src/gallium/drivers/r600/compute_memory_pool.c

   1 /*
   2  * Permission is hereby granted, free of charge, to any person obtaining a
   3  * copy of this software and associated documentation files (the "Software"),
   4  * to deal in the Software without restriction, including without limitation
   5  * on the rights to use, copy, modify, merge, publish, distribute, sub
   6  * license, and/or sell copies of the Software, and to permit persons to whom
   7  * the Software is furnished to do so, subject to the following conditions:
   8  *
   9  * The above copyright notice and this permission notice (including the next
  10  * paragraph) shall be included in all copies or substantial portions of the
  11  * Software.
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * Authors:
  22  *      Adam Rak <adam.rak@streamnovation.com>
  23  */
  24
  25 #include "pipe/p_defines.h"
  26 #include "pipe/p_state.h"
  27 #include "pipe/p_context.h"
  28 #include "util/u_blitter.h"
  29 #include "util/u_double_list.h"
  30 #include "util/u_transfer.h"
  31 #include "util/u_surface.h"
  32 #include "util/u_pack_color.h"
  33 #include "util/u_math.h"
  34 #include "util/u_memory.h"
  35 #include "util/u_inlines.h"
  36 #include "util/u_framebuffer.h"
  37 #include "r600_shader.h"
  38 #include "r600_pipe.h"
  39 #include "r600_formats.h"
  40 #include "compute_memory_pool.h"
  41 #include "evergreen_compute.h"
  42 #include "evergreen_compute_internal.h"
  43 #include <inttypes.h>
  44
  45 #define ITEM_ALIGNMENT 1024
  46 /**
  47  * Creates a new pool
  48  */
  49 struct compute_memory_pool* compute_memory_pool_new(
  50         struct r600_screen * rscreen)
  51 {
  52         struct compute_memory_pool* pool = (struct compute_memory_pool*)
  53                                 CALLOC(sizeof(struct compute_memory_pool), 1);
  54         if (pool == NULL)
  55                 return NULL;
  56
  57         COMPUTE_DBG(rscreen, "* compute_memory_pool_new()\n");
  58
  59         pool->screen = rscreen;
  60         pool->item_list = (struct list_head *)
  61                                 CALLOC(sizeof(struct list_head), 1);
  62         pool->unallocated_list = (struct list_head *)
  63                                 CALLOC(sizeof(struct list_head), 1);
  64         list_inithead(pool->item_list);
  65         list_inithead(pool->unallocated_list);
  66         return pool;
  67 }
  68
  69 static void compute_memory_pool_init(struct compute_memory_pool * pool,
  70         unsigned initial_size_in_dw)
  71 {
  72
  73         COMPUTE_DBG(pool->screen, "* compute_memory_pool_init() initial_size_in_dw = %ld\n",
  74                 initial_size_in_dw);
  75
  76         pool->shadow = (uint32_t*)CALLOC(initial_size_in_dw, 4);
  77         if (pool->shadow == NULL)
  78                 return;
  79
  80         pool->size_in_dw = initial_size_in_dw;
  81         pool->bo = (struct r600_resource*)r600_compute_buffer_alloc_vram(pool->screen,
  82                                                         pool->size_in_dw * 4);
  83 }
  84
  85 /**
  86  * Frees all stuff in the pool and the pool struct itself too
  87  */
  88 void compute_memory_pool_delete(struct compute_memory_pool* pool)
  89 {
  90         COMPUTE_DBG(pool->screen, "* compute_memory_pool_delete()\n");
  91         free(pool->shadow);
  92         if (pool->bo) {
  93                 pool->screen->b.b.resource_destroy((struct pipe_screen *)
  94                         pool->screen, (struct pipe_resource *)pool->bo);
  95         }
  96         free(pool);
  97 }
  98
  99 /**
 100  * Searches for an empty space in the pool, return with the pointer to the
 101  * allocatable space in the pool, returns -1 on failure.
 102  */
 103 int64_t compute_memory_prealloc_chunk(
 104         struct compute_memory_pool* pool,
 105         int64_t size_in_dw)
 106 {
 107         struct compute_memory_item *item;
 108
 109         int last_end = 0;
 110
 111         assert(size_in_dw <= pool->size_in_dw);
 112
 113         COMPUTE_DBG(pool->screen, "* compute_memory_prealloc_chunk() size_in_dw = %ld\n",
 114                 size_in_dw);
 115
 116         LIST_FOR_EACH_ENTRY(item, pool->item_list, link) {
 117                 if (last_end + size_in_dw <= item->start_in_dw) {
 118                         return last_end;
 119                 }
 120
 121                 last_end = item->start_in_dw + align(item->size_in_dw, ITEM_ALIGNMENT);
 122         }
 123
 124         if (pool->size_in_dw - last_end < size_in_dw) {
 125                 return -1;
 126         }
 127
 128         return last_end;
 129 }
 130
 131 /**
 132  *  Search for the chunk where we can link our new chunk after it.
 133  */
 134 struct list_head *compute_memory_postalloc_chunk(
 135         struct compute_memory_pool* pool,
 136         int64_t start_in_dw)
 137 {
 138         struct compute_memory_item *item;
 139         struct compute_memory_item *next;
 140         struct list_head *next_link;
 141
 142         COMPUTE_DBG(pool->screen, "* compute_memory_postalloc_chunck() start_in_dw = %ld\n",
 143                 start_in_dw);
 144
 145         /* Check if we can insert it in the front of the list */
 146         item = LIST_ENTRY(struct compute_memory_item, pool->item_list->next, link);
 147         if (LIST_IS_EMPTY(pool->item_list) || item->start_in_dw > start_in_dw) {
 148                 return pool->item_list;
 149         }
 150
 151         LIST_FOR_EACH_ENTRY(item, pool->item_list, link) {
 152                 next_link = item->link.next;
 153
 154                 if (next_link != pool->item_list) {
 155                         next = container_of(next_link, item, link);
 156                         if (item->start_in_dw < start_in_dw
 157                                 && next->start_in_dw > start_in_dw) {
 158                                 return &item->link;
 159                         }
 160                 }
 161                 else {
 162                         /* end of chain */
 163                         assert(item->start_in_dw < start_in_dw);
 164                         return &item->link;
 165                 }
 166         }
 167
 168         assert(0 && "unreachable");
 169         return NULL;
 170 }
 171
 172 /**
 173  * Reallocates pool, conserves data.
 174  * @returns -1 if it fails, 0 otherwise
 175  */
 176 int compute_memory_grow_pool(struct compute_memory_pool* pool,
 177         struct pipe_context * pipe, int new_size_in_dw)
 178 {
 179         COMPUTE_DBG(pool->screen, "* compute_memory_grow_pool() "
 180                 "new_size_in_dw = %d (%d bytes)\n",
 181                 new_size_in_dw, new_size_in_dw * 4);
 182
 183         assert(new_size_in_dw >= pool->size_in_dw);
 184
 185         if (!pool->bo) {
 186                 compute_memory_pool_init(pool, MAX2(new_size_in_dw, 1024 * 16));
 187                 if (pool->shadow == NULL)
 188                         return -1;
 189         } else {
 190                 new_size_in_dw = align(new_size_in_dw, ITEM_ALIGNMENT);
 191
 192                 COMPUTE_DBG(pool->screen, "  Aligned size = %d (%d bytes)\n",
 193                         new_size_in_dw, new_size_in_dw * 4);
 194
 195                 compute_memory_shadow(pool, pipe, 1);
 196                 pool->shadow = realloc(pool->shadow, new_size_in_dw*4);
 197                 if (pool->shadow == NULL)
 198                         return -1;
 199
 200                 pool->size_in_dw = new_size_in_dw;
 201                 pool->screen->b.b.resource_destroy(
 202                         (struct pipe_screen *)pool->screen,
 203                         (struct pipe_resource *)pool->bo);
 204                 pool->bo = (struct r600_resource*)r600_compute_buffer_alloc_vram(
 205                                                         pool->screen,
 206                                                         pool->size_in_dw * 4);
 207                 compute_memory_shadow(pool, pipe, 0);
 208         }
 209
 210         return 0;
 211 }
 212
 213 /**
 214  * Copy pool from device to host, or host to device.
 215  */
 216 void compute_memory_shadow(struct compute_memory_pool* pool,
 217         struct pipe_context * pipe, int device_to_host)
 218 {
 219         struct compute_memory_item chunk;
 220
 221         COMPUTE_DBG(pool->screen, "* compute_memory_shadow() device_to_host = %d\n",
 222                 device_to_host);
 223
 224         chunk.id = 0;
 225         chunk.start_in_dw = 0;
 226         chunk.size_in_dw = pool->size_in_dw;
 227         compute_memory_transfer(pool, pipe, device_to_host, &chunk,
 228                                 pool->shadow, 0, pool->size_in_dw*4);
 229 }
 230
 231 /**
 232  * Allocates pending allocations in the pool
 233  * @returns -1 if it fails, 0 otherwise
 234  */
 235 int compute_memory_finalize_pending(struct compute_memory_pool* pool,
 236         struct pipe_context * pipe)
 237 {
 238         struct compute_memory_item *item, *next;
 239
 240         int64_t allocated = 0;
 241         int64_t unallocated = 0;
 242
 243         int err = 0;
 244
 245         COMPUTE_DBG(pool->screen, "* compute_memory_finalize_pending()\n");
 246
 247         LIST_FOR_EACH_ENTRY(item, pool->item_list, link) {
 248                 COMPUTE_DBG(pool->screen, "  + list: offset = %i id = %i size = %i "
 249                         "(%i bytes)\n",item->start_in_dw, item->id,
 250                         item->size_in_dw, item->size_in_dw * 4);
 251         }
 252
 253         /* Calculate the total allocated size */
 254         LIST_FOR_EACH_ENTRY(item, pool->item_list, link) {
 255                 allocated += align(item->size_in_dw, ITEM_ALIGNMENT);
 256         }
 257
 258         /* Calculate the total unallocated size of the items that
 259          * will be promoted to the pool */
 260         LIST_FOR_EACH_ENTRY(item, pool->unallocated_list, link) {
 261                 if (item->status & ITEM_FOR_PROMOTING)
 262                         unallocated += align(item->size_in_dw, ITEM_ALIGNMENT);
 263         }
 264
 265         /* If we require more space than the size of the pool, then grow the
 266          * pool.
 267          *
 268          * XXX: I'm pretty sure this won't work.  Imagine this scenario:
 269          *
 270          * Offset Item Size
 271          *   0    A    50
 272          * 200    B    50
 273          * 400    C    50
 274          *
 275          * Total size = 450
 276          * Allocated size = 150
 277          * Pending Item D Size = 200
 278          *
 279          * In this case, there are 300 units of free space in the pool, but
 280          * they aren't contiguous, so it will be impossible to allocate Item D.
 281          */
 282         if (pool->size_in_dw < allocated + unallocated) {
 283                 err = compute_memory_grow_pool(pool, pipe, allocated + unallocated);
 284                 if (err == -1)
 285                         return -1;
 286         }
 287
 288         /* Loop through all the unallocated items, check if they are marked
 289          * for promoting, allocate space for them and add them to the item_list. */
 290         LIST_FOR_EACH_ENTRY_SAFE(item, next, pool->unallocated_list, link) {
 291                 if (item->status & ITEM_FOR_PROMOTING) {
 292                         err = compute_memory_promote_item(pool, item, pipe, allocated);
 293                         item->status ^= ITEM_FOR_PROMOTING;
 294
 295                         allocated += align(item->size_in_dw, ITEM_ALIGNMENT);
 296
 297                         if (err == -1)
 298                                 return -1;
 299                 }
 300         }
 301
 302         return 0;
 303 }
 304
 305 int compute_memory_promote_item(struct compute_memory_pool *pool,
 306                 struct compute_memory_item *item, struct pipe_context *pipe,
 307                 int64_t allocated)
 308 {
 309         struct pipe_screen *screen = (struct pipe_screen *)pool->screen;
 310         struct r600_context *rctx = (struct r600_context *)pipe;
 311         struct pipe_resource *dst = (struct pipe_resource *)pool->bo;
 312         struct pipe_resource *src = (struct pipe_resource *)item->real_buffer;
 313         struct pipe_box box;
 314
 315         struct list_head *pos;
 316         int64_t start_in_dw;
 317         int err = 0;
 318
 319
 320         /* Search for free space in the pool for this item. */
 321         while ((start_in_dw=compute_memory_prealloc_chunk(pool,
 322                                         item->size_in_dw)) == -1) {
 323                 int64_t need = item->size_in_dw + 2048 -
 324                         (pool->size_in_dw - allocated);
 325
 326                 if (need < 0) {
 327                         need = pool->size_in_dw / 10;
 328                 }
 329
 330                 need = align(need, ITEM_ALIGNMENT);
 331
 332                 err = compute_memory_grow_pool(pool,
 333                                 pipe,
 334                                 pool->size_in_dw + need);
 335
 336                 if (err == -1)
 337                         return -1;
 338         }
 339         COMPUTE_DBG(pool->screen, "  + Found space for Item %p id = %u "
 340                         "start_in_dw = %u (%u bytes) size_in_dw = %u (%u bytes)\n",
 341                         item, item->id, start_in_dw, start_in_dw * 4,
 342                         item->size_in_dw, item->size_in_dw * 4);
 343
 344         /* Remove the item from the unallocated list */
 345         list_del(&item->link);
 346
 347         /* Add it back to the item_list */
 348         pos = compute_memory_postalloc_chunk(pool, start_in_dw);
 349         list_add(&item->link, pos);
 350         item->start_in_dw = start_in_dw;
 351
 352         u_box_1d(0, item->size_in_dw * 4, &box);
 353
 354         rctx->b.b.resource_copy_region(pipe,
 355                         dst, 0, item->start_in_dw * 4, 0 ,0,
 356                         src, 0, &box);
 357
 358         /* We check if the item is mapped for reading.
 359          * In this case, we need to keep the temporary buffer 'alive'
 360          * because it is possible to keep a map active for reading
 361          * while a kernel (that reads from it) executes */
 362         if (!(item->status & ITEM_MAPPED_FOR_READING)) {
 363                 pool->screen->b.b.resource_destroy(screen, src);
 364                 item->real_buffer = NULL;
 365         }
 366
 367         return 0;
 368 }
 369
 370 void compute_memory_demote_item(struct compute_memory_pool *pool,
 371         struct compute_memory_item *item, struct pipe_context *pipe)
 372 {
 373         struct r600_context *rctx = (struct r600_context *)pipe;
 374         struct pipe_resource *src = (struct pipe_resource *)pool->bo;
 375         struct pipe_resource *dst;
 376         struct pipe_box box;
 377
 378         /* First, we remove the item from the item_list */
 379         list_del(&item->link);
 380
 381         /* Now we add it to the unallocated list */
 382         list_addtail(&item->link, pool->unallocated_list);
 383
 384         /* We check if the intermediate buffer exists, and if it
 385          * doesn't, we create it again */
 386         if (item->real_buffer == NULL) {
 387                 item->real_buffer = (struct r600_resource*)r600_compute_buffer_alloc_vram(
 388                                 pool->screen, item->size_in_dw * 4);
 389         }
 390
 391         dst = (struct pipe_resource *)item->real_buffer;
 392
 393         /* We transfer the memory from the item in the pool to the
 394          * temporary buffer */
 395         u_box_1d(item->start_in_dw * 4, item->size_in_dw * 4, &box);
 396
 397         rctx->b.b.resource_copy_region(pipe,
 398                 dst, 0, 0, 0, 0,
 399                 src, 0, &box);
 400
 401         /* Remember to mark the buffer as 'pending' by setting start_in_dw to -1 */
 402         item->start_in_dw = -1;
 403 }
 404
 405 void compute_memory_free(struct compute_memory_pool* pool, int64_t id)
 406 {
 407         struct compute_memory_item *item, *next;
 408         struct pipe_screen *screen = (struct pipe_screen *)pool->screen;
 409         struct pipe_resource *res;
 410
 411         COMPUTE_DBG(pool->screen, "* compute_memory_free() id + %ld \n", id);
 412
 413         LIST_FOR_EACH_ENTRY_SAFE(item, next, pool->item_list, link) {
 414
 415                 if (item->id == id) {
 416                         list_del(&item->link);
 417
 418                         if (item->real_buffer) {
 419                                 res = (struct pipe_resource *)item->real_buffer;
 420                                 pool->screen->b.b.resource_destroy(
 421                                                 screen, res);
 422                         }
 423
 424                         free(item);
 425
 426                         return;
 427                 }
