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