src/gallium/drivers/svga/svga_screen_cache.c

   1 /**********************************************************
   2  * Copyright 2008-2009 VMware, Inc.  All rights reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person
   5  * obtaining a copy of this software and associated documentation
   6  * files (the "Software"), to deal in the Software without
   7  * restriction, including without limitation the rights to use, copy,
   8  * modify, merge, publish, distribute, sublicense, and/or sell copies
   9  * of the Software, and to permit persons to whom the Software is
  10  * furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be
  13  * included in all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  **********************************************************/
  25
  26 #include "util/u_math.h"
  27 #include "util/u_memory.h"
  28 #include "util/u_hash.h"
  29
  30 #include "svga_debug.h"
  31 #include "svga_format.h"
  32 #include "svga_winsys.h"
  33 #include "svga_screen.h"
  34 #include "svga_screen_cache.h"
  35
  36
  37 #define SVGA_SURFACE_CACHE_ENABLED 1
  38
  39
  40 /**
  41  * Return the size of the surface described by the key (in bytes).
  42  */
  43 static unsigned
  44 surface_size(const struct svga_host_surface_cache_key *key)
  45 {
  46    unsigned bw, bh, bpb, total_size, i;
  47
  48    assert(key->numMipLevels > 0);
  49    assert(key->numFaces > 0);
  50
  51    if (key->format == SVGA3D_BUFFER) {
  52       /* Special case: we don't want to count vertex/index buffers
  53        * against the cache size limit, so view them as zero-sized.
  54        */
  55       return 0;
  56    }
  57
  58    svga_format_size(key->format, &bw, &bh, &bpb);
  59
  60    total_size = 0;
  61
  62    for (i = 0; i < key->numMipLevels; i++) {
  63       unsigned w = u_minify(key->size.width, i);
  64       unsigned h = u_minify(key->size.height, i);
  65       unsigned d = u_minify(key->size.depth, i);
  66       unsigned img_size = ((w + bw - 1) / bw) * ((h + bh - 1) / bh) * d * bpb;
  67       total_size += img_size;
  68    }
  69
  70    total_size *= key->numFaces;
  71
  72    return total_size;
  73 }
  74
  75
  76 /**
  77  * Compute the bucket for this key.
  78  */
  79 static inline unsigned
  80 svga_screen_cache_bucket(const struct svga_host_surface_cache_key *key)
  81 {
  82    return util_hash_crc32(key, sizeof *key) % SVGA_HOST_SURFACE_CACHE_BUCKETS;
  83 }
  84
  85
  86 /**
  87  * Search the cache for a surface that matches the key.  If a match is
  88  * found, remove it from the cache and return the surface pointer.
  89  * Return NULL otherwise.
  90  */
  91 static struct svga_winsys_surface *
  92 svga_screen_cache_lookup(struct svga_screen *svgascreen,
  93                          const struct svga_host_surface_cache_key *key)
  94 {
  95    struct svga_host_surface_cache *cache = &svgascreen->cache;
  96    struct svga_winsys_screen *sws = svgascreen->sws;
  97    struct svga_host_surface_cache_entry *entry;
  98    struct svga_winsys_surface *handle = NULL;
  99    struct list_head *curr, *next;
 100    unsigned bucket;
 101    unsigned tries = 0;
 102
 103    assert(key->cachable);
 104
 105    bucket = svga_screen_cache_bucket(key);
 106
 107    pipe_mutex_lock(cache->mutex);
 108
 109    curr = cache->bucket[bucket].next;
 110    next = curr->next;
 111    while (curr != &cache->bucket[bucket]) {
 112       ++tries;
 113
 114       entry = LIST_ENTRY(struct svga_host_surface_cache_entry, curr, bucket_head);
 115
 116       assert(entry->handle);
 117
 118       /* If the key matches and the fence is signalled (the surface is no
 119        * longer needed) the lookup was successful.  We found a surface that
 120        * can be reused.
 121        * We unlink the surface from the cache entry and we add the entry to
 122        * the 'empty' list.
