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         DBG("%p", batch);
 227
 228         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 229
 230         mtx_lock(&batch->ctx->screen->lock);
 231         fd_bc_invalidate_batch(batch, true);
 232         mtx_unlock(&batch->ctx->screen->lock);
 233
 234         batch_fini(batch);
 235
 236         batch_reset_resources(batch);
 237         debug_assert(batch->resources->entries == 0);
 238         _mesa_set_destroy(batch->resources, NULL);
 239
 240         batch_flush_reset_dependencies(batch, false);
 241         debug_assert(batch->dependents_mask == 0);
 242
 243         free(batch);
 244 }
 245
 246 void
 247 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 248 {
 249         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 250 }
 251
 252 void
 253 fd_batch_sync(struct fd_batch *batch)
 254 {
 255         if (!batch->ctx->screen->reorder)
 256                 return;
 257         util_queue_fence_wait(&batch->flush_fence);
 258 }
 259
 260 static void
 261 batch_flush_func(void *job, int id)
 262 {
 263         struct fd_batch *batch = job;
 264
 265         DBG("%p", batch);
 266
 267         fd_gmem_render_tiles(batch);
 268         batch_reset_resources(batch);
 269 }
 270
 271 static void
 272 batch_cleanup_func(void *job, int id)
 273 {
 274         struct fd_batch *batch = job;
 275         fd_batch_reference(&batch, NULL);
 276 }
 277
 278 static void
 279 batch_flush(struct fd_batch *batch, bool force)
 280 {
 281         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 282
 283         if (batch->flushed)
 284                 return;
 285
 286         batch->needs_flush = false;
 287
 288         /* close out the draw cmds by making sure any active queries are
 289          * paused:
 290          */
 291         fd_batch_set_stage(batch, FD_STAGE_NULL);
 292
 293         fd_context_all_dirty(batch->ctx);
 294         batch_flush_reset_dependencies(batch, true);
 295
 296         batch->flushed = true;
 297
 298         if (batch->ctx->screen->reorder) {
 299                 struct fd_batch *tmp = NULL;
 300                 fd_batch_reference(&tmp, batch);
 301
 302                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 303                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 304
 305                 util_queue_add_job(&batch->ctx->flush_queue,
 306                                 batch, &batch->flush_fence,
 307                                 batch_flush_func, batch_cleanup_func);
 308         } else {
 309                 fd_gmem_render_tiles(batch);
 310                 batch_reset_resources(batch);
 311         }
 312
 313         debug_assert(batch->reference.count > 0);
 314
 315         mtx_lock(&batch->ctx->screen->lock);
 316         fd_bc_invalidate_batch(batch, false);
 317         mtx_unlock(&batch->ctx->screen->lock);
 318 }
 319
 320 /* NOTE: could drop the last ref to batch
 321  *
 322  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 323  *   to kernel before this returns, as opposed to just being queued to be
 324  *   flushed
 325  * @force: force a flush even if no rendering, mostly useful if you need
 326  *   a fence to sync on
 327  */
 328 void
 329 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 330 {
 331         struct fd_batch *tmp = NULL;
 332         bool newbatch = false;
 333
 334         /* NOTE: we need to hold an extra ref across the body of flush,
 335          * since the last ref to this batch could be dropped when cleaning
 336          * up used_resources
 337          */
 338         fd_batch_reference(&tmp, batch);
 339
 340         if (batch == batch->ctx->batch) {
 341                 batch->ctx->batch = NULL;
 342                 newbatch = true;
 343         }
 344
 345         batch_flush(tmp, force);
 346
 347         if (newbatch) {
 348                 struct fd_context *ctx = batch->ctx;
 349                 struct fd_batch *new_batch =
 350                         fd_batch_from_fb(&ctx->screen->batch_cache, ctx, &batch->framebuffer);
 351
 352                 util_copy_framebuffer_state(&new_batch->framebuffer, &batch->framebuffer);
 353
 354                 fd_batch_reference(&batch, NULL);
 355                 ctx->batch = new_batch;
 356         }
 357
 358         if (sync)
 359                 fd_batch_sync(tmp);
 360
 361         fd_batch_reference(&tmp, NULL);
 362 }
 363
 364 /* does 'batch' depend directly or indirectly on 'other' ? */
 365 static bool
 366 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 367 {
 368         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 369         struct fd_batch *dep;
 370
 371         if (batch->dependents_mask & (1 << other->idx))
