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
.i_data
, prev
.i_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
.i_data
, self
.o_data
),
91 def connect_out(self
, nxt
):
92 """ helper function to connect stage to an output source. do not
93 use to connect stage-to-stage!
95 return [nxt
.o_valid
.eq(self
.o_valid
),
96 self
.i_ready
.eq(nxt
.i_ready
),
97 eq(nxt
.o_data
, self
.o_data
),
102 if not isinstance(o
, Sequence
):
105 for (ao
, ai
) in zip(o
, i
):
106 res
.append(ao
.eq(ai
))
111 """ Common functions for Pipeline API
113 def __init__(self
, stage
):
114 """ pass in a "stage" which may be either a static class or a class
115 instance, which has three functions:
116 * ispec: returns input signals according to the input specification
117 * ispec: returns output signals to the output specification
118 * process: takes an input instance and returns processed data
121 * add i_data member to PrevControl and
122 * add o_data member to NextControl
126 # set up input and output IO ACK (prev/next ready/valid)
127 self
.p
= PrevControl()
128 self
.n
= NextControl()
130 def connect_to_next(self
, nxt
):
131 """ helper function to connect to the next stage data/valid/ready.
133 return self
.n
.connect_to_next(nxt
.p
)
135 def connect_in(self
, prev
):
136 """ helper function to connect stage to an input source. do not
137 use to connect stage-to-stage!
139 return self
.p
.connect_in(prev
.p
)
141 def connect_out(self
, nxt
):
142 """ helper function to connect stage to an output source. do not
143 use to connect stage-to-stage!
145 return self
.n
.connect_out(nxt
.n
)
147 def set_input(self
, i
):
148 """ helper function to set the input data
150 return eq(self
.p
.i_data
, i
)
153 return [self
.p
.i_valid
, self
.n
.i_ready
,
154 self
.n
.o_valid
, self
.p
.o_ready
,
155 self
.p
.i_data
, self
.n
.o_data
159 class BufferedPipeline(PipelineBase
):
160 """ buffered pipeline stage. data and strobe signals travel in sync.
161 if ever the input is ready and the output is not, processed data
162 is stored in a temporary register.
164 stage-1 p.i_valid >>in stage n.o_valid out>> stage+1
165 stage-1 p.o_ready <<out stage n.i_ready <<in stage+1
166 stage-1 p.i_data >>in stage n.o_data out>> stage+1
172 input data p.i_data is read (only), is processed and goes into an
173 intermediate result store [process()]. this is updated combinatorially.
175 in a non-stall condition, the intermediate result will go into the
176 output (update_output). however if ever there is a stall, it goes
177 into r_data instead [update_buffer()].
179 when the non-stall condition is released, r_data is the first
180 to be transferred to the output [flush_buffer()], and the stall
183 on the next cycle (as long as stall is not raised again) the
184 input may begin to be processed and transferred directly to output.
186 def __init__(self
, stage
):
187 PipelineBase
.__init
__(self
, stage
)
189 # set up the input and output data
190 self
.p
.i_data
= stage
.ispec() # input type
191 self
.r_data
= stage
.ospec() # all these are output type
192 self
.result
= stage
.ospec()
193 self
.n
.o_data
= stage
.ospec()
195 def update_buffer(self
):
196 """ copies the result into the intermediate register r_data,
197 which will need to be outputted on a subsequent cycle
198 prior to allowing "normal" operation.
200 return eq(self
.r_data
, self
.result
)
202 def update_output(self
):
203 """ copies the (combinatorial) result into the output
205 return eq(self
.n
.o_data
, self
.result
)
207 def flush_buffer(self
):
208 """ copies the *intermediate* register r_data into the output
210 return eq(self
.n
.o_data
, self
.r_data
)
212 def elaborate(self
, platform
):
214 if hasattr(self
.stage
, "setup"):
215 self
.stage
.setup(m
, self
.p
.i_data
)
217 # establish some combinatorial temporaries
218 o_n_validn
= Signal(reset_less
=True)
219 i_p_valid_o_p_ready
= Signal(reset_less
=True)
220 m
.d
.comb
+= [o_n_validn
.eq(~self
.n
.o_valid
),
221 i_p_valid_o_p_ready
.eq(self
.p
.i_valid
& self
.p
.o_ready
),
224 # store result of processing in combinatorial temporary
225 with m
.If(self
.p
.i_valid
): # input is valid: process it
226 m
.d
.comb
+= eq(self
.result
, self
.stage
.process(self
.p
.i_data
))
227 # if not in stall condition, update the temporary register
228 with m
.If(self
.p
.o_ready
): # not stalled
229 m
.d
.sync
+= self
.update_buffer()
231 #with m.If(self.p.i_rst): # reset
232 # m.d.sync += self.n.o_valid.eq(0)
233 # m.d.sync += self.p.o_ready.eq(0)
234 with m
.If(self
.n
.i_ready
): # next stage is ready
235 with m
.If(self
.p
.o_ready
): # not stalled
236 # nothing in buffer: send (processed) input direct to output
237 m
.d
.sync
+= [self
.n
.o_valid
.eq(self
.p
.i_valid
),
238 self
.update_output(),
240 with m
.Else(): # p.o_ready is false, and something is in buffer.
