src/gallium/drivers/v3d/v3d_job.c

   1 /*
   2  * Copyright © 2014-2017 Broadcom
   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
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 /** @file v3d_job.c
  25  *
  26  * Functions for submitting VC5 render jobs to the kernel.
  27  */
  28
  29 #include <xf86drm.h>
  30 #include "v3d_context.h"
  31 /* The OQ/semaphore packets are the same across V3D versions. */
  32 #define V3D_VERSION 33
  33 #include "broadcom/cle/v3dx_pack.h"
  34 #include "broadcom/common/v3d_macros.h"
  35 #include "util/hash_table.h"
  36 #include "util/ralloc.h"
  37 #include "util/set.h"
  38 #include "broadcom/clif/clif_dump.h"
  39
  40 static void
  41 v3d_job_free(struct v3d_context *v3d, struct v3d_job *job)
  42 {
  43         set_foreach(job->bos, entry) {
  44                 struct v3d_bo *bo = (struct v3d_bo *)entry->key;
  45                 v3d_bo_unreference(&bo);
  46         }
  47
  48         _mesa_hash_table_remove_key(v3d->jobs, &job->key);
  49
  50         if (job->write_prscs) {
  51                 set_foreach(job->write_prscs, entry) {
  52                         const struct pipe_resource *prsc = entry->key;
  53
  54                         _mesa_hash_table_remove_key(v3d->write_jobs, prsc);
  55                 }
  56         }
  57
  58         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
  59                 if (job->cbufs[i]) {
  60                         _mesa_hash_table_remove_key(v3d->write_jobs,
  61                                                     job->cbufs[i]->texture);
  62                         pipe_surface_reference(&job->cbufs[i], NULL);
  63                 }
  64         }
  65         if (job->zsbuf) {
  66                 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
  67                 if (rsc->separate_stencil)
  68                         _mesa_hash_table_remove_key(v3d->write_jobs,
  69                                                     &rsc->separate_stencil->base);
  70
  71                 _mesa_hash_table_remove_key(v3d->write_jobs,
  72                                             job->zsbuf->texture);
  73                 pipe_surface_reference(&job->zsbuf, NULL);
  74         }
  75
  76         if (v3d->job == job)
  77                 v3d->job = NULL;
  78
  79         v3d_destroy_cl(&job->bcl);
  80         v3d_destroy_cl(&job->rcl);
  81         v3d_destroy_cl(&job->indirect);
  82         v3d_bo_unreference(&job->tile_alloc);
  83         v3d_bo_unreference(&job->tile_state);
  84
  85         ralloc_free(job);
  86 }
  87
  88 static struct v3d_job *
  89 v3d_job_create(struct v3d_context *v3d)
  90 {
  91         struct v3d_job *job = rzalloc(v3d, struct v3d_job);
  92
  93         job->v3d = v3d;
  94
  95         v3d_init_cl(job, &job->bcl);
  96         v3d_init_cl(job, &job->rcl);
  97         v3d_init_cl(job, &job->indirect);
  98
  99         job->draw_min_x = ~0;
 100         job->draw_min_y = ~0;
 101         job->draw_max_x = 0;
 102         job->draw_max_y = 0;
 103
 104         job->bos = _mesa_set_create(job,
 105                                     _mesa_hash_pointer,
 106                                     _mesa_key_pointer_equal);
 107         return job;
 108 }
 109
 110 void
 111 v3d_job_add_bo(struct v3d_job *job, struct v3d_bo *bo)
 112 {
 113         if (!bo)
 114                 return;
 115
 116         if (_mesa_set_search(job->bos, bo))
 117                 return;
 118
 119         v3d_bo_reference(bo);
 120         _mesa_set_add(job->bos, bo);
 121         job->referenced_size += bo->size;
 122
 123         uint32_t *bo_handles = (void *)(uintptr_t)job->submit.bo_handles;
 124
 125         if (job->submit.bo_handle_count >= job->bo_handles_size) {
 126                 job->bo_handles_size = MAX2(4, job->bo_handles_size * 2);
 127                 bo_handles = reralloc(job, bo_handles,
 128                                       uint32_t, job->bo_handles_size);
 129                 job->submit.bo_handles = (uintptr_t)(void *)bo_handles;
 130         }
 131         bo_handles[job->submit.bo_handle_count++] = bo->handle;
 132 }
 133
 134 void
 135 v3d_job_add_write_resource(struct v3d_job *job, struct pipe_resource *prsc)
 136 {
 137         struct v3d_context *v3d = job->v3d;
 138
 139         if (!job->write_prscs) {
 140                 job->write_prscs = _mesa_set_create(job,
 141                                                     _mesa_hash_pointer,
 142                                                     _mesa_key_pointer_equal);
 143         }
 144
 145         _mesa_set_add(job->write_prscs, prsc);
 146         _mesa_hash_table_insert(v3d->write_jobs, prsc, job);
 147 }
 148
 149 void
 150 v3d_flush_jobs_using_bo(struct v3d_context *v3d, struct v3d_bo *bo)
 151 {
 152         hash_table_foreach(v3d->jobs, entry) {
 153                 struct v3d_job *job = entry->data;
 154
 155                 if (_mesa_set_search(job->bos, bo))
 156                         v3d_job_submit(v3d, job);
 157         }
 158 }
 159
 160 void
 161 v3d_job_add_tf_write_resource(struct v3d_job *job, struct pipe_resource *prsc)
 162 {
 163         v3d_job_add_write_resource(job, prsc);
 164
 165         if (!job->tf_write_prscs)
 166                 job->tf_write_prscs = _mesa_pointer_set_create(job);
 167
 168         _mesa_set_add(job->tf_write_prscs, prsc);
 169 }
 170
 171 static bool
 172 v3d_job_writes_resource_from_tf(struct v3d_job *job,
 173                                 struct pipe_resource *prsc)
 174 {
 175         if (!job->tf_enabled)
 176                 return false;
 177
 178         if (!job->tf_write_prscs)
 179                 return false;
 180
 181         return _mesa_set_search(job->tf_write_prscs, prsc) != NULL;
 182 }
 183
 184 void
 185 v3d_flush_jobs_writing_resource(struct v3d_context *v3d,
 186                                 struct pipe_resource *prsc,
 187                                 enum v3d_flush_cond flush_cond)
 188 {
 189         struct hash_entry *entry = _mesa_hash_table_search(v3d->write_jobs,
 190                                                            prsc);
 191         if (!entry)
 192                 return;
 193
 194         struct v3d_job *job = entry->data;
 195
 196         bool needs_flush;
 197         switch (flush_cond) {
 198         case V3D_FLUSH_ALWAYS:
 199                 needs_flush = true;
 200                 break;
 201         case V3D_FLUSH_NOT_CURRENT_JOB:
 202                 needs_flush = !v3d->job || v3d->job != job;
 203                 break;
 204         case V3D_FLUSH_DEFAULT:
 205         default:
 206                 /* For writes from TF in the same job we use the "Wait for TF"
 207                  * feature provided by the hardware so we don't want to flush.
 208                  * The exception to this is when the caller is about to map the
 209                  * resource since in that case we don't have a 'Wait for TF'
 210                  * command the in command stream. In this scenario the caller
 211                  * is expected to set 'always_flush' to True.
 212                  */
 213                 needs_flush = !v3d_job_writes_resource_from_tf(job, prsc);
 214         }
 215
 216         if (needs_flush)
 217                 v3d_job_submit(v3d, job);
 218 }
 219
 220 void
 221 v3d_flush_jobs_reading_resource(struct v3d_context *v3d,
 222                                 struct pipe_resource *prsc,
 223                                 enum v3d_flush_cond flush_cond)
 224 {
 225         struct v3d_resource *rsc = v3d_resource(prsc);
 226
 227         /* We only need to force the flush on TF writes, which is the only
 228          * case where we might skip the flush to use the 'Wait for TF'
 229          * command. Here we are flushing for a read, which means that the
 230          * caller intends to write to the resource, so we don't care if
 231          * there was a previous TF write to it.
 232          */
 233         v3d_flush_jobs_writing_resource(v3d, prsc, flush_cond);
 234
 235         hash_table_foreach(v3d->jobs, entry) {
 236                 struct v3d_job *job = entry->data;
 237
 238                 if (!_mesa_set_search(job->bos, rsc->bo))
 239                         continue;
 240
 241                 bool needs_flush;
 242                 switch (flush_cond) {
 243                 case V3D_FLUSH_NOT_CURRENT_JOB:
 244                         needs_flush = !v3d->job || v3d->job != job;
 245                         break;
 246                 case V3D_FLUSH_ALWAYS:
 247                 case V3D_FLUSH_DEFAULT:
 248                 default:
 249                         needs_flush = true;
 250                 }
 251
 252                 if (needs_flush)
 253                         v3d_job_submit(v3d, job);
 254
 255                 /* Reminder: v3d->jobs is safe to keep iterating even
 256                  * after deletion of an entry.
