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 brw_context *brw, drm_intel_bo *query_bo, int idx)
  47 {
  48    struct intel_context *intel = &brw->intel;
  49    /* Emit workaround flushes: */
  50    if (intel->gen == 6) {
  51       /* The timestamp write below is a non-zero post-sync op, which on
  52        * Gen6 necessitates a CS stall.  CS stalls need stall at scoreboard
  53        * set.  See the comments for intel_emit_post_sync_nonzero_flush().
  54        */
  55       BEGIN_BATCH(4);
  56       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
  57       OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
  58       OUT_BATCH(0);
  59       OUT_BATCH(0);
  60       ADVANCE_BATCH();
  61    }
  62
  63    BEGIN_BATCH(5);
  64    OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  65    OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
  66    OUT_RELOC(query_bo,
  67              I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  68              PIPE_CONTROL_GLOBAL_GTT_WRITE |
  69              idx * sizeof(uint64_t));
  70    OUT_BATCH(0);
  71    OUT_BATCH(0);
  72    ADVANCE_BATCH();
  73 }
  74
  75 /**
  76  * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
  77  */
  78 static void
  79 write_depth_count(struct brw_context *brw, drm_intel_bo *query_bo, int idx)
  80 {
  81    struct intel_context *intel = &brw->intel;
  82    /* Emit Sandybridge workaround flush: */
  83    if (intel->gen == 6)
  84       intel_emit_post_sync_nonzero_flush(brw);
  85
  86    BEGIN_BATCH(5);
  87    OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
  88    OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
  89              PIPE_CONTROL_WRITE_DEPTH_COUNT);
  90    OUT_RELOC(query_bo,
  91              I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  92              PIPE_CONTROL_GLOBAL_GTT_WRITE |
  93              (idx * sizeof(uint64_t)));
  94    OUT_BATCH(0);
  95    OUT_BATCH(0);
  96    ADVANCE_BATCH();
  97 }
  98
  99 /*
 100  * Write an arbitrary 64-bit register to a buffer via MI_STORE_REGISTER_MEM.
 101  *
 102  * Only TIMESTAMP and PS_DEPTH_COUNT have special PIPE_CONTROL support; other
 103  * counters have to be read via the generic MI_STORE_REGISTER_MEM.  This
 104  * function also performs a pipeline flush for proper synchronization.
 105  */
 106 static void
 107 write_reg(struct brw_context *brw,
 108           drm_intel_bo *query_bo, uint32_t reg, int idx)
 109 {
 110    struct intel_context *intel = &brw->intel;
 111    assert(intel->gen >= 6);
 112
 113    intel_batchbuffer_emit_mi_flush(brw);
 114
 115    /* MI_STORE_REGISTER_MEM only stores a single 32-bit value, so to
 116     * read a full 64-bit register, we need to do two of them.
 117     */
 118    BEGIN_BATCH(3);
 119    OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
 120    OUT_BATCH(reg);
 121    OUT_RELOC(query_bo, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
 122              idx * sizeof(uint64_t));
 123    ADVANCE_BATCH();
 124
 125    BEGIN_BATCH(3);
 126    OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
 127    OUT_BATCH(reg + sizeof(uint32_t));
 128    OUT_RELOC(query_bo, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
 129              sizeof(uint32_t) + idx * sizeof(uint64_t));
 130    ADVANCE_BATCH();
 131 }
 132
 133 static void
 134 write_primitives_generated(struct brw_context *brw,
 135                            drm_intel_bo *query_bo, int idx)
 136 {
 137    write_reg(brw, query_bo, CL_INVOCATION_COUNT, idx);
 138 }
 139
 140 static void
 141 write_xfb_primitives_written(struct brw_context *brw,
 142                              drm_intel_bo *query_bo, int idx)
 143 {
 144    struct intel_context *intel = &brw->intel;
 145    if (intel->gen >= 7) {
 146       write_reg(brw, query_bo, SO_NUM_PRIMS_WRITTEN0_IVB, idx);
 147    } else {
 148       write_reg(brw, query_bo, SO_NUM_PRIMS_WRITTEN, idx);
 149    }
 150 }
 151
 152 /**
 153  * Wait on the query object's BO and calculate the final result.
 154  */
 155 static void
 156 gen6_queryobj_get_results(struct gl_context *ctx,
 157                           struct brw_query_object *query)
 158 {
 159    struct brw_context *brw = brw_context(ctx);
 160    struct intel_context *intel = intel_context(ctx);
 161
 162    if (query->bo == NULL)
 163       return;
 164
 165    /* If the application has requested the query result, but this batch is
 166     * still contributing to it, flush it now so the results will be present
 167     * when mapped.
 168     */
 169    if (drm_intel_bo_references(intel->batch.bo, query->bo))
 170       intel_batchbuffer_flush(brw);
 171
 172    if (unlikely(intel->perf_debug)) {
 173       if (drm_intel_bo_busy(query->bo)) {
 174          perf_debug("Stalling on the GPU waiting for a query object.\n");
 175       }
 176    }
 177
 178    drm_intel_bo_map(query->bo, false);
 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       query->Base.Result = results[1] - results[0];
 227       break;
 228
 229    default:
 230       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 231       break;
 232    }
 233    drm_intel_bo_unmap(query->bo);
 234
 235    /* Now that we've processed the data stored in the query's buffer object,
 236     * we can release it.
