3 not in any way intended for production use. connects up FunctionUnits to
4 Register Files in a brain-dead fashion that only permits one and only one
5 Function Unit to be operational.
7 the principle here is to take the Function Units, analyse their regspecs,
8 and turn their requirements for access to register file read/write ports
9 into groupings by Register File and Register File Port name.
11 under each grouping - by regfile/port - a list of Function Units that
12 need to connect to that port is created. as these are a contended
13 resource a "Broadcast Bus" per read/write port is then also created,
14 with access to it managed by a PriorityPicker.
16 the brain-dead part of this module is that even though there is no
17 conflict of access, regfile read/write hazards are *not* analysed,
18 and consequently it is safer to wait for the Function Unit to complete
19 before allowing a new instruction to proceed.
22 from nmigen
import Elaboratable
, Module
, Signal
, ResetSignal
, Cat
, Mux
23 from nmigen
.cli
import rtlil
25 from soc
.decoder
.power_decoder2
import PowerDecodeSubset
26 from soc
.decoder
.power_regspec_map
import regspec_decode_read
27 from soc
.decoder
.power_regspec_map
import regspec_decode_write
29 from nmutil
.picker
import PriorityPicker
30 from nmutil
.util
import treereduce
32 from soc
.fu
.compunits
.compunits
import AllFunctionUnits
33 from soc
.regfile
.regfiles
import RegFiles
34 from soc
.decoder
.decode2execute1
import Decode2ToExecute1Type
35 from soc
.decoder
.decode2execute1
import IssuerDecode2ToOperand
36 from soc
.decoder
.power_decoder2
import get_rdflags
37 from soc
.decoder
.decode2execute1
import Data
38 from soc
.experiment
.l0_cache
import TstL0CacheBuffer
# test only
39 from soc
.config
.test
.test_loadstore
import TestMemPspec
40 from soc
.decoder
.power_enums
import MicrOp
41 from soc
.config
.state
import CoreState
45 from nmutil
.util
import rising_edge
48 # helper function for reducing a list of signals down to a parallel
50 def ortreereduce(tree
, attr
="data_o"):
51 return treereduce(tree
, operator
.or_
, lambda x
: getattr(x
, attr
))
54 def ortreereduce_sig(tree
):
55 return treereduce(tree
, operator
.or_
, lambda x
: x
)
58 # helper function to place full regs declarations first
59 def sort_fuspecs(fuspecs
):
61 for (regname
, fspec
) in fuspecs
.items():
62 if regname
.startswith("full"):
63 res
.append((regname
, fspec
))
64 for (regname
, fspec
) in fuspecs
.items():
65 if not regname
.startswith("full"):
66 res
.append((regname
, fspec
))
67 return res
# enumerate(res)
70 class NonProductionCore(Elaboratable
):
71 def __init__(self
, pspec
):
74 # test to see if regfile ports should be reduced
75 self
.regreduce_en
= (hasattr(pspec
, "regreduce") and
76 (pspec
.regreduce
== True))
78 # single LD/ST funnel for memory access
79 self
.l0
= TstL0CacheBuffer(pspec
, n_units
=1)
80 pi
= self
.l0
.l0
.dports
[0]
82 # function units (only one each)
83 # only include mmu if enabled in pspec
84 self
.fus
= AllFunctionUnits(pspec
, pilist
=[pi
])
86 # register files (yes plural)
87 self
.regs
= RegFiles(pspec
)
89 # instruction decoder - needs a Trap-capable Record (captures EINT etc.)
