start putting LDSTSplitter together
[soc.git] / src / soc / scoreboard / fn_unit.py
1 from nmigen.compat.sim import run_simulation
2 from nmigen.cli import verilog, rtlil
3 from nmigen import Module, Signal, Cat, Array, Const, Elaboratable
4 from nmigen.lib.coding import Decoder
5
6 from nmutil.latch import SRLatch, latchregister
7
8 from soc.scoreboard.shadow import Shadow
9
10
11 class FnUnit(Elaboratable):
12 """ implements 11.4.8 function unit, p31
13 also implements optional shadowing 11.5.1, p55
14
15 shadowing can be used for branches as well as exceptions (interrupts),
16 load/store hold (exceptions again), and vector-element predication
17 (once the predicate is known, which it may not be at instruction issue)
18
19 Inputs
20
21 * :wid: register file width
22 * :shadow_wid: number of shadow/fail/good/go_die sets
23 * :n_dests: number of destination regfile(s) (index: rfile_sel_i)
24 * :wr_pend: if true, writable observes the g_wr_pend_i vector
25 otherwise observes g_rd_pend_i
26
27 notes:
28
29 * dest_i / src1_i / src2_i are in *binary*, whereas...
30 * ...g_rd_pend_i / g_wr_pend_i and rd_pend_o / wr_pend_o are UNARY
31 * req_rel_i (request release) is the direct equivalent of pipeline
32 "output valid" (valid_o)
33 * recover is a local python variable (actually go_die_o)
34 * when shadow_wid = 0, recover and shadown are Consts (i.e. do nothing)
35 * wr_pend is set False for the majority of uses: however for
36 use in a STORE Function Unit it is set to True
37 """
38 def __init__(self, wid, shadow_wid=0, n_dests=1, wr_pend=False):
39 self.reg_width = wid
40 self.n_dests = n_dests
41 self.shadow_wid = shadow_wid
42 self.wr_pend = wr_pend
43
44 # inputs
45 if n_dests > 1:
46 self.rfile_sel_i = Signal(max=n_dests, reset_less=True)
47 else:
48 self.rfile_sel_i = Const(0) # no selection. gets Array[0]
49 self.dest_i = Signal(max=wid, reset_less=True) # Dest R# in (top)
50 self.src1_i = Signal(max=wid, reset_less=True) # oper1 R# in (top)
51 self.src2_i = Signal(max=wid, reset_less=True) # oper2 R# in (top)
52 self.issue_i = Signal(reset_less=True) # Issue in (top)
53
54 self.go_wr_i = Signal(reset_less=True) # Go Write in (left)
55 self.go_rd_i = Signal(reset_less=True) # Go Read in (left)
56 self.req_rel_i = Signal(reset_less=True) # request release (left)
57
58 self.g_xx_pend_i = Array(Signal(wid, reset_less=True, name="g_pend_i") \
59 for i in range(n_dests)) # global rd (right)
60 self.g_wr_pend_i = Signal(wid, reset_less=True) # global wr (right)
61
62 if shadow_wid:
63 self.shadow_i = Signal(shadow_wid, reset_less=True)
64 self.s_fail_i = Signal(shadow_wid, reset_less=True)
65 self.s_good_i = Signal(shadow_wid, reset_less=True)
66 self.go_die_o = Signal(reset_less=True)
67
68 # outputs
69 self.readable_o = Signal(reset_less=True) # Readable out (right)
70 self.writable_o = Array(Signal(reset_less=True, name="writable_o") \
71 for i in range(n_dests)) # writable out (right)
72 self.busy_o = Signal(reset_less=True) # busy out (left)
