src/add/fpbase.py

   1 # IEEE Floating Point Adder (Single Precision)
   2 # Copyright (C) Jonathan P Dawson 2013
   3 # 2013-12-12
   4
   5 from nmigen import Signal, Cat, Const, Mux, Module
   6 from math import log
   7 from operator import or_
   8 from functools import reduce
   9
  10 class MultiShiftR:
  11
  12     def __init__(self, width):
  13         self.width = width
  14         self.smax = int(log(width) / log(2))
  15         self.i = Signal(width, reset_less=True)
  16         self.s = Signal(self.smax, reset_less=True)
  17         self.o = Signal(width, reset_less=True)
  18
  19     def elaborate(self, platform):
  20         m = Module()
  21         m.d.comb += self.o.eq(self.i >> self.s)
  22         return m
  23
  24
  25 class MultiShift:
  26     """ Generates variable-length single-cycle shifter from a series
  27         of conditional tests on each bit of the left/right shift operand.
  28         Each bit tested produces output shifted by that number of bits,
  29         in a binary fashion: bit 1 if set shifts by 1 bit, bit 2 if set
  30         shifts by 2 bits, each partial result cascading to the next Mux.
  31
  32         Could be adapted to do arithmetic shift by taking copies of the
  33         MSB instead of zeros.
  34     """
  35
  36     def __init__(self, width):
  37         self.width = width
  38         self.smax = int(log(width) / log(2))
  39
  40     def lshift(self, op, s):
  41         res = op << s
  42         return res[:len(op)]
  43         res = op
  44         for i in range(self.smax):
  45             zeros = [0] * (1<<i)
  46             res = Mux(s & (1<<i), Cat(zeros, res[0:-(1<<i)]), res)
  47         return res
  48
  49     def rshift(self, op, s):
  50         res = op >> s
  51         return res[:len(op)]
  52         res = op
  53         for i in range(self.smax):
  54             zeros = [0] * (1<<i)
  55             res = Mux(s & (1<<i), Cat(res[(1<<i):], zeros), res)
  56         return res
  57
  58
  59 class FPNumBase:
  60     """ Floating-point Base Number Class
  61     """
  62     def __init__(self, width, m_extra=True):
  63         self.width = width
  64         m_width = {16: 11, 32: 24, 64: 53}[width] # 1 extra bit (overflow)
  65         e_width = {16: 7,  32: 10, 64: 13}[width] # 2 extra bits (overflow)
  66         e_max = 1<<(e_width-3)
  67         self.rmw = m_width # real mantissa width (not including extras)
  68         self.e_max = e_max
  69         if m_extra:
  70             # mantissa extra bits (top,guard,round)
  71             self.m_extra = 3
  72             m_width += self.m_extra
  73         else:
  74             self.m_extra = 0
  75         #print (m_width, e_width, e_max, self.rmw, self.m_extra)
  76         self.m_width = m_width
  77         self.e_width = e_width
  78         self.e_start = self.rmw - 1
  79         self.e_end = self.rmw + self.e_width - 3 # for decoding
  80
  81         self.v = Signal(width, reset_less=True)      # Latched copy of value
  82         self.m = Signal(m_width, reset_less=True)    # Mantissa
  83         self.e = Signal((e_width, True), reset_less=True) # Exponent: IEEE754exp+2 bits, signed
  84         self.s = Signal(reset_less=True)           # Sign bit
  85
  86         self.mzero = Const(0, (m_width, False))
  87         self.m1s = Const(-1, (m_width, False))
  88         self.P128 = Const(e_max, (e_width, True))
  89         self.P127 = Const(e_max-1, (e_width, True))
  90         self.N127 = Const(-(e_max-1), (e_width, True))
  91         self.N126 = Const(-(e_max-2), (e_width, True))
  92
  93         self.is_nan = Signal(reset_less=True)
  94         self.is_zero = Signal(reset_less=True)
  95         self.is_inf = Signal(reset_less=True)
  96         self.is_overflowed = Signal(reset_less=True)
  97         self.is_denormalised = Signal(reset_less=True)
  98         self.exp_128 = Signal(reset_less=True)
  99         self.exp_gt127 = Signal(reset_less=True)
 100         self.exp_n127 = Signal(reset_less=True)
 101         self.exp_n126 = Signal(reset_less=True)
 102         self.m_zero = Signal(reset_less=True)
 103         self.m_msbzero = Signal(reset_less=True)
 104
 105     def elaborate(self, platform):
 106         m = Module()
 107         m.d.comb += self.is_nan.eq(self._is_nan())
 108         m.d.comb += self.is_zero.eq(self._is_zero())
 109         m.d.comb += self.is_inf.eq(self._is_inf())
 110         m.d.comb += self.is_overflowed.eq(self._is_overflowed())
 111         m.d.comb += self.is_denormalised.eq(self._is_denormalised())
 112         m.d.comb += self.exp_128.eq(self.e == self.P128)
 113         m.d.comb += self.exp_gt127.eq(self.e > self.P127)
 114         m.d.comb += self.exp_n127.eq(self.e == self.N127)
 115         m.d.comb += self.exp_n126.eq(self.e == self.N126)
 116         m.d.comb += self.m_zero.eq(self.m == self.mzero)
 117         m.d.comb += self.m_msbzero.eq(self.m[self.e_start] == 0)
 118
 119         return m
 120
 121     def _is_nan(self):
 122         return (self.exp_128) & (~self.m_zero)
 123
 124     def _is_inf(self):
 125         return (self.exp_128) & (self.m_zero)
 126
 127     def _is_zero(self):
 128         return (self.exp_n127) & (self.m_zero)
 129
 130     def _is_overflowed(self):
 131         return self.exp_gt127
 132
 133     def _is_denormalised(self):
 134         return (self.exp_n126) & (self.m_msbzero)
 135
 136
 137 class FPNumOut(FPNumBase):
 138     """ Floating-point Number Class
 139
 140         Contains signals for an incoming copy of the value, decoded into
 141         sign / exponent / mantissa.
 142         Also contains encoding functions, creation and recognition of
 143         zero, NaN and inf (all signed)
 144
 145         Four extra bits are included in the mantissa: the top bit
 146         (m[-1]) is effectively a carry-overflow.  The other three are
 147         guard (m[2]), round (m[1]), and sticky (m[0])
 148     """
 149     def __init__(self, width, m_extra=True):
 150         FPNumBase.__init__(self, width, m_extra)
 151
 152     def elaborate(self, platform):
 153         m = FPNumBase.elaborate(self, platform)
 154
 155         return m
 156
 157     def create(self, s, e, m):
 158         """ creates a value from sign / exponent / mantissa
 159
 160             bias is added here, to the exponent
 161         """
 162         return [
 163           self.v[-1].eq(s),          # sign
 164           self.v[self.e_start:self.e_end].eq(e + self.P127), # exp (add on bias)
 165           self.v[0:self.e_start].eq(m)         # mantissa
 166         ]
 167
 168     def nan(self, s):
 169         return self.create(s, self.P128, 1<<(self.e_start-1))
 170
 171     def inf(self, s):
 172         return self.create(s, self.P128, 0)
 173
 174     def zero(self, s):
 175         return self.create(s, self.N127, 0)
 176
 177
 178 class FPNumShift(FPNumBase):
 179     """ Floating-point Number Class for shifting
 180     """
 181     def __init__(self, mainm, op, inv, width, m_extra=True):
 182         FPNumBase.__init__(self, width, m_extra)
 183         self.latch_in = Signal()
 184         self.mainm = mainm
 185         self.inv = inv
 186         self.op = op
 187
 188     def elaborate(self, platform):
 189         m = FPNumBase.elaborate(self, platform)
 190
 191         m.d.comb += self.s.eq(op.s)
 192         m.d.comb += self.e.eq(op.e)
 193         m.d.comb += self.m.eq(op.m)
 194
 195         with self.mainm.State("align"):
 196             with m.If(self.e < self.inv.e):
 197                 m.d.sync += self.shift_down()
 198
 199         return m
 200
 201     def shift_down(self):
 202         """ shifts a mantissa down by one. exponent is increased to compensate
 203
 204             accuracy is lost as a result in the mantissa however there are 3
 205             guard bits (the latter of which is the "sticky" bit)
 206         """
 207         return [self.e.eq(self.e + 1),
 208                 self.m.eq(Cat(self.m[0] | self.m[1], self.m[2:], 0))
 209                ]
 210
 211     def shift_down_multi(self, diff):
 212         """ shifts a mantissa down. exponent is increased to compensate
 213
 214             accuracy is lost as a result in the mantissa however there are 3
 215             guard bits (the latter of which is the "sticky" bit)
 216
 217             this code works by variable-shifting the mantissa by up to
 218             its maximum bit-length: no point doing more (it'll still be
 219             zero).
 220
 221             the sticky bit is computed by shifting a batch of 1s by
 222             the same amount, which will introduce zeros.  it's then
 223             inverted and used as a mask to get the LSBs of the mantissa.
 224             those are then |'d into the sticky bit.
 225         """
 226         sm = MultiShift(self.width)
 227         mw = Const(self.m_width-1, len(diff))
 228         maxslen = Mux(diff > mw, mw, diff)
 229         rs = sm.rshift(self.m[1:], maxslen)
 230         maxsleni = mw - maxslen
 231         m_mask = sm.rshift(self.m1s[1:], maxsleni) # shift and invert
 232
 233         stickybits = reduce(or_, self.m[1:] & m_mask) | self.m[0]
 234         return [self.e.eq(self.e + diff),
 235                 self.m.eq(Cat(stickybits, rs))
 236                ]
 237
 238     def shift_up_multi(self, diff):
 239         """ shifts a mantissa up. exponent is decreased to compensate
 240         """
 241         sm = MultiShift(self.width)
 242         mw = Const(self.m_width, len(diff))
 243         maxslen = Mux(diff > mw, mw, diff)
 244
 245         return [self.e.eq(self.e - diff),
 246                 self.m.eq(sm.lshift(self.m, maxslen))
 247                ]
 248
 249 class FPNumIn(FPNumBase):
 250     """ Floating-point Number Class
 251
 252         Contains signals for an incoming copy of the value, decoded into
 253         sign / exponent / mantissa.
 254         Also contains encoding functions, creation and recognition of
 255         zero, NaN and inf (all signed)
 256
 257         Four extra bits are included in the mantissa: the top bit
 258         (m[-1]) is effectively a carry-overflow.  The other three are
 259         guard (m[2]), round (m[1]), and sticky (m[0])
 260     """
 261     def __init__(self, op, width, m_extra=True):
 262         FPNumBase.__init__(self, width, m_extra)
 263         self.latch_in = Signal()
 264         self.op = op
 265
 266     def elaborate(self, platform):
 267         m = FPNumBase.elaborate(self, platform)
 268
 269         #m.d.comb += self.latch_in.eq(self.op.ack & self.op.stb)
 270         #with m.If(self.latch_in):
 271         #    m.d.sync += self.decode(self.v)
 272
 273         return m
 274
 275     def decode(self, v):
 276         """ decodes a latched value into sign / exponent / mantissa
 277
 278             bias is subtracted here, from the exponent.  exponent
 279             is extended to 10 bits so that subtract 127 is done on
 280             a 10-bit number
 281         """
 282         args = [0] * self.m_extra + [v[0:self.e_start]] # pad with extra zeros
 283         #print ("decode", self.e_end)
 284         return [self.m.eq(Cat(*args)), # mantissa
 285                 self.e.eq(v[self.e_start:self.e_end] - self.P127), # exp
 286                 self.s.eq(v[-1]),                 # sign
 287                 ]
 288
 289     def shift_down(self):
 290         """ shifts a mantissa down by one. exponent is increased to compensate
 291
 292             accuracy is lost as a result in the mantissa however there are 3
 293             guard bits (the latter of which is the "sticky" bit)
 294         """
 295         return [self.e.eq(self.e + 1),
 296                 self.m.eq(Cat(self.m[0] | self.m[1], self.m[2:], 0))
 297                ]
 298
 299     def shift_down_multi(self, diff):
 300         """ shifts a mantissa down. exponent is increased to compensate
 301
 302             accuracy is lost as a result in the mantissa however there are 3
 303             guard bits (the latter of which is the "sticky" bit)
 304
 305             this code works by variable-shifting the mantissa by up to
 306             its maximum bit-length: no point doing more (it'll still be
 307             zero).
 308
 309             the sticky bit is computed by shifting a batch of 1s by
 310             the same amount, which will introduce zeros.  it's then
 311             inverted and used as a mask to get the LSBs of the mantissa.
 312             those are then |'d into the sticky bit.
 313         """
 314         sm = MultiShift(self.width)
 315         mw = Const(self.m_width-1, len(diff))
 316         maxslen = Mux(diff > mw, mw, diff)
 317         rs = sm.rshift(self.m[1:], maxslen)
 318         maxsleni = mw - maxslen
 319         m_mask = sm.rshift(self.m1s[1:], maxsleni) # shift and invert
 320
 321         stickybits = reduce(or_, self.m[1:] & m_mask) | self.m[0]
 322         return [self.e.eq(self.e + diff),
 323                 self.m.eq(Cat(stickybits, rs))
 324                ]
 325
 326     def shift_up_multi(self, diff):
 327         """ shifts a mantissa up. exponent is decreased to compensate
 328         """
 329         sm = MultiShift(self.width)
 330         mw = Const(self.m_width, len(diff))
 331         maxslen = Mux(diff > mw, mw, diff)
 332
 333         return [self.e.eq(self.e - diff),
 334                 self.m.eq(sm.lshift(self.m, maxslen))
 335                ]
 336
 337 class FPOp:
 338     def __init__(self, width):
 339         self.width = width
 340
 341         self.v   = Signal(width)
 342         self.stb = Signal()
 343         self.ack = Signal()
 344
 345     def chain_from(self, in_op):
 346         return [self.v.eq(in_op.v),          # receive value
 347                 self.stb.eq(in_op.stb),      # receive STB
 348                 in_op.ack.eq(self.ack), # send ACK
 349                ]
 350
 351     def ports(self):
 352         return [self.v, self.stb, self.ack]
 353
 354
 355 class Overflow:
 356     def __init__(self):
 357         self.guard = Signal(reset_less=True)     # tot[2]
 358         self.round_bit = Signal(reset_less=True) # tot[1]
 359         self.sticky = Signal(reset_less=True)    # tot[0]
 360         self.m0 = Signal(reset_less=True)        # mantissa zero bit
 361
 362         self.roundz = Signal(reset_less=True)
 363
 364     def elaborate(self, platform):
 365         m = Module()
 366         m.d.comb += self.roundz.eq(self.guard & \
 367                                    (self.round_bit | self.sticky | self.m0))
 368         return m
 369
 370
 371 class FPBase:
 372     """ IEEE754 Floating Point Base Class
 373
 374         contains common functions for FP manipulation, such as
 375         extracting and packing operands, normalisation, denormalisation,
 376         rounding etc.
 377     """
 378
 379     def get_op(self, m, op, v, next_state):
 380         """ this function moves to the next state and copies the operand
 381             when both stb and ack are 1.
 382             acknowledgement is sent by setting ack to ZERO.
 383         """
 384         with m.If((op.ack) & (op.stb)):
 385             m.next = next_state
 386             m.d.sync += [
 387                 # op is latched in from FPNumIn class on same ack/stb
 388                 v.decode(op.v),
 389                 op.ack.eq(0)
 390             ]
 391         with m.Else():
 392             m.d.sync += op.ack.eq(1)
 393
 394     def denormalise(self, m, a):
 395         """ denormalises a number.  this is probably the wrong name for
 396             this function.  for normalised numbers (exponent != minimum)
 397             one *extra* bit (the implicit 1) is added *back in*.
 398             for denormalised numbers, the mantissa is left alone
 399             and the exponent increased by 1.
 400
 401             both cases *effectively multiply the number stored by 2*,
 402             which has to be taken into account when extracting the result.
