# output (busy/done)
self.busy_o = Signal(name="cu_busy_o", reset_less=True) # fn busy out
self.done_o = Signal(name="cu_done_o", reset_less=True)
+ self.alu_done_o = Signal(name="cu_alu_done_o", reset_less=True)
class MultiCompUnit(RegSpecALUAPI, Elaboratable):
- def __init__(self, rwid, alu, opsubsetkls, n_src=2, n_dst=1, name=None):
+ def __init__(self, rwid, alu, opsubsetkls, n_src=2, n_dst=1, name=None,
+ sync_rw=True):
"""MultiCompUnit
* :rwid: width of register latches (TODO: allocate per regspec)
* :n_dst: number of destination operands
"""
RegSpecALUAPI.__init__(self, rwid, alu)
+ self.sync_rw = sync_rw
self.alu_name = name or "alu"
self.opsubsetkls = opsubsetkls
self.cu = cu = CompUnitRecord(opsubsetkls, rwid, n_src, n_dst,
self.wr = cu.wr
self.rdmaskn = cu.rdmaskn
self.wrmask = cu.wrmask
+ self.alu_done_o = cu.alu_done_o
self.go_rd_i = self.rd.go_i # temporary naming
self.go_wr_i = self.wr.go_i # temporary naming
self.rd_rel_o = self.rd.rel_o # temporary naming
def elaborate(self, platform):
m = Module()
+ if self.sync_rw:
+ rw_domain = m.d.sync
+ else:
+ rw_domain = m.d.comb
# add the ALU to the MultiCompUnit only if it is a "real" ALU
# see AllFunctionUnits as to why: a FunctionUnitBaseMulti
# only has one "real" ALU but multiple pseudo front-ends,
# ALU only proceeds when all src are ready. rd_rel_o is delayed
# so combine it with go_rd_i. if all bits are set we're good
all_rd = Signal(reset_less=True)
- m.d.comb += all_rd.eq(self.busy_o & rok_l.q &
+ m.d.comb += all_rd.eq(self.busy_o & # rok_l.q & # XXX LOOP
(((~self.rd.rel_o) | self.rd.go_i).all()))
# generate read-done pulse
all_rd_pulse = Signal(reset_less=True)
- m.d.comb += all_rd_pulse.eq(rising_edge(m, all_rd))
+ m.d.comb += all_rd_pulse.eq(rising_edge(m, all_rd)) # XXX LOOP
# create rising pulse from alu valid condition.
- alu_done = Signal(reset_less=True)
+ alu_done = self.cu.alu_done_o
alu_pulse = Signal(reset_less=True)
alu_pulsem = Signal(self.n_dst, reset_less=True)
m.d.comb += alu_done.eq(self.alu.n.o_valid)
# is enough, when combined with when read-phase is done (rst_l.q)
wr_any = Signal(reset_less=True)
req_done = Signal(reset_less=True)
- m.d.comb += self.done_o.eq(self.busy_o &
- ~((self.wr.rel_o & ~self.wrmask).bool()))
+ m.d.comb += self.done_o.eq(self.busy_o & ~(self.wr.rel_o).bool())
m.d.comb += wr_any.eq(self.wr.go_i.bool() | prev_wr_go.bool())
- m.d.comb += req_done.eq(wr_any & ~self.alu.n.i_ready &
- ((req_l.q & self.wrmask) == 0))
+ m.d.comb += req_done.eq(wr_any & ~self.alu.n.i_ready & (req_l.q == 0))
# argh, complicated hack: if there are no regs to write,
# instead of waiting for regs that are never going to happen,
# we indicate "done" when the ALU is "done"
m.d.comb += reset.eq(req_done | self.go_die_i)
m.d.comb += rst_r.eq(self.issue_i | self.go_die_i)
m.d.comb += reset_w.eq(self.wr.go_i | Repl(self.go_die_i, self.n_dst))
- m.d.comb += reset_r.eq(self.rd.go_i | Repl(self.go_die_i, self.n_src))
+ m.d.comb += reset_r.eq(self.rd.go_i | Repl(rst_r, self.n_src))
# read-done,wr-proceed latch
- m.d.sync += rok_l.s.eq(self.issue_i) # set up when issue starts
- m.d.sync += rok_l.r.eq(self.alu.n.o_valid & self.busy_o) # ALU done
+ rw_domain += rok_l.s.eq(self.issue_i) # set up when issue starts
+ rw_domain += rok_l.r.eq(self.alu.n.o_valid & self.busy_o) # ALUdone LOOP
# wr-done, back-to-start latch
- m.d.sync += rst_l.s.eq(all_rd) # set when read-phase is fully done
- m.d.sync += rst_l.r.eq(rst_r) # *off* on issue
+ rw_domain += rst_l.s.eq(all_rd) # set when read-phase is fully done
+ rw_domain += rst_l.r.eq(rst_r) # *off* on issue
# opcode latch (not using go_rd_i) - inverted so that busy resets to 0
m.d.sync += opc_l.s.eq(self.issue_i) # set on issue
m.d.sync += opc_l.r.eq(req_done) # reset on ALU
- # src operand latch (not using go_wr_i)
- m.d.sync += src_l.s.eq(Repl(self.issue_i, self.n_src))
+ # src operand latch (not using go_wr_i) ANDed with rdmask
+ rdmaskn = Signal(self.n_src)
+ latchregister(m, self.rdmaskn, rdmaskn, self.issue_i, name="rdmask_l")
+ m.d.sync += src_l.s.eq(Repl(self.issue_i, self.n_src) & ~rdmaskn)
m.d.sync += src_l.r.eq(reset_r)
# dest operand latch (not using issue_i)
- m.d.sync += req_l.s.eq(alu_pulsem & self.wrmask)
- m.d.sync += req_l.r.eq(reset_w | prev_wr_go)
+ rw_domain += req_l.s.eq(alu_pulsem & self.wrmask)
+ m.d.comb += req_l.r.eq(reset_w | prev_wr_go)
# pass operation to the ALU (sync: plenty time to wait for src reads)
op = self.get_op()
name = "data_r%d" % i
lro = self.get_out(i)
ok = Const(1, 1)
+ data_r_ok = Const(1, 1)
if isinstance(lro, Record):
+ print("wr fields", i, lro, lro.fields)
data_r = Record.like(lro, name=name)
- print("wr fields", i, lro, data_r.fields)
# bye-bye abstract interface design..
- fname = find_ok(data_r.fields)
+ fname = find_ok(lro.fields)
if fname:
ok = getattr(lro, fname)
+ data_r_ok = getattr(data_r, fname)
+ # write-ok based on incoming output *and* whether the latched
+ # data was ok.
+ # XXX fails - wrok.append((ok|data_r_ok) & self.busy_o)
+ wrok.append(ok & self.busy_o)
else:
- data_r = Signal.like(lro, name=name, reset_less=True)
- wrok.append(ok & self.busy_o)
- with m.If(alu_pulse):
- m.d.sync += data_r.eq(lro)
+ data_r = Signal.like(lro, name=name)
+ # really should retire this but it's part of unit tests
+ wrok.append(ok & self.busy_o)
+ #latchregister(m, lro, data_r, ok & self.busy_o, name=name)
+ latchregister(m, lro, data_r, alu_pulse, name=name)
with m.If(self.issue_i):
- m.d.sync += data_r.eq(0)
+ m.d.comb += data_r.eq(0)
drl.append(data_r)
# ok, above we collated anything with an "ok" on the output side
# create a latch/register for src1/src2 (even if it is a copy of imm)
for i in range(self.n_src):
src, alusrc, latch, _ = sl[i]
- latchregister(m, src, alusrc, latch, name="src_r%d" % i)
+ reg = latchregister(m, src, alusrc, latch, name="src_r%d" % i)
+ # rdmask stops src latches from being set. clear all if not busy
+ with m.If(~self.busy_o):
+ m.d.sync += reg.eq(0)
# -----
# ALU connection / interaction
m.submodules.alu_l = alu_l = SRLatch(False, name="alu")
m.d.comb += self.alu.n.i_ready.eq(alu_l.q)
m.d.sync += alu_l.r.eq(self.alu.n.o_valid & alu_l.q)
- m.d.comb += alu_l.s.eq(all_rd_pulse)
+ m.d.comb += alu_l.s.eq(all_rd_pulse) # XXX LOOP
# -----
# outputs
m.d.comb += self.busy_o.eq(opc_l.q) # busy out
# read-release gated by busy (and read-mask)
- bro = Repl(self.busy_o, self.n_src)
- m.d.comb += self.rd.rel_o.eq(src_l.q & bro & slg & ~self.rdmaskn)
+ if True: #self.sync_rw: - experiment (doesn't work)
+ bro = Repl(self.busy_o, self.n_src)
+ else:
+ bro = Repl(self.busy_o|self.issue_i, self.n_src)
+ m.d.comb += self.rd.rel_o.eq(src_l.q & bro & slg)
# write-release gated by busy and by shadow (and write-mask)
brd = Repl(self.busy_o & self.shadown_i, self.n_dst)
- m.d.comb += self.wr.rel_o.eq(req_l.q & brd & self.wrmask)
+ m.d.comb += self.wr.rel_o.eq(req_l.q_int & brd)
# output the data from the latch on go_write
for i in range(self.n_dst):
projects / soc.git / blobdiff