| 1 | #include "ggml-alloc.h" |
| 2 | #include "ggml-backend-impl.h" |
| 3 | #include "ggml.h" |
| 4 | #include "ggml-impl.h" |
| 5 | #include <assert.h> |
| 6 | #include <limits.h> |
| 7 | #include <stdarg.h> |
| 8 | #include <stdio.h> |
| 9 | #include <stdlib.h> |
| 10 | #include <string.h> |
| 11 | |
| 12 | #define MAX(a, b) ((a) > (b) ? (a) : (b)) |
| 13 | #define MAX_FREE_BLOCKS 256 |
| 14 | |
| 15 | //#define GGML_ALLOCATOR_DEBUG |
| 16 | |
| 17 | //#define AT_PRINTF(...) GGML_LOG_DEBUG(__VA_ARGS__) |
| 18 | #define AT_PRINTF(...) |
| 19 | |
| 20 | |
| 21 | static bool ggml_is_view(const struct ggml_tensor * t) { |
| 22 | return t->view_src != NULL; |
| 23 | } |
| 24 | |
| 25 | // ops that return true for this function must not use restrict pointers for their backend implementations |
| 26 | bool ggml_op_can_inplace(enum ggml_op op) { |
| 27 | switch (op) { |
| 28 | case GGML_OP_SCALE: |
| 29 | case GGML_OP_DIAG_MASK_ZERO: |
| 30 | case GGML_OP_DIAG_MASK_INF: |
| 31 | case GGML_OP_ADD: |
| 32 | case GGML_OP_ADD_ID: |
| 33 | case GGML_OP_ADD1: |
| 34 | case GGML_OP_SUB: |
| 35 | case GGML_OP_MUL: |
| 36 | case GGML_OP_DIV: |
| 37 | case GGML_OP_SQR: |
| 38 | case GGML_OP_SQRT: |
| 39 | case GGML_OP_LOG: |
| 40 | case GGML_OP_UNARY: |
| 41 | case GGML_OP_ROPE: |
| 42 | case GGML_OP_ROPE_BACK: |
| 43 | case GGML_OP_SILU_BACK: |
| 44 | case GGML_OP_RMS_NORM: |
| 45 | case GGML_OP_RMS_NORM_BACK: |
| 46 | case GGML_OP_SOFT_MAX: |
| 47 | case GGML_OP_SOFT_MAX_BACK: |
| 48 | return true; |
| 49 | |
| 50 | default: |
| 51 | return false; |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) { |
| 56 | assert(alignment && !(alignment & (alignment - 1))); // power of 2 |
| 57 | size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment; |
| 58 | return offset + align; |
| 59 | } |
| 60 | |
| 61 | // tallocr |
| 62 | |
| 63 | struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) { |
| 64 | void * base = ggml_backend_buffer_get_base(buffer); |
| 65 | size_t align = ggml_backend_buffer_get_alignment(buffer); |
| 66 | |
| 67 | assert(align && !(align & (align - 1))); // power of 2 |
| 68 | |
| 69 | struct ggml_tallocr talloc = (struct ggml_tallocr) { |
| 70 | /*.buffer = */ buffer, |
| 71 | /*.base = */ base, |
| 72 | /*.alignment = */ align, |
| 73 | /*.offset = */ aligned_offset(buffer: base, offset: 0, alignment: align), |
| 74 | }; |
| 75 | return talloc; |
| 76 | } |
| 77 | |
| 78 | enum ggml_status ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) { |
| 79 | size_t size = ggml_backend_buffer_get_alloc_size(buffer: talloc->buffer, tensor); |
| 80 | size = GGML_PAD(size, talloc->alignment); |
| 81 | |
| 82 | if (talloc->offset + size > ggml_backend_buffer_get_size(buffer: talloc->buffer)) { |
| 83 | GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n" , |
| 84 | __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset); |
| 85 | GGML_ABORT("not enough space in the buffer" ); |
| 86 | } |
| 87 | |
| 88 | void * addr = (char *)ggml_backend_buffer_get_base(buffer: talloc->buffer) + talloc->offset; |
| 89 | talloc->offset += size; |
| 90 | |
| 91 | assert(((uintptr_t)addr % talloc->alignment) == 0); |
| 92 | |
| 93 | return ggml_backend_tensor_alloc(buffer: talloc->buffer, tensor, addr); |
| 94 | } |
| 95 | |
| 96 | // dynamic tensor allocator |
| 97 | |
| 98 | #define GGML_VBUFFER_MAX_CHUNKS 16 |
| 99 | |
| 100 | // relative memory address within an allocation that can be split into multiple buffers (chunks) |
| 101 | struct buffer_address { |
| 102 | int chunk; // index of a backend buffer |
| 103 | size_t offset; // local memory offset within the buffer |
| 104 | }; |
| 105 | |
| 106 | static const struct buffer_address GGML_BUFFER_ADDRESS_INVALID = { -1, SIZE_MAX }; |
| 107 | |
| 108 | static bool ggml_buffer_address_less(struct buffer_address a, struct buffer_address b) { |
| 109 | return a.chunk != b.chunk ? a.chunk < b.chunk : a.offset < b.offset; |
| 110 | } |
| 111 | |
| 112 | struct free_block { |
| 113 | size_t offset; |
| 114 | size_t size; |
| 115 | }; |
| 116 | |
| 117 | struct tallocr_chunk { |
| 118 | struct free_block free_blocks[MAX_FREE_BLOCKS]; |
| 119 | int n_free_blocks; |
| 120 | size_t max_size; |
| 121 | }; |
| 122 | |
| 123 | struct ggml_dyn_tallocr { |
| 124 | size_t alignment; |
| 125 | size_t max_chunk_size; |
| 126 | struct tallocr_chunk * chunks[GGML_VBUFFER_MAX_CHUNKS]; |
| 127 | int n_chunks; |
| 128 | |
| 129 | #ifdef GGML_ALLOCATOR_DEBUG |
| 130 | struct { |
| 131 | const struct ggml_tensor * tensor; |
| 132 | struct buffer_address addr; |
| 133 | } allocated_tensors[1024]; |
| 134 | #endif |
| 135 | }; |
| 136 | |
| 137 | static void ggml_dyn_tallocr_insert_block(struct tallocr_chunk * chunk, size_t offset, size_t size) { |
| 138 | GGML_ASSERT(chunk->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks" ); |
| 139 | // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster) |
| 140 | int insert_pos = 0; |
| 141 | while (insert_pos < chunk->n_free_blocks && chunk->free_blocks[insert_pos].offset < offset) { |
| 142 | insert_pos++; |
| 143 | } |
| 144 | // shift all blocks from insert_pos onward to make room for the new block |
| 145 | for (int i = chunk->n_free_blocks; i > insert_pos; i--) { |
| 146 | chunk->free_blocks[i] = chunk->free_blocks[i-1]; |
| 147 | } |
| 148 | // insert the new block |
| 149 | chunk->free_blocks[insert_pos].offset = offset; |
| 150 | chunk->free_blocks[insert_pos].size = size; |
| 151 | chunk->n_free_blocks++; |
| 152 | } |
| 153 | |
| 154 | static void ggml_dyn_tallocr_remove_block(struct tallocr_chunk * chunk, int idx) { |
| 155 | // shift all elements after idx by 1 to the left, overwriting the element at idx |
| 156 | for (int i = idx; i < chunk->n_free_blocks; i++) { |
| 157 | chunk->free_blocks[i] = chunk->free_blocks[i+1]; |
| 158 | } |
| 159 | chunk->n_free_blocks--; |
| 160 | } |
| 161 | |
| 162 | static int ggml_dyn_tallocr_new_chunk(struct ggml_dyn_tallocr * alloc, size_t min_size) { |
| 163 | if (alloc->n_chunks >= GGML_VBUFFER_MAX_CHUNKS) { |
| 164 | return -1; |
| 165 | } |
| 166 | struct tallocr_chunk * chunk = calloc(nmemb: 1, size: sizeof(struct tallocr_chunk)); |