 403         """
 404         with m.If(a.e == a.N127):
 405             m.d.sync += a.e.eq(a.N126) # limit a exponent
 406         with m.Else():
 407             m.d.sync += a.m[-1].eq(1) # set top mantissa bit
 408
 409     def op_normalise(self, m, op, next_state):
 410         """ operand normalisation
 411             NOTE: just like "align", this one keeps going round every clock
 412                   until the result's exponent is within acceptable "range"
 413         """
 414         with m.If((op.m[-1] == 0)): # check last bit of mantissa
 415             m.d.sync +=[
 416                 op.e.eq(op.e - 1),  # DECREASE exponent
 417                 op.m.eq(op.m << 1), # shift mantissa UP
 418             ]
 419         with m.Else():
 420             m.next = next_state
 421
 422     def normalise_1(self, m, z, of, next_state):
 423         """ first stage normalisation
 424
 425             NOTE: just like "align", this one keeps going round every clock
 426                   until the result's exponent is within acceptable "range"
 427             NOTE: the weirdness of reassigning guard and round is due to
 428                   the extra mantissa bits coming from tot[0..2]
 429         """
 430         with m.If((z.m[-1] == 0) & (z.e > z.N126)):
 431             m.d.sync += [
 432                 z.e.eq(z.e - 1),  # DECREASE exponent
 433                 z.m.eq(z.m << 1), # shift mantissa UP
 434                 z.m[0].eq(of.guard),       # steal guard bit (was tot[2])
 435                 of.guard.eq(of.round_bit), # steal round_bit (was tot[1])
 436                 of.round_bit.eq(0),        # reset round bit
 437                 of.m0.eq(of.guard),
 438             ]
 439         with m.Else():
 440             m.next = next_state
 441
 442     def normalise_2(self, m, z, of, next_state):
 443         """ second stage normalisation
 444
 445             NOTE: just like "align", this one keeps going round every clock
 446                   until the result's exponent is within acceptable "range"
 447             NOTE: the weirdness of reassigning guard and round is due to
 448                   the extra mantissa bits coming from tot[0..2]
 449         """
 450         with m.If(z.e < z.N126):
 451             m.d.sync +=[
 452                 z.e.eq(z.e + 1),  # INCREASE exponent
 453                 z.m.eq(z.m >> 1), # shift mantissa DOWN
 454                 of.guard.eq(z.m[0]),
 455                 of.m0.eq(z.m[1]),
 456                 of.round_bit.eq(of.guard),
 457                 of.sticky.eq(of.sticky | of.round_bit)
 458             ]
 459         with m.Else():
 460             m.next = next_state
 461
 462     def roundz(self, m, z, of, next_state):
 463         """ performs rounding on the output.  TODO: different kinds of rounding
 464         """
 465         m.next = next_state
 466         with m.If(of.roundz):
 467             m.d.sync += z.m.eq(z.m + 1) # mantissa rounds up
 468             with m.If(z.m == z.m1s): # all 1s
 469                 m.d.sync += z.e.eq(z.e + 1) # exponent rounds up
 470
 471     def corrections(self, m, z, next_state):
 472         """ denormalisation and sign-bug corrections
 473         """
 474         m.next = next_state
 475         # denormalised, correct exponent to zero
 476         with m.If(z.is_denormalised):
 477             m.d.sync += z.e.eq(z.N127)
 478
 479     def pack(self, m, z, next_state):
 480         """ packs the result into the output (detects overflow->Inf)
 481         """
 482         m.next = next_state
 483         # if overflow occurs, return inf
 484         with m.If(z.is_overflowed):
 485             m.d.sync += z.inf(z.s)
 486         with m.Else():
 487             m.d.sync += z.create(z.s, z.e, z.m)
 488
 489     def put_z(self, m, z, out_z, next_state):
 490         """ put_z: stores the result in the output.  raises stb and waits
 491             for ack to be set to 1 before moving to the next state.
 492             resets stb back to zero when that occurs, as acknowledgement.
 493         """
 494         m.d.sync += [
 495           out_z.v.eq(z.v)
 496         ]
 497         with m.If(out_z.stb & out_z.ack):
 498             m.d.sync += out_z.stb.eq(0)
 499             m.next = next_state
 500         with m.Else():
 501             m.d.sync += out_z.stb.eq(1)
 502
 503