The remainder is the left-hand-side of the comparison minus the
right-hand-side of the comparison in the above formulas.
"""
-from nmigen import (Elaboratable, Module, Signal, Const, Mux)
+from nmigen import (Elaboratable, Module, Signal, Const, Mux, Cat)
import enum
-# TODO, move to new (suitable) location
-#from ieee754.fpcommon.getop import FPPipeContext
-
class DivPipeCoreConfig:
""" Configuration for core of the div/rem/sqrt/rsqrt pipeline.
+ f"{self.fract_width}, {self.log2_radix})"
@property
- def num_calculate_stages(self):
+ def n_stages(self):
""" Get the number of ``DivPipeCoreCalculateStage`` needed. """
return (self.bit_width + self.log2_radix - 1) // self.log2_radix
-class DivPipeCoreOperation(enum.IntEnum):
+class DivPipeCoreOperation(enum.Enum):
""" Operation for ``DivPipeCore``.
:attribute UDivRem: unsigned divide/remainder.
SqrtRem = 1
RSqrtRem = 2
+ def __int__(self):
+ """ Convert to int. """
+ return self.value
+
@classmethod
def create_signal(cls, *, src_loc_at=0, **kwargs):
""" Create a signal that can contain a ``DivPipeCoreOperation``. """
- return Signal(min=int(min(cls)),
- max=int(max(cls)),
+ return Signal(min=min(map(int, cls)),
+ max=max(map(int, cls)) + 2,
src_loc_at=(src_loc_at + 1),
- decoder=cls,
+ decoder=lambda v: str(cls(v)),
**kwargs)
-# TODO: move to suitable location
-class DivPipeBaseData:
- """ input data base type for ``DivPipe``.
- """
-
- def __init__(self, width, pspec):
- """ Create a ``DivPipeBaseData`` instance. """
- self.out_do_z = Signal(reset_less=True)
- self.oz = Signal(width, reset_less=True)
-
- self.ctx = FPPipeContext(width, pspec) # context: muxid, operator etc.
- self.muxid = self.ctx.muxid # annoying. complicated.
-
- def __iter__(self):
- """ Get member signals. """
- yield self.out_do_z
- yield self.oz
- yield from self.ctx
-
- def eq(self, rhs):
- """ Assign member signals. """
- return [self.out_do_z.eq(i.out_do_z), self.oz.eq(i.oz),
- self.ctx.eq(i.ctx)]
-
+DP = DivPipeCoreOperation
class DivPipeCoreInputData:
""" input data type for ``DivPipeCore``.
:attribute operation: the ``DivPipeCoreOperation`` to be computed.
"""
- def __init__(self, core_config):
+ def __init__(self, core_config, reset_less=True):
""" Create a ``DivPipeCoreInputData`` instance. """
self.core_config = core_config
self.dividend = Signal(core_config.bit_width + core_config.fract_width,
- reset_less=True)
- self.divisor_radicand = Signal(core_config.bit_width, reset_less=True)
-
- # FIXME: this goes into (is replaced by) self.ctx.op
- self.operation = DivPipeCoreOperation.create_signal(reset_less=True)
+ reset_less=reset_less)
+ self.divisor_radicand = Signal(core_config.bit_width,
+ reset_less=reset_less)
+ self.operation = DP.create_signal(reset_less=reset_less)
def __iter__(self):
""" Get member signals. """
yield self.dividend
yield self.divisor_radicand
- yield self.operation # FIXME: delete. already covered by self.ctx
- return
- yield self.z
- yield self.out_do_z
- yield self.oz
- yield from self.ctx
+ yield self.operation
def eq(self, rhs):
""" Assign member signals. """
return [self.dividend.eq(rhs.dividend),
self.divisor_radicand.eq(rhs.divisor_radicand),
- self.operation.eq(rhs.operation)] # FIXME: delete.
-
-
-# TODO: move to suitable location
-class DivPipeInputData(DivPipeCoreInputData, DivPipeBaseData):
- """ input data type for ``DivPipe``.
