src/ieee754/part_mul_add/multiply.py

   1 # SPDX-License-Identifier: LGPL-2.1-or-later
   2 # See Notices.txt for copyright information
   3 """Integer Multiplication."""
   4
   5 from nmigen import Signal, Module, Value, Elaboratable, Cat, C, Mux, Repl
   6 from nmigen.hdl.ast import Assign
   7 from abc import ABCMeta, abstractmethod
   8 from nmigen.cli import main
   9 from functools import reduce
  10 from operator import or_
  11
  12
  13 class PartitionPoints(dict):
  14     """Partition points and corresponding ``Value``s.
  15
  16     The points at where an ALU is partitioned along with ``Value``s that
  17     specify if the corresponding partition points are enabled.
  18
  19     For example: ``{1: True, 5: True, 10: True}`` with
  20     ``width == 16`` specifies that the ALU is split into 4 sections:
  21     * bits 0 <= ``i`` < 1
  22     * bits 1 <= ``i`` < 5
  23     * bits 5 <= ``i`` < 10
  24     * bits 10 <= ``i`` < 16
  25
  26     If the partition_points were instead ``{1: True, 5: a, 10: True}``
  27     where ``a`` is a 1-bit ``Signal``:
  28     * If ``a`` is asserted:
  29         * bits 0 <= ``i`` < 1
  30         * bits 1 <= ``i`` < 5
  31         * bits 5 <= ``i`` < 10
  32         * bits 10 <= ``i`` < 16
  33     * Otherwise
  34         * bits 0 <= ``i`` < 1
  35         * bits 1 <= ``i`` < 10
  36         * bits 10 <= ``i`` < 16
  37     """
  38
  39     def __init__(self, partition_points=None):
  40         """Create a new ``PartitionPoints``.
  41
  42         :param partition_points: the input partition points to values mapping.
  43         """
  44         super().__init__()
  45         if partition_points is not None:
  46             for point, enabled in partition_points.items():
  47                 if not isinstance(point, int):
  48                     raise TypeError("point must be a non-negative integer")
  49                 if point < 0:
  50                     raise ValueError("point must be a non-negative integer")
  51                 self[point] = Value.wrap(enabled)
  52
  53     def like(self, name=None, src_loc_at=0, mul=1):
  54         """Create a new ``PartitionPoints`` with ``Signal``s for all values.
  55
  56         :param name: the base name for the new ``Signal``s.
  57         :param mul: a multiplication factor on the indices
  58         """
  59         if name is None:
  60             name = Signal(src_loc_at=1+src_loc_at).name  # get variable name
  61         retval = PartitionPoints()
  62         for point, enabled in self.items():
  63             point *= mul
  64             retval[point] = Signal(enabled.shape(), name=f"{name}_{point}")
  65         return retval
  66
  67     def eq(self, rhs):
  68         """Assign ``PartitionPoints`` using ``Signal.eq``."""
  69         if set(self.keys()) != set(rhs.keys()):
  70             raise ValueError("incompatible point set")
  71         for point, enabled in self.items():
  72             yield enabled.eq(rhs[point])
  73
  74     def as_mask(self, width, mul=1):
  75         """Create a bit-mask from `self`.
  76
  77         Each bit in the returned mask is clear only if the partition point at
  78         the same bit-index is enabled.
  79
  80         :param width: the bit width of the resulting mask
  81         :param mul: a "multiplier" which in-place expands the partition points
  82                     typically set to "2" when used for multipliers
  83         """
  84         bits = []
  85         for i in range(width):
  86             i /= mul
  87             if i.is_integer() and int(i) in self:
  88                 bits.append(~self[i])
  89             else:
  90                 bits.append(True)
  91         return Cat(*bits)
  92
  93     def get_max_partition_count(self, width):
  94         """Get the maximum number of partitions.
  95
  96         Gets the number of partitions when all partition points are enabled.
  97         """
  98         retval = 1
  99         for point in self.keys():
 100             if point < width:
 101                 retval += 1
 102         return retval
 103
 104     def fits_in_width(self, width):
 105         """Check if all partition points are smaller than `width`."""
 106         for point in self.keys():
 107             if point >= width:
 108                 return False
 109         return True
 110
 111     def part_byte(self, index, mfactor=1): # mfactor used for "expanding"
 112         if index == -1 or index == 7:
 113             return C(True, 1)
 114         assert index >= 0 and index < 8
 115         return self[(index * 8 + 8)*mfactor]
 116
 117
 118 class FullAdder(Elaboratable):
 119     """Full Adder.
 120
 121     :attribute in0: the first input
 122     :attribute in1: the second input
 123     :attribute in2: the third input
 124     :attribute sum: the sum output
 125     :attribute carry: the carry output
 126
 127     Rather than do individual full adders (and have an array of them,
 128     which would be very slow to simulate), this module can specify the
 129     bit width of the inputs and outputs: in effect it performs multiple
 130     Full 3-2 Add operations "in parallel".
 131     """
 132
 133     def __init__(self, width):
 134         """Create a ``FullAdder``.
 135
 136         :param width: the bit width of the input and output
 137         """
 138         self.in0 = Signal(width, reset_less=True)
 139         self.in1 = Signal(width, reset_less=True)
 140         self.in2 = Signal(width, reset_less=True)
 141         self.sum = Signal(width, reset_less=True)
 142         self.carry = Signal(width, reset_less=True)
 143
 144     def elaborate(self, platform):
 145         """Elaborate this module."""
 146         m = Module()
 147         m.d.comb += self.sum.eq(self.in0 ^ self.in1 ^ self.in2)
 148         m.d.comb += self.carry.eq((self.in0 & self.in1)
 149                                   | (self.in1 & self.in2)
 150                                   | (self.in2 & self.in0))
 151         return m
 152
 153
 154 class MaskedFullAdder(Elaboratable):
 155     """Masked Full Adder.
 156
 157     :attribute mask: the carry partition mask
 158     :attribute in0: the first input
 159     :attribute in1: the second input
 160     :attribute in2: the third input
 161     :attribute sum: the sum output
 162     :attribute mcarry: the masked carry output
 163
 164     FullAdders are always used with a "mask" on the output.  To keep
 165     the graphviz "clean", this class performs the masking here rather
 166     than inside a large for-loop.
 167
 168     See the following discussion as to why this is no longer derived
 169     from FullAdder.  Each carry is shifted here *before* being ANDed
 170     with the mask, so that an AOI cell may be used (which is more
 171     gate-efficient)
 172     https://en.wikipedia.org/wiki/AND-OR-Invert
 173     https://groups.google.com/d/msg/comp.arch/fcq-GLQqvas/vTxmcA0QAgAJ
 174     """
 175
 176     def __init__(self, width):
 177         """Create a ``MaskedFullAdder``.
 178
 179         :param width: the bit width of the input and output
 180         """
 181         self.width = width
 182         self.mask = Signal(width, reset_less=True)
 183         self.mcarry = Signal(width, reset_less=True)
 184         self.in0 = Signal(width, reset_less=True)
 185         self.in1 = Signal(width, reset_less=True)
 186         self.in2 = Signal(width, reset_less=True)
 187         self.sum = Signal(width, reset_less=True)
 188
 189     def elaborate(self, platform):
 190         """Elaborate this module."""
