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 openpower.decoder.power_decoder2 import PowerDecodeSubset
+from openpower.decoder.power_regspec_map import regspec_decode_read
+from openpower.decoder.power_regspec_map import regspec_decode_write
+from openpower.sv.svp64 import SVP64Rec
from nmutil.picker import PriorityPicker
from nmutil.util import treereduce
+from nmutil.singlepipe import ControlBase
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 openpower.decoder.power_decoder2 import get_rdflags
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
+from openpower.decoder.power_enums import MicrOp, Function
+from soc.simple.core_data import CoreInput, CoreOutput
+from collections import defaultdict
import operator
from nmutil.util import rising_edge
# helper function for reducing a list of signals down to a parallel
# ORed single signal.
-def ortreereduce(tree, attr="data_o"):
+def ortreereduce(tree, attr="o_data"):
return treereduce(tree, operator.or_, lambda x: getattr(x, attr))
return res # enumerate(res)
-class NonProductionCore(Elaboratable):
+# derive from ControlBase rather than have a separate Stage instance,
+# this is simpler to do
+class NonProductionCore(ControlBase):
def __init__(self, pspec):
self.pspec = pspec
+ # test is SVP64 is to be enabled
+ self.svp64_en = hasattr(pspec, "svp64") and (pspec.svp64 == True)
+
+ # test to see if regfile ports should be reduced
+ self.regreduce_en = (hasattr(pspec, "regreduce") and
+ (pspec.regreduce == True))
+
+ # test core type
+ self.core_type = "fsm"
+ if hasattr(pspec, "core_type"):
+ self.core_type = pspec.core_type
+
+ super().__init__(stage=self)
+
# single LD/ST funnel for memory access
- self.l0 = TstL0CacheBuffer(pspec, n_units=1)
- pi = self.l0.l0.dports[0]
+ self.l0 = l0 = TstL0CacheBuffer(pspec, n_units=1)
+ pi = 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(pspec)
-
- # instruction decoder - needs a Trap-capable Record (captures EINT etc.)
- self.e = Decode2ToExecute1Type("core", opkls=IssuerDecode2ToOperand)
-
- # SVP64 RA_OR_ZERO needs to know if the relevant EXTRA2/3 field is zero
- self.sv_a_nz = Signal()
+ # link LoadStore1 into MMU
+ mmu = self.fus.get_fu('mmu0')
+ print ("core pspec", pspec.ldst_ifacetype)
+ print ("core mmu", mmu)
+ print ("core lsmem.lsi", l0.cmpi.lsmem.lsi)
+ if mmu is not None:
+ mmu.alu.set_ldst_interface(l0.cmpi.lsmem.lsi)
- # state and raw instruction
- self.state = CoreState("core")
- self.raw_insn_i = Signal(32) # raw instruction
- self.bigendian_i = Signal() # bigendian - TODO, set by MSR.BE
-
- # issue/valid/busy signalling
- self.ivalid_i = Signal(reset_less=True) # instruction is valid
- self.issue_i = Signal(reset_less=True)
- self.busy_o = Signal(name="corebusy_o", reset_less=True)
+ # register files (yes plural)
+ self.regs = RegFiles(pspec, make_hazard_vecs=True)
- # start/stop and terminated signalling
- self.core_stopped_i = Signal(reset_less=True)
- self.core_terminate_o = Signal(reset=0) # indicates stopped
+ # set up input and output: unusual requirement to set data directly
+ # (due to the way that the core is set up in a different domain,
+ # see TestIssuer.setup_peripherals
+ self.i, self.o = self.new_specs(None)
+ self.i, self.o = self.p.i_data, self.n.o_data
# create per-FU instruction decoders (subsetted)
self.decoders = {}
fnunit = fu.fnunit.value
opkls = fu.opsubsetkls
if f_name == 'TRAP':
+ # TRAP decoder is the *main* decoder
self.trapunit = funame
continue
self.decoders[funame] = PowerDecodeSubset(None, opkls, f_name,
final=True,
- state=self.state)
+ state=self.i.state,
+ svp64_en=self.svp64_en,
+ regreduce_en=self.regreduce_en)
self.des[funame] = self.decoders[funame].do
if "mmu0" in self.decoders:
self.decoders["mmu0"].mmu0_spr_dec = self.decoders["spr0"]
+ def setup(self, m, i):
+ pass
+
+ def ispec(self):
+ return CoreInput(self.pspec, self.svp64_en, self.regreduce_en)
+
+ def ospec(self):
+ return CoreOutput()
+
def elaborate(self, platform):
- m = Module()
