src/mesa/drivers/dri/i965/brw_performance_query.c

   1 /*
   2  * Copyright © 2013 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
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 /**
  25  * \file brw_performance_query.c
  26  *
  27  * Implementation of the GL_INTEL_performance_query extension.
  28  *
  29  * Currently there are two possible counter sources exposed here:
  30  *
  31  * On Gen6+ hardware we have numerous 64bit Pipeline Statistics Registers
  32  * that we can snapshot at the beginning and end of a query.
  33  *
  34  * On Gen7.5+ we have Observability Architecture counters which are
  35  * covered in separate document from the rest of the PRMs.  It is available at:
  36  * https://01.org/linuxgraphics/documentation/driver-documentation-prms
  37  * => 2013 Intel Core Processor Family => Observability Performance Counters
  38  * (This one volume covers Sandybridge, Ivybridge, Baytrail, and Haswell,
  39  * though notably we currently only support OA counters for Haswell+)
  40  */
  41
  42 #include <limits.h>
  43
  44 /* put before sys/types.h to silence glibc warnings */
  45 #ifdef MAJOR_IN_MKDEV
  46 #include <sys/mkdev.h>
  47 #endif
  48 #ifdef MAJOR_IN_SYSMACROS
  49 #include <sys/sysmacros.h>
  50 #endif
  51 #include <sys/types.h>
  52 #include <sys/stat.h>
  53 #include <fcntl.h>
  54 #include <sys/mman.h>
  55 #include <sys/ioctl.h>
  56
  57 #include <xf86drm.h>
  58 #include "drm-uapi/i915_drm.h"
  59
  60 #include "main/hash.h"
  61 #include "main/macros.h"
  62 #include "main/mtypes.h"
  63 #include "main/performance_query.h"
  64
  65 #include "util/bitset.h"
  66 #include "util/ralloc.h"
  67 #include "util/hash_table.h"
  68 #include "util/list.h"
  69 #include "util/u_math.h"
  70
  71 #include "brw_context.h"
  72 #include "brw_defines.h"
  73 #include "intel_batchbuffer.h"
  74
  75 #include "perf/gen_perf.h"
  76 #include "perf/gen_perf_mdapi.h"
  77
  78 #define FILE_DEBUG_FLAG DEBUG_PERFMON
  79
  80 #define OAREPORT_REASON_MASK           0x3f
  81 #define OAREPORT_REASON_SHIFT          19
  82 #define OAREPORT_REASON_TIMER          (1<<0)
  83 #define OAREPORT_REASON_TRIGGER1       (1<<1)
  84 #define OAREPORT_REASON_TRIGGER2       (1<<2)
  85 #define OAREPORT_REASON_CTX_SWITCH     (1<<3)
  86 #define OAREPORT_REASON_GO_TRANSITION  (1<<4)
  87
  88 struct brw_perf_query_object {
  89    struct gl_perf_query_object base;
  90    struct gen_perf_query_object *query;
  91 };
  92
  93 /** Downcasting convenience macro. */
  94 static inline struct brw_perf_query_object *
  95 brw_perf_query(struct gl_perf_query_object *o)
  96 {
  97    return (struct brw_perf_query_object *) o;
  98 }
  99
 100 #define MI_RPC_BO_SIZE              4096
 101 #define MI_RPC_BO_END_OFFSET_BYTES  (MI_RPC_BO_SIZE / 2)
 102 #define MI_FREQ_START_OFFSET_BYTES  (3072)
 103 #define MI_FREQ_END_OFFSET_BYTES    (3076)
 104
 105 /******************************************************************************/
 106
 107 static bool
 108 brw_is_perf_query_ready(struct gl_context *ctx,
 109                         struct gl_perf_query_object *o);
 110
 111 static void
 112 dump_perf_query_callback(GLuint id, void *query_void, void *brw_void)
 113 {
 114    struct brw_context *ctx = brw_void;
 115    struct gen_perf_context *perf_ctx = ctx->perf_ctx;
 116    struct gl_perf_query_object *o = query_void;
 117    struct brw_perf_query_object * brw_query = brw_perf_query(o);
 118    struct gen_perf_query_object *obj = brw_query->query;
 119
 120    DBG("%4d: %-6s %-8s ",
 121        id,
 122        o->Used ? "Dirty," : "New,",
 123        o->Active ? "Active," : (o->Ready ? "Ready," : "Pending,"));
 124    gen_perf_dump_query(perf_ctx, obj, &ctx->batch);
 125 }
 126
 127 static void
 128 dump_perf_queries(struct brw_context *brw)
 129 {
 130    struct gl_context *ctx = &brw->ctx;
 131    gen_perf_dump_query_count(brw->perf_ctx);
 132    _mesa_HashWalk(ctx->PerfQuery.Objects, dump_perf_query_callback, brw);
 133 }
 134
 135 /**
 136  * Driver hook for glGetPerfQueryInfoINTEL().
