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
23 from nmigen
.cli
import rtlil
25 from nmutil
.picker
import PriorityPicker
26 from nmutil
.util
import treereduce
28 from soc
.fu
.compunits
.compunits
import AllFunctionUnits
29 from soc
.regfile
.regfiles
import RegFiles
30 from soc
.decoder
.power_decoder
import create_pdecode
31 from soc
.decoder
.power_decoder2
import PowerDecode2
32 from soc
.experiment
.l0_cache
import TstL0CacheBuffer
# test only
36 # helper function for reducing a list of signals down to a parallel
38 def ortreereduce(tree
, attr
="data_o"):
39 return treereduce(tree
, operator
.or_
, lambda x
: getattr(x
, attr
))
41 # helper function to place full regs declarations first
42 def sort_fuspecs(fuspecs
):
44 for (regname
, fspec
) in fuspecs
.items():
45 if regname
.startswith("full"):
46 res
.append((regname
, fspec
))
47 for (regname
, fspec
) in fuspecs
.items():
48 if not regname
.startswith("full"):
49 res
.append((regname
, fspec
))
50 return res
# enumerate(res)
53 class NonProductionCore(Elaboratable
):
54 def __init__(self
, addrwid
=6):
55 self
.l0
= TstL0CacheBuffer(n_units
=1, regwid
=64, addrwid
=addrwid
)
56 pi
= self
.l0
.l0
.dports
[0].pi
58 self
.fus
= AllFunctionUnits(pilist
=[pi
], addrwid
=addrwid
)
59 self
.regs
= RegFiles()
60 self
.pdecode
= pdecode
= create_pdecode()
61 self
.pdecode2
= PowerDecode2(pdecode
) # instruction decoder
62 self
.ivalid_i
= self
.pdecode2
.e
.valid
# instruction is valid
63 self
.issue_i
= Signal(reset_less
=True)
64 self
.busy_o
= Signal(reset_less
=True)
66 def elaborate(self
, platform
):
69 m
.submodules
.pdecode2
= dec2
= self
.pdecode2
70 m
.submodules
.fus
= self
.fus
71 m
.submodules
.l0
= l0
= self
.l0
72 self
.regs
.elaborate_into(m
, platform
)
76 fu_bitdict
= self
.connect_instruction(m
)
77 self
.connect_rdports(m
, fu_bitdict
)
78 self
.connect_wrports(m
, fu_bitdict
)
82 def connect_instruction(self
, m
):
83 comb
, sync
= m
.d
.comb
, m
.d
.sync
87 # enable-signals for each FU, get one bit for each FU (by name)
88 fu_enable
= Signal(len(fus
), reset_less
=True)
90 for i
, funame
in enumerate(fus
.keys()):
91 fu_bitdict
[funame
] = fu_enable
[i
]
93 # connect up instructions. only one is enabled at any given time
94 for funame
, fu
in fus
.items():
95 fnunit
= fu
.fnunit
.value
96 enable
= Signal(name
="en_%s" % funame
, reset_less
=True)
97 comb
+= enable
.eq(self
.ivalid_i
& (dec2
.e
.fn_unit
& fnunit
).bool())
99 comb
+= fu
.oper_i
.eq_from_execute1(dec2
.e
)
100 comb
+= fu
.issue_i
.eq(self
.issue_i
)
101 comb
+= self
.busy_o
.eq(fu
.busy_o
)
102 rdmask
= dec2
.rdflags(fu
)
103 comb
+= fu
.rdmaskn
.eq(~rdmask
)
104 comb
+= fu_bitdict
[funame
].eq(enable
)
108 def connect_rdports(self
, m
, fu_bitdict
):
109 """connect read ports
111 orders the read regspecs into a dict-of-dicts, by regfile, by
112 regport name, then connects all FUs that want that regport by
113 way of a PriorityPicker.
