def __init__(self, interface, counter, fifo_depth):
data_width = rtlink.get_data_width(interface)
fine_ts_width = rtlink.get_fine_ts_width(interface)
ev_layout = []
if data_width:
ev_layout.append(("data", data_width))
if interface.timestamped:
ev_layout.append(("timestamp", counter.width + fine_ts_width))
self.ev = Record(ev_layout)
self.readable = Signal()
self.re = Signal()
self.overflow = Signal() # pulsed
# # #
fifo = ClockDomainsRenamer({"read": "rsys", "write": "rio"})(
AsyncFIFO(layout_len(ev_layout), fifo_depth))
self.submodules += fifo
fifo_in = Record(ev_layout)
fifo_out = Record(ev_layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# latency compensation
if interface.delay:
counter_rtio = Signal.like(counter.value_rtio, reset_less=True)
self.sync.rtio += counter_rtio.eq(counter.value_rtio -
(interface.delay + 1))
else:
counter_rtio = counter.value_rtio
# FIFO write
if data_width:
self.comb += fifo_in.data.eq(interface.data)
if interface.timestamped:
if fine_ts_width:
full_ts = Cat(interface.fine_ts, counter_rtio)
else:
full_ts = counter_rtio
self.comb += fifo_in.timestamp.eq(full_ts)
self.comb += fifo.we.eq(interface.stb)
# FIFO read
self.comb += [
self.ev.eq(fifo_out),
self.readable.eq(fifo.readable),
fifo.re.eq(self.re)
]
overflow_transfer = BlindTransfer()
self.submodules += overflow_transfer
self.comb += [
overflow_transfer.i.eq(fifo.we & ~fifo.writable),
self.overflow.eq(overflow_transfer.o),
]
def __init__(self, nchan=3, depth=8):
self.valid_i = Signal()
self.chan_synced = Signal()
self._r_channels_synced = CSRStatus()
lst_control = []
all_control = Signal()
for i in range(nchan):
name = "data_in" + str(i)
data_in = Record(channel_layout, name=name)
setattr(self, name, data_in)
name = "data_out" + str(i)
data_out = Record(channel_layout, name=name)
setattr(self, name, data_out)
###
syncbuffer = RenameClockDomains(_SyncBuffer(layout_len(channel_layout), depth), "pix")
self.submodules += syncbuffer
self.comb += [syncbuffer.din.eq(data_in.raw_bits()), data_out.raw_bits().eq(syncbuffer.dout)]
is_control = Signal()
self.comb += [is_control.eq(~data_out.de), syncbuffer.re.eq(~is_control | all_control)]
lst_control.append(is_control)
some_control = Signal()
self.comb += [all_control.eq(optree("&", lst_control)), some_control.eq(optree("|", lst_control))]
self.sync.pix += If(~self.valid_i, self.chan_synced.eq(0)).Else(
If(some_control, If(all_control, self.chan_synced.eq(1)).Else(self.chan_synced.eq(0)))
)
self.specials += MultiReg(self.chan_synced, self._r_channels_synced.status)
def __init__(self, interface, counter, fifo_depth):
data_width = rtlink.get_data_width(interface)
fine_ts_width = rtlink.get_fine_ts_width(interface)
ev_layout = []
if data_width:
ev_layout.append(("data", data_width))
if interface.timestamped:
ev_layout.append(("timestamp", counter.width + fine_ts_width))
self.ev = Record(ev_layout)
self.readable = Signal()
self.re = Signal()
self.overflow = Signal() # pulsed
# # #
fifo = ClockDomainsRenamer({"read": "rsys", "write": "rio"})(
AsyncFIFO(layout_len(ev_layout), fifo_depth))
self.submodules += fifo
fifo_in = Record(ev_layout)
fifo_out = Record(ev_layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# FIFO write
if data_width:
self.comb += fifo_in.data.eq(interface.data)
if interface.timestamped:
if fine_ts_width:
full_ts = Cat(interface.fine_ts, counter.value_rio)
else:
full_ts = counter.value_rio
self.comb += fifo_in.timestamp.eq(full_ts)
self.comb += fifo.we.eq(interface.stb)
# FIFO read
self.comb += [
self.ev.eq(fifo_out),
self.readable.eq(fifo.readable),
fifo.re.eq(self.re)
]
overflow_sync = PulseSynchronizer("rio", "rsys")
overflow_ack_sync = PulseSynchronizer("rsys", "rio")
self.submodules += overflow_sync, overflow_ack_sync
overflow_blind = Signal()
self.comb += overflow_sync.i.eq(fifo.we & ~fifo.writable & ~overflow_blind)
self.sync.rio += [
If(fifo.we & ~fifo.writable, overflow_blind.eq(1)),
If(overflow_ack_sync.o, overflow_blind.eq(0))
]
self.comb += [
overflow_ack_sync.i.eq(overflow_sync.o),
self.overflow.eq(overflow_sync.o)
]
def __init__(self, interface, counter, fifo_depth):
data_width = rtlink.get_data_width(interface)
fine_ts_width = rtlink.get_fine_ts_width(interface)
ev_layout = []
if data_width:
ev_layout.append(("data", data_width))
if interface.timestamped:
ev_layout.append(("timestamp", counter.width + fine_ts_width))
self.ev = Record(ev_layout)
self.readable = Signal()
self.re = Signal()
self.overflow = Signal() # pulsed
# # #
fifo = ClockDomainsRenamer({
"read": "rsys",
"write": "rio"
})(AsyncFIFO(layout_len(ev_layout), fifo_depth))
self.submodules += fifo
fifo_in = Record(ev_layout)
fifo_out = Record(ev_layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# FIFO write
if data_width:
self.comb += fifo_in.data.eq(interface.data)
if interface.timestamped:
if fine_ts_width:
full_ts = Cat(interface.fine_ts, counter.value_rtio)
else:
full_ts = counter.value_rtio
self.comb += fifo_in.timestamp.eq(full_ts)
self.comb += fifo.we.eq(interface.stb)
# FIFO read
self.comb += [
self.ev.eq(fifo_out),
self.readable.eq(fifo.readable),
fifo.re.eq(self.re)
]
overflow_transfer = _BlindTransfer()
self.submodules += overflow_transfer
self.comb += [
overflow_transfer.i.eq(fifo.we & ~fifo.writable),
self.overflow.eq(overflow_transfer.o),
]
def __init__(self, nchan=3, depth=8):
self.valid_i = Signal()
self.chan_synced = Signal()
self._channels_synced = CSRStatus()
lst_control = []
all_control = Signal()
for i in range(nchan):
name = "data_in" + str(i)
data_in = Record(channel_layout, name=name)
setattr(self, name, data_in)
name = "data_out" + str(i)
data_out = Record(channel_layout, name=name)
setattr(self, name, data_out)
# # #
syncbuffer = _SyncBuffer(layout_len(channel_layout), depth)
syncbuffer = ClockDomainsRenamer("pix")(syncbuffer)
self.submodules += syncbuffer
self.comb += [
syncbuffer.din.eq(data_in.raw_bits()),
data_out.raw_bits().eq(syncbuffer.dout)
]
is_control = Signal()
self.comb += [
is_control.eq(~data_out.de),
syncbuffer.re.eq(~is_control | all_control)
]
lst_control.append(is_control)
some_control = Signal()
self.comb += [
all_control.eq(reduce(and_, lst_control)),
some_control.eq(reduce(or_, lst_control))
]
self.sync.pix += \
If(~self.valid_i,
self.chan_synced.eq(0)
).Else(
If(some_control,
If(all_control,
self.chan_synced.eq(1)
).Else(
self.chan_synced.eq(0)
)
)
)
self.specials += MultiReg(self.chan_synced,
self._channels_synced.status)
def __init__(self, stream_slicer):
self.source = stream.Endpoint(record_layout)
self.end_marker_found = Signal()
self.flush = Signal()
hdrlen = (layout_len(record_layout) - 512)//8
record_raw = Record(record_layout)
self.comb += [
record_raw.raw_bits().eq(stream_slicer.source),
self.source.channel.eq(record_raw.channel),
self.source.timestamp.eq(record_raw.timestamp),
self.source.address.eq(record_raw.address),
Case(record_raw.length,
{hdrlen+i: self.source.data.eq(record_raw.data[:i*8])
for i in range(1, 512//8+1)}),
]
fsm = FSM(reset_state="FLOWING")
self.submodules += fsm
