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