src/mesa/main/texenvprogram.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "glheader.h"
  29 #include "macros.h"
  30 #include "enums.h"
  31 #include "shader/prog_parameter.h"
  32 #include "shader/prog_cache.h"
  33 #include "shader/prog_instruction.h"
  34 #include "shader/prog_print.h"
  35 #include "shader/prog_statevars.h"
  36 #include "shader/programopt.h"
  37 #include "texenvprogram.h"
  38
  39 /**
  40  * This MAX is probably a bit generous, but that's OK.  There can be
  41  * up to four instructions per texture unit (TEX + 3 for combine),
  42  * then there's fog and specular add.
  43  */
  44 #define MAX_INSTRUCTIONS ((MAX_TEXTURE_UNITS * 4) + 12)
  45
  46 #define DISASSEM (MESA_VERBOSE & VERBOSE_DISASSEM)
  47
  48 struct mode_opt {
  49    GLuint Source:4;
  50    GLuint Operand:3;
  51 };
  52
  53 struct state_key {
  54    GLbitfield enabled_units;
  55    GLuint separate_specular:1;
  56    GLuint fog_enabled:1;
  57    GLuint fog_mode:2;
  58
  59    struct {
  60       GLuint enabled:1;
  61       GLuint source_index:3;   /* one of TEXTURE_1D/2D/3D/CUBE/RECT_INDEX */
  62       GLuint ScaleShiftRGB:2;
  63       GLuint ScaleShiftA:2;
  64
  65       GLuint NumArgsRGB:2;
  66       GLuint ModeRGB:4;
  67       struct mode_opt OptRGB[3];
  68
  69       GLuint NumArgsA:2;
  70       GLuint ModeA:4;
  71       struct mode_opt OptA[3];
  72    } unit[8];
  73 };
  74
  75 #define FOG_LINEAR  0
  76 #define FOG_EXP     1
  77 #define FOG_EXP2    2
  78 #define FOG_UNKNOWN 3
  79
  80 static GLuint translate_fog_mode( GLenum mode )
  81 {
  82    switch (mode) {
  83    case GL_LINEAR: return FOG_LINEAR;
  84    case GL_EXP: return FOG_EXP;
  85    case GL_EXP2: return FOG_EXP2;
  86    default: return FOG_UNKNOWN;
  87    }
  88 }
  89
  90 #define OPR_SRC_COLOR           0
  91 #define OPR_ONE_MINUS_SRC_COLOR 1
  92 #define OPR_SRC_ALPHA           2
  93 #define OPR_ONE_MINUS_SRC_ALPHA 3
  94 #define OPR_ZERO                4
  95 #define OPR_ONE                 5
  96 #define OPR_UNKNOWN             7
  97
  98 static GLuint translate_operand( GLenum operand )
  99 {
 100    switch (operand) {
 101    case GL_SRC_COLOR: return OPR_SRC_COLOR;
 102    case GL_ONE_MINUS_SRC_COLOR: return OPR_ONE_MINUS_SRC_COLOR;
 103    case GL_SRC_ALPHA: return OPR_SRC_ALPHA;
 104    case GL_ONE_MINUS_SRC_ALPHA: return OPR_ONE_MINUS_SRC_ALPHA;
 105    case GL_ZERO: return OPR_ZERO;
 106    case GL_ONE: return OPR_ONE;
 107    default:     return OPR_UNKNOWN;
 108    }
 109 }
 110
 111 #define SRC_TEXTURE  0
 112 #define SRC_TEXTURE0 1
 113 #define SRC_TEXTURE1 2
 114 #define SRC_TEXTURE2 3
 115 #define SRC_TEXTURE3 4
 116 #define SRC_TEXTURE4 5
 117 #define SRC_TEXTURE5 6
 118 #define SRC_TEXTURE6 7
 119 #define SRC_TEXTURE7 8
 120 #define SRC_CONSTANT 9
 121 #define SRC_PRIMARY_COLOR 10
 122 #define SRC_PREVIOUS 11
 123 #define SRC_UNKNOWN  15
 124
 125 static GLuint translate_source( GLenum src )
 126 {
 127    switch (src) {
 128    case GL_TEXTURE: return SRC_TEXTURE;
 129    case GL_TEXTURE0:
 130    case GL_TEXTURE1:
 131    case GL_TEXTURE2:
 132    case GL_TEXTURE3:
 133    case GL_TEXTURE4:
 134    case GL_TEXTURE5:
 135    case GL_TEXTURE6:
 136    case GL_TEXTURE7: return SRC_TEXTURE0 + (src - GL_TEXTURE0);
 137    case GL_CONSTANT: return SRC_CONSTANT;
 138    case GL_PRIMARY_COLOR: return SRC_PRIMARY_COLOR;
 139    case GL_PREVIOUS: return SRC_PREVIOUS;
 140    default: return SRC_UNKNOWN;
 141    }
 142 }
 143
 144 #define MODE_REPLACE       0
 145 #define MODE_MODULATE      1
 146 #define MODE_ADD           2
 147 #define MODE_ADD_SIGNED    3
 148 #define MODE_INTERPOLATE   4
 149 #define MODE_SUBTRACT      5
 150 #define MODE_DOT3_RGB      6
 151 #define MODE_DOT3_RGB_EXT  7
 152 #define MODE_DOT3_RGBA     8
 153 #define MODE_DOT3_RGBA_EXT 9
 154 #define MODE_MODULATE_ADD_ATI           10
 155 #define MODE_MODULATE_SIGNED_ADD_ATI    11
 156 #define MODE_MODULATE_SUBTRACT_ATI      12
 157 #define MODE_UNKNOWN       15
 158
 159 static GLuint translate_mode( GLenum mode )
 160 {
 161    switch (mode) {
 162    case GL_REPLACE: return MODE_REPLACE;
 163    case GL_MODULATE: return MODE_MODULATE;
 164    case GL_ADD: return MODE_ADD;
 165    case GL_ADD_SIGNED: return MODE_ADD_SIGNED;
 166    case GL_INTERPOLATE: return MODE_INTERPOLATE;
 167    case GL_SUBTRACT: return MODE_SUBTRACT;
 168    case GL_DOT3_RGB: return MODE_DOT3_RGB;
