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 void
  50 brw_write_timestamp(struct brw_context *brw, drm_intel_bo *query_bo, int idx)
  51 {
  52    if (brw->gen == 6) {
  53       /* Emit Sandybridge workaround flush: */
  54       brw_emit_pipe_control_flush(brw,
  55                                   PIPE_CONTROL_CS_STALL |
  56                                   PIPE_CONTROL_STALL_AT_SCOREBOARD);
  57    }
  58
  59    uint32_t flags = PIPE_CONTROL_WRITE_TIMESTAMP;
  60
  61    if (brw->gen == 9 && brw->gt == 4)
  62       flags |= PIPE_CONTROL_CS_STALL;
  63
  64    brw_emit_pipe_control_write(brw, flags,
  65                                query_bo, idx * sizeof(uint64_t), 0, 0);
  66 }
  67
  68 /**
  69  * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
  70  */
  71 void
  72 brw_write_depth_count(struct brw_context *brw, drm_intel_bo *query_bo, int idx)
  73 {
  74    uint32_t flags = PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_DEPTH_STALL;
  75
  76    if (brw->gen == 9 && brw->gt == 4)
  77       flags |= PIPE_CONTROL_CS_STALL;
  78
  79    brw_emit_pipe_control_write(brw, flags,
  80                                query_bo, idx * sizeof(uint64_t),
  81                                0, 0);
  82 }
  83
  84 /**
  85  * Wait on the query object's BO and calculate the final result.
  86  */
  87 static void
  88 brw_queryobj_get_results(struct gl_context *ctx,
  89                          struct brw_query_object *query)
  90 {
  91    struct brw_context *brw = brw_context(ctx);
  92
  93    int i;
  94    uint64_t *results;
  95
  96    assert(brw->gen < 6);
  97
  98    if (query->bo == NULL)
  99       return;
 100
 101    /* If the application has requested the query result, but this batch is
 102     * still contributing to it, flush it now so the results will be present
 103     * when mapped.
 104     */
 105    if (drm_intel_bo_references(brw->batch.bo, query->bo))
 106       intel_batchbuffer_flush(brw);
 107
 108    if (unlikely(brw->perf_debug)) {
 109       if (drm_intel_bo_busy(query->bo)) {
 110          perf_debug("Stalling on the GPU waiting for a query object.\n");
 111       }
 112    }
 113
 114    drm_intel_bo_map(query->bo, false);
 115    results = query->bo->virtual;
 116    switch (query->Base.Target) {
 117    case GL_TIME_ELAPSED_EXT:
 118       /* The query BO contains the starting and ending timestamps.
 119        * Subtract the two and convert to nanoseconds.
 120        */
 121       query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
 122       break;
 123
 124    case GL_TIMESTAMP:
 125       /* The query BO contains a single timestamp value in results[0]. */
 126       query->Base.Result = 1000 * (results[0] >> 32);
 127       break;
 128
 129    case GL_SAMPLES_PASSED_ARB:
 130       /* Loop over pairs of values from the BO, which are the PS_DEPTH_COUNT
 131        * value at the start and end of the batchbuffer.  Subtract them to
 132        * get the number of fragments which passed the depth test in each
 133        * individual batch, and add those differences up to get the number
 134        * of fragments for the entire query.
 135        *
 136        * Note that query->Base.Result may already be non-zero.  We may have
 137        * run out of space in the query's BO and allocated a new one.  If so,
 138        * this function was already called to accumulate the results so far.
 139        */
 140       for (i = 0; i < query->last_index; i++) {
 141          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 142       }
 143       break;
 144
 145    case GL_ANY_SAMPLES_PASSED:
 146    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 147       /* If the starting and ending PS_DEPTH_COUNT from any of the batches
 148        * differ, then some fragments passed the depth test.
 149        */
 150       for (i = 0; i < query->last_index; i++) {
 151          if (results[i * 2 + 1] != results[i * 2]) {
 152             query->Base.Result = GL_TRUE;
 153             break;
 154          }
 155       }
 156       break;
 157
 158    default:
 159       unreachable("Unrecognized query target in brw_queryobj_get_results()");
 160    }
 161    drm_intel_bo_unmap(query->bo);
 162
 163    /* Now that we've processed the data stored in the query's buffer object,
 164     * we can release it.
 165     */
 166    drm_intel_bo_unreference(query->bo);
 167    query->bo = NULL;
 168 }
 169
 170 /**
 171  * The NewQueryObject() driver hook.
 172  *
 173  * Allocates and initializes a new query object.
 174  */
 175 static struct gl_query_object *
 176 brw_new_query_object(struct gl_context *ctx, GLuint id)
 177 {
 178    struct brw_query_object *query;
 179
 180    query = calloc(1, sizeof(struct brw_query_object));
 181
 182    query->Base.Id = id;
 183    query->Base.Result = 0;
 184    query->Base.Active = false;
 185    query->Base.Ready = true;
 186
 187    return &query->Base;
 188 }
 189
 190 /**
 191  * The DeleteQuery() driver hook.
 192  */
 193 static void
 194 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 195 {
 196    struct brw_query_object *query = (struct brw_query_object *)q;
 197
 198    drm_intel_bo_unreference(query->bo);
 199    free(query);
 200 }
 201
 202 /**
 203  * Gen4-5 driver hook for glBeginQuery().
 204  *
 205  * Initializes driver structures and emits any GPU commands required to begin
 206  * recording data for the query.
 207  */
 208 static void
 209 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 210 {
 211    struct brw_context *brw = brw_context(ctx);
 212    struct brw_query_object *query = (struct brw_query_object *)q;
 213
 214    assert(brw->gen < 6);
 215
 216    switch (query->Base.Target) {
 217    case GL_TIME_ELAPSED_EXT:
 218       /* For timestamp queries, we record the starting time right away so that
 219        * we measure the full time between BeginQuery and EndQuery.  There's
 220        * some debate about whether this is the right thing to do.  Our decision
 221        * is based on the following text from the ARB_timer_query extension:
 222        *
 223        * "(5) Should the extension measure total time elapsed between the full
 224        *      completion of the BeginQuery and EndQuery commands, or just time
 225        *      spent in the graphics library?
 226        *
 227        *  RESOLVED:  This extension will measure the total time elapsed
 228        *  between the full completion of these commands.  Future extensions
 229        *  may implement a query to determine time elapsed at different stages
 230        *  of the graphics pipeline."
 231        *
 232        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 233        * we'll write a second timestamp (at index 1), and subtract the two to
 234        * obtain the time elapsed.  Notably, this includes time elapsed while
 235        * the system was doing other work, such as running other applications.
 236        */
 237       drm_intel_bo_unreference(query->bo);
 238       query->bo = drm_intel_bo_alloc(brw->bufmgr, "timer query", 4096, 4096);
 239       brw_write_timestamp(brw, query->bo, 0);
 240       break;
 241
 242    case GL_ANY_SAMPLES_PASSED:
 243    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 244    case GL_SAMPLES_PASSED_ARB:
 245       /* For occlusion queries, we delay taking an initial sample until the
 246        * first drawing occurs in this batch.  See the reasoning in the comments
 247        * for brw_emit_query_begin() below.
 248        *
 249        * Since we're starting a new query, we need to be sure to throw away
 250        * any previous occlusion query results.
 251        */
 252       drm_intel_bo_unreference(query->bo);
 253       query->bo = NULL;
 254       query->last_index = -1;
 255
 256       brw->query.obj = query;
 257
 258       /* Depth statistics on Gen4 require strange workarounds, so we try to
 259        * avoid them when necessary.  They're required for occlusion queries,
 260        * so turn them on now.
 261        */
 262       brw->stats_wm++;
 263       brw->ctx.NewDriverState |= BRW_NEW_STATS_WM;
 264       break;
 265
 266    default:
 267       unreachable("Unrecognized query target in brw_begin_query()");
 268    }
 269 }
 270
 271 /**
 272  * Gen4-5 driver hook for glEndQuery().
 273  *
 274  * Emits GPU commands to record a final query value, ending any data capturing.
 275  * However, the final result isn't necessarily available until the GPU processes
 276  * those commands.  brw_queryobj_get_results() processes the captured data to
 277  * produce the final result.
