def mem_signal_logic(self):
if self.if_sram_cfg_s.wr_en:
if self.if_sram_cfg_s.wr_addr[self._params.bank_byte_offset -
1] == 0:
self.mem_wr_en = 1
self.mem_rd_en_w = 0
self.mem_addr = self.if_sram_cfg_s.wr_addr
self.mem_data_in = concat(
const(
0, self._params.bank_data_width -
self._params.axi_data_width),
self.if_sram_cfg_s.wr_data)
self.mem_data_in_bit_sel = concat(
const(
0, self._params.bank_data_width -
self._params.axi_data_width),
const(2**self._params.axi_data_width - 1,
self._params.axi_data_width))
else:
self.mem_wr_en = 1
self.mem_rd_en_w = 0
self.mem_addr = self.if_sram_cfg_s.wr_addr
self.mem_data_in = concat(
self.if_sram_cfg_s.wr_data[self._params.bank_data_width -
self._params.axi_data_width - 1,
0],
const(0, self._params.axi_data_width))
self.mem_data_in_bit_sel = concat(
const(
2**(self._params.bank_data_width -
self._params.axi_data_width) - 1,
self._params.bank_data_width -
self._params.axi_data_width),
const(0, self._params.axi_data_width))
elif self.if_sram_cfg_s.rd_en:
self.mem_wr_en = 0
self.mem_rd_en_w = 1
self.mem_addr = self.if_sram_cfg_s.rd_addr
self.mem_data_in = 0
self.mem_data_in_bit_sel = 0
elif self.packet_wr_en:
self.mem_wr_en = 1
self.mem_rd_en_w = 0
self.mem_addr = self.packet_wr_addr
self.mem_data_in = self.packet_wr_data
self.mem_data_in_bit_sel = self.packet_wr_data_bit_sel
elif self.packet_rd_en:
self.mem_wr_en = 0
self.mem_rd_en_w = 1
self.mem_addr = self.packet_rd_addr
self.mem_data_in = 0
self.mem_data_in_bit_sel = 0
else:
self.mem_wr_en = 0
self.mem_rd_en_w = 0
self.mem_addr = 0
self.mem_data_in = 0
self.mem_data_in_bit_sel = 0
def zext(gen, wire, size):
if wire.width >= size:
return wire
else:
zext_signal = gen.var(f"{wire.name}_zext", size)
gen.wire(zext_signal, kts.concat(kts.const(0, size - wire.width),
wire))
return zext_signal
def add_pipeline(self):
sram_signals_reset_high_in = concat(self.WEB, self.CEB, self.web_demux,
self.ceb_demux, self.BWEB)
sram_signals_reset_high_out = concat(self.WEB_d, self.CEB_d,
self.web_demux_d,
self.ceb_demux_d, self.BWEB_d)
self.sram_signals_reset_high_pipeline = Pipeline(
width=sram_signals_reset_high_in.width,
depth=self._params.sram_gen_pipeline_depth,
reset_high=True)
self.add_child("sram_signals_reset_high_pipeline",
self.sram_signals_reset_high_pipeline,
clk=self.CLK,
clk_en=const(1, 1),
reset=self.RESET,
in_=sram_signals_reset_high_in,
out_=sram_signals_reset_high_out)
sram_signals_in = concat(self.a_sram, self.sram_sel, self.D)
sram_signals_out = concat(self.a_sram_d, self.sram_sel_d, self.D_d)
self.sram_signals_pipeline = Pipeline(
width=sram_signals_in.width,
depth=self._params.sram_gen_pipeline_depth)
self.add_child("sram_signals_pipeline",
self.sram_signals_pipeline,
clk=self.CLK,
clk_en=const(1, 1),
reset=self.RESET,
in_=sram_signals_in,
out_=sram_signals_out)
self.sram_signals_output_pipeline = Pipeline(
width=self.sram_macro_width,
depth=self._params.sram_gen_output_pipeline_depth)
self.add_child("sram_signals_output_pipeline",
self.sram_signals_output_pipeline,
clk=self.CLK,
clk_en=const(1, 1),
reset=self.RESET,
in_=self.Q_w,
out_=self.Q)
def mem_ff(self):
if self.CEB == 0:
self.Q_w = concat(
self.mem[resize((self.A << 2) + 3, self.addr_width + 2)],
self.mem[resize((self.A << 2) + 2, self.addr_width + 2)],
self.mem[resize((self.A << 2) + 1,
self.addr_width + 2)], self.mem[resize(
(self.A << 2), self.addr_width + 2)])
if self.WEB == 0:
for i in range(self.data_width):
if self.BWEB[i] == 0:
self.mem[resize(
(self.A << 2) + i // 16,
self.addr_width + 2)][resize(
i % 16, clog2(
self._params.cgra_data_width))] = self.D[i]
def test_regression_concat():
from kratos import concat
parent = Generator("parent")
for i in range(2):
child = Generator("child")
in_ = child.input("in", 1)
out_ = child.output("out", 1)
child.wire(out_, in_)
parent.add_child("child{0}".format(i), child)
in_ = parent.input("in", 1)
out_2 = parent.output("out2", 2)
for i in range(2):
parent.wire(in_, parent["child{0}".format(i)].ports["in"])
parent.wire(out_2,
concat(parent["child0"].ports.out, parent["child1"].ports.out))
src = verilog(parent, optimize_passthrough=False)["parent"]
assert "assign out2 = {child0_out, child1_out};" in src
def __init__(self, agg_height, data_width, mem_width, max_agg_schedule):
super().__init__("aggregation_buffer")
self.agg_height = agg_height
self.data_width = data_width
self.mem_width = mem_width
self.max_agg_schedule = max_agg_schedule # This is the maximum length of the schedule
self.fw_int = int(self.mem_width / self.data_width)
# Clock and Reset
self._clk = self.clock("clk")
self._rst_n = self.reset("rst_n")
# Inputs
# Bring in a single element into an AggregationBuffer w/ valid signaling
self._data_in = self.input("data_in", self.data_width)
self._valid_in = self.input("valid_in", 1)
self._align = self.input("align", 1)
# Outputs
self._data_out = self.output("data_out", self.mem_width)
self._valid_out = self.output("valid_out", 1)
self._data_out_chop = []
for i in range(self.fw_int):
self._data_out_chop.append(
self.output(f"data_out_chop_{i}", self.data_width))
self.add_stmt(self._data_out_chop[i].assign(
self._data_out[(self.data_width * (i + 1)) - 1,
self.data_width * i]))
# CONFIG:
# We receive a periodic (doesn't need to be, but has a maximum schedule,
# so...possibly the schedule is a for loop?
# Tells us where to write successive elements...
self._in_schedule = self.input("in_sched",
max(1, clog2(self.agg_height)),
size=self.max_agg_schedule,
explicit_array=True,
packed=True)
doc = "Input schedule for aggregation buffer. Enumerate which" + \
f" of {self.agg_height} buffers to write to."