+ m = super().elaborate(platform)
+
# 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
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)
- # sigh due to SVP64 RA_OR_ZERO detection connect these too
- comb += v.sv_a_nz.eq(self.sv_a_nz)
+ self.connect_satellite_decoders(m)
# ssh, cheat: trap uses the main decoder because of the rewriting
- self.des[self.trapunit] = self.e.do
+ self.des[self.trapunit] = self.i.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)
+ fu_bitdict, fu_selected = self.connect_instruction(m)
+ self.connect_rdports(m, fu_selected)
+ self.connect_wrports(m, fu_selected)
+
+ # note if an exception happened. in a pipelined or OoO design
+ # this needs to be accompanied by "shadowing" (or stalling)
+ el = []
+ for exc in self.fus.excs.values():
+ el.append(exc.happened)
+ if len(el) > 0: # at least one exception
+ comb += self.o.exc_happened.eq(Cat(*el).bool())
return m
+ def connect_satellite_decoders(self, m):
+ comb = m.d.comb
+ for k, v in self.decoders.items():
+ # connect each satellite decoder and give it the instruction.
+ # as subset decoders this massively reduces wire fanout given
+ # the large number of ALUs
+ setattr(m.submodules, "dec_%s" % v.fn_name, v)
+ comb += v.dec.raw_opcode_in.eq(self.i.raw_insn_i)
+ comb += v.dec.bigendian.eq(self.i.bigendian_i)
+ # sigh due to SVP64 RA_OR_ZERO detection connect these too
+ comb += v.sv_a_nz.eq(self.i.sv_a_nz)
+ if self.svp64_en:
+ comb += v.pred_sm.eq(self.i.sv_pred_sm)
+ comb += v.pred_dm.eq(self.i.sv_pred_dm)
+ if k != self.trapunit:
+ comb += v.sv_rm.eq(self.i.sv_rm) # pass through SVP64 ReMap
+ comb += v.is_svp64_mode.eq(self.i.is_svp64_mode)
+ # only the LDST PowerDecodeSubset *actually* needs to
+ # know to use the alternative decoder. this is all
+ # a terrible hack
+ if k.lower().startswith("ldst"):
+ comb += v.use_svp64_ldst_dec.eq(
+ self.i.use_svp64_ldst_dec)
+
def connect_instruction(self, m):
"""connect_instruction
comb, sync = m.d.comb, m.d.sync
fus = self.fus.fus
- # enable-signals for each FU, get one bit for each FU (by name)
+ # indicate if core is busy
+ busy_o = self.o.busy_o
+
+ # enable/busy-signals for each FU, get one bit for each FU (by name)
fu_enable = Signal(len(fus), reset_less=True)
+ fu_busy = Signal(len(fus), reset_less=True)
fu_bitdict = {}
+ fu_selected = {}
for i, funame in enumerate(fus.keys()):
fu_bitdict[funame] = fu_enable[i]
-
- # enable the required Function Unit based on the opcode decode
- # note: this *only* works correctly for simple core when one and
- # *only* one FU is allocated per instruction
- for funame, fu in fus.items():
- fnunit = fu.fnunit.value
- enable = Signal(name="en_%s" % funame, reset_less=True)
- comb += enable.eq((self.e.do.fn_unit & fnunit).bool())
- comb += fu_bitdict[funame].eq(enable)
+ fu_selected[funame] = fu_busy[i]
+
+ # identify function units and create a list by fnunit so that
+ # PriorityPickers can be created for selecting one of them that
+ # isn't busy at the time the incoming instruction needs passing on
+ by_fnunit = defaultdict(list)
+ for fname, member in Function.__members__.items():
+ for funame, fu in fus.items():
+ fnunit = fu.fnunit.value
+ if member.value & fnunit: # this FU handles this type of op
+ by_fnunit[fname].append((funame, fu)) # add by Function
+
+ # ok now just print out the list of FUs by Function, because we can
+ for fname, fu_list in by_fnunit.items():
+ print ("FUs by type", fname, fu_list)
+
+ # now create a PriorityPicker per FU-type such that only one
+ # non-busy FU will be picked
+ issue_pps = {}
+ fu_found = Signal() # take a note if no Function Unit was available
+ for fname, fu_list in by_fnunit.items():
+ i_pp = PriorityPicker(len(fu_list))
+ m.submodules['i_pp_%s' % fname] = i_pp
+ i_l = []
+ for i, (funame, fu) in enumerate(fu_list):
+ # match the decoded instruction (e.do.fn_unit) against the
+ # "capability" of this FU, gate that by whether that FU is
+ # busy, and drop that into the PriorityPicker.