| 167 | chunk->n_free_blocks = 1; |
| 168 | chunk->free_blocks[0].offset = 0; |
| 169 | // available space in a chunk is limited to max_chunk_size, but can be higher if: |
| 170 | // 1. a single tensor exceeds the maximum, and cannot fit any other way |
| 171 | // 2. we are running out of chunks |
| 172 | // backends will either manage to allocate the larger size, or report an error. |
| 173 | chunk->free_blocks[0].size = MAX(min_size, alloc->max_chunk_size); |
| 174 | if (alloc->n_chunks == GGML_VBUFFER_MAX_CHUNKS - 1) { |
| 175 | chunk->free_blocks[0].size = SIZE_MAX/2; |
| 176 | } |
| 177 | alloc->chunks[alloc->n_chunks] = chunk; |
| 178 | alloc->n_chunks++; |
| 179 | return alloc->n_chunks - 1; |
| 180 | } |
| 181 | |
| 182 | #ifdef GGML_ALLOCATOR_DEBUG |
| 183 | static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, const struct ggml_tensor * tensor) { |
| 184 | for (int i = 0; i < 1024; i++) { |
| 185 | if (alloc->allocated_tensors[i].tensor == NULL) { |
| 186 | alloc->allocated_tensors[i].tensor = tensor; |
| 187 | alloc->allocated_tensors[i].addr = addr; |
| 188 | return; |
| 189 | } |
| 190 | } |
| 191 | GGML_ABORT("out of allocated_tensors" ); |
| 192 | } |
| 193 | static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, const struct ggml_tensor * tensor) { |
| 194 | for (int i = 0; i < 1024; i++) { |
| 195 | if (alloc->allocated_tensors[i].addr.chunk == addr.chunk && alloc->allocated_tensors[i].addr.offset == addr.offset) { |
| 196 | alloc->allocated_tensors[i].tensor = NULL; |
| 197 | return; |
| 198 | } |
| 199 | } |
| 200 | GGML_ABORT("tried to free tensor %s not found\n" , tensor->name); |
| 201 | } |
| 202 | #endif |
| 203 | |
| 204 | static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) { |
| 205 | size = aligned_offset(NULL, offset: size, alignment: alloc->alignment); |
| 206 | |
| 207 | AT_PRINTF("%s: allocating %s (%zu bytes) - " , __func__, tensor->name, size); |
| 208 | |
| 209 | int best_fit_chunk = -1; |
| 210 | int best_fit_block = -1; |
| 211 | size_t max_avail = 0; |
| 212 | |
| 213 | // find the best fitting free block besides the last block, within any chunk |
| 214 | for (int c = 0; c < alloc->n_chunks; ++c) { |
| 215 | struct tallocr_chunk * chunk = alloc->chunks[c]; |
| 216 | size_t best_fit_size = SIZE_MAX; |
| 217 | for (int i = 0; i < chunk->n_free_blocks - 1; i++) { |
| 218 | struct free_block * block = &chunk->free_blocks[i]; |
| 219 | max_avail = MAX(max_avail, block->size); |
| 220 | if (block->size >= size && block->size <= best_fit_size) { |
| 221 | best_fit_chunk = c; |
| 222 | best_fit_block = i; |
| 223 | best_fit_size = block->size; |
| 224 | } |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | if (best_fit_block == -1) { |
| 229 | // no suitable block found, try the last block (this may grow a chunks size) |
| 230 | int64_t best_reuse = INT64_MIN; |
| 231 | for (int c = 0; c < alloc->n_chunks; ++c) { |
| 232 | struct tallocr_chunk * chunk = alloc->chunks[c]; |
| 233 | if (chunk->n_free_blocks > 0) { |
| 234 | struct free_block * block = &chunk->free_blocks[chunk->n_free_blocks - 1]; |
| 235 | max_avail = MAX(max_avail, block->size); |
| 236 | int64_t reuse_factor = chunk->max_size - block->offset - size; |
| 237 | // reuse_factor < 0 : amount of extra memory that needs to be allocated |
| 238 | // reuse_factor = 0 : allocated free space exactly matches tensor size |
| 239 | // reuse_factor > 0 : superfluous memory that will remain unused |
| 240 | bool better_reuse = best_reuse < 0 && reuse_factor > best_reuse; |
| 241 | bool better_fit = reuse_factor >= 0 && reuse_factor < best_reuse; |
| 242 | if (block->size >= size && (better_reuse || better_fit)) { |
| 243 | best_fit_chunk = c; |
| 244 | best_fit_block = chunk->n_free_blocks - 1; |
| 245 | best_reuse = reuse_factor; |
| 246 | } |
| 247 | } |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | if (best_fit_block == -1) { |
| 252 | // none of the existing chunks have enough space left |
| 253 | best_fit_chunk = ggml_dyn_tallocr_new_chunk(alloc, min_size: size); |
| 254 | best_fit_block = 0; |
| 255 | } |
| 256 | if (best_fit_chunk == -1) { |
| 257 | // since the last chunk always has virtually endless memory, this should never happen |
| 258 | GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n" , |
| 259 | __func__, size, max_avail); |
| 260 | GGML_ABORT("graph allocation: failed to reserve memory" ); |
| 261 | } |
| 262 | |
| 263 | struct tallocr_chunk * chunk = alloc->chunks[best_fit_chunk]; |
| 264 | struct free_block * block = &chunk->free_blocks[best_fit_block]; |
| 265 | struct buffer_address addr = {.chunk = best_fit_chunk, .offset = block->offset }; |
| 266 | block->offset += size; |
| 267 | block->size -= size; |
| 268 | if (block->size == 0) { |
| 269 | // remove block if empty |
| 270 | ggml_dyn_tallocr_remove_block(chunk, idx: best_fit_block); |
| 271 | } |
| 272 | |
| 273 | AT_PRINTF("block %d, offset %zu, chunk %d\n" , best_fit_block, addr.offset, addr.chunk); |
| 274 | |
| 275 | #ifdef GGML_ALLOCATOR_DEBUG |
| 276 | add_allocated_tensor(alloc, addr, tensor); |
| 277 | size_t cur_max = addr.offset + size; |
| 278 | if (cur_max > chunk->max_size) { |
| 279 | // sort allocated_tensors by chunk/offset |
| 280 | for (int i = 0; i < 1024; i++) { |
| 281 | for (int j = i + 1; j < 1024; j++) { |
| 282 | if (ggml_buffer_address_less(alloc->allocated_tensors[j].addr, alloc->allocated_tensors[i].addr)) { |
| 283 | const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor; |
| 284 | struct buffer_address tmp_addr = alloc->allocated_tensors[i].addr; |
| 285 | alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor; |
| 286 | alloc->allocated_tensors[i].addr = alloc->allocated_tensors[j].addr; |
| 287 | alloc->allocated_tensors[j].tensor = tmp_tensor; |
| 288 | alloc->allocated_tensors[j].addr = tmp_addr; |
| 289 | } |
| 290 | } |
| 291 | } |
| 292 | GGML_LOG_DEBUG("max_size[%d] = %.2f MB: tensors: " , addr.chunk, cur_max / 1024.0 / 1024.0); |
| 293 | for (int i = 0; i < 1024; i++) { |
| 294 | if (alloc->allocated_tensors[i].tensor) { |
