""" nmigen implementation of buffered pipeline stage, based on zipcpu:
https://zipcpu.com/blog/2017/08/14/strategies-for-pipelining.html
+
+ this module requires quite a bit of thought to understand how it works
+ (and why it is needed in the first place). reading the above is
+ *strongly* recommended.
+
+ unlike john dawson's IEEE754 FPU STB/ACK signalling, which requires
+ the STB / ACK signals to raise and lower (on separate clocks) before
+ data may proceeed (thus only allowing one piece of data to proceed
+ on *ALTERNATE* cycles), the signalling here is a true pipeline
+ where data will flow on *every* clock when the conditions are right.
+
+ input acceptance conditions are when:
+ * incoming previous-stage strobe (i.p_valid) is HIGH
+ * outgoing previous-stage ready (o.p_ready) is LOW
+
+ output transmission conditions are when:
+ * outgoing next-stage strobe (o.n_valid) is HIGH
+ * outgoing next-stage ready (i.n_ready) is LOW
+
+ the tricky bit is when the input has valid data and the output is not
+ ready to accept it. if it wasn't for the clock synchronisation, it
+ would be possible to tell the input "hey don't send that data, we're
+ not ready". unfortunately, it's not possible to "change the past":
+ the previous stage *has no choice* but to pass on its data.
+
+ therefore, the incoming data *must* be accepted - and stored: that
+ is the responsibility / contract that this stage *must* accept.
+ on the same clock, it's possible to tell the input that it must
+ not send any more data. this is the "stall" condition.
+
+ we now effectively have *two* possible pieces of data to "choose" from:
+ the buffered data, and the incoming data. the decision as to which
+ to process and output is based on whether we are in "stall" or not.
+ i.e. when the next stage is no longer ready, the output comes from
+ the buffer if a stall had previously occurred, otherwise it comes
+ direct from processing the input.
+
+ this allows us to respect a synchronous "travelling STB" with what
+ dan calls a "buffered handshake".
+
+ it's quite a complex state machine!
"""
+
from nmigen import Signal, Cat, Const, Mux, Module
-from nmigen.compat.sim import run_simulation
from nmigen.cli import verilog, rtlil
+from collections.abc import Sequence
+
+
+class IOAckIn:
+
+ def __init__(self):
+ self.p_valid = Signal() # >>in - comes in from PREVIOUS stage
+ self.n_ready = Signal() # in<< - comes in from the NEXT stage
-class BufPipe:
- """ buffered pipeline stage
- stage-1 i_p_stb >>in stage o_n_stb out>> stage+1
- stage-1 o_p_busy <<out stage i_n_busy <<in stage+1
- stage-1 i_data >>in stage o_data out>> stage+1
+class IOAckOut:
+
+ def __init__(self):
+ self.n_valid = Signal() # out>> - goes out to the NEXT stage
+ self.p_ready = Signal() # <<out - goes out to the PREVIOUS stage
+
+
+def eq(o, i):
+ if not isinstance(o, Sequence):
+ o, i = [o], [i]
+ res = []
+ for (ao, ai) in zip(o, i):
+ res.append(ao.eq(ai))
+ return res
+
+
+class BufferedPipeline:
+ """ buffered pipeline stage. data and strobe signals travel in sync.
+ if ever the input is ready and the output is not, processed data
+ is stored in a temporary register.
+
+ stage-1 i.p_valid >>in stage o.n_valid out>> stage+1
+ stage-1 o.p_ready <<out stage i.n_ready <<in stage+1
+ stage-1 i.data >>in stage o.data out>> stage+1
| |
- +-------> process
+ process --->----^
| |
- +-- r_data ---+
+ +-- r_data ->-+
+
+ input data i_data is read (only), is processed and goes into an
+ intermediate result store [process()]. this is updated combinatorially.
+
+ in a non-stall condition, the intermediate result will go into the
+ output (update_output). however if ever there is a stall, it goes
+ into r_data instead [update_buffer()].
+
+ when the non-stall condition is released, r_data is the first
+ to be transferred to the output [flush_buffer()], and the stall
+ condition cleared.
+
+ on the next cycle (as long as stall is not raised again) the
+ input may begin to be processed and transferred directly to output.
