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 #include "intel_reg.h"
  42
  43 static inline void
  44 set_query_availability(struct brw_context *brw, struct brw_query_object *query,
  45                        bool available)
  46 {
  47    /* For platforms that support ARB_query_buffer_object, we write the
  48     * query availability for "pipelined" queries.
  49     *
  50     * Most counter snapshots are written by the command streamer, by
  51     * doing a CS stall and then MI_STORE_REGISTER_MEM.  For these
  52     * counters, the CS stall guarantees that the results will be
  53     * available when subsequent CS commands run.  So we don't need to
  54     * do any additional tracking.
  55     *
  56     * Other counters (occlusion queries and timestamp) are written by
  57     * PIPE_CONTROL, without a CS stall.  This means that we can't be
  58     * sure whether the writes have landed yet or not.  Performing a
  59     * PIPE_CONTROL with an immediate write will synchronize with
  60     * those earlier writes, so we write 1 when the value has landed.
  61     */
  62    if (brw->ctx.Extensions.ARB_query_buffer_object &&
  63        brw_is_query_pipelined(query)) {
  64       brw_emit_pipe_control_write(brw,
  65                                   PIPE_CONTROL_WRITE_IMMEDIATE,
  66                                   query->bo, 2 * sizeof(uint64_t),
  67                                   available, 0);
  68    }
  69 }
  70
  71 static void
  72 write_primitives_generated(struct brw_context *brw,
  73                            drm_intel_bo *query_bo, int stream, int idx)
  74 {
  75    brw_emit_mi_flush(brw);
  76
  77    if (brw->gen >= 7 && stream > 0) {
  78       brw_store_register_mem64(brw, query_bo,
  79                                GEN7_SO_PRIM_STORAGE_NEEDED(stream),
  80                                idx * sizeof(uint64_t));
  81    } else {
  82       brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT,
  83                                idx * sizeof(uint64_t));
  84    }
  85 }
  86
  87 static void
  88 write_xfb_primitives_written(struct brw_context *brw,
  89                              drm_intel_bo *bo, int stream, int idx)
  90 {
  91    brw_emit_mi_flush(brw);
  92
  93    if (brw->gen >= 7) {
  94       brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(stream),
  95                                idx * sizeof(uint64_t));
  96    } else {
  97       brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN,
  98                                idx * sizeof(uint64_t));
  99    }
 100 }
 101
 102 static inline const int
 103 pipeline_target_to_index(int target)
 104 {
 105    if (target == GL_GEOMETRY_SHADER_INVOCATIONS)
 106       return MAX_PIPELINE_STATISTICS - 1;
 107    else
 108       return target - GL_VERTICES_SUBMITTED_ARB;
 109 }
 110
 111 static void
 112 emit_pipeline_stat(struct brw_context *brw, drm_intel_bo *bo,
 113                    int stream, int target, int idx)
 114 {
 115    /* One source of confusion is the tessellation shader statistics. The
 116     * hardware has no statistics specific to the TE unit. Ideally we could have
 117     * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
 118     * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
 119     * Unfortunately we don't have HS primitives, we only have HS invocations.
 120     */
 121
 122    /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
 123     * lookup table
 124     */
 125    static const uint32_t target_to_register[] = {
 126       IA_VERTICES_COUNT,   /* VERTICES_SUBMITTED */
 127       IA_PRIMITIVES_COUNT, /* PRIMITIVES_SUBMITTED */
 128       VS_INVOCATION_COUNT, /* VERTEX_SHADER_INVOCATIONS */
 129       HS_INVOCATION_COUNT, /* TESS_CONTROL_SHADER_PATCHES */
 130       DS_INVOCATION_COUNT, /* TESS_EVALUATION_SHADER_INVOCATIONS */
 131       GS_PRIMITIVES_COUNT, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
 132       PS_INVOCATION_COUNT, /* FRAGMENT_SHADER_INVOCATIONS */
 133       CS_INVOCATION_COUNT, /* COMPUTE_SHADER_INVOCATIONS */
 134       CL_INVOCATION_COUNT, /* CLIPPING_INPUT_PRIMITIVES */
 135       CL_PRIMITIVES_COUNT, /* CLIPPING_OUTPUT_PRIMITIVES */
 136       GS_INVOCATION_COUNT /* This one is special... */
 137    };
 138    STATIC_ASSERT(ARRAY_SIZE(target_to_register) == MAX_PIPELINE_STATISTICS);
 139    uint32_t reg = target_to_register[pipeline_target_to_index(target)];
 140    /* Gen6 GS code counts full primitives, that is, it won't count individual
 141     * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
 142     */
 143    if (brw->gen == 6 && target == GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB)
 144       reg = CL_INVOCATION_COUNT;
 145    assert(reg != 0);
 146
 147    /* Emit a flush to make sure various parts of the pipeline are complete and
 148     * we get an accurate value
 149     */
 150    brw_emit_mi_flush(brw);
 151
 152    brw_store_register_mem64(brw, bo, reg, idx * sizeof(uint64_t));
 153 }
 154
 155
 156 /**
 157  * Wait on the query object's BO and calculate the final result.
 158  */
 159 static void
 160 gen6_queryobj_get_results(struct gl_context *ctx,
 161                           struct brw_query_object *query)
 162 {
 163    struct brw_context *brw = brw_context(ctx);
 164
 165    if (query->bo == NULL)
 166       return;
 167
 168    brw_bo_map(brw, query->bo, false, "query object");
 169    uint64_t *results = query->bo->virtual;
 170    switch (query->Base.Target) {
 171    case GL_TIME_ELAPSED:
 172       /* The query BO contains the starting and ending timestamps.
 173        * Subtract the two and convert to nanoseconds.
 174        */
 175       query->Base.Result += 80 * (results[1] - results[0]);
 176       break;
 177
 178    case GL_TIMESTAMP:
 179       /* Our timer is a clock that increments every 80ns (regardless of
 180        * other clock scaling in the system).  The timestamp register we can
 181        * read for glGetTimestamp() masks out the top 32 bits, so we do that
 182        * here too to let the two counters be compared against each other.
 183        *
 184        * If we just multiplied that 32 bits of data by 80, it would roll
 185        * over at a non-power-of-two, so an application couldn't use
 186        * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
 187        * report 36 bits and truncate at that (rolling over 5 times as often
 188        * as the HW counter), and when the 32-bit counter rolls over, it
 189        * happens to also be at a rollover in the reported value from near
 190        * (1<<36) to 0.
 191        *
 192        * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
 193        * rolls over every ~69 seconds.
 194        *
 195        * The query BO contains a single timestamp value in results[0].
 196        */
 197       query->Base.Result = 80 * (results[0] & 0xffffffff);
 198       query->Base.Result &= (1ull << 36) - 1;
 199       break;
 200
 201    case GL_SAMPLES_PASSED_ARB:
 202       /* We need to use += rather than = here since some BLT-based operations
 203        * may have added additional samples to our occlusion query value.
 204        */
 205       query->Base.Result += results[1] - results[0];
 206       break;
 207
 208    case GL_ANY_SAMPLES_PASSED:
 209    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 210       if (results[0] != results[1])
 211          query->Base.Result = true;
 212       break;
 213
 214    case GL_PRIMITIVES_GENERATED:
 215    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 216    case GL_VERTICES_SUBMITTED_ARB:
 217    case GL_PRIMITIVES_SUBMITTED_ARB:
 218    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 219    case GL_GEOMETRY_SHADER_INVOCATIONS:
 220    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 221    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 222    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 223    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 224    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 225    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 226       query->Base.Result = results[1] - results[0];
 227       break;
 228
 229    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 230       query->Base.Result = (results[1] - results[0]);
 231       /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
 232        * "Invocation counter is 4 times actual.  WA: SW to divide HW reported
 233        *  PS Invocations value by 4."
 234        *
 235        * Prior to Haswell, invocation count was counted by the WM, and it
 236        * buggily counted invocations in units of subspans (2x2 unit). To get the
 237        * correct value, the CS multiplied this by 4. With HSW the logic moved,
 238        * and correctly emitted the number of pixel shader invocations, but,
 239        * whomever forgot to undo the multiply by 4.
 240        */
 241       if (brw->gen == 8 || brw->is_haswell)
 242          query->Base.Result /= 4;
 243       break;
 244
 245    default:
 246       unreachable("Unrecognized query target in brw_queryobj_get_results()");
 247    }
 248    drm_intel_bo_unmap(query->bo);
 249
 250    /* Now that we've processed the data stored in the query's buffer object,
 251     * we can release it.
 252     */
 253    drm_intel_bo_unreference(query->bo);
 254    query->bo = NULL;
 255
 256    query->Base.Ready = true;
 257 }
 258
 259 /**
 260  * Driver hook for glBeginQuery().
 261  *
 262  * Initializes driver structures and emits any GPU commands required to begin
 263  * recording data for the query.
 264  */
 265 static void
 266 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 267 {
 268    struct brw_context *brw = brw_context(ctx);
 269    struct brw_query_object *query = (struct brw_query_object *)q;
 270
 271    /* Since we're starting a new query, we need to throw away old results. */
 272    drm_intel_bo_unreference(query->bo);
 273    query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
 274
 275    /* For ARB_query_buffer_object: The result is not available */
 276    set_query_availability(brw, query, false);
 277
 278    switch (query->Base.Target) {
 279    case GL_TIME_ELAPSED:
 280       /* For timestamp queries, we record the starting time right away so that
 281        * we measure the full time between BeginQuery and EndQuery.  There's
 282        * some debate about whether this is the right thing to do.  Our decision
 283        * is based on the following text from the ARB_timer_query extension:
 284        *
 285        * "(5) Should the extension measure total time elapsed between the full
 286        *      completion of the BeginQuery and EndQuery commands, or just time
 287        *      spent in the graphics library?
 288        *
 289        *  RESOLVED:  This extension will measure the total time elapsed
 290        *  between the full completion of these commands.  Future extensions
 291        *  may implement a query to determine time elapsed at different stages
 292        *  of the graphics pipeline."
 293        *
 294        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 295        * we'll write a second timestamp (at index 1), and subtract the two to
 296        * obtain the time elapsed.  Notably, this includes time elapsed while
 297        * the system was doing other work, such as running other applications.
 298        */
 299       brw_write_timestamp(brw, query->bo, 0);
 300       break;
 301
 302    case GL_ANY_SAMPLES_PASSED:
 303    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 304    case GL_SAMPLES_PASSED_ARB:
 305       brw_write_depth_count(brw, query->bo, 0);
 306       break;
 307
 308    case GL_PRIMITIVES_GENERATED:
 309       write_primitives_generated(brw, query->bo, query->Base.Stream, 0);
 310       if (query->Base.Stream == 0)
 311          ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD;
 312       break;
 313
 314    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 315       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0);
 316       break;
 317
 318    case GL_VERTICES_SUBMITTED_ARB:
 319    case GL_PRIMITIVES_SUBMITTED_ARB:
 320    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 321    case GL_GEOMETRY_SHADER_INVOCATIONS:
 322    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 323    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 324    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 325    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 326    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 327    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 328    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 329       emit_pipeline_stat(brw, query->bo, query->Base.Stream, query->Base.Target, 0);
 330       break;
 331
 332    default:
 333       unreachable("Unrecognized query target in brw_begin_query()");
 334    }
 335 }
 336
 337 /**
 338  * Driver hook for glEndQuery().
 339  *
 340  * Emits GPU commands to record a final query value, ending any data capturing.
 341  * However, the final result isn't necessarily available until the GPU processes
 342  * those commands.  brw_queryobj_get_results() processes the captured data to
 343  * produce the final result.
