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    tri->inputs.dadx[slot][i] = dadx;
 121    tri->inputs.dady[slot][i] = dady;
 122    tri->inputs.a0[slot][i] = (a1 -
 123                               (dadx * (v1[0][0] - 0.5f) +
 124                                dady * (v1[0][1] - 0.5f)));
 125 }
 126
 127
 128 /**
 129  * Special coefficient setup for gl_FragCoord.
 130  * X and Y are trivial, though Y has to be inverted for OpenGL.
 131  * Z and W are copied from position_coef which should have already been computed.
 132  * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
 133  */
 134 static void
 135 setup_fragcoord_coef(struct lp_rast_triangle *tri,
 136                      unsigned slot,
 137                      const float (*v1)[4],
 138                      const float (*v2)[4],
 139                      const float (*v3)[4])
 140 {
 141    /*X*/
 142    tri->inputs.a0[slot][0] = 0.0;
 143    tri->inputs.dadx[slot][0] = 1.0;
 144    tri->inputs.dady[slot][0] = 0.0;
 145    /*Y*/
 146    tri->inputs.a0[slot][1] = 0.0;
 147    tri->inputs.dadx[slot][1] = 0.0;
 148    tri->inputs.dady[slot][1] = 1.0;
 149    /*Z*/
 150    linear_coef(tri, slot, v1, v2, v3, 0, 2);
 151    /*W*/
 152    linear_coef(tri, slot, v1, v2, v3, 0, 3);
 153 }
 154
 155
 156 static void setup_facing_coef( struct lp_rast_triangle *tri,
 157                                unsigned slot,
 158                                boolean frontface )
 159 {
 160    constant_coef( tri, slot, 1.0f - frontface, 0 );
 161    constant_coef( tri, slot, 0.0f, 1 ); /* wasted */
 162    constant_coef( tri, slot, 0.0f, 2 ); /* wasted */
 163    constant_coef( tri, slot, 0.0f, 3 ); /* wasted */
 164 }
 165
 166
 167 /**
 168  * Compute the tri->coef[] array dadx, dady, a0 values.
 169  */
 170 static void setup_tri_coefficients( struct setup_context *setup,
 171                                     struct lp_rast_triangle *tri,
 172                                     const float (*v1)[4],
 173                                     const float (*v2)[4],
 174                                     const float (*v3)[4],
 175                                     boolean frontface )
 176 {
 177    unsigned slot;
 178
 179    /* Allocate space for the a0, dadx and dady arrays
 180     */
 181    {
 182       unsigned bytes;
 183       bytes = (setup->fs.nr_inputs + 1) * 4 * sizeof(float);
 184       tri->inputs.a0   = get_data_aligned( &setup->data, bytes, 16 );
 185       tri->inputs.dadx = get_data_aligned( &setup->data, bytes, 16 );
 186       tri->inputs.dady = get_data_aligned( &setup->data, bytes, 16 );
 187    }
 188
 189    /* The internal position input is in slot zero:
 190     */
 191    setup_fragcoord_coef(tri, 0, v1, v2, v3);
 192
 193    /* setup interpolation for all the remaining attributes:
 194     */
 195    for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
 196       unsigned vert_attr = setup->fs.input[slot].src_index;
 197       unsigned i;
 198
 199       switch (setup->fs.input[slot].interp) {
 200       case LP_INTERP_CONSTANT:
 201          for (i = 0; i < NUM_CHANNELS; i++)
 202             constant_coef(tri, slot+1, v3[vert_attr][i], i);
 203          break;
 204
 205       case LP_INTERP_LINEAR:
 206          for (i = 0; i < NUM_CHANNELS; i++)
 207             linear_coef(tri, slot+1, v1, v2, v3, vert_attr, i);
 208          break;
 209
 210       case LP_INTERP_PERSPECTIVE:
 211          for (i = 0; i < NUM_CHANNELS; i++)
 212             perspective_coef(tri, slot+1, v1, v2, v3, vert_attr, i);
 213          break;
 214
 215       case LP_INTERP_POSITION:
 216          /* XXX: fix me - duplicates the values in slot zero.
 217           */
 218          setup_fragcoord_coef(tri, slot+1, v1, v2, v3);
 219          break;
 220
 221       case LP_INTERP_FACING:
 222          setup_facing_coef(tri, slot+1, frontface);
 223          break;
 224
 225       default:
 226          assert(0);
 227       }
 228    }
 229 }
 230
 231
 232
 233 static inline int subpixel_snap( float a )
 234 {
 235    return util_iround(FIXED_ONE * a);
 236 }
 237
 238
 239 #define MIN3(a,b,c) MIN2(MIN2(a,b),c)
 240 #define MAX3(a,b,c) MAX2(MAX2(a,b),c)
 241
 242 /**
 243  * Do basic setup for triangle rasterization and determine which
 244  * framebuffer tiles are touched.  Put the triangle in the bins for the
 245  * tiles which we overlap.
 246  */
 247 static void
 248 do_triangle_ccw(struct setup_context *setup,
 249                 const float (*v1)[4],
 250                 const float (*v2)[4],
 251                 const float (*v3)[4],
 252                 boolean frontfacing )
 253 {
 254    /* x/y positions in fixed point */
 255    const int x1 = subpixel_snap(v1[0][0]);
 256    const int x2 = subpixel_snap(v2[0][0]);
 257    const int x3 = subpixel_snap(v3[0][0]);
 258    const int y1 = subpixel_snap(v1[0][1]);
 259    const int y2 = subpixel_snap(v2[0][1]);
 260    const int y3 = subpixel_snap(v3[0][1]);
 261
 262    struct lp_rast_triangle *tri = get_data( &setup->data, sizeof *tri );
 263    float area;
 264    int minx, maxx, miny, maxy;
 265
 266    tri->dx12 = x1 - x2;
 267    tri->dx23 = x2 - x3;
 268    tri->dx31 = x3 - x1;
 269
 270    tri->dy12 = y1 - y2;
 271    tri->dy23 = y2 - y3;
 272    tri->dy31 = y3 - y1;
 273
 274    area = (tri->dx12 * tri->dy31 -
 275            tri->dx31 * tri->dy12);
 276
 277    /* Cull non-ccw and zero-sized triangles.
 278     *
 279     * XXX: subject to overflow??
 280     */
 281    if (area <= 0) {
 282       putback_data( &setup->data, sizeof *tri );
 283       return;
 284    }
 285
 286    /* Bounding rectangle (in pixels) */
 287    tri->minx = (MIN3(x1, x2, x3) + 0xf) >> FIXED_ORDER;
 288    tri->maxx = (MAX3(x1, x2, x3) + 0xf) >> FIXED_ORDER;
 289    tri->miny = (MIN3(y1, y2, y3) + 0xf) >> FIXED_ORDER;
 290    tri->maxy = (MAX3(y1, y2, y3) + 0xf) >> FIXED_ORDER;
 291
 292    if (tri->miny == tri->maxy ||
 293        tri->minx == tri->maxx) {
 294       putback_data( &setup->data, sizeof *tri );
 295       return;
 296    }
 297
 298    /*
 299     */
 300    tri->oneoverarea = ((float)FIXED_ONE) / (float)area;
 301
 302    /* Setup parameter interpolants:
 303     */
 304    setup_tri_coefficients( setup, tri, v1, v2, v3, frontfacing );
 305
 306    /* half-edge constants, will be interated over the whole
 307     * rendertarget.
 308     */
 309    tri->c1 = tri->dy12 * x1 - tri->dx12 * y1;
 310    tri->c2 = tri->dy23 * x2 - tri->dx23 * y2;
 311    tri->c3 = tri->dy31 * x3 - tri->dx31 * y3;
 312
 313    /* correct for top-left fill convention:
 314     */
 315    if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) tri->c1++;
 316    if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) tri->c2++;
 317    if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) tri->c3++;
 318
 319    tri->dy12 *= FIXED_ONE;
 320    tri->dy23 *= FIXED_ONE;
 321    tri->dy31 *= FIXED_ONE;
 322
 323    tri->dx12 *= FIXED_ONE;
 324    tri->dx23 *= FIXED_ONE;
 325    tri->dx31 *= FIXED_ONE;
 326
 327    /* find trivial reject offsets for each edge for a single-pixel
 328     * sized block.  These will be scaled up at each recursive level to
 329     * match the active blocksize.  Scaling in this way works best if
 330     * the blocks are square.
 331     */
 332    tri->eo1 = 0;
 333    if (tri->dy12 < 0) tri->eo1 -= tri->dy12;
 334    if (tri->dx12 > 0) tri->eo1 += tri->dx12;
 335
 336    tri->eo2 = 0;
 337    if (tri->dy23 < 0) tri->eo2 -= tri->dy23;
 338    if (tri->dx23 > 0) tri->eo2 += tri->dx23;
 339
 340    tri->eo3 = 0;
 341    if (tri->dy31 < 0) tri->eo3 -= tri->dy31;
 342    if (tri->dx31 > 0) tri->eo3 += tri->dx31;
 343
 344    /* Calculate trivial accept offsets from the above.
 345     */
 346    tri->ei1 = tri->dx12 - tri->dy12 - tri->eo1;
 347    tri->ei2 = tri->dx23 - tri->dy23 - tri->eo2;
 348    tri->ei3 = tri->dx31 - tri->dy31 - tri->eo3;
 349
 350    {
 351       int xstep1 = -tri->dy12;
 352       int xstep2 = -tri->dy23;
 353       int xstep3 = -tri->dy31;
 354
 355       int ystep1 = tri->dx12;
 356       int ystep2 = tri->dx23;
 357       int ystep3 = tri->dx31;
 358
 359       int ix, iy;
 360       int i = 0;
 361
 362       int c1 = 0;
 363       int c2 = 0;
 364       int c3 = 0;
 365
 366       for (iy = 0; iy < 4; iy++) {
 367          int cx1 = c1;
 368          int cx2 = c2;
 369          int cx3 = c3;
 370
 371          for (ix = 0; ix < 4; ix++, i++) {
 372             tri->step[0][i] = cx1;
 373             tri->step[1][i] = cx2;
 374             tri->step[2][i] = cx3;
 375             cx1 += xstep1;
 376             cx2 += xstep2;
 377             cx3 += xstep3;
 378          }
 379
 380          c1 += ystep1;
 381          c2 += ystep2;
 382          c3 += ystep3;
 383       }
 384    }
 385
 386    /*
 387     * All fields of 'tri' are now set.  The remaining code here is
 388     * concerned with binning.
 389     */
 390
 391    /* Convert to tile coordinates:
 392     */
 393    minx = tri->minx / TILE_SIZE;
 394    miny = tri->miny / TILE_SIZE;
 395    maxx = tri->maxx / TILE_SIZE;
 396    maxy = tri->maxy / TILE_SIZE;
 397
 398    /* Determine which tile(s) intersect the triangle's bounding box
 399     */
 400    if (miny == maxy && minx == maxx)
 401    {
 402       /* Triangle is contained in a single tile:
 403        */
 404       bin_command( &setup->tile[minx][miny], lp_rast_triangle,
 405                    lp_rast_arg_triangle(tri) );
 406    }
 407    else
 408    {
 409       int c1 = (tri->c1 +
 410                 tri->dx12 * miny * TILE_SIZE -
 411                 tri->dy12 * minx * TILE_SIZE);
 412       int c2 = (tri->c2 +
 413                 tri->dx23 * miny * TILE_SIZE -
 414                 tri->dy23 * minx * TILE_SIZE);
 415       int c3 = (tri->c3 +
 416                 tri->dx31 * miny * TILE_SIZE -
 417                 tri->dy31 * minx * TILE_SIZE);
 418
 419       int ei1 = tri->ei1 << TILE_ORDER;
 420       int ei2 = tri->ei2 << TILE_ORDER;
 421       int ei3 = tri->ei3 << TILE_ORDER;
 422
 423       int eo1 = tri->eo1 << TILE_ORDER;
 424       int eo2 = tri->eo2 << TILE_ORDER;
 425       int eo3 = tri->eo3 << TILE_ORDER;
 426
 427       int xstep1 = -(tri->dy12 << TILE_ORDER);
 428       int xstep2 = -(tri->dy23 << TILE_ORDER);
 429       int xstep3 = -(tri->dy31 << TILE_ORDER);
 430
 431       int ystep1 = tri->dx12 << TILE_ORDER;
 432       int ystep2 = tri->dx23 << TILE_ORDER;
 433       int ystep3 = tri->dx31 << TILE_ORDER;
 434       int x, y;
 435
 436
 437       /* Trivially accept or reject blocks, else jump to per-pixel
 438        * examination above.
 439        */
 440       for (y = miny; y <= maxy; y++)
 441       {
 442          int cx1 = c1;
 443          int cx2 = c2;
 444          int cx3 = c3;
 445          int in = 0;
 446
 447          for (x = minx; x <= maxx; x++)
 448          {
 449             if (cx1 + eo1 < 0 ||
 450                 cx2 + eo2 < 0 ||
 451                 cx3 + eo3 < 0)
 452             {
 453                /* do nothing */
 454                if (in)
 455                   break;
 456             }
 457             else if (cx1 + ei1 > 0 &&
 458                      cx2 + ei2 > 0 &&
 459                      cx3 + ei3 > 0)
 460             {
 461                in = 1;
 462                /* triangle covers the whole tile- shade whole tile */
 463                bin_command( &setup->tile[x][y],
 464                             lp_rast_shade_tile,
 465                             lp_rast_arg_inputs(&tri->inputs) );
 466             }
 467             else
 468             {
 469                in = 1;
 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