1 """*Experimental* ALU: based on nmigen alu_hier.py, includes branch-compare ALU
3 This ALU is *deliberately* designed to add in (unnecessary) delays into
4 different operations so as to be able to test the 6600-style matrices
5 and the CompUnits. Countdown timers wait for (defined) periods before
6 indicating that the output is valid
8 A "real" integer ALU would place the answers onto the output bus after
12 from nmigen
import Elaboratable
, Signal
, Module
, Const
, Mux
, Array
13 from nmigen
.hdl
.rec
import Record
, Layout
14 from nmigen
.cli
import main
15 from nmigen
.cli
import verilog
, rtlil
16 from nmigen
.compat
.sim
import run_simulation
17 from nmutil
.extend
import exts
18 from nmutil
.gtkw
import write_gtkw
20 # NOTE: to use cxxsim, export NMIGEN_SIM_MODE=cxxsim from the shell
21 # Also, check out the cxxsim nmigen branch, and latest yosys from git
22 from nmutil
.sim_tmp_alternative
import (Simulator
, nmigen_sim_top_module
,
25 from soc
.decoder
.decode2execute1
import Data
26 from soc
.decoder
.power_enums
import MicrOp
, Function
, CryIn
28 from soc
.fu
.alu
.alu_input_record
import CompALUOpSubset
29 from soc
.fu
.cr
.cr_input_record
import CompCROpSubset
34 class Adder(Elaboratable
):
35 def __init__(self
, width
):
36 self
.invert_in
= Signal()
37 self
.a
= Signal(width
)
38 self
.b
= Signal(width
)
39 self
.o
= Signal(width
, name
="add_o")
41 def elaborate(self
, platform
):
43 with m
.If(self
.invert_in
):
44 m
.d
.comb
+= self
.o
.eq((~self
.a
) + self
.b
)
46 m
.d
.comb
+= self
.o
.eq(self
.a
+ self
.b
)
50 class Subtractor(Elaboratable
):
51 def __init__(self
, width
):
52 self
.a
= Signal(width
)
53 self
.b
= Signal(width
)
54 self
.o
= Signal(width
, name
="sub_o")
56 def elaborate(self
, platform
):
58 m
.d
.comb
+= self
.o
.eq(self
.a
- self
.b
)
62 class Multiplier(Elaboratable
):
63 def __init__(self
, width
):
64 self
.a
= Signal(width
)
65 self
.b
= Signal(width
)
66 self
.o
= Signal(width
, name
="mul_o")
68 def elaborate(self
, platform
):
70 m
.d
.comb
+= self
.o
.eq(self
.a
* self
.b
)
74 class Shifter(Elaboratable
):
75 def __init__(self
, width
):
77 self
.a
= Signal(width
)
78 self
.b
= Signal(width
)
79 self
.o
= Signal(width
, name
="shf_o")
81 def elaborate(self
, platform
):
83 btrunc
= Signal(self
.width
)
84 m
.d
.comb
+= btrunc
.eq(self
.b
& Const((1 << self
.width
)-1))
85 m
.d
.comb
+= self
.o
.eq(self
.a
>> btrunc
)
89 class SignExtend(Elaboratable
):
90 def __init__(self
, width
):
92 self
.a
= Signal(width
)
93 self
.o
= Signal(width
, name
="exts_o")
95 def elaborate(self
, platform
):
97 m
.d
.comb
+= self
.o
.eq(exts(self
.a
, 8, self
.width
))
105 class DummyALU(Elaboratable
):
106 def __init__(self
, width
):
107 self
.p
= Dummy() # make look like nmutil pipeline API
108 self
.p
.data_i
= Dummy()
109 self
.p
.data_i
.ctx
= Dummy()
110 self
.n
= Dummy() # make look like nmutil pipeline API
111 self
.n
.data_o
= Dummy()
112 self
.p
.valid_i
= Signal()
113 self
.p
.ready_o
= Signal()
114 self
.n
.ready_i
= Signal()
115 self
.n
.valid_o
= Signal()
116 self
.counter
= Signal(4)
117 self
.op
= CompCROpSubset()
119 i
.append(Signal(width
, name
="i1"))
120 i
.append(Signal(width
, name
="i2"))
121 i
.append(Signal(width
, name
="i3"))
123 self
.a
, self
.b
, self
.c
= i
[0], i
[1], i
[2]
124 self
.out
= Array([Signal(width
, name
="alu_o")])
127 # more "look like nmutil pipeline API"
128 self
.p
.data_i
.ctx
.op
= self
.op
129 self
.p
.data_i
.a
= self
.a
130 self
.p
.data_i
.b
= self
.b
131 self
.p
.data_i
.c
= self
.c
132 self
.n
.data_o
.o
= self
.o
134 def elaborate(self
, platform
):
137 go_now
= Signal(reset_less
=True) # testing no-delay ALU
139 with m
.If(self
.p
.valid_i
):
140 # input is valid. next check, if we already said "ready" or not
141 with m
.If(~self
.p
.ready_o
):
142 # we didn't say "ready" yet, so say so and initialise
143 m
.d
.sync
+= self
.p
.ready_o
.eq(1)
145 m
.d
.sync
+= self
.o
.eq(self
.a
)
146 m
.d
.comb
+= go_now
.eq(1)
147 m
.d
.sync
+= self
.counter
.eq(1)
150 # input says no longer valid, so drop ready as well.