self._in_schedule.add_attribute(ConfigRegAttr(doc))
self._in_period = self.input("in_period", clog2(self.max_agg_schedule))
doc = "Input period for aggregation buffer. 1 is a reasonable" + \
" setting for most applications"
self._in_period.add_attribute(ConfigRegAttr(doc))
# ...and which order to output the blocks
self._out_schedule = self.input("out_sched",
max(1, clog2(agg_height)),
size=self.max_agg_schedule,
explicit_array=True,
packed=True)
doc = "Output schedule for aggregation buffer. Enumerate which" + \
f" of {self.agg_height} buffers to write to SRAM from."
self._out_schedule.add_attribute(ConfigRegAttr(doc))
self._out_period = self.input("out_period",
clog2(self.max_agg_schedule))
self._out_period.add_attribute(
ConfigRegAttr("Output period for aggregation buffer"))
self._in_sched_ptr = self.var("in_sched_ptr",
clog2(self.max_agg_schedule))
self._out_sched_ptr = self.var("out_sched_ptr",
clog2(self.max_agg_schedule))
# Local Signals
self._aggs_out = self.var("aggs_out",
self.mem_width,
size=self.agg_height,
packed=True,
explicit_array=True)
self._aggs_sep = []
for i in range(self.agg_height):
self._aggs_sep.append(
self.var(f"aggs_sep_{i}",
self.data_width,
size=self.fw_int,
packed=True))
self._valid_demux = self.var("valid_demux", self.agg_height)
self._align_demux = self.var("align_demux", self.agg_height)
self._next_full = self.var("next_full", self.agg_height)
self._valid_out_mux = self.var("valid_out_mux", self.agg_height)
for i in range(self.agg_height):
# Add in the children aggregators...
self.add_child(f"agg_{i}",
Aggregator(self.data_width,
mem_word_width=self.fw_int),
clk=self._clk,
rst_n=self._rst_n,
in_pixels=self._data_in,
valid_in=self._valid_demux[i],
agg_out=self._aggs_sep[i],
valid_out=self._valid_out_mux[i],
next_full=self._next_full[i],
align=self._align_demux[i])
portlist = []
if self.fw_int == 1:
self.wire(self._aggs_out[i], self._aggs_sep[i])
else:
for j in range(self.fw_int):
portlist.append(self._aggs_sep[i][self.fw_int - 1 - j])
self.wire(self._aggs_out[i], kts.concat(*portlist))
# Sequential code blocks
self.add_code(self.update_in_sched_ptr)
self.add_code(self.update_out_sched_ptr)
# Combinational code blocks
self.add_code(self.valid_demux_comb)
self.add_code(self.align_demux_comb)
self.add_code(self.valid_out_comb)
self.add_code(self.output_data_comb)
def test_single_concat():
from kratos import concat
mod = Generator("mod")
a = mod.var("a", 1)
r = concat([a])
assert id(r) == id(a)
def __init__(self,
fetch_width=16,
data_width=16,
memory_depth=32,
num_tiles=1,
int_in_ports=2,
int_out_ports=2,
strg_wr_ports=2,
strg_rd_ports=2,
read_delay=0,
rw_same_cycle=True,
separate_addresses=True):
assert not (memory_depth & (memory_depth - 1)), "Memory depth needs to be a power of 2"
super().__init__("rw_arbiter")
# Absorb inputs
self.fetch_width = fetch_width
self.data_width = data_width
self.fw_int = int(self.fetch_width / self.data_width)
self.int_in_ports = int_in_ports
self.int_out_ports = int_out_ports
self.strg_wr_ports = strg_wr_ports
self.strg_rd_ports = strg_rd_ports
self.memory_depth = memory_depth
self.num_tiles = num_tiles
self.mem_addr_width = clog2(self.num_tiles * self.memory_depth)
self.read_delay = read_delay
self.rw_same_cycle = rw_same_cycle
self.separate_addresses = separate_addresses
# Clock and Reset
self._clk = self.clock("clk")
self._rst_n = self.reset("rst_n")
# Generate the packed struct of the right size
port_pkt_struct = create_port_pkt(self.fetch_width, self.int_out_ports)
# Inputs
self._wen_in = self.input("wen_in", self.strg_wr_ports)
# self._wen_en = self.input("wen_en", self.strg_wr_ports)
self._w_data = self.input("w_data",
self.data_width,
size=(self.strg_wr_ports,
self.fw_int),
explicit_array=True,
packed=True)
self._w_addr = self.input("w_addr", self.mem_addr_width,
size=self.strg_wr_ports,
explicit_array=True,
packed=True)
self._data_from_mem = self.input("data_from_mem",
self.data_width,
size=(self.strg_rd_ports,
self.fw_int),
explicit_array=True,
packed=True)
self._mem_valid_data = self.input("mem_valid_data",
self.strg_rd_ports)
self._out_mem_valid_data = self.output("out_mem_valid_data",
self.strg_rd_ports)
self._ren_in = self.input("ren_in", self.int_out_ports)
self._ren_en = self.input("ren_en", self.int_out_ports)
self._rd_addr = self.input("rd_addr",
self.mem_addr_width,
size=self.int_out_ports,
explicit_array=True,
packed=True)
self._rd_addr_sel = self.var("rd_addr_sel",
self.mem_addr_width,
size=self.strg_rd_ports,
explicit_array=True,
packed=True)
# Outputs
self._out_data = self.output("out_data",
self.data_width,
size=(self.strg_rd_ports,
self.fw_int),
explicit_array=True,
packed=True)
self._out_port = self.output("out_port", self.int_out_ports,
size=self.strg_rd_ports,
explicit_array=True,
packed=True)
self._out_valid = self.output("out_valid", self.strg_rd_ports)
self._cen_mem = self.output("cen_mem", self.strg_rd_ports)
self._wen_mem = self.output("wen_mem", self.strg_wr_ports)
self._data_to_mem = self.output("data_to_mem",
self.data_width,
size=(self.strg_wr_ports,
self.fw_int),
explicit_array=True,
packed=True)
# In this case, need separate addresses
if self.separate_addresses:
self._wr_addr_to_mem = self.output("wr_addr_to_mem",
self.mem_addr_width,
size=self.strg_wr_ports,
explicit_array=True,
packed=True)
self._rd_addr_to_mem = self.output("rd_addr_to_mem",
self.mem_addr_width,
size=self.strg_rd_ports,
explicit_array=True,
packed=True)
# If the addresses are combined, we better have in==out
else:
assert self.strg_rd_ports == self.strg_wr_ports, \
"Cannot have coalesced address with mismatch port count"
assert not self.rw_same_cycle, \
"Cannot read and write with a shared address...set rw_same_cycle to false"
self._addr_to_mem = self.output("addr_to_mem", self.mem_addr_width,
size=self.strg_rd_ports,
explicit_array=True,