 428         }
 429
 430         LIST_FOR_EACH_ENTRY_SAFE(item, next, pool->unallocated_list, link) {
 431
 432                 if (item->id == id) {
 433                         list_del(&item->link);
 434
 435                         if (item->real_buffer) {
 436                                 res = (struct pipe_resource *)item->real_buffer;
 437                                 pool->screen->b.b.resource_destroy(
 438                                                 screen, res);
 439                         }
 440
 441                         free(item);
 442
 443                         return;
 444                 }
 445         }
 446
 447         fprintf(stderr, "Internal error, invalid id %"PRIi64" "
 448                 "for compute_memory_free\n", id);
 449
 450         assert(0 && "error");
 451 }
 452
 453 /**
 454  * Creates pending allocations
 455  */
 456 struct compute_memory_item* compute_memory_alloc(
 457         struct compute_memory_pool* pool,
 458         int64_t size_in_dw)
 459 {
 460         struct compute_memory_item *new_item = NULL;
 461
 462         COMPUTE_DBG(pool->screen, "* compute_memory_alloc() size_in_dw = %ld (%ld bytes)\n",
 463                         size_in_dw, 4 * size_in_dw);
 464
 465         new_item = (struct compute_memory_item *)
 466                                 CALLOC(sizeof(struct compute_memory_item), 1);
 467         if (new_item == NULL)
 468                 return NULL;
 469
 470         new_item->size_in_dw = size_in_dw;
 471         new_item->start_in_dw = -1; /* mark pending */
 472         new_item->id = pool->next_id++;
 473         new_item->pool = pool;
 474         new_item->real_buffer = (struct r600_resource*)r600_compute_buffer_alloc_vram(
 475                                                         pool->screen, size_in_dw * 4);
 476
 477         list_addtail(&new_item->link, pool->unallocated_list);
 478
 479         COMPUTE_DBG(pool->screen, "  + Adding item %p id = %u size = %u (%u bytes)\n",
 480                         new_item, new_item->id, new_item->size_in_dw,
 481                         new_item->size_in_dw * 4);
 482         return new_item;
 483 }
 484
 485 /**
 486  * Transfer data host<->device, offset and size is in bytes
 487  */
 488 void compute_memory_transfer(
 489         struct compute_memory_pool* pool,
 490         struct pipe_context * pipe,
 491         int device_to_host,
 492         struct compute_memory_item* chunk,
 493         void* data,
 494         int offset_in_chunk,
 495         int size)
 496 {
 497         int64_t aligned_size = pool->size_in_dw;
 498         struct pipe_resource* gart = (struct pipe_resource*)pool->bo;
 499         int64_t internal_offset = chunk->start_in_dw*4 + offset_in_chunk;
 500
 501         struct pipe_transfer *xfer;
 502         uint32_t *map;
 503
 504         assert(gart);
 505
 506         COMPUTE_DBG(pool->screen, "* compute_memory_transfer() device_to_host = %d, "
 507                 "offset_in_chunk = %d, size = %d\n", device_to_host,
 508                 offset_in_chunk, size);
 509
 510         if (device_to_host) {
 511                 map = pipe->transfer_map(pipe, gart, 0, PIPE_TRANSFER_READ,
 512                         &(struct pipe_box) { .width = aligned_size * 4,
 513                         .height = 1, .depth = 1 }, &xfer);
 514                 assert(xfer);
 515                 assert(map);
 516                 memcpy(data, map + internal_offset, size);
 517                 pipe->transfer_unmap(pipe, xfer);
 518         } else {
 519                 map = pipe->transfer_map(pipe, gart, 0, PIPE_TRANSFER_WRITE,
 520                         &(struct pipe_box) { .width = aligned_size * 4,
 521                         .height = 1, .depth = 1 }, &xfer);
 522                 assert(xfer);
 523                 assert(map);
 524                 memcpy(map + internal_offset, data, size);
 525                 pipe->transfer_unmap(pipe, xfer);
 526         }
 527 }
 528
 529 /**
 530  * Transfer data between chunk<->data, it is for VRAM<->GART transfers
 531  */
 532 void compute_memory_transfer_direct(
 533         struct compute_memory_pool* pool,
 534         int chunk_to_data,
 535         struct compute_memory_item* chunk,
 536         struct r600_resource* data,
 537         int offset_in_chunk,
 538         int offset_in_data,
 539         int size)
 540 {
 541         ///TODO: DMA
 542 }