| 295 | GGML_LOG_DEBUG("%s [%d: %zx-%zx] (%.2f MB) " , alloc->allocated_tensors[i].tensor->name, |
| 296 | alloc->allocated_tensors[i].addr.chunk, |
| 297 | alloc->allocated_tensors[i].addr.offset, |
| 298 | alloc->allocated_tensors[i].addr.offset + ggml_nbytes(alloc->allocated_tensors[i].tensor), |
| 299 | ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0); |
| 300 | } |
| 301 | } |
| 302 | GGML_LOG_DEBUG("\n" ); |
| 303 | } |
| 304 | #endif |
| 305 | |
| 306 | chunk->max_size = MAX(chunk->max_size, addr.offset + size); |
| 307 | |
| 308 | return addr; |
| 309 | |
| 310 | GGML_UNUSED(tensor); |
| 311 | } |
| 312 | |
| 313 | // this is a very naive implementation, but for our case the number of free blocks should be very small |
| 314 | static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, size_t size, const struct ggml_tensor * tensor) { |
| 315 | size = aligned_offset(NULL, offset: size, alignment: alloc->alignment); |
| 316 | |
| 317 | AT_PRINTF("%s: freeing %s at {chunk=%d, offset=%zu} (%zu bytes) - n_free_blocks = %d\n" , |
| 318 | __func__, tensor->name, addr.chunk, addr.offset, size, alloc->chunks[addr.chunk]->n_free_blocks); |
| 319 | |
| 320 | #ifdef GGML_ALLOCATOR_DEBUG |
| 321 | remove_allocated_tensor(alloc, addr, tensor); |
| 322 | #endif |
| 323 | |
| 324 | struct tallocr_chunk * chunk = alloc->chunks[addr.chunk]; |
| 325 | |
| 326 | // see if we can merge with an existing block |
| 327 | for (int i = 0; i < chunk->n_free_blocks; i++) { |
| 328 | struct free_block * block = &chunk->free_blocks[i]; |
| 329 | // check if ptr is at the end of the block |
| 330 | if (block->offset + block->size == addr.offset) { |
| 331 | block->size += size; |
| 332 | // check if we can merge with the next block |
| 333 | if (i < chunk->n_free_blocks - 1) { |
| 334 | struct free_block * next = &chunk->free_blocks[i+1]; |
| 335 | if (block->offset + block->size == next->offset) { |
| 336 | block->size += next->size; |
| 337 | ggml_dyn_tallocr_remove_block(chunk, idx: i+1); |
| 338 | } |
| 339 | } |
| 340 | return; |
| 341 | } |
| 342 | // check if ptr is at the beginning of the block |
| 343 | if (addr.offset + size == block->offset) { |
| 344 | block->offset = addr.offset; |
| 345 | block->size += size; |
| 346 | // check if we can merge with the previous block |
| 347 | if (i > 0) { |
| 348 | struct free_block * prev = &chunk->free_blocks[i-1]; |
| 349 | if (prev->offset + prev->size == block->offset) { |
| 350 | prev->size += block->size; |
| 351 | ggml_dyn_tallocr_remove_block(chunk, idx: i); |
| 352 | } |
| 353 | } |
| 354 | return; |
| 355 | } |
| 356 | } |
| 357 | // otherwise, add a new block |
| 358 | ggml_dyn_tallocr_insert_block(chunk, offset: addr.offset, size); |
| 359 | |
| 360 | GGML_UNUSED(tensor); |
| 361 | } |
| 362 | |
| 363 | static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) { |
| 364 | for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS; i++) { |
| 365 | free(ptr: alloc->chunks[i]); |
| 366 | alloc->chunks[i] = NULL; |
| 367 | } |
| 368 | alloc->n_chunks = 0; |
| 369 | |
| 370 | #ifdef GGML_ALLOCATOR_DEBUG |
| 371 | for (int i = 0; i < 1024; i++) { |
| 372 | alloc->allocated_tensors[i].tensor = NULL; |
| 373 | } |
| 374 | #endif |
| 375 | } |
| 376 | |
| 377 | static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment, size_t max_buffer_size) { |
| 378 | struct ggml_dyn_tallocr * alloc = (struct ggml_dyn_tallocr *)malloc(size: sizeof(struct ggml_dyn_tallocr)); |
| 379 | |
| 380 | *alloc = (struct ggml_dyn_tallocr) { |
| 381 | /*.alignment = */ alignment, |
| 382 | /*.max_chunk_size = */ MIN(max_buffer_size, SIZE_MAX/2), // clamp to avoid overflows |
| 383 | /*.chunks = */ {NULL}, |
| 384 | /*.n_chunks = */ 0, |
| 385 | #ifdef GGML_ALLOCATOR_DEBUG |
| 386 | /*.allocated_tensors = */ {{0}}, |
| 387 | #endif |
| 388 | }; |
| 389 | |
| 390 | ggml_dyn_tallocr_reset(alloc); |
| 391 | |
| 392 | return alloc; |
| 393 | } |
| 394 | |
| 395 | static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) { |
| 396 | for (int i = 0; i < alloc->n_chunks; ++i) { |
| 397 | free(ptr: alloc->chunks[i]); |
| 398 | } |
| 399 | free(ptr: alloc); |
| 400 | } |
| 401 | |
| 402 | static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc, int chunk) { |
| 403 | return chunk < alloc->n_chunks ? alloc->chunks[chunk]->max_size : 0; |
| 404 | } |
| 405 | |
| 406 | |
| 407 | // virtual buffer with contiguous memory range, split into multiple backend buffers (chunks) |
| 408 | |
| 409 | struct vbuffer { |
| 410 | ggml_backend_buffer_t chunks[GGML_VBUFFER_MAX_CHUNKS]; |
| 411 | }; |
| 412 | |
| 413 | static void ggml_vbuffer_free(struct vbuffer * buf) { |
| 414 | if (buf == NULL) { |
| 415 | return; |
| 416 | } |
| 417 | for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS; ++i) { |
| 418 | ggml_backend_buffer_free(buffer: buf->chunks[i]); |
| 419 | } |
| 420 | free(ptr: buf); |
| 421 | } |
| 422 | |
| 423 | static size_t ggml_vbuffer_chunk_size(struct vbuffer * buf, int chunk) { |
| 424 | return buf->chunks[chunk] ? ggml_backend_buffer_get_size(buffer: buf->chunks[chunk]) : 0; |
| 425 | } |
| 426 | |
| 427 | static size_t ggml_vbuffer_size(struct vbuffer * buf) { |
| 428 | size_t size = 0; |
| 429 | for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[i]; ++i) { |
| 430 | size += ggml_backend_buffer_get_size(buffer: buf->chunks[i]); |
| 431 | } |
| 432 | return size; |
| 433 | } |
| 434 | |
| 435 | static struct vbuffer * ggml_vbuffer_alloc(ggml_backend_buffer_type_t buft, const struct ggml_dyn_tallocr * talloc, enum ggml_backend_buffer_usage usage) { |
| 436 | struct vbuffer * buf = (struct vbuffer *)calloc(nmemb: 1, size: sizeof(struct vbuffer)); |
| 437 | if (buf == NULL) { |
| 438 | return NULL; |
| 439 | } |
| 440 | |
| 441 | for (int n = 0; n < talloc->n_chunks; n++) { |
| 442 | size_t chunk_size = talloc->chunks[n]->max_size; |
| 443 | buf->chunks[n] = ggml_backend_buft_alloc_buffer(buft, size: chunk_size); |
| 444 | if (buf->chunks[n] == NULL) { |
| 445 | ggml_vbuffer_free(buf); |
| 446 | return NULL; |
| 447 | } |
| 448 | ggml_backend_buffer_set_usage(buffer: buf->chunks[n], usage); |
| 449 | } |
| 450 | return buf; |
| 451 | } |
| 452 | |
| 453 | static void ggml_vbuffer_tensor_alloc(struct vbuffer * buf, struct ggml_tensor * tensor, struct buffer_address buf_addr) { |
| 454 | void * base = ggml_backend_buffer_get_base(buffer: buf->chunks[buf_addr.chunk]); |
| 455 | void * addr = (char *)base + buf_addr.offset; |
| 456 | ggml_backend_tensor_alloc(buffer: buf->chunks[buf_addr.chunk], tensor, addr); |
| 457 | } |
| 458 | |
| 459 | static void ggml_vbuffer_reset(struct vbuffer * buf) { |
| 460 | for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[i]; ++i) { |
| 461 | ggml_backend_buffer_reset(buffer: buf->chunks[i]); |
| 462 | } |
| 463 | } |
| 464 | |
| 465 | |
| 466 | ///////////////////////////////////// |
| 467 | |
| 468 | // graph allocator |
| 469 | |
| 470 | struct hash_node { |
| 471 | int n_children; |
| 472 | int n_views; |
| 473 | int buffer_id; |
| 474 | struct buffer_address addr; |
| 475 | bool allocated; |
| 476 | }; |
| 477 | |
| 478 | struct tensor_alloc { |
| 479 | int buffer_id; |
| 480 | struct buffer_address addr; |
| 481 | size_t size_max; // 0 = pre-allocated, unused, or view |
| 482 | }; |
| 483 | |
| 484 | struct leaf_alloc { |
| 485 | struct tensor_alloc leaf; |
| 486 | }; |
| 487 | |
| 488 | struct node_alloc { |
| 489 | struct tensor_alloc dst; |
| 490 | struct tensor_alloc src[GGML_MAX_SRC]; |
| 491 | }; |
| 492 | |
| 493 | struct ggml_gallocr { |
| 494 | ggml_backend_buffer_type_t * bufts; // [n_buffers] |
| 495 | struct vbuffer ** buffers; // [n_buffers] |
| 496 | struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers] |
| 497 | int n_buffers; |
| 498 | |
| 499 | struct ggml_hash_set hash_set; |
| 500 | struct hash_node * hash_values; // [hash_set.size] |
| 501 | |
| 502 | struct node_alloc * node_allocs; // [n_nodes] |
| 503 | int n_nodes; |
| 504 | |
| 505 | struct leaf_alloc * leaf_allocs; // [n_leafs] |
| 506 | int n_leafs; |
| 507 | }; |
| 508 | |
| 509 | ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) { |
| 510 | ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(nmemb: 1, size: sizeof(struct ggml_gallocr)); |
| 511 | GGML_ASSERT(galloc != NULL); |
| 512 | |
| 513 | galloc->bufts = calloc(nmemb: n_bufs, size: sizeof(ggml_backend_buffer_type_t)); |
| 514 | GGML_ASSERT(galloc->bufts != NULL); |
| 515 | |
| 516 | galloc->buffers = calloc(nmemb: n_bufs, size: sizeof(struct vbuffer *)); |
| 517 | GGML_ASSERT(galloc->buffers != NULL); |
| 518 | |
| 519 | galloc->buf_tallocs = calloc(nmemb: n_bufs, size: sizeof(struct ggml_dyn_tallocr *)); |
| 520 | GGML_ASSERT(galloc->buf_tallocs != NULL); |
| 521 | |
| 522 | for (int i = 0; i < n_bufs; i++) { |
| 523 | galloc->bufts[i] = bufts[i]; |
| 524 | galloc->buffers[i] = NULL; |
| 525 | |
| 526 | // check if the same buffer type is used multiple times and reuse the same allocator |
| 527 | for (int j = 0; j < i; j++) { |
| 528 | if (bufts[i] == bufts[j]) { |
| 529 | galloc->buf_tallocs[i] = galloc->buf_tallocs[j]; |
| 530 | break; |
| 531 | } |
| 532 | } |
| 533 | |
| 534 | if (galloc->buf_tallocs[i] == NULL) { |
| 535 | size_t alignment = ggml_backend_buft_get_alignment(buft: bufts[i]); |
| 536 | size_t max_size = ggml_backend_buft_get_max_size(buft: bufts[i]); |
| 537 | galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment, max_buffer_size: max_size); |
| 538 | } |
| 539 | } |
| 540 | galloc->n_buffers = n_bufs; |
| 541 | |
| 542 | return galloc; |
| 543 | } |
| 544 | |
| 545 | ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft) { |
| 546 | return ggml_gallocr_new_n(bufts: &buft, n_bufs: 1); |
| 547 | } |
| 548 | |
| 549 | void ggml_gallocr_free(ggml_gallocr_t galloc) { |
| 550 | if (galloc == NULL) { |
| 551 | return; |
| 552 | } |
| 553 | |
| 554 | for (int i = 0; i < galloc->n_buffers; i++) { |
| 555 | if (galloc->buffers != NULL) { |
| 556 | // skip if already freed |
| 557 | bool freed = false; |
| 558 | for (int j = 0; j < i; j++) { |
| 559 | if (galloc->buffers[j] == galloc->buffers[i]) { |
| 560 | freed = true; |
| 561 | break; |
| 562 | } |
| 563 | } |
| 564 | if (!freed) { |
| 565 | ggml_vbuffer_free(buf: galloc->buffers[i]); |
| 566 | } |
| 567 | } |
| 568 | if (galloc->buf_tallocs != NULL) { |
| 569 | // skip if already freed |
| 570 | bool freed = false; |
| 571 | for (int j = 0; j < i; j++) { |
| 572 | if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) { |
| 573 | freed = true; |
| 574 | break; |
| 575 | } |
| 576 | } |
| 577 | if (!freed) { |
| 578 | ggml_dyn_tallocr_free(alloc: galloc->buf_tallocs[i]); |
| 579 | } |
| 580 | } |
| 581 | } |
| 582 | |
| 583 | ggml_hash_set_free(hash_set: &galloc->hash_set); |
| 584 | free(ptr: galloc->hash_values); |
| 585 | free(ptr: galloc->bufts); |
| 586 | free(ptr: galloc->buffers); |
| 587 | free(ptr: galloc->buf_tallocs); |
| 588 | free(ptr: galloc->node_allocs); |
| 589 | free(ptr: galloc->leaf_allocs); |
| 590 | free(ptr: galloc); |
| 591 | } |
| 592 | |
| 593 | typedef struct ggml_gallocr * ggml_gallocr_t; |
| 594 | |
| 595 | static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) { |
| 596 | size_t i = ggml_hash_find_or_insert(hash_set: &galloc->hash_set, key: t); |
| 597 | return &galloc->hash_values[i]; |
| 598 | } |
| 599 | |
| 600 | static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) { |
| 601 | return ggml_gallocr_hash_get(galloc, t)->allocated; |
| 602 | } |
| 603 | |
| 604 | static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) { |
| 605 | return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated; |
| 606 | } |
| 607 | |
| 608 | // free the extra space at the end if the new tensor is smaller |
| 609 | static void (ggml_gallocr_t galloc, struct ggml_tensor * node, struct ggml_tensor * parent) { |
| 610 | struct hash_node * hn = ggml_gallocr_hash_get(galloc, t: node); |
| 611 | struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, t: parent); |
| 612 | |
| 613 | size_t parent_size = ggml_backend_buft_get_alloc_size(buft: galloc->bufts[p_hn->buffer_id], tensor: parent); |
| 614 | size_t node_size = ggml_backend_buft_get_alloc_size(buft: galloc->bufts[hn->buffer_id], tensor: node); |
| 615 | |
