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_memory.h"
  34 #include "util/u_inlines.h"
  35 #include "util/u_framebuffer.h"
  36 #include "r600_shader.h"
  37 #include "r600_pipe.h"
  38 #include "r600_formats.h"
  39 #include "compute_memory_pool.h"
  40 #include "evergreen_compute.h"
  41 #include "evergreen_compute_internal.h"
  42 #include <inttypes.h>
  43
  44 /**
  45  * Creates a new pool
  46  */
  47 struct compute_memory_pool* compute_memory_pool_new(
  48         struct r600_screen * rscreen)
  49 {
  50         struct compute_memory_pool* pool = (struct compute_memory_pool*)
  51                                 CALLOC(sizeof(struct compute_memory_pool), 1);
  52
  53         COMPUTE_DBG(rscreen, "* compute_memory_pool_new()\n");
  54
  55         pool->screen = rscreen;
  56         return pool;
  57 }
  58
  59 static void compute_memory_pool_init(struct compute_memory_pool * pool,
  60         unsigned initial_size_in_dw)
  61 {
  62
  63         COMPUTE_DBG(pool->screen, "* compute_memory_pool_init() initial_size_in_dw = %ld\n",
  64                 initial_size_in_dw);
  65
  66         pool->shadow = (uint32_t*)CALLOC(initial_size_in_dw, 4);
  67         pool->next_id = 1;
  68         pool->size_in_dw = initial_size_in_dw;
  69         pool->bo = (struct r600_resource*)r600_compute_buffer_alloc_vram(pool->screen,
  70                                                         pool->size_in_dw * 4);
  71 }
  72
  73 /**
  74  * Frees all stuff in the pool and the pool struct itself too
  75  */
  76 void compute_memory_pool_delete(struct compute_memory_pool* pool)
  77 {
  78         COMPUTE_DBG(pool->screen, "* compute_memory_pool_delete()\n");
  79         free(pool->shadow);
  80         if (pool->bo) {
  81                 pool->screen->b.b.resource_destroy((struct pipe_screen *)
  82                         pool->screen, (struct pipe_resource *)pool->bo);
  83         }
  84         free(pool);
  85 }
  86
  87 /**
  88  * Searches for an empty space in the pool, return with the pointer to the
  89  * allocatable space in the pool, returns -1 on failure.
  90  */
  91 int64_t compute_memory_prealloc_chunk(
  92         struct compute_memory_pool* pool,
  93         int64_t size_in_dw)
  94 {
  95         struct compute_memory_item *item;
  96
  97         int last_end = 0;
  98
  99         assert(size_in_dw <= pool->size_in_dw);
 100
 101         COMPUTE_DBG(pool->screen, "* compute_memory_prealloc_chunk() size_in_dw = %ld\n",
 102                 size_in_dw);
 103
 104         for (item = pool->item_list; item; item = item->next) {
 105                 if (item->start_in_dw > -1) {
 106                         if (item->start_in_dw-last_end > size_in_dw) {
 107                                 return last_end;
 108                         }
 109
 110                         last_end = item->start_in_dw + item->size_in_dw;
 111                         last_end += (1024 - last_end % 1024);
 112                 }
 113         }
 114
 115         if (pool->size_in_dw - last_end < size_in_dw) {
 116                 return -1;
 117         }
 118
 119         return last_end;
 120 }
 121
 122 /**
 123  *  Search for the chunk where we can link our new chunk after it.
 124  */
 125 struct compute_memory_item* compute_memory_postalloc_chunk(
 126         struct compute_memory_pool* pool,
 127         int64_t start_in_dw)
 128 {
 129         struct compute_memory_item* item;
 130
 131         COMPUTE_DBG(pool->screen, "* compute_memory_postalloc_chunck() start_in_dw = %ld\n",
 132                 start_in_dw);
 133
 134         /* Check if we can insert it in the front of the list */
 135         if (pool->item_list && pool->item_list->start_in_dw > start_in_dw) {
 136                 return NULL;
 137         }
 138
 139         for (item = pool->item_list; item; item = item->next) {
 140                 if (item->next) {
 141                         if (item->start_in_dw < start_in_dw
 142                                 && item->next->start_in_dw > start_in_dw) {
 143                                 return item;
 144                         }
 145                 }
 146                 else {
 147                         /* end of chain */
 148                         assert(item->start_in_dw < start_in_dw);
 149                         return item;
 150                 }
 151         }
 152
 153         assert(0 && "unreachable");
 154         return NULL;
 155 }
 156
 157 /**
 158  * Reallocates pool, conserves data
 159  */
 160 void compute_memory_grow_pool(struct compute_memory_pool* pool,
 161         struct pipe_context * pipe, int new_size_in_dw)
 162 {
 163         COMPUTE_DBG(pool->screen, "* compute_memory_grow_pool() "
 164                 "new_size_in_dw = %d (%d bytes)\n",
 165                 new_size_in_dw, new_size_in_dw * 4);
 166
 167         assert(new_size_in_dw >= pool->size_in_dw);
 168
 169         if (!pool->bo) {
 170                 compute_memory_pool_init(pool, MAX2(new_size_in_dw, 1024 * 16));
 171         } else {
 172                 new_size_in_dw += 1024 - (new_size_in_dw % 1024);
 173
 174                 COMPUTE_DBG(pool->screen, "  Aligned size = %d (%d bytes)\n",
 175                         new_size_in_dw, new_size_in_dw * 4);
 176
 177                 compute_memory_shadow(pool, pipe, 1);
 178                 pool->shadow = realloc(pool->shadow, new_size_in_dw*4);
 179                 pool->size_in_dw = new_size_in_dw;
 180                 pool->screen->b.b.resource_destroy(
 181                         (struct pipe_screen *)pool->screen,
 182                         (struct pipe_resource *)pool->bo);
 183                 pool->bo = (struct r600_resource*)r600_compute_buffer_alloc_vram(
 184                                                         pool->screen,
 185                                                         pool->size_in_dw * 4);
 186                 compute_memory_shadow(pool, pipe, 0);
 187         }
 188 }
 189
 190 /**
 191  * Copy pool from device to host, or host to device.
 192  */
 193 void compute_memory_shadow(struct compute_memory_pool* pool,
 194         struct pipe_context * pipe, int device_to_host)
 195 {
 196         struct compute_memory_item chunk;
 197
 198         COMPUTE_DBG(pool->screen, "* compute_memory_shadow() device_to_host = %d\n",
 199                 device_to_host);
 200
 201         chunk.id = 0;
 202         chunk.start_in_dw = 0;
 203         chunk.size_in_dw = pool->size_in_dw;
 204         chunk.prev = chunk.next = NULL;
 205         compute_memory_transfer(pool, pipe, device_to_host, &chunk,
 206                                 pool->shadow, 0, pool->size_in_dw*4);
 207 }
 208
 209 /**
 210  * Allocates pending allocations in the pool
 211  */
 212 void compute_memory_finalize_pending(struct compute_memory_pool* pool,
 213         struct pipe_context * pipe)
 214 {
 215         struct compute_memory_item *pending_list = NULL, *end_p = NULL;
 216         struct compute_memory_item *item, *next;
 217
 218         int64_t allocated = 0;
 219         int64_t unallocated = 0;
 220
 221         int64_t start_in_dw = 0;
 222
 223         COMPUTE_DBG(pool->screen, "* compute_memory_finalize_pending()\n");
 224
 225         for (item = pool->item_list; item; item = item->next) {
 226                 COMPUTE_DBG(pool->screen, "  + list: offset = %i id = %i size = %i "
 227                         "(%i bytes)\n",item->start_in_dw, item->id,
 228                         item->size_in_dw, item->size_in_dw * 4);
 229         }
 230
 231         /* Search through the list of memory items in the pool */
 232         for (item = pool->item_list; item; item = next) {
 233                 next = item->next;
 234
 235                 /* Check if the item is pending. */
 236                 if (item->start_in_dw == -1) {
 237                         /* It is pending, so add it to the pending_list... */
 238                         if (end_p) {
 239                                 end_p->next = item;
 240                         }
 241                         else {
 242                                 pending_list = item;
 243                         }
 244
 245                         /* ... and then remove it from the item list. */
 246                         if (item->prev) {
 247                                 item->prev->next = next;
 248                         }
 249                         else {
 250                                 pool->item_list = next;
 251                         }
 252
 253                         if (next) {
 254                                 next->prev = item->prev;
 255                         }
 256
 257                         /* This sequence makes the item be at the end of the list */
 258                         item->prev = end_p;
 259                         item->next = NULL;
 260                         end_p = item;
 261
 262                         /* Update the amount of space we will need to allocate. */
 263                         unallocated += item->size_in_dw+1024;
 264                 }
 265                 else {
 266                         /* The item is not pendng, so update the amount of space
 267                          * that has already been allocated. */
 268                         allocated += item->size_in_dw;
 269                 }
 270         }
 271
 272         /* If we require more space than the size of the pool, then grow the
 273          * pool.
 274          *
 275          * XXX: I'm pretty sure this won't work.  Imagine this scenario:
 276          *
 277          * Offset Item Size
 278          *   0    A    50
 279          * 200    B    50
 280          * 400    C    50
 281          *
 282          * Total size = 450
 283          * Allocated size = 150
 284          * Pending Item D Size = 200
 285          *
 286          * In this case, there are 300 units of free space in the pool, but
 287          * they aren't contiguous, so it will be impossible to allocate Item D.
 288          */
 289         if (pool->size_in_dw < allocated+unallocated) {
 290                 compute_memory_grow_pool(pool, pipe, allocated+unallocated);
 291         }
 292
 293         /* Loop through all the pending items, allocate space for them and
 294          * add them back to the item_list. */
 295         for (item = pending_list; item; item = next) {
 296                 next = item->next;
 297
 298                 /* Search for free space in the pool for this item. */
 299                 while ((start_in_dw=compute_memory_prealloc_chunk(pool,
 300                                                 item->size_in_dw)) == -1) {
 301                         int64_t need = item->size_in_dw+2048 -
 302                                                 (pool->size_in_dw - allocated);
 303
 304                         need += 1024 - (need % 1024);
 305
 306                         if (need > 0) {
 307                                 compute_memory_grow_pool(pool,
 308                                                 pipe,
 309                                                 pool->size_in_dw + need);
 310                         }
 311                         else {
 312                                 need = pool->size_in_dw / 10;
 313                                 need += 1024 - (need % 1024);
 314                                 compute_memory_grow_pool(pool,
 315                                                 pipe,
 316                                                 pool->size_in_dw + need);
 317                         }
 318                 }
 319                 COMPUTE_DBG(pool->screen, "  + Found space for Item %p id = %u "
 320                         "start_in_dw = %u (%u bytes) size_in_dw = %u (%u bytes)\n",
 321                         item, item->id, start_in_dw, start_in_dw * 4,
 322                         item->size_in_dw, item->size_in_dw * 4);
 323
 324                 item->start_in_dw = start_in_dw;
 325                 item->next = NULL;
 326                 item->prev = NULL;
 327
 328                 if (pool->item_list) {
 329                         struct compute_memory_item *pos;
 330
 331                         pos = compute_memory_postalloc_chunk(pool, start_in_dw);
 332                         if (pos) {
 333                                 item->prev = pos;
 334                                 item->next = pos->next;
 335                                 pos->next = item;
 336                                 if (item->next) {
 337                                         item->next->prev = item;
 338                                 }
 339                         } else {
 340                                 /* Add item to the front of the list */
 341                                 item->next = pool->item_list;
 342                                 item->prev = pool->item_list->prev;
 343                                 pool->item_list->prev = item;
 344                                 pool->item_list = item;
 345                         }
 346                 }
 347                 else {
 348                         pool->item_list = item;
 349                 }
 350
 351                 allocated += item->size_in_dw;
 352         }
 353 }
 354
 355
 356 void compute_memory_free(struct compute_memory_pool* pool, int64_t id)
 357 {
 358         struct compute_memory_item *item, *next;
 359
 360         COMPUTE_DBG(pool->screen, "* compute_memory_free() id + %ld \n", id);
 361
 362         for (item = pool->item_list; item; item = next) {
 363                 next = item->next;
 364
 365                 if (item->id == id) {
 366                         if (item->prev) {
 367                                 item->prev->next = item->next;
 368                         }
 369                         else {
 370                                 pool->item_list = item->next;
 371                         }
 372
 373                         if (item->next) {
 374                                 item->next->prev = item->prev;
 375                         }
 376
 377                         free(item);
 378
 379                         return;
 380                 }
 381         }
 382
 383         fprintf(stderr, "Internal error, invalid id %"PRIi64" "
 384                 "for compute_memory_free\n", id);
 385
 386         assert(0 && "error");
 387 }
 388
 389 /**
 390  * Creates pending allocations
 391  */
 392 struct compute_memory_item* compute_memory_alloc(
 393         struct compute_memory_pool* pool,
 394         int64_t size_in_dw)
 395 {
 396         struct compute_memory_item *new_item = NULL, *last_item = NULL;
 397
 398         COMPUTE_DBG(pool->screen, "* compute_memory_alloc() size_in_dw = %ld (%ld bytes)\n",
 399                         size_in_dw, 4 * size_in_dw);
 400
 401         new_item = (struct compute_memory_item *)
 402                                 CALLOC(sizeof(struct compute_memory_item), 1);
 403         new_item->size_in_dw = size_in_dw;
 404         new_item->start_in_dw = -1; /* mark pending */
 405         new_item->id = pool->next_id++;
 406         new_item->pool = pool;
 407
 408         if (pool->item_list) {
 409                 for (last_item = pool->item_list; last_item->next;
 410                                                 last_item = last_item->next);
 411
 412                 last_item->next = new_item;
 413                 new_item->prev = last_item;
 414         }
 415         else {
 416                 pool->item_list = new_item;
 417         }
 418
 419         COMPUTE_DBG(pool->screen, "  + Adding item %p id = %u size = %u (%u bytes)\n",
 420                         new_item, new_item->id, new_item->size_in_dw,
 421                         new_item->size_in_dw * 4);
 422         return new_item;
 423 }
 424
 425 /**
 426  * Transfer data host<->device, offset and size is in bytes
 427  */
 428 void compute_memory_transfer(
 429         struct compute_memory_pool* pool,
 430         struct pipe_context * pipe,
 431         int device_to_host,
 432         struct compute_memory_item* chunk,
 433         void* data,
 434         int offset_in_chunk,
 435         int size)
 436 {
 437         int64_t aligned_size = pool->size_in_dw;
 438         struct pipe_resource* gart = (struct pipe_resource*)pool->bo;
 439         int64_t internal_offset = chunk->start_in_dw*4 + offset_in_chunk;
 440
 441         struct pipe_transfer *xfer;
 442         uint32_t *map;
 443
 444         assert(gart);
 445
 446         COMPUTE_DBG(pool->screen, "* compute_memory_transfer() device_to_host = %d, "
 447                 "offset_in_chunk = %d, size = %d\n", device_to_host,
 448                 offset_in_chunk, size);
 449
 450         if (device_to_host) {
 451                 map = pipe->transfer_map(pipe, gart, 0, PIPE_TRANSFER_READ,
 452                         &(struct pipe_box) { .width = aligned_size,
 453                         .height = 1, .depth = 1 }, &xfer);
 454                 assert(xfer);
 455                 assert(map);
 456                 memcpy(data, map + internal_offset, size);
 457                 pipe->transfer_unmap(pipe, xfer);
 458         } else {
 459                 map = pipe->transfer_map(pipe, gart, 0, PIPE_TRANSFER_WRITE,
 460                         &(struct pipe_box) { .width = aligned_size,
 461                         .height = 1, .depth = 1 }, &xfer);
 462                 assert(xfer);
 463                 assert(map);
 464                 memcpy(map + internal_offset, data, size);
 465                 pipe->transfer_unmap(pipe, xfer);
 466         }
 467 }
 468
 469 /**
 470  * Transfer data between chunk<->data, it is for VRAM<->GART transfers
 471  */
 472 void compute_memory_transfer_direct(
 473         struct compute_memory_pool* pool,
 474         int chunk_to_data,
 475         struct compute_memory_item* chunk,
 476         struct r600_resource* data,
 477         int offset_in_chunk,
 478         int offset_in_data,
 479         int size)
 480 {
 481         ///TODO: DMA
 482 }