packed=True)
self._out_ack = self.output("out_ack", self.int_out_ports)
# Local
# self._rd_data = self.var("rd_data", self.fetch_width)
self._wen_int = self.var("wen_int", self.strg_wr_ports)
self._ren_int = self.var("ren_int", self.int_out_ports)
self.wire(self._ren_int, self._ren_in & self._ren_en)
self.wire(self._wen_int, self._wen_in) # & self._wen_en)
self._rd_valid = self.var("rd_valid", self.strg_rd_ports)
self._rd_port = self.var("rd_port", self.int_out_ports,
size=self.strg_rd_ports,
explicit_array=True,
packed=True)
self._next_rd_port = self.var("next_rd_port", self.int_out_ports,
size=self.strg_rd_ports,
explicit_array=True,
packed=True)
# For demuxing the read ports
self._done = self.var("done", self.strg_rd_ports)
self.add_code(self.set_next_read_port_lowest)
if(self.strg_rd_ports > 1):
self._idx_cnt = self.var("idx_cnt", 5,
size=self.strg_rd_ports - 1,
explicit_array=True,
packed=True)
for i in range(self.strg_rd_ports - 1):
self.add_code(self.set_next_read_port_alt, index=i + 1)
# If we have more than one read port, we need to use slightly different logic
# to set the other reads...
self._next_rd_port_red = self.var("next_rd_port_red", self.int_out_ports)
for i in range(self.int_out_ports):
temp_port = self._next_rd_port[0][i]
for j in range(self.strg_rd_ports - 1):
temp_port = kts.concat(temp_port, self._next_rd_port[j + 1][i])
self.wire(self._next_rd_port_red[i], temp_port.r_or())
# The next read port can be used to acknowledge reads
# We do not need to gate the ack if we can read and write in the same cycle
if self.rw_same_cycle:
self.wire(self._out_ack,
self._next_rd_port_red)
else:
self.wire(self._out_ack,
self._next_rd_port_red & kts.concat(*([~self._wen_int] * self._out_ack.width)))
# self.add_code(self.mem_controls)
if self.separate_addresses:
for i in range(self.strg_wr_ports):
self.add_code(self.mem_controls_wr, idx=i)
for i in range(self.strg_rd_ports):
self.add_code(self.mem_controls_rd, idx=i)
else:
for i in range(self.strg_rd_ports):
self.add_code(self.mem_controls_combined, idx=i)
if self.read_delay == 1:
for i in range(self.strg_rd_ports):
self.add_code(self.next_read_valid, idx=i)
else:
for i in range(self.strg_rd_ports):
self.add_code(self.zero_delay_read, idx=i)
self.add_code(self.output_stage)
def __init__(self,
data_width,
config_addr_width,
addr_width,
fetch_width,
total_sets,
sets_per_macro):
super().__init__("storage_config_seq")
self.data_width = data_width
self.config_addr_width = config_addr_width
self.addr_width = addr_width
self.fetch_width = fetch_width
self.fw_int = int(self.fetch_width / self.data_width)
self.total_sets = total_sets
self.sets_per_macro = sets_per_macro
self.banks = int(self.total_sets / self.sets_per_macro)
self.set_addr_width = clog2(total_sets)
# self.storage_addr_width = self.
# Clock and Reset
self._clk = self.clock("clk")
self._rst_n = self.reset("rst_n")
# Inputs
# phases = [] TODO
# Take in the valid and data and attach an address + direct to a port
self._config_data_in = self.input("config_data_in",
self.data_width)
self._config_addr_in = self.input("config_addr_in",
self.config_addr_width)
self._config_wr = self.input("config_wr", 1)
self._config_rd = self.input("config_rd", 1)
self._config_en = self.input("config_en", self.total_sets)
self._clk_en = self.input("clk_en", 1)
self._rd_data_stg = self.input("rd_data_stg", self.data_width,
size=(self.banks,
self.fw_int),
explicit_array=True,
packed=True)
self._wr_data = self.output("wr_data",
self.data_width,
size=self.fw_int,
explicit_array=True,
packed=True)
self._rd_data_out = self.output("rd_data_out", self.data_width,
size=self.total_sets,
explicit_array=True,
packed=True)
self._addr_out = self.output("addr_out",
self.addr_width)
# One set per macro means we directly send the config address through
if self.sets_per_macro == 1:
width = self.addr_width - self.config_addr_width
if width > 0:
self.wire(self._addr_out, kts.concat(kts.const(0, width), self._config_addr_in))
else:
self.wire(self._addr_out, self._config_addr_in[self.addr_width - 1, 0])
else:
width = self.addr_width - self.config_addr_width - clog2(self.sets_per_macro)
self._set_to_addr = self.var("set_to_addr",
clog2(self.sets_per_macro))
self._reduce_en = self.var("reduce_en", self.sets_per_macro)
for i in range(self.sets_per_macro):
reduce_var = self._config_en[i]
for j in range(self.banks - 1):
reduce_var = kts.concat(reduce_var, self._config_en[i + (self.sets_per_macro * (j + 1))])
self.wire(self._reduce_en[i], reduce_var.r_or())
self.add_code(self.demux_set_addr)
if width > 0:
self.wire(self._addr_out, kts.concat(kts.const(0, width),
self._set_to_addr,
self._config_addr_in))
else:
self.wire(self._addr_out, kts.concat(self._set_to_addr, self._config_addr_in))
self._wen_out = self.output("wen_out", self.banks)
self._ren_out = self.output("ren_out", self.banks)
# Handle data passing
if self.fw_int == 1:
# If word width is same as data width, just pass everything through
self.wire(self._wr_data[0], self._config_data_in)
# self.wire(self._rd_data_out, self._rd_data_stg[0])
num = 0
for i in range(self.banks):
for j in range(self.sets_per_macro):
self.wire(self._rd_data_out[num], self._rd_data_stg[i])
num = num + 1
else:
self._data_wr_reg = self.var("data_wr_reg",
self.data_width,
size=self.fw_int - 1,
packed=True,
explicit_array=True)
# self._data_rd_reg = self.var("data_rd_reg",
# self.data_width,
# size=self.fw_int - 1,
# packed=True,
# explicit_array=True)
# Have word counter for repeated reads/writes
self._cnt = self.var("cnt", clog2(self.fw_int))
self._rd_cnt = self.var("rd_cnt", clog2(self.fw_int))
self.add_code(self.update_cnt)
self.add_code(self.update_rd_cnt)
# Gate wen if not about to finish the word
num = 0
for i in range(self.banks):
for j in range(self.sets_per_macro):
self.wire(self._rd_data_out[num], self._rd_data_stg[i][self._rd_cnt])
