self.output_chn, self.psum_output_chn)
# Instantiate GLB and GLB channels
self.ifmap_rd_chn = Channel(3)
#self.ifmap_glb = IFMapGLB(self.ifmap_wr_chn, self.ifmap_rd_chn, arr_y,
# ifmap_glb_depth, block_size, num_nonzero)
self.psum_rd_chn = Channel(3)
self.psum_noc_wr_chn = Channel()
self.psum_glb = PSumGLB(self.psum_wr_chn, self.psum_noc_wr_chn, self.psum_rd_chn,
psum_glb_depth, block_size, num_nonzero)
self.weights_rd_chn = Channel()
#self.weights_glb = WeightsGLB(self.weights_wr_chn, self.weights_rd_chn, weight_glb_depth, block_size)
self.ifmap_weights_glb = IFMapWeightsGLB(self.ifmap_wr_chn, self.ifmap_rd_chn,\
self.weights_wr_chn, self.weights_rd_chn, arr_y, ifmap_glb_depth,\
weight_glb_depth, block_size, num_nonzero)
# PE Array and local channel declaration
self.pe_array = ModuleList()
self.pe_ifmap_chns = ModuleList()
self.pe_filter_chns = ModuleList()
self.pe_psum_in_chns = ModuleList()
self.pe_psum_out_chns = ModuleList()
# Actual array instantiation
for y in range(self.arr_y):
self.pe_array.append(ModuleList())
self.pe_ifmap_chns.append(ModuleList())
self.pe_filter_chns.append(ModuleList())
self.pe_psum_in_chns.append(ModuleList())
self.pe_psum_out_chns.append(ModuleList())
for x in range(1):
self.pe_ifmap_chns[y].append(Channel(32))
self.pe_filter_chns[y].append(Channel(32))
self.pe_psum_in_chns[y].append(Channel(32))
self.pe_psum_out_chns[y].append(Channel(32))
self.pe_array[y].append(
PE(x, y,
self.pe_ifmap_chns[y][x],
self.pe_filter_chns[y][x],
self.pe_psum_in_chns[y][x],
self.pe_psum_out_chns[y][x]
)
)
# Setup NoC to deliver weights, ifmaps and psums
self.filter_noc = WeightsNoC(self.weights_rd_chn, self.pe_filter_chns, block_size)
self.ifmap_noc = IFMapNoC(self.ifmap_rd_chn, self.pe_ifmap_chns)
self.psum_rd_noc = PSumRdNoC(self.psum_rd_chn, self.pe_psum_in_chns, self.arr_y, block_size)
self.psum_wr_noc = PSumWrNoC(self.pe_psum_out_chns, self.psum_noc_wr_chn, self.psum_output_chn, self.arr_y, block_size)
def configure(self, image_size, filter_size, in_chn, out_chn):
in_sets = in_chn//self.block_size
out_sets = out_chn//self.arr_y
fmap_per_iteration = image_size[0]*image_size[1]
num_iteration = filter_size[0]*filter_size[1]
self.deserializer.configure(image_size, filter_size, in_chn, out_chn)
#self.ifmap_glb.configure(image_size, filter_size, in_chn, fmap_per_iteration)
self.psum_glb.configure(filter_size, out_chn, fmap_per_iteration)
#self.weights_glb.configure(filter_size, image_size, in_chn, out_chn)
self.filter_noc.configure(out_chn//self.block_size, self.arr_y, in_chn, out_chn)
self.ifmap_noc.configure(in_sets, self.arr_y)
self.psum_rd_noc.configure(out_chn//self.block_size)
self.psum_wr_noc.configure(fmap_per_iteration, self.arr_y//self.block_size)
self.ifmap_weights_glb.configure(image_size, filter_size, in_chn, out_chn,\
fmap_per_iteration)
for y in range(self.arr_y):
for x in range(1):
self.pe_array[y][x].configure(num_iteration*in_sets, fmap_per_iteration*out_sets)
class WSArch(Module):
def instantiate(self, arr_x, arr_y, input_chn, output_chn, chn_per_word,
ifmap_glb_depth, psum_glb_depth):
# PE static configuration (immutable)
self.name = 'chip'
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.stat_type = 'show'
# Instantiate DRAM IO channels
self.input_chn = input_chn
self.output_chn = output_chn
# Instantiate input deserializer and output serializer
self.ifmap_wr_chn = Channel()
self.psum_wr_chn = Channel()
self.weights_wr_chn = Channel()
self.deserializer = InputDeserializer(self.input_chn,
self.ifmap_wr_chn,
self.weights_wr_chn,
self.psum_wr_chn, arr_x, arr_y,
chn_per_word)
self.psum_output_chn = Channel()
self.serializer = OutputSerializer(self.output_chn,
self.psum_output_chn)
# Instantiate GLB and GLB channels
self.ifmap_rd_chn = Channel(3)
#self.hold_weights = Channel(9)
self.ifmap_glb = IFMapGLB(self.ifmap_wr_chn, self.ifmap_rd_chn,
ifmap_glb_depth,
chn_per_word) #, self.hold_weights)
self.psum_rd_chn = Channel(3)
self.psum_noc_wr_chn = Channel()
self.psum_glb = PSumGLB(self.psum_wr_chn, self.psum_noc_wr_chn,
self.psum_rd_chn, psum_glb_depth, chn_per_word)
self.weights_rd_chn = Channel()
self.weights_glb = WeightsGLB(
self.weights_wr_chn, self.weights_rd_chn) #, self.hold_weights)
# PE Array and local channel declaration
self.pe_array = ModuleList()
self.pe_ifmap_chns = ModuleList()
self.pe_filter_chns = ModuleList()
self.pe_psum_chns = ModuleList()
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_psum_chns[0].append(Channel(32))
# Actual array instantiation
for y in range(self.arr_y):
self.pe_array.append(ModuleList())
self.pe_ifmap_chns.append(ModuleList())
self.pe_filter_chns.append(ModuleList())
self.pe_psum_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_ifmap_chns[y].append(Channel(32))
self.pe_filter_chns[y].append(Channel(32))
self.pe_psum_chns[y + 1].append(Channel(32))
self.pe_array[y].append(
PE(x, y, self.pe_ifmap_chns[y][x],
self.pe_filter_chns[y][x], self.pe_psum_chns[y][x],
self.pe_psum_chns[y + 1][x]))
# Setup NoC to deliver weights, ifmaps and psums
self.filter_noc = WeightsNoC(self.weights_rd_chn, self.pe_filter_chns,
self.chn_per_word)
self.ifmap_noc = IFMapNoC(self.ifmap_rd_chn, self.pe_ifmap_chns,
self.arr_x, self.chn_per_word)
self.psum_rd_noc = PSumRdNoC(self.psum_rd_chn, self.pe_psum_chns[0],
self.chn_per_word)
self.psum_wr_noc = PSumWrNoC(self.pe_psum_chns[-1],
self.psum_noc_wr_chn,
self.psum_output_chn, self.chn_per_word)
def configure(self, image_size, filter_size, in_chn, out_chn):
# print inputs
print("image size:", image_size)
print("filter size:", filter_size)
print("in chn:", in_chn)
print("out chn:", out_chn)
in_sets = self.arr_y // self.chn_per_word
out_sets = self.arr_x // self.chn_per_word
fmap_per_iteration = image_size[0] * image_size[1]
fmap_per_iteration_in = image_size[0] * image_size[1]
fmap_per_iteration_out = 4 # (image_size[0]-filter_size[0]+1)*(image_size[1]-filter_size[1]+1)
num_iteration = filter_size[0] * filter_size[1]
self.deserializer.configure(image_size)
self.ifmap_glb.configure(image_size, filter_size, in_sets,
fmap_per_iteration_out)
self.psum_glb.configure(filter_size, out_sets, fmap_per_iteration_out)
self.filter_noc.configure(in_sets, self.arr_x)
self.ifmap_noc.configure(in_sets)
self.psum_rd_noc.configure(out_sets)
self.psum_wr_noc.configure(num_iteration, fmap_per_iteration_out,
out_sets)
print("PE array size:", self.arr_y * self.arr_x)
for y in range(self.arr_y):
for x in range(self.arr_x):
self.pe_array[y][x].configure(fmap_per_iteration_out,
num_iteration)
def instantiate(self, setup):
self.pe_array_row = setup['pe_array'][0]
self.pe_array_col = setup['pe_array'][1]
# ---------------------------------------------------------------------
# io data b/w offchip
# ---------------------------------------------------------------------
self.weights_in_chns = setup['io_chns']['weights']
