src/intel/vulkan/anv_pipeline_cache.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "util/blob.h"
  25 #include "util/hash_table.h"
  26 #include "util/debug.h"
  27 #include "util/disk_cache.h"
  28 #include "util/mesa-sha1.h"
  29 #include "nir/nir_serialize.h"
  30 #include "anv_private.h"
  31 #include "nir/nir_xfb_info.h"
  32
  33 struct anv_shader_bin *
  34 anv_shader_bin_create(struct anv_device *device,
  35                       const void *key_data, uint32_t key_size,
  36                       const void *kernel_data, uint32_t kernel_size,
  37                       const void *constant_data, uint32_t constant_data_size,
  38                       const struct brw_stage_prog_data *prog_data_in,
  39                       uint32_t prog_data_size, const void *prog_data_param_in,
  40                       const struct brw_compile_stats *stats, uint32_t num_stats,
  41                       const nir_xfb_info *xfb_info_in,
  42                       const struct anv_pipeline_bind_map *bind_map)
  43 {
  44    struct anv_shader_bin *shader;
  45    struct anv_shader_bin_key *key;
  46    struct brw_stage_prog_data *prog_data;
  47    uint32_t *prog_data_param;
  48    nir_xfb_info *xfb_info;
  49    struct anv_pipeline_binding *surface_to_descriptor, *sampler_to_descriptor;
  50
  51    ANV_MULTIALLOC(ma);
  52    anv_multialloc_add(&ma, &shader, 1);
  53    anv_multialloc_add_size(&ma, &key, sizeof(*key) + key_size);
  54    anv_multialloc_add_size(&ma, &prog_data, prog_data_size);
  55    anv_multialloc_add(&ma, &prog_data_param, prog_data_in->nr_params);
  56    if (xfb_info_in) {
  57       uint32_t xfb_info_size = nir_xfb_info_size(xfb_info_in->output_count);
  58       anv_multialloc_add_size(&ma, &xfb_info, xfb_info_size);
  59    }
  60    anv_multialloc_add(&ma, &surface_to_descriptor,
  61                            bind_map->surface_count);
  62    anv_multialloc_add(&ma, &sampler_to_descriptor,
  63                            bind_map->sampler_count);
  64
  65    if (!anv_multialloc_alloc(&ma, &device->alloc,
  66                              VK_SYSTEM_ALLOCATION_SCOPE_DEVICE))
  67       return NULL;
  68
  69    shader->ref_cnt = 1;
  70
  71    key->size = key_size;
  72    memcpy(key->data, key_data, key_size);
  73    shader->key = key;
  74
  75    shader->kernel =
  76       anv_state_pool_alloc(&device->instruction_state_pool, kernel_size, 64);
  77    memcpy(shader->kernel.map, kernel_data, kernel_size);
  78    shader->kernel_size = kernel_size;
  79
  80    if (constant_data_size) {
  81       shader->constant_data =
  82          anv_state_pool_alloc(&device->dynamic_state_pool,
  83                               constant_data_size, 32);
  84       memcpy(shader->constant_data.map, constant_data, constant_data_size);
  85    } else {
  86       shader->constant_data = ANV_STATE_NULL;
  87    }
  88    shader->constant_data_size = constant_data_size;
  89
  90    memcpy(prog_data, prog_data_in, prog_data_size);
  91    memcpy(prog_data_param, prog_data_param_in,
  92           prog_data->nr_params * sizeof(*prog_data_param));
  93    prog_data->param = prog_data_param;
  94    shader->prog_data = prog_data;
  95    shader->prog_data_size = prog_data_size;
  96
  97    assert(num_stats <= ARRAY_SIZE(shader->stats));
  98    typed_memcpy(shader->stats, stats, num_stats);
  99    shader->num_stats = num_stats;
 100
 101    if (xfb_info_in) {
 102       *xfb_info = *xfb_info_in;
 103       typed_memcpy(xfb_info->outputs, xfb_info_in->outputs,
 104                    xfb_info_in->output_count);
 105       shader->xfb_info = xfb_info;
 106    } else {
 107       shader->xfb_info = NULL;
 108    }
 109
 110    shader->bind_map = *bind_map;
 111    typed_memcpy(surface_to_descriptor, bind_map->surface_to_descriptor,
 112                 bind_map->surface_count);
 113    shader->bind_map.surface_to_descriptor = surface_to_descriptor;
 114    typed_memcpy(sampler_to_descriptor, bind_map->sampler_to_descriptor,
 115                 bind_map->sampler_count);
 116    shader->bind_map.sampler_to_descriptor = sampler_to_descriptor;
 117
 118    return shader;
 119 }
 120
 121 void
 122 anv_shader_bin_destroy(struct anv_device *device,
 123                        struct anv_shader_bin *shader)
 124 {
 125    assert(shader->ref_cnt == 0);
 126    anv_state_pool_free(&device->instruction_state_pool, shader->kernel);
 127    anv_state_pool_free(&device->dynamic_state_pool, shader->constant_data);
 128    vk_free(&device->alloc, shader);
 129 }
 130
 131 static bool
 132 anv_shader_bin_write_to_blob(const struct anv_shader_bin *shader,
 133                              struct blob *blob)
 134 {
 135    blob_write_uint32(blob, shader->key->size);
 136    blob_write_bytes(blob, shader->key->data, shader->key->size);
 137
 138    blob_write_uint32(blob, shader->kernel_size);
 139    blob_write_bytes(blob, shader->kernel.map, shader->kernel_size);
 140
 141    blob_write_uint32(blob, shader->constant_data_size);
 142    blob_write_bytes(blob, shader->constant_data.map,
 143                     shader->constant_data_size);
 144
 145    blob_write_uint32(blob, shader->prog_data_size);
 146    blob_write_bytes(blob, shader->prog_data, shader->prog_data_size);
 147    blob_write_bytes(blob, shader->prog_data->param,
 148                     shader->prog_data->nr_params *
 149                     sizeof(*shader->prog_data->param));
 150
 151    blob_write_uint32(blob, shader->num_stats);
 152    blob_write_bytes(blob, shader->stats,
 153                     shader->num_stats * sizeof(shader->stats[0]));
 154
 155    if (shader->xfb_info) {
 156       uint32_t xfb_info_size =
 157          nir_xfb_info_size(shader->xfb_info->output_count);
 158       blob_write_uint32(blob, xfb_info_size);
 159       blob_write_bytes(blob, shader->xfb_info, xfb_info_size);
 160    } else {
 161       blob_write_uint32(blob, 0);
 162    }
 163
 164    blob_write_uint32(blob, shader->bind_map.surface_count);
 165    blob_write_uint32(blob, shader->bind_map.sampler_count);
 166    blob_write_bytes(blob, shader->bind_map.surface_to_descriptor,
 167                     shader->bind_map.surface_count *
 168                     sizeof(*shader->bind_map.surface_to_descriptor));
 169    blob_write_bytes(blob, shader->bind_map.sampler_to_descriptor,
 170                     shader->bind_map.sampler_count *
 171                     sizeof(*shader->bind_map.sampler_to_descriptor));
 172    blob_write_bytes(blob, shader->bind_map.push_ranges,
 173                     sizeof(shader->bind_map.push_ranges));
 174
 175    return !blob->out_of_memory;
 176 }
 177
 178 static struct anv_shader_bin *
 179 anv_shader_bin_create_from_blob(struct anv_device *device,
 180                                 struct blob_reader *blob)
 181 {
 182    uint32_t key_size = blob_read_uint32(blob);
 183    const void *key_data = blob_read_bytes(blob, key_size);
 184
 185    uint32_t kernel_size = blob_read_uint32(blob);
 186    const void *kernel_data = blob_read_bytes(blob, kernel_size);
 187
 188    uint32_t constant_data_size = blob_read_uint32(blob);
 189    const void *constant_data = blob_read_bytes(blob, constant_data_size);
 190
 191    uint32_t prog_data_size = blob_read_uint32(blob);
 192    const struct brw_stage_prog_data *prog_data =
 193       blob_read_bytes(blob, prog_data_size);
 194    if (blob->overrun)
 195       return NULL;
 196    const void *prog_data_param =
