src/gallium/drivers/vc4/vc4_job.c

   1 /*
   2  * Copyright © 2014-2015 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 vc4_job.c
  25  *
  26  * Functions for submitting VC4 render jobs to the kernel.
  27  */
  28
  29 #include <xf86drm.h>
  30 #include "vc4_cl_dump.h"
  31 #include "vc4_context.h"
  32 #include "util/hash_table.h"
  33
  34 static void
  35 remove_from_ht(struct hash_table *ht, void *key)
  36 {
  37         struct hash_entry *entry = _mesa_hash_table_search(ht, key);
  38         _mesa_hash_table_remove(ht, entry);
  39 }
  40
  41 static void
  42 vc4_job_free(struct vc4_context *vc4, struct vc4_job *job)
  43 {
  44         struct vc4_bo **referenced_bos = job->bo_pointers.base;
  45         for (int i = 0; i < cl_offset(&job->bo_handles) / 4; i++) {
  46                 vc4_bo_unreference(&referenced_bos[i]);
  47         }
  48
  49         remove_from_ht(vc4->jobs, &job->key);
  50
  51         if (job->color_write) {
  52                 remove_from_ht(vc4->write_jobs, job->color_write->texture);
  53                 pipe_surface_reference(&job->color_write, NULL);
  54         }
  55         if (job->msaa_color_write) {
  56                 remove_from_ht(vc4->write_jobs, job->msaa_color_write->texture);
  57                 pipe_surface_reference(&job->msaa_color_write, NULL);
  58         }
  59         if (job->zs_write) {
  60                 remove_from_ht(vc4->write_jobs, job->zs_write->texture);
  61                 pipe_surface_reference(&job->zs_write, NULL);
  62         }
  63         if (job->msaa_zs_write) {
  64                 remove_from_ht(vc4->write_jobs, job->msaa_zs_write->texture);
  65                 pipe_surface_reference(&job->msaa_zs_write, NULL);
  66         }
  67
  68         pipe_surface_reference(&job->color_read, NULL);
  69         pipe_surface_reference(&job->zs_read, NULL);
  70
  71         if (vc4->job == job)
  72                 vc4->job = NULL;
  73
  74         ralloc_free(job);
  75 }
  76
  77 static struct vc4_job *
  78 vc4_job_create(struct vc4_context *vc4)
  79 {
  80         struct vc4_job *job = rzalloc(vc4, struct vc4_job);
  81
  82         vc4_init_cl(job, &job->bcl);
  83         vc4_init_cl(job, &job->shader_rec);
  84         vc4_init_cl(job, &job->uniforms);
  85         vc4_init_cl(job, &job->bo_handles);
  86         vc4_init_cl(job, &job->bo_pointers);
  87
  88         job->draw_min_x = ~0;
  89         job->draw_min_y = ~0;
  90         job->draw_max_x = 0;
  91         job->draw_max_y = 0;
  92
  93         job->last_gem_handle_hindex = ~0;
  94
  95         if (vc4->perfmon)
  96                 job->perfmon = vc4->perfmon;
  97
  98         return job;
  99 }
 100
 101 void
 102 vc4_flush_jobs_writing_resource(struct vc4_context *vc4,
 103                                 struct pipe_resource *prsc)
 104 {
 105         struct hash_entry *entry = _mesa_hash_table_search(vc4->write_jobs,
 106                                                            prsc);
 107         if (entry) {
 108                 struct vc4_job *job = entry->data;
 109                 vc4_job_submit(vc4, job);
 110         }
 111 }
 112
 113 void
 114 vc4_flush_jobs_reading_resource(struct vc4_context *vc4,
 115                                 struct pipe_resource *prsc)
 116 {
 117         struct vc4_resource *rsc = vc4_resource(prsc);
 118
 119         vc4_flush_jobs_writing_resource(vc4, prsc);
 120
 121         hash_table_foreach(vc4->jobs, entry) {
 122                 struct vc4_job *job = entry->data;
 123
 124                 struct vc4_bo **referenced_bos = job->bo_pointers.base;
 125                 bool found = false;
 126                 for (int i = 0; i < cl_offset(&job->bo_handles) / 4; i++) {
 127                         if (referenced_bos[i] == rsc->bo) {
 128                                 found = true;
 129                                 break;
 130                         }
 131                 }
 132                 if (found) {
 133                         vc4_job_submit(vc4, job);
 134                         continue;
 135                 }
 136
 137                 /* Also check for the Z/color buffers, since the references to
 138                  * those are only added immediately before submit.
 139                  */
 140                 if (job->color_read && !(job->cleared & PIPE_CLEAR_COLOR)) {
 141                         struct vc4_resource *ctex =
 142                                 vc4_resource(job->color_read->texture);
 143                         if (ctex->bo == rsc->bo) {
 144                                 vc4_job_submit(vc4, job);
 145                                 continue;
 146                         }
 147                 }
 148
 149                 if (job->zs_read && !(job->cleared &
 150                                       (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL))) {
 151                         struct vc4_resource *ztex =
 152                                 vc4_resource(job->zs_read->texture);
 153                         if (ztex->bo == rsc->bo) {
 154                                 vc4_job_submit(vc4, job);
 155                                 continue;
 156                         }
 157                 }
 158         }
 159 }
 160
 161 /**
 162  * Returns a vc4_job struture for tracking V3D rendering to a particular FBO.
 163  *
 164  * If we've already started rendering to this FBO, then return old same job,
 165  * otherwise make a new one.  If we're beginning rendering to an FBO, make
 166  * sure that any previous reads of the FBO (or writes to its color/Z surfaces)
 167  * have been flushed.
 168  */
 169 struct vc4_job *
 170 vc4_get_job(struct vc4_context *vc4,
 171             struct pipe_surface *cbuf, struct pipe_surface *zsbuf)
 172 {
 173         /* Return the existing job for this FBO if we have one */
 174         struct vc4_job_key local_key = {.cbuf = cbuf, .zsbuf = zsbuf};
 175         struct hash_entry *entry = _mesa_hash_table_search(vc4->jobs,
 176                                                            &local_key);
 177         if (entry)
 178                 return entry->data;
 179
 180         /* Creating a new job.  Make sure that any previous jobs reading or
 181          * writing these buffers are flushed.
 182          */
 183         if (cbuf)
 184                 vc4_flush_jobs_reading_resource(vc4, cbuf->texture);
 185         if (zsbuf)
 186                 vc4_flush_jobs_reading_resource(vc4, zsbuf->texture);
 187
 188         struct vc4_job *job = vc4_job_create(vc4);
 189
 190         if (cbuf) {
 191                 if (cbuf->texture->nr_samples > 1) {
 192                         job->msaa = true;
 193                         pipe_surface_reference(&job->msaa_color_write, cbuf);
 194                 } else {
 195                         pipe_surface_reference(&job->color_write, cbuf);
 196                 }
 197         }
 198
 199         if (zsbuf) {
 200                 if (zsbuf->texture->nr_samples > 1) {
 201                         job->msaa = true;
 202                         pipe_surface_reference(&job->msaa_zs_write, zsbuf);
 203                 } else {
 204                         pipe_surface_reference(&job->zs_write, zsbuf);
 205                 }
 206         }
 207
 208         if (job->msaa) {
 209                 job->tile_width = 32;
 210                 job->tile_height = 32;
 211         } else {
 212                 job->tile_width = 64;
 213                 job->tile_height = 64;
 214         }
 215
 216         if (cbuf)
 217                 _mesa_hash_table_insert(vc4->write_jobs, cbuf->texture, job);
 218         if (zsbuf)
 219                 _mesa_hash_table_insert(vc4->write_jobs, zsbuf->texture, job);
 220
 221         job->key.cbuf = cbuf;
 222         job->key.zsbuf = zsbuf;
 223         _mesa_hash_table_insert(vc4->jobs, &job->key, job);
 224
 225         return job;
 226 }
 227
 228 struct vc4_job *
 229 vc4_get_job_for_fbo(struct vc4_context *vc4)
 230 {
 231         if (vc4->job)
 232                 return vc4->job;
 233
 234         struct pipe_surface *cbuf = vc4->framebuffer.cbufs[0];
 235         struct pipe_surface *zsbuf = vc4->framebuffer.zsbuf;
 236         struct vc4_job *job = vc4_get_job(vc4, cbuf, zsbuf);
 237
 238         /* The dirty flags are tracking what's been updated while vc4->job has
 239          * been bound, so set them all to ~0 when switching between jobs.  We
 240          * also need to reset all state at the start of rendering.
 241          */
 242         vc4->dirty = ~0;
 243
 244         /* Set up the read surfaces in the job.  If they aren't actually
 245          * getting read (due to a clear starting the frame), job->cleared will
 246          * mask out the read.
 247          */
 248         pipe_surface_reference(&job->color_read, cbuf);
 249         pipe_surface_reference(&job->zs_read, zsbuf);
 250
 251         /* If we're binding to uninitialized buffers, no need to load their
 252          * contents before drawing.
 253          */
 254         if (cbuf) {
 255                 struct vc4_resource *rsc = vc4_resource(cbuf->texture);
 256                 if (!rsc->writes)
 257                         job->cleared |= PIPE_CLEAR_COLOR0;
 258         }
 259
 260         if (zsbuf) {
 261                 struct vc4_resource *rsc = vc4_resource(zsbuf->texture);
 262                 if (!rsc->writes)
 263                         job->cleared |= PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL;
 264         }
 265
 266         job->draw_tiles_x = DIV_ROUND_UP(vc4->framebuffer.width,
 267                                          job->tile_width);
 268         job->draw_tiles_y = DIV_ROUND_UP(vc4->framebuffer.height,
 269                                          job->tile_height);
 270
 271         /* Initialize the job with the raster order flags -- each draw will
 272          * check that we haven't changed the flags, since that requires a
 273          * flush.
 274          */
 275         if (vc4->rasterizer)
 276                 job->flags = vc4->rasterizer->tile_raster_order_flags;
 277
 278         vc4->job = job;
 279
 280         return job;
 281 }
 282
 283 static void
 284 vc4_submit_setup_rcl_surface(struct vc4_job *job,
 285                              struct drm_vc4_submit_rcl_surface *submit_surf,
 286                              struct pipe_surface *psurf,
 287                              bool is_depth, bool is_write)
 288 {
 289         struct vc4_surface *surf = vc4_surface(psurf);
 290
 291         if (!surf)
 292                 return;
 293
 294         struct vc4_resource *rsc = vc4_resource(psurf->texture);
 295         submit_surf->hindex = vc4_gem_hindex(job, rsc->bo);
 296         submit_surf->offset = surf->offset;
 297
 298         if (psurf->texture->nr_samples <= 1) {
 299                 if (is_depth) {
 300                         submit_surf->bits =
 301                                 VC4_SET_FIELD(VC4_LOADSTORE_TILE_BUFFER_ZS,
 302                                               VC4_LOADSTORE_TILE_BUFFER_BUFFER);
 303
 304                 } else {
 305                         submit_surf->bits =
 306                                 VC4_SET_FIELD(VC4_LOADSTORE_TILE_BUFFER_COLOR,
 307                                               VC4_LOADSTORE_TILE_BUFFER_BUFFER) |
 308                                 VC4_SET_FIELD(vc4_rt_format_is_565(psurf->format) ?
 309                                               VC4_LOADSTORE_TILE_BUFFER_BGR565 :
 310                                               VC4_LOADSTORE_TILE_BUFFER_RGBA8888,
 311                                               VC4_LOADSTORE_TILE_BUFFER_FORMAT);
 312                 }
 313                 submit_surf->bits |=
 314                         VC4_SET_FIELD(surf->tiling,
 315                                       VC4_LOADSTORE_TILE_BUFFER_TILING);
 316         } else {
 317                 assert(!is_write);
 318                 submit_surf->flags |= VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES;
 319         }
 320
 321         if (is_write)
 322                 rsc->writes++;
 323 }
 324
 325 static void
 326 vc4_submit_setup_rcl_render_config_surface(struct vc4_job *job,
 327                                            struct drm_vc4_submit_rcl_surface *submit_surf,
 328                                            struct pipe_surface *psurf)
 329 {
 330         struct vc4_surface *surf = vc4_surface(psurf);
 331
 332         if (!surf)
 333                 return;
 334
 335         struct vc4_resource *rsc = vc4_resource(psurf->texture);
 336         submit_surf->hindex = vc4_gem_hindex(job, rsc->bo);
 337         submit_surf->offset = surf->offset;
 338
 339         if (psurf->texture->nr_samples <= 1) {
 340                 submit_surf->bits =
 341                         VC4_SET_FIELD(vc4_rt_format_is_565(surf->base.format) ?
 342                                       VC4_RENDER_CONFIG_FORMAT_BGR565 :
 343                                       VC4_RENDER_CONFIG_FORMAT_RGBA8888,
 344                                       VC4_RENDER_CONFIG_FORMAT) |
 345                         VC4_SET_FIELD(surf->tiling,
 346                                       VC4_RENDER_CONFIG_MEMORY_FORMAT);
 347         }
 348
 349         rsc->writes++;
 350 }
 351
 352 static void
 353 vc4_submit_setup_rcl_msaa_surface(struct vc4_job *job,
 354                                   struct drm_vc4_submit_rcl_surface *submit_surf,
 355                                   struct pipe_surface *psurf)
 356 {
 357         struct vc4_surface *surf = vc4_surface(psurf);
 358
 359         if (!surf)
 360                 return;
 361
 362         struct vc4_resource *rsc = vc4_resource(psurf->texture);
 363         submit_surf->hindex = vc4_gem_hindex(job, rsc->bo);
 364         submit_surf->offset = surf->offset;
 365         submit_surf->bits = 0;
 366         rsc->writes++;
 367 }
 368
 369 /**
 370  * Submits the job to the kernel and then reinitializes it.
 371  */
 372 void
 373 vc4_job_submit(struct vc4_context *vc4, struct vc4_job *job)
 374 {
 375         if (!job->needs_flush)
 376                 goto done;
 377
 378         /* The RCL setup would choke if the draw bounds cause no drawing, so
 379          * just drop the drawing if that's the case.
 380          */
 381         if (job->draw_max_x <= job->draw_min_x ||
 382             job->draw_max_y <= job->draw_min_y) {
 383                 goto done;
 384         }
 385
 386         if (vc4_debug & VC4_DEBUG_CL) {
 387                 fprintf(stderr, "BCL:\n");
 388                 vc4_dump_cl(job->bcl.base, cl_offset(&job->bcl), false);
 389         }
 390
 391         if (cl_offset(&job->bcl) > 0) {
 392                 /* Increment the semaphore indicating that binning is done and
 393                  * unblocking the render thread.  Note that this doesn't act
 394                  * until the FLUSH completes.
 395                  */
 396                 cl_ensure_space(&job->bcl, 8);
 397                 cl_emit(&job->bcl, INCREMENT_SEMAPHORE, incr);
 398                 /* The FLUSH caps all of our bin lists with a
 399                  * VC4_PACKET_RETURN.
 400                  */
 401                 cl_emit(&job->bcl, FLUSH, flush);
 402         }
 403         struct drm_vc4_submit_cl submit = {
 404                 .color_read.hindex = ~0,
 405                 .zs_read.hindex = ~0,
 406                 .color_write.hindex = ~0,
 407                 .msaa_color_write.hindex = ~0,
 408                 .zs_write.hindex = ~0,
 409                 .msaa_zs_write.hindex = ~0,
 410         };
 411
 412         cl_ensure_space(&job->bo_handles, 6 * sizeof(uint32_t));
 413         cl_ensure_space(&job->bo_pointers, 6 * sizeof(struct vc4_bo *));
 414
 415         if (job->resolve & PIPE_CLEAR_COLOR) {
 416                 if (!(job->cleared & PIPE_CLEAR_COLOR)) {
 417                         vc4_submit_setup_rcl_surface(job, &submit.color_read,
 418                                                      job->color_read,
 419                                                      false, false);
 420                 }
 421                 vc4_submit_setup_rcl_render_config_surface(job,
 422                                                            &submit.color_write,
 423                                                            job->color_write);
 424                 vc4_submit_setup_rcl_msaa_surface(job,
 425                                                   &submit.msaa_color_write,
 426                                                   job->msaa_color_write);
 427         }
 428         if (job->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 429                 if (!(job->cleared & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL))) {
 430                         vc4_submit_setup_rcl_surface(job, &submit.zs_read,
 431                                                      job->zs_read, true, false);
 432                 }
 433                 vc4_submit_setup_rcl_surface(job, &submit.zs_write,
 434                                              job->zs_write, true, true);
 435                 vc4_submit_setup_rcl_msaa_surface(job, &submit.msaa_zs_write,
 436                                                   job->msaa_zs_write);
 437         }
 438
 439         if (job->msaa) {
 440                 /* This bit controls how many pixels the general
 441                  * (i.e. subsampled) loads/stores are iterating over
 442                  * (multisample loads replicate out to the other samples).
 443                  */
 444                 submit.color_write.bits |= VC4_RENDER_CONFIG_MS_MODE_4X;
 445                 /* Controls whether color_write's
 446                  * VC4_PACKET_STORE_MS_TILE_BUFFER does 4x decimation
 447                  */
 448                 submit.color_write.bits |= VC4_RENDER_CONFIG_DECIMATE_MODE_4X;
 449         }
 450
 451         submit.bo_handles = (uintptr_t)job->bo_handles.base;
 452         submit.bo_handle_count = cl_offset(&job->bo_handles) / 4;
 453         submit.bin_cl = (uintptr_t)job->bcl.base;
 454         submit.bin_cl_size = cl_offset(&job->bcl);
 455         submit.shader_rec = (uintptr_t)job->shader_rec.base;
 456         submit.shader_rec_size = cl_offset(&job->shader_rec);
 457         submit.shader_rec_count = job->shader_rec_count;
 458         submit.uniforms = (uintptr_t)job->uniforms.base;
 459         submit.uniforms_size = cl_offset(&job->uniforms);
 460         if (job->perfmon)
 461                 submit.perfmonid = job->perfmon->id;
 462
 463         assert(job->draw_min_x != ~0 && job->draw_min_y != ~0);
 464         submit.min_x_tile = job->draw_min_x / job->tile_width;
 465         submit.min_y_tile = job->draw_min_y / job->tile_height;
 466         submit.max_x_tile = (job->draw_max_x - 1) / job->tile_width;
 467         submit.max_y_tile = (job->draw_max_y - 1) / job->tile_height;
 468         submit.width = job->draw_width;
 469         submit.height = job->draw_height;
 470         if (job->cleared) {
 471                 submit.flags |= VC4_SUBMIT_CL_USE_CLEAR_COLOR;
 472                 submit.clear_color[0] = job->clear_color[0];
 473                 submit.clear_color[1] = job->clear_color[1];
 474                 submit.clear_z = job->clear_depth;
 475                 submit.clear_s = job->clear_stencil;
 476         }
 477         submit.flags |= job->flags;
 478
 479         if (vc4->screen->has_syncobj) {
 480                 submit.out_sync = vc4->job_syncobj;
 481
 482                 if (vc4->in_fence_fd >= 0) {
 483                         /* This replaces the fence in the syncobj. */
 484                         drmSyncobjImportSyncFile(vc4->fd, vc4->in_syncobj,
 485                                                  vc4->in_fence_fd);
 486                         submit.in_sync = vc4->in_syncobj;
 487                         close(vc4->in_fence_fd);
 488                         vc4->in_fence_fd = -1;
 489                 }
 490         }
 491
 492         if (!(vc4_debug & VC4_DEBUG_NORAST)) {
 493                 int ret;
 494
 495                 ret = vc4_ioctl(vc4->fd, DRM_IOCTL_VC4_SUBMIT_CL, &submit);
 496                 static bool warned = false;
 497                 if (ret && !warned) {
 498                         fprintf(stderr, "Draw call returned %s.  "
 499                                         "Expect corruption.\n", strerror(errno));
 500                         warned = true;
 501                 } else if (!ret) {
 502                         vc4->last_emit_seqno = submit.seqno;
 503                         if (job->perfmon)
 504                                 job->perfmon->last_seqno = submit.seqno;
 505                 }
 506         }
 507
 508         if (vc4->last_emit_seqno - vc4->screen->finished_seqno > 5) {
 509                 if (!vc4_wait_seqno(vc4->screen,
 510                                     vc4->last_emit_seqno - 5,
 511                                     PIPE_TIMEOUT_INFINITE,
 512                                     "job throttling")) {
 513                         fprintf(stderr, "Job throttling failed\n");
 514                 }
 515         }
 516
 517         if (vc4_debug & VC4_DEBUG_ALWAYS_SYNC) {
 518                 if (!vc4_wait_seqno(vc4->screen, vc4->last_emit_seqno,
 519                                     PIPE_TIMEOUT_INFINITE, "sync")) {
 520                         fprintf(stderr, "Wait failed.\n");
 521                         abort();
 522                 }
 523         }
 524
 525 done:
 526         vc4_job_free(vc4, job);
 527 }
 528
 529 static bool
 530 vc4_job_compare(const void *a, const void *b)
 531 {
 532         return memcmp(a, b, sizeof(struct vc4_job_key)) == 0;
 533 }
 534
 535 static uint32_t
 536 vc4_job_hash(const void *key)
 537 {
 538         return _mesa_hash_data(key, sizeof(struct vc4_job_key));
 539 }
 540
 541 int
 542 vc4_job_init(struct vc4_context *vc4)
 543 {
 544         vc4->jobs = _mesa_hash_table_create(vc4,
 545                                             vc4_job_hash,
 546                                             vc4_job_compare);
 547         vc4->write_jobs = _mesa_hash_table_create(vc4,
 548                                                   _mesa_hash_pointer,
 549                                                   _mesa_key_pointer_equal);
 550
 551         if (vc4->screen->has_syncobj) {
 552                 /* Create the syncobj as signaled since with no job executed
 553                  * there is nothing to wait on.
 554                  */
 555                 int ret = drmSyncobjCreate(vc4->fd,
 556                                            DRM_SYNCOBJ_CREATE_SIGNALED,
 557                                            &vc4->job_syncobj);
 558                 if (ret) {
 559                         /* If the screen indicated syncobj support, we should
 560                          * be able to create a signaled syncobj.
 561                          * At this point it is too late to pretend the screen
 562                          * has no syncobj support.
 563                          */
 564                         return ret;
 565                 }
 566         }
 567
 568         return 0;
 569 }
 570