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