| 1 | #include <ggml-alloc.h> |
| 2 | #include <ggml-backend-impl.h> |
| 3 | #include <ggml-cpp.h> |
| 4 | #include <ggml-impl.h> |
| 5 | #include <ggml.h> |
| 6 | |
| 7 | #include <algorithm> |
| 8 | #include <exception> |
| 9 | #include <memory> |
| 10 | #include <vector> |
| 11 | |
| 12 | // |
| 13 | // dummy backend with configurable max_buffer_size, tracks allocations |
| 14 | |
| 15 | uint8_t * const alloc_base = (uint8_t *) 16; |
| 16 | |
| 17 | struct dummy_backend_context { |
| 18 | size_t max_buffer_size = 64; |
| 19 | size_t alignment = 8; |
| 20 | |
| 21 | ggml_backend_buffer_i buffer_interface; |
| 22 | std::vector<ggml_backend_buffer_t> buffers; |
| 23 | |
| 24 | size_t allocated_total() const { |
| 25 | size_t n = 0; |
| 26 | for (ggml_backend_buffer_t buf : buffers) { |
| 27 | n += ggml_backend_buffer_get_size(buffer: buf); |
| 28 | } |
| 29 | return n; |
| 30 | } |
| 31 | }; |
| 32 | |
| 33 | // ggml_backend_buffer_type interface |
| 34 | |
| 35 | static const char * dummy_backend_buffer_type_get_name(ggml_backend_buffer_type_t) { |
| 36 | return "dummy_buffer_type" ; |
| 37 | } |
| 38 | |
| 39 | static ggml_backend_buffer_t dummy_backend_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { |
| 40 | dummy_backend_context * ctx = (dummy_backend_context *) buft->context; |
| 41 | ggml_backend_buffer_t & buffer = ctx->buffers.emplace_back(); |
| 42 | buffer = ggml_backend_buffer_init(buft, iface: ctx->buffer_interface, context: ctx, size); |
| 43 | return buffer; |
| 44 | } |
| 45 | |
| 46 | static size_t dummy_backend_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { |
| 47 | dummy_backend_context * ctx = (dummy_backend_context *) buft->context; |
| 48 | return ctx->alignment; |
| 49 | } |
| 50 | |
| 51 | static size_t dummy_backend_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { |
| 52 | dummy_backend_context * ctx = (dummy_backend_context *) buft->context; |
| 53 | return ctx->max_buffer_size; |
| 54 | } |
| 55 | |
| 56 | static bool dummy_backend_buffer_type_is_host(ggml_backend_buffer_type_t) { |
| 57 | return true; |
| 58 | } |
| 59 | |
| 60 | // ggml_backend_buffer interface |
| 61 | |
| 62 | static void dummy_backend_buffer_free_buffer(ggml_backend_buffer_t buffer) { |
| 63 | dummy_backend_context * ctx = (dummy_backend_context *) buffer->context; |
| 64 | |
| 65 | auto i = std::find(first: ctx->buffers.begin(), last: ctx->buffers.end(), val: buffer); |
| 66 | GGML_ASSERT(i != ctx->buffers.end()); |
| 67 | ctx->buffers.erase(position: i); |
| 68 | } |
| 69 | |
| 70 | static void * dummy_backend_buffer_get_base(ggml_backend_buffer_t) { |
| 71 | return alloc_base; |
| 72 | } |
| 73 | |
| 74 | static ggml_status dummy_backend_buffer_init_tensor(ggml_backend_buffer_t, ggml_tensor *) { |
| 75 | return GGML_STATUS_SUCCESS; |
| 76 | } |
| 77 | |
| 78 | static void dummy_backend_buffer_memset_tensor(ggml_backend_buffer_t, ggml_tensor *, uint8_t, size_t, size_t) {} |
| 79 | |
| 80 | static void dummy_backend_buffer_set_tensor(ggml_backend_buffer_t, ggml_tensor *, const void *, size_t, size_t) {} |
| 81 | |
| 82 | static void dummy_backend_buffer_get_tensor(ggml_backend_buffer_t, const ggml_tensor *, void *, size_t, size_t) {} |
| 83 | |
| 84 | static void dummy_backend_buffer_clear(ggml_backend_buffer_t, uint8_t) {} |
| 85 | |
| 86 | // dummy_backend (not really a full backend, just provides what gallocr needs) |
| 87 | |
| 88 | struct dummy_backend { |
| 89 | std::unique_ptr<dummy_backend_context> context; |
| 90 | ggml_backend_buffer_type buffer_type; |
| 91 | }; |
| 92 | |
| 93 | static dummy_backend dummy_backend_init(size_t max_buffer_size, size_t alignment = 8) { |
| 94 | dummy_backend b{}; |
| 95 | b.context = std::make_unique<dummy_backend_context>(); |
| 96 | b.context->alignment = alignment; |
| 97 | b.context->max_buffer_size = max_buffer_size; |
| 98 | |
| 99 | b.context->buffer_interface.free_buffer = dummy_backend_buffer_free_buffer; |
| 100 | b.context->buffer_interface.get_base = dummy_backend_buffer_get_base; |
| 101 | b.context->buffer_interface.init_tensor = dummy_backend_buffer_init_tensor; |
| 102 | b.context->buffer_interface.memset_tensor = dummy_backend_buffer_memset_tensor; |
| 103 | b.context->buffer_interface.set_tensor = dummy_backend_buffer_set_tensor; |
| 104 | b.context->buffer_interface.get_tensor = dummy_backend_buffer_get_tensor; |
| 105 | b.context->buffer_interface.clear = dummy_backend_buffer_clear; |
| 106 | |
| 107 | b.buffer_type.context = b.context.get(); |
| 108 | b.buffer_type.iface.get_name = dummy_backend_buffer_type_get_name; |
| 109 | b.buffer_type.iface.alloc_buffer = dummy_backend_buffer_type_alloc_buffer; |
| 110 | b.buffer_type.iface.get_alignment = dummy_backend_buffer_type_get_alignment; |
| 111 | b.buffer_type.iface.get_max_size = dummy_backend_buffer_type_get_max_size; |
| 112 | b.buffer_type.iface.is_host = dummy_backend_buffer_type_is_host; |
| 113 | return b; |
| 114 | } |
| 115 | |
| 116 | // |
| 117 | // test utilities |
| 118 | |
| 119 | struct test_context_with_graph { |
| 120 | ggml_context * ctx; |
| 121 | ggml_cgraph * graph; |
| 122 | ggml_context_ptr ctx_ptr; |
| 123 | }; |
| 124 | |
| 125 | static test_context_with_graph make_context() { |
| 126 | ggml_init_params params{}; |
| 127 | params.mem_size = 48 * ggml_tensor_overhead() + ggml_graph_overhead(); |
| 128 | params.no_alloc = true; |
| 129 | |
| 130 | ggml_context * ctx = ggml_init(params); |
| 131 | ggml_context_ptr ctx_ptr = ggml_context_ptr(ctx); |
| 132 | ggml_cgraph * graph = ggml_new_graph(ctx); |
| 133 | return { .ctx: ctx, .graph: graph, .ctx_ptr: std::move(ctx_ptr) }; |
| 134 | } |
| 135 | |
| 136 | static ggml_tensor * make_input_1d(ggml_context * ctx, int64_t n_elements) { |
| 137 | ggml_tensor * t = ggml_new_tensor_1d(ctx, type: GGML_TYPE_F32, ne0: n_elements); |
| 138 | ggml_set_input(tensor: t); |
| 139 | return t; |
| 140 | } |
| 141 | |
| 142 | static ggml_tensor * make_input_with_size(ggml_context * ctx, size_t size_bytes) { |
| 143 | GGML_ASSERT(size_bytes % 4 == 0); |
| 144 | return make_input_1d(ctx, n_elements: size_bytes / 4); |
| 145 | } |
| 146 | |
| 147 | static void assign_names(ggml_context * ctx, const char * prefix = "x" ) { |
| 148 | int i = 0; |
| 149 | for (ggml_tensor * t = ggml_get_first_tensor(ctx); t; t = ggml_get_next_tensor(ctx, tensor: t)) { |
| 150 | ggml_format_name(tensor: t, fmt: "%s%d" , prefix, i++); |
| 151 | } |
| 152 | } |
| 153 | |
| 154 | static int get_leaf_id(ggml_cgraph * graph, const char * tensor_name) { |
| 155 | for (int i = 0; i < graph->n_leafs; ++i) { |
| 156 | if (strncmp(s1: graph->leafs[i]->name, s2: tensor_name, GGML_MAX_NAME) == 0) { |
| 157 | return i; |
| 158 | } |
| 159 | } |
| 160 | fprintf(stderr, format: "leaf not found: %s\n" , tensor_name); |
| 161 | return -1; |
| 162 | } |
| 163 | |
| 164 | static int get_node_id(ggml_cgraph * graph, const char * tensor_name) { |
| 165 | for (int i = 0; i < graph->n_nodes; ++i) { |
| 166 | if (strncmp(s1: graph->nodes[i]->name, s2: tensor_name, GGML_MAX_NAME) == 0) { |
| 167 | return i; |
| 168 | } |
| 169 | } |
| 170 | fprintf(stderr, format: "node not found: %s" , tensor_name); |
| 171 | return -1; |
| 172 | } |
| 173 | |
| 174 | static ggml_gallocr_ptr allocate_graph(ggml_cgraph * graph, ggml_tensor * out, ggml_backend_buffer_type_t buft) { |
| 175 | ggml_set_output(tensor: out); |
| 176 | ggml_build_forward_expand(cgraph: graph, tensor: out); |
| 177 | |
| 178 | ggml_gallocr_ptr galloc = ggml_gallocr_ptr(ggml_gallocr_new(buft)); |
| 179 | bool result = ggml_gallocr_alloc_graph(galloc: galloc.get(), graph); |
| 180 | GGML_ASSERT(result); |
| 181 | return galloc; |
| 182 | } |
| 183 | |
| 184 | // |
| 185 | // correctness checks for result allocations |
| 186 | |
| 187 | static void check_all_allocated(ggml_cgraph * graph) { |
| 188 | for (int i = 0; i < ggml_graph_n_nodes(cgraph: graph); ++i) { |
| 189 | ggml_tensor * t = ggml_graph_node(cgraph: graph, i); |
| 190 | GGML_ASSERT(t->buffer != nullptr); |
| 191 | GGML_ASSERT(t->data != nullptr); |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | static void check_max_size(ggml_context * ctx) { |
| 196 | for (ggml_tensor * t = ggml_get_first_tensor(ctx); t; t = ggml_get_next_tensor(ctx, tensor: t)) { |
| 197 | auto buft = ggml_backend_buffer_get_type(buffer: t->buffer); |
| 198 | size_t max_size = ggml_backend_buft_get_max_size(buft); |
| 199 | size_t offset = (char *) t->data - (char *) ggml_backend_buffer_get_base(buffer: t->buffer); |
| 200 | GGML_ASSERT(t->data >= ggml_backend_buffer_get_base(t->buffer)); |
| 201 | GGML_ASSERT((size_t) offset + ggml_nbytes(t) <= max_size); |
| 202 | } |
| 203 | } |
| 204 | |
| 205 | static bool can_reuse_memory(ggml_cgraph * graph, int current_i, ggml_tensor * current, ggml_tensor * other) { |
| 206 | if (other->flags & GGML_TENSOR_FLAG_OUTPUT) { |
| 207 | return false; |
| 208 | } |
| 209 | // Check if `other` is still "alive", ie. an input to any node after the `current` op |
| 210 | for (int i = current_i; i < ggml_graph_n_nodes(cgraph: graph); ++i) { |
| 211 | ggml_tensor * t = ggml_graph_node(cgraph: graph, i); |
| 212 | for (int s = 0; s < GGML_MAX_SRC; s++) { |
| 213 | if (t == current && ggml_op_can_inplace(op: t->op)) { |
| 214 | continue; |
| 215 | } |
| 216 | if (t->src[s] == other) { |
| 217 | return false; |
| 218 | } |
| 219 | if (t->src[s] && t->src[s]->view_src == other) { |
| 220 | return false; |
| 221 | } |
| 222 | } |
| 223 | } |
| 224 | return true; |
| 225 | } |
| 226 | |
| 227 | static bool memory_overlap(ggml_tensor * a, ggml_tensor * b) { |
| 228 | if (a->buffer != b->buffer) { |
| 229 | return false; |
| 230 | } |
| 231 | int64_t a0 = (int64_t) a->data; |
| 232 | int64_t a1 = a0 + ggml_nbytes(tensor: a); |
| 233 | int64_t b0 = (int64_t) b->data; |
| 234 | int64_t b1 = b0 + ggml_nbytes(tensor: b); |
| 235 | return a1 > b0 && b1 > a0; |
| 236 | } |
| 237 | |
| 238 | static ggml_tensor * get_view_source(ggml_tensor * t) { |
| 239 | while (t->view_src) { |
| 240 | t = t->view_src; |
| 241 | } |
| 242 | return t; |
| 243 | } |
| 244 | |
| 245 | static void check_no_overlap(ggml_cgraph * graph) { |
| 246 | for (int i = 0; i < ggml_graph_n_nodes(cgraph: graph); ++i) { |
| 247 | for (int j = 0; j < i; ++j) { |
| 248 | ggml_tensor * t = ggml_graph_node(cgraph: graph, i); |
| 249 | ggml_tensor * o = ggml_graph_node(cgraph: graph, i: j); |
| 250 | GGML_ASSERT(t != o); |
| 251 | |
| 252 | if (get_view_source(t) == get_view_source(t: o)) { |
| 253 | continue; |
| 254 | } |
| 255 | if (memory_overlap(a: t, b: o)) { |
| 256 | GGML_ASSERT(can_reuse_memory(graph, i, t, o)); |
| 257 | } |
| 258 | } |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | // |
| 263 | // test cases |
| 264 | |
| 265 | // Scenario where the first backend buffer is completely exhausted and there are further |
| 266 | // tensors which require a second buffer |
| 267 | static void test_max_size_too_many_tensors() { |
| 268 | dummy_backend backend = dummy_backend_init(max_buffer_size: 16); |
| 269 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 270 | |
| 271 | ggml_tensor * x[7]; |
| 272 | x[0] = make_input_with_size(ctx, size_bytes: 8); |
| 273 | x[1] = make_input_with_size(ctx, size_bytes: 8); |
| 274 | x[2] = make_input_with_size(ctx, size_bytes: 8); |
| 275 | x[3] = ggml_mul(ctx, a: x[0], b: x[1]); |
| 276 | x[4] = ggml_add(ctx, a: x[1], b: x[2]); |
| 277 | x[5] = ggml_add(ctx, a: x[3], b: x[0]); |
| 278 | x[6] = ggml_add(ctx, a: x[4], b: x[5]); |
| 279 | assign_names(ctx); |
| 280 | |
| 281 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[6], buft: &backend.buffer_type); |
| 282 | check_all_allocated(graph); |
| 283 | check_no_overlap(graph); |
| 284 | check_max_size(ctx); |
| 285 | GGML_ASSERT(backend.context->allocated_total() <= 16 + 16); |
| 286 | } |
| 287 | |
| 288 | // Scenario where there is some space left in the first buffer, but not enough to accomodate |
| 289 | // a larger tensor, so a second buffer is required |
| 290 | static void test_max_size_tensor_too_large() { |
| 291 | dummy_backend backend = dummy_backend_init(max_buffer_size: 32); |
| 292 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 293 | |
| 294 | ggml_tensor * x[3]; |
| 295 | x[0] = make_input_with_size(ctx, size_bytes: 16); // chunk 0, [0 , 16) |
| 296 | x[1] = make_input_with_size(ctx, size_bytes: 8); // chunk 0, [16, 24) |
| 297 | x[2] = ggml_concat(ctx, a: x[0], b: x[1], dim: 0); // chunk 1, [0 , 24) |
| 298 | assign_names(ctx); |
| 299 | |
| 300 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[2], buft: &backend.buffer_type); |
| 301 | check_all_allocated(graph); |
| 302 | check_no_overlap(graph); |
| 303 | check_max_size(ctx); |
| 304 | GGML_ASSERT(backend.context->allocated_total() <= 32 + 24); |
| 305 | } |
| 306 | |
| 307 | // Scenario where a single tensor exceeds the max buffer size - in this case the allocator |
| 308 | // should try to create a bigger buffer anyway, and wait for the backend to throw an error. |
| 309 | // Backends may report an artificially lower max size in some cases for compatibility reasons. |
| 310 | static void test_tensor_larger_than_max_size() { |
| 311 | dummy_backend backend = dummy_backend_init(max_buffer_size: 16); |
| 312 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 313 | |
| 314 | ggml_tensor * x[2]; |
| 315 | x[0] = make_input_with_size(ctx, size_bytes: 24); |
| 316 | x[1] = ggml_scale(ctx, a: x[0], s: 2.0f); |
| 317 | assign_names(ctx); |
| 318 | |
| 319 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[1], buft: &backend.buffer_type); |
| 320 | check_all_allocated(graph); |
| 321 | check_no_overlap(graph); |
| 322 | GGML_ASSERT(backend.context->allocated_total() == 24); |
| 323 | } |
| 324 | |
| 325 | // This test assumes a max of 16 buffer chunks, and tries to allocate tensors that would |
| 326 | // require more. Expectation is that the last buffer should grow to fit everything, |
| 327 | // leaving it to the backend to error out if it can't allocate that much. |
| 328 | static void test_not_enough_chunks() { |
| 329 | const int max_chunks = 16; |
| 330 | const int max_size = 8; |
| 331 | |
| 332 | dummy_backend backend = dummy_backend_init(max_buffer_size: max_size); |
| 333 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 334 | |
| 335 | ggml_tensor * x[max_chunks + 1]; |
| 336 | for (int i = 0; i < max_chunks + 1; ++i) { |
| 337 | x[i] = make_input_with_size(ctx, size_bytes: max_size); |
| 338 | } |
| 339 | ggml_tensor * acc = x[0]; |
| 340 | for (int i = 0; i < max_chunks; ++i) { |
| 341 | acc = ggml_add(ctx, a: acc, b: x[i + 1]); |
| 342 | } |
| 343 | assign_names(ctx); |
| 344 | |
| 345 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: acc, buft: &backend.buffer_type); |
| 346 | check_all_allocated(graph); |
| 347 | check_no_overlap(graph); |
| 348 | GGML_ASSERT(backend.context->allocated_total() > max_chunks * max_size); |
| 349 | } |
| 350 | |
| 351 | // Fill up leftover unallocated space of a chunk after allocating a large tensor that |
| 352 | // requires a new chunk. |
| 353 | static void test_fill_leftover_space() { |
| 354 | dummy_backend backend = dummy_backend_init(max_buffer_size: 16); |
| 355 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 356 | |
| 357 | ggml_tensor * x[4]; |
| 358 | x[0] = make_input_with_size(ctx, size_bytes: 8); |
| 359 | x[1] = ggml_pad(ctx, a: x[0], p0: 2, p1: 0, p2: 0, p3: 0); |
| 360 | x[3] = ggml_mean(ctx, a: x[1]); |
| 361 | assign_names(ctx); |
| 362 | |
| 363 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[3], buft: &backend.buffer_type); |
| 364 | check_all_allocated(graph); |
| 365 | check_no_overlap(graph); |
| 366 | check_max_size(ctx); |
| 367 | GGML_ASSERT(backend.context->allocated_total() <= 12 + 16); |
| 368 | } |
| 369 | |
| 370 | // Check that views don't require any extra memory |
| 371 | static void test_view_inplace() { |
| 372 | dummy_backend backend = dummy_backend_init(max_buffer_size: 32); |
| 373 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 374 | |
| 375 | ggml_tensor * x[6]; |
| 376 | x[0] = make_input_1d(ctx, n_elements: 4); // chunk 0, [0, 16) |
| 377 | x[1] = ggml_reshape_2d(ctx, a: x[0], ne0: 2, ne1: 2); // view of x0 |
| 378 | x[2] = ggml_permute(ctx, a: x[1], axis0: 1, axis1: 0, axis2: 2, axis3: 3); // view of x0 |
| 379 | x[3] = ggml_view_1d(ctx, a: x[2], ne0: 2, offset: 4); // view of x0 |
| 380 | x[4] = make_input_1d(ctx, n_elements: 2); // chunk 0, [16, 24) |
| 381 | x[5] = ggml_add(ctx, a: x[3], b: x[4]); // reuse (inplace add) |
| 382 | assign_names(ctx); |
| 383 | |
| 384 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[5], buft: &backend.buffer_type); |
| 385 | check_all_allocated(graph); |
| 386 | check_no_overlap(graph); |
| 387 | check_max_size(ctx); |
| 388 | GGML_ASSERT(backend.context->allocated_total() <= 24); |
| 389 | } |
| 390 | |
| 391 | static void test_reuse_and_free() { |
| 392 | dummy_backend backend = dummy_backend_init(max_buffer_size: 40); |
| 393 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 394 | |
| 395 | ggml_tensor * x[9]; |
| 396 | x[0] = make_input_with_size(ctx, size_bytes: 24); |
| 397 | x[1] = make_input_with_size(ctx, size_bytes: 8); |
| 398 | x[2] = make_input_with_size(ctx, size_bytes: 8); |
| 399 | x[3] = ggml_add(ctx, a: x[1], b: x[2]); // reuse, free x2 |
| 400 | x[4] = ggml_pad(ctx, a: x[0], p0: 2, p1: 0, p2: 0, p3: 0); // alloc new buffer, free x0 |
| 401 | x[5] = ggml_scale(ctx, a: x[4], s: 2.0f); // alloc from free block |
| 402 | x[6] = ggml_add(ctx, a: x[4], b: x[5]); // reuse, free x5 |
| 403 | x[7] = ggml_view_1d(ctx, a: x[6], ne0: 2, offset: 8); // view |
| 404 | x[8] = ggml_add(ctx, a: x[3], b: x[7]); // reuse |
| 405 | assign_names(ctx); |
| 406 | |
| 407 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[8], buft: &backend.buffer_type); |
| 408 | check_all_allocated(graph); |
| 409 | check_no_overlap(graph); |
| 410 | check_max_size(ctx); |
| 411 | GGML_ASSERT(backend.context->allocated_total() <= 40 + 32 + 32); |
| 412 | } |
| 413 | |
| 414 | static void test_merge_free_block(size_t max_buffer_size) { |
| 415 | dummy_backend backend = dummy_backend_init(max_buffer_size); |
| 416 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 417 | |
| 418 | ggml_tensor * x[9]; |
| 419 | x[0] = make_input_with_size(ctx, size_bytes: 16); |
| 420 | x[1] = make_input_with_size(ctx, size_bytes: 16); |
| 421 | x[2] = make_input_with_size(ctx, size_bytes: 16); |
| 422 | x[3] = ggml_mean(ctx, a: x[0]); |
| 423 | x[4] = ggml_mean(ctx, a: x[1]); |
| 424 | x[5] = ggml_pad(ctx, a: x[2], p0: 2, p1: 0, p2: 0, p3: 0); |
| 425 | x[6] = ggml_add(ctx, a: x[3], b: x[4]); |
| 426 | x[7] = ggml_pad(ctx, a: x[6], p0: 5, p1: 0, p2: 0, p3: 0); |
| 427 | x[8] = ggml_add(ctx, a: x[5], b: x[7]); |
| 428 | assign_names(ctx); |
| 429 | |
| 430 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[8], buft: &backend.buffer_type); |
| 431 | check_all_allocated(graph); |
| 432 | check_no_overlap(graph); |
| 433 | check_max_size(ctx); |
| 434 | GGML_ASSERT(backend.context->allocated_total() <= 32 + 32 + 24); |
| 435 | } |
| 436 | |
| 437 | // Check that previously allocated but freed memory is preferred over allocating |
| 438 | // additional memory, even if the remaining space in a chunk would match tensor size better |
| 439 | static void test_prefer_already_allocated_memory() { |
| 440 | dummy_backend backend = dummy_backend_init(max_buffer_size: 32, /*align*/ alignment: 4); |
| 441 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 442 | |
| 443 | ggml_tensor * x[3]; |
| 444 | x[0] = make_input_with_size(ctx, size_bytes: 24); // [24b][8b unused] |
| 445 | x[1] = ggml_mean(ctx, a: x[0]); // [24b free][4b][4b unused] |
| 446 | x[2] = ggml_mean(ctx, a: x[1]); // should be allocated in the 24b block |
| 447 | assign_names(ctx); |
| 448 | |
| 449 | ggml_gallocr_ptr galloc = allocate_graph(graph, out: x[2], buft: &backend.buffer_type); |
| 450 | check_all_allocated(graph); |
| 451 | check_no_overlap(graph); |
| 452 | GGML_ASSERT(backend.context->allocated_total() <= 28); |
| 453 | } |
| 454 | |
| 455 | // test for allocating on multiple devices with some tensors in the graph |
| 456 | // allocated externally (not by gallocr). |
| 457 | static void test_multiple_buffer_types() { |
| 458 | dummy_backend backend_a = dummy_backend_init(max_buffer_size: 32); |
| 459 | dummy_backend backend_b = dummy_backend_init(SIZE_MAX); |
| 460 | |
| 461 | auto [ctx_a, _a, ctx_a_ptr] = make_context(); |
| 462 | auto [ctx_b, _b, ctx_b_ptr] = make_context(); |
| 463 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 464 | |
| 465 | ggml_tensor * a[2]; |
| 466 | a[0] = make_input_with_size(ctx: ctx_a, size_bytes: 16); |
| 467 | a[1] = make_input_with_size(ctx: ctx_a, size_bytes: 16); |
| 468 | assign_names(ctx: ctx_a, prefix: "a" ); |
| 469 | |
| 470 | ggml_tensor * b[2]; |
| 471 | b[0] = make_input_with_size(ctx: ctx_b, size_bytes: 24); |
| 472 | b[1] = make_input_with_size(ctx: ctx_b, size_bytes: 4); |
| 473 | assign_names(ctx: ctx_b, prefix: "b" ); |
| 474 | |
| 475 | ggml_tensor * x[9]; |
| 476 | x[0] = make_input_with_size(ctx, size_bytes: 16); |
| 477 | x[1] = ggml_mul(ctx, a: x[0], b: a[0]); |
| 478 | x[2] = ggml_pad(ctx, a: x[1], p0: 2, p1: 0, p2: 0, p3: 0); |
| 479 | x[3] = ggml_mul(ctx, a: x[2], b: b[0]); |
| 480 | x[4] = ggml_mean(ctx, a: x[3]); |
| 481 | x[5] = ggml_add(ctx, a: x[4], b: b[1]); |
| 482 | x[6] = ggml_pad(ctx, a: x[5], p0: 3, p1: 0, p2: 0, p3: 0); |
| 483 | x[7] = ggml_add(ctx, a: x[6], b: a[1]); |
| 484 | x[8] = ggml_scale(ctx, a: x[7], s: 2.0f); |
| 485 | assign_names(ctx, prefix: "x" ); |
| 486 | |
| 487 | ggml_backend_buffer_ptr buf_a(ggml_backend_alloc_ctx_tensors_from_buft(ctx: ctx_a, buft: &backend_a.buffer_type)); |
| 488 | ggml_backend_buffer_ptr buf_b(ggml_backend_alloc_ctx_tensors_from_buft(ctx: ctx_b, buft: &backend_b.buffer_type)); |
| 489 | ggml_backend_buffer_type_t bufts[2] = { &backend_a.buffer_type, &backend_b.buffer_type }; |
| 490 | |
| 491 | // assign buffer types manually to avoid extra complexity from backend scheduler |
| 492 | ggml_set_output(tensor: x[8]); |
| 493 | ggml_build_forward_expand(cgraph: graph, tensor: x[8]); |
| 494 | |
| 495 | GGML_ASSERT(graph->n_leafs == 5); |
| 496 | int leaf_buffer_ids[5]; |
| 497 | leaf_buffer_ids[get_leaf_id(graph, tensor_name: "a0" )] = 0; |
| 498 | leaf_buffer_ids[get_leaf_id(graph, tensor_name: "a1" )] = 0; |
| 499 | leaf_buffer_ids[get_leaf_id(graph, tensor_name: "b0" )] = 1; |
| 500 | leaf_buffer_ids[get_leaf_id(graph, tensor_name: "b1" )] = 1; |
| 501 | leaf_buffer_ids[get_leaf_id(graph, tensor_name: "x0" )] = 0; |
| 502 | |
| 503 | GGML_ASSERT(graph->n_nodes == 8); |
| 504 | int node_buffer_ids[8]; |
| 505 | node_buffer_ids[get_node_id(graph, tensor_name: "x1" )] = 0; |
| 506 | node_buffer_ids[get_node_id(graph, tensor_name: "x2" )] = 0; |
| 507 | node_buffer_ids[get_node_id(graph, tensor_name: "x3" )] = 1; |
| 508 | node_buffer_ids[get_node_id(graph, tensor_name: "x4" )] = 1; |
| 509 | node_buffer_ids[get_node_id(graph, tensor_name: "x5" )] = 1; |
| 510 | node_buffer_ids[get_node_id(graph, tensor_name: "x6" )] = 1; |
| 511 | node_buffer_ids[get_node_id(graph, tensor_name: "x7" )] = 0; |
| 512 | node_buffer_ids[get_node_id(graph, tensor_name: "x8" )] = 0; |
| 513 | |
| 514 | ggml_gallocr_ptr galloc(ggml_gallocr_new_n(bufts, n_bufs: 2)); |
| 515 | ggml_gallocr_reserve_n(galloc: galloc.get(), graph, node_buffer_ids, leaf_buffer_ids); |
| 516 | ggml_gallocr_alloc_graph(galloc: galloc.get(), graph); |
| 517 | |
| 518 | check_all_allocated(graph); |
| 519 | check_no_overlap(graph); |
| 520 | check_max_size(ctx); |
| 521 | GGML_ASSERT(backend_a.context->allocated_total() <= 32 + 32 + 24); |
| 522 | GGML_ASSERT(backend_b.context->allocated_total() <= 32 + 24); |
| 523 | } |
| 524 | |
| 525 | static void test_buffer_size_zero() { |
| 526 | dummy_backend backend_a = dummy_backend_init(SIZE_MAX); |
| 527 | dummy_backend backend_b = dummy_backend_init(SIZE_MAX); |
| 528 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 529 | |
| 530 | ggml_tensor * x[2]; |
| 531 | x[0] = make_input_with_size(ctx, size_bytes: 16); |
| 532 | x[1] = ggml_scale(ctx, a: x[0], s: 2.0f); |
| 533 | |
| 534 | ggml_set_output(tensor: x[1]); |
| 535 | ggml_build_forward_expand(cgraph: graph, tensor: x[1]); |
| 536 | |
| 537 | int leaf_buffer_ids[1] = { 0 }; |
| 538 | int node_buffer_ids[1] = { 0 }; |
| 539 | |
| 540 | ggml_backend_buffer_type_t bufts[2] = { &backend_a.buffer_type, &backend_b.buffer_type }; |
| 541 | ggml_gallocr_ptr galloc = ggml_gallocr_ptr(ggml_gallocr_new_n(bufts, n_bufs: 2)); |
| 542 | bool res1 = ggml_gallocr_reserve_n(galloc: galloc.get(), graph, node_buffer_ids, leaf_buffer_ids); |
| 543 | bool res2 = ggml_gallocr_alloc_graph(galloc: galloc.get(), graph); |
| 544 | GGML_ASSERT(res1 && res2); |
| 545 | |
| 546 | check_all_allocated(graph); |
| 547 | GGML_ASSERT(backend_a.context->allocated_total() == 16); |
| 548 | GGML_ASSERT(backend_b.context->allocated_total() == 0); |
| 549 | } |
| 550 | |
| 551 | // Test re-using gallocr for a different graph. The new graph has the same |
| 552 | // total size, but one of the chunks is larger, so reallocation is required. |
| 553 | static void test_reallocation() { |
| 554 | dummy_backend backend = dummy_backend_init(max_buffer_size: 32, /*align*/ alignment: 4); |
| 555 | ggml_gallocr_ptr galloc; |
| 556 | { |
| 557 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 558 | ggml_tensor * x[4]; |
| 559 | x[0] = make_input_with_size(ctx, size_bytes: 24); |
| 560 | x[1] = make_input_with_size(ctx, size_bytes: 16); |
| 561 | x[2] = ggml_view_1d(ctx, a: x[0], ne0: 4, offset: 0); |
| 562 | x[3] = ggml_add(ctx, a: x[2], b: x[1]); |
| 563 | assign_names(ctx); |
| 564 | |
| 565 | galloc = allocate_graph(graph, out: x[3], buft: &backend.buffer_type); |
| 566 | check_all_allocated(graph); |
| 567 | GGML_ASSERT(backend.context->allocated_total() == 40); |
| 568 | } |
| 569 | { |
| 570 | auto [ctx, graph, ctx_ptr] = make_context(); |
| 571 | ggml_tensor * x[3]; |
| 572 | x[0] = make_input_with_size(ctx, size_bytes: 20); |
| 573 | x[1] = make_input_with_size(ctx, size_bytes: 20); |
| 574 | x[2] = ggml_add(ctx, a: x[0], b: x[1]); |
| 575 | assign_names(ctx); |
| 576 | ggml_set_output(tensor: x[2]); |
| 577 | ggml_build_forward_expand(cgraph: graph, tensor: x[2]); |
| 578 | |
| 579 | bool result = ggml_gallocr_alloc_graph(galloc: galloc.get(), graph); |
| 580 | GGML_ASSERT(result); |
| 581 | check_all_allocated(graph); |
| 582 | GGML_ASSERT(backend.context->allocated_total() == 40); |
| 583 | } |
| 584 | } |
| 585 | |
| 586 | static void run(const char * name, void (*f)()) { |
| 587 | printf(format: "%s " , name); |
| 588 | fflush(stdout); |
| 589 | f(); |
| 590 | printf(format: "PASSED\n" ); |
| 591 | } |
| 592 | |
| 593 | int main() { |
| 594 | run(name: "test_max_size_too_many_tensors" , f: test_max_size_too_many_tensors); |
| 595 | run(name: "test_max_size_tensor_too_large" , f: test_max_size_tensor_too_large); |
| 596 | run(name: "test_tensor_larger_than_max_size" , f: test_tensor_larger_than_max_size); |
| 597 | run(name: "test_not_enough_chunks" , f: test_not_enough_chunks); |
| 598 | run(name: "test_fill_leftover_space" , f: test_fill_leftover_space); |
| 599 | run(name: "test_view_inplace" , f: test_view_inplace); |
| 600 | run(name: "test_reuse_and_free" , f: test_reuse_and_free); |
| 601 | run(name: "test_merge_free_block(32)" , f: []() { test_merge_free_block(max_buffer_size: 32); }); |
| 602 | run(name: "test_merge_free_block(SIZE_MAX)" , f: []() { test_merge_free_block(SIZE_MAX); }); |
| 603 | run(name: "test_prefer_already_allocated_memory" , f: test_prefer_already_allocated_memory); |
| 604 | run(name: "test_multiple_buffer_types" , f: test_multiple_buffer_types); |
| 605 | run(name: "test_buffer_size_zero" , f: test_buffer_size_zero); |
| 606 | run(name: "test_reallocation" , f: test_reallocation); |
| 607 | return 0; |
| 608 | } |
| 609 | |