Skip to content
Navigation Menu
{{ message }}
forked from ggml-org/llama.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtest-alloc.cpp
More file actions
608 lines (512 loc) · 22.1 KB
/
Copy pathtest-alloc.cpp
File metadata and controls
608 lines (512 loc) · 22.1 KB
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
#include <ggml-alloc.h>
#include <ggml-backend-impl.h>
#include <ggml-cpp.h>
#include <ggml-impl.h>
#include <ggml.h>
#include <algorithm>
#include <exception>
#include <memory>
#include <vector>
//
// dummy backend with configurable max_buffer_size, tracks allocations
uint8_t * const alloc_base = (uint8_t *) 16;
struct dummy_backend_context {
size_t max_buffer_size = 64;
size_t alignment = 8;
ggml_backend_buffer_i buffer_interface;
std::vector<ggml_backend_buffer_t> buffers;
size_t allocated_total() const {
size_t n = 0;
for (ggml_backend_buffer_t buf : buffers) {
n += ggml_backend_buffer_get_size(buf);
}
return n;
}
};
// ggml_backend_buffer_type interface
static const char * dummy_backend_buffer_type_get_name(ggml_backend_buffer_type_t) {
return "dummy_buffer_type";
}
static ggml_backend_buffer_t dummy_backend_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
dummy_backend_context * ctx = (dummy_backend_context *) buft->context;
ggml_backend_buffer_t & buffer = ctx->buffers.emplace_back();
buffer = ggml_backend_buffer_init(buft, ctx->buffer_interface, ctx, size);
return buffer;
}
static size_t dummy_backend_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
dummy_backend_context * ctx = (dummy_backend_context *) buft->context;
return ctx->alignment;
}
static size_t dummy_backend_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
dummy_backend_context * ctx = (dummy_backend_context *) buft->context;
return ctx->max_buffer_size;
}
static bool dummy_backend_buffer_type_is_host(ggml_backend_buffer_type_t) {
return true;
}
// ggml_backend_buffer interface
static void dummy_backend_buffer_free_buffer(ggml_backend_buffer_t buffer) {
dummy_backend_context * ctx = (dummy_backend_context *) buffer->context;
auto i = std::find(ctx->buffers.begin(), ctx->buffers.end(), buffer);
GGML_ASSERT(i != ctx->buffers.end());
ctx->buffers.erase(i);
}
static void * dummy_backend_buffer_get_base(ggml_backend_buffer_t) {
return alloc_base;
}
static ggml_status dummy_backend_buffer_init_tensor(ggml_backend_buffer_t, ggml_tensor *) {
return GGML_STATUS_SUCCESS;
}
static void dummy_backend_buffer_memset_tensor(ggml_backend_buffer_t, ggml_tensor *, uint8_t, size_t, size_t) {}
static void dummy_backend_buffer_set_tensor(ggml_backend_buffer_t, ggml_tensor *, const void *, size_t, size_t) {}
static void dummy_backend_buffer_get_tensor(ggml_backend_buffer_t, const ggml_tensor *, void *, size_t, size_t) {}
static void dummy_backend_buffer_clear(ggml_backend_buffer_t, uint8_t) {}
// dummy_backend (not really a full backend, just provides what gallocr needs)
struct dummy_backend {
std::unique_ptr<dummy_backend_context> context;
ggml_backend_buffer_type buffer_type;
};
static dummy_backend dummy_backend_init(size_t max_buffer_size, size_t alignment = 8) {
dummy_backend b{};
b.context = std::make_unique<dummy_backend_context>();
b.context->alignment = alignment;
b.context->max_buffer_size = max_buffer_size;
b.context->buffer_interface.free_buffer = dummy_backend_buffer_free_buffer;
b.context->buffer_interface.get_base = dummy_backend_buffer_get_base;
b.context->buffer_interface.init_tensor = dummy_backend_buffer_init_tensor;
b.context->buffer_interface.memset_tensor = dummy_backend_buffer_memset_tensor;
b.context->buffer_interface.set_tensor = dummy_backend_buffer_set_tensor;
b.context->buffer_interface.get_tensor = dummy_backend_buffer_get_tensor;
b.context->buffer_interface.clear = dummy_backend_buffer_clear;
b.buffer_type.context = b.context.get();
b.buffer_type.iface.get_name = dummy_backend_buffer_type_get_name;
b.buffer_type.iface.alloc_buffer = dummy_backend_buffer_type_alloc_buffer;
b.buffer_type.iface.get_alignment = dummy_backend_buffer_type_get_alignment;
b.buffer_type.iface.get_max_size = dummy_backend_buffer_type_get_max_size;
b.buffer_type.iface.is_host = dummy_backend_buffer_type_is_host;
return b;
}
//
// test utilities
struct test_context_with_graph {
ggml_context * ctx;
ggml_cgraph * graph;
ggml_context_ptr ctx_ptr;
};
static test_context_with_graph make_context() {
ggml_init_params params{};
params.mem_size = 48 * ggml_tensor_overhead() + ggml_graph_overhead();
params.no_alloc = true;
ggml_context * ctx = ggml_init(params);
ggml_context_ptr ctx_ptr = ggml_context_ptr(ctx);
ggml_cgraph * graph = ggml_new_graph(ctx);
return { ctx, graph, std::move(ctx_ptr) };
}
static ggml_tensor * make_input_1d(ggml_context * ctx, int64_t n_elements) {
ggml_tensor * t = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
ggml_set_input(t);
return t;
}
static ggml_tensor * make_input_with_size(ggml_context * ctx, size_t size_bytes) {
GGML_ASSERT(size_bytes % 4 == 0);
return make_input_1d(ctx, size_bytes / 4);
}
static void assign_names(ggml_context * ctx, const char * prefix = "x") {
int i = 0;
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t; t = ggml_get_next_tensor(ctx, t)) {
ggml_format_name(t, "%s%d", prefix, i++);
}
}
static int get_leaf_id(ggml_cgraph * graph, const char * tensor_name) {
for (int i = 0; i < graph->n_leafs; ++i) {
if (strncmp(graph->leafs[i]->name, tensor_name, GGML_MAX_NAME) == 0) {
return i;
}
}
fprintf(stderr, "leaf not found: %s\n", tensor_name);
return -1;
}
static int get_node_id(ggml_cgraph * graph, const char * tensor_name) {
for (int i = 0; i < graph->n_nodes; ++i) {
if (strncmp(graph->nodes[i]->name, tensor_name, GGML_MAX_NAME) == 0) {
return i;
}
}
fprintf(stderr, "node not found: %s", tensor_name);
return -1;
}
static ggml_gallocr_ptr allocate_graph(ggml_cgraph * graph, ggml_tensor * out, ggml_backend_buffer_type_t buft) {
ggml_set_output(out);
ggml_build_forward_expand(graph, out);
ggml_gallocr_ptr galloc = ggml_gallocr_ptr(ggml_gallocr_new(buft));
bool result = ggml_gallocr_alloc_graph(galloc.get(), graph);
GGML_ASSERT(result);
return galloc;
}
//
// correctness checks for result allocations
static void check_all_allocated(ggml_cgraph * graph) {
for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
ggml_tensor * t = ggml_graph_node(graph, i);
GGML_ASSERT(t->buffer != nullptr);
GGML_ASSERT(t->data != nullptr);
}
}
static void check_max_size(ggml_context * ctx) {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t; t = ggml_get_next_tensor(ctx, t)) {
auto buft = ggml_backend_buffer_get_type(t->buffer);
size_t max_size = ggml_backend_buft_get_max_size(buft);
size_t offset = (char *) t->data - (char *) ggml_backend_buffer_get_base(t->buffer);
GGML_ASSERT(t->data >= ggml_backend_buffer_get_base(t->buffer));
GGML_ASSERT((size_t) offset + ggml_nbytes(t) <= max_size);
}
}
static bool can_reuse_memory(ggml_cgraph * graph, int current_i, ggml_tensor * current, ggml_tensor * other) {
if (other->flags & GGML_TENSOR_FLAG_OUTPUT) {
return false;
}
// Check if `other` is still "alive", ie. an input to any node after the `current` op
for (int i = current_i; i < ggml_graph_n_nodes(graph); ++i) {
ggml_tensor * t = ggml_graph_node(graph, i);
for (int s = 0; s < GGML_MAX_SRC; s++) {
if (t == current && ggml_op_can_inplace(t->op)) {
continue;
}
if (t->src[s] == other) {
return false;
}
if (t->src[s] && t->src[s]->view_src == other) {
return false;
}
}
}
return true;
}
static bool memory_overlap(ggml_tensor * a, ggml_tensor * b) {
if (a->buffer != b->buffer) {
return false;
}
int64_t a0 = (int64_t) a->data;
int64_t a1 = a0 + ggml_nbytes(a);
int64_t b0 = (int64_t) b->data;
int64_t b1 = b0 + ggml_nbytes(b);
return a1 > b0 && b1 > a0;
}
static ggml_tensor * get_view_source(ggml_tensor * t) {
while (t->view_src) {
t = t->view_src;
}
return t;
}
static void check_no_overlap(ggml_cgraph * graph) {
for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
for (int j = 0; j < i; ++j) {
ggml_tensor * t = ggml_graph_node(graph, i);
ggml_tensor * o = ggml_graph_node(graph, j);
GGML_ASSERT(t != o);
if (get_view_source(t) == get_view_source(o)) {
continue;
}
if (memory_overlap(t, o)) {
GGML_ASSERT(can_reuse_memory(graph, i, t, o));
}
}
}
}
//
// test cases
// Scenario where the first backend buffer is completely exhausted and there are further
// tensors which require a second buffer
static void test_max_size_too_many_tensors() {
dummy_backend backend = dummy_backend_init(16);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[7];
x[0] = make_input_with_size(ctx, 8);
x[1] = make_input_with_size(ctx, 8);
x[2] = make_input_with_size(ctx, 8);
x[3] = ggml_mul(ctx, x[0], x[1]);
x[4] = ggml_add(ctx, x[1], x[2]);
x[5] = ggml_add(ctx, x[3], x[0]);
x[6] = ggml_add(ctx, x[4], x[5]);
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[6], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
check_max_size(ctx);
GGML_ASSERT(backend.context->allocated_total() <= 16 + 16);
}
// Scenario where there is some space left in the first buffer, but not enough to accommodate
// a larger tensor, so a second buffer is required
static void test_max_size_tensor_too_large() {
dummy_backend backend = dummy_backend_init(32);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[3];
x[0] = make_input_with_size(ctx, 16); // chunk 0, [0 , 16)
x[1] = make_input_with_size(ctx, 8); // chunk 0, [16, 24)
x[2] = ggml_concat(ctx, x[0], x[1], 0); // chunk 1, [0 , 24)
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[2], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
check_max_size(ctx);
GGML_ASSERT(backend.context->allocated_total() <= 32 + 24);
}
// Scenario where a single tensor exceeds the max buffer size - in this case the allocator
// should try to create a bigger buffer anyway, and wait for the backend to throw an error.
// Backends may report an artificially lower max size in some cases for compatibility reasons.
static void test_tensor_larger_than_max_size() {
dummy_backend backend = dummy_backend_init(16);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[2];
x[0] = make_input_with_size(ctx, 24);
x[1] = ggml_scale(ctx, x[0], 2.0f);
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[1], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
GGML_ASSERT(backend.context->allocated_total() == 24);
}
// This test assumes a max of 16 buffer chunks, and tries to allocate tensors that would
// require more. Expectation is that the last buffer should grow to fit everything,
// leaving it to the backend to error out if it can't allocate that much.
static void test_not_enough_chunks() {
const int max_chunks = 16;
const int max_size = 8;
dummy_backend backend = dummy_backend_init(max_size);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[max_chunks + 1];
for (int i = 0; i < max_chunks + 1; ++i) {
x[i] = make_input_with_size(ctx, max_size);
}
ggml_tensor * acc = x[0];
for (int i = 0; i < max_chunks; ++i) {
acc = ggml_add(ctx, acc, x[i + 1]);
}
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, acc, &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
GGML_ASSERT(backend.context->allocated_total() > max_chunks * max_size);
}
// Fill up leftover unallocated space of a chunk after allocating a large tensor that
// requires a new chunk.
static void test_fill_leftover_space() {
dummy_backend backend = dummy_backend_init(16);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[4];
x[0] = make_input_with_size(ctx, 8);
x[1] = ggml_pad(ctx, x[0], 2, 0, 0, 0);
x[3] = ggml_mean(ctx, x[1]);
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[3], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
check_max_size(ctx);
GGML_ASSERT(backend.context->allocated_total() <= 12 + 16);
}
// Check that views don't require any extra memory
static void test_view_inplace() {
dummy_backend backend = dummy_backend_init(32);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[6];
x[0] = make_input_1d(ctx, 4); // chunk 0, [0, 16)
x[1] = ggml_reshape_2d(ctx, x[0], 2, 2); // view of x0
x[2] = ggml_permute(ctx, x[1], 1, 0, 2, 3); // view of x0
x[3] = ggml_view_1d(ctx, x[2], 2, 4); // view of x0
x[4] = make_input_1d(ctx, 2); // chunk 0, [16, 24)
x[5] = ggml_add(ctx, x[3], x[4]); // reuse (inplace add)
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[5], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
check_max_size(ctx);
GGML_ASSERT(backend.context->allocated_total() <= 24);
}
static void test_reuse_and_free() {
dummy_backend backend = dummy_backend_init(40);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[9];
x[0] = make_input_with_size(ctx, 24);
x[1] = make_input_with_size(ctx, 8);
x[2] = make_input_with_size(ctx, 8);
x[3] = ggml_add(ctx, x[1], x[2]); // reuse, free x2
x[4] = ggml_pad(ctx, x[0], 2, 0, 0, 0); // alloc new buffer, free x0
x[5] = ggml_scale(ctx, x[4], 2.0f); // alloc from free block
x[6] = ggml_add(ctx, x[4], x[5]); // reuse, free x5
x[7] = ggml_view_1d(ctx, x[6], 2, 8); // view
x[8] = ggml_add(ctx, x[3], x[7]); // reuse
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[8], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
check_max_size(ctx);
GGML_ASSERT(backend.context->allocated_total() <= 40 + 32 + 32);
}
static void test_merge_free_block(size_t max_buffer_size) {
dummy_backend backend = dummy_backend_init(max_buffer_size);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[9];
x[0] = make_input_with_size(ctx, 16);
x[1] = make_input_with_size(ctx, 16);
x[2] = make_input_with_size(ctx, 16);
x[3] = ggml_mean(ctx, x[0]);
x[4] = ggml_mean(ctx, x[1]);
x[5] = ggml_pad(ctx, x[2], 2, 0, 0, 0);
x[6] = ggml_add(ctx, x[3], x[4]);
x[7] = ggml_pad(ctx, x[6], 5, 0, 0, 0);
x[8] = ggml_add(ctx, x[5], x[7]);
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[8], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
check_max_size(ctx);
GGML_ASSERT(backend.context->allocated_total() <= 32 + 32 + 24);
}
// Check that previously allocated but freed memory is preferred over allocating
// additional memory, even if the remaining space in a chunk would match tensor size better
static void test_prefer_already_allocated_memory() {
dummy_backend backend = dummy_backend_init(32, /*align*/ 4);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[3];
x[0] = make_input_with_size(ctx, 24); // [24b][8b unused]
x[1] = ggml_mean(ctx, x[0]); // [24b free][4b][4b unused]
x[2] = ggml_mean(ctx, x[1]); // should be allocated in the 24b block
assign_names(ctx);
ggml_gallocr_ptr galloc = allocate_graph(graph, x[2], &backend.buffer_type);
check_all_allocated(graph);
check_no_overlap(graph);
GGML_ASSERT(backend.context->allocated_total() <= 28);
}
// test for allocating on multiple devices with some tensors in the graph
// allocated externally (not by gallocr).
static void test_multiple_buffer_types() {
dummy_backend backend_a = dummy_backend_init(32);
dummy_backend backend_b = dummy_backend_init(SIZE_MAX);
auto [ctx_a, _a, ctx_a_ptr] = make_context();
auto [ctx_b, _b, ctx_b_ptr] = make_context();
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * a[2];
a[0] = make_input_with_size(ctx_a, 16);
a[1] = make_input_with_size(ctx_a, 16);
assign_names(ctx_a, "a");
ggml_tensor * b[2];
b[0] = make_input_with_size(ctx_b, 24);
b[1] = make_input_with_size(ctx_b, 4);
assign_names(ctx_b, "b");
ggml_tensor * x[9];
x[0] = make_input_with_size(ctx, 16);
x[1] = ggml_mul(ctx, x[0], a[0]);
x[2] = ggml_pad(ctx, x[1], 2, 0, 0, 0);
x[3] = ggml_mul(ctx, x[2], b[0]);
x[4] = ggml_mean(ctx, x[3]);
x[5] = ggml_add(ctx, x[4], b[1]);
x[6] = ggml_pad(ctx, x[5], 3, 0, 0, 0);
x[7] = ggml_add(ctx, x[6], a[1]);
x[8] = ggml_scale(ctx, x[7], 2.0f);
assign_names(ctx, "x");
ggml_backend_buffer_ptr buf_a(ggml_backend_alloc_ctx_tensors_from_buft(ctx_a, &backend_a.buffer_type));
ggml_backend_buffer_ptr buf_b(ggml_backend_alloc_ctx_tensors_from_buft(ctx_b, &backend_b.buffer_type));
ggml_backend_buffer_type_t bufts[2] = { &backend_a.buffer_type, &backend_b.buffer_type };
// assign buffer types manually to avoid extra complexity from backend scheduler
ggml_set_output(x[8]);
ggml_build_forward_expand(graph, x[8]);
GGML_ASSERT(graph->n_leafs == 5);
int leaf_buffer_ids[5];
leaf_buffer_ids[get_leaf_id(graph, "a0")] = 0;
leaf_buffer_ids[get_leaf_id(graph, "a1")] = 0;
leaf_buffer_ids[get_leaf_id(graph, "b0")] = 1;
leaf_buffer_ids[get_leaf_id(graph, "b1")] = 1;
leaf_buffer_ids[get_leaf_id(graph, "x0")] = 0;
GGML_ASSERT(graph->n_nodes == 8);
int node_buffer_ids[8];
node_buffer_ids[get_node_id(graph, "x1")] = 0;
node_buffer_ids[get_node_id(graph, "x2")] = 0;
node_buffer_ids[get_node_id(graph, "x3")] = 1;
node_buffer_ids[get_node_id(graph, "x4")] = 1;
node_buffer_ids[get_node_id(graph, "x5")] = 1;
node_buffer_ids[get_node_id(graph, "x6")] = 1;
node_buffer_ids[get_node_id(graph, "x7")] = 0;
node_buffer_ids[get_node_id(graph, "x8")] = 0;
ggml_gallocr_ptr galloc(ggml_gallocr_new_n(bufts, 2));
ggml_gallocr_reserve_n(galloc.get(), graph, node_buffer_ids, leaf_buffer_ids);
ggml_gallocr_alloc_graph(galloc.get(), graph);
check_all_allocated(graph);
check_no_overlap(graph);
check_max_size(ctx);
GGML_ASSERT(backend_a.context->allocated_total() <= 32 + 32 + 24);
GGML_ASSERT(backend_b.context->allocated_total() <= 32 + 24);
}
static void test_buffer_size_zero() {
dummy_backend backend_a = dummy_backend_init(SIZE_MAX);
dummy_backend backend_b = dummy_backend_init(SIZE_MAX);
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[2];
x[0] = make_input_with_size(ctx, 16);
x[1] = ggml_scale(ctx, x[0], 2.0f);
ggml_set_output(x[1]);
ggml_build_forward_expand(graph, x[1]);
int leaf_buffer_ids[1] = { 0 };
int node_buffer_ids[1] = { 0 };
ggml_backend_buffer_type_t bufts[2] = { &backend_a.buffer_type, &backend_b.buffer_type };
ggml_gallocr_ptr galloc = ggml_gallocr_ptr(ggml_gallocr_new_n(bufts, 2));
bool res1 = ggml_gallocr_reserve_n(galloc.get(), graph, node_buffer_ids, leaf_buffer_ids);
bool res2 = ggml_gallocr_alloc_graph(galloc.get(), graph);
GGML_ASSERT(res1 && res2);
check_all_allocated(graph);
GGML_ASSERT(backend_a.context->allocated_total() == 16);
GGML_ASSERT(backend_b.context->allocated_total() == 0);
}
// Test re-using gallocr for a different graph. The new graph has the same
// total size, but one of the chunks is larger, so reallocation is required.
static void test_reallocation() {
dummy_backend backend = dummy_backend_init(32, /*align*/ 4);
ggml_gallocr_ptr galloc;
{
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[4];
x[0] = make_input_with_size(ctx, 24);
x[1] = make_input_with_size(ctx, 16);
x[2] = ggml_view_1d(ctx, x[0], 4, 0);
x[3] = ggml_add(ctx, x[2], x[1]);
assign_names(ctx);
galloc = allocate_graph(graph, x[3], &backend.buffer_type);
check_all_allocated(graph);
GGML_ASSERT(backend.context->allocated_total() == 40);
}
{
auto [ctx, graph, ctx_ptr] = make_context();
ggml_tensor * x[3];
x[0] = make_input_with_size(ctx, 20);
x[1] = make_input_with_size(ctx, 20);
x[2] = ggml_add(ctx, x[0], x[1]);
assign_names(ctx);
ggml_set_output(x[2]);
ggml_build_forward_expand(graph, x[2]);
bool result = ggml_gallocr_alloc_graph(galloc.get(), graph);
GGML_ASSERT(result);
check_all_allocated(graph);
GGML_ASSERT(backend.context->allocated_total() == 40);
}
}
static void run(const char * name, void (*f)()) {
printf("%s ", name);
fflush(stdout);
f();
printf("PASSED\n");
}
int main() {
run("test_max_size_too_many_tensors", test_max_size_too_many_tensors);
run("test_max_size_tensor_too_large", test_max_size_tensor_too_large);
run("test_tensor_larger_than_max_size", test_tensor_larger_than_max_size);
run("test_not_enough_chunks", test_not_enough_chunks);
run("test_fill_leftover_space", test_fill_leftover_space);
run("test_view_inplace", test_view_inplace);
run("test_reuse_and_free", test_reuse_and_free);
run("test_merge_free_block(32)", []() { test_merge_free_block(32); });
run("test_merge_free_block(SIZE_MAX)", []() { test_merge_free_block(SIZE_MAX); });
run("test_prefer_already_allocated_memory", test_prefer_already_allocated_memory);
run("test_multiple_buffer_types", test_multiple_buffer_types);
run("test_buffer_size_zero", test_buffer_size_zero);
run("test_reallocation", test_reallocation);
return 0;
}
You can’t perform that action at this time.