241 # Flush the [already processed] buffer to the output port.
242 m
.d
.sync
+= [self
.n
.o_valid
.eq(1),
244 # clear stall condition, declare register empty.
245 self
.p
.o_ready
.eq(1),
247 # ignore input, since p.o_ready is also false.
249 # (n.i_ready) is false here: next stage is ready
250 with m
.Elif(o_n_validn
): # next stage being told "ready"
251 m
.d
.sync
+= [self
.n
.o_valid
.eq(self
.p
.i_valid
),
252 self
.p
.o_ready
.eq(1), # Keep the buffer empty
253 # set the output data (from comb result)
254 self
.update_output(),
256 # (n.i_ready) false and (n.o_valid) true:
257 with m
.Elif(i_p_valid_o_p_ready
):
258 # If next stage *is* ready, and not stalled yet, accept input
259 m
.d
.sync
+= self
.p
.o_ready
.eq(~
(self
.p
.i_valid
& self
.n
.o_valid
))
264 class ExampleAddStage
:
265 """ an example of how to use the buffered pipeline, as a class instance
269 """ returns a tuple of input signals which will be the incoming data
271 return (Signal(16), Signal(16))
274 """ returns an output signal which will happen to contain the sum
279 def process(self
, i
):
280 """ process the input data (sums the values in the tuple) and returns it
285 class ExampleBufPipeAdd(BufferedPipeline
):
286 """ an example of how to use the buffered pipeline, using a class instance
290 addstage
= ExampleAddStage()
291 BufferedPipeline
.__init
__(self
, addstage
)
295 """ an example of how to use the buffered pipeline, in a static class
306 """ process the input data and returns it (adds 1)
311 class ExampleBufPipe(BufferedPipeline
):
312 """ an example of how to use the buffered pipeline.
316 BufferedPipeline
.__init
__(self
, ExampleStage
)
319 class CombPipe(PipelineBase
):
320 """A simple pipeline stage containing combinational logic that can execute
321 completely in one clock cycle.
325 input_shape : int or tuple or None
326 the shape of ``input.data`` and ``comb_input``
327 output_shape : int or tuple or None
328 the shape of ``output.data`` and ``comb_output``
338 comb_input : Signal, input_shape
339 The input to the combinatorial logic
340 comb_output: Signal, output_shape
341 The output of the combinatorial logic
344 def __init__(self
, stage
):
345 PipelineBase
.__init
__(self
, stage
)
346 self
._data
_valid
= Signal()
348 # set up the input and output data
349 self
.p
.i_data
= stage
.ispec() # input type
350 self
.r_data
= stage
.ispec() # input type
351 self
.result
= stage
.ospec() # output data
352 self
.n
.o_data
= stage
.ospec() # output type
353 self
.n
.o_data
.name
= "outdata"
355 def elaborate(self
, platform
):
357 if hasattr(self
.stage
, "setup"):
358 self
.stage
.setup(m
, self
.r_data
)
359 m
.d
.comb
+= eq(self
.result
, self
.stage
.process(self
.r_data
))
360 m
.d
.comb
+= self
.n
.o_valid
.eq(self
._data
_valid
)
361 m
.d
.comb
+= self
.p
.o_ready
.eq(~self
._data
_valid | self
.n
.i_ready
)
362 m
.d
.sync
+= self
._data
_valid
.eq(self
.p
.i_valid | \
363 (~self
.n
.i_ready
& self
._data
_valid
))
364 with m
.If(self
.p
.i_valid
& self
.p
.o_ready
):
365 m
.d
.sync
+= eq(self
.r_data
, self
.p
.i_data
)
366 m
.d
.comb
+= eq(self
.n
.o_data
, self
.result
)
370 class ExampleCombPipe(CombPipe
):
371 """ an example of how to use the combinatorial pipeline.
375 CombPipe
.__init
__(self
, ExampleStage
)
378 if __name__
== '__main__':
379 dut
= ExampleBufPipe()
380 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
381 with
open("test_bufpipe.il", "w") as f
:
384 dut
= ExampleCombPipe()
385 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
386 with
open("test_combpipe.il", "w") as f
: