-
Notifications
You must be signed in to change notification settings - Fork 984
/
ggml-backend.c
2245 lines (1853 loc) · 83.3 KB
/
ggml-backend.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
#include "ggml-backend-impl.h"
#include "ggml-alloc.h"
#include "ggml-impl.h"
#include <assert.h>
#include <limits.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#define MAX(a, b) ((a) > (b) ? (a) : (b))
// backend buffer type
const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
return buft->iface.get_name(buft);
}
GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
return buft->iface.alloc_buffer(buft, size);
}
size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
return buft->iface.get_alignment(buft);
}
size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
// get_max_size is optional, defaults to SIZE_MAX
if (buft->iface.get_max_size) {
return buft->iface.get_max_size(buft);
}
return SIZE_MAX;
}
GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
// get_alloc_size is optional, defaults to ggml_nbytes
if (buft->iface.get_alloc_size) {
size_t size = buft->iface.get_alloc_size(buft, tensor);
assert(size >= ggml_nbytes(tensor));
return size;
}
return ggml_nbytes(tensor);
}
bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
if (buft->iface.is_host) {
return buft->iface.is_host(buft);
}
return false;
}
// backend buffer
GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
ggml_backend_buffer_type_t buft,
struct ggml_backend_buffer_i iface,
ggml_backend_buffer_context_t context,
size_t size) {
ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
(*buffer) = (struct ggml_backend_buffer) {
/* .interface = */ iface,
/* .buft = */ buft,
/* .context = */ context,
/* .size = */ size,
/* .usage = */ GGML_BACKEND_BUFFER_USAGE_ANY
};
return buffer;
}
const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name(buffer);
}
void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
if (buffer == NULL) {
return;
}
if (buffer->iface.free_buffer != NULL) {
buffer->iface.free_buffer(buffer);
}
free(buffer);
}
size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
return buffer->size;
}
void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
void * base = buffer->iface.get_base(buffer);
GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
return base;
}
GGML_CALL void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
// init_tensor is optional
if (buffer->iface.init_tensor) {
buffer->iface.init_tensor(buffer, tensor);
}
}
size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
}
size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
}
size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
}
void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
buffer->iface.clear(buffer, value);
}
bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
}
void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
buffer->usage = usage;
// FIXME: add a generic callback to the buffer interface
if (ggml_backend_buffer_is_multi_buffer(buffer)) {
ggml_backend_multi_buffer_set_usage(buffer, usage);
}
}
enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage(ggml_backend_buffer_t buffer) {
return buffer->usage;
}
ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
return buffer->buft;
}
void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
if (buffer->iface.reset) {
buffer->iface.reset(buffer);
}
}
bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
if (dst_buf->iface.cpy_tensor) {
return dst_buf->iface.cpy_tensor(dst_buf, src, dst);
}
return false;
}
// backend
ggml_guid_t ggml_backend_guid(ggml_backend_t backend) {
if (backend == NULL) {
return NULL;
}
return backend->guid;
}
const char * ggml_backend_name(ggml_backend_t backend) {
if (backend == NULL) {
return "NULL";
}
return backend->iface.get_name(backend);
}
void ggml_backend_free(ggml_backend_t backend) {
if (backend == NULL) {
return;
}
backend->iface.free(backend);
}
ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
return backend->iface.get_default_buffer_type(backend);
}
ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size);
}
size_t ggml_backend_get_alignment(ggml_backend_t backend) {
return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
}
size_t ggml_backend_get_max_size(ggml_backend_t backend) {
return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
}
void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
if (backend->iface.set_tensor_async == NULL) {
ggml_backend_tensor_set(tensor, data, offset, size);
} else {
backend->iface.set_tensor_async(backend, tensor, data, offset, size);
}
}
void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
if (backend->iface.get_tensor_async == NULL) {
ggml_backend_tensor_get(tensor, data, offset, size);
} else {
backend->iface.get_tensor_async(backend, tensor, data, offset, size);
}
}
GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(buf != NULL && "tensor buffer not set");
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
if (!size) {
return;
}
buf->iface.set_tensor(buf, tensor, data, offset, size);
}
GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(buf != NULL && "tensor buffer not set");
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
if (!size) {
return;
}
buf->iface.get_tensor(buf, tensor, data, offset, size);
}
void ggml_backend_synchronize(ggml_backend_t backend) {
if (backend->iface.synchronize == NULL) {
return;
}
backend->iface.synchronize(backend);
}
ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
GGML_ASSERT(backend->iface.graph_plan_create != NULL);
return backend->iface.graph_plan_create(backend, cgraph);
}
void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
GGML_ASSERT(backend->iface.graph_plan_free != NULL);
backend->iface.graph_plan_free(backend, plan);
}
enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
return backend->iface.graph_plan_compute(backend, plan);
}
enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
ggml_backend_synchronize(backend);
return err;
}
enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
return backend->iface.graph_compute(backend, cgraph);
}
bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
return backend->iface.supports_op(backend, op);
}
bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
return backend->iface.supports_buft(backend, buft);
}
bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
if (backend->iface.offload_op != NULL) {
return backend->iface.offload_op(backend, op);
}
return false;
}
// backend copy
static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
if (a->type != b->type) {
return false;
}
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (a->ne[i] != b->ne[i]) {
return false;
}
if (a->nb[i] != b->nb[i]) {
return false;
}
}
return true;
}
void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
if (src == dst) {
return;
}
if (ggml_backend_buffer_is_host(src->buffer)) {
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
} else if (ggml_backend_buffer_is_host(dst->buffer)) {
ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
} else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
#ifndef NDEBUG
fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
#endif
size_t nbytes = ggml_nbytes(src);
void * data = malloc(nbytes);
ggml_backend_tensor_get(src, data, 0, nbytes);
ggml_backend_tensor_set(dst, data, 0, nbytes);
free(data);
}
}
void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
if (src == dst) {
return;
}
if (backend_dst->iface.cpy_tensor_async != NULL) {
if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
return;
}
}
// an async copy would normally happen after all the queued operations on both backends are completed
// to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
ggml_backend_synchronize(backend_src);
ggml_backend_synchronize(backend_dst);
ggml_backend_tensor_copy(src, dst);
}
// events
ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) {
if (backend->iface.event_new == NULL) {
return NULL;
}
return backend->iface.event_new(backend);
}
void ggml_backend_event_free(ggml_backend_event_t event) {
if (event == NULL) {
return;
}
event->backend->iface.event_free(event);
}
void ggml_backend_event_record(ggml_backend_event_t event) {
GGML_ASSERT(event->backend->iface.event_record != NULL);
event->backend->iface.event_record(event);
}
void ggml_backend_event_synchronize(ggml_backend_event_t event) {
GGML_ASSERT(event->backend->iface.event_synchronize != NULL);
event->backend->iface.event_synchronize(event);
}
void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
GGML_ASSERT(backend->iface.event_wait != NULL);
backend->iface.event_wait(backend, event);
}
// backend registry
#define GGML_REG_MAX_BACKENDS 64
struct ggml_backend_reg {
char name[128];
ggml_backend_init_fn init_fn;
ggml_backend_buffer_type_t default_buffer_type;
void * user_data;
};
static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS];
static size_t ggml_backend_registry_count = 0;
GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
GGML_CALL static void ggml_backend_registry_init(void) {
static bool initialized = false;
if (initialized) {
return;
}
initialized = true;
ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);
// add forward decls here to avoid including the backend headers
#ifdef GGML_USE_CUDA
extern GGML_CALL void ggml_backend_cuda_reg_devices(void);
ggml_backend_cuda_reg_devices();
#endif
#ifdef GGML_USE_SYCL
extern void ggml_backend_sycl_reg_devices(void);
ggml_backend_sycl_reg_devices();
#endif
#ifdef GGML_USE_METAL
extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
#endif
#ifdef GGML_USE_VULKAN
extern GGML_CALL int ggml_backend_vk_reg_devices(void);
ggml_backend_vk_reg_devices();
#endif
#ifdef GGML_USE_KOMPUTE
extern GGML_CALL void ggml_backend_kompute_reg_devices(void);
ggml_backend_kompute_reg_devices();
#endif
#ifdef GGML_USE_CANN
extern GGML_CALL int ggml_backend_cann_reg_devices(void);
ggml_backend_cann_reg_devices();
#endif
}
GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS);
size_t id = ggml_backend_registry_count;
ggml_backend_registry[id] = (struct ggml_backend_reg) {
/* .name = */ {0},
/* .fn = */ init_fn,
/* .default_buffer_type = */ default_buffer_type,
/* .user_data = */ user_data,
};
snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name);
#ifndef NDEBUG
fprintf(stderr, "%s: registered backend %s\n", __func__, name);
#endif
ggml_backend_registry_count++;
}
size_t ggml_backend_reg_get_count(void) {
ggml_backend_registry_init();
return ggml_backend_registry_count;
}
size_t ggml_backend_reg_find_by_name(const char * name) {
ggml_backend_registry_init();
for (size_t i = 0; i < ggml_backend_registry_count; i++) {
// TODO: case insensitive in a portable way
if (strcmp(ggml_backend_registry[i].name, name) == 0) {
return i;
}
}
// not found
return SIZE_MAX;
}
// init from backend:params string
ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) {
ggml_backend_registry_init();
const char * params = strchr(backend_str, ':');
char backend_name[128];
if (params == NULL) {
snprintf(backend_name, sizeof(backend_name), "%s", backend_str);
params = "";
} else {
snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str);
params++;
}
size_t backend_i = ggml_backend_reg_find_by_name(backend_name);
if (backend_i == SIZE_MAX) {
fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name);
return NULL;
}
return ggml_backend_reg_init_backend(backend_i, params);
}
const char * ggml_backend_reg_get_name(size_t i) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_registry[i].name;
}
ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data);
}
ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_registry[i].default_buffer_type;
}
ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size);
}
// backend CPU
static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
return "CPU";
GGML_UNUSED(buffer);
}
GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
uintptr_t data = (uintptr_t)buffer->context;
// align the buffer
if (data % TENSOR_ALIGNMENT != 0) {
data = GGML_PAD(data, TENSOR_ALIGNMENT);
}
return (void *)data;
}
GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
}
GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
memcpy((char *)tensor->data + offset, data, size);
GGML_UNUSED(buffer);
}
GGML_CALL static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
memcpy(data, (const char *)tensor->data + offset, size);
GGML_UNUSED(buffer);
}
GGML_CALL static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
return true;
}
return false;
GGML_UNUSED(buffer);
}
GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
memset(buffer->context, value, buffer->size);
}
static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
/* .get_name = */ ggml_backend_cpu_buffer_name,
/* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer,
/* .get_base = */ ggml_backend_cpu_buffer_get_base,
/* .init_tensor = */ NULL, // no initialization required
/* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor,
/* .clear = */ ggml_backend_cpu_buffer_clear,
/* .reset = */ NULL,
};
// for buffers from ptr, free is not called
static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
/* .get_name = */ ggml_backend_cpu_buffer_name,
/* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
/* .get_base = */ ggml_backend_cpu_buffer_get_base,
/* .init_tensor = */ NULL, // no initialization required
/* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor,
/* .clear = */ ggml_backend_cpu_buffer_clear,
/* .reset = */ NULL,
};
GGML_CALL static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU";
GGML_UNUSED(buft);
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned
void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
if (data == NULL) {
fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
return NULL;
}
return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
}
GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return true;
GGML_UNUSED(buft);
}
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type;
}
#ifdef GGML_USE_CPU_HBM
// buffer type HBM
#include <hbwmalloc.h>
GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
return "CPU_HBM";
GGML_UNUSED(buf);
}
GGML_CALL static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
//void * ptr = hbw_malloc(size);
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
struct ggml_backend_cpu_context {
int n_threads;
void * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
free(cpu_ctx->work_data);
free(cpu_ctx);
free(backend);
}
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(backend);
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
if (cpu_plan->cplan.work_data == NULL) {
free(cpu_plan);
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return cpu_plan;
}
GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
free(cpu_plan->cplan.work_data);
free(cpu_plan);
GGML_UNUSED(backend);
}
GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
if (cpu_ctx->work_size < cplan.work_size) {
free(cpu_ctx->work_data);
cpu_ctx->work_data = malloc(cplan.work_size);
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}
GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_CPY:
return
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
default:
return true;
}
GGML_UNUSED(backend);
}
GGML_CALL static bool ggml_backend_cpu_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft);
GGML_UNUSED(backend);
}
static struct ggml_backend_i cpu_backend_i = {
/* .get_name = */ ggml_backend_cpu_name,
/* .free = */ ggml_backend_cpu_free,
/* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .supports_op = */ ggml_backend_cpu_supports_op,
/* .supports_buft = */ ggml_backend_cpu_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .event_synchronize = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
if (cpu_backend == NULL) {
free(ctx);
return NULL;
}
*cpu_backend = (struct ggml_backend) {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ cpu_backend_i,
/* .context = */ ctx
};
return cpu_backend;
}
GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
}
GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
return ggml_backend_cpu_init();
GGML_UNUSED(params);
GGML_UNUSED(user_data);
}
// multi-buffer buffer
struct ggml_backend_multi_buffer_context {
ggml_backend_buffer_t * buffers;
size_t n_buffers;
};
typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t;
GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
}
GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
for (size_t i = 0; i < ctx->n_buffers; i++) {
ggml_backend_buffer_free(ctx->buffers[i]);
}
free(ctx->buffers);
free(ctx);
}
GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
for (size_t i = 0; i < ctx->n_buffers; i++) {
ggml_backend_buffer_clear(ctx->buffers[i], value);
}
}
static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) {
static struct ggml_backend_buffer_i multi_backend_buffer_i = {
/* .get_name = */ ggml_backend_multi_buffer_get_name,
/* .free_buffer = */ ggml_backend_multi_buffer_free_buffer,
/* .get_base = */ NULL,
/* .init_tensor = */ NULL,
/* .set_tensor = */ NULL,
/* .get_tensor = */ NULL,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_multi_buffer_clear,
/* .reset = */ NULL,
};
return multi_backend_buffer_i;
}
GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context));
ctx->n_buffers = n_buffers;
ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
GGML_ASSERT(ctx->buffers != NULL);
size_t total_size = 0;
for (size_t i = 0; i < n_buffers; i++) {
ctx->buffers[i] = buffers[i];
total_size += ggml_backend_buffer_get_size(buffers[i]);
}
return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size);
}
GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
}
GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
for (size_t i = 0; i < ctx->n_buffers; i++) {
ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
}
}