src/amd/vulkan/radv_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 "util/disk_cache.h"
  27 #include "util/u_atomic.h"
  28 #include "radv_debug.h"
  29 #include "radv_private.h"
  30 #include "radv_shader.h"
  31
  32 #include "ac_nir_to_llvm.h"
  33
  34 struct cache_entry {
  35         union {
  36                 unsigned char sha1[20];
  37                 uint32_t sha1_dw[5];
  38         };
  39         uint32_t code_size;
  40         struct ac_shader_variant_info variant_info;
  41         struct ac_shader_config config;
  42         uint32_t rsrc1, rsrc2;
  43         struct radv_shader_variant *variant;
  44         uint32_t code[0];
  45 };
  46
  47 void
  48 radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
  49                          struct radv_device *device)
  50 {
  51         cache->device = device;
  52         pthread_mutex_init(&cache->mutex, NULL);
  53
  54         cache->modified = false;
  55         cache->kernel_count = 0;
  56         cache->total_size = 0;
  57         cache->table_size = 1024;
  58         const size_t byte_size = cache->table_size * sizeof(cache->hash_table[0]);
  59         cache->hash_table = malloc(byte_size);
  60
  61         /* We don't consider allocation failure fatal, we just start with a 0-sized
  62          * cache. */
  63         if (cache->hash_table == NULL ||
  64             (device->instance->debug_flags & RADV_DEBUG_NO_CACHE))
  65                 cache->table_size = 0;
  66         else
  67                 memset(cache->hash_table, 0, byte_size);
  68 }
  69
  70 void
  71 radv_pipeline_cache_finish(struct radv_pipeline_cache *cache)
  72 {
  73         for (unsigned i = 0; i < cache->table_size; ++i)
  74                 if (cache->hash_table[i]) {
  75                         if (cache->hash_table[i]->variant)
  76                                 radv_shader_variant_destroy(cache->device,
  77                                                             cache->hash_table[i]->variant);
  78                         vk_free(&cache->alloc, cache->hash_table[i]);
  79                 }
  80         pthread_mutex_destroy(&cache->mutex);
  81         free(cache->hash_table);
  82 }
  83
  84 static uint32_t
  85 entry_size(struct cache_entry *entry)
  86 {
  87         return sizeof(*entry) + entry->code_size;
  88 }
  89
  90 void
  91 radv_hash_shader(unsigned char *hash, struct radv_shader_module *module,
  92                  const char *entrypoint,
  93                  const VkSpecializationInfo *spec_info,
  94                  const struct radv_pipeline_layout *layout,
  95                  const struct ac_shader_variant_key *key,
  96                  uint32_t is_geom_copy_shader)
  97 {
  98         struct mesa_sha1 ctx;
  99
 100         _mesa_sha1_init(&ctx);
 101         if (key)
 102                 _mesa_sha1_update(&ctx, key, sizeof(*key));
 103         _mesa_sha1_update(&ctx, module->sha1, sizeof(module->sha1));
 104         _mesa_sha1_update(&ctx, entrypoint, strlen(entrypoint));
 105         if (layout)
 106                 _mesa_sha1_update(&ctx, layout->sha1, sizeof(layout->sha1));
 107         if (spec_info) {
 108                 _mesa_sha1_update(&ctx, spec_info->pMapEntries,
 109                                   spec_info->mapEntryCount * sizeof spec_info->pMapEntries[0]);
 110                 _mesa_sha1_update(&ctx, spec_info->pData, spec_info->dataSize);
 111         }
 112         _mesa_sha1_update(&ctx, &is_geom_copy_shader, 4);
 113         _mesa_sha1_final(&ctx, hash);
 114 }
 115
 116
 117 static struct cache_entry *
 118 radv_pipeline_cache_search_unlocked(struct radv_pipeline_cache *cache,
 119                                     const unsigned char *sha1)
 120 {
 121         const uint32_t mask = cache->table_size - 1;
 122         const uint32_t start = (*(uint32_t *) sha1);
 123
 124         if (cache->table_size == 0)
 125                 return NULL;
 126
 127         for (uint32_t i = 0; i < cache->table_size; i++) {
 128                 const uint32_t index = (start + i) & mask;
 129                 struct cache_entry *entry = cache->hash_table[index];
 130
 131                 if (!entry)
 132                         return NULL;
 133
 134                 if (memcmp(entry->sha1, sha1, sizeof(entry->sha1)) == 0) {
 135                         return entry;
 136                 }
 137         }
 138
 139         unreachable("hash table should never be full");
 140 }
 141
 142 static struct cache_entry *
 143 radv_pipeline_cache_search(struct radv_pipeline_cache *cache,
 144                            const unsigned char *sha1)
 145 {
 146         struct cache_entry *entry;
 147
 148         pthread_mutex_lock(&cache->mutex);
 149
 150         entry = radv_pipeline_cache_search_unlocked(cache, sha1);
 151
 152         pthread_mutex_unlock(&cache->mutex);
 153
 154         return entry;
 155 }
 156
 157 struct radv_shader_variant *
 158 radv_create_shader_variant_from_pipeline_cache(struct radv_device *device,
 159                                                struct radv_pipeline_cache *cache,
 160                                                const unsigned char *sha1)
 161 {
 162         struct cache_entry *entry = NULL;
 163
 164         if (cache)
 165                 entry = radv_pipeline_cache_search(cache, sha1);
 166         else
 167                 entry = radv_pipeline_cache_search(device->mem_cache, sha1);
 168
 169         if (!entry) {
 170                 uint8_t disk_sha1[20];
 171                 disk_cache_compute_key(device->physical_device->disk_cache,
 172                                        sha1, 20, disk_sha1);
 173                 entry = (struct cache_entry *)
 174                         disk_cache_get(device->physical_device->disk_cache,
 175                                        disk_sha1, NULL);
 176                 if (!entry)
 177                         return NULL;
 178         }
 179
 180         if (!entry->variant) {
 181                 struct radv_shader_variant *variant;
 182
 183                 variant = calloc(1, sizeof(struct radv_shader_variant));
 184                 if (!variant)
 185                         return NULL;
 186
 187                 variant->code_size = entry->code_size;
 188                 variant->config = entry->config;
 189                 variant->info = entry->variant_info;
 190                 variant->rsrc1 = entry->rsrc1;
 191                 variant->rsrc2 = entry->rsrc2;
 192                 variant->code_size = entry->code_size;
 193                 variant->ref_count = 1;
 194
 195                 void *ptr = radv_alloc_shader_memory(device, variant);
 196                 memcpy(ptr, entry->code, entry->code_size);
 197
 198                 entry->variant = variant;
 199         }
 200
 201         p_atomic_inc(&entry->variant->ref_count);
 202         return entry->variant;
 203 }
 204
 205
 206 static void
 207 radv_pipeline_cache_set_entry(struct radv_pipeline_cache *cache,
 208                               struct cache_entry *entry)
 209 {
 210         const uint32_t mask = cache->table_size - 1;
 211         const uint32_t start = entry->sha1_dw[0];
 212
 213         /* We'll always be able to insert when we get here. */
 214         assert(cache->kernel_count < cache->table_size / 2);
 215
 216         for (uint32_t i = 0; i < cache->table_size; i++) {
 217                 const uint32_t index = (start + i) & mask;
 218                 if (!cache->hash_table[index]) {
 219                         cache->hash_table[index] = entry;
 220                         break;
 221                 }
 222         }
 223
 224         cache->total_size += entry_size(entry);
 225         cache->kernel_count++;
 226 }
 227
 228
 229 static VkResult
 230 radv_pipeline_cache_grow(struct radv_pipeline_cache *cache)
 231 {
 232         const uint32_t table_size = cache->table_size * 2;
 233         const uint32_t old_table_size = cache->table_size;
 234         const size_t byte_size = table_size * sizeof(cache->hash_table[0]);
 235         struct cache_entry **table;
 236         struct cache_entry **old_table = cache->hash_table;
 237
 238         table = malloc(byte_size);
 239         if (table == NULL)
 240                 return VK_ERROR_OUT_OF_HOST_MEMORY;
 241
 242         cache->hash_table = table;
 243         cache->table_size = table_size;
 244         cache->kernel_count = 0;
 245         cache->total_size = 0;
 246
 247         memset(cache->hash_table, 0, byte_size);
 248         for (uint32_t i = 0; i < old_table_size; i++) {
 249                 struct cache_entry *entry = old_table[i];
 250                 if (!entry)
 251                         continue;
 252
 253                 radv_pipeline_cache_set_entry(cache, entry);
 254         }
 255
 256         free(old_table);
 257
 258         return VK_SUCCESS;
 259 }
 260
 261 static void
 262 radv_pipeline_cache_add_entry(struct radv_pipeline_cache *cache,
 263                               struct cache_entry *entry)
 264 {
 265         if (cache->kernel_count == cache->table_size / 2)
 266                 radv_pipeline_cache_grow(cache);
 267
 268         /* Failing to grow that hash table isn't fatal, but may mean we don't
 269          * have enough space to add this new kernel. Only add it if there's room.
 270          */
 271         if (cache->kernel_count < cache->table_size / 2)
 272                 radv_pipeline_cache_set_entry(cache, entry);
 273 }
 274
 275 struct radv_shader_variant *
 276 radv_pipeline_cache_insert_shader(struct radv_device *device,
 277                                   struct radv_pipeline_cache *cache,
 278                                   const unsigned char *sha1,
 279                                   struct radv_shader_variant *variant,
 280                                   const void *code, unsigned code_size)
 281 {
 282         if (!cache)
 283                 cache = device->mem_cache;
 284
 285         pthread_mutex_lock(&cache->mutex);
 286         struct cache_entry *entry = radv_pipeline_cache_search_unlocked(cache, sha1);
 287         if (entry) {
 288                 if (entry->variant) {
 289                         radv_shader_variant_destroy(cache->device, variant);
 290                         variant = entry->variant;
 291                 } else {
 292                         entry->variant = variant;
 293                 }
 294                 p_atomic_inc(&variant->ref_count);
 295                 pthread_mutex_unlock(&cache->mutex);
 296                 return variant;
 297         }
 298
 299         entry = vk_alloc(&cache->alloc, sizeof(*entry) + code_size, 8,
 300                            VK_SYSTEM_ALLOCATION_SCOPE_CACHE);
 301         if (!entry) {
 302                 pthread_mutex_unlock(&cache->mutex);
 303                 return variant;
 304         }
 305
 306         memcpy(entry->sha1, sha1, 20);
 307         memcpy(entry->code, code, code_size);
 308         entry->config = variant->config;
 309         entry->variant_info = variant->info;
 310         entry->rsrc1 = variant->rsrc1;
 311         entry->rsrc2 = variant->rsrc2;
 312         entry->code_size = code_size;
 313
 314         /* Set variant to NULL so we have reproducible cache items */
 315         entry->variant = NULL;
 316
 317         /* Always add cache items to disk. This will allow collection of
 318          * compiled shaders by third parties such as steam, even if the app
 319          * implements its own pipeline cache.
 320          */
 321         uint8_t disk_sha1[20];
 322         disk_cache_compute_key(device->physical_device->disk_cache, sha1, 20,
 323                                disk_sha1);
 324         disk_cache_put(device->physical_device->disk_cache,
 325                        disk_sha1, entry, entry_size(entry), NULL);
 326
 327         entry->variant = variant;
 328         p_atomic_inc(&variant->ref_count);
 329
 330         radv_pipeline_cache_add_entry(cache, entry);
 331
 332         cache->modified = true;
 333         pthread_mutex_unlock(&cache->mutex);
 334         return variant;
 335 }
 336
 337 struct cache_header {
 338         uint32_t header_size;
 339         uint32_t header_version;
 340         uint32_t vendor_id;
 341         uint32_t device_id;
 342         uint8_t  uuid[VK_UUID_SIZE];
 343 };
 344
 345 void
 346 radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
 347                          const void *data, size_t size)
 348 {
 349         struct radv_device *device = cache->device;
 350         struct cache_header header;
 351
 352         if (size < sizeof(header))
 353                 return;
 354         memcpy(&header, data, sizeof(header));
 355         if (header.header_size < sizeof(header))
 356                 return;
 357         if (header.header_version != VK_PIPELINE_CACHE_HEADER_VERSION_ONE)
 358                 return;
 359         if (header.vendor_id != ATI_VENDOR_ID)
 360                 return;
 361         if (header.device_id != device->physical_device->rad_info.pci_id)
 362                 return;
 363         if (memcmp(header.uuid, device->physical_device->cache_uuid, VK_UUID_SIZE) != 0)
 364                 return;
 365
 366         char *end = (void *) data + size;
 367         char *p = (void *) data + header.header_size;
 368
 369         while (end - p >= sizeof(struct cache_entry)) {
 370                 struct cache_entry *entry = (struct cache_entry*)p;
 371                 struct cache_entry *dest_entry;
 372                 if(end - p < sizeof(*entry) + entry->code_size)
 373                         break;
 374
 375                 dest_entry = vk_alloc(&cache->alloc, sizeof(*entry) + entry->code_size,
 376                                         8, VK_SYSTEM_ALLOCATION_SCOPE_CACHE);
 377                 if (dest_entry) {
 378                         memcpy(dest_entry, entry, sizeof(*entry) + entry->code_size);
 379                         dest_entry->variant = NULL;
 380                         radv_pipeline_cache_add_entry(cache, dest_entry);
 381                 }
 382                 p += sizeof (*entry) + entry->code_size;
 383         }
 384 }
 385
 386 VkResult radv_CreatePipelineCache(
 387         VkDevice                                    _device,
 388         const VkPipelineCacheCreateInfo*            pCreateInfo,
 389         const VkAllocationCallbacks*                pAllocator,
 390         VkPipelineCache*                            pPipelineCache)
 391 {
 392         RADV_FROM_HANDLE(radv_device, device, _device);
 393         struct radv_pipeline_cache *cache;
 394
 395         assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
 396         assert(pCreateInfo->flags == 0);
 397
 398         cache = vk_alloc2(&device->alloc, pAllocator,
 399                             sizeof(*cache), 8,
 400                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 401         if (cache == NULL)
 402                 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 403
 404         if (pAllocator)
 405                 cache->alloc = *pAllocator;
 406         else
 407                 cache->alloc = device->alloc;
 408
 409         radv_pipeline_cache_init(cache, device);
 410
 411         if (pCreateInfo->initialDataSize > 0) {
 412                 radv_pipeline_cache_load(cache,
 413                                          pCreateInfo->pInitialData,
 414                                          pCreateInfo->initialDataSize);
 415         }
 416
 417         *pPipelineCache = radv_pipeline_cache_to_handle(cache);
 418
 419         return VK_SUCCESS;
 420 }
 421
 422 void radv_DestroyPipelineCache(
 423         VkDevice                                    _device,
 424         VkPipelineCache                             _cache,
 425         const VkAllocationCallbacks*                pAllocator)
 426 {
 427         RADV_FROM_HANDLE(radv_device, device, _device);
 428         RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
 429
 430         if (!cache)
 431                 return;
 432         radv_pipeline_cache_finish(cache);
 433
 434         vk_free2(&device->alloc, pAllocator, cache);
 435 }
 436
 437 VkResult radv_GetPipelineCacheData(
 438         VkDevice                                    _device,
 439         VkPipelineCache                             _cache,
 440         size_t*                                     pDataSize,
 441         void*                                       pData)
 442 {
 443         RADV_FROM_HANDLE(radv_device, device, _device);
 444         RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
 445         struct cache_header *header;
 446         VkResult result = VK_SUCCESS;
 447         const size_t size = sizeof(*header) + cache->total_size;
 448         if (pData == NULL) {
 449                 *pDataSize = size;
 450                 return VK_SUCCESS;
 451         }
 452         if (*pDataSize < sizeof(*header)) {
 453                 *pDataSize = 0;
 454                 return VK_INCOMPLETE;
 455         }
 456         void *p = pData, *end = pData + *pDataSize;
 457         header = p;
 458         header->header_size = sizeof(*header);
 459         header->header_version = VK_PIPELINE_CACHE_HEADER_VERSION_ONE;
 460         header->vendor_id = ATI_VENDOR_ID;
 461         header->device_id = device->physical_device->rad_info.pci_id;
 462         memcpy(header->uuid, device->physical_device->cache_uuid, VK_UUID_SIZE);
 463         p += header->header_size;
 464
 465         struct cache_entry *entry;
 466         for (uint32_t i = 0; i < cache->table_size; i++) {
 467                 if (!cache->hash_table[i])
 468                         continue;
 469                 entry = cache->hash_table[i];
 470                 const uint32_t size = entry_size(entry);
 471                 if (end < p + size) {
 472                         result = VK_INCOMPLETE;
 473                         break;
 474                 }
 475
 476                 memcpy(p, entry, size);
 477                 ((struct cache_entry*)p)->variant = NULL;
 478                 p += size;
 479         }
 480         *pDataSize = p - pData;
 481
 482         return result;
 483 }
 484
 485 static void
 486 radv_pipeline_cache_merge(struct radv_pipeline_cache *dst,
 487                           struct radv_pipeline_cache *src)
 488 {
 489         for (uint32_t i = 0; i < src->table_size; i++) {
 490                 struct cache_entry *entry = src->hash_table[i];
 491                 if (!entry || radv_pipeline_cache_search(dst, entry->sha1))
 492                         continue;
 493
 494                 radv_pipeline_cache_add_entry(dst, entry);
 495
 496                 src->hash_table[i] = NULL;
 497         }
 498 }
 499
 500 VkResult radv_MergePipelineCaches(
 501         VkDevice                                    _device,
 502         VkPipelineCache                             destCache,
 503         uint32_t                                    srcCacheCount,
 504         const VkPipelineCache*                      pSrcCaches)
 505 {
 506         RADV_FROM_HANDLE(radv_pipeline_cache, dst, destCache);
 507
 508         for (uint32_t i = 0; i < srcCacheCount; i++) {
 509                 RADV_FROM_HANDLE(radv_pipeline_cache, src, pSrcCaches[i]);
 510
 511                 radv_pipeline_cache_merge(dst, src);
 512         }
 513
 514         return VK_SUCCESS;
 515 }