+ # this will give us an output of the first available *non-busy*
+ # Function Unit (Reservation Statio) capable of handling this
+ # instruction.
+ fnunit = fu.fnunit.value
+ en_req = Signal(name="issue_en_%s" % funame, reset_less=True)
+ fnmatch = (self.i.e.do.fn_unit & fnunit).bool()
+ comb += en_req.eq(fnmatch & ~fu.busy_o & self.p.i_valid)
+ i_l.append(en_req) # store in list for doing the Cat-trick
+ # picker output, gated by enable: store in fu_bitdict
+ po = Signal(name="o_issue_pick_"+funame) # picker output
+ comb += po.eq(i_pp.o[i] & i_pp.en_o)
+ comb += fu_bitdict[funame].eq(po)
+ comb += fu_selected[funame].eq(fu.busy_o | po)
+ # if we don't do this, then when there are no FUs available,
+ # the "p.o_ready" signal will go back "ok we accepted this
+ # instruction" which of course isn't true.
+ comb += fu_found.eq(~fnmatch | i_pp.en_o)
+ # for each input, Cat them together and drop them into the picker
+ comb += i_pp.i.eq(Cat(*i_l))
# sigh - need a NOP counter
counter = Signal(2)
with m.If(counter != 0):
sync += counter.eq(counter - 1)
- comb += self.busy_o.eq(1)
+ comb += busy_o.eq(1)
- with m.If(self.ivalid_i): # run only when valid
- with m.Switch(self.e.do.insn_type):
+ with m.If(self.p.i_valid): # run only when valid
+ with m.Switch(self.i.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)
+ m.d.sync += self.o.core_terminate_o.eq(1)
+ # fake NOP - this isn't really used (Issuer detects NOP)
with m.Case(MicrOp.OP_NOP):
sync += counter.eq(2)
- comb += self.busy_o.eq(1)
+ comb += busy_o.eq(1)
with m.Default():
# connect up instructions. only one enabled at a time
with m.If(enable):
# 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)
+ comb += fu.issue_i.eq(1) # issue when input valid
# rdmask, which is for registers, needs to come
# from the *main* decoder
- rdmask = get_rdflags(self.e, fu)
+ rdmask = get_rdflags(self.i.e, fu)
comb += fu.rdmaskn.eq(~rdmask)
- return fu_bitdict
+ # if instruction is busy, set busy output for core.
+ busys = map(lambda fu: fu.busy_o, fus.values())
+ comb += busy_o.eq(Cat(*busys).bool())
+
+ # ready/valid signalling. if busy, means refuse incoming issue.
+ # (this is a global signal, TODO, change to one which allows
+ # overlapping instructions)
+ # also, if there was no fu found we must not send back a valid
+ # indicator. BUT, of course, when there is no instruction
+ # we must ignore the fu_found flag, otherwise o_ready will never
+ # be set when everything is idle
+ comb += self.p.o_ready.eq(fu_found | ~self.p.i_valid)
+
+ # return both the function unit "enable" dict as well as the "busy".