 169    case GL_DOT3_RGB_EXT: return MODE_DOT3_RGB_EXT;
 170    case GL_DOT3_RGBA: return MODE_DOT3_RGBA;
 171    case GL_DOT3_RGBA_EXT: return MODE_DOT3_RGBA_EXT;
 172    case GL_MODULATE_ADD_ATI: return MODE_MODULATE_ADD_ATI;
 173    case GL_MODULATE_SIGNED_ADD_ATI: return MODE_MODULATE_SIGNED_ADD_ATI;
 174    case GL_MODULATE_SUBTRACT_ATI: return MODE_MODULATE_SUBTRACT_ATI;
 175    default: return MODE_UNKNOWN;
 176    }
 177 }
 178
 179 #define TEXTURE_UNKNOWN_INDEX 7
 180 static GLuint translate_tex_src_bit( GLbitfield bit )
 181 {
 182    switch (bit) {
 183    case TEXTURE_1D_BIT:   return TEXTURE_1D_INDEX;
 184    case TEXTURE_2D_BIT:   return TEXTURE_2D_INDEX;
 185    case TEXTURE_RECT_BIT: return TEXTURE_RECT_INDEX;
 186    case TEXTURE_3D_BIT:   return TEXTURE_3D_INDEX;
 187    case TEXTURE_CUBE_BIT: return TEXTURE_CUBE_INDEX;
 188    default: return TEXTURE_UNKNOWN_INDEX;
 189    }
 190 }
 191
 192 /**
 193  * Examine current texture environment state and generate a unique
 194  * key to identify it.
 195  */
 196 static void make_state_key( GLcontext *ctx,  struct state_key *key )
 197 {
 198    GLuint i, j;
 199
 200    memset(key, 0, sizeof(*key));
 201
 202    for (i=0;i<MAX_TEXTURE_UNITS;i++) {
 203       const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[i];
 204
 205       if (!texUnit->_ReallyEnabled)
 206          continue;
 207
 208       key->unit[i].enabled = 1;
 209       key->enabled_units |= (1<<i);
 210
 211       key->unit[i].source_index =
 212          translate_tex_src_bit(texUnit->_ReallyEnabled);
 213
 214       key->unit[i].NumArgsRGB = texUnit->_CurrentCombine->_NumArgsRGB;
 215       key->unit[i].NumArgsA = texUnit->_CurrentCombine->_NumArgsA;
 216
 217       key->unit[i].ModeRGB =
 218          translate_mode(texUnit->_CurrentCombine->ModeRGB);
 219       key->unit[i].ModeA =
 220          translate_mode(texUnit->_CurrentCombine->ModeA);
 221
 222       key->unit[i].ScaleShiftRGB = texUnit->_CurrentCombine->ScaleShiftRGB;
 223       key->unit[i].ScaleShiftA = texUnit->_CurrentCombine->ScaleShiftA;
 224
 225       for (j=0;j<3;j++) {
 226          key->unit[i].OptRGB[j].Operand =
 227             translate_operand(texUnit->_CurrentCombine->OperandRGB[j]);
 228          key->unit[i].OptA[j].Operand =
 229             translate_operand(texUnit->_CurrentCombine->OperandA[j]);
 230          key->unit[i].OptRGB[j].Source =
 231             translate_source(texUnit->_CurrentCombine->SourceRGB[j]);
 232          key->unit[i].OptA[j].Source =
 233             translate_source(texUnit->_CurrentCombine->SourceA[j]);
 234       }
 235    }
 236
 237    if (ctx->_TriangleCaps & DD_SEPARATE_SPECULAR)
 238       key->separate_specular = 1;
 239
 240    if (ctx->Fog.Enabled) {
 241       key->fog_enabled = 1;
 242       key->fog_mode = translate_fog_mode(ctx->Fog.Mode);
 243    }
 244 }
 245
 246 /* Use uregs to represent registers internally, translate to Mesa's
 247  * expected formats on emit.
 248  *
 249  * NOTE: These are passed by value extensively in this file rather
 250  * than as usual by pointer reference.  If this disturbs you, try
 251  * remembering they are just 32bits in size.
 252  *
 253  * GCC is smart enough to deal with these dword-sized structures in
 254  * much the same way as if I had defined them as dwords and was using
 255  * macros to access and set the fields.  This is much nicer and easier
 256  * to evolve.
 257  */
 258 struct ureg {
 259    GLuint file:4;
 260    GLuint idx:8;
 261    GLuint negatebase:1;
 262    GLuint abs:1;
 263    GLuint negateabs:1;
 264    GLuint swz:12;
 265    GLuint pad:5;
 266 };
 267
 268 static const struct ureg undef = {
 269    PROGRAM_UNDEFINED,
 270    ~0,
 271    0,
 272    0,
 273    0,
 274    0,
 275    0
 276 };
 277
 278
 279 /* State used to build the fragment program:
 280  */
 281 struct texenv_fragment_program {
 282    struct gl_fragment_program *program;
 283    GLcontext *ctx;
 284    struct state_key *state;
 285
 286    GLbitfield alu_temps;        /* Track texture indirections, see spec. */
 287    GLbitfield temps_output;     /* Track texture indirections, see spec. */
 288    GLbitfield temp_in_use;      /* Tracks temporary regs which are in use. */
 289    GLboolean error;
 290
 291    struct ureg src_texture[MAX_TEXTURE_UNITS];
 292    /* Reg containing each texture unit's sampled texture color,
 293     * else undef.
