llvmpipe/query: add support for indexed queries
[mesa.git] / src / gallium / drivers / llvmpipe / lp_query.c
1 /**************************************************************************
2 *
3 * Copyright 2007 VMware, Inc.
4 * Copyright 2010 VMware, Inc.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL THE AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29 /* Authors:
30 * Keith Whitwell, Qicheng Christopher Li, Brian Paul
31 */
32
33 #include "draw/draw_context.h"
34 #include "pipe/p_defines.h"
35 #include "util/u_memory.h"
36 #include "util/os_time.h"
37 #include "lp_context.h"
38 #include "lp_flush.h"
39 #include "lp_fence.h"
40 #include "lp_query.h"
41 #include "lp_screen.h"
42 #include "lp_state.h"
43 #include "lp_rast.h"
44
45
46 static struct llvmpipe_query *llvmpipe_query( struct pipe_query *p )
47 {
48 return (struct llvmpipe_query *)p;
49 }
50
51 static struct pipe_query *
52 llvmpipe_create_query(struct pipe_context *pipe,
53 unsigned type,
54 unsigned index)
55 {
56 struct llvmpipe_query *pq;
57
58 assert(type < PIPE_QUERY_TYPES);
59
60 pq = CALLOC_STRUCT( llvmpipe_query );
61
62 if (pq) {
63 pq->type = type;
64 pq->index = index;
65 }
66
67 return (struct pipe_query *) pq;
68 }
69
70
71 static void
72 llvmpipe_destroy_query(struct pipe_context *pipe, struct pipe_query *q)
73 {
74 struct llvmpipe_query *pq = llvmpipe_query(q);
75
76 /* Ideally we would refcount queries & not get destroyed until the
77 * last scene had finished with us.
78 */
79 if (pq->fence) {
80 if (!lp_fence_issued(pq->fence))
81 llvmpipe_flush(pipe, NULL, __FUNCTION__);
82
83 if (!lp_fence_signalled(pq->fence))
84 lp_fence_wait(pq->fence);
85
86 lp_fence_reference(&pq->fence, NULL);
87 }
88
89 FREE(pq);
90 }
91
92
93 static bool
94 llvmpipe_get_query_result(struct pipe_context *pipe,
95 struct pipe_query *q,
96 bool wait,
97 union pipe_query_result *vresult)
98 {
99 struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
100 unsigned num_threads = MAX2(1, screen->num_threads);
101 struct llvmpipe_query *pq = llvmpipe_query(q);
102 uint64_t *result = (uint64_t *)vresult;
103 int i;
104
105 if (pq->fence) {
106 /* only have a fence if there was a scene */
107 if (!lp_fence_signalled(pq->fence)) {
108 if (!lp_fence_issued(pq->fence))
109 llvmpipe_flush(pipe, NULL, __FUNCTION__);
110
111 if (!wait)
112 return false;
113
114 lp_fence_wait(pq->fence);
115 }
116 }
117
118 /* Sum the results from each of the threads:
119 */
120 *result = 0;
121
122 switch (pq->type) {
123 case PIPE_QUERY_OCCLUSION_COUNTER:
124 for (i = 0; i < num_threads; i++) {
125 *result += pq->end[i];
126 }
127 break;
128 case PIPE_QUERY_OCCLUSION_PREDICATE:
129 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
130 for (i = 0; i < num_threads; i++) {
131 /* safer (still not guaranteed) when there's an overflow */
132 vresult->b = vresult->b || pq->end[i];
133 }
134 break;
135 case PIPE_QUERY_TIMESTAMP:
136 for (i = 0; i < num_threads; i++) {
137 if (pq->end[i] > *result) {
138 *result = pq->end[i];
139 }
140 }
141 break;
142 case PIPE_QUERY_TIMESTAMP_DISJOINT: {
143 struct pipe_query_data_timestamp_disjoint *td =
144 (struct pipe_query_data_timestamp_disjoint *)vresult;
145 /* os_get_time_nano return nanoseconds */
146 td->frequency = UINT64_C(1000000000);
147 td->disjoint = false;
148 }
149 break;
150 case PIPE_QUERY_GPU_FINISHED:
151 vresult->b = true;
152 break;
153 case PIPE_QUERY_PRIMITIVES_GENERATED:
154 *result = pq->num_primitives_generated;
155 break;
156 case PIPE_QUERY_PRIMITIVES_EMITTED:
157 *result = pq->num_primitives_written;
158 break;
159 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
160 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
161 vresult->b = pq->num_primitives_generated > pq->num_primitives_written;
162 break;
163 case PIPE_QUERY_SO_STATISTICS: {
164 struct pipe_query_data_so_statistics *stats =
165 (struct pipe_query_data_so_statistics *)vresult;
166 stats->num_primitives_written = pq->num_primitives_written;
167 stats->primitives_storage_needed = pq->num_primitives_generated;
168 }
169 break;
170 case PIPE_QUERY_PIPELINE_STATISTICS: {
171 struct pipe_query_data_pipeline_statistics *stats =
172 (struct pipe_query_data_pipeline_statistics *)vresult;
173 /* only ps_invocations come from binned query */
174 for (i = 0; i < num_threads; i++) {
175 pq->stats.ps_invocations += pq->end[i];
176 }
177 pq->stats.ps_invocations *= LP_RASTER_BLOCK_SIZE * LP_RASTER_BLOCK_SIZE;
178 *stats = pq->stats;
179 }
180 break;
181 default:
182 assert(0);
183 break;
184 }
185
186 return true;
187 }
188
189 static void
190 llvmpipe_get_query_result_resource(struct pipe_context *pipe,
191 struct pipe_query *q,
192 bool wait,
193 enum pipe_query_value_type result_type,
194 int index,
195 struct pipe_resource *resource,
196 unsigned offset)
197 {
198 struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
199 unsigned num_threads = MAX2(1, screen->num_threads);
200 struct llvmpipe_query *pq = llvmpipe_query(q);
201 struct llvmpipe_resource *lpr = llvmpipe_resource(resource);
202 bool unflushed = false;
203 bool unsignalled = false;
204 if (pq->fence) {
205 /* only have a fence if there was a scene */
206 if (!lp_fence_signalled(pq->fence)) {
207 unsignalled = true;
208 if (!lp_fence_issued(pq->fence))
209 unflushed = true;
210 }
211 }
212
213
214 uint64_t value = 0;
215 if (index == -1)
216 if (unsignalled)
217 value = 0;
218 else
219 value = 1;
220 else {
221 unsigned i;
222
223 if (unflushed) {
224 llvmpipe_flush(pipe, NULL, __FUNCTION__);
225
226 if (!wait)
227 return;
228
229 lp_fence_wait(pq->fence);
230 }
231
232 switch (pq->type) {
233 case PIPE_QUERY_OCCLUSION_COUNTER:
234 for (i = 0; i < num_threads; i++) {
235 value += pq->end[i];
236 }
237 break;
238 case PIPE_QUERY_OCCLUSION_PREDICATE:
239 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
240 for (i = 0; i < num_threads; i++) {
241 /* safer (still not guaranteed) when there's an overflow */
242 value = value || pq->end[i];
243 }
244 break;
245 case PIPE_QUERY_PRIMITIVES_GENERATED:
246 value = pq->num_primitives_generated;
247 break;
248 case PIPE_QUERY_PRIMITIVES_EMITTED:
249 value = pq->num_primitives_written;
250 break;
251 case PIPE_QUERY_TIMESTAMP:
252 for (i = 0; i < num_threads; i++) {
253 if (pq->end[i] > value) {
254 value = pq->end[i];
255 }
256 }
257 break;
258 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
259 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
260 value = !!(pq->num_primitives_generated > pq->num_primitives_written);
261 break;
262 case PIPE_QUERY_PIPELINE_STATISTICS:
263 switch ((enum pipe_statistics_query_index)index) {
264 case PIPE_STAT_QUERY_IA_VERTICES:
265 value = pq->stats.ia_vertices;
266 break;
267 case PIPE_STAT_QUERY_IA_PRIMITIVES:
268 value = pq->stats.ia_primitives;
269 break;
270 case PIPE_STAT_QUERY_VS_INVOCATIONS:
271 value = pq->stats.vs_invocations;
272 break;
273 case PIPE_STAT_QUERY_GS_INVOCATIONS:
274 value = pq->stats.gs_invocations;
275 break;
276 case PIPE_STAT_QUERY_GS_PRIMITIVES:
277 value = pq->stats.gs_primitives;
278 break;
279 case PIPE_STAT_QUERY_C_INVOCATIONS:
280 value = pq->stats.c_invocations;
281 break;
282 case PIPE_STAT_QUERY_C_PRIMITIVES:
283 value = pq->stats.c_primitives;
284 break;
285 case PIPE_STAT_QUERY_PS_INVOCATIONS:
286 value = 0;
287 for (i = 0; i < num_threads; i++) {
288 value += pq->end[i];
289 }
290 value *= LP_RASTER_BLOCK_SIZE * LP_RASTER_BLOCK_SIZE;
291 break;
292 case PIPE_STAT_QUERY_HS_INVOCATIONS:
293 value = pq->stats.hs_invocations;
294 break;
295 case PIPE_STAT_QUERY_DS_INVOCATIONS:
296 value = pq->stats.ds_invocations;
297 break;
298 case PIPE_STAT_QUERY_CS_INVOCATIONS:
299 value = pq->stats.cs_invocations;
300 break;
301 }
302 break;
303 default:
304 fprintf(stderr, "Unknown query type %d\n", pq->type);
305 break;
306 }
307 }
308
309 void *dst = (uint8_t *)lpr->data + offset;
310 switch (result_type) {
311 case PIPE_QUERY_TYPE_I32: {
312 int32_t *iptr = (int32_t *)dst;
313 if (value > 0x7fffffff)
314 *iptr = 0x7fffffff;
315 else
316 *iptr = (int32_t)value;
317 break;
318 }
319 case PIPE_QUERY_TYPE_U32: {
320 uint32_t *uptr = (uint32_t *)dst;
321 if (value > 0xffffffff)
322 *uptr = 0xffffffff;
323 else
324 *uptr = (uint32_t)value;
325 break;
326 }
327 case PIPE_QUERY_TYPE_I64: {
328 int64_t *iptr = (int64_t *)dst;
329 *iptr = (int64_t)value;
330 break;
331 }
332 case PIPE_QUERY_TYPE_U64: {
333 uint64_t *uptr = (uint64_t *)dst;
334 *uptr = (uint64_t)value;
335 break;
336 }
337 }
338 }
339
340 static bool
341 llvmpipe_begin_query(struct pipe_context *pipe, struct pipe_query *q)
342 {
343 struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe );
344 struct llvmpipe_query *pq = llvmpipe_query(q);
345
346 /* Check if the query is already in the scene. If so, we need to
347 * flush the scene now. Real apps shouldn't re-use a query in a
348 * frame of rendering.
349 */
350 if (pq->fence && !lp_fence_issued(pq->fence)) {
351 llvmpipe_finish(pipe, __FUNCTION__);
352 }
353
354
355 memset(pq->start, 0, sizeof(pq->start));
356 memset(pq->end, 0, sizeof(pq->end));
357 lp_setup_begin_query(llvmpipe->setup, pq);
358
359 switch (pq->type) {
360 case PIPE_QUERY_PRIMITIVES_EMITTED:
361 pq->num_primitives_written = llvmpipe->so_stats[pq->index].num_primitives_written;
362 break;
363 case PIPE_QUERY_PRIMITIVES_GENERATED:
364 pq->num_primitives_generated = llvmpipe->so_stats[pq->index].primitives_storage_needed;
365 llvmpipe->active_primgen_queries++;
366 break;
367 case PIPE_QUERY_SO_STATISTICS:
368 pq->num_primitives_written = llvmpipe->so_stats[pq->index].num_primitives_written;
369 pq->num_primitives_generated = llvmpipe->so_stats[pq->index].primitives_storage_needed;
370 break;
371 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
372 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
373 pq->num_primitives_written = llvmpipe->so_stats[pq->index].num_primitives_written;
374 pq->num_primitives_generated = llvmpipe->so_stats[pq->index].primitives_storage_needed;
375 break;
376 case PIPE_QUERY_PIPELINE_STATISTICS:
377 /* reset our cache */
378 if (llvmpipe->active_statistics_queries == 0) {
379 memset(&llvmpipe->pipeline_statistics, 0,
380 sizeof(llvmpipe->pipeline_statistics));
381 }
382 memcpy(&pq->stats, &llvmpipe->pipeline_statistics, sizeof(pq->stats));
383 llvmpipe->active_statistics_queries++;
384 break;
385 case PIPE_QUERY_OCCLUSION_COUNTER:
386 case PIPE_QUERY_OCCLUSION_PREDICATE:
387 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
388 llvmpipe->active_occlusion_queries++;
389 llvmpipe->dirty |= LP_NEW_OCCLUSION_QUERY;
390 break;
391 default:
392 break;
393 }
394 return true;
395 }
396
397
398 static bool
399 llvmpipe_end_query(struct pipe_context *pipe, struct pipe_query *q)
400 {
401 struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe );
402 struct llvmpipe_query *pq = llvmpipe_query(q);
403
404 lp_setup_end_query(llvmpipe->setup, pq);
405
406 switch (pq->type) {
407
408 case PIPE_QUERY_PRIMITIVES_EMITTED:
409 pq->num_primitives_written =
410 llvmpipe->so_stats[pq->index].num_primitives_written - pq->num_primitives_written;
411 break;
412 case PIPE_QUERY_PRIMITIVES_GENERATED:
413 assert(llvmpipe->active_primgen_queries);
414 llvmpipe->active_primgen_queries--;
415 pq->num_primitives_generated =
416 llvmpipe->so_stats[pq->index].primitives_storage_needed - pq->num_primitives_generated;
417 break;
418 case PIPE_QUERY_SO_STATISTICS:
419 pq->num_primitives_written =
420 llvmpipe->so_stats[pq->index].num_primitives_written - pq->num_primitives_written;
421 pq->num_primitives_generated =
422 llvmpipe->so_stats[pq->index].primitives_storage_needed - pq->num_primitives_generated;
423 break;
424 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
425 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
426 pq->num_primitives_written =
427 llvmpipe->so_stats[pq->index].num_primitives_written - pq->num_primitives_written;
428 pq->num_primitives_generated =
429 llvmpipe->so_stats[pq->index].primitives_storage_needed - pq->num_primitives_generated;
430 break;
431 case PIPE_QUERY_PIPELINE_STATISTICS:
432 pq->stats.ia_vertices =
433 llvmpipe->pipeline_statistics.ia_vertices - pq->stats.ia_vertices;
434 pq->stats.ia_primitives =
435 llvmpipe->pipeline_statistics.ia_primitives - pq->stats.ia_primitives;
436 pq->stats.vs_invocations =
437 llvmpipe->pipeline_statistics.vs_invocations - pq->stats.vs_invocations;
438 pq->stats.gs_invocations =
439 llvmpipe->pipeline_statistics.gs_invocations - pq->stats.gs_invocations;
440 pq->stats.gs_primitives =
441 llvmpipe->pipeline_statistics.gs_primitives - pq->stats.gs_primitives;
442 pq->stats.c_invocations =
443 llvmpipe->pipeline_statistics.c_invocations - pq->stats.c_invocations;
444 pq->stats.c_primitives =
445 llvmpipe->pipeline_statistics.c_primitives - pq->stats.c_primitives;
446 pq->stats.ps_invocations =
447 llvmpipe->pipeline_statistics.ps_invocations - pq->stats.ps_invocations;
448 pq->stats.cs_invocations =
449 llvmpipe->pipeline_statistics.cs_invocations - pq->stats.cs_invocations;
450 llvmpipe->active_statistics_queries--;
451 break;
452 case PIPE_QUERY_OCCLUSION_COUNTER:
453 case PIPE_QUERY_OCCLUSION_PREDICATE:
454 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
455 assert(llvmpipe->active_occlusion_queries);
456 llvmpipe->active_occlusion_queries--;
457 llvmpipe->dirty |= LP_NEW_OCCLUSION_QUERY;
458 break;
459 default:
460 break;
461 }
462
463 return true;
464 }
465
466 boolean
467 llvmpipe_check_render_cond(struct llvmpipe_context *lp)
468 {
469 struct pipe_context *pipe = &lp->pipe;
470 boolean b, wait;
471 uint64_t result;
472
473 if (!lp->render_cond_query)
474 return TRUE; /* no query predicate, draw normally */
475
476 wait = (lp->render_cond_mode == PIPE_RENDER_COND_WAIT ||
477 lp->render_cond_mode == PIPE_RENDER_COND_BY_REGION_WAIT);
478
479 b = pipe->get_query_result(pipe, lp->render_cond_query, wait, (void*)&result);
480 if (b)
481 return ((!result) == lp->render_cond_cond);
482 else
483 return TRUE;
484 }
485
486 static void
487 llvmpipe_set_active_query_state(struct pipe_context *pipe, bool enable)
488 {
489 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
490
491 llvmpipe->queries_disabled = !enable;
492 /* for OQs we need to regenerate the fragment shader */
493 llvmpipe->dirty |= LP_NEW_OCCLUSION_QUERY;
494 }
495
496 void llvmpipe_init_query_funcs(struct llvmpipe_context *llvmpipe )
497 {
498 llvmpipe->pipe.create_query = llvmpipe_create_query;
499 llvmpipe->pipe.destroy_query = llvmpipe_destroy_query;
500 llvmpipe->pipe.begin_query = llvmpipe_begin_query;
501 llvmpipe->pipe.end_query = llvmpipe_end_query;
502 llvmpipe->pipe.get_query_result = llvmpipe_get_query_result;
503 llvmpipe->pipe.get_query_result_resource = llvmpipe_get_query_result_resource;
504 llvmpipe->pipe.set_active_query_state = llvmpipe_set_active_query_state;
505 }
506
507