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
import Module
, Signal
, Mux
, Elaboratable
, Repl
, Cat
, Const
14 from nmigen
.hdl
.rec
import (Record
, DIR_FANIN
, DIR_FANOUT
)
16 from nmutil
.latch
import SRLatch
, latchregister
17 from nmutil
.iocontrol
import RecordObject
18 from nmutil
.util
import rising_edge
20 from soc
.fu
.regspec
import RegSpec
, RegSpecALUAPI
24 """find_ok helper function - finds field ending in "_ok"
26 for field_name
in fields
:
27 if field_name
.endswith("_ok"):
32 def go_record(n
, name
):
33 r
= Record([('go_i', n
, DIR_FANIN
),
34 ('rel_o', n
, DIR_FANOUT
)], name
=name
)
35 r
.go_i
.reset_less
= True
36 r
.rel_o
.reset_less
= True
40 # see https://libre-soc.org/3d_gpu/architecture/regfile/ section on regspecs
42 class CompUnitRecord(RegSpec
, RecordObject
):
45 base class for Computation Units, to provide a uniform API
46 and allow "record.connect" etc. to be used, particularly when
47 it comes to connecting multiple Computation Units up as a block
50 LDSTCompUnitRecord should derive from this class and add the
51 additional signals it requires
53 :subkls: the class (not an instance) needed to construct the opcode
54 :rwid: either an integer (specifies width of all regs) or a "regspec"
56 see https://libre-soc.org/3d_gpu/architecture/regfile/ section on regspecs
59 def __init__(self
, subkls
, rwid
, n_src
=None, n_dst
=None, name
=None):
60 RegSpec
.__init
__(self
, rwid
, n_src
, n_dst
)
62 RecordObject
.__init
__(self
)
64 n_src
, n_dst
= self
._n
_src
, self
._n
_dst
66 # create source operands
68 for i
in range(n_src
):
69 j
= i
+ 1 # name numbering to match src1/src2
71 rw
= self
._get
_srcwid
(i
)
72 sreg
= Signal(rw
, name
=sname
, reset_less
=True)
73 setattr(self
, sname
, sreg
)
77 # create dest operands
79 for i
in range(n_dst
):
80 j
= i
+ 1 # name numbering to match dest1/2...
81 dname
= "dest%d_o" % j
82 rw
= self
._get
_dstwid
(i
)
83 # dreg = Data(rw, name=name) XXX ??? output needs to be a Data type?
84 dreg
= Signal(rw
, name
=dname
, reset_less
=True)
85 setattr(self
, dname
, dreg
)
89 # operation / data input
90 self
.oper_i
= subkls(name
="oper_i_%s" % name
) # operand
92 # create read/write and other scoreboard signalling
93 self
.rd
= go_record(n_src
, name
="cu_rd") # read in, req out
94 self
.wr
= go_record(n_dst
, name
="cu_wr") # write in, req out
96 self
.rdmaskn
= Signal(n_src
, name
="cu_rdmaskn_i", reset_less
=True)
97 self
.wrmask
= Signal(n_dst
, name
="cu_wrmask_o", reset_less
=True)
100 self
.issue_i
= Signal(name
="cu_issue_i", reset_less
=True)
101 # shadow function, defaults to ON
102 self
.shadown_i
= Signal(name
="cu_shadown_i", reset
=1)
104 self
.go_die_i
= Signal(name
="cu_go_die_i")
107 self
.busy_o
= Signal(name
="cu_busy_o", reset_less
=True) # fn busy out
108 self
.done_o
= Signal(name
="cu_done_o", reset_less
=True)
109 self
.alu_done_o
= Signal(name
="cu_alu_done_o", reset_less
=True)
112 class MultiCompUnit(RegSpecALUAPI
, Elaboratable
):
113 def __init__(self
, rwid
, alu
, opsubsetkls
, n_src
=2, n_dst
=1, name
=None):
116 * :rwid: width of register latches (TODO: allocate per regspec)
117 * :alu: ALU (pipeline, FSM) - must conform to nmutil Pipe API
118 * :opsubsetkls: subset of Decode2ExecuteType
119 * :n_src: number of src operands
120 * :n_dst: number of destination operands
122 RegSpecALUAPI
.__init
__(self
, rwid
, alu
)
123 self
.alu_name
= name
or "alu"
124 self
.opsubsetkls
= opsubsetkls
125 self
.cu
= cu
= CompUnitRecord(opsubsetkls
, rwid
, n_src
, n_dst
,
127 n_src
, n_dst
= self
.n_src
, self
.n_dst
= cu
._n
_src
, cu
._n
_dst
128 print("n_src %d n_dst %d" % (self
.n_src
, self
.n_dst
))
130 # convenience names for src operands
131 for i
in range(n_src
):
132 j
= i
+ 1 # name numbering to match src1/src2
134 setattr(self
, name
, getattr(cu
, name
))
136 # convenience names for dest operands
137 for i
in range(n_dst
):
138 j
= i
+ 1 # name numbering to match dest1/2...
