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 /** Waits on the query object's BO and totals the results for this query */
  49 static void
  50 brw_queryobj_get_results(struct gl_context *ctx,
  51                          struct brw_query_object *query)
  52 {
  53    struct intel_context *intel = intel_context(ctx);
  54
  55    int i;
  56    uint64_t *results;
  57
  58    if (query->bo == NULL)
  59       return;
  60
  61    drm_intel_bo_map(query->bo, false);
  62    results = query->bo->virtual;
  63    switch (query->Base.Target) {
  64    case GL_TIME_ELAPSED_EXT:
  65       if (intel->gen >= 6)
  66          query->Base.Result += 80 * (results[1] - results[0]);
  67       else
  68          query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
  69       break;
  70
  71    case GL_SAMPLES_PASSED_ARB:
  72       /* Map and count the pixels from the current query BO */
  73       for (i = query->first_index; i <= query->last_index; i++) {
  74          query->Base.Result += results[i * 2 + 1] - results[i * 2];
  75       }
  76       break;
  77
  78    case GL_PRIMITIVES_GENERATED:
  79    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
  80       /* We don't actually query the hardware for this value, so query->bo
  81        * should always be NULL and execution should never reach here.
  82        */
  83       assert(!"Unreachable");
  84       break;
  85
  86    default:
  87       assert(!"Unrecognized query target in brw_queryobj_get_results()");
  88       break;
  89    }
  90    drm_intel_bo_unmap(query->bo);
  91
  92    drm_intel_bo_unreference(query->bo);
  93    query->bo = NULL;
  94 }
  95
  96 static struct gl_query_object *
  97 brw_new_query_object(struct gl_context *ctx, GLuint id)
  98 {
  99    struct brw_query_object *query;
 100
 101    query = calloc(1, sizeof(struct brw_query_object));
 102
 103    query->Base.Id = id;
 104    query->Base.Result = 0;
 105    query->Base.Active = false;
 106    query->Base.Ready = true;
 107
 108    return &query->Base;
 109 }
 110
 111 static void
 112 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
 113 {
 114    struct brw_query_object *query = (struct brw_query_object *)q;
 115
 116    drm_intel_bo_unreference(query->bo);
 117    free(query);
 118 }
 119
 120 static void
 121 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 122 {
 123    struct brw_context *brw = brw_context(ctx);
 124    struct intel_context *intel = intel_context(ctx);
 125    struct brw_query_object *query = (struct brw_query_object *)q;
 126
 127    switch (query->Base.Target) {
 128    case GL_TIME_ELAPSED_EXT:
 129       drm_intel_bo_unreference(query->bo);
 130       query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query",
 131                                      4096, 4096);
 132
 133       if (intel->gen >= 6) {
 134           BEGIN_BATCH(4);
 135           OUT_BATCH(_3DSTATE_PIPE_CONTROL);
 136           OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
 137           OUT_RELOC(query->bo,
 138                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 139                   PIPE_CONTROL_GLOBAL_GTT_WRITE |
 140                   0);
 141           OUT_BATCH(0);
 142           ADVANCE_BATCH();
 143
 144       } else {
 145           BEGIN_BATCH(4);
 146           OUT_BATCH(_3DSTATE_PIPE_CONTROL |
 147                   PIPE_CONTROL_WRITE_TIMESTAMP);
 148           OUT_RELOC(query->bo,
 149                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 150                   PIPE_CONTROL_GLOBAL_GTT_WRITE |
 151                   0);
 152           OUT_BATCH(0);
 153           OUT_BATCH(0);
 154           ADVANCE_BATCH();
 155       }
 156       break;
 157
 158    case GL_SAMPLES_PASSED_ARB:
 159       /* Reset our driver's tracking of query state. */
 160       drm_intel_bo_unreference(query->bo);
 161       query->bo = NULL;
 162       query->first_index = -1;
 163       query->last_index = -1;
 164
 165       brw->query.obj = query;
 166       intel->stats_wm++;
 167       break;
 168
 169    case GL_PRIMITIVES_GENERATED:
 170       /* We don't actually query the hardware for this value; we keep track of
 171        * it a software counter.  So just reset the counter.
 172        */
 173       brw->sol.primitives_generated = 0;
 174       brw->sol.counting_primitives_generated = true;
 175       break;
 176
 177    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 178       /* We don't actually query the hardware for this value; we keep track of
 179        * it a software counter.  So just reset the counter.
 180        */
 181       brw->sol.primitives_written = 0;
 182       brw->sol.counting_primitives_written = true;
 183       break;
 184
 185    default:
 186       assert(!"Unrecognized query target in brw_begin_query()");
 187       break;
 188    }
 189 }
 190
 191 /**
 192  * Begin the ARB_occlusion_query query on a query object.