num = num + 1
# Deal with writing to the data buffer
self.add_code(self.write_buffer)
# Wire the reg + such to this guy
for i in range(self.fw_int - 1):
self.wire(self._wr_data[i], self._data_wr_reg[i])
self.wire(self._wr_data[self.fw_int - 1], self._config_data_in)
# If we have one bank, we can just always rd/wr from that one
if self.banks == 1:
if self.fw_int == 1:
self.wire(self._wen_out, self._config_wr)
else:
self.wire(self._wen_out,
self._config_wr & (self._cnt == (self.fw_int - 1)))
self.wire(self._ren_out, self._config_rd)
# Otherwise we need to extract the bank from the set
else:
if self.fw_int == 1:
for i in range(self.banks):
width = self.sets_per_macro
self.wire(self._wen_out[i], self._config_wr &
self._config_en[(i + 1) * width - 1, i * width].r_or())
else:
for i in range(self.banks):
width = self.sets_per_macro
self.wire(self._wen_out[i],
self._config_wr & self._config_en[(i + 1) * width - 1, i * width].r_or() &
(self._cnt == (self.fw_int - 1)))
for i in range(self.banks):
width = self.sets_per_macro
self.wire(self._ren_out[i],
self._config_rd & self._config_en[(i + 1) * width - 1, i * width].r_or())
def __init__(self,
interconnect_input_ports,
interconnect_output_ports,
depth_width=16,
sprt_stcl_valid=False,
stcl_cnt_width=16,
stcl_iter_support=4):
super().__init__("app_ctrl", debug=True)
self.int_in_ports = interconnect_input_ports
self.int_out_ports = interconnect_output_ports
self.depth_width = depth_width
self.sprt_stcl_valid = sprt_stcl_valid
self.stcl_cnt_width = stcl_cnt_width
self.stcl_iter_support = stcl_iter_support
# Clock and Reset
self._clk = self.clock("clk")
self._rst_n = self.reset("rst_n")
# IO
self._wen_in = self.input("wen_in", self.int_in_ports)
self._ren_in = self.input("ren_in", self.int_out_ports)
self._ren_update = self.input("ren_update", self.int_out_ports)
self._tb_valid = self.input("tb_valid", self.int_out_ports)
self._valid_out_data = self.output("valid_out_data", self.int_out_ports)
self._valid_out_stencil = self.output("valid_out_stencil", self.int_out_ports)
# Send tb valid to valid out for now...
if self.sprt_stcl_valid:
# Add the config registers to watch
self._ranges = self.input("ranges", self.stcl_cnt_width,
size=self.stcl_iter_support,
packed=True,
explicit_array=True)
self._ranges.add_attribute(ConfigRegAttr("Ranges of stencil valid generator"))
self._threshold = self.input("threshold", self.stcl_cnt_width,
size=self.stcl_iter_support,
packed=True,
explicit_array=True)
self._threshold.add_attribute(ConfigRegAttr("Threshold of stencil valid generator"))
self._dim_counter = self.var("dim_counter", self.stcl_cnt_width,
size=self.stcl_iter_support,
packed=True,
explicit_array=True)
self._update = self.var("update", self.stcl_iter_support)
self.wire(self._update[0], const(1, 1))
for i in range(self.stcl_iter_support - 1):
self.wire(self._update[i + 1],
(self._dim_counter[i] == (self._ranges[i] - 1)) & self._update[i])
for i in range(self.stcl_iter_support):
self.add_code(self.dim_counter_update, idx=i)
# Now we need to just compute stencil valid
threshold_comps = [self._dim_counter[_i] >= self._threshold[_i] for _i in range(self.stcl_iter_support)]
self.wire(self._valid_out_stencil[0], kts.concat(*threshold_comps).r_and())
for i in range(self.int_out_ports - 1):
# self.wire(self._valid_out_stencil[i + 1], 0)
# for multiple ports
self.wire(self._valid_out_stencil[i + 1], kts.concat(*threshold_comps).r_and())
else:
self.wire(self._valid_out_stencil, self._tb_valid)
# Now gate the valid with stencil valid
self.wire(self._valid_out_data, self._tb_valid & self._valid_out_stencil)
self._wr_delay_state_n = self.var("wr_delay_state_n", self.int_out_ports)
self._wen_out = self.output("wen_out", self.int_in_ports)
self._ren_out = self.output("ren_out", self.int_out_ports)
self._write_depth_wo = self.input("write_depth_wo", self.depth_width,
size=self.int_in_ports,
explicit_array=True,
packed=True)
self._write_depth_wo.add_attribute(ConfigRegAttr("Depth of writes"))
self._write_depth_ss = self.input("write_depth_ss", self.depth_width,
size=self.int_in_ports,
explicit_array=True,
packed=True)
self._write_depth_ss.add_attribute(ConfigRegAttr("Depth of writes"))
self._write_depth = self.var("write_depth", self.depth_width,
size=self.int_in_ports,
explicit_array=True,
packed=True)
for i in range(self.int_in_ports):
self.wire(self._write_depth[i],
kts.ternary(self._wr_delay_state_n[i],
self._write_depth_ss[i],
self._write_depth_wo[i]))
self._read_depth = self.input("read_depth", self.depth_width,
size=self.int_out_ports,
explicit_array=True,
packed=True)
self._read_depth.add_attribute(ConfigRegAttr("Depth of reads"))
self._write_count = self.var("write_count", self.depth_width,
size=self.int_in_ports,
explicit_array=True,
packed=True)
self._read_count = self.var("read_count", self.depth_width,
size=self.int_out_ports,
explicit_array=True,
packed=True)
self._write_done = self.var("write_done", self.int_in_ports)
self._write_done_ff = self.var("write_done_ff", self.int_in_ports)
self._read_done = self.var("read_done", self.int_out_ports)
self._read_done_ff = self.var("read_done_ff", self.int_out_ports)
self.in_port_bits = max(1, kts.clog2(self.int_in_ports))
self._input_port = self.input("input_port", self.in_port_bits,
size=self.int_out_ports,
explicit_array=True,
packed=True)
self._input_port.add_attribute(ConfigRegAttr("Relative input port for an output port"))
self.out_port_bits = max(1, kts.clog2(self.int_out_ports))
self._output_port = self.input("output_port", self.out_port_bits,
size=self.int_in_ports,
explicit_array=True,
packed=True)
self._output_port.add_attribute(ConfigRegAttr("Relative output port for an input port"))
self._prefill = self.input("prefill", self.int_out_ports)
self._prefill.add_attribute(ConfigRegAttr("Is the input stream prewritten?"))
for i in range(self.int_out_ports):
self.add_code(self.set_read_done, idx=i)
if self.int_in_ports == 1:
self.add_code(self.set_read_done_ff_one_wr, idx=i)
else:
self.add_code(self.set_read_done_ff, idx=i)
# self._write_done_comb = self.var("write_done_comb", self.int_in_ports)
for i in range(self.int_in_ports):
self.add_code(self.set_write_done, idx=i)
self.add_code(self.set_write_done_ff, idx=i)
for i in range(self.int_in_ports):
self.add_code(self.set_write_cnt, idx=i)
for i in range(self.int_out_ports):
if self.int_in_ports == 1:
self.add_code(self.set_read_cnt_one_wr, idx=i)
else:
self.add_code(self.set_read_cnt, idx=i)
for i in range(self.int_out_ports):
if self.int_in_ports == 1:
self.add_code(self.set_wr_delay_state_one_wr, idx=i)
else:
self.add_code(self.set_wr_delay_state, idx=i)
self._read_on = self.var("read_on", self.int_out_ports)
for i in range(self.int_out_ports):
self.wire(self._read_on[i], self._read_depth[i].r_or())
# If we have prefill enabled, we are skipping the initial delay step...
self.wire(self._ren_out,
(self._wr_delay_state_n | self._prefill) & ~self._read_done_ff & self._ren_in &
self._read_on)
self.wire(self._wen_out, ~self._write_done_ff & self._wen_in)
def __init__(self, _params: GlobalBufferParams):
super().__init__("global_buffer")
self._params = _params
self.header = GlbHeader(self._params)
self.clk = self.clock("clk")
self.stall = self.input("stall", self._params.num_glb_tiles)
self.reset = self.reset("reset")
# TODO: Why cgra_stall has same width as num_glb_tiles
self.cgra_stall_in = self.input("cgra_stall_in",
self._params.num_glb_tiles)
self.cgra_stall = self.output(
"cgra_stall",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.proc_wr_en = self.input("proc_wr_en", 1)
self.proc_wr_strb = self.input("proc_wr_strb",
self._params.bank_data_width // 8)
self.proc_wr_addr = self.input("proc_wr_addr",
self._params.glb_addr_width)
self.proc_wr_data = self.input("proc_wr_data",
self._params.bank_data_width)
self.proc_rd_en = self.input("proc_rd_en", 1)
self.proc_rd_addr = self.input("proc_rd_addr",
self._params.glb_addr_width)
self.proc_rd_data = self.output("proc_rd_data",
self._params.bank_data_width)
self.proc_rd_data_valid = self.output("proc_rd_data_valid", 1)
self.if_cfg_wr_en = self.input("if_cfg_wr_en", 1)
self.if_cfg_wr_addr = self.input("if_cfg_wr_addr",
self._params.axi_addr_width)
self.if_cfg_wr_data = self.input("if_cfg_wr_data",
self._params.axi_data_width)
self.if_cfg_rd_en = self.input("if_cfg_rd_en", 1)
self.if_cfg_rd_addr = self.input("if_cfg_rd_addr",
self._params.axi_addr_width)
self.if_cfg_rd_data = self.output("if_cfg_rd_data",
self._params.axi_data_width)
self.if_cfg_rd_data_valid = self.output("if_cfg_rd_data_valid", 1)
self.if_sram_cfg_wr_en = self.input("if_sram_cfg_wr_en", 1)
self.if_sram_cfg_wr_addr = self.input("if_sram_cfg_wr_addr",
self._params.glb_addr_width)
self.if_sram_cfg_wr_data = self.input("if_sram_cfg_wr_data",
self._params.axi_data_width)
self.if_sram_cfg_rd_en = self.input("if_sram_cfg_rd_en", 1)
self.if_sram_cfg_rd_addr = self.input("if_sram_cfg_rd_addr",
self._params.glb_addr_width)
self.if_sram_cfg_rd_data = self.output("if_sram_cfg_rd_data",
self._params.axi_data_width)
self.if_sram_cfg_rd_data_valid = self.output(
"if_sram_cfg_rd_data_valid", 1)
self.cgra_cfg_jtag_gc2glb_wr_en = self.input(
"cgra_cfg_jtag_gc2glb_wr_en", 1)
self.cgra_cfg_jtag_gc2glb_rd_en = self.input(
"cgra_cfg_jtag_gc2glb_rd_en", 1)
self.cgra_cfg_jtag_gc2glb_addr = self.input(
"cgra_cfg_jtag_gc2glb_addr", self._params.cgra_cfg_addr_width)
self.cgra_cfg_jtag_gc2glb_data = self.input(
"cgra_cfg_jtag_gc2glb_data", self._params.cgra_cfg_data_width)
self.stream_data_f2g = self.input(
"stream_data_f2g",
self._params.cgra_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.stream_data_valid_f2g = self.input(
"stream_data_valid_f2g",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.stream_data_g2f = self.output(
"stream_data_g2f",
self._params.cgra_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.stream_data_valid_g2f = self.output(
"stream_data_valid_g2f",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_wr_en = self.output(
"cgra_cfg_g2f_cfg_wr_en",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_rd_en = self.output(
"cgra_cfg_g2f_cfg_rd_en",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_addr = self.output(
"cgra_cfg_g2f_cfg_addr",
self._params.cgra_cfg_addr_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_data = self.output(
"cgra_cfg_g2f_cfg_data",
self._params.cgra_cfg_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.strm_g2f_start_pulse = self.input("strm_g2f_start_pulse",
self._params.num_glb_tiles)
self.strm_f2g_start_pulse = self.input("strm_f2g_start_pulse",
self._params.num_glb_tiles)
self.pcfg_start_pulse = self.input("pcfg_start_pulse",
self._params.num_glb_tiles)
self.strm_f2g_interrupt_pulse = self.output("strm_f2g_interrupt_pulse",
self._params.num_glb_tiles)
self.strm_g2f_interrupt_pulse = self.output("strm_g2f_interrupt_pulse",
self._params.num_glb_tiles)
self.pcfg_g2f_interrupt_pulse = self.output("pcfg_g2f_interrupt_pulse",
self._params.num_glb_tiles)
# local variables
self.cgra_cfg_jtag_gc2glb_wr_en_d = self.var(
"cgra_cfg_jtag_gc2glb_wr_en_d", 1)
self.cgra_cfg_jtag_gc2glb_rd_en_d = self.var(
