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 "util/u_rect.h"
  35 #include "lp_perf.h"
  36 #include "lp_setup_context.h"
  37 #include "lp_setup_coef.h"
  38 #include "lp_rast.h"
  39 #include "lp_state_fs.h"
  40
  41 #define NUM_CHANNELS 4
  42
  43
  44
  45 static INLINE int
  46 subpixel_snap(float a)
  47 {
  48    return util_iround(FIXED_ONE * a);
  49 }
  50
  51 static INLINE float
  52 fixed_to_float(int a)
  53 {
  54    return a * (1.0 / FIXED_ONE);
  55 }
  56
  57
  58
  59
  60
  61
  62
  63 /**
  64  * Alloc space for a new triangle plus the input.a0/dadx/dady arrays
  65  * immediately after it.
  66  * The memory is allocated from the per-scene pool, not per-tile.
  67  * \param tri_size  returns number of bytes allocated
  68  * \param nr_inputs  number of fragment shader inputs
  69  * \return pointer to triangle space
  70  */
  71 struct lp_rast_triangle *
  72 lp_setup_alloc_triangle(struct lp_scene *scene,
  73                         unsigned nr_inputs,
  74                         unsigned nr_planes,
  75                         unsigned *tri_size)
  76 {
  77    unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
  78    struct lp_rast_triangle *tri;
  79    unsigned tri_bytes, bytes;
  80    char *inputs;
  81
  82    tri_bytes = align(Offset(struct lp_rast_triangle, plane[nr_planes]), 16);
  83    bytes = tri_bytes + (3 * input_array_sz);
  84
  85    tri = lp_scene_alloc_aligned( scene, bytes, 16 );
  86
  87    if (tri) {
  88       inputs = ((char *)tri) + tri_bytes;
  89       tri->inputs.a0   = (float (*)[4]) inputs;
  90       tri->inputs.dadx = (float (*)[4]) (inputs + input_array_sz);
  91       tri->inputs.dady = (float (*)[4]) (inputs + 2 * input_array_sz);
  92
  93       *tri_size = bytes;
  94    }
  95
  96    return tri;
  97 }
  98
  99 void
 100 lp_setup_print_vertex(struct lp_setup_context *setup,
 101                       const char *name,
 102                       const float (*v)[4])
 103 {
 104    int i, j;
 105
 106    debug_printf("   wpos (%s[0]) xyzw %f %f %f %f\n",
 107                 name,
 108                 v[0][0], v[0][1], v[0][2], v[0][3]);
 109
 110    for (i = 0; i < setup->fs.nr_inputs; i++) {
 111       const float *in = v[setup->fs.input[i].src_index];
 112
 113       debug_printf("  in[%d] (%s[%d]) %s%s%s%s ",
 114                    i,
 115                    name, setup->fs.input[i].src_index,
 116                    (setup->fs.input[i].usage_mask & 0x1) ? "x" : " ",
 117                    (setup->fs.input[i].usage_mask & 0x2) ? "y" : " ",
 118                    (setup->fs.input[i].usage_mask & 0x4) ? "z" : " ",
 119                    (setup->fs.input[i].usage_mask & 0x8) ? "w" : " ");
 120
 121       for (j = 0; j < 4; j++)
 122          if (setup->fs.input[i].usage_mask & (1<<j))
 123             debug_printf("%.5f ", in[j]);
 124
 125       debug_printf("\n");
 126    }
 127 }
 128
 129
 130 /**
 131  * Print triangle vertex attribs (for debug).
 132  */
 133 void
 134 lp_setup_print_triangle(struct lp_setup_context *setup,
 135                         const float (*v0)[4],
 136                         const float (*v1)[4],
 137                         const float (*v2)[4])
 138 {
 139    debug_printf("triangle\n");
 140
 141    {
 142       const float ex = v0[0][0] - v2[0][0];
 143       const float ey = v0[0][1] - v2[0][1];
 144       const float fx = v1[0][0] - v2[0][0];
 145       const float fy = v1[0][1] - v2[0][1];
 146
 147       /* det = cross(e,f).z */
 148       const float det = ex * fy - ey * fx;
 149       if (det < 0.0f)
 150          debug_printf("   - ccw\n");
 151       else if (det > 0.0f)
 152          debug_printf("   - cw\n");
 153       else
 154          debug_printf("   - zero area\n");
 155    }
 156
 157    lp_setup_print_vertex(setup, "v0", v0);
 158    lp_setup_print_vertex(setup, "v1", v1);
 159    lp_setup_print_vertex(setup, "v2", v2);
 160 }
 161
 162
 163 lp_rast_cmd lp_rast_tri_tab[9] = {
 164    NULL,               /* should be impossible */
 165    lp_rast_triangle_1,
 166    lp_rast_triangle_2,
 167    lp_rast_triangle_3,
 168    lp_rast_triangle_4,
 169    lp_rast_triangle_5,
 170    lp_rast_triangle_6,
 171    lp_rast_triangle_7,
 172    lp_rast_triangle_8
 173 };
 174
 175 /**
 176  * Do basic setup for triangle rasterization and determine which
 177  * framebuffer tiles are touched.  Put the triangle in the scene's
 178  * bins for the tiles which we overlap.
 179  */
 180 static void
 181 do_triangle_ccw(struct lp_setup_context *setup,
 182                 const float (*v0)[4],
 183                 const float (*v1)[4],
 184                 const float (*v2)[4],
 185                 boolean frontfacing )
 186 {
 187
 188    struct lp_scene *scene = lp_setup_get_current_scene(setup);
 189    struct lp_fragment_shader_variant *variant = setup->fs.current.variant;
 190    struct lp_rast_triangle *tri;
 191    int x[3];
 192    int y[3];
 193    float dy01, dy20;
 194    float dx01, dx20;
 195    float oneoverarea;
 196    struct lp_tri_info info;
 197    int area;
 198    struct u_rect bbox;
 199    int ix0, ix1, iy0, iy1;
 200    unsigned tri_bytes;
 201    int i;
 202    int nr_planes = 3;
 203
 204    if (0)
 205       lp_setup_print_triangle(setup, v0, v1, v2);
 206
 207    if (setup->scissor_test) {
 208       nr_planes = 7;
 209    }
 210    else {
 211       nr_planes = 3;
 212    }
 213
 214    /* x/y positions in fixed point */
 215    x[0] = subpixel_snap(v0[0][0] - setup->pixel_offset);
 216    x[1] = subpixel_snap(v1[0][0] - setup->pixel_offset);
 217    x[2] = subpixel_snap(v2[0][0] - setup->pixel_offset);
 218    y[0] = subpixel_snap(v0[0][1] - setup->pixel_offset);
 219    y[1] = subpixel_snap(v1[0][1] - setup->pixel_offset);
 220    y[2] = subpixel_snap(v2[0][1] - setup->pixel_offset);
 221
 222
 223    /* Bounding rectangle (in pixels) */
 224    {
 225       /* Yes this is necessary to accurately calculate bounding boxes
 226        * with the two fill-conventions we support.  GL (normally) ends
 227        * up needing a bottom-left fill convention, which requires
 228        * slightly different rounding.
 229        */
 230       int adj = (setup->pixel_offset != 0) ? 1 : 0;
 231
 232       bbox.x0 = (MIN3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
 233       bbox.x1 = (MAX3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
 234       bbox.y0 = (MIN3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
 235       bbox.y1 = (MAX3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
 236
 237       /* Inclusive coordinates:
 238        */
 239       bbox.x1--;
 240       bbox.y1--;
 241    }
 242
 243    if (bbox.x1 < bbox.x0 ||
 244        bbox.y1 < bbox.y0) {
 245       if (0) debug_printf("empty bounding box\n");
 246       LP_COUNT(nr_culled_tris);
 247       return;
 248    }
 249
 250    if (!u_rect_test_intersection(&setup->draw_region, &bbox)) {
 251       if (0) debug_printf("offscreen\n");
 252       LP_COUNT(nr_culled_tris);
 253       return;
 254    }
 255
 256    u_rect_find_intersection(&setup->draw_region, &bbox);
 257
 258    tri = lp_setup_alloc_triangle(scene,
 259                                  setup->fs.nr_inputs,
 260                                  nr_planes,
 261                                  &tri_bytes);
 262    if (!tri)
 263       return;
 264
 265 #ifdef DEBUG
 266    tri->v[0][0] = v0[0][0];
 267    tri->v[1][0] = v1[0][0];
 268    tri->v[2][0] = v2[0][0];
 269    tri->v[0][1] = v0[0][1];
 270    tri->v[1][1] = v1[0][1];
 271    tri->v[2][1] = v2[0][1];
 272 #endif
 273
 274    tri->plane[0].dcdy = x[0] - x[1];
 275    tri->plane[1].dcdy = x[1] - x[2];
 276    tri->plane[2].dcdy = x[2] - x[0];
 277
 278    tri->plane[0].dcdx = y[0] - y[1];
 279    tri->plane[1].dcdx = y[1] - y[2];
 280    tri->plane[2].dcdx = y[2] - y[0];
 281
 282    area = (tri->plane[0].dcdy * tri->plane[2].dcdx -
 283            tri->plane[2].dcdy * tri->plane[0].dcdx);
 284
 285    LP_COUNT(nr_tris);
 286
 287    /* Cull non-ccw and zero-sized triangles.
