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