src/gallium/drivers/freedreno/freedreno_query_hw.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #include "pipe/p_state.h"
  30 #include "util/u_memory.h"
  31 #include "util/u_inlines.h"
  32
  33 #include "freedreno_query_hw.h"
  34 #include "freedreno_context.h"
  35 #include "freedreno_util.h"
  36
  37 struct fd_hw_sample_period {
  38         struct fd_hw_sample *start, *end;
  39         struct list_head list;
  40 };
  41
  42 /* maps query_type to sample provider idx: */
  43 static int pidx(unsigned query_type)
  44 {
  45         switch (query_type) {
  46         case PIPE_QUERY_OCCLUSION_COUNTER:
  47                 return 0;
  48         case PIPE_QUERY_OCCLUSION_PREDICATE:
  49                 return 1;
  50         case PIPE_QUERY_TIME_ELAPSED:
  51                 return 2;
  52         default:
  53                 return -1;
  54         }
  55 }
  56
  57 static struct fd_hw_sample *
  58 get_sample(struct fd_context *ctx, struct fd_ringbuffer *ring,
  59                 unsigned query_type)
  60 {
  61         struct fd_hw_sample *samp = NULL;
  62         int idx = pidx(query_type);
  63
  64         assume(idx >= 0);   /* query never would have been created otherwise */
  65
  66         if (!ctx->sample_cache[idx]) {
  67                 ctx->sample_cache[idx] =
  68                         ctx->sample_providers[idx]->get_sample(ctx, ring);
  69         }
  70
  71         fd_hw_sample_reference(ctx, &samp, ctx->sample_cache[idx]);
  72
  73         return samp;
  74 }
  75
  76 static void
  77 clear_sample_cache(struct fd_context *ctx)
  78 {
  79         int i;
  80
  81         for (i = 0; i < ARRAY_SIZE(ctx->sample_cache); i++)
  82                 fd_hw_sample_reference(ctx, &ctx->sample_cache[i], NULL);
  83 }
  84
  85 static bool
  86 is_active(struct fd_hw_query *hq, enum fd_render_stage stage)
  87 {
  88         return !!(hq->provider->active & stage);
  89 }
  90
  91
  92 static void
  93 resume_query(struct fd_context *ctx, struct fd_hw_query *hq,
  94                 struct fd_ringbuffer *ring)
  95 {
  96         int idx = pidx(hq->provider->query_type);
  97         assert(!hq->period);
  98         ctx->active_providers |= (1 << idx);
  99         hq->period = util_slab_alloc(&ctx->sample_period_pool);
 100         list_inithead(&hq->period->list);
 101         hq->period->start = get_sample(ctx, ring, hq->base.type);
 102         /* NOTE: util_slab_alloc() does not zero out the buffer: */
 103         hq->period->end = NULL;
 104 }
 105
 106 static void
 107 pause_query(struct fd_context *ctx, struct fd_hw_query *hq,
 108                 struct fd_ringbuffer *ring)
 109 {
 110         int idx = pidx(hq->provider->query_type);
 111         assert(hq->period && !hq->period->end);
 112         assert(ctx->active_providers & (1 << idx));
 113         hq->period->end = get_sample(ctx, ring, hq->base.type);
 114         list_addtail(&hq->period->list, &hq->current_periods);
 115         hq->period = NULL;
 116 }
 117
 118 static void
 119 destroy_periods(struct fd_context *ctx, struct list_head *list)
 120 {
 121         struct fd_hw_sample_period *period, *s;
 122         LIST_FOR_EACH_ENTRY_SAFE(period, s, list, list) {
 123                 fd_hw_sample_reference(ctx, &period->start, NULL);
 124                 fd_hw_sample_reference(ctx, &period->end, NULL);
 125                 list_del(&period->list);
 126                 util_slab_free(&ctx->sample_period_pool, period);
 127         }
 128 }
 129
 130 static void
 131 fd_hw_destroy_query(struct fd_context *ctx, struct fd_query *q)
 132 {
 133         struct fd_hw_query *hq = fd_hw_query(q);
 134
 135         destroy_periods(ctx, &hq->periods);
 136         destroy_periods(ctx, &hq->current_periods);
 137         list_del(&hq->list);
 138
 139         free(hq);
 140 }
 141
 142 static boolean
 143 fd_hw_begin_query(struct fd_context *ctx, struct fd_query *q)
 144 {
 145         struct fd_hw_query *hq = fd_hw_query(q);
 146         if (q->active)
 147                 return false;
 148
 149         /* begin_query() should clear previous results: */
