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->gmem    = fd_ringbuffer_new(ctx->pipe, size);
  60
  61                 fd_ringbuffer_set_parent(batch->gmem, NULL);
  62                 fd_ringbuffer_set_parent(batch->draw, batch->gmem);
  63
  64                 if (ctx->screen->gpu_id < 600) {
  65                         batch->binning = fd_ringbuffer_new(ctx->pipe, size);
  66                         fd_ringbuffer_set_parent(batch->binning, batch->gmem);
  67                 }
  68         } else {
  69                 fd_ringbuffer_set_parent(batch->draw, NULL);
  70         }
  71
  72         batch->in_fence_fd = -1;
  73         batch->fence = fd_fence_create(batch);
  74
  75         batch->cleared = 0;
  76         batch->invalidated = 0;
  77         batch->restore = batch->resolve = 0;
  78         batch->needs_flush = false;
  79         batch->flushed = false;
  80         batch->gmem_reason = 0;
  81         batch->num_draws = 0;
  82         batch->stage = FD_STAGE_NULL;
  83
  84         fd_reset_wfi(batch);
  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                 if (batch->binning)
 138                         fd_ringbuffer_del(batch->binning);
 139                 fd_ringbuffer_del(batch->gmem);
 140         } else {
 141                 debug_assert(!batch->binning);
 142                 debug_assert(!batch->gmem);
 143         }
 144         if (batch->lrz_clear) {
 145                 fd_ringbuffer_del(batch->lrz_clear);
 146                 batch->lrz_clear = NULL;
 147         }
 148
 149         util_dynarray_fini(&batch->draw_patches);
 150
 151         if (is_a3xx(batch->ctx->screen))
 152                 util_dynarray_fini(&batch->rbrc_patches);
 153
 154         util_dynarray_fini(&batch->gmem_patches);
 155
 156         while (batch->samples.size > 0) {
 157                 struct fd_hw_sample *samp =
 158                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 159                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 160         }
 161         util_dynarray_fini(&batch->samples);
 162
 163         if (batch->ctx->screen->reorder)
 164                 util_queue_fence_destroy(&batch->flush_fence);
 165 }
 166
 167 static void
 168 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 169 {
 170         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 171         struct fd_batch *dep;
 172
 173         foreach_batch(dep, cache, batch->dependents_mask) {
 174                 if (flush)
 175                         fd_batch_flush(dep, false, false);
 176                 fd_batch_reference(&dep, NULL);
 177         }
 178
 179         batch->dependents_mask = 0;
 180 }
 181
 182 static void
 183 batch_reset_resources_locked(struct fd_batch *batch)
 184 {
 185         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 186
 187         set_foreach(batch->resources, entry) {
 188                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 189                 _mesa_set_remove(batch->resources, entry);
 190                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 191                 rsc->batch_mask &= ~(1 << batch->idx);
 192                 if (rsc->write_batch == batch)
 193                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 194         }
 195 }
 196
 197 static void
 198 batch_reset_resources(struct fd_batch *batch)
 199 {
 200         mtx_lock(&batch->ctx->screen->lock);
 201         batch_reset_resources_locked(batch);
 202         mtx_unlock(&batch->ctx->screen->lock);
 203 }
 204
 205 static void
 206 batch_reset(struct fd_batch *batch)
 207 {
 208         DBG("%p", batch);
 209
 210         fd_batch_sync(batch);
 211
 212         batch_flush_reset_dependencies(batch, false);
 213         batch_reset_resources(batch);
 214
 215         batch_fini(batch);
 216         batch_init(batch);
 217 }
 218
 219 void
 220 fd_batch_reset(struct fd_batch *batch)
 221 {
 222         if (batch->needs_flush)
 223                 batch_reset(batch);
 224 }
 225
 226 void
 227 __fd_batch_destroy(struct fd_batch *batch)
 228 {
 229         struct fd_context *ctx = batch->ctx;
 230
 231         DBG("%p", batch);
 232
 233         fd_context_assert_locked(batch->ctx);
 234
 235         fd_bc_invalidate_batch(batch, true);
 236
 237         batch_reset_resources_locked(batch);
 238         debug_assert(batch->resources->entries == 0);
 239         _mesa_set_destroy(batch->resources, NULL);
 240
 241         fd_context_unlock(ctx);
 242         batch_flush_reset_dependencies(batch, false);
 243         debug_assert(batch->dependents_mask == 0);
 244
