src/gallium/drivers/llvmpipe/lp_rast_tri.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007-2009 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 #include <limits.h>
  33 #include "util/u_math.h"
  34 #include "lp_debug.h"
  35 #include "lp_perf.h"
  36 #include "lp_rast_priv.h"
  37
  38
  39
  40
  41 /**
  42  * Shade all pixels in a 4x4 block.
  43  */
  44 static void
  45 block_full_4(struct lp_rasterizer_task *task,
  46              const struct lp_rast_triangle *tri,
  47              int x, int y)
  48 {
  49    lp_rast_shade_quads_all(task, &tri->inputs, x, y);
  50 }
  51
  52
  53 /**
  54  * Shade all pixels in a 16x16 block.
  55  */
  56 static void
  57 block_full_16(struct lp_rasterizer_task *task,
  58               const struct lp_rast_triangle *tri,
  59               int x, int y)
  60 {
  61    unsigned ix, iy;
  62    assert(x % 16 == 0);
  63    assert(y % 16 == 0);
  64    for (iy = 0; iy < 16; iy += 4)
  65       for (ix = 0; ix < 16; ix += 4)
  66          block_full_4(task, tri, x + ix, y + iy);
  67 }
  68
  69 #if !defined(PIPE_ARCH_SSE)
  70
  71 static INLINE unsigned
  72 build_mask_linear(int c, int dcdx, int dcdy)
  73 {
  74    int mask = 0;
  75
  76    int c0 = c;
  77    int c1 = c0 + dcdy;
  78    int c2 = c1 + dcdy;
  79    int c3 = c2 + dcdy;
  80
  81    mask |= ((c0 + 0 * dcdx) >> 31) & (1 << 0);
  82    mask |= ((c0 + 1 * dcdx) >> 31) & (1 << 1);
  83    mask |= ((c0 + 2 * dcdx) >> 31) & (1 << 2);
  84    mask |= ((c0 + 3 * dcdx) >> 31) & (1 << 3);
  85    mask |= ((c1 + 0 * dcdx) >> 31) & (1 << 4);
  86    mask |= ((c1 + 1 * dcdx) >> 31) & (1 << 5);
  87    mask |= ((c1 + 2 * dcdx) >> 31) & (1 << 6);
  88    mask |= ((c1 + 3 * dcdx) >> 31) & (1 << 7);
  89    mask |= ((c2 + 0 * dcdx) >> 31) & (1 << 8);
  90    mask |= ((c2 + 1 * dcdx) >> 31) & (1 << 9);
  91    mask |= ((c2 + 2 * dcdx) >> 31) & (1 << 10);
  92    mask |= ((c2 + 3 * dcdx) >> 31) & (1 << 11);
  93    mask |= ((c3 + 0 * dcdx) >> 31) & (1 << 12);
  94    mask |= ((c3 + 1 * dcdx) >> 31) & (1 << 13);
  95    mask |= ((c3 + 2 * dcdx) >> 31) & (1 << 14);
  96    mask |= ((c3 + 3 * dcdx) >> 31) & (1 << 15);
  97
  98    return mask;
  99 }
 100
 101
 102 static INLINE void
 103 build_masks(int c,
 104             int cdiff,
 105             int dcdx,
 106             int dcdy,
 107             unsigned *outmask,
 108             unsigned *partmask)
 109 {
 110    *outmask |= build_mask_linear(c, dcdx, dcdy);
 111    *partmask |= build_mask_linear(c + cdiff, dcdx, dcdy);
 112 }
 113
 114 void
 115 lp_rast_triangle_3_16(struct lp_rasterizer_task *task,
 116                       const union lp_rast_cmd_arg arg)
 117 {
 118    union lp_rast_cmd_arg arg2;
 119    arg2.triangle.tri = arg.triangle.tri;
 120    arg2.triangle.plane_mask = (1<<3)-1;
 121    lp_rast_triangle_3(task, arg2);
 122 }
 123
 124 void
 125 lp_rast_triangle_4_16(struct lp_rasterizer_task *task,
 126                       const union lp_rast_cmd_arg arg)
 127 {
 128    union lp_rast_cmd_arg arg2;
 129    arg2.triangle.tri = arg.triangle.tri;
 130    arg2.triangle.plane_mask = (1<<4)-1;
 131    lp_rast_triangle_4(task, arg2);
 132 }
 133
 134 void
 135 lp_rast_triangle_3_4(struct lp_rasterizer_task *task,
 136                       const union lp_rast_cmd_arg arg)
 137 {
 138    lp_rast_triangle_3_16(task, arg);
 139 }
 140
 141 #else
 142 #include <emmintrin.h>
 143 #include "util/u_sse.h"
 144
 145
 146 static INLINE void
 147 build_masks(int c,
 148             int cdiff,
 149             int dcdx,
 150             int dcdy,
 151             unsigned *outmask,
 152             unsigned *partmask)
 153 {
 154    __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
 155    __m128i xdcdy = _mm_set1_epi32(dcdy);
 156
 157    /* Get values across the quad
 158     */
 159    __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
 160    __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
 161    __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
 162
 163    {
 164       __m128i cstep01, cstep23, result;
 165
 166       cstep01 = _mm_packs_epi32(cstep0, cstep1);
 167       cstep23 = _mm_packs_epi32(cstep2, cstep3);
 168       result = _mm_packs_epi16(cstep01, cstep23);
 169
 170       *outmask |= _mm_movemask_epi8(result);
 171    }
 172
 173
 174    {
 175       __m128i cio4 = _mm_set1_epi32(cdiff);
 176       __m128i cstep01, cstep23, result;
 177
 178       cstep0 = _mm_add_epi32(cstep0, cio4);
 179       cstep1 = _mm_add_epi32(cstep1, cio4);
 180       cstep2 = _mm_add_epi32(cstep2, cio4);
 181       cstep3 = _mm_add_epi32(cstep3, cio4);
 182
 183       cstep01 = _mm_packs_epi32(cstep0, cstep1);
 184       cstep23 = _mm_packs_epi32(cstep2, cstep3);
 185       result = _mm_packs_epi16(cstep01, cstep23);
 186
 187       *partmask |= _mm_movemask_epi8(result);
 188    }
 189 }
 190
 191
 192 static INLINE unsigned
 193 build_mask_linear(int c, int dcdx, int dcdy)
 194 {
 195    __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
 196    __m128i xdcdy = _mm_set1_epi32(dcdy);
 197
 198    /* Get values across the quad
 199     */
 200    __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
 201    __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
 202    __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
 203
 204    /* pack pairs of results into epi16
 205     */
 206    __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
 207    __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
 208
 209    /* pack into epi8, preserving sign bits
 210     */
 211    __m128i result = _mm_packs_epi16(cstep01, cstep23);
 212
 213    /* extract sign bits to create mask
 214     */
 215    return _mm_movemask_epi8(result);
 216 }
 217
 218 static INLINE unsigned
 219 sign_bits4(const __m128i *cstep, int cdiff)
 220 {
 221
 222    /* Adjust the step values
 223     */
 224    __m128i cio4 = _mm_set1_epi32(cdiff);
 225    __m128i cstep0 = _mm_add_epi32(cstep[0], cio4);
 226    __m128i cstep1 = _mm_add_epi32(cstep[1], cio4);
 227    __m128i cstep2 = _mm_add_epi32(cstep[2], cio4);
 228    __m128i cstep3 = _mm_add_epi32(cstep[3], cio4);
 229
 230    /* Pack down to epi8
 231     */
 232    __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
 233    __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
 234    __m128i result = _mm_packs_epi16(cstep01, cstep23);
 235
 236    /* Extract the sign bits
 237     */
 238    return _mm_movemask_epi8(result);
 239 }
 240
 241
 242 #define NR_PLANES 3
 243
 244
 245
 246
 247
 248
 249
 250 void
 251 lp_rast_triangle_3_16(struct lp_rasterizer_task *task,
 252                       const union lp_rast_cmd_arg arg)
 253 {
 254    const struct lp_rast_triangle *tri = arg.triangle.tri;
 255    const struct lp_rast_plane *plane = GET_PLANES(tri);
 256    int x = (arg.triangle.plane_mask & 0xff) + task->x;
 257    int y = (arg.triangle.plane_mask >> 8) + task->y;
 258    unsigned i, j;
 259
 260    struct { unsigned mask:16; unsigned i:8; unsigned j:8; } out[16];
 261    unsigned nr = 0;
 262
 263    __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* c, dcdx, dcdy, eo */
 264    __m128i p1 = _mm_load_si128((__m128i *)&plane[1]); /* c, dcdx, dcdy, eo */
 265    __m128i p2 = _mm_load_si128((__m128i *)&plane[2]); /* c, dcdx, dcdy, eo */
 266    __m128i zero = _mm_setzero_si128();
 267
 268    __m128i c;
 269    __m128i dcdx;
 270    __m128i dcdy;
 271    __m128i rej4;
 272
 273    __m128i dcdx2;
 274    __m128i dcdx3;
 275
 276    __m128i span_0;                /* 0,dcdx,2dcdx,3dcdx for plane 0 */
 277    __m128i span_1;                /* 0,dcdx,2dcdx,3dcdx for plane 1 */
 278    __m128i span_2;                /* 0,dcdx,2dcdx,3dcdx for plane 2 */
 279    __m128i unused;
 280
 281    transpose4_epi32(&p0, &p1, &p2, &zero,
 282                     &c, &dcdx, &dcdy, &rej4);
 283
 284    /* Adjust dcdx;
 285     */
 286    dcdx = _mm_sub_epi32(zero, dcdx);
 287
 288    c = _mm_add_epi32(c, mm_mullo_epi32(dcdx, _mm_set1_epi32(x)));
 289    c = _mm_add_epi32(c, mm_mullo_epi32(dcdy, _mm_set1_epi32(y)));
 290    rej4 = _mm_slli_epi32(rej4, 2);
 291
 292    /* Adjust so we can just check the sign bit (< 0 comparison), instead of having to do a less efficient <= 0 comparison */
 293    c = _mm_sub_epi32(c, _mm_set1_epi32(1));
 294    rej4 = _mm_add_epi32(rej4, _mm_set1_epi32(1));
 295
 296    dcdx2 = _mm_add_epi32(dcdx, dcdx);
 297    dcdx3 = _mm_add_epi32(dcdx2, dcdx);
 298
 299    transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
 300                     &span_0, &span_1, &span_2, &unused);
 301
 302    for (i = 0; i < 4; i++) {
 303       __m128i cx = c;
 304
 305       for (j = 0; j < 4; j++) {
 306          __m128i c4rej = _mm_add_epi32(cx, rej4);
 307          __m128i rej_masks = _mm_srai_epi32(c4rej, 31);
 308
 309          /* if (is_zero(rej_masks)) */
 310          if (_mm_movemask_epi8(rej_masks) == 0) {
 311             __m128i c0_0 = _mm_add_epi32(SCALAR_EPI32(cx, 0), span_0);
 312             __m128i c1_0 = _mm_add_epi32(SCALAR_EPI32(cx, 1), span_1);
 313             __m128i c2_0 = _mm_add_epi32(SCALAR_EPI32(cx, 2), span_2);
 314
 315             __m128i c_0 = _mm_or_si128(_mm_or_si128(c0_0, c1_0), c2_0);
 316
 317             __m128i c0_1 = _mm_add_epi32(c0_0, SCALAR_EPI32(dcdy, 0));
 318             __m128i c1_1 = _mm_add_epi32(c1_0, SCALAR_EPI32(dcdy, 1));
 319             __m128i c2_1 = _mm_add_epi32(c2_0, SCALAR_EPI32(dcdy, 2));
 320
 321             __m128i c_1 = _mm_or_si128(_mm_or_si128(c0_1, c1_1), c2_1);
 322             __m128i c_01 = _mm_packs_epi32(c_0, c_1);
 323
 324             __m128i c0_2 = _mm_add_epi32(c0_1, SCALAR_EPI32(dcdy, 0));
 325             __m128i c1_2 = _mm_add_epi32(c1_1, SCALAR_EPI32(dcdy, 1));
 326             __m128i c2_2 = _mm_add_epi32(c2_1, SCALAR_EPI32(dcdy, 2));
 327
 328             __m128i c_2 = _mm_or_si128(_mm_or_si128(c0_2, c1_2), c2_2);
 329
 330             __m128i c0_3 = _mm_add_epi32(c0_2, SCALAR_EPI32(dcdy, 0));
 331             __m128i c1_3 = _mm_add_epi32(c1_2, SCALAR_EPI32(dcdy, 1));
 332             __m128i c2_3 = _mm_add_epi32(c2_2, SCALAR_EPI32(dcdy, 2));
 333
 334             __m128i c_3 = _mm_or_si128(_mm_or_si128(c0_3, c1_3), c2_3);
 335             __m128i c_23 = _mm_packs_epi32(c_2, c_3);
 336             __m128i c_0123 = _mm_packs_epi16(c_01, c_23);
 337
 338             unsigned mask = _mm_movemask_epi8(c_0123);
 339
 340             out[nr].i = i;
 341             out[nr].j = j;
 342             out[nr].mask = mask;
 343             if (mask != 0xffff)
 344                nr++;
 345          }
 346          cx = _mm_add_epi32(cx, _mm_slli_epi32(dcdx, 2));
 347       }
 348
 349       c = _mm_add_epi32(c, _mm_slli_epi32(dcdy, 2));
 350    }
 351
 352    for (i = 0; i < nr; i++)
 353       lp_rast_shade_quads_mask(task,
 354                                &tri->inputs,
 355                                x + 4 * out[i].j,
 356                                y + 4 * out[i].i,
 357                                0xffff & ~out[i].mask);
 358 }
 359
 360
 361
 362
 363
 364 void
 365 lp_rast_triangle_3_4(struct lp_rasterizer_task *task,
 366                      const union lp_rast_cmd_arg arg)
 367 {
 368    const struct lp_rast_triangle *tri = arg.triangle.tri;
 369    const struct lp_rast_plane *plane = GET_PLANES(tri);
 370    unsigned x = (arg.triangle.plane_mask & 0xff) + task->x;
 371    unsigned y = (arg.triangle.plane_mask >> 8) + task->y;
 372
 373    __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* c, dcdx, dcdy, eo */
 374    __m128i p1 = _mm_load_si128((__m128i *)&plane[1]); /* c, dcdx, dcdy, eo */
 375    __m128i p2 = _mm_load_si128((__m128i *)&plane[2]); /* c, dcdx, dcdy, eo */
 376    __m128i zero = _mm_setzero_si128();
 377
 378    __m128i c;
 379    __m128i dcdx;
 380    __m128i dcdy;
 381
 382    __m128i dcdx2;
 383    __m128i dcdx3;
 384
 385    __m128i span_0;                /* 0,dcdx,2dcdx,3dcdx for plane 0 */
 386    __m128i span_1;                /* 0,dcdx,2dcdx,3dcdx for plane 1 */
 387    __m128i span_2;                /* 0,dcdx,2dcdx,3dcdx for plane 2 */
 388    __m128i unused;
 389
 390    transpose4_epi32(&p0, &p1, &p2, &zero,
 391                     &c, &dcdx, &dcdy, &unused);
 392
 393    /* Adjust dcdx;
 394     */
 395    dcdx = _mm_sub_epi32(zero, dcdx);
 396
 397    c = _mm_add_epi32(c, mm_mullo_epi32(dcdx, _mm_set1_epi32(x)));
 398    c = _mm_add_epi32(c, mm_mullo_epi32(dcdy, _mm_set1_epi32(y)));
 399
 400    /* Adjust so we can just check the sign bit (< 0 comparison), instead of having to do a less efficient <= 0 comparison */
 401    c = _mm_sub_epi32(c, _mm_set1_epi32(1));
 402
 403    dcdx2 = _mm_add_epi32(dcdx, dcdx);
 404    dcdx3 = _mm_add_epi32(dcdx2, dcdx);
 405
 406    transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
 407                     &span_0, &span_1, &span_2, &unused);
 408
 409
 410    {
 411       __m128i c0_0 = _mm_add_epi32(SCALAR_EPI32(c, 0), span_0);
 412       __m128i c1_0 = _mm_add_epi32(SCALAR_EPI32(c, 1), span_1);
 413       __m128i c2_0 = _mm_add_epi32(SCALAR_EPI32(c, 2), span_2);
 414
 415       __m128i c_0 = _mm_or_si128(_mm_or_si128(c0_0, c1_0), c2_0);
 416
 417       __m128i c0_1 = _mm_add_epi32(c0_0, SCALAR_EPI32(dcdy, 0));
 418       __m128i c1_1 = _mm_add_epi32(c1_0, SCALAR_EPI32(dcdy, 1));
 419       __m128i c2_1 = _mm_add_epi32(c2_0, SCALAR_EPI32(dcdy, 2));
 420
 421       __m128i c_1 = _mm_or_si128(_mm_or_si128(c0_1, c1_1), c2_1);
 422       __m128i c_01 = _mm_packs_epi32(c_0, c_1);
 423
 424       __m128i c0_2 = _mm_add_epi32(c0_1, SCALAR_EPI32(dcdy, 0));
 425       __m128i c1_2 = _mm_add_epi32(c1_1, SCALAR_EPI32(dcdy, 1));
 426       __m128i c2_2 = _mm_add_epi32(c2_1, SCALAR_EPI32(dcdy, 2));
 427
 428       __m128i c_2 = _mm_or_si128(_mm_or_si128(c0_2, c1_2), c2_2);
 429
 430       __m128i c0_3 = _mm_add_epi32(c0_2, SCALAR_EPI32(dcdy, 0));
 431       __m128i c1_3 = _mm_add_epi32(c1_2, SCALAR_EPI32(dcdy, 1));
 432       __m128i c2_3 = _mm_add_epi32(c2_2, SCALAR_EPI32(dcdy, 2));
 433
 434       __m128i c_3 = _mm_or_si128(_mm_or_si128(c0_3, c1_3), c2_3);
 435       __m128i c_23 = _mm_packs_epi32(c_2, c_3);
 436       __m128i c_0123 = _mm_packs_epi16(c_01, c_23);
 437
 438       unsigned mask = _mm_movemask_epi8(c_0123);
 439
 440       if (mask != 0xffff)
 441          lp_rast_shade_quads_mask(task,
 442                                   &tri->inputs,
 443                                   x,
 444                                   y,
 445                                   0xffff & ~mask);
 446    }
 447 }
 448
 449 #undef NR_PLANES
 450 #endif
 451
 452
 453
 454
 455 #define TAG(x) x##_1
 456 #define NR_PLANES 1
 457 #include "lp_rast_tri_tmp.h"
 458
 459 #define TAG(x) x##_2
 460 #define NR_PLANES 2
 461 #include "lp_rast_tri_tmp.h"
 462
 463 #define TAG(x) x##_3
 464 #define NR_PLANES 3
 465 /*#define TRI_4 lp_rast_triangle_3_4*/
 466 /*#define TRI_16 lp_rast_triangle_3_16*/
 467 #include "lp_rast_tri_tmp.h"
 468
 469 #define TAG(x) x##_4
 470 #define NR_PLANES 4
 471 #define TRI_16 lp_rast_triangle_4_16
 472 #include "lp_rast_tri_tmp.h"
 473
 474 #define TAG(x) x##_5
 475 #define NR_PLANES 5
 476 #include "lp_rast_tri_tmp.h"
 477
 478 #define TAG(x) x##_6
 479 #define NR_PLANES 6
 480 #include "lp_rast_tri_tmp.h"
 481
 482 #define TAG(x) x##_7
 483 #define NR_PLANES 7
 484 #include "lp_rast_tri_tmp.h"
 485
 486 #define TAG(x) x##_8
 487 #define NR_PLANES 8
 488 #include "lp_rast_tri_tmp.h"
 489