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 workaround flushes: */
  54       if (brw->gen == 6) {
  55          brw_emit_pipe_control_flush(brw,
  56                                      PIPE_CONTROL_CS_STALL |
  57                                      PIPE_CONTROL_STALL_AT_SCOREBOARD);
  58       }
  59
  60       BEGIN_BATCH(5);
  61       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  62       OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
  63       OUT_RELOC(query_bo,
  64                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  65                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  66                 idx * sizeof(uint64_t));
  67       OUT_BATCH(0);
  68       OUT_BATCH(0);
  69       ADVANCE_BATCH();
  70    } else {
  71       BEGIN_BATCH(4);
  72       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
  73                 PIPE_CONTROL_WRITE_TIMESTAMP);
  74       OUT_RELOC(query_bo,
  75                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  76                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  77                 idx * sizeof(uint64_t));
  78       OUT_BATCH(0);
  79       OUT_BATCH(0);
  80       ADVANCE_BATCH();
  81    }
  82 }
  83
  84 /**
  85  * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
  86  */
  87 static void
  88 write_depth_count(struct brw_context *brw, drm_intel_bo *query_bo, int idx)
  89 {
  90    assert(brw->gen < 6);
  91
  92    BEGIN_BATCH(4);
  93    OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
  94              PIPE_CONTROL_DEPTH_STALL | PIPE_CONTROL_WRITE_DEPTH_COUNT);
  95    /* This object could be mapped cacheable, but we don't have an exposed
  96     * mechanism to support that.  Since it's going uncached, tell GEM that
  97     * we're writing to it.  The usual clflush should be all that's required
  98     * to pick up the results.
  99     */
 100    OUT_RELOC(query_bo,
 101              I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 102              PIPE_CONTROL_GLOBAL_GTT_WRITE |
 103              (idx * sizeof(uint64_t)));
 104    OUT_BATCH(0);
 105    OUT_BATCH(0);
 106    ADVANCE_BATCH();
 107 }
 108
 109 /**
 110  * Wait on the query object's BO and calculate the final result.
 111  */
 112 static void
 113 brw_queryobj_get_results(struct gl_context *ctx,
 114                          struct brw_query_object *query)
 115 {
 116    struct brw_context *brw = brw_context(ctx);
 117
 118    int i;
 119    uint64_t *results;
 120
 121    assert(brw->gen < 6);
 122
 123    if (query->bo == NULL)
 124       return;
 125
 126    /* If the application has requested the query result, but this batch is
 127     * still contributing to it, flush it now so the results will be present
 128     * when mapped.
 129     */
 130    if (drm_intel_bo_references(brw->batch.bo, query->bo))
 131       intel_batchbuffer_flush(brw);
 132
 133    if (unlikely(brw->perf_debug)) {
 134       if (drm_intel_bo_busy(query->bo)) {
 135          perf_debug("Stalling on the GPU waiting for a query object.\n");
 136       }
 137    }
 138
 139    drm_intel_bo_map(query->bo, false);
 140    results = query->bo->virtual;
 141    switch (query->Base.Target) {
 142    case GL_TIME_ELAPSED_EXT:
 143       /* The query BO contains the starting and ending timestamps.
 144        * Subtract the two and convert to nanoseconds.
 145        */
 146       query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
 147       break;
 148
 149    case GL_TIMESTAMP:
 150       /* The query BO contains a single timestamp value in results[0]. */
 151       query->Base.Result = 1000 * (results[0] >> 32);
 152       break;
 153
 154    case GL_SAMPLES_PASSED_ARB:
 155       /* Loop over pairs of values from the BO, which are the PS_DEPTH_COUNT
 156        * value at the start and end of the batchbuffer.  Subtract them to
 157        * get the number of fragments which passed the depth test in each
 158        * individual batch, and add those differences up to get the number
 159        * of fragments for the entire query.
 160        *
 161        * Note that query->Base.Result may already be non-zero.  We may have
 162        * run out of space in the query's BO and allocated a new one.  If so,
 163        * this function was already called to accumulate the results so far.
 164        */
 165       for (i = 0; i < query->last_index; i++) {
 166          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 167       }
 168       break;
 169
 170    case GL_ANY_SAMPLES_PASSED:
 171    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 172       /* If the starting and ending PS_DEPTH_COUNT from any of the batches
 173        * differ, then some fragments passed the depth test.