 344  */
 345 static void
 346 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 347 {
 348    struct brw_context *brw = brw_context(ctx);
 349    struct brw_query_object *query = (struct brw_query_object *)q;
 350
 351    switch (query->Base.Target) {
 352    case GL_TIME_ELAPSED:
 353       brw_write_timestamp(brw, query->bo, 1);
 354       break;
 355
 356    case GL_ANY_SAMPLES_PASSED:
 357    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 358    case GL_SAMPLES_PASSED_ARB:
 359       brw_write_depth_count(brw, query->bo, 1);
 360       break;
 361
 362    case GL_PRIMITIVES_GENERATED:
 363       write_primitives_generated(brw, query->bo, query->Base.Stream, 1);
 364       if (query->Base.Stream == 0)
 365          ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD;
 366       break;
 367
 368    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 369       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1);
 370       break;
 371
 372    case GL_VERTICES_SUBMITTED_ARB:
 373    case GL_PRIMITIVES_SUBMITTED_ARB:
 374    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 375    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 376    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 377    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 378    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 379    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 380    case GL_GEOMETRY_SHADER_INVOCATIONS:
 381    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 382    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 383       emit_pipeline_stat(brw, query->bo,
 384                          query->Base.Stream, query->Base.Target, 1);
 385       break;
 386
 387    default:
 388       unreachable("Unrecognized query target in brw_end_query()");
 389    }
 390
 391    /* The current batch contains the commands to handle EndQuery(),
 392     * but they won't actually execute until it is flushed.
 393     */
 394    query->flushed = false;
 395
 396    /* For ARB_query_buffer_object: The result is now available */
 397    set_query_availability(brw, query, true);
 398 }
 399
 400 /**
 401  * Flush the batch if it still references the query object BO.
 402  */
 403 static void
 404 flush_batch_if_needed(struct brw_context *brw, struct brw_query_object *query)
 405 {
 406    /* If the batch doesn't reference the BO, it must have been flushed
 407     * (for example, due to being full).  Record that it's been flushed.
 408     */
 409    query->flushed = query->flushed ||
 410       !drm_intel_bo_references(brw->batch.bo, query->bo);
 411
 412    if (!query->flushed)
 413       intel_batchbuffer_flush(brw);
 414 }
 415
 416 /**
 417  * The WaitQuery() driver hook.
 418  *
 419  * Wait for a query result to become available and return it.  This is the
 420  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 421  */
 422 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 423 {
 424    struct brw_context *brw = brw_context(ctx);
 425    struct brw_query_object *query = (struct brw_query_object *)q;
 426
 427    /* If the application has requested the query result, but this batch is
 428     * still contributing to it, flush it now to finish that work so the
 429     * result will become available (eventually).
 430     */
 431    flush_batch_if_needed(brw, query);
 432
 433    gen6_queryobj_get_results(ctx, query);
 434 }
 435
 436 /**
 437  * The CheckQuery() driver hook.
 438  *
 439  * Checks whether a query result is ready yet.  If not, flushes.
 440  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 441  */
 442 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
 443 {
 444    struct brw_context *brw = brw_context(ctx);
 445    struct brw_query_object *query = (struct brw_query_object *)q;
 446
 447    /* If query->bo is NULL, we've already gathered the results - this is a
 448     * redundant CheckQuery call.  Ignore it.
 449     */
 450    if (query->bo == NULL)
 451       return;
 452
 453    /* From the GL_ARB_occlusion_query spec:
 454     *
 455     *     "Instead of allowing for an infinite loop, performing a
 456     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 457     *      not ready yet on the first time it is queried.  This ensures that
 458     *      the async query will return true in finite time.
 459     */
 460    flush_batch_if_needed(brw, query);
 461
 462    if (!drm_intel_bo_busy(query->bo)) {
 463       gen6_queryobj_get_results(ctx, query);
 464    }
 465 }
 466
 467 static void
 468 gen6_query_counter(struct gl_context *ctx, struct gl_query_object *q)
 469 {
 470    struct brw_context *brw = brw_context(ctx);
 471    struct brw_query_object *query = (struct brw_query_object *)q;
 472    brw_query_counter(ctx, q);
 473    set_query_availability(brw, query, true);
 474 }
 475
 476 /* Initialize Gen6+-specific query object functions. */
 477 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 478 {
 479    functions->BeginQuery = gen6_begin_query;
 480    functions->EndQuery = gen6_end_query;
 481    functions->CheckQuery = gen6_check_query;
 482    functions->WaitQuery = gen6_wait_query;
 483    functions->QueryCounter = gen6_query_counter;
 484 }