 191         m = Module()
 192         s1 = Signal(self.width, reset_less=True)
 193         s2 = Signal(self.width, reset_less=True)
 194         s3 = Signal(self.width, reset_less=True)
 195         c1 = Signal(self.width, reset_less=True)
 196         c2 = Signal(self.width, reset_less=True)
 197         c3 = Signal(self.width, reset_less=True)
 198         m.d.comb += self.sum.eq(self.in0 ^ self.in1 ^ self.in2)
 199         m.d.comb += s1.eq(Cat(0, self.in0))
 200         m.d.comb += s2.eq(Cat(0, self.in1))
 201         m.d.comb += s3.eq(Cat(0, self.in2))
 202         m.d.comb += c1.eq(s1 & s2 & self.mask)
 203         m.d.comb += c2.eq(s2 & s3 & self.mask)
 204         m.d.comb += c3.eq(s3 & s1 & self.mask)
 205         m.d.comb += self.mcarry.eq(c1 | c2 | c3)
 206         return m
 207
 208
 209 class PartitionedAdder(Elaboratable):
 210     """Partitioned Adder.
 211
 212     Performs the final add.  The partition points are included in the
 213     actual add (in one of the operands only), which causes a carry over
 214     to the next bit.  Then the final output *removes* the extra bits from
 215     the result.
 216
 217     partition: .... P... P... P... P... (32 bits)
 218     a        : .... .... .... .... .... (32 bits)
 219     b        : .... .... .... .... .... (32 bits)
 220     exp-a    : ....P....P....P....P.... (32+4 bits, P=1 if no partition)
 221     exp-b    : ....0....0....0....0.... (32 bits plus 4 zeros)
 222     exp-o    : ....xN...xN...xN...xN... (32+4 bits - x to be discarded)
 223     o        : .... N... N... N... N... (32 bits - x ignored, N is carry-over)
 224
 225     :attribute width: the bit width of the input and output. Read-only.
 226     :attribute a: the first input to the adder
 227     :attribute b: the second input to the adder
 228     :attribute output: the sum output
 229     :attribute partition_points: the input partition points. Modification not
 230         supported, except for by ``Signal.eq``.
 231     """
 232
 233     def __init__(self, width, partition_points, partition_step=1):
 234         """Create a ``PartitionedAdder``.
 235
 236         :param width: the bit width of the input and output
 237         :param partition_points: the input partition points
 238         :param partition_step: a multiplier (typically double) step
 239                                which in-place "expands" the partition points
 240         """
 241         self.width = width
 242         self.pmul = partition_step
 243         self.a = Signal(width, reset_less=True)
 244         self.b = Signal(width, reset_less=True)
 245         self.output = Signal(width, reset_less=True)
 246         self.partition_points = PartitionPoints(partition_points)
 247         if not self.partition_points.fits_in_width(width):
 248             raise ValueError("partition_points doesn't fit in width")
 249         expanded_width = 0
 250         for i in range(self.width):
 251             if i in self.partition_points:
 252                 expanded_width += 1
 253             expanded_width += 1
 254         self._expanded_width = expanded_width
 255
 256     def elaborate(self, platform):
 257         """Elaborate this module."""
 258         m = Module()
 259         expanded_a = Signal(self._expanded_width, reset_less=True)
 260         expanded_b = Signal(self._expanded_width, reset_less=True)
 261         expanded_o = Signal(self._expanded_width, reset_less=True)
 262
 263         expanded_index = 0
 264         # store bits in a list, use Cat later.  graphviz is much cleaner
 265         al, bl, ol, ea, eb, eo = [],[],[],[],[],[]
 266
 267         # partition points are "breaks" (extra zeros or 1s) in what would
 268         # otherwise be a massive long add.  when the "break" points are 0,
 269         # whatever is in it (in the output) is discarded.  however when
 270         # there is a "1", it causes a roll-over carry to the *next* bit.
 271         # we still ignore the "break" bit in the [intermediate] output,
 272         # however by that time we've got the effect that we wanted: the
 273         # carry has been carried *over* the break point.
 274
 275         for i in range(self.width):
 276             pi = i/self.pmul # double the range of the partition point test
 277             if pi.is_integer() and pi in self.partition_points:
 278                 # add extra bit set to 0 + 0 for enabled partition points
 279                 # and 1 + 0 for disabled partition points
 280                 ea.append(expanded_a[expanded_index])
 281                 al.append(~self.partition_points[pi]) # add extra bit in a
 282                 eb.append(expanded_b[expanded_index])
 283                 bl.append(C(0)) # yes, add a zero
 284                 expanded_index += 1 # skip the extra point.  NOT in the output
 285             ea.append(expanded_a[expanded_index])
 286             eb.append(expanded_b[expanded_index])
 287             eo.append(expanded_o[expanded_index])
 288             al.append(self.a[i])
 289             bl.append(self.b[i])
 290             ol.append(self.output[i])
 291             expanded_index += 1
 292
 293         # combine above using Cat
 294         m.d.comb += Cat(*ea).eq(Cat(*al))
 295         m.d.comb += Cat(*eb).eq(Cat(*bl))
 296         m.d.comb += Cat(*ol).eq(Cat(*eo))
 297
 298         # use only one addition to take advantage of look-ahead carry and
 299         # special hardware on FPGAs
 300         m.d.comb += expanded_o.eq(expanded_a + expanded_b)
 301         return m
 302
 303
 304 FULL_ADDER_INPUT_COUNT = 3
 305
 306 class AddReduceData:
 307
 308     def __init__(self, part_pts, n_inputs, output_width, n_parts):
 309         self.part_ops = [Signal(2, name=f"part_ops_{i}", reset_less=True)
 310                           for i in range(n_parts)]
 311         self.terms = [Signal(output_width, name=f"inputs_{i}",
 312                               reset_less=True)
 313                         for i in range(n_inputs)]
 314         self.part_pts = part_pts.like()
 315
 316     def eq_from(self, part_pts, inputs, part_ops):
 317         return [self.part_pts.eq(part_pts)] + \
 318                [self.terms[i].eq(inputs[i])
 319                                      for i in range(len(self.terms))] + \
 320                [self.part_ops[i].eq(part_ops[i])
 321                                      for i in range(len(self.part_ops))]
 322
 323     def eq(self, rhs):
 324         return self.eq_from(rhs.part_pts, rhs.terms, rhs.part_ops)
 325
 326
 327 class FinalReduceData:
 328
 329     def __init__(self, part_pts, output_width, n_parts):
 330         self.part_ops = [Signal(2, name=f"part_ops_{i}", reset_less=True)
 331                           for i in range(n_parts)]
 332         self.output = Signal(output_width, reset_less=True)
 333         self.part_pts = part_pts.like()
 334
 335     def eq_from(self, part_pts, output, part_ops):
 336         return [self.part_pts.eq(part_pts)] + \
 337                [self.output.eq(output)] + \
 338                [self.part_ops[i].eq(part_ops[i])
 339                                      for i in range(len(self.part_ops))]
 340
 341     def eq(self, rhs):
 342         return self.eq_from(rhs.part_pts, rhs.output, rhs.part_ops)
 343
 344
 345 class FinalAdd(Elaboratable):
 346     """ Final stage of add reduce
 347     """
 348
 349     def __init__(self, lidx, n_inputs, output_width, n_parts, partition_points,
 350                        partition_step=1):
 351         self.lidx = lidx
 352         self.partition_step = partition_step
 353         self.output_width = output_width
 354         self.n_inputs = n_inputs
 355         self.n_parts = n_parts
 356         self.partition_points = PartitionPoints(partition_points)
 357         if not self.partition_points.fits_in_width(output_width):
 358             raise ValueError("partition_points doesn't fit in output_width")
 359
 360         self.i = self.ispec()
 361         self.o = self.ospec()
 362
 363     def ispec(self):
 364         return AddReduceData(self.partition_points, self.n_inputs,
 365                              self.output_width, self.n_parts)
 366
 367     def ospec(self):
 368         return FinalReduceData(self.partition_points,
 369                                  self.output_width, self.n_parts)
 370
 371     def setup(self, m, i):
 372         m.submodules.finaladd = self
 373         m.d.comb += self.i.eq(i)
 374
 375     def process(self, i):
 376         return self.o
 377
 378     def elaborate(self, platform):
 379         """Elaborate this module."""
 380         m = Module()
 381
 382         output_width = self.output_width
 383         output = Signal(output_width, reset_less=True)
 384         if self.n_inputs == 0:
 385             # use 0 as the default output value
 386             m.d.comb += output.eq(0)
 387         elif self.n_inputs == 1:
 388             # handle single input
 389             m.d.comb += output.eq(self.i.terms[0])
 390         else:
 391             # base case for adding 2 inputs
 392             assert self.n_inputs == 2
 393             adder = PartitionedAdder(output_width,
 394                                      self.i.part_pts, self.partition_step)
 395             m.submodules.final_adder = adder
 396             m.d.comb += adder.a.eq(self.i.terms[0])
 397             m.d.comb += adder.b.eq(self.i.terms[1])
 398             m.d.comb += output.eq(adder.output)
 399
 400         # create output
 401         m.d.comb += self.o.eq_from(self.i.part_pts, output,
 402                                    self.i.part_ops)
 403
 404         return m
 405
 406
 407 class AddReduceSingle(Elaboratable):
 408     """Add list of numbers together.
 409
 410     :attribute inputs: input ``Signal``s to be summed. Modification not
 411         supported, except for by ``Signal.eq``.
 412     :attribute register_levels: List of nesting levels that should have
 413         pipeline registers.
 414     :attribute output: output sum.
 415     :attribute partition_points: the input partition points. Modification not
 416         supported, except for by ``Signal.eq``.
 417     """
 418
 419     def __init__(self, lidx, n_inputs, output_width, n_parts, partition_points,
 420                        partition_step=1):
 421         """Create an ``AddReduce``.
 422
 423         :param inputs: input ``Signal``s to be summed.
 424         :param output_width: bit-width of ``output``.
 425         :param partition_points: the input partition points.
 426         """
 427         self.lidx = lidx
 428         self.partition_step = partition_step
 429         self.n_inputs = n_inputs
 430         self.n_parts = n_parts
 431         self.output_width = output_width
 432         self.partition_points = PartitionPoints(partition_points)
 433         if not self.partition_points.fits_in_width(output_width):
 434             raise ValueError("partition_points doesn't fit in output_width")
 435
 436         self.groups = AddReduceSingle.full_adder_groups(n_inputs)
 437         self.n_terms = AddReduceSingle.calc_n_inputs(n_inputs, self.groups)
 438
 439         self.i = self.ispec()
 440         self.o = self.ospec()
 441
 442     def ispec(self):
 443         return AddReduceData(self.partition_points, self.n_inputs,
 444                              self.output_width, self.n_parts)
 445
 446     def ospec(self):
 447         return AddReduceData(self.partition_points, self.n_terms,
 448                              self.output_width, self.n_parts)
 449
 450     def setup(self, m, i):
 451         setattr(m.submodules, "addreduce_%d" % self.lidx, self)
 452         m.d.comb += self.i.eq(i)
 453
 454     def process(self, i):
 455         return self.o
 456
 457     @staticmethod
 458     def calc_n_inputs(n_inputs, groups):
 459         retval = len(groups)*2
 460         if n_inputs % FULL_ADDER_INPUT_COUNT == 1:
 461             retval += 1
 462         elif n_inputs % FULL_ADDER_INPUT_COUNT == 2:
 463             retval += 2
 464         else:
 465             assert n_inputs % FULL_ADDER_INPUT_COUNT == 0
 466         return retval
 467
 468     @staticmethod
 469     def get_max_level(input_count):
 470         """Get the maximum level.
 471
 472         All ``register_levels`` must be less than or equal to the maximum
 473         level.
 474         """
 475         retval = 0
 476         while True:
 477             groups = AddReduceSingle.full_adder_groups(input_count)
 478             if len(groups) == 0:
 479                 return retval
 480             input_count %= FULL_ADDER_INPUT_COUNT
 481             input_count += 2 * len(groups)
 482             retval += 1
 483
 484     @staticmethod
 485     def full_adder_groups(input_count):
 486         """Get ``inputs`` indices for which a full adder should be built."""
 487         return range(0,
 488                      input_count - FULL_ADDER_INPUT_COUNT + 1,
 489                      FULL_ADDER_INPUT_COUNT)
 490
 491     def create_next_terms(self):
 492         """ create next intermediate terms, for linking up in elaborate, below
 493         """
 494         terms = []
 495         adders = []
 496
 497         # create full adders for this recursive level.
 498         # this shrinks N terms to 2 * (N // 3) plus the remainder
 499         for i in self.groups:
 500             adder_i = MaskedFullAdder(self.output_width)
 501             adders.append((i, adder_i))
 502             # add both the sum and the masked-carry to the next level.
 503             # 3 inputs have now been reduced to 2...
 504             terms.append(adder_i.sum)
 505             terms.append(adder_i.mcarry)
 506         # handle the remaining inputs.
 507         if self.n_inputs % FULL_ADDER_INPUT_COUNT == 1:
 508             terms.append(self.i.terms[-1])
 509         elif self.n_inputs % FULL_ADDER_INPUT_COUNT == 2:
 510             # Just pass the terms to the next layer, since we wouldn't gain
 511             # anything by using a half adder since there would still be 2 terms
 512             # and just passing the terms to the next layer saves gates.
 513             terms.append(self.i.terms[-2])
 514             terms.append(self.i.terms[-1])
 515         else:
 516             assert self.n_inputs % FULL_ADDER_INPUT_COUNT == 0
 517
 518         return terms, adders
 519
 520     def elaborate(self, platform):
 521         """Elaborate this module."""
 522         m = Module()
 523
 524         terms, adders = self.create_next_terms()
 525
 526         # copy the intermediate terms to the output
 527         for i, value in enumerate(terms):
 528             m.d.comb += self.o.terms[i].eq(value)
 529
 530         # copy reg part points and part ops to output
 531         m.d.comb += self.o.part_pts.eq(self.i.part_pts)
 532         m.d.comb += [self.o.part_ops[i].eq(self.i.part_ops[i])
 533                                      for i in range(len(self.i.part_ops))]
 534
 535         # set up the partition mask (for the adders)
 536         part_mask = Signal(self.output_width, reset_less=True)
 537
 538         # get partition points as a mask
 539         mask = self.i.part_pts.as_mask(self.output_width,
 540                                        mul=self.partition_step)
 541         m.d.comb += part_mask.eq(mask)
 542
 543         # add and link the intermediate term modules
 544         for i, (iidx, adder_i) in enumerate(adders):
 545             setattr(m.submodules, f"adder_{i}", adder_i)
 546
 547             m.d.comb += adder_i.in0.eq(self.i.terms[iidx])
 548             m.d.comb += adder_i.in1.eq(self.i.terms[iidx + 1])
 549             m.d.comb += adder_i.in2.eq(self.i.terms[iidx + 2])
 550             m.d.comb += adder_i.mask.eq(part_mask)
 551
 552         return m
 553
 554
 555 class AddReduceInternal:
 556     """Recursively Add list of numbers together.
 557
 558     :attribute inputs: input ``Signal``s to be summed. Modification not
 559         supported, except for by ``Signal.eq``.
 560     :attribute register_levels: List of nesting levels that should have
 561         pipeline registers.
 562     :attribute output: output sum.
 563     :attribute partition_points: the input partition points. Modification not
 564         supported, except for by ``Signal.eq``.
 565     """
 566
 567     def __init__(self, i, output_width, partition_step=1):
 568         """Create an ``AddReduce``.
 569
 570         :param inputs: input ``Signal``s to be summed.
 571         :param output_width: bit-width of ``output``.
 572         :param partition_points: the input partition points.
 573         """
 574         self.i = i
 575         self.inputs = i.terms
 576         self.part_ops = i.part_ops
 577         self.output_width = output_width
 578         self.partition_points = i.part_pts
 579         self.partition_step = partition_step
 580
 581         self.create_levels()
 582
 583     def create_levels(self):
 584         """creates reduction levels"""
 585
 586         mods = []
 587         partition_points = self.partition_points
 588         part_ops = self.part_ops
 589         n_parts = len(part_ops)
 590         inputs = self.inputs
 591         ilen = len(inputs)
 592         while True:
 593             groups = AddReduceSingle.full_adder_groups(len(inputs))
 594             if len(groups) == 0:
 595                 break
 596             lidx = len(mods)
 597             next_level = AddReduceSingle(lidx, ilen, self.output_width, n_parts,
 598                                          partition_points,
 599                                          self.partition_step)
 600             mods.append(next_level)
 601             partition_points = next_level.i.part_pts
 602             inputs = next_level.o.terms
 603             ilen = len(inputs)
 604             part_ops = next_level.i.part_ops
 605
 606         lidx = len(mods)
 607         next_level = FinalAdd(lidx, ilen, self.output_width, n_parts,
 608                               partition_points, self.partition_step)
 609         mods.append(next_level)
 610
 611         self.levels = mods
 612
 613
 614 class AddReduce(AddReduceInternal, Elaboratable):
 615     """Recursively Add list of numbers together.
 616
 617     :attribute inputs: input ``Signal``s to be summed. Modification not
 618         supported, except for by ``Signal.eq``.
 619     :attribute register_levels: List of nesting levels that should have
 620         pipeline registers.
 621     :attribute output: output sum.
 622     :attribute partition_points: the input partition points. Modification not
 623         supported, except for by ``Signal.eq``.
 624     """
 625
 626     def __init__(self, inputs, output_width, register_levels, part_pts,
 627                        part_ops, partition_step=1):
 628         """Create an ``AddReduce``.
 629
 630         :param inputs: input ``Signal``s to be summed.
 631         :param output_width: bit-width of ``output``.
 632         :param register_levels: List of nesting levels that should have
 633             pipeline registers.
 634         :param partition_points: the input partition points.
 635         """
 636         self._inputs = inputs
 637         self._part_pts = part_pts
 638         self._part_ops = part_ops
 639         n_parts = len(part_ops)
 640         self.i = AddReduceData(part_pts, len(inputs),
 641                              output_width, n_parts)
 642         AddReduceInternal.__init__(self, self.i, output_width, partition_step)
 643         self.o = FinalReduceData(part_pts, output_width, n_parts)
 644         self.register_levels = register_levels
 645
 646     @staticmethod
 647     def get_max_level(input_count):
 648         return AddReduceSingle.get_max_level(input_count)
 649
 650     @staticmethod
 651     def next_register_levels(register_levels):
 652         """``Iterable`` of ``register_levels`` for next recursive level."""
 653         for level in register_levels:
 654             if level > 0:
 655                 yield level - 1
 656
 657     def elaborate(self, platform):
 658         """Elaborate this module."""
 659         m = Module()
 660
 661         m.d.comb += self.i.eq_from(self._part_pts, self._inputs, self._part_ops)
 662
 663         for i, next_level in enumerate(self.levels):
 664             setattr(m.submodules, "next_level%d" % i, next_level)
 665
 666         i = self.i
 667         for idx in range(len(self.levels)):
 668             mcur = self.levels[idx]
 669             if idx in self.register_levels:
 670                 m.d.sync += mcur.i.eq(i)
 671             else:
 672                 m.d.comb += mcur.i.eq(i)
 673             i = mcur.o # for next loop
 674
 675         # output comes from last module
 676         m.d.comb += self.o.eq(i)
 677
 678         return m
 679
 680
 681 OP_MUL_LOW = 0
 682 OP_MUL_SIGNED_HIGH = 1
 683 OP_MUL_SIGNED_UNSIGNED_HIGH = 2  # a is signed, b is unsigned
 684 OP_MUL_UNSIGNED_HIGH = 3
 685
 686
 687 def get_term(value, shift=0, enabled=None):
 688     if enabled is not None:
 689         value = Mux(enabled, value, 0)
 690     if shift > 0:
 691         value = Cat(Repl(C(0, 1), shift), value)
 692     else:
 693         assert shift == 0
 694     return value
 695
 696
 697 class ProductTerm(Elaboratable):
 698     """ this class creates a single product term (a[..]*b[..]).
 699         it has a design flaw in that is the *output* that is selected,
 700         where the multiplication(s) are combinatorially generated
 701         all the time.
 702     """
 703
 704     def __init__(self, width, twidth, pbwid, a_index, b_index):
 705         self.a_index = a_index
 706         self.b_index = b_index
 707         shift = 8 * (self.a_index + self.b_index)
 708         self.pwidth = width
 709         self.twidth = twidth
 710         self.width = width*2
 711         self.shift = shift
 712
 713         self.ti = Signal(self.width, reset_less=True)
 714         self.term = Signal(twidth, reset_less=True)
 715         self.a = Signal(twidth//2, reset_less=True)
 716         self.b = Signal(twidth//2, reset_less=True)
 717         self.pb_en = Signal(pbwid, reset_less=True)
 718
 719         self.tl = tl = []
 720         min_index = min(self.a_index, self.b_index)
 721         max_index = max(self.a_index, self.b_index)
 722         for i in range(min_index, max_index):
 723             tl.append(self.pb_en[i])
 724         name = "te_%d_%d" % (self.a_index, self.b_index)
 725         if len(tl) > 0:
 726             term_enabled = Signal(name=name, reset_less=True)
 727         else:
 728             term_enabled = None
 729         self.enabled = term_enabled
 730         self.term.name = "term_%d_%d" % (a_index, b_index) # rename
 731
 732     def elaborate(self, platform):
 733
 734         m = Module()
 735         if self.enabled is not None:
 736             m.d.comb += self.enabled.eq(~(Cat(*self.tl).bool()))
 737
 738         bsa = Signal(self.width, reset_less=True)
 739         bsb = Signal(self.width, reset_less=True)
 740         a_index, b_index = self.a_index, self.b_index
 741         pwidth = self.pwidth
 742         m.d.comb += bsa.eq(self.a.bit_select(a_index * pwidth, pwidth))
 743         m.d.comb += bsb.eq(self.b.bit_select(b_index * pwidth, pwidth))
 744         m.d.comb += self.ti.eq(bsa * bsb)
 745         m.d.comb += self.term.eq(get_term(self.ti, self.shift, self.enabled))
 746         """
 747         #TODO: sort out width issues, get inputs a/b switched on/off.
 748         #data going into Muxes is 1/2 the required width
 749
 750         pwidth = self.pwidth
 751         width = self.width
 752         bsa = Signal(self.twidth//2, reset_less=True)
 753         bsb = Signal(self.twidth//2, reset_less=True)
 754         asel = Signal(width, reset_less=True)
 755         bsel = Signal(width, reset_less=True)
 756         a_index, b_index = self.a_index, self.b_index
 757         m.d.comb += asel.eq(self.a.bit_select(a_index * pwidth, pwidth))
 758         m.d.comb += bsel.eq(self.b.bit_select(b_index * pwidth, pwidth))
 759         m.d.comb += bsa.eq(get_term(asel, self.shift, self.enabled))
 760         m.d.comb += bsb.eq(get_term(bsel, self.shift, self.enabled))
 761         m.d.comb += self.ti.eq(bsa * bsb)
 762         m.d.comb += self.term.eq(self.ti)
 763         """
 764
 765         return m
 766
 767
 768 class ProductTerms(Elaboratable):
 769     """ creates a bank of product terms.  also performs the actual bit-selection
 770         this class is to be wrapped with a for-loop on the "a" operand.
 771         it creates a second-level for-loop on the "b" operand.
 772     """
 773     def __init__(self, width, twidth, pbwid, a_index, blen):
 774         self.a_index = a_index
 775         self.blen = blen
 776         self.pwidth = width
 777         self.twidth = twidth
 778         self.pbwid = pbwid
 779         self.a = Signal(twidth//2, reset_less=True)
 780         self.b = Signal(twidth//2, reset_less=True)
 781         self.pb_en = Signal(pbwid, reset_less=True)
 782         self.terms = [Signal(twidth, name="term%d"%i, reset_less=True) \
 783                             for i in range(blen)]
 784
 785     def elaborate(self, platform):
 786
 787         m = Module()
 788
 789         for b_index in range(self.blen):
 790             t = ProductTerm(self.pwidth, self.twidth, self.pbwid,
 791                             self.a_index, b_index)
 792             setattr(m.submodules, "term_%d" % b_index, t)
 793
 794             m.d.comb += t.a.eq(self.a)
 795             m.d.comb += t.b.eq(self.b)
 796             m.d.comb += t.pb_en.eq(self.pb_en)
 797
 798             m.d.comb += self.terms[b_index].eq(t.term)
 799
 800         return m
 801
 802
 803 class LSBNegTerm(Elaboratable):
 804
 805     def __init__(self, bit_width):
 806         self.bit_width = bit_width
 807         self.part = Signal(reset_less=True)
 808         self.signed = Signal(reset_less=True)
 809         self.op = Signal(bit_width, reset_less=True)
 810         self.msb = Signal(reset_less=True)
 811         self.nt = Signal(bit_width*2, reset_less=True)
 812         self.nl = Signal(bit_width*2, reset_less=True)
 813
 814     def elaborate(self, platform):
 815         m = Module()
 816         comb = m.d.comb
 817         bit_wid = self.bit_width
 818         ext = Repl(0, bit_wid) # extend output to HI part
 819
 820         # determine sign of each incoming number *in this partition*
 821         enabled = Signal(reset_less=True)
 822         m.d.comb += enabled.eq(self.part & self.msb & self.signed)
 823
 824         # for 8-bit values: form a * 0xFF00 by using -a * 0x100, the
 825         # negation operation is split into a bitwise not and a +1.
 826         # likewise for 16, 32, and 64-bit values.
 827
 828         # width-extended 1s complement if a is signed, otherwise zero
 829         comb += self.nt.eq(Mux(enabled, Cat(ext, ~self.op), 0))
 830
 831         # add 1 if signed, otherwise add zero
 832         comb += self.nl.eq(Cat(ext, enabled, Repl(0, bit_wid-1)))
 833
 834         return m
 835
 836
 837 class Parts(Elaboratable):
 838
 839     def __init__(self, pbwid, part_pts, n_parts):
 840         self.pbwid = pbwid
 841         # inputs
 842         self.part_pts = PartitionPoints.like(part_pts)
 843         # outputs
 844         self.parts = [Signal(name=f"part_{i}", reset_less=True)
 845                       for i in range(n_parts)]
 846
 847     def elaborate(self, platform):
 848         m = Module()
 849
 850         part_pts, parts = self.part_pts, self.parts
 851         # collect part-bytes (double factor because the input is extended)
 852         pbs = Signal(self.pbwid, reset_less=True)
 853         tl = []
 854         for i in range(self.pbwid):
 855             pb = Signal(name="pb%d" % i, reset_less=True)
 856             m.d.comb += pb.eq(part_pts.part_byte(i))
 857             tl.append(pb)
 858         m.d.comb += pbs.eq(Cat(*tl))
 859
 860         # negated-temporary copy of partition bits
 861         npbs = Signal.like(pbs, reset_less=True)
 862         m.d.comb += npbs.eq(~pbs)
 863         byte_count = 8 // len(parts)
 864         for i in range(len(parts)):
 865             pbl = []
 866             pbl.append(npbs[i * byte_count - 1])
 867             for j in range(i * byte_count, (i + 1) * byte_count - 1):
 868                 pbl.append(pbs[j])
 869             pbl.append(npbs[(i + 1) * byte_count - 1])
 870             value = Signal(len(pbl), name="value_%d" % i, reset_less=True)
 871             m.d.comb += value.eq(Cat(*pbl))
 872             m.d.comb += parts[i].eq(~(value).bool())
 873
 874         return m
 875
 876
 877 class Part(Elaboratable):
 878     """ a key class which, depending on the partitioning, will determine
 879         what action to take when parts of the output are signed or unsigned.
 880
 881         this requires 2 pieces of data *per operand, per partition*:
 882         whether the MSB is HI/LO (per partition!), and whether a signed
 883         or unsigned operation has been *requested*.
 884
 885         once that is determined, signed is basically carried out
 886         by splitting 2's complement into 1's complement plus one.
 887         1's complement is just a bit-inversion.
 888
 889         the extra terms - as separate terms - are then thrown at the
 890         AddReduce alongside the multiplication part-results.
 891     """
 892     def __init__(self, part_pts, width, n_parts, pbwid):
 893
 894         self.pbwid = pbwid
 895         self.part_pts = part_pts
 896
 897         # inputs
 898         self.a = Signal(64, reset_less=True)
 899         self.b = Signal(64, reset_less=True)
 900         self.a_signed = [Signal(name=f"a_signed_{i}", reset_less=True)
 901                             for i in range(8)]
 902         self.b_signed = [Signal(name=f"_b_signed_{i}", reset_less=True)
 903                             for i in range(8)]
 904         self.pbs = Signal(pbwid, reset_less=True)
 905
 906         # outputs
 907         self.parts = [Signal(name=f"part_{i}", reset_less=True)
 908                             for i in range(n_parts)]
 909
 910         self.not_a_term = Signal(width, reset_less=True)
 911         self.neg_lsb_a_term = Signal(width, reset_less=True)
 912         self.not_b_term = Signal(width, reset_less=True)
 913         self.neg_lsb_b_term = Signal(width, reset_less=True)
 914
 915     def elaborate(self, platform):
 916         m = Module()
 917
 918         pbs, parts = self.pbs, self.parts
 919         part_pts = self.part_pts
 920         m.submodules.p = p = Parts(self.pbwid, part_pts, len(parts))
 921         m.d.comb += p.part_pts.eq(part_pts)
 922         parts = p.parts
 923
 924         byte_count = 8 // len(parts)
 925
 926         not_a_term, neg_lsb_a_term, not_b_term, neg_lsb_b_term = (
 927                 self.not_a_term, self.neg_lsb_a_term,
 928                 self.not_b_term, self.neg_lsb_b_term)
 929
 930         byte_width = 8 // len(parts) # byte width
 931         bit_wid = 8 * byte_width     # bit width
 932         nat, nbt, nla, nlb = [], [], [], []
 933         for i in range(len(parts)):
 934             # work out bit-inverted and +1 term for a.
 935             pa = LSBNegTerm(bit_wid)
 936             setattr(m.submodules, "lnt_%d_a_%d" % (bit_wid, i), pa)
 937             m.d.comb += pa.part.eq(parts[i])
 938             m.d.comb += pa.op.eq(self.a.bit_select(bit_wid * i, bit_wid))
 939             m.d.comb += pa.signed.eq(self.b_signed[i * byte_width]) # yes b
 940             m.d.comb += pa.msb.eq(self.b[(i + 1) * bit_wid - 1]) # really, b
 941             nat.append(pa.nt)
 942             nla.append(pa.nl)
 943
 944             # work out bit-inverted and +1 term for b
 945             pb = LSBNegTerm(bit_wid)
 946             setattr(m.submodules, "lnt_%d_b_%d" % (bit_wid, i), pb)
 947             m.d.comb += pb.part.eq(parts[i])
 948             m.d.comb += pb.op.eq(self.b.bit_select(bit_wid * i, bit_wid))
 949             m.d.comb += pb.signed.eq(self.a_signed[i * byte_width]) # yes a
 950             m.d.comb += pb.msb.eq(self.a[(i + 1) * bit_wid - 1]) # really, a
 951             nbt.append(pb.nt)
 952             nlb.append(pb.nl)
 953
 954         # concatenate together and return all 4 results.
 955         m.d.comb += [not_a_term.eq(Cat(*nat)),
 956                      not_b_term.eq(Cat(*nbt)),
 957                      neg_lsb_a_term.eq(Cat(*nla)),
 958                      neg_lsb_b_term.eq(Cat(*nlb)),
 959                     ]
 960
 961         return m
 962
 963
 964 class IntermediateOut(Elaboratable):
 965     """ selects the HI/LO part of the multiplication, for a given bit-width
 966         the output is also reconstructed in its SIMD (partition) lanes.
 967     """
 968     def __init__(self, width, out_wid, n_parts):
 969         self.width = width
 970         self.n_parts = n_parts
 971         self.part_ops = [Signal(2, name="dpop%d" % i, reset_less=True)
 972                                      for i in range(8)]
 973         self.intermed = Signal(out_wid, reset_less=True)
 974         self.output = Signal(out_wid//2, reset_less=True)
 975
 976     def elaborate(self, platform):
 977         m = Module()
 978
 979         ol = []
 980         w = self.width
 981         sel = w // 8
 982         for i in range(self.n_parts):
 983             op = Signal(w, reset_less=True, name="op%d_%d" % (w, i))
 984             m.d.comb += op.eq(
 985                 Mux(self.part_ops[sel * i] == OP_MUL_LOW,
 986                     self.intermed.bit_select(i * w*2, w),
 987                     self.intermed.bit_select(i * w*2 + w, w)))
 988             ol.append(op)
 989         m.d.comb += self.output.eq(Cat(*ol))
 990
 991         return m
 992
 993
 994 class FinalOut(Elaboratable):
 995     """ selects the final output based on the partitioning.
 996
 997         each byte is selectable independently, i.e. it is possible
 998         that some partitions requested 8-bit computation whilst others
 999         requested 16 or 32 bit.
1000     """
1001     def __init__(self, output_width, n_parts, part_pts):
1002         self.part_pts = part_pts
1003         self.output_width = output_width
1004         self.n_parts = n_parts
1005         self.out_wid = output_width//2
1006
1007         self.i = self.ispec()
1008         self.o = self.ospec()
1009
1010     def ispec(self):
1011         return IntermediateData(self.part_pts, self.output_width, self.n_parts)
1012
1013     def ospec(self):
1014         return OutputData()
1015
1016     def setup(self, m, i):
1017         m.submodules.finalout = self
1018         m.d.comb += self.i.eq(i)
1019
1020     def process(self, i):
1021         return self.o
1022
1023     def elaborate(self, platform):
1024         m = Module()
1025
1026         part_pts = self.part_pts
1027         m.submodules.p_8 = p_8 = Parts(8, part_pts, 8)
1028         m.submodules.p_16 = p_16 = Parts(8, part_pts, 4)
1029         m.submodules.p_32 = p_32 = Parts(8, part_pts, 2)
1030         m.submodules.p_64 = p_64 = Parts(8, part_pts, 1)
1031
1032         out_part_pts = self.i.part_pts
1033
1034         # temporaries
1035         d8 = [Signal(name=f"d8_{i}", reset_less=True) for i in range(8)]
1036         d16 = [Signal(name=f"d16_{i}", reset_less=True) for i in range(4)]
1037         d32 = [Signal(name=f"d32_{i}", reset_less=True) for i in range(2)]
1038
1039         i8 = Signal(self.out_wid, reset_less=True)
1040         i16 = Signal(self.out_wid, reset_less=True)
1041         i32 = Signal(self.out_wid, reset_less=True)
1042         i64 = Signal(self.out_wid, reset_less=True)
1043
1044         m.d.comb += p_8.part_pts.eq(out_part_pts)
1045         m.d.comb += p_16.part_pts.eq(out_part_pts)
1046         m.d.comb += p_32.part_pts.eq(out_part_pts)
1047         m.d.comb += p_64.part_pts.eq(out_part_pts)
1048
1049         for i in range(len(p_8.parts)):
1050             m.d.comb += d8[i].eq(p_8.parts[i])
1051         for i in range(len(p_16.parts)):
1052             m.d.comb += d16[i].eq(p_16.parts[i])
1053         for i in range(len(p_32.parts)):
1054             m.d.comb += d32[i].eq(p_32.parts[i])
1055         m.d.comb += i8.eq(self.i.outputs[0])
1056         m.d.comb += i16.eq(self.i.outputs[1])
1057         m.d.comb += i32.eq(self.i.outputs[2])
1058         m.d.comb += i64.eq(self.i.outputs[3])
1059
1060         ol = []
1061         for i in range(8):
1062             # select one of the outputs: d8 selects i8, d16 selects i16
1063             # d32 selects i32, and the default is i64.
1064             # d8 and d16 are ORed together in the first Mux
1065             # then the 2nd selects either i8 or i16.
1066             # if neither d8 nor d16 are set, d32 selects either i32 or i64.
1067             op = Signal(8, reset_less=True, name="op_%d" % i)
1068             m.d.comb += op.eq(
1069                 Mux(d8[i] | d16[i // 2],
1070                     Mux(d8[i], i8.bit_select(i * 8, 8),
1071                                i16.bit_select(i * 8, 8)),
1072                     Mux(d32[i // 4], i32.bit_select(i * 8, 8),
1073                                       i64.bit_select(i * 8, 8))))
1074             ol.append(op)
1075
1076         # create outputs
1077         m.d.comb += self.o.output.eq(Cat(*ol))
1078         m.d.comb += self.o.intermediate_output.eq(self.i.intermediate_output)
1079
1080         return m
1081
1082
1083 class OrMod(Elaboratable):
1084     """ ORs four values together in a hierarchical tree
1085     """
1086     def __init__(self, wid):
1087         self.wid = wid
1088         self.orin = [Signal(wid, name="orin%d" % i, reset_less=True)
1089                      for i in range(4)]
1090         self.orout = Signal(wid, reset_less=True)
1091
1092     def elaborate(self, platform):
1093         m = Module()
1094         or1 = Signal(self.wid, reset_less=True)
1095         or2 = Signal(self.wid, reset_less=True)
1096         m.d.comb += or1.eq(self.orin[0] | self.orin[1])
1097         m.d.comb += or2.eq(self.orin[2] | self.orin[3])
1098         m.d.comb += self.orout.eq(or1 | or2)
1099
1100         return m
1101
1102
1103 class Signs(Elaboratable):
1104     """ determines whether a or b are signed numbers
1105         based on the required operation type (OP_MUL_*)
1106     """
1107
1108     def __init__(self):
1109         self.part_ops = Signal(2, reset_less=True)
1110         self.a_signed = Signal(reset_less=True)
1111         self.b_signed = Signal(reset_less=True)
1112
1113     def elaborate(self, platform):
1114
1115         m = Module()
1116
1117         asig = self.part_ops != OP_MUL_UNSIGNED_HIGH
1118         bsig = (self.part_ops == OP_MUL_LOW) \
1119                     | (self.part_ops == OP_MUL_SIGNED_HIGH)
1120         m.d.comb += self.a_signed.eq(asig)
1121         m.d.comb += self.b_signed.eq(bsig)
1122
1123         return m
1124
1125
1126 class IntermediateData:
1127
1128     def __init__(self, part_pts, output_width, n_parts):
1129         self.part_ops = [Signal(2, name=f"part_ops_{i}", reset_less=True)
1130                           for i in range(n_parts)]
1131         self.part_pts = part_pts.like()
1132         self.outputs = [Signal(output_width, name="io%d" % i, reset_less=True)
1133                           for i in range(4)]
1134         # intermediates (needed for unit tests)
1135         self.intermediate_output = Signal(output_width)
1136
1137     def eq_from(self, part_pts, outputs, intermediate_output,
1138                       part_ops):
1139         return [self.part_pts.eq(part_pts)] + \
1140                [self.intermediate_output.eq(intermediate_output)] + \
1141                [self.outputs[i].eq(outputs[i])
1142                                      for i in range(4)] + \
1143                [self.part_ops[i].eq(part_ops[i])
1144                                      for i in range(len(self.part_ops))]
1145
1146     def eq(self, rhs):
1147         return self.eq_from(rhs.part_pts, rhs.outputs,
1148                             rhs.intermediate_output, rhs.part_ops)
1149
1150
1151 class InputData:
1152
1153     def __init__(self):
1154         self.a = Signal(64)
1155         self.b = Signal(64)
1156         self.part_pts = PartitionPoints()
1157         for i in range(8, 64, 8):
1158             self.part_pts[i] = Signal(name=f"part_pts_{i}")
1159         self.part_ops = [Signal(2, name=f"part_ops_{i}") for i in range(8)]
1160
1161     def eq_from(self, part_pts, a, b, part_ops):
1162         return [self.part_pts.eq(part_pts)] + \
1163                [self.a.eq(a), self.b.eq(b)] + \
1164                [self.part_ops[i].eq(part_ops[i])
1165                                      for i in range(len(self.part_ops))]
1166
1167     def eq(self, rhs):
1168         return self.eq_from(rhs.part_pts, rhs.a, rhs.b, rhs.part_ops)
1169
1170
1171 class OutputData:
1172
1173     def __init__(self):
1174         self.intermediate_output = Signal(128) # needed for unit tests
1175         self.output = Signal(64)
1176
1177     def eq(self, rhs):
1178         return [self.intermediate_output.eq(rhs.intermediate_output),
1179                 self.output.eq(rhs.output)]
1180
1181
1182 class AllTerms(Elaboratable):
1183     """Set of terms to be added together
1184     """
1185
1186     def __init__(self, n_inputs, output_width, n_parts):
1187         """Create an ``AddReduce``.
1188
1189         :param inputs: input ``Signal``s to be summed.
1190         :param output_width: bit-width of ``output``.
1191         :param register_levels: List of nesting levels that should have
1192             pipeline registers.
1193         :param partition_points: the input partition points.
1194         """
1195         self.n_inputs = n_inputs
1196         self.n_parts = n_parts
1197         self.output_width = output_width
1198
1199         self.i = self.ispec()
1200         self.o = self.ospec()
1201
1202     def setup(self, m, i):
1203         m.submodules.allterms = self
1204         m.d.comb += self.i.eq(i)
1205
1206     def process(self, i):
1207         return self.o
1208
1209     def ispec(self):
1210         return InputData()
1211
1212     def ospec(self):
1213         return AddReduceData(self.i.part_pts, self.n_inputs,
1214                              self.output_width, self.n_parts)
1215
1216     def elaborate(self, platform):
1217         m = Module()
1218
1219         eps = self.i.part_pts
1220
1221         # collect part-bytes
1222         pbs = Signal(8, reset_less=True)
1223         tl = []
1224         for i in range(8):
1225             pb = Signal(name="pb%d" % i, reset_less=True)
1226             m.d.comb += pb.eq(eps.part_byte(i))
1227             tl.append(pb)
1228         m.d.comb += pbs.eq(Cat(*tl))
1229
1230         # local variables
1231         signs = []
1232         for i in range(8):
1233             s = Signs()
1234             signs.append(s)
1235             setattr(m.submodules, "signs%d" % i, s)
1236             m.d.comb += s.part_ops.eq(self.i.part_ops[i])
1237
1238         m.submodules.part_8 = part_8 = Part(eps, 128, 8, 8)
1239         m.submodules.part_16 = part_16 = Part(eps, 128, 4, 8)
1240         m.submodules.part_32 = part_32 = Part(eps, 128, 2, 8)
1241         m.submodules.part_64 = part_64 = Part(eps, 128, 1, 8)
1242         nat_l, nbt_l, nla_l, nlb_l = [], [], [], []
1243         for mod in [part_8, part_16, part_32, part_64]:
1244             m.d.comb += mod.a.eq(self.i.a)
1245             m.d.comb += mod.b.eq(self.i.b)
1246             for i in range(len(signs)):
1247                 m.d.comb += mod.a_signed[i].eq(signs[i].a_signed)
1248                 m.d.comb += mod.b_signed[i].eq(signs[i].b_signed)
1249             m.d.comb += mod.pbs.eq(pbs)
1250             nat_l.append(mod.not_a_term)
1251             nbt_l.append(mod.not_b_term)
1252             nla_l.append(mod.neg_lsb_a_term)
1253             nlb_l.append(mod.neg_lsb_b_term)
1254
1255         terms = []
1256
1257         for a_index in range(8):
1258             t = ProductTerms(8, 128, 8, a_index, 8)
1259             setattr(m.submodules, "terms_%d" % a_index, t)
1260
1261             m.d.comb += t.a.eq(self.i.a)
1262             m.d.comb += t.b.eq(self.i.b)
1263             m.d.comb += t.pb_en.eq(pbs)
1264
1265             for term in t.terms:
1266                 terms.append(term)
1267
1268         # it's fine to bitwise-or data together since they are never enabled
1269         # at the same time
1270         m.submodules.nat_or = nat_or = OrMod(128)
1271         m.submodules.nbt_or = nbt_or = OrMod(128)
1272         m.submodules.nla_or = nla_or = OrMod(128)
1273         m.submodules.nlb_or = nlb_or = OrMod(128)
1274         for l, mod in [(nat_l, nat_or),
1275                              (nbt_l, nbt_or),
1276                              (nla_l, nla_or),
1277                              (nlb_l, nlb_or)]:
1278             for i in range(len(l)):
1279                 m.d.comb += mod.orin[i].eq(l[i])
1280             terms.append(mod.orout)
1281
1282         # copy the intermediate terms to the output
1283         for i, value in enumerate(terms):
1284             m.d.comb += self.o.terms[i].eq(value)
1285
1286         # copy reg part points and part ops to output
1287         m.d.comb += self.o.part_pts.eq(eps)
1288         m.d.comb += [self.o.part_ops[i].eq(self.i.part_ops[i])
1289                                      for i in range(len(self.i.part_ops))]
1290
1291         return m
1292
1293
1294 class Intermediates(Elaboratable):
1295     """ Intermediate output modules
1296     """
1297
1298     def __init__(self, output_width, n_parts, part_pts):
1299         self.part_pts = part_pts
1300         self.output_width = output_width
1301         self.n_parts = n_parts
1302
1303         self.i = self.ispec()
1304         self.o = self.ospec()
1305
1306     def ispec(self):
1307         return FinalReduceData(self.part_pts, self.output_width, self.n_parts)
1308
1309     def ospec(self):
1310         return IntermediateData(self.part_pts, self.output_width, self.n_parts)
1311
1312     def setup(self, m, i):
1313         m.submodules.intermediates = self
1314         m.d.comb += self.i.eq(i)
1315
1316     def process(self, i):
1317         return self.o
1318
1319     def elaborate(self, platform):
1320         m = Module()
1321
1322         out_part_ops = self.i.part_ops
1323         out_part_pts = self.i.part_pts
1324
1325         # create _output_64
1326         m.submodules.io64 = io64 = IntermediateOut(64, 128, 1)
1327         m.d.comb += io64.intermed.eq(self.i.output)
1328         for i in range(8):
1329             m.d.comb += io64.part_ops[i].eq(out_part_ops[i])
1330         m.d.comb += self.o.outputs[3].eq(io64.output)
1331
1332         # create _output_32
1333         m.submodules.io32 = io32 = IntermediateOut(32, 128, 2)
1334         m.d.comb += io32.intermed.eq(self.i.output)
1335         for i in range(8):
1336             m.d.comb += io32.part_ops[i].eq(out_part_ops[i])
1337         m.d.comb += self.o.outputs[2].eq(io32.output)
1338
1339         # create _output_16
1340         m.submodules.io16 = io16 = IntermediateOut(16, 128, 4)
1341         m.d.comb += io16.intermed.eq(self.i.output)
1342         for i in range(8):
1343             m.d.comb += io16.part_ops[i].eq(out_part_ops[i])
1344         m.d.comb += self.o.outputs[1].eq(io16.output)
1345
1346         # create _output_8
1347         m.submodules.io8 = io8 = IntermediateOut(8, 128, 8)
1348         m.d.comb += io8.intermed.eq(self.i.output)
1349         for i in range(8):
1350             m.d.comb += io8.part_ops[i].eq(out_part_ops[i])
1351         m.d.comb += self.o.outputs[0].eq(io8.output)
1352
1353         for i in range(8):
1354             m.d.comb += self.o.part_ops[i].eq(out_part_ops[i])
1355         m.d.comb += self.o.part_pts.eq(out_part_pts)
1356         m.d.comb += self.o.intermediate_output.eq(self.i.output)
1357
1358         return m
1359
1360
1361 class Mul8_16_32_64(Elaboratable):
1362     """Signed/Unsigned 8/16/32/64-bit partitioned integer multiplier.
1363
1364     Supports partitioning into any combination of 8, 16, 32, and 64-bit
1365     partitions on naturally-aligned boundaries. Supports the operation being
1366     set for each partition independently.
1367
1368     :attribute part_pts: the input partition points. Has a partition point at
1369         multiples of 8 in 0 < i < 64. Each partition point's associated
1370         ``Value`` is a ``Signal``. Modification not supported, except for by
1371         ``Signal.eq``.
1372     :attribute part_ops: the operation for each byte. The operation for a
1373         particular partition is selected by assigning the selected operation
1374         code to each byte in the partition. The allowed operation codes are:
1375
1376         :attribute OP_MUL_LOW: the LSB half of the product. Equivalent to
1377             RISC-V's `mul` instruction.
1378         :attribute OP_MUL_SIGNED_HIGH: the MSB half of the product where both
1379             ``a`` and ``b`` are signed. Equivalent to RISC-V's `mulh`
1380             instruction.
1381         :attribute OP_MUL_SIGNED_UNSIGNED_HIGH: the MSB half of the product
1382             where ``a`` is signed and ``b`` is unsigned. Equivalent to RISC-V's
1383             `mulhsu` instruction.
1384         :attribute OP_MUL_UNSIGNED_HIGH: the MSB half of the product where both
1385             ``a`` and ``b`` are unsigned. Equivalent to RISC-V's `mulhu`
1386             instruction.
1387     """
1388
1389     def __init__(self, register_levels=()):
1390         """ register_levels: specifies the points in the cascade at which
1391             flip-flops are to be inserted.
1392         """
1393
1394         # parameter(s)
1395         self.register_levels = list(register_levels)
1396
1397         self.i = self.ispec()
1398         self.o = self.ospec()
1399
1400         # inputs
1401         self.part_pts = self.i.part_pts
1402         self.part_ops = self.i.part_ops
1403         self.a = self.i.a
1404         self.b = self.i.b
1405
1406         # output
1407         self.intermediate_output = self.o.intermediate_output
1408         self.output = self.o.output
1409
1410     def ispec(self):
1411         return InputData()
1412
1413     def ospec(self):
1414         return OutputData()
1415
1416     def elaborate(self, platform):
1417         m = Module()
1418
1419         part_pts = self.part_pts
1420
1421         n_inputs = 64 + 4
1422         n_parts = 8
1423         t = AllTerms(n_inputs, 128, n_parts)
1424         t.setup(m, self.i)
1425
1426         terms = t.o.terms
1427
1428         at = AddReduceInternal(t.process(self.i), 128, partition_step=2)
1429
1430         i = at.i
1431         for idx in range(len(at.levels)):
1432             mcur = at.levels[idx]
1433             mcur.setup(m, i)
1434             o = mcur.ospec()
1435             if idx in self.register_levels:
1436                 m.d.sync += o.eq(mcur.process(i))
1437             else:
1438                 m.d.comb += o.eq(mcur.process(i))
1439             i = o # for next loop
1440
1441         interm = Intermediates(128, 8, part_pts)
1442         interm.setup(m, i)
1443         o = interm.process(interm.i)
1444
1445         # final output
1446         finalout = FinalOut(128, 8, part_pts)
1447         finalout.setup(m, o)
1448         m.d.comb += self.o.eq(finalout.process(o))
1449
1450         return m
1451
1452
1453 if __name__ == "__main__":
1454     m = Mul8_16_32_64()
1455     main(m, ports=[m.a,
1456                    m.b,
1457                    m.intermediate_output,
1458                    m.output,
1459                    *m.part_ops,
1460                    *m.part_pts.values()])