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