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 PS_DEPTH_COUNT register into a buffer.
  44  */
  45 static void
  46 write_depth_count(struct brw_context *brw, drm_intel_bo *query_bo, int idx)
  47 {
  48    /* Emit Sandybridge workaround flush: */
  49    if (brw->gen == 6)
  50       intel_emit_post_sync_nonzero_flush(brw);
  51
  52    brw_emit_pipe_control_write(brw,
  53                                PIPE_CONTROL_WRITE_DEPTH_COUNT
  54                                | PIPE_CONTROL_DEPTH_STALL,
  55                                query_bo, idx * sizeof(uint64_t), 0, 0);
  56 }
  57
  58 /*
  59  * Write an arbitrary 64-bit register to a buffer via MI_STORE_REGISTER_MEM.
  60  *
  61  * Only TIMESTAMP and PS_DEPTH_COUNT have special PIPE_CONTROL support; other
  62  * counters have to be read via the generic MI_STORE_REGISTER_MEM.
  63  *
  64  * Callers must explicitly flush the pipeline to ensure the desired value is
  65  * available.
  66  */
  67 void
  68 brw_store_register_mem64(struct brw_context *brw,
  69                          drm_intel_bo *bo, uint32_t reg, int idx)
  70 {
  71    assert(brw->gen >= 6);
  72
  73    /* MI_STORE_REGISTER_MEM only stores a single 32-bit value, so to
  74     * read a full 64-bit register, we need to do two of them.
  75     */
  76    if (brw->gen >= 8) {
  77       BEGIN_BATCH(8);
  78       OUT_BATCH(MI_STORE_REGISTER_MEM | (4 - 2));
  79       OUT_BATCH(reg);
  80       OUT_RELOC64(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  81                   idx * sizeof(uint64_t));
  82       OUT_BATCH(MI_STORE_REGISTER_MEM | (4 - 2));
  83       OUT_BATCH(reg + sizeof(uint32_t));
  84       OUT_RELOC64(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  85                   sizeof(uint32_t) + idx * sizeof(uint64_t));
  86       ADVANCE_BATCH();
  87    } else {
  88       BEGIN_BATCH(6);
  89       OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
  90       OUT_BATCH(reg);
  91       OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  92                 idx * sizeof(uint64_t));
  93       OUT_BATCH(MI_STORE_REGISTER_MEM | (3 - 2));
  94       OUT_BATCH(reg + sizeof(uint32_t));
  95       OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
  96                 sizeof(uint32_t) + idx * sizeof(uint64_t));
  97       ADVANCE_BATCH();
  98    }
  99 }
 100
 101 static void
 102 write_primitives_generated(struct brw_context *brw,
 103                            drm_intel_bo *query_bo, int idx)
 104 {
 105    intel_batchbuffer_emit_mi_flush(brw);
 106
 107    brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT, idx);
 108 }
 109
 110 static void
 111 write_xfb_primitives_written(struct brw_context *brw,
 112                              drm_intel_bo *bo, int idx)
 113 {
 114    intel_batchbuffer_emit_mi_flush(brw);
 115
 116    if (brw->gen >= 7) {
 117       brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(0), idx);
 118    } else {
 119       brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN, idx);
 120    }
 121 }
 122
 123 /**
 124  * Wait on the query object's BO and calculate the final result.
 125  */
 126 static void
 127 gen6_queryobj_get_results(struct gl_context *ctx,
 128                           struct brw_query_object *query)
 129 {
 130    struct brw_context *brw = brw_context(ctx);
 131
 132    if (query->bo == NULL)
 133       return;
 134
 135    /* If the application has requested the query result, but this batch is
 136     * still contributing to it, flush it now so the results will be present
 137     * when mapped.