 123        */
 124       if (memcmp(&entry->key, key, sizeof *key) == 0 &&
 125           sws->fence_signalled(sws, entry->fence, 0) == 0) {
 126          unsigned surf_size;
 127
 128          assert(sws->surface_is_flushed(sws, entry->handle));
 129
 130          handle = entry->handle; /* Reference is transfered here. */
 131          entry->handle = NULL;
 132
 133          /* Remove from hash table */
 134          LIST_DEL(&entry->bucket_head);
 135
 136          /* remove from LRU list */
 137          LIST_DEL(&entry->head);
 138
 139          /* Add the cache entry (but not the surface!) to the empty list */
 140          LIST_ADD(&entry->head, &cache->empty);
 141
 142          /* update the cache size */
 143          surf_size = surface_size(&entry->key);
 144          assert(surf_size <= cache->total_size);
 145          if (surf_size > cache->total_size)
 146             cache->total_size = 0; /* should never happen, but be safe */
 147          else
 148             cache->total_size -= surf_size;
 149
 150          break;
 151       }
 152
 153       curr = next;
 154       next = curr->next;
 155    }
 156
 157    pipe_mutex_unlock(cache->mutex);
 158
 159    if (SVGA_DEBUG & DEBUG_DMA)
 160       debug_printf("%s: cache %s after %u tries (bucket %d)\n", __FUNCTION__,
 161                    handle ? "hit" : "miss", tries, bucket);
 162
 163    return handle;
 164 }
 165
 166
 167 /**
 168  * Free the least recently used entries in the surface cache until the
 169  * cache size is <= the target size OR there are no unused entries left
 170  * to discard.  We don't do any flushing to try to free up additional
 171  * surfaces.
 172  */
 173 static void
 174 svga_screen_cache_shrink(struct svga_screen *svgascreen,
 175                          unsigned target_size)
 176 {
 177    struct svga_host_surface_cache *cache = &svgascreen->cache;
 178    struct svga_winsys_screen *sws = svgascreen->sws;
 179    struct svga_host_surface_cache_entry *entry = NULL, *next_entry;
 180
 181    /* Walk over the list of unused buffers in reverse order: from oldest
 182     * to newest.
 183     */
 184    LIST_FOR_EACH_ENTRY_SAFE_REV(entry, next_entry, &cache->unused, head) {
 185       if (entry->key.format != SVGA3D_BUFFER) {
 186          /* we don't want to discard vertex/index buffers */
 187
 188          cache->total_size -= surface_size(&entry->key);
 189
 190          assert(entry->handle);
 191          sws->surface_reference(sws, &entry->handle, NULL);
 192
 193          LIST_DEL(&entry->bucket_head);
 194          LIST_DEL(&entry->head);
 195          LIST_ADD(&entry->head, &cache->empty);
 196
 197          if (cache->total_size <= target_size) {
 198             /* all done */
 199             break;
 200          }
 201       }
 202    }
 203 }
 204
 205
 206 /**
 207  * Add a surface to the cache.  This is done when the driver deletes
 208  * the surface.  Note: transfers a handle reference.
 209  */
 210 static void
 211 svga_screen_cache_add(struct svga_screen *svgascreen,
 212                       const struct svga_host_surface_cache_key *key,
 213                       struct svga_winsys_surface **p_handle)
 214 {
 215    struct svga_host_surface_cache *cache = &svgascreen->cache;
 216    struct svga_winsys_screen *sws = svgascreen->sws;
 217    struct svga_host_surface_cache_entry *entry = NULL;
 218    struct svga_winsys_surface *handle = *p_handle;
 219    unsigned surf_size;
 220
 221    assert(key->cachable);
 222
 223    if (!handle)
 224       return;
 225
 226    surf_size = surface_size(key);
 227
 228    *p_handle = NULL;
 229    pipe_mutex_lock(cache->mutex);
 230
 231    if (surf_size >= SVGA_HOST_SURFACE_CACHE_BYTES) {
 232       /* this surface is too large to cache, just free it */
 233       sws->surface_reference(sws, &handle, NULL);
 234       pipe_mutex_unlock(cache->mutex);
 235       return;
 236    }
 237
 238    if (cache->total_size + surf_size > SVGA_HOST_SURFACE_CACHE_BYTES) {
 239       /* Adding this surface would exceed the cache size.
 240        * Try to discard least recently used entries until we hit the
 241        * new target cache size.
 242        */
 243       unsigned target_size = SVGA_HOST_SURFACE_CACHE_BYTES - surf_size;
 244
 245       svga_screen_cache_shrink(svgascreen, target_size);
 246
 247       if (cache->total_size > target_size) {
 248          /* we weren't able to shrink the cache as much as we wanted so
 249           * just discard this surface.
 250           */
 251          sws->surface_reference(sws, &handle, NULL);
 252          pipe_mutex_unlock(cache->mutex);
 253          return;
 254       }
 255    }
 256
 257    if (!LIST_IS_EMPTY(&cache->empty)) {
 258       /* An empty entry has no surface associated with it.
 259        * Use the first empty entry.
 260        */
 261       entry = LIST_ENTRY(struct svga_host_surface_cache_entry,
 262                          cache->empty.next, head);
 263
 264       /* Remove from LRU list */
 265       LIST_DEL(&entry->head);
 266    }
 267    else if (!LIST_IS_EMPTY(&cache->unused)) {
 268       /* free the last used buffer and reuse its entry */
 269       entry = LIST_ENTRY(struct svga_host_surface_cache_entry,
 270                          cache->unused.prev, head);
 271       SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 272                "unref sid %p (make space)\n", entry->handle);
 273
 274       cache->total_size -= surface_size(&entry->key);
 275
 276       sws->surface_reference(sws, &entry->handle, NULL);
 277
 278       /* Remove from hash table */
 279       LIST_DEL(&entry->bucket_head);
 280
 281       /* Remove from LRU list */
 282       LIST_DEL(&entry->head);
 283    }
 284
 285    if (entry) {
 286       assert(entry->handle == NULL);
 287       entry->handle = handle;
 288       memcpy(&entry->key, key, sizeof entry->key);
 289
 290       SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 291                "cache sid %p\n", entry->handle);
 292       LIST_ADD(&entry->head, &cache->validated);
 293
 294       cache->total_size += surf_size;
 295    }
 296    else {
 297       /* Couldn't cache the buffer -- this really shouldn't happen */
 298       SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 299                "unref sid %p (couldn't find space)\n", handle);
 300       sws->surface_reference(sws, &handle, NULL);
 301    }
 302
 303    pipe_mutex_unlock(cache->mutex);
 304 }
 305
 306
 307 /**
 308  * Called during the screen flush to move all buffers not in a validate list
 309  * into the unused list.
 310  */
 311 void
 312 svga_screen_cache_flush(struct svga_screen *svgascreen,
 313                         struct pipe_fence_handle *fence)
 314 {
 315    struct svga_host_surface_cache *cache = &svgascreen->cache;
 316    struct svga_winsys_screen *sws = svgascreen->sws;
 317    struct svga_host_surface_cache_entry *entry;
 318    struct list_head *curr, *next;
 319    unsigned bucket;
 320
 321    pipe_mutex_lock(cache->mutex);
 322
 323    /* Loop over entries in the validated list */
 324    curr = cache->validated.next;
 325    next = curr->next;
 326    while (curr != &cache->validated) {
 327       entry = LIST_ENTRY(struct svga_host_surface_cache_entry, curr, head);
 328
 329       assert(entry->handle);
 330
 331       if (sws->surface_is_flushed(sws, entry->handle)) {
 332          /* remove entry from LRU list */
 333          LIST_DEL(&entry->head);
 334
 335          svgascreen->sws->fence_reference(svgascreen->sws, &entry->fence, fence);
 336
 337          /* Add entry to the unused list */
 338          LIST_ADD(&entry->head, &cache->unused);
 339
 340          /* Add entry to the hash table bucket */
 341          bucket = svga_screen_cache_bucket(&entry->key);
 342          LIST_ADD(&entry->bucket_head, &cache->bucket[bucket]);
 343       }
 344
 345       curr = next;
 346       next = curr->next;
 347    }
 348
 349    pipe_mutex_unlock(cache->mutex);
 350 }
 351
 352
 353 /**
 354  * Free all the surfaces in the cache.
 355  * Called when destroying the svga screen object.
 356  */
 357 void
 358 svga_screen_cache_cleanup(struct svga_screen *svgascreen)
 359 {
 360    struct svga_host_surface_cache *cache = &svgascreen->cache;
 361    struct svga_winsys_screen *sws = svgascreen->sws;
 362    unsigned i;
 363
 364    for (i = 0; i < SVGA_HOST_SURFACE_CACHE_SIZE; ++i) {
 365       if (cache->entries[i].handle) {
 366          SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 367                   "unref sid %p (shutdown)\n", cache->entries[i].handle);
 368          sws->surface_reference(sws, &cache->entries[i].handle, NULL);
 369
 370          cache->total_size -= surface_size(&cache->entries[i].key);
 371       }
 372
 373       if (cache->entries[i].fence)
 374          svgascreen->sws->fence_reference(svgascreen->sws,
 375                                           &cache->entries[i].fence, NULL);
 376    }
 377
 378    pipe_mutex_destroy(cache->mutex);
 379 }
 380
 381
 382 enum pipe_error
 383 svga_screen_cache_init(struct svga_screen *svgascreen)
 384 {
 385    struct svga_host_surface_cache *cache = &svgascreen->cache;
 386    unsigned i;
 387
 388    assert(cache->total_size == 0);
 389
 390    pipe_mutex_init(cache->mutex);
 391
 392    for (i = 0; i < SVGA_HOST_SURFACE_CACHE_BUCKETS; ++i)
 393       LIST_INITHEAD(&cache->bucket[i]);
 394
 395    LIST_INITHEAD(&cache->unused);
 396
 397    LIST_INITHEAD(&cache->validated);
 398
 399    LIST_INITHEAD(&cache->empty);
 400    for (i = 0; i < SVGA_HOST_SURFACE_CACHE_SIZE; ++i)
 401       LIST_ADDTAIL(&cache->entries[i].head, &cache->empty);
 402
 403    return PIPE_OK;
 404 }
 405
 406
 407 /**
 408  * Allocate a new host-side surface.  If the surface is marked as cachable,
 409  * first try re-using a surface in the cache of freed surfaces.  Otherwise,
 410  * allocate a new surface.
 411  * \param bind_flags  bitmask of PIPE_BIND_x flags
 412  * \param usage  one of PIPE_USAGE_x values
 413  */
 414 struct svga_winsys_surface *
 415 svga_screen_surface_create(struct svga_screen *svgascreen,
 416                            unsigned bind_flags, unsigned usage,
 417                            struct svga_host_surface_cache_key *key)
 418 {
 419    struct svga_winsys_screen *sws = svgascreen->sws;
 420    struct svga_winsys_surface *handle = NULL;
 421    boolean cachable = SVGA_SURFACE_CACHE_ENABLED && key->cachable;
 422
 423    SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 424             "%s sz %dx%dx%d mips %d faces %d arraySize %d cachable %d\n",
 425             __FUNCTION__,
 426             key->size.width,
 427             key->size.height,
 428             key->size.depth,
 429             key->numMipLevels,
 430             key->numFaces,
 431             key->arraySize,
 432             key->cachable);
 433
 434    if (cachable) {
 435       if (key->format == SVGA3D_BUFFER) {
 436          SVGA3dSurfaceFlags hint_flag;
 437
 438          /* For buffers, round the buffer size up to the nearest power
 439           * of two to increase the probability of cache hits.  Keep
 440           * texture surface dimensions unchanged.
 441           */
 442          uint32_t size = 1;
 443          while (size < key->size.width)
 444             size <<= 1;
 445          key->size.width = size;
 446
 447          /* Determine whether the buffer is static or dynamic.
 448           * This is a bit of a heuristic which can be tuned as needed.
 449           */
 450          if (usage == PIPE_USAGE_DEFAULT ||
 451              usage == PIPE_USAGE_IMMUTABLE) {
 452             hint_flag = SVGA3D_SURFACE_HINT_STATIC;
 453          }
 454          else if (bind_flags & PIPE_BIND_INDEX_BUFFER) {
 455             /* Index buffers don't change too often.  Mark them as static.
 456              */
 457             hint_flag = SVGA3D_SURFACE_HINT_STATIC;
 458          }
 459          else {
 460             /* Since we're reusing buffers we're effectively transforming all
 461              * of them into dynamic buffers.
 462              *
 463              * It would be nice to not cache long lived static buffers. But there
 464              * is no way to detect the long lived from short lived ones yet. A
 465              * good heuristic would be buffer size.
 466              */
 467             hint_flag = SVGA3D_SURFACE_HINT_DYNAMIC;
 468          }
 469
 470          key->flags &= ~(SVGA3D_SURFACE_HINT_STATIC |
 471                          SVGA3D_SURFACE_HINT_DYNAMIC);
 472          key->flags |= hint_flag;
 473       }
 474
 475       handle = svga_screen_cache_lookup(svgascreen, key);
 476       if (handle) {
 477          if (key->format == SVGA3D_BUFFER)
 478             SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 479                      "reuse sid %p sz %d (buffer)\n", handle,
 480                      key->size.width);
 481          else
 482             SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 483                      "reuse sid %p sz %dx%dx%d mips %d faces %d arraySize %d\n", handle,
 484                      key->size.width,
 485                      key->size.height,
 486                      key->size.depth,
 487                      key->numMipLevels,
 488                      key->numFaces,
 489                      key->arraySize);
 490       }
 491    }
 492
 493    if (!handle) {
 494       unsigned usage = 0;
 495
 496       if (!key->cachable)
 497          usage |= SVGA_SURFACE_USAGE_SHARED;
 498       if (key->scanout)
 499          usage |= SVGA_SURFACE_USAGE_SCANOUT;
 500
 501       handle = sws->surface_create(sws,
 502                                    key->flags,
 503                                    key->format,
 504                                    usage,
 505                                    key->size,
 506                                    key->numFaces * key->arraySize,
 507                                    key->numMipLevels,
 508                                    key->sampleCount);
 509       if (handle)
 510          SVGA_DBG(DEBUG_CACHE|DEBUG_DMA,
 511                   "  CREATE sid %p sz %dx%dx%d\n",
 512                   handle,
 513                   key->size.width,
 514                   key->size.height,
 515                   key->size.depth);
 516    }
 517
 518    return handle;
 519 }
 520
 521
 522 /**
 523  * Release a surface.  We don't actually free the surface- we put
 524  * it into the cache of freed surfaces (if it's cachable).
 525  */
 526 void
 527 svga_screen_surface_destroy(struct svga_screen *svgascreen,
 528                             const struct svga_host_surface_cache_key *key,
 529                             struct svga_winsys_surface **p_handle)
 530 {
 531    struct svga_winsys_screen *sws = svgascreen->sws;
 532
 533    /* We only set the cachable flag for surfaces of which we are the
 534     * exclusive owner.  So just hold onto our existing reference in
 535     * that case.
 536     */
 537    if (SVGA_SURFACE_CACHE_ENABLED && key->cachable) {
 538       svga_screen_cache_add(svgascreen, key, p_handle);
 539    }
 540    else {
 541       SVGA_DBG(DEBUG_DMA,
 542                "unref sid %p (uncachable)\n", *p_handle);
 543       sws->surface_reference(sws, p_handle, NULL);
 544    }
 545 }
 546
 547
 548 /**
 549  * Print/dump the contents of the screen cache.  For debugging.
 550  */
 551 void
 552 svga_screen_cache_dump(const struct svga_screen *svgascreen)
 553 {
 554    const struct svga_host_surface_cache *cache = &svgascreen->cache;
 555    unsigned bucket;
 556    unsigned count = 0;
 557
 558    debug_printf("svga3d surface cache:\n");
 559    for (bucket = 0; bucket < SVGA_HOST_SURFACE_CACHE_BUCKETS; bucket++) {
 560       struct list_head *curr;
 561       curr = cache->bucket[bucket].next;
 562       while (curr && curr != &cache->bucket[bucket]) {
 563          struct svga_host_surface_cache_entry *entry =
 564             LIST_ENTRY(struct svga_host_surface_cache_entry,
 565                        curr, bucket_head);
 566          if (entry->key.format != 37) {
 567             debug_printf("  %u x %u x %u format %u\n",
 568                          entry->key.size.width,
 569                          entry->key.size.height,
 570                          entry->key.size.depth,
 571                          entry->key.format);
 572          }
 573          curr = curr->next;
 574          count++;
 575       }
 576    }
 577
 578    debug_printf("%u surfaces, %u bytes\n", count, cache->total_size);
 579 }