# Copyright (C) Jonathan P Dawson 2013
# 2013-12-12
-from nmigen import Module, Signal, Cat, Mux
+from nmigen import Module, Signal, Cat, Mux, Array, Const
from nmigen.lib.coding import PriorityEncoder
from nmigen.cli import main, verilog
+from math import log
from fpbase import FPNumIn, FPNumOut, FPOp, Overflow, FPBase, FPNumBase
from fpbase import MultiShiftRMerge, Trigger
#from fpbase import FPNumShiftMultiRight
+
class FPState(FPBase):
def __init__(self, state_from):
self.state_from = state_from
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)
""" links module to inputs and outputs
"""
setattr(m.submodules, self.state_from, self.mod)
- m.d.comb += self.mod.in_op.copy(in_op)
+ m.d.comb += self.mod.in_op.eq(in_op)
#m.d.comb += self.out_op.eq(self.mod.out_op)
m.d.comb += self.out_decode.eq(self.mod.out_decode)
self.mod.ack.eq(0),
#self.out_op1.v.eq(self.mod.out_op1.v),
#self.out_op2.v.eq(self.mod.out_op2.v),
- self.out_op1.copy(self.mod.out_op1),
- self.out_op2.copy(self.mod.out_op2)
+ self.out_op1.eq(self.mod.out_op1),
+ self.out_op2.eq(self.mod.out_op2)
]
with m.Else():
m.d.sync += self.mod.ack.eq(1)
+class FPNumBase2Ops:
+
+ def __init__(self, width, m_extra=True):
+ self.a = FPNumBase(width, m_extra)
+ self.b = FPNumBase(width, m_extra)
+
+ def eq(self, i):
+ return [self.a.eq(i.a), self.b.eq(i.b)]
+
class FPAddSpecialCasesMod:
""" special cases: NaNs, infs, zeros, denormalised
"""
def __init__(self, width):
- self.in_a = FPNumBase(width)
- self.in_b = FPNumBase(width)
- self.out_z = FPNumOut(width, False)
+ self.width = width
+ self.i = self.ispec()
+ self.out_z = self.ospec()
self.out_do_z = Signal(reset_less=True)
+ def ispec(self):
+ return FPNumBase2Ops(self.width)
+
+ def ospec(self):
+ return FPNumOut(self.width, False)
+
+ def setup(self, m, in_a, in_b, out_do_z):
+ """ links module to inputs and outputs
+ """
+ m.submodules.specialcases = self
+ m.d.comb += self.i.a.eq(in_a)
+ m.d.comb += self.i.b.eq(in_b)
+ m.d.comb += out_do_z.eq(self.out_do_z)
+
def elaborate(self, platform):
m = Module()
- m.submodules.sc_in_a = self.in_a
- m.submodules.sc_in_b = self.in_b
+ m.submodules.sc_in_a = self.i.a
+ m.submodules.sc_in_b = self.i.b
m.submodules.sc_out_z = self.out_z
s_nomatch = Signal()
- m.d.comb += s_nomatch.eq(self.in_a.s != self.in_b.s)
+ m.d.comb += s_nomatch.eq(self.i.a.s != self.i.b.s)
m_match = Signal()
- m.d.comb += m_match.eq(self.in_a.m == self.in_b.m)
+ m.d.comb += m_match.eq(self.i.a.m == self.i.b.m)
# if a is NaN or b is NaN return NaN
- with m.If(self.in_a.is_nan | self.in_b.is_nan):
+ with m.If(self.i.a.is_nan | self.i.b.is_nan):
m.d.comb += self.out_do_z.eq(1)
m.d.comb += self.out_z.nan(0)
# m.d.comb += z.create(a.s & b.s, a.e, Cat(a.m[3:-2], 1))
# if a is inf return inf (or NaN)
- with m.Elif(self.in_a.is_inf):
+ with m.Elif(self.i.a.is_inf):
m.d.comb += self.out_do_z.eq(1)
- m.d.comb += self.out_z.inf(self.in_a.s)
+ m.d.comb += self.out_z.inf(self.i.a.s)
# if a is inf and signs don't match return NaN
- with m.If(self.in_b.exp_128 & s_nomatch):
+ with m.If(self.i.b.exp_128 & s_nomatch):
m.d.comb += self.out_z.nan(0)
# if b is inf return inf
- with m.Elif(self.in_b.is_inf):
+ with m.Elif(self.i.b.is_inf):
m.d.comb += self.out_do_z.eq(1)
- m.d.comb += self.out_z.inf(self.in_b.s)
+ m.d.comb += self.out_z.inf(self.i.b.s)
# if a is zero and b zero return signed-a/b
- with m.Elif(self.in_a.is_zero & self.in_b.is_zero):
+ with m.Elif(self.i.a.is_zero & self.i.b.is_zero):
