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_resource.h"
  35 #include "freedreno_query_hw.h"
  36
  37 static void
  38 batch_init(struct fd_batch *batch)
  39 {
  40         struct fd_context *ctx = batch->ctx;
  41         unsigned size = 0;
  42
  43         if (ctx->screen->reorder)
  44                 util_queue_fence_init(&batch->flush_fence);
  45
  46         /* if kernel is too old to support unlimited # of cmd buffers, we
  47          * have no option but to allocate large worst-case sizes so that
  48          * we don't need to grow the ringbuffer.  Performance is likely to
  49          * suffer, but there is no good alternative.
  50          */
  51         if ((fd_device_version(ctx->screen->dev) < FD_VERSION_UNLIMITED_CMDS) ||
  52                         (fd_mesa_debug & FD_DBG_NOGROW)){
  53                 size = 0x100000;
  54         }
  55
  56         batch->draw    = fd_ringbuffer_new(ctx->screen->pipe, size);
  57         batch->binning = fd_ringbuffer_new(ctx->screen->pipe, size);
  58         batch->gmem    = fd_ringbuffer_new(ctx->screen->pipe, size);
  59
  60         fd_ringbuffer_set_parent(batch->gmem, NULL);
  61         fd_ringbuffer_set_parent(batch->draw, batch->gmem);
  62         fd_ringbuffer_set_parent(batch->binning, batch->gmem);
  63
  64         batch->in_fence_fd = -1;
  65
  66         batch->cleared = batch->partial_cleared = 0;
  67         batch->restore = batch->resolve = 0;
  68         batch->needs_flush = false;
  69         batch->gmem_reason = 0;
  70         batch->num_draws = 0;
  71         batch->stage = FD_STAGE_NULL;
  72
  73         fd_reset_wfi(batch);
  74
  75         /* reset maximal bounds: */
  76         batch->max_scissor.minx = batch->max_scissor.miny = ~0;
  77         batch->max_scissor.maxx = batch->max_scissor.maxy = 0;
  78
  79         util_dynarray_init(&batch->draw_patches, NULL);
  80
  81         if (is_a3xx(ctx->screen))
  82                 util_dynarray_init(&batch->rbrc_patches, NULL);
  83
  84         assert(batch->resources->entries == 0);
  85
  86         util_dynarray_init(&batch->samples, NULL);
  87 }
  88
  89 struct fd_batch *
  90 fd_batch_create(struct fd_context *ctx)
  91 {
  92         struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
  93
  94         if (!batch)
  95                 return NULL;
  96
  97         DBG("%p", batch);
  98
  99         pipe_reference_init(&batch->reference, 1);
 100         batch->ctx = ctx;
 101
 102         batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 103                         _mesa_key_pointer_equal);
 104
 105         batch_init(batch);
 106
 107         return batch;
 108 }
 109
 110 static void
 111 batch_fini(struct fd_batch *batch)
 112 {
 113         pipe_resource_reference(&batch->query_buf, NULL);
 114
 115         if (batch->in_fence_fd != -1)
 116                 close(batch->in_fence_fd);
 117
 118         fd_ringbuffer_del(batch->draw);
 119         fd_ringbuffer_del(batch->binning);
 120         fd_ringbuffer_del(batch->gmem);
 121         if (batch->lrz_clear) {
 122                 fd_ringbuffer_del(batch->lrz_clear);
 123                 batch->lrz_clear = NULL;
 124         }
 125
 126         util_dynarray_fini(&batch->draw_patches);
 127
 128         if (is_a3xx(batch->ctx->screen))
 129                 util_dynarray_fini(&batch->rbrc_patches);
 130
 131         while (batch->samples.size > 0) {
 132                 struct fd_hw_sample *samp =
 133                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 134                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 135         }
 136         util_dynarray_fini(&batch->samples);
 137
 138         if (batch->ctx->screen->reorder)
 139                 util_queue_fence_destroy(&batch->flush_fence);
 140 }
 141
 142 static void
 143 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 144 {
 145         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 146         struct fd_batch *dep;
 147
 148         foreach_batch(dep, cache, batch->dependents_mask) {
 149                 if (flush)
 150                         fd_batch_flush(dep, false);
 151                 fd_batch_reference(&dep, NULL);
 152         }
 153
 154         batch->dependents_mask = 0;
 155 }
 156
 157 static void
 158 batch_reset_resources_locked(struct fd_batch *batch)
 159 {
 160         struct set_entry *entry;
 161
 162         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 163
 164         set_foreach(batch->resources, entry) {
 165                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 166                 _mesa_set_remove(batch->resources, entry);
 167                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 168                 rsc->batch_mask &= ~(1 << batch->idx);
 169                 if (rsc->write_batch == batch)
 170                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 171         }
 172 }
 173
 174 static void
 175 batch_reset_resources(struct fd_batch *batch)
 176 {