 288     *
 289     * XXX: subject to overflow??
 290     */
 291    if (area <= 0) {
 292       lp_scene_putback_data( scene, tri_bytes );
 293       LP_COUNT(nr_culled_tris);
 294       return;
 295    }
 296
 297
 298    /*
 299     */
 300    dx01 = v0[0][0] - v1[0][0];
 301    dy01 = v0[0][1] - v1[0][1];
 302    dx20 = v2[0][0] - v0[0][0];
 303    dy20 = v2[0][1] - v0[0][1];
 304    oneoverarea = 1.0f / (dx01 * dy20 - dx20 * dy01);
 305
 306    info.v0 = v0;
 307    info.v1 = v1;
 308    info.v2 = v2;
 309    info.frontfacing = frontfacing;
 310    info.x0_center = v0[0][0] - setup->pixel_offset;
 311    info.y0_center = v0[0][1] - setup->pixel_offset;
 312    info.dx01_ooa  = dx01 * oneoverarea;
 313    info.dx20_ooa  = dx20 * oneoverarea;
 314    info.dy01_ooa  = dy01 * oneoverarea;
 315    info.dy20_ooa  = dy20 * oneoverarea;
 316
 317    /* Setup parameter interpolants:
 318     */
 319    lp_setup_tri_coef( setup, &tri->inputs, &info );
 320
 321    tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
 322    tri->inputs.state = setup->fs.stored;
 323
 324
 325
 326    for (i = 0; i < 3; i++) {
 327       struct lp_rast_plane *plane = &tri->plane[i];
 328
 329       /* half-edge constants, will be interated over the whole render
 330        * target.
 331        */
 332       plane->c = plane->dcdx * x[i] - plane->dcdy * y[i];
 333
 334       /* correct for top-left vs. bottom-left fill convention.
 335        *
 336        * note that we're overloading gl_rasterization_rules to mean
 337        * both (0.5,0.5) pixel centers *and* bottom-left filling
 338        * convention.
 339        *
 340        * GL actually has a top-left filling convention, but GL's
 341        * notion of "top" differs from gallium's...
 342        *
 343        * Also, sometimes (in FBO cases) GL will render upside down
 344        * to its usual method, in which case it will probably want
 345        * to use the opposite, top-left convention.
 346        */
 347       if (plane->dcdx < 0) {
 348          /* both fill conventions want this - adjust for left edges */
 349          plane->c++;
 350       }
 351       else if (plane->dcdx == 0) {
 352          if (setup->pixel_offset == 0) {
 353             /* correct for top-left fill convention:
 354              */
 355             if (plane->dcdy > 0) plane->c++;
 356          }
 357          else {
 358             /* correct for bottom-left fill convention:
 359              */
 360             if (plane->dcdy < 0) plane->c++;
 361          }
 362       }
 363
 364       plane->dcdx *= FIXED_ONE;
 365       plane->dcdy *= FIXED_ONE;
 366
 367       /* find trivial reject offsets for each edge for a single-pixel
 368        * sized block.  These will be scaled up at each recursive level to
 369        * match the active blocksize.  Scaling in this way works best if
 370        * the blocks are square.
 371        */
 372       plane->eo = 0;
 373       if (plane->dcdx < 0) plane->eo -= plane->dcdx;
 374       if (plane->dcdy > 0) plane->eo += plane->dcdy;
 375
 376       /* Calculate trivial accept offsets from the above.
 377        */
 378       plane->ei = plane->dcdy - plane->dcdx - plane->eo;
 379    }
 380
 381
 382    /*
 383     * When rasterizing scissored tris, use the intersection of the
 384     * triangle bounding box and the scissor rect to generate the
 385     * scissor planes.
 386     *
 387     * This permits us to cut off the triangle "tails" that are present
 388     * in the intermediate recursive levels caused when two of the
 389     * triangles edges don't diverge quickly enough to trivially reject
 390     * exterior blocks from the triangle.
 391     *
 392     * It's not really clear if it's worth worrying about these tails,
 393     * but since we generate the planes for each scissored tri, it's
 394     * free to trim them in this case.
 395     *
 396     * Note that otherwise, the scissor planes only vary in 'C' value,
 397     * and even then only on state-changes.  Could alternatively store
 398     * these planes elsewhere.
 399     */
 400    if (nr_planes == 7) {
 401       tri->plane[3].dcdx = -1;
 402       tri->plane[3].dcdy = 0;
 403       tri->plane[3].c = 1-bbox.x0;
 404       tri->plane[3].ei = 0;
 405       tri->plane[3].eo = 1;
 406
 407       tri->plane[4].dcdx = 1;
 408       tri->plane[4].dcdy = 0;
 409       tri->plane[4].c = bbox.x1+1;
 410       tri->plane[4].ei = -1;
 411       tri->plane[4].eo = 0;
 412
 413       tri->plane[5].dcdx = 0;
 414       tri->plane[5].dcdy = 1;
 415       tri->plane[5].c = 1-bbox.y0;
 416       tri->plane[5].ei = 0;
 417       tri->plane[5].eo = 1;
 418
 419       tri->plane[6].dcdx = 0;
 420       tri->plane[6].dcdy = -1;
 421       tri->plane[6].c = bbox.y1+1;
 422       tri->plane[6].ei = -1;
 423       tri->plane[6].eo = 0;
 424    }
 425
 426
 427    /*
 428     * All fields of 'tri' are now set.  The remaining code here is
 429     * concerned with binning.
 430     */
 431
 432    /* Convert to tile coordinates, and inclusive ranges:
 433     */
 434    if (nr_planes == 3) {
 435       int ix0 = bbox.x0 / 16;
 436       int iy0 = bbox.y0 / 16;
 437       int ix1 = bbox.x1 / 16;
 438       int iy1 = bbox.y1 / 16;
 439
 440       if (iy0 == iy1 && ix0 == ix1)
 441       {
 442
 443          /* Triangle is contained in a single 16x16 block:
 444           */
 445          int mask = (ix0 & 3) | ((iy0 & 3) << 4);
 446
 447          lp_scene_bin_command( scene, ix0/4, iy0/4,
 448                                lp_rast_triangle_3_16,
 449                                lp_rast_arg_triangle(tri, mask) );
 450          return;
 451       }
 452    }
 453
 454    ix0 = bbox.x0 / TILE_SIZE;
 455    iy0 = bbox.y0 / TILE_SIZE;
 456    ix1 = bbox.x1 / TILE_SIZE;
 457    iy1 = bbox.y1 / TILE_SIZE;
 458
 459    /*
 460     * Clamp to framebuffer size
 461     */
 462    assert(ix0 == MAX2(ix0, 0));
 463    assert(iy0 == MAX2(iy0, 0));
 464    assert(ix1 == MIN2(ix1, scene->tiles_x - 1));
 465    assert(iy1 == MIN2(iy1, scene->tiles_y - 1));
 466
 467    /* Determine which tile(s) intersect the triangle's bounding box
 468     */
 469    if (iy0 == iy1 && ix0 == ix1)
 470    {
 471       /* Triangle is contained in a single tile:
 472        */
 473       lp_scene_bin_command( scene, ix0, iy0,
 474                             lp_rast_tri_tab[nr_planes],
 475                             lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
 476    }
 477    else
 478    {
 479       int c[7];
 480       int ei[7];
 481       int eo[7];
 482       int xstep[7];
 483       int ystep[7];
 484       int x, y;
 485
 486       for (i = 0; i < nr_planes; i++) {
 487          c[i] = (tri->plane[i].c +
 488                  tri->plane[i].dcdy * iy0 * TILE_SIZE -
 489                  tri->plane[i].dcdx * ix0 * TILE_SIZE);
 490
 491          ei[i] = tri->plane[i].ei << TILE_ORDER;
 492          eo[i] = tri->plane[i].eo << TILE_ORDER;
 493          xstep[i] = -(tri->plane[i].dcdx << TILE_ORDER);
 494          ystep[i] = tri->plane[i].dcdy << TILE_ORDER;
 495       }
 496
 497
 498
 499       /* Test tile-sized blocks against the triangle.
 500        * Discard blocks fully outside the tri.  If the block is fully
 501        * contained inside the tri, bin an lp_rast_shade_tile command.
 502        * Else, bin a lp_rast_triangle command.
