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.hdl.ast import Rose
  13 from nmigen.lib.cdc import FFSynchronizer
  14 from nmigen.utils import log2_int
  15
  16 from lambdasoc.periph import Peripheral
  17
  18 from gram.common import *
  19 from gram.phy.dfi import Interface
  20 from gram.compat import Timeline
  21
  22 __all__ = ["ECP5DDRPHY"]
  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         update = Signal()
  36         freeze = Signal()
  37
  38         # DDRDLLA instance -------------------------------------------------------------------------
  39         _lock = Signal()
  40         lock = Signal()
  41         lock_d = 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=self.delay,
  48             o_LOCK=_lock)
  49         m.submodules += FFSynchronizer(_lock, lock, o_domain="init")
  50         m.d.init += lock_d.eq(lock)
  51
  52         # DDRDLLA/DDQBUFM/ECLK initialization sequence ---------------------------------------------
  53         t = 8  # in cycles
  54         tl = Timeline([
  55             (1*t,  [freeze.eq(1)]),  # Freeze DDRDLLA
  56             (2*t,  [self.stop.eq(1)]),   # Stop ECLK domain
  57             (3*t,  [self.reset.eq(1)]),  # Reset ECLK domain
  58             (4*t,  [self.reset.eq(0)]),  # Release ECLK domain reset
  59             (5*t,  [self.stop.eq(0)]),   # Release ECLK domain stop
  60             (6*t,  [freeze.eq(0)]),  # Release DDRDLLA freeze
  61             (7*t,  [self.pause.eq(1)]),  # Pause DQSBUFM
  62             (8*t,  [update.eq(1)]),  # Update DDRDLLA
  63             (9*t,  [update.eq(0)]),  # Release DDRDMMA update
  64             (10*t, [self.pause.eq(0)]),  # Release DQSBUFM pause
  65         ])
  66         m.d.comb += tl.trigger.eq(lock & ~lock_d) # Trigger timeline on lock rising edge
  67         m.submodules += DomainRenamer("init")(tl)
  68
  69         return m
  70
  71
  72 class _DQSBUFMSettingManager(Elaboratable):
  73     """DQSBUFM setting manager.
  74
  75     The DQSBUFM primitive requires a very basic sequence when updating
  76     read delay or other parameters. This elaboratable generates this
  77     sequence from CSR events.
  78
  79     Parameters
  80     ----------
  81     rdly_slr : CSR
  82         CSR storing the rdly value.
  83
  84     Attributes
  85     ----------
  86     pause : Signal(), out
  87         Pause signal for DQSBUFM.
  88     readclksel : Signal(3), out
  89         Readclksel signal for DQSBUFM.
  90     """
  91     def __init__(self, rdly_csr):
  92         self.rdly_csr = rdly_csr
  93
  94         self.pause = Signal()
  95         self.readclksel = Signal(3)
  96
  97     def elaborate(self, platform):
  98         m = Module()
  99
 100         with m.FSM():
 101             with m.State("Idle"):
 102                 with m.If(self.rdly_csr.w_stb):
 103                     m.d.sync += self.pause.eq(1)
 104                     m.next = "RdlyUpdateRequested"
 105
 106             with m.State("RdlyUpdateRequested"):
 107                 m.d.sync += self.readclksel.eq(self.rdly_csr.w_data)
 108                 m.next = "ResetPause"
 109
 110             with m.State("ResetPause"):
 111                 m.d.sync += self.pause.eq(0)
 112                 m.next = "Idle"
 113
 114         return m
 115
 116
 117 class ECP5DDRPHY(Peripheral, Elaboratable):
 118     def __init__(self, pads, sys_clk_freq=100e6):
 119         super().__init__(name="phy")
 120
 121         self.pads = pads
 122         self._sys_clk_freq = sys_clk_freq
 123
 124         databits = len(self.pads.dq.io)
 125         if databits % 8 != 0:
 126             raise ValueError("DQ pads should come in a multiple of 8")
 127
 128         # CSR
 129         bank = self.csr_bank()
 130
 131         self.burstdet = bank.csr(databits//8, "rw")
 132
 133         self.rdly = []
 134         self.rdly += [bank.csr(3, "rw", name="rdly_p0")]
 135         self.rdly += [bank.csr(3, "rw", name="rdly_p1")]
 136