 245         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 246         batch_fini(batch);
 247         free(batch);
 248         fd_context_lock(ctx);
 249 }
 250
 251 void
 252 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 253 {
 254         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 255 }
 256
 257 void
 258 fd_batch_sync(struct fd_batch *batch)
 259 {
 260         if (!batch->ctx->screen->reorder)
 261                 return;
 262         util_queue_fence_wait(&batch->flush_fence);
 263 }
 264
 265 static void
 266 batch_flush_func(void *job, int id)
 267 {
 268         struct fd_batch *batch = job;
 269
 270         DBG("%p", batch);
 271
 272         fd_gmem_render_tiles(batch);
 273         batch_reset_resources(batch);
 274 }
 275
 276 static void
 277 batch_cleanup_func(void *job, int id)
 278 {
 279         struct fd_batch *batch = job;
 280         fd_batch_reference(&batch, NULL);
 281 }
 282
 283 static void
 284 batch_flush(struct fd_batch *batch, bool force)
 285 {
 286         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 287
 288         if (batch->flushed)
 289                 return;
 290
 291         batch->needs_flush = false;
 292
 293         /* close out the draw cmds by making sure any active queries are
 294          * paused:
 295          */
 296         fd_batch_set_stage(batch, FD_STAGE_NULL);
 297
 298         batch_flush_reset_dependencies(batch, true);
 299
 300         batch->flushed = true;
 301
 302         if (batch->ctx->screen->reorder) {
 303                 struct fd_batch *tmp = NULL;
 304                 fd_batch_reference(&tmp, batch);
 305
 306                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 307                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 308
 309                 util_queue_add_job(&batch->ctx->flush_queue,
 310                                 batch, &batch->flush_fence,
 311                                 batch_flush_func, batch_cleanup_func);
 312         } else {
 313                 fd_gmem_render_tiles(batch);
 314                 batch_reset_resources(batch);
 315         }
 316
 317         debug_assert(batch->reference.count > 0);
 318
 319         mtx_lock(&batch->ctx->screen->lock);
 320         fd_bc_invalidate_batch(batch, false);
 321         mtx_unlock(&batch->ctx->screen->lock);
 322 }
 323
 324 /* NOTE: could drop the last ref to batch
 325  *
 326  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 327  *   to kernel before this returns, as opposed to just being queued to be
 328  *   flushed
 329  * @force: force a flush even if no rendering, mostly useful if you need
 330  *   a fence to sync on
 331  */
 332 void
 333 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 334 {
 335         struct fd_batch *tmp = NULL;
 336         bool newbatch = false;
 337
 338         /* NOTE: we need to hold an extra ref across the body of flush,
 339          * since the last ref to this batch could be dropped when cleaning
 340          * up used_resources
 341          */
 342         fd_batch_reference(&tmp, batch);
 343
 344         if (batch == batch->ctx->batch) {
 345                 batch->ctx->batch = NULL;
 346                 newbatch = true;
 347         }
 348
 349         batch_flush(tmp, force);
 350
 351         if (newbatch) {
 352                 struct fd_context *ctx = batch->ctx;
 353                 struct fd_batch *new_batch;
 354
 355                 if (ctx->screen->reorder) {
 356                         /* defer allocating new batch until one is needed for rendering
 357                          * to avoid unused batches for apps that create many contexts
 358                          */
 359                         new_batch = NULL;
 360                 } else {
 361                         new_batch = fd_batch_create(ctx, false);
 362                         util_copy_framebuffer_state(&new_batch->framebuffer, &batch->framebuffer);
 363                 }
 364
 365                 fd_batch_reference(&batch, NULL);
 366                 ctx->batch = new_batch;
 367         }
 368
 369         if (sync)
 370                 fd_batch_sync(tmp);
 371
 372         fd_batch_reference(&tmp, NULL);
 373 }
 374
 375 /* does 'batch' depend directly or indirectly on 'other' ? */
 376 static bool
 377 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 378 {
 379         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 380         struct fd_batch *dep;
 381
 382         if (batch->dependents_mask & (1 << other->idx))
 383                 return true;
 384
 385         foreach_batch(dep, cache, batch->dependents_mask)
 386                 if (batch_depends_on(batch, dep))