 237     */
 238    drm_intel_bo_unreference(query->bo);
 239    query->bo = NULL;
 240 }
 241
 242 /**
 243  * Driver hook for glBeginQuery().
 244  *
 245  * Initializes driver structures and emits any GPU commands required to begin
 246  * recording data for the query.
 247  */
 248 static void
 249 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 250 {
 251    struct brw_context *brw = brw_context(ctx);
 252    struct brw_query_object *query = (struct brw_query_object *)q;
 253
 254    /* Since we're starting a new query, we need to throw away old results. */
 255    drm_intel_bo_unreference(query->bo);
 256    query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
 257
 258    switch (query->Base.Target) {
 259    case GL_TIME_ELAPSED:
 260       /* For timestamp queries, we record the starting time right away so that
 261        * we measure the full time between BeginQuery and EndQuery.  There's
 262        * some debate about whether this is the right thing to do.  Our decision
 263        * is based on the following text from the ARB_timer_query extension:
 264        *
 265        * "(5) Should the extension measure total time elapsed between the full
 266        *      completion of the BeginQuery and EndQuery commands, or just time
 267        *      spent in the graphics library?
 268        *
 269        *  RESOLVED:  This extension will measure the total time elapsed
 270        *  between the full completion of these commands.  Future extensions
 271        *  may implement a query to determine time elapsed at different stages
 272        *  of the graphics pipeline."
 273        *
 274        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 275        * we'll write a second timestamp (at index 1), and subtract the two to
 276        * obtain the time elapsed.  Notably, this includes time elapsed while
 277        * the system was doing other work, such as running other applications.
 278        */
 279       write_timestamp(brw, query->bo, 0);
 280       break;
 281
 282    case GL_ANY_SAMPLES_PASSED:
 283    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 284    case GL_SAMPLES_PASSED_ARB:
 285       write_depth_count(brw, query->bo, 0);
 286       break;
 287
 288    case GL_PRIMITIVES_GENERATED:
 289       write_primitives_generated(brw, query->bo, 0);
 290       break;
 291
 292    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 293       write_xfb_primitives_written(brw, query->bo, 0);
 294       break;
 295
 296    default:
 297       assert(!"Unrecognized query target in brw_begin_query()");
 298       break;
 299    }
 300 }
 301
 302 /**
 303  * Driver hook for glEndQuery().
 304  *
 305  * Emits GPU commands to record a final query value, ending any data capturing.
 306  * However, the final result isn't necessarily available until the GPU processes
 307  * those commands.  brw_queryobj_get_results() processes the captured data to
 308  * produce the final result.
 309  */
 310 static void
 311 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 312 {
 313    struct brw_context *brw = brw_context(ctx);
 314    struct brw_query_object *query = (struct brw_query_object *)q;
 315
 316    switch (query->Base.Target) {
 317    case GL_TIME_ELAPSED:
 318       write_timestamp(brw, query->bo, 1);
 319       break;
 320
 321    case GL_ANY_SAMPLES_PASSED:
 322    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 323    case GL_SAMPLES_PASSED_ARB:
 324       write_depth_count(brw, query->bo, 1);
 325       break;
 326
 327    case GL_PRIMITIVES_GENERATED:
 328       write_primitives_generated(brw, query->bo, 1);
 329       break;
 330
 331    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 332       write_xfb_primitives_written(brw, query->bo, 1);
 333       break;
 334
 335    default:
 336       assert(!"Unrecognized query target in brw_end_query()");
 337       break;
 338    }
 339 }
 340
 341 /**
 342  * The WaitQuery() driver hook.
 343  *
 344  * Wait for a query result to become available and return it.  This is the
 345  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 346  */
 347 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 348 {
 349    struct brw_query_object *query = (struct brw_query_object *)q;
 350
 351    gen6_queryobj_get_results(ctx, query);
 352    query->Base.Ready = true;
 353 }
 354
 355 /**
 356  * The CheckQuery() driver hook.
 357  *
 358  * Checks whether a query result is ready yet.  If not, flushes.
 359  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 360  */
 361 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
 362 {
 363    struct brw_context *brw = brw_context(ctx);
 364    struct intel_context *intel = intel_context(ctx);
 365    struct brw_query_object *query = (struct brw_query_object *)q;
 366
 367    /* From the GL_ARB_occlusion_query spec:
 368     *
 369     *     "Instead of allowing for an infinite loop, performing a
 370     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 371     *      not ready yet on the first time it is queried.  This ensures that
 372     *      the async query will return true in finite time.
 373     */
 374    if (query->bo && drm_intel_bo_references(intel->batch.bo, query->bo))
 375       intel_batchbuffer_flush(brw);
 376
 377    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 378       gen6_queryobj_get_results(ctx, query);
 379       query->Base.Ready = true;
 380    }
 381 }
 382
 383 /* Initialize Gen6+-specific query object functions. */
 384 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 385 {
 386    functions->BeginQuery = gen6_begin_query;
 387    functions->EndQuery = gen6_end_query;
 388    functions->CheckQuery = gen6_check_query;
 389    functions->WaitQuery = gen6_wait_query;
 390 }