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