big (single-purpose) update: move width arg into pspec
[ieee754fpu.git] / src / ieee754 / fpdiv / pipeline.py
index 2cb53b90cc4bab51c37c5a851475c0a32d4c8ef3..58fa8e9b0df0e73e64799d7d1ce57dd07d78ddf8 100644 (file)
@@ -1,4 +1,60 @@
-# IEEE Floating Point Divider Pipeline
+"""IEEE Floating Point Divider Pipeline
+
+Relevant bugreport: http://bugs.libre-riscv.org/show_bug.cgi?id=99
+
+Stack looks like this:
+
+scnorm   - FPDIVSpecialCasesDeNorm ispec FPADDBaseData
+------                             ospec FPSCData
+
+                StageChain: FPDIVSpecialCasesMod,
+                            FPAddDeNormMod
+
+pipediv0 - FPDivStagesSetup        ispec FPSCData
+--------                           ospec DivPipeCoreInterstageData
+
+                StageChain: FPDivStage0Mod,
+                            DivPipeSetupStage,
+                            DivPipeCalculateStage,
+                            ...
+                            DivPipeCalculateStage
+
+pipediv1 - FPDivStagesIntermediate ispec DivPipeCoreInterstageData
+--------                           ospec DivPipeCoreInterstageData
+
+                StageChain: DivPipeCalculateStage,
+                            ...
+                            DivPipeCalculateStage
+...
+...
+
+pipediv5 - FPDivStageFinal         ispec FPDivStage0Data
+--------                           ospec FPAddStage1Data
+
+                StageChain: DivPipeCalculateStage,
+                            ...
+                            DivPipeCalculateStage,
+                            DivPipeFinalStage,
+                            FPDivStage2Mod
+
+normpack - FPNormToPack            ispec FPAddStage1Data
+--------                           ospec FPPackData
+
+                StageChain: Norm1ModSingle,
+                            RoundMod,
+                            CorrectionsMod,
+                            PackMod
+
+the number of combinatorial StageChains (n_comb_stages) in
+FPDivStages is an argument arranged to get the length of the whole
+pipeline down to sane numbers.
+
+the reason for keeping the number of stages down is that for every
+pipeline clock delay, a corresponding ReservationStation is needed.
+if there are 24 pipeline stages, we need a whopping TWENTY FOUR
+RS's.  that's far too many.  6 is just about an acceptable number.
+even 8 is starting to get alarmingly high.
+"""
 
 from nmigen import Module
 from nmigen.cli import main, verilog
@@ -10,26 +66,58 @@ from ieee754.fpcommon.getop import FPADDBaseData
 from ieee754.fpcommon.denorm import FPSCData
 from ieee754.fpcommon.pack import FPPackData
 from ieee754.fpcommon.normtopack import FPNormToPack
-from .specialcases import FPDivSpecialCasesDeNorm
-from .divstages import FPDivStages
-
+from .specialcases import FPDIVSpecialCasesDeNorm
+from .divstages import (FPDivStagesSetup,
+                        FPDivStagesIntermediate,
+                        FPDivStagesFinal)
 
 
 class FPDIVBasePipe(ControlBase):
-    def __init__(self, width, id_wid):
+    def __init__(self, pspec):
         ControlBase.__init__(self)
-        self.pipe1 = FPDivSpecialCasesDeNorm(width, id_wid)
-        self.pipe2 = FPDivStages(width, id_wid)
-        self.pipe3 = FPNormToPack(width, id_wid)
-
-        self._eqs = self.connect([self.pipe1, self.pipe2, self.pipe3])
+        self.pspec = pspec
 
     def elaborate(self, platform):
         m = ControlBase.elaborate(self, platform)
-        m.submodules.scnorm = self.pipe1
-        m.submodules.divstages = self.pipe2
-        m.submodules.normpack = self.pipe3
-        m.d.comb += self._eqs
+
+        pipechain = []
+        n_stages = 6      # TODO (depends on width)
+        n_comb_stages = 3 # TODO (depends on how many RS's we want)
+                          # to which the answer: "as few as possible"
+                          # is required.  too many ReservationStations
+                          # means "big problems".
+
+        for i in range(n_stages):
+
+            # needs to convert input from pipestart ospec
+            if i == 0:
+                kls = FPDivStagesSetup
+                n_comb_stages -= 1 # reduce due to work done at start
+
+            # needs to convert output to pipeend ispec
+            elif i == n_stages - 1:
+                kls = FPDivStagesFinal
+                n_comb_stages -= 1 # FIXME - reduce due to work done at end?
+
+            # intermediary stage
+            else:
+                kls = FPDivStagesIntermediate
+
+            pipechain.append(kls(self.pspec, n_comb_stages))
+
+        # start and end: unpack/specialcases then normalisation/packing
+        pipestart = FPDIVSpecialCasesDeNorm(self.pspec)
+        pipeend = FPNormToPack(self.pspec)
+
+        # add submodules
+        m.submodules.scnorm = pipestart
+        for i, p in enumerate(pipechain):
+            setattr(m.submodules, "pipediv%d" % i, p)
+        m.submodules.normpack = pipeend
+
+        # ControlBase.connect creates the "eqs" needed to connect each pipe
+        m.d.comb += self.connect([pipestart] + pipechain + [pipeend])
+
         return m
 
 
@@ -41,15 +129,18 @@ class FPDIVMuxInOut(ReservationStations):
         * fan-out on outputs (an array of FPPackData: z,mid)
 
         Fan-in and Fan-out are combinatorial.
+
+        :op_wid: - set this to the width of an operator which can
+                   then be used to change the behaviour of the pipeline.
     """
-    def __init__(self, width, num_rows):
-        self.width = width
+    def __init__(self, width, num_rows, op_wid=0):
         self.id_wid = num_bits(width)
-        self.alu = FPDIVBasePipe(width, self.id_wid)
+        self.pspec = {'width': width, 'id_wid': self.id_wid, 'op_wid': op_wid}
+        self.alu = FPDIVBasePipe(self.pspec)
         ReservationStations.__init__(self, num_rows)
 
     def i_specfn(self):
-        return FPADDBaseData(self.width, self.id_wid)
+        return FPADDBaseData(self.pspec)
 
     def o_specfn(self):
-        return FPPackData(self.width, self.id_wid)
+        return FPPackData(self.pspec)