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    default:
 191       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 192       break;
 193    }
 194    drm_intel_bo_unmap(query->bo);
 195
 196    /* Now that we've processed the data stored in the query's buffer object,
 197     * we can release it.
 198     */
 199    drm_intel_bo_unreference(query->bo);
 200    query->bo = NULL;
 201 }
 202
 203 /**
 204  * The NewQueryObject() driver hook.
 205  *
 206  * Allocates and initializes a new query object.
 207  */
 208 static struct gl_query_object *
 209 brw_new_query_object(struct gl_context *ctx, GLuint id)
 210 {
 211    struct brw_query_object *query;
 212
 213    query = calloc(1, sizeof(struct brw_query_object));
 214
 215    query->Base.Id = id;
 216    query->Base.Result = 0;
 217    query->Base.Active = false;
 218    query->Base.Ready = true;
 219
 220    return &query->Base;
 221 }
 222
 223 /**
 224  * The DeleteQuery() driver hook.
 225  */
 226 static void
 227 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 228 {
 229    struct brw_query_object *query = (struct brw_query_object *)q;
 230
 231    drm_intel_bo_unreference(query->bo);
 232    free(query);
 233 }
 234
 235 /**
 236  * Gen4-5 driver hook for glBeginQuery().
 237  *
 238  * Initializes driver structures and emits any GPU commands required to begin
 239  * recording data for the query.
 240  */
 241 static void
 242 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 243 {
 244    struct brw_context *brw = brw_context(ctx);
 245    struct intel_context *intel = intel_context(ctx);
 246    struct brw_query_object *query = (struct brw_query_object *)q;
 247
 248    assert(intel->gen < 6);
 249
 250    switch (query->Base.Target) {
 251    case GL_TIME_ELAPSED_EXT:
 252       /* For timestamp queries, we record the starting time right away so that
 253        * we measure the full time between BeginQuery and EndQuery.  There's
 254        * some debate about whether this is the right thing to do.  Our decision
 255        * is based on the following text from the ARB_timer_query extension:
 256        *
 257        * "(5) Should the extension measure total time elapsed between the full
 258        *      completion of the BeginQuery and EndQuery commands, or just time
 259        *      spent in the graphics library?
 260        *
 261        *  RESOLVED:  This extension will measure the total time elapsed
 262        *  between the full completion of these commands.  Future extensions
 263        *  may implement a query to determine time elapsed at different stages
 264        *  of the graphics pipeline."
 265        *
 266        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 267        * we'll write a second timestamp (at index 1), and subtract the two to
 268        * obtain the time elapsed.  Notably, this includes time elapsed while
 269        * the system was doing other work, such as running other applications.
 270        */
 271       drm_intel_bo_unreference(query->bo);
 272       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
 273       write_timestamp(intel, query->bo, 0);
 274       break;
 275
 276    case GL_ANY_SAMPLES_PASSED:
 277    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 278    case GL_SAMPLES_PASSED_ARB:
 279       /* For occlusion queries, we delay taking an initial sample until the
 280        * first drawing occurs in this batch.  See the reasoning in the comments
 281        * for brw_emit_query_begin() below.
 282        *
 283        * Since we're starting a new query, we need to be sure to throw away
 284        * any previous occlusion query results.
 285        */
 286       drm_intel_bo_unreference(query->bo);
 287       query->bo = NULL;
 288       query->last_index = -1;
 289
 290       brw->query.obj = query;
 291
 292       /* Depth statistics on Gen4 require strange workarounds, so we try to
 293        * avoid them when necessary.  They're required for occlusion queries,
 294        * so turn them on now.
 295        */
 296       intel->stats_wm++;
 297       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 298       break;
 299
 300    default:
 301       assert(!"Unrecognized query target in brw_begin_query()");
 302       break;
 303    }
 304 }
 305
 306 /**
 307  * Gen4-5 driver hook for glEndQuery().
 308  *
 309  * Emits GPU commands to record a final query value, ending any data capturing.
 310  * However, the final result isn't necessarily available until the GPU processes
 311  * those commands.  brw_queryobj_get_results() processes the captured data to
 312  * produce the final result.
