Documented at http://libre-soc.org/3d_gpu/architecture/compunit
"""
-from nmigen.compat.sim import run_simulation, Settle
-from nmigen.cli import verilog, rtlil
-from nmigen import Module, Signal, Mux, Elaboratable, Repl, Array, Cat, Const
+from nmigen import Module, Signal, Mux, Elaboratable, Repl, Cat, Const
from nmigen.hdl.rec import (Record, DIR_FANIN, DIR_FANOUT)
from nmutil.latch import SRLatch, latchregister
from nmutil.iocontrol import RecordObject
-from soc.decoder.power_decoder2 import Data
-from soc.decoder.power_enums import InternalOp
from soc.fu.regspec import RegSpec, RegSpecALUAPI
+def find_ok(fields):
+ """find_ok helper function - finds field ending in "_ok"
+ """
+ for field_name in fields:
+ if field_name.endswith("_ok"):
+ return field_name
+ return None
+
+
def go_record(n, name):
r = Record([('go', n, DIR_FANIN),
('rel', n, DIR_FANOUT)], name=name)
r.rel.reset_less = True
return r
+
# see https://libre-soc.org/3d_gpu/architecture/regfile/ section on regspecs
class CompUnitRecord(RegSpec, RecordObject):
j = i + 1 # name numbering to match dest1/2...
name = "dest%d_o" % j
rw = self._get_dstwid(i)
+ #dreg = Data(rw, name=name) XXX ??? output needs to be a Data type?
dreg = Signal(rw, name=name, reset_less=True)
setattr(self, name, dreg)
dst.append(dreg)
# create read/write and other scoreboard signalling
self.rd = go_record(n_src, name="rd") # read in, req out
self.wr = go_record(n_dst, name="wr") # write in, req out
+ self.rdmaskn = Signal(n_src, reset_less=True) # read mask
+ self.wrmask = Signal(n_dst, reset_less=True) # write mask
self.issue_i = Signal(reset_less=True) # fn issue in
self.shadown_i = Signal(reset=1) # shadow function, defaults to ON
self.go_die_i = Signal() # go die (reset)
class MultiCompUnit(RegSpecALUAPI, Elaboratable):
- def __init__(self, rwid, alu, opsubsetkls, n_src=2, n_dst=1):
+ def __init__(self, rwid, alu, opsubsetkls, n_src=2, n_dst=1, name=None):
"""MultiCompUnit
* :rwid: width of register latches (TODO: allocate per regspec)
* :n_dst: number of destination operands
"""
RegSpecALUAPI.__init__(self, rwid, alu)
+ self.alu_name = name or "alu"
self.opsubsetkls = opsubsetkls
self.cu = cu = CompUnitRecord(opsubsetkls, rwid, n_src, n_dst)
n_src, n_dst = self.n_src, self.n_dst = cu._n_src, cu._n_dst
# more convenience names
self.rd = cu.rd
self.wr = cu.wr
+ self.rdmaskn = cu.rdmaskn
+ self.wrmask = cu.wrmask
self.go_rd_i = self.rd.go # temporary naming
self.go_wr_i = self.wr.go # temporary naming
self.rd_rel_o = self.rd.rel # temporary naming
self.data_o = self.dest[0] # Dest out
self.done_o = cu.done_o
-
def _mux_op(self, m, sl, op_is_imm, imm, i):
# select imm if opcode says so. however also change the latch
# to trigger *from* the opcode latch instead.
def elaborate(self, platform):
m = Module()
- m.submodules.alu = self.alu
+ setattr(m.submodules, self.alu_name, self.alu)
m.submodules.src_l = src_l = SRLatch(False, self.n_src, name="src")
m.submodules.opc_l = opc_l = SRLatch(sync=False, name="opc")
m.submodules.req_l = req_l = SRLatch(False, self.n_dst, name="req")
m.d.comb += all_rd.eq(self.busy_o & rok_l.q &
(((~self.rd.rel) | self.rd.go).all()))
- # write_requests all done
- # req_done works because any one of the last of the writes
- # 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.bool()))
- m.d.comb += wr_any.eq(self.wr.go.bool())
- m.d.comb += req_done.eq(rst_l.q & wr_any)
+ # generate read-done pulse
+ all_rd_dly = Signal(reset_less=True)
+ all_rd_pulse = Signal(reset_less=True)
+ m.d.sync += all_rd_dly.eq(all_rd)
+ m.d.comb += all_rd_pulse.eq(all_rd & ~all_rd_dly)
# create rising pulse from alu valid condition.
alu_done = Signal(reset_less=True)
m.d.comb += alu_pulse.eq(alu_done & ~alu_done_dly)
m.d.comb += alu_pulsem.eq(Repl(alu_pulse, self.n_dst))
+ # sigh bug where req_l gets both set and reset raised at same time
+ prev_wr_go = Signal(self.n_dst)
+ brd = Repl(self.busy_o, self.n_dst)
+ m.d.sync += prev_wr_go.eq(self.wr.go & brd)
+
+ # write_requests all done
+ # req_done works because any one of the last of the writes
+ # 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 & ~self.wrmask).bool()))
+ m.d.comb += wr_any.eq(self.wr.go.bool() | prev_wr_go.bool())
+ m.d.comb += req_done.eq(wr_any & ~self.alu.n.ready_i & \
+ ((req_l.q & self.wrmask) == 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"
+ with m.If((self.wrmask == 0) & \
+ self.alu.n.ready_i & self.alu.n.valid_o & self.busy_o):
+ m.d.comb += req_done.eq(1)
+
# shadow/go_die
reset = Signal(reset_less=True)
rst_r = Signal(reset_less=True) # reset latch off
# read-done,wr-proceed latch
m.d.comb += rok_l.s.eq(self.issue_i) # set up when issue starts
- m.d.comb += rok_l.r.eq(self.alu.p.ready_o) # off when ALU acknowledges
+ m.d.sync += rok_l.r.eq(self.alu.n.valid_o & self.busy_o) # ALU done
# wr-done, back-to-start latch
m.d.comb += rst_l.s.eq(all_rd) # set when read-phase is fully done
# 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(self.alu.n.valid_o & req_done) # reset on ALU
+ 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))
m.d.sync += src_l.r.eq(reset_r)
# dest operand latch (not using issue_i)
- m.d.comb += req_l.s.eq(alu_pulsem)
- m.d.sync += req_l.r.eq(reset_w)
+ m.d.comb += req_l.s.eq(alu_pulsem & self.wrmask)
+ m.d.comb += req_l.r.eq(reset_w | prev_wr_go)
# create a latch/register for the operand
oper_r = self.opsubsetkls(name="oper_r")
# and for each output from the ALU: capture when ALU output is valid
drl = []
+ wrok = []
for i in range(self.n_dst):
name = "data_r%d" % i
- data_r = Signal(self.cu._get_dstwid(i), name=name, reset_less=True)
- latchregister(m, self.get_out(i), data_r, alu_pulsem, name + "_l")
+ lro = self.get_out(i)
+ ok = Const(1, 1)
+ if isinstance(lro, Record):
+ 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)
+ if fname:
+ ok = data_r[fname]
+ else:
+ data_r = Signal.like(lro, name=name, reset_less=True)
+ wrok.append(ok)
+ latchregister(m, lro, data_r, alu_pulsem, name + "_l")
drl.append(data_r)
+ # ok, above we collated anything with an "ok" on the output side
+ # now actually use those to create a write-mask. this basically
+ # is now the Function Unit API tells the Comp Unit "do not request
+ # a regfile port because this particular output is not valid"
+ m.d.comb += self.wrmask.eq(Cat(*wrok))
+
# pass the operation to the ALU
m.d.comb += self.get_op().eq(oper_r)
latchregister(m, src, alusrc, latch, name="src_r%d" % i)
# -----
- # outputs
+ # ALU connection / interaction
# -----
- slg = Cat(*map(lambda x: x[3], sl)) # get req gate conditions
- # all request signals gated by busy_o. prevents picker problems
- m.d.comb += self.busy_o.eq(opc_l.q) # busy out
- bro = Repl(self.busy_o, self.n_src)
- m.d.comb += self.rd.rel.eq(src_l.q & bro & slg) # src1/src2 req rel
-
- # write-release gated by busy and by shadow
- brd = Repl(self.busy_o & self.shadown_i, self.n_dst)
- m.d.comb += self.wr.rel.eq(req_l.q & brd)
-
- # generate read-done pulse
- all_rd_dly = Signal(reset_less=True)
- all_rd_pulse = Signal(reset_less=True)
- m.d.sync += all_rd_dly.eq(all_rd)
- m.d.comb += all_rd_pulse.eq(all_rd & ~all_rd_dly)
-
# on a go_read, tell the ALU we're accepting data.
- # NOTE: this spells TROUBLE if the ALU isn't ready!
- # go_read is only valid for one clock!
m.submodules.alui_l = alui_l = SRLatch(False, name="alui")
m.d.comb += self.alu.p.valid_i.eq(alui_l.q)
- m.d.comb += alui_l.r.eq(self.alu.p.ready_o) # valid for one extra
+ m.d.sync += alui_l.r.eq(self.alu.p.ready_o & alui_l.q)
m.d.comb += alui_l.s.eq(all_rd_pulse)
# ALU output "ready" side. alu "ready" indication stays hi until
# ALU says "valid".
m.submodules.alu_l = alu_l = SRLatch(False, name="alu")
- m.d.comb += self.alu.n.ready_i.eq(alu_l.qn)
- m.d.sync += alu_l.r.eq(self.alu.n.valid_o) # valid for one extra
+ m.d.comb += self.alu.n.ready_i.eq(alu_l.q)
+ m.d.sync += alu_l.r.eq(self.alu.n.valid_o & alu_l.q)
m.d.comb += alu_l.s.eq(all_rd_pulse)
+ # -----
+ # outputs
+ # -----
+
+ slg = Cat(*map(lambda x: x[3], sl)) # get req gate conditions
+ # all request signals gated by busy_o. prevents picker problems
+ 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.eq(src_l.q & bro & slg & ~self.rdmaskn)
+
+ # 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.eq(req_l.q & brd & self.wrmask)
+
# output the data from the latch on go_write
for i in range(self.n_dst):
with m.If(self.wr.go[i]):
return list(self)
-def op_sim(dut, a, b, op, inv_a=0, imm=0, imm_ok=0, zero_a=0):
- yield dut.issue_i.eq(0)
- yield
- yield dut.src_i[0].eq(a)
- yield dut.src_i[1].eq(b)
- yield dut.oper_i.insn_type.eq(op)
- yield dut.oper_i.invert_a.eq(inv_a)
- yield dut.oper_i.imm_data.imm.eq(imm)
- yield dut.oper_i.imm_data.imm_ok.eq(imm_ok)
- yield dut.oper_i.zero_a.eq(zero_a)
- yield dut.issue_i.eq(1)
- yield
- yield dut.issue_i.eq(0)
- yield
- if not imm_ok or not zero_a:
- yield dut.rd.go.eq(0b11)
- while True:
- yield
- rd_rel_o = yield dut.rd.rel
- print ("rd_rel", rd_rel_o)
- if rd_rel_o:
- break
- yield dut.rd.go.eq(0)
- if len(dut.src_i) == 3:
- yield dut.rd.go.eq(0b100)
- while True:
- yield
- rd_rel_o = yield dut.rd.rel
- print ("rd_rel", rd_rel_o)
- if rd_rel_o:
- break
- yield dut.rd.go.eq(0)
-
- req_rel_o = yield dut.wr.rel
- result = yield dut.data_o
- print ("req_rel", req_rel_o, result)
- while True:
- req_rel_o = yield dut.wr.rel
- result = yield dut.data_o
- print ("req_rel", req_rel_o, result)
- if req_rel_o:
- break
- yield
- yield dut.wr.go[0].eq(1)
- yield
- result = yield dut.data_o
- print ("result", result)
- yield dut.wr.go[0].eq(0)
- yield
- return result
-
-
-def scoreboard_sim_dummy(dut):
- result = yield from op_sim(dut, 5, 2, InternalOp.OP_NOP, inv_a=0,
- imm=8, imm_ok=1)
- assert result == 5, result
-
- result = yield from op_sim(dut, 9, 2, InternalOp.OP_NOP, inv_a=0,
- imm=8, imm_ok=1)
- assert result == 9, result
-
-
-def scoreboard_sim(dut):
- result = yield from op_sim(dut, 5, 2, InternalOp.OP_ADD, inv_a=0,
- imm=8, imm_ok=1)
- assert result == 13
-
- result = yield from op_sim(dut, 5, 2, InternalOp.OP_ADD)
- assert result == 7
-
- result = yield from op_sim(dut, 5, 2, InternalOp.OP_ADD, inv_a=1)
- assert result == 65532
-
- result = yield from op_sim(dut, 5, 2, InternalOp.OP_ADD, zero_a=1,
- imm=8, imm_ok=1)
- assert result == 8
-
- result = yield from op_sim(dut, 5, 2, InternalOp.OP_ADD, zero_a=1)
- assert result == 2
-
-
-def test_compunit():
- from alu_hier import ALU
- from soc.fu.alu.alu_input_record import CompALUOpSubset
-
- m = Module()
- alu = ALU(16)
- dut = MultiCompUnit(16, alu, CompALUOpSubset)
- m.submodules.cu = dut
-
- vl = rtlil.convert(dut, ports=dut.ports())
- with open("test_compunit1.il", "w") as f:
- f.write(vl)
-
- run_simulation(m, scoreboard_sim(dut), vcd_name='test_compunit1.vcd')
-
-
-class CompUnitParallelTest:
- def __init__(self, dut):
- self.dut = dut
-
- # Operation cycle should not take longer than this:
- self.MAX_BUSY_WAIT = 50
-
- # Minimum duration in which issue_i will be kept inactive,
- # during which busy_o must remain low.
- self.MIN_BUSY_LOW = 5
-
- # Number of cycles to stall until the assertion of go.
- # One value, for each port. Can be zero, for no delay.
- self.RD_GO_DELAY = [0, 3]
-
- # store common data for the input operation of the processes
- # input operation:
- self.op = 0
- self.inv_a = self.zero_a = 0
- self.imm = self.imm_ok = 0
- # input data:
- self.a = self.b = 0
-
- def driver(self):
- print("Begin parallel test.")
- yield from self.operation(5, 2, InternalOp.OP_ADD, inv_a=0,
- imm=8, imm_ok=1)
-
- def operation(self, a, b, op, inv_a=0, imm=0, imm_ok=0, zero_a=0):
- # store data for the operation
- self.a = a
- self.b = b
- self.op = op
- self.inv_a = inv_a
- self.imm = imm
- self.imm_ok = imm_ok
- self.zero_a = zero_a
-
- # trigger operation cycle
- yield from self.issue()
-
- def issue(self):
- # issue_i starts inactive
- yield self.dut.issue_i.eq(0)
-
- for n in range(self.MIN_BUSY_LOW):
- yield
- # busy_o must remain inactive. It cannot rise on its own.
- busy_o = yield self.dut.busy_o
- assert not busy_o
-
- # activate issue_i to begin the operation cycle
- yield self.dut.issue_i.eq(1)
-
- # at the same time, present the operation
- yield self.dut.oper_i.insn_type.eq(self.op)
- yield self.dut.oper_i.invert_a.eq(self.inv_a)
- yield self.dut.oper_i.imm_data.imm.eq(self.imm)
- yield self.dut.oper_i.imm_data.imm_ok.eq(self.imm_ok)
- yield self.dut.oper_i.zero_a.eq(self.zero_a)
-
- # give one cycle for the CompUnit to latch the data
- yield
-
- # busy_o must keep being low in this cycle, because issue_i was
- # low on the previous cycle.
- # It cannot rise on its own.
- # Also, busy_o and issue_i must never be active at the same time, ever.
- busy_o = yield self.dut.busy_o
- assert not busy_o
-
- # Lower issue_i
- yield self.dut.issue_i.eq(0)
-
- # deactivate inputs along with issue_i, so we can be sure the data
- # was latched at the correct cycle
- yield self.dut.oper_i.insn_type.eq(0)
- yield self.dut.oper_i.invert_a.eq(0)
- yield self.dut.oper_i.imm_data.imm.eq(0)
- yield self.dut.oper_i.imm_data.imm_ok.eq(0)
- yield self.dut.oper_i.zero_a.eq(0)
- yield
-
- # wait for busy_o to lower
- # timeout after self.MAX_BUSY_WAIT cycles
- for n in range(self.MAX_BUSY_WAIT):
- # sample busy_o in the current cycle
- busy_o = yield self.dut.busy_o
- if not busy_o:
- # operation cycle ends when busy_o becomes inactive
- break
- yield
-
- # if busy_o is still active, a timeout has occurred
- # TODO: Uncomment this, once the test is complete:
- # assert not busy_o
-
- if busy_o:
- print("If you are reading this, "
- "it's because the above test failed, as expected,\n"
- "with a timeout. It must pass, once the test is complete.")
- return
-
- print("If you are reading this, "
- "it's because the above test unexpectedly passed.")
-
- def rd(self, rd_idx):
- # wait for issue_i to rise
- while True:
- issue_i = yield self.dut.issue_i
- if issue_i:
- break
- # issue_i has not risen yet, so rd must keep low
- rel = yield self.dut.rd.rel[rd_idx]
- assert not rel
- yield
-
- # we do not want rd to rise on an immediate operand
- # if it is immediate, exit the process
- # TODO: don't exit the process, monitor rd instead to ensure it
- # doesn't rise on its own
- if (self.zero_a and rd_idx == 0) or (self.imm_ok and rd_idx == 1):
- return
-
- # issue_i has risen. rel must rise on the next cycle
- rel = yield self.dut.rd.rel[rd_idx]
- assert not rel
-
- # stall for additional cycles. Check that rel doesn't fall on its own
- for n in range(self.RD_GO_DELAY[rd_idx]):
- yield
- rel = yield self.dut.rd.rel[rd_idx]
- assert rel
-
- # Before asserting "go", make sure "rel" has risen.
- # The use of Settle allows "go" to be set combinatorially,
- # rising on the same cycle as "rel".
- yield Settle()
- rel = yield self.dut.rd.rel[rd_idx]
- assert rel
-
- # assert go for one cycle
- yield self.dut.rd.go[rd_idx].eq(1)
- yield
-
- # rel must keep high, since go was inactive in the last cycle
- rel = yield self.dut.rd.rel[rd_idx]
- assert rel
-
- # finish the go one-clock pulse
- yield self.dut.rd.go[rd_idx].eq(0)
- yield
-
- # rel must have gone low in response to go being high
- # on the previous cycle
- rel = yield self.dut.rd.rel[rd_idx]
- assert not rel
-
- # TODO: also when dut.rd.go is set, put the expected value into
- # the src_i. use dut.get_in[rd_idx] to do so
-
- def wr(self, wr_idx):
- # monitor self.dut.wr.req[rd_idx] and sets dut.wr.go[idx] for one cycle
- yield
- # TODO: also when dut.wr.go is set, check the output against the
- # self.expected_o and assert. use dut.get_out(wr_idx) to do so.
-
- def run_simulation(self, vcd_name):
- run_simulation(self.dut, [self.driver(),
- self.rd(0), # one read port (a)
- self.rd(1), # one read port (b)
- self.wr(0), # one write port (o)
- ],
- vcd_name=vcd_name)
-
-
-def test_compunit_regspec3():
- from alu_hier import DummyALU
- from soc.fu.alu.alu_input_record import CompALUOpSubset
-
- inspec = [('INT', 'a', '0:15'),
- ('INT', 'b', '0:15'),
- ('INT', 'c', '0:15')]
- outspec = [('INT', 'o', '0:15'),
- ]
-
- regspec = (inspec, outspec)
-
- m = Module()
- alu = DummyALU(16)
- dut = MultiCompUnit(regspec, alu, CompALUOpSubset)
- m.submodules.cu = dut
-
- run_simulation(m, scoreboard_sim_dummy(dut),
- vcd_name='test_compunit_regspec3.vcd')
-
-
-def test_compunit_regspec1():
- from alu_hier import ALU
- from soc.fu.alu.alu_input_record import CompALUOpSubset
-
- inspec = [('INT', 'a', '0:15'),
- ('INT', 'b', '0:15')]
- outspec = [('INT', 'o', '0:15'),
- ]
-
- regspec = (inspec, outspec)
-
- m = Module()
- alu = ALU(16)
- dut = MultiCompUnit(regspec, alu, CompALUOpSubset)
- m.submodules.cu = dut
-
- vl = rtlil.convert(dut, ports=dut.ports())
- with open("test_compunit_regspec1.il", "w") as f:
- f.write(vl)
-
- run_simulation(m, scoreboard_sim(dut),
- vcd_name='test_compunit_regspec1.vcd')
-
- test = CompUnitParallelTest(dut)
- test.run_simulation("test_compunit_parallel.vcd")
-
-
-if __name__ == '__main__':
- test_compunit()
- test_compunit_regspec1()
- test_compunit_regspec3()