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_program_cache.c
34 * This file implements a simple program cache for 965. The consumers can
35 * query the hash table of programs using a cache_id and program key, and
36 * receive the corresponding program buffer object (plus associated auxiliary
37 * data) in return. Objects in the cache may not have relocations
38 * (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 "main/streaming-load-memcpy.h"
49 #include "x86/common_x86_asm.h"
50 #include "intel_batchbuffer.h"
51 #include "brw_state.h"
55 #include "brw_program.h"
56 #include "compiler/brw_eu.h"
58 #define FILE_DEBUG_FLAG DEBUG_STATE
60 struct brw_cache_item
{
62 * Effectively part of the key, cache_id identifies what kind of state
63 * buffer is involved, and also which dirty flag should set.
65 enum brw_cache_id cache_id
;
67 /** 32-bit hash of the key data */
70 /** for variable-sized keys */
78 struct brw_cache_item
*next
;
82 get_program_string_id(enum brw_cache_id cache_id
, const void *key
)
85 case BRW_CACHE_VS_PROG
:
86 return ((struct brw_vs_prog_key
*) key
)->program_string_id
;
87 case BRW_CACHE_TCS_PROG
:
88 return ((struct brw_tcs_prog_key
*) key
)->program_string_id
;
89 case BRW_CACHE_TES_PROG
:
90 return ((struct brw_tes_prog_key
*) key
)->program_string_id
;
91 case BRW_CACHE_GS_PROG
:
92 return ((struct brw_gs_prog_key
*) key
)->program_string_id
;
93 case BRW_CACHE_CS_PROG
:
94 return ((struct brw_cs_prog_key
*) key
)->program_string_id
;
95 case BRW_CACHE_FS_PROG
:
96 return ((struct brw_wm_prog_key
*) key
)->program_string_id
;
98 unreachable("no program string id for this kind of program");
103 hash_key(struct brw_cache_item
*item
)
105 GLuint
*ikey
= (GLuint
*)item
->key
;
106 GLuint hash
= item
->cache_id
, i
;
108 assert(item
->key_size
% 4 == 0);
110 /* I'm sure this can be improved on:
112 for (i
= 0; i
< item
->key_size
/4; i
++) {
114 hash
= (hash
<< 5) | (hash
>> 27);
121 brw_cache_item_equals(const struct brw_cache_item
*a
,
122 const struct brw_cache_item
*b
)
124 return a
->cache_id
== b
->cache_id
&&
125 a
->hash
== b
->hash
&&
126 a
->key_size
== b
->key_size
&&
127 (memcmp(a
->key
, b
->key
, a
->key_size
) == 0);
130 static struct brw_cache_item
*
131 search_cache(struct brw_cache
*cache
, GLuint hash
,
132 struct brw_cache_item
*lookup
)
134 struct brw_cache_item
*c
;
139 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
)
142 fprintf(stderr
, "bucket %d/%d = %d/%d items\n", hash
% cache
->size
,
143 cache
->size
, bucketcount
, cache
->n_items
);
146 for (c
= cache
->items
[hash
% cache
->size
]; c
; c
= c
->next
) {
147 if (brw_cache_item_equals(lookup
, c
))
156 rehash(struct brw_cache
*cache
)
158 struct brw_cache_item
**items
;
159 struct brw_cache_item
*c
, *next
;
162 size
= cache
->size
* 3;
163 items
= calloc(size
, sizeof(*items
));
165 for (i
= 0; i
< cache
->size
; i
++)
166 for (c
= cache
->items
[i
]; c
; c
= next
) {
168 c
->next
= items
[c
->hash
% size
];
169 items
[c
->hash
% size
] = c
;
173 cache
->items
= items
;
179 * Returns the buffer object matching cache_id and key, or NULL.
