X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fieee754%2Fpart_mul_add%2Fmultiply.py;h=64994ad36c9feff26d84cd6b52acd05193456fa9;hb=a750daabd40ac4fe2054bb88b632ab720d985c4e;hp=32e817a1b8f99de9bdda87eede9365cee8843deb;hpb=2a8f9f77461367ee6852dca2286f7f7a83c24f9c;p=ieee754fpu.git diff --git a/src/ieee754/part_mul_add/multiply.py b/src/ieee754/part_mul_add/multiply.py index 32e817a1..64994ad3 100644 --- a/src/ieee754/part_mul_add/multiply.py +++ b/src/ieee754/part_mul_add/multiply.py @@ -103,6 +103,12 @@ class PartitionPoints(dict): return False return True + def part_byte(self, index, mfactor=1): # mfactor used for "expanding" + if index == -1 or index == 7: + return C(True, 1) + assert index >= 0 and index < 8 + return self[(index * 8 + 8)*mfactor] + class FullAdder(Elaboratable): """Full Adder. @@ -316,7 +322,7 @@ class AddReduceSingle(Elaboratable): :param partition_points: the input partition points. """ self.part_ops = part_ops - self._part_ops = [Signal(2, name=f"part_ops_{i}") + 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 = [ @@ -335,6 +341,10 @@ 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(len(self.inputs)) self._intermediate_terms = [] if len(self.groups) != 0: @@ -371,7 +381,7 @@ class AddReduceSingle(Elaboratable): # pipeline registers resized_input_assignments = [self._resized_inputs[i].eq(self.inputs[i]) for i in range(len(self.inputs))] - copy_part_ops = [self._part_ops[i].eq(self.part_ops[i]) + 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 @@ -485,6 +495,8 @@ 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) self.output_width = output_width self.register_levels = register_levels @@ -492,6 +504,10 @@ class AddReduce(Elaboratable): self.create_levels() + @staticmethod + def get_max_level(input_count): + return AddReduceSingle.get_max_level(input_count) + @staticmethod def next_register_levels(register_levels): """``Iterable`` of ``register_levels`` for next recursive level.""" @@ -506,15 +522,17 @@ class AddReduce(Elaboratable): next_levels = self.register_levels partition_points = self.partition_points inputs = self.inputs + part_ops = self.part_ops while True: next_level = AddReduceSingle(inputs, self.output_width, next_levels, - partition_points, self.part_ops) + partition_points, part_ops) 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 self.levels = mods @@ -527,6 +545,9 @@ class AddReduce(Elaboratable): # 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 += copy_part_ops return m @@ -592,8 +613,8 @@ class ProductTerm(Elaboratable): bsb = Signal(self.width, reset_less=True) a_index, b_index = self.a_index, self.b_index pwidth = self.pwidth - m.d.comb += bsa.eq(self.a.bit_select(a_index * pwidth, pwidth)) - m.d.comb += bsb.eq(self.b.bit_select(b_index * pwidth, pwidth)) + m.d.comb += bsa.eq(self.a.part(a_index * pwidth, pwidth)) + m.d.comb += bsb.eq(self.b.part(b_index * pwidth, pwidth)) m.d.comb += self.ti.eq(bsa * bsb) m.d.comb += self.term.eq(get_term(self.ti, self.shift, self.enabled)) """ @@ -607,8 +628,8 @@ class ProductTerm(Elaboratable): asel = Signal(width, reset_less=True) bsel = Signal(width, reset_less=True) a_index, b_index = self.a_index, self.b_index - m.d.comb += asel.eq(self.a.bit_select(a_index * pwidth, pwidth)) - m.d.comb += bsel.eq(self.b.bit_select(b_index * pwidth, pwidth)) + m.d.comb += asel.eq(self.a.part(a_index * pwidth, pwidth)) + m.d.comb += bsel.eq(self.b.part(b_index * pwidth, pwidth)) m.d.comb += bsa.eq(get_term(asel, self.shift, self.enabled)) m.d.comb += bsb.eq(get_term(bsel, self.shift, self.enabled)) m.d.comb += self.ti.eq(bsa * bsb) @@ -687,6 +708,45 @@ class LSBNegTerm(Elaboratable): return m +class Parts(Elaboratable): + + def __init__(self, pbwid, epps, n_parts): + self.pbwid = pbwid + # inputs + self.epps = PartitionPoints.like(epps, name="epps") # expanded points + # outputs + self.parts = [Signal(name=f"part_{i}") for i in range(n_parts)] + + def elaborate(self, platform): + m = Module() + + epps, parts = self.epps, 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 + tl.append(pb) + m.d.comb += pbs.eq(Cat(*tl)) + + # negated-temporary copy of partition bits + npbs = Signal.like(pbs, reset_less=True) + m.d.comb += npbs.eq(~pbs) + byte_count = 8 // len(parts) + for i in range(len(parts)): + pbl = [] + pbl.append(npbs[i * byte_count - 1]) + for j in range(i * byte_count, (i + 1) * byte_count - 1): + pbl.append(pbs[j]) + pbl.append(npbs[(i + 1) * byte_count - 1]) + value = Signal(len(pbl), name="value_%d" % i, reset_less=True) + m.d.comb += value.eq(Cat(*pbl)) + m.d.comb += parts[i].eq(~(value).bool()) + + return m + + class Part(Elaboratable): """ a key class which, depending on the partitioning, will determine what action to take when parts of the output are signed or unsigned. @@ -760,7 +820,7 @@ class Part(Elaboratable): pa = LSBNegTerm(bit_wid) setattr(m.submodules, "lnt_%d_a_%d" % (bit_wid, i), pa) m.d.comb += pa.part.eq(parts[i]) - m.d.comb += pa.op.eq(self.a.bit_select(bit_wid * i, bit_wid)) + m.d.comb += pa.op.eq(self.a.part(bit_wid * i, bit_wid)) m.d.comb += pa.signed.eq(self.b_signed[i * byte_width]) # yes b m.d.comb += pa.msb.eq(self.b[(i + 1) * bit_wid - 1]) # really, b nat.append(pa.nt) @@ -770,7 +830,7 @@ class Part(Elaboratable): pb = LSBNegTerm(bit_wid) setattr(m.submodules, "lnt_%d_b_%d" % (bit_wid, i), pb) m.d.comb += pb.part.eq(parts[i]) - m.d.comb += pb.op.eq(self.b.bit_select(bit_wid * i, bit_wid)) + m.d.comb += pb.op.eq(self.b.part(bit_wid * i, bit_wid)) m.d.comb += pb.signed.eq(self.a_signed[i * byte_width]) # yes a m.d.comb += pb.msb.eq(self.a[(i + 1) * bit_wid - 1]) # really, a nbt.append(pb.nt) @@ -793,7 +853,7 @@ class IntermediateOut(Elaboratable): def __init__(self, width, out_wid, n_parts): self.width = width self.n_parts = n_parts - self.delayed_part_ops = [Signal(2, name="dpop%d" % i, reset_less=True) + self.part_ops = [Signal(2, name="dpop%d" % i, reset_less=True) for i in range(8)] self.intermed = Signal(out_wid, reset_less=True) self.output = Signal(out_wid//2, reset_less=True) @@ -807,9 +867,9 @@ class IntermediateOut(Elaboratable): for i in range(self.n_parts): op = Signal(w, reset_less=True, name="op%d_%d" % (w, i)) m.d.comb += op.eq( - Mux(self.delayed_part_ops[sel * i] == OP_MUL_LOW, - self.intermed.bit_select(i * w*2, w), - self.intermed.bit_select(i * w*2 + w, w))) + Mux(self.part_ops[sel * i] == OP_MUL_LOW, + self.intermed.part(i * w*2, w), + self.intermed.part(i * w*2 + w, w))) ol.append(op) m.d.comb += self.output.eq(Cat(*ol)) @@ -849,10 +909,10 @@ class FinalOut(Elaboratable): 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.bit_select(i * 8, 8), - self.i16.bit_select(i * 8, 8)), - Mux(self.d32[i // 4], self.i32.bit_select(i * 8, 8), - self.i64.bit_select(i * 8, 8)))) + 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)))) ol.append(op) m.d.comb += self.out.eq(Cat(*ol)) return m @@ -951,12 +1011,6 @@ class Mul8_16_32_64(Elaboratable): # output self.output = Signal(64) - def _part_byte(self, index): - if index == -1 or index == 7: - return C(True, 1) - assert index >= 0 and index < 8 - return self.part_pts[index * 8 + 8] - def elaborate(self, platform): m = Module() @@ -965,10 +1019,17 @@ class Mul8_16_32_64(Elaboratable): tl = [] for i in range(8): pb = Signal(name="pb%d" % i, reset_less=True) - m.d.comb += pb.eq(self._part_byte(i)) + m.d.comb += pb.eq(self.part_pts.part_byte(i)) tl.append(pb) m.d.comb += pbs.eq(Cat(*tl)) + # create (doubled) PartitionPoints (output is double input width) + expanded_part_pts = 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): @@ -977,15 +1038,6 @@ class Mul8_16_32_64(Elaboratable): setattr(m.submodules, "signs%d" % i, s) m.d.comb += s.part_ops.eq(self.part_ops[i]) - delayed_part_ops = [ - [Signal(2, name=f"_delayed_part_ops_{delay}_{i}") - for i in range(8)] - for delay in range(1 + len(self.register_levels))] - for i in range(len(self.part_ops)): - m.d.comb += delayed_part_ops[0][i].eq(self.part_ops[i]) - m.d.sync += [delayed_part_ops[j + 1][i].eq(delayed_part_ops[j][i]) - for j in range(len(self.register_levels))] - n_levels = len(self.register_levels)+1 m.submodules.part_8 = part_8 = Part(128, 8, n_levels, 8) m.submodules.part_16 = part_16 = Part(128, 4, n_levels, 8) @@ -1031,43 +1083,39 @@ class Mul8_16_32_64(Elaboratable): m.d.comb += mod.orin[i].eq(l[i]) terms.append(mod.orout) - expanded_part_pts = PartitionPoints() - for i, v in self.part_pts.items(): - signal = Signal(name=f"expanded_part_pts_{i*2}", reset_less=True) - expanded_part_pts[i * 2] = signal - m.d.comb += signal.eq(v) - add_reduce = AddReduce(terms, 128, self.register_levels, expanded_part_pts, self.part_ops) + out_part_ops = add_reduce.levels[-1].out_part_ops + 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) for i in range(8): - m.d.comb += io64.delayed_part_ops[i].eq(delayed_part_ops[-1][i]) + m.d.comb += io64.part_ops[i].eq(out_part_ops[i]) # create _output_32 m.submodules.io32 = io32 = IntermediateOut(32, 128, 2) m.d.comb += io32.intermed.eq(self._intermediate_output) for i in range(8): - m.d.comb += io32.delayed_part_ops[i].eq(delayed_part_ops[-1][i]) + m.d.comb += io32.part_ops[i].eq(out_part_ops[i]) # create _output_16 m.submodules.io16 = io16 = IntermediateOut(16, 128, 4) m.d.comb += io16.intermed.eq(self._intermediate_output) for i in range(8): - m.d.comb += io16.delayed_part_ops[i].eq(delayed_part_ops[-1][i]) + m.d.comb += io16.part_ops[i].eq(out_part_ops[i]) # create _output_8 m.submodules.io8 = io8 = IntermediateOut(8, 128, 8) m.d.comb += io8.intermed.eq(self._intermediate_output) for i in range(8): - m.d.comb += io8.delayed_part_ops[i].eq(delayed_part_ops[-1][i]) + m.d.comb += io8.part_ops[i].eq(out_part_ops[i]) # final output m.submodules.finalout = finalout = FinalOut(64)