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_buffer_objects.h"
  41
  42 static inline void
  43 set_query_availability(struct brw_context *brw, struct brw_query_object *query,
  44                        bool available)
  45 {
  46    /* For platforms that support ARB_query_buffer_object, we write the
  47     * query availability for "pipelined" queries.
  48     *
  49     * Most counter snapshots are written by the command streamer, by
  50     * doing a CS stall and then MI_STORE_REGISTER_MEM.  For these
  51     * counters, the CS stall guarantees that the results will be
  52     * available when subsequent CS commands run.  So we don't need to
  53     * do any additional tracking.
  54     *
  55     * Other counters (occlusion queries and timestamp) are written by
  56     * PIPE_CONTROL, without a CS stall.  This means that we can't be
  57     * sure whether the writes have landed yet or not.  Performing a
  58     * PIPE_CONTROL with an immediate write will synchronize with
  59     * those earlier writes, so we write 1 when the value has landed.
  60     */
  61    if (brw->ctx.Extensions.ARB_query_buffer_object &&
  62        brw_is_query_pipelined(query)) {
  63       brw_emit_pipe_control_write(brw,
  64                                   PIPE_CONTROL_WRITE_IMMEDIATE,
  65                                   query->bo, 2 * sizeof(uint64_t),
  66                                   available, 0);
  67    }
  68 }
  69
  70 static void
  71 write_primitives_generated(struct brw_context *brw,
  72                            drm_bacon_bo *query_bo, int stream, int idx)
  73 {
  74    brw_emit_mi_flush(brw);
  75
  76    if (brw->gen >= 7 && stream > 0) {
  77       brw_store_register_mem64(brw, query_bo,
  78                                GEN7_SO_PRIM_STORAGE_NEEDED(stream),
  79                                idx * sizeof(uint64_t));
  80    } else {
  81       brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT,
  82                                idx * sizeof(uint64_t));
  83    }
  84 }
  85
  86 static void
  87 write_xfb_primitives_written(struct brw_context *brw,
  88                              drm_bacon_bo *bo, int stream, int idx)
  89 {
  90    brw_emit_mi_flush(brw);
  91
  92    if (brw->gen >= 7) {
  93       brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(stream),
  94                                idx * sizeof(uint64_t));
  95    } else {
  96       brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN,
  97                                idx * sizeof(uint64_t));
  98    }
  99 }
 100
 101 static void
 102 write_xfb_overflow_streams(struct gl_context *ctx,
 103                            drm_bacon_bo *bo, int stream, int count,
 104                            int idx)
 105 {
 106    struct brw_context *brw = brw_context(ctx);
 107
 108    brw_emit_mi_flush(brw);
 109
 110    for (int i = 0; i < count; i++) {
 111       int w_idx = 4 * i + idx;
 112       int g_idx = 4 * i + idx + 2;
 113
 114       if (brw->gen >= 7) {
 115          brw_store_register_mem64(brw, bo,
 116                                   GEN7_SO_NUM_PRIMS_WRITTEN(stream + i),
 117                                   g_idx * sizeof(uint64_t));
 118          brw_store_register_mem64(brw, bo,
 119                                   GEN7_SO_PRIM_STORAGE_NEEDED(stream + i),
 120                                   w_idx * sizeof(uint64_t));
 121       } else {
 122          brw_store_register_mem64(brw, bo,
 123                                   GEN6_SO_NUM_PRIMS_WRITTEN,
 124                                   g_idx * sizeof(uint64_t));
 125          brw_store_register_mem64(brw, bo,
 126                                   GEN6_SO_PRIM_STORAGE_NEEDED,
 127                                   w_idx * sizeof(uint64_t));
 128       }
 129    }
 130 }
 131
 132 static bool
 133 check_xfb_overflow_streams(uint64_t *results, int count)
 134 {
 135    bool overflow = false;
 136
 137    for (int i = 0; i < count; i++) {
 138       uint64_t *result_i = &results[4 * i];
 139
 140       if ((result_i[3] - result_i[2]) != (result_i[1] - result_i[0])) {
 141          overflow = true;
 142          break;
 143       }
 144    }
 145
 146    return overflow;
 147 }
 148
 149 static inline int
 150 pipeline_target_to_index(int target)
 151 {
 152    if (target == GL_GEOMETRY_SHADER_INVOCATIONS)
 153       return MAX_PIPELINE_STATISTICS - 1;
 154    else
 155       return target - GL_VERTICES_SUBMITTED_ARB;
 156 }
 157
 158 static void
 159 emit_pipeline_stat(struct brw_context *brw, drm_bacon_bo *bo,
 160                    int stream, int target, int idx)
 161 {
 162    /* One source of confusion is the tessellation shader statistics. The
 163     * hardware has no statistics specific to the TE unit. Ideally we could have
 164     * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
 165     * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
 166     * Unfortunately we don't have HS primitives, we only have HS invocations.
 167     */
 168
 169    /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
 170     * lookup table
 171     */
 172    static const uint32_t target_to_register[] = {
 173       IA_VERTICES_COUNT,   /* VERTICES_SUBMITTED */
 174       IA_PRIMITIVES_COUNT, /* PRIMITIVES_SUBMITTED */
 175       VS_INVOCATION_COUNT, /* VERTEX_SHADER_INVOCATIONS */
 176       HS_INVOCATION_COUNT, /* TESS_CONTROL_SHADER_PATCHES */
 177       DS_INVOCATION_COUNT, /* TESS_EVALUATION_SHADER_INVOCATIONS */
 178       GS_PRIMITIVES_COUNT, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
 179       PS_INVOCATION_COUNT, /* FRAGMENT_SHADER_INVOCATIONS */
 180       CS_INVOCATION_COUNT, /* COMPUTE_SHADER_INVOCATIONS */
 181       CL_INVOCATION_COUNT, /* CLIPPING_INPUT_PRIMITIVES */
 182       CL_PRIMITIVES_COUNT, /* CLIPPING_OUTPUT_PRIMITIVES */
 183       GS_INVOCATION_COUNT /* This one is special... */
 184    };
 185    STATIC_ASSERT(ARRAY_SIZE(target_to_register) == MAX_PIPELINE_STATISTICS);
 186    uint32_t reg = target_to_register[pipeline_target_to_index(target)];
 187    /* Gen6 GS code counts full primitives, that is, it won't count individual
 188     * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
 189     */
 190    if (brw->gen == 6 && target == GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB)
 191       reg = CL_INVOCATION_COUNT;
 192    assert(reg != 0);
 193
 194    /* Emit a flush to make sure various parts of the pipeline are complete and
 195     * we get an accurate value
 196     */
 197    brw_emit_mi_flush(brw);
 198
 199    brw_store_register_mem64(brw, bo, reg, idx * sizeof(uint64_t));
 200 }
 201
 202
 203 /**
 204  * Wait on the query object's BO and calculate the final result.
 205  */
 206 static void
 207 gen6_queryobj_get_results(struct gl_context *ctx,
 208                           struct brw_query_object *query)
 209 {
 210    struct brw_context *brw = brw_context(ctx);
 211
 212    if (query->bo == NULL)
 213       return;
 214
 215    drm_bacon_bo_map(query->bo, false);
 216    uint64_t *results = query->bo->virtual;
 217    switch (query->Base.Target) {
 218    case GL_TIME_ELAPSED:
 219       /* The query BO contains the starting and ending timestamps.
 220        * Subtract the two and convert to nanoseconds.
 221        */
 222       query->Base.Result = brw_raw_timestamp_delta(brw, results[0], results[1]);
 223       query->Base.Result = brw_timebase_scale(brw, query->Base.Result);
 224       break;
 225
 226    case GL_TIMESTAMP:
 227       /* The query BO contains a single timestamp value in results[0]. */
 228       query->Base.Result = brw_timebase_scale(brw, results[0]);
 229
 230       /* Ensure the scaled timestamp overflows according to
 231        * GL_QUERY_COUNTER_BITS
 232        */
 233       query->Base.Result &= (1ull << ctx->Const.QueryCounterBits.Timestamp) - 1;
 234       break;
 235
 236    case GL_SAMPLES_PASSED_ARB:
 237       /* We need to use += rather than = here since some BLT-based operations
 238        * may have added additional samples to our occlusion query value.
 239        */
 240       query->Base.Result += results[1] - results[0];
 241       break;
 242
 243    case GL_ANY_SAMPLES_PASSED:
 244    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 245       if (results[0] != results[1])
 246          query->Base.Result = true;
 247       break;
 248
 249    case GL_PRIMITIVES_GENERATED:
 250    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 251    case GL_VERTICES_SUBMITTED_ARB:
 252    case GL_PRIMITIVES_SUBMITTED_ARB:
 253    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 254    case GL_GEOMETRY_SHADER_INVOCATIONS:
 255    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 256    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 257    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 258    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 259    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 260    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 261       query->Base.Result = results[1] - results[0];
 262       break;
 263
 264    case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB:
 265       query->Base.Result = check_xfb_overflow_streams(results, 1);
 266       break;
 267
 268    case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB:
 269       query->Base.Result = check_xfb_overflow_streams(results, MAX_VERTEX_STREAMS);
 270       break;
 271
 272    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 273       query->Base.Result = (results[1] - results[0]);
 274       /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
 275        * "Invocation counter is 4 times actual.  WA: SW to divide HW reported
 276        *  PS Invocations value by 4."
 277        *
 278        * Prior to Haswell, invocation count was counted by the WM, and it
 279        * buggily counted invocations in units of subspans (2x2 unit). To get the
 280        * correct value, the CS multiplied this by 4. With HSW the logic moved,
 281        * and correctly emitted the number of pixel shader invocations, but,
 282        * whomever forgot to undo the multiply by 4.
 283        */
 284       if (brw->gen == 8 || brw->is_haswell)
 285          query->Base.Result /= 4;
 286       break;
 287
 288    default:
 289       unreachable("Unrecognized query target in brw_queryobj_get_results()");
 290    }
 291    drm_bacon_bo_unmap(query->bo);
 292
 293    /* Now that we've processed the data stored in the query's buffer object,
 294     * we can release it.
 295     */
 296    drm_bacon_bo_unreference(query->bo);
 297    query->bo = NULL;
 298
 299    query->Base.Ready = true;
 300 }
 301
 302 /**
 303  * Driver hook for glBeginQuery().
 304  *
 305  * Initializes driver structures and emits any GPU commands required to begin
 306  * recording data for the query.
 307  */
 308 static void
 309 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 310 {
 311    struct brw_context *brw = brw_context(ctx);
 312    struct brw_query_object *query = (struct brw_query_object *)q;
 313
 314    /* Since we're starting a new query, we need to throw away old results. */
 315    drm_bacon_bo_unreference(query->bo);
 316    query->bo = drm_bacon_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
 317
 318    /* For ARB_query_buffer_object: The result is not available */
 319    set_query_availability(brw, query, false);
 320
 321    switch (query->Base.Target) {
 322    case GL_TIME_ELAPSED:
 323       /* For timestamp queries, we record the starting time right away so that
 324        * we measure the full time between BeginQuery and EndQuery.  There's
 325        * some debate about whether this is the right thing to do.  Our decision
 326        * is based on the following text from the ARB_timer_query extension:
 327        *
 328        * "(5) Should the extension measure total time elapsed between the full
 329        *      completion of the BeginQuery and EndQuery commands, or just time
 330        *      spent in the graphics library?
 331        *
 332        *  RESOLVED:  This extension will measure the total time elapsed
 333        *  between the full completion of these commands.  Future extensions
 334        *  may implement a query to determine time elapsed at different stages
 335        *  of the graphics pipeline."
 336        *
 337        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 338        * we'll write a second timestamp (at index 1), and subtract the two to
 339        * obtain the time elapsed.  Notably, this includes time elapsed while
 340        * the system was doing other work, such as running other applications.
 341        */
 342       brw_write_timestamp(brw, query->bo, 0);
 343       break;
 344
 345    case GL_ANY_SAMPLES_PASSED:
 346    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 347    case GL_SAMPLES_PASSED_ARB:
 348       brw_write_depth_count(brw, query->bo, 0);
 349       break;
 350
 351    case GL_PRIMITIVES_GENERATED:
 352       write_primitives_generated(brw, query->bo, query->Base.Stream, 0);
 353       if (query->Base.Stream == 0)
 354          ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD;
 355       break;
 356
 357    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 358       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0);
 359       break;
 360
 361    case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB:
 362       write_xfb_overflow_streams(ctx, query->bo, query->Base.Stream, 1, 0);
 363       break;
 364
 365    case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB:
 366       write_xfb_overflow_streams(ctx, query->bo, 0, MAX_VERTEX_STREAMS, 0);
 367       break;
 368
 369    case GL_VERTICES_SUBMITTED_ARB:
 370    case GL_PRIMITIVES_SUBMITTED_ARB:
 371    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 372    case GL_GEOMETRY_SHADER_INVOCATIONS:
 373    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 374    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 375    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 376    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 377    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 378    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 379    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 380       emit_pipeline_stat(brw, query->bo, query->Base.Stream, query->Base.Target, 0);
 381       break;
 382
 383    default:
 384       unreachable("Unrecognized query target in brw_begin_query()");
 385    }
 386 }
 387
 388 /**
 389  * Driver hook for glEndQuery().
 390  *
 391  * Emits GPU commands to record a final query value, ending any data capturing.
 392  * However, the final result isn't necessarily available until the GPU processes
 393  * those commands.  brw_queryobj_get_results() processes the captured data to
 394  * produce the final result.
