gram/phy/ecp5ddrphy.py

   1 # This file is Copyright (c) 2019 David Shah <dave@ds0.me>
   2 # This file is Copyright (c) 2019-2020 Florent Kermarrec <florent@enjoy-digital.fr>
   3 # This file is Copyright (c) 2020 LambdaConcept <contact@lambdaconcept.com>
   4 # License: BSD
   5
   6 # 1:2 frequency-ratio DDR3 PHY for Lattice's ECP5
   7 # DDR3: 800 MT/s
   8
   9 import math
  10
  11 from nmigen import *
  12 from nmigen.lib.cdc import FFSynchronizer
  13 from nmigen.utils import log2_int
  14
  15 from lambdasoc.periph import Peripheral
  16
  17 import gram.stream as stream
  18 from gram.common import *
  19 from gram.phy.dfi import Interface
  20 from gram.compat import Timeline
  21
  22 # Lattice ECP5 DDR PHY Initialization --------------------------------------------------------------
  23
  24
  25 class ECP5DDRPHYInit(Elaboratable):
  26     def __init__(self):
  27         self.pause = Signal()
  28         self.stop = Signal()
  29         self.delay = Signal()
  30         self.reset = Signal()
  31
  32     def elaborate(self, platform):
  33         m = Module()
  34
  35         new_lock = Signal()
  36         update = Signal()
  37         freeze = Signal()
  38
  39         # DDRDLLA instance -------------------------------------------------------------------------
  40         _lock = Signal()
  41         delay = Signal()
  42         m.submodules += Instance("DDRDLLA",
  43             i_CLK=ClockSignal("sync2x"),
  44             i_RST=ResetSignal("init"),
  45             i_UDDCNTLN=~update,
  46             i_FREEZE=freeze,
  47             o_DDRDEL=delay,
  48             o_LOCK=_lock)
  49         lock = Signal()
  50         lock_d = Signal()
  51         m.submodules += FFSynchronizer(_lock, lock, o_domain="init")
  52         m.d.init += lock_d.eq(lock)
  53         m.d.sync += new_lock.eq(lock & ~lock_d)
  54
  55         # DDRDLLA/DDQBUFM/ECLK initialization sequence ---------------------------------------------
  56         t = 8  # in cycles
  57         tl = Timeline([
  58             (1*t,  [freeze.eq(1)]),  # Freeze DDRDLLA
  59             (2*t,  [self.stop.eq(1)]),   # Stop ECLK domain
  60             (3*t,  [self.reset.eq(1)]),  # Reset ECLK domain
  61             (4*t,  [self.reset.eq(0)]),  # Release ECLK domain reset
  62             (5*t,  [self.stop.eq(0)]),   # Release ECLK domain stop
  63             (6*t,  [freeze.eq(0)]),  # Release DDRDLLA freeze
  64             (7*t,  [self.pause.eq(1)]),  # Pause DQSBUFM
  65             (8*t,  [update.eq(1)]),  # Update DDRDLLA
  66             (9*t,  [update.eq(0)]),  # Release DDRDMMA update
  67             (10*t, [self.pause.eq(0)]),  # Release DQSBUFM pause
  68         ])
  69         m.submodules += tl
  70         # Wait DDRDLLA Lock
  71         m.d.comb += tl.trigger.eq(new_lock)
  72
  73         m.d.comb += self.delay.eq(delay)
  74
  75         return m
  76
  77 # Lattice ECP5 DDR PHY -----------------------------------------------------------------------------
  78
  79
  80 class ECP5DDRPHY(Peripheral, Elaboratable):
  81     def __init__(self, pads, sys_clk_freq=100e6):
  82         super().__init__(name="phy")
  83
  84         self.pads = pads
  85         self._sys_clk_freq = sys_clk_freq
  86
  87         databits = len(self.pads.dq.io)
  88         assert databits % 8 == 0
  89
  90         # CSR
  91         bank = self.csr_bank()
  92
  93         self.burstdet = bank.csr(databits//8, "rw")
  94
  95         self.rdly = []
  96         self.rdly += [bank.csr(3, "rw", name="rdly_p0")]
  97         self.rdly += [bank.csr(3, "rw", name="rdly_p1")]
  98
  99         self._bridge = self.bridge(data_width=32, granularity=8, alignment=2)
 100         self.bus = self._bridge.bus
 101
 102         addressbits = len(self.pads.a.o0)
