from fpbase import FPNumIn, FPNumOut, FPOp, Overflow, FPBase, FPNumBase
from fpbase import MultiShiftRMerge, Trigger
-from singlepipe import (ControlBase, StageChain, UnbufferedPipeline)
+from singlepipe import (ControlBase, StageChain, UnbufferedPipeline,
+ PassThroughStage)
from multipipe import CombMuxOutPipe
from multipipe import PriorityCombMuxInPipe
-#from fpbase import FPNumShiftMultiRight
-
-
-class FPState(FPBase):
- def __init__(self, state_from):
- self.state_from = state_from
-
- def set_inputs(self, inputs):
- self.inputs = inputs
- for k,v in inputs.items():
- setattr(self, k, v)
-
- def set_outputs(self, outputs):
- self.outputs = outputs
- for k,v in outputs.items():
- setattr(self, k, v)
-
-
-class FPGetSyncOpsMod:
- def __init__(self, width, num_ops=2):
- self.width = width
- self.num_ops = num_ops
- inops = []
- outops = []
- for i in range(num_ops):
- inops.append(Signal(width, reset_less=True))
- outops.append(Signal(width, reset_less=True))
- self.in_op = inops
- self.out_op = outops
- self.stb = Signal(num_ops)
- self.ack = Signal()
- self.ready = Signal(reset_less=True)
- self.out_decode = Signal(reset_less=True)
-
- def elaborate(self, platform):
- m = Module()
- m.d.comb += self.ready.eq(self.stb == Const(-1, (self.num_ops, False)))
- m.d.comb += self.out_decode.eq(self.ack & self.ready)
- with m.If(self.out_decode):
- for i in range(self.num_ops):
- m.d.comb += [
- self.out_op[i].eq(self.in_op[i]),
- ]
- return m
-
- def ports(self):
- return self.in_op + self.out_op + [self.stb, self.ack]
-
-
-class FPOps(Trigger):
- def __init__(self, width, num_ops):
- Trigger.__init__(self)
- self.width = width
- self.num_ops = num_ops
-
- res = []
- for i in range(num_ops):
- res.append(Signal(width))
- self.v = Array(res)
-
- def ports(self):
- res = []
- for i in range(self.num_ops):
- res.append(self.v[i])
- res.append(self.ack)
- res.append(self.stb)
- return res
-
-
-class InputGroup:
- def __init__(self, width, num_ops=2, num_rows=4):
- self.width = width
- self.num_ops = num_ops
- self.num_rows = num_rows
- self.mmax = int(log(self.num_rows) / log(2))
- self.rs = []
- self.mid = Signal(self.mmax, reset_less=True) # multiplex id
- for i in range(num_rows):
- self.rs.append(FPGetSyncOpsMod(width, num_ops))
- self.rs = Array(self.rs)
-
- self.out_op = FPOps(width, num_ops)
-
- def elaborate(self, platform):
- m = Module()
-
- pe = PriorityEncoder(self.num_rows)
- m.submodules.selector = pe
- m.submodules.out_op = self.out_op
- m.submodules += self.rs
-
- # connect priority encoder
- in_ready = []
- for i in range(self.num_rows):
- in_ready.append(self.rs[i].ready)
- m.d.comb += pe.i.eq(Cat(*in_ready))
-
- active = Signal(reset_less=True)
- out_en = Signal(reset_less=True)
- m.d.comb += active.eq(~pe.n) # encoder active
- m.d.comb += out_en.eq(active & self.out_op.trigger)
-
- # encoder active: ack relevant input, record MID, pass output
- with m.If(out_en):
- rs = self.rs[pe.o]
- m.d.sync += self.mid.eq(pe.o)
- m.d.sync += rs.ack.eq(0)
- m.d.sync += self.out_op.stb.eq(0)
- for j in range(self.num_ops):
- m.d.sync += self.out_op.v[j].eq(rs.out_op[j])
- with m.Else():
- m.d.sync += self.out_op.stb.eq(1)
- # acks all default to zero
- for i in range(self.num_rows):
- m.d.sync += self.rs[i].ack.eq(1)
-
- return m
-
- def ports(self):
- res = []
- for i in range(self.num_rows):
- inop = self.rs[i]
- res += inop.in_op + [inop.stb]
- return self.out_op.ports() + res + [self.mid]
-
-
-class FPGetOpMod:
- def __init__(self, width):
- self.in_op = FPOp(width)
- self.out_op = Signal(width)
- self.out_decode = Signal(reset_less=True)
-
- def elaborate(self, platform):
- m = Module()
- m.d.comb += self.out_decode.eq((self.in_op.ack) & (self.in_op.stb))
- m.submodules.get_op_in = self.in_op
- #m.submodules.get_op_out = self.out_op
- with m.If(self.out_decode):
- m.d.comb += [
- self.out_op.eq(self.in_op.v),
- ]
- return m
-
-
-class FPGetOp(FPState):
- """ gets operand
- """
-
- def __init__(self, in_state, out_state, in_op, width):
- FPState.__init__(self, in_state)
- self.out_state = out_state
- self.mod = FPGetOpMod(width)
- self.in_op = in_op
- self.out_op = Signal(width)
- self.out_decode = Signal(reset_less=True)
-
- def setup(self, m, in_op):
- """ links module to inputs and outputs
