forked from ggml-org/llama.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathggml-vulkan.cpp
10154 lines (8716 loc) · 516 KB
/
ggml-vulkan.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
#include "ggml-vulkan.h"
#include <vulkan/vulkan_core.h>
#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF) || defined(GGML_VULKAN_CHECK_RESULTS)
#include <chrono>
#include "ggml-cpu.h"
#endif
#include <vulkan/vulkan.hpp>
#include <algorithm>
#include <cmath>
#include <iomanip>
#include <iostream>
#include <tuple>
#include <vector>
#include <sstream>
#include <utility>
#include <memory>
#include <limits>
#include <map>
#include <unordered_map>
#include <memory>
#include <mutex>
#include <future>
#include <thread>
#if defined(_MSC_VER)
# define NOMINMAX 1
# include <windows.h>
# define YIELD() YieldProcessor()
#elif defined(__clang__) || defined(__GNUC__)
# if defined(__x86_64__) ||defined(__i386__)
# include <immintrin.h>
# define YIELD() _mm_pause()
# elif defined(__arm__) || defined(__aarch64__)
# if defined(__clang__)
# include <arm_acle.h>
# define YIELD() __yield()
# else
# define YIELD() asm volatile("yield")
# endif
# endif
#endif
#if !defined(YIELD)
#define YIELD()
#endif
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-vulkan-shaders.hpp"
#define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
#define VK_VENDOR_ID_AMD 0x1002
#define VK_VENDOR_ID_APPLE 0x106b
#define VK_VENDOR_ID_INTEL 0x8086
#define VK_VENDOR_ID_NVIDIA 0x10de
#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 32
#define GGML_VK_MAX_NODES 8192
#define MAX_VK_BUFFERS 256
#define VK_CHECK(err, msg) \
do { \
vk::Result err_ = (err); \
if (err_ != vk::Result::eSuccess) { \
fprintf(stderr, "ggml_vulkan: %s error %s at %s:%d\n", \
#err, to_string(err_).c_str(), __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
#ifdef GGML_VULKAN_DEBUG
#define VK_LOG_DEBUG(msg) std::cerr << msg << std::endl
#else
#define VK_LOG_DEBUG(msg) ((void) 0)
#endif // GGML_VULKAN_DEBUG
struct ggml_backend_vk_context;
struct vk_queue {
uint32_t queue_family_index;
vk::Queue queue;
vk::CommandPool pool;
uint32_t cmd_buffer_idx;
std::vector<vk::CommandBuffer> cmd_buffers;
vk::PipelineStageFlags stage_flags;
bool transfer_only;
};
struct vk_pipeline_struct {
std::string name;
vk::ShaderModule shader_module;
vk::DescriptorSetLayout dsl;
std::vector<vk::DescriptorPool> descriptor_pools;
std::vector<vk::DescriptorSet> descriptor_sets;
uint32_t descriptor_set_idx;
vk::PipelineLayout layout;
vk::Pipeline pipeline;
uint32_t push_constant_size;
uint32_t parameter_count;
std::array<uint32_t, 3> wg_denoms;
uint32_t align;
// set to true to request the pipeline is compiled after the dryrun
bool needed {};
// set to true when the shader has been compiled
bool compiled {};
};
typedef std::shared_ptr<vk_pipeline_struct> vk_pipeline;
typedef std::weak_ptr<vk_pipeline_struct> vk_pipeline_ref;
static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
struct vk_matmul_pipeline_struct {
vk_pipeline l, m, s;
vk_pipeline a_l, a_m, a_s;
};
typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
struct vk_matmul_pipeline2 {
vk_matmul_pipeline2() {
f16acc = std::make_shared<vk_matmul_pipeline_struct>();
f32acc = std::make_shared<vk_matmul_pipeline_struct>();
}
vk_matmul_pipeline f32acc;
vk_matmul_pipeline f16acc;
};
struct vk_device_struct;
typedef std::shared_ptr<vk_device_struct> vk_device;
typedef std::weak_ptr<vk_device_struct> vk_device_ref;
struct vk_buffer_struct;
typedef std::shared_ptr<vk_buffer_struct> vk_buffer;
typedef std::weak_ptr<vk_buffer_struct> vk_buffer_ref;
struct ggml_backend_vk_buffer_type_context {
std::string name;
vk_device device;
};
static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft);
static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft);
static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor);
static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
/* .get_name = */ ggml_backend_vk_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_vk_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_vk_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_vk_buffer_type_get_max_size,
/* .get_alloc_size = */ ggml_backend_vk_buffer_type_get_alloc_size,
/* .is_host = */ NULL,
};
#ifdef GGML_VULKAN_MEMORY_DEBUG
class vk_memory_logger;
#endif
#ifdef GGML_VULKAN_PERF
class vk_perf_logger;
#endif
static void ggml_vk_destroy_buffer(vk_buffer& buf);
static constexpr uint32_t mul_mat_vec_max_cols = 8;
static constexpr uint32_t p021_max_gqa_ratio = 8;
enum vk_device_architecture {
OTHER,
AMD_GCN,
AMD_RDNA1,
AMD_RDNA2,
AMD_RDNA3,
};
static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
vk::PhysicalDeviceProperties props = device.getProperties();
if (props.vendorID == VK_VENDOR_ID_AMD) {
const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
bool amd_shader_core_properties = false;
bool integer_dot_product = false;
bool subgroup_size_control = false;
for (const auto& properties : ext_props) {
if (strcmp("VK_AMD_shader_core_properties", properties.extensionName) == 0) {
amd_shader_core_properties = true;
} else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0) {
integer_dot_product = true;
} else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
subgroup_size_control = true;
}
}
if (!amd_shader_core_properties || !integer_dot_product || !subgroup_size_control) {
return vk_device_architecture::OTHER;
}
vk::PhysicalDeviceProperties2 props2;
vk::PhysicalDeviceShaderCorePropertiesAMD shader_core_props_amd;
vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;
vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
props2.pNext = &shader_core_props_amd;
shader_core_props_amd.pNext = &integer_dot_props;
integer_dot_props.pNext = &subgroup_size_control_props;
device.getProperties2(&props2);
if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 64) {
return vk_device_architecture::AMD_GCN;
}
if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 32) {
// RDNA
if (shader_core_props_amd.wavefrontsPerSimd == 20) {
return vk_device_architecture::AMD_RDNA1;
}
if (integer_dot_props.integerDotProduct4x8BitPackedMixedSignednessAccelerated) {
return vk_device_architecture::AMD_RDNA3;
}
return vk_device_architecture::AMD_RDNA2;
}
}
return vk_device_architecture::OTHER;
}
struct vk_device_struct {
std::mutex mutex;
vk::PhysicalDevice physical_device;
vk::PhysicalDeviceProperties properties;
std::string name;
uint64_t max_memory_allocation_size;
uint64_t suballocation_block_size;
bool fp16;
bool pipeline_robustness;
vk::Device device;
uint32_t vendor_id;
vk_device_architecture architecture;
vk_queue compute_queue;
vk_queue transfer_queue;
bool single_queue;
uint32_t subgroup_size;
uint32_t shader_core_count;
bool uma;
bool prefer_host_memory;
bool float_controls_rte_fp16;
bool subgroup_add;
bool integer_dot_product;
bool subgroup_size_control;
uint32_t subgroup_min_size;
uint32_t subgroup_max_size;
bool subgroup_require_full_support;
bool coopmat_support;
bool coopmat_acc_f32_support;
bool coopmat_acc_f16_support;
uint32_t coopmat_m;
uint32_t coopmat_n;
uint32_t coopmat_k;
bool coopmat_int_support;
uint32_t coopmat_int_m;
uint32_t coopmat_int_n;
uint32_t coopmat_int_k;
bool coopmat2;
size_t idx;
bool mul_mat_l[GGML_TYPE_COUNT];
bool mul_mat_m[GGML_TYPE_COUNT];
bool mul_mat_s[GGML_TYPE_COUNT];
bool mul_mat_id_l[GGML_TYPE_COUNT];
bool mul_mat_id_m[GGML_TYPE_COUNT];
bool mul_mat_id_s[GGML_TYPE_COUNT];
// set to true to indicate that some shaders need to be compiled after the dryrun
bool need_compiles {};
vk_matmul_pipeline pipeline_matmul_f32 {};
vk_matmul_pipeline pipeline_matmul_f32_f16 {};
vk_matmul_pipeline2 pipeline_matmul_f16;
vk_matmul_pipeline2 pipeline_matmul_f16_f32;
vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_COUNT];
vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT];
vk_matmul_pipeline pipeline_matmul_id_f32 {};
vk_matmul_pipeline2 pipeline_matmul_id_f16;
vk_matmul_pipeline2 pipeline_matmul_id_f16_f32;
vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
vk_pipeline pipeline_matmul_split_k_reduce;
vk_pipeline pipeline_quantize_q8_1;
vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
vk_pipeline pipeline_acc_f32;
vk_pipeline pipeline_add_f32, pipeline_add_f32_norepeat;
vk_pipeline pipeline_add_f16_f32_f16, pipeline_add_f16_f32_f16_norepeat;
vk_pipeline pipeline_sub_f32, pipeline_sub_f32_norepeat;
vk_pipeline pipeline_mul_f32, pipeline_mul_f32_norepeat;
vk_pipeline pipeline_div_f32, pipeline_div_f32_norepeat;
vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
vk_pipeline pipeline_upscale_f32;
vk_pipeline pipeline_scale_f32;
vk_pipeline pipeline_sqr_f32;
vk_pipeline pipeline_sin_f32;
vk_pipeline pipeline_cos_f32;
vk_pipeline pipeline_clamp_f32;
vk_pipeline pipeline_pad_f32;
vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32;
vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
vk_pipeline pipeline_norm_f32;
vk_pipeline pipeline_group_norm_f32;
vk_pipeline pipeline_rms_norm_f32;
vk_pipeline pipeline_rms_norm_back_f32;
vk_pipeline pipeline_l2_norm_f32;
vk_pipeline pipeline_gelu_f32;
vk_pipeline pipeline_gelu_quick_f32;
vk_pipeline pipeline_silu_f32;
vk_pipeline pipeline_silu_back_f32;
vk_pipeline pipeline_relu_f32;
vk_pipeline pipeline_leaky_relu_f32;
vk_pipeline pipeline_tanh_f32;
vk_pipeline pipeline_sigmoid_f32;
vk_pipeline pipeline_diag_mask_inf_f32;
vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
vk_pipeline pipeline_soft_max_back_f32;
vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
vk_pipeline pipeline_argsort_f32;
vk_pipeline pipeline_sum_rows_f32;
vk_pipeline pipeline_argmax_f32;
vk_pipeline pipeline_count_equal_i32;
vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
vk_pipeline pipeline_timestep_embedding_f32;
vk_pipeline pipeline_pool2d_f32;
vk_pipeline pipeline_rwkv_wkv6_f32;
vk_pipeline pipeline_rwkv_wkv7_f32;
vk_pipeline pipeline_opt_step_adamw_f32;
// [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2];
vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2];
vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2];
vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2];
vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2];
vk_pipeline pipeline_flash_attn_split_k_reduce;
std::unordered_map<std::string, vk_pipeline_ref> pipelines;
std::unordered_map<std::string, uint64_t> pipeline_descriptor_set_requirements;
std::vector<std::tuple<void*, size_t, vk_buffer>> pinned_memory;
vk::Fence fence;
vk_buffer sync_staging;
ggml_backend_buffer_type buffer_type;
#ifdef GGML_VULKAN_MEMORY_DEBUG
std::unique_ptr<vk_memory_logger> memory_logger;
#endif
#ifdef GGML_VULKAN_PERF
std::unique_ptr<vk_perf_logger> perf_logger;
#endif
~vk_device_struct() {
VK_LOG_DEBUG("destroy device " << name);
device.destroyFence(fence);
ggml_vk_destroy_buffer(sync_staging);
device.destroyCommandPool(compute_queue.pool);
if (!single_queue) {
device.destroyCommandPool(transfer_queue.pool);
}
for (auto& pipeline : pipelines) {
if (pipeline.second.expired()) {
continue;
}
vk_pipeline pl = pipeline.second.lock();
ggml_vk_destroy_pipeline(device, pl);
}
pipelines.clear();
device.destroy();
}
};
struct vk_buffer_struct {
vk::Buffer buffer = VK_NULL_HANDLE;
vk::DeviceMemory device_memory = VK_NULL_HANDLE;
vk::MemoryPropertyFlags memory_property_flags;
void * ptr;
size_t size = 0;
vk_device device;
~vk_buffer_struct() {
if (size == 0) {
return;
}
VK_LOG_DEBUG("~vk_buffer_struct(" << buffer << ", " << size << ")");
device->device.freeMemory(device_memory);
device->device.destroyBuffer(buffer);
}
};
struct vk_subbuffer {
vk_buffer buffer;
uint64_t offset;
uint64_t size;
operator vk::DescriptorBufferInfo() const {
return { buffer->buffer, offset, size };
}
};
struct vk_semaphore {
vk::Semaphore s;
uint64_t value;
};
struct vk_submission {
vk::CommandBuffer buffer;
std::vector<vk_semaphore> wait_semaphores;
std::vector<vk_semaphore> signal_semaphores;
};
typedef std::vector<vk_submission> vk_sequence;
struct vk_mat_mat_push_constants {
uint32_t M; uint32_t N; uint32_t K;
uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
uint32_t k_split;
uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
uint32_t padded_N;
};
struct vk_mat_vec_push_constants {
uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
};
struct vk_mat_mat_id_push_constants {
uint32_t M; uint32_t N; uint32_t K;
uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
uint32_t nei0; uint32_t nei1; uint32_t nbi1; uint32_t ne11;
uint32_t padded_N;
};
struct vk_mat_vec_id_push_constants {
uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
uint32_t nei0; uint32_t ne11;
};
struct vk_flash_attn_push_constants {
uint32_t N;
uint32_t KV;
uint32_t ne1;
uint32_t ne2;
uint32_t ne3;
uint32_t neq2;
uint32_t neq3;
uint32_t nek2;
uint32_t nek3;
uint32_t nev2;
uint32_t nev3;
uint32_t nem1;
uint32_t nb01;
uint32_t nb02;
uint32_t nb03;
uint32_t nb11;
uint32_t nb12;
uint32_t nb13;
uint32_t nb21;
uint32_t nb22;
uint32_t nb23;
uint32_t nb31;
float scale;
float max_bias;
float logit_softcap;
uint32_t mask;
uint32_t n_head_log2;
float m0;
float m1;
uint32_t gqa_ratio;
uint32_t split_kv;
uint32_t k_num;
};
struct vk_op_push_constants {
uint32_t KX;
uint32_t KY;
float param1;
float param2;
};
struct vk_op_unary_push_constants {
uint32_t ne;
uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
uint32_t misalign_offsets;
float param1; float param2;
uint32_t ne0_012mp; uint32_t ne0_012L;
uint32_t ne0_01mp; uint32_t ne0_01L;
uint32_t ne0_0mp; uint32_t ne0_0L;
uint32_t ne1_012mp; uint32_t ne1_012L;
uint32_t ne1_01mp; uint32_t ne1_01L;
uint32_t ne1_0mp; uint32_t ne1_0L;
};
static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_push_constants) must be <= 128");
// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
// Precompute mp (m' in the paper) and L such that division
// can be computed using a multiply (high 32b of 64b result)
// and a shift:
//
// n/d = (mulhi(n, mp) + n) >> L;
static void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L)
{
// compute L = ceil(log2(d));
L = 0;
while (L < 32 && (uint32_t{1} << L) < d) {
L++;
}
mp = (uint32_t)((uint64_t{1} << 32) * ((uint64_t{1} << L) - d) / d + 1);
}
template <typename T> void init_pushconst_fastdiv(T &p) {
GGML_UNUSED(p);
static_assert(!std::is_const<T>::value, "unexpected type");
}
template <> void init_pushconst_fastdiv(vk_op_unary_push_constants &p) {
// Compute magic values to divide by these six numbers.
init_fastdiv_values(p.ne02*p.ne01*p.ne00, p.ne0_012mp, p.ne0_012L);
init_fastdiv_values(p.ne01*p.ne00, p.ne0_01mp, p.ne0_01L);
init_fastdiv_values(p.ne00, p.ne0_0mp, p.ne0_0L);
init_fastdiv_values(p.ne12*p.ne11*p.ne10, p.ne1_012mp, p.ne1_012L);
init_fastdiv_values(p.ne11*p.ne10, p.ne1_01mp, p.ne1_01L);
init_fastdiv_values(p.ne10, p.ne1_0mp, p.ne1_0L);
}
struct vk_op_binary_push_constants {
uint32_t ne;
uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
uint32_t misalign_offsets;
float param1; float param2; int32_t param3;
};
struct vk_op_diag_mask_push_constants {
uint32_t ncols;
uint32_t rows_per_channel;
int32_t n_past;
};
struct vk_op_rope_push_constants {
uint32_t ncols;
uint32_t n_dims;
float freq_scale;
uint32_t p_delta_rows;
float freq_base;
float ext_factor;
float attn_factor;
float corr_dims[2];
float theta_scale;
uint32_t has_ff;
uint32_t ne02;
uint32_t s1;
uint32_t s2;
int32_t sections[4];
uint32_t is_back;
};
struct vk_op_soft_max_push_constants {
uint32_t KX;
uint32_t KY;
float scale;
float max_bias;
float m0;
float m1;
uint32_t n_head_log2;
uint32_t nrows_x;
};
struct vk_op_argsort_push_constants {
uint32_t ncols;
uint32_t ncols_pad;
int32_t order;
};
struct vk_op_im2col_push_constants {
uint32_t batch_offset; uint32_t offset_delta;
uint32_t IC;
uint32_t IW; uint32_t IH;
uint32_t OW; uint32_t OH;
uint32_t KW; uint32_t KH;
uint32_t pelements;
uint32_t CHW;
int32_t s0; int32_t s1;
int32_t p0; int32_t p1;
int32_t d0; int32_t d1;
};
struct vk_op_timestep_embedding_push_constants {
uint32_t nb1;
uint32_t dim;
uint32_t max_period;
};
struct vk_op_pool2d_push_constants {
uint32_t IW; uint32_t IH;
uint32_t OW; uint32_t OH;
uint32_t OC;
uint32_t pelements;
uint32_t op;
int32_t k0; int32_t k1;
int32_t s0; int32_t s1;
int32_t p0; int32_t p1;
};
struct vk_op_rwkv_wkv6_push_constants {
uint32_t B;
uint32_t T;
uint32_t C;
uint32_t H;
};
struct vk_op_rwkv_wkv7_push_constants {
uint32_t B;
uint32_t T;
uint32_t C;
uint32_t H;
};
struct vk_op_upscale_push_constants {
uint32_t ne; uint32_t a_offset; uint32_t d_offset;
uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
float sf0; float sf1; float sf2; float sf3;
};
// Allow pre-recording command buffers
struct vk_staging_memcpy {
vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
void * dst;
const void * src;
size_t n;
};
struct vk_context_struct {
vk_submission * s;
std::vector<vk_sequence> seqs;
int exit_tensor_idx;
std::vector<vk_staging_memcpy> in_memcpys;
std::vector<vk_staging_memcpy> out_memcpys;
vk_queue * q;
};
typedef std::shared_ptr<vk_context_struct> vk_context;
typedef std::weak_ptr<vk_context_struct> vk_context_ref;
struct ggml_vk_garbage_collector {
std::vector<vk_semaphore> tl_semaphores;
std::vector<vk_semaphore> semaphores;
std::vector<vk::Event> events;
std::vector<vk_buffer> temp_buffers;
std::vector<vk_context> contexts;
};
#if defined(GGML_VULKAN_MEMORY_DEBUG) || defined(GGML_VULKAN_DEBUG)
#define VK_LOG_MEMORY(msg) std::cerr << "ggml_vulkan memory: " << msg << std::endl
static std::string format_size(size_t size) {
const size_t kib = 1024;
const size_t mib = kib * 1024;
const size_t gib = mib * 1024;
std::ostringstream oss;
oss << std::fixed << std::setprecision(2);
if (size >= gib) {
oss << static_cast<double>(size) / gib << " GiB";
} else if (size >= mib) {
oss << static_cast<double>(size) / mib << " MiB";
} else if (size >= kib) {
oss << static_cast<double>(size) / kib << " KiB";
} else {
oss << size << " B";
}
return oss.str();
}
static std::mutex log_mutex;
class vk_memory_logger {
public:
vk_memory_logger(): total_device(0), total_host(0) {}
void log_allocation(vk_buffer_ref buf_ref, size_t size);
void log_deallocation(vk_buffer_ref buf_ref);
private:
std::map<vk::Buffer, size_t> allocations; // Track allocations
size_t total_device;
size_t total_host;
};
#else
#define VK_LOG_MEMORY(msg) ((void) 0)
#endif // GGML_VULKAN_MEMORY_DEBUG
#if defined(GGML_VULKAN_PERF)
class vk_perf_logger {
public:
void print_timings() {
std::cerr << "----------------\nVulkan Timings:" << std::endl;
for (const auto& t : timings) {
uint64_t total = 0;
for (const auto& time : t.second) {
total += time;
}
std::cerr << t.first << ": " << t.second.size() << " x " << (total / t.second.size() / 1000.0) << " ms" << std::endl;
}
timings.clear();
}
void log_timing(const ggml_tensor * node, uint64_t time) {
if (node->op == GGML_OP_UNARY) {
timings[ggml_unary_op_name(ggml_get_unary_op(node))].push_back(time);
return;
}
if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
const uint64_t m = node->src[0]->ne[1];
const uint64_t n = node->src[1]->ne[1];
const uint64_t k = node->src[1]->ne[0];
std::string name = ggml_op_name(node->op);
if (n == 1) {
name += "_VEC m=" + std::to_string(m) + " k=" + std::to_string(k);
} else {
name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
}
timings[name].push_back(time);
return;
}
timings[ggml_op_name(node->op)].push_back(time);
}
private:
std::map<std::string, std::vector<uint64_t>> timings;
};
#endif // GGML_VULKAN_PERF
struct ggml_backend_vk_context {
std::string name;
vk_device device;
size_t semaphore_idx, event_idx;
ggml_vk_garbage_collector gc;
size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k;
vk_buffer prealloc_x, prealloc_y, prealloc_split_k;
vk::Fence fence, almost_ready_fence;
bool almost_ready_fence_pending {};
vk_buffer buffer_pool[MAX_VK_BUFFERS];
vk_context_ref compute_ctx;
vk_context_ref transfer_ctx;
std::vector<vk_context_ref> tensor_ctxs;
};
static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000; // NOLINT
static uint64_t vk_tensor_offset(const ggml_tensor * tensor) {
if (tensor->view_src) {
return (uint8_t *) tensor->view_src->data - (uint8_t *) vk_ptr_base;
}
return (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
}
struct ggml_backend_vk_buffer_context {
vk_device_ref device;
vk_buffer dev_buffer;
std::string name;
ggml_backend_vk_buffer_context(vk_device_ref device, vk_buffer&& dev_buffer, std::string& name) :
device(device),
dev_buffer(dev_buffer),
name(name) {
}
~ggml_backend_vk_buffer_context() {
ggml_vk_destroy_buffer(dev_buffer);
}
};
#ifdef GGML_VULKAN_MEMORY_DEBUG
void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
std::lock_guard<std::mutex> guard(log_mutex);
vk_buffer buf = buf_ref.lock();
const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
const std::string type = device ? "device" : "host";
allocations[buf->buffer] = size;
total_device += device ? size : 0;
total_host += device ? 0 : size;
VK_LOG_MEMORY(buf->device->name << ": +" << format_size(size) << " " << type << " at " << buf->buffer << ". Total device: " << format_size(total_device) << ", total host: " << format_size(total_host));
}
void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
if (buf_ref.expired() || buf_ref.lock()->size == 0) {
return;
}
std::lock_guard<std::mutex> guard(log_mutex);
vk_buffer buf = buf_ref.lock();
const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
std::string type = device ? "device" : "host";
auto it = allocations.find(buf->buffer);
total_device -= device ? it->second : 0;
total_host -= device ? 0 : it->second;
if (it != allocations.end()) {
VK_LOG_MEMORY(buf->device->name << ": -" << format_size(it->second) << " " << type << " at " << buf->buffer << ". Total device: " << format_size(total_device) << ", total host: " << format_size(total_host));
allocations.erase(it);
} else {
VK_LOG_MEMORY("ERROR " << buf->device->name << ": Attempted to deallocate unknown " << type << " memory at " << buf->buffer);
}
}
#endif // GGML_VULKAN_MEMORY_DEBUG
struct vk_instance_t {
vk::Instance instance;
std::vector<size_t> device_indices;
vk_device devices[GGML_VK_MAX_DEVICES];
};
static bool vk_instance_initialized = false;
static vk_instance_t vk_instance;
#ifdef GGML_VULKAN_CHECK_RESULTS
static size_t vk_skip_checks;
static size_t vk_output_tensor;
static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name);
static void ggml_vk_check_results_0(ggml_tensor * tensor);
static void ggml_vk_check_results_1(ggml_tensor * tensor);
#endif
typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
static void ggml_backend_vk_free(ggml_backend_t backend);
// Wait for ctx->fence to be signaled.
static void ggml_vk_wait_for_fence(ggml_backend_vk_context * ctx) {
// Use waitForFences while most of the graph executes. Hopefully the CPU can sleep
// during this wait.
if (ctx->almost_ready_fence_pending) {
VK_CHECK(ctx->device->device.waitForFences({ ctx->almost_ready_fence }, true, UINT64_MAX), "almost_ready_fence");
ctx->device->device.resetFences({ ctx->almost_ready_fence });
ctx->almost_ready_fence_pending = false;
}
// Spin (w/pause) waiting for the graph to finish executing.
vk::Result result;
while ((result = ctx->device->device.getFenceStatus(ctx->fence)) != vk::Result::eSuccess) {
if (result != vk::Result::eNotReady) {
fprintf(stderr, "ggml_vulkan: error %s at %s:%d\n", to_string(result).c_str(), __FILE__, __LINE__);
exit(1);
}
for (uint32_t i = 0; i < 100; ++i) {
YIELD();
YIELD();
YIELD();
YIELD();
YIELD();
YIELD();
YIELD();
YIELD();
YIELD();
YIELD();
}
}
ctx->device->device.resetFences({ ctx->fence });
}
// variables to track number of compiles in progress
static uint32_t compile_count = 0;
static std::mutex compile_count_mutex;
static std::condition_variable compile_count_cond;
static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, size_t spv_size, const void* spv_data, const std::string entrypoint,
uint32_t parameter_count, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants,
bool disable_robustness, bool require_full_subgroups, uint32_t required_subgroup_size) {
VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << pipeline->name << ", " << entrypoint << ", " << parameter_count <<
", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " <<
disable_robustness << ", " << require_full_subgroups << ", " << required_subgroup_size << ")");
GGML_ASSERT(parameter_count > 0);
GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data));
pipeline->shader_module = device->device.createShaderModule(shader_module_create_info);
std::vector<vk::DescriptorSetLayoutBinding> dsl_binding;
std::vector<vk::DescriptorBindingFlags> dsl_binding_flags;
for (uint32_t i = 0; i < parameter_count; i++) {
dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute});
dsl_binding_flags.push_back({});
}
vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags };
vk::PushConstantRange pcr(
vk::ShaderStageFlagBits::eCompute,
0,
pipeline->push_constant_size
);
vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
{},
dsl_binding);
descriptor_set_layout_create_info.setPNext(&dslbfci);
pipeline->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info);
vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
pipeline->descriptor_set_idx = 0;
vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline->dsl, pcr);
pipeline->layout = device->device.createPipelineLayout(pipeline_layout_create_info);
std::vector<vk::SpecializationMapEntry> specialization_entries(specialization_constants.size());
for (size_t i = 0; i < specialization_constants.size(); i++) {
specialization_entries[i].constantID = i;
specialization_entries[i].offset = i * sizeof(uint32_t);
specialization_entries[i].size = sizeof(uint32_t);
}
vk::SpecializationInfo specialization_info(
specialization_entries.size(),
specialization_entries.data(),
specialization_constants.size() * sizeof(uint32_t),
specialization_constants.data()
);
vk::PipelineShaderStageCreateFlags pipeline_shader_stage_create_flags{};