X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fieee754%2Fpart_mul_add%2Fmultiply.py;h=672bbfd33acdef4510167e6ec1d8c78e0bf3603f;hb=5e1e5b602e087661f3ee58d608ae1396ecbc14b3;hp=17eec4e5d4e78d3d0d763a04734eb9d709e4110c;hpb=43db70eac45b2e86180faf1b9288817b2fd80a46;p=ieee754fpu.git diff --git a/src/ieee754/part_mul_add/multiply.py b/src/ieee754/part_mul_add/multiply.py index 17eec4e5..672bbfd3 100644 --- a/src/ieee754/part_mul_add/multiply.py +++ b/src/ieee754/part_mul_add/multiply.py @@ -71,17 +71,20 @@ class PartitionPoints(dict): for point, enabled in self.items(): yield enabled.eq(rhs[point]) - def as_mask(self, width): + def as_mask(self, width, mul=1): """Create a bit-mask from `self`. Each bit in the returned mask is clear only if the partition point at the same bit-index is enabled. :param width: the bit width of the resulting mask + :param mul: a "multiplier" which in-place expands the partition points + typically set to "2" when used for multipliers """ bits = [] for i in range(width): - if i in self: + i /= mul + if i.is_integer() and int(i) in self: bits.append(~self[i]) else: bits.append(True) @@ -132,11 +135,11 @@ class FullAdder(Elaboratable): :param width: the bit width of the input and output """ - self.in0 = Signal(width) - self.in1 = Signal(width) - self.in2 = Signal(width) - self.sum = Signal(width) - self.carry = Signal(width) + self.in0 = Signal(width, reset_less=True) + self.in1 = Signal(width, reset_less=True) + self.in2 = Signal(width, reset_less=True) + self.sum = Signal(width, reset_less=True) + self.carry = Signal(width, reset_less=True) def elaborate(self, platform): """Elaborate this module.""" @@ -227,16 +230,19 @@ class PartitionedAdder(Elaboratable): supported, except for by ``Signal.eq``. """ - def __init__(self, width, partition_points): + def __init__(self, width, partition_points, partition_step=1): """Create a ``PartitionedAdder``. :param width: the bit width of the input and output :param partition_points: the input partition points + :param partition_step: a multiplier (typically double) step + which in-place "expands" the partition points """ self.width = width - self.a = Signal(width) - self.b = Signal(width) - self.output = Signal(width) + self.pmul = partition_step + self.a = Signal(width, reset_less=True) + self.b = Signal(width, reset_less=True) + self.output = Signal(width, reset_less=True) self.partition_points = PartitionPoints(partition_points) if not self.partition_points.fits_in_width(width): raise ValueError("partition_points doesn't fit in width") @@ -246,17 +252,14 @@ class PartitionedAdder(Elaboratable): expanded_width += 1 expanded_width += 1 self._expanded_width = expanded_width - # XXX these have to remain here due to some horrible nmigen - # simulation bugs involving sync. it is *not* necessary to - # have them here, they should (under normal circumstances) - # be moved into elaborate, as they are entirely local - self._expanded_a = Signal(expanded_width) # includes extra part-points - self._expanded_b = Signal(expanded_width) # likewise. - self._expanded_o = Signal(expanded_width) # likewise. def elaborate(self, platform): """Elaborate this module.""" m = Module() + expanded_a = Signal(self._expanded_width, reset_less=True) + expanded_b = Signal(self._expanded_width, reset_less=True) + expanded_o = Signal(self._expanded_width, reset_less=True) + expanded_index = 0 # store bits in a list, use Cat later. graphviz is much cleaner al, bl, ol, ea, eb, eo = [],[],[],[],[],[] @@ -270,17 +273,18 @@ class PartitionedAdder(Elaboratable): # carry has been carried *over* the break point. for i in range(self.width): - if i in self.partition_points: + pi = i/self.pmul # double the range of the partition point test + if pi.is_integer() and pi in self.partition_points: # add extra bit set to 0 + 0 for enabled partition points # and 1 + 0 for disabled partition points - ea.append(self._expanded_a[expanded_index]) - al.append(~self.partition_points[i]) # add extra bit in a - eb.append(self._expanded_b[expanded_index]) + ea.append(expanded_a[expanded_index]) + al.append(~self.partition_points[pi]) # add extra bit in a + eb.append(expanded_b[expanded_index]) bl.append(C(0)) # yes, add a zero expanded_index += 1 # skip