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 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 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                                 bool is_compute_pipeline)
 189 {
 190         struct hash_entry *entry = _mesa_hash_table_search(v3d->write_jobs,
 191                                                            prsc);
 192         struct v3d_resource *rsc = v3d_resource(prsc);
 193
 194         /* We need to sync if graphics pipeline reads a resource written
 195          * by the compute pipeline. The same would be needed for the case of
 196          * graphics-compute dependency but nowadays all compute jobs
 197          * are serialized with the previous submitted job.
 198          */
 199         if (!is_compute_pipeline && rsc->bo != NULL && rsc->compute_written) {
 200            v3d->sync_on_last_compute_job = true;
 201            rsc->compute_written = false;
 202         }
 203
 204         if (!entry)
 205                 return;
 206
 207         struct v3d_job *job = entry->data;
 208
 209         bool needs_flush;
 210         switch (flush_cond) {
 211         case V3D_FLUSH_ALWAYS:
 212                 needs_flush = true;
 213                 break;
 214         case V3D_FLUSH_NOT_CURRENT_JOB:
 215                 needs_flush = !v3d->job || v3d->job != job;
 216                 break;
 217         case V3D_FLUSH_DEFAULT:
 218         default:
 219                 /* For writes from TF in the same job we use the "Wait for TF"
 220                  * feature provided by the hardware so we don't want to flush.
 221                  * The exception to this is when the caller is about to map the
 222                  * resource since in that case we don't have a 'Wait for TF'
 223                  * command the in command stream. In this scenario the caller
 224                  * is expected to set 'always_flush' to True.
 225                  */
 226                 needs_flush = !v3d_job_writes_resource_from_tf(job, prsc);
 227         }
 228
 229         if (needs_flush)
 230                 v3d_job_submit(v3d, job);
 231 }
 232
 233 void
 234 v3d_flush_jobs_reading_resource(struct v3d_context *v3d,
 235                                 struct pipe_resource *prsc,
 236                                 enum v3d_flush_cond flush_cond,
 237                                 bool is_compute_pipeline)
 238 {
 239         struct v3d_resource *rsc = v3d_resource(prsc);
 240
 241         /* We only need to force the flush on TF writes, which is the only
 242          * case where we might skip the flush to use the 'Wait for TF'
 243          * command. Here we are flushing for a read, which means that the
 244          * caller intends to write to the resource, so we don't care if
 245          * there was a previous TF write to it.
 246          */
 247         v3d_flush_jobs_writing_resource(v3d, prsc, flush_cond,
 248                                         is_compute_pipeline);
 249
 250         hash_table_foreach(v3d->jobs, entry) {
 251                 struct v3d_job *job = entry->data;
 252
 253                 if (!_mesa_set_search(job->bos, rsc->bo))
 254                         continue;
 255
 256                 bool needs_flush;
 257                 switch (flush_cond) {
 258                 case V3D_FLUSH_NOT_CURRENT_JOB:
 259                         needs_flush = !v3d->job || v3d->job != job;
 260                         break;
 261                 case V3D_FLUSH_ALWAYS:
 262                 case V3D_FLUSH_DEFAULT:
 263                 default:
 264                         needs_flush = true;
 265                 }
 266
 267                 if (needs_flush)
 268                         v3d_job_submit(v3d, job);
 269
 270                 /* Reminder: v3d->jobs is safe to keep iterating even
 271                  * after deletion of an entry.
