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 "brw_state.h"
60 #include "brw_batchbuffer.h"
62 /* XXX: Fixme - have to include these to get the sizes of the prog_key
73 hash_key(const void *key
, GLuint key_size
,
74 struct brw_winsys_buffer
**reloc_bufs
, GLuint nr_reloc_bufs
)
76 GLuint
*ikey
= (GLuint
*)key
;
79 assert(key_size
% 4 == 0);
81 /* I'm sure this can be improved on:
83 for (i
= 0; i
< key_size
/4; i
++) {
85 hash
= (hash
<< 5) | (hash
>> 27);
88 /* Include the BO pointers as key data as well */
89 ikey
= (GLuint
*)reloc_bufs
;
90 key_size
= nr_reloc_bufs
* sizeof(struct brw_winsys_buffer
*);
91 for (i
= 0; i
< key_size
/4; i
++) {
93 hash
= (hash
<< 5) | (hash
>> 27);
101 * Marks a new buffer as being chosen for the given cache id.
104 update_cache_last(struct brw_cache
*cache
, enum brw_cache_id cache_id
,
105 struct brw_winsys_buffer
*bo
)
107 if (bo
== cache
->last_bo
[cache_id
])
108 return; /* no change */
110 brw
->sws
->bo_unreference(cache
->last_bo
[cache_id
]);
111 cache
->last_bo
[cache_id
] = bo
;
112 brw
->sws
->bo_reference(cache
->last_bo
[cache_id
]);
113 cache
->brw
->state
.dirty
.cache
|= 1 << cache_id
;
117 static struct brw_cache_item
*
118 search_cache(struct brw_cache
*cache
, enum brw_cache_id cache_id
,
119 GLuint hash
, const void *key
, GLuint key_size
,
120 struct brw_winsys_buffer
**reloc_bufs
, GLuint nr_reloc_bufs
)
122 struct brw_cache_item
*c
;
127 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
)
130 fprintf(stderr
, "bucket %d/%d = %d/%d items\n", hash
% cache
->size
,
131 cache
->size
, bucketcount
, cache
->n_items
);
134 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
) {
135 if (c
->cache_id
== cache_id
&&
137 c
->key_size
== key_size
&&
138 memcmp(c
->key
, key
, key_size
) == 0 &&
139 c
->nr_reloc_bufs
== nr_reloc_bufs
&&
140 memcmp(c
->reloc_bufs
, reloc_bufs
,
141 nr_reloc_bufs
* sizeof(struct brw_winsys_buffer
*)) == 0)
150 rehash(struct brw_cache
*cache
)
152 struct brw_cache_item
**items
;
153 struct brw_cache_item
*c
, *next
;
156 size
= cache
->size
* 3;
157 items
= (struct brw_cache_item
**) _mesa_calloc(size
* sizeof(*items
));
159 for (i
= 0; i
< cache
->size
; i
++)
160 for (c
= cache
->items
[i
]; c
; c
= next
) {
162 c
->next
= items
[c
->hash
% size
];
163 items
[c
->hash
% size
] = c
;
167 cache
->items
= items
;
173 * Returns the buffer object matching cache_id and key, or NULL.
175 struct brw_winsys_buffer
*
176 brw_search_cache(struct brw_cache
*cache
,
177 enum brw_cache_id cache_id
,
180 struct brw_winsys_buffer
**reloc_bufs
, GLuint nr_reloc_bufs
,
183 struct brw_cache_item
*item
;
184 GLuint hash
= hash_key(key
, key_size
, reloc_bufs
, nr_reloc_bufs
);
186 item
= search_cache(cache
, cache_id
, hash
, key
, key_size
,
187 reloc_bufs
, nr_reloc_bufs
);
193 *(void **)aux_return
= (void *)((char *)item
->key
+ item
->key_size
);
195 update_cache_last(cache
, cache_id
, item
->bo
);
197 brw
->sws
->bo_reference(item
->bo
);
202 struct brw_winsys_buffer
*
203 brw_upload_cache( struct brw_cache
*cache
,
204 enum brw_cache_id cache_id
,
207 struct brw_winsys_buffer
**reloc_bufs
,
208 GLuint nr_reloc_bufs
,
214 struct brw_cache_item
*item
= CALLOC_STRUCT(brw_cache_item
);
215 GLuint hash
= hash_key(key
, key_size
, reloc_bufs
, nr_reloc_bufs
);
216 GLuint relocs_size
= nr_reloc_bufs
* sizeof(struct brw_winsys_buffer