 137  */
 138 static void
 139 brw_get_perf_query_info(struct gl_context *ctx,
 140                         unsigned query_index,
 141                         const char **name,
 142                         GLuint *data_size,
 143                         GLuint *n_counters,
 144                         GLuint *n_active)
 145 {
 146    struct brw_context *brw = brw_context(ctx);
 147    struct gen_perf_context *perf_ctx = brw->perf_ctx;
 148    struct gen_perf_config *perf_cfg = gen_perf_config(perf_ctx);
 149    const struct gen_perf_query_info *query = &perf_cfg->queries[query_index];
 150
 151    *name = query->name;
 152    *data_size = query->data_size;
 153    *n_counters = query->n_counters;
 154    *n_active = gen_perf_active_queries(perf_ctx, query);
 155 }
 156
 157 static GLuint
 158 gen_counter_type_enum_to_gl_type(enum gen_perf_counter_type type)
 159 {
 160    switch (type) {
 161    case GEN_PERF_COUNTER_TYPE_EVENT: return GL_PERFQUERY_COUNTER_EVENT_INTEL;
 162    case GEN_PERF_COUNTER_TYPE_DURATION_NORM: return GL_PERFQUERY_COUNTER_DURATION_NORM_INTEL;
 163    case GEN_PERF_COUNTER_TYPE_DURATION_RAW: return GL_PERFQUERY_COUNTER_DURATION_RAW_INTEL;
 164    case GEN_PERF_COUNTER_TYPE_THROUGHPUT: return GL_PERFQUERY_COUNTER_THROUGHPUT_INTEL;
 165    case GEN_PERF_COUNTER_TYPE_RAW: return GL_PERFQUERY_COUNTER_RAW_INTEL;
 166    case GEN_PERF_COUNTER_TYPE_TIMESTAMP: return GL_PERFQUERY_COUNTER_TIMESTAMP_INTEL;
 167    default:
 168       unreachable("Unknown counter type");
 169    }
 170 }
 171
 172 static GLuint
 173 gen_counter_data_type_to_gl_type(enum gen_perf_counter_data_type type)
 174 {
 175    switch (type) {
 176    case GEN_PERF_COUNTER_DATA_TYPE_BOOL32: return GL_PERFQUERY_COUNTER_DATA_BOOL32_INTEL;
 177    case GEN_PERF_COUNTER_DATA_TYPE_UINT32: return GL_PERFQUERY_COUNTER_DATA_UINT32_INTEL;
 178    case GEN_PERF_COUNTER_DATA_TYPE_UINT64: return GL_PERFQUERY_COUNTER_DATA_UINT64_INTEL;
 179    case GEN_PERF_COUNTER_DATA_TYPE_FLOAT: return GL_PERFQUERY_COUNTER_DATA_FLOAT_INTEL;
 180    case GEN_PERF_COUNTER_DATA_TYPE_DOUBLE: return GL_PERFQUERY_COUNTER_DATA_DOUBLE_INTEL;
 181    default:
 182       unreachable("Unknown counter data type");
 183    }
 184 }
 185
 186 /**
 187  * Driver hook for glGetPerfCounterInfoINTEL().
