src/gallium/state_trackers/g3dvl/vl_r16snorm_mc_buf_shaders.inc

   1 static int vlCreateVertexShaderIMB
   2 (
   3         struct vlR16SnormBufferedMC *mc
   4 )
   5 {
   6         const unsigned int              max_tokens = 50;
   7
   8         struct pipe_context             *pipe;
   9         struct pipe_shader_state        vs;
  10         struct tgsi_token               *tokens;
  11         struct tgsi_header              *header;
  12
  13         struct tgsi_full_declaration    decl;
  14         struct tgsi_full_instruction    inst;
  15
  16         unsigned int                    ti;
  17         unsigned int                    i;
  18
  19         assert(mc);
  20
  21         pipe = mc->pipe;
  22         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
  23
  24         /* Version */
  25         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
  26         /* Header */
  27         header = (struct tgsi_header*)&tokens[1];
  28         *header = tgsi_build_header();
  29         /* Processor */
  30         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_VERTEX, header);
  31
  32         ti = 3;
  33
  34         /*
  35          * decl i0              ; Vertex pos
  36          * decl i1              ; Luma texcoords
  37          * decl i2              ; Chroma Cb texcoords
  38          * decl i3              ; Chroma Cr texcoords
  39          */
  40         for (i = 0; i < 4; i++)
  41         {
  42                 decl = vl_decl_input(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
  43                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
  44         }
  45
  46         /*
  47          * decl o0              ; Vertex pos
  48          * decl o1              ; Luma texcoords
  49          * decl o2              ; Chroma Cb texcoords
  50          * decl o3              ; Chroma Cr texcoords
  51          */
  52         for (i = 0; i < 4; i++)
  53         {
  54                 decl = vl_decl_output(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
  55                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
  56         }
  57
  58         /*
  59          * mov o0, i0           ; Move input vertex pos to output
  60          * mov o1, i1           ; Move input luma texcoords to output
  61          * mov o2, i2           ; Move input chroma Cb texcoords to output
  62          * mov o3, i3           ; Move input chroma Cr texcoords to output
  63          */
  64         for (i = 0; i < 4; ++i)
  65         {
  66                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_OUTPUT, i, TGSI_FILE_INPUT, i);
  67                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
  68         }
  69
  70         /* end */
  71         inst = vl_end();
  72         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
  73
  74         vs.tokens = tokens;
  75         mc->i_vs = pipe->create_vs_state(pipe, &vs);
  76         free(tokens);
  77
  78         return 0;
  79 }
  80
  81 static int vlCreateFragmentShaderIMB
  82 (
  83         struct vlR16SnormBufferedMC *mc
  84 )
  85 {
  86         const unsigned int              max_tokens = 100;
  87
  88         struct pipe_context             *pipe;
  89         struct pipe_shader_state        fs;
  90         struct tgsi_token               *tokens;
  91         struct tgsi_header              *header;
  92
  93         struct tgsi_full_declaration    decl;
  94         struct tgsi_full_instruction    inst;
  95
  96         unsigned int                    ti;
  97         unsigned int                    i;
  98
  99         assert(mc);
 100
 101         pipe = mc->pipe;
 102         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 103
 104         /* Version */
 105         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 106         /* Header */
 107         header = (struct tgsi_header*)&tokens[1];
 108         *header = tgsi_build_header();
 109         /* Processor */
 110         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_FRAGMENT, header);
 111
 112         ti = 3;
 113
 114         /*
 115          * decl i0                      ; Luma texcoords
 116          * decl i1                      ; Chroma Cb texcoords
 117          * decl i2                      ; Chroma Cr texcoords
 118          */
 119         for (i = 0; i < 3; ++i)
 120         {
 121                 decl = vl_decl_interpolated_input(TGSI_SEMANTIC_GENERIC, i + 1, i, i, TGSI_INTERPOLATE_LINEAR);
 122                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 123         }
 124
 125         /* decl c0                      ; Scaling factor, rescales 16-bit snorm to 9-bit snorm */
 126         decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 0);
 127         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 128
 129         /* decl o0                      ; Fragment color */
 130         decl = vl_decl_output(TGSI_SEMANTIC_COLOR, 0, 0, 0);
 131         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 132
 133         /* decl t0, t1 */
 134         decl = vl_decl_temps(0, 1);
 135         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 136
 137         /*
 138          * decl s0                      ; Sampler for luma texture
 139          * decl s1                      ; Sampler for chroma Cb texture
 140          * decl s2                      ; Sampler for chroma Cr texture
 141          */
 142         for (i = 0; i < 3; ++i)
 143         {
 144                 decl = vl_decl_samplers(i, i);
 145                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header,max_tokens - ti);
 146         }
 147
 148         /*
 149          * tex2d t1, i0, s0             ; Read texel from luma texture
 150          * mov t0.x, t1.x               ; Move luma sample into .x component
 151          * tex2d t1, i1, s1             ; Read texel from chroma Cb texture
 152          * mov t0.y, t1.x               ; Move Cb sample into .y component
 153          * tex2d t1, i2, s2             ; Read texel from chroma Cr texture
 154          * mov t0.z, t1.x               ; Move Cr sample into .z component
 155          */
 156         for (i = 0; i < 3; ++i)
 157         {
 158                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_INPUT, i, TGSI_FILE_SAMPLER, i);
 159                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 160
 161                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
 162                 inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 163                 inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 164                 inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
 165                 inst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X << i;