*);
217 GLuint aux_size
= cache
->aux_size
[cache_id
];
219 struct brw_winsys_buffer
*bo
;
222 /* Create the buffer object to contain the data */
223 bo
= dri_bo_alloc(cache
->brw
->intel
.bufmgr
,
224 cache
->name
[cache_id
], data_size
, 1 << 6);
227 /* Set up the memory containing the key, aux_data, and reloc_bufs */
228 tmp
= _mesa_malloc(key_size
+ aux_size
+ relocs_size
);
230 memcpy(tmp
, key
, key_size
);
231 memcpy(tmp
+ key_size
, aux
, cache
->aux_size
[cache_id
]);
232 memcpy(tmp
+ key_size
+ aux_size
, reloc_bufs
, relocs_size
);
233 for (i
= 0; i
< nr_reloc_bufs
; i
++) {
234 if (reloc_bufs
[i
] != NULL
)
235 brw
->sws
->bo_reference(reloc_bufs
[i
]);
238 item
->cache_id
= cache_id
;
241 item
->key_size
= key_size
;
242 item
->reloc_bufs
= tmp
+ key_size
+ aux_size
;
243 item
->nr_reloc_bufs
= nr_reloc_bufs
;
246 brw
->sws
->bo_reference(bo
);
247 item
->data_size
= data_size
;
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 * This doesn't really work with aux data. Use search/upload instead
279 struct brw_winsys_buffer
*
280 brw_cache_data_sz(struct brw_cache
*cache
,
281 enum brw_cache_id cache_id
,
284 struct brw_winsys_buffer
**reloc_bufs
,
285 GLuint nr_reloc_bufs
)
287 struct brw_winsys_buffer
*bo
;
288 struct brw_cache_item
*item
;
289 GLuint hash
= hash_key(data
, data_size
, reloc_bufs
, nr_reloc_bufs
);
291 item
= search_cache(cache
, cache_id
, hash
, data
, data_size
,
292 reloc_bufs
, nr_reloc_bufs
);
294 update_cache_last(cache
, cache_id
, item
->bo
);
295 brw
->sws
->bo_reference(item
->bo
);
299 bo
= brw_upload_cache(cache
, cache_id
,
301 reloc_bufs
, nr_reloc_bufs
,
310 * Wrapper around brw_cache_data_sz using the cache_id's canonical key size.
312 * If nr_reloc_bufs is nonzero, brw_search_cache()/brw_upload_cache() would be
313 * better to use, as the potentially changing offsets in the data-used-as-key
314 * will result in excessive cache misses.
316 struct brw_winsys_buffer
*
317 brw_cache_data(struct brw_cache
*cache
,
318 enum brw_cache_id cache_id
,
320 struct brw_winsys_buffer
**reloc_bufs
,
321 GLuint nr_reloc_bufs
)
323 return brw_cache_data_sz(cache
, cache_id
, data
, cache
->key_size
[cache_id
],
324 reloc_bufs
, nr_reloc_bufs
);
334 brw_init_cache_id(struct brw_cache
*cache
,
336 enum brw_cache_id id
,
340 cache
->name
[id
] = strdup(name
);
341 cache
->key_size
[id
] = key_size
;
342 cache
->aux_size
[id
] = aux_size
;
347 brw_init_non_surface_cache(struct brw_context
*brw
)
349 struct brw_cache
*cache
= &brw
->cache
;
355 cache
->items
= (struct brw_cache_item
**)
356 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
358 brw_init_cache_id(cache
,
361 sizeof(struct brw_cc_viewport
),
364 brw_init_cache_id(cache
,
367 sizeof(struct brw_cc_unit_state
),
370 brw_init_cache_id(cache
,
373 sizeof(struct brw_wm_prog_key
),
374 sizeof(struct brw_wm_prog_data
));
376 brw_init_cache_id(cache
,
377 "SAMPLER_DEFAULT_COLOR",
378 BRW_SAMPLER_DEFAULT_COLOR
,
379 sizeof(struct brw_sampler_default_color
),
382 brw_init_cache_id(cache
,
385 0, /* variable key/data size */
388 brw_init_cache_id(cache
,
391 sizeof(struct brw_wm_unit_state
),
394 brw_init_cache_id(cache
,
397 sizeof(struct brw_sf_prog_key
),
398 sizeof(struct brw_sf_prog_data
));
400 brw_init_cache_id(cache
,
403 sizeof(struct brw_sf_viewport
),
406 brw_init_cache_id(cache
,
409 sizeof(struct brw_sf_unit_state
),
412 brw_init_cache_id(cache
,
415 sizeof(struct brw_vs_unit_state
),
418 brw_init_cache_id(cache
,
421 sizeof(struct brw_vs_prog_key
),
422 sizeof(struct brw_vs_prog_data
));
424 brw_init_cache_id(cache
,
427 sizeof(struct brw_clip_unit_state
),
430 brw_init_cache_id(cache
,
433 sizeof(struct brw_clip_prog_key
),
434 sizeof(struct brw_clip_prog_data
));
436 brw_init_cache_id(cache
,
439 sizeof(struct brw_gs_unit_state
),
442 brw_init_cache_id(cache
,
445 sizeof(struct brw_gs_prog_key
),
446 sizeof(struct brw_gs_prog_data
));
451 brw_init_surface_cache(struct brw_context
*brw
)
453 struct brw_cache
*cache
= &brw
->surface_cache
;
459 cache
->items
= (struct brw_cache_item
**)
460 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
462 brw_init_cache_id(cache
,
465 sizeof(struct brw_surface_state
),
468 brw_init_cache_id(cache
,
477 brw_init_caches(struct brw_context
*brw
)
479 brw_init_non_surface_cache(brw
);
480 brw_init_surface_cache(brw
);
485 brw_clear_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
487 struct brw_cache_item
*c
, *next
;
490 if (INTEL_DEBUG
& DEBUG_STATE
)
491 _mesa_printf("%s\n", __FUNCTION__
);
493 for (i
= 0; i
< cache
->size
; i
++) {
494 for (c
= cache
->items
[i
]; c
; c
= next
) {
498 for (j
= 0; j
< c
->nr_reloc_bufs
; j
++)
499 brw
->sws
->bo_unreference(c
->reloc_bufs
[j
]);
500 brw
->sws
->bo_unreference(c
->bo
);
501 free((void *)c
->key
);
504 cache
->items
[i
] = NULL
;
509 if (brw
->curbe
.last_buf
) {
510 _mesa_free(brw
->curbe
.last_buf
);
511 brw
->curbe
.last_buf
= NULL
;
514 brw
->state
.dirty
.mesa
|= ~0;
515 brw
->state
.dirty
.brw
|= ~0;
516 brw
->state
.dirty
.cache
|= ~0;
519 /* Clear all entries from the cache that point to the given bo.
521 * This lets us release memory for reuse earlier for known-dead buffers,
522 * at the cost of walking the entire hash table.
525 brw_state_cache_bo_delete(struct brw_cache
*cache
, struct brw_winsys_buffer
*bo
)
527 struct brw_cache_item
**prev
;
530 if (INTEL_DEBUG
& DEBUG_STATE
)
531 _mesa_printf("%s\n", __FUNCTION__
);
533 for (i
= 0; i
< cache
->size
; i
++) {
534 for (prev
= &cache
->items
[i
]; *prev
;) {
535 struct brw_cache_item
*c
= *prev
;
537 if (cache
->sws
->bo_references(c
->bo
, bo
)) {
542 for (j
= 0; j
< c
->nr_reloc_bufs
; j
++)
543 brw
->sws
->bo_unreference(c
->reloc_bufs
[j
]);
544 brw
->sws
->bo_unreference(c
->bo
);
545 free((void *)c
->key
);
556 brw_state_cache_check_size(struct brw_context
*brw
)
558 if (INTEL_DEBUG
& DEBUG_STATE
)
559 _mesa_printf("%s (n_items=%d)\n", __FUNCTION__
, brw
->cache
.n_items
);
561 /* un-tuned guess. We've got around 20 state objects for a total of around
562 * 32k, so 1000 of them is around 1.5MB.
564 if (brw
->cache
.n_items
> 1000)
565 brw_clear_cache(brw
, &brw
->cache
);
567 if (brw
->surface_cache
.n_items
> 1000)
568 brw_clear_cache(brw
, &brw
->surface_cache
);
573 brw_destroy_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
577 if (INTEL_DEBUG
& DEBUG_STATE
)
578 _mesa_printf("%s\n", __FUNCTION__
);
580 brw_clear_cache(brw
, cache
);
581 for (i
= 0; i
< BRW_MAX_CACHE
; i
++) {
582 brw
->sws
->bo_unreference(cache
->last_bo
[i
]);
583 free(cache
->name
[i
]);
592 brw_destroy_caches(struct brw_context
*brw
)
594 brw_destroy_cache(brw
, &brw
->cache
);
595 brw_destroy_cache(brw
, &brw
->surface_cache
);