"""
- def __init__(self):
- # input
- #self.i_p_rst = Signal() # >>in - comes in from PREVIOUS stage
- self.i_p_stb = Signal() # >>in - comes in from PREVIOUS stage
- self.i_n_busy = Signal() # in<< - comes in from the NEXT stage
- self.i_data = Signal(32) # >>in - comes in from the PREVIOUS stage
- #self.i_rst = Signal()
+ def __init__(self, stage):
+ """ pass in a "stage" which may be either a static class or a class
+ instance, which has three functions:
+ * ispec: returns input signals according to the input specification
+ * ispec: returns output signals to the output specification
+ * process: takes an input instance and returns processed data
- # buffered
- self.r_data = Signal(32)
+ i_data -> process() -> result --> o.data
+ | ^
+ | |
+ +-> r_data -+
+ """
+ self.stage = stage
- # output
- self.o_n_stb = Signal() # out>> - goes out to the NEXT stage
- self.o_p_busy = Signal() # <<out - goes out to the PREVIOUS stage
- self.o_data = Signal(32) # out>> - goes out to the NEXT stage
+ # set up input and output IO ACK (prev/next ready/valid)
+ self.i = IOAckIn()
+ self.o = IOAckOut()
- def pre_process(self, d_in):
- return d_in | 0xf0000
+ # set up the input and output data
+ self.i.data = stage.ispec() # input type
+ self.r_data = stage.ospec() # all these are output type
+ self.result = stage.ospec()
+ self.o.data = stage.ospec()
+
+ def connect_next(self, nxt):
+ """ helper function to connect to the next stage data/valid/ready.
+ data/valid is passed *TO* nxt, and ready comes *IN* from nxt.
+ """
+ return [nxt.i.p_valid.eq(self.o.n_valid),
+ self.i.n_ready.eq(nxt.o.p_ready),
+ eq(nxt.i.data, self.o.data),
+ ]
- def process(self, d_in):
- return d_in + 1
+ def connect_in(self, prev):
+ """ helper function to connect stage to an input source. do not
+ use to connect stage-to-stage!
+ """
+ return [self.i.p_valid.eq(prev.i.p_valid),
+ prev.o.p_ready.eq(self.o.p_ready),
+ eq(self.i.data, prev.i.data),
+ ]
+
+ def connect_out(self, nxt):
+ """ helper function to connect stage to an output source. do not
+ use to connect stage-to-stage!
+ """
+ return [nxt.o.n_valid.eq(self.o.n_valid),
+ self.i.n_ready.eq(nxt.i.n_ready),
+ eq(nxt.o.data, self.o.data),
+ ]
+
+ def set_input(self, i):
+ """ helper function to set the input data
+ """
+ return eq(self.i.data, i)
+
+ def update_buffer(self):
+ """ copies the result into the intermediate register r_data,
+ which will need to be outputted on a subsequent cycle
+ prior to allowing "normal" operation.
+ """
+ return eq(self.r_data, self.result)
+
+ def update_output(self):
+ """ copies the (combinatorial) result into the output
+ """
+ return eq(self.o.data, self.result)
+
+ def flush_buffer(self):
+ """ copies the *intermediate* register r_data into the output
+ """
+ return eq(self.o.data, self.r_data)
+
+ def ports(self):
+ return [self.i.data, self.o.data]
def elaborate(self, platform):
m = Module()
- o_p_busyn = Signal(reset_less=True)
- i_p_stb_o_p_busyn = Signal(reset_less=True)
- m.d.comb += o_p_busyn.eq(~self.o_p_busy)
- m.d.comb += i_p_stb_o_p_busyn.eq(self.i_p_stb & o_p_busyn)
-
- result = Signal(32)
- m.d.comb += result.eq(self.process(self.i_data))
- with m.If(o_p_busyn): # not stalled
- m.d.sync += self.r_data.eq(result)
-
- #with m.If(self.i_p_rst): # reset
- # m.d.sync += self.o_n_stb.eq(0)
- # m.d.sync += self.o_p_busy.eq(0)
- with m.If(~self.i_n_busy): # previous stage is not busy
- with m.If(o_p_busyn): # not stalled
- # nothing in buffer: send input direct to output
- m.d.sync += self.o_n_stb.eq(self.i_p_stb)
- m.d.sync += self.o_data.eq(result)
- with m.Else(): # o_p_busy is true, and something is in our buffer.
- # Flush the buffer to the output port.
- m.d.sync += self.o_n_stb.eq(1)
- m.d.sync += self.o_data.eq(self.r_data)
- # ignore input, since o_p_busy is also true.
- # also clear stall condition, declare register to be empty.
- m.d.sync += self.o_p_busy.eq(0)
-
- # (i_n_busy) is true here: previous stage is busy
- with m.Elif(~self.o_n_stb): # next stage being told "not busy"
- m.d.sync += self.o_n_stb.eq(self.i_p_stb)
- m.d.sync += self.o_p_busy.eq(0) # Keep the buffer empty
- # Apply the logic to the input data, and set the output data
- m.d.sync += self.o_data.eq(result)
-
- # (i_n_busy) and (o_n_stb) both true:
- with m.Elif(i_p_stb_o_p_busyn):
- # If next stage *is* busy, and not stalled yet, accept input
- m.d.sync += self.o_p_busy.eq(self.i_p_stb & self.o_n_stb)
-
- with m.If(o_p_busyn): # not stalled
- # turns out that from all of the above conditions, just
- # always put result into buffer if not busy
- m.d.sync += self.r_data.eq(result)
+ # establish some combinatorial temporaries
+ o_n_validn = Signal(reset_less=True)
+ i_p_valid_o_p_ready = Signal(reset_less=True)
+ m.d.comb += [o_n_validn.eq(~self.o.n_valid),
+ i_p_valid_o_p_ready.eq(self.i.p_valid & self.o.p_ready),
+ ]
+
+ # store result of processing in combinatorial temporary
+ with m.If(self.i.p_valid): # input is valid: process it
+ m.d.comb += eq(self.result, self.stage.process(self.i.data))
+ # if not in stall condition, update the temporary register
+ with m.If(self.o.p_ready): # not stalled
+ m.d.sync += self.update_buffer()
+
+ #with m.If(self.i.p_rst): # reset
+ # m.d.sync += self.o.n_valid.eq(0)
+ # m.d.sync += self.o.p_ready.eq(0)
+ with m.If(self.i.n_ready): # next stage is ready
+ with m.If(self.o.p_ready): # not stalled
+ # nothing in buffer: send (processed) input direct to output
+ m.d.sync += [self.o.n_valid.eq(self.i.p_valid),
+ self.update_output(),
+ ]
+ with m.Else(): # o.p_ready is false, and something is in buffer.
+ # Flush the [already processed] buffer to the output port.
+ m.d.sync += [self.o.n_valid.eq(1),
+ self.flush_buffer(),
+ # clear stall condition, declare register empty.
+ self.o.p_ready.eq(1),
+ ]
+ # ignore input, since o.p_ready is also false.
+
+ # (i.n_ready) is false here: next stage is ready
+ with m.Elif(o_n_validn): # next stage being told "ready"
+ m.d.sync += [self.o.n_valid.eq(self.i.p_valid),
+ self.o.p_ready.eq(1), # Keep the buffer empty
+ # set the output data (from comb result)
+ self.update_output(),
+ ]
+ # (i.n_ready) false and (o.n_valid) true:
+ with m.Elif(i_p_valid_o_p_ready):
+ # If next stage *is* ready, and not stalled yet, accept input
+ m.d.sync += self.o.p_ready.eq(~(self.i.p_valid & self.o.n_valid))
return m
def ports(self):
- return [self.i_p_stb, self.i_n_busy, self.i_data,
- self.r_data,
- self.o_n_stb, self.o_p_busy, self.o_data
+ return [self.i.p_valid, self.i.n_ready,
+ self.o.n_valid, self.o.p_ready,
]
-def testbench(dut):
- #yield dut.i_p_rst.eq(1)
- yield dut.i_n_busy.eq(1)
- yield dut.o_p_busy.eq(1)
- yield
- yield
- #yield dut.i_p_rst.eq(0)
- yield dut.i_n_busy.eq(0)
- yield dut.i_data.eq(5)
- yield dut.i_p_stb.eq(1)
- yield
- yield dut.i_data.eq(7)
- yield
- yield dut.i_data.eq(2)
- yield
- yield dut.i_n_busy.eq(1)
- yield dut.i_data.eq(9)
- yield
- yield dut.i_p_stb.eq(0)
- yield dut.i_data.eq(12)
- yield
- yield dut.i_data.eq(32)
- yield dut.i_n_busy.eq(0)
- yield
- yield
- yield
- yield
+class ExampleAddStage:
+ """ an example of how to use the buffered pipeline, as a class instance
+ """
+
+ def ispec(self):
+ """ returns a tuple of input signals which will be the incoming data
+ """
+ return (Signal(16), Signal(16))
+
+ def ospec(self):
+ """ returns an output signal which will happen to contain the sum
+ of the two inputs
+ """
+ return Signal(16)
+
+ def process(self, i):
+ """ process the input data (sums the values in the tuple) and returns it
+ """
+ return i[0] + i[1]
+
+
+class ExampleBufPipeAdd(BufferedPipeline):
+ """ an example of how to use the buffered pipeline, using a class instance
+ """
+
+ def __init__(self):
+ addstage = ExampleAddStage()
+ BufferedPipeline.__init__(self, addstage)
+
+
+class ExampleStage:
+ """ an example of how to use the buffered pipeline, in a static class
+ fashion
+ """
+
+ def ispec():
+ return Signal(16)
+
+ def ospec():
+ return Signal(16)
+
+ def process(i):
+ """ process the input data and returns it (adds 1)
+ """
+ return i + 1
+
+
+class ExampleBufPipe(BufferedPipeline):
+ """ an example of how to use the buffered pipeline.
+ """
+
+ def __init__(self):
+ BufferedPipeline.__init__(self, ExampleStage)
if __name__ == '__main__':
- dut = BufPipe()
+ dut = ExampleBufPipe()
vl = rtlil.convert(dut, ports=dut.ports())
with open("test_bufpipe.il", "w") as f:
f.write(vl)
- run_simulation(dut, testbench(dut), vcd_name="test_bufpipe.vcd")
-
projects / ieee754fpu.git / blobdiff