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 nmigen
.hdl
.rec
import Record
, Layout
50 from collections
.abc
import Sequence
54 """ contains signals that come *from* the previous stage (both in and out)
55 * i_valid: previous stage indicating all incoming data is valid.
56 may be a multi-bit signal, where all bits are required
57 to be asserted to indicate "valid".
58 * o_ready: output to next stage indicating readiness to accept data
59 * i_data : an input - added by the user of this class
62 def __init__(self
, i_width
=1):
63 self
.i_valid
= Signal(i_width
, name
="p_i_valid") # prev >>in self
64 self
.o_ready
= Signal(name
="p_o_ready") # prev <<out self
66 def connect_in(self
, prev
):
67 """ helper function to connect stage to an input source. do not
68 use to connect stage-to-stage!
70 return [self
.i_valid
.eq(prev
.i_valid
),
71 prev
.o_ready
.eq(self
.o_ready
),
72 eq(self
.i_data
, prev
.i_data
),
77 """ contains the signals that go *to* the next stage (both in and out)
78 * o_valid: output indicating to next stage that data is valid
79 * i_ready: input from next stage indicating that it can accept data
80 * o_data : an output - added by the user of this class
83 self
.o_valid
= Signal(name
="n_o_valid") # self out>> next
84 self
.i_ready
= Signal(name
="n_i_ready") # self <<in next
86 def connect_to_next(self
, nxt
):
87 """ helper function to connect to the next stage data/valid/ready.
88 data/valid is passed *TO* nxt, and ready comes *IN* from nxt.
90 return [nxt
.i_valid
.eq(self
.o_valid
),
91 self
.i_ready
.eq(nxt
.o_ready
),
92 eq(nxt
.i_data
, self
.o_data
),
95 def connect_out(self
, nxt
):
96 """ helper function to connect stage to an output source. do not
97 use to connect stage-to-stage!
99 return [nxt
.o_valid
.eq(self
.o_valid
),
100 self
.i_ready
.eq(nxt
.i_ready
),
101 eq(nxt
.o_data
, self
.o_data
),
106 """ makes signals equal: a helper routine which identifies if it is being
107 passsed a list (or tuple) of objects, and calls the objects' eq
110 complex objects (classes) can be used: they must follow the
111 convention of having an eq member function, which takes the
112 responsibility of further calling eq and returning a list of
115 Record is a special (unusual, recursive) case, where the input
116 is specified as a dictionary (which may contain further dictionaries,
117 recursively), where the field names of the dictionary must match
118 the Record's field spec.
120 if not isinstance(o
, Sequence
):
123 for (ao
, ai
) in zip(o
, i
):
124 #print ("eq", ao, ai)
125 if isinstance(ao
, Record
):
126 for idx
, (field_name
, field_shape
, _
) in enumerate(ao
.layout
):
127 if isinstance(field_shape
, Layout
):
128 rres
= eq(ao
.fields
[field_name
], ai
.fields
[field_name
])
130 rres
= eq(ao
.fields
[field_name
], ai
[field_name
])
133 res
.append(ao
.eq(ai
))
138 """ Common functions for Pipeline API
140 def __init__(self
, stage
, in_multi
=None):
141 """ pass in a "stage" which may be either a static class or a class
142 instance, which has four functions (one optional):
143 * ispec: returns input signals according to the input specification
144 * ispec: returns output signals to the output specification
145 * process: takes an input instance and returns processed data
146 * setup: performs any module linkage if the stage uses one.
149 * add i_data member to PrevControl and
150 * add o_data member to NextControl
154 # set up input and output IO ACK (prev/next ready/valid)
155 self
.p
= PrevControl(in_multi
)
156 self
.n
= NextControl()
158 def connect_to_next(self
, nxt
):
159 """ helper function to connect to the next stage data/valid/ready.
161 return self
.n
.connect_to_next(nxt
.p
)
163 def connect_in(self
, prev
):
164 """ helper function to connect stage to an input source. do not
165 use to connect stage-to-stage!
