1 /*
2  * Copyright © 2008 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *
26  */
27 
28 /** @file brw_queryobj.c
29  *
30  * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query,
31  * GL_EXT_transform_feedback, and friends).
32  *
33  * The hardware provides a PIPE_CONTROL command that can report the number of
34  * fragments that passed the depth test, or the hardware timer.  They are
35  * appropriately synced with the stage of the pipeline for our extensions'
36  * needs.
37  *
38  * To avoid getting samples from another context's rendering in our results,
39  * we capture the counts at the start and end of every batchbuffer while the
40  * query is active, and sum up the differences.  (We should do so for
41  * GL_TIME_ELAPSED as well, but don't).
42  */
43 #include "main/imports.h"
44 
45 #include "brw_context.h"
46 #include "brw_defines.h"
47 #include "brw_state.h"
48 #include "intel_batchbuffer.h"
49 #include "intel_reg.h"
50 
51 static void
write_timestamp(struct intel_context * intel,drm_intel_bo * query_bo,int idx)52 write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
53 {
54    if (intel->gen >= 6) {
55       /* Emit workaround flushes: */
56       if (intel->gen == 6) {
57          /* The timestamp write below is a non-zero post-sync op, which on
58           * Gen6 necessitates a CS stall.  CS stalls need stall at scoreboard
59           * set.  See the comments for intel_emit_post_sync_nonzero_flush().
60           */
61          BEGIN_BATCH(4);
62          OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
63          OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
64          OUT_BATCH(0);
65          OUT_BATCH(0);
66          ADVANCE_BATCH();
67       }
68 
69       BEGIN_BATCH(5);
70       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
71       OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
72       OUT_RELOC(query_bo,
73                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
74                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
75                 idx * sizeof(uint64_t));
76       OUT_BATCH(0);
77       OUT_BATCH(0);
78       ADVANCE_BATCH();
79    } else {
80       BEGIN_BATCH(4);
81       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
82                 PIPE_CONTROL_WRITE_TIMESTAMP);
83       OUT_RELOC(query_bo,
84                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
85                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
86                 idx * sizeof(uint64_t));
87       OUT_BATCH(0);
88       OUT_BATCH(0);
89       ADVANCE_BATCH();
90    }
91 }
92 
93 static void
write_depth_count(struct intel_context * intel,drm_intel_bo * query_bo,int idx)94 write_depth_count(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
95 {
96    if (intel->gen >= 6) {
97       /* Emit Sandybridge workaround flush: */
98       if (intel->gen == 6)
99          intel_emit_post_sync_nonzero_flush(intel);
100 
101       BEGIN_BATCH(5);
102       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
103       OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
104                 PIPE_CONTROL_WRITE_DEPTH_COUNT);
105       OUT_RELOC(query_bo,
106                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
107                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
108                 (idx * sizeof(uint64_t)));
109       OUT_BATCH(0);
110       OUT_BATCH(0);
111       ADVANCE_BATCH();
112    } else {
113       BEGIN_BATCH(4);
114       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
115                 PIPE_CONTROL_DEPTH_STALL |
116                 PIPE_CONTROL_WRITE_DEPTH_COUNT);
117       /* This object could be mapped cacheable, but we don't have an exposed
118        * mechanism to support that.  Since it's going uncached, tell GEM that
119        * we're writing to it.  The usual clflush should be all that's required
120        * to pick up the results.
121        */
122       OUT_RELOC(query_bo,
123                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
124                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
125                 (idx * sizeof(uint64_t)));
126       OUT_BATCH(0);
127       OUT_BATCH(0);
128       ADVANCE_BATCH();
129    }
130 }
131 
132 /** Waits on the query object's BO and totals the results for this query */
133 static void
brw_queryobj_get_results(struct gl_context * ctx,struct brw_query_object * query)134 brw_queryobj_get_results(struct gl_context *ctx,
135 			 struct brw_query_object *query)
136 {
137    struct intel_context *intel = intel_context(ctx);
138 
139    int i;
140    uint64_t *results;
141 
142    if (query->bo == NULL)
143       return;
144 
145    if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
146       if (drm_intel_bo_busy(query->bo)) {
147          perf_debug("Stalling on the GPU waiting for a query object.\n");
148       }
149    }
150 
151    drm_intel_bo_map(query->bo, false);
152    results = query->bo->virtual;
153    switch (query->Base.Target) {
154    case GL_TIME_ELAPSED_EXT:
155       if (intel->gen >= 6)
156 	 query->Base.Result += 80 * (results[1] - results[0]);
157       else
158 	 query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
159       break;
160 
161    case GL_TIMESTAMP:
162       if (intel->gen >= 6) {
163          /* Our timer is a clock that increments every 80ns (regardless of
164           * other clock scaling in the system).  The timestamp register we can
165           * read for glGetTimestamp() masks out the top 32 bits, so we do that
166           * here too to let the two counters be compared against each other.
167           *
168           * If we just multiplied that 32 bits of data by 80, it would roll
169           * over at a non-power-of-two, so an application couldn't use
170           * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
171           * report 36 bits and truncate at that (rolling over 5 times as often
172           * as the HW counter), and when the 32-bit counter rolls over, it
173           * happens to also be at a rollover in the reported value from near
174           * (1<<36) to 0.
