-
Notifications
You must be signed in to change notification settings - Fork 41
/
SIMD.py
527 lines (478 loc) · 22.8 KB
/
SIMD.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
""" Convert Expression to SIMD Compiler Intrinsics """
# Authors: Ken Sible & Zachariah Etienne
# Emails: ksible *at* outlook *dot** com
# assumpcaothiago *at* gmail *dot** com
# zachetie *at* gmail *dot** com
from sympy import (Integer, Rational, Float, Function, Symbol,
Add, Mul, Pow, Abs, S, sign, srepr, simplify,
var, sin, cos, exp, log, preorder_traversal)
from expr_tree import ExprTree
from cse_helpers import cse_preprocess
# Basic Arithmetic Operations (Debugging)
def ConstSIMD_check(a):
return Float(a, 34)
def AbsSIMD_check(a):
return Abs(a)
def nrpyAbsSIMD_check(a):
return Abs(a)
def AddSIMD_check(a, b):
return a + b
def SubSIMD_check(a, b):
return a - b
def MulSIMD_check(a, b):
return a * b
def FusedMulAddSIMD_check(a, b, c):
return a*b + c
def FusedMulSubSIMD_check(a, b, c):
return a*b - c
def NegFusedMulAddSIMD_check(a, b, c):
return -a*b + c
def NegFusedMulSubSIMD_check(a, b, c):
return -a*b - c
def DivSIMD_check(a, b):
return a / b
def signSIMD_check(a):
return sign(a)
# Transcendental Operations (Debugging)
def PowSIMD_check(a, b):
return a**b
def SqrtSIMD_check(a):
return a**(Rational(1, 2))
def CbrtSIMD_check(a):
return a**(Rational(1, 3))
def ExpSIMD_check(a):
return exp(a)
def LogSIMD_check(a):
return log(a)
def SinSIMD_check(a):
return sin(a)
def CosSIMD_check(a):
return cos(a)
def expr_convert_to_SIMD_intrins(expr, map_sym_to_rat=None, prefix="", SIMD_find_more_FMAsFMSs="True", debug="False"):
""" Convert expression to SIMD compiler intrinsics
:arg: SymPy expression
:arg: symbol to rational dictionary
:arg: option to find more FMA/FMS patterns
:arg: back-substitute and check difference
:return: expression containing SIMD compiler intrinsics
>>> from sympy.abc import a, b, c, d
>>> from cse_helpers import cse_preprocess
>>> convert = expr_convert_to_SIMD_intrins
>>> convert(a**2)
MulSIMD(a, a)
>>> convert(a**(-2))
DivSIMD(_Integer_1, MulSIMD(a, a))
>>> convert(a**(1/2))
SqrtSIMD(a)
>>> convert(a**(-1/2))
DivSIMD(_Integer_1, SqrtSIMD(a))
>>> convert(a**(-3/2))
DivSIMD(_Integer_1, MulSIMD(a, SqrtSIMD(a)))
>>> convert(a**(-5/2))
DivSIMD(_Integer_1, MulSIMD(MulSIMD(a, a), SqrtSIMD(a)))
>>> from sympy import Rational
>>> convert(a**Rational(1, 3))
CbrtSIMD(a)
>>> convert(a**b)
PowSIMD(a, b)
>>> convert(a - b)
SubSIMD(a, b)
>>> convert(a + b - c)
AddSIMD(b, SubSIMD(a, c))
>>> convert(a + b + c)
AddSIMD(a, AddSIMD(b, c))
>>> convert(a + b + c + d)
AddSIMD(AddSIMD(a, b), AddSIMD(c, d))
>>> convert(a*b*c)
MulSIMD(a, MulSIMD(b, c))
>>> convert(a*b*c*d)
MulSIMD(MulSIMD(a, b), MulSIMD(c, d))
>>> convert(a/b)
DivSIMD(a, b)
>>> convert(a*b + c)
FusedMulAddSIMD(a, b, c)
>>> convert(a*b - c)
FusedMulSubSIMD(a, b, c)
>>> convert(-a*b + c)
NegFusedMulAddSIMD(a, b, c)
>>> convert(-a*b - c)
NegFusedMulSubSIMD(a, b, c)
"""
for item in preorder_traversal(expr):
for arg in item.args:
if isinstance(arg, Symbol):
