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