-
Notifications
You must be signed in to change notification settings - Fork 11.5k
/
Copy pathggml-sycl.cpp
4304 lines (3675 loc) · 164 KB
/
ggml-sycl.cpp
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
//
// MIT license
// Copyright (C) 2024 Intel Corporation
// SPDX-License-Identifier: MIT
//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
#include <algorithm>
#include <assert.h>
#include <atomic>
#include <cinttypes>
#include <cstddef>
#include <cstdint>
#include <cstdlib>
#include <float.h>
#include <limits>
#include <stdint.h>
#include <stdio.h>
#include <vector>
#include <cmath>
#include <iostream>
#include <fstream>
#include <stdio.h>
#include <stdlib.h>
#include <regex>
#include <sycl/sycl.hpp>
#include <sycl/half_type.hpp>
#include "ggml-sycl.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-sycl/backend.hpp"
#include "ggml-sycl/common.hpp"
#include "ggml-sycl/presets.hpp"
#include "ggml-sycl/gemm.hpp"
#include "ggml-sycl/sycl_hw.hpp"
#include "ggml-sycl/getrows.hpp"
#include "ggml.h"
static bool g_sycl_loaded = false;
int g_ggml_sycl_debug = 0;
int g_ggml_sycl_disable_optimize = 0;
int g_ggml_sycl_disable_graph = 0;
static ggml_sycl_device_info ggml_sycl_init() {
ggml_sycl_device_info info = {};
info.device_count = dpct::dev_mgr::instance().device_count();
if (info.device_count == 0) {
GGML_LOG_ERROR("%s: failed to initialize: %s\n", GGML_SYCL_NAME, __func__);
return info;
}
GGML_ASSERT(info.device_count <= GGML_SYCL_MAX_DEVICES);
int64_t total_vram = 0;
/* This is a bit misleading; reserved for later */
// #if defined(SYCL_USE_XMX)
// GGML_LOG_INFO("%s: SYCL_USE_XMX: yes\n", __func__);
// #else
// GGML_LOG_INFO("%s: SYCL_USE_XMX: no\n", __func__);
// #endif
for (int i = 0; i < info.device_count; ++i) {
info.devices[i].vmm = 0;
dpct::device_info prop;
sycl::device device = dpct::dev_mgr::instance().get_device(i);
SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
prop, device)));
info.default_tensor_split[i] = total_vram;
total_vram += prop.get_global_mem_size();
info.devices[i].cc =
100 * prop.get_major_version() + 10 * prop.get_minor_version();
info.devices[i].hw_info = get_device_hw_info(&device);
info.devices[i].opt_feature = check_gpu_optimize_feature(info.devices[i].hw_info.arch);
info.max_work_group_sizes[i] = prop.get_max_work_group_size();
}
for (int id = 0; id < info.device_count; ++id) {
info.default_tensor_split[id] /= total_vram;
}
return info;
}
const ggml_sycl_device_info & ggml_sycl_info() {
static ggml_sycl_device_info info = ggml_sycl_init();
return info;
}
static void print_device_detail(int id, sycl::device &device, std::string device_type) {
dpct::device_info prop;
SYCL_CHECK(CHECK_TRY_ERROR(
dpct::get_device_info(prop, device)));
std::string version;
version += std::to_string(prop.get_major_version());
version += ".";
version += std::to_string(prop.get_minor_version());
device_type = std::regex_replace(device_type, std::regex("ext_oneapi_"), "");
std::string name = std::string(prop.get_name());
name = std::regex_replace(name, std::regex("\\(R\\)"), "");
name = std::regex_replace(name, std::regex("\\(TM\\)"), "");
auto global_mem_size = prop.get_global_mem_size()/1000000;
GGML_LOG_INFO("|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(),
name.c_str(), version.c_str(), prop.get_max_compute_units(),
prop.get_max_work_group_size(), prop.get_max_sub_group_size(),
global_mem_size, device.get_info<sycl::info::device::driver_version>().c_str());
}
static void print_device_opt_feature(int device_count) {