 137         self._bridge = self.bridge(data_width=32, granularity=8, alignment=2)
 138         self.bus = self._bridge.bus
 139
 140         addressbits = len(self.pads.a.o0)
 141         bankbits = len(self.pads.ba.o0)
 142         nranks = 1
 143         if hasattr(self.pads, "cs_n") and hasattr(self.pads.cs_n, "o0"):
 144             nranks = len(self.pads.cs_n.o0)
 145         databits = len(self.pads.dq.io)
 146         self.dfi = Interface(addressbits, bankbits, nranks, 4*databits, 4,
 147                              name="ecp5phy")
 148
 149         # PHY settings -----------------------------------------------------------------------------
 150         tck = 1/(2*self._sys_clk_freq)
 151         nphases = 2
 152         databits = len(self.pads.dq.io)
 153         cl, cwl = get_cl_cw("DDR3", tck)
 154         cl_sys_latency = get_sys_latency(nphases, cl)
 155         cwl_sys_latency = get_sys_latency(nphases, cwl)
 156         rdcmdphase, rdphase = get_sys_phases(nphases, cl_sys_latency, cl)
 157         wrcmdphase, wrphase = get_sys_phases(nphases, cwl_sys_latency, cwl)
 158         self.settings = PhySettings(
 159             phytype="ECP5DDRPHY",
 160             memtype="DDR3",
 161             databits=databits,
 162             dfi_databits=4*databits,
 163             nranks=nranks,
 164             nphases=nphases,
 165             rdphase=rdphase,
 166             wrphase=wrphase,
 167             rdcmdphase=rdcmdphase,
 168             wrcmdphase=wrcmdphase,
 169             cl=cl,
 170             cwl=cwl,
 171             read_latency=2 + cl_sys_latency + 2 + log2_int(4//nphases) + 4,
 172             write_latency=cwl_sys_latency
 173         )
 174
 175     def elaborate(self, platform):
 176         m = Module()
 177         comb, sync = m.d.comb, m.d.sync
 178
 179         m.submodules.bridge = self._bridge
 180
 181         tck = 1/(2*self._sys_clk_freq)
 182         nphases = 2
 183         databits = len(self.pads.dq.io)
 184
 185         burstdet_reg = Signal(databits//8, reset_less=True)
 186         m.d.comb += self.burstdet.r_data.eq(burstdet_reg)
 187
 188         # Burstdet clear
 189         with m.If(self.burstdet.w_stb):
 190             m.d.sync += burstdet_reg.eq(0)
 191
 192         # Init -------------------------------------------------------------------------------------
 193         m.submodules.init = init = ECP5DDRPHYInit()
 194
 195         # Parameters -------------------------------------------------------------------------------
 196         cl, cwl = get_cl_cw("DDR3", tck)
 197         cl_sys_latency = get_sys_latency(nphases, cl)
 198         cwl_sys_latency = get_sys_latency(nphases, cwl)
 199
 200         # DFI Interface ----------------------------------------------------------------------------
 201         dfi = self.dfi
 202
 203         bl8_chunk = Signal()
 204
 205         # Clock --------------------------------------------------------------------------------
 206         m.d.comb += [
 207             self.pads.clk.o_clk.eq(ClockSignal("dramsync")),
 208             self.pads.clk.o_fclk.eq(ClockSignal("sync2x")),
 209         ]
 210         for i in range(len(self.pads.clk.o0)):
 211             m.d.comb += [
 212                 self.pads.clk.o0[i].eq(0),
 213                 self.pads.clk.o1[i].eq(1),
 214                 self.pads.clk.o2[i].eq(0),
 215                 self.pads.clk.o3[i].eq(1),
 216             ]
 217
 218         # Addresses and Commands ---------------------------------------------------------------
 219         m.d.comb += [
 220             self.pads.a.o_clk.eq(ClockSignal("dramsync")),
 221             self.pads.a.o_fclk.eq(ClockSignal("sync2x")),
 222             self.pads.ba.o_clk.eq(ClockSignal("dramsync")),
 223             self.pads.ba.o_fclk.eq(ClockSignal("sync2x")),
 224         ]
 225         for i in range(len(self.pads.a.o0)):
 226             m.d.comb += [
 227                 self.pads.a.o0[i].eq(dfi.phases[0].address[i]),