fsm.act("FLOWING",
If(stream_slicer.source_stb,
If(record_raw.length == 0,
NextState("END_MARKER_FOUND")
).Else(
self.source.stb.eq(1)
)
),
If(self.source.ack,
stream_slicer.source_consume.eq(record_raw.length)
)
)
fsm.act("END_MARKER_FOUND",
self.end_marker_found.eq(1),
If(self.flush,
stream_slicer.flush.eq(1),
NextState("WAIT_FLUSH")
)
)
fsm.act("WAIT_FLUSH",
If(stream_slicer.flush_done,
NextState("SEND_EOP")
)
)
fsm.act("SEND_EOP",
self.source.eop.eq(1),
self.source.stb.eq(1),
If(self.source.ack, NextState("FLOWING"))
)
def __init__(self, stream_slicer):
self.source = stream.Endpoint(record_layout)
self.end_marker_found = Signal()
self.flush = Signal()
hdrlen = (layout_len(record_layout) - 512)//8
record_raw = Record(record_layout)
self.comb += [
record_raw.raw_bits().eq(stream_slicer.source),
self.source.channel.eq(record_raw.channel),
self.source.timestamp.eq(record_raw.timestamp),
self.source.address.eq(record_raw.address),
Case(record_raw.length,
{hdrlen+i: self.source.data.eq(record_raw.data[:i*8])
for i in range(1, 512//8+1)}),
]
fsm = FSM(reset_state="FLOWING")
self.submodules += fsm
fsm.act("FLOWING",
If(stream_slicer.source_stb,
If(record_raw.length == 0,
NextState("END_MARKER_FOUND")
).Else(
self.source.stb.eq(1)
)
),
If(self.source.ack,
stream_slicer.source_consume.eq(record_raw.length)
)
)
fsm.act("END_MARKER_FOUND",
self.end_marker_found.eq(1),
If(self.flush,
stream_slicer.flush.eq(1),
NextState("WAIT_FLUSH")
)
)
fsm.act("WAIT_FLUSH",
If(stream_slicer.flush_done,
NextState("SEND_EOP")
)
)
fsm.act("SEND_EOP",
self.source.eop.eq(1),
self.source.stb.eq(1),
If(self.source.ack, NextState("FLOWING"))
)
def __init__(self):
# in pix clock domain
self.valid_i = Signal()
self.vsync = Signal()
self.de = Signal()
self.r = Signal(8)
self.g = Signal(8)
self.b = Signal(8)
# in sys clock domain
self.frame = Source(frame_layout)
self.busy = Signal()
###
fifo_stb = Signal()
fifo_in = Record(frame_layout)
self.comb += [
fifo_stb.eq(self.valid_i & self.de),
fifo_in.r.eq(self.r),
fifo_in.g.eq(self.g),
fifo_in.b.eq(self.b),
]
vsync_r = Signal()
self.sync.pix += [
If(self.vsync & ~vsync_r, fifo_in.parity.eq(~fifo_in.parity)),
vsync_r.eq(self.vsync)
]
fifo = AsyncFIFO(layout_len(frame_layout), 256)
self.add_submodule(fifo, {"write": "pix", "read": "sys"})
self.comb += [
fifo.we.eq(fifo_stb),
fifo.din.eq(fifo_in.raw_bits()),
self.frame.stb.eq(fifo.readable),
self.frame.payload.raw_bits().eq(fifo.dout),
fifo.re.eq(self.frame.ack),
self.busy.eq(0)
]
def __init__(self, cachesize, lasmim):
self.wishbone = wishbone.Interface()
###
if lasmim.dw <= 32:
raise ValueError("LASMI data width must be strictly larger than 32")
if (lasmim.dw % 32) != 0:
raise ValueError("LASMI data width must be a multiple of 32")
# Split address:
# TAG | LINE NUMBER | LINE OFFSET
offsetbits = log2_int(lasmim.dw//32)
addressbits = lasmim.aw + offsetbits
linebits = log2_int(cachesize) - offsetbits
tagbits = addressbits - linebits
adr_offset, adr_line, adr_tag = split(self.wishbone.adr, offsetbits, linebits, tagbits)
# Data memory
data_mem = Memory(lasmim.dw, 2**linebits)
data_port = data_mem.get_port(write_capable=True, we_granularity=8)
self.specials += data_mem, data_port
write_from_lasmi = Signal()
write_to_lasmi = Signal()
adr_offset_r = Signal(offsetbits)
self.comb += [
data_port.adr.eq(adr_line),
If(write_from_lasmi,
data_port.dat_w.eq(lasmim.dat_r),
data_port.we.eq(Replicate(1, lasmim.dw//8))
).Else(
data_port.dat_w.eq(Replicate(self.wishbone.dat_w, lasmim.dw//32)),
If(self.wishbone.cyc & self.wishbone.stb & self.wishbone.we & self.wishbone.ack,
displacer(self.wishbone.sel, adr_offset, data_port.we, 2**offsetbits, reverse=True)
)
),
If(write_to_lasmi,
lasmim.dat_w.eq(data_port.dat_r),
lasmim.dat_we.eq(2**(lasmim.dw//8)-1)
),
chooser(data_port.dat_r, adr_offset_r, self.wishbone.dat_r, reverse=True)
]
self.sync += adr_offset_r.eq(adr_offset)
# Tag memory
tag_layout = [("tag", tagbits), ("dirty", 1)]
tag_mem = Memory(layout_len(tag_layout), 2**linebits)
tag_port = tag_mem.get_port(write_capable=True)
self.specials += tag_mem, tag_port
tag_do = Record(tag_layout)
tag_di = Record(tag_layout)
self.comb += [
tag_do.raw_bits().eq(tag_port.dat_r),
tag_port.dat_w.eq(tag_di.raw_bits())
]
self.comb += [
tag_port.adr.eq(adr_line),
tag_di.tag.eq(adr_tag),
lasmim.adr.eq(Cat(adr_line, tag_do.tag))
]
# Control FSM
assert(lasmim.write_latency >= 1 and lasmim.read_latency >= 1)
fsm = FSM("IDLE", "TEST_HIT",
"EVICT_REQUEST", "EVICT_WAIT_DATA_ACK", "EVICT_DATA",
"REFILL_WRTAG", "REFILL_REQUEST", "REFILL_WAIT_DATA_ACK", "REFILL_DATA",
delayed_enters=[
("EVICT_DATAD", "EVICT_DATA", lasmim.write_latency-1),
("REFILL_DATAD", "REFILL_DATA", lasmim.read_latency-1)
])
self.submodules += fsm
fsm.act(fsm.IDLE,
If(self.wishbone.cyc & self.wishbone.stb, fsm.next_state(fsm.TEST_HIT))
)
fsm.act(fsm.TEST_HIT,
If(tag_do.tag == adr_tag,
self.wishbone.ack.eq(1),
If(self.wishbone.we,
tag_di.dirty.eq(1),
tag_port.we.eq(1)
),
fsm.next_state(fsm.IDLE)
).Else(
If(tag_do.dirty,
fsm.next_state(fsm.EVICT_REQUEST)
).Else(
fsm.next_state(fsm.REFILL_WRTAG)
)
)
)
fsm.act(fsm.EVICT_REQUEST,
lasmim.stb.eq(1),
lasmim.we.eq(1),
If(lasmim.req_ack, fsm.next_state(fsm.EVICT_WAIT_DATA_ACK))
)
fsm.act(fsm.EVICT_WAIT_DATA_ACK,
If(lasmim.dat_ack, fsm.next_state(fsm.EVICT_DATAD))
)
fsm.act(fsm.EVICT_DATA,
write_to_lasmi.eq(1),
fsm.next_state(fsm.REFILL_WRTAG)
)
fsm.act(fsm.REFILL_WRTAG,
# Write the tag first to set the LASMI address
tag_port.we.eq(1),
fsm.next_state(fsm.REFILL_REQUEST)
)
fsm.act(fsm.REFILL_REQUEST,
lasmim.stb.eq(1),
If(lasmim.req_ack, fsm.next_state(fsm.REFILL_WAIT_DATA_ACK))
)
fsm.act(fsm.REFILL_WAIT_DATA_ACK,
If(lasmim.dat_ack, fsm.next_state(fsm.REFILL_DATAD))
)
fsm.act(fsm.REFILL_DATA,
write_from_lasmi.eq(1),
fsm.next_state(fsm.TEST_HIT)
)
def __init__(self, interface, counter, fifo_depth, guard_io_cycles):
data_width = rtlink.get_data_width(interface)
address_width = rtlink.get_address_width(interface)
fine_ts_width = rtlink.get_fine_ts_width(interface)
ev_layout = []
if data_width:
ev_layout.append(("data", data_width))
if address_width:
ev_layout.append(("address", address_width))
ev_layout.append(("timestamp", counter.width + fine_ts_width))
# ev must be valid 1 cycle before we to account for the latency in
# generating replace, sequence_error and collision
self.ev = Record(ev_layout)
self.writable = Signal()
self.we = Signal() # maximum throughput 1/2
self.underflow = Signal() # valid 1 cycle after we, pulsed
self.sequence_error = Signal()
self.collision = Signal()
self.busy = Signal() # pulsed
# # #
# FIFO
fifo = ClockDomainsRenamer({
"write": "rsys",
"read": "rio"
})(AsyncFIFO(layout_len(ev_layout), fifo_depth))
self.submodules += fifo
fifo_in = Record(ev_layout)
fifo_out = Record(ev_layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# Buffer
buf_pending = Signal()
buf = Record(ev_layout)
buf_just_written = Signal()
# Special cases
replace = Signal()
sequence_error = Signal()
collision = Signal()
any_error = Signal()
if interface.enable_replace:
# Note: replace may be asserted at the same time as collision
# when addresses are different. In that case, it is a collision.
self.sync.rsys += replace.eq(self.ev.timestamp == buf.timestamp)
# Detect sequence errors on coarse timestamps only
# so that they are mutually exclusive with collision errors.
self.sync.rsys += sequence_error.eq(
self.ev.timestamp[fine_ts_width:] < buf.timestamp[fine_ts_width:])
if interface.enable_replace:
if address_width:
different_addresses = self.ev.address != buf.address
else:
different_addresses = 0
if fine_ts_width:
self.sync.rsys += collision.eq(
(self.ev.timestamp[fine_ts_width:] ==
buf.timestamp[fine_ts_width:])
& ((self.ev.timestamp[:fine_ts_width] !=
buf.timestamp[:fine_ts_width])
| different_addresses))
else:
self.sync.rsys += collision.eq(self.ev.timestamp[fine_ts_width:] ==
buf.timestamp[fine_ts_width:])
self.comb += [
any_error.eq(sequence_error | collision),
self.sequence_error.eq(self.we & sequence_error),
self.collision.eq(self.we & collision)
]
# Buffer read and FIFO write
self.comb += fifo_in.eq(buf)
in_guard_time = Signal()
self.comb += in_guard_time.eq(
buf.timestamp[fine_ts_width:] < counter.value_sys +
guard_io_cycles)
self.sync.rsys += If(in_guard_time, buf_pending.eq(0))
self.comb += \
If(buf_pending,
If(in_guard_time,
If(buf_just_written,
self.underflow.eq(1)
).Else(
fifo.we.eq(1)
)
),
If(self.we & ~replace & ~any_error,
fifo.we.eq(1)
)
)
# Buffer write
# Must come after read to handle concurrent read+write properly
self.sync.rsys += [
buf_just_written.eq(0),
If(self.we & ~any_error, buf_just_written.eq(1), buf_pending.eq(1),
buf.eq(self.ev))
]
self.comb += self.writable.eq(fifo.writable)
# Buffer output of FIFO to improve timing
dout_stb = Signal()
dout_ack = Signal()
dout = Record(ev_layout)
self.sync.rio += \
If(fifo.re,
dout_stb.eq(1),
dout.eq(fifo_out)
).Elif(dout_ack,
dout_stb.eq(0)
)
self.comb += fifo.re.eq(fifo.readable & (~dout_stb | dout_ack))
# latency compensation
if interface.delay:
counter_rtio = Signal.like(counter.value_rtio)
self.sync.rtio += counter_rtio.eq(counter.value_rtio -
interface.delay + 1)
else:
counter_rtio = counter.value_rtio
# FIFO read through buffer
self.comb += [
dout_ack.eq(dout.timestamp[fine_ts_width:] == counter_rtio),
interface.stb.eq(dout_stb & dout_ack)
]
busy_transfer = BlindTransfer()
self.submodules += busy_transfer
self.comb += [
busy_transfer.i.eq(interface.stb & interface.busy),
self.busy.eq(busy_transfer.o),
]
if data_width:
self.comb += interface.data.eq(dout.data)
if address_width:
self.comb += interface.address.eq(dout.address)
if fine_ts_width:
self.comb += interface.fine_ts.eq(dout.timestamp[:fine_ts_width])
def __init__(self, tsc, cri, enable):
self.source = stream.Endpoint([("data", message_len)])
self.overflow = CSRStatus()
self.overflow_reset = CSR()
# # #
read_wait_event = cri.i_status[2]
read_wait_event_r = Signal()
read_done = Signal()
read_overflow = Signal()
self.sync += read_wait_event_r.eq(read_wait_event)
self.comb += \
If(read_wait_event_r & ~read_wait_event,
If(~cri.i_status[0], read_done.eq(1)),
If(cri.i_status[1], read_overflow.eq(1))
)
input_output_stb = Signal()
input_output = Record(input_output_layout)
self.comb += [
input_output.channel.eq(cri.chan_sel),
input_output.address_padding.eq(cri.o_address),
input_output.rtio_counter.eq(tsc.full_ts_cri),
If(cri.cmd == cri_commands["write"],
input_output.message_type.eq(MessageType.output.value),
input_output.timestamp.eq(cri.o_timestamp),
input_output.data.eq(cri.o_data)
).Else(
input_output.message_type.eq(MessageType.input.value),
input_output.timestamp.eq(cri.i_timestamp),
input_output.data.eq(cri.i_data)
),
input_output_stb.eq((cri.cmd == cri_commands["write"]) | read_done)
]
exception_stb = Signal()
exception = Record(exception_layout)
self.comb += [
exception.message_type.eq(MessageType.exception.value),
exception.channel.eq(cri.chan_sel),
exception.rtio_counter.eq(tsc.full_ts_cri),
]
just_written = Signal()
self.sync += just_written.eq(cri.cmd == cri_commands["write"])
self.comb += [
If(just_written & cri.o_status[1],
exception_stb.eq(1),
exception.exception_type.eq(ExceptionType.o_underflow.value)
),
If(read_overflow,
exception_stb.eq(1),
exception.exception_type.eq(ExceptionType.i_overflow.value)
)
]
stopped = Record(stopped_layout)
self.comb += [
stopped.message_type.eq(MessageType.stopped.value),
stopped.rtio_counter.eq(tsc.full_ts_cri),
]
enable_r = Signal()
stopping = Signal()
self.sync += [
enable_r.eq(enable),
If(~enable & enable_r, stopping.eq(1)),
If(~stopping,
If(exception_stb,
self.source.data.eq(exception.raw_bits())
).Else(
self.source.data.eq(input_output.raw_bits())
),
self.source.eop.eq(0),
self.source.stb.eq(enable &
(input_output_stb | exception_stb)),
If(self.overflow_reset.re, self.overflow.status.eq(0)),
If(self.source.stb & ~self.source.ack,
self.overflow.status.eq(1)
)
).Else(
self.source.data.eq(stopped.raw_bits()),
self.source.eop.eq(1),
self.source.stb.eq(1),
If(self.source.ack, stopping.eq(0))
)
]
def __init__(self, cachesize, lasmim):
self.wishbone = wishbone.Interface()
###
data_width = flen(self.wishbone.dat_r)
if lasmim.dw > data_width and (lasmim.dw % data_width) != 0:
raise ValueError(
"LASMI data width must be a multiple of {dw}".format(
dw=data_width))
if lasmim.dw < data_width and (data_width % lasmim.dw) != 0:
raise ValueError(
"WISHBONE data width must be a multiple of {dw}".format(
dw=lasmim.dw))
# Split address:
# TAG | LINE NUMBER | LINE OFFSET
offsetbits = log2_int(max(lasmim.dw // data_width, 1))
addressbits = lasmim.aw + offsetbits
linebits = log2_int(cachesize) - offsetbits
tagbits = addressbits - linebits
wordbits = log2_int(max(data_width // lasmim.dw, 1))
adr_offset, adr_line, adr_tag = split(self.wishbone.adr, offsetbits,
linebits, tagbits)
word = Signal(wordbits) if wordbits else None
# Data memory
data_mem = Memory(lasmim.dw * 2**wordbits, 2**linebits)
data_port = data_mem.get_port(write_capable=True, we_granularity=8)
self.specials += data_mem, data_port
write_from_lasmi = Signal()
write_to_lasmi = Signal()
if adr_offset is None:
adr_offset_r = None
else:
adr_offset_r = Signal(offsetbits)
self.sync += adr_offset_r.eq(adr_offset)
self.comb += [
data_port.adr.eq(adr_line),
If(write_from_lasmi, displacer(lasmim.dat_r, word,
data_port.dat_w),
displacer(Replicate(1, lasmim.dw // 8), word,
data_port.we)).Else(
data_port.dat_w.eq(
Replicate(self.wishbone.dat_w,
max(lasmim.dw // data_width, 1))),
If(
self.wishbone.cyc & self.wishbone.stb
& self.wishbone.we & self.wishbone.ack,
displacer(self.wishbone.sel,
adr_offset,
data_port.we,
2**offsetbits,
reverse=True))),
If(write_to_lasmi, chooser(data_port.dat_r, word, lasmim.dat_w),
lasmim.dat_we.eq(2**(lasmim.dw // 8) - 1)),
chooser(data_port.dat_r,
adr_offset_r,
self.wishbone.dat_r,
reverse=True)
]
# Tag memory
tag_layout = [("tag", tagbits), ("dirty", 1)]
tag_mem = Memory(layout_len(tag_layout), 2**linebits)
tag_port = tag_mem.get_port(write_capable=True)
self.specials += tag_mem, tag_port
tag_do = Record(tag_layout)
tag_di = Record(tag_layout)
self.comb += [
tag_do.raw_bits().eq(tag_port.dat_r),
tag_port.dat_w.eq(tag_di.raw_bits())
]
self.comb += [tag_port.adr.eq(adr_line), tag_di.tag.eq(adr_tag)]
if word is not None:
self.comb += lasmim.adr.eq(Cat(word, adr_line, tag_do.tag))
else:
self.comb += lasmim.adr.eq(Cat(adr_line, tag_do.tag))
