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_bytes(blob, shader->bind_map.surface_sha1,
 165                     sizeof(shader->bind_map.surface_sha1));
 166    blob_write_bytes(blob, shader->bind_map.sampler_sha1,
 167                     sizeof(shader->bind_map.sampler_sha1));
 168    blob_write_bytes(blob, shader->bind_map.push_sha1,
 169                     sizeof(shader->bind_map.push_sha1));
 170    blob_write_uint32(blob, shader->bind_map.surface_count);
 171    blob_write_uint32(blob, shader->bind_map.sampler_count);
 172    blob_write_bytes(blob, shader->bind_map.surface_to_descriptor,
 173                     shader->bind_map.surface_count *
 174                     sizeof(*shader->bind_map.surface_to_descriptor));
 175    blob_write_bytes(blob, shader->bind_map.sampler_to_descriptor,
 176                     shader->bind_map.sampler_count *
 177                     sizeof(*shader->bind_map.sampler_to_descriptor));
 178    blob_write_bytes(blob, shader->bind_map.push_ranges,
 179                     sizeof(shader->bind_map.push_ranges));
 180
 181    return !blob->out_of_memory;
 182 }
 183
 184 static struct anv_shader_bin *
 185 anv_shader_bin_create_from_blob(struct anv_device *device,
 186                                 struct blob_reader *blob)
 187 {
 188    uint32_t key_size = blob_read_uint32(blob);
 189    const void *key_data = blob_read_bytes(blob, key_size);
 190
 191    uint32_t kernel_size = blob_read_uint32(blob);
 192    const void *kernel_data = blob_read_bytes(blob, kernel_size);
 193
 194    uint32_t constant_data_size = blob_read_uint32(blob);
 195    const void *constant_data = blob_read_bytes(blob, constant_data_size);
 196
 197    uint32_t prog_data_size = blob_read_uint32(blob);
 198    const struct brw_stage_prog_data *prog_data =
 199       blob_read_bytes(blob, prog_data_size);
 200    if (blob->overrun)
 201       return NULL;
 202    const void *prog_data_param =
 203       blob_read_bytes(blob, prog_data->nr_params * sizeof(*prog_data->param));
 204
 205    uint32_t num_stats = blob_read_uint32(blob);
 206    const struct brw_compile_stats *stats =
 207       blob_read_bytes(blob, num_stats * sizeof(stats[0]));
 208
 209    const nir_xfb_info *xfb_info = NULL;
 210    uint32_t xfb_size = blob_read_uint32(blob);
 211    if (xfb_size)
 212       xfb_info = blob_read_bytes(blob, xfb_size);
 213
 214    struct anv_pipeline_bind_map bind_map;
 215    blob_copy_bytes(blob, bind_map.surface_sha1, sizeof(bind_map.surface_sha1));
 216    blob_copy_bytes(blob, bind_map.sampler_sha1, sizeof(bind_map.sampler_sha1));
 217    blob_copy_bytes(blob, bind_map.push_sha1, sizeof(bind_map.push_sha1));
 218    bind_map.surface_count = blob_read_uint32(blob);
 219    bind_map.sampler_count = blob_read_uint32(blob);
 220    bind_map.surface_to_descriptor = (void *)
 221       blob_read_bytes(blob, bind_map.surface_count *
 222                             sizeof(*bind_map.surface_to_descriptor));
 223    bind_map.sampler_to_descriptor = (void *)
 224       blob_read_bytes(blob, bind_map.sampler_count *
 225                             sizeof(*bind_map.sampler_to_descriptor));
 226    blob_copy_bytes(blob, bind_map.push_ranges, sizeof(bind_map.push_ranges));
 227
 228    if (blob->overrun)
 229       return NULL;
 230
 231    return anv_shader_bin_create(device,
 232                                 key_data, key_size,
 233                                 kernel_data, kernel_size,
 234                                 constant_data, constant_data_size,
 235                                 prog_data, prog_data_size, prog_data_param,
 236                                 stats, num_stats, xfb_info, &bind_map);
 237 }
 238
 239 /* Remaining work:
 240  *
 241  * - Compact binding table layout so it's tight and not dependent on
 242  *   descriptor set layout.
