src/soc/experiment/l0_cache.py

   1 """L0 Cache/Buffer
   2
   3 This first version is intended for prototyping and test purposes:
   4 it has "direct" access to Memory.
   5
   6 The intention is that this version remains an integral part of the
   7 test infrastructure, and, just as with minerva's memory arrangement,
   8 a dynamic runtime config *selects* alternative memory arrangements
   9 rather than *replaces and discards* this code.
  10
  11 Links:
  12
  13 * https://bugs.libre-soc.org/show_bug.cgi?id=216
  14 * https://libre-soc.org/3d_gpu/architecture/memory_and_cache/
  15
  16 """
  17
  18 from nmigen.compat.sim import run_simulation
  19 from nmigen.cli import verilog, rtlil
  20 from nmigen import Module, Signal, Mux, Elaboratable, Array, Cat
  21 from nmutil.iocontrol import RecordObject
  22 from nmigen.utils import log2_int
  23
  24 from nmutil.latch import SRLatch, latchregister
  25 from soc.decoder.power_decoder2 import Data
  26 from soc.decoder.power_enums import InternalOp
  27
  28 from soc.experiment.compldst import CompLDSTOpSubset
  29 from soc.decoder.power_decoder2 import Data
  30 #from nmutil.picker import PriorityPicker
  31 from nmigen.lib.coding import PriorityEncoder
  32
  33 # for testing purposes
  34 from soc.experiment.testmem import TestMemory
  35
  36
  37 class PortInterface(RecordObject):
  38     """PortInterface
  39
  40     defines the interface - the API - that the LDSTCompUnit connects
  41     to.  note that this is NOT a "fire-and-forget" interface.  the
  42     LDSTCompUnit *must* be kept appraised that the request is in
  43     progress, and only when it has a 100% successful completion rate
  44     can the notification be given (busy dropped).
  45
  46     The interface FSM rules are as follows:
  47
  48     * if busy_o is asserted, a LD/ST is in progress.  further
  49       requests may not be made until busy_o is deasserted.
  50
  51     * only one of is_ld_i or is_st_i may be asserted.  busy_o
  52       will immediately be asserted and remain asserted.
  53
  54     * addr.ok is to be asserted when the LD/ST address is known.
  55       addr.data is to be valid on the same cycle.
  56
  57       addr.ok and addr.data must REMAIN asserted until busy_o
  58       is de-asserted.  this ensures that there is no need
  59       for the L0 Cache/Buffer to have an additional address latch
  60       (because the LDSTCompUnit already has it)
  61
  62     * addr_ok_o (or addr_exc_o) must be waited for.  these will
  63       be asserted *only* for one cycle and one cycle only.
  64
  65     * addr_exc_o will be asserted if there is no chance that the
  66       memory request may be fulfilled.
  67
  68       busy_o is deasserted on the same cycle as addr_exc_o is asserted.
  69
  70     * conversely: addr_ok_o must *ONLY* be asserted if there is a
  71       HUNDRED PERCENT guarantee that the memory request will be
  72       fulfilled.
  73
  74     * for a LD, ld.ok will be asserted - for only one clock cycle -
  75       at any point in the future that is acceptable to the underlying
  76       Memory subsystem.  the recipient MUST latch ld.data on that cycle.
  77
  78       busy_o is deasserted on the same cycle as ld.ok is asserted.
  79
  80     * for a ST, st.ok may be asserted only after addr_ok_o had been
  81       asserted, alongside valid st.data at the same time.  st.ok
  82       must only be asserted for one cycle.
  83
  84       the underlying Memory is REQUIRED to pick up that data and
  85       guarantee its delivery.  no back-acknowledgement is required.
  86
  87       busy_o is deasserted on the cycle AFTER st.ok is asserted.
  88     """
  89
  90     def __init__(self, name=None, regwid=64, addrwid=48):
  91
  92         self._regwid = regwid
  93         self._addrwid = addrwid
  94
  95         RecordObject.__init__(self, name=name)
  96
  97         # distinguish op type (ld/st)
  98         self.is_ld_i = Signal(reset_less=True)
  99         self.is_st_i = Signal(reset_less=True)
 100         self.op = CompLDSTOpSubset() # hm insn_type ld/st duplicates here
 101
 102         # common signals
 103         self.busy_o = Signal(reset_less=True)     # do not use if busy
 104         self.go_die_i = Signal(reset_less=True)   # back to reset
 105         self.addr = Data(addrwid, "addr_i")            # addr/addr-ok
 106         self.addr_ok_o = Signal(reset_less=True)  # addr is valid (TLB, L1 etc.)
