src/mesa/drivers/dri/i965/gen6_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  *    Kenneth Graunke <kenneth@whitecape.org>
  26  */
  27
  28 /** @file gen6_queryobj.c
  29  *
  30  * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query,
  31  * GL_EXT_transform_feedback, and friends) on platforms that support
  32  * hardware contexts (Gen6+).
  33  */
  34 #include "main/imports.h"
  35
  36 #include "brw_context.h"
  37 #include "brw_defines.h"
  38 #include "brw_state.h"
  39 #include "intel_batchbuffer.h"
  40 #include "intel_reg.h"
  41
  42 /**
  43  * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
  44  */
  45 static void
  46 write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  47 {
  48    /* Emit workaround flushes: */
  49    if (intel->gen == 6) {
  50       /* The timestamp write below is a non-zero post-sync op, which on
  51        * Gen6 necessitates a CS stall.  CS stalls need stall at scoreboard
  52        * set.  See the comments for intel_emit_post_sync_nonzero_flush().
  53        */
  54       BEGIN_BATCH(4);
  55       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
  56       OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
  57       OUT_BATCH(0);
  58       OUT_BATCH(0);
  59       ADVANCE_BATCH();
  60    }
  61
  62    BEGIN_BATCH(5);
  63    OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  64    OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
  65    OUT_RELOC(query_bo,
  66              I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  67              PIPE_CONTROL_GLOBAL_GTT_WRITE |
  68              idx * sizeof(uint64_t));
  69    OUT_BATCH(0);
  70    OUT_BATCH(0);
  71    ADVANCE_BATCH();
  72 }
  73
  74 /**
  75  * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
  76  */
  77 static void
  78 write_depth_count(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
  79 {
  80    /* Emit Sandybridge workaround flush: */
  81    if (intel->gen == 6)
  82       intel_emit_post_sync_nonzero_flush(intel);
  83
  84    BEGIN_BATCH(5);
  85    OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  86    OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
  87              PIPE_CONTROL_WRITE_DEPTH_COUNT);
  88    OUT_RELOC(query_bo,
  89              I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  90              PIPE_CONTROL_GLOBAL_GTT_WRITE |
  91              (idx * sizeof(uint64_t)));
  92    OUT_BATCH(0);
  93    OUT_BATCH(0);
  94    ADVANCE_BATCH();
  95 }
  96
  97 /*
  98  * Write an arbitrary 64-bit register to a buffer via MI_STORE_REGISTER_MEM.
  99  *
 100  * Only TIMESTAMP and PS_DEPTH_COUNT have special PIPE_CONTROL support; other
 101  * counters have to be read via the generic MI_STORE_REGISTER_MEM.  This
 102  * function also performs a pipeline flush for proper synchronization.
 103  */
 104 static void
 105 write_reg(struct intel_context *intel,
 106           drm_intel_bo *query_bo, uint32_t reg, int idx)
 107 {
 108    assert(intel->gen >= 6);
 109
 110    intel_batchbuffer_emit_mi_flush(intel);
 111
 112    /* MI_STORE_REGISTER_MEM only stores a single 32-bit value, so to
 113     * read a full 64-bit register, we need to do two of them.
 114     */
 115    BEGIN_BATCH(3);
 116    OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
 117    OUT_BATCH(reg);
 118    OUT_RELOC(query_bo, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
 119              idx * sizeof(uint64_t));
 120    ADVANCE_BATCH();
 121
 122    BEGIN_BATCH(3);
 123    OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
 124    OUT_BATCH(reg + sizeof(uint32_t));
 125    OUT_RELOC(query_bo, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
 126              sizeof(uint32_t) + idx * sizeof(uint64_t));
 127    ADVANCE_BATCH();
 128 }
 129
 130 static void
 131 write_primitives_generated(struct intel_context *intel,
 132                            drm_intel_bo *query_bo, int idx)
 133 {
 134    write_reg(intel, query_bo, CL_INVOCATION_COUNT, idx);
 135 }
 136
 137 static void
 138 write_xfb_primitives_written(struct intel_context *intel,
 139                              drm_intel_bo *query_bo, int idx)
 140 {
 141    if (intel->gen >= 7) {
 142       write_reg(intel, query_bo, SO_NUM_PRIMS_WRITTEN0_IVB, idx);
 143    } else {
 144       write_reg(intel, query_bo, SO_NUM_PRIMS_WRITTEN, idx);
 145    }
 146 }
 147
 148 /**
 149  * Wait on the query object's BO and calculate the final result.
