Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
4358 | Serge | 1 | /********************************************************** |
2 | * Copyright 2008-2009 VMware, Inc. All rights reserved. |
||
3 | * |
||
4 | * Permission is hereby granted, free of charge, to any person |
||
5 | * obtaining a copy of this software and associated documentation |
||
6 | * files (the "Software"), to deal in the Software without |
||
7 | * restriction, including without limitation the rights to use, copy, |
||
8 | * modify, merge, publish, distribute, sublicense, and/or sell copies |
||
9 | * of the Software, and to permit persons to whom the Software is |
||
10 | * furnished to do so, subject to the following conditions: |
||
11 | * |
||
12 | * The above copyright notice and this permission notice shall be |
||
13 | * included in all copies or substantial portions of the Software. |
||
14 | * |
||
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, |
||
16 | * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF |
||
17 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND |
||
18 | * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS |
||
19 | * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN |
||
20 | * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN |
||
21 | * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
||
22 | * SOFTWARE. |
||
23 | * |
||
24 | **********************************************************/ |
||
25 | |||
26 | #include "util/u_math.h" |
||
27 | #include "util/u_memory.h" |
||
28 | #include "util/u_hash.h" |
||
29 | |||
30 | #include "svga_debug.h" |
||
31 | #include "svga_format.h" |
||
32 | #include "svga_winsys.h" |
||
33 | #include "svga_screen.h" |
||
34 | #include "svga_screen_cache.h" |
||
35 | |||
36 | |||
37 | #define SVGA_SURFACE_CACHE_ENABLED 1 |
||
38 | |||
39 | |||
40 | /** |
||
41 | * Return the size of the surface described by the key (in bytes). |
||
42 | */ |
||
43 | static unsigned |
||
44 | surface_size(const struct svga_host_surface_cache_key *key) |
||
45 | { |
||
46 | unsigned bw, bh, bpb, total_size, i; |
||
47 | |||
48 | assert(key->numMipLevels > 0); |
||
49 | assert(key->numFaces > 0); |
||
50 | |||
51 | if (key->format == SVGA3D_BUFFER) { |
||
52 | /* Special case: we don't want to count vertex/index buffers |
||
53 | * against the cache size limit, so view them as zero-sized. |
||
54 | */ |
||
55 | return 0; |
||
56 | } |
||
57 | |||
58 | svga_format_size(key->format, &bw, &bh, &bpb); |
||
59 | |||
60 | total_size = 0; |
||
61 | |||
62 | for (i = 0; i < key->numMipLevels; i++) { |
||
63 | unsigned w = u_minify(key->size.width, i); |
||
64 | unsigned h = u_minify(key->size.height, i); |
||
65 | unsigned d = u_minify(key->size.depth, i); |
||
66 | unsigned img_size = ((w + bw - 1) / bw) * ((h + bh - 1) / bh) * d * bpb; |
||
67 | total_size += img_size; |
||
68 | } |
||
69 | |||
70 | total_size *= key->numFaces; |
||
71 | |||
72 | return total_size; |
||
73 | } |
||
74 | |||
75 | |||
76 | /** |
||
77 | * Compute the bucket for this key. |
||
78 | */ |
||
79 | static INLINE unsigned |
||
80 | svga_screen_cache_bucket(const struct svga_host_surface_cache_key *key) |
||
81 | { |
||
82 | return util_hash_crc32(key, sizeof *key) % SVGA_HOST_SURFACE_CACHE_BUCKETS; |
||
83 | } |
||
84 | |||
85 | |||
86 | /** |
||
87 | * Search the cache for a surface that matches the key. If a match is |
||
88 | * found, remove it from the cache and return the surface pointer. |
||
89 | * Return NULL otherwise. |
||
90 | */ |
||
91 | static INLINE struct svga_winsys_surface * |
||
92 | svga_screen_cache_lookup(struct svga_screen *svgascreen, |
||
93 | const struct svga_host_surface_cache_key *key) |
||
94 | { |
||
95 | struct svga_host_surface_cache *cache = &svgascreen->cache; |
||