 272                  */
 273                 continue;
 274         }
 275 }
 276
 277 static void
 278 v3d_job_set_tile_buffer_size(struct v3d_job *job)
 279 {
 280         static const uint8_t tile_sizes[] = {
 281                 64, 64,
 282                 64, 32,
 283                 32, 32,
 284                 32, 16,
 285                 16, 16,
 286         };
 287         int tile_size_index = 0;
 288         if (job->msaa)
 289                 tile_size_index += 2;
 290
 291         if (job->cbufs[3] || job->cbufs[2])
 292                 tile_size_index += 2;
 293         else if (job->cbufs[1])
 294                 tile_size_index++;
 295
 296         int max_bpp = RENDER_TARGET_MAXIMUM_32BPP;
 297         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 298                 if (job->cbufs[i]) {
 299                         struct v3d_surface *surf = v3d_surface(job->cbufs[i]);
 300                         max_bpp = MAX2(max_bpp, surf->internal_bpp);
 301                 }
 302         }
 303         job->internal_bpp = max_bpp;
 304         STATIC_ASSERT(RENDER_TARGET_MAXIMUM_32BPP == 0);
 305         tile_size_index += max_bpp;
 306
 307         assert(tile_size_index < ARRAY_SIZE(tile_sizes));
 308         job->tile_width = tile_sizes[tile_size_index * 2 + 0];
 309         job->tile_height = tile_sizes[tile_size_index * 2 + 1];
 310 }
 311
 312 /**
 313  * Returns a v3d_job struture for tracking V3D rendering to a particular FBO.
 314  *
 315  * If we've already started rendering to this FBO, then return the same job,
 316  * otherwise make a new one.  If we're beginning rendering to an FBO, make
 317  * sure that any previous reads of the FBO (or writes to its color/Z surfaces)
 318  * have been flushed.
 319  */
 320 struct v3d_job *
 321 v3d_get_job(struct v3d_context *v3d,
 322             struct pipe_surface **cbufs, struct pipe_surface *zsbuf)
 323 {
 324         /* Return the existing job for this FBO if we have one */
 325         struct v3d_job_key local_key = {
 326                 .cbufs = {
 327                         cbufs[0],
 328                         cbufs[1],
 329                         cbufs[2],
 330                         cbufs[3],
 331                 },
 332                 .zsbuf = zsbuf,
 333         };
 334         struct hash_entry *entry = _mesa_hash_table_search(v3d->jobs,
 335                                                            &local_key);
 336         if (entry)
 337                 return entry->data;
 338
 339         /* Creating a new job.  Make sure that any previous jobs reading or
 340          * writing these buffers are flushed.
 341          */
 342         struct v3d_job *job = v3d_job_create(v3d);
 343
 344         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 345                 if (cbufs[i]) {
 346                         v3d_flush_jobs_reading_resource(v3d, cbufs[i]->texture,
 347                                                         V3D_FLUSH_DEFAULT,
 348                                                         false);
 349                         pipe_surface_reference(&job->cbufs[i], cbufs[i]);
 350
 351                         if (cbufs[i]->texture->nr_samples > 1)
 352                                 job->msaa = true;
 353                 }
 354         }
 355         if (zsbuf) {
 356                 v3d_flush_jobs_reading_resource(v3d, zsbuf->texture,
 357                                                 V3D_FLUSH_DEFAULT,
 358                                                 false);
 359                 pipe_surface_reference(&job->zsbuf, zsbuf);
 360                 if (zsbuf->texture->nr_samples > 1)
 361                         job->msaa = true;
 362         }
 363
 364         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 365                 if (cbufs[i])
 366                         _mesa_hash_table_insert(v3d->write_jobs,
 367                                                 cbufs[i]->texture, job);
 368         }
 369         if (zsbuf) {
 370                 _mesa_hash_table_insert(v3d->write_jobs, zsbuf->texture, job);
 371
 372                 struct v3d_resource *rsc = v3d_resource(zsbuf->texture);
 373                 if (rsc->separate_stencil) {
 374                         v3d_flush_jobs_reading_resource(v3d,
 375                                                         &rsc->separate_stencil->base,
 376                                                         V3D_FLUSH_DEFAULT,
 377                                                         false);
 378                         _mesa_hash_table_insert(v3d->write_jobs,
 379                                                 &rsc->separate_stencil->base,
 380                                                 job);
 381                 }
 382         }
 383
 384         memcpy(&job->key, &local_key, sizeof(local_key));
 385         _mesa_hash_table_insert(v3d->jobs, &job->key, job);
 386
 387         return job;
 388 }
 389
 390 struct v3d_job *
 391 v3d_get_job_for_fbo(struct v3d_context *v3d)
 392 {
 393         if (v3d->job)
 394                 return v3d->job;
 395
 396         struct pipe_surface **cbufs = v3d->framebuffer.cbufs;
 397         struct pipe_surface *zsbuf = v3d->framebuffer.zsbuf;
 398         struct v3d_job *job = v3d_get_job(v3d, cbufs, zsbuf);
 399
 400         if (v3d->framebuffer.samples >= 1)
 401                 job->msaa = true;
 402
 403         v3d_job_set_tile_buffer_size(job);
 404
 405         /* The dirty flags are tracking what's been updated while v3d->job has
 406          * been bound, so set them all to ~0 when switching between jobs.  We
 407          * also need to reset all state at the start of rendering.
 408          */
 409         v3d->dirty = ~0;
 410
 411         /* If we're binding to uninitialized buffers, no need to load their
 412          * contents before drawing.
 413          */
 414         for (int i = 0; i < 4; i++) {
 415                 if (cbufs[i]) {
 416                         struct v3d_resource *rsc = v3d_resource(cbufs[i]->texture);
 417                         if (!rsc->writes)
 418                                 job->clear |= PIPE_CLEAR_COLOR0 << i;
 419                 }
 420         }
 421
 422         if (zsbuf) {
 423                 struct v3d_resource *rsc = v3d_resource(zsbuf->texture);
 424                 if (!rsc->writes)
 425                         job->clear |= PIPE_CLEAR_DEPTH;
 426
 427                 if (rsc->separate_stencil)
 428                         rsc = rsc->separate_stencil;
 429
 430                 if (!rsc->writes)
 431                         job->clear |= PIPE_CLEAR_STENCIL;
 432         }
 433
 434         job->draw_tiles_x = DIV_ROUND_UP(v3d->framebuffer.width,
 435                                          job->tile_width);
 436         job->draw_tiles_y = DIV_ROUND_UP(v3d->framebuffer.height,
 437                                          job->tile_height);
 438
 439         v3d->job = job;
 440
 441         return job;
 442 }
 443
 444 static void
 445 v3d_clif_dump(struct v3d_context *v3d, struct v3d_job *job)
 446 {
 447         if (!(V3D_DEBUG & (V3D_DEBUG_CL | V3D_DEBUG_CLIF)))
 448                 return;
 449
 450         struct clif_dump *clif = clif_dump_init(&v3d->screen->devinfo,
 451                                                 stderr,
 452                                                 V3D_DEBUG & V3D_DEBUG_CL);
 453
 454         set_foreach(job->bos, entry) {
 455                 struct v3d_bo *bo = (void *)entry->key;
 456                 char *name = ralloc_asprintf(NULL, "%s_0x%x",
 457                                              bo->name, bo->offset);
 458
 459                 v3d_bo_map(bo);
 460                 clif_dump_add_bo(clif, name, bo->offset, bo->size, bo->map);
 461
 462                 ralloc_free(name);
 463         }
 464
 465         clif_dump(clif, &job->submit);
 466
 467         clif_dump_destroy(clif);
 468 }
 469
 470 static void
 471 v3d_read_and_accumulate_primitive_counters(struct v3d_context *v3d)
 472 {
 473         assert(v3d->prim_counts);
 474
 475         perf_debug("stalling on TF counts readback\n");
 476         struct v3d_resource *rsc = v3d_resource(v3d->prim_counts);
 477         if (v3d_bo_wait(rsc->bo, PIPE_TIMEOUT_INFINITE, "prim-counts")) {
 478                 uint32_t *map = v3d_bo_map(rsc->bo) + v3d->prim_counts_offset;
 479                 v3d->tf_prims_generated += map[V3D_PRIM_COUNTS_TF_WRITTEN];
 480                 /* When we only have a vertex shader we determine the primitive
 481                  * count in the CPU so don't update it here again.
 482                  */
 483                 if (v3d->prog.gs)
 484                         v3d->prims_generated += map[V3D_PRIM_COUNTS_WRITTEN];
 485         }
 486 }
 487
 488 /**
 489  * Submits the job to the kernel and then reinitializes it.
 490  */
 491 void
 492 v3d_job_submit(struct v3d_context *v3d, struct v3d_job *job)
 493 {
 494         struct v3d_screen *screen = v3d->screen;
 495
 496         if (!job->needs_flush)
 497                 goto done;
 498
 499         if (screen->devinfo.ver >= 41)
 500                 v3d41_emit_rcl(job);
 501         else
 502                 v3d33_emit_rcl(job);
 503
 504         if (cl_offset(&job->bcl) > 0) {
 505                 if (screen->devinfo.ver >= 41)
 506                         v3d41_bcl_epilogue(v3d, job);
 507                 else
 508                         v3d33_bcl_epilogue(v3d, job);
 509         }
 510
 511         /* While the RCL will implicitly depend on the last RCL to have
 512          * finished, we also need to block on any previous TFU job we may have
 513          * dispatched.
 514          */
 515         job->submit.in_sync_rcl = v3d->out_sync;
 516
 517         /* Update the sync object for the last rendering by our context. */
 518         job->submit.out_sync = v3d->out_sync;
 519
 520         job->submit.bcl_end = job->bcl.bo->offset + cl_offset(&job->bcl);
 521         job->submit.rcl_end = job->rcl.bo->offset + cl_offset(&job->rcl);
 522
 523         job->submit.flags = 0;
 524         if (job->tmu_dirty_rcl && screen->has_cache_flush)
 525                 job->submit.flags |= DRM_V3D_SUBMIT_CL_FLUSH_CACHE;
 526
 527         /* On V3D 4.1, the tile alloc/state setup moved to register writes
 528          * instead of binner packets.
 529          */
 530         if (screen->devinfo.ver >= 41) {
 531                 v3d_job_add_bo(job, job->tile_alloc);
 532                 job->submit.qma = job->tile_alloc->offset;
 533                 job->submit.qms = job->tile_alloc->size;
 534
 535                 v3d_job_add_bo(job, job->tile_state);
 536                 job->submit.qts = job->tile_state->offset;
 537         }
 538
 539         v3d_clif_dump(v3d, job);
 540
 541         if (!(V3D_DEBUG & V3D_DEBUG_NORAST)) {
 542                 int ret;
 543
 544                 ret = v3d_ioctl(v3d->fd, DRM_IOCTL_V3D_SUBMIT_CL, &job->submit);
 545                 static bool warned = false;
 546                 if (ret && !warned) {
 547                         fprintf(stderr, "Draw call returned %s.  "
 548                                         "Expect corruption.\n", strerror(errno));
 549                         warned = true;
 550                 }
 551
 552                 /* If we are submitting a job in the middle of transform
 553                  * feedback we need to read the primitive counts and accumulate
 554                  * them, otherwise they will be reset at the start of the next
 555                  * draw when we emit the Tile Binning Mode Configuration packet.
 556                  *
 557                  * If the job doesn't have any TF draw calls, then we know
 558                  * the primitive count must be zero and we can skip stalling
 559                  * for this. This also fixes a problem because it seems that
 560                  * in this scenario the counters are not reset with the Tile
 561                  * Binning Mode Configuration packet, which would translate
 562                  * to us reading an obsolete (possibly non-zero) value from
 563                  * the GPU counters.
 564                  */
 565                 if (v3d->streamout.num_targets && job->tf_draw_calls_queued > 0)
 566                         v3d_read_and_accumulate_primitive_counters(v3d);
 567         }
 568
 569 done:
 570         v3d_job_free(v3d, job);
 571 }
 572
 573 static bool
 574 v3d_job_compare(const void *a, const void *b)
 575 {
 576         return memcmp(a, b, sizeof(struct v3d_job_key)) == 0;
 577 }
 578
 579 static uint32_t
 580 v3d_job_hash(const void *key)
 581 {
 582         return _mesa_hash_data(key, sizeof(struct v3d_job_key));
 583 }
 584
 585 void
 586 v3d_job_init(struct v3d_context *v3d)
 587 {
 588         v3d->jobs = _mesa_hash_table_create(v3d,
 589                                             v3d_job_hash,
 590                                             v3d_job_compare);
 591         v3d->write_jobs = _mesa_hash_table_create(v3d,
 592                                                   _mesa_hash_pointer,
 593                                                   _mesa_key_pointer_equal);
 594 }
 595