 294     */
 295
 296    struct ureg src_previous;    /* Reg containing color from previous
 297                                  * stage.  May need to be decl'd.
 298                                  */
 299
 300    GLuint last_tex_stage;       /* Number of last enabled texture unit */
 301
 302    struct ureg half;
 303    struct ureg one;
 304    struct ureg zero;
 305 };
 306
 307
 308
 309 static struct ureg make_ureg(GLuint file, GLuint idx)
 310 {
 311    struct ureg reg;
 312    reg.file = file;
 313    reg.idx = idx;
 314    reg.negatebase = 0;
 315    reg.abs = 0;
 316    reg.negateabs = 0;
 317    reg.swz = SWIZZLE_NOOP;
 318    reg.pad = 0;
 319    return reg;
 320 }
 321
 322 static struct ureg swizzle( struct ureg reg, int x, int y, int z, int w )
 323 {
 324    reg.swz = MAKE_SWIZZLE4(GET_SWZ(reg.swz, x),
 325                            GET_SWZ(reg.swz, y),
 326                            GET_SWZ(reg.swz, z),
 327                            GET_SWZ(reg.swz, w));
 328
 329    return reg;
 330 }
 331
 332 static struct ureg swizzle1( struct ureg reg, int x )
 333 {
 334    return swizzle(reg, x, x, x, x);
 335 }
 336
 337 static struct ureg negate( struct ureg reg )
 338 {
 339    reg.negatebase ^= 1;
 340    return reg;
 341 }
 342
 343 static GLboolean is_undef( struct ureg reg )
 344 {
 345    return reg.file == PROGRAM_UNDEFINED;
 346 }
 347
 348
 349 static struct ureg get_temp( struct texenv_fragment_program *p )
 350 {
 351    GLint bit;
 352
 353    /* First try and reuse temps which have been used already:
 354     */
 355    bit = _mesa_ffs( ~p->temp_in_use & p->alu_temps );
 356
 357    /* Then any unused temporary:
 358     */
 359    if (!bit)
 360       bit = _mesa_ffs( ~p->temp_in_use );
 361
 362    if (!bit) {
 363       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 364       _mesa_exit(1);
 365    }
 366
 367    if ((GLuint) bit > p->program->Base.NumTemporaries)
 368       p->program->Base.NumTemporaries = bit;
 369
 370    p->temp_in_use |= 1<<(bit-1);
 371    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 372 }
 373
 374 static struct ureg get_tex_temp( struct texenv_fragment_program *p )
 375 {
 376    int bit;
 377
 378    /* First try to find availble temp not previously used (to avoid
 379     * starting a new texture indirection).  According to the spec, the
 380     * ~p->temps_output isn't necessary, but will keep it there for
 381     * now:
 382     */
 383    bit = _mesa_ffs( ~p->temp_in_use & ~p->alu_temps & ~p->temps_output );
 384
 385    /* Then any unused temporary:
 386     */
 387    if (!bit)
 388       bit = _mesa_ffs( ~p->temp_in_use );
 389
 390    if (!bit) {
 391       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 392       _mesa_exit(1);
 393    }
 394
 395    if ((GLuint) bit > p->program->Base.NumTemporaries)
 396       p->program->Base.NumTemporaries = bit;
 397
 398    p->temp_in_use |= 1<<(bit-1);
 399    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 400 }
 401
 402
 403 static void release_temps(GLcontext *ctx, struct texenv_fragment_program *p )
 404 {
 405    GLuint max_temp = ctx->Const.FragmentProgram.MaxTemps;
 406
 407    /* KW: To support tex_env_crossbar, don't release the registers in
 408     * temps_output.
