diffusion_rs_common 0.1.0

Blazingly fast inference of diffusion models.
Documentation
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
#include <metal_stdlib>

using namespace metal;

#if defined(__HAVE_BFLOAT__)

typedef bfloat bfloat16_t;

#else

/////////////////////////////////////////////////////////////////////////////
// Helpers
/////////////////////////////////////////////////////////////////////////////

constexpr METAL_FUNC uint16_t float_to_bfloat_bits(float x) {
  // Check for nan
  if ((as_type<uint32_t>(x) & ~_fp_encoding_traits<float>::sign_mask) >
      _fp_encoding_traits<float>::inf_mask) {
    return uint16_t(as_type<uint32_t>(0x7FC0));
  }
  // Take bits
  uint32_t float_bits = as_type<uint32_t>(x);

  // Round to nearest even
  float_bits += ((float_bits >> 16) & 1) + as_type<uint32_t>(0x7FFF);

  // Take upper 16 bits
  return float_bits >> 16;
}

constexpr METAL_FUNC float bfloat_bits_to_float(uint16_t x) {
  // Upper 16 bits are the data and lower 16 bits are 0s
  return as_type<float>((uint32_t)x << 16);
}

struct _MLX_BFloat16;

template <typename T>
static constexpr constant bool can_convert_to_bfloat =
    !is_same_v<T, _MLX_BFloat16> && is_convertible_v<T, float>;

template <typename T>
static constexpr constant bool can_convert_from_bfloat =
    !is_same_v<T, _MLX_BFloat16> && is_convertible_v<float, T>;

/////////////////////////////////////////////////////////////////////////////
// Bfloat struct
/////////////////////////////////////////////////////////////////////////////

struct _MLX_BFloat16 {
  /////////////////////////////////////////////////////////////////////////////
  // Constructors
  uint16_t bits_;
  _MLX_BFloat16() thread = default;
  _MLX_BFloat16() threadgroup = default;
  _MLX_BFloat16() device = default;
  _MLX_BFloat16() constant = default;

  struct bits_to_bfloat_struct {};
  static constexpr METAL_FUNC bits_to_bfloat_struct bits_to_bfloat() {
    return bits_to_bfloat_struct();
  }
  constexpr METAL_FUNC _MLX_BFloat16(uint16_t bits, bits_to_bfloat_struct)
      : bits_(bits) {}

  /////////////////////////////////////////////////////////////////////////////
  // Conversions to bfloat

  template <
      typename T,
      typename = typename enable_if<can_convert_to_bfloat<T>>::type>
  constexpr METAL_FUNC _MLX_BFloat16(T x) thread
      : bits_(float_to_bfloat_bits(static_cast<float>(x))) {}

  template <
      typename T,
      typename = typename enable_if<can_convert_to_bfloat<T>>::type>
  constexpr METAL_FUNC _MLX_BFloat16(T x) threadgroup
      : bits_(float_to_bfloat_bits(static_cast<float>(x))) {}

  template <
      typename T,
      typename = typename enable_if<can_convert_to_bfloat<T>>::type>
  constexpr METAL_FUNC _MLX_BFloat16(T x) device
      : bits_(float_to_bfloat_bits(static_cast<float>(x))) {}

  template <
      typename T,
      typename = typename enable_if<can_convert_to_bfloat<T>>::type>
  constexpr METAL_FUNC _MLX_BFloat16(T x) constant
      : bits_(float_to_bfloat_bits(static_cast<float>(x))) {}

  /////////////////////////////////////////////////////////////////////////////
  // Conversions from bfloat

  template <
      typename T,
      typename = typename enable_if<can_convert_from_bfloat<T>>::type>
  constexpr METAL_FUNC operator T() const thread {
    return static_cast<T>(bfloat_bits_to_float(bits_));
  }

  template <
      typename T,
      typename = typename enable_if<can_convert_from_bfloat<T>>::type>
  constexpr METAL_FUNC operator T() const threadgroup {
    return static_cast<T>(bfloat_bits_to_float(bits_));
  }

  template <
      typename T,
      typename = typename enable_if<can_convert_from_bfloat<T>>::type>
  constexpr METAL_FUNC operator T() const device {
    return static_cast<T>(bfloat_bits_to_float(bits_));
  }

  template <
      typename T,
      typename = typename enable_if<can_convert_from_bfloat<T>>::type>
  constexpr METAL_FUNC operator T() const constant {
    return static_cast<T>(bfloat_bits_to_float(bits_));
  }
};