 243  *
 244  * - Review prog_data struct for size and cacheability: struct
 245  *   brw_stage_prog_data has binding_table which uses a lot of uint32_t for 8
 246  *   bit quantities etc; use bit fields for all bools, eg dual_src_blend.
 247  */
 248
 249 static uint32_t
 250 shader_bin_key_hash_func(const void *void_key)
 251 {
 252    const struct anv_shader_bin_key *key = void_key;
 253    return _mesa_hash_data(key->data, key->size);
 254 }
 255
 256 static bool
 257 shader_bin_key_compare_func(const void *void_a, const void *void_b)
 258 {
 259    const struct anv_shader_bin_key *a = void_a, *b = void_b;
 260    if (a->size != b->size)
 261       return false;
 262
 263    return memcmp(a->data, b->data, a->size) == 0;
 264 }
 265
 266 static uint32_t
 267 sha1_hash_func(const void *sha1)
 268 {
 269    return _mesa_hash_data(sha1, 20);
 270 }
 271
 272 static bool
 273 sha1_compare_func(const void *sha1_a, const void *sha1_b)
 274 {
 275    return memcmp(sha1_a, sha1_b, 20) == 0;
 276 }
 277
 278 void
 279 anv_pipeline_cache_init(struct anv_pipeline_cache *cache,
 280                         struct anv_device *device,
 281                         bool cache_enabled)
 282 {
 283    cache->device = device;
 284    pthread_mutex_init(&cache->mutex, NULL);
 285
 286    if (cache_enabled) {
 287       cache->cache = _mesa_hash_table_create(NULL, shader_bin_key_hash_func,
 288                                              shader_bin_key_compare_func);
 289       cache->nir_cache = _mesa_hash_table_create(NULL, sha1_hash_func,
 290                                                  sha1_compare_func);
 291    } else {
 292       cache->cache = NULL;
 293       cache->nir_cache = NULL;
 294    }
 295 }
 296
 297 void
 298 anv_pipeline_cache_finish(struct anv_pipeline_cache *cache)
 299 {
 300    pthread_mutex_destroy(&cache->mutex);
 301
 302    if (cache->cache) {
 303       /* This is a bit unfortunate.  In order to keep things from randomly
 304        * going away, the shader cache has to hold a reference to all shader
 305        * binaries it contains.  We unref them when we destroy the cache.
 306        */
 307       hash_table_foreach(cache->cache, entry)
 308          anv_shader_bin_unref(cache->device, entry->data);
 309
 310       _mesa_hash_table_destroy(cache->cache, NULL);
 311    }
 312
 313    if (cache->nir_cache) {
 314       hash_table_foreach(cache->nir_cache, entry)
 315          ralloc_free(entry->data);
 316
 317       _mesa_hash_table_destroy(cache->nir_cache, NULL);
 318    }
 319 }
 320
 321 static struct anv_shader_bin *
 322 anv_pipeline_cache_search_locked(struct anv_pipeline_cache *cache,
 323                                  const void *key_data, uint32_t key_size)
 324 {
 325    uint32_t vla[1 + DIV_ROUND_UP(key_size, sizeof(uint32_t))];
 326    struct anv_shader_bin_key *key = (void *)vla;
 327    key->size = key_size;
 328    memcpy(key->data, key_data, key_size);
 329
 330    struct hash_entry *entry = _mesa_hash_table_search(cache->cache, key);
 331    if (entry)
 332       return entry->data;
 333    else
 334       return NULL;
 335 }
 336
 337 struct anv_shader_bin *
 338 anv_pipeline_cache_search(struct anv_pipeline_cache *cache,
 339                           const void *key_data, uint32_t key_size)
 340 {
 341    if (!cache->cache)
 342       return NULL;
 343
 344    pthread_mutex_lock(&cache->mutex);
 345
 346    struct anv_shader_bin *shader =
 347       anv_pipeline_cache_search_locked(cache, key_data, key_size);
 348
 349    pthread_mutex_unlock(&cache->mutex);
 350
 351    /* We increment refcount before handing it to the caller */
 352    if (shader)
 353       anv_shader_bin_ref(shader);
 354
 355    return shader;
 356 }
 357
 358 static void
 359 anv_pipeline_cache_add_shader_bin(struct anv_pipeline_cache *cache,
 360                                   struct anv_shader_bin *bin)
 361 {
 362    if (!cache->cache)
 363       return;
 364
 365    pthread_mutex_lock(&cache->mutex);
 366
 367    struct hash_entry *entry = _mesa_hash_table_search(cache->cache, bin->key);
 368    if (entry == NULL) {
 369       /* Take a reference for the cache */
 370       anv_shader_bin_ref(bin);
 371       _mesa_hash_table_insert(cache->cache, bin->key, bin);
