2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **********************************************************************/
29 * Keith Whitwell <keith@tungstengraphics.com>
32 /** @file brw_state_cache.c
34 * This file implements a simple static state cache for 965. The consumers
35 * can query the hash table of state using a cache_id, opaque key data,
36 * and list of buffers that will be used in relocations, and receive the
37 * corresponding state buffer object of state (plus associated auxiliary
40 * The inner workings are a simple hash table based on a CRC of the key data.
41 * The cache_id and relocation target buffers associated with the state
42 * buffer are included as auxiliary key data, but are not part of the hash
43 * value (this should be fixed, but will likely be fixed instead by making
44 * consumers use structured keys).
46 * Replacement is not implemented. Instead, when the cache gets too big, at
47 * a safe point (unlock) we throw out all of the cache data and let it
48 * regenerate for the next rendering operation.
50 * The reloc_buf pointers need to be included as key data, otherwise the
51 * non-unique values stuffed in the offset in key data through
52 * brw_cache_data() may result in successful probe for state buffers
53 * even when the buffer being referenced doesn't match. The result would be
54 * that the same state cache entry is used twice for different buffers,
55 * only one of the two buffers referenced gets put into the offset, and the
56 * incorrect program is run for the other instance.
59 #include "main/imports.h"
60 #include "brw_state.h"
61 #include "intel_batchbuffer.h"
63 /* XXX: Fixme - have to include these to get the sizes of the prog_key
74 hash_key(const void *key
, GLuint key_size
,
75 dri_bo
**reloc_bufs
, GLuint nr_reloc_bufs
)
77 GLuint
*ikey
= (GLuint
*)key
;
80 assert(key_size
% 4 == 0);
82 /* I'm sure this can be improved on:
84 for (i
= 0; i
< key_size
/4; i
++) {
86 hash
= (hash
<< 5) | (hash
>> 27);
89 /* Include the BO pointers as key data as well */
90 ikey
= (GLuint
*)reloc_bufs
;
91 key_size
= nr_reloc_bufs
* sizeof(dri_bo
*);
92 for (i
= 0; i
< key_size
/4; i
++) {
94 hash
= (hash
<< 5) | (hash
>> 27);
102 * Marks a new buffer as being chosen for the given cache id.
105 update_cache_last(struct brw_cache
*cache
, enum brw_cache_id cache_id
,
108 if (bo
== cache
->last_bo
[cache_id
])
109 return; /* no change */
111 dri_bo_unreference(cache
->last_bo
[cache_id
]);
112 cache
->last_bo
[cache_id
] = bo
;
113 dri_bo_reference(cache
->last_bo
[cache_id
]);
114 cache
->brw
->state
.dirty
.cache
|= 1 << cache_id
;
118 static struct brw_cache_item
*
119 search_cache(struct brw_cache
*cache
, enum brw_cache_id cache_id
,
120 GLuint hash
, const void *key
, GLuint key_size
,
121 dri_bo
**reloc_bufs
, GLuint nr_reloc_bufs
)
123 struct brw_cache_item
*c
;
128 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
)
131 fprintf(stderr
, "bucket %d/%d = %d/%d items\n", hash
% cache
->size
,
132 cache
->size
, bucketcount
, cache
->n_items
);
135 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
) {
136 if (c
->cache_id
== cache_id
&&
138 c
->key_size
== key_size
&&
139 memcmp(c
->key
, key
, key_size
) == 0 &&
140 c
->nr_reloc_bufs
== nr_reloc_bufs
&&
141 memcmp(c
->reloc_bufs
, reloc_bufs
,
142 nr_reloc_bufs
* sizeof(dri_bo
*)) == 0)
151 rehash(struct brw_cache
*cache
)
153 struct brw_cache_item
**items
;
154 struct brw_cache_item
*c
, *next
;
157 size
= cache
->size
* 3;
158 items
= (struct brw_cache_item
**) _mesa_calloc(size
* sizeof(*items
));
160 for (i
= 0; i
< cache
->size
; i
++)
161 for (c
= cache
->items
[i
]; c
; c
= next
) {
163 c
->next
= items
[c
->hash
% size
];
164 items
[c
->hash
% size
] = c
;
168 cache
->items
= items
;
174 * Returns the buffer object matching cache_id and key, or NULL.