 197       blob_read_bytes(blob, prog_data->nr_params * sizeof(*prog_data->param));
 198
 199    uint32_t num_stats = blob_read_uint32(blob);
 200    const struct brw_compile_stats *stats =
 201       blob_read_bytes(blob, num_stats * sizeof(stats[0]));
 202
 203    const nir_xfb_info *xfb_info = NULL;
 204    uint32_t xfb_size = blob_read_uint32(blob);
 205    if (xfb_size)
 206       xfb_info = blob_read_bytes(blob, xfb_size);
 207
 208    struct anv_pipeline_bind_map bind_map;
 209    bind_map.surface_count = blob_read_uint32(blob);
 210    bind_map.sampler_count = blob_read_uint32(blob);
 211    bind_map.surface_to_descriptor = (void *)
 212       blob_read_bytes(blob, bind_map.surface_count *
 213                             sizeof(*bind_map.surface_to_descriptor));
 214    bind_map.sampler_to_descriptor = (void *)
 215       blob_read_bytes(blob, bind_map.sampler_count *
 216                             sizeof(*bind_map.sampler_to_descriptor));
 217    blob_copy_bytes(blob, bind_map.push_ranges, sizeof(bind_map.push_ranges));
 218
 219    if (blob->overrun)
 220       return NULL;
 221
 222    return anv_shader_bin_create(device,
 223                                 key_data, key_size,
 224                                 kernel_data, kernel_size,
 225                                 constant_data, constant_data_size,
 226                                 prog_data, prog_data_size, prog_data_param,
 227                                 stats, num_stats, xfb_info, &bind_map);
 228 }
 229
 230 /* Remaining work:
 231  *
 232  * - Compact binding table layout so it's tight and not dependent on
 233  *   descriptor set layout.
 234  *
 235  * - Review prog_data struct for size and cacheability: struct
 236  *   brw_stage_prog_data has binding_table which uses a lot of uint32_t for 8
 237  *   bit quantities etc; use bit fields for all bools, eg dual_src_blend.
 238  */
 239
 240 static uint32_t
 241 shader_bin_key_hash_func(const void *void_key)
 242 {
 243    const struct anv_shader_bin_key *key = void_key;
 244    return _mesa_hash_data(key->data, key->size);
 245 }
 246
 247 static bool
 248 shader_bin_key_compare_func(const void *void_a, const void *void_b)
 249 {
 250    const struct anv_shader_bin_key *a = void_a, *b = void_b;
 251    if (a->size != b->size)
 252       return false;
 253
 254    return memcmp(a->data, b->data, a->size) == 0;
 255 }
 256
 257 static uint32_t
 258 sha1_hash_func(const void *sha1)
 259 {
 260    return _mesa_hash_data(sha1, 20);
 261 }
 262
 263 static bool
 264 sha1_compare_func(const void *sha1_a, const void *sha1_b)
 265 {
 266    return memcmp(sha1_a, sha1_b, 20) == 0;
 267 }
 268
 269 void
 270 anv_pipeline_cache_init(struct anv_pipeline_cache *cache,
 271                         struct anv_device *device,
 272                         bool cache_enabled)
 273 {
 274    cache->device = device;
 275    pthread_mutex_init(&cache->mutex, NULL);
 276
 277    if (cache_enabled) {
 278       cache->cache = _mesa_hash_table_create(NULL, shader_bin_key_hash_func,
 279                                              shader_bin_key_compare_func);
 280       cache->nir_cache = _mesa_hash_table_create(NULL, sha1_hash_func,
 281                                                  sha1_compare_func);
 282    } else {
 283       cache->cache = NULL;
 284       cache->nir_cache = NULL;
 285    }
 286 }
 287
 288 void
 289 anv_pipeline_cache_finish(struct anv_pipeline_cache *cache)
 290 {
 291    pthread_mutex_destroy(&cache->mutex);
 292
 293    if (cache->cache) {
 294       /* This is a bit unfortunate.  In order to keep things from randomly
 295        * going away, the shader cache has to hold a reference to all shader
 296        * binaries it contains.  We unref them when we destroy the cache.
