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 /**
 113  * Wait on the query object's BO and calculate the final result.
 114  */
 115 static void
 116 gen6_queryobj_get_results(struct gl_context *ctx,
 117                           struct brw_query_object *query)
 118 {
 119    struct brw_context *brw = brw_context(ctx);
 120
 121    if (query->bo == NULL)
 122       return;
 123
 124    brw_bo_map(brw, query->bo, false, "query object");
 125    uint64_t *results = query->bo->virtual;
 126    switch (query->Base.Target) {
 127    case GL_TIME_ELAPSED:
 128       /* The query BO contains the starting and ending timestamps.
 129        * Subtract the two and convert to nanoseconds.
 130        */
 131       query->Base.Result += 80 * (results[1] - results[0]);
 132       break;
 133
 134    case GL_TIMESTAMP:
 135       /* Our timer is a clock that increments every 80ns (regardless of
 136        * other clock scaling in the system).  The timestamp register we can
 137        * read for glGetTimestamp() masks out the top 32 bits, so we do that
 138        * here too to let the two counters be compared against each other.
 139        *
 140        * If we just multiplied that 32 bits of data by 80, it would roll
 141        * over at a non-power-of-two, so an application couldn't use
 142        * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
 143        * report 36 bits and truncate at that (rolling over 5 times as often
 144        * as the HW counter), and when the 32-bit counter rolls over, it
 145        * happens to also be at a rollover in the reported value from near
 146        * (1<<36) to 0.
 147        *
 148        * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
 149        * rolls over every ~69 seconds.
 150        *
 151        * The query BO contains a single timestamp value in results[0].
 152        */
 153       query->Base.Result = 80 * (results[0] & 0xffffffff);
 154       query->Base.Result &= (1ull << 36) - 1;
 155       break;
 156
 157    case GL_SAMPLES_PASSED_ARB:
 158       /* We need to use += rather than = here since some BLT-based operations
 159        * may have added additional samples to our occlusion query value.
 160        */
 161       query->Base.Result += results[1] - results[0];
 162       break;
 163
 164    case GL_ANY_SAMPLES_PASSED:
 165    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 166       if (results[0] != results[1])
 167          query->Base.Result = true;
 168       break;
 169
 170    case GL_PRIMITIVES_GENERATED:
 171    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 172       query->Base.Result = results[1] - results[0];
 173       break;
 174
 175    default:
 176       unreachable("Unrecognized query target in brw_queryobj_get_results()");
 177    }
 178    drm_intel_bo_unmap(query->bo);
 179
 180    /* Now that we've processed the data stored in the query's buffer object,
 181     * we can release it.
 182     */
 183    drm_intel_bo_unreference(query->bo);
 184    query->bo = NULL;
 185
 186    query->Base.Ready = true;
 187 }
 188
 189 /**
 190  * Driver hook for glBeginQuery().
 191  *
 192  * Initializes driver structures and emits any GPU commands required to begin
 193  * recording data for the query.
 194  */
 195 static void
 196 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 197 {
 198    struct brw_context *brw = brw_context(ctx);
 199    struct brw_query_object *query = (struct brw_query_object *)q;
 200
 201    /* Since we're starting a new query, we need to throw away old results. */
 202    drm_intel_bo_unreference(query->bo);
 203    query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
 204
 205    switch (query->Base.Target) {
 206    case GL_TIME_ELAPSED:
 207       /* For timestamp queries, we record the starting time right away so that
 208        * we measure the full time between BeginQuery and EndQuery.  There's
 209        * some debate about whether this is the right thing to do.  Our decision
 210        * is based on the following text from the ARB_timer_query extension:
 211        *
 212        * "(5) Should the extension measure total time elapsed between the full
 213        *      completion of the BeginQuery and EndQuery commands, or just time
 214        *      spent in the graphics library?
 215        *
 216        *  RESOLVED:  This extension will measure the total time elapsed
 217        *  between the full completion of these commands.  Future extensions
 218        *  may implement a query to determine time elapsed at different stages
 219        *  of the graphics pipeline."
 220        *
 221        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 222        * we'll write a second timestamp (at index 1), and subtract the two to
 223        * obtain the time elapsed.  Notably, this includes time elapsed while
 224        * the system was doing other work, such as running other applications.
 225        */
 226       brw_write_timestamp(brw, query->bo, 0);
 227       break;
 228
 229    case GL_ANY_SAMPLES_PASSED:
 230    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 231    case GL_SAMPLES_PASSED_ARB:
 232       brw_write_depth_count(brw, query->bo, 0);
 233       break;
 234
 235    case GL_PRIMITIVES_GENERATED:
 236       write_primitives_generated(brw, query->bo, query->Base.Stream, 0);
 237       break;
 238
 239    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 240       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0);
 241       break;
 242
 243    default:
 244       unreachable("Unrecognized query target in brw_begin_query()");
 245    }
 246 }
 247
 248 /**
 249  * Driver hook for glEndQuery().
 250  *
 251  * Emits GPU commands to record a final query value, ending any data capturing.
 252  * However, the final result isn't necessarily available until the GPU processes
 253  * those commands.  brw_queryobj_get_results() processes the captured data to
 254  * produce the final result.
 255  */
 256 static void
 257 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 258 {
 259    struct brw_context *brw = brw_context(ctx);
 260    struct brw_query_object *query = (struct brw_query_object *)q;
 261
 262    switch (query->Base.Target) {
 263    case GL_TIME_ELAPSED:
 264       brw_write_timestamp(brw, query->bo, 1);
 265       break;
 266
 267    case GL_ANY_SAMPLES_PASSED:
 268    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 269    case GL_SAMPLES_PASSED_ARB:
 270       brw_write_depth_count(brw, query->bo, 1);
 271       break;
 272
 273    case GL_PRIMITIVES_GENERATED:
 274       write_primitives_generated(brw, query->bo, query->Base.Stream, 1);
 275       break;
 276
 277    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 278       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1);
 279       break;
 280
 281    default:
 282       unreachable("Unrecognized query target in brw_end_query()");
 283    }
 284
 285    /* The current batch contains the commands to handle EndQuery(),
 286     * but they won't actually execute until it is flushed.
 287     */
 288    query->flushed = false;
 289 }
 290
 291 /**
 292  * Flush the batch if it still references the query object BO.
 293  */
 294 static void
 295 flush_batch_if_needed(struct brw_context *brw, struct brw_query_object *query)
 296 {
 297    /* If the batch doesn't reference the BO, it must have been flushed
 298     * (for example, due to being full).  Record that it's been flushed.
 299     */
 300    query->flushed = query->flushed ||
 301       !drm_intel_bo_references(brw->batch.bo, query->bo);
 302
 303    if (!query->flushed)
 304       intel_batchbuffer_flush(brw);
 305 }
 306
 307 /**
 308  * The WaitQuery() driver hook.
 309  *
 310  * Wait for a query result to become available and return it.  This is the
 311  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 312  */
 313 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 314 {
 315    struct brw_context *brw = brw_context(ctx);
 316    struct brw_query_object *query = (struct brw_query_object *)q;
 317
 318    /* If the application has requested the query result, but this batch is
 319     * still contributing to it, flush it now to finish that work so the
 320     * result will become available (eventually).
 321     */
 322    flush_batch_if_needed(brw, query);
 323
 324    gen6_queryobj_get_results(ctx, query);
 325 }
 326
 327 /**
 328  * The CheckQuery() driver hook.
 329  *
 330  * Checks whether a query result is ready yet.  If not, flushes.
 331  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 332  */
 333 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
 334 {
 335    struct brw_context *brw = brw_context(ctx);
 336    struct brw_query_object *query = (struct brw_query_object *)q;
 337
 338    /* If query->bo is NULL, we've already gathered the results - this is a
 339     * redundant CheckQuery call.  Ignore it.
 340     */
 341    if (query->bo == NULL)
 342       return;
 343
 344    /* From the GL_ARB_occlusion_query spec:
 345     *
 346     *     "Instead of allowing for an infinite loop, performing a
 347     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 348     *      not ready yet on the first time it is queried.  This ensures that
 349     *      the async query will return true in finite time.
 350     */
 351    flush_batch_if_needed(brw, query);
 352
 353    if (!drm_intel_bo_busy(query->bo)) {
 354       gen6_queryobj_get_results(ctx, query);
 355    }
 356 }
 357
 358 /* Initialize Gen6+-specific query object functions. */
 359 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 360 {
 361    functions->BeginQuery = gen6_begin_query;
 362    functions->EndQuery = gen6_end_query;
 363    functions->CheckQuery = gen6_check_query;
 364    functions->WaitQuery = gen6_wait_query;
 365 }