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 idx)
  88 {
  89    intel_batchbuffer_emit_mi_flush(brw);
  90
  91    brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT, idx);
  92 }
  93
  94 static void
  95 write_xfb_primitives_written(struct brw_context *brw,
  96                              drm_intel_bo *bo, int stream, int idx)
  97 {
  98    intel_batchbuffer_emit_mi_flush(brw);
  99
 100    if (brw->gen >= 7) {
 101       brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(stream), idx);
 102    } else {
 103       brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN, idx);
 104    }
 105 }
 106
 107 /**
 108  * Wait on the query object's BO and calculate the final result.
 109  */
 110 static void
 111 gen6_queryobj_get_results(struct gl_context *ctx,
 112                           struct brw_query_object *query)
 113 {
 114    struct brw_context *brw = brw_context(ctx);
 115
 116    if (query->bo == NULL)
 117       return;
 118
 119    /* If the application has requested the query result, but this batch is
 120     * still contributing to it, flush it now so the results will be present
 121     * when mapped.
 122     */
 123    if (drm_intel_bo_references(brw->batch.bo, query->bo))
 124       intel_batchbuffer_flush(brw);
 125
 126    if (unlikely(brw->perf_debug)) {
 127       if (drm_intel_bo_busy(query->bo)) {
 128          perf_debug("Stalling on the GPU waiting for a query object.\n");
 129       }
 130    }
 131
 132    drm_intel_bo_map(query->bo, false);
 133    uint64_t *results = query->bo->virtual;
 134    switch (query->Base.Target) {
 135    case GL_TIME_ELAPSED:
 136       /* The query BO contains the starting and ending timestamps.
 137        * Subtract the two and convert to nanoseconds.
 138        */
 139       query->Base.Result += 80 * (results[1] - results[0]);
 140       break;
 141
 142    case GL_TIMESTAMP:
 143       /* Our timer is a clock that increments every 80ns (regardless of
 144        * other clock scaling in the system).  The timestamp register we can
 145        * read for glGetTimestamp() masks out the top 32 bits, so we do that
 146        * here too to let the two counters be compared against each other.
 147        *
 148        * If we just multiplied that 32 bits of data by 80, it would roll
 149        * over at a non-power-of-two, so an application couldn't use
 150        * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
 151        * report 36 bits and truncate at that (rolling over 5 times as often
 152        * as the HW counter), and when the 32-bit counter rolls over, it
 153        * happens to also be at a rollover in the reported value from near
 154        * (1<<36) to 0.
 155        *
 156        * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
 157        * rolls over every ~69 seconds.
 158        *
 159        * The query BO contains a single timestamp value in results[0].
 160        */
 161       query->Base.Result = 80 * (results[0] & 0xffffffff);
 162       query->Base.Result &= (1ull << 36) - 1;
 163       break;
 164
 165    case GL_SAMPLES_PASSED_ARB:
 166       /* We need to use += rather than = here since some BLT-based operations
 167        * may have added additional samples to our occlusion query value.
 168        */
 169       query->Base.Result += results[1] - results[0];
 170       break;
 171
 172    case GL_ANY_SAMPLES_PASSED:
 173    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 174       if (results[0] != results[1])
 175          query->Base.Result = true;
 176       break;
 177
 178    case GL_PRIMITIVES_GENERATED:
 179    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 180       query->Base.Result = results[1] - results[0];
 181       break;
 182
 183    default:
 184       unreachable("Unrecognized query target in brw_queryobj_get_results()");
 185    }
 186    drm_intel_bo_unmap(query->bo);
 187
 188    /* Now that we've processed the data stored in the query's buffer object,
 189     * we can release it.
 190     */
 191    drm_intel_bo_unreference(query->bo);
 192    query->bo = NULL;
 193 }
 194
 195 /**
 196  * Driver hook for glBeginQuery().
 197  *
 198  * Initializes driver structures and emits any GPU commands required to begin
 199  * recording data for the query.
