forked from ggerganov/llama.cpp
-
Notifications
You must be signed in to change notification settings - Fork 1
/
ggml-vulkan.cpp
6442 lines (5411 loc) · 310 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"
#ifdef GGML_VULKAN_RUN_TESTS
#include <chrono>
#endif
#include <vulkan/vulkan.hpp>
#include <algorithm>
#include <cmath>
#include <iostream>
#include <limits>
#include <tuple>
#include <vector>
#include <sstream>
#include <utility>
#include <memory>
#include "ggml.h"
#include "ggml-backend-impl.h"
#include "ggml-vulkan-shaders.hpp"
#define VK_API_VERSION VK_API_VERSION_1_2
#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
#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_MODE_UNKNOWN 0
#define VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI 1
#define VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE 2
#define VK_NUM_TYPES 16
#define GGML_VK_MAX_NODES 8192
#define MAX_VK_BUFFERS 256
#ifndef K_QUANTS_PER_ITERATION
#define K_QUANTS_PER_ITERATION 1
#else
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
#endif
#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)
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;
};
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;
};
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_device {
vk::PhysicalDevice physical_device;
vk::PhysicalDeviceProperties properties;
std::string name;
uint64_t max_memory_allocation_size;
bool fp16;
vk::Device device;
uint32_t vendor_id;
vk_queue compute_queue;
vk_queue transfer_queue;
bool single_queue;
uint32_t descriptor_set_mode;
uint32_t subgroup_size;
bool uma;
bool initialized;
size_t idx;
vk_matmul_pipeline pipeline_matmul_f32;
vk_matmul_pipeline pipeline_matmul_f16;
vk_matmul_pipeline pipeline_matmul_f16_f32;
vk_pipeline pipeline_matmul_split_k_reduce;
vk_matmul_pipeline pipeline_dequant_mul_mat_mat[VK_NUM_TYPES];
vk_pipeline pipeline_dequant[VK_NUM_TYPES];
vk_pipeline pipeline_dequant_mul_mat_vec_f32[VK_NUM_TYPES];
vk_pipeline pipeline_mul_mat_vec_p021_f16_f32;
vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
vk_pipeline pipeline_get_rows[VK_NUM_TYPES];
vk_pipeline pipeline_get_rows_f32[VK_NUM_TYPES];
vk_pipeline pipeline_mul_f32;
vk_pipeline pipeline_add_f32;
vk_pipeline pipeline_scale_f32;
vk_pipeline pipeline_sqr_f32;
vk_pipeline pipeline_clamp_f32;
vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
vk_pipeline pipeline_norm_f32;
vk_pipeline pipeline_rms_norm_f32;
vk_pipeline pipeline_gelu_f32;
vk_pipeline pipeline_silu_f32;
vk_pipeline pipeline_relu_f32;
vk_pipeline pipeline_diag_mask_inf_f32;
vk_pipeline pipeline_soft_max_f32;
vk_pipeline pipeline_rope_f32, pipeline_rope_f16;
vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
vk_pipeline pipeline_argsort_f32;
std::vector<vk_pipeline_ref> pipelines;
~vk_device() {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "destroy device " << name << std::endl;
#endif
device.destroyCommandPool(compute_queue.pool);
if (!single_queue) {
device.destroyCommandPool(transfer_queue.pool);
}
for (auto& pipeline : pipelines) {
if (pipeline.expired()) {
continue;
}
vk_pipeline pl = pipeline.lock();
ggml_vk_destroy_pipeline(device, pl);
}
pipelines.clear();
device.destroy();
}
};
struct vk_buffer_struct {
vk::Buffer buffer;
vk::DeviceMemory device_memory;
vk::MemoryPropertyFlags memory_property_flags;
void * ptr;
size_t size = 0;
ggml_backend_vk_context * ctx;
std::shared_ptr<vk_device> device;
~vk_buffer_struct() {
if (size == 0) {
return;
}
#ifdef GGML_VULKAN_DEBUG
std::cerr << "~vk_buffer_struct(" << buffer << ", " << size << ")" << std::endl;
#endif
device->device.freeMemory(device_memory);
device->device.destroyBuffer(buffer);
}
};
typedef std::shared_ptr<vk_buffer_struct> vk_buffer;
typedef std::weak_ptr<vk_buffer_struct> vk_buffer_ref;
struct vk_subbuffer {
vk_buffer buffer;
uint64_t offset;
uint64_t 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_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 d_offset;
float param1; float param2;
};
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 d_offset;
float param1; float param2;
};
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;
float freq_scale;
uint32_t p_delta_rows;
float freq_base;
float ext_factor;
float attn_factor;
float corr_dims[4];
};
struct vk_op_rope_neox_push_constants {
uint32_t ncols;
uint32_t ndims;
float freq_scale;
uint32_t p_delta_rows;
float freq_base;
float ext_factor;
float attn_factor;
float corr_dims[4];
float theta_scale;
float inv_ndims;
};
struct vk_op_soft_max_push_constants {
uint32_t KX;
uint32_t KY;
uint32_t KZ;
float scale;
float max_bias;
float m0;
float m1;
uint32_t n_head_log2;
};
struct vk_op_argsort_push_constants {
uint32_t ncols;
bool ascending;
};
// 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 {
size_t idx;
vk_submission * s;
std::vector<vk_sequence> seqs;
ggml_tensor * exit_tensor;
std::vector<vk_staging_memcpy> in_memcpys;
std::vector<vk_staging_memcpy> out_memcpys;
vk_queue * q;
};
struct ggml_tensor_extra_gpu {
bool ready;
size_t ctx_idx;
vk_buffer_ref buffer_gpu;
uint64_t offset;
void reset() {
ready = false;
ctx_idx = 0;
buffer_gpu.reset();
offset = 0;
}
};
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;
};
struct ggml_backend_vk_context {
std::string name;
std::shared_ptr<vk_device> device;
size_t semaphore_idx, event_idx;
ggml_vk_garbage_collector gc;
std::vector<std::tuple<void*, size_t, vk_buffer>> pinned_memory;
size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k;
vk_buffer prealloc_x, prealloc_y, prealloc_split_k;
vk::Fence fence;
vk_buffer staging;
size_t staging_size;
size_t staging_offset;
vk_buffer sync_staging;
vk_buffer buffer_pool[MAX_VK_BUFFERS];
vk_context * compute_ctx;
vk_context * transfer_ctx;
bool disable;
bool initialized;
size_t idx;
};
struct vk_instance {
vk::Instance instance;
std::vector<size_t> device_indices;
ggml_backend_t backends[GGML_VK_MAX_DEVICES];
ggml_backend_vk_context contexts[GGML_VK_MAX_DEVICES];
ggml_backend_buffer_type buffer_types[GGML_VK_MAX_DEVICES];
bool initialized[GGML_VK_MAX_DEVICES];
};
static std::shared_ptr<vk_device> ggml_vk_get_device(size_t idx) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_get_device(" << idx << ")" << std::endl;
#endif
static std::weak_ptr<vk_device> devices[GGML_VK_MAX_DEVICES];
if (devices[idx].expired()) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "Initializing new vk_device" << std::endl;
#endif
std::shared_ptr<vk_device> device = std::make_shared<vk_device>();
device->initialized = false;
devices[idx] = device;
return device;
}
return devices[idx].lock();
}
#ifdef GGML_VULKAN_CHECK_RESULTS
static size_t vk_skip_checks;
static size_t vk_output_tensor;
static void ggml_vk_print_tensor(ggml_backend * ctx, const ggml_tensor * tensor, const char * name);
static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor);
static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_params * params, 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 bool vk_instance_initialized = false;
static vk_instance vk_instance;
GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend);
static void ggml_vk_create_pipeline(ggml_backend_vk_context * ctx, vk_pipeline& pipeline, const std::string& name, size_t spv_size, const void* spv_data, const std::string& entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_pipeline(" << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")" << std::endl;
#endif
GGML_ASSERT(parameter_count > 0);
GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
pipeline = std::make_shared<vk_pipeline_struct>();
pipeline->name = name;
pipeline->parameter_count = parameter_count;
pipeline->push_constant_size = push_constant_size;
pipeline->wg_denoms = wg_denoms;
pipeline->align = align;
vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data));
pipeline->shader_module = ctx->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 = ctx->device->device.createDescriptorSetLayout(descriptor_set_layout_create_info);
// Check if device supports multiple descriptors per pool
if (ctx->device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN) {
const uint32_t alloc_count = 2;
// Try allocating multiple sets from one pool
// This fails on AMD for some reason, so add a fall back to allocating one pool per set
vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count);
vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, alloc_count, descriptor_pool_size);
vk::DescriptorPool pool = ctx->device->device.createDescriptorPool(descriptor_pool_create_info);
std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
for (uint32_t i = 0; i < alloc_count; i++) {
layouts[i] = pipeline->dsl;
}
try {
vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pool, alloc_count, layouts.data());
std::vector<vk::DescriptorSet> sets = ctx->device->device.allocateDescriptorSets(descriptor_set_alloc_info);
} catch(vk::OutOfPoolMemoryError const&) {
ctx->device->descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE;
}
ctx->device->device.destroyDescriptorPool(pool);
}
if (ctx->device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) {
vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count);
vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 128, descriptor_pool_size);
pipeline->descriptor_pools.push_back(ctx->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 = ctx->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::PipelineShaderStageCreateInfo pipeline_shader_create_info(
vk::PipelineShaderStageCreateFlags(),
vk::ShaderStageFlagBits::eCompute,
pipeline->shader_module,
entrypoint.c_str(),
&specialization_info);
vk::ComputePipelineCreateInfo compute_pipeline_create_info(
vk::PipelineCreateFlags(),
pipeline_shader_create_info,
pipeline->layout);
pipeline->pipeline = ctx->device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
ctx->device->pipelines.push_back(pipeline);
}
static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_pipeline_destroy_pipeline(" << pipeline->name << ")" << std::endl;
#endif
for (auto& pool : pipeline->descriptor_pools) {
device.destroyDescriptorPool(pool);
}
pipeline->descriptor_pools.clear();
pipeline->descriptor_sets.clear();
pipeline->descriptor_set_idx = 0;
device.destroyDescriptorSetLayout(pipeline->dsl);
device.destroyPipelineLayout(pipeline->layout);
device.destroyShaderModule(pipeline->shader_module);
device.destroyPipeline(pipeline->pipeline);
}
static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx, vk_pipeline& pipeline, uint32_t n) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_pipeline_allocate_descriptor_sets(" << pipeline->name << ", " << n << ")" << std::endl;
#endif
if (pipeline->descriptor_sets.size() >= pipeline->descriptor_set_idx + n) {
// Enough descriptors are available
return;
}
if (ctx->device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) {
const uint32_t alloc_count = pipeline->descriptor_set_idx + n - pipeline->descriptor_sets.size();
std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
for (uint32_t i = 0; i < alloc_count; i++) {
layouts[i] = pipeline->dsl;
}
vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[0], alloc_count, layouts.data());
std::vector<vk::DescriptorSet> sets = ctx->device->device.allocateDescriptorSets(descriptor_set_alloc_info);
pipeline->descriptor_sets.insert(pipeline->descriptor_sets.end(), sets.begin(), sets.end());
} else {
for (uint32_t i = pipeline->descriptor_sets.size(); i < pipeline->descriptor_set_idx + n; i++) {
vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count);
vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 1, descriptor_pool_size);
pipeline->descriptor_pools.push_back(ctx->device->device.createDescriptorPool(descriptor_pool_create_info));
vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[i], 1, &pipeline->dsl);
std::vector<vk::DescriptorSet> sets = ctx->device->device.allocateDescriptorSets(descriptor_set_alloc_info);
pipeline->descriptor_sets.push_back(sets[0]);
}
}
}
static void ggml_pipeline_cleanup(vk_pipeline& pipeline) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_pipeline_cleanup(" << pipeline->name << ")" << std::endl;
#endif
pipeline->descriptor_set_idx = 0;
}
static vk::CommandBuffer ggml_vk_create_cmd_buffer(ggml_backend_vk_context * ctx, vk_queue& q) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_cmd_buffer()" << std::endl;
#endif
if (q.cmd_buffers.size() > q.cmd_buffer_idx) {
// Reuse command buffer
return q.cmd_buffers[q.cmd_buffer_idx++];
}
vk::CommandBufferAllocateInfo command_buffer_alloc_info(
q.pool,
vk::CommandBufferLevel::ePrimary,
1);
const std::vector<vk::CommandBuffer> cmd_buffers = ctx->device->device.allocateCommandBuffers(command_buffer_alloc_info);
auto buf = cmd_buffers.front();
q.cmd_buffers.push_back(buf);
q.cmd_buffer_idx++;
return buf;
}
static vk_submission ggml_vk_create_submission(ggml_backend_vk_context * ctx, vk_queue& q, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_submission()" << std::endl;
#endif
vk_submission s;
s.buffer = ggml_vk_create_cmd_buffer(ctx, q);
s.wait_semaphores = std::move(wait_semaphores);
s.signal_semaphores = std::move(signal_semaphores);
return s;
}
static void ggml_vk_submit(vk_context * ctx, vk::Fence fence) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_submit(" << ctx->seqs.size() << ", " << fence << ")" << std::endl;
#endif
if (ctx->seqs.empty()) {
return;
}
std::vector<std::vector<uint64_t>> tl_wait_vals;
std::vector<std::vector<uint64_t>> tl_signal_vals;
std::vector<std::vector<vk::Semaphore>> tl_wait_semaphores;
std::vector<std::vector<vk::Semaphore>> tl_signal_semaphores;
std::vector<vk::TimelineSemaphoreSubmitInfo> tl_submit_infos;
std::vector<vk::SubmitInfo> submit_infos;
int idx = -1;
std::vector<std::vector<vk::PipelineStageFlags>> stage_flags;
size_t reserve = 0;
for (const auto& sequence : ctx->seqs) {
reserve += sequence.size();
}
// Pre-reserve vectors to prevent reallocation, which invalidates pointers
tl_wait_semaphores.reserve(reserve);
tl_wait_vals.reserve(reserve);
tl_signal_semaphores.reserve(reserve);
tl_signal_vals.reserve(reserve);
tl_submit_infos.reserve(reserve);
submit_infos.reserve(reserve);
stage_flags.reserve(reserve);
for (const auto& sequence : ctx->seqs) {
for (const auto& submission : sequence) {
stage_flags.push_back({});
idx++;
tl_wait_vals.push_back({});
tl_wait_semaphores.push_back({});
tl_signal_vals.push_back({});
tl_signal_semaphores.push_back({});
for (size_t i = 0; i < submission.wait_semaphores.size(); i++) {
stage_flags[idx].push_back(ctx->q->stage_flags);
tl_wait_vals[idx].push_back(submission.wait_semaphores[i].value);
tl_wait_semaphores[idx].push_back(submission.wait_semaphores[i].s);
}
for (size_t i = 0; i < submission.signal_semaphores.size(); i++) {
tl_signal_vals[idx].push_back(submission.signal_semaphores[i].value);
tl_signal_semaphores[idx].push_back(submission.signal_semaphores[i].s);
}
tl_submit_infos.push_back({
(uint32_t) submission.wait_semaphores.size(),
tl_wait_vals[idx].data(),
(uint32_t) submission.signal_semaphores.size(),
tl_signal_vals[idx].data(),
});
tl_submit_infos[idx].sType = vk::StructureType::eTimelineSemaphoreSubmitInfo;
tl_submit_infos[idx].pNext = nullptr;
vk::SubmitInfo si{
(uint32_t) submission.wait_semaphores.size(),
tl_wait_semaphores[idx].data(),
stage_flags[idx].data(),
1,
&submission.buffer,
(uint32_t) submission.signal_semaphores.size(),
tl_signal_semaphores[idx].data(),
};
si.setPNext(&tl_submit_infos[idx]);
submit_infos.push_back(si);
}
}
ctx->q->queue.submit(submit_infos, fence);
ctx->seqs.clear();
}
static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyProperties>& queue_family_props, const vk::QueueFlags& required, const vk::QueueFlags& avoid, int32_t compute_index, uint32_t min_num_queues) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_find_queue_family_index()" << std::endl;
#endif
const uint32_t qfsize = queue_family_props.size();
// Try with avoid preferences first
for (uint32_t i = 0; i < qfsize; i++) {
if (queue_family_props[i].queueCount >= min_num_queues && (compute_index < 0 || i != (uint32_t) compute_index) && queue_family_props[i].queueFlags & required && !(queue_family_props[i].queueFlags & avoid)) {
return i;
}
}
// Fall back to only required
for (size_t i = 0; i < qfsize; i++) {
if (queue_family_props[i].queueCount >= min_num_queues && (compute_index < 0 || i != (uint32_t) compute_index) && queue_family_props[i].queueFlags & required) {
return i;
}
}
// Fall back to reusing compute queue
for (size_t i = 0; i < qfsize; i++) {
if (queue_family_props[i].queueCount >= min_num_queues && queue_family_props[i].queueFlags & required) {
return i;
}
}
// Fall back to ignoring min_num_queries
for (size_t i = 0; i < qfsize; i++) {
if (queue_family_props[i].queueFlags & required) {
return i;
}
}
// All commands that are allowed on a queue that supports transfer operations are also allowed on a queue that supports either graphics or compute operations.
// Thus, if the capabilities of a queue family include VK_QUEUE_GRAPHICS_BIT or VK_QUEUE_COMPUTE_BIT, then reporting the VK_QUEUE_TRANSFER_BIT capability separately for that queue family is optional.
if (compute_index >= 0) {
return compute_index;
}
std::cerr << "ggml_vulkan: No suitable queue family index found." << std::endl;
for(auto &q_family : queue_family_props) {
std::cerr << "Queue number: " + std::to_string(q_family.queueCount) << " flags: " + to_string(q_family.queueFlags) << std::endl;
}
abort();
}
static void ggml_vk_create_queue(ggml_backend_vk_context * ctx, vk_queue& q, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_queue()" << std::endl;
#endif
q.queue_family_index = queue_family_index;
vk::CommandPoolCreateInfo command_pool_create_info_compute(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), queue_family_index);
q.pool = ctx->device->device.createCommandPool(command_pool_create_info_compute);
q.cmd_buffer_idx = 0;
q.queue = ctx->device->device.getQueue(queue_family_index, queue_index);
q.stage_flags = stage_flags;
}
static vk_context * ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_context()" << std::endl;
#endif
ctx->gc.contexts.emplace_back();
vk_context * result = &ctx->gc.contexts[ctx->gc.contexts.size() - 1];
memset((void *) result, 0, sizeof(vk_context));
result->idx = ctx->gc.contexts.size() - 1;
result->q = &q;
return result;
}
static vk_semaphore * ggml_vk_create_binary_semaphore(ggml_backend_vk_context * ctx) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_timeline_semaphore()" << std::endl;
#endif
vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eBinary, 0 };
vk::SemaphoreCreateInfo ci{};
ci.setPNext(&tci);
vk::Semaphore semaphore = ctx->device->device.createSemaphore(ci);
ctx->gc.semaphores.push_back({ semaphore, 0 });
return &ctx->gc.semaphores[ctx->gc.semaphores.size() - 1];
}
static vk_semaphore * ggml_vk_create_timeline_semaphore(ggml_backend_vk_context * ctx) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_timeline_semaphore()" << std::endl;
#endif
if (ctx->semaphore_idx >= ctx->gc.tl_semaphores.size()) {
vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 };
vk::SemaphoreCreateInfo ci{};
ci.setPNext(&tci);
vk::Semaphore semaphore = ctx->device->device.createSemaphore(ci);
ctx->gc.tl_semaphores.push_back({ semaphore, 0 });
}
return &ctx->gc.tl_semaphores[ctx->semaphore_idx++];
}
static vk::Event ggml_vk_create_event(ggml_backend_vk_context * ctx) {
if (ctx->event_idx >= ctx->gc.events.size()) {
ctx->gc.events.push_back(ctx->device->device.createEvent({}));
}
return ctx->gc.events[ctx->event_idx++];
}
static void ggml_vk_queue_cleanup(ggml_backend_vk_context * ctx, vk_queue& q) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_queue_cleanup()" << std::endl;
#endif
// Requires command buffers to be done
ctx->device->device.resetCommandPool(q.pool);
q.cmd_buffer_idx = 0;
}
static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
vk::MemoryType memory_type = mem_props->memoryTypes[i];
if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
(flags & memory_type.propertyFlags) == flags &&
mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
return static_cast<int32_t>(i);
}
}
return UINT32_MAX;
}
static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_buffer(device " << ctx->idx << ", " << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags) << ")" << std::endl;
#endif
vk_buffer buf = std::make_shared<vk_buffer_struct>();
if (size == 0) {
buf->size = 0;
return buf;
}
buf->size = size;
vk::BufferCreateInfo buffer_create_info{
vk::BufferCreateFlags(),
size,
vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc | vk::BufferUsageFlagBits::eTransferDst,
vk::SharingMode::eExclusive,
0,
nullptr,
};
buf->buffer = ctx->device->device.createBuffer(buffer_create_info);
vk::MemoryRequirements mem_req = ctx->device->device.getBufferMemoryRequirements(buf->buffer);
vk::PhysicalDeviceMemoryProperties mem_props = ctx->device->physical_device.getMemoryProperties();
uint32_t memory_type_index = UINT32_MAX;
memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
buf->memory_property_flags = req_flags;
if (memory_type_index == UINT32_MAX && fallback_flags) {
memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
buf->memory_property_flags = fallback_flags;
}
if (memory_type_index == UINT32_MAX) {
ctx->device->device.destroyBuffer(buf->buffer);
buf->size = 0;
throw vk::OutOfDeviceMemoryError("No suitable memory type found");
}
try {
buf->device_memory = ctx->device->device.allocateMemory({ mem_req.size, memory_type_index });
} catch (const vk::SystemError& e) {
// Out of Host/Device memory, clean up buffer
ctx->device->device.destroyBuffer(buf->buffer);
buf->size = 0;
throw e;
}
buf->ptr = nullptr;
if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
buf->ptr = ctx->device->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE);
}
ctx->device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0);
buf->ctx = ctx;
buf->device = ctx->device;
#ifdef GGML_VULKAN_DEBUG
std::cerr << "Created buffer " << buf->buffer << std::endl;
#endif
return buf;
}
static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
try {
return ggml_vk_create_buffer(ctx, size, req_flags, fallback_flags);
} catch (const vk::SystemError& e) {
std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
throw e;
}
}
static vk_buffer ggml_vk_create_buffer_device(ggml_backend_vk_context * ctx, size_t size) {
vk_buffer buf;
try {
if (ctx->device->uma) {
// Fall back to host memory type
buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
} else {
buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
}
} catch (const vk::SystemError& e) {
std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
throw e;
}
return buf;
}
static void ggml_vk_destroy_buffer(vk_buffer& buf) {
buf.reset();
}
static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
return { buf, 0, VK_WHOLE_SIZE };
}
static void ggml_vk_sync_buffers(vk_context * ctx) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_sync_buffers()" << std::endl;
#endif
const std::vector<vk::MemoryBarrier> mem_barriers{ { { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite }, { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite } } };
ctx->s->buffer.pipelineBarrier(
ctx->q->stage_flags,
ctx->q->stage_flags,
{},
mem_barriers,
{},
{}
);
}
static void ggml_vk_wait_events(vk_context * ctx, std::vector<vk::Event>&& events) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_wait_events()" << std::endl;
#endif
if (events.empty()) {
return;
}
ctx->s->buffer.waitEvents(
events,
ctx->q->stage_flags,
ctx->q->stage_flags,
{},
{},
{}
);
}
static bool ggml_vk_build_shader(ggml_type type) {
switch(type) {
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
return true;
default:
return false;
}
}
static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_load_shaders(" << ctx->name << ")" << std::endl;
#endif
const std::shared_ptr<vk_device> device = ctx->device;
// mulmat
std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
std::initializer_list<uint32_t> warptile_m = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
std::initializer_list<uint32_t> warptile_s = { device->subgroup_size, 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
std::initializer_list<uint32_t> warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
std::initializer_list<uint32_t> warptile_mmq_m = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
std::initializer_list<uint32_t> warptile_mmq_s = { device->subgroup_size, 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 };
std::array<uint32_t, 3> m_wg_denoms = { 64, 64, 1 };
std::array<uint32_t, 3> s_wg_denoms = { 32, 32, 1 };
uint32_t l_align = 128;
uint32_t m_align = 64;
uint32_t s_align = 32;
ctx->device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();