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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
|
#include "ggml-opencl.h"
#include <array>
#include <atomic>
#include <sstream>
#include <vector>
#include <limits>
#define CL_TARGET_OPENCL_VERSION 110
#include <clblast.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include "ggml.h"
#define CL_DMMV_BLOCK_SIZE 32;
#define MULTILINE_QUOTE(...) #__VA_ARGS__
static std::string program_source = MULTILINE_QUOTE(
typedef char int8_t;
typedef uchar uint8_t;
typedef int int32_t;
typedef uint uint32_t;
struct __attribute__ ((packed)) block_q4_0
{
half d;
uint8_t qs[QK4_0 / 2];
};
struct __attribute__ ((packed)) block_q4_1
{
half d;
half m;
uint8_t qs[QK4_1 / 2];
};
struct __attribute__ ((packed)) block_q5_0
{
half d;
uint32_t qh;
uint8_t qs[QK5_0 / 2];
};
struct __attribute__ ((packed)) block_q5_1
{
half d;
half m;
uint32_t qh;
uint8_t qs[QK5_1 / 2];
};
struct __attribute__ ((packed)) block_q8_0
{
half d;
int8_t qs[QK8_0];
};
__kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
const uint i = get_global_id(0);
y[i] = vload_half(0, &x[i]);
}
void dequantize_q4_0(__global const struct block_q4_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
*v0 = (vi0 - 8)*d;
*v1 = (vi1 - 8)*d;
}
void dequantize_q4_1(__global const struct block_q4_1* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const float m = vload_half(0, &x[ib].m);
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
*v0 = vi0*d + m;
*v1 = vi1*d + m;
}
void dequantize_q5_0(__global const struct block_q5_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
uint32_t qh = x[ib].qh;
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1) - 16;
*v0 = x0*d;
*v1 = x1*d;
}
void dequantize_q5_1(__global const struct block_q5_1* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const float m = vload_half(0, &x[ib].m);
uint32_t qh = x[ib].qh;
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1);
*v0 = x0*d + m;
*v1 = x1*d + m;
}
void dequantize_q8_0(__global const struct block_q8_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const int8_t vi0 = x[ib].qs[iqs + 0];
const int8_t vi1 = x[ib].qs[iqs + 1];
*v0 = vi0*d;
*v1 = vi1*d;
}
void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float* v1){
*v0 = vload_half(0, &x[ib + 0]);
*v1 = vload_half(0, &x[ib + 1]);
}
);
std::string dequant_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0)*2;
if (i >= get_global_size(0)) {
return;
}
const uint qk = QUANT_K;
const uint qr = QUANT_R;
const int ib = i/qk; // block index
const int iqs = (i%qk)/qr; // quant index
const int iybs = i - i%qk; // y block start index
const int y_offset = qr == 1 ? 1 : qk/2;
// dequantize
float v0, v1;
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
y[iybs + iqs + 0] = v0;
y[iybs + iqs + y_offset] = v1;
}
);
std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
const int block_size = get_local_size(0);
const int row = get_global_id(0) / block_size;
const int tid = get_local_id(0);
const uint qk = QUANT_K;
const uint qr = QUANT_R;
const int y_offset = qr == 1 ? 1 : qk/2;
tmp[tid] = 0;
for (int i = 0; i < ncols/block_size; i += 2) {
const int col = i*block_size + 2*tid;
const int ib = (row*ncols + col)/qk; // block index
const int iqs = (col%qk)/qr; // quant index
const int iybs = col - col%qk; // y block start index
// dequantize
float v0, v1;
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
// matrix multiplication
tmp[tid] += v0 * y[iybs + iqs + 0];
tmp[tid] += v1 * y[iybs + iqs + y_offset];
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=block_size/2; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
);
std::string mul_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
if (i >= get_global_size(0)) {
return;
}
dst[dst_offset + i] = x[x_offset + i] * y[y_offset + i%ky];
}
);
#define CL_CHECK(err) \
do { \
cl_int err_ = (err); \
if (err_ != CL_SUCCESS) { \
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
#err, err_, __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
#define CLBLAST_CHECK(err) \
do { \
CLBlastStatusCode err_ = (err); \
if (err_ != CLBlastSuccess) { \
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
#err, err_, __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
std::array<std::string, 5> dequant_str_keys = {
"KERNEL_NAME", "X_TYPE", "QUANT_K", "QUANT_R", "DEQUANT_FUNC"
};
std::array<std::string, 30> dequant_str_values = {
"dequantize_row_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
"dequantize_row_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
"dequantize_row_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
"dequantize_row_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
"dequantize_row_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
"convert_row_f16", "half", "1", "1", "convert_f16"
};
std::array<std::string, 30> dequant_mul_mat_vec_str_values = {
"dequantize_mul_mat_vec_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
"dequantize_mul_mat_vec_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
"dequantize_mul_mat_vec_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
"dequantize_mul_mat_vec_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
"dequantize_mul_mat_vec_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
"convert_mul_mat_vec_f16", "half", "1", "1", "convert_f16"
};
std::array<std::string, 2> mul_str_keys = {
"KERNEL_NAME", "TYPE"
};
std::array<std::string, 2> mul_str_values = {
"mul_f32", "float"
};
std::string& replace(std::string& s, const std::string& from, const std::string& to) {
size_t pos = 0;
while ((pos = s.find(from, pos)) != std::string::npos) {
s.replace(pos, from.length(), to);
pos += to.length();
}
return s;
}
std::string generate_kernels() {
std::stringstream src;
src << program_source << '\n';
for (size_t i = 0; i < dequant_str_values.size(); i += dequant_str_keys.size()) {
std::string dequant_kernel = dequant_template;
std::string dmmv_kernel = dequant_mul_mat_vec_template;
for (size_t j = 0; j < dequant_str_keys.size(); j++) {
replace(dequant_kernel, dequant_str_keys[j], dequant_str_values[i + j]);
replace(dmmv_kernel, dequant_str_keys[j], dequant_mul_mat_vec_str_values[i + j]);
}
src << dequant_kernel << '\n';
src << dmmv_kernel << '\n';
}
for (size_t i = 0; i < mul_str_values.size(); i += mul_str_keys.size()) {
std::string mul_kernel = mul_template;
for (size_t j = 0; j < mul_str_keys.size(); j++) {
replace(mul_kernel, mul_str_keys[j], mul_str_values[i + j]);
}
src << mul_kernel << '\n';
}
return src.str();
}
static cl_platform_id platform;
static cl_device_id device;
static cl_context context;
static cl_command_queue queue;
static cl_program program;
static cl_kernel convert_row_f16_cl;
static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, dequantize_mul_mat_vec_q5_0_cl, dequantize_mul_mat_vec_q5_1_cl, dequantize_mul_mat_vec_q8_0_cl, convert_mul_mat_vec_f16_cl;
static cl_kernel mul_f32_cl;
static bool fp16_support;
static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
cl_program p;
char *program_log;
size_t program_size;
size_t log_size;
int err;
program_size = strlen(program_buffer);
p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
if(err < 0) {
fprintf(stderr, "OpenCL error creating program");
exit(1);
}
const char* compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math "
"-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1";
err = clBuildProgram(p, 0, NULL, compile_opts, NULL, NULL);
if(err < 0) {
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
program_log = (char*) malloc(log_size + 1);
program_log[log_size] = '\0';
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
fprintf(stderr, "ggml_opencl: kernel compile error:\n\n%s\n", program_log);
free(program_log);
exit(1);
}
return p;
}
void ggml_cl_init(void) {
cl_int err;
struct cl_device;
struct cl_platform {
cl_platform_id id;
unsigned number;
char name[128];
char vendor[128];
struct cl_device * devices;
unsigned n_devices;
struct cl_device * default_device;
};
struct cl_device {
struct cl_platform * platform;
cl_device_id id;
unsigned number;
cl_device_type type;
char name[128];
};
enum { NPLAT = 16, NDEV = 16 };
struct cl_platform platforms[NPLAT];
unsigned n_platforms = 0;
struct cl_device devices[NDEV];
unsigned n_devices = 0;
struct cl_device * default_device = NULL;
platform = NULL;
device = NULL;
cl_platform_id platform_ids[NPLAT];
CL_CHECK(clGetPlatformIDs(NPLAT, platform_ids, &n_platforms));
for (unsigned i = 0; i < n_platforms; i++) {