177 brw_search_cache(struct brw_cache
*cache
,
178 enum brw_cache_id cache_id
,
181 dri_bo
**reloc_bufs
, GLuint nr_reloc_bufs
,
184 struct brw_cache_item
*item
;
185 GLuint hash
= hash_key(key
, key_size
, reloc_bufs
, nr_reloc_bufs
);
187 item
= search_cache(cache
, cache_id
, hash
, key
, key_size
,
188 reloc_bufs
, nr_reloc_bufs
);
194 *(void **)aux_return
= (void *)((char *)item
->key
+ item
->key_size
);
196 update_cache_last(cache
, cache_id
, item
->bo
);
198 dri_bo_reference(item
->bo
);
204 brw_upload_cache( struct brw_cache
*cache
,
205 enum brw_cache_id cache_id
,
209 GLuint nr_reloc_bufs
,
215 struct brw_cache_item
*item
= CALLOC_STRUCT(brw_cache_item
);
216 GLuint hash
= hash_key(key
, key_size
, reloc_bufs
, nr_reloc_bufs
);
217 GLuint relocs_size
= nr_reloc_bufs
* sizeof(dri_bo
*);
218 GLuint aux_size
= cache
->aux_size
[cache_id
];
223 /* Create the buffer object to contain the data */
224 bo
= dri_bo_alloc(cache
->brw
->intel
.bufmgr
,
225 cache
->name
[cache_id
], data_size
, 1 << 6);
228 /* Set up the memory containing the key, aux_data, and reloc_bufs */
229 tmp
= _mesa_malloc(key_size
+ aux_size
+ relocs_size
);
231 memcpy(tmp
, key
, key_size
);
232 memcpy(tmp
+ key_size
, aux
, cache
->aux_size
[cache_id
]);
233 memcpy(tmp
+ key_size
+ aux_size
, reloc_bufs
, relocs_size
);
234 for (i
= 0; i
< nr_reloc_bufs
; i
++) {
235 if (reloc_bufs
[i
] != NULL
)
236 dri_bo_reference(reloc_bufs
[i
]);
239 item
->cache_id
= cache_id
;
242 item
->key_size
= key_size
;
243 item
->reloc_bufs
= tmp
+ key_size
+ aux_size
;
244 item
->nr_reloc_bufs
= nr_reloc_bufs
;
247 dri_bo_reference(bo
);
249 if (cache
->n_items
> cache
->size
* 1.5)
253 item
->next
= cache
->items
[hash
];
254 cache
->items
[hash
] = item
;
258 assert(cache
->aux_size
[cache_id
]);
259 *(void **)aux_return
= (void *)((char *)item
->key
+ item
->key_size
);
262 if (INTEL_DEBUG
& DEBUG_STATE
)
263 _mesa_printf("upload %s: %d bytes to cache id %d\n",
264 cache
->name
[cache_id
],
265 data_size
, cache_id
);
267 /* Copy data to the buffer */
268 dri_bo_subdata(bo
, 0, data_size
, data
);
270 update_cache_last(cache
, cache_id
, bo
);
277 * Wrapper around brw_cache_data_sz using the cache_id's canonical key size.
279 * If nr_reloc_bufs is nonzero, brw_search_cache()/brw_upload_cache() would be
280 * better to use, as the potentially changing offsets in the data-used-as-key
281 * will result in excessive cache misses.
283 * If aux data is involved, use search/upload instead.
