src/ieee754/fpdiv/pipeline.py

   1 """IEEE Floating Point Divider Pipeline
   2
   3 Relevant bugreport: http://bugs.libre-riscv.org/show_bug.cgi?id=99
   4
   5 Stack looks like this:
   6
   7 scnorm   - FPDIVSpecialCasesDeNorm ispec FPADDBaseData
   8 ------                             ospec FPSCData
   9
  10                 StageChain: FPDIVSpecialCasesMod,
  11                             FPAddDeNormMod
  12
  13 pipediv0 - FPDivStagesSetup        ispec FPSCData
  14 --------                           ospec DivPipeInterstageData
  15
  16                 StageChain: FPDivStage0Mod,
  17                             DivPipeSetupStage,
  18                             DivPipeCalculateStage,
  19                             ...
  20                             DivPipeCalculateStage
  21
  22 pipediv1 - FPDivStagesIntermediate ispec DivPipeInterstageData
  23 --------                           ospec DivPipeInterstageData
  24
  25                 StageChain: DivPipeCalculateStage,
  26                             ...
  27                             DivPipeCalculateStage
  28 ...
  29 ...
  30
  31 pipediv5 - FPDivStageFinal         ispec FPDivStage0Data
  32 --------                           ospec FPAddStage1Data
  33
  34                 StageChain: DivPipeCalculateStage,
  35                             ...
  36                             DivPipeCalculateStage,
  37                             DivPipeFinalStage,
  38                             FPDivStage2Mod
  39
  40 normpack - FPNormToPack            ispec FPAddStage1Data
  41 --------                           ospec FPPackData
  42
  43                 StageChain: Norm1ModSingle,
  44                             RoundMod,
  45                             CorrectionsMod,
  46                             PackMod
  47
  48 the number of combinatorial StageChains (n_comb_stages) in
  49 FPDivStages is an argument arranged to get the length of the whole
  50 pipeline down to sane numbers.  it specifies the number of "blocks"
  51 that will be combinatorially chained together.
  52
  53 the reason for keeping the number of stages down is that for every
  54 pipeline clock delay, a corresponding ReservationStation is needed.
  55 if there are 24 pipeline stages, we need a whopping TWENTY FOUR
  56 RS's.  that's far too many.  6 is just about an acceptable number.
  57 even 8 is starting to get alarmingly high.
  58 """
  59
  60 from nmigen import Module
  61 from nmigen.cli import main, verilog
  62
  63 from nmutil.singlepipe import ControlBase
  64 from nmutil.concurrentunit import ReservationStations, num_bits
  65
  66 from ieee754.fpcommon.getop import FPADDBaseData
  67 from ieee754.fpcommon.denorm import FPSCData
  68 from ieee754.fpcommon.fpbase import FPFormat
  69 from ieee754.fpcommon.pack import FPPackData
  70 from ieee754.fpcommon.normtopack import FPNormToPack
  71 from ieee754.fpdiv.specialcases import FPDIVSpecialCasesDeNorm
  72 from ieee754.fpdiv.divstages import (FPDivStagesSetup,
  73                                      FPDivStagesIntermediate,
  74                                      FPDivStagesFinal)
  75 from ieee754.pipeline import PipelineSpec
  76 from ieee754.div_rem_sqrt_rsqrt.core import DivPipeCoreConfig
  77
  78
  79 class FPDIVBasePipe(ControlBase):
  80     def __init__(self, pspec):
  81         self.pspec = pspec
  82         ControlBase.__init__(self)
  83
  84         pipechain = []
  85         # to which the answer: "as few as possible"
  86         # is required.  too many ReservationStations
  87         # means "big problems".
  88
  89         # get number of stages, set up loop.