- """
-
- def __init__(self, core_config):
- """ Create a ``DivPipeInputData`` instance. """
- DivPipeCoreInputData.__init__(self, core_config)
- DivPipeBaseData.__init__(self, width, pspec) # XXX TODO args
- self.out_do_z = Signal(reset_less=True)
- self.oz = Signal(width, reset_less=True)
-
- self.ctx = FPPipeContext(width, pspec) # context: muxid, operator etc.
- self.muxid = self.ctx.muxid # annoying. complicated.
-
- def __iter__(self):
- """ Get member signals. """
- yield from DivPipeCoreInputData.__iter__(self)
- yield from DivPipeBaseData.__iter__(self)
-
- def eq(self, rhs):
- """ Assign member signals. """
- return DivPipeBaseData.eq(self, rhs) + \
- DivPipeCoreInputData.eq(self, rhs)
-
+ self.operation.eq(rhs.operation),
+ ]
class DivPipeCoreInterstageData:
``core_config.fract_width * 3`` bits.
"""
- def __init__(self, core_config):
+ def __init__(self, core_config, reset_less=True):
""" Create a ``DivPipeCoreInterstageData`` instance. """
self.core_config = core_config
- self.divisor_radicand = Signal(core_config.bit_width, reset_less=True)
- # XXX FIXME: delete. already covered by self.ctx.op
- self.operation = DivPipeCoreOperation.create_signal(reset_less=True)
- self.quotient_root = Signal(core_config.bit_width, reset_less=True)
+ self.divisor_radicand = Signal(core_config.bit_width,
+ reset_less=reset_less)
+ self.operation = DP.create_signal(reset_less=reset_less)
+ self.quotient_root = Signal(core_config.bit_width,
+ reset_less=reset_less)
self.root_times_radicand = Signal(core_config.bit_width * 2,
- reset_less=True)
- self.compare_lhs = Signal(core_config.bit_width * 3, reset_less=True)
- self.compare_rhs = Signal(core_config.bit_width * 3, reset_less=True)
+ reset_less=reset_less)
+ self.compare_lhs = Signal(core_config.bit_width * 3,
+ reset_less=reset_less)
+ self.compare_rhs = Signal(core_config.bit_width * 3,
+ reset_less=reset_less)
def __iter__(self):
""" Get member signals. """
yield self.divisor_radicand
- yield self.operation # XXX FIXME: delete. already in self.ctx.op
+ yield self.operation
yield self.quotient_root
yield self.root_times_radicand
yield self.compare_lhs
def eq(self, rhs):
""" Assign member signals. """
return [self.divisor_radicand.eq(rhs.divisor_radicand),
- self.operation.eq(rhs.operation), # FIXME: delete.
+ self.operation.eq(rhs.operation),
self.quotient_root.eq(rhs.quotient_root),
self.root_times_radicand.eq(rhs.root_times_radicand),
self.compare_lhs.eq(rhs.compare_lhs),
self.compare_rhs.eq(rhs.compare_rhs)]
-# TODO: move to suitable location
-class DivPipeInterstageData(DivPipeCoreInterstageData, DivPipeBaseData):
- """ interstage data type for ``DivPipe``.
-
- :attribute core_config: ``DivPipeCoreConfig`` instance describing the
- configuration to be used.
- """
-
- def __init__(self, core_config):
- """ Create a ``DivPipeCoreInterstageData`` instance. """
- DivPipeCoreInterstageData.__init__(self, core_config)
- DivPipeBaseData.__init__(self, width, pspec) # XXX TODO args
-
- def __iter__(self):
- """ Get member signals. """
- yield from DivPipeInterstageData.__iter__(self)
- yield from DivPipeBaseData.__iter__(self)
-
- def eq(self, rhs):
- """ Assign member signals. """
- return DivPipeBaseData.eq(self, rhs) + \
- DivPipeCoreInterstageData.eq(self, rhs)
-
-
class DivPipeCoreOutputData:
""" output data type for ``DivPipeCore``.
fract-width of ``core_config.fract_width * 3`` bits.