287 brw_cache_data(struct brw_cache
*cache
,
288 enum brw_cache_id cache_id
,
292 GLuint nr_reloc_bufs
)
295 struct brw_cache_item
*item
;
296 GLuint hash
= hash_key(data
, data_size
, reloc_bufs
, nr_reloc_bufs
);
298 item
= search_cache(cache
, cache_id
, hash
, data
, data_size
,
299 reloc_bufs
, nr_reloc_bufs
);
301 update_cache_last(cache
, cache_id
, item
->bo
);
302 dri_bo_reference(item
->bo
);
306 bo
= brw_upload_cache(cache
, cache_id
,
308 reloc_bufs
, nr_reloc_bufs
,
322 brw_init_cache_id(struct brw_cache
*cache
,
324 enum brw_cache_id id
,
327 cache
->name
[id
] = strdup(name
);
328 cache
->aux_size
[id
] = aux_size
;
333 brw_init_non_surface_cache(struct brw_context
*brw
)
335 struct brw_cache
*cache
= &brw
->cache
;
341 cache
->items
= (struct brw_cache_item
**)
342 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
344 brw_init_cache_id(cache
,
349 brw_init_cache_id(cache
,
354 brw_init_cache_id(cache
,
357 sizeof(struct brw_wm_prog_data
));
359 brw_init_cache_id(cache
,
360 "SAMPLER_DEFAULT_COLOR",
361 BRW_SAMPLER_DEFAULT_COLOR
,
364 brw_init_cache_id(cache
,
369 brw_init_cache_id(cache
,
374 brw_init_cache_id(cache
,
377 sizeof(struct brw_sf_prog_data
));
379 brw_init_cache_id(cache
,
384 brw_init_cache_id(cache
,
389 brw_init_cache_id(cache
,
394 brw_init_cache_id(cache
,
397 sizeof(struct brw_vs_prog_data
));
399 brw_init_cache_id(cache
,
404 brw_init_cache_id(cache
,
407 sizeof(struct brw_clip_prog_data
));
409 brw_init_cache_id(cache
,
414 brw_init_cache_id(cache
,
417 sizeof(struct brw_gs_prog_data
));
422 brw_init_surface_cache(struct brw_context
*brw
)
424 struct brw_cache
*cache
= &brw
->surface_cache
;
430 cache
->items
= (struct brw_cache_item
**)
431 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
433 brw_init_cache_id(cache
,
438 brw_init_cache_id(cache
,
446 brw_init_caches(struct brw_context
*brw
)
448 brw_init_non_surface_cache(brw
);
449 brw_init_surface_cache(brw
);
454 brw_clear_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
456 struct brw_cache_item
*c
, *next
;
459 if (INTEL_DEBUG
& DEBUG_STATE
)
460 _mesa_printf("%s\n", __FUNCTION__
);
462 for (i
= 0; i
< cache
->size
; i
++) {
463 for (c
= cache
->items
[i
]; c
; c
= next
) {
467 for (j
= 0; j
< c
->nr_reloc_bufs
; j
++)
468 dri_bo_unreference(c
->reloc_bufs
[j
]);
469 dri_bo_unreference(c
->bo
);
470 free((void *)c
->key
);
473 cache
->items
[i
] = NULL
;
478 if (brw
->curbe
.last_buf
) {
479 _mesa_free(brw
->curbe
.last_buf
);
480 brw
->curbe
.last_buf
= NULL
;
483 brw
->state
.dirty
.mesa
|= ~0;
484 brw
->state
.dirty
.brw
|= ~0;
485 brw
->state
.dirty
.cache
|= ~0;
488 /* Clear all entries from the cache that point to the given bo.
490 * This lets us release memory for reuse earlier for known-dead buffers,
491 * at the cost of walking the entire hash table.
494 brw_state_cache_bo_delete(struct brw_cache
*cache
, dri_bo
*bo
)
496 struct brw_cache_item
**prev
;
499 if (INTEL_DEBUG
& DEBUG_STATE
)
500 _mesa_printf("%s\n", __FUNCTION__
);
502 for (i
= 0; i
< cache
->size
; i
++) {
503 for (prev
= &cache
->items
[i
]; *prev
;) {
504 struct brw_cache_item
*c
= *prev
;
506 if (drm_intel_bo_references(c
->bo
, bo
)) {
511 for (j
= 0; j
< c
->nr_reloc_bufs
; j
++)
512 dri_bo_unreference(c
->reloc_bufs
[j
]);
513 dri_bo_unreference(c
->bo
);
514 free((void *)c
->key
);
525 brw_state_cache_check_size(struct brw_context
*brw
)
527 if (INTEL_DEBUG
& DEBUG_STATE
)
528 _mesa_printf("%s (n_items=%d)\n", __FUNCTION__
, brw
->cache
.n_items
);
530 /* un-tuned guess. We've got around 20 state objects for a total of around
531 * 32k, so 1000 of them is around 1.5MB.
533 if (brw
->cache
.n_items
> 1000)
534 brw_clear_cache(brw
, &brw
->cache
);
536 if (brw
->surface_cache
.n_items
> 1000)
537 brw_clear_cache(brw
, &brw
->surface_cache
);
542 brw_destroy_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
546 if (INTEL_DEBUG
& DEBUG_STATE
)
547 _mesa_printf("%s\n", __FUNCTION__
);
549 brw_clear_cache(brw
, cache
);
550 for (i
= 0; i
< BRW_MAX_CACHE
; i
++) {
551 dri_bo_unreference(cache
->last_bo
[i
]);
552 free(cache
->name
[i
]);
561 brw_destroy_caches(struct brw_context
*brw
)
563 brw_destroy_cache(brw
, &brw
->cache
);
564 brw_destroy_cache(brw
, &brw
->surface_cache
);