1 # IEEE Floating Point Divider (Single Precision)
2 # Copyright (C) Jonathan P Dawson 2013
5 from nmigen
import Module
, Signal
, Const
, Cat
6 from nmigen
.cli
import main
, verilog
8 from fpbase
import FPNumIn
, FPNumOut
, FPOp
, Overflow
, FPBase
11 def __init__(self
, width
):
13 self
.quot
= Signal(width
) # quotient
14 self
.dor
= Signal(width
) # divisor
15 self
.dend
= Signal(width
) # dividend
16 self
.rem
= Signal(width
) # remainder
17 self
.count
= Signal(7) # loop count
19 self
.czero
= Const(0, width
)
23 self
.quot
.eq(self
.czero
),
24 self
.rem
.eq(self
.czero
),
25 self
.count
.eq(Const(0, 7))
31 def __init__(self
, width
):
35 self
.in_a
= FPOp(width
)
36 self
.in_b
= FPOp(width
)
37 self
.out_z
= FPOp(width
)
39 def get_fragment(self
, platform
=None):
40 """ creates the HDL code-fragment for FPDiv
45 a
= FPNumIn(None, self
.width
, False)
46 b
= FPNumIn(None, self
.width
, False)
47 z
= FPNumOut(self
.width
, False)
49 div
= Div(a
.m_width
*2 + 3) # double the mantissa width plus g/r/sticky
62 with m
.State("get_a"):
63 self
.get_op(m
, self
.in_a
, a
, "get_b")
68 with m
.State("get_b"):
69 self
.get_op(m
, self
.in_b
, b
, "special_cases")
72 # special cases: NaNs, infs, zeros, denormalised
73 # NOTE: some of these are unique to div. see "Special Operations"
74 # https://steve.hollasch.net/cgindex/coding/ieeefloat.html
76 with m
.State("special_cases"):
78 # if a is NaN or b is NaN return NaN
79 with m
.If(a
.is_nan | b
.is_nan
):
83 # if a is Inf and b is Inf return NaN
84 with m
.Elif(a
.is_inf
& b
.is_inf
):
88 # if a is inf return inf (or NaN if b is zero)
89 with m
.Elif(a
.is_inf
):
91 m
.d
.sync
+= z
.inf(a
.s ^ b
.s
)
93 # if b is inf return zero
94 with m
.Elif(b
.is_inf
):
96 m
.d
.sync
+= z
.zero(a
.s ^ b
.s
)
98 # if a is zero return zero (or NaN if b is zero)
99 with m
.Elif(a
.is_zero
):
101 # if b is zero return NaN
102 with m
.If(b
.is_zero
):
105 m
.d
.sync
+= z
.zero(a
.s ^ b
.s
)
107 # if b is zero return Inf
108 with m
.Elif(b
.is_zero
):
110 m
.d
.sync
+= z
.inf(a
.s ^ b
.s
)
112 # Denormalised Number checks
114 m
.next
= "normalise_a"
115 self
.denormalise(m
, a
)
116 self
.denormalise(m
, b
)
121 with m
.State("normalise_a"):
122 self
.op_normalise(m
, a
, "normalise_b")
127 with m
.State("normalise_b"):
128 self
.op_normalise(m
, b
, "divide_0")
131 # First stage of divide. initialise state
133 with m
.State("divide_0"):
136 z
.s
.eq(a
.s ^ b
.s
), # sign
137 z
.e
.eq(a
.e
- b
.e
), # exponent
138 div
.dend
.eq(a
.m
<<(a
.m_width
+3)), # 3 bits for g/r/sticky
144 # Second stage of divide.
146 with m
.State("divide_1"):
149 div
.quot
.eq(div
.quot
<< 1),
150 div
.rem
.eq(Cat(div
.dend
[-1], div
.rem
[0:])),
151 div
.dend
.eq(div
.dend
<< 1),
155 # Third stage of divide.
156 # This stage ends by jumping out to divide_3
157 # However it defaults to jumping to divide_1 (which comes back here)
159 with m
.State("divide_2"):
160 with m
.If(div
.rem
>= div
.dor
):
163 div
.rem
.eq(div
.rem
- div
.dor
),
165 with m
.If(div
.count
== div
.width
-2):
170 div
.count
.eq(div
.count
+ 1),
174 # Fourth stage of divide.
176 with m
.State("divide_3"):
177 m
.next
= "normalise_1"
179 z
.m
.eq(div
.quot
[3:]),
180 of
.guard
.eq(div
.quot
[2]),
181 of
.round_bit
.eq(div
.quot
[1]),
182 of
.sticky
.eq(div
.quot
[0] |
(div
.rem
!= 0))
186 # First stage of normalisation.
188 with m
.State("normalise_1"):
189 self
.normalise_1(m
, z
, of
, "normalise_2")
192 # Second stage of normalisation.
194 with m
.State("normalise_2"):
195 self
.normalise_2(m
, z
, of
, "round")
200 with m
.State("round"):
201 self
.roundz(m
, z
, of
.roundz
)
202 m
.next
= "corrections"
207 with m
.State("corrections"):
208 self
.corrections(m
, z
, "pack")
213 with m
.State("pack"):
214 self
.pack(m
, z
, "put_z")
219 with m
.State("put_z"):
220 self
.put_z(m
, z
, self
.out_z
, "get_a")
225 if __name__
== "__main__":
226 alu
= FPDIV(width
=32)
227 main(alu
, ports
=alu
.in_a
.ports() + alu
.in_b
.ports() + alu
.out_z
.ports())
230 # works... but don't use, just do "python fname.py convert -t v"
231 #print (verilog.convert(alu, ports=[
232 # ports=alu.in_a.ports() + \
233 # alu.in_b.ports() + \