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  * Up to nine instructions per tex unit, plus fog, specular color.
  41  */
  42 #define MAX_INSTRUCTIONS ((MAX_TEXTURE_UNITS * 9) + 12)
  43
  44 #define DISASSEM (MESA_VERBOSE & VERBOSE_DISASSEM)
  45
  46 struct mode_opt {
  47    GLubyte Source:4;
  48    GLubyte Operand:3;
  49 };
  50
  51 struct state_key {
  52    GLuint nr_enabled_units:8;
  53    GLuint enabled_units:8;
  54    GLuint separate_specular:1;
  55    GLuint fog_enabled:1;
  56    GLuint fog_mode:2;
  57    GLuint inputs_available:12;
  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       GLuint NumArgsA:2;
  68       GLuint ModeA:4;
  69
  70       struct mode_opt OptRGB[3];
  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 available 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 /** Mark a temp reg as being no longer allocatable. */
 404 static void reserve_temp( struct texenv_fragment_program *p, struct ureg r )
 405 {
 406    if (r.file == PROGRAM_TEMPORARY)
 407       p->temps_output |= (1 << r.idx);
 408 }
 409
 410
 411 static void release_temps(GLcontext *ctx, struct texenv_fragment_program *p )
 412 {
 413    GLuint max_temp = ctx->Const.FragmentProgram.MaxTemps;
 414
 415    /* KW: To support tex_env_crossbar, don't release the registers in
 416     * temps_output.
 417     */
 418    if (max_temp >= sizeof(int) * 8)
 419       p->temp_in_use = p->temps_output;
 420    else
 421       p->temp_in_use = ~((1<<max_temp)-1) | p->temps_output;
 422 }
 423
 424
 425 static struct ureg register_param5( struct texenv_fragment_program *p,
 426                                     GLint s0,
 427                                     GLint s1,
 428                                     GLint s2,
 429                                     GLint s3,
 430                                     GLint s4)
 431 {
 432    gl_state_index tokens[STATE_LENGTH];
 433    GLuint idx;
 434    tokens[0] = s0;
 435    tokens[1] = s1;
 436    tokens[2] = s2;
 437    tokens[3] = s3;
 438    tokens[4] = s4;
 439    idx = _mesa_add_state_reference( p->program->Base.Parameters, tokens );
 440    return make_ureg(PROGRAM_STATE_VAR, idx);
 441 }
 442
 443
 444 #define register_param1(p,s0)          register_param5(p,s0,0,0,0,0)
 445 #define register_param2(p,s0,s1)       register_param5(p,s0,s1,0,0,0)
 446 #define register_param3(p,s0,s1,s2)    register_param5(p,s0,s1,s2,0,0)
 447 #define register_param4(p,s0,s1,s2,s3) register_param5(p,s0,s1,s2,s3,0)
 448
 449
 450 static struct ureg register_input( struct texenv_fragment_program *p, GLuint input )
 451 {
 452    p->program->Base.InputsRead |= (1 << input);
 453    return make_ureg(PROGRAM_INPUT, input);
 454 }
 455
 456
 457 static void emit_arg( struct prog_src_register *reg,
 458                       struct ureg ureg )
 459 {
 460    reg->File = ureg.file;
 461    reg->Index = ureg.idx;
 462    reg->Swizzle = ureg.swz;
 463    reg->NegateBase = ureg.negatebase ? 0xf : 0x0;
 464    reg->Abs = ureg.abs;
 465    reg->NegateAbs = ureg.negateabs;
 466 }
 467
 468 static void emit_dst( struct prog_dst_register *dst,
 469                       struct ureg ureg, GLuint mask )
 470 {
 471    dst->File = ureg.file;
 472    dst->Index = ureg.idx;
 473    dst->WriteMask = mask;