the extra point. NOT in the output - ea.append(self._expanded_a[expanded_index]) - eb.append(self._expanded_b[expanded_index]) - eo.append(self._expanded_o[expanded_index]) + ea.append(expanded_a[expanded_index]) + eb.append(expanded_b[expanded_index]) + eo.append(expanded_o[expanded_index]) al.append(self.a[i]) bl.append(self.b[i]) ol.append(self.output[i]) @@ -293,8 +297,7 @@ class PartitionedAdder(Elaboratable): # use only one addition to take advantage of look-ahead carry and # special hardware on FPGAs - m.d.comb += self._expanded_o.eq( - self._expanded_a + self._expanded_b) + m.d.comb += expanded_o.eq(expanded_a + expanded_b) return m @@ -302,20 +305,42 @@ FULL_ADDER_INPUT_COUNT = 3 class AddReduceData: - def __init__(self, ppoints, n_inputs, output_width, n_parts): - self.part_ops = [Signal(2, name=f"part_ops_{i}") + def __init__(self, part_pts, n_inputs, output_width, n_parts): + self.part_ops = [Signal(2, name=f"part_ops_{i}", reset_less=True) for i in range(n_parts)] - self.inputs = [Signal(output_width, name=f"inputs[{i}]") - for i in range(n_inputs)] - self.reg_partition_points = ppoints.like() + self.terms = [Signal(output_width, name=f"inputs_{i}", + reset_less=True) + for i in range(n_inputs)] + self.part_pts = part_pts.like() + + def eq_from(self, part_pts, inputs, part_ops): + return [self.part_pts.eq(part_pts)] + \ + [self.terms[i].eq(inputs[i]) + for i in range(len(self.terms))] + \ + [self.part_ops[i].eq(part_ops[i]) + for i in range(len(self.part_ops))] def eq(self, rhs): - return [self.reg_partition_points.eq(rhs.reg_partition_points)] + \ - [self.inputs[i].eq(rhs.inputs[i]) - for i in range(len(self.inputs))] + \ - [self.part_ops[i].eq(rhs.part_ops[i]) + return self.eq_from(rhs.part_pts, rhs.terms, rhs.part_ops) + + +class FinalReduceData: + + def __init__(self, part_pts, output_width, n_parts): + self.part_ops = [Signal(2, name=f"part_ops_{i}", reset_less=True) + for i in range(n_parts)] + self.output = Signal(output_width, reset_less=True) + self.part_pts = part_pts.like() + + def eq_from(self, part_pts, output, part_ops): + return [self.part_pts.eq(part_pts)] + \ + [self.output.eq(output)] + \ + [self.part_ops[i].eq(part_ops[i]) for i in range(len(self.part_ops))] + def eq(self, rhs): + return self.eq_from(rhs.part_pts, rhs.output, rhs.part_ops) + class FinalAdd(Elaboratable): """ Final stage of add reduce @@ -325,35 +350,41 @@ class FinalAdd(Elaboratable): partition_points): self.i = AddReduceData(partition_points, n_inputs, output_width, n_parts) + self.o = FinalReduceData(partition_points, output_width, n_parts) + self.output_width = output_width self.n_inputs = n_inputs self.n_parts = n_parts - self._resized_inputs = self.i.inputs self.register_levels = list(register_levels) - self.output = Signal(output_width) self.partition_points = PartitionPoints(partition_points) if not self.partition_points.fits_in_width(output_width): raise ValueError("partition_points doesn't fit in output_width") - self.intermediate_terms = [] def elaborate(self, platform): """Elaborate this module.""" m = Module() + output_width = self.output_width + output = Signal(output_width, reset_less=True) if self.n_inputs == 0: # use 0 as the default output value - m.d.comb += self.output.eq(0) + m.d.comb += output.eq(0) elif self.n_inputs == 1: # handle single input - m.d.comb += self.output.eq(self._resized_inputs[0]) + m.d.comb += output.eq(self.i.terms[0]) else: # base case for adding 2 inputs assert self.n_inputs == 2 - adder = PartitionedAdder(len(self.output), - self.i.reg_partition_points) + adder = PartitionedAdder(output_width, + self.i.part_pts, 2) m.submodules.final_adder = adder - m.d.comb += adder.a.eq(self._resized_inputs[0]) - m.d.comb += adder.b.eq(self._resized_inputs[1]) - m.d.comb += self.output.eq(adder.output) + m.d.comb += adder.a.eq(self.i.terms[0]) + m.d.comb += adder.b.eq(self.i.terms[1]) + m.d.comb += output.eq(adder.output) + + # create output + m.d.comb += self.o.eq_from(self.i.part_pts, output, + self.i.part_ops) + return m @@ -384,7 +415,6 @@ class AddReduceSingle(Elaboratable): self.output_width = output_width self.i = AddReduceData(partition_points, n_inputs, output_width, n_parts) - self._resized_inputs = self.i.inputs self.register_levels = list(register_levels) self.partition_points = PartitionPoints(partition_points) if not self.partition_points.fits_in_width(output_width): @@ -396,14 +426,20 @@ class AddReduceSingle(Elaboratable): raise ValueError( "not enough adder levels for specified register levels") - # this is annoying. we have to create the modules (and terms) - # because we need to know what they are (in order to set up the - # interconnects back in AddReduce), but cannot do the m.d.comb += - # etc because this is not in elaboratable. self.groups = AddReduceSingle.full_adder_groups(n_inputs) - self._intermediate_terms = [] - if len(self.groups) != 0: - self.create_next_terms() + n_terms = AddReduceSingle.calc_n_inputs(n_inputs, self.groups) + self.o = AddReduceData(partition_points, n_terms, output_width, n_parts) + + @staticmethod + def calc_n_inputs(n_inputs, groups): + retval = len(groups)*2 + if n_inputs % FULL_ADDER_INPUT_COUNT == 1: + retval += 1 + elif n_inputs % FULL_ADDER_INPUT_COUNT == 2: + retval += 2 + else: + assert n_inputs % FULL_ADDER_INPUT_COUNT == 0 + return retval @staticmethod def get_max_level(input_count): @@ -428,67 +464,67 @@ class AddReduceSingle(Elaboratable): input_count - FULL_ADDER_INPUT_COUNT + 1, FULL_ADDER_INPUT_COUNT) - def elaborate(self, platform): - """Elaborate this module.""" - m = Module() - - for (value, term) in self._intermediate_terms: - m.d.comb += term.eq(value) - - mask = self.i.reg_partition_points.as_mask(self.output_width) - m.d.comb += self.part_mask.eq(mask) - - # add and link the intermediate term modules - for i, (iidx, adder_i) in enumerate(self.adders): - setattr(m.submodules, f"adder_{i}", adder_i) - - m.d.comb += adder_i.in0.eq(self._resized_inputs[iidx]) - m.d.comb += adder_i.in1.eq(self._resized_inputs[iidx + 1]) - m.d.comb += adder_i.in2.eq(self._resized_inputs[iidx + 2]) - m.d.comb += adder_i.mask.eq(self.part_mask) - - return m - def create_next_terms(self): - - # go on to prepare recursive case - intermediate_terms = [] - _intermediate_terms = [] - - def add_intermediate_term(value): - intermediate_term = Signal( - self.output_width, - name=f"intermediate_terms[{len(intermediate_terms)}]") - _intermediate_terms.append((value, intermediate_term)) - intermediate_terms.append(intermediate_term) - - # store mask in intermediary (simplifies graph) - self.part_mask = Signal(self.output_width, reset_less=True) + """ create next intermediate terms, for linking up in elaborate, below + """ + terms = [] + adders = [] # create full adders for this recursive level. # this shrinks N terms to 2 * (N // 3) plus the remainder - self.adders = [] for i in self.groups: adder_i = MaskedFullAdder(self.output_width) - self.adders.append((i, adder_i)) + adders.append((i, adder_i)) # add both the sum and the masked-carry to the next level. # 3 inputs have now been reduced to 2... - add_intermediate_term(adder_i.sum) - add_intermediate_term(adder_i.mcarry) + terms.append(adder_i.sum) + terms.append(adder_i.mcarry) # handle the remaining inputs. if self.n_inputs % FULL_ADDER_INPUT_COUNT == 1: - add_intermediate_term(self._resized_inputs[-1]) + terms.append(self.i.terms[-1]) elif self.n_inputs % FULL_ADDER_INPUT_COUNT == 2: # Just pass the terms to the next layer, since we wouldn't gain # anything by using a half adder since there would still be 2 terms # and just passing the terms to the next layer saves gates. - add_intermediate_term(self._resized_inputs[-2]) - add_intermediate_term(self._resized_inputs[-1]) + terms.append(self.i.terms[-2]) + terms.append(self.i.terms[-1]) else: assert self.n_inputs % FULL_ADDER_INPUT_COUNT == 0 - self.intermediate_terms = intermediate_terms - self._intermediate_terms = _intermediate_terms + return terms, adders + + def elaborate(self, platform): + """Elaborate this module.""" + m = Module() + + terms, adders = self.create_next_terms() + + # copy the intermediate terms to the output + for i, value in enumerate(terms): + m.d.comb += self.o.terms[i].eq(value) + + # copy reg part points and part ops to output + m.d.comb += self.o.part_pts.eq(self.i.part_pts) + m.d.comb += [self.o.part_ops[i].eq(self.i.part_ops[i]) + for i in range(len(self.i.part_ops))] + + # set up the partition mask (for the adders) + part_mask = Signal(self.output_width, reset_less=True) + + # get partition points as a mask + mask = self.i.part_pts.as_mask(self.output_width, mul=2) + m.d.comb += part_mask.eq(mask) + + # add and link the intermediate term modules + for i, (iidx, adder_i) in enumerate(adders): + setattr(m.submodules, f"adder_{i}", adder_i) + + m.d.comb += adder_i.in0.eq(self.i.terms[iidx]) + m.d.comb += adder_i.in1.eq(self.i.terms[iidx + 1]) + m.d.comb += adder_i.in2.eq(self.i.terms[iidx + 2]) + m.d.comb += adder_i.mask.eq(part_mask) + + return m class AddReduce(Elaboratable): @@ -515,9 +551,8 @@ class AddReduce(Elaboratable): """ self.inputs = inputs self.part_ops = part_ops - self.out_part_ops = [Signal(2, name=f"out_part_ops_{i}") - for i in range(len(part_ops))] - self.output = Signal(output_width) + n_parts = len(part_ops) + self.o = FinalReduceData(partition_points, output_width, n_parts) self.output_width = output_width self.register_levels = register_levels self.partition_points = partition_points @@ -541,27 +576,26 @@ class AddReduce(Elaboratable): mods = [] next_levels = self.register_levels partition_points = self.partition_points - inputs = self.inputs part_ops = self.part_ops n_parts = len(part_ops) + inputs = self.inputs + ilen = len(inputs) while True: - ilen = len(inputs) + groups = AddReduceSingle.full_adder_groups(len(inputs)) + if len(groups) == 0: + break next_level = AddReduceSingle(ilen, self.output_width, n_parts, next_levels, partition_points) mods.append(next_level) next_levels = list(AddReduce.next_register_levels(next_levels)) - partition_points = next_level.i.reg_partition_points - inputs = next_level.intermediate_terms + partition_points = next_level.i.part_pts + inputs = next_level.o.terms ilen = len(inputs) part_ops = next_level.i.part_ops - groups = AddReduceSingle.full_adder_groups(len(inputs)) - if len(groups) == 0: - break - if ilen != 0: - next_level = FinalAdd(ilen, self.output_width, n_parts, - next_levels, partition_points) - mods.append(next_level) + next_level = FinalAdd(ilen, self.output_width, n_parts, + next_levels, partition_points) + mods.append(next_level) self.levels = mods @@ -575,29 +609,21 @@ class AddReduce(Elaboratable): partition_points = self.partition_points inputs = self.inputs part_ops = self.part_ops - for i in range(len(self.levels)): - mcur = self.levels[i] - inassign = [mcur._resized_inputs[i].eq(inputs[i]) - for i in range(len(inputs))] - copy_part_ops = [mcur.i.part_ops[i].eq(part_ops[i]) - for i in range(len(part_ops))] + n_parts = len(part_ops) + n_inputs = len(inputs) + output_width = self.output_width + i = AddReduceData(partition_points, n_inputs, output_width, n_parts) + m.d.comb += i.eq_from(partition_points, inputs, part_ops) + for idx in range(len(self.levels)): + mcur = self.levels[idx] if 0 in mcur.register_levels: - m.d.sync += copy_part_ops - m.d.sync += inassign - m.d.sync += mcur.i.reg_partition_points.eq(partition_points) + m.d.sync += mcur.i.eq(i) else: - m.d.comb += copy_part_ops - m.d.comb += inassign - m.d.comb += mcur.i.reg_partition_points.eq(partition_points) - partition_points = mcur.i.reg_partition_points - inputs = mcur.intermediate_terms - part_ops = mcur.i.part_ops + m.d.comb += mcur.i.eq(i) + i = mcur.o # for next loop # output comes from last module - m.d.comb += self.output.eq(next_level.output) - copy_part_ops = [self.out_part_ops[i].eq(next_level.i.part_ops[i]) - for i in range(len(self.part_ops))] - m.d.comb += copy_part_ops + m.d.comb += self.o.eq(i) return m @@ -760,23 +786,24 @@ class LSBNegTerm(Elaboratable): class Parts(Elaboratable): - def __init__(self, pbwid, epps, n_parts): + def __init__(self, pbwid, part_pts, n_parts): self.pbwid = pbwid # inputs - self.epps = PartitionPoints.like(epps, name="epps") # expanded points + self.part_pts = PartitionPoints.like(part_pts) # outputs - self.parts = [Signal(name=f"part_{i}") for i in range(n_parts)] + self.parts = [Signal(name=f"part_{i}", reset_less=True) + for i in range(n_parts)] def elaborate(self, platform): m = Module() - epps, parts = self.epps, self.parts + part_pts, parts = self.part_pts, self.parts # collect part-bytes (double factor because the input is extended) pbs = Signal(self.pbwid, reset_less=True) tl = [] for i in range(self.pbwid): pb = Signal(name="pb%d" % i, reset_less=True) - m.d.comb += pb.eq(epps.part_byte(i, mfactor=2)) # double + m.d.comb += pb.eq(part_pts.part_byte(i)) tl.append(pb) m.d.comb += pbs.eq(Cat(*tl)) @@ -812,33 +839,36 @@ class Part(Elaboratable): the extra terms - as separate terms - are then thrown at the AddReduce alongside the multiplication part-results. """ - def __init__(self, epps, width, n_parts, n_levels, pbwid): + def __init__(self, part_pts, width, n_parts, n_levels, pbwid): self.pbwid = pbwid - self.epps = epps + self.part_pts = part_pts # inputs - self.a = Signal(64) - self.b = Signal(64) - self.a_signed = [Signal(name=f"a_signed_{i}") for i in range(8)] - self.b_signed = [Signal(name=f"_b_signed_{i}") for i in range(8)] + self.a = Signal(64, reset_less=True) + self.b = Signal(64, reset_less=True) + self.a_signed = [Signal(name=f"a_signed_{i}", reset_less=True) + for i in range(8)] + self.b_signed = [Signal(name=f"_b_signed_{i}", reset_less=True) + for i in range(8)] self.pbs = Signal(pbwid, reset_less=True) # outputs - self.parts = [Signal(name=f"part_{i}") for i in range(n_parts)] + self.parts = [Signal(name=f"part_{i}", reset_less=True) + for i in range(n_parts)] - self.not_a_term = Signal(width) - self.neg_lsb_a_term = Signal(width) - self.not_b_term = Signal(width) - self.neg_lsb_b_term = Signal(width) + self.not_a_term = Signal(width, reset_less=True) + self.neg_lsb_a_term = Signal(width, reset_less=True) + self.not_b_term = Signal(width, reset_less=True) + self.neg_lsb_b_term = Signal(width, reset_less=True) def elaborate(self, platform): m = Module() pbs, parts = self.pbs, self.parts - epps = self.epps - m.submodules.p = p = Parts(self.pbwid, epps, len(parts)) - m.d.comb += p.epps.eq(epps) + part_pts = self.part_pts + m.submodules.p = p = Parts(self.pbwid, part_pts, len(parts)) + m.d.comb += p.part_pts.eq(part_pts) parts = p.parts byte_count = 8 // len(parts) @@ -918,22 +948,51 @@ class FinalOut(Elaboratable): that some partitions requested 8-bit computation whilst others requested 16 or 32 bit. """ - def __init__(self, out_wid): - # inputs - self.d8 = [Signal(name=f"d8_{i}", reset_less=True) for i in range(8)] - self.d16 = [Signal(name=f"d16_{i}", reset_less=True) for i in range(4)] - self.d32 = [Signal(name=f"d32_{i}", reset_less=True) for i in range(2)] - - self.i8 = Signal(out_wid, reset_less=True) - self.i16 = Signal(out_wid, reset_less=True) - self.i32 = Signal(out_wid, reset_less=True) - self.i64 = Signal(out_wid, reset_less=True) - + def __init__(self, output_width, n_parts, part_pts): + self.part_pts = part_pts + self.i = IntermediateData(part_pts, output_width, n_parts) + self.out_wid = output_width//2 # output - self.out = Signal(out_wid, reset_less=True) + self.out = Signal(self.out_wid, reset_less=True) + self.intermediate_output = Signal(output_width, reset_less=True) def elaborate(self, platform): m = Module() + + part_pts = self.part_pts + m.submodules.p_8 = p_8 = Parts(8, part_pts, 8) + m.submodules.p_16 = p_16 = Parts(8, part_pts, 4) + m.submodules.p_32 = p_32 = Parts(8, part_pts, 2) + m.submodules.p_64 = p_64 = Parts(8, part_pts, 1) + + out_part_pts = self.i.part_pts + + # temporaries + d8 = [Signal(name=f"d8_{i}", reset_less=True) for i in range(8)] + d16 = [Signal(name=f"d16_{i}", reset_less=True) for i in range(4)] + d32 = [Signal(name=f"d32_{i}", reset_less=True) for i in range(2)] + + i8 = Signal(self.out_wid, reset_less=True) + i16 = Signal(self.out_wid, reset_less=True) + i32 = Signal(self.out_wid, reset_less=True) + i64 = Signal(self.out_wid, reset_less=True) + + m.d.comb += p_8.part_pts.eq(out_part_pts) + m.d.comb += p_16.part_pts.eq(out_part_pts) + m.d.comb += p_32.part_pts.eq(out_part_pts) + m.d.comb += p_64.part_pts.eq(out_part_pts) + + for i in range(len(p_8.parts)): + m.d.comb += d8[i].eq(p_8.parts[i]) + for i in range(len(p_16.parts)): + m.d.comb += d16[i].eq(p_16.parts[i]) + for i in range(len(p_32.parts)): + m.d.comb += d32[i].eq(p_32.parts[i]) + m.d.comb += i8.eq(self.i.outputs[0]) + m.d.comb += i16.eq(self.i.outputs[1]) + m.d.comb += i32.eq(self.i.outputs[2]) + m.d.comb += i64.eq(self.i.outputs[3]) + ol = [] for i in range(8): # select one of the outputs: d8 selects i8, d16 selects i16 @@ -943,13 +1002,12 @@ class FinalOut(Elaboratable): # if neither d8 nor d16 are set, d32 selects either i32 or i64. op = Signal(8, reset_less=True, name="op_%d" % i) m.d.comb += op.eq( - Mux(self.d8[i] | self.d16[i // 2], - Mux(self.d8[i], self.i8.part(i * 8, 8), - self.i16.part(i * 8, 8)), - Mux(self.d32[i // 4], self.i32.part(i * 8, 8), - self.i64.part(i * 8, 8)))) + Mux(d8[i] | d16[i // 2], + Mux(d8[i], i8.part(i * 8, 8), i16.part(i * 8, 8)), + Mux(d32[i // 4], i32.part(i * 8, 8), i64.part(i * 8, 8)))) ol.append(op) m.d.comb += self.out.eq(Cat(*ol)) + m.d.comb += self.intermediate_output.eq(self.i.intermediate_output) return m @@ -996,82 +1054,92 @@ class Signs(Elaboratable): return m -class Mul8_16_32_64(Elaboratable): - """Signed/Unsigned 8/16/32/64-bit partitioned integer multiplier. - - Supports partitioning into any combination of 8, 16, 32, and 64-bit - partitions on naturally-aligned boundaries. Supports the operation being - set for each partition independently. +class IntermediateData: - :attribute part_pts: the input partition points. Has a partition point at - multiples of 8 in 0 < i < 64. Each partition point's associated - ``Value`` is a ``Signal``. Modification not supported, except for by - ``Signal.eq``. - :attribute part_ops: the operation for each byte. The operation for a - particular partition is selected by assigning the selected operation - code to each byte in the partition. The allowed operation codes are: + def __init__(self, part_pts, output_width, n_parts): + self.part_ops = [Signal(2, name=f"part_ops_{i}", reset_less=True) + for i in range(n_parts)] + self.part_pts = part_pts.like() + self.outputs = [Signal(output_width, name="io%d" % i, reset_less=True) + for i in range(4)] + # intermediates (needed for unit tests) + self.intermediate_output = Signal(output_width) + + def eq_from(self, part_pts, outputs, intermediate_output, + part_ops): + return [self.part_pts.eq(part_pts)] + \ + [self.intermediate_output.eq(intermediate_output)] + \ + [self.outputs[i].eq(outputs[i]) + for i in range(4)] + \ + [self.part_ops[i].eq(part_ops[i]) + for i in range(len(self.part_ops))] - :attribute OP_MUL_LOW: the LSB half of the product. Equivalent to - RISC-V's `mul` instruction. - :attribute OP_MUL_SIGNED_HIGH: the MSB half of the product where both - ``a`` and ``b`` are signed. Equivalent to RISC-V's `mulh` - instruction. - :attribute OP_MUL_SIGNED_UNSIGNED_HIGH: the MSB half of the product - where ``a`` is signed and ``b`` is unsigned. Equivalent to RISC-V's - `mulhsu` instruction. - :attribute OP_MUL_UNSIGNED_HIGH: the MSB half of the product where both - ``a`` and ``b`` are unsigned. Equivalent to RISC-V's `mulhu` - instruction. - """ + def eq(self, rhs): + return self.eq_from(rhs.part_pts, rhs.outputs, + rhs.intermediate_output, rhs.part_ops) - def __init__(self, register_levels=()): - """ register_levels: specifies the points in the cascade at which - flip-flops are to be inserted. - """ - # parameter(s) - self.register_levels = list(register_levels) +class AllTermsData: - # inputs - self.part_pts = PartitionPoints() - for i in range(8, 64, 8): - self.part_pts[i] = Signal(name=f"part_pts_{i}") - self.part_ops = [Signal(2, name=f"part_ops_{i}") for i in range(8)] + def __init__(self, partition_points): self.a = Signal(64) self.b = Signal(64) + self.part_pts = partition_points.like() + self.part_ops = [Signal(2, name=f"part_ops_{i}") for i in range(8)] - # intermediates (needed for unit tests) - self._intermediate_output = Signal(128) + def eq_from(self, part_pts, inputs, part_ops): + return [self.part_pts.eq(part_pts)] + \ + [self.a.eq(a), self.b.eq(b)] + \ + [self.part_ops[i].eq(part_ops[i]) + for i in range(len(self.part_ops))] - # output - self.output = Signal(64) + def eq(self, rhs): + return self.eq_from(rhs.part_pts, rhs.a, rhs.b, rhs.part_ops) + + +class AllTerms(Elaboratable): + """Set of terms to be added together + """ + + def __init__(self, n_inputs, output_width, n_parts, register_levels, + partition_points): + """Create an ``AddReduce``. + + :param inputs: input ``Signal``s to be summed. + :param output_width: bit-width of ``output``. + :param register_levels: List of nesting levels that should have + pipeline registers. + :param partition_points: the input partition points. + """ + self.i = AllTermsData(partition_points) + self.register_levels = register_levels + self.n_inputs = n_inputs + self.n_parts = n_parts + self.output_width = output_width + self.o = AddReduceData(self.i.part_pts, n_inputs, + output_width, n_parts) def elaborate(self, platform): m = Module() + eps = self.i.part_pts + # collect part-bytes pbs = Signal(8, reset_less=True) tl = [] for i in range(8): pb = Signal(name="pb%d" % i, reset_less=True) - m.d.comb += pb.eq(self.part_pts.part_byte(i)) + m.d.comb += pb.eq(eps.part_byte(i)) tl.append(pb) m.d.comb += pbs.eq(Cat(*tl)) - # create (doubled) PartitionPoints (output is double input width) - expanded_part_pts = eps = PartitionPoints() - for i, v in self.part_pts.items(): - ep = Signal(name=f"expanded_part_pts_{i*2}", reset_less=True) - expanded_part_pts[i * 2] = ep - m.d.comb += ep.eq(v) - # local variables signs = [] for i in range(8): s = Signs() signs.append(s) setattr(m.submodules, "signs%d" % i, s) - m.d.comb += s.part_ops.eq(self.part_ops[i]) + m.d.comb += s.part_ops.eq(self.i.part_ops[i]) n_levels = len(self.register_levels)+1 m.submodules.part_8 = part_8 = Part(eps, 128, 8, n_levels, 8) @@ -1080,8 +1148,8 @@ class Mul8_16_32_64(Elaboratable): m.submodules.part_64 = part_64 = Part(eps, 128, 1, n_levels, 8) nat_l, nbt_l, nla_l, nlb_l = [], [], [], [] for mod in [part_8, part_16, part_32, part_64]: - m.d.comb += mod.a.eq(self.a) - m.d.comb += mod.b.eq(self.b) + m.d.comb += mod.a.eq(self.i.a) + m.d.comb += mod.b.eq(self.i.b) for i in range(len(signs)): m.d.comb += mod.a_signed[i].eq(signs[i].a_signed) m.d.comb += mod.b_signed[i].eq(signs[i].b_signed) @@ -1097,8 +1165,8 @@ class Mul8_16_32_64(Elaboratable): t = ProductTerms(8, 128, 8, a_index, 8) setattr(m.submodules, "terms_%d" % a_index, t) - m.d.comb += t.a.eq(self.a) - m.d.comb += t.b.eq(self.b) + m.d.comb += t.a.eq(self.i.a) + m.d.comb += t.b.eq(self.i.b) m.d.comb += t.pb_en.eq(pbs) for term in t.terms: @@ -1118,64 +1186,155 @@ class Mul8_16_32_64(Elaboratable): m.d.comb += mod.orin[i].eq(l[i]) terms.append(mod.orout) - add_reduce = AddReduce(terms, - 128, - self.register_levels, - expanded_part_pts, - self.part_ops) + # copy the intermediate terms to the output + for i, value in enumerate(terms): + m.d.comb += self.o.terms[i].eq(value) - out_part_ops = add_reduce.levels[-1].i.part_ops - out_part_pts = add_reduce.levels[-1].i.reg_partition_points + # copy reg part points and part ops to output + m.d.comb += self.o.part_pts.eq(eps) + m.d.comb += [self.o.part_ops[i].eq(self.i.part_ops[i]) + for i in range(len(self.i.part_ops))] + + return m + + +class Intermediates(Elaboratable): + """ Intermediate output modules + """ + + def __init__(self, output_width, n_parts, partition_points): + self.i = FinalReduceData(partition_points, output_width, n_parts) + self.o = IntermediateData(partition_points, output_width, n_parts) + + def elaborate(self, platform): + m = Module() + + out_part_ops = self.i.part_ops + out_part_pts = self.i.part_pts - m.submodules.add_reduce = add_reduce - m.d.comb += self._intermediate_output.eq(add_reduce.output) # create _output_64 m.submodules.io64 = io64 = IntermediateOut(64, 128, 1) - m.d.comb += io64.intermed.eq(self._intermediate_output) + m.d.comb += io64.intermed.eq(self.i.output) for i in range(8): m.d.comb += io64.part_ops[i].eq(out_part_ops[i]) + m.d.comb += self.o.outputs[3].eq(io64.output) # create _output_32 m.submodules.io32 = io32 = IntermediateOut(32, 128, 2) - m.d.comb += io32.intermed.eq(self._intermediate_output) + m.d.comb += io32.intermed.eq(self.i.output) for i in range(8): m.d.comb += io32.part_ops[i].eq(out_part_ops[i]) + m.d.comb += self.o.outputs[2].eq(io32.output) # create _output_16 m.submodules.io16 = io16 = IntermediateOut(16, 128, 4) - m.d.comb += io16.intermed.eq(self._intermediate_output) + m.d.comb += io16.intermed.eq(self.i.output) for i in range(8): m.d.comb += io16.part_ops[i].eq(out_part_ops[i]) + m.d.comb += self.o.outputs[1].eq(io16.output) # create _output_8 m.submodules.io8 = io8 = IntermediateOut(8, 128, 8) - m.d.comb += io8.intermed.eq(self._intermediate_output) + m.d.comb += io8.intermed.eq(self.i.output) for i in range(8): m.d.comb += io8.part_ops[i].eq(out_part_ops[i]) + m.d.comb += self.o.outputs[0].eq(io8.output) + + for i in range(8): + m.d.comb += self.o.part_ops[i].eq(out_part_ops[i]) + m.d.comb += self.o.part_pts.eq(out_part_pts) + m.d.comb += self.o.intermediate_output.eq(self.i.output) + + return m + + +class Mul8_16_32_64(Elaboratable): + """Signed/Unsigned 8/16/32/64-bit partitioned integer multiplier. + + Supports partitioning into any combination of 8, 16, 32, and 64-bit + partitions on naturally-aligned boundaries. Supports the operation being + set for each partition independently. + + :attribute part_pts: the input partition points. Has a partition point at + multiples of 8 in 0 < i < 64. Each partition point's associated + ``Value`` is a ``Signal``. Modification not supported, except for by + ``Signal.eq``. + :attribute part_ops: the operation for each byte. The operation for a + particular partition is selected by assigning the selected operation + code to each byte in the partition. The allowed operation codes are: + + :attribute OP_MUL_LOW: the LSB half of the product. Equivalent to + RISC-V's `mul` instruction. + :attribute OP_MUL_SIGNED_HIGH: the MSB half of the product where both + ``a`` and ``b`` are signed. Equivalent to RISC-V's `mulh` + instruction. + :attribute OP_MUL_SIGNED_UNSIGNED_HIGH: the MSB half of the product + where ``a`` is signed and ``b`` is unsigned. Equivalent to RISC-V's + `mulhsu` instruction. + :attribute OP_MUL_UNSIGNED_HIGH: the MSB half of the product where both + ``a`` and ``b`` are unsigned. Equivalent to RISC-V's `mulhu` + instruction. + """ + + def __init__(self, register_levels=()): + """ register_levels: specifies the points in the cascade at which + flip-flops are to be inserted. + """ + + # parameter(s) + self.register_levels = list(register_levels) + + # inputs + self.part_pts = PartitionPoints() + for i in range(8, 64, 8): + self.part_pts[i] = Signal(name=f"part_pts_{i}") + self.part_ops = [Signal(2, name=f"part_ops_{i}") for i in range(8)] + self.a = Signal(64) + self.b = Signal(64) + + # intermediates (needed for unit tests) + self.intermediate_output = Signal(128) + + # output + self.output = Signal(64) + + def elaborate(self, platform): + m = Module() + + part_pts = self.part_pts + + n_inputs = 64 + 4 + n_parts = 8 #len(self.part_pts) + t = AllTerms(n_inputs, 128, n_parts, self.register_levels, part_pts) + m.submodules.allterms = t + m.d.comb += t.i.a.eq(self.a) + m.d.comb += t.i.b.eq(self.b) + m.d.comb += t.i.part_pts.eq(part_pts) + for i in range(8): + m.d.comb += t.i.part_ops[i].eq(self.part_ops[i]) - m.submodules.p_8 = p_8 = Parts(8, eps, len(part_8.parts)) - m.submodules.p_16 = p_16 = Parts(8, eps, len(part_16.parts)) - m.submodules.p_32 = p_32 = Parts(8, eps, len(part_32.parts)) - m.submodules.p_64 = p_64 = Parts(8, eps, len(part_64.parts)) + terms = t.o.terms + + add_reduce = AddReduce(terms, + 128, + self.register_levels, + t.o.part_pts, + t.o.part_ops) + + out_part_ops = add_reduce.o.part_ops + out_part_pts = add_reduce.o.part_pts + + m.submodules.add_reduce = add_reduce - m.d.comb += p_8.epps.eq(out_part_pts) - m.d.comb += p_16.epps.eq(out_part_pts) - m.d.comb += p_32.epps.eq(out_part_pts) - m.d.comb += p_64.epps.eq(out_part_pts) + interm = Intermediates(128, 8, part_pts) + m.submodules.intermediates = interm + m.d.comb += interm.i.eq(add_reduce.o) # final output - m.submodules.finalout = finalout = FinalOut(64) - for i in range(len(part_8.parts)): - m.d.comb += finalout.d8[i].eq(p_8.parts[i]) - for i in range(len(part_16.parts)): - m.d.comb += finalout.d16[i].eq(p_16.parts[i]) - for i in range(len(part_32.parts)): - m.d.comb += finalout.d32[i].eq(p_32.parts[i]) - m.d.comb += finalout.i8.eq(io8.output) - m.d.comb += finalout.i16.eq(io16.output) - m.d.comb += finalout.i32.eq(io32.output) - m.d.comb += finalout.i64.eq(io64.output) + m.submodules.finalout = finalout = FinalOut(128, 8, part_pts) + m.d.comb += finalout.i.eq(interm.o) m.d.comb += self.output.eq(finalout.out) + m.d.comb += self.intermediate_output.eq(finalout.intermediate_output) return m @@ -1184,7 +1343,7 @@ if __name__ == "__main__": m = Mul8_16_32_64() main(m, ports=[m.a, m.b, - m._intermediate_output, + m.intermediate_output, m.output, *m.part_ops, *m.part_pts.values()])