src/add/nmigen_add_experiment.py

   1 # IEEE Floating Point Adder (Single Precision)
   2 # Copyright (C) Jonathan P Dawson 2013
   3 # 2013-12-12
   4
   5 from nmigen import Module, Signal, Cat
   6 from nmigen.cli import main, verilog
   7
   8 from fpbase import FPNumIn, FPNumOut, FPOp, Overflow, FPBase, FPNumBase
   9
  10
  11 class FPState(FPBase):
  12     def __init__(self, state_from):
  13         self.state_from = state_from
  14
  15     def set_inputs(self, inputs):
  16         self.inputs = inputs
  17         for k,v in inputs.items():
  18             setattr(self, k, v)
  19
  20     def set_outputs(self, outputs):
  21         self.outputs = outputs
  22         for k,v in outputs.items():
  23             setattr(self, k, v)
  24
  25
  26 class FPGetOpA(FPState):
  27     """ gets operand a
  28     """
  29
  30     def __init__(self, in_a, width):
  31         FPState.__init__(self, "get_a")
  32         self.in_a = in_a
  33         self.a = FPNumIn(in_a, width)
  34
  35     def action(self, m):
  36         self.get_op(m, self.in_a, self.a, "get_b")
  37
  38
  39 class FPGetOpB(FPState):
  40     """ gets operand b
  41     """
  42
  43     def action(self, m):
  44         self.get_op(m, self.in_b, self.b, "special_cases")
  45
  46
  47 class FPAddSpecialCasesMod:
  48     """ special cases: NaNs, infs, zeros, denormalised
  49         NOTE: some of these are unique to add.  see "Special Operations"
  50         https://steve.hollasch.net/cgindex/coding/ieeefloat.html
  51     """
  52
  53     def __init__(self, width):
  54         self.in_a = FPNumBase(width)
  55         self.in_b = FPNumBase(width)
  56         self.out_z = FPNumOut(width, False)
  57         self.out_do_z = Signal(reset_less=True)
  58
  59     def setup(self, m, in_a, in_b, out_z, out_do_z):
  60         """ links module to inputs and outputs
  61         """
  62         m.d.comb += self.in_a.copy(in_a)
  63         m.d.comb += self.in_b.copy(in_b)
  64         m.d.comb += out_z.v.eq(self.out_z.v)
  65         m.d.comb += out_do_z.eq(self.out_do_z)
  66
  67     def elaborate(self, platform):
  68         m = Module()
  69
  70         m.submodules.sc_in_a = self.in_a
  71         m.submodules.sc_in_b = self.in_b
  72         m.submodules.sc_out_z = self.out_z
  73
  74         s_nomatch = Signal()
  75         m.d.comb += s_nomatch.eq(self.in_a.s != self.in_b.s)
  76
  77         m_match = Signal()
  78         m.d.comb += m_match.eq(self.in_a.m == self.in_b.m)
  79
  80         # if a is NaN or b is NaN return NaN
  81         with m.If(self.in_a.is_nan | self.in_b.is_nan):
  82             m.d.comb += self.out_do_z.eq(1)
  83             m.d.comb += self.out_z.nan(1)
  84
  85         # XXX WEIRDNESS for FP16 non-canonical NaN handling
  86         # under review
  87
  88         ## if a is zero and b is NaN return -b
  89         #with m.If(a.is_zero & (a.s==0) & b.is_nan):
  90         #    m.d.comb += self.out_do_z.eq(1)
  91         #    m.d.comb += z.create(b.s, b.e, Cat(b.m[3:-2], ~b.m[0]))
  92
  93         ## if b is zero and a is NaN return -a
  94         #with m.Elif(b.is_zero & (b.s==0) & a.is_nan):
  95         #    m.d.comb += self.out_do_z.eq(1)
  96         #    m.d.comb += z.create(a.s, a.e, Cat(a.m[3:-2], ~a.m[0]))
  97
  98         ## if a is -zero and b is NaN return -b
  99         #with m.Elif(a.is_zero & (a.s==1) & b.is_nan):
 100         #    m.d.comb += self.out_do_z.eq(1)
 101         #    m.d.comb += z.create(a.s & b.s, b.e, Cat(b.m[3:-2], 1))
 102
 103         ## if b is -zero and a is NaN return -a
 104         #with m.Elif(b.is_zero & (b.s==1) & a.is_nan):
 105         #    m.d.comb += self.out_do_z.eq(1)
