amaranth/amaranth/lib/fifo.py
Wanda 115954b4d9 lib.fifo: add Memory as submodules instead of its ports. [NFC]
This makes the generated netlist very slightly nicer.
2024-01-31 21:14:08 +00:00

580 lines
22 KiB
Python

"""First-in first-out queues."""
import warnings
from .. import *
from ..asserts import *
from ..utils import ceil_log2
from .coding import GrayEncoder, GrayDecoder
from .cdc import FFSynchronizer, AsyncFFSynchronizer
__all__ = ["FIFOInterface", "SyncFIFO", "SyncFIFOBuffered", "AsyncFIFO", "AsyncFIFOBuffered"]
class FIFOInterface:
_doc_template = """
{description}
Parameters
----------
width : int
Bit width of data entries.
depth : int
Depth of the queue. If zero, the FIFO cannot be read from or written to.
{parameters}
Attributes
----------
{attributes}
w_data : Signal(width), in
Input data.
w_rdy : Signal(1), out
Asserted if there is space in the queue, i.e. ``w_en`` can be asserted to write
a new entry.
w_en : Signal(1), in
Write strobe. Latches ``w_data`` into the queue. Does nothing if ``w_rdy`` is not asserted.
w_level : Signal(range(depth + 1)), out
Number of unread entries.
{w_attributes}
r_data : Signal(width), out
Output data. {r_data_valid}
r_rdy : Signal(1), out
Asserted if there is an entry in the queue, i.e. ``r_en`` can be asserted to read
an existing entry.
r_en : Signal(1), in
Read strobe. Makes the next entry (if any) available on ``r_data`` at the next cycle.
Does nothing if ``r_rdy`` is not asserted.
r_level : Signal(range(depth + 1)), out
Number of unread entries.
{r_attributes}
"""
__doc__ = _doc_template.format(description="""
Data written to the input interface (``w_data``, ``w_rdy``, ``w_en``) is buffered and can be
read at the output interface (``r_data``, ``r_rdy``, ``r_en``). The data entry written first
to the input also appears first on the output.
""",
parameters="",
r_data_valid="The conditions in which ``r_data`` is valid depends on the type of the queue.",
attributes="",
w_attributes="",
r_attributes="")
def __init__(self, *, width, depth):
if not isinstance(width, int) or width < 0:
raise TypeError("FIFO width must be a non-negative integer, not {!r}"
.format(width))
if not isinstance(depth, int) or depth < 0:
raise TypeError("FIFO depth must be a non-negative integer, not {!r}"
.format(depth))
self.width = width
self.depth = depth
self.w_data = Signal(width, reset_less=True)
self.w_rdy = Signal() # writable; not full
self.w_en = Signal()
self.w_level = Signal(range(depth + 1))
self.r_data = Signal(width, reset_less=True)
self.r_rdy = Signal() # readable; not empty
self.r_en = Signal()
self.r_level = Signal(range(depth + 1))
def _incr(signal, modulo):
if modulo == 2 ** len(signal):
return signal + 1
else:
return Mux(signal == modulo - 1, 0, signal + 1)
class SyncFIFO(Elaboratable, FIFOInterface):
__doc__ = FIFOInterface._doc_template.format(
description="""
Synchronous first in, first out queue.
Read and write interfaces are accessed from the same clock domain. If different clock domains
are needed, use :class:`AsyncFIFO`.
""".strip(),
parameters="",
r_data_valid="Valid if ``r_rdy`` is asserted.",
attributes="""
level : Signal(range(depth + 1)), out
Number of unread entries. This level is the same between read and write for synchronous FIFOs.
""".strip(),
r_attributes="",
w_attributes="")
def __init__(self, *, width, depth):
super().__init__(width=width, depth=depth)
self.level = Signal(range(depth + 1))
def elaborate(self, platform):
m = Module()
if self.depth == 0:
m.d.comb += [
self.w_rdy.eq(0),
self.r_rdy.eq(0),
]
return m
m.d.comb += [
self.w_rdy.eq(self.level != self.depth),
self.r_rdy.eq(self.level != 0),
self.w_level.eq(self.level),
self.r_level.eq(self.level),
]
do_read = self.r_rdy & self.r_en
do_write = self.w_rdy & self.w_en
storage = m.submodules.storage = Memory(width=self.width, depth=self.depth)
w_port = storage.write_port()
r_port = storage.read_port(domain="comb")
produce = Signal(range(self.depth))
consume = Signal(range(self.depth))
m.d.comb += [
w_port.addr.eq(produce),
w_port.data.eq(self.w_data),
w_port.en.eq(self.w_en & self.w_rdy),
]
with m.If(do_write):
m.d.sync += produce.eq(_incr(produce, self.depth))
m.d.comb += [
r_port.addr.eq(consume),
self.r_data.eq(r_port.data),
]
with m.If(do_read):
m.d.sync += consume.eq(_incr(consume, self.depth))
with m.If(do_write & ~do_read):
m.d.sync += self.level.eq(self.level + 1)
with m.If(do_read & ~do_write):
m.d.sync += self.level.eq(self.level - 1)
if platform == "formal":
# TODO: move this logic to SymbiYosys
with m.If(Initial()):
m.d.comb += [
Assume(produce < self.depth),
Assume(consume < self.depth),
]
with m.If(produce == consume):
m.d.comb += Assume((self.level == 0) | (self.level == self.depth))
with m.If(produce > consume):
m.d.comb += Assume(self.level == (produce - consume))
with m.If(produce < consume):
m.d.comb += Assume(self.level == (self.depth + produce - consume))
with m.Else():
m.d.comb += [
Assert(produce < self.depth),
Assert(consume < self.depth),
]
with m.If(produce == consume):
m.d.comb += Assert((self.level == 0) | (self.level == self.depth))
with m.If(produce > consume):
m.d.comb += Assert(self.level == (produce - consume))
with m.If(produce < consume):
m.d.comb += Assert(self.level == (self.depth + produce - consume))
return m
class SyncFIFOBuffered(Elaboratable, FIFOInterface):
__doc__ = FIFOInterface._doc_template.format(
description="""
Buffered synchronous first in, first out queue.
This queue's interface is identical to :class:`SyncFIFO`, but it
does not use asynchronous memory reads, which are incompatible with FPGA block RAMs.
In exchange, the latency between an entry being written to an empty queue and that entry
becoming available on the output is increased by one cycle compared to :class:`SyncFIFO`.
""".strip(),
parameters="",
attributes="""
level : Signal(range(depth + 1)), out
Number of unread entries. This level is the same between read and write for synchronous FIFOs.
""".strip(),
r_data_valid="Valid if ``r_rdy`` is asserted.",
r_attributes="",
w_attributes="")
def __init__(self, *, width, depth):
super().__init__(width=width, depth=depth)
self.level = Signal(range(depth + 1))
def elaborate(self, platform):
m = Module()
if self.depth == 0:
m.d.comb += [
self.w_rdy.eq(0),
self.r_rdy.eq(0),
]
return m
do_write = self.w_rdy & self.w_en
do_read = self.r_rdy & self.r_en
m.d.comb += [
self.w_level.eq(self.level),
self.r_level.eq(self.level),
]
if self.depth == 1:
# Special case: a single register. Note that, by construction, this will
# only be able to push a value every other cycle (alternating between
# full and empty).
m.d.comb += [
self.w_rdy.eq(self.level == 0),
self.r_rdy.eq(self.level == 1),
]
with m.If(do_write):
m.d.sync += [
self.r_data.eq(self.w_data),
self.level.eq(1),
]
with m.If(do_read):
m.d.sync += [
self.level.eq(0),
]
return m
inner_depth = self.depth - 1
inner_level = Signal(range(inner_depth + 1))
inner_r_rdy = Signal()
m.d.comb += [
self.w_rdy.eq(inner_level != inner_depth),
inner_r_rdy.eq(inner_level != 0),
]
do_inner_read = inner_r_rdy & (~self.r_rdy | self.r_en)
storage = m.submodules.storage = Memory(width=self.width, depth=inner_depth)
w_port = storage.write_port()
r_port = storage.read_port(domain="sync", transparent=False)
produce = Signal(range(inner_depth))
consume = Signal(range(inner_depth))
m.d.comb += [
w_port.addr.eq(produce),
w_port.data.eq(self.w_data),
w_port.en.eq(do_write),
]
with m.If(do_write):
m.d.sync += produce.eq(_incr(produce, inner_depth))
m.d.comb += [
r_port.addr.eq(consume),
self.r_data.eq(r_port.data),
r_port.en.eq(do_inner_read)
]
with m.If(do_inner_read):
m.d.sync += consume.eq(_incr(consume, inner_depth))
with m.If(do_write & ~do_inner_read):
m.d.sync += inner_level.eq(inner_level + 1)
with m.If(do_inner_read & ~do_write):
m.d.sync += inner_level.eq(inner_level - 1)
with m.If(do_inner_read):
m.d.sync += self.r_rdy.eq(1)
with m.Elif(self.r_en):
m.d.sync += self.r_rdy.eq(0)
m.d.comb += [
self.level.eq(inner_level + self.r_rdy),
]
if platform == "formal":
# TODO: move this logic to SymbiYosys
with m.If(Initial()):
m.d.comb += [
Assume(produce < inner_depth),
Assume(consume < inner_depth),
]
with m.If(produce == consume):
m.d.comb += Assume((inner_level == 0) | (inner_level == inner_depth))
with m.If(produce > consume):
m.d.comb += Assume(inner_level == (produce - consume))
with m.If(produce < consume):
m.d.comb += Assume(inner_level == (inner_depth + produce - consume))
with m.Else():
m.d.comb += [
Assert(produce < inner_depth),
Assert(consume < inner_depth),
]
with m.If(produce == consume):
m.d.comb += Assert((inner_level == 0) | (inner_level == inner_depth))
with m.If(produce > consume):
m.d.comb += Assert(inner_level == (produce - consume))
with m.If(produce < consume):
m.d.comb += Assert(inner_level == (inner_depth + produce - consume))
return m
class AsyncFIFO(Elaboratable, FIFOInterface):
__doc__ = FIFOInterface._doc_template.format(
description="""
Asynchronous first in, first out queue.
Read and write interfaces are accessed from different clock domains, which can be set when
constructing the FIFO.
:class:`AsyncFIFO` can be reset from the write clock domain. When the write domain reset is
asserted, the FIFO becomes empty. When the read domain is reset, data remains in the FIFO - the
read domain logic should correctly handle this case.
:class:`AsyncFIFO` only supports power of 2 depths. Unless ``exact_depth`` is specified,
the ``depth`` parameter is rounded up to the next power of 2.
""".strip(),
parameters="""
r_domain : str
Read clock domain.
w_domain : str
Write clock domain.
""".strip(),
attributes="",
r_data_valid="Valid if ``r_rdy`` is asserted.",
r_attributes="""
r_rst : Signal(1), out
Asserted, for at least one read-domain clock cycle, after the FIFO has been reset by
the write-domain reset.
""".strip(),
w_attributes="")
def __init__(self, *, width, depth, r_domain="read", w_domain="write", exact_depth=False):
if depth != 0:
depth_bits = ceil_log2(depth)
if exact_depth and depth != 1 << depth_bits:
raise ValueError("AsyncFIFO only supports depths that are powers of 2; requested "
"exact depth {} is not"
.format(depth)) from None
depth = 1 << depth_bits
else:
depth_bits = 0
super().__init__(width=width, depth=depth)
self.r_rst = Signal()
self._r_domain = r_domain
self._w_domain = w_domain
self._ctr_bits = depth_bits + 1
def elaborate(self, platform):
m = Module()
if self.depth == 0:
m.d.comb += [
self.w_rdy.eq(0),
self.r_rdy.eq(0),
]
return m
# The design of this queue is the "style #2" from Clifford E. Cummings' paper "Simulation
# and Synthesis Techniques for Asynchronous FIFO Design":
# http://www.sunburst-design.com/papers/CummingsSNUG2002SJ_FIFO1.pdf
do_write = self.w_rdy & self.w_en
do_read = self.r_rdy & self.r_en
# TODO: extract this pattern into lib.cdc.GrayCounter
produce_w_bin = Signal(self._ctr_bits)
produce_w_nxt = Signal(self._ctr_bits)
m.d.comb += produce_w_nxt.eq(produce_w_bin + do_write)
m.d[self._w_domain] += produce_w_bin.eq(produce_w_nxt)
# Note: Both read-domain counters must be reset_less (see comments below)
consume_r_bin = Signal(self._ctr_bits, reset_less=True)
consume_r_nxt = Signal(self._ctr_bits)
m.d.comb += consume_r_nxt.eq(consume_r_bin + do_read)
m.d[self._r_domain] += consume_r_bin.eq(consume_r_nxt)
produce_w_gry = Signal(self._ctr_bits)
produce_r_gry = Signal(self._ctr_bits)
produce_enc = m.submodules.produce_enc = \
GrayEncoder(self._ctr_bits)
produce_cdc = m.submodules.produce_cdc = \
FFSynchronizer(produce_w_gry, produce_r_gry, o_domain=self._r_domain)
m.d.comb += produce_enc.i.eq(produce_w_nxt),
m.d[self._w_domain] += produce_w_gry.eq(produce_enc.o)
consume_r_gry = Signal(self._ctr_bits, reset_less=True)
consume_w_gry = Signal(self._ctr_bits)
consume_enc = m.submodules.consume_enc = \
GrayEncoder(self._ctr_bits)
consume_cdc = m.submodules.consume_cdc = \
FFSynchronizer(consume_r_gry, consume_w_gry, o_domain=self._w_domain)
m.d.comb += consume_enc.i.eq(consume_r_nxt)
m.d[self._r_domain] += consume_r_gry.eq(consume_enc.o)
consume_w_bin = Signal(self._ctr_bits)
consume_dec = m.submodules.consume_dec = \
GrayDecoder(self._ctr_bits)
m.d.comb += consume_dec.i.eq(consume_w_gry),
m.d[self._w_domain] += consume_w_bin.eq(consume_dec.o)
produce_r_bin = Signal(self._ctr_bits)
produce_dec = m.submodules.produce_dec = \
GrayDecoder(self._ctr_bits)
m.d.comb += produce_dec.i.eq(produce_r_gry),
m.d.comb += produce_r_bin.eq(produce_dec.o)
w_full = Signal()
r_empty = Signal()
m.d.comb += [
w_full.eq((produce_w_gry[-1] != consume_w_gry[-1]) &
(produce_w_gry[-2] != consume_w_gry[-2]) &
(produce_w_gry[:-2] == consume_w_gry[:-2])),
r_empty.eq(consume_r_gry == produce_r_gry),
]
m.d[self._w_domain] += self.w_level.eq(produce_w_bin - consume_w_bin)
m.d.comb += self.r_level.eq(produce_r_bin - consume_r_bin)
storage = m.submodules.storage = Memory(width=self.width, depth=self.depth)
w_port = storage.write_port(domain=self._w_domain)
r_port = storage.read_port (domain=self._r_domain, transparent=False)
m.d.comb += [
w_port.addr.eq(produce_w_bin[:-1]),
w_port.data.eq(self.w_data),
w_port.en.eq(do_write),
self.w_rdy.eq(~w_full),
]
m.d.comb += [
r_port.addr.eq(consume_r_nxt[:-1]),
self.r_data.eq(r_port.data),
r_port.en.eq(1),
self.r_rdy.eq(~r_empty),
]
# Reset handling to maintain FIFO and CDC invariants in the presence of a write-domain
# reset.
# There is a CDC hazard associated with resetting an async FIFO - Gray code counters which
# are reset to 0 violate their Gray code invariant. One way to handle this is to ensure
# that both sides of the FIFO are asynchronously reset by the same signal. We adopt a
# slight variation on this approach - reset control rests entirely with the write domain.
# The write domain's reset signal is used to asynchronously reset the read domain's
# counters and force the FIFO to be empty when the write domain's reset is asserted.
# This requires the two read domain counters to be marked as "reset_less", as they are
# reset through another mechanism. See https://github.com/amaranth-lang/amaranth/issues/181
# for the full discussion.
w_rst = ResetSignal(domain=self._w_domain, allow_reset_less=True)
r_rst = Signal()
# Async-set-sync-release synchronizer avoids CDC hazards
rst_cdc = m.submodules.rst_cdc = \
AsyncFFSynchronizer(w_rst, r_rst, o_domain=self._r_domain)
# Decode Gray code counter synchronized from write domain to overwrite binary
# counter in read domain.
rst_dec = m.submodules.rst_dec = \
GrayDecoder(self._ctr_bits)
m.d.comb += rst_dec.i.eq(produce_r_gry)
with m.If(r_rst):
m.d.comb += r_empty.eq(1)
m.d[self._r_domain] += consume_r_gry.eq(produce_r_gry)
m.d[self._r_domain] += consume_r_bin.eq(rst_dec.o)
m.d[self._r_domain] += self.r_rst.eq(1)
with m.Else():
m.d[self._r_domain] += self.r_rst.eq(0)
if platform == "formal":
with m.If(Initial()):
m.d.comb += Assume(produce_w_gry == (produce_w_bin ^ produce_w_bin[1:]))
m.d.comb += Assume(consume_r_gry == (consume_r_bin ^ consume_r_bin[1:]))
return m
class AsyncFIFOBuffered(Elaboratable, FIFOInterface):
__doc__ = FIFOInterface._doc_template.format(
description="""
Buffered asynchronous first in, first out queue.
Read and write interfaces are accessed from different clock domains, which can be set when
constructing the FIFO.
:class:`AsyncFIFOBuffered` only supports power of 2 plus one depths. Unless ``exact_depth``
is specified, the ``depth`` parameter is rounded up to the next power of 2 plus one.
(The output buffer acts as an additional queue element.)
This queue's interface is identical to :class:`AsyncFIFO`, but it has an additional register
on the output, improving timing in case of block RAM that has large clock-to-output delay.
In exchange, the latency between an entry being written to an empty queue and that entry
becoming available on the output is increased by one cycle compared to :class:`AsyncFIFO`.
""".strip(),
parameters="""
r_domain : str
Read clock domain.
w_domain : str
Write clock domain.
""".strip(),
attributes="",
r_data_valid="Valid if ``r_rdy`` is asserted.",
r_attributes="""
r_rst : Signal(1), out
Asserted, for at least one read-domain clock cycle, after the FIFO has been reset by
the write-domain reset.
""".strip(),
w_attributes="")
def __init__(self, *, width, depth, r_domain="read", w_domain="write", exact_depth=False):
if depth != 0:
depth_bits = ceil_log2(max(0, depth - 1))
if exact_depth and depth != (1 << depth_bits) + 1:
raise ValueError("AsyncFIFOBuffered only supports depths that are one higher "
"than powers of 2; requested exact depth {} is not"
.format(depth)) from None
depth = (1 << depth_bits) + 1
super().__init__(width=width, depth=depth)
self.r_rst = Signal()
self._r_domain = r_domain
self._w_domain = w_domain
def elaborate(self, platform):
m = Module()
if self.depth == 0:
m.d.comb += [
self.w_rdy.eq(0),
self.r_rdy.eq(0),
]
return m
m.submodules.unbuffered = fifo = AsyncFIFO(width=self.width, depth=self.depth - 1,
r_domain=self._r_domain, w_domain=self._w_domain)
m.d.comb += [
fifo.w_data.eq(self.w_data),
self.w_rdy.eq(fifo.w_rdy),
fifo.w_en.eq(self.w_en),
]
r_consume_buffered = Signal()
m.d.comb += r_consume_buffered.eq((self.r_rdy - self.r_en) & self.r_rdy)
m.d[self._r_domain] += self.r_level.eq(fifo.r_level + r_consume_buffered)
w_consume_buffered = Signal()
m.submodules.consume_buffered_cdc = FFSynchronizer(r_consume_buffered, w_consume_buffered, o_domain=self._w_domain, stages=4)
m.d.comb += self.w_level.eq(fifo.w_level + w_consume_buffered)
with m.If(self.r_en | ~self.r_rdy):
m.d[self._r_domain] += [
self.r_data.eq(fifo.r_data),
self.r_rdy.eq(fifo.r_rdy),
self.r_rst.eq(fifo.r_rst),
]
m.d.comb += [
fifo.r_en.eq(1)
]
return m