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 support for ARB_query_object
  29  *
  30  * ARB_query_object is implemented by using the PIPE_CONTROL command to stall
  31  * execution on the completion of previous depth tests, and write the
  32  * current PS_DEPTH_COUNT to a buffer object.
  33  *
  34  * We use before and after counts when drawing during a query so that
  35  * we don't pick up other clients' query data in ours.  To reduce overhead,
  36  * a single BO is used to record the query data for all active queries at
  37  * once.  This also gives us a simple bound on how much batchbuffer space is
  38  * required for handling queries, so that we can be sure that we won't
  39  * have to emit a batchbuffer without getting the ending PS_DEPTH_COUNT.
  40  */
  41 #include "main/imports.h"
  42
  43 #include "brw_context.h"
  44 #include "brw_state.h"
  45 #include "intel_batchbuffer.h"
  46 #include "intel_reg.h"
  47
  48 static void
  49 write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  50 {
  51    if (intel->gen >= 6) {
  52       /* Emit workaround flushes: */
  53       if (intel->gen == 6) {
  54          /* The timestamp write below is a non-zero post-sync op, which on
  55           * Gen6 necessitates a CS stall.  CS stalls need stall at scoreboard
  56           * set.  See the comments for intel_emit_post_sync_nonzero_flush().
  57           */
  58          BEGIN_BATCH(4);
  59          OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
  60          OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
  61          OUT_BATCH(0);
  62          OUT_BATCH(0);
  63          ADVANCE_BATCH();
  64       }
  65
  66       BEGIN_BATCH(5);
  67       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  68       OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
  69       OUT_RELOC(query_bo,
  70                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  71                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  72                 idx * sizeof(uint64_t));
  73       OUT_BATCH(0);
  74       OUT_BATCH(0);
  75       ADVANCE_BATCH();
  76    } else {
  77       BEGIN_BATCH(4);
  78       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
  79                 PIPE_CONTROL_WRITE_TIMESTAMP);
  80       OUT_RELOC(query_bo,
  81                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  82                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
  83                 idx * sizeof(uint64_t));
  84       OUT_BATCH(0);
  85       OUT_BATCH(0);
  86       ADVANCE_BATCH();
  87    }
  88 }
  89
  90 static void
  91 write_depth_count(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  92 {
  93    if (intel->gen >= 6) {
  94       /* Emit Sandybridge workaround flush: */
  95       if (intel->gen == 6)
  96          intel_emit_post_sync_nonzero_flush(intel);
  97
  98       BEGIN_BATCH(5);
  99       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
 100       OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
 101                 PIPE_CONTROL_WRITE_DEPTH_COUNT);
 102       OUT_RELOC(query_bo,
 103                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 104                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
 105                 (idx * sizeof(uint64_t)));
 106       OUT_BATCH(0);
 107       OUT_BATCH(0);
 108       ADVANCE_BATCH();
 109    } else {
 110       BEGIN_BATCH(4);
 111       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
 112                 PIPE_CONTROL_DEPTH_STALL |
 113                 PIPE_CONTROL_WRITE_DEPTH_COUNT);
 114       /* This object could be mapped cacheable, but we don't have an exposed
 115        * mechanism to support that.  Since it's going uncached, tell GEM that
 116        * we're writing to it.  The usual clflush should be all that's required
 117        * to pick up the results.
 118        */
 119       OUT_RELOC(query_bo,
 120                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 121                 PIPE_CONTROL_GLOBAL_GTT_WRITE |
 122                 (idx * sizeof(uint64_t)));
 123       OUT_BATCH(0);
 124       OUT_BATCH(0);
 125       ADVANCE_BATCH();
 126    }
 127 }
 128
 129 /** Waits on the query object's BO and totals the results for this query */
 130 static void
 131 brw_queryobj_get_results(struct gl_context *ctx,
 132                          struct brw_query_object *query)
 133 {
 134    struct intel_context *intel = intel_context(ctx);
 135
 136    int i;
 137    uint64_t *results;
 138
 139    if (query->bo == NULL)
 140       return;
 141
 142    if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
 143       if (drm_intel_bo_busy(query->bo)) {
 144          perf_debug("Stalling on the GPU waiting for a query object.\n");
 145       }
 146    }
 147
 148    drm_intel_bo_map(query->bo, false);
 149    results = query->bo->virtual;
 150    switch (query->Base.Target) {
 151    case GL_TIME_ELAPSED_EXT:
 152       if (intel->gen >= 6)
 153          query->Base.Result += 80 * (results[1] - results[0]);
 154       else
 155          query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
 156       break;
 157
 158    case GL_SAMPLES_PASSED_ARB:
 159       /* Map and count the pixels from the current query BO */
 160       for (i = query->first_index; i <= query->last_index; i++) {
 161          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 162       }
 163       break;
 164
 165    case GL_PRIMITIVES_GENERATED:
 166    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 167       /* We don't actually query the hardware for this value, so query->bo
 168        * should always be NULL and execution should never reach here.
 169        */
 170       assert(!"Unreachable");
 171       break;
 172
 173    default:
 174       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 175       break;
 176    }
 177    drm_intel_bo_unmap(query->bo);
 178
 179    drm_intel_bo_unreference(query->bo);
 180    query->bo = NULL;
 181 }
 182
 183 static struct gl_query_object *
 184 brw_new_query_object(struct gl_context *ctx, GLuint id)
 185 {
 186    struct brw_query_object *query;
 187
 188    query = calloc(1, sizeof(struct brw_query_object));
 189
 190    query->Base.Id = id;
 191    query->Base.Result = 0;
 192    query->Base.Active = false;
 193    query->Base.Ready = true;
 194
 195    return &query->Base;
 196 }
 197
 198 static void
 199 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 200 {
 201    struct brw_query_object *query = (struct brw_query_object *)q;
 202
 203    drm_intel_bo_unreference(query->bo);
 204    free(query);
 205 }
 206
 207 static void
 208 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 209 {
 210    struct brw_context *brw = brw_context(ctx);
 211    struct intel_context *intel = intel_context(ctx);
 212    struct brw_query_object *query = (struct brw_query_object *)q;
 213
 214    switch (query->Base.Target) {
 215    case GL_TIME_ELAPSED_EXT:
 216       drm_intel_bo_unreference(query->bo);
 217       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
 218       write_timestamp(intel, query->bo, 0);
 219       break;
 220
 221    case GL_SAMPLES_PASSED_ARB:
 222       /* Reset our driver's tracking of query state. */
 223       drm_intel_bo_unreference(query->bo);
 224       query->bo = NULL;
 225       query->first_index = -1;
 226       query->last_index = -1;
 227
 228       brw->query.obj = query;
 229       intel->stats_wm++;
 230       break;
 231
 232    case GL_PRIMITIVES_GENERATED:
 233       /* We don't actually query the hardware for this value; we keep track of
 234        * it a software counter.  So just reset the counter.
 235        */
 236       brw->sol.primitives_generated = 0;
 237       brw->sol.counting_primitives_generated = true;
 238       break;
 239
 240    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 241       /* We don't actually query the hardware for this value; we keep track of
 242        * it a software counter.  So just reset the counter.
 243        */
 244       brw->sol.primitives_written = 0;
 245       brw->sol.counting_primitives_written = true;
 246       break;
 247
 248    default:
 249       assert(!"Unrecognized query target in brw_begin_query()");
 250       break;
 251    }
 252 }
 253
 254 /**
 255  * Begin the ARB_occlusion_query query on a query object.
 256  */
 257 static void
 258 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 259 {
 260    struct brw_context *brw = brw_context(ctx);
 261    struct intel_context *intel = intel_context(ctx);
 262    struct brw_query_object *query = (struct brw_query_object *)q;
 263
 264    switch (query->Base.Target) {
 265    case GL_TIME_ELAPSED_EXT:
 266       write_timestamp(intel, query->bo, 1);
 267       intel_batchbuffer_flush(intel);
 268       break;
 269
 270    case GL_SAMPLES_PASSED_ARB:
 271       /* Flush the batchbuffer in case it has writes to our query BO.
 272        * Have later queries write to a new query BO so that further rendering
 273        * doesn't delay the collection of our results.