"cgra_cfg_jtag_gc2glb_rd_en_d", 1)
self.cgra_cfg_jtag_gc2glb_addr_d = self.var(
"cgra_cfg_jtag_gc2glb_addr_d", self._params.cgra_cfg_addr_width)
self.cgra_cfg_jtag_gc2glb_data_d = self.var(
"cgra_cfg_jtag_gc2glb_data_d", self._params.cgra_cfg_data_width)
self.proc_packet_d = self.var("proc_packet_d", self.header.packet_t)
self.proc_packet_e2w_esti = self.var("proc_packet_e2w_esti",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.proc_packet_w2e_wsti = self.var("proc_packet_w2e_wsti",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.proc_packet_e2w_wsto = self.var("proc_packet_e2w_wsto",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.proc_packet_w2e_esto = self.var("proc_packet_w2e_esto",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.strm_packet_e2w_esti = self.var("strm_packet_e2w_esti",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.strm_packet_w2e_wsti = self.var("strm_packet_w2e_wsti",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.strm_packet_e2w_wsto = self.var("strm_packet_e2w_wsto",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.strm_packet_w2e_esto = self.var("strm_packet_w2e_esto",
self.header.packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.pcfg_packet_e2w_esti = self.var("pcfg_packet_e2w_esti",
self.header.rd_packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.pcfg_packet_w2e_wsti = self.var("pcfg_packet_w2e_wsti",
self.header.rd_packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.pcfg_packet_e2w_wsto = self.var("pcfg_packet_e2w_wsto",
self.header.rd_packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.pcfg_packet_w2e_esto = self.var("pcfg_packet_w2e_esto",
self.header.rd_packet_t,
size=self._params.num_glb_tiles,
packed=True)
self.cfg_tile_connected = self.var("cfg_tile_connected",
self._params.num_glb_tiles + 1)
self.cfg_pcfg_tile_connected = self.var("cfg_pcfg_tile_connected",
self._params.num_glb_tiles + 1)
self.wire(self.cfg_tile_connected[0], 0)
self.wire(self.cfg_pcfg_tile_connected[0], 0)
self.cgra_cfg_jtag_wsti_wr_en = self.var(
"cgra_cfg_jtag_wsti_wr_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_wsti_rd_en = self.var(
"cgra_cfg_jtag_wsti_rd_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_wsti_addr = self.var(
"cgra_cfg_jtag_wsti_addr",
self._params.cgra_cfg_addr_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_wsti_data = self.var(
"cgra_cfg_jtag_wsti_data",
self._params.cgra_cfg_data_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_esto_wr_en = self.var(
"cgra_cfg_jtag_esto_wr_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_esto_rd_en = self.var(
"cgra_cfg_jtag_esto_rd_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_esto_addr = self.var(
"cgra_cfg_jtag_esto_addr",
self._params.cgra_cfg_addr_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_esto_data = self.var(
"cgra_cfg_jtag_esto_data",
self._params.cgra_cfg_data_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_wsti_rd_en_bypass = self.var(
"cgra_cfg_jtag_wsti_rd_en_bypass",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_wsti_addr_bypass = self.var(
"cgra_cfg_jtag_wsti_addr_bypass",
self._params.cgra_cfg_addr_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_esto_rd_en_bypass = self.var(
"cgra_cfg_jtag_esto_rd_en_bypass",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_jtag_esto_addr_bypass = self.var(
"cgra_cfg_jtag_esto_addr_bypass",
self._params.cgra_cfg_addr_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_wsti_wr_en = self.var(
"cgra_cfg_pcfg_wsti_wr_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_wsti_rd_en = self.var(
"cgra_cfg_pcfg_wsti_rd_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_wsti_addr = self.var(
"cgra_cfg_pcfg_wsti_addr",
self._params.cgra_cfg_addr_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_wsti_data = self.var(
"cgra_cfg_pcfg_wsti_data",
self._params.cgra_cfg_data_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_esto_wr_en = self.var(
"cgra_cfg_pcfg_esto_wr_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_esto_rd_en = self.var(
"cgra_cfg_pcfg_esto_rd_en",
1,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_esto_addr = self.var(
"cgra_cfg_pcfg_esto_addr",
self._params.cgra_cfg_addr_width,
size=self._params.num_glb_tiles,
packed=True)
self.cgra_cfg_pcfg_esto_data = self.var(
"cgra_cfg_pcfg_esto_data",
self._params.cgra_cfg_data_width,
size=self._params.num_glb_tiles,
packed=True)
self.stall_w = self.var("stall_w", self._params.num_glb_tiles)
self.stall_d = self.var("stall_d", self._params.num_glb_tiles)
self.wire(self.stall_w, self.stall)
self.cgra_stall_in_w = self.var("cgra_stall_in_w",
self._params.num_glb_tiles)
self.cgra_stall_in_d = self.var("cgra_stall_in_d",
self._params.num_glb_tiles)
self.wire(self.cgra_stall_in_w, self.cgra_stall_in)
for i in range(self._params.num_glb_tiles):
self.wire(
self.cgra_stall[i],
concat(*[self.cgra_stall_in_d[i]] * self._params.cgra_per_glb))
self.strm_g2f_start_pulse_w = self.var("strm_g2f_start_pulse_w",
self._params.num_glb_tiles)
self.strm_g2f_start_pulse_d = self.var("strm_g2f_start_pulse_d",
self._params.num_glb_tiles)
self.wire(self.strm_g2f_start_pulse, self.strm_g2f_start_pulse_w)
self.strm_f2g_start_pulse_w = self.var("strm_f2g_start_pulse_w",
self._params.num_glb_tiles)