 297        */
 298       hash_table_foreach(cache->cache, entry)
 299          anv_shader_bin_unref(cache->device, entry->data);
 300
 301       _mesa_hash_table_destroy(cache->cache, NULL);
 302    }
 303
 304    if (cache->nir_cache) {
 305       hash_table_foreach(cache->nir_cache, entry)
 306          ralloc_free(entry->data);
 307
 308       _mesa_hash_table_destroy(cache->nir_cache, NULL);
 309    }
 310 }
 311
 312 static struct anv_shader_bin *
 313 anv_pipeline_cache_search_locked(struct anv_pipeline_cache *cache,
 314                                  const void *key_data, uint32_t key_size)
 315 {
 316    uint32_t vla[1 + DIV_ROUND_UP(key_size, sizeof(uint32_t))];
 317    struct anv_shader_bin_key *key = (void *)vla;
 318    key->size = key_size;
 319    memcpy(key->data, key_data, key_size);
 320
 321    struct hash_entry *entry = _mesa_hash_table_search(cache->cache, key);
 322    if (entry)
 323       return entry->data;
 324    else
 325       return NULL;
 326 }
 327
 328 struct anv_shader_bin *
 329 anv_pipeline_cache_search(struct anv_pipeline_cache *cache,
 330                           const void *key_data, uint32_t key_size)
 331 {
 332    if (!cache->cache)
 333       return NULL;
 334
 335    pthread_mutex_lock(&cache->mutex);
 336
 337    struct anv_shader_bin *shader =
 338       anv_pipeline_cache_search_locked(cache, key_data, key_size);
 339
 340    pthread_mutex_unlock(&cache->mutex);
 341
 342    /* We increment refcount before handing it to the caller */
 343    if (shader)
 344       anv_shader_bin_ref(shader);
 345
 346    return shader;
 347 }
 348
 349 static void
 350 anv_pipeline_cache_add_shader_bin(struct anv_pipeline_cache *cache,
 351                                   struct anv_shader_bin *bin)
 352 {
 353    if (!cache->cache)
 354       return;
 355
 356    pthread_mutex_lock(&cache->mutex);
 357
 358    struct hash_entry *entry = _mesa_hash_table_search(cache->cache, bin->key);
 359    if (entry == NULL) {
 360       /* Take a reference for the cache */
 361       anv_shader_bin_ref(bin);
 362       _mesa_hash_table_insert(cache->cache, bin->key, bin);
 363    }
 364
 365    pthread_mutex_unlock(&cache->mutex);
 366 }
 367
 368 static struct anv_shader_bin *
 369 anv_pipeline_cache_add_shader_locked(struct anv_pipeline_cache *cache,
 370                                      const void *key_data, uint32_t key_size,
 371                                      const void *kernel_data,
 372                                      uint32_t kernel_size,
 373                                      const void *constant_data,
 374                                      uint32_t constant_data_size,
 375                                      const struct brw_stage_prog_data *prog_data,
 376                                      uint32_t prog_data_size,
 377                                      const void *prog_data_param,
 378                                      const struct brw_compile_stats *stats,
 379                                      uint32_t num_stats,
 380                                      const nir_xfb_info *xfb_info,
 381                                      const struct anv_pipeline_bind_map *bind_map)
 382 {
 383    struct anv_shader_bin *shader =
 384       anv_pipeline_cache_search_locked(cache, key_data, key_size);
 385    if (shader)
 386       return shader;
 387
 388    struct anv_shader_bin *bin =
 389       anv_shader_bin_create(cache->device, key_data, key_size,
 390                             kernel_data, kernel_size,
 391                             constant_data, constant_data_size,
 392                             prog_data, prog_data_size, prog_data_param,
 393                             stats, num_stats, xfb_info, bind_map);
 394    if (!bin)
 395       return NULL;
 396
 397    _mesa_hash_table_insert(cache->cache, bin->key, bin);
 398
 399    return bin;
 400 }
 401
 402 struct anv_shader_bin *
 403 anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache,
 404                                  const void *key_data, uint32_t key_size,
 405                                  const void *kernel_data, uint32_t kernel_size,
 406                                  const void *constant_data,
 407                                  uint32_t constant_data_size,
 408                                  const struct brw_stage_prog_data *prog_data,
 409                                  uint32_t prog_data_size,
 410                                  const struct brw_compile_stats *stats,
 411                                  uint32_t num_stats,
 412                                  const nir_xfb_info *xfb_info,
 413                                  const struct anv_pipeline_bind_map *bind_map)
 414 {
 415    if (cache->cache) {
 416       pthread_mutex_lock(&cache->mutex);
 417
 418       struct anv_shader_bin *bin =
 419          anv_pipeline_cache_add_shader_locked(cache, key_data, key_size,
 420                                               kernel_data, kernel_size,
 421                                               constant_data, constant_data_size,
 422                                               prog_data, prog_data_size,
 423                                               prog_data->param,
 424                                               stats, num_stats,
 425                                               xfb_info, bind_map);
 426
 427       pthread_mutex_unlock(&cache->mutex);
 428
 429       /* We increment refcount before handing it to the caller */
 430       if (bin)
 431          anv_shader_bin_ref(bin);
 432
 433       return bin;
 434    } else {
 435       /* In this case, we're not caching it so the caller owns it entirely */
 436       return anv_shader_bin_create(cache->device, key_data, key_size,
 437                                    kernel_data, kernel_size,
 438                                    constant_data, constant_data_size,
 439                                    prog_data, prog_data_size,
 440                                    prog_data->param,
 441                                    stats, num_stats,
 442                                    xfb_info, bind_map);
 443    }
 444 }
 445
 446 struct cache_header {
 447    uint32_t header_size;
 448    uint32_t header_version;
 449    uint32_t vendor_id;
 450    uint32_t device_id;
 451    uint8_t  uuid[VK_UUID_SIZE];
 452 };
 453
 454 static void
 455 anv_pipeline_cache_load(struct anv_pipeline_cache *cache,
 456                         const void *data, size_t size)
 457 {
 458    struct anv_device *device = cache->device;
 459    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
 460
 461    if (cache->cache == NULL)
 462       return;
 463
 464    struct blob_reader blob;
 465    blob_reader_init(&blob, data, size);
 466
 467    struct cache_header header;
 468    blob_copy_bytes(&blob, &header, sizeof(header));
 469    uint32_t count = blob_read_uint32(&blob);
 470    if (blob.overrun)
 471       return;
 472
 473    if (header.header_size < sizeof(header))
 474       return;
 475    if (header.header_version != VK_PIPELINE_CACHE_HEADER_VERSION_ONE)
 476       return;
 477    if (header.vendor_id != 0x8086)
 478       return;
 479    if (header.device_id != device->chipset_id)
 480       return;
 481    if (memcmp(header.uuid, pdevice->pipeline_cache_uuid, VK_UUID_SIZE) != 0)
 482       return;
 483
 484    for (uint32_t i = 0; i < count; i++) {
 485       struct anv_shader_bin *bin =
 486          anv_shader_bin_create_from_blob(device, &blob);
 487       if (!bin)
 488          break;
 489       _mesa_hash_table_insert(cache->cache, bin->key, bin);
 490    }
 491 }
 492
 493 VkResult anv_CreatePipelineCache(
 494     VkDevice                                    _device,
 495     const VkPipelineCacheCreateInfo*            pCreateInfo,
 496     const VkAllocationCallbacks*                pAllocator,
 497     VkPipelineCache*                            pPipelineCache)
 498 {
 499    ANV_FROM_HANDLE(anv_device, device, _device);
 500    struct anv_pipeline_cache *cache;
 501
 502    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
 503    assert(pCreateInfo->flags == 0);
 504
 505    cache = vk_alloc2(&device->alloc, pAllocator,
 506                        sizeof(*cache), 8,
 507                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 508    if (cache == NULL)
 509       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 510
 511    anv_pipeline_cache_init(cache, device,
 512                            device->instance->pipeline_cache_enabled);
 513
 514    if (pCreateInfo->initialDataSize > 0)
 515       anv_pipeline_cache_load(cache,
 516                               pCreateInfo->pInitialData,
 517                               pCreateInfo->initialDataSize);
 518
 519    *pPipelineCache = anv_pipeline_cache_to_handle(cache);
 520
 521    return VK_SUCCESS;
 522 }
 523
 524 void anv_DestroyPipelineCache(
 525     VkDevice                                    _device,
 526     VkPipelineCache                             _cache,
 527     const VkAllocationCallbacks*                pAllocator)
 528 {
 529    ANV_FROM_HANDLE(anv_device, device, _device);
 530    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 531
 532    if (!cache)
 533       return;
 534
 535    anv_pipeline_cache_finish(cache);
 536
 537    vk_free2(&device->alloc, pAllocator, cache);
 538 }
 539
 540 VkResult anv_GetPipelineCacheData(
 541     VkDevice                                    _device,
 542     VkPipelineCache                             _cache,
 543     size_t*                                     pDataSize,
 544     void*                                       pData)
 545 {
 546    ANV_FROM_HANDLE(anv_device, device, _device);
 547    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 548    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
 549
 550    struct blob blob;
 551    if (pData) {
 552       blob_init_fixed(&blob, pData, *pDataSize);
 553    } else {
 554       blob_init_fixed(&blob, NULL, SIZE_MAX);
 555    }
 556
 557    struct cache_header header = {
 558       .header_size = sizeof(struct cache_header),
 559       .header_version = VK_PIPELINE_CACHE_HEADER_VERSION_ONE,
 560       .vendor_id = 0x8086,
 561       .device_id = device->chipset_id,
 562    };
 563    memcpy(header.uuid, pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
 564    blob_write_bytes(&blob, &header, sizeof(header));
 565
 566    uint32_t count = 0;
 567    intptr_t count_offset = blob_reserve_uint32(&blob);
 568    if (count_offset < 0) {
 569       *pDataSize = 0;
 570       blob_finish(&blob);
 571       return VK_INCOMPLETE;
 572    }
 573
 574    VkResult result = VK_SUCCESS;
 575    if (cache->cache) {
 576       hash_table_foreach(cache->cache, entry) {
 577          struct anv_shader_bin *shader = entry->data;
 578
 579          size_t save_size = blob.size;
 580          if (!anv_shader_bin_write_to_blob(shader, &blob)) {
 581             /* If it fails reset to the previous size and bail */
 582             blob.size = save_size;
 583             result = VK_INCOMPLETE;
 584             break;
 585          }
 586
 587          count++;
 588       }
 589    }
 590
 591    blob_overwrite_uint32(&blob, count_offset, count);
 592
 593    *pDataSize = blob.size;
 594
 595    blob_finish(&blob);
 596
 597    return result;
 598 }
 599
 600 VkResult anv_MergePipelineCaches(
 601     VkDevice                                    _device,
 602     VkPipelineCache                             destCache,
 603     uint32_t                                    srcCacheCount,
 604     const VkPipelineCache*                      pSrcCaches)
 605 {
 606    ANV_FROM_HANDLE(anv_pipeline_cache, dst, destCache);
 607
 608    if (!dst->cache)
 609       return VK_SUCCESS;
 610
 611    for (uint32_t i = 0; i < srcCacheCount; i++) {
 612       ANV_FROM_HANDLE(anv_pipeline_cache, src, pSrcCaches[i]);
 613       if (!src->cache)
 614          continue;
 615
 616       hash_table_foreach(src->cache, entry) {
 617          struct anv_shader_bin *bin = entry->data;
 618          assert(bin);
 619
 620          if (_mesa_hash_table_search(dst->cache, bin->key))
 621             continue;
 622
 623          anv_shader_bin_ref(bin);
 624          _mesa_hash_table_insert(dst->cache, bin->key, bin);
 625       }
 626    }
 627
 628    return VK_SUCCESS;
 629 }
 630
 631 struct anv_shader_bin *
 632 anv_device_search_for_kernel(struct anv_device *device,
 633                              struct anv_pipeline_cache *cache,
 634                              const void *key_data, uint32_t key_size,
 635                              bool *user_cache_hit)
 636 {
 637    struct anv_shader_bin *bin;
 638
 639    *user_cache_hit = false;
 640
 641    if (cache) {
 642       bin = anv_pipeline_cache_search(cache, key_data, key_size);
 643       if (bin) {
 644          *user_cache_hit = cache != &device->default_pipeline_cache;
 645          return bin;
 646       }
 647    }
 648
 649 #ifdef ENABLE_SHADER_CACHE
 650    struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache;
 651    if (disk_cache && device->instance->pipeline_cache_enabled) {
 652       cache_key cache_key;
 653       disk_cache_compute_key(disk_cache, key_data, key_size, cache_key);
 654
 655       size_t buffer_size;
 656       uint8_t *buffer = disk_cache_get(disk_cache, cache_key, &buffer_size);
 657       if (buffer) {
 658          struct blob_reader blob;
 659          blob_reader_init(&blob, buffer, buffer_size);
 660          bin = anv_shader_bin_create_from_blob(device, &blob);
 661          free(buffer);
 662
 663          if (bin) {
 664             if (cache)
 665                anv_pipeline_cache_add_shader_bin(cache, bin);
 666             return bin;
 667          }
 668       }
 669    }
 670 #endif
 671
 672    return NULL;
 673 }
 674
 675 struct anv_shader_bin *
 676 anv_device_upload_kernel(struct anv_device *device,
 677                          struct anv_pipeline_cache *cache,
 678                          const void *key_data, uint32_t key_size,
 679                          const void *kernel_data, uint32_t kernel_size,
 680                          const void *constant_data,
 681                          uint32_t constant_data_size,
 682                          const struct brw_stage_prog_data *prog_data,
 683                          uint32_t prog_data_size,
 684                          const struct brw_compile_stats *stats,
 685                          uint32_t num_stats,
 686                          const nir_xfb_info *xfb_info,
 687                          const struct anv_pipeline_bind_map *bind_map)
 688 {
 689    struct anv_shader_bin *bin;
 690    if (cache) {
 691       bin = anv_pipeline_cache_upload_kernel(cache, key_data, key_size,
 692                                              kernel_data, kernel_size,
 693                                              constant_data, constant_data_size,
 694                                              prog_data, prog_data_size,
 695                                              stats, num_stats,
 696                                              xfb_info, bind_map);
 697    } else {
 698       bin = anv_shader_bin_create(device, key_data, key_size,
 699                                   kernel_data, kernel_size,
