use FPPackData in specialcases mod
[ieee754fpu.git] / src / add / nmigen_add_experiment.py
1 # IEEE Floating Point Adder (Single Precision)
2 # Copyright (C) Jonathan P Dawson 2013
3 # 2013-12-12
4
5 from nmigen import Module, Signal, Cat, Mux, Array, Const
6 from nmigen.lib.coding import PriorityEncoder
7 from nmigen.cli import main, verilog
8 from math import log
9
10 from fpbase import FPNumIn, FPNumOut, FPOp, Overflow, FPBase, FPNumBase
11 from fpbase import MultiShiftRMerge, Trigger
12 #from fpbase import FPNumShiftMultiRight
13
14
15 class FPState(FPBase):
16 def __init__(self, state_from):
17 self.state_from = state_from
18
19 def set_inputs(self, inputs):
20 self.inputs = inputs
21 for k,v in inputs.items():
22 setattr(self, k, v)
23
24 def set_outputs(self, outputs):
25 self.outputs = outputs
26 for k,v in outputs.items():
27 setattr(self, k, v)
28
29
30 class FPGetSyncOpsMod:
31 def __init__(self, width, num_ops=2):
32 self.width = width
33 self.num_ops = num_ops
34 inops = []
35 outops = []
36 for i in range(num_ops):
37 inops.append(Signal(width, reset_less=True))
38 outops.append(Signal(width, reset_less=True))
39 self.in_op = inops
40 self.out_op = outops
41 self.stb = Signal(num_ops)
42 self.ack = Signal()
43 self.ready = Signal(reset_less=True)
44 self.out_decode = Signal(reset_less=True)
45
46 def elaborate(self, platform):
47 m = Module()
48 m.d.comb += self.ready.eq(self.stb == Const(-1, (self.num_ops, False)))
49 m.d.comb += self.out_decode.eq(self.ack & self.ready)
50 with m.If(self.out_decode):
51 for i in range(self.num_ops):
52 m.d.comb += [
53 self.out_op[i].eq(self.in_op[i]),
54 ]
55 return m
56
57 def ports(self):
58 return self.in_op + self.out_op + [self.stb, self.ack]
59
60
61 class FPOps(Trigger):
62 def __init__(self, width, num_ops):
63 Trigger.__init__(self)
64 self.width = width
65 self.num_ops = num_ops
66
67 res = []
68 for i in range(num_ops):
69 res.append(Signal(width))
70 self.v = Array(res)
71
72 def ports(self):
73 res = []
74 for i in range(self.num_ops):
75 res.append(self.v[i])
76 res.append(self.ack)
77 res.append(self.stb)
78 return res
79
80
81 class InputGroup:
82 def __init__(self, width, num_ops=2, num_rows=4):
83 self.width = width
84 self.num_ops = num_ops
85 self.num_rows = num_rows
86 self.mmax = int(log(self.num_rows) / log(2))
87 self.rs = []
88 self.mid = Signal(self.mmax, reset_less=True) # multiplex id
89 for i in range(num_rows):
90 self.rs.append(FPGetSyncOpsMod(width, num_ops))
91 self.rs = Array(self.rs)
92
93 self.out_op = FPOps(width, num_ops)
94
95 def elaborate(self, platform):
96 m = Module()
97
98 pe = PriorityEncoder(self.num_rows)
99 m.submodules.selector = pe
100 m.submodules.out_op = self.out_op
101 m.submodules += self.rs
102
103 # connect priority encoder
104 in_ready = []
105 for i in range(self.num_rows):
106 in_ready.append(self.rs[i].ready)
107 m.d.comb += pe.i.eq(Cat(*in_ready))
108
109 active = Signal(reset_less=True)
110 out_en = Signal(reset_less=True)
111 m.d.comb += active.eq(~pe.n) # encoder active
112 m.d.comb += out_en.eq(active & self.out_op.trigger)
113
114 # encoder active: ack relevant input, record MID, pass output
115 with m.If(out_en):
116 rs = self.rs[pe.o]
117 m.d.sync += self.mid.eq(pe.o)
118 m.d.sync += rs.ack.eq(0)
119 m.d.sync += self.out_op.stb.eq(0)
120 for j in range(self.num_ops):
121 m.d.sync += self.out_op.v[j].eq(rs.out_op[j])
122 with m.Else():
123 m.d.sync += self.out_op.stb.eq(1)
124 # acks all default to zero
125 for i in range(self.num_rows):
126 m.d.sync += self.rs[i].ack.eq(1)
127
128 return m
129
130 def ports(self):
131 res = []
132 for i in range(self.num_rows):
133 inop = self.rs[i]
134 res += inop.in_op + [inop.stb]
135 return self.out_op.ports() + res + [self.mid]
136
137
138 class FPGetOpMod:
139 def __init__(self, width):
140 self.in_op = FPOp(width)
141 self.out_op = Signal(width)
142 self.out_decode = Signal(reset_less=True)
143
144 def elaborate(self, platform):
145 m = Module()
146 m.d.comb += self.out_decode.eq((self.in_op.ack) & (self.in_op.stb))
147 m.submodules.get_op_in = self.in_op
148 #m.submodules.get_op_out = self.out_op
149 with m.If(self.out_decode):
150 m.d.comb += [
151 self.out_op.eq(self.in_op.v),
152 ]
153 return m
154
155
156 class FPGetOp(FPState):
157 """ gets operand
158 """
159
160 def __init__(self, in_state, out_state, in_op, width):
161 FPState.__init__(self, in_state)
162 self.out_state = out_state
163 self.mod = FPGetOpMod(width)
164 self.in_op = in_op
165 self.out_op = Signal(width)
166 self.out_decode = Signal(reset_less=True)
167
168 def setup(self, m, in_op):
169 """ links module to inputs and outputs
170 """
171 setattr(m.submodules, self.state_from, self.mod)
172 m.d.comb += self.mod.in_op.eq(in_op)
173 #m.d.comb += self.out_op.eq(self.mod.out_op)
174 m.d.comb += self.out_decode.eq(self.mod.out_decode)
175
176 def action(self, m):
177 with m.If(self.out_decode):
178 m.next = self.out_state
179 m.d.sync += [
180 self.in_op.ack.eq(0),
181 self.out_op.eq(self.mod.out_op)
182 ]
183 with m.Else():
184 m.d.sync += self.in_op.ack.eq(1)
185
186
187 class FPGet2OpMod(Trigger):
188 def __init__(self, width):
189 Trigger.__init__(self)
190 self.in_op1 = Signal(width, reset_less=True)
191 self.in_op2 = Signal(width, reset_less=True)
192 self.out_op1 = FPNumIn(None, width)
193 self.out_op2 = FPNumIn(None, width)
194
195 def elaborate(self, platform):
196 m = Trigger.elaborate(self, platform)
197 #m.submodules.get_op_in = self.in_op
198 m.submodules.get_op1_out = self.out_op1
199 m.submodules.get_op2_out = self.out_op2
200 with m.If(self.trigger):
201 m.d.comb += [
202 self.out_op1.decode(self.in_op1),
203 self.out_op2.decode(self.in_op2),
204 ]
205 return m
206
207
208 class FPGet2Op(FPState):
209 """ gets operands
210 """
211
212 def __init__(self, in_state, out_state, in_op1, in_op2, width):
213 FPState.__init__(self, in_state)
214 self.out_state = out_state
215 self.mod = FPGet2OpMod(width)
216 self.in_op1 = in_op1
217 self.in_op2 = in_op2
218 self.out_op1 = FPNumIn(None, width)
219 self.out_op2 = FPNumIn(None, width)
220 self.in_stb = Signal(reset_less=True)
221 self.out_ack = Signal(reset_less=True)
222 self.out_decode = Signal(reset_less=True)
223
224 def setup(self, m, in_op1, in_op2, in_stb, in_ack):
225 """ links module to inputs and outputs
226 """
227 m.submodules.get_ops = self.mod
228 m.d.comb += self.mod.in_op1.eq(in_op1)
229 m.d.comb += self.mod.in_op2.eq(in_op2)
230 m.d.comb += self.mod.stb.eq(in_stb)
231 m.d.comb += self.out_ack.eq(self.mod.ack)
232 m.d.comb += self.out_decode.eq(self.mod.trigger)
233 m.d.comb += in_ack.eq(self.mod.ack)
234
235 def action(self, m):
236 with m.If(self.out_decode):
237 m.next = self.out_state
238 m.d.sync += [
239 self.mod.ack.eq(0),
240 #self.out_op1.v.eq(self.mod.out_op1.v),
241 #self.out_op2.v.eq(self.mod.out_op2.v),
242 self.out_op1.eq(self.mod.out_op1),
243 self.out_op2.eq(self.mod.out_op2)
244 ]
245 with m.Else():
246 m.d.sync += self.mod.ack.eq(1)
247
248 class FPNumBase2Ops:
249
250 def __init__(self, width, id_wid, m_extra=True):
251 self.a = FPNumBase(width, m_extra)
252 self.b = FPNumBase(width, m_extra)
253 self.mid = Signal(id_wid, reset_less=True)
254
255 def eq(self, i):
256 return [self.a.eq(i.a), self.b.eq(i.b), self.mid.eq(i.mid)]
257
258
259 class FPAddSpecialCasesMod:
260 """ special cases: NaNs, infs, zeros, denormalised
261 NOTE: some of these are unique to add. see "Special Operations"
262 https://steve.hollasch.net/cgindex/coding/ieeefloat.html
263 """
264
265 def __init__(self, width, id_wid):
266 self.width = width
267 self.id_wid = id_wid
268 self.i = self.ispec()
269 self.o = self.ospec()
270 self.out_do_z = Signal(reset_less=True)
271
272 def ispec(self):
273 return FPNumBase2Ops(self.width, self.id_wid)
274
275 def ospec(self):
276 return FPPackData(self.width, self.id_wid)
277
278 def setup(self, m, in_a, in_b, out_do_z):
279 """ links module to inputs and outputs
280 """
281 m.submodules.specialcases = self
282 m.d.comb += self.i.a.eq(in_a)
283 m.d.comb += self.i.b.eq(in_b)
284 m.d.comb += out_do_z.eq(self.out_do_z)
285
286 def elaborate(self, platform):
287 m = Module()
288
289 m.submodules.sc_in_a = self.i.a
290 m.submodules.sc_in_b = self.i.b
291 m.submodules.sc_out_z = self.o.z
292
293 s_nomatch = Signal()
294 m.d.comb += s_nomatch.eq(self.i.a.s != self.i.b.s)
295
296 m_match = Signal()
297 m.d.comb += m_match.eq(self.i.a.m == self.i.b.m)
298
299 # if a is NaN or b is NaN return NaN
300 with m.If(self.i.a.is_nan | self.i.b.is_nan):
301 m.d.comb += self.out_do_z.eq(1)
302 m.d.comb += self.o.z.nan(0)
303
304 # XXX WEIRDNESS for FP16 non-canonical NaN handling
305 # under review
306
307 ## if a is zero and b is NaN return -b
308 #with m.If(a.is_zero & (a.s==0) & b.is_nan):
309 # m.d.comb += self.out_do_z.eq(1)
310 # m.d.comb += z.create(b.s, b.e, Cat(b.m[3:-2], ~b.m[0]))
311
312 ## if b is zero and a is NaN return -a
313 #with m.Elif(b.is_zero & (b.s==0) & a.is_nan):
314 # m.d.comb += self.out_do_z.eq(1)
315 # m.d.comb += z.create(a.s, a.e, Cat(a.m[3:-2], ~a.m[0]))
316
317 ## if a is -zero and b is NaN return -b
318 #with m.Elif(a.is_zero & (a.s==1) & b.is_nan):
319 # m.d.comb += self.out_do_z.eq(1)
320 # m.d.comb += z.create(a.s & b.s, b.e, Cat(b.m[3:-2], 1))
321
322 ## if b is -zero and a is NaN return -a
323 #with m.Elif(b.is_zero & (b.s==1) & a.is_nan):
324 # m.d.comb += self.out_do_z.eq(1)
325 # m.d.comb += z.create(a.s & b.s, a.e, Cat(a.m[3:-2], 1))
326
327 # if a is inf return inf (or NaN)
328 with m.Elif(self.i.a.is_inf):
329 m.d.comb += self.out_do_z.eq(1)
330 m.d.comb += self.o.z.inf(self.i.a.s)
331 # if a is inf and signs don't match return NaN
332 with m.If(self.i.b.exp_128 & s_nomatch):
333 m.d.comb += self.o.z.nan(0)
334
335 # if b is inf return inf
336 with m.Elif(self.i.b.is_inf):
337 m.d.comb += self.out_do_z.eq(1)
338 m.d.comb += self.o.z.inf(self.i.b.s)
339
340 # if a is zero and b zero return signed-a/b
341 with m.Elif(self.i.a.is_zero & self.i.b.is_zero):
342 m.d.comb += self.out_do_z.eq(1)
343 m.d.comb += self.o.z.create(self.i.a.s & self.i.b.s,
344 self.i.b.e,
345 self.i.b.m[3:-1])
346
347 # if a is zero return b
348 with m.Elif(self.i.a.is_zero):
349 m.d.comb += self.out_do_z.eq(1)
350 m.d.comb += self.o.z.create(self.i.b.s, self.i.b.e,
351 self.i.b.m[3:-1])
352
353 # if b is zero return a
354 with m.Elif(self.i.b.is_zero):
355 m.d.comb += self.out_do_z.eq(1)
356 m.d.comb += self.o.z.create(self.i.a.s, self.i.a.e,
357 self.i.a.m[3:-1])
358
359 # if a equal to -b return zero (+ve zero)
360 with m.Elif(s_nomatch & m_match & (self.i.a.e == self.i.b.e)):
361 m.d.comb += self.out_do_z.eq(1)
362 m.d.comb += self.o.z.zero(0)
363
364 # Denormalised Number checks
365 with m.Else():
366 m.d.comb += self.out_do_z.eq(0)
367
368 return m
369
370
371 class FPID:
372 def __init__(self, id_wid):
373 self.id_wid = id_wid
374 if self.id_wid:
375 self.in_mid = Signal(id_wid, reset_less=True)
376 self.out_mid = Signal(id_wid, reset_less=True)
377 else:
378 self.in_mid = None
379 self.out_mid = None
380
381 def idsync(self, m):
382 if self.id_wid is not None:
383 m.d.sync += self.out_mid.eq(self.in_mid)
384
385
386 class FPAddSpecialCases(FPState, FPID):
387 """ special cases: NaNs, infs, zeros, denormalised
388 NOTE: some of these are unique to add. see "Special Operations"
389 https://steve.hollasch.net/cgindex/coding/ieeefloat.html
390 """
391
392 def __init__(self, width, id_wid):
393 FPState.__init__(self, "special_cases")
394 FPID.__init__(self, id_wid)
395 self.mod = FPAddSpecialCasesMod(width)
396 self.out_z = self.mod.ospec()
397 self.out_do_z = Signal(reset_less=True)
398
399 def setup(self, m, in_a, in_b, in_mid):
400 """ links module to inputs and outputs
401 """
402 self.mod.setup(m, in_a, in_b, self.out_do_z)
403 if self.in_mid is not None:
404 m.d.comb += self.in_mid.eq(in_mid)
405
406 def action(self, m):
407 self.idsync(m)
408 with m.If(self.out_do_z):
409 m.d.sync += self.out_z.v.eq(self.mod.out_z.v) # only take the output
410 m.next = "put_z"
411 with m.Else():
412 m.next = "denormalise"
413
414
415 class FPAddSpecialCasesDeNorm(FPState, FPID):
416 """ special cases: NaNs, infs, zeros, denormalised
417 NOTE: some of these are unique to add. see "Special Operations"
418 https://steve.hollasch.net/cgindex/coding/ieeefloat.html
419 """
420
421 def __init__(self, width, id_wid):
422 FPState.__init__(self, "special_cases")
423 FPID.__init__(self, id_wid)
424 self.smod = FPAddSpecialCasesMod(width, id_wid)
425 self.out_z = self.smod.ospec()
426 self.out_do_z = Signal(reset_less=True)
427
428 self.dmod = FPAddDeNormMod(width, id_wid)
429 self.o = self.dmod.ospec()
430
431 def setup(self, m, in_a, in_b, in_mid):
432 """ links module to inputs and outputs
433 """
434 self.smod.setup(m, in_a, in_b, self.out_do_z)
435 self.dmod.setup(m, in_a, in_b)
436 if self.in_mid is not None:
437 m.d.comb += self.in_mid.eq(in_mid)
438
439 def action(self, m):
440 self.idsync(m)
441 with m.If(self.out_do_z):
442 m.d.sync += self.out_z.z.v.eq(self.smod.o.z.v) # only take output
443 m.next = "put_z"
444 with m.Else():
445 m.next = "align"
446 m.d.sync += self.o.a.eq(self.dmod.o.a)
447 m.d.sync += self.o.b.eq(self.dmod.o.b)
448
449
450 class FPAddDeNormMod(FPState):
451
452 def __init__(self, width, id_wid):
453 self.width = width
454 self.id_wid = id_wid
455 self.i = self.ispec()
456 self.o = self.ospec()
457
458 def ispec(self):
459 return FPNumBase2Ops(self.width, self.id_wid)
460
461 def ospec(self):
462 return FPNumBase2Ops(self.width, self.id_wid)
463
464 def setup(self, m, in_a, in_b):
465 """ links module to inputs and outputs
466 """
467 m.submodules.denormalise = self
468 m.d.comb += self.i.a.eq(in_a)
469 m.d.comb += self.i.b.eq(in_b)
470
471 def elaborate(self, platform):
472 m = Module()
473 m.submodules.denorm_in_a = self.i.a
474 m.submodules.denorm_in_b = self.i.b
475 m.submodules.denorm_out_a = self.o.a
476 m.submodules.denorm_out_b = self.o.b
477 # hmmm, don't like repeating identical code
478 m.d.comb += self.o.a.eq(self.i.a)
479 with m.If(self.i.a.exp_n127):
480 m.d.comb += self.o.a.e.eq(self.i.a.N126) # limit a exponent
481 with m.Else():
482 m.d.comb += self.o.a.m[-1].eq(1) # set top mantissa bit
483
484 m.d.comb += self.o.b.eq(self.i.b)
485 with m.If(self.i.b.exp_n127):
486 m.d.comb += self.o.b.e.eq(self.i.b.N126) # limit a exponent
487 with m.Else():
488 m.d.comb += self.o.b.m[-1].eq(1) # set top mantissa bit
489
490 return m
491
492
493 class FPAddDeNorm(FPState, FPID):
494
495 def __init__(self, width, id_wid):
496 FPState.__init__(self, "denormalise")
497 FPID.__init__(self, id_wid)
498 self.mod = FPAddDeNormMod(width)
499 self.out_a = FPNumBase(width)
500 self.out_b = FPNumBase(width)
501
502 def setup(self, m, in_a, in_b, in_mid):
503 """ links module to inputs and outputs
504 """
505 self.mod.setup(m, in_a, in_b)
506 if self.in_mid is not None:
507 m.d.comb += self.in_mid.eq(in_mid)
508
509 def action(self, m):
510 self.idsync(m)
511 # Denormalised Number checks
512 m.next = "align"
513 m.d.sync += self.out_a.eq(self.mod.out_a)
514 m.d.sync += self.out_b.eq(self.mod.out_b)
515
516
517 class FPAddAlignMultiMod(FPState):
518
519 def __init__(self, width):
520 self.in_a = FPNumBase(width)
521 self.in_b = FPNumBase(width)
522 self.out_a = FPNumIn(None, width)
523 self.out_b = FPNumIn(None, width)
524 self.exp_eq = Signal(reset_less=True)
525
526 def elaborate(self, platform):
527 # This one however (single-cycle) will do the shift
528 # in one go.
529
530 m = Module()
531
532 m.submodules.align_in_a = self.in_a
533 m.submodules.align_in_b = self.in_b
534 m.submodules.align_out_a = self.out_a
535 m.submodules.align_out_b = self.out_b
536
537 # NOTE: this does *not* do single-cycle multi-shifting,
538 # it *STAYS* in the align state until exponents match
539
540 # exponent of a greater than b: shift b down
541 m.d.comb += self.exp_eq.eq(0)
542 m.d.comb += self.out_a.eq(self.in_a)
543 m.d.comb += self.out_b.eq(self.in_b)
544 agtb = Signal(reset_less=True)
545 altb = Signal(reset_less=True)
546 m.d.comb += agtb.eq(self.in_a.e > self.in_b.e)
547 m.d.comb += altb.eq(self.in_a.e < self.in_b.e)
548 with m.If(agtb):
549 m.d.comb += self.out_b.shift_down(self.in_b)
550 # exponent of b greater than a: shift a down
551 with m.Elif(altb):
552 m.d.comb += self.out_a.shift_down(self.in_a)
553 # exponents equal: move to next stage.
554 with m.Else():
555 m.d.comb += self.exp_eq.eq(1)
556 return m
557
558
559 class FPAddAlignMulti(FPState, FPID):
560
561 def __init__(self, width, id_wid):
562 FPID.__init__(self, id_wid)
563 FPState.__init__(self, "align")
564 self.mod = FPAddAlignMultiMod(width)
565 self.out_a = FPNumIn(None, width)
566 self.out_b = FPNumIn(None, width)
567 self.exp_eq = Signal(reset_less=True)
568
569 def setup(self, m, in_a, in_b, in_mid):
570 """ links module to inputs and outputs
571 """
572 m.submodules.align = self.mod
573 m.d.comb += self.mod.in_a.eq(in_a)
574 m.d.comb += self.mod.in_b.eq(in_b)
575 #m.d.comb += self.out_a.eq(self.mod.out_a)
576 #m.d.comb += self.out_b.eq(self.mod.out_b)
577 m.d.comb += self.exp_eq.eq(self.mod.exp_eq)
578 if self.in_mid is not None:
579 m.d.comb += self.in_mid.eq(in_mid)
580
581 def action(self, m):
582 self.idsync(m)
583 m.d.sync += self.out_a.eq(self.mod.out_a)
584 m.d.sync += self.out_b.eq(self.mod.out_b)
585 with m.If(self.exp_eq):
586 m.next = "add_0"
587
588
589 class FPNumIn2Ops:
590
591 def __init__(self, width, id_wid):
592 self.a = FPNumIn(None, width)
593 self.b = FPNumIn(None, width)
594 self.mid = Signal(id_wid, reset_less=True)
595
596 def eq(self, i):
597 return [self.a.eq(i.a), self.b.eq(i.b), self.mid.eq(i.mid)]
598
599
600 class FPAddAlignSingleMod:
601
602 def __init__(self, width, id_wid):
603 self.width = width
604 self.id_wid = id_wid
605 self.i = self.ispec()
606 self.o = self.ospec()
607
608 def ispec(self):
609 return FPNumBase2Ops(self.width, self.id_wid)
610
611 def ospec(self):
612 return FPNumIn2Ops(self.width, self.id_wid)
613
614 def setup(self, m, in_a, in_b):
615 """ links module to inputs and outputs
616 """
617 m.submodules.align = self
618 m.d.comb += self.i.a.eq(in_a)
619 m.d.comb += self.i.b.eq(in_b)
620
621 def elaborate(self, platform):
622 """ Aligns A against B or B against A, depending on which has the
623 greater exponent. This is done in a *single* cycle using
624 variable-width bit-shift
625
626 the shifter used here is quite expensive in terms of gates.
627 Mux A or B in (and out) into temporaries, as only one of them
628 needs to be aligned against the other
629 """
630 m = Module()
631
632 m.submodules.align_in_a = self.i.a
633 m.submodules.align_in_b = self.i.b
634 m.submodules.align_out_a = self.o.a
635 m.submodules.align_out_b = self.o.b
636
637 # temporary (muxed) input and output to be shifted
638 t_inp = FPNumBase(self.width)
639 t_out = FPNumIn(None, self.width)
640 espec = (len(self.i.a.e), True)
641 msr = MultiShiftRMerge(self.i.a.m_width, espec)
642 m.submodules.align_t_in = t_inp
643 m.submodules.align_t_out = t_out
644 m.submodules.multishift_r = msr
645
646 ediff = Signal(espec, reset_less=True)
647 ediffr = Signal(espec, reset_less=True)
648 tdiff = Signal(espec, reset_less=True)
649 elz = Signal(reset_less=True)
650 egz = Signal(reset_less=True)
651
652 # connect multi-shifter to t_inp/out mantissa (and tdiff)
653 m.d.comb += msr.inp.eq(t_inp.m)
654 m.d.comb += msr.diff.eq(tdiff)
655 m.d.comb += t_out.m.eq(msr.m)
656 m.d.comb += t_out.e.eq(t_inp.e + tdiff)
657 m.d.comb += t_out.s.eq(t_inp.s)
658
659 m.d.comb += ediff.eq(self.i.a.e - self.i.b.e)
660 m.d.comb += ediffr.eq(self.i.b.e - self.i.a.e)
661 m.d.comb += elz.eq(self.i.a.e < self.i.b.e)
662 m.d.comb += egz.eq(self.i.a.e > self.i.b.e)
663
664 # default: A-exp == B-exp, A and B untouched (fall through)
665 m.d.comb += self.o.a.eq(self.i.a)
666 m.d.comb += self.o.b.eq(self.i.b)
667 # only one shifter (muxed)
668 #m.d.comb += t_out.shift_down_multi(tdiff, t_inp)
669 # exponent of a greater than b: shift b down
670 with m.If(egz):
671 m.d.comb += [t_inp.eq(self.i.b),
672 tdiff.eq(ediff),
673 self.o.b.eq(t_out),
674 self.o.b.s.eq(self.i.b.s), # whoops forgot sign
675 ]
676 # exponent of b greater than a: shift a down
677 with m.Elif(elz):
678 m.d.comb += [t_inp.eq(self.i.a),
679 tdiff.eq(ediffr),
680 self.o.a.eq(t_out),
681 self.o.a.s.eq(self.i.a.s), # whoops forgot sign
682 ]
683 return m
684
685
686 class FPAddAlignSingle(FPState, FPID):
687
688 def __init__(self, width, id_wid):
689 FPState.__init__(self, "align")
690 FPID.__init__(self, id_wid)
691 self.mod = FPAddAlignSingleMod(width, id_wid)
692 self.out_a = FPNumIn(None, width)
693 self.out_b = FPNumIn(None, width)
694
695 def setup(self, m, in_a, in_b, in_mid):
696 """ links module to inputs and outputs
697 """
698 self.mod.setup(m, in_a, in_b)
699 if self.in_mid is not None:
700 m.d.comb += self.in_mid.eq(in_mid)
701
702 def action(self, m):
703 self.idsync(m)
704 # NOTE: could be done as comb
705 m.d.sync += self.out_a.eq(self.mod.out_a)
706 m.d.sync += self.out_b.eq(self.mod.out_b)
707 m.next = "add_0"
708
709
710 class FPAddAlignSingleAdd(FPState, FPID):
711
712 def __init__(self, width, id_wid):
713 FPState.__init__(self, "align")
714 FPID.__init__(self, id_wid)
715 self.mod = FPAddAlignSingleMod(width, id_wid)
716 self.o = self.mod.ospec()
717
718 self.a0mod = FPAddStage0Mod(width, id_wid)
719 self.a0o = self.a0mod.ospec()
720
721 self.a1mod = FPAddStage1Mod(width, id_wid)
722 self.a1o = self.a1mod.ospec()
723
724 def setup(self, m, in_a, in_b, in_mid):
725 """ links module to inputs and outputs
726 """
727 self.mod.setup(m, in_a, in_b)
728 m.d.comb += self.o.eq(self.mod.o)
729
730 self.a0mod.setup(m, self.o.a, self.o.b)
731 m.d.comb += self.a0o.eq(self.a0mod.o)
732
733 self.a1mod.setup(m, self.a0o.tot, self.a0o.z)
734
735 if self.in_mid is not None:
736 m.d.comb += self.in_mid.eq(in_mid)
737
738 def action(self, m):
739 self.idsync(m)
740 m.d.sync += self.a1o.eq(self.a1mod.o)
741 m.next = "normalise_1"
742
743
744 class FPAddStage0Data:
745
746 def __init__(self, width, id_wid):
747 self.z = FPNumBase(width, False)
748 self.tot = Signal(self.z.m_width + 4, reset_less=True)
749 self.mid = Signal(id_wid, reset_less=True)
750
751 def eq(self, i):
752 return [self.z.eq(i.z), self.tot.eq(i.tot), self.mid.eq(i.mid)]
753
754
755 class FPAddStage0Mod:
756
757 def __init__(self, width, id_wid):
758 self.width = width
759 self.id_wid = id_wid
760 self.i = self.ispec()
761 self.o = self.ospec()
762
763 def ispec(self):
764 return FPNumBase2Ops(self.width, self.id_wid)
765
766 def ospec(self):
767 return FPAddStage0Data(self.width, self.id_wid)
768
769 def setup(self, m, in_a, in_b):
770 """ links module to inputs and outputs
771 """
772 m.submodules.add0 = self
773 m.d.comb += self.i.a.eq(in_a)
774 m.d.comb += self.i.b.eq(in_b)
775
776 def elaborate(self, platform):
777 m = Module()
778 m.submodules.add0_in_a = self.i.a
779 m.submodules.add0_in_b = self.i.b
780 m.submodules.add0_out_z = self.o.z
781
782 m.d.comb += self.o.z.e.eq(self.i.a.e)
783
784 # store intermediate tests (and zero-extended mantissas)
785 seq = Signal(reset_less=True)
786 mge = Signal(reset_less=True)
787 am0 = Signal(len(self.i.a.m)+1, reset_less=True)
788 bm0 = Signal(len(self.i.b.m)+1, reset_less=True)
789 m.d.comb += [seq.eq(self.i.a.s == self.i.b.s),
790 mge.eq(self.i.a.m >= self.i.b.m),
791 am0.eq(Cat(self.i.a.m, 0)),
792 bm0.eq(Cat(self.i.b.m, 0))
793 ]
794 # same-sign (both negative or both positive) add mantissas
795 with m.If(seq):
796 m.d.comb += [
797 self.o.tot.eq(am0 + bm0),
798 self.o.z.s.eq(self.i.a.s)
799 ]
800 # a mantissa greater than b, use a
801 with m.Elif(mge):
802 m.d.comb += [
803 self.o.tot.eq(am0 - bm0),
804 self.o.z.s.eq(self.i.a.s)
805 ]
806 # b mantissa greater than a, use b
807 with m.Else():
808 m.d.comb += [
809 self.o.tot.eq(bm0 - am0),
810 self.o.z.s.eq(self.i.b.s)
811 ]
812 return m
813
814
815 class FPAddStage0(FPState, FPID):
816 """ First stage of add. covers same-sign (add) and subtract
817 special-casing when mantissas are greater or equal, to
818 give greatest accuracy.
819 """
820
821 def __init__(self, width, id_wid):
822 FPState.__init__(self, "add_0")
823 FPID.__init__(self, id_wid)
824 self.mod = FPAddStage0Mod(width)
825 self.o = self.mod.ospec()
826
827 def setup(self, m, in_a, in_b, in_mid):
828 """ links module to inputs and outputs
829 """
830 self.mod.setup(m, in_a, in_b)
831 if self.in_mid is not None:
832 m.d.comb += self.in_mid.eq(in_mid)
833
834 def action(self, m):
835 self.idsync(m)
836 # NOTE: these could be done as combinatorial (merge add0+add1)
837 m.d.sync += self.o.eq(self.mod.o)
838 m.next = "add_1"
839
840
841 class FPAddStage1Data:
842
843 def __init__(self, width, id_wid):
844 self.z = FPNumBase(width, False)
845 self.of = Overflow()
846 self.mid = Signal(id_wid, reset_less=True)
847
848 def eq(self, i):
849 return [self.z.eq(i.z), self.of.eq(i.of), self.mid.eq(i.mid)]
850
851
852
853 class FPAddStage1Mod(FPState):
854 """ Second stage of add: preparation for normalisation.
855 detects when tot sum is too big (tot[27] is kinda a carry bit)
856 """
857
858 def __init__(self, width, id_wid):
859 self.width = width
860 self.id_wid = id_wid
861 self.i = self.ispec()
862 self.o = self.ospec()
863
864 def ispec(self):
865 return FPAddStage0Data(self.width, self.id_wid)
866
867 def ospec(self):
868 return FPAddStage1Data(self.width, self.id_wid)
869
870 def setup(self, m, in_tot, in_z):
871 """ links module to inputs and outputs
872 """
873 m.submodules.add1 = self
874 m.submodules.add1_out_overflow = self.o.of
875
876 m.d.comb += self.i.z.eq(in_z)
877 m.d.comb += self.i.tot.eq(in_tot)
878
879 def elaborate(self, platform):
880 m = Module()
881 #m.submodules.norm1_in_overflow = self.in_of
882 #m.submodules.norm1_out_overflow = self.out_of
883 #m.submodules.norm1_in_z = self.in_z
884 #m.submodules.norm1_out_z = self.out_z
885 m.d.comb += self.o.z.eq(self.i.z)
886 # tot[-1] (MSB) gets set when the sum overflows. shift result down
887 with m.If(self.i.tot[-1]):
888 m.d.comb += [
889 self.o.z.m.eq(self.i.tot[4:]),
890 self.o.of.m0.eq(self.i.tot[4]),
891 self.o.of.guard.eq(self.i.tot[3]),
892 self.o.of.round_bit.eq(self.i.tot[2]),
893 self.o.of.sticky.eq(self.i.tot[1] | self.i.tot[0]),
894 self.o.z.e.eq(self.i.z.e + 1)
895 ]
896 # tot[-1] (MSB) zero case
897 with m.Else():
898 m.d.comb += [
899 self.o.z.m.eq(self.i.tot[3:]),
900 self.o.of.m0.eq(self.i.tot[3]),
901 self.o.of.guard.eq(self.i.tot[2]),
902 self.o.of.round_bit.eq(self.i.tot[1]),
903 self.o.of.sticky.eq(self.i.tot[0])
904 ]
905 return m
906
907
908 class FPAddStage1(FPState, FPID):
909
910 def __init__(self, width, id_wid):
911 FPState.__init__(self, "add_1")
912 FPID.__init__(self, id_wid)
913 self.mod = FPAddStage1Mod(width)
914 self.out_z = FPNumBase(width, False)
915 self.out_of = Overflow()
916 self.norm_stb = Signal()
917
918 def setup(self, m, in_tot, in_z, in_mid):
919 """ links module to inputs and outputs
920 """
921 self.mod.setup(m, in_tot, in_z)
922
923 m.d.sync += self.norm_stb.eq(0) # sets to zero when not in add1 state
924
925 if self.in_mid is not None:
926 m.d.comb += self.in_mid.eq(in_mid)
927
928 def action(self, m):
929 self.idsync(m)
930 m.d.sync += self.out_of.eq(self.mod.out_of)
931 m.d.sync += self.out_z.eq(self.mod.out_z)
932 m.d.sync += self.norm_stb.eq(1)
933 m.next = "normalise_1"
934
935
936 class FPNormaliseModSingle:
937
938 def __init__(self, width):
939 self.width = width
940 self.in_z = self.ispec()
941 self.out_z = self.ospec()
942
943 def ispec(self):
944 return FPNumBase(self.width, False)
945
946 def ospec(self):
947 return FPNumBase(self.width, False)
948
949 def setup(self, m, in_z, out_z):
950 """ links module to inputs and outputs
951 """
952 m.submodules.normalise = self
953 m.d.comb += self.in_z.eq(in_z)
954 m.d.comb += out_z.eq(self.out_z)
955
956 def elaborate(self, platform):
957 m = Module()
958
959 mwid = self.out_z.m_width+2
960 pe = PriorityEncoder(mwid)
961 m.submodules.norm_pe = pe
962
963 m.submodules.norm1_out_z = self.out_z
964 m.submodules.norm1_in_z = self.in_z
965
966 in_z = FPNumBase(self.width, False)
967 in_of = Overflow()
968 m.submodules.norm1_insel_z = in_z
969 m.submodules.norm1_insel_overflow = in_of
970
971 espec = (len(in_z.e), True)
972 ediff_n126 = Signal(espec, reset_less=True)
973 msr = MultiShiftRMerge(mwid, espec)
974 m.submodules.multishift_r = msr
975
976 m.d.comb += in_z.eq(self.in_z)
977 m.d.comb += in_of.eq(self.in_of)
978 # initialise out from in (overridden below)
979 m.d.comb += self.out_z.eq(in_z)
980 m.d.comb += self.out_of.eq(in_of)
981 # normalisation decrease condition
982 decrease = Signal(reset_less=True)
983 m.d.comb += decrease.eq(in_z.m_msbzero)
984 # decrease exponent
985 with m.If(decrease):
986 # *sigh* not entirely obvious: count leading zeros (clz)
987 # with a PriorityEncoder: to find from the MSB
988 # we reverse the order of the bits.
989 temp_m = Signal(mwid, reset_less=True)
990 temp_s = Signal(mwid+1, reset_less=True)
991 clz = Signal((len(in_z.e), True), reset_less=True)
992 m.d.comb += [
993 # cat round and guard bits back into the mantissa
994 temp_m.eq(Cat(in_of.round_bit, in_of.guard, in_z.m)),
995 pe.i.eq(temp_m[::-1]), # inverted
996 clz.eq(pe.o), # count zeros from MSB down
997 temp_s.eq(temp_m << clz), # shift mantissa UP
998 self.out_z.e.eq(in_z.e - clz), # DECREASE exponent
999 self.out_z.m.eq(temp_s[2:]), # exclude bits 0&1
1000 ]
1001
1002 return m
1003
1004 class FPNorm1Data:
1005
1006 def __init__(self, width, id_wid):
1007 self.roundz = Signal(reset_less=True)
1008 self.z = FPNumBase(width, False)
1009 self.mid = Signal(id_wid, reset_less=True)
1010
1011 def eq(self, i):
1012 return [self.z.eq(i.z), self.roundz.eq(i.roundz), self.mid.eq(i.mid)]
1013
1014
1015 class FPNorm1ModSingle:
1016
1017 def __init__(self, width, id_wid):
1018 self.width = width
1019 self.id_wid = id_wid
1020 self.i = self.ispec()
1021 self.o = self.ospec()
1022
1023 def ispec(self):
1024 return FPAddStage1Data(self.width, self.id_wid)
1025
1026 def ospec(self):
1027 return FPNorm1Data(self.width, self.id_wid)
1028
1029 def setup(self, m, in_z, in_of, out_z):
1030 """ links module to inputs and outputs
1031 """
1032 m.submodules.normalise_1 = self
1033
1034 m.d.comb += self.i.z.eq(in_z)
1035 m.d.comb += self.i.of.eq(in_of)
1036
1037 m.d.comb += out_z.eq(self.o.z)
1038
1039 def elaborate(self, platform):
1040 m = Module()
1041
1042 mwid = self.o.z.m_width+2
1043 pe = PriorityEncoder(mwid)
1044 m.submodules.norm_pe = pe
1045
1046 of = Overflow()
1047 m.d.comb += self.o.roundz.eq(of.roundz)
1048
1049 m.submodules.norm1_out_z = self.o.z
1050 m.submodules.norm1_out_overflow = of
1051 m.submodules.norm1_in_z = self.i.z
1052 m.submodules.norm1_in_overflow = self.i.of
1053
1054 i = self.ispec()
1055 m.submodules.norm1_insel_z = i.z
1056 m.submodules.norm1_insel_overflow = i.of
1057
1058 espec = (len(i.z.e), True)
1059 ediff_n126 = Signal(espec, reset_less=True)
1060 msr = MultiShiftRMerge(mwid, espec)
1061 m.submodules.multishift_r = msr
1062
1063 m.d.comb += i.eq(self.i)
1064 # initialise out from in (overridden below)
1065 m.d.comb += self.o.z.eq(i.z)
1066 m.d.comb += of.eq(i.of)
1067 # normalisation increase/decrease conditions
1068 decrease = Signal(reset_less=True)
1069 increase = Signal(reset_less=True)
1070 m.d.comb += decrease.eq(i.z.m_msbzero & i.z.exp_gt_n126)
1071 m.d.comb += increase.eq(i.z.exp_lt_n126)
1072 # decrease exponent
1073 with m.If(decrease):
1074 # *sigh* not entirely obvious: count leading zeros (clz)
1075 # with a PriorityEncoder: to find from the MSB
1076 # we reverse the order of the bits.
1077 temp_m = Signal(mwid, reset_less=True)
1078 temp_s = Signal(mwid+1, reset_less=True)
1079 clz = Signal((len(i.z.e), True), reset_less=True)
1080 # make sure that the amount to decrease by does NOT
1081 # go below the minimum non-INF/NaN exponent
1082 limclz = Mux(i.z.exp_sub_n126 > pe.o, pe.o,
1083 i.z.exp_sub_n126)
1084 m.d.comb += [
1085 # cat round and guard bits back into the mantissa
1086 temp_m.eq(Cat(i.of.round_bit, i.of.guard, i.z.m)),
1087 pe.i.eq(temp_m[::-1]), # inverted
1088 clz.eq(limclz), # count zeros from MSB down
1089 temp_s.eq(temp_m << clz), # shift mantissa UP
1090 self.o.z.e.eq(i.z.e - clz), # DECREASE exponent
1091 self.o.z.m.eq(temp_s[2:]), # exclude bits 0&1
1092 of.m0.eq(temp_s[2]), # copy of mantissa[0]
1093 # overflow in bits 0..1: got shifted too (leave sticky)
1094 of.guard.eq(temp_s[1]), # guard
1095 of.round_bit.eq(temp_s[0]), # round
1096 ]
1097 # increase exponent
1098 with m.Elif(increase):
1099 temp_m = Signal(mwid+1, reset_less=True)
1100 m.d.comb += [
1101 temp_m.eq(Cat(i.of.sticky, i.of.round_bit, i.of.guard,
1102 i.z.m)),
1103 ediff_n126.eq(i.z.N126 - i.z.e),
1104 # connect multi-shifter to inp/out mantissa (and ediff)
1105 msr.inp.eq(temp_m),
1106 msr.diff.eq(ediff_n126),
1107 self.o.z.m.eq(msr.m[3:]),
1108 of.m0.eq(temp_s[3]), # copy of mantissa[0]
1109 # overflow in bits 0..1: got shifted too (leave sticky)
1110 of.guard.eq(temp_s[2]), # guard
1111 of.round_bit.eq(temp_s[1]), # round
1112 of.sticky.eq(temp_s[0]), # sticky
1113 self.o.z.e.eq(i.z.e + ediff_n126),
1114 ]
1115
1116 return m
1117
1118
1119 class FPNorm1ModMulti:
1120
1121 def __init__(self, width, single_cycle=True):
1122 self.width = width
1123 self.in_select = Signal(reset_less=True)
1124 self.in_z = FPNumBase(width, False)
1125 self.in_of = Overflow()
1126 self.temp_z = FPNumBase(width, False)
1127 self.temp_of = Overflow()
1128 self.out_z = FPNumBase(width, False)
1129 self.out_of = Overflow()
1130
1131 def elaborate(self, platform):
1132 m = Module()
1133
1134 m.submodules.norm1_out_z = self.out_z
1135 m.submodules.norm1_out_overflow = self.out_of
1136 m.submodules.norm1_temp_z = self.temp_z
1137 m.submodules.norm1_temp_of = self.temp_of
1138 m.submodules.norm1_in_z = self.in_z
1139 m.submodules.norm1_in_overflow = self.in_of
1140
1141 in_z = FPNumBase(self.width, False)
1142 in_of = Overflow()
1143 m.submodules.norm1_insel_z = in_z
1144 m.submodules.norm1_insel_overflow = in_of
1145
1146 # select which of temp or in z/of to use
1147 with m.If(self.in_select):
1148 m.d.comb += in_z.eq(self.in_z)
1149 m.d.comb += in_of.eq(self.in_of)
1150 with m.Else():
1151 m.d.comb += in_z.eq(self.temp_z)
1152 m.d.comb += in_of.eq(self.temp_of)
1153 # initialise out from in (overridden below)
1154 m.d.comb += self.out_z.eq(in_z)
1155 m.d.comb += self.out_of.eq(in_of)
1156 # normalisation increase/decrease conditions
1157 decrease = Signal(reset_less=True)
1158 increase = Signal(reset_less=True)
1159 m.d.comb += decrease.eq(in_z.m_msbzero & in_z.exp_gt_n126)
1160 m.d.comb += increase.eq(in_z.exp_lt_n126)
1161 m.d.comb += self.out_norm.eq(decrease | increase) # loop-end
1162 # decrease exponent
1163 with m.If(decrease):
1164 m.d.comb += [
1165 self.out_z.e.eq(in_z.e - 1), # DECREASE exponent
1166 self.out_z.m.eq(in_z.m << 1), # shift mantissa UP
1167 self.out_z.m[0].eq(in_of.guard), # steal guard (was tot[2])
1168 self.out_of.guard.eq(in_of.round_bit), # round (was tot[1])
1169 self.out_of.round_bit.eq(0), # reset round bit
1170 self.out_of.m0.eq(in_of.guard),
1171 ]
1172 # increase exponent
1173 with m.Elif(increase):
1174 m.d.comb += [
1175 self.out_z.e.eq(in_z.e + 1), # INCREASE exponent
1176 self.out_z.m.eq(in_z.m >> 1), # shift mantissa DOWN
1177 self.out_of.guard.eq(in_z.m[0]),
1178 self.out_of.m0.eq(in_z.m[1]),
1179 self.out_of.round_bit.eq(in_of.guard),
1180 self.out_of.sticky.eq(in_of.sticky | in_of.round_bit)
1181 ]
1182
1183 return m
1184
1185
1186 class FPNorm1Single(FPState, FPID):
1187
1188 def __init__(self, width, id_wid, single_cycle=True):
1189 FPID.__init__(self, id_wid)
1190 FPState.__init__(self, "normalise_1")
1191 self.mod = FPNorm1ModSingle(width)
1192 self.out_z = FPNumBase(width, False)
1193 self.out_roundz = Signal(reset_less=True)
1194
1195 def setup(self, m, in_z, in_of, in_mid):
1196 """ links module to inputs and outputs
1197 """
1198 self.mod.setup(m, in_z, in_of, self.out_z)
1199
1200 if self.in_mid is not None:
1201 m.d.comb += self.in_mid.eq(in_mid)
1202
1203 def action(self, m):
1204 self.idsync(m)
1205 m.d.sync += self.out_roundz.eq(self.mod.out_of.roundz)
1206 m.next = "round"
1207
1208
1209 class FPNorm1Multi(FPState, FPID):
1210
1211 def __init__(self, width, id_wid):
1212 FPID.__init__(self, id_wid)
1213 FPState.__init__(self, "normalise_1")
1214 self.mod = FPNorm1ModMulti(width)
1215 self.stb = Signal(reset_less=True)
1216 self.ack = Signal(reset=0, reset_less=True)
1217 self.out_norm = Signal(reset_less=True)
1218 self.in_accept = Signal(reset_less=True)
1219 self.temp_z = FPNumBase(width)
1220 self.temp_of = Overflow()
1221 self.out_z = FPNumBase(width)
1222 self.out_roundz = Signal(reset_less=True)
1223
1224 def setup(self, m, in_z, in_of, norm_stb, in_mid):
1225 """ links module to inputs and outputs
1226 """
1227 self.mod.setup(m, in_z, in_of, norm_stb,
1228 self.in_accept, self.temp_z, self.temp_of,
1229 self.out_z, self.out_norm)
1230
1231 m.d.comb += self.stb.eq(norm_stb)
1232 m.d.sync += self.ack.eq(0) # sets to zero when not in normalise_1 state
1233
1234 if self.in_mid is not None:
1235 m.d.comb += self.in_mid.eq(in_mid)
1236
1237 def action(self, m):
1238 self.idsync(m)
1239 m.d.comb += self.in_accept.eq((~self.ack) & (self.stb))
1240 m.d.sync += self.temp_of.eq(self.mod.out_of)
1241 m.d.sync += self.temp_z.eq(self.out_z)
1242 with m.If(self.out_norm):
1243 with m.If(self.in_accept):
1244 m.d.sync += [
1245 self.ack.eq(1),
1246 ]
1247 with m.Else():
1248 m.d.sync += self.ack.eq(0)
1249 with m.Else():
1250 # normalisation not required (or done).
1251 m.next = "round"
1252 m.d.sync += self.ack.eq(1)
1253 m.d.sync += self.out_roundz.eq(self.mod.out_of.roundz)
1254
1255
1256 class FPNormToPack(FPState, FPID):
1257
1258 def __init__(self, width, id_wid):
1259 FPID.__init__(self, id_wid)
1260 FPState.__init__(self, "normalise_1")
1261 self.width = width
1262
1263 def setup(self, m, in_z, in_of, in_mid):
1264 """ links module to inputs and outputs
1265 """
1266
1267 # Normalisation (chained to input in_z+in_of)
1268 nmod = FPNorm1ModSingle(self.width, self.id_wid)
1269 n_out = nmod.ospec()
1270 nmod.setup(m, in_z, in_of, n_out.z)
1271 m.d.comb += n_out.roundz.eq(nmod.o.roundz)
1272
1273 # Rounding (chained to normalisation)
1274 rmod = FPRoundMod(self.width, self.id_wid)
1275 r_out_z = rmod.ospec()
1276 rmod.setup(m, n_out.z, n_out.roundz)
1277 m.d.comb += r_out_z.eq(rmod.out_z)
1278
1279 # Corrections (chained to rounding)
1280 cmod = FPCorrectionsMod(self.width, self.id_wid)
1281 c_out_z = cmod.ospec()
1282 cmod.setup(m, r_out_z)
1283 m.d.comb += c_out_z.eq(cmod.out_z)
1284
1285 # Pack (chained to corrections)
1286 self.pmod = FPPackMod(self.width, self.id_wid)
1287 self.out_z = self.pmod.ospec()
1288 self.pmod.setup(m, c_out_z)
1289
1290 # Multiplex ID
1291 if self.in_mid is not None:
1292 m.d.comb += self.in_mid.eq(in_mid)
1293
1294 def action(self, m):
1295 self.idsync(m) # copies incoming ID to outgoing
1296 m.d.sync += self.out_z.z.v.eq(self.pmod.o.z.v) # outputs packed result
1297 m.next = "pack_put_z"
1298
1299
1300 class FPRoundData:
1301
1302 def __init__(self, width, id_wid):
1303 self.z = FPNumBase(width, False)
1304 self.mid = Signal(id_wid, reset_less=True)
1305
1306 def eq(self, i):
1307 return [self.z.eq(i.z), self.mid.eq(i.mid)]
1308
1309
1310 class FPRoundMod:
1311
1312 def __init__(self, width, id_wid):
1313 self.width = width
1314 self.id_wid = id_wid
1315 self.i = self.ispec()
1316 self.out_z = self.ospec()
1317
1318 def ispec(self):
1319 return FPNorm1Data(self.width, self.id_wid)
1320
1321 def ospec(self):
1322 return FPRoundData(self.width, self.id_wid)
1323
1324 def setup(self, m, in_z, roundz):
1325 m.submodules.roundz = self
1326
1327 m.d.comb += self.i.z.eq(in_z)
1328 m.d.comb += self.i.roundz.eq(roundz)
1329
1330 def elaborate(self, platform):
1331 m = Module()
1332 m.d.comb += self.out_z.eq(self.i)
1333 with m.If(self.i.roundz):
1334 m.d.comb += self.out_z.z.m.eq(self.i.z.m + 1) # mantissa rounds up
1335 with m.If(self.i.z.m == self.i.z.m1s): # all 1s
1336 m.d.comb += self.out_z.z.e.eq(self.i.z.e + 1) # exponent up
1337 return m
1338
1339
1340 class FPRound(FPState, FPID):
1341
1342 def __init__(self, width, id_wid):
1343 FPState.__init__(self, "round")
1344 FPID.__init__(self, id_wid)
1345 self.mod = FPRoundMod(width)
1346 self.out_z = self.mod.ospec()
1347
1348 def setup(self, m, in_z, roundz, in_mid):
1349 """ links module to inputs and outputs
1350 """
1351 self.mod.setup(m, in_z, roundz)
1352
1353 if self.in_mid is not None:
1354 m.d.comb += self.in_mid.eq(in_mid)
1355
1356 def action(self, m):
1357 self.idsync(m)
1358 m.d.sync += self.out_z.eq(self.mod.out_z)
1359 m.next = "corrections"
1360
1361
1362 class FPCorrectionsMod:
1363
1364 def __init__(self, width, id_wid):
1365 self.width = width
1366 self.id_wid = id_wid
1367 self.in_z = self.ispec()
1368 self.out_z = self.ospec()
1369
1370 def ispec(self):
1371 return FPRoundData(self.width, self.id_wid)
1372
1373 def ospec(self):
1374 return FPRoundData(self.width, self.id_wid)
1375
1376 def setup(self, m, in_z):
1377 """ links module to inputs and outputs
1378 """
1379 m.submodules.corrections = self
1380 m.d.comb += self.in_z.eq(in_z)
1381
1382 def elaborate(self, platform):
1383 m = Module()
1384 m.submodules.corr_in_z = self.in_z.z
1385 m.submodules.corr_out_z = self.out_z.z
1386 m.d.comb += self.out_z.eq(self.in_z)
1387 with m.If(self.in_z.z.is_denormalised):
1388 m.d.comb += self.out_z.z.e.eq(self.in_z.z.N127)
1389 return m
1390
1391
1392 class FPCorrections(FPState, FPID):
1393
1394 def __init__(self, width, id_wid):
1395 FPState.__init__(self, "corrections")
1396 FPID.__init__(self, id_wid)
1397 self.mod = FPCorrectionsMod(width)
1398 self.out_z = self.mod.ospec()
1399
1400 def setup(self, m, in_z, in_mid):
1401 """ links module to inputs and outputs
1402 """
1403 self.mod.setup(m, in_z)
1404 if self.in_mid is not None:
1405 m.d.comb += self.in_mid.eq(in_mid)
1406
1407 def action(self, m):
1408 self.idsync(m)
1409 m.d.sync += self.out_z.eq(self.mod.out_z)
1410 m.next = "pack"
1411
1412
1413 class FPPackData:
1414
1415 def __init__(self, width, id_wid):
1416 self.z = FPNumOut(width, False)
1417 self.mid = Signal(id_wid, reset_less=True)
1418
1419 def eq(self, i):
1420 return [self.z.eq(i.z), self.mid.eq(i.mid)]
1421
1422
1423 class FPPackMod:
1424
1425 def __init__(self, width, id_wid):
1426 self.width = width
1427 self.id_wid = id_wid
1428 self.i = self.ispec()
1429 self.o = self.ospec()
1430
1431 def ispec(self):
1432 return FPRoundData(self.width, self.id_wid)
1433
1434 def ospec(self):
1435 return FPPackData(self.width, self.id_wid)
1436
1437 def setup(self, m, in_z):
1438 """ links module to inputs and outputs
1439 """
1440 m.submodules.pack = self
1441 m.d.comb += self.i.eq(in_z)
1442
1443 def elaborate(self, platform):
1444 m = Module()
1445 m.submodules.pack_in_z = self.i.z
1446 with m.If(self.i.z.is_overflowed):
1447 m.d.comb += self.o.z.inf(self.i.z.s)
1448 with m.Else():
1449 m.d.comb += self.o.z.create(self.i.z.s, self.i.z.e, self.i.z.m)
1450 return m
1451
1452
1453 class FPPackData:
1454 def __init__(self, width, id_wid):
1455 self.z = FPNumOut(width, False)
1456 self.mid = Signal(id_wid, reset_less=True)
1457
1458 def eq(self, i):
1459 return [self.z.eq(i.z), self.mid.eq(i.mid)]
1460
1461
1462 class FPPack(FPState, FPID):
1463
1464 def __init__(self, width, id_wid):
1465 FPState.__init__(self, "pack")
1466 FPID.__init__(self, id_wid)
1467 self.mod = FPPackMod(width)
1468 self.out_z = self.ospec()
1469
1470 def ispec(self):
1471 return self.mod.ispec()
1472
1473 def ospec(self):
1474 return self.mod.ospec()
1475
1476 def setup(self, m, in_z, in_mid):
1477 """ links module to inputs and outputs
1478 """
1479 self.mod.setup(m, in_z)
1480 if self.in_mid is not None:
1481 m.d.comb += self.in_mid.eq(in_mid)
1482
1483 def action(self, m):
1484 self.idsync(m)
1485 m.d.sync += self.out_z.v.eq(self.mod.out_z.v)
1486 m.next = "pack_put_z"
1487
1488
1489 class FPPutZ(FPState):
1490
1491 def __init__(self, state, in_z, out_z, in_mid, out_mid, to_state=None):
1492 FPState.__init__(self, state)
1493 if to_state is None:
1494 to_state = "get_ops"
1495 self.to_state = to_state
1496 self.in_z = in_z
1497 self.out_z = out_z
1498 self.in_mid = in_mid
1499 self.out_mid = out_mid
1500
1501 def action(self, m):
1502 if self.in_mid is not None:
1503 m.d.sync += self.out_mid.eq(self.in_mid)
1504 m.d.sync += [
1505 self.out_z.v.eq(self.in_z.v)
1506 ]
1507 with m.If(self.out_z.stb & self.out_z.ack):
1508 m.d.sync += self.out_z.stb.eq(0)
1509 m.next = self.to_state
1510 with m.Else():
1511 m.d.sync += self.out_z.stb.eq(1)
1512
1513
1514 class FPPutZIdx(FPState):
1515
1516 def __init__(self, state, in_z, out_zs, in_mid, to_state=None):
1517 FPState.__init__(self, state)
1518 if to_state is None:
1519 to_state = "get_ops"
1520 self.to_state = to_state
1521 self.in_z = in_z
1522 self.out_zs = out_zs
1523 self.in_mid = in_mid
1524
1525 def action(self, m):
1526 outz_stb = Signal(reset_less=True)
1527 outz_ack = Signal(reset_less=True)
1528 m.d.comb += [outz_stb.eq(self.out_zs[self.in_mid].stb),
1529 outz_ack.eq(self.out_zs[self.in_mid].ack),
1530 ]
1531 m.d.sync += [
1532 self.out_zs[self.in_mid].v.eq(self.in_z.v)
1533 ]
1534 with m.If(outz_stb & outz_ack):
1535 m.d.sync += self.out_zs[self.in_mid].stb.eq(0)
1536 m.next = self.to_state
1537 with m.Else():
1538 m.d.sync += self.out_zs[self.in_mid].stb.eq(1)
1539
1540
1541 class FPADDBaseMod(FPID):
1542
1543 def __init__(self, width, id_wid=None, single_cycle=False, compact=True):
1544 """ IEEE754 FP Add
1545
1546 * width: bit-width of IEEE754. supported: 16, 32, 64
1547 * id_wid: an identifier that is sync-connected to the input
1548 * single_cycle: True indicates each stage to complete in 1 clock
1549 * compact: True indicates a reduced number of stages
1550 """
1551 FPID.__init__(self, id_wid)
1552 self.width = width
1553 self.single_cycle = single_cycle
1554 self.compact = compact
1555
1556 self.in_t = Trigger()
1557 self.in_a = Signal(width)
1558 self.in_b = Signal(width)
1559 self.out_z = FPOp(width)
1560
1561 self.states = []
1562
1563 def add_state(self, state):
1564 self.states.append(state)
1565 return state
1566
1567 def get_fragment(self, platform=None):
1568 """ creates the HDL code-fragment for FPAdd
1569 """
1570 m = Module()
1571 m.submodules.out_z = self.out_z
1572 m.submodules.in_t = self.in_t
1573 if self.compact:
1574 self.get_compact_fragment(m, platform)
1575 else:
1576 self.get_longer_fragment(m, platform)
1577
1578 with m.FSM() as fsm:
1579
1580 for state in self.states:
1581 with m.State(state.state_from):
1582 state.action(m)
1583
1584 return m
1585
1586 def get_longer_fragment(self, m, platform=None):
1587
1588 get = self.add_state(FPGet2Op("get_ops", "special_cases",
1589 self.in_a, self.in_b, self.width))
1590 get.setup(m, self.in_a, self.in_b, self.in_t.stb, self.in_t.ack)
1591 a = get.out_op1
1592 b = get.out_op2
1593
1594 sc = self.add_state(FPAddSpecialCases(self.width, self.id_wid))
1595 sc.setup(m, a, b, self.in_mid)
1596
1597 dn = self.add_state(FPAddDeNorm(self.width, self.id_wid))
1598 dn.setup(m, a, b, sc.in_mid)
1599
1600 if self.single_cycle:
1601 alm = self.add_state(FPAddAlignSingle(self.width, self.id_wid))
1602 alm.setup(m, dn.out_a, dn.out_b, dn.in_mid)
1603 else:
1604 alm = self.add_state(FPAddAlignMulti(self.width, self.id_wid))
1605 alm.setup(m, dn.out_a, dn.out_b, dn.in_mid)
1606
1607 add0 = self.add_state(FPAddStage0(self.width, self.id_wid))
1608 add0.setup(m, alm.out_a, alm.out_b, alm.in_mid)
1609
1610 add1 = self.add_state(FPAddStage1(self.width, self.id_wid))
1611 add1.setup(m, add0.out_tot, add0.out_z, add0.in_mid)
1612
1613 if self.single_cycle:
1614 n1 = self.add_state(FPNorm1Single(self.width, self.id_wid))
1615 n1.setup(m, add1.out_z, add1.out_of, add0.in_mid)
1616 else:
1617 n1 = self.add_state(FPNorm1Multi(self.width, self.id_wid))
1618 n1.setup(m, add1.out_z, add1.out_of, add1.norm_stb, add0.in_mid)
1619
1620 rn = self.add_state(FPRound(self.width, self.id_wid))
1621 rn.setup(m, n1.out_z, n1.out_roundz, n1.in_mid)
1622
1623 cor = self.add_state(FPCorrections(self.width, self.id_wid))
1624 cor.setup(m, rn.out_z, rn.in_mid)
1625
1626 pa = self.add_state(FPPack(self.width, self.id_wid))
1627 pa.setup(m, cor.out_z, rn.in_mid)
1628
1629 ppz = self.add_state(FPPutZ("pack_put_z", pa.out_z, self.out_z,
1630 pa.in_mid, self.out_mid))
1631
1632 pz = self.add_state(FPPutZ("put_z", sc.out_z, self.out_z,
1633 pa.in_mid, self.out_mid))
1634
1635 def get_compact_fragment(self, m, platform=None):
1636
1637 get = self.add_state(FPGet2Op("get_ops", "special_cases",
1638 self.in_a, self.in_b, self.width))
1639 get.setup(m, self.in_a, self.in_b, self.in_t.stb, self.in_t.ack)
1640 a = get.out_op1
1641 b = get.out_op2
1642
1643 sc = self.add_state(FPAddSpecialCasesDeNorm(self.width, self.id_wid))
1644 sc.setup(m, a, b, self.in_mid)
1645
1646 alm = self.add_state(FPAddAlignSingleAdd(self.width, self.id_wid))
1647 alm.setup(m, sc.o.a, sc.o.b, sc.in_mid)
1648
1649 n1 = self.add_state(FPNormToPack(self.width, self.id_wid))
1650 n1.setup(m, alm.a1o.z, alm.a1o.of, alm.in_mid)
1651
1652 ppz = self.add_state(FPPutZ("pack_put_z", n1.out_z.z, self.out_z,
1653 n1.in_mid, self.out_mid))
1654
1655 pz = self.add_state(FPPutZ("put_z", sc.out_z.z, self.out_z,
1656 sc.in_mid, self.out_mid))
1657
1658
1659 class FPADDBase(FPState, FPID):
1660
1661 def __init__(self, width, id_wid=None, single_cycle=False):
1662 """ IEEE754 FP Add
1663
1664 * width: bit-width of IEEE754. supported: 16, 32, 64
1665 * id_wid: an identifier that is sync-connected to the input
1666 * single_cycle: True indicates each stage to complete in 1 clock
1667 """
1668 FPID.__init__(self, id_wid)
1669 FPState.__init__(self, "fpadd")
1670 self.width = width
1671 self.single_cycle = single_cycle
1672 self.mod = FPADDBaseMod(width, id_wid, single_cycle)
1673
1674 self.in_t = Trigger()
1675 self.in_a = Signal(width)
1676 self.in_b = Signal(width)
1677 #self.out_z = FPOp(width)
1678
1679 self.z_done = Signal(reset_less=True) # connects to out_z Strobe
1680 self.in_accept = Signal(reset_less=True)
1681 self.add_stb = Signal(reset_less=True)
1682 self.add_ack = Signal(reset=0, reset_less=True)
1683
1684 def setup(self, m, a, b, add_stb, in_mid, out_z, out_mid):
1685 self.out_z = out_z
1686 self.out_mid = out_mid
1687 m.d.comb += [self.in_a.eq(a),
1688 self.in_b.eq(b),
1689 self.mod.in_a.eq(self.in_a),
1690 self.mod.in_b.eq(self.in_b),
1691 self.in_mid.eq(in_mid),
1692 self.mod.in_mid.eq(self.in_mid),
1693 self.z_done.eq(self.mod.out_z.trigger),
1694 #self.add_stb.eq(add_stb),
1695 self.mod.in_t.stb.eq(self.in_t.stb),
1696 self.in_t.ack.eq(self.mod.in_t.ack),
1697 self.out_mid.eq(self.mod.out_mid),
1698 self.out_z.v.eq(self.mod.out_z.v),
1699 self.out_z.stb.eq(self.mod.out_z.stb),
1700 self.mod.out_z.ack.eq(self.out_z.ack),
1701 ]
1702
1703 m.d.sync += self.add_stb.eq(add_stb)
1704 m.d.sync += self.add_ack.eq(0) # sets to zero when not in active state
1705 m.d.sync += self.out_z.ack.eq(0) # likewise
1706 #m.d.sync += self.in_t.stb.eq(0)
1707
1708 m.submodules.fpadd = self.mod
1709
1710 def action(self, m):
1711
1712 # in_accept is set on incoming strobe HIGH and ack LOW.
1713 m.d.comb += self.in_accept.eq((~self.add_ack) & (self.add_stb))
1714
1715 #with m.If(self.in_t.ack):
1716 # m.d.sync += self.in_t.stb.eq(0)
1717 with m.If(~self.z_done):
1718 # not done: test for accepting an incoming operand pair
1719 with m.If(self.in_accept):
1720 m.d.sync += [
1721 self.add_ack.eq(1), # acknowledge receipt...
1722 self.in_t.stb.eq(1), # initiate add
1723 ]
1724 with m.Else():
1725 m.d.sync += [self.add_ack.eq(0),
1726 self.in_t.stb.eq(0),
1727 self.out_z.ack.eq(1),
1728 ]
1729 with m.Else():
1730 # done: acknowledge, and write out id and value
1731 m.d.sync += [self.add_ack.eq(1),
1732 self.in_t.stb.eq(0)
1733 ]
1734 m.next = "put_z"
1735
1736 return
1737
1738 if self.in_mid is not None:
1739 m.d.sync += self.out_mid.eq(self.mod.out_mid)
1740
1741 m.d.sync += [
1742 self.out_z.v.eq(self.mod.out_z.v)
1743 ]
1744 # move to output state on detecting z ack
1745 with m.If(self.out_z.trigger):
1746 m.d.sync += self.out_z.stb.eq(0)
1747 m.next = "put_z"
1748 with m.Else():
1749 m.d.sync += self.out_z.stb.eq(1)
1750
1751 class ResArray:
1752 def __init__(self, width, id_wid):
1753 self.width = width
1754 self.id_wid = id_wid
1755 res = []
1756 for i in range(rs_sz):
1757 out_z = FPOp(width)
1758 out_z.name = "out_z_%d" % i
1759 res.append(out_z)
1760 self.res = Array(res)
1761 self.in_z = FPOp(width)
1762 self.in_mid = Signal(self.id_wid, reset_less=True)
1763
1764 def setup(self, m, in_z, in_mid):
1765 m.d.comb += [self.in_z.eq(in_z),
1766 self.in_mid.eq(in_mid)]
1767
1768 def get_fragment(self, platform=None):
1769 """ creates the HDL code-fragment for FPAdd
1770 """
1771 m = Module()
1772 m.submodules.res_in_z = self.in_z
1773 m.submodules += self.res
1774
1775 return m
1776
1777 def ports(self):
1778 res = []
1779 for z in self.res:
1780 res += z.ports()
1781 return res
1782
1783
1784 class FPADD(FPID):
1785 """ FPADD: stages as follows:
1786
1787 FPGetOp (a)
1788 |
1789 FPGetOp (b)
1790 |
1791 FPAddBase---> FPAddBaseMod
1792 | |
1793 PutZ GetOps->Specials->Align->Add1/2->Norm->Round/Pack->PutZ
1794
1795 FPAddBase is tricky: it is both a stage and *has* stages.
1796 Connection to FPAddBaseMod therefore requires an in stb/ack
1797 and an out stb/ack. Just as with Add1-Norm1 interaction, FPGetOp
1798 needs to be the thing that raises the incoming stb.
1799 """
1800
1801 def __init__(self, width, id_wid=None, single_cycle=False, rs_sz=2):
1802 """ IEEE754 FP Add
1803
1804 * width: bit-width of IEEE754. supported: 16, 32, 64
1805 * id_wid: an identifier that is sync-connected to the input
1806 * single_cycle: True indicates each stage to complete in 1 clock
1807 """
1808 self.width = width
1809 self.id_wid = id_wid
1810 self.single_cycle = single_cycle
1811
1812 #self.out_z = FPOp(width)
1813 self.ids = FPID(id_wid)
1814
1815 rs = []
1816 for i in range(rs_sz):
1817 in_a = FPOp(width)
1818 in_b = FPOp(width)
1819 in_a.name = "in_a_%d" % i
1820 in_b.name = "in_b_%d" % i
1821 rs.append((in_a, in_b))
1822 self.rs = Array(rs)
1823
1824 res = []
1825 for i in range(rs_sz):
1826 out_z = FPOp(width)
1827 out_z.name = "out_z_%d" % i
1828 res.append(out_z)
1829 self.res = Array(res)
1830
1831 self.states = []
1832
1833 def add_state(self, state):
1834 self.states.append(state)
1835 return state
1836
1837 def get_fragment(self, platform=None):
1838 """ creates the HDL code-fragment for FPAdd
1839 """
1840 m = Module()
1841 m.submodules += self.rs
1842
1843 in_a = self.rs[0][0]
1844 in_b = self.rs[0][1]
1845
1846 out_z = FPOp(self.width)
1847 out_mid = Signal(self.id_wid, reset_less=True)
1848 m.submodules.out_z = out_z
1849
1850 geta = self.add_state(FPGetOp("get_a", "get_b",
1851 in_a, self.width))
1852 geta.setup(m, in_a)
1853 a = geta.out_op
1854
1855 getb = self.add_state(FPGetOp("get_b", "fpadd",
1856 in_b, self.width))
1857 getb.setup(m, in_b)
1858 b = getb.out_op
1859
1860 ab = FPADDBase(self.width, self.id_wid, self.single_cycle)
1861 ab = self.add_state(ab)
1862 ab.setup(m, a, b, getb.out_decode, self.ids.in_mid,
1863 out_z, out_mid)
1864
1865 pz = self.add_state(FPPutZIdx("put_z", ab.out_z, self.res,
1866 out_mid, "get_a"))
1867
1868 with m.FSM() as fsm:
1869
1870 for state in self.states:
1871 with m.State(state.state_from):
1872 state.action(m)
1873
1874 return m
1875
1876
1877 if __name__ == "__main__":
1878 if True:
1879 alu = FPADD(width=32, id_wid=5, single_cycle=True)
1880 main(alu, ports=alu.rs[0][0].ports() + \
1881 alu.rs[0][1].ports() + \
1882 alu.res[0].ports() + \
1883 [alu.ids.in_mid, alu.ids.out_mid])
1884 else:
1885 alu = FPADDBase(width=32, id_wid=5, single_cycle=True)
1886 main(alu, ports=[alu.in_a, alu.in_b] + \
1887 alu.in_t.ports() + \
1888 alu.out_z.ports() + \
1889 [alu.in_mid, alu.out_mid])
1890
1891
1892 # works... but don't use, just do "python fname.py convert -t v"
1893 #print (verilog.convert(alu, ports=[
1894 # ports=alu.in_a.ports() + \
1895 # alu.in_b.ports() + \
1896 # alu.out_z.ports())