- """
- setattr(m.submodules, self.state_from, self.mod)
- m.d.comb += self.mod.in_op.eq(in_op)
- m.d.comb += self.out_decode.eq(self.mod.out_decode)
-
- def action(self, m):
- with m.If(self.out_decode):
- m.next = self.out_state
- m.d.sync += [
- self.in_op.ack.eq(0),
- self.out_op.eq(self.mod.out_op)
- ]
- with m.Else():
- m.d.sync += self.in_op.ack.eq(1)
-
-
-class FPNumBase2Ops:
-
- def __init__(self, width, id_wid, m_extra=True):
- self.a = FPNumBase(width, m_extra)
- self.b = FPNumBase(width, m_extra)
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.a.eq(i.a), self.b.eq(i.b), self.mid.eq(i.mid)]
-
- def ports(self):
- return [self.a, self.b, self.mid]
-
-
-class FPADDBaseData:
-
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
- self.a = Signal(width)
- self.b = Signal(width)
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.a.eq(i.a), self.b.eq(i.b), self.mid.eq(i.mid)]
-
- def ports(self):
- return [self.a, self.b, self.mid]
-
-
-class FPGet2OpMod(Trigger):
- def __init__(self, width, id_wid):
- Trigger.__init__(self)
- self.width = width
- self.id_wid = id_wid
- self.i = self.ispec()
- self.o = self.ospec()
-
- def ispec(self):
- return FPADDBaseData(self.width, self.id_wid)
-
- def ospec(self):
- return FPADDBaseData(self.width, self.id_wid)
-
- def process(self, i):
- return self.o
-
- def elaborate(self, platform):
- m = Trigger.elaborate(self, platform)
- with m.If(self.trigger):
- m.d.comb += [
- self.o.eq(self.i),
- ]
- return m
-
-
-class FPGet2Op(FPState):
- """ gets operands
- """
-
- def __init__(self, in_state, out_state, width, id_wid):
- FPState.__init__(self, in_state)
- self.out_state = out_state
- self.mod = FPGet2OpMod(width, id_wid)
- self.o = self.mod.ospec()
- self.in_stb = Signal(reset_less=True)
- self.out_ack = Signal(reset_less=True)
- self.out_decode = Signal(reset_less=True)
-
- def setup(self, m, i, in_stb, in_ack):
- """ links module to inputs and outputs
- """
- m.submodules.get_ops = self.mod
- m.d.comb += self.mod.i.eq(i)
- m.d.comb += self.mod.stb.eq(in_stb)
- m.d.comb += self.out_ack.eq(self.mod.ack)
- m.d.comb += self.out_decode.eq(self.mod.trigger)
- m.d.comb += in_ack.eq(self.mod.ack)
-
- def action(self, m):
- with m.If(self.out_decode):
- m.next = self.out_state
- m.d.sync += [
- self.mod.ack.eq(0),
- self.o.eq(self.mod.o),
- ]
- with m.Else():
- m.d.sync += self.mod.ack.eq(1)
-
-
-class FPSCData:
-
- def __init__(self, width, id_wid):
- self.a = FPNumBase(width, True)
- self.b = FPNumBase(width, True)
- self.z = FPNumOut(width, False)
- self.oz = Signal(width, reset_less=True)
- self.out_do_z = Signal(reset_less=True)
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.z.eq(i.z), self.out_do_z.eq(i.out_do_z), self.oz.eq(i.oz),
- self.a.eq(i.a), self.b.eq(i.b), self.mid.eq(i.mid)]
+from fpbase import FPState, FPID
+from fpcommon.getop import (FPGetOpMod, FPGetOp, FPNumBase2Ops, FPADDBaseData, FPGet2OpMod, FPGet2Op)
+from fpcommon.denorm import (FPSCData, FPAddDeNormMod, FPAddDeNorm)
+from fpcommon.postcalc import FPAddStage1Data
+from fpcommon.postnormalise import (FPNorm1Data, FPNorm1ModSingle,
+ FPNorm1ModMulti, FPNorm1Single, FPNorm1Multi)
+from fpcommon.roundz import (FPRoundData, FPRoundMod, FPRound)
+from fpcommon.corrections import (FPCorrectionsMod, FPCorrections)
+from fpcommon.pack import (FPPackData, FPPackMod, FPPack)
+from fpcommon.normtopack import FPNormToPack
class FPAddSpecialCasesMod:
return m
-class FPID:
- def __init__(self, id_wid):
- self.id_wid = id_wid
- if self.id_wid:
- self.in_mid = Signal(id_wid, reset_less=True)
- self.out_mid = Signal(id_wid, reset_less=True)
- else:
- self.in_mid = None
- self.out_mid = None
-
- def idsync(self, m):
- if self.id_wid is not None:
- m.d.sync += self.out_mid.eq(self.in_mid)
-
-
class FPAddSpecialCases(FPState):
""" special cases: NaNs, infs, zeros, denormalised
NOTE: some of these are unique to add. see "Special Operations"
def __init__(self, width, id_wid):
FPState.__init__(self, "special_cases")
- self.smod = FPAddSpecialCasesMod(width, id_wid)
- self.dmod = FPAddDeNormMod(width, id_wid)
+ self.width = width
+ self.id_wid = id_wid
UnbufferedPipeline.__init__(self, self) # pipe is its own stage
- self.o = self.ospec()
+ self.out = self.ospec()