 372    }
 373
 374    pthread_mutex_unlock(&cache->mutex);
 375 }
 376
 377 static struct anv_shader_bin *
 378 anv_pipeline_cache_add_shader_locked(struct anv_pipeline_cache *cache,
 379                                      const void *key_data, uint32_t key_size,
 380                                      const void *kernel_data,
 381                                      uint32_t kernel_size,
 382                                      const void *constant_data,
 383                                      uint32_t constant_data_size,
 384                                      const struct brw_stage_prog_data *prog_data,
 385                                      uint32_t prog_data_size,
 386                                      const void *prog_data_param,
 387                                      const struct brw_compile_stats *stats,
 388                                      uint32_t num_stats,
 389                                      const nir_xfb_info *xfb_info,
 390                                      const struct anv_pipeline_bind_map *bind_map)
 391 {
 392    struct anv_shader_bin *shader =
 393       anv_pipeline_cache_search_locked(cache, key_data, key_size);
 394    if (shader)
 395       return shader;
 396
 397    struct anv_shader_bin *bin =
 398       anv_shader_bin_create(cache->device, key_data, key_size,
 399                             kernel_data, kernel_size,
 400                             constant_data, constant_data_size,
 401                             prog_data, prog_data_size, prog_data_param,
 402                             stats, num_stats, xfb_info, bind_map);
 403    if (!bin)
 404       return NULL;
 405
 406    _mesa_hash_table_insert(cache->cache, bin->key, bin);
 407
 408    return bin;
 409 }
 410
 411 struct anv_shader_bin *
 412 anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache,
 413                                  const void *key_data, uint32_t key_size,
 414                                  const void *kernel_data, uint32_t kernel_size,
 415                                  const void *constant_data,
 416                                  uint32_t constant_data_size,
 417                                  const struct brw_stage_prog_data *prog_data,
 418                                  uint32_t prog_data_size,
 419                                  const struct brw_compile_stats *stats,
 420                                  uint32_t num_stats,
 421                                  const nir_xfb_info *xfb_info,
 422                                  const struct anv_pipeline_bind_map *bind_map)
 423 {
 424    if (cache->cache) {
 425       pthread_mutex_lock(&cache->mutex);
 426
 427       struct anv_shader_bin *bin =
 428          anv_pipeline_cache_add_shader_locked(cache, key_data, key_size,
 429                                               kernel_data, kernel_size,
 430                                               constant_data, constant_data_size,
 431                                               prog_data, prog_data_size,
 432                                               prog_data->param,
 433                                               stats, num_stats,
 434                                               xfb_info, bind_map);
 435
 436       pthread_mutex_unlock(&cache->mutex);
 437
 438       /* We increment refcount before handing it to the caller */
 439       if (bin)
 440          anv_shader_bin_ref(bin);
 441
 442       return bin;
 443    } else {
 444       /* In this case, we're not caching it so the caller owns it entirely */
 445       return anv_shader_bin_create(cache->device, key_data, key_size,
 446                                    kernel_data, kernel_size,
 447                                    constant_data, constant_data_size,
 448                                    prog_data, prog_data_size,
 449                                    prog_data->param,
 450                                    stats, num_stats,
 451                                    xfb_info, bind_map);
 452    }
 453 }
 454
 455 struct cache_header {
 456    uint32_t header_size;
 457    uint32_t header_version;
 458    uint32_t vendor_id;
 459    uint32_t device_id;
 460    uint8_t  uuid[VK_UUID_SIZE];
 461 };
 462
 463 static void
 464 anv_pipeline_cache_load(struct anv_pipeline_cache *cache,
 465                         const void *data, size_t size)
 466 {
 467    struct anv_device *device = cache->device;
 468    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
 469
 470    if (cache->cache == NULL)
 471       return;
 472
 473    struct blob_reader blob;
 474    blob_reader_init(&blob, data, size);
 475
 476    struct cache_header header;
 477    blob_copy_bytes(&blob, &header, sizeof(header));
 478    uint32_t count = blob_read_uint32(&blob);
 479    if (blob.overrun)
 480       return;
 481
 482    if (header.header_size < sizeof(header))
 483       return;
 484    if (header.header_version != VK_PIPELINE_CACHE_HEADER_VERSION_ONE)
 485       return;
 486    if (header.vendor_id != 0x8086)
 487       return;
 488    if (header.device_id != device->chipset_id)
 489       return;
 490    if (memcmp(header.uuid, pdevice->pipeline_cache_uuid, VK_UUID_SIZE) != 0)
 491       return;
 492
 493    for (uint32_t i = 0; i < count; i++) {
 494       struct anv_shader_bin *bin =
 495          anv_shader_bin_create_from_blob(device, &blob);
 496       if (!bin)
 497          break;
 498       _mesa_hash_table_insert(cache->cache, bin->key, bin);
 499    }
 500 }
 501
 502 VkResult anv_CreatePipelineCache(
 503     VkDevice                                    _device,
 504     const VkPipelineCacheCreateInfo*            pCreateInfo,
 505     const VkAllocationCallbacks*                pAllocator,
 506     VkPipelineCache*                            pPipelineCache)
 507 {
 508    ANV_FROM_HANDLE(anv_device, device, _device);
 509    struct anv_pipeline_cache *cache;
 510
 511    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
 512    assert(pCreateInfo->flags == 0);
 513
 514    cache = vk_alloc2(&device->alloc, pAllocator,
 515                        sizeof(*cache), 8,
 516                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 517    if (cache == NULL)
 518       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 519
 520    anv_pipeline_cache_init(cache, device,
 521                            device->instance->pipeline_cache_enabled);
 522
 523    if (pCreateInfo->initialDataSize > 0)
 524       anv_pipeline_cache_load(cache,
 525                               pCreateInfo->pInitialData,
 526                               pCreateInfo->initialDataSize);
 527
 528    *pPipelineCache = anv_pipeline_cache_to_handle(cache);
 529
 530    return VK_SUCCESS;
 531 }
 532
 533 void anv_DestroyPipelineCache(
 534     VkDevice                                    _device,
 535     VkPipelineCache                             _cache,
 536     const VkAllocationCallbacks*                pAllocator)
 537 {
 538    ANV_FROM_HANDLE(anv_device, device, _device);
 539    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 540
 541    if (!cache)
 542       return;
 543
 544    anv_pipeline_cache_finish(cache);
 545
 546    vk_free2(&device->alloc, pAllocator, cache);
 547 }
 548
 549 VkResult anv_GetPipelineCacheData(
 550     VkDevice                                    _device,
 551     VkPipelineCache                             _cache,
 552     size_t*                                     pDataSize,
 553     void*                                       pData)
 554 {
 555    ANV_FROM_HANDLE(anv_device, device, _device);
 556    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 557    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
 558
 559    struct blob blob;
 560    if (pData) {
 561       blob_init_fixed(&blob, pData, *pDataSize);
 562    } else {
 563       blob_init_fixed(&blob, NULL, SIZE_MAX);
 564    }
 565
 566    struct cache_header header = {
 567       .header_size = sizeof(struct cache_header),
 568       .header_version = VK_PIPELINE_CACHE_HEADER_VERSION_ONE,
 569       .vendor_id = 0x8086,
 570       .device_id = device->chipset_id,
 571    };
 572    memcpy(header.uuid, pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
 573    blob_write_bytes(&blob, &header, sizeof(header));
 574
 575    uint32_t count = 0;
 576    intptr_t count_offset = blob_reserve_uint32(&blob);
 577    if (count_offset < 0) {
 578       *pDataSize = 0;
 579       blob_finish(&blob);
 580       return VK_INCOMPLETE;
 581    }
 582
 583    VkResult result = VK_SUCCESS;
 584    if (cache->cache) {
 585       hash_table_foreach(cache->cache, entry) {
 586          struct anv_shader_bin *shader = entry->data;
 587
 588          size_t save_size = blob.size;
 589          if (!anv_shader_bin_write_to_blob(shader, &blob)) {
 590             /* If it fails reset to the previous size and bail */
 591             blob.size = save_size;
 592             result = VK_INCOMPLETE;
 593             break;
 594          }
 595
 596          count++;
 597       }
 598    }
 599
 600    blob_overwrite_uint32(&blob, count_offset, count);
 601
 602    *pDataSize = blob.size;
 603
 604    blob_finish(&blob);
 605
 606    return result;
 607 }
 608
 609 VkResult anv_MergePipelineCaches(
 610     VkDevice                                    _device,
 611     VkPipelineCache                             destCache,
 612     uint32_t                                    srcCacheCount,
 613     const VkPipelineCache*                      pSrcCaches)
 614 {
 615    ANV_FROM_HANDLE(anv_pipeline_cache, dst, destCache);
 616
 617    if (!dst->cache)
 618       return VK_SUCCESS;
 619
 620    for (uint32_t i = 0; i < srcCacheCount; i++) {
 621       ANV_FROM_HANDLE(anv_pipeline_cache, src, pSrcCaches[i]);
 622       if (!src->cache)
 623          continue;
 624
 625       hash_table_foreach(src->cache, entry) {
 626          struct anv_shader_bin *bin = entry->data;
 627          assert(bin);
 628
 629          if (_mesa_hash_table_search(dst->cache, bin->key))
 630             continue;
 631
 632          anv_shader_bin_ref(bin);
 633          _mesa_hash_table_insert(dst->cache, bin->key, bin);
 634       }
 635    }
 636
 637    return VK_SUCCESS;
 638 }
 639
 640 struct anv_shader_bin *
 641 anv_device_search_for_kernel(struct anv_device *device,
 642                              struct anv_pipeline_cache *cache,
 643                              const void *key_data, uint32_t key_size,
 644                              bool *user_cache_hit)
 645 {
 646    struct anv_shader_bin *bin;
 647
 648    *user_cache_hit = false;
 649
 650    if (cache) {
 651       bin = anv_pipeline_cache_search(cache, key_data, key_size);
 652       if (bin) {
 653          *user_cache_hit = cache != &device->default_pipeline_cache;
 654          return bin;
 655       }
 656    }
 657
 658 #ifdef ENABLE_SHADER_CACHE
 659    struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache;
 660    if (disk_cache && device->instance->pipeline_cache_enabled) {
 661       cache_key cache_key;
 662       disk_cache_compute_key(disk_cache, key_data, key_size, cache_key);
 663
 664       size_t buffer_size;
 665       uint8_t *buffer = disk_cache_get(disk_cache, cache_key, &buffer_size);
 666       if (buffer) {
 667          struct blob_reader blob;
 668          blob_reader_init(&blob, buffer, buffer_size);
 669          bin = anv_shader_bin_create_from_blob(device, &blob);
 670          free(buffer);
 671
 672          if (bin) {
 673             if (cache)
 674                anv_pipeline_cache_add_shader_bin(cache, bin);
 675             return bin;
 676          }
 677       }
 678    }
 679 #endif
 680
 681    return NULL;
 682 }
 683
 684 struct anv_shader_bin *
 685 anv_device_upload_kernel(struct anv_device *device,
 686                          struct anv_pipeline_cache *cache,
 687                          const void *key_data, uint32_t key_size,
 688                          const void *kernel_data, uint32_t kernel_size,
 689                          const void *constant_data,
 690                          uint32_t constant_data_size,
 691                          const struct brw_stage_prog_data *prog_data,
 692                          uint32_t prog_data_size,
 693                          const struct brw_compile_stats *stats,
 694                          uint32_t num_stats,
 695                          const nir_xfb_info *xfb_info,
 696                          const struct anv_pipeline_bind_map *bind_map)
 697 {
 698    struct anv_shader_bin *bin;
 699    if (cache) {
 700       bin = anv_pipeline_cache_upload_kernel(cache, key_data, key_size,
 701                                              kernel_data, kernel_size,
