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         batch->binning = fd_ringbuffer_new(ctx->pipe, size);
  59         batch->gmem    = fd_ringbuffer_new(ctx->pipe, size);
  60
  61         fd_ringbuffer_set_parent(batch->gmem, NULL);
  62         fd_ringbuffer_set_parent(batch->draw, batch->gmem);
  63         fd_ringbuffer_set_parent(batch->binning, batch->gmem);
  64
  65         batch->in_fence_fd = -1;
  66         batch->fence = fd_fence_create(batch);
  67
  68         batch->cleared = batch->partial_cleared = 0;
  69         batch->restore = batch->resolve = 0;
  70         batch->needs_flush = false;
  71         batch->gmem_reason = 0;
  72         batch->num_draws = 0;
  73         batch->stage = FD_STAGE_NULL;
  74
  75         fd_reset_wfi(batch);
  76
  77         /* reset maximal bounds: */
  78         batch->max_scissor.minx = batch->max_scissor.miny = ~0;
  79         batch->max_scissor.maxx = batch->max_scissor.maxy = 0;
  80
  81         util_dynarray_init(&batch->draw_patches, NULL);
  82
  83         if (is_a3xx(ctx->screen))
  84                 util_dynarray_init(&batch->rbrc_patches, NULL);
  85
  86         assert(batch->resources->entries == 0);
  87
  88         util_dynarray_init(&batch->samples, NULL);
  89 }
  90
  91 struct fd_batch *
  92 fd_batch_create(struct fd_context *ctx)
  93 {
  94         struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
  95
  96         if (!batch)
  97                 return NULL;
  98
  99         DBG("%p", batch);
 100
 101         pipe_reference_init(&batch->reference, 1);
 102         batch->ctx = ctx;
 103
 104         batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 105                         _mesa_key_pointer_equal);
 106
 107         batch_init(batch);
 108
 109         return batch;
 110 }
 111
 112 static void
 113 batch_fini(struct fd_batch *batch)
 114 {
 115         pipe_resource_reference(&batch->query_buf, NULL);
 116
 117         if (batch->in_fence_fd != -1)
 118                 close(batch->in_fence_fd);
 119
 120         /* in case batch wasn't flushed but fence was created: */
 121         fd_fence_populate(batch->fence, 0, -1);
 122
 123         fd_fence_ref(NULL, &batch->fence, NULL);
 124
 125         fd_ringbuffer_del(batch->draw);
 126         fd_ringbuffer_del(batch->binning);
 127         fd_ringbuffer_del(batch->gmem);
 128         if (batch->lrz_clear) {
 129                 fd_ringbuffer_del(batch->lrz_clear);
 130                 batch->lrz_clear = NULL;
 131         }
 132
 133         util_dynarray_fini(&batch->draw_patches);
 134
 135         if (is_a3xx(batch->ctx->screen))
 136                 util_dynarray_fini(&batch->rbrc_patches);
 137
 138         while (batch->samples.size > 0) {
 139                 struct fd_hw_sample *samp =
 140                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 141                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 142         }
 143         util_dynarray_fini(&batch->samples);
 144
 145         if (batch->ctx->screen->reorder)
 146                 util_queue_fence_destroy(&batch->flush_fence);
 147 }
 148
 149 static void
 150 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 151 {
 152         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 153         struct fd_batch *dep;
 154
 155         foreach_batch(dep, cache, batch->dependents_mask) {
 156                 if (flush)
 157                         fd_batch_flush(dep, false, false);
 158                 fd_batch_reference(&dep, NULL);
 159         }
 160
 161         batch->dependents_mask = 0;
 162 }
 163
 164 static void
 165 batch_reset_resources_locked(struct fd_batch *batch)
 166 {
 167         struct set_entry *entry;
 168
 169         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 170
 171         set_foreach(batch->resources, entry) {
 172                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 173                 _mesa_set_remove(batch->resources, entry);
 174                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 175                 rsc->batch_mask &= ~(1 << batch->idx);
 176                 if (rsc->write_batch == batch)
 177                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 178         }
 179 }
 180
 181 static void
 182 batch_reset_resources(struct fd_batch *batch)
 183 {
 184         mtx_lock(&batch->ctx->screen->lock);
 185         batch_reset_resources_locked(batch);
 186         mtx_unlock(&batch->ctx->screen->lock);
 187 }
 188
 189 static void
 190 batch_reset(struct fd_batch *batch)
 191 {
 192         DBG("%p", batch);
 193
 194         fd_batch_sync(batch);
 195
 196         batch_flush_reset_dependencies(batch, false);
 197         batch_reset_resources(batch);
 198
 199         batch_fini(batch);
 200         batch_init(batch);
 201 }
 202
 203 void
 204 fd_batch_reset(struct fd_batch *batch)
 205 {
 206         if (batch->needs_flush)
 207                 batch_reset(batch);
 208 }
 209
 210 void
 211 __fd_batch_destroy(struct fd_batch *batch)
 212 {
 213         DBG("%p", batch);
 214
 215         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 216
 217         mtx_lock(&batch->ctx->screen->lock);
 218         fd_bc_invalidate_batch(batch, true);
 219         mtx_unlock(&batch->ctx->screen->lock);
 220
 221         batch_fini(batch);
 222
 223         batch_reset_resources(batch);
 224         debug_assert(batch->resources->entries == 0);
 225         _mesa_set_destroy(batch->resources, NULL);
 226
 227         batch_flush_reset_dependencies(batch, false);
 228         debug_assert(batch->dependents_mask == 0);
 229
 230         free(batch);
 231 }
 232
 233 void
 234 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 235 {
 236         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 237 }
 238
 239 void
 240 fd_batch_sync(struct fd_batch *batch)
 241 {
 242         if (!batch->ctx->screen->reorder)
 243                 return;
 244         util_queue_fence_wait(&batch->flush_fence);
 245 }
 246
 247 static void
 248 batch_flush_func(void *job, int id)
 249 {
 250         struct fd_batch *batch = job;
 251
 252         fd_gmem_render_tiles(batch);
 253         batch_reset_resources(batch);
 254 }
 255
 256 static void
 257 batch_cleanup_func(void *job, int id)
 258 {
 259         struct fd_batch *batch = job;
 260         fd_batch_reference(&batch, NULL);
 261 }
 262
 263 static void
 264 batch_flush(struct fd_batch *batch, bool force)
 265 {
 266         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 267
 268         if (!batch->needs_flush) {
 269                 if (force) {
 270                         fd_gmem_render_noop(batch);
 271                         goto out;
 272                 }
 273                 return;
 274         }
 275
 276         batch->needs_flush = false;
 277
 278         /* close out the draw cmds by making sure any active queries are
 279          * paused:
 280          */
 281         fd_batch_set_stage(batch, FD_STAGE_NULL);
 282
 283         fd_context_all_dirty(batch->ctx);
 284         batch_flush_reset_dependencies(batch, true);
 285
 286         if (batch->ctx->screen->reorder) {
 287                 struct fd_batch *tmp = NULL;
 288                 fd_batch_reference(&tmp, batch);
 289
 290                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 291                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 292
 293                 util_queue_add_job(&batch->ctx->flush_queue,
 294                                 batch, &batch->flush_fence,
 295                                 batch_flush_func, batch_cleanup_func);
 296         } else {
 297                 fd_gmem_render_tiles(batch);
 298                 batch_reset_resources(batch);
 299         }
 300
 301         debug_assert(batch->reference.count > 0);
 302
 303 out:
 304         if (batch == batch->ctx->batch) {
 305                 batch_reset(batch);
 306         } else {
 307                 mtx_lock(&batch->ctx->screen->lock);
 308                 fd_bc_invalidate_batch(batch, false);
 309                 mtx_unlock(&batch->ctx->screen->lock);
 310         }
 311 }
 312
 313 /* NOTE: could drop the last ref to batch
 314  *
 315  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 316  *   to kernel before this returns, as opposed to just being queued to be
 317  *   flushed
 318  * @force: force a flush even if no rendering, mostly useful if you need
 319  *   a fence to sync on
 320  */
 321 void
 322 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 323 {
 324         /* NOTE: we need to hold an extra ref across the body of flush,
 325          * since the last ref to this batch could be dropped when cleaning
 326          * up used_resources
 327          */
 328         struct fd_batch *tmp = NULL;
 329         fd_batch_reference(&tmp, batch);
 330         batch_flush(tmp, force);
 331         if (sync)
 332                 fd_batch_sync(tmp);
 333         fd_batch_reference(&tmp, NULL);
 334 }
 335
 336 /* does 'batch' depend directly or indirectly on 'other' ? */
 337 static bool
 338 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 339 {
 340         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 341         struct fd_batch *dep;
 342
 343         if (batch->dependents_mask & (1 << other->idx))
 344                 return true;
 345
 346         foreach_batch(dep, cache, batch->dependents_mask)
 347                 if (batch_depends_on(batch, dep))
 348                         return true;
 349
 350         return false;
 351 }
 352
 353 void
 354 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 355 {
 356         if (batch->dependents_mask & (1 << dep->idx))
 357                 return;
 358
 359         /* if the new depedency already depends on us, we need to flush
 360          * to avoid a loop in the dependency graph.
 361          */
 362         if (batch_depends_on(dep, batch)) {
 363                 DBG("%p: flush forced on %p!", batch, dep);
 364                 mtx_unlock(&batch->ctx->screen->lock);
 365                 fd_batch_flush(dep, false, false);
 366                 mtx_lock(&batch->ctx->screen->lock);
 367         } else {
 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
 375 void
 376 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 377 {
 378         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 379
 380         if (rsc->stencil)
 381                 fd_batch_resource_used(batch, rsc->stencil, write);
 382
 383         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 384
 385         /* note, invalidate write batch, to avoid further writes to rsc
 386          * resulting in a write-after-read hazard.
 387          */
 388
 389         if (write) {
 390                 /* if we are pending read or write by any other batch: */
 391                 if (rsc->batch_mask != (1 << batch->idx)) {
 392                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 393                         struct fd_batch *dep;
 394                         foreach_batch(dep, cache, rsc->batch_mask) {
 395                                 struct fd_batch *b = NULL;
 396                                 /* note that batch_add_dep could flush and unref dep, so
 397                                  * we need to hold a reference to keep it live for the
 398                                  * fd_bc_invalidate_batch()
 399                                  */
 400                                 fd_batch_reference(&b, dep);
 401                                 fd_batch_add_dep(batch, b);
 402                                 fd_bc_invalidate_batch(b, false);
 403                                 fd_batch_reference_locked(&b, NULL);
 404                         }
 405                 }
 406                 fd_batch_reference_locked(&rsc->write_batch, batch);
 407         } else {
 408                 if (rsc->write_batch) {
 409                         fd_batch_add_dep(batch, rsc->write_batch);
 410                         fd_bc_invalidate_batch(rsc->write_batch, false);
 411                 }
 412         }
 413
 414         if (rsc->batch_mask & (1 << batch->idx))
 415                 return;
 416
 417         debug_assert(!_mesa_set_search(batch->resources, rsc));
 418
 419         _mesa_set_add(batch->resources, rsc);
 420         rsc->batch_mask |= (1 << batch->idx);
 421 }
 422
 423 void
 424 fd_batch_check_size(struct fd_batch *batch)
 425 {
 426         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 427                 return;
 428
 429         struct fd_ringbuffer *ring = batch->draw;
 430         if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
 431                         (fd_mesa_debug & FD_DBG_FLUSH))
 432                 fd_batch_flush(batch, true, false);
 433 }
 434
 435 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 436  * been one since last draw:
 437  */
 438 void
 439 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 440 {
 441         if (batch->needs_wfi) {
 442                 if (batch->ctx->screen->gpu_id >= 500)
 443                         OUT_WFI5(ring);
 444                 else
 445                         OUT_WFI(ring);
 446                 batch->needs_wfi = false;
 447         }
 448 }