struct cl_platform * p = &platforms[i];
p->number = i;
p->id = platform_ids[i];
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_NAME, sizeof(p->name), &p->name, NULL));
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_VENDOR, sizeof(p->vendor), &p->vendor, NULL));
cl_device_id device_ids[NDEV];
cl_int clGetDeviceIDsError = clGetDeviceIDs(p->id, CL_DEVICE_TYPE_ALL, NDEV, device_ids, &p->n_devices);
if (clGetDeviceIDsError == CL_DEVICE_NOT_FOUND) {
p->n_devices = 0;
} else {
CL_CHECK(clGetDeviceIDsError);
}
p->devices = p->n_devices > 0 ? &devices[n_devices] : NULL;
p->default_device = NULL;
for (unsigned j = 0; j < p->n_devices; j++) {
struct cl_device * d = &devices[n_devices];
d->number = n_devices++;
d->id = device_ids[j];
d->platform = p;
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
p->default_device = d;
}
}
if (default_device == NULL && p->default_device != NULL) {
default_device = p->default_device;
}
}
if (n_devices == 0) {
fprintf(stderr, "ggml_opencl: could find any OpenCL devices.\n");
exit(1);
}
char * user_platform_string = getenv("GGML_OPENCL_PLATFORM");
char * user_device_string = getenv("GGML_OPENCL_DEVICE");
int user_platform_number = -1;
int user_device_number = -1;
unsigned n;
if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
user_platform_number = (int)n;
}
if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1 && n < n_devices) {
user_device_number = (int)n;
}
if (user_platform_number != -1 && user_device_number != -1) {
cl_platform* platform = &platforms[user_platform_number];
if ((unsigned)user_device_number >= platform->n_devices) {
fprintf(stderr, "ggml_opencl: invalid device number %d\n", user_device_number);
exit(1);
}
default_device = &platform->devices[user_device_number];
} else {
struct cl_device * selected_devices = devices;
unsigned n_selected_devices = n_devices;
if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
for (unsigned i = 0; i < n_platforms; i++) {
struct cl_platform * p = &platforms[i];
if (strstr(p->name, user_platform_string) != NULL ||
strstr(p->vendor, user_platform_string) != NULL) {
user_platform_number = (int)i;
break;
}
}
if (user_platform_number == -1) {
fprintf(stderr, "ggml_opencl: no platform matching '%s' was found.\n", user_platform_string);
exit(1);
}
}
if (user_platform_number != -1) {
struct cl_platform * p = &platforms[user_platform_number];
selected_devices = p->devices;
n_selected_devices = p->n_devices;
default_device = p->default_device;
if (n_selected_devices == 0) {
fprintf(stderr, "ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
exit(1);
}
}
if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
for (unsigned i = 0; i < n_selected_devices; i++) {
struct cl_device * d = &selected_devices[i];
if (strstr(d->name, user_device_string) != NULL) {
user_device_number = d->number;
break;
}
}
if (user_device_number == -1) {
fprintf(stderr, "ggml_opencl: no device matching '%s' was found.\n", user_device_string);
exit(1);
}
}
if (user_device_number != -1) {
selected_devices = &devices[user_device_number];
n_selected_devices = 1;
default_device = &selected_devices[0];
}
GGML_ASSERT(n_selected_devices > 0);
if (default_device == NULL) {
default_device = &selected_devices[0];
}
}
fprintf(stderr, "ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
fprintf(stderr, "ggml_opencl: selecting device: '%s'\n", default_device->name);
if (default_device->type != CL_DEVICE_TYPE_GPU) {
fprintf(stderr, "ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
}
platform = default_device->platform->id;
device = default_device->id;
size_t ext_str_size;
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
char *ext_buffer = (char *)alloca(ext_str_size + 1);
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
// Check if ext_buffer contains cl_khr_fp16
fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
cl_context_properties properties[] = {
(intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
};
CL_CHECK((context = clCreateContext(properties, 1, &device, NULL, NULL, &err), err));
CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
(err != CL_INVALID_QUEUE_PROPERTIES && err != CL_INVALID_VALUE ? err :
(queue = clCreateCommandQueue(context, device, 0, &err), err)
)));
const std::string kernel_src = generate_kernels();
program = build_program_from_source(context, device, kernel_src.c_str());
// FP16 to FP32 kernel
CL_CHECK((convert_row_f16_cl = clCreateKernel(program, "convert_row_f16", &err), err));
// Dequantize kernels
CL_CHECK((dequantize_row_q4_0_cl = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
CL_CHECK((dequantize_row_q4_1_cl = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
// dequant mul mat kernel
CL_CHECK((dequantize_mul_mat_vec_q4_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_0", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q4_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_1", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_0", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_1", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q8_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q8_0", &err), err));
CL_CHECK((convert_mul_mat_vec_f16_cl = clCreateKernel(program, "convert_mul_mat_vec_f16", &err), err));
// mul kernel
CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
}
static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_row_q4_0_cl;
case GGML_TYPE_Q4_1:
return &dequantize_row_q4_1_cl;
case GGML_TYPE_Q5_0:
return &dequantize_row_q5_0_cl;
case GGML_TYPE_Q5_1:
return &dequantize_row_q5_1_cl;
case GGML_TYPE_Q8_0:
return &dequantize_row_q8_0_cl;
case GGML_TYPE_F16:
return &convert_row_f16_cl;
default:
return nullptr;
}
}
static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_mul_mat_vec_q4_0_cl;
case GGML_TYPE_Q4_1:
return &dequantize_mul_mat_vec_q4_1_cl;
case GGML_TYPE_Q5_0:
return &dequantize_mul_mat_vec_q5_0_cl;
case GGML_TYPE_Q5_1:
return &dequantize_mul_mat_vec_q5_1_cl;
case GGML_TYPE_Q8_0:
return &dequantize_mul_mat_vec_q8_0_cl;
case GGML_TYPE_F16:
return &convert_mul_mat_vec_f16_cl;
default:
return nullptr;
}
}
// buffer pool for cl
#define MAX_CL_BUFFERS 256
struct scoped_spin_lock {
std::atomic_flag& lock;
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
while (lock.test_and_set(std::memory_order_acquire)) {
; // spin
}
}
~scoped_spin_lock() {
lock.clear(std::memory_order_release);
}
scoped_spin_lock(const scoped_spin_lock&) = delete;
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
};
struct cl_buffer {
cl_mem mem;
size_t size = 0;
};
static cl_buffer g_cl_buffer_pool[MAX_CL_BUFFERS];
static std::atomic_flag g_cl_pool_lock = ATOMIC_FLAG_INIT;
static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size) {
scoped_spin_lock lock(g_cl_pool_lock);
cl_int err;
int best_i = -1;
size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
int worst_i = -1;
size_t worst_size = 0; //largest unused buffer seen so far
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer &b = g_cl_buffer_pool[i];
if (b.size > 0 && b.size >= size && b.size < best_size)
{
best_i = i;
best_size = b.size;
}
if (b.size > 0 && b.size > worst_size)
{
worst_i = i;
worst_size = b.size;
}
}
if(best_i!=-1) //found the smallest buffer that fits our needs
{
cl_buffer& b = g_cl_buffer_pool[best_i];
cl_mem mem = b.mem;
*actual_size = b.size;
b.size = 0;
return mem;
}
if(worst_i!=-1) //no buffer that fits our needs, resize largest one to save memory
{
cl_buffer& b = g_cl_buffer_pool[worst_i];
cl_mem mem = b.mem;
b.size = 0;
clReleaseMemObject(mem);
}
cl_mem mem;
CL_CHECK((mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
*actual_size = size;
return mem;
}
static void ggml_cl_pool_free(cl_mem mem, size_t size) {
scoped_spin_lock lock(g_cl_pool_lock);
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer& b = g_cl_buffer_pool[i];
if (b.size == 0) {
b.mem = mem;
b.size = size;
return;
}
}
fprintf(stderr, "WARNING: cl buffer pool full, increase MAX_CL_BUFFERS\n");
clReleaseMemObject(mem);
}
static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
cl_int err;
const uint64_t ne0 = src->ne[0];
const uint64_t ne1 = src->ne[1];
const uint64_t nb0 = src->nb[0];
const uint64_t nb1 = src->nb[1];
const uint64_t nb2 = src->nb[2];
const uint64_t nb3 = src->nb[3];
const enum ggml_type type = src->type;
const size_t ts = ggml_type_size(type);
const size_t bs = ggml_blck_size(type);
const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
if (nb0 == ts && nb1 == ts*ne0/bs) {
err = clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*nb1, x, 0, NULL, ev);
return err;
}
if (nb0 == ts) {
const size_t buffer_origin[3] = { offset, 0, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { ts*ne0/bs, ne1, 1 };
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts*ne0/bs, 0, nb1, 0, x, 0, NULL, ev);
return err;
}
for (uint64_t i1 = 0; i1 < ne1; i1++) {
// pretend the row is a matrix with cols=1
const size_t buffer_origin[3] = { offset, i1, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { ts/bs, ne0, 1 };
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, 0, 0, nb0, 0, ((const char *)x) + i1*nb0, 0, NULL, ev);
if (err != CL_SUCCESS) {
break;
}
}
return err;
}
static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[2];
const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int64_t nb10 = src1->nb[0];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
size_t x_size;
size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted.
cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
const int i0 = i03*ne02 + i02;
cl_event ev;
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, i0, src0, i03, i02, &ev));
if (nb10 == sizeof(float)) {
// Contiguous, avoid overhead from queueing many kernel runs
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int i1 = i13*ne12*ne11 + i12*ne11;
cl_int x_offset = 0;
cl_int y_offset = i1*ne10;
cl_int d_offset = 0;
size_t global = ne00 * ne01;
cl_int ky = ne10;
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
} else {
for (int64_t i01 = 0; i01 < ne01; i01++) {
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int64_t i11 = i01%ne11;
const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
cl_int x_offset = i01*ne00;
cl_int y_offset = i1*ne10;
cl_int d_offset = i01*ne00;
// compute
size_t global = ne00;
cl_int ky = ne10;
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
}
}
CL_CHECK(clReleaseEvent(ev));
CL_CHECK(clFinish(queue));
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
}
}
ggml_cl_pool_free(d_X, x_size);
ggml_cl_pool_free(d_D, d_size);
}
void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
ggml_cl_mul_f32(src0, src1, dst);
}
static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
d_X = (cl_mem) src0->data;
} else {
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy data to device
if (src0->backend != GGML_BACKEND_GPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
}
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
if (src0->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
GGML_ASSERT(fp16_support);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb10 = src1->nb[0];
const int nb11 = src1->nb[1];
const int nb12 = src1->nb[2];
const int nb13 = src1->nb[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
d_X = (cl_mem) src0->data;
} else {
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size);
bool src1_cont_rows = nb10 == sizeof(float);
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device
if (src0->backend != GGML_BACKEND_GPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
}
// convert src1 to fp16
// TODO: use multiple threads
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
if (src1_cont_rows) {
if (src1_cont_cols) {
ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
}
}
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
for (int64_t i00 = 0; i00 < ne10; i00++) {
// very slow due to no inlining
tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
}
}
}
// copy src1 to device
CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host, then convert to float
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
}
}
if (src0->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_type type = src0->type;
const bool mul_mat_vec = ne11 == 1;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
size_t x_size;
size_t y_size;
size_t d_size;
size_t q_size;
cl_mem d_X;
if (!mul_mat_vec) {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
cl_mem d_Q;
if (src0->backend == GGML_BACKEND_CPU) {
d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
}
cl_kernel* to_fp32_cl = ggml_get_to_fp32_cl(type);
cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
GGML_ASSERT(to_fp32_cl != nullptr);
size_t ev_idx = 0;
std::vector<cl_event> events;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_CPU) {
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
} else if (src0->backend == GGML_BACKEND_GPU) {
d_Q = (cl_mem) src0->data;
} else {
GGML_ASSERT(false);
}
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
// copy src1 to device
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, events.data() + ev_idx++));
// compute
const size_t global = ne01 * CL_DMMV_BLOCK_SIZE;
const size_t local = CL_DMMV_BLOCK_SIZE;
const cl_int ncols = ne00;
events.emplace_back();
CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, NULL, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
} else { // general dequantization kernel + CLBlast matrix matrix multiplication
// convert src0 to fp32 on device
const size_t global = x_ne;
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
events.emplace_back();
// wait for conversion
CL_CHECK(clFinish(queue));
// compute
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, events.data() + ev_idx++);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
for (auto *event : events) {
clReleaseEvent(event);
}
ev_idx = 0;
events.clear();
}
}
if (!mul_mat_vec) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
if (src0->backend == GGML_BACKEND_CPU) {
ggml_cl_pool_free(d_Q, q_size);
}
}
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
// TODO: find the optimal values for these
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
src1->type == GGML_TYPE_F32 &&
dst->type == GGML_TYPE_F32 &&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
return true;
}
return false;
}
bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
// If device doesn't support FP16
if (!fp16_support) {
return false;
}
size_t src0_sz = ggml_nbytes(src0);
size_t src1_sz = ggml_nbytes(src1);
// mul_mat_q: src0 is converted to fp32 on device
size_t mul_mat_q_transfer = src0_sz + src1_sz;
// mul_mat_f16: src1 is converted to fp16 on cpu
size_t mul_mat_f16_transfer = src0_sz + sizeof(ggml_fp16_t) * ggml_nelements(src1);
// choose the smaller one to transfer to the device
// TODO: this is not always the best choice due to the overhead of converting to fp16
return mul_mat_f16_transfer < mul_mat_q_transfer;
}
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));
if (src0->type == GGML_TYPE_F32) {
ggml_cl_mul_mat_f32(src0, src1, dst);
}
else if (src0->type == GGML_TYPE_F16) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
ggml_cl_mul_mat_f16(src0, src1, dst, wdata, wsize);
}
else {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
}
else if (ggml_is_quantized(src0->type)) {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
else {
GGML_ASSERT(false);
}
}
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
return ggml_nelements(src1) * sizeof(ggml_fp16_t);
}
return 0;
}
void ggml_cl_transform_tensor(ggml_tensor * tensor) {
const int64_t ne0 = tensor->ne[0];
const int64_t ne1 = tensor->ne[1];
const int64_t ne2 = tensor->ne[2];
const int64_t ne3 = tensor->ne[3];
const ggml_type type = tensor->type;
const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
size_t q_size;
cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
// copy tensor to device
for (int64_t i3 = 0; i3 < ne3; i3++) {
for (int64_t i2 = 0; i2 < ne2; i2++) {
int i = i3*ne2 + i2;
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, i*ne0*ne1, tensor, i3, i2, NULL));
}
}
CL_CHECK(clFinish(queue));
tensor->data = dst;
tensor->backend = GGML_BACKEND_GPU;
}
void ggml_cl_load_data(const char * fname, struct ggml_tensor * tensor, const size_t offset) {
cl_int err;
FILE * fp = fopen(fname, "rb");
const size_t size = ggml_nbytes(tensor);
cl_mem dst;
CL_CHECK((dst = clCreateBuffer(context, CL_MEM_READ_ONLY, size, nullptr, &err), err));
void * buf_host = malloc(size);
#ifdef _WIN32
int ret = _fseeki64(fp, (__int64) offset, SEEK_SET);
#else
int ret = fseek(fp, (long) offset, SEEK_SET);
#endif
GGML_ASSERT(ret == 0); // same
size_t ret2 = fread(buf_host, size, 1, fp);
if (ret2 != 1) {
fprintf(stderr, "unexpectedly reached end of file");
exit(1);
}
clEnqueueWriteBuffer(queue, dst, CL_TRUE, 0, size, buf_host, 0, nullptr, nullptr);
tensor->data = dst;
free(buf_host);
fclose(fp);
}
|