"""
- def __init__(self, core_config):
+ def __init__(self, core_config, reset_less=True):
""" Create a ``DivPipeCoreOutputData`` instance. """
self.core_config = core_config
- self.quotient_root = Signal(core_config.bit_width, reset_less=True)
- self.remainder = Signal(core_config.bit_width * 3, reset_less=True)
+ self.quotient_root = Signal(core_config.bit_width,
+ reset_less=reset_less)
+ self.remainder = Signal(core_config.bit_width * 3,
+ reset_less=reset_less)
def __iter__(self):
""" Get member signals. """
self.remainder.eq(rhs.remainder)]
-# TODO: move to suitable location
-class DivPipeOutputData(DivPipeCoreOutputData, DivPipeBaseData):
- """ interstage data type for ``DivPipe``.
-
- :attribute core_config: ``DivPipeCoreConfig`` instance describing the
- configuration to be used.
- """
-
- def __init__(self, core_config):
- """ Create a ``DivPipeCoreOutputData`` instance. """
- DivPipeCoreOutputData.__init__(self, core_config)
- DivPipeBaseData.__init__(self, width, pspec) # XXX TODO args
-
- def __iter__(self):
- """ Get member signals. """
- yield from DivPipeOutputData.__iter__(self)
- yield from DivPipeBaseData.__iter__(self)
-
- def eq(self, rhs):
- """ Assign member signals. """
- return DivPipeBaseData.eq(self, rhs) + \
- DivPipeCoreOutputData.eq(self, rhs)
-
-
-class DivPipeBaseStage:
- """ Base Mix-in for DivPipe*Stage """
-
- def _elaborate(self, m, platform):
- m.d.comb += self.o.oz.eq(self.i.oz)
- m.d.comb += self.o.out_do_z.eq(self.i.out_do_z)
- m.d.comb += self.o.ctx.eq(self.i.ctx)
-
-
class DivPipeCoreSetupStage(Elaboratable):
""" Setup Stage of the core of the div/rem/sqrt/rsqrt pipeline. """
m.d.comb += self.o.quotient_root.eq(0)
m.d.comb += self.o.root_times_radicand.eq(0)
- with m.If(self.i.operation == DivPipeCoreOperation.UDivRem):
+ with m.If(self.i.operation == int(DP.UDivRem)):
m.d.comb += self.o.compare_lhs.eq(self.i.dividend
<< self.core_config.fract_width)
- with m.Elif(self.i.operation == DivPipeCoreOperation.SqrtRem):
+ with m.Elif(self.i.operation == int(DP.SqrtRem)):
m.d.comb += self.o.compare_lhs.eq(
self.i.divisor_radicand << (self.core_config.fract_width * 2))
with m.Else(): # DivPipeCoreOperation.RSqrtRem
return m
- # XXX in DivPipeSetupStage
- DivPipeBaseStage._elaborate(self, m, platform)
+
+class Trial(Elaboratable):
+ def __init__(self, core_config, trial_bits, current_shift, log2_radix):
+ self.core_config = core_config
+ self.trial_bits = trial_bits
+ self.current_shift = current_shift
+ self.log2_radix = log2_radix
+ bw = core_config.bit_width
+ self.divisor_radicand = Signal(bw, reset_less=True)
+ self.quotient_root = Signal(bw, reset_less=True)
+ self.root_times_radicand = Signal(bw * 2, reset_less=True)
+ self.compare_rhs = Signal(bw * 3, reset_less=True)
+ self.trial_compare_rhs = Signal(bw * 3, reset_less=True)
+ self.operation = DP.create_signal(reset_less=True)
+
+ def elaborate(self, platform):
+
+ m = Module()
+
+ dr = self.divisor_radicand
+ qr = self.quotient_root
+ rr = self.root_times_radicand
+
+ trial_bits_sig = Const(self.trial_bits, self.log2_radix)
