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>
25 * Kenneth Graunke <kenneth@whitecape.org>
28 /** @file gen6_queryobj.c
30 * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query,
31 * GL_EXT_transform_feedback, and friends) on platforms that support
32 * hardware contexts (Gen6+).
34 #include "main/imports.h"
36 #include "brw_context.h"
37 #include "brw_defines.h"
38 #include "brw_state.h"
39 #include "perf/gen_perf.h"
40 #include "intel_batchbuffer.h"
41 #include "intel_buffer_objects.h"
44 set_query_availability(struct brw_context
*brw
, struct brw_query_object
*query
,
47 /* For platforms that support ARB_query_buffer_object, we write the
48 * query availability for "pipelined" queries.
50 * Most counter snapshots are written by the command streamer, by
51 * doing a CS stall and then MI_STORE_REGISTER_MEM. For these
52 * counters, the CS stall guarantees that the results will be
53 * available when subsequent CS commands run. So we don't need to
54 * do any additional tracking.
56 * Other counters (occlusion queries and timestamp) are written by
57 * PIPE_CONTROL, without a CS stall. This means that we can't be
58 * sure whether the writes have landed yet or not. Performing a
59 * PIPE_CONTROL with an immediate write will synchronize with
60 * those earlier writes, so we write 1 when the value has landed.
62 if (brw
->ctx
.Extensions
.ARB_query_buffer_object
&&
63 brw_is_query_pipelined(query
)) {
64 unsigned flags
= PIPE_CONTROL_WRITE_IMMEDIATE
;
67 /* Order available *after* the query results. */
68 flags
|= PIPE_CONTROL_FLUSH_ENABLE
;
70 /* Make it unavailable *before* any pipelined reads. */
71 flags
|= PIPE_CONTROL_CS_STALL
;
74 brw_emit_pipe_control_write(brw
, flags
,
75 query
->bo
, 2 * sizeof(uint64_t),
81 write_primitives_generated(struct brw_context
*brw
,
82 struct brw_bo
*query_bo
, int stream
, int idx
)
84 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
86 brw_emit_mi_flush(brw
);
88 if (devinfo
->gen
>= 7 && stream
> 0) {
89 brw_store_register_mem64(brw
, query_bo
,
90 GEN7_SO_PRIM_STORAGE_NEEDED(stream
),
91 idx
* sizeof(uint64_t));
93 brw_store_register_mem64(brw
, query_bo
, CL_INVOCATION_COUNT
,
94 idx
* sizeof(uint64_t));
99 write_xfb_primitives_written(struct brw_context
*brw
,
100 struct brw_bo
*bo
, int stream
, int idx
)
102 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
104 brw_emit_mi_flush(brw
);
106 if (devinfo
->gen
>= 7) {
107 brw_store_register_mem64(brw
, bo
, GEN7_SO_NUM_PRIMS_WRITTEN(stream
),
108 idx
* sizeof(uint64_t));
110 brw_store_register_mem64(brw
, bo
, GEN6_SO_NUM_PRIMS_WRITTEN
,
111 idx
* sizeof(uint64_t));
116 write_xfb_overflow_streams(struct gl_context
*ctx
,
117 struct brw_bo
*bo
, int stream
, int count
,
120 struct brw_context
*brw
= brw_context(ctx
);
121 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
123 brw_emit_mi_flush(brw
);
125 for (int i
= 0; i
< count
; i
++) {
126 int w_idx
= 4 * i
+ idx
;
127 int g_idx
= 4 * i
+ idx
+ 2;
129 if (devinfo
->gen
>= 7) {
130 brw_store_register_mem64(brw
, bo
,
131 GEN7_SO_NUM_PRIMS_WRITTEN(stream
+ i
),
132 g_idx
* sizeof(uint64_t));
133 brw_store_register_mem64(brw
, bo
,
134 GEN7_SO_PRIM_STORAGE_NEEDED(stream
+ i
),
135 w_idx
* sizeof(uint64_t));
137 brw_store_register_mem64(brw
, bo
,
138 GEN6_SO_NUM_PRIMS_WRITTEN
,
139 g_idx
* sizeof(uint64_t));
140 brw_store_register_mem64(brw
, bo
,
141 GEN6_SO_PRIM_STORAGE_NEEDED
,
142 w_idx
* sizeof(uint64_t));
148 check_xfb_overflow_streams(uint64_t *results
, int count
)
150 bool overflow
= false;
152 for (int i
= 0; i
< count
; i
++) {
153 uint64_t *result_i
= &results
[4 * i
];
155 if ((result_i
[3] - result_i
[2]) != (result_i
[1] - result_i
[0])) {
165 pipeline_target_to_index(int target
)
167 if (target
== GL_GEOMETRY_SHADER_INVOCATIONS
)
168 return MAX_PIPELINE_STATISTICS
- 1;
170 return target
- GL_VERTICES_SUBMITTED_ARB
;
174 emit_pipeline_stat(struct brw_context
*brw
, struct brw_bo
*bo
,
175 int stream
, int target
, int idx
)
177 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
179 /* One source of confusion is the tessellation shader statistics. The
180 * hardware has no statistics specific to the TE unit. Ideally we could have
181 * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
182 * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
183 * Unfortunately we don't have HS primitives, we only have HS invocations.
186 /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
189 static const uint32_t target_to_register
[] = {
190 IA_VERTICES_COUNT
, /* VERTICES_SUBMITTED */
191 IA_PRIMITIVES_COUNT
, /* PRIMITIVES_SUBMITTED */
192 VS_INVOCATION_COUNT
, /* VERTEX_SHADER_INVOCATIONS */
193 HS_INVOCATION_COUNT
, /* TESS_CONTROL_SHADER_PATCHES */
194 DS_INVOCATION_COUNT
, /* TESS_EVALUATION_SHADER_INVOCATIONS */
195 GS_PRIMITIVES_COUNT
, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
196 PS_INVOCATION_COUNT
, /* FRAGMENT_SHADER_INVOCATIONS */
197 CS_INVOCATION_COUNT
, /* COMPUTE_SHADER_INVOCATIONS */
198 CL_INVOCATION_COUNT
, /* CLIPPING_INPUT_PRIMITIVES */
199 CL_PRIMITIVES_COUNT
, /* CLIPPING_OUTPUT_PRIMITIVES */
200 GS_INVOCATION_COUNT
/* This one is special... */
202 STATIC_ASSERT(ARRAY_SIZE(target_to_register
) == MAX_PIPELINE_STATISTICS
);
203 uint32_t reg
= target_to_register
[pipeline_target_to_index(target
)];
204 /* Gen6 GS code counts full primitives, that is, it won't count individual
205 * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
207 if (devinfo
->gen
== 6 && target
== GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
)
208 reg
= CL_INVOCATION_COUNT
;
211 /* Emit a flush to make sure various parts of the pipeline are complete and
212 * we get an accurate value
214 brw_emit_mi_flush(brw
);
216 brw_store_register_mem64(brw
, bo
, reg
, idx
* sizeof(uint64_t));
221 * Wait on the query object's BO and calculate the final result.
224 gen6_queryobj_get_results(struct gl_context
*ctx
,
225 struct brw_query_object
*query
)
227 struct brw_context
*brw
= brw_context(ctx
);
228 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
230 if (query
->bo
== NULL
)
233 uint64_t *results
= brw_bo_map(brw
, query
->bo
, MAP_READ
);
234 switch (query
->Base
.Target
) {
235 case GL_TIME_ELAPSED
:
236 /* The query BO contains the starting and ending timestamps.
237 * Subtract the two and convert to nanoseconds.
239 query
->Base
.Result
= brw_raw_timestamp_delta(brw
, results
[0], results
[1]);
240 query
->Base
.Result
= gen_device_info_timebase_scale(devinfo
, query
->Base
.Result
);
244 /* The query BO contains a single timestamp value in results[0]. */
245 query
->Base
.Result
= gen_device_info_timebase_scale(devinfo
, results
[0]);
247 /* Ensure the scaled timestamp overflows according to
248 * GL_QUERY_COUNTER_BITS
250 query
->Base
.Result
&= (1ull << ctx
->Const
.QueryCounterBits
.Timestamp
) - 1;
253 case GL_SAMPLES_PASSED_ARB
:
254 /* We need to use += rather than = here since some BLT-based operations
255 * may have added additional samples to our occlusion query value.
257 query
->Base
.Result
+= results
[1] - results
[0];
260 case GL_ANY_SAMPLES_PASSED
:
261 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
262 if (results
[0] != results
[1])
263 query
->Base
.Result
= true;
266 case GL_PRIMITIVES_GENERATED
:
267 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
268 case GL_VERTICES_SUBMITTED_ARB
:
269 case GL_PRIMITIVES_SUBMITTED_ARB
:
270 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
271 case GL_GEOMETRY_SHADER_INVOCATIONS
:
272 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
273 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
274 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
275 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
276 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
277 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
278 query
->Base
.Result
= results
[1] - results
[0];
281 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
282 query
->Base
.Result
= check_xfb_overflow_streams(results
, 1);
285 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
286 query
->Base
.Result
= check_xfb_overflow_streams(results
, MAX_VERTEX_STREAMS
);
289 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
290 query
->Base
.Result
= (results
[1] - results
[0]);
291 /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
292 * "Invocation counter is 4 times actual. WA: SW to divide HW reported
293 * PS Invocations value by 4."
295 * Prior to Haswell, invocation count was counted by the WM, and it
296 * buggily counted invocations in units of subspans (2x2 unit). To get the
297 * correct value, the CS multiplied this by 4. With HSW the logic moved,
298 * and correctly emitted the number of pixel shader invocations, but,
299 * whomever forgot to undo the multiply by 4.
301 if (devinfo
->gen
== 8 || devinfo
->is_haswell
)
302 query
->Base
.Result
/= 4;
306 unreachable("Unrecognized query target in brw_queryobj_get_results()");
308 brw_bo_unmap(query
->bo
);
310 /* Now that we've processed the data stored in the query's buffer object,
313 brw_bo_unreference(query
->bo
);
316 query
->Base
.Ready
= true;
320 * Driver hook for glBeginQuery().
322 * Initializes driver structures and emits any GPU commands required to begin
323 * recording data for the query.
326 gen6_begin_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
328 struct brw_context
*brw
= brw_context(ctx
);
329 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
331 /* Since we're starting a new query, we need to throw away old results. */
332 brw_bo_unreference(query
->bo
);
334 brw_bo_alloc(brw
->bufmgr
, "query results", 4096, BRW_MEMZONE_OTHER
);
336 /* For ARB_query_buffer_object: The result is not available */
337 set_query_availability(brw
, query
, false);
339 switch (query
->Base
.Target
) {
340 case GL_TIME_ELAPSED
:
341 /* For timestamp queries, we record the starting time right away so that
342 * we measure the full time between BeginQuery and EndQuery. There's
343 * some debate about whether this is the right thing to do. Our decision
344 * is based on the following text from the ARB_timer_query extension:
346 * "(5) Should the extension measure total time elapsed between the full
347 * completion of the BeginQuery and EndQuery commands, or just time
348 * spent in the graphics library?
350 * RESOLVED: This extension will measure the total time elapsed
351 * between the full completion of these commands. Future extensions
352 * may implement a query to determine time elapsed at different stages
353 * of the graphics pipeline."
355 * We write a starting timestamp now (at index 0). At EndQuery() time,
356 * we'll write a second timestamp (at index 1), and subtract the two to
357 * obtain the time elapsed. Notably, this includes time elapsed while
358 * the system was doing other work, such as running other applications.
360 brw_write_timestamp(brw
, query
->bo
, 0);
363 case GL_ANY_SAMPLES_PASSED
:
364 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
365 case GL_SAMPLES_PASSED_ARB
:
366 brw_write_depth_count(brw
, query
->bo
, 0);
369 case GL_PRIMITIVES_GENERATED
:
370 write_primitives_generated(brw
, query
->bo
, query
->Base
.Stream
, 0);
371 if (query
->Base
.Stream
== 0)
372 ctx
->NewDriverState
|= BRW_NEW_RASTERIZER_DISCARD
;
375 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
376 write_xfb_primitives_written(brw
, query
->bo
, query
->Base
.Stream
, 0);
379 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
380 write_xfb_overflow_streams(ctx
, query
->bo
, query
->Base
.Stream
, 1, 0);
383 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
384 write_xfb_overflow_streams(ctx
, query
->bo
, 0, MAX_VERTEX_STREAMS
, 0);
387 case GL_VERTICES_SUBMITTED_ARB
:
388 case GL_PRIMITIVES_SUBMITTED_ARB
:
389 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
390 case GL_GEOMETRY_SHADER_INVOCATIONS
:
391 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
392 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
393 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
394 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
395 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
396 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
397 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
398 emit_pipeline_stat(brw
, query
->bo
, query
->Base
.Stream
, query
->Base
.Target
, 0);
402 unreachable("Unrecognized query target in brw_begin_query()");
407 * Driver hook for glEndQuery().
409 * Emits GPU commands to record a final query value, ending any data capturing.
410 * However, the final result isn't necessarily available until the GPU processes
411 * those commands. brw_queryobj_get_results() processes the captured data to
412 * produce the final result.
415 gen6_end_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
417 struct brw_context
*brw
= brw_context(ctx
);
418 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
420 switch (query
->Base
.Target
) {
421 case GL_TIME_ELAPSED
:
422 brw_write_timestamp(brw
, query
->bo
, 1);
425 case GL_ANY_SAMPLES_PASSED
:
426 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
427 case GL_SAMPLES_PASSED_ARB
:
428 brw_write_depth_count(brw
, query
->bo
, 1);
431 case GL_PRIMITIVES_GENERATED
:
432 write_primitives_generated(brw
, query
->bo
, query
->Base
.Stream
, 1);
433 if (query
->Base
.Stream
== 0)
434 ctx
->NewDriverState
|= BRW_NEW_RASTERIZER_DISCARD
;
437 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
438 write_xfb_primitives_written(brw
, query
->bo
, query
->Base
.Stream
, 1);
441 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
442 write_xfb_overflow_streams(ctx
, query
->bo
, query
->Base
.Stream
, 1, 1);
445 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
446 write_xfb_overflow_streams(ctx
, query
->bo
, 0, MAX_VERTEX_STREAMS
, 1);
449 /* calculate overflow here */
450 case GL_VERTICES_SUBMITTED_ARB
:
451 case GL_PRIMITIVES_SUBMITTED_ARB
:
452 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
453 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
454 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
455 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
456 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
457 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
458 case GL_GEOMETRY_SHADER_INVOCATIONS
:
459 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
460 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
461 emit_pipeline_stat(brw
, query
->bo
,
462 query
->Base
.Stream
, query
->Base
.Target
, 1);
466 unreachable("Unrecognized query target in brw_end_query()");
469 /* The current batch contains the commands to handle EndQuery(),
470 * but they won't actually execute until it is flushed.
472 query
->flushed
= false;
474 /* For ARB_query_buffer_object: The result is now available */
475 set_query_availability(brw
, query
, true);
479 * Flush the batch if it still references the query object BO.
482 flush_batch_if_needed(struct brw_context
*brw
, struct brw_query_object
*query
)
484 /* If the batch doesn't reference the BO, it must have been flushed
485 * (for example, due to being full). Record that it's been flushed.
487 query
->flushed
= query
->flushed
||
488 !brw_batch_references(&brw
->batch
, query
->bo
);
491 intel_batchbuffer_flush(brw
);
495 * The WaitQuery() driver hook.
497 * Wait for a query result to become available and return it. This is the
498 * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
500 static void gen6_wait_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
502 struct brw_context
*brw
= brw_context(ctx
);
503 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
505 /* If the application has requested the query result, but this batch is
506 * still contributing to it, flush it now to finish that work so the
507 * result will become available (eventually).
509 flush_batch_if_needed(brw
, query
);
511 gen6_queryobj_get_results(ctx
, query
);
515 * The CheckQuery() driver hook.
517 * Checks whether a query result is ready yet. If not, flushes.
518 * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
520 static void gen6_check_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
522 struct brw_context
*brw
= brw_context(ctx
);
523 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
525 /* If query->bo is NULL, we've already gathered the results - this is a
526 * redundant CheckQuery call. Ignore it.
528 if (query
->bo
== NULL
)
531 /* From the GL_ARB_occlusion_query spec:
533 * "Instead of allowing for an infinite loop, performing a
534 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
535 * not ready yet on the first time it is queried. This ensures that
536 * the async query will return true in finite time.
538 flush_batch_if_needed(brw
, query
);
540 if (!brw_bo_busy(query
->bo
)) {
541 gen6_queryobj_get_results(ctx
, query
);
546 gen6_query_counter(struct gl_context
*ctx
, struct gl_query_object
*q
)
548 struct brw_context
*brw
= brw_context(ctx
);
549 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
550 brw_query_counter(ctx
, q
);
551 set_query_availability(brw
, query
, true);
554 /* Initialize Gen6+-specific query object functions. */
555 void gen6_init_queryobj_functions(struct dd_function_table
*functions
)
557 functions
->BeginQuery
= gen6_begin_query
;
558 functions
->EndQuery
= gen6_end_query
;
559 functions
->CheckQuery
= gen6_check_query
;
560 functions
->WaitQuery
= gen6_wait_query
;
561 functions
->QueryCounter
= gen6_query_counter
;