self.strm_f2g_start_pulse_d = self.var("strm_f2g_start_pulse_d",
self._params.num_glb_tiles)
self.wire(self.strm_f2g_start_pulse, self.strm_f2g_start_pulse_w)
self.pcfg_start_pulse_w = self.var("pcfg_start_pulse_w",
self._params.num_glb_tiles)
self.pcfg_start_pulse_d = self.var("pcfg_start_pulse_d",
self._params.num_glb_tiles)
self.wire(self.pcfg_start_pulse, self.pcfg_start_pulse_w)
self.strm_f2g_interrupt_pulse_w = self.var(
"strm_f2g_interrupt_pulse_w", self._params.num_glb_tiles)
self.strm_f2g_interrupt_pulse_d = self.var(
"strm_f2g_interrupt_pulse_d", self._params.num_glb_tiles)
self.wire(self.strm_f2g_interrupt_pulse_d,
self.strm_f2g_interrupt_pulse)
self.strm_g2f_interrupt_pulse_w = self.var(
"strm_g2f_interrupt_pulse_w", self._params.num_glb_tiles)
self.strm_g2f_interrupt_pulse_d = self.var(
"strm_g2f_interrupt_pulse_d", self._params.num_glb_tiles)
self.wire(self.strm_g2f_interrupt_pulse_d,
self.strm_g2f_interrupt_pulse)
self.pcfg_g2f_interrupt_pulse_w = self.var(
"pcfg_g2f_interrupt_pulse_w", self._params.num_glb_tiles)
self.pcfg_g2f_interrupt_pulse_d = self.var(
"pcfg_g2f_interrupt_pulse_d", self._params.num_glb_tiles)
self.wire(self.pcfg_g2f_interrupt_pulse_d,
self.pcfg_g2f_interrupt_pulse)
self.cgra_cfg_g2f_cfg_wr_en_w = self.var(
"cgra_cfg_g2f_cfg_wr_en_w",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_wr_en_d = self.var(
"cgra_cfg_g2f_cfg_wr_en_d",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.cgra_cfg_g2f_cfg_wr_en_d, self.cgra_cfg_g2f_cfg_wr_en)
self.cgra_cfg_g2f_cfg_rd_en_w = self.var(
"cgra_cfg_g2f_cfg_rd_en_w",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_rd_en_d = self.var(
"cgra_cfg_g2f_cfg_rd_en_d",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.cgra_cfg_g2f_cfg_rd_en_d, self.cgra_cfg_g2f_cfg_rd_en)
self.cgra_cfg_g2f_cfg_addr_w = self.var(
"cgra_cfg_g2f_cfg_addr_w",
self._params.cgra_cfg_addr_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_addr_d = self.var(
"cgra_cfg_g2f_cfg_addr_d",
self._params.cgra_cfg_addr_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.cgra_cfg_g2f_cfg_addr_d, self.cgra_cfg_g2f_cfg_addr)
self.cgra_cfg_g2f_cfg_data_w = self.var(
"cgra_cfg_g2f_cfg_data_w",
self._params.cgra_cfg_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.cgra_cfg_g2f_cfg_data_d = self.var(
"cgra_cfg_g2f_cfg_data_d",
self._params.cgra_cfg_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.cgra_cfg_g2f_cfg_data_d, self.cgra_cfg_g2f_cfg_data)
self.stream_data_f2g_w = self.var(
"stream_data_f2g_w",
self._params.cgra_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.stream_data_f2g_d = self.var(
"stream_data_f2g_d",
self._params.cgra_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.stream_data_f2g, self.stream_data_f2g_w)
self.stream_data_valid_f2g_w = self.var(
"stream_data_valid_f2g_w",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.stream_data_valid_f2g_d = self.var(
"stream_data_valid_f2g_d",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.stream_data_valid_f2g, self.stream_data_valid_f2g_w)
self.stream_data_g2f_w = self.var(
"stream_data_g2f_w",
self._params.cgra_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.stream_data_g2f_d = self.var(
"stream_data_g2f_d",
self._params.cgra_data_width,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.stream_data_g2f_d, self.stream_data_g2f)
self.stream_data_valid_g2f_w = self.var(
"stream_data_valid_g2f_w",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.stream_data_valid_g2f_d = self.var(
"stream_data_valid_g2f_d",
1,
size=[self._params.num_glb_tiles, self._params.cgra_per_glb],
packed=True)
self.wire(self.stream_data_valid_g2f_d, self.stream_data_valid_g2f)
# interface
if_cfg_tile2tile = GlbConfigInterface(
addr_width=self._params.axi_addr_width,
data_width=self._params.axi_data_width)
if_sram_cfg_tile2tile = GlbConfigInterface(
addr_width=self._params.glb_addr_width,
data_width=self._params.axi_data_width)
self.if_cfg_list = []
self.if_sram_cfg_list = []
for i in range(self._params.num_glb_tiles + 1):
self.if_cfg_list.append(
self.interface(if_cfg_tile2tile, f"if_cfg_tile2tile_{i}"))
self.if_sram_cfg_list.append(
self.interface(if_sram_cfg_tile2tile,
f"if_sram_cfg_tile2tile_{i}"))
self.glb_tile = []
for i in range(self._params.num_glb_tiles):
self.glb_tile.append(GlbTile(_params=self._params))
self.wire(self.if_cfg_list[-1].rd_data, 0)
self.wire(self.if_cfg_list[-1].rd_data_valid, 0)
self.wire(self.if_sram_cfg_list[-1].rd_data, 0)
self.wire(self.if_sram_cfg_list[-1].rd_data_valid, 0)
self.add_glb_tile()
self.add_always(self.left_edge_proc_ff)
self.add_always(self.left_edge_cfg_ff)
self.add_always(self.left_edge_sram_cfg_ff)
self.add_always(self.left_edge_cgra_cfg_ff)
self.tile2tile_e2w_wiring()
self.tile2tile_w2e_wiring()
self.add_always(self.tile2tile_w2e_cfg_wiring)
self.add_always(self.interrupt_pipeline)
self.add_always(self.start_pulse_pipeline)
self.add_always(self.stall_pipeline)
self.add_always(self.stream_data_pipeline)
self.add_always(self.cgra_cfg_pcfg_pipeline)
def __init__(self,
interconnect_input_ports=2,
mem_depth=32,
num_tiles=1,
banks=1,
iterator_support=6,
address_width=5,
data_width=16,
fetch_width=16,
multiwrite=1,
strg_wr_ports=2,
config_width=16):
super().__init__("input_addr_ctrl", debug=True)
assert multiwrite >= 1, "Multiwrite must be at least 1..."