var(str(arg))
def lookup_rational(arg):
if arg.func == Symbol:
try: arg = map_sym_to_rat[arg]
except KeyError: pass
return arg
if map_sym_to_rat is None:
cse_preprocessed_expr_list, map_sym_to_rat = cse_preprocess(expr)
expr = cse_preprocessed_expr_list[0]
map_rat_to_sym = {map_sym_to_rat[v]:v for v in map_sym_to_rat}
expr_orig, tree = expr, ExprTree(expr)
AbsSIMD = Function("AbsSIMD")
AddSIMD = Function("AddSIMD")
SubSIMD = Function("SubSIMD")
MulSIMD = Function("MulSIMD")
FusedMulAddSIMD = Function("FusedMulAddSIMD")
FusedMulSubSIMD = Function("FusedMulSubSIMD")
NegFusedMulAddSIMD = Function("NegFusedMulAddSIMD")
NegFusedMulSubSIMD = Function("NegFusedMulSubSIMD")
DivSIMD = Function("DivSIMD")
SignSIMD = Function("SignSIMD")
PowSIMD = Function("PowSIMD")
SqrtSIMD = Function("SqrtSIMD")
CbrtSIMD = Function("CbrtSIMD")
ExpSIMD = Function("ExpSIMD")
LogSIMD = Function("LogSIMD")
SinSIMD = Function("SinSIMD")
CosSIMD = Function("CosSIMD")
# Step 1: Replace transcendental functions, power functions, and division expressions.
# Note: SymPy does not represent fractional integers as rationals since
# those are explicitly declared using the rational class, and hence
# the following algorithm does not affect fractional integers.
# SymPy: srepr(a**(-2)) = Pow(a, -2)
# NRPy: srepr(a**(-2)) = DivSIMD(1, MulSIMD(a, a))
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
if func == Abs:
subtree.expr = AbsSIMD(args[0])
elif func == exp:
subtree.expr = ExpSIMD(args[0])
elif func == log:
subtree.expr = LogSIMD(args[0])
elif func == sin:
subtree.expr = SinSIMD(args[0])
elif func == cos:
subtree.expr = CosSIMD(args[0])
elif func == sign:
subtree.expr = SignSIMD(args[0])
tree.reconstruct()
def IntegerPowSIMD(a, n):
# Recursive Helper Function: Construct Integer Powers
if n == 2:
return MulSIMD(a, a)
if n > 2:
return MulSIMD(IntegerPowSIMD(a, n - 1), a)
if n <= -2:
one = Symbol(prefix + '_Integer_1')
try: map_rat_to_sym[1]
except KeyError:
map_sym_to_rat[one], map_rat_to_sym[1] = S.One, one
return DivSIMD(one, IntegerPowSIMD(a, -n))
if n == -1:
one = Symbol(prefix + '_Integer_1')
try: map_rat_to_sym[1]
except KeyError:
map_sym_to_rat[one], map_rat_to_sym[1] = S.One, one
return DivSIMD(one, a)
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
if func == Pow:
one = Symbol(prefix + "_Integer_1")
exponent = lookup_rational(args[1])
if exponent == 0.5:
subtree.expr = SqrtSIMD(args[0])
subtree.children.pop(1)
elif exponent == -0.5:
subtree.expr = DivSIMD(one, SqrtSIMD(args[0]))
tree.build(subtree)
elif exponent == -1.5:
pow_1p5 = (args[0])*SqrtSIMD(args[0])
subtree.expr = DivSIMD(one, pow_1p5)
tree.build(subtree)
elif exponent == -2.5:
pow_2p5 = (args[0])*(args[0])*SqrtSIMD(args[0])
subtree.expr = DivSIMD(one, pow_2p5)
tree.build(subtree)
elif exponent == Rational(1, 3):
subtree.expr = CbrtSIMD(args[0])
subtree.children.pop(1)
elif isinstance(exponent, Integer):
subtree.expr = IntegerPowSIMD(args[0], exponent)
tree.build(subtree)
else:
subtree.expr = PowSIMD(*args)
tree.reconstruct()