 138     */
 139    if (drm_intel_bo_references(brw->batch.bo, query->bo))
 140       intel_batchbuffer_flush(brw);
 141
 142    if (unlikely(brw->perf_debug)) {
 143       if (drm_intel_bo_busy(query->bo)) {
 144          perf_debug("Stalling on the GPU waiting for a query object.\n");
 145       }
 146    }
 147
 148    drm_intel_bo_map(query->bo, false);
 149    uint64_t *results = query->bo->virtual;
 150    switch (query->Base.Target) {
 151    case GL_TIME_ELAPSED:
 152       /* The query BO contains the starting and ending timestamps.
 153        * Subtract the two and convert to nanoseconds.
 154        */
 155       query->Base.Result += 80 * (results[1] - results[0]);
 156       break;
 157
 158    case GL_TIMESTAMP:
 159       /* Our timer is a clock that increments every 80ns (regardless of
 160        * other clock scaling in the system).  The timestamp register we can
 161        * read for glGetTimestamp() masks out the top 32 bits, so we do that
 162        * here too to let the two counters be compared against each other.
 163        *
 164        * If we just multiplied that 32 bits of data by 80, it would roll
 165        * over at a non-power-of-two, so an application couldn't use
 166        * GL_QUERY_COUNTER_BITS to handle rollover correctly.  Instead, we
 167        * report 36 bits and truncate at that (rolling over 5 times as often
 168        * as the HW counter), and when the 32-bit counter rolls over, it
 169        * happens to also be at a rollover in the reported value from near
 170        * (1<<36) to 0.
 171        *
 172        * The low 32 bits rolls over in ~343 seconds.  Our 36-bit result
 173        * rolls over every ~69 seconds.
 174        *
 175        * The query BO contains a single timestamp value in results[0].
 176        */
 177       query->Base.Result = 80 * (results[0] & 0xffffffff);
 178       query->Base.Result &= (1ull << 36) - 1;
 179       break;
 180
 181    case GL_SAMPLES_PASSED_ARB:
 182       /* We need to use += rather than = here since some BLT-based operations
 183        * may have added additional samples to our occlusion query value.
 184        */
 185       query->Base.Result += results[1] - results[0];
 186       break;
 187
 188    case GL_ANY_SAMPLES_PASSED:
 189    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 190       if (results[0] != results[1])
 191          query->Base.Result = true;
 192       break;
 193
 194    case GL_PRIMITIVES_GENERATED:
 195    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 196       query->Base.Result = results[1] - results[0];
 197       break;
 198
 199    default:
 200       assert(!"Unrecognized query target in brw_queryobj_get_results()");
 201       break;
 202    }
 203    drm_intel_bo_unmap(query->bo);
 204
 205    /* Now that we've processed the data stored in the query's buffer object,
 206     * we can release it.
 207     */
 208    drm_intel_bo_unreference(query->bo);
 209    query->bo = NULL;
 210 }
 211
 212 /**
 213  * Driver hook for glBeginQuery().
 214  *
 215  * Initializes driver structures and emits any GPU commands required to begin
 216  * recording data for the query.
 217  */
 218 static void
 219 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
 220 {
 221    struct brw_context *brw = brw_context(ctx);
 222    struct brw_query_object *query = (struct brw_query_object *)q;
 223
 224    /* Since we're starting a new query, we need to throw away old results. */
 225    drm_intel_bo_unreference(query->bo);
 226    query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
 227
 228    switch (query->Base.Target) {
 229    case GL_TIME_ELAPSED:
 230       /* For timestamp queries, we record the starting time right away so that
 231        * we measure the full time between BeginQuery and EndQuery.  There's
 232        * some debate about whether this is the right thing to do.  Our decision
 233        * is based on the following text from the ARB_timer_query extension:
 234        *
 235        * "(5) Should the extension measure total time elapsed between the full
 236        *      completion of the BeginQuery and EndQuery commands, or just time
 237        *      spent in the graphics library?
 238        *
 239        *  RESOLVED:  This extension will measure the total time elapsed
 240        *  between the full completion of these commands.  Future extensions
 241        *  may implement a query to determine time elapsed at different stages
 242        *  of the graphics pipeline."