 166                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 167         }
 168
 169         /* mul o0, t0, c0               ; Rescale texel to correct range */
 170         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_OUTPUT, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_CONSTANT, 0);
 171         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 172
 173         /* end */
 174         inst = vl_end();
 175         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 176
 177         fs.tokens = tokens;
 178         mc->i_fs = pipe->create_fs_state(pipe, &fs);
 179         free(tokens);
 180
 181         return 0;
 182 }
 183
 184 static int vlCreateVertexShaderFramePMB
 185 (
 186         struct vlR16SnormBufferedMC *mc
 187 )
 188 {
 189         const unsigned int              max_tokens = 100;
 190
 191         struct pipe_context             *pipe;
 192         struct pipe_shader_state        vs;
 193         struct tgsi_token               *tokens;
 194         struct tgsi_header              *header;
 195
 196         struct tgsi_full_declaration    decl;
 197         struct tgsi_full_instruction    inst;
 198
 199         unsigned int                    ti;
 200         unsigned int                    i;
 201
 202         assert(mc);
 203
 204         pipe = mc->pipe;
 205         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 206
 207         /* Version */
 208         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 209         /* Header */
 210         header = (struct tgsi_header*)&tokens[1];
 211         *header = tgsi_build_header();
 212         /* Processor */
 213         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_VERTEX, header);
 214
 215         ti = 3;
 216
 217         /*
 218          * decl i0              ; Vertex pos
 219          * decl i1              ; Luma texcoords
 220          * decl i2              ; Chroma Cb texcoords
 221          * decl i3              ; Chroma Cr texcoords
 222          * decl i4              ; Ref surface top field texcoords
 223          * decl i5              ; Ref surface bottom field texcoords (unused, packed in the same stream)
 224          */
 225         for (i = 0; i < 6; i++)
 226         {
 227                 decl = vl_decl_input(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 228                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 229         }
 230
 231         /*
 232          * decl o0              ; Vertex pos
 233          * decl o1              ; Luma texcoords
 234          * decl o2              ; Chroma Cb texcoords
 235          * decl o3              ; Chroma Cr texcoords
 236          * decl o4              ; Ref macroblock texcoords
 237          */
 238         for (i = 0; i < 5; i++)
 239         {
 240                 decl = vl_decl_output(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 241                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 242         }
 243
 244         /*
 245          * mov o0, i0           ; Move input vertex pos to output
 246          * mov o1, i1           ; Move input luma texcoords to output
 247          * mov o2, i2           ; Move input chroma Cb texcoords to output
 248          * mov o3, i3           ; Move input chroma Cr texcoords to output
 249          */
 250         for (i = 0; i < 4; ++i)
 251         {
 252                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_OUTPUT, i, TGSI_FILE_INPUT, i);
 253                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 254         }
 255
 256         /* add o4, i0, i4       ; Translate vertex pos by motion vec to form ref macroblock texcoords */
 257         inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, 4, TGSI_FILE_INPUT, 0, TGSI_FILE_INPUT, 4);
 258         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 259
 260         /* end */
 261         inst = vl_end();
 262         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 263
 264         vs.tokens = tokens;
 265         mc->p_vs[0] = pipe->create_vs_state(pipe, &vs);
 266         free(tokens);
 267
 268         return 0;
 269 }
 270
 271 static int vlCreateVertexShaderFieldPMB
 272 (
 273         struct vlR16SnormBufferedMC *mc
 274 )
 275 {
 276         const unsigned int              max_tokens = 100;
 277
 278         struct pipe_context             *pipe;
 279         struct pipe_shader_state        vs;
 280         struct tgsi_token               *tokens;
 281         struct tgsi_header              *header;
 282
 283         struct tgsi_full_declaration    decl;
 284         struct tgsi_full_instruction    inst;
 285
 286         unsigned int                    ti;
 287         unsigned int                    i;
 288
 289         assert(mc);
 290
 291         pipe = mc->pipe;
 292         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 293
 294         /* Version */
 295         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 296         /* Header */
 297         header = (struct tgsi_header*)&tokens[1];
 298         *header = tgsi_build_header();
 299         /* Processor */
 300         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_VERTEX, header);
 301
 302         ti = 3;
 303
 304         /*
 305          * decl i0              ; Vertex pos
 306          * decl i1              ; Luma texcoords
 307          * decl i2              ; Chroma Cb texcoords
 308          * decl i3              ; Chroma Cr texcoords
 309          * decl i4              ; Ref macroblock top field texcoords
 310          * decl i5              ; Ref macroblock bottom field texcoords
 311          */
 312         for (i = 0; i < 6; i++)
 313         {
 314                 decl = vl_decl_input(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 315                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 316         }
 317
 318         /* decl c0              ; Render target dimensions */
 319         decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 0);
 320         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 321
 322         /*
 323          * decl o0              ; Vertex pos
 324          * decl o1              ; Luma texcoords
 325          * decl o2              ; Chroma Cb texcoords
 326          * decl o3              ; Chroma Cr texcoords
 327          * decl o4              ; Ref macroblock top field texcoords
 328          * decl o5              ; Ref macroblock bottom field texcoords
 329          * decl o6              ; Denormalized vertex pos
 330          */
 331         for (i = 0; i < 7; i++)
 332         {
 333                 decl = vl_decl_output(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 334                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 335         }
 336
 337         /*
 338          * mov o0, i0           ; Move input vertex pos to output
 339          * mov o1, i1           ; Move input luma texcoords to output
 340          * mov o2, i2           ; Move input chroma Cb texcoords to output
 341          * mov o3, i3           ; Move input chroma Cr texcoords to output
 342          */
 343         for (i = 0; i < 4; ++i)
 344         {
 345                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_OUTPUT, i, TGSI_FILE_INPUT, i);
 346                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 347         }
 348
 349         /*
 350          * add o4, i0, i4       ; Translate vertex pos by motion vec to form top field macroblock texcoords
 351          * add o5, i0, i5       ; Translate vertex pos by motion vec to form bottom field macroblock texcoords
 352          */
 353         for (i = 0; i < 2; ++i)
 354         {
 355                 inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, i + 4, TGSI_FILE_INPUT, 0, TGSI_FILE_INPUT, i + 4);
 356                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 357         }
 358
 359         /* mul o6, i0, c0       ; Denorm vertex pos */
 360         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_OUTPUT, 6, TGSI_FILE_INPUT, 0, TGSI_FILE_CONSTANT, 0);
 361         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 362
 363         /* end */
 364         inst = vl_end();
 365         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 366
 367         vs.tokens = tokens;
 368         mc->p_vs[1] = pipe->create_vs_state(pipe, &vs);
 369         free(tokens);
 370
 371         return 0;
 372 }
 373
 374 static int vlCreateFragmentShaderFramePMB
 375 (
 376         struct vlR16SnormBufferedMC *mc
 377 )
 378 {
 379         const unsigned int              max_tokens = 100;
 380
 381         struct pipe_context             *pipe;
 382         struct pipe_shader_state        fs;
 383         struct tgsi_token               *tokens;
 384         struct tgsi_header              *header;
 385
 386         struct tgsi_full_declaration    decl;
 387         struct tgsi_full_instruction    inst;
 388
 389         unsigned int                    ti;
 390         unsigned int                    i;
 391
 392         assert(mc);
 393
 394         pipe = mc->pipe;
 395         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 396
 397         /* Version */
 398         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 399         /* Header */
 400         header = (struct tgsi_header*)&tokens[1];
 401         *header = tgsi_build_header();
 402         /* Processor */
 403         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_FRAGMENT, header);
 404
 405         ti = 3;
 406
 407         /*
 408          * decl i0                      ; Luma texcoords
 409          * decl i1                      ; Chroma Cb texcoords
 410          * decl i2                      ; Chroma Cr texcoords
 411          * decl i3                      ; Ref macroblock texcoords
 412          */
 413         for (i = 0; i < 4; ++i)
 414         {
 415                 decl = vl_decl_interpolated_input(TGSI_SEMANTIC_GENERIC, i + 1, i, i, TGSI_INTERPOLATE_LINEAR);
 416                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 417         }
 418
 419         /* decl c0                      ; Scaling factor, rescales 16-bit snorm to 9-bit snorm */
 420         decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 0);
 421         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 422
 423         /* decl o0                      ; Fragment color */
 424         decl = vl_decl_output(TGSI_SEMANTIC_COLOR, 0, 0, 0);
 425         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 426
 427         /* decl t0, t1 */
 428         decl = vl_decl_temps(0, 1);
 429         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 430
 431         /*
 432          * decl s0                      ; Sampler for luma texture
 433          * decl s1                      ; Sampler for chroma Cb texture
 434          * decl s2                      ; Sampler for chroma Cr texture
 435          * decl s3                      ; Sampler for ref surface texture
 436          */
 437         for (i = 0; i < 4; ++i)
 438         {
 439                 decl = vl_decl_samplers(i, i);
 440                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 441         }
 442
 443         /*
 444          * tex2d t1, i0, s0             ; Read texel from luma texture
 445          * mov t0.x, t1.x               ; Move luma sample into .x component
 446          * tex2d t1, i1, s1             ; Read texel from chroma Cb texture
 447          * mov t0.y, t1.x               ; Move Cb sample into .y component
 448          * tex2d t1, i2, s2             ; Read texel from chroma Cr texture
 449          * mov t0.z, t1.x               ; Move Cr sample into .z component
 450          */
 451         for (i = 0; i < 3; ++i)
 452         {
 453                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_INPUT, i, TGSI_FILE_SAMPLER, i);
 454                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 455
 456                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
 457                 inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 458                 inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 459                 inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
 460                 inst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X << i;
 461                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 462         }
 463
 464         /* mul t0, t0, c0               ; Rescale texel to correct range */
 465         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_CONSTANT, 0);
 466         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 467
 468         /* tex2d t1, i3, s3             ; Read texel from ref macroblock */
 469         inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_INPUT, 3, TGSI_FILE_SAMPLER, 3);
 470         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 471
 472         /* add o0, t0, t1               ; Add ref and differential to form final output */
 473         inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
 474         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 475
 476         /* end */
 477         inst = vl_end();
 478         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 479
 480         fs.tokens = tokens;
 481         mc->p_fs[0] = pipe->create_fs_state(pipe, &fs);
 482         free(tokens);
 483
 484         return 0;
 485 }
 486
 487 static int vlCreateFragmentShaderFieldPMB
 488 (
 489         struct vlR16SnormBufferedMC *mc