# Step 2: Replace subtraction expressions.
# Note: SymPy: srepr(a - b) = Add(a, Mul(-1, b))
# NRPy: srepr(a - b) = SubSIMD(a, b)
for subtree in tree.preorder():
func = subtree.expr.func
args = list(subtree.expr.args)
if func == Add:
try:
# Find the first occurrence of a negative product inside the addition
i = next(i for i, arg in enumerate(args) if arg.func == Mul and \
any(lookup_rational(arg) == -1 for arg in args[i].args))
# Find the first occurrence of a negative symbol inside the product
j = next(j for j, arg in enumerate(args[i].args) if lookup_rational(arg) == -1)
# Find the first non-negative argument of the product
k = next(k for k in range(len(args)) if k != i)
# Remove the negative symbol from the product
subargs = list(args[i].args); subargs.pop(j)
# Build the subtraction expression for replacement
subexpr = SubSIMD(args[k], Mul(*subargs))
args = [arg for arg in args if arg not in (args[i], args[k])]
if len(args) > 0:
subexpr = Add(subexpr, *args)
subtree.expr = subexpr
tree.build(subtree)
except StopIteration: pass
tree.reconstruct()
# Step 3: Replace addition and multiplication expressions.
# Note: SIMD addition and multiplication compiler intrinsics can read
# only two arguments at once, whereas SymPy's Mul() and Add()
# operators can read an arbitrary number of arguments.
# SymPy: srepr(a*b*c*d) = Mul(a, b, c, d)
# NRPy: srepr(a*b*c*d) = MulSIMD(MulSIMD(a, b), MulSIMD(c, d))
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
if func in (Mul, Add):
func = MulSIMD if func == Mul else AddSIMD
subexpr = func(*args[-2:])
args, N = args[:-2], len(args) - 2
for i in range(0, N, 2):
if N - i > 1:
tmpexpr = func(args[i], args[i + 1])
subexpr = func(tmpexpr, subexpr, evaluate=False)
else:
subexpr = func(args[i], subexpr, evaluate=False)
subtree.expr = subexpr
tree.build(subtree)
tree.reconstruct()
# Step 4: Replace the pattern Mul(Div(1, b), a) or Mul(a, Div(1, b)) with Div(a, b).
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# MulSIMD(DivSIMD(1, b), a) >> DivSIMD(a, b)
if func == MulSIMD and args[0].func == DivSIMD and \
lookup_rational(args[0].args[0]) == 1:
subtree.expr = DivSIMD(args[1], args[0].args[1])
tree.build(subtree)
# MulSIMD(a, DivSIMD(1, b)) >> DivSIMD(a, b)
elif func == MulSIMD and args[1].func == DivSIMD and \
lookup_rational(args[1].args[0]) == 1:
subtree.expr = DivSIMD(args[0], args[1].args[1])
tree.build(subtree)
tree.reconstruct()
