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 "anv_private.h"
30 * - Compact binding table layout so it's tight and not dependent on
31 * descriptor set layout.
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
41 anv_pipeline_cache_init(struct anv_pipeline_cache
*cache
,
42 struct anv_device
*device
)
44 cache
->device
= device
;
45 anv_state_stream_init(&cache
->program_stream
,
46 &device
->instruction_block_pool
);
47 pthread_mutex_init(&cache
->mutex
, NULL
);
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
);
55 /* We don't consider allocation failure fatal, we just start with a 0-sized
57 if (cache
->hash_table
== NULL
)
58 cache
->table_size
= 0;
60 memset(cache
->hash_table
, 0xff, byte_size
);
64 anv_pipeline_cache_finish(struct anv_pipeline_cache
*cache
)
66 anv_state_stream_finish(&cache
->program_stream
);
67 pthread_mutex_destroy(&cache
->mutex
);
68 free(cache
->hash_table
);
72 unsigned char sha1
[20];
73 uint32_t prog_data_size
;
77 /* kernel follows prog_data at next 64 byte aligned address */
81 entry_size(struct cache_entry
*entry
)
83 /* This returns the number of bytes needed to serialize an entry, which
84 * doesn't include the alignment padding bytes.
87 return sizeof(*entry
) + entry
->prog_data_size
+ entry
->kernel_size
;
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
)
96 struct mesa_sha1
*ctx
;
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? */
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
);
108 _mesa_sha1_final(ctx
, hash
);
112 anv_pipeline_cache_search(struct anv_pipeline_cache
*cache
,
113 const unsigned char *sha1
,
114 const struct brw_stage_prog_data
**prog_data
)
116 const uint32_t mask
= cache
->table_size
- 1;
117 const uint32_t start
= (*(uint32_t *) sha1
);
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
];
126 struct cache_entry
*entry
=
127 cache
->program_stream
.block_pool
->map
+ offset
;
128 if (memcmp(entry
->sha1
, sha1
, sizeof(entry
->sha1
)) == 0) {
130 *prog_data
= (const struct brw_stage_prog_data
*) entry
->prog_data
;
132 const uint32_t preamble_size
=
133 align_u32(sizeof(*entry
) + entry
->prog_data_size
, 64);
135 return offset
+ preamble_size
;
143 anv_pipeline_cache_set_entry(struct anv_pipeline_cache
*cache
,
144 struct cache_entry
*entry
, uint32_t entry_offset
)
146 const uint32_t mask
= cache
->table_size
- 1;
147 const uint32_t start
= (*(uint32_t *) entry
->sha1
);
149 /* We'll always be able to insert when we get here. */
150 assert(cache
->kernel_count
< cache
->table_size
/ 2);
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
;
160 cache
->total_size
+= entry_size(entry
);
161 cache
->kernel_count
++;
165 anv_pipeline_cache_grow(struct anv_pipeline_cache
*cache
)
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]);
171 uint32_t *old_table
= cache
->hash_table
;
173 table
= malloc(byte_size
);
175 return VK_ERROR_OUT_OF_HOST_MEMORY
;
177 cache
->hash_table
= table
;
178 cache
->table_size
= table_size
;
179 cache
->kernel_count
= 0;
180 cache
->total_size
= 0;
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
];
188 struct cache_entry
*entry
=
189 cache
->program_stream
.block_pool
->map
+ offset
;
190 anv_pipeline_cache_set_entry(cache
, entry
, offset
);
199 anv_pipeline_cache_add_entry(struct anv_pipeline_cache
*cache
,
200 struct cache_entry
*entry
, uint32_t entry_offset
)
202 if (cache
->kernel_count
== cache
->table_size
/ 2)
203 anv_pipeline_cache_grow(cache
);
205 /* Failing to grow that hash table isn't fatal, but may mean we don't
206 * have enough space to add this new kernel. Only add it if there's room.
208 if (cache
->kernel_count
< cache
->table_size
/ 2)
209 anv_pipeline_cache_set_entry(cache
, entry
, entry_offset
);
213 anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache
*cache
,
214 const unsigned char *sha1
,
215 const void *kernel
, size_t kernel_size
,
216 const struct brw_stage_prog_data
**prog_data
,
217 size_t prog_data_size
)
219 pthread_mutex_lock(&cache
->mutex
);
220 struct cache_entry
*entry
;
222 const uint32_t preamble_size
=
223 align_u32(sizeof(*entry
) + prog_data_size
, 64);
225 const uint32_t size
= preamble_size
+ kernel_size
;
227 assert(size
< cache
->program_stream
.block_pool
->block_size
);
228 const struct anv_state state
=
229 anv_state_stream_alloc(&cache
->program_stream
, size
, 64);
232 entry
->prog_data_size
= prog_data_size
;
233 memcpy(entry
->prog_data
, *prog_data
, prog_data_size
);
234 *prog_data
= (const struct brw_stage_prog_data
*) entry
->prog_data
;
235 entry
->kernel_size
= kernel_size
;
237 if (sha1
&& env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", false)) {
238 assert(anv_pipeline_cache_search(cache
, sha1
, NULL
) == NO_KERNEL
);
240 memcpy(entry
->sha1
, sha1
, sizeof(entry
->sha1
));
241 anv_pipeline_cache_add_entry(cache
, entry
, state
.offset
);
244 pthread_mutex_unlock(&cache
->mutex
);
246 memcpy(state
.map
+ preamble_size
, kernel
, kernel_size
);
248 if (!cache
->device
->info
.has_llc
)
249 anv_state_clflush(state
);
251 return state
.offset
+ preamble_size
;
254 struct cache_header
{
255 uint32_t header_size
;
256 uint32_t header_version
;
259 uint8_t uuid
[VK_UUID_SIZE
];
263 anv_pipeline_cache_load(struct anv_pipeline_cache
*cache
,
264 const void *data
, size_t size
)
266 struct anv_device
*device
= cache
->device
;
267 struct cache_header header
;
268 uint8_t uuid
[VK_UUID_SIZE
];
270 if (size
< sizeof(header
))
272 memcpy(&header
, data
, sizeof(header
));
273 if (header
.header_size
< sizeof(header
))
275 if (header
.header_version
!= VK_PIPELINE_CACHE_HEADER_VERSION_ONE
)
277 if (header
.vendor_id
!= 0x8086)
279 if (header
.device_id
!= device
->chipset_id
)
281 anv_device_get_cache_uuid(uuid
);
282 if (memcmp(header
.uuid
, uuid
, VK_UUID_SIZE
) != 0)
285 const void *end
= data
+ size
;
286 const void *p
= data
+ header
.header_size
;
289 /* The kernels aren't 64 byte aligned in the serialized format so
290 * they're always right after the prog_data.
292 const struct cache_entry
*entry
= p
;
293 const void *kernel
= &entry
->prog_data
[entry
->prog_data_size
];
295 const struct brw_stage_prog_data
*prog_data
=
296 (const struct brw_stage_prog_data
*) entry
->prog_data
;
298 anv_pipeline_cache_upload_kernel(cache
, entry
->sha1
,
299 kernel
, entry
->kernel_size
,
301 entry
->prog_data_size
);
302 p
= kernel
+ entry
->kernel_size
;
306 VkResult
anv_CreatePipelineCache(
308 const VkPipelineCacheCreateInfo
* pCreateInfo
,
309 const VkAllocationCallbacks
* pAllocator
,
310 VkPipelineCache
* pPipelineCache
)
312 ANV_FROM_HANDLE(anv_device
, device
, _device
);
313 struct anv_pipeline_cache
*cache
;
315 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
);
316 assert(pCreateInfo
->flags
== 0);
318 cache
= anv_alloc2(&device
->alloc
, pAllocator
,
320 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
322 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
324 anv_pipeline_cache_init(cache
, device
);
326 if (pCreateInfo
->initialDataSize
> 0)
327 anv_pipeline_cache_load(cache
,
328 pCreateInfo
->pInitialData
,
329 pCreateInfo
->initialDataSize
);
331 *pPipelineCache
= anv_pipeline_cache_to_handle(cache
);
336 void anv_DestroyPipelineCache(
338 VkPipelineCache _cache
,
339 const VkAllocationCallbacks
* pAllocator
)
341 ANV_FROM_HANDLE(anv_device
, device
, _device
);
342 ANV_FROM_HANDLE(anv_pipeline_cache
, cache
, _cache
);
344 anv_pipeline_cache_finish(cache
);
346 anv_free2(&device
->alloc
, pAllocator
, cache
);
349 VkResult
anv_GetPipelineCacheData(
351 VkPipelineCache _cache
,
355 ANV_FROM_HANDLE(anv_device
, device
, _device
);
356 ANV_FROM_HANDLE(anv_pipeline_cache
, cache
, _cache
);
357 struct cache_header
*header
;
359 const size_t size
= sizeof(*header
) + cache
->total_size
;
366 if (*pDataSize
< sizeof(*header
)) {
368 return VK_INCOMPLETE
;
371 void *p
= pData
, *end
= pData
+ *pDataSize
;
373 header
->header_size
= sizeof(*header
);
374 header
->header_version
= VK_PIPELINE_CACHE_HEADER_VERSION_ONE
;
375 header
->vendor_id
= 0x8086;
376 header
->device_id
= device
->chipset_id
;
377 anv_device_get_cache_uuid(header
->uuid
);
378 p
+= header
->header_size
;
380 struct cache_entry
*entry
;
381 for (uint32_t i
= 0; i
< cache
->table_size
; i
++) {
382 if (cache
->hash_table
[i
] == ~0)
385 entry
= cache
->program_stream
.block_pool
->map
+ cache
->hash_table
[i
];
386 if (end
< p
+ entry_size(entry
))
389 memcpy(p
, entry
, sizeof(*entry
) + entry
->prog_data_size
);
390 p
+= sizeof(*entry
) + entry
->prog_data_size
;
392 void *kernel
= (void *) entry
+
393 align_u32(sizeof(*entry
) + entry
->prog_data_size
, 64);
395 memcpy(p
, kernel
, entry
->kernel_size
);
396 p
+= entry
->kernel_size
;
399 *pDataSize
= p
- pData
;
405 anv_pipeline_cache_merge(struct anv_pipeline_cache
*dst
,
406 struct anv_pipeline_cache
*src
)
408 for (uint32_t i
= 0; i
< src
->table_size
; i
++) {
409 if (src
->hash_table
[i
] == ~0)
412 struct cache_entry
*entry
=
413 src
->program_stream
.block_pool
->map
+ src
->hash_table
[i
];
415 if (anv_pipeline_cache_search(dst
, entry
->sha1
, NULL
) != NO_KERNEL
)
418 const void *kernel
= (void *) entry
+
419 align_u32(sizeof(*entry
) + entry
->prog_data_size
, 64);
420 const struct brw_stage_prog_data
*prog_data
=
421 (const struct brw_stage_prog_data
*) entry
->prog_data
;
423 anv_pipeline_cache_upload_kernel(dst
, entry
->sha1
,
424 kernel
, entry
->kernel_size
,
425 &prog_data
, entry
->prog_data_size
);
429 VkResult
anv_MergePipelineCaches(
431 VkPipelineCache destCache
,
432 uint32_t srcCacheCount
,
433 const VkPipelineCache
* pSrcCaches
)
435 ANV_FROM_HANDLE(anv_pipeline_cache
, dst
, destCache
);
437 for (uint32_t i
= 0; i
< srcCacheCount
; i
++) {
438 ANV_FROM_HANDLE(anv_pipeline_cache
, src
, pSrcCaches
[i
]);
440 anv_pipeline_cache_merge(dst
, src
);