115 comb
, sync
= m
.d
.comb
, m
.d
.sync
119 # dictionary of lists of regfile read ports
120 byregfiles_rd
, byregfiles_rdspec
= self
.get_byregfiles(True)
122 # okaay, now we need a PriorityPicker per regfile per regfile port
123 # loootta pickers... peter piper picked a pack of pickled peppers...
125 for regfile
, spec
in byregfiles_rd
.items():
126 fuspecs
= byregfiles_rdspec
[regfile
]
127 rdpickers
[regfile
] = {}
129 # for each named regfile port, connect up all FUs to that port
130 for (regname
, fspec
) in sort_fuspecs(fuspecs
):
131 print ("connect rd", regname
, fspec
)
133 # get the regfile specs for this regfile port
134 (rf
, read
, write
, wid
, fuspec
) = fspec
135 name
= "rdflag_%s_%s" % (regfile
, regname
)
136 rdflag
= Signal(name
=name
, reset_less
=True)
137 comb
+= rdflag
.eq(rf
)
139 # select the required read port. these are pre-defined sizes
140 print (rpidx
, regfile
, regs
.rf
.keys())
141 rport
= regs
.rf
[regfile
.lower()].r_ports
[rpidx
]
143 # create a priority picker to manage this port
144 rdpickers
[regfile
][rpidx
] = rdpick
= PriorityPicker(len(fuspec
))
145 setattr(m
.submodules
, "rdpick_%s_%s" % (regfile
, rpidx
), rdpick
)
147 # connect the regspec "reg select" number to this port
148 with m
.If(rdpick
.en_o
):
149 comb
+= rport
.ren
.eq(read
)
151 # connect up the FU req/go signals, and the reg-read to the FU
152 # and create a Read Broadcast Bus
153 for pi
, (funame
, fu
, idx
) in enumerate(fuspec
):
156 # connect request-read to picker input, and output to go-rd
157 fu_active
= fu_bitdict
[funame
]
158 pick
= fu
.rd_rel_o
[idx
] & fu_active
& rdflag
159 comb
+= rdpick
.i
[pi
].eq(pick
)
160 comb
+= fu
.go_rd_i
[idx
].eq(rdpick
.o
[pi
])
162 # connect regfile port to input, creating a Broadcast Bus
163 print ("reg connect widths",
164 regfile
, regname
, pi
, funame
,
165 src
.shape(), rport
.data_o
.shape())
166 comb
+= src
.eq(rport
.data_o
) # all FUs connect to same port
168 def connect_wrports(self
, m
, fu_bitdict
):
169 """connect write ports
171 orders the write regspecs into a dict-of-dicts, by regfile,
172 by regport name, then connects all FUs that want that regport
173 by way of a PriorityPicker.
175 note that the write-port wen, write-port data, and go_wr_i all need to
176 be on the exact same clock cycle. as there is a combinatorial loop bug
177 at the moment, these all use sync.
179 comb
, sync
= m
.d
.comb
, m
.d
.sync
182 # dictionary of lists of regfile write ports
183 byregfiles_wr
, byregfiles_wrspec
= self
.get_byregfiles(False)
185 # same for write ports.
186 # BLECH! complex code-duplication! BLECH!
