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, Cat)
+from nmigen import (Elaboratable, Module, Signal, Const, Mux, Cat, Array)
+from nmigen.lib.coding import PriorityEncoder
import enum
+ 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
def create_signal(cls, *, src_loc_at=0, **kwargs):
""" Create a signal that can contain a ``DivPipeCoreOperation``. """
return Signal(min=min(map(int, cls)),
- max=max(map(int, cls)),
+ max=max(map(int, cls)) + 2,
src_loc_at=(src_loc_at + 1),
decoder=lambda v: str(cls(v)),
**kwargs)
+DP = DivPipeCoreOperation
+
+
class DivPipeCoreInputData:
""" input data type for ``DivPipeCore``.
reset_less=reset_less)
self.divisor_radicand = Signal(core_config.bit_width,
reset_less=reset_less)
-
- # FIXME: this goes into (is replaced by) self.ctx.op
- self.operation = \
- DivPipeCoreOperation.create_signal(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
+ 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.
+ self.operation.eq(rhs.operation),
]
self.core_config = core_config
self.divisor_radicand = Signal(core_config.bit_width,
reset_less=reset_less)
- # FIXME: delete self.operation. already covered by self.ctx.op
- self.operation = \
- DivPipeCoreOperation.create_signal(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,
def __iter__(self):
""" Get member signals. """
yield self.divisor_radicand
- yield self.operation # 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),
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 == int(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 == int(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
+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):
""" Calculate Stage of the core of the div/rem/sqrt/rsqrt pipeline. """
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()
def elaborate(self, platform):
""" Elaborate into ``Module``. """
m = Module()
+
+ # copy invariant inputs to outputs (for next stage)
m.d.comb += self.o.divisor_radicand.eq(self.i.divisor_radicand)
m.d.comb += self.o.operation.eq(self.i.operation)
m.d.comb += self.o.compare_lhs.eq(self.i.compare_lhs)
+
+ # constants
log2_radix = self.core_config.log2_radix
current_shift = self.core_config.bit_width
current_shift -= self.stage_index * log2_radix
assert log2_radix > 0
current_shift -= log2_radix
radix = 1 << log2_radix
+
+ # trials within this radix range. carried out by Trial module,
+ # results stored in pass_flags. pass_flags are unary priority.
trial_compare_rhs_values = []
- pass_flags = []
+ pfl = []
for trial_bits in range(radix):
- tb = trial_bits << current_shift
- tb_width = log2_radix + current_shift
- shifted_trial_bits = Const(tb, tb_width)
- shifted_trial_bits2 = Const(tb*2, tb_width+1)
- shifted_trial_bits_sqrd = Const(tb * tb, tb_width * 2)
-
- # UDivRem
- div_rhs = self.i.compare_rhs
- if tb != 0: # no point adding stuff that's multiplied by zero
- div_factor1 = self.i.divisor_radicand * shifted_trial_bits2
- div_rhs += div_factor1 << self.core_config.fract_width
-
- # SqrtRem
- sqrt_rhs = self.i.compare_rhs
- if tb != 0: # no point adding stuff that's multiplied by zero
- sqrt_factor1 = self.i.quotient_root * shifted_trial_bits2
- 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
- if tb != 0: # no point adding stuff that's multiplied by zero
- rsqrt_rhs += self.i.root_times_radicand * shifted_trial_bits2
- rsqrt_rhs += self.i.divisor_radicand * shifted_trial_bits_sqrd
-
- trial_compare_rhs = Signal.like(
- self.o.compare_rhs, name=f"trial_compare_rhs_{trial_bits}",
- reset_less=True)
-
- with m.If(self.i.operation == int(DivPipeCoreOperation.UDivRem)):
- m.d.comb += trial_compare_rhs.eq(div_rhs)
- with m.Elif(self.i.operation == int(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)
+ 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)
+
+ # get the trial output
+ trial_compare_rhs_values.append(t.trial_compare_rhs)
+
+ # make the trial comparison against the [invariant] lhs.
+ # trial_compare_rhs is always decreasing as trial_bits increases
pass_flag = Signal(name=f"pass_flag_{trial_bits}", reset_less=True)
- m.d.comb += pass_flag.eq(self.i.compare_lhs >= trial_compare_rhs)
- pass_flags.append(pass_flag)
+ m.d.comb += pass_flag.eq(self.i.compare_lhs >= t.trial_compare_rhs)
+ pfl.append(pass_flag)
+
+ # Cat all the pass flags list together (easier to handle, below)
+ pass_flags = Signal(radix, reset_less=True)
+ m.d.comb += pass_flags.eq(Cat(*pfl))
- # convert pass_flags to next_bits.
+ # convert pass_flags (unary priority) to next_bits (binary index)
#
# Assumes that for each set bit in pass_flag, all previous bits are
# also set.
# Assumes that pass_flag[0] is always set (since
# compare_lhs >= compare_rhs is a pipeline invariant).
+ m.submodules.pe = pe = PriorityEncoder(radix)
next_bits = Signal(log2_radix, reset_less=True)
- 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)
-
- next_compare_rhs = Signal(radix, reset_less=True)
- l = []
- for i in range(radix):
- next_flag = pass_flags[i + 1] if (i + 1 < radix) else Const(0)
- flag = Signal(reset_less=True, name=f"flag{i}")
- test = Signal(reset_less=True, name=f"test{i}")
- # XXX TODO: check the width on this
- m.d.comb += test.eq((pass_flags[i] & ~next_flag))
- m.d.comb += flag.eq(Mux(test, trial_compare_rhs_values[i], 0))
- l.append(flag)
-
- m.d.comb += next_compare_rhs.eq(Cat(*l))
- m.d.comb += self.o.compare_rhs.eq(next_compare_rhs.bool())
+ m.d.comb += pe.i.eq(~pass_flags)
+ with m.If(~pe.n):
+ m.d.comb += next_bits.eq(pe.o-1)
+ with m.Else():
+ m.d.comb += next_bits.eq(radix-1)
+
+ # get the highest passing rhs trial (indexed by next_bits)
+ ta = Array(trial_compare_rhs_values)
+ m.d.comb += self.o.compare_rhs.eq(ta[next_bits])
+
+ # create outputs for next phase
m.d.comb += self.o.root_times_radicand.eq(self.i.root_times_radicand
+ ((self.i.divisor_radicand
* next_bits)