def ispec(self):
- return self.smod.ispec()
+ return FPADDBaseData(self.width, self.id_wid) # SpecialCases ispec
def ospec(self):
- return self.dmod.ospec()
+ return FPSCData(self.width, self.id_wid) # DeNorm ospec
def setup(self, m, i):
""" links module to inputs and outputs
"""
- # these only needed for break-out (early-out)
- # out_z = self.smod.ospec()
- # out_do_z = Signal(reset_less=True)
- self.smod.setup(m, i)
- self.dmod.setup(m, self.smod.o)
- #m.d.comb += out_do_z.eq(self.smod.o.out_do_z)
+ smod = FPAddSpecialCasesMod(self.width, self.id_wid)
+ dmod = FPAddDeNormMod(self.width, self.id_wid)
+
+ chain = StageChain([smod, dmod])
+ chain.setup(m, i)
- # out_do_z=True, only needed for early-out (split pipeline)
- #m.d.sync += out_z.z.v.eq(self.smod.o.z.v) # only take output
- #m.d.sync += out_z.mid.eq(self.smod.o.mid) # (and mid)
+ # only needed for break-out (early-out)
+ # self.out_do_z = smod.o.out_do_z
- # out_do_z=False
- m.d.comb += self.o.eq(self.dmod.o)
+ self.o = dmod.o
def process(self, i):
return self.o
def action(self, m):
+ # for break-out (early-out)
#with m.If(self.out_do_z):
# m.next = "put_z"
#with m.Else():
+ m.d.sync += self.out.eq(self.process(None))
m.next = "align"
-class FPAddDeNormMod(FPState):
-
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
- self.i = self.ispec()
- self.o = self.ospec()
-
- def ispec(self):
- return FPSCData(self.width, self.id_wid)
-
- def ospec(self):
- return FPSCData(self.width, self.id_wid)
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- m.submodules.denormalise = self
- m.d.comb += self.i.eq(i)
-
- def elaborate(self, platform):
- m = Module()
- m.submodules.denorm_in_a = self.i.a
- m.submodules.denorm_in_b = self.i.b
- m.submodules.denorm_out_a = self.o.a
- m.submodules.denorm_out_b = self.o.b
-
- with m.If(~self.i.out_do_z):
- # XXX hmmm, don't like repeating identical code
- m.d.comb += self.o.a.eq(self.i.a)
- with m.If(self.i.a.exp_n127):
- m.d.comb += self.o.a.e.eq(self.i.a.N126) # limit a exponent
- with m.Else():
- m.d.comb += self.o.a.m[-1].eq(1) # set top mantissa bit
-
- m.d.comb += self.o.b.eq(self.i.b)
- with m.If(self.i.b.exp_n127):
- m.d.comb += self.o.b.e.eq(self.i.b.N126) # limit a exponent
- with m.Else():
- m.d.comb += self.o.b.m[-1].eq(1) # set top mantissa bit
-
- m.d.comb += self.o.mid.eq(self.i.mid)
- m.d.comb += self.o.z.eq(self.i.z)
- m.d.comb += self.o.out_do_z.eq(self.i.out_do_z)
- m.d.comb += self.o.oz.eq(self.i.oz)
-
- return m
-
-
-class FPAddDeNorm(FPState):
-
- def __init__(self, width, id_wid):
- FPState.__init__(self, "denormalise")
- self.mod = FPAddDeNormMod(width)
- self.out_a = FPNumBase(width)
- self.out_b = FPNumBase(width)
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- self.mod.setup(m, i)
-
- m.d.sync += self.out_a.eq(self.mod.out_a)
- m.d.sync += self.out_b.eq(self.mod.out_b)
-
- def action(self, m):
- # Denormalised Number checks
- m.next = "align"
-
-
class FPAddAlignMultiMod(FPState):
def __init__(self, width):
m.submodules.align = self.mod
m.d.comb += self.mod.in_a.eq(in_a)
m.d.comb += self.mod.in_b.eq(in_b)
- #m.d.comb += self.out_a.eq(self.mod.out_a)
- #m.d.comb += self.out_b.eq(self.mod.out_b)
m.d.comb += self.exp_eq.eq(self.mod.exp_eq)
m.d.sync += self.out_a.eq(self.mod.out_a)
m.d.sync += self.out_b.eq(self.mod.out_b)
def ispec(self):
return FPSCData(self.width, self.id_wid)
- #return FPNumBase2Ops(self.width, self.id_wid) # AlignSingle ispec
def ospec(self):
return FPAddStage1Data(self.width, self.id_wid) # AddStage1 ospec
chain = StageChain([mod, a0mod, a1mod])
chain.setup(m, i)
- m.d.comb += self.a1o.eq(a1mod.o)
+ self.o = a1mod.o
def process(self, i):
- return self.a1o
+ return self.o
def action(self, m):
+ m.d.sync += self.a1o.eq(self.process(None))
m.next = "normalise_1"
m.next = "add_1"
-class FPAddStage1Data:
-
- def __init__(self, width, id_wid):
- self.z = FPNumBase(width, False)
- self.out_do_z = Signal(reset_less=True)
- self.oz = Signal(width, reset_less=True)
- self.of = Overflow()
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.z.eq(i.z), self.out_do_z.eq(i.out_do_z), self.oz.eq(i.oz),
- self.of.eq(i.of), self.mid.eq(i.mid)]
-
-
-
class FPAddStage1Mod(FPState):
""" Second stage of add: preparation for normalisation.
