1 """ nmigen implementation of buffered pipeline stage, based on zipcpu:
2 https://zipcpu.com/blog/2017/08/14/strategies-for-pipelining.html
4 this module requires quite a bit of thought to understand how it works
5 (and why it is needed in the first place). reading the above is
6 *strongly* recommended.
8 unlike john dawson's IEEE754 FPU STB/ACK signalling, which requires
9 the STB / ACK signals to raise and lower (on separate clocks) before
10 data may proceeed (thus only allowing one piece of data to proceed
11 on *ALTERNATE* cycles), the signalling here is a true pipeline
12 where data will flow on *every* clock when the conditions are right.
14 input acceptance conditions are when:
15 * incoming previous-stage strobe (p.i_valid) is HIGH
16 * outgoing previous-stage ready (p.o_ready) is LOW
18 output transmission conditions are when:
19 * outgoing next-stage strobe (n.o_valid) is HIGH
20 * outgoing next-stage ready (n.i_ready) is LOW
22 the tricky bit is when the input has valid data and the output is not
23 ready to accept it. if it wasn't for the clock synchronisation, it
24 would be possible to tell the input "hey don't send that data, we're
25 not ready". unfortunately, it's not possible to "change the past":
26 the previous stage *has no choice* but to pass on its data.
28 therefore, the incoming data *must* be accepted - and stored: that
29 is the responsibility / contract that this stage *must* accept.
30 on the same clock, it's possible to tell the input that it must
31 not send any more data. this is the "stall" condition.
33 we now effectively have *two* possible pieces of data to "choose" from:
34 the buffered data, and the incoming data. the decision as to which
35 to process and output is based on whether we are in "stall" or not.
36 i.e. when the next stage is no longer ready, the output comes from
37 the buffer if a stall had previously occurred, otherwise it comes
38 direct from processing the input.
40 this allows us to respect a synchronous "travelling STB" with what
41 dan calls a "buffered handshake".
43 it's quite a complex state machine!
46 from nmigen
import Signal
, Cat
, Const
, Mux
, Module
47 from nmigen
.cli
import verilog
, rtlil
48 from collections
.abc
import Sequence
52 """ contains signals that come *from* the previous stage (both in and out)
53 * i_valid: input from previous stage indicating incoming data is valid
54 * o_ready: output to next stage indicating readiness to accept data
55 * i_data : an input - added by the user of this class
59 self
.i_valid
= Signal(name
="p_i_valid") # >>in
60 self
.o_ready
= Signal(name
="p_o_ready") # <<out
62 def connect_in(self
, prev
):
63 """ helper function to connect stage to an input source. do not
64 use to connect stage-to-stage!
66 return [self
.i_valid
.eq(prev
.i_valid
),
67 prev
.o_ready
.eq(self
.o_ready
),
68 eq(self
.data
, prev
.data
),
73 """ contains the signals that go *to* the next stage (both in and out)
74 * o_valid: output indicating to next stage that data is valid
75 * i_ready: input from next stage indicating that it can accept data
76 * o_data : an output - added by the user of this class
79 self
.o_valid
= Signal(name
="n_o_valid") # out>>
80 self
.i_ready
= Signal(name
="n_i_ready") # <<in
82 def connect_to_next(self
, nxt
):
83 """ helper function to connect to the next stage data/valid/ready.
84 data/valid is passed *TO* nxt, and ready comes *IN* from nxt.
86 return [nxt
.i_valid
.eq(self
.o_valid
),
87 self
.i_ready
.eq(nxt
.o_ready
),
88 eq(nxt
.data
, self
.data
),
93 if not isinstance(o
, Sequence
):
96 for (ao
, ai
) in zip(o
, i
):
101 class BufferedPipeline
:
102 """ buffered pipeline stage. data and strobe signals travel in sync.
103 if ever the input is ready and the output is not, processed data
104 is stored in a temporary register.
106 stage-1 p.i_valid >>in stage n.o_valid out>> stage+1
107 stage-1 p.o_ready <<out stage n.i_ready <<in stage+1
108 stage-1 p.data >>in stage n.data out>> stage+1
114 input data p.data is read (only), is processed and goes into an
115 intermediate result store [process()]. this is updated combinatorially.
117 in a non-stall condition, the intermediate result will go into the
118 output (update_output). however if ever there is a stall, it goes
119 into r_data instead [update_buffer()].
121 when the non-stall condition is released, r_data is the first
122 to be transferred to the output [flush_buffer()], and the stall
125 on the next cycle (as long as stall is not raised again) the
126 input may begin to be processed and transferred directly to output.
128 def __init__(self
, stage
):
129 """ pass in a "stage" which may be either a static class or a class
130 instance, which has three functions:
131 * ispec: returns input signals according to the input specification
132 * ispec: returns output signals to the output specification
133 * process: takes an input instance and returns processed data
135 p.data -> process() -> result --> n.data
142 # set up input and output IO ACK (prev/next ready/valid)
143 self
.p
= PrevControl()
144 self
.n
= NextControl()
146 # set up the input and output data
147 self
.p
.data
= stage
.ispec() # input type
148 self
.r_data
= stage
.ospec() # all these are output type
149 self
.result
= stage
.ospec()
150 self
.n
.data
= stage
.ospec()
152 def connect_to_next(self
, nxt
):
153 """ helper function to connect to the next stage data/valid/ready.
154 data/valid is passed *TO* nxt, and ready comes *IN* from nxt.
156 return self
.n
.connect_to_next(nxt
.p
)
158 def connect_in(self
, prev
):
159 """ helper function to connect stage to an input source. do not
160 use to connect stage-to-stage!
162 return self
.p
.connect_in(prev
.p
)
164 def connect_out(self
, nxt
):
165 """ helper function to connect stage to an output source. do not
166 use to connect stage-to-stage!
168 return [nxt
.n
.o_valid
.eq(self
.n
.o_valid
),
169 self
.n
.i_ready
.eq(nxt
.n
.i_ready
),
170 eq(nxt
.n
.data
, self
.n
.data
),
173 def set_input(self
, i
):
174 """ helper function to set the input data
176 return eq(self
.p
.data
, i
)
178 def update_buffer(self
):
179 """ copies the result into the intermediate register r_data,
180 which will need to be outputted on a subsequent cycle
181 prior to allowing "normal" operation.
183 return eq(self
.r_data
, self
.result
)
185 def update_output(self
):
186 """ copies the (combinatorial) result into the output
188 return eq(self
.n
.data
, self
.result
)
190 def flush_buffer(self
):
191 """ copies the *intermediate* register r_data into the output
193 return eq(self
.n
.data
, self
.r_data
)
196 return [self
.p
.data
, self
.n
.data
]
198 def elaborate(self
, platform
):
200 if hasattr(self
.stage
, "setup"):
201 self
.stage
.setup(m
, self
.p
.data
)
203 # establish some combinatorial temporaries
204 o_n_validn
= Signal(reset_less
=True)
205 i_p_valid_o_p_ready
= Signal(reset_less
=True)
206 m
.d
.comb
+= [o_n_validn
.eq(~self
.n
.o_valid
),
207 i_p_valid_o_p_ready
.eq(self
.p
.i_valid
& self
.p
.o_ready
),
210 # store result of processing in combinatorial temporary
211 with m
.If(self
.p
.i_valid
): # input is valid: process it
212 m
.d
.comb
+= eq(self
.result
, self
.stage
.process(self
.p
.data
))
213 # if not in stall condition, update the temporary register
214 with m
.If(self
.p
.o_ready
): # not stalled
215 m
.d
.sync
+= self
.update_buffer()
217 #with m.If(self.p.i_rst): # reset
218 # m.d.sync += self.n.o_valid.eq(0)
219 # m.d.sync += self.p.o_ready.eq(0)
220 with m
.If(self
.n
.i_ready
): # next stage is ready
221 with m
.If(self
.p
.o_ready
): # not stalled
222 # nothing in buffer: send (processed) input direct to output
223 m
.d
.sync
+= [self
.n
.o_valid
.eq(self
.p
.i_valid
),
224 self
.update_output(),
226 with m
.Else(): # p.o_ready is false, and something is in buffer.
227 # Flush the [already processed] buffer to the output port.
228 m
.d
.sync
+= [self
.n
.o_valid
.eq(1),
230 # clear stall condition, declare register empty.
