1 """Computation Unit (aka "ALU Manager").
3 Manages a Pipeline or FSM, ensuring that the start and end time are 100%
4 monitored. At no time may the ALU proceed without this module notifying
5 the Dependency Matrices. At no time is a result production "abandoned".
6 This module blocks (indicates busy) starting from when it first receives
7 an opcode until it receives notification that
8 its result(s) have been successfully stored in the regfile(s)
10 Documented at http://libre-soc.org/3d_gpu/architecture/compunit
13 from nmigen
.compat
.sim
import run_simulation
, Settle
14 from nmigen
.cli
import rtlil
15 from nmigen
import Module
17 from soc
.decoder
.power_enums
import InternalOp
19 from soc
.experiment
.compalu_multi
import MultiCompUnit
20 from soc
.experiment
.alu_hier
import ALU
, DummyALU
21 from soc
.fu
.alu
.alu_input_record
import CompALUOpSubset
24 def op_sim(dut
, a
, b
, op
, inv_a
=0, imm
=0, imm_ok
=0, zero_a
=0):
25 yield dut
.issue_i
.eq(0)
27 yield dut
.src_i
[0].eq(a
)
28 yield dut
.src_i
[1].eq(b
)
29 yield dut
.oper_i
.insn_type
.eq(op
)
30 yield dut
.oper_i
.invert_a
.eq(inv_a
)
31 yield dut
.oper_i
.imm_data
.imm
.eq(imm
)
32 yield dut
.oper_i
.imm_data
.imm_ok
.eq(imm_ok
)
33 yield dut
.oper_i
.zero_a
.eq(zero_a
)
34 yield dut
.issue_i
.eq(1)
36 yield dut
.issue_i
.eq(0)
38 if not imm_ok
or not zero_a
:
39 yield dut
.rd
.go
.eq(0b11)
42 rd_rel_o
= yield dut
.rd
.rel
43 print ("rd_rel", rd_rel_o
)
47 if len(dut
.src_i
) == 3:
48 yield dut
.rd
.go
.eq(0b100)
51 rd_rel_o
= yield dut
.rd
.rel
52 print ("rd_rel", rd_rel_o
)
57 req_rel_o
= yield dut
.wr
.rel
58 result
= yield dut
.data_o
59 print ("req_rel", req_rel_o
, result
)
61 req_rel_o
= yield dut
.wr
.rel
62 result
= yield dut
.data_o
63 print ("req_rel", req_rel_o
, result
)
67 yield dut
.wr
.go
[0].eq(1)
69 result
= yield dut
.data_o
71 print ("result", result
)
72 yield dut
.wr
.go
[0].eq(0)
77 def scoreboard_sim_dummy(dut
):
78 result
= yield from op_sim(dut
, 5, 2, InternalOp
.OP_NOP
, inv_a
=0,
80 assert result
== 5, result
82 result
= yield from op_sim(dut
, 9, 2, InternalOp
.OP_NOP
, inv_a
=0,
84 assert result
== 9, result
87 def scoreboard_sim(dut
):
88 result
= yield from op_sim(dut
, 5, 2, InternalOp
.OP_ADD
, inv_a
=0,
92 result
= yield from op_sim(dut
, 5, 2, InternalOp
.OP_ADD
)
95 result
= yield from op_sim(dut
, 5, 2, InternalOp
.OP_ADD
, inv_a
=1)
96 assert result
== 65532
98 result
= yield from op_sim(dut
, 5, 2, InternalOp
.OP_ADD
, zero_a
=1,
102 result
= yield from op_sim(dut
, 5, 2, InternalOp
.OP_ADD
, zero_a
=1)
110 dut
= MultiCompUnit(16, alu
, CompALUOpSubset
)
111 m
.submodules
.cu
= dut
113 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
114 with
open("test_compunit1.il", "w") as f
:
117 run_simulation(m
, scoreboard_sim(dut
), vcd_name
='test_compunit1.vcd')
120 class CompUnitParallelTest
:
121 def __init__(self
, dut
):
124 # Operation cycle should not take longer than this:
125 self
.MAX_BUSY_WAIT
= 50
127 # Minimum duration in which issue_i will be kept inactive,
128 # during which busy_o must remain low.