self.ifmap_in_chns = setup['io_chns']['ifmap']
self.psum_in_chns = setup['io_chns']['psum_in']
self.psum_out_chn = setup['io_chns']['psum_out']
# ---------------------------------------------------------------------
# onchip channels
# ---------------------------------------------------------------------
# >> buffer output channels
self.weights_out_chns = ModuleList(Channel())
self.ifmap_out_chns = ModuleList(Channel())
self.ifmap_out_chns = ModuleList(Channel())
self.psum_out_chns = ModuleList(Channel())
# >> update data to buffer
self.psum_update_chns = ModuleList(Channel())
# ===================================================================
# GLBs
# ===================================================================
# ------------------------ WEIGHTS -------------------------------
weights_glb_setup = {'fill_data_ichns': self.weights_in_chns,\
'drain_data_ochns': self.weights_out_chns, \
'num_logical_managers': 1,\
'SRAM': {\
'depth': setup['depth']['WeightsGLB'],\
'width': setup['width']['WeightsGLB'],\
'data_width': setup['data_width']['WeightsGLB'],\
'nbanks': setup['nbanks']['WeightsGLB'],\
'nports': setup['nports']['WeightsGLB'],\
'port_type': setup['port_type']['WeightsGLB'],\
},\
'debug': ' WeightsGLB'}
self.weights_glb = WeightsGLB(weights_glb_setup)
# ------------------------ IFMAP -------------------------------
ifmap_glb_setup = {'fill_data_ichns': self.ifmap_in_chns,\
'drain_data_ochns': self.ifmap_out_chns, \
'num_logical_managers': 1,\
'SRAM': {\
'depth': setup['depth']['IFmapGLB'],\
'width': setup['width']['IFmapGLB'],\
'data_width': setup['data_width']['IFmapGLB'],\
'nbanks': setup['nbanks']['IFmapGLB'],\
'nports': setup['nports']['IFmapGLB'],\
'port_type': setup['port_type']['IFmapGLB'],\
},\
'debug': ' IFmapGLB'}
self.ifmap_glb = IFmapGLB(ifmap_glb_setup)
# ------------------------ PSUM -------------------------------
psum_glb_setup = {'fill_data_ichns': self.psum_in_chns,\
'update_data_ichns': self.psum_update_chns,\
'drain_data_ochns': self.psum_out_chns, \
'num_logical_managers': 1,\
'SRAM': {\
'depth': setup['depth']['PsumGLB'],\
'width': setup['width']['PsumGLB'],\
'data_width': setup['data_width']['PsumGLB'],\
'nbanks': setup['nbanks']['PsumGLB'],\
'nports': setup['nports']['PsumGLB'],\
'port_type': setup['port_type']['PsumGLB'],\
},\
'debug': ' PsumGLB'}
self.psum_glb = PsumGLB(psum_glb_setup)
# ===================================================================
# PE Array Channels
# ===================================================================
self.ifmap_pe_data_chns = ModuleList()
self.weights_pe_data_chns = ModuleList()
self.psum_data_chns = ModuleList()
self.pe_data_chns = {'ifmap': self.ifmap_pe_data_chns,\
'weights': self.weights_pe_data_chns,\
'psum': self.psum_data_chns}
for pe_row in range(self.pe_array_row):
for chn_type, chn_row in self.pe_data_chns.items():
chn_row.append(ModuleList())
for pe_col in range(self.pe_array_col):
for chn_type, chn_col in self.pe_data_chns.items():
chn_col[pe_row].append(Channel())
self.psum_data_chns.append(ModuleList())
for pe_col in range(self.pe_array_col):
self.psum_data_chns[-1].append(Channel())
# ===================================================================
# Destination Calculators for NoC
# ===================================================================
self.ifmap_NoC_destination_chn = ModuleList(Channel())
ifmap_NoC_destination_calculator_setup = \
{'out_chn': self.ifmap_NoC_destination_chn[0],\
'out_channel_width': 1,\
'debug': 'IFmapNoCDestCalc'}
self.ifmap_NoC_destination_calculator = IFmapNoCDestCalc(
ifmap_NoC_destination_calculator_setup)
self.weights_NoC_destination_chn = ModuleList(Channel())
weights_NoC_desitnation_calculator_setup = \
{'out_chn': self.weights_NoC_destination_chn[0],\
'out_channel_width': 1,\
'debug': 'WeightsNoCDestCalc'}
self.weights_NoC_destionation_calculator = WeightsNoCDestCalc(
weights_NoC_desitnation_calculator_setup)
self.psum_in_NoC_destination_chn = ModuleList(Channel())
psum_in_NoC_desitnation_calculator_setup = \
{'out_chn': self.psum_in_NoC_destination_chn[0],\
'out_channel_width': 1,\
'debug': 'PsumInNoCDestCalc'}
self.psum_in_NoC_destionation_calculator = PsumInNoCDestCalc(
psum_in_NoC_desitnation_calculator_setup)
# ===================================================================
# NoCs
# ===================================================================
# weights serializer 4:1
self.weights_serializered_data_chn = Channel()
weights_serializer_setup = {'in_chn': self.weights_out_chns[0],\
'out_chn': self.weights_serializered_data_chn,\
'ratio': setup['weights_seri_ratio'],\
'debug': 'weights_serialzer'}
self.weights_serializer = WeightsSerializer(weights_serializer_setup)
# Weights NoC: weightGLB -> PEs
weights_noc_setup = {'rd_chns': self.pe_data_chns['weights'],\
'wr_chns': self.weights_serializered_data_chn,\
'dest_chns': self.weights_NoC_destination_chn,\
'debug': 'WeightsNoC'}
self.weightsNoC = WeightsNoC(weights_noc_setup)
# -------------- IFmap NoC: IfmapGLB -> PEs -----------------------
# ifmap serializer 4:1
self.ifmap_serialized_data_chn = Channel()
ifmap_serializer_setup = {'in_chn': self.ifmap_out_chns[0],\
'out_chn': self.ifmap_serialized_data_chn,\
'ratio': setup['ifmap_seri_ratio'],\
'debug': 'ifmap_serializer'}
self.ifmap_serilizer = IfmapSerializer(ifmap_serializer_setup)
# ifmap NoC
ifmap_noc_wr_chns = ModuleList()
ifmap_noc_wr_chns.append(self.ifmap_serialized_data_chn)
ifmap_noc_setup = {'rd_chns': self.pe_data_chns['ifmap'],\
'wr_chns': ifmap_noc_wr_chns,\
'dest_chns': self.ifmap_NoC_destination_chn,\
'debug': 'IFmapNoC'}
self.ifmapNoC = IFMapNoC(ifmap_noc_setup)
# ---------------------------------------------------------------------
# -------------- PsumRd NoC: Psum GLB -> PEs ----------------------
# psum serializer 4:1
self.psum_serialized_data_chn = Channel()
psum_serializer_setup = {'in_chn': self.psum_out_chns[0],\
'out_chn': self.psum_serialized_data_chn,\
'ratio': setup['psum_seri_ratio'],\
'debug': 'psum_serialzer'}
self.psum_serializer = PsumSerializer(psum_serializer_setup)
# psum read noc
pe_noc_wr_chn = ModuleList(self.psum_serialized_data_chn)
psum_rd_noc_setup = {'rd_chns': self.psum_data_chns,\
'wr_chns': pe_noc_wr_chn,\
'dest_chns': self.psum_in_NoC_destination_chn,\
'debug': 'PsumRdNoC'}
self.psumRdNoC = PsumRdNoC(psum_rd_noc_setup)
# ---------------------------------------------------------------------
# -------------- PsumWrNoC: PEs -> ifmapPsum GLB ---------------------
self.psum_out_noc_rd_chns = ModuleList()
self.psum_out_noc_rd_chns.append(
self.psum_update_chns[0]) # write back to GLB
self.psum_out_noc_rd_chns.append(self.psum_out_chn) # write offchip