 193  */
 194 static void
 195 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
 196 {
 197    struct brw_context *brw = brw_context(ctx);
 198    struct intel_context *intel = intel_context(ctx);
 199    struct brw_query_object *query = (struct brw_query_object *)q;
 200
 201    switch (query->Base.Target) {
 202    case GL_TIME_ELAPSED_EXT:
 203       if (intel->gen >= 6) {
 204           BEGIN_BATCH(4);
 205           OUT_BATCH(_3DSTATE_PIPE_CONTROL);
 206           OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
 207           OUT_RELOC(query->bo,
 208                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 209                   PIPE_CONTROL_GLOBAL_GTT_WRITE |
 210                   8);
 211           OUT_BATCH(0);
 212           ADVANCE_BATCH();
 213
 214       } else {
 215           BEGIN_BATCH(4);
 216           OUT_BATCH(_3DSTATE_PIPE_CONTROL |
 217                   PIPE_CONTROL_WRITE_TIMESTAMP);
 218           OUT_RELOC(query->bo,
 219                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 220                   PIPE_CONTROL_GLOBAL_GTT_WRITE |
 221                   8);
 222           OUT_BATCH(0);
 223           OUT_BATCH(0);
 224           ADVANCE_BATCH();
 225       }
 226
 227       intel_batchbuffer_flush(intel);
 228       break;
 229
 230    case GL_SAMPLES_PASSED_ARB:
 231       /* Flush the batchbuffer in case it has writes to our query BO.
 232        * Have later queries write to a new query BO so that further rendering
 233        * doesn't delay the collection of our results.
 234        */
 235       if (query->bo) {
 236          brw_emit_query_end(brw);
 237          intel_batchbuffer_flush(intel);
 238
 239          drm_intel_bo_unreference(brw->query.bo);
 240          brw->query.bo = NULL;
 241       }
 242
 243       brw->query.obj = NULL;
 244
 245       intel->stats_wm--;
 246       break;
 247
 248    case GL_PRIMITIVES_GENERATED:
 249       /* We don't actually query the hardware for this value; we keep track of
 250        * it in a software counter.  So just read the counter and store it in
 251        * the query object.
 252        */
 253       query->Base.Result = brw->sol.primitives_generated;
 254       brw->sol.counting_primitives_generated = false;
 255
 256       /* And set brw->query.obj to NULL so that this query won't try to wait
 257        * for any rendering to complete.
 258        */
 259       query->bo = NULL;
 260       break;
 261
 262    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 263       /* We don't actually query the hardware for this value; we keep track of
 264        * it in a software counter.  So just read the counter and store it in
 265        * the query object.
 266        */
 267       query->Base.Result = brw->sol.primitives_written;
 268       brw->sol.counting_primitives_written = false;
 269
 270       /* And set brw->query.obj to NULL so that this query won't try to wait
 271        * for any rendering to complete.
 272        */
 273       query->bo = NULL;
 274       break;
 275
 276    default:
 277       assert(!"Unrecognized query target in brw_end_query()");
 278       break;
 279    }
 280 }
 281
 282 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 283 {
 284    struct brw_query_object *query = (struct brw_query_object *)q;
 285
 286    brw_queryobj_get_results(ctx, query);
 287    query->Base.Ready = true;
 288 }
 289
 290 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
 291 {
 292    struct brw_query_object *query = (struct brw_query_object *)q;
 293
 294    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 295       brw_queryobj_get_results(ctx, query);
 296       query->Base.Ready = true;
 297    }
 298 }
 299
 300 /** Called to set up the query BO and account for its aperture space */
 301 void
 302 brw_prepare_query_begin(struct brw_context *brw)
 303 {
 304    struct intel_context *intel = &brw->intel;
 305
 306    /* Skip if we're not doing any queries. */
 307    if (!brw->query.obj)
 308       return;
 309
 310    /* Get a new query BO if we're going to need it. */
 311    if (brw->query.bo == NULL ||
 312        brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
 313       drm_intel_bo_unreference(brw->query.bo);
 314       brw->query.bo = NULL;
 315
 316       brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
 317
 318       /* clear target buffer */
 319       drm_intel_bo_map(brw->query.bo, true);
 320       memset((char *)brw->query.bo->virtual, 0, 4096);
 321       drm_intel_bo_unmap(brw->query.bo);
 322
 323       brw->query.index = 0;
 324    }
 325 }
 326
 327 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
 328 void
 329 brw_emit_query_begin(struct brw_context *brw)
 330 {
 331    struct intel_context *intel = &brw->intel;
 332    struct gl_context *ctx = &intel->ctx;
 333    struct brw_query_object *query = brw->query.obj;
 334
 335    /* Skip if we're not doing any queries, or we've emitted the start. */
 336    if (!query || brw->query.active)
 337       return;
 338
 339    if (intel->gen >= 6) {
 340        BEGIN_BATCH(8);
 341
 342        /* workaround: CS stall required before depth stall. */
 343        OUT_BATCH(_3DSTATE_PIPE_CONTROL);
 344        OUT_BATCH(PIPE_CONTROL_CS_STALL);
 345        OUT_BATCH(0); /* write address */
 346        OUT_BATCH(0); /* write data */
 347
 348        OUT_BATCH(_3DSTATE_PIPE_CONTROL);
 349        OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
 350                  PIPE_CONTROL_WRITE_DEPTH_COUNT);
 351        OUT_RELOC(brw->query.bo,
 352                  I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 353                  PIPE_CONTROL_GLOBAL_GTT_WRITE |
 354                  ((brw->query.index * 2) * sizeof(uint64_t)));
 355        OUT_BATCH(0);
 356        ADVANCE_BATCH();
 357
 358    } else {
 359        BEGIN_BATCH(4);
 360        OUT_BATCH(_3DSTATE_PIPE_CONTROL |
 361                PIPE_CONTROL_DEPTH_STALL |
 362                PIPE_CONTROL_WRITE_DEPTH_COUNT);
 363        /* This object could be mapped cacheable, but we don't have an exposed
 364         * mechanism to support that.  Since it's going uncached, tell GEM that
 365         * we're writing to it.  The usual clflush should be all that's required
 366         * to pick up the results.
 367         */
 368        OUT_RELOC(brw->query.bo,
 369                I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 370                PIPE_CONTROL_GLOBAL_GTT_WRITE |
 371                ((brw->query.index * 2) * sizeof(uint64_t)));
 372        OUT_BATCH(0);
 373        OUT_BATCH(0);
 374        ADVANCE_BATCH();
 375    }
 376
 377    if (query->bo != brw->query.bo) {
 378       if (query->bo != NULL)
 379          brw_queryobj_get_results(ctx, query);
 380       drm_intel_bo_reference(brw->query.bo);
 381       query->bo = brw->query.bo;
 382       query->first_index = brw->query.index;
 383    }
 384    query->last_index = brw->query.index;
 385    brw->query.active = true;
 386 }
 387
 388 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
 389 void
 390 brw_emit_query_end(struct brw_context *brw)
 391 {
 392    struct intel_context *intel = &brw->intel;
 393
 394    if (!brw->query.active)
 395       return;
 396
 397    if (intel->gen >= 6) {
 398        BEGIN_BATCH(8);
 399        /* workaround: CS stall required before depth stall. */
 400        OUT_BATCH(_3DSTATE_PIPE_CONTROL);
 401        OUT_BATCH(PIPE_CONTROL_CS_STALL);
 402        OUT_BATCH(0); /* write address */
 403        OUT_BATCH(0); /* write data */
 404
 405        OUT_BATCH(_3DSTATE_PIPE_CONTROL);
 406        OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
 407                  PIPE_CONTROL_WRITE_DEPTH_COUNT);
 408        OUT_RELOC(brw->query.bo,
 409                  I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 410                  PIPE_CONTROL_GLOBAL_GTT_WRITE |
 411                  ((brw->query.index * 2 + 1) * sizeof(uint64_t)));
 412        OUT_BATCH(0);
 413        ADVANCE_BATCH();
 414
 415    } else {
 416        BEGIN_BATCH(4);
 417        OUT_BATCH(_3DSTATE_PIPE_CONTROL |
 418                PIPE_CONTROL_DEPTH_STALL |
 419                PIPE_CONTROL_WRITE_DEPTH_COUNT);
 420        OUT_RELOC(brw->query.bo,
 421                I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 422                PIPE_CONTROL_GLOBAL_GTT_WRITE |
 423                ((brw->query.index * 2 + 1) * sizeof(uint64_t)));
 424        OUT_BATCH(0);
 425        OUT_BATCH(0);
 426        ADVANCE_BATCH();
 427    }
 428
 429    brw->query.active = false;
 430    brw->query.index++;
 431 }
 432
 433 void brw_init_queryobj_functions(struct dd_function_table *functions)
 434 {
 435    functions->NewQueryObject = brw_new_query_object;
 436    functions->DeleteQuery = brw_delete_query;
 437    functions->BeginQuery = brw_begin_query;
 438    functions->EndQuery = brw_end_query;
 439    functions->CheckQuery = brw_check_query;
 440    functions->WaitQuery = brw_wait_query;
 441 }