# Lasmim word computation, word_clr and word_inc will be simplified
# at synthesis when wordbits=0
word_clr = Signal()
word_inc = Signal()
if word is not None:
self.sync += \
If(word_clr,
word.eq(0),
).Elif(word_inc,
word.eq(word+1)
)
def word_is_last(word):
if word is not None:
return word == 2**wordbits - 1
else:
return 1
# Control FSM
assert (lasmim.write_latency >= 1 and lasmim.read_latency >= 1)
fsm = FSM(reset_state="IDLE")
self.submodules += fsm
fsm.delayed_enter("EVICT_DATAD", "EVICT_DATA",
lasmim.write_latency - 1)
fsm.delayed_enter("REFILL_DATAD", "REFILL_DATA",
lasmim.read_latency - 1)
fsm.act(
"IDLE",
If(self.wishbone.cyc & self.wishbone.stb, NextState("TEST_HIT")))
fsm.act(
"TEST_HIT", word_clr.eq(1),
If(tag_do.tag == adr_tag, self.wishbone.ack.eq(1),
If(self.wishbone.we, tag_di.dirty.eq(1), tag_port.we.eq(1)),
NextState("IDLE")).Else(
If(tag_do.dirty, NextState("EVICT_REQUEST")).Else(
NextState("REFILL_WRTAG"))))
fsm.act("EVICT_REQUEST", lasmim.stb.eq(1), lasmim.we.eq(1),
If(lasmim.req_ack, NextState("EVICT_WAIT_DATA_ACK")))
fsm.act("EVICT_WAIT_DATA_ACK",
If(lasmim.dat_ack, NextState("EVICT_DATAD")))
fsm.act(
"EVICT_DATA", write_to_lasmi.eq(1), word_inc.eq(1),
If(
word_is_last(word),
NextState("REFILL_WRTAG"),
).Else(NextState("EVICT_REQUEST")))
fsm.act(
"REFILL_WRTAG",
# Write the tag first to set the LASMI address
tag_port.we.eq(1),
word_clr.eq(1),
NextState("REFILL_REQUEST"))
fsm.act("REFILL_REQUEST", lasmim.stb.eq(1),
If(lasmim.req_ack, NextState("REFILL_WAIT_DATA_ACK")))
fsm.act("REFILL_WAIT_DATA_ACK",
If(lasmim.dat_ack, NextState("REFILL_DATAD")))
fsm.act(
"REFILL_DATA", write_from_lasmi.eq(1), word_inc.eq(1),
If(
word_is_last(word),
NextState("TEST_HIT"),
).Else(NextState("REFILL_REQUEST")))
def __init__(self, interface, counter, fifo_depth, guard_io_cycles):
data_width = rtlink.get_data_width(interface)
address_width = rtlink.get_address_width(interface)
fine_ts_width = rtlink.get_fine_ts_width(interface)
ev_layout = []
if data_width:
ev_layout.append(("data", data_width))
if address_width:
ev_layout.append(("address", address_width))
ev_layout.append(("timestamp", counter.width + fine_ts_width))
# ev must be valid 1 cycle before we to account for the latency in
# generating replace, sequence_error and nop
self.ev = Record(ev_layout)
self.writable = Signal()
self.we = Signal() # maximum throughput 1/2
self.underflow = Signal() # valid 1 cycle after we, pulsed
self.sequence_error = Signal()
self.collision_error = Signal()
# # #
# FIFO
fifo = ClockDomainsRenamer({"write": "rsys", "read": "rio"})(
AsyncFIFO(layout_len(ev_layout), fifo_depth))
self.submodules += fifo
fifo_in = Record(ev_layout)
fifo_out = Record(ev_layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# Buffer
buf_pending = Signal()
buf = Record(ev_layout)
buf_just_written = Signal()
# Special cases
replace = Signal()
sequence_error = Signal()
collision_error = Signal()
any_error = Signal()
nop = Signal()
self.sync.rsys += [
# Note: replace does not perform any RTLink address checks,
# i.e. a write to a different address will be silently replaced
# as well.
replace.eq(self.ev.timestamp == buf.timestamp),
# Detect sequence errors on coarse timestamps only
# so that they are mutually exclusive with collision errors.
sequence_error.eq(self.ev.timestamp[fine_ts_width:]
< buf.timestamp[fine_ts_width:])
]
if fine_ts_width:
self.sync.rsys += collision_error.eq(
(self.ev.timestamp[fine_ts_width:] == buf.timestamp[fine_ts_width:])
& (self.ev.timestamp[:fine_ts_width] != buf.timestamp[:fine_ts_width]))
self.comb += any_error.eq(sequence_error | collision_error)
if interface.suppress_nop:
# disable NOP at reset: do not suppress a first write with all 0s
nop_en = Signal(reset=0)
addresses_equal = [getattr(self.ev, a) == getattr(buf, a)
for a in ("data", "address")
if hasattr(self.ev, a)]
if addresses_equal:
self.sync.rsys += nop.eq(
nop_en & reduce(and_, addresses_equal))
else:
self.comb.eq(nop.eq(0))
self.sync.rsys += [
# buf now contains valid data. enable NOP.
If(self.we & ~any_error, nop_en.eq(1)),
# underflows cancel the write. allow it to be retried.
If(self.underflow, nop_en.eq(0))
]
self.comb += [
self.sequence_error.eq(self.we & sequence_error),
self.collision_error.eq(self.we & collision_error)
]
# Buffer read and FIFO write
self.comb += fifo_in.eq(buf)
in_guard_time = Signal()
self.comb += in_guard_time.eq(
buf.timestamp[fine_ts_width:]
< counter.value_sys + guard_io_cycles)
self.sync.rsys += If(in_guard_time, buf_pending.eq(0))
self.comb += \
If(buf_pending,
If(in_guard_time,
If(buf_just_written,
self.underflow.eq(1)
).Else(
fifo.we.eq(1)
)
),
If(self.we & ~replace & ~nop & ~any_error,
fifo.we.eq(1)
)
)
# Buffer write
# Must come after read to handle concurrent read+write properly
self.sync.rsys += [
buf_just_written.eq(0),
If(self.we & ~nop & ~any_error,
buf_just_written.eq(1),
buf_pending.eq(1),
buf.eq(self.ev)
)
]
self.comb += self.writable.eq(fifo.writable)
# Buffer output of FIFO to improve timing
dout_stb = Signal()
dout_ack = Signal()
dout = Record(ev_layout)
self.sync.rio += \
If(fifo.re,
dout_stb.eq(1),
dout.eq(fifo_out)
).Elif(dout_ack,
dout_stb.eq(0)
)
self.comb += fifo.re.eq(fifo.readable & (~dout_stb | dout_ack))
# FIFO read through buffer
# TODO: report error on stb & busy
self.comb += [
dout_ack.eq(
dout.timestamp[fine_ts_width:] == counter.value_rio),
interface.stb.eq(dout_stb & dout_ack)
]
if data_width:
self.comb += interface.data.eq(dout.data)
if address_width:
self.comb += interface.address.eq(dout.address)
if fine_ts_width:
self.comb += interface.fine_ts.eq(dout.timestamp[:fine_ts_width])
def __init__(self, tsc, channels, mode, quash_channels=[], interface=None):
if interface is None:
interface = cri.Interface()
self.cri = interface
# # #
if mode == "sync":
fifo_factory = SyncFIFOBuffered
sync_io = self.sync
sync_cri = self.sync
elif mode == "async":
fifo_factory = lambda *args: ClockDomainsRenamer({"write": "rio", "read": "rsys"})(AsyncFIFO(*args))
sync_io = self.sync.rio
sync_cri = self.sync.rsys
else:
raise ValueError
i_statuses, i_datas, i_timestamps = [], [], []
i_ack = Signal()
sel = self.cri.chan_sel[:16]
for n, channel in enumerate(channels):
iif = channel.interface.i
if iif is None or n in quash_channels:
i_datas.append(0)
i_timestamps.append(0)
i_statuses.append(0)
continue
# FIFO
layout = get_channel_layout(len(tsc.coarse_ts), iif)
fifo = fifo_factory(layout_len(layout), channel.ififo_depth)
self.submodules += fifo
fifo_in = Record(layout)
fifo_out = Record(layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# FIFO write
if iif.delay:
counter_rtio = Signal.like(tsc.coarse_ts, reset_less=True)
sync_io += counter_rtio.eq(tsc.coarse_ts - (iif.delay + 1))
else:
counter_rtio = tsc.coarse_ts
if hasattr(fifo_in, "data"):
self.comb += fifo_in.data.eq(iif.data)
if hasattr(fifo_in, "timestamp"):
if hasattr(iif, "fine_ts"):
full_ts = Cat(iif.fine_ts, counter_rtio)
else:
full_ts = counter_rtio
self.comb += fifo_in.timestamp.eq(full_ts)
self.comb += fifo.we.eq(iif.stb)
overflow_io = Signal()
self.comb += overflow_io.eq(fifo.we & ~fifo.writable)
if mode == "sync":
overflow_trigger = overflow_io
elif mode == "async":
overflow_transfer = BlindTransfer()
self.submodules += overflow_transfer
self.comb += overflow_transfer.i.eq(overflow_io)
overflow_trigger = overflow_transfer.o
else:
raise ValueError
# FIFO read, CRI connection
if hasattr(fifo_out, "data"):
