radeonsi: fix occlusion queries on Hawaii
[mesa.git] / src / gallium / drivers / freedreno / freedreno_query_hw.c
1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
2
3 /*
4 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
5 *
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:
12 *
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
15 * Software.
16 *
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
23 * SOFTWARE.
24 *
25 * Authors:
26 * Rob Clark <robclark@freedesktop.org>
27 */
28
29 #include "pipe/p_state.h"
30 #include "util/u_memory.h"
31 #include "util/u_inlines.h"
32
33 #include "freedreno_query_hw.h"
34 #include "freedreno_context.h"
35 #include "freedreno_util.h"
36
37 struct fd_hw_sample_period {
38 struct fd_hw_sample *start, *end;
39 struct list_head list;
40 };
41
42 /* maps query_type to sample provider idx: */
43 static int pidx(unsigned query_type)
44 {
45 switch (query_type) {
46 case PIPE_QUERY_OCCLUSION_COUNTER:
47 return 0;
48 case PIPE_QUERY_OCCLUSION_PREDICATE:
49 return 1;
50 default:
51 return -1;
52 }
53 }
54
55 static struct fd_hw_sample *
56 get_sample(struct fd_context *ctx, struct fd_ringbuffer *ring,
57 unsigned query_type)
58 {
59 struct fd_hw_sample *samp = NULL;
60 int idx = pidx(query_type);
61
62 if (!ctx->sample_cache[idx]) {
63 ctx->sample_cache[idx] =
64 ctx->sample_providers[idx]->get_sample(ctx, ring);
65 }
66
67 fd_hw_sample_reference(ctx, &samp, ctx->sample_cache[idx]);
68
69 return samp;
70 }
71
72 static void
73 clear_sample_cache(struct fd_context *ctx)
74 {
75 int i;
76
77 for (i = 0; i < ARRAY_SIZE(ctx->sample_cache); i++)
78 fd_hw_sample_reference(ctx, &ctx->sample_cache[i], NULL);
79 }
80
81 static bool
82 is_active(struct fd_hw_query *hq, enum fd_render_stage stage)
83 {
84 return !!(hq->provider->active & stage);
85 }
86
87
88 static void
89 resume_query(struct fd_context *ctx, struct fd_hw_query *hq,
90 struct fd_ringbuffer *ring)
91 {
92 assert(!hq->period);
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;
98 }
99
100 static void
101 pause_query(struct fd_context *ctx, struct fd_hw_query *hq,
102 struct fd_ringbuffer *ring)
103 {
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);
107 hq->period = NULL;
108 }
109
110 static void
111 destroy_periods(struct fd_context *ctx, struct list_head *list)
112 {
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);
119 }
120 }
121
122 static void
123 fd_hw_destroy_query(struct fd_context *ctx, struct fd_query *q)
124 {
125 struct fd_hw_query *hq = fd_hw_query(q);
126
127 destroy_periods(ctx, &hq->periods);
128 destroy_periods(ctx, &hq->current_periods);
129 list_del(&hq->list);
130
131 free(hq);
132 }
133
134 static void
135 fd_hw_begin_query(struct fd_context *ctx, struct fd_query *q)
136 {
137 struct fd_hw_query *hq = fd_hw_query(q);
138 if (q->active)
139 return;
140
141 /* begin_query() should clear previous results: */
142 destroy_periods(ctx, &hq->periods);
143
144 if (is_active(hq, ctx->stage))
145 resume_query(ctx, hq, ctx->ring);
146
147 q->active = true;
148
149 /* add to active list: */
150 list_del(&hq->list);
151 list_addtail(&hq->list, &ctx->active_queries);
152 }
153
154 static void
155 fd_hw_end_query(struct fd_context *ctx, struct fd_query *q)
156 {
157 struct fd_hw_query *hq = fd_hw_query(q);
158 if (!q->active)
159 return;
160 if (is_active(hq, ctx->stage))
161 pause_query(ctx, hq, ctx->ring);
162 q->active = false;
163 /* move to current list: */
164 list_del(&hq->list);
165 list_addtail(&hq->list, &ctx->current_queries);
166 }
167
168 /* helper to get ptr to specified sample: */
169 static void * sampptr(struct fd_hw_sample *samp, uint32_t n, void *ptr)
170 {
171 return ((char *)ptr) + (samp->tile_stride * n) + samp->offset;
172 }
173
174 static boolean
175 fd_hw_get_query_result(struct fd_context *ctx, struct fd_query *q,
176 boolean wait, union pipe_query_result *result)
177 {
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;
181
182 if (q->active)
183 return false;
184
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:
188 */
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);
193 }
194
195 util_query_clear_result(result, q->type);
196
197 if (LIST_IS_EMPTY(&hq->periods))
198 return true;
199
200 assert(LIST_IS_EMPTY(&hq->list));
201 assert(LIST_IS_EMPTY(&hq->current_periods));
202 assert(!hq->period);
203
204 if (LIST_IS_EMPTY(&hq->periods))
205 return true;
206
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:
209 */
210 if (!wait) {
211 int ret;
212
213 period = LIST_ENTRY(struct fd_hw_sample_period,
214 hq->periods.prev, list);
215
216 ret = fd_bo_cpu_prep(period->end->bo, ctx->screen->pipe,
217 DRM_FREEDRENO_PREP_READ | DRM_FREEDRENO_PREP_NOSYNC);
218 if (ret)
219 return false;
220
221 fd_bo_cpu_fini(period->end->bo);
222 }
223
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;
228 unsigned i;
229
230 /* start and end samples should be from same batch: */
231 assert(start->bo == end->bo);
232 assert(start->num_tiles == end->num_tiles);
233
234 for (i = 0; i < start->num_tiles; i++) {
235 void *ptr;
236
237 fd_bo_cpu_prep(start->bo, ctx->screen->pipe,
238 DRM_FREEDRENO_PREP_READ);
239
240 ptr = fd_bo_map(start->bo);
241
242 p->accumulate_result(ctx, sampptr(period->start, i, ptr),
243 sampptr(period->end, i, ptr), result);
244
245 fd_bo_cpu_fini(start->bo);
246 }
247 }
248
249 return true;
250 }
251
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,
257 };
258
259 struct fd_query *
260 fd_hw_create_query(struct fd_context *ctx, unsigned query_type)
261 {
262 struct fd_hw_query *hq;
263 struct fd_query *q;
264 int idx = pidx(query_type);
265
266 if ((idx < 0) || !ctx->sample_providers[idx])
267 return NULL;
268
269 hq = CALLOC_STRUCT(fd_hw_query);
270 if (!hq)
271 return NULL;
272
273 hq->provider = ctx->sample_providers[idx];
274
275 list_inithead(&hq->periods);
276 list_inithead(&hq->current_periods);
277 list_inithead(&hq->list);
278
279 q = &hq->base;
280 q->funcs = &hw_query_funcs;
281 q->type = query_type;
282
283 return q;
284 }
285
286 struct fd_hw_sample *
287 fd_hw_sample_init(struct fd_context *ctx, uint32_t size)
288 {
289 struct fd_hw_sample *samp = util_slab_alloc(&ctx->sample_pool);
290 pipe_reference_init(&samp->reference, 1);
291 samp->size = size;
292 samp->offset = ctx->next_sample_offset;
293 /* NOTE: util_slab_alloc() does not zero out the buffer: */
294 samp->bo = NULL;
295 samp->num_tiles = 0;
296 samp->tile_stride = 0;
297 ctx->next_sample_offset += size;
298 return samp;
299 }
300
301 void
302 __fd_hw_sample_destroy(struct fd_context *ctx, struct fd_hw_sample *samp)
303 {
304 if (samp->bo)
305 fd_bo_del(samp->bo);
306 util_slab_free(&ctx->sample_pool, samp);
307 }
308
309 static void
310 prepare_sample(struct fd_hw_sample *samp, struct fd_bo *bo,
311 uint32_t num_tiles, uint32_t tile_stride)
312 {
313 if (samp->bo) {
314 assert(samp->bo == bo);
315 assert(samp->num_tiles == num_tiles);
316 assert(samp->tile_stride == tile_stride);
317 return;
318 }
319 samp->bo = bo;
320 samp->num_tiles = num_tiles;
321 samp->tile_stride = tile_stride;
322 }
323
324 static void
325 prepare_query(struct fd_hw_query *hq, struct fd_bo *bo,
326 uint32_t num_tiles, uint32_t tile_stride)
327 {
328 struct fd_hw_sample_period *period, *s;
329
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);
334
335 /* move from current_periods list to periods list: */
336 list_del(&period->list);
337 list_addtail(&period->list, &hq->periods);
338 }
339 }
340
341 static void
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)
345 {
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);
349 if (remove)
350 list_delinit(&hq->list);
351 }
352 }
353
354 /* called from gmem code once total storage requirements are known (ie.
355 * number of samples times number of tiles)
356 */
357 void
358 fd_hw_query_prepare(struct fd_context *ctx, uint32_t num_tiles)
359 {
360 uint32_t tile_stride = ctx->next_sample_offset;
361 struct fd_bo *bo;
362
363 if (ctx->query_bo)
364 fd_bo_del(ctx->query_bo);
365
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);
370 } else {
371 bo = NULL;
372 }
373
374 ctx->query_bo = bo;
375 ctx->query_tile_stride = tile_stride;
376
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);
381
382 /* reset things for next batch: */
383 ctx->next_sample_offset = 0;
384 }
385
386 void
387 fd_hw_query_prepare_tile(struct fd_context *ctx, uint32_t n,
388 struct fd_ringbuffer *ring)
389 {
390 uint32_t tile_stride = ctx->query_tile_stride;
391 uint32_t offset = tile_stride * n;
392
393 /* bail if no queries: */
394 if (tile_stride == 0)
395 return;
396
397 fd_wfi(ctx, ring);
398 OUT_PKT0 (ring, HW_QUERY_BASE_REG, 1);
399 OUT_RELOCW(ring, ctx->query_bo, offset, 0, 0);
400 }
401
402 void
403 fd_hw_query_set_stage(struct fd_context *ctx, struct fd_ringbuffer *ring,
404 enum fd_render_stage stage)
405 {
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
410 * blits:
411 */
412 if ((ctx->stage == FD_STAGE_BLIT) &&
413 (stage != FD_STAGE_NULL))
414 return;
415
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);
421
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);
426 }
427 }
428 clear_sample_cache(ctx);
429 ctx->stage = stage;
430 }
431
432 void
433 fd_hw_query_register_provider(struct pipe_context *pctx,
434 const struct fd_hw_sample_provider *provider)
435 {
436 struct fd_context *ctx = fd_context(pctx);
437 int idx = pidx(provider->query_type);
438
439 assert((0 <= idx) && (idx < MAX_HW_SAMPLE_PROVIDERS));
440 assert(!ctx->sample_providers[idx]);
441
442 ctx->sample_providers[idx] = provider;
443 }
444
445 void
446 fd_hw_query_init(struct pipe_context *pctx)
447 {
448 struct fd_context *ctx = fd_context(pctx);
449
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);
456 }
457
458 void
459 fd_hw_query_fini(struct pipe_context *pctx)
460 {
461 struct fd_context *ctx = fd_context(pctx);
462
463 util_slab_destroy(&ctx->sample_pool);
464 util_slab_destroy(&ctx->sample_period_pool);
465 }