src/mesa/main/texenvprogram.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "glheader.h"
  29 #include "macros.h"
  30 #include "enums.h"
  31 #include "texenvprogram.h"
  32
  33 #include "shader/program.h"
  34 #include "shader/program_instruction.h"
  35 #include "shader/arbfragparse.h"
  36
  37
  38 #define DISASSEM (MESA_VERBOSE & VERBOSE_DISASSEM)
  39
  40 struct mode_opt {
  41    unsigned Source:4;
  42    unsigned Operand:3;
  43 };
  44
  45 struct state_key {
  46    GLuint enabled_units;
  47    unsigned separate_specular:1;
  48    unsigned fog_enabled:1;
  49    unsigned fog_mode:2;
  50
  51    struct {
  52       unsigned enabled:1;
  53       unsigned source_index:3;
  54       unsigned ScaleShiftRGB:2;
  55       unsigned ScaleShiftA:2;
  56
  57       unsigned NumArgsRGB:2;
  58       unsigned ModeRGB:4;
  59       struct mode_opt OptRGB[3];
  60
  61       unsigned NumArgsA:2;
  62       unsigned ModeA:4;
  63       struct mode_opt OptA[3];
  64    } unit[8];
  65 };
  66
  67 #define FOG_LINEAR  0
  68 #define FOG_EXP     1
  69 #define FOG_EXP2    2
  70 #define FOG_UNKNOWN 3
  71
  72 static GLuint translate_fog_mode( GLenum mode )
  73 {
  74    switch (mode) {
  75    case GL_LINEAR: return FOG_LINEAR;
  76    case GL_EXP: return FOG_EXP;
  77    case GL_EXP2: return FOG_EXP2;
  78    default: return FOG_UNKNOWN;
  79    }
  80 }
  81
  82 #define OPR_SRC_COLOR           0
  83 #define OPR_ONE_MINUS_SRC_COLOR 1
  84 #define OPR_SRC_ALPHA           2
  85 #define OPR_ONE_MINUS_SRC_ALPHA 3
  86 #define OPR_ZERO                4
  87 #define OPR_ONE                 5
  88 #define OPR_UNKNOWN             7
  89
  90 static GLuint translate_operand( GLenum operand )
  91 {
  92    switch (operand) {
  93    case GL_SRC_COLOR: return OPR_SRC_COLOR;
  94    case GL_ONE_MINUS_SRC_COLOR: return OPR_ONE_MINUS_SRC_COLOR;
  95    case GL_SRC_ALPHA: return OPR_SRC_ALPHA;
  96    case GL_ONE_MINUS_SRC_ALPHA: return OPR_ONE_MINUS_SRC_ALPHA;
  97    case GL_ZERO: return OPR_ZERO;
  98    case GL_ONE: return OPR_ONE;
  99    default:     return OPR_UNKNOWN;
 100    }
 101 }
 102
 103 #define SRC_TEXTURE  0
 104 #define SRC_TEXTURE0 1
 105 #define SRC_TEXTURE1 2
 106 #define SRC_TEXTURE2 3
 107 #define SRC_TEXTURE3 4
 108 #define SRC_TEXTURE4 5
 109 #define SRC_TEXTURE5 6
 110 #define SRC_TEXTURE6 7
 111 #define SRC_TEXTURE7 8
 112 #define SRC_CONSTANT 9
 113 #define SRC_PRIMARY_COLOR 10
 114 #define SRC_PREVIOUS 11
 115 #define SRC_UNKNOWN  15
 116
 117 static GLuint translate_source( GLenum src )
 118 {
 119    switch (src) {
 120    case GL_TEXTURE: return SRC_TEXTURE;
 121    case GL_TEXTURE0:
 122    case GL_TEXTURE1:
 123    case GL_TEXTURE2:
 124    case GL_TEXTURE3:
 125    case GL_TEXTURE4:
 126    case GL_TEXTURE5:
 127    case GL_TEXTURE6:
 128    case GL_TEXTURE7: return SRC_TEXTURE0 + (src - GL_TEXTURE0);
 129    case GL_CONSTANT: return SRC_CONSTANT;
 130    case GL_PRIMARY_COLOR: return SRC_PRIMARY_COLOR;
 131    case GL_PREVIOUS: return SRC_PREVIOUS;
 132    default: return SRC_UNKNOWN;
 133    }
 134 }
 135
 136 #define MODE_REPLACE       0
 137 #define MODE_MODULATE      1
 138 #define MODE_ADD           2
 139 #define MODE_ADD_SIGNED    3
 140 #define MODE_INTERPOLATE   4
 141 #define MODE_SUBTRACT      5
 142 #define MODE_DOT3_RGB      6
 143 #define MODE_DOT3_RGB_EXT  7
 144 #define MODE_DOT3_RGBA     8
 145 #define MODE_DOT3_RGBA_EXT 9
 146 #define MODE_MODULATE_ADD_ATI           10
 147 #define MODE_MODULATE_SIGNED_ADD_ATI    11
 148 #define MODE_MODULATE_SUBTRACT_ATI      12
 149 #define MODE_UNKNOWN       15
 150
 151 static GLuint translate_mode( GLenum mode )
 152 {
 153    switch (mode) {
 154    case GL_REPLACE: return MODE_REPLACE;
 155    case GL_MODULATE: return MODE_MODULATE;
 156    case GL_ADD: return MODE_ADD;
 157    case GL_ADD_SIGNED: return MODE_ADD_SIGNED;
 158    case GL_INTERPOLATE: return MODE_INTERPOLATE;
 159    case GL_SUBTRACT: return MODE_SUBTRACT;
 160    case GL_DOT3_RGB: return MODE_DOT3_RGB;
 161    case GL_DOT3_RGB_EXT: return MODE_DOT3_RGB_EXT;
 162    case GL_DOT3_RGBA: return MODE_DOT3_RGBA;
 163    case GL_DOT3_RGBA_EXT: return MODE_DOT3_RGBA_EXT;
 164    case GL_MODULATE_ADD_ATI: return MODE_MODULATE_ADD_ATI;