psum_wr_noc_setup = {'rd_chns': self.psum_out_noc_rd_chns,\
'wr_chns': self.psum_data_chns,\
'debug': 'PsumWrNoC'}
self.psumWrNoC = PsumWrNoC(psum_wr_noc_setup)
# ===================================================================
# PE Array
# ===================================================================
# general setup of a PE
PE_setup = setup['PE']
self.PE = ModuleList()
for pe_row in range(self.pe_array_row):
self.PE.append(ModuleList())
for pe_col in range(self.pe_array_col):
# PE specific setup
PE_setup['row'] = pe_row
PE_setup['col'] = pe_col
PE_setup['weights_data_in_chn'] = self.weights_pe_data_chns[
pe_row][pe_col]
PE_setup['ifmap_data_in_chn'] = self.ifmap_pe_data_chns[
pe_row][pe_col]
PE_setup['psum_data_in_chn'] = self.psum_data_chns[pe_row][
pe_col]
PE_setup['psum_data_out_chn'] = self.psum_data_chns[pe_row +
1][pe_col]
self.PE[pe_row].append(PE(PE_setup))
class chip(Module):
def instantiate(self, setup):
self.pe_array_row = setup['pe_array'][0]
self.pe_array_col = setup['pe_array'][1]
# ---------------------------------------------------------------------
# io data b/w offchip
# ---------------------------------------------------------------------
self.weights_in_chns = setup['io_chns']['weights']
self.ifmap_in_chns = setup['io_chns']['ifmap']
self.psum_in_chns = setup['io_chns']['psum_in']
self.psum_out_chn = setup['io_chns']['psum_out']
# ---------------------------------------------------------------------
# onchip channels
# ---------------------------------------------------------------------
# >> buffer output channels
self.weights_out_chns = ModuleList(Channel())
self.ifmap_out_chns = ModuleList(Channel())
self.ifmap_out_chns = ModuleList(Channel())
self.psum_out_chns = ModuleList(Channel())
# >> update data to buffer
self.psum_update_chns = ModuleList(Channel())
# ===================================================================
# GLBs
# ===================================================================
# ------------------------ WEIGHTS -------------------------------
weights_glb_setup = {'fill_data_ichns': self.weights_in_chns,\
'drain_data_ochns': self.weights_out_chns, \
'num_logical_managers': 1,\
'SRAM': {\
'depth': setup['depth']['WeightsGLB'],\
'width': setup['width']['WeightsGLB'],\
'data_width': setup['data_width']['WeightsGLB'],\
'nbanks': setup['nbanks']['WeightsGLB'],\
'nports': setup['nports']['WeightsGLB'],\
'port_type': setup['port_type']['WeightsGLB'],\
},\
'debug': ' WeightsGLB'}
self.weights_glb = WeightsGLB(weights_glb_setup)
# ------------------------ IFMAP -------------------------------
ifmap_glb_setup = {'fill_data_ichns': self.ifmap_in_chns,\
'drain_data_ochns': self.ifmap_out_chns, \
'num_logical_managers': 1,\
'SRAM': {\
'depth': setup['depth']['IFmapGLB'],\
'width': setup['width']['IFmapGLB'],\
'data_width': setup['data_width']['IFmapGLB'],\
'nbanks': setup['nbanks']['IFmapGLB'],\
'nports': setup['nports']['IFmapGLB'],\
'port_type': setup['port_type']['IFmapGLB'],\
},\
'debug': ' IFmapGLB'}
self.ifmap_glb = IFmapGLB(ifmap_glb_setup)
# ------------------------ PSUM -------------------------------
psum_glb_setup = {'fill_data_ichns': self.psum_in_chns,\
'update_data_ichns': self.psum_update_chns,\
'drain_data_ochns': self.psum_out_chns, \
'num_logical_managers': 1,\
'SRAM': {\
'depth': setup['depth']['PsumGLB'],\
'width': setup['width']['PsumGLB'],\
'data_width': setup['data_width']['PsumGLB'],\
'nbanks': setup['nbanks']['PsumGLB'],\
'nports': setup['nports']['PsumGLB'],\
'port_type': setup['port_type']['PsumGLB'],\
},\
'debug': ' PsumGLB'}
self.psum_glb = PsumGLB(psum_glb_setup)
# ===================================================================
# PE Array Channels
# ===================================================================
self.ifmap_pe_data_chns = ModuleList()
self.weights_pe_data_chns = ModuleList()
self.psum_data_chns = ModuleList()
self.pe_data_chns = {'ifmap': self.ifmap_pe_data_chns,\
'weights': self.weights_pe_data_chns,\
'psum': self.psum_data_chns}
for pe_row in range(self.pe_array_row):
for chn_type, chn_row in self.pe_data_chns.items():
chn_row.append(ModuleList())
for pe_col in range(self.pe_array_col):
for chn_type, chn_col in self.pe_data_chns.items():
chn_col[pe_row].append(Channel())
self.psum_data_chns.append(ModuleList())
for pe_col in range(self.pe_array_col):
self.psum_data_chns[-1].append(Channel())
# ===================================================================
# Destination Calculators for NoC
# ===================================================================
self.ifmap_NoC_destination_chn = ModuleList(Channel())
ifmap_NoC_destination_calculator_setup = \
{'out_chn': self.ifmap_NoC_destination_chn[0],\
'out_channel_width': 1,\
'debug': 'IFmapNoCDestCalc'}
self.ifmap_NoC_destination_calculator = IFmapNoCDestCalc(
ifmap_NoC_destination_calculator_setup)
self.weights_NoC_destination_chn = ModuleList(Channel())
weights_NoC_desitnation_calculator_setup = \
{'out_chn': self.weights_NoC_destination_chn[0],\
'out_channel_width': 1,\
'debug': 'WeightsNoCDestCalc'}
self.weights_NoC_destionation_calculator = WeightsNoCDestCalc(
weights_NoC_desitnation_calculator_setup)
self.psum_in_NoC_destination_chn = ModuleList(Channel())
psum_in_NoC_desitnation_calculator_setup = \
{'out_chn': self.psum_in_NoC_destination_chn[0],\
'out_channel_width': 1,\
'debug': 'PsumInNoCDestCalc'}
self.psum_in_NoC_destionation_calculator = PsumInNoCDestCalc(
psum_in_NoC_desitnation_calculator_setup)
# ===================================================================
# NoCs
# ===================================================================
# weights serializer 4:1
self.weights_serializered_data_chn = Channel()
weights_serializer_setup = {'in_chn': self.weights_out_chns[0],\
'out_chn': self.weights_serializered_data_chn,\
'ratio': setup['weights_seri_ratio'],\
'debug': 'weights_serialzer'}
self.weights_serializer = WeightsSerializer(weights_serializer_setup)
# Weights NoC: weightGLB -> PEs
weights_noc_setup = {'rd_chns': self.pe_data_chns['weights'],\
'wr_chns': self.weights_serializered_data_chn,\
'dest_chns': self.weights_NoC_destination_chn,\
'debug': 'WeightsNoC'}
self.weightsNoC = WeightsNoC(weights_noc_setup)
# -------------- IFmap NoC: IfmapGLB -> PEs -----------------------
# ifmap serializer 4:1
self.ifmap_serialized_data_chn = Channel()
ifmap_serializer_setup = {'in_chn': self.ifmap_out_chns[0],\
'out_chn': self.ifmap_serialized_data_chn,\
'ratio': setup['ifmap_seri_ratio'],\
'debug': 'ifmap_serializer'}
self.ifmap_serilizer = IfmapSerializer(ifmap_serializer_setup)
# ifmap NoC
ifmap_noc_wr_chns = ModuleList()
ifmap_noc_wr_chns.append(self.ifmap_serialized_data_chn)
ifmap_noc_setup = {'rd_chns': self.pe_data_chns['ifmap'],\
'wr_chns': ifmap_noc_wr_chns,\
'dest_chns': self.ifmap_NoC_destination_chn,\
'debug': 'IFmapNoC'}
self.ifmapNoC = IFMapNoC(ifmap_noc_setup)
# ---------------------------------------------------------------------
# -------------- PsumRd NoC: Psum GLB -> PEs ----------------------
# psum serializer 4:1
self.psum_serialized_data_chn = Channel()
psum_serializer_setup = {'in_chn': self.psum_out_chns[0],\
'out_chn': self.psum_serialized_data_chn,\
'ratio': setup['psum_seri_ratio'],\
'debug': 'psum_serialzer'}
self.psum_serializer = PsumSerializer(psum_serializer_setup)
# psum read noc
pe_noc_wr_chn = ModuleList(self.psum_serialized_data_chn)
psum_rd_noc_setup = {'rd_chns': self.psum_data_chns,\
'wr_chns': pe_noc_wr_chn,\
'dest_chns': self.psum_in_NoC_destination_chn,\
'debug': 'PsumRdNoC'}
self.psumRdNoC = PsumRdNoC(psum_rd_noc_setup)
# ---------------------------------------------------------------------
# -------------- PsumWrNoC: PEs -> ifmapPsum GLB ---------------------
self.psum_out_noc_rd_chns = ModuleList()
self.psum_out_noc_rd_chns.append(
self.psum_update_chns[0]) # write back to GLB
self.psum_out_noc_rd_chns.append(self.psum_out_chn) # write offchip
psum_wr_noc_setup = {'rd_chns': self.psum_out_noc_rd_chns,\
'wr_chns': self.psum_data_chns,\
'debug': 'PsumWrNoC'}
self.psumWrNoC = PsumWrNoC(psum_wr_noc_setup)
# ===================================================================
# PE Array
# ===================================================================
# general setup of a PE
PE_setup = setup['PE']
self.PE = ModuleList()
for pe_row in range(self.pe_array_row):
self.PE.append(ModuleList())
for pe_col in range(self.pe_array_col):
# PE specific setup
PE_setup['row'] = pe_row
PE_setup['col'] = pe_col
PE_setup['weights_data_in_chn'] = self.weights_pe_data_chns[
pe_row][pe_col]
PE_setup['ifmap_data_in_chn'] = self.ifmap_pe_data_chns[
pe_row][pe_col]
PE_setup['psum_data_in_chn'] = self.psum_data_chns[pe_row][
pe_col]
PE_setup['psum_data_out_chn'] = self.psum_data_chns[pe_row +
1][pe_col]
self.PE[pe_row].append(PE(PE_setup))
def configure(self, config):
# extract the shape and mapping information
self.mapping = config['mapping']
self.shape = config['shape']
# --------------------------------------------------------------------
# Configure Global Buffers
# --------------------------------------------------------------------
self.weights_glb.configure(config['WeightsGLB'])
self.ifmap_glb.configure(config['IFmapGLB'])
self.psum_glb.configure(config['PsumGLB'])
# --------------------------------------------------------------------
# Configure Destination Calculators
# --------------------------------------------------------------------
shape_mapping_info = {'mapping': self.mapping, 'shape': self.shape}
self.ifmap_NoC_destination_calculator.configure(shape_mapping_info)
self.weights_NoC_destionation_calculator.configure(shape_mapping_info)
self.psum_in_NoC_destionation_calculator.configure(shape_mapping_info)
# --------------------------------------------------------------------
# Configure NoCs
# --------------------------------------------------------------------
self.weightsNoC.configure()
self.ifmapNoC.configure()
self.psumRdNoC.configure()
# >> self defined NoC module
self.psumWrNoC.configure({
'mapping': self.mapping,
'shape': self.shape
})