 228                 self.pads.a.o1[i].eq(dfi.phases[0].address[i]),
 229                 self.pads.a.o2[i].eq(dfi.phases[1].address[i]),
 230                 self.pads.a.o3[i].eq(dfi.phases[1].address[i]),
 231             ]
 232         for i in range(len(self.pads.ba.o0)):
 233             m.d.comb += [
 234                 self.pads.ba.o0[i].eq(dfi.phases[0].bank[i]),
 235                 self.pads.ba.o1[i].eq(dfi.phases[0].bank[i]),
 236                 self.pads.ba.o2[i].eq(dfi.phases[1].bank[i]),
 237                 self.pads.ba.o3[i].eq(dfi.phases[1].bank[i]),
 238             ]
 239
 240         # Control pins: all of thees have to be declared "xdr 4" when
 241         # requesting the resource:
 242         # ddr_pins = platform.request("ddr3", 0, xdr={"clk":4, "odt":4, ... })
 243         controls = ["ras", "cas", "we", "clk_en", "odt"]
 244         if hasattr(self.pads, "rst"): # this gets renamed later to match dfi
 245             controls.append("rst")
 246         if hasattr(self.pads, "cs"):
 247             controls.append("cs")
 248         for name in controls:
 249             print ("clock", name, getattr(self.pads, name))
 250             pad = getattr(self.pads, name)
 251             # sigh, convention in nmigen_boards is "rst" but in
 252             # dfi.Interface it is "reset"
 253             dfi2pads = {'rst': 'reset', 'cs': 'cs_n'}
 254             name = dfi2pads.get(name, name) # remap if exists
 255             m.d.comb += [
 256                 pad.o_clk.eq(ClockSignal("dramsync")),
 257                 pad.o_fclk.eq(ClockSignal("sync2x")),
 258             ]
 259             for i in range(len(pad.o0)):
 260                 m.d.comb += [
 261                     pad.o0[i].eq(getattr(dfi.phases[0], name)[i]),
 262                     pad.o1[i].eq(getattr(dfi.phases[0], name)[i]),
 263                     pad.o2[i].eq(getattr(dfi.phases[1], name)[i]),
 264                     pad.o3[i].eq(getattr(dfi.phases[1], name)[i]),
 265                 ]
 266
 267         # DQ ---------------------------------------------------------------------------------------
 268         dq_oe = Signal()
 269         dqs_re = Signal()
 270         dqs_oe = Signal()
 271         dqs_postamble = Signal()
 272         dqs_preamble = Signal()
 273         for i in range(databits//8):
 274             # DQSBUFM
 275             dqs_i = Signal()
 276             dqsr90 = Signal()
 277             dqsw270 = Signal()
 278             dqsw = Signal()
 279             rdpntr = Signal(3)
 280             wrpntr = Signal(3)
 281             burstdet = Signal()
 282             datavalid = Signal()
 283             datavalid_prev = Signal()
 284             m.d.sync += datavalid_prev.eq(datavalid)
 285
 286             dqsbufm_manager = _DQSBUFMSettingManager(self.rdly[i])
 287             setattr(m.submodules, f"dqsbufm_manager{i}", dqsbufm_manager)
 288
 289             m.submodules += Instance("DQSBUFM",
 290                 p_DQS_LI_DEL_ADJ="MINUS",
 291                 p_DQS_LI_DEL_VAL=1,
 292                 p_DQS_LO_DEL_ADJ="MINUS",
 293                 p_DQS_LO_DEL_VAL=4,
 294
 295                 # Delay
 296                 i_DYNDELAY0=0,
 297                 i_DYNDELAY1=0,
 298                 i_DYNDELAY2=0,
 299                 i_DYNDELAY3=0,
 300                 i_DYNDELAY4=0,
 301                 i_DYNDELAY5=0,
 302                 i_DYNDELAY6=0,
 303                 i_DYNDELAY7=0,
 304
 305                 # Clocks / Reset
 306                 i_SCLK=ClockSignal("sync"),
 307                 i_ECLK=ClockSignal("sync2x"),
 308                 i_RST=ResetSignal("dramsync"),
 309                 i_DDRDEL=init.delay,
 310                 i_PAUSE=init.pause | dqsbufm_manager.pause,
 311
 312                 # Control
 313                 # Assert LOADNs to use DDRDEL control
 314                 i_RDLOADN=0,
 315                 i_RDMOVE=0,
 316                 i_RDDIRECTION=1,
 317                 i_WRLOADN=0,
