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, partition_points):
+ self.expanded_part_points = partition_points
+ self.i = IntermediateData(partition_points, 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()
+
+ eps = self.expanded_part_points
+ m.submodules.p_8 = p_8 = Parts(8, eps, 8)
+ m.submodules.p_16 = p_16 = Parts(8, eps, 4)
+ m.submodules.p_32 = p_32 = Parts(8, eps, 2)
+ m.submodules.p_64 = p_64 = Parts(8, eps, 1)
+
+ out_part_pts = self.i.reg_partition_points
+
+ # 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.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)
+
+ 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
# 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
return m
+class IntermediateData:
+
+ def __init__(self, ppoints, output_width, n_parts):
+ self.part_ops = [Signal(2, name=f"part_ops_{i}", reset_less=True)
+ for i in range(n_parts)]
+ self.reg_partition_points = ppoints.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, reg_partition_points, outputs, intermediate_output,
+ part_ops):
+ return [self.reg_partition_points.eq(reg_partition_points)] + \
+ [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))]
+
+ def eq(self, rhs):
+ return self.eq_from(rhs.reg_partition_points, rhs.outputs,
+ rhs.intermediate_output, rhs.part_ops)
+
+
+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.reg_partition_points
+
+ # create _output_64
+ m.submodules.io64 = io64 = IntermediateOut(64, 128, 1)
+ 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.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.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.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.reg_partition_points.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.
expanded_part_pts[i * 2] = ep
m.d.comb += ep.eq(v)
+ n_inputs = 64 + 4
+ n_parts = 8 #len(self.part_pts)
+ t = AllTerms(8, n_inputs, 128, n_parts, self.register_levels,
+ eps)
+ m.submodules.allterms = t
+ m.d.comb += t.a.eq(self.a)
+ m.d.comb += t.b.eq(self.b)
+ m.d.comb += t.pbs.eq(pbs)
+ m.d.comb += t.epps.eq(eps)
+ for i in range(8):
+ m.d.comb += t.part_ops[i].eq(self.part_ops[i])
+
+ terms = t.o.inputs
+
+ add_reduce = AddReduce(terms,
+ 128,
+ self.register_levels,
+ t.o.reg_partition_points,
+ t.o.part_ops)
+
+ out_part_ops = add_reduce.o.part_ops
+ out_part_pts = add_reduce.o.reg_partition_points
+
+ m.submodules.add_reduce = add_reduce
+ m.d.comb += self.intermediate_output.eq(add_reduce.o.output)
+
+ interm = Intermediates(128, 8, expanded_part_pts)
+ m.submodules.intermediates = interm
+ m.d.comb += interm.i.eq(add_reduce.o)
+
+ # final output
+ m.submodules.finalout = finalout = FinalOut(128, 8, expanded_part_pts)
+ m.d.comb += finalout.i.eq(interm.o)
+ m.d.comb += self.output.eq(finalout.out)
+
+ return m
+
+
+class AllTerms(Elaboratable):
+ """Set of terms to be added together
+ """
+
+ def __init__(self, pbwid, 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.epps = partition_points.like()
+ self.register_levels = register_levels
+ self.pbwid = pbwid
+ self.n_inputs = n_inputs
+ self.n_parts = n_parts
+ self.output_width = output_width
+ self.o = AddReduceData(self.epps, n_inputs,
+ output_width, n_parts)
+
+ self.a = Signal(64)
+ self.b = Signal(64)
+
+ self.pbs = Signal(pbwid, reset_less=True)
+ self.part_ops = [Signal(2, name=f"part_ops_{i}") for i in range(8)]
+
+ def elaborate(self, platform):
+ m = Module()
+
+ pbs = self.pbs
+ eps = self.epps
+
# local variables
signs = []
for i in range(8):
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)
-
- out_part_ops = add_reduce.o.part_ops
- out_part_pts = add_reduce.o.reg_partition_points
-
- m.submodules.add_reduce = add_reduce
- m.d.comb += self.intermediate_output.eq(add_reduce.o.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.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.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.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.part_ops[i].eq(out_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))
-
- 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)
+ # copy the intermediate terms to the output
+ for i, value in enumerate(terms):
+ m.d.comb += self.o.inputs[i].eq(value)
- # 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.d.comb += self.output.eq(finalout.out)
+ # copy reg part points and part ops to output
+ m.d.comb += self.o.reg_partition_points.eq(eps)
+ m.d.comb += [self.o.part_ops[i].eq(self.part_ops[i])
+ for i in range(len(self.part_ops))]
return m