 103         bankbits = len(self.pads.ba.o0)
 104         nranks = 1 if not hasattr(self.pads, "cs") else len(self.pads.cs.o0)
 105         databits = len(self.pads.dq.io)
 106         self.dfi = Interface(addressbits, bankbits, nranks, 4*databits, 4)
 107
 108         # PHY settings -----------------------------------------------------------------------------
 109         tck = 1/(2*self._sys_clk_freq)
 110         nphases = 2
 111         databits = len(self.pads.dq.io)
 112         nranks = 1 if not hasattr(self.pads, "cs") else len(self.pads.cs.o0)
 113         cl, cwl = get_cl_cw("DDR3", tck)
 114         cl_sys_latency = get_sys_latency(nphases, cl)
 115         cwl_sys_latency = get_sys_latency(nphases, cwl)
 116         rdcmdphase, rdphase = get_sys_phases(nphases, cl_sys_latency, cl)
 117         wrcmdphase, wrphase = get_sys_phases(nphases, cwl_sys_latency, cwl)
 118         self.settings = PhySettings(
 119             phytype="ECP5DDRPHY",
 120             memtype="DDR3",
 121             databits=databits,
 122             dfi_databits=4*databits,
 123             nranks=nranks,
 124             nphases=nphases,
 125             rdphase=rdphase,
 126             wrphase=wrphase,
 127             rdcmdphase=rdcmdphase,
 128             wrcmdphase=wrcmdphase,
 129             cl=cl,
 130             cwl=cwl,
 131             read_latency=2 + cl_sys_latency + 2 + log2_int(4//nphases) + 4,
 132             write_latency=cwl_sys_latency
 133         )
 134
 135     def elaborate(self, platform):
 136         m = Module()
 137
 138         m.submodules.bridge = self._bridge
 139
 140         tck = 1/(2*self._sys_clk_freq)
 141         nphases = 2
 142         databits = len(self.pads.dq.io)
 143
 144         burstdet_reg = Signal(databits//8)
 145         m.d.comb += self.burstdet.r_data.eq(burstdet_reg)
 146
 147         # Burstdet clear
 148         with m.If(self.burstdet.w_stb):
 149             m.d.sync += burstdet_reg.eq(0)
 150
 151         # Init -------------------------------------------------------------------------------------
 152         m.submodules.init = init = ECP5DDRPHYInit()
 153
 154         # Parameters -------------------------------------------------------------------------------
 155         cl, cwl = get_cl_cw("DDR3", tck)
 156         cl_sys_latency = get_sys_latency(nphases, cl)
 157         cwl_sys_latency = get_sys_latency(nphases, cwl)
 158
 159         # DFI Interface ----------------------------------------------------------------------------
 160         dfi = self.dfi
 161
 162         bl8_chunk = Signal()
 163
 164         # Clock --------------------------------------------------------------------------------
 165         m.d.comb += [
 166             self.pads.clk.o_clk.eq(ClockSignal("dramsync")),
 167             self.pads.clk.o_fclk.eq(ClockSignal("sync2x")),
 168         ]
 169         for i in range(len(self.pads.clk.o0)):
 170             m.d.comb += [
 171                 self.pads.clk.o0[i].eq(0),
 172                 self.pads.clk.o1[i].eq(1),
 173                 self.pads.clk.o2[i].eq(0),
 174                 self.pads.clk.o3[i].eq(1),
 175             ]
 176
 177         # Addresses and Commands ---------------------------------------------------------------
 178         m.d.comb += [
 179             self.pads.a.o_clk.eq(ClockSignal("dramsync")),
 180             self.pads.a.o_fclk.eq(ClockSignal("sync2x")),
 181             self.pads.ba.o_clk.eq(ClockSignal("dramsync")),
 182             self.pads.ba.o_fclk.eq(ClockSignal("sync2x")),
 183         ]
 184         for i in range(len(self.pads.a.o0)):
 185             m.d.comb += [
 186                 self.pads.a.o0[i].eq(dfi.phases[0].address[i]),
 187                 self.pads.a.o1[i].eq(dfi.phases[0].address[i]),
