src/soc/simple/core.py

   1 """simple core
   2
   3 not in any way intended for production use.  connects up FunctionUnits to
   4 Register Files in a brain-dead fashion that only permits one and only one
   5 Function Unit to be operational.
   6
   7 the principle here is to take the Function Units, analyse their regspecs,
   8 and turn their requirements for access to register file read/write ports
   9 into groupings by Register File and Register File Port name.
  10
  11 under each grouping - by regfile/port - a list of Function Units that
  12 need to connect to that port is created.  as these are a contended
  13 resource a "Broadcast Bus" per read/write port is then also created,
  14 with access to it managed by a PriorityPicker.
  15
  16 the brain-dead part of this module is that even though there is no
  17 conflict of access, regfile read/write hazards are *not* analysed,
  18 and consequently it is safer to wait for the Function Unit to complete
  19 before allowing a new instruction to proceed.
  20 """
  21
  22 from nmigen import Elaboratable, Module, Signal, ResetSignal, Cat
  23 from nmigen.cli import rtlil
  24
  25 from nmutil.picker import PriorityPicker
  26 from nmutil.util import treereduce
  27
  28 from soc.fu.compunits.compunits import AllFunctionUnits
  29 from soc.regfile.regfiles import RegFiles
  30 from soc.decoder.power_decoder import create_pdecode
  31 from soc.decoder.power_decoder2 import PowerDecode2
  32 from soc.decoder.decode2execute1 import Data
  33 from soc.experiment.l0_cache import TstL0CacheBuffer  # test only
  34 from soc.config.test.test_loadstore import TestMemPspec
  35 from soc.decoder.power_enums import MicrOp
  36 import operator
  37
  38 from nmutil.util import rising_edge
  39
  40
  41 # helper function for reducing a list of signals down to a parallel
  42 # ORed single signal.
  43 def ortreereduce(tree, attr="data_o"):
  44     return treereduce(tree, operator.or_, lambda x: getattr(x, attr))
  45
  46
  47 def ortreereduce_sig(tree):
  48     return treereduce(tree, operator.or_, lambda x: x)
  49
  50
  51 # helper function to place full regs declarations first
  52 def sort_fuspecs(fuspecs):
  53     res = []
  54     for (regname, fspec) in fuspecs.items():
  55         if regname.startswith("full"):
  56             res.append((regname, fspec))
  57     for (regname, fspec) in fuspecs.items():
  58         if not regname.startswith("full"):
  59             res.append((regname, fspec))
  60     return res  # enumerate(res)
  61
  62
  63 class NonProductionCore(Elaboratable):
  64     def __init__(self, pspec):
  65         # single LD/ST funnel for memory access
  66         self.l0 = TstL0CacheBuffer(pspec, n_units=1)
  67         pi = self.l0.l0.dports[0]
  68
  69         # function units (only one each)
  70         self.fus = AllFunctionUnits(pspec, pilist=[pi])
  71
  72         # register files (yes plural)
  73         self.regs = RegFiles()
  74
  75         # instruction decoder
  76         pdecode = create_pdecode()
  77         self.pdecode2 = PowerDecode2(pdecode)   # instruction decoder
  78
  79         # issue/valid/busy signalling
  80         self.ivalid_i = self.pdecode2.valid   # instruction is valid
  81         self.issue_i = Signal(reset_less=True)
  82         self.busy_o = Signal(name="corebusy_o", reset_less=True)
  83
  84         # instruction input
  85         self.bigendian_i = self.pdecode2.dec.bigendian
  86         self.raw_opcode_i = self.pdecode2.dec.raw_opcode_in
  87
  88         # start/stop and terminated signalling
  89         self.core_stopped_i = Signal(reset_less=True)
  90         self.core_reset_i = Signal()
  91         self.core_terminate_o = Signal(reset=0)  # indicates stopped
  92
  93     def elaborate(self, platform):
  94         m = Module()
  95
  96         m.submodules.pdecode2 = dec2 = self.pdecode2
  97         m.submodules.fus = self.fus
  98         m.submodules.l0 = l0 = self.l0
  99         self.regs.elaborate_into(m, platform)
 100         regs = self.regs
 101         fus = self.fus.fus
 102
 103         # connect up Function Units, then read/write ports
 104         fu_bitdict = self.connect_instruction(m)
 105         self.connect_rdports(m, fu_bitdict)
 106         self.connect_wrports(m, fu_bitdict)
 107
 108         # connect up reset
 109         m.d.comb += ResetSignal().eq(self.core_reset_i)
 110
 111         return m
 112
 113     def connect_instruction(self, m):
 114         """connect_instruction
 115
 116         uses decoded (from PowerOp) function unit information from CSV files
 117         to ascertain which Function Unit should deal with the current
 118         instruction.