i_datas.append(fifo_out.data)
else:
i_datas.append(0)
if hasattr(fifo_out, "timestamp"):
ts_shift = 64 - len(fifo_out.timestamp)
i_timestamps.append(fifo_out.timestamp << ts_shift)
else:
i_timestamps.append(0)
selected = Signal()
self.comb += selected.eq(sel == n)
overflow = Signal()
sync_cri += [
If(selected & i_ack,
overflow.eq(0)),
If(overflow_trigger,
overflow.eq(1))
]
self.comb += fifo.re.eq(selected & i_ack & ~overflow)
i_statuses.append(Cat(fifo.readable & ~overflow, overflow))
i_status_raw = Signal(2)
self.comb += i_status_raw.eq(Array(i_statuses)[sel])
input_timeout = Signal.like(self.cri.i_timeout, reset_less=True)
input_pending = Signal()
self.cri.i_data.reset_less = True
self.cri.i_timestamp.reset_less = True
sync_cri += [
i_ack.eq(0),
If(i_ack,
self.cri.i_status.eq(Cat(~i_status_raw[0], i_status_raw[1], 0)),
self.cri.i_data.eq(Array(i_datas)[sel]),
self.cri.i_timestamp.eq(Array(i_timestamps)[sel]),
),
If((tsc.full_ts_cri >= input_timeout) | (i_status_raw != 0),
If(input_pending, i_ack.eq(1)),
input_pending.eq(0)
),
If(self.cri.cmd == cri.commands["read"],
input_timeout.eq(self.cri.i_timeout),
input_pending.eq(1),
self.cri.i_status.eq(0b100)
)
]
def __init__(self, rtio_core):
self.source = stream.Endpoint([("data", 256)])
self.overflow = CSRStatus()
self.overflow_reset = CSR()
# # #
kcsrs = rtio_core.kcsrs
input_output_stb = Signal()
input_output = Record(input_output_layout)
if hasattr(kcsrs, "o_data"):
o_data = kcsrs.o_data.storage
else:
o_data = 0
if hasattr(kcsrs, "o_address"):
o_address = kcsrs.o_address.storage
else:
o_address = 0
if hasattr(kcsrs, "i_data"):
i_data = kcsrs.i_data.status
else:
i_data = 0
self.comb += [
input_output.channel.eq(kcsrs.chan_sel.storage),
input_output.address_padding.eq(o_address),
input_output.rtio_counter.eq(
rtio_core.counter.value_sys << rtio_core.fine_ts_width),
If(kcsrs.o_we.re,
input_output.message_type.eq(MessageType.output.value),
input_output.timestamp.eq(kcsrs.o_timestamp.storage),
input_output.data.eq(o_data)
).Else(
input_output.message_type.eq(MessageType.input.value),
input_output.timestamp.eq(kcsrs.i_timestamp.status),
input_output.data.eq(i_data)
),
input_output_stb.eq(kcsrs.o_we.re | kcsrs.i_re.re)
]
exception_stb = Signal()
exception = Record(exception_layout)
self.comb += [
exception.message_type.eq(MessageType.exception.value),
exception.channel.eq(kcsrs.chan_sel.storage),
exception.rtio_counter.eq(
rtio_core.counter.value_sys << rtio_core.fine_ts_width),
]
for ename in ("o_underflow_reset", "o_sequence_error_reset",
"o_collision_error_reset", "i_overflow_reset"):
self.comb += \
If(getattr(kcsrs, ename).re,
exception_stb.eq(1),
exception.exception_type.eq(
getattr(ExceptionType, ename).value)
)
for rname in "reset", "reset_phy":
r_d = Signal(reset=1)
r = getattr(kcsrs, rname).storage
self.sync += r_d.eq(r)
self.comb += [
If(r & ~r_d,
exception_stb.eq(1),
exception.exception_type.eq(
getattr(ExceptionType, rname+"_rising").value)
),
If(~r & r_d,
exception_stb.eq(1),
exception.exception_type.eq(
getattr(ExceptionType, rname+"_falling").value)
)
]
self.sync += [
If(exception_stb,
self.source.data.eq(exception.raw_bits())
).Else(
self.source.data.eq(input_output.raw_bits())
),
self.source.stb.eq(input_output_stb | exception_stb)
]
self.sync += [
If(self.overflow_reset.re, self.overflow.status.eq(0)),
If(self.source.stb & ~self.source.ack,
self.overflow.status.eq(1)
)
]
def __init__(self, tsc, cri, enable):
self.source = stream.Endpoint([("data", message_len)])
self.overflow = CSRStatus()
self.overflow_reset = CSR()
# # #
read_wait_event = cri.i_status[2]
read_wait_event_r = Signal()
read_done = Signal()
read_overflow = Signal()
self.sync += read_wait_event_r.eq(read_wait_event)
self.comb += \
If(read_wait_event_r & ~read_wait_event,
If(~cri.i_status[0], read_done.eq(1)),
If(cri.i_status[1], read_overflow.eq(1))
)
input_output_stb = Signal()
input_output = Record(input_output_layout)
self.comb += [
input_output.channel.eq(cri.chan_sel),
input_output.address_padding.eq(cri.o_address),
input_output.rtio_counter.eq(tsc.full_ts_cri),
If(cri.cmd == cri_commands["write"],
input_output.message_type.eq(MessageType.output.value),
input_output.timestamp.eq(cri.o_timestamp),
input_output.data.eq(cri.o_data)).Else(
input_output.message_type.eq(MessageType.input.value),
input_output.timestamp.eq(cri.i_timestamp),
input_output.data.eq(cri.i_data)),
input_output_stb.eq((cri.cmd == cri_commands["write"]) | read_done)
]
exception_stb = Signal()
exception = Record(exception_layout)
self.comb += [
exception.message_type.eq(MessageType.exception.value),
exception.channel.eq(cri.chan_sel),
exception.rtio_counter.eq(tsc.full_ts_cri),
]
just_written = Signal()
self.sync += just_written.eq(cri.cmd == cri_commands["write"])
self.comb += [
If(just_written & cri.o_status[1], exception_stb.eq(1),
exception.exception_type.eq(ExceptionType.o_underflow.value)),
If(read_overflow, exception_stb.eq(1),
exception.exception_type.eq(ExceptionType.i_overflow.value))
]
stopped = Record(stopped_layout)
self.comb += [
stopped.message_type.eq(MessageType.stopped.value),
stopped.rtio_counter.eq(tsc.full_ts_cri),
]
enable_r = Signal()
stopping = Signal()
self.sync += [
enable_r.eq(enable),
If(~enable & enable_r, stopping.eq(1)),
If(
~stopping,
If(exception_stb,
self.source.data.eq(exception.raw_bits())).Else(
self.source.data.eq(input_output.raw_bits())),
self.source.eop.eq(0),
self.source.stb.eq(enable
& (input_output_stb | exception_stb)),
If(self.overflow_reset.re, self.overflow.status.eq(0)),
If(self.source.stb & ~self.source.ack,
self.overflow.status.eq(1))).Else(
self.source.data.eq(stopped.raw_bits()),
self.source.eop.eq(1), self.source.stb.eq(1),
If(self.source.ack, stopping.eq(0)))
]
def __init__(self, rtio_core, enable):
self.source = stream.Endpoint([("data", message_len)])
self.overflow = CSRStatus()
self.overflow_reset = CSR()
# # #
kcsrs = rtio_core.kcsrs
input_output_stb = Signal()
input_output = Record(input_output_layout)
if hasattr(kcsrs, "o_data"):
o_data = kcsrs.o_data.storage
else:
o_data = 0
if hasattr(kcsrs, "o_address"):
o_address = kcsrs.o_address.storage
else:
o_address = 0
if hasattr(kcsrs, "i_data"):
i_data = kcsrs.i_data.status
else:
i_data = 0
self.comb += [
input_output.channel.eq(kcsrs.chan_sel.storage),
input_output.address_padding.eq(o_address),
input_output.rtio_counter.eq(
rtio_core.counter.value_sys << rtio_core.fine_ts_width),
If(kcsrs.o_we.re,
input_output.message_type.eq(MessageType.output.value),
input_output.timestamp.eq(kcsrs.o_timestamp.storage),
input_output.data.eq(o_data)).Else(
input_output.message_type.eq(MessageType.input.value),
input_output.timestamp.eq(kcsrs.i_timestamp.status),
input_output.data.eq(i_data)),
input_output_stb.eq(kcsrs.o_we.re | kcsrs.i_re.re)
]
exception_stb = Signal()
exception = Record(exception_layout)
self.comb += [
exception.message_type.eq(MessageType.exception.value),
exception.channel.eq(kcsrs.chan_sel.storage),
exception.rtio_counter.eq(
rtio_core.counter.value_sys << rtio_core.fine_ts_width),
]
for ename in ("o_underflow_reset", "o_sequence_error_reset",
"o_collision_reset", "i_overflow_reset"):
self.comb += \
If(getattr(kcsrs, ename).re,
exception_stb.eq(1),
exception.exception_type.eq(
getattr(ExceptionType, ename).value)
)
for rname in "reset", "reset_phy":
r_d = Signal(reset=1)
r = getattr(kcsrs, rname).storage
self.sync += r_d.eq(r)
self.comb += [
If(
r & ~r_d, exception_stb.eq(1),
exception.exception_type.eq(
getattr(ExceptionType, rname + "_rising").value)),
If(
~r & r_d, exception_stb.eq(1),
exception.exception_type.eq(
getattr(ExceptionType, rname + "_falling").value))
]
stopped = Record(stopped_layout)
self.comb += [
stopped.message_type.eq(MessageType.stopped.value),
stopped.rtio_counter.eq(
rtio_core.counter.value_sys << rtio_core.fine_ts_width),
]
enable_r = Signal()
stopping = Signal()
self.sync += [
enable_r.eq(enable),
If(~enable & enable_r, stopping.eq(1)),
If(
~stopping,
If(exception_stb,
self.source.data.eq(exception.raw_bits())).Else(
self.source.data.eq(input_output.raw_bits())),
self.source.eop.eq(0),
self.source.stb.eq(enable
& (input_output_stb | exception_stb)),
If(self.overflow_reset.re, self.overflow.status.eq(0)),
If(self.source.stb & ~self.source.ack,
self.overflow.status.eq(1))).Else(
self.source.data.eq(stopped.raw_bits()),
self.source.eop.eq(1), self.source.stb.eq(1),
If(self.source.ack, stopping.eq(0)))
]
def __init__(self, cachesize, lasmim):
self.wishbone = wishbone.Interface()
###
data_width = flen(self.wishbone.dat_r)
if lasmim.dw > data_width and (lasmim.dw % data_width) != 0:
raise ValueError("LASMI data width must be a multiple of {dw}".format(dw=data_width))
if lasmim.dw < data_width and (data_width % lasmim.dw) != 0:
raise ValueError("WISHBONE data width must be a multiple of {dw}".format(dw=lasmim.dw))
# Split address:
# TAG | LINE NUMBER | LINE OFFSET
offsetbits = log2_int(max(lasmim.dw//data_width, 1))
addressbits = lasmim.aw + offsetbits
linebits = log2_int(cachesize) - offsetbits
tagbits = addressbits - linebits
wordbits = data_width//lasmim.dw
adr_offset, adr_line, adr_tag = split(self.wishbone.adr, offsetbits, linebits, tagbits)
word = Signal(wordbits) if wordbits else None
# Data memory
data_mem = Memory(lasmim.dw*2**wordbits, 2**linebits)
data_port = data_mem.get_port(write_capable=True, we_granularity=8)
self.specials += data_mem, data_port
write_from_lasmi = Signal()
write_to_lasmi = Signal()
if adr_offset is None:
adr_offset_r = None
else:
adr_offset_r = Signal(offsetbits)
self.sync += adr_offset_r.eq(adr_offset)
self.comb += [
data_port.adr.eq(adr_line),
If(write_from_lasmi,
displacer(lasmim.dat_r, word, data_port.dat_w),
displacer(Replicate(1, lasmim.dw//8), word, data_port.we)
).Else(
data_port.dat_w.eq(Replicate(self.wishbone.dat_w, max(lasmim.dw//data_width, 1))),
If(self.wishbone.cyc & self.wishbone.stb & self.wishbone.we & self.wishbone.ack,
displacer(self.wishbone.sel, adr_offset, data_port.we, 2**offsetbits, reverse=True)
)
),
If(write_to_lasmi,
chooser(data_port.dat_r, word, lasmim.dat_w),
lasmim.dat_we.eq(2**(lasmim.dw//8)-1)
),
chooser(data_port.dat_r, adr_offset_r, self.wishbone.dat_r, reverse=True)
]
# Tag memory
tag_layout = [("tag", tagbits), ("dirty", 1)]
tag_mem = Memory(layout_len(tag_layout), 2**linebits)
tag_port = tag_mem.get_port(write_capable=True)
self.specials += tag_mem, tag_port
tag_do = Record(tag_layout)
tag_di = Record(tag_layout)
self.comb += [
tag_do.raw_bits().eq(tag_port.dat_r),
tag_port.dat_w.eq(tag_di.raw_bits())
]
self.comb += [
tag_port.adr.eq(adr_line),
tag_di.tag.eq(adr_tag)
]
if word is not None:
self.comb += lasmim.adr.eq(Cat(word, adr_line, tag_do.tag))
else:
self.comb += lasmim.adr.eq(Cat(adr_line, tag_do.tag))
# Lasmim word computation, word_clr and word_inc will be simplified
# at synthesis when wordbits=0
word_clr = Signal()
word_inc = Signal()
if word is not None:
self.sync += \
If(word_clr,
word.eq(0),
).Elif(word_inc,
word.eq(word+1)
)
def word_is_last(word):
if word is not None:
return word == 2**wordbits-1
else:
return 1
# Control FSM
assert(lasmim.write_latency >= 1 and lasmim.read_latency >= 1)
fsm = FSM(reset_state="IDLE")
self.submodules += fsm
fsm.delayed_enter("EVICT_DATAD", "EVICT_DATA", lasmim.write_latency-1)
fsm.delayed_enter("REFILL_DATAD", "REFILL_DATA", lasmim.read_latency-1)
fsm.act("IDLE",
If(self.wishbone.cyc & self.wishbone.stb, NextState("TEST_HIT"))
)
fsm.act("TEST_HIT",
word_clr.eq(1),
If(tag_do.tag == adr_tag,
self.wishbone.ack.eq(1),
If(self.wishbone.we,
tag_di.dirty.eq(1),
tag_port.we.eq(1)
),
NextState("IDLE")
).Else(
If(tag_do.dirty,
NextState("EVICT_REQUEST")
).Else(
NextState("REFILL_WRTAG")
)
)
)
fsm.act("EVICT_REQUEST",
lasmim.stb.eq(1),
lasmim.we.eq(1),
If(lasmim.req_ack, NextState("EVICT_WAIT_DATA_ACK"))
)
fsm.act("EVICT_WAIT_DATA_ACK",
If(lasmim.dat_ack, NextState("EVICT_DATAD"))
)
fsm.act("EVICT_DATA",
write_to_lasmi.eq(1),
word_inc.eq(1),
If(word_is_last(word),
NextState("REFILL_WRTAG"),
).Else(
NextState("EVICT_REQUEST")
)
)
fsm.act("REFILL_WRTAG",
# Write the tag first to set the LASMI address
tag_port.we.eq(1),
word_clr.eq(1),
NextState("REFILL_REQUEST")
)
fsm.act("REFILL_REQUEST",
lasmim.stb.eq(1),
If(lasmim.req_ack, NextState("REFILL_WAIT_DATA_ACK"))
)
fsm.act("REFILL_WAIT_DATA_ACK",
If(lasmim.dat_ack, NextState("REFILL_DATAD"))
)
fsm.act("REFILL_DATA",
write_from_lasmi.eq(1),
word_inc.eq(1),
If(word_is_last(word),
NextState("TEST_HIT"),
).Else(
NextState("REFILL_REQUEST")
)
)
def get_fragment(self):
comb = []
sync = []
aaw = self.asmiport.hub.aw
adw = self.asmiport.hub.dw
# Split address:
# TAG | LINE NUMBER | LINE OFFSET
offsetbits = log2_int(adw//32)
addressbits = aaw + offsetbits
linebits = log2_int(self.cachesize) - offsetbits
tagbits = addressbits - linebits
adr_offset, adr_line, adr_tag = split(self.wishbone.adr, offsetbits, linebits, tagbits)
# Data memory
data_mem = Memory(adw, 2**linebits)
data_port = data_mem.get_port(write_capable=True, we_granularity=8)
write_from_asmi = Signal()
write_to_asmi = Signal()
adr_offset_r = Signal(offsetbits)
comb += [
data_port.adr.eq(adr_line),
If(write_from_asmi,
data_port.dat_w.eq(self.asmiport.dat_r),
data_port.we.eq(Replicate(1, adw//8))
).Else(
data_port.dat_w.eq(Replicate(self.wishbone.dat_w, adw//32)),
If(self.wishbone.cyc & self.wishbone.stb & self.wishbone.we & self.wishbone.ack,
displacer(self.wishbone.sel, adr_offset, data_port.we, 2**offsetbits, reverse=True)
)
),
If(write_to_asmi, self.asmiport.dat_w.eq(data_port.dat_r)),
self.asmiport.dat_wm.eq(0),
chooser(data_port.dat_r, adr_offset_r, self.wishbone.dat_r, reverse=True)
]
sync += [
adr_offset_r.eq(adr_offset)
]
# Tag memory
tag_layout = [("tag", tagbits), ("dirty", 1)]
tag_mem = Memory(layout_len(tag_layout), 2**linebits)
tag_port = tag_mem.get_port(write_capable=True)
tag_do = Record(tag_layout)
tag_di = Record(tag_layout)
comb += [
tag_do.raw_bits().eq(tag_port.dat_r),
tag_port.dat_w.eq(tag_di.raw_bits())
]
comb += [
tag_port.adr.eq(adr_line),
tag_di.tag.eq(adr_tag),
self.asmiport.adr.eq(Cat(adr_line, tag_do.tag))
]
# Control FSM
write_to_asmi_pre = Signal()
sync.append(write_to_asmi.eq(write_to_asmi_pre))
fsm = FSM("IDLE", "TEST_HIT",
"EVICT_ISSUE", "EVICT_WAIT",
"REFILL_WRTAG", "REFILL_ISSUE", "REFILL_WAIT", "REFILL_COMPLETE")
fsm.act(fsm.IDLE,
If(self.wishbone.cyc & self.wishbone.stb, fsm.next_state(fsm.TEST_HIT))
)
fsm.act(fsm.TEST_HIT,
If(tag_do.tag == adr_tag,
self.wishbone.ack.eq(1),
If(self.wishbone.we,
tag_di.dirty.eq(1),
tag_port.we.eq(1)
),
fsm.next_state(fsm.IDLE)
).Else(
If(tag_do.dirty,
fsm.next_state(fsm.EVICT_ISSUE)
).Else(
fsm.next_state(fsm.REFILL_WRTAG)
)
)
)
fsm.act(fsm.EVICT_ISSUE,
self.asmiport.stb.eq(1),
self.asmiport.we.eq(1),
If(self.asmiport.ack, fsm.next_state(fsm.EVICT_WAIT))
)
fsm.act(fsm.EVICT_WAIT,
# Data is actually sampled by the memory controller in the next state.
# But since the data memory has one cycle latency, it gets the data
# at the address given during this cycle.
If(self.asmiport.get_call_expression(),
write_to_asmi_pre.eq(1),
fsm.next_state(fsm.REFILL_WRTAG)
)
)
fsm.act(fsm.REFILL_WRTAG,
# Write the tag first to set the ASMI address
tag_port.we.eq(1),
fsm.next_state(fsm.REFILL_ISSUE)
)
fsm.act(fsm.REFILL_ISSUE,
self.asmiport.stb.eq(1),
If(self.asmiport.ack, fsm.next_state(fsm.REFILL_WAIT))
)
fsm.act(fsm.REFILL_WAIT,
If(self.asmiport.get_call_expression(), fsm.next_state(fsm.REFILL_COMPLETE))
)
fsm.act(fsm.REFILL_COMPLETE,
write_from_asmi.eq(1),
fsm.next_state(fsm.TEST_HIT)
)
return Fragment(comb, sync, specials={data_mem, tag_mem}) \
+ fsm.get_fragment()
def __init__(self, kcsrs, rtio_counter, enable):
self.source = stream.Endpoint([("data", message_len)])
self.overflow = CSRStatus()
self.overflow_reset = CSR()
# # #
input_output_stb = Signal()
input_output = Record(input_output_layout)
o_data = kcsrs.o_data.storage
o_address = kcsrs.o_address.storage
i_data = kcsrs.i_data.status
self.comb += [
input_output.channel.eq(kcsrs.chan_sel.storage),
input_output.address_padding.eq(o_address),
input_output.rtio_counter.eq(rtio_counter),
If(kcsrs.o_we.re,
input_output.message_type.eq(MessageType.output.value),
input_output.timestamp.eq(kcsrs.o_timestamp.storage),
input_output.data.eq(o_data)).Else(
input_output.message_type.eq(MessageType.input.value),
input_output.timestamp.eq(kcsrs.i_timestamp.status),
input_output.data.eq(i_data)),
input_output_stb.eq(kcsrs.o_we.re | kcsrs.i_re.re)
]
exception_stb = Signal()
exception = Record(exception_layout)
self.comb += [
exception.message_type.eq(MessageType.exception.value),
exception.channel.eq(kcsrs.chan_sel.storage),
exception.rtio_counter.eq(rtio_counter),
]
for ename in ("o_underflow_reset", "o_sequence_error_reset",
"o_collision_reset", "i_overflow_reset"):
self.comb += \
If(getattr(kcsrs, ename).re,
exception_stb.eq(1),
exception.exception_type.eq(
getattr(ExceptionType, ename).value)
)
stopped = Record(stopped_layout)
self.comb += [
stopped.message_type.eq(MessageType.stopped.value),
stopped.rtio_counter.eq(rtio_counter),
]
enable_r = Signal()
stopping = Signal()
self.sync += [
enable_r.eq(enable),
If(~enable & enable_r, stopping.eq(1)),
If(
~stopping,
If(exception_stb,
self.source.data.eq(exception.raw_bits())).Else(
self.source.data.eq(input_output.raw_bits())),
self.source.eop.eq(0),
self.source.stb.eq(enable
& (input_output_stb | exception_stb)),
If(self.overflow_reset.re, self.overflow.status.eq(0)),
If(self.source.stb & ~self.source.ack,
self.overflow.status.eq(1))).Else(
self.source.data.eq(stopped.raw_bits()),
self.source.eop.eq(1), self.source.stb.eq(1),
If(self.source.ack, stopping.eq(0)))
]
def __init__(self, interface, counter, fifo_depth, guard_io_cycles):
data_width = rtlink.get_data_width(interface)
address_width = rtlink.get_address_width(interface)
fine_ts_width = rtlink.get_fine_ts_width(interface)
ev_layout = []
if data_width:
ev_layout.append(("data", data_width))
if address_width:
ev_layout.append(("address", address_width))
ev_layout.append(("timestamp", counter.width + fine_ts_width))
# ev must be valid 1 cycle before we to account for the latency in
# generating replace, sequence_error and collision
self.ev = Record(ev_layout)
self.writable = Signal()
self.we = Signal() # maximum throughput 1/2
self.underflow = Signal() # valid 1 cycle after we, pulsed
self.sequence_error = Signal()
self.collision = Signal()
self.busy = Signal() # pulsed
# # #
# FIFO
fifo = ClockDomainsRenamer({"write": "rsys", "read": "rio"})(
AsyncFIFO(layout_len(ev_layout), fifo_depth))
self.submodules += fifo
fifo_in = Record(ev_layout)
fifo_out = Record(ev_layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# Buffer
buf_pending = Signal()
buf = Record(ev_layout)
buf_just_written = Signal()
# Special cases
replace = Signal()
sequence_error = Signal()
collision = Signal()
any_error = Signal()
if interface.enable_replace:
# Note: replace may be asserted at the same time as collision
# when addresses are different. In that case, it is a collision.
self.sync.rsys += replace.eq(self.ev.timestamp == buf.timestamp)
# Detect sequence errors on coarse timestamps only
# so that they are mutually exclusive with collision errors.
self.sync.rsys += sequence_error.eq(self.ev.timestamp[fine_ts_width:] <
buf.timestamp[fine_ts_width:])
if interface.enable_replace:
if address_width:
different_addresses = self.ev.address != buf.address
else:
different_addresses = 0
if fine_ts_width:
self.sync.rsys += collision.eq(
(self.ev.timestamp[fine_ts_width:] == buf.timestamp[fine_ts_width:])
& ((self.ev.timestamp[:fine_ts_width] != buf.timestamp[:fine_ts_width])
|different_addresses))
else:
self.sync.rsys += collision.eq(
self.ev.timestamp[fine_ts_width:] == buf.timestamp[fine_ts_width:])
self.comb += [
any_error.eq(sequence_error | collision),
self.sequence_error.eq(self.we & sequence_error),
self.collision.eq(self.we & collision)
]
# Buffer read and FIFO write
self.comb += fifo_in.eq(buf)
in_guard_time = Signal()
self.comb += in_guard_time.eq(
buf.timestamp[fine_ts_width:]
< counter.value_sys + guard_io_cycles)
self.sync.rsys += If(in_guard_time, buf_pending.eq(0))
self.comb += \
If(buf_pending,
If(in_guard_time,
If(buf_just_written,
self.underflow.eq(1)
).Else(
fifo.we.eq(1)
)
),
If(self.we & ~replace & ~any_error,
fifo.we.eq(1)
)
)
# Buffer write
# Must come after read to handle concurrent read+write properly
self.sync.rsys += [
buf_just_written.eq(0),
If(self.we & ~any_error,
buf_just_written.eq(1),
buf_pending.eq(1),
buf.eq(self.ev)
)
]
self.comb += self.writable.eq(fifo.writable)
# Buffer output of FIFO to improve timing
dout_stb = Signal()
dout_ack = Signal()
dout = Record(ev_layout)
self.sync.rio += \
If(fifo.re,
dout_stb.eq(1),
dout.eq(fifo_out)
).Elif(dout_ack,
dout_stb.eq(0)
)
self.comb += fifo.re.eq(fifo.readable & (~dout_stb | dout_ack))
# FIFO read through buffer
self.comb += [
dout_ack.eq(
dout.timestamp[fine_ts_width:] == counter.value_rtio),
interface.stb.eq(dout_stb & dout_ack)
]
busy_transfer = _BlindTransfer()
self.submodules += busy_transfer
self.comb += [
busy_transfer.i.eq(interface.stb & interface.busy),
self.busy.eq(busy_transfer.o),
]
if data_width:
self.comb += interface.data.eq(dout.data)
if address_width:
self.comb += interface.address.eq(dout.address)
if fine_ts_width:
self.comb += interface.fine_ts.eq(dout.timestamp[:fine_ts_width])
def __init__(self,
channels,
glbl_fine_ts_width,
mode,
quash_channels=[],
interface=None):
if interface is None:
interface = cri.Interface()
self.cri = interface
self.coarse_timestamp = Signal(64 - glbl_fine_ts_width)
# # #
if mode == "sync":
fifo_factory = SyncFIFOBuffered
sync_io = self.sync
sync_cri = self.sync
elif mode == "async":
fifo_factory = lambda *args: ClockDomainsRenamer({
"write": "rio",
"read": "rsys"
})(AsyncFIFO(*args))
sync_io = self.sync.rio
sync_cri = self.sync.rsys
else:
raise ValueError
i_statuses, i_datas, i_timestamps = [], [], []
i_ack = Signal()
sel = self.cri.chan_sel[:16]
for n, channel in enumerate(channels):
iif = channel.interface.i
if iif is None or n in quash_channels:
i_datas.append(0)
i_timestamps.append(0)
i_statuses.append(0)
continue
# FIFO
layout = get_channel_layout(len(self.coarse_timestamp), iif)
fifo = fifo_factory(layout_len(layout), channel.ififo_depth)
self.submodules += fifo
fifo_in = Record(layout)
fifo_out = Record(layout)
self.comb += [
fifo.din.eq(fifo_in.raw_bits()),
fifo_out.raw_bits().eq(fifo.dout)
]
# FIFO write
if iif.delay:
counter_rtio = Signal.like(self.coarse_timestamp,
reset_less=True)
sync_io += counter_rtio.eq(self.coarse_timestamp -
(iif.delay + 1))
else:
counter_rtio = self.coarse_timestamp
if hasattr(fifo_in, "data"):
self.comb += fifo_in.data.eq(iif.data)
if hasattr(fifo_in, "timestamp"):
if hasattr(iif, "fine_ts"):
full_ts = Cat(iif.fine_ts, counter_rtio)
else:
full_ts = counter_rtio
self.comb += fifo_in.timestamp.eq(full_ts)
self.comb += fifo.we.eq(iif.stb)
overflow_io = Signal()
self.comb += overflow_io.eq(fifo.we & ~fifo.writable)
if mode == "sync":
overflow_trigger = overflow_io
elif mode == "async":
overflow_transfer = BlindTransfer()
self.submodules += overflow_transfer
self.comb += overflow_transfer.i.eq(overflow_io)
overflow_trigger = overflow_transfer.o
else:
raise ValueError
# FIFO read, CRI connection
if hasattr(fifo_out, "data"):
i_datas.append(fifo_out.data)
else:
i_datas.append(0)
if hasattr(fifo_out, "timestamp"):
ts_shift = 64 - len(fifo_out.timestamp)
i_timestamps.append(fifo_out.timestamp << ts_shift)
else:
i_timestamps.append(0)
selected = Signal()
self.comb += selected.eq(sel == n)
overflow = Signal()
sync_cri += [
If(selected & i_ack, overflow.eq(0)),
If(overflow_trigger, overflow.eq(1))
]
self.comb += fifo.re.eq(selected & i_ack & ~overflow)
i_statuses.append(Cat(fifo.readable & ~overflow, overflow))
i_status_raw = Signal(2)
self.comb += i_status_raw.eq(Array(i_statuses)[sel])
input_timeout = Signal.like(self.cri.timestamp, reset_less=True)
input_pending = Signal()
self.cri.i_data.reset_less = True
self.cri.i_timestamp.reset_less = True
sync_cri += [
i_ack.eq(0),
If(
i_ack,
self.cri.i_status.eq(Cat(~i_status_raw[0], i_status_raw[1],
0)),
self.cri.i_data.eq(Array(i_datas)[sel]),
self.cri.i_timestamp.eq(Array(i_timestamps)[sel]),
),
If((self.cri.counter >= input_timeout) | (i_status_raw != 0),
If(input_pending, i_ack.eq(1)), input_pending.eq(0)),
If(self.cri.cmd == cri.commands["read"],
input_timeout.eq(self.cri.timestamp), input_pending.eq(1),
self.cri.i_status.eq(0b100))
]
def __init__(self, cachesize, lasmim, wbm=None):
if wbm is None:
wbm = wishbone.Interface()
self.wishbone = wbm
###
data_width = flen(self.wishbone.dat_r)
if lasmim.dw < data_width:
raise ValueError("LASMI data width must be >= {dw}".format(dw=data_width))
if (lasmim.dw % data_width) != 0:
raise ValueError("LASMI data width must be a multiple of {dw}".format(dw=data_width))
# Split address:
# TAG | LINE NUMBER | LINE OFFSET
offsetbits = log2_int(lasmim.dw//data_width)
addressbits = lasmim.aw + offsetbits
linebits = log2_int(cachesize) - offsetbits
tagbits = addressbits - linebits
adr_offset, adr_line, adr_tag = split(self.wishbone.adr, offsetbits, linebits, tagbits)
# Data memory
data_mem = Memory(lasmim.dw, 2**linebits)
data_port = data_mem.get_port(write_capable=True, we_granularity=8)
self.specials += data_mem, data_port
write_from_lasmi = Signal()
write_to_lasmi = Signal()
if adr_offset is None:
adr_offset_r = None
else:
adr_offset_r = Signal(offsetbits)
self.sync += adr_offset_r.eq(adr_offset)
self.comb += [
data_port.adr.eq(adr_line),
If(write_from_lasmi,
data_port.dat_w.eq(lasmim.dat_r),
data_port.we.eq(Replicate(1, lasmim.dw//8))
).Else(
data_port.dat_w.eq(Replicate(self.wishbone.dat_w, lasmim.dw//data_width)),
If(self.wishbone.cyc & self.wishbone.stb & self.wishbone.we & self.wishbone.ack,
displacer(self.wishbone.sel, adr_offset, data_port.we, 2**offsetbits, reverse=True)
)
),
If(write_to_lasmi,
lasmim.dat_w.eq(data_port.dat_r),
lasmim.dat_we.eq(2**(lasmim.dw//8)-1)
),
chooser(data_port.dat_r, adr_offset_r, self.wishbone.dat_r, reverse=True)
]
# Tag memory
tag_layout = [("tag", tagbits), ("dirty", 1)]
tag_mem = Memory(layout_len(tag_layout), 2**linebits)
tag_port = tag_mem.get_port(write_capable=True)
self.specials += tag_mem, tag_port
tag_do = Record(tag_layout)
tag_di = Record(tag_layout)
self.comb += [
tag_do.raw_bits().eq(tag_port.dat_r),
tag_port.dat_w.eq(tag_di.raw_bits())
]
self.comb += [
tag_port.adr.eq(adr_line),
tag_di.tag.eq(adr_tag),
lasmim.adr.eq(Cat(adr_line, tag_do.tag))
]
# Control FSM
assert(lasmim.write_latency >= 1 and lasmim.read_latency >= 1)
fsm = FSM()
self.submodules += fsm
fsm.delayed_enter("EVICT_DATAD", "EVICT_DATA", lasmim.write_latency-1)
fsm.delayed_enter("REFILL_DATAD", "REFILL_DATA", lasmim.read_latency-1)
fsm.act("IDLE",
If(self.wishbone.cyc & self.wishbone.stb, NextState("TEST_HIT"))
)
fsm.act("TEST_HIT",
If(tag_do.tag == adr_tag,
self.wishbone.ack.eq(1),
If(self.wishbone.we,
tag_di.dirty.eq(1),
tag_port.we.eq(1)
),
NextState("IDLE")
).Else(
If(tag_do.dirty,
NextState("EVICT_REQUEST")
).Else(
NextState("REFILL_WRTAG")
)
)
)
fsm.act("EVICT_REQUEST",
lasmim.stb.eq(1),
lasmim.we.eq(1),
If(lasmim.req_ack, NextState("EVICT_WAIT_DATA_ACK"))
)
fsm.act("EVICT_WAIT_DATA_ACK",
If(lasmim.dat_ack, NextState("EVICT_DATAD"))
)
fsm.act("EVICT_DATA",
write_to_lasmi.eq(1),
NextState("REFILL_WRTAG")
)
fsm.act("REFILL_WRTAG",
# Write the tag first to set the LASMI address
tag_port.we.eq(1),
NextState("REFILL_REQUEST")
)
fsm.act("REFILL_REQUEST",
lasmim.stb.eq(1),
If(lasmim.req_ack, NextState("REFILL_WAIT_DATA_ACK"))
)
fsm.act("REFILL_WAIT_DATA_ACK",
If(lasmim.dat_ack, NextState("REFILL_DATAD"))
)
fsm.act("REFILL_DATA",
write_from_lasmi.eq(1),
NextState("TEST_HIT")
)