src/gallium/drivers/freedreno/freedreno_batch.c

   1 /*
   2  * Copyright (C) 2016 Rob Clark <robclark@freedesktop.org>
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "util/list.h"
  28 #include "util/set.h"
  29 #include "util/hash_table.h"
  30 #include "util/u_string.h"
  31
  32 #include "freedreno_batch.h"
  33 #include "freedreno_context.h"
  34 #include "freedreno_fence.h"
  35 #include "freedreno_resource.h"
  36 #include "freedreno_query_hw.h"
  37
  38 static void
  39 batch_init(struct fd_batch *batch)
  40 {
  41         struct fd_context *ctx = batch->ctx;
  42         unsigned size = 0;
  43
  44         if (ctx->screen->reorder)
  45                 util_queue_fence_init(&batch->flush_fence);
  46
  47         /* if kernel is too old to support unlimited # of cmd buffers, we
  48          * have no option but to allocate large worst-case sizes so that
  49          * we don't need to grow the ringbuffer.  Performance is likely to
  50          * suffer, but there is no good alternative.
  51          *
  52          * XXX I think we can just require new enough kernel for this?
  53          */
  54         if ((fd_device_version(ctx->screen->dev) < FD_VERSION_UNLIMITED_CMDS) ||
  55                         (fd_mesa_debug & FD_DBG_NOGROW)){
  56                 size = 0x100000;
  57         }
  58
  59         batch->submit = fd_submit_new(ctx->pipe);
  60         if (batch->nondraw) {
  61                 batch->draw = fd_submit_new_ringbuffer(batch->submit, size,
  62                                 FD_RINGBUFFER_PRIMARY | FD_RINGBUFFER_GROWABLE);
  63         } else {
  64                 batch->gmem = fd_submit_new_ringbuffer(batch->submit, size,
  65                                 FD_RINGBUFFER_PRIMARY | FD_RINGBUFFER_GROWABLE);
  66                 batch->draw = fd_submit_new_ringbuffer(batch->submit, size,
  67                                 FD_RINGBUFFER_GROWABLE);
  68
  69                 if (ctx->screen->gpu_id < 600) {
  70                         batch->binning = fd_submit_new_ringbuffer(batch->submit,
  71                                         size, FD_RINGBUFFER_GROWABLE);
  72                 }
  73         }
  74
  75         batch->in_fence_fd = -1;
  76         batch->fence = fd_fence_create(batch);
  77
  78         batch->cleared = 0;
  79         batch->invalidated = 0;
  80         batch->restore = batch->resolve = 0;
  81         batch->needs_flush = false;
  82         batch->flushed = false;
  83         batch->gmem_reason = 0;
  84         batch->num_draws = 0;
  85         batch->stage = FD_STAGE_NULL;
  86
  87         fd_reset_wfi(batch);
  88
  89         util_dynarray_init(&batch->draw_patches, NULL);
  90
  91         if (is_a3xx(ctx->screen))
  92                 util_dynarray_init(&batch->rbrc_patches, NULL);
  93
  94         util_dynarray_init(&batch->gmem_patches, NULL);
  95
  96         assert(batch->resources->entries == 0);
  97
  98         util_dynarray_init(&batch->samples, NULL);
  99 }
 100
 101 struct fd_batch *
 102 fd_batch_create(struct fd_context *ctx, bool nondraw)
 103 {
 104         struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
 105
 106         if (!batch)
 107                 return NULL;
 108
 109         DBG("%p", batch);
 110
 111         pipe_reference_init(&batch->reference, 1);
 112         batch->ctx = ctx;
 113         batch->nondraw = nondraw;
 114
 115         batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 116                         _mesa_key_pointer_equal);
 117
 118         batch_init(batch);
 119
 120         return batch;
 121 }
 122
 123 static void
 124 batch_fini(struct fd_batch *batch)
 125 {
 126         DBG("%p", batch);
 127
 128         pipe_resource_reference(&batch->query_buf, NULL);
 129
 130         if (batch->in_fence_fd != -1)
 131                 close(batch->in_fence_fd);
 132
 133         /* in case batch wasn't flushed but fence was created: */
 134         fd_fence_populate(batch->fence, 0, -1);
 135
 136         fd_fence_ref(NULL, &batch->fence, NULL);
 137
 138         fd_ringbuffer_del(batch->draw);
 139         if (!batch->nondraw) {
 140                 if (batch->binning)
 141                         fd_ringbuffer_del(batch->binning);
 142                 fd_ringbuffer_del(batch->gmem);
 143         } else {
 144                 debug_assert(!batch->binning);
 145                 debug_assert(!batch->gmem);
