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 #include "main/imports.h"
40 #include "brw_context.h"
41 #include "brw_defines.h"
42 #include "brw_state.h"
43 #include "intel_batchbuffer.h"
44 #include "intel_reg.h"
47 * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
50 write_timestamp(struct brw_context
*brw
, drm_intel_bo
*query_bo
, int idx
)
53 /* Emit workaround flushes: */
55 /* The timestamp write below is a non-zero post-sync op, which on
56 * Gen6 necessitates a CS stall. CS stalls need stall at scoreboard
57 * set. See the comments for intel_emit_post_sync_nonzero_flush().
60 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (4 - 2));
61 OUT_BATCH(PIPE_CONTROL_CS_STALL
| PIPE_CONTROL_STALL_AT_SCOREBOARD
);
68 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (5 - 2));
69 OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP
);
71 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
72 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
73 idx
* sizeof(uint64_t));
79 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (4 - 2) |
80 PIPE_CONTROL_WRITE_TIMESTAMP
);
82 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
83 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
84 idx
* sizeof(uint64_t));
92 * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
95 write_depth_count(struct brw_context
*brw
, drm_intel_bo
*query_bo
, int idx
)
100 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (4 - 2) |
101 PIPE_CONTROL_DEPTH_STALL
| PIPE_CONTROL_WRITE_DEPTH_COUNT
);
102 /* This object could be mapped cacheable, but we don't have an exposed
103 * mechanism to support that. Since it's going uncached, tell GEM that
104 * we're writing to it. The usual clflush should be all that's required
105 * to pick up the results.
108 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
109 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
110 (idx
* sizeof(uint64_t)));
117 * Wait on the query object's BO and calculate the final result.
120 brw_queryobj_get_results(struct gl_context
*ctx
,
121 struct brw_query_object
*query
)
123 struct brw_context
*brw
= brw_context(ctx
);
128 assert(brw
->gen
< 6);
130 if (query
->bo
== NULL
)
133 /* If the application has requested the query result, but this batch is
134 * still contributing to it, flush it now so the results will be present
137 if (drm_intel_bo_references(brw
->batch
.bo
, query
->bo
))
138 intel_batchbuffer_flush(brw
);
140 if (unlikely(brw
->perf_debug
)) {
141 if (drm_intel_bo_busy(query
->bo
)) {
142 perf_debug("Stalling on the GPU waiting for a query object.\n");
146 drm_intel_bo_map(query
->bo
, false);
147 results
= query
->bo
->virtual;
148 switch (query
->Base
.Target
) {
149 case GL_TIME_ELAPSED_EXT
:
150 /* The query BO contains the starting and ending timestamps.
151 * Subtract the two and convert to nanoseconds.
153 query
->Base
.Result
+= 1000 * ((results
[1] >> 32) - (results
[0] >> 32));
157 /* The query BO contains a single timestamp value in results[0]. */
158 query
->Base
.Result
= 1000 * (results
[0] >> 32);
161 case GL_SAMPLES_PASSED_ARB
:
162 /* Loop over pairs of values from the BO, which are the PS_DEPTH_COUNT
163 * value at the start and end of the batchbuffer. Subtract them to
164 * get the number of fragments which passed the depth test in each
165 * individual batch, and add those differences up to get the number
166 * of fragments for the entire query.
168 * Note that query->Base.Result may already be non-zero. We may have
169 * run out of space in the query's BO and allocated a new one. If so,
170 * this function was already called to accumulate the results so far.
172 for (i
= 0; i
< query
->last_index
; i
++) {
173 query
->Base
.Result
+= results
[i
* 2 + 1] - results
[i
* 2];
177 case GL_ANY_SAMPLES_PASSED
:
178 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
179 /* If the starting and ending PS_DEPTH_COUNT from any of the batches
180 * differ, then some fragments passed the depth test.
182 for (i
= 0; i
< query
->last_index
; i
++) {
183 if (results
[i
* 2 + 1] != results
[i
* 2]) {
184 query
->Base
.Result
= GL_TRUE
;
191 assert(!"Unrecognized query target in brw_queryobj_get_results()");
194 drm_intel_bo_unmap(query
->bo
);
196 /* Now that we've processed the data stored in the query's buffer object,
199 drm_intel_bo_unreference(query
->bo
);
204 * The NewQueryObject() driver hook.
206 * Allocates and initializes a new query object.
208 static struct gl_query_object
*
209 brw_new_query_object(struct gl_context
*ctx
, GLuint id
)
211 struct brw_query_object
*query
;
213 query
= calloc(1, sizeof(struct brw_query_object
));
216 query
->Base
.Result
= 0;
217 query
->Base
.Active
= false;
218 query
->Base
.Ready
= true;
224 * The DeleteQuery() driver hook.
227 brw_delete_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
229 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
231 drm_intel_bo_unreference(query
->bo
);
236 * Gen4-5 driver hook for glBeginQuery().
238 * Initializes driver structures and emits any GPU commands required to begin
239 * recording data for the query.
242 brw_begin_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
244 struct brw_context
*brw
= brw_context(ctx
);
245 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
247 assert(brw
->gen
< 6);
249 switch (query
->Base
.Target
) {
250 case GL_TIME_ELAPSED_EXT
:
251 /* For timestamp queries, we record the starting time right away so that
252 * we measure the full time between BeginQuery and EndQuery. There's
253 * some debate about whether this is the right thing to do. Our decision
254 * is based on the following text from the ARB_timer_query extension:
256 * "(5) Should the extension measure total time elapsed between the full
257 * completion of the BeginQuery and EndQuery commands, or just time
258 * spent in the graphics library?
260 * RESOLVED: This extension will measure the total time elapsed
261 * between the full completion of these commands. Future extensions
262 * may implement a query to determine time elapsed at different stages
263 * of the graphics pipeline."
265 * We write a starting timestamp now (at index 0). At EndQuery() time,
266 * we'll write a second timestamp (at index 1), and subtract the two to
267 * obtain the time elapsed. Notably, this includes time elapsed while
268 * the system was doing other work, such as running other applications.
270 drm_intel_bo_unreference(query
->bo
);
271 query
->bo
= drm_intel_bo_alloc(brw
->bufmgr
, "timer query", 4096, 4096);
272 write_timestamp(brw
, query
->bo
, 0);
275 case GL_ANY_SAMPLES_PASSED
:
276 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
277 case GL_SAMPLES_PASSED_ARB
:
278 /* For occlusion queries, we delay taking an initial sample until the
279 * first drawing occurs in this batch. See the reasoning in the comments
280 * for brw_emit_query_begin() below.
282 * Since we're starting a new query, we need to be sure to throw away
283 * any previous occlusion query results.
285 drm_intel_bo_unreference(query
->bo
);
287 query
->last_index
= -1;
289 brw
->query
.obj
= query
;
291 /* Depth statistics on Gen4 require strange workarounds, so we try to
292 * avoid them when necessary. They're required for occlusion queries,
293 * so turn them on now.
296 brw
->state
.dirty
.brw
|= BRW_NEW_STATS_WM
;
300 assert(!"Unrecognized query target in brw_begin_query()");
306 * Gen4-5 driver hook for glEndQuery().
308 * Emits GPU commands to record a final query value, ending any data capturing.
309 * However, the final result isn't necessarily available until the GPU processes
310 * those commands. brw_queryobj_get_results() processes the captured data to
311 * produce the final result.
314 brw_end_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
316 struct brw_context
*brw
= brw_context(ctx
);
317 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
319 assert(brw
->gen
< 6);
321 switch (query
->Base
.Target
) {
322 case GL_TIME_ELAPSED_EXT
:
323 /* Write the final timestamp. */
324 write_timestamp(brw
, query
->bo
, 1);
327 case GL_ANY_SAMPLES_PASSED
:
328 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
329 case GL_SAMPLES_PASSED_ARB
:
331 /* No query->bo means that EndQuery was called after BeginQuery with no
332 * intervening drawing. Rather than doing nothing at all here in this
333 * case, we emit the query_begin and query_end state to the
334 * hardware. This is to guarantee that waiting on the result of this
335 * empty state will cause all previous queries to complete at all, as
336 * required by the specification:
338 * It must always be true that if any query object
339 * returns a result available of TRUE, all queries of the
340 * same type issued prior to that query must also return
341 * TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
344 brw_emit_query_begin(brw
);
349 brw_emit_query_end(brw
);
351 brw
->query
.obj
= NULL
;
354 brw
->state
.dirty
.brw
|= BRW_NEW_STATS_WM
;
358 assert(!"Unrecognized query target in brw_end_query()");
364 * The Gen4-5 WaitQuery() driver hook.
366 * Wait for a query result to become available and return it. This is the
367 * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
369 static void brw_wait_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
371 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
373 assert(brw_context(ctx
)->gen
< 6);
375 brw_queryobj_get_results(ctx
, query
);
376 query
->Base
.Ready
= true;
380 * The Gen4-5 CheckQuery() driver hook.
382 * Checks whether a query result is ready yet. If not, flushes.
383 * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
385 static void brw_check_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
387 struct brw_context
*brw
= brw_context(ctx
);
388 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
390 assert(brw
->gen
< 6);
392 /* From the GL_ARB_occlusion_query spec:
394 * "Instead of allowing for an infinite loop, performing a
395 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
396 * not ready yet on the first time it is queried. This ensures that
397 * the async query will return true in finite time.
399 if (query
->bo
&& drm_intel_bo_references(brw
->batch
.bo
, query
->bo
))
400 intel_batchbuffer_flush(brw
);
402 if (query
->bo
== NULL
|| !drm_intel_bo_busy(query
->bo
)) {
403 brw_queryobj_get_results(ctx
, query
);
404 query
->Base
.Ready
= true;
409 * Ensure there query's BO has enough space to store a new pair of values.
411 * If not, gather the existing BO's results and create a new buffer of the
415 ensure_bo_has_space(struct gl_context
*ctx
, struct brw_query_object
*query
)
417 struct brw_context
*brw
= brw_context(ctx
);
419 assert(brw
->gen
< 6);
421 if (!query
->bo
|| query
->last_index
* 2 + 1 >= 4096 / sizeof(uint64_t)) {
423 if (query
->bo
!= NULL
) {
424 /* The old query BO did not have enough space, so we allocated a new
425 * one. Gather the results so far (adding up the differences) and
426 * release the old BO.
428 brw_queryobj_get_results(ctx
, query
);
431 query
->bo
= drm_intel_bo_alloc(brw
->bufmgr
, "query", 4096, 1);
432 query
->last_index
= 0;
437 * Record the PS_DEPTH_COUNT value (for occlusion queries) just before
440 * In a pre-hardware context world, the single PS_DEPTH_COUNT register is
441 * shared among all applications using the GPU. However, our query value
442 * needs to only include fragments generated by our application/GL context.
444 * To accommodate this, we record PS_DEPTH_COUNT at the start and end of
445 * each batchbuffer (technically, the first primitive drawn and flush time).
446 * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT
447 * caused by a batchbuffer. Since there is no preemption inside batches,
448 * this is guaranteed to only measure the effects of our current application.
450 * Adding each of these differences (in case drawing is done over many batches)
451 * produces the final expected value.
453 * In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
454 * as part of the context state, so this is unnecessary, and skipped.
457 brw_emit_query_begin(struct brw_context
*brw
)
459 struct gl_context
*ctx
= &brw
->ctx
;
460 struct brw_query_object
*query
= brw
->query
.obj
;
465 /* Skip if we're not doing any queries, or we've already recorded the
466 * initial query value for this batchbuffer.
468 if (!query
|| brw
->query
.begin_emitted
)
471 ensure_bo_has_space(ctx
, query
);
473 write_depth_count(brw
, query
->bo
, query
->last_index
* 2);
475 brw
->query
.begin_emitted
= true;
479 * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
480 * (for non-hardware context platforms).
482 * See the explanation in brw_emit_query_begin().
485 brw_emit_query_end(struct brw_context
*brw
)
487 struct brw_query_object
*query
= brw
->query
.obj
;
492 if (!brw
->query
.begin_emitted
)
495 write_depth_count(brw
, query
->bo
, query
->last_index
* 2 + 1);
497 brw
->query
.begin_emitted
= false;
502 * Driver hook for glQueryCounter().
504 * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
505 * current GPU time. This is unlike GL_TIME_ELAPSED, which measures the
506 * time while the query is active.
509 brw_query_counter(struct gl_context
*ctx
, struct gl_query_object
*q
)
511 struct brw_context
*brw
= brw_context(ctx
);
512 struct brw_query_object
*query
= (struct brw_query_object
*) q
;
514 assert(q
->Target
== GL_TIMESTAMP
);
516 drm_intel_bo_unreference(query
->bo
);
517 query
->bo
= drm_intel_bo_alloc(brw
->bufmgr
, "timestamp query", 4096, 4096);
518 write_timestamp(brw
, query
->bo
, 0);
522 * Read the TIMESTAMP register immediately (in a non-pipelined fashion).
524 * This is used to implement the GetTimestamp() driver hook.
527 brw_get_timestamp(struct gl_context
*ctx
)
529 struct brw_context
*brw
= brw_context(ctx
);
532 drm_intel_reg_read(brw
->bufmgr
, TIMESTAMP
, &result
);
534 /* See logic in brw_queryobj_get_results() */
535 result
= result
>> 32;
537 result
&= (1ull << 36) - 1;
542 /* Initialize query object functions used on all generations. */
543 void brw_init_common_queryobj_functions(struct dd_function_table
*functions
)
545 functions
->NewQueryObject
= brw_new_query_object
;
546 functions
->DeleteQuery
= brw_delete_query
;
547 functions
->QueryCounter
= brw_query_counter
;
548 functions
->GetTimestamp
= brw_get_timestamp
;
551 /* Initialize Gen4/5-specific query object functions. */
552 void gen4_init_queryobj_functions(struct dd_function_table
*functions
)
554 functions
->BeginQuery
= brw_begin_query
;
555 functions
->EndQuery
= brw_end_query
;
556 functions
->CheckQuery
= brw_check_query
;
557 functions
->WaitQuery
= brw_wait_query
;