139 name
= "dest%d_o" % j
140 setattr(self
, name
, getattr(cu
, name
))
142 # more convenience names
145 self
.rdmaskn
= cu
.rdmaskn
146 self
.wrmask
= cu
.wrmask
147 self
.alu_done_o
= cu
.alu_done_o
148 self
.go_rd_i
= self
.rd
.go_i
# temporary naming
149 self
.go_wr_i
= self
.wr
.go_i
# temporary naming
150 self
.rd_rel_o
= self
.rd
.rel_o
# temporary naming
151 self
.req_rel_o
= self
.wr
.rel_o
# temporary naming
152 self
.issue_i
= cu
.issue_i
153 self
.shadown_i
= cu
.shadown_i
154 self
.go_die_i
= cu
.go_die_i
156 # operation / data input
157 self
.oper_i
= cu
.oper_i
158 self
.src_i
= cu
._src
_i
160 self
.busy_o
= cu
.busy_o
162 self
.o_data
= self
.dest
[0] # Dest out
163 self
.done_o
= cu
.done_o
165 def _mux_op(self
, m
, sl
, op_is_imm
, imm
, i
):
166 # select imm if opcode says so. however also change the latch
167 # to trigger *from* the opcode latch instead.
168 src_or_imm
= Signal(self
.cu
._get
_srcwid
(i
), reset_less
=True)
169 src_sel
= Signal(reset_less
=True)
170 m
.d
.comb
+= src_sel
.eq(Mux(op_is_imm
, self
.opc_l
.q
, sl
[i
][2]))
171 m
.d
.comb
+= src_or_imm
.eq(Mux(op_is_imm
, imm
, self
.src_i
[i
]))
172 # overwrite 1st src-latch with immediate-muxed stuff
173 sl
[i
][0] = src_or_imm
175 sl
[i
][3] = ~op_is_imm
# change rd.rel[i] gate condition
177 def elaborate(self
, platform
):
179 # add the ALU to the MultiCompUnit only if it is a "real" ALU
180 # see AllFunctionUnits as to why: a FunctionUnitBaseMulti
181 # only has one "real" ALU but multiple pseudo front-ends,
182 # aka "ReservationStations" (ALUProxy "fronts")
183 if isinstance(self
.alu
, Elaboratable
):
184 setattr(m
.submodules
, self
.alu_name
, self
.alu
)
185 m
.submodules
.src_l
= src_l
= SRLatch(False, self
.n_src
, name
="src")
186 m
.submodules
.opc_l
= opc_l
= SRLatch(sync
=False, name
="opc")
187 m
.submodules
.req_l
= req_l
= SRLatch(False, self
.n_dst
, name
="req")
188 m
.submodules
.rst_l
= rst_l
= SRLatch(sync
=False, name
="rst")
189 m
.submodules
.rok_l
= rok_l
= SRLatch(sync
=False, name
="rdok")
190 self
.opc_l
, self
.src_l
= opc_l
, src_l
192 # ALU only proceeds when all src are ready. rd_rel_o is delayed
193 # so combine it with go_rd_i. if all bits are set we're good
194 all_rd
= Signal(reset_less
=True)
195 m
.d
.comb
+= all_rd
.eq(self
.busy_o
& rok_l
.q
&
196 (((~self
.rd
.rel_o
) | self
.rd
.go_i
).all()))
198 # generate read-done pulse
199 all_rd_pulse
= Signal(reset_less
=True)
200 m
.d
.comb
+= all_rd_pulse
.eq(rising_edge(m
, all_rd
))
202 # create rising pulse from alu valid condition.