 490 )
 491 {
 492         const unsigned int              max_tokens = 200;
 493
 494         struct pipe_context             *pipe;
 495         struct pipe_shader_state        fs;
 496         struct tgsi_token               *tokens;
 497         struct tgsi_header              *header;
 498
 499         struct tgsi_full_declaration    decl;
 500         struct tgsi_full_instruction    inst;
 501
 502         unsigned int                    ti;
 503         unsigned int                    i;
 504
 505         assert(mc);
 506
 507         pipe = mc->pipe;
 508         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 509
 510         /* Version */
 511         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 512         /* Header */
 513         header = (struct tgsi_header*)&tokens[1];
 514         *header = tgsi_build_header();
 515         /* Processor */
 516         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_FRAGMENT, header);
 517
 518         ti = 3;
 519
 520         /*
 521          * decl i0                      ; Luma texcoords
 522          * decl i1                      ; Chroma Cb texcoords
 523          * decl i2                      ; Chroma Cr texcoords
 524          * decl i3                      ; Ref macroblock top field texcoords
 525          * decl i4                      ; Ref macroblock bottom field texcoords
 526          * decl i5                      ; Denormalized vertex pos
 527          */
 528         for (i = 0; i < 6; ++i)
 529         {
 530                 decl = vl_decl_interpolated_input(TGSI_SEMANTIC_GENERIC, i + 1, i, i, TGSI_INTERPOLATE_LINEAR);
 531                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 532         }
 533
 534         /*
 535          * decl c0                      ; Scaling factor, rescales 16-bit snorm to 9-bit snorm
 536          * decl c1                      ; Constants 1/2 & 2 in .x, .y channels for Y-mod-2 top/bottom field selection
 537          */
 538         decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 1);
 539         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 540
 541         /* decl o0                      ; Fragment color */
 542         decl = vl_decl_output(TGSI_SEMANTIC_COLOR, 0, 0, 0);
 543         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 544
 545         /* decl t0-t4 */
 546         decl = vl_decl_temps(0, 4);
 547         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 548
 549         /*
 550          * decl s0                      ; Sampler for luma texture
 551          * decl s1                      ; Sampler for chroma Cb texture
 552          * decl s2                      ; Sampler for chroma Cr texture
 553          * decl s3                      ; Sampler for ref surface texture
 554          */
 555         for (i = 0; i < 4; ++i)
 556         {
 557                 decl = vl_decl_samplers(i, i);
 558                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 559         }
 560
 561         /*
 562          * tex2d t1, i0, s0             ; Read texel from luma texture
 563          * mov t0.x, t1.x               ; Move luma sample into .x component
 564          * tex2d t1, i1, s1             ; Read texel from chroma Cb texture
 565          * mov t0.y, t1.x               ; Move Cb sample into .y component
 566          * tex2d t1, i2, s2             ; Read texel from chroma Cr texture
 567          * mov t0.z, t1.x               ; Move Cr sample into .z component
 568          */
 569         for (i = 0; i < 3; ++i)
 570         {
 571                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_INPUT, i, TGSI_FILE_SAMPLER, i);
 572                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 573
 574                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
 575                 inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 576                 inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 577                 inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
 578                 inst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X << i;
 579                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 580         }
 581
 582         /* mul t0, t0, c0               ; Rescale texel to correct range */
 583         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_CONSTANT, 0);
 584         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 585
 586         /*
 587          * tex2d t1, i3, s3             ; Read texel from ref macroblock top field
 588          * tex2d t2, i4, s3             ; Read texel from ref macroblock bottom field
 589          */
 590         for (i = 0; i < 2; ++i)
 591         {
 592                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, i + 1, TGSI_FILE_INPUT, i + 3, TGSI_FILE_SAMPLER, 3);
 593                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 594         }
 595
 596         /* XXX: Pos values off by 0.5? */
 597         /* sub t4, i5.y, c1.x           ; Sub 0.5 from denormalized pos */
 598         inst = vl_inst3(TGSI_OPCODE_SUB, TGSI_FILE_TEMPORARY, 4, TGSI_FILE_INPUT, 5, TGSI_FILE_CONSTANT, 1);
 599         inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y;
 600         inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
 601         inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Y;
 602         inst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y;
 603         inst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 604         inst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 605         inst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
 606         inst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_X;
 607         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 608
 609         /* mul t3, t4, c1.x             ; Multiply pos Y-coord by 1/2 */
 610         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 4, TGSI_FILE_CONSTANT, 1);
 611         inst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 612         inst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 613         inst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
 614         inst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_X;
 615         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 616
 617         /* floor t3, t3                 ; Get rid of fractional part */
 618         inst = vl_inst2(TGSI_OPCODE_FLR, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 3);
 619         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 620
 621         /* mul t3, t3, c1.y             ; Multiply by 2 */
 622         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_CONSTANT, 1);
 623         inst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y;
 624         inst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
 625         inst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Y;
 626         inst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y;
 627         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 628
 629         /* sub t3, t4, t3               ; Subtract from original Y to get Y % 2 */
 630         inst = vl_inst3(TGSI_OPCODE_SUB, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 4, TGSI_FILE_TEMPORARY, 3);
 631         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 632
 633         /* TODO: Move to conditional tex fetch on t3 instead of lerp */
 634         /* lerp t1, t3, t1, t2          ; Choose between top and bottom fields based on Y % 2 */
 635         inst = vl_inst4(TGSI_OPCODE_LRP, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_TEMPORARY, 2);
 636         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 637
 638         /* add o0, t0, t1               ; Add ref and differential to form final output */
 639         inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
 640         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 641
 642         /* end */
 643         inst = vl_end();
 644         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 645
 646         fs.tokens = tokens;
 647         mc->p_fs[1] = pipe->create_fs_state(pipe, &fs);
 648         free(tokens);
 649
 650         return 0;
 651 }
 652
 653 static int vlCreateVertexShaderFrameBMB
 654 (
 655         struct vlR16SnormBufferedMC *mc
 656 )
 657 {
 658         const unsigned int              max_tokens = 100;
 659
 660         struct pipe_context             *pipe;
 661         struct pipe_shader_state        vs;
 662         struct tgsi_token               *tokens;
 663         struct tgsi_header              *header;
 664
 665         struct tgsi_full_declaration    decl;
 666         struct tgsi_full_instruction    inst;
 667
 668         unsigned int                    ti;
 669         unsigned int                    i;
 670
 671         assert(mc);
 672
 673         pipe = mc->pipe;
 674         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 675
 676         /* Version */
 677         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 678         /* Header */
 679         header = (struct tgsi_header*)&tokens[1];
 680         *header = tgsi_build_header();
 681         /* Processor */
 682         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_VERTEX, header);
 683
 684         ti = 3;
 685
 686         /*
 687          * decl i0              ; Vertex pos
 688          * decl i1              ; Luma texcoords
 689          * decl i2              ; Chroma Cb texcoords
 690          * decl i3              ; Chroma Cr texcoords
 691          * decl i4              ; First ref macroblock top field texcoords
 692          * decl i5              ; First ref macroblock bottom field texcoords (unused, packed in the same stream)
 693          * decl i6              ; Second ref macroblock top field texcoords
 694          * decl i7              ; Second ref macroblock bottom field texcoords (unused, packed in the same stream)
 695          */
 696         for (i = 0; i < 8; i++)
 697         {
 698                 decl = vl_decl_input(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 699                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 700         }
 701
 702         /*
 703          * decl o0              ; Vertex pos
 704          * decl o1              ; Luma texcoords
 705          * decl o2              ; Chroma Cb texcoords
 706          * decl o3              ; Chroma Cr texcoords
 707          * decl o4              ; First ref macroblock texcoords
 708          * decl o5              ; Second ref macroblock texcoords
 709          */
 710         for (i = 0; i < 6; i++)
 711         {
 712                 decl = vl_decl_output(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 713                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 714         }
 715
 716         /*
 717          * mov o0, i0           ; Move input vertex pos to output
 718          * mov o1, i1           ; Move input luma texcoords to output
 719          * mov o2, i2           ; Move input chroma Cb texcoords to output
 720          * mov o3, i3           ; Move input chroma Cr texcoords to output
 721          */
 722         for (i = 0; i < 4; ++i)
 723         {
 724                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_OUTPUT, i, TGSI_FILE_INPUT, i);
 725                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 726         }
 727
 728         /*
 729          * add o4, i0, i4       ; Translate vertex pos by motion vec to form first ref macroblock texcoords
 730          * add o5, i0, i6       ; Translate vertex pos by motion vec to form second ref macroblock texcoords
 731          */
 732         for (i = 0; i < 2; ++i)
 733         {
 734                 inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, i + 4, TGSI_FILE_INPUT, 0, TGSI_FILE_INPUT, (i + 2) * 2);
 735                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 736         }
 737
 738         /* end */
 739         inst = vl_end();
 740         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 741
 742         vs.tokens = tokens;
 743         mc->b_vs[0] = pipe->create_vs_state(pipe, &vs);
 744         free(tokens);
 745
 746         return 0;
 747 }
 748
 749 static int vlCreateVertexShaderFieldBMB
 750 (
 751         struct vlR16SnormBufferedMC *mc
 752 )
 753 {
 754         const unsigned int              max_tokens = 100;
 755
 756         struct pipe_context             *pipe;
 757         struct pipe_shader_state        vs;
 758         struct tgsi_token               *tokens;
 759         struct tgsi_header              *header;
 760
 761         struct tgsi_full_declaration    decl;
 762         struct tgsi_full_instruction    inst;
 763
 764         unsigned int                    ti;
 765         unsigned int                    i;
 766
 767         assert(mc);
 768
 769         pipe = mc->pipe;
 770         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 771
 772         /* Version */
 773         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 774         /* Header */
 775         header = (struct tgsi_header*)&tokens[1];
 776         *header = tgsi_build_header();
 777         /* Processor */
 778         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_VERTEX, header);
 779
 780         ti = 3;
 781
 782         /*
 783          * decl i0              ; Vertex pos
 784          * decl i1              ; Luma texcoords
 785          * decl i2              ; Chroma Cb texcoords
 786          * decl i3              ; Chroma Cr texcoords
 787          * decl i4              ; First ref macroblock top field texcoords
 788          * decl i5              ; First ref macroblock bottom field texcoords
 789          * decl i6              ; Second ref macroblock top field texcoords
 790          * decl i7              ; Second ref macroblock bottom field texcoords
 791          */
 792         for (i = 0; i < 8; i++)
 793         {
 794                 decl = vl_decl_input(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 795                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 796         }
 797
 798         /* decl c0              ; Render target dimensions */
 799         decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 0);
 800         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 801
 802         /*
 803          * decl o0              ; Vertex pos
 804          * decl o1              ; Luma texcoords
 805          * decl o2              ; Chroma Cb texcoords
 806          * decl o3              ; Chroma Cr texcoords
 807          * decl o4              ; First ref macroblock top field texcoords
 808          * decl o5              ; First ref macroblock Bottom field texcoords
 809          * decl o6              ; Second ref macroblock top field texcoords
 810          * decl o7              ; Second ref macroblock Bottom field texcoords
 811          * decl o8              ; Denormalized vertex pos
 812          */
 813         for (i = 0; i < 9; i++)
 814         {
 815                 decl = vl_decl_output(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 816                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 817         }
 818
 819         /* decl t0, t1 */
 820         decl = vl_decl_temps(0, 1);
 821         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 822
 823         /*
 824          * mov o0, i0           ; Move input vertex pos to output
 825          * mov o1, i1           ; Move input luma texcoords to output
 826          * mov o2, i2           ; Move input chroma Cb texcoords to output
 827          * mov o3, i3           ; Move input chroma Cr texcoords to output
 828          */
 829         for (i = 0; i < 4; ++i)
 830         {
 831                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_OUTPUT, i, TGSI_FILE_INPUT, i);
 832                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 833         }
 834
 835         /*
 836          * add o4, i0, i4       ; Translate vertex pos by motion vec to form first top field macroblock texcoords
 837          * add o5, i0, i5       ; Translate vertex pos by motion vec to form first bottom field macroblock texcoords
 838          * add o6, i0, i6       ; Translate vertex pos by motion vec to form second top field macroblock texcoords
 839          * add o7, i0, i7       ; Translate vertex pos by motion vec to form second bottom field macroblock texcoords
 840          */
 841         for (i = 0; i < 4; ++i)
 842         {
 843                 inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, i + 4, TGSI_FILE_INPUT, 0, TGSI_FILE_INPUT, i + 4);
 844                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 845         }
 846
 847         /* mul o8, i0, c0       ; Denorm vertex pos */
 848         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_OUTPUT, 8, TGSI_FILE_INPUT, 0, TGSI_FILE_CONSTANT, 0);
 849         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 850
 851         /* end */
 852         inst = vl_end();
 853         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 854
 855         vs.tokens = tokens;
 856         mc->b_vs[1] = pipe->create_vs_state(pipe, &vs);
 857         free(tokens);
 858
 859         return 0;
 860 }
 861
 862 static int vlCreateFragmentShaderFrameBMB
 863 (
 864         struct vlR16SnormBufferedMC *mc
 865 )
 866 {
 867         const unsigned int              max_tokens = 100;
 868
 869         struct pipe_context             *pipe;
 870         struct pipe_shader_state        fs;
 871         struct tgsi_token               *tokens;
 872         struct tgsi_header              *header;
 873
 874         struct tgsi_full_declaration    decl;
 875         struct tgsi_full_instruction    inst;
 876
 877         unsigned int                    ti;
 878         unsigned int                    i;
 879
 880         assert(mc);
 881
 882         pipe = mc->pipe;
 883         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
 884
 885         /* Version */
 886         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
 887         /* Header */
 888         header = (struct tgsi_header*)&tokens[1];
 889         *header = tgsi_build_header();
 890         /* Processor */
 891         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_FRAGMENT, header);
 892
 893         ti = 3;
 894
 895         /*
 896          * decl i0                      ; Luma texcoords
 897          * decl i1                      ; Chroma Cb texcoords
 898          * decl i2                      ; Chroma Cr texcoords
 899          * decl i3                      ; First ref macroblock texcoords
 900          * decl i4                      ; Second ref macroblock texcoords
 901          */
 902         for (i = 0; i < 5; ++i)
 903         {
 904                 decl = vl_decl_interpolated_input(TGSI_SEMANTIC_GENERIC, i + 1, i, i, TGSI_INTERPOLATE_LINEAR);
 905                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 906         }
 907
 908         /*
 909          * decl c0                      ; Scaling factor, rescales 16-bit snorm to 9-bit snorm
 910          * decl c1                      ; Constant 1/2 in .x channel to use as weight to blend past and future texels
 911          */
 912         decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 1);
 913         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 914
 915         /* decl o0                      ; Fragment color */
 916         decl = vl_decl_output(TGSI_SEMANTIC_COLOR, 0, 0, 0);
 917         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 918
 919         /* decl t0-t2 */
 920         decl = vl_decl_temps(0, 2);
 921         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 922
 923         /*
 924          * decl s0                      ; Sampler for luma texture
 925          * decl s1                      ; Sampler for chroma Cb texture
 926          * decl s2                      ; Sampler for chroma Cr texture
 927          * decl s3                      ; Sampler for first ref surface texture
 928          * decl s4                      ; Sampler for second ref surface texture
 929          */
 930         for (i = 0; i < 5; ++i)
 931         {
 932                 decl = vl_decl_samplers(i, i);
 933                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 934         }
 935
 936         /*
 937          * tex2d t1, i0, s0             ; Read texel from luma texture
 938          * mov t0.x, t1.x               ; Move luma sample into .x component
 939          * tex2d t1, i1, s1             ; Read texel from chroma Cb texture
 940          * mov t0.y, t1.x               ; Move Cb sample into .y component
 941          * tex2d t1, i2, s2             ; Read texel from chroma Cr texture
 942          * mov t0.z, t1.x               ; Move Cr sample into .z component
 943          */
 944         for (i = 0; i < 3; ++i)
 945         {
 946                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_INPUT, i, TGSI_FILE_SAMPLER, i);
 947                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 948
 949                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
 950                 inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 951                 inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 952                 inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
 953                 inst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X << i;
 954                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 955         }
 956
 957         /* mul t0, t0, c0               ; Rescale texel to correct range */
 958         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_CONSTANT, 0);
 959         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 960
 961         /*
 962          * tex2d t1, i3, s3             ; Read texel from first ref macroblock
 963          * tex2d t2, i4, s4             ; Read texel from second ref macroblock
 964          */
 965         for (i = 0; i < 2; ++i)
 966         {
 967                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, i + 1, TGSI_FILE_INPUT, i + 3, TGSI_FILE_SAMPLER, i + 3);
 968                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 969         }
 970
 971         /* lerp t1, c1.x, t1, t2        ; Blend past and future texels */
 972         inst = vl_inst4(TGSI_OPCODE_LRP, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_CONSTANT, 1, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_TEMPORARY, 2);
 973         inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 974         inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 975         inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
 976         inst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_X;
 977         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 978
 979         /* add o0, t0, t1               ; Add past/future ref and differential to form final output */
 980         inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
 981         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 982
 983         /* end */
 984         inst = vl_end();
 985         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 986
 987         fs.tokens = tokens;
 988         mc->b_fs[0] = pipe->create_fs_state(pipe, &fs);
 989         free(tokens);
 990
 991         return 0;
 992 }
 993
 994 static int vlCreateFragmentShaderFieldBMB
 995 (
 996         struct vlR16SnormBufferedMC *mc
 997 )
 998 {
 999         const unsigned int              max_tokens = 200;
1000
1001         struct pipe_context             *pipe;
1002         struct pipe_shader_state        fs;
1003         struct tgsi_token               *tokens;
1004         struct tgsi_header              *header;
1005
1006         struct tgsi_full_declaration    decl;
1007         struct tgsi_full_instruction    inst;
1008
1009         unsigned int                    ti;
1010         unsigned int                    i;
1011
1012         assert(mc);
1013
1014         pipe = mc->pipe;
1015         tokens = (struct tgsi_token*)malloc(max_tokens * sizeof(struct tgsi_token));
1016
1017         /* Version */
1018         *(struct tgsi_version*)&tokens[0] = tgsi_build_version();
1019         /* Header */
1020         header = (struct tgsi_header*)&tokens[1];
1021         *header = tgsi_build_header();
1022         /* Processor */
1023         *(struct tgsi_processor*)&tokens[2] = tgsi_build_processor(TGSI_PROCESSOR_FRAGMENT, header);
1024
1025         ti = 3;
1026
1027         /*
1028          * decl i0                      ; Luma texcoords
1029          * decl i1                      ; Chroma Cb texcoords
1030          * decl i2                      ; Chroma Cr texcoords
1031          * decl i3                      ; First ref macroblock top field texcoords
1032          * decl i4                      ; First ref macroblock bottom field texcoords
1033          * decl i5                      ; Second ref macroblock top field texcoords
1034          * decl i6                      ; Second ref macroblock bottom field texcoords
1035          * decl i7                      ; Denormalized vertex pos
1036          */
1037         for (i = 0; i < 8; ++i)
1038         {
1039                 decl = vl_decl_interpolated_input(TGSI_SEMANTIC_GENERIC, i + 1, i, i, TGSI_INTERPOLATE_LINEAR);
1040                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
1041         }
1042
1043         /*
1044          * decl c0                      ; Scaling factor, rescales 16-bit snorm to 9-bit snorm
1045          * decl c1                      ; Constants 1/2 & 2 in .x, .y channels to use as weight to blend past and future texels
1046          *                              ; and for Y-mod-2 top/bottom field selection
1047          */
1048         decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 1);
1049         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
1050
1051         /* decl o0                      ; Fragment color */
1052         decl = vl_decl_output(TGSI_SEMANTIC_COLOR, 0, 0, 0);
1053         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
1054
1055         /* decl t0-t5 */
1056         decl = vl_decl_temps(0, 5);
1057         ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
1058
1059         /*
1060          * decl s0                      ; Sampler for luma texture
1061          * decl s1                      ; Sampler for chroma Cb texture
1062          * decl s2                      ; Sampler for chroma Cr texture
1063          * decl s3                      ; Sampler for first ref surface texture
1064          * decl s4                      ; Sampler for second ref surface texture
1065          */
1066         for (i = 0; i < 5; ++i)
1067         {
1068                 decl = vl_decl_samplers(i, i);
1069                 ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
1070         }
1071
1072         /*
1073          * tex2d t1, i0, s0             ; Read texel from luma texture
1074          * mov t0.x, t1.x               ; Move luma sample into .x component
1075          * tex2d t1, i1, s1             ; Read texel from chroma Cb texture
1076          * mov t0.y, t1.x               ; Move Cb sample into .y component
1077          * tex2d t1, i2, s2             ; Read texel from chroma Cr texture
1078          * mov t0.z, t1.x               ; Move Cr sample into .z component
1079          */
1080         for (i = 0; i < 3; ++i)
1081         {
1082                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_INPUT, i, TGSI_FILE_SAMPLER, i);
1083                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1084
1085                 inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
1086                 inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
1087                 inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
1088                 inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
1089                 inst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X << i;
1090                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1091         }
1092
1093         /* mul t0, t0, c0               ; Rescale texel to correct range */
1094         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_CONSTANT, 0);
1095         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1096
1097         /* XXX: Pos values off by 0.5? */
1098         /* sub t4, i7.y, c1.x           ; Sub 0.5 from denormalized pos */
1099         inst = vl_inst3(TGSI_OPCODE_SUB, TGSI_FILE_TEMPORARY, 4, TGSI_FILE_INPUT, 7, TGSI_FILE_CONSTANT, 1);
1100         inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y;
1101         inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
1102         inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Y;
1103         inst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y;
1104         inst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
1105         inst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
1106         inst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
1107         inst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_X;
1108         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1109
1110         /* mul t3, t4, c1.x             ; Multiply pos Y-coord by 1/2 */
1111         inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 4, TGSI_FILE_CONSTANT, 1);
1112         inst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
1113         inst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
1114         inst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
1115         inst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_X;
1116         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1117
1118         /* floor t3, t3                 ; Get rid of fractional part */
1119         inst = vl_inst2(TGSI_OPCODE_FLR, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 3);
1120         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1121
1122         /* mul t3, t3, c1.y             ; Multiply by 2 */
1123         inst = vl_inst3( TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_CONSTANT, 1);
1124         inst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y;
1125         inst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
1126         inst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Y;
1127         inst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y;
1128         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1129
1130         /* sub t3, t4, t3               ; Subtract from original Y to get Y % 2 */
1131         inst = vl_inst3(TGSI_OPCODE_SUB, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 4, TGSI_FILE_TEMPORARY, 3);
1132         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1133
1134         /*
1135          * tex2d t1, i3, s3             ; Read texel from past ref macroblock top field
1136          * tex2d t2, i4, s3             ; Read texel from past ref macroblock bottom field
1137          */
1138         for (i = 0; i < 2; ++i)
1139         {
1140                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, i + 1, TGSI_FILE_INPUT, i + 3, TGSI_FILE_SAMPLER, 3);
1141                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1142         }
1143
1144         /* TODO: Move to conditional tex fetch on t3 instead of lerp */
1145         /* lerp t1, t3, t1, t2          ; Choose between top and bottom fields based on Y % 2 */
1146         inst = vl_inst4(TGSI_OPCODE_LRP, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_TEMPORARY, 2);
1147         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1148
1149         /*
1150          * tex2d t4, i5, s4             ; Read texel from future ref macroblock top field
1151          * tex2d t5, i6, s4             ; Read texel from future ref macroblock bottom field
1152          */
1153         for (i = 0; i < 2; ++i)
1154         {
1155                 inst = vl_tex(TGSI_TEXTURE_2D, TGSI_FILE_TEMPORARY, i + 4, TGSI_FILE_INPUT, i + 5, TGSI_FILE_SAMPLER, 4);
1156                 ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1157         }
1158
1159         /* TODO: Move to conditional tex fetch on t3 instead of lerp */
1160         /* lerp t2, t3, t4, t5          ; Choose between top and bottom fields based on Y % 2 */
1161         inst = vl_inst4(TGSI_OPCODE_LRP, TGSI_FILE_TEMPORARY, 2, TGSI_FILE_TEMPORARY, 3, TGSI_FILE_TEMPORARY, 4, TGSI_FILE_TEMPORARY, 5);
1162         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1163
1164         /* lerp t1, c1.x, t1, t2        ; Blend past and future texels */
1165         inst = vl_inst4(TGSI_OPCODE_LRP, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_CONSTANT, 1, TGSI_FILE_TEMPORARY, 1, TGSI_FILE_TEMPORARY, 2);
1166         inst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
1167         inst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
1168         inst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_X;
1169         inst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_X;
1170         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1171
1172         /* add o0, t0, t1               ; Add past/future ref and differential to form final output */
1173         inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, 0, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_TEMPORARY, 1);
1174         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1175
1176         /* end */
1177         inst = vl_end();
1178         ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
1179
1180         fs.tokens = tokens;
1181         mc->b_fs[1] = pipe->create_fs_state(pipe, &fs);
1182         free(tokens);
1183
1184         return 0;
1185 }