 188                 self.pads.a.o2[i].eq(dfi.phases[1].address[i]),
 189                 self.pads.a.o3[i].eq(dfi.phases[1].address[i]),
 190             ]
 191         for i in range(len(self.pads.ba.o0)):
 192             m.d.comb += [
 193                 self.pads.ba.o0[i].eq(dfi.phases[0].bank[i]),
 194                 self.pads.ba.o1[i].eq(dfi.phases[0].bank[i]),
 195                 self.pads.ba.o2[i].eq(dfi.phases[1].bank[i]),
 196                 self.pads.ba.o3[i].eq(dfi.phases[1].bank[i]),
 197             ]
 198
 199         # Control pins
 200         controls = ["ras", "cas", "we", "clk_en", "odt"]
 201         if hasattr(self.pads, "reset"):
 202             controls.append("reset")
 203         if hasattr(self.pads, "cs"):
 204             controls.append("cs")
 205         for name in controls:
 206             m.d.comb += [
 207                 getattr(self.pads, name).o_clk.eq(ClockSignal("dramsync")),
 208                 getattr(self.pads, name).o_fclk.eq(ClockSignal("sync2x")),
 209             ]
 210             for i in range(len(getattr(self.pads, name).o0)):
 211                 m.d.comb += [
 212                     getattr(self.pads, name).o0[i].eq(getattr(dfi.phases[0], name)[i]),
 213                     getattr(self.pads, name).o1[i].eq(getattr(dfi.phases[0], name)[i]),
 214                     getattr(self.pads, name).o2[i].eq(getattr(dfi.phases[1], name)[i]),
 215                     getattr(self.pads, name).o3[i].eq(getattr(dfi.phases[1], name)[i]),
 216                 ]
 217
 218         # DQ ---------------------------------------------------------------------------------------
 219         dq_oe = Signal()
 220         dqs_re = Signal()
 221         dqs_oe = Signal()
 222         dqs_postamble = Signal()
 223         dqs_preamble = Signal()
 224         for i in range(databits//8):
 225             # DQSBUFM
 226             dqs_i = Signal()
 227             dqsr90 = Signal()
 228             dqsw270 = Signal()
 229             dqsw = Signal()
 230             rdpntr = Signal(3)
 231             wrpntr = Signal(3)
 232             burstdet = Signal()
 233
 234             m.submodules += Instance("DQSBUFM",
 235                 p_DQS_LI_DEL_ADJ="MINUS",
 236                 p_DQS_LI_DEL_VAL=1,
 237                 p_DQS_LO_DEL_ADJ="MINUS",
 238                 p_DQS_LO_DEL_VAL=4,
 239
 240                 # Delay
 241                 i_DYNDELAY0=0,
 242                 i_DYNDELAY1=0,
 243                 i_DYNDELAY2=0,
 244                 i_DYNDELAY3=0,
 245                 i_DYNDELAY4=0,
 246                 i_DYNDELAY5=0,
 247                 i_DYNDELAY6=0,
 248                 i_DYNDELAY7=0,
 249
 250                 # Clocks / Reset
 251                 i_SCLK=ClockSignal("sync"),
 252                 i_ECLK=ClockSignal("sync2x"),
 253                 i_RST=ResetSignal("dramsync"),
 254                 i_DDRDEL=init.delay,
 255                 i_PAUSE=init.pause | self.rdly[i].w_stb,
 256
 257                 # Control
 258                 # Assert LOADNs to use DDRDEL control
 259                 i_RDLOADN=0,
 260                 i_RDMOVE=0,
 261                 i_RDDIRECTION=1,
 262                 i_WRLOADN=0,
 263                 i_WRMOVE=0,
 264                 i_WRDIRECTION=1,
 265
 266                 # Reads (generate shifted DQS clock for reads)
 267                 i_READ0=dqs_re,
 268                 i_READ1=dqs_re,
 269                 i_READCLKSEL0=self.rdly[i].w_data[0],
 270                 i_READCLKSEL1=self.rdly[i].w_data[1],
 271                 i_READCLKSEL2=self.rdly[i].w_data[2],
 272                 i_DQSI=dqs_i,
 273                 o_DQSR90=dqsr90,
 274                 o_RDPNTR0=rdpntr[0],
 275                 o_RDPNTR1=rdpntr[1],