73
74 self.src1_pend_o = Signal(wid, reset_less=True) # src1 pending
75 self.src2_pend_o = Signal(wid, reset_less=True) # src1 pending
76 self.rd_pend_o = Signal(wid, reset_less=True) # rd pending (right)
77 self.xx_pend_o = Array(Signal(wid, reset_less=True, name="pend_o") \
78 for i in range(n_dests))# wr pending (right)
79
80 def elaborate(self, platform):
81 m = Module()
82 m.submodules.rd_l = rd_l = SRLatch(sync=False)
83 m.submodules.wr_l = wr_l = SRLatch(sync=False)
84 m.submodules.dest_d = dest_d = Decoder(self.reg_width)
85 m.submodules.src1_d = src1_d = Decoder(self.reg_width)
86 m.submodules.src2_d = src2_d = Decoder(self.reg_width)
87
88 # shadow / recover (optional: shadow_wid > 0)
89 m.submodules.shadow = shadow = Shadow(self.shadow_wid)
90 if self.shadow_wid:
91 m.d.comb += shadow.issue_i.eq(self.issue_i)
92 m.d.comb += shadow.s_fail_i.eq(self.s_fail_i)
93 m.d.comb += shadow.s_good_i.eq(self.s_good_i)
94 m.d.comb += shadow.shadow_i.eq(self.shadow_i)
95 shadown = shadow.shadown_o
96 recover = shadow.go_die_o
97
98 # selector
99 xx_pend_o = self.xx_pend_o[self.rfile_sel_i]
100 writable_o = self.writable_o[self.rfile_sel_i]
101 g_pend_i = self.g_xx_pend_i[self.rfile_sel_i]
102
103 for i in range(self.n_dests):
104 m.d.comb += self.xx_pend_o[i].eq(0) # initialise all array
105 m.d.comb += self.writable_o[i].eq(0) # to zero
106 m.d.comb += self.readable_o.eq(0) # to zero
107
108 # go_wr latch: reset on go_wr HI, set on issue
109 m.d.comb += wr_l.s.eq(self.issue_i)
110 m.d.comb += wr_l.r.eq(self.go_wr_i | recover)
111
112 # src1 latch: reset on go_rd HI, set on issue
113 m.d.comb += rd_l.s.eq(self.issue_i)
114 m.d.comb += rd_l.r.eq(self.go_rd_i | recover)
115
116 # latch/registers for dest / src1 / src2
117 dest_r = Signal(max=self.reg_width, reset_less=True)
118 src1_r = Signal(max=self.reg_width, reset_less=True)
119 src2_r = Signal(max=self.reg_width, reset_less=True)
120 # XXX latch based on *issue* rather than !latch (as in book)
121 latchregister(m, self.dest_i, dest_r, self.issue_i) #wr_l.qn)
122 latchregister(m, self.src1_i, src1_r, self.issue_i) #wr_l.qn)
123 latchregister(m, self.src2_i, src2_r, self.issue_i) #wr_l.qn)
124
125 # dest decoder (use dest reg as input): write-pending out
126 m.d.comb += dest_d.i.eq(dest_r)
127 m.d.comb += dest_d.n.eq(wr_l.qn) # decode is inverted
128 m.d.comb += self.busy_o.eq(wr_l.q) # busy if set
129 m.d.comb += xx_pend_o.eq(dest_d.o)
130
131 # src1/src2 decoder (use src1/2 regs as input): read-pending out
132 m.d.comb += src1_d.i.eq(src1_r)
133 m.d.comb += src1_d.n.eq(rd_l.qn) # decode is inverted
134 m.d.comb += src2_d.i.eq(src2_r)
135 m.d.comb += src2_d.n.eq(rd_l.qn) # decode is inverted
136 m.d.comb += self.src1_pend_o.eq(src1_d.o)
137 m.d.comb += self.src2_pend_o.eq(src2_d.o)
138 m.d.comb += self.rd_pend_o.eq(src1_d.o | src2_d.o)
139
140 # readable output signal
141 g_rd = Signal(self.reg_width, reset_less=True)
142 ro = Signal(reset_less=True)