 188  */
 189 static void
 190 brw_get_perf_counter_info(struct gl_context *ctx,
 191                           unsigned query_index,
 192                           unsigned counter_index,
 193                           const char **name,
 194                           const char **desc,
 195                           GLuint *offset,
 196                           GLuint *data_size,
 197                           GLuint *type_enum,
 198                           GLuint *data_type_enum,
 199                           GLuint64 *raw_max)
 200 {
 201    struct brw_context *brw = brw_context(ctx);
 202    struct gen_perf_config *perf_cfg = gen_perf_config(brw->perf_ctx);
 203    const struct gen_perf_query_info *query =
 204       &perf_cfg->queries[query_index];
 205    const struct gen_perf_query_counter *counter =
 206       &query->counters[counter_index];
 207
 208    *name = counter->name;
 209    *desc = counter->desc;
 210    *offset = counter->offset;
 211    *data_size = gen_perf_query_counter_get_size(counter);
 212    *type_enum = gen_counter_type_enum_to_gl_type(counter->type);
 213    *data_type_enum = gen_counter_data_type_to_gl_type(counter->data_type);
 214    *raw_max = counter->raw_max;
 215 }
 216
 217 enum OaReadStatus {
 218    OA_READ_STATUS_ERROR,
 219    OA_READ_STATUS_UNFINISHED,
 220    OA_READ_STATUS_FINISHED,
 221 };
 222
 223 /******************************************************************************/
 224
 225 static void
 226 capture_frequency_stat_register(struct brw_context *brw,
 227                                 struct brw_bo *bo,
 228                                 uint32_t bo_offset)
 229 {
 230    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 231
 232    if (devinfo->gen >= 7 && devinfo->gen <= 8 &&
 233        !devinfo->is_baytrail && !devinfo->is_cherryview) {
 234       brw_store_register_mem32(brw, bo, GEN7_RPSTAT1, bo_offset);
 235    } else if (devinfo->gen >= 9) {
 236       brw_store_register_mem32(brw, bo, GEN9_RPSTAT0, bo_offset);
 237    }
 238 }
 239
 240 /**
 241  * Driver hook for glBeginPerfQueryINTEL().
 242  */
 243 static bool
 244 brw_begin_perf_query(struct gl_context *ctx,
 245                      struct gl_perf_query_object *o)
 246 {
 247    struct brw_context *brw = brw_context(ctx);
 248    struct brw_perf_query_object *brw_query = brw_perf_query(o);
 249    struct gen_perf_query_object *obj = brw_query->query;
 250    struct gen_perf_context *perf_ctx = brw->perf_ctx;
 251
 252    /* We can assume the frontend hides mistaken attempts to Begin a
 253     * query object multiple times before its End. Similarly if an
 254     * application reuses a query object before results have arrived
 255     * the frontend will wait for prior results so we don't need
 256     * to support abandoning in-flight results.
 257     */
 258    assert(!o->Active);
 259    assert(!o->Used || o->Ready); /* no in-flight query to worry about */
 260
 261    DBG("Begin(%d)\n", o->Id);
 262
 263    gen_perf_begin_query(perf_ctx, obj);
 264
 265    if (INTEL_DEBUG & DEBUG_PERFMON)
 266       dump_perf_queries(brw);
 267
 268    return true;
 269 }
 270
 271 /**
 272  * Driver hook for glEndPerfQueryINTEL().
 273  */
 274 static void
 275 brw_end_perf_query(struct gl_context *ctx,
 276                      struct gl_perf_query_object *o)
 277 {
 278    struct brw_context *brw = brw_context(ctx);
 279    struct brw_perf_query_object *brw_query = brw_perf_query(o);
 280    struct gen_perf_query_object *obj = brw_query->query;
 281    struct gen_perf_context *perf_ctx = brw->perf_ctx;
 282
 283    DBG("End(%d)\n", o->Id);
 284    gen_perf_end_query(perf_ctx, obj);
 285 }
 286
 287 static void
 288 brw_wait_perf_query(struct gl_context *ctx, struct gl_perf_query_object *o)
 289 {
 290    struct brw_context *brw = brw_context(ctx);
 291    struct brw_perf_query_object *brw_query = brw_perf_query(o);
 292    struct gen_perf_query_object *obj = brw_query->query;
 293
 294    assert(!o->Ready);
 295
 296    gen_perf_wait_query(brw->perf_ctx, obj, &brw->batch);
 297 }
 298
 299 static bool
 300 brw_is_perf_query_ready(struct gl_context *ctx,
 301                         struct gl_perf_query_object *o)
 302 {
 303    struct brw_context *brw = brw_context(ctx);
 304    struct brw_perf_query_object *brw_query = brw_perf_query(o);
 305    struct gen_perf_query_object *obj = brw_query->query;
 306
 307    if (o->Ready)
 308       return true;
 309
 310    return gen_perf_is_query_ready(brw->perf_ctx, obj, &brw->batch);
 311 }
 312
 313 /**
 314  * Driver hook for glGetPerfQueryDataINTEL().
 315  */
 316 static void
 317 brw_get_perf_query_data(struct gl_context *ctx,
 318                         struct gl_perf_query_object *o,
 319                         GLsizei data_size,
 320                         GLuint *data,
 321                         GLuint *bytes_written)
 322 {
 323    struct brw_context *brw = brw_context(ctx);
 324    struct brw_perf_query_object *brw_query = brw_perf_query(o);
 325    struct gen_perf_query_object *obj = brw_query->query;
 326
 327    assert(brw_is_perf_query_ready(ctx, o));
 328
 329    DBG("GetData(%d)\n", o->Id);
 330
 331    if (INTEL_DEBUG & DEBUG_PERFMON)
 332       dump_perf_queries(brw);
 333
 334    /* We expect that the frontend only calls this hook when it knows
 335     * that results are available.