90 self
.e
= Decode2ToExecute1Type("core", opkls
=IssuerDecode2ToOperand
)
92 # SVP64 RA_OR_ZERO needs to know if the relevant EXTRA2/3 field is zero
93 self
.sv_a_nz
= Signal()
95 # state and raw instruction
96 self
.state
= CoreState("core")
97 self
.raw_insn_i
= Signal(32) # raw instruction
98 self
.bigendian_i
= Signal() # bigendian - TODO, set by MSR.BE
100 # issue/valid/busy signalling
101 self
.ivalid_i
= Signal(reset_less
=True) # instruction is valid
102 self
.issue_i
= Signal(reset_less
=True)
103 self
.busy_o
= Signal(name
="corebusy_o", reset_less
=True)
105 # start/stop and terminated signalling
106 self
.core_stopped_i
= Signal(reset_less
=True)
107 self
.core_terminate_o
= Signal(reset
=0) # indicates stopped
109 # create per-FU instruction decoders (subsetted)
113 for funame
, fu
in self
.fus
.fus
.items():
114 f_name
= fu
.fnunit
.name
115 fnunit
= fu
.fnunit
.value
116 opkls
= fu
.opsubsetkls
118 self
.trapunit
= funame
120 self
.decoders
[funame
] = PowerDecodeSubset(None, opkls
, f_name
,
123 self
.des
[funame
] = self
.decoders
[funame
].do
125 if "mmu0" in self
.decoders
:
126 self
.decoders
["mmu0"].mmu0_spr_dec
= self
.decoders
["spr0"]
128 def elaborate(self
, platform
):
130 # for testing purposes, to cut down on build time in coriolis2
131 if hasattr(self
.pspec
, "nocore") and self
.pspec
.nocore
== True:
132 x
= Signal() # dummy signal
137 m
.submodules
.fus
= self
.fus
138 m
.submodules
.l0
= l0
= self
.l0
139 self
.regs
.elaborate_into(m
, platform
)
144 for k
, v
in self
.decoders
.items():
145 setattr(m
.submodules
, "dec_%s" % v
.fn_name
, v
)
146 comb
+= v
.dec
.raw_opcode_in
.eq(self
.raw_insn_i
)
147 comb
+= v
.dec
.bigendian
.eq(self
.bigendian_i
)
148 # sigh due to SVP64 RA_OR_ZERO detection connect these too
149 comb
+= v
.sv_a_nz
.eq(self
.sv_a_nz
)
151 # ssh, cheat: trap uses the main decoder because of the rewriting
152 self
.des
[self
.trapunit
] = self
.e
.do
154 # connect up Function Units, then read/write ports
155 fu_bitdict
= self
.connect_instruction(m
)
156 self
.connect_rdports(m
, fu_bitdict
)
157 self
.connect_wrports(m
, fu_bitdict
)
161 def connect_instruction(self
, m
):
162 """connect_instruction
164 uses decoded (from PowerOp) function unit information from CSV files
165 to ascertain which Function Unit should deal with the current
168 some (such as OP_ATTN, OP_NOP) are dealt with here, including
169 ignoring it and halting the processor. OP_NOP is a bit annoying
170 because the issuer expects busy flag still to be raised then lowered.
171 (this requires a fake counter to be set).
173 comb
, sync
= m
.d
.comb
, m
.d
.sync
176 # enable-signals for each FU, get one bit for each FU (by name)
177 fu_enable
= Signal(len(fus
), reset_less
=True)
179 for i
, funame
in enumerate(fus
.keys()):
180 fu_bitdict
[funame
] = fu_enable
[i
]
182 # enable the required Function Unit based on the opcode decode
183 # note: this *only* works correctly for simple core when one and
184 # *only* one FU is allocated per instruction
185 for funame
, fu
in fus
.items():
186 fnunit
= fu
.fnunit
.value
187 enable
= Signal(name
="en_%s" % funame
, reset_less
=True)
188 comb
+= enable
.eq((self
.e
.do
.fn_unit
& fnunit
).bool())
189 comb
+= fu_bitdict
[funame
].eq(enable
)
191 # sigh - need a NOP counter
193 with m
.If(counter
!= 0):
194 sync
+= counter
.eq(counter
- 1)
195 comb
+= self
.busy_o
.eq(1)
197 with m
.If(self
.ivalid_i
): # run only when valid
198 with m
.Switch(self
.e
.do
.insn_type
):
199 # check for ATTN: halt if true
200 with m
.Case(MicrOp
.OP_ATTN
):
201 m
.d
.sync
+= self
.core_terminate_o
.eq(1)
203 with m
.Case(MicrOp
.OP_NOP
):
204 sync
+= counter
.eq(2)
205 comb
+= self
.busy_o
.eq(1)
208 # connect up instructions. only one enabled at a time
