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 #include "intel_reg.h"
  45
  46 /**
  47  * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
  48  */
  49 static void
  50 write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  51 {
  52    if (intel->gen >= 6) {
  53       /* Emit workaround flushes: */
  54       if (intel->gen == 6) {
  55          /* The timestamp write below is a non-zero post-sync op, which on
  56           * Gen6 necessitates a CS stall.  CS stalls need stall at scoreboard
  57           * set.  See the comments for intel_emit_post_sync_nonzero_flush().
  58           */
  59          BEGIN_BATCH(4);
  60          OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
  61          OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
  62          OUT_BATCH(0);
  63          OUT_BATCH(0);
  64          ADVANCE_BATCH();
  65       }
  66
  67       BEGIN_BATCH(5);
  68       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  69       OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
  70       OUT_RELOC(query_bo,
  71                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  72                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  73                 idx * sizeof(uint64_t));
  74       OUT_BATCH(0);
  75       OUT_BATCH(0);
  76       ADVANCE_BATCH();
  77    } else {
  78       BEGIN_BATCH(4);
  79       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
  80                 PIPE_CONTROL_WRITE_TIMESTAMP);
  81       OUT_RELOC(query_bo,
  82                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  83                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  84                 idx * sizeof(uint64_t));
  85       OUT_BATCH(0);
  86       OUT_BATCH(0);
  87       ADVANCE_BATCH();
  88    }
  89 }
  90
  91 /**
  92  * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
  93  */
  94 static void
  95 write_depth_count(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  96 {
  97    assert(intel->gen < 6);
  98
  99    BEGIN_BATCH(4);
 100    OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
 101              PIPE_CONTROL_DEPTH_STALL | PIPE_CONTROL_WRITE_DEPTH_COUNT);
 102    /* This object could be mapped cacheable, but we don't have an exposed
 103     * mechanism to support that.  Since it's going uncached, tell GEM that
 104     * we're writing to it.  The usual clflush should be all that's required
 105     * to pick up the results.
 106     */
 107    OUT_RELOC(query_bo,
 108              I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 109              PIPE_CONTROL_GLOBAL_GTT_WRITE |
 110              (idx * sizeof(uint64_t)));
 111    OUT_BATCH(0);
 112    OUT_BATCH(0);
 113    ADVANCE_BATCH();
 114 }
 115
 116 /**
 117  * Wait on the query object's BO and calculate the final result.
 118  */
 119 static void
 120 brw_queryobj_get_results(struct gl_context *ctx,
 121                          struct brw_query_object *query)
 122 {
 123    struct intel_context *intel = intel_context(ctx);
 124
 125    int i;
 126    uint64_t *results;
 127
 128    assert(intel->gen < 6);
 129
 130    if (query->bo == NULL)
 131       return;
 132
 133    /* If the application has requested the query result, but this batch is
 134     * still contributing to it, flush it now so the results will be present
 135     * when mapped.
 136     */
 137    if (drm_intel_bo_references(intel->batch.bo, query->bo))
 138       intel_batchbuffer_flush(intel);
 139
 140    if (unlikely(intel->perf_debug)) {
 141       if (drm_intel_bo_busy(query->bo)) {
 142          perf_debug("Stalling on the GPU waiting for a query object.\n");
 143       }
 144    }
 145
 146    drm_intel_bo_map(query->bo, false);
 147    results = query->bo->virtual;
 148    switch (query->Base.Target) {
 149    case GL_TIME_ELAPSED_EXT:
 150       /* The query BO contains the starting and ending timestamps.
 151        * Subtract the two and convert to nanoseconds.
 152        */
 153       query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
 154       break;
 155
 156    case GL_TIMESTAMP:
 157       /* The query BO contains a single timestamp value in results[0]. */
 158       query->Base.Result = 1000 * (results[0] >> 32);
 159       break;
 160
 161    case GL_SAMPLES_PASSED_ARB:
 162       /* Loop over pairs of values from the BO, which are the PS_DEPTH_COUNT
 163        * value at the start and end of the batchbuffer.  Subtract them to
 164        * get the number of fragments which passed the depth test in each
 165        * individual batch, and add those differences up to get the number
 166        * of fragments for the entire query.
 167        *
 168        * Note that query->Base.Result may already be non-zero.  We may have
 169        * run out of space in the query's BO and allocated a new one.  If so,
 170        * this function was already called to accumulate the results so far.