 278  */
 279 static void
 280 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 281 {
 282    struct brw_context *brw = brw_context(ctx);
 283    struct brw_query_object *query = (struct brw_query_object *)q;
 284
 285    assert(brw->gen < 6);
 286
 287    switch (query->Base.Target) {
 288    case GL_TIME_ELAPSED_EXT:
 289       /* Write the final timestamp. */
 290       brw_write_timestamp(brw, query->bo, 1);
 291       break;
 292
 293    case GL_ANY_SAMPLES_PASSED:
 294    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 295    case GL_SAMPLES_PASSED_ARB:
 296
 297       /* No query->bo means that EndQuery was called after BeginQuery with no
 298        * intervening drawing. Rather than doing nothing at all here in this
 299        * case, we emit the query_begin and query_end state to the
 300        * hardware. This is to guarantee that waiting on the result of this
 301        * empty state will cause all previous queries to complete at all, as
 302        * required by the specification:
 303        *
 304        *        It must always be true that if any query object
 305        *        returns a result available of TRUE, all queries of the
 306        *        same type issued prior to that query must also return
 307        *        TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
 308        */
 309       if (!query->bo) {
 310          brw_emit_query_begin(brw);
 311       }
 312
 313       assert(query->bo);
 314
 315       brw_emit_query_end(brw);
 316
 317       brw->query.obj = NULL;
 318
 319       brw->stats_wm--;
 320       brw->ctx.NewDriverState |= BRW_NEW_STATS_WM;
 321       break;
 322
 323    default:
 324       unreachable("Unrecognized query target in brw_end_query()");
 325    }
 326 }
 327
 328 /**
 329  * The Gen4-5 WaitQuery() driver hook.
 330  *
 331  * Wait for a query result to become available and return it.  This is the
 332  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 333  */
 334 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 335 {
 336    struct brw_query_object *query = (struct brw_query_object *)q;
 337
 338    assert(brw_context(ctx)->gen < 6);
 339
 340    brw_queryobj_get_results(ctx, query);
 341    query->Base.Ready = true;
 342 }
 343
 344 /**
 345  * The Gen4-5 CheckQuery() driver hook.
 346  *
 347  * Checks whether a query result is ready yet.  If not, flushes.
 348  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 349  */
 350 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 351 {
 352    struct brw_context *brw = brw_context(ctx);
 353    struct brw_query_object *query = (struct brw_query_object *)q;
 354
 355    assert(brw->gen < 6);
 356
 357    /* From the GL_ARB_occlusion_query spec:
 358     *
 359     *     "Instead of allowing for an infinite loop, performing a
 360     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 361     *      not ready yet on the first time it is queried.  This ensures that
 362     *      the async query will return true in finite time.
 363     */
 364    if (query->bo && drm_intel_bo_references(brw->batch.bo, query->bo))
 365       intel_batchbuffer_flush(brw);
 366
 367    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 368       brw_queryobj_get_results(ctx, query);
 369       query->Base.Ready = true;
 370    }
 371 }
 372
 373 /**
 374  * Ensure there query's BO has enough space to store a new pair of values.
 375  *
 376  * If not, gather the existing BO's results and create a new buffer of the
 377  * same size.
 378  */
 379 static void
 380 ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query)
 381 {
 382    struct brw_context *brw = brw_context(ctx);
 383
 384    assert(brw->gen < 6);
 385
 386    if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 387
 388       if (query->bo != NULL) {
 389          /* The old query BO did not have enough space, so we allocated a new
 390           * one.  Gather the results so far (adding up the differences) and
 391           * release the old BO.
 392           */
 393          brw_queryobj_get_results(ctx, query);
 394       }
 395
 396       query->bo = drm_intel_bo_alloc(brw->bufmgr, "query", 4096, 1);
 397       query->last_index = 0;
 398    }
 399 }
 400
 401 /**
 402  * Record the PS_DEPTH_COUNT value (for occlusion queries) just before
 403  * primitive drawing.
 404  *
 405  * In a pre-hardware context world, the single PS_DEPTH_COUNT register is
 406  * shared among all applications using the GPU.  However, our query value
 407  * needs to only include fragments generated by our application/GL context.
 408  *
 409  * To accommodate this, we record PS_DEPTH_COUNT at the start and end of
 410  * each batchbuffer (technically, the first primitive drawn and flush time).
 411  * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT
 412  * caused by a batchbuffer.  Since there is no preemption inside batches,
 413  * this is guaranteed to only measure the effects of our current application.
 414  *
 415  * Adding each of these differences (in case drawing is done over many batches)
 416  * produces the final expected value.
 417  *
 418  * In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
 419  * as part of the context state, so this is unnecessary, and skipped.
 420  */
 421 void
 422 brw_emit_query_begin(struct brw_context *brw)
 423 {
 424    struct gl_context *ctx = &brw->ctx;
 425    struct brw_query_object *query = brw->query.obj;
 426
 427    if (brw->hw_ctx)
 428       return;
 429
 430    /* Skip if we're not doing any queries, or we've already recorded the
 431     * initial query value for this batchbuffer.
 432     */
 433    if (!query || brw->query.begin_emitted)
 434       return;
 435
 436    ensure_bo_has_space(ctx, query);
 437
 438    brw_write_depth_count(brw, query->bo, query->last_index * 2);
 439
 440    brw->query.begin_emitted = true;
 441 }
 442
 443 /**
 444  * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
 445  * (for non-hardware context platforms).
 446  *
 447  * See the explanation in brw_emit_query_begin().
 448  */
 449 void
 450 brw_emit_query_end(struct brw_context *brw)
 451 {
 452    struct brw_query_object *query = brw->query.obj;
 453
 454    if (brw->hw_ctx)
 455       return;
 456
 457    if (!brw->query.begin_emitted)
 458       return;
 459
 460    brw_write_depth_count(brw, query->bo, query->last_index * 2 + 1);
 461
 462    brw->query.begin_emitted = false;
 463    query->last_index++;
 464 }
 465
 466 /**
 467  * Driver hook for glQueryCounter().
 468  *
 469  * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
 470  * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
 471  * time while the query is active.
 472  */
 473 void
 474 brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 475 {
 476    struct brw_context *brw = brw_context(ctx);
 477    struct brw_query_object *query = (struct brw_query_object *) q;
 478
 479    assert(q->Target == GL_TIMESTAMP);
 480
 481    drm_intel_bo_unreference(query->bo);
 482    query->bo = drm_intel_bo_alloc(brw->bufmgr, "timestamp query", 4096, 4096);
 483    brw_write_timestamp(brw, query->bo, 0);
 484
 485    query->flushed = false;
 486 }
 487
 488 /**
 489  * Read the TIMESTAMP register immediately (in a non-pipelined fashion).
 490  *
 491  * This is used to implement the GetTimestamp() driver hook.
 492  */
 493 static uint64_t
 494 brw_get_timestamp(struct gl_context *ctx)
 495 {
 496    struct brw_context *brw = brw_context(ctx);
 497    uint64_t result = 0;
 498
 499    switch (brw->intelScreen->hw_has_timestamp) {
 500    case 3: /* New kernel, always full 36bit accuracy */
 501       drm_intel_reg_read(brw->bufmgr, TIMESTAMP | 1, &result);
 502       break;
 503    case 2: /* 64bit kernel, result is left-shifted by 32bits, losing 4bits */
 504       drm_intel_reg_read(brw->bufmgr, TIMESTAMP, &result);
 505       result = result >> 32;
 506       break;
 507    case 1: /* 32bit kernel, result is 36bit wide but may be inaccurate! */
 508       drm_intel_reg_read(brw->bufmgr, TIMESTAMP, &result);
 509       break;
 510    }
 511
 512    /* See logic in brw_queryobj_get_results() */
 513    result *= 80;
 514    result &= (1ull << 36) - 1;
 515    return result;
 516 }
 517
 518 /**
 519  * Is this type of query written by PIPE_CONTROL?
 520  */
 521 bool
 522 brw_is_query_pipelined(struct brw_query_object *query)
 523 {
 524    switch (query->Base.Target) {
 525    case GL_TIMESTAMP:
 526    case GL_TIME_ELAPSED:
 527    case GL_ANY_SAMPLES_PASSED:
 528    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 529    case GL_SAMPLES_PASSED_ARB:
 530       return true;
 531
 532    case GL_PRIMITIVES_GENERATED:
 533    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 534    case GL_VERTICES_SUBMITTED_ARB:
 535    case GL_PRIMITIVES_SUBMITTED_ARB:
 536    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 537    case GL_GEOMETRY_SHADER_INVOCATIONS:
 538    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 539    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 540    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 541    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 542    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 543    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 544    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 545       return false;
 546
 547    default:
 548       unreachable("Unrecognized query target in is_query_pipelined()");
 549    }
 550 }
 551
 552 /* Initialize query object functions used on all generations. */
 553 void brw_init_common_queryobj_functions(struct dd_function_table *functions)
 554 {
 555    functions->NewQueryObject = brw_new_query_object;
 556    functions->DeleteQuery = brw_delete_query;
 557    functions->GetTimestamp = brw_get_timestamp;
 558 }
 559
 560 /* Initialize Gen4/5-specific query object functions. */
 561 void gen4_init_queryobj_functions(struct dd_function_table *functions)
 562 {
 563    functions->BeginQuery = brw_begin_query;
 564    functions->EndQuery = brw_end_query;
 565    functions->CheckQuery = brw_check_query;
 566    functions->WaitQuery = brw_wait_query;
 567    functions->QueryCounter = brw_query_counter;
 568 }