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    struct lp_scene *scene = lp_setup_get_current_scene(setup);
 188    struct lp_rast_triangle *tri;
 189    int x[3];
 190    int y[3];
 191    float dy01, dy20;
 192    float dx01, dx20;
 193    float oneoverarea;
 194    struct lp_tri_info info;
 195    int area;
 196    struct u_rect bbox;
 197    unsigned tri_bytes;
 198    int i;
 199    int nr_planes = 3;
 200
 201    if (0)
 202       lp_setup_print_triangle(setup, v0, v1, v2);
 203
 204    if (setup->scissor_test) {
 205       nr_planes = 7;
 206    }
 207    else {
 208       nr_planes = 3;
 209    }
 210
 211    /* x/y positions in fixed point */
 212    x[0] = subpixel_snap(v0[0][0] - setup->pixel_offset);
 213    x[1] = subpixel_snap(v1[0][0] - setup->pixel_offset);
 214    x[2] = subpixel_snap(v2[0][0] - setup->pixel_offset);
 215    y[0] = subpixel_snap(v0[0][1] - setup->pixel_offset);
 216    y[1] = subpixel_snap(v1[0][1] - setup->pixel_offset);
 217    y[2] = subpixel_snap(v2[0][1] - setup->pixel_offset);
 218
 219
 220    /* Bounding rectangle (in pixels) */
 221    {
 222       /* Yes this is necessary to accurately calculate bounding boxes
 223        * with the two fill-conventions we support.  GL (normally) ends
 224        * up needing a bottom-left fill convention, which requires
 225        * slightly different rounding.
 226        */
 227       int adj = (setup->pixel_offset != 0) ? 1 : 0;
 228
 229       bbox.x0 = (MIN3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
 230       bbox.x1 = (MAX3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
 231       bbox.y0 = (MIN3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
 232       bbox.y1 = (MAX3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
 233
 234       /* Inclusive coordinates:
 235        */
 236       bbox.x1--;
 237       bbox.y1--;
 238    }
 239
 240    if (bbox.x1 < bbox.x0 ||
 241        bbox.y1 < bbox.y0) {
 242       if (0) debug_printf("empty bounding box\n");
 243       LP_COUNT(nr_culled_tris);
 244       return;
 245    }
 246
 247    if (!u_rect_test_intersection(&setup->draw_region, &bbox)) {
 248       if (0) debug_printf("offscreen\n");
 249       LP_COUNT(nr_culled_tris);
 250       return;
 251    }
 252
 253    u_rect_find_intersection(&setup->draw_region, &bbox);
 254
 255    tri = lp_setup_alloc_triangle(scene,
 256                                  setup->fs.nr_inputs,
 257                                  nr_planes,
 258                                  &tri_bytes);
 259    if (!tri)
 260       return;
 261
 262 #ifdef DEBUG
 263    tri->v[0][0] = v0[0][0];
 264    tri->v[1][0] = v1[0][0];
 265    tri->v[2][0] = v2[0][0];
 266    tri->v[0][1] = v0[0][1];
 267    tri->v[1][1] = v1[0][1];
 268    tri->v[2][1] = v2[0][1];
 269 #endif
 270
 271    tri->plane[0].dcdy = x[0] - x[1];
 272    tri->plane[1].dcdy = x[1] - x[2];
 273    tri->plane[2].dcdy = x[2] - x[0];
 274
 275    tri->plane[0].dcdx = y[0] - y[1];
 276    tri->plane[1].dcdx = y[1] - y[2];
 277    tri->plane[2].dcdx = y[2] - y[0];
 278
 279    area = (tri->plane[0].dcdy * tri->plane[2].dcdx -
 280            tri->plane[2].dcdy * tri->plane[0].dcdx);
 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, tri_bytes );
 290       LP_COUNT(nr_culled_tris);
 291       return;
 292    }
 293
 294
 295    /*
 296     */
 297    dx01 = v0[0][0] - v1[0][0];
 298    dy01 = v0[0][1] - v1[0][1];
 299    dx20 = v2[0][0] - v0[0][0];
 300    dy20 = v2[0][1] - v0[0][1];
 301    oneoverarea = 1.0f / (dx01 * dy20 - dx20 * dy01);
 302
 303    info.v0 = v0;
 304    info.v1 = v1;
 305    info.v2 = v2;
 306    info.frontfacing = frontfacing;
 307    info.x0_center = v0[0][0] - setup->pixel_offset;
 308    info.y0_center = v0[0][1] - setup->pixel_offset;
 309    info.dx01_ooa  = dx01 * oneoverarea;
 310    info.dx20_ooa  = dx20 * oneoverarea;
 311    info.dy01_ooa  = dy01 * oneoverarea;
 312    info.dy20_ooa  = dy20 * oneoverarea;
 313
 314    /* Setup parameter interpolants:
 315     */
 316    lp_setup_tri_coef( setup, &tri->inputs, &info );
 317
 318    tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
 319    tri->inputs.state = setup->fs.stored;
 320
 321
 322
 323    for (i = 0; i < 3; i++) {
 324       struct lp_rast_plane *plane = &tri->plane[i];
 325
 326       /* half-edge constants, will be interated over the whole render
 327        * target.