151 # a "proper" ALU would have had to sync in the opcode and a/b ops
152 m
.d
.sync
+= self
.p
.ready_o
.eq(0)
154 # ok so the counter's running: when it gets to 1, fire the output
155 with m
.If((self
.counter
== 1) | go_now
):
156 # set the output as valid if the recipient is ready for it
157 m
.d
.sync
+= self
.n
.valid_o
.eq(1)
158 with m
.If(self
.n
.ready_i
& self
.n
.valid_o
):
159 m
.d
.sync
+= self
.n
.valid_o
.eq(0)
160 # recipient said it was ready: reset back to known-good.
161 m
.d
.sync
+= self
.counter
.eq(0) # reset the counter
162 m
.d
.sync
+= self
.o
.eq(0) # clear the output for tidiness sake
164 # countdown to 1 (transition from 1 to 0 only on acknowledgement)
165 with m
.If(self
.counter
> 1):
166 m
.d
.sync
+= self
.counter
.eq(self
.counter
- 1)
171 yield from self
.op
.ports()
181 class ALU(Elaboratable
):
182 def __init__(self
, width
):
183 self
.p
= Dummy() # make look like nmutil pipeline API
184 self
.p
.data_i
= Dummy()
185 self
.p
.data_i
.ctx
= Dummy()
186 self
.n
= Dummy() # make look like nmutil pipeline API
187 self
.n
.data_o
= Dummy()
188 self
.p
.valid_i
= Signal()
189 self
.p
.ready_o
= Signal()
190 self
.n
.ready_i
= Signal()
191 self
.n
.valid_o
= Signal()
192 self
.counter
= Signal(4)
193 self
.op
= CompALUOpSubset(name
="op")
195 i
.append(Signal(width
, name
="i1"))
196 i
.append(Signal(width
, name
="i2"))
198 self
.a
, self
.b
= i
[0], i
[1]
199 self
.out
= Array([Data(width
, name
="alu_o")])
202 # more "look like nmutil pipeline API"
203 self
.p
.data_i
.ctx
.op
= self
.op
204 self
.p
.data_i
.a
= self
.a
205 self
.p
.data_i
.b
= self
.b
206 self
.n
.data_o
.o
= self
.o
208 def elaborate(self
, platform
):
210 add
= Adder(self
.width
)
211 mul
= Multiplier(self
.width
)
212 shf
= Shifter(self
.width
)
213 sub
= Subtractor(self
.width
)
214 ext_sign
= SignExtend(self
.width
)
216 m
.submodules
.add
= add
217 m
.submodules
.mul
= mul
218 m
.submodules
.shf
= shf
219 m
.submodules
.sub
= sub
220 m
.submodules
.ext_sign
= ext_sign
222 # really should not activate absolutely all ALU inputs like this
223 for mod
in [add
, mul
, shf
, sub
]:
228 # EXTS sign extends the first input
229 with m
.If(self
.op
.insn_type
== MicrOp
.OP_EXTS
):
230 m
.d
.comb
+= ext_sign
.a
.eq(self
.a
)
231 # EXTSWSLI sign extends the second input
232 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_EXTSWSLI
):
233 m
.d
.comb
+= ext_sign
.a
.eq(self
.b
)
235 # pass invert (and carry later)
236 m
.d
.comb
+= add
.invert_in
.eq(self
.op
.invert_in
)
238 go_now
= Signal(reset_less
=True) # testing no-delay ALU
240 # ALU sequencer is idle when the count is zero
241 alu_idle
= Signal(reset_less
=True)
242 m
.d
.comb
+= alu_idle
.eq(self
.counter
== 0)
244 # ALU sequencer is done when the count is one
245 alu_done
= Signal(reset_less
=True)
246 m
.d
.comb
+= alu_done
.eq(self
.counter
== 1)
248 # select handshake handling according to ALU type
250 # with a combinatorial, no-delay ALU, just pass through
251 # the handshake signals to the other side
252 m
.d
.comb
+= self
.p
.ready_o
.eq(self
.n
.ready_i
)
253 m
.d
.comb
+= self
.n
.valid_o
.eq(self
.p
.valid_i
)
255 # sequential ALU handshake:
256 # ready_o responds to valid_i, but only if the ALU is idle
