-
Notifications
You must be signed in to change notification settings - Fork 2.5k
/
rng_mrg.py
1618 lines (1393 loc) · 58.9 KB
/
rng_mrg.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
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
"""
Implementation of MRG31k3p random number generator for Theano.
Generator code in SSJ package (L'Ecuyer & Simard).
http://www.iro.umontreal.ca/~simardr/ssj/indexe.html
"""
from __future__ import absolute_import, print_function, division
import warnings
import numpy
from six import integer_types
from six.moves import xrange
from theano import Op, Apply, shared, config, Variable
from theano import gradient, function
from theano import tensor
from theano.tensor import (TensorType, as_tensor_variable, get_vector_length,
cast, opt, scal)
from theano.tensor import sqrt, log, sin, cos, join, prod
from theano.compile import optdb
from theano.gof import local_optimizer
from . import multinomial
import theano.sandbox.cuda
from theano.sandbox.cuda import GpuOp
from theano.sandbox.cuda.basic_ops import as_cuda_ndarray_variable
from theano.gpuarray.basic_ops import GpuKernelBase, Kernel, infer_context_name, as_gpuarray_variable
from theano.gpuarray.type import GpuArrayType
from theano.gpuarray.fp16_help import write_w
from theano.gpuarray.opt import (register_opt as register_gpua,
register_opt2,
host_from_gpu as host_from_gpua)
if theano.sandbox.cuda.cuda_available:
from theano.sandbox.cuda import (CudaNdarrayType,
float32_shared_constructor)
def matVecModM(A, s, m):
# TODO : need description for method, parameter and return
assert A.dtype == 'int64'
return numpy.int32(numpy.sum((A * s) % m, 1) % m)
def multMatVect(v, A, m1, B, m2):
# TODO : need description for parameter and return
"""
Multiply the first half of v by A with a modulo of m1 and the second half
by B with a modulo of m2.
Notes
-----
The parameters of dot_modulo are passed implicitly because passing them
explicitly takes more time than running the function's C-code.
"""
if multMatVect.dot_modulo is None:
A_sym = tensor.lmatrix('A')
s_sym = tensor.ivector('s')
m_sym = tensor.iscalar('m')
A2_sym = tensor.lmatrix('A2')
s2_sym = tensor.ivector('s2')
m2_sym = tensor.iscalar('m2')
o = DotModulo()(A_sym, s_sym, m_sym, A2_sym, s2_sym, m2_sym)
multMatVect.dot_modulo = function(
[A_sym, s_sym, m_sym, A2_sym, s2_sym, m2_sym], o, profile=False)
# This way of calling the Theano fct is done to bypass Theano overhead.
f = multMatVect.dot_modulo
f.input_storage[0].storage[0] = A
f.input_storage[1].storage[0] = v[:3]
f.input_storage[2].storage[0] = m1
f.input_storage[3].storage[0] = B
f.input_storage[4].storage[0] = v[3:]
f.input_storage[5].storage[0] = m2
f.fn()
r = f.output_storage[0].storage[0]
return r
multMatVect.dot_modulo = None
class DotModulo(Op):
"""
Efficient and numerically stable implementation of a dot product followed
by a modulo operation. This performs the same function as matVecModM.
We do this 2 times on 2 triple inputs and concatenating the output.
"""
__props__ = ()
def make_node(self, A, s, m, A2, s2, m2):
return Apply(self, [A, s, m, A2, s2, m2], [s.type()])
def perform(self, node, inputs, outputs):
(A, s, m, A2, s2, m2) = inputs
(out,) = outputs
o1 = matVecModM(A, s, m)
o2 = matVecModM(A2, s2, m2)
out[0] = numpy.concatenate((o1, o2))
def c_code_cache_version(self):
return (6,)
def c_code(self, node, name, inputs, outputs, sub):
(_A, _s, _m, _A2, _s2, _m2) = inputs
(_z,) = outputs
return """
int osize = -1;
if (PyArray_NDIM(%(_A)s) != 2) {PyErr_SetString(PyExc_NotImplementedError, "rank(A) != 2"); %(fail)s;}
if (PyArray_NDIM(%(_s)s) != 1) {PyErr_SetString(PyExc_NotImplementedError, "rank(v) != 1"); %(fail)s;}
if (PyArray_NDIM(%(_m)s) != 0) {PyErr_SetString(PyExc_NotImplementedError, "rank(m) != 0"); %(fail)s;}
if (PyArray_NDIM(%(_A2)s) != 2) {PyErr_SetString(PyExc_NotImplementedError, "rank(A2) != 2"); %(fail)s;}
if (PyArray_NDIM(%(_s2)s) != 1) {PyErr_SetString(PyExc_NotImplementedError, "rank(v2) != 1"); %(fail)s;}
if (PyArray_NDIM(%(_m2)s) != 0) {PyErr_SetString(PyExc_NotImplementedError, "rank(m2) != 0"); %(fail)s;}
if( PyArray_DIMS(%(_A)s)[1] != PyArray_DIMS(%(_s)s)[0])
{PyErr_SetString(PyExc_NotImplementedError, "A and s shapes don't agree."); %(fail)s;}
if( PyArray_DIMS(%(_A2)s)[1] != PyArray_DIMS(%(_s2)s)[0])
{PyErr_SetString(PyExc_NotImplementedError, "A2 and s2 shapes don't agree."); %(fail)s;}
osize = PyArray_DIMS(%(_A)s)[0] + PyArray_DIMS(%(_A2)s)[0];
if (!%(_z)s
|| (PyArray_DIMS(%(_z)s)[0] != osize))
{
{Py_XDECREF(%(_z)s);}
npy_intp dims[] = {0,};
dims[0] = osize;
%(_z)s = (PyArrayObject*) PyArray_SimpleNew(1, dims, PyArray_TYPE(%(_s)s));
}
if(!%(_z)s){%(fail)s;}
{ //makes it compile even though labels jump over variable definitions.
// A has size MxN, s has N, output M
npy_intp M = PyArray_DIMS(%(_A)s)[0];
npy_intp N = PyArray_DIMS(%(_A)s)[1];
const dtype_%(_A)s* __restrict__ DA = (dtype_%(_A)s*)PyArray_DATA(%(_A)s);
dtype_%(_s)s* __restrict__ Ds = (dtype_%(_s)s*)PyArray_DATA(%(_s)s);
dtype_%(_z)s* __restrict__ Dz = (dtype_%(_z)s*)PyArray_DATA(%(_z)s);
const dtype_%(_m)s m = ((dtype_%(_m)s*)PyArray_DATA(%(_m)s))[0];
npy_intp SA = PyArray_STRIDES(%(_A)s)[1] / PyArray_DESCR(%(_A)s)->elsize;
npy_intp Ss = PyArray_STRIDES(%(_s)s)[0] / PyArray_DESCR(%(_s)s)->elsize;
npy_intp Sz = PyArray_STRIDES(%(_z)s)[0] / PyArray_DESCR(%(_z)s)->elsize;
for (npy_int32 i = 0; i < M; ++i)
{
const dtype_%(_A)s* __restrict__ Ak = (dtype_%(_A)s*)(PyArray_BYTES(%(_A)s) + PyArray_STRIDES(%(_A)s)[0] * i);
npy_int64 r = 0;
for (npy_int32 j = 0; j < N; ++j)
{
r += (npy_int64)(Ds[j * Ss] * (npy_int64)(Ak[j * SA])) %% m;
}
Dz[i * Sz] = r %% m;
}
}
//redo it with the second triple of inputs
{
// A has size MxN, s has N, output M
npy_intp M = PyArray_DIMS(%(_A2)s)[0];
npy_intp N = PyArray_DIMS(%(_A2)s)[1];
const dtype_%(_A2)s* __restrict__ DA = (dtype_%(_A2)s*)PyArray_DATA(%(_A2)s);
dtype_%(_s2)s* __restrict__ Ds = (dtype_%(_s2)s*)PyArray_DATA(%(_s2)s);
const dtype_%(_m2)s m = ((dtype_%(_m2)s*)PyArray_DATA(%(_m2)s))[0];
npy_intp SA = PyArray_STRIDES(%(_A2)s)[1] / PyArray_DESCR(%(_A2)s)->elsize;
npy_intp Ss = PyArray_STRIDES(%(_s2)s)[0] / PyArray_DESCR(%(_s2)s)->elsize;
npy_intp Sz = PyArray_STRIDES(%(_z)s)[0] / PyArray_DESCR(%(_z)s)->elsize;
dtype_%(_z)s* __restrict__ Dz = (dtype_%(_z)s*)PyArray_DATA(%(_z)s) + PyArray_DIMS(%(_A)s)[0] * Sz;
for (npy_int32 i = 0; i < M; ++i)
{
const dtype_%(_A2)s* __restrict__ Ak = (dtype_%(_A2)s*)(PyArray_BYTES(%(_A2)s) + PyArray_STRIDES(%(_A2)s)[0] * i);
npy_int64 r = 0;
for (npy_int32 j = 0; j < N; ++j)
{
r += (npy_int64)(Ds[j * Ss] * (npy_int64)(Ak[j * SA])) %% m;
}
Dz[i * Sz] = r %% m;
}
}
""" % dict(locals(), **sub)
# MRG31k3p
# generator constants :
M1 = numpy.asarray(numpy.int32(2147483647)) # 2^31 - 1
M2 = numpy.asarray(numpy.int32(2147462579)) # 2^31 - 21069
MASK12 = numpy.int32(511) # 2^9 - 1
MASK13 = numpy.int32(16777215) # 2^24 - 1
MASK2 = numpy.int32(65535) # 2^16 - 1
MULT2 = numpy.int32(21069)
NORM = 4.656612873077392578125e-10 # 1./2^31
# A1p0 = numpy.asarray([[0, 4194304, 129], [1, 0, 0], [0, 1, 0]],
# dtype='int64')
# A2p0 = numpy.asarray([[32768, 0, 32769], [1, 0, 0], [0, 1, 0]],
# dtype='int64')
A1p72 = numpy.asarray([[1516919229, 758510237, 499121365],
[1884998244, 1516919229, 335398200],
[601897748, 1884998244, 358115744]],
dtype='int64')
A2p72 = numpy.asarray([[1228857673, 1496414766, 954677935],
[1133297478, 1407477216, 1496414766],
[2002613992, 1639496704, 1407477216]],
dtype='int64')
A1p134 = numpy.asarray(
[[1702500920, 1849582496, 1656874625],
[828554832, 1702500920, 1512419905],
[1143731069, 828554832, 102237247]],
dtype='int64')
A2p134 = numpy.asarray(
[[796789021, 1464208080, 607337906],
[1241679051, 1431130166, 1464208080],
[1401213391, 1178684362, 1431130166]],
dtype='int64')
np_int32_vals = [numpy.int32(i) for i in (0, 7, 9, 15, 16, 22, 24)]
def ff_2p134(rstate):
# TODO : need description for method, parameter and return
return multMatVect(rstate, A1p134, M1, A2p134, M2)
def ff_2p72(rstate):
# TODO : need description for method, parameter and return
return multMatVect(rstate, A1p72, M1, A2p72, M2)
def mrg_next_value(rstate, new_rstate):
# TODO : need description for method, parameter and return
x11, x12, x13, x21, x22, x23 = rstate
assert type(x11) == numpy.int32
i0, i7, i9, i15, i16, i22, i24 = np_int32_vals
# first component
y1 = (((x12 & MASK12) << i22) + (x12 >> i9) +
((x13 & MASK13) << i7) + (x13 >> i24))
assert type(y1) == numpy.int32
if (y1 < 0 or y1 >= M1): # must also check overflow
y1 -= M1
y1 += x13
if (y1 < 0 or y1 >= M1):
y1 -= M1
x13 = x12
x12 = x11
x11 = y1
# second component
y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16))
assert type(y1) == numpy.int32
if (y1 < 0 or y1 >= M2):
y1 -= M2
y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16))
assert type(y2) == numpy.int32
if (y2 < 0 or y2 >= M2):
y2 -= M2
y2 += x23
if (y2 < 0 or y2 >= M2):
y2 -= M2
y2 += y1
if (y2 < 0 or y2 >= M2):
y2 -= M2
x23 = x22
x22 = x21
x21 = y2
# Must never return either 0 or M1+1
new_rstate[...] = [x11, x12, x13, x21, x22, x23]
assert new_rstate.dtype == numpy.int32
if (x11 <= x21):
return (x11 - x21 + M1) * NORM
else:
return (x11 - x21) * NORM
class mrg_uniform_base(Op):
# TODO : need description for class, parameter
__props__ = ("output_type", "inplace")
def __init__(self, output_type, inplace=False):
Op.__init__(self)
self.output_type = output_type
self.inplace = inplace
if inplace:
self.destroy_map = {0: [0]}
self.warned_numpy_version = False
def __str__(self):
if self.inplace:
s = "inplace"
else:
s = "no_inplace"
return self.__class__.__name__ + "{%s,%s}" % (self.output_type, s)
def grad(self, inputs, ograd):
return [gradient.grad_undefined(self, k, inp,
'No gradient defined through '
'random sampling op')
for k, inp in enumerate(inputs)]
def R_op(self, inputs, eval_points):
return [None for i in eval_points]
class mrg_uniform(mrg_uniform_base):
# CPU VERSION
def make_node(self, rstate, size):
# error checking slightly redundant here, since
# this op should not be called directly.
#
# call through MRG_RandomStreams instead.
broad = []
for i in range(self.output_type.ndim):
broad.append(tensor.extract_constant(size[i]) == 1)
output_type = self.output_type.clone(broadcastable=broad)()
rstate = as_tensor_variable(rstate)
return Apply(self,
[rstate, size],
[rstate.type(), output_type])
@classmethod
def new(cls, rstate, ndim, dtype, size):
v_size = as_tensor_variable(size)
if ndim is None:
ndim = get_vector_length(v_size)
op = cls(TensorType(dtype, (False,) * ndim))
return op(rstate, v_size)
def perform(self, node, inp, out):
rstate, size = inp
o_rstate, o_sample = out
n_elements = 1
for s in size:
n_elements *= s
if n_elements > M1:
# The limit is on the C and GPU code. This perform don't
# have this limit. But to have all of them behave the
# same (and have DebugMode don't use too much memory for
# some rng_mrg tests) I also add this limit here.
raise ValueError("rng_mrg does not support more then (2**31 -1) samples")
rstate = numpy.asarray(rstate) # bring state from GPU if necessary
if not self.inplace:
rstate = rstate.copy()
n_streams, _ = rstate.shape
rval = numpy.zeros(n_elements, dtype=self.output_type.dtype)
err_orig = numpy.seterr(over='ignore')
try:
for i in xrange(n_elements):
sample = mrg_next_value(rstate[i % n_streams],
rstate[i % n_streams])
rval[i] = sample
finally:
numpy.seterr(**err_orig)
# send to GPU if necessary
o_rstate[0] = node.outputs[0].type.filter(rstate)
o_sample[0] = node.outputs[1].type.filter(rval.reshape(size))
def c_code(self, node, name, inp, out, sub):
rstate, size = inp
# If we try to use the C code here with something else than a
# TensorType, something is wrong (likely one of the GPU ops
# not defining C code correctly).
assert isinstance(node.inputs[0].type, TensorType)
o_rstate, o_sample = out
if self.inplace:
o_rstate_requirement = (
'NPY_ARRAY_C_CONTIGUOUS|NPY_ARRAY_ALIGNED')
else:
o_rstate_requirement = (
'NPY_ARRAY_ENSURECOPY|NPY_ARRAY_C_CONTIGUOUS|'
'NPY_ARRAY_ALIGNED')
ndim = self.output_type.ndim
o_type_num = numpy.asarray(0, dtype=self.output_type.dtype).dtype.num
fail = sub['fail']
if self.output_type.dtype == 'float32':
otype = 'float'
NORM = '4.6566126e-10f' # numpy.float32(1.0/(2**31+65))
# this was determined by finding the biggest number such that
# numpy.float32(number * M1) < 1.0
else:
otype = 'double'
NORM = '4.656612873077392578125e-10'
return """
//////// <code generated by mrg_uniform>
// The +1 is to avoid odims[0] which fails on windows
// We have to read size[i] as an int64, but odims has to be intp*
// for NumPy on 32-bit platforms.
npy_intp odims[%(ndim)s+1];
npy_int64 odims_i;
npy_int64 n_elements = 1;
int n_streams = 0;
int must_alloc_sample = ((NULL == %(o_sample)s)
|| (PyArray_NDIM(%(o_sample)s) != %(ndim)s)
|| !(PyArray_ISCONTIGUOUS(%(o_sample)s)));
%(otype)s * sample_data;
npy_int32 * state_data;
const npy_int32 i0 = 0;
const npy_int32 i7 = 7;
const npy_int32 i9 = 9;
const npy_int32 i15 = 15;
const npy_int32 i16 = 16;
const npy_int32 i22 = 22;
const npy_int32 i24 = 24;
const npy_int32 M1 = 2147483647; //2^31 - 1
const npy_int32 M2 = 2147462579; //2^31 - 21069
const npy_int32 MASK12 = 511; //2^9 - 1
const npy_int32 MASK13 = 16777215; //2^24 - 1
const npy_int32 MASK2 = 65535; //2^16 - 1
const npy_int32 MULT2 = 21069;
if (PyArray_NDIM(%(size)s) != 1)
{
PyErr_SetString(PyExc_ValueError, "size must be vector");
%(fail)s
}
if (PyArray_DIMS(%(size)s)[0] != %(ndim)s)
{
PyErr_Format(PyExc_ValueError, "size must have length %%i (not %%i)",
%(ndim)s, int(PyArray_DIMS(%(size)s)[0]));
%(fail)s
}
for (int i = 0; i < %(ndim)s; ++i)
{
odims_i = *(dtype_%(size)s *)PyArray_GETPTR1(%(size)s, i);
odims[i] = odims_i;
n_elements *= odims_i;
must_alloc_sample = must_alloc_sample || (PyArray_DIMS(%(o_sample)s)[i] != odims[i]);
//fprintf(stderr, "size %%i %%i\\n", i, (int)odims[i]);
//printf("%%li", n_elements);
}
//fprintf(stderr, "n_elements %%lld\\n", (long long)n_elements);
if (n_elements > M1)
{
PyErr_SetString(
PyExc_ValueError,
"rng_mrg cpu-implementation does not support more than (2**31 -1) samples");
%(fail)s
}
if (must_alloc_sample)
{
Py_XDECREF(%(o_sample)s);
%(o_sample)s = (PyArrayObject*)PyArray_SimpleNew(%(ndim)s, odims, %(o_type_num)s);
if(!%(o_sample)s) {
PyErr_SetString(PyExc_MemoryError, "failed to alloc mrg_uniform output");
%(fail)s
}
}
Py_XDECREF(%(o_rstate)s);
%(o_rstate)s = (PyArrayObject*)PyArray_FromAny(
(PyObject*)%(rstate)s,
NULL, 0, 0, %(o_rstate_requirement)s,NULL);
if (PyArray_NDIM(%(o_rstate)s) != 2)
{
PyErr_SetString(PyExc_ValueError, "rstate must be matrix");
%(fail)s
}
if (PyArray_DIMS(%(o_rstate)s)[1] != 6)
{
PyErr_Format(PyExc_ValueError, "rstate must have 6 columns");
%(fail)s
}
if (PyArray_DESCR(%(o_rstate)s)->type_num != NPY_INT32)
{
PyErr_SetString(PyExc_ValueError, "rstate must be int32");
%(fail)s
}
n_streams = PyArray_DIMS(%(o_rstate)s)[0];
sample_data = (%(otype)s *) PyArray_DATA(%(o_sample)s);
state_data = (npy_int32 *) PyArray_DATA(%(o_rstate)s);
for (int i = 0; i < n_elements; ++i)
{
npy_int32 * state_data_i = state_data + (i%%n_streams)*6;
npy_int32 y1, y2, x11, x12, x13, x21, x22, x23;
x11 = state_data_i[0];
x12 = state_data_i[1];
x13 = state_data_i[2];
x21 = state_data_i[3];
x22 = state_data_i[4];
x23 = state_data_i[5];
y1 = ((x12 & MASK12) << i22) + (x12 >> i9) + ((x13 & MASK13) << i7) + (x13 >> i24);
if ((y1 < 0 || y1 >= M1)) //must also check overflow
y1 -= M1;
y1 += x13;
if ((y1 < 0 or y1 >= M1))
y1 -= M1;
x13 = x12;
x12 = x11;
x11 = y1;
y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16));
if (y1 < 0 || y1 >= M2)
y1 -= M2;
y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16));
if (y2 < 0 || y2 >= M2)
y2 -= M2;
y2 += x23;
if (y2 < 0 || y2 >= M2)
y2 -= M2;
y2 += y1;
if (y2 < 0 or y2 >= M2)
y2 -= M2;
x23 = x22;
x22 = x21;
x21 = y2;
if (x11 <= x21) {
assert((x11 - x21 + M1) <= M1);
sample_data[i] = (x11 - x21 + M1) * %(NORM)s;
}
else
{
assert(x11 - x21 <= M1);
sample_data[i] = (x11 - x21) * %(NORM)s;
}
state_data_i[0]= x11;
state_data_i[1]= x12;
state_data_i[2]= x13;
state_data_i[3]= x21;
state_data_i[4]= x22;
state_data_i[5]= x23;
}
//////// </ code generated by mrg_uniform>
""" % locals()
def c_code_cache_version(self):
return (8, )
class GPU_mrg_uniform(mrg_uniform_base, GpuOp):
# GPU VERSION
def make_node(self, rstate, size):
# error checking slightly redundant here, since
# this op should not be called directly.
#
# call through MRG_RandomStreams instead.
broad = []
for i in range(self.output_type.ndim):
broad.append(tensor.extract_constant(size[i]) == 1)
output_type = self.output_type.clone(broadcastable=broad)()
rstate = as_cuda_ndarray_variable(rstate)
return Apply(self,
[rstate, size],
[rstate.type(), output_type])
@classmethod
def new(cls, rstate, ndim, dtype, size):
v_size = as_tensor_variable(size)
if ndim is None:
ndim = get_vector_length(v_size)
op = cls(CudaNdarrayType((False,) * ndim))
return op(rstate, v_size)
def c_support_code_apply(self, node, nodename):
if self.output_type.dtype == 'float32':
otype = 'float'
NORM = '4.6566126e-10f' # numpy.float32(1.0/(2**31+65))
# this was determined by finding the biggest number such that
# numpy.float32(number * M1) < 1.0
else:
otype = 'double'
NORM = '4.656612873077392578125e-10'
return """
// FB: I disable the printing of the warning, as we
//receive too much email about this and this don't help
//people. I'm not even sure if the "fix" to give the info about
//the shape statically give a speed up. So I consider this
//warning as useless until proved it can speed the user code.
static int %(nodename)s_printed_warning = 1;
static __global__ void %(nodename)s_mrg_uniform(
%(otype)s*sample_data,
npy_int32*state_data,
const int Nsamples,
const int Nstreams_used)
{
const npy_int32 i0 = 0;
const npy_int32 i7 = 7;
const npy_int32 i9 = 9;
const npy_int32 i15 = 15;
const npy_int32 i16 = 16;
const npy_int32 i22 = 22;
const npy_int32 i24 = 24;
const npy_int32 M1 = 2147483647; //2^31 - 1
const npy_int32 M2 = 2147462579; //2^31 - 21069
const npy_int32 MASK12 = 511; //2^9 - 1
const npy_int32 MASK13 = 16777215; //2^24 - 1
const npy_int32 MASK2 = 65535; //2^16 - 1
const npy_int32 MULT2 = 21069;
const unsigned int numThreads = blockDim.x * gridDim.x;
const unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
npy_int32 y1, y2, x11, x12, x13, x21, x22, x23;
if (idx < Nstreams_used)
{
x11 = state_data[idx*6+0];
x12 = state_data[idx*6+1];
x13 = state_data[idx*6+2];
x21 = state_data[idx*6+3];
x22 = state_data[idx*6+4];
x23 = state_data[idx*6+5];
for (int i = idx; i < Nsamples; i += Nstreams_used)
{
y1 = ((x12 & MASK12) << i22) + (x12 >> i9) + ((x13 & MASK13) << i7) + (x13 >> i24);
y1 -= (y1 < 0 || y1 >= M1) ? M1 : 0;
y1 += x13;
y1 -= (y1 < 0 || y1 >= M1) ? M1 : 0;
x13 = x12;
x12 = x11;
x11 = y1;
y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16));
y1 -= (y1 < 0 || y1 >= M2) ? M2 : 0;
y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16));
y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
y2 += x23;
y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
y2 += y1;
y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
x23 = x22;
x22 = x21;
x21 = y2;
if (x11 <= x21) {
sample_data[i] = (x11 - x21 + M1) * %(NORM)s;
}
else
{
sample_data[i] = (x11 - x21) * %(NORM)s;
}
}
state_data[idx*6+0]= x11;
state_data[idx*6+1]= x12;
state_data[idx*6+2]= x13;
state_data[idx*6+3]= x21;
state_data[idx*6+4]= x22;
state_data[idx*6+5]= x23;
}
}
""" % locals()
def c_code(self, node, nodename, inp, out, sub):
rstate, size = inp
o_rstate, o_sample = out
inplace = int(self.inplace)
ndim = self.output_type.ndim
o_type_num = numpy.asarray(0, dtype=self.output_type.dtype).dtype.num
fail = sub['fail']
if self.output_type.dtype == 'float32':
otype = 'float'
else:
otype = 'double'
SYNC = "CNDA_THREAD_SYNC"
return """
//////// <code generated by mrg_uniform>
npy_int64 M1 = 2147483647; //2^31 - 1
// The +1 is to avoid odims[0] which fails on windows
npy_int64 odims[%(ndim)s+1];
npy_int64 n_elements = 1;
int n_streams, n_streams_used_in_this_call;
int must_alloc_sample = ((NULL == %(o_sample)s)
|| !CudaNdarray_Check((PyObject*)%(o_sample)s)
|| !CudaNdarray_is_c_contiguous(%(o_sample)s)
|| (CudaNdarray_NDIM(%(o_sample)s) != %(ndim)s));
if (PyArray_NDIM(%(size)s) != 1)
{
PyErr_SetString(PyExc_ValueError, "size must be vector");
%(fail)s
}
if (PyArray_DIMS(%(size)s)[0] != %(ndim)s)
{
PyErr_Format(PyExc_ValueError, "size must have length %%i (not %%i)",
%(ndim)s, PyArray_DIMS(%(size)s)[0]);
%(fail)s
}
for (int i = 0; i < %(ndim)s; ++i)
{
odims[i] = *(dtype_%(size)s *)PyArray_GETPTR1(%(size)s, i);
n_elements *= odims[i];
must_alloc_sample = (must_alloc_sample
|| CudaNdarray_HOST_DIMS(%(o_sample)s)[i] != odims[i]);
}
if (n_elements > M1)
{
PyErr_SetString(
PyExc_ValueError,
"rng_mrg gpu implementation does not support more than (2**31 -1) samples");
%(fail)s
}
if (must_alloc_sample)
{
Py_XDECREF(%(o_sample)s);
%(o_sample)s = (CudaNdarray*)CudaNdarray_NewDims(%(ndim)s, odims);
if(!%(o_sample)s)
{
%(fail)s;
}
}
if (!CudaNdarray_Check((PyObject*)%(rstate)s))
{
PyErr_Format(PyExc_ValueError, "rstate must be cudandarray");
%(fail)s;
}
Py_XDECREF(%(o_rstate)s);
if (%(inplace)s)
{
Py_INCREF(%(rstate)s);
%(o_rstate)s = %(rstate)s;
}
else
{
%(o_rstate)s = (CudaNdarray*)CudaNdarray_Copy(%(rstate)s);
if (!%(o_rstate)s) {
PyErr_SetString(PyExc_RuntimeError, "GPU_mrg_uniform: "
"could not copy rstate");
%(fail)s
}
}
if (CudaNdarray_NDIM(%(o_rstate)s) != 1)
{
PyErr_SetString(PyExc_ValueError, "rstate must be vector");
%(fail)s;
}
if (CudaNdarray_HOST_DIMS(%(o_rstate)s)[0] %% 6)
{
PyErr_Format(PyExc_ValueError, "rstate len must be multiple of 6");
%(fail)s;
}
n_streams = CudaNdarray_HOST_DIMS(%(o_rstate)s)[0]/6;
n_streams_used_in_this_call = std::min(n_streams, (int)n_elements);
{
unsigned int threads_per_block = std::min((unsigned int)n_streams_used_in_this_call, (unsigned int)NUM_VECTOR_OP_THREADS_PER_BLOCK);
unsigned int n_blocks = std::min(ceil_intdiv((unsigned int)n_streams_used_in_this_call, threads_per_block), (unsigned int)NUM_VECTOR_OP_BLOCKS);
if (n_streams > (unsigned int)NUM_VECTOR_OP_THREADS_PER_BLOCK * (unsigned int)NUM_VECTOR_OP_BLOCKS)
{
PyErr_Format(PyExc_ValueError, "On GPU, n_streams should be at most %%u",
(unsigned int)NUM_VECTOR_OP_THREADS_PER_BLOCK * (unsigned int)NUM_VECTOR_OP_BLOCKS);
%(fail)s;
}
if (threads_per_block * n_blocks < n_streams)
{
if (! %(nodename)s_printed_warning)
fprintf(stderr, "WARNING: unused streams above %%i (Tune GPU_mrg get_n_streams)\\n", threads_per_block * n_blocks );
%(nodename)s_printed_warning = 1;
}
%(nodename)s_mrg_uniform<<<n_blocks,threads_per_block>>>(
CudaNdarray_DEV_DATA(%(o_sample)s),
(npy_int32*)CudaNdarray_DEV_DATA(%(o_rstate)s),
n_elements, n_streams_used_in_this_call);
}
%(SYNC)s;
{
cudaError_t err = cudaGetLastError();
if( cudaSuccess != err)
{
PyErr_Format(PyExc_RuntimeError, "Cuda error: %%s: %%s.\\n", "mrg_uniform", cudaGetErrorString(err));
%(fail)s;
}
}
//////// </ code generated by mrg_uniform>
""" % locals()
def c_code_cache_version(self):
return (12,)
class GPUA_mrg_uniform(GpuKernelBase, mrg_uniform_base):
# GpuArray version
_f16_ok = True
def make_node(self, rstate, size):
# error checking slightly redundant here, since
# this op should not be called directly.
#
# call through MRG_RandomStreams instead.
broad = []
for i in range(self.output_type.ndim):
broad.append(tensor.extract_constant(size[i]) == 1)
output_type = self.output_type.clone(broadcastable=broad)()
rstate = as_gpuarray_variable(rstate, infer_context_name(rstate))
return Apply(self,
[rstate, size],
[rstate.type(), output_type])
def get_params(self, node):
return node.inputs[0].type.context
@classmethod
def new(cls, rstate, ndim, dtype, size):
v_size = as_tensor_variable(size)
if ndim is None:
ndim = get_vector_length(v_size)
op = cls(GpuArrayType(dtype, (False,) * ndim))
return op(rstate, v_size)
def c_headers(self):
return super(GPUA_mrg_uniform, self).c_headers() + ['numpy_compat.h']
def gpu_kernels(self, node, name):
write = write_w(self.output_type.dtype)
if self.output_type.dtype == 'float16':
otype = 'ga_half'
# limit the values of the state that we use.
mask = '& 0x7fff'
NORM = '3.0518e-05f' # numpy.float16(1.0/(2**15+8))
# this was determined by finding the biggest number such that
# numpy.float16(number * (M1 & 0x7fff)) < 1.0
elif self.output_type.dtype == 'float32':
otype = 'float'
mask = ''
NORM = '4.6566126e-10f' # numpy.float32(1.0/(2**31+65))
# this was determined by finding the biggest number such that
# numpy.float32(number * M1) < 1.0
elif self.output_type.dtype == 'float64':
otype = 'double'
mask = ''
NORM = '4.656612873077392578125e-10'
else:
raise ValueError('Unsupported data type for output',
self.output_type.dtype)
code = """
KERNEL void mrg_uniform(
GLOBAL_MEM %(otype)s *sample_data,
GLOBAL_MEM ga_int *state_data,
const ga_uint Nsamples,
const ga_uint Nstreams_used)
{
/*
* The cluda backend makes sure that ga_int corresponds to
* a 32 bit signed type on the target device. It is not a
* variable width type.
*/
const ga_int i7 = 7;
const ga_int i9 = 9;
const ga_int i15 = 15;
const ga_int i16 = 16;
const ga_int i22 = 22;
const ga_int i24 = 24;
const ga_int M1 = 2147483647; //2^31 - 1
const ga_int M2 = 2147462579; //2^31 - 21069
const ga_int MASK12 = 511; //2^9 - 1
const ga_int MASK13 = 16777215; //2^24 - 1
const ga_int MASK2 = 65535; //2^16 - 1
const ga_int MULT2 = 21069;
const ga_uint idx = GID_0 * LDIM_0 + LID_0;
ga_int y1, y2, x11, x12, x13, x21, x22, x23;
if (idx < Nstreams_used)
{
x11 = state_data[idx*6+0];
x12 = state_data[idx*6+1];
x13 = state_data[idx*6+2];
x21 = state_data[idx*6+3];
x22 = state_data[idx*6+4];
x23 = state_data[idx*6+5];
for (ga_uint i = idx; i < Nsamples; i += Nstreams_used)
{
y1 = ((x12 & MASK12) << i22) + (x12 >> i9) + ((x13 & MASK13) << i7) + (x13 >> i24);
y1 -= (y1 < 0 || y1 >= M1) ? M1 : 0;
y1 += x13;
y1 -= (y1 < 0 || y1 >= M1) ? M1 : 0;
x13 = x12;
x12 = x11;
x11 = y1;
y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16));
y1 -= (y1 < 0 || y1 >= M2) ? M2 : 0;
y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16));
y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
y2 += x23;
y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
y2 += y1;
y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
x23 = x22;
x22 = x21;
x21 = y2;
if (x11 <= x21) {
sample_data[i] = %(write)s(((x11 - x21 + M1) %(mask)s) * %(NORM)s);
}
else
{
sample_data[i] = %(write)s(((x11 - x21) %(mask)s) * %(NORM)s);
}
}
state_data[idx*6+0]= x11;
state_data[idx*6+1]= x12;
state_data[idx*6+2]= x13;
state_data[idx*6+3]= x21;
state_data[idx*6+4]= x22;
state_data[idx*6+5]= x23;
}
}
""" % locals()
# we shouldn't get to this line if it's about to fail
from pygpu import gpuarray
return [Kernel(code=code, name="mrg_uniform",
params=[gpuarray.GpuArray, gpuarray.GpuArray,
'uint32', 'uint32'],
flags=Kernel.get_flags(self.output_type.dtype, 'int32'))
]
def c_code(self, node, nodename, inp, out, sub):
rstate, size = inp
o_rstate, o_sample = out
inplace = int(self.inplace)
ndim = self.output_type.ndim
o_type_num = numpy.asarray(0, dtype=self.output_type.dtype).dtype.num
fail = sub['fail']
ctx = sub['params']
kname = self.gpu_kernels(node, nodename)[0].objvar
otypecode = str(self.output_type.typecode)
return """
npy_int64 M1 = 2147483647; //2^31 - 1
// The +1 is to avoid odims[0] which fails on windows
size_t odims[%(ndim)s+1];
size_t n_elements = 1;
unsigned int n_streams;
int must_alloc_sample = ((NULL == %(o_sample)s)
|| !pygpu_GpuArray_Check((PyObject*)%(o_sample)s)
|| !(%(o_sample)s->ga.flags & GA_C_CONTIGUOUS)
|| (PyGpuArray_NDIM(%(o_sample)s) != %(ndim)s));
if (PyArray_NDIM(%(size)s) != 1)
{
PyErr_SetString(PyExc_ValueError, "size must be vector");
%(fail)s
}
if (PyArray_DIMS(%(size)s)[0] != %(ndim)s)
{
PyErr_Format(PyExc_ValueError, "size must have length %%i (not %%li)",
%(ndim)s, PyArray_DIMS(%(size)s)[0]);
%(fail)s
}