 336     */
 337    assert(o->Ready);
 338
 339    gen_perf_get_query_data(brw->perf_ctx, obj,
 340                            data_size, data, bytes_written);
 341 }
 342
 343 static struct gl_perf_query_object *
 344 brw_new_perf_query_object(struct gl_context *ctx, unsigned query_index)
 345 {
 346    struct brw_context *brw = brw_context(ctx);
 347    struct gen_perf_context *perf_ctx = brw->perf_ctx;
 348    struct gen_perf_query_object * obj = gen_perf_new_query(perf_ctx, query_index);
 349    if (unlikely(!obj))
 350       return NULL;
 351
 352    struct brw_perf_query_object *brw_query = calloc(1, sizeof(struct brw_perf_query_object));
 353    if (unlikely(!brw_query)) {
 354       gen_perf_delete_query(perf_ctx, obj);
 355       return NULL;
 356    }
 357
 358    brw_query->query = obj;
 359    return &brw_query->base;
 360 }
 361
 362 /**
 363  * Driver hook for glDeletePerfQueryINTEL().
 364  */
 365 static void
 366 brw_delete_perf_query(struct gl_context *ctx,
 367                       struct gl_perf_query_object *o)
 368 {
 369    struct brw_context *brw = brw_context(ctx);
 370    struct brw_perf_query_object *brw_query = brw_perf_query(o);
 371    struct gen_perf_query_object *obj = brw_query->query;
 372    struct gen_perf_context *perf_ctx = brw->perf_ctx;
 373
 374    /* We can assume that the frontend waits for a query to complete
 375     * before ever calling into here, so we don't have to worry about
 376     * deleting an in-flight query object.
 377     */
 378    assert(!o->Active);
 379    assert(!o->Used || o->Ready);
 380
 381    DBG("Delete(%d)\n", o->Id);
 382
 383    gen_perf_delete_query(perf_ctx, obj);
 384    free(brw_query);
 385 }
 386
 387 /******************************************************************************/
 388 /* gen_device_info will have incorrect default topology values for unsupported kernels.
 389  * verify kernel support to ensure OA metrics are accurate.
 390  */
 391 static bool
 392 oa_metrics_kernel_support(int fd, const struct gen_device_info *devinfo)
 393 {
 394    if (devinfo->gen >= 10) {
 395       /* topology uAPI required for CNL+ (kernel 4.17+) make a call to the api
 396        * to verify support
 397        */
 398       struct drm_i915_query_item item = {
 399          .query_id = DRM_I915_QUERY_TOPOLOGY_INFO,
 400       };
 401       struct drm_i915_query query = {
 402          .num_items = 1,
 403          .items_ptr = (uintptr_t) &item,
 404       };
 405
 406       /* kernel 4.17+ supports the query */
 407       return drmIoctl(fd, DRM_IOCTL_I915_QUERY, &query) == 0;
 408    }
 409
 410    if (devinfo->gen >= 8) {
 411       /* 4.13+ api required for gen8 - gen9 */
 412       int mask;
 413       struct drm_i915_getparam gp = {
 414          .param = I915_PARAM_SLICE_MASK,
 415          .value = &mask,
 416       };
 417       /* kernel 4.13+ supports this parameter */
 418       return drmIoctl(fd, DRM_IOCTL_I915_GETPARAM, &gp) == 0;
 419    }
 420
 421    if (devinfo->gen == 7)
 422       /* default topology values are correct for HSW */
 423       return true;
 424
 425    /* oa not supported before gen 7*/
 426    return false;
 427 }
 428
 429 static void *
 430 brw_oa_bo_alloc(void *bufmgr, const char *name, uint64_t size)
 431 {
 432    return brw_bo_alloc(bufmgr, name, size, BRW_MEMZONE_OTHER);
 433 }
 434
 435 static void
 436 brw_oa_emit_mi_report_perf_count(void *c,
 437                                  void *bo,
 438                                  uint32_t offset_in_bytes,
 439                                  uint32_t report_id)
 440 {
 441    struct brw_context *ctx = c;