self.interconnect_input_ports = interconnect_input_ports
self.mem_depth = mem_depth
self.num_tiles = num_tiles
self.banks = banks
self.iterator_support = iterator_support
self.address_width = address_width
self.port_sched_width = max(1, clog2(self.interconnect_input_ports))
self.data_width = data_width
self.fetch_width = fetch_width
self.fw_int = int(self.fetch_width / self.data_width)
self.multiwrite = multiwrite
self.strg_wr_ports = strg_wr_ports
self.config_width = config_width
self.mem_addr_width = clog2(self.num_tiles * self.mem_depth)
if self.banks > 1:
self.bank_addr_width = clog2(self.banks)
else:
self.bank_addr_width = 0
self.address_width = self.mem_addr_width + self.bank_addr_width
# Clock and Reset
self._clk = self.clock("clk")
self._rst_n = self.reset("rst_n")
# Inputs
# phases = [] TODO
# Take in the valid and data and attach an address + direct to a port
self._valid_in = self.input("valid_in", self.interconnect_input_ports)
self._wen_en = self.input("wen_en", self.interconnect_input_ports)
self._wen_en_saved = self.var("wen_en_saved",
self.interconnect_input_ports)
self._data_in = self.input("data_in",
self.data_width,
size=(self.interconnect_input_ports,
self.fw_int),
explicit_array=True,
packed=True)
self._data_in_saved = self.var("data_in_saved",
self.data_width,
size=(self.interconnect_input_ports,
self.fw_int),
explicit_array=True,
packed=True)
# Outputs
self._wen = self.output("wen_to_sram",
self.strg_wr_ports,
size=self.banks,
explicit_array=True,
packed=True)
wen_full_size = (self.interconnect_input_ports, self.multiwrite)
self._wen_full = self.var("wen_full",
self.banks,
size=wen_full_size,
explicit_array=True,
packed=True)
self._wen_reduced = self.var("wen_reduced",
self.banks,
size=self.interconnect_input_ports,
explicit_array=True,
packed=True)
self._wen_reduced_saved = self.var("wen_reduced_saved",
self.banks,
size=self.interconnect_input_ports,
explicit_array=True,
packed=True)
self._addresses = self.output("addr_out",
self.mem_addr_width,
size=(self.banks, self.strg_wr_ports),
explicit_array=True,
packed=True)
self._data_out = self.output("data_out",
self.data_width,
size=(self.banks, self.strg_wr_ports,
self.fw_int),
explicit_array=True,
packed=True)
self._port_out_exp = self.var("port_out_exp",
self.interconnect_input_ports,
size=self.banks,
explicit_array=True,
packed=True)
self._port_out = self.output("port_out", self.interconnect_input_ports)
self._counter = self.var("counter", self.port_sched_width)
# Wire to port out
for i in range(self.interconnect_input_ports):
new_tmp = []
for j in range(self.banks):
new_tmp.append(self._port_out_exp[j][i])
self.wire(self._port_out[i], kts.concat(*new_tmp).r_or())
self._done = self.var("done",
self.strg_wr_ports,
size=self.banks,
explicit_array=True,
packed=True)
# LOCAL VARS
self._local_addrs = self.var("local_addrs",
self.address_width,
size=(self.interconnect_input_ports,
self.multiwrite),
packed=True,
explicit_array=True)
self._local_addrs_saved = self.var("local_addrs_saved",
self.address_width,
size=(self.interconnect_input_ports,
self.multiwrite),
packed=True,
explicit_array=True)
for i in range(self.interconnect_input_ports):
for j in range(self.banks):
concat_ports = []
for k in range(self.multiwrite):
concat_ports.append(self._wen_full[i][k][j])
self.wire(self._wen_reduced[i][j],
kts.concat(*concat_ports).r_or())
if self.banks == 1 and self.interconnect_input_ports == 1:
self.wire(self._wen_full[0][0][0], self._valid_in)
elif self.banks == 1 and self.interconnect_input_ports > 1:
self.add_code(self.set_wen_single)
else:
self.add_code(self.set_wen_mult)
# MAIN
# Iterate through all banks to priority decode the wen
self.add_code(self.decode_out_lowest)
# Also set the write ports on the storage
if self.strg_wr_ports > 1:
self._idx_cnt = self.var("idx_cnt",
8,
size=(self.banks, self.strg_wr_ports - 1),
explicit_array=True,
packed=True)
for i in range(self.strg_wr_ports - 1):
self.add_code(self.decode_out_alt, idx=i + 1)
# Now we should instantiate the child address generators
# (1 per input port) to send to the sram banks
for i in range(self.interconnect_input_ports):
self.add_child(f"address_gen_{i}",
AddrGen(iterator_support=self.iterator_support,
config_width=self.config_width),
clk=self._clk,
rst_n=self._rst_n,
clk_en=const(1, 1),
flush=const(0, 1),
step=self._valid_in[i])
# Need to check that the address falls into the bank for implicit banking
# Then, obey the input schedule to send the proper Aggregator to the output
# The wen to sram should be that the valid for the selected port is high
# Do the same thing for the output address
assert self.multiwrite <= self.banks and self.multiwrite > 0,\
"Multiwrite should be between 1 and banks"
if self.multiwrite > 1:
size = (self.interconnect_input_ports, self.multiwrite - 1)
self._offsets_cfg = self.input("offsets_cfg",
self.address_width,
size=size,
packed=True,
explicit_array=True)
doc = "These offsets provide the ability to write to multiple banks explicitly"
self._offsets_cfg.add_attribute(ConfigRegAttr(doc))
self.add_code(self.set_multiwrite_addrs)
# to handle multiple input ports going to fewer SRAM write ports
self.add_code(self.set_int_ports_counter)
self.add_code(self.save_mult_int_signals)