X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fieee754%2Fpart_mul_add%2Fmultiply.py;h=e0fe069d5b44139b12373fbfec72849d1396d813;hb=e297dd6f682d8d149cad6ff45a56fa55da694b3c;hp=de40a89dd7e7cb6a79986f9c74660f8a8c141629;hpb=58ed1a5b82eb23003e3a787d0da59cf81aae253f;p=ieee754fpu.git diff --git a/src/ieee754/part_mul_add/multiply.py b/src/ieee754/part_mul_add/multiply.py index de40a89d..e0fe069d 100644 --- a/src/ieee754/part_mul_add/multiply.py +++ b/src/ieee754/part_mul_add/multiply.py @@ -302,20 +302,86 @@ FULL_ADDER_INPUT_COUNT = 3 class AddReduceData: - def __init__(self, ppoints, output_width, n_parts) + def __init__(self, ppoints, n_inputs, output_width, n_parts): self.part_ops = [Signal(2, name=f"part_ops_{i}") for i in range(n_parts)] self.inputs = [Signal(output_width, name=f"inputs[{i}]") - for i in range(len(self.inputs))] - self.reg_partition_points = partition_points.like() + for i in range(n_inputs)] + self.reg_partition_points = ppoints.like() - def eq(self, rhs): - return [self.reg_partition_points.eq(rhs.reg_partition_points)] + \ - [self.inputs[i].eq(rhs.inputs[i]) + def eq_from(self, reg_partition_points, inputs, part_ops): + return [self.reg_partition_points.eq(reg_partition_points)] + \ + [self.inputs[i].eq(inputs[i]) for i in range(len(self.inputs))] + \ - [self.part_ops[i].eq(rhs.part_ops[i]) + [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.reg_partition_points, rhs.inputs, rhs.part_ops) + + +class FinalReduceData: + + def __init__(self, ppoints, output_width, n_parts): + self.part_ops = [Signal(2, name=f"part_ops_{i}") + for i in range(n_parts)] + self.output = Signal(output_width) + self.reg_partition_points = ppoints.like() + + def eq_from(self, reg_partition_points, output, part_ops): + return [self.reg_partition_points.eq(reg_partition_points)] + \ + [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.reg_partition_points, rhs.output, rhs.part_ops) + + +class FinalAdd(Elaboratable): + """ Final stage of add reduce + """ + + def __init__(self, n_inputs, output_width, n_parts, register_levels, + 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.register_levels = list(register_levels) + 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") + + def elaborate(self, platform): + """Elaborate this module.""" + m = Module() + + output_width = self.output_width + output = Signal(output_width) + if self.n_inputs == 0: + # use 0 as the default output value + m.d.comb += output.eq(0) + elif self.n_inputs == 1: + # handle single input + m.d.comb += output.eq(self.i.inputs[0]) + else: + # base case for adding 2 inputs + assert self.n_inputs == 2 + adder = PartitionedAdder(output_width, self.i.reg_partition_points) + m.submodules.final_adder = adder + m.d.comb += adder.a.eq(self.i.inputs[0]) + m.d.comb += adder.b.eq(self.i.inputs[1]) + m.d.comb += output.eq(adder.output) + + # create output + m.d.comb += self.o.eq_from(self.i.reg_partition_points, output, + self.i.part_ops) + + return m + class AddReduceSingle(Elaboratable): """Add list of numbers together. @@ -329,8 +395,8 @@ class AddReduceSingle(Elaboratable): supported, except for by ``Signal.eq``. """ - def __init__(self, inputs, output_width, register_levels, partition_points, - part_ops): + def __init__(self, n_inputs, output_width, n_parts, register_levels, + partition_points): """Create an ``AddReduce``. :param inputs: input ``Signal``s to be summed. @@ -339,21 +405,17 @@ class AddReduceSingle(Elaboratable): pipeline registers. :param partition_points: the input partition points. """ - 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.inputs = list(inputs) - self._resized_inputs = [ - Signal(output_width, name=f"resized_inputs[{i}]") - for i in range(len(self.inputs))] + self.n_inputs = n_inputs + self.n_parts = n_parts + self.output_width = output_width + self.i = AddReduceData(partition_points, n_inputs, + output_width, n_parts) 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._reg_partition_points = self.partition_points.like() - max_level = AddReduceSingle.get_max_level(len(self.inputs)) + max_level = AddReduceSingle.get_max_level(n_inputs) for level in self.register_levels: if level > max_level: raise ValueError( @@ -363,11 +425,14 @@ class AddReduceSingle(Elaboratable): # 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(len(self.inputs)) + self.groups = AddReduceSingle.full_adder_groups(n_inputs) self._intermediate_terms = [] if len(self.groups) != 0: self.create_next_terms() + self.o = AddReduceData(partition_points, len(self._intermediate_terms), + output_width, n_parts) + @staticmethod def get_max_level(input_count): """Get the maximum level. @@ -395,97 +460,62 @@ class AddReduceSingle(Elaboratable): """Elaborate this module.""" m = Module() - # resize inputs to correct bit-width and optionally add in - # pipeline registers - resized_input_assignments = [self._resized_inputs[i].eq(self.inputs[i]) - for i in range(len(self.inputs))] - copy_part_ops = [self.out_part_ops[i].eq(self.part_ops[i]) - for i in range(len(self.part_ops))] - if 0 in self.register_levels: - m.d.sync += copy_part_ops - m.d.sync += resized_input_assignments - m.d.sync += self._reg_partition_points.eq(self.partition_points) - else: - m.d.comb += copy_part_ops - m.d.comb += resized_input_assignments - m.d.comb += self._reg_partition_points.eq(self.partition_points) - - for (value, term) in self._intermediate_terms: - m.d.comb += term.eq(value) - - # if there are no full adders to create, then we handle the base cases - # and return, otherwise we go on to the recursive case - if len(self.groups) == 0: - if len(self.inputs) == 0: - # use 0 as the default output value - m.d.comb += self.output.eq(0) - elif len(self.inputs) == 1: - # handle single input - m.d.comb += self.output.eq(self._resized_inputs[0]) - else: - # base case for adding 2 inputs - assert len(self.inputs) == 2 - adder = PartitionedAdder(len(self.output), - self._reg_partition_points) - 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) - return m - - mask = self._reg_partition_points.as_mask(len(self.output)) + # copy the intermediate terms to the output + for i, value in enumerate(self._intermediate_terms): + m.d.comb += self.o.inputs[i].eq(value) + + # copy reg part points and part ops to output + m.d.comb += self.o.reg_partition_points.eq(self.i.reg_partition_points) + 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) + 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.in0.eq(self.i.inputs[iidx]) + m.d.comb += adder_i.in1.eq(self.i.inputs[iidx + 1]) + m.d.comb += adder_i.in2.eq(self.i.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( - len(self.output), - name=f"intermediate_terms[{len(intermediate_terms)}]") - _intermediate_terms.append((value, intermediate_term)) - intermediate_terms.append(intermediate_term) + _intermediate_terms.append(value) # store mask in intermediary (simplifies graph) - self.part_mask = Signal(len(self.output), reset_less=True) + self.part_mask = Signal(self.output_width, reset_less=True) # 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(len(self.output)) + adder_i = MaskedFullAdder(self.output_width) self.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) # handle the remaining inputs. - if len(self.inputs) % FULL_ADDER_INPUT_COUNT == 1: - add_intermediate_term(self._resized_inputs[-1]) - elif len(self.inputs) % FULL_ADDER_INPUT_COUNT == 2: + if self.n_inputs % FULL_ADDER_INPUT_COUNT == 1: + add_intermediate_term(self.i.inputs[-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]) + add_intermediate_term(self.i.inputs[-2]) + add_intermediate_term(self.i.inputs[-1]) else: - assert len(self.inputs) % FULL_ADDER_INPUT_COUNT == 0 + assert self.n_inputs % FULL_ADDER_INPUT_COUNT == 0 - self.intermediate_terms = intermediate_terms self._intermediate_terms = _intermediate_terms @@ -539,18 +569,27 @@ 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: - next_level = AddReduceSingle(inputs, self.output_width, next_levels, - partition_points, part_ops) + next_level = AddReduceSingle(ilen, self.output_width, n_parts, + next_levels, partition_points) mods.append(next_level) - if len(next_level.groups) == 0: - break next_levels = list(AddReduce.next_register_levels(next_levels)) - partition_points = next_level._reg_partition_points - inputs = next_level.intermediate_terms - part_ops = next_level.out_part_ops + partition_points = next_level.i.reg_partition_points + inputs = next_level.o.inputs + 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) self.levels = mods @@ -561,10 +600,26 @@ class AddReduce(Elaboratable): for i, next_level in enumerate(self.levels): setattr(m.submodules, "next_level%d" % i, next_level) + partition_points = self.partition_points + inputs = self.inputs + part_ops = self.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 += mcur.i.eq(i) + else: + 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.out_part_ops[i]) - for i in range(len(self.part_ops))] + m.d.comb += self.output.eq(i.output) + copy_part_ops = [self.out_part_ops[idx].eq(i.part_ops[idx]) + for idx in range(len(self.part_ops))] m.d.comb += copy_part_ops return m @@ -1092,8 +1147,8 @@ class Mul8_16_32_64(Elaboratable): expanded_part_pts, self.part_ops) - out_part_ops = add_reduce.levels[-1].out_part_ops - out_part_pts = add_reduce.levels[-1]._reg_partition_points + out_part_ops = add_reduce.out_part_ops + out_part_pts = add_reduce.levels[-1].o.reg_partition_points m.submodules.add_reduce = add_reduce m.d.comb += self._intermediate_output.eq(add_reduce.output)