forked from ggerganov/llama.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathggml.c
13380 lines (10881 loc) · 417 KB
/
ggml.c
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
// Defines CLOCK_MONOTONIC on Linux
#define _GNU_SOURCE
#include "ggml.h"
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <malloc.h> // using malloc.h with MSC/MINGW
#elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
#include <alloca.h>
#endif
#include <assert.h>
#include <errno.h>
#include <time.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <inttypes.h>
#include <stdio.h>
#include <float.h>
#include <limits.h>
// if C99 - static_assert is noop
// ref: https://stackoverflow.com/a/53923785/4039976
#ifndef static_assert
#define static_assert(cond, msg) struct global_scope_noop_trick
#endif
#if defined(_WIN32)
#include <windows.h>
typedef volatile LONG atomic_int;
typedef atomic_int atomic_bool;
static void atomic_store(atomic_int* ptr, LONG val) {
InterlockedExchange(ptr, val);
}
static LONG atomic_load(atomic_int* ptr) {
return InterlockedCompareExchange(ptr, 0, 0);
}
static LONG atomic_fetch_add(atomic_int* ptr, LONG inc) {
return InterlockedExchangeAdd(ptr, inc);
}
static LONG atomic_fetch_sub(atomic_int* ptr, LONG dec) {
return atomic_fetch_add(ptr, -(dec));
}
typedef HANDLE pthread_t;
typedef DWORD thread_ret_t;
static int pthread_create(pthread_t* out, void* unused, thread_ret_t(*func)(void*), void* arg) {
(void) unused;
HANDLE handle = CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE) func, arg, 0, NULL);
if (handle == NULL)
{
return EAGAIN;
}
*out = handle;
return 0;
}
static int pthread_join(pthread_t thread, void* unused) {
(void) unused;
return (int) WaitForSingleObject(thread, INFINITE);
}
static int sched_yield (void) {
Sleep (0);
return 0;
}
#else
#include <pthread.h>
#include <stdatomic.h>
typedef void* thread_ret_t;
#endif
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
#define __FMA__
#endif
#ifndef __F16C__
#define __F16C__
#endif
#ifndef __SSE3__
#define __SSE3__
#endif
#endif
#ifdef __HAIKU__
#define static_assert(cond, msg) _Static_assert(cond, msg)
#endif
/*#define GGML_PERF*/
#define GGML_DEBUG 0
#define GGML_GELU_FP16
#define GGML_SILU_FP16
#define GGML_SOFT_MAX_UNROLL 4
#define GGML_VEC_DOT_UNROLL 2
#ifdef GGML_USE_ACCELERATE
// uncomment to use vDSP for soft max computation
// note: not sure if it is actually faster
//#define GGML_SOFT_MAX_ACCELERATE
#endif
#if UINTPTR_MAX == 0xFFFFFFFF
#define GGML_MEM_ALIGN 4
#else
#define GGML_MEM_ALIGN 16
#endif
#if defined(_MSC_VER) || defined(__MINGW32__)
#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr)
#else
inline static void* ggml_aligned_malloc(size_t size) {
void* aligned_memory = NULL;
int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size);
if (result != 0) {
// Handle allocation failure
return NULL;
}
return aligned_memory;
}
#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size)
#define GGML_ALIGNED_FREE(ptr) free(ptr)
#endif
#define UNUSED(x) (void)(x)
#define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
#define GGML_ASSERT(x) \
do { \
if (!(x)) { \
fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
abort(); \
} \
} while (0)
#if defined(GGML_USE_ACCELERATE)
#include <Accelerate/Accelerate.h>
#elif defined(GGML_USE_OPENBLAS)
#include <cblas.h>
#elif defined(GGML_USE_CUBLAS)
#include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h"
#endif
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
// floating point type used to accumulate sums
typedef double ggml_float;
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
#ifdef __ARM_NEON
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
#define GGML_FP16_TO_FP32(x) ((float) (x))
#define GGML_FP32_TO_FP16(x) (x)
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#include <immintrin.h>
#endif
#endif
#ifdef __F16C__
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
register float f;
register double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // __F16C__
#endif // __ARM_NEON
//
// global data
//
// precomputed gelu table for f16 (128 KB)
static ggml_fp16_t table_gelu_f16[1 << 16];
// precomputed silu table for f16 (128 KB)
static ggml_fp16_t table_silu_f16[1 << 16];
// precomputed exp table for f16 (128 KB)
static ggml_fp16_t table_exp_f16[1 << 16];
// precomputed f32 table for f16 (256 KB)
static float table_f32_f16[1 << 16];
#if defined(__ARM_NEON)
#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s
#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
#define B8(c,s ) B7(c,s, c), B7(c,s, s)
// precomputed tables for expanding 8bits to 8 bytes (shl 4)
static const uint64_t table_b2b_u[1 << 8] = { B8(00, 10) };
#endif
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return table_f32_f16[s];
}
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
// note: do not use these inside ggml.c
// these are meant to be used via the ggml.h API
float ggml_fp16_to_fp32(ggml_fp16_t x) {
return (float) GGML_FP16_TO_FP32(x);
}
ggml_fp16_t ggml_fp32_to_fp16(float x) {
return GGML_FP32_TO_FP16(x);
}
//
// timing
//
#if defined(_MSC_VER) || defined(__MINGW32__)
static int64_t timer_freq;
void ggml_time_init(void) {
LARGE_INTEGER frequency;
QueryPerformanceFrequency(&frequency);
timer_freq = frequency.QuadPart;
}
int64_t ggml_time_ms(void) {
LARGE_INTEGER t;
QueryPerformanceCounter(&t);
return (t.QuadPart * 1000) / timer_freq;
}
int64_t ggml_time_us(void) {
LARGE_INTEGER t;
QueryPerformanceCounter(&t);
return (t.QuadPart * 1000000) / timer_freq;
}
#else
void ggml_time_init(void) {}
int64_t ggml_time_ms(void) {
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return (int64_t)ts.tv_sec*1000 + (int64_t)ts.tv_nsec/1000000;
}
int64_t ggml_time_us(void) {
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return (int64_t)ts.tv_sec*1000000 + (int64_t)ts.tv_nsec/1000;
}
#endif
int64_t ggml_cycles(void) {
return clock();
}
int64_t ggml_cycles_per_ms(void) {
return CLOCKS_PER_SEC/1000;
}
#ifdef GGML_PERF
#define ggml_perf_time_ms() ggml_time_ms()
#define ggml_perf_time_us() ggml_time_us()
#define ggml_perf_cycles() ggml_cycles()
#define ggml_perf_cycles_per_ms() ggml_cycles_per_ms()
#else
#define ggml_perf_time_ms() 0
#define ggml_perf_time_us() 0
#define ggml_perf_cycles() 0
#define ggml_perf_cycles_per_ms() 0
#endif
//
// cache line
//
#if defined(__cpp_lib_hardware_interference_size)
#define CACHE_LINE_SIZE hardware_destructive_interference_size
#else
#if defined(__POWER9_VECTOR__)
#define CACHE_LINE_SIZE 128
#else
#define CACHE_LINE_SIZE 64
#endif
#endif
static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
//
// quantization
//
#if __AVX__ || __AVX2__ || __AVX512F__
// Unpack 16 4-bit fields into 16 bytes
// The output vector contains 16 bytes, each one in [ 0 .. 15 ] interval
static inline __m128i bytes_from_nibbles_16(const uint8_t * rsi)
{
// Load 8 bytes from memory
__m128i tmp = _mm_loadl_epi64( ( const __m128i* )rsi );
// Expand bytes into uint16_t values
__m128i bytes = _mm_cvtepu8_epi16( tmp );
// Unpack values into individual bytes
const __m128i lowMask = _mm_set1_epi8( 0xF );
__m128i high = _mm_andnot_si128( lowMask, bytes );
__m128i low = _mm_and_si128( lowMask, bytes );
high = _mm_slli_epi16( high, 4 );
bytes = _mm_or_si128( low, high );
return bytes;
}
// horizontally add 8 floats
static inline float hsum_float_8(const __m256 x) {
__m128 res = _mm256_extractf128_ps(x, 1);
res = _mm_add_ps(res, _mm256_castps256_ps128(x));
res = _mm_add_ps(res, _mm_movehl_ps(res, res));
res = _mm_add_ss(res, _mm_movehdup_ps(res));
return _mm_cvtss_f32(res);
}
// horizontally add 8 int32_t
static inline int hsum_i32_8(const __m256i a) {
const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
const __m128i sum64 = _mm_add_epi32(hi64, sum128);
const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
}
// horizontally add 4 int32_t
static inline int hsum_i32_4(const __m128i a) {
const __m128i hi64 = _mm_unpackhi_epi64(a, a);
const __m128i sum64 = _mm_add_epi32(hi64, a);
const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
}
#if __AVX2__ || __AVX512F__
// spread 32 bits to 32 bytes { 0x00, 0xFF }
static inline __m256i bytes_from_bits_32(const uint8_t * x) {
uint32_t x32;
memcpy(&x32, x, sizeof(uint32_t));
const __m256i shuf_mask = _mm256_set_epi64x(
0x0303030303030303, 0x0202020202020202,
0x0101010101010101, 0x0000000000000000);
__m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
bytes = _mm256_or_si256(bytes, bit_mask);
return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
}
// Unpack 32 4-bit fields into 32 bytes
// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
{
// Load 16 bytes from memory
__m128i tmp = _mm_loadu_si128( ( const __m128i* )rsi );
// Expand bytes into uint16_t values
__m256i bytes = _mm256_cvtepu8_epi16( tmp );
// Unpack values into individual bytes
const __m256i lowMask = _mm256_set1_epi8( 0xF );
__m256i high = _mm256_andnot_si256( lowMask, bytes );
__m256i low = _mm256_and_si256( lowMask, bytes );
high = _mm256_slli_epi16( high, 4 );
bytes = _mm256_or_si256( low, high );
return bytes;
}
// add int16_t pairwise and return as float vector
static inline __m256 sum_i16_pairs_float(const __m256i x) {
const __m256i ones = _mm256_set1_epi16(1);
const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
return _mm256_cvtepi32_ps(summed_pairs);
}
// multiply int8_t, add results pairwise twice and return as float vector
static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
// Get absolute values of x vectors
const __m256i ax = _mm256_sign_epi8(x, x);
// Sign the values of the y vectors
const __m256i sy = _mm256_sign_epi8(y, x);
#if __AVXVNNI__
const __m256i zero = _mm256_setzero_si256();
const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
return _mm256_cvtepi32_ps(summed_pairs);
#else
// Perform multiplication and create 16-bit values
const __m256i dot = _mm256_maddubs_epi16(ax, sy);
return sum_i16_pairs_float(dot);
#endif
}
static inline __m128i packNibbles( __m256i bytes )
{
// Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
#if __AVX512F__
const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000
bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh
return _mm256_cvtepi16_epi8(bytes); // abcd_efgh
#else
const __m256i lowByte = _mm256_set1_epi16( 0xFF );
__m256i high = _mm256_andnot_si256( lowByte, bytes );
__m256i low = _mm256_and_si256( lowByte, bytes );
high = _mm256_srli_epi16( high, 4 );
bytes = _mm256_or_si256( low, high );
// Compress uint16_t lanes into bytes
__m128i r0 = _mm256_castsi256_si128( bytes );
__m128i r1 = _mm256_extracti128_si256( bytes, 1 );
return _mm_packus_epi16( r0, r1 );
#endif
}
#else
static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
{
// Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
const __m128i lowByte = _mm_set1_epi16( 0xFF );
__m128i high = _mm_andnot_si128( lowByte, bytes1 );
__m128i low = _mm_and_si128( lowByte, bytes1 );
high = _mm_srli_epi16( high, 4 );
bytes1 = _mm_or_si128( low, high );
high = _mm_andnot_si128( lowByte, bytes2 );
low = _mm_and_si128( lowByte, bytes2 );
high = _mm_srli_epi16( high, 4 );
bytes2 = _mm_or_si128( low, high );
return _mm_packus_epi16( bytes1, bytes2);
}
#endif
#endif // __AVX__ || __AVX2__ || __AVX512F__
#if __ARM_NEON
#if !defined(__aarch64__)
inline static uint16_t vaddvq_u8(uint8x16_t v) {
return
(uint16_t)vgetq_lane_u8(v, 0) + (uint16_t)vgetq_lane_u8(v, 1) +
(uint16_t)vgetq_lane_u8(v, 2) + (uint16_t)vgetq_lane_u8(v, 3) +
(uint16_t)vgetq_lane_u8(v, 4) + (uint16_t)vgetq_lane_u8(v, 5) +
(uint16_t)vgetq_lane_u8(v, 6) + (uint16_t)vgetq_lane_u8(v, 7) +
(uint16_t)vgetq_lane_u8(v, 8) + (uint16_t)vgetq_lane_u8(v, 9) +
(uint16_t)vgetq_lane_u8(v, 10) + (uint16_t)vgetq_lane_u8(v, 11) +
(uint16_t)vgetq_lane_u8(v, 12) + (uint16_t)vgetq_lane_u8(v, 13) +
(uint16_t)vgetq_lane_u8(v, 14) + (uint16_t)vgetq_lane_u8(v, 15);
}
inline static int16_t vaddvq_s8(int8x16_t v) {
return
(int16_t)vgetq_lane_s8(v, 0) + (int16_t)vgetq_lane_s8(v, 1) +
(int16_t)vgetq_lane_s8(v, 2) + (int16_t)vgetq_lane_s8(v, 3) +
(int16_t)vgetq_lane_s8(v, 4) + (int16_t)vgetq_lane_s8(v, 5) +
(int16_t)vgetq_lane_s8(v, 6) + (int16_t)vgetq_lane_s8(v, 7) +
(int16_t)vgetq_lane_s8(v, 8) + (int16_t)vgetq_lane_s8(v, 9) +
(int16_t)vgetq_lane_s8(v, 10) + (int16_t)vgetq_lane_s8(v, 11) +
(int16_t)vgetq_lane_s8(v, 12) + (int16_t)vgetq_lane_s8(v, 13) +
(int16_t)vgetq_lane_s8(v, 14) + (int16_t)vgetq_lane_s8(v, 15);
}
inline static int32_t vaddvq_s16(int16x8_t v) {
return
(int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
(int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
(int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
(int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
}
inline static uint32_t vaddvq_u16(uint16x8_t v) {
return
(uint32_t)vgetq_lane_u16(v, 0) + (uint32_t)vgetq_lane_u16(v, 1) +
(uint32_t)vgetq_lane_u16(v, 2) + (uint32_t)vgetq_lane_u16(v, 3) +
(uint32_t)vgetq_lane_u16(v, 4) + (uint32_t)vgetq_lane_u16(v, 5) +
(uint32_t)vgetq_lane_u16(v, 6) + (uint32_t)vgetq_lane_u16(v, 7);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
float vminvq_f32(float32x4_t v) {
return
MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) {
return vget_low_s8(vcombine_s8(a, b));
}
int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) {
return vget_high_s8(vcombine_s8(a, b));
}
uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
return vget_low_u8(vcombine_u8(a, b));
}
uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
return vget_high_u8(vcombine_u8(a, b));
}
#endif
#endif
#define QK4_0 32
typedef struct {
float d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
typedef struct {
float d; // delta
float m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK4_2 16
typedef struct {
ggml_fp16_t d; // delta
uint8_t qs[QK4_2 / 2]; // nibbles / quants
} block_q4_2;
static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
#define QK4_3 16
typedef struct {
ggml_fp16_t d; // delta
ggml_fp16_t m; // min
uint8_t qs[QK4_3 / 2]; // nibbles / quants
} block_q4_3;
static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding");
#define QK5_0 32
typedef struct {
ggml_fp16_t d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
typedef struct {
ggml_fp16_t d; // delta
ggml_fp16_t m; // min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
typedef struct {
float d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
typedef struct {
float d; // delta
float s0; // d * sum(qs[i]) low
float s1; // d * sum(qs[i]) high
int8_t qs[QK8_1]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 3*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
// reference implementation for deterministic creation of model files
static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) {
assert(k % QK4_0 == 0);
const int nb = k / QK4_0;
uint8_t pp[QK4_0/2];
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
float max = 0.0f;
for (int l = 0; l < QK4_0; l++) {
const float v = x[i*QK4_0 + l];
if (amax < fabsf(v)) {
amax = fabsf(v);
max = v;
}
}
const float d = max / -8;
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
for (int l = 0; l < QK4_0; l += 2) {
const float v0 = x[i*QK4_0 + l + 0]*id;
const float v1 = x[i*QK4_0 + l + 1]*id;
const uint8_t vi0 = MIN(15, (int8_t)roundf(v0) + 8);
const uint8_t vi1 = MIN(15, (int8_t)roundf(v1) + 8);
assert(vi0 < 16);
assert(vi1 < 16);
pp[l/2] = vi0 | (vi1 << 4);
}
memcpy(y[i].qs, pp, sizeof(pp));
}
}
static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int k) {
assert(k % QK4_0 == 0);
const int nb = k / QK4_0;
block_q4_0 * restrict y = vy;
#if defined(__POWER9_VECTOR__)
const vector float v85 = vec_splats(8.5f);
const vector signed int v15 = vec_splats(15);
for (int i = 0; i < nb; i++) {
float max = 0.0f;
float min = 0.0f;
vector float srcv [8];
vector float maxv[8];
vector float minv[8];
for (int l = 0; l < 8; l++) srcv[l] = *(vector float *)(x + i*32 + 4*l);
//for (int l = 0; l < 8; l++) asrcv[l] = vec_abs(srcv[l]);
for (int l = 0; l < 4; l++) maxv[2*l] = vec_max(asrcv[2*l], asrcv[2*l+1]);
//for (int l = 0; l < 2; l++) maxv[4*l] = vec_max(maxv[4*l], maxv[4*l+2]);
maxv[0] = vec_max(maxv[0], maxv[2]);
maxv[4] = vec_max(maxv[4], maxv[6]);
//for (int l = 0; l < 1; l++) maxv[8*l] = vec_max(maxv[8*l], maxv[8*l+4]);
maxv[0] = vec_max(maxv[0], maxv[4]);
for (int l = 0; l < 4; l++) minv[2*l] = vec_min(asrcv[2*l], asrcv[2*l+1]);
//for (int l = 0; l < 2; l++) minv[4*l] = vec_min(minv[4*l], minv[4*l+2]);
minv[0] = vec_min(minv[0], minv[2]);
minv[4] = vec_min(minv[4], minv[6]);
//for (int l = 0; l < 1; l++) minv[8*l] = vec_min(minv[8*l], minv[8*l+4]);
minv[0] = vec_min(minv[0], minv[4]);
max = MAX(
MAX(vec_extract(maxv[0], 0), vec_extract(maxv[0], 1)),
MAX(vec_extract(maxv[0], 2), vec_extract(maxv[0], 3)));
min = MIN(
MIN(vec_extract(minv[0], 0), vec_extract(minv[0], 1)),
MIN(vec_extract(minv[0], 2), vec_extract(minv[0], 3)));
const float magnitude = max >= fabsf(min) ? max : min;
const float d = magnitude / -8;
const float id = d ? 1.0/d : 0.0;
y[i].d = d;
const vector float vid = vec_splats(id);
uint8_t * restrict pb = y[i].qs;
for (int l = 0; l < 8; l++) {
const vector float vf = vec_madd(srcv[l], vid, v85);
const vector signed int vi = vec_signed(vf);
const vector signed int vc = vec_min(vi, v15);
pb[2*l + 0] = vec_extract(vc, 0) | (vec_extract(vc, 1) << 4);
pb[2*l + 1] = vec_extract(vc, 2) | (vec_extract(vc, 3) << 4);
}
}
#elif __ARM_NEON
for (int i = 0; i < nb; i++) {
float32x4_t srcv [8];
float32x4_t maxv[8];
float32x4_t minv[8];
for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l);
for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l+1]);
for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l+2]);
for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l+4]);
for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l+1]);
for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l+2]);
for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l+4]);
const float max = vmaxvq_f32(maxv[0]);
const float min = vminvq_f32(minv[0]);
const float magnitude = max >= fabsf(min) ? max : min;
const float d = magnitude / -8;
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
for (int l = 0; l < 8; l++) {
const float32x4_t v = vmulq_n_f32(srcv[l], id);
const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(8.5f));
const int32x4_t vi = vcvtq_s32_f32(vf);
const int32x4_t vc = vminq_s32(vi, vdupq_n_s32(15));
y[i].qs[2*l + 0] = vgetq_lane_s32(vc, 0) | (vgetq_lane_s32(vc, 1) << 4);
y[i].qs[2*l + 1] = vgetq_lane_s32(vc, 2) | (vgetq_lane_s32(vc, 3) << 4);
}
}
#elif defined(__AVX2__)
for (int i = 0; i < nb; i++) {
// Load elements into 4 AVX vectors
__m256 v0 = _mm256_loadu_ps( x );
__m256 v1 = _mm256_loadu_ps( x + 8 );
__m256 v2 = _mm256_loadu_ps( x + 16 );
__m256 v3 = _mm256_loadu_ps( x + 24 );
x += 32;
// Compute max for the block
__m256 max = _mm256_max_ps( v0, v1 );
__m256 maxTmp = _mm256_max_ps( v2, v3 );
max = _mm256_max_ps( max, maxTmp );
__m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) );
max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
const float maxScalar = _mm_cvtss_f32( max4 );
// Compute min for the block
__m256 min = _mm256_min_ps( v0, v1 );
__m256 minTmp = _mm256_min_ps( v2, v3 );
min = _mm256_min_ps( min, minTmp );
__m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) );
min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) );
min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) );
const float minScalar = _mm_cvtss_f32( min4 );
// Quantize these floats
const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar;
const float d = magnitude / -8.0f;
y[i].d = d;
const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
// Apply the multiplier
v0 = _mm256_mul_ps( v0, mul );
v1 = _mm256_mul_ps( v1, mul );
v2 = _mm256_mul_ps( v2, mul );
v3 = _mm256_mul_ps( v3, mul );
// Round to nearest integer
v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
// Convert floats to integers
__m256i i0 = _mm256_cvtps_epi32( v0 );
__m256i i1 = _mm256_cvtps_epi32( v1 );
__m256i i2 = _mm256_cvtps_epi32( v2 );
__m256i i3 = _mm256_cvtps_epi32( v3 );
// Convert int32 to int16
i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31
// Convert int16 to int8
i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
// We got our precious signed bytes, but the order is now wrong
// These AVX2 pack instructions process 16-byte pieces independently
// The following instruction is fixing the order
const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
i0 = _mm256_permutevar8x32_epi32( i0, perm );
// Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ]
const __m256i off = _mm256_set1_epi8( 8 );
i0 = _mm256_add_epi8( i0, off );
const __m256i maxNibble = _mm256_set1_epi8( 15 );
i0 = _mm256_min_epi8( i0, maxNibble );
// Compress the vector into 4 bit/value, and store
__m128i res = packNibbles( i0 );
_mm_storeu_si128( ( __m128i* )y[i].qs, res );
}
#elif defined(__AVX__)
for (int i = 0; i < nb; i++) {
// Load elements into 4 AVX vectors
__m256 v0 = _mm256_loadu_ps( x );
__m256 v1 = _mm256_loadu_ps( x + 8 );
__m256 v2 = _mm256_loadu_ps( x + 16 );
__m256 v3 = _mm256_loadu_ps( x + 24 );
x += 32;
// Compute max for the block
__m256 max = _mm256_max_ps( v0, v1 );
__m256 maxTmp = _mm256_max_ps( v2, v3 );
max = _mm256_max_ps( max, maxTmp );
__m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) );
max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
const float maxScalar = _mm_cvtss_f32( max4 );
// Compute min for the block
__m256 min = _mm256_min_ps( v0, v1 );
__m256 minTmp = _mm256_min_ps( v2, v3 );
min = _mm256_min_ps( min, minTmp );
__m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) );
min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) );
min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) );
const float minScalar = _mm_cvtss_f32( min4 );
// Quantize these floats
const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar;
const float d = magnitude / -8.0f;
y[i].d = d;
const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
// Apply the multiplier
v0 = _mm256_mul_ps( v0, mul );
v1 = _mm256_mul_ps( v1, mul );
v2 = _mm256_mul_ps( v2, mul );
v3 = _mm256_mul_ps( v3, mul );
// Round to nearest integer
v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
// Convert floats to integers
__m256i i0 = _mm256_cvtps_epi32( v0 );
__m256i i1 = _mm256_cvtps_epi32( v1 );
__m256i i2 = _mm256_cvtps_epi32( v2 );