 165    case GL_MODULATE_SIGNED_ADD_ATI: return MODE_MODULATE_SIGNED_ADD_ATI;
 166    case GL_MODULATE_SUBTRACT_ATI: return MODE_MODULATE_SUBTRACT_ATI;
 167    default: return MODE_UNKNOWN;
 168    }
 169 }
 170
 171 #define TEXTURE_UNKNOWN_INDEX 7
 172 static GLuint translate_tex_src_bit( GLbitfield bit )
 173 {
 174    switch (bit) {
 175    case TEXTURE_1D_BIT:   return TEXTURE_1D_INDEX;
 176    case TEXTURE_2D_BIT:   return TEXTURE_2D_INDEX;
 177    case TEXTURE_RECT_BIT: return TEXTURE_RECT_INDEX;
 178    case TEXTURE_3D_BIT:   return TEXTURE_3D_INDEX;
 179    case TEXTURE_CUBE_BIT: return TEXTURE_CUBE_INDEX;
 180    default: return TEXTURE_UNKNOWN_INDEX;
 181    }
 182 }
 183
 184 static struct state_key *make_state_key( GLcontext *ctx )
 185 {
 186    struct state_key *key = CALLOC_STRUCT(state_key);
 187    GLuint i, j;
 188
 189    for (i=0;i<MAX_TEXTURE_UNITS;i++) {
 190       struct gl_texture_unit *texUnit = &ctx->Texture.Unit[i];
 191
 192       if (!texUnit->_ReallyEnabled)
 193          continue;
 194
 195       key->unit[i].enabled = 1;
 196       key->enabled_units |= (1<<i);
 197
 198       key->unit[i].source_index =
 199          translate_tex_src_bit(texUnit->_ReallyEnabled);
 200
 201       key->unit[i].NumArgsRGB = texUnit->_CurrentCombine->_NumArgsRGB;
 202       key->unit[i].NumArgsA = texUnit->_CurrentCombine->_NumArgsA;
 203
 204       key->unit[i].ModeRGB =
 205          translate_mode(texUnit->_CurrentCombine->ModeRGB);
 206       key->unit[i].ModeA =
 207          translate_mode(texUnit->_CurrentCombine->ModeA);
 208
 209       key->unit[i].ScaleShiftRGB = texUnit->_CurrentCombine->ScaleShiftRGB;
 210       key->unit[i].ScaleShiftA = texUnit->_CurrentCombine->ScaleShiftRGB;
 211
 212       for (j=0;j<3;j++) {
 213          key->unit[i].OptRGB[j].Operand =
 214             translate_operand(texUnit->_CurrentCombine->OperandRGB[j]);
 215          key->unit[i].OptA[j].Operand =
 216             translate_operand(texUnit->_CurrentCombine->OperandA[j]);
 217          key->unit[i].OptRGB[j].Source =
 218             translate_source(texUnit->_CurrentCombine->SourceRGB[j]);
 219          key->unit[i].OptA[j].Source =
 220             translate_source(texUnit->_CurrentCombine->SourceA[j]);
 221       }
 222    }
 223
 224    if (ctx->_TriangleCaps & DD_SEPARATE_SPECULAR)
 225       key->separate_specular = 1;
 226
 227    if (ctx->Fog.Enabled) {
 228       key->fog_enabled = 1;
 229       key->fog_mode = translate_fog_mode(ctx->Fog.Mode);
 230    }
 231
 232    return key;
 233 }
 234
 235 /* Use uregs to represent registers internally, translate to Mesa's
 236  * expected formats on emit.
 237  *
 238  * NOTE: These are passed by value extensively in this file rather
 239  * than as usual by pointer reference.  If this disturbs you, try
 240  * remembering they are just 32bits in size.
 241  *
 242  * GCC is smart enough to deal with these dword-sized structures in
 243  * much the same way as if I had defined them as dwords and was using
 244  * macros to access and set the fields.  This is much nicer and easier
 245  * to evolve.
 246  */
 247 struct ureg {
 248    GLuint file:4;
 249    GLuint idx:8;
 250    GLuint negatebase:1;
 251    GLuint abs:1;
 252    GLuint negateabs:1;
 253    GLuint swz:12;
 254    GLuint pad:5;
 255 };
 256
 257 const static struct ureg undef = {
 258    ~0,
 259    ~0,
 260    0,
 261    0,
 262    0,
 263    0,
 264    0
 265 };
 266
 267 #define X    0
 268 #define Y    1
 269 #define Z    2
 270 #define W    3
 271
 272 /* State used to build the fragment program:
 273  */
 274 struct texenv_fragment_program {
 275    struct fragment_program *program;
 276    GLcontext *ctx;
 277    struct state_key *state;
 278
 279    GLuint alu_temps;            /* Track texture indirections, see spec. */
 280    GLuint temps_output;         /* Track texture indirections, see spec. */
 281
 282    GLuint temp_in_use;          /* Tracks temporary regs which are in
 283                                  * use.
 284                                  */
 285
 286
 287    GLboolean error;
 288
 289    struct ureg src_texture[MAX_TEXTURE_UNITS];
 290    /* Reg containing each texture unit's sampled texture color,
 291     * else undef.