 106         #    m.d.comb += z.create(a.s & b.s, a.e, Cat(a.m[3:-2], 1))
 107
 108         # if a is inf return inf (or NaN)
 109         with m.Elif(self.in_a.is_inf):
 110             m.d.comb += self.out_do_z.eq(1)
 111             m.d.comb += self.out_z.inf(self.in_a.s)
 112             # if a is inf and signs don't match return NaN
 113             with m.If(self.in_b.exp_128 & s_nomatch):
 114                 m.d.comb += self.out_z.nan(1)
 115
 116         # if b is inf return inf
 117         with m.Elif(self.in_b.is_inf):
 118             m.d.comb += self.out_do_z.eq(1)
 119             m.d.comb += self.out_z.inf(self.in_b.s)
 120
 121         # if a is zero and b zero return signed-a/b
 122         with m.Elif(self.in_a.is_zero & self.in_b.is_zero):
 123             m.d.comb += self.out_do_z.eq(1)
 124             m.d.comb += self.out_z.create(self.in_a.s & self.in_b.s,
 125                                           self.in_b.e,
 126                                           self.in_b.m[3:-1])
 127
 128         # if a is zero return b
 129         with m.Elif(self.in_a.is_zero):
 130             m.d.comb += self.out_do_z.eq(1)
 131             m.d.comb += self.out_z.create(self.in_b.s, self.in_b.e,
 132                                       self.in_b.m[3:-1])
 133
 134         # if b is zero return a
 135         with m.Elif(self.in_b.is_zero):
 136             m.d.comb += self.out_do_z.eq(1)
 137             m.d.comb += self.out_z.create(self.in_a.s, self.in_a.e,
 138                                       self.in_a.m[3:-1])
 139
 140         # if a equal to -b return zero (+ve zero)
 141         with m.Elif(s_nomatch & m_match & (self.in_a.e == self.in_b.e)):
 142             m.d.comb += self.out_do_z.eq(1)
 143             m.d.comb += self.out_z.zero(0)
 144
 145         # Denormalised Number checks
 146         with m.Else():
 147             m.d.comb += self.out_do_z.eq(0)
 148
 149         return m
 150
 151
 152 class FPAddSpecialCases(FPState):
 153     """ special cases: NaNs, infs, zeros, denormalised
 154         NOTE: some of these are unique to add.  see "Special Operations"
 155         https://steve.hollasch.net/cgindex/coding/ieeefloat.html
 156     """
 157
 158     def __init__(self, width):
 159         FPState.__init__(self, "special_cases")
 160         self.mod = FPAddSpecialCasesMod(width)
 161         self.out_z = FPNumOut(width, False)
 162         self.out_do_z = Signal(reset_less=True)
 163
 164     def action(self, m):
 165         with m.If(self.out_do_z):
 166             m.d.sync += self.z.v.eq(self.out_z.v) # only take the output
 167             m.next = "put_z"
 168         with m.Else():
 169             m.next = "denormalise"
 170
 171
 172 class FPAddDeNorm(FPState):
 173
 174     def action(self, m):
 175         # Denormalised Number checks
 176         m.next = "align"
 177         self.denormalise(m, self.a)
 178         self.denormalise(m, self.b)
 179
 180
 181 class FPAddAlignMultiMod(FPState):
 182
 183     def __init__(self, width):
 184         self.in_a = FPNumBase(width)
 185         self.in_b = FPNumBase(width)
 186         self.out_a = FPNumIn(None, width)
 187         self.out_b = FPNumIn(None, width)
 188         self.exp_eq = Signal(reset_less=True)
 189
 190     def setup(self, m, in_a, in_b, out_a, out_b, exp_eq):
 191         """ links module to inputs and outputs
 192         """
 193         m.d.comb += self.in_a.copy(in_a)
 194         m.d.comb += self.in_b.copy(in_b)
 195         m.d.comb += out_a.copy(self.out_a)
 196         m.d.comb += out_b.copy(self.out_b)
 197         m.d.comb += exp_eq.eq(self.exp_eq)
 198
 199     def elaborate(self, platform):
 200         # This one however (single-cycle) will do the shift
 201         # in one go.
 202
 203         m = Module()
 204
 205         #m.submodules.align_in_a = self.in_a
 206         #m.submodules.align_in_b = self.in_b
 207         m.submodules.align_out_a = self.out_a
 208         m.submodules.align_out_b = self.out_b
 209
 210         # NOTE: this does *not* do single-cycle multi-shifting,
 211         #       it *STAYS* in the align state until exponents match
 212
 213         # exponent of a greater than b: shift b down
 214         m.d.comb += self.exp_eq.eq(0)
 215         m.d.comb += self.out_a.copy(self.in_a)
 216         m.d.comb += self.out_b.copy(self.in_b)
 217         with m.If(self.in_a.e > self.in_b.e):
 218             m.d.comb += self.out_b.shift_down(self.in_b)
 219         # exponent of b greater than a: shift a down
 220         with m.Elif(self.in_a.e < self.in_b.e):
 221             m.d.comb += self.out_a.shift_down(self.in_a)
 222         # exponents equal: move to next stage.
 223         with m.Else():
 224             m.d.comb += self.exp_eq.eq(1)
 225         return m
 226
 227
 228 class FPAddAlignMulti(FPState):
 229
 230     def __init__(self, width):
 231         FPState.__init__(self, "align")
 232         self.mod = FPAddAlignMultiMod(width)
 233         self.out_a = FPNumIn(None, width)
 234         self.out_b = FPNumIn(None, width)
 235         self.exp_eq = Signal(reset_less=True)
 236
 237     def action(self, m):
 238         m.d.sync += self.a.copy(self.out_a)
 239         m.d.sync += self.b.copy(self.out_b)
 240         with m.If(self.exp_eq):
 241             m.next = "add_0"
 242
 243
 244 class FPAddAlignSingleMod:
 245
 246     def __init__(self, width):
 247         self.in_a = FPNumBase(width)
 248         self.in_b = FPNumBase(width)
 249         self.out_a = FPNumIn(None, width)
 250         self.out_b = FPNumIn(None, width)
 251         #self.out_a = FPNumBase(width)
 252         #self.out_b = FPNumBase(width)
 253
 254     def setup(self, m, in_a, in_b, out_a, out_b):
 255         """ links module to inputs and outputs
 256         """
 257         m.d.comb += self.in_a.copy(in_a)
 258         m.d.comb += self.in_b.copy(in_b)
 259         m.d.comb += out_a.copy(self.out_a)
 260         m.d.comb += out_b.copy(self.out_b)
 261
 262     def elaborate(self, platform):
 263         # This one however (single-cycle) will do the shift
 264         # in one go.
 265
 266         m = Module()
 267
 268         #m.submodules.align_in_a = self.in_a
 269         #m.submodules.align_in_b = self.in_b
 270         m.submodules.align_out_a = self.out_a
 271         m.submodules.align_out_b = self.out_b
 272
 273         # XXX TODO: the shifter used here is quite expensive
 274         # having only one would be better
 275
 276         ediff = Signal((len(self.in_a.e), True), reset_less=True)
 277         ediffr = Signal((len(self.in_a.e), True), reset_less=True)
 278         m.d.comb += ediff.eq(self.in_a.e - self.in_b.e)
 279         m.d.comb += ediffr.eq(self.in_b.e - self.in_a.e)
 280         m.d.comb += self.out_a.copy(self.in_a)
 281         m.d.comb += self.out_b.copy(self.in_b)
 282         with m.If(ediff > 0):
 283             m.d.comb += self.out_b.shift_down_multi(ediff)
 284         # exponent of b greater than a: shift a down
 285         with m.Elif(ediff < 0):
 286             m.d.comb += self.out_a.shift_down_multi(ediffr)
 287         return m
 288
 289
 290 class FPAddAlignSingle(FPState):
 291
 292     def __init__(self, width):
 293         FPState.__init__(self, "align")
 294         self.mod = FPAddAlignSingleMod(width)
 295         self.out_a = FPNumIn(None, width)
 296         self.out_b = FPNumIn(None, width)
 297
 298     def action(self, m):
 299         m.d.sync += self.a.copy(self.out_a)
 300         m.d.sync += self.b.copy(self.out_b)
 301         m.next = "add_0"
 302
 303
 304 class FPAddStage0Mod:
 305
 306     def __init__(self, width):
 307         self.in_a = FPNumBase(width)
 308         self.in_b = FPNumBase(width)
 309         self.in_z = FPNumBase(width, False)
 310         self.out_z = FPNumBase(width, False)
 311         self.out_tot = Signal(self.out_z.m_width + 4, reset_less=True)
 312
 313     def setup(self, m, in_a, in_b, in_z, out_z, out_tot):
 314         """ links module to inputs and outputs
 315         """
 316         m.d.comb += self.in_a.copy(in_a)
 317         m.d.comb += self.in_b.copy(in_b)
 318         m.d.comb += self.in_z.copy(in_z)
 319         m.d.comb += out_z.copy(self.out_z)
 320         m.d.comb += out_tot.eq(self.out_tot)
 321
 322     def elaborate(self, platform):
 323         m = Module()
 324         #m.submodules.add0_in_a = self.in_a
 325         #m.submodules.add0_in_b = self.in_b
 326         #m.submodules.add0_in_z = self.in_z
 327         #m.submodules.add0_out_z = self.out_z
 328
 329         m.d.comb += self.out_z.e.eq(self.in_a.e)
 330         # same-sign (both negative or both positive) add mantissas
 331         with m.If(self.in_a.s == self.in_b.s):
 332             m.d.comb += [
 333                 self.out_tot.eq(Cat(self.in_a.m, 0) + Cat(self.in_b.m, 0)),
 334                 self.out_z.s.eq(self.in_a.s)
 335             ]
 336         # a mantissa greater than b, use a
 337         with m.Elif(self.in_a.m >= self.in_b.m):
 338             m.d.comb += [
 339                 self.out_tot.eq(Cat(self.in_a.m, 0) - Cat(self.in_b.m, 0)),
 340                 self.out_z.s.eq(self.in_a.s)
 341             ]
 342         # b mantissa greater than a, use b
 343         with m.Else():
 344             m.d.comb += [
 345                 self.out_tot.eq(Cat(self.in_b.m, 0) - Cat(self.in_a.m, 0)),
 346                 self.out_z.s.eq(self.in_b.s)
 347         ]
 348         return m
 349
 350
 351 class FPAddStage0(FPState):
 352     """ First stage of add.  covers same-sign (add) and subtract
 353         special-casing when mantissas are greater or equal, to
 354         give greatest accuracy.
 355     """
 356
 357     def __init__(self, width):
 358         FPState.__init__(self, "add_0")
 359         self.mod = FPAddStage0Mod(width)
 360         self.out_z = FPNumBase(width, False)
 361         self.out_tot = Signal(self.out_z.m_width + 4, reset_less=True)
 362
 363     def action(self, m):
 364         m.next = "add_1"
 365         m.d.sync += self.z.copy(self.out_z)
 366
 367
 368 class FPAddStage1Mod(FPState):
 369     """ Second stage of add: preparation for normalisation.
 370         detects when tot sum is too big (tot[27] is kinda a carry bit)
 371     """
 372
 373     def __init__(self, width):
 374         self.out_norm = Signal(reset_less=True)
 375         self.in_z = FPNumBase(width, False)
 376         self.in_tot = Signal(self.in_z.m_width + 4, reset_less=True)
 377         self.out_z = FPNumBase(width, False)
 378         self.out_of = Overflow()
 379
 380     def setup(self, m, in_tot, in_z, out_z, out_of):
 381         """ links module to inputs and outputs
 382         """
 383         m.d.comb += self.in_z.copy(in_z)
 384         m.d.comb += self.in_tot.eq(in_tot)
 385         m.d.comb += out_z.copy(self.out_z)
 386         m.d.comb += out_of.copy(self.out_of)
 387
 388     def elaborate(self, platform):
 389         m = Module()
 390         m.submodules.add1_out_overflow = self.out_of
 391         #m.submodules.norm1_in_overflow = self.in_of
 392         #m.submodules.norm1_in_z = self.in_z
 393         #m.submodules.norm1_out_z = self.out_z
 394         m.d.comb += self.out_z.copy(self.in_z)
 395         # tot[27] gets set when the sum overflows. shift result down
 396         with m.If(self.in_tot[-1]):
 397             m.d.comb += [
 398                 self.out_z.m.eq(self.in_tot[4:]),
 399                 self.out_of.m0.eq(self.in_tot[4]),
 400                 self.out_of.guard.eq(self.in_tot[3]),
 401                 self.out_of.round_bit.eq(self.in_tot[2]),
 402                 self.out_of.sticky.eq(self.in_tot[1] | self.in_tot[0]),
 403                 self.out_z.e.eq(self.in_z.e + 1)
 404         ]
 405         # tot[27] zero case
 406         with m.Else():
 407             m.d.comb += [
 408                 self.out_z.m.eq(self.in_tot[3:]),
 409                 self.out_of.m0.eq(self.in_tot[3]),
 410                 self.out_of.guard.eq(self.in_tot[2]),
 411                 self.out_of.round_bit.eq(self.in_tot[1]),
 412                 self.out_of.sticky.eq(self.in_tot[0])
 413         ]
 414         return m
 415
 416
 417 class FPAddStage1(FPState):
 418
 419     def __init__(self, width):
 420         FPState.__init__(self, "add_1")
 421         self.mod = FPAddStage1Mod(width)
 422         self.out_z = FPNumBase(width, False)
 423         self.out_of = Overflow()
 424
 425     def action(self, m):
 426         #m.d.sync += self.of.copy(self.out_of)
 427         m.d.sync += self.z.copy(self.out_z)
 428         m.next = "normalise_1"
 429
 430
 431 class FPNorm1Mod:
 432
 433     def __init__(self, width):
 434         self.out_norm = Signal(reset_less=True)
 435         self.in_z = FPNumBase(width, False)
 436         self.out_z = FPNumBase(width, False)
 437         self.in_of = Overflow()
 438         self.out_of = Overflow()
 439
 440     def setup(self, m, in_z, out_z, in_of, out_of, out_norm):
 441         """ links module to inputs and outputs
 442         """
 443         m.d.comb += self.in_z.copy(in_z)
 444         m.d.comb += out_z.copy(self.out_z)
 445         m.d.comb += self.in_of.copy(in_of)
 446         m.d.comb += out_of.copy(self.out_of)
 447         m.d.comb += out_norm.eq(self.out_norm)
 448
 449     def elaborate(self, platform):
 450         m = Module()
 451         m.submodules.norm1_in_overflow = self.in_of
 452         m.submodules.norm1_out_overflow = self.out_of
 453         m.submodules.norm1_in_z = self.in_z
 454         m.submodules.norm1_out_z = self.out_z
 455         m.d.comb += self.out_z.copy(self.in_z)
 456         m.d.comb += self.out_of.copy(self.in_of)
 457         m.d.comb += self.out_norm.eq((self.in_z.m_msbzero) & \
 458                                      (self.in_z.exp_gt_n126))
 459         with m.If(self.out_norm):
 460             m.d.comb += [
 461                 self.out_z.e.eq(self.in_z.e - 1),  # DECREASE exponent
 462                 self.out_z.m.eq(self.in_z.m << 1), # shift mantissa UP
 463                 self.out_z.m[0].eq(self.in_of.guard), # steal guard (was tot[2])
 464                 self.out_of.guard.eq(self.in_of.round_bit), # round (was tot[1])
 465                 self.out_of.round_bit.eq(0),        # reset round bit
 466                 self.out_of.m0.eq(self.in_of.guard),
 467             ]
 468
 469         return m
 470
 471
 472 class FPNorm1(FPState):
 473
 474     def __init__(self, width):
 475         FPState.__init__(self, "normalise_1")
 476         self.mod = FPNorm1Mod(width)
 477         self.out_norm = Signal(reset_less=True)
 478         self.out_z = FPNumBase(width)
 479         self.out_of = Overflow()
 480
 481     def action(self, m):
 482         #m.d.sync += self.of.copy(self.out_of)
 483         m.d.sync += self.z.copy(self.out_z)
 484         with m.If(~self.out_norm):
 485             m.next = "normalise_2"
 486
 487
 488 class FPNorm2Mod:
 489
 490     def __init__(self, width):
 491         self.out_norm = Signal(reset_less=True)
 492         self.in_z = FPNumBase(width, False)
 493         self.out_z = FPNumBase(width, False)
 494         self.in_of = Overflow()
 495         self.out_of = Overflow()
 496
 497     def setup(self, m, in_z, out_z, in_of, out_of, out_norm):
 498         """ links module to inputs and outputs
 499         """
 500         m.d.comb += self.in_z.copy(in_z)
 501         m.d.comb += out_z.copy(self.out_z)
 502         m.d.comb += self.in_of.copy(in_of)
 503         m.d.comb += out_of.copy(self.out_of)
 504         m.d.comb += out_norm.eq(self.out_norm)
 505
 506     def elaborate(self, platform):
 507         m = Module()
 508         m.submodules.norm2_in_overflow = self.in_of
 509         m.submodules.norm2_out_overflow = self.out_of
 510         m.submodules.norm2_in_z = self.in_z
 511         m.submodules.norm2_out_z = self.out_z
 512         m.d.comb += self.out_z.copy(self.in_z)
 513         m.d.comb += self.out_of.copy(self.in_of)
 514         m.d.comb += self.out_norm.eq(self.in_z.exp_lt_n126)
 515         with m.If(self.out_norm):
 516             m.d.comb += [
 517                 self.out_z.e.eq(self.in_z.e + 1),  # INCREASE exponent
 518                 self.out_z.m.eq(self.in_z.m >> 1), # shift mantissa DOWN
 519                 self.out_of.guard.eq(self.in_z.m[0]),
 520                 self.out_of.m0.eq(self.in_z.m[1]),
 521                 self.out_of.round_bit.eq(self.in_of.guard),
 522                 self.out_of.sticky.eq(self.in_of.sticky | self.in_of.round_bit)
 523             ]
 524
 525         return m
 526
 527
 528 class FPNorm2(FPState):
 529
 530     def __init__(self, width):
 531         FPState.__init__(self, "normalise_2")
 532         self.mod = FPNorm2Mod(width)
 533         self.out_norm = Signal(reset_less=True)
 534         self.out_z = FPNumBase(width)
 535         self.out_of = Overflow()
 536
 537     def action(self, m):
 538         m.submodules.norm_of = self.out_of
 539         #m.d.sync += self.of.copy(self.out_of)
 540         m.d.sync += self.z.copy(self.out_z)
 541         with m.If(~self.out_norm):
 542             m.next = "round"
 543
 544
 545 class FPRoundMod:
 546
 547     def __init__(self, width):
 548         self.in_roundz = Signal(reset_less=True)
 549         self.in_z = FPNumBase(width, False)
 550         self.out_z = FPNumBase(width, False)
 551
 552     def setup(self, m, in_z, out_z, in_of):
 553         """ links module to inputs and outputs
 554         """
 555         m.d.comb += self.in_z.copy(in_z)
 556         m.d.comb += out_z.copy(self.out_z)
 557         m.d.comb += self.in_roundz.eq(in_of.roundz)
 558
 559     def elaborate(self, platform):
 560         m = Module()
 561         m.d.comb += self.out_z.copy(self.in_z)
 562         m.submodules.round_in_z = self.in_z
 563         m.submodules.round_out_z = self.out_z
 564         with m.If(self.in_roundz):
 565             m.d.comb += self.out_z.m.eq(self.in_z.m + 1) # mantissa rounds up
 566             with m.If(self.in_z.m == self.in_z.m1s): # all 1s
 567                 m.d.comb += self.out_z.e.eq(self.in_z.e + 1) # exponent up
 568         return m
 569
 570
 571 class FPRound(FPState):
 572
 573     def __init__(self, width):
 574         FPState.__init__(self, "round")
 575         self.mod = FPRoundMod(width)
 576         self.out_z = FPNumBase(width)
 577
 578     def action(self, m):
 579         m.d.sync += self.z.copy(self.out_z)
 580         m.next = "corrections"
 581
 582
 583 class FPCorrectionsMod:
 584
 585     def __init__(self, width):
 586         self.in_z = FPNumOut(width, False)
 587         self.out_z = FPNumOut(width, False)
 588
 589     def setup(self, m, in_z, out_z):
 590         """ links module to inputs and outputs
 591         """
 592         m.d.comb += self.in_z.copy(in_z)
 593         m.d.comb += out_z.copy(self.out_z)
 594
 595     def elaborate(self, platform):
 596         m = Module()
 597         m.submodules.corr_in_z = self.in_z
 598         m.submodules.corr_out_z = self.out_z
 599         m.d.comb += self.out_z.copy(self.in_z)
 600         with m.If(self.in_z.is_denormalised):
 601             m.d.comb += self.out_z.e.eq(self.in_z.N127)
 602
 603         #        with m.If(self.in_z.is_overflowed):
 604         #            m.d.comb += self.out_z.inf(self.in_z.s)
 605         #        with m.Else():
 606         #            m.d.comb += self.out_z.create(self.in_z.s, self.in_z.e, self.in_z.m)
 607         return m
 608
 609
 610 class FPCorrections(FPState):
 611
 612     def __init__(self, width):
 613         FPState.__init__(self, "corrections")
 614         self.mod = FPCorrectionsMod(width)
 615         self.out_z = FPNumBase(width)
 616
 617     def action(self, m):
 618         m.d.sync += self.z.copy(self.out_z)
 619         m.next = "pack"
 620
 621
 622 class FPPackMod:
 623
 624     def __init__(self, width):
 625         self.in_z = FPNumOut(width, False)
 626         self.out_z = FPNumOut(width, False)
 627
 628     def setup(self, m, in_z, out_z):
 629         """ links module to inputs and outputs
 630         """
 631         m.d.comb += self.in_z.copy(in_z)
 632         m.d.comb += out_z.v.eq(self.out_z.v)
 633
 634     def elaborate(self, platform):
 635         m = Module()
 636         m.submodules.pack_in_z = self.in_z
 637         with m.If(self.in_z.is_overflowed):
 638             m.d.comb += self.out_z.inf(self.in_z.s)
 639         with m.Else():
 640             m.d.comb += self.out_z.create(self.in_z.s, self.in_z.e, self.in_z.m)
 641         return m
 642
 643
 644 class FPPack(FPState):
 645
 646     def __init__(self, width):
 647         FPState.__init__(self, "pack")
 648         self.mod = FPPackMod(width)
 649         self.out_z = FPNumOut(width, False)
 650
 651     def action(self, m):
 652         m.d.sync += self.z.v.eq(self.out_z.v)
 653         m.next = "put_z"
 654
 655
 656 class FPPutZ(FPState):
 657
 658     def action(self, m):
 659         self.put_z(m, self.z, self.out_z, "get_a")
 660
 661
 662 class FPADD:
 663
 664     def __init__(self, width, single_cycle=False):
 665         self.width = width
 666         self.single_cycle = single_cycle
 667
 668         self.in_a  = FPOp(width)
 669         self.in_b  = FPOp(width)
 670         self.out_z = FPOp(width)
 671
 672         self.states = []
 673
 674     def add_state(self, state):
 675         self.states.append(state)
 676         return state
 677
 678     def get_fragment(self, platform=None):
 679         """ creates the HDL code-fragment for FPAdd
 680         """
 681         m = Module()
 682
 683         # Latches
 684         #a = FPNumIn(self.in_a, self.width)
 685         b = FPNumIn(self.in_b, self.width)
 686         z = FPNumOut(self.width, False)
 687
 688         m.submodules.fpnum_b = b
 689         m.submodules.fpnum_z = z
 690
 691         w = z.m_width + 4
 692
 693         geta = self.add_state(FPGetOpA(self.in_a, self.width))
 694         #geta.set_inputs({"in_a": self.in_a})
 695         #geta.set_outputs({"a": a})
 696         a = geta.a
 697         # XXX m.d.comb += a.v.eq(self.in_a.v) # links in_a to a
 698         m.submodules.fpnum_a = a
 699
 700         getb = self.add_state(FPGetOpB("get_b"))
 701         getb.set_inputs({"in_b": self.in_b})
 702         getb.set_outputs({"b": b})
 703         # XXX m.d.comb += b.v.eq(self.in_b.v) # links in_b to b
 704
 705         sc = self.add_state(FPAddSpecialCases(self.width))
 706         sc.set_inputs({"a": a, "b": b})
 707         sc.set_outputs({"z": z})
 708         sc.mod.setup(m, a, b, sc.out_z, sc.out_do_z)
 709         m.submodules.specialcases = sc.mod
 710
 711         dn = self.add_state(FPAddDeNorm("denormalise"))
 712         dn.set_inputs({"a": a, "b": b})
 713         dn.set_outputs({"a": a, "b": b}) # XXX outputs same as inputs
 714
 715         if self.single_cycle:
 716             alm = self.add_state(FPAddAlignSingle(self.width))
 717             alm.set_inputs({"a": a, "b": b})
 718             alm.set_outputs({"a": a, "b": b}) # XXX outputs same as inputs
 719             alm.mod.setup(m, a, b, alm.out_a, alm.out_b)
 720         else:
 721             alm = self.add_state(FPAddAlignMulti(self.width))
 722             alm.set_inputs({"a": a, "b": b})
 723             alm.set_outputs({"a": a, "b": b}) # XXX outputs same as inputs
 724             alm.mod.setup(m, a, b, alm.out_a, alm.out_b, alm.exp_eq)
 725         m.submodules.align = alm.mod
 726
 727         add0 = self.add_state(FPAddStage0(self.width))
 728         add0.set_inputs({"a": a, "b": b})
 729         add0.set_outputs({"z": z})
 730         add0.mod.setup(m, a, b, z, add0.out_z, add0.out_tot)
 731         m.submodules.add0 = add0.mod
 732
 733         add1 = self.add_state(FPAddStage1(self.width))
 734         add1.set_inputs({"tot": add0.out_tot, "z": add0.out_z})
 735         add1.set_outputs({"z": z})  # XXX Z as output
 736         add1.mod.setup(m, add0.out_tot, add0.out_z, add1.out_z, add1.out_of)
 737         m.submodules.add1 = add1.mod
 738
 739         n1 = self.add_state(FPNorm1(self.width))
 740         n1.set_inputs({"z": z, "of": add1.out_of})  # XXX Z as output
 741         n1.set_outputs({"z": z})  # XXX Z as output
 742         n1.mod.setup(m, z, n1.out_z, add1.out_of, n1.out_of, n1.out_norm)
 743         m.submodules.normalise_1 = n1.mod
 744
 745         n2 = self.add_state(FPNorm2(self.width))
 746         n2.set_inputs({"z": n1.out_z, "of": n1.out_of})
 747         n2.set_outputs({"z": z})
 748         n2.mod.setup(m, n1.out_z, n2.out_z, n1.out_of, n2.out_of, n2.out_norm)
 749         m.submodules.normalise_2 = n2.mod
 750
 751         rn = self.add_state(FPRound(self.width))
 752         rn.set_inputs({"z": n2.out_z, "of": n2.out_of})
 753         rn.set_outputs({"z": z})
 754         rn.mod.setup(m, n2.out_z, rn.out_z, n2.out_of)
 755         m.submodules.roundz = rn.mod
 756
 757         cor = self.add_state(FPCorrections(self.width))
 758         cor.set_inputs({"z": rn.out_z})  # XXX Z as output
 759         cor.set_outputs({"z": z})  # XXX Z as output
 760         cor.mod.setup(m, z, cor.out_z)
 761         m.submodules.corrections = cor.mod
 762
 763         pa = self.add_state(FPPack(self.width))
 764         pa.set_inputs({"z": cor.out_z})  # XXX Z as output
 765         pa.set_outputs({"z": z})  # XXX Z as output
 766         pa.mod.setup(m, cor.out_z, pa.out_z)
 767         m.submodules.pack = pa.mod
 768
 769         pz = self.add_state(FPPutZ("put_z"))
 770         pz.set_inputs({"z": z})
 771         pz.set_outputs({"out_z": self.out_z})
 772
 773         with m.FSM() as fsm:
 774
 775             for state in self.states:
 776                 with m.State(state.state_from):
 777                     state.action(m)
 778
 779         return m
 780
 781
 782 if __name__ == "__main__":
 783     alu = FPADD(width=32)
 784     main(alu, ports=alu.in_a.ports() + alu.in_b.ports() + alu.out_z.ports())
 785
 786
 787     # works... but don't use, just do "python fname.py convert -t v"
 788     #print (verilog.convert(alu, ports=[
 789     #                        ports=alu.in_a.ports() + \
 790     #                              alu.in_b.ports() + \
 791     #                              alu.out_z.ports())