 107         self.addr_exc_o = Signal(reset_less=True) # TODO, "type" of exception
 108
 109         # LD/ST
 110         self.ld = Data(regwid, "ld_data_o") # ok to be set by L0 Cache/Buf
 111         self.st = Data(regwid, "st_data_i") # ok to be set by CompUnit
 112
 113 # TODO:
 114 class DualPortSplitter(Elaboratable):
 115     """DualPortSplitter
 116
 117     * one incoming PortInterface
 118     * two *OUTGOING* PortInterfaces
 119     * uses LDSTSplitter to do it
 120
 121     (actually, thinking about it LDSTSplitter could simply be
 122      modified to conform to PortInterface: one in, two out)
 123
 124     once that is done each pair of ports may be wired directly
 125     to the dual ports of L0CacheBuffer
 126     """
 127     pass
 128
 129
 130 class LDSTPort(Elaboratable):
 131     def __init__(self, idx, regwid=64, addrwid=48):
 132         self.pi = PortInterface("ldst_port%d" % idx, regwid, addrwid)
 133
 134     def elaborate(self, platform):
 135         m = Module()
 136         comb, sync = m.d.comb, m.d.sync
 137
 138         # latches
 139         m.submodules.busy_l = busy_l = SRLatch(False, name="busy")
 140         m.submodules.cyc_l = cyc_l = SRLatch(True, name="cyc")
 141         comb += cyc_l.s.eq(0)
 142         comb += cyc_l.r.eq(0)
 143
 144         # this is a little weird: we let the L0Cache/Buffer set
 145         # the outputs: this module just monitors "state".
 146
 147         # LD/ST requested activates "busy"
 148         with m.If(self.pi.is_ld_i | self.pi.is_st_i):
 149             comb += busy_l.s.eq(1)
 150
 151         # monitor for an exception or the completion of LD.
 152         with m.If(self.pi.addr_exc_o):
 153             comb += busy_l.r.eq(1)
 154
 155         # however ST needs one cycle before busy is reset
 156         with m.If(self.pi.st.ok | self.pi.ld.ok):
 157             comb += cyc_l.s.eq(1)
 158
 159         with m.If(cyc_l.q):
 160             comb += cyc_l.r.eq(1)
 161             comb += busy_l.r.eq(1)
 162
 163         # busy latch outputs to interface
 164         comb += self.pi.busy_o.eq(busy_l.q)
 165
 166         return m
 167
 168     def __iter__(self):
 169         yield self.pi.is_ld_i
 170         yield self.pi.is_st_i
 171         yield from self.pi.op.ports()
 172         yield self.pi.busy_o
 173         yield self.pi.go_die_i
 174         yield from self.pi.addr.ports()
 175         yield self.pi.addr_ok_o
 176         yield self.pi.addr_exc_o
 177
 178         yield from self.pi.ld.ports()
 179         yield from self.pi.st.ports()
 180
 181     def ports(self):
 182         return list(self)
 183
 184
 185 class L0CacheBuffer(Elaboratable):
 186     """L0 Cache / Buffer
 187
 188     Note that the final version will have *two* interfaces per LDSTCompUnit,
 189     to cover mis-aligned requests, as well as *two* 128-bit L1 Cache
 190     interfaces: one for odd (addr[4] == 1) and one for even (addr[4] == 1).
 191
 192     This version is to be used for test purposes (and actively maintained
 193     for such, rather than "replaced")
 194
 195     There are much better ways to implement this.  However it's only
 196     a "demo" / "test" class, and one important aspect: it responds
 197     combinatorially, where a nmigen FSM's state-changes only activate
 198     on clock-sync boundaries.
 199     """
 200     def __init__(self, n_units, mem, regwid=64, addrwid=48):
 201         self.n_units = n_units
 202         self.mem = mem
 203         ul = []
 204         for i in range(n_units):
 205             ul.append(LDSTPort(i, regwid, addrwid))
 206         self.dports = Array(ul)
 207
 208     def elaborate(self, platform):
 209         m = Module()
 210         comb, sync = m.d.comb, m.d.sync
 211
 212         # connect the ports as modules
 213         for i in range(self.n_units):
 214             setattr(m.submodules, "port%d" % i, self.dports[i])
 215
 216         # state-machine latches
 217         m.submodules.st_active = st_active = SRLatch(False, name="st_active")
 218         m.submodules.ld_active = ld_active = SRLatch(False, name="ld_active")
 219         m.submodules.reset_l = reset_l = SRLatch(True, name="reset")
 220         m.submodules.idx_l   = idx_l   = SRLatch(False, name="idx_l")
 221         m.submodules.adrok_l = adrok_l = SRLatch(False, name="addr_acked")
 222
 223         # find one LD (or ST) and do it.  only one per cycle.