 177         mtx_lock(&batch->ctx->screen->lock);
 178         batch_reset_resources_locked(batch);
 179         mtx_unlock(&batch->ctx->screen->lock);
 180 }
 181
 182 static void
 183 batch_reset(struct fd_batch *batch)
 184 {
 185         DBG("%p", batch);
 186
 187         fd_batch_sync(batch);
 188
 189         batch_flush_reset_dependencies(batch, false);
 190         batch_reset_resources(batch);
 191
 192         batch_fini(batch);
 193         batch_init(batch);
 194 }
 195
 196 void
 197 fd_batch_reset(struct fd_batch *batch)
 198 {
 199         if (batch->needs_flush)
 200                 batch_reset(batch);
 201 }
 202
 203 void
 204 __fd_batch_destroy(struct fd_batch *batch)
 205 {
 206         DBG("%p", batch);
 207
 208         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 209
 210         mtx_lock(&batch->ctx->screen->lock);
 211         fd_bc_invalidate_batch(batch, true);
 212         mtx_unlock(&batch->ctx->screen->lock);
 213
 214         batch_fini(batch);
 215
 216         batch_reset_resources(batch);
 217         debug_assert(batch->resources->entries == 0);
 218         _mesa_set_destroy(batch->resources, NULL);
 219
 220         batch_flush_reset_dependencies(batch, false);
 221         debug_assert(batch->dependents_mask == 0);
 222
 223         free(batch);
 224 }
 225
 226 void
 227 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 228 {
 229         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 230 }
 231
 232 void
 233 fd_batch_sync(struct fd_batch *batch)
 234 {
 235         if (!batch->ctx->screen->reorder)
 236                 return;
 237         util_queue_fence_wait(&batch->flush_fence);
 238 }
 239
 240 static void
 241 batch_flush_func(void *job, int id)
 242 {
 243         struct fd_batch *batch = job;
 244
 245         fd_gmem_render_tiles(batch);
 246         batch_reset_resources(batch);
 247 }
 248
 249 static void
 250 batch_cleanup_func(void *job, int id)
 251 {
 252         struct fd_batch *batch = job;
 253         fd_batch_reference(&batch, NULL);
 254 }
 255
 256 static void
 257 batch_flush(struct fd_batch *batch)
 258 {
 259         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 260
 261         if (!batch->needs_flush)
 262                 return;
 263
 264         batch->needs_flush = false;
 265
 266         /* close out the draw cmds by making sure any active queries are
 267          * paused:
 268          */
 269         fd_batch_set_stage(batch, FD_STAGE_NULL);
 270
 271         fd_context_all_dirty(batch->ctx);
 272         batch_flush_reset_dependencies(batch, true);
 273
 274         if (batch->ctx->screen->reorder) {
 275                 struct fd_batch *tmp = NULL;
 276                 fd_batch_reference(&tmp, batch);
 277
 278                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 279                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
 280
 281                 util_queue_add_job(&batch->ctx->flush_queue,
 282                                 batch, &batch->flush_fence,
 283                                 batch_flush_func, batch_cleanup_func);
 284         } else {
 285                 fd_gmem_render_tiles(batch);
 286                 batch_reset_resources(batch);
 287         }
 288
 289         debug_assert(batch->reference.count > 0);
 290
 291         if (batch == batch->ctx->batch) {
 292                 batch_reset(batch);
 293         } else {
 294                 mtx_lock(&batch->ctx->screen->lock);
 295                 fd_bc_invalidate_batch(batch, false);
 296                 mtx_unlock(&batch->ctx->screen->lock);
 297         }
 298 }
 299
 300 /* NOTE: could drop the last ref to batch */
 301 void
 302 fd_batch_flush(struct fd_batch *batch, bool sync)
 303 {
 304         /* NOTE: we need to hold an extra ref across the body of flush,
 305          * since the last ref to this batch could be dropped when cleaning
 306          * up used_resources
 307          */
 308         struct fd_batch *tmp = NULL;
 309         fd_batch_reference(&tmp, batch);
 310         batch_flush(tmp);
 311         if (sync)
 312                 fd_batch_sync(tmp);
 313         fd_batch_reference(&tmp, NULL);
 314 }
 315
 316 /* does 'batch' depend directly or indirectly on 'other' ? */
 317 static bool
 318 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 319 {
 320         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 321         struct fd_batch *dep;
 322
 323         if (batch->dependents_mask & (1 << other->idx))
 324                 return true;
 325
 326         foreach_batch(dep, cache, batch->dependents_mask)
 327                 if (batch_depends_on(batch, dep))
 328                         return true;
 329
 330         return false;
 331 }
 332
 333 static void
 334 batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 335 {
 336         if (batch->dependents_mask & (1 << dep->idx))
 337                 return;
 338
 339         /* if the new depedency already depends on us, we need to flush
 340          * to avoid a loop in the dependency graph.
 341          */
 342         if (batch_depends_on(dep, batch)) {
 343                 DBG("%p: flush forced on %p!", batch, dep);
 344                 mtx_unlock(&batch->ctx->screen->lock);
 345                 fd_batch_flush(dep, false);
 346                 mtx_lock(&batch->ctx->screen->lock);
 347         } else {
 348                 struct fd_batch *other = NULL;
 349                 fd_batch_reference_locked(&other, dep);
 350                 batch->dependents_mask |= (1 << dep->idx);
 351                 DBG("%p: added dependency on %p", batch, dep);
 352         }
 353 }
 354
 355 void
 356 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 357 {
 358         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 359
 360         if (rsc->stencil)
 361                 fd_batch_resource_used(batch, rsc->stencil, write);
 362
 363         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 364
 365         /* note, invalidate write batch, to avoid further writes to rsc
 366          * resulting in a write-after-read hazard.
 367          */
 368
 369         if (write) {
 370                 /* if we are pending read or write by any other batch: */
 371                 if (rsc->batch_mask != (1 << batch->idx)) {
 372                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 373                         struct fd_batch *dep;
 374                         foreach_batch(dep, cache, rsc->batch_mask) {
 375                                 struct fd_batch *b = NULL;
 376                                 /* note that batch_add_dep could flush and unref dep, so
 377                                  * we need to hold a reference to keep it live for the
 378                                  * fd_bc_invalidate_batch()
 379                                  */
 380                                 fd_batch_reference(&b, dep);
 381                                 batch_add_dep(batch, b);
 382                                 fd_bc_invalidate_batch(b, false);
 383                                 fd_batch_reference_locked(&b, NULL);
 384                         }
 385                 }
 386                 fd_batch_reference_locked(&rsc->write_batch, batch);
 387         } else {
 388                 if (rsc->write_batch) {
 389                         batch_add_dep(batch, rsc->write_batch);
 390                         fd_bc_invalidate_batch(rsc->write_batch, false);
 391                 }
 392         }
 393
 394         if (rsc->batch_mask & (1 << batch->idx))
 395                 return;
 396
 397         debug_assert(!_mesa_set_search(batch->resources, rsc));
 398
 399         _mesa_set_add(batch->resources, rsc);
 400         rsc->batch_mask |= (1 << batch->idx);
 401 }
 402
 403 void
 404 fd_batch_check_size(struct fd_batch *batch)
 405 {
 406         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 407                 return;
 408
 409         struct fd_ringbuffer *ring = batch->draw;
 410         if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
 411                         (fd_mesa_debug & FD_DBG_FLUSH))
 412                 fd_batch_flush(batch, true);
 413 }
 414
 415 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 416  * been one since last draw:
 417  */
 418 void
 419 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 420 {
 421         if (batch->needs_wfi) {
 422                 if (batch->ctx->screen->gpu_id >= 500)
 423                         OUT_WFI5(ring);
 424                 else
 425                         OUT_WFI(ring);
 426                 batch->needs_wfi = false;
 427         }
 428 }