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 "brw_context.h"
35 #include "brw_defines.h"
36 #include "brw_state.h"
37 #include "perf/gen_perf_regs.h"
38 #include "intel_batchbuffer.h"
39 #include "intel_buffer_objects.h"
42 set_query_availability(struct brw_context
*brw
, struct brw_query_object
*query
,
45 /* For platforms that support ARB_query_buffer_object, we write the
46 * query availability for "pipelined" queries.
48 * Most counter snapshots are written by the command streamer, by
49 * doing a CS stall and then MI_STORE_REGISTER_MEM. For these
50 * counters, the CS stall guarantees that the results will be
51 * available when subsequent CS commands run. So we don't need to
52 * do any additional tracking.
54 * Other counters (occlusion queries and timestamp) are written by
55 * PIPE_CONTROL, without a CS stall. This means that we can't be
56 * sure whether the writes have landed yet or not. Performing a
57 * PIPE_CONTROL with an immediate write will synchronize with
58 * those earlier writes, so we write 1 when the value has landed.
60 if (brw
->ctx
.Extensions
.ARB_query_buffer_object
&&
61 brw_is_query_pipelined(query
)) {
62 unsigned flags
= PIPE_CONTROL_WRITE_IMMEDIATE
;
65 /* Order available *after* the query results. */
66 flags
|= PIPE_CONTROL_FLUSH_ENABLE
;
68 /* Make it unavailable *before* any pipelined reads. */
69 flags
|= PIPE_CONTROL_CS_STALL
;
72 brw_emit_pipe_control_write(brw
, flags
,
73 query
->bo
, 2 * sizeof(uint64_t),
79 write_primitives_generated(struct brw_context
*brw
,
80 struct brw_bo
*query_bo
, int stream
, int idx
)
82 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
84 brw_emit_mi_flush(brw
);
86 if (devinfo
->gen
>= 7 && stream
> 0) {
87 brw_store_register_mem64(brw
, query_bo
,
88 GEN7_SO_PRIM_STORAGE_NEEDED(stream
),
89 idx
* sizeof(uint64_t));
91 brw_store_register_mem64(brw
, query_bo
, CL_INVOCATION_COUNT
,
92 idx
* sizeof(uint64_t));
97 write_xfb_primitives_written(struct brw_context
*brw
,
98 struct brw_bo
*bo
, int stream
, int idx
)
100 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
102 brw_emit_mi_flush(brw
);
104 if (devinfo
->gen
>= 7) {
105 brw_store_register_mem64(brw
, bo
, GEN7_SO_NUM_PRIMS_WRITTEN(stream
),
106 idx
* sizeof(uint64_t));
108 brw_store_register_mem64(brw
, bo
, GEN6_SO_NUM_PRIMS_WRITTEN
,
109 idx
* sizeof(uint64_t));
114 write_xfb_overflow_streams(struct gl_context
*ctx
,
115 struct brw_bo
*bo
, int stream
, int count
,
118 struct brw_context
*brw
= brw_context(ctx
);
119 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
121 brw_emit_mi_flush(brw
);
123 for (int i
= 0; i
< count
; i
++) {
124 int w_idx
= 4 * i
+ idx
;
125 int g_idx
= 4 * i
+ idx
+ 2;
127 if (devinfo
->gen
>= 7) {
128 brw_store_register_mem64(brw
, bo
,
129 GEN7_SO_NUM_PRIMS_WRITTEN(stream
+ i
),
130 g_idx
* sizeof(uint64_t));
131 brw_store_register_mem64(brw
, bo
,
132 GEN7_SO_PRIM_STORAGE_NEEDED(stream
+ i
),
133 w_idx
* sizeof(uint64_t));
135 brw_store_register_mem64(brw
, bo
,
136 GEN6_SO_NUM_PRIMS_WRITTEN
,
137 g_idx
* sizeof(uint64_t));
138 brw_store_register_mem64(brw
, bo
,
139 GEN6_SO_PRIM_STORAGE_NEEDED
,
140 w_idx
* sizeof(uint64_t));
146 check_xfb_overflow_streams(uint64_t *results
, int count
)
148 bool overflow
= false;
150 for (int i
= 0; i
< count
; i
++) {
151 uint64_t *result_i
= &results
[4 * i
];
153 if ((result_i
[3] - result_i
[2]) != (result_i
[1] - result_i
[0])) {
163 pipeline_target_to_index(int target
)
165 if (target
== GL_GEOMETRY_SHADER_INVOCATIONS
)
166 return MAX_PIPELINE_STATISTICS
- 1;
168 return target
- GL_VERTICES_SUBMITTED_ARB
;
172 emit_pipeline_stat(struct brw_context
*brw
, struct brw_bo
*bo
,
173 int stream
, int target
, int idx
)
175 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
177 /* One source of confusion is the tessellation shader statistics. The
178 * hardware has no statistics specific to the TE unit. Ideally we could have
179 * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
180 * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
181 * Unfortunately we don't have HS primitives, we only have HS invocations.