 224         # TODO: in the "live" (production) L0Cache/Buffer, merge multiple
 225         # LD/STs using mask-expansion - see LenExpand class
 226
 227         m.submodules.ldpick = ldpick = PriorityEncoder(self.n_units)
 228         m.submodules.stpick = stpick = PriorityEncoder(self.n_units)
 229
 230         lds = Signal(self.n_units, reset_less=True)
 231         sts = Signal(self.n_units, reset_less=True)
 232         ldi = []
 233         sti = []
 234         for i in range(self.n_units):
 235             pi = self.dports[i].pi
 236             ldi.append(pi.is_ld_i & pi.busy_o) # accumulate ld-req signals
 237             sti.append(pi.is_st_i & pi.busy_o) # accumulate st-req signals
 238         # put the requests into the priority-pickers
 239         comb += ldpick.i.eq(Cat(*ldi))
 240         comb += stpick.i.eq(Cat(*sti))
 241
 242         # hmm, have to select (record) the right port index
 243         nbits = log2_int(self.n_units, False)
 244         ld_idx = Signal(nbits, reset_less=False)
 245         st_idx = Signal(nbits, reset_less=False)
 246         # use these because of the sync-and-comb pass-through capability
 247         latchregister(m, ldpick.o, ld_idx, idx_l.qn, name="ld_idx")
 248         latchregister(m, stpick.o, st_idx, idx_l.qn, name="st_idx")
 249
 250         # convenience variables to reference the "picked" port
 251         ldport = self.dports[ld_idx].pi
 252         stport = self.dports[st_idx].pi
 253         # and the memory ports
 254         rdport = self.mem.rdport
 255         wrport = self.mem.wrport
 256
 257         # Priority-Pickers pick one and only one request, capture its index.
 258         # from that point on this code *only* "listens" to that port.
 259
 260         sync += adrok_l.s.eq(0)
 261         comb += adrok_l.r.eq(0)
 262         with m.If(~ldpick.n):
 263             comb += ld_active.s.eq(1) # activate LD mode
 264             comb += idx_l.r.eq(1)  # pick (and capture) the port index
 265         with m.Elif(~stpick.n):
 266             comb += st_active.s.eq(1) # activate ST mode
 267             comb += idx_l.r.eq(1)  # pick (and capture) the port index
 268
 269         # from this point onwards, with the port "picked", it stays picked
 270         # until ld_active (or st_active) are de-asserted.
 271
 272         # if now in "LD" mode: wait for addr_ok, then send the address out
 273         # to memory, acknowledge address, and send out LD data
 274         with m.If(ld_active.q):
 275             with m.If(ldport.addr.ok):
 276                 comb += rdport.addr.eq(ldport.addr.data) # addr ok, send thru
 277                 with m.If(adrok_l.qn):
 278                     comb += ldport.addr_ok_o.eq(1) # acknowledge addr ok
 279                     sync += adrok_l.s.eq(1)       # and pull "ack" latch
 280
 281         # if now in "ST" mode: likewise do the same but with "ST"
 282         # to memory, acknowledge address, and send out LD data
 283         with m.If(st_active.q):
 284             with m.If(stport.addr.ok):
 285                 comb += wrport.addr.eq(stport.addr.data) # addr ok, send thru
 286                 with m.If(adrok_l.qn):
 287                     comb += stport.addr_ok_o.eq(1) # acknowledge addr ok
 288                     sync += adrok_l.s.eq(1)       # and pull "ack" latch
 289
 290         # NOTE: in both these, below, the port itself takes care
 291         # of de-asserting its "busy_o" signal, based on either ld.ok going
 292         # high (by us, here) or by st.ok going high (by the LDSTCompUnit).
 293
 294         # for LD mode, when addr has been "ok'd", assume that (because this
 295         # is a "Memory" test-class) the memory read data is valid.