 119
 120         some (such as OP_ATTN, OP_NOP) are dealt with here, including
 121         ignoring it and halting the processor.  OP_NOP is a bit annoying
 122         because the issuer expects busy flag still to be raised then lowered.
 123         (this requires a fake counter to be set).
 124         """
 125         comb, sync = m.d.comb, m.d.sync
 126         fus = self.fus.fus
 127         dec2 = self.pdecode2
 128
 129         # enable-signals for each FU, get one bit for each FU (by name)
 130         fu_enable = Signal(len(fus), reset_less=True)
 131         fu_bitdict = {}
 132         for i, funame in enumerate(fus.keys()):
 133             fu_bitdict[funame] = fu_enable[i]
 134
 135         # enable the required Function Unit based on the opcode decode
 136         # note: this *only* works correctly for simple core when one and
 137         # *only* one FU is allocated per instruction
 138         for funame, fu in fus.items():
 139             fnunit = fu.fnunit.value
 140             enable = Signal(name="en_%s" % funame, reset_less=True)
 141             comb += enable.eq((dec2.e.do.fn_unit & fnunit).bool())
 142             comb += fu_bitdict[funame].eq(enable)
 143
 144         # sigh - need a NOP counter
 145         counter = Signal(2)
 146         with m.If(counter != 0):
 147             sync += counter.eq(counter - 1)
 148             comb += self.busy_o.eq(1)
 149
 150         with m.If(self.ivalid_i): # run only when valid
 151             with m.Switch(dec2.e.do.insn_type):
 152                 # check for ATTN: halt if true
 153                 with m.Case(MicrOp.OP_ATTN):
 154                     m.d.sync += self.core_terminate_o.eq(1)
 155
 156                 with m.Case(MicrOp.OP_NOP):
 157                     sync += counter.eq(2)
 158                     comb += self.busy_o.eq(1)
 159
 160                 with m.Default():
 161                     # connect up instructions.  only one enabled at a time
 162                     for funame, fu in fus.items():
 163                         enable = fu_bitdict[funame]
 164
 165                         # run this FunctionUnit if enabled
 166                         with m.If(enable):
 167                             # route op, issue, busy, read flags and mask to FU
 168                             comb += fu.oper_i.eq_from_execute1(dec2.e)
 169                             comb += fu.issue_i.eq(self.issue_i)
 170                             comb += self.busy_o.eq(fu.busy_o)
 171                             rdmask = dec2.rdflags(fu)
 172                             comb += fu.rdmaskn.eq(~rdmask)
 173
 174         return fu_bitdict
 175
 176     def connect_rdport(self, m, fu_bitdict, rdpickers, regfile, regname, fspec):
 177         comb, sync = m.d.comb, m.d.sync
 178         fus = self.fus.fus
 179         regs = self.regs
 180
 181         rpidx = regname
 182
 183         # select the required read port.  these are pre-defined sizes
 184         print(rpidx, regfile, regs.rf.keys())
 185         rport = regs.rf[regfile.lower()].r_ports[rpidx]
 186
 187         fspecs = fspec
 188         if not isinstance(fspecs, list):
 189             fspecs = [fspecs]
 190
 191         rdflags = []
 192         pplen = 0
 193         reads = []
 194         ppoffs = []
 195         for i, fspec in enumerate(fspecs):
 196             # get the regfile specs for this regfile port
 197             (rf, read, write, wid, fuspec) = fspec
 198             print ("fpsec", i, fspec, len(fuspec))
 199             ppoffs.append(pplen) # record offset for picker
 200             pplen += len(fuspec)
 201             name = "rdflag_%s_%s_%d" % (regfile, regname, i)
 202             rdflag = Signal(name=name, reset_less=True)
 203             comb += rdflag.eq(rf)
 204             rdflags.append(rdflag)
 205             reads.append(read)
 206
 207         print ("pplen", pplen)
 208
 209         # create a priority picker to manage this port
 210         rdpickers[regfile][rpidx] = rdpick = PriorityPicker(pplen)
 211         setattr(m.submodules, "rdpick_%s_%s" % (regfile, rpidx), rdpick)
 212
 213         for i, fspec in enumerate(fspecs):
 214             (rf, read, write, wid, fuspec) = fspec
 215             # connect up the FU req/go signals, and the reg-read to the FU
 216             # and create a Read Broadcast Bus
 217             for pi, (funame, fu, idx) in enumerate(fuspec):
 218                 pi += ppoffs[i]
 219                 src = fu.src_i[idx]
 220
 221                 # connect request-read to picker input, and output to go-rd
 222                 fu_active = fu_bitdict[funame]
 223                 pick = Signal()
 224                 comb += pick.eq(fu.rd_rel_o[idx] & fu_active & rdflags[i])
 225                 print (pick, len(pick))
 226                 print (rdpick.i, len(rdpick.i), pi)
 227                 comb += rdpick.i[pi].eq(pick)
 228                 comb += fu.go_rd_i[idx].eq(rdpick.o[pi])
 229
 230                 # if picked, select read-port "reg select" number to port
 231                 with m.If(rdpick.o[pi] & rdpick.en_o):
 232                     comb += rport.ren.eq(reads[i])
 233
 234                     # connect regfile port to input, creating a Broadcast Bus
 235                     print("reg connect widths",
 236                           regfile, regname, pi, funame,
 237                           src.shape(), rport.data_o.shape())
 238                     # all FUs connect to same port
 239                     comb += src.eq(rport.data_o)
 240
 241     def connect_rdports(self, m, fu_bitdict):
 242         """connect read ports
 243
 244         orders the read regspecs into a dict-of-dicts, by regfile, by
 245         regport name, then connects all FUs that want that regport by
 246         way of a PriorityPicker.
 247         """
 248         comb, sync = m.d.comb, m.d.sync
 249         fus = self.fus.fus
 250         regs = self.regs
 251
 252         # dictionary of lists of regfile read ports
 253         byregfiles_rd, byregfiles_rdspec = self.get_byregfiles(True)
 254
 255         # okaay, now we need a PriorityPicker per regfile per regfile port
 256         # loootta pickers... peter piper picked a pack of pickled peppers...
 257         rdpickers = {}
 258         for regfile, spec in byregfiles_rd.items():
 259             fuspecs = byregfiles_rdspec[regfile]
 260             rdpickers[regfile] = {}
 261
 262             # argh.  an experiment to merge RA and RB in the INT regfile
 263             # (we have too many read/write ports)
 264             if regfile == 'INT':
 265                 fuspecs['rbc'] = [fuspecs.pop('rb')]
 266                 fuspecs['rbc'].append(fuspecs.pop('rc'))
 267             if regfile == 'FAST':
 268                 fuspecs['fast1'] = [fuspecs.pop('fast1')]
 269                 fuspecs['fast1'].append(fuspecs.pop('fast2'))
 270
 271             # for each named regfile port, connect up all FUs to that port
 272             for (regname, fspec) in sort_fuspecs(fuspecs):
 273                 print("connect rd", regname, fspec)
 274                 self.connect_rdport(m, fu_bitdict, rdpickers, regfile,
 275                                        regname, fspec)
 276
 277     def connect_wrport(self, m, fu_bitdict, wrpickers, regfile, regname, fspec):
 278         comb, sync = m.d.comb, m.d.sync
 279         fus = self.fus.fus
 280         regs = self.regs
 281
 282         print("connect wr", regname, fspec)
 283         rpidx = regname
 284         # get the regfile specs for this regfile port
 285         (rf, read, write, wid, fuspec) = fspec
 286
 287         # select the required write port.  these are pre-defined sizes
 288         print(regfile, regs.rf.keys())
 289         wport = regs.rf[regfile.lower()].w_ports[rpidx]
 290
 291         # create a priority picker to manage this port
 292         wrpickers[regfile][rpidx] = wrpick = PriorityPicker(
 293             len(fuspec))
 294         setattr(m.submodules, "wrpick_%s_%s" %
 295                 (regfile, rpidx), wrpick)
 296
 297         # connect the regspec write "reg select" number to this port
 298         # only if one FU actually requests (and is granted) the port
 299         # will the write-enable be activated
 300         with m.If(wrpick.en_o):
 301             comb += wport.wen.eq(write)
 302         with m.Else():
 303             comb += wport.wen.eq(0)
 304
 305         # connect up the FU req/go signals and the reg-read to the FU
 306         # these are arbitrated by Data.ok signals
 307         wsigs = []
 308         for pi, (funame, fu, idx) in enumerate(fuspec):
 309             # write-request comes from dest.ok
 310             dest = fu.get_out(idx)
 311             fu_dest_latch = fu.get_fu_out(idx)  # latched output
 312             name = "wrflag_%s_%s_%d" % (funame, regname, idx)
 313             wrflag = Signal(name=name, reset_less=True)
 314             comb += wrflag.eq(dest.ok & fu.busy_o)
 315
 316             # connect request-write to picker input, and output to go-wr
 317             fu_active = fu_bitdict[funame]
 318             pick = fu.wr.rel_o[idx] & fu_active  # & wrflag
 319             comb += wrpick.i[pi].eq(pick)
 320             # create a single-pulse go write from the picker output
 321             wr_pick = Signal()
 322             comb += wr_pick.eq(wrpick.o[pi] & wrpick.en_o)
 323             comb += fu.go_wr_i[idx].eq(rising_edge(m, wr_pick))
 324             # connect regfile port to input
 325             print("reg connect widths",
 326                   regfile, regname, pi, funame,
 327                   dest.shape(), wport.data_i.shape())
 328             wsigs.append(fu_dest_latch)
 329
 330         # here is where we create the Write Broadcast Bus. simple, eh?
