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  * Write an arbitrary 64-bit register to a buffer via MI_STORE_REGISTER_MEM.
  44  *
  45  * Only TIMESTAMP and PS_DEPTH_COUNT have special PIPE_CONTROL support; other
  46  * counters have to be read via the generic MI_STORE_REGISTER_MEM.
  47  *
  48  * Callers must explicitly flush the pipeline to ensure the desired value is
  49  * available.
  50  */
  51 void
  52 brw_store_register_mem64(struct brw_context *brw,
  53                          drm_intel_bo *bo, uint32_t reg, int idx)
  54 {
  55    assert(brw->gen >= 6);
  56
  57    /* MI_STORE_REGISTER_MEM only stores a single 32-bit value, so to
  58     * read a full 64-bit register, we need to do two of them.
  59     */
  60    if (brw->gen >= 8) {
  61       BEGIN_BATCH(8);
  62       OUT_BATCH(MI_STORE_REGISTER_MEM | (4 - 2));
  63       OUT_BATCH(reg);
  64       OUT_RELOC64(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  65                   idx * sizeof(uint64_t));
  66       OUT_BATCH(MI_STORE_REGISTER_MEM | (4 - 2));
  67       OUT_BATCH(reg + sizeof(uint32_t));
  68       OUT_RELOC64(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  69                   sizeof(uint32_t) + idx * sizeof(uint64_t));
  70       ADVANCE_BATCH();
  71    } else {
  72       BEGIN_BATCH(6);
  73       OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
  74       OUT_BATCH(reg);
  75       OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  76                 idx * sizeof(uint64_t));
  77       OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
  78       OUT_BATCH(reg + sizeof(uint32_t));
  79       OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  80                 sizeof(uint32_t) + idx * sizeof(uint64_t));
  81       ADVANCE_BATCH();
  82    }
  83 }
  84
  85 static void
  86 write_primitives_generated(struct brw_context *brw,
  87                            drm_intel_bo *query_bo, int stream, int idx)
  88 {
  89    intel_batchbuffer_emit_mi_flush(brw);
  90
  91    if (brw->gen >= 7 && stream > 0) {
  92       brw_store_register_mem64(brw, query_bo,
  93                                GEN7_SO_PRIM_STORAGE_NEEDED(stream), idx);
  94    } else {
  95       brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT, idx);
  96    }
  97 }
  98
  99 static void
 100 write_xfb_primitives_written(struct brw_context *brw,
 101                              drm_intel_bo *bo, int stream, int idx)
 102 {
 103    intel_batchbuffer_emit_mi_flush(brw);
 104
 105    if (brw->gen >= 7) {
 106       brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(stream), idx);
 107    } else {
 108       brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN, idx);
 109    }
 110 }
 111
 112 static inline const int
 113 pipeline_target_to_index(int target)
 114 {
 115    if (target == GL_GEOMETRY_SHADER_INVOCATIONS)
 116       return MAX_PIPELINE_STATISTICS - 1;
 117    else
 118       return target - GL_VERTICES_SUBMITTED_ARB;
 119 }
 120
 121 static void
 122 emit_pipeline_stat(struct brw_context *brw, drm_intel_bo *bo,
 123                    int stream, int target, int idx)
 124 {
 125    /* One source of confusion is the tessellation shader statistics. The
 126     * hardware has no statistics specific to the TE unit. Ideally we could have
 127     * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
 128     * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
 129     * Unfortunately we don't have HS primitives, we only have HS invocations.
 130     */
 131
 132    /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
 133     * lookup table
 134     */
 135    static const uint32_t target_to_register[] = {
 136       IA_VERTICES_COUNT,   /* VERTICES_SUBMITTED */
 137       IA_PRIMITIVES_COUNT, /* PRIMITIVES_SUBMITTED */
 138       VS_INVOCATION_COUNT, /* VERTEX_SHADER_INVOCATIONS */
 139       0, /* HS_INVOCATION_COUNT,*/  /* TESS_CONTROL_SHADER_PATCHES */
 140       0, /* DS_INVOCATION_COUNT,*/  /* TESS_EVALUATION_SHADER_INVOCATIONS */
 141       GS_PRIMITIVES_COUNT, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
 142       PS_INVOCATION_COUNT, /* FRAGMENT_SHADER_INVOCATIONS */
 143       CS_INVOCATION_COUNT, /* COMPUTE_SHADER_INVOCATIONS */
 144       CL_INVOCATION_COUNT, /* CLIPPING_INPUT_PRIMITIVES */
 145       CL_PRIMITIVES_COUNT, /* CLIPPING_OUTPUT_PRIMITIVES */
 146       GS_INVOCATION_COUNT /* This one is special... */
 147    };
 148    STATIC_ASSERT(ARRAY_SIZE(target_to_register) == MAX_PIPELINE_STATISTICS);
 149    uint32_t reg = target_to_register[pipeline_target_to_index(target)];
 150    /* Gen6 GS code counts full primitives, that is, it won't count individual
 151     * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
 152     */
 153    if (brw->gen == 6 && target == GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB)
 154       reg = CL_INVOCATION_COUNT;
 155    assert(reg != 0);
 156
 157    /* Emit a flush to make sure various parts of the pipeline are complete and
 158     * we get an accurate value
 159     */
 160    intel_batchbuffer_emit_mi_flush(brw);
 161
 162    brw_store_register_mem64(brw, bo, reg, idx);
 163 }
 164
 165
 166 /**
 167  * Wait on the query object's BO and calculate the final result.