 318                 i_WRMOVE=0,
 319                 i_WRDIRECTION=1,
 320
 321                 # Reads (generate shifted DQS clock for reads)
 322                 i_READ0=dqs_re,
 323                 i_READ1=dqs_re,
 324                 i_READCLKSEL0=dqsbufm_manager.readclksel[0],
 325                 i_READCLKSEL1=dqsbufm_manager.readclksel[1],
 326                 i_READCLKSEL2=dqsbufm_manager.readclksel[2],
 327                 i_DQSI=dqs_i,
 328                 o_DQSR90=dqsr90,
 329                 o_RDPNTR0=rdpntr[0],
 330                 o_RDPNTR1=rdpntr[1],
 331                 o_RDPNTR2=rdpntr[2],
 332                 o_WRPNTR0=wrpntr[0],
 333                 o_WRPNTR1=wrpntr[1],
 334                 o_WRPNTR2=wrpntr[2],
 335                 o_BURSTDET=burstdet,
 336                 o_DATAVALID=datavalid,
 337
 338                 # Writes (generate shifted ECLK clock for writes)
 339                 o_DQSW270=dqsw270,
 340                 o_DQSW=dqsw)
 341
 342             with m.If(Rose(burstdet)):
 343                 m.d.sync += burstdet_reg[i].eq(1)
 344
 345             # DQS and DM ---------------------------------------------------------------------------
 346             dm_o_data = Signal(8)
 347             dm_o_data_d = Signal(8, reset_less=True)
 348             dm_o_data_muxed = Signal(4, reset_less=True)
 349             m.d.comb += dm_o_data.eq(Cat(
 350                 dfi.phases[0].wrdata_mask[0*databits//8+i],
 351                 dfi.phases[0].wrdata_mask[1*databits//8+i],
 352                 dfi.phases[0].wrdata_mask[2*databits//8+i],
 353                 dfi.phases[0].wrdata_mask[3*databits//8+i],
 354
 355                 dfi.phases[1].wrdata_mask[0*databits//8+i],
 356                 dfi.phases[1].wrdata_mask[1*databits//8+i],
 357                 dfi.phases[1].wrdata_mask[2*databits//8+i],
 358                 dfi.phases[1].wrdata_mask[3*databits//8+i]),
 359             )
 360             m.d.sync += dm_o_data_d.eq(dm_o_data)
 361
 362             with m.If(bl8_chunk):
 363                 m.d.sync += dm_o_data_muxed.eq(dm_o_data_d[4:])
 364             with m.Else():
 365                 m.d.sync += dm_o_data_muxed.eq(dm_o_data[:4])
 366
 367             m.submodules += Instance("ODDRX2DQA",
 368                 i_RST=ResetSignal("dramsync"),
 369                 i_ECLK=ClockSignal("sync2x"),
 370                 i_SCLK=ClockSignal("dramsync"),
 371                 i_DQSW270=dqsw270,
 372                 i_D0=dm_o_data_muxed[0],
 373                 i_D1=dm_o_data_muxed[1],
 374                 i_D2=dm_o_data_muxed[2],
 375                 i_D3=dm_o_data_muxed[3],
 376                 o_Q=self.pads.dm.o[i])
 377
 378             dqs = Signal()
 379             dqs_oe_n = Signal()
 380             m.submodules += [
 381                 Instance("ODDRX2DQSB",
 382                     i_RST=ResetSignal("dramsync"),
 383                     i_ECLK=ClockSignal("sync2x"),
 384                     i_SCLK=ClockSignal(),
 385                     i_DQSW=dqsw,
 386                     i_D0=0,
 387                     i_D1=1,
 388                     i_D2=0,
 389                     i_D3=1,
 390                     o_Q=dqs),
 391                 Instance("TSHX2DQSA",
 392                     i_RST=ResetSignal("dramsync"),
 393                     i_ECLK=ClockSignal("sync2x"),
 394                     i_SCLK=ClockSignal(),
 395                     i_DQSW=dqsw,
 396                     i_T0=~(dqs_oe | dqs_postamble),
 397                     i_T1=~(dqs_oe | dqs_preamble),
 398                     o_Q=dqs_oe_n),
 399                 Instance("BB",
 400                     i_I=dqs,
 401                     i_T=dqs_oe_n,
 402                     o_O=dqs_i,
 403                     io_B=self.pads.dqs.p[i]),
 404             ]
 405
 406             for j in range(8*i, 8*(i+1)):
 407                 dq_o = Signal(name="dq_o_%d" % j)
 408                 dq_i = Signal(name="dq_i_%d" % j)
 409                 dq_oe_n = Signal(name="dq_oe_n_%d" % j)
 410                 dq_i_delayed = Signal(name="dq_i_delayed_%d" % j)
 411                 dq_i_data = Signal(4, name="dq_i_data_%d" % j)
 412                 dq_o_data = Signal(8, name="dq_o_data_%d" % j)
 413                 dq_o_data_d = Signal(8, reset_less=True)
 414                 dq_o_data_muxed = Signal(4, reset_less=True)
 415                 m.d.comb += dq_o_data.eq(Cat(
 416                     dfi.phases[0].wrdata[0*databits+j],
 417                     dfi.phases[0].wrdata[1*databits+j],
 418                     dfi.phases[0].wrdata[2*databits+j],
 419                     dfi.phases[0].wrdata[3*databits+j],
 420                     dfi.phases[1].wrdata[0*databits+j],
 421                     dfi.phases[1].wrdata[1*databits+j],
 422                     dfi.phases[1].wrdata[2*databits+j],
 423                     dfi.phases[1].wrdata[3*databits+j])
 424                 )
 425
 426                 m.d.sync += dq_o_data_d.eq(dq_o_data)
 427                 with m.If(bl8_chunk):
 428                     m.d.sync += dq_o_data_muxed.eq(dq_o_data_d[4:])
 429                 with m.Else():
 430                     m.d.sync += dq_o_data_muxed.eq(dq_o_data[:4])
 431
 432                 m.submodules += [
 433                     Instance("ODDRX2DQA",
 434                         i_RST=ResetSignal("dramsync"),
 435                         i_ECLK=ClockSignal("sync2x"),
 436                         i_SCLK=ClockSignal(),
 437                         i_DQSW270=dqsw270,
 438                         i_D0=dq_o_data_muxed[0],
 439                         i_D1=dq_o_data_muxed[1],
 440                         i_D2=dq_o_data_muxed[2],
 441                         i_D3=dq_o_data_muxed[3],
 442                         o_Q=dq_o),
 443                     Instance("DELAYF",
 444                         p_DEL_MODE="DQS_ALIGNED_X2",
 445                         i_LOADN=1,
 446                         i_MOVE=0,
 447                         i_DIRECTION=0,
 448                         i_A=dq_i,
 449                         o_Z=dq_i_delayed),
 450                     Instance("IDDRX2DQA",
 451                         i_RST=ResetSignal("dramsync"),
 452                         i_ECLK=ClockSignal("sync2x"),
 453                         i_SCLK=ClockSignal(),
 454                         i_DQSR90=dqsr90,
 455                         i_RDPNTR0=rdpntr[0],
 456                         i_RDPNTR1=rdpntr[1],
 457                         i_RDPNTR2=rdpntr[2],
 458                         i_WRPNTR0=wrpntr[0],
 459                         i_WRPNTR1=wrpntr[1],
 460                         i_WRPNTR2=wrpntr[2],
 461                         i_D=dq_i_delayed,
 462                         o_Q0=dq_i_data[0],
 463                         o_Q1=dq_i_data[1],
 464                         o_Q2=dq_i_data[2],
 465                         o_Q3=dq_i_data[3]),
 466                     Instance("TSHX2DQA",
 467                         i_RST=ResetSignal("dramsync"),
 468                         i_ECLK=ClockSignal("sync2x"),
 469                         i_SCLK=ClockSignal(),
 470                         i_DQSW270=dqsw270,
 471                         i_T0=~dq_oe,
 472                         i_T1=~dq_oe,
 473                         o_Q=dq_oe_n),
 474                     Instance("BB",
 475                         i_I=dq_o,
 476                         i_T=dq_oe_n,
 477                         o_O=dq_i,
 478                         io_B=self.pads.dq.io[j])
 479                 ]
 480                 # shift-register delay on the incoming read data
 481                 dq_i_bs = BitSlip(4, Const(0), Const(0), cycles=1)
 482                 m.submodules['dq_i_bitslip_%d' % j] = dq_i_bs
 483                 dq_i_bs_o = Signal(4, name="dq_i_bs_o_%d" % j)
 484                 dq_i_bs_o_d = Signal(4, name="dq_i_bs_o_d_%d" % j)
 485                 comb += dq_i_bs.i.eq(dq_i_data)
 486                 comb += dq_i_bs_o.eq(dq_i_bs.o)
 487                 sync += dq_i_bs_o_d.eq(dq_i_bs_o) # delay by 1 clock
 488                 #with m.If(~datavalid_prev & datavalid):
 489                 comb += [
 490                     dfi.phases[0].rddata[0*databits+j].eq(dq_i_bs_o_d[0]),
 491                     dfi.phases[0].rddata[1*databits+j].eq(dq_i_bs_o_d[1]),
 492                     dfi.phases[0].rddata[2*databits+j].eq(dq_i_bs_o_d[2]),
 493                     dfi.phases[0].rddata[3*databits+j].eq(dq_i_bs_o_d[3]),
 494                 ]
 495                 #with m.Elif(datavalid):
 496                 comb += [
 497                     dfi.phases[1].rddata[0*databits+j].eq(dq_i_bs_o[0]),
 498                     dfi.phases[1].rddata[1*databits+j].eq(dq_i_bs_o[1]),
 499                     dfi.phases[1].rddata[2*databits+j].eq(dq_i_bs_o[2]),
 500                     dfi.phases[1].rddata[3*databits+j].eq(dq_i_bs_o[3]),
 501                 ]
 502
 503         # Read Control Path ------------------------------------------------------------------------
 504         # Creates a shift register of read commands coming from the DFI interface. This shift register
 505         # is used to control DQS read (internal read pulse of the DQSBUF) and to indicate to the
 506         # DFI interface that the read data is valid.
 507         #
 508         # The DQS read must be asserted for 2 sys_clk cycles before the read data is coming back from
 509         # the DRAM (see 6.2.4 READ Pulse Positioning Optimization of FPGA-TN-02035-1.2)
 510         #
 511         # The read data valid is asserted for 1 sys_clk cycle when the data is available on the DFI
 512         # interface, the latency is the sum of the ODDRX2DQA, CAS, IDDRX2DQA latencies.
 513         rddata_en = Signal(self.settings.read_latency)
 514         rddata_en_last = Signal.like(rddata_en)
 515         m.d.comb += rddata_en.eq(Cat(dfi.phases[self.settings.rdphase].rddata_en, rddata_en_last))
 516         m.d.sync += rddata_en_last.eq(rddata_en)
 517         for phase in dfi.phases:
 518             m.d.sync += phase.rddata_valid.eq(rddata_en[-1])
 519         m.d.comb += dqs_re.eq(rddata_en[cl_sys_latency + 1] | rddata_en[cl_sys_latency + 2])
 520
 521
 522         # Write Control Path -----------------------------------------------------------------------
 523         # Creates a shift register of write commands coming from the DFI interface. This shift register
 524         # is used to control DQ/DQS tristates and to select write data of the DRAM burst from the DFI
 525         # interface: The PHY is operating in halfrate mode (so provide 4 datas every sys_clk cycles:
 526         # 2x for DDR, 2x for halfrate) but DDR3 requires a burst of 8 datas (BL8) for best efficiency.
 527         # Writes are then performed in 2 sys_clk cycles and data needs to be selected for each cycle.
 528         # FIXME: understand +2
 529         wrdata_en = Signal(cwl_sys_latency + 4)
 530         wrdata_en_last = Signal.like(wrdata_en)
 531         m.d.comb += wrdata_en.eq(Cat(dfi.phases[self.settings.wrphase].wrdata_en, wrdata_en_last))
 532         m.d.sync += wrdata_en_last.eq(wrdata_en)
 533         m.d.comb += dq_oe.eq(wrdata_en[cwl_sys_latency + 1] | wrdata_en[cwl_sys_latency + 2])
 534         m.d.comb += bl8_chunk.eq(wrdata_en[cwl_sys_latency + 1])
 535         m.d.comb += dqs_oe.eq(dq_oe)
 536
 537         # Write DQS Postamble/Preamble Control Path ------------------------------------------------
 538         # Generates DQS Preamble 1 cycle before the first write and Postamble 1 cycle after the last
 539         # write. During writes, DQS tristate is configured as output for at least 4 sys_clk cycles:
 540         # 1 for Preamble, 2 for the Write and 1 for the Postamble.
 541         m.d.comb += dqs_preamble.eq(wrdata_en[cwl_sys_latency + 0] & ~wrdata_en[cwl_sys_latency + 1])
 542         m.d.comb += dqs_postamble.eq(wrdata_en[cwl_sys_latency + 3] & ~wrdata_en[cwl_sys_latency + 2])
 543
 544         return m