209 for funame
, fu
in fus
.items():
210 do
= self
.des
[funame
]
211 enable
= fu_bitdict
[funame
]
213 # run this FunctionUnit if enabled
214 # route op, issue, busy, read flags and mask to FU
216 # operand comes from the *local* decoder
217 comb
+= fu
.oper_i
.eq_from(do
)
218 #comb += fu.oper_i.eq_from_execute1(e)
219 comb
+= fu
.issue_i
.eq(self
.issue_i
)
220 comb
+= self
.busy_o
.eq(fu
.busy_o
)
221 # rdmask, which is for registers, needs to come
222 # from the *main* decoder
223 rdmask
= get_rdflags(self
.e
, fu
)
224 comb
+= fu
.rdmaskn
.eq(~rdmask
)
228 def connect_rdport(self
, m
, fu_bitdict
, rdpickers
, regfile
, regname
, fspec
):
229 comb
, sync
= m
.d
.comb
, m
.d
.sync
235 # select the required read port. these are pre-defined sizes
236 rfile
= regs
.rf
[regfile
.lower()]
237 rport
= rfile
.r_ports
[rpidx
]
238 print("read regfile", rpidx
, regfile
, regs
.rf
.keys(),
242 if not isinstance(fspecs
, list):
249 for i
, fspec
in enumerate(fspecs
):
250 # get the regfile specs for this regfile port
251 (rf
, read
, write
, wid
, fuspec
) = fspec
252 print ("fpsec", i
, fspec
, len(fuspec
))
253 ppoffs
.append(pplen
) # record offset for picker
255 name
= "rdflag_%s_%s_%d" % (regfile
, regname
, i
)
256 rdflag
= Signal(name
=name
, reset_less
=True)
257 comb
+= rdflag
.eq(rf
)
258 rdflags
.append(rdflag
)
261 print ("pplen", pplen
)
263 # create a priority picker to manage this port
264 rdpickers
[regfile
][rpidx
] = rdpick
= PriorityPicker(pplen
)
265 setattr(m
.submodules
, "rdpick_%s_%s" % (regfile
, rpidx
), rdpick
)
269 for i
, fspec
in enumerate(fspecs
):
270 (rf
, read
, write
, wid
, fuspec
) = fspec
271 # connect up the FU req/go signals, and the reg-read to the FU
272 # and create a Read Broadcast Bus
273 for pi
, (funame
, fu
, idx
) in enumerate(fuspec
):
276 # connect request-read to picker input, and output to go-rd
277 fu_active
= fu_bitdict
[funame
]
278 name
= "%s_%s_%s_%i" % (regfile
, rpidx
, funame
, pi
)
279 addr_en
= Signal
.like(reads
[i
], name
="addr_en_"+name
)
280 pick
= Signal(name
="pick_"+name
) # picker input
281 rp
= Signal(name
="rp_"+name
) # picker output
282 delay_pick
= Signal(name
="dp_"+name
) # read-enable "underway"
284 # exclude any currently-enabled read-request (mask out active)
285 comb
+= pick
.eq(fu
.rd_rel_o
[idx
] & fu_active
& rdflags
[i
] &
287 comb
+= rdpick
.i
[pi
].eq(pick
)
288 comb
+= fu
.go_rd_i
[idx
].eq(delay_pick
) # pass in *delayed* pick
290 # if picked, select read-port "reg select" number to port
291 comb
+= rp
.eq(rdpick
.o
[pi
] & rdpick
.en_o
)
292 sync
+= delay_pick
.eq(rp
) # delayed "pick"
293 comb
+= addr_en
.eq(Mux(rp
, reads
[i
], 0))
295 # the read-enable happens combinatorially (see mux-bus below)
296 # but it results in the data coming out on a one-cycle delay.
300 addrs
.append(addr_en
)
303 # use the *delayed* pick signal to put requested data onto bus
304 with m
.If(delay_pick
):
305 # connect regfile port to input, creating fan-out Bus
307 print("reg connect widths",
308 regfile
, regname
, pi
, funame
,
309 src
.shape(), rport
.data_o
.shape())
310 # all FUs connect to same port
311 comb
+= src
.eq(rport
.data_o
)
313 # or-reduce the muxed read signals
315 # for unary-addressed
316 comb
+= rport
.ren
.eq(ortreereduce_sig(rens
))
318 # for binary-addressed
319 comb
+= rport
.addr
.eq(ortreereduce_sig(addrs
))
320 comb
+= rport
.ren
.eq(Cat(*rens
).bool())
321 print ("binary", regfile
, rpidx
, rport
, rport
.ren
, rens
, addrs
)
323 def connect_rdports(self
, m
, fu_bitdict
):
324 """connect read ports
326 orders the read regspecs into a dict-of-dicts, by regfile, by
327 regport name, then connects all FUs that want that regport by
328 way of a PriorityPicker.
330 comb
, sync
= m
.d
.comb
, m
.d
.sync
334 # dictionary of lists of regfile read ports
335 byregfiles_rd
, byregfiles_rdspec
= self
.get_byregfiles(True)
337 # okaay, now we need a PriorityPicker per regfile per regfile port
338 # loootta pickers... peter piper picked a pack of pickled peppers...
340 for regfile
, spec
in byregfiles_rd
.items():
341 fuspecs
= byregfiles_rdspec
[regfile
]
342 rdpickers
[regfile
] = {}
344 # argh. an experiment to merge RA and RB in the INT regfile
345 # (we have too many read/write ports)
346 if self
.regreduce_en
:
348 fuspecs
['rabc'] = [fuspecs
.pop('rb')]
349 fuspecs
['rabc'].append(fuspecs
.pop('rc'))
350 fuspecs
['rabc'].append(fuspecs
.pop('ra'))
351 if regfile
== 'FAST':
352 fuspecs
['fast1'] = [fuspecs
.pop('fast1')]
353 if 'fast2' in fuspecs
:
354 fuspecs
['fast1'].append(fuspecs
.pop('fast2'))
356 # for each named regfile port, connect up all FUs to that port
357 for (regname
, fspec
) in sort_fuspecs(fuspecs
):
358 print("connect rd", regname
, fspec
)
359 self
.connect_rdport(m
, fu_bitdict
, rdpickers
, regfile
,
362 def connect_wrport(self
, m
, fu_bitdict
, wrpickers
, regfile
, regname
, fspec
):
363 comb
, sync
= m
.d
.comb
, m
.d
.sync
367 print("connect wr", regname
, fspec
)
370 # select the required write port. these are pre-defined sizes
371 print(regfile
, regs
.rf
.keys())
372 rfile
= regs
.rf
[regfile
.lower()]
373 wport
= rfile
.w_ports
[rpidx
]
376 if not isinstance(fspecs
, list):
382 for i
, fspec
in enumerate(fspecs
):
383 # get the regfile specs for this regfile port
384 (rf
, read
, write
, wid
, fuspec
) = fspec
385 print ("fpsec", i
, fspec
, len(fuspec
))
386 ppoffs
.append(pplen
) # record offset for picker
389 # create a priority picker to manage this port
390 wrpickers
[regfile
][rpidx
] = wrpick
= PriorityPicker(pplen
)
391 setattr(m
.submodules
, "wrpick_%s_%s" % (regfile
, rpidx
), wrpick
)
396 for i
, fspec
in enumerate(fspecs
):
397 # connect up the FU req/go signals and the reg-read to the FU
398 # these are arbitrated by Data.ok signals
399 (rf
, read
, write
, wid
, fuspec
) = fspec
400 for pi
, (funame
, fu
, idx
) in enumerate(fuspec
):
403 # write-request comes from dest.ok
404 dest
= fu
.get_out(idx
)
405 fu_dest_latch
= fu
.get_fu_out(idx
) # latched output
406 name
= "wrflag_%s_%s_%d" % (funame
, regname
, idx
)
407 wrflag
= Signal(name
=name
, reset_less
=True)
408 comb
+= wrflag
.eq(dest
.ok
& fu
.busy_o
)
410 # connect request-write to picker input, and output to go-wr
411 fu_active
= fu_bitdict
[funame
]
412 pick
= fu
.wr
.rel_o
[idx
] & fu_active
# & wrflag
413 comb
+= wrpick
.i
[pi
].eq(pick
)
414 # create a single-pulse go write from the picker output
416 comb
+= wr_pick
.eq(wrpick
.o
[pi
] & wrpick
.en_o
)
417 comb
+= fu
.go_wr_i
[idx
].eq(rising_edge(m
, wr_pick
))
419 # connect the regspec write "reg select" number to this port
420 # only if one FU actually requests (and is granted) the port
421 # will the write-enable be activated
422 addr_en
= Signal
.like(write
)
424 comb
+= wp
.eq(wr_pick
& wrpick
.en_o
)
425 comb
+= addr_en
.eq(Mux(wp
, write
, 0))
429 addrs
.append(addr_en
)
432 # connect regfile port to input
433 print("reg connect widths",
434 regfile
, regname
, pi
, funame
,
435 dest
.shape(), wport
.data_i
.shape())
436 wsigs
.append(fu_dest_latch
)
438 # here is where we create the Write Broadcast Bus. simple, eh?