 328        */
 329       plane->c = plane->dcdx * x[i] - plane->dcdy * y[i];
 330
 331       /* correct for top-left vs. bottom-left fill convention.
 332        *
 333        * note that we're overloading gl_rasterization_rules to mean
 334        * both (0.5,0.5) pixel centers *and* bottom-left filling
 335        * convention.
 336        *
 337        * GL actually has a top-left filling convention, but GL's
 338        * notion of "top" differs from gallium's...
 339        *
 340        * Also, sometimes (in FBO cases) GL will render upside down
 341        * to its usual method, in which case it will probably want
 342        * to use the opposite, top-left convention.
 343        */
 344       if (plane->dcdx < 0) {
 345          /* both fill conventions want this - adjust for left edges */
 346          plane->c++;
 347       }
 348       else if (plane->dcdx == 0) {
 349          if (setup->pixel_offset == 0) {
 350             /* correct for top-left fill convention:
 351              */
 352             if (plane->dcdy > 0) plane->c++;
 353          }
 354          else {
 355             /* correct for bottom-left fill convention:
 356              */
 357             if (plane->dcdy < 0) plane->c++;
 358          }
 359       }
 360
 361       plane->dcdx *= FIXED_ONE;
 362       plane->dcdy *= FIXED_ONE;
 363
 364       /* find trivial reject offsets for each edge for a single-pixel
 365        * sized block.  These will be scaled up at each recursive level to
 366        * match the active blocksize.  Scaling in this way works best if
 367        * the blocks are square.
 368        */
 369       plane->eo = 0;
 370       if (plane->dcdx < 0) plane->eo -= plane->dcdx;
 371       if (plane->dcdy > 0) plane->eo += plane->dcdy;
 372
 373       /* Calculate trivial accept offsets from the above.
 374        */
 375       plane->ei = plane->dcdy - plane->dcdx - plane->eo;
 376    }
 377
 378
 379    /*
 380     * When rasterizing scissored tris, use the intersection of the
 381     * triangle bounding box and the scissor rect to generate the
 382     * scissor planes.
 383     *
 384     * This permits us to cut off the triangle "tails" that are present
 385     * in the intermediate recursive levels caused when two of the
 386     * triangles edges don't diverge quickly enough to trivially reject
 387     * exterior blocks from the triangle.
 388     *
 389     * It's not really clear if it's worth worrying about these tails,
 390     * but since we generate the planes for each scissored tri, it's
 391     * free to trim them in this case.
 392     *
 393     * Note that otherwise, the scissor planes only vary in 'C' value,
 394     * and even then only on state-changes.  Could alternatively store
 395     * these planes elsewhere.
 396     */
 397    if (nr_planes == 7) {
 398       tri->plane[3].dcdx = -1;
 399       tri->plane[3].dcdy = 0;
 400       tri->plane[3].c = 1-bbox.x0;
 401       tri->plane[3].ei = 0;
 402       tri->plane[3].eo = 1;
 403
 404       tri->plane[4].dcdx = 1;
 405       tri->plane[4].dcdy = 0;
 406       tri->plane[4].c = bbox.x1+1;
 407       tri->plane[4].ei = -1;
 408       tri->plane[4].eo = 0;
 409
 410       tri->plane[5].dcdx = 0;
 411       tri->plane[5].dcdy = 1;
 412       tri->plane[5].c = 1-bbox.y0;
 413       tri->plane[5].ei = 0;
 414       tri->plane[5].eo = 1;
 415
 416       tri->plane[6].dcdx = 0;
 417       tri->plane[6].dcdy = -1;
 418       tri->plane[6].c = bbox.y1+1;
 419       tri->plane[6].ei = -1;
 420       tri->plane[6].eo = 0;
 421    }
 422
 423    lp_setup_bin_triangle( setup, tri, &bbox, nr_planes );
 424 }
 425
 426
 427 void
 428 lp_setup_bin_triangle( struct lp_setup_context *setup,
 429                        struct lp_rast_triangle *tri,
 430                        const struct u_rect *bbox,
 431                        int nr_planes )
 432 {
 433    struct lp_scene *scene = setup->scene;
 434    struct lp_fragment_shader_variant *variant = setup->fs.current.variant;
 435    int ix0, ix1, iy0, iy1;
 436    int i;
 437
 438    /*
 439     * All fields of 'tri' are now set.  The remaining code here is
 440     * concerned with binning.
 441     */
 442
 443    /* Convert to tile coordinates, and inclusive ranges:
 444     */
 445    if (nr_planes == 3) {
 446       int ix0 = bbox->x0 / 16;
 447       int iy0 = bbox->y0 / 16;
 448       int ix1 = bbox->x1 / 16;
 449       int iy1 = bbox->y1 / 16;
 450
 451       if (iy0 == iy1 && ix0 == ix1)
 452       {
 453
 454          /* Triangle is contained in a single 16x16 block:
 455           */
 456          int mask = (ix0 & 3) | ((iy0 & 3) << 4);
 457
 458          lp_scene_bin_command( scene, ix0/4, iy0/4,
 459                                lp_rast_triangle_3_16,
 460                                lp_rast_arg_triangle(tri, mask) );
 461          return;
 462       }
 463    }
 464
 465    ix0 = bbox->x0 / TILE_SIZE;
 466    iy0 = bbox->y0 / TILE_SIZE;
 467    ix1 = bbox->x1 / TILE_SIZE;
 468    iy1 = bbox->y1 / TILE_SIZE;
 469
 470    /*
 471     * Clamp to framebuffer size
 472     */
 473    assert(ix0 == MAX2(ix0, 0));
 474    assert(iy0 == MAX2(iy0, 0));
 475    assert(ix1 == MIN2(ix1, scene->tiles_x - 1));
 476    assert(iy1 == MIN2(iy1, scene->tiles_y - 1));
 477
 478    /* Determine which tile(s) intersect the triangle's bounding box
 479     */
 480    if (iy0 == iy1 && ix0 == ix1)
 481    {
 482       /* Triangle is contained in a single tile:
 483        */
 484       lp_scene_bin_command( scene, ix0, iy0,
 485                             lp_rast_tri_tab[nr_planes],
 486                             lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
 487    }
 488    else
 489    {
 490       int c[7];
 491       int ei[7];
 492       int eo[7];
 493       int xstep[7];
 494       int ystep[7];
 495       int x, y;
 496
 497       for (i = 0; i < nr_planes; i++) {
 498          c[i] = (tri->plane[i].c +
 499                  tri->plane[i].dcdy * iy0 * TILE_SIZE -
 500                  tri->plane[i].dcdx * ix0 * TILE_SIZE);
 501
 502          ei[i] = tri->plane[i].ei << TILE_ORDER;
 503          eo[i] = tri->plane[i].eo << TILE_ORDER;
 504          xstep[i] = -(tri->plane[i].dcdx << TILE_ORDER);
 505          ystep[i] = tri->plane[i].dcdy << TILE_ORDER;
 506       }
 507
 508
 509
 510       /* Test tile-sized blocks against the triangle.
 511        * Discard blocks fully outside the tri.  If the block is fully
 512        * contained inside the tri, bin an lp_rast_shade_tile command.
 513        * Else, bin a lp_rast_triangle command.
 514        */
 515       for (y = iy0; y <= iy1; y++)
 516       {
 517          boolean in = FALSE;  /* are we inside the triangle? */
 518          int cx[7];
 519
 520          for (i = 0; i < nr_planes; i++)
 521             cx[i] = c[i];
 522
 523          for (x = ix0; x <= ix1; x++)
 524          {
 525             int out = 0;
 526             int partial = 0;
 527
 528             for (i = 0; i < nr_planes; i++) {
 529                int planeout = cx[i] + eo[i];
 530                int planepartial = cx[i] + ei[i] - 1;
 531                out |= (planeout >> 31);
 532                partial |= (planepartial >> 31) & (1<<i);
 533             }
 534
 535             if (out) {
 536                /* do nothing */
 537                if (in)
 538                   break;  /* exiting triangle, all done with this row */
 539                LP_COUNT(nr_empty_64);
 540             }
 541             else if (partial) {
 542                /* Not trivially accepted by at least one plane -
 543                 * rasterize/shade partial tile
 544                 */
 545                int count = util_bitcount(partial);
 546                in = TRUE;
 547                lp_scene_bin_command( scene, x, y,
 548                                      lp_rast_tri_tab[count],
 549                                      lp_rast_arg_triangle(tri, partial) );
 550
 551                LP_COUNT(nr_partially_covered_64);
 552             }
 553             else {
 554                /* triangle covers the whole tile- shade whole tile */
 555                LP_COUNT(nr_fully_covered_64);
 556                in = TRUE;
 557                if (variant->opaque &&
 558                    !setup->fb.zsbuf) {
 559                   lp_scene_bin_reset( scene, x, y );
 560                }
 561                lp_scene_bin_command( scene, x, y,
 562                                      lp_rast_shade_tile,
 563                                      lp_rast_arg_inputs(&tri->inputs) );
 564             }
 565
 566             /* Iterate cx values across the region:
 567              */
 568             for (i = 0; i < nr_planes; i++)
 569                cx[i] += xstep[i];
 570          }
 571
 572          /* Iterate c values down the region:
 573           */
 574          for (i = 0; i < nr_planes; i++)
 575             c[i] += ystep[i];
 576       }
 577    }
 578 }
 579
 580
 581 /**
 582  * Draw triangle if it's CW, cull otherwise.
 583  */
 584 static void triangle_cw( struct lp_setup_context *setup,
 585                          const float (*v0)[4],
 586                          const float (*v1)[4],
 587                          const float (*v2)[4] )
 588 {
 589    do_triangle_ccw( setup, v1, v0, v2, !setup->ccw_is_frontface );
 590 }
 591
 592
 593 /**
 594  * Draw triangle if it's CCW, cull otherwise.
 595  */
 596 static void triangle_ccw( struct lp_setup_context *setup,
 597                          const float (*v0)[4],
 598                          const float (*v1)[4],
 599                          const float (*v2)[4] )
 600 {
 601    do_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface );
 602 }
 603
 604
 605
 606 /**
 607  * Draw triangle whether it's CW or CCW.
 608  */
 609 static void triangle_both( struct lp_setup_context *setup,
 610                            const float (*v0)[4],
 611                            const float (*v1)[4],
 612                            const float (*v2)[4] )
 613 {
 614    /* edge vectors e = v0 - v2, f = v1 - v2 */
 615    const float ex = v0[0][0] - v2[0][0];
 616    const float ey = v0[0][1] - v2[0][1];
 617    const float fx = v1[0][0] - v2[0][0];
 618    const float fy = v1[0][1] - v2[0][1];
 619
 620    /* det = cross(e,f).z */
 621    const float det = ex * fy - ey * fx;
 622    if (det < 0.0f)
 623       triangle_ccw( setup, v0, v1, v2 );
 624    else if (det > 0.0f)
 625       triangle_cw( setup, v0, v1, v2 );
 626 }
 627
 628
 629 static void triangle_nop( struct lp_setup_context *setup,
 630                           const float (*v0)[4],
 631                           const float (*v1)[4],
 632                           const float (*v2)[4] )
 633 {
 634 }
 635
 636
 637 void
 638 lp_setup_choose_triangle( struct lp_setup_context *setup )
 639 {
 640    switch (setup->cullmode) {
 641    case PIPE_FACE_NONE:
 642       setup->triangle = triangle_both;
 643       break;
 644    case PIPE_FACE_BACK:
 645       setup->triangle = setup->ccw_is_frontface ? triangle_ccw : triangle_cw;
 646       break;
 647    case PIPE_FACE_FRONT:
 648       setup->triangle = setup->ccw_is_frontface ? triangle_cw : triangle_ccw;
 649       break;
 650    default:
 651       setup->triangle = triangle_nop;
 652       break;
 653    }
 654 }