src/gallium/drivers/llvmpipe/lp_tile_soa.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  * RGBA/float tile get/put functions.
  30  * Usable both by drivers and state trackers.
  31  */
  32
  33
  34 #include "pipe/p_defines.h"
  35 #include "pipe/p_inlines.h"
  36
  37 #include "util/u_math.h"
  38 #include "util/u_memory.h"
  39 #include "util/u_rect.h"
  40 #include "util/u_tile.h"
  41 #include "lp_tile_cache.h"
  42 #include "lp_tile_soa.h"
  43
  44
  45 #define PIXEL(_p, _x, _y, _c) ((_p)[(_c)*TILE_SIZE*TILE_SIZE + (_y)*TILE_SIZE + (_x)])
  46
  47
  48
  49 /*** PIPE_FORMAT_A8R8G8B8_UNORM ***/
  50
  51 static void
  52 a8r8g8b8_get_tile_rgba(const unsigned *src,
  53                        unsigned w, unsigned h,
  54                        uint8_t *p)
  55 {
  56    unsigned i, j;
  57
  58    for (i = 0; i < h; i++) {
  59       for (j = 0; j < w; j++) {
  60          const unsigned pixel = *src++;
  61          PIXEL(p, j, i, 0) = (pixel >> 16) & 0xff;
  62          PIXEL(p, j, i, 1) = (pixel >>  8) & 0xff;
  63          PIXEL(p, j, i, 2) = (pixel >>  0) & 0xff;
  64          PIXEL(p, j, i, 3) = (pixel >> 24) & 0xff;
  65       }
  66    }
  67 }
  68
  69
  70 static void
  71 a8r8g8b8_put_tile_rgba(unsigned *dst,
  72                        unsigned w, unsigned h,
  73                        const uint8_t *p)
  74 {
  75    unsigned i, j;
  76
  77    for (i = 0; i < h; i++) {
  78       for (j = 0; j < w; j++) {
  79          unsigned r, g, b, a;
  80          r = PIXEL(p, j, i, 0);
  81          g = PIXEL(p, j, i, 1);
  82          b = PIXEL(p, j, i, 2);
  83          a = PIXEL(p, j, i, 3);
  84          *dst++ = (a << 24) | (r << 16) | (g << 8) | b;
  85       }
  86    }
  87 }
  88
  89
  90 /*** PIPE_FORMAT_A8R8G8B8_UNORM ***/
  91
  92 static void
  93 x8r8g8b8_get_tile_rgba(const unsigned *src,
  94                        unsigned w, unsigned h,
  95                        uint8_t *p)
  96 {
  97    unsigned i, j;
  98
  99    for (i = 0; i < h; i++) {
 100       for (j = 0; j < w; j++) {
 101          const unsigned pixel = *src++;
 102          PIXEL(p, j, i, 0) = (pixel >> 16) & 0xff;
 103          PIXEL(p, j, i, 1) = (pixel >>  8) & 0xff;
 104          PIXEL(p, j, i, 2) = (pixel >>  0) & 0xff;
 105          PIXEL(p, j, i, 3) = 0xff;
 106       }
 107    }
 108 }
 109
 110
 111 static void
 112 x8r8g8b8_put_tile_rgba(unsigned *dst,
 113                        unsigned w, unsigned h,
 114                        const uint8_t *p)
 115 {
 116    unsigned i, j;
 117
 118    for (i = 0; i < h; i++) {
 119       for (j = 0; j < w; j++) {
 120          unsigned r, g, b;
 121          r = PIXEL(p, j, i, 0);
 122          g = PIXEL(p, j, i, 1);
 123          b = PIXEL(p, j, i, 2);
 124          *dst++ = (0xff << 24) | (r << 16) | (g << 8) | b;
 125       }
 126    }
 127 }
 128
 129
 130 /*** PIPE_FORMAT_B8G8R8A8_UNORM ***/
 131
 132 static void
 133 b8g8r8a8_get_tile_rgba(const unsigned *src,
 134                        unsigned w, unsigned h,
 135                        uint8_t *p)
 136 {
 137    unsigned i, j;
 138
 139    for (i = 0; i < h; i++) {
 140       for (j = 0; j < w; j++) {
 141          const unsigned pixel = *src++;
 142          PIXEL(p, j, i, 0) = (pixel >>  8) & 0xff;
 143          PIXEL(p, j, i, 1) = (pixel >> 16) & 0xff;
 144          PIXEL(p, j, i, 2) = (pixel >> 24) & 0xff;
 145          PIXEL(p, j, i, 3) = (pixel >>  0) & 0xff;
 146       }
 147    }
 148 }
 149
 150
 151 static void
 152 b8g8r8a8_put_tile_rgba(unsigned *dst,
 153                        unsigned w, unsigned h,
 154                        const uint8_t *p)
 155 {
 156    unsigned i, j;
 157
 158    for (i = 0; i < h; i++) {
 159       for (j = 0; j < w; j++) {
 160          unsigned r, g, b, a;
 161          r = PIXEL(p, j, i, 0);
 162          g = PIXEL(p, j, i, 1);
 163          b = PIXEL(p, j, i, 2);
 164          a = PIXEL(p, j, i, 3);
 165          *dst++ = (b << 24) | (g << 16) | (r << 8) | a;
 166       }
 167    }
 168 }
 169
 170
 171 /*** PIPE_FORMAT_A1R5G5B5_UNORM ***/
 172
 173 static void
 174 a1r5g5b5_get_tile_rgba(const ushort *src,
 175                        unsigned w, unsigned h,
 176                        uint8_t *p)
 177 {
 178    unsigned i, j;
 179
 180    for (i = 0; i < h; i++) {
 181       for (j = 0; j < w; j++) {
 182          const ushort pixel = *src++;
 183          PIXEL(p, j, i, 0) = ((pixel >> 10) & 0x1f) * 255 / 31;
 184          PIXEL(p, j, i, 1) = ((pixel >>  5) & 0x1f) * 255 / 31;
 185          PIXEL(p, j, i, 2) = ((pixel      ) & 0x1f) * 255 / 31;
 186          PIXEL(p, j, i, 3) = ((pixel >> 15)       ) * 255;
 187       }
 188    }
 189 }
 190
 191
 192 static void
 193 a1r5g5b5_put_tile_rgba(ushort *dst,
 194                        unsigned w, unsigned h,
 195                        const uint8_t *p)
 196 {
 197    unsigned i, j;
 198
 199    for (i = 0; i < h; i++) {
 200       for (j = 0; j < w; j++) {
 201          unsigned r, g, b, a;
 202          r = PIXEL(p, j, i, 0);
 203          g = PIXEL(p, j, i, 1);
 204          b = PIXEL(p, j, i, 2);
 205          a = PIXEL(p, j, i, 3);
 206          r = r >> 3;  /* 5 bits */
 207          g = g >> 3;  /* 5 bits */
 208          b = b >> 3;  /* 5 bits */
 209          a = a >> 7;  /* 1 bit */
 210          *dst++ = (a << 15) | (r << 10) | (g << 5) | b;
 211       }
 212    }
 213 }
 214
 215
 216 /*** PIPE_FORMAT_A4R4G4B4_UNORM ***/
 217
 218 static void
 219 a4r4g4b4_get_tile_rgba(const ushort *src,
 220                        unsigned w, unsigned h,
 221                        uint8_t *p)
 222 {
 223    unsigned i, j;
 224
 225    for (i = 0; i < h; i++) {
 226       for (j = 0; j < w; j++) {
 227          const ushort pixel = *src++;
 228          PIXEL(p, j, i, 0) = ((pixel >>  8) & 0xf) * 255 / 15;
 229          PIXEL(p, j, i, 1) = ((pixel >>  4) & 0xf) * 255 / 15;
 230          PIXEL(p, j, i, 2) = ((pixel      ) & 0xf) * 255 / 15;
 231          PIXEL(p, j, i, 3) = ((pixel >> 12)      ) * 255 / 15;
 232       }
 233    }
 234 }
 235
 236
 237 static void
 238 a4r4g4b4_put_tile_rgba(ushort *dst,
 239                        unsigned w, unsigned h,
 240                        const uint8_t *p)
 241 {
 242    unsigned i, j;
 243
 244    for (i = 0; i < h; i++) {
 245       for (j = 0; j < w; j++) {
 246          unsigned r, g, b, a;
 247          r = PIXEL(p, j, i, 0);