 387                         return true;
 388
 389         return false;
 390 }
 391
 392 void
 393 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 394 {
 395         if (batch->dependents_mask & (1 << dep->idx))
 396                 return;
 397
 398         /* a loop should not be possible */
 399         debug_assert(!batch_depends_on(dep, batch));
 400
 401         struct fd_batch *other = NULL;
 402         fd_batch_reference_locked(&other, dep);
 403         batch->dependents_mask |= (1 << dep->idx);
 404         DBG("%p: added dependency on %p", batch, dep);
 405 }
 406
 407 static void
 408 flush_write_batch(struct fd_resource *rsc)
 409 {
 410         struct fd_batch *b = NULL;
 411         fd_batch_reference(&b, rsc->write_batch);
 412
 413         mtx_unlock(&b->ctx->screen->lock);
 414         fd_batch_flush(b, true, false);
 415         mtx_lock(&b->ctx->screen->lock);
 416
 417         fd_bc_invalidate_batch(b, false);
 418         fd_batch_reference_locked(&b, NULL);
 419 }
 420
 421 void
 422 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 423 {
 424         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 425
 426         if (rsc->stencil)
 427                 fd_batch_resource_used(batch, rsc->stencil, write);
 428
 429         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 430
 431         if (write)
 432                 rsc->valid = true;
 433
 434         /* note, invalidate write batch, to avoid further writes to rsc
 435          * resulting in a write-after-read hazard.
 436          */
 437
 438         if (write) {
 439                 /* if we are pending read or write by any other batch: */
 440                 if (rsc->batch_mask & ~(1 << batch->idx)) {
 441                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 442                         struct fd_batch *dep;
 443
 444                         if (rsc->write_batch && rsc->write_batch != batch)
 445                                 flush_write_batch(rsc);
 446
 447                         foreach_batch(dep, cache, rsc->batch_mask) {
 448                                 struct fd_batch *b = NULL;
 449                                 if (dep == batch)
 450                                         continue;
 451                                 /* note that batch_add_dep could flush and unref dep, so
 452                                  * we need to hold a reference to keep it live for the
 453                                  * fd_bc_invalidate_batch()
 454                                  */
 455                                 fd_batch_reference(&b, dep);
 456                                 fd_batch_add_dep(batch, b);
 457                                 fd_bc_invalidate_batch(b, false);
 458                                 fd_batch_reference_locked(&b, NULL);
 459                         }
 460                 }
 461                 fd_batch_reference_locked(&rsc->write_batch, batch);
 462         } else {
 463                 /* If reading a resource pending a write, go ahead and flush the
 464                  * writer.  This avoids situations where we end up having to
 465                  * flush the current batch in _resource_used()
 466                  */
 467                 if (rsc->write_batch && rsc->write_batch != batch)
 468                         flush_write_batch(rsc);
 469         }
 470
 471         if (rsc->batch_mask & (1 << batch->idx))
 472                 return;
 473
 474         debug_assert(!_mesa_set_search(batch->resources, rsc));
 475
 476         _mesa_set_add(batch->resources, rsc);
 477         rsc->batch_mask |= (1 << batch->idx);
 478 }
 479
 480 void
 481 fd_batch_check_size(struct fd_batch *batch)
 482 {
 483         debug_assert(!batch->flushed);
 484
 485         if (unlikely(fd_mesa_debug & FD_DBG_FLUSH)) {
 486                 fd_batch_flush(batch, true, false);
 487                 return;
 488         }
 489
 490         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 491                 return;
 492
 493         struct fd_ringbuffer *ring = batch->draw;
 494         if ((ring->cur - ring->start) > (ring->size/4 - 0x1000))
 495                 fd_batch_flush(batch, true, false);
 496 }
 497
 498 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 499  * been one since last draw:
 500  */
 501 void
 502 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 503 {
 504         if (batch->needs_wfi) {
 505                 if (batch->ctx->screen->gpu_id >= 500)
 506                         OUT_WFI5(ring);
 507                 else
 508                         OUT_WFI(ring);
 509                 batch->needs_wfi = false;
 510         }
 511 }