src/mesa/drivers/dri/i965/brw_queryobj.c

   1 /*
   2  * Copyright © 2008 Intel Corporation
   3  *
   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:
  10  *
  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
  13  * Software.
  14  *
  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
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eric Anholt <eric@anholt.net>
  25  *
  26  */
  27
  28 /** @file brw_queryobj.c
  29  *
  30  * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query,
  31  * GL_EXT_transform_feedback, and friends).
  32  *
  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'
  36  * needs.
  37  */
  38 #include "main/imports.h"
  39
  40 #include "brw_context.h"
  41 #include "brw_defines.h"
  42 #include "brw_state.h"
  43 #include "intel_batchbuffer.h"
  44
  45 uint64_t
  46 brw_timebase_scale(struct brw_context *brw, uint64_t gpu_timestamp)
  47 {
  48    const struct gen_device_info *devinfo = &brw->screen->devinfo;
  49
  50    return (1000000000ull * gpu_timestamp) / devinfo->timestamp_frequency;
  51 }
  52
  53 /* As best we know currently, the Gen HW timestamps are 36bits across
  54  * all platforms, which we need to account for when calculating a
  55  * delta to measure elapsed time.
  56  *
  57  * The timestamps read via glGetTimestamp() / brw_get_timestamp() sometimes
  58  * only have 32bits due to a kernel bug and so in that case we make sure to
  59  * treat all raw timestamps as 32bits so they overflow consistently and remain
  60  * comparable. (Note: the timestamps being passed here are not from the kernel
  61  * so we don't need to be taking the upper 32bits in this buggy kernel case we
  62  * are just clipping to 32bits here for consistency.)
  63  */
  64 uint64_t
  65 brw_raw_timestamp_delta(struct brw_context *brw, uint64_t time0, uint64_t time1)
  66 {
  67    if (brw->screen->hw_has_timestamp == 2) {
  68       /* Kernel clips timestamps to 32bits in this case, so we also clip
  69        * PIPE_CONTROL timestamps for consistency.
  70        */
  71       return (uint32_t)time1 - (uint32_t)time0;
  72    } else {
  73       if (time0 > time1) {
  74          return (1ULL << 36) + time1 - time0;
  75       } else {
  76          return time1 - time0;
  77       }
  78    }
  79 }
  80
  81 /**
  82  * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
  83  */
  84 void
  85 brw_write_timestamp(struct brw_context *brw, struct brw_bo *query_bo, int idx)
  86 {
  87    if (brw->gen == 6) {
  88       /* Emit Sandybridge workaround flush: */
  89       brw_emit_pipe_control_flush(brw,
  90                                   PIPE_CONTROL_CS_STALL |
  91                                   PIPE_CONTROL_STALL_AT_SCOREBOARD);
  92    }
  93
  94    uint32_t flags = PIPE_CONTROL_WRITE_TIMESTAMP;
  95
  96    if (brw->gen == 9 && brw->gt == 4)
  97       flags |= PIPE_CONTROL_CS_STALL;
  98
  99    brw_emit_pipe_control_write(brw, flags,
 100                                query_bo, idx * sizeof(uint64_t), 0);
 101 }
 102
 103 /**
 104  * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
 105  */
 106 void
 107 brw_write_depth_count(struct brw_context *brw, struct brw_bo *query_bo, int idx)
 108 {
 109    uint32_t flags = PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_DEPTH_STALL;
 110
 111    if (brw->gen == 9 && brw->gt == 4)
 112       flags |= PIPE_CONTROL_CS_STALL;
 113
 114    if (brw->gen >= 10) {
 115       /* "Driver must program PIPE_CONTROL with only Depth Stall Enable bit set
 116        * prior to programming a PIPE_CONTROL with Write PS Depth Count Post sync
 117        * operation."
 118        */
 119       brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_STALL);
 120    }
 121
 122    brw_emit_pipe_control_write(brw, flags,
 123                                query_bo, idx * sizeof(uint64_t), 0);
 124 }
 125
 126 /**
 127  * Wait on the query object's BO and calculate the final result.