 150         destroy_periods(ctx, &hq->periods);
 151
 152         if (is_active(hq, ctx->stage))
 153                 resume_query(ctx, hq, ctx->ring);
 154
 155         q->active = true;
 156
 157         /* add to active list: */
 158         list_del(&hq->list);
 159         list_addtail(&hq->list, &ctx->active_queries);
 160    return true;
 161 }
 162
 163 static void
 164 fd_hw_end_query(struct fd_context *ctx, struct fd_query *q)
 165 {
 166         struct fd_hw_query *hq = fd_hw_query(q);
 167         /* there are a couple special cases, which don't have
 168          * a matching ->begin_query():
 169          */
 170         if (skip_begin_query(q->type) && !q->active) {
 171                 fd_hw_begin_query(ctx, q);
 172         }
 173         if (!q->active)
 174                 return;
 175         if (is_active(hq, ctx->stage))
 176                 pause_query(ctx, hq, ctx->ring);
 177         q->active = false;
 178         /* move to current list: */
 179         list_del(&hq->list);
 180         list_addtail(&hq->list, &ctx->current_queries);
 181 }
 182
 183 /* helper to get ptr to specified sample: */
 184 static void * sampptr(struct fd_hw_sample *samp, uint32_t n, void *ptr)
 185 {
 186         return ((char *)ptr) + (samp->tile_stride * n) + samp->offset;
 187 }
 188
 189 static boolean
 190 fd_hw_get_query_result(struct fd_context *ctx, struct fd_query *q,
 191                 boolean wait, union pipe_query_result *result)
 192 {
 193         struct fd_hw_query *hq = fd_hw_query(q);
 194         const struct fd_hw_sample_provider *p = hq->provider;
 195         struct fd_hw_sample_period *period;
 196
 197         if (q->active)
 198                 return false;
 199
 200         /* if the app tries to read back the query result before the
 201          * batch is submitted, that forces us to flush so that there
 202          * are actually results to wait for:
 203          */
 204         if (!LIST_IS_EMPTY(&hq->list)) {
 205                 /* if app didn't actually trigger any cmdstream, then
 206                  * we have nothing to do:
 207                  */
 208                 if (!ctx->needs_flush)
 209                         return true;
 210                 DBG("reading query result forces flush!");
 211                 fd_context_render(&ctx->base);
 212         }
 213
 214         util_query_clear_result(result, q->type);
 215
 216         if (LIST_IS_EMPTY(&hq->periods))
 217                 return true;
 218
 219         assert(LIST_IS_EMPTY(&hq->list));
 220         assert(LIST_IS_EMPTY(&hq->current_periods));
 221         assert(!hq->period);
 222
 223         /* if !wait, then check the last sample (the one most likely to
 224          * not be ready yet) and bail if it is not ready:
 225          */
 226         if (!wait) {
 227                 int ret;
 228
 229                 period = LIST_ENTRY(struct fd_hw_sample_period,
 230                                 hq->periods.prev, list);
 231
 232                 ret = fd_bo_cpu_prep(period->end->bo, ctx->screen->pipe,
 233                                 DRM_FREEDRENO_PREP_READ | DRM_FREEDRENO_PREP_NOSYNC);
 234                 if (ret)
 235                         return false;
 236
 237                 fd_bo_cpu_fini(period->end->bo);
 238         }
 239
 240         /* sum the result across all sample periods: */
 241         LIST_FOR_EACH_ENTRY(period, &hq->periods, list) {
 242                 struct fd_hw_sample *start = period->start;
 243                 struct fd_hw_sample *end = period->end;
 244                 unsigned i;
 245
 246                 /* start and end samples should be from same batch: */
 247                 assert(start->bo == end->bo);
 248                 assert(start->num_tiles == end->num_tiles);
 249
 250                 for (i = 0; i < start->num_tiles; i++) {
 251                         void *ptr;
 252
 253                         fd_bo_cpu_prep(start->bo, ctx->screen->pipe,
 254                                         DRM_FREEDRENO_PREP_READ);
 255
 256                         ptr = fd_bo_map(start->bo);
 257
