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