b29f9d409ced07411dcd595b8fe82ca36943f3dd
[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 * batch 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 /* if app didn't actually trigger any cmdstream, then
191 * we have nothing to do:
192 */
193 if (!ctx->needs_flush)
194 return true;
195 DBG("reading query result forces flush!");
196 fd_context_render(&ctx->base);
197 }
198
199 util_query_clear_result(result, q->type);
200
201 if (LIST_IS_EMPTY(&hq->periods))
202 return true;
203
204 assert(LIST_IS_EMPTY(&hq->list));
205 assert(LIST_IS_EMPTY(&hq->current_periods));
206 assert(!hq->period);
207
208 /* if !wait, then check the last sample (the one most likely to
209 * not be ready yet) and bail if it is not ready:
210 */
211 if (!wait) {
212 int ret;
213
214 period = LIST_ENTRY(struct fd_hw_sample_period,
215 hq->periods.prev, list);
216
217 ret = fd_bo_cpu_prep(period->end->bo, ctx->screen->pipe,
218 DRM_FREEDRENO_PREP_READ | DRM_FREEDRENO_PREP_NOSYNC);
219 if (ret)
220 return false;
221
222 fd_bo_cpu_fini(period->end->bo);
223 }
224
225 /* sum the result across all sample periods: */
226 LIST_FOR_EACH_ENTRY(period, &hq->periods, list) {
227 struct fd_hw_sample *start = period->start;
228 struct fd_hw_sample *end = period->end;
229 unsigned i;
230
231 /* start and end samples should be from same batch: */
232 assert(start->bo == end->bo);
233 assert(start->num_tiles == end->num_tiles);
234
235 for (i = 0; i < start->num_tiles; i++) {
236 void *ptr;
237
238 fd_bo_cpu_prep(start->bo, ctx->screen->pipe,
239 DRM_FREEDRENO_PREP_READ);
240
241 ptr = fd_bo_map(start->bo);
242
243 p->accumulate_result(ctx, sampptr(period->start, i, ptr),
244 sampptr(period->end, i, ptr), result);
245
246 fd_bo_cpu_fini(start->bo);
247 }
248 }
249
250 return true;
251 }
252
253 static const struct fd_query_funcs hw_query_funcs = {
254 .destroy_query = fd_hw_destroy_query,
255 .begin_query = fd_hw_begin_query,
256 .end_query = fd_hw_end_query,
257 .get_query_result = fd_hw_get_query_result,
258 };
259
260 struct fd_query *
261 fd_hw_create_query(struct fd_context *ctx, unsigned query_type)
262 {
263 struct fd_hw_query *hq;
264 struct fd_query *q;
265 int idx = pidx(query_type);
266
267 if ((idx < 0) || !ctx->sample_providers[idx])
268 return NULL;
269
270 hq = CALLOC_STRUCT(fd_hw_query);
271 if (!hq)
272 return NULL;
273
274 hq->provider = ctx->sample_providers[idx];
275
276 list_inithead(&hq->periods);
277 list_inithead(&hq->current_periods);
278 list_inithead(&hq->list);
279
280 q = &hq->base;
281 q->funcs = &hw_query_funcs;
282 q->type = query_type;
283
284 return q;
285 }
286
287 struct fd_hw_sample *
288 fd_hw_sample_init(struct fd_context *ctx, uint32_t size)
289 {
290 struct fd_hw_sample *samp = util_slab_alloc(&ctx->sample_pool);
291 pipe_reference_init(&samp->reference, 1);
292 samp->size = size;
293 samp->offset = ctx->next_sample_offset;
294 /* NOTE: util_slab_alloc() does not zero out the buffer: */
295 samp->bo = NULL;
296 samp->num_tiles = 0;
297 samp->tile_stride = 0;
298 ctx->next_sample_offset += size;
299 return samp;
300 }
301
302 void
303 __fd_hw_sample_destroy(struct fd_context *ctx, struct fd_hw_sample *samp)
304 {
305 if (samp->bo)
306 fd_bo_del(samp->bo);
307 util_slab_free(&ctx->sample_pool, samp);
308 }
309
310 static void
311 prepare_sample(struct fd_hw_sample *samp, struct fd_bo *bo,
312 uint32_t num_tiles, uint32_t tile_stride)
313 {
314 if (samp->bo) {
315 assert(samp->bo == bo);
316 assert(samp->num_tiles == num_tiles);
317 assert(samp->tile_stride == tile_stride);
318 return;
319 }
320 samp->bo = bo;
321 samp->num_tiles = num_tiles;
322 samp->tile_stride = tile_stride;
323 }
324
325 static void
326 prepare_query(struct fd_hw_query *hq, struct fd_bo *bo,
327 uint32_t num_tiles, uint32_t tile_stride)
328 {
329 struct fd_hw_sample_period *period, *s;
330
331 /* prepare all the samples in the query: */
332 LIST_FOR_EACH_ENTRY_SAFE(period, s, &hq->current_periods, list) {
333 prepare_sample(period->start, bo, num_tiles, tile_stride);
334 prepare_sample(period->end, bo, num_tiles, tile_stride);
335
336 /* move from current_periods list to periods list: */
337 list_del(&period->list);
338 list_addtail(&period->list, &hq->periods);
339 }
340 }
341
342 static void
343 prepare_queries(struct fd_context *ctx, struct fd_bo *bo,
344 uint32_t num_tiles, uint32_t tile_stride,
345 struct list_head *list, bool remove)
346 {
347 struct fd_hw_query *hq, *s;
348 LIST_FOR_EACH_ENTRY_SAFE(hq, s, list, list) {
349 prepare_query(hq, bo, num_tiles, tile_stride);
350 if (remove)
351 list_delinit(&hq->list);
352 }
353 }
354
355 /* called from gmem code once total storage requirements are known (ie.
356 * number of samples times number of tiles)
357 */
358 void
359 fd_hw_query_prepare(struct fd_context *ctx, uint32_t num_tiles)
360 {
361 uint32_t tile_stride = ctx->next_sample_offset;
362 struct fd_bo *bo;
363
364 if (ctx->query_bo)
365 fd_bo_del(ctx->query_bo);
366
367 if (tile_stride > 0) {
368 bo = fd_bo_new(ctx->dev, tile_stride * num_tiles,
369 DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
370 DRM_FREEDRENO_GEM_TYPE_KMEM);
371 } else {
372 bo = NULL;
373 }
374
375 ctx->query_bo = bo;
376 ctx->query_tile_stride = tile_stride;
377
378 prepare_queries(ctx, bo, num_tiles, tile_stride,
379 &ctx->active_queries, false);
380 prepare_queries(ctx, bo, num_tiles, tile_stride,
381 &ctx->current_queries, true);
382
383 /* reset things for next batch: */
384 ctx->next_sample_offset = 0;
385 }
386
387 void
388 fd_hw_query_prepare_tile(struct fd_context *ctx, uint32_t n,
389 struct fd_ringbuffer *ring)
390 {
391 uint32_t tile_stride = ctx->query_tile_stride;
392 uint32_t offset = tile_stride * n;
393
394 /* bail if no queries: */
395 if (tile_stride == 0)
396 return;
397
398 fd_wfi(ctx, ring);
399 OUT_PKT0 (ring, HW_QUERY_BASE_REG, 1);
400 OUT_RELOCW(ring, ctx->query_bo, offset, 0, 0);
401 }
402
403 void
404 fd_hw_query_set_stage(struct fd_context *ctx, struct fd_ringbuffer *ring,
405 enum fd_render_stage stage)
406 {
407 /* special case: internal blits (like mipmap level generation)
408 * go through normal draw path (via util_blitter_blit()).. but
409 * we need to ignore the FD_STAGE_DRAW which will be set, so we
410 * don't enable queries which should be paused during internal
411 * blits:
412 */
413 if ((ctx->stage == FD_STAGE_BLIT) &&
414 (stage != FD_STAGE_NULL))
415 return;
416
417 if (stage != ctx->stage) {
418 struct fd_hw_query *hq;
419 LIST_FOR_EACH_ENTRY(hq, &ctx->active_queries, list) {
420 bool was_active = is_active(hq, ctx->stage);
421 bool now_active = is_active(hq, stage);
422
423 if (now_active && !was_active)
424 resume_query(ctx, hq, ring);
425 else if (was_active && !now_active)
426 pause_query(ctx, hq, ring);
427 }
428 }
429 clear_sample_cache(ctx);
430 ctx->stage = stage;
431 }
432
433 void
434 fd_hw_query_register_provider(struct pipe_context *pctx,
435 const struct fd_hw_sample_provider *provider)
436 {
437 struct fd_context *ctx = fd_context(pctx);
438 int idx = pidx(provider->query_type);
439
440 assert((0 <= idx) && (idx < MAX_HW_SAMPLE_PROVIDERS));
441 assert(!ctx->sample_providers[idx]);
442
443 ctx->sample_providers[idx] = provider;
444 }
445
446 void
447 fd_hw_query_init(struct pipe_context *pctx)
448 {
449 struct fd_context *ctx = fd_context(pctx);
450
451 util_slab_create(&ctx->sample_pool, sizeof(struct fd_hw_sample),
452 16, UTIL_SLAB_SINGLETHREADED);
453 util_slab_create(&ctx->sample_period_pool, sizeof(struct fd_hw_sample_period),
454 16, UTIL_SLAB_SINGLETHREADED);
455 list_inithead(&ctx->active_queries);
456 list_inithead(&ctx->current_queries);
457 }
458
459 void
460 fd_hw_query_fini(struct pipe_context *pctx)
461 {
462 struct fd_context *ctx = fd_context(pctx);
463
464 util_slab_destroy(&ctx->sample_pool);
465 util_slab_destroy(&ctx->sample_period_pool);
466 }