i965: Always pass the size argument to brw_cache_data.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_state_cache.c
1 /*
2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
5
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:
13
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.
17
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.
25
26 **********************************************************************/
27 /*
28 * Authors:
29 * Keith Whitwell <keith@tungstengraphics.com>
30 */
31
32 /** @file brw_state_cache.c
33 *
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
38 * data) in return.
39 *
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).
45 *
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.
49 *
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.
57 */
58
59 #include "main/imports.h"
60 #include "brw_state.h"
61 #include "intel_batchbuffer.h"
62
63 /* XXX: Fixme - have to include these to get the sizes of the prog_key
64 * structs:
65 */
66 #include "brw_wm.h"
67 #include "brw_vs.h"
68 #include "brw_clip.h"
69 #include "brw_sf.h"
70 #include "brw_gs.h"
71
72
73 static GLuint
74 hash_key(const void *key, GLuint key_size,
75 dri_bo **reloc_bufs, GLuint nr_reloc_bufs)
76 {
77 GLuint *ikey = (GLuint *)key;
78 GLuint hash = 0, i;
79
80 assert(key_size % 4 == 0);
81
82 /* I'm sure this can be improved on:
83 */
84 for (i = 0; i < key_size/4; i++) {
85 hash ^= ikey[i];
86 hash = (hash << 5) | (hash >> 27);
87 }
88
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++) {
93 hash ^= ikey[i];
94 hash = (hash << 5) | (hash >> 27);
95 }
96
97 return hash;
98 }
99
100
101 /**
102 * Marks a new buffer as being chosen for the given cache id.
103 */
104 static void
105 update_cache_last(struct brw_cache *cache, enum brw_cache_id cache_id,
106 dri_bo *bo)
107 {
108 if (bo == cache->last_bo[cache_id])
109 return; /* no change */
110
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;
115 }
116
117
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)
122 {
123 struct brw_cache_item *c;
124
125 #if 0
126 int bucketcount = 0;
127
128 for (c = cache->items[hash % cache->size]; c; c = c->next)
129 bucketcount++;
130
131 fprintf(stderr, "bucket %d/%d = %d/%d items\n", hash % cache->size,
132 cache->size, bucketcount, cache->n_items);
133 #endif
134
135 for (c = cache->items[hash % cache->size]; c; c = c->next) {
136 if (c->cache_id == cache_id &&
137 c->hash == hash &&
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)
143 return c;
144 }
145
146 return NULL;
147 }
148
149
150 static void
151 rehash(struct brw_cache *cache)
152 {
153 struct brw_cache_item **items;
154 struct brw_cache_item *c, *next;
155 GLuint size, i;
156
157 size = cache->size * 3;
158 items = (struct brw_cache_item**) _mesa_calloc(size * sizeof(*items));
159
160 for (i = 0; i < cache->size; i++)
161 for (c = cache->items[i]; c; c = next) {
162 next = c->next;
163 c->next = items[c->hash % size];
164 items[c->hash % size] = c;
165 }
166
167 FREE(cache->items);
168 cache->items = items;
169 cache->size = size;
170 }
171
172
173 /**
174 * Returns the buffer object matching cache_id and key, or NULL.
175 */
176 dri_bo *
177 brw_search_cache(struct brw_cache *cache,
178 enum brw_cache_id cache_id,
179 const void *key,
180 GLuint key_size,
181 dri_bo **reloc_bufs, GLuint nr_reloc_bufs,
182 void *aux_return)
183 {
184 struct brw_cache_item *item;
185 GLuint hash = hash_key(key, key_size, reloc_bufs, nr_reloc_bufs);
186
187 item = search_cache(cache, cache_id, hash, key, key_size,
188 reloc_bufs, nr_reloc_bufs);
189
190 if (item == NULL)
191 return NULL;
192
193 if (aux_return)
194 *(void **)aux_return = (void *)((char *)item->key + item->key_size);
195
196 update_cache_last(cache, cache_id, item->bo);
197
198 dri_bo_reference(item->bo);
199 return item->bo;
200 }
201
202
203 dri_bo *
204 brw_upload_cache( struct brw_cache *cache,
205 enum brw_cache_id cache_id,
206 const void *key,
207 GLuint key_size,
208 dri_bo **reloc_bufs,
209 GLuint nr_reloc_bufs,
210 const void *data,
211 GLuint data_size,
212 const void *aux,
213 void *aux_return )
214 {
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];
219 void *tmp;
220 dri_bo *bo;
221 int i;
222
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);
226
227
228 /* Set up the memory containing the key, aux_data, and reloc_bufs */
229 tmp = _mesa_malloc(key_size + aux_size + relocs_size);
230
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]);
237 }
238
239 item->cache_id = cache_id;
240 item->key = tmp;
241 item->hash = hash;
242 item->key_size = key_size;
243 item->reloc_bufs = tmp + key_size + aux_size;
244 item->nr_reloc_bufs = nr_reloc_bufs;
245
246 item->bo = bo;
247 dri_bo_reference(bo);
248 item->data_size = data_size;
249
250 if (cache->n_items > cache->size * 1.5)
251 rehash(cache);
252
253 hash %= cache->size;
254 item->next = cache->items[hash];
255 cache->items[hash] = item;
256 cache->n_items++;
257
258 if (aux_return) {
259 assert(cache->aux_size[cache_id]);
260 *(void **)aux_return = (void *)((char *)item->key + item->key_size);
261 }
262
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);
267
268 /* Copy data to the buffer */
269 dri_bo_subdata(bo, 0, data_size, data);
270
271 update_cache_last(cache, cache_id, bo);
272
273 return bo;
274 }
275
276
277 /**
278 * Wrapper around brw_cache_data_sz using the cache_id's canonical key size.