203 alu_done
= self
.cu
.alu_done_o
204 alu_pulse
= Signal(reset_less
=True)
205 alu_pulsem
= Signal(self
.n_dst
, reset_less
=True)
206 m
.d
.comb
+= alu_done
.eq(self
.alu
.n
.o_valid
)
207 m
.d
.comb
+= alu_pulse
.eq(rising_edge(m
, alu_done
))
208 m
.d
.comb
+= alu_pulsem
.eq(Repl(alu_pulse
, self
.n_dst
))
210 # sigh bug where req_l gets both set and reset raised at same time
211 prev_wr_go
= Signal(self
.n_dst
)
212 brd
= Repl(self
.busy_o
, self
.n_dst
)
213 m
.d
.sync
+= prev_wr_go
.eq(self
.wr
.go_i
& brd
)
215 # write_requests all done
216 # req_done works because any one of the last of the writes
217 # is enough, when combined with when read-phase is done (rst_l.q)
218 wr_any
= Signal(reset_less
=True)
219 req_done
= Signal(reset_less
=True)
220 m
.d
.comb
+= self
.done_o
.eq(self
.busy_o
& ~
(self
.wr
.rel_o
).bool())
221 m
.d
.comb
+= wr_any
.eq(self
.wr
.go_i
.bool() | prev_wr_go
.bool())
222 m
.d
.comb
+= req_done
.eq(wr_any
& ~self
.alu
.n
.i_ready
& (req_l
.q
== 0))
223 # argh, complicated hack: if there are no regs to write,
224 # instead of waiting for regs that are never going to happen,
225 # we indicate "done" when the ALU is "done"
226 with m
.If((self
.wrmask
== 0) &
227 self
.alu
.n
.i_ready
& self
.alu
.n
.o_valid
& self
.busy_o
):
228 m
.d
.comb
+= req_done
.eq(1)
231 reset
= Signal(reset_less
=True)
232 rst_r
= Signal(reset_less
=True) # reset latch off
233 reset_w
= Signal(self
.n_dst
, reset_less
=True)
234 reset_r
= Signal(self
.n_src
, reset_less
=True)
235 m
.d
.comb
+= reset
.eq(req_done | self
.go_die_i
)
236 m
.d
.comb
+= rst_r
.eq(self
.issue_i | self
.go_die_i
)
237 m
.d
.comb
+= reset_w
.eq(self
.wr
.go_i |
Repl(self
.go_die_i
, self
.n_dst
))
238 m
.d
.comb
+= reset_r
.eq(self
.rd
.go_i |
Repl(self
.go_die_i
, self
.n_src
))
240 # read-done,wr-proceed latch
241 m
.d
.sync
+= rok_l
.s
.eq(self
.issue_i
) # set up when issue starts
242 m
.d
.sync
+= rok_l
.r
.eq(self
.alu
.n
.o_valid
& self
.busy_o
) # ALU done
244 # wr-done, back-to-start latch
245 m
.d
.sync
+= rst_l
.s
.eq(all_rd
) # set when read-phase is fully done
246 m
.d
.sync
+= rst_l
.r
.eq(rst_r
) # *off* on issue
248 # opcode latch (not using go_rd_i) - inverted so that busy resets to 0
249 m
.d
.sync
+= opc_l
.s
.eq(self
.issue_i
) # set on issue
250 m
.d
.sync
+= opc_l
.r
.eq(req_done
) # reset on ALU
252 # src operand latch (not using go_wr_i)
253 m
.d
.sync
+= src_l
.s
.eq(Repl(self
.issue_i
, self
.n_src
) & ~self
.rdmaskn
)
254 m
.d
.sync
+= src_l
.r
.eq(reset_r
)
256 # dest operand latch (not using issue_i)
257 m
.d
.sync
+= req_l
.s
.eq(alu_pulsem
& self
.wrmask
)
258 m
.d
.sync
+= req_l
.r
.eq(reset_w | prev_wr_go
)
260 # pass operation to the ALU (sync: plenty time to wait for src reads)
262 with m
.If(self
.issue_i
):
263 m
.d
.sync
+= op
.eq(self
.oper_i
)
265 # and for each output from the ALU: capture when ALU output is valid
268 for i
in range(self
.n_dst
):
269 name
= "data_r%d" % i
270 lro
= self
.get_out(i
)
272 data_r_ok
= Const(1, 1)
273 if isinstance(lro
, Record
):
274 data_r
= Record
.like(lro
, name
=name
)
275 print("wr fields", i
, lro
, data_r
.fields
)
276 # bye-bye abstract interface design..