 168  */
 169 static void
 170 gen6_queryobj_get_results(struct gl_context *ctx,
 171                           struct brw_query_object *query)
 172 {
 173    struct brw_context *brw = brw_context(ctx);
 174
 175    if (query->bo == NULL)
 176       return;
 177
 178    brw_bo_map(brw, query->bo, false, "query object");
 179    uint64_t *results = query->bo->virtual;
 180    switch (query->Base.Target) {
 181    case GL_TIME_ELAPSED:
 182       /* The query BO contains the starting and ending timestamps.
 183        * Subtract the two and convert to nanoseconds.
 184        */
 185       query->Base.Result += 80 * (results[1] - results[0]);
 186       break;
 187
 188    case GL_TIMESTAMP:
 189       /* Our timer is a clock that increments every 80ns (regardless of
 190        * other clock scaling in the system).  The timestamp register we can
 191        * read for glGetTimestamp() masks out the top 32 bits, so we do that
 192        * here too to let the two counters be compared against each other.
 193        *
 194        * If we just multiplied that 32 bits of data by 80, it would roll
 195        * over at a non-power-of-two, so an application couldn't use
 196        * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
 197        * report 36 bits and truncate at that (rolling over 5 times as often
 198        * as the HW counter), and when the 32-bit counter rolls over, it
 199        * happens to also be at a rollover in the reported value from near
 200        * (1<<36) to 0.
 201        *
 202        * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
 203        * rolls over every ~69 seconds.
 204        *
 205        * The query BO contains a single timestamp value in results[0].
 206        */
 207       query->Base.Result = 80 * (results[0] & 0xffffffff);
 208       query->Base.Result &= (1ull << 36) - 1;
 209       break;
 210
 211    case GL_SAMPLES_PASSED_ARB:
 212       /* We need to use += rather than = here since some BLT-based operations
 213        * may have added additional samples to our occlusion query value.
 214        */
 215       query->Base.Result += results[1] - results[0];
 216       break;
 217
 218    case GL_ANY_SAMPLES_PASSED:
 219    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 220       if (results[0] != results[1])
 221          query->Base.Result = true;
 222       break;
 223
 224    case GL_PRIMITIVES_GENERATED:
 225    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 226    case GL_VERTICES_SUBMITTED_ARB:
 227    case GL_PRIMITIVES_SUBMITTED_ARB:
 228    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 229    case GL_GEOMETRY_SHADER_INVOCATIONS:
 230    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 231    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 232    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 233    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 234       query->Base.Result = results[1] - results[0];
 235       break;
 236
 237    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 238       query->Base.Result = (results[1] - results[0]);
 239       /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
 240        * "Invocation counter is 4 times actual.  WA: SW to divide HW reported
 241        *  PS Invocations value by 4."
 242        *
 243        * Prior to Haswell, invocation count was counted by the WM, and it
 244        * buggily counted invocations in units of subspans (2x2 unit). To get the
 245        * correct value, the CS multiplied this by 4. With HSW the logic moved,
 246        * and correctly emitted the number of pixel shader invocations, but,
 247        * whomever forgot to undo the multiply by 4.
 248        */
 249       if (brw->gen == 8 || brw->is_haswell)
 250          query->Base.Result /= 4;
 251       break;
 252
 253    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 254    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 255    default:
 256       unreachable("Unrecognized query target in brw_queryobj_get_results()");
 257    }
 258    drm_intel_bo_unmap(query->bo);
 259
 260    /* Now that we've processed the data stored in the query's buffer object,
 261     * we can release it.
 262     */
 263    drm_intel_bo_unreference(query->bo);
 264    query->bo = NULL;
 265
 266    query->Base.Ready = true;
 267 }
 268
 269 /**
 270  * Driver hook for glBeginQuery().
 271  *
 272  * Initializes driver structures and emits any GPU commands required to begin
 273  * recording data for the query.
 274  */
 275 static void
 276 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 277 {
 278    struct brw_context *brw = brw_context(ctx);
 279    struct brw_query_object *query = (struct brw_query_object *)q;
 280
 281    /* Since we're starting a new query, we need to throw away old results. */
 282    drm_intel_bo_unreference(query->bo);
 283    query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
 284
 285    switch (query->Base.Target) {
 286    case GL_TIME_ELAPSED:
 287       /* For timestamp queries, we record the starting time right away so that
 288        * we measure the full time between BeginQuery and EndQuery.  There's
 289        * some debate about whether this is the right thing to do.  Our decision
 290        * is based on the following text from the ARB_timer_query extension:
 291        *
 292        * "(5) Should the extension measure total time elapsed between the full
 293        *      completion of the BeginQuery and EndQuery commands, or just time
 294        *      spent in the graphics library?
 295        *
 296        *  RESOLVED:  This extension will measure the total time elapsed
 297        *  between the full completion of these commands.  Future extensions
 298        *  may implement a query to determine time elapsed at different stages
 299        *  of the graphics pipeline."
 300        *
 301        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 302        * we'll write a second timestamp (at index 1), and subtract the two to
 303        * obtain the time elapsed.  Notably, this includes time elapsed while
 304        * the system was doing other work, such as running other applications.
 305        */
 306       brw_write_timestamp(brw, query->bo, 0);
 307       break;
 308
 309    case GL_ANY_SAMPLES_PASSED:
 310    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 311    case GL_SAMPLES_PASSED_ARB:
 312       brw_write_depth_count(brw, query->bo, 0);
 313       break;
 314
 315    case GL_PRIMITIVES_GENERATED:
 316       write_primitives_generated(brw, query->bo, query->Base.Stream, 0);
 317       break;
 318
 319    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 320       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0);
 321       break;
 322
 323    case GL_VERTICES_SUBMITTED_ARB:
 324    case GL_PRIMITIVES_SUBMITTED_ARB:
 325    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 326    case GL_GEOMETRY_SHADER_INVOCATIONS:
 327    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 328    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 329    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 330    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 331    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 332       emit_pipeline_stat(brw, query->bo, query->Base.Stream, query->Base.Target, 0);
 333       break;
 334
 335    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 336    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 337    default:
 338       unreachable("Unrecognized query target in brw_begin_query()");
 339    }
 340 }
 341
 342 /**
 343  * Driver hook for glEndQuery().
 344  *
 345  * Emits GPU commands to record a final query value, ending any data capturing.
 346  * However, the final result isn't necessarily available until the GPU processes
 347  * those commands.  brw_queryobj_get_results() processes the captured data to
 348  * produce the final result.
 349  */
 350 static void
 351 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 352 {
 353    struct brw_context *brw = brw_context(ctx);
 354    struct brw_query_object *query = (struct brw_query_object *)q;
 355
 356    switch (query->Base.Target) {
 357    case GL_TIME_ELAPSED:
 358       brw_write_timestamp(brw, query->bo, 1);
 359       break;
 360
 361    case GL_ANY_SAMPLES_PASSED:
 362    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 363    case GL_SAMPLES_PASSED_ARB:
 364       brw_write_depth_count(brw, query->bo, 1);
 365       break;
 366
 367    case GL_PRIMITIVES_GENERATED:
 368       write_primitives_generated(brw, query->bo, query->Base.Stream, 1);
 369       break;
 370
 371    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 372       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1);
 373       break;
 374
 375    case GL_VERTICES_SUBMITTED_ARB:
 376    case GL_PRIMITIVES_SUBMITTED_ARB:
 377    case GL_VERTEX_SHADER_INVOCATIONS_ARB:
 378    case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
 379    case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
 380    case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
 381    case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
 382    case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
 383    case GL_GEOMETRY_SHADER_INVOCATIONS:
 384       emit_pipeline_stat(brw, query->bo,
 385                          query->Base.Stream, query->Base.Target, 1);
 386       break;
 387
 388    case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
 389    case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
 390    default:
 391       unreachable("Unrecognized query target in brw_end_query()");
 392    }
 393
 394    /* The current batch contains the commands to handle EndQuery(),
 395     * but they won't actually execute until it is flushed.
 396     */
 397    query->flushed = false;
 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 /* Initialize Gen6+-specific query object functions. */
 468 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 469 {
 470    functions->BeginQuery = gen6_begin_query;
 471    functions->EndQuery = gen6_end_query;
 472    functions->CheckQuery = gen6_check_query;
 473    functions->WaitQuery = gen6_wait_query;
 474 }