 146         }
 147
 148         if (batch->lrz_clear) {
 149                 fd_ringbuffer_del(batch->lrz_clear);
 150                 batch->lrz_clear = NULL;
 151         }
 152
 153         if (batch->tile_setup) {
 154                 fd_ringbuffer_del(batch->tile_setup);
 155                 batch->tile_setup = NULL;
 156         }
 157
 158         if (batch->tile_fini) {
 159                 fd_ringbuffer_del(batch->tile_fini);
 160                 batch->tile_fini = NULL;
 161         }
 162
 163         fd_submit_del(batch->submit);
 164
 165         util_dynarray_fini(&batch->draw_patches);
 166
 167         if (is_a3xx(batch->ctx->screen))
 168                 util_dynarray_fini(&batch->rbrc_patches);
 169
 170         util_dynarray_fini(&batch->gmem_patches);
 171
 172         while (batch->samples.size > 0) {
 173                 struct fd_hw_sample *samp =
 174                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 175                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 176         }
 177         util_dynarray_fini(&batch->samples);
 178
 179         if (batch->ctx->screen->reorder)
 180                 util_queue_fence_destroy(&batch->flush_fence);
 181 }
 182
 183 static void
 184 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 185 {
 186         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 187         struct fd_batch *dep;
 188
 189         foreach_batch(dep, cache, batch->dependents_mask) {
 190                 if (flush)
 191                         fd_batch_flush(dep, false, false);
 192                 fd_batch_reference(&dep, NULL);
 193         }
 194
 195         batch->dependents_mask = 0;
 196 }
 197
 198 static void
 199 batch_reset_resources_locked(struct fd_batch *batch)
 200 {
 201         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 202
 203         set_foreach(batch->resources, entry) {
 204                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 205                 _mesa_set_remove(batch->resources, entry);
 206                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 207                 rsc->batch_mask &= ~(1 << batch->idx);
 208                 if (rsc->write_batch == batch)
 209                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 210         }
 211 }
 212
 213 static void
 214 batch_reset_resources(struct fd_batch *batch)
 215 {
 216         mtx_lock(&batch->ctx->screen->lock);
 217         batch_reset_resources_locked(batch);
 218         mtx_unlock(&batch->ctx->screen->lock);
 219 }
 220
 221 static void
 222 batch_reset(struct fd_batch *batch)
 223 {
 224         DBG("%p", batch);
 225
 226         fd_batch_sync(batch);
 227
 228         batch_flush_reset_dependencies(batch, false);
 229         batch_reset_resources(batch);
 230
 231         batch_fini(batch);
 232         batch_init(batch);
 233 }
 234
 235 void
 236 fd_batch_reset(struct fd_batch *batch)
 237 {
 238         if (batch->needs_flush)
 239                 batch_reset(batch);
 240 }
 241
 242 void
 243 __fd_batch_destroy(struct fd_batch *batch)
 244 {
 245         struct fd_context *ctx = batch->ctx;
 246
 247         DBG("%p", batch);
 248
 249         fd_context_assert_locked(batch->ctx);
 250
 251         fd_bc_invalidate_batch(batch, true);
 252
 253         batch_reset_resources_locked(batch);
 254         debug_assert(batch->resources->entries == 0);
 255         _mesa_set_destroy(batch->resources, NULL);
 256
 257         fd_context_unlock(ctx);
 258         batch_flush_reset_dependencies(batch, false);
 259         debug_assert(batch->dependents_mask == 0);
 260
 261         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 262         batch_fini(batch);
 263         free(batch);
 264         fd_context_lock(ctx);
 265 }
 266
 267 void
 268 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 269 {
 270         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 271 }
 272
 273 void
 274 fd_batch_sync(struct fd_batch *batch)
 275 {
 276         if (!batch->ctx->screen->reorder)
 277                 return;
 278         util_queue_fence_wait(&batch->flush_fence);
 279 }
 280
 281 static void
 282 batch_flush_func(void *job, int id)
 283 {
 284         struct fd_batch *batch = job;
 285
 286         DBG("%p", batch);
 287
 288         fd_gmem_render_tiles(batch);
 289         batch_reset_resources(batch);
