src/gallium/state_trackers/xorg/xorg_exa_tgsi.c

   1 #include "xorg_exa_tgsi.h"
   2
   3 /*### stupidity defined in X11/extensions/XI.h */
   4 #undef Absolute
   5
   6 #include "pipe/p_format.h"
   7 #include "pipe/p_context.h"
   8 #include "pipe/p_state.h"
   9 #include "pipe/p_inlines.h"
  10 #include "pipe/p_shader_tokens.h"
  11
  12 #include "util/u_memory.h"
  13 #include "util/u_simple_shaders.h"
  14
  15 #include "tgsi/tgsi_ureg.h"
  16
  17 #include "cso_cache/cso_context.h"
  18 #include "cso_cache/cso_hash.h"
  19
  20 /* Vertex shader:
  21  * IN[0]    = vertex pos
  22  * IN[1]    = src tex coord | solid fill color
  23  * IN[2]    = mask tex coord
  24  * IN[3]    = dst tex coord
  25  * CONST[0] = (2/dst_width, 2/dst_height, 1, 1)
  26  * CONST[1] = (-1, -1, 0, 0)
  27  *
  28  * OUT[0]   = vertex pos
  29  * OUT[1]   = src tex coord | solid fill color
  30  * OUT[2]   = mask tex coord
  31  * OUT[3]   = dst tex coord
  32  */
  33
  34 /* Fragment shader:
  35  * SAMP[0]  = src
  36  * SAMP[1]  = mask
  37  * SAMP[2]  = dst
  38  * IN[0]    = pos src | solid fill color
  39  * IN[1]    = pos mask
  40  * IN[2]    = pos dst
  41  * CONST[0] = (0, 0, 0, 1)
  42  *
  43  * OUT[0] = color
  44  */
  45
  46 struct xorg_shaders {
  47    struct xorg_renderer *r;
  48
  49    struct cso_hash *vs_hash;
  50    struct cso_hash *fs_hash;
  51 };
  52
  53 static INLINE void
  54 src_in_mask(struct ureg_program *ureg,
  55             struct ureg_dst dst,
  56             struct ureg_src src,
  57             struct ureg_src mask,
  58             int component_alpha)
  59 {
  60    if (component_alpha == FS_CA_FULL) {
  61       ureg_MUL(ureg, dst, src, mask);
  62    } else if (component_alpha == FS_CA_SRCALPHA) {
  63       ureg_MUL(ureg, dst,
  64                ureg_scalar(src, TGSI_SWIZZLE_W), mask);
  65    }
  66    else {
  67       ureg_MUL(ureg, dst, src,
  68                ureg_scalar(mask, TGSI_SWIZZLE_X));
  69    }
  70 }
  71
  72 static struct ureg_src
  73 vs_normalize_coords(struct ureg_program *ureg, struct ureg_src coords,
  74                     struct ureg_src const0, struct ureg_src const1)
  75 {
  76    struct ureg_dst tmp = ureg_DECL_temporary(ureg);
  77    struct ureg_src ret;
  78    ureg_MAD(ureg, tmp, coords, const0, const1);
  79    ret = ureg_src(tmp);
  80    ureg_release_temporary(ureg, tmp);
  81    return ret;
  82 }
  83
  84 static void
  85 linear_gradient(struct ureg_program *ureg,
  86                 struct ureg_dst out,
  87                 struct ureg_src pos,
  88                 struct ureg_src sampler,
  89                 struct ureg_src coords,
  90                 struct ureg_src const0124,
  91                 struct ureg_src matrow0,
  92                 struct ureg_src matrow1,
  93                 struct ureg_src matrow2)
  94 {
  95    struct ureg_dst temp0 = ureg_DECL_temporary(ureg);
  96    struct ureg_dst temp1 = ureg_DECL_temporary(ureg);
  97    struct ureg_dst temp2 = ureg_DECL_temporary(ureg);
  98    struct ureg_dst temp3 = ureg_DECL_temporary(ureg);
  99    struct ureg_dst temp4 = ureg_DECL_temporary(ureg);
 100    struct ureg_dst temp5 = ureg_DECL_temporary(ureg);
 101
 102    ureg_MOV(ureg,
 103             ureg_writemask(temp0, TGSI_WRITEMASK_XY), pos);
 104    ureg_MOV(ureg,
 105             ureg_writemask(temp0, TGSI_WRITEMASK_Z),
 106             ureg_scalar(const0124, TGSI_SWIZZLE_Y));
 107
 108    ureg_DP3(ureg, temp1, matrow0, ureg_src(temp0));
 109    ureg_DP3(ureg, temp2, matrow1, ureg_src(temp0));
 110    ureg_DP3(ureg, temp3, matrow2, ureg_src(temp0));
 111    ureg_RCP(ureg, temp3, ureg_src(temp3));
 112    ureg_MUL(ureg, temp1, ureg_src(temp1), ureg_src(temp3));
 113    ureg_MUL(ureg, temp2, ureg_src(temp2), ureg_src(temp3));
 114
 115    ureg_MOV(ureg, ureg_writemask(temp4, TGSI_WRITEMASK_X),
 116             ureg_src(temp1));
 117    ureg_MOV(ureg, ureg_writemask(temp4, TGSI_WRITEMASK_Y),
 118             ureg_src(temp2));
 119
 120    ureg_MUL(ureg, temp0,
 121             ureg_scalar(coords, TGSI_SWIZZLE_Y),
 122             ureg_scalar(ureg_src(temp4), TGSI_SWIZZLE_Y));
 123    ureg_MAD(ureg, temp1,
 124             ureg_scalar(coords, TGSI_SWIZZLE_X),
 125             ureg_scalar(ureg_src(temp4), TGSI_SWIZZLE_X),
 126             ureg_src(temp0));
 127
 128    ureg_MUL(ureg, temp2,
 129             ureg_src(temp1),
 130             ureg_scalar(coords, TGSI_SWIZZLE_Z));
 131
 132    ureg_TEX(ureg, out,
 133             TGSI_TEXTURE_1D, ureg_src(temp2), sampler);
 134
 135    ureg_release_temporary(ureg, temp0);
 136    ureg_release_temporary(ureg, temp1);
 137    ureg_release_temporary(ureg, temp2);
 138    ureg_release_temporary(ureg, temp3);
 139    ureg_release_temporary(ureg, temp4);
 140    ureg_release_temporary(ureg, temp5);
 141 }
 142
 143
 144 static void
 145 radial_gradient(struct ureg_program *ureg,
 146                 struct ureg_dst out,
 147                 struct ureg_src pos,
 148                 struct ureg_src sampler,
 149                 struct ureg_src coords,
 150                 struct ureg_src const0124,
 151                 struct ureg_src matrow0,
 152                 struct ureg_src matrow1,
 153                 struct ureg_src matrow2)
 154 {
 155    struct ureg_dst temp0 = ureg_DECL_temporary(ureg);
 156    struct ureg_dst temp1 = ureg_DECL_temporary(ureg);
 157    struct ureg_dst temp2 = ureg_DECL_temporary(ureg);
 158    struct ureg_dst temp3 = ureg_DECL_temporary(ureg);
 159    struct ureg_dst temp4 = ureg_DECL_temporary(ureg);
 160    struct ureg_dst temp5 = ureg_DECL_temporary(ureg);
 161
 162    ureg_MOV(ureg,
 163             ureg_writemask(temp0, TGSI_WRITEMASK_XY),
 164             pos);
 165    ureg_MOV(ureg,
 166             ureg_writemask(temp0, TGSI_WRITEMASK_Z),
 167             ureg_scalar(const0124, TGSI_SWIZZLE_Y));
 168
 169    ureg_DP3(ureg, temp1, matrow0, ureg_src(temp0));
 170    ureg_DP3(ureg, temp2, matrow1, ureg_src(temp0));
 171    ureg_DP3(ureg, temp3, matrow2, ureg_src(temp0));
 172    ureg_RCP(ureg, temp3, ureg_src(temp3));
 173    ureg_MUL(ureg, temp1, ureg_src(temp1), ureg_src(temp3));
 174    ureg_MUL(ureg, temp2, ureg_src(temp2), ureg_src(temp3));
 175
 176    ureg_MOV(ureg, ureg_writemask(temp5, TGSI_WRITEMASK_X),
 177             ureg_src(temp1));
 178    ureg_MOV(ureg, ureg_writemask(temp5, TGSI_WRITEMASK_Y),
 179             ureg_src(temp2));
 180
 181    ureg_MUL(ureg, temp0, ureg_scalar(coords, TGSI_SWIZZLE_Y),
 182             ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_Y));
 183    ureg_MAD(ureg, temp1,
 184             ureg_scalar(coords, TGSI_SWIZZLE_X),
 185             ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_X),
 186             ureg_src(temp0));
 187    ureg_ADD(ureg, temp1,
 188             ureg_src(temp1), ureg_src(temp1));
 189    ureg_MUL(ureg, temp3,
 190             ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_Y),
 191             ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_Y));
 192    ureg_MAD(ureg, temp4,
 193             ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_X),
 194             ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_X),
 195             ureg_src(temp3));
 196    ureg_MOV(ureg, temp4, ureg_negate(ureg_src(temp4)));
 197    ureg_MUL(ureg, temp2,
 198             ureg_scalar(coords, TGSI_SWIZZLE_Z),
 199             ureg_src(temp4));
 200    ureg_MUL(ureg, temp0,
 201             ureg_scalar(const0124, TGSI_SWIZZLE_W),
 202             ureg_src(temp2));
 203    ureg_MUL(ureg, temp3,
 204             ureg_src(temp1), ureg_src(temp1));
 205    ureg_SUB(ureg, temp2,
 206             ureg_src(temp3), ureg_src(temp0));
 207    ureg_RSQ(ureg, temp2, ureg_abs(ureg_src(temp2)));
 208    ureg_RCP(ureg, temp2, ureg_src(temp2));
 209    ureg_SUB(ureg, temp1,
 210             ureg_src(temp2), ureg_src(temp1));
 211    ureg_ADD(ureg, temp0,
 212             ureg_scalar(coords, TGSI_SWIZZLE_Z),
 213             ureg_scalar(coords, TGSI_SWIZZLE_Z));
 214    ureg_RCP(ureg, temp0, ureg_src(temp0));
 215    ureg_MUL(ureg, temp2,
 216             ureg_src(temp1), ureg_src(temp0));
 217    ureg_TEX(ureg, out, TGSI_TEXTURE_1D,
 218             ureg_src(temp2), sampler);
 219
 220    ureg_release_temporary(ureg, temp0);
 221    ureg_release_temporary(ureg, temp1);
 222    ureg_release_temporary(ureg, temp2);
 223    ureg_release_temporary(ureg, temp3);
 224    ureg_release_temporary(ureg, temp4);
 225    ureg_release_temporary(ureg, temp5);
 226 }
 227
 228 static void *
 229 create_vs(struct pipe_context *pipe,
 230           unsigned vs_traits)
 231 {
 232    struct ureg_program *ureg;
 233    struct ureg_src src;
 234    struct ureg_dst dst;
 235    struct ureg_src const0, const1;
 236    boolean is_fill = vs_traits & VS_FILL;
 237    boolean is_composite = vs_traits & VS_COMPOSITE;
 238    boolean has_mask = vs_traits & VS_MASK;
 239    unsigned input_slot = 0;
 240
 241    ureg = ureg_create(TGSI_PROCESSOR_VERTEX);
 242    if (ureg == NULL)
 243       return 0;
 244
 245    const0 = ureg_DECL_constant(ureg, 0);
 246    const1 = ureg_DECL_constant(ureg, 1);
 247
 248    /* it has to be either a fill or a composite op */
 249    debug_assert(is_fill ^ is_composite);
 250
 251    src = ureg_DECL_vs_input(ureg, input_slot++);
 252    dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0);
 253    src = vs_normalize_coords(ureg, src,
 254                              const0, const1);
 255    ureg_MOV(ureg, dst, src);
 256
 257    if (is_composite) {
 258       src = ureg_DECL_vs_input(ureg, input_slot++);
 259       dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, 0);
 260       ureg_MOV(ureg, dst, src);
 261    }
 262
 263    if (is_fill) {
 264       src = ureg_DECL_vs_input(ureg, input_slot++);
 265       dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 266       ureg_MOV(ureg, dst, src);