+ trial_bits_sqrd_sig = Const(self.trial_bits * self.trial_bits,
+ self.log2_radix * 2)
+
+ tblen = self.core_config.bit_width+self.log2_radix
+ tblen2 = self.core_config.bit_width+self.log2_radix*2
+ dr_times_trial_bits_sqrd = Signal(tblen2, reset_less=True)
+ m.d.comb += dr_times_trial_bits_sqrd.eq(dr * trial_bits_sqrd_sig)
+
+ # UDivRem
+ with m.If(self.operation == int(DP.UDivRem)):
+ dr_times_trial_bits = Signal(tblen, reset_less=True)
+ m.d.comb += dr_times_trial_bits.eq(dr * trial_bits_sig)
+ div_rhs = self.compare_rhs
+
+ div_term1 = dr_times_trial_bits
+ div_term1_shift = self.core_config.fract_width
+ div_term1_shift += self.current_shift
+ div_rhs += div_term1 << div_term1_shift
+
+ m.d.comb += self.trial_compare_rhs.eq(div_rhs)
+
+ # SqrtRem
+ with m.Elif(self.operation == int(DP.SqrtRem)):
+ qr_times_trial_bits = Signal((tblen+1)*2, reset_less=True)
+ m.d.comb += qr_times_trial_bits.eq(qr * trial_bits_sig)
+ sqrt_rhs = self.compare_rhs
+
+ sqrt_term1 = qr_times_trial_bits
+ sqrt_term1_shift = self.core_config.fract_width
+ sqrt_term1_shift += self.current_shift + 1
+ sqrt_rhs += sqrt_term1 << sqrt_term1_shift
+ sqrt_term2 = trial_bits_sqrd_sig
+ sqrt_term2_shift = self.core_config.fract_width
+ sqrt_term2_shift += self.current_shift * 2
+ sqrt_rhs += sqrt_term2 << sqrt_term2_shift
+
+ m.d.comb += self.trial_compare_rhs.eq(sqrt_rhs)
+
+ # RSqrtRem
+ with m.Else():
+ rr_times_trial_bits = Signal((tblen+1)*3, reset_less=True)
+ m.d.comb += rr_times_trial_bits.eq(rr * trial_bits_sig)
+ rsqrt_rhs = self.compare_rhs
+
+ rsqrt_term1 = rr_times_trial_bits
+ rsqrt_term1_shift = self.current_shift + 1
+ rsqrt_rhs += rsqrt_term1 << rsqrt_term1_shift
+ rsqrt_term2 = dr_times_trial_bits_sqrd
+ rsqrt_term2_shift = self.current_shift * 2
+ rsqrt_rhs += rsqrt_term2 << rsqrt_term2_shift
+
+ m.d.comb += self.trial_compare_rhs.eq(rsqrt_rhs)
+
+ return m
class DivPipeCoreCalculateStage(Elaboratable):
def __init__(self, core_config, stage_index):
""" Create a ``DivPipeCoreSetupStage`` instance. """
self.core_config = core_config
- assert stage_index in range(core_config.num_calculate_stages)
+ assert stage_index in range(core_config.n_stages)
self.stage_index = stage_index
self.i = self.ispec()
self.o = self.ospec()
current_shift -= log2_radix
radix = 1 << log2_radix
trial_compare_rhs_values = []
- pass_flags = []
+ pfl = []
for trial_bits in range(radix):
- shifted_trial_bits = Const(trial_bits, log2_radix) << current_shift
- shifted_trial_bits_sqrd = shifted_trial_bits * shifted_trial_bits
-
- # UDivRem
- div_rhs = self.i.compare_rhs
- div_factor1 = self.i.divisor_radicand * shifted_trial_bits
- div_rhs += div_factor1 << self.core_config.fract_width
-
- # SqrtRem
- sqrt_rhs = self.i.compare_rhs
- sqrt_factor1 = self.i.quotient_root * (shifted_trial_bits << 1)
- sqrt_rhs += sqrt_factor1 << self.core_config.fract_width
- sqrt_factor2 = shifted_trial_bits_sqrd
- sqrt_rhs += sqrt_factor2 << self.core_config.fract_width
-
- # RSqrtRem
- rsqrt_rhs = self.i.compare_rhs