277 fname
= find_ok(data_r
.fields
)
279 ok
= getattr(lro
, fname
)
280 data_r_ok
= getattr(data_r
, fname
)
281 # write-ok based on incoming output *and* whether the latched
283 # XXX fails - wrok.append((ok|data_r_ok) & self.busy_o)
284 wrok
.append(ok
& self
.busy_o
)
286 # really should retire this but it's part of unit tests
287 data_r
= Signal
.like(lro
, name
=name
, reset_less
=True)
288 wrok
.append(ok
& self
.busy_o
)
289 with m
.If(alu_pulse
):
290 m
.d
.sync
+= data_r
.eq(lro
)
291 with m
.If(self
.issue_i
):
292 m
.d
.sync
+= data_r
.eq(0)
295 # ok, above we collated anything with an "ok" on the output side
296 # now actually use those to create a write-mask. this basically
297 # is now the Function Unit API tells the Comp Unit "do not request
298 # a regfile port because this particular output is not valid"
299 m
.d
.comb
+= self
.wrmask
.eq(Cat(*wrok
))
301 # create list of src/alu-src/src-latch. override 1st and 2nd one below.
302 # in the case, for ALU and Logical pipelines, we assume RB is the
303 # 2nd operand in the input "regspec". see for example
304 # soc.fu.alu.pipe_data.ALUInputData
306 print("src_i", self
.src_i
)
307 for i
in range(self
.n_src
):
308 sl
.append([self
.src_i
[i
], self
.get_in(i
), src_l
.q
[i
], Const(1, 1)])
310 # if the operand subset has "zero_a" we implicitly assume that means
311 # src_i[0] is an INT reg type where zero can be multiplexed in, instead.
312 # see https://bugs.libre-soc.org/show_bug.cgi?id=336
313 if hasattr(op
, "zero_a"):
314 # select zero imm if opcode says so. however also change the latch
315 # to trigger *from* the opcode latch instead.
316 self
._mux
_op
(m
, sl
, op
.zero_a
, 0, 0)
318 # if the operand subset has "imm_data" we implicitly assume that means
319 # "this is an INT ALU/Logical FU jobbie, RB is muxed with the immediate"
320 if hasattr(op
, "imm_data"):
321 # select immediate if opcode says so. however also change the latch
322 # to trigger *from* the opcode latch instead.
323 op_is_imm
= op
.imm_data
.ok
324 imm
= op
.imm_data
.data
325 self
._mux
_op
(m
, sl
, op_is_imm
, imm
, 1)
327 # create a latch/register for src1/src2 (even if it is a copy of imm)
328 for i
in range(self
.n_src
):
329 src
, alusrc
, latch
, _
= sl
[i
]
330 reg
= latchregister(m
, src
, alusrc
, latch
, name
="src_r%d" % i
)
331 # rdmask stops src latches from being set. clear all if not busy
332 with m
.If(~self
.busy_o
):
333 m
.d
.sync
+= reg
.eq(0)
336 # ALU connection / interaction
339 # on a go_read, tell the ALU we're accepting data.
340 m
.submodules
.alui_l
= alui_l
= SRLatch(False, name
="alui")
341 m
.d
.comb
+= self
.alu
.p
.i_valid
.eq(alui_l
.q
)
342 m
.d
.sync
+= alui_l
.r
.eq(self
.alu
.p
.o_ready
& alui_l
.q
)
343 m
.d
.comb
+= alui_l
.s
.eq(all_rd_pulse
)
345 # ALU output "ready" side. alu "ready" indication stays hi until
347 m
.submodules
.alu_l
= alu_l
= SRLatch(False, name
="alu")
348 m
.d
.comb
+= self
.alu
.n
.i_ready
.eq(alu_l
.q
)
349 m
.d
.sync
+= alu_l
.r
.eq(self
.alu
.n
.o_valid
& alu_l
.q
)
350 m
.d
.comb
+= alu_l
.s
.eq(all_rd_pulse
)
356 slg
= Cat(*map(lambda x
: x
[3], sl
)) # get req gate conditions
357 # all request signals gated by busy_o. prevents picker problems
358 m
.d
.comb
+= self
.busy_o
.eq(opc_l
.q
) # busy out
360 # read-release gated by busy (and read-mask)
361 bro
= Repl(self
.busy_o
, self
.n_src
)
362 m
.d
.comb
+= self
.rd
.rel_o
.eq(src_l
.q
& bro
& slg
)
364 # write-release gated by busy and by shadow (and write-mask)
365 brd
= Repl(self
.busy_o
& self
.shadown_i
, self
.n_dst
)
366 m
.d
.comb
+= self
.wr
.rel_o
.eq(req_l
.q
& brd
)
368 # output the data from the latch on go_write
369 for i
in range(self
.n_dst
):
370 with m
.If(self
.wr
.go_i
[i
] & self
.busy_o
):
371 m
.d
.comb
+= self
.dest
[i
].eq(drl
[i
])
375 def get_fu_out(self
, i
):
384 yield from self
.oper_i
.ports()