96 | struct svga_winsys_screen *sws = svgascreen->sws; |
||
97 | struct svga_host_surface_cache_entry *entry; |
||
98 | struct svga_winsys_surface *handle = NULL; |
||
99 | struct list_head *curr, *next; |
||
100 | unsigned bucket; |
||
101 | unsigned tries = 0; |
||
102 | |||
103 | assert(key->cachable); |
||
104 | |||
105 | bucket = svga_screen_cache_bucket(key); |
||
106 | |||
107 | pipe_mutex_lock(cache->mutex); |
||
108 | |||
109 | curr = cache->bucket[bucket].next; |
||
110 | next = curr->next; |
||
111 | while (curr != &cache->bucket[bucket]) { |
||
112 | ++tries; |
||
113 | |||
114 | entry = LIST_ENTRY(struct svga_host_surface_cache_entry, curr, bucket_head); |
||
115 | |||
116 | assert(entry->handle); |
||
117 | |||
118 | if (memcmp(&entry->key, key, sizeof *key) == 0 && |
||
119 | sws->fence_signalled(sws, entry->fence, 0) == 0) { |
||
120 | unsigned surf_size; |
||
121 | |||
122 | assert(sws->surface_is_flushed(sws, entry->handle)); |
||
123 | |||
124 | handle = entry->handle; /* Reference is transfered here. */ |
||
125 | entry->handle = NULL; |
||
126 | |||
127 | LIST_DEL(&entry->bucket_head); |
||
128 | |||
129 | LIST_DEL(&entry->head); |
||
130 | |||
131 | LIST_ADD(&entry->head, &cache->empty); |
||
132 | |||
133 | /* update the cache size */ |
||
134 | surf_size = surface_size(&entry->key); |
||
135 | assert(surf_size <= cache->total_size); |
||
136 | if (surf_size > cache->total_size) |
||
137 | cache->total_size = 0; /* should never happen, but be safe */ |
||
138 | else |
||
139 | cache->total_size -= surf_size; |
||
140 | |||
141 | break; |
||
142 | } |
||
143 | |||
144 | curr = next; |
||
145 | next = curr->next; |
||
146 | } |
||
147 | |||
148 | pipe_mutex_unlock(cache->mutex); |
||
149 | |||
150 | if (SVGA_DEBUG & DEBUG_DMA) |
||
151 | debug_printf("%s: cache %s after %u tries (bucket %d)\n", __FUNCTION__, |
||
152 | handle ? "hit" : "miss", tries, bucket); |
||
153 | |||
154 | return handle; |
||
155 | } |
||
156 | |||
157 | |||
158 | /** |
||
159 | * Free the least recently used entries in the surface cache until the |
||
160 | * cache size is <= the target size OR there are no unused entries left |
||
161 | * to discard. We don't do any flushing to try to free up additional |
||
162 | * surfaces. |
||
163 | */ |
||
164 | static void |
||
165 | svga_screen_cache_shrink(struct svga_screen *svgascreen, |
||
166 | unsigned target_size) |
||
167 | { |
||
168 | struct svga_host_surface_cache *cache = &svgascreen->cache; |
||
169 | struct svga_winsys_screen *sws = svgascreen->sws; |
||
170 | struct svga_host_surface_cache_entry *entry = NULL, *next_entry; |
||
171 | |||
172 | /* Walk over the list of unused buffers in reverse order: from oldest |
||
173 | * to newest. |
||
174 | */ |
||
175 | LIST_FOR_EACH_ENTRY_SAFE_REV(entry, next_entry, &cache->unused, head) { |
||
176 | if (entry->key.format != SVGA3D_BUFFER) { |
||
177 | /* we don't want to discard vertex/index buffers */ |
||
178 | |||
179 | cache->total_size -= surface_size(&entry->key); |
||
180 | |||
181 | assert(entry->handle); |
||
182 | sws->surface_reference(sws, &entry->handle, NULL); |
||
183 | |||
184 | LIST_DEL(&entry->bucket_head); |
||
185 | LIST_DEL(&entry->head); |
||
186 | LIST_ADD(&entry->head, &cache->empty); |
||
187 | |||
188 | if (cache->total_size <= target_size) { |
||
189 | /* all done */ |
||
190 | break; |
||
191 | } |
||
192 | } |
||
193 | } |
||
194 | } |
||
195 | |||
196 | |||
197 | /** |
||
198 | * Transfers a handle reference. |
||
199 | */ |
||
200 | static INLINE void |
||
201 | svga_screen_cache_add(struct svga_screen *svgascreen, |
||
202 | const struct svga_host_surface_cache_key *key, |
||
203 | struct svga_winsys_surface **p_handle) |
||
204 | { |
||
205 | struct svga_host_surface_cache *cache = &svgascreen->cache; |
||
206 | struct svga_winsys_screen *sws = svgascreen->sws; |
||
207 | struct svga_host_surface_cache_entry *entry = NULL; |
||
208 | struct svga_winsys_surface *handle = *p_handle; |
||
209 | unsigned surf_size; |
||
210 | |||
211 | assert(key->cachable); |
||
212 | |||
213 | assert(handle); |
||
214 | if (!handle) |
||
215 | return; |
||
216 | |||
217 | surf_size = surface_size(key); |
||
218 | |||
219 | *p_handle = NULL; |
||
220 | pipe_mutex_lock(cache->mutex); |
||
221 | |||
222 | if (surf_size >= SVGA_HOST_SURFACE_CACHE_BYTES) { |
||
223 | /* this surface is too large to cache, just free it */ |
||
224 | sws->surface_reference(sws, &handle, NULL); |
||
225 | pipe_mutex_unlock(cache->mutex); |
||
226 | return; |
||
227 | } |
||
228 | |||
229 | if (cache->total_size + surf_size > SVGA_HOST_SURFACE_CACHE_BYTES) { |
||
230 | /* Adding this surface would exceed the cache size. |
||
231 | * Try to discard least recently used entries until we hit the |
||
232 | * new target cache size. |
||
233 | */ |
||
234 | unsigned target_size = SVGA_HOST_SURFACE_CACHE_BYTES - surf_size; |
||
235 | |||
236 | svga_screen_cache_shrink(svgascreen, target_size); |
||
237 | |||
238 | if (cache->total_size > target_size) { |
||
239 | /* we weren't able to shrink the cache as much as we wanted so |
||
240 | * just discard this surface. |
||
241 | */ |
||
242 | sws->surface_reference(sws, &handle, NULL); |
||
243 | pipe_mutex_unlock(cache->mutex); |
||
244 | return; |
||
245 | } |
||
246 | } |
||
247 | |||
248 | if (!LIST_IS_EMPTY(&cache->empty)) { |
||
249 | /* use the first empty entry */ |
||
250 | entry = LIST_ENTRY(struct svga_host_surface_cache_entry, |
||
251 | cache->empty.next, head); |
||
252 | |||
253 | LIST_DEL(&entry->head); |
||
254 | } |
||
255 | else if (!LIST_IS_EMPTY(&cache->unused)) { |
||
256 | /* free the last used buffer and reuse its entry */ |
||
257 | entry = LIST_ENTRY(struct svga_host_surface_cache_entry, |
||
258 | cache->unused.prev, head); |
||
259 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
260 | "unref sid %p (make space)\n", entry->handle); |
||
261 | |||
262 | cache->total_size -= surface_size(&entry->key); |
||
263 | |||
264 | sws->surface_reference(sws, &entry->handle, NULL); |
||
265 | |||
266 | LIST_DEL(&entry->bucket_head); |
||
267 | |||
268 | LIST_DEL(&entry->head); |
||
269 | } |
||
270 | |||
271 | if (entry) { |
||
272 | entry->handle = handle; |
||
273 | memcpy(&entry->key, key, sizeof entry->key); |
||
274 | |||
275 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
276 | "cache sid %p\n", entry->handle); |
||
277 | LIST_ADD(&entry->head, &cache->validated); |
||
278 | |||
279 | cache->total_size += surf_size; |
||
280 | } |
||
281 | else { |
||
282 | /* Couldn't cache the buffer -- this really shouldn't happen */ |
||
283 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
284 | "unref sid %p (couldn't find space)\n", handle); |
||
285 | sws->surface_reference(sws, &handle, NULL); |
||
286 | } |
||
287 | |||
288 | pipe_mutex_unlock(cache->mutex); |
||
289 | } |
||
290 | |||
291 | |||
292 | /** |
||
293 | * Called during the screen flush to move all buffers not in a validate list |
||
294 | * into the unused list. |
||
295 | */ |
||
296 | void |
||
297 | svga_screen_cache_flush(struct svga_screen *svgascreen, |
||
298 | struct pipe_fence_handle *fence) |
||
299 | { |
||
300 | struct svga_host_surface_cache *cache = &svgascreen->cache; |
||
301 | struct svga_winsys_screen *sws = svgascreen->sws; |
||
302 | struct svga_host_surface_cache_entry *entry; |
||
303 | struct list_head *curr, *next; |
||
304 | unsigned bucket; |
||
305 | |||
306 | pipe_mutex_lock(cache->mutex); |
||
307 | |||
308 | curr = cache->validated.next; |
||
309 | next = curr->next; |
||
310 | while (curr != &cache->validated) { |
||