175           *
176           * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
177           * rolls over every ~69 seconds.
178           */
179 	 query->Base.Result = 80 * (results[1] & 0xffffffff);
180          query->Base.Result &= (1ull << 36) - 1;
181       } else {
182 	 query->Base.Result = 1000 * (results[1] >> 32);
183       }
184 
185       break;
186 
187    case GL_SAMPLES_PASSED_ARB:
188       /* Map and count the pixels from the current query BO */
189       for (i = query->first_index; i <= query->last_index; i++) {
190 	 query->Base.Result += results[i * 2 + 1] - results[i * 2];
191       }
192       break;
193 
194    case GL_ANY_SAMPLES_PASSED:
195       /* Set true if any of the sub-queries passed. */
196       for (i = query->first_index; i <= query->last_index; i++) {
197 	 if (results[i * 2 + 1] != results[i * 2]) {
198             query->Base.Result = GL_TRUE;
199             break;
200          }
201       }
202       break;
203 
204    case GL_PRIMITIVES_GENERATED:
205    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
206       /* We don't actually query the hardware for this value, so query->bo
207        * should always be NULL and execution should never reach here.
208        */
209       assert(!"Unreachable");
210       break;
211 
212    default:
213       assert(!"Unrecognized query target in brw_queryobj_get_results()");
214       break;
215    }
216    drm_intel_bo_unmap(query->bo);
217 
218    drm_intel_bo_unreference(query->bo);
219    query->bo = NULL;
220 }
221 
222 static struct gl_query_object *
brw_new_query_object(struct gl_context * ctx,GLuint id)223 brw_new_query_object(struct gl_context *ctx, GLuint id)
224 {
225    struct brw_query_object *query;
226 
227    query = calloc(1, sizeof(struct brw_query_object));
228 
229    query->Base.Id = id;
230    query->Base.Result = 0;
231    query->Base.Active = false;
232    query->Base.Ready = true;
233 
234    return &query->Base;
235 }
236 
237 static void
brw_delete_query(struct gl_context * ctx,struct gl_query_object * q)238 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
239 {
240    struct brw_query_object *query = (struct brw_query_object *)q;
241 
242    drm_intel_bo_unreference(query->bo);
243    free(query);
244 }
245 
246 static void
brw_begin_query(struct gl_context * ctx,struct gl_query_object * q)247 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
248 {
249    struct brw_context *brw = brw_context(ctx);
250    struct intel_context *intel = intel_context(ctx);
251    struct brw_query_object *query = (struct brw_query_object *)q;
252 
253    switch (query->Base.Target) {
254    case GL_TIME_ELAPSED_EXT:
255       drm_intel_bo_unreference(query->bo);
256       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
257       write_timestamp(intel, query->bo, 0);
258       break;
259 
260    case GL_ANY_SAMPLES_PASSED:
261    case GL_SAMPLES_PASSED_ARB:
262       /* Reset our driver's tracking of query state. */
263       drm_intel_bo_unreference(query->bo);
264       query->bo = NULL;
265       query->first_index = -1;
266       query->last_index = -1;
267 
268       brw->query.obj = query;
269       intel->stats_wm++;
270       break;
271 
272    case GL_PRIMITIVES_GENERATED:
273       /* We don't actually query the hardware for this value; we keep track of
274        * it a software counter.  So just reset the counter.
275        */
276       brw->sol.primitives_generated = 0;
277       brw->sol.counting_primitives_generated = true;
278       break;
279 
280    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
281       /* We don't actually query the hardware for this value; we keep track of
282        * it a software counter.  So just reset the counter.
283        */
284       brw->sol.primitives_written = 0;
285       brw->sol.counting_primitives_written = true;
286       break;
287 
288    default:
289       assert(!"Unrecognized query target in brw_begin_query()");
290       break;
291    }
292 }
293 
294 /**
295  * Begin the ARB_occlusion_query query on a query object.
296  */
297 static void
brw_end_query(struct gl_context * ctx,struct gl_query_object * q)298 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
299 {
300    struct brw_context *brw = brw_context(ctx);
301    struct intel_context *intel = intel_context(ctx);
302    struct brw_query_object *query = (struct brw_query_object *)q;
303 
304    switch (query->Base.Target) {
305    case GL_TIMESTAMP:
306       drm_intel_bo_unreference(query->bo);
307       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query",
308 				     4096, 4096);
309       /* FALLTHROUGH */
310 
311    case GL_TIME_ELAPSED_EXT:
312       write_timestamp(intel, query->bo, 1);
313       intel_batchbuffer_flush(intel);
314       break;
315 
316    case GL_ANY_SAMPLES_PASSED:
317    case GL_SAMPLES_PASSED_ARB:
318       /* Flush the batchbuffer in case it has writes to our query BO.
319        * Have later queries write to a new query BO so that further rendering
320        * doesn't delay the collection of our results.
321        */
322       if (query->bo) {
323 	 brw_emit_query_end(brw);
324 	 intel_batchbuffer_flush(intel);
325 
326 	 drm_intel_bo_unreference(brw->query.bo);
327 	 brw->query.bo = NULL;
328       }
329 
330       brw->query.obj = NULL;
331 
332       intel->stats_wm--;
333       break;
334 
335    case GL_PRIMITIVES_GENERATED:
336       /* We don't actually query the hardware for this value; we keep track of
337        * it in a software counter.  So just read the counter and store it in
338        * the query object.
339        */
340       query->Base.Result = brw->sol.primitives_generated;
341       brw->sol.counting_primitives_generated = false;
342 
343       /* And set brw->query.obj to NULL so that this query won't try to wait
344        * for any rendering to complete.
345        */
346       query->bo = NULL;
347       break;
348 
349    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
350       /* We don't actually query the hardware for this value; we keep track of
351        * it in a software counter.  So just read the counter and store it in
352        * the query object.
353        */
354       query->Base.Result = brw->sol.primitives_written;
355       brw->sol.counting_primitives_written = false;
356 
357       /* And set brw->query.obj to NULL so that this query won't try to wait
358        * for any rendering to complete.
359        */
360       query->bo = NULL;
361       break;
362 
363    default:
364       assert(!"Unrecognized query target in brw_end_query()");
365       break;
366    }
367 }
368 
brw_wait_query(struct gl_context * ctx,struct gl_query_object * q)369 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
370 {
371    struct brw_query_object *query = (struct brw_query_object *)q;
372 
373    brw_queryobj_get_results(ctx, query);
374    query->Base.Ready = true;
375 }
376 
brw_check_query(struct gl_context * ctx,struct gl_query_object * q)377 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
378 {
379    struct brw_query_object *query = (struct brw_query_object *)q;
380 
381    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
382       brw_queryobj_get_results(ctx, query);
383       query->Base.Ready = true;
384    }
385 }
386 
387 /** Called to set up the query BO and account for its aperture space */
388 void
brw_prepare_query_begin(struct brw_context * brw)389 brw_prepare_query_begin(struct brw_context *brw)
390 {
391    struct intel_context *intel = &brw->intel;
392 
393    /* Skip if we're not doing any queries. */
394    if (!brw->query.obj)
395       return;
396 
397    /* Get a new query BO if we're going to need it. */
398    if (brw->query.bo == NULL ||
399        brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
400       drm_intel_bo_unreference(brw->query.bo);
401       brw->query.bo = NULL;
402 
403       brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
404 
405       /* clear target buffer */
406       drm_intel_bo_map(brw->query.bo, true);
407       memset((char *)brw->query.bo->virtual, 0, 4096);
408       drm_intel_bo_unmap(brw->query.bo);
409 
410       brw->query.index = 0;
411    }
412 }
413 
414 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
415 void
brw_emit_query_begin(struct brw_context * brw)416 brw_emit_query_begin(struct brw_context *brw)
417 {
418    struct intel_context *intel = &brw->intel;
419    struct gl_context *ctx = &intel->ctx;
420    struct brw_query_object *query = brw->query.obj;
421 
422    /* Skip if we're not doing any queries, or we've emitted the start. */
423    if (!query || brw->query.active)
424       return;
425 
426    write_depth_count(intel, brw->query.bo, brw->query.index * 2);
427 
428    if (query->bo != brw->query.bo) {
429       if (query->bo != NULL)
430 	 brw_queryobj_get_results(ctx, query);
431       drm_intel_bo_reference(brw->query.bo);
432       query->bo = brw->query.bo;
433       query->first_index = brw->query.index;
434    }
435    query->last_index = brw->query.index;
436    brw->query.active = true;
437 }
438 
439 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
440 void
brw_emit_query_end(struct brw_context * brw)441 brw_emit_query_end(struct brw_context *brw)
442 {
443    struct intel_context *intel = &brw->intel;
444 
445    if (!brw->query.active)
446       return;
447 
448    write_depth_count(intel, brw->query.bo, brw->query.index * 2 + 1);
449 
450    brw->query.active = false;
451    brw->query.index++;
452 }
453 
454 static uint64_t
brw_get_timestamp(struct gl_context * ctx)455 brw_get_timestamp(struct gl_context *ctx)
456 {
457    struct intel_context *intel = intel_context(ctx);
458    uint64_t result = 0;
459 
460    drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
461 
462    /* See logic in brw_queryobj_get_results() */
463    result = result >> 32;
464    result *= 80;
465    result &= (1ull << 36) - 1;
466 
467    return result;
468 }
469 
brw_init_queryobj_functions(struct dd_function_table * functions)470 void brw_init_queryobj_functions(struct dd_function_table *functions)
471 {
472    functions->NewQueryObject = brw_new_query_object;
473    functions->DeleteQuery = brw_delete_query;
474    functions->BeginQuery = brw_begin_query;
475    functions->EndQuery = brw_end_query;
476    functions->CheckQuery = brw_check_query;
477    functions->WaitQuery = brw_wait_query;
478    functions->GetTimestamp = brw_get_timestamp;
479 }
480