m.d.comb += self.out_do_z.eq(1)
- m.d.comb += self.out_z.create(self.in_a.s & self.in_b.s,
- self.in_b.e,
- self.in_b.m[3:-1])
+ m.d.comb += self.out_z.create(self.i.a.s & self.i.b.s,
+ self.i.b.e,
+ self.i.b.m[3:-1])
# if a is zero return b
- with m.Elif(self.in_a.is_zero):
+ with m.Elif(self.i.a.is_zero):
m.d.comb += self.out_do_z.eq(1)
- m.d.comb += self.out_z.create(self.in_b.s, self.in_b.e,
- self.in_b.m[3:-1])
+ m.d.comb += self.out_z.create(self.i.b.s, self.i.b.e,
+ self.i.b.m[3:-1])
# if b is zero return a
- with m.Elif(self.in_b.is_zero):
+ with m.Elif(self.i.b.is_zero):
m.d.comb += self.out_do_z.eq(1)
- m.d.comb += self.out_z.create(self.in_a.s, self.in_a.e,
- self.in_a.m[3:-1])
+ m.d.comb += self.out_z.create(self.i.a.s, self.i.a.e,
+ self.i.a.m[3:-1])
# if a equal to -b return zero (+ve zero)
- with m.Elif(s_nomatch & m_match & (self.in_a.e == self.in_b.e)):
+ with m.Elif(s_nomatch & m_match & (self.i.a.e == self.i.b.e)):
m.d.comb += self.out_do_z.eq(1)
m.d.comb += self.out_z.zero(0)
FPState.__init__(self, "special_cases")
FPID.__init__(self, id_wid)
self.mod = FPAddSpecialCasesMod(width)
- self.out_z = FPNumOut(width, False)
+ self.out_z = self.mod.ospec()
self.out_do_z = Signal(reset_less=True)
def setup(self, m, in_a, in_b, in_mid):
""" links module to inputs and outputs
"""
- m.submodules.specialcases = self.mod
- m.d.comb += self.mod.in_a.copy(in_a)
- m.d.comb += self.mod.in_b.copy(in_b)
- #m.d.comb += self.out_z.v.eq(self.mod.out_z.v)
- m.d.comb += self.out_do_z.eq(self.mod.out_do_z)
+ self.mod.setup(m, in_a, in_b, self.out_do_z)
if self.in_mid is not None:
m.d.comb += self.in_mid.eq(in_mid)
m.next = "denormalise"
+class FPAddSpecialCasesDeNorm(FPState, FPID):
+ """ special cases: NaNs, infs, zeros, denormalised
+ NOTE: some of these are unique to add. see "Special Operations"
+ https://steve.hollasch.net/cgindex/coding/ieeefloat.html
+ """
+
+ def __init__(self, width, id_wid):
+ FPState.__init__(self, "special_cases")
+ FPID.__init__(self, id_wid)
+ self.smod = FPAddSpecialCasesMod(width)
+ self.out_z = self.smod.ospec()
+ self.out_do_z = Signal(reset_less=True)
+
+ self.dmod = FPAddDeNormMod(width)
+ self.o = self.dmod.ospec()
+
+ def setup(self, m, in_a, in_b, in_mid):
+ """ links module to inputs and outputs
+ """
+ self.smod.setup(m, in_a, in_b, self.out_do_z)
+ self.dmod.setup(m, in_a, in_b)
+ if self.in_mid is not None:
+ m.d.comb += self.in_mid.eq(in_mid)
+
+ def action(self, m):
+ self.idsync(m)
+ with m.If(self.out_do_z):
+ m.d.sync += self.out_z.v.eq(self.smod.out_z.v) # only take output
+ m.next = "put_z"
+ with m.Else():
+ m.next = "align"
+ m.d.sync += self.o.a.eq(self.dmod.o.a)
+ m.d.sync += self.o.b.eq(self.dmod.o.b)
+
+
class FPAddDeNormMod(FPState):
def __init__(self, width):
- self.in_a = FPNumBase(width)
- self.in_b = FPNumBase(width)
- self.out_a = FPNumBase(width)
- self.out_b = FPNumBase(width)
+ self.width = width
+ self.i = self.ispec()
+ self.o = self.ospec()
+
+ def ispec(self):
+ return FPNumBase2Ops(self.width)
+
+ def ospec(self):
+ return FPNumBase2Ops(self.width)
+
+ def setup(self, m, in_a, in_b):
+ """ links module to inputs and outputs
+ """
+ m.submodules.denormalise = self
+ m.d.comb += self.i.a.eq(in_a)
+ m.d.comb += self.i.b.eq(in_b)
def elaborate(self, platform):
m = Module()
- m.submodules.denorm_in_a = self.in_a
- m.submodules.denorm_in_b = self.in_b
- m.submodules.denorm_out_a = self.out_a
- m.submodules.denorm_out_b = self.out_b
+ 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