439 comb
+= wport
.data_i
.eq(ortreereduce_sig(wsigs
))
441 # for unary-addressed
442 comb
+= wport
.wen
.eq(ortreereduce_sig(wens
))
444 # for binary-addressed
445 comb
+= wport
.addr
.eq(ortreereduce_sig(addrs
))
446 comb
+= wport
.wen
.eq(ortreereduce_sig(wens
))
448 def connect_wrports(self
, m
, fu_bitdict
):
449 """connect write ports
451 orders the write regspecs into a dict-of-dicts, by regfile,
452 by regport name, then connects all FUs that want that regport
453 by way of a PriorityPicker.
455 note that the write-port wen, write-port data, and go_wr_i all need to
456 be on the exact same clock cycle. as there is a combinatorial loop bug
457 at the moment, these all use sync.
459 comb
, sync
= m
.d
.comb
, m
.d
.sync
462 # dictionary of lists of regfile write ports
463 byregfiles_wr
, byregfiles_wrspec
= self
.get_byregfiles(False)
465 # same for write ports.
466 # BLECH! complex code-duplication! BLECH!
468 for regfile
, spec
in byregfiles_wr
.items():
469 fuspecs
= byregfiles_wrspec
[regfile
]
470 wrpickers
[regfile
] = {}
472 if self
.regreduce_en
:
473 # argh, more port-merging
475 fuspecs
['o'] = [fuspecs
.pop('o')]
476 fuspecs
['o'].append(fuspecs
.pop('o1'))
477 if regfile
== 'FAST':
478 fuspecs
['fast1'] = [fuspecs
.pop('fast1')]
479 if 'fast2' in fuspecs
:
480 fuspecs
['fast1'].append(fuspecs
.pop('fast2'))
482 for (regname
, fspec
) in sort_fuspecs(fuspecs
):
483 self
.connect_wrport(m
, fu_bitdict
, wrpickers
,
484 regfile
, regname
, fspec
)
486 def get_byregfiles(self
, readmode
):
488 mode
= "read" if readmode
else "write"
491 e
= self
.e
# decoded instruction to execute
493 # dictionary of lists of regfile ports
496 for (funame
, fu
) in fus
.items():
497 print("%s ports for %s" % (mode
, funame
))
498 for idx
in range(fu
.n_src
if readmode
else fu
.n_dst
):
500 (regfile
, regname
, wid
) = fu
.get_in_spec(idx
)
502 (regfile
, regname
, wid
) = fu
.get_out_spec(idx
)
503 print(" %d %s %s %s" % (idx
, regfile
, regname
, str(wid
)))
505 rdflag
, read
= regspec_decode_read(e
, regfile
, regname
)
508 rdflag
, read
= None, None
509 wrport
, write
= regspec_decode_write(e
, regfile
, regname
)
510 if regfile
not in byregfiles
:
511 byregfiles
[regfile
] = {}
512 byregfiles_spec
[regfile
] = {}
513 if regname
not in byregfiles_spec
[regfile
]:
514 byregfiles_spec
[regfile
][regname
] = \
515 (rdflag
, read
, write
, wid
, [])
516 # here we start to create "lanes"
517 if idx
not in byregfiles
[regfile
]:
518 byregfiles
[regfile
][idx
] = []
519 fuspec
= (funame
, fu
, idx
)
520 byregfiles
[regfile
][idx
].append(fuspec
)
521 byregfiles_spec
[regfile
][regname
][4].append(fuspec
)
523 # ok just print that out, for convenience
524 for regfile
, spec
in byregfiles
.items():
525 print("regfile %s ports:" % mode
, regfile
)
526 fuspecs
= byregfiles_spec
[regfile
]
527 for regname
, fspec
in fuspecs
.items():
528 [rdflag
, read
, write
, wid
, fuspec
] = fspec
529 print(" rf %s port %s lane: %s" % (mode
, regfile
, regname
))
530 print(" %s" % regname
, wid
, read
, write
, rdflag
)
531 for (funame
, fu
, idx
) in fuspec
:
532 fusig
= fu
.src_i
[idx
] if readmode
else fu
.dest
[idx
]
533 print(" ", funame
, fu
, idx
, fusig
)
536 return byregfiles
, byregfiles_spec
539 yield from self
.fus
.ports()
540 yield from self
.e
.ports()
541 yield from self
.l0
.ports()
548 if __name__
== '__main__':
549 pspec
= TestMemPspec(ldst_ifacetype
='testpi',
554 dut
= NonProductionCore(pspec
)
555 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
556 with
open("test_core.il", "w") as f
: