2 * Copyright © 2015 Intel Corporation
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:
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
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
24 #include "util/mesa-sha1.h"
25 #include "util/debug.h"
26 #include "util/u_atomic.h"
27 #include "radv_debug.h"
28 #include "radv_private.h"
30 #include "ac_nir_to_llvm.h"
34 unsigned char sha1
[20];
38 struct ac_shader_variant_info variant_info
;
39 struct ac_shader_config config
;
40 uint32_t rsrc1
, rsrc2
;
41 struct radv_shader_variant
*variant
;
46 radv_pipeline_cache_init(struct radv_pipeline_cache
*cache
,
47 struct radv_device
*device
)
49 cache
->device
= device
;
50 pthread_mutex_init(&cache
->mutex
, NULL
);
52 cache
->modified
= false;
53 cache
->kernel_count
= 0;
54 cache
->total_size
= 0;
55 cache
->table_size
= 1024;
56 const size_t byte_size
= cache
->table_size
* sizeof(cache
->hash_table
[0]);
57 cache
->hash_table
= malloc(byte_size
);
59 /* We don't consider allocation failure fatal, we just start with a 0-sized
61 if (cache
->hash_table
== NULL
||
62 (device
->debug_flags
& RADV_DEBUG_NO_CACHE
))
63 cache
->table_size
= 0;
65 memset(cache
->hash_table
, 0, byte_size
);
69 radv_pipeline_cache_finish(struct radv_pipeline_cache
*cache
)
71 for (unsigned i
= 0; i
< cache
->table_size
; ++i
)
72 if (cache
->hash_table
[i
]) {
73 if (cache
->hash_table
[i
]->variant
)
74 radv_shader_variant_destroy(cache
->device
,
75 cache
->hash_table
[i
]->variant
);
76 vk_free(&cache
->alloc
, cache
->hash_table
[i
]);
78 pthread_mutex_destroy(&cache
->mutex
);
79 free(cache
->hash_table
);
83 entry_size(struct cache_entry
*entry
)
85 return sizeof(*entry
) + entry
->code_size
;
89 radv_hash_shader(unsigned char *hash
, struct radv_shader_module
*module
,
90 const char *entrypoint
,
91 const VkSpecializationInfo
*spec_info
,
92 const struct radv_pipeline_layout
*layout
,
93 const struct ac_shader_variant_key
*key
,
94 uint32_t is_geom_copy_shader
)
98 _mesa_sha1_init(&ctx
);
100 _mesa_sha1_update(&ctx
, key
, sizeof(*key
));
101 _mesa_sha1_update(&ctx
, module
->sha1
, sizeof(module
->sha1
));
102 _mesa_sha1_update(&ctx
, entrypoint
, strlen(entrypoint
));
104 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
106 _mesa_sha1_update(&ctx
, spec_info
->pMapEntries
,
107 spec_info
->mapEntryCount
* sizeof spec_info
->pMapEntries
[0]);
108 _mesa_sha1_update(&ctx
, spec_info
->pData
, spec_info
->dataSize
);
110 _mesa_sha1_update(&ctx
, &is_geom_copy_shader
, 4);
111 _mesa_sha1_final(&ctx
, hash
);
115 static struct cache_entry
*
116 radv_pipeline_cache_search_unlocked(struct radv_pipeline_cache
*cache
,
117 const unsigned char *sha1
)
119 const uint32_t mask
= cache
->table_size
- 1;
120 const uint32_t start
= (*(uint32_t *) sha1
);
122 if (cache
->table_size
== 0)
125 for (uint32_t i
= 0; i
< cache
->table_size
; i
++) {
126 const uint32_t index
= (start
+ i
) & mask
;
127 struct cache_entry
*entry
= cache
->hash_table
[index
];
132 if (memcmp(entry
->sha1
, sha1
, sizeof(entry
->sha1
)) == 0) {
137 unreachable("hash table should never be full");
140 static struct cache_entry
*
141 radv_pipeline_cache_search(struct radv_pipeline_cache
*cache
,
142 const unsigned char *sha1
)
144 struct cache_entry
*entry
;
146 pthread_mutex_lock(&cache
->mutex
);
148 entry
= radv_pipeline_cache_search_unlocked(cache
, sha1
);
150 pthread_mutex_unlock(&cache
->mutex
);
155 struct radv_shader_variant
*
156 radv_create_shader_variant_from_pipeline_cache(struct radv_device
*device
,
157 struct radv_pipeline_cache
*cache
,
158 const unsigned char *sha1
)
160 struct cache_entry
*entry
= NULL
;
163 entry
= radv_pipeline_cache_search(cache
, sha1
);
168 if (!entry
->variant
) {
169 struct radv_shader_variant
*variant
;
171 variant
= calloc(1, sizeof(struct radv_shader_variant
));
175 variant
->code_size
= entry
->code_size
;
176 variant
->config
= entry
->config
;
177 variant
->info
= entry
->variant_info
;
178 variant
->rsrc1
= entry
->rsrc1
;
179 variant
->rsrc2
= entry
->rsrc2
;
180 variant
->code_size
= entry
->code_size
;
181 variant
->ref_count
= 1;
183 void *ptr
= radv_alloc_shader_memory(device
, variant
);
184 memcpy(ptr
, entry
->code
, entry
->code_size
);
186 entry
->variant
= variant
;
189 p_atomic_inc(&entry
->variant
->ref_count
);
190 return entry
->variant
;
195 radv_pipeline_cache_set_entry(struct radv_pipeline_cache
*cache
,
196 struct cache_entry
*entry
)
198 const uint32_t mask
= cache
->table_size
- 1;
199 const uint32_t start
= entry
->sha1_dw
[0];
201 /* We'll always be able to insert when we get here. */
202 assert(cache
->kernel_count
< cache
->table_size
/ 2);
204 for (uint32_t i
= 0; i
< cache
->table_size
; i
++) {
205 const uint32_t index
= (start
+ i
) & mask
;
206 if (!cache
->hash_table
[index
]) {
207 cache
->hash_table
[index
] = entry
;
212 cache
->total_size
+= entry_size(entry
);
213 cache
->kernel_count
++;
218 radv_pipeline_cache_grow(struct radv_pipeline_cache
*cache
)
220 const uint32_t table_size
= cache
->table_size
* 2;
221 const uint32_t old_table_size
= cache
->table_size
;
222 const size_t byte_size
= table_size
* sizeof(cache
->hash_table
[0]);
223 struct cache_entry
**table
;
224 struct cache_entry
**old_table
= cache
->hash_table
;
226 table
= malloc(byte_size
);
228 return VK_ERROR_OUT_OF_HOST_MEMORY
;
230 cache
->hash_table
= table
;
231 cache
->table_size
= table_size
;
232 cache
->kernel_count
= 0;
233 cache
->total_size
= 0;
235 memset(cache
->hash_table
, 0, byte_size
);
236 for (uint32_t i
= 0; i
< old_table_size
; i
++) {
237 struct cache_entry
*entry
= old_table
[i
];
241 radv_pipeline_cache_set_entry(cache
, entry
);
250 radv_pipeline_cache_add_entry(struct radv_pipeline_cache
*cache
,
251 struct cache_entry
*entry
)
253 if (cache
->kernel_count
== cache
->table_size
/ 2)
254 radv_pipeline_cache_grow(cache
);
256 /* Failing to grow that hash table isn't fatal, but may mean we don't
257 * have enough space to add this new kernel. Only add it if there's room.
259 if (cache
->kernel_count
< cache
->table_size
/ 2)
260 radv_pipeline_cache_set_entry(cache
, entry
);
263 struct radv_shader_variant
*
264 radv_pipeline_cache_insert_shader(struct radv_pipeline_cache
*cache
,
265 const unsigned char *sha1
,
266 struct radv_shader_variant
*variant
,
267 const void *code
, unsigned code_size
)
272 pthread_mutex_lock(&cache
->mutex
);
273 struct cache_entry
*entry
= radv_pipeline_cache_search_unlocked(cache
, sha1
);
275 if (entry
->variant
) {
276 radv_shader_variant_destroy(cache
->device
, variant
);
277 variant
= entry
->variant
;
279 entry
->variant
= variant
;
281 p_atomic_inc(&variant
->ref_count
);
282 pthread_mutex_unlock(&cache
->mutex
);
286 entry
= vk_alloc(&cache
->alloc
, sizeof(*entry
) + code_size
, 8,
287 VK_SYSTEM_ALLOCATION_SCOPE_CACHE
);
289 pthread_mutex_unlock(&cache
->mutex
);
293 memcpy(entry
->sha1
, sha1
, 20);
294 memcpy(entry
->code
, code
, code_size
);
295 entry
->config
= variant
->config
;
296 entry
->variant_info
= variant
->info
;
297 entry
->rsrc1
= variant
->rsrc1
;
298 entry
->rsrc2
= variant
->rsrc2
;
299 entry
->code_size
= code_size
;
300 entry
->variant
= variant
;
301 p_atomic_inc(&variant
->ref_count
);
303 radv_pipeline_cache_add_entry(cache
, entry
);
305 cache
->modified
= true;
306 pthread_mutex_unlock(&cache
->mutex
);
310 struct cache_header
{
311 uint32_t header_size
;
312 uint32_t header_version
;
315 uint8_t uuid
[VK_UUID_SIZE
];
319 radv_pipeline_cache_load(struct radv_pipeline_cache
*cache
,
320 const void *data
, size_t size
)
322 struct radv_device
*device
= cache
->device
;
323 struct cache_header header
;
325 if (size
< sizeof(header
))
327 memcpy(&header
, data
, sizeof(header
));
328 if (header
.header_size
< sizeof(header
))
330 if (header
.header_version
!= VK_PIPELINE_CACHE_HEADER_VERSION_ONE
)
332 if (header
.vendor_id
!= 0x1002)
334 if (header
.device_id
!= device
->physical_device
->rad_info
.pci_id
)
336 if (memcmp(header
.uuid
, device
->physical_device
->cache_uuid
, VK_UUID_SIZE
) != 0)
339 char *end
= (void *) data
+ size
;
340 char *p
= (void *) data
+ header
.header_size
;
342 while (end
- p
>= sizeof(struct cache_entry
)) {
343 struct cache_entry
*entry
= (struct cache_entry
*)p
;
344 struct cache_entry
*dest_entry
;
345 if(end
- p
< sizeof(*entry
) + entry
->code_size
)
348 dest_entry
= vk_alloc(&cache
->alloc
, sizeof(*entry
) + entry
->code_size
,
349 8, VK_SYSTEM_ALLOCATION_SCOPE_CACHE
);
351 memcpy(dest_entry
, entry
, sizeof(*entry
) + entry
->code_size
);
352 dest_entry
->variant
= NULL
;
353 radv_pipeline_cache_add_entry(cache
, dest_entry
);
355 p
+= sizeof (*entry
) + entry
->code_size
;
359 VkResult
radv_CreatePipelineCache(
361 const VkPipelineCacheCreateInfo
* pCreateInfo
,
362 const VkAllocationCallbacks
* pAllocator
,
363 VkPipelineCache
* pPipelineCache
)
365 RADV_FROM_HANDLE(radv_device
, device
, _device
);
366 struct radv_pipeline_cache
*cache
;
368 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
);
369 assert(pCreateInfo
->flags
== 0);
371 cache
= vk_alloc2(&device
->alloc
, pAllocator
,
373 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
375 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
378 cache
->alloc
= *pAllocator
;
380 cache
->alloc
= device
->alloc
;
382 radv_pipeline_cache_init(cache
, device
);
384 if (pCreateInfo
->initialDataSize
> 0) {
385 radv_pipeline_cache_load(cache
,
386 pCreateInfo
->pInitialData
,
387 pCreateInfo
->initialDataSize
);
390 *pPipelineCache
= radv_pipeline_cache_to_handle(cache
);
395 void radv_DestroyPipelineCache(
397 VkPipelineCache _cache
,
398 const VkAllocationCallbacks
* pAllocator
)
400 RADV_FROM_HANDLE(radv_device
, device
, _device
);
401 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
405 radv_pipeline_cache_finish(cache
);
407 vk_free2(&device
->alloc
, pAllocator
, cache
);
410 VkResult
radv_GetPipelineCacheData(
412 VkPipelineCache _cache
,
416 RADV_FROM_HANDLE(radv_device
, device
, _device
);
417 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
418 struct cache_header
*header
;
419 VkResult result
= VK_SUCCESS
;
420 const size_t size
= sizeof(*header
) + cache
->total_size
;
425 if (*pDataSize
< sizeof(*header
)) {
427 return VK_INCOMPLETE
;
429 void *p
= pData
, *end
= pData
+ *pDataSize
;
431 header
->header_size
= sizeof(*header
);
432 header
->header_version
= VK_PIPELINE_CACHE_HEADER_VERSION_ONE
;
433 header
->vendor_id
= 0x1002;
434 header
->device_id
= device
->physical_device
->rad_info
.pci_id
;
435 memcpy(header
->uuid
, device
->physical_device
->cache_uuid
, VK_UUID_SIZE
);
436 p
+= header
->header_size
;
438 struct cache_entry
*entry
;
439 for (uint32_t i
= 0; i
< cache
->table_size
; i
++) {
440 if (!cache
->hash_table
[i
])
442 entry
= cache
->hash_table
[i
];
443 const uint32_t size
= entry_size(entry
);
444 if (end
< p
+ size
) {
445 result
= VK_INCOMPLETE
;
449 memcpy(p
, entry
, size
);
450 ((struct cache_entry
*)p
)->variant
= NULL
;
453 *pDataSize
= p
- pData
;
459 radv_pipeline_cache_merge(struct radv_pipeline_cache
*dst
,
460 struct radv_pipeline_cache
*src
)
462 for (uint32_t i
= 0; i
< src
->table_size
; i
++) {
463 struct cache_entry
*entry
= src
->hash_table
[i
];
464 if (!entry
|| radv_pipeline_cache_search(dst
, entry
->sha1
))
467 radv_pipeline_cache_add_entry(dst
, entry
);
469 src
->hash_table
[i
] = NULL
;
473 VkResult
radv_MergePipelineCaches(
475 VkPipelineCache destCache
,
476 uint32_t srcCacheCount
,
477 const VkPipelineCache
* pSrcCaches
)
479 RADV_FROM_HANDLE(radv_pipeline_cache
, dst
, destCache
);
481 for (uint32_t i
= 0; i
< srcCacheCount
; i
++) {
482 RADV_FROM_HANDLE(radv_pipeline_cache
, src
, pSrcCaches
[i
]);
484 radv_pipeline_cache_merge(dst
, src
);