167 return self
.p
.connect_in(prev
.p
)
169 def connect_out(self
, nxt
):
170 """ helper function to connect stage to an output source. do not
171 use to connect stage-to-stage!
173 return self
.n
.connect_out(nxt
.n
)
175 def set_input(self
, i
):
176 """ helper function to set the input data
178 return eq(self
.p
.i_data
, i
)
181 return [self
.p
.i_valid
, self
.n
.i_ready
,
182 self
.n
.o_valid
, self
.p
.o_ready
,
183 self
.p
.i_data
, self
.n
.o_data
# XXX need flattening!
187 class BufferedPipeline(PipelineBase
):
188 """ buffered pipeline stage. data and strobe signals travel in sync.
189 if ever the input is ready and the output is not, processed data
190 is stored in a temporary register.
192 stage-1 p.i_valid >>in stage n.o_valid out>> stage+1
193 stage-1 p.o_ready <<out stage n.i_ready <<in stage+1
194 stage-1 p.i_data >>in stage n.o_data out>> stage+1
200 input data p.i_data is read (only), is processed and goes into an
201 intermediate result store [process()]. this is updated combinatorially.
203 in a non-stall condition, the intermediate result will go into the
204 output (update_output). however if ever there is a stall, it goes
205 into r_data instead [update_buffer()].
207 when the non-stall condition is released, r_data is the first
208 to be transferred to the output [flush_buffer()], and the stall
211 on the next cycle (as long as stall is not raised again) the
212 input may begin to be processed and transferred directly to output.
214 def __init__(self
, stage
):
215 PipelineBase
.__init
__(self
, stage
)
217 # set up the input and output data
218 self
.p
.i_data
= stage
.ispec() # input type
219 self
.n
.o_data
= stage
.ospec()
221 def elaborate(self
, platform
):
224 result
= self
.stage
.ospec()
225 r_data
= self
.stage
.ospec()
226 if hasattr(self
.stage
, "setup"):
227 self
.stage
.setup(m
, self
.p
.i_data
)
229 # establish some combinatorial temporaries
230 p_i_valid
= Signal(reset_less
=True)
231 o_n_validn
= Signal(reset_less
=True)
232 i_p_valid_o_p_ready
= Signal(reset_less
=True)
233 vlen
= len(self
.p
.i_valid
)
234 if vlen
> 1: # multi-bit case: valid only when i_valid is all 1s
235 all1s
= Const(-1, (len(self
.p
.i_valid
), False))
236 m
.d
.comb
+= p_i_valid
.eq(self
.p
.i_valid
== all1s
)
237 else: # single-bit i_valid case
238 m
.d
.comb
+= p_i_valid
.eq(self
.p
.i_valid
)
239 m
.d
.comb
+= [ o_n_validn
.eq(~self
.n
.o_valid
),
240 i_p_valid_o_p_ready
.eq(p_i_valid
& self
.p
.o_ready
),
243 # store result of processing in combinatorial temporary
244 #with m.If(self.p.i_valid): # input is valid: process it
245 m
.d
.comb
+= eq(result
, self
.stage
.process(self
.p
.i_data
))
246 # if not in stall condition, update the temporary register
247 with m
.If(self
.p
.o_ready
): # not stalled
248 m
.d
.sync
+= eq(r_data
, result
) # update buffer
250 #with m.If(self.p.i_rst): # reset
251 # m.d.sync += self.n.o_valid.eq(0)
252 # m.d.sync += self.p.o_ready.eq(0)
253 with m
.If(self
.n
.i_ready
): # next stage is ready
254 with m
.If(self
.p
.o_ready
): # not stalled
255 # nothing in buffer: send (processed) input direct to output
256 m
.d
.sync
+= [self
.n
.o_valid
.eq(p_i_valid
),
257 eq(self
.n
.o_data
, result
), # update output
259 with m
.Else(): # p.o_ready is false, and something is in buffer.