257 m
.d
.comb
+= self
.p
.ready_o
.eq(alu_idle
)
258 # select the internally generated valid_o, above
259 m
.d
.comb
+= self
.n
.valid_o
.eq(alu_done
)
261 # hold the ALU result until ready_o is asserted
262 alu_r
= Signal(self
.width
)
264 # NOP doesn't output anything
265 with m
.If(self
.op
.insn_type
!= MicrOp
.OP_NOP
):
266 m
.d
.comb
+= self
.o
.ok
.eq(1)
268 with m
.If(self
.p
.valid_i
):
270 # as this is a "fake" pipeline, just grab the output right now
271 with m
.If(self
.op
.insn_type
== MicrOp
.OP_ADD
):
272 m
.d
.sync
+= alu_r
.eq(add
.o
)
273 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_MUL_L64
):
274 m
.d
.sync
+= alu_r
.eq(mul
.o
)
275 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_SHR
):
276 m
.d
.sync
+= alu_r
.eq(shf
.o
)
277 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_EXTS
):
278 m
.d
.sync
+= alu_r
.eq(ext_sign
.o
)
279 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_EXTSWSLI
):
280 m
.d
.sync
+= alu_r
.eq(ext_sign
.o
)
281 # SUB is zero-delay, no need to register
283 # NOTE: all of these are fake, just something to test
285 # MUL, to take 5 instructions
286 with m
.If(self
.op
.insn_type
== MicrOp
.OP_MUL_L64
):
287 m
.d
.sync
+= self
.counter
.eq(5)
288 # SHIFT to take 1, straight away
289 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_SHR
):
290 m
.d
.sync
+= self
.counter
.eq(1)
292 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_ADD
):
293 m
.d
.sync
+= self
.counter
.eq(3)
295 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_EXTS
):
296 m
.d
.sync
+= self
.counter
.eq(1)
298 with m
.Elif(self
.op
.insn_type
== MicrOp
.OP_EXTSWSLI
):
299 m
.d
.sync
+= self
.counter
.eq(1)
300 # others to take no delay
302 m
.d
.comb
+= go_now
.eq(1)
304 with m
.Elif(~alu_done | self
.n
.ready_i
):
305 # decrement the counter while the ALU is neither idle nor finished
306 m
.d
.sync
+= self
.counter
.eq(self
.counter
- 1)
308 # choose between zero-delay output, or registered
310 m
.d
.comb
+= self
.o
.data
.eq(sub
.o
)
311 # only present the result at the last computation cycle
312 with m
.Elif(alu_done
):
313 m
.d
.comb
+= self
.o
.data
.eq(alu_r
)
318 yield from self
.op
.ports()
321 yield from self
.o
.ports()
331 class BranchOp(Elaboratable
):
332 def __init__(self
, width
, op
):
333 self
.a
= Signal(width
)
334 self
.b
= Signal(width
)
335 self
.o
= Signal(width
)
338 def elaborate(self
, platform
):
340 m
.d
.comb
+= self
.o
.eq(Mux(self
.op(self
.a
, self
.b
), 1, 0))
344 class BranchALU(Elaboratable
):
345 def __init__(self
, width
):
346 self
.p
= Dummy() # make look like nmutil pipeline API
347 self
.p
.data_i
= Dummy()
348 self
.p
.data_i
.ctx
= Dummy()
349 self
.n
= Dummy() # make look like nmutil pipeline API
350 self
.n
.data_o
= Dummy()
351 self
.p
.valid_i
= Signal()
352 self
.p
.ready_o
= Signal()
353 self
.n
.ready_i
= Signal()
354 self
.n
.valid_o
= Signal()
355 self
.counter
= Signal(4)
358 i
.append(Signal(width
, name
="i1"))
359 i
.append(Signal(width
, name
="i2"))
361 self
.a
, self
.b
= i
[0], i
[1]
362 self
.out
= Array([Signal(width
)])
366 def elaborate(self
, platform
):
368 bgt
= BranchOp(self
.width
, operator
.gt
)
369 blt
= BranchOp(self
.width
, operator
.lt
)
370 beq
= BranchOp(self
.width
, operator
.eq
)
371 bne
= BranchOp(self
.width
, operator
.ne
)
373 m
.submodules
.bgt
= bgt
374 m
.submodules
.blt
= blt
375 m
.submodules
.beq
= beq
376 m
.submodules
.bne
= bne
377 for mod
in [bgt
, blt
, beq
, bne
]:
383 go_now
= Signal(reset_less
=True) # testing no-delay ALU
384 with m
.If(self
.p
.valid_i
):
385 # input is valid. next check, if we already said "ready" or not
386 with m
.If(~self
.p
.ready_o
):
387 # we didn't say "ready" yet, so say so and initialise
388 m
.d
.sync
+= self
.p
.ready_o
.eq(1)
390 # as this is a "fake" pipeline, just grab the output right now
391 with m
.Switch(self
.op
):
392 for i
, mod
in enumerate([bgt
, blt
, beq
, bne
]):
394 m
.d
.sync
+= self
.o
.eq(mod
.o
)
395 # branch to take 5 cycles (fake)
396 m
.d
.sync
+= self
.counter
.eq(5)
397 #m.d.comb += go_now.eq(1)
399 # input says no longer valid, so drop ready as well.
400 # a "proper" ALU would have had to sync in the opcode and a/b ops
401 m
.d
.sync
+= self
.p
.ready_o
.eq(0)
403 # ok so the counter's running: when it gets to 1, fire the output
404 with m
.If((self
.counter
== 1) | go_now
):
405 # set the output as valid if the recipient is ready for it
406 m
.d
.sync
+= self
.n
.valid_o
.eq(1)
407 with m
.If(self
.n
.ready_i
& self
.n
.valid_o
):
408 m
.d
.sync
+= self
.n
.valid_o
.eq(0)
409 # recipient said it was ready: reset back to known-good.
410 m
.d
.sync
+= self
.counter
.eq(0) # reset the counter
411 m
.d
.sync
+= self
.o
.eq(0) # clear the output for tidiness sake
413 # countdown to 1 (transition from 1 to 0 only on acknowledgement)
414 with m
.If(self
.counter
> 1):
415 m
.d
.sync
+= self
.counter
.eq(self
.counter
- 1)
429 def run_op(dut
, a
, b
, op
, inv_a
=0):
432 yield dut
.op
.insn_type
.eq(op
)
433 yield dut
.op
.invert_in
.eq(inv_a
)
434 yield dut
.n
.ready_i
.eq(0)
435 yield dut
.p
.valid_i
.eq(1)
436 yield dut
.n
.ready_i
.eq(1)
439 # wait for the ALU to accept our input data
440 while not (yield dut
.p
.ready_o
):
443 yield dut
.p
.valid_i
.eq(0)
446 yield dut
.op
.insn_type
.eq(0)
447 yield dut
.op
.invert_in
.eq(0)
449 # wait for the ALU to present the output data
450 while not (yield dut
.n
.valid_o
):
453 # latch the result and lower read_i
454 result
= yield dut
.o
.data
455 yield dut
.n
.ready_i
.eq(0)
461 result
= yield from run_op(dut
, 5, 3, MicrOp
.OP_ADD
)
462 print("alu_sim add", result
)
465 result
= yield from run_op(dut
, 2, 3, MicrOp
.OP_MUL_L64
)
466 print("alu_sim mul", result
)
469 result
= yield from run_op(dut
, 5, 3, MicrOp
.OP_ADD
, inv_a
=1)
470 print("alu_sim add-inv", result
)
471 assert (result
== 65533)
473 # test zero-delay ALU
474 # don't have OP_SUB, so use any other
475 result
= yield from run_op(dut
, 5, 3, MicrOp
.OP_CMP
)
476 print("alu_sim sub", result
)
479 result
= yield from run_op(dut
, 13, 2, MicrOp
.OP_SHR
)
480 print("alu_sim shr", result
)
486 write_alu_gtkw("test_alusim.gtkw", clk_period
=10e-9)
487 run_simulation(alu
, {"sync": alu_sim(alu
)}, vcd_name
='test_alusim.vcd')
489 vl
= rtlil
.convert(alu
, ports
=alu
.ports())
490 with
open("test_alu.il", "w") as f
:
494 def test_alu_parallel():
495 # Compare with the sequential test implementation, above.
