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         if (batch->lrz_clear) {
 148                 fd_ringbuffer_del(batch->lrz_clear);
 149                 batch->lrz_clear = NULL;
 150         }
 151
 152         fd_submit_del(batch->submit);
 153
 154         util_dynarray_fini(&batch->draw_patches);
 155
 156         if (is_a3xx(batch->ctx->screen))
 157                 util_dynarray_fini(&batch->rbrc_patches);
 158
 159         util_dynarray_fini(&batch->gmem_patches);
 160
 161         while (batch->samples.size > 0) {
 162                 struct fd_hw_sample *samp =
 163                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 164                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 165         }
 166         util_dynarray_fini(&batch->samples);
 167
 168         if (batch->ctx->screen->reorder)
 169                 util_queue_fence_destroy(&batch->flush_fence);
 170 }
 171
 172 static void
 173 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 174 {
 175         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 176         struct fd_batch *dep;
 177
 178         foreach_batch(dep, cache, batch->dependents_mask) {
 179                 if (flush)
 180                         fd_batch_flush(dep, false, false);
 181                 fd_batch_reference(&dep, NULL);
 182         }
 183
 184         batch->dependents_mask = 0;
 185 }
 186
 187 static void
 188 batch_reset_resources_locked(struct fd_batch *batch)
 189 {
 190         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 191
 192         set_foreach(batch->resources, entry) {
 193                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 194                 _mesa_set_remove(batch->resources, entry);
 195                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 196                 rsc->batch_mask &= ~(1 << batch->idx);
 197                 if (rsc->write_batch == batch)
 198                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 199         }
 200 }
 201
 202 static void
 203 batch_reset_resources(struct fd_batch *batch)
 204 {
 205         mtx_lock(&batch->ctx->screen->lock);
 206         batch_reset_resources_locked(batch);
 207         mtx_unlock(&batch->ctx->screen->lock);
 208 }
 209
 210 static void
 211 batch_reset(struct fd_batch *batch)
 212 {
 213         DBG("%p", batch);
 214
 215         fd_batch_sync(batch);
 216
 217         batch_flush_reset_dependencies(batch, false);
 218         batch_reset_resources(batch);
 219
 220         batch_fini(batch);
 221         batch_init(batch);
 222 }
 223
 224 void
 225 fd_batch_reset(struct fd_batch *batch)
 226 {
 227         if (batch->needs_flush)
 228                 batch_reset(batch);
 229 }
 230
 231 void
 232 __fd_batch_destroy(struct fd_batch *batch)
 233 {
 234         struct fd_context *ctx = batch->ctx;
 235
 236         DBG("%p", batch);
 237
 238         fd_context_assert_locked(batch->ctx);
 239
 240         fd_bc_invalidate_batch(batch, true);
 241
 242         batch_reset_resources_locked(batch);
 243         debug_assert(batch->resources->entries == 0);
 244         _mesa_set_destroy(batch->resources, NULL);
 245
 246         fd_context_unlock(ctx);
 247         batch_flush_reset_dependencies(batch, false);
 248         debug_assert(batch->dependents_mask == 0);
 249
 250         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 251         batch_fini(batch);
 252         free(batch);
 253         fd_context_lock(ctx);
 254 }
 255
 256 void
 257 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 258 {
 259         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 260 }
 261
 262 void
 263 fd_batch_sync(struct fd_batch *batch)
 264 {
 265         if (!batch->ctx->screen->reorder)
 266                 return;
 267         util_queue_fence_wait(&batch->flush_fence);
 268 }
 269
 270 static void
 271 batch_flush_func(void *job, int id)
 272 {
 273         struct fd_batch *batch = job;
 274
 275         DBG("%p", batch);
 276
 277         fd_gmem_render_tiles(batch);
 278         batch_reset_resources(batch);
 279 }
 280
 281 static void
 282 batch_cleanup_func(void *job, int id)
 283 {
 284         struct fd_batch *batch = job;
 285         fd_batch_reference(&batch, NULL);
 286 }
 287
 288 static void
 289 batch_flush(struct fd_batch *batch, bool force)
 290 {
 291         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 292
 293         if (batch->flushed)
 294                 return;
 295
 296         batch->needs_flush = false;
 297
 298         /* close out the draw cmds by making sure any active queries are
 299          * paused:
 300          */
 301         fd_batch_set_stage(batch, FD_STAGE_NULL);
 302
 303         batch_flush_reset_dependencies(batch, true);
 304
 305         batch->flushed = true;
 306
 307         if (batch->ctx->screen->reorder) {
 308                 struct fd_batch *tmp = NULL;
 309                 fd_batch_reference(&tmp, batch);
 310
 311                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 312                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 313
 314                 util_queue_add_job(&batch->ctx->flush_queue,
 315                                 batch, &batch->flush_fence,
 316                                 batch_flush_func, batch_cleanup_func);
 317         } else {
 318                 fd_gmem_render_tiles(batch);
 319                 batch_reset_resources(batch);
 320         }
 321
 322         debug_assert(batch->reference.count > 0);
 323
 324         mtx_lock(&batch->ctx->screen->lock);
 325         fd_bc_invalidate_batch(batch, false);
 326         mtx_unlock(&batch->ctx->screen->lock);
 327 }
 328
 329 /* NOTE: could drop the last ref to batch
 330  *
 331  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 332  *   to kernel before this returns, as opposed to just being queued to be
 333  *   flushed
 334  * @force: force a flush even if no rendering, mostly useful if you need
 335  *   a fence to sync on
 336  */
 337 void
 338 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 339 {
 340         struct fd_batch *tmp = NULL;
 341         bool newbatch = false;
 342
 343         /* NOTE: we need to hold an extra ref across the body of flush,
 344          * since the last ref to this batch could be dropped when cleaning
 345          * up used_resources
 346          */
 347         fd_batch_reference(&tmp, batch);
 348