GGML_LOG_INFO("SYCL Optimization Feature:\n");
GGML_LOG_INFO(
"|ID| Device Type|Reorder|\n");
GGML_LOG_INFO(
"|--|-------------------|-------|\n");
std::map<std::string, size_t> DeviceNums;
for (int id = 0; id < device_count; ++id) {
sycl::device device = dpct::dev_mgr::instance().get_device(id);
std::string backend_type = get_device_backend_and_type(device);
int type_id = DeviceNums[backend_type]++;
std::stringstream device_type;
device_type << "[" << backend_type << ":" << std::to_string(type_id)
<< "]";
std::string device_type_s = device_type.str();
device_type_s = std::regex_replace(device_type_s, std::regex("ext_oneapi_"), "");
GGML_LOG_INFO("|%2d|%19s|%7s|\n", id, device_type_s.c_str(),
ggml_sycl_info().devices[id].opt_feature.reorder ? "Y": "N");
}
}
void ggml_backend_sycl_print_sycl_devices() {
GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_print_sycl_devices\n");
int device_count = dpct::dev_mgr::instance().device_count();
std::map<std::string, size_t> DeviceNums;
GGML_LOG_INFO("Found %d SYCL devices:\n", device_count);
GGML_LOG_INFO(
"| | | | "
" |Max | |Max |Global | |\n");
GGML_LOG_INFO(
"| | | | "
" |compute|Max work|sub |mem | |\n");
GGML_LOG_INFO(
"|ID| Device Type| "
"Name|Version|units |group |group|size | Driver version|\n");
GGML_LOG_INFO(
"|--|-------------------|---------------------------------------|------"
"-|-------|--------|-----|-------|---------------------|\n");
for (int id = 0; id < device_count; ++id) {
sycl::device device = dpct::dev_mgr::instance().get_device(id);
std::string backend_type = get_device_backend_and_type(device);
int type_id = DeviceNums[backend_type]++;
std::stringstream device_type;
device_type << "[" << backend_type << ":" << std::to_string(type_id)
<< "]";
print_device_detail(id, device, device_type.str());
}
print_device_opt_feature(device_count);
}
static inline int get_sycl_env(const char *env_name, int default_val) {
char *user_device_string = getenv(env_name);
int user_number = default_val;
unsigned n;
if (user_device_string != NULL &&
sscanf(user_device_string, " %u", &n) == 1) {
user_number = (int)n;
} else {
user_number = default_val;
}
return user_number;
}
static void ggml_check_sycl() try {
static bool initialized = false;
if (!initialized) {
g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
GGML_LOG_INFO("Running with Environment Variables:\n");
GGML_LOG_INFO(" GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
GGML_LOG_INFO(" GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize);
GGML_LOG_INFO(" GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph);
GGML_LOG_INFO("Build with Macros:\n");
#if defined(GGML_SYCL_FORCE_MMQ)
GGML_LOG_INFO(" GGML_SYCL_FORCE_MMQ: yes\n");
#else
GGML_LOG_INFO(" GGML_SYCL_FORCE_MMQ: no\n");
#endif
#if defined(GGML_SYCL_F16)
GGML_LOG_INFO(" GGML_SYCL_F16: yes\n");
#else
GGML_LOG_INFO(" GGML_SYCL_F16: no\n");
#endif
/* NOT REMOVE, keep it for next optimize for XMX.
#if defined(SYCL_USE_XMX)
fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__);
#else
fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
#endif
*/
if (CHECK_TRY_ERROR(g_all_sycl_device_count =
dpct::dev_mgr::instance().device_count()) != 0) {
initialized = true;
g_sycl_loaded = false;
return;
}
GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
initialized = true;
g_sycl_loaded = true;
ggml_backend_sycl_print_sycl_devices();
}
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
/*
device_index: device index from 0 to n (continue numbers).
It is used for device select/set in SYCL backend internal data structure.
*/
inline void check_allow_gpu_index(const int device_index) {
if (device_index >= ggml_sycl_info().device_count) {
char error_buf[256];
snprintf(
error_buf,
sizeof(error_buf),
"%s error: device_index:%d is out of range: [0-%d]",
__func__,
device_index,
ggml_sycl_info().device_count - 1);
GGML_LOG_ERROR("%s\n", error_buf);
assert(false);
}
}
GGML_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len) try {
GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_gpu_list\n");
for(int i=0;i<max_len;i++) id_list[i] = -1;
for (int i=0;i< ggml_sycl_info().device_count;i++){
if (i>=max_len) break;
id_list[i] = i;
}
return;
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
// sycl buffer
struct ggml_backend_sycl_buffer_context {
int device;
void * dev_ptr = nullptr;
queue_ptr stream;
std::string name;
optimize_feature opt_feature;
std::vector<ggml_tensor_extra_gpu *> tensor_extras;
ggml_backend_sycl_buffer_context(int device, void * dev_ptr, queue_ptr stream) :
device(device), dev_ptr(dev_ptr), stream(stream) {
check_allow_gpu_index(device);
name = (GGML_SYCL_NAME + std::to_string(device));
opt_feature = ggml_sycl_info().devices[device].opt_feature;
}
~ggml_backend_sycl_buffer_context() {
if (dev_ptr != nullptr) {
ggml_sycl_set_device(device);
SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(dev_ptr, *stream)));
}
//release extra used by tensors
for (ggml_tensor_extra_gpu * extra : tensor_extras) {
release_extra_gpu(extra);
}
}
};
static const char * ggml_backend_sycl_buffer_type_get_name(ggml_backend_buffer_type_t buft);
static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) {
return buffer->buft->iface.get_name == ggml_backend_sycl_buffer_type_get_name;
}
static void
ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
ggml_sycl_set_device(ctx->device);
delete ctx;
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) {
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
return ctx->dev_ptr;
}
static enum ggml_status
ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor) try {
ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
if (tensor->view_src != NULL) {
assert(tensor->view_src->buffer->buft == buffer->buft);
return GGML_STATUS_SUCCESS;
}
if (tensor->type == GGML_TYPE_Q4_0) {
ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
tensor->extra = extra;
ctx->tensor_extras.push_back(extra); //used to release it when destroy ctx.
}
if (ggml_is_quantized(tensor->type)) {
// initialize padding to 0 to avoid possible NaN values
size_t original_size = ggml_nbytes(tensor);
size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
if (padded_size > original_size && tensor->view_src == nullptr) {
SYCL_CHECK(CHECK_TRY_ERROR(ctx->stream->memset(
(char *)tensor->data + original_size, 0,
padded_size - original_size).wait()));
}
}
return GGML_STATUS_SUCCESS;
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor,
const void *data, size_t offset,
size_t size) try {
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
ggml_sycl_set_device(ctx->device);
auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue());
SYCL_CHECK(
CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw()));
// Note: Use host buffer to save the data from mmap(), then copy to device. It's workaround for mmap() issue on PVC GPU.
// This function will be called during load model from disk. Use memory buffer replace dynamic won't save more time and brings potential memory leak risk here.
char* host_buf = (char*)malloc(size);
memcpy(host_buf, data, size);
SYCL_CHECK(
CHECK_TRY_ERROR((*stream).memcpy((char *)tensor->data + offset, host_buf, size)
.wait()));
free(host_buf);
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor *tensor,
void *data, size_t offset,
size_t size) try {
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
ggml_sycl_set_device(ctx->device);
auto stream = dpct::dev_mgr::instance().get_device(ctx->device).default_queue();
SYCL_CHECK(CHECK_TRY_ERROR(
stream.memcpy(data, (const char *)tensor->data + offset, size)
.wait()));
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst,
const void *ptr_src, size_t size) {
char *host_buf = (char *)malloc(size);
q_src.memcpy(host_buf, (const char *)ptr_src, size).wait();
q_dst.memcpy((char *)ptr_dst, host_buf, size).wait();
free(host_buf);
}
static bool
ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor *src,
ggml_tensor *dst) try {
if (ggml_backend_buffer_is_sycl(src->buffer)) {
ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context;
ggml_backend_sycl_buffer_context * dst_ctx = (ggml_backend_sycl_buffer_context *)dst->buffer->context;
ggml_sycl_set_device(src_ctx->device);
/*
DPCT1009:198: SYCL uses exceptions to report errors and does not use the
error codes. The original code was commented out and a warning string
was inserted. You need to rewrite this code.
*/
SYCL_CHECK(CHECK_TRY_ERROR(
dpct::dev_mgr::instance().get_device(src_ctx->device).queues_wait_and_throw()));
ggml_sycl_set_device(dst_ctx->device);
/*
DPCT1009:199: SYCL uses exceptions to report errors and does not use the
error codes. The original code was commented out and a warning string
was inserted. You need to rewrite this code.
*/
SYCL_CHECK(CHECK_TRY_ERROR(
dpct::dev_mgr::instance().get_device(dst_ctx->device).queues_wait_and_throw()));
/*
DPCT1009:200: SYCL uses exceptions to report errors and does not use the
error codes. The original code was commented out and a warning string
was inserted. You need to rewrite this code.
*/
queue_ptr stream_dst = dst_ctx->stream;
queue_ptr stream_src = src_ctx->stream;
size_t size = ggml_nbytes(src);
//todo. it's dirty solutino to walkaroud known issue:device2device cross GPUs.
dev2dev_memcpy(*stream_dst, *stream_src, dst->data, src->data, size);
//todo, it's known issue：error in device2device cross GPUs. reused when the issue is fixed. DON"T remove
#if 0
SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(
(char *)dst->data, (const char *)src->data, size).wait()));
/*
DPCT1009:201: SYCL uses exceptions to report errors and does not use the
error codes. The original code was commented out and a warning string
was inserted. You need to rewrite this code.
*/
SYCL_CHECK(CHECK_TRY_ERROR(
dpct::dev_mgr::instance().get_device(dst_ctx->device).queues_wait_and_throw()));
#endif
return true;
}
return false;
GGML_UNUSED(buffer);
} catch (const sycl::exception & exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer,
uint8_t value) try {
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
ggml_sycl_set_device(ctx->device);
queue_ptr stream = ctx->stream;
SYCL_CHECK(
CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw()));
SYCL_CHECK(CHECK_TRY_ERROR((*stream)
.memset(ctx->dev_ptr, value, buffer->size)
.wait()));
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value,
size_t offset, size_t size) {
GGML_SYCL_DEBUG(" [SYCL] call %s\n", __func__);
ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
SYCL_CHECK(ggml_sycl_set_device(ctx->device));
auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue());
if (size == 0) {
return; // Nothing to do
}
if (tensor->data == nullptr) {
GGML_ABORT("Error: Tensor data pointer is null.\n");
}
void * target_ptr = static_cast<char *>(tensor->data) + offset;
SYCL_CHECK(CHECK_TRY_ERROR((*stream).memset(target_ptr, value, size)));
SYCL_CHECK(CHECK_TRY_ERROR((*stream).wait()));
}
static void ggml_backend_sycl_buffer_reset(ggml_backend_buffer_t buffer) {
GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__);
if (buffer == nullptr) {
return;
}
ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
if (ctx != nullptr) {
for (ggml_tensor_extra_gpu * extra : ctx->tensor_extras) {
release_extra_gpu(extra);
}
ctx->tensor_extras.clear(); // reset the tensor_extras vector
}
}
static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = {
/* .free_buffer = */ ggml_backend_sycl_buffer_free_buffer,
/* .get_base = */ ggml_backend_sycl_buffer_get_base,
/* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor,
/* .memset_tensor = */ ggml_backend_sycl_buffer_memset_tensor,
/* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_sycl_buffer_cpy_tensor,
/* .clear = */ ggml_backend_sycl_buffer_clear,
/* .reset = */ ggml_backend_sycl_buffer_reset,
};
// sycl buffer type
struct ggml_backend_sycl_buffer_type_context {
int device;
std::string name;
// each buffer type has its own stream
queue_ptr stream = nullptr;
};
static const char * ggml_backend_sycl_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
return ctx->name.c_str();
}
static ggml_backend_buffer_t
ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
size_t size) try {
ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
ggml_sycl_set_device(buft_ctx->device);
const queue_ptr stream = buft_ctx->stream;
size = std::max(size, (size_t)1); // syclMalloc returns null for size 0
void * dev_ptr;
SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device(
size, *stream)));
if (!dev_ptr) {
GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
return nullptr;
}
ggml_backend_sycl_buffer_context * ctx = new ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr, buft_ctx->stream);
return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size);
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 128;
GGML_UNUSED(buft);
}
static size_t ggml_backend_sycl_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
return dpct::get_current_device().get_max_mem_alloc_size();
GGML_UNUSED(buft);
}
static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
size_t size = ggml_nbytes(tensor);
int64_t ne0 = tensor->ne[0];
if (ggml_is_quantized(tensor->type)) {
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
}
return size;
GGML_UNUSED(buft);
}
static const ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = {
/* .get_name = */ ggml_backend_sycl_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_sycl_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_sycl_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_sycl_buffer_type_get_max_size,
/* .get_alloc_size = */ ggml_backend_sycl_buffer_type_get_alloc_size,
/* .is_host = */ NULL,
};
ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
auto dev_count = ggml_backend_sycl_get_device_count();
if (device>=dev_count or device<0) {
GGML_LOG_ERROR("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
device, dev_count-1);
GGML_ASSERT(device<dev_count);
}
static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES];
static bool ggml_backend_sycl_buffer_type_initialized = false;
if (!ggml_backend_sycl_buffer_type_initialized) {
for (int i = 0; i < dev_count; i++) {
auto & device_i = dpct::dev_mgr::instance().get_device(i);
queue_ptr stream = &(device_i.default_queue());
ggml_backend_sycl_buffer_types[i] = {
/* .iface = */ ggml_backend_sycl_buffer_type_interface,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_sycl_reg(), i),
/* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), stream},
};
}
ggml_backend_sycl_buffer_type_initialized = true;
}
return &ggml_backend_sycl_buffer_types[device];
}
static ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_context * ctx) {
GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n");
int device = ctx->device;
if (device>=ggml_sycl_info().device_count or device<0) {
GGML_LOG_ERROR("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
device, ggml_sycl_info().device_count-1);
GGML_ASSERT(device<ggml_sycl_info().device_count);
}
static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES];
static bool ggml_backend_sycl_buffer_type_initialized = false;
if (!ggml_backend_sycl_buffer_type_initialized) {
for (int i = 0; i < ggml_sycl_info().device_count; i++) {
ggml_backend_sycl_buffer_types[i] = {
/* .iface = */ ggml_backend_sycl_buffer_type_interface,
/* .device = */ nullptr,
/* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), ctx->stream(i, 0)},
};
}
ggml_backend_sycl_buffer_type_initialized = true;
}
return &ggml_backend_sycl_buffer_types[device];
}
// sycl split buffer
static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_SYCL_MAX_DEVICES> & tensor_split) {
int64_t min_compute_capability = INT_MAX;
int64_t max_compute_capability = INT_MIN;
for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
if (tensor_split[i] < (i + 1 < ggml_sycl_info().device_count ? tensor_split[i + 1] : 1.0f)) {
if (min_compute_capability > ggml_sycl_info().devices[i].cc) {
min_compute_capability = ggml_sycl_info().devices[i].cc;
}
if (max_compute_capability < ggml_sycl_info().devices[i].cc) {
max_compute_capability = ggml_sycl_info().devices[i].cc;
}
}
}
switch(type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
return max_compute_capability >= VER_GEN9 ? 128 : 64;
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
return 64;
case GGML_TYPE_F16:
case GGML_TYPE_F32:
return 1;
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ4_NL:
return max_compute_capability >= VER_GEN9 ? 128 : 64;
case GGML_TYPE_IQ3_S:
return max_compute_capability >= VER_GEN9 ? 128 : 64;
case GGML_TYPE_Q6_K:
return 64;
default:
GGML_ABORT("fatal error");
}
}
static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_SYCL_MAX_DEVICES> & tensor_split, int id) {
const int64_t nrows = ggml_nrows(tensor);
const int64_t rounding = get_row_rounding(tensor->type, tensor_split);
*row_low = id == 0 ? 0 : nrows*tensor_split[id];
*row_low -= *row_low % rounding;
if (id == ggml_sycl_info().device_count - 1) {
*row_high = nrows;
} else {
*row_high = nrows*tensor_split[id + 1];
*row_high -= *row_high % rounding;
}
}
static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
}
struct ggml_backend_sycl_split_buffer_type_context {
std::array<float, GGML_SYCL_MAX_DEVICES> tensor_split;
};
struct ggml_backend_sycl_split_buffer_context {
~ggml_backend_sycl_split_buffer_context() try {
for (ggml_tensor_extra_gpu * extra : tensor_extras) {
release_extra_gpu(extra, streams);
}
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
std::vector<ggml_tensor_extra_gpu *> tensor_extras;
std::vector<queue_ptr> streams;
};
static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
delete ctx;
}
static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
// the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
return (void *)0x1000;
GGML_UNUSED(buffer);
}
static enum ggml_status
ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor) try {
GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context;
const int64_t ne0 = tensor->ne[0];
ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
ctx->tensor_extras.push_back(extra);
ctx->streams.push_back(&(dpct::get_current_device().default_queue()));
for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
int64_t row_low, row_high;
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i);
int64_t nrows_split = row_high - row_low;
if (nrows_split == 0) {
continue;
}
size_t size = ggml_nbytes_split(tensor, nrows_split);
const size_t original_size = size;
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
// FIXME: do not crash if SYCL Buffer alloc fails
// currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
ggml_sycl_set_device(i);
const queue_ptr stream = ctx->streams[i];
char * buf;
/*
DPCT1009:208: SYCL uses exceptions to report errors and does not use the
error codes. The original code was commented out and a warning string
was inserted. You need to rewrite this code.
*/
SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)sycl::malloc_device(
size, *stream)));
if (!buf) {
char err_buf[1024];
snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
throw std::runtime_error(err_buf);
}
// set padding to 0 to avoid possible NaN values
if (size > original_size) {
/*
DPCT1009:209: SYCL uses exceptions to report errors and does not use
the error codes. The original code was commented out and a warning
string was inserted. You need to rewrite this code.
*/
SYCL_CHECK(CHECK_TRY_ERROR(
(*stream)
.memset(buf + original_size, 0, size - original_size)
.wait()));
}
extra->data_device[i] = buf;
for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) {
/*
DPCT1009:210: SYCL uses exceptions to report errors and does not use
the error codes. The original code was commented out and a warning
string was inserted. You need to rewrite this code.
*/
SYCL_CHECK(
CHECK_TRY_ERROR(extra->events[i][is] = new sycl::event()));
}
}
tensor->extra = extra;
return GGML_STATUS_SUCCESS;
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void
ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor, const void *data,
size_t offset, size_t size) try {
// split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context;
const int64_t ne0 = tensor->ne[0];
const size_t nb1 = tensor->nb[1];
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
int64_t row_low, row_high;
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i);
int64_t nrows_split = row_high - row_low;
if (nrows_split == 0) {
continue;
}
const size_t offset_split = row_low*nb1;
size_t size = ggml_nbytes_split(tensor, nrows_split);
const size_t original_size = size;
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
const char * buf_host = (const char *)data + offset_split;
/*
DPCT1009:211: SYCL uses exceptions to report errors and does not use the
error codes. The original code was commented out and a warning string
was inserted. You need to rewrite this code.
*/
ggml_sycl_set_device(i);
const queue_ptr stream = ctx->streams[i];
SYCL_CHECK(CHECK_TRY_ERROR(
(*stream)
.memcpy(extra->data_device[i], buf_host, original_size)
.wait()));
}
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void
ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor *tensor, void *data,
size_t offset, size_t size) try {
// split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context;
const int64_t ne0 = tensor->ne[0];
const size_t nb1 = tensor->nb[1];
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
int64_t row_low, row_high;
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i);
int64_t nrows_split = row_high - row_low;
if (nrows_split == 0) {
continue;
}
const size_t offset_split = row_low*nb1;
size_t size = ggml_nbytes_split(tensor, nrows_split);
const size_t original_size = size;
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
char * buf_host = (char *)data + offset_split;
/*
DPCT1009:212: SYCL uses exceptions to report errors and does not use the
error codes. The original code was commented out and a warning string
was inserted. You need to rewrite this code.
*/
ggml_sycl_set_device(i);
const queue_ptr stream = ctx->streams[i];
SYCL_CHECK(CHECK_TRY_ERROR(
(*stream)
.memcpy(buf_host, extra->data_device[i], original_size)
.wait()));
}
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
GGML_UNUSED(buffer);
GGML_UNUSED(value);
}
static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = {
/* .free_buffer = */ ggml_backend_sycl_split_buffer_free_buffer,
/* .get_base = */ ggml_backend_sycl_split_buffer_get_base,
/* .init_tensor = */ ggml_backend_sycl_split_buffer_init_tensor,
/* .memset_tensor = */ NULL,
/* .set_tensor = */ ggml_backend_sycl_split_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_sycl_split_buffer_get_tensor,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_sycl_split_buffer_clear,
/* .reset = */ NULL,
};
// sycl split buffer type
static const char * ggml_backend_sycl_split_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return GGML_SYCL_NAME "_Split";
GGML_UNUSED(buft);
}
static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) {
return buffer->buft->iface.get_name == ggml_backend_sycl_split_buffer_type_get_name;
}
static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
// since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
// instead, we allocate them for each tensor separately in init_tensor
// however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
// as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct.
ggml_backend_sycl_split_buffer_context * ctx = new ggml_backend_sycl_split_buffer_context();
return ggml_backend_buffer_init(buft, ggml_backend_sycl_split_buffer_interface, ctx, size);
}