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