forked from ggerganov/llama.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ggml.c
10884 lines (8859 loc) · 329 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 _POSIX_C_SOURCE 199309L
#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 <time.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <stdio.h>
#include <float.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 _MSC_VER || defined(__MINGW32__)
#if !defined(__MINGW32__)
#include <Windows.h>
#else
// ref: https://github.com/ggerganov/whisper.cpp/issues/168
#include <windows.h>
#include <errno.h>
#endif
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) {
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) {
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
#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
#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)
#ifdef GGML_USE_ACCELERATE
#include <Accelerate/Accelerate.h>
#elif GGML_USE_OPENBLAS
#include <cblas.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) (x)
#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
#define GGML_FP16_TO_FP32(x) (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__
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#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];
// 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.
#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 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
//
#define QK 32
// AVX routines provided by GH user Const-me
// ref: https://github.com/ggerganov/ggml/pull/27#issuecomment-1464934600
#if __AVX2__ || __AVX512F__
// Unpack 32 4-bit fields into 32 bytes
// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
static inline __m256i bytesFromNibbles( 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;
}
static inline __m128i packNibbles( __m256i bytes )
{
// Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
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
// method 5
// blocks of QK elements
// represented with a single float (delta) and QK/2 8-bit ints (i.e QK 4-bit signed integer factors)
// reference implementation for deterministic creation of model files
static void quantize_row_q4_0_reference(const float * restrict x, void * restrict y, int k) {
assert(k % QK == 0);
const int nb = k / QK;
const size_t bs = sizeof(float) + QK/2;
uint8_t * restrict pd = ((uint8_t *)y + 0*bs);
uint8_t * restrict pb = ((uint8_t *)y + 0*bs + sizeof(float));
uint8_t pp[QK/2];
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
for (int l = 0; l < QK; l++) {
const float v = x[i*QK + l];
amax = MAX(amax, fabsf(v));
}
const float d = amax / ((1 << 3) - 1);
const float id = d ? 1.0f/d : 0.0f;
*(float *)pd = d;
pd += bs;
for (int l = 0; l < QK; l += 2) {
const float v0 = x[i*QK + l + 0]*id;
const float v1 = x[i*QK + l + 1]*id;
const uint8_t vi0 = ((int8_t) (round(v0))) + 8;
const uint8_t vi1 = ((int8_t) (round(v1))) + 8;
assert(vi0 >= 0 && vi0 < 16);
assert(vi1 >= 0 && vi1 < 16);
pp[l/2] = vi0 | (vi1 << 4);
}
memcpy(pb, pp, sizeof(pp));
pb += bs;
}
}
void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
assert(k % QK == 0);
#if __ARM_NEON || defined(__AVX2__) || defined(__wasm_simd128__)
const int nb = k / QK;
const size_t bs = sizeof(float) + QK/2;
uint8_t * restrict pd = ((uint8_t *)y + 0*bs);
uint8_t * restrict pb = ((uint8_t *)y + 0*bs + sizeof(float));
uint8_t pp[QK/2];
#endif
#if __ARM_NEON
#if QK == 32
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
float32x4_t srcv [8];
float32x4_t asrcv[8];
float32x4_t amaxv[8];
for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l);
for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]);
for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]);
for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]);
for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]);
amax = MAX(
MAX(vgetq_lane_f32(amaxv[0], 0), vgetq_lane_f32(amaxv[0], 1)),
MAX(vgetq_lane_f32(amaxv[0], 2), vgetq_lane_f32(amaxv[0], 3)));
const float d = amax / ((1 << 3) - 1);
const float id = d ? 1.0/d : 0.0;
*(float *)pd = d;
pd += bs;
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);
pp[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4);
pp[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4);
}
memcpy(pb, pp, sizeof(pp));
pb += bs;
}
#else
#error "not implemented for QK"
#endif
#elif defined(__AVX2__)
#if QK == 32
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(abs(e)) for the block
const __m256 signBit = _mm256_set1_ps( -0.0f );
__m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
__m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
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 );
// Quantize these floats
const float d = maxScalar / 7.0f;
*(float *)pd = d;
pd += bs;
const float id = ( maxScalar != 0.0f ) ? 7.0f / maxScalar : 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 to translate the range from [ -7 .. +7 ] into [ +1 .. +15 ]
const __m256i off = _mm256_set1_epi8( 8 );
i0 = _mm256_add_epi8( i0, off );
// Compress the vector into 4 bit/value, and store
__m128i res = packNibbles( i0 );
_mm_storeu_si128( ( __m128i* )pb, res );
pb += bs;
}
#else
#error "not implemented for QK"
#endif
#elif defined(__wasm_simd128__)
#if QK == 32
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
v128_t srcv [8];
v128_t asrcv[8];
v128_t amaxv[8];
for (int l = 0; l < 8; l++) srcv[l] = wasm_v128_load(x + i*32 + 4*l);
for (int l = 0; l < 8; l++) asrcv[l] = wasm_f32x4_abs(srcv[l]);
for (int l = 0; l < 4; l++) amaxv[2*l] = wasm_f32x4_max(asrcv[2*l], asrcv[2*l+1]);
for (int l = 0; l < 2; l++) amaxv[4*l] = wasm_f32x4_max(amaxv[4*l], amaxv[4*l+2]);
for (int l = 0; l < 1; l++) amaxv[8*l] = wasm_f32x4_max(amaxv[8*l], amaxv[8*l+4]);
amax = MAX(
MAX(wasm_f32x4_extract_lane(amaxv[0], 0), wasm_f32x4_extract_lane(amaxv[0], 1)),
MAX(wasm_f32x4_extract_lane(amaxv[0], 2), wasm_f32x4_extract_lane(amaxv[0], 3)));
const float d = amax / ((1 << 3) - 1);
const float id = d ? 1.0/d : 0.0;
*(float *)pd = d;
pd += bs;
for (int l = 0; l < 8; l++) {
const v128_t v = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id));
const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f));
const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf);
pp[2*l + 0] = wasm_i32x4_extract_lane(vi, 0) | (wasm_i32x4_extract_lane(vi, 1) << 4);
pp[2*l + 1] = wasm_i32x4_extract_lane(vi, 2) | (wasm_i32x4_extract_lane(vi, 3) << 4);
}
memcpy(pb, pp, sizeof(pp));
pb += bs;
}
#else
#error "not implemented for QK"
#endif
#else
// scalar
quantize_row_q4_0_reference(x, y, k);
#endif
}
// method 4
// blocks of QK elements
// represented with 2 floats (min + delta) and QK/2 8-bit ints (i.e QK 4-bit unsigned integer factors)
void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) {
assert(k % QK == 0);
const int nb = k / QK;
const size_t bs = 2*sizeof(float) + QK/2;
uint8_t * restrict pd = ((uint8_t *)y + 0*bs);
uint8_t * restrict pm = ((uint8_t *)y + 0*bs + sizeof(float));
uint8_t * restrict pb = ((uint8_t *)y + 0*bs + 2*sizeof(float));
uint8_t pp[QK/2];
for (int i = 0; i < nb; i++) {
float min = FLT_MAX;
float max = -FLT_MAX;
for (int l = 0; l < QK; l++) {
const float v = x[i*QK + l];
if (v < min) min = v;
if (v > max) max = v;
}
const float d = (max - min) / ((1 << 4) - 1);
const float id = d ? 1.0f/d : 0.0f;
*(float *)pm = min;
*(float *)pd = d;
pm += bs;
pd += bs;
for (int l = 0; l < QK; l += 2) {
const float v0 = (x[i*QK + l + 0] - min)*id;
const float v1 = (x[i*QK + l + 1] - min)*id;
const uint8_t vi0 = round(v0);
const uint8_t vi1 = round(v1);
assert(vi0 >= 0 && vi0 < 16);
assert(vi1 >= 0 && vi1 < 16);
pp[l/2] = vi0 | (vi1 << 4);
}
memcpy(pb, pp, sizeof(pp));
pb += bs;
}
}
// TODO: vectorize
void dequantize_row_q4_0(const void * restrict x, float * restrict y, int k) {
assert(k % QK == 0);
const int nb = k / QK;
const size_t bs = sizeof(float) + QK/2;
const uint8_t * restrict pd = ((const uint8_t *)x + 0*bs);
const uint8_t * restrict pb = ((const uint8_t *)x + 0*bs + sizeof(float));
// scalar
for (int i = 0; i < nb; i++) {
const float d = *(const float *) (pd + i*bs);
const uint8_t * restrict pp = pb + i*bs;
for (int l = 0; l < QK; l += 2) {
const uint8_t vi = pp[l/2];
const int8_t vi0 = vi & 0xf;
const int8_t vi1 = vi >> 4;
const float v0 = (vi0 - 8)*d;
const float v1 = (vi1 - 8)*d;
//printf("d = %f, vi = %d, vi0 = %d, vi1 = %d, v0 = %f, v1 = %f\n", d, vi, vi0, vi1, v0, v1);
y[i*QK + l + 0] = v0;
y[i*QK + l + 1] = v1;
assert(!isnan(y[i*QK + l + 0]));
assert(!isnan(y[i*QK + l + 1]));
}
}
}
void dequantize_row_q4_1(const void * restrict x, float * restrict y, int k) {
assert(k % QK == 0);
const int nb = k / QK;
const size_t bs = 2*sizeof(float) + QK/2;
const uint8_t * restrict pd = ((const uint8_t *)x + 0*bs);
const uint8_t * restrict pm = ((const uint8_t *)x + 0*bs + sizeof(float));
const uint8_t * restrict pb = ((const uint8_t *)x + 0*bs + 2*sizeof(float));
for (int i = 0; i < nb; i++) {
const float d = *(const float *) (pd + i*bs);
const float m = *(const float *) (pm + i*bs);
const uint8_t * restrict pp = pb + i*bs;
for (int l = 0; l < QK; l += 2) {
const uint8_t vi = pp[l/2];
const int8_t vi0 = vi & 0xf;
const int8_t vi1 = vi >> 4;
const float v0 = vi0*d + m;
const float v1 = vi1*d + m;
y[i*QK + l + 0] = v0;
y[i*QK + l + 1] = v1;
assert(!isnan(y[i*QK + l + 0]));
assert(!isnan(y[i*QK + l + 1]));
}
}
}
//
// simd mappings
//
// we define a common set of C macros which map to specific intrinsics based on the current architecture
// we then implement the fundamental computation operations below using only these macros
// adding support for new architectures requires to define the corresponding SIMD macros
//
// GGML_F32_STEP / GGML_F16_STEP
// number of elements to process in a single step
//
// GGML_F32_EPR / GGML_F16_EPR
// number of elements to fit in a single register
//
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
#define GGML_SIMD
// F32 NEON
#define GGML_F32_STEP 16
#define GGML_F32_EPR 4
#define GGML_F32x4 float32x4_t
#define GGML_F32x4_ZERO vdupq_n_f32(0.0f)
#define GGML_F32x4_SET1(x) vdupq_n_f32(x)
#define GGML_F32x4_LOAD vld1q_f32
#define GGML_F32x4_STORE vst1q_f32
#define GGML_F32x4_FMA(a, b, c) vfmaq_f32(a, b, c)
#define GGML_F32x4_ADD vaddq_f32
#define GGML_F32x4_MUL vmulq_f32
#if defined(__ARM_FEATURE_QRDMX)
#define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
#else
#define GGML_F32x4_REDUCE_ONE(x) \
(vgetq_lane_f32(x, 0) + \
vgetq_lane_f32(x, 1) + \
vgetq_lane_f32(x, 2) + \
vgetq_lane_f32(x, 3))
#endif
#define GGML_F32x4_REDUCE(res, x) \
{ \
for (int i = 0; i < GGML_F32_ARR/2; ++i) { \
x[2*i] = vaddq_f32(x[2*i], x[2*i+1]); \
} \
for (int i = 0; i < GGML_F32_ARR/4; ++i) { \
x[4*i] = vaddq_f32(x[4*i], x[4*i+2]); \
} \
for (int i = 0; i < GGML_F32_ARR/8; ++i) { \
x[8*i] = vaddq_f32(x[8*i], x[8*i+4]); \
} \
res = GGML_F32x4_REDUCE_ONE(x[0]); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 NEON
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
#define GGML_F16_STEP 32
#define GGML_F16_EPR 8
#define GGML_F16x8 float16x8_t
#define GGML_F16x8_ZERO vdupq_n_f16(0.0f)
#define GGML_F16x8_SET1(x) vdupq_n_f16(x)
#define GGML_F16x8_LOAD vld1q_f16
#define GGML_F16x8_STORE vst1q_f16
#define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
#define GGML_F16x8_ADD vaddq_f16
#define GGML_F16x8_MUL vmulq_f16
#define GGML_F16x8_REDUCE(res, x) \
{ \
for (int i = 0; i < GGML_F16_ARR/2; ++i) { \
x[2*i] = vaddq_f16(x[2*i], x[2*i+1]); \
} \
for (int i = 0; i < GGML_F16_ARR/4; ++i) { \
x[4*i] = vaddq_f16(x[4*i], x[4*i+2]); \
} \
for (int i = 0; i < GGML_F16_ARR/8; ++i) { \
x[8*i] = vaddq_f16(x[8*i], x[8*i+4]); \
} \
const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 (x[0])); \
const float32x4_t t1 = vcvt_f32_f16(vget_high_f16(x[0])); \
res = vaddvq_f32(vaddq_f32(t0, t1)); \
}
#define GGML_F16_VEC GGML_F16x8
#define GGML_F16_VEC_ZERO GGML_F16x8_ZERO
#define GGML_F16_VEC_SET1 GGML_F16x8_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F16x8_FMA
#define GGML_F16_VEC_ADD GGML_F16x8_ADD
#define GGML_F16_VEC_MUL GGML_F16x8_MUL
#define GGML_F16_VEC_REDUCE GGML_F16x8_REDUCE
#else
// if FP16 vector arithmetic is not supported, we use FP32 instead
// and take advantage of the vcvt_ functions to convert to/from FP16
#define GGML_F16_STEP 16
#define GGML_F16_EPR 4
#define GGML_F32Cx4 float32x4_t
#define GGML_F32Cx4_ZERO vdupq_n_f32(0.0f)
#define GGML_F32Cx4_SET1(x) vdupq_n_f32(x)
#define GGML_F32Cx4_LOAD(x) vcvt_f32_f16(vld1_f16(x))
#define GGML_F32Cx4_STORE(x, y) vst1_f16(x, vcvt_f16_f32(y))
#define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
#define GGML_F32Cx4_ADD vaddq_f32
#define GGML_F32Cx4_MUL vmulq_f32
#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
#define GGML_F16_VEC GGML_F32Cx4
#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
#endif
#elif defined(__AVX__)
#define GGML_SIMD
// F32 AVX
#define GGML_F32_STEP 32
#define GGML_F32_EPR 8
#define GGML_F32x8 __m256
#define GGML_F32x8_ZERO _mm256_setzero_ps()
#define GGML_F32x8_SET1(x) _mm256_set1_ps(x)
#define GGML_F32x8_LOAD _mm256_loadu_ps
#define GGML_F32x8_STORE _mm256_storeu_ps
#if defined(__FMA__)
#define GGML_F32x8_FMA(a, b, c) _mm256_fmadd_ps(b, c, a)
#else
#define GGML_F32x8_FMA(a, b, c) _mm256_add_ps(_mm256_mul_ps(b, c), a)
#endif
#define GGML_F32x8_ADD _mm256_add_ps
#define GGML_F32x8_MUL _mm256_mul_ps
#define GGML_F32x8_REDUCE(res, x) \
{ \
for (int i = 0; i < GGML_F32_ARR/2; ++i) { \
x[2*i] = _mm256_add_ps(x[2*i], x[2*i+1]); \
} \
for (int i = 0; i < GGML_F32_ARR/4; ++i) { \
x[4*i] = _mm256_add_ps(x[4*i], x[4*i+2]); \
} \
for (int i = 0; i < GGML_F32_ARR/8; ++i) { \
x[8*i] = _mm256_add_ps(x[8*i], x[8*i+4]); \
} \
const __m128 t0 = _mm_add_ps(_mm256_castps256_ps128(x[0]), \
_mm256_extractf128_ps(x[0], 1)); \
const __m128 t1 = _mm_hadd_ps(t0, t0); \
res = _mm_cvtss_f32(_mm_hadd_ps(t1, t1)); \
}
// TODO: is this optimal ?
#define GGML_F32_VEC GGML_F32x8
#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x8_SET1
#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD
#define GGML_F32_VEC_STORE GGML_F32x8_STORE
#define GGML_F32_VEC_FMA GGML_F32x8_FMA
#define GGML_F32_VEC_ADD GGML_F32x8_ADD
#define GGML_F32_VEC_MUL GGML_F32x8_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE
// F16 AVX
#define GGML_F16_STEP 32
#define GGML_F16_EPR 8
// F16 arithmetic is not supported by AVX, so we use F32 instead
// we take advantage of the _mm256_cvt intrinsics to convert F16 <-> F32
#define GGML_F32Cx8 __m256
#define GGML_F32Cx8_ZERO _mm256_setzero_ps()
#define GGML_F32Cx8_SET1(x) _mm256_set1_ps(x)
#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((__m128i *)(x)))
#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
#define GGML_F32Cx8_FMA GGML_F32x8_FMA
#define GGML_F32Cx8_ADD _mm256_add_ps
#define GGML_F32Cx8_MUL _mm256_mul_ps
#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE
#define GGML_F16_VEC GGML_F32Cx8
#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE
#elif defined(__POWER9_VECTOR__)
#define GGML_SIMD
// F32 POWER9
#define GGML_F32_STEP 32
#define GGML_F32_EPR 4
#define GGML_F32x4 vector float
#define GGML_F32x4_ZERO 0.0f
#define GGML_F32x4_SET1 vec_splats
#define GGML_F32x4_LOAD(p) vec_xl(0, p)
#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p)
#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
#define GGML_F32x4_ADD vec_add
#define GGML_F32x4_MUL vec_mul
#define GGML_F32x4_REDUCE(res, x) \
{ \
for (int i = 0; i < GGML_F32_ARR/2; ++i) { \
x[2*i] = vec_add(x[2*i], x[2*i+1]); \
} \
for (int i = 0; i < GGML_F32_ARR/4; ++i) { \
x[4*i] = vec_add(x[4*i], x[4*i+2]); \
} \
for (int i = 0; i < GGML_F32_ARR/8; ++i) { \
x[8*i] = vec_add(x[8*i], x[8*i+4]); \
} \
res = vec_extract(x[0], 0) + \
vec_extract(x[0], 1) + \
vec_extract(x[0], 2) + \
vec_extract(x[0], 3); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 POWER9
#define GGML_F16_STEP GGML_F32_STEP
#define GGML_F16_EPR GGML_F32_EPR
#define GGML_F16_VEC GGML_F32x4