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