coarse.comp raw

   1  // SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense
   2  
   3  // The coarse rasterizer stage of the pipeline.
   4  //
   5  // As input we have the ordered partitions of paths from the binning phase and
   6  // the annotated tile list of segments and backdrop per path.
   7  //
   8  // Each workgroup operating on one bin by stream compacting
   9  // the elements corresponding to the bin.
  10  //
  11  // As output we have an ordered command stream per tile. Every tile from a path (backdrop + segment list) will be encoded.
  12  
  13  #version 450
  14  #extension GL_GOOGLE_include_directive : enable
  15  
  16  #include "mem.h"
  17  #include "setup.h"
  18  
  19  layout(local_size_x = N_TILE, local_size_y = 1) in;
  20  
  21  layout(set = 0, binding = 1) readonly buffer ConfigBuf {
  22      Config conf;
  23  };
  24  
  25  #include "annotated.h"
  26  #include "bins.h"
  27  #include "tile.h"
  28  #include "ptcl.h"
  29  
  30  #define LG_N_PART_READ (7 + LG_WG_FACTOR)
  31  #define N_PART_READ (1 << LG_N_PART_READ)
  32  
  33  shared uint sh_elements[N_TILE];
  34  
  35  // Number of elements in the partition; prefix sum.
  36  shared uint sh_part_count[N_PART_READ];
  37  shared Alloc sh_part_elements[N_PART_READ];
  38  
  39  shared uint sh_bitmaps[N_SLICE][N_TILE];
  40  
  41  shared uint sh_tile_count[N_TILE];
  42  // The width of the tile rect for the element, intersected with this bin
  43  shared uint sh_tile_width[N_TILE];
  44  shared uint sh_tile_x0[N_TILE];
  45  shared uint sh_tile_y0[N_TILE];
  46  
  47  // These are set up so base + tile_y * stride + tile_x points to a Tile.
  48  shared uint sh_tile_base[N_TILE];
  49  shared uint sh_tile_stride[N_TILE];
  50  
  51  #ifdef MEM_DEBUG
  52  // Store allocs only when MEM_DEBUG to save shared memory traffic.
  53  shared Alloc sh_tile_alloc[N_TILE];
  54  
  55  void write_tile_alloc(uint el_ix, Alloc a) {
  56      sh_tile_alloc[el_ix] = a;
  57  }
  58  
  59  Alloc read_tile_alloc(uint el_ix, bool mem_ok) {
  60      return sh_tile_alloc[el_ix];
  61  }
  62  #else
  63  void write_tile_alloc(uint el_ix, Alloc a) {
  64      // No-op
  65  }
  66  
  67  Alloc read_tile_alloc(uint el_ix, bool mem_ok) {
  68      // All memory.
  69      return new_alloc(0, memory.length()*4, mem_ok);
  70  }
  71  #endif
  72  
  73  // The maximum number of commands per annotated element.
  74  #define ANNO_COMMANDS 2
  75  
  76  // Perhaps cmd_alloc should be a global? This is a style question.
  77  bool alloc_cmd(inout Alloc cmd_alloc, inout CmdRef cmd_ref, inout uint cmd_limit) {
  78      if (cmd_ref.offset < cmd_limit) {
  79          return true;
  80      }
  81      MallocResult new_cmd = malloc(PTCL_INITIAL_ALLOC);
  82      if (new_cmd.failed) {
  83          return false;
  84      }
  85      CmdJump jump = CmdJump(new_cmd.alloc.offset);
  86      Cmd_Jump_write(cmd_alloc, cmd_ref, jump);
  87      cmd_alloc = new_cmd.alloc;
  88      cmd_ref = CmdRef(cmd_alloc.offset);
  89      // Reserve space for the maximum number of commands and a potential jump.
  90      cmd_limit = cmd_alloc.offset + PTCL_INITIAL_ALLOC - (ANNO_COMMANDS + 1) * Cmd_size;
  91      return true;
  92  }
  93  
  94  void write_fill(Alloc alloc, inout CmdRef cmd_ref, uint flags, Tile tile, float linewidth) {
  95      if (fill_mode_from_flags(flags) == MODE_NONZERO) {
  96          if (tile.tile.offset != 0) {
  97              CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
  98              Cmd_Fill_write(alloc, cmd_ref, cmd_fill);
  99              cmd_ref.offset += 4 + CmdFill_size;
 100          } else {
 101              Cmd_Solid_write(alloc, cmd_ref);
 102              cmd_ref.offset += 4;
 103          }
 104      } else {
 105          CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * linewidth);
 106          Cmd_Stroke_write(alloc, cmd_ref, cmd_stroke);
 107          cmd_ref.offset += 4 + CmdStroke_size;
 108      }
 109  }
 110  
 111  void main() {
 112      // Could use either linear or 2d layouts for both dispatch and
 113      // invocations within the workgroup. We'll use variables to abstract.
 114      uint width_in_bins = (conf.width_in_tiles + N_TILE_X - 1)/N_TILE_X;
 115      uint bin_ix = width_in_bins * gl_WorkGroupID.y + gl_WorkGroupID.x;
 116      uint partition_ix = 0;
 117      uint n_partitions = (conf.n_elements + N_TILE - 1) / N_TILE;
 118      uint th_ix = gl_LocalInvocationID.x;
 119  
 120      // Coordinates of top left of bin, in tiles.
 121      uint bin_tile_x = N_TILE_X * gl_WorkGroupID.x;
 122      uint bin_tile_y = N_TILE_Y * gl_WorkGroupID.y;
 123  
 124      // Per-tile state
 125      uint tile_x = gl_LocalInvocationID.x % N_TILE_X;
 126      uint tile_y = gl_LocalInvocationID.x / N_TILE_X;
 127      uint this_tile_ix = (bin_tile_y + tile_y) * conf.width_in_tiles + bin_tile_x + tile_x;
 128      Alloc cmd_alloc = slice_mem(conf.ptcl_alloc, this_tile_ix * PTCL_INITIAL_ALLOC, PTCL_INITIAL_ALLOC);
 129      CmdRef cmd_ref = CmdRef(cmd_alloc.offset);
 130      // Reserve space for the maximum number of commands and a potential jump.
 131      uint cmd_limit = cmd_ref.offset + PTCL_INITIAL_ALLOC - (ANNO_COMMANDS + 1) * Cmd_size;
 132      // The nesting depth of the clip stack
 133      uint clip_depth = 0;
 134      // State for the "clip zero" optimization. If it's nonzero, then we are
 135      // currently in a clip for which the entire tile has an alpha of zero, and
 136      // the value is the depth after the "begin clip" of that element.
 137      uint clip_zero_depth = 0;
 138      // State for the "clip one" optimization. If bit `i` is set, then that means
 139      // that the clip pushed at depth `i` has an alpha of all one.
 140      uint clip_one_mask = 0;
 141  
 142      // I'm sure we can figure out how to do this with at least one fewer register...
 143      // Items up to rd_ix have been read from sh_elements
 144      uint rd_ix = 0;
 145      // Items up to wr_ix have been written into sh_elements
 146      uint wr_ix = 0;
 147      // Items between part_start_ix and ready_ix are ready to be transferred from sh_part_elements
 148      uint part_start_ix = 0;
 149      uint ready_ix = 0;
 150  
 151      // Leave room for the fine rasterizer scratch allocation.
 152      Alloc scratch_alloc = slice_mem(cmd_alloc, 0, Alloc_size);
 153      cmd_ref.offset += Alloc_size;
 154  
 155      uint num_begin_slots = 0;
 156      uint begin_slot = 0;
 157      bool mem_ok = mem_error == NO_ERROR;
 158      while (true) {
 159          for (uint i = 0; i < N_SLICE; i++) {
 160              sh_bitmaps[i][th_ix] = 0;
 161          }
 162  
 163          // parallel read of input partitions
 164          do {
 165              if (ready_ix == wr_ix && partition_ix < n_partitions) {
 166                  part_start_ix = ready_ix;
 167                  uint count = 0;
 168                  if (th_ix < N_PART_READ && partition_ix + th_ix < n_partitions) {
 169                      uint in_ix = (conf.bin_alloc.offset >> 2) + ((partition_ix + th_ix) * N_TILE + bin_ix) * 2;
 170                      count = read_mem(conf.bin_alloc, in_ix);
 171                      uint offset = read_mem(conf.bin_alloc, in_ix + 1);
 172                      sh_part_elements[th_ix] = new_alloc(offset, count*BinInstance_size, mem_ok);
 173                  }
 174                  // prefix sum of counts
 175                  for (uint i = 0; i < LG_N_PART_READ; i++) {
 176                      if (th_ix < N_PART_READ) {
 177                          sh_part_count[th_ix] = count;
 178                      }
 179                      barrier();
 180                      if (th_ix < N_PART_READ) {
 181                          if (th_ix >= (1 << i)) {
 182                              count += sh_part_count[th_ix - (1 << i)];
 183                          }
 184                      }
 185                      barrier();
 186                  }
 187                  if (th_ix < N_PART_READ) {
 188                      sh_part_count[th_ix] = part_start_ix + count;
 189                  }
 190                  barrier();
 191                  ready_ix = sh_part_count[N_PART_READ - 1];
 192                  partition_ix += N_PART_READ;
 193              }
 194              // use binary search to find element to read
 195              uint ix = rd_ix + th_ix;
 196              if (ix >= wr_ix && ix < ready_ix && mem_ok) {
 197                  uint part_ix = 0;
 198                  for (uint i = 0; i < LG_N_PART_READ; i++) {
 199                      uint probe = part_ix + ((N_PART_READ / 2) >> i);
 200                      if (ix >= sh_part_count[probe - 1]) {
 201                          part_ix = probe;
 202                      }
 203                  }
 204                  ix -= part_ix > 0 ? sh_part_count[part_ix - 1] : part_start_ix;
 205                  Alloc bin_alloc = sh_part_elements[part_ix];
 206                  BinInstanceRef inst_ref = BinInstanceRef(bin_alloc.offset);
 207                  BinInstance inst = BinInstance_read(bin_alloc, BinInstance_index(inst_ref, ix));
 208                  sh_elements[th_ix] = inst.element_ix;
 209              }
 210              barrier();
 211  
 212              wr_ix = min(rd_ix + N_TILE, ready_ix);
 213          } while (wr_ix - rd_ix < N_TILE && (wr_ix < ready_ix || partition_ix < n_partitions));
 214  
 215          // We've done the merge and filled the buffer.
 216  
 217          // Read one element, compute coverage.
 218          uint tag = Annotated_Nop;
 219          uint element_ix;
 220          AnnotatedRef ref;
 221          if (th_ix + rd_ix < wr_ix) {
 222              element_ix = sh_elements[th_ix];
 223              ref = AnnotatedRef(conf.anno_alloc.offset + element_ix * Annotated_size);
 224              tag = Annotated_tag(conf.anno_alloc, ref).tag;
 225          }
 226  
 227          // Bounding box of element in pixel coordinates.
 228          uint tile_count;
 229          switch (tag) {
 230          case Annotated_Color:
 231          case Annotated_Image:
 232          case Annotated_BeginClip:
 233          case Annotated_EndClip:
 234              // We have one "path" for each element, even if the element isn't
 235              // actually a path (currently EndClip, but images etc in the future).
 236              uint path_ix = element_ix;
 237              Path path = Path_read(conf.tile_alloc, PathRef(conf.tile_alloc.offset + path_ix * Path_size));
 238              uint stride = path.bbox.z - path.bbox.x;
 239              sh_tile_stride[th_ix] = stride;
 240              int dx = int(path.bbox.x) - int(bin_tile_x);
 241              int dy = int(path.bbox.y) - int(bin_tile_y);
 242              int x0 = clamp(dx, 0, N_TILE_X);
 243              int y0 = clamp(dy, 0, N_TILE_Y);
 244              int x1 = clamp(int(path.bbox.z) - int(bin_tile_x), 0, N_TILE_X);
 245              int y1 = clamp(int(path.bbox.w) - int(bin_tile_y), 0, N_TILE_Y);
 246              sh_tile_width[th_ix] = uint(x1 - x0);
 247              sh_tile_x0[th_ix] = x0;
 248              sh_tile_y0[th_ix] = y0;
 249              tile_count = uint(x1 - x0) * uint(y1 - y0);
 250              // base relative to bin
 251              uint base = path.tiles.offset - uint(dy * stride + dx) * Tile_size;
 252              sh_tile_base[th_ix] = base;
 253              Alloc path_alloc = new_alloc(path.tiles.offset, (path.bbox.z - path.bbox.x) * (path.bbox.w - path.bbox.y) * Tile_size, mem_ok);
 254              write_tile_alloc(th_ix, path_alloc);
 255              break;
 256          default:
 257              tile_count = 0;
 258              break;
 259          }
 260  
 261          // Prefix sum of sh_tile_count
 262          sh_tile_count[th_ix] = tile_count;
 263          for (uint i = 0; i < LG_N_TILE; i++) {
 264              barrier();
 265              if (th_ix >= (1 << i)) {
 266                  tile_count += sh_tile_count[th_ix - (1 << i)];
 267              }
 268              barrier();
 269              sh_tile_count[th_ix] = tile_count;
 270          }
 271          barrier();
 272          uint total_tile_count = sh_tile_count[N_TILE - 1];
 273          for (uint ix = th_ix; ix < total_tile_count; ix += N_TILE) {
 274              // Binary search to find element
 275              uint el_ix = 0;
 276              for (uint i = 0; i < LG_N_TILE; i++) {
 277                  uint probe = el_ix + ((N_TILE / 2) >> i);
 278                  if (ix >= sh_tile_count[probe - 1]) {
 279                      el_ix = probe;
 280                  }
 281              }
 282              AnnotatedRef ref = AnnotatedRef(conf.anno_alloc.offset + sh_elements[el_ix] * Annotated_size);
 283              uint tag = Annotated_tag(conf.anno_alloc, ref).tag;
 284              uint seq_ix = ix - (el_ix > 0 ? sh_tile_count[el_ix - 1] : 0);
 285              uint width = sh_tile_width[el_ix];
 286              uint x = sh_tile_x0[el_ix] + seq_ix % width;
 287              uint y = sh_tile_y0[el_ix] + seq_ix / width;
 288              bool include_tile = false;
 289              if (tag == Annotated_BeginClip || tag == Annotated_EndClip) {
 290                  include_tile = true;
 291              } else if (mem_ok) {
 292                  Tile tile = Tile_read(read_tile_alloc(el_ix, mem_ok), TileRef(sh_tile_base[el_ix] + (sh_tile_stride[el_ix] * y + x) * Tile_size));
 293                  // Include the path in the tile if
 294                  // - the tile contains at least a segment (tile offset non-zero)
 295                  // - the tile is completely covered (backdrop non-zero)
 296                  include_tile = tile.tile.offset != 0 || tile.backdrop != 0;
 297              }
 298              if (include_tile) {
 299                  uint el_slice = el_ix / 32;
 300                  uint el_mask = 1 << (el_ix & 31);
 301                  atomicOr(sh_bitmaps[el_slice][y * N_TILE_X + x], el_mask);
 302              }
 303          }
 304  
 305          barrier();
 306  
 307          // Output non-segment elements for this tile. The thread does a sequential walk
 308          // through the non-segment elements.
 309          uint slice_ix = 0;
 310          uint bitmap = sh_bitmaps[0][th_ix];
 311          while (mem_ok) {
 312              if (bitmap == 0) {
 313                  slice_ix++;
 314                  if (slice_ix == N_SLICE) {
 315                      break;
 316                  }
 317                  bitmap = sh_bitmaps[slice_ix][th_ix];
 318                  if (bitmap == 0) {
 319                      continue;
 320                  }
 321              }
 322              uint element_ref_ix = slice_ix * 32 + findLSB(bitmap);
 323              uint element_ix = sh_elements[element_ref_ix];
 324  
 325              // Clear LSB
 326              bitmap &= bitmap - 1;
 327  
 328              // At this point, we read the element again from global memory.
 329              // If that turns out to be expensive, maybe we can pack it into
 330              // shared memory (or perhaps just the tag).
 331              ref = AnnotatedRef(conf.anno_alloc.offset + element_ix * Annotated_size);
 332              AnnotatedTag tag = Annotated_tag(conf.anno_alloc, ref);
 333  
 334              if (clip_zero_depth == 0) {
 335                  switch (tag.tag) {
 336                  case Annotated_Color:
 337                      Tile tile = Tile_read(read_tile_alloc(element_ref_ix, mem_ok), TileRef(sh_tile_base[element_ref_ix]
 338                          + (sh_tile_stride[element_ref_ix] * tile_y + tile_x) * Tile_size));
 339                      AnnoColor fill = Annotated_Color_read(conf.anno_alloc, ref);
 340                      if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
 341                          break;
 342                      }
 343                      write_fill(cmd_alloc, cmd_ref, tag.flags, tile, fill.linewidth);
 344                      Cmd_Color_write(cmd_alloc, cmd_ref, CmdColor(fill.rgba_color));
 345                      cmd_ref.offset += 4 + CmdColor_size;
 346                      break;
 347                  case Annotated_Image:
 348                      tile = Tile_read(read_tile_alloc(element_ref_ix, mem_ok), TileRef(sh_tile_base[element_ref_ix]
 349                          + (sh_tile_stride[element_ref_ix] * tile_y + tile_x) * Tile_size));
 350                      AnnoImage fill_img = Annotated_Image_read(conf.anno_alloc, ref);
 351                      if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
 352                          break;
 353                      }
 354                      write_fill(cmd_alloc, cmd_ref, tag.flags, tile, fill_img.linewidth);
 355                      Cmd_Image_write(cmd_alloc, cmd_ref, CmdImage(fill_img.index, fill_img.offset));
 356                      cmd_ref.offset += 4 + CmdImage_size;
 357                      break;
 358                  case Annotated_BeginClip:
 359                      tile = Tile_read(read_tile_alloc(element_ref_ix, mem_ok), TileRef(sh_tile_base[element_ref_ix]
 360                          + (sh_tile_stride[element_ref_ix] * tile_y + tile_x) * Tile_size));
 361                      if (tile.tile.offset == 0 && tile.backdrop == 0) {
 362                          clip_zero_depth = clip_depth + 1;
 363                      } else if (tile.tile.offset == 0 && clip_depth < 32) {
 364                          clip_one_mask |= (1 << clip_depth);
 365                      } else {
 366                          AnnoBeginClip begin_clip = Annotated_BeginClip_read(conf.anno_alloc, ref);
 367                          if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
 368                              break;
 369                          }
 370                          write_fill(cmd_alloc, cmd_ref, tag.flags, tile, begin_clip.linewidth);
 371                          Cmd_BeginClip_write(cmd_alloc, cmd_ref);
 372                          cmd_ref.offset += 4;
 373                          if (clip_depth < 32) {
 374                              clip_one_mask &= ~(1 << clip_depth);
 375                          }
 376                          begin_slot++;
 377                          num_begin_slots = max(num_begin_slots, begin_slot);
 378                      }
 379                      clip_depth++;
 380                      break;
 381                  case Annotated_EndClip:
 382                      clip_depth--;
 383                      if (clip_depth >= 32 || (clip_one_mask & (1 << clip_depth)) == 0) {
 384                          if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
 385                              break;
 386                          }
 387                          Cmd_Solid_write(cmd_alloc, cmd_ref);
 388                          cmd_ref.offset += 4;
 389                          begin_slot--;
 390                          Cmd_EndClip_write(cmd_alloc, cmd_ref);
 391                          cmd_ref.offset += 4;
 392                      }
 393                      break;
 394                  }
 395              } else {
 396                  // In "clip zero" state, suppress all drawing
 397                  switch (tag.tag) {
 398                  case Annotated_BeginClip:
 399                      clip_depth++;
 400                      break;
 401                  case Annotated_EndClip:
 402                      if (clip_depth == clip_zero_depth) {
 403                          clip_zero_depth = 0;
 404                      }
 405                      clip_depth--;
 406                      break;
 407                  }
 408              }
 409          }
 410          barrier();
 411  
 412          rd_ix += N_TILE;
 413          if (rd_ix >= ready_ix && partition_ix >= n_partitions) break;
 414      }
 415      if (bin_tile_x + tile_x < conf.width_in_tiles && bin_tile_y + tile_y < conf.height_in_tiles) {
 416          Cmd_End_write(cmd_alloc, cmd_ref);
 417          if (num_begin_slots > 0) {
 418              // Write scratch allocation: one state per BeginClip per rasterizer chunk.
 419              uint scratch_size = num_begin_slots * TILE_WIDTH_PX * TILE_HEIGHT_PX * CLIP_STATE_SIZE * 4;
 420              MallocResult scratch = malloc(scratch_size);
 421              // Ignore scratch.failed; we don't use the allocation and kernel4
 422              // checks for memory overflow before using it.
 423              alloc_write(scratch_alloc, scratch_alloc.offset, scratch.alloc);
 424          }
 425      }
 426  }
 427