-"""IEEE754 Floating Point Divider
+"""IEEE754 Floating Point Divider / Square-Root / Reciprocal-Square-Root
+
+Copyright (C) 2019 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
+Copyright (C) 2019 Jacob Lifshay
+
+Relevant bugreports:
+* http://bugs.libre-riscv.org/show_bug.cgi?id=99
+* http://bugs.libre-riscv.org/show_bug.cgi?id=43
+* http://bugs.libre-riscv.org/show_bug.cgi?id=44
-Relevant bugreport: http://bugs.libre-riscv.org/show_bug.cgi?id=99
"""
-from nmigen import Module, Signal, Cat, Elaboratable, Const
+from nmigen import Module, Signal, Cat, Elaboratable, Const, Mux
from nmigen.cli import main, verilog
-from ieee754.fpcommon.fpbase import (FPNumBaseRecord, Overflow)
-from ieee754.fpcommon.fpbase import FPState
+from nmutil.pipemodbase import PipeModBase
+from ieee754.fpcommon.fpbase import FPNumBaseRecord
from ieee754.fpcommon.denorm import FPSCData
from ieee754.fpcommon.getop import FPPipeContext
from ieee754.div_rem_sqrt_rsqrt.div_pipe import DivPipeInputData
+from ieee754.div_rem_sqrt_rsqrt.core import DivPipeCoreOperation as DPCOp
-class FPDivStage0Mod(Elaboratable):
+class FPDivPreFPAdjust(PipeModBase):
+ """ DIV/SQRT/RSQRT "preparation" module.
+ adjusts mantissa and exponent (sqrt/rsqrt exponent must be even),
+ puts exponent (and sign) into data structures for passing through to
+ the end, and puts the (adjusted) mantissa into the processing engine.
+
+ no *actual* processing occurs here: it is *purely* preparation work.
+ """
def __init__(self, pspec):
- self.pspec = pspec
- self.i = self.ispec()
- self.o = self.ospec()
+ super().__init__(pspec, "pre_fp_adjust")
def ispec(self):
return FPSCData(self.pspec, False)
def ospec(self):
return DivPipeInputData(self.pspec)
- def process(self, i):
- return self.o
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- m.submodules.div0 = self
- m.d.comb += self.i.eq(i)
-
def elaborate(self, platform):
m = Module()
-
- # XXX TODO, actual DIV code here. this class would be
- # "step one" which takes the pre-normalised data (see ispec) and
- # *begins* the processing phase (enters the massive DIV
- # pipeline chain) - see ospec.
-
- # INPUT SPEC: FPSCData
- # OUTPUT SPEC: DivPipeInputData
-
- # NOTE: this stage does *NOT* do *ACTUAL* DIV processing,
- # it is PURELY the *ENTRY* point into the chain, performing
- # "preparation" work.
-
- with m.If(~self.i.out_do_z):
- # do conversion here, of both self.i.a and self.i.b,
- # into DivPipeInputData dividend and divisor.
-
- if self.pspec.width == 16:
- if self.pspec.log2_radix == 1:
- extra = 2
- elif self.pspec.log2_radix == 3:
- extra = 2
- else:
- extra = 3
- elif self.pspec.width == 32:
- if self.pspec.log2_radix == 1:
- extra = 3
- else:
- extra = 4
- elif self.pspec.width == 64:
- if self.pspec.log2_radix == 1:
- extra = 2
- elif self.pspec.log2_radix == 3:
- extra = 2
- else:
- extra = 3
- # the mantissas, having been de-normalised (and containing
- # a "1" in the MSB) represent numbers in the range 0.5 to
- # 0.9999999-recurring. the min and max range of the
- # result is therefore 0.4999999 (0.5/0.99999) and 1.9999998
- # (0.99999/0.5).
