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[1] & 0xffffffff);
 183          query->Base.Result &= (1ull << 36) - 1;
 184       } else {
 185          query->Base.Result = 1000 * (results[1] >> 32);
 186       }
 187
 188       break;
 189
 190    case GL_SAMPLES_PASSED_ARB:
 191       /* Map and count the pixels from the current query BO */
 192       for (i = query->first_index; i <= query->last_index; i++) {
 193          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 194       }
 195       break;
 196
 197    case GL_ANY_SAMPLES_PASSED:
 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_SAMPLES_PASSED_ARB:
 265       /* Reset our driver's tracking of query state. */
 266       drm_intel_bo_unreference(query->bo);
 267       query->bo = NULL;
 268       query->first_index = -1;
 269       query->last_index = -1;
 270
 271       brw->query.obj = query;
 272       intel->stats_wm++;
 273       break;
 274
 275    case GL_PRIMITIVES_GENERATED:
 276       /* We don't actually query the hardware for this value; we keep track of
 277        * it a software counter.  So just reset the counter.
 278        */
 279       brw->sol.primitives_generated = 0;
 280       brw->sol.counting_primitives_generated = true;
 281       break;
 282
 283    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 284       /* We don't actually query the hardware for this value; we keep track of
 285        * it a software counter.  So just reset the counter.
 286        */
 287       brw->sol.primitives_written = 0;
 288       brw->sol.counting_primitives_written = true;
 289       break;
 290
 291    default:
 292       assert(!"Unrecognized query target in brw_begin_query()");
 293       break;
 294    }
 295 }
 296
 297 /**
 298  * Begin the ARB_occlusion_query query on a query object.
 299  */
 300 static void
 301 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 302 {
 303    struct brw_context *brw = brw_context(ctx);
 304    struct intel_context *intel = intel_context(ctx);
 305    struct brw_query_object *query = (struct brw_query_object *)q;
 306
 307    switch (query->Base.Target) {
 308    case GL_TIMESTAMP:
 309       drm_intel_bo_unreference(query->bo);
 310       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query",
 311                                      4096, 4096);
 312       /* FALLTHROUGH */
 313
 314    case GL_TIME_ELAPSED_EXT:
 315       write_timestamp(intel, query->bo, 1);
 316       break;
 317
 318    case GL_ANY_SAMPLES_PASSED:
 319    case GL_SAMPLES_PASSED_ARB:
 320       if (query->bo) {
 321          brw_emit_query_end(brw);
 322
 323          drm_intel_bo_unreference(brw->query.bo);
 324          brw->query.bo = NULL;
 325       }
 326
 327       brw->query.obj = NULL;
 328
 329       intel->stats_wm--;
 330       break;
 331
 332    case GL_PRIMITIVES_GENERATED:
 333       /* We don't actually query the hardware for this value; we keep track of
 334        * it in a software counter.  So just read the counter and store it in
 335        * the query object.
 336        */
 337       query->Base.Result = brw->sol.primitives_generated;
 338       brw->sol.counting_primitives_generated = false;
 339
 340       /* And set brw->query.obj to NULL so that this query won't try to wait
 341        * for any rendering to complete.
 342        */
 343       query->bo = NULL;
 344       break;
 345
 346    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 347       /* We don't actually query the hardware for this value; we keep track of
 348        * it in a software counter.  So just read the counter and store it in
 349        * the query object.
 350        */
 351       query->Base.Result = brw->sol.primitives_written;
 352       brw->sol.counting_primitives_written = false;
 353
 354       /* And set brw->query.obj to NULL so that this query won't try to wait
 355        * for any rendering to complete.
 356        */
 357       query->bo = NULL;
 358       break;
 359
 360    default:
 361       assert(!"Unrecognized query target in brw_end_query()");
 362       break;
 363    }
 364 }
 365
 366 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 367 {
 368    struct brw_query_object *query = (struct brw_query_object *)q;
 369
 370    brw_queryobj_get_results(ctx, query);
 371    query->Base.Ready = true;
 372 }
 373
 374 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 375 {
 376    struct intel_context *intel = intel_context(ctx);
 377    struct brw_query_object *query = (struct brw_query_object *)q;
 378
 379    /* From the GL_ARB_occlusion_query spec:
 380     *
 381     *     "Instead of allowing for an infinite loop, performing a
 382     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 383     *      not ready yet on the first time it is queried.  This ensures that
 384     *      the async query will return true in finite time.
 385     */
 386    if (query->bo && drm_intel_bo_references(intel->batch.bo, query->bo))
 387       intel_batchbuffer_flush(intel);
 388
 389    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 390       brw_queryobj_get_results(ctx, query);
 391       query->Base.Ready = true;
 392    }
 393 }
 394
 395 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
 396 void
 397 brw_emit_query_begin(struct brw_context *brw)
 398 {
 399    struct intel_context *intel = &brw->intel;
 400    struct gl_context *ctx = &intel->ctx;
 401    struct brw_query_object *query = brw->query.obj;
 402
 403    /* Skip if we're not doing any queries, or we've emitted the start. */
 404    if (!query || brw->query.begin_emitted)
 405       return;
 406
 407    /* Get a new query BO if we're going to need it. */
 408    if (brw->query.bo == NULL ||
 409        brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 410       drm_intel_bo_unreference(brw->query.bo);
 411       brw->query.bo = NULL;
 412
 413       brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
 414
 415       /* clear target buffer */
 416       drm_intel_bo_map(brw->query.bo, true);
 417       memset((char *)brw->query.bo->virtual, 0, 4096);
 418       drm_intel_bo_unmap(brw->query.bo);
 419
 420       brw->query.index = 0;
 421    }
 422
 423    write_depth_count(intel, brw->query.bo, brw->query.index * 2);
 424
 425    if (query->bo != brw->query.bo) {
 426       if (query->bo != NULL)
 427          brw_queryobj_get_results(ctx, query);
 428       drm_intel_bo_reference(brw->query.bo);
 429       query->bo = brw->query.bo;
 430       query->first_index = brw->query.index;
 431    }
 432    query->last_index = brw->query.index;
 433    brw->query.begin_emitted = true;
 434 }
 435
 436 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
 437 void
 438 brw_emit_query_end(struct brw_context *brw)
 439 {
 440    struct intel_context *intel = &brw->intel;
 441
 442    if (!brw->query.begin_emitted)
 443       return;
 444
 445    write_depth_count(intel, brw->query.bo, brw->query.index * 2 + 1);
 446
 447    brw->query.begin_emitted = false;
 448    brw->query.index++;
 449 }
 450
 451 static uint64_t
 452 brw_get_timestamp(struct gl_context *ctx)
 453 {
 454    struct intel_context *intel = intel_context(ctx);
 455    uint64_t result = 0;
 456
 457    drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
 458
 459    /* See logic in brw_queryobj_get_results() */
 460    result = result >> 32;
 461    result *= 80;
 462    result &= (1ull << 36) - 1;
 463
 464    return result;
 465 }
 466
 467 void brw_init_queryobj_functions(struct dd_function_table *functions)
 468 {
 469    functions->NewQueryObject = brw_new_query_object;
 470    functions->DeleteQuery = brw_delete_query;
 471    functions->BeginQuery = brw_begin_query;
 472    functions->EndQuery = brw_end_query;
 473    functions->CheckQuery = brw_check_query;
 474    functions->WaitQuery = brw_wait_query;
 475    functions->GetTimestamp = brw_get_timestamp;
 476 }