143 m.d.comb += g_rd.eq(~self.g_wr_pend_i & self.rd_pend_o)
144 m.d.comb += ro.eq(~g_rd.bool())
145 m.d.comb += self.readable_o.eq(ro)
146
147 # writable output signal
148 g_wr_v = Signal(self.reg_width, reset_less=True)
149 g_wr = Signal(reset_less=True)
150 wo = Signal(reset_less=True)
151 m.d.comb += g_wr_v.eq(g_pend_i & xx_pend_o)
152 m.d.comb += g_wr.eq(~g_wr_v.bool())
153 m.d.comb += wo.eq(g_wr & rd_l.qn & self.req_rel_i & shadown)
154 m.d.comb += writable_o.eq(wo)
155
156 return m
157
158 def __iter__(self):
159 yield self.dest_i
160 yield self.src1_i
161 yield self.src2_i
162 yield self.issue_i
163 yield self.go_wr_i
164 yield self.go_rd_i
165 yield self.req_rel_i
166 yield from self.g_xx_pend_i
167 yield self.g_wr_pend_i
168 yield self.readable_o
169 yield from self.writable_o
170 yield self.rd_pend_o
171 yield from self.xx_pend_o
172
173 def ports(self):
174 return list(self)
175
176 ############# ###############
177 # --- --- #
178 # --- renamed / redirected from base class --- #
179 # --- --- #
180 # --- below are convenience classes which match the names --- #
181 # --- of the various mitch alsup book chapter gate diagrams --- #
182 # --- --- #
183 ############# ###############
184
185
186 class IntFnUnit(FnUnit):
187 def __init__(self, wid, shadow_wid=0):
188 FnUnit.__init__(self, wid, shadow_wid)
189 self.int_rd_pend_o = self.rd_pend_o
190 self.int_wr_pend_o = self.xx_pend_o[0]
191 self.g_int_wr_pend_i = self.g_wr_pend_i
192 self.g_int_rd_pend_i = self.g_xx_pend_i[0]
193 self.int_readable_o = self.readable_o
194 self.int_writable_o = self.writable_o[0]
195
196 self.int_rd_pend_o.name = "int_rd_pend_o"
197 self.int_wr_pend_o.name = "int_wr_pend_o"
198 self.g_int_rd_pend_i.name = "g_int_rd_pend_i"
199 self.g_int_wr_pend_i.name = "g_int_wr_pend_i"
200 self.int_readable_o.name = "int_readable_o"
201 self.int_writable_o.name = "int_writable_o"
202
203
204 class FPFnUnit(FnUnit):
205 def __init__(self, wid, shadow_wid=0):
206 FnUnit.__init__(self, wid, shadow_wid)
207 self.fp_rd_pend_o = self.rd_pend_o
208 self.fp_wr_pend_o = self.xx_pend_o[0]
209 self.g_fp_wr_pend_i = self.g_wr_pend_i
210 self.g_fp_rd_pend_i = self.g_xx_pend_i[0]
211 self.fp_writable_o = self.writable_o[0]
212 self.fp_readable_o = self.readable_o
213
214 self.fp_rd_pend_o.name = "fp_rd_pend_o"
215 self.fp_wr_pend_o.name = "fp_wr_pend_o"
216 self.g_fp_rd_pend_i.name = "g_fp_rd_pend_i"
217 self.g_fp_wr_pend_i.name = "g_fp_wr_pend_i"
218 self.fp_writable_o.name = "fp_writable_o"
219 self.fp_readable_o.name = "fp_readable_o"
220
221
222 class LDFnUnit(FnUnit):
223 """ number of dest selectors: 2. assumes len(int_regfile) == len(fp_regfile)
224 * when rfile_sel_i == 0, int_wr_pend_o is set
225 * when rfile_sel_i == 1, fp_wr_pend_o is set
226 """
227 def __init__(self, wid, shadow_wid=0):
228 FnUnit.__init__(self, wid, shadow_wid, n_dests=2)
229 self.int_rd_pend_o = self.rd_pend_o
230 self.int_wr_pend_o = self.xx_pend_o[0]
231 self.fp_wr_pend_o = self.xx_pend_o[1]