 395  */
 396 static void
 397 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 398 {
 399    struct brw_context *brw = brw_context(ctx);
 400    struct brw_query_object *query = (struct brw_query_object *)q;
 401
 402    switch (query->Base.Target) {
 403    case GL_TIME_ELAPSED:
 404       brw_write_timestamp(brw, query->bo, 1);
 405       break;
 406
 407    case GL_ANY_SAMPLES_PASSED:
 408    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 409    case GL_SAMPLES_PASSED_ARB:
 410       brw_write_depth_count(brw, query->bo, 1);
 411       break;
 412
 413    case GL_PRIMITIVES_GENERATED:
 414       write_primitives_generated(brw, query->bo, query->Base.Stream, 1);
 415       if (query->Base.Stream == 0)
 416          ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD;
 417       break;
 418
 419    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 420       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1);
 421       break;
 422
 423    case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB:
 424       write_xfb_overflow_streams(ctx, query->bo, query->Base.Stream, 1, 1);
 425       break;
 426
 427    case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB:
 428       write_xfb_overflow_streams(ctx, query->bo, 0, MAX_VERTEX_STREAMS, 1);
 429       break;
 430
 431       /* calculate overflow here */
 432    case GL_VERTICES_SUBMITTED_ARB:
 433    case GL_PRIMITIVES_SUBMITTED_ARB:
 434    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 435    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 436    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 437    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 438    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 439    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 440    case GL_GEOMETRY_SHADER_INVOCATIONS:
 441    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 442    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 443       emit_pipeline_stat(brw, query->bo,
 444                          query->Base.Stream, query->Base.Target, 1);
 445       break;
 446
 447    default:
 448       unreachable("Unrecognized query target in brw_end_query()");
 449    }
 450
 451    /* The current batch contains the commands to handle EndQuery(),
 452     * but they won't actually execute until it is flushed.
 453     */
 454    query->flushed = false;
 455
 456    /* For ARB_query_buffer_object: The result is now available */
 457    set_query_availability(brw, query, true);
 458 }
 459
 460 /**
 461  * Flush the batch if it still references the query object BO.
 462  */
 463 static void
 464 flush_batch_if_needed(struct brw_context *brw, struct brw_query_object *query)
 465 {
 466    /* If the batch doesn't reference the BO, it must have been flushed
 467     * (for example, due to being full).  Record that it's been flushed.
 468     */
 469    query->flushed = query->flushed ||
 470                     !brw_batch_references(&brw->batch, query->bo);
 471
 472    if (!query->flushed)
 473       intel_batchbuffer_flush(brw);
 474 }
 475
 476 /**
 477  * The WaitQuery() driver hook.
 478  *
 479  * Wait for a query result to become available and return it.  This is the
 480  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 481  */
 482 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 483 {
 484    struct brw_context *brw = brw_context(ctx);
 485    struct brw_query_object *query = (struct brw_query_object *)q;
 486
 487    /* If the application has requested the query result, but this batch is
 488     * still contributing to it, flush it now to finish that work so the
 489     * result will become available (eventually).
 490     */
 491    flush_batch_if_needed(brw, query);
 492
 493    gen6_queryobj_get_results(ctx, query);
 494 }
 495
 496 /**
 497  * The CheckQuery() driver hook.
 498  *
 499  * Checks whether a query result is ready yet.  If not, flushes.
 500  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 501  */
 502 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
 503 {
 504    struct brw_context *brw = brw_context(ctx);
 505    struct brw_query_object *query = (struct brw_query_object *)q;
 506
 507    /* If query->bo is NULL, we've already gathered the results - this is a
 508     * redundant CheckQuery call.  Ignore it.
 509     */
 510    if (query->bo == NULL)
 511       return;
 512
 513    /* From the GL_ARB_occlusion_query spec:
 514     *
 515     *     "Instead of allowing for an infinite loop, performing a
 516     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 517     *      not ready yet on the first time it is queried.  This ensures that
 518     *      the async query will return true in finite time.
 519     */
 520    flush_batch_if_needed(brw, query);
 521
 522    if (!drm_bacon_bo_busy(query->bo)) {
 523       gen6_queryobj_get_results(ctx, query);
 524    }
 525 }
 526
 527 static void
 528 gen6_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 529 {
 530    struct brw_context *brw = brw_context(ctx);
 531    struct brw_query_object *query = (struct brw_query_object *)q;
 532    brw_query_counter(ctx, q);
 533    set_query_availability(brw, query, true);
 534 }
 535
 536 /* Initialize Gen6+-specific query object functions. */
 537 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 538 {
 539    functions->BeginQuery = gen6_begin_query;
 540    functions->EndQuery = gen6_end_query;
 541    functions->CheckQuery = gen6_check_query;
 542    functions->WaitQuery = gen6_wait_query;
 543    functions->QueryCounter = gen6_query_counter;
 544 }