 296         comb += reset_l.s.eq(0)
 297         comb += reset_l.r.eq(0)
 298         with m.If(ld_active.q & adrok_l.q):
 299             comb += ldport.ld.data.eq(rdport.data) # put data out
 300             comb += ldport.ld.ok.eq(1)             # indicate data valid
 301             comb += reset_l.s.eq(1)   # reset mode after 1 cycle
 302
 303         # for ST mode, when addr has been "ok'd", wait for incoming "ST ok"
 304         with m.If(st_active.q & stport.st.ok):
 305             comb += wrport.data.eq(stport.st.data) # write st to mem
 306             comb += wrport.en.eq(1)                # enable write
 307             comb += reset_l.s.eq(1)   # reset mode after 1 cycle
 308
 309         # after waiting one cycle (reset_l is "sync" mode), reset the port
 310         with m.If(reset_l.q):
 311             comb += idx_l.s.eq(1)  # deactivate port-index selector
 312             comb += ld_active.r.eq(1)   # leave the ST active for 1 cycle
 313             comb += st_active.r.eq(1)   # leave the ST active for 1 cycle
 314             comb += reset_l.r.eq(1)     # clear reset
 315             comb += adrok_l.r.eq(1)     # address reset
 316
 317         return m
 318
 319     def ports(self):
 320         for p in self.dports:
 321             yield from p.ports()
 322
 323
 324 class TstL0CacheBuffer(Elaboratable):
 325     def __init__(self, n_units=3, regwid=16, addrwid=4):
 326         self.mem = TestMemory(regwid, addrwid)
 327         self.l0 = L0CacheBuffer(n_units, self.mem, regwid, addrwid)
 328
 329     def elaborate(self, platform):
 330         m = Module()
 331         m.submodules.mem = self.mem
 332         m.submodules.l0 = self.l0
 333
 334         return m
 335
 336     def ports(self):
 337         yield from self.l0.ports()
 338         yield self.mem.rdport.addr
 339         yield self.mem.rdport.data
 340         yield self.mem.wrport.addr
 341         yield self.mem.wrport.data
 342         # TODO: mem ports
 343
 344
 345 def wait_busy(port, no=False):
 346     while True:
 347         busy = yield port.pi.busy_o
 348         print ("busy", no, busy)
 349         if bool(busy) == no:
 350             break
 351         yield
 352
 353
 354 def wait_addr(port):
 355     while True:
 356         addr_ok = yield port.pi.addr_ok_o
 357         print ("addrok", addr_ok)
 358         if not addr_ok:
 359             break
 360         yield
 361
 362
 363 def wait_ldok(port):
 364     while True:
 365         ldok = yield port.pi.ld.ok
 366         print ("ldok", ldok)
 367         if ldok:
 368             break
 369         yield
 370
 371
 372 def l0_cache_st(dut, addr, data):
 373     l0 = dut.l0
 374     mem = dut.mem
 375     port0 = l0.dports[0]
 376     port1 = l0.dports[1]
 377
 378     # have to wait until not busy
 379     yield from wait_busy(port1, no=False)    # wait until not busy
 380
 381     # set up a ST on the port.  address first:
 382     yield port1.pi.is_st_i.eq(1) # indicate LD
 383
 384     yield port1.pi.addr.data.eq(addr) # set address
 385     yield port1.pi.addr.ok.eq(1) # set ok
 386     yield from wait_addr(port1)             # wait until addr ok
 387     #yield # not needed, just for checking
 388     #yield # not needed, just for checking
 389     # assert "ST" for one cycle (required by the API)
 390     yield port1.pi.st.data.eq(data)
 391     yield port1.pi.st.ok.eq(1)
 392     yield
 393     yield port1.pi.st.ok.eq(0)
 394
 395     # can go straight to reset.
 396     yield port1.pi.is_st_i.eq(0) #end
 397     yield port1.pi.addr.ok.eq(0) # set !ok
 398     #yield from wait_busy(port1, False)    # wait until not busy
 399
 400
 401 def l0_cache_ld(dut, addr, expected):
 402
 403     l0 = dut.l0
 404     mem = dut.mem
 405     port0 = l0.dports[0]
 406     port1 = l0.dports[1]
 407
 408     # have to wait until not busy
 409     yield from wait_busy(port1, no=False)    # wait until not busy
 410
 411     # set up a LD on the port.  address first:
 412     yield port1.pi.is_ld_i.eq(1) # indicate LD
 413
 414     yield port1.pi.addr.data.eq(addr) # set address
 415     yield port1.pi.addr.ok.eq(1) # set ok
 416     yield from wait_addr(port1)             # wait until addr ok
 417
 418     yield from wait_ldok(port1)             # wait until ld ok
 419     data = yield port1.pi.ld.data
 420
 421     # cleanup
 422     yield port1.pi.is_ld_i.eq(0) #end
 423     yield port1.pi.addr.ok.eq(0) # set !ok
 424     #yield from wait_busy(port1, no=False)    # wait until not busy
 425
 426     return data
 427
 428
 429 def l0_cache_ldst(dut):
 430     yield
 431     addr = 0x2
 432     data = 0xbeef
 433     data2 = 0xf00f
 434     #data = 0x4
 435     yield from l0_cache_st(dut, 0x2, data)
 436     yield from l0_cache_st(dut, 0x3, data2)
 437     result = yield from l0_cache_ld(dut, 0x2, data)
 438     result2 = yield from l0_cache_ld(dut, 0x3, data2)
 439     yield
 440     assert data == result, "data %x != %x" % (result, data)
 441     assert data2 == result2, "data2 %x != %x" % (result2, data2)
 442
 443
 444 def test_l0_cache():
 445
 446     dut = TstL0CacheBuffer()
 447     vl = rtlil.convert(dut, ports=dut.ports())
 448     with open("test_basic_l0_cache.il", "w") as f:
 449         f.write(vl)
 450
 451     run_simulation(dut, l0_cache_ldst(dut),
 452                    vcd_name='test_l0_cache_basic.vcd')
 453
 454
 455 if __name__ == '__main__':
 456     test_l0_cache()