src/gallium/drivers/llvmpipe/lp_setup_tri.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 /*
  29  * Binning code for triangles
  30  */
  31
  32 #include "util/u_math.h"
  33 #include "util/u_memory.h"
  34 #include "lp_perf.h"
  35 #include "lp_setup_context.h"
  36 #include "lp_rast.h"
  37
  38 #define NUM_CHANNELS 4
  39
  40
  41 /**
  42  * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
  43  */
  44 static void constant_coef( struct lp_rast_triangle *tri,
  45                            unsigned slot,
  46                            const float value,
  47                            unsigned i )
  48 {
  49    tri->inputs.a0[slot][i] = value;
  50    tri->inputs.dadx[slot][i] = 0.0f;
  51    tri->inputs.dady[slot][i] = 0.0f;
  52 }
  53
  54
  55 /**
  56  * Compute a0, dadx and dady for a linearly interpolated coefficient,
  57  * for a triangle.
  58  */
  59 static void linear_coef( struct lp_rast_triangle *tri,
  60                          float oneoverarea,
  61                          unsigned slot,
  62                          const float (*v1)[4],
  63                          const float (*v2)[4],
  64                          const float (*v3)[4],
  65                          unsigned vert_attr,
  66                          unsigned i)
  67 {
  68    float a1 = v1[vert_attr][i];
  69    float a2 = v2[vert_attr][i];
  70    float a3 = v3[vert_attr][i];
  71
  72    float da12 = a1 - a2;
  73    float da31 = a3 - a1;
  74    float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * oneoverarea;
  75    float dady = (da31 * tri->dx12 - tri->dx31 * da12) * oneoverarea;
  76
  77    tri->inputs.dadx[slot][i] = dadx;
  78    tri->inputs.dady[slot][i] = dady;
  79
  80    /* calculate a0 as the value which would be sampled for the
  81     * fragment at (0,0), taking into account that we want to sample at
  82     * pixel centers, in other words (0.5, 0.5).
  83     *
  84     * this is neat but unfortunately not a good way to do things for
  85     * triangles with very large values of dadx or dady as it will
  86     * result in the subtraction and re-addition from a0 of a very
  87     * large number, which means we'll end up loosing a lot of the
  88     * fractional bits and precision from a0.  the way to fix this is
  89     * to define a0 as the sample at a pixel center somewhere near vmin
  90     * instead - i'll switch to this later.
  91     */
  92    tri->inputs.a0[slot][i] = (a1 -
  93                               (dadx * (v1[0][0] - 0.5f) +
  94                                dady * (v1[0][1] - 0.5f)));
  95 }
  96
  97
  98 /**
  99  * Compute a0, dadx and dady for a perspective-corrected interpolant,
 100  * for a triangle.
 101  * We basically multiply the vertex value by 1/w before computing
 102  * the plane coefficients (a0, dadx, dady).
 103  * Later, when we compute the value at a particular fragment position we'll
 104  * divide the interpolated value by the interpolated W at that fragment.
 105  */
 106 static void perspective_coef( struct lp_rast_triangle *tri,
 107                               float oneoverarea,
 108                               unsigned slot,
 109                               const float (*v1)[4],
 110                               const float (*v2)[4],
 111                               const float (*v3)[4],
 112                               unsigned vert_attr,
 113                               unsigned i)
 114 {
 115    /* premultiply by 1/w  (v[0][3] is always 1/w):
 116     */
 117    float a1 = v1[vert_attr][i] * v1[0][3];
 118    float a2 = v2[vert_attr][i] * v2[0][3];
 119    float a3 = v3[vert_attr][i] * v3[0][3];
 120    float da12 = a1 - a2;
 121    float da31 = a3 - a1;
 122    float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * oneoverarea;
 123    float dady = (da31 * tri->dx12 - tri->dx31 * da12) * oneoverarea;
 124
 125    tri->inputs.dadx[slot][i] = dadx;
 126    tri->inputs.dady[slot][i] = dady;
 127    tri->inputs.a0[slot][i] = (a1 -
 128                               (dadx * (v1[0][0] - 0.5f) +
 129                                dady * (v1[0][1] - 0.5f)));
 130 }
 131
 132
 133 /**
 134  * Special coefficient setup for gl_FragCoord.
 135  * X and Y are trivial
 136  * Z and W are copied from position_coef which should have already been computed.
 137  * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
 138  */
 139 static void
 140 setup_fragcoord_coef(struct lp_rast_triangle *tri,
 141                      float oneoverarea,
 142                      unsigned slot,
 143                      const float (*v1)[4],
 144                      const float (*v2)[4],
 145                      const float (*v3)[4])
 146 {
 147    /*X*/
 148    tri->inputs.a0[slot][0] = 0.0;
 149    tri->inputs.dadx[slot][0] = 1.0;
 150    tri->inputs.dady[slot][0] = 0.0;
 151    /*Y*/
 152    tri->inputs.a0[slot][1] = 0.0;
 153    tri->inputs.dadx[slot][1] = 0.0;
 154    tri->inputs.dady[slot][1] = 1.0;
 155    /*Z*/
 156    linear_coef(tri, oneoverarea, slot, v1, v2, v3, 0, 2);
 157    /*W*/
 158    linear_coef(tri, oneoverarea, slot, v1, v2, v3, 0, 3);
 159 }
 160
 161
 162 static void setup_facing_coef( struct lp_rast_triangle *tri,
 163                                unsigned slot,
 164                                boolean frontface )
 165 {
 166    constant_coef( tri, slot, 1.0f - frontface, 0 );
 167    constant_coef( tri, slot, 0.0f, 1 ); /* wasted */
 168    constant_coef( tri, slot, 0.0f, 2 ); /* wasted */
 169    constant_coef( tri, slot, 0.0f, 3 ); /* wasted */
 170 }
 171
 172
 173 /**
 174  * Compute the tri->coef[] array dadx, dady, a0 values.
 175  */
 176 static void setup_tri_coefficients( struct setup_context *setup,
 177                                     struct lp_rast_triangle *tri,
 178                                     float oneoverarea,
 179                                     const float (*v1)[4],
 180                                     const float (*v2)[4],
 181                                     const float (*v3)[4],
 182                                     boolean frontface)
 183 {
 184    struct lp_scene *scene = lp_setup_get_current_scene(setup);
 185    unsigned slot;
 186
 187    /* Allocate space for the a0, dadx and dady arrays
 188     */
 189    {
 190       unsigned bytes = (setup->fs.nr_inputs + 1) * 4 * sizeof(float);
 191       tri->inputs.a0   = lp_scene_alloc_aligned( scene, bytes, 16 );
 192       tri->inputs.dadx = lp_scene_alloc_aligned( scene, bytes, 16 );
 193       tri->inputs.dady = lp_scene_alloc_aligned( scene, bytes, 16 );
 194    }
 195
 196    /* The internal position input is in slot zero:
 197     */
 198    setup_fragcoord_coef(tri, oneoverarea, 0, v1, v2, v3);
 199
 200    /* setup interpolation for all the remaining attributes:
 201     */
 202    for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
 203       unsigned vert_attr = setup->fs.input[slot].src_index;
 204       unsigned i;
 205
 206       switch (setup->fs.input[slot].interp) {
 207       case LP_INTERP_CONSTANT:
 208          for (i = 0; i < NUM_CHANNELS; i++)
 209             constant_coef(tri, slot+1, v3[vert_attr][i], i);
 210          break;
 211
 212       case LP_INTERP_LINEAR:
 213          for (i = 0; i < NUM_CHANNELS; i++)
 214             linear_coef(tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
 215          break;
 216
 217       case LP_INTERP_PERSPECTIVE:
 218          for (i = 0; i < NUM_CHANNELS; i++)
 219             perspective_coef(tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
 220          break;
 221
 222       case LP_INTERP_POSITION:
 223          /* XXX: fix me - duplicates the values in slot zero.
 224           */
 225          setup_fragcoord_coef(tri, oneoverarea, slot+1, v1, v2, v3);
 226          break;
 227
 228       case LP_INTERP_FACING:
 229          setup_facing_coef(tri, slot+1, frontface);
 230          break;
 231
 232       default:
 233          assert(0);
 234       }
 235    }
 236 }
 237
 238
 239
 240 static inline int subpixel_snap( float a )
 241 {
 242    return util_iround(FIXED_ONE * a - (FIXED_ONE / 2));
 243 }
 244
 245
 246 /**
 247  * Do basic setup for triangle rasterization and determine which
 248  * framebuffer tiles are touched.  Put the triangle in the scene's
 249  * bins for the tiles which we overlap.
 250  */
 251 static void
 252 do_triangle_ccw(struct setup_context *setup,
 253                 const float (*v1)[4],
 254                 const float (*v2)[4],
 255                 const float (*v3)[4],
 256                 boolean frontfacing )
 257 {
 258    /* x/y positions in fixed point */
 259    const int x1 = subpixel_snap(v1[0][0]);
 260    const int x2 = subpixel_snap(v2[0][0]);
 261    const int x3 = subpixel_snap(v3[0][0]);
 262    const int y1 = subpixel_snap(v1[0][1]);
 263    const int y2 = subpixel_snap(v2[0][1]);
 264    const int y3 = subpixel_snap(v3[0][1]);
 265
 266    struct lp_scene *scene = lp_setup_get_current_scene(setup);
 267    struct lp_rast_triangle *tri = lp_scene_alloc_aligned( scene, sizeof *tri, 16 );
 268    int area;
 269    float oneoverarea;
 270    int minx, maxx, miny, maxy;
 271
 272    tri->dx12 = x1 - x2;
 273    tri->dx23 = x2 - x3;
 274    tri->dx31 = x3 - x1;
 275
 276    tri->dy12 = y1 - y2;
 277    tri->dy23 = y2 - y3;
 278    tri->dy31 = y3 - y1;
 279
 280    area = (tri->dx12 * tri->dy31 - tri->dx31 * tri->dy12);
 281
 282    LP_COUNT(nr_tris);
 283
 284    /* Cull non-ccw and zero-sized triangles.
 285     *
 286     * XXX: subject to overflow??
 287     */
 288    if (area <= 0) {
 289       lp_scene_putback_data( scene, sizeof *tri );
 290       LP_COUNT(nr_culled_tris);
 291       return;
 292    }
 293
 294    /* Bounding rectangle (in pixels) */
 295    minx = (MIN3(x1, x2, x3) + (FIXED_ONE-1)) >> FIXED_ORDER;
 296    maxx = (MAX3(x1, x2, x3) + (FIXED_ONE-1)) >> FIXED_ORDER;
 297    miny = (MIN3(y1, y2, y3) + (FIXED_ONE-1)) >> FIXED_ORDER;
 298    maxy = (MAX3(y1, y2, y3) + (FIXED_ONE-1)) >> FIXED_ORDER;
 299
 300    if (setup->scissor_test) {
 301       minx = MAX2(minx, setup->scissor.current.minx);
 302       maxx = MIN2(maxx, setup->scissor.current.maxx);
 303       miny = MAX2(miny, setup->scissor.current.miny);
 304       maxy = MIN2(maxy, setup->scissor.current.maxy);
 305    }
 306
 307    if (miny == maxy ||
 308        minx == maxx) {
 309       lp_scene_putback_data( scene, sizeof *tri );
 310       LP_COUNT(nr_culled_tris);
 311       return;
 312    }
 313
 314    /*
 315     */
 316    oneoverarea = ((float)FIXED_ONE) / (float)area;
 317
 318    /* Setup parameter interpolants:
 319     */
 320    setup_tri_coefficients( setup, tri, oneoverarea, v1, v2, v3, frontfacing );
 321
 322    /* half-edge constants, will be interated over the whole render target.
 323     */
 324    tri->c1 = tri->dy12 * x1 - tri->dx12 * y1;
 325    tri->c2 = tri->dy23 * x2 - tri->dx23 * y2;
 326    tri->c3 = tri->dy31 * x3 - tri->dx31 * y3;
 327
 328    /* correct for top-left fill convention:
 329     */
 330    if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) tri->c1++;
 331    if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) tri->c2++;
 332    if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) tri->c3++;
 333
 334    tri->dy12 *= FIXED_ONE;
 335    tri->dy23 *= FIXED_ONE;
 336    tri->dy31 *= FIXED_ONE;
 337
 338    tri->dx12 *= FIXED_ONE;
 339    tri->dx23 *= FIXED_ONE;
 340    tri->dx31 *= FIXED_ONE;
 341
 342    /* find trivial reject offsets for each edge for a single-pixel
 343     * sized block.  These will be scaled up at each recursive level to
 344     * match the active blocksize.  Scaling in this way works best if
 345     * the blocks are square.
 346     */
 347    tri->eo1 = 0;
 348    if (tri->dy12 < 0) tri->eo1 -= tri->dy12;
 349    if (tri->dx12 > 0) tri->eo1 += tri->dx12;
 350
 351    tri->eo2 = 0;
 352    if (tri->dy23 < 0) tri->eo2 -= tri->dy23;
 353    if (tri->dx23 > 0) tri->eo2 += tri->dx23;
 354
 355    tri->eo3 = 0;
 356    if (tri->dy31 < 0) tri->eo3 -= tri->dy31;
 357    if (tri->dx31 > 0) tri->eo3 += tri->dx31;
 358
 359    /* Calculate trivial accept offsets from the above.
 360     */
 361    tri->ei1 = tri->dx12 - tri->dy12 - tri->eo1;
 362    tri->ei2 = tri->dx23 - tri->dy23 - tri->eo2;
 363    tri->ei3 = tri->dx31 - tri->dy31 - tri->eo3;
 364
 365    /* Fill in the inputs.step[][] arrays.
 366     * We've manually unrolled some loops here.
 367     */
 368    {
 369       const int xstep1 = -tri->dy12;
 370       const int xstep2 = -tri->dy23;
 371       const int xstep3 = -tri->dy31;
 372       const int ystep1 = tri->dx12;
 373       const int ystep2 = tri->dx23;
 374       const int ystep3 = tri->dx31;
 375
 376 #define SETUP_STEP(i, x, y)                                \
 377       do {                                                 \
 378          tri->inputs.step[0][i] = x * xstep1 + y * ystep1; \
 379          tri->inputs.step[1][i] = x * xstep2 + y * ystep2; \
 380          tri->inputs.step[2][i] = x * xstep3 + y * ystep3; \
 381       } while (0)
 382
 383       SETUP_STEP(0, 0, 0);
 384       SETUP_STEP(1, 1, 0);
 385       SETUP_STEP(2, 0, 1);
 386       SETUP_STEP(3, 1, 1);
 387
 388       SETUP_STEP(4, 2, 0);
 389       SETUP_STEP(5, 3, 0);
 390       SETUP_STEP(6, 2, 1);
 391       SETUP_STEP(7, 3, 1);
 392
 393       SETUP_STEP(8, 0, 2);
 394       SETUP_STEP(9, 1, 2);
 395       SETUP_STEP(10, 0, 3);
 396       SETUP_STEP(11, 1, 3);
 397
 398       SETUP_STEP(12, 2, 2);
 399       SETUP_STEP(13, 3, 2);
 400       SETUP_STEP(14, 2, 3);
 401       SETUP_STEP(15, 3, 3);
 402 #undef STEP
 403    }
 404
 405    /*
 406     * All fields of 'tri' are now set.  The remaining code here is
 407     * concerned with binning.
 408     */
 409
 410    /* Convert to tile coordinates:
 411     */
 412    minx = minx / TILE_SIZE;
 413    miny = miny / TILE_SIZE;
 414    maxx = maxx / TILE_SIZE;
 415    maxy = maxy / TILE_SIZE;
 416
 417    /* Clamp maxx, maxy to framebuffer size
 418     */
 419    maxx = MIN2(maxx, scene->tiles_x - 1);
 420    maxy = MIN2(maxy, scene->tiles_y - 1);
 421
 422    /* Determine which tile(s) intersect the triangle's bounding box
 423     */
 424    if (miny == maxy && minx == maxx)
 425    {
 426       /* Triangle is contained in a single tile:
 427        */
 428       lp_scene_bin_command( scene, minx, miny, lp_rast_triangle,
 429                             lp_rast_arg_triangle(tri) );
 430    }
 431    else
 432    {
 433       int c1 = (tri->c1 +
 434                 tri->dx12 * miny * TILE_SIZE -
 435                 tri->dy12 * minx * TILE_SIZE);
 436       int c2 = (tri->c2 +
 437                 tri->dx23 * miny * TILE_SIZE -
 438                 tri->dy23 * minx * TILE_SIZE);
 439       int c3 = (tri->c3 +
 440                 tri->dx31 * miny * TILE_SIZE -
 441                 tri->dy31 * minx * TILE_SIZE);
 442
 443       int ei1 = tri->ei1 << TILE_ORDER;
 444       int ei2 = tri->ei2 << TILE_ORDER;
 445       int ei3 = tri->ei3 << TILE_ORDER;
 446
 447       int eo1 = tri->eo1 << TILE_ORDER;
 448       int eo2 = tri->eo2 << TILE_ORDER;
 449       int eo3 = tri->eo3 << TILE_ORDER;
 450
 451       int xstep1 = -(tri->dy12 << TILE_ORDER);
 452       int xstep2 = -(tri->dy23 << TILE_ORDER);
 453       int xstep3 = -(tri->dy31 << TILE_ORDER);
 454
 455       int ystep1 = tri->dx12 << TILE_ORDER;
 456       int ystep2 = tri->dx23 << TILE_ORDER;
 457       int ystep3 = tri->dx31 << TILE_ORDER;
 458       int x, y;
 459
 460
 461       /* Test tile-sized blocks against the triangle.
 462        * Discard blocks fully outside the tri.  If the block is fully
 463        * contained inside the tri, bin an lp_rast_shade_tile command.
 464        * Else, bin a lp_rast_triangle command.
 465        */
 466       for (y = miny; y <= maxy; y++)
 467       {
 468          int cx1 = c1;
 469          int cx2 = c2;
 470          int cx3 = c3;
 471          boolean in = FALSE;  /* are we inside the triangle? */
 472
 473          for (x = minx; x <= maxx; x++)
 474          {
 475             if (cx1 + eo1 < 0 ||
 476                 cx2 + eo2 < 0 ||
 477                 cx3 + eo3 < 0)
 478             {
 479                /* do nothing */
 480                LP_COUNT(nr_empty_64);
 481                if (in)
 482                   break;  /* exiting triangle, all done with this row */
 483             }
 484             else if (cx1 + ei1 > 0 &&
 485                      cx2 + ei2 > 0 &&
 486                      cx3 + ei3 > 0)
 487             {
 488                /* triangle covers the whole tile- shade whole tile */
 489                LP_COUNT(nr_fully_covered_64);
 490                in = TRUE;
 491                if(setup->fs.current.opaque) {
 492                   lp_scene_bin_reset( scene, x, y );
 493                   lp_scene_bin_command( scene, x, y,
 494                                         lp_rast_set_state,
 495                                         lp_rast_arg_state(setup->fs.stored) );
 496                }
 497                lp_scene_bin_command( scene, x, y,
 498                                      lp_rast_shade_tile,
 499                                      lp_rast_arg_inputs(&tri->inputs) );
 500             }
 501             else
 502             {
 503                /* rasterizer/shade partial tile */
 504                LP_COUNT(nr_partially_covered_64);
 505                in = TRUE;
 506                lp_scene_bin_command( scene, x, y,
 507                                      lp_rast_triangle,
 508                                      lp_rast_arg_triangle(tri) );
 509             }
 510
 511             /* Iterate cx values across the region:
 512              */
 513             cx1 += xstep1;
 514             cx2 += xstep2;
 515             cx3 += xstep3;
 516          }
 517
 518          /* Iterate c values down the region:
 519           */
 520          c1 += ystep1;
 521          c2 += ystep2;
 522          c3 += ystep3;
 523       }
 524    }
 525 }
 526
 527
 528 static void triangle_cw( struct setup_context *setup,
 529                          const float (*v0)[4],
 530                          const float (*v1)[4],
 531                          const float (*v2)[4] )
 532 {
 533    do_triangle_ccw( setup, v1, v0, v2, !setup->ccw_is_frontface );
 534 }
 535
 536
 537 static void triangle_ccw( struct setup_context *setup,
 538                          const float (*v0)[4],
 539                          const float (*v1)[4],
 540                          const float (*v2)[4] )
 541 {
 542    do_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface );
 543 }
 544
 545
 546 static void triangle_both( struct setup_context *setup,
 547                            const float (*v0)[4],
 548                            const float (*v1)[4],
 549                            const float (*v2)[4] )
 550 {
 551    /* edge vectors e = v0 - v2, f = v1 - v2 */
 552    const float ex = v0[0][0] - v2[0][0];
 553    const float ey = v0[0][1] - v2[0][1];
 554    const float fx = v1[0][0] - v2[0][0];
 555    const float fy = v1[0][1] - v2[0][1];
 556
 557    /* det = cross(e,f).z */
 558    if (ex * fy - ey * fx < 0.0f)
 559       triangle_ccw( setup, v0, v1, v2 );
 560    else
 561       triangle_cw( setup, v0, v1, v2 );
 562 }
 563
 564
 565 static void triangle_nop( struct setup_context *setup,
 566                           const float (*v0)[4],
 567                           const float (*v1)[4],
 568                           const float (*v2)[4] )
 569 {
 570 }
 571
 572
 573 void
 574 lp_setup_choose_triangle( struct setup_context *setup )
 575 {
 576    switch (setup->cullmode) {
 577    case PIPE_WINDING_NONE:
 578       setup->triangle = triangle_both;
 579       break;
 580    case PIPE_WINDING_CCW:
 581       setup->triangle = triangle_cw;
 582       break;
 583    case PIPE_WINDING_CW:
 584       setup->triangle = triangle_ccw;
 585       break;
 586    default:
 587       setup->triangle = triangle_nop;
 588       break;
 589    }
 590 }