src/nmutil/queue.py

   1 # Copyright (c) 2014 - 2019 The Regents of the University of
   2 # California (Regents). All Rights Reserved.  Redistribution and use in
   3 # source and binary forms, with or without modification, are permitted
   4 # provided that the following conditions are met:
   5 #    * Redistributions of source code must retain the above
   6 #      copyright notice, this list of conditions and the following
   7 #      two paragraphs of disclaimer.
   8 #    * Redistributions in binary form must reproduce the above
   9 #      copyright notice, this list of conditions and the following
  10 #      two paragraphs of disclaimer in the documentation and/or other materials
  11 #      provided with the distribution.
  12 #    * Neither the name of the Regents nor the names of its contributors
  13 #      may be used to endorse or promote products derived from this
  14 #      software without specific prior written permission.
  15 # IN NO EVENT SHALL REGENTS BE LIABLE TO ANY PARTY FOR DIRECT, INDIRECT,
  16 # SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, INCLUDING LOST PROFITS,
  17 # ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN IF
  18 # REGENTS HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  19 # REGENTS SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING, BUT NOT
  20 # LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  21 # A PARTICULAR PURPOSE. THE SOFTWARE AND ACCOMPANYING DOCUMENTATION, IF
  22 # ANY, PROVIDED HEREUNDER IS PROVIDED "AS IS". REGENTS HAS NO OBLIGATION
  23 # TO PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR
  24 # MODIFICATIONS.
  25
  26 from nmigen import Module, Signal, Mux, Elaboratable, Memory
  27 from nmigen.utils import bits_for
  28 from nmigen.cli import main
  29 from nmigen.lib.fifo import FIFOInterface
  30
  31 # translated from https://github.com/freechipsproject/chisel3/blob/a4a29e29c3f1eed18f851dcf10bdc845571dfcb6/src/main/scala/chisel3/util/Decoupled.scala#L185   # noqa
  32
  33
  34 class Queue(FIFOInterface, Elaboratable):
  35     def __init__(self, width, depth, fwft=True, pipe=False):
  36         """ Queue (FIFO) with pipe mode and first-write fall-through capability
  37
  38             * :width: width of Queue data in/out
  39             * :depth: queue depth.  NOTE: may be set to 0 (this is ok)
  40             * :fwft : first-write, fall-through mode (Chisel Queue "flow" mode)
  41             * :pipe : pipe mode.  NOTE: this mode can cause unanticipated
  42                       problems.  when read is enabled, so is w_rdy.
  43                       therefore if read is enabled, the data ABSOLUTELY MUST
  44                       be read.
  45
  46             fwft mode = True basically means that the data may be transferred
  47             combinatorially from input to output.
  48
  49             Attributes:
  50             * level: available free space (number of unread entries)
  51
  52             w_data  = enq_data, w_rdy  = enq_ready, w_en = enq_valid
  53             r_data = deq_data, r_en = deq_ready, r_rdy = deq_valid
  54         """
  55         FIFOInterface.__init__(self, width=width, depth=depth, fwft=fwft)
  56         self.pipe = pipe
  57         self.depth = depth
  58         self.level = Signal(bits_for(depth))
  59
  60     def elaborate(self, platform):
  61         m = Module()
  62
  63         # set up an SRAM.  XXX bug in Memory: cannot create SRAM of depth 1
  64         ram = Memory(width=self.width,
  65                      depth=self.depth if self.depth > 1 else 2)
  66         m.submodules.ram_read = ram_read = ram.read_port(domain="comb")
  67         m.submodules.ram_write = ram_write = ram.write_port()
  68
  69         # convenience names, for people familiar with ready/valid terminology
  70         # "p" stands for "previous stage", "n" stands for "next stage"
  71         # for people familiar with the chisel Decoupled library:
  72         # enq is "enqueue" (data in, aka "prev stage"),
  73         # deq is "dequeue" (data out, aka "next stage")
  74         p_o_ready = self.w_rdy
  75         p_i_valid = self.w_en
  76         enq_data = self.w_data  # aka p_i_data
  77
  78         n_o_valid = self.r_rdy
  79         n_i_ready = self.r_en
  80         deq_data = self.r_data  # aka n_o_data
  81
  82         # intermediaries
  83         ptr_width = bits_for(self.depth - 1) if self.depth > 1 else 0
  84         # cyclic pointer to "insert" point (wrport)
  85         enq_ptr = Signal(ptr_width)
  86         # cyclic pointer to "remove" point (rdport)
  87         deq_ptr = Signal(ptr_width)
  88         maybe_full = Signal()  # not reset_less (set by sync)
  89
  90         # temporaries
  91         do_enq = Signal(reset_less=True)
  92         do_deq = Signal(reset_less=True)
  93         ptr_diff = Signal(ptr_width)
  94         ptr_match = Signal(reset_less=True)
  95         empty = Signal(reset_less=True)
  96         full = Signal(reset_less=True)
  97         enq_max = Signal(reset_less=True)
  98         deq_max = Signal(reset_less=True)
  99
 100         m.d.comb += [ptr_match.eq(enq_ptr == deq_ptr),  # read-ptr = write-ptr
 101                      ptr_diff.eq(enq_ptr - deq_ptr),
 102                      enq_max.eq(enq_ptr == self.depth - 1),
 103                      deq_max.eq(deq_ptr == self.depth - 1),
 104                      empty.eq(ptr_match & ~maybe_full),
 105                      full.eq(ptr_match & maybe_full),
 106                      do_enq.eq(p_o_ready & p_i_valid),  # write conditions ok
 107                      do_deq.eq(n_i_ready & n_o_valid),  # read conditions ok
 108
 109                      # set r_rdy and w_rdy (NOTE: see pipe mode below)
 110                      n_o_valid.eq(~empty),  # cannot read if empty!
 111                      p_o_ready.eq(~full),  # cannot write if full!
 112
 113                      # set up memory and connect to input and output
 114                      ram_write.addr.eq(enq_ptr),
 115                      ram_write.data.eq(enq_data),
 116                      ram_write.en.eq(do_enq),
 117                      ram_read.addr.eq(deq_ptr),
 118                      # NOTE: overridden in fwft mode
 119                      deq_data.eq(ram_read.data)
 120                      ]
 121
 122         # under write conditions, SRAM write-pointer moves on next clock
 123         with m.If(do_enq):
 124             m.d.sync += enq_ptr.eq(Mux(enq_max, 0, enq_ptr+1))
 125
 126         # under read conditions, SRAM read-pointer moves on next clock
 127         with m.If(do_deq):
 128             m.d.sync += deq_ptr.eq(Mux(deq_max, 0, deq_ptr+1))
 129
 130         # if read-but-not-write or write-but-not-read, maybe_full set
 131         with m.If(do_enq != do_deq):
 132             m.d.sync += maybe_full.eq(do_enq)
 133
 134         # first-word fall-through: same as "flow" parameter in Chisel3 Queue
 135         # basically instead of relying on the Memory characteristics (which
 136         # in FPGAs do not have write-through), then when the queue is empty
 137         # take the output directly from the input, i.e. *bypass* the SRAM.
 138         # this done combinatorially to give the exact same characteristics
 139         # as Memory "write-through"... without relying on a changing API
 140         if self.fwft:
 141             with m.If(p_i_valid):
 142                 m.d.comb += n_o_valid.eq(1)
 143             with m.If(empty):
 144                 m.d.comb += deq_data.eq(enq_data)
 145                 m.d.comb += do_deq.eq(0)
 146                 with m.If(n_i_ready):
 147                     m.d.comb += do_enq.eq(0)
 148
 149         # pipe mode: if next stage says it's ready (r_rdy), w_en
 150         #            *must* declare the input ready (w_rdy).
 151         if self.pipe:
 152             with m.If(n_i_ready):
 153                 m.d.comb += p_o_ready.eq(1)
 154
 155         # set the count (available free space), optimise on power-of-two
 156         if self.depth == 1 << ptr_width:  # is depth a power of 2
 157             m.d.comb += self.level.eq(
 158                 Mux(maybe_full & ptr_match, self.depth, 0) | ptr_diff)
 159         else:
 160             m.d.comb += self.level.eq(Mux(ptr_match,
 161                                           Mux(maybe_full, self.depth, 0),
 162                                           Mux(deq_ptr > enq_ptr,
 163                                               self.depth + ptr_diff,
 164                                               ptr_diff)))
 165
 166         return m
 167
 168
 169 if __name__ == "__main__":
 170     reg_stage = Queue(1, 1, pipe=True)
 171     break_ready_chain_stage = Queue(1, 1, pipe=True, fwft=True)
 172     m = Module()
 173     ports = []
 174
 175     def queue_ports(queue, name_prefix):
 176         retval = []
 177         for name in ["level",
 178                      "r_data",
 179                      "r_rdy",
 180                      "w_rdy"]:
 181             port = getattr(queue, name)
 182             signal = Signal(port.shape(), name=name_prefix+name)
 183             m.d.comb += signal.eq(port)
 184             retval.append(signal)
 185         for name in ["r_en",
 186                      "w_data",
 187                      "w_en"]:
 188             port = getattr(queue, name)
 189             signal = Signal(port.shape(), name=name_prefix+name)
 190             m.d.comb += port.eq(signal)
 191             retval.append(signal)
 192         return retval
 193
 194     m.submodules.reg_stage = reg_stage
 195     ports += queue_ports(reg_stage, "reg_stage_")
 196     m.submodules.break_ready_chain_stage = break_ready_chain_stage
 197     ports += queue_ports(break_ready_chain_stage, "break_ready_chain_stage_")
 198     main(m, ports=ports)