 442    ctx->vtbl.emit_mi_report_perf_count(ctx,
 443                                        bo,
 444                                        offset_in_bytes,
 445                                        report_id);
 446 }
 447
 448 typedef void (*bo_unreference_t)(void *);
 449 typedef void *(*bo_map_t)(void *, void *, unsigned flags);
 450 typedef void (*bo_unmap_t)(void *);
 451 typedef void (* emit_mi_report_t)(void *, void *, uint32_t, uint32_t);
 452 typedef void (*emit_mi_flush_t)(void *);
 453
 454 static void
 455 brw_oa_batchbuffer_flush(void *c, const char *file, int line)
 456 {
 457    struct brw_context *ctx = c;
 458    _intel_batchbuffer_flush_fence(ctx, -1, NULL, file,  line);
 459 }
 460
 461 typedef void (*capture_frequency_stat_register_t)(void *, void *, uint32_t );
 462 typedef void (*store_register_mem64_t)(void *ctx, void *bo,
 463                                        uint32_t reg, uint32_t offset);
 464 typedef bool (*batch_references_t)(void *batch, void *bo);
 465 typedef void (*bo_wait_rendering_t)(void *bo);
 466 typedef int (*bo_busy_t)(void *bo);
 467
 468 static unsigned
 469 brw_init_perf_query_info(struct gl_context *ctx)
 470 {
 471    struct brw_context *brw = brw_context(ctx);
 472    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 473
 474    struct gen_perf_context *perf_ctx = brw->perf_ctx;
 475    struct gen_perf_config *perf_cfg = gen_perf_config(perf_ctx);
 476
 477    if (perf_cfg)
 478       return perf_cfg->n_queries;
 479
 480    if (!oa_metrics_kernel_support(brw->screen->driScrnPriv->fd, devinfo))
 481       return 0;
 482
 483    perf_cfg = gen_perf_new(ctx);
 484
 485    perf_cfg->vtbl.bo_alloc = brw_oa_bo_alloc;
 486    perf_cfg->vtbl.bo_unreference = (bo_unreference_t)brw_bo_unreference;
 487    perf_cfg->vtbl.bo_map = (bo_map_t)brw_bo_map;
 488    perf_cfg->vtbl.bo_unmap = (bo_unmap_t)brw_bo_unmap;
 489    perf_cfg->vtbl.emit_mi_flush = (emit_mi_flush_t)brw_emit_mi_flush;
 490    perf_cfg->vtbl.emit_mi_report_perf_count =
 491       (emit_mi_report_t)brw_oa_emit_mi_report_perf_count;
 492    perf_cfg->vtbl.batchbuffer_flush = brw_oa_batchbuffer_flush;
 493    perf_cfg->vtbl.capture_frequency_stat_register =
 494       (capture_frequency_stat_register_t) capture_frequency_stat_register;
 495    perf_cfg->vtbl.store_register_mem64 =
 496       (store_register_mem64_t) brw_store_register_mem64;
 497    perf_cfg->vtbl.batch_references = (batch_references_t)brw_batch_references;
 498    perf_cfg->vtbl.bo_wait_rendering = (bo_wait_rendering_t)brw_bo_wait_rendering;
 499    perf_cfg->vtbl.bo_busy = (bo_busy_t)brw_bo_busy;
 500
 501    gen_perf_init_context(perf_ctx, perf_cfg, brw, brw->bufmgr, devinfo,
 502                          brw->hw_ctx, brw->screen->driScrnPriv->fd);
 503    gen_perf_init_metrics(perf_cfg, devinfo, brw->screen->driScrnPriv->fd);
 504
 505    return perf_cfg->n_queries;
 506 }
 507
 508 void
 509 brw_init_performance_queries(struct brw_context *brw)
 510 {
 511    struct gl_context *ctx = &brw->ctx;
 512
 513    ctx->Driver.InitPerfQueryInfo = brw_init_perf_query_info;
 514    ctx->Driver.GetPerfQueryInfo = brw_get_perf_query_info;
 515    ctx->Driver.GetPerfCounterInfo = brw_get_perf_counter_info;
 516    ctx->Driver.NewPerfQueryObject = brw_new_perf_query_object;
 517    ctx->Driver.DeletePerfQuery = brw_delete_perf_query;
 518    ctx->Driver.BeginPerfQuery = brw_begin_perf_query;
 519    ctx->Driver.EndPerfQuery = brw_end_perf_query;
 520    ctx->Driver.WaitPerfQuery = brw_wait_perf_query;
 521    ctx->Driver.IsPerfQueryReady = brw_is_perf_query_ready;
 522    ctx->Driver.GetPerfQueryData = brw_get_perf_query_data;
 523 }