# hmmm, don't like repeating identical code
- m.d.comb += self.out_a.copy(self.in_a)
- with m.If(self.in_a.exp_n127):
- m.d.comb += self.out_a.e.eq(self.in_a.N126) # limit a exponent
+ 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.out_a.m[-1].eq(1) # set top mantissa bit
+ m.d.comb += self.o.a.m[-1].eq(1) # set top mantissa bit
- m.d.comb += self.out_b.copy(self.in_b)
- with m.If(self.in_b.exp_n127):
- m.d.comb += self.out_b.e.eq(self.in_b.N126) # limit a exponent
+ 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.out_b.m[-1].eq(1) # set top mantissa bit
+ m.d.comb += self.o.b.m[-1].eq(1) # set top mantissa bit
return m
def setup(self, m, in_a, in_b, in_mid):
""" links module to inputs and outputs
"""
- m.submodules.denormalise = self.mod
- m.d.comb += self.mod.in_a.copy(in_a)
- m.d.comb += self.mod.in_b.copy(in_b)
+ self.mod.setup(m, in_a, in_b)
if self.in_mid is not None:
m.d.comb += self.in_mid.eq(in_mid)
self.idsync(m)
# Denormalised Number checks
m.next = "align"
- m.d.sync += self.out_a.copy(self.mod.out_a)
- m.d.sync += self.out_b.copy(self.mod.out_b)
+ m.d.sync += self.out_a.eq(self.mod.out_a)
+ m.d.sync += self.out_b.eq(self.mod.out_b)
class FPAddAlignMultiMod(FPState):
# exponent of a greater than b: shift b down
m.d.comb += self.exp_eq.eq(0)
- m.d.comb += self.out_a.copy(self.in_a)
- m.d.comb += self.out_b.copy(self.in_b)
+ m.d.comb += self.out_a.eq(self.in_a)
+ m.d.comb += self.out_b.eq(self.in_b)
agtb = Signal(reset_less=True)
altb = Signal(reset_less=True)
m.d.comb += agtb.eq(self.in_a.e > self.in_b.e)
""" links module to inputs and outputs
"""
m.submodules.align = self.mod
- m.d.comb += self.mod.in_a.copy(in_a)
- m.d.comb += self.mod.in_b.copy(in_b)
- #m.d.comb += self.out_a.copy(self.mod.out_a)
- #m.d.comb += self.out_b.copy(self.mod.out_b)
+ 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)
if self.in_mid is not None:
m.d.comb += self.in_mid.eq(in_mid)
def action(self, m):
self.idsync(m)
- m.d.sync += self.out_a.copy(self.mod.out_a)
- m.d.sync += self.out_b.copy(self.mod.out_b)
+ m.d.sync += self.out_a.eq(self.mod.out_a)
+ m.d.sync += self.out_b.eq(self.mod.out_b)
with m.If(self.exp_eq):
m.next = "add_0"
+class FPNumIn2Ops:
+
+ def __init__(self, width):
+ self.a = FPNumIn(None, width)
+ self.b = FPNumIn(None, width)
+
+ def eq(self, i):
+ return [self.a.eq(i.a), self.b.eq(i.b)]
+
+
class FPAddAlignSingleMod:
def __init__(self, width):
self.width = width
- self.in_a = FPNumBase(width)
- self.in_b = FPNumBase(width)
- self.out_a = FPNumIn(None, width)
- self.out_b = FPNumIn(None, width)
+ self.i = self.ispec()
+ self.o = self.ospec()
+
+ def ispec(self):
+ return FPNumBase2Ops(self.width)
+
+ def ospec(self):
+ return FPNumIn2Ops(self.width)
+
+ def setup(self, m, in_a, in_b):
+ """ links module to inputs and outputs
+ """
+ m.submodules.align = self
+ m.d.comb += self.i.a.eq(in_a)
+ m.d.comb += self.i.b.eq(in_b)
def elaborate(self, platform):
""" Aligns A against B or B against A, depending on which has the
"""
m = Module()
- m.submodules.align_in_a = self.in_a
- m.submodules.align_in_b = self.in_b
- m.submodules.align_out_a = self.out_a
- m.submodules.align_out_b = self.out_b
+ m.submodules.align_in_a = self.i.a
+ m.submodules.align_in_b = self.i.b
+ m.submodules.align_out_a = self.o.a
+ m.submodules.align_out_b = self.o.b
# temporary (muxed) input and output to be shifted
t_inp = FPNumBase(self.width)
t_out = FPNumIn(None, self.width)