 150  */
 151 static void
 152 gen6_queryobj_get_results(struct gl_context *ctx,
 153                           struct brw_query_object *query)
 154 {
 155    struct intel_context *intel = intel_context(ctx);
 156
 157    if (query->bo == NULL)
 158       return;
 159
 160    /* If the application has requested the query result, but this batch is
 161     * still contributing to it, flush it now so the results will be present
 162     * when mapped.
 163     */
 164    if (drm_intel_bo_references(intel->batch.bo, query->bo))
 165       intel_batchbuffer_flush(intel);
 166
 167    if (unlikely(intel->perf_debug)) {
 168       if (drm_intel_bo_busy(query->bo)) {
 169          perf_debug("Stalling on the GPU waiting for a query object.\n");
 170       }
 171    }
 172
 173    drm_intel_bo_map(query->bo, false);
 174    uint64_t *results = query->bo->virtual;
 175    switch (query->Base.Target) {
 176    case GL_TIME_ELAPSED:
 177       /* The query BO contains the starting and ending timestamps.
 178        * Subtract the two and convert to nanoseconds.
 179        */
 180       query->Base.Result += 80 * (results[1] - results[0]);
 181       break;
 182
 183    case GL_TIMESTAMP:
 184       /* Our timer is a clock that increments every 80ns (regardless of
 185        * other clock scaling in the system).  The timestamp register we can
 186        * read for glGetTimestamp() masks out the top 32 bits, so we do that
 187        * here too to let the two counters be compared against each other.
 188        *
 189        * If we just multiplied that 32 bits of data by 80, it would roll
 190        * over at a non-power-of-two, so an application couldn't use
 191        * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
 192        * report 36 bits and truncate at that (rolling over 5 times as often
 193        * as the HW counter), and when the 32-bit counter rolls over, it
 194        * happens to also be at a rollover in the reported value from near
 195        * (1<<36) to 0.
 196        *
 197        * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
 198        * rolls over every ~69 seconds.
 199        *
 200        * The query BO contains a single timestamp value in results[0].
 201        */
 202       query->Base.Result = 80 * (results[0] & 0xffffffff);
 203       query->Base.Result &= (1ull << 36) - 1;
 204       break;
 205
 206    case GL_SAMPLES_PASSED_ARB:
 207       /* We need to use += rather than = here since some BLT-based operations
 208        * may have added additional samples to our occlusion query value.
 209        */
 210       query->Base.Result += results[1] - results[0];
 211       break;
 212
 213    case GL_ANY_SAMPLES_PASSED:
 214    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 215       if (results[0] != results[1])
 216          query->Base.Result = true;
 217       break;
 218
 219    case GL_PRIMITIVES_GENERATED:
 220    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 221       query->Base.Result = results[1] - results[0];
 222       break;
 223
 224    default:
 225       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 226       break;
 227    }
 228    drm_intel_bo_unmap(query->bo);
 229
 230    /* Now that we've processed the data stored in the query's buffer object,
 231     * we can release it.
 232     */
 233    drm_intel_bo_unreference(query->bo);
 234    query->bo = NULL;
 235 }
 236
 237 /**
 238  * Driver hook for glBeginQuery().
 239  *
 240  * Initializes driver structures and emits any GPU commands required to begin
 241  * recording data for the query.
 242  */
 243 static void
 244 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 245 {
 246    struct intel_context *intel = intel_context(ctx);
 247    struct brw_query_object *query = (struct brw_query_object *)q;
 248
 249    /* Since we're starting a new query, we need to throw away old results. */
 250    drm_intel_bo_unreference(query->bo);
 251    query->bo = drm_intel_bo_alloc(intel->bufmgr, "query results", 4096, 4096);
 252
 253    switch (query->Base.Target) {
 254    case GL_TIME_ELAPSED:
 255       /* For timestamp queries, we record the starting time right away so that
 256        * we measure the full time between BeginQuery and EndQuery.  There's
 257        * some debate about whether this is the right thing to do.  Our decision
 258        * is based on the following text from the ARB_timer_query extension:
 259        *
 260        * "(5) Should the extension measure total time elapsed between the full
 261        *      completion of the BeginQuery and EndQuery commands, or just time
 262        *      spent in the graphics library?
 263        *
 264        *  RESOLVED:  This extension will measure the total time elapsed
 265        *  between the full completion of these commands.  Future extensions
 266        *  may implement a query to determine time elapsed at different stages
 267        *  of the graphics pipeline."
 268        *
 269        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 270        * we'll write a second timestamp (at index 1), and subtract the two to
 271        * obtain the time elapsed.  Notably, this includes time elapsed while
 272        * the system was doing other work, such as running other applications.
 273        */
 274       write_timestamp(intel, query->bo, 0);
 275       break;
 276
 277    case GL_ANY_SAMPLES_PASSED:
 278    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 279    case GL_SAMPLES_PASSED_ARB:
 280       write_depth_count(intel, query->bo, 0);
 281       break;
 282
 283    case GL_PRIMITIVES_GENERATED:
 284       write_primitives_generated(intel, query->bo, 0);
 285       break;
 286
 287    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 288       write_xfb_primitives_written(intel, query->bo, 0);
 289       break;
 290
 291    default:
 292       assert(!"Unrecognized query target in brw_begin_query()");
 293       break;
 294    }
 295 }
 296
 297 /**
 298  * Driver hook for glEndQuery().
 299  *
 300  * Emits GPU commands to record a final query value, ending any data capturing.
 301  * However, the final result isn't necessarily available until the GPU processes
 302  * those commands.  brw_queryobj_get_results() processes the captured data to
 303  * produce the final result.
 304  */
 305 static void
 306 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 307 {
 308    struct intel_context *intel = intel_context(ctx);
 309    struct brw_query_object *query = (struct brw_query_object *)q;
 310
 311    switch (query->Base.Target) {
 312    case GL_TIME_ELAPSED:
 313       write_timestamp(intel, query->bo, 1);
 314       break;
 315
 316    case GL_ANY_SAMPLES_PASSED:
 317    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 318    case GL_SAMPLES_PASSED_ARB:
 319       write_depth_count(intel, query->bo, 1);
 320       break;
 321
 322    case GL_PRIMITIVES_GENERATED:
 323       write_primitives_generated(intel, query->bo, 1);
 324       break;
 325
 326    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 327       write_xfb_primitives_written(intel, query->bo, 1);
 328       break;
 329
 330    default:
 331       assert(!"Unrecognized query target in brw_end_query()");
 332       break;
 333    }
 334 }
 335
 336 /**
 337  * The WaitQuery() driver hook.
 338  *
 339  * Wait for a query result to become available and return it.  This is the
 340  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 341  */
 342 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 343 {
 344    struct brw_query_object *query = (struct brw_query_object *)q;
 345
 346    gen6_queryobj_get_results(ctx, query);
 347    query->Base.Ready = true;
 348 }
 349
 350 /**
 351  * The CheckQuery() driver hook.
 352  *
 353  * Checks whether a query result is ready yet.  If not, flushes.
 354  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 355  */
 356 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
 357 {
 358    struct intel_context *intel = intel_context(ctx);
 359    struct brw_query_object *query = (struct brw_query_object *)q;
 360
 361    /* From the GL_ARB_occlusion_query spec:
 362     *
 363     *     "Instead of allowing for an infinite loop, performing a
 364     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 365     *      not ready yet on the first time it is queried.  This ensures that
 366     *      the async query will return true in finite time.
 367     */
 368    if (query->bo && drm_intel_bo_references(intel->batch.bo, query->bo))
 369       intel_batchbuffer_flush(intel);
 370
 371    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 372       gen6_queryobj_get_results(ctx, query);
 373       query->Base.Ready = true;
 374    }
 375 }
 376
 377 /* Initialize Gen6+-specific query object functions. */
 378 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 379 {
 380    functions->BeginQuery = gen6_begin_query;
 381    functions->EndQuery = gen6_end_query;
 382    functions->CheckQuery = gen6_check_query;
 383    functions->WaitQuery = gen6_wait_query;
 384 }