 171        */
 172       for (i = 0; i < query->last_index; i++) {
 173          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 174       }
 175       break;
 176
 177    case GL_ANY_SAMPLES_PASSED:
 178    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 179       /* If the starting and ending PS_DEPTH_COUNT from any of the batches
 180        * differ, then some fragments passed the depth test.
 181        */
 182       for (i = 0; i < query->last_index; i++) {
 183          if (results[i * 2 + 1] != results[i * 2]) {
 184             query->Base.Result = GL_TRUE;
 185             break;
 186          }
 187       }
 188       break;
 189
 190    case GL_PRIMITIVES_GENERATED:
 191    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 192       /* We don't actually query the hardware for this value, so query->bo
 193        * should always be NULL and execution should never reach here.
 194        */
 195       assert(!"Unreachable");
 196       break;
 197
 198    default:
 199       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 200       break;
 201    }
 202    drm_intel_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    drm_intel_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    drm_intel_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 intel_context *intel = intel_context(ctx);
 254    struct brw_query_object *query = (struct brw_query_object *)q;
 255
 256    assert(intel->gen < 6);
 257
 258    switch (query->Base.Target) {
 259    case GL_TIME_ELAPSED_EXT:
 260       /* For timestamp queries, we record the starting time right away so that
 261        * we measure the full time between BeginQuery and EndQuery.  There's
 262        * some debate about whether this is the right thing to do.  Our decision
 263        * is based on the following text from the ARB_timer_query extension:
 264        *
 265        * "(5) Should the extension measure total time elapsed between the full
 266        *      completion of the BeginQuery and EndQuery commands, or just time
 267        *      spent in the graphics library?
 268        *
 269        *  RESOLVED:  This extension will measure the total time elapsed
 270        *  between the full completion of these commands.  Future extensions
 271        *  may implement a query to determine time elapsed at different stages
 272        *  of the graphics pipeline."
 273        *
 274        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 275        * we'll write a second timestamp (at index 1), and subtract the two to
 276        * obtain the time elapsed.  Notably, this includes time elapsed while
 277        * the system was doing other work, such as running other applications.
 278        */
 279       drm_intel_bo_unreference(query->bo);
 280       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
 281       write_timestamp(intel, query->bo, 0);
 282       break;
 283
 284    case GL_ANY_SAMPLES_PASSED:
 285    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 286    case GL_SAMPLES_PASSED_ARB:
 287       /* For occlusion queries, we delay taking an initial sample until the
 288        * first drawing occurs in this batch.  See the reasoning in the comments
 289        * for brw_emit_query_begin() below.
 290        *
 291        * Since we're starting a new query, we need to be sure to throw away
 292        * any previous occlusion query results.
 293        */
 294       drm_intel_bo_unreference(query->bo);
 295       query->bo = NULL;
 296       query->last_index = -1;
 297
 298       brw->query.obj = query;
 299
 300       /* Depth statistics on Gen4 require strange workarounds, so we try to
 301        * avoid them when necessary.  They're required for occlusion queries,
 302        * so turn them on now.
 303        */
 304       intel->stats_wm++;
 305       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 306       break;
 307
 308    case GL_PRIMITIVES_GENERATED:
 309       /* We don't actually query the hardware for this value; we keep track of
 310        * it a software counter.  So just reset the counter.
 311        */
 312       brw->sol.primitives_generated = 0;
 313       brw->sol.counting_primitives_generated = true;
 314       break;
 315
 316    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 317       /* We don't actually query the hardware for this value; we keep track of
 318        * it a software counter.  So just reset the counter.
 319        */
 320       brw->sol.primitives_written = 0;
 321       brw->sol.counting_primitives_written = true;
 322       break;
 323
 324    default:
 325       assert(!"Unrecognized query target in brw_begin_query()");
 326       break;
 327    }
 328 }
 329
 330 /**
 331  * Gen4-5 driver hook for glEndQuery().
 332  *
 333  * Emits GPU commands to record a final query value, ending any data capturing.
 334  * However, the final result isn't necessarily available until the GPU processes
 335  * those commands.  brw_queryobj_get_results() processes the captured data to
 336  * produce the final result.
