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 * This doesn't really work with aux data. Use search/upload instead
281 brw_cache_data_sz(struct brw_cache
*cache
,
282 enum brw_cache_id cache_id
,
286 GLuint nr_reloc_bufs
)
289 struct brw_cache_item
*item
;
290 GLuint hash
= hash_key(data
, data_size
, reloc_bufs
, nr_reloc_bufs
);
292 item
= search_cache(cache
, cache_id
, hash
, data
, data_size
,
293 reloc_bufs
, nr_reloc_bufs
);
295 update_cache_last(cache
, cache_id
, item
->bo
);
296 dri_bo_reference(item
->bo
);
300 bo
= brw_upload_cache(cache
, cache_id
,
302 reloc_bufs
, nr_reloc_bufs
,
311 * Wrapper around brw_cache_data_sz using the cache_id's canonical key size.
313 * If nr_reloc_bufs is nonzero, brw_search_cache()/brw_upload_cache() would be
314 * better to use, as the potentially changing offsets in the data-used-as-key
315 * will result in excessive cache misses.
318 brw_cache_data(struct brw_cache
*cache
,
319 enum brw_cache_id cache_id
,
322 GLuint nr_reloc_bufs
)
324 return brw_cache_data_sz(cache
, cache_id
, data
, cache
->key_size
[cache_id
],
325 reloc_bufs
, nr_reloc_bufs
);
335 brw_init_cache_id(struct brw_cache
*cache
,
337 enum brw_cache_id id
,
341 cache
->name
[id
] = strdup(name
);
342 cache
->key_size
[id
] = key_size
;
343 cache
->aux_size
[id
] = aux_size
;
348 brw_init_non_surface_cache(struct brw_context
*brw
)
350 struct brw_cache
*cache
= &brw
->cache
;
356 cache
->items
= (struct brw_cache_item
**)
357 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
359 brw_init_cache_id(cache
,
362 sizeof(struct brw_cc_viewport
),
365 brw_init_cache_id(cache
,
368 sizeof(struct brw_cc_unit_state
),
371 brw_init_cache_id(cache
,
374 sizeof(struct brw_wm_prog_key
),
375 sizeof(struct brw_wm_prog_data
));
377 brw_init_cache_id(cache
,
378 "SAMPLER_DEFAULT_COLOR",
379 BRW_SAMPLER_DEFAULT_COLOR
,
380 sizeof(struct brw_sampler_default_color
),
383 brw_init_cache_id(cache
,
386 0, /* variable key/data size */
389 brw_init_cache_id(cache
,
392 sizeof(struct brw_wm_unit_state
),
395 brw_init_cache_id(cache
,
398 sizeof(struct brw_sf_prog_key
),
399 sizeof(struct brw_sf_prog_data
));
401 brw_init_cache_id(cache
,
404 sizeof(struct brw_sf_viewport
),
407 brw_init_cache_id(cache
,
410 sizeof(struct brw_sf_unit_state
),
413 brw_init_cache_id(cache
,
416 sizeof(struct brw_vs_unit_state
),
419 brw_init_cache_id(cache
,
422 sizeof(struct brw_vs_prog_key
),
423 sizeof(struct brw_vs_prog_data
));
425 brw_init_cache_id(cache
,
428 sizeof(struct brw_clip_unit_state
),
431 brw_init_cache_id(cache
,
434 sizeof(struct brw_clip_prog_key
),
435 sizeof(struct brw_clip_prog_data
));
437 brw_init_cache_id(cache
,
440 sizeof(struct brw_gs_unit_state
),
443 brw_init_cache_id(cache
,
446 sizeof(struct brw_gs_prog_key
),
447 sizeof(struct brw_gs_prog_data
));
452 brw_init_surface_cache(struct brw_context
*brw
)
454 struct brw_cache
*cache
= &brw
->surface_cache
;
460 cache
->items
= (struct brw_cache_item
**)
461 _mesa_calloc(cache
->size
* sizeof(struct brw_cache_item
));
463 brw_init_cache_id(cache
,
466 sizeof(struct brw_surface_state
),
469 brw_init_cache_id(cache
,
478 brw_init_caches(struct brw_context
*brw
)
480 brw_init_non_surface_cache(brw
);
481 brw_init_surface_cache(brw
);
486 brw_clear_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
488 struct brw_cache_item
*c
, *next
;
491 if (INTEL_DEBUG
& DEBUG_STATE
)
492 _mesa_printf("%s\n", __FUNCTION__
);
494 for (i
= 0; i
< cache
->size
; i
++) {
495 for (c
= cache
->items
[i
]; c
; c
= next
) {
499 for (j
= 0; j
< c
->nr_reloc_bufs
; j
++)
500 dri_bo_unreference(c
->reloc_bufs
[j
]);
501 dri_bo_unreference(c
->bo
);
502 free((void *)c
->key
);
505 cache
->items
[i
] = NULL
;
510 if (brw
->curbe
.last_buf
) {
511 _mesa_free(brw
->curbe
.last_buf
);
512 brw
->curbe
.last_buf
= NULL
;
515 brw
->state
.dirty
.mesa
|= ~0;
516 brw
->state
.dirty
.brw
|= ~0;
517 brw
->state
.dirty
.cache
|= ~0;
522 brw_state_cache_check_size(struct brw_context
*brw
)
524 /* un-tuned guess. We've got around 20 state objects for a total of around
525 * 32k, so 1000 of them is around 1.5MB.
527 if (brw
->cache
.n_items
> 1000)
528 brw_clear_cache(brw
, &brw
->cache
);
530 if (brw
->surface_cache
.n_items
> 1000)
531 brw_clear_cache(brw
, &brw
->surface_cache
);
536 brw_destroy_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
540 brw_clear_cache(brw
, cache
);
541 for (i
= 0; i
< BRW_MAX_CACHE
; i
++) {
542 dri_bo_unreference(cache
->last_bo
[i
]);
543 free(cache
->name
[i
]);
552 brw_destroy_caches(struct brw_context
*brw
)
554 brw_destroy_cache(brw
, &brw
->cache
);
555 brw_destroy_cache(brw
, &brw
->surface_cache
);