/
mkl_gru_seq_wx.py
610 lines (512 loc) · 21.7 KB
/
mkl_gru_seq_wx.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
import theano
from theano import tensor, gof
from theano.tensor.blas import ldflags
class GRU_Inference(gof.Op):
__props__ = ('hid', 'step', 'dim', 'return_sequences', 'max_len')
def __init__(self, hid, step=None, dim=None, return_sequences=True, max_len=None):
self.hid = hid
self.step = step
self.dim = dim
self.return_sequences = return_sequences
self.max_len=max_len
super(GRU_Inference, self).__init__()
def make_node(self, *inputs):
"""
inputs: X, Wx, Wh, hid_init, bias. bias is optional.
"""
if len(inputs) in (4, 5):
inp = list(map(tensor.as_tensor_variable, inputs))
else:
raise ValueError('GRU: number of parameter is wrong.')
if len(inputs) == 5:
self.bias = True
assert inp[-1].ndim is 1
else:
self.bias = False
assert inp[0].ndim is 3
assert inp[1].ndim is 2
assert inp[2].ndim is 2
assert inp[3].ndim is 2
if self.return_sequences:
out = [inp[0].type()]
else:
bcast = [inp[0].type.broadcastable[1], inp[0].type.broadcastable[2]]
out = [tensor.tensor(dtype=inp[0].type.dtype, broadcastable=bcast)()]
return gof.Apply(self, inp, out)
def c_headers(self):
headers = ['<mkl.h>', '<omp.h>']
return headers
def c_libraries(self):
return ldflags()
def c_support_code_struct(self, node, name):
if node.inputs[0].type.dtype == 'float32':
dtype = 'float'
elif node.inputs[0].type.dtype == 'float64':
dtype = 'double'
else:
raise TypeError('GRU: dtype %s is not supported.'
% (node.inputs[0].type.dtype))
ccode = """
%(dtype)s** A;
%(dtype)s** B;
%(dtype)s** C;
MKL_INT lda[1];
MKL_INT ldb[1];
MKL_INT ldc[1];
MKL_INT m[1];
MKL_INT n[1];
MKL_INT k[1];
CBLAS_TRANSPOSE transA[1];
CBLAS_TRANSPOSE transB[1];
%(dtype)s alpha[1];
%(dtype)s beta[1];
MKL_INT size_per_grp[1];
size_t time_step;
size_t batch_size;
size_t embed_dims;
size_t max_len;
%(dtype)s* temp;
%(dtype)s* x_hzr;
%(dtype)s* zt_ptr;
%(dtype)s* rt_ptr;
""" % locals()
return ccode
def c_init_code_struct(self, node, name, sub):
ccode = """
A = NULL;
B = NULL;
C = NULL;
lda[0] = 0;
ldb[0] = 0;
ldc[0] = 0;
m[0] = 0;
n[0] = 0;
k[0] = 0;
alpha[0] = 1.0;
beta[0] = 1.0;
transA[0] = CblasNoTrans;
transB[0] = CblasNoTrans;
size_per_grp[0] = 1;
time_step = 0;
batch_size = 0;
embed_dims = 0;
max_len = 0;
temp = NULL;
x_hzr = NULL;
zt_ptr = NULL;
rt_ptr = NULL;
""" % locals()
return ccode
def c_cleanup_code_struct(self, node, name):
ccode = """
// printf("c_clean_code_struct\\n");
if (A) {
mkl_free (A);
A = NULL;
}
if (B) {
mkl_free (B);
B = NULL;
}
if (C) {
mkl_free (C);
C = NULL;
}
if (temp) {
mkl_free(temp);
}
if (zt_ptr) {
mkl_free(zt_ptr);
}
if (rt_ptr) {
mkl_free(rt_ptr);
}
if (x_hzr) {
mkl_free(x_hzr);
}
"""
return ccode
def c_code(self, node, name, inputs, outputs, sub):
if len(inputs) is 4:
with_bias = 0
X, W_x, W_h, hid_init = inputs
elif len(inputs) is 5:
with_bias = 1
X, W_x, W_h, hid_init, b = inputs
else:
raise TypeError('GRU: wrong number of arguments, expecting for 4 or 5')
z, = outputs
hid = self.hid
if self.return_sequences:
return_sequences = 1
else:
return_sequences = 0
if self.max_len:
max_len = int(self.max_len)
else:
max_len = 0
if node.inputs[0].type.dtype == 'float32':
dtype = 's'
d = 'float'
elif node.inputs[0].type.dtype == 'float64':
dtype = 'd'
d = 'double'
else:
raise TypeError('GRU: dtype %s is not supported.'
% (node.inputs[0].type.dtype))
ccode = """
double tic = dsecnd();
time_step = PyArray_DIMS(%(X)s)[0];
embed_dims = PyArray_DIMS(%(X)s)[1];
batch_size = PyArray_DIMS(%(X)s)[2];
assert (%(hid)s == PyArray_DIMS(%(W_x)s)[0] / 3);
npy_intp dims[3] = {0, 0, 0};
%(d)s* x_ptr = NULL;
%(d)s* w_x_ptr = NULL;
%(d)s* w_h_ptr = NULL;
%(d)s* b_ptr = NULL; // If no bias input, keep b_ptr as NULL.
%(d)s* hinit_ptr = NULL;
// if max_len is given, and it's small than time_step,
// a runtime error will be threw out.
if (%(max_len)s && %(max_len)s < time_step) {
PyErr_Format(PyExc_RuntimeError,
\"GRU: length of input sequence is too long: %%d,\"
\" max length is %%d\",
time_step, %(max_len)s);
goto gru_fail;
}
if (%(max_len)s) {
max_len = %(max_len)s;
}
// vmlSetMode(vmlGetMode() & 0xFFFFFFF0 | VML_EP);
// if max_len is not input as a constraint, we need to
// recreate buffers and output objects if time_step increased
// among iterations. Otherwise, we will reuse those buffers
// and output objects.
if (max_len < time_step) {
printf(\"GRU: no max_len constraint, time_step increased: %%d -> %%d\\n\", max_len, time_step);
mkl_free(A); A = NULL;
mkl_free(B); B = NULL;
mkl_free(C); C = NULL;
mkl_free(x_hzr); x_hzr = NULL;
max_len = time_step;
}
if (NULL == A) {
A = (%(d)s**)mkl_malloc(3 * max_len * sizeof (%(d)s*), 64);
}
if (NULL == B) {
B = (%(d)s**)mkl_malloc(3 * max_len * sizeof (%(d)s*), 64);
}
if (NULL == C) {
C = (%(d)s**)mkl_malloc(3 * max_len * sizeof (%(d)s*), 64);
}
if (NULL == A || NULL == B || NULL == C) {
PyErr_SetString(PyExc_MemoryError,
\"GRU: malloc A, B, C buffers for gemm_batch failed\");
goto gru_fail;
}
PyArrayObject* x_src = NULL;
PyArrayObject* w_x_src = NULL;
PyArrayObject* w_h_src = NULL;
PyArrayObject* b_src = NULL; // If no bias input, keep it with NULL
PyArrayObject* hinit_src = NULL;
if (!PyArray_IS_C_CONTIGUOUS(%(X)s)) {
printf(\"Warning: GRU need convert X to C-Contiguous\\n\");
x_src = (PyArrayObject*)PyArray_ContiguousFromAny((PyObject*)%(X)s,
PyArray_TYPE(%(X)s),
PyArray_NDIM(%(X)s),
PyArray_NDIM(%(X)s));
if (!x_src) {
PyErr_SetString(PyExc_RuntimeError, \"GRU: fail to cast X to contiguous array\");
goto gru_fail;
}
x_ptr = (%(d)s*) PyArray_DATA(x_src);
} else {
x_ptr = (%(d)s*) PyArray_DATA(%(X)s);
}
//// step 1. Dot(X, (W_x))
if (embed_dims != PyArray_DIMS(%(W_x)s)[1]) {
PyErr_SetString(PyExc_RuntimeError, \"GRU: W_x * X size error\");
goto gru_fail;
}
if (!PyArray_IS_C_CONTIGUOUS(%(W_x)s)) {
printf(\"Warning: GRU need convert W_x to C-Contiguous\\n\");
w_x_src = (PyArrayObject*)PyArray_ContiguousFromAny((PyObject*)%(W_x)s,
PyArray_TYPE(%(W_x)s),
PyArray_NDIM(%(W_x)s),
PyArray_NDIM(%(W_x)s));
if (!w_x_src) {
PyErr_SetString(PyExc_RuntimeError, \"GRU: fail to cast W_x to contiguous array\");
goto gru_fail;
}
w_x_ptr = (%(d)s*) PyArray_DATA(w_x_src);
} else {
w_x_ptr = (%(d)s*) PyArray_DATA(%(W_x)s);
}
if (!PyArray_IS_C_CONTIGUOUS(%(W_h)s)) {
printf(\"Warning: GRU need convert W_h to C-Contiguous\\n\");
w_h_src = (PyArrayObject*)PyArray_ContiguousFromAny((PyObject*)%(W_h)s,
PyArray_TYPE(%(W_h)s),
PyArray_NDIM(%(W_h)s),
PyArray_NDIM(%(W_h)s));
if (!w_h_src) {
PyErr_SetString(PyExc_RuntimeError, \"GRU: fail to cast W_h to contiguous array\");
goto gru_fail;
}
w_h_ptr = (%(d)s*) PyArray_DATA(w_h_src);
} else {
w_h_ptr = (%(d)s*) PyArray_DATA(%(W_h)s);
}
if (!PyArray_IS_C_CONTIGUOUS(%(hid_init)s)) {
printf(\"Warning: GRU need convert hid_init to C-Contiguous\\n\");
hinit_src = (PyArrayObject*)PyArray_ContiguousFromAny((PyObject*)%(hid_init)s,
PyArray_TYPE(%(hid_init)s),
PyArray_NDIM(%(hid_init)s),
PyArray_NDIM(%(hid_init)s));
if (!hinit_src) {
PyErr_SetString(PyExc_RuntimeError, \"GRU: fail to cast hid_init to contiguous array\");
goto gru_fail;
}
hinit_ptr = (%(d)s*) PyArray_DATA(hinit_src);
} else {
hinit_ptr = (%(d)s*) PyArray_DATA(%(hid_init)s);
}
// x_hzr has shape (3 * max_len, batch_size, %(hid)s)
if (NULL == x_hzr) {
x_hzr = (%(d)s*)mkl_calloc(max_len * 3 * batch_size * %(hid)s, sizeof (%(d)s), 64);
}
if (NULL == x_hzr) {
PyErr_SetString(PyExc_MemoryError,
\"GRU: fail to allocate buffer for x_hzr\");
goto gru_fail;
}
""" % locals()
if with_bias:
ccode += """
if (!PyArray_IS_C_CONTIGUOUS(%(b)s)) {
printf(\"Warning: Need convert bias to C-Contiguous\\n\");
b_src = (PyArrayObject*)PyArray_ContiguousFromAny((PyObject*)%(b)s,
PyArray_TYPE(%(b)s),
PyArray_NDIM(%(b)s),
PyArray_NDIM(%(b)s));
if (!b_src) {
PyErr_SetString(PyExc_RuntimeError, \"GRU: fail to case bias to contiguous array\");
goto gru_fail;
}
b_ptr = (%(d)s*) PyArray_DATA(b_src);
} else {
b_ptr = (%(d)s*) PyArray_DATA(%(b)s);
}
#pragma omp parallel for
for (int i = 0; i < time_step; i++) {
for (int j = 0; j < batch_size; j++) {
for (int k = 0; k < %(hid)s; k++) }
size_t ind1 = i * batch_size * %(hid)s + j * %(hid)s + k;
size_t ind2 = (i + time_stemp) * batch_size * %(hid)s + j * %(hid)s + k;
size_t ind3 = (i + 2 * time_step) * batch_size * %(hid)s + j * %(hid)s + k;
x_hzr[ind1] = b_ptr[k];
x_hzr[ind2] = b_ptr[k + time_stemp];
x_hzr[ind3] = b_ptr[k + 2 * time_step];
}
}
}
""" % locals()
else:
ccode += """
// x_hzr is allocated by mkl_calloc and set to zero already.
// memset((char*)x_hzr, 0, max_len * 3 * batch_size * %(hid)s * sizeof (%(d)s));
""" % locals()
ccode += """
// printf(\"head: %%.8f\\n\", (dsecnd() - tic) * 1000.0);
m[0] = %(hid)s;
k[0] = embed_dims;
n[0] = batch_size;
#pragma omp parallel for
for (int i = 0; i < time_step; i++) {
A[i] = w_x_ptr; // w_xh
A[i + time_step] = w_x_ptr + embed_dims * %(hid)s; // w_xz
A[i + 2 * time_step] = w_x_ptr + 2 * embed_dims * %(hid)s; // w_xr
B[i] = x_ptr + i * k[0] * n[0];
B[i + time_step] = B[i];
B[i + 2 * time_step] = B[i];
C[i] = x_hzr + i * m[0] * n[0];
C[i + time_step] = x_hzr + (i + time_step) * m[0] * n[0];
C[i + 2 * time_step] = x_hzr + (i + 2 * time_step) * m[0] * n[0];
}
if (%(with_bias)s) {
beta[0] = 1.0;
} else {
beta[0] = 0.0;
}
size_per_grp[0] = 3 * time_step;
tic = dsecnd();
cblas_%(dtype)sgemm_batch(CblasRowMajor, transA, transB, m, n, k,
alpha, A, k, B, n, beta, C, n, 1, size_per_grp);
// printf(\"batchgemm: %%.8f\\n\", (dsecnd() - tic) * 1000.0);
//// step 2. construct output
tic = dsecnd();
if ( NULL == %(z)s) {
if (%(return_sequences)s) {
dims[0] = time_step;
dims[1] = %(hid)s;
dims[2] = batch_size;
%(z)s = (PyArrayObject*) PyArray_ZEROS(3, dims, PyArray_TYPE(%(X)s), 0);
} else {
dims[0] = %(hid)s;
dims[1] = batch_size;
%(z)s = (PyArrayObject*) PyArray_ZEROS(2, dims, PyArray_TYPE(%(X)s), 0);
}
} else {
if (%(return_sequences)s) {
if (PyArray_NDIM(%(z)s) != 3 ||
PyArray_DIMS(%(z)s)[0] != time_step ||
PyArray_DIMS(%(z)s)[1] != %(hid)s ||
PyArray_DIMS(%(z)s)[2] != batch_size) {
// printf(\"delete output and recreate it\\n\");
Py_DECREF(%(z)s);
dims[0] = time_step;
dims[1] = %(hid)s;
dims[2] = batch_size;
%(z)s = (PyArrayObject*) PyArray_ZEROS(3, dims, PyArray_TYPE(%(X)s), 0);
}
} else {
if (PyArray_NDIM(%(z)s) != 2 ||
PyArray_DIMS(%(z)s)[0] != %(hid)s ||
PyArray_DIMS(%(z)s)[1] != batch_size) {
Py_DECREF(%(z)s);
dims[0] = %(hid)s;
dims[1] = batch_size;
%(z)s = (PyArrayObject*) PyArray_ZEROS(2, dims, PyArray_TYPE(%(X)s), 0);
}
}
}
if (NULL == %(z)s) {
PyErr_SetString(PyExc_RuntimeError, \"GRU: create output array failed\");
goto gru_fail;
}
//// step 3: step on time_step
// loop on step
if (NULL == temp) {
temp = (%(d)s*)mkl_malloc(batch_size * %(hid)s * sizeof (%(d)s), 64);
}
if (NULL == temp) {
PyErr_SetString(PyExc_MemoryError,
\"GRU: fail to allocate buffer for temp\");
goto gru_fail;
}
memset((void*)temp, 0, batch_size * %(hid)s * sizeof (%(d)s));
if (NULL == zt_ptr) {
zt_ptr = (%(d)s*)mkl_malloc(batch_size * %(hid)s * sizeof (%(d)s), 64);
}
if (NULL == zt_ptr) {
PyErr_SetString(PyExc_MemoryError,
\"GRU: fail to allocate buffer for zt_ptr\");
goto gru_fail;
}
if (NULL == rt_ptr) {
rt_ptr = (%(d)s*)mkl_malloc(batch_size * %(hid)s * sizeof (%(d)s), 64);
}
if (NULL == rt_ptr) {
PyErr_SetString(PyExc_MemoryError,
\"GRU: fail to allocate buffer for rt_ptr\");
goto gru_fail;
}
// printf(\"alloc: %%.8f\\n\", (dsecnd() - tic) * 1000.0);
m[0] = %(hid)s;
k[0] = %(hid)s;
n[0] = batch_size;
beta[0] = 1.0;
size_per_grp[0] = 2;
%(d)s* wh_z = (%(d)s*)PyArray_DATA(%(W_h)s) + %(hid)s * %(hid)s;
%(d)s* wh_r = (%(d)s*)PyArray_DATA(%(W_h)s) + 2 * %(hid)s * %(hid)s;
%(d)s* wh_h = (%(d)s*)PyArray_DATA(%(W_h)s);
A[0] = wh_z;
A[1] = wh_r;
A[2] = wh_h;
B[0] = hinit_ptr;
B[1] = hinit_ptr;
B[2] = hinit_ptr;
double acc1 = 0.0f;
double acc2 = 0.0f;
double acc3 = 0.0f;
double acc4 = 0.0f;
double acc5 = 0.0f;
for (int i = 0; i < time_step; i++) {
//z_t, r_t
C[0] = x_hzr + (i + time_step) * m[0] * n[0];
C[1] = x_hzr + (i + 2 * time_step) * m[0] * n[0];
C[2] = temp;
// do below two function with batch-gemm first, then sigmoid respectively
// z_t = K.sigmoid(x_z + K.dot(h_tm1, self.W_hz) + self.b_z)
// r_t = K.sigmoid(x_r + K.dot(h_tm1, self.W_hr) + self.b_r)
tic = dsecnd();
cblas_%(dtype)sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m[0], n[0], k[0],
alpha[0], A[0], k[0], B[0], n[0], beta[0], C[0], n[0]);
cblas_%(dtype)sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m[0], n[0], k[0],
alpha[0], A[1], k[0], B[1], n[0], beta[0], C[1], n[0]);
acc1 += dsecnd() - tic;
// sigmoid(C[0]), sigmoid(C[1])
// v%(dtype)sExp(m[0] * n[0], C[0], C[0]);
// v%(dtype)sExp(m[0] * n[0], C[1], C[1]);
size_t mn = m[0] * n[0];
int t = 0;
tic = dsecnd();
#pragma omp parallel for
for (t = 0; t < mn; t++) {
double exp_zt = exp((double)(C[0][t]));
double exp_rt = exp((double)(C[1][t]));
// Save rt and zt for workspace
zt_ptr[t] = (%(d)s)(exp_zt / ((double)1.0 + exp_zt));
rt_ptr[t] = (%(d)s)(exp_rt / ((double)1.0 + exp_rt));
}
acc2 += dsecnd() - tic;
// GEMM -> Mul -> Add -> tanh
// can_h_t = K.tanh(x_h + r_t * K.dot(h_tm1, self.W_hh) + self.b_h)
tic = dsecnd();
cblas_%(dtype)sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m[0], n[0], k[0],
alpha[0], A[2], k[0], B[2], n[0], beta[0], C[2], n[0]);
acc3 += dsecnd() - tic;
// v%(dtype)sMul(batch_size * %(hid)s, rt_ptr, temp, temp);
// v%(dtype)sAdd(batch_size * %(hid)s, x_hzr + i * m[0] * n[0], temp, temp);
// tanh(temp), save hcan(t)
// v%(dtype)sTanh(m[0] * n[0], temp, hcan_ptr);
tic = dsecnd();
#pragma omp parallel for
for (t = 0; t < mn; t++) {
double foo = (double)(rt_ptr[t]) * (double)(temp[t]) + (double)((x_hzr + i * mn)[t]);
// foo = (double)(temp[t]) + foo;
temp[t] = tanh(foo);
}
acc4 += dsecnd() - tic;
// h_t = (1. - z_t) * h_tm1 + z_t * can_h_t
mn = batch_size * %(hid)s;
%(d)s* z_ptr = NULL;
if (%(return_sequences)s) {
z_ptr = (%(d)s*) PyArray_DATA(%(z)s) + i * batch_size * %(hid)s;
} else {
z_ptr = (%(d)s*) PyArray_DATA(%(z)s);
}
// update
tic = dsecnd();
#pragma omp parallel for
for (int j = 0; j < mn; j++) {
z_ptr[j] = (%(d)s)( ((double)1.0 - (double)(zt_ptr[j])) * (double)(A[0][j]) + (double)(zt_ptr[j]) * (double)(temp[j]));
}
acc5 += dsecnd() - tic;
B[0] = z_ptr;
B[1] = A[0];
B[2] = A[0];
}
// printf(\"loop: %%.8f, %%.8f, %%.8f, %%.8f, %%.8f\\n\", acc1 * 1000.0, acc2 * 1000.0, acc3 * 1000.0, acc4 * 1000.0, acc5 * 1000.0);
gru_fail:
Py_XDECREF(x_src);
Py_XDECREF(w_x_src);
Py_XDECREF(w_h_src);
Py_XDECREF(b_src);
Py_XDECREF(hinit_src);
""" % locals()
return ccode
def grad(self, inp, grads):
raise NotImplemented('GRU_Inference: gradient function is not implemented.')
def c_code_cache_version(self):
return (1, 0, 1)