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