+ # the "busy-or-issued" can be passed in to the Read/Write port
+ # connecters to give them permission to request access to regfiles
+ return fu_bitdict, fu_selected
def connect_rdport(self, m, fu_bitdict, rdpickers, regfile, regname, fspec):
comb, sync = m.d.comb, m.d.sync
src = fu.src_i[idx]
print("reg connect widths",
regfile, regname, pi, funame,
- src.shape(), rport.data_o.shape())
+ src.shape(), rport.o_data.shape())
# all FUs connect to same port
- comb += src.eq(rport.data_o)
+ comb += src.eq(rport.o_data)
# or-reduce the muxed read signals
if rfile.unary:
# argh. an experiment to merge RA and RB in the INT regfile
# (we have too many read/write ports)
- #if regfile == 'INT':
- #fuspecs['rabc'] = [fuspecs.pop('rb')]
- #fuspecs['rabc'].append(fuspecs.pop('rc'))
- #fuspecs['rabc'].append(fuspecs.pop('ra'))
- #if regfile == 'FAST':
- # fuspecs['fast1'] = [fuspecs.pop('fast1')]
- # if 'fast2' in fuspecs:
- # fuspecs['fast1'].append(fuspecs.pop('fast2'))
+ if self.regreduce_en:
+ if regfile == 'INT':
+ fuspecs['rabc'] = [fuspecs.pop('rb')]
+ fuspecs['rabc'].append(fuspecs.pop('rc'))
+ fuspecs['rabc'].append(fuspecs.pop('ra'))
+ if regfile == 'FAST':
+ fuspecs['fast1'] = [fuspecs.pop('fast1')]
+ if 'fast2' in fuspecs:
+ fuspecs['fast1'].append(fuspecs.pop('fast2'))
+ if 'fast3' in fuspecs:
+ fuspecs['fast1'].append(fuspecs.pop('fast3'))
# for each named regfile port, connect up all FUs to that port
for (regname, fspec) in sort_fuspecs(fuspecs):
pick = fu.wr.rel_o[idx] & fu_active # & wrflag
comb += wrpick.i[pi].eq(pick)
# create a single-pulse go write from the picker output
- wr_pick = Signal()
+ wr_pick = Signal(name="wpick_%s_%s_%d" % (funame, regname, idx))
comb += wr_pick.eq(wrpick.o[pi] & wrpick.en_o)
comb += fu.go_wr_i[idx].eq(rising_edge(m, wr_pick))
# connect regfile port to input
print("reg connect widths",
regfile, regname, pi, funame,
- dest.shape(), wport.data_i.shape())
+ dest.shape(), wport.i_data.shape())
wsigs.append(fu_dest_latch)
# here is where we create the Write Broadcast Bus. simple, eh?
- comb += wport.data_i.eq(ortreereduce_sig(wsigs))
+ comb += wport.i_data.eq(ortreereduce_sig(wsigs))
if rfile.unary:
# for unary-addressed
comb += wport.wen.eq(ortreereduce_sig(wens))
fuspecs = byregfiles_wrspec[regfile]
wrpickers[regfile] = {}
- # argh, more port-merging
- if regfile == 'INT':
- fuspecs['o'] = [fuspecs.pop('o')]
- fuspecs['o'].append(fuspecs.pop('o1'))
- if regfile == 'FAST':
- fuspecs['fast1'] = [fuspecs.pop('fast1')]
- if 'fast2' in fuspecs:
- fuspecs['fast1'].append(fuspecs.pop('fast2'))
+ if self.regreduce_en:
+ # argh, more port-merging
+ if regfile == 'INT':
+ fuspecs['o'] = [fuspecs.pop('o')]
+ fuspecs['o'].append(fuspecs.pop('o1'))
+ if regfile == 'FAST':
+ fuspecs['fast1'] = [fuspecs.pop('fast1')]
+ if 'fast2' in fuspecs:
+ fuspecs['fast1'].append(fuspecs.pop('fast2'))
+ if 'fast3' in fuspecs:
+ fuspecs['fast1'].append(fuspecs.pop('fast3'))
for (regname, fspec) in sort_fuspecs(fuspecs):
self.connect_wrport(m, fu_bitdict, wrpickers,
mode = "read" if readmode else "write"
regs = self.regs
fus = self.fus.fus
- e = self.e # decoded instruction to execute
+ e = self.i.e # decoded instruction to execute
# dictionary of lists of regfile ports
byregfiles = {}
def __iter__(self):
yield from self.fus.ports()
- yield from self.e.ports()
+ yield from self.i.e.ports()
yield from self.l0.ports()
# TODO: regs