188 for regfile
, spec
in byregfiles_wr
.items():
189 fuspecs
= byregfiles_wrspec
[regfile
]
190 wrpickers
[regfile
] = {}
191 for (regname
, fspec
) in sort_fuspecs(fuspecs
):
192 print ("connect wr", regname
, fspec
)
194 # get the regfile specs for this regfile port
195 (rf
, read
, write
, wid
, fuspec
) = fspec
197 # select the required write port. these are pre-defined sizes
198 print (regfile
, regs
.rf
.keys())
199 wport
= regs
.rf
[regfile
.lower()].w_ports
[rpidx
]
201 # create a priority picker to manage this port
202 wrpickers
[regfile
][rpidx
] = wrpick
= PriorityPicker(len(fuspec
))
203 setattr(m
.submodules
, "wrpick_%s_%s" % (regfile
, rpidx
), wrpick
)
205 # connect the regspec write "reg select" number to this port
206 # only if one FU actually requests (and is granted) the port
207 # will the write-enable be activated
208 with m
.If(wrpick
.en_o
):
209 sync
+= wport
.wen
.eq(write
)
211 sync
+= wport
.wen
.eq(0)
213 # connect up the FU req/go signals and the reg-read to the FU
214 # these are arbitrated by Data.ok signals
216 for pi
, (funame
, fu
, idx
) in enumerate(fuspec
):
217 # write-request comes from dest.ok
218 dest
= fu
.get_out(idx
)
219 name
= "wrflag_%s_%s_%d" % (funame
, regname
, idx
)
220 wrflag
= Signal(name
=name
, reset_less
=True)
221 comb
+= wrflag
.eq(dest
.ok
)
223 # connect request-read to picker input, and output to go-wr
224 fu_active
= fu_bitdict
[funame
]
225 pick
= fu
.wr
.rel
[idx
] & fu_active
#& wrflag
226 comb
+= wrpick
.i
[pi
].eq(pick
)
227 sync
+= fu
.go_wr_i
[idx
].eq(wrpick
.o
[pi
] & wrpick
.en_o
)
228 # connect regfile port to input
229 print ("reg connect widths",
230 regfile
, regname
, pi
, funame
,
231 dest
.shape(), wport
.data_i
.shape())
234 # here is where we create the Write Broadcast Bus. simple, eh?
235 sync
+= wport
.data_i
.eq(ortreereduce(wsigs
, "data"))
237 def get_byregfiles(self
, readmode
):
239 mode
= "read" if readmode
else "write"
244 # dictionary of lists of regfile ports
247 for (funame
, fu
) in fus
.items():
248 print ("%s ports for %s" % (mode
, funame
))
249 for idx
in range(fu
.n_src
if readmode
else fu
.n_dst
):
251 (regfile
, regname
, wid
) = fu
.get_in_spec(idx
)
253 (regfile
, regname
, wid
) = fu
.get_out_spec(idx
)
254 print (" %d %s %s %s" % (idx
, regfile
, regname
, str(wid
)))
256 rdflag
, read
= dec2
.regspecmap_read(regfile
, regname
)
259 rdflag
, read
= None, None
260 wrport
, write
= dec2
.regspecmap_write(regfile
, regname
)
261 if regfile
not in byregfiles
:
262 byregfiles
[regfile
] = {}
263 byregfiles_spec
[regfile
] = {}
264 if regname
not in byregfiles_spec
[regfile
]:
265 byregfiles_spec
[regfile
][regname
] = \
266 [rdflag
, read
, write
, wid
, []]
267 # here we start to create "lanes"
268 if idx
not in byregfiles
[regfile
]:
269 byregfiles
[regfile
][idx
] = []
270 fuspec
= (funame
, fu
, idx
)
271 byregfiles
[regfile
][idx
].append(fuspec
)
272 byregfiles_spec
[regfile
][regname
][4].append(fuspec
)
274 # ok just print that out, for convenience
275 for regfile
, spec
in byregfiles
.items():
276 print ("regfile %s ports:" % mode
, regfile
)
277 fuspecs
= byregfiles_spec
[regfile
]
278 for regname
, fspec
in fuspecs
.items():
279 [rdflag
, read
, write
, wid
, fuspec
] = fspec
280 print (" rf %s port %s lane: %s" % (mode
, regfile
, regname
))
281 print (" %s" % regname
, wid
, read
, write
, rdflag
)
282 for (funame
, fu
, idx
) in fuspec
:
283 fusig
= fu
.src_i
[idx
] if readmode
else fu
.dest
[idx
]
284 print (" ", funame
, fu
, idx
, fusig
)
287 return byregfiles
, byregfiles_spec
290 yield from self
.fus
.ports()
291 yield from self
.pdecode2
.ports()
298 if __name__
== '__main__':
299 dut
= NonProductionCore()
300 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
301 with
open("non_production_core.il", "w") as f
: