src/gallium/drivers/panfrost/pan_job.c

   1 /*
   2  * Copyright (C) 2019 Alyssa Rosenzweig
   3  * Copyright (C) 2014-2017 Broadcom
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  */
  25
  26 #include <assert.h>
  27
  28 #include "pan_context.h"
  29 #include "util/hash_table.h"
  30 #include "util/ralloc.h"
  31 #include "util/u_format.h"
  32 #include "util/u_pack_color.h"
  33
  34 static struct panfrost_batch *
  35 panfrost_create_batch(struct panfrost_context *ctx,
  36                       const struct pipe_framebuffer_state *key)
  37 {
  38         struct panfrost_batch *batch = rzalloc(ctx, struct panfrost_batch);
  39
  40         batch->ctx = ctx;
  41
  42         batch->bos = _mesa_set_create(batch,
  43                                       _mesa_hash_pointer,
  44                                       _mesa_key_pointer_equal);
  45
  46         batch->minx = batch->miny = ~0;
  47         batch->maxx = batch->maxy = 0;
  48         batch->transient_offset = 0;
  49
  50         util_dynarray_init(&batch->headers, batch);
  51         util_dynarray_init(&batch->gpu_headers, batch);
  52         util_copy_framebuffer_state(&batch->key, key);
  53
  54         return batch;
  55 }
  56
  57 static void
  58 panfrost_free_batch(struct panfrost_batch *batch)
  59 {
  60         if (!batch)
  61                 return;
  62
  63         struct panfrost_context *ctx = batch->ctx;
  64
  65         set_foreach(batch->bos, entry) {
  66                 struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
  67                 panfrost_bo_unreference(ctx->base.screen, bo);
  68         }
  69
  70         _mesa_hash_table_remove_key(ctx->batches, &batch->key);
  71
  72         if (ctx->batch == batch)
  73                 ctx->batch = NULL;
  74
  75         util_unreference_framebuffer_state(&batch->key);
  76         ralloc_free(batch);
  77 }
  78
  79 static struct panfrost_batch *
  80 panfrost_get_batch(struct panfrost_context *ctx,
  81                    const struct pipe_framebuffer_state *key)
  82 {
  83         /* Lookup the job first */
  84         struct hash_entry *entry = _mesa_hash_table_search(ctx->batches, key);
  85
  86         if (entry)
  87                 return entry->data;
  88
  89         /* Otherwise, let's create a job */
  90
  91         struct panfrost_batch *batch = panfrost_create_batch(ctx, key);
  92
  93         /* Save the created job */
  94         _mesa_hash_table_insert(ctx->batches, &batch->key, batch);
  95
  96         return batch;
  97 }
  98
  99 /* Get the job corresponding to the FBO we're currently rendering into */
 100
 101 struct panfrost_batch *
 102 panfrost_get_batch_for_fbo(struct panfrost_context *ctx)
 103 {
 104         /* If we're wallpapering, we special case to workaround
 105          * u_blitter abuse */
 106
 107         if (ctx->wallpaper_batch)
 108                 return ctx->wallpaper_batch;
 109
 110         /* If we already began rendering, use that */
 111
 112         if (ctx->batch) {
 113                 assert(util_framebuffer_state_equal(&ctx->batch->key,
 114                                                     &ctx->pipe_framebuffer));
 115                 return ctx->batch;
 116         }
 117
 118         /* If not, look up the job */
 119         struct panfrost_batch *batch = panfrost_get_batch(ctx,
 120                                                           &ctx->pipe_framebuffer);
 121
 122         /* Set this job as the current FBO job. Will be reset when updating the
 123          * FB state and when submitting or releasing a job.