detects when tot sum is too big (tot[27] is kinda a carry bit)
def elaborate(self, platform):
m = Module()
- #m.submodules.norm1_in_overflow = self.in_of
- #m.submodules.norm1_out_overflow = self.out_of
- #m.submodules.norm1_in_z = self.in_z
- #m.submodules.norm1_out_z = self.out_z
m.d.comb += self.o.z.eq(self.i.z)
# tot[-1] (MSB) gets set when the sum overflows. shift result down
with m.If(~self.i.out_do_z):
m.next = "normalise_1"
-class FPNormaliseModSingle:
- def __init__(self, width):
- self.width = width
- self.in_z = self.ispec()
- self.out_z = self.ospec()
+class FPPutZ(FPState):
- def ispec(self):
- return FPNumBase(self.width, False)
+ def __init__(self, state, in_z, out_z, in_mid, out_mid, to_state=None):
+ FPState.__init__(self, state)
+ if to_state is None:
+ to_state = "get_ops"
+ self.to_state = to_state
+ self.in_z = in_z
+ self.out_z = out_z
+ self.in_mid = in_mid
+ self.out_mid = out_mid
- def ospec(self):
- return FPNumBase(self.width, False)
+ def action(self, m):
+ if self.in_mid is not None:
+ m.d.sync += self.out_mid.eq(self.in_mid)
+ m.d.sync += [
+ self.out_z.z.v.eq(self.in_z)
+ ]
+ with m.If(self.out_z.z.stb & self.out_z.z.ack):
+ m.d.sync += self.out_z.z.stb.eq(0)
+ m.next = self.to_state
+ with m.Else():
+ m.d.sync += self.out_z.z.stb.eq(1)
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- m.submodules.normalise = self
- m.d.comb += self.i.eq(i)
- def elaborate(self, platform):
- m = Module()
+class FPPutZIdx(FPState):
- mwid = self.out_z.m_width+2
- pe = PriorityEncoder(mwid)
- m.submodules.norm_pe = pe
+ def __init__(self, state, in_z, out_zs, in_mid, to_state=None):
+ FPState.__init__(self, state)
+ if to_state is None:
+ to_state = "get_ops"
+ self.to_state = to_state
+ self.in_z = in_z
+ self.out_zs = out_zs
+ self.in_mid = in_mid
- m.submodules.norm1_out_z = self.out_z
- m.submodules.norm1_in_z = self.in_z
+ def action(self, m):
+ outz_stb = Signal(reset_less=True)
+ outz_ack = Signal(reset_less=True)
+ m.d.comb += [outz_stb.eq(self.out_zs[self.in_mid].stb),
+ outz_ack.eq(self.out_zs[self.in_mid].ack),
+ ]
+ m.d.sync += [
+ self.out_zs[self.in_mid].v.eq(self.in_z.v)
+ ]
+ with m.If(outz_stb & outz_ack):
+ m.d.sync += self.out_zs[self.in_mid].stb.eq(0)
+ m.next = self.to_state
+ with m.Else():
+ m.d.sync += self.out_zs[self.in_mid].stb.eq(1)
- in_z = FPNumBase(self.width, False)
- in_of = Overflow()
- m.submodules.norm1_insel_z = in_z
- m.submodules.norm1_insel_overflow = in_of
- espec = (len(in_z.e), True)
- ediff_n126 = Signal(espec, reset_less=True)
- msr = MultiShiftRMerge(mwid, espec)
- m.submodules.multishift_r = msr
-
- m.d.comb += in_z.eq(self.in_z)
- m.d.comb += in_of.eq(self.in_of)
- # initialise out from in (overridden below)
- m.d.comb += self.out_z.eq(in_z)
- m.d.comb += self.out_of.eq(in_of)
- # normalisation decrease condition
- decrease = Signal(reset_less=True)
- m.d.comb += decrease.eq(in_z.m_msbzero)
- # decrease exponent
- with m.If(decrease):
- # *sigh* not entirely obvious: count leading zeros (clz)
- # with a PriorityEncoder: to find from the MSB
- # we reverse the order of the bits.
- temp_m = Signal(mwid, reset_less=True)
- temp_s = Signal(mwid+1, reset_less=True)
- clz = Signal((len(in_z.e), True), reset_less=True)
- m.d.comb += [
- # cat round and guard bits back into the mantissa
- temp_m.eq(Cat(in_of.round_bit, in_of.guard, in_z.m)),
- pe.i.eq(temp_m[::-1]), # inverted
- clz.eq(pe.o), # count zeros from MSB down
- temp_s.eq(temp_m << clz), # shift mantissa UP
- self.out_z.e.eq(in_z.e - clz), # DECREASE exponent
- self.out_z.m.eq(temp_s[2:]), # exclude bits 0&1
- ]
-
- return m
-
-class FPNorm1Data:
-
- def __init__(self, width, id_wid):
- self.roundz = Signal(reset_less=True)
- self.z = FPNumBase(width, False)
- self.out_do_z = Signal(reset_less=True)
- self.oz = Signal(width, reset_less=True)
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.z.eq(i.z), self.out_do_z.eq(i.out_do_z), self.oz.eq(i.oz),
- self.roundz.eq(i.roundz), self.mid.eq(i.mid)]
-
-
-class FPNorm1ModSingle:
-
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
- self.i = self.ispec()
- self.o = self.ospec()
-
- def ispec(self):
- return FPAddStage1Data(self.width, self.id_wid)
-
- def ospec(self):
- return FPNorm1Data(self.width, self.id_wid)
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- m.submodules.normalise_1 = self
- m.d.comb += self.i.eq(i)
-
- def process(self, i):
- return self.o
-
- def elaborate(self, platform):
- m = Module()
-
- mwid = self.o.z.m_width+2
- pe = PriorityEncoder(mwid)
- m.submodules.norm_pe = pe
-
- of = Overflow()
- m.d.comb += self.o.roundz.eq(of.roundz)
-
- m.submodules.norm1_out_z = self.o.z
- m.submodules.norm1_out_overflow = of
- m.submodules.norm1_in_z = self.i.z
- m.submodules.norm1_in_overflow = self.i.of
-
- i = self.ispec()
- m.submodules.norm1_insel_z = i.z
- m.submodules.norm1_insel_overflow = i.of
-
- espec = (len(i.z.e), True)
- ediff_n126 = Signal(espec, reset_less=True)
- msr = MultiShiftRMerge(mwid, espec)
- m.submodules.multishift_r = msr
-
- m.d.comb += i.eq(self.i)
- # initialise out from in (overridden below)
- m.d.comb += self.o.z.eq(i.z)
- m.d.comb += of.eq(i.of)
- # normalisation increase/decrease conditions
- decrease = Signal(reset_less=True)
- increase = Signal(reset_less=True)
- m.d.comb += decrease.eq(i.z.m_msbzero & i.z.exp_gt_n126)
- m.d.comb += increase.eq(i.z.exp_lt_n126)
- # decrease exponent
- with m.If(~self.i.out_do_z):
- with m.If(decrease):
- # *sigh* not entirely obvious: count leading zeros (clz)
- # with a PriorityEncoder: to find from the MSB
- # we reverse the order of the bits.
