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 struct fd_ringbuffer *
  39 alloc_ring(struct fd_batch *batch, unsigned sz, enum fd_ringbuffer_flags flags)
  40 {
  41         struct fd_context *ctx = batch->ctx;
  42
  43         /* if kernel is too old to support unlimited # of cmd buffers, we
  44          * have no option but to allocate large worst-case sizes so that
  45          * we don't need to grow the ringbuffer.  Performance is likely to
  46          * suffer, but there is no good alternative.
  47          *
  48          * Otherwise if supported, allocate a growable ring with initial
  49          * size of zero.
  50          */
  51         if ((fd_device_version(ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS) &&
  52                         !(fd_mesa_debug & FD_DBG_NOGROW)){
  53                 flags |= FD_RINGBUFFER_GROWABLE;
  54                 sz = 0;
  55         }
  56
  57         return fd_submit_new_ringbuffer(batch->submit, sz, flags);
  58 }
  59
  60 static void
  61 batch_init(struct fd_batch *batch)
  62 {
  63         struct fd_context *ctx = batch->ctx;
  64
  65         batch->submit = fd_submit_new(ctx->pipe);
  66         if (batch->nondraw) {
  67                 batch->gmem = alloc_ring(batch, 0x1000, FD_RINGBUFFER_PRIMARY);
  68                 batch->draw = alloc_ring(batch, 0x100000, 0);
  69         } else {
  70                 batch->gmem = alloc_ring(batch, 0x100000, FD_RINGBUFFER_PRIMARY);
  71                 batch->draw = alloc_ring(batch, 0x100000, 0);
  72
  73                 /* a6xx+ re-uses draw rb for both draw and binning pass: */
  74                 if (ctx->screen->gpu_id < 600) {
  75                         batch->binning = alloc_ring(batch, 0x100000, 0);
  76                 }
  77         }
  78
  79         batch->in_fence_fd = -1;
  80         batch->fence = fd_fence_create(batch);
  81
  82         batch->cleared = 0;
  83         batch->fast_cleared = 0;
  84         batch->invalidated = 0;
  85         batch->restore = batch->resolve = 0;
  86         batch->needs_flush = false;
  87         batch->flushed = false;
  88         batch->gmem_reason = 0;
  89         batch->num_draws = 0;
  90         batch->num_vertices = 0;
  91         batch->num_bins_per_pipe = 0;
  92         batch->prim_strm_bits = 0;
  93         batch->draw_strm_bits = 0;
  94         batch->stage = FD_STAGE_NULL;
  95
  96         fd_reset_wfi(batch);
  97
  98         util_dynarray_init(&batch->draw_patches, NULL);
  99         util_dynarray_init(&batch->fb_read_patches, NULL);
 100
 101         if (is_a2xx(ctx->screen)) {
 102                 util_dynarray_init(&batch->shader_patches, NULL);
 103                 util_dynarray_init(&batch->gmem_patches, NULL);
 104         }
 105
 106         if (is_a3xx(ctx->screen))
 107                 util_dynarray_init(&batch->rbrc_patches, NULL);
 108
 109         assert(batch->resources->entries == 0);
 110
 111         util_dynarray_init(&batch->samples, NULL);
 112
 113         list_inithead(&batch->log_chunks);
 114 }
 115
 116 struct fd_batch *
 117 fd_batch_create(struct fd_context *ctx, bool nondraw)
 118 {
 119         struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
 120
 121         if (!batch)
 122                 return NULL;
 123
 124         DBG("%p", batch);
 125
 126         pipe_reference_init(&batch->reference, 1);
 127         batch->ctx = ctx;
 128         batch->nondraw = nondraw;
 129
 130         batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 131                         _mesa_key_pointer_equal);
 132
 133         batch_init(batch);
 134
 135         fd_screen_assert_locked(ctx->screen);
 136         if (BATCH_DEBUG) {
 137                 _mesa_set_add(ctx->screen->live_batches, batch);
 138         }
 139
 140         return batch;
 141 }
 142
 143 static void
 144 batch_fini(struct fd_batch *batch)
 145 {
 146         DBG("%p", batch);
 147
 148         pipe_resource_reference(&batch->query_buf, NULL);
 149
 150         if (batch->in_fence_fd != -1)
 151                 close(batch->in_fence_fd);
 152
 153         /* in case batch wasn't flushed but fence was created: */
 154         fd_fence_populate(batch->fence, 0, -1);
 155
 156         fd_fence_ref(&batch->fence, NULL);
 157
 158         fd_ringbuffer_del(batch->draw);
 159         fd_ringbuffer_del(batch->gmem);
 160
 161         if (batch->binning) {
 162                 fd_ringbuffer_del(batch->binning);
 163                 batch->binning = NULL;
 164         }
 165
 166         if (batch->prologue) {
 167                 fd_ringbuffer_del(batch->prologue);
 168                 batch->prologue = NULL;
 169         }
 170
 171         if (batch->epilogue) {
 172                 fd_ringbuffer_del(batch->epilogue);
 173                 batch->epilogue = NULL;
 174         }
 175
 176         if (batch->tile_setup) {
 177                 fd_ringbuffer_del(batch->tile_setup);
 178                 batch->tile_setup = NULL;
 179         }
 180
 181         if (batch->tile_fini) {
 182                 fd_ringbuffer_del(batch->tile_fini);
 183                 batch->tile_fini = NULL;
 184         }
 185
 186         if (batch->tessellation) {
 187                 fd_bo_del(batch->tessfactor_bo);
 188                 fd_bo_del(batch->tessparam_bo);
 189                 fd_ringbuffer_del(batch->tess_addrs_constobj);
 190         }
 191
 192         fd_submit_del(batch->submit);
 193
 194         util_dynarray_fini(&batch->draw_patches);
 195         util_dynarray_fini(&batch->fb_read_patches);
 196
 197         if (is_a2xx(batch->ctx->screen)) {
 198                 util_dynarray_fini(&batch->shader_patches);
 199                 util_dynarray_fini(&batch->gmem_patches);
 200         }
 201
 202         if (is_a3xx(batch->ctx->screen))
 203                 util_dynarray_fini(&batch->rbrc_patches);
 204
 205         while (batch->samples.size > 0) {
 206                 struct fd_hw_sample *samp =
 207                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 208                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 209         }
 210         util_dynarray_fini(&batch->samples);
 211
 212         assert(list_is_empty(&batch->log_chunks));
 213 }
 214
 215 static void
 216 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 217 {
 218         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 219         struct fd_batch *dep;
 220
 221         foreach_batch(dep, cache, batch->dependents_mask) {
 222                 if (flush)
 223                         fd_batch_flush(dep);
 224                 fd_batch_reference(&dep, NULL);
 225         }
 226
 227         batch->dependents_mask = 0;
 228 }
 229
 230 static void
 231 batch_reset_resources_locked(struct fd_batch *batch)
 232 {
 233         fd_screen_assert_locked(batch->ctx->screen);
 234
 235         set_foreach(batch->resources, entry) {
 236                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 237                 _mesa_set_remove(batch->resources, entry);
 238                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 239                 rsc->batch_mask &= ~(1 << batch->idx);
 240                 if (rsc->write_batch == batch)
 241                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 242         }
 243 }
 244
 245 static void
 246 batch_reset_resources(struct fd_batch *batch)
 247 {
 248         fd_screen_lock(batch->ctx->screen);
 249         batch_reset_resources_locked(batch);
 250         fd_screen_unlock(batch->ctx->screen);
 251 }
 252
 253 static void
 254 batch_reset(struct fd_batch *batch)
 255 {
 256         DBG("%p", batch);
 257
 258         batch_flush_reset_dependencies(batch, false);
 259         batch_reset_resources(batch);
 260
 261         batch_fini(batch);
 262         batch_init(batch);
 263 }
 264
 265 void
 266 fd_batch_reset(struct fd_batch *batch)
 267 {
 268         if (batch->needs_flush)
 269                 batch_reset(batch);
 270 }
 271
 272 void
 273 __fd_batch_destroy(struct fd_batch *batch)
 274 {
 275         struct fd_context *ctx = batch->ctx;
 276
 277         DBG("%p", batch);
 278
 279         fd_context_assert_locked(batch->ctx);
 280
 281         if (BATCH_DEBUG) {
 282                 _mesa_set_remove_key(ctx->screen->live_batches, batch);
 283         }
 284
 285         fd_bc_invalidate_batch(batch, true);
 286
 287         batch_reset_resources_locked(batch);
 288         debug_assert(batch->resources->entries == 0);
 289         _mesa_set_destroy(batch->resources, NULL);
 290
 291         fd_context_unlock(ctx);
 292         batch_flush_reset_dependencies(batch, false);
 293         debug_assert(batch->dependents_mask == 0);
 294
 295         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 296         batch_fini(batch);
 297         free(batch);
 298         fd_context_lock(ctx);
 299 }
 300
 301 void
 302 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 303 {
 304         sprintf(buf, "fd_batch<%u>", batch->seqno);
 305 }
 306
 307 static void
 308 batch_flush(struct fd_batch *batch)
 309 {
 310         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 311
 312         if (batch->flushed)
 313                 return;
 314
 315         batch->needs_flush = false;
 316
 317         /* close out the draw cmds by making sure any active queries are
 318          * paused:
 319          */
 320         fd_batch_set_stage(batch, FD_STAGE_NULL);
 321
 322         batch_flush_reset_dependencies(batch, true);
 323
 324         batch->flushed = true;
 325
 326         fd_fence_ref(&batch->ctx->last_fence, batch->fence);
 327
 328         fd_gmem_render_tiles(batch);
 329         batch_reset_resources(batch);
 330
 331         debug_assert(batch->reference.count > 0);
 332
 333         fd_screen_lock(batch->ctx->screen);
 334         fd_bc_invalidate_batch(batch, false);
 335         fd_screen_unlock(batch->ctx->screen);
 336 }
 337
 338 /* Get per-batch prologue */
 339 struct fd_ringbuffer *
 340 fd_batch_get_prologue(struct fd_batch *batch)
 341 {
 342         if (!batch->prologue)
 343                 batch->prologue = alloc_ring(batch, 0x1000, 0);
 344         return batch->prologue;
 345 }
 346
 347 /* NOTE: could drop the last ref to batch
 348  *
 349  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 350  *   to kernel before this returns, as opposed to just being queued to be
 351  *   flushed
 352  * @force: force a flush even if no rendering, mostly useful if you need
 353  *   a fence to sync on
 354  */
 355 void
 356 fd_batch_flush(struct fd_batch *batch)
 357 {
 358         struct fd_batch *tmp = NULL;
 359
 360         /* NOTE: we need to hold an extra ref across the body of flush,
 361          * since the last ref to this batch could be dropped when cleaning
 362          * up used_resources
 363          */
 364         fd_batch_reference(&tmp, batch);
 365
 366         batch_flush(tmp);
 367
 368         if (batch == batch->ctx->batch) {
 369                 fd_batch_reference(&batch->ctx->batch, NULL);
 370         }
 371
 372         fd_batch_reference(&tmp, NULL);
 373 }
 374
 375 /* find a batches dependents mask, including recursive dependencies: */
 376 static uint32_t
 377 recursive_dependents_mask(struct fd_batch *batch)
 378 {
 379         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 380         struct fd_batch *dep;
 381         uint32_t dependents_mask = batch->dependents_mask;
 382
 383         foreach_batch(dep, cache, batch->dependents_mask)
 384                 dependents_mask |= recursive_dependents_mask(dep);
 385
 386         return dependents_mask;
 387 }
 388
 389 void
 390 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 391 {
 392         fd_screen_assert_locked(batch->ctx->screen);
 393
 394         if (batch->dependents_mask & (1 << dep->idx))
 395                 return;
 396
 397         /* a loop should not be possible */
 398         debug_assert(!((1 << batch->idx) & recursive_dependents_mask(dep)));
 399
 400         struct fd_batch *other = NULL;
 401         fd_batch_reference_locked(&other, dep);
 402         batch->dependents_mask |= (1 << dep->idx);
 403         DBG("%p: added dependency on %p", batch, dep);
 404 }
 405
 406 static void
 407 flush_write_batch(struct fd_resource *rsc)
 408 {
 409         struct fd_batch *b = NULL;
 410         fd_batch_reference_locked(&b, rsc->write_batch);
 411
 412         fd_screen_unlock(b->ctx->screen);
 413         fd_batch_flush(b);
 414         fd_screen_lock(b->ctx->screen);
 415
 416         fd_bc_invalidate_batch(b, false);
 417         fd_batch_reference_locked(&b, NULL);
 418 }
 419
 420 static void
 421 fd_batch_add_resource(struct fd_batch *batch, struct fd_resource *rsc)
 422 {
 423
 424         if (likely(fd_batch_references_resource(batch, rsc))) {
 425                 debug_assert(_mesa_set_search(batch->resources, rsc));
 426                 return;
 427         }
 428
 429         debug_assert(!_mesa_set_search(batch->resources, rsc));
 430
 431         _mesa_set_add(batch->resources, rsc);
 432         rsc->batch_mask |= (1 << batch->idx);
 433 }
 434
 435 void
 436 fd_batch_resource_write(struct fd_batch *batch, struct fd_resource *rsc)
 437 {
 438         fd_screen_assert_locked(batch->ctx->screen);
 439
 440         fd_batch_write_prep(batch, rsc);
 441
 442         if (rsc->stencil)
 443                 fd_batch_resource_write(batch, rsc->stencil);
 444
 445         DBG("%p: write %p", batch, rsc);
 446
 447         rsc->valid = true;
 448
 449         /* note, invalidate write batch, to avoid further writes to rsc
 450          * resulting in a write-after-read hazard.
 451          */
 452         /* if we are pending read or write by any other batch: */
 453         if (unlikely(rsc->batch_mask & ~(1 << batch->idx))) {
 454                 struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 455                 struct fd_batch *dep;
 456
 457                 if (rsc->write_batch && rsc->write_batch != batch)
 458                         flush_write_batch(rsc);
 459
 460                 foreach_batch(dep, cache, rsc->batch_mask) {
 461                         struct fd_batch *b = NULL;
 462                         if (dep == batch)
 463                                 continue;
 464                         /* note that batch_add_dep could flush and unref dep, so
 465                          * we need to hold a reference to keep it live for the
 466                          * fd_bc_invalidate_batch()
 467                          */
 468                         fd_batch_reference(&b, dep);
 469                         fd_batch_add_dep(batch, b);
 470                         fd_bc_invalidate_batch(b, false);
 471                         fd_batch_reference_locked(&b, NULL);
 472                 }
 473         }
 474         fd_batch_reference_locked(&rsc->write_batch, batch);
 475
 476         fd_batch_add_resource(batch, rsc);
 477 }
 478
 479 void
 480 fd_batch_resource_read_slowpath(struct fd_batch *batch, struct fd_resource *rsc)
 481 {
 482         fd_screen_assert_locked(batch->ctx->screen);
 483
 484         if (rsc->stencil)
 485                 fd_batch_resource_read(batch, rsc->stencil);
 486
 487         DBG("%p: read %p", batch, rsc);
 488
 489         /* If reading a resource pending a write, go ahead and flush the
 490          * writer.  This avoids situations where we end up having to
 491          * flush the current batch in _resource_used()
 492          */
 493         if (unlikely(rsc->write_batch && rsc->write_batch != batch))
 494                 flush_write_batch(rsc);
 495
 496         fd_batch_add_resource(batch, rsc);
 497 }
 498
 499 void
 500 fd_batch_check_size(struct fd_batch *batch)
 501 {
 502         debug_assert(!batch->flushed);
 503
 504         if (unlikely(fd_mesa_debug & FD_DBG_FLUSH)) {
 505                 fd_batch_flush(batch);
 506                 return;
 507         }
 508
 509         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 510                 return;
 511
 512         struct fd_ringbuffer *ring = batch->draw;
 513         if ((ring->cur - ring->start) > (ring->size/4 - 0x1000))
 514                 fd_batch_flush(batch);
 515 }
 516
 517 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 518  * been one since last draw:
 519  */
 520 void
 521 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 522 {
 523         if (batch->needs_wfi) {
 524                 if (batch->ctx->screen->gpu_id >= 500)
 525                         OUT_WFI5(ring);
 526                 else
 527                         OUT_WFI(ring);
 528                 batch->needs_wfi = false;
 529         }
 530 }