 349         if (batch == batch->ctx->batch) {
 350                 batch->ctx->batch = NULL;
 351                 newbatch = true;
 352         }
 353
 354         batch_flush(tmp, force);
 355
 356         if (newbatch) {
 357                 struct fd_context *ctx = batch->ctx;
 358                 struct fd_batch *new_batch;
 359
 360                 if (ctx->screen->reorder) {
 361                         /* defer allocating new batch until one is needed for rendering
 362                          * to avoid unused batches for apps that create many contexts
 363                          */
 364                         new_batch = NULL;
 365                 } else {
 366                         new_batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, false);
 367                         util_copy_framebuffer_state(&new_batch->framebuffer, &batch->framebuffer);
 368                 }
 369
 370                 fd_batch_reference(&batch, NULL);
 371                 ctx->batch = new_batch;
 372                 fd_context_all_dirty(ctx);
 373         }
 374
 375         if (sync)
 376                 fd_batch_sync(tmp);
 377
 378         fd_batch_reference(&tmp, NULL);
 379 }
 380
 381 /* does 'batch' depend directly or indirectly on 'other' ? */
 382 static bool
 383 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 384 {
 385         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 386         struct fd_batch *dep;
 387
 388         if (batch->dependents_mask & (1 << other->idx))
 389                 return true;
 390
 391         foreach_batch(dep, cache, batch->dependents_mask)
 392                 if (batch_depends_on(batch, dep))
 393                         return true;
 394
 395         return false;
 396 }
 397
 398 void
 399 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 400 {
 401         if (batch->dependents_mask & (1 << dep->idx))
 402                 return;
 403
 404         /* a loop should not be possible */
 405         debug_assert(!batch_depends_on(dep, batch));
 406
 407         struct fd_batch *other = NULL;
 408         fd_batch_reference_locked(&other, dep);
 409         batch->dependents_mask |= (1 << dep->idx);
 410         DBG("%p: added dependency on %p", batch, dep);
 411 }
 412
 413 static void
 414 flush_write_batch(struct fd_resource *rsc)
 415 {
 416         struct fd_batch *b = NULL;
 417         fd_batch_reference(&b, rsc->write_batch);
 418
 419         mtx_unlock(&b->ctx->screen->lock);
 420         fd_batch_flush(b, true, false);
 421         mtx_lock(&b->ctx->screen->lock);
 422
 423         fd_bc_invalidate_batch(b, false);
 424         fd_batch_reference_locked(&b, NULL);
 425 }
 426
 427 void
 428 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 429 {
 430         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 431
 432         if (rsc->stencil)
 433                 fd_batch_resource_used(batch, rsc->stencil, write);
 434
 435         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 436
 437         if (write)
 438                 rsc->valid = true;
 439
 440         /* note, invalidate write batch, to avoid further writes to rsc
 441          * resulting in a write-after-read hazard.
 442          */
 443
 444         if (write) {
 445                 /* if we are pending read or write by any other batch: */
 446                 if (rsc->batch_mask & ~(1 << batch->idx)) {
 447                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 448                         struct fd_batch *dep;
 449
 450                         if (rsc->write_batch && rsc->write_batch != batch)
 451                                 flush_write_batch(rsc);
 452
 453                         foreach_batch(dep, cache, rsc->batch_mask) {
 454                                 struct fd_batch *b = NULL;
 455                                 if (dep == batch)
 456                                         continue;
 457                                 /* note that batch_add_dep could flush and unref dep, so
 458                                  * we need to hold a reference to keep it live for the
 459                                  * fd_bc_invalidate_batch()
 460                                  */
 461                                 fd_batch_reference(&b, dep);
 462                                 fd_batch_add_dep(batch, b);
 463                                 fd_bc_invalidate_batch(b, false);
 464                                 fd_batch_reference_locked(&b, NULL);
 465                         }
 466                 }
 467                 fd_batch_reference_locked(&rsc->write_batch, batch);
 468         } else {
 469                 /* If reading a resource pending a write, go ahead and flush the
 470                  * writer.  This avoids situations where we end up having to
 471                  * flush the current batch in _resource_used()
 472                  */
 473                 if (rsc->write_batch && rsc->write_batch != batch)
 474                         flush_write_batch(rsc);
 475         }
 476
 477         if (rsc->batch_mask & (1 << batch->idx)) {
 478                 debug_assert(_mesa_set_search(batch->resources, rsc));
 479                 return;
 480         }
 481
 482         debug_assert(!_mesa_set_search(batch->resources, rsc));
 483
 484         _mesa_set_add(batch->resources, rsc);
 485         rsc->batch_mask |= (1 << batch->idx);
 486 }
 487
 488 void
 489 fd_batch_check_size(struct fd_batch *batch)
 490 {
 491         debug_assert(!batch->flushed);
 492
 493         if (unlikely(fd_mesa_debug & FD_DBG_FLUSH)) {
 494                 fd_batch_flush(batch, true, false);
 495                 return;
 496         }
 497
 498         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 499                 return;
 500
 501         struct fd_ringbuffer *ring = batch->draw;
 502         if ((ring->cur - ring->start) > (ring->size/4 - 0x1000))
 503                 fd_batch_flush(batch, true, false);
 504 }
 505
 506 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 507  * been one since last draw:
 508  */
 509 void
 510 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 511 {
 512         if (batch->needs_wfi) {
 513                 if (batch->ctx->screen->gpu_id >= 500)
 514                         OUT_WFI5(ring);
 515                 else
 516                         OUT_WFI(ring);
 517                 batch->needs_wfi = false;
 518         }
 519 }