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
);
248 item
->data_size
= data_size
;
250 if (cache
->n_items
> cache
->size
* 1.5)
254 item
->next
= cache
->items
[hash
];
255 cache
->items
[hash
] = item
;
259 assert(cache
->aux_size
[cache_id
]);
260 *(void **)aux_return
= (void *)((char *)item
->key
+ item
->key_size
);
263 if (INTEL_DEBUG
& DEBUG_STATE
)
264 _mesa_printf("upload %s: %d bytes to cache id %d\n",
265 cache
->name
[cache_id
],
266 data_size
, cache_id
);
268 /* Copy data to the buffer */
269 dri_bo_subdata(bo
, 0, data_size
, data
);
271 update_cache_last(cache
, cache_id
, bo
);
278 * Wrapper around brw_cache_data_sz using the cache_id's canonical key size.
280 * If nr_reloc_bufs is nonzero, brw_search_cache()/brw_upload_cache() would be
281 * better to use, as the potentially changing offsets in the data-used-as-key
282 * will result in excessive cache misses.
284 * If aux data is involved, use search/upload instead.
288 brw_cache_data(struct brw_cache
*cache
,
289 enum brw_cache_id cache_id
,
293 GLuint nr_reloc_bufs
)
296 struct brw_cache_item
*item
;
297 GLuint hash
= hash_key(data
, data_size
, reloc_bufs
, nr_reloc_bufs
);
299 item
= search_cache(cache
, cache_id
, hash
, data
, data_size
,
300 reloc_bufs
, nr_reloc_bufs
);
302 update_cache_last(cache
, cache_id
, item
->bo
);
303 dri_bo_reference(item
->bo
);
307 bo
= brw_upload_cache(cache
, cache_id
,
309 reloc_bufs
, nr_reloc_bufs
,
323 brw_init_cache_id(struct brw_cache
*cache
,
325 enum brw_cache_id id
,
328 cache
->name
[id
] = strdup(name
);
329 cache
->aux_size
[id
] = aux_size
;
334 brw_init_non_surface_cache(struct brw_context
*brw
)
336 struct brw_cache
*cache
= &brw
->cache
;
342 cache
->items
= (struct brw_cache_item
**)
343 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
345 brw_init_cache_id(cache
,
350 brw_init_cache_id(cache
,
355 brw_init_cache_id(cache
,
358 sizeof(struct brw_wm_prog_data
));
360 brw_init_cache_id(cache
,
361 "SAMPLER_DEFAULT_COLOR",
362 BRW_SAMPLER_DEFAULT_COLOR
,
365 brw_init_cache_id(cache
,
370 brw_init_cache_id(cache
,
375 brw_init_cache_id(cache
,
378 sizeof(struct brw_sf_prog_data
));
380 brw_init_cache_id(cache
,
385 brw_init_cache_id(cache
,
390 brw_init_cache_id(cache
,
395 brw_init_cache_id(cache
,
398 sizeof(struct brw_vs_prog_data
));
400 brw_init_cache_id(cache
,
405 brw_init_cache_id(cache
,
408 sizeof(struct brw_clip_prog_data
));
410 brw_init_cache_id(cache
,
415 brw_init_cache_id(cache
,
418 sizeof(struct brw_gs_prog_data
));
423 brw_init_surface_cache(struct brw_context
*brw
)
425 struct brw_cache
*cache
= &brw
->surface_cache
;
431 cache
->items
= (struct brw_cache_item
**)
432 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
434 brw_init_cache_id(cache
,
439 brw_init_cache_id(cache
,
447 brw_init_caches(struct brw_context
*brw
)
449 brw_init_non_surface_cache(brw
);
450 brw_init_surface_cache(brw
);
455 brw_clear_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
457 struct brw_cache_item
*c
, *next
;
460 if (INTEL_DEBUG
& DEBUG_STATE
)
461 _mesa_printf("%s\n", __FUNCTION__
);
463 for (i
= 0; i
< cache
->size
; i
++) {
464 for (c
= cache
->items
[i
]; c
; c
= next
) {
468 for (j
= 0; j
< c
->nr_reloc_bufs
; j
++)
469 dri_bo_unreference(c
->reloc_bufs
[j
]);
470 dri_bo_unreference(c
->bo
);
471 free((void *)c
->key
);
474 cache
->items
[i
] = NULL
;
479 if (brw
->curbe
.last_buf
) {
480 _mesa_free(brw
->curbe
.last_buf
);
481 brw
->curbe
.last_buf
= NULL
;
484 brw
->state
.dirty
.mesa
|= ~0;
485 brw
->state
.dirty
.brw
|= ~0;
486 brw
->state
.dirty
.cache
|= ~0;
489 /* Clear all entries from the cache that point to the given bo.
491 * This lets us release memory for reuse earlier for known-dead buffers,
492 * at the cost of walking the entire hash table.
495 brw_state_cache_bo_delete(struct brw_cache
*cache
, dri_bo
*bo
)
497 struct brw_cache_item
**prev
;
500 if (INTEL_DEBUG
& DEBUG_STATE
)
501 _mesa_printf("%s\n", __FUNCTION__
);
503 for (i
= 0; i
< cache
->size
; i
++) {
504 for (prev
= &cache
->items
[i
]; *prev
;) {
505 struct brw_cache_item
*c
= *prev
;
507 if (drm_intel_bo_references(c
->bo
, bo
)) {
512 for (j
= 0; j
< c
->nr_reloc_bufs
; j
++)
513 dri_bo_unreference(c
->reloc_bufs
[j
]);
514 dri_bo_unreference(c
->bo
);
515 free((void *)c
->key
);
526 brw_state_cache_check_size(struct brw_context
*brw
)
528 if (INTEL_DEBUG
& DEBUG_STATE
)
529 _mesa_printf("%s (n_items=%d)\n", __FUNCTION__
, brw
->cache
.n_items
);
531 /* un-tuned guess. We've got around 20 state objects for a total of around
532 * 32k, so 1000 of them is around 1.5MB.
534 if (brw
->cache
.n_items
> 1000)
535 brw_clear_cache(brw
, &brw
->cache
);
537 if (brw
->surface_cache
.n_items
> 1000)
538 brw_clear_cache(brw
, &brw
->surface_cache
);
543 brw_destroy_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
547 if (INTEL_DEBUG
& DEBUG_STATE
)
548 _mesa_printf("%s\n", __FUNCTION__
);
550 brw_clear_cache(brw
, cache
);
551 for (i
= 0; i
< BRW_MAX_CACHE
; i
++) {
552 dri_bo_unreference(cache
->last_bo
[i
]);
553 free(cache
->name
[i
]);
562 brw_destroy_caches(struct brw_context
*brw
)
564 brw_destroy_cache(brw
, &brw
->cache
);
565 brw_destroy_cache(brw
, &brw
->surface_cache
);