- temp_m = Signal(mwid, reset_less=True)
- temp_s = Signal(mwid+1, reset_less=True)
- clz = Signal((len(i.z.e), True), reset_less=True)
- # make sure that the amount to decrease by does NOT
- # go below the minimum non-INF/NaN exponent
- limclz = Mux(i.z.exp_sub_n126 > pe.o, pe.o,
- i.z.exp_sub_n126)
- m.d.comb += [
- # cat round and guard bits back into the mantissa
- temp_m.eq(Cat(i.of.round_bit, i.of.guard, i.z.m)),
- pe.i.eq(temp_m[::-1]), # inverted
- clz.eq(limclz), # count zeros from MSB down
- temp_s.eq(temp_m << clz), # shift mantissa UP
- self.o.z.e.eq(i.z.e - clz), # DECREASE exponent
- self.o.z.m.eq(temp_s[2:]), # exclude bits 0&1
- of.m0.eq(temp_s[2]), # copy of mantissa[0]
- # overflow in bits 0..1: got shifted too (leave sticky)
- of.guard.eq(temp_s[1]), # guard
- of.round_bit.eq(temp_s[0]), # round
- ]
- # increase exponent
- with m.Elif(increase):
- temp_m = Signal(mwid+1, reset_less=True)
- m.d.comb += [
- temp_m.eq(Cat(i.of.sticky, i.of.round_bit, i.of.guard,
- i.z.m)),
- ediff_n126.eq(i.z.N126 - i.z.e),
- # connect multi-shifter to inp/out mantissa (and ediff)
- msr.inp.eq(temp_m),
- msr.diff.eq(ediff_n126),
- self.o.z.m.eq(msr.m[3:]),
- of.m0.eq(temp_s[3]), # copy of mantissa[0]
- # overflow in bits 0..1: got shifted too (leave sticky)
- of.guard.eq(temp_s[2]), # guard
- of.round_bit.eq(temp_s[1]), # round
- of.sticky.eq(temp_s[0]), # sticky
- self.o.z.e.eq(i.z.e + ediff_n126),
- ]
-
- m.d.comb += self.o.mid.eq(self.i.mid)
- m.d.comb += self.o.out_do_z.eq(self.i.out_do_z)
- m.d.comb += self.o.oz.eq(self.i.oz)
-
- return m
-
-
-class FPNorm1ModMulti:
-
- def __init__(self, width, single_cycle=True):
- self.width = width
- self.in_select = Signal(reset_less=True)
- self.in_z = FPNumBase(width, False)
- self.in_of = Overflow()
- self.temp_z = FPNumBase(width, False)
- self.temp_of = Overflow()
- self.out_z = FPNumBase(width, False)
- self.out_of = Overflow()
-
- def elaborate(self, platform):
- m = Module()
-
- m.submodules.norm1_out_z = self.out_z
- m.submodules.norm1_out_overflow = self.out_of
- m.submodules.norm1_temp_z = self.temp_z
- m.submodules.norm1_temp_of = self.temp_of
- m.submodules.norm1_in_z = self.in_z
- m.submodules.norm1_in_overflow = self.in_of
-
- in_z = FPNumBase(self.width, False)
- in_of = Overflow()
- m.submodules.norm1_insel_z = in_z
- m.submodules.norm1_insel_overflow = in_of
-
- # select which of temp or in z/of to use
- with m.If(self.in_select):
- m.d.comb += in_z.eq(self.in_z)
- m.d.comb += in_of.eq(self.in_of)
- with m.Else():
- m.d.comb += in_z.eq(self.temp_z)
- m.d.comb += in_of.eq(self.temp_of)
- # initialise out from in (overridden below)
- m.d.comb += self.out_z.eq(in_z)
- m.d.comb += self.out_of.eq(in_of)
- # normalisation increase/decrease conditions
- decrease = Signal(reset_less=True)
- increase = Signal(reset_less=True)
- m.d.comb += decrease.eq(in_z.m_msbzero & in_z.exp_gt_n126)
- m.d.comb += increase.eq(in_z.exp_lt_n126)
- m.d.comb += self.out_norm.eq(decrease | increase) # loop-end
- # decrease exponent
- with m.If(decrease):
- m.d.comb += [
- self.out_z.e.eq(in_z.e - 1), # DECREASE exponent
- self.out_z.m.eq(in_z.m << 1), # shift mantissa UP
- self.out_z.m[0].eq(in_of.guard), # steal guard (was tot[2])
- self.out_of.guard.eq(in_of.round_bit), # round (was tot[1])
- self.out_of.round_bit.eq(0), # reset round bit
- self.out_of.m0.eq(in_of.guard),
- ]
- # increase exponent
- with m.Elif(increase):
- m.d.comb += [
- self.out_z.e.eq(in_z.e + 1), # INCREASE exponent
- self.out_z.m.eq(in_z.m >> 1), # shift mantissa DOWN
- self.out_of.guard.eq(in_z.m[0]),
- self.out_of.m0.eq(in_z.m[1]),
- self.out_of.round_bit.eq(in_of.guard),
- self.out_of.sticky.eq(in_of.sticky | in_of.round_bit)
- ]
-
- return m
-
-
-class FPNorm1Single(FPState):
-
- def __init__(self, width, id_wid, single_cycle=True):
- FPState.__init__(self, "normalise_1")
- self.mod = FPNorm1ModSingle(width)
- self.o = self.ospec()
- self.out_z = FPNumBase(width, False)
- self.out_roundz = Signal(reset_less=True)
-
- def ispec(self):
- return self.mod.ispec()
-
- def ospec(self):
- return self.mod.ospec()
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- self.mod.setup(m, i)
-
- def action(self, m):
- m.next = "round"
-
-
-class FPNorm1Multi(FPState):
-
- def __init__(self, width, id_wid):
- FPState.__init__(self, "normalise_1")
- self.mod = FPNorm1ModMulti(width)
- self.stb = Signal(reset_less=True)
- self.ack = Signal(reset=0, reset_less=True)
- self.out_norm = Signal(reset_less=True)
- self.in_accept = Signal(reset_less=True)
- self.temp_z = FPNumBase(width)
- self.temp_of = Overflow()
- self.out_z = FPNumBase(width)
- self.out_roundz = Signal(reset_less=True)
-
- def setup(self, m, in_z, in_of, norm_stb):
- """ links module to inputs and outputs
- """
- self.mod.setup(m, in_z, in_of, norm_stb,
- self.in_accept, self.temp_z, self.temp_of,
- self.out_z, self.out_norm)
-
- m.d.comb += self.stb.eq(norm_stb)
- m.d.sync += self.ack.eq(0) # sets to zero when not in normalise_1 state
-
- def action(self, m):
- m.d.comb += self.in_accept.eq((~self.ack) & (self.stb))
- m.d.sync += self.temp_of.eq(self.mod.out_of)
- m.d.sync += self.temp_z.eq(self.out_z)
- with m.If(self.out_norm):
- with m.If(self.in_accept):
- m.d.sync += [
- self.ack.eq(1),
- ]
- with m.Else():
- m.d.sync += self.ack.eq(0)
- with m.Else():
- # normalisation not required (or done).
- m.next = "round"
- m.d.sync += self.ack.eq(1)
- m.d.sync += self.out_roundz.eq(self.mod.out_of.roundz)
-
-
-class FPNormToPack(FPState, UnbufferedPipeline):
-
- def __init__(self, width, id_wid):
- FPState.__init__(self, "normalise_1")
- self.id_wid = id_wid
- self.width = width
- UnbufferedPipeline.__init__(self, self) # pipeline is its own stage
-
- def ispec(self):
- return FPAddStage1Data(self.width, self.id_wid) # Norm1ModSingle ispec
-
- def ospec(self):
- return FPPackData(self.width, self.id_wid) # FPPackMod ospec
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
-
- # Normalisation, Rounding Corrections, Pack - in a chain
- nmod = FPNorm1ModSingle(self.width, self.id_wid)
- rmod = FPRoundMod(self.width, self.id_wid)
- cmod = FPCorrectionsMod(self.width, self.id_wid)
- pmod = FPPackMod(self.width, self.id_wid)
- chain = StageChain([nmod, rmod, cmod, pmod])
- chain.setup(m, i)
- self.out_z = pmod.ospec()
-
- m.d.comb += self.out_z.mid.eq(pmod.o.mid)
- m.d.comb += self.out_z.z.eq(pmod.o.z) # outputs packed result
-
- def process(self, i):
- return self.out_z
-
- def action(self, m):
- m.next = "pack_put_z"
-
-
-class FPRoundData:
-
- def __init__(self, width, id_wid):
- self.z = FPNumBase(width, False)
- self.out_do_z = Signal(reset_less=True)
- self.oz = Signal(width, reset_less=True)
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.z.eq(i.z), self.out_do_z.eq(i.out_do_z), self.oz.eq(i.oz),
- self.mid.eq(i.mid)]
-
-
-class FPRoundMod:
-
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
- self.i = self.ispec()
- self.out_z = self.ospec()
-
- def ispec(self):
- return FPNorm1Data(self.width, self.id_wid)
-
- def ospec(self):
- return FPRoundData(self.width, self.id_wid)
-
- def process(self, i):
- return self.out_z
-
- def setup(self, m, i):
- m.submodules.roundz = self
- m.d.comb += self.i.eq(i)
-
- def elaborate(self, platform):
- m = Module()
- m.d.comb += self.out_z.eq(self.i) # copies mid, z, out_do_z
- with m.If(~self.i.out_do_z):
- with m.If(self.i.roundz):
- m.d.comb += self.out_z.z.m.eq(self.i.z.m + 1) # mantissa up
- with m.If(self.i.z.m == self.i.z.m1s): # all 1s
- m.d.comb += self.out_z.z.e.eq(self.i.z.e + 1) # exponent up
-
- return m
-
-
-class FPRound(FPState):
-
- def __init__(self, width, id_wid):
- FPState.__init__(self, "round")
- self.mod = FPRoundMod(width)
- self.out_z = self.ospec()
-
- def ispec(self):
- return self.mod.ispec()
-
- def ospec(self):
- return self.mod.ospec()
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- self.mod.setup(m, i)
-
- self.idsync(m)
- m.d.sync += self.out_z.eq(self.mod.out_z)
- m.d.sync += self.out_z.mid.eq(self.mod.o.mid)
-
- def action(self, m):
- m.next = "corrections"
-
-
-class FPCorrectionsMod:
-
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
- self.i = self.ispec()
- self.out_z = self.ospec()
-
- def ispec(self):
- return FPRoundData(self.width, self.id_wid)
-
- def ospec(self):
- return FPRoundData(self.width, self.id_wid)
-
- def process(self, i):
- return self.out_z
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- m.submodules.corrections = self
- m.d.comb += self.i.eq(i)
-
- def elaborate(self, platform):
- m = Module()
- m.submodules.corr_in_z = self.i.z
- m.submodules.corr_out_z = self.out_z.z
- m.d.comb += self.out_z.eq(self.i) # copies mid, z, out_do_z
- with m.If(~self.i.out_do_z):
- with m.If(self.i.z.is_denormalised):
- m.d.comb += self.out_z.z.e.eq(self.i.z.N127)
- return m
-
-
-class FPCorrections(FPState):
-
- def __init__(self, width, id_wid):
- FPState.__init__(self, "corrections")
- self.mod = FPCorrectionsMod(width)
- self.out_z = self.ospec()
-
- def ispec(self):
- return self.mod.ispec()
-
- def ospec(self):
- return self.mod.ospec()
-
- def setup(self, m, in_z):
- """ links module to inputs and outputs
- """
- self.mod.setup(m, in_z)
-
- m.d.sync += self.out_z.eq(self.mod.out_z)
- m.d.sync += self.out_z.mid.eq(self.mod.o.mid)
-
- def action(self, m):
- m.next = "pack"
-
-
-class FPPackData:
-
- def __init__(self, width, id_wid):
- self.z = Signal(width, reset_less=True)
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.z.eq(i.z), self.mid.eq(i.mid)]
-
-
-class FPPackMod:
-
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
- self.i = self.ispec()
- self.o = self.ospec()
-
- def ispec(self):
- return FPRoundData(self.width, self.id_wid)
-
- def ospec(self):
- return FPPackData(self.width, self.id_wid)
-
- def process(self, i):
- return self.o
-
- def setup(self, m, in_z):
- """ links module to inputs and outputs
- """
- m.submodules.pack = self
- m.d.comb += self.i.eq(in_z)
-
- def elaborate(self, platform):
- m = Module()
- z = FPNumOut(self.width, False)
- m.submodules.pack_in_z = self.i.z
- m.submodules.pack_out_z = z
- m.d.comb += self.o.mid.eq(self.i.mid)
- with m.If(~self.i.out_do_z):
- with m.If(self.i.z.is_overflowed):
- m.d.comb += z.inf(self.i.z.s)
- with m.Else():
- m.d.comb += z.create(self.i.z.s, self.i.z.e, self.i.z.m)
- with m.Else():
- m.d.comb += z.v.eq(self.i.oz)
- m.d.comb += self.o.z.eq(z.v)
- return m
-
-
-class FPPack(FPState):
-
- def __init__(self, width, id_wid):
- FPState.__init__(self, "pack")
- self.mod = FPPackMod(width)
- self.out_z = self.ospec()
-
- def ispec(self):
- return self.mod.ispec()
-
- def ospec(self):
- return self.mod.ospec()
-
- def setup(self, m, in_z):
- """ links module to inputs and outputs
- """
- self.mod.setup(m, in_z)
-
- m.d.sync += self.out_z.v.eq(self.mod.out_z.v)
- m.d.sync += self.out_z.mid.eq(self.mod.o.mid)
-
- def action(self, m):
- m.next = "pack_put_z"
-
-
-class FPPutZ(FPState):
-
- def __init__(self, state, in_z, out_z, in_mid, out_mid, to_state=None):
- FPState.__init__(self, state)
- if to_state is None:
- to_state = "get_ops"
- self.to_state = to_state
- self.in_z = in_z
- self.out_z = out_z
- self.in_mid = in_mid
- self.out_mid = out_mid
-
- def action(self, m):
- if self.in_mid is not None:
- m.d.sync += self.out_mid.eq(self.in_mid)
- m.d.sync += [
- self.out_z.z.v.eq(self.in_z)
- ]
- with m.If(self.out_z.z.stb & self.out_z.z.ack):
- m.d.sync += self.out_z.z.stb.eq(0)
- m.next = self.to_state
- with m.Else():
- m.d.sync += self.out_z.z.stb.eq(1)
-
-
-class FPPutZIdx(FPState):
-
- def __init__(self, state, in_z, out_zs, in_mid, to_state=None):
- FPState.__init__(self, state)
- if to_state is None:
- to_state = "get_ops"
- self.to_state = to_state
- self.in_z = in_z
- self.out_zs = out_zs
- self.in_mid = in_mid
-
- def action(self, m):
- outz_stb = Signal(reset_less=True)
- outz_ack = Signal(reset_less=True)
- m.d.comb += [outz_stb.eq(self.out_zs[self.in_mid].stb),
- outz_ack.eq(self.out_zs[self.in_mid].ack),
- ]
- m.d.sync += [
- self.out_zs[self.in_mid].v.eq(self.in_z.v)
- ]
- with m.If(outz_stb & outz_ack):
- m.d.sync += self.out_zs[self.in_mid].stb.eq(0)
- m.next = self.to_state
- with m.Else():
- m.d.sync += self.out_zs[self.in_mid].stb.eq(1)
-
-class FPOpData:
- def __init__(self, width, id_wid):
- self.z = FPOp(width)
- self.mid = Signal(id_wid, reset_less=True)
+class FPOpData:
+ def __init__(self, width, id_wid):
+ self.z = FPOp(width)
+ self.mid = Signal(id_wid, reset_less=True)
def eq(self, i):
return [self.z.eq(i.z), self.mid.eq(i.mid)]
get = self.add_state(FPGet2Op("get_ops", "special_cases",
self.width))
- get.setup(m, self.i, self.in_t.stb, self.in_t.ack)
+ get.setup(m, self.i)
a = get.out_op1
b = get.out_op2
+ get.trigger_setup(m, self.in_t.stb, self.in_t.ack)
sc = self.add_state(FPAddSpecialCases(self.width, self.id_wid))
sc.setup(m, a, b, self.in_mid)
def get_compact_fragment(self, m, platform=None):
- get = self.add_state(FPGet2Op("get_ops", "special_cases",
- self.width, self.id_wid))
- get.setup(m, self.i, self.in_t.stb, self.in_t.ack)
- sc = self.add_state(FPAddSpecialCasesDeNorm(self.width, self.id_wid))
- sc.setup(m, get.o)
+ get = FPGet2Op("get_ops", "special_cases", self.width, self.id_wid)
+ sc = FPAddSpecialCasesDeNorm(self.width, self.id_wid)
+ alm = FPAddAlignSingleAdd(self.width, self.id_wid)
+ n1 = FPNormToPack(self.width, self.id_wid)
- alm = self.add_state(FPAddAlignSingleAdd(self.width, self.id_wid))
- alm.setup(m, sc.o)
+ get.trigger_setup(m, self.in_t.stb, self.in_t.ack)
- n1 = self.add_state(FPNormToPack(self.width, self.id_wid))
- n1.setup(m, alm.a1o)
+ chainlist = [get, sc, alm, n1]
+ chain = StageChain(chainlist, specallocate=True)
+ chain.setup(m, self.i)
+
+ for mod in chainlist:
+ sc = self.add_state(mod)
ppz = self.add_state(FPPutZ("pack_put_z", n1.out_z.z, self.o,
n1.out_z.mid, self.o.mid))
m.d.sync += self.out_z.stb.eq(1)
-class FPADDStageOut:
- def __init__(self, width, id_wid):
- self.z = Signal(width)
- self.mid = Signal(id_wid, reset_less=True)
-
- def eq(self, i):
- return [self.z.eq(i.z), self.mid.eq(i.mid)]
-
- def ports(self):
- return [self.z, self.mid]
-
-
-# matches the format of FPADDStageOut, allows eq function to do assignments
-class PlaceHolder: pass
-
-
-class FPAddBaseStage:
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
-
- def ispec(self):
- return FPADDBaseData(self.width, self.id_wid)
-
- def ospec(self):
- return FPADDStageOut(self.width, self.id_wid)
-
- def process(self, i):
- o = PlaceHolder()
- o.z = i.a + i.b
- o.mid = i.mid
- return o
-
-
-class FPADDBasePipe1(UnbufferedPipeline):
- def __init__(self, width, id_wid):
- stage = FPAddBaseStage(width, id_wid)
- UnbufferedPipeline.__init__(self, stage)
-
-
class FPADDBasePipe(ControlBase):
def __init__(self, width, id_wid):
ControlBase.__init__(self)
- #self.pipe1 = FPADDBasePipe1(width, id_wid)
self.pipe1 = FPAddSpecialCasesDeNorm(width, id_wid)
self.pipe2 = FPAddAlignSingleAdd(width, id_wid)
self.pipe3 = FPNormToPack(width, id_wid)
return m
-class FPAddInPassThruStage:
- def __init__(self, width, id_wid):
- self.width, self.id_wid = width, id_wid
- def ispec(self): return FPADDBaseData(self.width, self.id_wid)
- def ospec(self): return self.ispec()
- def process(self, i): return i
-
-
class FPADDInMuxPipe(PriorityCombMuxInPipe):
- def __init__(self, width, id_width, num_rows):
+ def __init__(self, width, id_wid, num_rows):
self.num_rows = num_rows
- stage = FPAddInPassThruStage(width, id_width)
+ def iospec(): return FPADDBaseData(width, id_wid)
+ stage = PassThroughStage(iospec)
PriorityCombMuxInPipe.__init__(self, stage, p_len=self.num_rows)
- def ports(self):
- res = []
- for i in range(len(self.p)):
- res += [self.p[i].i_valid, self.p[i].o_ready] + \
- self.p[i].i_data.ports()
- res += [self.n.i_ready, self.n.o_valid] + \
- self.n.o_data.ports()
- return res
-
-
-class FPAddOutPassThruStage:
- def __init__(self, width, id_wid):
- self.width, self.id_wid = width, id_wid
- def ispec(self): return FPADDStageOut(self.width, self.id_wid)
- def ospec(self): return self.ispec()
- def process(self, i): return i
-
class FPADDMuxOutPipe(CombMuxOutPipe):
def __init__(self, width, id_wid, num_rows):
self.num_rows = num_rows
- stage = FPAddOutPassThruStage(width, id_wid)
+ def iospec(): return FPPackData(width, id_wid)
+ stage = PassThroughStage(iospec)
CombMuxOutPipe.__init__(self, stage, n_len=self.num_rows)
- def ports(self):
- res = [self.p.i_valid, self.p.o_ready] + \
- self.p.i_data.ports()
- for i in range(len(self.n)):
- res += [self.n[i].i_ready, self.n[i].o_valid] + \
- self.n[i].o_data.ports()
- return res
-
class FPADDMuxInOut:
""" Reservation-Station version of FPADD pipeline.
- fan-in on
+ * fan-in on inputs (an array of FPADDBaseData: a,b,mid)
+ * 3-stage adder pipeline
+ * fan-out on outputs (an array of FPPackData: z,mid)
+
+ Fan-in and Fan-out are combinatorial.
"""
def __init__(self, width, id_wid, num_rows):
self.num_rows = num_rows
return self._ports
-class ResArray:
- def __init__(self, width, id_wid):
- self.width = width
- self.id_wid = id_wid
- res = []
- for i in range(rs_sz):
- out_z = FPOp(width)
- out_z.name = "out_z_%d" % i
- res.append(out_z)
- self.res = Array(res)
- self.in_z = FPOp(width)
- self.in_mid = Signal(self.id_wid, reset_less=True)
-
- def setup(self, m, in_z, in_mid):
- m.d.comb += [self.in_z.eq(in_z),
- self.in_mid.eq(in_mid)]
-
- def get_fragment(self, platform=None):
- """ creates the HDL code-fragment for FPAdd
- """
- m = Module()
- m.submodules.res_in_z = self.in_z
- m.submodules += self.res
-
- return m
-
- def ports(self):
- res = []
- for z in self.res:
- res += z.ports()
- return res
-
-
class FPADD(FPID):
""" FPADD: stages as follows: