src/gallium/state_trackers/xa/xa_tgsi.c

   1 /**********************************************************
   2  * Copyright 2009-2011 VMware, Inc. All rights reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person
   5  * obtaining a copy of this software and associated documentation
   6  * files (the "Software"), to deal in the Software without
   7  * restriction, including without limitation the rights to use, copy,
   8  * modify, merge, publish, distribute, sublicense, and/or sell copies
   9  * of the Software, and to permit persons to whom the Software is
  10  * furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be
  13  * included in all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  *********************************************************
  25  * Authors:
  26  * Zack Rusin <zackr-at-vmware-dot-com>
  27  */
  28 #include "xa_priv.h"
  29
  30 #include "pipe/p_format.h"
  31 #include "pipe/p_context.h"
  32 #include "pipe/p_state.h"
  33 #include "pipe/p_shader_tokens.h"
  34
  35 #include "util/u_memory.h"
  36
  37 #include "tgsi/tgsi_ureg.h"
  38
  39 #include "cso_cache/cso_context.h"
  40 #include "cso_cache/cso_hash.h"
  41
  42 /* Vertex shader:
  43  * IN[0]    = vertex pos
  44  * IN[1]    = src tex coord | solid fill color
  45  * IN[2]    = mask tex coord
  46  * IN[3]    = dst tex coord
  47  * CONST[0] = (2/dst_width, 2/dst_height, 1, 1)
  48  * CONST[1] = (-1, -1, 0, 0)
  49  *
  50  * OUT[0]   = vertex pos
  51  * OUT[1]   = src tex coord | solid fill color
  52  * OUT[2]   = mask tex coord
  53  * OUT[3]   = dst tex coord
  54  */
  55
  56 /* Fragment shader:
  57  * SAMP[0]  = src
  58  * SAMP[1]  = mask
  59  * SAMP[2]  = dst
  60  * IN[0]    = pos src | solid fill color
  61  * IN[1]    = pos mask
  62  * IN[2]    = pos dst
  63  * CONST[0] = (0, 0, 0, 1)
  64  *
  65  * OUT[0] = color
  66  */
  67
  68 static void
  69 print_fs_traits(int fs_traits)
  70 {
  71     const char *strings[] = {
  72         "FS_COMPOSITE",         /* = 1 << 0, */
  73         "FS_MASK",              /* = 1 << 1, */
  74         "FS_SOLID_FILL",        /* = 1 << 2, */
  75         "FS_LINGRAD_FILL",      /* = 1 << 3, */
  76         "FS_RADGRAD_FILL",      /* = 1 << 4, */
  77         "FS_CA_FULL",           /* = 1 << 5, *//* src.rgba * mask.rgba */
  78         "FS_CA_SRCALPHA",       /* = 1 << 6, *//* src.aaaa * mask.rgba */
  79         "FS_YUV",               /* = 1 << 7, */
  80         "FS_SRC_REPEAT_NONE",   /* = 1 << 8, */
  81         "FS_MASK_REPEAT_NONE",  /* = 1 << 9, */
  82         "FS_SRC_SWIZZLE_RGB",   /* = 1 << 10, */
  83         "FS_MASK_SWIZZLE_RGB",  /* = 1 << 11, */
  84         "FS_SRC_SET_ALPHA",     /* = 1 << 12, */
  85         "FS_MASK_SET_ALPHA",    /* = 1 << 13, */
  86         "FS_SRC_LUMINANCE",     /* = 1 << 14, */
  87         "FS_MASK_LUMINANCE",    /* = 1 << 15, */
  88     };
  89     int i, k;
  90
  91     debug_printf("%s: ", __func__);
  92
  93     for (i = 0, k = 1; k < (1 << 16); i++, k <<= 1) {
  94         if (fs_traits & k)
  95             debug_printf("%s, ", strings[i]);
  96     }
  97
  98     debug_printf("\n");
  99 }
 100
 101 struct xa_shaders {
 102     struct xa_context *r;
 103
 104     struct cso_hash *vs_hash;
 105     struct cso_hash *fs_hash;
 106 };
 107
 108 static INLINE void
 109 src_in_mask(struct ureg_program *ureg,
 110             struct ureg_dst dst,
 111             struct ureg_src src,
 112             struct ureg_src mask,
 113             unsigned component_alpha, unsigned mask_luminance)
 114 {
 115     if (component_alpha == FS_CA_FULL) {
 116         ureg_MUL(ureg, dst, src, mask);
 117     } else if (component_alpha == FS_CA_SRCALPHA) {
 118         ureg_MUL(ureg, dst, ureg_scalar(src, TGSI_SWIZZLE_W), mask);
 119     } else {
 120         if (mask_luminance)
 121             ureg_MUL(ureg, dst, src, ureg_scalar(mask, TGSI_SWIZZLE_X));
 122         else
 123             ureg_MUL(ureg, dst, src, ureg_scalar(mask, TGSI_SWIZZLE_W));
 124     }
 125 }
 126
 127 static struct ureg_src
 128 vs_normalize_coords(struct ureg_program *ureg,
 129                     struct ureg_src coords,
 130                     struct ureg_src const0, struct ureg_src const1)
 131 {
 132     struct ureg_dst tmp = ureg_DECL_temporary(ureg);
 133     struct ureg_src ret;
 134
 135     ureg_MAD(ureg, tmp, coords, const0, const1);
 136     ret = ureg_src(tmp);
 137     ureg_release_temporary(ureg, tmp);
 138     return ret;
 139 }
 140
 141 static void
 142 linear_gradient(struct ureg_program *ureg,
 143                 struct ureg_dst out,
 144                 struct ureg_src pos,
 145                 struct ureg_src sampler,
 146                 struct ureg_src coords,
 147                 struct ureg_src const0124,
 148                 struct ureg_src matrow0,
 149                 struct ureg_src matrow1, struct ureg_src matrow2)
 150 {
 151     struct ureg_dst temp0 = ureg_DECL_temporary(ureg);
 152     struct ureg_dst temp1 = ureg_DECL_temporary(ureg);
 153     struct ureg_dst temp2 = ureg_DECL_temporary(ureg);
 154     struct ureg_dst temp3 = ureg_DECL_temporary(ureg);
 155     struct ureg_dst temp4 = ureg_DECL_temporary(ureg);
 156     struct ureg_dst temp5 = ureg_DECL_temporary(ureg);
 157
 158     ureg_MOV(ureg, ureg_writemask(temp0, TGSI_WRITEMASK_XY), pos);
 159     ureg_MOV(ureg,
 160              ureg_writemask(temp0, TGSI_WRITEMASK_Z),
 161              ureg_scalar(const0124, TGSI_SWIZZLE_Y));
 162
 163     ureg_DP3(ureg, temp1, matrow0, ureg_src(temp0));
 164     ureg_DP3(ureg, temp2, matrow1, ureg_src(temp0));
 165     ureg_DP3(ureg, temp3, matrow2, ureg_src(temp0));
 166     ureg_RCP(ureg, temp3, ureg_src(temp3));
 167     ureg_MUL(ureg, temp1, ureg_src(temp1), ureg_src(temp3));
 168     ureg_MUL(ureg, temp2, ureg_src(temp2), ureg_src(temp3));
 169
 170     ureg_MOV(ureg, ureg_writemask(temp4, TGSI_WRITEMASK_X), ureg_src(temp1));
 171     ureg_MOV(ureg, ureg_writemask(temp4, TGSI_WRITEMASK_Y), ureg_src(temp2));
 172
 173     ureg_MUL(ureg, temp0,
 174              ureg_scalar(coords, TGSI_SWIZZLE_Y),
 175              ureg_scalar(ureg_src(temp4), TGSI_SWIZZLE_Y));
 176     ureg_MAD(ureg, temp1,
 177              ureg_scalar(coords, TGSI_SWIZZLE_X),
 178              ureg_scalar(ureg_src(temp4), TGSI_SWIZZLE_X), ureg_src(temp0));
 179
 180     ureg_MUL(ureg, temp2, ureg_src(temp1), ureg_scalar(coords, TGSI_SWIZZLE_Z));
 181
 182     ureg_TEX(ureg, out, TGSI_TEXTURE_1D, ureg_src(temp2), sampler);
 183
 184     ureg_release_temporary(ureg, temp0);
 185     ureg_release_temporary(ureg, temp1);
 186     ureg_release_temporary(ureg, temp2);
 187     ureg_release_temporary(ureg, temp3);
 188     ureg_release_temporary(ureg, temp4);
 189     ureg_release_temporary(ureg, temp5);
 190 }
 191
 192 static void
 193 radial_gradient(struct ureg_program *ureg,
 194                 struct ureg_dst out,
 195                 struct ureg_src pos,
 196                 struct ureg_src sampler,
 197                 struct ureg_src coords,
 198                 struct ureg_src const0124,
 199                 struct ureg_src matrow0,
 200                 struct ureg_src matrow1, struct ureg_src matrow2)
 201 {
 202     struct ureg_dst temp0 = ureg_DECL_temporary(ureg);
 203     struct ureg_dst temp1 = ureg_DECL_temporary(ureg);
 204     struct ureg_dst temp2 = ureg_DECL_temporary(ureg);
 205     struct ureg_dst temp3 = ureg_DECL_temporary(ureg);
 206     struct ureg_dst temp4 = ureg_DECL_temporary(ureg);
 207     struct ureg_dst temp5 = ureg_DECL_temporary(ureg);
 208
 209     ureg_MOV(ureg, ureg_writemask(temp0, TGSI_WRITEMASK_XY), pos);
 210     ureg_MOV(ureg,
 211              ureg_writemask(temp0, TGSI_WRITEMASK_Z),
 212              ureg_scalar(const0124, TGSI_SWIZZLE_Y));
 213
 214     ureg_DP3(ureg, temp1, matrow0, ureg_src(temp0));
 215     ureg_DP3(ureg, temp2, matrow1, ureg_src(temp0));
 216     ureg_DP3(ureg, temp3, matrow2, ureg_src(temp0));
 217     ureg_RCP(ureg, temp3, ureg_src(temp3));
 218     ureg_MUL(ureg, temp1, ureg_src(temp1), ureg_src(temp3));
 219     ureg_MUL(ureg, temp2, ureg_src(temp2), ureg_src(temp3));
 220
 221     ureg_MOV(ureg, ureg_writemask(temp5, TGSI_WRITEMASK_X), ureg_src(temp1));
 222     ureg_MOV(ureg, ureg_writemask(temp5, TGSI_WRITEMASK_Y), ureg_src(temp2));
 223
 224     ureg_MUL(ureg, temp0, ureg_scalar(coords, TGSI_SWIZZLE_Y),
 225              ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_Y));
 226     ureg_MAD(ureg, temp1,
 227              ureg_scalar(coords, TGSI_SWIZZLE_X),
 228              ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_X), ureg_src(temp0));
 229     ureg_ADD(ureg, temp1, ureg_src(temp1), ureg_src(temp1));
 230     ureg_MUL(ureg, temp3,
 231              ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_Y),
 232              ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_Y));
 233     ureg_MAD(ureg, temp4,
 234              ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_X),
 235              ureg_scalar(ureg_src(temp5), TGSI_SWIZZLE_X), ureg_src(temp3));
 236     ureg_MOV(ureg, temp4, ureg_negate(ureg_src(temp4)));
 237     ureg_MUL(ureg, temp2, ureg_scalar(coords, TGSI_SWIZZLE_Z), ureg_src(temp4));
 238     ureg_MUL(ureg, temp0,
 239              ureg_scalar(const0124, TGSI_SWIZZLE_W), ureg_src(temp2));
 240     ureg_MUL(ureg, temp3, ureg_src(temp1), ureg_src(temp1));
 241     ureg_SUB(ureg, temp2, ureg_src(temp3), ureg_src(temp0));
 242     ureg_RSQ(ureg, temp2, ureg_abs(ureg_src(temp2)));
 243     ureg_RCP(ureg, temp2, ureg_src(temp2));
 244     ureg_SUB(ureg, temp1, ureg_src(temp2), ureg_src(temp1));
 245     ureg_ADD(ureg, temp0,
 246              ureg_scalar(coords, TGSI_SWIZZLE_Z),
 247              ureg_scalar(coords, TGSI_SWIZZLE_Z));
 248     ureg_RCP(ureg, temp0, ureg_src(temp0));
 249     ureg_MUL(ureg, temp2, ureg_src(temp1), ureg_src(temp0));
 250     ureg_TEX(ureg, out, TGSI_TEXTURE_1D, ureg_src(temp2), sampler);
 251
 252     ureg_release_temporary(ureg, temp0);
 253     ureg_release_temporary(ureg, temp1);
 254     ureg_release_temporary(ureg, temp2);
 255     ureg_release_temporary(ureg, temp3);
 256     ureg_release_temporary(ureg, temp4);
 257     ureg_release_temporary(ureg, temp5);
 258 }
 259
 260 static void *
 261 create_vs(struct pipe_context *pipe, unsigned vs_traits)
 262 {
 263     struct ureg_program *ureg;
 264     struct ureg_src src;
 265     struct ureg_dst dst;
 266     struct ureg_src const0, const1;
 267     boolean is_fill = (vs_traits & VS_FILL) != 0;
 268     boolean is_composite = (vs_traits & VS_COMPOSITE) != 0;
 269     boolean has_mask = (vs_traits & VS_MASK) != 0;
 270     boolean is_yuv = (vs_traits & VS_YUV) != 0;
 271     unsigned input_slot = 0;
 272
 273     ureg = ureg_create(TGSI_PROCESSOR_VERTEX);
 274     if (ureg == NULL)
 275         return 0;
 276
 277     const0 = ureg_DECL_constant(ureg, 0);
 278     const1 = ureg_DECL_constant(ureg, 1);
 279
 280     /* it has to be either a fill or a composite op */
 281     debug_assert((is_fill ^ is_composite) ^ is_yuv);
 282
 283     src = ureg_DECL_vs_input(ureg, input_slot++);
 284     dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0);
 285     src = vs_normalize_coords(ureg, src, const0, const1);
 286     ureg_MOV(ureg, dst, src);
 287
 288     if (is_yuv) {
 289         src = ureg_DECL_vs_input(ureg, input_slot++);
 290         dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, 0);
 291         ureg_MOV(ureg, dst, src);
 292     }
 293
 294     if (is_composite) {
 295         src = ureg_DECL_vs_input(ureg, input_slot++);
 296         dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, 0);
 297         ureg_MOV(ureg, dst, src);
 298     }
 299
 300     if (is_fill) {
 301         src = ureg_DECL_vs_input(ureg, input_slot++);
 302         dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 303         ureg_MOV(ureg, dst, src);
 304     }
 305
 306     if (has_mask) {
 307         src = ureg_DECL_vs_input(ureg, input_slot++);
 308         dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, 1);
 309         ureg_MOV(ureg, dst, src);
 310     }
 311
 312     ureg_END(ureg);
 313
 314     return ureg_create_shader_and_destroy(ureg, pipe);
 315 }
 316
 317 static void *
 318 create_yuv_shader(struct pipe_context *pipe, struct ureg_program *ureg)
 319 {
 320     struct ureg_src y_sampler, u_sampler, v_sampler;
 321     struct ureg_src pos;
 322     struct ureg_src matrow0, matrow1, matrow2;
 323     struct ureg_dst y, u, v, rgb;
 324     struct ureg_dst out = ureg_DECL_output(ureg,
 325                                            TGSI_SEMANTIC_COLOR,
 326                                            0);
 327
 328     pos = ureg_DECL_fs_input(ureg,
 329                              TGSI_SEMANTIC_GENERIC, 0,
 330                              TGSI_INTERPOLATE_PERSPECTIVE);
 331
 332     rgb = ureg_DECL_temporary(ureg);
 333     y = ureg_DECL_temporary(ureg);
 334     u = ureg_DECL_temporary(ureg);
 335     v = ureg_DECL_temporary(ureg);
 336
 337     y_sampler = ureg_DECL_sampler(ureg, 0);
 338     u_sampler = ureg_DECL_sampler(ureg, 1);
 339     v_sampler = ureg_DECL_sampler(ureg, 2);
 340
 341     matrow0 = ureg_DECL_constant(ureg, 0);
 342     matrow1 = ureg_DECL_constant(ureg, 1);
 343     matrow2 = ureg_DECL_constant(ureg, 2);
 344
 345     ureg_TEX(ureg, y, TGSI_TEXTURE_2D, pos, y_sampler);
 346     ureg_TEX(ureg, u, TGSI_TEXTURE_2D, pos, u_sampler);
 347     ureg_TEX(ureg, v, TGSI_TEXTURE_2D, pos, v_sampler);
 348
 349     ureg_SUB(ureg, u, ureg_src(u), ureg_scalar(matrow0, TGSI_SWIZZLE_W));
 350     ureg_SUB(ureg, v, ureg_src(v), ureg_scalar(matrow0, TGSI_SWIZZLE_W));
 351
 352     ureg_MUL(ureg, rgb, ureg_scalar(ureg_src(y), TGSI_SWIZZLE_X), matrow0);
 353     ureg_MAD(ureg, rgb,
 354              ureg_scalar(ureg_src(u), TGSI_SWIZZLE_X), matrow1, ureg_src(rgb));
 355     ureg_MAD(ureg, rgb,
 356              ureg_scalar(ureg_src(v), TGSI_SWIZZLE_X), matrow2, ureg_src(rgb));
 357
 358     /* rgb.a = 1; */
 359     ureg_MOV(ureg, ureg_writemask(rgb, TGSI_WRITEMASK_W),
 360              ureg_scalar(matrow0, TGSI_SWIZZLE_X));
 361
 362     ureg_MOV(ureg, out, ureg_src(rgb));
 363
 364     ureg_release_temporary(ureg, rgb);
 365     ureg_release_temporary(ureg, y);
 366     ureg_release_temporary(ureg, u);
 367     ureg_release_temporary(ureg, v);
 368
 369     ureg_END(ureg);
 370
 371     return ureg_create_shader_and_destroy(ureg, pipe);
 372 }
 373
 374 static INLINE void
 375 xrender_tex(struct ureg_program *ureg,
 376             struct ureg_dst dst,
 377             struct ureg_src coords,
 378             struct ureg_src sampler,
 379             struct ureg_src imm0,
 380             boolean repeat_none, boolean swizzle, boolean set_alpha)
 381 {
 382     if (repeat_none) {
 383         struct ureg_dst tmp0 = ureg_DECL_temporary(ureg);
 384         struct ureg_dst tmp1 = ureg_DECL_temporary(ureg);
 385
 386         ureg_SGT(ureg, tmp1, ureg_swizzle(coords,
 387                                           TGSI_SWIZZLE_X,
 388                                           TGSI_SWIZZLE_Y,
 389                                           TGSI_SWIZZLE_X,
 390                                           TGSI_SWIZZLE_Y), ureg_scalar(imm0,
 391                                                                        TGSI_SWIZZLE_X));
 392         ureg_SLT(ureg, tmp0,
 393                  ureg_swizzle(coords, TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y,
 394                               TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y), ureg_scalar(imm0,
 395                                                                            TGSI_SWIZZLE_W));
 396         ureg_MIN(ureg, tmp0, ureg_src(tmp0), ureg_src(tmp1));
 397         ureg_MIN(ureg, tmp0, ureg_scalar(ureg_src(tmp0), TGSI_SWIZZLE_X),
 398                  ureg_scalar(ureg_src(tmp0), TGSI_SWIZZLE_Y));
 399         ureg_TEX(ureg, tmp1, TGSI_TEXTURE_2D, coords, sampler);
 400         if (swizzle)
 401             ureg_MOV(ureg, tmp1, ureg_swizzle(ureg_src(tmp1),
 402                                               TGSI_SWIZZLE_Z,
 403                                               TGSI_SWIZZLE_Y, TGSI_SWIZZLE_X,
 404                                               TGSI_SWIZZLE_W));
 405         if (set_alpha)
 406             ureg_MOV(ureg,
 407                      ureg_writemask(tmp1, TGSI_WRITEMASK_W),
 408                      ureg_scalar(imm0, TGSI_SWIZZLE_W));
 409         ureg_MUL(ureg, dst, ureg_src(tmp1), ureg_src(tmp0));
 410         ureg_release_temporary(ureg, tmp0);
 411         ureg_release_temporary(ureg, tmp1);
 412     } else {
 413         if (swizzle) {
 414             struct ureg_dst tmp = ureg_DECL_temporary(ureg);
 415
 416             ureg_TEX(ureg, tmp, TGSI_TEXTURE_2D, coords, sampler);
 417             ureg_MOV(ureg, dst, ureg_swizzle(ureg_src(tmp),
 418                                              TGSI_SWIZZLE_Z,
 419                                              TGSI_SWIZZLE_Y, TGSI_SWIZZLE_X,
 420                                              TGSI_SWIZZLE_W));
 421             ureg_release_temporary(ureg, tmp);
 422         } else {
 423             ureg_TEX(ureg, dst, TGSI_TEXTURE_2D, coords, sampler);
 424         }
 425         if (set_alpha)
 426             ureg_MOV(ureg,
 427                      ureg_writemask(dst, TGSI_WRITEMASK_W),
 428                      ureg_scalar(imm0, TGSI_SWIZZLE_W));
 429     }
 430 }
 431
 432 static void *
 433 create_fs(struct pipe_context *pipe, unsigned fs_traits)
 434 {
 435     struct ureg_program *ureg;
 436     struct ureg_src /*dst_sampler, */ src_sampler, mask_sampler;
 437     struct ureg_src /*dst_pos, */ src_input, mask_pos;
 438     struct ureg_dst src, mask;
 439     struct ureg_dst out;
 440     struct ureg_src imm0 = { 0 };
 441     unsigned has_mask = (fs_traits & FS_MASK) != 0;
 442     unsigned is_fill = (fs_traits & FS_FILL) != 0;
 443     unsigned is_composite = (fs_traits & FS_COMPOSITE) != 0;
 444     unsigned is_solid = (fs_traits & FS_SOLID_FILL) != 0;
 445     unsigned is_lingrad = (fs_traits & FS_LINGRAD_FILL) != 0;
 446     unsigned is_radgrad = (fs_traits & FS_RADGRAD_FILL) != 0;
 447     unsigned comp_alpha_mask = fs_traits & FS_COMPONENT_ALPHA;
 448     unsigned is_yuv = (fs_traits & FS_YUV) != 0;
 449     unsigned src_repeat_none = (fs_traits & FS_SRC_REPEAT_NONE) != 0;
 450     unsigned mask_repeat_none = (fs_traits & FS_MASK_REPEAT_NONE) != 0;
 451     unsigned src_swizzle = (fs_traits & FS_SRC_SWIZZLE_RGB) != 0;
 452     unsigned mask_swizzle = (fs_traits & FS_MASK_SWIZZLE_RGB) != 0;
 453     unsigned src_set_alpha = (fs_traits & FS_SRC_SET_ALPHA) != 0;
 454     unsigned mask_set_alpha = (fs_traits & FS_MASK_SET_ALPHA) != 0;
 455     unsigned src_luminance = (fs_traits & FS_SRC_LUMINANCE) != 0;
 456     unsigned mask_luminance = (fs_traits & FS_MASK_LUMINANCE) != 0;
 457
 458 #if 0
 459     print_fs_traits(fs_traits);
 460 #else
 461     (void)print_fs_traits;
 462 #endif
 463
 464     ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 465     if (ureg == NULL)
 466         return 0;
 467
 468     /* it has to be either a fill, a composite op or a yuv conversion */
 469     debug_assert((is_fill ^ is_composite) ^ is_yuv);
 470     (void)is_yuv;
 471
 472     out = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 473
 474     if (src_repeat_none || mask_repeat_none ||
 475         src_set_alpha || mask_set_alpha || src_luminance) {
 476         imm0 = ureg_imm4f(ureg, 0, 0, 0, 1);
 477     }
 478     if (is_composite) {
 479         src_sampler = ureg_DECL_sampler(ureg, 0);
 480         src_input = ureg_DECL_fs_input(ureg,
 481                                        TGSI_SEMANTIC_GENERIC, 0,
 482                                        TGSI_INTERPOLATE_PERSPECTIVE);
 483     } else if (is_fill) {
 484         if (is_solid)
 485             src_input = ureg_DECL_fs_input(ureg,
 486                                            TGSI_SEMANTIC_COLOR, 0,
 487                                            TGSI_INTERPOLATE_PERSPECTIVE);
 488         else
 489             src_input = ureg_DECL_fs_input(ureg,
 490                                            TGSI_SEMANTIC_POSITION, 0,
 491                                            TGSI_INTERPOLATE_PERSPECTIVE);
 492     } else {
 493         debug_assert(is_yuv);
 494         return create_yuv_shader(pipe, ureg);
 495     }
 496
 497     if (has_mask) {
 498         mask_sampler = ureg_DECL_sampler(ureg, 1);
 499         mask_pos = ureg_DECL_fs_input(ureg,
 500                                       TGSI_SEMANTIC_GENERIC, 1,
 501                                       TGSI_INTERPOLATE_PERSPECTIVE);
 502     }
 503 #if 0                           /* unused right now */
 504     dst_sampler = ureg_DECL_sampler(ureg, 2);
 505     dst_pos = ureg_DECL_fs_input(ureg,
 506                                  TGSI_SEMANTIC_POSITION, 2,
 507                                  TGSI_INTERPOLATE_PERSPECTIVE);
 508 #endif
 509
 510     if (is_composite) {
 511         if (has_mask || src_luminance)
 512             src = ureg_DECL_temporary(ureg);
 513         else
 514             src = out;
 515         xrender_tex(ureg, src, src_input, src_sampler, imm0,
 516                     src_repeat_none, src_swizzle, src_set_alpha);
 517     } else if (is_fill) {
 518         if (is_solid) {
 519             if (has_mask || src_luminance)
 520                 src = ureg_dst(src_input);
 521             else
 522                 ureg_MOV(ureg, out, src_input);
 523         } else if (is_lingrad || is_radgrad) {
 524             struct ureg_src coords, const0124, matrow0, matrow1, matrow2;
 525
 526             if (has_mask || src_luminance)
 527                 src = ureg_DECL_temporary(ureg);
 528             else
 529                 src = out;
 530
 531             coords = ureg_DECL_constant(ureg, 0);
 532             const0124 = ureg_DECL_constant(ureg, 1);
 533             matrow0 = ureg_DECL_constant(ureg, 2);
 534             matrow1 = ureg_DECL_constant(ureg, 3);
 535             matrow2 = ureg_DECL_constant(ureg, 4);
 536
 537             if (is_lingrad) {
 538                 linear_gradient(ureg, src,
 539                                 src_input, src_sampler,
 540                                 coords, const0124, matrow0, matrow1, matrow2);
 541             } else if (is_radgrad) {
 542                 radial_gradient(ureg, src,
 543                                 src_input, src_sampler,
 544                                 coords, const0124, matrow0, matrow1, matrow2);
 545             }
 546         } else
 547             debug_assert(!"Unknown fill type!");
 548     }
 549     if (src_luminance) {
 550         ureg_MOV(ureg, src, ureg_scalar(ureg_src(src), TGSI_SWIZZLE_X));
 551         ureg_MOV(ureg, ureg_writemask(src, TGSI_WRITEMASK_XYZ),
 552                  ureg_scalar(imm0, TGSI_SWIZZLE_X));
 553         if (!has_mask)
 554             ureg_MOV(ureg, out, ureg_src(src));
 555     }
 556
 557     if (has_mask) {
 558         mask = ureg_DECL_temporary(ureg);
 559         xrender_tex(ureg, mask, mask_pos, mask_sampler, imm0,
 560                     mask_repeat_none, mask_swizzle, mask_set_alpha);
 561         /* src IN mask */
 562         src_in_mask(ureg, out, ureg_src(src), ureg_src(mask),
 563                     comp_alpha_mask, mask_luminance);
 564         ureg_release_temporary(ureg, mask);
 565     }
 566
 567     ureg_END(ureg);
 568
 569     return ureg_create_shader_and_destroy(ureg, pipe);
 570 }
 571
 572 struct xa_shaders *
 573 xa_shaders_create(struct xa_context *r)
 574 {
 575     struct xa_shaders *sc = CALLOC_STRUCT(xa_shaders);
 576
 577     sc->r = r;
 578     sc->vs_hash = cso_hash_create();
 579     sc->fs_hash = cso_hash_create();
 580
 581     return sc;
 582 }
 583
 584 static void
 585 cache_destroy(struct cso_context *cso,
 586               struct cso_hash *hash, unsigned processor)
 587 {
 588     struct cso_hash_iter iter = cso_hash_first_node(hash);
 589
 590     while (!cso_hash_iter_is_null(iter)) {
 591         void *shader = (void *)cso_hash_iter_data(iter);
 592
 593         if (processor == PIPE_SHADER_FRAGMENT) {
 594             cso_delete_fragment_shader(cso, shader);
 595         } else if (processor == PIPE_SHADER_VERTEX) {
 596             cso_delete_vertex_shader(cso, shader);
 597         }
 598         iter = cso_hash_erase(hash, iter);
 599     }
 600     cso_hash_delete(hash);
 601 }
 602
 603 void
 604 xa_shaders_destroy(struct xa_shaders *sc)
 605 {
 606     cache_destroy(sc->r->cso, sc->vs_hash, PIPE_SHADER_VERTEX);
 607     cache_destroy(sc->r->cso, sc->fs_hash, PIPE_SHADER_FRAGMENT);
 608
 609     FREE(sc);
 610 }
 611
 612 static INLINE void *
 613 shader_from_cache(struct pipe_context *pipe,
 614                   unsigned type, struct cso_hash *hash, unsigned key)
 615 {
 616     void *shader = 0;
 617
 618     struct cso_hash_iter iter = cso_hash_find(hash, key);
 619
 620     if (cso_hash_iter_is_null(iter)) {
 621         if (type == PIPE_SHADER_VERTEX)
 622             shader = create_vs(pipe, key);
 623         else
 624             shader = create_fs(pipe, key);
 625         cso_hash_insert(hash, key, shader);
 626     } else
 627         shader = (void *)cso_hash_iter_data(iter);
 628
 629     return shader;
 630 }
 631
 632 struct xa_shader
 633 xa_shaders_get(struct xa_shaders *sc, unsigned vs_traits, unsigned fs_traits)
 634 {
 635     struct xa_shader shader = { NULL, NULL };
 636     void *vs, *fs;
 637
 638     vs = shader_from_cache(sc->r->pipe, PIPE_SHADER_VERTEX,
 639                            sc->vs_hash, vs_traits);
 640     fs = shader_from_cache(sc->r->pipe, PIPE_SHADER_FRAGMENT,
 641                            sc->fs_hash, fs_traits);
 642
 643     debug_assert(vs && fs);
 644     if (!vs || !fs)
 645         return shader;
 646
 647     shader.vs = vs;
 648     shader.fs = fs;
 649
 650     return shader;
 651 }