 702                                              constant_data, constant_data_size,
 703                                              prog_data, prog_data_size,
 704                                              stats, num_stats,
 705                                              xfb_info, bind_map);
 706    } else {
 707       bin = anv_shader_bin_create(device, key_data, key_size,
 708                                   kernel_data, kernel_size,
 709                                   constant_data, constant_data_size,
 710                                   prog_data, prog_data_size,
 711                                   prog_data->param,
 712                                   stats, num_stats,
 713                                   xfb_info, bind_map);
 714    }
 715
 716    if (bin == NULL)
 717       return NULL;
 718
 719 #ifdef ENABLE_SHADER_CACHE
 720    struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache;
 721    if (disk_cache) {
 722       struct blob binary;
 723       blob_init(&binary);
 724       if (anv_shader_bin_write_to_blob(bin, &binary)) {
 725          cache_key cache_key;
 726          disk_cache_compute_key(disk_cache, key_data, key_size, cache_key);
 727
 728          disk_cache_put(disk_cache, cache_key, binary.data, binary.size, NULL);
 729       }
 730
 731       blob_finish(&binary);
 732    }
 733 #endif
 734
 735    return bin;
 736 }
 737
 738 struct serialized_nir {
 739    unsigned char sha1_key[20];
 740    size_t size;
 741    char data[0];
 742 };
 743
 744 struct nir_shader *
 745 anv_device_search_for_nir(struct anv_device *device,
 746                           struct anv_pipeline_cache *cache,
 747                           const nir_shader_compiler_options *nir_options,
 748                           unsigned char sha1_key[20],
 749                           void *mem_ctx)
 750 {
 751    if (cache && cache->nir_cache) {
 752       const struct serialized_nir *snir = NULL;
 753
 754       pthread_mutex_lock(&cache->mutex);
 755       struct hash_entry *entry =
 756          _mesa_hash_table_search(cache->nir_cache, sha1_key);
 757       if (entry)
 758          snir = entry->data;
 759       pthread_mutex_unlock(&cache->mutex);
 760
 761       if (snir) {
 762          struct blob_reader blob;
 763          blob_reader_init(&blob, snir->data, snir->size);
 764
 765          nir_shader *nir = nir_deserialize(mem_ctx, nir_options, &blob);
 766          if (blob.overrun) {
 767             ralloc_free(nir);
 768          } else {
 769             return nir;
 770          }
 771       }
 772    }
 773
 774    return NULL;
 775 }
 776
 777 void
 778 anv_device_upload_nir(struct anv_device *device,
 779                       struct anv_pipeline_cache *cache,
 780                       const struct nir_shader *nir,
 781                       unsigned char sha1_key[20])
 782 {
 783    if (cache && cache->nir_cache) {
 784       pthread_mutex_lock(&cache->mutex);
 785       struct hash_entry *entry =
 786          _mesa_hash_table_search(cache->nir_cache, sha1_key);
 787       pthread_mutex_unlock(&cache->mutex);
 788       if (entry)
 789          return;
 790
 791       struct blob blob;
 792       blob_init(&blob);
 793
 794       nir_serialize(&blob, nir, false);
 795       if (blob.out_of_memory) {
 796          blob_finish(&blob);
 797          return;
 798       }
 799
 800       pthread_mutex_lock(&cache->mutex);
 801       /* Because ralloc isn't thread-safe, we have to do all this inside the
 802        * lock.  We could unlock for the big memcpy but it's probably not worth
 803        * the hassle.
 804        */
 805       entry = _mesa_hash_table_search(cache->nir_cache, sha1_key);
 806       if (entry) {
 807          blob_finish(&blob);
 808          pthread_mutex_unlock(&cache->mutex);
 809          return;
 810       }
 811
 812       struct serialized_nir *snir =
 813          ralloc_size(cache->nir_cache, sizeof(*snir) + blob.size);
 814       memcpy(snir->sha1_key, sha1_key, 20);
 815       snir->size = blob.size;
 816       memcpy(snir->data, blob.data, blob.size);
 817
 818       blob_finish(&blob);
 819
 820       _mesa_hash_table_insert(cache->nir_cache, snir->sha1_key, snir);
 821
 822       pthread_mutex_unlock(&cache->mutex);
 823    }
 824 }