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       BEGIN_BATCH(9);
  95
  96       /* workaround: CS stall required before depth stall. */
  97       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
  98       OUT_BATCH(PIPE_CONTROL_CS_STALL);
  99       OUT_BATCH(0); /* write address */
 100       OUT_BATCH(0); /* write data */
 101
 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    drm_intel_bo_map(query->bo, false);
 146    results = query->bo->virtual;
 147    switch (query->Base.Target) {
 148    case GL_TIME_ELAPSED_EXT:
 149       if (intel->gen >= 6)
 150          query->Base.Result += 80 * (results[1] - results[0]);
 151       else
 152          query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
 153       break;
 154
 155    case GL_SAMPLES_PASSED_ARB:
 156       /* Map and count the pixels from the current query BO */
 157       for (i = query->first_index; i <= query->last_index; i++) {
 158          query->Base.Result += results[i * 2 + 1] - results[i * 2];
 159       }
 160       break;
 161
 162    case GL_PRIMITIVES_GENERATED:
 163    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 164       /* We don't actually query the hardware for this value, so query->bo
 165        * should always be NULL and execution should never reach here.
 166        */
 167       assert(!"Unreachable");
 168       break;
 169
 170    default:
 171       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 172       break;
 173    }
 174    drm_intel_bo_unmap(query->bo);
 175
 176    drm_intel_bo_unreference(query->bo);
 177    query->bo = NULL;
 178 }
 179
 180 static struct gl_query_object *
 181 brw_new_query_object(struct gl_context *ctx, GLuint id)
 182 {
 183    struct brw_query_object *query;
 184
 185    query = calloc(1, sizeof(struct brw_query_object));
 186
 187    query->Base.Id = id;
 188    query->Base.Result = 0;
 189    query->Base.Active = false;
 190    query->Base.Ready = true;
 191
 192    return &query->Base;
 193 }
 194
 195 static void
 196 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 197 {
 198    struct brw_query_object *query = (struct brw_query_object *)q;
 199
 200    drm_intel_bo_unreference(query->bo);
 201    free(query);
 202 }
 203
 204 static void
 205 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 206 {
 207    struct brw_context *brw = brw_context(ctx);
 208    struct intel_context *intel = intel_context(ctx);
 209    struct brw_query_object *query = (struct brw_query_object *)q;
 210
 211    switch (query->Base.Target) {
 212    case GL_TIME_ELAPSED_EXT:
 213       drm_intel_bo_unreference(query->bo);
 214       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
 215       write_timestamp(intel, query->bo, 0);
 216       break;
 217
 218    case GL_SAMPLES_PASSED_ARB:
 219       /* Reset our driver's tracking of query state. */
 220       drm_intel_bo_unreference(query->bo);
 221       query->bo = NULL;
 222       query->first_index = -1;
 223       query->last_index = -1;
 224
 225       brw->query.obj = query;
 226       intel->stats_wm++;
 227       break;
 228
 229    case GL_PRIMITIVES_GENERATED:
 230       /* We don't actually query the hardware for this value; we keep track of
 231        * it a software counter.  So just reset the counter.
 232        */
 233       brw->sol.primitives_generated = 0;
 234       brw->sol.counting_primitives_generated = true;
 235       break;
 236
 237    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 238       /* We don't actually query the hardware for this value; we keep track of
 239        * it a software counter.  So just reset the counter.
 240        */
 241       brw->sol.primitives_written = 0;
 242       brw->sol.counting_primitives_written = true;
 243       break;
 244
 245    default:
 246       assert(!"Unrecognized query target in brw_begin_query()");
 247       break;
 248    }
 249 }
 250
 251 /**
 252  * Begin the ARB_occlusion_query query on a query object.
 253  */
 254 static void
 255 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 256 {
 257    struct brw_context *brw = brw_context(ctx);
 258    struct intel_context *intel = intel_context(ctx);
 259    struct brw_query_object *query = (struct brw_query_object *)q;
 260
 261    switch (query->Base.Target) {
 262    case GL_TIME_ELAPSED_EXT:
 263       write_timestamp(intel, query->bo, 1);
 264       intel_batchbuffer_flush(intel);
 265       break;
 266
 267    case GL_SAMPLES_PASSED_ARB:
 268       /* Flush the batchbuffer in case it has writes to our query BO.
 269        * Have later queries write to a new query BO so that further rendering
 270        * doesn't delay the collection of our results.
 271        */
 272       if (query->bo) {
 273          brw_emit_query_end(brw);
 274          intel_batchbuffer_flush(intel);
 275
 276          drm_intel_bo_unreference(brw->query.bo);
 277          brw->query.bo = NULL;
 278       }
 279
 280       brw->query.obj = NULL;
 281
 282       intel->stats_wm--;
 283       break;
 284
 285    case GL_PRIMITIVES_GENERATED:
 286       /* We don't actually query the hardware for this value; we keep track of
 287        * it in a software counter.  So just read the counter and store it in
 288        * the query object.
 289        */
 290       query->Base.Result = brw->sol.primitives_generated;
 291       brw->sol.counting_primitives_generated = false;
 292
 293       /* And set brw->query.obj to NULL so that this query won't try to wait
 294        * for any rendering to complete.
 295        */
 296       query->bo = NULL;
 297       break;
 298
 299    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 300       /* We don't actually query the hardware for this value; we keep track of
 301        * it in a software counter.  So just read the counter and store it in
 302        * the query object.
 303        */
 304       query->Base.Result = brw->sol.primitives_written;
 305       brw->sol.counting_primitives_written = false;
 306
 307       /* And set brw->query.obj to NULL so that this query won't try to wait
 308        * for any rendering to complete.
 309        */
 310       query->bo = NULL;
 311       break;
 312
 313    default:
 314       assert(!"Unrecognized query target in brw_end_query()");
 315       break;
 316    }
 317 }
 318
 319 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 320 {
 321    struct brw_query_object *query = (struct brw_query_object *)q;
 322
 323    brw_queryobj_get_results(ctx, query);
 324    query->Base.Ready = true;
 325 }
 326
 327 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 328 {
 329    struct brw_query_object *query = (struct brw_query_object *)q;
 330
 331    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 332       brw_queryobj_get_results(ctx, query);
 333       query->Base.Ready = true;
 334    }
 335 }
 336
 337 /** Called to set up the query BO and account for its aperture space */
 338 void
 339 brw_prepare_query_begin(struct brw_context *brw)
 340 {
 341    struct intel_context *intel = &brw->intel;
 342
 343    /* Skip if we're not doing any queries. */
 344    if (!brw->query.obj)
 345       return;
 346
 347    /* Get a new query BO if we're going to need it. */
 348    if (brw->query.bo == NULL ||
 349        brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 350       drm_intel_bo_unreference(brw->query.bo);
 351       brw->query.bo = NULL;
 352
 353       brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
 354
 355       /* clear target buffer */
 356       drm_intel_bo_map(brw->query.bo, true);
 357       memset((char *)brw->query.bo->virtual, 0, 4096);
 358       drm_intel_bo_unmap(brw->query.bo);
 359
 360       brw->query.index = 0;
 361    }
 362 }
 363
 364 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
 365 void
 366 brw_emit_query_begin(struct brw_context *brw)
 367 {
 368    struct intel_context *intel = &brw->intel;
 369    struct gl_context *ctx = &intel->ctx;
 370    struct brw_query_object *query = brw->query.obj;
 371
 372    /* Skip if we're not doing any queries, or we've emitted the start. */
 373    if (!query || brw->query.active)
 374       return;
 375
 376    write_depth_count(intel, brw->query.bo, brw->query.index * 2);
 377
 378    if (query->bo != brw->query.bo) {
 379       if (query->bo != NULL)
 380          brw_queryobj_get_results(ctx, query);
 381       drm_intel_bo_reference(brw->query.bo);
 382       query->bo = brw->query.bo;
 383       query->first_index = brw->query.index;
 384    }
 385    query->last_index = brw->query.index;
 386    brw->query.active = true;
 387 }
 388
 389 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
 390 void
 391 brw_emit_query_end(struct brw_context *brw)
 392 {
 393    struct intel_context *intel = &brw->intel;
 394
 395    if (!brw->query.active)
 396       return;
 397
 398    write_depth_count(intel, brw->query.bo, brw->query.index * 2 + 1);
 399
 400    brw->query.active = false;
 401    brw->query.index++;
 402 }
 403
 404 void brw_init_queryobj_functions(struct dd_function_table *functions)
 405 {
 406    functions->NewQueryObject = brw_new_query_object;
 407    functions->DeleteQuery = brw_delete_query;
 408    functions->BeginQuery = brw_begin_query;
 409    functions->EndQuery = brw_end_query;
 410    functions->CheckQuery = brw_check_query;
 411    functions->WaitQuery = brw_wait_query;
 412 }