311 | entry = LIST_ENTRY(struct svga_host_surface_cache_entry, curr, head); |
||
312 | |||
313 | assert(entry->handle); |
||
314 | |||
315 | if (sws->surface_is_flushed(sws, entry->handle)) { |
||
316 | LIST_DEL(&entry->head); |
||
317 | |||
318 | svgascreen->sws->fence_reference(svgascreen->sws, &entry->fence, fence); |
||
319 | |||
320 | LIST_ADD(&entry->head, &cache->unused); |
||
321 | |||
322 | bucket = svga_screen_cache_bucket(&entry->key); |
||
323 | LIST_ADD(&entry->bucket_head, &cache->bucket[bucket]); |
||
324 | } |
||
325 | |||
326 | curr = next; |
||
327 | next = curr->next; |
||
328 | } |
||
329 | |||
330 | pipe_mutex_unlock(cache->mutex); |
||
331 | } |
||
332 | |||
333 | |||
334 | /** |
||
335 | * Free all the surfaces in the cache. |
||
336 | * Called when destroying the svga screen object. |
||
337 | */ |
||
338 | void |
||
339 | svga_screen_cache_cleanup(struct svga_screen *svgascreen) |
||
340 | { |
||
341 | struct svga_host_surface_cache *cache = &svgascreen->cache; |
||
342 | struct svga_winsys_screen *sws = svgascreen->sws; |
||
343 | unsigned i; |
||
344 | |||
345 | for (i = 0; i < SVGA_HOST_SURFACE_CACHE_SIZE; ++i) { |
||
346 | if (cache->entries[i].handle) { |
||
347 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
348 | "unref sid %p (shutdown)\n", cache->entries[i].handle); |
||
349 | sws->surface_reference(sws, &cache->entries[i].handle, NULL); |
||
350 | |||
351 | cache->total_size -= surface_size(&cache->entries[i].key); |
||
352 | } |
||
353 | |||
354 | if (cache->entries[i].fence) |
||
355 | svgascreen->sws->fence_reference(svgascreen->sws, |
||
356 | &cache->entries[i].fence, NULL); |
||
357 | } |
||
358 | |||
359 | pipe_mutex_destroy(cache->mutex); |
||
360 | } |
||
361 | |||
362 | |||
363 | enum pipe_error |
||
364 | svga_screen_cache_init(struct svga_screen *svgascreen) |
||
365 | { |
||
366 | struct svga_host_surface_cache *cache = &svgascreen->cache; |
||
367 | unsigned i; |
||
368 | |||
369 | assert(cache->total_size == 0); |
||
370 | |||
371 | pipe_mutex_init(cache->mutex); |
||
372 | |||
373 | for (i = 0; i < SVGA_HOST_SURFACE_CACHE_BUCKETS; ++i) |
||
374 | LIST_INITHEAD(&cache->bucket[i]); |
||
375 | |||
376 | LIST_INITHEAD(&cache->unused); |
||
377 | |||
378 | LIST_INITHEAD(&cache->validated); |
||
379 | |||
380 | LIST_INITHEAD(&cache->empty); |
||
381 | for (i = 0; i < SVGA_HOST_SURFACE_CACHE_SIZE; ++i) |
||
382 | LIST_ADDTAIL(&cache->entries[i].head, &cache->empty); |
||
383 | |||
384 | return PIPE_OK; |
||
385 | } |
||
386 | |||
387 | |||
388 | /** |
||
389 | * Allocate a new host-side surface. If the surface is marked as cachable, |
||
390 | * first try re-using a surface in the cache of freed surfaces. Otherwise, |
||
391 | * allocate a new surface. |
||
392 | */ |
||
393 | struct svga_winsys_surface * |
||
394 | svga_screen_surface_create(struct svga_screen *svgascreen, |
||
395 | struct svga_host_surface_cache_key *key) |
||
396 | { |
||
397 | struct svga_winsys_screen *sws = svgascreen->sws; |
||
398 | struct svga_winsys_surface *handle = NULL; |
||
399 | boolean cachable = SVGA_SURFACE_CACHE_ENABLED && key->cachable; |
||
400 | |||
401 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
402 | "%s sz %dx%dx%d mips %d faces %d cachable %d\n", |
||
403 | __FUNCTION__, |
||
404 | key->size.width, |
||
405 | key->size.height, |
||
406 | key->size.depth, |
||
407 | key->numMipLevels, |
||
408 | key->numFaces, |
||
409 | key->cachable); |
||
410 | |||
411 | if (cachable) { |
||
412 | if (key->format == SVGA3D_BUFFER) { |
||
413 | /* For buffers, round the buffer size up to the nearest power |
||
414 | * of two to increase the probability of cache hits. Keep |
||
415 | * texture surface dimensions unchanged. |
||
416 | */ |
||
417 | uint32_t size = 1; |
||
418 | while (size < key->size.width) |
||
419 | size <<= 1; |
||
420 | key->size.width = size; |
||
421 | /* Since we're reusing buffers we're effectively transforming all |
||
422 | * of them into dynamic buffers. |
||
423 | * |
||
424 | * It would be nice to not cache long lived static buffers. But there |
||
425 | * is no way to detect the long lived from short lived ones yet. A |
||
426 | * good heuristic would be buffer size. |
||
427 | */ |
||
428 | key->flags &= ~SVGA3D_SURFACE_HINT_STATIC; |
||
429 | key->flags |= SVGA3D_SURFACE_HINT_DYNAMIC; |
||
430 | } |
||
431 | |||
432 | handle = svga_screen_cache_lookup(svgascreen, key); |
||
433 | if (handle) { |
||
434 | if (key->format == SVGA3D_BUFFER) |
||
435 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
436 | "reuse sid %p sz %d (buffer)\n", handle, |
||
437 | key->size.width); |
||
438 | else |
||
439 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
440 | "reuse sid %p sz %dx%dx%d mips %d faces %d\n", handle, |
||
441 | key->size.width, |
||
442 | key->size.height, |
||
443 | key->size.depth, |
||
444 | key->numMipLevels, |
||
445 | key->numFaces); |
||
446 | } |
||
447 | } |
||
448 | |||
449 | if (!handle) { |
||
450 | handle = sws->surface_create(sws, |
||
451 | key->flags, |
||
452 | key->format, |
||
453 | key->size, |
||
454 | key->numFaces, |
||
455 | key->numMipLevels); |
||
456 | if (handle) |
||
457 | SVGA_DBG(DEBUG_CACHE|DEBUG_DMA, |
||
458 | " CREATE sid %p sz %dx%dx%d\n", |
||
459 | handle, |
||
460 | key->size.width, |
||
461 | key->size.height, |
||
462 | key->size.depth); |
||
463 | } |
||
464 | |||
465 | return handle; |
||
466 | } |
||
467 | |||
468 | |||
469 | /** |
||
470 | * Release a surface. We don't actually free the surface- we put |
||
471 | * it into the cache of freed surfaces (if it's cachable). |
||
472 | */ |
||
473 | void |
||
474 | svga_screen_surface_destroy(struct svga_screen *svgascreen, |
||
475 | const struct svga_host_surface_cache_key *key, |
||
476 | struct svga_winsys_surface **p_handle) |
||
477 | { |
||
478 | struct svga_winsys_screen *sws = svgascreen->sws; |
||
479 | |||
480 | /* We only set the cachable flag for surfaces of which we are the |
||
481 | * exclusive owner. So just hold onto our existing reference in |
||
482 | * that case. |
||
483 | */ |
||
484 | if (SVGA_SURFACE_CACHE_ENABLED && key->cachable) { |
||
485 | svga_screen_cache_add(svgascreen, key, p_handle); |
||
486 | } |
||
487 | else { |
||
488 | SVGA_DBG(DEBUG_DMA, |
||
489 | "unref sid %p (uncachable)\n", *p_handle); |
||
490 | sws->surface_reference(sws, p_handle, NULL); |
||
491 | } |
||
492 | } |
||
493 | |||
494 | |||
495 | /** |
||
496 | * Print/dump the contents of the screen cache. For debugging. |
||
497 | */ |
||
498 | void |
||
499 | svga_screen_cache_dump(const struct svga_screen *svgascreen) |
||
500 | { |
||
501 | const struct svga_host_surface_cache *cache = &svgascreen->cache; |
||
502 | unsigned bucket; |
||
503 | unsigned count = 0; |
||
504 | |||
505 | debug_printf("svga3d surface cache:\n"); |
||
506 | for (bucket = 0; bucket < SVGA_HOST_SURFACE_CACHE_BUCKETS; bucket++) { |
||
507 | struct list_head *curr; |
||
508 | curr = cache->bucket[bucket].next; |
||
509 | while (curr && curr != &cache->bucket[bucket]) { |
||
510 | struct svga_host_surface_cache_entry *entry = |
||
511 | LIST_ENTRY(struct svga_host_surface_cache_entry, |
||
512 | curr, bucket_head); |
||
513 | if (entry->key.format != 37) { |
||
514 | debug_printf(" %u x %u x %u format %u\n", |
||
515 | entry->key.size.width, |
||
516 | entry->key.size.height, |
||
517 | entry->key.size.depth, |
||
518 | entry->key.format); |
||
519 | } |
||
520 | curr = curr->next; |
||
521 | count++; |
||
522 | } |
||
523 | } |
||
524 | |||
525 | debug_printf("%u surfaces, %u bytes\n", count, cache->total_size); |
||
526 | }>=><=>>>>>=>=>=>> |