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 "intel_batchbuffer.h"
40 #include "intel_buffer_objects.h"
43 set_query_availability(struct brw_context
*brw
, struct brw_query_object
*query
,
46 /* For platforms that support ARB_query_buffer_object, we write the
47 * query availability for "pipelined" queries.
49 * Most counter snapshots are written by the command streamer, by
50 * doing a CS stall and then MI_STORE_REGISTER_MEM. For these
51 * counters, the CS stall guarantees that the results will be
52 * available when subsequent CS commands run. So we don't need to
53 * do any additional tracking.
55 * Other counters (occlusion queries and timestamp) are written by
56 * PIPE_CONTROL, without a CS stall. This means that we can't be
57 * sure whether the writes have landed yet or not. Performing a
58 * PIPE_CONTROL with an immediate write will synchronize with
59 * those earlier writes, so we write 1 when the value has landed.
61 if (brw
->ctx
.Extensions
.ARB_query_buffer_object
&&
62 brw_is_query_pipelined(query
)) {
63 unsigned flags
= PIPE_CONTROL_WRITE_IMMEDIATE
;
66 /* Order available *after* the query results. */
67 flags
|= PIPE_CONTROL_FLUSH_ENABLE
;
69 /* Make it unavailable *before* any pipelined reads. */
70 flags
|= PIPE_CONTROL_CS_STALL
;
73 brw_emit_pipe_control_write(brw
, flags
,
74 query
->bo
, 2 * sizeof(uint64_t),
80 write_primitives_generated(struct brw_context
*brw
,
81 struct brw_bo
*query_bo
, int stream
, int idx
)
83 brw_emit_mi_flush(brw
);
85 if (brw
->gen
>= 7 && stream
> 0) {
86 brw_store_register_mem64(brw
, query_bo
,
87 GEN7_SO_PRIM_STORAGE_NEEDED(stream
),
88 idx
* sizeof(uint64_t));
90 brw_store_register_mem64(brw
, query_bo
, CL_INVOCATION_COUNT
,
91 idx
* sizeof(uint64_t));
96 write_xfb_primitives_written(struct brw_context
*brw
,
97 struct brw_bo
*bo
, int stream
, int idx
)
99 brw_emit_mi_flush(brw
);
102 brw_store_register_mem64(brw
, bo
, GEN7_SO_NUM_PRIMS_WRITTEN(stream
),
103 idx
* sizeof(uint64_t));
105 brw_store_register_mem64(brw
, bo
, GEN6_SO_NUM_PRIMS_WRITTEN
,
106 idx
* sizeof(uint64_t));
111 write_xfb_overflow_streams(struct gl_context
*ctx
,
112 struct brw_bo
*bo
, int stream
, int count
,
115 struct brw_context
*brw
= brw_context(ctx
);
117 brw_emit_mi_flush(brw
);
119 for (int i
= 0; i
< count
; i
++) {
120 int w_idx
= 4 * i
+ idx
;
121 int g_idx
= 4 * i
+ idx
+ 2;
124 brw_store_register_mem64(brw
, bo
,
125 GEN7_SO_NUM_PRIMS_WRITTEN(stream
+ i
),
126 g_idx
* sizeof(uint64_t));
127 brw_store_register_mem64(brw
, bo
,
128 GEN7_SO_PRIM_STORAGE_NEEDED(stream
+ i
),
129 w_idx
* sizeof(uint64_t));
131 brw_store_register_mem64(brw
, bo
,
132 GEN6_SO_NUM_PRIMS_WRITTEN
,
133 g_idx
* sizeof(uint64_t));
134 brw_store_register_mem64(brw
, bo
,
135 GEN6_SO_PRIM_STORAGE_NEEDED
,
136 w_idx
* sizeof(uint64_t));
142 check_xfb_overflow_streams(uint64_t *results
, int count
)
144 bool overflow
= false;
146 for (int i
= 0; i
< count
; i
++) {
147 uint64_t *result_i
= &results
[4 * i
];
149 if ((result_i
[3] - result_i
[2]) != (result_i
[1] - result_i
[0])) {
159 pipeline_target_to_index(int target
)
161 if (target
== GL_GEOMETRY_SHADER_INVOCATIONS
)
162 return MAX_PIPELINE_STATISTICS
- 1;
164 return target
- GL_VERTICES_SUBMITTED_ARB
;
168 emit_pipeline_stat(struct brw_context
*brw
, struct brw_bo
*bo
,
169 int stream
, int target
, int idx
)
171 /* One source of confusion is the tessellation shader statistics. The
172 * hardware has no statistics specific to the TE unit. Ideally we could have
173 * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
174 * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
175 * Unfortunately we don't have HS primitives, we only have HS invocations.
178 /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
181 static const uint32_t target_to_register
[] = {
182 IA_VERTICES_COUNT
, /* VERTICES_SUBMITTED */
183 IA_PRIMITIVES_COUNT
, /* PRIMITIVES_SUBMITTED */
184 VS_INVOCATION_COUNT
, /* VERTEX_SHADER_INVOCATIONS */
185 HS_INVOCATION_COUNT
, /* TESS_CONTROL_SHADER_PATCHES */
186 DS_INVOCATION_COUNT
, /* TESS_EVALUATION_SHADER_INVOCATIONS */
187 GS_PRIMITIVES_COUNT
, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
188 PS_INVOCATION_COUNT
, /* FRAGMENT_SHADER_INVOCATIONS */
189 CS_INVOCATION_COUNT
, /* COMPUTE_SHADER_INVOCATIONS */
190 CL_INVOCATION_COUNT
, /* CLIPPING_INPUT_PRIMITIVES */
191 CL_PRIMITIVES_COUNT
, /* CLIPPING_OUTPUT_PRIMITIVES */
192 GS_INVOCATION_COUNT
/* This one is special... */
194 STATIC_ASSERT(ARRAY_SIZE(target_to_register
) == MAX_PIPELINE_STATISTICS
);
195 uint32_t reg
= target_to_register
[pipeline_target_to_index(target
)];
196 /* Gen6 GS code counts full primitives, that is, it won't count individual
197 * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
199 if (brw
->gen
== 6 && target
== GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
)
200 reg
= CL_INVOCATION_COUNT
;
203 /* Emit a flush to make sure various parts of the pipeline are complete and
204 * we get an accurate value
206 brw_emit_mi_flush(brw
);
208 brw_store_register_mem64(brw
, bo
, reg
, idx
* sizeof(uint64_t));
213 * Wait on the query object's BO and calculate the final result.
216 gen6_queryobj_get_results(struct gl_context
*ctx
,
217 struct brw_query_object
*query
)
219 struct brw_context
*brw
= brw_context(ctx
);
221 if (query
->bo
== NULL
)
224 uint64_t *results
= brw_bo_map(brw
, query
->bo
, MAP_READ
);
225 switch (query
->Base
.Target
) {
226 case GL_TIME_ELAPSED
:
227 /* The query BO contains the starting and ending timestamps.
228 * Subtract the two and convert to nanoseconds.
230 query
->Base
.Result
= brw_raw_timestamp_delta(brw
, results
[0], results
[1]);
231 query
->Base
.Result
= brw_timebase_scale(brw
, query
->Base
.Result
);
235 /* The query BO contains a single timestamp value in results[0]. */
236 query
->Base
.Result
= brw_timebase_scale(brw
, results
[0]);
238 /* Ensure the scaled timestamp overflows according to
239 * GL_QUERY_COUNTER_BITS
241 query
->Base
.Result
&= (1ull << ctx
->Const
.QueryCounterBits
.Timestamp
) - 1;
244 case GL_SAMPLES_PASSED_ARB
:
245 /* We need to use += rather than = here since some BLT-based operations
246 * may have added additional samples to our occlusion query value.
248 query
->Base
.Result
+= results
[1] - results
[0];
251 case GL_ANY_SAMPLES_PASSED
:
252 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
253 if (results
[0] != results
[1])
254 query
->Base
.Result
= true;
257 case GL_PRIMITIVES_GENERATED
:
258 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
259 case GL_VERTICES_SUBMITTED_ARB
:
260 case GL_PRIMITIVES_SUBMITTED_ARB
:
261 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
262 case GL_GEOMETRY_SHADER_INVOCATIONS
:
263 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
264 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
265 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
266 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
267 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
268 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
269 query
->Base
.Result
= results
[1] - results
[0];
272 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
273 query
->Base
.Result
= check_xfb_overflow_streams(results
, 1);
276 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
277 query
->Base
.Result
= check_xfb_overflow_streams(results
, MAX_VERTEX_STREAMS
);
280 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
281 query
->Base
.Result
= (results
[1] - results
[0]);
282 /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
283 * "Invocation counter is 4 times actual. WA: SW to divide HW reported
284 * PS Invocations value by 4."
286 * Prior to Haswell, invocation count was counted by the WM, and it
287 * buggily counted invocations in units of subspans (2x2 unit). To get the
288 * correct value, the CS multiplied this by 4. With HSW the logic moved,
289 * and correctly emitted the number of pixel shader invocations, but,
290 * whomever forgot to undo the multiply by 4.
292 if (brw
->gen
== 8 || brw
->is_haswell
)
293 query
->Base
.Result
/= 4;
297 unreachable("Unrecognized query target in brw_queryobj_get_results()");
299 brw_bo_unmap(query
->bo
);
301 /* Now that we've processed the data stored in the query's buffer object,
304 brw_bo_unreference(query
->bo
);
307 query
->Base
.Ready
= true;
311 * Driver hook for glBeginQuery().
313 * Initializes driver structures and emits any GPU commands required to begin
314 * recording data for the query.
317 gen6_begin_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
319 struct brw_context
*brw
= brw_context(ctx
);
320 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
322 /* Since we're starting a new query, we need to throw away old results. */
323 brw_bo_unreference(query
->bo
);
324 query
->bo
= brw_bo_alloc(brw
->bufmgr
, "query results", 4096, 4096);
326 /* For ARB_query_buffer_object: The result is not available */
327 set_query_availability(brw
, query
, false);
329 switch (query
->Base
.Target
) {
330 case GL_TIME_ELAPSED
:
331 /* For timestamp queries, we record the starting time right away so that
332 * we measure the full time between BeginQuery and EndQuery. There's
333 * some debate about whether this is the right thing to do. Our decision
334 * is based on the following text from the ARB_timer_query extension:
336 * "(5) Should the extension measure total time elapsed between the full
337 * completion of the BeginQuery and EndQuery commands, or just time
338 * spent in the graphics library?
340 * RESOLVED: This extension will measure the total time elapsed
341 * between the full completion of these commands. Future extensions
342 * may implement a query to determine time elapsed at different stages
343 * of the graphics pipeline."
345 * We write a starting timestamp now (at index 0). At EndQuery() time,
346 * we'll write a second timestamp (at index 1), and subtract the two to
347 * obtain the time elapsed. Notably, this includes time elapsed while
348 * the system was doing other work, such as running other applications.
350 brw_write_timestamp(brw
, query
->bo
, 0);
353 case GL_ANY_SAMPLES_PASSED
:
354 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
355 case GL_SAMPLES_PASSED_ARB
:
356 brw_write_depth_count(brw
, query
->bo
, 0);
359 case GL_PRIMITIVES_GENERATED
:
360 write_primitives_generated(brw
, query
->bo
, query
->Base
.Stream
, 0);
361 if (query
->Base
.Stream
== 0)
362 ctx
->NewDriverState
|= BRW_NEW_RASTERIZER_DISCARD
;
365 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
366 write_xfb_primitives_written(brw
, query
->bo
, query
->Base
.Stream
, 0);
369 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
370 write_xfb_overflow_streams(ctx
, query
->bo
, query
->Base
.Stream
, 1, 0);
373 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
374 write_xfb_overflow_streams(ctx
, query
->bo
, 0, MAX_VERTEX_STREAMS
, 0);
377 case GL_VERTICES_SUBMITTED_ARB
:
378 case GL_PRIMITIVES_SUBMITTED_ARB
:
379 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
380 case GL_GEOMETRY_SHADER_INVOCATIONS
:
381 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
382 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
383 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
384 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
385 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
386 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
387 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
388 emit_pipeline_stat(brw
, query
->bo
, query
->Base
.Stream
, query
->Base
.Target
, 0);
392 unreachable("Unrecognized query target in brw_begin_query()");
397 * Driver hook for glEndQuery().
399 * Emits GPU commands to record a final query value, ending any data capturing.
400 * However, the final result isn't necessarily available until the GPU processes
401 * those commands. brw_queryobj_get_results() processes the captured data to
402 * produce the final result.
405 gen6_end_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
407 struct brw_context
*brw
= brw_context(ctx
);
408 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
410 switch (query
->Base
.Target
) {
411 case GL_TIME_ELAPSED
:
412 brw_write_timestamp(brw
, query
->bo
, 1);
415 case GL_ANY_SAMPLES_PASSED
:
416 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
417 case GL_SAMPLES_PASSED_ARB
:
418 brw_write_depth_count(brw
, query
->bo
, 1);
421 case GL_PRIMITIVES_GENERATED
:
422 write_primitives_generated(brw
, query
->bo
, query
->Base
.Stream
, 1);
423 if (query
->Base
.Stream
== 0)
424 ctx
->NewDriverState
|= BRW_NEW_RASTERIZER_DISCARD
;
427 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
428 write_xfb_primitives_written(brw
, query
->bo
, query
->Base
.Stream
, 1);
431 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB
:
432 write_xfb_overflow_streams(ctx
, query
->bo
, query
->Base
.Stream
, 1, 1);
435 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB
:
436 write_xfb_overflow_streams(ctx
, query
->bo
, 0, MAX_VERTEX_STREAMS
, 1);
439 /* calculate overflow here */
440 case GL_VERTICES_SUBMITTED_ARB
:
441 case GL_PRIMITIVES_SUBMITTED_ARB
:
442 case GL_VERTEX_SHADER_INVOCATIONS_ARB
:
443 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB
:
444 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB
:
445 case GL_COMPUTE_SHADER_INVOCATIONS_ARB
:
446 case GL_CLIPPING_INPUT_PRIMITIVES_ARB
:
447 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB
:
448 case GL_GEOMETRY_SHADER_INVOCATIONS
:
449 case GL_TESS_CONTROL_SHADER_PATCHES_ARB
:
450 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB
:
451 emit_pipeline_stat(brw
, query
->bo
,
452 query
->Base
.Stream
, query
->Base
.Target
, 1);
456 unreachable("Unrecognized query target in brw_end_query()");
459 /* The current batch contains the commands to handle EndQuery(),
460 * but they won't actually execute until it is flushed.
462 query
->flushed
= false;
464 /* For ARB_query_buffer_object: The result is now available */
465 set_query_availability(brw
, query
, true);
469 * Flush the batch if it still references the query object BO.
472 flush_batch_if_needed(struct brw_context
*brw
, struct brw_query_object
*query
)
474 /* If the batch doesn't reference the BO, it must have been flushed
475 * (for example, due to being full). Record that it's been flushed.
477 query
->flushed
= query
->flushed
||
478 !brw_batch_references(&brw
->batch
, query
->bo
);
481 intel_batchbuffer_flush(brw
);
485 * The WaitQuery() driver hook.
487 * Wait for a query result to become available and return it. This is the
488 * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
490 static void gen6_wait_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
492 struct brw_context
*brw
= brw_context(ctx
);
493 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
495 /* If the application has requested the query result, but this batch is
496 * still contributing to it, flush it now to finish that work so the
497 * result will become available (eventually).
499 flush_batch_if_needed(brw
, query
);
501 gen6_queryobj_get_results(ctx
, query
);
505 * The CheckQuery() driver hook.
507 * Checks whether a query result is ready yet. If not, flushes.
508 * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
510 static void gen6_check_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
512 struct brw_context
*brw
= brw_context(ctx
);
513 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
515 /* If query->bo is NULL, we've already gathered the results - this is a
516 * redundant CheckQuery call. Ignore it.
518 if (query
->bo
== NULL
)
521 /* From the GL_ARB_occlusion_query spec:
523 * "Instead of allowing for an infinite loop, performing a
524 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
525 * not ready yet on the first time it is queried. This ensures that
526 * the async query will return true in finite time.
528 flush_batch_if_needed(brw
, query
);
530 if (!brw_bo_busy(query
->bo
)) {
531 gen6_queryobj_get_results(ctx
, query
);
536 gen6_query_counter(struct gl_context
*ctx
, struct gl_query_object
*q
)
538 struct brw_context
*brw
= brw_context(ctx
);
539 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
540 brw_query_counter(ctx
, q
);
541 set_query_availability(brw
, query
, true);
544 /* Initialize Gen6+-specific query object functions. */
545 void gen6_init_queryobj_functions(struct dd_function_table
*functions
)
547 functions
->BeginQuery
= gen6_begin_query
;
548 functions
->EndQuery
= gen6_end_query
;
549 functions
->CheckQuery
= gen6_check_query
;
550 functions
->WaitQuery
= gen6_wait_query
;
551 functions
->QueryCounter
= gen6_query_counter
;