 257                  */
 258                 continue;
 259         }
 260 }
 261
 262 static void
 263 v3d_job_set_tile_buffer_size(struct v3d_job *job)
 264 {
 265         static const uint8_t tile_sizes[] = {
 266                 64, 64,
 267                 64, 32,
 268                 32, 32,
 269                 32, 16,
 270                 16, 16,
 271         };
 272         int tile_size_index = 0;
 273         if (job->msaa)
 274                 tile_size_index += 2;
 275
 276         if (job->cbufs[3] || job->cbufs[2])
 277                 tile_size_index += 2;
 278         else if (job->cbufs[1])
 279                 tile_size_index++;
 280
 281         int max_bpp = RENDER_TARGET_MAXIMUM_32BPP;
 282         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 283                 if (job->cbufs[i]) {
 284                         struct v3d_surface *surf = v3d_surface(job->cbufs[i]);
 285                         max_bpp = MAX2(max_bpp, surf->internal_bpp);
 286                 }
 287         }
 288         job->internal_bpp = max_bpp;
 289         STATIC_ASSERT(RENDER_TARGET_MAXIMUM_32BPP == 0);
 290         tile_size_index += max_bpp;
 291
 292         assert(tile_size_index < ARRAY_SIZE(tile_sizes));
 293         job->tile_width = tile_sizes[tile_size_index * 2 + 0];
 294         job->tile_height = tile_sizes[tile_size_index * 2 + 1];
 295 }
 296
 297 /**
 298  * Returns a v3d_job struture for tracking V3D rendering to a particular FBO.
 299  *
 300  * If we've already started rendering to this FBO, then return the same job,
 301  * otherwise make a new one.  If we're beginning rendering to an FBO, make
 302  * sure that any previous reads of the FBO (or writes to its color/Z surfaces)
 303  * have been flushed.
 304  */
 305 struct v3d_job *
 306 v3d_get_job(struct v3d_context *v3d,
 307             struct pipe_surface **cbufs, struct pipe_surface *zsbuf)
 308 {
 309         /* Return the existing job for this FBO if we have one */
 310         struct v3d_job_key local_key = {
 311                 .cbufs = {
 312                         cbufs[0],
 313                         cbufs[1],
 314                         cbufs[2],
 315                         cbufs[3],
 316                 },
 317                 .zsbuf = zsbuf,
 318         };
 319         struct hash_entry *entry = _mesa_hash_table_search(v3d->jobs,
 320                                                            &local_key);
 321         if (entry)
 322                 return entry->data;
 323
 324         /* Creating a new job.  Make sure that any previous jobs reading or
 325          * writing these buffers are flushed.
 326          */
 327         struct v3d_job *job = v3d_job_create(v3d);
 328
 329         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 330                 if (cbufs[i]) {
 331                         v3d_flush_jobs_reading_resource(v3d, cbufs[i]->texture,
 332                                                         V3D_FLUSH_DEFAULT);
 333                         pipe_surface_reference(&job->cbufs[i], cbufs[i]);
 334
 335                         if (cbufs[i]->texture->nr_samples > 1)
 336                                 job->msaa = true;
 337                 }
 338         }
 339         if (zsbuf) {
 340                 v3d_flush_jobs_reading_resource(v3d, zsbuf->texture,
 341                                                 V3D_FLUSH_DEFAULT);
 342                 pipe_surface_reference(&job->zsbuf, zsbuf);
 343                 if (zsbuf->texture->nr_samples > 1)
 344                         job->msaa = true;
 345         }
 346
 347         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 348                 if (cbufs[i])
 349                         _mesa_hash_table_insert(v3d->write_jobs,
 350                                                 cbufs[i]->texture, job);
 351         }
 352         if (zsbuf) {
 353                 _mesa_hash_table_insert(v3d->write_jobs, zsbuf->texture, job);
 354
 355                 struct v3d_resource *rsc = v3d_resource(zsbuf->texture);
 356                 if (rsc->separate_stencil) {
 357                         v3d_flush_jobs_reading_resource(v3d,
 358                                                         &rsc->separate_stencil->base,
 359                                                         V3D_FLUSH_DEFAULT);
 360                         _mesa_hash_table_insert(v3d->write_jobs,
 361                                                 &rsc->separate_stencil->base,
 362                                                 job);
 363                 }
 364         }
 365
 366         memcpy(&job->key, &local_key, sizeof(local_key));
 367         _mesa_hash_table_insert(v3d->jobs, &job->key, job);
 368
 369         return job;
 370 }
 371
 372 struct v3d_job *
 373 v3d_get_job_for_fbo(struct v3d_context *v3d)
 374 {
 375         if (v3d->job)
 376                 return v3d->job;
 377
 378         struct pipe_surface **cbufs = v3d->framebuffer.cbufs;
 379         struct pipe_surface *zsbuf = v3d->framebuffer.zsbuf;
 380         struct v3d_job *job = v3d_get_job(v3d, cbufs, zsbuf);
 381
 382         if (v3d->framebuffer.samples >= 1)
 383                 job->msaa = true;
 384
 385         v3d_job_set_tile_buffer_size(job);
 386
 387         /* The dirty flags are tracking what's been updated while v3d->job has
 388          * been bound, so set them all to ~0 when switching between jobs.  We
 389          * also need to reset all state at the start of rendering.
 390          */
 391         v3d->dirty = ~0;
 392
 393         /* If we're binding to uninitialized buffers, no need to load their
 394          * contents before drawing.
 395          */
 396         for (int i = 0; i < 4; i++) {
 397                 if (cbufs[i]) {
 398                         struct v3d_resource *rsc = v3d_resource(cbufs[i]->texture);
 399                         if (!rsc->writes)
 400                                 job->clear |= PIPE_CLEAR_COLOR0 << i;
 401                 }
 402         }
 403
 404         if (zsbuf) {
 405                 struct v3d_resource *rsc = v3d_resource(zsbuf->texture);
 406                 if (!rsc->writes)
 407                         job->clear |= PIPE_CLEAR_DEPTH;
 408
 409                 if (rsc->separate_stencil)
 410                         rsc = rsc->separate_stencil;
 411
 412                 if (!rsc->writes)
 413                         job->clear |= PIPE_CLEAR_STENCIL;
 414         }
 415
 416         job->draw_tiles_x = DIV_ROUND_UP(v3d->framebuffer.width,
 417                                          job->tile_width);
 418         job->draw_tiles_y = DIV_ROUND_UP(v3d->framebuffer.height,
 419                                          job->tile_height);
 420
 421         v3d->job = job;
 422
 423         return job;
 424 }
 425
 426 static void
 427 v3d_clif_dump(struct v3d_context *v3d, struct v3d_job *job)
 428 {
 429         if (!(V3D_DEBUG & (V3D_DEBUG_CL | V3D_DEBUG_CLIF)))
 430                 return;
 431
 432         struct clif_dump *clif = clif_dump_init(&v3d->screen->devinfo,
 433                                                 stderr,
 434                                                 V3D_DEBUG & V3D_DEBUG_CL);
 435
 436         set_foreach(job->bos, entry) {
 437                 struct v3d_bo *bo = (void *)entry->key;
 438                 char *name = ralloc_asprintf(NULL, "%s_0x%x",
 439                                              bo->name, bo->offset);
 440
 441                 v3d_bo_map(bo);
 442                 clif_dump_add_bo(clif, name, bo->offset, bo->size, bo->map);
 443
 444                 ralloc_free(name);
 445         }
 446
 447         clif_dump(clif, &job->submit);
 448
 449         clif_dump_destroy(clif);
 450 }
 451
 452 static void
 453 v3d_read_and_accumulate_primitive_counters(struct v3d_context *v3d)
 454 {
 455         assert(v3d->prim_counts);
 456
 457         perf_debug("stalling on TF counts readback\n");
 458         struct v3d_resource *rsc = v3d_resource(v3d->prim_counts);
 459         if (v3d_bo_wait(rsc->bo, PIPE_TIMEOUT_INFINITE, "prim-counts")) {
 460                 uint32_t *map = v3d_bo_map(rsc->bo) + v3d->prim_counts_offset;
 461                 v3d->tf_prims_generated += map[V3D_PRIM_COUNTS_TF_WRITTEN];
 462         }
 463 }
 464
 465 /**
 466  * Submits the job to the kernel and then reinitializes it.
 467  */
 468 void
 469 v3d_job_submit(struct v3d_context *v3d, struct v3d_job *job)
 470 {
 471         struct v3d_screen *screen = v3d->screen;
 472
 473         if (!job->needs_flush)
 474                 goto done;
 475
 476         if (screen->devinfo.ver >= 41)
 477                 v3d41_emit_rcl(job);
 478         else
 479                 v3d33_emit_rcl(job);
 480
 481         if (cl_offset(&job->bcl) > 0) {
 482                 if (screen->devinfo.ver >= 41)
 483                         v3d41_bcl_epilogue(v3d, job);
 484                 else
 485                         v3d33_bcl_epilogue(v3d, job);
 486         }
 487
 488         /* While the RCL will implicitly depend on the last RCL to have
 489          * finished, we also need to block on any previous TFU job we may have
 490          * dispatched.
 491          */
 492         job->submit.in_sync_rcl = v3d->out_sync;
 493
 494         /* Update the sync object for the last rendering by our context. */
 495         job->submit.out_sync = v3d->out_sync;
 496
 497         job->submit.bcl_end = job->bcl.bo->offset + cl_offset(&job->bcl);
 498         job->submit.rcl_end = job->rcl.bo->offset + cl_offset(&job->rcl);
 499
 500         /* On V3D 4.1, the tile alloc/state setup moved to register writes
 501          * instead of binner packets.
 502          */
 503         if (screen->devinfo.ver >= 41) {
 504                 v3d_job_add_bo(job, job->tile_alloc);
 505                 job->submit.qma = job->tile_alloc->offset;
 506                 job->submit.qms = job->tile_alloc->size;
 507
 508                 v3d_job_add_bo(job, job->tile_state);
 509                 job->submit.qts = job->tile_state->offset;
 510         }
 511
 512         v3d_clif_dump(v3d, job);
 513
 514         if (!(V3D_DEBUG & V3D_DEBUG_NORAST)) {
 515                 int ret;
 516
 517                 ret = v3d_ioctl(v3d->fd, DRM_IOCTL_V3D_SUBMIT_CL, &job->submit);
 518                 static bool warned = false;
 519                 if (ret && !warned) {
 520                         fprintf(stderr, "Draw call returned %s.  "
 521                                         "Expect corruption.\n", strerror(errno));
 522                         warned = true;
 523                 }
 524
 525                 /* If we are submitting a job in the middle of transform
 526                  * feedback we need to read the primitive counts and accumulate
 527                  * them, otherwise they will be reset at the start of the next
 528                  * draw when we emit the Tile Binning Mode Configuration packet.
 529                  *
 530                  * If the job doesn't have any TF draw calls, then we know
 531                  * the primitive count must be zero and we can skip stalling
 532                  * for this. This also fixes a problem because it seems that
 533                  * in this scenario the counters are not reset with the Tile
 534                  * Binning Mode Configuration packet, which would translate
 535                  * to us reading an obsolete (possibly non-zero) value from
 536                  * the GPU counters.
 537                  */
 538                 if (v3d->streamout.num_targets && job->tf_draw_calls_queued > 0)
 539                         v3d_read_and_accumulate_primitive_counters(v3d);
 540         }
 541
 542 done:
 543         v3d_job_free(v3d, job);
 544 }
 545
 546 static bool
 547 v3d_job_compare(const void *a, const void *b)
 548 {
 549         return memcmp(a, b, sizeof(struct v3d_job_key)) == 0;
 550 }
 551
 552 static uint32_t
 553 v3d_job_hash(const void *key)
 554 {
 555         return _mesa_hash_data(key, sizeof(struct v3d_job_key));
 556 }
 557
 558 void
 559 v3d_job_init(struct v3d_context *v3d)
 560 {
 561         v3d->jobs = _mesa_hash_table_create(v3d,
 562                                             v3d_job_hash,
 563                                             v3d_job_compare);
 564         v3d->write_jobs = _mesa_hash_table_create(v3d,
 565                                                   _mesa_hash_pointer,
 566                                                   _mesa_key_pointer_equal);
 567 }
 568