 503        */
 504       for (y = iy0; y <= iy1; y++)
 505       {
 506          boolean in = FALSE;  /* are we inside the triangle? */
 507          int cx[7];
 508
 509          for (i = 0; i < nr_planes; i++)
 510             cx[i] = c[i];
 511
 512          for (x = ix0; x <= ix1; x++)
 513          {
 514             int out = 0;
 515             int partial = 0;
 516
 517             for (i = 0; i < nr_planes; i++) {
 518                int planeout = cx[i] + eo[i];
 519                int planepartial = cx[i] + ei[i] - 1;
 520                out |= (planeout >> 31);
 521                partial |= (planepartial >> 31) & (1<<i);
 522             }
 523
 524             if (out) {
 525                /* do nothing */
 526                if (in)
 527                   break;  /* exiting triangle, all done with this row */
 528                LP_COUNT(nr_empty_64);
 529             }
 530             else if (partial) {
 531                /* Not trivially accepted by at least one plane -
 532                 * rasterize/shade partial tile
 533                 */
 534                int count = util_bitcount(partial);
 535                in = TRUE;
 536                lp_scene_bin_command( scene, x, y,
 537                                      lp_rast_tri_tab[count],
 538                                      lp_rast_arg_triangle(tri, partial) );
 539
 540                LP_COUNT(nr_partially_covered_64);
 541             }
 542             else {
 543                /* triangle covers the whole tile- shade whole tile */
 544                LP_COUNT(nr_fully_covered_64);
 545                in = TRUE;
 546                if (variant->opaque &&
 547                    !setup->fb.zsbuf) {
 548                   lp_scene_bin_reset( scene, x, y );
 549                }
 550                lp_scene_bin_command( scene, x, y,
 551                                      lp_rast_shade_tile,
 552                                      lp_rast_arg_inputs(&tri->inputs) );
 553             }
 554
 555             /* Iterate cx values across the region:
 556              */
 557             for (i = 0; i < nr_planes; i++)
 558                cx[i] += xstep[i];
 559          }
 560
 561          /* Iterate c values down the region:
 562           */
 563          for (i = 0; i < nr_planes; i++)
 564             c[i] += ystep[i];
 565       }
 566    }
 567 }
 568
 569
 570 /**
 571  * Draw triangle if it's CW, cull otherwise.
 572  */
 573 static void triangle_cw( struct lp_setup_context *setup,
 574                          const float (*v0)[4],
 575                          const float (*v1)[4],
 576                          const float (*v2)[4] )
 577 {
 578    do_triangle_ccw( setup, v1, v0, v2, !setup->ccw_is_frontface );
 579 }
 580
 581
 582 /**
 583  * Draw triangle if it's CCW, cull otherwise.
 584  */
 585 static void triangle_ccw( struct lp_setup_context *setup,
 586                          const float (*v0)[4],
 587                          const float (*v1)[4],
 588                          const float (*v2)[4] )
 589 {
 590    do_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface );
 591 }
 592
 593
 594
 595 /**
 596  * Draw triangle whether it's CW or CCW.
 597  */
 598 static void triangle_both( struct lp_setup_context *setup,
 599                            const float (*v0)[4],
 600                            const float (*v1)[4],
 601                            const float (*v2)[4] )
 602 {
 603    /* edge vectors e = v0 - v2, f = v1 - v2 */
 604    const float ex = v0[0][0] - v2[0][0];
 605    const float ey = v0[0][1] - v2[0][1];
 606    const float fx = v1[0][0] - v2[0][0];
 607    const float fy = v1[0][1] - v2[0][1];
 608
 609    /* det = cross(e,f).z */
 610    const float det = ex * fy - ey * fx;
 611    if (det < 0.0f)
 612       triangle_ccw( setup, v0, v1, v2 );
 613    else if (det > 0.0f)
 614       triangle_cw( setup, v0, v1, v2 );
 615 }
 616
 617
 618 static void triangle_nop( struct lp_setup_context *setup,
 619                           const float (*v0)[4],
 620                           const float (*v1)[4],
 621                           const float (*v2)[4] )
 622 {
 623 }
 624
 625
 626 void
 627 lp_setup_choose_triangle( struct lp_setup_context *setup )
 628 {
 629    switch (setup->cullmode) {
 630    case PIPE_FACE_NONE:
 631       setup->triangle = triangle_both;
 632       break;
 633    case PIPE_FACE_BACK:
 634       setup->triangle = setup->ccw_is_frontface ? triangle_ccw : triangle_cw;
 635       break;
 636    case PIPE_FACE_FRONT:
 637       setup->triangle = setup->ccw_is_frontface ? triangle_cw : triangle_ccw;
 638       break;
 639    default:
 640       setup->triangle = triangle_nop;
 641       break;
 642    }
 643 }