182 brw_search_cache(struct brw_cache
*cache
,
183 enum brw_cache_id cache_id
,
184 const void *key
, GLuint key_size
,
185 uint32_t *inout_offset
, void *inout_aux
)
187 struct brw_context
*brw
= cache
->brw
;
188 struct brw_cache_item
*item
;
189 struct brw_cache_item lookup
;
192 lookup
.cache_id
= cache_id
;
194 lookup
.key_size
= key_size
;
195 hash
= hash_key(&lookup
);
198 item
= search_cache(cache
, hash
, &lookup
);
203 void *aux
= ((char *) item
->key
) + item
->key_size
;
205 if (item
->offset
!= *inout_offset
|| aux
!= *((void **) inout_aux
)) {
206 brw
->ctx
.NewDriverState
|= (1 << cache_id
);
207 *inout_offset
= item
->offset
;
208 *((void **) inout_aux
) = aux
;
215 brw_cache_new_bo(struct brw_cache
*cache
, uint32_t new_size
)
217 struct brw_context
*brw
= cache
->brw
;
218 struct brw_bo
*new_bo
;
220 perf_debug("Copying to larger program cache: %u kB -> %u kB\n",
221 (unsigned) cache
->bo
->size
/ 1024, new_size
/ 1024);
223 new_bo
= brw_bo_alloc(brw
->bufmgr
, "program cache", new_size
, 64);
224 if (can_do_exec_capture(brw
->screen
))
225 new_bo
->kflags
= EXEC_OBJECT_CAPTURE
;
227 void *map
= brw_bo_map(brw
, new_bo
, MAP_READ
| MAP_WRITE
|
228 MAP_ASYNC
| MAP_PERSISTENT
);
230 /* Copy any existing data that needs to be saved. */
231 if (cache
->next_offset
!= 0) {
233 if (!cache
->bo
->cache_coherent
&& cpu_has_sse4_1
)
234 _mesa_streaming_load_memcpy(map
, cache
->map
, cache
->next_offset
);
237 memcpy(map
, cache
->map
, cache
->next_offset
);
240 brw_bo_unmap(cache
->bo
);
241 brw_bo_unreference(cache
->bo
);
245 /* Since we have a new BO in place, we need to signal the units
246 * that depend on it (state base address on gen5+, or unit state before).
248 brw
->ctx
.NewDriverState
|= BRW_NEW_PROGRAM_CACHE
;
249 brw
->batch
.state_base_address_emitted
= false;
253 * Attempts to find an item in the cache with identical data.
255 static const struct brw_cache_item
*
256 brw_lookup_prog(const struct brw_cache
*cache
,
257 enum brw_cache_id cache_id
,
258 const void *data
, unsigned data_size
)
261 const struct brw_cache_item
*item
;
263 for (i
= 0; i
< cache
->size
; i
++) {
264 for (item
= cache
->items
[i
]; item
; item
= item
->next
) {
265 if (item
->cache_id
!= cache_id
|| item
->size
!= data_size
||
266 memcmp(cache
->map
+ item
->offset
, data
, item
->size
) != 0)
277 brw_alloc_item_data(struct brw_cache
*cache
, uint32_t size
)
281 /* Allocate space in the cache BO for our new program. */
282 if (cache
->next_offset
+ size
> cache
->bo
->size
) {
283 uint32_t new_size
= cache
->bo
->size
* 2;
285 while (cache
->next_offset
+ size
> new_size
)
288 brw_cache_new_bo(cache
, new_size
);
291 offset
= cache
->next_offset
;
293 /* Programs are always 64-byte aligned, so set up the next one now */
294 cache
->next_offset
= ALIGN(offset
+ size
, 64);
300 brw_find_previous_compile(struct brw_cache
*cache
,
301 enum brw_cache_id cache_id
,
302 unsigned program_string_id
)
304 for (unsigned i
= 0; i
< cache
->size
; i
++) {
305 for (struct brw_cache_item
*c
= cache
->items
[i
]; c
; c
= c
->next
) {
306 if (c
->cache_id
== cache_id
&&
307 get_program_string_id(cache_id
, c
->key
) == program_string_id
) {
317 brw_upload_cache(struct brw_cache
*cache
,
318 enum brw_cache_id cache_id
,
325 uint32_t *out_offset
,
328 struct brw_cache_item
*item
= CALLOC_STRUCT(brw_cache_item
);
329 const struct brw_cache_item
*matching_data
=
330 brw_lookup_prog(cache
, cache_id
, data
, data_size
);
334 item
->cache_id
= cache_id
;
335 item
->size
= data_size
;
337 item
->key_size
= key_size
;
338 item
->aux_size
= aux_size
;
339 hash
= hash_key(item
);
342 /* If we can find a matching prog in the cache already, then reuse the
343 * existing stuff without creating new copy into the underlying buffer
344 * object. This is notably useful for programs generating shaders at
345 * runtime, where multiple shaders may compile to the same thing in our
349 item
->offset
= matching_data
->offset
;
351 item
->offset
= brw_alloc_item_data(cache
, data_size
);
353 /* Copy data to the buffer */
354 memcpy(cache
->map
+ item
->offset
, data
, data_size
);
357 /* Set up the memory containing the key and aux_data */
358 tmp
= malloc(key_size
+ aux_size
);
360 memcpy(tmp
, key
, key_size
);
361 memcpy(tmp
+ key_size
, aux
, aux_size
);
365 if (cache
->n_items
> cache
->size
* 1.5f
)
369 item
->next
= cache
->items
[hash
];
370 cache
->items
[hash
] = item
;
373 *out_offset
= item
->offset
;
374 *(void **)out_aux
= (void *)((char *)item
->key
+ item
->key_size
);
375 cache
->brw
->ctx
.NewDriverState
|= 1 << cache_id
;
379 brw_init_caches(struct brw_context
*brw
)
381 struct brw_cache
*cache
= &brw
->cache
;
388 calloc(cache
->size
, sizeof(struct brw_cache_item
*));
390 cache
->bo
= brw_bo_alloc(brw
->bufmgr
, "program cache", 16384, 64);
391 if (can_do_exec_capture(brw
->screen
))
392 cache
->bo
->kflags
= EXEC_OBJECT_CAPTURE
;
394 cache
->map
= brw_bo_map(brw
, cache
->bo
, MAP_READ
| MAP_WRITE
|
395 MAP_ASYNC
| MAP_PERSISTENT
);
399 brw_clear_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
401 struct brw_cache_item
*c
, *next
;
404 DBG("%s\n", __func__
);
406 for (i
= 0; i
< cache
->size
; i
++) {
407 for (c
= cache
->items
[i
]; c
; c
= next
) {
409 if (c
->cache_id
== BRW_CACHE_VS_PROG
||
410 c
->cache_id
== BRW_CACHE_TCS_PROG
||
411 c
->cache_id
== BRW_CACHE_TES_PROG
||
412 c
->cache_id
== BRW_CACHE_GS_PROG
||
413 c
->cache_id
== BRW_CACHE_FS_PROG
||
414 c
->cache_id
== BRW_CACHE_CS_PROG
) {
415 const void *item_aux
= c
->key
+ c
->key_size
;
416 brw_stage_prog_data_free(item_aux
);
418 free((void *)c
->key
);
421 cache
->items
[i
] = NULL
;
426 /* Start putting programs into the start of the BO again, since
427 * we'll never find the old results.
429 cache
->next_offset
= 0;
431 /* We need to make sure that the programs get regenerated, since
432 * any offsets leftover in brw_context will no longer be valid.
434 brw
->NewGLState
= ~0;
435 brw
->ctx
.NewDriverState
= ~0ull;
436 brw
->state
.pipelines
[BRW_RENDER_PIPELINE
].mesa
= ~0;
437 brw
->state
.pipelines
[BRW_RENDER_PIPELINE
].brw
= ~0ull;
438 brw
->state
.pipelines
[BRW_COMPUTE_PIPELINE
].mesa
= ~0;
439 brw
->state
.pipelines
[BRW_COMPUTE_PIPELINE
].brw
= ~0ull;
441 /* Also, NULL out any stale program pointers. */
442 brw
->vs
.base
.prog_data
= NULL
;
443 brw
->tcs
.base
.prog_data
= NULL
;
444 brw
->tes
.base
.prog_data
= NULL
;
445 brw
->gs
.base
.prog_data
= NULL
;
446 brw
->wm
.base
.prog_data
= NULL
;
447 brw
->cs
.base
.prog_data
= NULL
;
449 intel_batchbuffer_flush(brw
);
453 brw_program_cache_check_size(struct brw_context
*brw
)
455 /* un-tuned guess. Each object is generally a page, so 2000 of them is 8 MB of
458 if (brw
->cache
.n_items
> 2000) {
459 perf_debug("Exceeded state cache size limit. Clearing the set "
460 "of compiled programs, which will trigger recompiles\n");
461 brw_clear_cache(brw
, &brw
->cache
);
467 brw_destroy_cache(struct brw_context
*brw
, struct brw_cache
*cache
)
470 DBG("%s\n", __func__
);
472 /* This can be NULL if context creation failed early on */
474 brw_bo_unmap(cache
->bo
);
475 brw_bo_unreference(cache
->bo
);
479 brw_clear_cache(brw
, cache
);
487 brw_destroy_caches(struct brw_context
*brw
)
489 brw_destroy_cache(brw
, &brw
->cache
);
493 cache_name(enum brw_cache_id cache_id
)
496 case BRW_CACHE_VS_PROG
:
498 case BRW_CACHE_TCS_PROG
:
500 case BRW_CACHE_TES_PROG
:
502 case BRW_CACHE_FF_GS_PROG
:
503 return "Fixed-function GS kernel";
504 case BRW_CACHE_GS_PROG
:
506 case BRW_CACHE_CLIP_PROG
:
507 return "CLIP kernel";
508 case BRW_CACHE_SF_PROG
:
510 case BRW_CACHE_FS_PROG
:
512 case BRW_CACHE_CS_PROG
:
520 brw_print_program_cache(struct brw_context
*brw
)
522 const struct brw_cache
*cache
= &brw
->cache
;
523 struct brw_cache_item
*item
;
525 for (unsigned i
= 0; i
< cache
->size
; i
++) {
526 for (item
= cache
->items
[i
]; item
; item
= item
->next
) {
527 fprintf(stderr
, "%s:\n", cache_name(i
));
528 brw_disassemble(&brw
->screen
->devinfo
, cache
->map
,
529 item
->offset
, item
->size
, stderr
);