# Step 5: Now that all multiplication and addition functions only take two
# arguments, we can define fused-multiply-add functions,
# where AddSIMD(a, MulSIMD(b, c)) = b*c + a = FusedMulAddSIMD(b, c, a),
# or AddSIMD(MulSIMD(b, c), a) = b*c + a = FusedMulAddSIMD(b, c, a).
# Note: Fused-multiply-add (FMA3) is standard on Intel CPUs with the AVX2
# instruction set, starting with Haswell processors in 2013:
# https://en.wikipedia.org/wiki/Haswell_(microarchitecture)
# Step 5.a: Find double FMA patterns first [e.g. FMA(a, b, FMA(c, d, e))].
# Note: Double FMA simplifications do not guarantee a significant performance impact when solving BSSN equations
if SIMD_find_more_FMAsFMSs == "True":
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# a + b*c + d*e -> FMA(b,c,FMA(d,e,a))
# AddSIMD(a, AddSIMD(MulSIMD(b,c), MulSIMD(d,e))) >> FusedMulAddSIMD(b, c, FusedMulAddSIMD(d,e,a))
# Validate:
# x = a + b*c + d*e
# outputC(x,"x", params="enable_SIMD=True,SIMD_debug=True")
if (func == AddSIMD and args[1].func == AddSIMD and args[1].args[0].func == MulSIMD and args[1].args[1].func == MulSIMD):
subtree.expr = FusedMulAddSIMD( args[1].args[0].args[0], args[1].args[0].args[1],
FusedMulAddSIMD(args[1].args[1].args[0], args[1].args[1].args[1],
args[0]))
tree.build(subtree)
# b*c + d*e + a -> FMA(b,c,FMA(d,e,a))
# Validate:
# x = b*c + d*e + a
# outputC(x,"x", params="enable_SIMD=True,SIMD_debug=True")
# AddSIMD(AddSIMD(MulSIMD(b,c), MulSIMD(d,e)),a) >> FusedMulAddSIMD(b, c, FusedMulAddSIMD(d,e,a))
elif func == AddSIMD and args[0].func == AddSIMD and args[0].args[0].func == MulSIMD and args[0].args[1].func == MulSIMD:
subtree.expr = FusedMulAddSIMD( args[0].args[0].args[0], args[0].args[0].args[1],
FusedMulAddSIMD(args[0].args[1].args[0], args[0].args[1].args[1],
args[1]))
tree.build(subtree)
tree.reconstruct()
# Step 5.b: Find single FMA patterns.
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# AddSIMD(MulSIMD(b, c), a) >> FusedMulAddSIMD(b, c, a)
if func == AddSIMD and args[0].func == MulSIMD:
subtree.expr = FusedMulAddSIMD(args[0].args[0], args[0].args[1], args[1])
tree.build(subtree)
# AddSIMD(a, MulSIMD(b, c)) >> FusedMulAddSIMD(b, c, a)
elif func == AddSIMD and args[1].func == MulSIMD:
subtree.expr = FusedMulAddSIMD(args[1].args[0], args[1].args[1], args[0])
tree.build(subtree)
# SubSIMD(MulSIMD(b, c), a) >> FusedMulSubSIMD(b, c, a)
elif func == SubSIMD and args[0].func == MulSIMD:
subtree.expr = FusedMulSubSIMD(args[0].args[0], args[0].args[1], args[1])
tree.build(subtree)
# SubSIMD(a, MulSIMD(b, c)) >> NegativeFusedMulAddSIMD(b, c, a)
elif func == SubSIMD and args[1].func == MulSIMD:
subtree.expr = NegFusedMulAddSIMD(args[1].args[0], args[1].args[1], args[0])
tree.build(subtree)
# FMS(-1, MulSIMD(a, b), c) >> NegativeFusedMulSubSIMD(b, c, a)
func = subtree.expr.func
args = subtree.expr.args
if func == FusedMulSubSIMD and args[1].func == MulSIMD and lookup_rational(args[0]) == -1:
subtree.expr = NegFusedMulSubSIMD(args[1].args[0], args[1].args[1], args[2])
tree.build(subtree)
tree.reconstruct()
# Step 5.c: Remaining double FMA patterns that previously in Step 5.a were difficult to find.
# Note: Double FMA simplifications do not guarantee a significant performance impact when solving BSSN equations
if SIMD_find_more_FMAsFMSs == "True":
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# (b*c - d*e) + a -> AddSIMD(a, FusedMulSubSIMD(b, c, MulSIMD(d, e))) >> FusedMulSubSIMD(b, c, FusedMulSubSIMD(d,e,a))
# Validate:
# x = (b*c - d*e) + a
# outputC(x,"x", params="enable_SIMD=True,SIMD_debug=True")
if func == AddSIMD and args[1].func == FusedMulSubSIMD and args[1].args[2].func == MulSIMD:
subtree.expr = FusedMulSubSIMD( args[1].args[0] ,args[1].args[1],
FusedMulSubSIMD(args[1].args[2].args[0],args[1].args[2].args[1],
args[0]))
tree.build(subtree)
# b*c - (a - d*e) -> SubSIMD(FusedMulAddSIMD(b, c, MulSIMD(d, e)), a) >> FMA(b,c,FMS(d,e,a))
# Validate:
# x = b * c - (a - d * e)
# outputC(x, "x", params="enable_SIMD=True,SIMD_debug=True")
elif func == SubSIMD and args[0].func == FusedMulAddSIMD and args[0].args[2].func == MulSIMD:
subtree.expr = FusedMulAddSIMD(args[0].args[0], args[0].args[1],
FusedMulSubSIMD(args[0].args[2].args[0], args[0].args[2].args[1],
args[1]))
tree.build(subtree)
# (b*c - d*e) - a -> SubSIMD(FusedMulSubSIMD(b, c, MulSIMD(d, e)), a) >> FMS(b,c,FMA(d,e,a))
# Validate:
# x = (b*c - d*e) - a
# outputC(x,"x", params="enable_SIMD=True,SIMD_debug=True")
elif func == SubSIMD and args[0].func == FusedMulSubSIMD and args[0].args[2].func == MulSIMD:
subtree.expr = FusedMulSubSIMD(args[0].args[0], args[0].args[1],
FusedMulAddSIMD(args[0].args[2].args[0], args[0].args[2].args[1],
args[1]))
tree.build(subtree)
tree.reconstruct()
# Step 5.d: NegFusedMulAddSIMD(a,b,c) = -a*b + c:
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# FMA(a,Mul(-1,b),c) >> NFMA(a,b,c)
if func == FusedMulAddSIMD and args[1].func == MulSIMD and \
lookup_rational(args[1].args[0]) == -1:
subtree.expr = NegFusedMulAddSIMD(args[0],args[1].args[1],args[2])
tree.build(subtree)
# FMA(a,Mul(b,-1),c) >> NFMA(a,b,c)
elif func == FusedMulAddSIMD and args[1].func == MulSIMD and \
lookup_rational(args[1].args[1]) == -1:
subtree.expr = NegFusedMulAddSIMD(args[0],args[1].args[0],args[2])
tree.build(subtree)
# FMA(Mul(-1,a), b,c) >> NFMA(a,b,c)
elif func == FusedMulAddSIMD and args[0].func == MulSIMD and \
lookup_rational(args[0].args[0]) == -1:
subtree.expr = NegFusedMulAddSIMD(args[0].args[1],args[1],args[2])
tree.build(subtree)
# FMA(Mul(a,-1), b,c) >> NFMA(a,b,c)
elif func == FusedMulAddSIMD and args[0].func == MulSIMD and \
lookup_rational(args[0].args[1]) == -1:
subtree.expr = NegFusedMulAddSIMD(args[0].args[0],args[1],args[2])
tree.build(subtree)
tree.reconstruct()
# Step 5.e: Replace e.g., FMA(-1,b,c) with SubSIMD(c,b) and similar patterns
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# FMA(-1,b,c) >> SubSIMD(c,b)
if func == FusedMulAddSIMD and lookup_rational(args[0]) == -1:
subtree.expr = SubSIMD(args[2], args[1])
tree.build(subtree)
# FMA(a,-1,c) >> SubSIMD(c,a)
elif func == FusedMulAddSIMD and lookup_rational(args[1]) == -1:
subtree.expr = SubSIMD(args[2], args[0])
tree.build(subtree)
# FMS(a,-1,c) >> MulSIMD(-1,AddSIMD(a,c))
elif func == FusedMulSubSIMD and lookup_rational(args[1]) == -1:
subtree.expr = MulSIMD(args[1], AddSIMD(args[0], args[2]))
tree.build(subtree)
# FMS(-1,b,c) >> MulSIMD(-1,AddSIMD(b,c))
elif func == FusedMulSubSIMD and lookup_rational(args[0]) == -1:
subtree.expr = MulSIMD(args[0], AddSIMD(args[1], args[2]))
tree.build(subtree)
tree.reconstruct()
# Step 5.f: NegFusedMulSubSIMD(a,b,c) = -a*b - c:
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# NFMA(a,b,Mul(-1,c)) >> NFMS(a,b,c)
if func == NegFusedMulAddSIMD and args[2].func == MulSIMD and \
lookup_rational(args[2].args[0]) == -1:
subtree.expr = NegFusedMulSubSIMD(args[0],args[1],args[2].args[1])
tree.build(subtree)
# NFMA(a,b,Mul(c,-1)) >> NFMS(a,b,c)
elif func == NegFusedMulAddSIMD and args[2].func == MulSIMD and \
lookup_rational(args[2].args[1]) == -1:
subtree.expr = NegFusedMulSubSIMD(args[0],args[1],args[2].args[0])
tree.build(subtree)
# FMS(a,Mul(-1,b),c) >> NFMS(a,b,c)
elif func == FusedMulSubSIMD and args[1].func == MulSIMD and \
lookup_rational(args[1].args[0]) == -1:
subtree.expr = NegFusedMulSubSIMD(args[0],args[1].args[1],args[2])
tree.build(subtree)
# FMS(a,Mul(b,-1),c) >> NFMS(a,b,c)
elif func == FusedMulSubSIMD and args[1].func == MulSIMD and \
lookup_rational(args[1].args[1]) == -1:
subtree.expr = NegFusedMulSubSIMD(args[0],args[1].args[0],args[2])
tree.build(subtree)
# FMS(a,Mul([something],Mul(-1,b)),c) >> NFMS(a,Mul([something],b),c)
elif func == FusedMulSubSIMD and args[1].func == MulSIMD and \
args[1].args[1].func == MulSIMD and lookup_rational(args[1].args[1].args[0]) == -1:
subtree.expr = NegFusedMulSubSIMD(args[0], MulSIMD(args[1].args[0],args[1].args[1].args[1]), args[2])
tree.build(subtree)
# FMS(a,Mul([something],Mul(b,-1)),c) >> NFMS(a,Mul([something],b),c)
elif func == FusedMulSubSIMD and args[1].func == MulSIMD and \
args[1].args[1].func == MulSIMD and lookup_rational(args[1].args[1].args[1]) == -1:
subtree.expr = NegFusedMulSubSIMD(args[0], MulSIMD(args[1].args[0],args[1].args[1].args[0]), args[2])
tree.build(subtree)
tree.reconstruct()
# Step 5.g: Find single FMA patterns again, as some new ones might be found.
for subtree in tree.preorder():
func = subtree.expr.func
args = subtree.expr.args
# AddSIMD(MulSIMD(b, c), a) >> FusedMulAddSIMD(b, c, a)
if func == AddSIMD and args[0].func == MulSIMD:
subtree.expr = FusedMulAddSIMD(args[0].args[0], args[0].args[1], args[1])
tree.build(subtree)
# AddSIMD(a, MulSIMD(b, c)) >> FusedMulAddSIMD(b, c, a)
elif func == AddSIMD and args[1].func == MulSIMD:
subtree.expr = FusedMulAddSIMD(args[1].args[0], args[1].args[1], args[0])
tree.build(subtree)
# SubSIMD(MulSIMD(b, c), a) >> FusedMulSubSIMD(b, c, a)
elif func == SubSIMD and args[0].func == MulSIMD:
subtree.expr = FusedMulSubSIMD(args[0].args[0], args[0].args[1], args[1])
tree.build(subtree)
expr = tree.reconstruct()
if debug == "True":
expr_check = eval(str(expr).replace("SIMD", "SIMD_check"))
expr_check = expr_check.subs(-1, Symbol('_NegativeOne_'))
expr_diff = expr_check - expr_orig
# The eval(str(srepr())) below normalizes the expression,
# fixing a cancellation issue in SymPy ~0.7.4.
expr_diff = eval(str(srepr(expr_diff)))
tree_diff = ExprTree(expr_diff)
for subtree in tree_diff.preorder():
subexpr = subtree.expr
if subexpr.func == Float:
if abs(subexpr - Integer(subexpr)) < 1.0e-14*subexpr:
subtree.expr = Integer(subexpr)
expr_diff = tree_diff.reconstruct()
if expr_diff != 0:
simp_expr_diff = simplify(expr_diff)
if simp_expr_diff != 0:
raise Warning('Expression Difference: ' + str(simp_expr_diff))
return(expr)
if __name__ == "__main__":
import doctest
doctest.testmod()