 313  */
 314 static void
 315 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 316 {
 317    struct brw_context *brw = brw_context(ctx);
 318    struct intel_context *intel = intel_context(ctx);
 319    struct brw_query_object *query = (struct brw_query_object *)q;
 320
 321    assert(intel->gen < 6);
 322
 323    switch (query->Base.Target) {
 324    case GL_TIME_ELAPSED_EXT:
 325       /* Write the final timestamp. */
 326       write_timestamp(intel, query->bo, 1);
 327       break;
 328
 329    case GL_ANY_SAMPLES_PASSED:
 330    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 331    case GL_SAMPLES_PASSED_ARB:
 332
 333       /* No query->bo means that EndQuery was called after BeginQuery with no
 334        * intervening drawing. Rather than doing nothing at all here in this
 335        * case, we emit the query_begin and query_end state to the
 336        * hardware. This is to guarantee that waiting on the result of this
 337        * empty state will cause all previous queries to complete at all, as
 338        * required by the specification:
 339        *
 340        *        It must always be true that if any query object
 341        *        returns a result available of TRUE, all queries of the
 342        *        same type issued prior to that query must also return
 343        *        TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
 344        */
 345       if (!query->bo) {
 346          brw_emit_query_begin(brw);
 347       }
 348
 349       assert(query->bo);
 350
 351       brw_emit_query_end(brw);
 352
 353       brw->query.obj = NULL;
 354
 355       intel->stats_wm--;
 356       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 357       break;
 358
 359    default:
 360       assert(!"Unrecognized query target in brw_end_query()");
 361       break;
 362    }
 363 }
 364
 365 /**
 366  * The Gen4-5 WaitQuery() driver hook.
 367  *
 368  * Wait for a query result to become available and return it.  This is the
 369  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 370  */
 371 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 372 {
 373    struct brw_query_object *query = (struct brw_query_object *)q;
 374
 375    assert(intel_context(ctx)->gen < 6);
 376
 377    brw_queryobj_get_results(ctx, query);
 378    query->Base.Ready = true;
 379 }
 380
 381 /**
 382  * The Gen4-5 CheckQuery() driver hook.
 383  *
 384  * Checks whether a query result is ready yet.  If not, flushes.
 385  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 386  */
 387 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 388 {
 389    struct intel_context *intel = intel_context(ctx);
 390    struct brw_query_object *query = (struct brw_query_object *)q;
 391
 392    assert(intel->gen < 6);
 393
 394    /* From the GL_ARB_occlusion_query spec:
 395     *
 396     *     "Instead of allowing for an infinite loop, performing a
 397     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 398     *      not ready yet on the first time it is queried.  This ensures that
 399     *      the async query will return true in finite time.
 400     */
 401    if (query->bo && drm_intel_bo_references(intel->batch.bo, query->bo))
 402       intel_batchbuffer_flush(intel);
 403
 404    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 405       brw_queryobj_get_results(ctx, query);
 406       query->Base.Ready = true;
 407    }
 408 }
 409
 410 /**
 411  * Ensure there query's BO has enough space to store a new pair of values.
 412  *
 413  * If not, gather the existing BO's results and create a new buffer of the
 414  * same size.
 415  */
 416 static void
 417 ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query)
 418 {
 419    struct intel_context *intel = intel_context(ctx);
 420
 421    assert(intel->gen < 6);
 422
 423    if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 424
 425       if (query->bo != NULL) {
 426          /* The old query BO did not have enough space, so we allocated a new
 427           * one.  Gather the results so far (adding up the differences) and
 428           * release the old BO.
 429           */
 430          brw_queryobj_get_results(ctx, query);
 431       }
 432
 433       query->bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
 434       query->last_index = 0;
 435    }
 436 }
 437
 438 /**
 439  * Record the PS_DEPTH_COUNT value (for occlusion queries) just before
 440  * primitive drawing.
 441  *
 442  * In a pre-hardware context world, the single PS_DEPTH_COUNT register is
 443  * shared among all applications using the GPU.  However, our query value
 444  * needs to only include fragments generated by our application/GL context.
 445  *
 446  * To accommodate this, we record PS_DEPTH_COUNT at the start and end of
 447  * each batchbuffer (technically, the first primitive drawn and flush time).
 448  * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT
 449  * caused by a batchbuffer.  Since there is no preemption inside batches,
 450  * this is guaranteed to only measure the effects of our current application.
 451  *
 452  * Adding each of these differences (in case drawing is done over many batches)
 453  * produces the final expected value.
 454  *
 455  * In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
 456  * as part of the context state, so this is unnecessary, and skipped.
 457  */
 458 void
 459 brw_emit_query_begin(struct brw_context *brw)
 460 {
 461    struct intel_context *intel = &brw->intel;
 462    struct gl_context *ctx = &intel->ctx;
 463    struct brw_query_object *query = brw->query.obj;
 464
 465    if (intel->hw_ctx)
 466       return;
 467
 468    /* Skip if we're not doing any queries, or we've already recorded the
 469     * initial query value for this batchbuffer.
 470     */
 471    if (!query || brw->query.begin_emitted)
 472       return;
 473
 474    ensure_bo_has_space(ctx, query);
 475
 476    write_depth_count(intel, query->bo, query->last_index * 2);
 477
 478    brw->query.begin_emitted = true;
 479 }
 480
 481 /**
 482  * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
 483  * (for non-hardware context platforms).
 484  *
 485  * See the explanation in brw_emit_query_begin().
 486  */
 487 void
 488 brw_emit_query_end(struct brw_context *brw)
 489 {
 490    struct intel_context *intel = &brw->intel;
 491    struct brw_query_object *query = brw->query.obj;
 492
 493    if (intel->hw_ctx)
 494       return;
 495
 496    if (!brw->query.begin_emitted)
 497       return;
 498
 499    write_depth_count(intel, query->bo, query->last_index * 2 + 1);
 500
 501    brw->query.begin_emitted = false;
 502    query->last_index++;
 503 }
 504
 505 /**
 506  * Driver hook for glQueryCounter().
 507  *
 508  * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
 509  * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
 510  * time while the query is active.
 511  */
 512 static void
 513 brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 514 {
 515    struct intel_context *intel = intel_context(ctx);
 516    struct brw_query_object *query = (struct brw_query_object *) q;
 517
 518    assert(q->Target == GL_TIMESTAMP);
 519
 520    drm_intel_bo_unreference(query->bo);
 521    query->bo = drm_intel_bo_alloc(intel->bufmgr, "timestamp query", 4096, 4096);
 522    write_timestamp(intel, query->bo, 0);
 523 }
 524
 525 /**
 526  * Read the TIMESTAMP register immediately (in a non-pipelined fashion).
 527  *
 528  * This is used to implement the GetTimestamp() driver hook.
 529  */
 530 static uint64_t
 531 brw_get_timestamp(struct gl_context *ctx)
 532 {
 533    struct intel_context *intel = intel_context(ctx);
 534    uint64_t result = 0;
 535
 536    drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
 537
 538    /* See logic in brw_queryobj_get_results() */
 539    result = result >> 32;
 540    result *= 80;
 541    result &= (1ull << 36) - 1;
 542
 543    return result;
 544 }
 545
 546 /* Initialize query object functions used on all generations. */
 547 void brw_init_common_queryobj_functions(struct dd_function_table *functions)
 548 {
 549    functions->NewQueryObject = brw_new_query_object;
 550    functions->DeleteQuery = brw_delete_query;
 551    functions->QueryCounter = brw_query_counter;
 552    functions->GetTimestamp = brw_get_timestamp;
 553 }
 554
 555 /* Initialize Gen4/5-specific query object functions. */
 556 void gen4_init_queryobj_functions(struct dd_function_table *functions)
 557 {
 558    functions->BeginQuery = brw_begin_query;
 559    functions->EndQuery = brw_end_query;
 560    functions->CheckQuery = brw_check_query;
 561    functions->WaitQuery = brw_wait_query;
 562 }