 292     */
 293
 294    struct ureg src_previous;    /* Reg containing color from previous
 295                                  * stage.  May need to be decl'd.
 296                                  */
 297
 298    GLuint last_tex_stage;       /* Number of last enabled texture unit */
 299
 300    struct ureg half;
 301    struct ureg one;
 302    struct ureg zero;
 303 };
 304
 305
 306
 307 static struct ureg make_ureg(GLuint file, GLuint idx)
 308 {
 309    struct ureg reg;
 310    reg.file = file;
 311    reg.idx = idx;
 312    reg.negatebase = 0;
 313    reg.abs = 0;
 314    reg.negateabs = 0;
 315    reg.swz = SWIZZLE_NOOP;
 316    reg.pad = 0;
 317    return reg;
 318 }
 319
 320 static struct ureg swizzle( struct ureg reg, int x, int y, int z, int w )
 321 {
 322    reg.swz = MAKE_SWIZZLE4(GET_SWZ(reg.swz, x),
 323                            GET_SWZ(reg.swz, y),
 324                            GET_SWZ(reg.swz, z),
 325                            GET_SWZ(reg.swz, w));
 326
 327    return reg;
 328 }
 329
 330 static struct ureg swizzle1( struct ureg reg, int x )
 331 {
 332    return swizzle(reg, x, x, x, x);
 333 }
 334
 335 static struct ureg negate( struct ureg reg )
 336 {
 337    reg.negatebase ^= 1;
 338    return reg;
 339 }
 340
 341 static GLboolean is_undef( struct ureg reg )
 342 {
 343    return reg.file == 0xf;
 344 }
 345
 346
 347 static struct ureg get_temp( struct texenv_fragment_program *p )
 348 {
 349    int bit;
 350
 351    /* First try and reuse temps which have been used already:
 352     */
 353    bit = _mesa_ffs( ~p->temp_in_use & p->alu_temps );
 354
 355    /* Then any unused temporary:
 356     */
 357    if (!bit)
 358       bit = _mesa_ffs( ~p->temp_in_use );
 359
 360    if (!bit) {
 361       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 362       _mesa_exit(1);
 363    }
 364
 365    if (bit > p->program->Base.NumTemporaries)
 366       p->program->Base.NumTemporaries = bit;
 367
 368    p->temp_in_use |= 1<<(bit-1);
 369    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 370 }
 371
 372 static struct ureg get_tex_temp( struct texenv_fragment_program *p )
 373 {
 374    int bit;
 375
 376    /* First try to find availble temp not previously used (to avoid
 377     * starting a new texture indirection).  According to the spec, the
 378     * ~p->temps_output isn't necessary, but will keep it there for
 379     * now:
 380     */
 381    bit = _mesa_ffs( ~p->temp_in_use & ~p->alu_temps & ~p->temps_output );
 382
 383    /* Then any unused temporary:
 384     */
 385    if (!bit)
 386       bit = _mesa_ffs( ~p->temp_in_use );
 387
 388    if (!bit) {
 389       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 390       _mesa_exit(1);
 391    }
 392
 393    if (bit > p->program->Base.NumTemporaries)
 394       p->program->Base.NumTemporaries = bit;
 395
 396    p->temp_in_use |= 1<<(bit-1);
 397    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 398 }
 399
 400
 401 static void release_temps( struct texenv_fragment_program *p )
 402 {
 403    GLuint max_temp = p->ctx->Const.FragmentProgram.MaxTemps;
 404
 405    /* KW: To support tex_env_crossbar, don't release the registers in
 406     * temps_output.
 407     */
 408    if (max_temp >= sizeof(int) * 8)
 409       p->temp_in_use = p->temps_output;
 410    else
 411       p->temp_in_use = ~((1<<max_temp)-1) | p->temps_output;
 412 }
 413
 414
 415 static struct ureg register_param6( struct texenv_fragment_program *p,
 416                                     GLint s0,
 417                                     GLint s1,
 418                                     GLint s2,
 419                                     GLint s3,
 420                                     GLint s4,
 421                                     GLint s5)
 422 {
 423    GLint tokens[6];
 424    GLuint idx;
 425    tokens[0] = s0;
 426    tokens[1] = s1;
 427    tokens[2] = s2;
 428    tokens[3] = s3;
 429    tokens[4] = s4;
 430    tokens[5] = s5;
 431    idx = _mesa_add_state_reference( p->program->Base.Parameters, tokens );
 432    return make_ureg(PROGRAM_STATE_VAR, idx);
 433 }
 434
 435
 436 #define register_param1(p,s0)          register_param6(p,s0,0,0,0,0,0)
 437 #define register_param2(p,s0,s1)       register_param6(p,s0,s1,0,0,0,0)
 438 #define register_param3(p,s0,s1,s2)    register_param6(p,s0,s1,s2,0,0,0)
 439 #define register_param4(p,s0,s1,s2,s3) register_param6(p,s0,s1,s2,s3,0,0)
 440
 441
 442 static struct ureg register_input( struct texenv_fragment_program *p, GLuint input )
 443 {
 444    p->program->Base.InputsRead |= (1 << input);
 445    return make_ureg(PROGRAM_INPUT, input);
 446 }
 447
 448
 449 static void emit_arg( struct prog_src_register *reg,
 450                       struct ureg ureg )
 451 {
 452    reg->File = ureg.file;
 453    reg->Index = ureg.idx;
 454    reg->Swizzle = ureg.swz;
 455    reg->NegateBase = ureg.negatebase ? 0xf : 0x0;
 456    reg->Abs = ureg.abs;
 457    reg->NegateAbs = ureg.negateabs;
 458 }
 459
 460 static void emit_dst( struct prog_dst_register *dst,
 461                       struct ureg ureg, GLuint mask )
 462 {
 463    dst->File = ureg.file;
 464    dst->Index = ureg.idx;
 465    dst->WriteMask = mask;
 466    dst->CondMask = 0;
 467    dst->CondSwizzle = 0;
 468 }
 469