 274        */
 275       if (query->bo) {
 276          brw_emit_query_end(brw);
 277          intel_batchbuffer_flush(intel);
 278
 279          drm_intel_bo_unreference(brw->query.bo);
 280          brw->query.bo = NULL;
 281       }
 282
 283       brw->query.obj = NULL;
 284
 285       intel->stats_wm--;
 286       break;
 287
 288    case GL_PRIMITIVES_GENERATED:
 289       /* We don't actually query the hardware for this value; we keep track of
 290        * it in a software counter.  So just read the counter and store it in
 291        * the query object.
 292        */
 293       query->Base.Result = brw->sol.primitives_generated;
 294       brw->sol.counting_primitives_generated = false;
 295
 296       /* And set brw->query.obj to NULL so that this query won't try to wait
 297        * for any rendering to complete.
 298        */
 299       query->bo = NULL;
 300       break;
 301
 302    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 303       /* We don't actually query the hardware for this value; we keep track of
 304        * it in a software counter.  So just read the counter and store it in
 305        * the query object.
 306        */
 307       query->Base.Result = brw->sol.primitives_written;
 308       brw->sol.counting_primitives_written = false;
 309
 310       /* And set brw->query.obj to NULL so that this query won't try to wait
 311        * for any rendering to complete.
 312        */
 313       query->bo = NULL;
 314       break;
 315
 316    default:
 317       assert(!"Unrecognized query target in brw_end_query()");
 318       break;
 319    }
 320 }
 321
 322 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 323 {
 324    struct brw_query_object *query = (struct brw_query_object *)q;
 325
 326    brw_queryobj_get_results(ctx, query);
 327    query->Base.Ready = true;
 328 }
 329
 330 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 331 {
 332    struct brw_query_object *query = (struct brw_query_object *)q;
 333
 334    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 335       brw_queryobj_get_results(ctx, query);
 336       query->Base.Ready = true;
 337    }
 338 }
 339
 340 /** Called to set up the query BO and account for its aperture space */
 341 void
 342 brw_prepare_query_begin(struct brw_context *brw)
 343 {
 344    struct intel_context *intel = &brw->intel;
 345
 346    /* Skip if we're not doing any queries. */
 347    if (!brw->query.obj)
 348       return;
 349
 350    /* Get a new query BO if we're going to need it. */
 351    if (brw->query.bo == NULL ||
 352        brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 353       drm_intel_bo_unreference(brw->query.bo);
 354       brw->query.bo = NULL;
 355
 356       brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
 357
 358       /* clear target buffer */
 359       drm_intel_bo_map(brw->query.bo, true);
 360       memset((char *)brw->query.bo->virtual, 0, 4096);
 361       drm_intel_bo_unmap(brw->query.bo);
 362
 363       brw->query.index = 0;
 364    }
 365 }
 366
 367 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
 368 void
 369 brw_emit_query_begin(struct brw_context *brw)
 370 {
 371    struct intel_context *intel = &brw->intel;
 372    struct gl_context *ctx = &intel->ctx;
 373    struct brw_query_object *query = brw->query.obj;
 374
 375    /* Skip if we're not doing any queries, or we've emitted the start. */
 376    if (!query || brw->query.active)
 377       return;
 378
 379    write_depth_count(intel, brw->query.bo, brw->query.index * 2);
 380
 381    if (query->bo != brw->query.bo) {
 382       if (query->bo != NULL)
 383          brw_queryobj_get_results(ctx, query);
 384       drm_intel_bo_reference(brw->query.bo);
 385       query->bo = brw->query.bo;
 386       query->first_index = brw->query.index;
 387    }
 388    query->last_index = brw->query.index;
 389    brw->query.active = true;
 390 }
 391
 392 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
 393 void
 394 brw_emit_query_end(struct brw_context *brw)
 395 {
 396    struct intel_context *intel = &brw->intel;
 397
 398    if (!brw->query.active)
 399       return;
 400
 401    write_depth_count(intel, brw->query.bo, brw->query.index * 2 + 1);
 402
 403    brw->query.active = false;
 404    brw->query.index++;
 405 }
 406
 407 void brw_init_queryobj_functions(struct dd_function_table *functions)
 408 {
 409    functions->NewQueryObject = brw_new_query_object;
 410    functions->DeleteQuery = brw_delete_query;
 411    functions->BeginQuery = brw_begin_query;
 412    functions->EndQuery = brw_end_query;
 413    functions->CheckQuery = brw_check_query;
 414    functions->WaitQuery = brw_wait_query;
 415 }