- espec = (len(self.in_a.e), True)
- msr = MultiShiftRMerge(self.in_a.m_width, espec)
+ espec = (len(self.i.a.e), True)
+ msr = MultiShiftRMerge(self.i.a.m_width, espec)
m.submodules.align_t_in = t_inp
m.submodules.align_t_out = t_out
m.submodules.multishift_r = msr
m.d.comb += t_out.e.eq(t_inp.e + tdiff)
m.d.comb += t_out.s.eq(t_inp.s)
- m.d.comb += ediff.eq(self.in_a.e - self.in_b.e)
- m.d.comb += ediffr.eq(self.in_b.e - self.in_a.e)
- m.d.comb += elz.eq(self.in_a.e < self.in_b.e)
- m.d.comb += egz.eq(self.in_a.e > self.in_b.e)
+ m.d.comb += ediff.eq(self.i.a.e - self.i.b.e)
+ m.d.comb += ediffr.eq(self.i.b.e - self.i.a.e)
+ m.d.comb += elz.eq(self.i.a.e < self.i.b.e)
+ m.d.comb += egz.eq(self.i.a.e > self.i.b.e)
# default: A-exp == B-exp, A and B untouched (fall through)
- m.d.comb += self.out_a.copy(self.in_a)
- m.d.comb += self.out_b.copy(self.in_b)
+ m.d.comb += self.o.a.eq(self.i.a)
+ m.d.comb += self.o.b.eq(self.i.b)
# only one shifter (muxed)
#m.d.comb += t_out.shift_down_multi(tdiff, t_inp)
# exponent of a greater than b: shift b down
with m.If(egz):
- m.d.comb += [t_inp.copy(self.in_b),
+ m.d.comb += [t_inp.eq(self.i.b),
tdiff.eq(ediff),
- self.out_b.copy(t_out),
- self.out_b.s.eq(self.in_b.s), # whoops forgot sign
+ self.o.b.eq(t_out),
+ self.o.b.s.eq(self.i.b.s), # whoops forgot sign
]
# exponent of b greater than a: shift a down
with m.Elif(elz):
- m.d.comb += [t_inp.copy(self.in_a),
+ m.d.comb += [t_inp.eq(self.i.a),
tdiff.eq(ediffr),
- self.out_a.copy(t_out),
- self.out_a.s.eq(self.in_a.s), # whoops forgot sign
+ self.o.a.eq(t_out),
+ self.o.a.s.eq(self.i.a.s), # whoops forgot sign
]
return m
def setup(self, m, in_a, in_b, in_mid):
""" links module to inputs and outputs
"""
- m.submodules.align = self.mod
- m.d.comb += self.mod.in_a.copy(in_a)
- m.d.comb += self.mod.in_b.copy(in_b)
+ self.mod.setup(m, in_a, in_b)
if self.in_mid is not None:
m.d.comb += self.in_mid.eq(in_mid)
def action(self, m):
self.idsync(m)
# NOTE: could be done as comb
- m.d.sync += self.out_a.copy(self.mod.out_a)
- m.d.sync += self.out_b.copy(self.mod.out_b)
+ m.d.sync += self.out_a.eq(self.mod.out_a)
+ m.d.sync += self.out_b.eq(self.mod.out_b)
m.next = "add_0"
+class FPAddAlignSingleAdd(FPState, FPID):
+
+ def __init__(self, width, id_wid):
+ FPState.__init__(self, "align")
+ FPID.__init__(self, id_wid)
+ self.mod = FPAddAlignSingleMod(width)
+ self.o = self.mod.ospec()
+
+ self.a0mod = FPAddStage0Mod(width)
+ self.a0_out_z = FPNumBase(width, False)
+ self.out_tot = Signal(self.a0_out_z.m_width + 4, reset_less=True)
+ self.a0_out_z = FPNumBase(width, False)
+
+ self.a1mod = FPAddStage1Mod(width)
+ self.a1o = self.a1mod.ospec()
+
+ def setup(self, m, in_a, in_b, in_mid):
+ """ links module to inputs and outputs
+ """
+ self.mod.setup(m, in_a, in_b)
+ m.d.comb += self.o.eq(self.mod.o)
+
+ self.a0mod.setup(m, self.o.a, self.o.b)
+ m.d.comb += self.a0_out_z.eq(self.a0mod.o.z)
+ m.d.comb += self.out_tot.eq(self.a0mod.o.tot)
+
+ self.a1mod.setup(m, self.out_tot, self.a0_out_z)
+
+ if self.in_mid is not None:
+ m.d.comb += self.in_mid.eq(in_mid)
+
+ def action(self, m):
+ self.idsync(m)
+ m.d.sync += self.a1o.eq(self.a1mod.o)
+ m.next = "normalise_1"
+
+
+class FPAddStage0Data:
+
+ def __init__(self, width):
+ self.z = FPNumBase(width, False)
+ self.tot = Signal(self.z.m_width + 4, reset_less=True)
+
+ def eq(self, i):
+ return [self.z.eq(i.z), self.tot.eq(i.tot)]
+
+
class FPAddStage0Mod:
def __init__(self, width):
- self.in_a = FPNumBase(width)
