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_uint32(blob, shader->bind_map.image_param_count);
 158    blob_write_bytes(blob, shader->bind_map.surface_to_descriptor,
 159                     shader->bind_map.surface_count *
 160                     sizeof(*shader->bind_map.surface_to_descriptor));
 161    blob_write_bytes(blob, shader->bind_map.sampler_to_descriptor,
 162                     shader->bind_map.sampler_count *
 163                     sizeof(*shader->bind_map.sampler_to_descriptor));
 164
 165    return !blob->out_of_memory;
 166 }
 167
 168 static struct anv_shader_bin *
 169 anv_shader_bin_create_from_blob(struct anv_device *device,
 170                                 struct blob_reader *blob)
 171 {
 172    uint32_t key_size = blob_read_uint32(blob);
 173    const void *key_data = blob_read_bytes(blob, key_size);
 174
 175    uint32_t kernel_size = blob_read_uint32(blob);
 176    const void *kernel_data = blob_read_bytes(blob, kernel_size);
 177
 178    uint32_t constant_data_size = blob_read_uint32(blob);
 179    const void *constant_data = blob_read_bytes(blob, constant_data_size);
 180
 181    uint32_t prog_data_size = blob_read_uint32(blob);
 182    const struct brw_stage_prog_data *prog_data =
 183       blob_read_bytes(blob, prog_data_size);
 184    if (blob->overrun)
 185       return NULL;
 186    const void *prog_data_param =
 187       blob_read_bytes(blob, prog_data->nr_params * sizeof(*prog_data->param));
 188
 189    const nir_xfb_info *xfb_info = NULL;
 190    uint32_t xfb_size = blob_read_uint32(blob);
 191    if (xfb_size)
 192       xfb_info = blob_read_bytes(blob, xfb_size);
 193
 194    struct anv_pipeline_bind_map bind_map;
 195    bind_map.surface_count = blob_read_uint32(blob);
 196    bind_map.sampler_count = blob_read_uint32(blob);
 197    bind_map.image_param_count = blob_read_uint32(blob);
 198    bind_map.surface_to_descriptor = (void *)
 199       blob_read_bytes(blob, bind_map.surface_count *
 200                             sizeof(*bind_map.surface_to_descriptor));
 201    bind_map.sampler_to_descriptor = (void *)
 202       blob_read_bytes(blob, bind_map.sampler_count *
 203                             sizeof(*bind_map.sampler_to_descriptor));
 204
 205    if (blob->overrun)
 206       return NULL;
 207
 208    return anv_shader_bin_create(device,
 209                                 key_data, key_size,
 210                                 kernel_data, kernel_size,
 211                                 constant_data, constant_data_size,
 212                                 prog_data, prog_data_size, prog_data_param,
 213                                 xfb_info, &bind_map);
 214 }
 215
 216 /* Remaining work:
 217  *
 218  * - Compact binding table layout so it's tight and not dependent on
 219  *   descriptor set layout.
 220  *
 221  * - Review prog_data struct for size and cacheability: struct
 222  *   brw_stage_prog_data has binding_table which uses a lot of uint32_t for 8
 223  *   bit quantities etc; use bit fields for all bools, eg dual_src_blend.
 224  */
 225
 226 static uint32_t
 227 shader_bin_key_hash_func(const void *void_key)
 228 {
 229    const struct anv_shader_bin_key *key = void_key;
 230    return _mesa_hash_data(key->data, key->size);
 231 }
 232
 233 static bool
 234 shader_bin_key_compare_func(const void *void_a, const void *void_b)
 235 {
 236    const struct anv_shader_bin_key *a = void_a, *b = void_b;
 237    if (a->size != b->size)
 238       return false;
 239
 240    return memcmp(a->data, b->data, a->size) == 0;
 241 }
 242
 243 static uint32_t
 244 sha1_hash_func(const void *sha1)
 245 {
 246    return _mesa_hash_data(sha1, 20);
 247 }
 248
 249 static bool
 250 sha1_compare_func(const void *sha1_a, const void *sha1_b)
 251 {
 252    return memcmp(sha1_a, sha1_b, 20) == 0;
 253 }
 254
 255 void
 256 anv_pipeline_cache_init(struct anv_pipeline_cache *cache,
 257                         struct anv_device *device,
 258                         bool cache_enabled)
 259 {
 260    cache->device = device;
 261    pthread_mutex_init(&cache->mutex, NULL);
 262
 263    if (cache_enabled) {
 264       cache->cache = _mesa_hash_table_create(NULL, shader_bin_key_hash_func,
 265                                              shader_bin_key_compare_func);
 266       cache->nir_cache = _mesa_hash_table_create(NULL, sha1_hash_func,
 267                                                  sha1_compare_func);
 268    } else {
 269       cache->cache = NULL;
 270       cache->nir_cache = NULL;
 271    }
 272 }
 273
 274 void
 275 anv_pipeline_cache_finish(struct anv_pipeline_cache *cache)
 276 {
 277    pthread_mutex_destroy(&cache->mutex);
 278
 279    if (cache->cache) {
 280       /* This is a bit unfortunate.  In order to keep things from randomly
 281        * going away, the shader cache has to hold a reference to all shader
 282        * binaries it contains.  We unref them when we destroy the cache.
 283        */
 284       hash_table_foreach(cache->cache, entry)
 285          anv_shader_bin_unref(cache->device, entry->data);
 286
 287       _mesa_hash_table_destroy(cache->cache, NULL);
 288    }
 289
 290    if (cache->nir_cache) {
 291       hash_table_foreach(cache->nir_cache, entry)
 292          ralloc_free(entry->data);
 293
 294       _mesa_hash_table_destroy(cache->nir_cache, NULL);
 295    }
 296 }
 297
 298 static struct anv_shader_bin *
 299 anv_pipeline_cache_search_locked(struct anv_pipeline_cache *cache,
 300                                  const void *key_data, uint32_t key_size)
 301 {
 302    uint32_t vla[1 + DIV_ROUND_UP(key_size, sizeof(uint32_t))];
 303    struct anv_shader_bin_key *key = (void *)vla;
 304    key->size = key_size;
 305    memcpy(key->data, key_data, key_size);
 306
 307    struct hash_entry *entry = _mesa_hash_table_search(cache->cache, key);
 308    if (entry)
 309       return entry->data;
 310    else
 311       return NULL;
 312 }
 313
 314 struct anv_shader_bin *
 315 anv_pipeline_cache_search(struct anv_pipeline_cache *cache,
 316                           const void *key_data, uint32_t key_size)
 317 {
 318    if (!cache->cache)
 319       return NULL;
 320
 321    pthread_mutex_lock(&cache->mutex);
 322
 323    struct anv_shader_bin *shader =
 324       anv_pipeline_cache_search_locked(cache, key_data, key_size);
 325
 326    pthread_mutex_unlock(&cache->mutex);
 327
 328    /* We increment refcount before handing it to the caller */
 329    if (shader)
 330       anv_shader_bin_ref(shader);
 331
 332    return shader;
 333 }
 334
 335 static void
 336 anv_pipeline_cache_add_shader_bin(struct anv_pipeline_cache *cache,
 337                                   struct anv_shader_bin *bin)
 338 {
 339    if (!cache->cache)
 340       return;
 341
 342    pthread_mutex_lock(&cache->mutex);
 343
 344    struct hash_entry *entry = _mesa_hash_table_search(cache->cache, bin->key);
 345    if (entry == NULL) {
 346       /* Take a reference for the cache */
 347       anv_shader_bin_ref(bin);
 348       _mesa_hash_table_insert(cache->cache, bin->key, bin);
 349    }
 350
 351    pthread_mutex_unlock(&cache->mutex);
 352 }
 353
 354 static struct anv_shader_bin *
 355 anv_pipeline_cache_add_shader_locked(struct anv_pipeline_cache *cache,
 356                                      const void *key_data, uint32_t key_size,
 357                                      const void *kernel_data,
 358                                      uint32_t kernel_size,
 359                                      const void *constant_data,
 360                                      uint32_t constant_data_size,
 361                                      const struct brw_stage_prog_data *prog_data,
 362                                      uint32_t prog_data_size,
 363                                      const void *prog_data_param,
 364                                      const nir_xfb_info *xfb_info,
 365                                      const struct anv_pipeline_bind_map *bind_map)
 366 {
 367    struct anv_shader_bin *shader =
 368       anv_pipeline_cache_search_locked(cache, key_data, key_size);
 369    if (shader)
 370       return shader;
 371
 372    struct anv_shader_bin *bin =
 373       anv_shader_bin_create(cache->device, key_data, key_size,
 374                             kernel_data, kernel_size,
 375                             constant_data, constant_data_size,
 376                             prog_data, prog_data_size, prog_data_param,
 377                             xfb_info, bind_map);
 378    if (!bin)
 379       return NULL;
 380
 381    _mesa_hash_table_insert(cache->cache, bin->key, bin);
 382
 383    return bin;
 384 }
 385
 386 struct anv_shader_bin *
 387 anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache,
 388                                  const void *key_data, uint32_t key_size,
 389                                  const void *kernel_data, uint32_t kernel_size,
 390                                  const void *constant_data,
 391                                  uint32_t constant_data_size,
 392                                  const struct brw_stage_prog_data *prog_data,
 393                                  uint32_t prog_data_size,
 394                                  const nir_xfb_info *xfb_info,
 395                                  const struct anv_pipeline_bind_map *bind_map)
 396 {
 397    if (cache->cache) {
 398       pthread_mutex_lock(&cache->mutex);
 399
 400       struct anv_shader_bin *bin =
 401          anv_pipeline_cache_add_shader_locked(cache, key_data, key_size,
 402                                               kernel_data, kernel_size,
 403                                               constant_data, constant_data_size,
 404                                               prog_data, prog_data_size,
 405                                               prog_data->param,
 406                                               xfb_info, bind_map);
 407
 408       pthread_mutex_unlock(&cache->mutex);
 409
 410       /* We increment refcount before handing it to the caller */
 411       if (bin)
 412          anv_shader_bin_ref(bin);
 413
 414       return bin;
 415    } else {
 416       /* In this case, we're not caching it so the caller owns it entirely */
 417       return anv_shader_bin_create(cache->device, key_data, key_size,
 418                                    kernel_data, kernel_size,
 419                                    constant_data, constant_data_size,
 420                                    prog_data, prog_data_size,
 421                                    prog_data->param,
 422                                    xfb_info, bind_map);
 423    }
 424 }
 425
 426 struct cache_header {
 427    uint32_t header_size;
 428    uint32_t header_version;
 429    uint32_t vendor_id;
 430    uint32_t device_id;
 431    uint8_t  uuid[VK_UUID_SIZE];
 432 };
 433
 434 static void
 435 anv_pipeline_cache_load(struct anv_pipeline_cache *cache,
 436                         const void *data, size_t size)
 437 {
 438    struct anv_device *device = cache->device;
 439    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
 440
 441    if (cache->cache == NULL)
 442       return;
 443
 444    struct blob_reader blob;
 445    blob_reader_init(&blob, data, size);
 446
 447    struct cache_header header;
 448    blob_copy_bytes(&blob, &header, sizeof(header));
 449    uint32_t count = blob_read_uint32(&blob);
 450    if (blob.overrun)
 451       return;
 452
 453    if (header.header_size < sizeof(header))
 454       return;
 455    if (header.header_version != VK_PIPELINE_CACHE_HEADER_VERSION_ONE)
 456       return;
 457    if (header.vendor_id != 0x8086)
 458       return;
 459    if (header.device_id != device->chipset_id)
 460       return;
 461    if (memcmp(header.uuid, pdevice->pipeline_cache_uuid, VK_UUID_SIZE) != 0)
 462       return;
 463
 464    for (uint32_t i = 0; i < count; i++) {
 465       struct anv_shader_bin *bin =
 466          anv_shader_bin_create_from_blob(device, &blob);
 467       if (!bin)
 468          break;
 469       _mesa_hash_table_insert(cache->cache, bin->key, bin);
 470    }
 471 }
 472
 473 VkResult anv_CreatePipelineCache(
 474     VkDevice                                    _device,
 475     const VkPipelineCacheCreateInfo*            pCreateInfo,
 476     const VkAllocationCallbacks*                pAllocator,
 477     VkPipelineCache*                            pPipelineCache)
 478 {
 479    ANV_FROM_HANDLE(anv_device, device, _device);
 480    struct anv_pipeline_cache *cache;
 481
 482    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
 483    assert(pCreateInfo->flags == 0);
 484
 485    cache = vk_alloc2(&device->alloc, pAllocator,
 486                        sizeof(*cache), 8,
 487                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 488    if (cache == NULL)
 489       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 490
 491    anv_pipeline_cache_init(cache, device,
 492                            device->instance->pipeline_cache_enabled);
 493
 494    if (pCreateInfo->initialDataSize > 0)
 495       anv_pipeline_cache_load(cache,
 496                               pCreateInfo->pInitialData,
 497                               pCreateInfo->initialDataSize);
 498
 499    *pPipelineCache = anv_pipeline_cache_to_handle(cache);
 500
 501    return VK_SUCCESS;
 502 }
 503
 504 void anv_DestroyPipelineCache(
 505     VkDevice                                    _device,
 506     VkPipelineCache                             _cache,
 507     const VkAllocationCallbacks*                pAllocator)
 508 {
 509    ANV_FROM_HANDLE(anv_device, device, _device);
 510    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 511
 512    if (!cache)
 513       return;
 514
 515    anv_pipeline_cache_finish(cache);
 516
 517    vk_free2(&device->alloc, pAllocator, cache);
 518 }
 519
 520 VkResult anv_GetPipelineCacheData(
 521     VkDevice                                    _device,
 522     VkPipelineCache                             _cache,
 523     size_t*                                     pDataSize,
 524     void*                                       pData)
 525 {
 526    ANV_FROM_HANDLE(anv_device, device, _device);
 527    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 528    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
 529
 530    struct blob blob;
 531    if (pData) {
 532       blob_init_fixed(&blob, pData, *pDataSize);
 533    } else {
 534       blob_init_fixed(&blob, NULL, SIZE_MAX);
 535    }
 536
 537    struct cache_header header = {
 538       .header_size = sizeof(struct cache_header),
 539       .header_version = VK_PIPELINE_CACHE_HEADER_VERSION_ONE,
 540       .vendor_id = 0x8086,
 541       .device_id = device->chipset_id,
 542    };
 543    memcpy(header.uuid, pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
 544    blob_write_bytes(&blob, &header, sizeof(header));
 545
 546    uint32_t count = 0;
 547    intptr_t count_offset = blob_reserve_uint32(&blob);
 548    if (count_offset < 0) {
 549       *pDataSize = 0;
 550       blob_finish(&blob);
 551       return VK_INCOMPLETE;
 552    }
 553
 554    VkResult result = VK_SUCCESS;
 555    if (cache->cache) {
 556       hash_table_foreach(cache->cache, entry) {
 557          struct anv_shader_bin *shader = entry->data;
 558
 559          size_t save_size = blob.size;
 560          if (!anv_shader_bin_write_to_blob(shader, &blob)) {
 561             /* If it fails reset to the previous size and bail */
 562             blob.size = save_size;
 563             result = VK_INCOMPLETE;
 564             break;
 565          }
 566
 567          count++;
 568       }
 569    }
 570
 571    blob_overwrite_uint32(&blob, count_offset, count);
 572
 573    *pDataSize = blob.size;
 574
 575    blob_finish(&blob);
 576
 577    return result;
 578 }
 579
 580 VkResult anv_MergePipelineCaches(
 581     VkDevice                                    _device,
 582     VkPipelineCache                             destCache,
 583     uint32_t                                    srcCacheCount,
 584     const VkPipelineCache*                      pSrcCaches)
 585 {
 586    ANV_FROM_HANDLE(anv_pipeline_cache, dst, destCache);
 587
 588    if (!dst->cache)
 589       return VK_SUCCESS;
 590
 591    for (uint32_t i = 0; i < srcCacheCount; i++) {
 592       ANV_FROM_HANDLE(anv_pipeline_cache, src, pSrcCaches[i]);
 593       if (!src->cache)
 594          continue;
 595
 596       hash_table_foreach(src->cache, entry) {
 597          struct anv_shader_bin *bin = entry->data;
 598          assert(bin);
 599
 600          if (_mesa_hash_table_search(dst->cache, bin->key))
 601             continue;
 602
 603          anv_shader_bin_ref(bin);
 604          _mesa_hash_table_insert(dst->cache, bin->key, bin);
 605       }
 606    }
 607
 608    return VK_SUCCESS;
 609 }
 610
 611 struct anv_shader_bin *
 612 anv_device_search_for_kernel(struct anv_device *device,
 613                              struct anv_pipeline_cache *cache,
 614                              const void *key_data, uint32_t key_size)
 615 {
 616    struct anv_shader_bin *bin;
 617
 618    if (cache) {
 619       bin = anv_pipeline_cache_search(cache, key_data, key_size);
 620       if (bin)
 621          return bin;
 622    }
 623
 624 #ifdef ENABLE_SHADER_CACHE
 625    struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache;
 626    if (disk_cache && device->instance->pipeline_cache_enabled) {
 627       cache_key cache_key;
 628       disk_cache_compute_key(disk_cache, key_data, key_size, cache_key);
 629
 630       size_t buffer_size;
 631       uint8_t *buffer = disk_cache_get(disk_cache, cache_key, &buffer_size);
 632       if (buffer) {
 633          struct blob_reader blob;
 634          blob_reader_init(&blob, buffer, buffer_size);
 635          bin = anv_shader_bin_create_from_blob(device, &blob);
 636          free(buffer);
 637
 638          if (bin) {
 639             if (cache)
 640                anv_pipeline_cache_add_shader_bin(cache, bin);
 641             return bin;
 642          }
 643       }
 644    }
 645 #endif
 646
 647    return NULL;
 648 }
 649
 650 struct anv_shader_bin *
 651 anv_device_upload_kernel(struct anv_device *device,
 652                          struct anv_pipeline_cache *cache,
 653                          const void *key_data, uint32_t key_size,
 654                          const void *kernel_data, uint32_t kernel_size,
 655                          const void *constant_data,
 656                          uint32_t constant_data_size,
 657                          const struct brw_stage_prog_data *prog_data,
 658                          uint32_t prog_data_size,
 659                          const nir_xfb_info *xfb_info,
 660                          const struct anv_pipeline_bind_map *bind_map)
 661 {
 662    struct anv_shader_bin *bin;
 663    if (cache) {
 664       bin = anv_pipeline_cache_upload_kernel(cache, key_data, key_size,
 665                                              kernel_data, kernel_size,
 666                                              constant_data, constant_data_size,
 667                                              prog_data, prog_data_size,
 668                                              xfb_info, bind_map);
 669    } else {
 670       bin = anv_shader_bin_create(device, key_data, key_size,
 671                                   kernel_data, kernel_size,
 672                                   constant_data, constant_data_size,
 673                                   prog_data, prog_data_size,
 674                                   prog_data->param,
 675                                   xfb_info, bind_map);
 676    }
 677
 678    if (bin == NULL)
 679       return NULL;
 680
 681 #ifdef ENABLE_SHADER_CACHE
 682    struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache;
 683    if (disk_cache) {
 684       struct blob binary;
 685       blob_init(&binary);
 686       if (anv_shader_bin_write_to_blob(bin, &binary)) {
 687          cache_key cache_key;
 688          disk_cache_compute_key(disk_cache, key_data, key_size, cache_key);
 689
 690          disk_cache_put(disk_cache, cache_key, binary.data, binary.size, NULL);
 691       }
 692
 693       blob_finish(&binary);
 694    }
 695 #endif
 696
 697    return bin;
 698 }
 699
 700 struct serialized_nir {
 701    unsigned char sha1_key[20];
 702    size_t size;
 703    char data[0];
 704 };
 705
 706 struct nir_shader *
 707 anv_device_search_for_nir(struct anv_device *device,
 708                           struct anv_pipeline_cache *cache,
 709                           const nir_shader_compiler_options *nir_options,
 710                           unsigned char sha1_key[20],
 711                           void *mem_ctx)
 712 {
 713    if (cache && cache->nir_cache) {
 714       const struct serialized_nir *snir = NULL;
 715
 716       pthread_mutex_lock(&cache->mutex);
 717       struct hash_entry *entry =
 718          _mesa_hash_table_search(cache->nir_cache, sha1_key);
 719       if (entry)
 720          snir = entry->data;
 721       pthread_mutex_unlock(&cache->mutex);
 722
 723       if (snir) {
 724          struct blob_reader blob;
 725          blob_reader_init(&blob, snir->data, snir->size);
 726
 727          nir_shader *nir = nir_deserialize(mem_ctx, nir_options, &blob);
 728          if (blob.overrun) {
 729             ralloc_free(nir);
 730          } else {
 731             return nir;
 732          }
 733       }
 734    }
 735
 736    return NULL;
 737 }
 738
 739 void
 740 anv_device_upload_nir(struct anv_device *device,
 741                       struct anv_pipeline_cache *cache,
 742                       const struct nir_shader *nir,
 743                       unsigned char sha1_key[20])
 744 {
 745    if (cache && cache->nir_cache) {
 746       pthread_mutex_lock(&cache->mutex);
 747       struct hash_entry *entry =
 748          _mesa_hash_table_search(cache->nir_cache, sha1_key);
 749       pthread_mutex_unlock(&cache->mutex);
 750       if (entry)
 751          return;
 752
 753       struct blob blob;
 754       blob_init(&blob);
 755
 756       nir_serialize(&blob, nir);
 757       if (blob.out_of_memory) {
 758          blob_finish(&blob);
 759          return;
 760       }
 761
 762       pthread_mutex_lock(&cache->mutex);
 763       /* Because ralloc isn't thread-safe, we have to do all this inside the
 764        * lock.  We could unlock for the big memcpy but it's probably not worth
 765        * the hassle.
 766        */
 767       entry = _mesa_hash_table_search(cache->nir_cache, sha1_key);
 768       if (entry) {
 769          blob_finish(&blob);
 770          pthread_mutex_unlock(&cache->mutex);
 771          return;
 772       }
 773
 774       struct serialized_nir *snir =
 775          ralloc_size(cache->nir_cache, sizeof(*snir) + blob.size);
 776       memcpy(snir->sha1_key, sha1_key, 20);
 777       snir->size = blob.size;
 778       memcpy(snir->data, blob.data, blob.size);
 779
 780       blob_finish(&blob);
 781
 782       _mesa_hash_table_insert(cache->nir_cache, snir->sha1_key, snir);
 783
 784       pthread_mutex_unlock(&cache->mutex);
 785    }
 786 }