 267    }
 268
 269    if (has_mask) {
 270       src = ureg_DECL_vs_input(ureg, input_slot++);
 271       dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, 1);
 272       ureg_MOV(ureg, dst, src);
 273    }
 274
 275    ureg_END(ureg);
 276
 277    return ureg_create_shader_and_destroy(ureg, pipe);
 278 }
 279
 280 static void *
 281 create_yuv_shader(struct pipe_context *pipe, struct ureg_program *ureg)
 282 {
 283    struct ureg_src y_sampler, u_sampler, v_sampler;
 284    struct ureg_src pos;
 285    struct ureg_src matrow0, matrow1, matrow2;
 286    struct ureg_dst y, u, v, rgb;
 287    struct ureg_dst out = ureg_DECL_output(ureg,
 288                                           TGSI_SEMANTIC_COLOR,
 289                                           0);
 290
 291    pos = ureg_DECL_fs_input(ureg,
 292                             TGSI_SEMANTIC_GENERIC,
 293                             0,
 294                             TGSI_INTERPOLATE_PERSPECTIVE);
 295
 296    rgb = ureg_DECL_temporary(ureg);
 297    y = ureg_DECL_temporary(ureg);
 298    u = ureg_DECL_temporary(ureg);
 299    v = ureg_DECL_temporary(ureg);
 300
 301    y_sampler = ureg_DECL_sampler(ureg, 0);
 302    u_sampler = ureg_DECL_sampler(ureg, 1);
 303    v_sampler = ureg_DECL_sampler(ureg, 2);
 304
 305    matrow0 = ureg_DECL_constant(ureg, 0);
 306    matrow1 = ureg_DECL_constant(ureg, 1);
 307    matrow2 = ureg_DECL_constant(ureg, 2);
 308
 309    ureg_TEX(ureg, y,
 310             TGSI_TEXTURE_2D, pos, y_sampler);
 311    ureg_TEX(ureg, u,
 312             TGSI_TEXTURE_2D, pos, u_sampler);
 313    ureg_TEX(ureg, v,
 314             TGSI_TEXTURE_2D, pos, v_sampler);
 315
 316    ureg_MUL(ureg, rgb,
 317             ureg_scalar(ureg_src(y), TGSI_SWIZZLE_X),
 318             matrow0);
 319    ureg_MAD(ureg, rgb,
 320             ureg_scalar(ureg_src(u), TGSI_SWIZZLE_X),
 321             matrow1,
 322             ureg_src(rgb));
 323    ureg_MAD(ureg, rgb,
 324             ureg_scalar(ureg_src(v), TGSI_SWIZZLE_X),
 325             matrow2,
 326             ureg_src(rgb));
 327
 328    /* rgb.a = 1; */
 329    ureg_MOV(ureg, ureg_writemask(rgb, TGSI_WRITEMASK_W),
 330             ureg_scalar(matrow0, TGSI_SWIZZLE_X));
 331
 332    ureg_MOV(ureg, out, ureg_src(rgb));
 333
 334    ureg_release_temporary(ureg, rgb);
 335    ureg_release_temporary(ureg, y);
 336    ureg_release_temporary(ureg, u);
 337    ureg_release_temporary(ureg, v);
 338
 339    ureg_END(ureg);
 340
 341    return ureg_create_shader_and_destroy(ureg, pipe);
 342 }
 343
 344 static void *
 345 create_fs(struct pipe_context *pipe,
 346           unsigned fs_traits)
 347 {
 348    struct ureg_program *ureg;
 349    struct ureg_src /*dst_sampler,*/ src_sampler, mask_sampler;
 350    struct ureg_src /*dst_pos,*/ src_input, mask_pos;
 351    struct ureg_dst src, mask;
 352    struct ureg_dst out;
 353    boolean has_mask = fs_traits & FS_MASK;
 354    boolean is_fill = fs_traits & FS_FILL;
 355    boolean is_composite = fs_traits & FS_COMPOSITE;
 356    boolean is_solid   = fs_traits & FS_SOLID_FILL;
 357    boolean is_lingrad = fs_traits & FS_LINGRAD_FILL;
 358    boolean is_radgrad = fs_traits & FS_RADGRAD_FILL;
 359    unsigned comp_alpha = fs_traits & FS_COMPONENT_ALPHA;
 360    boolean is_yuv = fs_traits & FS_YUV;
 361
 362    ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 363    if (ureg == NULL)
 364       return 0;
 365
 366    /* it has to be either a fill, a composite op or a yuv conversion */
 367    debug_assert((is_fill ^ is_composite) ^ is_yuv);
 368
 369    out = ureg_DECL_output(ureg,
 370                           TGSI_SEMANTIC_COLOR,
 371                           0);
 372
 373    if (is_composite) {
 374       src_sampler = ureg_DECL_sampler(ureg, 0);
 375       src_input = ureg_DECL_fs_input(ureg,
 376                                      TGSI_SEMANTIC_GENERIC,
 377                                      0,
 378                                      TGSI_INTERPOLATE_PERSPECTIVE);
 379    } else if (is_fill) {
 380       if (is_solid)
 381          src_input = ureg_DECL_fs_input(ureg,
 382                                         TGSI_SEMANTIC_COLOR,
 383                                         0,
 384                                         TGSI_INTERPOLATE_PERSPECTIVE);
 385       else
 386          src_input = ureg_DECL_fs_input(ureg,
 387                                         TGSI_SEMANTIC_POSITION,
 388                                         0,
 389                                         TGSI_INTERPOLATE_PERSPECTIVE);
 390    } else {
 391       debug_assert(is_yuv);
 392       return create_yuv_shader(pipe, ureg);
 393    }
 394
 395    if (has_mask) {
 396       mask_sampler = ureg_DECL_sampler(ureg, 1);
 397       mask_pos = ureg_DECL_fs_input(ureg,
 398                                     TGSI_SEMANTIC_GENERIC,
 399                                     1,
 400                                     TGSI_INTERPOLATE_PERSPECTIVE);
 401    }
 402
 403 #if 0  /* unused right now */
 404    dst_sampler = ureg_DECL_sampler(ureg, 2);
 405    dst_pos = ureg_DECL_fs_input(ureg,
 406                                 TGSI_SEMANTIC_POSITION,
 407                                 2,
 408                                 TGSI_INTERPOLATE_PERSPECTIVE);
 409 #endif
 410
 411    if (is_composite) {
 412       if (has_mask)
 413          src = ureg_DECL_temporary(ureg);
 414       else
 415          src = out;
 416       ureg_TEX(ureg, src,
 417                TGSI_TEXTURE_2D, src_input, src_sampler);
 418    } else if (is_fill) {
 419       if (is_solid) {
 420          if (has_mask)
 421             src = ureg_dst(src_input);
 422          else
 423             ureg_MOV(ureg, out, src_input);
 424       } else if (is_lingrad || is_radgrad) {
 425          struct ureg_src coords, const0124,
 426             matrow0, matrow1, matrow2;
 427
 428          if (has_mask)
 429             src = ureg_DECL_temporary(ureg);
 430          else
 431             src = out;
 432
 433          coords = ureg_DECL_constant(ureg, 0);
 434          const0124 = ureg_DECL_constant(ureg, 1);
 435          matrow0 = ureg_DECL_constant(ureg, 2);
 436          matrow1 = ureg_DECL_constant(ureg, 3);
 437          matrow2 = ureg_DECL_constant(ureg, 4);
 438
 439          if (is_lingrad) {
 440             linear_gradient(ureg, src,
 441                             src_input, src_sampler,
 442                             coords, const0124,
 443                             matrow0, matrow1, matrow2);
 444          } else if (is_radgrad) {
 445             radial_gradient(ureg, src,
 446                             src_input, src_sampler,
 447                             coords, const0124,
 448                             matrow0, matrow1, matrow2);
 449          }
 450       } else
 451          debug_assert(!"Unknown fill type!");
 452    }
 453
 454    if (has_mask) {
 455       mask = ureg_DECL_temporary(ureg);
 456       ureg_TEX(ureg, mask,
 457                TGSI_TEXTURE_2D, mask_pos, mask_sampler);
 458       /* src IN mask */
 459       src_in_mask(ureg, out, ureg_src(src), ureg_src(mask), comp_alpha);
 460       ureg_release_temporary(ureg, mask);
 461    }
 462
 463    ureg_END(ureg);
 464
 465    return ureg_create_shader_and_destroy(ureg, pipe);
 466 }
 467
 468 struct xorg_shaders * xorg_shaders_create(struct xorg_renderer *r)
 469 {
 470    struct xorg_shaders *sc = CALLOC_STRUCT(xorg_shaders);
 471
 472    sc->r = r;
 473    sc->vs_hash = cso_hash_create();
 474    sc->fs_hash = cso_hash_create();
 475
 476    return sc;
 477 }
 478
 479 static void
 480 cache_destroy(struct cso_context *cso,
 481               struct cso_hash *hash,
 482               unsigned processor)
 483 {
 484    struct cso_hash_iter iter = cso_hash_first_node(hash);
 485    while (!cso_hash_iter_is_null(iter)) {
 486       void *shader = (void *)cso_hash_iter_data(iter);
 487       if (processor == PIPE_SHADER_FRAGMENT) {
 488          cso_delete_fragment_shader(cso, shader);
 489       } else if (processor == PIPE_SHADER_VERTEX) {
 490          cso_delete_vertex_shader(cso, shader);
 491       }
 492       iter = cso_hash_erase(hash, iter);
 493    }
 494    cso_hash_delete(hash);
 495 }
 496
 497 void xorg_shaders_destroy(struct xorg_shaders *sc)
 498 {
 499    cache_destroy(sc->r->cso, sc->vs_hash,
 500                  PIPE_SHADER_VERTEX);
 501    cache_destroy(sc->r->cso, sc->fs_hash,
 502                  PIPE_SHADER_FRAGMENT);
 503
 504    free(sc);
 505 }
 506
 507 static INLINE void *
 508 shader_from_cache(struct pipe_context *pipe,
 509                   unsigned type,
 510                   struct cso_hash *hash,
 511                   unsigned key)
 512 {
 513    void *shader = 0;
 514
 515    struct cso_hash_iter iter = cso_hash_find(hash, key);
 516
 517    if (cso_hash_iter_is_null(iter)) {
 518       if (type == PIPE_SHADER_VERTEX)
 519          shader = create_vs(pipe, key);
 520       else
 521          shader = create_fs(pipe, key);
 522       cso_hash_insert(hash, key, shader);
 523    } else
 524       shader = (void *)cso_hash_iter_data(iter);
 525
 526    return shader;
 527 }
 528
 529 struct xorg_shader xorg_shaders_get(struct xorg_shaders *sc,
 530                                     unsigned vs_traits,
 531                                     unsigned fs_traits)
 532 {
 533    struct xorg_shader shader = { NULL, NULL };
 534    void *vs, *fs;
 535
 536    vs = shader_from_cache(sc->r->pipe, PIPE_SHADER_VERTEX,
 537                           sc->vs_hash, vs_traits);
 538    fs = shader_from_cache(sc->r->pipe, PIPE_SHADER_FRAGMENT,
 539                           sc->fs_hash, fs_traits);
 540
 541    debug_assert(vs && fs);
 542    if (!vs || !fs)
 543       return shader;
 544
 545    shader.vs = vs;
 546    shader.fs = fs;
 547
 548    return shader;
 549 }