- self.in_b = FPNumBase(width)
- self.in_z = FPNumBase(width, False)
- self.out_z = FPNumBase(width, False)
- self.out_tot = Signal(self.out_z.m_width + 4, reset_less=True)
+ self.width = width
+ self.i = self.ispec()
+ self.o = self.ospec()
+
+ def ispec(self):
+ return FPNumBase2Ops(self.width)
+
+ def ospec(self):
+ return FPAddStage0Data(self.width)
+
+ def setup(self, m, in_a, in_b):
+ """ links module to inputs and outputs
+ """
+ m.submodules.add0 = self
+ m.d.comb += self.i.a.eq(in_a)
+ m.d.comb += self.i.b.eq(in_b)
def elaborate(self, platform):
m = Module()
- m.submodules.add0_in_a = self.in_a
- m.submodules.add0_in_b = self.in_b
- m.submodules.add0_out_z = self.out_z
+ m.submodules.add0_in_a = self.i.a
+ m.submodules.add0_in_b = self.i.b
+ m.submodules.add0_out_z = self.o.z
- m.d.comb += self.out_z.e.eq(self.in_a.e)
+ m.d.comb += self.o.z.e.eq(self.i.a.e)
# store intermediate tests (and zero-extended mantissas)
seq = Signal(reset_less=True)
mge = Signal(reset_less=True)
- am0 = Signal(len(self.in_a.m)+1, reset_less=True)
- bm0 = Signal(len(self.in_b.m)+1, reset_less=True)
- m.d.comb += [seq.eq(self.in_a.s == self.in_b.s),
- mge.eq(self.in_a.m >= self.in_b.m),
- am0.eq(Cat(self.in_a.m, 0)),
- bm0.eq(Cat(self.in_b.m, 0))
+ am0 = Signal(len(self.i.a.m)+1, reset_less=True)
+ bm0 = Signal(len(self.i.b.m)+1, reset_less=True)
+ m.d.comb += [seq.eq(self.i.a.s == self.i.b.s),
+ mge.eq(self.i.a.m >= self.i.b.m),
+ am0.eq(Cat(self.i.a.m, 0)),
+ bm0.eq(Cat(self.i.b.m, 0))
]
# same-sign (both negative or both positive) add mantissas
with m.If(seq):
m.d.comb += [
- self.out_tot.eq(am0 + bm0),
- self.out_z.s.eq(self.in_a.s)
+ self.o.tot.eq(am0 + bm0),
+ self.o.z.s.eq(self.i.a.s)
]
# a mantissa greater than b, use a
with m.Elif(mge):
m.d.comb += [
- self.out_tot.eq(am0 - bm0),
- self.out_z.s.eq(self.in_a.s)
+ self.o.tot.eq(am0 - bm0),
+ self.o.z.s.eq(self.i.a.s)
]
# b mantissa greater than a, use b
with m.Else():
m.d.comb += [
- self.out_tot.eq(bm0 - am0),
- self.out_z.s.eq(self.in_b.s)
+ self.o.tot.eq(bm0 - am0),
+ self.o.z.s.eq(self.i.b.s)
]
return m
FPState.__init__(self, "add_0")
FPID.__init__(self, id_wid)
self.mod = FPAddStage0Mod(width)
- self.out_z = FPNumBase(width, False)
- self.out_tot = Signal(self.out_z.m_width + 4, reset_less=True)
+ self.o = self.mod.ospec()
def setup(self, m, in_a, in_b, in_mid):
""" 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)
+ self.mod.setup(m, in_a, in_b)
if self.in_mid is not None:
m.d.comb += self.in_mid.eq(in_mid)
def action(self, m):
self.idsync(m)
# 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)
+ m.d.sync += self.o.eq(self.mod.o)
m.next = "add_1"
+class FPAddStage1Data:
+
+ def __init__(self, width):
+ self.z = FPNumBase(width, False)
+ self.of = Overflow()
+
+ def eq(self, i):
+ return [self.z.eq(i.z), self.of.eq(i.of)]
+
+
+
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()
+ self.width = width
+ self.i = self.ispec()
+ self.o = self.ospec()
+
+ def ispec(self):
+ return FPAddStage0Data(self.width)
+
+ def ospec(self):
+ return FPAddStage1Data(self.width)
+
+ def setup(self, m, in_tot, in_z):
+ """ links module to inputs and outputs
+ """
+ m.submodules.add1 = self
+ m.submodules.add1_out_overflow = self.o.of
+
+ m.d.comb += self.i.z.eq(in_z)
+ m.d.comb += self.i.tot.eq(in_tot)
def elaborate(self, platform):
m = Module()
#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.o.z.eq(self.i.z)
+ # tot[-1] (MSB) gets set when the sum overflows. shift result down
