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->gmem_reason = 0;
  76         batch->num_draws = 0;
  77         batch->stage = FD_STAGE_NULL;
  78
  79         fd_reset_wfi(batch);
  80
  81         /* reset maximal bounds: */
  82         batch->max_scissor.minx = batch->max_scissor.miny = ~0;
  83         batch->max_scissor.maxx = batch->max_scissor.maxy = 0;
  84
  85         util_dynarray_init(&batch->draw_patches, NULL);
  86
  87         if (is_a3xx(ctx->screen))
  88                 util_dynarray_init(&batch->rbrc_patches, NULL);
  89
  90         assert(batch->resources->entries == 0);
  91
  92         util_dynarray_init(&batch->samples, NULL);
  93 }
  94
  95 struct fd_batch *
  96 fd_batch_create(struct fd_context *ctx, bool nondraw)
  97 {
  98         struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
  99
 100         if (!batch)
 101                 return NULL;
 102
 103         DBG("%p", batch);
 104
 105         pipe_reference_init(&batch->reference, 1);
 106         batch->ctx = ctx;
 107         batch->nondraw = nondraw;
 108
 109         batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 110                         _mesa_key_pointer_equal);
 111
 112         batch_init(batch);
 113
 114         return batch;
 115 }
 116
 117 static void
 118 batch_fini(struct fd_batch *batch)
 119 {
 120         pipe_resource_reference(&batch->query_buf, NULL);
 121
 122         if (batch->in_fence_fd != -1)
 123                 close(batch->in_fence_fd);
 124
 125         /* in case batch wasn't flushed but fence was created: */
 126         fd_fence_populate(batch->fence, 0, -1);
 127
 128         fd_fence_ref(NULL, &batch->fence, NULL);
 129
 130         fd_ringbuffer_del(batch->draw);
 131         if (!batch->nondraw) {
 132                 fd_ringbuffer_del(batch->binning);
 133                 fd_ringbuffer_del(batch->gmem);
 134         } else {
 135                 debug_assert(!batch->binning);
 136                 debug_assert(!batch->gmem);
 137         }
 138         if (batch->lrz_clear) {
 139                 fd_ringbuffer_del(batch->lrz_clear);
 140                 batch->lrz_clear = NULL;
 141         }
 142
 143         util_dynarray_fini(&batch->draw_patches);
 144
 145         if (is_a3xx(batch->ctx->screen))
 146                 util_dynarray_fini(&batch->rbrc_patches);
 147
 148         while (batch->samples.size > 0) {
 149                 struct fd_hw_sample *samp =
 150                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 151                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 152         }
 153         util_dynarray_fini(&batch->samples);
 154
 155         if (batch->ctx->screen->reorder)
 156                 util_queue_fence_destroy(&batch->flush_fence);
 157 }
 158
 159 static void
 160 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 161 {
 162         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 163         struct fd_batch *dep;
 164
 165         foreach_batch(dep, cache, batch->dependents_mask) {
 166                 if (flush)
 167                         fd_batch_flush(dep, false, false);
 168                 fd_batch_reference(&dep, NULL);
 169         }
 170
 171         batch->dependents_mask = 0;
 172 }
 173
 174 static void
 175 batch_reset_resources_locked(struct fd_batch *batch)
 176 {
 177         struct set_entry *entry;
 178
 179         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 180
 181         set_foreach(batch->resources, entry) {
 182                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 183                 _mesa_set_remove(batch->resources, entry);
 184                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 185                 rsc->batch_mask &= ~(1 << batch->idx);
 186                 if (rsc->write_batch == batch)
 187                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 188         }
 189 }
 190
 191 static void
 192 batch_reset_resources(struct fd_batch *batch)
 193 {
 194         mtx_lock(&batch->ctx->screen->lock);
 195         batch_reset_resources_locked(batch);
 196         mtx_unlock(&batch->ctx->screen->lock);
 197 }
 198
 199 static void
 200 batch_reset(struct fd_batch *batch)
 201 {
 202         DBG("%p", batch);
 203
 204         fd_batch_sync(batch);
 205
 206         batch_flush_reset_dependencies(batch, false);
 207         batch_reset_resources(batch);
 208
 209         batch_fini(batch);
 210         batch_init(batch);
 211 }
 212
 213 void
 214 fd_batch_reset(struct fd_batch *batch)
 215 {
 216         if (batch->needs_flush)
 217                 batch_reset(batch);
 218 }
 219
 220 void
 221 __fd_batch_destroy(struct fd_batch *batch)
 222 {
 223         DBG("%p", batch);
 224
 225         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 226
 227         mtx_lock(&batch->ctx->screen->lock);
 228         fd_bc_invalidate_batch(batch, true);
 229         mtx_unlock(&batch->ctx->screen->lock);
 230
 231         batch_fini(batch);
 232
 233         batch_reset_resources(batch);
 234         debug_assert(batch->resources->entries == 0);
 235         _mesa_set_destroy(batch->resources, NULL);
 236
 237         batch_flush_reset_dependencies(batch, false);
 238         debug_assert(batch->dependents_mask == 0);
 239
 240         free(batch);
 241 }
 242
 243 void
 244 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 245 {
 246         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 247 }
 248
 249 void
 250 fd_batch_sync(struct fd_batch *batch)
 251 {
 252         if (!batch->ctx->screen->reorder)
 253                 return;
 254         util_queue_fence_wait(&batch->flush_fence);
 255 }
 256
 257 static void
 258 batch_flush_func(void *job, int id)
 259 {
 260         struct fd_batch *batch = job;
 261
 262         fd_gmem_render_tiles(batch);
 263         batch_reset_resources(batch);
 264 }
 265
 266 static void
 267 batch_cleanup_func(void *job, int id)
 268 {
 269         struct fd_batch *batch = job;
 270         fd_batch_reference(&batch, NULL);
 271 }
 272
 273 static void
 274 batch_flush(struct fd_batch *batch, bool force)
 275 {
 276         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 277
 278         if (!batch->needs_flush) {
 279                 return;
 280         }
 281
 282         batch->needs_flush = false;
 283
 284         /* close out the draw cmds by making sure any active queries are
 285          * paused:
 286          */
 287         fd_batch_set_stage(batch, FD_STAGE_NULL);
 288
 289         fd_context_all_dirty(batch->ctx);
 290         batch_flush_reset_dependencies(batch, true);
 291
 292         if (batch->ctx->screen->reorder) {
 293                 struct fd_batch *tmp = NULL;
 294                 fd_batch_reference(&tmp, batch);
 295
 296                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 297                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 298
 299                 util_queue_add_job(&batch->ctx->flush_queue,
 300                                 batch, &batch->flush_fence,
 301                                 batch_flush_func, batch_cleanup_func);
 302         } else {
 303                 fd_gmem_render_tiles(batch);
 304                 batch_reset_resources(batch);
 305         }
 306
 307         debug_assert(batch->reference.count > 0);
 308
 309         if (batch == batch->ctx->batch) {
 310                 batch_reset(batch);
 311         } else {
 312                 mtx_lock(&batch->ctx->screen->lock);
 313                 fd_bc_invalidate_batch(batch, false);
 314                 mtx_unlock(&batch->ctx->screen->lock);
 315         }
 316 }
 317
 318 /* NOTE: could drop the last ref to batch
 319  *
 320  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 321  *   to kernel before this returns, as opposed to just being queued to be
 322  *   flushed
 323  * @force: force a flush even if no rendering, mostly useful if you need
 324  *   a fence to sync on
 325  */
 326 void
 327 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 328 {
 329         /* NOTE: we need to hold an extra ref across the body of flush,
 330          * since the last ref to this batch could be dropped when cleaning
 331          * up used_resources
 332          */
 333         struct fd_batch *tmp = NULL;
 334
 335         fd_batch_reference(&tmp, batch);
 336         batch_flush(tmp, force);
 337         if (sync)
 338                 fd_batch_sync(tmp);
 339         fd_batch_reference(&tmp, NULL);
 340 }
 341
 342 /* does 'batch' depend directly or indirectly on 'other' ? */
 343 static bool
 344 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 345 {
 346         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 347         struct fd_batch *dep;
 348
 349         if (batch->dependents_mask & (1 << other->idx))
 350                 return true;
 351
 352         foreach_batch(dep, cache, batch->dependents_mask)
 353                 if (batch_depends_on(batch, dep))
 354                         return true;
 355
 356         return false;
 357 }
 358
 359 void
 360 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 361 {
 362         if (batch->dependents_mask & (1 << dep->idx))
 363                 return;
 364
 365         /* a loop should not be possible */
 366         debug_assert(!batch_depends_on(dep, batch));
 367
 368         struct fd_batch *other = NULL;
 369         fd_batch_reference_locked(&other, dep);
 370         batch->dependents_mask |= (1 << dep->idx);
 371         DBG("%p: added dependency on %p", batch, dep);
 372 }
 373
 374 static void
 375 flush_write_batch(struct fd_resource *rsc)
 376 {
 377         struct fd_batch *b = NULL;
 378         fd_batch_reference(&b, rsc->write_batch);
 379
 380         mtx_unlock(&b->ctx->screen->lock);
 381         fd_batch_flush(b, true, false);
 382         mtx_lock(&b->ctx->screen->lock);
 383
 384         fd_bc_invalidate_batch(b, false);
 385         fd_batch_reference_locked(&b, NULL);
 386 }
 387
 388 void
 389 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 390 {
 391         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 392
 393         if (rsc->stencil)
 394                 fd_batch_resource_used(batch, rsc->stencil, write);
 395
 396         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 397
 398         if (write)
 399                 rsc->valid = true;
 400
 401         /* note, invalidate write batch, to avoid further writes to rsc
 402          * resulting in a write-after-read hazard.
 403          */
 404
 405         if (write) {
 406                 /* if we are pending read or write by any other batch: */
 407                 if (rsc->batch_mask & ~(1 << batch->idx)) {
 408                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 409                         struct fd_batch *dep;
 410
 411                         if (rsc->write_batch && rsc->write_batch != batch)
 412                                 flush_write_batch(rsc);
 413
 414                         foreach_batch(dep, cache, rsc->batch_mask) {
 415                                 struct fd_batch *b = NULL;
 416                                 if (dep == batch)
 417                                         continue;
 418                                 /* note that batch_add_dep could flush and unref dep, so
 419                                  * we need to hold a reference to keep it live for the
 420                                  * fd_bc_invalidate_batch()
 421                                  */
 422                                 fd_batch_reference(&b, dep);
 423                                 fd_batch_add_dep(batch, b);
 424                                 fd_bc_invalidate_batch(b, false);
 425                                 fd_batch_reference_locked(&b, NULL);
 426                         }
 427                 }
 428                 fd_batch_reference_locked(&rsc->write_batch, batch);
 429         } else {
 430                 /* If reading a resource pending a write, go ahead and flush the
 431                  * writer.  This avoids situations where we end up having to
 432                  * flush the current batch in _resource_used()
 433                  */
 434                 if (rsc->write_batch && rsc->write_batch != batch)
 435                         flush_write_batch(rsc);
 436         }
 437
 438         if (rsc->batch_mask & (1 << batch->idx))
 439                 return;
 440
 441         debug_assert(!_mesa_set_search(batch->resources, rsc));
 442
 443         _mesa_set_add(batch->resources, rsc);
 444         rsc->batch_mask |= (1 << batch->idx);
 445 }
 446
 447 void
 448 fd_batch_check_size(struct fd_batch *batch)
 449 {
 450         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 451                 return;
 452
 453         struct fd_ringbuffer *ring = batch->draw;
 454         if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
 455                         (fd_mesa_debug & FD_DBG_FLUSH))
 456                 fd_batch_flush(batch, true, false);
 457 }
 458
 459 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 460  * been one since last draw:
 461  */
 462 void
 463 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 464 {
 465         if (batch->needs_wfi) {
 466                 if (batch->ctx->screen->gpu_id >= 500)
 467                         OUT_WFI5(ring);
 468                 else
 469                         OUT_WFI(ring);
 470                 batch->needs_wfi = false;
 471         }
 472 }