 258                         p->accumulate_result(ctx, sampptr(period->start, i, ptr),
 259                                         sampptr(period->end, i, ptr), result);
 260
 261                         fd_bo_cpu_fini(start->bo);
 262                 }
 263         }
 264
 265         return true;
 266 }
 267
 268 static const struct fd_query_funcs hw_query_funcs = {
 269                 .destroy_query    = fd_hw_destroy_query,
 270                 .begin_query      = fd_hw_begin_query,
 271                 .end_query        = fd_hw_end_query,
 272                 .get_query_result = fd_hw_get_query_result,
 273 };
 274
 275 struct fd_query *
 276 fd_hw_create_query(struct fd_context *ctx, unsigned query_type)
 277 {
 278         struct fd_hw_query *hq;
 279         struct fd_query *q;
 280         int idx = pidx(query_type);
 281
 282         if ((idx < 0) || !ctx->sample_providers[idx])
 283                 return NULL;
 284
 285         hq = CALLOC_STRUCT(fd_hw_query);
 286         if (!hq)
 287                 return NULL;
 288
 289         hq->provider = ctx->sample_providers[idx];
 290
 291         list_inithead(&hq->periods);
 292         list_inithead(&hq->current_periods);
 293         list_inithead(&hq->list);
 294
 295         q = &hq->base;
 296         q->funcs = &hw_query_funcs;
 297         q->type = query_type;
 298
 299         return q;
 300 }
 301
 302 struct fd_hw_sample *
 303 fd_hw_sample_init(struct fd_context *ctx, uint32_t size)
 304 {
 305         struct fd_hw_sample *samp = util_slab_alloc(&ctx->sample_pool);
 306         pipe_reference_init(&samp->reference, 1);
 307         samp->size = size;
 308         debug_assert(util_is_power_of_two(size));
 309         ctx->next_sample_offset = align(ctx->next_sample_offset, size);
 310         samp->offset = ctx->next_sample_offset;
 311         /* NOTE: util_slab_alloc() does not zero out the buffer: */
 312         samp->bo = NULL;
 313         samp->num_tiles = 0;
 314         samp->tile_stride = 0;
 315         ctx->next_sample_offset += size;
 316         return samp;
 317 }
 318
 319 void
 320 __fd_hw_sample_destroy(struct fd_context *ctx, struct fd_hw_sample *samp)
 321 {
 322         if (samp->bo)
 323                 fd_bo_del(samp->bo);
 324         util_slab_free(&ctx->sample_pool, samp);
 325 }
 326
 327 static void
 328 prepare_sample(struct fd_hw_sample *samp, struct fd_bo *bo,
 329                 uint32_t num_tiles, uint32_t tile_stride)
 330 {
 331         if (samp->bo) {
 332                 assert(samp->bo == bo);
 333                 assert(samp->num_tiles == num_tiles);
 334                 assert(samp->tile_stride == tile_stride);
 335                 return;
 336         }
 337         samp->bo = fd_bo_ref(bo);
 338         samp->num_tiles = num_tiles;
 339         samp->tile_stride = tile_stride;
 340 }
 341
 342 static void
 343 prepare_query(struct fd_hw_query *hq, struct fd_bo *bo,
 344                 uint32_t num_tiles, uint32_t tile_stride)
 345 {
 346         struct fd_hw_sample_period *period, *s;
 347
 348         /* prepare all the samples in the query: */
 349         LIST_FOR_EACH_ENTRY_SAFE(period, s, &hq->current_periods, list) {
 350                 prepare_sample(period->start, bo, num_tiles, tile_stride);
 351                 prepare_sample(period->end, bo, num_tiles, tile_stride);
 352
 353                 /* move from current_periods list to periods list: */
 354                 list_del(&period->list);
 355                 list_addtail(&period->list, &hq->periods);
 356         }
 357 }
 358
 359 static void
 360 prepare_queries(struct fd_context *ctx, struct fd_bo *bo,
 361                 uint32_t num_tiles, uint32_t tile_stride,
 362                 struct list_head *list, bool remove)
 363 {
 364         struct fd_hw_query *hq, *s;
 365         LIST_FOR_EACH_ENTRY_SAFE(hq, s, list, list) {
 366                 prepare_query(hq, bo, num_tiles, tile_stride);
 367                 if (remove)
 368                         list_delinit(&hq->list);
 369         }
 370 }
 371
 372 /* called from gmem code once total storage requirements are known (ie.