 128  */
 129 static void
 130 brw_queryobj_get_results(struct gl_context *ctx,
 131                          struct brw_query_object *query)
 132 {
 133    struct brw_context *brw = brw_context(ctx);
 134
 135    int i;
 136    uint64_t *results;
 137
 138    assert(brw->gen < 6);
 139
 140    if (query->bo == NULL)
 141       return;
 142
 143    /* If the application has requested the query result, but this batch is
 144     * still contributing to it, flush it now so the results will be present
 145     * when mapped.
 146     */
 147    if (brw_batch_references(&brw->batch, query->bo))
 148       intel_batchbuffer_flush(brw);
 149
 150    if (unlikely(brw->perf_debug)) {
 151       if (brw_bo_busy(query->bo)) {
 152          perf_debug("Stalling on the GPU waiting for a query object.\n");
 153       }
 154    }
 155
 156    results = brw_bo_map(brw, query->bo, MAP_READ);
 157    switch (query->Base.Target) {
 158    case GL_TIME_ELAPSED_EXT:
 159       /* The query BO contains the starting and ending timestamps.
 160        * Subtract the two and convert to nanoseconds.
 161        */
 162       query->Base.Result = brw_raw_timestamp_delta(brw, results[0], results[1]);
 163       query->Base.Result = brw_timebase_scale(brw, query->Base.Result);
 164       break;
 165
 166    case GL_TIMESTAMP:
 167       /* The query BO contains a single timestamp value in results[0]. */
 168       query->Base.Result = brw_timebase_scale(brw, results[0]);
 169
 170       /* Ensure the scaled timestamp overflows according to
 171        * GL_QUERY_COUNTER_BITS
 172        */
 173       query->Base.Result &= (1ull << ctx->Const.QueryCounterBits.Timestamp) - 1;
 174       break;
 175
 176    case GL_SAMPLES_PASSED_ARB:
 177       /* Loop over pairs of values from the BO, which are the PS_DEPTH_COUNT
 178        * value at the start and end of the batchbuffer.  Subtract them to
 179        * get the number of fragments which passed the depth test in each
 180        * individual batch, and add those differences up to get the number
 181        * of fragments for the entire query.
 182        *
 183        * Note that query->Base.Result may already be non-zero.  We may have
 184        * run out of space in the query's BO and allocated a new one.  If so,
 185        * this function was already called to accumulate the results so far.
 186        */
 187       for (i = 0; i < query->last_index; i++) {
 188          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 189       }
 190       break;
 191
 192    case GL_ANY_SAMPLES_PASSED:
 193    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 194       /* If the starting and ending PS_DEPTH_COUNT from any of the batches
 195        * differ, then some fragments passed the depth test.
 196        */
 197       for (i = 0; i < query->last_index; i++) {
 198          if (results[i * 2 + 1] != results[i * 2]) {
 199             query->Base.Result = GL_TRUE;
 200             break;
 201          }
 202       }
 203       break;
 204
 205    default:
 206       unreachable("Unrecognized query target in brw_queryobj_get_results()");
 207    }
 208    brw_bo_unmap(query->bo);
 209
 210    /* Now that we've processed the data stored in the query's buffer object,
 211     * we can release it.
 212     */
 213    brw_bo_unreference(query->bo);
 214    query->bo = NULL;
 215 }
 216
 217 /**
 218  * The NewQueryObject() driver hook.
 219  *
 220  * Allocates and initializes a new query object.
 221  */
 222 static struct gl_query_object *
 223 brw_new_query_object(struct gl_context *ctx, GLuint id)
 224 {
 225    struct brw_query_object *query;
 226
 227    query = calloc(1, sizeof(struct brw_query_object));
 228
 229    query->Base.Id = id;
 230    query->Base.Result = 0;
 231    query->Base.Active = false;
 232    query->Base.Ready = true;
 233
 234    return &query->Base;
 235 }
 236
 237 /**
 238  * The DeleteQuery() driver hook.
 239  */
 240 static void
 241 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 242 {
 243    struct brw_query_object *query = (struct brw_query_object *)q;
 244
 245    brw_bo_unreference(query->bo);
 246    free(query);
 247 }
 248
 249 /**
 250  * Gen4-5 driver hook for glBeginQuery().
 251  *
 252  * Initializes driver structures and emits any GPU commands required to begin
 253  * recording data for the query.
 254  */
 255 static void
 256 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 257 {
 258    struct brw_context *brw = brw_context(ctx);
 259    struct brw_query_object *query = (struct brw_query_object *)q;
 260
 261    assert(brw->gen < 6);
 262
 263    switch (query->Base.Target) {
 264    case GL_TIME_ELAPSED_EXT:
 265       /* For timestamp queries, we record the starting time right away so that
 266        * we measure the full time between BeginQuery and EndQuery.  There's
 267        * some debate about whether this is the right thing to do.  Our decision
 268        * is based on the following text from the ARB_timer_query extension:
 269        *
 270        * "(5) Should the extension measure total time elapsed between the full
 271        *      completion of the BeginQuery and EndQuery commands, or just time
 272        *      spent in the graphics library?
 273        *
 274        *  RESOLVED:  This extension will measure the total time elapsed
 275        *  between the full completion of these commands.  Future extensions
 276        *  may implement a query to determine time elapsed at different stages
 277        *  of the graphics pipeline."
 278        *
 279        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 280        * we'll write a second timestamp (at index 1), and subtract the two to
 281        * obtain the time elapsed.  Notably, this includes time elapsed while
 282        * the system was doing other work, such as running other applications.
 283        */
 284       brw_bo_unreference(query->bo);
 285       query->bo = brw_bo_alloc(brw->bufmgr, "timer query", 4096, 4096);
 286       brw_write_timestamp(brw, query->bo, 0);
 287       break;
 288
 289    case GL_ANY_SAMPLES_PASSED:
 290    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 291    case GL_SAMPLES_PASSED_ARB:
 292       /* For occlusion queries, we delay taking an initial sample until the
 293        * first drawing occurs in this batch.  See the reasoning in the comments
 294        * for brw_emit_query_begin() below.
 295        *
 296        * Since we're starting a new query, we need to be sure to throw away
 297        * any previous occlusion query results.
 298        */
 299       brw_bo_unreference(query->bo);
 300       query->bo = NULL;
 301       query->last_index = -1;
 302
 303       brw->query.obj = query;
 304
 305       /* Depth statistics on Gen4 require strange workarounds, so we try to
 306        * avoid them when necessary.  They're required for occlusion queries,
 307        * so turn them on now.
 308        */
 309       brw->stats_wm++;
 310       brw->ctx.NewDriverState |= BRW_NEW_STATS_WM;
 311       break;
 312
 313    default:
 314       unreachable("Unrecognized query target in brw_begin_query()");
 315    }
 316 }
 317
 318 /**
 319  * Gen4-5 driver hook for glEndQuery().
 320  *
 321  * Emits GPU commands to record a final query value, ending any data capturing.
 322  * However, the final result isn't necessarily available until the GPU processes
 323  * those commands.  brw_queryobj_get_results() processes the captured data to
 324  * produce the final result.
 325  */
 326 static void
 327 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 328 {
 329    struct brw_context *brw = brw_context(ctx);
 330    struct brw_query_object *query = (struct brw_query_object *)q;
 331
 332    assert(brw->gen < 6);
 333
 334    switch (query->Base.Target) {
 335    case GL_TIME_ELAPSED_EXT:
 336       /* Write the final timestamp. */
 337       brw_write_timestamp(brw, query->bo, 1);
 338       break;
 339
 340    case GL_ANY_SAMPLES_PASSED:
 341    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 342    case GL_SAMPLES_PASSED_ARB:
 343
 344       /* No query->bo means that EndQuery was called after BeginQuery with no
 345        * intervening drawing. Rather than doing nothing at all here in this
 346        * case, we emit the query_begin and query_end state to the
 347        * hardware. This is to guarantee that waiting on the result of this
 348        * empty state will cause all previous queries to complete at all, as
 349        * required by the specification:
 350        *
 351        *        It must always be true that if any query object
 352        *        returns a result available of TRUE, all queries of the
 353        *        same type issued prior to that query must also return
 354        *        TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
 355        */
 356       if (!query->bo) {
 357          brw_emit_query_begin(brw);
 358       }
 359
 360       assert(query->bo);
 361
 362       brw_emit_query_end(brw);
 363
 364       brw->query.obj = NULL;
 365
 366       brw->stats_wm--;
 367       brw->ctx.NewDriverState |= BRW_NEW_STATS_WM;
 368       break;
 369
 370    default:
 371       unreachable("Unrecognized query target in brw_end_query()");
 372    }
 373 }
 374
 375 /**
 376  * The Gen4-5 WaitQuery() driver hook.
 377  *
 378  * Wait for a query result to become available and return it.  This is the
 379  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 380  */
 381 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 382 {
 383    struct brw_query_object *query = (struct brw_query_object *)q;
 384
 385    assert(brw_context(ctx)->gen < 6);
 386
 387    brw_queryobj_get_results(ctx, query);
 388    query->Base.Ready = true;
 389 }
 390
 391 /**
 392  * The Gen4-5 CheckQuery() driver hook.
 393  *
 394  * Checks whether a query result is ready yet.  If not, flushes.
 395  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 396  */
 397 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 398 {
 399    struct brw_context *brw = brw_context(ctx);
 400    struct brw_query_object *query = (struct brw_query_object *)q;
 401
 402    assert(brw->gen < 6);
 403
 404    /* From the GL_ARB_occlusion_query spec:
 405     *
 406     *     "Instead of allowing for an infinite loop, performing a
 407     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 408     *      not ready yet on the first time it is queried.  This ensures that
 409     *      the async query will return true in finite time.
 410     */
 411    if (query->bo && brw_batch_references(&brw->batch, query->bo))
 412       intel_batchbuffer_flush(brw);
 413
 414    if (query->bo == NULL || !brw_bo_busy(query->bo)) {
 415       brw_queryobj_get_results(ctx, query);
 416       query->Base.Ready = true;
 417    }
 418 }
 419
 420 /**
 421  * Ensure there query's BO has enough space to store a new pair of values.
 422  *
 423  * If not, gather the existing BO's results and create a new buffer of the
 424  * same size.
 425  */
 426 static void
 427 ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query)
 428 {
 429    struct brw_context *brw = brw_context(ctx);
 430
 431    assert(brw->gen < 6);
 432
 433    if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 434
 435       if (query->bo != NULL) {
 436          /* The old query BO did not have enough space, so we allocated a new
 437           * one.  Gather the results so far (adding up the differences) and
 438           * release the old BO.
 439           */
 440          brw_queryobj_get_results(ctx, query);
 441       }
 442
 443       query->bo = brw_bo_alloc(brw->bufmgr, "query", 4096, 1);
 444       query->last_index = 0;
 445    }
 446 }
 447
 448 /**
 449  * Record the PS_DEPTH_COUNT value (for occlusion queries) just before
 450  * primitive drawing.
 451  *
 452  * In a pre-hardware context world, the single PS_DEPTH_COUNT register is
 453  * shared among all applications using the GPU.  However, our query value
 454  * needs to only include fragments generated by our application/GL context.
 455  *
 456  * To accommodate this, we record PS_DEPTH_COUNT at the start and end of
 457  * each batchbuffer (technically, the first primitive drawn and flush time).
 458  * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT
 459  * caused by a batchbuffer.  Since there is no preemption inside batches,
 460  * this is guaranteed to only measure the effects of our current application.
 461  *
 462  * Adding each of these differences (in case drawing is done over many batches)
 463  * produces the final expected value.
 464  *
 465  * In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
 466  * as part of the context state, so this is unnecessary, and skipped.
 467  */
 468 void
 469 brw_emit_query_begin(struct brw_context *brw)
 470 {
 471    struct gl_context *ctx = &brw->ctx;
 472    struct brw_query_object *query = brw->query.obj;
 473
 474    if (brw->hw_ctx)
 475       return;
 476
 477    /* Skip if we're not doing any queries, or we've already recorded the
 478     * initial query value for this batchbuffer.
 479     */
 480    if (!query || brw->query.begin_emitted)
 481       return;
 482
 483    ensure_bo_has_space(ctx, query);
 484
 485    brw_write_depth_count(brw, query->bo, query->last_index * 2);
 486
 487    brw->query.begin_emitted = true;
 488 }
 489
 490 /**
 491  * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
 492  * (for non-hardware context platforms).
 493  *
 494  * See the explanation in brw_emit_query_begin().
