src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h

   1 /**************************************************************************
   2  *
   3  * Copyright 2007-2010 VMware, Inc.
   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 VMWARE 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  * Rasterization for binned triangles within a tile
  30  */
  31
  32
  33
  34 /**
  35  * Prototype for a 8 plane rasterizer function.  Will codegenerate
  36  * several of these.
  37  *
  38  * XXX: Varients for more/fewer planes.
  39  * XXX: Need ways of dropping planes as we descend.
  40  * XXX: SIMD
  41  */
  42 static void
  43 TAG(do_block_4)(struct lp_rasterizer_task *task,
  44                 const struct lp_rast_triangle *tri,
  45                 const struct lp_rast_plane *plane,
  46                 int x, int y,
  47                 const int64_t *c)
  48 {
  49    int j;
  50 #ifndef MULTISAMPLE
  51    unsigned mask = 0xffff;
  52 #else
  53    uint64_t mask = UINT64_MAX;
  54 #endif
  55
  56    for (j = 0; j < NR_PLANES; j++) {
  57 #ifndef MULTISAMPLE
  58 #ifdef RASTER_64
  59       mask &= ~BUILD_MASK_LINEAR(((c[j] - 1) >> (int64_t)FIXED_ORDER),
  60                                  -plane[j].dcdx >> FIXED_ORDER,
  61                                  plane[j].dcdy >> FIXED_ORDER);
  62 #else
  63       mask &= ~BUILD_MASK_LINEAR((c[j] - 1),
  64                                  -plane[j].dcdx,
  65                                  plane[j].dcdy);
  66 #endif
  67 #else
  68       for (unsigned s = 0; s < 4; s++) {
  69          int64_t new_c = (c[j]) + ((IMUL64(task->scene->fixed_sample_pos[s][1], plane[j].dcdy) + IMUL64(task->scene->fixed_sample_pos[s][0], -plane[j].dcdx)) >> FIXED_ORDER);
  70          uint32_t build_mask;
  71 #ifdef RASTER_64
  72          build_mask = BUILD_MASK_LINEAR((int32_t)((new_c - 1) >> (int64_t)FIXED_ORDER),
  73                                         -plane[j].dcdx >> FIXED_ORDER,
  74                                         plane[j].dcdy >> FIXED_ORDER);
  75 #else
  76          build_mask = BUILD_MASK_LINEAR((new_c - 1),
  77                                         -plane[j].dcdx,
  78                                         plane[j].dcdy);
  79 #endif
  80          mask &= ~((uint64_t)build_mask << (s * 16));
  81       }
  82 #endif
  83    }
  84
  85    /* Now pass to the shader:
  86     */
  87    if (mask)
  88       lp_rast_shade_quads_mask_sample(task, &tri->inputs, x, y, mask);
  89 }
  90
  91 /**
  92  * Evaluate a 16x16 block of pixels to determine which 4x4 subblocks are in/out
  93  * of the triangle's bounds.
  94  */
  95 static void
  96 TAG(do_block_16)(struct lp_rasterizer_task *task,
  97                  const struct lp_rast_triangle *tri,
  98                  const struct lp_rast_plane *plane,
  99                  int x, int y,
 100                  const int64_t *c)
 101 {
 102    unsigned outmask, inmask, partmask, partial_mask;
 103    unsigned j;
 104
 105    outmask = 0;                 /* outside one or more trivial reject planes */
 106    partmask = 0;                /* outside one or more trivial accept planes */
 107
 108    for (j = 0; j < NR_PLANES; j++) {
 109 #ifdef RASTER_64
 110       int32_t dcdx = -plane[j].dcdx >> FIXED_ORDER;
 111       int32_t dcdy = plane[j].dcdy >> FIXED_ORDER;
 112       const int32_t cox = plane[j].eo >> FIXED_ORDER;
 113       const int32_t ei = (dcdy + dcdx - cox) << 2;
 114       const int32_t cox_s = cox << 2;
 115       const int32_t co = (int32_t)(c[j] >> (int64_t)FIXED_ORDER) + cox_s;
 116       int32_t cdiff;
 117       cdiff = ei - cox_s + ((int32_t)((c[j] - 1) >> (int64_t)FIXED_ORDER) -
 118                             (int32_t)(c[j] >> (int64_t)FIXED_ORDER));
 119       dcdx <<= 2;
 120       dcdy <<= 2;
 121 #else
 122       const int64_t dcdx = -IMUL64(plane[j].dcdx, 4);
 123       const int64_t dcdy = IMUL64(plane[j].dcdy, 4);
 124       const int64_t cox = IMUL64(plane[j].eo, 4);
 125       const int32_t ei = plane[j].dcdy - plane[j].dcdx - (int64_t)plane[j].eo;
 126       const int64_t cio = IMUL64(ei, 4) - 1;
 127       int32_t co, cdiff;
 128       co = c[j] + cox;
 129       cdiff = cio - cox;
 130 #endif
 131
 132       BUILD_MASKS(co, cdiff,
 133                   dcdx, dcdy,
 134                   &outmask,   /* sign bits from c[i][0..15] + cox */
 135                   &partmask); /* sign bits from c[i][0..15] + cio */
 136    }
 137
 138    if (outmask == 0xffff)
 139       return;
 140
 141    /* Mask of sub-blocks which are inside all trivial accept planes:
 142     */
 143    inmask = ~partmask & 0xffff;
 144
 145    /* Mask of sub-blocks which are inside all trivial reject planes,
 146     * but outside at least one trivial accept plane:
 147     */
 148    partial_mask = partmask & ~outmask;
 149
 150    assert((partial_mask & inmask) == 0);
 151
 152    LP_COUNT_ADD(nr_empty_4, util_bitcount(0xffff & ~(partial_mask | inmask)));
 153
 154    /* Iterate over partials:
 155     */
 156    while (partial_mask) {
 157       int i = ffs(partial_mask) - 1;
 158       int ix = (i & 3) * 4;
 159       int iy = (i >> 2) * 4;
 160       int px = x + ix;
 161       int py = y + iy;
 162       int64_t cx[NR_PLANES];
 163
 164       partial_mask &= ~(1 << i);
 165
 166       LP_COUNT(nr_partially_covered_4);
 167
 168       for (j = 0; j < NR_PLANES; j++)
 169          cx[j] = (c[j]
 170                   - IMUL64(plane[j].dcdx, ix)
 171                   + IMUL64(plane[j].dcdy, iy));
 172
 173       TAG(do_block_4)(task, tri, plane, px, py, cx);
 174    }
 175
 176    /* Iterate over fulls:
 177     */
 178    while (inmask) {
 179       int i = ffs(inmask) - 1;
 180       int ix = (i & 3) * 4;
 181       int iy = (i >> 2) * 4;
 182       int px = x + ix;
 183       int py = y + iy;
 184
 185       inmask &= ~(1 << i);
 186
 187       LP_COUNT(nr_fully_covered_4);
 188       block_full_4(task, tri, px, py);
 189    }
 190 }
 191
 192
 193 /**
 194  * Scan the tile in chunks and figure out which pixels to rasterize
 195  * for this triangle.
 196  */
 197 void
 198 TAG(lp_rast_triangle)(struct lp_rasterizer_task *task,
 199                       const union lp_rast_cmd_arg arg)
 200 {
 201    const struct lp_rast_triangle *tri = arg.triangle.tri;
 202    unsigned plane_mask = arg.triangle.plane_mask;
 203    const struct lp_rast_plane *tri_plane = GET_PLANES(tri);
 204    const int x = task->x, y = task->y;
 205    struct lp_rast_plane plane[NR_PLANES];
 206    int64_t c[NR_PLANES];
 207    unsigned outmask, inmask, partmask, partial_mask;
 208    unsigned j = 0;
 209
 210    if (tri->inputs.disable) {
 211       /* This triangle was partially binned and has been disabled */
 212       return;
 213    }
 214
 215    outmask = 0;                 /* outside one or more trivial reject planes */
 216    partmask = 0;                /* outside one or more trivial accept planes */
 217
 218    while (plane_mask) {
 219       int i = ffs(plane_mask) - 1;
 220       plane[j] = tri_plane[i];
 221       plane_mask &= ~(1 << i);
 222       c[j] = plane[j].c + IMUL64(plane[j].dcdy, y) - IMUL64(plane[j].dcdx, x);
 223
 224       {
 225 #ifdef RASTER_64
 226          /*
 227           * Strip off lower FIXED_ORDER bits. Note that those bits from
 228           * dcdx, dcdy, eo are always 0 (by definition).
 229           * c values, however, are not. This means that for every
 230           * addition of the form c + n*dcdx the lower FIXED_ORDER bits will
 231           * NOT change. And those bits are not relevant to the sign bit (which
 232           * is only what we need!) that is,
 233           * sign(c + n*dcdx) == sign((c >> FIXED_ORDER) + n*(dcdx >> FIXED_ORDER))
 234           * This means we can get away with using 32bit math for the most part.
 235           * Only tricky part is the -1 adjustment for cdiff.
 236           */
 237          int32_t dcdx = -plane[j].dcdx >> FIXED_ORDER;
 238          int32_t dcdy = plane[j].dcdy >> FIXED_ORDER;
 239          const int32_t cox = plane[j].eo >> FIXED_ORDER;
 240          const int32_t ei = (dcdy + dcdx - cox) << 4;
 241          const int32_t cox_s = cox << 4;
 242          const int32_t co = (int32_t)(c[j] >> (int64_t)FIXED_ORDER) + cox_s;
 243          int32_t cdiff;
 244          /*
 245           * Plausibility check to ensure the 32bit math works.
