src/gallium/drivers/radeonsi/si_clear.c

   1 /*
   2  * Copyright 2017 Advanced Micro Devices, Inc.
   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  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "si_pipe.h"
  25 #include "sid.h"
  26
  27 #include "util/u_format.h"
  28 #include "util/u_pack_color.h"
  29 #include "util/u_surface.h"
  30
  31 enum {
  32         SI_CLEAR         = SI_SAVE_FRAGMENT_STATE,
  33         SI_CLEAR_SURFACE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE,
  34 };
  35
  36 static void si_alloc_separate_cmask(struct si_screen *sscreen,
  37                                     struct r600_texture *rtex)
  38 {
  39         if (rtex->cmask_buffer)
  40                 return;
  41
  42         assert(rtex->cmask.size == 0);
  43
  44         si_texture_get_cmask_info(sscreen, rtex, &rtex->cmask);
  45         if (!rtex->cmask.size)
  46                 return;
  47
  48         rtex->cmask_buffer = (struct r600_resource *)
  49                 si_aligned_buffer_create(&sscreen->b,
  50                                          R600_RESOURCE_FLAG_UNMAPPABLE,
  51                                          PIPE_USAGE_DEFAULT,
  52                                          rtex->cmask.size,
  53                                          rtex->cmask.alignment);
  54         if (rtex->cmask_buffer == NULL) {
  55                 rtex->cmask.size = 0;
  56                 return;
  57         }
  58
  59         /* update colorbuffer state bits */
  60         rtex->cmask.base_address_reg = rtex->cmask_buffer->gpu_address >> 8;
  61
  62         rtex->cb_color_info |= S_028C70_FAST_CLEAR(1);
  63
  64         p_atomic_inc(&sscreen->compressed_colortex_counter);
  65 }
  66
  67 static void si_set_clear_color(struct r600_texture *rtex,
  68                                enum pipe_format surface_format,
  69                                const union pipe_color_union *color)
  70 {
  71         union util_color uc;
  72
  73         memset(&uc, 0, sizeof(uc));
  74
  75         if (rtex->surface.bpe == 16) {
  76                 /* DCC fast clear only:
  77                  *   CLEAR_WORD0 = R = G = B
  78                  *   CLEAR_WORD1 = A
  79                  */
  80                 assert(color->ui[0] == color->ui[1] &&
  81                        color->ui[0] == color->ui[2]);
  82                 uc.ui[0] = color->ui[0];
  83                 uc.ui[1] = color->ui[3];
  84         } else if (util_format_is_pure_uint(surface_format)) {
  85                 util_format_write_4ui(surface_format, color->ui, 0, &uc, 0, 0, 0, 1, 1);
  86         } else if (util_format_is_pure_sint(surface_format)) {
  87                 util_format_write_4i(surface_format, color->i, 0, &uc, 0, 0, 0, 1, 1);
  88         } else {
  89                 util_pack_color(color->f, surface_format, &uc);
  90         }
  91
  92         memcpy(rtex->color_clear_value, &uc, 2 * sizeof(uint32_t));
  93 }
  94
  95 static bool vi_get_fast_clear_parameters(enum pipe_format surface_format,
  96                                          const union pipe_color_union *color,
  97                                          uint32_t* reset_value,
  98                                          bool* clear_words_needed)
  99 {
 100         bool values[4] = {};
 101         int i;
 102         bool main_value = false;
 103         bool extra_value = false;
 104         int extra_channel;
 105
 106         /* This is needed to get the correct DCC clear value for luminance formats.
 107          * 1) Get the linear format (because the next step can't handle L8_SRGB).
 108          * 2) Convert luminance to red. (the real hw format for luminance)
 109          */
 110         surface_format = util_format_linear(surface_format);
 111         surface_format = util_format_luminance_to_red(surface_format);
 112
 113         const struct util_format_description *desc = util_format_description(surface_format);
 114
 115         if (desc->block.bits == 128 &&
 116             (color->ui[0] != color->ui[1] ||
 117              color->ui[0] != color->ui[2]))
 118                 return false;
 119
 120         *clear_words_needed = true;
 121         *reset_value = 0x20202020U;
 122
 123         /* If we want to clear without needing a fast clear eliminate step, we
 124          * can set each channel to 0 or 1 (or 0/max for integer formats). We
 125          * have two sets of flags, one for the last or first channel(extra) and
 126          * one for the other channels(main).
 127          */
 128
 129         if (surface_format == PIPE_FORMAT_R11G11B10_FLOAT ||
 130             surface_format == PIPE_FORMAT_B5G6R5_UNORM ||
 131             surface_format == PIPE_FORMAT_B5G6R5_SRGB ||
 132             util_format_is_alpha(surface_format)) {
 133                 extra_channel = -1;
 134         } else if (desc->layout == UTIL_FORMAT_LAYOUT_PLAIN) {
 135                 if (si_translate_colorswap(surface_format, false) <= 1)
 136                         extra_channel = desc->nr_channels - 1;
 137                 else
 138                         extra_channel = 0;
 139         } else
 140                 return true;
 141
 142         for (i = 0; i < 4; ++i) {
 143                 int index = desc->swizzle[i] - PIPE_SWIZZLE_X;
 144
 145                 if (desc->swizzle[i] < PIPE_SWIZZLE_X ||
 146                     desc->swizzle[i] > PIPE_SWIZZLE_W)
 147                         continue;
 148
 149                 if (desc->channel[i].pure_integer &&
 150                     desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
 151                         /* Use the maximum value for clamping the clear color. */
 152                         int max = u_bit_consecutive(0, desc->channel[i].size - 1);
 153
 154                         values[i] = color->i[i] != 0;
 155                         if (color->i[i] != 0 && MIN2(color->i[i], max) != max)
 156                                 return true;
 157                 } else if (desc->channel[i].pure_integer &&
 158                            desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
 159                         /* Use the maximum value for clamping the clear color. */
 160                         unsigned max = u_bit_consecutive(0, desc->channel[i].size);
 161
 162                         values[i] = color->ui[i] != 0U;
 163                         if (color->ui[i] != 0U && MIN2(color->ui[i], max) != max)
 164                                 return true;
 165                 } else {
 166                         values[i] = color->f[i] != 0.0F;
 167                         if (color->f[i] != 0.0F && color->f[i] != 1.0F)
 168                                 return true;
 169                 }
 170
 171                 if (index == extra_channel)
 172                         extra_value = values[i];
 173                 else
 174                         main_value = values[i];
 175         }
 176
 177         for (int i = 0; i < 4; ++i)
 178                 if (values[i] != main_value &&
 179                     desc->swizzle[i] - PIPE_SWIZZLE_X != extra_channel &&
 180                     desc->swizzle[i] >= PIPE_SWIZZLE_X &&
 181                     desc->swizzle[i] <= PIPE_SWIZZLE_W)
 182                         return true;
 183
 184         *clear_words_needed = false;
 185         if (main_value)
 186                 *reset_value |= 0x80808080U;
 187
 188         if (extra_value)
 189                 *reset_value |= 0x40404040U;
 190         return true;
 191 }
 192
 193 void vi_dcc_clear_level(struct si_context *sctx,
 194                         struct r600_texture *rtex,
 195                         unsigned level, unsigned clear_value)
 196 {
 197         struct pipe_resource *dcc_buffer;
 198         uint64_t dcc_offset, clear_size;
 199
 200         assert(vi_dcc_enabled(rtex, level));
 201
 202         if (rtex->dcc_separate_buffer) {
 203                 dcc_buffer = &rtex->dcc_separate_buffer->b.b;
 204                 dcc_offset = 0;
 205         } else {
 206                 dcc_buffer = &rtex->resource.b.b;
 207                 dcc_offset = rtex->dcc_offset;
 208         }
 209
 210         if (sctx->b.chip_class >= GFX9) {
 211                 /* Mipmap level clears aren't implemented. */
 212                 assert(rtex->resource.b.b.last_level == 0);
 213                 /* MSAA needs a different clear size. */
 214                 assert(rtex->resource.b.b.nr_samples <= 1);
 215                 clear_size = rtex->surface.dcc_size;
 216         } else {
 217                 unsigned num_layers = util_num_layers(&rtex->resource.b.b, level);
 218
 219                 /* If this is 0, fast clear isn't possible. (can occur with MSAA) */
 220                 assert(rtex->surface.u.legacy.level[level].dcc_fast_clear_size);
 221                 /* Layered MSAA DCC fast clears need to clear dcc_fast_clear_size
 222                  * bytes for each layer. This is not currently implemented, and
 223                  * therefore MSAA DCC isn't even enabled with multiple layers.
 224                  */
 225                 assert(rtex->resource.b.b.nr_samples <= 1 || num_layers == 1);
 226
 227                 dcc_offset += rtex->surface.u.legacy.level[level].dcc_offset;
 228                 clear_size = rtex->surface.u.legacy.level[level].dcc_fast_clear_size *
 229                              num_layers;
 230         }
 231
 232         si_clear_buffer(&sctx->b.b, dcc_buffer, dcc_offset, clear_size,
 233                         clear_value, R600_COHERENCY_CB_META);
 234 }
 235
 236 /* Set the same micro tile mode as the destination of the last MSAA resolve.
 237  * This allows hitting the MSAA resolve fast path, which requires that both
 238  * src and dst micro tile modes match.
 239  */
 240 static void si_set_optimal_micro_tile_mode(struct si_screen *sscreen,
 241                                            struct r600_texture *rtex)
 242 {
 243         if (rtex->resource.b.is_shared ||
 244             rtex->resource.b.b.nr_samples <= 1 ||
 245             rtex->surface.micro_tile_mode == rtex->last_msaa_resolve_target_micro_mode)
 246                 return;
 247
 248         assert(sscreen->info.chip_class >= GFX9 ||
 249                rtex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
 250         assert(rtex->resource.b.b.last_level == 0);
 251
 252         if (sscreen->info.chip_class >= GFX9) {
 253                 /* 4K or larger tiles only. 0 is linear. 1-3 are 256B tiles. */
 254                 assert(rtex->surface.u.gfx9.surf.swizzle_mode >= 4);
 255
 256                 /* If you do swizzle_mode % 4, you'll get:
 257                  *   0 = Depth
 258                  *   1 = Standard,
 259                  *   2 = Displayable
 260                  *   3 = Rotated
 261                  *
 262                  * Depth-sample order isn't allowed:
 263                  */
 264                 assert(rtex->surface.u.gfx9.surf.swizzle_mode % 4 != 0);
 265
 266                 switch (rtex->last_msaa_resolve_target_micro_mode) {
 267                 case RADEON_MICRO_MODE_DISPLAY:
 268                         rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
 269                         rtex->surface.u.gfx9.surf.swizzle_mode += 2; /* D */
 270                         break;
 271                 case RADEON_MICRO_MODE_THIN:
 272                         rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
 273                         rtex->surface.u.gfx9.surf.swizzle_mode += 1; /* S */
 274                         break;
 275                 case RADEON_MICRO_MODE_ROTATED:
 276                         rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
 277                         rtex->surface.u.gfx9.surf.swizzle_mode += 3; /* R */
 278                         break;
 279                 default: /* depth */
 280                         assert(!"unexpected micro mode");
 281                         return;
 282                 }
 283         } else if (sscreen->info.chip_class >= CIK) {
 284                 /* These magic numbers were copied from addrlib. It doesn't use
 285                  * any definitions for them either. They are all 2D_TILED_THIN1
 286                  * modes with different bpp and micro tile mode.
 287                  */
 288                 switch (rtex->last_msaa_resolve_target_micro_mode) {
 289                 case RADEON_MICRO_MODE_DISPLAY:
 290                         rtex->surface.u.legacy.tiling_index[0] = 10;
 291                         break;
 292                 case RADEON_MICRO_MODE_THIN:
 293                         rtex->surface.u.legacy.tiling_index[0] = 14;
 294                         break;
 295                 case RADEON_MICRO_MODE_ROTATED:
 296                         rtex->surface.u.legacy.tiling_index[0] = 28;
 297                         break;
 298                 default: /* depth, thick */
 299                         assert(!"unexpected micro mode");
 300                         return;
 301                 }
 302         } else { /* SI */
 303                 switch (rtex->last_msaa_resolve_target_micro_mode) {
 304                 case RADEON_MICRO_MODE_DISPLAY:
 305                         switch (rtex->surface.bpe) {
 306                         case 1:
 307                             rtex->surface.u.legacy.tiling_index[0] = 10;
 308                             break;
 309                         case 2:
 310                             rtex->surface.u.legacy.tiling_index[0] = 11;
 311                             break;
 312                         default: /* 4, 8 */
 313                             rtex->surface.u.legacy.tiling_index[0] = 12;
 314                             break;
 315                         }
 316                         break;
 317                 case RADEON_MICRO_MODE_THIN:
 318                         switch (rtex->surface.bpe) {
 319                         case 1:
 320                                 rtex->surface.u.legacy.tiling_index[0] = 14;
 321                                 break;
 322                         case 2:
 323                                 rtex->surface.u.legacy.tiling_index[0] = 15;
 324                                 break;
 325                         case 4:
 326                                 rtex->surface.u.legacy.tiling_index[0] = 16;
 327                                 break;
 328                         default: /* 8, 16 */
 329                                 rtex->surface.u.legacy.tiling_index[0] = 17;
 330                                 break;
 331                         }
 332                         break;
 333                 default: /* depth, thick */
 334                         assert(!"unexpected micro mode");
 335                         return;
 336                 }
 337         }
 338
 339         rtex->surface.micro_tile_mode = rtex->last_msaa_resolve_target_micro_mode;
 340
 341         p_atomic_inc(&sscreen->dirty_tex_counter);
 342 }
 343
 344 static void si_do_fast_color_clear(struct si_context *sctx,
 345                                    unsigned *buffers,
 346                                    const union pipe_color_union *color)
 347 {
 348         struct pipe_framebuffer_state *fb = &sctx->framebuffer.state;
 349         int i;
 350
 351         /* This function is broken in BE, so just disable this path for now */
 352 #ifdef PIPE_ARCH_BIG_ENDIAN
 353         return;
 354 #endif
 355
 356         if (sctx->b.render_cond)
 357                 return;
 358
 359         for (i = 0; i < fb->nr_cbufs; i++) {
 360                 struct r600_texture *tex;
 361                 unsigned clear_bit = PIPE_CLEAR_COLOR0 << i;
 362
 363                 if (!fb->cbufs[i])
 364                         continue;
 365
 366                 /* if this colorbuffer is not being cleared */
 367                 if (!(*buffers & clear_bit))
 368                         continue;
 369
 370                 unsigned level = fb->cbufs[i]->u.tex.level;
 371                 tex = (struct r600_texture *)fb->cbufs[i]->texture;
 372
 373                 /* the clear is allowed if all layers are bound */
 374                 if (fb->cbufs[i]->u.tex.first_layer != 0 ||
 375                     fb->cbufs[i]->u.tex.last_layer != util_max_layer(&tex->resource.b.b, 0)) {
 376                         continue;
 377                 }
 378
 379                 /* cannot clear mipmapped textures */
 380                 if (fb->cbufs[i]->texture->last_level != 0) {
 381                         continue;
 382                 }
 383
 384                 /* only supported on tiled surfaces */
 385                 if (tex->surface.is_linear) {
 386                         continue;
 387                 }
 388
 389                 /* shared textures can't use fast clear without an explicit flush,
 390                  * because there is no way to communicate the clear color among
 391                  * all clients
 392                  */
 393                 if (tex->resource.b.is_shared &&
 394                     !(tex->resource.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
 395                         continue;
 396
 397                 /* fast color clear with 1D tiling doesn't work on old kernels and CIK */
 398                 if (sctx->b.chip_class == CIK &&
 399                     tex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_1D &&
 400                     sctx->screen->info.drm_major == 2 &&
 401                     sctx->screen->info.drm_minor < 38) {
 402                         continue;
 403                 }
 404
 405                 /* Fast clear is the most appropriate place to enable DCC for
 406                  * displayable surfaces.
 407                  */
 408                 if (sctx->b.chip_class >= VI &&
 409                     !(sctx->screen->debug_flags & DBG(NO_DCC_FB))) {
 410                         vi_separate_dcc_try_enable(sctx, tex);
 411
 412                         /* RB+ isn't supported with a CMASK clear only on Stoney,
 413                          * so all clears are considered to be hypothetically slow
 414                          * clears, which is weighed when determining whether to
 415                          * enable separate DCC.
 416                          */
 417                         if (tex->dcc_gather_statistics &&
 418                             sctx->b.family == CHIP_STONEY)
 419                                 tex->num_slow_clears++;
 420                 }
 421
 422                 bool need_decompress_pass = false;
 423
 424                 /* Use a slow clear for small surfaces where the cost of
 425                  * the eliminate pass can be higher than the benefit of fast
 426                  * clear. The closed driver does this, but the numbers may differ.
 427                  *
 428                  * This helps on both dGPUs and APUs, even small APUs like Mullins.
 429                  */
 430                 bool too_small = tex->resource.b.b.nr_samples <= 1 &&
 431                                  tex->resource.b.b.width0 *
 432                                  tex->resource.b.b.height0 <= 512 * 512;
 433
 434                 /* Try to clear DCC first, otherwise try CMASK. */
 435                 if (vi_dcc_enabled(tex, 0)) {
 436                         uint32_t reset_value;
 437                         bool clear_words_needed;
 438
 439                         if (sctx->screen->debug_flags & DBG(NO_DCC_CLEAR))
 440                                 continue;
 441
 442                         /* This can only occur with MSAA. */
 443                         if (sctx->b.chip_class == VI &&
 444                             !tex->surface.u.legacy.level[level].dcc_fast_clear_size)
 445                                 continue;
 446
 447                         if (!vi_get_fast_clear_parameters(fb->cbufs[i]->format,
 448                                                           color, &reset_value,
 449                                                           &clear_words_needed))
 450                                 continue;
 451
 452                         if (clear_words_needed && too_small)
 453                                 continue;
 454
 455                         /* DCC fast clear with MSAA should clear CMASK to 0xC. */
 456                         if (tex->resource.b.b.nr_samples >= 2 && tex->cmask.size) {
 457                                 /* TODO: This doesn't work with MSAA. */
 458                                 if (clear_words_needed)
 459                                         continue;
 460
 461                                 si_clear_buffer(&sctx->b.b, &tex->cmask_buffer->b.b,
 462                                                 tex->cmask.offset, tex->cmask.size,
 463                                                 0xCCCCCCCC, R600_COHERENCY_CB_META);
 464                                 need_decompress_pass = true;
 465                         }
 466
 467                         vi_dcc_clear_level(sctx, tex, 0, reset_value);
 468
 469                         if (clear_words_needed)
 470                                 need_decompress_pass = true;
 471
 472                         tex->separate_dcc_dirty = true;
 473                 } else {
 474                         if (too_small)
 475                                 continue;
 476
 477                         /* 128-bit formats are unusupported */
 478                         if (tex->surface.bpe > 8) {
 479                                 continue;
 480                         }
 481
 482                         /* RB+ doesn't work with CMASK fast clear on Stoney. */
 483                         if (sctx->b.family == CHIP_STONEY)
 484                                 continue;
 485
 486                         /* ensure CMASK is enabled */
 487                         si_alloc_separate_cmask(sctx->screen, tex);
 488                         if (tex->cmask.size == 0) {
 489                                 continue;
 490                         }
 491
 492                         /* Do the fast clear. */
 493                         si_clear_buffer(&sctx->b.b, &tex->cmask_buffer->b.b,
 494                                         tex->cmask.offset, tex->cmask.size, 0,
 495                                         R600_COHERENCY_CB_META);
 496                         need_decompress_pass = true;
 497                 }
 498
 499                 if (need_decompress_pass &&
 500                     !(tex->dirty_level_mask & (1 << level))) {
 501                         tex->dirty_level_mask |= 1 << level;
 502                         p_atomic_inc(&sctx->screen->compressed_colortex_counter);
 503                 }
 504
 505                 /* We can change the micro tile mode before a full clear. */
 506                 si_set_optimal_micro_tile_mode(sctx->screen, tex);
 507
 508                 si_set_clear_color(tex, fb->cbufs[i]->format, color);
 509
 510                 sctx->framebuffer.dirty_cbufs |= 1 << i;
 511                 si_mark_atom_dirty(sctx, &sctx->framebuffer.atom);
 512                 *buffers &= ~clear_bit;
 513         }
 514 }
 515
 516 static void si_clear(struct pipe_context *ctx, unsigned buffers,
 517                      const union pipe_color_union *color,
 518                      double depth, unsigned stencil)
 519 {
 520         struct si_context *sctx = (struct si_context *)ctx;
 521         struct pipe_framebuffer_state *fb = &sctx->framebuffer.state;
 522         struct pipe_surface *zsbuf = fb->zsbuf;
 523         struct r600_texture *zstex =
 524                 zsbuf ? (struct r600_texture*)zsbuf->texture : NULL;
 525
 526         if (buffers & PIPE_CLEAR_COLOR) {
 527                 si_do_fast_color_clear(sctx, &buffers, color);
 528                 if (!buffers)
 529                         return; /* all buffers have been fast cleared */
 530         }
 531
 532         if (buffers & PIPE_CLEAR_COLOR) {
 533                 int i;
 534
 535                 /* These buffers cannot use fast clear, make sure to disable expansion. */
 536                 for (i = 0; i < fb->nr_cbufs; i++) {
 537                         struct r600_texture *tex;
 538
 539                         /* If not clearing this buffer, skip. */
 540                         if (!(buffers & (PIPE_CLEAR_COLOR0 << i)))
 541                                 continue;
 542
 543                         if (!fb->cbufs[i])
 544                                 continue;
 545
 546                         tex = (struct r600_texture *)fb->cbufs[i]->texture;
 547                         if (tex->fmask.size == 0)
 548                                 tex->dirty_level_mask &= ~(1 << fb->cbufs[i]->u.tex.level);
 549                 }
 550         }
 551
 552         if (zstex &&
 553             si_htile_enabled(zstex, zsbuf->u.tex.level) &&
 554             zsbuf->u.tex.first_layer == 0 &&
 555             zsbuf->u.tex.last_layer == util_max_layer(&zstex->resource.b.b, 0)) {
 556                 /* TC-compatible HTILE only supports depth clears to 0 or 1. */
 557                 if (buffers & PIPE_CLEAR_DEPTH &&
 558                     (!zstex->tc_compatible_htile ||
 559                      depth == 0 || depth == 1)) {
 560                         /* Need to disable EXPCLEAR temporarily if clearing
 561                          * to a new value. */
 562                         if (!zstex->depth_cleared || zstex->depth_clear_value != depth) {
 563                                 sctx->db_depth_disable_expclear = true;
 564                         }
 565
 566                         zstex->depth_clear_value = depth;
 567                         sctx->framebuffer.dirty_zsbuf = true;
 568                         si_mark_atom_dirty(sctx, &sctx->framebuffer.atom); /* updates DB_DEPTH_CLEAR */
 569                         sctx->db_depth_clear = true;
 570                         si_mark_atom_dirty(sctx, &sctx->db_render_state);
 571                 }
 572
 573                 /* TC-compatible HTILE only supports stencil clears to 0. */
 574                 if (buffers & PIPE_CLEAR_STENCIL &&
 575                     (!zstex->tc_compatible_htile || stencil == 0)) {
 576                         stencil &= 0xff;
 577
 578                         /* Need to disable EXPCLEAR temporarily if clearing
 579                          * to a new value. */
 580                         if (!zstex->stencil_cleared || zstex->stencil_clear_value != stencil) {
 581                                 sctx->db_stencil_disable_expclear = true;
 582                         }
 583
 584                         zstex->stencil_clear_value = stencil;
 585                         sctx->framebuffer.dirty_zsbuf = true;
 586                         si_mark_atom_dirty(sctx, &sctx->framebuffer.atom); /* updates DB_STENCIL_CLEAR */
 587                         sctx->db_stencil_clear = true;
 588                         si_mark_atom_dirty(sctx, &sctx->db_render_state);
 589                 }
 590
 591                 /* TODO: Find out what's wrong here. Fast depth clear leads to
 592                  * corruption in ARK: Survival Evolved, but that may just be
 593                  * a coincidence and the root cause is elsewhere.
 594                  *
 595                  * The corruption can be fixed by putting the DB flush before
 596                  * or after the depth clear. (surprisingly)
 597                  *
 598                  * https://bugs.freedesktop.org/show_bug.cgi?id=102955 (apitrace)
 599                  *
 600                  * This hack decreases back-to-back ClearDepth performance.
 601                  */
 602                 if (sctx->screen->clear_db_cache_before_clear) {
 603                         sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_DB;
 604                 }
 605         }
 606
 607         si_blitter_begin(ctx, SI_CLEAR);
 608         util_blitter_clear(sctx->blitter, fb->width, fb->height,
 609                            util_framebuffer_get_num_layers(fb),
 610                            buffers, color, depth, stencil);
 611         si_blitter_end(ctx);
 612
 613         if (sctx->db_depth_clear) {
 614                 sctx->db_depth_clear = false;
 615                 sctx->db_depth_disable_expclear = false;
 616                 zstex->depth_cleared = true;
 617                 si_mark_atom_dirty(sctx, &sctx->db_render_state);
 618         }
 619
 620         if (sctx->db_stencil_clear) {
 621                 sctx->db_stencil_clear = false;
 622                 sctx->db_stencil_disable_expclear = false;
 623                 zstex->stencil_cleared = true;
 624                 si_mark_atom_dirty(sctx, &sctx->db_render_state);
 625         }
 626 }
 627
 628 static void si_clear_render_target(struct pipe_context *ctx,
 629                                    struct pipe_surface *dst,
 630                                    const union pipe_color_union *color,
 631                                    unsigned dstx, unsigned dsty,
 632                                    unsigned width, unsigned height,
 633                                    bool render_condition_enabled)
 634 {
 635         struct si_context *sctx = (struct si_context *)ctx;
 636
 637         si_blitter_begin(ctx, SI_CLEAR_SURFACE |
 638                          (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
 639         util_blitter_clear_render_target(sctx->blitter, dst, color,
 640                                          dstx, dsty, width, height);
 641         si_blitter_end(ctx);
 642 }
 643
 644 static void si_clear_depth_stencil(struct pipe_context *ctx,
 645                                    struct pipe_surface *dst,
 646                                    unsigned clear_flags,
 647                                    double depth,
 648                                    unsigned stencil,
 649                                    unsigned dstx, unsigned dsty,
 650                                    unsigned width, unsigned height,
 651                                    bool render_condition_enabled)
 652 {
 653         struct si_context *sctx = (struct si_context *)ctx;
 654
 655         si_blitter_begin(ctx, SI_CLEAR_SURFACE |
 656                          (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
 657         util_blitter_clear_depth_stencil(sctx->blitter, dst, clear_flags, depth, stencil,
 658                                          dstx, dsty, width, height);
 659         si_blitter_end(ctx);
 660 }
 661
 662 static void si_clear_texture(struct pipe_context *pipe,
 663                              struct pipe_resource *tex,
 664                              unsigned level,
 665                              const struct pipe_box *box,
 666                              const void *data)
 667 {
 668         struct pipe_screen *screen = pipe->screen;
 669         struct r600_texture *rtex = (struct r600_texture*)tex;
 670         struct pipe_surface tmpl = {{0}};
 671         struct pipe_surface *sf;
 672         const struct util_format_description *desc =
 673                 util_format_description(tex->format);
 674
 675         tmpl.format = tex->format;
 676         tmpl.u.tex.first_layer = box->z;
 677         tmpl.u.tex.last_layer = box->z + box->depth - 1;
 678         tmpl.u.tex.level = level;
 679         sf = pipe->create_surface(pipe, tex, &tmpl);
 680         if (!sf)
 681                 return;
 682
 683         if (rtex->is_depth) {
 684                 unsigned clear;
 685                 float depth;
 686                 uint8_t stencil = 0;
 687
 688                 /* Depth is always present. */
 689                 clear = PIPE_CLEAR_DEPTH;
 690                 desc->unpack_z_float(&depth, 0, data, 0, 1, 1);
 691
 692                 if (rtex->surface.has_stencil) {
 693                         clear |= PIPE_CLEAR_STENCIL;
 694                         desc->unpack_s_8uint(&stencil, 0, data, 0, 1, 1);
 695                 }
 696
 697                 si_clear_depth_stencil(pipe, sf, clear, depth, stencil,
 698                                        box->x, box->y,
 699                                        box->width, box->height, false);
 700         } else {
 701                 union pipe_color_union color;
 702
 703                 /* pipe_color_union requires the full vec4 representation. */
 704                 if (util_format_is_pure_uint(tex->format))
 705                         desc->unpack_rgba_uint(color.ui, 0, data, 0, 1, 1);
 706                 else if (util_format_is_pure_sint(tex->format))
 707                         desc->unpack_rgba_sint(color.i, 0, data, 0, 1, 1);
 708                 else
 709                         desc->unpack_rgba_float(color.f, 0, data, 0, 1, 1);
 710
 711                 if (screen->is_format_supported(screen, tex->format,
 712                                                 tex->target, 0,
 713                                                 PIPE_BIND_RENDER_TARGET)) {
 714                         si_clear_render_target(pipe, sf, &color,
 715                                                box->x, box->y,
 716                                                box->width, box->height, false);
 717                 } else {
 718                         /* Software fallback - just for R9G9B9E5_FLOAT */
 719                         util_clear_render_target(pipe, sf, &color,
 720                                                  box->x, box->y,
 721                                                  box->width, box->height);
 722                 }
 723         }
 724         pipe_surface_reference(&sf, NULL);
 725 }
 726
 727 void si_init_clear_functions(struct si_context *sctx)
 728 {
 729         sctx->b.b.clear = si_clear;
 730         sctx->b.b.clear_render_target = si_clear_render_target;
 731         sctx->b.b.clear_depth_stencil = si_clear_depth_stencil;
 732         sctx->b.b.clear_texture = si_clear_texture;
 733 }