 495  */
 496 void
 497 brw_emit_query_end(struct brw_context *brw)
 498 {
 499    struct brw_query_object *query = brw->query.obj;
 500
 501    if (brw->hw_ctx)
 502       return;
 503
 504    if (!brw->query.begin_emitted)
 505       return;
 506
 507    brw_write_depth_count(brw, query->bo, query->last_index * 2 + 1);
 508
 509    brw->query.begin_emitted = false;
 510    query->last_index++;
 511 }
 512
 513 /**
 514  * Driver hook for glQueryCounter().
 515  *
 516  * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
 517  * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
 518  * time while the query is active.
 519  */
 520 void
 521 brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 522 {
 523    struct brw_context *brw = brw_context(ctx);
 524    struct brw_query_object *query = (struct brw_query_object *) q;
 525
 526    assert(q->Target == GL_TIMESTAMP);
 527
 528    brw_bo_unreference(query->bo);
 529    query->bo = brw_bo_alloc(brw->bufmgr, "timestamp query", 4096, 4096);
 530    brw_write_timestamp(brw, query->bo, 0);
 531
 532    query->flushed = false;
 533 }
 534
 535 /**
 536  * Read the TIMESTAMP register immediately (in a non-pipelined fashion).
 537  *
 538  * This is used to implement the GetTimestamp() driver hook.
 539  */
 540 static uint64_t
 541 brw_get_timestamp(struct gl_context *ctx)
 542 {
 543    struct brw_context *brw = brw_context(ctx);
 544    uint64_t result = 0;
 545
 546    switch (brw->screen->hw_has_timestamp) {
 547    case 3: /* New kernel, always full 36bit accuracy */
 548       brw_reg_read(brw->bufmgr, TIMESTAMP | 1, &result);
 549       break;
 550    case 2: /* 64bit kernel, result is left-shifted by 32bits, losing 4bits */
 551       brw_reg_read(brw->bufmgr, TIMESTAMP, &result);
 552       result = result >> 32;
 553       break;
 554    case 1: /* 32bit kernel, result is 36bit wide but may be inaccurate! */
 555       brw_reg_read(brw->bufmgr, TIMESTAMP, &result);
 556       break;
 557    }
 558
 559    /* Scale to nanosecond units */
 560    result = brw_timebase_scale(brw, result);
 561
 562    /* Ensure the scaled timestamp overflows according to
 563     * GL_QUERY_COUNTER_BITS.  Technically this isn't required if
 564     * querying GL_TIMESTAMP via glGetInteger but it seems best to keep
 565     * QueryObject and GetInteger timestamps consistent.
 566     */
 567    result &= (1ull << ctx->Const.QueryCounterBits.Timestamp) - 1;
 568    return result;
 569 }
 570
 571 /**
 572  * Is this type of query written by PIPE_CONTROL?
 573  */
 574 bool
 575 brw_is_query_pipelined(struct brw_query_object *query)
 576 {
 577    switch (query->Base.Target) {
 578    case GL_TIMESTAMP:
 579    case GL_TIME_ELAPSED:
 580    case GL_ANY_SAMPLES_PASSED:
 581    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 582    case GL_SAMPLES_PASSED_ARB:
 583       return true;
 584
 585    case GL_PRIMITIVES_GENERATED:
 586    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 587    case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB:
 588    case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB:
 589    case GL_VERTICES_SUBMITTED_ARB:
 590    case GL_PRIMITIVES_SUBMITTED_ARB:
 591    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 592    case GL_GEOMETRY_SHADER_INVOCATIONS:
 593    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 594    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 595    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 596    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 597    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 598    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 599    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 600       return false;
 601
 602    default:
 603       unreachable("Unrecognized query target in is_query_pipelined()");
 604    }
 605 }
 606
 607 /* Initialize query object functions used on all generations. */
 608 void brw_init_common_queryobj_functions(struct dd_function_table *functions)
 609 {
 610    functions->NewQueryObject = brw_new_query_object;
 611    functions->DeleteQuery = brw_delete_query;
 612    functions->GetTimestamp = brw_get_timestamp;
 613 }
 614
 615 /* Initialize Gen4/5-specific query object functions. */
 616 void gen4_init_queryobj_functions(struct dd_function_table *functions)
 617 {
 618    functions->BeginQuery = brw_begin_query;
 619    functions->EndQuery = brw_end_query;
 620    functions->CheckQuery = brw_check_query;
 621    functions->WaitQuery = brw_wait_query;
 622    functions->QueryCounter = brw_query_counter;
 623 }