# --------------------------------------------------------------------
# Configure Serializers/Deserilizers
# --------------------------------------------------------------------
self.ifmap_serilizer.configure()
self.psum_serializer.configure()
self.weights_serializer.configure()
# --------------------------------------------------------------------
# Configure PE Arrays
# --------------------------------------------------------------------
PE_config = config['PE']
for pe_row in range(self.pe_array_row):
for pe_col in range(self.pe_array_col):
# determine if the PE needs to be used for this layer
PE_config['clk_gated'] = True if pe_col >= self.mapping['M0'] or \
pe_row >= self.mapping['C0']\
else False
self.PE[pe_row][pe_col].configure(PE_config)
class OSArch(Module):
def instantiate(self, arr_x, arr_y,
input_chn, output_chn,
chn_per_word,
ifmap_glb_depth, weight_glb_depth):
# PE static configuration (immutable)
self.name = 'chip'
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.stat_type = 'show'
# Instantiate DRAM IO channels
self.input_chn = input_chn
self.output_chn = output_chn
# Instantiate input deserializer and output serializer
self.ifmap_wr_chn = Channel(name='ifmap_wr_chn')
self.weights_wr_chn = Channel(name='weights_wr_chn')
self.bias_wr_chn = Channel(name='bias_wr_chn')
self.deserializer = InputDeserializer(self.input_chn, self.ifmap_wr_chn,
self.weights_wr_chn, self.bias_wr_chn, arr_x, arr_y,
chn_per_word)
# Instantiate GLB and GLB channels
self.ifmap_rd_chn = Channel(3, name='ifmap_rd_chn')
self.ifmap_glb = GLB(self.ifmap_wr_chn, self.ifmap_rd_chn,
ifmap_glb_depth, self.arr_y, chn_per_word, name='ifmap_glb')
self.weights_rd_chn = Channel(name='weights_rd_chn')
self.weights_glb = GLB(self.weights_wr_chn, self.weights_rd_chn, weight_glb_depth, self.arr_x, self.chn_per_word, name='weight_glb')
# PE Array and local channel declaration
self.pe_array = ModuleList()
self.pe_ifmap_chns = ModuleList()
self.pe_weight_chns = ModuleList()
self.pe_bias_chns = ModuleList()
self.pe_out_chns = ModuleList()
# Actual array instantiation
for y in range(self.arr_y):
self.pe_array.append(ModuleList())
self.pe_ifmap_chns.append(ModuleList())
self.pe_weight_chns.append(ModuleList())
self.pe_out_chns.append(ModuleList())
for x in range(self.arr_x):
self.pe_ifmap_chns[y].append(Channel(32, name='pe_ifmap_chns_{}_{}'.format(x, y)))
self.pe_weight_chns[y].append(Channel(32, name='pe_filter_chns_{}_{}'.format(x, y)))
self.pe_out_chns[y].append(Channel(32, name='pe_psum_chns_{}_{}'.format(x, y)))
self.pe_array[y].append(
PE(x, y,
self.pe_ifmap_chns[y][x],
self.pe_weight_chns[y][x],
self.pe_out_chns[y][x],
)
)
# Setup NoC to deliver weights, ifmaps and psums
self.weight_noc = WeightsNoC(self.weights_rd_chn, self.pe_weight_chns, self.arr_x, self.arr_y)
self.ifmap_noc = IFMapNoC(self.ifmap_rd_chn, self.pe_ifmap_chns, self.arr_x, self.arr_y)
self.serializer = OutputSerializer(self.output_chn, self.pe_out_chns, self.arr_x, self.arr_y, chn_per_word)
def configure(self, batch_size, input_size, output_size):
self.deserializer.configure(batch_size, input_size, output_size)
self.ifmap_glb.configure(batch_size * input_size // self.arr_y,
output_size // self.arr_x,
batch_size)
self.weights_glb.configure((input_size+1) * output_size // self.arr_x,
batch_size // self.arr_y,
self.arr_x)
self.weight_noc.configure()
self.ifmap_noc.configure()
for y in range(self.arr_y):
for x in range(self.arr_x):
self.pe_array[y][x].configure(input_size)