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);
  80
  81         if (is_a3xx(ctx->screen))
  82                 util_dynarray_init(&batch->rbrc_patches);
  83
  84         assert(batch->resources->entries == 0);
  85
  86         util_dynarray_init(&batch->samples);
  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
 122         util_dynarray_fini(&batch->draw_patches);
 123
 124         if (is_a3xx(batch->ctx->screen))
 125                 util_dynarray_fini(&batch->rbrc_patches);
 126
 127         while (batch->samples.size > 0) {
 128                 struct fd_hw_sample *samp =
 129                         util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
 130                 fd_hw_sample_reference(batch->ctx, &samp, NULL);
 131         }
 132         util_dynarray_fini(&batch->samples);
 133
 134         if (batch->ctx->screen->reorder)
 135                 util_queue_fence_destroy(&batch->flush_fence);
 136 }
 137
 138 static void
 139 batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 140 {
 141         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 142         struct fd_batch *dep;
 143
 144         foreach_batch(dep, cache, batch->dependents_mask) {
 145                 if (flush)
 146                         fd_batch_flush(dep, false);
 147                 fd_batch_reference(&dep, NULL);
 148         }
 149
 150         batch->dependents_mask = 0;
 151 }
 152
 153 static void
 154 batch_reset_resources_locked(struct fd_batch *batch)
 155 {
 156         struct set_entry *entry;
 157
 158         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 159
 160         set_foreach(batch->resources, entry) {
 161                 struct fd_resource *rsc = (struct fd_resource *)entry->key;
 162                 _mesa_set_remove(batch->resources, entry);
 163                 debug_assert(rsc->batch_mask & (1 << batch->idx));
 164                 rsc->batch_mask &= ~(1 << batch->idx);
 165                 if (rsc->write_batch == batch)
 166                         fd_batch_reference_locked(&rsc->write_batch, NULL);
 167         }
 168 }
 169
 170 static void
 171 batch_reset_resources(struct fd_batch *batch)
 172 {
 173         mtx_lock(&batch->ctx->screen->lock);
 174         batch_reset_resources_locked(batch);
 175         pipe_mutex_unlock(batch->ctx->screen->lock);
 176 }
 177
 178 static void
 179 batch_reset(struct fd_batch *batch)
 180 {
 181         DBG("%p", batch);
 182
 183         fd_batch_sync(batch);
 184
 185         batch_flush_reset_dependencies(batch, false);
 186         batch_reset_resources(batch);
 187
 188         batch_fini(batch);
 189         batch_init(batch);
 190 }
 191
 192 void
 193 fd_batch_reset(struct fd_batch *batch)
 194 {
 195         if (batch->needs_flush)
 196                 batch_reset(batch);
 197 }
 198
 199 void
 200 __fd_batch_destroy(struct fd_batch *batch)
 201 {
 202         DBG("%p", batch);
 203
 204         util_copy_framebuffer_state(&batch->framebuffer, NULL);
 205
 206         mtx_lock(&batch->ctx->screen->lock);
 207         fd_bc_invalidate_batch(batch, true);
 208         pipe_mutex_unlock(batch->ctx->screen->lock);
 209
 210         batch_fini(batch);
 211
 212         batch_reset_resources(batch);
 213         debug_assert(batch->resources->entries == 0);
 214         _mesa_set_destroy(batch->resources, NULL);
 215
 216         batch_flush_reset_dependencies(batch, false);
 217         debug_assert(batch->dependents_mask == 0);
 218
 219         free(batch);
 220 }
 221
 222 void
 223 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 224 {
 225         util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 226 }
 227
 228 void
 229 fd_batch_sync(struct fd_batch *batch)
 230 {
 231         if (!batch->ctx->screen->reorder)
 232                 return;
 233         util_queue_fence_wait(&batch->flush_fence);
 234 }
 235
 236 static void
 237 batch_flush_func(void *job, int id)
 238 {
 239         struct fd_batch *batch = job;
 240
 241         fd_gmem_render_tiles(batch);
 242         batch_reset_resources(batch);
 243 }
 244
 245 static void
 246 batch_cleanup_func(void *job, int id)
 247 {
 248         struct fd_batch *batch = job;
 249         fd_batch_reference(&batch, NULL);
 250 }
 251
 252 static void
 253 batch_flush(struct fd_batch *batch)
 254 {
 255         DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 256
 257         if (!batch->needs_flush)
 258                 return;
 259
 260         batch->needs_flush = false;
 261
 262         /* close out the draw cmds by making sure any active queries are
 263          * paused:
 264          */
 265         fd_hw_query_set_stage(batch, batch->draw, FD_STAGE_NULL);
 266
 267         batch->ctx->dirty = ~0;
 268         batch_flush_reset_dependencies(batch, true);
 269
 270         if (batch->ctx->screen->reorder) {
 271                 struct fd_batch *tmp = NULL;
 272                 fd_batch_reference(&tmp, batch);
 273
 274                 if (!util_queue_is_initialized(&batch->ctx->flush_queue))
 275                         util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1);
 276
 277                 util_queue_add_job(&batch->ctx->flush_queue,
 278                                 batch, &batch->flush_fence,
 279                                 batch_flush_func, batch_cleanup_func);
 280         } else {
 281                 fd_gmem_render_tiles(batch);
 282                 batch_reset_resources(batch);
 283         }
 284
 285         debug_assert(batch->reference.count > 0);
 286
 287         if (batch == batch->ctx->batch) {
 288                 batch_reset(batch);
 289         } else {
 290                 mtx_lock(&batch->ctx->screen->lock);
 291                 fd_bc_invalidate_batch(batch, false);
 292                 pipe_mutex_unlock(batch->ctx->screen->lock);
 293         }
 294 }
 295
 296 /* NOTE: could drop the last ref to batch */
 297 void
 298 fd_batch_flush(struct fd_batch *batch, bool sync)
 299 {
 300         /* NOTE: we need to hold an extra ref across the body of flush,
 301          * since the last ref to this batch could be dropped when cleaning
 302          * up used_resources
 303          */
 304         struct fd_batch *tmp = NULL;
 305         fd_batch_reference(&tmp, batch);
 306         batch_flush(tmp);
 307         if (sync)
 308                 fd_batch_sync(tmp);
 309         fd_batch_reference(&tmp, NULL);
 310 }
 311
 312 /* does 'batch' depend directly or indirectly on 'other' ? */
 313 static bool
 314 batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 315 {
 316         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 317         struct fd_batch *dep;
 318
 319         if (batch->dependents_mask & (1 << other->idx))
 320                 return true;
 321
 322         foreach_batch(dep, cache, batch->dependents_mask)
 323                 if (batch_depends_on(batch, dep))
 324                         return true;
 325
 326         return false;
 327 }
 328
 329 static void
 330 batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 331 {
 332         if (batch->dependents_mask & (1 << dep->idx))
 333                 return;
 334
 335         /* if the new depedency already depends on us, we need to flush
 336          * to avoid a loop in the dependency graph.
 337          */
 338         if (batch_depends_on(dep, batch)) {
 339                 DBG("%p: flush forced on %p!", batch, dep);
 340                 pipe_mutex_unlock(batch->ctx->screen->lock);
 341                 fd_batch_flush(dep, false);
 342                 mtx_lock(&batch->ctx->screen->lock);
 343         } else {
 344                 struct fd_batch *other = NULL;
 345                 fd_batch_reference_locked(&other, dep);
 346                 batch->dependents_mask |= (1 << dep->idx);
 347                 DBG("%p: added dependency on %p", batch, dep);
 348         }
 349 }
 350
 351 void
 352 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 353 {
 354         pipe_mutex_assert_locked(batch->ctx->screen->lock);
 355
 356         if (rsc->stencil)
 357                 fd_batch_resource_used(batch, rsc->stencil, write);
 358
 359         DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 360
 361         /* note, invalidate write batch, to avoid further writes to rsc
 362          * resulting in a write-after-read hazard.
 363          */
 364
 365         if (write) {
 366                 /* if we are pending read or write by any other batch: */
 367                 if (rsc->batch_mask != (1 << batch->idx)) {
 368                         struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 369                         struct fd_batch *dep;
 370                         foreach_batch(dep, cache, rsc->batch_mask) {
 371                                 struct fd_batch *b = NULL;
 372                                 /* note that batch_add_dep could flush and unref dep, so
 373                                  * we need to hold a reference to keep it live for the
 374                                  * fd_bc_invalidate_batch()
 375                                  */
 376                                 fd_batch_reference(&b, dep);
 377                                 batch_add_dep(batch, b);
 378                                 fd_bc_invalidate_batch(b, false);
 379                                 fd_batch_reference_locked(&b, NULL);
 380                         }
 381                 }
 382                 fd_batch_reference_locked(&rsc->write_batch, batch);
 383         } else {
 384                 if (rsc->write_batch) {
 385                         batch_add_dep(batch, rsc->write_batch);
 386                         fd_bc_invalidate_batch(rsc->write_batch, false);
 387                 }
 388         }
 389
 390         if (rsc->batch_mask & (1 << batch->idx))
 391                 return;
 392
 393         debug_assert(!_mesa_set_search(batch->resources, rsc));
 394
 395         _mesa_set_add(batch->resources, rsc);
 396         rsc->batch_mask |= (1 << batch->idx);
 397 }
 398
 399 void
 400 fd_batch_check_size(struct fd_batch *batch)
 401 {
 402         if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 403                 return;
 404
 405         struct fd_ringbuffer *ring = batch->draw;
 406         if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
 407                         (fd_mesa_debug & FD_DBG_FLUSH))
 408                 fd_batch_flush(batch, true);
 409 }
 410
 411 /* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
 412  * been one since last draw:
 413  */
 414 void
 415 fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
 416 {
 417         if (batch->needs_wfi) {
 418                 if (batch->ctx->screen->gpu_id >= 500)
 419                         OUT_WFI5(ring);
 420                 else
 421                         OUT_WFI(ring);
 422                 batch->needs_wfi = false;
 423         }
 424 }