 337  */
 338 static void
 339 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 340 {
 341    struct brw_context *brw = brw_context(ctx);
 342    struct intel_context *intel = intel_context(ctx);
 343    struct brw_query_object *query = (struct brw_query_object *)q;
 344
 345    assert(intel->gen < 6);
 346
 347    switch (query->Base.Target) {
 348    case GL_TIME_ELAPSED_EXT:
 349       /* Write the final timestamp. */
 350       write_timestamp(intel, query->bo, 1);
 351       break;
 352
 353    case GL_ANY_SAMPLES_PASSED:
 354    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 355    case GL_SAMPLES_PASSED_ARB:
 356
 357       /* No query->bo means that EndQuery was called after BeginQuery with no
 358        * intervening drawing. Rather than doing nothing at all here in this
 359        * case, we emit the query_begin and query_end state to the
 360        * hardware. This is to guarantee that waiting on the result of this
 361        * empty state will cause all previous queries to complete at all, as
 362        * required by the specification:
 363        *
 364        *        It must always be true that if any query object
 365        *        returns a result available of TRUE, all queries of the
 366        *        same type issued prior to that query must also return
 367        *        TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
 368        */
 369       if (!query->bo) {
 370          brw_emit_query_begin(brw);
 371       }
 372
 373       assert(query->bo);
 374
 375       brw_emit_query_end(brw);
 376
 377       brw->query.obj = NULL;
 378
 379       intel->stats_wm--;
 380       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 381       break;
 382
 383    case GL_PRIMITIVES_GENERATED:
 384       /* We don't actually query the hardware for this value; we keep track of
 385        * it in a software counter.  So just read the counter and store it in
 386        * the query object.
 387        */
 388       query->Base.Result = brw->sol.primitives_generated;
 389       brw->sol.counting_primitives_generated = false;
 390
 391       /* And set query->bo to NULL so that this query won't try to wait
 392        * for any rendering to complete.
 393        */
 394       query->bo = NULL;
 395       break;
 396
 397    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 398       /* We don't actually query the hardware for this value; we keep track of
 399        * it in a software counter.  So just read the counter and store it in
 400        * the query object.
 401        */
 402       query->Base.Result = brw->sol.primitives_written;
 403       brw->sol.counting_primitives_written = false;
 404
 405       /* And set query->bo to NULL so that this query won't try to wait
 406        * for any rendering to complete.
 407        */
 408       query->bo = NULL;
 409       break;
 410
 411    default:
 412       assert(!"Unrecognized query target in brw_end_query()");
 413       break;
 414    }
 415 }
 416
 417 /**
 418  * The Gen4-5 WaitQuery() driver hook.
 419  *
 420  * Wait for a query result to become available and return it.  This is the
 421  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 422  */
 423 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 424 {
 425    struct brw_query_object *query = (struct brw_query_object *)q;
 426
 427    assert(intel_context(ctx)->gen < 6);
 428
 429    brw_queryobj_get_results(ctx, query);
 430    query->Base.Ready = true;
 431 }
 432
 433 /**
 434  * The Gen4-5 CheckQuery() driver hook.
 435  *
 436  * Checks whether a query result is ready yet.  If not, flushes.
 437  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 438  */
 439 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 440 {
 441    struct intel_context *intel = intel_context(ctx);
 442    struct brw_query_object *query = (struct brw_query_object *)q;
 443
 444    assert(intel->gen < 6);
 445
 446    /* From the GL_ARB_occlusion_query spec:
 447     *
 448     *     "Instead of allowing for an infinite loop, performing a
 449     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 450     *      not ready yet on the first time it is queried.  This ensures that
 451     *      the async query will return true in finite time.
 452     */
 453    if (query->bo && drm_intel_bo_references(intel->batch.bo, query->bo))
 454       intel_batchbuffer_flush(intel);
 455
 456    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 457       brw_queryobj_get_results(ctx, query);
 458       query->Base.Ready = true;
 459    }
 460 }
 461
 462 /**
 463  * Ensure there query's BO has enough space to store a new pair of values.
 464  *
 465  * If not, gather the existing BO's results and create a new buffer of the
 466  * same size.
 467  */
 468 static void
 469 ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query)
 470 {
 471    struct intel_context *intel = intel_context(ctx);
 472
 473    assert(intel->gen < 6);
 474
 475    if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 476
 477       if (query->bo != NULL) {
 478          /* The old query BO did not have enough space, so we allocated a new
 479           * one.  Gather the results so far (adding up the differences) and
 480           * release the old BO.
 481           */
 482          brw_queryobj_get_results(ctx, query);
 483       }
 484
 485       query->bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
 486       query->last_index = 0;
 487    }
 488 }
 489
 490 /**
 491  * Record the PS_DEPTH_COUNT value (for occlusion queries) just before
 492  * primitive drawing.
 493  *
 494  * In a pre-hardware context world, the single PS_DEPTH_COUNT register is
 495  * shared among all applications using the GPU.  However, our query value
 496  * needs to only include fragments generated by our application/GL context.
 497  *
 498  * To accommodate this, we record PS_DEPTH_COUNT at the start and end of
 499  * each batchbuffer (technically, the first primitive drawn and flush time).
 500  * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT
 501  * caused by a batchbuffer.  Since there is no preemption inside batches,
 502  * this is guaranteed to only measure the effects of our current application.
 503  *
 504  * Adding each of these differences (in case drawing is done over many batches)
 505  * produces the final expected value.
 506  *
 507  * In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
 508  * as part of the context state, so this is unnecessary, and skipped.
 509  */
 510 void
 511 brw_emit_query_begin(struct brw_context *brw)
 512 {
 513    struct intel_context *intel = &brw->intel;
 514    struct gl_context *ctx = &intel->ctx;
 515    struct brw_query_object *query = brw->query.obj;
 516
 517    if (intel->hw_ctx)
 518       return;
 519
 520    /* Skip if we're not doing any queries, or we've already recorded the
 521     * initial query value for this batchbuffer.
 522     */
 523    if (!query || brw->query.begin_emitted)
 524       return;
 525
 526    ensure_bo_has_space(ctx, query);
 527
 528    write_depth_count(intel, query->bo, query->last_index * 2);
 529
 530    brw->query.begin_emitted = true;
 531 }
 532
 533 /**
 534  * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
 535  * (for non-hardware context platforms).
 536  *
 537  * See the explanation in brw_emit_query_begin().
 538  */
 539 void
 540 brw_emit_query_end(struct brw_context *brw)
 541 {
 542    struct intel_context *intel = &brw->intel;
 543    struct brw_query_object *query = brw->query.obj;
 544
 545    if (intel->hw_ctx)
 546       return;
 547
 548    if (!brw->query.begin_emitted)
 549       return;
 550
 551    write_depth_count(intel, query->bo, query->last_index * 2 + 1);
 552
 553    brw->query.begin_emitted = false;
 554    query->last_index++;
 555 }
 556
 557 /**
 558  * Driver hook for glQueryCounter().
 559  *
 560  * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
 561  * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
 562  * time while the query is active.
 563  */
 564 static void
 565 brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 566 {
 567    struct intel_context *intel = intel_context(ctx);
 568    struct brw_query_object *query = (struct brw_query_object *) q;
 569
 570    assert(q->Target == GL_TIMESTAMP);
 571
 572    drm_intel_bo_unreference(query->bo);
 573    query->bo = drm_intel_bo_alloc(intel->bufmgr, "timestamp query", 4096, 4096);
 574    write_timestamp(intel, query->bo, 0);
 575 }
 576
 577 /**
 578  * Read the TIMESTAMP register immediately (in a non-pipelined fashion).
 579  *
 580  * This is used to implement the GetTimestamp() driver hook.
 581  */
 582 static uint64_t
 583 brw_get_timestamp(struct gl_context *ctx)
 584 {
 585    struct intel_context *intel = intel_context(ctx);
 586    uint64_t result = 0;
 587
 588    drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
 589
 590    /* See logic in brw_queryobj_get_results() */
 591    result = result >> 32;
 592    result *= 80;
 593    result &= (1ull << 36) - 1;
 594
 595    return result;
 596 }
 597
 598 /* Initialize query object functions used on all generations. */
 599 void brw_init_common_queryobj_functions(struct dd_function_table *functions)
 600 {
 601    functions->NewQueryObject = brw_new_query_object;
 602    functions->DeleteQuery = brw_delete_query;
 603    functions->QueryCounter = brw_query_counter;
 604    functions->GetTimestamp = brw_get_timestamp;
 605 }
 606
 607 /* Initialize Gen4/5-specific query object functions. */
 608 void gen4_init_queryobj_functions(struct dd_function_table *functions)
 609 {
 610    functions->BeginQuery = brw_begin_query;
 611    functions->EndQuery = brw_end_query;
 612    functions->CheckQuery = brw_check_query;
 613    functions->WaitQuery = brw_wait_query;
 614 }