 700                                   constant_data, constant_data_size,
 701                                   prog_data, prog_data_size,
 702                                   prog_data->param,
 703                                   stats, num_stats,
 704                                   xfb_info, bind_map);
 705    }
 706
 707    if (bin == NULL)
 708       return NULL;
 709
 710 #ifdef ENABLE_SHADER_CACHE
 711    struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache;
 712    if (disk_cache) {
 713       struct blob binary;
 714       blob_init(&binary);
 715       if (anv_shader_bin_write_to_blob(bin, &binary)) {
 716          cache_key cache_key;
 717          disk_cache_compute_key(disk_cache, key_data, key_size, cache_key);
 718
 719          disk_cache_put(disk_cache, cache_key, binary.data, binary.size, NULL);
 720       }
 721
 722       blob_finish(&binary);
 723    }
 724 #endif
 725
 726    return bin;
 727 }
 728
 729 struct serialized_nir {
 730    unsigned char sha1_key[20];
 731    size_t size;
 732    char data[0];
 733 };
 734
 735 struct nir_shader *
 736 anv_device_search_for_nir(struct anv_device *device,
 737                           struct anv_pipeline_cache *cache,
 738                           const nir_shader_compiler_options *nir_options,
 739                           unsigned char sha1_key[20],
 740                           void *mem_ctx)
 741 {
 742    if (cache && cache->nir_cache) {
 743       const struct serialized_nir *snir = NULL;
 744
 745       pthread_mutex_lock(&cache->mutex);
 746       struct hash_entry *entry =
 747          _mesa_hash_table_search(cache->nir_cache, sha1_key);
 748       if (entry)
 749          snir = entry->data;
 750       pthread_mutex_unlock(&cache->mutex);
 751
 752       if (snir) {
 753          struct blob_reader blob;
 754          blob_reader_init(&blob, snir->data, snir->size);
 755
 756          nir_shader *nir = nir_deserialize(mem_ctx, nir_options, &blob);
 757          if (blob.overrun) {
 758             ralloc_free(nir);
 759          } else {
 760             return nir;
 761          }
 762       }
 763    }
 764
 765    return NULL;
 766 }
 767
 768 void
 769 anv_device_upload_nir(struct anv_device *device,
 770                       struct anv_pipeline_cache *cache,
 771                       const struct nir_shader *nir,
 772                       unsigned char sha1_key[20])
 773 {
 774    if (cache && cache->nir_cache) {
 775       pthread_mutex_lock(&cache->mutex);
 776       struct hash_entry *entry =
 777          _mesa_hash_table_search(cache->nir_cache, sha1_key);
 778       pthread_mutex_unlock(&cache->mutex);
 779       if (entry)
 780          return;
 781
 782       struct blob blob;
 783       blob_init(&blob);
 784
 785       nir_serialize(&blob, nir, false);
 786       if (blob.out_of_memory) {
 787          blob_finish(&blob);
 788          return;
 789       }
 790
 791       pthread_mutex_lock(&cache->mutex);
 792       /* Because ralloc isn't thread-safe, we have to do all this inside the
 793        * lock.  We could unlock for the big memcpy but it's probably not worth
 794        * the hassle.
 795        */
 796       entry = _mesa_hash_table_search(cache->nir_cache, sha1_key);
 797       if (entry) {
 798          blob_finish(&blob);
 799          pthread_mutex_unlock(&cache->mutex);
 800          return;
 801       }
 802
 803       struct serialized_nir *snir =
 804          ralloc_size(cache->nir_cache, sizeof(*snir) + blob.size);
 805       memcpy(snir->sha1_key, sha1_key, 20);
 806       snir->size = blob.size;
 807       memcpy(snir->data, blob.data, blob.size);
 808
 809       blob_finish(&blob);
 810
 811       _mesa_hash_table_insert(cache->nir_cache, snir->sha1_key, snir);
 812
 813       pthread_mutex_unlock(&cache->mutex);
 814    }
 815 }