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 brw_context *brw, drm_intel_bo *query_bo, int idx)
  51 {
  52    struct intel_context *intel = &brw->intel;
  53    if (intel->gen >= 6) {
  54       /* Emit workaround flushes: */
  55       if (intel->gen == 6) {
  56          /* The timestamp write below is a non-zero post-sync op, which on
  57           * Gen6 necessitates a CS stall.  CS stalls need stall at scoreboard
  58           * set.  See the comments for intel_emit_post_sync_nonzero_flush().
  59           */
  60          BEGIN_BATCH(4);
  61          OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
  62          OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
  63          OUT_BATCH(0);
  64          OUT_BATCH(0);
  65          ADVANCE_BATCH();
  66       }
  67
  68       BEGIN_BATCH(5);
  69       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  70       OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
  71       OUT_RELOC(query_bo,
  72                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  73                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  74                 idx * sizeof(uint64_t));
  75       OUT_BATCH(0);
  76       OUT_BATCH(0);
  77       ADVANCE_BATCH();
  78    } else {
  79       BEGIN_BATCH(4);
  80       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
  81                 PIPE_CONTROL_WRITE_TIMESTAMP);
  82       OUT_RELOC(query_bo,
  83                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  84                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  85                 idx * sizeof(uint64_t));
  86       OUT_BATCH(0);
  87       OUT_BATCH(0);
  88       ADVANCE_BATCH();
  89    }
  90 }
  91
  92 /**
  93  * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
  94  */
  95 static void
  96 write_depth_count(struct brw_context *brw, drm_intel_bo *query_bo, int idx)
  97 {
  98    struct intel_context *intel = &brw->intel;
  99    assert(intel->gen < 6);
 100
 101    BEGIN_BATCH(4);
 102    OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
 103              PIPE_CONTROL_DEPTH_STALL | PIPE_CONTROL_WRITE_DEPTH_COUNT);
 104    /* This object could be mapped cacheable, but we don't have an exposed
 105     * mechanism to support that.  Since it's going uncached, tell GEM that
 106     * we're writing to it.  The usual clflush should be all that's required
 107     * to pick up the results.
 108     */
 109    OUT_RELOC(query_bo,
 110              I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 111              PIPE_CONTROL_GLOBAL_GTT_WRITE |
 112              (idx * sizeof(uint64_t)));
 113    OUT_BATCH(0);
 114    OUT_BATCH(0);
 115    ADVANCE_BATCH();
 116 }
 117
 118 /**
 119  * Wait on the query object's BO and calculate the final result.
 120  */
 121 static void
 122 brw_queryobj_get_results(struct gl_context *ctx,
 123                          struct brw_query_object *query)
 124 {
 125    struct brw_context *brw = brw_context(ctx);
 126    struct intel_context *intel = intel_context(ctx);
 127
 128    int i;
 129    uint64_t *results;
 130
 131    assert(intel->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 (drm_intel_bo_references(brw->batch.bo, query->bo))
 141       intel_batchbuffer_flush(brw);
 142
 143    if (unlikely(intel->perf_debug)) {
 144       if (drm_intel_bo_busy(query->bo)) {
 145          perf_debug("Stalling on the GPU waiting for a query object.\n");
 146       }
 147    }
 148
 149    drm_intel_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 += 1000 * ((results[1] >> 32) - (results[0] >> 32));
 157       break;
 158
 159    case GL_TIMESTAMP:
 160       /* The query BO contains a single timestamp value in results[0]. */
 161       query->Base.Result = 1000 * (results[0] >> 32);
 162       break;
 163
 164    case GL_SAMPLES_PASSED_ARB:
 165       /* Loop over pairs of values from the BO, which are the PS_DEPTH_COUNT
 166        * value at the start and end of the batchbuffer.  Subtract them to
 167        * get the number of fragments which passed the depth test in each
 168        * individual batch, and add those differences up to get the number
 169        * of fragments for the entire query.
 170        *
 171        * Note that query->Base.Result may already be non-zero.  We may have
 172        * run out of space in the query's BO and allocated a new one.  If so,
 173        * this function was already called to accumulate the results so far.
 174        */
 175       for (i = 0; i < query->last_index; i++) {
 176          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 177       }
 178       break;
 179
 180    case GL_ANY_SAMPLES_PASSED:
 181    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 182       /* If the starting and ending PS_DEPTH_COUNT from any of the batches
 183        * differ, then some fragments passed the depth test.
 184        */
 185       for (i = 0; i < query->last_index; i++) {
 186          if (results[i * 2 + 1] != results[i * 2]) {
 187             query->Base.Result = GL_TRUE;
 188             break;
 189          }
 190       }
 191       break;
 192
 193    default:
 194       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 195       break;
 196    }
 197    drm_intel_bo_unmap(query->bo);
 198
 199    /* Now that we've processed the data stored in the query's buffer object,
 200     * we can release it.
 201     */
 202    drm_intel_bo_unreference(query->bo);
 203    query->bo = NULL;
 204 }
 205
 206 /**
 207  * The NewQueryObject() driver hook.
 208  *
 209  * Allocates and initializes a new query object.
 210  */
 211 static struct gl_query_object *
 212 brw_new_query_object(struct gl_context *ctx, GLuint id)
 213 {
 214    struct brw_query_object *query;
 215
 216    query = calloc(1, sizeof(struct brw_query_object));
 217
 218    query->Base.Id = id;
 219    query->Base.Result = 0;
 220    query->Base.Active = false;
 221    query->Base.Ready = true;
 222
 223    return &query->Base;
 224 }
 225
 226 /**
 227  * The DeleteQuery() driver hook.
 228  */
 229 static void
 230 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 231 {
 232    struct brw_query_object *query = (struct brw_query_object *)q;
 233
 234    drm_intel_bo_unreference(query->bo);
 235    free(query);
 236 }
 237
 238 /**
 239  * Gen4-5 driver hook for glBeginQuery().
 240  *
 241  * Initializes driver structures and emits any GPU commands required to begin
 242  * recording data for the query.
 243  */
 244 static void
 245 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 246 {
 247    struct brw_context *brw = brw_context(ctx);
 248    struct intel_context *intel = intel_context(ctx);
 249    struct brw_query_object *query = (struct brw_query_object *)q;
 250
 251    assert(intel->gen < 6);
 252
 253    switch (query->Base.Target) {
 254    case GL_TIME_ELAPSED_EXT:
 255       /* For timestamp queries, we record the starting time right away so that
 256        * we measure the full time between BeginQuery and EndQuery.  There's
 257        * some debate about whether this is the right thing to do.  Our decision
 258        * is based on the following text from the ARB_timer_query extension:
 259        *
 260        * "(5) Should the extension measure total time elapsed between the full
 261        *      completion of the BeginQuery and EndQuery commands, or just time
 262        *      spent in the graphics library?
 263        *
 264        *  RESOLVED:  This extension will measure the total time elapsed
 265        *  between the full completion of these commands.  Future extensions
 266        *  may implement a query to determine time elapsed at different stages
 267        *  of the graphics pipeline."
 268        *
 269        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 270        * we'll write a second timestamp (at index 1), and subtract the two to
 271        * obtain the time elapsed.  Notably, this includes time elapsed while
 272        * the system was doing other work, such as running other applications.
 273        */
 274       drm_intel_bo_unreference(query->bo);
 275       query->bo = drm_intel_bo_alloc(brw->bufmgr, "timer query", 4096, 4096);
 276       write_timestamp(brw, query->bo, 0);
 277       break;
 278
 279    case GL_ANY_SAMPLES_PASSED:
 280    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 281    case GL_SAMPLES_PASSED_ARB:
 282       /* For occlusion queries, we delay taking an initial sample until the
 283        * first drawing occurs in this batch.  See the reasoning in the comments
 284        * for brw_emit_query_begin() below.
 285        *
 286        * Since we're starting a new query, we need to be sure to throw away
 287        * any previous occlusion query results.
 288        */
 289       drm_intel_bo_unreference(query->bo);
 290       query->bo = NULL;
 291       query->last_index = -1;
 292
 293       brw->query.obj = query;
 294
 295       /* Depth statistics on Gen4 require strange workarounds, so we try to
 296        * avoid them when necessary.  They're required for occlusion queries,
 297        * so turn them on now.
 298        */
 299       intel->stats_wm++;
 300       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 301       break;
 302
 303    default:
 304       assert(!"Unrecognized query target in brw_begin_query()");
 305       break;
 306    }
 307 }
 308
 309 /**
 310  * Gen4-5 driver hook for glEndQuery().
 311  *
 312  * Emits GPU commands to record a final query value, ending any data capturing.
 313  * However, the final result isn't necessarily available until the GPU processes
 314  * those commands.  brw_queryobj_get_results() processes the captured data to
 315  * produce the final result.
 316  */
 317 static void
 318 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 319 {
 320    struct brw_context *brw = brw_context(ctx);
 321    struct intel_context *intel = intel_context(ctx);
 322    struct brw_query_object *query = (struct brw_query_object *)q;
 323
 324    assert(intel->gen < 6);
 325
 326    switch (query->Base.Target) {
 327    case GL_TIME_ELAPSED_EXT:
 328       /* Write the final timestamp. */
 329       write_timestamp(brw, query->bo, 1);
 330       break;
 331
 332    case GL_ANY_SAMPLES_PASSED:
 333    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 334    case GL_SAMPLES_PASSED_ARB:
 335
 336       /* No query->bo means that EndQuery was called after BeginQuery with no
 337        * intervening drawing. Rather than doing nothing at all here in this
 338        * case, we emit the query_begin and query_end state to the
 339        * hardware. This is to guarantee that waiting on the result of this
 340        * empty state will cause all previous queries to complete at all, as
 341        * required by the specification:
 342        *
 343        *        It must always be true that if any query object
 344        *        returns a result available of TRUE, all queries of the
 345        *        same type issued prior to that query must also return
 346        *        TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
 347        */
 348       if (!query->bo) {
 349          brw_emit_query_begin(brw);
 350       }
 351
 352       assert(query->bo);
 353
 354       brw_emit_query_end(brw);
 355
 356       brw->query.obj = NULL;
 357
 358       intel->stats_wm--;
 359       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 360       break;
 361
 362    default:
 363       assert(!"Unrecognized query target in brw_end_query()");
 364       break;
 365    }
 366 }
 367
 368 /**
 369  * The Gen4-5 WaitQuery() driver hook.
 370  *
 371  * Wait for a query result to become available and return it.  This is the
 372  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 373  */
 374 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 375 {
 376    struct brw_query_object *query = (struct brw_query_object *)q;
 377
 378    assert(intel_context(ctx)->gen < 6);
 379
 380    brw_queryobj_get_results(ctx, query);
 381    query->Base.Ready = true;
 382 }
 383
 384 /**
 385  * The Gen4-5 CheckQuery() driver hook.
 386  *
 387  * Checks whether a query result is ready yet.  If not, flushes.
 388  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 389  */
 390 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 391 {
 392    struct brw_context *brw = brw_context(ctx);
 393    struct intel_context *intel = intel_context(ctx);
 394    struct brw_query_object *query = (struct brw_query_object *)q;
 395
 396    assert(intel->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 && drm_intel_bo_references(brw->batch.bo, query->bo))
 406       intel_batchbuffer_flush(brw);
 407
 408    if (query->bo == NULL || !drm_intel_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    struct intel_context *intel = intel_context(ctx);
 425
 426    assert(intel->gen < 6);
 427
 428    if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 429
 430       if (query->bo != NULL) {
 431          /* The old query BO did not have enough space, so we allocated a new
 432           * one.  Gather the results so far (adding up the differences) and
 433           * release the old BO.