129 self
.MIN_BUSY_LOW
= 5
131 # Number of cycles to stall until the assertion of go.
132 # One value, for each port. Can be zero, for no delay.
133 self
.RD_GO_DELAY
= [0, 3]
135 # store common data for the input operation of the processes
138 self
.inv_a
= self
.zero_a
= 0
139 self
.imm
= self
.imm_ok
= 0
140 self
.imm_control
= (0, 0)
141 self
.rdmaskn
= (0, 0)
143 self
.operands
= (0, 0)
146 print("Begin parallel test.")
147 yield from self
.operation(5, 2, InternalOp
.OP_ADD
, inv_a
=0,
148 imm
=8, imm_ok
=0, rdmaskn
=(1, 0))
150 def operation(self
, a
, b
, op
, inv_a
=0, imm
=0, imm_ok
=0, zero_a
=0,
152 # store data for the operation
153 self
.operands
= (a
, b
)
159 self
.imm_control
= (zero_a
, imm_ok
)
160 self
.rdmaskn
= rdmaskn
162 # trigger operation cycle
163 yield from self
.issue()
166 # issue_i starts inactive
167 yield self
.dut
.issue_i
.eq(0)
169 for n
in range(self
.MIN_BUSY_LOW
):
171 # busy_o must remain inactive. It cannot rise on its own.
172 busy_o
= yield self
.dut
.busy_o
175 # activate issue_i to begin the operation cycle
176 yield self
.dut
.issue_i
.eq(1)
178 # at the same time, present the operation
179 yield self
.dut
.oper_i
.insn_type
.eq(self
.op
)
180 yield self
.dut
.oper_i
.invert_a
.eq(self
.inv_a
)
181 yield self
.dut
.oper_i
.imm_data
.imm
.eq(self
.imm
)
182 yield self
.dut
.oper_i
.imm_data
.imm_ok
.eq(self
.imm_ok
)
183 yield self
.dut
.oper_i
.zero_a
.eq(self
.zero_a
)
184 rdmaskn
= self
.rdmaskn
[0] |
(self
.rdmaskn
[1] << 1)
185 yield self
.dut
.rdmaskn
.eq(rdmaskn
)
187 # give one cycle for the CompUnit to latch the data
190 # busy_o must keep being low in this cycle, because issue_i was
191 # low on the previous cycle.
192 # It cannot rise on its own.
193 # Also, busy_o and issue_i must never be active at the same time, ever.
194 busy_o
= yield self
.dut
.busy_o
198 yield self
.dut
.issue_i
.eq(0)
200 # deactivate inputs along with issue_i, so we can be sure the data
201 # was latched at the correct cycle
202 # note: rdmaskn must be held, while busy_o is active
203 # TODO: deactivate rdmaskn when the busy_o cycle ends
204 yield self
.dut
.oper_i
.insn_type
.eq(0)
205 yield self
.dut
.oper_i
.invert_a
.eq(0)
206 yield self
.dut
.oper_i
.imm_data
.imm
.eq(0)
207 yield self
.dut
.oper_i
.imm_data
.imm_ok
.eq(0)
208 yield self
.dut
.oper_i
.zero_a
.eq(0)
211 # wait for busy_o to lower
212 # timeout after self.MAX_BUSY_WAIT cycles
213 for n
in range(self
.MAX_BUSY_WAIT
):
214 # sample busy_o in the current cycle
215 busy_o
= yield self
.dut
.busy_o
217 # operation cycle ends when busy_o becomes inactive
221 # if busy_o is still active, a timeout has occurred
222 # TODO: Uncomment this, once the test is complete:
226 print("If you are reading this, "
227 "it's because the above test failed, as expected,\n"
228 "with a timeout. It must pass, once the test is complete.")
231 print("If you are reading this, "
232 "it's because the above test unexpectedly passed.")