- rsqrt_rhs += self.i.root_times_radicand * (shifted_trial_bits << 1)
- rsqrt_rhs += self.i.divisor_radicand * shifted_trial_bits_sqrd
-
- trial_compare_rhs = self.o.compare_rhs.like(
- name=f"trial_compare_rhs_{trial_bits}")
-
- with m.If(self.i.operation == DivPipeCoreOperation.UDivRem):
- m.d.comb += trial_compare_rhs.eq(div_rhs)
- with m.Elif(self.i.operation == DivPipeCoreOperation.SqrtRem):
- m.d.comb += trial_compare_rhs.eq(sqrt_rhs)
- with m.Else(): # DivPipeCoreOperation.RSqrtRem
- m.d.comb += trial_compare_rhs.eq(rsqrt_rhs)
- trial_compare_rhs_values.append(trial_compare_rhs)
-
- pass_flag = Signal(name=f"pass_flag_{trial_bits}")
- m.d.comb += pass_flag.eq(self.i.compare_lhs >= trial_compare_rhs)
- pass_flags.append(pass_flag)
+ t = Trial(self.core_config, trial_bits,
+ current_shift, log2_radix)
+ setattr(m.submodules, "trial%d" % trial_bits, t)
+ m.d.comb += t.divisor_radicand.eq(self.i.divisor_radicand)
+ m.d.comb += t.quotient_root.eq(self.i.quotient_root)
+ m.d.comb += t.root_times_radicand.eq(self.i.root_times_radicand)
+ m.d.comb += t.compare_rhs.eq(self.i.compare_rhs)
+ m.d.comb += t.operation.eq(self.i.operation)
+
+ trial_compare_rhs_values.append(t.trial_compare_rhs)
+
+ pass_flag = Signal(name=f"pass_flag_{trial_bits}", reset_less=True)
+ m.d.comb += pass_flag.eq(self.i.compare_lhs >= t.trial_compare_rhs)
+ pfl.append(pass_flag)
+ pass_flags = Signal(radix, reset_less=True)
+ m.d.comb += pass_flags.eq(Cat(*pfl))
# convert pass_flags to next_bits.
#
# Assumes that pass_flag[0] is always set (since
# compare_lhs >= compare_rhs is a pipeline invariant).
- next_bits = Signal(log2_radix)
+ next_bits = Signal(log2_radix, reset_less=True)
+ l = []
for i in range(log2_radix):
bit_value = 1
for j in range(0, radix, 1 << i):
bit_value ^= pass_flags[j]
- m.d.comb += next_bits.part(i, 1).eq(bit_value)
+ bv = Signal(reset_less=True)
+ m.d.comb += bv.eq(bit_value)
+ l.append(bv)
+ m.d.comb += next_bits.eq(Cat(*l))
+ # merge/select multi-bit trial_compare_rhs_values, to go
+ # into compare_rhs. XXX (only one of these will succeed?)
next_compare_rhs = 0
for i in range(radix):
- next_flag = pass_flags[i + 1] if i + 1 < radix else 0
- next_compare_rhs |= Mux(pass_flags[i] & ~next_flag,
+ next_flag = Signal(name=f"next_flag{i}", reset_less=True)
+ selected = Signal(name=f"selected_{i}", reset_less=True)
+ m.d.comb += next_flag.eq(~pass_flags[i + 1] if i + 1 < radix else 1)
+ m.d.comb += selected.eq(pass_flags[i] & next_flag)
+ next_compare_rhs |= Mux(selected,
trial_compare_rhs_values[i],
0)
| (next_bits << current_shift))
return m
- # XXX in DivPipeCalculateStage
- DivPipeBaseStage._elaborate(self, m, platform)
-
-
class DivPipeCoreFinalStage(Elaboratable):
""" Final Stage of the core of the div/rem/sqrt/rsqrt pipeline. """
def setup(self, m, i):
""" Pipeline stage setup. """
- m.submodules.div_pipe_core_setup = self
+ m.submodules.div_pipe_core_final = self
m.d.comb += self.i.eq(i)
def process(self, i):
- self.i.compare_rhs)
return m
-
- # XXX in DivPipeFinalStage
- DivPipeBaseStage._elaborate(self, m, platform)
-