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