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 = 0;
  73         batch->restore = batch->resolve = 0;
  74         batch->needs_flush = false;
  75         batch->flushed = false;
  76         batch->gmem_reason = 0;
  77         batch->num_draws = 0;
  78         batch->stage = FD_STAGE_NULL;
  79
  80         fd_reset_wfi(batch);
  81
  82         util_dynarray_init(&batch->draw_patches, NULL);
  83
  84         if (is_a3xx(ctx->screen))
  85                 util_dynarray_init(&batch->rbrc_patches, NULL);
  86
  87         util_dynarray_init(&batch->gmem_patches, NULL);
  88
  89         assert(batch->resources->entries == 0);
  90
  91         util_dynarray_init(&batch->samples, NULL);
  92 }
  93
  94 struct fd_batch *
  95 fd_batch_create(struct fd_context *ctx, bool nondraw)
  96 {
  97         struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
  98
  99         if (!batch)
 100                 return NULL;
 101
 102         DBG("%p", batch);
 103
 104         pipe_reference_init(&batch->reference, 1);
 105         batch->ctx = ctx;
 106         batch->nondraw = nondraw;
 107
 108         batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 109                         _mesa_key_pointer_equal);
 110
 111         batch_init(batch);
 112
 113         return batch;
 114 }
 115
 116 static void
 117 batch_fini(struct fd_batch *batch)
 118 {
 119         DBG("%p", batch);
 120
 121         pipe_resource_reference(&batch->query_buf, NULL);
 122
 123         if (batch->in_fence_fd != -1)
 124                 close(batch->in_fence_fd);
 125
 126         /* in case batch wasn't flushed but fence was created: */
 127         fd_fence_populate(batch->fence, 0, -1);
 128
 129         fd_fence_ref(NULL, &batch->fence, NULL);
 130
 131         fd_ringbuffer_del(batch->draw);
 132         if (!batch->nondraw) {
 133                 fd_ringbuffer_del(batch->binning);
 134                 fd_ringbuffer_del(batch->gmem);
 135         } else {
 136                 debug_assert(!batch->binning);
 137                 debug_assert(!batch->gmem);
 138         }
 139         if (batch->lrz_clear) {
 140                 fd_ringbuffer_del(batch->lrz_clear);
 141                 batch->lrz_clear = NULL;
 142         }
 143
 144         util_dynarray_fini(&batch->draw_patches);
 145
 146         if (is_a3xx(batch->ctx->screen))
 147                 util_dynarray_fini(&batch->rbrc_patches);
 148
 149         util_dynarray_fini(&batch->gmem_patches);
 150
 151         while (batch->samples.size > 0) {
 152                 struct fd_hw_sample *samp =
 153                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 154                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 155         }
 156         util_dynarray_fini(&batch->samples);
 157
 158         if (batch->ctx->screen->reorder)
 159                 util_queue_fence_destroy(&batch->flush_fence);
 160 }
 161
 162 static void
 163 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 164 {
 165         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 166         struct fd_batch *dep;
 167
 168         foreach_batch(dep, cache, batch->dependents_mask) {
 169                 if (flush)
 170                         fd_batch_flush(dep, false, false);
 171                 fd_batch_reference(&dep, NULL);
 172         }
 173
 174         batch->dependents_mask = 0;
 175 }
 176
 177 static void
 178 batch_reset_resources_locked(struct fd_batch *batch)
 179 {
 180         struct set_entry *entry;
 181
 182         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 183
 184         set_foreach(batch->resources, entry) {
 185                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 186                 _mesa_set_remove(batch->resources, entry);
 187                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 188                 rsc->batch_mask &= ~(1 << batch->idx);
 189                 if (rsc->write_batch == batch)
 190                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 191         }
 192 }
 193
 194 static void
 195 batch_reset_resources(struct fd_batch *batch)
 196 {
 197         mtx_lock(&batch->ctx->screen->lock);
 198         batch_reset_resources_locked(batch);
 199         mtx_unlock(&batch->ctx->screen->lock);
 200 }
 201
 202 static void
 203 batch_reset(struct fd_batch *batch)
 204 {
 205         DBG("%p", batch);
 206
 207         fd_batch_sync(batch);
 208
 209         batch_flush_reset_dependencies(batch, false);
 210         batch_reset_resources(batch);
 211
 212         batch_fini(batch);
 213         batch_init(batch);
 214 }
 215
 216 void
 217 fd_batch_reset(struct fd_batch *batch)
 218 {
 219         if (batch->needs_flush)
 220                 batch_reset(batch);
 221 }
 222
 223 void
 224 __fd_batch_destroy(struct fd_batch *batch)
 225 {
 226         struct fd_context *ctx = batch->ctx;
 227
 228         DBG("%p", batch);
 229
 230         fd_context_assert_locked(batch->ctx);
 231
 232         fd_bc_invalidate_batch(batch, true);
 233
 234         batch_reset_resources_locked(batch);
 235         debug_assert(batch->resources->entries == 0);
 236         _mesa_set_destroy(batch->resources, NULL);
 237
 238         fd_context_unlock(ctx);
 239         batch_flush_reset_dependencies(batch, false);
 240         debug_assert(batch->dependents_mask == 0);
 241
 242         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 243         batch_fini(batch);
 244         free(batch);
 245         fd_context_lock(ctx);
 246 }
 247
 248 void
 249 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 250 {
 251         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 252 }
 253
 254 void
 255 fd_batch_sync(struct fd_batch *batch)
 256 {
 257         if (!batch->ctx->screen->reorder)
 258                 return;
 259         util_queue_fence_wait(&batch->flush_fence);
 260 }
 261
 262 static void
 263 batch_flush_func(void *job, int id)
 264 {
 265         struct fd_batch *batch = job;
 266
 267         DBG("%p", batch);
 268
 269         fd_gmem_render_tiles(batch);
 270         batch_reset_resources(batch);
 271 }
 272
 273 static void
 274 batch_cleanup_func(void *job, int id)
 275 {
 276         struct fd_batch *batch = job;
 277         fd_batch_reference(&batch, NULL);
 278 }
 279
 280 static void
 281 batch_flush(struct fd_batch *batch, bool force)
 282 {
 283         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 284
 285         if (batch->flushed)
 286                 return;
 287
 288         batch->needs_flush = false;
 289
 290         /* close out the draw cmds by making sure any active queries are
 291          * paused:
 292          */
 293         fd_batch_set_stage(batch, FD_STAGE_NULL);
 294
 295         batch_flush_reset_dependencies(batch, true);
 296
 297         batch->flushed = true;
 298
 299         if (batch->ctx->screen->reorder) {
 300                 struct fd_batch *tmp = NULL;
 301                 fd_batch_reference(&tmp, batch);
 302
 303                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 304                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 305
 306                 util_queue_add_job(&batch->ctx->flush_queue,
 307                                 batch, &batch->flush_fence,
 308                                 batch_flush_func, batch_cleanup_func);
 309         } else {
 310                 fd_gmem_render_tiles(batch);
 311                 batch_reset_resources(batch);
 312         }
 313
 314         debug_assert(batch->reference.count > 0);
 315
 316         mtx_lock(&batch->ctx->screen->lock);
 317         fd_bc_invalidate_batch(batch, false);
 318         mtx_unlock(&batch->ctx->screen->lock);
 319 }
 320
 321 /* NOTE: could drop the last ref to batch
 322  *
 323  * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
 324  *   to kernel before this returns, as opposed to just being queued to be
 325  *   flushed
 326  * @force: force a flush even if no rendering, mostly useful if you need
 327  *   a fence to sync on
 328  */
 329 void
 330 fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 331 {
 332         struct fd_batch *tmp = NULL;
 333         bool newbatch = false;
 334
 335         /* NOTE: we need to hold an extra ref across the body of flush,
 336          * since the last ref to this batch could be dropped when cleaning
 337          * up used_resources
 338          */
 339         fd_batch_reference(&tmp, batch);
 340
 341         if (batch == batch->ctx->batch) {
 342                 batch->ctx->batch = NULL;
 343                 newbatch = true;
 344         }
 345
 346         batch_flush(tmp, force);
 347
 348         if (newbatch) {
 349                 struct fd_context *ctx = batch->ctx;
 350                 struct fd_batch *new_batch;
 351
 352                 if (ctx->screen->reorder) {
 353                         /* defer allocating new batch until one is needed for rendering
 354                          * to avoid unused batches for apps that create many contexts
 355                          */
 356                         new_batch = NULL;
 357                 } else {
 358                         new_batch = fd_batch_create(ctx, false);
 359                         util_copy_framebuffer_state(&new_batch->framebuffer, &batch->framebuffer);
 360                 }
 361
 362                 fd_batch_reference(&batch, NULL);
 363                 ctx->batch = new_batch;
 364         }
 365
 366         if (sync)
 367                 fd_batch_sync(tmp);
 368
 369         fd_batch_reference(&tmp, NULL);
 370 }
 371
 372 /* does 'batch' depend directly or indirectly on 'other' ? */
 373 static bool
 374 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 375 {
 376         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 377         struct fd_batch *dep;
 378
 379         if (batch->dependents_mask & (1 << other->idx))
 380                 return true;
 381
 382         foreach_batch(dep, cache, batch->dependents_mask)
 383                 if (batch_depends_on(batch, dep))
 384                         return true;
 385
 386         return false;
 387 }
 388
 389 void
 390 fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 391 {
 392         if (batch->dependents_mask & (1 << dep->idx))
 393                 return;
 394
 395         /* a loop should not be possible */
 396         debug_assert(!batch_depends_on(dep, batch));
 397
 398         struct fd_batch *other = NULL;
 399         fd_batch_reference_locked(&other, dep);
 400         batch->dependents_mask |= (1 << dep->idx);
 401         DBG("%p: added dependency on %p", batch, dep);
 402 }
 403
 404 static void
 405 flush_write_batch(struct fd_resource *rsc)
 406 {
 407         struct fd_batch *b = NULL;
 408         fd_batch_reference(&b, rsc->write_batch);
 409
 410         mtx_unlock(&b->ctx->screen->lock);
 411         fd_batch_flush(b, true, false);
 412         mtx_lock(&b->ctx->screen->lock);
 413
 414         fd_bc_invalidate_batch(b, false);
 415         fd_batch_reference_locked(&b, NULL);
 416 }
 417
 418 void
 419 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 420 {
 421         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 422
 423         if (rsc->stencil)
 424                 fd_batch_resource_used(batch, rsc->stencil, write);
 425
 426         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 427
 428         if (write)
 429                 rsc->valid = true;
 430
 431         /* note, invalidate write batch, to avoid further writes to rsc
 432          * resulting in a write-after-read hazard.
 433          */
 434
 435         if (write) {
 436                 /* if we are pending read or write by any other batch: */
 437                 if (rsc->batch_mask & ~(1 << batch->idx)) {
 438                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 439                         struct fd_batch *dep;
 440
 441                         if (rsc->write_batch && rsc->write_batch != batch)
 442                                 flush_write_batch(rsc);
 443
 444                         foreach_batch(dep, cache, rsc->batch_mask) {
 445                                 struct fd_batch *b = NULL;
 446                                 if (dep == batch)
 447                                         continue;
 448                                 /* note that batch_add_dep could flush and unref dep, so
 449                                  * we need to hold a reference to keep it live for the
 450                                  * fd_bc_invalidate_batch()
 451                                  */
 452                                 fd_batch_reference(&b, dep);
 453                                 fd_batch_add_dep(batch, b);
 454                                 fd_bc_invalidate_batch(b, false);
 455                                 fd_batch_reference_locked(&b, NULL);
 456                         }
 457                 }
 458                 fd_batch_reference_locked(&rsc->write_batch, batch);
 459         } else {
 460                 /* If reading a resource pending a write, go ahead and flush the
 461                  * writer.  This avoids situations where we end up having to
 462                  * flush the current batch in _resource_used()
 463                  */
 464                 if (rsc->write_batch && rsc->write_batch != batch)
 465                         flush_write_batch(rsc);
 466         }
 467
 468         if (rsc->batch_mask & (1 << batch->idx))
 469                 return;
 470
 471         debug_assert(!_mesa_set_search(batch->resources, rsc));
 472
 473         _mesa_set_add(batch->resources, rsc);
 474         rsc->batch_mask |= (1 << batch->idx);
 475 }
 476
 477 void
 478 fd_batch_check_size(struct fd_batch *batch)
 479 {
 480         debug_assert(!batch->flushed);
 481
 482         if (unlikely(fd_mesa_debug & FD_DBG_FLUSH)) {
 483                 fd_batch_flush(batch, true, false);
 484                 return;
 485         }
 486
 487         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 488                 return;
 489
 490         struct fd_ringbuffer *ring = batch->draw;
 491         if ((ring->cur - ring->start) > (ring->size/4 - 0x1000))
 492                 fd_batch_flush(batch, true, false);
 493 }
 494
 495 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 496  * been one since last draw:
 497  */
 498 void
 499 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 500 {
 501         if (batch->needs_wfi) {
 502                 if (batch->ctx->screen->gpu_id >= 500)
 503                         OUT_WFI5(ring);
 504                 else
 505                         OUT_WFI(ring);
 506                 batch->needs_wfi = false;
 507         }
 508 }