 276                 o_RDPNTR2=rdpntr[2],
 277                 o_WRPNTR0=wrpntr[0],
 278                 o_WRPNTR1=wrpntr[1],
 279                 o_WRPNTR2=wrpntr[2],
 280                 o_BURSTDET=burstdet,
 281
 282                 # Writes (generate shifted ECLK clock for writes)
 283                 o_DQSW270=dqsw270,
 284                 o_DQSW=dqsw)
 285
 286             with m.If(burstdet):
 287                 m.d.sync += burstdet_reg[i].eq(1)
 288
 289             # DQS and DM ---------------------------------------------------------------------------
 290             dm_o_data = Signal(8)
 291             dm_o_data_d = Signal(8)
 292             dm_o_data_muxed = Signal(4)
 293             m.d.comb += dm_o_data.eq(Cat(
 294                 dfi.phases[0].wrdata_mask[0*databits//8+i],
 295                 dfi.phases[0].wrdata_mask[1*databits//8+i],
 296                 dfi.phases[0].wrdata_mask[2*databits//8+i],
 297                 dfi.phases[0].wrdata_mask[3*databits//8+i],
 298
 299                 dfi.phases[1].wrdata_mask[0*databits//8+i],
 300                 dfi.phases[1].wrdata_mask[1*databits//8+i],
 301                 dfi.phases[1].wrdata_mask[2*databits//8+i],
 302                 dfi.phases[1].wrdata_mask[3*databits//8+i]),
 303             )
 304             m.d.sync += dm_o_data_d.eq(dm_o_data)
 305
 306             with m.If(bl8_chunk):
 307                 m.d.sync += dm_o_data_muxed.eq(dm_o_data_d[4:])
 308             with m.Else():
 309                 m.d.sync += dm_o_data_muxed.eq(dm_o_data[:4])
 310
 311             m.submodules += Instance("ODDRX2DQA",
 312                 i_RST=ResetSignal("dramsync"),
 313                 i_ECLK=ClockSignal("sync2x"),
 314                 i_SCLK=ClockSignal("dramsync"),
 315                 i_DQSW270=dqsw270,
 316                 i_D0=dm_o_data_muxed[0],
 317                 i_D1=dm_o_data_muxed[1],
 318                 i_D2=dm_o_data_muxed[2],
 319                 i_D3=dm_o_data_muxed[3],
 320                 o_Q=self.pads.dm.o[i])
 321
 322             dqs = Signal()
 323             dqs_oe_n = Signal()
 324             m.submodules += [
 325                 Instance("ODDRX2DQSB",
 326                     i_RST=ResetSignal("dramsync"),
 327                     i_ECLK=ClockSignal("sync2x"),
 328                     i_SCLK=ClockSignal(),
 329                     i_DQSW=dqsw,
 330                     i_D0=0,
 331                     i_D1=1,
 332                     i_D2=0,
 333                     i_D3=1,
 334                     o_Q=dqs),
 335                 Instance("TSHX2DQSA",
 336                     i_RST=ResetSignal("dramsync"),
 337                     i_ECLK=ClockSignal("sync2x"),
 338                     i_SCLK=ClockSignal(),
 339                     i_DQSW=dqsw,
 340                     i_T0=~(dqs_oe | dqs_postamble),
 341                     i_T1=~(dqs_oe | dqs_preamble),
 342                     o_Q=dqs_oe_n),
 343                 Instance("BB",
 344                     i_I=dqs,
 345                     i_T=dqs_oe_n,
 346                     o_O=dqs_i,
 347                     io_B=self.pads.dqs.p[i]),
 348             ]
 349
 350             for j in range(8*i, 8*(i+1)):
 351                 dq_o = Signal()
 352                 dq_i = Signal()
 353                 dq_oe_n = Signal()
 354                 dq_i_delayed = Signal()
 355                 dq_i_data = Signal(4)
 356                 dq_o_data = Signal(8)
 357                 dq_o_data_d = Signal(8)
 358                 dq_o_data_muxed = Signal(4)
 359                 m.d.comb += dq_o_data.eq(Cat(
 360                     dfi.phases[0].wrdata[0*databits+j],
 361                     dfi.phases[0].wrdata[1*databits+j],
 362                     dfi.phases[0].wrdata[2*databits+j],
 363                     dfi.phases[0].wrdata[3*databits+j],
 364                     dfi.phases[1].wrdata[0*databits+j],
 365                     dfi.phases[1].wrdata[1*databits+j],
 366                     dfi.phases[1].wrdata[2*databits+j],
 367                     dfi.phases[1].wrdata[3*databits+j])
 368                 )
 369
 370                 m.d.sync += dq_o_data_d.eq(dq_o_data)
 371                 with m.If(bl8_chunk):
 372                     m.d.sync += dq_o_data_muxed.eq(dq_o_data_d[4:])
 373                 with m.Else():
 374                     m.d.sync += dq_o_data_muxed.eq(dq_o_data[:4])
 375
 376                 m.submodules += [
 377                     Instance("ODDRX2DQA",
 378                         i_RST=ResetSignal("dramsync"),
 379                         i_ECLK=ClockSignal("sync2x"),
 380                         i_SCLK=ClockSignal(),
 381                         i_DQSW270=dqsw270,
 382                         i_D0=dq_o_data_muxed[0],
 383                         i_D1=dq_o_data_muxed[1],
 384                         i_D2=dq_o_data_muxed[2],
 385                         i_D3=dq_o_data_muxed[3],
 386                         o_Q=dq_o),
 387                     Instance("DELAYF",
 388                         p_DEL_MODE="DQS_ALIGNED_X2",
 389                         i_LOADN=1,
 390                         i_MOVE=0,
 391                         i_DIRECTION=0,
 392                         i_A=dq_i,
 393                         o_Z=dq_i_delayed),
 394                     Instance("IDDRX2DQA",
 395                         i_RST=ResetSignal("dramsync"),
 396                         i_ECLK=ClockSignal("sync2x"),
 397                         i_SCLK=ClockSignal(),
 398                         i_DQSR90=dqsr90,
 399                         i_RDPNTR0=rdpntr[0],
 400                         i_RDPNTR1=rdpntr[1],
 401                         i_RDPNTR2=rdpntr[2],
 402                         i_WRPNTR0=wrpntr[0],
 403                         i_WRPNTR1=wrpntr[1],
 404                         i_WRPNTR2=wrpntr[2],
 405                         i_D=dq_i_delayed,
 406                         o_Q0=dq_i_data[0],
 407                         o_Q1=dq_i_data[1],
 408                         o_Q2=dq_i_data[2],
 409                         o_Q3=dq_i_data[3]),
 410                 ]
 411                 m.submodules += [
 412                     Instance("TSHX2DQA",
 413                         i_RST=ResetSignal("dramsync"),
 414                         i_ECLK=ClockSignal("sync2x"),
 415                         i_SCLK=ClockSignal(),
 416                         i_DQSW270=dqsw270,
 417                         i_T0=~dq_oe,
 418                         i_T1=~dq_oe,
 419                         o_Q=dq_oe_n),
 420                     Instance("BB",
 421                         i_I=dq_o,
 422                         i_T=dq_oe_n,
 423                         o_O=dq_i,
 424                         io_B=self.pads.dq.io[j])
 425                 ]
 426                 m.d.sync += [
 427                     dfi.phases[1].rddata[j].eq(dq_i_data[0]),
 428                     dfi.phases[1].rddata[1*databits+j].eq(dq_i_data[1]),
 429                     dfi.phases[1].rddata[2*databits+j].eq(dq_i_data[2]),
 430                     dfi.phases[1].rddata[3*databits+j].eq(dq_i_data[3]),
 431                 ]
 432         m.d.sync += [
 433             dfi.phases[0].rddata.eq(dfi.phases[1].rddata),
 434         ]
 435
 436         # Read Control Path ------------------------------------------------------------------------
 437         # Creates a shift register of read commands coming from the DFI interface. This shift register
 438         # is used to control DQS read (internal read pulse of the DQSBUF) and to indicate to the
 439         # DFI interface that the read data is valid.
 440         #
 441         # The DQS read must be asserted for 2 sys_clk cycles before the read data is coming back from
 442         # the DRAM (see 6.2.4 READ Pulse Positioning Optimization of FPGA-TN-02035-1.2)
 443         #
 444         # The read data valid is asserted for 1 sys_clk cycle when the data is available on the DFI
 445         # interface, the latency is the sum of the ODDRX2DQA, CAS, IDDRX2DQA latencies.
 446         rddata_en = Signal(self.settings.read_latency)
 447         rddata_en_last = Signal.like(rddata_en)
 448         m.d.comb += rddata_en.eq(Cat(dfi.phases[self.settings.rdphase].rddata_en, rddata_en_last))
 449         m.d.sync += rddata_en_last.eq(rddata_en)
 450         m.d.sync += [phase.rddata_valid.eq(rddata_en[-1]) for phase in dfi.phases]
 451         m.d.comb += dqs_re.eq(rddata_en[cl_sys_latency + 1] | rddata_en[cl_sys_latency + 2])
 452
 453         # Write Control Path -----------------------------------------------------------------------
 454         # Creates a shift register of write commands coming from the DFI interface. This shift register
 455         # is used to control DQ/DQS tristates and to select write data of the DRAM burst from the DFI
 456         # interface: The PHY is operating in halfrate mode (so provide 4 datas every sys_clk cycles:
 457         # 2x for DDR, 2x for halfrate) but DDR3 requires a burst of 8 datas (BL8) for best efficiency.
 458         # Writes are then performed in 2 sys_clk cycles and data needs to be selected for each cycle.
 459         # FIXME: understand +2
 460         wrdata_en = Signal(cwl_sys_latency + 4)
 461         wrdata_en_last = Signal.like(wrdata_en)
 462         m.d.comb += wrdata_en.eq(Cat(dfi.phases[self.settings.wrphase].wrdata_en, wrdata_en_last))
 463         m.d.sync += wrdata_en_last.eq(wrdata_en)
 464         m.d.comb += dq_oe.eq(wrdata_en[cwl_sys_latency + 1] | wrdata_en[cwl_sys_latency + 2])
 465         m.d.comb += bl8_chunk.eq(wrdata_en[cwl_sys_latency + 1])
 466         m.d.comb += dqs_oe.eq(dq_oe)
 467
 468         # Write DQS Postamble/Preamble Control Path ------------------------------------------------
 469         # Generates DQS Preamble 1 cycle before the first write and Postamble 1 cycle after the last
 470         # write. During writes, DQS tristate is configured as output for at least 4 sys_clk cycles:
 471         # 1 for Preamble, 2 for the Write and 1 for the Postamble.
 472         m.d.comb += dqs_preamble.eq(wrdata_en[cwl_sys_latency + 0] & ~wrdata_en[cwl_sys_latency + 1])
 473         m.d.comb += dqs_postamble.eq(wrdata_en[cwl_sys_latency + 3] & ~wrdata_en[cwl_sys_latency + 2])
 474
 475         return m