+
+class FPAddStage0(FPState):
+ """ First stage of add. covers same-sign (add) and subtract
+ special-casing when mantissas are greater or equal, to
+ give greatest accuracy.
+ """
+
+ def __init__(self, width):
+ FPState.__init__(self, "add_0")
+ self.mod = FPAddStage0Mod(width)
+ self.out_z = FPNumBase(width, False)
+ self.out_tot = Signal(self.out_z.m_width + 4, reset_less=True)
+
+ def setup(self, m, in_a, in_b):
+ """ links module to inputs and outputs
+ """
+ m.submodules.add0 = self.mod
+
+ m.d.comb += self.mod.in_a.copy(in_a)
+ m.d.comb += self.mod.in_b.copy(in_b)
+
+ def action(self, m):
+ m.next = "add_1"
+ # NOTE: these could be done as combinatorial (merge add0+add1)
+ m.d.sync += self.out_z.copy(self.mod.out_z)
+ m.d.sync += self.out_tot.eq(self.mod.out_tot)
+
+
+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 __init__(self, width):
+ self.out_norm = Signal(reset_less=True)
+ self.in_z = FPNumBase(width, False)
+ self.in_tot = Signal(self.in_z.m_width + 4, reset_less=True)
+ self.out_z = FPNumBase(width, False)
+ self.out_of = Overflow()
+
+ 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.out_z.copy(self.in_z)
+ # tot[27] gets set when the sum overflows. shift result down
+ with m.If(self.in_tot[-1]):
+ m.d.comb += [
+ self.out_z.m.eq(self.in_tot[4:]),
+ self.out_of.m0.eq(self.in_tot[4]),
+ self.out_of.guard.eq(self.in_tot[3]),
+ self.out_of.round_bit.eq(self.in_tot[2]),
+ self.out_of.sticky.eq(self.in_tot[1] | self.in_tot[0]),
+ self.out_z.e.eq(self.in_z.e + 1)
+ ]
+ # tot[27] zero case
+ with m.Else():
+ m.d.comb += [
+ self.out_z.m.eq(self.in_tot[3:]),
+ self.out_of.m0.eq(self.in_tot[3]),
+ self.out_of.guard.eq(self.in_tot[2]),
+ self.out_of.round_bit.eq(self.in_tot[1]),
+ self.out_of.sticky.eq(self.in_tot[0])
+ ]
+ return m
+
+
+class FPAddStage1(FPState):
+
+ def __init__(self, width):
+ FPState.__init__(self, "add_1")
+ self.mod = FPAddStage1Mod(width)
+ self.out_z = FPNumBase(width, False)
+ self.out_of = Overflow()
+ self.norm_stb = Signal()
+
+ def setup(self, m, in_tot, in_z):
+ """ links module to inputs and outputs
+ """
+ m.submodules.add1 = self.mod
+
+ m.d.comb += self.mod.in_z.copy(in_z)
+ m.d.comb += self.mod.in_tot.eq(in_tot)
+
+ m.d.sync += self.norm_stb.eq(0) # sets to zero when not in add1 state
+
+ def action(self, m):
+ m.submodules.add1_out_overflow = self.out_of
+ m.d.sync += self.out_of.copy(self.mod.out_of)
+ m.d.sync += self.out_z.copy(self.mod.out_z)
+ m.d.sync += self.norm_stb.eq(1)
+ m.next = "normalise_1"
+
+
+class FPNorm1ModSingle:
+
+ def __init__(self, width):
+ self.width = width
+ self.in_select = Signal(reset_less=True)
+ self.out_norm = 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()
+
+ mwid = self.out_z.m_width+2
+ pe = PriorityEncoder(mwid)
+ m.submodules.norm_pe = pe
+
+ 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
+
+ espec = (len(in_z.e), True)
+ ediff_n126 = Signal(espec, reset_less=True)
+ msr = MultiShiftRMerge(mwid, espec)
+ m.submodules.multishift_r = msr
+
+ # select which of temp or in z/of to use
+ with m.If(self.in_select):
+ m.d.comb += in_z.copy(self.in_z)
+ m.d.comb += in_of.copy(self.in_of)
+ with m.Else():
+ m.d.comb += in_z.copy(self.temp_z)
+ m.d.comb += in_of.copy(self.temp_of)
+ # initialise out from in (overridden below)
+ m.d.comb += self.out_z.copy(in_z)
+ m.d.comb += self.out_of.copy(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(0) # loop-end condition
+ # 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)
+ # make sure that the amount to decrease by does NOT
+ # go below the minimum non-INF/NaN exponent
+ limclz = Mux(in_z.exp_sub_n126 > pe.o, pe.o,
+ in_z.exp_sub_n126)
+ 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(limclz), # 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
+ self.out_of.m0.eq(temp_s[2]), # copy of mantissa[0]
+ # overflow in bits 0..1: got shifted too (leave sticky)
+ self.out_of.guard.eq(temp_s[1]), # guard
+ self.out_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(in_of.sticky, in_of.round_bit, in_of.guard,
+ in_z.m)),
+ ediff_n126.eq(in_z.N126 - in_z.e),
+ # connect multi-shifter to inp/out mantissa (and ediff)
+ msr.inp.eq(temp_m),
+ msr.diff.eq(ediff_n126),
+ self.out_z.m.eq(msr.m[3:]),
+ self.out_of.m0.eq(temp_s[3]), # copy of mantissa[0]
+ # overflow in bits 0..1: got shifted too (leave sticky)
+ self.out_of.guard.eq(temp_s[2]), # guard
+ self.out_of.round_bit.eq(temp_s[1]), # round
+ self.out_of.sticky.eq(temp_s[0]), # sticky
+ self.out_z.e.eq(in_z.e + ediff_n126),
+ ]
+
+ return m
+
+
+class FPNorm1ModMulti:
+
+ def __init__(self, width, single_cycle=True):
+ self.width = width
+ self.in_select = Signal(reset_less=True)
+ self.out_norm = 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.copy(self.in_z)
+ m.d.comb += in_of.copy(self.in_of)
+ with m.Else():
+ m.d.comb += in_z.copy(self.temp_z)
+ m.d.comb += in_of.copy(self.temp_of)
+ # initialise out from in (overridden below)
+ m.d.comb += self.out_z.copy(in_z)
+ m.d.comb += self.out_of.copy(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 FPNorm1(FPState):
+
+ def __init__(self, width, single_cycle=True):
+ FPState.__init__(self, "normalise_1")
+ if single_cycle:
+ self.mod = FPNorm1ModSingle(width)
+ else:
+ 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
+ """
+ m.submodules.normalise_1 = self.mod
+
+ m.d.comb += self.mod.in_z.copy(in_z)
+ m.d.comb += self.mod.in_of.copy(in_of)
+
+ m.d.comb += self.mod.in_select.eq(self.in_accept)
+ m.d.comb += self.mod.temp_z.copy(self.temp_z)
+ m.d.comb += self.mod.temp_of.copy(self.temp_of)
+
+ m.d.comb += self.out_z.copy(self.mod.out_z)
+ m.d.comb += self.out_norm.eq(self.mod.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.copy(self.mod.out_of)
+ m.d.sync += self.temp_z.copy(self.out_z)
+ with m.If(self.out_norm):
+ with m.If(self.in_accept):
+ m.d.sync += [
+ self.ack.eq(1),