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.h"
  37 #include "r600_resource.h"
  38 #include "r600_shader.h"
  39 #include "r600_pipe.h"
  40 #include "r600_formats.h"
  41 #include "compute_memory_pool.h"
  42 #include "evergreen_compute_internal.h"
  43 #include <inttypes.h>
  44
  45 /**
  46  * Creates a new pool
  47  */
  48 struct compute_memory_pool* compute_memory_pool_new(
  49         struct r600_screen * rscreen)
  50 {
  51         struct compute_memory_pool* pool = (struct compute_memory_pool*)
  52                                 CALLOC(sizeof(struct compute_memory_pool), 1);
  53
  54         COMPUTE_DBG("* compute_memory_pool_new()\n");
  55
  56         pool->screen = rscreen;
  57         return pool;
  58 }
  59
  60 static void compute_memory_pool_init(struct compute_memory_pool * pool,
  61         unsigned initial_size_in_dw)
  62 {
  63
  64         COMPUTE_DBG("* compute_memory_pool_init() initial_size_in_dw = %ld\n",
  65                 initial_size_in_dw);
  66
  67         pool->shadow = (uint32_t*)CALLOC(initial_size_in_dw, 4);
  68         pool->next_id = 1;
  69         pool->size_in_dw = initial_size_in_dw;
  70         pool->bo = (struct r600_resource*)r600_compute_buffer_alloc_vram(pool->screen,
  71                                                         pool->size_in_dw * 4);
  72 }
  73
  74 /**
  75  * Frees all stuff in the pool and the pool struct itself too
  76  */
  77 void compute_memory_pool_delete(struct compute_memory_pool* pool)
  78 {
  79         COMPUTE_DBG("* compute_memory_pool_delete()\n");
  80         free(pool->shadow);
  81         if (pool->bo) {
  82                 pool->screen->screen.resource_destroy((struct pipe_screen *)
  83                         pool->screen, (struct pipe_resource *)pool->bo);
  84         }
  85         free(pool);
  86 }
  87
  88 /**
  89  * Searches for an empty space in the pool, return with the pointer to the
  90  * allocatable space in the pool, returns -1 on failure.
  91  */
  92 int64_t compute_memory_prealloc_chunk(
  93         struct compute_memory_pool* pool,
  94         int64_t size_in_dw)
  95 {
  96         struct compute_memory_item *item;
  97
  98         int last_end = 0;
  99
 100         assert(size_in_dw <= pool->size_in_dw);
 101
 102         COMPUTE_DBG("* compute_memory_prealloc_chunk() size_in_dw = %ld\n",
 103                 size_in_dw);
 104
 105         for (item = pool->item_list; item; item = item->next) {
 106                 if (item->start_in_dw > -1) {
 107                         if (item->start_in_dw-last_end > size_in_dw) {
 108                                 return last_end;
 109                         }
 110
 111                         last_end = item->start_in_dw + item->size_in_dw;
 112                         last_end += (1024 - last_end % 1024);
 113                 }
 114         }
 115
 116         if (pool->size_in_dw - last_end < size_in_dw) {
 117                 return -1;
 118         }
 119
 120         return last_end;
 121 }
 122
 123 /**
 124  *  Search for the chunk where we can link our new chunk after it.
 125  */
 126 struct compute_memory_item* compute_memory_postalloc_chunk(
 127         struct compute_memory_pool* pool,
 128         int64_t start_in_dw)
 129 {
 130         struct compute_memory_item* item;
 131
 132         COMPUTE_DBG("* compute_memory_postalloc_chunck() start_in_dw = %ld\n",
 133                 start_in_dw);
 134
 135         /* Check if we can insert it in the front of the list */
 136         if (pool->item_list && pool->item_list->start_in_dw > start_in_dw) {
 137                 return NULL;
 138         }
 139
 140         for (item = pool->item_list; item; item = item->next) {
 141                 if (item->next) {
 142                         if (item->start_in_dw < start_in_dw
 143                                 && item->next->start_in_dw > start_in_dw) {
 144                                 return item;
 145                         }
 146                 }
 147                 else {
 148                         /* end of chain */
 149                         assert(item->start_in_dw < start_in_dw);
 150                         return item;
 151                 }
 152         }
 153
 154         assert(0 && "unreachable");
 155         return NULL;
 156 }
 157
 158 /**
 159  * Reallocates pool, conserves data
 160  */
 161 void compute_memory_grow_pool(struct compute_memory_pool* pool,
 162         struct pipe_context * pipe, int new_size_in_dw)
 163 {
 164         COMPUTE_DBG("* compute_memory_grow_pool() new_size_in_dw = %d\n",
 165                 new_size_in_dw);
 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("  Aligned size = %d\n", new_size_in_dw);
 175
 176                 compute_memory_shadow(pool, pipe, 1);
 177                 pool->shadow = realloc(pool->shadow, new_size_in_dw*4);
 178                 pool->size_in_dw = new_size_in_dw;
 179                 pool->screen->screen.resource_destroy(
 180                         (struct pipe_screen *)pool->screen,
 181                         (struct pipe_resource *)pool->bo);
 182                 pool->bo = (struct r600_resource*)r600_compute_buffer_alloc_vram(
 183                                                         pool->screen,
 184                                                         pool->size_in_dw * 4);
 185                 compute_memory_shadow(pool, pipe, 0);
 186         }
 187 }
 188
 189 /**
 190  * Copy pool from device to host, or host to device.
 191  */
 192 void compute_memory_shadow(struct compute_memory_pool* pool,
 193         struct pipe_context * pipe, int device_to_host)
 194 {
 195         struct compute_memory_item chunk;
 196
 197         COMPUTE_DBG("* compute_memory_shadow() device_to_host = %d\n",
 198                 device_to_host);
 199
 200         chunk.id = 0;
 201         chunk.start_in_dw = 0;
 202         chunk.size_in_dw = pool->size_in_dw;
 203         chunk.prev = chunk.next = NULL;
 204         compute_memory_transfer(pool, pipe, device_to_host, &chunk,
 205                                 pool->shadow, 0, pool->size_in_dw*4);
 206 }
 207
 208 /**
 209  * Allocates pending allocations in the pool
 210  */
 211 void compute_memory_finalize_pending(struct compute_memory_pool* pool,
 212         struct pipe_context * pipe)
 213 {
 214         struct compute_memory_item *pending_list = NULL, *end_p = NULL;
 215         struct compute_memory_item *item, *next;
 216
 217         int64_t allocated = 0;
 218         int64_t unallocated = 0;
 219
 220         int64_t start_in_dw = 0;
 221
 222         COMPUTE_DBG("* compute_memory_finalize_pending()\n");
 223
 224         for (item = pool->item_list; item; item = item->next) {
 225                 COMPUTE_DBG("  + list: offset = %i id = %i size = %i "
 226                         "(%i bytes)\n",item->start_in_dw, item->id,
 227                         item->size_in_dw, item->size_in_dw * 4);
 228         }
 229
 230         /* Search through the list of memory items in the pool */
 231         for (item = pool->item_list; item; item = next) {
 232                 next = item->next;
 233
 234                 /* Check if the item is pending. */
 235                 if (item->start_in_dw == -1) {
 236                         /* It is pending, so add it to the pending_list... */
 237                         if (end_p) {
 238                                 end_p->next = item;
 239                         }
 240                         else {
 241                                 pending_list = item;
 242                         }
 243
 244                         /* ... and then remove it from the item list. */
 245                         if (item->prev) {
 246                                 item->prev->next = next;
 247                         }
 248                         else {
 249                                 pool->item_list = next;
 250                         }
 251
 252                         if (next) {
 253                                 next->prev = item->prev;
 254                         }
 255
 256                         /* This sequence makes the item be at the end of the list */
 257                         item->prev = end_p;
 258                         item->next = NULL;
 259                         end_p = item;
 260
 261                         /* Update the amount of space we will need to allocate. */
 262                         unallocated += item->size_in_dw+1024;
 263                 }
 264                 else {
 265                         /* The item is not pendng, so update the amount of space
 266                          * that has already been allocated. */
 267                         allocated += item->size_in_dw;
 268                 }
 269         }
 270
 271         /* If we require more space than the size of the pool, then grow the
 272          * pool.
 273          *
 274          * XXX: I'm pretty sure this won't work.  Imagine this scenario:
 275          *
 276          * Offset Item Size
 277          *   0    A    50
 278          * 200    B    50
 279          * 400    C    50
 280          *
 281          * Total size = 450
 282          * Allocated size = 150
 283          * Pending Item D Size = 200
 284          *
 285          * In this case, there are 300 units of free space in the pool, but
 286          * they aren't contiguous, so it will be impossible to allocate Item D.
 287          */
 288         if (pool->size_in_dw < allocated+unallocated) {
 289                 compute_memory_grow_pool(pool, pipe, allocated+unallocated);
 290         }
 291
 292         /* Loop through all the pending items, allocate space for them and
 293          * add them back to the item_list. */
 294         for (item = pending_list; item; item = next) {
 295                 next = item->next;
 296
 297                 /* Search for free space in the pool for this item. */
 298                 while ((start_in_dw=compute_memory_prealloc_chunk(pool,
 299                                                 item->size_in_dw)) == -1) {
 300                         int64_t need = item->size_in_dw+2048 -
 301                                                 (pool->size_in_dw - allocated);
 302
 303                         need += 1024 - (need % 1024);
 304
 305                         if (need > 0) {
 306                                 compute_memory_grow_pool(pool,
 307                                                 pipe,
 308                                                 pool->size_in_dw + need);
 309                         }
 310                         else {
 311                                 need = pool->size_in_dw / 10;
 312                                 need += 1024 - (need % 1024);
 313                                 compute_memory_grow_pool(pool,
 314                                                 pipe,
 315                                                 pool->size_in_dw + need);
 316                         }
 317                 }
 318                 COMPUTE_DBG("  + Found space for Item %p id = %u "
 319                         "start_in_dw = %u (%u bytes) size_in_dw = %u (%u bytes)\n",
 320                         item, item->id, start_in_dw, start_in_dw * 4,
 321                         item->size_in_dw, item->size_in_dw * 4);
 322
 323                 item->start_in_dw = start_in_dw;
 324                 item->next = NULL;
 325                 item->prev = NULL;
 326
 327                 if (pool->item_list) {
 328                         struct compute_memory_item *pos;
 329
 330                         pos = compute_memory_postalloc_chunk(pool, start_in_dw);
 331                         if (pos) {
 332                                 item->prev = pos;
 333                                 item->next = pos->next;
 334                                 pos->next = item;
 335                                 if (item->next) {
 336                                         item->next->prev = item;
 337                                 }
 338                         } else {
 339                                 /* Add item to the front of the list */
 340                                 item->next = pool->item_list->next;
 341                                 if (pool->item_list->next) {
 342                                         pool->item_list->next->prev = item;
 343                                 }
 344                                 item->prev = pool->item_list->prev;
 345                                 if (pool->item_list->prev) {
 346                                         pool->item_list->prev->next = item;
 347                                 }
 348                                 pool->item_list = item;
 349                         }
 350                 }
 351                 else {
 352                         pool->item_list = item;
 353                 }
 354
 355                 allocated += item->size_in_dw;
 356         }
 357 }
 358
 359
 360 void compute_memory_free(struct compute_memory_pool* pool, int64_t id)
 361 {
 362         struct compute_memory_item *item, *next;
 363
 364         COMPUTE_DBG("* compute_memory_free() id + %ld \n", id);
 365
 366         for (item = pool->item_list; item; item = next) {
 367                 next = item->next;
 368
 369                 if (item->id == id) {
 370                         if (item->prev) {
 371                                 item->prev->next = item->next;
 372                         }
 373                         else {
 374                                 pool->item_list = item->next;
 375                         }
 376
 377                         if (item->next) {
 378                                 item->next->prev = item->prev;
 379                         }
 380
 381                         free(item);
 382
 383                         return;
 384                 }
 385         }
 386
 387         fprintf(stderr, "Internal error, invalid id %"PRIi64" "
 388                 "for compute_memory_free\n", id);
 389
 390         assert(0 && "error");
 391 }
 392
 393 /**
 394  * Creates pending allocations
 395  */
 396 struct compute_memory_item* compute_memory_alloc(
 397         struct compute_memory_pool* pool,
 398         int64_t size_in_dw)
 399 {
 400         struct compute_memory_item *new_item = NULL, *last_item = NULL;
 401
 402         COMPUTE_DBG("* compute_memory_alloc() size_in_dw = %ld (%ld bytes)\n",
 403                         size_in_dw, 4 * size_in_dw);
 404
 405         new_item = (struct compute_memory_item *)
 406                                 CALLOC(sizeof(struct compute_memory_item), 1);
 407         new_item->size_in_dw = size_in_dw;
 408         new_item->start_in_dw = -1; /* mark pending */
 409         new_item->id = pool->next_id++;
 410         new_item->pool = pool;
 411
 412         if (pool->item_list) {
 413                 for (last_item = pool->item_list; last_item->next;
 414                                                 last_item = last_item->next);
 415
 416                 last_item->next = new_item;
 417                 new_item->prev = last_item;
 418         }
 419         else {
 420                 pool->item_list = new_item;
 421         }
 422
 423         COMPUTE_DBG("  + Adding item %p id = %u size = %u (%u bytes)\n",
 424                         new_item, new_item->id, new_item->size_in_dw,
 425                         new_item->size_in_dw * 4);
 426         return new_item;
 427 }
 428
 429 /**
 430  * Transfer data host<->device, offset and size is in bytes
 431  */
 432 void compute_memory_transfer(
 433         struct compute_memory_pool* pool,
 434         struct pipe_context * pipe,
 435         int device_to_host,
 436         struct compute_memory_item* chunk,
 437         void* data,
 438         int offset_in_chunk,
 439         int size)
 440 {
 441         int64_t aligned_size = pool->size_in_dw;
 442         struct pipe_resource* gart = (struct pipe_resource*)pool->bo;
 443         int64_t internal_offset = chunk->start_in_dw*4 + offset_in_chunk;
 444
 445         struct pipe_transfer *xfer;
 446         uint32_t *map;
 447
 448         assert(gart);
 449
 450         COMPUTE_DBG("* compute_memory_transfer() device_to_host = %d, "
 451                 "offset_in_chunk = %d, size = %d\n", device_to_host,
 452                 offset_in_chunk, size);
 453
 454         if (device_to_host) {
 455                 map = pipe->transfer_map(pipe, gart, 0, PIPE_TRANSFER_READ,
 456                         &(struct pipe_box) { .width = aligned_size,
 457                         .height = 1, .depth = 1 }, &xfer);
 458                 assert(xfer);
 459                 assert(map);
 460                 memcpy(data, map + internal_offset, size);
 461                 pipe->transfer_unmap(pipe, xfer);
 462         } else {
 463                 map = pipe->transfer_map(pipe, gart, 0, PIPE_TRANSFER_WRITE,
 464                         &(struct pipe_box) { .width = aligned_size,
 465                         .height = 1, .depth = 1 }, &xfer);
 466                 assert(xfer);
 467                 assert(map);
 468                 memcpy(map + internal_offset, data, size);
 469                 pipe->transfer_unmap(pipe, xfer);
 470         }
 471 }
 472
 473 /**
 474  * Transfer data between chunk<->data, it is for VRAM<->GART transfers
 475  */
 476 void compute_memory_transfer_direct(
 477         struct compute_memory_pool* pool,
 478         int chunk_to_data,
 479         struct compute_memory_item* chunk,
 480         struct r600_resource* data,
 481         int offset_in_chunk,
 482         int offset_in_data,
 483         int size)
 484 {
 485         ///TODO: DMA
 486 }