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