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 "lp_setup_context.h"
  33 #include "lp_rast.h"
  34 #include "util/u_math.h"
  35 #include "util/u_memory.h"
  36
  37 #define NUM_CHANNELS 4
  38
  39 /**
  40  * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
  41  */
  42 static void constant_coef( struct lp_rast_triangle *tri,
  43                            unsigned slot,
  44                            const float value,
  45                            unsigned i )
  46 {
  47    tri->inputs.a0[slot][i] = value;
  48    tri->inputs.dadx[slot][i] = 0;
  49    tri->inputs.dady[slot][i] = 0;
  50 }
  51
  52 /**
  53  * Compute a0, dadx and dady for a linearly interpolated coefficient,
  54  * for a triangle.
  55  */
  56 static void linear_coef( struct lp_rast_triangle *tri,
  57                          unsigned slot,
  58                          const float (*v1)[4],
  59                          const float (*v2)[4],
  60                          const float (*v3)[4],
  61                          unsigned vert_attr,
  62                          unsigned i)
  63 {
  64    float a1 = v1[vert_attr][i];
  65    float a2 = v2[vert_attr][i];
  66    float a3 = v3[vert_attr][i];
  67
  68    float da12 = a1 - a2;
  69    float da31 = a3 - a1;
  70    float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * tri->oneoverarea;
  71    float dady = (da31 * tri->dx12 - tri->dx31 * da12) * tri->oneoverarea;
  72
  73    tri->inputs.dadx[slot][i] = dadx;
  74    tri->inputs.dady[slot][i] = dady;
  75
  76    /* calculate a0 as the value which would be sampled for the
  77     * fragment at (0,0), taking into account that we want to sample at
  78     * pixel centers, in other words (0.5, 0.5).
  79     *
  80     * this is neat but unfortunately not a good way to do things for
  81     * triangles with very large values of dadx or dady as it will
  82     * result in the subtraction and re-addition from a0 of a very
  83     * large number, which means we'll end up loosing a lot of the
  84     * fractional bits and precision from a0.  the way to fix this is
  85     * to define a0 as the sample at a pixel center somewhere near vmin
  86     * instead - i'll switch to this later.
  87     */
  88    tri->inputs.a0[slot][i] = (v1[vert_attr][i] -
  89                               (dadx * (v1[0][0] - 0.5f) +
  90                                dady * (v1[0][1] - 0.5f)));
  91 }
  92
  93
  94 /**
  95  * Compute a0, dadx and dady for a perspective-corrected interpolant,
  96  * for a triangle.
  97  * We basically multiply the vertex value by 1/w before computing
  98  * the plane coefficients (a0, dadx, dady).
  99  * Later, when we compute the value at a particular fragment position we'll
 100  * divide the interpolated value by the interpolated W at that fragment.
 101  */
 102 static void perspective_coef( struct lp_rast_triangle *tri,
 103                               unsigned slot,
 104                               const float (*v1)[4],
 105                               const float (*v2)[4],
 106                               const float (*v3)[4],
 107                               unsigned vert_attr,
 108                               unsigned i)
 109 {
 110    /* premultiply by 1/w  (v[0][3] is always 1/w):
 111     */
 112    float a1 = v1[vert_attr][i] * v1[0][3];
 113    float a2 = v2[vert_attr][i] * v2[0][3];
 114    float a3 = v3[vert_attr][i] * v3[0][3];
 115    float da12 = a1 - a2;
 116    float da31 = a3 - a1;
 117    float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * tri->oneoverarea;
 118    float dady = (da31 * tri->dx12 - tri->dx31 * da12) * tri->oneoverarea;
 119
 120
 121    tri->inputs.dadx[slot][i] = dadx;
 122    tri->inputs.dady[slot][i] = dady;
 123    tri->inputs.a0[slot][i] = (a1 -
 124                               (dadx * (v1[0][0] - 0.5f) +
 125                                dady * (v1[0][1] - 0.5f)));
 126 }
 127
 128
 129 /**
 130  * Special coefficient setup for gl_FragCoord.
 131  * X and Y are trivial, though Y has to be inverted for OpenGL.
 132  * Z and W are copied from position_coef which should have already been computed.
 133  * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
 134  */
 135 static void
 136 setup_fragcoord_coef(struct lp_rast_triangle *tri,
 137                      unsigned slot,
 138                      const float (*v1)[4],
 139                      const float (*v2)[4],
 140                      const float (*v3)[4])
 141 {
 142    /*X*/
 143    tri->inputs.a0[slot][0] = 0.0;
 144    tri->inputs.dadx[slot][0] = 1.0;
 145    tri->inputs.dady[slot][0] = 0.0;
 146    /*Y*/
 147    tri->inputs.a0[slot][1] = 0.0;
 148    tri->inputs.dadx[slot][1] = 0.0;
 149    tri->inputs.dady[slot][1] = 1.0;
 150    /*Z*/
 151    linear_coef(tri, slot, v1, v2, v3, 0, 2);
 152    /*W*/
 153    linear_coef(tri, slot, v1, v2, v3, 0, 3);
 154 }
 155
 156
 157 static void setup_facing_coef( struct lp_rast_triangle *tri,
 158                                unsigned slot,
 159                                boolean frontface )
 160 {
 161    constant_coef( tri, slot, 1.0f - frontface, 0 );
 162    constant_coef( tri, slot, 0.0f, 1 ); /* wasted */
 163    constant_coef( tri, slot, 0.0f, 2 ); /* wasted */
 164    constant_coef( tri, slot, 0.0f, 3 ); /* wasted */
 165 }
 166
 167
 168 /**
 169  * Compute the tri->coef[] array dadx, dady, a0 values.
 170  */
 171 static void setup_tri_coefficients( struct setup_context *setup,
 172                                     struct lp_rast_triangle *tri,
 173                                     const float (*v1)[4],
 174                                     const float (*v2)[4],
 175                                     const float (*v3)[4],
 176                                     boolean frontface )
 177 {
 178    unsigned slot;
 179
 180    /* The internal position input is in slot zero:
 181     */
 182    setup_fragcoord_coef(tri, 0, v1, v2, v3);
 183
 184    /* setup interpolation for all the remaining attrbutes:
 185     */
 186    for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
 187       unsigned vert_attr = setup->fs.input[slot].src_index;
 188       unsigned i;
 189
 190       switch (setup->fs.input[slot].interp) {
 191       case LP_INTERP_CONSTANT:
 192          for (i = 0; i < NUM_CHANNELS; i++)
 193             constant_coef(tri, slot+1, v3[vert_attr][i], i);
 194          break;
 195
 196       case LP_INTERP_LINEAR:
 197          for (i = 0; i < NUM_CHANNELS; i++)
 198             linear_coef(tri, slot+1, v1, v2, v3, vert_attr, i);
 199          break;
 200
 201       case LP_INTERP_PERSPECTIVE:
 202          for (i = 0; i < NUM_CHANNELS; i++)
 203             perspective_coef(tri, slot+1, v1, v2, v3, vert_attr, i);
 204          break;
 205
 206       case LP_INTERP_POSITION:
 207          /* XXX: fix me - duplicates the values in slot zero.
 208           */
 209          setup_fragcoord_coef(tri, slot+1, v1, v2, v3);
 210          break;
 211
 212       case LP_INTERP_FACING:
 213          setup_facing_coef(tri, slot+1, frontface);
 214          break;
 215
 216       default:
 217          assert(0);
 218       }
 219    }
 220 }
 221
 222
 223
 224 /* XXX: do this by add/subtracting a large floating point number:
 225  */
 226 static inline int subpixel_snap( float a )
 227 {
 228    return util_iround(FIXED_ONE * a);
 229 }
 230
 231
 232 static INLINE void bin_triangle( struct cmd_block_list *list,
 233                                  const struct lp_rast_triangle arg )
 234 {
 235 }
 236
 237
 238 /* to avoid having to allocate power-of-four, square render targets,
 239  * end up having a specialized version of the above that runs only at
 240  * the topmost level.
 241  *
 242  * at the topmost level there may be an arbitary number of steps on
 243  * either dimension, so this loop needs to be either separately
 244  * code-generated and unrolled for each render target size, or kept as
 245  * generic looping code:
 246  */
 247
 248 #define MIN3(a,b,c) MIN2(MIN2(a,b),c)
 249 #define MAX3(a,b,c) MAX2(MAX2(a,b),c)
 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
 259    const int y1 = subpixel_snap(v1[0][1]);
 260    const int y2 = subpixel_snap(v2[0][1]);
 261    const int y3 = subpixel_snap(v3[0][1]);
 262
 263    const int x1 = subpixel_snap(v1[0][0]);
 264    const int x2 = subpixel_snap(v2[0][0]);
 265    const int x3 = subpixel_snap(v3[0][0]);
 266
 267    struct lp_rast_triangle *tri = get_data( &setup->data, sizeof *tri );
 268    float area;
 269    int minx, maxx, miny, maxy;
 270
 271    tri->dx12 = x1 - x2;
 272    tri->dx23 = x2 - x3;
 273    tri->dx31 = x3 - x1;
 274
 275    tri->dy12 = y1 - y2;
 276    tri->dy23 = y2 - y3;
 277    tri->dy31 = y3 - y1;
 278
 279    area = (tri->dx12 * tri->dy31 -
 280            tri->dx31 * tri->dy12);
 281
 282    /* Cull non-ccw and zero-sized triangles.
 283     *
 284     * XXX: subject to overflow??
 285     */
 286    if (area <= 0) {
 287       putback_data( &setup->data, sizeof *tri );
 288       return;
 289    }
 290
 291    // Bounding rectangle
 292    tri->minx = (MIN3(x1, x2, x3) + 0xf) >> FIXED_ORDER;
 293    tri->maxx = (MAX3(x1, x2, x3) + 0xf) >> FIXED_ORDER;
 294    tri->miny = (MIN3(y1, y2, y3) + 0xf) >> FIXED_ORDER;
 295    tri->maxy = (MAX3(y1, y2, y3) + 0xf) >> FIXED_ORDER;
 296
 297    if (tri->miny == tri->maxy ||
 298        tri->minx == tri->maxx) {
 299       putback_data( &setup->data, sizeof *tri );
 300       return;
 301    }
 302
 303    tri->inputs.state = setup->fs.stored;
 304
 305    /*
 306     */
 307    tri->oneoverarea = ((float)FIXED_ONE) / (float)area;
 308
 309    /* Setup parameter interpolants:
 310     */
 311    setup_tri_coefficients( setup, tri, v1, v2, v3, frontfacing );
 312
 313    /* half-edge constants, will be interated over the whole
 314     * rendertarget.
 315     */
 316    tri->c1 = tri->dy12 * x1 - tri->dx12 * y1;
 317    tri->c2 = tri->dy23 * x2 - tri->dx23 * y2;
 318    tri->c3 = tri->dy31 * x3 - tri->dx31 * y3;
 319
 320    /* correct for top-left fill convention:
 321     */
 322    if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) tri->c1++;
 323    if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) tri->c2++;
 324    if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) tri->c3++;
 325
 326    tri->dy12 *= FIXED_ONE;
 327    tri->dy23 *= FIXED_ONE;
 328    tri->dy31 *= FIXED_ONE;
 329
 330    tri->dx12 *= FIXED_ONE;
 331    tri->dx23 *= FIXED_ONE;
 332    tri->dx31 *= FIXED_ONE;
 333
 334    /* find trivial reject offsets for each edge for a single-pixel
 335     * sized block.  These will be scaled up at each recursive level to
 336     * match the active blocksize.  Scaling in this way works best if
 337     * the blocks are square.
 338     */
 339    tri->eo1 = 0;
 340    if (tri->dy12 < 0) tri->eo1 -= tri->dy12;
 341    if (tri->dx12 > 0) tri->eo1 += tri->dx12;
 342
 343    tri->eo2 = 0;
 344    if (tri->dy23 < 0) tri->eo2 -= tri->dy23;
 345    if (tri->dx23 > 0) tri->eo2 += tri->dx23;
 346
 347    tri->eo3 = 0;
 348    if (tri->dy31 < 0) tri->eo3 -= tri->dy31;
 349    if (tri->dx31 > 0) tri->eo3 += tri->dx31;
 350
 351    /* Calculate trivial accept offsets from the above.
 352     */
 353    tri->ei1 = tri->dx12 - tri->dy12 - tri->eo1;
 354    tri->ei2 = tri->dx23 - tri->dy23 - tri->eo2;
 355    tri->ei3 = tri->dx31 - tri->dy31 - tri->eo3;
 356
 357    minx = tri->minx / TILESIZE;
 358    miny = tri->miny / TILESIZE;
 359    maxx = tri->maxx / TILESIZE;
 360    maxy = tri->maxy / TILESIZE;
 361
 362    {
 363       int xstep1 = -tri->dy12;
 364       int xstep2 = -tri->dy23;
 365       int xstep3 = -tri->dy31;
 366
 367       int ystep1 = tri->dx12;
 368       int ystep2 = tri->dx23;
 369       int ystep3 = tri->dx31;
 370
 371       int ix, iy;
 372       int qx, qy;
 373       int i = 0;
 374
 375       for (qy = 0; qy < 4; qy += 2) {
 376          for (qx = 0; qx < 4; qx += 2) {
 377             for (iy = 0; iy < 2; iy++) {
 378                for (ix = 0; ix < 2; ix++, i++) {
 379                   tri->step[0][i] = (xstep1 * (qx+ix)) + (ystep1 * (qy+iy));
 380                   tri->step[1][i] = (xstep2 * (qx+ix)) + (ystep2 * (qy+iy));
 381                   tri->step[2][i] = (xstep3 * (qx+ix)) + (ystep3 * (qy+iy));
 382                }
 383             }
 384          }
 385       }
 386    }
 387
 388    /* Convert to tile coordinates:
 389     */
 390    if (miny == maxy && minx == maxx)
 391    {
 392       /* Triangle is contained in a single tile:
 393        */
 394       bin_command( &setup->tile[minx][miny], lp_rast_triangle,
 395                    lp_rast_arg_triangle(tri) );
 396    }
 397    else
 398    {
 399       int c1 = (tri->c1 +
 400                 tri->dx12 * miny * TILESIZE -
 401                 tri->dy12 * minx * TILESIZE);
 402       int c2 = (tri->c2 +
 403                 tri->dx23 * miny * TILESIZE -
 404                 tri->dy23 * minx * TILESIZE);
 405       int c3 = (tri->c3 +
 406                 tri->dx31 * miny * TILESIZE -
 407                 tri->dy31 * minx * TILESIZE);
 408
 409       int ei1 = tri->ei1 << TILE_ORDER;
 410       int ei2 = tri->ei2 << TILE_ORDER;
 411       int ei3 = tri->ei3 << TILE_ORDER;
 412
 413       int eo1 = tri->eo1 << TILE_ORDER;
 414       int eo2 = tri->eo2 << TILE_ORDER;
 415       int eo3 = tri->eo3 << TILE_ORDER;
 416
 417       int xstep1 = -(tri->dy12 << TILE_ORDER);
 418       int xstep2 = -(tri->dy23 << TILE_ORDER);
 419       int xstep3 = -(tri->dy31 << TILE_ORDER);
 420
 421       int ystep1 = tri->dx12 << TILE_ORDER;
 422       int ystep2 = tri->dx23 << TILE_ORDER;
 423       int ystep3 = tri->dx31 << TILE_ORDER;
 424       int x, y;
 425
 426
 427       /* Subdivide space into NxM blocks, where each block is square and
 428        * power-of-four in dimension.
 429        *
 430        * Trivially accept or reject blocks, else jump to per-pixel
 431        * examination above.
 432        */
 433       for (y = miny; y <= maxy; y++)
 434       {
 435          int cx1 = c1;
 436          int cx2 = c2;
 437          int cx3 = c3;
 438
 439          for (x = minx; x <= maxx; x++)
 440          {
 441             assert(cx1 ==
 442                    tri->c1 +
 443                    tri->dx12 * y * TILESIZE -
 444                    tri->dy12 * x * TILESIZE);
 445             assert(cx2 ==
 446                    tri->c2 +
 447                    tri->dx23 * y * TILESIZE -
 448                    tri->dy23 * x * TILESIZE);
 449             assert(cx3 ==
 450                    tri->c3 +
 451                    tri->dx31 * y * TILESIZE -
 452                    tri->dy31 * x * TILESIZE);
 453
 454             if (cx1 + eo1 < 0 ||
 455                 cx2 + eo2 < 0 ||
 456                 cx3 + eo3 < 0)
 457             {
 458                /* do nothing */
 459             }
 460             else if (cx1 + ei1 > 0 &&
 461                      cx2 + ei2 > 0 &&
 462                      cx3 + ei3 > 0)
 463             {
 464                /* shade whole tile */
 465                bin_command( &setup->tile[x][y], lp_rast_shade_tile,
 466                             lp_rast_arg_inputs(&tri->inputs) );
 467             }
 468             else
 469             {
 470                /* shade partial tile */
 471                bin_command( &setup->tile[x][y],
 472                             lp_rast_triangle,
 473                             lp_rast_arg_triangle(tri) );
 474             }
 475
 476             /* Iterate cx values across the region:
 477              */
 478             cx1 += xstep1;
 479             cx2 += xstep2;
 480             cx3 += xstep3;
 481          }
 482
 483          /* Iterate c values down the region:
 484           */
 485          c1 += ystep1;
 486          c2 += ystep2;
 487          c3 += ystep3;
 488       }
 489    }
 490 }
 491
 492 static void triangle_cw( struct setup_context *setup,
 493                          const float (*v0)[4],
 494                          const float (*v1)[4],
 495                          const float (*v2)[4] )
 496 {
 497    do_triangle_ccw( setup, v1, v0, v2, !setup->ccw_is_frontface );
 498 }
 499
 500 static void triangle_ccw( struct setup_context *setup,
 501                          const float (*v0)[4],
 502                          const float (*v1)[4],
 503                          const float (*v2)[4] )
 504 {
 505    do_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface );
 506 }
 507
 508 static void triangle_both( struct setup_context *setup,
 509                            const float (*v0)[4],
 510                            const float (*v1)[4],
 511                            const float (*v2)[4] )
 512 {
 513    /* edge vectors e = v0 - v2, f = v1 - v2 */
 514    const float ex = v0[0][0] - v2[0][0];
 515    const float ey = v0[0][1] - v2[0][1];
 516    const float fx = v1[0][0] - v2[0][0];
 517    const float fy = v1[0][1] - v2[0][1];
 518
 519    /* det = cross(e,f).z */
 520    if (ex * fy - ey * fx < 0)
 521       triangle_ccw( setup, v0, v1, v2 );
 522    else
 523       triangle_cw( setup, v0, v1, v2 );
 524 }
 525
 526 static void triangle_nop( struct setup_context *setup,
 527                           const float (*v0)[4],
 528                           const float (*v1)[4],
 529                           const float (*v2)[4] )
 530 {
 531 }
 532
 533
 534 void
 535 lp_setup_choose_triangle( struct setup_context *setup )
 536 {
 537    switch (setup->cullmode) {
 538    case PIPE_WINDING_NONE:
 539       setup->triangle = triangle_both;
 540       break;
 541    case PIPE_WINDING_CCW:
 542       setup->triangle = triangle_cw;
 543       break;
 544    case PIPE_WINDING_CW:
 545       setup->triangle = triangle_ccw;
 546       break;
 547    default:
 548       setup->triangle = triangle_nop;
 549       break;
 550    }
 551 }
 552
 553