+ with m.If(self.i.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)
+ self.o.z.m.eq(self.i.tot[4:]),
+ self.o.of.m0.eq(self.i.tot[4]),
+ self.o.of.guard.eq(self.i.tot[3]),
+ self.o.of.round_bit.eq(self.i.tot[2]),
+ self.o.of.sticky.eq(self.i.tot[1] | self.i.tot[0]),
+ self.o.z.e.eq(self.i.z.e + 1)
]
- # tot[27] zero case
+ # tot[-1] (MSB) 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])
+ self.o.z.m.eq(self.i.tot[3:]),
+ self.o.of.m0.eq(self.i.tot[3]),
+ self.o.of.guard.eq(self.i.tot[2]),
+ self.o.of.round_bit.eq(self.i.tot[1]),
+ self.o.of.sticky.eq(self.i.tot[0])
]
return m
def setup(self, m, in_tot, in_z, in_mid):
""" links module to inputs and outputs
"""
- m.submodules.add1 = self.mod
- m.submodules.add1_out_overflow = self.out_of
-
- m.d.comb += self.mod.in_z.copy(in_z)
- m.d.comb += self.mod.in_tot.eq(in_tot)
+ self.mod.setup(m, in_tot, in_z)
m.d.sync += self.norm_stb.eq(0) # sets to zero when not in add1 state
def action(self, m):
self.idsync(m)
- 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.out_of.eq(self.mod.out_of)
+ m.d.sync += self.out_z.eq(self.mod.out_z)
m.d.sync += self.norm_stb.eq(1)
m.next = "normalise_1"
+class FPNormaliseModSingle:
+
+ def __init__(self, width):
+ self.width = width
+ self.in_z = FPNumBase(width, False)
+ self.out_z = FPNumBase(width, False)
+
+ def setup(self, m, in_z, out_z, modname):
+ """ links module to inputs and outputs
+ """
+ m.submodules.normalise = self
+ m.d.comb += self.in_z.eq(in_z)
+ m.d.comb += out_z.eq(self.out_z)
+
+ 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_in_z = self.in_z
+
+ 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 increase/decrease conditions
+ 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 FPNorm1ModSingle:
def __init__(self, width):
"""
m.submodules.normalise_1 = self
- m.d.comb += self.in_z.copy(in_z)
- m.d.comb += self.in_of.copy(in_of)
+ m.d.comb += self.in_z.eq(in_z)
+ m.d.comb += self.in_of.eq(in_of)
- m.d.comb += out_z.copy(self.out_z)
+ m.d.comb += out_z.eq(self.out_z)
def elaborate(self, platform):
m = Module()
msr = MultiShiftRMerge(mwid, espec)
m.submodules.multishift_r = msr
- m.d.comb += in_z.copy(self.in_z)
- m.d.comb += in_of.copy(self.in_of)
+ 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.copy(in_z)
- m.d.comb += self.out_of.copy(in_of)
+ 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)
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)
# 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)
+ 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.copy(self.temp_z)
- m.d.comb += in_of.copy(self.temp_of)
+ 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.copy(in_z)
- m.d.comb += self.out_of.copy(in_of)
+ 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)
def action(self, m):
self.idsync(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)
+ 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 += [
r_out_z = FPNumBase(self.width)
rmod.setup(m, n_out_z, n_out_roundz)
m.d.comb += n_out_roundz.eq(nmod.out_of.roundz)
- m.d.comb += r_out_z.copy(rmod.out_z)
+ m.d.comb += r_out_z.eq(rmod.out_z)
# Corrections (chained to rounding)
cmod = FPCorrectionsMod(self.width)
c_out_z = FPNumBase(self.width)
cmod.setup(m, r_out_z)
- m.d.comb += c_out_z.copy(cmod.out_z)
+ m.d.comb += c_out_z.eq(cmod.out_z)
# Pack (chained to corrections)
self.pmod = FPPackMod(self.width)
def setup(self, m, in_z, roundz):
m.submodules.roundz = self
- m.d.comb += self.in_z.copy(in_z)