260 # Flush the [already processed] buffer to the output port.
261 m
.d
.sync
+= [self
.n
.o_valid
.eq(1),
262 eq(self
.n
.o_data
, r_data
), # flush buffer
263 # clear stall condition, declare register empty.
264 self
.p
.o_ready
.eq(1),
266 # ignore input, since p.o_ready is also false.
268 # (n.i_ready) is false here: next stage is ready
269 with m
.Elif(o_n_validn
): # next stage being told "ready"
270 m
.d
.sync
+= [self
.n
.o_valid
.eq(p_i_valid
),
271 self
.p
.o_ready
.eq(1), # Keep the buffer empty
272 # set the output data (from comb result)
273 eq(self
.n
.o_data
, result
),
275 # (n.i_ready) false and (n.o_valid) true:
276 with m
.Elif(i_p_valid_o_p_ready
):
277 # If next stage *is* ready, and not stalled yet, accept input
278 m
.d
.sync
+= self
.p
.o_ready
.eq(~
(p_i_valid
& self
.n
.o_valid
))
283 class ExampleAddStage
:
284 """ an example of how to use the buffered pipeline, as a class instance
288 """ returns a tuple of input signals which will be the incoming data
290 return (Signal(16), Signal(16))
293 """ returns an output signal which will happen to contain the sum
298 def process(self
, i
):
299 """ process the input data (sums the values in the tuple) and returns it
304 class ExampleBufPipeAdd(BufferedPipeline
):
305 """ an example of how to use the buffered pipeline, using a class instance
309 addstage
= ExampleAddStage()
310 BufferedPipeline
.__init
__(self
, addstage
)
314 """ an example of how to use the buffered pipeline, in a static class
325 """ process the input data and returns it (adds 1)
330 class ExampleBufPipe(BufferedPipeline
):
331 """ an example of how to use the buffered pipeline.
335 BufferedPipeline
.__init
__(self
, ExampleStage
)
338 class CombPipe(PipelineBase
):
339 """A simple pipeline stage containing combinational logic that can execute
340 completely in one clock cycle.
348 r_data : Signal, input_shape
349 A temporary (buffered) copy of a prior (valid) input
350 result: Signal, output_shape
351 The output of the combinatorial logic
354 def __init__(self
, stage
):
355 PipelineBase
.__init
__(self
, stage
)
356 self
._data
_valid
= Signal()
358 # set up the input and output data
359 self
.p
.i_data
= stage
.ispec() # input type
360 self
.n
.o_data
= stage
.ospec() # output type
362 def elaborate(self
, platform
):
365 r_data
= self
.stage
.ispec() # input type
366 result
= self
.stage
.ospec() # output data
367 if hasattr(self
.stage
, "setup"):
368 self
.stage
.setup(m
, r_data
)
370 m
.d
.comb
+= eq(result
, self
.stage
.process(r_data
))
371 m
.d
.comb
+= self
.n
.o_valid
.eq(self
._data
_valid
)
372 m
.d
.comb
+= self
.p
.o_ready
.eq(~self
._data
_valid | self
.n
.i_ready
)
373 m
.d
.sync
+= self
._data
_valid
.eq(self
.p
.i_valid | \
374 (~self
.n
.i_ready
& self
._data
_valid
))
375 with m
.If(self
.p
.i_valid
& self
.p
.o_ready
):
376 m
.d
.sync
+= eq(r_data
, self
.p
.i_data
)
377 m
.d
.comb
+= eq(self
.n
.o_data
, result
)
381 class ExampleCombPipe(CombPipe
):
382 """ an example of how to use the combinatorial pipeline.
386 CombPipe
.__init
__(self
, ExampleStage
)
389 if __name__
== '__main__':
390 dut
= ExampleBufPipe()
391 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
392 with
open("test_bufpipe.il", "w") as f
:
395 dut
= ExampleCombPipe()
396 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
397 with
open("test_combpipe.il", "w") as f
: