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  * To avoid getting samples from another context's rendering in our results,
  39  * we capture the counts at the start and end of every batchbuffer while the
  40  * query is active, and sum up the differences.  (We should do so for
  41  * GL_TIME_ELAPSED as well, but don't).
  42  */
  43 #include "main/imports.h"
  44
  45 #include "brw_context.h"
  46 #include "brw_defines.h"
  47 #include "brw_state.h"
  48 #include "intel_batchbuffer.h"
  49 #include "intel_reg.h"
  50
  51 static void
  52 write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  53 {
  54    if (intel->gen >= 6) {
  55       /* Emit workaround flushes: */
  56       if (intel->gen == 6) {
  57          /* The timestamp write below is a non-zero post-sync op, which on
  58           * Gen6 necessitates a CS stall.  CS stalls need stall at scoreboard
  59           * set.  See the comments for intel_emit_post_sync_nonzero_flush().
  60           */
  61          BEGIN_BATCH(4);
  62          OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
  63          OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
  64          OUT_BATCH(0);
  65          OUT_BATCH(0);
  66          ADVANCE_BATCH();
  67       }
  68
  69       BEGIN_BATCH(5);
  70       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  71       OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
  72       OUT_RELOC(query_bo,
  73                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  74                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  75                 idx * sizeof(uint64_t));
  76       OUT_BATCH(0);
  77       OUT_BATCH(0);
  78       ADVANCE_BATCH();
  79    } else {
  80       BEGIN_BATCH(4);
  81       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
  82                 PIPE_CONTROL_WRITE_TIMESTAMP);
  83       OUT_RELOC(query_bo,
  84                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  85                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  86                 idx * sizeof(uint64_t));
  87       OUT_BATCH(0);
  88       OUT_BATCH(0);
  89       ADVANCE_BATCH();
  90    }
  91 }
  92
  93 static void
  94 write_depth_count(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  95 {
  96    if (intel->gen >= 6) {
  97       /* Emit Sandybridge workaround flush: */
  98       if (intel->gen == 6)
  99          intel_emit_post_sync_nonzero_flush(intel);
 100
 101       BEGIN_BATCH(5);
 102       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
 103       OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
 104                 PIPE_CONTROL_WRITE_DEPTH_COUNT);
 105       OUT_RELOC(query_bo,
 106                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 107                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
 108                 (idx * sizeof(uint64_t)));
 109       OUT_BATCH(0);
 110       OUT_BATCH(0);
 111       ADVANCE_BATCH();
 112    } else {
 113       BEGIN_BATCH(4);
 114       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
 115                 PIPE_CONTROL_DEPTH_STALL |
 116                 PIPE_CONTROL_WRITE_DEPTH_COUNT);
 117       /* This object could be mapped cacheable, but we don't have an exposed
 118        * mechanism to support that.  Since it's going uncached, tell GEM that
 119        * we're writing to it.  The usual clflush should be all that's required
 120        * to pick up the results.
 121        */
 122       OUT_RELOC(query_bo,
 123                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 124                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
 125                 (idx * sizeof(uint64_t)));
 126       OUT_BATCH(0);
 127       OUT_BATCH(0);
 128       ADVANCE_BATCH();
 129    }
 130 }
 131
 132 /** Waits on the query object's BO and totals the results for this query */
 133 static void
 134 brw_queryobj_get_results(struct gl_context *ctx,
 135                          struct brw_query_object *query)
 136 {
 137    struct intel_context *intel = intel_context(ctx);
 138
 139    int i;
 140    uint64_t *results;
 141
 142    if (query->bo == NULL)
 143       return;
 144
 145    if (drm_intel_bo_references(intel->batch.bo, query->bo))
 146       intel_batchbuffer_flush(intel);
 147
 148    if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
 149       if (drm_intel_bo_busy(query->bo)) {
 150          perf_debug("Stalling on the GPU waiting for a query object.\n");
 151       }
 152    }
 153
 154    drm_intel_bo_map(query->bo, false);
 155    results = query->bo->virtual;
 156    switch (query->Base.Target) {
 157    case GL_TIME_ELAPSED_EXT:
 158       if (intel->gen >= 6)
 159          query->Base.Result += 80 * (results[1] - results[0]);
 160       else
 161          query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
 162       break;
 163
 164    case GL_TIMESTAMP:
 165       if (intel->gen >= 6) {
 166          /* Our timer is a clock that increments every 80ns (regardless of
 167           * other clock scaling in the system).  The timestamp register we can
 168           * read for glGetTimestamp() masks out the top 32 bits, so we do that
 169           * here too to let the two counters be compared against each other.
 170           *
 171           * If we just multiplied that 32 bits of data by 80, it would roll
 172           * over at a non-power-of-two, so an application couldn't use
 173           * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
 174           * report 36 bits and truncate at that (rolling over 5 times as often
 175           * as the HW counter), and when the 32-bit counter rolls over, it
 176           * happens to also be at a rollover in the reported value from near
 177           * (1<<36) to 0.
 178           *
 179           * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
 180           * rolls over every ~69 seconds.
 181           */
 182          query->Base.Result = 80 * (results[0] & 0xffffffff);
 183          query->Base.Result &= (1ull << 36) - 1;
 184       } else {
 185          query->Base.Result = 1000 * (results[0] >> 32);
 186       }
 187       break;
 188
 189    case GL_SAMPLES_PASSED_ARB:
 190       /* Map and count the pixels from the current query BO */
 191       for (i = query->first_index; i <= query->last_index; i++) {
 192          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 193       }
 194       break;
 195
 196    case GL_ANY_SAMPLES_PASSED:
 197    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 198       /* Set true if any of the sub-queries passed. */
 199       for (i = query->first_index; i <= query->last_index; i++) {
 200          if (results[i * 2 + 1] != results[i * 2]) {
 201             query->Base.Result = GL_TRUE;
 202             break;
 203          }
 204       }
 205       break;
 206
 207    case GL_PRIMITIVES_GENERATED:
 208    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 209       /* We don't actually query the hardware for this value, so query->bo
 210        * should always be NULL and execution should never reach here.
 211        */
 212       assert(!"Unreachable");
 213       break;
 214
 215    default:
 216       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 217       break;
 218    }
 219    drm_intel_bo_unmap(query->bo);
 220
 221    drm_intel_bo_unreference(query->bo);
 222    query->bo = NULL;
 223 }
 224
 225 static struct gl_query_object *
 226 brw_new_query_object(struct gl_context *ctx, GLuint id)
 227 {
 228    struct brw_query_object *query;
 229
 230    query = calloc(1, sizeof(struct brw_query_object));
 231
 232    query->Base.Id = id;
 233    query->Base.Result = 0;
 234    query->Base.Active = false;
 235    query->Base.Ready = true;
 236
 237    return &query->Base;
 238 }
 239
 240 static void
 241 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 242 {
 243    struct brw_query_object *query = (struct brw_query_object *)q;
 244
 245    drm_intel_bo_unreference(query->bo);
 246    free(query);
 247 }
 248
 249 static void
 250 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 251 {
 252    struct brw_context *brw = brw_context(ctx);
 253    struct intel_context *intel = intel_context(ctx);
 254    struct brw_query_object *query = (struct brw_query_object *)q;
 255
 256    switch (query->Base.Target) {
 257    case GL_TIME_ELAPSED_EXT:
 258       drm_intel_bo_unreference(query->bo);
 259       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
 260       write_timestamp(intel, query->bo, 0);
 261       break;
 262
 263    case GL_ANY_SAMPLES_PASSED:
 264    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 265    case GL_SAMPLES_PASSED_ARB:
 266       /* Reset our driver's tracking of query state. */
 267       drm_intel_bo_unreference(query->bo);
 268       query->bo = NULL;
 269       query->first_index = -1;
 270       query->last_index = -1;
 271
 272       brw->query.obj = query;
 273       intel->stats_wm++;
 274       break;
 275
 276    case GL_PRIMITIVES_GENERATED:
 277       /* We don't actually query the hardware for this value; we keep track of
 278        * it a software counter.  So just reset the counter.
 279        */
 280       brw->sol.primitives_generated = 0;
 281       brw->sol.counting_primitives_generated = true;
 282       break;
 283
 284    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 285       /* We don't actually query the hardware for this value; we keep track of
 286        * it a software counter.  So just reset the counter.
 287        */
 288       brw->sol.primitives_written = 0;
 289       brw->sol.counting_primitives_written = true;
 290       break;
 291
 292    default:
 293       assert(!"Unrecognized query target in brw_begin_query()");
 294       break;
 295    }
 296 }
 297
 298 /**
 299  * Begin the ARB_occlusion_query query on a query object.
 300  */
 301 static void
 302 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 303 {
 304    struct brw_context *brw = brw_context(ctx);
 305    struct intel_context *intel = intel_context(ctx);
 306    struct brw_query_object *query = (struct brw_query_object *)q;
 307
 308    switch (query->Base.Target) {
 309    case GL_TIME_ELAPSED_EXT:
 310       write_timestamp(intel, query->bo, 1);
 311       break;
 312
 313    case GL_ANY_SAMPLES_PASSED:
 314    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 315    case GL_SAMPLES_PASSED_ARB:
 316
 317       /* No query->bo means that EndQuery was called after BeginQuery with no
 318        * intervening drawing. Rather than doing nothing at all here in this
 319        * case, we emit the query_begin and query_end state to the
 320        * hardware. This is to guarantee that waiting on the result of this
 321        * empty state will cause all previous queries to complete at all, as
 322        * required by the specification:
 323        *
 324        *        It must always be true that if any query object
 325        *        returns a result available of TRUE, all queries of the
 326        *        same type issued prior to that query must also return
 327        *        TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1]
 328        */
 329       if (!query->bo) {
 330          brw_emit_query_begin(brw);
 331       }
 332
 333       if (query->bo) {
 334          brw_emit_query_end(brw);
 335
 336          drm_intel_bo_unreference(brw->query.bo);
 337          brw->query.bo = NULL;
 338       }
 339
 340       brw->query.obj = NULL;
 341
 342       intel->stats_wm--;
 343       break;
 344
 345    case GL_PRIMITIVES_GENERATED:
 346       /* We don't actually query the hardware for this value; we keep track of
 347        * it in a software counter.  So just read the counter and store it in
 348        * the query object.
 349        */
 350       query->Base.Result = brw->sol.primitives_generated;
 351       brw->sol.counting_primitives_generated = false;
 352
 353       /* And set brw->query.obj to NULL so that this query won't try to wait
 354        * for any rendering to complete.
 355        */
 356       query->bo = NULL;
 357       break;
 358
 359    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 360       /* We don't actually query the hardware for this value; we keep track of
 361        * it in a software counter.  So just read the counter and store it in
 362        * the query object.
 363        */
 364       query->Base.Result = brw->sol.primitives_written;
 365       brw->sol.counting_primitives_written = false;
 366
 367       /* And set brw->query.obj to NULL so that this query won't try to wait
 368        * for any rendering to complete.
 369        */
 370       query->bo = NULL;
 371       break;
 372
 373    default:
 374       assert(!"Unrecognized query target in brw_end_query()");
 375       break;
 376    }
 377 }
 378
 379 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 380 {
 381    struct brw_query_object *query = (struct brw_query_object *)q;
 382
 383    brw_queryobj_get_results(ctx, query);
 384    query->Base.Ready = true;
 385 }
 386
 387 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 388 {
 389    struct intel_context *intel = intel_context(ctx);
 390    struct brw_query_object *query = (struct brw_query_object *)q;
 391
 392    /* From the GL_ARB_occlusion_query spec:
 393     *
 394     *     "Instead of allowing for an infinite loop, performing a
 395     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 396     *      not ready yet on the first time it is queried.  This ensures that
 397     *      the async query will return true in finite time.
 398     */
 399    if (query->bo && drm_intel_bo_references(intel->batch.bo, query->bo))
 400       intel_batchbuffer_flush(intel);
 401
 402    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 403       brw_queryobj_get_results(ctx, query);
 404       query->Base.Ready = true;
 405    }
 406 }
 407
 408 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
 409 void
 410 brw_emit_query_begin(struct brw_context *brw)
 411 {
 412    struct intel_context *intel = &brw->intel;
 413    struct gl_context *ctx = &intel->ctx;
 414    struct brw_query_object *query = brw->query.obj;
 415
 416    /* Skip if we're not doing any queries, or we've emitted the start. */
 417    if (!query || brw->query.begin_emitted)
 418       return;
 419
 420    /* Get a new query BO if we're going to need it. */
 421    if (brw->query.bo == NULL ||
 422        brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 423       drm_intel_bo_unreference(brw->query.bo);
 424       brw->query.bo = NULL;
 425
 426       brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
 427
 428       /* clear target buffer */
 429       drm_intel_bo_map(brw->query.bo, true);
 430       memset((char *)brw->query.bo->virtual, 0, 4096);
 431       drm_intel_bo_unmap(brw->query.bo);
 432
 433       brw->query.index = 0;
 434    }
 435
 436    write_depth_count(intel, brw->query.bo, brw->query.index * 2);
 437
 438    if (query->bo != brw->query.bo) {
 439       if (query->bo != NULL)
 440          brw_queryobj_get_results(ctx, query);
 441       drm_intel_bo_reference(brw->query.bo);
 442       query->bo = brw->query.bo;
 443       query->first_index = brw->query.index;
 444    }
 445    query->last_index = brw->query.index;
 446    brw->query.begin_emitted = true;
 447 }
 448
 449 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
 450 void
 451 brw_emit_query_end(struct brw_context *brw)
 452 {
 453    struct intel_context *intel = &brw->intel;
 454
 455    if (!brw->query.begin_emitted)
 456       return;
 457
 458    write_depth_count(intel, brw->query.bo, brw->query.index * 2 + 1);
 459
 460    brw->query.begin_emitted = false;
 461    brw->query.index++;
 462 }
 463
 464 /**
 465  * Driver hook for glQueryCounter().
 466  *
 467  * This handles GL_TIMESTAMP queries, which perform a pipelined read of the
 468  * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
 469  * time while the query is active.
 470  */
 471 static void
 472 brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 473 {
 474    struct intel_context *intel = intel_context(ctx);
 475    struct brw_query_object *query = (struct brw_query_object *) q;
 476
 477    assert(q->Target == GL_TIMESTAMP);
 478
 479    drm_intel_bo_unreference(query->bo);
 480    query->bo = drm_intel_bo_alloc(intel->bufmgr, "timestamp query", 4096, 4096);
 481    write_timestamp(intel, query->bo, 0);
 482 }
 483
 484 static uint64_t
 485 brw_get_timestamp(struct gl_context *ctx)
 486 {
 487    struct intel_context *intel = intel_context(ctx);
 488    uint64_t result = 0;
 489
 490    drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
 491
 492    /* See logic in brw_queryobj_get_results() */
 493    result = result >> 32;
 494    result *= 80;
 495    result &= (1ull << 36) - 1;
 496
 497    return result;
 498 }
 499
 500 void brw_init_queryobj_functions(struct dd_function_table *functions)
 501 {
 502    functions->NewQueryObject = brw_new_query_object;
 503    functions->DeleteQuery = brw_delete_query;
 504    functions->BeginQuery = brw_begin_query;
 505    functions->EndQuery = brw_end_query;
 506    functions->QueryCounter = brw_query_counter;
 507    functions->CheckQuery = brw_check_query;
 508    functions->WaitQuery = brw_wait_query;
 509    functions->GetTimestamp = brw_get_timestamp;
 510 }