-
- # zero-extend the mantissas (room for sticky/guard)
- # plus the extra MSB. See DivPipeBaseStage.get_core_config
- am0 = Signal(len(self.i.a.m)+3, reset_less=True)
- bm0 = Signal(len(self.i.b.m)+3, reset_less=True)
- m.d.comb += [
- am0.eq(Cat(0,0,0,self.i.a.m, 0)),
- bm0.eq(Cat(0,0,0,self.i.b.m, 0)),
- #am0.eq(0x392),
- #bm0.eq(0x1110),
- ]
-
- m.d.comb += [self.o.z.e.eq(self.i.a.e - self.i.b.e + 1),
- self.o.z.s.eq(self.i.a.s ^ self.i.b.s),
- self.o.dividend[len(self.i.a.m)+extra:].eq(am0), # TODO: check
- self.o.divisor_radicand.eq(bm0), # TODO: check
- self.o.operation.eq(Const(0)) # TODO check: DIV
- ]
-
- # these are required and must not be touched
- 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)
+ comb = m.d.comb
+
+ # mantissas start in the range [1.0, 2.0)
+
+ # intermediary temp signals
+ is_div = Signal(reset_less=True)
+ need_exp_adj = Signal(reset_less=True)
+
+ # "adjusted" - ``self.i.a.rmw`` fractional bits and 2 integer bits
+ adj_a_mw = self.i.a.rmw
+ adj_a_m = Signal(self.i.a.rmw + 2, reset_less=True)
+ adj_a_e = Signal((len(self.i.a.e), True), reset_less=True)
+
+ # adjust (shift) the exponent so that it is even, but only for [r]sqrt
+ comb += [is_div.eq(self.i.ctx.op == int(DPCOp.UDivRem)),
+ need_exp_adj.eq(~is_div & self.i.a.e[0]), # even? !div? adjust
+ adj_a_m.eq(self.i.a.m << need_exp_adj),
+ adj_a_e.eq(self.i.a.e - need_exp_adj)]
+
+ # adj_a_m now in the range [1.0, 4.0) for sqrt/rsqrt
+ # and [1.0, 2.0) for div
+
+ fw = self.pspec.core_config.fract_width
+ divr_rad = Signal(len(self.o.divisor_radicand), reset_less=True)
+
+ # real mantissa fractional widths
+ a_mw = self.i.a.rmw
+ b_mw = self.i.b.rmw
+
+ comb += [self.o.dividend.eq(self.i.a.m << (fw*2 - a_mw)),
+ divr_rad.eq(Mux(is_div, self.i.b.m << (fw - b_mw),
+ adj_a_m << (fw - adj_a_mw))),
+ self.o.divisor_radicand.eq(divr_rad),
+ ]
+
+ ############# DIV #############
+ with m.If(self.i.ctx.op == int(DPCOp.UDivRem)):
+ # DIV: subtract exponents, XOR sign
+ comb += [self.o.z.e.eq(self.i.a.e - self.i.b.e),
+ self.o.z.s.eq(self.i.a.s ^ self.i.b.s),
+ self.o.operation.eq(int(DPCOp.UDivRem))
+ ]
+
+ ############# SQRT #############
+ with m.Elif(self.i.ctx.op == int(DPCOp.SqrtRem)):
+ # SQRT: sign is the same, [adjusted] exponent is halved
+ comb += [self.o.z.e.eq(adj_a_e >> 1), # halve
+ self.o.z.s.eq(self.i.a.s),
+ self.o.operation.eq(int(DPCOp.SqrtRem))
+ ]
+
+ ############# RSQRT #############
+ with m.Elif(self.i.ctx.op == int(DPCOp.RSqrtRem)):
+ # RSQRT: sign same, [adjusted] exponent halved and inverted
+ comb += [self.o.z.e.eq(-(adj_a_e >> 1)), # NEGATE and halve
+ self.o.z.s.eq(self.i.a.s),
+ self.o.operation.eq(int(DPCOp.RSqrtRem))
+ ]
+
+ # pass through context
+ comb += self.o.oz.eq(self.i.oz)
+ comb += self.o.out_do_z.eq(self.i.out_do_z)
+ comb += self.o.ctx.eq(self.i.ctx)
return m
-class FPDivStage0(FPState):
- """ First stage of div.
- """
-
- def __init__(self, pspec):
- FPState.__init__(self, "divider_0")
- self.mod = FPDivStage0Mod(pspec)
- self.o = self.mod.ospec()
-
- def setup(self, m, i):
- """ links module to inputs and outputs
- """
- self.mod.setup(m, i)
-
- # NOTE: these could be done as combinatorial (merge div0+div1)
- m.d.sync += self.o.eq(self.mod.o)
-
- def action(self, m):
- m.next = "divider_1"
projects / ieee754fpu.git / blobdiff