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