 174        */
 175       for (i = 0; i < query->last_index; i++) {
 176          if (results[i * 2 + 1] != results[i * 2]) {
 177             query->Base.Result = GL_TRUE;
 178             break;
 179          }
 180       }
 181       break;
 182
 183    default:
 184       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 185       break;
 186    }
 187    drm_intel_bo_unmap(query->bo);
 188
 189    /* Now that we've processed the data stored in the query's buffer object,
 190     * we can release it.
 191     */
 192    drm_intel_bo_unreference(query->bo);
 193    query->bo = NULL;
 194 }
 195
 196 /**
 197  * The NewQueryObject() driver hook.
 198  *
 199  * Allocates and initializes a new query object.
 200  */
 201 static struct gl_query_object *
 202 brw_new_query_object(struct gl_context *ctx, GLuint id)
 203 {
 204    struct brw_query_object *query;
 205
 206    query = calloc(1, sizeof(struct brw_query_object));
 207
 208    query->Base.Id = id;
 209    query->Base.Result = 0;
 210    query->Base.Active = false;
 211    query->Base.Ready = true;
 212
 213    return &query->Base;
 214 }
 215
 216 /**
 217  * The DeleteQuery() driver hook.
 218  */
 219 static void
 220 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 221 {
 222    struct brw_query_object *query = (struct brw_query_object *)q;
 223
 224    drm_intel_bo_unreference(query->bo);
 225    free(query);
 226 }
 227
 228 /**
 229  * Gen4-5 driver hook for glBeginQuery().
 230  *
 231  * Initializes driver structures and emits any GPU commands required to begin
 232  * recording data for the query.
 233  */
 234 static void
 235 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 236 {
 237    struct brw_context *brw = brw_context(ctx);
 238    struct brw_query_object *query = (struct brw_query_object *)q;
 239
 240    assert(brw->gen < 6);
 241
 242    switch (query->Base.Target) {
 243    case GL_TIME_ELAPSED_EXT:
 244       /* For timestamp queries, we record the starting time right away so that
 245        * we measure the full time between BeginQuery and EndQuery.  There's
 246        * some debate about whether this is the right thing to do.  Our decision
 247        * is based on the following text from the ARB_timer_query extension:
 248        *
 249        * "(5) Should the extension measure total time elapsed between the full
 250        *      completion of the BeginQuery and EndQuery commands, or just time
 251        *      spent in the graphics library?
 252        *
 253        *  RESOLVED:  This extension will measure the total time elapsed
 254        *  between the full completion of these commands.  Future extensions
 255        *  may implement a query to determine time elapsed at different stages
 256        *  of the graphics pipeline."
 257        *
 258        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 259        * we'll write a second timestamp (at index 1), and subtract the two to
 260        * obtain the time elapsed.  Notably, this includes time elapsed while
 261        * the system was doing other work, such as running other applications.
 262        */
 263       drm_intel_bo_unreference(query->bo);
 264       query->bo = drm_intel_bo_alloc(brw->bufmgr, "timer query", 4096, 4096);
 265       brw_write_timestamp(brw, query->bo, 0);
 266       break;
 267
 268    case GL_ANY_SAMPLES_PASSED:
 269    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 270    case GL_SAMPLES_PASSED_ARB:
 271       /* For occlusion queries, we delay taking an initial sample until the
 272        * first drawing occurs in this batch.  See the reasoning in the comments
 273        * for brw_emit_query_begin() below.
 274        *
 275        * Since we're starting a new query, we need to be sure to throw away
 276        * any previous occlusion query results.
 277        */
 278       drm_intel_bo_unreference(query->bo);
 279       query->bo = NULL;
 280       query->last_index = -1;
 281
 282       brw->query.obj = query;
 283
 284       /* Depth statistics on Gen4 require strange workarounds, so we try to
 285        * avoid them when necessary.  They're required for occlusion queries,
 286        * so turn them on now.
 287        */
 288       brw->stats_wm++;
 289       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 290       break;
 291
 292    default:
 293       assert(!"Unrecognized query target in brw_begin_query()");
 294       break;
 295    }
 296 }
 297
 298 /**
 299  * Gen4-5 driver hook for glEndQuery().
 300  *
 301  * Emits GPU commands to record a final query value, ending any data capturing.
 302  * However, the final result isn't necessarily available until the GPU processes
 303  * those commands.  brw_queryobj_get_results() processes the captured data to
 304  * produce the final result.
 305  */
 306 static void
 307 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 308 {
 309    struct brw_context *brw = brw_context(ctx);
 310    struct brw_query_object *query = (struct brw_query_object *)q;
 311
 312    assert(brw->gen < 6);
 313
 314    switch (query->Base.Target) {
 315    case GL_TIME_ELAPSED_EXT:
 316       /* Write the final timestamp. */
 317       brw_write_timestamp(brw, query->bo, 1);
 318       break;
 319
 320    case GL_ANY_SAMPLES_PASSED:
 321    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 322    case GL_SAMPLES_PASSED_ARB:
 323
 324       /* No query->bo means that EndQuery was called after BeginQuery with no
 325        * intervening drawing. Rather than doing nothing at all here in this
 326        * case, we emit the query_begin and query_end state to the
 327        * hardware. This is to guarantee that waiting on the result of this
 328        * empty state will cause all previous queries to complete at all, as
 329        * required by the specification:
 330        *
 331        *        It must always be true that if any query object
 332        *        returns a result available of TRUE, all queries of the
 333        *        same type issued prior to that query must also return
 334        *        TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
 335        */
 336       if (!query->bo) {
 337          brw_emit_query_begin(brw);
 338       }
 339
 340       assert(query->bo);
 341
 342       brw_emit_query_end(brw);
 343
 344       brw->query.obj = NULL;
 345
 346       brw->stats_wm--;
 347       brw->state.dirty.brw |= BRW_NEW_STATS_WM;
 348       break;
 349
 350    default:
 351       assert(!"Unrecognized query target in brw_end_query()");
 352       break;
 353    }
 354 }
 355
 356 /**
 357  * The Gen4-5 WaitQuery() driver hook.
 358  *
 359  * Wait for a query result to become available and return it.  This is the
 360  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 361  */
 362 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 363 {
 364    struct brw_query_object *query = (struct brw_query_object *)q;
 365
 366    assert(brw_context(ctx)->gen < 6);
 367
 368    brw_queryobj_get_results(ctx, query);
 369    query->Base.Ready = true;
 370 }
 371
 372 /**
 373  * The Gen4-5 CheckQuery() driver hook.
 374  *
 375  * Checks whether a query result is ready yet.  If not, flushes.
 376  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 377  */
 378 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 379 {
 380    struct brw_context *brw = brw_context(ctx);
 381    struct brw_query_object *query = (struct brw_query_object *)q;
 382
 383    assert(brw->gen < 6);
 384
 385    /* From the GL_ARB_occlusion_query spec:
 386     *
 387     *     "Instead of allowing for an infinite loop, performing a
 388     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 389     *      not ready yet on the first time it is queried.  This ensures that
 390     *      the async query will return true in finite time.
 391     */
 392    if (query->bo && drm_intel_bo_references(brw->batch.bo, query->bo))
 393       intel_batchbuffer_flush(brw);
 394
 395    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 396       brw_queryobj_get_results(ctx, query);
 397       query->Base.Ready = true;
 398    }
 399 }
 400
 401 /**
 402  * Ensure there query's BO has enough space to store a new pair of values.
 403  *
 404  * If not, gather the existing BO's results and create a new buffer of the
 405  * same size.
 406  */
 407 static void
 408 ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query)
 409 {
 410    struct brw_context *brw = brw_context(ctx);
 411
 412    assert(brw->gen < 6);
 413
 414    if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 415
 416       if (query->bo != NULL) {
 417          /* The old query BO did not have enough space, so we allocated a new
 418           * one.  Gather the results so far (adding up the differences) and
 419           * release the old BO.
 420           */
 421          brw_queryobj_get_results(ctx, query);
 422       }
 423
 424       query->bo = drm_intel_bo_alloc(brw->bufmgr, "query", 4096, 1);
 425       query->last_index = 0;
 426    }
 427 }
 428
 429 /**
 430  * Record the PS_DEPTH_COUNT value (for occlusion queries) just before
 431  * primitive drawing.
 432  *
 433  * In a pre-hardware context world, the single PS_DEPTH_COUNT register is
 434  * shared among all applications using the GPU.  However, our query value
 435  * needs to only include fragments generated by our application/GL context.
 436  *
 437  * To accommodate this, we record PS_DEPTH_COUNT at the start and end of
 438  * each batchbuffer (technically, the first primitive drawn and flush time).
 439  * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT
 440  * caused by a batchbuffer.  Since there is no preemption inside batches,
 441  * this is guaranteed to only measure the effects of our current application.
 442  *
 443  * Adding each of these differences (in case drawing is done over many batches)
 444  * produces the final expected value.
 445  *
 446  * In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
 447  * as part of the context state, so this is unnecessary, and skipped.
 448  */
 449 void
 450 brw_emit_query_begin(struct brw_context *brw)
 451 {
 452    struct gl_context *ctx = &brw->ctx;
 453    struct brw_query_object *query = brw->query.obj;
 454
 455    if (brw->hw_ctx)
 456       return;
 457
 458    /* Skip if we're not doing any queries, or we've already recorded the
 459     * initial query value for this batchbuffer.
 460     */
 461    if (!query || brw->query.begin_emitted)
 462       return;
 463
 464    ensure_bo_has_space(ctx, query);
 465
 466    write_depth_count(brw, query->bo, query->last_index * 2);
 467
 468    brw->query.begin_emitted = true;
 469 }
 470
 471 /**
 472  * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
 473  * (for non-hardware context platforms).
 474  *
 475  * See the explanation in brw_emit_query_begin().
 476  */
 477 void
 478 brw_emit_query_end(struct brw_context *brw)
 479 {
 480    struct brw_query_object *query = brw->query.obj;
 481
 482    if (brw->hw_ctx)
 483       return;
 484
 485    if (!brw->query.begin_emitted)
 486       return;
 487
 488    write_depth_count(brw, query->bo, query->last_index * 2 + 1);
 489
 490    brw->query.begin_emitted = false;
 491    query->last_index++;
 492 }
 493
 494 /**
 495  * Driver hook for glQueryCounter().
 496  *
 497  * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
 498  * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
 499  * time while the query is active.
 500  */
 501 static void
 502 brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 503 {
 504    struct brw_context *brw = brw_context(ctx);
 505    struct brw_query_object *query = (struct brw_query_object *) q;
 506
 507    assert(q->Target == GL_TIMESTAMP);
 508
 509    drm_intel_bo_unreference(query->bo);
 510    query->bo = drm_intel_bo_alloc(brw->bufmgr, "timestamp query", 4096, 4096);
 511    brw_write_timestamp(brw, query->bo, 0);
 512 }
 513
 514 /**
 515  * Read the TIMESTAMP register immediately (in a non-pipelined fashion).
 516  *
 517  * This is used to implement the GetTimestamp() driver hook.
 518  */
 519 static uint64_t
 520 brw_get_timestamp(struct gl_context *ctx)
 521 {
 522    struct brw_context *brw = brw_context(ctx);
 523    uint64_t result = 0;
 524
 525    drm_intel_reg_read(brw->bufmgr, TIMESTAMP, &result);
 526
 527    /* See logic in brw_queryobj_get_results() */
 528    result = result >> 32;
 529    result *= 80;
 530    result &= (1ull << 36) - 1;
 531
 532    return result;
 533 }
 534
 535 /* Initialize query object functions used on all generations. */
 536 void brw_init_common_queryobj_functions(struct dd_function_table *functions)
 537 {
 538    functions->NewQueryObject = brw_new_query_object;
 539    functions->DeleteQuery = brw_delete_query;
 540    functions->QueryCounter = brw_query_counter;
 541    functions->GetTimestamp = brw_get_timestamp;
 542 }
 543
 544 /* Initialize Gen4/5-specific query object functions. */
 545 void gen4_init_queryobj_functions(struct dd_function_table *functions)
 546 {
 547    functions->BeginQuery = brw_begin_query;
 548    functions->EndQuery = brw_end_query;
 549    functions->CheckQuery = brw_check_query;
 550    functions->WaitQuery = brw_wait_query;
 551 }