 246           * Note that within a tile, the max we can move the edge function
 247           * is essentially dcdx * TILE_SIZE + dcdy * TILE_SIZE.
 248           * TILE_SIZE is 64, dcdx/dcdy are nominally 21 bit (for 8192 max size
 249           * and 8 subpixel bits), I'd be happy with 2 bits more too (1 for
 250           * increasing fb size to 16384, the required d3d11 value, another one
 251           * because I'm not quite sure we can't be _just_ above the max value
 252           * here). This gives us 30 bits max - hence if c would exceed that here
 253           * that means the plane is either trivial reject for the whole tile
 254           * (in which case the tri will not get binned), or trivial accept for
 255           * the whole tile (in which case plane_mask will not include it).
 256           */
 257          assert((c[j] >> (int64_t)FIXED_ORDER) > (int32_t)0xb0000000 &&
 258                 (c[j] >> (int64_t)FIXED_ORDER) < (int32_t)0x3fffffff);
 259          /*
 260           * Note the fixup part is constant throughout the tile - thus could
 261           * just calculate this and avoid _all_ 64bit math in rasterization
 262           * (except exactly this fixup calc).
 263           * In fact theoretically could move that even to setup, albeit that
 264           * seems tricky (pre-bin certainly can have values larger than 32bit,
 265           * and would need to communicate that fixup value through).
 266           * And if we want to support msaa, we'd probably don't want to do the
 267           * downscaling in setup in any case...
 268           */
 269          cdiff = ei - cox_s + ((int32_t)((c[j] - 1) >> (int64_t)FIXED_ORDER) -
 270                                (int32_t)(c[j] >> (int64_t)FIXED_ORDER));
 271          dcdx <<= 4;
 272          dcdy <<= 4;
 273 #else
 274          const int32_t dcdx = -plane[j].dcdx << 4;
 275          const int32_t dcdy = plane[j].dcdy << 4;
 276          const int32_t cox = plane[j].eo << 4;
 277          const int32_t ei = plane[j].dcdy - plane[j].dcdx - (int32_t)plane[j].eo;
 278          const int32_t cio = (ei << 4) - 1;
 279          int32_t co, cdiff;
 280          co = c[j] + cox;
 281          cdiff = cio - cox;
 282 #endif
 283          BUILD_MASKS(co, cdiff,
 284                      dcdx, dcdy,
 285                      &outmask,   /* sign bits from c[i][0..15] + cox */
 286                      &partmask); /* sign bits from c[i][0..15] + cio */
 287       }
 288
 289       j++;
 290    }
 291
 292    if (outmask == 0xffff)
 293       return;
 294
 295    /* Mask of sub-blocks which are inside all trivial accept planes:
 296     */
 297    inmask = ~partmask & 0xffff;
 298
 299    /* Mask of sub-blocks which are inside all trivial reject planes,
 300     * but outside at least one trivial accept plane:
 301     */
 302    partial_mask = partmask & ~outmask;
 303
 304    assert((partial_mask & inmask) == 0);
 305
 306    LP_COUNT_ADD(nr_empty_16, util_bitcount(0xffff & ~(partial_mask | inmask)));
 307
 308    /* Iterate over partials:
 309     */
 310    while (partial_mask) {
 311       int i = ffs(partial_mask) - 1;
 312       int ix = (i & 3) * 16;
 313       int iy = (i >> 2) * 16;
 314       int px = x + ix;
 315       int py = y + iy;
 316       int64_t cx[NR_PLANES];
 317
 318       for (j = 0; j < NR_PLANES; j++)
 319          cx[j] = (c[j]
 320                   - IMUL64(plane[j].dcdx, ix)
 321                   + IMUL64(plane[j].dcdy, iy));
 322
 323       partial_mask &= ~(1 << i);
 324
 325       LP_COUNT(nr_partially_covered_16);
 326       TAG(do_block_16)(task, tri, plane, px, py, cx);
 327    }
 328
 329    /* Iterate over fulls:
 330     */
 331    while (inmask) {
 332       int i = ffs(inmask) - 1;
 333       int ix = (i & 3) * 16;
 334       int iy = (i >> 2) * 16;
 335       int px = x + ix;
 336       int py = y + iy;
 337
 338       inmask &= ~(1 << i);
 339
 340       LP_COUNT(nr_fully_covered_16);
 341       block_full_16(task, tri, px, py);
 342    }
 343 }
 344
 345 #if defined(PIPE_ARCH_SSE) && defined(TRI_16)
 346 /* XXX: special case this when intersection is not required.
 347  *      - tile completely within bbox,
 348  *      - bbox completely within tile.
 349  */
 350 void
 351 TRI_16(struct lp_rasterizer_task *task,
 352        const union lp_rast_cmd_arg arg)
 353 {
 354    const struct lp_rast_triangle *tri = arg.triangle.tri;
 355    const struct lp_rast_plane *plane = GET_PLANES(tri);
 356    unsigned mask = arg.triangle.plane_mask;
 357    unsigned outmask, partial_mask;
 358    unsigned j;
 359    __m128i cstep4[NR_PLANES][4];
 360
 361    int x = (mask & 0xff);
 362    int y = (mask >> 8);
 363
 364    outmask = 0;                 /* outside one or more trivial reject planes */
 365
 366    x += task->x;
 367    y += task->y;
 368
 369    for (j = 0; j < NR_PLANES; j++) {
 370       const int dcdx = -plane[j].dcdx * 4;
 371       const int dcdy = plane[j].dcdy * 4;
 372       __m128i xdcdy = _mm_set1_epi32(dcdy);
 373
 374       cstep4[j][0] = _mm_setr_epi32(0, dcdx, dcdx*2, dcdx*3);
 375       cstep4[j][1] = _mm_add_epi32(cstep4[j][0], xdcdy);
 376       cstep4[j][2] = _mm_add_epi32(cstep4[j][1], xdcdy);
 377       cstep4[j][3] = _mm_add_epi32(cstep4[j][2], xdcdy);
 378
 379       {
 380          const int c = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x;
 381          const int cox = plane[j].eo * 4;
 382
 383          outmask |= sign_bits4(cstep4[j], c + cox);
 384       }
 385    }
 386
 387    if (outmask == 0xffff)
 388       return;
 389
 390
 391    /* Mask of sub-blocks which are inside all trivial reject planes,
 392     * but outside at least one trivial accept plane:
 393     */
 394    partial_mask = 0xffff & ~outmask;
 395
 396    /* Iterate over partials:
 397     */
 398    while (partial_mask) {
 399       int i = ffs(partial_mask) - 1;
 400       int ix = (i & 3) * 4;
 401       int iy = (i >> 2) * 4;
 402       int px = x + ix;
 403       int py = y + iy;
 404       unsigned mask = 0xffff;
 405
 406       partial_mask &= ~(1 << i);
 407
 408       for (j = 0; j < NR_PLANES; j++) {
 409          const int cx = (plane[j].c - 1
 410                          - plane[j].dcdx * px
 411                          + plane[j].dcdy * py) * 4;
 412
 413          mask &= ~sign_bits4(cstep4[j], cx);
 414       }
 415
 416       if (mask)
 417          lp_rast_shade_quads_mask(task, &tri->inputs, px, py, mask);
 418    }
 419 }
 420 #endif
 421
 422 #if defined(PIPE_ARCH_SSE) && defined(TRI_4)
 423 void
 424 TRI_4(struct lp_rasterizer_task *task,
 425       const union lp_rast_cmd_arg arg)
 426 {
 427    const struct lp_rast_triangle *tri = arg.triangle.tri;
 428    const struct lp_rast_plane *plane = GET_PLANES(tri);
 429    unsigned mask = arg.triangle.plane_mask;
 430    const int x = task->x + (mask & 0xff);
 431    const int y = task->y + (mask >> 8);
 432    unsigned j;
 433
 434    /* Iterate over partials:
 435     */
 436    {
 437       unsigned mask = 0xffff;
 438
 439       for (j = 0; j < NR_PLANES; j++) {
 440          const int cx = (plane[j].c
 441                          - plane[j].dcdx * x
 442                          + plane[j].dcdy * y);
 443
 444          const int dcdx = -plane[j].dcdx;
 445          const int dcdy = plane[j].dcdy;
 446          __m128i xdcdy = _mm_set1_epi32(dcdy);
 447
 448          __m128i cstep0 = _mm_setr_epi32(cx, cx + dcdx, cx + dcdx*2, cx + dcdx*3);
 449          __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
 450          __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
 451          __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
 452
 453          __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
 454          __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
 455          __m128i result = _mm_packs_epi16(cstep01, cstep23);
 456
 457          /* Extract the sign bits
 458           */
 459          mask &= ~_mm_movemask_epi8(result);
 460       }
 461
 462       if (mask)
 463          lp_rast_shade_quads_mask(task, &tri->inputs, x, y, mask);
 464    }
 465 }
 466 #endif
 467
 468
 469
 470 #undef TAG
 471 #undef TRI_4
 472 #undef TRI_16
 473 #undef NR_PLANES
 474