/////////////////////////////////////////////////////////////////////////////
// Bfloat operators
/////////////////////////////////////////////////////////////////////////////

/////////////////////////////////////////////////////////////////////////////
// Unary ops
constexpr METAL_FUNC _MLX_BFloat16 operator-(_MLX_BFloat16 x) {
  return -static_cast<float>(x);
}

/////////////////////////////////////////////////////////////////////////////
// Binary operators
#define bfloat_binop_base(__op__, __operator__, otype, atype, btype, ctype) \
  constexpr METAL_FUNC otype __operator__(atype lhs, btype rhs) {           \
    return static_cast<ctype>(lhs) __op__ static_cast<ctype>(rhs);          \
  }

#define bfloat_binop_helper(__op__, __operator__, otype, itype, ctype)    \
  constexpr METAL_FUNC otype __operator__(_MLX_BFloat16 lhs, itype rhs) { \
    return static_cast<ctype>(lhs) __op__ static_cast<ctype>(rhs);        \
  }                                                                       \
  constexpr METAL_FUNC otype __operator__(itype lhs, _MLX_BFloat16 rhs) { \
    return static_cast<ctype>(lhs) __op__ static_cast<ctype>(rhs);        \
  }

/////////////////////////////////////////////////////////////////////////////
// Arithmetic Operators
#define bfloat_binop(_op_, _operator_)                                       \
  bfloat_binop_base(                                                         \
      _op_, _operator_, _MLX_BFloat16, _MLX_BFloat16, _MLX_BFloat16, float); \
  bfloat_binop_helper(_op_, _operator_, float, float, float);                \
  bfloat_binop_helper(_op_, _operator_, float, half, float);                 \
  bfloat_binop_helper(_op_, _operator_, _MLX_BFloat16, int32_t, float);      \
  bfloat_binop_helper(_op_, _operator_, _MLX_BFloat16, uint32_t, float);     \
  bfloat_binop_helper(_op_, _operator_, _MLX_BFloat16, int64_t, float);      \
  bfloat_binop_helper(_op_, _operator_, _MLX_BFloat16, uint64_t, float);

bfloat_binop(+, operator+);
bfloat_binop(-, operator-);
bfloat_binop(*, operator*);
bfloat_binop(/, operator/);

/////////////////////////////////////////////////////////////////////////////
// Comparison ops
#define bfloat_compop(__op__, __operator__)                             \
  bfloat_binop_base(                                                    \
      __op__, __operator__, bool, _MLX_BFloat16, _MLX_BFloat16, float); \
  bfloat_binop_helper(__op__, __operator__, bool, float, float);        \
  bfloat_binop_helper(__op__, __operator__, bool, half, float);         \
  bfloat_binop_helper(__op__, __operator__, bool, int32_t, float);      \
  bfloat_binop_helper(__op__, __operator__, bool, uint32_t, float);     \
  bfloat_binop_helper(__op__, __operator__, bool, int64_t, float);      \
  bfloat_binop_helper(__op__, __operator__, bool, uint64_t, float);

bfloat_compop(>, operator>);
bfloat_compop(<, operator<);
bfloat_compop(>=, operator>=);
bfloat_compop(<=, operator<=);
bfloat_compop(==, operator==);
bfloat_compop(!=, operator!=);

#undef bfloat_compop
#undef bfloat_binop_base
#undef bfloat_binop_helper
#undef bfloat_binop

/////////////////////////////////////////////////////////////////////////////
// Inplace Operators
#define bfloat_inplace_op_helper(__op__, __operator__, itype, addr_space) \
  constexpr METAL_FUNC addr_space _MLX_BFloat16& __operator__(            \
      addr_space _MLX_BFloat16& lhs, itype rhs) {                         \
    lhs = static_cast<float>(lhs) __op__ static_cast<float>(rhs);         \
    return lhs;                                                           \
  }                                                                       \
  constexpr METAL_FUNC addr_space itype& __operator__(                    \
      addr_space itype& lhs, _MLX_BFloat16 rhs) {                         \
    lhs = static_cast<float>(lhs) __op__ static_cast<float>(rhs);         \
    return lhs;                                                           \
  }

#define bfloat_inplace_op_addr_space_helper(__op__, __operator__, itype) \
  bfloat_inplace_op_helper(__op__, __operator__, itype, device);         \
  bfloat_inplace_op_helper(__op__, __operator__, itype, thread);         \
  bfloat_inplace_op_helper(__op__, __operator__, itype, threadgroup);

#define bfloat_inplace_op(itype)                             \
  bfloat_inplace_op_addr_space_helper(+, operator+=, itype); \
  bfloat_inplace_op_addr_space_helper(-, operator-=, itype); \
  bfloat_inplace_op_addr_space_helper(*, operator*=, itype); \
  bfloat_inplace_op_addr_space_helper(/, operator/=, itype);

bfloat_inplace_op(float);
bfloat_inplace_op(half);
bfloat_inplace_op(int16_t);
bfloat_inplace_op(int32_t);
bfloat_inplace_op(int64_t);
bfloat_inplace_op(uint16_t);
bfloat_inplace_op(uint32_t);
bfloat_inplace_op(uint64_t);

#undef bfloat_inplace_op_helper
#undef bfloat_inplace_op_addr_space_helper
#undef bfloat_inplace_op

#define bfloat_inplace_op_helper(__op__, __operator__, addr_space) \
  constexpr METAL_FUNC addr_space _MLX_BFloat16& __operator__(     \
      addr_space _MLX_BFloat16& lhs, _MLX_BFloat16 rhs) {          \
    lhs = static_cast<float>(lhs) __op__ static_cast<float>(rhs);  \
    return lhs;                                                    \
  }

#define bfloat_inplace_op_addr_space_helper(__op__, __operator__) \
  bfloat_inplace_op_helper(__op__, __operator__, device);         \
  bfloat_inplace_op_helper(__op__, __operator__, thread);         \
  bfloat_inplace_op_helper(__op__, __operator__, threadgroup);

bfloat_inplace_op_addr_space_helper(+, operator+=);
bfloat_inplace_op_addr_space_helper(-, operator-=);
bfloat_inplace_op_addr_space_helper(*, operator*=);
bfloat_inplace_op_addr_space_helper(/, operator/=);

#undef bfloat_inplace_op_helper
#undef bfloat_inplace_op_addr_space_helper

/////////////////////////////////////////////////////////////////////////////
// Bfloat typedef
/////////////////////////////////////////////////////////////////////////////

typedef struct _MLX_BFloat16 bfloat16_t;

#endif



///////////////////////////////////////////////////////////////////////////////
// Metal math for bfloat16
///////////////////////////////////////////////////////////////////////////////

/*

Following the Metal Shading Language Specification (Metal 3.1)

"bfloat is an extended itypeing point type that only allows implicit conversion
 to a type of greater itypeing point rank. While bfloat can be implicitly
 converted to itype, it cannot be implicitly converted to half, and neither
 itype nor half can be implicitly converted to bfloat."

Further, as far as I can tell, the stdlib math/simd functions are not defined
for bfloat and calling with an argument of type bfloat will result in that
argument getting implicitly converted to itype which then returns an output
that is (likely) a itype which cannot be implicitly converted into a bfloat

This leads to situations where
bfloat a = 5.0bf;
bfloat b = metal::abs(a); // this will throw an error since abs return itype
bfloat c = static_cast<bfloat>(metal::abs(a)); // this is fine

For the moment, I will be adding overloaded instantiations of the math
functions to accordingly automatically handle the casting

*/

#define instantiate_metal_math_funcs(itype, otype, ctype, mfast)               \
                                                                               \
  METAL_FUNC otype abs(itype x) {                                              \
    return static_cast<otype>(__metal_fabs(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype acos(itype x) {                                             \
    return static_cast<otype>(__metal_acos(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype acosh(itype x) {                                            \
    return static_cast<otype>(__metal_acosh(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype asin(itype x) {                                             \
    return static_cast<otype>(__metal_asin(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype asinh(itype x) {                                            \
    return static_cast<otype>(__metal_asinh(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype atan(itype y_over_x) {                                      \
    return static_cast<otype>(                                                 \
        __metal_atan(static_cast<ctype>(y_over_x), mfast));                    \
  }                                                                            \
  METAL_FUNC otype atan2(itype y, itype x) {                                   \
    return static_cast<otype>(                                                 \
        __metal_atan2(static_cast<ctype>(y), static_cast<ctype>(x), mfast));   \
  }                                                                            \
  METAL_FUNC otype atanh(itype x) {                                            \
    return static_cast<otype>(__metal_atanh(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype ceil(itype x) {                                             \
    return static_cast<otype>(__metal_ceil(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype cos(itype x) {                                              \
    return static_cast<otype>(__metal_cos(static_cast<ctype>(x), mfast));      \
  }                                                                            \
  METAL_FUNC otype cosh(itype x) {                                             \
    return static_cast<otype>(__metal_cosh(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype cospi(itype x) {                                            \
    return static_cast<otype>(__metal_cospi(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype divide(itype x, itype y) {                                  \
    return static_cast<otype>(                                                 \
        __metal_divide(static_cast<ctype>(x), static_cast<ctype>(y), mfast));  \
  }                                                                            \
  METAL_FUNC otype exp(itype x) {                                              \
    return static_cast<otype>(__metal_exp(static_cast<ctype>(x), mfast));      \
  }                                                                            \
  METAL_FUNC otype exp10(itype x) {                                            \
    return static_cast<otype>(__metal_exp10(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype exp2(itype x) {                                             \
    return static_cast<otype>(__metal_exp2(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype fabs(itype x) {                                             \
    return static_cast<otype>(__metal_fabs(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype fdim(itype x, itype y) {                                    \
    ctype t = static_cast<ctype>(x - y);                                       \
    return static_cast<otype>(select(t, ctype(0), t < ctype(0) || x == y));    \
  }                                                                            \
  METAL_FUNC otype floor(itype x) {                                            \
    return static_cast<otype>(__metal_floor(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype fma(itype x, itype y, itype z) {                            \
    return static_cast<otype>(__metal_fma(                                     \
        static_cast<ctype>(x), static_cast<ctype>(y), static_cast<ctype>(z))); \
  }                                                                            \
  METAL_FUNC otype fmax(itype x, itype y) {                                    \
    return static_cast<otype>(                                                 \
        __metal_fmax(static_cast<ctype>(x), static_cast<ctype>(y), mfast));    \
  }                                                                            \
  METAL_FUNC otype fmax3(itype x, itype y, itype z) {                          \
    return static_cast<otype>(__metal_fmax3(                                   \
        static_cast<ctype>(x),                                                 \
        static_cast<ctype>(y),                                                 \
        static_cast<ctype>(z),                                                 \
        mfast));                                                               \
  }                                                                            \
  METAL_FUNC otype fmedian3(itype x, itype y, itype z) {                       \
    return static_cast<otype>(__metal_fmedian3(                                \
        static_cast<ctype>(x),                                                 \
        static_cast<ctype>(y),                                                 \
        static_cast<ctype>(z),                                                 \
        mfast));                                                               \
  }                                                                            \
  METAL_FUNC otype fmin(itype x, itype y) {                                    \
    return static_cast<otype>(                                                 \
        __metal_fmin(static_cast<ctype>(x), static_cast<ctype>(y), mfast));    \
  }                                                                            \
  METAL_FUNC otype fmin3(itype x, itype y, itype z) {                          \
    return static_cast<otype>(__metal_fmin3(                                   \
        static_cast<ctype>(x),                                                 \
        static_cast<ctype>(y),                                                 \
        static_cast<ctype>(z),                                                 \
        mfast));                                                               \
  }                                                                            \
  METAL_FUNC otype fmod(itype x, itype y) {                                    \
    return static_cast<otype>(                                                 \
        __metal_fmod(static_cast<ctype>(x), static_cast<ctype>(y), mfast));    \
  }                                                                            \
  METAL_FUNC otype fract(itype x) {                                            \
    return static_cast<otype>(__metal_fract(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype frexp(itype x, thread int& exp) {                           \
    return static_cast<otype>(__metal_frexp(static_cast<ctype>(x), &exp));     \
  }                                                                            \
  METAL_FUNC otype ldexp(itype x, int k) {                                     \
    return static_cast<otype>(__metal_ldexp(static_cast<ctype>(x), k, mfast)); \
  }                                                                            \
  METAL_FUNC otype log(itype x) {                                              \
    return static_cast<otype>(__metal_log(static_cast<ctype>(x), mfast));      \
  }                                                                            \
  METAL_FUNC otype log10(itype x) {                                            \
    return static_cast<otype>(__metal_log10(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype log2(itype x) {                                             \
    return static_cast<otype>(__metal_log2(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype max(itype x, itype y) {                                     \
    return static_cast<otype>(                                                 \
        __metal_fmax(static_cast<ctype>(x), static_cast<ctype>(y), mfast));    \
  }                                                                            \
  METAL_FUNC otype max3(itype x, itype y, itype z) {                           \
    return static_cast<otype>(__metal_fmax3(                                   \
        static_cast<ctype>(x),                                                 \
        static_cast<ctype>(y),                                                 \
        static_cast<ctype>(z),                                                 \
        mfast));                                                               \
  }                                                                            \
  METAL_FUNC otype median3(itype x, itype y, itype z) {                        \
    return static_cast<otype>(__metal_fmedian3(                                \
        static_cast<ctype>(x),                                                 \
        static_cast<ctype>(y),                                                 \
        static_cast<ctype>(z),                                                 \
        mfast));                                                               \
  }                                                                            \
  METAL_FUNC otype min(itype x, itype y) {                                     \
    return static_cast<otype>(                                                 \
        __metal_fmin(static_cast<ctype>(x), static_cast<ctype>(y), mfast));    \
  }                                                                            \
  METAL_FUNC otype min3(itype x, itype y, itype z) {                           \
    return static_cast<otype>(__metal_fmin3(                                   \
        static_cast<ctype>(x),                                                 \
        static_cast<ctype>(y),                                                 \
        static_cast<ctype>(z),                                                 \
        mfast));                                                               \
  }                                                                            \
  METAL_FUNC otype nextafter(itype x, itype y) {                               \
    return static_cast<otype>(                                                 \
        __metal_nextafter(static_cast<ctype>(x), static_cast<ctype>(y)));      \
  }                                                                            \
  METAL_FUNC otype pow(itype x, itype y) {                                     \
    return static_cast<otype>(                                                 \
        __metal_pow(static_cast<ctype>(x), static_cast<ctype>(y), mfast));     \
  }                                                                            \
  METAL_FUNC otype powr(itype x, itype y) {                                    \
    return static_cast<otype>(                                                 \
        __metal_powr(static_cast<ctype>(x), static_cast<ctype>(y), mfast));    \
  }                                                                            \
  METAL_FUNC otype rint(itype x) {                                             \
    return static_cast<otype>(__metal_rint(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype round(itype x) {                                            \
    return static_cast<otype>(__metal_round(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype rsqrt(itype x) {                                            \
    return static_cast<otype>(__metal_rsqrt(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype sin(itype x) {                                              \
    return static_cast<otype>(__metal_sin(static_cast<ctype>(x), mfast));      \
  }                                                                            \
  METAL_FUNC otype sinh(itype x) {                                             \
    return static_cast<otype>(__metal_sinh(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype sinpi(itype x) {                                            \
    return static_cast<otype>(__metal_sinpi(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype sqrt(itype x) {                                             \
    return static_cast<otype>(__metal_sqrt(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype tan(itype x) {                                              \
    return static_cast<otype>(__metal_tan(static_cast<ctype>(x), mfast));      \
  }                                                                            \
  METAL_FUNC otype tanh(itype x) {                                             \
    return static_cast<otype>(__metal_tanh(static_cast<ctype>(x), mfast));     \
  }                                                                            \
  METAL_FUNC otype tanpi(itype x) {                                            \
    return static_cast<otype>(__metal_tanpi(static_cast<ctype>(x), mfast));    \
  }                                                                            \
  METAL_FUNC otype trunc(itype x) {                                            \
    return static_cast<otype>(__metal_trunc(static_cast<ctype>(x), mfast));    \
  }

namespace metal {

instantiate_metal_math_funcs(
    bfloat16_t,
    bfloat16_t,
    float,
    __METAL_MAYBE_FAST_MATH__);

namespace fast {

instantiate_metal_math_funcs(
    bfloat16_t,
    bfloat16_t,
    float,
    __METAL_FAST_MATH__);

} // namespace fast

namespace precise {

instantiate_metal_math_funcs(
    bfloat16_t,
    bfloat16_t,
    float,
    __METAL_PRECISE_MATH__);

} // namespace precise

} // namespace metal

METAL_FUNC uint get_strided_index(
    uint idx,
    constant size_t &num_dims,
    constant size_t *dims,
    constant size_t *strides
) {
    uint strided_i = 0;
    for (uint d = 0; d < num_dims; d++) {
        uint dim_idx = num_dims - 1 - d;
        strided_i += (idx % dims[dim_idx]) * strides[dim_idx];
        idx /= dims[dim_idx];
    }
    return strided_i;
}

#define AFFINE(FN_NAME, T) \
kernel void FN_NAME( \
    constant size_t &dim, \
    constant float &mul, \
    constant float &add, \
    device const T *input,  \
    device T *output, \
    uint id [[ thread_position_in_grid ]] \
) { \
    if (id >= dim) { \
        return; \
    } \
    output[id] = T(fma(float(input[id]), mul, add)); \
} \
kernel void FN_NAME##_strided( \
    constant size_t &dim, \
    constant size_t &num_dims, \
    constant size_t *dims, \
    constant size_t *strides, \
    constant float &mul, \
    constant float &add, \
    device const T *input,  \
    device T *output, \
    uint id [[ thread_position_in_grid ]] \
) { \
    if (id >= dim) { \
        return; \
    } \
    output[id] = T(fma(float(input[get_strided_index(id, num_dims, dims, strides)]), mul, add)); \
}

#define POWF(FN_NAME, TYPENAME) \
kernel void FN_NAME( \
    constant size_t &dim, \
    constant float &mul, \
    device const TYPENAME *input,  \
    device TYPENAME *output, \
    uint id [[ thread_position_in_grid ]] \
) { \
    if (id >= dim) { \
        return; \
    } \
    output[id] = TYPENAME(pow(input[id], TYPENAME(mul))); \
} \
kernel void FN_NAME##_strided( \
    constant size_t &dim, \
    constant size_t &num_dims, \
    constant size_t *dims, \
    constant size_t *strides, \
    constant float &mul, \
    device const TYPENAME *input,  \
    device TYPENAME *output, \
    uint id [[ thread_position_in_grid ]] \
) { \
    if (id >= dim) { \
        return; \
    } \
    output[id] = TYPENAME(pow(input[get_strided_index(id, num_dims, dims, strides)], TYPENAME(mul))); \
}

#define ELU(FN_NAME, TYPENAME) \
kernel void FN_NAME( \
    constant size_t &dim, \
    constant float &mul, \
    device const TYPENAME *input,  \
    device TYPENAME *output, \
    uint id [[ thread_position_in_grid ]] \
) { \
    if (id >= dim) { \
        return; \
    } \
    const TYPENAME x = input[id]; \
    output[id] = TYPENAME((x > 0)?x: static_cast<TYPENAME>(mul) * (exp(x) - 1)); \
} \
kernel void FN_NAME##_strided( \
    constant size_t &dim, \
    constant size_t &num_dims, \
    constant size_t *dims, \
    constant size_t *strides, \
    constant float &mul, \
    device const TYPENAME *input,  \
    device TYPENAME *output, \
    uint id [[ thread_position_in_grid ]] \
) { \
    if (id >= dim) { \
        return; \
    } \
    const TYPENAME x = input[get_strided_index(id, num_dims, dims, strides)]; \
    output[id] = TYPENAME((x > 0)?x: static_cast<TYPENAME>(mul) * (exp(x) - 1)); \
} \


AFFINE(affine_i8, int8_t)
AFFINE(affine_u8, uint8_t)
AFFINE(affine_u32, uint32_t)
AFFINE(affine_f32, float)
AFFINE(affine_f16, half)
POWF(powf_f32, float)
POWF(powf_f16, half)
ELU(elu_f32, float)
ELU(elu_f16, half)

AFFINE(affine_bf16, bfloat16_t);
POWF(powf_bf16, bfloat16_t);
ELU(elu_bf16, bfloat16_t);