 290 }
 291
 292 static void
 293 batch_cleanup_func(void *job, int id)
 294 {
 295         struct fd_batch *batch = job;
 296         fd_batch_reference(&batch, NULL);
 297 }
 298
 299 static void
 300 batch_flush(struct fd_batch *batch, bool force)
 301 {
 302         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 303
 304         if (batch->flushed)
 305                 return;
 306
 307         batch->needs_flush = false;
 308
 309         /* close out the draw cmds by making sure any active queries are
 310          * paused:
 311          */
 312         fd_batch_set_stage(batch, FD_STAGE_NULL);
 313
 314         batch_flush_reset_dependencies(batch, true);
 315
 316         batch->flushed = true;
 317
 318         if (batch->ctx->screen->reorder) {
 319                 struct fd_batch *tmp = NULL;
 320                 fd_batch_reference(&tmp, batch);
 321
 322                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 323                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 324
 325                 util_queue_add_job(&batch->ctx->flush_queue,
 326                                 batch, &batch->flush_fence,
 327                                 batch_flush_func, batch_cleanup_func);
 328         } else {
 329                 fd_gmem_render_tiles(batch);
 330                 batch_reset_resources(batch);
 331         }
 332
 333         debug_assert(batch->reference.count > 0);
 334
 335         mtx_lock(&batch->ctx->screen->lock);
 336         fd_bc_invalidate_batch(batch, false);
 337         mtx_unlock(&batch->ctx->screen->lock);
 338 }
 339
 340 /* NOTE: could drop the last ref to batch
 341  *
 342  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 343  *   to kernel before this returns, as opposed to just being queued to be
 344  *   flushed
 345  * @force: force a flush even if no rendering, mostly useful if you need
 346  *   a fence to sync on
 347  */
 348 void
 349 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 350 {
 351         struct fd_batch *tmp = NULL;
 352         bool newbatch = false;
 353
 354         /* NOTE: we need to hold an extra ref across the body of flush,
 355          * since the last ref to this batch could be dropped when cleaning
 356          * up used_resources
 357          */
 358         fd_batch_reference(&tmp, batch);
 359
 360         if (batch == batch->ctx->batch) {
 361                 batch->ctx->batch = NULL;
 362                 newbatch = true;
 363         }
 364
 365         batch_flush(tmp, force);
 366
 367         if (newbatch) {
 368                 struct fd_context *ctx = batch->ctx;
 369                 struct fd_batch *new_batch;
 370
 371                 if (ctx->screen->reorder) {
 372                         /* defer allocating new batch until one is needed for rendering
 373                          * to avoid unused batches for apps that create many contexts
 374                          */
 375                         new_batch = NULL;
 376                 } else {
 377                         new_batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, false);
 378                         util_copy_framebuffer_state(&new_batch->framebuffer, &batch->framebuffer);
 379                 }
 380
 381                 fd_batch_reference(&batch, NULL);
 382                 ctx->batch = new_batch;
 383                 fd_context_all_dirty(ctx);
 384         }
 385
 386         if (sync)
 387                 fd_batch_sync(tmp);
 388
 389         fd_batch_reference(&tmp, NULL);
 390 }
 391
 392 /* does 'batch' depend directly or indirectly on 'other' ? */
 393 static bool
 394 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 395 {
 396         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 397         struct fd_batch *dep;
 398
 399         if (batch->dependents_mask & (1 << other->idx))
 400                 return true;
 401
 402         foreach_batch(dep, cache, batch->dependents_mask)
 403                 if (batch_depends_on(batch, dep))
 404                         return true;
 405
 406         return false;
 407 }
 408
 409 void
 410 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 411 {
 412         if (batch->dependents_mask & (1 << dep->idx))
 413                 return;
 414
 415         /* a loop should not be possible */
 416         debug_assert(!batch_depends_on(dep, batch));
 417