+ m.d.comb += self.in_z.eq(in_z)
m.d.comb += self.in_roundz.eq(roundz)
def elaborate(self, platform):
m = Module()
- m.d.comb += self.out_z.copy(self.in_z)
+ m.d.comb += self.out_z.eq(self.in_z)
with m.If(self.in_roundz):
m.d.comb += self.out_z.m.eq(self.in_z.m + 1) # mantissa rounds up
with m.If(self.in_z.m == self.in_z.m1s): # all 1s
def action(self, m):
self.idsync(m)
- m.d.sync += self.out_z.copy(self.mod.out_z)
+ m.d.sync += self.out_z.eq(self.mod.out_z)
m.next = "corrections"
""" links module to inputs and outputs
"""
m.submodules.corrections = self
- m.d.comb += self.in_z.copy(in_z)
+ m.d.comb += self.in_z.eq(in_z)
def elaborate(self, platform):
m = Module()
m.submodules.corr_in_z = self.in_z
m.submodules.corr_out_z = self.out_z
- m.d.comb += self.out_z.copy(self.in_z)
+ m.d.comb += self.out_z.eq(self.in_z)
with m.If(self.in_z.is_denormalised):
m.d.comb += self.out_z.e.eq(self.in_z.N127)
return m
def action(self, m):
self.idsync(m)
- m.d.sync += self.out_z.copy(self.mod.out_z)
+ m.d.sync += self.out_z.eq(self.mod.out_z)
m.next = "pack"
""" links module to inputs and outputs
"""
m.submodules.pack = self
- m.d.comb += self.in_z.copy(in_z)
+ m.d.comb += self.in_z.eq(in_z)
def elaborate(self, platform):
m = Module()
class FPPutZ(FPState):
- def __init__(self, state, in_z, out_z, in_mid, out_mid):
+ 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
]
with m.If(self.out_z.stb & self.out_z.ack):
m.d.sync += self.out_z.stb.eq(0)
- m.next = "get_ops"
+ m.next = self.to_state
with m.Else():
m.d.sync += self.out_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 FPADDBaseMod(FPID):
def __init__(self, width, id_wid=None, single_cycle=False, compact=True):
a = get.out_op1
b = get.out_op2
- sc = self.add_state(FPAddSpecialCases(self.width, self.id_wid))
+ sc = self.add_state(FPAddSpecialCasesDeNorm(self.width, self.id_wid))
sc.setup(m, a, b, self.in_mid)
- dn = self.add_state(FPAddDeNorm(self.width, self.id_wid))
- dn.setup(m, a, b, sc.in_mid)
-
- if self.single_cycle:
- alm = self.add_state(FPAddAlignSingle(self.width, self.id_wid))
- alm.setup(m, dn.out_a, dn.out_b, dn.in_mid)
- else:
- alm = self.add_state(FPAddAlignMulti(self.width, self.id_wid))
- alm.setup(m, dn.out_a, dn.out_b, dn.in_mid)
-
- add0 = self.add_state(FPAddStage0(self.width, self.id_wid))
- add0.setup(m, alm.out_a, alm.out_b, alm.in_mid)
-
- add1 = self.add_state(FPAddStage1(self.width, self.id_wid))
- add1.setup(m, add0.out_tot, add0.out_z, add0.in_mid)
+ alm = self.add_state(FPAddAlignSingleAdd(self.width, self.id_wid))
+ alm.setup(m, sc.o.a, sc.o.b, sc.in_mid)
n1 = self.add_state(FPNormToPack(self.width, self.id_wid))
- n1.setup(m, add1.out_z, add1.out_of, add0.in_mid)
+ n1.setup(m, alm.a1o.z, alm.a1o.of, alm.in_mid)
ppz = self.add_state(FPPutZ("pack_put_z", n1.out_z, self.out_z,
n1.in_mid, self.out_mid))
m.d.sync += self.add_stb.eq(add_stb)
m.d.sync += self.add_ack.eq(0) # sets to zero when not in active state
+ m.d.sync += self.out_z.ack.eq(0) # likewise
#m.d.sync += self.in_t.stb.eq(0)
m.submodules.fpadd = self.mod
with m.Else():
m.d.sync += [self.add_ack.eq(0),
self.in_t.stb.eq(0),
+ self.out_z.ack.eq(1),
]
with m.Else():
# done: acknowledge, and write out id and value
m.d.sync += [self.add_ack.eq(1),
self.in_t.stb.eq(0)
]
- m.next = "get_a"
+ m.next = "put_z"
return
with m.Else():
m.d.sync += self.out_z.stb.eq(1)
+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:
needs to be the thing that raises the incoming stb.