 372                 return true;
 373
 374         foreach_batch(dep, cache, batch->dependents_mask)
 375                 if (batch_depends_on(batch, dep))
 376                         return true;
 377
 378         return false;
 379 }
 380
 381 void
 382 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 383 {
 384         if (batch->dependents_mask & (1 << dep->idx))
 385                 return;
 386
 387         /* a loop should not be possible */
 388         debug_assert(!batch_depends_on(dep, batch));
 389
 390         struct fd_batch *other = NULL;
 391         fd_batch_reference_locked(&other, dep);
 392         batch->dependents_mask |= (1 << dep->idx);
 393         DBG("%p: added dependency on %p", batch, dep);
 394 }
 395
 396 static void
 397 flush_write_batch(struct fd_resource *rsc)
 398 {
 399         struct fd_batch *b = NULL;
 400         fd_batch_reference(&b, rsc->write_batch);
 401
 402         mtx_unlock(&b->ctx->screen->lock);
 403         fd_batch_flush(b, true, false);
 404         mtx_lock(&b->ctx->screen->lock);
 405
 406         fd_bc_invalidate_batch(b, false);
 407         fd_batch_reference_locked(&b, NULL);
 408 }
 409
 410 void
 411 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 412 {
 413         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 414
 415         if (rsc->stencil)
 416                 fd_batch_resource_used(batch, rsc->stencil, write);
 417
 418         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 419
 420         if (write)
 421                 rsc->valid = true;
 422
 423         /* note, invalidate write batch, to avoid further writes to rsc
 424          * resulting in a write-after-read hazard.
 425          */
 426
 427         if (write) {
 428                 /* if we are pending read or write by any other batch: */
 429                 if (rsc->batch_mask & ~(1 << batch->idx)) {
 430                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 431                         struct fd_batch *dep;
 432
 433                         if (rsc->write_batch && rsc->write_batch != batch)
 434                                 flush_write_batch(rsc);
 435
 436                         foreach_batch(dep, cache, rsc->batch_mask) {
 437                                 struct fd_batch *b = NULL;
 438                                 if (dep == batch)
 439                                         continue;
 440                                 /* note that batch_add_dep could flush and unref dep, so
 441                                  * we need to hold a reference to keep it live for the
 442                                  * fd_bc_invalidate_batch()
 443                                  */
 444                                 fd_batch_reference(&b, dep);
 445                                 fd_batch_add_dep(batch, b);
 446                                 fd_bc_invalidate_batch(b, false);
 447                                 fd_batch_reference_locked(&b, NULL);
 448                         }
 449                 }
 450                 fd_batch_reference_locked(&rsc->write_batch, batch);
 451         } else {
 452                 /* If reading a resource pending a write, go ahead and flush the
 453                  * writer.  This avoids situations where we end up having to
 454                  * flush the current batch in _resource_used()
 455                  */
 456                 if (rsc->write_batch && rsc->write_batch != batch)
 457                         flush_write_batch(rsc);
 458         }
 459
 460         if (rsc->batch_mask & (1 << batch->idx))
 461                 return;
 462
 463         debug_assert(!_mesa_set_search(batch->resources, rsc));
 464
 465         _mesa_set_add(batch->resources, rsc);
 466         rsc->batch_mask |= (1 << batch->idx);
 467 }
 468
 469 void
 470 fd_batch_check_size(struct fd_batch *batch)
 471 {
 472         debug_assert(!batch->flushed);
 473
 474         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 475                 return;
 476
 477         struct fd_ringbuffer *ring = batch->draw;
 478         if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
 479                         (fd_mesa_debug & FD_DBG_FLUSH))
 480                 fd_batch_flush(batch, true, false);
 481 }
 482
 483 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 484  * been one since last draw:
 485  */
 486 void
 487 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 488 {
 489         if (batch->needs_wfi) {
 490                 if (batch->ctx->screen->gpu_id >= 500)
 491                         OUT_WFI5(ring);
 492                 else
 493                         OUT_WFI(ring);
 494                 batch->needs_wfi = false;
 495         }
 496 }