184 /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
187 static const uint32_t target_to_register
[] = {
188 IA_VERTICES_COUNT
, /* VERTICES_SUBMITTED */
189 IA_PRIMITIVES_COUNT
, /* PRIMITIVES_SUBMITTED */
190 VS_INVOCATION_COUNT
, /* VERTEX_SHADER_INVOCATIONS */
191 HS_INVOCATION_COUNT
, /* TESS_CONTROL_SHADER_PATCHES */
192 DS_INVOCATION_COUNT
, /* TESS_EVALUATION_SHADER_INVOCATIONS */
193 GS_PRIMITIVES_COUNT
, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
194 PS_INVOCATION_COUNT
, /* FRAGMENT_SHADER_INVOCATIONS */
195 CS_INVOCATION_COUNT
, /* COMPUTE_SHADER_INVOCATIONS */
196 CL_INVOCATION_COUNT
, /* CLIPPING_INPUT_PRIMITIVES */
197 CL_PRIMITIVES_COUNT
, /* CLIPPING_OUTPUT_PRIMITIVES */
198 GS_INVOCATION_COUNT
/* This one is special... */
200 STATIC_ASSERT(ARRAY_SIZE(target_to_register
) == MAX_PIPELINE_STATISTICS
);
201 uint32_t reg
= target_to_register
[pipeline_target_to_index(target
)];
202 /* Gen6 GS code counts full primitives, that is, it won't count individual
203 * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
205 if (devinfo
->gen
== 6 && target
== GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
)
206 reg
= CL_INVOCATION_COUNT
;
209 /* Emit a flush to make sure various parts of the pipeline are complete and
210 * we get an accurate value
212 brw_emit_mi_flush(brw
);
214 brw_store_register_mem64(brw
, bo
, reg
, idx
* sizeof(uint64_t));
219 * Wait on the query object's BO and calculate the final result.
222 gen6_queryobj_get_results(struct gl_context
*ctx
,
223 struct brw_query_object
*query
)
225 struct brw_context
*brw
= brw_context(ctx
);
226 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
228 if (query
->bo
== NULL
)
231 uint64_t *results
= brw_bo_map(brw
, query
->bo
, MAP_READ
);
232 switch (query
->Base
.Target
) {
233 case GL_TIME_ELAPSED
:
234 /* The query BO contains the starting and ending timestamps.
235 * Subtract the two and convert to nanoseconds.
237 query
->Base
.Result
= brw_raw_timestamp_delta(brw
, results
[0], results
[1]);
238 query
->Base
.Result
= gen_device_info_timebase_scale(devinfo
, query
->Base
.Result
);
242 /* The query BO contains a single timestamp value in results[0]. */
243 query
->Base
.Result
= gen_device_info_timebase_scale(devinfo
, results
[0]);
245 /* Ensure the scaled timestamp overflows according to
246 * GL_QUERY_COUNTER_BITS
248 query
->Base
.Result
&= (1ull << ctx
->Const
.QueryCounterBits
.Timestamp
) - 1;
251 case GL_SAMPLES_PASSED_ARB
:
252 /* We need to use += rather than = here since some BLT-based operations
253 * may have added additional samples to our occlusion query value.
255 query
->Base
.Result
+= results
[1] - results
[0];
258 case GL_ANY_SAMPLES_PASSED
:
259 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
260 if (results
[0] != results
[1])
261 query
->Base
.Result
= true;
264 case GL_PRIMITIVES_GENERATED
:
265 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
266 case GL_VERTICES_SUBMITTED_ARB
:
267 case GL_PRIMITIVES_SUBMITTED_ARB
:
268 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
269 case GL_GEOMETRY_SHADER_INVOCATIONS
:
270 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
271 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
272 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
273 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
274 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
275 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
276 query
->Base
.Result
= results
[1] - results
[0];
279 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
280 query
->Base
.Result
= check_xfb_overflow_streams(results
, 1);
283 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
284 query
->Base
.Result
= check_xfb_overflow_streams(results
, MAX_VERTEX_STREAMS
);
287 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
288 query
->Base
.Result
= (results
[1] - results
[0]);
289 /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
290 * "Invocation counter is 4 times actual. WA: SW to divide HW reported
291 * PS Invocations value by 4."
293 * Prior to Haswell, invocation count was counted by the WM, and it
294 * buggily counted invocations in units of subspans (2x2 unit). To get the
295 * correct value, the CS multiplied this by 4. With HSW the logic moved,
296 * and correctly emitted the number of pixel shader invocations, but,
297 * whomever forgot to undo the multiply by 4.
299 if (devinfo
->gen
== 8 || devinfo
->is_haswell
)
300 query
->Base
.Result
/= 4;
304 unreachable("Unrecognized query target in brw_queryobj_get_results()");
306 brw_bo_unmap(query
->bo
);
308 /* Now that we've processed the data stored in the query's buffer object,
311 brw_bo_unreference(query
->bo
);
314 query
->Base
.Ready
= true;
318 * Driver hook for glBeginQuery().
320 * Initializes driver structures and emits any GPU commands required to begin
321 * recording data for the query.
324 gen6_begin_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
326 struct brw_context
*brw
= brw_context(ctx
);
327 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
329 /* Since we're starting a new query, we need to throw away old results. */
330 brw_bo_unreference(query
->bo
);
332 brw_bo_alloc(brw
->bufmgr
, "query results", 4096, BRW_MEMZONE_OTHER
);
334 /* For ARB_query_buffer_object: The result is not available */
335 set_query_availability(brw
, query
, false);
337 switch (query
->Base
.Target
) {
338 case GL_TIME_ELAPSED
:
339 /* For timestamp queries, we record the starting time right away so that
340 * we measure the full time between BeginQuery and EndQuery. There's
341 * some debate about whether this is the right thing to do. Our decision
342 * is based on the following text from the ARB_timer_query extension:
344 * "(5) Should the extension measure total time elapsed between the full
345 * completion of the BeginQuery and EndQuery commands, or just time
346 * spent in the graphics library?
348 * RESOLVED: This extension will measure the total time elapsed
349 * between the full completion of these commands. Future extensions
350 * may implement a query to determine time elapsed at different stages
351 * of the graphics pipeline."
353 * We write a starting timestamp now (at index 0). At EndQuery() time,
354 * we'll write a second timestamp (at index 1), and subtract the two to
355 * obtain the time elapsed. Notably, this includes time elapsed while
356 * the system was doing other work, such as running other applications.
358 brw_write_timestamp(brw
, query
->bo
, 0);
361 case GL_ANY_SAMPLES_PASSED
:
362 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
363 case GL_SAMPLES_PASSED_ARB
:
364 brw_write_depth_count(brw
, query
->bo
, 0);
367 case GL_PRIMITIVES_GENERATED
:
368 write_primitives_generated(brw
, query
->bo
, query
->Base
.Stream
, 0);
369 if (query
->Base
.Stream
== 0)
370 ctx
->NewDriverState
|= BRW_NEW_RASTERIZER_DISCARD
;
373 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
374 write_xfb_primitives_written(brw
, query
->bo
, query
->Base
.Stream
, 0);
377 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
378 write_xfb_overflow_streams(ctx
, query
->bo
, query
->Base
.Stream
, 1, 0);
381 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
382 write_xfb_overflow_streams(ctx
, query
->bo
, 0, MAX_VERTEX_STREAMS
, 0);
385 case GL_VERTICES_SUBMITTED_ARB
:
386 case GL_PRIMITIVES_SUBMITTED_ARB
:
387 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
388 case GL_GEOMETRY_SHADER_INVOCATIONS
:
389 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
390 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
391 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
392 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
393 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
394 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
395 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
396 emit_pipeline_stat(brw
, query
->bo
, query
->Base
.Stream
, query
->Base
.Target
, 0);
400 unreachable("Unrecognized query target in brw_begin_query()");
405 * Driver hook for glEndQuery().
407 * Emits GPU commands to record a final query value, ending any data capturing.
408 * However, the final result isn't necessarily available until the GPU processes
409 * those commands. brw_queryobj_get_results() processes the captured data to
410 * produce the final result.
413 gen6_end_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
415 struct brw_context
*brw
= brw_context(ctx
);
416 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
418 switch (query
->Base
.Target
) {
419 case GL_TIME_ELAPSED
:
420 brw_write_timestamp(brw
, query
->bo
, 1);
423 case GL_ANY_SAMPLES_PASSED
:
424 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
425 case GL_SAMPLES_PASSED_ARB
:
426 brw_write_depth_count(brw
, query
->bo
, 1);
429 case GL_PRIMITIVES_GENERATED
:
430 write_primitives_generated(brw
, query
->bo
, query
->Base
.Stream
, 1);
431 if (query
->Base
.Stream
== 0)
432 ctx
->NewDriverState
|= BRW_NEW_RASTERIZER_DISCARD
;
435 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
436 write_xfb_primitives_written(brw
, query
->bo
, query
->Base
.Stream
, 1);
439 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
440 write_xfb_overflow_streams(ctx
, query
->bo
, query
->Base
.Stream
, 1, 1);
443 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
444 write_xfb_overflow_streams(ctx
, query
->bo
, 0, MAX_VERTEX_STREAMS
, 1);
447 /* calculate overflow here */
448 case GL_VERTICES_SUBMITTED_ARB
:
449 case GL_PRIMITIVES_SUBMITTED_ARB
:
450 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
451 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
452 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
453 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
454 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
455 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
456 case GL_GEOMETRY_SHADER_INVOCATIONS
:
457 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
458 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
459 emit_pipeline_stat(brw
, query
->bo
,
460 query
->Base
.Stream
, query
->Base
.Target
, 1);
464 unreachable("Unrecognized query target in brw_end_query()");
467 /* The current batch contains the commands to handle EndQuery(),
468 * but they won't actually execute until it is flushed.
470 query
->flushed
= false;
472 /* For ARB_query_buffer_object: The result is now available */
473 set_query_availability(brw
, query
, true);
477 * Flush the batch if it still references the query object BO.
480 flush_batch_if_needed(struct brw_context
*brw
, struct brw_query_object
*query
)
482 /* If the batch doesn't reference the BO, it must have been flushed
483 * (for example, due to being full). Record that it's been flushed.
485 query
->flushed
= query
->flushed
||
486 !brw_batch_references(&brw
->batch
, query
->bo
);
489 intel_batchbuffer_flush(brw
);
493 * The WaitQuery() driver hook.
495 * Wait for a query result to become available and return it. This is the
496 * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
498 static void gen6_wait_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
500 struct brw_context
*brw
= brw_context(ctx
);
501 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
503 /* If the application has requested the query result, but this batch is
504 * still contributing to it, flush it now to finish that work so the
505 * result will become available (eventually).
507 flush_batch_if_needed(brw
, query
);
509 gen6_queryobj_get_results(ctx
, query
);
513 * The CheckQuery() driver hook.
515 * Checks whether a query result is ready yet. If not, flushes.
516 * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
518 static void gen6_check_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
520 struct brw_context
*brw
= brw_context(ctx
);
521 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
523 /* If query->bo is NULL, we've already gathered the results - this is a
524 * redundant CheckQuery call. Ignore it.
526 if (query
->bo
== NULL
)
529 /* From the GL_ARB_occlusion_query spec:
531 * "Instead of allowing for an infinite loop, performing a
532 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
533 * not ready yet on the first time it is queried. This ensures that
534 * the async query will return true in finite time.
536 flush_batch_if_needed(brw
, query
);
538 if (!brw_bo_busy(query
->bo
)) {
539 gen6_queryobj_get_results(ctx
, query
);
544 gen6_query_counter(struct gl_context
*ctx
, struct gl_query_object
*q
)
546 struct brw_context
*brw
= brw_context(ctx
);
547 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
548 brw_query_counter(ctx
, q
);
549 set_query_availability(brw
, query
, true);
552 /* Initialize Gen6+-specific query object functions. */
553 void gen6_init_queryobj_functions(struct dd_function_table
*functions
)
555 functions
->BeginQuery
= gen6_begin_query
;
556 functions
->EndQuery
= gen6_end_query
;
557 functions
->CheckQuery
= gen6_check_query
;
558 functions
->WaitQuery
= gen6_wait_query
;
559 functions
->QueryCounter
= gen6_query_counter
;