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->fast_cleared = 0;
  80         batch->invalidated = 0;
  81         batch->restore = batch->resolve = 0;
  82         batch->needs_flush = false;
  83         batch->flushed = false;
  84         batch->gmem_reason = 0;
  85         batch->num_draws = 0;
  86         batch->num_vertices = 0;
  87         batch->stage = FD_STAGE_NULL;
  88
  89         fd_reset_wfi(batch);
  90
  91         util_dynarray_init(&batch->draw_patches, NULL);
  92
  93         if (is_a2xx(ctx->screen))
  94                 util_dynarray_init(&batch->shader_patches, NULL);
  95
  96         if (is_a3xx(ctx->screen))
  97                 util_dynarray_init(&batch->rbrc_patches, NULL);
  98
  99         assert(batch->resources->entries == 0);
 100
 101         util_dynarray_init(&batch->samples, NULL);
 102 }
 103
 104 struct fd_batch *
 105 fd_batch_create(struct fd_context *ctx, bool nondraw)
 106 {
 107         struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
 108
 109         if (!batch)
 110                 return NULL;
 111
 112         DBG("%p", batch);
 113
 114         pipe_reference_init(&batch->reference, 1);
 115         batch->ctx = ctx;
 116         batch->nondraw = nondraw;
 117
 118         batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 119                         _mesa_key_pointer_equal);
 120
 121         batch_init(batch);
 122
 123         return batch;
 124 }
 125
 126 static void
 127 batch_fini(struct fd_batch *batch)
 128 {
 129         DBG("%p", batch);
 130
 131         pipe_resource_reference(&batch->query_buf, NULL);
 132
 133         if (batch->in_fence_fd != -1)
 134                 close(batch->in_fence_fd);
 135
 136         /* in case batch wasn't flushed but fence was created: */
 137         fd_fence_populate(batch->fence, 0, -1);
 138
 139         fd_fence_ref(NULL, &batch->fence, NULL);
 140
 141         fd_ringbuffer_del(batch->draw);
 142         if (!batch->nondraw) {
 143                 if (batch->binning)
 144                         fd_ringbuffer_del(batch->binning);
 145                 fd_ringbuffer_del(batch->gmem);
 146         } else {
 147                 debug_assert(!batch->binning);
 148                 debug_assert(!batch->gmem);
 149         }
 150
 151         if (batch->lrz_clear) {
 152                 fd_ringbuffer_del(batch->lrz_clear);
 153                 batch->lrz_clear = NULL;
 154         }
 155
 156         if (batch->tile_setup) {
 157                 fd_ringbuffer_del(batch->tile_setup);
 158                 batch->tile_setup = NULL;
 159         }
 160
 161         if (batch->tile_fini) {
 162                 fd_ringbuffer_del(batch->tile_fini);
 163                 batch->tile_fini = NULL;
 164         }
 165
 166         fd_submit_del(batch->submit);
 167
 168         util_dynarray_fini(&batch->draw_patches);
 169
 170         if (is_a2xx(batch->ctx->screen))
 171                 util_dynarray_fini(&batch->shader_patches);
 172
 173         if (is_a3xx(batch->ctx->screen))
 174                 util_dynarray_fini(&batch->rbrc_patches);
 175
 176         while (batch->samples.size > 0) {
 177                 struct fd_hw_sample *samp =
 178                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 179                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 180         }
 181         util_dynarray_fini(&batch->samples);
 182
 183         if (batch->ctx->screen->reorder)
 184                 util_queue_fence_destroy(&batch->flush_fence);
 185 }
 186
 187 static void
 188 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 189 {
 190         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 191         struct fd_batch *dep;
 192
 193         foreach_batch(dep, cache, batch->dependents_mask) {
 194                 if (flush)
 195                         fd_batch_flush(dep, false, false);
 196                 fd_batch_reference(&dep, NULL);
 197         }
 198
 199         batch->dependents_mask = 0;
 200 }
 201
 202 static void
 203 batch_reset_resources_locked(struct fd_batch *batch)
 204 {
 205         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 206
 207         set_foreach(batch->resources, entry) {
 208                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 209                 _mesa_set_remove(batch->resources, entry);
 210                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 211                 rsc->batch_mask &= ~(1 << batch->idx);
 212                 if (rsc->write_batch == batch)
 213                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 214         }
 215 }
 216
 217 static void
 218 batch_reset_resources(struct fd_batch *batch)
 219 {
 220         mtx_lock(&batch->ctx->screen->lock);
 221         batch_reset_resources_locked(batch);
 222         mtx_unlock(&batch->ctx->screen->lock);
 223 }
 224
 225 static void
 226 batch_reset(struct fd_batch *batch)
 227 {
 228         DBG("%p", batch);
 229
 230         fd_batch_sync(batch);
 231
 232         batch_flush_reset_dependencies(batch, false);
 233         batch_reset_resources(batch);
 234
 235         batch_fini(batch);
 236         batch_init(batch);
 237 }
 238
 239 void
 240 fd_batch_reset(struct fd_batch *batch)
 241 {
 242         if (batch->needs_flush)
 243                 batch_reset(batch);
 244 }
 245
 246 void
 247 __fd_batch_destroy(struct fd_batch *batch)
 248 {
 249         struct fd_context *ctx = batch->ctx;
 250
 251         DBG("%p", batch);
 252
 253         fd_context_assert_locked(batch->ctx);
 254
 255         fd_bc_invalidate_batch(batch, true);
 256
 257         batch_reset_resources_locked(batch);
 258         debug_assert(batch->resources->entries == 0);
 259         _mesa_set_destroy(batch->resources, NULL);
 260
 261         fd_context_unlock(ctx);
 262         batch_flush_reset_dependencies(batch, false);
 263         debug_assert(batch->dependents_mask == 0);
 264
 265         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 266         batch_fini(batch);
 267         free(batch);
 268         fd_context_lock(ctx);
 269 }
 270
 271 void
 272 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 273 {
 274         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 275 }
 276
 277 void
 278 fd_batch_sync(struct fd_batch *batch)
 279 {
 280         if (!batch->ctx->screen->reorder)
 281                 return;
 282         util_queue_fence_wait(&batch->flush_fence);
 283 }
 284
 285 static void
 286 batch_flush_func(void *job, int id)
 287 {
 288         struct fd_batch *batch = job;
 289
 290         DBG("%p", batch);
 291
 292         fd_gmem_render_tiles(batch);
 293         batch_reset_resources(batch);
 294 }
 295
 296 static void
 297 batch_cleanup_func(void *job, int id)
 298 {
 299         struct fd_batch *batch = job;
 300         fd_batch_reference(&batch, NULL);
 301 }
 302
 303 static void
 304 batch_flush(struct fd_batch *batch, bool force)
 305 {
 306         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 307
 308         if (batch->flushed)
 309                 return;
 310
 311         batch->needs_flush = false;
 312
 313         /* close out the draw cmds by making sure any active queries are
 314          * paused:
 315          */
 316         fd_batch_set_stage(batch, FD_STAGE_NULL);
 317
 318         batch_flush_reset_dependencies(batch, true);
 319
 320         batch->flushed = true;
 321
 322         if (batch->ctx->screen->reorder) {
 323                 struct fd_batch *tmp = NULL;
 324                 fd_batch_reference(&tmp, batch);
 325
 326                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 327                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 328
 329                 util_queue_add_job(&batch->ctx->flush_queue,
 330                                 batch, &batch->flush_fence,
 331                                 batch_flush_func, batch_cleanup_func);
 332         } else {
 333                 fd_gmem_render_tiles(batch);
 334                 batch_reset_resources(batch);
 335         }
 336
 337         debug_assert(batch->reference.count > 0);
 338
 339         mtx_lock(&batch->ctx->screen->lock);
 340         fd_bc_invalidate_batch(batch, false);
 341         mtx_unlock(&batch->ctx->screen->lock);
 342 }
 343
 344 /* NOTE: could drop the last ref to batch
 345  *
 346  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 347  *   to kernel before this returns, as opposed to just being queued to be
 348  *   flushed
 349  * @force: force a flush even if no rendering, mostly useful if you need
 350  *   a fence to sync on
 351  */
 352 void
 353 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 354 {
 355         struct fd_batch *tmp = NULL;
 356         bool newbatch = false;
 357
 358         /* NOTE: we need to hold an extra ref across the body of flush,
 359          * since the last ref to this batch could be dropped when cleaning
 360          * up used_resources
 361          */
 362         fd_batch_reference(&tmp, batch);
 363
 364         if (batch == batch->ctx->batch) {
 365                 batch->ctx->batch = NULL;
 366                 newbatch = true;
 367         }
 368
 369         batch_flush(tmp, force);
 370
 371         if (newbatch) {
 372                 struct fd_context *ctx = batch->ctx;
 373                 struct fd_batch *new_batch;
 374
 375                 if (ctx->screen->reorder) {
 376                         /* defer allocating new batch until one is needed for rendering
 377                          * to avoid unused batches for apps that create many contexts
 378                          */
 379                         new_batch = NULL;
 380                 } else {
 381                         new_batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, false);
 382                         util_copy_framebuffer_state(&new_batch->framebuffer, &batch->framebuffer);
 383                 }
 384
 385                 fd_batch_reference(&batch, NULL);
 386                 ctx->batch = new_batch;
 387                 fd_context_all_dirty(ctx);
 388         }
 389
 390         if (sync)
 391                 fd_batch_sync(tmp);
 392
 393         fd_batch_reference(&tmp, NULL);
 394 }
 395
 396 /* does 'batch' depend directly or indirectly on 'other' ? */
 397 static bool
