from soc.fu.compunits.compunits import AllFunctionUnits
from soc.regfile.regfiles import RegFiles
-from openpower.decoder.decode2execute1 import Decode2ToExecute1Type
-from openpower.decoder.decode2execute1 import IssuerDecode2ToOperand
from openpower.decoder.power_decoder2 import get_rdflags
-from openpower.decoder.decode2execute1 import Data
from soc.experiment.l0_cache import TstL0CacheBuffer # test only
from soc.config.test.test_loadstore import TestMemPspec
-from openpower.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
return res # enumerate(res)
-class CoreInput:
- def __init__(self, pspec, svp64_en, regreduce_en):
- self.pspec = pspec
- self.svp64_en = svp64_en
- self.e = Decode2ToExecute1Type("core", opkls=IssuerDecode2ToOperand,
- regreduce_en=regreduce_en)
-
- # SVP64 RA_OR_ZERO needs to know if the relevant EXTRA2/3 field is zero
- self.sv_a_nz = Signal()
-
- # state and raw instruction (and SVP64 ReMap fields)
- self.state = CoreState("core")
- self.raw_insn_i = Signal(32) # raw instruction
- self.bigendian_i = Signal() # bigendian - TODO, set by MSR.BE
- if svp64_en:
- self.sv_rm = SVP64Rec(name="core_svp64_rm") # SVP64 RM field
- self.is_svp64_mode = Signal() # set if SVP64 mode is enabled
- self.use_svp64_ldst_dec = Signal() # use alternative LDST decoder
- self.sv_pred_sm = Signal() # TODO: SIMD width
- self.sv_pred_dm = Signal() # TODO: SIMD width
-
- def eq(self, i):
- self.e.eq(i.e)
- self.sv_a_nz.eq(i.sv_a_nz)
- self.state.eq(i.state)
- self.raw_insn_i.eq(i.raw_insn_i)
- self.bigendian_i.eq(i.bigendian_i)
- if not self.svp64_en:
- return
- self.sv_rm.eq(i.sv_rm)
- self.is_svp64_mode.eq(i.is_svp64_mode)
- self.use_svp64_ldst_dec.eq(i.use_svp64_ldst_dec)
- self.sv_pred_sm.eq(i.sv_pred_sm)
- self.sv_pred_dm.eq(i.sv_pred_dm)
-
-
-class CoreOutput:
- def __init__(self):
- # start/stop and terminated signalling
- self.core_terminate_o = Signal(reset=0) # indicates stopped
- self.exc_happened = Signal() # exception happened
-
- def eq(self, i):
- self.core_terminate_o.eq(i.core_terminate_o)
- self.exc_happened.eq(i.exc_happened)
-
-
# derive from ControlBase rather than have a separate Stage instance,
# this is simpler to do
class NonProductionCore(ControlBase):
self.regreduce_en = (hasattr(pspec, "regreduce") and
(pspec.regreduce == True))
+ # test core type
+ self.make_hazard_vecs = True
+ 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
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:
+ print ("core lsmem.lsi", l0.cmpi.lsmem.lsi)
mmu.alu.set_ldst_interface(l0.cmpi.lsmem.lsi)
# register files (yes plural)
- self.regs = RegFiles(pspec)
+ self.regs = RegFiles(pspec, make_hazard_vecs=self.make_hazard_vecs)
# 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,
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)
+ # create per-FU instruction decoders (subsetted). these "satellite"
+ # decoders reduce wire fan-out from the one (main) PowerDecoder2
+ # (used directly by the trap unit) to the *twelve* (or more)
+ # Function Units. we can either have 32 wires (the instruction)
+ # to each, or we can have well over a 200 wire fan-out (to 12
+ # ALUs). it's an easy choice to make.
self.decoders = {}
self.des = {}
regreduce_en=self.regreduce_en)
self.des[funame] = self.decoders[funame].do
+ # share the SPR decoder with the MMU if it exists
if "mmu0" in self.decoders:
self.decoders["mmu0"].mmu0_spr_dec = self.decoders["spr0"]
+ # next 3 functions are Stage API Compliance
def setup(self, m, i):
pass
def ospec(self):
return CoreOutput()
+ # elaborate function to create HDL
def elaborate(self, platform):
m = super().elaborate(platform)
fus = self.fus.fus
# connect decoders
+ self.connect_satellite_decoders(m)
+
+ # ssh, cheat: trap uses the main decoder because of the rewriting
+ self.des[self.trapunit] = self.i.e.do
+
+ # connect up Function Units, then read/write ports
+ 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
comb += v.use_svp64_ldst_dec.eq(
self.i.use_svp64_ldst_dec)
- # ssh, cheat: trap uses the main decoder because of the rewriting
- 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)
-
- # 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_instruction(self, m):
"""connect_instruction
fus = self.fus.fus
# indicate if core is busy
- busy_o = Signal(name="corebusy_o", reset_less=True)
+ busy_o = self.o.busy_o
- # enable-signals for each FU, get one bit for each FU (by name)
+ # 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.i.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)
# operand comes from the *local* decoder
comb += fu.oper_i.eq_from(do)
comb += fu.issue_i.eq(1) # issue when input valid
- comb += busy_o.eq(fu.busy_o)
# rdmask, which is for registers, needs to come
# from the *main* decoder
rdmask = get_rdflags(self.i.e, fu)
comb += fu.rdmaskn.eq(~rdmask)
- # if instruction is busy, set busy output for core. also
- # continue to hold each fu rdmask
- for funame, fu in fus.items():
- with m.If(fu.busy_o):
- comb += busy_o.eq(fu.busy_o)
- # rdmask, which is for registers, needs to come
- # from the *main* decoder
- rdmask = get_rdflags(self.i.e, fu)
- comb += fu.rdmaskn.eq(~rdmask)
+ # 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())
- # set ready/valid signalling. if busy, means refuse incoming issue
- comb += self.p.o_ready.eq(~busy_o)
+ # 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 fu_bitdict
+ # 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
ppoffs = []
for i, fspec in enumerate(fspecs):
# get the regfile specs for this regfile port
- (rf, read, write, wid, fuspec) = fspec
+ (rf, wf, read, write, wid, fuspec) = fspec
print ("fpsec", i, fspec, len(fuspec))
ppoffs.append(pplen) # record offset for picker
pplen += len(fuspec)
rens = []
addrs = []
+ wvens = []
+
for i, fspec in enumerate(fspecs):
- (rf, read, write, wid, fuspec) = fspec
+ (rf, wf, read, write, wid, fuspec) = fspec
# connect up the FU req/go signals, and the reg-read to the FU
# and create a Read Broadcast Bus
for pi, (funame, fu, idx) in enumerate(fuspec):
self.connect_rdport(m, fu_bitdict, rdpickers, regfile,
regname, fspec)
+ def make_hazards(self, m, regfile, rfile, wvclr, wvset,
+ funame, regname, idx,
+ addr_en, wp, fu, fu_active, wrflag, write):
+ """make_hazards: a setter and a clearer for the regfile write ports
+
+ setter is at issue time (using PowerDecoder2 regfile write numbers)
+ clearer is at regfile write time (when FU has said what to write to)
+
+ there is *one* unusual case here which has to be dealt with:
+ when the Function Unit does *NOT* request a write to the regfile
+ (has its data.ok bit CLEARED). this is perfectly legitimate.
+ and a royal pain.
+ """
+ comb = m.d.comb
+ name = "%s_%s_%d" % (funame, regname, idx)
+
+ # deal with write vector clear: this kicks in when the regfile
+ # is written to, and clears the corresponding bitvector entry
+ print ("write vector", regfile, wvclr)
+ wvaddr_en = Signal(len(wvclr.wen), name="wvaddr_en_"+name)
+ if rfile.unary:
+ comb += wvaddr_en.eq(addr_en)
+ else:
+ with m.If(wp):
+ comb += wvaddr_en.eq(1<<addr_en)
+
+ # now connect up the bitvector write hazard. unlike the
+ # regfile writeports, a ONE must be written to the corresponding
+ # bit of the hazard bitvector (to indicate the existence of
+ # the hazard)
+
+ # the detection of what shall be written to is based
+ # on *issue*
+ print ("write vector (for regread)", regfile, wvset)
+ wviaddr_en = Signal(len(wvset.wen), name="wv_issue_addr_en_"+name)
+ issue_active = Signal(name="iactive_"+name)
+ comb += issue_active.eq(fu.issue_i & fu_active & wrflag)
+ with m.If(issue_active):
+ if rfile.unary:
+ comb += wviaddr_en.eq(write)
+ else:
+ comb += wviaddr_en.eq(1<<write)
+
+ return wvaddr_en, wviaddr_en
+
def connect_wrport(self, m, fu_bitdict, wrpickers, regfile, regname, fspec):
comb, sync = m.d.comb, m.d.sync
fus = self.fus.fus
regs = self.regs
- print("connect wr", regname, fspec)
rpidx = regname
# select the required write port. these are pre-defined sizes
- print(regfile, regs.rf.keys())
rfile = regs.rf[regfile.lower()]
wport = rfile.w_ports[rpidx]
+ print("connect wr", regname, "unary", rfile.unary, fspec)
+ print(regfile, regs.rf.keys())
+
+ # select the write-protection hazard vector. note that this still
+ # requires to WRITE to the hazard bitvector! read-requests need
+ # to RAISE the bitvector (set it to 1), which, duh, requires a WRITE
+ if self.make_hazard_vecs:
+ wv = regs.wv[regfile.lower()]
+ wvset = wv.w_ports["set"] # write-vec bit-level hazard ctrl
+ wvclr = wv.w_ports["clr"] # write-vec bit-level hazard ctrl
+
fspecs = fspec
if not isinstance(fspecs, list):
fspecs = [fspecs]
pplen = 0
writes = []
ppoffs = []
+ rdflags = []
+ wrflags = []
for i, fspec in enumerate(fspecs):
# get the regfile specs for this regfile port
- (rf, read, write, wid, fuspec) = fspec
- print ("fpsec", i, fspec, len(fuspec))
+ (rf, wf, read, write, wid, fuspec) = fspec
+ print ("fpsec", i, "wrflag", wf, fspec, len(fuspec))
ppoffs.append(pplen) # record offset for picker
pplen += len(fuspec)
+ name = "%s_%s_%d" % (regfile, regname, i)
+ rdflag = Signal(name="rd_flag_"+name)
+ wrflag = Signal(name="wr_flag_"+name)
+ if rf is not None:
+ comb += rdflag.eq(rf)
+ else:
+ comb += rdflag.eq(0)
+ if wf is not None:
+ comb += wrflag.eq(wf)
+ else:
+ comb += wrflag.eq(0)
+ rdflags.append(rdflag)
+ wrflags.append(wrflag)
+
# create a priority picker to manage this port
wrpickers[regfile][rpidx] = wrpick = PriorityPicker(pplen)
setattr(m.submodules, "wrpick_%s_%s" % (regfile, rpidx), wrpick)
wsigs = []
wens = []
+ wvsets = []
+ wvseten = []
+ wvclren = []
addrs = []
for i, fspec in enumerate(fspecs):
# connect up the FU req/go signals and the reg-read to the FU
# these are arbitrated by Data.ok signals
- (rf, read, write, wid, fuspec) = fspec
+ (rf, wf, read, _write, wid, fuspec) = fspec
+ wrname = "write_%s_%s_%d" % (regfile, regname, i)
+ write = Signal.like(_write, name=wrname)
+ comb += write.eq(_write)
for pi, (funame, fu, idx) in enumerate(fuspec):
pi += ppoffs[i]
# connect the regspec write "reg select" number to this port
# only if one FU actually requests (and is granted) the port
# will the write-enable be activated
- addr_en = Signal.like(write)
+ wname = "waddr_en_%s_%s_%d" % (funame, regname, idx)
+ addr_en = Signal.like(write, name=wname)
wp = Signal()
comb += wp.eq(wr_pick & wrpick.en_o)
comb += addr_en.eq(Mux(wp, write, 0))
dest.shape(), wport.i_data.shape())
wsigs.append(fu_dest_latch)
+ # now connect up the bitvector write hazard
+ if not self.make_hazard_vecs:
+ continue
+ res = self.make_hazards(m, regfile, rfile, wvclr, wvset,
+ funame, regname, idx,
+ addr_en, wp, fu, fu_active,
+ wrflags[i], write)
+ wvaddr_en, wv_issue_en = res
+ wvclren.append(wvaddr_en) # set only: no data => clear bit
+ wvseten.append(wv_issue_en) # set data same as enable
+ wvsets.append(wv_issue_en) # because enable needs a 1
+
# here is where we create the Write Broadcast Bus. simple, eh?
comb += wport.i_data.eq(ortreereduce_sig(wsigs))
if rfile.unary:
comb += wport.addr.eq(ortreereduce_sig(addrs))
comb += wport.wen.eq(ortreereduce_sig(wens))
+ # for write-vectors
+ comb += wvclr.wen.eq(ortreereduce_sig(wvclren)) # clear (regfile write)
+ comb += wvset.wen.eq(ortreereduce_sig(wvseten)) # set (issue time)
+ comb += wvset.i_data.eq(ortreereduce_sig(wvsets))
+
def connect_wrports(self, m, fu_bitdict):
"""connect write ports
print(" %d %s %s %s" % (idx, regfile, regname, str(wid)))
if readmode:
rdflag, read = regspec_decode_read(e, regfile, regname)
- write = None
+ wrport, write = None, None
else:
rdflag, read = None, None
wrport, write = regspec_decode_write(e, regfile, regname)
byregfiles_spec[regfile] = {}
if regname not in byregfiles_spec[regfile]:
byregfiles_spec[regfile][regname] = \
- (rdflag, read, write, wid, [])
+ (rdflag, wrport, read, write, wid, [])
# here we start to create "lanes"
if idx not in byregfiles[regfile]:
byregfiles[regfile][idx] = []
fuspec = (funame, fu, idx)
byregfiles[regfile][idx].append(fuspec)
- byregfiles_spec[regfile][regname][4].append(fuspec)
+ byregfiles_spec[regfile][regname][5].append(fuspec)
# ok just print that out, for convenience
for regfile, spec in byregfiles.items():
print("regfile %s ports:" % mode, regfile)
fuspecs = byregfiles_spec[regfile]
for regname, fspec in fuspecs.items():
- [rdflag, read, write, wid, fuspec] = fspec
+ [rdflag, wrflag, read, write, wid, fuspec] = fspec
print(" rf %s port %s lane: %s" % (mode, regfile, regname))
- print(" %s" % regname, wid, read, write, rdflag)
+ print(" %s" % regname, wid, read, write, rdflag, wrflag)
for (funame, fu, idx) in fuspec:
fusig = fu.src_i[idx] if readmode else fu.dest[idx]
- print(" ", funame, fu, idx, fusig)
+ print(" ", funame, fu.__class__.__name__, idx, fusig)
print()
return byregfiles, byregfiles_spec