"""
- def __init__(self, width, id_wid=None, single_cycle=False):
+ def __init__(self, width, id_wid=None, single_cycle=False, rs_sz=2):
""" IEEE754 FP Add
* width: bit-width of IEEE754. supported: 16, 32, 64
* id_wid: an identifier that is sync-connected to the input
* single_cycle: True indicates each stage to complete in 1 clock
"""
- FPID.__init__(self, id_wid)
self.width = width
self.id_wid = id_wid
self.single_cycle = single_cycle
- self.in_a = FPOp(width)
- self.in_b = FPOp(width)
- self.out_z = FPOp(width)
+ #self.out_z = FPOp(width)
+ self.ids = FPID(id_wid)
+
+ rs = []
+ for i in range(rs_sz):
+ in_a = FPOp(width)
+ in_b = FPOp(width)
+ in_a.name = "in_a_%d" % i
+ in_b.name = "in_b_%d" % i
+ rs.append((in_a, in_b))
+ self.rs = Array(rs)
+
+ 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.states = []
""" creates the HDL code-fragment for FPAdd
"""
m = Module()
- m.submodules.in_a = self.in_a
- m.submodules.in_b = self.in_b
- m.submodules.out_z = self.out_z
+ m.submodules += self.rs
+
+ in_a = self.rs[0][0]
+ in_b = self.rs[0][1]
+
+ out_z = FPOp(self.width)
+ out_mid = Signal(self.id_wid, reset_less=True)
+ m.submodules.out_z = out_z
geta = self.add_state(FPGetOp("get_a", "get_b",
- self.in_a, self.width))
- geta.setup(m, self.in_a)
+ in_a, self.width))
+ geta.setup(m, in_a)
a = geta.out_op
getb = self.add_state(FPGetOp("get_b", "fpadd",
- self.in_b, self.width))
- getb.setup(m, self.in_b)
+ in_b, self.width))
+ getb.setup(m, in_b)
b = getb.out_op
ab = FPADDBase(self.width, self.id_wid, self.single_cycle)
ab = self.add_state(ab)
- ab.setup(m, a, b, getb.out_decode, self.in_mid,
- self.out_z, self.out_mid)
+ ab.setup(m, a, b, getb.out_decode, self.ids.in_mid,
+ out_z, out_mid)
- #pz = self.add_state(FPPutZ("put_z", ab.out_z, self.out_z,
- # ab.out_mid, self.out_mid))
+ pz = self.add_state(FPPutZIdx("put_z", ab.out_z, self.res,
+ out_mid, "get_a"))
with m.FSM() as fsm:
if __name__ == "__main__":
if True:
alu = FPADD(width=32, id_wid=5, single_cycle=True)
- main(alu, ports=alu.in_a.ports() + \
- alu.in_b.ports() + \
- alu.out_z.ports() + \
- [alu.in_mid, alu.out_mid])
+ main(alu, ports=alu.rs[0][0].ports() + \
+ alu.rs[0][1].ports() + \
+ alu.res[0].ports() + \
+ [alu.ids.in_mid, alu.ids.out_mid])
else:
alu = FPADDBase(width=32, id_wid=5, single_cycle=True)
main(alu, ports=[alu.in_a, alu.in_b] + \