 418         struct fd_batch *other = NULL;
 419         fd_batch_reference_locked(&other, dep);
 420         batch->dependents_mask |= (1 << dep->idx);
 421         DBG("%p: added dependency on %p", batch, dep);
 422 }
 423
 424 static void
 425 flush_write_batch(struct fd_resource *rsc)
 426 {
 427         struct fd_batch *b = NULL;
 428         fd_batch_reference(&b, rsc->write_batch);
 429
 430         mtx_unlock(&b->ctx->screen->lock);
 431         fd_batch_flush(b, true, false);
 432         mtx_lock(&b->ctx->screen->lock);
 433
 434         fd_bc_invalidate_batch(b, false);
 435         fd_batch_reference_locked(&b, NULL);
 436 }
 437
 438 void
 439 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 440 {
 441         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 442
 443         if (rsc->stencil)
 444                 fd_batch_resource_used(batch, rsc->stencil, write);
 445
 446         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 447
 448         if (write)
 449                 rsc->valid = true;
 450
 451         /* note, invalidate write batch, to avoid further writes to rsc
 452          * resulting in a write-after-read hazard.
 453          */
 454
 455         if (write) {
 456                 /* if we are pending read or write by any other batch: */
 457                 if (rsc->batch_mask & ~(1 << batch->idx)) {
 458                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 459                         struct fd_batch *dep;
 460
 461                         if (rsc->write_batch && rsc->write_batch != batch)
 462                                 flush_write_batch(rsc);
 463
 464                         foreach_batch(dep, cache, rsc->batch_mask) {
 465                                 struct fd_batch *b = NULL;
 466                                 if (dep == batch)
 467                                         continue;
 468                                 /* note that batch_add_dep could flush and unref dep, so
 469                                  * we need to hold a reference to keep it live for the
 470                                  * fd_bc_invalidate_batch()
 471                                  */
 472                                 fd_batch_reference(&b, dep);
 473                                 fd_batch_add_dep(batch, b);
 474                                 fd_bc_invalidate_batch(b, false);
 475                                 fd_batch_reference_locked(&b, NULL);
 476                         }
 477                 }
 478                 fd_batch_reference_locked(&rsc->write_batch, batch);
 479         } else {
 480                 /* If reading a resource pending a write, go ahead and flush the
 481                  * writer.  This avoids situations where we end up having to
 482                  * flush the current batch in _resource_used()
 483                  */
 484                 if (rsc->write_batch && rsc->write_batch != batch)
 485                         flush_write_batch(rsc);
 486         }
 487
 488         if (rsc->batch_mask & (1 << batch->idx)) {
 489                 debug_assert(_mesa_set_search(batch->resources, rsc));
 490                 return;
 491         }
 492
 493         debug_assert(!_mesa_set_search(batch->resources, rsc));
 494
 495         _mesa_set_add(batch->resources, rsc);
 496         rsc->batch_mask |= (1 << batch->idx);
 497 }
 498
 499 void
 500 fd_batch_check_size(struct fd_batch *batch)
 501 {
 502         debug_assert(!batch->flushed);
 503
 504         if (unlikely(fd_mesa_debug & FD_DBG_FLUSH)) {
 505                 fd_batch_flush(batch, true, false);
 506                 return;
 507         }
 508
 509         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 510                 return;
 511
 512         struct fd_ringbuffer *ring = batch->draw;
 513         if ((ring->cur - ring->start) > (ring->size/4 - 0x1000))
 514                 fd_batch_flush(batch, true, false);
 515 }
 516
 517 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 518  * been one since last draw:
 519  */
 520 void
 521 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 522 {
 523         if (batch->needs_wfi) {
 524                 if (batch->ctx->screen->gpu_id >= 500)
 525                         OUT_WFI5(ring);
 526                 else
 527                         OUT_WFI(ring);
 528                 batch->needs_wfi = false;
 529         }
 530 }