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