 373  * number of samples times number of tiles)
 374  */
 375 void
 376 fd_hw_query_prepare(struct fd_context *ctx, uint32_t num_tiles)
 377 {
 378         uint32_t tile_stride = ctx->next_sample_offset;
 379         struct fd_bo *bo;
 380
 381         if (ctx->query_bo)
 382                 fd_bo_del(ctx->query_bo);
 383
 384         if (tile_stride > 0) {
 385                 bo = fd_bo_new(ctx->dev, tile_stride * num_tiles,
 386                                 DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
 387                                 DRM_FREEDRENO_GEM_TYPE_KMEM);
 388         } else {
 389                 bo = NULL;
 390         }
 391
 392         ctx->query_bo = bo;
 393         ctx->query_tile_stride = tile_stride;
 394
 395         prepare_queries(ctx, bo, num_tiles, tile_stride,
 396                         &ctx->active_queries, false);
 397         prepare_queries(ctx, bo, num_tiles, tile_stride,
 398                         &ctx->current_queries, true);
 399
 400         /* reset things for next batch: */
 401         ctx->next_sample_offset = 0;
 402 }
 403
 404 void
 405 fd_hw_query_prepare_tile(struct fd_context *ctx, uint32_t n,
 406                 struct fd_ringbuffer *ring)
 407 {
 408         uint32_t tile_stride = ctx->query_tile_stride;
 409         uint32_t offset = tile_stride * n;
 410
 411         /* bail if no queries: */
 412         if (tile_stride == 0)
 413                 return;
 414
 415         fd_wfi(ctx, ring);
 416         OUT_PKT0 (ring, HW_QUERY_BASE_REG, 1);
 417         OUT_RELOCW(ring, ctx->query_bo, offset, 0, 0);
 418 }
 419
 420 void
 421 fd_hw_query_set_stage(struct fd_context *ctx, struct fd_ringbuffer *ring,
 422                 enum fd_render_stage stage)
 423 {
 424         /* special case: internal blits (like mipmap level generation)
 425          * go through normal draw path (via util_blitter_blit()).. but
 426          * we need to ignore the FD_STAGE_DRAW which will be set, so we
 427          * don't enable queries which should be paused during internal
 428          * blits:
 429          */
 430         if ((ctx->stage == FD_STAGE_BLIT) &&
 431                         (stage != FD_STAGE_NULL))
 432                 return;
 433
 434         if (stage != ctx->stage) {
 435                 struct fd_hw_query *hq;
 436                 LIST_FOR_EACH_ENTRY(hq, &ctx->active_queries, list) {
 437                         bool was_active = is_active(hq, ctx->stage);
 438                         bool now_active = is_active(hq, stage);
 439
 440                         if (now_active && !was_active)
 441                                 resume_query(ctx, hq, ring);
 442                         else if (was_active && !now_active)
 443                                 pause_query(ctx, hq, ring);
 444                 }
 445         }
 446         clear_sample_cache(ctx);
 447         ctx->stage = stage;
 448 }
 449
 450 /* call the provider->enable() for all the hw queries that were active
 451  * in the current batch.  This sets up perfctr selector regs statically
 452  * for the duration of the batch.
 453  */
 454 void
 455 fd_hw_query_enable(struct fd_context *ctx, struct fd_ringbuffer *ring)
 456 {
 457         for (int idx = 0; idx < MAX_HW_SAMPLE_PROVIDERS; idx++) {
 458                 if (ctx->active_providers & (1 << idx)) {
 459                         assert(ctx->sample_providers[idx]);
 460                         if (ctx->sample_providers[idx]->enable)
 461                                 ctx->sample_providers[idx]->enable(ctx, ring);
 462                 }
 463         }
 464         ctx->active_providers = 0;  /* clear it for next frame */
 465 }
 466
 467 void
 468 fd_hw_query_register_provider(struct pipe_context *pctx,
 469                 const struct fd_hw_sample_provider *provider)
 470 {
 471         struct fd_context *ctx = fd_context(pctx);
 472         int idx = pidx(provider->query_type);
 473
 474         assert((0 <= idx) && (idx < MAX_HW_SAMPLE_PROVIDERS));
 475         assert(!ctx->sample_providers[idx]);
 476
 477         ctx->sample_providers[idx] = provider;
 478 }
 479
 480 void
 481 fd_hw_query_init(struct pipe_context *pctx)
 482 {
 483         struct fd_context *ctx = fd_context(pctx);
 484
 485         util_slab_create(&ctx->sample_pool, sizeof(struct fd_hw_sample),
 486                         16, UTIL_SLAB_SINGLETHREADED);
 487         util_slab_create(&ctx->sample_period_pool, sizeof(struct fd_hw_sample_period),
 488                         16, UTIL_SLAB_SINGLETHREADED);
 489         list_inithead(&ctx->active_queries);
 490         list_inithead(&ctx->current_queries);
 491 }
 492
 493 void
 494 fd_hw_query_fini(struct pipe_context *pctx)
 495 {
 496         struct fd_context *ctx = fd_context(pctx);
 497
 498         util_slab_destroy(&ctx->sample_pool);
 499         util_slab_destroy(&ctx->sample_period_pool);
 500 }