f7c0a2037d944c90dff5a53499067ac73c82e468
2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics 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 <keithw@vmware.com>
32 /** @file brw_state_cache.c
34 * This file implements a simple static state cache for 965. The
35 * consumers can query the hash table of state using a cache_id,
36 * opaque key data, and receive the corresponding state buffer object
37 * of state (plus associated auxiliary data) in return. Objects in
38 * the cache may not have relocations (pointers to other BOs) in them.
40 * The inner workings are a simple hash table based on a CRC of the
43 * Replacement is not implemented. Instead, when the cache gets too
44 * big we throw out all of the cache data and let it get regenerated.
47 #include "main/imports.h"
48 #include "intel_batchbuffer.h"
49 #include "brw_state.h"
55 #define FILE_DEBUG_FLAG DEBUG_STATE
58 hash_key(struct brw_cache_item
*item
)
60 GLuint
*ikey
= (GLuint
*)item
->key
;
61 GLuint hash
= item
->cache_id
, i
;
63 assert(item
->key_size
% 4 == 0);
65 /* I'm sure this can be improved on:
67 for (i
= 0; i
< item
->key_size
/4; i
++) {
69 hash
= (hash
<< 5) | (hash
>> 27);
76 brw_cache_item_equals(const struct brw_cache_item
*a
,
77 const struct brw_cache_item
*b
)
79 return a
->cache_id
== b
->cache_id
&&
81 a
->key_size
== b
->key_size
&&
82 (memcmp(a
->key
, b
->key
, a
->key_size
) == 0);
85 static struct brw_cache_item
*
86 search_cache(struct brw_cache
*cache
, GLuint hash
,
87 struct brw_cache_item
*lookup
)
89 struct brw_cache_item
*c
;
94 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
)
97 fprintf(stderr
, "bucket %d/%d = %d/%d items\n", hash
% cache
->size
,
98 cache
->size
, bucketcount
, cache
->n_items
);
101 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
) {
102 if (brw_cache_item_equals(lookup
, c
))
111 rehash(struct brw_cache
*cache
)
113 struct brw_cache_item
**items
;
114 struct brw_cache_item
*c
, *next
;
117 size
= cache
->size
* 3;
118 items
= calloc(size
, sizeof(*items
));
120 for (i
= 0; i
< cache
->size
; i
++)
121 for (c
= cache
->items
[i
]; c
; c
= next
) {
123 c
->next
= items
[c
->hash
% size
];
124 items
[c
->hash
% size
] = c
;
128 cache
->items
= items
;
134 * Returns the buffer object matching cache_id and key, or NULL.
137 brw_search_cache(struct brw_cache
*cache
,
138 enum brw_cache_id cache_id
,
139 const void *key
, GLuint key_size
,
140 uint32_t *inout_offset
, void *inout_aux
)
142 struct brw_context
*brw
= cache
->brw
;
143 struct brw_cache_item
*item
;
144 struct brw_cache_item lookup
;
147 lookup
.cache_id
= cache_id
;
149 lookup
.key_size
= key_size
;
150 hash
= hash_key(&lookup
);
153 item
= search_cache(cache
, hash
, &lookup
);
158 void *aux
= ((char *) item
->key
) + item
->key_size
;
160 if (item
->offset
!= *inout_offset
|| aux
!= *((void **) inout_aux
)) {
161 brw
->ctx
.NewDriverState
|= (1 << cache_id
);
162 *inout_offset
= item
->offset
;
163 *((void **) inout_aux
) = aux
;
170 brw_cache_new_bo(struct brw_cache
*cache
, uint32_t new_size
)
172 struct brw_context
*brw
= cache
->brw
;
173 drm_intel_bo
*new_bo
;
175 new_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "program cache", new_size
, 64);
177 drm_intel_gem_bo_map_unsynchronized(new_bo
);
179 /* Copy any existing data that needs to be saved. */
180 if (cache
->next_offset
!= 0) {
182 memcpy(new_bo
->virtual, cache
->bo
->virtual, cache
->next_offset
);
184 drm_intel_bo_map(cache
->bo
, false);
185 drm_intel_bo_subdata(new_bo
, 0, cache
->next_offset
,
187 drm_intel_bo_unmap(cache
->bo
);
192 drm_intel_bo_unmap(cache
->bo
);
193 drm_intel_bo_unreference(cache
->bo
);
195 cache
->bo_used_by_gpu
= false;
197 /* Since we have a new BO in place, we need to signal the units
198 * that depend on it (state base address on gen5+, or unit state before).
200 brw
->ctx
.NewDriverState
|= BRW_NEW_PROGRAM_CACHE
;
204 * Attempts to find an item in the cache with identical data.
206 static const struct brw_cache_item
*
207 brw_lookup_prog(const struct brw_cache
*cache
,
208 enum brw_cache_id cache_id
,
209 const void *data
, unsigned data_size
)
211 const struct brw_context
*brw
= cache
->brw
;
213 const struct brw_cache_item
*item
;
215 for (i
= 0; i
< cache
->size
; i
++) {
216 for (item
= cache
->items
[i
]; item
; item
= item
->next
) {
219 if (item
->cache_id
!= cache_id
|| item
->size
!= data_size
)
223 drm_intel_bo_map(cache
->bo
, false);
224 ret
= memcmp(cache
->bo
->virtual + item
->offset
, data
, item
->size
);
226 drm_intel_bo_unmap(cache
->bo
);
238 brw_alloc_item_data(struct brw_cache
*cache
, uint32_t size
)
241 struct brw_context
*brw
= cache
->brw
;
243 /* Allocate space in the cache BO for our new program. */
244 if (cache
->next_offset
+ size
> cache
->bo
->size
) {
245 uint32_t new_size
= cache
->bo
->size
* 2;
247 while (cache
->next_offset
+ size
> new_size
)
250 brw_cache_new_bo(cache
, new_size
);
253 /* If we would block on writing to an in-use program BO, just
256 if (!brw
->has_llc
&& cache
->bo_used_by_gpu
) {
257 perf_debug("Copying busy program cache buffer.\n");
258 brw_cache_new_bo(cache
, cache
->bo
->size
);
261 offset
= cache
->next_offset
;
263 /* Programs are always 64-byte aligned, so set up the next one now */
264 cache
->next_offset
= ALIGN(offset
+ size
, 64);
270 brw_upload_cache(struct brw_cache
*cache
,
271 enum brw_cache_id cache_id
,
278 uint32_t *out_offset
,
281 struct brw_context
*brw
= cache
->brw
;
282 struct brw_cache_item
*item
= CALLOC_STRUCT(brw_cache_item
);
283 const struct brw_cache_item
*matching_data
=
284 brw_lookup_prog(cache
, cache_id
, data
, data_size
);
288 item
->cache_id
= cache_id
;
289 item
->size
= data_size
;
291 item
->key_size
= key_size
;
292 item
->aux_size
= aux_size
;
293 hash
= hash_key(item
);
296 /* If we can find a matching prog in the cache already, then reuse the
297 * existing stuff without creating new copy into the underlying buffer
298 * object. This is notably useful for programs generating shaders at
299 * runtime, where multiple shaders may compile to the same thing in our
303 item
->offset
= matching_data
->offset
;
305 item
->offset
= brw_alloc_item_data(cache
, data_size
);
307 /* Copy data to the buffer */
309 memcpy((char *)cache
->bo
->virtual + item
->offset
, data
, data_size
);
311 drm_intel_bo_subdata(cache
->bo
, item
->offset
, data_size
, data
);
315 /* Set up the memory containing the key and aux_data */
316 tmp
= malloc(key_size
+ aux_size
);
318 memcpy(tmp
, key
, key_size
);
319 memcpy(tmp
+ key_size
, aux
, aux_size
);
323 if (cache
->n_items
> cache
->size
* 1.5f
)
327 item
->next
= cache
->items
[hash
];
328 cache
->items
[hash
] = item
;
331 *out_offset
= item
->offset
;
332 *(void **)out_aux
= (void *)((char *)item
->key
+ item
->key_size
);
333 cache
->brw
->ctx
.NewDriverState
|= 1 << cache_id
;
337 brw_init_caches(struct brw_context
*brw
)
339 struct brw_cache
*cache
= &brw
->cache
;
346 calloc(cache
->size
, sizeof(struct brw_cache_item
*));
348 cache
->bo
= drm_intel_bo_alloc(brw
->bufmgr
,
352 drm_intel_gem_bo_map_unsynchronized(cache
->bo
);
356 brw_clear_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
358 struct brw_cache_item
*c
, *next
;
361 DBG("%s\n", __func__
);
363 for (i
= 0; i
< cache
->size
; i
++) {
364 for (c
= cache
->items
[i
]; c
; c
= next
) {
366 if (c
->cache_id
== BRW_CACHE_VS_PROG
||
367 c
->cache_id
== BRW_CACHE_GS_PROG
||
368 c
->cache_id
== BRW_CACHE_FS_PROG
||
369 c
->cache_id
== BRW_CACHE_CS_PROG
) {
370 const void *item_aux
= c
->key
+ c
->key_size
;
371 brw_stage_prog_data_free(item_aux
);
373 free((void *)c
->key
);
376 cache
->items
[i
] = NULL
;
381 /* Start putting programs into the start of the BO again, since
382 * we'll never find the old results.
384 cache
->next_offset
= 0;
386 /* We need to make sure that the programs get regenerated, since
387 * any offsets leftover in brw_context will no longer be valid.
389 brw
->NewGLState
|= ~0;
390 brw
->ctx
.NewDriverState
|= ~0ull;
391 intel_batchbuffer_flush(brw
);
395 brw_state_cache_check_size(struct brw_context
*brw
)
397 /* un-tuned guess. Each object is generally a page, so 2000 of them is 8 MB of
400 if (brw
->cache
.n_items
> 2000) {
401 perf_debug("Exceeded state cache size limit. Clearing the set "
402 "of compiled programs, which will trigger recompiles\n");
403 brw_clear_cache(brw
, &brw
->cache
);
409 brw_destroy_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
412 DBG("%s\n", __func__
);
415 drm_intel_bo_unmap(cache
->bo
);
416 drm_intel_bo_unreference(cache
->bo
);
418 brw_clear_cache(brw
, cache
);
426 brw_destroy_caches(struct brw_context
*brw
)
428 brw_destroy_cache(brw
, &brw
->cache
);