2 * Copyright © 2008 Intel Corporation
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:
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
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
24 * Eric Anholt <eric@anholt.net>
28 /** @file brw_queryobj.c
30 * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query,
31 * GL_EXT_transform_feedback, and friends).
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'
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).
43 #include "main/imports.h"
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"
52 write_timestamp(struct intel_context
*intel
, drm_intel_bo
*query_bo
, int idx
)
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().
62 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (4 - 2));
63 OUT_BATCH(PIPE_CONTROL_CS_STALL
| PIPE_CONTROL_STALL_AT_SCOREBOARD
);
70 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (5 - 2));
71 OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP
);
73 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
74 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
75 idx
* sizeof(uint64_t));
81 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (4 - 2) |
82 PIPE_CONTROL_WRITE_TIMESTAMP
);
84 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
85 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
86 idx
* sizeof(uint64_t));
94 write_depth_count(struct intel_context
*intel
, drm_intel_bo
*query_bo
, int idx
)
96 if (intel
->gen
>= 6) {
97 /* Emit Sandybridge workaround flush: */
99 intel_emit_post_sync_nonzero_flush(intel
);
102 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (5 - 2));
103 OUT_BATCH(PIPE_CONTROL_DEPTH_STALL
|
104 PIPE_CONTROL_WRITE_DEPTH_COUNT
);
106 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
107 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
108 (idx
* sizeof(uint64_t)));
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.
123 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
124 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
125 (idx
* sizeof(uint64_t)));
132 /** Waits on the query object's BO and totals the results for this query */
134 brw_queryobj_get_results(struct gl_context
*ctx
,
135 struct brw_query_object
*query
)
137 struct intel_context
*intel
= intel_context(ctx
);
142 if (query
->bo
== NULL
)
145 if (drm_intel_bo_references(intel
->batch
.bo
, query
->bo
))
146 intel_batchbuffer_flush(intel
);
148 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
)) {
149 if (drm_intel_bo_busy(query
->bo
)) {
150 perf_debug("Stalling on the GPU waiting for a query object.\n");
154 drm_intel_bo_map(query
->bo
, false);
155 results
= query
->bo
->virtual;
156 switch (query
->Base
.Target
) {
157 case GL_TIME_ELAPSED_EXT
:
159 query
->Base
.Result
+= 80 * (results
[1] - results
[0]);
161 query
->Base
.Result
+= 1000 * ((results
[1] >> 32) - (results
[0] >> 32));
165 if (intel
->gen
>= 6) {
166 /* Our timer is a clock that increments every 80ns (regardless of
167 * other clock scaling in the system). The timestamp register we can
168 * read for glGetTimestamp() masks out the top 32 bits, so we do that
169 * here too to let the two counters be compared against each other.
171 * If we just multiplied that 32 bits of data by 80, it would roll
172 * over at a non-power-of-two, so an application couldn't use
173 * GL_QUERY_COUNTER_BITS to handle rollover correctly. Instead, we
174 * report 36 bits and truncate at that (rolling over 5 times as often
175 * as the HW counter), and when the 32-bit counter rolls over, it
176 * happens to also be at a rollover in the reported value from near
179 * The low 32 bits rolls over in ~343 seconds. Our 36-bit result
180 * rolls over every ~69 seconds.
182 query
->Base
.Result
= 80 * (results
[1] & 0xffffffff);
183 query
->Base
.Result
&= (1ull << 36) - 1;
185 query
->Base
.Result
= 1000 * (results
[1] >> 32);
190 case GL_SAMPLES_PASSED_ARB
:
191 /* Map and count the pixels from the current query BO */
192 for (i
= query
->first_index
; i
<= query
->last_index
; i
++) {
193 query
->Base
.Result
+= results
[i
* 2 + 1] - results
[i
* 2];
197 case GL_ANY_SAMPLES_PASSED
:
198 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
199 /* Set true if any of the sub-queries passed. */
200 for (i
= query
->first_index
; i
<= query
->last_index
; i
++) {
201 if (results
[i
* 2 + 1] != results
[i
* 2]) {
202 query
->Base
.Result
= GL_TRUE
;
208 case GL_PRIMITIVES_GENERATED
:
209 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
210 /* We don't actually query the hardware for this value, so query->bo
211 * should always be NULL and execution should never reach here.
213 assert(!"Unreachable");
217 assert(!"Unrecognized query target in brw_queryobj_get_results()");
220 drm_intel_bo_unmap(query
->bo
);
222 drm_intel_bo_unreference(query
->bo
);
226 static struct gl_query_object
*
227 brw_new_query_object(struct gl_context
*ctx
, GLuint id
)
229 struct brw_query_object
*query
;
231 query
= calloc(1, sizeof(struct brw_query_object
));
234 query
->Base
.Result
= 0;
235 query
->Base
.Active
= false;
236 query
->Base
.Ready
= true;
242 brw_delete_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
244 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
246 drm_intel_bo_unreference(query
->bo
);
251 brw_begin_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
253 struct brw_context
*brw
= brw_context(ctx
);
254 struct intel_context
*intel
= intel_context(ctx
);
255 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
257 switch (query
->Base
.Target
) {
258 case GL_TIME_ELAPSED_EXT
:
259 drm_intel_bo_unreference(query
->bo
);
260 query
->bo
= drm_intel_bo_alloc(intel
->bufmgr
, "timer query", 4096, 4096);
261 write_timestamp(intel
, query
->bo
, 0);
264 case GL_ANY_SAMPLES_PASSED
:
265 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
266 case GL_SAMPLES_PASSED_ARB
:
267 /* Reset our driver's tracking of query state. */
268 drm_intel_bo_unreference(query
->bo
);
270 query
->first_index
= -1;
271 query
->last_index
= -1;
273 brw
->query
.obj
= query
;
277 case GL_PRIMITIVES_GENERATED
:
278 /* We don't actually query the hardware for this value; we keep track of
279 * it a software counter. So just reset the counter.
281 brw
->sol
.primitives_generated
= 0;
282 brw
->sol
.counting_primitives_generated
= true;
285 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
286 /* We don't actually query the hardware for this value; we keep track of
287 * it a software counter. So just reset the counter.
289 brw
->sol
.primitives_written
= 0;
290 brw
->sol
.counting_primitives_written
= true;
294 assert(!"Unrecognized query target in brw_begin_query()");
300 * Begin the ARB_occlusion_query query on a query object.
303 brw_end_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
305 struct brw_context
*brw
= brw_context(ctx
);
306 struct intel_context
*intel
= intel_context(ctx
);
307 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
309 switch (query
->Base
.Target
) {
311 drm_intel_bo_unreference(query
->bo
);
312 query
->bo
= drm_intel_bo_alloc(intel
->bufmgr
, "timer query",
316 case GL_TIME_ELAPSED_EXT
:
317 write_timestamp(intel
, query
->bo
, 1);
320 case GL_ANY_SAMPLES_PASSED
:
321 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
322 case GL_SAMPLES_PASSED_ARB
:
324 /* No query->bo means that EndQuery was called after BeginQuery with no
325 * intervening drawing. Rather than doing nothing at all here in this
326 * case, we emit the query_begin and query_end state to the
327 * hardware. This is to guarantee that waiting on the result of this
328 * empty state will cause all previous queries to complete at all, as
329 * required by the specification:
331 * It must always be true that if any query object
332 * returns a result available of TRUE, all queries of the
333 * same type issued prior to that query must also return
334 * TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
337 brw_emit_query_begin(brw
);
341 brw_emit_query_end(brw
);
343 drm_intel_bo_unreference(brw
->query
.bo
);
344 brw
->query
.bo
= NULL
;
347 brw
->query
.obj
= NULL
;
352 case GL_PRIMITIVES_GENERATED
:
353 /* We don't actually query the hardware for this value; we keep track of
354 * it in a software counter. So just read the counter and store it in
357 query
->Base
.Result
= brw
->sol
.primitives_generated
;
358 brw
->sol
.counting_primitives_generated
= false;
360 /* And set brw->query.obj to NULL so that this query won't try to wait
361 * for any rendering to complete.
366 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
367 /* We don't actually query the hardware for this value; we keep track of
368 * it in a software counter. So just read the counter and store it in
371 query
->Base
.Result
= brw
->sol
.primitives_written
;
372 brw
->sol
.counting_primitives_written
= false;
374 /* And set brw->query.obj to NULL so that this query won't try to wait
375 * for any rendering to complete.
381 assert(!"Unrecognized query target in brw_end_query()");
386 static void brw_wait_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
388 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
390 brw_queryobj_get_results(ctx
, query
);
391 query
->Base
.Ready
= true;
394 static void brw_check_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
396 struct intel_context
*intel
= intel_context(ctx
);
397 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
399 /* From the GL_ARB_occlusion_query spec:
401 * "Instead of allowing for an infinite loop, performing a
402 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
403 * not ready yet on the first time it is queried. This ensures that
404 * the async query will return true in finite time.
406 if (query
->bo
&& drm_intel_bo_references(intel
->batch
.bo
, query
->bo
))
407 intel_batchbuffer_flush(intel
);
409 if (query
->bo
== NULL
|| !drm_intel_bo_busy(query
->bo
)) {
410 brw_queryobj_get_results(ctx
, query
);
411 query
->Base
.Ready
= true;
415 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
417 brw_emit_query_begin(struct brw_context
*brw
)
419 struct intel_context
*intel
= &brw
->intel
;
420 struct gl_context
*ctx
= &intel
->ctx
;
421 struct brw_query_object
*query
= brw
->query
.obj
;
423 /* Skip if we're not doing any queries, or we've emitted the start. */
424 if (!query
|| brw
->query
.begin_emitted
)
427 /* Get a new query BO if we're going to need it. */
428 if (brw
->query
.bo
== NULL
||
429 brw
->query
.index
* 2 + 1 >= 4096 / sizeof(uint64_t)) {
430 drm_intel_bo_unreference(brw
->query
.bo
);
431 brw
->query
.bo
= NULL
;
433 brw
->query
.bo
= drm_intel_bo_alloc(intel
->bufmgr
, "query", 4096, 1);
435 /* clear target buffer */
436 drm_intel_bo_map(brw
->query
.bo
, true);
437 memset((char *)brw
->query
.bo
->virtual, 0, 4096);
438 drm_intel_bo_unmap(brw
->query
.bo
);
440 brw
->query
.index
= 0;
443 write_depth_count(intel
, brw
->query
.bo
, brw
->query
.index
* 2);
445 if (query
->bo
!= brw
->query
.bo
) {
446 if (query
->bo
!= NULL
)
447 brw_queryobj_get_results(ctx
, query
);
448 drm_intel_bo_reference(brw
->query
.bo
);
449 query
->bo
= brw
->query
.bo
;
450 query
->first_index
= brw
->query
.index
;
452 query
->last_index
= brw
->query
.index
;
453 brw
->query
.begin_emitted
= true;
456 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
458 brw_emit_query_end(struct brw_context
*brw
)
460 struct intel_context
*intel
= &brw
->intel
;
462 if (!brw
->query
.begin_emitted
)
465 write_depth_count(intel
, brw
->query
.bo
, brw
->query
.index
* 2 + 1);
467 brw
->query
.begin_emitted
= false;
472 brw_get_timestamp(struct gl_context
*ctx
)
474 struct intel_context
*intel
= intel_context(ctx
);
477 drm_intel_reg_read(intel
->bufmgr
, TIMESTAMP
, &result
);
479 /* See logic in brw_queryobj_get_results() */
480 result
= result
>> 32;
482 result
&= (1ull << 36) - 1;
487 void brw_init_queryobj_functions(struct dd_function_table
*functions
)
489 functions
->NewQueryObject
= brw_new_query_object
;
490 functions
->DeleteQuery
= brw_delete_query
;
491 functions
->BeginQuery
= brw_begin_query
;
492 functions
->EndQuery
= brw_end_query
;
493 functions
->CheckQuery
= brw_check_query
;
494 functions
->WaitQuery
= brw_wait_query
;
495 functions
->GetTimestamp
= brw_get_timestamp
;