1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
4 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 * Rob Clark <robclark@freedesktop.org>
29 #include "pipe/p_state.h"
30 #include "util/u_memory.h"
31 #include "util/u_inlines.h"
33 #include "freedreno_query_hw.h"
34 #include "freedreno_context.h"
35 #include "freedreno_util.h"
37 struct fd_hw_sample_period
{
38 struct fd_hw_sample
*start
, *end
;
39 struct list_head list
;
42 /* maps query_type to sample provider idx: */
43 static int pidx(unsigned query_type
)
46 case PIPE_QUERY_OCCLUSION_COUNTER
:
48 case PIPE_QUERY_OCCLUSION_PREDICATE
:
55 static struct fd_hw_sample
*
56 get_sample(struct fd_context
*ctx
, struct fd_ringbuffer
*ring
,
59 struct fd_hw_sample
*samp
= NULL
;
60 int idx
= pidx(query_type
);
62 if (!ctx
->sample_cache
[idx
]) {
63 ctx
->sample_cache
[idx
] =
64 ctx
->sample_providers
[idx
]->get_sample(ctx
, ring
);
67 fd_hw_sample_reference(ctx
, &samp
, ctx
->sample_cache
[idx
]);
73 clear_sample_cache(struct fd_context
*ctx
)
77 for (i
= 0; i
< ARRAY_SIZE(ctx
->sample_cache
); i
++)
78 fd_hw_sample_reference(ctx
, &ctx
->sample_cache
[i
], NULL
);
82 is_active(struct fd_hw_query
*hq
, enum fd_render_stage stage
)
84 return !!(hq
->provider
->active
& stage
);
89 resume_query(struct fd_context
*ctx
, struct fd_hw_query
*hq
,
90 struct fd_ringbuffer
*ring
)
93 hq
->period
= util_slab_alloc(&ctx
->sample_period_pool
);
94 list_inithead(&hq
->period
->list
);
95 hq
->period
->start
= get_sample(ctx
, ring
, hq
->base
.type
);
96 /* NOTE: util_slab_alloc() does not zero out the buffer: */
97 hq
->period
->end
= NULL
;
101 pause_query(struct fd_context
*ctx
, struct fd_hw_query
*hq
,
102 struct fd_ringbuffer
*ring
)
104 assert(hq
->period
&& !hq
->period
->end
);
105 hq
->period
->end
= get_sample(ctx
, ring
, hq
->base
.type
);
106 list_addtail(&hq
->period
->list
, &hq
->current_periods
);
111 destroy_periods(struct fd_context
*ctx
, struct list_head
*list
)
113 struct fd_hw_sample_period
*period
, *s
;
114 LIST_FOR_EACH_ENTRY_SAFE(period
, s
, list
, list
) {
115 fd_hw_sample_reference(ctx
, &period
->start
, NULL
);
116 fd_hw_sample_reference(ctx
, &period
->end
, NULL
);
117 list_del(&period
->list
);
118 util_slab_free(&ctx
->sample_period_pool
, period
);
123 fd_hw_destroy_query(struct fd_context
*ctx
, struct fd_query
*q
)
125 struct fd_hw_query
*hq
= fd_hw_query(q
);
127 destroy_periods(ctx
, &hq
->periods
);
128 destroy_periods(ctx
, &hq
->current_periods
);
135 fd_hw_begin_query(struct fd_context
*ctx
, struct fd_query
*q
)
137 struct fd_hw_query
*hq
= fd_hw_query(q
);
141 /* begin_query() should clear previous results: */
142 destroy_periods(ctx
, &hq
->periods
);
144 if (is_active(hq
, ctx
->stage
))
145 resume_query(ctx
, hq
, ctx
->ring
);
149 /* add to active list: */
151 list_addtail(&hq
->list
, &ctx
->active_queries
);
155 fd_hw_end_query(struct fd_context
*ctx
, struct fd_query
*q
)
157 struct fd_hw_query
*hq
= fd_hw_query(q
);
160 if (is_active(hq
, ctx
->stage
))
161 pause_query(ctx
, hq
, ctx
->ring
);
163 /* move to current list: */
165 list_addtail(&hq
->list
, &ctx
->current_queries
);
168 /* helper to get ptr to specified sample: */
169 static void * sampptr(struct fd_hw_sample
*samp
, uint32_t n
, void *ptr
)
171 return ((char *)ptr
) + (samp
->tile_stride
* n
) + samp
->offset
;
175 fd_hw_get_query_result(struct fd_context
*ctx
, struct fd_query
*q
,
176 boolean wait
, union pipe_query_result
*result
)
178 struct fd_hw_query
*hq
= fd_hw_query(q
);
179 const struct fd_hw_sample_provider
*p
= hq
->provider
;
180 struct fd_hw_sample_period
*period
;
185 /* if the app tries to read back the query result before the
186 * back is submitted, that forces us to flush so that there
187 * are actually results to wait for:
189 if (!LIST_IS_EMPTY(&hq
->list
)) {
190 DBG("reading query result forces flush!");
191 ctx
->needs_flush
= true;
192 fd_context_render(&ctx
->base
);
195 util_query_clear_result(result
, q
->type
);
197 if (LIST_IS_EMPTY(&hq
->periods
))
200 assert(LIST_IS_EMPTY(&hq
->list
));
201 assert(LIST_IS_EMPTY(&hq
->current_periods
));
204 if (LIST_IS_EMPTY(&hq
->periods
))
207 /* if !wait, then check the last sample (the one most likely to
208 * not be ready yet) and bail if it is not ready:
213 period
= LIST_ENTRY(struct fd_hw_sample_period
,
214 hq
->periods
.prev
, list
);
216 ret
= fd_bo_cpu_prep(period
->end
->bo
, ctx
->screen
->pipe
,
217 DRM_FREEDRENO_PREP_READ
| DRM_FREEDRENO_PREP_NOSYNC
);
221 fd_bo_cpu_fini(period
->end
->bo
);
224 /* sum the result across all sample periods: */
225 LIST_FOR_EACH_ENTRY(period
, &hq
->periods
, list
) {
226 struct fd_hw_sample
*start
= period
->start
;
227 struct fd_hw_sample
*end
= period
->end
;
230 /* start and end samples should be from same batch: */
231 assert(start
->bo
== end
->bo
);
232 assert(start
->num_tiles
== end
->num_tiles
);
234 for (i
= 0; i
< start
->num_tiles
; i
++) {
237 fd_bo_cpu_prep(start
->bo
, ctx
->screen
->pipe
,
238 DRM_FREEDRENO_PREP_READ
);
240 ptr
= fd_bo_map(start
->bo
);
242 p
->accumulate_result(ctx
, sampptr(period
->start
, i
, ptr
),
243 sampptr(period
->end
, i
, ptr
), result
);
245 fd_bo_cpu_fini(start
->bo
);
252 static const struct fd_query_funcs hw_query_funcs
= {
253 .destroy_query
= fd_hw_destroy_query
,
254 .begin_query
= fd_hw_begin_query
,
255 .end_query
= fd_hw_end_query
,
256 .get_query_result
= fd_hw_get_query_result
,
260 fd_hw_create_query(struct fd_context
*ctx
, unsigned query_type
)
262 struct fd_hw_query
*hq
;
264 int idx
= pidx(query_type
);
266 if ((idx
< 0) || !ctx
->sample_providers
[idx
])
269 hq
= CALLOC_STRUCT(fd_hw_query
);
273 hq
->provider
= ctx
->sample_providers
[idx
];
275 list_inithead(&hq
->periods
);
276 list_inithead(&hq
->current_periods
);
277 list_inithead(&hq
->list
);
280 q
->funcs
= &hw_query_funcs
;
281 q
->type
= query_type
;
286 struct fd_hw_sample
*
287 fd_hw_sample_init(struct fd_context
*ctx
, uint32_t size
)
289 struct fd_hw_sample
*samp
= util_slab_alloc(&ctx
->sample_pool
);
290 pipe_reference_init(&samp
->reference
, 1);
292 samp
->offset
= ctx
->next_sample_offset
;
293 /* NOTE: util_slab_alloc() does not zero out the buffer: */
296 samp
->tile_stride
= 0;
297 ctx
->next_sample_offset
+= size
;
302 __fd_hw_sample_destroy(struct fd_context
*ctx
, struct fd_hw_sample
*samp
)
306 util_slab_free(&ctx
->sample_pool
, samp
);
310 prepare_sample(struct fd_hw_sample
*samp
, struct fd_bo
*bo
,
311 uint32_t num_tiles
, uint32_t tile_stride
)
314 assert(samp
->bo
== bo
);
315 assert(samp
->num_tiles
== num_tiles
);
316 assert(samp
->tile_stride
== tile_stride
);
320 samp
->num_tiles
= num_tiles
;
321 samp
->tile_stride
= tile_stride
;
325 prepare_query(struct fd_hw_query
*hq
, struct fd_bo
*bo
,
326 uint32_t num_tiles
, uint32_t tile_stride
)
328 struct fd_hw_sample_period
*period
, *s
;
330 /* prepare all the samples in the query: */
331 LIST_FOR_EACH_ENTRY_SAFE(period
, s
, &hq
->current_periods
, list
) {
332 prepare_sample(period
->start
, bo
, num_tiles
, tile_stride
);
333 prepare_sample(period
->end
, bo
, num_tiles
, tile_stride
);
335 /* move from current_periods list to periods list: */
336 list_del(&period
->list
);
337 list_addtail(&period
->list
, &hq
->periods
);
342 prepare_queries(struct fd_context
*ctx
, struct fd_bo
*bo
,
343 uint32_t num_tiles
, uint32_t tile_stride
,
344 struct list_head
*list
, bool remove
)
346 struct fd_hw_query
*hq
, *s
;
347 LIST_FOR_EACH_ENTRY_SAFE(hq
, s
, list
, list
) {
348 prepare_query(hq
, bo
, num_tiles
, tile_stride
);
350 list_delinit(&hq
->list
);
354 /* called from gmem code once total storage requirements are known (ie.
355 * number of samples times number of tiles)
358 fd_hw_query_prepare(struct fd_context
*ctx
, uint32_t num_tiles
)
360 uint32_t tile_stride
= ctx
->next_sample_offset
;
364 fd_bo_del(ctx
->query_bo
);
366 if (tile_stride
> 0) {
367 bo
= fd_bo_new(ctx
->dev
, tile_stride
* num_tiles
,
368 DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
369 DRM_FREEDRENO_GEM_TYPE_KMEM
);
375 ctx
->query_tile_stride
= tile_stride
;
377 prepare_queries(ctx
, bo
, num_tiles
, tile_stride
,
378 &ctx
->active_queries
, false);
379 prepare_queries(ctx
, bo
, num_tiles
, tile_stride
,
380 &ctx
->current_queries
, true);
382 /* reset things for next batch: */
383 ctx
->next_sample_offset
= 0;
387 fd_hw_query_prepare_tile(struct fd_context
*ctx
, uint32_t n
,
388 struct fd_ringbuffer
*ring
)
390 uint32_t tile_stride
= ctx
->query_tile_stride
;
391 uint32_t offset
= tile_stride
* n
;
393 /* bail if no queries: */
394 if (tile_stride
== 0)
398 OUT_PKT0 (ring
, HW_QUERY_BASE_REG
, 1);
399 OUT_RELOCW(ring
, ctx
->query_bo
, offset
, 0, 0);
403 fd_hw_query_set_stage(struct fd_context
*ctx
, struct fd_ringbuffer
*ring
,
404 enum fd_render_stage stage
)
406 /* special case: internal blits (like mipmap level generation)
407 * go through normal draw path (via util_blitter_blit()).. but
408 * we need to ignore the FD_STAGE_DRAW which will be set, so we
409 * don't enable queries which should be paused during internal
412 if ((ctx
->stage
== FD_STAGE_BLIT
) &&
413 (stage
!= FD_STAGE_NULL
))
416 if (stage
!= ctx
->stage
) {
417 struct fd_hw_query
*hq
;
418 LIST_FOR_EACH_ENTRY(hq
, &ctx
->active_queries
, list
) {
419 bool was_active
= is_active(hq
, ctx
->stage
);
420 bool now_active
= is_active(hq
, stage
);
422 if (now_active
&& !was_active
)
423 resume_query(ctx
, hq
, ring
);
424 else if (was_active
&& !now_active
)
425 pause_query(ctx
, hq
, ring
);
428 clear_sample_cache(ctx
);
433 fd_hw_query_register_provider(struct pipe_context
*pctx
,
434 const struct fd_hw_sample_provider
*provider
)
436 struct fd_context
*ctx
= fd_context(pctx
);
437 int idx
= pidx(provider
->query_type
);
439 assert((0 <= idx
) && (idx
< MAX_HW_SAMPLE_PROVIDERS
));
440 assert(!ctx
->sample_providers
[idx
]);
442 ctx
->sample_providers
[idx
] = provider
;
446 fd_hw_query_init(struct pipe_context
*pctx
)
448 struct fd_context
*ctx
= fd_context(pctx
);
450 util_slab_create(&ctx
->sample_pool
, sizeof(struct fd_hw_sample
),
451 16, UTIL_SLAB_SINGLETHREADED
);
452 util_slab_create(&ctx
->sample_period_pool
, sizeof(struct fd_hw_sample_period
),
453 16, UTIL_SLAB_SINGLETHREADED
);
454 list_inithead(&ctx
->active_queries
);
455 list_inithead(&ctx
->current_queries
);
459 fd_hw_query_fini(struct pipe_context
*pctx
)
461 struct fd_context
*ctx
= fd_context(pctx
);
463 util_slab_destroy(&ctx
->sample_pool
);
464 util_slab_destroy(&ctx
->sample_period_pool
);