232 self.g_int_wr_pend_i = self.g_wr_pend_i
233 self.g_int_rd_pend_i = self.g_xx_pend_i[0]
234 self.g_fp_rd_pend_i = self.g_xx_pend_i[1]
235 self.int_readable_o = self.readable_o
236 self.int_writable_o = self.writable_o[0]
237 self.fp_writable_o = self.writable_o[1]
238
239 self.int_rd_pend_o.name = "int_rd_pend_o"
240 self.int_wr_pend_o.name = "int_wr_pend_o"
241 self.fp_wr_pend_o.name = "fp_wr_pend_o"
242 self.g_int_wr_pend_i.name = "g_int_wr_pend_i"
243 self.g_int_rd_pend_i.name = "g_int_rd_pend_i"
244 self.g_fp_rd_pend_i.name = "g_fp_rd_pend_i"
245 self.int_readable_o.name = "int_readable_o"
246 self.int_writable_o.name = "int_writable_o"
247 self.fp_writable_o.name = "fp_writable_o"
248
249
250 class STFnUnit(FnUnit):
251 """ number of dest selectors: 2. assumes len(int_regfile) == len(fp_regfile)
252 * wr_pend=False indicates to observe global fp write pending
253 * when rfile_sel_i == 0, int_wr_pend_o is set
254 * when rfile_sel_i == 1, fp_wr_pend_o is set
255 *
256 """
257 def __init__(self, wid, shadow_wid=0):
258 FnUnit.__init__(self, wid, shadow_wid, n_dests=2, wr_pend=True)
259 self.int_rd_pend_o = self.rd_pend_o # 1st int read-pending vector
260 self.int2_rd_pend_o = self.xx_pend_o[0] # 2nd int read-pending vector
261 self.fp_rd_pend_o = self.xx_pend_o[1] # 1x FP read-pending vector
262 # yes overwrite FnUnit base class g_wr_pend_i vector
263 self.g_int_wr_pend_i = self.g_wr_pend_i = self.g_xx_pend_i[0]
264 self.g_fp_wr_pend_i = self.g_xx_pend_i[1]
265 self.int_readable_o = self.readable_o
266 self.int_writable_o = self.writable_o[0]
267 self.fp_writable_o = self.writable_o[1]
268
269 self.int_rd_pend_o.name = "int_rd_pend_o"
270 self.int2_rd_pend_o.name = "int2_rd_pend_o"
271 self.fp_rd_pend_o.name = "fp_rd_pend_o"
272 self.g_int_wr_pend_i.name = "g_int_wr_pend_i"
273 self.g_fp_wr_pend_i.name = "g_fp_wr_pend_i"
274 self.int_readable_o.name = "int_readable_o"
275 self.int_writable_o.name = "int_writable_o"
276 self.fp_writable_o.name = "fp_writable_o"
277
278
279
280 def int_fn_unit_sim(dut):
281 yield dut.dest_i.eq(1)
282 yield dut.issue_i.eq(1)
283 yield
284 yield dut.issue_i.eq(0)
285 yield
286 yield dut.src1_i.eq(1)
287 yield dut.issue_i.eq(1)
288 yield
289 yield
290 yield
291 yield dut.issue_i.eq(0)
292 yield
293 yield dut.go_rd_i.eq(1)
294 yield
295 yield dut.go_rd_i.eq(0)
296 yield
297 yield dut.go_wr_i.eq(1)
298 yield
299 yield dut.go_wr_i.eq(0)
300 yield
301
302 def test_int_fn_unit():
303 dut = FnUnit(32, 2, 2)
304 vl = rtlil.convert(dut, ports=dut.ports())
305 with open("test_fn_unit.il", "w") as f:
306 f.write(vl)
307
308 dut = LDFnUnit(32, 2)
309 vl = rtlil.convert(dut, ports=dut.ports())
310 with open("test_ld_fn_unit.il", "w") as f:
311 f.write(vl)
312
313 dut = STFnUnit(32, 0)
314 vl = rtlil.convert(dut, ports=dut.ports())
315 with open("test_st_fn_unit.il", "w") as f:
316 f.write(vl)
317
318 run_simulation(dut, int_fn_unit_sim(dut), vcd_name='test_fn_unit.vcd')
319
320 if __name__ == '__main__':
321 test_int_fn_unit()