 200  */
 201 static void
 202 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 203 {
 204    struct brw_context *brw = brw_context(ctx);
 205    struct brw_query_object *query = (struct brw_query_object *)q;
 206
 207    /* Since we're starting a new query, we need to throw away old results. */
 208    drm_intel_bo_unreference(query->bo);
 209    query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
 210
 211    switch (query->Base.Target) {
 212    case GL_TIME_ELAPSED:
 213       /* For timestamp queries, we record the starting time right away so that
 214        * we measure the full time between BeginQuery and EndQuery.  There's
 215        * some debate about whether this is the right thing to do.  Our decision
 216        * is based on the following text from the ARB_timer_query extension:
 217        *
 218        * "(5) Should the extension measure total time elapsed between the full
 219        *      completion of the BeginQuery and EndQuery commands, or just time
 220        *      spent in the graphics library?
 221        *
 222        *  RESOLVED:  This extension will measure the total time elapsed
 223        *  between the full completion of these commands.  Future extensions
 224        *  may implement a query to determine time elapsed at different stages
 225        *  of the graphics pipeline."
 226        *
 227        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 228        * we'll write a second timestamp (at index 1), and subtract the two to
 229        * obtain the time elapsed.  Notably, this includes time elapsed while
 230        * the system was doing other work, such as running other applications.
 231        */
 232       brw_write_timestamp(brw, query->bo, 0);
 233       break;
 234
 235    case GL_ANY_SAMPLES_PASSED:
 236    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 237    case GL_SAMPLES_PASSED_ARB:
 238       brw_write_depth_count(brw, query->bo, 0);
 239       break;
 240
 241    case GL_PRIMITIVES_GENERATED:
 242       write_primitives_generated(brw, query->bo, 0);
 243       break;
 244
 245    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 246       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0);
 247       break;
 248
 249    default:
 250       unreachable("Unrecognized query target in brw_begin_query()");
 251    }
 252 }
 253
 254 /**
 255  * Driver hook for glEndQuery().
 256  *
 257  * Emits GPU commands to record a final query value, ending any data capturing.
 258  * However, the final result isn't necessarily available until the GPU processes
 259  * those commands.  brw_queryobj_get_results() processes the captured data to
 260  * produce the final result.
 261  */
 262 static void
 263 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 264 {
 265    struct brw_context *brw = brw_context(ctx);
 266    struct brw_query_object *query = (struct brw_query_object *)q;
 267
 268    switch (query->Base.Target) {
 269    case GL_TIME_ELAPSED:
 270       brw_write_timestamp(brw, query->bo, 1);
 271       break;
 272
 273    case GL_ANY_SAMPLES_PASSED:
 274    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 275    case GL_SAMPLES_PASSED_ARB:
 276       brw_write_depth_count(brw, query->bo, 1);
 277       break;
 278
 279    case GL_PRIMITIVES_GENERATED:
 280       write_primitives_generated(brw, query->bo, 1);
 281       break;
 282
 283    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 284       write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1);
 285       break;
 286
 287    default:
 288       unreachable("Unrecognized query target in brw_end_query()");
 289    }
 290 }
 291
 292 /**
 293  * The WaitQuery() driver hook.
 294  *
 295  * Wait for a query result to become available and return it.  This is the
 296  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 297  */
 298 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 299 {
 300    struct brw_query_object *query = (struct brw_query_object *)q;
 301
 302    gen6_queryobj_get_results(ctx, query);
 303    query->Base.Ready = true;
 304 }
 305
 306 /**
 307  * The CheckQuery() driver hook.
 308  *
 309  * Checks whether a query result is ready yet.  If not, flushes.
 310  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 311  */
 312 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
 313 {
 314    struct brw_context *brw = brw_context(ctx);
 315    struct brw_query_object *query = (struct brw_query_object *)q;
 316
 317    /* From the GL_ARB_occlusion_query spec:
 318     *
 319     *     "Instead of allowing for an infinite loop, performing a
 320     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 321     *      not ready yet on the first time it is queried.  This ensures that
 322     *      the async query will return true in finite time.
 323     */
 324    if (query->bo && drm_intel_bo_references(brw->batch.bo, query->bo))
 325       intel_batchbuffer_flush(brw);
 326
 327    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 328       gen6_queryobj_get_results(ctx, query);
 329       query->Base.Ready = true;
 330    }
 331 }
 332
 333 /* Initialize Gen6+-specific query object functions. */
 334 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 335 {
 336    functions->BeginQuery = gen6_begin_query;
 337    functions->EndQuery = gen6_end_query;
 338    functions->CheckQuery = gen6_check_query;
 339    functions->WaitQuery = gen6_wait_query;
 340 }