  90         n_stages = pspec.core_config.n_stages
  91         max_n_comb_stages = self.pspec.n_comb_stages
  92         print("n_stages", n_stages)
  93         stage_idx = 0
  94
  95         end = False
  96         while not end:
  97
  98             n_comb_stages = max_n_comb_stages
  99             # needs to convert input from pipestart ospec
 100             if stage_idx == 0:
 101                 n_comb_stages -= 1
 102                 kls = FPDivStagesSetup  # does n_comb_stages-1 calcs as well
 103
 104             # needs to convert output to pipeend ispec
 105             elif stage_idx + n_comb_stages >= n_stages:
 106                 kls = FPDivStagesFinal  # does n_comb_stages-1 calcs as well
 107                 end = True
 108                 n_comb_stages = n_stages - stage_idx
 109
 110             # intermediary stage
 111             else:
 112                 kls = FPDivStagesIntermediate  # does n_comb_stages calcs
 113
 114             # create (in each pipe) a StageChain n_comb_stages in length
 115             pipechain.append(kls(self.pspec, n_comb_stages, stage_idx))
 116             stage_idx += n_comb_stages  # increment so that each CalcStage
 117             # gets a (correct) unique index
 118
 119         self.pipechain = pipechain
 120
 121         # start and end: unpack/specialcases then normalisation/packing
 122         self.pipestart = pipestart = FPDIVSpecialCasesDeNorm(self.pspec)
 123         self.pipeend = pipeend = FPNormToPack(self.pspec)
 124
 125         self._eqs = self.connect([pipestart] + pipechain + [pipeend])
 126
 127     def elaborate(self, platform):
 128         m = ControlBase.elaborate(self, platform)
 129
 130         # add submodules
 131         m.submodules.scnorm = self.pipestart
 132         for i, p in enumerate(self.pipechain):
 133             setattr(m.submodules, "pipediv%d" % i, p)
 134         m.submodules.normpack = self.pipeend
 135
 136         # ControlBase.connect creates the "eqs" needed to connect each pipe
 137         m.d.comb += self._eqs
 138
 139         return m
 140
 141
 142 def roundup(x, mod):
 143     return x if x % mod == 0 else x + mod - x % mod
 144
 145
 146 class FPDIVMuxInOut(ReservationStations):
 147     """ Reservation-Station version of FPDIV pipeline.
 148
 149         * fan-in on inputs (an array of FPADDBaseData: a,b,mid)
 150         * N-stage divider pipeline
 151         * fan-out on outputs (an array of FPPackData: z,mid)
 152
 153         Fan-in and Fan-out are combinatorial.
 154
 155         :op_wid: - set this to the width of an operator which can
 156                    then be used to change the behaviour of the pipeline.
 157     """
 158
 159     def __init__(self, width, num_rows, op_wid=2):
 160         self.id_wid = num_bits(num_rows)
 161         self.pspec = PipelineSpec(width, self.id_wid, op_wid)
 162
 163         # get the standard mantissa width, store in the pspec
 164         fmt = FPFormat.standard(width)
 165         log2_radix = 3     # tested options so far: 1, 2 and 3.
 166         n_comb_stages = 2  # 2 compute stages per pipeline stage
 167
 168         # extra bits needed: guard + round (sticky comes from remainer.bool())
 169         fraction_width = fmt.fraction_width
 170         fraction_width += 2
 171
 172         # rounding width to a multiple of log2_radix is not needed,
 173         # DivPipeCoreCalculateStage just internally reduces log2_radix on
 174         # the last stage
 175         cfg = DivPipeCoreConfig(fmt.width, fraction_width, log2_radix)
 176
 177         self.pspec.fpformat = fmt
 178         self.pspec.n_comb_stages = n_comb_stages
 179         self.pspec.core_config = cfg
 180
 181         # XXX TODO - a class (or function?) that takes the pspec (right here)
 182         # and creates... "something".  that "something" MUST have an eq function
 183         # new_pspec = deepcopy(self.pspec)
 184         # new_pspec.opkls = DivPipeCoreOperation
 185         # self.alu = FPDIVBasePipe(new_pspec)
 186         self.alu = FPDIVBasePipe(self.pspec)
 187         ReservationStations.__init__(self, num_rows)
 188
 189     def i_specfn(self):
 190         return FPADDBaseData(self.pspec)
 191
 192     def o_specfn(self):
 193         return FPPackData(self.pspec)