from nmigen import Elaboratable, Module, Signal, ResetSignal, Cat, Mux
from nmigen.cli import rtlil
+from soc.decoder.power_decoder2 import PowerDecodeSubset
from soc.decoder.power_regspec_map import regspec_decode_read
from soc.decoder.power_regspec_map import regspec_decode_write
from soc.fu.compunits.compunits import AllFunctionUnits
from soc.regfile.regfiles import RegFiles
from soc.decoder.decode2execute1 import Decode2ToExecute1Type
+from soc.decoder.decode2execute1 import IssuerDecode2ToOperand
from soc.decoder.power_decoder2 import get_rdflags
from soc.decoder.decode2execute1 import Data
from soc.experiment.l0_cache import TstL0CacheBuffer # test only
from soc.config.test.test_loadstore import TestMemPspec
from soc.decoder.power_enums import MicrOp
+from soc.config.state import CoreState
+
import operator
from nmutil.util import rising_edge
class NonProductionCore(Elaboratable):
def __init__(self, pspec):
+ self.pspec = pspec
+
# single LD/ST funnel for memory access
self.l0 = TstL0CacheBuffer(pspec, n_units=1)
pi = self.l0.l0.dports[0]
# function units (only one each)
+ # only include mmu if enabled in pspec
self.fus = AllFunctionUnits(pspec, pilist=[pi])
# register files (yes plural)
self.regs = RegFiles()
- # instruction decoder
- self.e = Decode2ToExecute1Type() # decoded instruction
+ # instruction decoder - needs a Trap-capable Record (captures EINT etc.)
+ self.e = Decode2ToExecute1Type("core", opkls=IssuerDecode2ToOperand)
+
+ self.state = CoreState("core")
+ self.raw_insn_i = Signal(32) # raw instruction
+ self.bigendian_i = Signal() # bigendian
# issue/valid/busy signalling
self.ivalid_i = Signal(reset_less=True) # instruction is valid
# start/stop and terminated signalling
self.core_stopped_i = Signal(reset_less=True)
- self.core_reset_i = Signal()
self.core_terminate_o = Signal(reset=0) # indicates stopped
+ # create per-FU instruction decoders (subsetted)
+ self.decoders = {}
+ self.des = {}
+
+ for funame, fu in self.fus.fus.items():
+ f_name = fu.fnunit.name
+ fnunit = fu.fnunit.value
+ opkls = fu.opsubsetkls
+ if f_name == 'TRAP':
+ self.trapunit = funame
+ continue
+ self.decoders[funame] = PowerDecodeSubset(None, opkls, f_name,
+ final=True,
+ state=self.state)
+ self.des[funame] = self.decoders[funame].do
+
+ if "mmu0" in self.decoders:
+ self.decoders["mmu0"].mmu0_spr_dec = self.decoders["spr0"]
+
def elaborate(self, platform):
m = Module()
+ # for testing purposes, to cut down on build time in coriolis2
+ if hasattr(self.pspec, "nocore") and self.pspec.nocore == True:
+ x = Signal() # dummy signal
+ m.d.sync += x.eq(~x)
+ return m
+ comb = m.d.comb
m.submodules.fus = self.fus
m.submodules.l0 = l0 = self.l0
regs = self.regs
fus = self.fus.fus
+ # connect decoders
+ for k, v in self.decoders.items():
+ setattr(m.submodules, "dec_%s" % v.fn_name, v)
+ comb += v.dec.raw_opcode_in.eq(self.raw_insn_i)
+ comb += v.dec.bigendian.eq(self.bigendian_i)
+
+ # ssh, cheat: trap uses the main decoder because of the rewriting
+ self.des[self.trapunit] = self.e.do
+
# connect up Function Units, then read/write ports
fu_bitdict = self.connect_instruction(m)
self.connect_rdports(m, fu_bitdict)
