brw_queryobj.c revision 7540f25a345ed8d35d97ba4fbf66227a281eff6a 
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
51static void
52write_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
93static void
94write_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 */
133static void
134brw_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_PRIMITIVES_GENERATED:
195   case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
196      /* We don't actually query the hardware for this value, so query->bo
197       * should always be NULL and execution should never reach here.
198       */
199      assert(!"Unreachable");
200      break;
201
202   default:
203      assert(!"Unrecognized query target in brw_queryobj_get_results()");
204      break;
205   }
206   drm_intel_bo_unmap(query->bo);
207
208   drm_intel_bo_unreference(query->bo);
209   query->bo = NULL;
210}
211
212static struct gl_query_object *
213brw_new_query_object(struct gl_context *ctx, GLuint id)
214{
215   struct brw_query_object *query;
216
217   query = calloc(1, sizeof(struct brw_query_object));
218
219   query->Base.Id = id;
220   query->Base.Result = 0;
221   query->Base.Active = false;
222   query->Base.Ready = true;
223
224   return &query->Base;
225}
226
227static void
228brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
229{
230   struct brw_query_object *query = (struct brw_query_object *)q;
231
232   drm_intel_bo_unreference(query->bo);
233   free(query);
234}
235
236static void
237brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
238{
239   struct brw_context *brw = brw_context(ctx);
240   struct intel_context *intel = intel_context(ctx);
241   struct brw_query_object *query = (struct brw_query_object *)q;
242
243   switch (query->Base.Target) {
244   case GL_TIME_ELAPSED_EXT:
245      drm_intel_bo_unreference(query->bo);
246      query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
247      write_timestamp(intel, query->bo, 0);
248      break;
249
250   case GL_SAMPLES_PASSED_ARB:
251      /* Reset our driver's tracking of query state. */
252      drm_intel_bo_unreference(query->bo);
253      query->bo = NULL;
254      query->first_index = -1;
255      query->last_index = -1;
256
257      brw->query.obj = query;
258      intel->stats_wm++;
259      break;
260
261   case GL_PRIMITIVES_GENERATED:
262      /* We don't actually query the hardware for this value; we keep track of
263       * it a software counter.  So just reset the counter.
264       */
265      brw->sol.primitives_generated = 0;
266      brw->sol.counting_primitives_generated = true;
267      break;
268
269   case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
270      /* We don't actually query the hardware for this value; we keep track of
271       * it a software counter.  So just reset the counter.
272       */
273      brw->sol.primitives_written = 0;
274      brw->sol.counting_primitives_written = true;
275      break;
276
277   default:
278      assert(!"Unrecognized query target in brw_begin_query()");
279      break;
280   }
281}
282
283/**
284 * Begin the ARB_occlusion_query query on a query object.
285 */
286static void
287brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
288{
289   struct brw_context *brw = brw_context(ctx);
290   struct intel_context *intel = intel_context(ctx);
291   struct brw_query_object *query = (struct brw_query_object *)q;
292
293   switch (query->Base.Target) {
294   case GL_TIMESTAMP:
295      drm_intel_bo_unreference(query->bo);
296      query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query",
297				     4096, 4096);
298      /* FALLTHROUGH */
299
300   case GL_TIME_ELAPSED_EXT:
301      write_timestamp(intel, query->bo, 1);
302      intel_batchbuffer_flush(intel);
303      break;
304
305   case GL_SAMPLES_PASSED_ARB:
306      /* Flush the batchbuffer in case it has writes to our query BO.
307       * Have later queries write to a new query BO so that further rendering
308       * doesn't delay the collection of our results.
309       */
310      if (query->bo) {
311	 brw_emit_query_end(brw);
312	 intel_batchbuffer_flush(intel);
313
314	 drm_intel_bo_unreference(brw->query.bo);
315	 brw->query.bo = NULL;
316      }
317
318      brw->query.obj = NULL;
319
320      intel->stats_wm--;
321      break;
322
323   case GL_PRIMITIVES_GENERATED:
324      /* We don't actually query the hardware for this value; we keep track of
325       * it in a software counter.  So just read the counter and store it in
326       * the query object.
327       */
328      query->Base.Result = brw->sol.primitives_generated;
329      brw->sol.counting_primitives_generated = false;
330
331      /* And set brw->query.obj to NULL so that this query won't try to wait
332       * for any rendering to complete.
333       */
334      query->bo = NULL;
335      break;
336
337   case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
338      /* We don't actually query the hardware for this value; we keep track of
339       * it in a software counter.  So just read the counter and store it in
340       * the query object.
341       */
342      query->Base.Result = brw->sol.primitives_written;
343      brw->sol.counting_primitives_written = false;
344
345      /* And set brw->query.obj to NULL so that this query won't try to wait
346       * for any rendering to complete.
347       */
348      query->bo = NULL;
349      break;
350
351   default:
352      assert(!"Unrecognized query target in brw_end_query()");
353      break;
354   }
355}
356
357static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
358{
359   struct brw_query_object *query = (struct brw_query_object *)q;
360
361   brw_queryobj_get_results(ctx, query);
362   query->Base.Ready = true;
363}
364
365static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
366{
367   struct brw_query_object *query = (struct brw_query_object *)q;
368
369   if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
370      brw_queryobj_get_results(ctx, query);
371      query->Base.Ready = true;
372   }
373}
374
375/** Called to set up the query BO and account for its aperture space */
376void
377brw_prepare_query_begin(struct brw_context *brw)
378{
379   struct intel_context *intel = &brw->intel;
380
381   /* Skip if we're not doing any queries. */
382   if (!brw->query.obj)
383      return;
384
385   /* Get a new query BO if we're going to need it. */
386   if (brw->query.bo == NULL ||
387       brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
388      drm_intel_bo_unreference(brw->query.bo);
389      brw->query.bo = NULL;
390
391      brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
392
393      /* clear target buffer */
394      drm_intel_bo_map(brw->query.bo, true);
395      memset((char *)brw->query.bo->virtual, 0, 4096);
396      drm_intel_bo_unmap(brw->query.bo);
397
398      brw->query.index = 0;
399   }
400}
401
402/** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
403void
404brw_emit_query_begin(struct brw_context *brw)
405{
406   struct intel_context *intel = &brw->intel;
407   struct gl_context *ctx = &intel->ctx;
408   struct brw_query_object *query = brw->query.obj;
409
410   /* Skip if we're not doing any queries, or we've emitted the start. */
411   if (!query || brw->query.active)
412      return;
413
414   write_depth_count(intel, brw->query.bo, brw->query.index * 2);
415
416   if (query->bo != brw->query.bo) {
417      if (query->bo != NULL)
418	 brw_queryobj_get_results(ctx, query);
419      drm_intel_bo_reference(brw->query.bo);
420      query->bo = brw->query.bo;
421      query->first_index = brw->query.index;
422   }
423   query->last_index = brw->query.index;
424   brw->query.active = true;
425}
426
427/** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
428void
429brw_emit_query_end(struct brw_context *brw)
430{
431   struct intel_context *intel = &brw->intel;
432
433   if (!brw->query.active)
434      return;
435
436   write_depth_count(intel, brw->query.bo, brw->query.index * 2 + 1);
437
438   brw->query.active = false;
439   brw->query.index++;
440}
441
442static uint64_t
443brw_get_timestamp(struct gl_context *ctx)
444{
445   struct intel_context *intel = intel_context(ctx);
446   uint64_t result = 0;
447
448   drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
449
450   /* See logic in brw_queryobj_get_results() */
451   result = result >> 32;
452   result *= 80;
453   result &= (1ull << 36) - 1;
454
455   return result;
456}
457
458void brw_init_queryobj_functions(struct dd_function_table *functions)
459{
460   functions->NewQueryObject = brw_new_query_object;
461   functions->DeleteQuery = brw_delete_query;
462   functions->BeginQuery = brw_begin_query;
463   functions->EndQuery = brw_end_query;
464   functions->CheckQuery = brw_check_query;
465   functions->WaitQuery = brw_wait_query;
466   functions->GetTimestamp = brw_get_timestamp;
467}
468