234 def rd(self
, rd_idx
):
235 # wait for issue_i to rise
237 issue_i
= yield self
.dut
.issue_i
240 # issue_i has not risen yet, so rd must keep low
241 rel
= yield self
.dut
.rd
.rel
[rd_idx
]
245 # we do not want rd to rise on an immediate operand
246 # if it is immediate, exit the process
247 # likewise, if the read mask is active
248 # TODO: don't exit the process, monitor rd instead to ensure it
249 # doesn't rise on its own
250 if self
.rdmaskn
[rd_idx
] or self
.imm_control
[rd_idx
]:
253 # issue_i has risen. rel must rise on the next cycle
254 rel
= yield self
.dut
.rd
.rel
[rd_idx
]
257 # stall for additional cycles. Check that rel doesn't fall on its own
258 for n
in range(self
.RD_GO_DELAY
[rd_idx
]):
260 rel
= yield self
.dut
.rd
.rel
[rd_idx
]
263 # Before asserting "go", make sure "rel" has risen.
264 # The use of Settle allows "go" to be set combinatorially,
265 # rising on the same cycle as "rel".
267 rel
= yield self
.dut
.rd
.rel
[rd_idx
]
270 # assert go for one cycle, passing along the operand value
271 yield self
.dut
.rd
.go
[rd_idx
].eq(1)
272 yield self
.dut
.src_i
[rd_idx
].eq(self
.operands
[rd_idx
])
273 # check that the operand was sent to the alu
274 # TODO: Properly check the alu protocol
276 alu_input
= yield self
.dut
.get_in(rd_idx
)
277 assert alu_input
== self
.operands
[rd_idx
]
280 # rel must keep high, since go was inactive in the last cycle
281 rel
= yield self
.dut
.rd
.rel
[rd_idx
]
284 # finish the go one-clock pulse
285 yield self
.dut
.rd
.go
[rd_idx
].eq(0)
286 yield self
.dut
.src_i
[rd_idx
].eq(0)
289 # rel must have gone low in response to go being high
290 # on the previous cycle
291 rel
= yield self
.dut
.rd
.rel
[rd_idx
]
294 def wr(self
, wr_idx
):
295 # monitor self.dut.wr.req[rd_idx] and sets dut.wr.go[idx] for one cycle
297 # TODO: also when dut.wr.go is set, check the output against the
298 # self.expected_o and assert. use dut.get_out(wr_idx) to do so.
300 def run_simulation(self
, vcd_name
):
301 run_simulation(self
.dut
, [self
.driver(),
302 self
.rd(0), # one read port (a)
303 self
.rd(1), # one read port (b)
304 self
.wr(0), # one write port (o)
309 def test_compunit_regspec3():
311 inspec
= [('INT', 'a', '0:15'),
312 ('INT', 'b', '0:15'),
313 ('INT', 'c', '0:15')]
314 outspec
= [('INT', 'o', '0:15'),
317 regspec
= (inspec
, outspec
)
321 dut
= MultiCompUnit(regspec
, alu
, CompALUOpSubset
)
322 m
.submodules
.cu
= dut
324 run_simulation(m
, scoreboard_sim_dummy(dut
),
325 vcd_name
='test_compunit_regspec3.vcd')
328 def test_compunit_regspec1():
330 inspec
= [('INT', 'a', '0:15'),
331 ('INT', 'b', '0:15')]
332 outspec
= [('INT', 'o', '0:15'),
335 regspec
= (inspec
, outspec
)
339 dut
= MultiCompUnit(regspec
, alu
, CompALUOpSubset
)
340 m
.submodules
.cu
= dut
342 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
343 with
open("test_compunit_regspec1.il", "w") as f
:
346 run_simulation(m
, scoreboard_sim(dut
),
347 vcd_name
='test_compunit_regspec1.vcd')
349 test
= CompUnitParallelTest(dut
)
350 test
.run_simulation("test_compunit_parallel.vcd")
353 if __name__
== '__main__':
355 test_compunit_regspec1()
356 test_compunit_regspec3()