| 616 | GGML_ASSERT(parent_size >= node_size); |
| 617 | |
| 618 | if (parent_size > node_size) { |
| 619 | struct ggml_dyn_tallocr * p_alloc = galloc->buf_tallocs[p_hn->buffer_id]; |
| 620 | struct buffer_address p_addr = p_hn->addr; |
| 621 | p_addr.offset += node_size; |
| 622 | size_t = parent_size - node_size; |
| 623 | AT_PRINTF("freeing extra %zu bytes from parent %s for %s\n" , extra_size, parent->name, node->name); |
| 624 | ggml_dyn_tallocr_free_tensor(alloc: p_alloc, addr: p_addr, size: extra_size, tensor: parent); |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) { |
| 629 | GGML_ASSERT(buffer_id >= 0); |
| 630 | struct hash_node * hn = ggml_gallocr_hash_get(galloc, t: node); |
| 631 | |
| 632 | if (!ggml_gallocr_is_allocated(galloc, t: node) && !ggml_is_view(t: node)) { |
| 633 | hn->allocated = true; |
| 634 | assert(hn->addr.offset == 0); |
| 635 | |
| 636 | // try to reuse a parent's buffer (inplace) |
| 637 | if (ggml_op_can_inplace(op: node->op)) { |
| 638 | for (int i = 0; i < GGML_MAX_SRC; i++) { |
| 639 | struct ggml_tensor * parent = node->src[i]; |
| 640 | if (parent == NULL) { |
| 641 | continue; |
| 642 | } |
| 643 | |
| 644 | // if the node's data is external, then we cannot re-use it |
| 645 | if (!ggml_gallocr_is_own(galloc, t: parent)) { |
| 646 | AT_PRINTF("not reusing parent %s for %s as %p is external\n" , parent->name, node->name, parent->data); |
| 647 | continue; |
| 648 | } |
| 649 | |
| 650 | // outputs cannot be reused |
| 651 | if (parent->flags & GGML_TENSOR_FLAG_OUTPUT || (parent->view_src != NULL && parent->view_src->flags & GGML_TENSOR_FLAG_OUTPUT)) { |
| 652 | AT_PRINTF("not reusing parent %s for %s as it is an output\n" , parent->name, node->name); |
| 653 | continue; |
| 654 | } |
| 655 | |
| 656 | if (!ggml_are_same_layout(a: node, b: parent)) { |
| 657 | AT_PRINTF("not reusing parent %s for %s as layouts are different\n" , parent->name, node->name); |
| 658 | continue; |
| 659 | } |
| 660 | |
| 661 | struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, t: parent); |
| 662 | if (p_hn->n_children == 1 && p_hn->n_views == 0) { |
| 663 | if (ggml_is_view(t: parent)) { |
| 664 | struct ggml_tensor * view_src = parent->view_src; |
| 665 | struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, t: view_src); |
| 666 | if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) { |
| 667 | AT_PRINTF("reusing view parent %s (%s) for %s\n" , parent->name, view_src->name, node->name); |
| 668 | assert(view_src_hn->addr.chunk == p_hn->addr.chunk && view_src_hn->addr.offset == p_hn->addr.offset); |
| 669 | hn->buffer_id = p_hn->buffer_id; |
| 670 | hn->addr = p_hn->addr; |
| 671 | p_hn->allocated = false; // avoid freeing the parent |
| 672 | view_src_hn->allocated = false; |
| 673 | ggml_gallocr_free_extra_space(galloc, node, parent: view_src); |
| 674 | return; |
| 675 | } |
| 676 | } else { |
| 677 | AT_PRINTF("reusing parent %s for %s\n" , parent->name, node->name); |
| 678 | hn->buffer_id = p_hn->buffer_id; |
| 679 | hn->addr = p_hn->addr; |
| 680 | p_hn->allocated = false; // avoid freeing the parent |
| 681 | ggml_gallocr_free_extra_space(galloc, node, parent); |
| 682 | return; |
| 683 | } |
| 684 | } |
| 685 | } |
| 686 | } |
| 687 | // allocate tensor from the buffer |
| 688 | struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id]; |
| 689 | ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id]; |
| 690 | size_t size = ggml_backend_buft_get_alloc_size(buft, tensor: node); |
| 691 | hn->buffer_id = buffer_id; |
| 692 | hn->addr = ggml_dyn_tallocr_alloc(alloc, size, tensor: node); |
| 693 | } |
| 694 | } |
| 695 | |
| 696 | static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) { |
| 697 | // graph outputs are never freed |
| 698 | if (node->flags & GGML_TENSOR_FLAG_OUTPUT) { |
| 699 | AT_PRINTF("not freeing output %s\n" , node->name); |
| 700 | return; |
| 701 | } |
| 702 | |
| 703 | struct hash_node * hn = ggml_gallocr_hash_get(galloc, t: node); |
| 704 | int buffer_id = hn->buffer_id; |
| 705 | struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id]; |
| 706 | ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id]; |
| 707 | size_t size = ggml_backend_buft_get_alloc_size(buft, tensor: node); |
| 708 | ggml_dyn_tallocr_free_tensor(alloc, addr: hn->addr, size, tensor: node); |
| 709 | hn->allocated = false; |
| 710 | } |
| 711 | |
| 712 | static int get_node_buffer_id(const int * node_buffer_ids, int i) { |
| 713 | return node_buffer_ids ? node_buffer_ids[i] : 0; |
| 714 | } |
| 715 | |
| 716 | static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) { |
| 717 | // clear hash tables |
| 718 | ggml_hash_set_reset(hash_set: &galloc->hash_set); |
| 719 | memset(s: galloc->hash_values, c: 0, n: sizeof(struct hash_node) * galloc->hash_set.size); |
| 720 | |
| 721 | // allocate leafs |
| 722 | // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes |
| 723 | for (int i = 0; i < graph->n_leafs; i++) { |
| 724 | struct ggml_tensor * leaf = graph->leafs[i]; |
| 725 | ggml_gallocr_allocate_node(galloc, node: leaf, buffer_id: get_node_buffer_id(node_buffer_ids: leaf_buffer_ids, i)); |
| 726 | } |
| 727 | |
| 728 | // count number of children and views |
| 729 | // allocate other graph inputs and leafs first to avoid overwriting them |
| 730 | for (int i = 0; i < graph->n_nodes; i++) { |
| 731 | struct ggml_tensor * node = graph->nodes[i]; |
| 732 | |
| 733 | // TODO: better way to add external dependencies |
| 734 | // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to |
| 735 | // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node |
| 736 | // itself is never used and should not be considered a dependency |
| 737 | if (ggml_is_view(t: node) && node->op != GGML_OP_NONE) { |
| 738 | struct ggml_tensor * view_src = node->view_src; |
| 739 | ggml_gallocr_hash_get(galloc, t: view_src)->n_views += 1; |
| 740 | } |
| 741 | |
| 742 | if (node->flags & GGML_TENSOR_FLAG_INPUT) { |
| 743 | ggml_gallocr_allocate_node(galloc, node: graph->nodes[i], buffer_id: get_node_buffer_id(node_buffer_ids, i)); |
| 744 | } |
| 745 | |
| 746 | for (int j = 0; j < GGML_MAX_SRC; j++) { |
| 747 | struct ggml_tensor * src = node->src[j]; |
| 748 | if (src == NULL) { |
| 749 | continue; |
| 750 | } |
| 751 | |
| 752 | ggml_gallocr_hash_get(galloc, t: src)->n_children += 1; |
| 753 | |
| 754 | // allocate explicit inputs |
| 755 | if (src->flags & GGML_TENSOR_FLAG_INPUT) { |
| 756 | ggml_gallocr_allocate_node(galloc, node: src, buffer_id: get_node_buffer_id(node_buffer_ids, i)); |
| 757 | } |
| 758 | } |
| 759 | } |
| 760 | |
| 761 | // allocate tensors |
| 762 | for (int i = 0; i < graph->n_nodes; i++) { |
| 763 | struct ggml_tensor * node = graph->nodes[i]; |
| 764 | int buffer_id = get_node_buffer_id(node_buffer_ids, i); |
| 765 | |
| 766 | // allocate parents (only leafs need to be allocated at this point) |
| 767 | for (int j = 0; j < GGML_MAX_SRC; j++) { |
| 768 | struct ggml_tensor * parent = node->src[j]; |
| 769 | if (parent == NULL) { |
| 770 | continue; |
| 771 | } |
| 772 | ggml_gallocr_allocate_node(galloc, node: parent, buffer_id); |
| 773 | } |
| 774 | |
| 775 | // allocate node |
| 776 | ggml_gallocr_allocate_node(galloc, node, buffer_id); |
| 777 | |
| 778 | AT_PRINTF("exec: %s (%s) <= " , ggml_op_desc(node), node->name); |
| 779 | for (int j = 0; j < GGML_MAX_SRC; j++) { |
| 780 | struct ggml_tensor * parent = node->src[j]; |
| 781 | if (parent == NULL) { |
| 782 | continue; |
| 783 | } |
| 784 | AT_PRINTF("%s" , parent->name); |
| 785 | if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) { |
| 786 | AT_PRINTF(", " ); |
| 787 | } |
| 788 | } |
| 789 | AT_PRINTF("\n" ); |
| 790 | |
| 791 | // update parents |
| 792 | for (int j = 0; j < GGML_MAX_SRC; j++) { |
| 793 | struct ggml_tensor * parent = node->src[j]; |
| 794 | if (parent == NULL) { |
| 795 | continue; |
| 796 | } |
| 797 | struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, t: parent); |
| 798 | p_hn->n_children -= 1; |
| 799 | |
| 800 | AT_PRINTF("parent %s: %d children, %d views, allocated: %d\n" , |
| 801 | parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated); |
| 802 | |
| 803 | if (p_hn->n_children == 0 && p_hn->n_views == 0) { |
| 804 | if (ggml_is_view(t: parent)) { |
| 805 | struct ggml_tensor * view_src = parent->view_src; |
| 806 | struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, t: view_src); |
| 807 | view_src_hn->n_views -= 1; |
| 808 | AT_PRINTF("view_src %s: %d children, %d views\n" , |
| 809 | view_src->name, view_src_hn->n_children, view_src_hn->n_views); |
| 810 | if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src_hn->allocated) { |
| 811 | ggml_gallocr_free_node(galloc, node: view_src); |
| 812 | } |
| 813 | } |
| 814 | else if (p_hn->allocated) { |
| 815 | ggml_gallocr_free_node(galloc, node: parent); |
| 816 | } |
| 817 | } |
| 818 | AT_PRINTF("\n" ); |
| 819 | } |
| 820 | } |
| 821 | } |
| 822 | |
| 823 | bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) { |
| 824 | size_t min_hash_size = graph->n_nodes + graph->n_leafs; |
| 825 | // add 25% margin to avoid hash collisions |
| 826 | min_hash_size += min_hash_size / 4; |
| 827 | |
| 828 | // initialize hash table |
| 829 | if (galloc->hash_set.size < min_hash_size) { |
| 830 | ggml_hash_set_free(hash_set: &galloc->hash_set); |
| 831 | galloc->hash_set = ggml_hash_set_new(size: min_hash_size); |
| 832 | GGML_ASSERT(galloc->hash_set.keys != NULL); |
| 833 | |
| 834 | free(ptr: galloc->hash_values); |
| 835 | galloc->hash_values = malloc(size: sizeof(struct hash_node) * galloc->hash_set.size); |
| 836 | GGML_ASSERT(galloc->hash_values != NULL); |
| 837 | } |
| 838 | |
| 839 | // reset allocators |
| 840 | for (int i = 0; i < galloc->n_buffers; i++) { |
| 841 | ggml_dyn_tallocr_reset(alloc: galloc->buf_tallocs[i]); |
| 842 | } |
| 843 | |
| 844 | // allocate in hash table |
| 845 | ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids); |
| 846 | |
| 847 | // set the node_allocs from the hash table |
| 848 | if (galloc->n_nodes < graph->n_nodes) { |
| 849 | free(ptr: galloc->node_allocs); |
| 850 | galloc->node_allocs = calloc(nmemb: graph->n_nodes, size: sizeof(struct node_alloc)); |
| 851 | GGML_ASSERT(galloc->node_allocs != NULL); |
| 852 | } |
| 853 | galloc->n_nodes = graph->n_nodes; |
| 854 | for (int i = 0; i < graph->n_nodes; i++) { |
| 855 | struct ggml_tensor * node = graph->nodes[i]; |
| 856 | struct node_alloc * node_alloc = &galloc->node_allocs[i]; |
| 857 | if (node->view_src || node->data) { |
| 858 | node_alloc->dst.buffer_id = -1; |
| 859 | node_alloc->dst.addr = GGML_BUFFER_ADDRESS_INVALID; |
| 860 | node_alloc->dst.size_max = 0; |
| 861 | } else { |
| 862 | struct hash_node * hn = ggml_gallocr_hash_get(galloc, t: node); |
| 863 | node_alloc->dst.buffer_id = hn->buffer_id; |
| 864 | node_alloc->dst.addr = hn->addr; |
| 865 | node_alloc->dst.size_max = ggml_backend_buft_get_alloc_size(buft: galloc->bufts[hn->buffer_id], tensor: node); |
| 866 | } |
| 867 | for (int j = 0; j < GGML_MAX_SRC; j++) { |
| 868 | struct ggml_tensor * src = node->src[j]; |
| 869 | if (!src || src->view_src || src->data) { |
| 870 | node_alloc->src[j].buffer_id = -1; |
| 871 | node_alloc->src[j].addr = GGML_BUFFER_ADDRESS_INVALID; |
| 872 | node_alloc->src[j].size_max = 0; |
| 873 | } else { |
| 874 | struct hash_node * hn = ggml_gallocr_hash_get(galloc, t: src); |
| 875 | node_alloc->src[j].buffer_id = hn->buffer_id; |
| 876 | node_alloc->src[j].addr = hn->addr; |
| 877 | node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(buft: galloc->bufts[hn->buffer_id], tensor: src); |
| 878 | } |
| 879 | } |
| 880 | } |
| 881 | if (galloc->n_leafs < graph->n_leafs) { |
| 882 | free(ptr: galloc->leaf_allocs); |
| 883 | galloc->leaf_allocs = calloc(nmemb: graph->n_leafs, size: sizeof(galloc->leaf_allocs[0])); |
| 884 | GGML_ASSERT(galloc->leaf_allocs != NULL); |
| 885 | } |
| 886 | galloc->n_leafs = graph->n_leafs; |
| 887 | for (int i = 0; i < graph->n_leafs; i++) { |
| 888 | struct ggml_tensor * leaf = graph->leafs[i]; |
| 889 | struct hash_node * hn = ggml_gallocr_hash_get(galloc, t: leaf); |
| 890 | if (leaf->view_src || leaf->data) { |
| 891 | galloc->leaf_allocs[i].leaf.buffer_id = -1; |
| 892 | galloc->leaf_allocs[i].leaf.addr = GGML_BUFFER_ADDRESS_INVALID; |
| 893 | galloc->leaf_allocs[i].leaf.size_max = 0; |
| 894 | } else { |
| 895 | galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id; |
| 896 | galloc->leaf_allocs[i].leaf.addr = hn->addr; |
| 897 | galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(buft: galloc->bufts[hn->buffer_id], tensor: leaf); |
| 898 | } |
| 899 | } |
| 900 | |
| 901 | // reallocate buffers if needed |
| 902 | for (int i = 0; i < galloc->n_buffers; i++) { |
| 903 | // if the buffer type is used multiple times, we reuse the same buffer |
| 904 | for (int j = 0; j < i; j++) { |
| 905 | if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) { |
| 906 | galloc->buffers[i] = galloc->buffers[j]; |
| 907 | break; |
| 908 | } |
| 909 | } |
| 910 | |
| 911 | // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views |
| 912 | bool realloc = galloc->buffers[i] == NULL; |
| 913 | size_t new_size = 0; |
| 914 | for (int c = 0; c < galloc->buf_tallocs[i]->n_chunks; c++) { |
| 915 | size_t cur_chunk_size = galloc->buffers[i] ? ggml_vbuffer_chunk_size(buf: galloc->buffers[i], chunk: c) : 0; |
| 916 | size_t new_chunk_size = ggml_dyn_tallocr_max_size(alloc: galloc->buf_tallocs[i], chunk: c); |
| 917 | new_size += new_chunk_size; |
| 918 | if (new_chunk_size > cur_chunk_size) { |
| 919 | realloc = true; |
| 920 | } |
| 921 | } |
| 922 | if (realloc) { |
| 923 | #ifndef NDEBUG |
| 924 | size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0; |
| 925 | GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n" , __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0); |
| 926 | #endif |
| 927 | |
| 928 | ggml_vbuffer_free(buf: galloc->buffers[i]); |
| 929 | galloc->buffers[i] = ggml_vbuffer_alloc(buft: galloc->bufts[i], talloc: galloc->buf_tallocs[i], usage: GGML_BACKEND_BUFFER_USAGE_COMPUTE); |
| 930 | if (galloc->buffers[i] == NULL) { |
| 931 | GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n" , __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size); |
| 932 | return false; |
| 933 | } |
| 934 | } |
| 935 | } |
| 936 | |
| 937 | return true; |
| 938 | } |
| 939 | |
| 940 | bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) { |
| 941 | return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL); |
| 942 | } |
| 943 | |
| 944 | static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, struct tensor_alloc * tensor_alloc) { |
| 945 | int buffer_id = tensor_alloc->buffer_id; |
| 946 | assert(tensor->data || tensor->view_src || ggml_backend_buft_get_alloc_size(galloc->bufts[buffer_id], tensor) <= tensor_alloc->size_max); |
| 947 | |
| 948 | if (tensor->view_src != NULL) { |
| 949 | if (tensor->buffer == NULL) { |
| 950 | assert(tensor_alloc->addr.offset == SIZE_MAX); |
| 951 | if (tensor->view_src->buffer == NULL) { |
| 952 | // this tensor was allocated without ggml-backend |
| 953 | return; |
| 954 | } |
| 955 | ggml_backend_view_init(tensor); |
| 956 | } |
| 957 | } else { |
| 958 | if (tensor->data == NULL) { |
| 959 | assert(tensor_alloc->addr.offset != SIZE_MAX); |
| 960 | assert(ggml_backend_buft_get_alloc_size(galloc->bufts[buffer_id], tensor) <= tensor_alloc->size_max); |
| 961 | ggml_vbuffer_tensor_alloc(buf: galloc->buffers[buffer_id], tensor, buf_addr: tensor_alloc->addr); |
| 962 | } else { |
| 963 | if (tensor->buffer == NULL) { |
| 964 | // this tensor was allocated without ggml-backend |
| 965 | return; |
| 966 | } |
| 967 | } |
| 968 | } |
| 969 | } |
| 970 | |
| 971 | static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) { |
| 972 | size_t node_size = 0; |
| 973 | if (!node->data && !node->view_src) { |
| 974 | // If we previously had data but don't now then reallocate |
| 975 | if (talloc->buffer_id < 0) { |
| 976 | return false; |
| 977 | } |
| 978 | node_size = ggml_backend_buft_get_alloc_size(buft: galloc->bufts[talloc->buffer_id], tensor: node); |
| 979 | } |
| 980 | return talloc->size_max >= node_size; |
| 981 | } |
| 982 | |
| 983 | static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) { |
| 984 | if (galloc->n_nodes != graph->n_nodes) { |
| 985 | #ifndef NDEBUG |
| 986 | GGML_LOG_DEBUG("%s: graph has different number of nodes\n" , __func__); |
| 987 | #endif |
| 988 | return true; |
| 989 | } |
| 990 | |
| 991 | if (galloc->n_leafs != graph->n_leafs) { |
| 992 | #ifndef NDEBUG |
| 993 | GGML_LOG_DEBUG("%s: graph has different number of leafs\n" , __func__); |
| 994 | #endif |
| 995 | return true; |
| 996 | } |
| 997 | |
| 998 | for (int i = 0; i < graph->n_nodes; i++) { |
| 999 | struct ggml_tensor * node = graph->nodes[i]; |
| 1000 | struct node_alloc * node_alloc = &galloc->node_allocs[i]; |
| 1001 | |
| 1002 | if (!ggml_gallocr_node_needs_realloc(galloc, node, talloc: &node_alloc->dst)) { |
| 1003 | #ifndef NDEBUG |
| 1004 | GGML_LOG_DEBUG("%s: node %s is not valid\n" , __func__, node->name); |
| 1005 | #endif |
| 1006 | return true; |
| 1007 | } |
| 1008 | |
| 1009 | for (int j = 0; j < GGML_MAX_SRC; j++) { |
| 1010 | struct ggml_tensor * src = node->src[j]; |
| 1011 | if (src == NULL) { |
| 1012 | continue; |
| 1013 | } |
| 1014 | if (!ggml_gallocr_node_needs_realloc(galloc, node: src, talloc: &node_alloc->src[j])) { |
| 1015 | #ifndef NDEBUG |
| 1016 | GGML_LOG_DEBUG("%s: src %d (%s) of node %s is not valid\n" , __func__, j, src->name, node->name); |
| 1017 | #endif |
| 1018 | return true; |
| 1019 | } |
| 1020 | } |
| 1021 | } |
| 1022 | |
| 1023 | return false; |
| 1024 | } |
| 1025 | |
| 1026 | bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph) { |
| 1027 | if (ggml_gallocr_needs_realloc(galloc, graph)) { |
| 1028 | if (galloc->n_buffers == 1) { |
| 1029 | #ifndef NDEBUG |
| 1030 | GGML_LOG_DEBUG("%s: reallocating buffers automatically\n" , __func__); |
| 1031 | #endif |
| 1032 | if (!ggml_gallocr_reserve(galloc, graph)) { |
| 1033 | return false; |
| 1034 | } |
| 1035 | } else { |
| 1036 | #ifndef NDEBUG |
| 1037 | GGML_LOG_DEBUG("%s: cannot reallocate multi buffer graph automatically, call reserve\n" , __func__); |
| 1038 | #endif |
| 1039 | return false; |
| 1040 | } |
| 1041 | } |
| 1042 | |
| 1043 | // reset buffers |
| 1044 | for (int i = 0; i < galloc->n_buffers; i++) { |
| 1045 | if (galloc->buffers[i] != NULL) { |
| 1046 | ggml_vbuffer_reset(buf: galloc->buffers[i]); |
| 1047 | } |
| 1048 | } |
| 1049 | |
| 1050 | // allocate the graph tensors from the previous assignments |
| 1051 | // leafs |
| 1052 | for (int i = 0; i < graph->n_leafs; i++) { |
| 1053 | struct ggml_tensor * leaf = graph->leafs[i]; |
| 1054 | struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i]; |
| 1055 | ggml_gallocr_init_tensor(galloc, tensor: leaf, tensor_alloc: &leaf_alloc->leaf); |
| 1056 | } |
| 1057 | // nodes |
| 1058 | for (int i = 0; i < graph->n_nodes; i++) { |
| 1059 | struct ggml_tensor * node = graph->nodes[i]; |
| 1060 | struct node_alloc * node_alloc = &galloc->node_allocs[i]; |
| 1061 | for (int j = 0; j < GGML_MAX_SRC; j++) { |
| 1062 | struct ggml_tensor * src = node->src[j]; |
| 1063 | if (src == NULL) { |
| 1064 | continue; |
| 1065 | } |
| 1066 | ggml_gallocr_init_tensor(galloc, tensor: src, tensor_alloc: &node_alloc->src[j]); |
| 1067 | } |
| 1068 | ggml_gallocr_init_tensor(galloc, tensor: node, tensor_alloc: &node_alloc->dst); |
| 1069 | } |
| 1070 | |
| 1071 | return true; |
| 1072 | } |
| 1073 | |
| 1074 | size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) { |
| 1075 | GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers); |
| 1076 | |
| 1077 | if (galloc->buffers[buffer_id] == NULL) { |
| 1078 | return 0; |
| 1079 | } |
| 1080 | |
| 1081 | for (int i = 0; i < buffer_id; i++) { |
| 1082 | if (galloc->buffers[i] == galloc->buffers[buffer_id]) { |
| 1083 | // this buffer is the same as a previous one due to the same buffer type being used multiple times |
| 1084 | // only return the buffer size the first time it appears to avoid double counting |
| 1085 | return 0; |
| 1086 | } |
| 1087 | } |
| 1088 | |
| 1089 | return ggml_vbuffer_size(buf: galloc->buffers[buffer_id]); |
| 1090 | } |
| 1091 | |
| 1092 | // utils |
| 1093 | |
| 1094 | static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) { |
| 1095 | for (size_t i = 0; i < *n_buffers; i++) { |
| 1096 | ggml_backend_buffer_free(buffer: (*buffers)[i]); |
| 1097 | } |
| 1098 | free(ptr: *buffers); |
| 1099 | } |
| 1100 | |
| 1101 | static bool alloc_tensor_range(struct ggml_context * ctx, |
| 1102 | struct ggml_tensor * first, struct ggml_tensor * last, |
| 1103 | ggml_backend_buffer_type_t buft, size_t size, |
| 1104 | ggml_backend_buffer_t ** buffers, size_t * n_buffers) { |
| 1105 | |
| 1106 | ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size); |
| 1107 | if (buffer == NULL) { |
| 1108 | GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n" , __func__, ggml_backend_buft_name(buft), size); |
| 1109 | free_buffers(buffers, n_buffers); |
| 1110 | return false; |
| 1111 | } |
| 1112 | |
| 1113 | *buffers = realloc(ptr: *buffers, size: sizeof(ggml_backend_buffer_t) * (*n_buffers + 1)); |
| 1114 | (*buffers)[(*n_buffers)++] = buffer; |
| 1115 | |
| 1116 | struct ggml_tallocr tallocr = ggml_tallocr_new(buffer); |
| 1117 | |
| 1118 | for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, tensor: t)) { |
| 1119 | enum ggml_status status = GGML_STATUS_SUCCESS; |
| 1120 | if (t->data == NULL) { |
| 1121 | if (t->view_src == NULL) { |
| 1122 | status = ggml_tallocr_alloc(talloc: &tallocr, tensor: t); |
| 1123 | } else if (t->buffer == NULL) { |
| 1124 | status = ggml_backend_view_init(tensor: t); |
| 1125 | } |
| 1126 | } else { |
| 1127 | if (t->view_src != NULL && t->buffer == NULL) { |
| 1128 | // view of a pre-allocated tensor |
| 1129 | status = ggml_backend_view_init(tensor: t); |
| 1130 | } |
| 1131 | } |
| 1132 | if (status != GGML_STATUS_SUCCESS) { |
| 1133 | GGML_LOG_ERROR("%s: failed to initialize tensor %s\n" , __func__, t->name); |
| 1134 | free_buffers(buffers, n_buffers); |
| 1135 | return false; |
| 1136 | } |
| 1137 | } |
| 1138 | |
| 1139 | return true; |
| 1140 | } |
| 1141 | |
| 1142 | ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) { |
| 1143 | GGML_ASSERT(ggml_get_no_alloc(ctx) == true); |
| 1144 | |
| 1145 | size_t alignment = ggml_backend_buft_get_alignment(buft); |
| 1146 | size_t max_size = ggml_backend_buft_get_max_size(buft); |
| 1147 | |
| 1148 | ggml_backend_buffer_t * buffers = NULL; |
| 1149 | size_t n_buffers = 0; |
| 1150 | |
| 1151 | size_t cur_buf_size = 0; |
| 1152 | struct ggml_tensor * first = ggml_get_first_tensor(ctx); |
| 1153 | for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, tensor: t)) { |
| 1154 | size_t this_size = 0; |
| 1155 | if (t->data == NULL && t->view_src == NULL) { |
| 1156 | this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment); |
| 1157 | } |
| 1158 | |
| 1159 | if (cur_buf_size > 0 && (cur_buf_size + this_size) > max_size) { |
| 1160 | // allocate tensors in the current buffer |
| 1161 | if (!alloc_tensor_range(ctx, first, last: t, buft, size: cur_buf_size, buffers: &buffers, n_buffers: &n_buffers)) { |
| 1162 | return NULL; |
| 1163 | } |
| 1164 | first = t; |
| 1165 | cur_buf_size = this_size; |
| 1166 | } else { |
| 1167 | cur_buf_size += this_size; |
| 1168 | } |
| 1169 | } |
| 1170 | |
| 1171 | // allocate remaining tensors |
| 1172 | if (cur_buf_size > 0) { |
| 1173 | if (!alloc_tensor_range(ctx, first, NULL, buft, size: cur_buf_size, buffers: &buffers, n_buffers: &n_buffers)) { |
| 1174 | return NULL; |
| 1175 | } |
| 1176 | } |
| 1177 | |
| 1178 | if (n_buffers == 0) { |
| 1179 | #ifndef NDEBUG |
| 1180 | GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n" , __func__); |
| 1181 | #endif |
| 1182 | return NULL; |
| 1183 | } |
| 1184 | |
| 1185 | ggml_backend_buffer_t buffer; |
| 1186 | if (n_buffers == 1) { |
| 1187 | buffer = buffers[0]; |
| 1188 | } else { |
| 1189 | buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers); |
| 1190 | } |
| 1191 | free(ptr: buffers); |
| 1192 | return buffer; |
| 1193 | } |
| 1194 | |
| 1195 | ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) { |
| 1196 | return ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft: ggml_backend_get_default_buffer_type(backend)); |
| 1197 | } |
| 1198 | |