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/mesa-sha1.h"
  25 #include "util/debug.h"
  26 #include "anv_private.h"
  27
  28 /* Remaining work:
  29  *
  30  * - Compact binding table layout so it's tight and not dependent on
  31  *   descriptor set layout.
  32  *
  33  * - Review prog_data struct for size and cacheability: struct
  34  *   brw_stage_prog_data has binding_table which uses a lot of uint32_t for 8
  35  *   bit quantities etc; param, pull_param, and image_params are pointers, we
  36  *   just need the compation map. use bit fields for all bools, eg
  37  *   dual_src_blend.
  38  */
  39
  40 void
  41 anv_pipeline_cache_init(struct anv_pipeline_cache *cache,
  42                         struct anv_device *device)
  43 {
  44    cache->device = device;
  45    anv_state_stream_init(&cache->program_stream,
  46                          &device->instruction_block_pool);
  47    pthread_mutex_init(&cache->mutex, NULL);
  48
  49    cache->kernel_count = 0;
  50    cache->total_size = 0;
  51    cache->table_size = 1024;
  52    const size_t byte_size = cache->table_size * sizeof(cache->hash_table[0]);
  53    cache->hash_table = malloc(byte_size);
  54
  55    /* We don't consider allocation failure fatal, we just start with a 0-sized
  56     * cache. */
  57    if (cache->hash_table == NULL)
  58       cache->table_size = 0;
  59    else
  60       memset(cache->hash_table, 0xff, byte_size);
  61 }
  62
  63 void
  64 anv_pipeline_cache_finish(struct anv_pipeline_cache *cache)
  65 {
  66    anv_state_stream_finish(&cache->program_stream);
  67    pthread_mutex_destroy(&cache->mutex);
  68    free(cache->hash_table);
  69 }
  70
  71 struct cache_entry {
  72    unsigned char sha1[20];
  73    uint32_t prog_data_size;
  74    uint32_t kernel_size;
  75    char prog_data[0];
  76
  77    /* kernel follows prog_data at next 64 byte aligned address */
  78 };
  79
  80 static uint32_t
  81 entry_size(struct cache_entry *entry)
  82 {
  83    /* This returns the number of bytes needed to serialize an entry, which
  84     * doesn't include the alignment padding bytes.
  85     */
  86
  87    return sizeof(*entry) + entry->prog_data_size + entry->kernel_size;
  88 }
  89
  90 void
  91 anv_hash_shader(unsigned char *hash, const void *key, size_t key_size,
  92                 struct anv_shader_module *module,
  93                 const char *entrypoint,
  94                 const VkSpecializationInfo *spec_info)
  95 {
  96    struct mesa_sha1 *ctx;
  97
  98    ctx = _mesa_sha1_init();
  99    _mesa_sha1_update(ctx, key, key_size);
 100    _mesa_sha1_update(ctx, module->sha1, sizeof(module->sha1));
 101    _mesa_sha1_update(ctx, entrypoint, strlen(entrypoint));
 102    /* hash in shader stage, pipeline layout? */
 103    if (spec_info) {
 104       _mesa_sha1_update(ctx, spec_info->pMapEntries,
 105                         spec_info->mapEntryCount * sizeof spec_info->pMapEntries[0]);
 106       _mesa_sha1_update(ctx, spec_info->pData, spec_info->dataSize);
 107    }
 108    _mesa_sha1_final(ctx, hash);
 109 }
 110
 111 uint32_t
 112 anv_pipeline_cache_search(struct anv_pipeline_cache *cache,
 113                           const unsigned char *sha1,
 114                           const struct brw_stage_prog_data **prog_data)
 115 {
 116    const uint32_t mask = cache->table_size - 1;
 117    const uint32_t start = (*(uint32_t *) sha1);
 118
 119    for (uint32_t i = 0; i < cache->table_size; i++) {
 120       const uint32_t index = (start + i) & mask;
 121       const uint32_t offset = cache->hash_table[index];
 122
 123       if (offset == ~0)
 124          return NO_KERNEL;
 125
 126       struct cache_entry *entry =
 127          cache->program_stream.block_pool->map + offset;
 128       if (memcmp(entry->sha1, sha1, sizeof(entry->sha1)) == 0) {
 129          if (prog_data)
 130             *prog_data = (const struct brw_stage_prog_data *) entry->prog_data;
 131
 132          const uint32_t preamble_size =
 133             align_u32(sizeof(*entry) + entry->prog_data_size, 64);
 134
 135          return offset + preamble_size;
 136       }
 137    }
 138
 139    return NO_KERNEL;
 140 }
 141
 142 static void
 143 anv_pipeline_cache_add_entry(struct anv_pipeline_cache *cache,
 144                              struct cache_entry *entry, uint32_t entry_offset)
 145 {
 146    const uint32_t mask = cache->table_size - 1;
 147    const uint32_t start = (*(uint32_t *) entry->sha1);
 148
 149    /* We'll always be able to insert when we get here. */
 150    assert(cache->kernel_count < cache->table_size / 2);
 151
 152    for (uint32_t i = 0; i < cache->table_size; i++) {
 153       const uint32_t index = (start + i) & mask;
 154       if (cache->hash_table[index] == ~0) {
 155          cache->hash_table[index] = entry_offset;
 156          break;
 157       }
 158    }
 159
 160    cache->total_size += entry_size(entry);
 161    cache->kernel_count++;
 162 }
 163
 164 static VkResult
 165 anv_pipeline_cache_grow(struct anv_pipeline_cache *cache)
 166 {
 167    const uint32_t table_size = cache->table_size * 2;
 168    const uint32_t old_table_size = cache->table_size;
 169    const size_t byte_size = table_size * sizeof(cache->hash_table[0]);
 170    uint32_t *table;
 171    uint32_t *old_table = cache->hash_table;
 172
 173    table = malloc(byte_size);
 174    if (table == NULL)
 175       return VK_ERROR_OUT_OF_HOST_MEMORY;
 176
 177    cache->hash_table = table;
 178    cache->table_size = table_size;
 179    cache->kernel_count = 0;
 180    cache->total_size = 0;
 181
 182    memset(cache->hash_table, 0xff, byte_size);
 183    for (uint32_t i = 0; i < old_table_size; i++) {
 184       const uint32_t offset = old_table[i];
 185       if (offset == ~0)
 186          continue;
 187
 188       struct cache_entry *entry =
 189          cache->program_stream.block_pool->map + offset;
 190       anv_pipeline_cache_add_entry(cache, entry, offset);
 191    }
 192
 193    free(old_table);
 194
 195    return VK_SUCCESS;
 196 }
 197
 198 uint32_t
 199 anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache,
 200                                  const unsigned char *sha1,
 201                                  const void *kernel, size_t kernel_size,
 202                                  const struct brw_stage_prog_data **prog_data,
 203                                  size_t prog_data_size)
 204 {
 205    pthread_mutex_lock(&cache->mutex);
 206    struct cache_entry *entry;
 207
 208    const uint32_t preamble_size =
 209       align_u32(sizeof(*entry) + prog_data_size, 64);
 210
 211    const uint32_t size = preamble_size + kernel_size;
 212
 213    assert(size < cache->program_stream.block_pool->block_size);
 214    const struct anv_state state =
 215       anv_state_stream_alloc(&cache->program_stream, size, 64);
 216
 217    entry = state.map;
 218    entry->prog_data_size = prog_data_size;
 219    memcpy(entry->prog_data, *prog_data, prog_data_size);
 220    *prog_data = (const struct brw_stage_prog_data *) entry->prog_data;
 221    entry->kernel_size = kernel_size;
 222
 223    if (sha1 && env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", false)) {
 224       assert(anv_pipeline_cache_search(cache, sha1, NULL) == NO_KERNEL);
 225
 226       memcpy(entry->sha1, sha1, sizeof(entry->sha1));
 227       if (cache->kernel_count == cache->table_size / 2)
 228          anv_pipeline_cache_grow(cache);
 229
 230       /* Failing to grow that hash table isn't fatal, but may mean we don't
 231        * have enough space to add this new kernel. Only add it if there's room.
 232        */
 233       if (cache->kernel_count < cache->table_size / 2)
 234          anv_pipeline_cache_add_entry(cache, entry, state.offset);
 235    }
 236
 237    pthread_mutex_unlock(&cache->mutex);
 238
 239    memcpy(state.map + preamble_size, kernel, kernel_size);
 240
 241    if (!cache->device->info.has_llc)
 242       anv_state_clflush(state);
 243
 244    return state.offset + preamble_size;
 245 }
 246
 247 struct cache_header {
 248    uint32_t header_size;
 249    uint32_t header_version;
 250    uint32_t vendor_id;
 251    uint32_t device_id;
 252    uint8_t  uuid[VK_UUID_SIZE];
 253 };
 254
 255 static void
 256 anv_pipeline_cache_load(struct anv_pipeline_cache *cache,
 257                         const void *data, size_t size)
 258 {
 259    struct anv_device *device = cache->device;
 260    struct cache_header header;
 261    uint8_t uuid[VK_UUID_SIZE];
 262
 263    if (size < sizeof(header))
 264       return;
 265    memcpy(&header, data, sizeof(header));
 266    if (header.header_size < sizeof(header))
 267       return;
 268    if (header.header_version != VK_PIPELINE_CACHE_HEADER_VERSION_ONE)
 269       return;
 270    if (header.vendor_id != 0x8086)
 271       return;
 272    if (header.device_id != device->chipset_id)
 273       return;
 274    anv_device_get_cache_uuid(uuid);
 275    if (memcmp(header.uuid, uuid, VK_UUID_SIZE) != 0)
 276       return;
 277
 278    const void *end = data + size;
 279    const void *p = data + header.header_size;
 280
 281    while (p < end) {
 282       /* The kernels aren't 64 byte aligned in the serialized format so
 283        * they're always right after the prog_data.
 284        */
 285       const struct cache_entry *entry = p;
 286       const void *kernel = &entry->prog_data[entry->prog_data_size];
 287
 288       const struct brw_stage_prog_data *prog_data =
 289          (const struct brw_stage_prog_data *) entry->prog_data;
 290
 291       anv_pipeline_cache_upload_kernel(cache, entry->sha1,
 292                                        kernel, entry->kernel_size,
 293                                        &prog_data,
 294                                        entry->prog_data_size);
 295       p = kernel + entry->kernel_size;
 296    }
 297 }
 298
 299 VkResult anv_CreatePipelineCache(
 300     VkDevice                                    _device,
 301     const VkPipelineCacheCreateInfo*            pCreateInfo,
 302     const VkAllocationCallbacks*                pAllocator,
 303     VkPipelineCache*                            pPipelineCache)
 304 {
 305    ANV_FROM_HANDLE(anv_device, device, _device);
 306    struct anv_pipeline_cache *cache;
 307
 308    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
 309    assert(pCreateInfo->flags == 0);
 310
 311    cache = anv_alloc2(&device->alloc, pAllocator,
 312                        sizeof(*cache), 8,
 313                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 314    if (cache == NULL)
 315       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 316
 317    anv_pipeline_cache_init(cache, device);
 318
 319    if (pCreateInfo->initialDataSize > 0)
 320       anv_pipeline_cache_load(cache,
 321                               pCreateInfo->pInitialData,
 322                               pCreateInfo->initialDataSize);
 323
 324    *pPipelineCache = anv_pipeline_cache_to_handle(cache);
 325
 326    return VK_SUCCESS;
 327 }
 328
 329 void anv_DestroyPipelineCache(
 330     VkDevice                                    _device,
 331     VkPipelineCache                             _cache,
 332     const VkAllocationCallbacks*                pAllocator)
 333 {
 334    ANV_FROM_HANDLE(anv_device, device, _device);
 335    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 336
 337    anv_pipeline_cache_finish(cache);
 338
 339    anv_free2(&device->alloc, pAllocator, cache);
 340 }
 341
 342 VkResult anv_GetPipelineCacheData(
 343     VkDevice                                    _device,
 344     VkPipelineCache                             _cache,
 345     size_t*                                     pDataSize,
 346     void*                                       pData)
 347 {
 348    ANV_FROM_HANDLE(anv_device, device, _device);
 349    ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 350    struct cache_header *header;
 351
 352    const size_t size = sizeof(*header) + cache->total_size;
 353
 354    if (pData == NULL) {
 355       *pDataSize = size;
 356       return VK_SUCCESS;
 357    }
 358
 359    if (*pDataSize < sizeof(*header)) {
 360       *pDataSize = 0;
 361       return VK_INCOMPLETE;
 362    }
 363
 364    void *p = pData, *end = pData + *pDataSize;
 365    header = p;
 366    header->header_size = sizeof(*header);
 367    header->header_version = VK_PIPELINE_CACHE_HEADER_VERSION_ONE;
 368    header->vendor_id = 0x8086;
 369    header->device_id = device->chipset_id;
 370    anv_device_get_cache_uuid(header->uuid);
 371    p += header->header_size;
 372
 373    struct cache_entry *entry;
 374    for (uint32_t i = 0; i < cache->table_size; i++) {
 375       if (cache->hash_table[i] == ~0)
 376          continue;
 377
 378       entry = cache->program_stream.block_pool->map + cache->hash_table[i];
 379       if (end < p + entry_size(entry))
 380          break;
 381
 382       memcpy(p, entry, sizeof(*entry) + entry->prog_data_size);
 383       p += sizeof(*entry) + entry->prog_data_size;
 384
 385       void *kernel = (void *) entry +
 386          align_u32(sizeof(*entry) + entry->prog_data_size, 64);
 387
 388       memcpy(p, kernel, entry->kernel_size);
 389       p += entry->kernel_size;
 390    }
 391
 392    *pDataSize = p - pData;
 393
 394    return VK_SUCCESS;
 395 }
 396
 397 static void
 398 anv_pipeline_cache_merge(struct anv_pipeline_cache *dst,
 399                          struct anv_pipeline_cache *src)
 400 {
 401    for (uint32_t i = 0; i < src->table_size; i++) {
 402       if (src->hash_table[i] == ~0)
 403          continue;
 404
 405       struct cache_entry *entry =
 406          src->program_stream.block_pool->map + src->hash_table[i];
 407
 408       if (anv_pipeline_cache_search(dst, entry->sha1, NULL) != NO_KERNEL)
 409          continue;
 410
 411       const void *kernel = (void *) entry +
 412          align_u32(sizeof(*entry) + entry->prog_data_size, 64);
 413       const struct brw_stage_prog_data *prog_data =
 414          (const struct brw_stage_prog_data *) entry->prog_data;
 415
 416       anv_pipeline_cache_upload_kernel(dst, entry->sha1,
 417                                        kernel, entry->kernel_size,
 418                                        &prog_data, entry->prog_data_size);
 419    }
 420 }
 421
 422 VkResult anv_MergePipelineCaches(
 423     VkDevice                                    _device,
 424     VkPipelineCache                             destCache,
 425     uint32_t                                    srcCacheCount,
 426     const VkPipelineCache*                      pSrcCaches)
 427 {
 428    ANV_FROM_HANDLE(anv_pipeline_cache, dst, destCache);
 429
 430    for (uint32_t i = 0; i < srcCacheCount; i++) {
 431       ANV_FROM_HANDLE(anv_pipeline_cache, src, pSrcCaches[i]);
 432
 433       anv_pipeline_cache_merge(dst, src);
 434    }
 435
 436    return VK_SUCCESS;
 437 }