 243        *
 244        * We write a starting timestamp now (at index 0).  At EndQuery() time,
 245        * we'll write a second timestamp (at index 1), and subtract the two to
 246        * obtain the time elapsed.  Notably, this includes time elapsed while
 247        * the system was doing other work, such as running other applications.
 248        */
 249       brw_write_timestamp(brw, query->bo, 0);
 250       break;
 251
 252    case GL_ANY_SAMPLES_PASSED:
 253    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 254    case GL_SAMPLES_PASSED_ARB:
 255       write_depth_count(brw, query->bo, 0);
 256       break;
 257
 258    case GL_PRIMITIVES_GENERATED:
 259       write_primitives_generated(brw, query->bo, 0);
 260       break;
 261
 262    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 263       write_xfb_primitives_written(brw, query->bo, 0);
 264       break;
 265
 266    default:
 267       assert(!"Unrecognized query target in brw_begin_query()");
 268       break;
 269    }
 270 }
 271
 272 /**
 273  * Driver hook for glEndQuery().
 274  *
 275  * Emits GPU commands to record a final query value, ending any data capturing.
 276  * However, the final result isn't necessarily available until the GPU processes
 277  * those commands.  brw_queryobj_get_results() processes the captured data to
 278  * produce the final result.
 279  */
 280 static void
 281 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
 282 {
 283    struct brw_context *brw = brw_context(ctx);
 284    struct brw_query_object *query = (struct brw_query_object *)q;
 285
 286    switch (query->Base.Target) {
 287    case GL_TIME_ELAPSED:
 288       brw_write_timestamp(brw, query->bo, 1);
 289       break;
 290
 291    case GL_ANY_SAMPLES_PASSED:
 292    case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
 293    case GL_SAMPLES_PASSED_ARB:
 294       write_depth_count(brw, query->bo, 1);
 295       break;
 296
 297    case GL_PRIMITIVES_GENERATED:
 298       write_primitives_generated(brw, query->bo, 1);
 299       break;
 300
 301    case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 302       write_xfb_primitives_written(brw, query->bo, 1);
 303       break;
 304
 305    default:
 306       assert(!"Unrecognized query target in brw_end_query()");
 307       break;
 308    }
 309 }
 310
 311 /**
 312  * The WaitQuery() driver hook.
 313  *
 314  * Wait for a query result to become available and return it.  This is the
 315  * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
 316  */
 317 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
 318 {
 319    struct brw_query_object *query = (struct brw_query_object *)q;
 320
 321    gen6_queryobj_get_results(ctx, query);
 322    query->Base.Ready = true;
 323 }
 324
 325 /**
 326  * The CheckQuery() driver hook.
 327  *
 328  * Checks whether a query result is ready yet.  If not, flushes.
 329  * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
 330  */
 331 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
 332 {
 333    struct brw_context *brw = brw_context(ctx);
 334    struct brw_query_object *query = (struct brw_query_object *)q;
 335
 336    /* From the GL_ARB_occlusion_query spec:
 337     *
 338     *     "Instead of allowing for an infinite loop, performing a
 339     *      QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
 340     *      not ready yet on the first time it is queried.  This ensures that
 341     *      the async query will return true in finite time.
 342     */
 343    if (query->bo && drm_intel_bo_references(brw->batch.bo, query->bo))
 344       intel_batchbuffer_flush(brw);
 345
 346    if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
 347       gen6_queryobj_get_results(ctx, query);
 348       query->Base.Ready = true;
 349    }
 350 }
 351
 352 /* Initialize Gen6+-specific query object functions. */
 353 void gen6_init_queryobj_functions(struct dd_function_table *functions)
 354 {
 355    functions->BeginQuery = gen6_begin_query;
 356    functions->EndQuery = gen6_end_query;
 357    functions->CheckQuery = gen6_check_query;
 358    functions->WaitQuery = gen6_wait_query;
 359 }