 398 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 399 {
 400         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 401         struct fd_batch *dep;
 402
 403         if (batch->dependents_mask & (1 << other->idx))
 404                 return true;
 405
 406         foreach_batch(dep, cache, batch->dependents_mask)
 407                 if (batch_depends_on(batch, dep))
 408                         return true;
 409
 410         return false;
 411 }
 412
 413 void
 414 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 415 {
 416         if (batch->dependents_mask & (1 << dep->idx))
 417                 return;
 418
 419         /* a loop should not be possible */
 420         debug_assert(!batch_depends_on(dep, batch));
 421
 422         struct fd_batch *other = NULL;
 423         fd_batch_reference_locked(&other, dep);
 424         batch->dependents_mask |= (1 << dep->idx);
 425         DBG("%p: added dependency on %p", batch, dep);
 426 }
 427
 428 static void
 429 flush_write_batch(struct fd_resource *rsc)
 430 {
 431         struct fd_batch *b = NULL;
 432         fd_batch_reference(&b, rsc->write_batch);
 433
 434         mtx_unlock(&b->ctx->screen->lock);
 435         fd_batch_flush(b, true, false);
 436         mtx_lock(&b->ctx->screen->lock);
 437
 438         fd_bc_invalidate_batch(b, false);
 439         fd_batch_reference_locked(&b, NULL);
 440 }
 441
 442 void
 443 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 444 {
 445         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 446
 447         if (rsc->stencil)
 448                 fd_batch_resource_used(batch, rsc->stencil, write);
 449
 450         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 451
 452         if (write)
 453                 rsc->valid = true;
 454
 455         /* note, invalidate write batch, to avoid further writes to rsc
 456          * resulting in a write-after-read hazard.
 457          */
 458
 459         if (write) {
 460                 /* if we are pending read or write by any other batch: */
 461                 if (rsc->batch_mask & ~(1 << batch->idx)) {
 462                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 463                         struct fd_batch *dep;
 464
 465                         if (rsc->write_batch && rsc->write_batch != batch)
 466                                 flush_write_batch(rsc);
 467
 468                         foreach_batch(dep, cache, rsc->batch_mask) {
 469                                 struct fd_batch *b = NULL;
 470                                 if (dep == batch)
 471                                         continue;
 472                                 /* note that batch_add_dep could flush and unref dep, so
 473                                  * we need to hold a reference to keep it live for the
 474                                  * fd_bc_invalidate_batch()
 475                                  */
 476                                 fd_batch_reference(&b, dep);
 477                                 fd_batch_add_dep(batch, b);
 478                                 fd_bc_invalidate_batch(b, false);
 479                                 fd_batch_reference_locked(&b, NULL);
 480                         }
 481                 }
 482                 fd_batch_reference_locked(&rsc->write_batch, batch);
 483         } else {
 484                 /* If reading a resource pending a write, go ahead and flush the
 485                  * writer.  This avoids situations where we end up having to
 486                  * flush the current batch in _resource_used()
 487                  */
 488                 if (rsc->write_batch && rsc->write_batch != batch)
 489                         flush_write_batch(rsc);
 490         }
 491
 492         if (rsc->batch_mask & (1 << batch->idx)) {
 493                 debug_assert(_mesa_set_search(batch->resources, rsc));
 494                 return;
 495         }
 496
 497         debug_assert(!_mesa_set_search(batch->resources, rsc));
 498
 499         _mesa_set_add(batch->resources, rsc);
 500         rsc->batch_mask |= (1 << batch->idx);
 501 }
 502
 503 void
 504 fd_batch_check_size(struct fd_batch *batch)
 505 {
 506         debug_assert(!batch->flushed);
 507
 508         if (unlikely(fd_mesa_debug & FD_DBG_FLUSH)) {
 509                 fd_batch_flush(batch, true, false);
 510                 return;
 511         }
 512
 513         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 514                 return;
 515
 516         struct fd_ringbuffer *ring = batch->draw;
 517         if ((ring->cur - ring->start) > (ring->size/4 - 0x1000))
 518                 fd_batch_flush(batch, true, false);
 519 }
 520
 521 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 522  * been one since last draw:
 523  */
 524 void
 525 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 526 {
 527         if (batch->needs_wfi) {
 528                 if (batch->ctx->screen->gpu_id >= 500)
 529                         OUT_WFI5(ring);
 530                 else
 531                         OUT_WFI(ring);
 532                 batch->needs_wfi = false;
 533         }
 534 }