 331         comb += wport.data_i.eq(ortreereduce_sig(wsigs))
 332
 333     def connect_wrports(self, m, fu_bitdict):
 334         """connect write ports
 335
 336         orders the write regspecs into a dict-of-dicts, by regfile,
 337         by regport name, then connects all FUs that want that regport
 338         by way of a PriorityPicker.
 339
 340         note that the write-port wen, write-port data, and go_wr_i all need to
 341         be on the exact same clock cycle.  as there is a combinatorial loop bug
 342         at the moment, these all use sync.
 343         """
 344         comb, sync = m.d.comb, m.d.sync
 345         fus = self.fus.fus
 346         regs = self.regs
 347         # dictionary of lists of regfile write ports
 348         byregfiles_wr, byregfiles_wrspec = self.get_byregfiles(False)
 349
 350         # same for write ports.
 351         # BLECH!  complex code-duplication! BLECH!
 352         wrpickers = {}
 353         for regfile, spec in byregfiles_wr.items():
 354             fuspecs = byregfiles_wrspec[regfile]
 355             wrpickers[regfile] = {}
 356             for (regname, fspec) in sort_fuspecs(fuspecs):
 357                 self.connect_wrport(m, fu_bitdict, wrpickers,
 358                                         regfile, regname, fspec)
 359
 360     def get_byregfiles(self, readmode):
 361
 362         mode = "read" if readmode else "write"
 363         dec2 = self.pdecode2
 364         regs = self.regs
 365         fus = self.fus.fus
 366
 367         # dictionary of lists of regfile ports
 368         byregfiles = {}
 369         byregfiles_spec = {}
 370         for (funame, fu) in fus.items():
 371             print("%s ports for %s" % (mode, funame))
 372             for idx in range(fu.n_src if readmode else fu.n_dst):
 373                 if readmode:
 374                     (regfile, regname, wid) = fu.get_in_spec(idx)
 375                 else:
 376                     (regfile, regname, wid) = fu.get_out_spec(idx)
 377                 print("    %d %s %s %s" % (idx, regfile, regname, str(wid)))
 378                 if readmode:
 379                     rdflag, read = dec2.regspecmap_read(regfile, regname)
 380                     write = None
 381                 else:
 382                     rdflag, read = None, None
 383                     wrport, write = dec2.regspecmap_write(regfile, regname)
 384                 if regfile not in byregfiles:
 385                     byregfiles[regfile] = {}
 386                     byregfiles_spec[regfile] = {}
 387                 if regname not in byregfiles_spec[regfile]:
 388                     byregfiles_spec[regfile][regname] = \
 389                         (rdflag, read, write, wid, [])
 390                 # here we start to create "lanes"
 391                 if idx not in byregfiles[regfile]:
 392                     byregfiles[regfile][idx] = []
 393                 fuspec = (funame, fu, idx)
 394                 byregfiles[regfile][idx].append(fuspec)
 395                 byregfiles_spec[regfile][regname][4].append(fuspec)
 396
 397         # ok just print that out, for convenience
 398         for regfile, spec in byregfiles.items():
 399             print("regfile %s ports:" % mode, regfile)
 400             fuspecs = byregfiles_spec[regfile]
 401             for regname, fspec in fuspecs.items():
 402                 [rdflag, read, write, wid, fuspec] = fspec
 403                 print("  rf %s port %s lane: %s" % (mode, regfile, regname))
 404                 print("  %s" % regname, wid, read, write, rdflag)
 405                 for (funame, fu, idx) in fuspec:
 406                     fusig = fu.src_i[idx] if readmode else fu.dest[idx]
 407                     print("    ", funame, fu, idx, fusig)
 408                     print()
 409
 410         return byregfiles, byregfiles_spec
 411
 412     def __iter__(self):
 413         yield from self.fus.ports()
 414         yield from self.pdecode2.ports()
 415         yield from self.l0.ports()
 416         # TODO: regs
 417
 418     def ports(self):
 419         return list(self)
 420
 421
 422 if __name__ == '__main__':
 423     pspec = TestMemPspec(ldst_ifacetype='testpi',
 424                          imem_ifacetype='',
 425                          addr_wid=48,
 426                          mask_wid=8,
 427                          reg_wid=64)
 428     dut = NonProductionCore(pspec)
 429     vl = rtlil.convert(dut, ports=dut.ports())
 430     with open("test_core.il", "w") as f:
 431         f.write(vl)