 248          g = PIXEL(p, j, i, 1);
 249          b = PIXEL(p, j, i, 2);
 250          a = PIXEL(p, j, i, 3);
 251          r >>= 4;
 252          g >>= 4;
 253          b >>= 4;
 254          a >>= 4;
 255          *dst++ = (a << 12) | (r << 16) | (g << 4) | b;
 256       }
 257    }
 258 }
 259
 260
 261 /*** PIPE_FORMAT_R5G6B5_UNORM ***/
 262
 263 static void
 264 r5g6b5_get_tile_rgba(const ushort *src,
 265                      unsigned w, unsigned h,
 266                      uint8_t *p)
 267 {
 268    unsigned i, j;
 269
 270    for (i = 0; i < h; i++) {
 271       for (j = 0; j < w; j++) {
 272          const ushort pixel = *src++;
 273          PIXEL(p, j, i, 0) = ((pixel >> 11) & 0x1f) * 255 / 31;
 274          PIXEL(p, j, i, 1) = ((pixel >>  5) & 0x3f) * 255 / 63;
 275          PIXEL(p, j, i, 2) = ((pixel      ) & 0x1f) * 255 / 31;
 276          PIXEL(p, j, i, 3) = 255;
 277       }
 278    }
 279 }
 280
 281
 282 static void
 283 r5g6b5_put_tile_rgba(ushort *dst,
 284                      unsigned w, unsigned h,
 285                      const uint8_t *p)
 286 {
 287    unsigned i, j;
 288
 289    for (i = 0; i < h; i++) {
 290       for (j = 0; j < w; j++) {
 291          uint r = (uint) PIXEL(p, j, i, 0) * 31 / 255;
 292          uint g = (uint) PIXEL(p, j, i, 1) * 63 / 255;
 293          uint b = (uint) PIXEL(p, j, i, 2) * 31 / 255;
 294          *dst++ = (r << 11) | (g << 5) | (b);
 295       }
 296    }
 297 }
 298
 299
 300
 301 /*** PIPE_FORMAT_Z16_UNORM ***/
 302
 303 /**
 304  * Return each Z value as four floats in [0,1].
 305  */
 306 static void
 307 z16_get_tile_rgba(const ushort *src,
 308                   unsigned w, unsigned h,
 309                   uint8_t *p)
 310 {
 311    const float scale = 1.0f / 65535.0f;
 312    unsigned i, j;
 313
 314    for (i = 0; i < h; i++) {
 315       for (j = 0; j < w; j++) {
 316          PIXEL(p, j, i, 0) =
 317          PIXEL(p, j, i, 1) =
 318          PIXEL(p, j, i, 2) =
 319          PIXEL(p, j, i, 3) = *src++ * scale;
 320       }
 321    }
 322 }
 323
 324
 325
 326
 327 /*** PIPE_FORMAT_L8_UNORM ***/
 328
 329 static void
 330 l8_get_tile_rgba(const ubyte *src,
 331                  unsigned w, unsigned h,
 332                  uint8_t *p)
 333 {
 334    unsigned i, j;
 335
 336    for (i = 0; i < h; i++) {
 337       for (j = 0; j < w; j++, src++) {
 338          PIXEL(p, j, i, 0) =
 339          PIXEL(p, j, i, 1) =
 340          PIXEL(p, j, i, 2) = *src;
 341          PIXEL(p, j, i, 3) = 255;
 342       }
 343    }
 344 }
 345
 346
 347 static void
 348 l8_put_tile_rgba(ubyte *dst,
 349                  unsigned w, unsigned h,
 350                  const uint8_t *p)
 351 {
 352    unsigned i, j;
 353
 354    for (i = 0; i < h; i++) {
 355       for (j = 0; j < w; j++) {
 356          unsigned r;
 357          r = PIXEL(p, j, i, 0);
 358          *dst++ = (ubyte) r;
 359       }
 360    }
 361 }
 362
 363
 364
 365 /*** PIPE_FORMAT_A8_UNORM ***/
 366
 367 static void
 368 a8_get_tile_rgba(const ubyte *src,
 369                  unsigned w, unsigned h,
 370                  uint8_t *p)
 371 {
 372    unsigned i, j;
 373
 374    for (i = 0; i < h; i++) {
 375       for (j = 0; j < w; j++, src++) {
 376          PIXEL(p, j, i, 0) =
 377          PIXEL(p, j, i, 1) =
 378          PIXEL(p, j, i, 2) = 0;
 379          PIXEL(p, j, i, 3) = *src;
 380       }
 381    }
 382 }
 383
 384
 385 static void
 386 a8_put_tile_rgba(ubyte *dst,
 387                  unsigned w, unsigned h,
 388                  const uint8_t *p)
 389 {
 390    unsigned i, j;
 391
 392    for (i = 0; i < h; i++) {
 393       for (j = 0; j < w; j++) {
 394          unsigned a;
 395          a = PIXEL(p, j, i, 3);
 396          *dst++ = (ubyte) a;
 397       }
 398    }
 399 }
 400
 401
 402
 403 /*** PIPE_FORMAT_R16_SNORM ***/
 404
 405 static void
 406 r16_get_tile_rgba(const short *src,
 407                   unsigned w, unsigned h,
 408                   uint8_t *p)
 409 {
 410    unsigned i, j;
 411
 412    for (i = 0; i < h; i++) {
 413       for (j = 0; j < w; j++, src++) {
 414          PIXEL(p, j, i, 0) = MAX2(src[0] >> 7, 0);
 415          PIXEL(p, j, i, 1) =
 416          PIXEL(p, j, i, 2) = 0;
 417          PIXEL(p, j, i, 3) = 255;
 418       }
 419    }
 420 }
 421
 422
 423 static void
 424 r16_put_tile_rgba(short *dst,
 425                   unsigned w, unsigned h,
 426                   const uint8_t *p)
 427 {
 428    unsigned i, j;
 429
 430    for (i = 0; i < h; i++) {
 431       for (j = 0; j < w; j++, dst++) {
 432          dst[0] = PIXEL(p, j, i, 0) << 7;
 433       }
 434    }
 435 }
 436
 437
 438 /*** PIPE_FORMAT_R16G16B16A16_SNORM ***/
 439
 440 static void
 441 r16g16b16a16_get_tile_rgba(const short *src,
 442                            unsigned w, unsigned h,
 443                            uint8_t *p)
 444 {
 445    unsigned i, j;
 446
 447    for (i = 0; i < h; i++) {
 448       for (j = 0; j < w; j++, src += 4) {
 449          PIXEL(p, j, i, 0) = src[0] >> 8;
 450          PIXEL(p, j, i, 1) = src[1] >> 8;
 451          PIXEL(p, j, i, 2) = src[2] >> 8;
 452          PIXEL(p, j, i, 3) = src[3] >> 8;
 453       }
 454    }
 455 }
 456
 457
 458 static void
 459 r16g16b16a16_put_tile_rgba(short *dst,
 460                            unsigned w, unsigned h,
 461                            const uint8_t *p)
 462 {
 463    unsigned i, j;
 464
 465    for (i = 0; i < h; i++) {
 466       for (j = 0; j < w; j++, dst += 4) {
 467          dst[0] = PIXEL(p, j, i, 0) << 8;
 468          dst[1] = PIXEL(p, j, i, 1) << 8;
 469          dst[2] = PIXEL(p, j, i, 2) << 8;
 470          dst[3] = PIXEL(p, j, i, 3) << 8;
 471       }
 472    }
 473 }
 474
 475
 476
 477 /*** PIPE_FORMAT_I8_UNORM ***/
 478
 479 static void
 480 i8_get_tile_rgba(const ubyte *src,
 481                  unsigned w, unsigned h,
 482                  uint8_t *p)
 483 {
 484    unsigned i, j;
 485
 486    for (i = 0; i < h; i++) {
 487       for (j = 0; j < w; j++, src++) {
 488          PIXEL(p, j, i, 0) =
 489          PIXEL(p, j, i, 1) =
 490          PIXEL(p, j, i, 2) =
 491          PIXEL(p, j, i, 3) = *src;
 492       }
 493    }
 494 }
 495
 496
 497 static void
 498 i8_put_tile_rgba(ubyte *dst,
 499                  unsigned w, unsigned h,
 500                  const uint8_t *p)
 501 {
 502    unsigned i, j;
 503
 504    for (i = 0; i < h; i++) {
 505       for (j = 0; j < w; j++) {
 506          unsigned r;
 507          r = PIXEL(p, j, i, 0);
 508          *dst++ = (ubyte) r;
 509       }
 510    }
 511 }
 512
 513
 514 /*** PIPE_FORMAT_A8L8_UNORM ***/
 515
 516 static void
 517 a8l8_get_tile_rgba(const ushort *src,
 518                    unsigned w, unsigned h,
 519                    uint8_t *p)
 520 {
 521    unsigned i, j;
 522
 523    for (i = 0; i < h; i++) {
 524       for (j = 0; j < w; j++) {
 525          ushort ra = *src++;
 526          PIXEL(p, j, i, 0) =
 527          PIXEL(p, j, i, 1) =
 528          PIXEL(p, j, i, 2) = ra & 0xff;
 529          PIXEL(p, j, i, 3) = ra >> 8;
 530       }
 531    }
 532 }
 533
 534
 535 static void
 536 a8l8_put_tile_rgba(ushort *dst,
 537                    unsigned w, unsigned h,
 538                    const uint8_t *p)
 539 {
 540    unsigned i, j;
 541
 542    for (i = 0; i < h; i++) {
 543       for (j = 0; j < w; j++) {
 544          unsigned r, a;
 545          r = PIXEL(p, j, i, 0);
 546          a = PIXEL(p, j, i, 3);
 547          *dst++ = (a << 8) | r;
 548       }
 549    }
 550 }
 551
 552
 553
 554
 555 /*** PIPE_FORMAT_Z32_UNORM ***/
 556
 557 /**
 558  * Return each Z value as four floats in [0,1].
 559  */
 560 static void
 561 z32_get_tile_rgba(const unsigned *src,
 562                   unsigned w, unsigned h,
 563                   uint8_t *p)
 564 {
 565    const double scale = 1.0 / (double) 0xffffffff;
 566    unsigned i, j;
 567
 568    for (i = 0; i < h; i++) {
 569       for (j = 0; j < w; j++) {
 570          PIXEL(p, j, i, 0) =
 571          PIXEL(p, j, i, 1) =
 572          PIXEL(p, j, i, 2) =
 573          PIXEL(p, j, i, 3) = (float) (*src++ * scale);
 574       }
 575    }
 576 }
 577
 578
 579 /*** PIPE_FORMAT_S8Z24_UNORM ***/
 580
 581 /**
 582  * Return Z component as four float in [0,1].  Stencil part ignored.
 583  */
 584 static void
 585 s8z24_get_tile_rgba(const unsigned *src,
 586                     unsigned w, unsigned h,
 587                     uint8_t *p)
 588 {
 589    const double scale = 1.0 / ((1 << 24) - 1);
 590    unsigned i, j;
 591
 592    for (i = 0; i < h; i++) {
 593       for (j = 0; j < w; j++) {
 594          PIXEL(p, j, i, 0) =
 595          PIXEL(p, j, i, 1) =
 596          PIXEL(p, j, i, 2) =
 597          PIXEL(p, j, i, 3) = (float) (scale * (*src++ & 0xffffff));
 598       }
 599    }
 600 }
 601
 602
 603 /*** PIPE_FORMAT_Z24S8_UNORM ***/
 604
 605 /**
 606  * Return Z component as four float in [0,1].  Stencil part ignored.
 607  */
 608 static void
 609 z24s8_get_tile_rgba(const unsigned *src,
 610                     unsigned w, unsigned h,
 611                     uint8_t *p)
 612 {
 613    const double scale = 1.0 / ((1 << 24) - 1);
 614    unsigned i, j;
 615
 616    for (i = 0; i < h; i++) {
 617       for (j = 0; j < w; j++) {
 618          PIXEL(p, j, i, 0) =
 619          PIXEL(p, j, i, 1) =
 620          PIXEL(p, j, i, 2) =
 621          PIXEL(p, j, i, 3) = (float) (scale * (*src++ >> 8));
 622       }
 623    }
 624 }
 625
 626
 627 /*** PIPE_FORMAT_Z32_FLOAT ***/
 628
 629 /**
 630  * Return each Z value as four floats in [0,1].
 631  */
 632 static void
 633 z32f_get_tile_rgba(const float *src,
 634                    unsigned w, unsigned h,
 635                    uint8_t *p)
 636 {
 637    unsigned i, j;
 638
 639    for (i = 0; i < h; i++) {
 640       for (j = 0; j < w; j++) {
 641          PIXEL(p, j, i, 0) =
 642          PIXEL(p, j, i, 1) =
 643          PIXEL(p, j, i, 2) =
 644          PIXEL(p, j, i, 3) = *src++;
 645       }
 646    }
 647 }
 648
 649
 650 /*** PIPE_FORMAT_YCBCR / PIPE_FORMAT_YCBCR_REV ***/
 651
 652 /**
 653  * Convert YCbCr (or YCrCb) to RGBA.
 654  */
 655 static void
 656 ycbcr_get_tile_rgba(const ushort *src,
 657                     unsigned w, unsigned h,
 658                     uint8_t *p,
 659                     boolean rev)
 660 {
 661    unsigned i, j;
 662
 663    for (i = 0; i < h; i++) {
 664       /* do two texels at a time */
 665       for (j = 0; j < (w & ~1); j += 2, src += 2) {
 666          const ushort t0 = src[0];
 667          const ushort t1 = src[1];
 668          const ubyte y0 = (t0 >> 8) & 0xff;  /* luminance */
 669          const ubyte y1 = (t1 >> 8) & 0xff;  /* luminance */
 670          ubyte cb, cr;
 671          float r, g, b;
 672
 673          if (rev) {
 674             cb = t1 & 0xff;         /* chroma U */
 675             cr = t0 & 0xff;         /* chroma V */
 676          }
 677          else {
 678             cb = t0 & 0xff;         /* chroma U */
 679             cr = t1 & 0xff;         /* chroma V */
 680          }
 681
 682          /* even pixel: y0,cr,cb */
 683          r = 1.164f * (y0-16) + 1.596f * (cr-128);
 684          g = 1.164f * (y0-16) - 0.813f * (cr-128) - 0.391f * (cb-128);
 685          b = 1.164f * (y0-16) + 2.018f * (cb-128);
 686          PIXEL(p, j, i, 0) = r;
 687          PIXEL(p, j, i, 1) = g;
 688          PIXEL(p, j, i, 2) = b;
 689          PIXEL(p, j, i, 3) = 255;
 690
 691          /* odd pixel: use y1,cr,cb */
 692          r = 1.164f * (y1-16) + 1.596f * (cr-128);
 693          g = 1.164f * (y1-16) - 0.813f * (cr-128) - 0.391f * (cb-128);
 694          b = 1.164f * (y1-16) + 2.018f * (cb-128);
 695          PIXEL(p, j + 1, i, 0) = r;
 696          PIXEL(p, j + 1, i, 1) = g;
 697          PIXEL(p, j + 1, i, 2) = b;
 698          PIXEL(p, j + 1, i, 3) = 255;
 699       }
 700       /* do the last texel */
 701       if (w & 1) {
 702          const ushort t0 = src[0];
 703          const ushort t1 = src[1];
 704          const ubyte y0 = (t0 >> 8) & 0xff;  /* luminance */
 705          ubyte cb, cr;
 706          float r, g, b;
 707
 708          if (rev) {
 709             cb = t1 & 0xff;         /* chroma U */
 710             cr = t0 & 0xff;         /* chroma V */
 711          }
 712          else {
 713             cb = t0 & 0xff;         /* chroma U */
 714             cr = t1 & 0xff;         /* chroma V */
 715          }
 716
 717          /* even pixel: y0,cr,cb */
 718          r = 1.164f * (y0-16) + 1.596f * (cr-128);
 719          g = 1.164f * (y0-16) - 0.813f * (cr-128) - 0.391f * (cb-128);
 720          b = 1.164f * (y0-16) + 2.018f * (cb-128);
 721          PIXEL(p, j, i, 0) = r;
 722          PIXEL(p, j, i, 1) = g;
 723          PIXEL(p, j, i, 2) = b;
 724          PIXEL(p, j, i, 3) = 255;
 725       }
 726    }
 727 }
 728
 729
 730 static void
 731 fake_get_tile_rgba(const ushort *src,
 732                    unsigned w, unsigned h,
 733                    uint8_t *p)
 734 {
 735    unsigned i, j;
 736
 737    for (i = 0; i < h; i++) {
 738       for (j = 0; j < w; j++) {
 739          PIXEL(p, j, i, 0) =
 740          PIXEL(p, j, i, 1) =
 741          PIXEL(p, j, i, 2) =
 742          PIXEL(p, j, i, 3) = (i ^ j) & 1 ? 255 : 0;
 743       }
 744    }
 745 }
 746
 747
 748 static void
 749 lp_tile_raw_to_rgba_soa(enum pipe_format format,
 750                         void *src,
 751                         uint w, uint h,
 752                         uint8_t *p)
 753 {
 754    switch (format) {
 755    case PIPE_FORMAT_A8R8G8B8_UNORM:
 756       a8r8g8b8_get_tile_rgba((unsigned *) src, w, h, p);
 757       break;
 758    case PIPE_FORMAT_X8R8G8B8_UNORM:
 759       x8r8g8b8_get_tile_rgba((unsigned *) src, w, h, p);
 760       break;
 761    case PIPE_FORMAT_B8G8R8A8_UNORM:
 762       b8g8r8a8_get_tile_rgba((unsigned *) src, w, h, p);
 763       break;
 764    case PIPE_FORMAT_A1R5G5B5_UNORM:
 765       a1r5g5b5_get_tile_rgba((ushort *) src, w, h, p);
 766       break;
 767    case PIPE_FORMAT_A4R4G4B4_UNORM:
 768       a4r4g4b4_get_tile_rgba((ushort *) src, w, h, p);
 769       break;
 770    case PIPE_FORMAT_R5G6B5_UNORM:
 771       r5g6b5_get_tile_rgba((ushort *) src, w, h, p);
 772       break;
 773    case PIPE_FORMAT_L8_UNORM:
 774       l8_get_tile_rgba((ubyte *) src, w, h, p);
 775       break;
 776    case PIPE_FORMAT_A8_UNORM:
 777       a8_get_tile_rgba((ubyte *) src, w, h, p);
 778       break;
 779    case PIPE_FORMAT_I8_UNORM:
 780       i8_get_tile_rgba((ubyte *) src, w, h, p);
 781       break;
 782    case PIPE_FORMAT_A8L8_UNORM:
 783       a8l8_get_tile_rgba((ushort *) src, w, h, p);
 784       break;
 785    case PIPE_FORMAT_R16_SNORM:
 786       r16_get_tile_rgba((short *) src, w, h, p);
 787       break;
 788    case PIPE_FORMAT_R16G16B16A16_SNORM:
 789       r16g16b16a16_get_tile_rgba((short *) src, w, h, p);
 790       break;
 791    case PIPE_FORMAT_Z16_UNORM:
 792       z16_get_tile_rgba((ushort *) src, w, h, p);
 793       break;
 794    case PIPE_FORMAT_Z32_UNORM:
 795       z32_get_tile_rgba((unsigned *) src, w, h, p);
 796       break;
 797    case PIPE_FORMAT_S8Z24_UNORM:
 798    case PIPE_FORMAT_X8Z24_UNORM:
 799       s8z24_get_tile_rgba((unsigned *) src, w, h, p);
 800       break;
 801    case PIPE_FORMAT_Z24S8_UNORM:
 802    case PIPE_FORMAT_Z24X8_UNORM:
 803       z24s8_get_tile_rgba((unsigned *) src, w, h, p);
 804       break;
 805    case PIPE_FORMAT_Z32_FLOAT:
 806       z32f_get_tile_rgba((float *) src, w, h, p);
 807       break;
 808    case PIPE_FORMAT_YCBCR:
 809       ycbcr_get_tile_rgba((ushort *) src, w, h, p, FALSE);
 810       break;
 811    case PIPE_FORMAT_YCBCR_REV:
 812       ycbcr_get_tile_rgba((ushort *) src, w, h, p, TRUE);
 813       break;
 814    default:
 815       debug_printf("%s: unsupported format %s\n", __FUNCTION__, pf_name(format));
 816       fake_get_tile_rgba(src, w, h, p);
 817    }
 818 }
 819
 820
 821 void
 822 lp_get_tile_rgba_soa(struct pipe_transfer *pt,
 823                      uint x, uint y,
 824                      uint8_t *p)
 825 {
 826    uint w = TILE_SIZE, h = TILE_SIZE;
 827    void *packed;
 828
 829    if (pipe_clip_tile(x, y, &w, &h, pt))
 830       return;
 831
 832    packed = MALLOC(pf_get_nblocks(&pt->block, w, h) * pt->block.size);
 833
 834    if (!packed)
 835       return;
 836
 837    if(pt->format == PIPE_FORMAT_YCBCR || pt->format == PIPE_FORMAT_YCBCR_REV)
 838       assert((x & 1) == 0);
 839
 840    pipe_get_tile_raw(pt, x, y, w, h, packed, 0);
 841
 842    lp_tile_raw_to_rgba_soa(pt->format, packed, w, h, p);
 843
 844    FREE(packed);
 845 }
 846
 847
 848 void
 849 lp_put_tile_rgba_soa(struct pipe_transfer *pt,
 850                      uint x, uint y,
 851                      const uint8_t *p)
 852 {
 853    uint w = TILE_SIZE, h = TILE_SIZE;
 854    void *packed;
 855
 856    if (pipe_clip_tile(x, y, &w, &h, pt))
 857       return;
 858
 859    packed = MALLOC(pf_get_nblocks(&pt->block, w, h) * pt->block.size);
 860
 861    if (!packed)
 862       return;
 863
 864    switch (pt->format) {
 865    case PIPE_FORMAT_A8R8G8B8_UNORM:
 866       a8r8g8b8_put_tile_rgba((unsigned *) packed, w, h, p);
 867       break;
 868    case PIPE_FORMAT_X8R8G8B8_UNORM:
 869       x8r8g8b8_put_tile_rgba((unsigned *) packed, w, h, p);
 870       break;
 871    case PIPE_FORMAT_B8G8R8A8_UNORM:
 872       b8g8r8a8_put_tile_rgba((unsigned *) packed, w, h, p);
 873       break;
 874    case PIPE_FORMAT_A1R5G5B5_UNORM:
 875       a1r5g5b5_put_tile_rgba((ushort *) packed, w, h, p);
 876       break;
 877    case PIPE_FORMAT_R5G6B5_UNORM:
 878       r5g6b5_put_tile_rgba((ushort *) packed, w, h, p);
 879       break;
 880    case PIPE_FORMAT_R8G8B8A8_UNORM:
 881       assert(0);
 882       break;
 883    case PIPE_FORMAT_A4R4G4B4_UNORM:
 884       a4r4g4b4_put_tile_rgba((ushort *) packed, w, h, p);
 885       break;
 886    case PIPE_FORMAT_L8_UNORM:
 887       l8_put_tile_rgba((ubyte *) packed, w, h, p);
 888       break;
 889    case PIPE_FORMAT_A8_UNORM:
 890       a8_put_tile_rgba((ubyte *) packed, w, h, p);
 891       break;
 892    case PIPE_FORMAT_I8_UNORM:
 893       i8_put_tile_rgba((ubyte *) packed, w, h, p);
 894       break;
 895    case PIPE_FORMAT_A8L8_UNORM:
 896       a8l8_put_tile_rgba((ushort *) packed, w, h, p);
 897       break;
 898    case PIPE_FORMAT_R16_SNORM:
 899       r16_put_tile_rgba((short *) packed, w, h, p);
 900       break;
 901    case PIPE_FORMAT_R16G16B16A16_SNORM:
 902       r16g16b16a16_put_tile_rgba((short *) packed, w, h, p);
 903       break;
 904    case PIPE_FORMAT_Z16_UNORM:
 905       /*z16_put_tile_rgba((ushort *) packed, w, h, p);*/
 906       break;
 907    case PIPE_FORMAT_Z32_UNORM:
 908       /*z32_put_tile_rgba((unsigned *) packed, w, h, p);*/
 909       break;
 910    case PIPE_FORMAT_S8Z24_UNORM:
 911    case PIPE_FORMAT_X8Z24_UNORM:
 912       /*s8z24_put_tile_rgba((unsigned *) packed, w, h, p);*/
 913       break;
 914    case PIPE_FORMAT_Z24S8_UNORM:
 915    case PIPE_FORMAT_Z24X8_UNORM:
 916       /*z24s8_put_tile_rgba((unsigned *) packed, w, h, p);*/
 917       break;
 918    default:
 919       debug_printf("%s: unsupported format %s\n", __FUNCTION__, pf_name(pt->format));
 920    }
 921
 922    pipe_put_tile_raw(pt, x, y, w, h, packed, 0);
 923
 924    FREE(packed);
 925 }
 926
 927