self.connect_wrports(m, fu_bitdict)
- # connect up reset
- m.d.comb += ResetSignal().eq(self.core_reset_i)
-
return m
def connect_instruction(self, m):
"""
comb, sync = m.d.comb, m.d.sync
fus = self.fus.fus
- e = self.e # to execute
# enable-signals for each FU, get one bit for each FU (by name)
fu_enable = Signal(len(fus), reset_less=True)
for funame, fu in fus.items():
fnunit = fu.fnunit.value
enable = Signal(name="en_%s" % funame, reset_less=True)
- comb += enable.eq((e.do.fn_unit & fnunit).bool())
+ comb += enable.eq((self.e.do.fn_unit & fnunit).bool())
comb += fu_bitdict[funame].eq(enable)
# sigh - need a NOP counter
comb += self.busy_o.eq(1)
with m.If(self.ivalid_i): # run only when valid
- with m.Switch(e.do.insn_type):
+ with m.Switch(self.e.do.insn_type):
# check for ATTN: halt if true
with m.Case(MicrOp.OP_ATTN):
m.d.sync += self.core_terminate_o.eq(1)
with m.Default():
# connect up instructions. only one enabled at a time
for funame, fu in fus.items():
+ do = self.des[funame]
enable = fu_bitdict[funame]
# run this FunctionUnit if enabled
+ # route op, issue, busy, read flags and mask to FU
with m.If(enable):
- # route op, issue, busy, read flags and mask to FU
- comb += fu.oper_i.eq_from_execute1(e)
+ # operand comes from the *local* decoder
+ comb += fu.oper_i.eq_from(do)
+ #comb += fu.oper_i.eq_from_execute1(e)
comb += fu.issue_i.eq(self.issue_i)
comb += self.busy_o.eq(fu.busy_o)
- rdmask = get_rdflags(e, fu)
+ # rdmask, which is for registers, needs to come
+ # from the *main* decoder
+ rdmask = get_rdflags(self.e, fu)
comb += fu.rdmaskn.eq(~rdmask)
return fu_bitdict
fu_active = fu_bitdict[funame]
name = "%s_%s_%s_%i" % (regfile, rpidx, funame, pi)
addr_en = Signal.like(reads[i], name="addr_en_"+name)
- rp = Signal(name="rp_"+name)
- pick = Signal()
+ pick = Signal(name="pick_"+name) # picker input
+ rp = Signal(name="rp_"+name) # picker output
+ delay_pick = Signal(name="dp_"+name) # read-enable "underway"
- comb += pick.eq(fu.rd_rel_o[idx] & fu_active & rdflags[i])
+ # exclude any currently-enabled read-request (mask out active)
+ comb += pick.eq(fu.rd_rel_o[idx] & fu_active & rdflags[i] &
+ ~delay_pick)
comb += rdpick.i[pi].eq(pick)
- sync += fu.go_rd_i[idx].eq(rising_edge(m, rp))
+ comb += fu.go_rd_i[idx].eq(delay_pick) # pass in *delayed* pick
+
# if picked, select read-port "reg select" number to port
comb += rp.eq(rdpick.o[pi] & rdpick.en_o)
+ sync += delay_pick.eq(rp) # delayed "pick"
comb += addr_en.eq(Mux(rp, reads[i], 0))
+
+ # the read-enable happens combinatorially (see mux-bus below)
+ # but it results in the data coming out on a one-cycle delay.
if rfile.unary:
rens.append(addr_en)
else:
addrs.append(addr_en)
rens.append(rp)
- with m.If(rp):
+ # use the *delayed* pick signal to put requested data onto bus
+ with m.If(delay_pick):
# connect regfile port to input, creating fan-out Bus
src = fu.src_i[idx]
print("reg connect widths",
regfile, regname, pi, funame,
src.shape(), rport.data_o.shape())
# all FUs connect to same port
- sync += src.eq(rport.data_o)
+ comb += src.eq(rport.data_o)
# or-reduce the muxed read signals
if rfile.unary:
projects / soc.git / blobdiff