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
#include "ops_binary.h"
#include "simd_utils.h"
#include "thread_utils.h"
#include <math.h>
// ============================================================================
// BINARY OPERATION IMPLEMENTATION MACROS
// ============================================================================
//
// This file implements element-wise binary operations for tensors.
// Operations are optimized by checking tensor contiguity to use faster loops.
//
// Four optimization paths:
// 1. Both tensors contiguous: Simple linear iteration
// 2. LHS contiguous, RHS strided: Compute RHS indices only
// 3. LHS strided, RHS contiguous: Compute LHS indices only
// 4. Both strided: Compute both indices
//
// Metadata layout:
// - metadata[0]: num_els (total number of elements)
// - metadata[1]: num_dims (number of dimensions)
// - metadata[2..2+num_dims]: lhs_shape
// - metadata[2+num_dims..2+2*num_dims]: rhs_shape
// - metadata[2+2*num_dims..2+3*num_dims]: lhs_strides
// - metadata[2+3*num_dims..2+4*num_dims]: rhs_strides
// - metadata[2+4*num_dims]: lhs_offset
// - metadata[2+4*num_dims+1]: rhs_offset
/// Macro to implement a binary operation returning the same type
///
/// Generates optimized implementations with contiguity checks.
///
/// @param TYPE C type for the operation (e.g., f32_t, i32_t)
/// @param TYPE_SUFFIX Suffix for function naming (e.g., f32, i32)
/// @param OP_NAME Operation name (e.g., add, mul)
/// @param FUNC Expression defining the operation (e.g., x + y, x * y)
///
/// Edge cases:
/// - Division by zero: Handled in FUNC expression (e.g., div uses x / y which produces inf/nan)
/// - Integer overflow: Not checked for performance; wraps according to C semantics
/// - Unsigned underflow: Checked explicitly (e.g., sub uses (x > y) ? (x - y) : 0)
#define IMPL_BINARY_OP(TYPE, TYPE_SUFFIX, OP_NAME, FUNC) \
typedef struct { \
const TYPE *lhs; \
const TYPE *rhs; \
TYPE *output; \
size_t start; \
size_t end; \
size_t lhs_offset; \
size_t rhs_offset; \
} binary_##OP_NAME##_##TYPE_SUFFIX##_args_t; \
\
static void *binary_##OP_NAME##_##TYPE_SUFFIX##_worker(void *arg) { \
binary_##OP_NAME##_##TYPE_SUFFIX##_args_t *args = \
(binary_##OP_NAME##_##TYPE_SUFFIX##_args_t *)arg; \
for (size_t i = args->start; i < args->end; i++) { \
TYPE x = args->lhs[args->lhs_offset + i]; \
TYPE y = args->rhs[args->rhs_offset + i]; \
args->output[i] = FUNC; \
} \
return NULL; \
} \
\
void OP_NAME##_##TYPE_SUFFIX(const void *lhs, const void *rhs, void *output, \
const size_t *metadata) { \
const TYPE *l = (const TYPE *)lhs; \
const TYPE *r = (const TYPE *)rhs; \
TYPE *out = (TYPE *)output; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
const size_t *lhs_shape = metadata + 2; \
const size_t *rhs_shape = metadata + 2 + num_dims; \
const size_t *lhs_strides = metadata + 2 + 2 * num_dims; \
const size_t *rhs_strides = metadata + 2 + 3 * num_dims; \
const size_t lhs_offset = metadata[2 + 4 * num_dims]; \
const size_t rhs_offset = metadata[2 + 4 * num_dims + 1]; \
\
bool lhs_cont = is_contiguous(num_dims, lhs_shape, lhs_strides); \
bool rhs_cont = is_contiguous(num_dims, rhs_shape, rhs_strides); \
\
if (lhs_cont && rhs_cont) { \
const size_t min_work_per_thread = 100000; \
size_t num_threads = get_optimal_threads(num_els, min_work_per_thread); \
\
if (num_threads > 1) { \
thread_t threads[num_threads]; \
binary_##OP_NAME##_##TYPE_SUFFIX##_args_t args[num_threads]; \
\
size_t chunk_size = num_els / num_threads; \
size_t remaining = num_els % num_threads; \
\
for (size_t t = 0; t < num_threads; t++) { \
args[t].lhs = l; \
args[t].rhs = r; \
args[t].output = out; \
args[t].lhs_offset = lhs_offset; \
args[t].rhs_offset = rhs_offset; \
args[t].start = t * chunk_size + (t < remaining ? t : remaining); \
args[t].end = args[t].start + chunk_size + (t < remaining ? 1 : 0); \
thread_create(&threads[t], binary_##OP_NAME##_##TYPE_SUFFIX##_worker, \
&args[t]); \
} \
\
for (size_t t = 0; t < num_threads; t++) { \
thread_join(threads[t]); \
} \
} else { \
for (size_t i = 0; i < num_els; i++) { \
TYPE x = l[lhs_offset + i]; \
TYPE y = r[rhs_offset + i]; \
out[i] = FUNC; \
} \
} \
} else if (lhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
TYPE x = l[lhs_offset + i]; \
TYPE y = r[rhs_i]; \
out[i] = FUNC; \
} \
} else if (rhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
TYPE x = l[lhs_i]; \
TYPE y = r[rhs_offset + i]; \
out[i] = FUNC; \
} \
} else { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
TYPE x = l[lhs_i]; \
TYPE y = r[rhs_i]; \
out[i] = FUNC; \
} \
} \
}
/// Macro to implement a binary operation returning boolean (uint8_t)
///
/// Used for logical and comparison operations. Output is always uint8_t
/// where 0 represents false and 1 represents true.
///
/// @param TYPE C type for the input operands
/// @param TYPE_SUFFIX Suffix for function naming
/// @param OP_NAME Operation name
/// @param FUNC Boolean expression to evaluate (e.g., x < y, x && y)
#define IMPL_BINARY_TO_BOOL(TYPE, TYPE_SUFFIX, OP_NAME, FUNC) \
void OP_NAME##_##TYPE_SUFFIX(const void *lhs, const void *rhs, void *output, \
const size_t *metadata) { \
const TYPE *l = (const TYPE *)lhs; \
const TYPE *r = (const TYPE *)rhs; \
uint8_t *out = (uint8_t *)output; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
const size_t *lhs_shape = metadata + 2; \
const size_t *rhs_shape = metadata + 2 + num_dims; \
const size_t *lhs_strides = metadata + 2 + 2 * num_dims; \
const size_t *rhs_strides = metadata + 2 + 3 * num_dims; \
const size_t lhs_offset = metadata[2 + 4 * num_dims]; \
const size_t rhs_offset = metadata[2 + 4 * num_dims + 1]; \
\
bool lhs_cont = (metadata == NULL) || is_contiguous(num_dims, lhs_shape, lhs_strides); \
bool rhs_cont = (metadata == NULL) || is_contiguous(num_dims, rhs_shape, rhs_strides); \
\
if (lhs_cont && rhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
TYPE x = l[lhs_offset + i]; \
TYPE y = r[rhs_offset + i]; \
out[i] = (FUNC) ? 1 : 0; \
} \
} else { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_cont \
? (lhs_offset + i) \
: (lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides)); \
size_t rhs_i = \
rhs_cont \
? (rhs_offset + i) \
: (rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides)); \
TYPE x = l[lhs_i]; \
TYPE y = r[rhs_i]; \
out[i] = (FUNC) ? 1 : 0; \
} \
} \
}
/// Macro to implement a binary operation with type conversion for reduced-precision floats
///
/// For FP8, FP16, and BF16 types, operations are performed in FP32 for accuracy.
/// Values are converted to float, operation is performed, then converted back.
///
/// @param TYPE Storage type (uint8_t for FP8, uint16_t for FP16/BF16)
/// @param TYPE_SUFFIX Suffix for function naming
/// @param OP_NAME Operation name
/// @param FUNC Expression defining the operation (in float precision)
/// @param TO_FLOAT Function to convert from storage type to float
/// @param FROM_FLOAT Function to convert from float to storage type
// Macros for FP8/FP16/BF16 with float conversion
#define IMPL_BINARY_OP_CONVERT(TYPE, TYPE_SUFFIX, OP_NAME, FUNC, TO_FLOAT, FROM_FLOAT) \
void OP_NAME##_##TYPE_SUFFIX(const void *lhs, const void *rhs, void *output, \
const size_t *metadata) { \
const TYPE *l = (const TYPE *)lhs; \
const TYPE *r = (const TYPE *)rhs; \
TYPE *out = (TYPE *)output; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
const size_t *lhs_shape = metadata + 2; \
const size_t *rhs_shape = metadata + 2 + num_dims; \
const size_t *lhs_strides = metadata + 2 + 2 * num_dims; \
const size_t *rhs_strides = metadata + 2 + 3 * num_dims; \
const size_t lhs_offset = metadata[2 + 4 * num_dims]; \
const size_t rhs_offset = metadata[2 + 4 * num_dims + 1]; \
\
bool lhs_cont = (metadata == NULL) || is_contiguous(num_dims, lhs_shape, lhs_strides); \
bool rhs_cont = (metadata == NULL) || is_contiguous(num_dims, rhs_shape, rhs_strides); \
\
if (lhs_cont && rhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
float x = TO_FLOAT(l[lhs_offset + i]); \
float y = TO_FLOAT(r[rhs_offset + i]); \
out[i] = FROM_FLOAT(FUNC); \
} \
} else if (lhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
float x = TO_FLOAT(l[lhs_offset + i]); \
float y = TO_FLOAT(r[rhs_i]); \
out[i] = FROM_FLOAT(FUNC); \
} \
} else if (rhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
float x = TO_FLOAT(l[lhs_i]); \
float y = TO_FLOAT(r[rhs_offset + i]); \
out[i] = FROM_FLOAT(FUNC); \
} \
} else { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
float x = TO_FLOAT(l[lhs_i]); \
float y = TO_FLOAT(r[rhs_i]); \
out[i] = FROM_FLOAT(FUNC); \
} \
} \
}
/// Macro to implement a boolean-returning operation with type conversion
///
/// Combines IMPL_BINARY_TO_BOOL and type conversion for reduced-precision floats.
///
/// @param TYPE Storage type (uint8_t for FP8, uint16_t for FP16/BF16)
/// @param TYPE_SUFFIX Suffix for function naming
/// @param OP_NAME Operation name
/// @param FUNC Boolean expression (evaluated in float precision)
/// @param TO_FLOAT Function to convert from storage type to float
#define IMPL_BINARY_TO_BOOL_CONVERT(TYPE, TYPE_SUFFIX, OP_NAME, FUNC, TO_FLOAT) \
void OP_NAME##_##TYPE_SUFFIX(const void *lhs, const void *rhs, void *output, \
const size_t *metadata) { \
const TYPE *l = (const TYPE *)lhs; \
const TYPE *r = (const TYPE *)rhs; \
uint8_t *out = (uint8_t *)output; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
const size_t *lhs_shape = metadata + 2; \
const size_t *rhs_shape = metadata + 2 + num_dims; \
const size_t *lhs_strides = metadata + 2 + 2 * num_dims; \
const size_t *rhs_strides = metadata + 2 + 3 * num_dims; \
const size_t lhs_offset = metadata[2 + 4 * num_dims]; \
const size_t rhs_offset = metadata[2 + 4 * num_dims + 1]; \
\
bool lhs_cont = (metadata == NULL) || is_contiguous(num_dims, lhs_shape, lhs_strides); \
bool rhs_cont = (metadata == NULL) || is_contiguous(num_dims, rhs_shape, rhs_strides); \
\
if (lhs_cont && rhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
float x = TO_FLOAT(l[lhs_offset + i]); \
float y = TO_FLOAT(r[rhs_offset + i]); \
out[i] = (FUNC) ? 1 : 0; \
} \
} else { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_cont \
? (lhs_offset + i) \
: (lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides)); \
size_t rhs_i = \
rhs_cont \
? (rhs_offset + i) \
: (rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides)); \
float x = TO_FLOAT(l[lhs_i]); \
float y = TO_FLOAT(r[rhs_i]); \
out[i] = (FUNC) ? 1 : 0; \
} \
} \
}
// ============================================================================
// SIMD-OPTIMIZED F32 OPERATIONS
// ============================================================================
/// SIMD-optimized binary operations for F32
/// Falls back to scalar if SIMD not available
#if SIMD_F32_WIDTH > 1
#define IMPL_BINARY_OP_F32_SIMD(OP_NAME, SIMD_OP, SCALAR_OP) \
void OP_NAME##_f32(const void *lhs, const void *rhs, void *output, const size_t *metadata) { \
const f32_t *l = (const f32_t *)lhs; \
const f32_t *r = (const f32_t *)rhs; \
f32_t *out = (f32_t *)output; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
const size_t *lhs_shape = metadata + 2; \
const size_t *rhs_shape = metadata + 2 + num_dims; \
const size_t *lhs_strides = metadata + 2 + 2 * num_dims; \
const size_t *rhs_strides = metadata + 2 + 3 * num_dims; \
const size_t lhs_offset = metadata[2 + 4 * num_dims]; \
const size_t rhs_offset = metadata[2 + 4 * num_dims + 1]; \
\
bool lhs_cont = is_contiguous(num_dims, lhs_shape, lhs_strides); \
bool rhs_cont = is_contiguous(num_dims, rhs_shape, rhs_strides); \
\
if (lhs_cont && rhs_cont) { \
/* SIMD path for contiguous data */ \
const f32_t *lhs_data = l + lhs_offset; \
const f32_t *rhs_data = r + rhs_offset; \
size_t i = 0; \
const size_t simd_end = (num_els / SIMD_F32_WIDTH) * SIMD_F32_WIDTH; \
\
for (; i < simd_end; i += SIMD_F32_WIDTH) { \
simd_f32_t va = simd_f32_load(&lhs_data[i]); \
simd_f32_t vb = simd_f32_load(&rhs_data[i]); \
simd_f32_t vc = SIMD_OP(va, vb); \
simd_f32_store(&out[i], vc); \
} \
\
/* Scalar remainder */ \
for (; i < num_els; i++) { \
f32_t x = lhs_data[i]; \
f32_t y = rhs_data[i]; \
out[i] = SCALAR_OP; \
} \
} else if (lhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
f32_t x = l[lhs_offset + i]; \
f32_t y = r[rhs_i]; \
out[i] = SCALAR_OP; \
} \
} else if (rhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
f32_t x = l[lhs_i]; \
f32_t y = r[rhs_offset + i]; \
out[i] = SCALAR_OP; \
} \
} else { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
f32_t x = l[lhs_i]; \
f32_t y = r[rhs_i]; \
out[i] = SCALAR_OP; \
} \
} \
}
IMPL_BINARY_OP_F32_SIMD(add, simd_f32_add, x + y)
IMPL_BINARY_OP_F32_SIMD(sub, simd_f32_sub, x - y)
IMPL_BINARY_OP_F32_SIMD(mul, simd_f32_mul, x *y)
IMPL_BINARY_OP_F32_SIMD(div, simd_f32_div, x / y)
#else
// Fallback to scalar implementation
IMPL_BINARY_OP(f32_t, f32, add, x + y)
IMPL_BINARY_OP(f32_t, f32, sub, x - y)
IMPL_BINARY_OP(f32_t, f32, mul, x *y)
IMPL_BINARY_OP(f32_t, f32, div, x / y)
#endif
IMPL_BINARY_OP(f32_t, f32, pow, powf_opt(x, y))
IMPL_BINARY_OP(f32_t, f32, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(f32_t, f32, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(f32_t, f32, logical_and, (x != 0.0f && y != 0.0f))
IMPL_BINARY_TO_BOOL(f32_t, f32, logical_or, (x != 0.0f || y != 0.0f))
IMPL_BINARY_TO_BOOL(f32_t, f32, logical_xor, (x != 0.0f) != (y != 0.0f))
IMPL_BINARY_TO_BOOL(f32_t, f32, eq, x == y)
IMPL_BINARY_TO_BOOL(f32_t, f32, ne, x != y)
IMPL_BINARY_TO_BOOL(f32_t, f32, lt, x < y)
IMPL_BINARY_TO_BOOL(f32_t, f32, le, x <= y)
IMPL_BINARY_TO_BOOL(f32_t, f32, gt, x > y)
IMPL_BINARY_TO_BOOL(f32_t, f32, ge, x >= y)
// ============================================================================
// SIMD-OPTIMIZED F64 OPERATIONS
// ============================================================================
#if SIMD_F64_WIDTH > 1
#define IMPL_BINARY_OP_F64_SIMD(OP_NAME, SIMD_OP, SCALAR_OP) \
void OP_NAME##_f64(const void *lhs, const void *rhs, void *output, const size_t *metadata) { \
const f64_t *l = (const f64_t *)lhs; \
const f64_t *r = (const f64_t *)rhs; \
f64_t *out = (f64_t *)output; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
const size_t *lhs_shape = metadata + 2; \
const size_t *rhs_shape = metadata + 2 + num_dims; \
const size_t *lhs_strides = metadata + 2 + 2 * num_dims; \
const size_t *rhs_strides = metadata + 2 + 3 * num_dims; \
const size_t lhs_offset = metadata[2 + 4 * num_dims]; \
const size_t rhs_offset = metadata[2 + 4 * num_dims + 1]; \
\
bool lhs_cont = is_contiguous(num_dims, lhs_shape, lhs_strides); \
bool rhs_cont = is_contiguous(num_dims, rhs_shape, rhs_strides); \
\
if (lhs_cont && rhs_cont) { \
const f64_t *lhs_data = l + lhs_offset; \
const f64_t *rhs_data = r + rhs_offset; \
size_t i = 0; \
const size_t simd_end = (num_els / SIMD_F64_WIDTH) * SIMD_F64_WIDTH; \
\
for (; i < simd_end; i += SIMD_F64_WIDTH) { \
simd_f64_t va = simd_f64_load(&lhs_data[i]); \
simd_f64_t vb = simd_f64_load(&rhs_data[i]); \
simd_f64_t vc = SIMD_OP(va, vb); \
simd_f64_store(&out[i], vc); \
} \
\
for (; i < num_els; i++) { \
f64_t x = lhs_data[i]; \
f64_t y = rhs_data[i]; \
out[i] = SCALAR_OP; \
} \
} else if (lhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
f64_t x = l[lhs_offset + i]; \
f64_t y = r[rhs_i]; \
out[i] = SCALAR_OP; \
} \
} else if (rhs_cont) { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
f64_t x = l[lhs_i]; \
f64_t y = r[rhs_offset + i]; \
out[i] = SCALAR_OP; \
} \
} else { \
for (size_t i = 0; i < num_els; i++) { \
size_t lhs_i = \
lhs_offset + get_strided_index(i, num_dims, lhs_shape, lhs_strides); \
size_t rhs_i = \
rhs_offset + get_strided_index(i, num_dims, rhs_shape, rhs_strides); \
f64_t x = l[lhs_i]; \
f64_t y = r[rhs_i]; \
out[i] = SCALAR_OP; \
} \
} \
}
IMPL_BINARY_OP_F64_SIMD(add, simd_f64_add, x + y)
IMPL_BINARY_OP_F64_SIMD(sub, simd_f64_sub, x - y)
IMPL_BINARY_OP_F64_SIMD(mul, simd_f64_mul, x *y)
IMPL_BINARY_OP_F64_SIMD(div, simd_f64_div, x / y)
#else
// Fallback to scalar
IMPL_BINARY_OP(f64_t, f64, add, x + y)
IMPL_BINARY_OP(f64_t, f64, sub, x - y)
IMPL_BINARY_OP(f64_t, f64, mul, x *y)
IMPL_BINARY_OP(f64_t, f64, div, x / y)
#endif
IMPL_BINARY_OP(f64_t, f64, pow, pow_opt(x, y))
IMPL_BINARY_OP(f64_t, f64, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(f64_t, f64, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(f64_t, f64, logical_and, (x != 0.0 && y != 0.0))
IMPL_BINARY_TO_BOOL(f64_t, f64, logical_or, (x != 0.0 || y != 0.0))
IMPL_BINARY_TO_BOOL(f64_t, f64, logical_xor, (x != 0.0) != (y != 0.0))
IMPL_BINARY_TO_BOOL(f64_t, f64, eq, x == y)
IMPL_BINARY_TO_BOOL(f64_t, f64, ne, x != y)
IMPL_BINARY_TO_BOOL(f64_t, f64, lt, x < y)
IMPL_BINARY_TO_BOOL(f64_t, f64, le, x <= y)
IMPL_BINARY_TO_BOOL(f64_t, f64, gt, x > y)
IMPL_BINARY_TO_BOOL(f64_t, f64, ge, x >= y)
// ============================================================================
// BOOL OPERATIONS
// ============================================================================
IMPL_BINARY_OP(uint8_t, bool, add, x || y)
IMPL_BINARY_OP(uint8_t, bool, sub, x && !y)
IMPL_BINARY_OP(uint8_t, bool, mul, x &&y)
IMPL_BINARY_OP(uint8_t, bool, div, x &&y)
IMPL_BINARY_OP(uint8_t, bool, pow, x || !y)
IMPL_BINARY_OP(uint8_t, bool, maximum, x || y)
IMPL_BINARY_OP(uint8_t, bool, minimum, x &&y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, logical_and, x &&y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, logical_or, x || y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, logical_xor, x != y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, eq, x == y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, ne, x != y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, lt, !x && y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, le, !x || y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, gt, x && !y)
IMPL_BINARY_TO_BOOL(uint8_t, bool, ge, x || !y)
// ============================================================================
// FP8 E4M3 OPERATIONS
// ============================================================================
IMPL_BINARY_OP_CONVERT(uint8_t, f8e4m3, add, x + y, fp8_e4m3_to_float, float_to_fp8_e4m3)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e4m3, sub, x - y, fp8_e4m3_to_float, float_to_fp8_e4m3)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e4m3, mul, x *y, fp8_e4m3_to_float, float_to_fp8_e4m3)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e4m3, div, x / y, fp8_e4m3_to_float, float_to_fp8_e4m3)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e4m3, pow, powf(x, y), fp8_e4m3_to_float, float_to_fp8_e4m3)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e4m3, maximum, MAXIMUM(x, y), fp8_e4m3_to_float,
float_to_fp8_e4m3)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e4m3, minimum, MINIMUM(x, y), fp8_e4m3_to_float,
float_to_fp8_e4m3)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, logical_and, (x != 0.0f && y != 0.0f),
fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, logical_or, (x != 0.0f || y != 0.0f),
fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, logical_xor, (x != 0.0f) != (y != 0.0f),
fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, eq, x == y, fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, ne, x != y, fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, lt, x < y, fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, le, x <= y, fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, gt, x > y, fp8_e4m3_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e4m3, ge, x >= y, fp8_e4m3_to_float)
// ============================================================================
// FP8 E5M2 OPERATIONS
// ============================================================================
IMPL_BINARY_OP_CONVERT(uint8_t, f8e5m2, add, x + y, fp8_e5m2_to_float, float_to_fp8_e5m2)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e5m2, sub, x - y, fp8_e5m2_to_float, float_to_fp8_e5m2)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e5m2, mul, x *y, fp8_e5m2_to_float, float_to_fp8_e5m2)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e5m2, div, x / y, fp8_e5m2_to_float, float_to_fp8_e5m2)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e5m2, pow, powf(x, y), fp8_e5m2_to_float, float_to_fp8_e5m2)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e5m2, maximum, MAXIMUM(x, y), fp8_e5m2_to_float,
float_to_fp8_e5m2)
IMPL_BINARY_OP_CONVERT(uint8_t, f8e5m2, minimum, MINIMUM(x, y), fp8_e5m2_to_float,
float_to_fp8_e5m2)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, logical_and, (x != 0.0f && y != 0.0f),
fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, logical_or, (x != 0.0f || y != 0.0f),
fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, logical_xor, (x != 0.0f) != (y != 0.0f),
fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, eq, x == y, fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, ne, x != y, fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, lt, x < y, fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, le, x <= y, fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, gt, x > y, fp8_e5m2_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint8_t, f8e5m2, ge, x >= y, fp8_e5m2_to_float)
// ============================================================================
// BF16 OPERATIONS
// ============================================================================
IMPL_BINARY_OP_CONVERT(uint16_t, bf16, add, x + y, bf16_to_float, float_to_bf16)
IMPL_BINARY_OP_CONVERT(uint16_t, bf16, sub, x - y, bf16_to_float, float_to_bf16)
IMPL_BINARY_OP_CONVERT(uint16_t, bf16, mul, x *y, bf16_to_float, float_to_bf16)
IMPL_BINARY_OP_CONVERT(uint16_t, bf16, div, x / y, bf16_to_float, float_to_bf16)
IMPL_BINARY_OP_CONVERT(uint16_t, bf16, pow, powf(x, y), bf16_to_float, float_to_bf16)
IMPL_BINARY_OP_CONVERT(uint16_t, bf16, maximum, MAXIMUM(x, y), bf16_to_float, float_to_bf16)
IMPL_BINARY_OP_CONVERT(uint16_t, bf16, minimum, MINIMUM(x, y), bf16_to_float, float_to_bf16)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, logical_and, (x != 0.0f && y != 0.0f), bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, logical_or, (x != 0.0f || y != 0.0f), bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, logical_xor, (x != 0.0f) != (y != 0.0f), bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, eq, x == y, bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, ne, x != y, bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, lt, x < y, bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, le, x <= y, bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, gt, x > y, bf16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, bf16, ge, x >= y, bf16_to_float)
// ============================================================================
// FP16 OPERATIONS
// ============================================================================
IMPL_BINARY_OP_CONVERT(uint16_t, f16, add, x + y, fp16_to_float, float_to_fp16)
IMPL_BINARY_OP_CONVERT(uint16_t, f16, sub, x - y, fp16_to_float, float_to_fp16)
IMPL_BINARY_OP_CONVERT(uint16_t, f16, mul, x *y, fp16_to_float, float_to_fp16)
IMPL_BINARY_OP_CONVERT(uint16_t, f16, div, x / y, fp16_to_float, float_to_fp16)
IMPL_BINARY_OP_CONVERT(uint16_t, f16, pow, powf(x, y), fp16_to_float, float_to_fp16)
IMPL_BINARY_OP_CONVERT(uint16_t, f16, maximum, MAXIMUM(x, y), fp16_to_float, float_to_fp16)
IMPL_BINARY_OP_CONVERT(uint16_t, f16, minimum, MINIMUM(x, y), fp16_to_float, float_to_fp16)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, logical_and, (x != 0.0f && y != 0.0f), fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, logical_or, (x != 0.0f || y != 0.0f), fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, logical_xor, (x != 0.0f) != (y != 0.0f), fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, eq, x == y, fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, ne, x != y, fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, lt, x < y, fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, le, x <= y, fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, gt, x > y, fp16_to_float)
IMPL_BINARY_TO_BOOL_CONVERT(uint16_t, f16, ge, x >= y, fp16_to_float)
// ============================================================================
// UNSIGNED INTEGER OPERATIONS (U8, U16, U32, U64)
// ============================================================================
// U8
IMPL_BINARY_OP(u8_t, u8, add, x + y)
IMPL_BINARY_OP(u8_t, u8, sub, (x > y) ? (x - y) : 0)
IMPL_BINARY_OP(u8_t, u8, mul, x *y)
IMPL_BINARY_OP(u8_t, u8, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(u8_t, u8, pow, (u8_t)ipow_u64(x, y))
IMPL_BINARY_OP(u8_t, u8, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(u8_t, u8, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(u8_t, u8, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(u8_t, u8, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(u8_t, u8, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(u8_t, u8, eq, x == y)
IMPL_BINARY_TO_BOOL(u8_t, u8, ne, x != y)
IMPL_BINARY_TO_BOOL(u8_t, u8, lt, x < y)
IMPL_BINARY_TO_BOOL(u8_t, u8, le, x <= y)
IMPL_BINARY_TO_BOOL(u8_t, u8, gt, x > y)
IMPL_BINARY_TO_BOOL(u8_t, u8, ge, x >= y)
// U16
IMPL_BINARY_OP(u16_t, u16, add, x + y)
IMPL_BINARY_OP(u16_t, u16, sub, (x > y) ? (x - y) : 0)
IMPL_BINARY_OP(u16_t, u16, mul, x *y)
IMPL_BINARY_OP(u16_t, u16, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(u16_t, u16, pow, (u16_t)ipow_u64(x, y))
IMPL_BINARY_OP(u16_t, u16, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(u16_t, u16, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(u16_t, u16, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(u16_t, u16, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(u16_t, u16, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(u16_t, u16, eq, x == y)
IMPL_BINARY_TO_BOOL(u16_t, u16, ne, x != y)
IMPL_BINARY_TO_BOOL(u16_t, u16, lt, x < y)
IMPL_BINARY_TO_BOOL(u16_t, u16, le, x <= y)
IMPL_BINARY_TO_BOOL(u16_t, u16, gt, x > y)
IMPL_BINARY_TO_BOOL(u16_t, u16, ge, x >= y)
// U32
IMPL_BINARY_OP(u32_t, u32, add, x + y)
IMPL_BINARY_OP(u32_t, u32, sub, (x > y) ? (x - y) : 0)
IMPL_BINARY_OP(u32_t, u32, mul, x *y)
IMPL_BINARY_OP(u32_t, u32, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(u32_t, u32, pow, (u32_t)ipow_u64(x, y))
IMPL_BINARY_OP(u32_t, u32, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(u32_t, u32, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(u32_t, u32, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(u32_t, u32, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(u32_t, u32, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(u32_t, u32, eq, x == y)
IMPL_BINARY_TO_BOOL(u32_t, u32, ne, x != y)
IMPL_BINARY_TO_BOOL(u32_t, u32, lt, x < y)
IMPL_BINARY_TO_BOOL(u32_t, u32, le, x <= y)
IMPL_BINARY_TO_BOOL(u32_t, u32, gt, x > y)
IMPL_BINARY_TO_BOOL(u32_t, u32, ge, x >= y)
// U64
IMPL_BINARY_OP(u64_t, u64, add, x + y)
IMPL_BINARY_OP(u64_t, u64, sub, (x > y) ? (x - y) : 0)
IMPL_BINARY_OP(u64_t, u64, mul, x *y)
IMPL_BINARY_OP(u64_t, u64, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(u64_t, u64, pow, ipow_u64(x, y))
IMPL_BINARY_OP(u64_t, u64, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(u64_t, u64, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(u64_t, u64, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(u64_t, u64, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(u64_t, u64, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(u64_t, u64, eq, x == y)
IMPL_BINARY_TO_BOOL(u64_t, u64, ne, x != y)
IMPL_BINARY_TO_BOOL(u64_t, u64, lt, x < y)
IMPL_BINARY_TO_BOOL(u64_t, u64, le, x <= y)
IMPL_BINARY_TO_BOOL(u64_t, u64, gt, x > y)
IMPL_BINARY_TO_BOOL(u64_t, u64, ge, x >= y)
// ============================================================================
// SIGNED INTEGER OPERATIONS (I8, I16, I32, I64)
// ============================================================================
// I8
IMPL_BINARY_OP(i8_t, i8, add, x + y)
IMPL_BINARY_OP(i8_t, i8, sub, x - y)
IMPL_BINARY_OP(i8_t, i8, mul, x *y)
IMPL_BINARY_OP(i8_t, i8, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(i8_t, i8, pow, (i8_t)ipow_i64(x, y))
IMPL_BINARY_OP(i8_t, i8, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(i8_t, i8, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(i8_t, i8, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(i8_t, i8, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(i8_t, i8, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(i8_t, i8, eq, x == y)
IMPL_BINARY_TO_BOOL(i8_t, i8, ne, x != y)
IMPL_BINARY_TO_BOOL(i8_t, i8, lt, x < y)
IMPL_BINARY_TO_BOOL(i8_t, i8, le, x <= y)
IMPL_BINARY_TO_BOOL(i8_t, i8, gt, x > y)
IMPL_BINARY_TO_BOOL(i8_t, i8, ge, x >= y)
// I16
IMPL_BINARY_OP(i16_t, i16, add, x + y)
IMPL_BINARY_OP(i16_t, i16, sub, x - y)
IMPL_BINARY_OP(i16_t, i16, mul, x *y)
IMPL_BINARY_OP(i16_t, i16, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(i16_t, i16, pow, (i16_t)ipow_i64(x, y))
IMPL_BINARY_OP(i16_t, i16, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(i16_t, i16, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(i16_t, i16, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(i16_t, i16, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(i16_t, i16, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(i16_t, i16, eq, x == y)
IMPL_BINARY_TO_BOOL(i16_t, i16, ne, x != y)
IMPL_BINARY_TO_BOOL(i16_t, i16, lt, x < y)
IMPL_BINARY_TO_BOOL(i16_t, i16, le, x <= y)
IMPL_BINARY_TO_BOOL(i16_t, i16, gt, x > y)
IMPL_BINARY_TO_BOOL(i16_t, i16, ge, x >= y)
// I32
IMPL_BINARY_OP(i32_t, i32, add, x + y)
IMPL_BINARY_OP(i32_t, i32, sub, x - y)
IMPL_BINARY_OP(i32_t, i32, mul, x *y)
IMPL_BINARY_OP(i32_t, i32, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(i32_t, i32, pow, (i32_t)ipow_i64(x, y))
IMPL_BINARY_OP(i32_t, i32, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(i32_t, i32, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(i32_t, i32, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(i32_t, i32, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(i32_t, i32, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(i32_t, i32, eq, x == y)
IMPL_BINARY_TO_BOOL(i32_t, i32, ne, x != y)
IMPL_BINARY_TO_BOOL(i32_t, i32, lt, x < y)
IMPL_BINARY_TO_BOOL(i32_t, i32, le, x <= y)
IMPL_BINARY_TO_BOOL(i32_t, i32, gt, x > y)
IMPL_BINARY_TO_BOOL(i32_t, i32, ge, x >= y)
// I64
IMPL_BINARY_OP(i64_t, i64, add, x + y)
IMPL_BINARY_OP(i64_t, i64, sub, x - y)
IMPL_BINARY_OP(i64_t, i64, mul, x *y)
IMPL_BINARY_OP(i64_t, i64, div, (y != 0) ? (x / y) : 0)
IMPL_BINARY_OP(i64_t, i64, pow, ipow_i64(x, y))
IMPL_BINARY_OP(i64_t, i64, maximum, MAXIMUM(x, y))
IMPL_BINARY_OP(i64_t, i64, minimum, MINIMUM(x, y))
IMPL_BINARY_TO_BOOL(i64_t, i64, logical_and, (x != 0 && y != 0))
IMPL_BINARY_TO_BOOL(i64_t, i64, logical_or, (x != 0 || y != 0))
IMPL_BINARY_TO_BOOL(i64_t, i64, logical_xor, (x != 0) != (y != 0))
IMPL_BINARY_TO_BOOL(i64_t, i64, eq, x == y)
IMPL_BINARY_TO_BOOL(i64_t, i64, ne, x != y)
IMPL_BINARY_TO_BOOL(i64_t, i64, lt, x < y)
IMPL_BINARY_TO_BOOL(i64_t, i64, le, x <= y)
IMPL_BINARY_TO_BOOL(i64_t, i64, gt, x > y)
IMPL_BINARY_TO_BOOL(i64_t, i64, ge, x >= y)