 474    dst->CondMask = COND_TR;  /* always pass cond test */
 475    dst->CondSwizzle = SWIZZLE_NOOP;
 476 }
 477
 478 static struct prog_instruction *
 479 emit_op(struct texenv_fragment_program *p,
 480         enum prog_opcode op,
 481         struct ureg dest,
 482         GLuint mask,
 483         GLboolean saturate,
 484         struct ureg src0,
 485         struct ureg src1,
 486         struct ureg src2 )
 487 {
 488    GLuint nr = p->program->Base.NumInstructions++;
 489    struct prog_instruction *inst = &p->program->Base.Instructions[nr];
 490
 491    assert(nr < MAX_INSTRUCTIONS);
 492
 493    _mesa_init_instructions(inst, 1);
 494    inst->Opcode = op;
 495
 496    emit_arg( &inst->SrcReg[0], src0 );
 497    emit_arg( &inst->SrcReg[1], src1 );
 498    emit_arg( &inst->SrcReg[2], src2 );
 499
 500    inst->SaturateMode = saturate ? SATURATE_ZERO_ONE : SATURATE_OFF;
 501
 502    emit_dst( &inst->DstReg, dest, mask );
 503
 504 #if 0
 505    /* Accounting for indirection tracking:
 506     */
 507    if (dest.file == PROGRAM_TEMPORARY)
 508       p->temps_output |= 1 << dest.idx;
 509 #endif
 510
 511    return inst;
 512 }
 513
 514
 515 static struct ureg emit_arith( struct texenv_fragment_program *p,
 516                                enum prog_opcode op,
 517                                struct ureg dest,
 518                                GLuint mask,
 519                                GLboolean saturate,
 520                                struct ureg src0,
 521                                struct ureg src1,
 522                                struct ureg src2 )
 523 {
 524    emit_op(p, op, dest, mask, saturate, src0, src1, src2);
 525
 526    /* Accounting for indirection tracking:
 527     */
 528    if (src0.file == PROGRAM_TEMPORARY)
 529       p->alu_temps |= 1 << src0.idx;
 530
 531    if (!is_undef(src1) && src1.file == PROGRAM_TEMPORARY)
 532       p->alu_temps |= 1 << src1.idx;
 533
 534    if (!is_undef(src2) && src2.file == PROGRAM_TEMPORARY)
 535       p->alu_temps |= 1 << src2.idx;
 536
 537    if (dest.file == PROGRAM_TEMPORARY)
 538       p->alu_temps |= 1 << dest.idx;
 539
 540    p->program->Base.NumAluInstructions++;
 541    return dest;
 542 }
 543
 544 static struct ureg emit_texld( struct texenv_fragment_program *p,
 545                                enum prog_opcode op,
 546                                struct ureg dest,
 547                                GLuint destmask,
 548                                GLuint tex_unit,
 549                                GLuint tex_idx,
 550                                struct ureg coord )
 551 {
 552    struct prog_instruction *inst = emit_op( p, op,
 553                                           dest, destmask,
 554                                           GL_FALSE,     /* don't saturate? */
 555                                           coord,        /* arg 0? */
 556                                           undef,
 557                                           undef);
 558
 559    inst->TexSrcTarget = tex_idx;
 560    inst->TexSrcUnit = tex_unit;
 561
 562    p->program->Base.NumTexInstructions++;
 563
 564    /* Accounting for indirection tracking:
 565     */
 566    reserve_temp(p, dest);
 567
 568    /* Is this a texture indirection?
 569     */
 570    if ((coord.file == PROGRAM_TEMPORARY &&
 571         (p->temps_output & (1<<coord.idx))) ||
 572        (dest.file == PROGRAM_TEMPORARY &&
 573         (p->alu_temps & (1<<dest.idx)))) {
 574       p->program->Base.NumTexIndirections++;
 575       p->temps_output = 1<<coord.idx;
 576       p->alu_temps = 0;
 577       assert(0);                /* KW: texture env crossbar */
 578    }
 579
 580    return dest;
 581 }
 582
 583
 584 static struct ureg register_const4f( struct texenv_fragment_program *p,
 585                                      GLfloat s0,
 586                                      GLfloat s1,
 587                                      GLfloat s2,
 588                                      GLfloat s3)
 589 {
 590    GLfloat values[4];
 591    GLuint idx, swizzle;
 592    struct ureg r;
 593    values[0] = s0;
 594    values[1] = s1;
 595    values[2] = s2;
 596    values[3] = s3;
 597    idx = _mesa_add_unnamed_constant( p->program->Base.Parameters, values, 4,
 598                                      &swizzle );
 599    r = make_ureg(PROGRAM_CONSTANT, idx);
 600    r.swz = swizzle;
 601    return r;
 602 }
 603
 604 #define register_scalar_const(p, s0)    register_const4f(p, s0, s0, s0, s0)
 605 #define register_const1f(p, s0)         register_const4f(p, s0, 0, 0, 1)
 606 #define register_const2f(p, s0, s1)     register_const4f(p, s0, s1, 0, 1)
 607 #define register_const3f(p, s0, s1, s2) register_const4f(p, s0, s1, s2, 1)
 608
 609
 610 static struct ureg get_one( struct texenv_fragment_program *p )
 611 {
 612    if (is_undef(p->one))
 613       p->one = register_scalar_const(p, 1.0);
 614    return p->one;
 615 }
 616
 617 static struct ureg get_half( struct texenv_fragment_program *p )
 618 {
 619    if (is_undef(p->half))
 620       p->half = register_scalar_const(p, 0.5);
 621    return p->half;
 622 }
 623
 624 static struct ureg get_zero( struct texenv_fragment_program *p )
 625 {
 626    if (is_undef(p->zero))
 627       p->zero = register_scalar_const(p, 0.0);
 628    return p->zero;
 629 }
 630
 631
 632 static void program_error( struct texenv_fragment_program *p, const char *msg )
 633 {
 634    _mesa_problem(NULL, msg);
 635    p->error = 1;
 636 }
 637
 638 static struct ureg get_source( struct texenv_fragment_program *p,
 639                                GLuint src, GLuint unit )
 640 {
 641    switch (src) {
 642    case SRC_TEXTURE:
 643       assert(!is_undef(p->src_texture[unit]));
 644       return p->src_texture[unit];
 645
 646    case SRC_TEXTURE0:
 647    case SRC_TEXTURE1:
 648    case SRC_TEXTURE2:
 649    case SRC_TEXTURE3:
 650    case SRC_TEXTURE4:
 651    case SRC_TEXTURE5:
 652    case SRC_TEXTURE6:
 653    case SRC_TEXTURE7:
 654       assert(!is_undef(p->src_texture[src - SRC_TEXTURE0]));
 655       return p->src_texture[src - SRC_TEXTURE0];
 656
 657    case SRC_CONSTANT:
 658       return register_param2(p, STATE_TEXENV_COLOR, unit);
 659
 660    case SRC_PRIMARY_COLOR:
 661       return register_input(p, FRAG_ATTRIB_COL0);
 662
 663    case SRC_PREVIOUS:
 664    default:
 665       if (is_undef(p->src_previous))
 666          return register_input(p, FRAG_ATTRIB_COL0);
 667       else
 668          return p->src_previous;
 669    }
 670 }
 671
 672 static struct ureg emit_combine_source( struct texenv_fragment_program *p,
 673                                         GLuint mask,
 674                                         GLuint unit,
 675                                         GLuint source,
 676                                         GLuint operand )
 677 {
 678    struct ureg arg, src, one;
 679
 680    src = get_source(p, source, unit);
 681
 682    switch (operand) {
 683    case OPR_ONE_MINUS_SRC_COLOR:
 684       /* Get unused tmp,
 685        * Emit tmp = 1.0 - arg.xyzw
 686        */
 687       arg = get_temp( p );
 688       one = get_one( p );
 689       return emit_arith( p, OPCODE_SUB, arg, mask, 0, one, src, undef);
 690
 691    case OPR_SRC_ALPHA:
 692       if (mask == WRITEMASK_W)
 693          return src;
 694       else
 695          return swizzle1( src, SWIZZLE_W );
 696    case OPR_ONE_MINUS_SRC_ALPHA:
 697       /* Get unused tmp,
 698        * Emit tmp = 1.0 - arg.wwww
 699        */
 700       arg = get_temp(p);
 701       one = get_one(p);
 702       return emit_arith(p, OPCODE_SUB, arg, mask, 0,
 703                         one, swizzle1(src, SWIZZLE_W), undef);
 704    case OPR_ZERO:
 705       return get_zero(p);
 706    case OPR_ONE:
 707       return get_one(p);
 708    case OPR_SRC_COLOR:
 709    default:
 710       return src;
 711    }
 712 }
 713
 714 static GLboolean args_match( struct state_key *key, GLuint unit )
 715 {
 716    GLuint i, nr = key->unit[unit].NumArgsRGB;
 717
 718    for (i = 0 ; i < nr ; i++) {
 719       if (key->unit[unit].OptA[i].Source != key->unit[unit].OptRGB[i].Source)
 720          return GL_FALSE;
 721
 722       switch(key->unit[unit].OptA[i].Operand) {
 723       case OPR_SRC_ALPHA:
 724          switch(key->unit[unit].OptRGB[i].Operand) {
 725          case OPR_SRC_COLOR:
 726          case OPR_SRC_ALPHA:
 727             break;
 728          default:
 729             return GL_FALSE;
 730          }
 731          break;
 732       case OPR_ONE_MINUS_SRC_ALPHA:
 733          switch(key->unit[unit].OptRGB[i].Operand) {
 734          case OPR_ONE_MINUS_SRC_COLOR:
 735          case OPR_ONE_MINUS_SRC_ALPHA:
 736             break;
 737          default:
 738             return GL_FALSE;
 739          }
 740          break;
 741       default:
 742          return GL_FALSE;       /* impossible */
 743       }
 744    }
 745
 746    return GL_TRUE;
 747 }
 748
 749 static struct ureg emit_combine( struct texenv_fragment_program *p,
 750                                  struct ureg dest,
 751                                  GLuint mask,
 752                                  GLboolean saturate,
 753                                  GLuint unit,
 754                                  GLuint nr,
 755                                  GLuint mode,
 756                                  const struct mode_opt *opt)
 757 {
 758    struct ureg src[3];
 759    struct ureg tmp, half;
 760    GLuint i;
 761
 762    tmp = undef; /* silence warning (bug 5318) */
 763
 764    for (i = 0; i < nr; i++)
 765       src[i] = emit_combine_source( p, mask, unit, opt[i].Source, opt[i].Operand );
 766
 767    switch (mode) {
 768    case MODE_REPLACE:
 769       if (mask == WRITEMASK_XYZW && !saturate)
 770          return src[0];
 771       else
 772          return emit_arith( p, OPCODE_MOV, dest, mask, saturate, src[0], undef, undef );
 773    case MODE_MODULATE:
 774       return emit_arith( p, OPCODE_MUL, dest, mask, saturate,
 775                          src[0], src[1], undef );
 776    case MODE_ADD:
 777       return emit_arith( p, OPCODE_ADD, dest, mask, saturate,
 778                          src[0], src[1], undef );
 779    case MODE_ADD_SIGNED:
 780       /* tmp = arg0 + arg1
 781        * result = tmp - .5
 782        */
 783       half = get_half(p);
 784       tmp = get_temp( p );
 785       emit_arith( p, OPCODE_ADD, tmp, mask, 0, src[0], src[1], undef );
 786       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp, half, undef );
 787       return dest;
 788    case MODE_INTERPOLATE:
 789       /* Arg0 * (Arg2) + Arg1 * (1-Arg2) -- note arguments are reordered:
 790        */
 791       return emit_arith( p, OPCODE_LRP, dest, mask, saturate, src[2], src[0], src[1] );
 792
 793    case MODE_SUBTRACT:
 794       return emit_arith( p, OPCODE_SUB, dest, mask, saturate, src[0], src[1], undef );
 795
 796    case MODE_DOT3_RGBA:
 797    case MODE_DOT3_RGBA_EXT:
 798    case MODE_DOT3_RGB_EXT:
 799    case MODE_DOT3_RGB: {
 800       struct ureg tmp0 = get_temp( p );
 801       struct ureg tmp1 = get_temp( p );
 802       struct ureg neg1 = register_scalar_const(p, -1);
 803       struct ureg two  = register_scalar_const(p, 2);
 804
 805       /* tmp0 = 2*src0 - 1
 806        * tmp1 = 2*src1 - 1
 807        *
 808        * dst = tmp0 dot3 tmp1
 809        */
 810       emit_arith( p, OPCODE_MAD, tmp0, WRITEMASK_XYZW, 0,
 811                   two, src[0], neg1);
 812
 813       if (_mesa_memcmp(&src[0], &src[1], sizeof(struct ureg)) == 0)
 814          tmp1 = tmp0;
 815       else
 816          emit_arith( p, OPCODE_MAD, tmp1, WRITEMASK_XYZW, 0,
 817                      two, src[1], neg1);
 818       emit_arith( p, OPCODE_DP3, dest, mask, saturate, tmp0, tmp1, undef);
 819       return dest;
 820    }
 821    case MODE_MODULATE_ADD_ATI:
 822       /* Arg0 * Arg2 + Arg1 */
 823       return emit_arith( p, OPCODE_MAD, dest, mask, saturate,
 824                          src[0], src[2], src[1] );
 825    case MODE_MODULATE_SIGNED_ADD_ATI: {
 826       /* Arg0 * Arg2 + Arg1 - 0.5 */
 827       struct ureg tmp0 = get_temp(p);
 828       half = get_half(p);
 829       emit_arith( p, OPCODE_MAD, tmp0, mask, 0, src[0], src[2], src[1] );
 830       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp0, half, undef );
 831       return dest;
 832    }
 833    case MODE_MODULATE_SUBTRACT_ATI:
 834       /* Arg0 * Arg2 - Arg1 */
 835       emit_arith( p, OPCODE_MAD, dest, mask, 0, src[0], src[2], negate(src[1]) );
 836       return dest;
 837    default:
 838       return src[0];
 839    }
 840 }
 841
 842
 843 /**
 844  * Generate instructions for one texture unit's env/combiner mode.
 845  */
 846 static struct ureg
 847 emit_texenv(struct texenv_fragment_program *p, GLuint unit)
 848 {
 849    struct state_key *key = p->state;
 850    GLboolean saturate = (unit < p->last_tex_stage);
 851    GLuint rgb_shift, alpha_shift;
 852    struct ureg out, shift;
 853    struct ureg dest;
 854
 855    if (!key->unit[unit].enabled) {
 856       return get_source(p, SRC_PREVIOUS, 0);
 857    }
 858
 859    switch (key->unit[unit].ModeRGB) {
 860    case MODE_DOT3_RGB_EXT:
 861       alpha_shift = key->unit[unit].ScaleShiftA;
 862       rgb_shift = 0;
 863       break;
 864    case MODE_DOT3_RGBA_EXT:
 865       alpha_shift = 0;
 866       rgb_shift = 0;
 867       break;
 868    default:
 869       rgb_shift = key->unit[unit].ScaleShiftRGB;
 870       alpha_shift = key->unit[unit].ScaleShiftA;
 871       break;
 872    }
 873
 874    /* If this is the very last calculation, emit direct to output reg:
 875     */
 876    if (key->separate_specular ||
 877        unit != p->last_tex_stage ||
 878        alpha_shift ||
 879        rgb_shift)
 880       dest = get_temp( p );
 881    else
 882       dest = make_ureg(PROGRAM_OUTPUT, FRAG_RESULT_COLR);
 883
 884    /* Emit the RGB and A combine ops
 885     */
 886    if (key->unit[unit].ModeRGB == key->unit[unit].ModeA &&
 887        args_match(key, unit)) {
 888       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 889                           unit,
 890                           key->unit[unit].NumArgsRGB,
 891                           key->unit[unit].ModeRGB,
 892                           key->unit[unit].OptRGB);
 893    }
 894    else if (key->unit[unit].ModeRGB == MODE_DOT3_RGBA_EXT ||
 895             key->unit[unit].ModeRGB == MODE_DOT3_RGBA) {
 896
 897       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 898                           unit,
 899                           key->unit[unit].NumArgsRGB,
 900                           key->unit[unit].ModeRGB,
 901                           key->unit[unit].OptRGB);
 902    }
 903    else {
 904       /* Need to do something to stop from re-emitting identical
 905        * argument calculations here:
 906        */
 907       out = emit_combine( p, dest, WRITEMASK_XYZ, saturate,
 908                           unit,
 909                           key->unit[unit].NumArgsRGB,
 910                           key->unit[unit].ModeRGB,
 911                           key->unit[unit].OptRGB);
 912       out = emit_combine( p, dest, WRITEMASK_W, saturate,
 913                           unit,
 914                           key->unit[unit].NumArgsA,
 915                           key->unit[unit].ModeA,
 916                           key->unit[unit].OptA);
 917    }
 918
 919    /* Deal with the final shift:
 920     */
 921    if (alpha_shift || rgb_shift) {
 922       if (rgb_shift == alpha_shift) {
 923          shift = register_scalar_const(p, (GLfloat)(1<<rgb_shift));
 924       }
 925       else {
 926          shift = register_const4f(p,
 927                                   (GLfloat)(1<<rgb_shift),
 928                                   (GLfloat)(1<<rgb_shift),
 929                                   (GLfloat)(1<<rgb_shift),
 930                                   (GLfloat)(1<<alpha_shift));
 931       }
 932       return emit_arith( p, OPCODE_MUL, dest, WRITEMASK_XYZW,
 933                          saturate, out, shift, undef );
 934    }
 935    else
 936       return out;
 937 }
 938
 939
 940 /**
 941  * Generate instruction for getting a texture source term.
 942  */
 943 static void load_texture( struct texenv_fragment_program *p, GLuint unit )
 944 {
 945    if (is_undef(p->src_texture[unit])) {
 946       GLuint dim = p->state->unit[unit].source_index;
 947       struct ureg texcoord = register_input(p, FRAG_ATTRIB_TEX0+unit);
 948       struct ureg tmp = get_tex_temp( p );
 949
 950       if (dim == TEXTURE_UNKNOWN_INDEX)
 951          program_error(p, "TexSrcBit");
 952
 953       /* TODO: Use D0_MASK_XY where possible.
 954        */
 955       if (p->state->unit[unit].enabled) {
 956          p->src_texture[unit] = emit_texld( p, OPCODE_TXP,
 957                                             tmp, WRITEMASK_XYZW,
 958                                             unit, dim, texcoord );
 959          p->program->Base.SamplersUsed |= (1 << unit);
 960          /* This identity mapping should already be in place
 961           * (see _mesa_init_program_struct()) but let's be safe.
 962           */
 963          p->program->Base.SamplerUnits[unit] = unit;
 964       }
 965       else
 966          p->src_texture[unit] = get_zero(p);
 967    }
 968 }
 969
 970 static GLboolean load_texenv_source( struct texenv_fragment_program *p,
 971                                      GLuint src, GLuint unit )
 972 {
 973    switch (src) {
 974    case SRC_TEXTURE:
 975       load_texture(p, unit);
 976       break;
 977
 978    case SRC_TEXTURE0:
 979    case SRC_TEXTURE1:
 980    case SRC_TEXTURE2:
 981    case SRC_TEXTURE3:
 982    case SRC_TEXTURE4:
 983    case SRC_TEXTURE5:
 984    case SRC_TEXTURE6:
 985    case SRC_TEXTURE7:
 986       load_texture(p, src - SRC_TEXTURE0);
 987       break;
 988
 989    default:
 990       break;
 991    }
 992
 993    return GL_TRUE;
 994 }
 995
 996
 997 /**
 998  * Generate instructions for loading all texture source terms.
 999  */
1000 static GLboolean
1001 load_texunit_sources( struct texenv_fragment_program *p, int unit )
1002 {
1003    struct state_key *key = p->state;
1004    GLuint i;
1005
1006    for (i = 0; i < key->unit[unit].NumArgsRGB; i++) {
1007       load_texenv_source( p, key->unit[unit].OptRGB[i].Source, unit);
1008    }
1009
1010    for (i = 0; i < key->unit[unit].NumArgsA; i++) {
1011       load_texenv_source( p, key->unit[unit].OptA[i].Source, unit );
1012    }
1013
1014    return GL_TRUE;
1015 }
1016
1017
1018 /**
1019  * Generate a new fragment program which implements the context's
1020  * current texture env/combine mode.
1021  */
1022 static void
1023 create_new_program(GLcontext *ctx, struct state_key *key,
1024                    struct gl_fragment_program *program)
1025 {
1026    struct prog_instruction instBuffer[MAX_INSTRUCTIONS];
1027    struct texenv_fragment_program p;
1028    GLuint unit;
1029    struct ureg cf, out;
1030
1031    _mesa_memset(&p, 0, sizeof(p));
1032    p.ctx = ctx;
1033    p.state = key;
1034    p.program = program;
1035
1036    /* During code generation, use locally-allocated instruction buffer,
1037     * then alloc dynamic storage below.
1038     */
1039    p.program->Base.Instructions = instBuffer;
1040    p.program->Base.Target = GL_FRAGMENT_PROGRAM_ARB;
1041    p.program->Base.NumTexIndirections = 1;      /* correct? */
1042    p.program->Base.NumTexInstructions = 0;
1043    p.program->Base.NumAluInstructions = 0;
1044    p.program->Base.String = NULL;
1045    p.program->Base.NumInstructions =
1046    p.program->Base.NumTemporaries =
1047    p.program->Base.NumParameters =
1048    p.program->Base.NumAttributes = p.program->Base.NumAddressRegs = 0;
1049    p.program->Base.Parameters = _mesa_new_parameter_list();
1050
1051    p.program->Base.InputsRead = 0;
1052    p.program->Base.OutputsWritten = 1 << FRAG_RESULT_COLR;
1053
1054    for (unit = 0; unit < MAX_TEXTURE_UNITS; unit++)
1055       p.src_texture[unit] = undef;
1056
1057    p.src_previous = undef;
1058    p.half = undef;
1059    p.zero = undef;
1060    p.one = undef;
1061
1062    p.last_tex_stage = 0;
1063    release_temps(ctx, &p);
1064
1065    if (key->enabled_units) {
1066       /* First pass - to support texture_env_crossbar, first identify
1067        * all referenced texture sources and emit texld instructions
1068        * for each:
1069        */
1070       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits ; unit++)
1071          if (key->unit[unit].enabled) {
1072             load_texunit_sources( &p, unit );
1073             p.last_tex_stage = unit;
1074          }
1075
1076       /* Second pass - emit combine instructions to build final color:
1077        */
1078       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits; unit++)
1079          if (key->enabled_units & (1<<unit)) {
1080             p.src_previous = emit_texenv( &p, unit );
1081             reserve_temp(&p, p.src_previous); /* don't re-use this temp reg */
1082             release_temps(ctx, &p);     /* release all temps */
1083          }
1084    }
1085
1086    cf = get_source( &p, SRC_PREVIOUS, 0 );
1087    out = make_ureg( PROGRAM_OUTPUT, FRAG_RESULT_COLR );
1088
1089    if (key->separate_specular) {
1090       /* Emit specular add.
1091        */
1092       struct ureg s = register_input(&p, FRAG_ATTRIB_COL1);
1093       emit_arith( &p, OPCODE_ADD, out, WRITEMASK_XYZ, 0, cf, s, undef );
1094       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_W, 0, cf, undef, undef );
1095    }
1096    else if (_mesa_memcmp(&cf, &out, sizeof(cf)) != 0) {
1097       /* Will wind up in here if no texture enabled or a couple of
1098        * other scenarios (GL_REPLACE for instance).
1099        */
1100       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_XYZW, 0, cf, undef, undef );
1101    }
1102
1103    /* Finish up:
1104     */
1105    emit_arith( &p, OPCODE_END, undef, WRITEMASK_XYZW, 0, undef, undef, undef);
1106
1107    if (key->fog_enabled) {
1108       /* Pull fog mode from GLcontext, the value in the state key is
1109        * a reduced value and not what is expected in FogOption
1110        */
1111       p.program->FogOption = ctx->Fog.Mode;
1112       p.program->Base.InputsRead |= FRAG_BIT_FOGC; /* XXX new */
1113    } else
1114       p.program->FogOption = GL_NONE;
1115
1116    if (p.program->Base.NumTexIndirections > ctx->Const.FragmentProgram.MaxTexIndirections)
1117       program_error(&p, "Exceeded max nr indirect texture lookups");
1118
1119    if (p.program->Base.NumTexInstructions > ctx->Const.FragmentProgram.MaxTexInstructions)
1120       program_error(&p, "Exceeded max TEX instructions");
1121
1122    if (p.program->Base.NumAluInstructions > ctx->Const.FragmentProgram.MaxAluInstructions)
1123       program_error(&p, "Exceeded max ALU instructions");
1124
1125    ASSERT(p.program->Base.NumInstructions <= MAX_INSTRUCTIONS);
1126
1127    /* Allocate final instruction array */
1128    p.program->Base.Instructions
1129       = _mesa_alloc_instructions(p.program->Base.NumInstructions);
1130    if (!p.program->Base.Instructions) {
1131       _mesa_error(ctx, GL_OUT_OF_MEMORY,
1132                   "generating tex env program");
1133       return;
1134    }
1135    _mesa_copy_instructions(p.program->Base.Instructions, instBuffer,
1136                            p.program->Base.NumInstructions);
1137
1138    if (p.program->FogOption) {
1139       _mesa_append_fog_code(ctx, p.program);
1140       p.program->FogOption = GL_NONE;
1141    }
1142
1143
1144    /* Notify driver the fragment program has (actually) changed.
1145     */
1146    if (ctx->Driver.ProgramStringNotify) {
1147       ctx->Driver.ProgramStringNotify( ctx, GL_FRAGMENT_PROGRAM_ARB,
1148                                        &p.program->Base );
1149    }
1150
1151    if (DISASSEM) {
1152       _mesa_print_program(&p.program->Base);
1153       _mesa_printf("\n");
1154    }
1155 }
1156
1157
1158 /**
1159  * Return a fragment program which implements the current
1160  * fixed-function texture, fog and color-sum operations.
1161  */
1162 struct gl_fragment_program *
1163 _mesa_get_fixed_func_fragment_program(GLcontext *ctx)
1164 {
1165    struct gl_fragment_program *prog;
1166    struct state_key key;
1167
1168    make_state_key(ctx, &key);
1169
1170    prog = (struct gl_fragment_program *)
1171       _mesa_search_program_cache(ctx->FragmentProgram.Cache,
1172                                  &key, sizeof(key));
1173
1174    if (!prog) {
1175       prog = (struct gl_fragment_program *)
1176          ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
1177
1178       create_new_program(ctx, &key, prog);
1179
1180       _mesa_program_cache_insert(ctx, ctx->FragmentProgram.Cache,
1181                                  &key, sizeof(key), &prog->Base);
1182    }
1183
1184    return prog;
1185 }
1186
1187
1188
1189 /**
1190  * If _MaintainTexEnvProgram is set we'll generate a fragment program that
1191  * implements the current texture env/combine mode.
1192  * This function generates that program and puts it into effect.
1193  */
1194 void
1195 _mesa_UpdateTexEnvProgram( GLcontext *ctx )
1196 {
1197    const struct gl_fragment_program *prev = ctx->FragmentProgram._Current;
1198
1199    ASSERT(ctx->FragmentProgram._MaintainTexEnvProgram);
1200
1201    /* If a conventional fragment program/shader isn't in effect... */
1202    if (!ctx->FragmentProgram._Enabled &&
1203        (!ctx->Shader.CurrentProgram ||
1204         !ctx->Shader.CurrentProgram->FragmentProgram) ) {
1205
1206       ctx->FragmentProgram._Current
1207          = ctx->FragmentProgram._TexEnvProgram
1208          = _mesa_get_fixed_func_fragment_program(ctx);
1209    }
1210
1211    /* Tell the driver about the change.  Could define a new target for
1212     * this?
1213     */
1214    if (ctx->FragmentProgram._Current != prev && ctx->Driver.BindProgram) {
1215       ctx->Driver.BindProgram(ctx, GL_FRAGMENT_PROGRAM_ARB,
1216                          (struct gl_program *) ctx->FragmentProgram._Current);
1217    }
1218 }