 434           */
 435          brw_queryobj_get_results(ctx, query);
 436       }
 437
 438       query->bo = drm_intel_bo_alloc(brw->bufmgr, "query", 4096, 1);
 439       query->last_index = 0;
 440    }
 441 }
 442
 443 /**
 444  * Record the PS_DEPTH_COUNT value (for occlusion queries) just before
 445  * primitive drawing.
 446  *
 447  * In a pre-hardware context world, the single PS_DEPTH_COUNT register is
 448  * shared among all applications using the GPU.  However, our query value
 449  * needs to only include fragments generated by our application/GL context.
 450  *
 451  * To accommodate this, we record PS_DEPTH_COUNT at the start and end of
 452  * each batchbuffer (technically, the first primitive drawn and flush time).
 453  * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT
 454  * caused by a batchbuffer.  Since there is no preemption inside batches,
 455  * this is guaranteed to only measure the effects of our current application.
 456  *
 457  * Adding each of these differences (in case drawing is done over many batches)
 458  * produces the final expected value.
 459  *
 460  * In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
 461  * as part of the context state, so this is unnecessary, and skipped.
 462  */
 463 void
 464 brw_emit_query_begin(struct brw_context *brw)
 465 {
 466    struct intel_context *intel = &brw->intel;
 467    struct gl_context *ctx = &intel->ctx;
 468    struct brw_query_object *query = brw->query.obj;
 469
 470    if (brw->hw_ctx)
 471       return;
 472
 473    /* Skip if we're not doing any queries, or we've already recorded the
 474     * initial query value for this batchbuffer.
 475     */
 476    if (!query || brw->query.begin_emitted)
 477       return;
 478
 479    ensure_bo_has_space(ctx, query);
 480
 481    write_depth_count(brw, query->bo, query->last_index * 2);
 482
 483    brw->query.begin_emitted = true;
 484 }
 485
 486 /**
 487  * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
 488  * (for non-hardware context platforms).
 489  *
 490  * See the explanation in brw_emit_query_begin().
 491  */
 492 void
 493 brw_emit_query_end(struct brw_context *brw)
 494 {
 495    struct brw_query_object *query = brw->query.obj;
 496
 497    if (brw->hw_ctx)
 498       return;
 499
 500    if (!brw->query.begin_emitted)
 501       return;
 502
 503    write_depth_count(brw, query->bo, query->last_index * 2 + 1);
 504
 505    brw->query.begin_emitted = false;
 506    query->last_index++;
 507 }
 508
 509 /**
 510  * Driver hook for glQueryCounter().
 511  *
 512  * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
 513  * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
 514  * time while the query is active.
 515  */
 516 static void
 517 brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 518 {
 519    struct brw_context *brw = brw_context(ctx);
 520    struct brw_query_object *query = (struct brw_query_object *) q;
 521
 522    assert(q->Target == GL_TIMESTAMP);
 523
 524    drm_intel_bo_unreference(query->bo);
 525    query->bo = drm_intel_bo_alloc(brw->bufmgr, "timestamp query", 4096, 4096);
 526    write_timestamp(brw, query->bo, 0);
 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    drm_intel_reg_read(brw->bufmgr, TIMESTAMP, &result);
 541
 542    /* See logic in brw_queryobj_get_results() */
 543    result = result >> 32;
 544    result *= 80;
 545    result &= (1ull << 36) - 1;
 546
 547    return result;
 548 }
 549
 550 /* Initialize query object functions used on all generations. */
 551 void brw_init_common_queryobj_functions(struct dd_function_table *functions)
 552 {
 553    functions->NewQueryObject = brw_new_query_object;
 554    functions->DeleteQuery = brw_delete_query;
 555    functions->QueryCounter = brw_query_counter;
 556    functions->GetTimestamp = brw_get_timestamp;
 557 }
 558
 559 /* Initialize Gen4/5-specific query object functions. */
 560 void gen4_init_queryobj_functions(struct dd_function_table *functions)
 561 {
 562    functions->BeginQuery = brw_begin_query;
 563    functions->EndQuery = brw_end_query;
 564    functions->CheckQuery = brw_check_query;
 565    functions->WaitQuery = brw_wait_query;
 566 }