231 self
.p
.o_ready
.eq(1),
233 # ignore input, since p.o_ready is also false.
235 # (n.i_ready) is false here: next stage is ready
236 with m
.Elif(o_n_validn
): # next stage being told "ready"
237 m
.d
.sync
+= [self
.n
.o_valid
.eq(self
.p
.i_valid
),
238 self
.p
.o_ready
.eq(1), # Keep the buffer empty
239 # set the output data (from comb result)
240 self
.update_output(),
242 # (n.i_ready) false and (n.o_valid) true:
243 with m
.Elif(i_p_valid_o_p_ready
):
244 # If next stage *is* ready, and not stalled yet, accept input
245 m
.d
.sync
+= self
.p
.o_ready
.eq(~
(self
.p
.i_valid
& self
.n
.o_valid
))
250 return [self
.p
.i_valid
, self
.n
.i_ready
,
251 self
.n
.o_valid
, self
.p
.o_ready
,
255 class ExampleAddStage
:
256 """ an example of how to use the buffered pipeline, as a class instance
260 """ returns a tuple of input signals which will be the incoming data
262 return (Signal(16), Signal(16))
265 """ returns an output signal which will happen to contain the sum
270 def process(self
, i
):
271 """ process the input data (sums the values in the tuple) and returns it
276 class ExampleBufPipeAdd(BufferedPipeline
):
277 """ an example of how to use the buffered pipeline, using a class instance
281 addstage
= ExampleAddStage()
282 BufferedPipeline
.__init
__(self
, addstage
)
286 """ an example of how to use the buffered pipeline, in a static class
297 """ process the input data and returns it (adds 1)
302 class ExampleBufPipe(BufferedPipeline
):
303 """ an example of how to use the buffered pipeline.
307 BufferedPipeline
.__init
__(self
, ExampleStage
)
311 """A simple pipeline stage containing combinational logic that can execute
312 completely in one clock cycle.
316 input_shape : int or tuple or None
317 the shape of ``input.data`` and ``comb_input``
318 output_shape : int or tuple or None
319 the shape of ``output.data`` and ``comb_output``
329 comb_input : Signal, input_shape
330 The input to the combinatorial logic
331 comb_output: Signal, output_shape
332 The output of the combinatorial logic
335 def __init__(self
, stage
):
337 self
._data
_valid
= Signal()
338 # set up input and output IO ACK (prev/next ready/valid)
339 self
.p
= PrevControl()
340 self
.n
= NextControl()
342 # set up the input and output data
343 self
.p
.data
= stage
.ispec() # input type
344 self
.r_data
= stage
.ispec() # input type
345 self
.result
= stage
.ospec() # output data
346 self
.n
.data
= stage
.ospec() # output type
347 self
.n
.data
.name
= "outdata"
349 def set_input(self
, i
):
350 """ helper function to set the input data
352 return eq(self
.p
.data
, i
)
354 def elaborate(self
, platform
):
356 if hasattr(self
.stage
, "setup"):
357 self
.stage
.setup(m
, self
.r_data
)
358 m
.d
.comb
+= eq(self
.result
, self
.stage
.process(self
.r_data
))
359 m
.d
.comb
+= self
.n
.o_valid
.eq(self
._data
_valid
)
360 m
.d
.comb
+= self
.p
.o_ready
.eq(~self
._data
_valid | self
.n
.i_ready
)
361 m
.d
.sync
+= self
._data
_valid
.eq(self
.p
.i_valid | \
362 (~self
.n
.i_ready
& self
._data
_valid
))
363 with m
.If(self
.p
.i_valid
& self
.p
.o_ready
):
364 m
.d
.sync
+= eq(self
.r_data
, self
.p
.data
)
365 m
.d
.comb
+= eq(self
.n
.data
, self
.result
)
369 return [self
.p
.data
, self
.n
.data
]
372 class ExampleCombPipe(CombPipe
):
373 """ an example of how to use the combinatorial pipeline.
377 CombPipe
.__init
__(self
, ExampleStage
)
380 if __name__
== '__main__':
381 dut
= ExampleBufPipe()
382 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
383 with
open("test_bufpipe.il", "w") as f
:
386 dut
= ExampleCombPipe()
387 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
388 with
open("test_combpipe.il", "w") as f
: