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