 409     */
 410    if (max_temp >= sizeof(int) * 8)
 411       p->temp_in_use = p->temps_output;
 412    else
 413       p->temp_in_use = ~((1<<max_temp)-1) | p->temps_output;
 414 }
 415
 416
 417 static struct ureg register_param5( struct texenv_fragment_program *p,
 418                                     GLint s0,
 419                                     GLint s1,
 420                                     GLint s2,
 421                                     GLint s3,
 422                                     GLint s4)
 423 {
 424    gl_state_index tokens[STATE_LENGTH];
 425    GLuint idx;
 426    tokens[0] = s0;
 427    tokens[1] = s1;
 428    tokens[2] = s2;
 429    tokens[3] = s3;
 430    tokens[4] = s4;
 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_param5(p,s0,0,0,0,0)
 437 #define register_param2(p,s0,s1)       register_param5(p,s0,s1,0,0,0)
 438 #define register_param3(p,s0,s1,s2)    register_param5(p,s0,s1,s2,0,0)
 439 #define register_param4(p,s0,s1,s2,s3) register_param5(p,s0,s1,s2,s3,0)
 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 = COND_TR;  /* always pass cond test */
 467    dst->CondSwizzle = SWIZZLE_NOOP;
 468 }
 469
 470 static struct prog_instruction *
 471 emit_op(struct texenv_fragment_program *p,
 472         enum prog_opcode op,
 473         struct ureg dest,
 474         GLuint mask,
 475         GLboolean 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    assert(nr < MAX_INSTRUCTIONS);
 484
 485    _mesa_init_instructions(inst, 1);
 486    inst->Opcode = op;
 487
 488    emit_arg( &inst->SrcReg[0], src0 );
 489    emit_arg( &inst->SrcReg[1], src1 );
 490    emit_arg( &inst->SrcReg[2], src2 );
 491
 492    inst->SaturateMode = saturate ? SATURATE_ZERO_ONE : SATURATE_OFF;
 493
 494    emit_dst( &inst->DstReg, dest, mask );
 495
 496    /* Accounting for indirection tracking:
 497     */
 498    if (dest.file == PROGRAM_TEMPORARY)
 499       p->temps_output |= 1 << dest.idx;
 500
 501    return inst;
 502 }
 503
 504
 505 static struct ureg emit_arith( struct texenv_fragment_program *p,
 506                                enum prog_opcode op,
 507                                struct ureg dest,
 508                                GLuint mask,
 509                                GLboolean saturate,
 510                                struct ureg src0,
 511                                struct ureg src1,
 512                                struct ureg src2 )
 513 {
 514    emit_op(p, op, dest, mask, saturate, src0, src1, src2);
 515
 516    /* Accounting for indirection tracking:
 517     */
 518    if (src0.file == PROGRAM_TEMPORARY)
 519       p->alu_temps |= 1 << src0.idx;
 520
 521    if (!is_undef(src1) && src1.file == PROGRAM_TEMPORARY)
 522       p->alu_temps |= 1 << src1.idx;
 523
 524    if (!is_undef(src2) && src2.file == PROGRAM_TEMPORARY)
 525       p->alu_temps |= 1 << src2.idx;
 526
 527    if (dest.file == PROGRAM_TEMPORARY)
 528       p->alu_temps |= 1 << dest.idx;
 529
 530    p->program->Base.NumAluInstructions++;
 531    return dest;
 532 }
 533
 534 static struct ureg emit_texld( struct texenv_fragment_program *p,
 535                                enum prog_opcode op,
 536                                struct ureg dest,
 537                                GLuint destmask,
 538                                GLuint tex_unit,
 539                                GLuint tex_idx,
 540                                struct ureg coord )
 541 {
 542    struct prog_instruction *inst = emit_op( p, op,
 543                                           dest, destmask,
 544                                           GL_FALSE,     /* don't saturate? */
 545                                           coord,        /* arg 0? */
 546                                           undef,
 547                                           undef);
 548
 549    inst->TexSrcTarget = tex_idx;
 550    inst->TexSrcUnit = tex_unit;
 551
 552    p->program->Base.NumTexInstructions++;
 553
 554    /* Is this a texture indirection?
 555     */
 556    if ((coord.file == PROGRAM_TEMPORARY &&
 557         (p->temps_output & (1<<coord.idx))) ||
 558        (dest.file == PROGRAM_TEMPORARY &&
 559         (p->alu_temps & (1<<dest.idx)))) {
 560       p->program->Base.NumTexIndirections++;
 561       p->temps_output = 1<<coord.idx;
 562       p->alu_temps = 0;
 563       assert(0);                /* KW: texture env crossbar */
 564    }
 565
 566    return dest;
 567 }
 568
 569
 570 static struct ureg register_const4f( struct texenv_fragment_program *p,
 571                                      GLfloat s0,
 572                                      GLfloat s1,
 573                                      GLfloat s2,
 574                                      GLfloat s3)
 575 {
 576    GLfloat values[4];
 577    GLuint idx, swizzle;
 578    values[0] = s0;
 579    values[1] = s1;
 580    values[2] = s2;
 581    values[3] = s3;
 582    idx = _mesa_add_unnamed_constant( p->program->Base.Parameters, values, 4,
 583                                      &swizzle );
 584    ASSERT(swizzle == SWIZZLE_NOOP);
 585    return make_ureg(PROGRAM_CONSTANT, idx);
 586 }
 587
 588 #define register_scalar_const(p, s0)    register_const4f(p, s0, s0, s0, s0)
 589 #define register_const1f(p, s0)         register_const4f(p, s0, 0, 0, 1)
 590 #define register_const2f(p, s0, s1)     register_const4f(p, s0, s1, 0, 1)
 591 #define register_const3f(p, s0, s1, s2) register_const4f(p, s0, s1, s2, 1)
 592
 593
 594 static struct ureg get_one( struct texenv_fragment_program *p )
 595 {
 596    if (is_undef(p->one))
 597       p->one = register_scalar_const(p, 1.0);
 598    return p->one;
 599 }
 600
 601 static struct ureg get_half( struct texenv_fragment_program *p )
 602 {
 603    if (is_undef(p->half))
 604       p->half = register_scalar_const(p, 0.5);
 605    return p->half;
 606 }
 607
 608 static struct ureg get_zero( struct texenv_fragment_program *p )
 609 {
 610    if (is_undef(p->zero))
 611       p->zero = register_scalar_const(p, 0.0);
 612    return p->zero;
 613 }
 614
 615
 616 static void program_error( struct texenv_fragment_program *p, const char *msg )
 617 {
 618    _mesa_problem(NULL, msg);
 619    p->error = 1;
 620 }
 621
 622 static struct ureg get_source( struct texenv_fragment_program *p,
 623                                GLuint src, GLuint unit )
 624 {
 625    switch (src) {
 626    case SRC_TEXTURE:
 627       assert(!is_undef(p->src_texture[unit]));
 628       return p->src_texture[unit];
 629
 630    case SRC_TEXTURE0:
 631    case SRC_TEXTURE1:
 632    case SRC_TEXTURE2:
 633    case SRC_TEXTURE3:
 634    case SRC_TEXTURE4:
 635    case SRC_TEXTURE5:
 636    case SRC_TEXTURE6:
 637    case SRC_TEXTURE7:
 638       assert(!is_undef(p->src_texture[src - SRC_TEXTURE0]));
 639       return p->src_texture[src - SRC_TEXTURE0];
 640
 641    case SRC_CONSTANT:
 642       return register_param2(p, STATE_TEXENV_COLOR, unit);
 643
 644    case SRC_PRIMARY_COLOR:
 645       return register_input(p, FRAG_ATTRIB_COL0);
 646
 647    case SRC_PREVIOUS:
 648    default:
 649       if (is_undef(p->src_previous))
 650          return register_input(p, FRAG_ATTRIB_COL0);
 651       else
 652          return p->src_previous;
 653    }
 654 }
 655
 656 static struct ureg emit_combine_source( struct texenv_fragment_program *p,
 657                                         GLuint mask,
 658                                         GLuint unit,
 659                                         GLuint source,
 660                                         GLuint operand )
 661 {
 662    struct ureg arg, src, one;
 663
 664    src = get_source(p, source, unit);
 665
 666    switch (operand) {
 667    case OPR_ONE_MINUS_SRC_COLOR:
 668       /* Get unused tmp,
 669        * Emit tmp = 1.0 - arg.xyzw
 670        */
 671       arg = get_temp( p );
 672       one = get_one( p );
 673       return emit_arith( p, OPCODE_SUB, arg, mask, 0, one, src, undef);
 674
 675    case OPR_SRC_ALPHA:
 676       if (mask == WRITEMASK_W)
 677          return src;
 678       else
 679          return swizzle1( src, SWIZZLE_W );
 680    case OPR_ONE_MINUS_SRC_ALPHA:
 681       /* Get unused tmp,
 682        * Emit tmp = 1.0 - arg.wwww
 683        */
 684       arg = get_temp(p);
 685       one = get_one(p);
 686       return emit_arith(p, OPCODE_SUB, arg, mask, 0,
 687                         one, swizzle1(src, SWIZZLE_W), undef);
 688    case OPR_ZERO:
 689       return get_zero(p);
 690    case OPR_ONE:
 691       return get_one(p);
 692    case OPR_SRC_COLOR:
 693    default:
 694       return src;
 695    }
 696 }
 697
 698 static GLboolean args_match( struct state_key *key, GLuint unit )
 699 {
 700    GLuint i, nr = key->unit[unit].NumArgsRGB;
 701
 702    for (i = 0 ; i < nr ; i++) {
 703       if (key->unit[unit].OptA[i].Source != key->unit[unit].OptRGB[i].Source)
 704          return GL_FALSE;
 705
 706       switch(key->unit[unit].OptA[i].Operand) {
 707       case OPR_SRC_ALPHA:
 708          switch(key->unit[unit].OptRGB[i].Operand) {
 709          case OPR_SRC_COLOR:
 710          case OPR_SRC_ALPHA:
 711             break;
 712          default:
 713             return GL_FALSE;
 714          }
 715          break;
 716       case OPR_ONE_MINUS_SRC_ALPHA:
 717          switch(key->unit[unit].OptRGB[i].Operand) {
 718          case OPR_ONE_MINUS_SRC_COLOR:
 719          case OPR_ONE_MINUS_SRC_ALPHA:
 720             break;
 721          default:
 722             return GL_FALSE;
 723          }
 724          break;
 725       default:
 726          return GL_FALSE;       /* impossible */
 727       }
 728    }
 729
 730    return GL_TRUE;
 731 }
 732
 733 static struct ureg emit_combine( struct texenv_fragment_program *p,
 734                                  struct ureg dest,
 735                                  GLuint mask,
 736                                  GLboolean saturate,
 737                                  GLuint unit,
 738                                  GLuint nr,
 739                                  GLuint mode,
 740                                  const struct mode_opt *opt)
 741 {
 742    struct ureg src[3];
 743    struct ureg tmp, half;
 744    GLuint i;
 745
 746    tmp = undef; /* silence warning (bug 5318) */
 747
 748    for (i = 0; i < nr; i++)
 749       src[i] = emit_combine_source( p, mask, unit, opt[i].Source, opt[i].Operand );
 750
 751    switch (mode) {
 752    case MODE_REPLACE:
 753       if (mask == WRITEMASK_XYZW && !saturate)
 754          return src[0];
 755       else
 756          return emit_arith( p, OPCODE_MOV, dest, mask, saturate, src[0], undef, undef );
 757    case MODE_MODULATE:
 758       return emit_arith( p, OPCODE_MUL, dest, mask, saturate,
 759                          src[0], src[1], undef );
 760    case MODE_ADD:
 761       return emit_arith( p, OPCODE_ADD, dest, mask, saturate,
 762                          src[0], src[1], undef );
 763    case MODE_ADD_SIGNED:
 764       /* tmp = arg0 + arg1
 765        * result = tmp - .5
 766        */
 767       half = get_half(p);
 768       tmp = get_temp( p );
 769       emit_arith( p, OPCODE_ADD, tmp, mask, 0, src[0], src[1], undef );
 770       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp, half, undef );
 771       return dest;
 772    case MODE_INTERPOLATE:
 773       /* Arg0 * (Arg2) + Arg1 * (1-Arg2) -- note arguments are reordered:
 774        */
 775       return emit_arith( p, OPCODE_LRP, dest, mask, saturate, src[2], src[0], src[1] );
 776
 777    case MODE_SUBTRACT:
 778       return emit_arith( p, OPCODE_SUB, dest, mask, saturate, src[0], src[1], undef );
 779
 780    case MODE_DOT3_RGBA:
 781    case MODE_DOT3_RGBA_EXT:
 782    case MODE_DOT3_RGB_EXT:
 783    case MODE_DOT3_RGB: {
 784       struct ureg tmp0 = get_temp( p );
 785       struct ureg tmp1 = get_temp( p );
 786       struct ureg neg1 = register_scalar_const(p, -1);
 787       struct ureg two  = register_scalar_const(p, 2);
 788
 789       /* tmp0 = 2*src0 - 1
 790        * tmp1 = 2*src1 - 1
 791        *
 792        * dst = tmp0 dot3 tmp1
 793        */
 794       emit_arith( p, OPCODE_MAD, tmp0, WRITEMASK_XYZW, 0,
 795                   two, src[0], neg1);
 796
 797       if (_mesa_memcmp(&src[0], &src[1], sizeof(struct ureg)) == 0)
 798          tmp1 = tmp0;
 799       else
 800          emit_arith( p, OPCODE_MAD, tmp1, WRITEMASK_XYZW, 0,
 801                      two, src[1], neg1);
 802       emit_arith( p, OPCODE_DP3, dest, mask, saturate, tmp0, tmp1, undef);
 803       return dest;
 804    }
 805    case MODE_MODULATE_ADD_ATI:
 806       /* Arg0 * Arg2 + Arg1 */
 807       return emit_arith( p, OPCODE_MAD, dest, mask, saturate,
 808                          src[0], src[2], src[1] );
 809    case MODE_MODULATE_SIGNED_ADD_ATI: {
 810       /* Arg0 * Arg2 + Arg1 - 0.5 */
 811       struct ureg tmp0 = get_temp(p);
 812       half = get_half(p);
 813       emit_arith( p, OPCODE_MAD, tmp0, mask, 0, src[0], src[2], src[1] );
 814       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp0, half, undef );
 815       return dest;
 816    }
 817    case MODE_MODULATE_SUBTRACT_ATI:
 818       /* Arg0 * Arg2 - Arg1 */
 819       emit_arith( p, OPCODE_MAD, dest, mask, 0, src[0], src[2], negate(src[1]) );
 820       return dest;
 821    default:
 822       return src[0];
 823    }
 824 }
 825
 826
 827 /**
 828  * Generate instructions for one texture unit's env/combiner mode.
 829  */
 830 static struct ureg
 831 emit_texenv(struct texenv_fragment_program *p, GLuint unit)
 832 {
 833    struct state_key *key = p->state;
 834    GLboolean saturate = (unit < p->last_tex_stage);
 835    GLuint rgb_shift, alpha_shift;
 836    struct ureg out, shift;
 837    struct ureg dest;
 838
 839    if (!key->unit[unit].enabled) {
 840       return get_source(p, SRC_PREVIOUS, 0);
 841    }
 842
 843    switch (key->unit[unit].ModeRGB) {
 844    case MODE_DOT3_RGB_EXT:
 845       alpha_shift = key->unit[unit].ScaleShiftA;
 846       rgb_shift = 0;
 847       break;
 848    case MODE_DOT3_RGBA_EXT:
 849       alpha_shift = 0;
 850       rgb_shift = 0;
 851       break;
 852    default:
 853       rgb_shift = key->unit[unit].ScaleShiftRGB;
 854       alpha_shift = key->unit[unit].ScaleShiftA;
 855       break;
 856    }
 857
 858    /* If this is the very last calculation, emit direct to output reg:
 859     */
 860    if (key->separate_specular ||
 861        unit != p->last_tex_stage ||
 862        alpha_shift ||
 863        rgb_shift)
 864       dest = get_temp( p );
 865    else
 866       dest = make_ureg(PROGRAM_OUTPUT, FRAG_RESULT_COLR);
 867
 868    /* Emit the RGB and A combine ops
 869     */
 870    if (key->unit[unit].ModeRGB == key->unit[unit].ModeA &&
 871        args_match(key, unit)) {
 872       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 873                           unit,
 874                           key->unit[unit].NumArgsRGB,
 875                           key->unit[unit].ModeRGB,
 876                           key->unit[unit].OptRGB);
 877    }
 878    else if (key->unit[unit].ModeRGB == MODE_DOT3_RGBA_EXT ||
 879             key->unit[unit].ModeRGB == MODE_DOT3_RGBA) {
 880
 881       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 882                           unit,
 883                           key->unit[unit].NumArgsRGB,
 884                           key->unit[unit].ModeRGB,
 885                           key->unit[unit].OptRGB);
 886    }
 887    else {
 888       /* Need to do something to stop from re-emitting identical
 889        * argument calculations here:
 890        */
 891       out = emit_combine( p, dest, WRITEMASK_XYZ, saturate,
 892                           unit,
 893                           key->unit[unit].NumArgsRGB,
 894                           key->unit[unit].ModeRGB,
 895                           key->unit[unit].OptRGB);
 896       out = emit_combine( p, dest, WRITEMASK_W, saturate,
 897                           unit,
 898                           key->unit[unit].NumArgsA,
 899                           key->unit[unit].ModeA,
 900                           key->unit[unit].OptA);
 901    }
 902
 903    /* Deal with the final shift:
 904     */
 905    if (alpha_shift || rgb_shift) {
 906       if (rgb_shift == alpha_shift) {
 907          shift = register_scalar_const(p, (GLfloat)(1<<rgb_shift));
 908       }
 909       else {
 910          shift = register_const4f(p,
 911                                   (GLfloat)(1<<rgb_shift),
 912                                   (GLfloat)(1<<rgb_shift),
 913                                   (GLfloat)(1<<rgb_shift),
 914                                   (GLfloat)(1<<alpha_shift));
 915       }
 916       return emit_arith( p, OPCODE_MUL, dest, WRITEMASK_XYZW,
 917                          saturate, out, shift, undef );
 918    }
 919    else
 920       return out;
 921 }
 922
 923
 924 /**
 925  * Generate instruction for getting a texture source term.
 926  */
 927 static void load_texture( struct texenv_fragment_program *p, GLuint unit )
 928 {
 929    if (is_undef(p->src_texture[unit])) {
 930       GLuint dim = p->state->unit[unit].source_index;
 931       struct ureg texcoord = register_input(p, FRAG_ATTRIB_TEX0+unit);
 932       struct ureg tmp = get_tex_temp( p );
 933
 934       if (dim == TEXTURE_UNKNOWN_INDEX)
 935          program_error(p, "TexSrcBit");
 936
 937       /* TODO: Use D0_MASK_XY where possible.
 938        */
 939       if (p->state->unit[unit].enabled) {
 940          p->src_texture[unit] = emit_texld( p, OPCODE_TXP,
 941                                             tmp, WRITEMASK_XYZW,
 942                                             unit, dim, texcoord );
 943          p->program->Base.SamplersUsed |= (1 << unit);
 944          /* This identity mapping should already be in place
 945           * (see _mesa_init_program_struct()) but let's be safe.
 946           */
 947          p->program->Base.SamplerUnits[unit] = unit;
 948       }
 949       else
 950          p->src_texture[unit] = get_zero(p);
 951    }
 952 }
 953
 954 static GLboolean load_texenv_source( struct texenv_fragment_program *p,
 955                                      GLuint src, GLuint unit )
 956 {
 957    switch (src) {
 958    case SRC_TEXTURE:
 959       load_texture(p, unit);
 960       break;
 961
 962    case SRC_TEXTURE0:
 963    case SRC_TEXTURE1:
 964    case SRC_TEXTURE2:
 965    case SRC_TEXTURE3:
 966    case SRC_TEXTURE4:
 967    case SRC_TEXTURE5:
 968    case SRC_TEXTURE6:
 969    case SRC_TEXTURE7:
 970       load_texture(p, src - SRC_TEXTURE0);
 971       break;
 972
 973    default:
 974       break;
 975    }
 976
 977    return GL_TRUE;
 978 }
 979
 980
 981 /**
 982  * Generate instructions for loading all texture source terms.
 983  */
 984 static GLboolean
 985 load_texunit_sources( struct texenv_fragment_program *p, int unit )
 986 {
 987    struct state_key *key = p->state;
 988    GLuint i;
 989
 990    for (i = 0; i < key->unit[unit].NumArgsRGB; i++) {
 991       load_texenv_source( p, key->unit[unit].OptRGB[i].Source, unit);
 992    }
 993
 994    for (i = 0; i < key->unit[unit].NumArgsA; i++) {
 995       load_texenv_source( p, key->unit[unit].OptA[i].Source, unit );
 996    }
 997
 998    return GL_TRUE;
 999 }
1000
1001
1002 /**
1003  * Generate a new fragment program which implements the context's
1004  * current texture env/combine mode.
1005  */
1006 static void
1007 create_new_program(GLcontext *ctx, struct state_key *key,
1008                    struct gl_fragment_program *program)
1009 {
1010    struct prog_instruction instBuffer[MAX_INSTRUCTIONS];
1011    struct texenv_fragment_program p;
1012    GLuint unit;
1013    struct ureg cf, out;
1014
1015    _mesa_memset(&p, 0, sizeof(p));
1016    p.ctx = ctx;
1017    p.state = key;
1018    p.program = program;
1019
1020    /* During code generation, use locally-allocated instruction buffer,
1021     * then alloc dynamic storage below.
1022     */
1023    p.program->Base.Instructions = instBuffer;
1024    p.program->Base.Target = GL_FRAGMENT_PROGRAM_ARB;
1025    p.program->Base.NumTexIndirections = 1;      /* correct? */
1026    p.program->Base.NumTexInstructions = 0;
1027    p.program->Base.NumAluInstructions = 0;
1028    p.program->Base.String = NULL;
1029    p.program->Base.NumInstructions =
1030    p.program->Base.NumTemporaries =
1031    p.program->Base.NumParameters =
1032    p.program->Base.NumAttributes = p.program->Base.NumAddressRegs = 0;
1033    p.program->Base.Parameters = _mesa_new_parameter_list();
1034
1035    p.program->Base.InputsRead = 0;
1036    p.program->Base.OutputsWritten = 1 << FRAG_RESULT_COLR;
1037
1038    for (unit = 0; unit < MAX_TEXTURE_UNITS; unit++)
1039       p.src_texture[unit] = undef;
1040
1041    p.src_previous = undef;
1042    p.half = undef;
1043    p.zero = undef;
1044    p.one = undef;
1045
1046    p.last_tex_stage = 0;
1047    release_temps(ctx, &p);
1048
1049    if (key->enabled_units) {
1050       /* First pass - to support texture_env_crossbar, first identify
1051        * all referenced texture sources and emit texld instructions
1052        * for each:
1053        */
1054       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits ; unit++)
1055          if (key->unit[unit].enabled) {
1056             load_texunit_sources( &p, unit );
1057             p.last_tex_stage = unit;
1058          }
1059
1060       /* Second pass - emit combine instructions to build final color:
1061        */
1062       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits; unit++)
1063          if (key->enabled_units & (1<<unit)) {
1064             p.src_previous = emit_texenv( &p, unit );
1065             release_temps(ctx, &p);     /* release all temps */
1066          }
1067    }
1068
1069    cf = get_source( &p, SRC_PREVIOUS, 0 );
1070    out = make_ureg( PROGRAM_OUTPUT, FRAG_RESULT_COLR );
1071
1072    if (key->separate_specular) {
1073       /* Emit specular add.
1074        */
1075       struct ureg s = register_input(&p, FRAG_ATTRIB_COL1);
1076       emit_arith( &p, OPCODE_ADD, out, WRITEMASK_XYZ, 0, cf, s, undef );
1077       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_W, 0, cf, undef, undef );
1078    }
1079    else if (_mesa_memcmp(&cf, &out, sizeof(cf)) != 0) {
1080       /* Will wind up in here if no texture enabled or a couple of
1081        * other scenarios (GL_REPLACE for instance).
1082        */
1083       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_XYZW, 0, cf, undef, undef );
1084    }
1085
1086    /* Finish up:
1087     */
1088    emit_arith( &p, OPCODE_END, undef, WRITEMASK_XYZW, 0, undef, undef, undef);
1089
1090    if (key->fog_enabled) {
1091       /* Pull fog mode from GLcontext, the value in the state key is
1092        * a reduced value and not what is expected in FogOption
1093        */
1094       p.program->FogOption = ctx->Fog.Mode;
1095       p.program->Base.InputsRead |= FRAG_BIT_FOGC; /* XXX new */
1096    } else
1097       p.program->FogOption = GL_NONE;
1098
1099    if (p.program->Base.NumTexIndirections > ctx->Const.FragmentProgram.MaxTexIndirections)
1100       program_error(&p, "Exceeded max nr indirect texture lookups");
1101
1102    if (p.program->Base.NumTexInstructions > ctx->Const.FragmentProgram.MaxTexInstructions)
1103       program_error(&p, "Exceeded max TEX instructions");
1104
1105    if (p.program->Base.NumAluInstructions > ctx->Const.FragmentProgram.MaxAluInstructions)
1106       program_error(&p, "Exceeded max ALU instructions");
1107
1108    ASSERT(p.program->Base.NumInstructions <= MAX_INSTRUCTIONS);
1109
1110    /* Allocate final instruction array */
1111    p.program->Base.Instructions
1112       = _mesa_alloc_instructions(p.program->Base.NumInstructions);
1113    if (!p.program->Base.Instructions) {
1114       _mesa_error(ctx, GL_OUT_OF_MEMORY,
1115                   "generating tex env program");
1116       return;
1117    }
1118    _mesa_copy_instructions(p.program->Base.Instructions, instBuffer,
1119                            p.program->Base.NumInstructions);
1120
1121    if (p.program->FogOption) {
1122       _mesa_append_fog_code(ctx, p.program);
1123       p.program->FogOption = GL_NONE;
1124    }
1125
1126
1127    /* Notify driver the fragment program has (actually) changed.
1128     */
1129    if (ctx->Driver.ProgramStringNotify) {
1130       ctx->Driver.ProgramStringNotify( ctx, GL_FRAGMENT_PROGRAM_ARB,
1131                                        &p.program->Base );
1132    }
1133
1134    if (DISASSEM) {
1135       _mesa_print_program(&p.program->Base);
1136       _mesa_printf("\n");
1137    }
1138 }
1139
1140
1141 /**
1142  * Return a fragment program which implements the current
1143  * fixed-function texture, fog and color-sum operations.
1144  */
1145 struct gl_fragment_program *
1146 _mesa_get_fixed_func_fragment_program(GLcontext *ctx)
1147 {
1148    struct gl_fragment_program *prog;
1149    struct state_key key;
1150
1151    make_state_key(ctx, &key);
1152
1153    prog = (struct gl_fragment_program *)
1154       _mesa_search_program_cache(ctx->FragmentProgram.Cache,
1155                                  &key, sizeof(key));
1156
1157    if (!prog) {
1158       prog = (struct gl_fragment_program *)
1159          ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
1160
1161       create_new_program(ctx, &key, prog);
1162
1163       _mesa_program_cache_insert(ctx, ctx->FragmentProgram.Cache,
1164                                  &key, sizeof(key), &prog->Base);
1165    }
1166
1167    return prog;
1168 }
1169
1170
1171
1172 /**
1173  * If _MaintainTexEnvProgram is set we'll generate a fragment program that
1174  * implements the current texture env/combine mode.
1175  * This function generates that program and puts it into effect.
1176  */
1177 void
1178 _mesa_UpdateTexEnvProgram( GLcontext *ctx )
1179 {
1180    const struct gl_fragment_program *prev = ctx->FragmentProgram._Current;
1181
1182    ASSERT(ctx->FragmentProgram._MaintainTexEnvProgram);
1183
1184    /* If a conventional fragment program/shader isn't in effect... */
1185    if (!ctx->FragmentProgram._Enabled &&
1186        (!ctx->Shader.CurrentProgram ||
1187         !ctx->Shader.CurrentProgram->FragmentProgram) ) {
1188
1189       ctx->FragmentProgram._Current
1190          = ctx->FragmentProgram._TexEnvProgram
1191          = _mesa_get_fixed_func_fragment_program(ctx);
1192    }
1193
1194    /* Tell the driver about the change.  Could define a new target for
1195     * this?
1196     */
1197    if (ctx->FragmentProgram._Current != prev && ctx->Driver.BindProgram) {
1198       ctx->Driver.BindProgram(ctx, GL_FRAGMENT_PROGRAM_ARB,
1199                          (struct gl_program *) ctx->FragmentProgram._Current);
1200    }
1201 }