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