more random experimenting
[ieee754fpu.git] / 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.
51
52 the reason for keeping the number of stages down is that for every
53 pipeline clock delay, a corresponding ReservationStation is needed.
54 if there are 24 pipeline stages, we need a whopping TWENTY FOUR
55 RS's. that's far too many. 6 is just about an acceptable number.
56 even 8 is starting to get alarmingly high.
57 """
58
59 from nmigen import Module
60 from nmigen.cli import main, verilog
61
62 from nmutil.singlepipe import ControlBase
63 from nmutil.concurrentunit import ReservationStations, num_bits
64
65 from ieee754.fpcommon.getop import FPADDBaseData
66 from ieee754.fpcommon.denorm import FPSCData
67 from ieee754.fpcommon.fpbase import FPFormat
68 from ieee754.fpcommon.pack import FPPackData
69 from ieee754.fpcommon.normtopack import FPNormToPack
70 from ieee754.fpdiv.specialcases import FPDIVSpecialCasesDeNorm
71 from ieee754.fpdiv.divstages import (FPDivStagesSetup,
72 FPDivStagesIntermediate,
73 FPDivStagesFinal)
74 from ieee754.pipeline import PipelineSpec
75 from ieee754.div_rem_sqrt_rsqrt.core import DivPipeCoreConfig
76
77
78 class FPDIVBasePipe(ControlBase):
79 def __init__(self, pspec):
80 self.pspec = pspec
81 ControlBase.__init__(self)
82
83 pipechain = []
84 max_n_comb_stages = 3 # TODO (depends on how many RS's we want)
85 # to which the answer: "as few as possible"
86 # is required. too many ReservationStations
87 # means "big problems".
88
89 # XXX BUG - subtracting 4 from number of stages stops assert
90 # probably related to having to add 4 in FPDivMuxInOut
91 radix = pspec.log2_radix
92 n_stages = pspec.core_config.n_stages // max_n_comb_stages
93 stage_idx = 0
94
95 for i in range(n_stages):
96
97 n_comb_stages = max_n_comb_stages
98 # needs to convert input from pipestart ospec
99 if i == 0:
100 kls = FPDivStagesSetup
101 #n_comb_stages -= 1 # reduce due to work done at start?
102
103 # needs to convert output to pipeend ispec
104 elif i == n_stages - 1:
105 kls = FPDivStagesFinal
106 #n_comb_stages -= 1 # FIXME - reduce due to work done at end?
107
108 # intermediary stage
109 else:
110 kls = FPDivStagesIntermediate
111
112 pipechain.append(kls(self.pspec, n_comb_stages, stage_idx))
113 stage_idx += n_comb_stages # increment so that each CalcStage
114 # gets a (correct) unique index
115
116 self.pipechain = pipechain
117
118 # start and end: unpack/specialcases then normalisation/packing
119 self.pipestart = pipestart = FPDIVSpecialCasesDeNorm(self.pspec)
120 self.pipeend = pipeend = FPNormToPack(self.pspec)
121
122 self._eqs = self.connect([pipestart] + pipechain + [pipeend])
123
124 def elaborate(self, platform):
125 m = ControlBase.elaborate(self, platform)
126
127 # add submodules
128 m.submodules.scnorm = self.pipestart
129 for i, p in enumerate(self.pipechain):
130 setattr(m.submodules, "pipediv%d" % i, p)
131 m.submodules.normpack = self.pipeend
132
133 # ControlBase.connect creates the "eqs" needed to connect each pipe
134 m.d.comb += self._eqs
135
136 return m
137
138 def roundup(x, mod):
139 return x if x % mod == 0 else x + mod - x % mod
140
141
142 class FPDIVMuxInOut(ReservationStations):
143 """ Reservation-Station version of FPDIV pipeline.
144
145 * fan-in on inputs (an array of FPADDBaseData: a,b,mid)
146 * N-stage divider pipeline
147 * fan-out on outputs (an array of FPPackData: z,mid)
148
149 Fan-in and Fan-out are combinatorial.
150
151 :op_wid: - set this to the width of an operator which can
152 then be used to change the behaviour of the pipeline.
153 """
154
155 def __init__(self, width, num_rows, op_wid=1):
156 self.id_wid = num_bits(width)
157 self.pspec = PipelineSpec(width, self.id_wid, op_wid)
158 # get the standard mantissa width, store in the pspec HOWEVER...
159 fmt = FPFormat.standard(width)
160 log2_radix = 2
161
162 # ...5 extra bits on the mantissa: MSB is zero, MSB-1 is 1
163 # then there is guard, round and sticky at the LSB end.
164 # also: round up to nearest radix
165 fmt.m_width = roundup(fmt.m_width + 5, log2_radix)
166
167 cfg = DivPipeCoreConfig(fmt.m_width, fmt.fraction_width, log2_radix)
168
169 self.pspec.fpformat = fmt
170 self.pspec.log2_radix = log2_radix
171 self.pspec.core_config = cfg
172
173 # XXX TODO - a class (or function?) that takes the pspec (right here)
174 # and creates... "something". that "something" MUST have an eq function
175 # new_pspec = deepcopy(self.pspec)
176 # new_pspec.opkls = DivPipeCoreOperation
177 # self.alu = FPDIVBasePipe(new_pspec)
178 self.alu = FPDIVBasePipe(self.pspec)
179 ReservationStations.__init__(self, num_rows)
180
181 def i_specfn(self):
182 return FPADDBaseData(self.pspec)
183
184 def o_specfn(self):
185 return FPPackData(self.pspec)