""" Final stage of add reduce
"""
- def __init__(self, n_inputs, output_width, n_parts, register_levels,
- partition_points):
+ def __init__(self, n_inputs, output_width, n_parts, 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")
supported, except for by ``Signal.eq``.
"""
- def __init__(self, n_inputs, output_width, n_parts, register_levels,
- partition_points):
+ def __init__(self, n_inputs, output_width, n_parts, 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.n_inputs = n_inputs
self.output_width = output_width
self.i = AddReduceData(partition_points, n_inputs,
output_width, 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")
- max_level = AddReduceSingle.get_max_level(n_inputs)
- for level in self.register_levels:
- if level > max_level:
- raise ValueError(
- "not enough adder levels for specified register levels")
-
self.groups = AddReduceSingle.full_adder_groups(n_inputs)
n_terms = AddReduceSingle.calc_n_inputs(n_inputs, self.groups)
self.o = AddReduceData(partition_points, n_terms, output_width, n_parts)
"""creates reduction levels"""
mods = []
- next_levels = self.register_levels
partition_points = self.partition_points
part_ops = self.part_ops
n_parts = len(part_ops)
if len(groups) == 0:
break
next_level = AddReduceSingle(ilen, self.output_width, n_parts,
- next_levels, partition_points)
+ partition_points)
mods.append(next_level)
- next_levels = list(AddReduce.next_register_levels(next_levels))
partition_points = next_level.i.part_pts
inputs = next_level.o.terms
ilen = len(inputs)
part_ops = next_level.i.part_ops
next_level = FinalAdd(ilen, self.output_width, n_parts,
- next_levels, partition_points)
+ partition_points)
mods.append(next_level)
self.levels = mods
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:
+ if idx in self.register_levels:
m.d.sync += mcur.i.eq(i)
else:
m.d.comb += mcur.i.eq(i)
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}", 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))
+ m.d.comb += pb.eq(part_pts.part_byte(i))
tl.append(pb)
m.d.comb += pbs.eq(Cat(*tl))
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, reset_less=True)
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)
def elaborate(self, platform):
m = Module()
- pps = self.part_pts
- m.submodules.p_8 = p_8 = Parts(8, pps, 8)
- m.submodules.p_16 = p_16 = Parts(8, pps, 4)
- m.submodules.p_32 = p_32 = Parts(8, pps, 2)
- m.submodules.p_64 = p_64 = Parts(8, pps, 1)
+ 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
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)
+ 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])
rhs.intermediate_output, rhs.part_ops)
-class AllTermsData:
+class InputData:
- def __init__(self, partition_points):
+ def __init__(self):
self.a = Signal(64)
self.b = Signal(64)
- self.epps = partition_points.like()
+ 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 eq_from(self, epps, inputs, part_ops):
- return [self.epps.eq(epps)] + \
+ 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))]
def eq(self, rhs):
- return self.eq_from(rhs.epps, rhs.a, rhs.b, rhs.part_ops)
+ 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):
+ def __init__(self, n_inputs, output_width, n_parts, register_levels):
"""Create an ``AddReduce``.
:param inputs: input ``Signal``s to be summed.
pipeline registers.
:param partition_points: the input partition points.
"""
- self.i = AllTermsData(partition_points)
+ self.i = InputData()
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.epps, n_inputs,
+ self.o = AddReduceData(self.i.part_pts, n_inputs,
output_width, n_parts)
def elaborate(self, platform):
m = Module()
- eps = self.i.epps
+ eps = self.i.part_pts
# collect part-bytes
pbs = Signal(8, reset_less=True)
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)
+ self.i = InputData()
+ self.part_pts = self.i.part_pts
+ self.part_ops = self.i.part_ops
+ self.a = self.i.a
+ self.b = self.i.b
# intermediates (needed for unit tests)
self.intermediate_output = Signal(128)
def elaborate(self, platform):
m = Module()
- pps = self.part_pts
+ 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, pps)
+ t = AllTerms(n_inputs, 128, n_parts, self.register_levels)
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.epps.eq(pps)
+ 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.add_reduce = add_reduce
- interm = Intermediates(128, 8, pps)
+ 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(128, 8, pps)
+ 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)