279 *
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.
283 *
284 * If aux data is involved, use search/upload instead.
285
286 */
287 dri_bo *
288 brw_cache_data(struct brw_cache *cache,
289 enum brw_cache_id cache_id,
290 const void *data,
291 GLuint data_size,
292 dri_bo **reloc_bufs,
293 GLuint nr_reloc_bufs)
294 {
295 dri_bo *bo;
296 struct brw_cache_item *item;
297 GLuint hash = hash_key(data, data_size, reloc_bufs, nr_reloc_bufs);
298
299 item = search_cache(cache, cache_id, hash, data, data_size,
300 reloc_bufs, nr_reloc_bufs);
301 if (item) {
302 update_cache_last(cache, cache_id, item->bo);
303 dri_bo_reference(item->bo);
304 return item->bo;
305 }
306
307 bo = brw_upload_cache(cache, cache_id,
308 data, data_size,
309 reloc_bufs, nr_reloc_bufs,
310 data, data_size,
311 NULL, NULL);
312
313 return bo;
314 }
315
316 enum pool_type {
317 DW_SURFACE_STATE,
318 DW_GENERAL_STATE
319 };
320
321
322 static void
323 brw_init_cache_id(struct brw_cache *cache,
324 const char *name,
325 enum brw_cache_id id,
326 GLuint aux_size)
327 {
328 cache->name[id] = strdup(name);
329 cache->aux_size[id] = aux_size;
330 }
331
332
333 static void
334 brw_init_non_surface_cache(struct brw_context *brw)
335 {
336 struct brw_cache *cache = &brw->cache;
337
338 cache->brw = brw;
339
340 cache->size = 7;
341 cache->n_items = 0;
342 cache->items = (struct brw_cache_item **)
343 _mesa_calloc(cache->size * sizeof(struct brw_cache_item));
344
345 brw_init_cache_id(cache,
346 "CC_VP",
347 BRW_CC_VP,
348 0);
349
350 brw_init_cache_id(cache,
351 "CC_UNIT",
352 BRW_CC_UNIT,
353 0);
354
355 brw_init_cache_id(cache,
356 "WM_PROG",
357 BRW_WM_PROG,
358 sizeof(struct brw_wm_prog_data));
359
360 brw_init_cache_id(cache,
361 "SAMPLER_DEFAULT_COLOR",
362 BRW_SAMPLER_DEFAULT_COLOR,
363 0);
364
365 brw_init_cache_id(cache,
366 "SAMPLER",
367 BRW_SAMPLER,
368 0);
369
370 brw_init_cache_id(cache,
371 "WM_UNIT",
372 BRW_WM_UNIT,
373 0);
374
375 brw_init_cache_id(cache,
376 "SF_PROG",
377 BRW_SF_PROG,
378 sizeof(struct brw_sf_prog_data));
379
380 brw_init_cache_id(cache,
381 "SF_VP",
382 BRW_SF_VP,
383 0);
384
385 brw_init_cache_id(cache,
386 "SF_UNIT",
387 BRW_SF_UNIT,
388 0);
389
390 brw_init_cache_id(cache,
391 "VS_UNIT",
392 BRW_VS_UNIT,
393 0);
394
395 brw_init_cache_id(cache,
396 "VS_PROG",
397 BRW_VS_PROG,
398 sizeof(struct brw_vs_prog_data));
399
400 brw_init_cache_id(cache,
401 "CLIP_UNIT",
402 BRW_CLIP_UNIT,
403 0);
404
405 brw_init_cache_id(cache,
406 "CLIP_PROG",
407 BRW_CLIP_PROG,
408 sizeof(struct brw_clip_prog_data));
409
410 brw_init_cache_id(cache,
411 "GS_UNIT",
412 BRW_GS_UNIT,
413 0);
414
415 brw_init_cache_id(cache,
416 "GS_PROG",
417 BRW_GS_PROG,
418 sizeof(struct brw_gs_prog_data));
419 }
420
421
422 static void
423 brw_init_surface_cache(struct brw_context *brw)
424 {
425 struct brw_cache *cache = &brw->surface_cache;
426
427 cache->brw = brw;
428
429 cache->size = 7;
430 cache->n_items = 0;
431 cache->items = (struct brw_cache_item **)
432 _mesa_calloc(cache->size * sizeof(struct brw_cache_item));
433
434 brw_init_cache_id(cache,
435 "SS_SURFACE",
436 BRW_SS_SURFACE,
437 0);
438
439 brw_init_cache_id(cache,
440 "SS_SURF_BIND",
441 BRW_SS_SURF_BIND,
442 0);
443 }
444
445
446 void
447 brw_init_caches(struct brw_context *brw)
448 {
449 brw_init_non_surface_cache(brw);
450 brw_init_surface_cache(brw);
451 }
452
453
454 static void
455 brw_clear_cache(struct brw_context *brw, struct brw_cache *cache)
456 {
457 struct brw_cache_item *c, *next;
458 GLuint i;
459
460 if (INTEL_DEBUG & DEBUG_STATE)
461 _mesa_printf("%s\n", __FUNCTION__);
462
463 for (i = 0; i < cache->size; i++) {
464 for (c = cache->items[i]; c; c = next) {
465 int j;
466
467 next = 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);
472 free(c);
473 }
474 cache->items[i] = NULL;
475 }
476
477 cache->n_items = 0;
478
479 if (brw->curbe.last_buf) {
480 _mesa_free(brw->curbe.last_buf);
481 brw->curbe.last_buf = NULL;
482 }
483
484 brw->state.dirty.mesa |= ~0;
485 brw->state.dirty.brw |= ~0;
486 brw->state.dirty.cache |= ~0;
487 }
488
489 /* Clear all entries from the cache that point to the given bo.
490 *
491 * This lets us release memory for reuse earlier for known-dead buffers,
492 * at the cost of walking the entire hash table.
493 */
494 void
495 brw_state_cache_bo_delete(struct brw_cache *cache, dri_bo *bo)
496 {
497 struct brw_cache_item **prev;
498 GLuint i;
499
500 if (INTEL_DEBUG & DEBUG_STATE)
501 _mesa_printf("%s\n", __FUNCTION__);
502
503 for (i = 0; i < cache->size; i++) {
504 for (prev = &cache->items[i]; *prev;) {
505 struct brw_cache_item *c = *prev;
506
507 if (drm_intel_bo_references(c->bo, bo)) {
508 int j;
509
510 *prev = c->next;
511
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);
516 free(c);
517 cache->n_items--;
518 } else {
519 prev = &c->next;
520 }
521 }
522 }
523 }
524
525 void
526 brw_state_cache_check_size(struct brw_context *brw)
527 {
528 if (INTEL_DEBUG & DEBUG_STATE)
529 _mesa_printf("%s (n_items=%d)\n", __FUNCTION__, brw->cache.n_items);
530
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.
533 */
534 if (brw->cache.n_items > 1000)
535 brw_clear_cache(brw, &brw->cache);
536
537 if (brw->surface_cache.n_items > 1000)
538 brw_clear_cache(brw, &brw->surface_cache);
539 }
540
541
542 static void
543 brw_destroy_cache(struct brw_context *brw, struct brw_cache *cache)
544 {
545 GLuint i;
546
547 if (INTEL_DEBUG & DEBUG_STATE)
548 _mesa_printf("%s\n", __FUNCTION__);
549
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]);
554 }
555 free(cache->items);
556 cache->items = NULL;
557 cache->size = 0;
558 }
559
560
561 void
562 brw_destroy_caches(struct brw_context *brw)
563 {
564 brw_destroy_cache(brw, &brw->cache);
565 brw_destroy_cache(brw, &brw->surface_cache);
566 }