 124          */
 125         ctx->batch = batch;
 126         return batch;
 127 }
 128
 129 void
 130 panfrost_batch_add_bo(struct panfrost_batch *batch, struct panfrost_bo *bo)
 131 {
 132         if (!bo)
 133                 return;
 134
 135         if (_mesa_set_search(batch->bos, bo))
 136                 return;
 137
 138         panfrost_bo_reference(bo);
 139         _mesa_set_add(batch->bos, bo);
 140 }
 141
 142 /* Returns the polygon list's GPU address if available, or otherwise allocates
 143  * the polygon list.  It's perfectly fast to use allocate/free BO directly,
 144  * since we'll hit the BO cache and this is one-per-batch anyway. */
 145
 146 mali_ptr
 147 panfrost_batch_get_polygon_list(struct panfrost_batch *batch, unsigned size)
 148 {
 149         if (batch->polygon_list) {
 150                 assert(batch->polygon_list->size >= size);
 151         } else {
 152                 struct panfrost_screen *screen = pan_screen(batch->ctx->base.screen);
 153
 154                 /* Create the BO as invisible, as there's no reason to map */
 155
 156                 batch->polygon_list = panfrost_drm_create_bo(screen,
 157                                 size, PAN_ALLOCATE_INVISIBLE);
 158                 panfrost_batch_add_bo(batch, batch->polygon_list);
 159
 160                 /* A BO reference has been retained by panfrost_batch_add_bo(),
 161                  * so we need to unreference it here if we want the BO to be
 162                  * automatically released when the batch is destroyed.
 163                  */
 164                 panfrost_bo_unreference(&screen->base, batch->polygon_list);
 165         }
 166
 167         return batch->polygon_list->gpu;
 168 }
 169
 170 static void
 171 panfrost_batch_draw_wallpaper(struct panfrost_batch *batch)
 172 {
 173         /* Nothing to reload? TODO: MRT wallpapers */
 174         if (batch->key.cbufs[0] == NULL)
 175                 return;
 176
 177         /* Check if the buffer has any content on it worth preserving */
 178
 179         struct pipe_surface *surf = batch->key.cbufs[0];
 180         struct panfrost_resource *rsrc = pan_resource(surf->texture);
 181         unsigned level = surf->u.tex.level;
 182
 183         if (!rsrc->slices[level].initialized)
 184                 return;
 185
 186         batch->ctx->wallpaper_batch = batch;
 187
 188         /* Clamp the rendering area to the damage extent. The
 189          * KHR_partial_update() spec states that trying to render outside of
 190          * the damage region is "undefined behavior", so we should be safe.
 191          */
 192         unsigned damage_width = (rsrc->damage.extent.maxx - rsrc->damage.extent.minx);
 193         unsigned damage_height = (rsrc->damage.extent.maxy - rsrc->damage.extent.miny);
 194
 195         if (damage_width && damage_height) {
 196                 panfrost_batch_intersection_scissor(batch,
 197                                                     rsrc->damage.extent.minx,
 198                                                     rsrc->damage.extent.miny,
 199                                                     rsrc->damage.extent.maxx,
 200                                                     rsrc->damage.extent.maxy);
 201         }
 202
 203         /* FIXME: Looks like aligning on a tile is not enough, but
 204          * aligning on twice the tile size seems to works. We don't
 205          * know exactly what happens here but this deserves extra
 206          * investigation to figure it out.
 207          */
 208         batch->minx = batch->minx & ~((MALI_TILE_LENGTH * 2) - 1);
 209         batch->miny = batch->miny & ~((MALI_TILE_LENGTH * 2) - 1);
 210         batch->maxx = MIN2(ALIGN_POT(batch->maxx, MALI_TILE_LENGTH * 2),
 211                            rsrc->base.width0);
 212         batch->maxy = MIN2(ALIGN_POT(batch->maxy, MALI_TILE_LENGTH * 2),
 213                            rsrc->base.height0);
 214
 215         struct pipe_scissor_state damage;
 216         struct pipe_box rects[4];
 217
 218         /* Clamp the damage box to the rendering area. */
 219         damage.minx = MAX2(batch->minx, rsrc->damage.biggest_rect.x);
 220         damage.miny = MAX2(batch->miny, rsrc->damage.biggest_rect.y);
 221         damage.maxx = MIN2(batch->maxx,
 222                            rsrc->damage.biggest_rect.x +
 223                            rsrc->damage.biggest_rect.width);
 224         damage.maxy = MIN2(batch->maxy,
 225                            rsrc->damage.biggest_rect.y +
 226                            rsrc->damage.biggest_rect.height);
 227
 228         /* One damage rectangle means we can end up with at most 4 reload
 229          * regions:
 230          * 1: left region, only exists if damage.x > 0
 231          * 2: right region, only exists if damage.x + damage.width < fb->width
 232          * 3: top region, only exists if damage.y > 0. The intersection with
 233          *    the left and right regions are dropped
 234          * 4: bottom region, only exists if damage.y + damage.height < fb->height.
 235          *    The intersection with the left and right regions are dropped
 236          *
 237          *                    ____________________________
 238          *                    |       |     3     |      |
 239          *                    |       |___________|      |
 240          *                    |       |   damage  |      |
 241          *                    |   1   |    rect   |   2  |
 242          *                    |       |___________|      |
 243          *                    |       |     4     |      |
 244          *                    |_______|___________|______|
 245          */
 246         u_box_2d(batch->minx, batch->miny, damage.minx - batch->minx,
 247                  batch->maxy - batch->miny, &rects[0]);
 248         u_box_2d(damage.maxx, batch->miny, batch->maxx - damage.maxx,
 249                  batch->maxy - batch->miny, &rects[1]);
 250         u_box_2d(damage.minx, batch->miny, damage.maxx - damage.minx,
 251                  damage.miny - batch->miny, &rects[2]);
 252         u_box_2d(damage.minx, damage.maxy, damage.maxx - damage.minx,
 253                  batch->maxy - damage.maxy, &rects[3]);
 254
 255         for (unsigned i = 0; i < 4; i++) {
 256                 /* Width and height are always >= 0 even if width is declared as a
 257                  * signed integer: u_box_2d() helper takes unsigned args and
 258                  * panfrost_set_damage_region() is taking care of clamping
 259                  * negative values.
 260                  */
 261                 if (!rects[i].width || !rects[i].height)
 262                         continue;
 263
 264                 /* Blit the wallpaper in */
 265                 panfrost_blit_wallpaper(batch->ctx, &rects[i]);
 266         }
 267         batch->ctx->wallpaper_batch = NULL;
 268 }
 269
 270 void
 271 panfrost_batch_submit(struct panfrost_batch *batch)
 272 {
 273         assert(batch);
 274
 275         struct panfrost_context *ctx = batch->ctx;
 276         int ret;
 277
 278         /* Nothing to do! */
 279         if (!batch->last_job.gpu && !batch->clear)
 280                 goto out;
 281
 282         if (!batch->clear && batch->last_tiler.gpu)
 283                 panfrost_batch_draw_wallpaper(batch);
 284
 285         panfrost_scoreboard_link_batch(batch);
 286
 287         bool has_draws = batch->last_job.gpu;
 288
 289         ret = panfrost_drm_submit_vs_fs_batch(batch, has_draws);
 290
 291         if (ret)
 292                 fprintf(stderr, "panfrost_batch_submit failed: %d\n", ret);
 293
 294 out:
 295         /* If this is the bound batch, the panfrost_context parameters are
 296          * relevant so submitting it invalidates those paramaters, but if it's
 297          * not bound, the context parameters are for some other batch so we
 298          * can't invalidate them.
 299          */
 300         if (ctx->batch == batch)
 301                 panfrost_invalidate_frame(ctx);
 302
 303         /* The job has been submitted, let's invalidate the current FBO job
 304          * cache.
 305          */
 306         assert(!ctx->batch || batch == ctx->batch);
 307         ctx->batch = NULL;
 308
 309         /* We always stall the pipeline for correct results since pipelined
 310          * rendering is quite broken right now (to be fixed by the panfrost_job
 311          * refactor, just take the perf hit for correctness)
 312          */
 313         drmSyncobjWait(pan_screen(ctx->base.screen)->fd, &ctx->out_sync, 1,
 314                        INT64_MAX, 0, NULL);
 315         panfrost_free_batch(batch);
 316 }
 317
 318 void
 319 panfrost_batch_set_requirements(struct panfrost_batch *batch)
 320 {
 321         struct panfrost_context *ctx = batch->ctx;
 322
 323         if (ctx->rasterizer && ctx->rasterizer->base.multisample)
 324                 batch->requirements |= PAN_REQ_MSAA;
 325
 326         if (ctx->depth_stencil && ctx->depth_stencil->depth.writemask)
 327                 batch->requirements |= PAN_REQ_DEPTH_WRITE;
 328 }
 329
 330 /* Helper to smear a 32-bit color across 128-bit components */
 331
 332 static void
 333 pan_pack_color_32(uint32_t *packed, uint32_t v)
 334 {
 335         for (unsigned i = 0; i < 4; ++i)
 336                 packed[i] = v;
 337 }
 338
 339 static void
 340 pan_pack_color_64(uint32_t *packed, uint32_t lo, uint32_t hi)
 341 {
 342         for (unsigned i = 0; i < 4; i += 2) {
 343                 packed[i + 0] = lo;
 344                 packed[i + 1] = hi;
 345         }
 346 }
 347
 348 static void
 349 pan_pack_color(uint32_t *packed, const union pipe_color_union *color, enum pipe_format format)
 350 {
 351         /* Alpha magicked to 1.0 if there is no alpha */
 352
 353         bool has_alpha = util_format_has_alpha(format);
 354         float clear_alpha = has_alpha ? color->f[3] : 1.0f;
 355
 356         /* Packed color depends on the framebuffer format */
 357
 358         const struct util_format_description *desc =
 359                 util_format_description(format);
 360
 361         if (util_format_is_rgba8_variant(desc)) {
 362                 pan_pack_color_32(packed,
 363                                   (float_to_ubyte(clear_alpha) << 24) |
 364                                   (float_to_ubyte(color->f[2]) << 16) |
 365                                   (float_to_ubyte(color->f[1]) <<  8) |
 366                                   (float_to_ubyte(color->f[0]) <<  0));
 367         } else if (format == PIPE_FORMAT_B5G6R5_UNORM) {
 368                 /* First, we convert the components to R5, G6, B5 separately */
 369                 unsigned r5 = CLAMP(color->f[0], 0.0, 1.0) * 31.0;
 370                 unsigned g6 = CLAMP(color->f[1], 0.0, 1.0) * 63.0;
 371                 unsigned b5 = CLAMP(color->f[2], 0.0, 1.0) * 31.0;
 372
 373                 /* Then we pack into a sparse u32. TODO: Why these shifts? */
 374                 pan_pack_color_32(packed, (b5 << 25) | (g6 << 14) | (r5 << 5));
 375         } else if (format == PIPE_FORMAT_B4G4R4A4_UNORM) {
 376                 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
 377                 unsigned r4 = CLAMP(color->f[0], 0.0, 1.0) * 240.0;
 378                 unsigned g4 = CLAMP(color->f[1], 0.0, 1.0) * 240.0;
 379                 unsigned b4 = CLAMP(color->f[2], 0.0, 1.0) * 240.0;
 380                 unsigned a4 = CLAMP(clear_alpha, 0.0, 1.0) * 240.0;
 381
 382                 /* Pack on *byte* intervals */
 383                 pan_pack_color_32(packed, (a4 << 24) | (b4 << 16) | (g4 << 8) | r4);
 384         } else if (format == PIPE_FORMAT_B5G5R5A1_UNORM) {
 385                 /* Scale as expected but shift oddly */
 386                 unsigned r5 = round(CLAMP(color->f[0], 0.0, 1.0)) * 31.0;
 387                 unsigned g5 = round(CLAMP(color->f[1], 0.0, 1.0)) * 31.0;
 388                 unsigned b5 = round(CLAMP(color->f[2], 0.0, 1.0)) * 31.0;
 389                 unsigned a1 = round(CLAMP(clear_alpha, 0.0, 1.0)) * 1.0;
 390
 391                 pan_pack_color_32(packed, (a1 << 31) | (b5 << 25) | (g5 << 15) | (r5 << 5));
 392         } else {
 393                 /* Try Gallium's generic default path. Doesn't work for all
 394                  * formats but it's a good guess. */
 395
 396                 union util_color out;
 397
 398                 if (util_format_is_pure_integer(format)) {
 399                         memcpy(out.ui, color->ui, 16);
 400                 } else {
 401                         util_pack_color(color->f, format, &out);
 402                 }
 403
 404                 unsigned size = util_format_get_blocksize(format);
 405
 406                 if (size == 1) {
 407                         unsigned b = out.ui[0];
 408                         unsigned s = b | (b << 8);
 409                         pan_pack_color_32(packed, s | (s << 16));
 410                 } else if (size == 2)
 411                         pan_pack_color_32(packed, out.ui[0] | (out.ui[0] << 16));
 412                 else if (size == 4)
 413                         pan_pack_color_32(packed, out.ui[0]);
 414                 else if (size == 8)
 415                         pan_pack_color_64(packed, out.ui[0], out.ui[1]);
 416                 else if (size == 16)
 417                         memcpy(packed, out.ui, 16);
 418                 else
 419                         unreachable("Unknown generic format size packing clear colour");
 420         }
 421 }
 422
 423 void
 424 panfrost_batch_clear(struct panfrost_batch *batch,
 425                      unsigned buffers,
 426                      const union pipe_color_union *color,
 427                      double depth, unsigned stencil)
 428 {
 429         struct panfrost_context *ctx = batch->ctx;
 430
 431         if (buffers & PIPE_CLEAR_COLOR) {
 432                 for (unsigned i = 0; i < PIPE_MAX_COLOR_BUFS; ++i) {
 433                         if (!(buffers & (PIPE_CLEAR_COLOR0 << i)))
 434                                 continue;
 435
 436                         enum pipe_format format = ctx->pipe_framebuffer.cbufs[i]->format;
 437                         pan_pack_color(batch->clear_color[i], color, format);
 438                 }
 439         }
 440
 441         if (buffers & PIPE_CLEAR_DEPTH) {
 442                 batch->clear_depth = depth;
 443         }
 444
 445         if (buffers & PIPE_CLEAR_STENCIL) {
 446                 batch->clear_stencil = stencil;
 447         }
 448
 449         batch->clear |= buffers;
 450
 451         /* Clearing affects the entire framebuffer (by definition -- this is
 452          * the Gallium clear callback, which clears the whole framebuffer. If
 453          * the scissor test were enabled from the GL side, the state tracker
 454          * would emit a quad instead and we wouldn't go down this code path) */
 455
 456         panfrost_batch_union_scissor(batch, 0, 0,
 457                                      ctx->pipe_framebuffer.width,
 458                                      ctx->pipe_framebuffer.height);
 459 }
 460
 461 static bool
 462 panfrost_batch_compare(const void *a, const void *b)
 463 {
 464         return util_framebuffer_state_equal(a, b);
 465 }
 466
 467 static uint32_t
 468 panfrost_batch_hash(const void *key)
 469 {
 470         return _mesa_hash_data(key, sizeof(struct pipe_framebuffer_state));
 471 }
 472
 473 /* Given a new bounding rectangle (scissor), let the job cover the union of the
 474  * new and old bounding rectangles */
 475
 476 void
 477 panfrost_batch_union_scissor(struct panfrost_batch *batch,
 478                              unsigned minx, unsigned miny,
 479                              unsigned maxx, unsigned maxy)
 480 {
 481         batch->minx = MIN2(batch->minx, minx);
 482         batch->miny = MIN2(batch->miny, miny);
 483         batch->maxx = MAX2(batch->maxx, maxx);
 484         batch->maxy = MAX2(batch->maxy, maxy);
 485 }
 486
 487 void
 488 panfrost_batch_intersection_scissor(struct panfrost_batch *batch,
 489                                   unsigned minx, unsigned miny,
 490                                   unsigned maxx, unsigned maxy)
 491 {
 492         batch->minx = MAX2(batch->minx, minx);
 493         batch->miny = MAX2(batch->miny, miny);
 494         batch->maxx = MIN2(batch->maxx, maxx);
 495         batch->maxy = MIN2(batch->maxy, maxy);
 496 }
 497
 498 /* Are we currently rendering to the screen (rather than an FBO)? */
 499
 500 bool
 501 panfrost_batch_is_scanout(struct panfrost_batch *batch)
 502 {
 503         /* If there is no color buffer, it's an FBO */
 504         if (batch->key.nr_cbufs != 1)
 505                 return false;
 506
 507         /* If we're too early that no framebuffer was sent, it's scanout */
 508         if (!batch->key.cbufs[0])
 509                 return true;
 510
 511         return batch->key.cbufs[0]->texture->bind & PIPE_BIND_DISPLAY_TARGET ||
 512                batch->key.cbufs[0]->texture->bind & PIPE_BIND_SCANOUT ||
 513                batch->key.cbufs[0]->texture->bind & PIPE_BIND_SHARED;
 514 }
 515
 516 void
 517 panfrost_batch_init(struct panfrost_context *ctx)
 518 {
 519         ctx->batches = _mesa_hash_table_create(ctx,
 520                                                panfrost_batch_hash,
 521                                                panfrost_batch_compare);
 522 }