 470 static struct prog_instruction *
 471 emit_op(struct texenv_fragment_program *p,
 472         GLuint op,
 473         struct ureg dest,
 474         GLuint mask,
 475         GLuint saturate,
 476         struct ureg src0,
 477         struct ureg src1,
 478         struct ureg src2 )
 479 {
 480    GLuint nr = p->program->Base.NumInstructions++;
 481    struct prog_instruction *inst = &p->program->Base.Instructions[nr];
 482
 483    _mesa_memset(inst, 0, sizeof(*inst));
 484    inst->Opcode = op;
 485
 486    emit_arg( &inst->SrcReg[0], src0 );
 487    emit_arg( &inst->SrcReg[1], src1 );
 488    emit_arg( &inst->SrcReg[2], src2 );
 489
 490    inst->Saturate = saturate;
 491
 492    emit_dst( &inst->DstReg, dest, mask );
 493
 494    /* Accounting for indirection tracking:
 495     */
 496    if (dest.file == PROGRAM_TEMPORARY)
 497       p->temps_output |= 1 << dest.idx;
 498
 499    return inst;
 500 }
 501
 502
 503 static struct ureg emit_arith( struct texenv_fragment_program *p,
 504                                GLuint op,
 505                                struct ureg dest,
 506                                GLuint mask,
 507                                GLuint saturate,
 508                                struct ureg src0,
 509                                struct ureg src1,
 510                                struct ureg src2 )
 511 {
 512    emit_op(p, op, dest, mask, saturate, src0, src1, src2);
 513
 514    /* Accounting for indirection tracking:
 515     */
 516    if (src0.file == PROGRAM_TEMPORARY)
 517       p->alu_temps |= 1 << src0.idx;
 518
 519    if (!is_undef(src1) && src1.file == PROGRAM_TEMPORARY)
 520       p->alu_temps |= 1 << src1.idx;
 521
 522    if (!is_undef(src2) && src2.file == PROGRAM_TEMPORARY)
 523       p->alu_temps |= 1 << src2.idx;
 524
 525    if (dest.file == PROGRAM_TEMPORARY)
 526       p->alu_temps |= 1 << dest.idx;
 527
 528    p->program->NumAluInstructions++;
 529    return dest;
 530 }
 531
 532 static struct ureg emit_texld( struct texenv_fragment_program *p,
 533                                GLuint op,
 534                                struct ureg dest,
 535                                GLuint destmask,
 536                                GLuint tex_unit,
 537                                GLuint tex_idx,
 538                                struct ureg coord )
 539 {
 540    struct prog_instruction *inst = emit_op( p, op,
 541                                           dest, destmask,
 542                                           0,            /* don't saturate? */
 543                                           coord,        /* arg 0? */
 544                                           undef,
 545                                           undef);
 546
 547    inst->TexSrcTarget = tex_idx;
 548    inst->TexSrcUnit = tex_unit;
 549
 550    p->program->NumTexInstructions++;
 551
 552    /* Is this a texture indirection?
 553     */
 554    if ((coord.file == PROGRAM_TEMPORARY &&
 555         (p->temps_output & (1<<coord.idx))) ||
 556        (dest.file == PROGRAM_TEMPORARY &&
 557         (p->alu_temps & (1<<dest.idx)))) {
 558       p->program->NumTexIndirections++;
 559       p->temps_output = 1<<coord.idx;
 560       p->alu_temps = 0;
 561       assert(0);                /* KW: texture env crossbar */
 562    }
 563
 564    return dest;
 565 }
 566
 567
 568 static struct ureg register_const4f( struct texenv_fragment_program *p,
 569                                      GLfloat s0,
 570                                      GLfloat s1,
 571                                      GLfloat s2,
 572                                      GLfloat s3)
 573 {
 574    GLfloat values[4];
 575    GLuint idx;
 576    values[0] = s0;
 577    values[1] = s1;
 578    values[2] = s2;
 579    values[3] = s3;
 580    idx = _mesa_add_unnamed_constant( p->program->Base.Parameters, values );
 581    return make_ureg(PROGRAM_STATE_VAR, idx);
 582 }
 583
 584 #define register_scalar_const(p, s0)    register_const4f(p, s0, s0, s0, s0)
 585 #define register_const1f(p, s0)         register_const4f(p, s0, 0, 0, 1)
 586 #define register_const2f(p, s0, s1)     register_const4f(p, s0, s1, 0, 1)
 587 #define register_const3f(p, s0, s1, s2) register_const4f(p, s0, s1, s2, 1)
 588
 589
 590
 591
 592 static struct ureg get_one( struct texenv_fragment_program *p )
 593 {
 594    if (is_undef(p->one))
 595       p->one = register_scalar_const(p, 1.0);
 596    return p->one;
 597 }
 598
 599 static struct ureg get_half( struct texenv_fragment_program *p )
 600 {
 601    if (is_undef(p->half))
 602       p->one = register_scalar_const(p, 0.5);
 603    return p->half;
 604 }
 605
 606 static struct ureg get_zero( struct texenv_fragment_program *p )
 607 {
 608    if (is_undef(p->zero))
 609       p->one = register_scalar_const(p, 0.0);
 610    return p->zero;
 611 }
 612
 613
 614
 615
 616
 617 static void program_error( struct texenv_fragment_program *p, const char *msg )
 618 {
 619    _mesa_problem(NULL, msg);
 620    p->error = 1;
 621 }
 622
 623 static struct ureg get_source( struct texenv_fragment_program *p,
 624                                GLuint src, GLuint unit )
 625 {
 626    switch (src) {
 627    case SRC_TEXTURE:
 628       assert(!is_undef(p->src_texture[unit]));
 629       return p->src_texture[unit];
 630
 631    case SRC_TEXTURE0:
 632    case SRC_TEXTURE1:
 633    case SRC_TEXTURE2:
 634    case SRC_TEXTURE3:
 635    case SRC_TEXTURE4:
 636    case SRC_TEXTURE5:
 637    case SRC_TEXTURE6:
 638    case SRC_TEXTURE7:
 639       assert(!is_undef(p->src_texture[src - SRC_TEXTURE0]));
 640       return p->src_texture[src - SRC_TEXTURE0];
 641
 642    case SRC_CONSTANT:
 643       return register_param2(p, STATE_TEXENV_COLOR, unit);
 644
 645    case SRC_PRIMARY_COLOR:
 646       return register_input(p, FRAG_ATTRIB_COL0);
 647
 648    case SRC_PREVIOUS:
 649    default:
 650       if (is_undef(p->src_previous))
 651          return register_input(p, FRAG_ATTRIB_COL0);
 652       else
 653          return p->src_previous;
 654    }
 655 }
 656
 657 static struct ureg emit_combine_source( struct texenv_fragment_program *p,
 658                                         GLuint mask,
 659                                         GLuint unit,
 660                                         GLuint source,
 661                                         GLuint operand )
 662 {
 663    struct ureg arg, src, one;
 664
 665    src = get_source(p, source, unit);
 666
 667    switch (operand) {
 668    case OPR_ONE_MINUS_SRC_COLOR:
 669       /* Get unused tmp,
 670        * Emit tmp = 1.0 - arg.xyzw
 671        */
 672       arg = get_temp( p );
 673       one = get_one( p );
 674       return emit_arith( p, OPCODE_SUB, arg, mask, 0, one, src, undef);
 675
 676    case OPR_SRC_ALPHA:
 677       if (mask == WRITEMASK_W)
 678          return src;
 679       else
 680          return swizzle1( src, W );
 681    case OPR_ONE_MINUS_SRC_ALPHA:
 682       /* Get unused tmp,
 683        * Emit tmp = 1.0 - arg.wwww
 684        */
 685       arg = get_temp(p);
 686       one = get_one(p);
 687       return emit_arith(p, OPCODE_SUB, arg, mask, 0,
 688                         one, swizzle1(src, W), undef);
 689    case OPR_ZERO:
 690       return get_zero(p);
 691    case OPR_ONE:
 692       return get_one(p);
 693    case OPR_SRC_COLOR:
 694    default:
 695       return src;
 696    }
 697 }
 698
 699 static GLboolean args_match( struct state_key *key, GLuint unit )
 700 {
 701    int i, nr = key->unit[unit].NumArgsRGB;
 702
 703    for (i = 0 ; i < nr ; i++) {
 704       if (key->unit[unit].OptA[i].Source != key->unit[unit].OptRGB[i].Source)
 705          return GL_FALSE;
 706
 707       switch(key->unit[unit].OptA[i].Operand) {
 708       case OPR_SRC_ALPHA:
 709          switch(key->unit[unit].OptRGB[i].Operand) {
 710          case OPR_SRC_COLOR:
 711          case OPR_SRC_ALPHA:
 712             break;
 713          default:
 714             return GL_FALSE;
 715          }
 716          break;
 717       case OPR_ONE_MINUS_SRC_ALPHA:
 718          switch(key->unit[unit].OptRGB[i].Operand) {
 719          case OPR_ONE_MINUS_SRC_COLOR:
 720          case OPR_ONE_MINUS_SRC_ALPHA:
 721             break;
 722          default:
 723             return GL_FALSE;
 724          }
 725          break;
 726       default:
 727          return GL_FALSE;       /* impossible */
 728       }
 729    }
 730
 731    return GL_TRUE;
 732 }
 733
 734 static struct ureg emit_combine( struct texenv_fragment_program *p,
 735                                  struct ureg dest,
 736                                  GLuint mask,
 737                                  GLuint saturate,
 738                                  GLuint unit,
 739                                  GLuint nr,
 740                                  GLuint mode,
 741                                  struct mode_opt *opt)
 742 {
 743    struct ureg src[3];
 744    struct ureg tmp, half;
 745    int i;
 746
 747    for (i = 0; i < nr; i++)
 748       src[i] = emit_combine_source( p, mask, unit, opt[i].Source, opt[i].Operand );
 749
 750    switch (mode) {
 751    case MODE_REPLACE:
 752       if (mask == WRITEMASK_XYZW && !saturate)
 753          return src[0];
 754       else
 755          return emit_arith( p, OPCODE_MOV, dest, mask, saturate, src[0], undef, undef );
 756    case MODE_MODULATE:
 757       return emit_arith( p, OPCODE_MUL, dest, mask, saturate,
 758                          src[0], src[1], undef );
 759    case MODE_ADD:
 760       return emit_arith( p, OPCODE_ADD, dest, mask, saturate,
 761                          src[0], src[1], undef );
 762    case MODE_ADD_SIGNED:
 763       /* tmp = arg0 + arg1
 764        * result = tmp - .5
 765        */
 766       half = get_half(p);
 767       emit_arith( p, OPCODE_ADD, tmp, mask, 0, src[0], src[1], undef );
 768       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp, half, undef );
 769       return dest;
 770    case MODE_INTERPOLATE:
 771       /* Arg0 * (Arg2) + Arg1 * (1-Arg2) -- note arguments are reordered:
 772        */
 773       return emit_arith( p, OPCODE_LRP, dest, mask, saturate, src[2], src[0], src[1] );
 774
 775    case MODE_SUBTRACT:
 776       return emit_arith( p, OPCODE_SUB, dest, mask, saturate, src[0], src[1], undef );
 777
 778    case MODE_DOT3_RGBA:
 779    case MODE_DOT3_RGBA_EXT:
 780    case MODE_DOT3_RGB_EXT:
 781    case MODE_DOT3_RGB: {
 782       struct ureg tmp0 = get_temp( p );
 783       struct ureg tmp1 = get_temp( p );
 784       struct ureg neg1 = register_scalar_const(p, -1);
 785       struct ureg two  = register_scalar_const(p, 2);
 786
 787       /* tmp0 = 2*src0 - 1
 788        * tmp1 = 2*src1 - 1
 789        *
 790        * dst = tmp0 dot3 tmp1
 791        */
 792       emit_arith( p, OPCODE_MAD, tmp0, WRITEMASK_XYZW, 0,
 793                   two, src[0], neg1);
 794
 795       if (_mesa_memcmp(&src[0], &src[1], sizeof(struct ureg)) == 0)
 796          tmp1 = tmp0;
 797       else
 798          emit_arith( p, OPCODE_MAD, tmp1, WRITEMASK_XYZW, 0,
 799                      two, src[1], neg1);
 800       emit_arith( p, OPCODE_DP3, dest, mask, saturate, tmp0, tmp1, undef);
 801       return dest;
 802    }
 803    case MODE_MODULATE_ADD_ATI:
 804       /* Arg0 * Arg2 + Arg1 */
 805       return emit_arith( p, OPCODE_MAD, dest, mask, saturate,
 806                          src[0], src[2], src[1] );
 807    case MODE_MODULATE_SIGNED_ADD_ATI: {
 808       /* Arg0 * Arg2 + Arg1 - 0.5 */
 809       struct ureg tmp0 = get_temp(p);
 810       half = get_half(p);
 811       emit_arith( p, OPCODE_MAD, tmp0, mask, 0, src[0], src[2], src[1] );
 812       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp0, half, undef );
 813       return dest;
 814    }
 815    case MODE_MODULATE_SUBTRACT_ATI:
 816       /* Arg0 * Arg2 - Arg1 */
 817       emit_arith( p, OPCODE_MAD, dest, mask, 0, src[0], src[2], negate(src[1]) );
 818       return dest;
 819    default:
 820       return src[0];
 821    }
 822 }
 823
 824
 825 static struct ureg emit_texenv( struct texenv_fragment_program *p, int unit )
 826 {
 827    struct state_key *key = p->state;
 828    GLuint saturate = (unit < p->last_tex_stage);
 829    GLuint rgb_shift, alpha_shift;
 830    struct ureg out, shift;
 831    struct ureg dest;
 832
 833    if (!key->unit[unit].enabled) {
 834       return get_source(p, SRC_PREVIOUS, 0);
 835    }
 836
 837    switch (key->unit[unit].ModeRGB) {
 838    case MODE_DOT3_RGB_EXT:
 839       alpha_shift = key->unit[unit].ScaleShiftA;
 840       rgb_shift = 0;
 841       break;
 842    case MODE_DOT3_RGBA_EXT:
 843       alpha_shift = 0;
 844       rgb_shift = 0;
 845       break;
 846    default:
 847       rgb_shift = key->unit[unit].ScaleShiftRGB;
 848       alpha_shift = key->unit[unit].ScaleShiftA;
 849       break;
 850    }
 851
 852    /* If this is the very last calculation, emit direct to output reg:
 853     */
 854    if (key->separate_specular ||
 855        unit != p->last_tex_stage ||
 856        alpha_shift ||
 857        rgb_shift)
 858       dest = get_temp( p );
 859    else
 860       dest = make_ureg(PROGRAM_OUTPUT, FRAG_RESULT_COLR);
 861
 862    /* Emit the RGB and A combine ops
 863     */
 864    if (key->unit[unit].ModeRGB == key->unit[unit].ModeA &&
 865        args_match(key, unit)) {
 866       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 867                           unit,
 868                           key->unit[unit].NumArgsRGB,
 869                           key->unit[unit].ModeRGB,
 870                           key->unit[unit].OptRGB);
 871    }
 872    else if (key->unit[unit].ModeRGB == MODE_DOT3_RGBA_EXT ||
 873             key->unit[unit].ModeA == MODE_DOT3_RGBA) {
 874
 875       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 876                           unit,
 877                           key->unit[unit].NumArgsRGB,
 878                           key->unit[unit].ModeRGB,
 879                           key->unit[unit].OptRGB);
 880    }
 881    else {
 882       /* Need to do something to stop from re-emitting identical
 883        * argument calculations here:
 884        */
 885       out = emit_combine( p, dest, WRITEMASK_XYZ, saturate,
 886                           unit,
 887                           key->unit[unit].NumArgsRGB,
 888                           key->unit[unit].ModeRGB,
 889                           key->unit[unit].OptRGB);
 890       out = emit_combine( p, dest, WRITEMASK_W, saturate,
 891                           unit,
 892                           key->unit[unit].NumArgsA,
 893                           key->unit[unit].ModeA,
 894                           key->unit[unit].OptA);
 895    }
 896
 897    /* Deal with the final shift:
 898     */
 899    if (alpha_shift || rgb_shift) {
 900       if (rgb_shift == alpha_shift) {
 901          shift = register_scalar_const(p, 1<<rgb_shift);
 902       }
 903       else {
 904          shift = register_const4f(p,
 905                                   1<<rgb_shift,
 906                                   1<<rgb_shift,
 907                                   1<<rgb_shift,
 908                                   1<<alpha_shift);
 909       }
 910       return emit_arith( p, OPCODE_MUL, dest, WRITEMASK_XYZW,
 911                          saturate, out, shift, undef );
 912    }
 913    else
 914       return out;
 915 }
 916
 917
 918
 919 static void load_texture( struct texenv_fragment_program *p, GLuint unit )
 920 {
 921    if (is_undef(p->src_texture[unit])) {
 922       GLuint dim = p->state->unit[unit].source_index;
 923       struct ureg texcoord = register_input(p, FRAG_ATTRIB_TEX0+unit);
 924       struct ureg tmp = get_tex_temp( p );
 925
 926       if (dim == TEXTURE_UNKNOWN_INDEX)
 927          program_error(p, "TexSrcBit");
 928
 929       /* TODO: Use D0_MASK_XY where possible.
 930        */
 931       p->src_texture[unit] = emit_texld( p, OPCODE_TXP,
 932                                          tmp, WRITEMASK_XYZW,
 933                                          unit, dim, texcoord );
 934    }
 935 }
 936
 937 static GLboolean load_texenv_source( struct texenv_fragment_program *p,
 938                                      GLuint src, GLuint unit )
 939 {
 940    switch (src) {
 941    case SRC_TEXTURE:
 942       load_texture(p, unit);
 943       break;
 944
 945    case SRC_TEXTURE0:
 946    case SRC_TEXTURE1:
 947    case SRC_TEXTURE2:
 948    case SRC_TEXTURE3:
 949    case SRC_TEXTURE4:
 950    case SRC_TEXTURE5:
 951    case SRC_TEXTURE6:
 952    case SRC_TEXTURE7:
 953       if (!p->state->unit[src - SRC_TEXTURE0].enabled)
 954          return GL_FALSE;
 955       load_texture(p, src - SRC_TEXTURE0);
 956       break;
 957
 958    default:
 959       break;
 960    }
 961
 962    return GL_TRUE;
 963 }
 964
 965 static GLboolean load_texunit_sources( struct texenv_fragment_program *p, int unit )
 966 {
 967    struct state_key *key = p->state;
 968    int i, nr = key->unit[unit].NumArgsRGB;
 969    for (i = 0; i < nr; i++) {
 970       if (!load_texenv_source( p, key->unit[unit].OptRGB[i].Source, unit) ||
 971           !load_texenv_source( p, key->unit[unit].OptA[i].Source, unit ))
 972          return GL_FALSE;
 973    }
 974    return GL_TRUE;
 975 }
 976
 977 static void create_new_program(struct state_key *key, GLcontext *ctx,
 978                                struct fragment_program *program)
 979 {
 980    struct texenv_fragment_program p;
 981    GLuint unit;
 982    struct ureg cf, out;
 983
 984    _mesa_memset(&p, 0, sizeof(p));
 985    p.ctx = ctx;
 986    p.state = key;
 987    p.program = program;
 988
 989    p.program->Base.Instructions = MALLOC(sizeof(struct prog_instruction) * 100);
 990    p.program->Base.NumInstructions = 0;
 991    p.program->Base.Target = GL_FRAGMENT_PROGRAM_ARB;
 992    p.program->NumTexIndirections = 1;   /* correct? */
 993    p.program->NumTexInstructions = 0;
 994    p.program->NumAluInstructions = 0;
 995    p.program->Base.String = 0;
 996    p.program->Base.NumInstructions =
 997       p.program->Base.NumTemporaries =
 998       p.program->Base.NumParameters =
 999       p.program->Base.NumAttributes = p.program->Base.NumAddressRegs = 0;
1000    p.program->Base.Parameters = _mesa_new_parameter_list();
1001
1002    p.program->Base.InputsRead = 0;
1003    p.program->Base.OutputsWritten = 1 << FRAG_RESULT_COLR;
1004
1005    for (unit = 0; unit < MAX_TEXTURE_UNITS; unit++)
1006       p.src_texture[unit] = undef;
1007
1008    p.src_previous = undef;
1009    p.last_tex_stage = 0;
1010    release_temps(&p);
1011
1012    if (key->enabled_units) {
1013       /* First pass - to support texture_env_crossbar, first identify
1014        * all referenced texture sources and emit texld instructions
1015        * for each:
1016        */
1017       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits ; unit++)
1018          if (key->unit[unit].enabled) {
1019             if (load_texunit_sources( &p, unit ))
1020                p.last_tex_stage = unit;
1021          }
1022
1023       /* Second pass - emit combine instructions to build final color:
1024        */
1025       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits; unit++)
1026          if (key->enabled_units & (1<<unit)) {
1027             p.src_previous = emit_texenv( &p, unit );
1028             release_temps(&p);  /* release all temps */
1029          }
1030    }
1031
1032    cf = get_source( &p, SRC_PREVIOUS, 0 );
1033    out = make_ureg( PROGRAM_OUTPUT, FRAG_RESULT_COLR );
1034
1035    if (key->separate_specular) {
1036       /* Emit specular add.
1037        */
1038       struct ureg s = register_input(&p, FRAG_ATTRIB_COL1);
1039       emit_arith( &p, OPCODE_ADD, out, WRITEMASK_XYZ, 0, cf, s, undef );
1040       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_W, 0, cf, undef, undef );
1041    }
1042    else if (_mesa_memcmp(&cf, &out, sizeof(cf)) != 0) {
1043       /* Will wind up in here if no texture enabled or a couple of
1044        * other scenarios (GL_REPLACE for instance).
1045        */
1046       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_XYZW, 0, cf, undef, undef );
1047    }
1048
1049    /* Finish up:
1050     */
1051    emit_arith( &p, OPCODE_END, undef, WRITEMASK_XYZW, 0, undef, undef, undef);
1052
1053    if (key->fog_enabled) {
1054       /* Pull fog mode from GLcontext, the value in the state key is
1055        * a reduced value and not what is expected in FogOption
1056        */
1057       p.program->FogOption = ctx->Fog.Mode;
1058    } else
1059       p.program->FogOption = GL_NONE;
1060
1061    if (p.program->NumTexIndirections > ctx->Const.FragmentProgram.MaxTexIndirections)
1062       program_error(&p, "Exceeded max nr indirect texture lookups");
1063
1064    if (p.program->NumTexInstructions > ctx->Const.FragmentProgram.MaxTexInstructions)
1065       program_error(&p, "Exceeded max TEX instructions");
1066
1067    if (p.program->NumAluInstructions > ctx->Const.FragmentProgram.MaxAluInstructions)
1068       program_error(&p, "Exceeded max ALU instructions");
1069
1070
1071    /* Notify driver the fragment program has (actually) changed.
1072     */
1073    if (ctx->Driver.ProgramStringNotify || DISASSEM) {
1074       if (ctx->Driver.ProgramStringNotify)
1075          ctx->Driver.ProgramStringNotify( ctx, GL_FRAGMENT_PROGRAM_ARB,
1076                                           &p.program->Base );
1077
1078       if (DISASSEM) {
1079          _mesa_print_program(&p.program->Base);
1080          _mesa_printf("\n");
1081       }
1082
1083    }
1084
1085 }
1086
1087 static void *search_cache( struct texenvprog_cache *cache,
1088                            GLuint hash,
1089                            const void *key,
1090                            GLuint keysize)
1091 {
1092    struct texenvprog_cache *c;
1093
1094    for (c = cache; c; c = c->next) {
1095       if (c->hash == hash && _mesa_memcmp(c->key, key, keysize) == 0)
1096          return c->data;
1097    }
1098
1099    return NULL;
1100 }
1101
1102 static void cache_item( struct texenvprog_cache **cache,
1103                         GLuint hash,
1104                         void *key,
1105                         void *data )
1106 {
1107    struct texenvprog_cache *c = MALLOC(sizeof(*c));
1108    c->hash = hash;
1109    c->key = key;
1110    c->data = data;
1111    c->next = *cache;
1112    *cache = c;
1113 }
1114
1115 static GLuint hash_key( struct state_key *key )
1116 {
1117    GLuint *ikey = (GLuint *)key;
1118    GLuint hash = 0, i;
1119
1120    /* I'm sure this can be improved on, but speed is important:
1121     */
1122    for (i = 0; i < sizeof(*key)/sizeof(GLuint); i++)
1123       hash ^= ikey[i];
1124
1125    return hash;
1126 }
1127
1128 void _mesa_UpdateTexEnvProgram( GLcontext *ctx )
1129 {
1130    struct state_key *key;
1131    GLuint hash;
1132    struct fragment_program *prev = ctx->FragmentProgram._Current;
1133
1134    if (!ctx->FragmentProgram._Enabled) {
1135       key = make_state_key(ctx);
1136       hash = hash_key(key);
1137
1138       ctx->FragmentProgram._Current = ctx->_TexEnvProgram =
1139          (struct fragment_program *)
1140          search_cache(ctx->Texture.env_fp_cache, hash, key, sizeof(*key));
1141
1142       if (!ctx->_TexEnvProgram) {
1143          if (0) _mesa_printf("Building new texenv proggy for key %x\n", hash);
1144
1145          ctx->FragmentProgram._Current = ctx->_TexEnvProgram =
1146             (struct fragment_program *)
1147             ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
1148
1149          create_new_program(key, ctx, ctx->_TexEnvProgram);
1150
1151          cache_item(&ctx->Texture.env_fp_cache, hash, key, ctx->_TexEnvProgram);
1152       } else {
1153          FREE(key);
1154          if (0) _mesa_printf("Found existing texenv program for key %x\n", hash);
1155       }
1156    }
1157    else {
1158       ctx->FragmentProgram._Current = ctx->FragmentProgram.Current;
1159    }
1160
1161    /* Tell the driver about the change.  Could define a new target for
1162     * this?
1163     */
1164    if (ctx->FragmentProgram._Current != prev)
1165       ctx->Driver.BindProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, (struct program *)
1166                               ctx->FragmentProgram._Current);
1167 }
1168
1169 void _mesa_TexEnvProgramCacheDestroy( GLcontext *ctx )
1170 {
1171    struct texenvprog_cache *a, *tmp;
1172
1173    for (a = ctx->Texture.env_fp_cache; a; a = tmp) {
1174       tmp = a->next;
1175       FREE(a->key);
1176       FREE(a->data);
1177       FREE(a);
1178    }
1179 }
1180