497 m
.submodules
.alu
= dut
= ALU(width
=16)
498 write_alu_gtkw("test_alu_parallel.gtkw", sub_module
='alu',
499 pysim
=is_engine_pysim())
504 def send(a
, b
, op
, inv_a
=0):
505 # present input data and assert valid_i
508 yield dut
.op
.insn_type
.eq(op
)
509 yield dut
.op
.invert_in
.eq(inv_a
)
510 yield dut
.p
.valid_i
.eq(1)
512 # wait for ready_o to be asserted
513 while not (yield dut
.p
.ready_o
):
515 # clear input data and negate valid_i
516 # if send is called again immediately afterwards, there will be no
517 # visible transition (they will not be negated, after all)
518 yield dut
.p
.valid_i
.eq(0)
521 yield dut
.op
.insn_type
.eq(0)
522 yield dut
.op
.invert_in
.eq(0)
525 # signal readiness to receive data
526 yield dut
.n
.ready_i
.eq(1)
528 # wait for valid_o to be asserted
529 while not (yield dut
.n
.valid_o
):
532 result
= yield dut
.o
.data
534 # if receive is called again immediately afterwards, there will be no
535 # visible transition (it will not be negated, after all)
536 yield dut
.n
.ready_i
.eq(0)
540 # send a few test cases, interspersed with wait states
541 # note that, for this test, we do not wait for the result to be ready,
542 # before presenting the next input
544 yield from send(5, 3, MicrOp
.OP_ADD
)
548 yield from send(2, 3, MicrOp
.OP_MUL_L64
)
550 yield from send(5, 3, MicrOp
.OP_ADD
, inv_a
=1)
553 # note that this is a zero-delay operation
554 yield from send(5, 3, MicrOp
.OP_NOP
)
558 yield from send(13, 2, MicrOp
.OP_SHR
)
560 yield from send(13, 2, MicrOp
.OP_EXTS
)
561 # sign extend -128 (8 bits)
562 yield from send(0x80, 2, MicrOp
.OP_EXTS
)
563 # sign extend -128 (8 bits)
564 yield from send(2, 0x80, MicrOp
.OP_EXTSWSLI
)
567 # receive and check results, interspersed with wait states
568 # the consumer is not in step with the producer, but the
569 # order of the results are preserved
572 result
= yield from receive()
575 result
= yield from receive()
580 result
= yield from receive()
581 assert (result
== 65533) # unsigned equivalent to -2
583 # note that this is a zero-delay operation
584 # this, and the previous result, will be received back-to-back
585 # (check the output waveform to see this)
586 result
= yield from receive()
591 result
= yield from receive()
593 # sign extent 13 = 13
594 result
= yield from receive()
595 assert (result
== 13)
596 # sign extend -128 (8 bits) = -128 (16 bits)
597 result
= yield from receive()
598 assert (result
== 0xFF80)
599 # sign extend -128 (8 bits) = -128 (16 bits)
600 result
= yield from receive()
601 assert (result
== 0xFF80)
603 sim
.add_sync_process(producer
)
604 sim
.add_sync_process(consumer
)
605 sim_writer
= sim
.write_vcd("test_alu_parallel.vcd")
610 def write_alu_gtkw(gtkw_name
, clk_period
=1e-6, sub_module
=None,
612 """Common function to write the GTKWave documents for this module"""
617 'op__insn_type' if pysim
else 'op__insn_type[6:0]',
625 # determine the module name of the DUT
627 if sub_module
is not None:
628 module
= nmigen_sim_top_module
+ sub_module
629 vcd_name
= gtkw_name
.replace('.gtkw', '.vcd')
630 write_gtkw(gtkw_name
, vcd_name
, gtkwave_desc
, module
=module
,
631 loc
=__file__
, clk_period
=clk_period
, base
='signed')
634 if __name__
== "__main__":
638 # alu = BranchALU(width=16)
639 # vl = rtlil.convert(alu, ports=alu.ports())
640 # with open("test_branch_alu.il", "w") as f: