from ieee754.fpcommon.pack import FPPackData
from ieee754.fpcommon.normtopack import FPNormToPack
from ieee754.fpcommon.postcalc import FPAddStage1Data
+from ieee754.fpcommon.msbhigh import FPMSBHigh
from nmigen import Module, Signal, Elaboratable
from ieee754.pipeline import PipelineSpec
+class FPCVTIntToFloatMod(Elaboratable):
+ """ FP integer conversion: copes with 16/32/64 int to 16/32/64 fp.
+
+ self.ctx.i.op & 0x1 == 0x1 : SIGNED int
+ self.ctx.i.op & 0x1 == 0x0 : UNSIGNED int
+ """
+ def __init__(self, in_pspec, out_pspec):
+ self.in_pspec = in_pspec
+ self.out_pspec = out_pspec
+ self.i = self.ispec()
+ self.o = self.ospec()
+
+ def ispec(self):
+ return FPADDBaseData(self.in_pspec)
+
+ def ospec(self):
+ return FPAddStage1Data(self.out_pspec, e_extra=True)
+
+ def setup(self, m, i):
+ """ links module to inputs and outputs
+ """
+ m.submodules.intconvert = self
+ m.d.comb += self.i.eq(i)
+
+ def process(self, i):
+ return self.o
+
+ def elaborate(self, platform):
+ m = Module()
+
+ #m.submodules.sc_out_z = self.o.z
+
+ # decode: XXX really should move to separate stage
+ print("in_width out", self.in_pspec.width,
+ self.out_pspec.width)
+ print("a1", self.in_pspec.width)
+ z1 = self.o.z
+ print("z1", z1.width, z1.rmw, z1.e_width, z1.e_start, z1.e_end)
+
+ me = self.in_pspec.width
+ mz = self.o.z.rmw
+ ms = mz - me
+ print("ms-me", ms, me, mz)
+
+ # 3 extra bits for guard/round/sticky
+ msb = FPMSBHigh(me+3, z1.e_width)
+ m.submodules.norm_msb = msb
+
+ # signed or unsigned, use operator context
+ signed = Signal(reset_less=True)
+ m.d.comb += signed.eq(self.i.ctx.op[0])
+
+ # copy of mantissa (one less bit if signed)
+ mantissa = Signal(me, reset_less=True)
+
+ # detect signed/unsigned. key case: -ve numbers need inversion
+ # to +ve because the FP sign says if it's -ve or not.
+ with m.If(signed):
+ m.d.comb += z1.s.eq(self.i.a[-1]) # sign in top bit of a
+ with m.If(z1.s):
+ m.d.comb += mantissa.eq(-self.i.a) # invert input if sign -ve
+ with m.Else():
+ m.d.comb += mantissa.eq(self.i.a) # leave as-is
+ with m.Else():
+ m.d.comb += mantissa.eq(self.i.a) # unsigned, use full a
+ m.d.comb += z1.s.eq(0)
+
+ # set input from full INT
+ m.d.comb += msb.m_in.eq(Cat(0, 0, 0, mantissa)) # g/r/s + input
+ m.d.comb += msb.e_in.eq(me) # exp = int width
+
+ if ms < 0:
+ # larger int to smaller FP (uint32/64 -> fp16 most likely)
+ m.d.comb += z1.e.eq(msb.e_out-1)
+ m.d.comb += z1.m[ms-1:].eq(msb.m_out[-mz-1:])
+ else:
+ # smaller int to larger FP
+ m.d.comb += z1.e.eq(msb.e_out)
+ m.d.comb += z1.m[ms:].eq(msb.m_out[3:])
+ m.d.comb += z1.create(z1.s, z1.e, z1.m) # ... here
+
+ # note: post-normalisation actually appears to be capable of
+ # detecting overflow to infinity (FPPackMod). so it's ok to
+ # drop the bits into the mantissa (with a fixed exponent),
+ # do some rounding (which might result in exceeding the
+ # range of the target FP by re-increasing the exponent),
+ # and basically *not* have to do any kind of range-checking
+ # here: just set up guard/round/sticky, drop the INT into the
+ # mantissa, and away we go. XXX TODO: see if FPNormaliseMod
+ # is even necessary. it probably isn't
+
+ # initialise rounding (but only activate if needed)
+ if ms < 0:
+ # larger int to smaller FP (uint32/64 -> fp16 most likely)
+ m.d.comb += self.o.of.guard.eq(msb.m_out[-mz-2])
+ m.d.comb += self.o.of.round_bit.eq(msb.m_out[-mz-3])
+ m.d.comb += self.o.of.sticky.eq(msb.m_out[:-mz-3].bool())
+ m.d.comb += self.o.of.m0.eq(msb.m_out[-mz-1])
+ else:
+ # smaller int to larger FP
+ m.d.comb += self.o.of.guard.eq(msb.m_out[2])
+ m.d.comb += self.o.of.round_bit.eq(msb.m_out[1])
+ m.d.comb += self.o.of.sticky.eq(msb.m_out[:1].bool())
+ m.d.comb += self.o.of.m0.eq(msb.m_out[3])
+
+ # special cases active by default
+ m.d.comb += self.o.out_do_z.eq(1)
+
+ # detect zero
+ with m.If(~self.i.a.bool()):
+ m.d.comb += self.o.z.zero(0)
+ with m.Else():
+ m.d.comb += self.o.out_do_z.eq(0) # activate normalisation
+
+ # copy the context (muxid, operator)
+ m.d.comb += self.o.oz.eq(self.o.z.v)
+ m.d.comb += self.o.ctx.eq(self.i.ctx)
+
+ return m
+
+
class FPCVTUpConvertMod(Elaboratable):
""" FP up-conversion (lower to higher bitwidth)
"""
return FPADDBaseData(self.in_pspec)
def ospec(self):
- return FPAddStage1Data(self.out_pspec, e_extra=True)
+ return FPAddStage1Data(self.out_pspec, e_extra=False)
def setup(self, m, i):
""" links module to inputs and outputs
me = a1.rmw
ms = self.o.z.rmw - a1.rmw
- print("ms-me", ms, me)
-
- # intermediaries
- exp_sub_n126 = Signal((a1.e_width, True), reset_less=True)
- exp_gt127 = Signal(reset_less=True)
- # constants from z1, at the bit-width of a1.
- N126 = Const(z1.fp.N126.value, (a1.e_width, True))
- P127 = Const(z1.fp.P127.value, (a1.e_width, True))
- m.d.comb += exp_sub_n126.eq(a1.e - N126)
- m.d.comb += exp_gt127.eq(a1.e > P127)
+ print("ms-me", ms, me, self.o.z.rmw, a1.rmw)
+ # conversion can mostly be done manually...
m.d.comb += self.o.z.s.eq(a1.s)
m.d.comb += self.o.z.e.eq(a1.e)
- m.d.comb += self.o.z.m[-self.o.z.rmw-1:].eq(a1.m)
- m.d.comb += self.o.z.create(a1.s, a1.e, self.o.z.m)
+ m.d.comb += self.o.z.m[ms:].eq(a1.m)
+ m.d.comb += self.o.z.create(a1.s, a1.e, self.o.z.m) # ... here
- # if exp == top
+ # initialise rounding to all zeros (deactivate)
+ m.d.comb += self.o.of.guard.eq(0)
+ m.d.comb += self.o.of.round_bit.eq(0)
+ m.d.comb += self.o.of.sticky.eq(0)
+ m.d.comb += self.o.of.m0.eq(a1.m[0])
+
+ # most special cases active (except tiny-number normalisation, below)
+ m.d.comb += self.o.out_do_z.eq(1)
+
+ # detect NaN/Inf first
with m.If(a1.exp_128):
- m.d.comb += self.o.z.create(a1.s, self.o.z.P128, self.o.z.m)
- m.d.comb += self.o.out_do_z.eq(1)
+ with m.If(~a1.m_zero):
+ m.d.comb += self.o.z.nan(0) # RISC-V wants normalised NaN
+ with m.Else():
+ m.d.comb += self.o.z.inf(a1.s) # RISC-V wants signed INF
with m.Else():
- with m.If(a1.exp_zero):
- with m.If(a1.m[:a1.rmw].bool()):
- m.d.comb += self.o.of.guard.eq(a1.m[a1.rmw-1])
- m.d.comb += self.o.of.round_bit.eq(a1.m[a1.rmw-2])
- m.d.comb += self.o.of.sticky.eq(1)
- m.d.comb += self.o.of.m0.eq(a1.m[a1.rmw]) # bit of a1
+ with m.If(a1.exp_n127):
+ with m.If(~a1.m_zero):
+ m.d.comb += self.o.z.m[ms:].eq(Cat(0, a1.m))
+ m.d.comb += self.o.out_do_z.eq(0) # activate normalisation
with m.Else():
- m.d.comb += self.o.out_do_z.eq(1)
+ # RISC-V zero needs actual zero
+ m.d.comb += self.o.z.zero(a1.s)
# copy the context (muxid, operator)
m.d.comb += self.o.oz.eq(self.o.z.v)
return m
+class FPCVTIntToFloat(FPState):
+ """ Up-conversion
+ """
+
+ def __init__(self, in_width, out_width, id_wid):
+ FPState.__init__(self, "inttofloat")
+ self.mod = FPCVTIntToFloatMod(in_width, out_width)
+ self.out_z = self.mod.ospec()
+ self.out_do_z = Signal(reset_less=True)
+
+ def setup(self, m, i):
+ """ links module to inputs and outputs
+ """
+ self.mod.setup(m, i, self.out_do_z)
+ m.d.sync += self.out_z.v.eq(self.mod.out_z.v) # only take the output
+ m.d.sync += self.out_z.ctx.eq(self.mod.o.ctx) # (and context)
+
+ def action(self, m):
+ self.idsync(m)
+ with m.If(self.out_do_z):
+ m.next = "put_z"
+ with m.Else():
+ m.next = "denormalise"
+
+
class FPCVTUpConvert(FPState):
""" Up-conversion
"""
m.next = "denormalise"
+class FPCVTIntToFloatDeNorm(FPState, SimpleHandshake):
+ """ Upconvert
+ """
+
+ def __init__(self, in_pspec, out_pspec):
+ FPState.__init__(self, "inttofloat")
+ sc = FPCVTIntToFloatMod(in_pspec, out_pspec)
+ SimpleHandshake.__init__(self, sc)
+ self.out = self.ospec(None)
+
+
class FPCVTUpConvertDeNorm(FPState, SimpleHandshake):
""" Upconvert
"""
self.out = self.ospec(None)
+class FPCVTIntBasePipe(ControlBase):
+ def __init__(self, in_pspec, out_pspec):
+ ControlBase.__init__(self)
+ self.pipe1 = FPCVTIntToFloatDeNorm(in_pspec, out_pspec)
+ self.pipe2 = FPNormToPack(out_pspec, e_extra=True)
+
+ self._eqs = self.connect([self.pipe1, self.pipe2])
+
+ def elaborate(self, platform):
+ m = ControlBase.elaborate(self, platform)
+ m.submodules.toint = self.pipe1
+ m.submodules.normpack = self.pipe2
+ m.d.comb += self._eqs
+ return m
+
+
class FPCVTUpBasePipe(ControlBase):
def __init__(self, in_pspec, out_pspec):
ControlBase.__init__(self)
self.pipe1 = FPCVTUpConvertDeNorm(in_pspec, out_pspec)
- self.pipe2 = FPNormToPack(out_pspec, e_extra=True)
+ self.pipe2 = FPNormToPack(out_pspec, e_extra=False)
self._eqs = self.connect([self.pipe1, self.pipe2])
return m
+class FPCVTIntMuxInOut(ReservationStations):
+ """ Reservation-Station version of FPCVT int-to-float pipeline.
+
+ * fan-in on inputs (an array of FPADDBaseData: a,b,mid)
+ * 2-stage multiplier pipeline
+ * fan-out on outputs (an array of FPPackData: z,mid)
+
+ Fan-in and Fan-out are combinatorial.
+ """
+
+ def __init__(self, in_width, out_width, num_rows, op_wid=0):
+ self.op_wid = op_wid
+ self.id_wid = num_bits(in_width)
+ self.out_id_wid = num_bits(out_width)
+
+ self.in_pspec = PipelineSpec(in_width, self.id_wid, self.op_wid)
+ self.out_pspec = PipelineSpec(out_width, self.out_id_wid, op_wid)
+
+ self.alu = FPCVTIntBasePipe(self.in_pspec, self.out_pspec)
+ ReservationStations.__init__(self, num_rows)
+
+ def i_specfn(self):
+ return FPADDBaseData(self.in_pspec)
+
+ def o_specfn(self):
+ return FPPackData(self.out_pspec)
+
+
class FPCVTUpMuxInOut(ReservationStations):
""" Reservation-Station version of FPCVT up pipeline.
def o_specfn(self):
return FPPackData(self.out_pspec)
+
class FPCVTDownMuxInOut(ReservationStations):
""" Reservation-Station version of FPCVT pipeline.