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/program.h"
  32 #include "shader/prog_parameter.h"
  33 #include "shader/prog_cache.h"
  34 #include "shader/prog_instruction.h"
  35 #include "shader/prog_print.h"
  36 #include "shader/prog_statevars.h"
  37 #include "shader/programopt.h"
  38 #include "texenvprogram.h"
  39
  40
  41 struct texenvprog_cache_item
  42 {
  43    GLuint hash;
  44    void *key;
  45    struct gl_fragment_program *data;
  46    struct texenvprog_cache_item *next;
  47 };
  48
  49
  50 /**
  51  * This MAX is probably a bit generous, but that's OK.  There can be
  52  * up to four instructions per texture unit (TEX + 3 for combine),
  53  * then there's fog and specular add.
  54  */
  55 #define MAX_INSTRUCTIONS ((MAX_TEXTURE_UNITS * 4) + 12)
  56
  57 #define DISASSEM (MESA_VERBOSE & VERBOSE_DISASSEM)
  58
  59 struct mode_opt {
  60    GLuint Source:4;
  61    GLuint Operand:3;
  62 };
  63
  64 struct state_key {
  65    GLbitfield enabled_units;
  66    GLuint separate_specular:1;
  67    GLuint fog_enabled:1;
  68    GLuint fog_mode:2;
  69
  70    struct {
  71       GLuint enabled:1;
  72       GLuint source_index:3;   /* one of TEXTURE_1D/2D/3D/CUBE/RECT_INDEX */
  73       GLuint shadow:1;
  74       GLuint ScaleShiftRGB:2;
  75       GLuint ScaleShiftA:2;
  76
  77       GLuint NumArgsRGB:2;
  78       GLuint ModeRGB:4;
  79       struct mode_opt OptRGB[3];
  80
  81       GLuint NumArgsA:2;
  82       GLuint ModeA:4;
  83       struct mode_opt OptA[3];
  84    } unit[8];
  85 };
  86
  87 #define FOG_LINEAR  0
  88 #define FOG_EXP     1
  89 #define FOG_EXP2    2
  90 #define FOG_UNKNOWN 3
  91
  92 static GLuint translate_fog_mode( GLenum mode )
  93 {
  94    switch (mode) {
  95    case GL_LINEAR: return FOG_LINEAR;
  96    case GL_EXP: return FOG_EXP;
  97    case GL_EXP2: return FOG_EXP2;
  98    default: return FOG_UNKNOWN;
  99    }
 100 }
 101
 102 #define OPR_SRC_COLOR           0
 103 #define OPR_ONE_MINUS_SRC_COLOR 1
 104 #define OPR_SRC_ALPHA           2
 105 #define OPR_ONE_MINUS_SRC_ALPHA 3
 106 #define OPR_ZERO                4
 107 #define OPR_ONE                 5
 108 #define OPR_UNKNOWN             7
 109
 110 static GLuint translate_operand( GLenum operand )
 111 {
 112    switch (operand) {
 113    case GL_SRC_COLOR: return OPR_SRC_COLOR;
 114    case GL_ONE_MINUS_SRC_COLOR: return OPR_ONE_MINUS_SRC_COLOR;
 115    case GL_SRC_ALPHA: return OPR_SRC_ALPHA;
 116    case GL_ONE_MINUS_SRC_ALPHA: return OPR_ONE_MINUS_SRC_ALPHA;
 117    case GL_ZERO: return OPR_ZERO;
 118    case GL_ONE: return OPR_ONE;
 119    default:     return OPR_UNKNOWN;
 120    }
 121 }
 122
 123 #define SRC_TEXTURE  0
 124 #define SRC_TEXTURE0 1
 125 #define SRC_TEXTURE1 2
 126 #define SRC_TEXTURE2 3
 127 #define SRC_TEXTURE3 4
 128 #define SRC_TEXTURE4 5
 129 #define SRC_TEXTURE5 6
 130 #define SRC_TEXTURE6 7
 131 #define SRC_TEXTURE7 8
 132 #define SRC_CONSTANT 9
 133 #define SRC_PRIMARY_COLOR 10
 134 #define SRC_PREVIOUS 11
 135 #define SRC_UNKNOWN  15
 136
 137 static GLuint translate_source( GLenum src )
 138 {
 139    switch (src) {
 140    case GL_TEXTURE: return SRC_TEXTURE;
 141    case GL_TEXTURE0:
 142    case GL_TEXTURE1:
 143    case GL_TEXTURE2:
 144    case GL_TEXTURE3:
 145    case GL_TEXTURE4:
 146    case GL_TEXTURE5:
 147    case GL_TEXTURE6:
 148    case GL_TEXTURE7: return SRC_TEXTURE0 + (src - GL_TEXTURE0);
 149    case GL_CONSTANT: return SRC_CONSTANT;
 150    case GL_PRIMARY_COLOR: return SRC_PRIMARY_COLOR;
 151    case GL_PREVIOUS: return SRC_PREVIOUS;
 152    default: return SRC_UNKNOWN;
 153    }
 154 }
 155
 156 #define MODE_REPLACE       0
 157 #define MODE_MODULATE      1
 158 #define MODE_ADD           2
 159 #define MODE_ADD_SIGNED    3
 160 #define MODE_INTERPOLATE   4
 161 #define MODE_SUBTRACT      5
 162 #define MODE_DOT3_RGB      6
 163 #define MODE_DOT3_RGB_EXT  7
 164 #define MODE_DOT3_RGBA     8
 165 #define MODE_DOT3_RGBA_EXT 9
 166 #define MODE_MODULATE_ADD_ATI           10
 167 #define MODE_MODULATE_SIGNED_ADD_ATI    11
 168 #define MODE_MODULATE_SUBTRACT_ATI      12
 169 #define MODE_UNKNOWN       15
 170
 171 static GLuint translate_mode( GLenum mode )
 172 {
 173    switch (mode) {
 174    case GL_REPLACE: return MODE_REPLACE;
 175    case GL_MODULATE: return MODE_MODULATE;
 176    case GL_ADD: return MODE_ADD;
 177    case GL_ADD_SIGNED: return MODE_ADD_SIGNED;
 178    case GL_INTERPOLATE: return MODE_INTERPOLATE;
 179    case GL_SUBTRACT: return MODE_SUBTRACT;
 180    case GL_DOT3_RGB: return MODE_DOT3_RGB;
 181    case GL_DOT3_RGB_EXT: return MODE_DOT3_RGB_EXT;
 182    case GL_DOT3_RGBA: return MODE_DOT3_RGBA;
 183    case GL_DOT3_RGBA_EXT: return MODE_DOT3_RGBA_EXT;
 184    case GL_MODULATE_ADD_ATI: return MODE_MODULATE_ADD_ATI;
 185    case GL_MODULATE_SIGNED_ADD_ATI: return MODE_MODULATE_SIGNED_ADD_ATI;
 186    case GL_MODULATE_SUBTRACT_ATI: return MODE_MODULATE_SUBTRACT_ATI;
 187    default: return MODE_UNKNOWN;
 188    }
 189 }
 190
 191 #define TEXTURE_UNKNOWN_INDEX 7
 192 static GLuint translate_tex_src_bit( GLbitfield bit )
 193 {
 194    switch (bit) {
 195    case TEXTURE_1D_BIT:   return TEXTURE_1D_INDEX;
 196    case TEXTURE_2D_BIT:   return TEXTURE_2D_INDEX;
 197    case TEXTURE_RECT_BIT: return TEXTURE_RECT_INDEX;
 198    case TEXTURE_3D_BIT:   return TEXTURE_3D_INDEX;
 199    case TEXTURE_CUBE_BIT: return TEXTURE_CUBE_INDEX;
 200    default: return TEXTURE_UNKNOWN_INDEX;
 201    }
 202 }
 203
 204 /**
 205  * Examine current texture environment state and generate a unique
 206  * key to identify it.
 207  */
 208 static void make_state_key( GLcontext *ctx,  struct state_key *key )
 209 {
 210    GLuint i, j;
 211
 212    memset(key, 0, sizeof(*key));
 213
 214    for (i=0;i<MAX_TEXTURE_UNITS;i++) {
 215       const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[i];
 216       GLenum format;
 217
 218       if (!texUnit->_ReallyEnabled || !texUnit->Enabled)
 219          continue;
 220
 221       format = texUnit->_Current->Image[0][texUnit->_Current->BaseLevel]->_BaseFormat;
 222
 223       key->unit[i].enabled = 1;
 224       key->enabled_units |= (1<<i);
 225
 226       key->unit[i].source_index =
 227          translate_tex_src_bit(texUnit->_ReallyEnabled);
 228       key->unit[i].shadow = ((texUnit->_Current->CompareMode == GL_COMPARE_R_TO_TEXTURE) &&
 229                              ((format == GL_DEPTH_COMPONENT) ||
 230                               (format == GL_DEPTH_STENCIL_EXT)));
 231
 232       key->unit[i].NumArgsRGB = texUnit->_CurrentCombine->_NumArgsRGB;
 233       key->unit[i].NumArgsA = texUnit->_CurrentCombine->_NumArgsA;
 234
 235       key->unit[i].ModeRGB =
 236          translate_mode(texUnit->_CurrentCombine->ModeRGB);
 237       key->unit[i].ModeA =
 238          translate_mode(texUnit->_CurrentCombine->ModeA);
 239
 240       key->unit[i].ScaleShiftRGB = texUnit->_CurrentCombine->ScaleShiftRGB;
 241       key->unit[i].ScaleShiftA = texUnit->_CurrentCombine->ScaleShiftA;
 242
 243       for (j=0;j<3;j++) {
 244          key->unit[i].OptRGB[j].Operand =
 245             translate_operand(texUnit->_CurrentCombine->OperandRGB[j]);
 246          key->unit[i].OptA[j].Operand =
 247             translate_operand(texUnit->_CurrentCombine->OperandA[j]);
 248          key->unit[i].OptRGB[j].Source =
 249             translate_source(texUnit->_CurrentCombine->SourceRGB[j]);
 250          key->unit[i].OptA[j].Source =
 251             translate_source(texUnit->_CurrentCombine->SourceA[j]);
 252       }
 253    }
 254
 255    if (ctx->_TriangleCaps & DD_SEPARATE_SPECULAR)
 256       key->separate_specular = 1;
 257
 258    if (ctx->Fog.Enabled) {
 259       key->fog_enabled = 1;
 260       key->fog_mode = translate_fog_mode(ctx->Fog.Mode);
 261    }
 262 }
 263
 264 /* Use uregs to represent registers internally, translate to Mesa's
 265  * expected formats on emit.
 266  *
 267  * NOTE: These are passed by value extensively in this file rather
 268  * than as usual by pointer reference.  If this disturbs you, try
 269  * remembering they are just 32bits in size.
 270  *
 271  * GCC is smart enough to deal with these dword-sized structures in
 272  * much the same way as if I had defined them as dwords and was using
 273  * macros to access and set the fields.  This is much nicer and easier
 274  * to evolve.
 275  */
 276 struct ureg {
 277    GLuint file:4;
 278    GLuint idx:8;
 279    GLuint negatebase:1;
 280    GLuint abs:1;
 281    GLuint negateabs:1;
 282    GLuint swz:12;
 283    GLuint pad:5;
 284 };
 285
 286 static const struct ureg undef = {
 287    PROGRAM_UNDEFINED,
 288    ~0,
 289    0,
 290    0,
 291    0,
 292    0,
 293    0
 294 };
 295
 296
 297 /* State used to build the fragment program:
 298  */
 299 struct texenv_fragment_program {
 300    struct gl_fragment_program *program;
 301    GLcontext *ctx;
 302    struct state_key *state;
 303
 304    GLbitfield alu_temps;        /* Track texture indirections, see spec. */
 305    GLbitfield temps_output;     /* Track texture indirections, see spec. */
 306    GLbitfield temp_in_use;      /* Tracks temporary regs which are in use. */
 307    GLboolean error;
 308
 309    struct ureg src_texture[MAX_TEXTURE_UNITS];
 310    /* Reg containing each texture unit's sampled texture color,
 311     * else undef.
 312     */
 313
 314    struct ureg src_previous;    /* Reg containing color from previous
 315                                  * stage.  May need to be decl'd.
 316                                  */
 317
 318    GLuint last_tex_stage;       /* Number of last enabled texture unit */
 319
 320    struct ureg half;
 321    struct ureg one;
 322    struct ureg zero;
 323 };
 324
 325
 326
 327 static struct ureg make_ureg(GLuint file, GLuint idx)
 328 {
 329    struct ureg reg;
 330    reg.file = file;
 331    reg.idx = idx;
 332    reg.negatebase = 0;
 333    reg.abs = 0;
 334    reg.negateabs = 0;
 335    reg.swz = SWIZZLE_NOOP;
 336    reg.pad = 0;
 337    return reg;
 338 }
 339
 340 static struct ureg swizzle( struct ureg reg, int x, int y, int z, int w )
 341 {
 342    reg.swz = MAKE_SWIZZLE4(GET_SWZ(reg.swz, x),
 343                            GET_SWZ(reg.swz, y),
 344                            GET_SWZ(reg.swz, z),
 345                            GET_SWZ(reg.swz, w));
 346
 347    return reg;
 348 }
 349
 350 static struct ureg swizzle1( struct ureg reg, int x )
 351 {
 352    return swizzle(reg, x, x, x, x);
 353 }
 354
 355 static struct ureg negate( struct ureg reg )
 356 {
 357    reg.negatebase ^= 1;
 358    return reg;
 359 }
 360
 361 static GLboolean is_undef( struct ureg reg )
 362 {
 363    return reg.file == PROGRAM_UNDEFINED;
 364 }
 365
 366
 367 static struct ureg get_temp( struct texenv_fragment_program *p )
 368 {
 369    GLint bit;
 370
 371    /* First try and reuse temps which have been used already:
 372     */
 373    bit = _mesa_ffs( ~p->temp_in_use & p->alu_temps );
 374
 375    /* Then any unused temporary:
 376     */
 377    if (!bit)
 378       bit = _mesa_ffs( ~p->temp_in_use );
 379
 380    if (!bit) {
 381       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 382       _mesa_exit(1);
 383    }
 384
 385    if ((GLuint) bit > p->program->Base.NumTemporaries)
 386       p->program->Base.NumTemporaries = bit;
 387
 388    p->temp_in_use |= 1<<(bit-1);
 389    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 390 }
 391
 392 static struct ureg get_tex_temp( struct texenv_fragment_program *p )
 393 {
 394    int bit;
 395
 396    /* First try to find availble temp not previously used (to avoid
 397     * starting a new texture indirection).  According to the spec, the
 398     * ~p->temps_output isn't necessary, but will keep it there for
 399     * now:
 400     */
 401    bit = _mesa_ffs( ~p->temp_in_use & ~p->alu_temps & ~p->temps_output );
 402
 403    /* Then any unused temporary:
 404     */
 405    if (!bit)
 406       bit = _mesa_ffs( ~p->temp_in_use );
 407
 408    if (!bit) {
 409       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 410       _mesa_exit(1);
 411    }
 412
 413    if ((GLuint) bit > p->program->Base.NumTemporaries)
 414       p->program->Base.NumTemporaries = bit;
 415
 416    p->temp_in_use |= 1<<(bit-1);
 417    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 418 }
 419
 420
 421 static void release_temps(GLcontext *ctx, struct texenv_fragment_program *p )
 422 {
 423    GLuint max_temp = ctx->Const.FragmentProgram.MaxTemps;
 424
 425    /* KW: To support tex_env_crossbar, don't release the registers in
 426     * temps_output.
 427     */
 428    if (max_temp >= sizeof(int) * 8)
 429       p->temp_in_use = p->temps_output;
 430    else
 431       p->temp_in_use = ~((1<<max_temp)-1) | p->temps_output;
 432 }
 433
 434
 435 static struct ureg register_param5( struct texenv_fragment_program *p,
 436                                     GLint s0,
 437                                     GLint s1,
 438                                     GLint s2,
 439                                     GLint s3,
 440                                     GLint s4)
 441 {
 442    gl_state_index tokens[STATE_LENGTH];
 443    GLuint idx;
 444    tokens[0] = s0;
 445    tokens[1] = s1;
 446    tokens[2] = s2;
 447    tokens[3] = s3;
 448    tokens[4] = s4;
 449    idx = _mesa_add_state_reference( p->program->Base.Parameters, tokens );
 450    return make_ureg(PROGRAM_STATE_VAR, idx);
 451 }
 452
 453
 454 #define register_param1(p,s0)          register_param5(p,s0,0,0,0,0)
 455 #define register_param2(p,s0,s1)       register_param5(p,s0,s1,0,0,0)
 456 #define register_param3(p,s0,s1,s2)    register_param5(p,s0,s1,s2,0,0)
 457 #define register_param4(p,s0,s1,s2,s3) register_param5(p,s0,s1,s2,s3,0)
 458
 459
 460 static struct ureg register_input( struct texenv_fragment_program *p, GLuint input )
 461 {
 462    p->program->Base.InputsRead |= (1 << input);
 463    return make_ureg(PROGRAM_INPUT, input);
 464 }
 465
 466
 467 static void emit_arg( struct prog_src_register *reg,
 468                       struct ureg ureg )
 469 {
 470    reg->File = ureg.file;
 471    reg->Index = ureg.idx;
 472    reg->Swizzle = ureg.swz;
 473    reg->NegateBase = ureg.negatebase ? 0xf : 0x0;
 474    reg->Abs = ureg.abs;
 475    reg->NegateAbs = ureg.negateabs;
 476 }
 477
 478 static void emit_dst( struct prog_dst_register *dst,
 479                       struct ureg ureg, GLuint mask )
 480 {
 481    dst->File = ureg.file;
 482    dst->Index = ureg.idx;
 483    dst->WriteMask = mask;
 484    dst->CondMask = COND_TR;  /* always pass cond test */
 485    dst->CondSwizzle = SWIZZLE_NOOP;
 486 }
 487
 488 static struct prog_instruction *
 489 emit_op(struct texenv_fragment_program *p,
 490         enum prog_opcode op,
 491         struct ureg dest,
 492         GLuint mask,
 493         GLboolean saturate,
 494         struct ureg src0,
 495         struct ureg src1,
 496         struct ureg src2 )
 497 {
 498    GLuint nr = p->program->Base.NumInstructions++;
 499    struct prog_instruction *inst = &p->program->Base.Instructions[nr];
 500
 501    assert(nr < MAX_INSTRUCTIONS);
 502
 503    _mesa_init_instructions(inst, 1);
 504    inst->Opcode = op;
 505
 506    emit_arg( &inst->SrcReg[0], src0 );
 507    emit_arg( &inst->SrcReg[1], src1 );
 508    emit_arg( &inst->SrcReg[2], src2 );
 509
 510    inst->SaturateMode = saturate ? SATURATE_ZERO_ONE : SATURATE_OFF;
 511
 512    emit_dst( &inst->DstReg, dest, mask );
 513
 514    /* Accounting for indirection tracking:
 515     */
 516    if (dest.file == PROGRAM_TEMPORARY)
 517       p->temps_output |= 1 << dest.idx;
 518
 519    return inst;
 520 }
 521
 522
 523 static struct ureg emit_arith( struct texenv_fragment_program *p,
 524                                enum prog_opcode op,
 525                                struct ureg dest,
 526                                GLuint mask,
 527                                GLboolean saturate,
 528                                struct ureg src0,
 529                                struct ureg src1,
 530                                struct ureg src2 )
 531 {
 532    emit_op(p, op, dest, mask, saturate, src0, src1, src2);
 533
 534    /* Accounting for indirection tracking:
 535     */
 536    if (src0.file == PROGRAM_TEMPORARY)
 537       p->alu_temps |= 1 << src0.idx;
 538
 539    if (!is_undef(src1) && src1.file == PROGRAM_TEMPORARY)
 540       p->alu_temps |= 1 << src1.idx;
 541
 542    if (!is_undef(src2) && src2.file == PROGRAM_TEMPORARY)
 543       p->alu_temps |= 1 << src2.idx;
 544
 545    if (dest.file == PROGRAM_TEMPORARY)
 546       p->alu_temps |= 1 << dest.idx;
 547
 548    p->program->Base.NumAluInstructions++;
 549    return dest;
 550 }
 551
 552 static struct ureg emit_texld( struct texenv_fragment_program *p,
 553                                enum prog_opcode op,
 554                                struct ureg dest,
 555                                GLuint destmask,
 556                                GLuint tex_unit,
 557                                GLuint tex_idx,
 558                                struct ureg coord )
 559 {
 560    struct prog_instruction *inst = emit_op( p, op,
 561                                           dest, destmask,
 562                                           GL_FALSE,     /* don't saturate? */
 563                                           coord,        /* arg 0? */
 564                                           undef,
 565                                           undef);
 566
 567    inst->TexSrcTarget = tex_idx;
 568    inst->TexSrcUnit = tex_unit;
 569
 570    p->program->Base.NumTexInstructions++;
 571
 572    /* Is this a texture indirection?
 573     */
 574    if ((coord.file == PROGRAM_TEMPORARY &&
 575         (p->temps_output & (1<<coord.idx))) ||
 576        (dest.file == PROGRAM_TEMPORARY &&
 577         (p->alu_temps & (1<<dest.idx)))) {
 578       p->program->Base.NumTexIndirections++;
 579       p->temps_output = 1<<coord.idx;
 580       p->alu_temps = 0;
 581       assert(0);                /* KW: texture env crossbar */
 582    }
 583
 584    return dest;
 585 }
 586
 587
 588 static struct ureg register_const4f( struct texenv_fragment_program *p,
 589                                      GLfloat s0,
 590                                      GLfloat s1,
 591                                      GLfloat s2,
 592                                      GLfloat s3)
 593 {
 594    GLfloat values[4];
 595    GLuint idx, swizzle;
 596    values[0] = s0;
 597    values[1] = s1;
 598    values[2] = s2;
 599    values[3] = s3;
 600    idx = _mesa_add_unnamed_constant( p->program->Base.Parameters, values, 4,
 601                                      &swizzle );
 602    ASSERT(swizzle == SWIZZLE_NOOP);
 603    return make_ureg(PROGRAM_CONSTANT, idx);
 604 }
 605
 606 #define register_scalar_const(p, s0)    register_const4f(p, s0, s0, s0, s0)
 607 #define register_const1f(p, s0)         register_const4f(p, s0, 0, 0, 1)
 608 #define register_const2f(p, s0, s1)     register_const4f(p, s0, s1, 0, 1)
 609 #define register_const3f(p, s0, s1, s2) register_const4f(p, s0, s1, s2, 1)
 610
 611
 612 static struct ureg get_one( struct texenv_fragment_program *p )
 613 {
 614    if (is_undef(p->one))
 615       p->one = register_scalar_const(p, 1.0);
 616    return p->one;
 617 }
 618
 619 static struct ureg get_half( struct texenv_fragment_program *p )
 620 {
 621    if (is_undef(p->half))
 622       p->half = register_scalar_const(p, 0.5);
 623    return p->half;
 624 }
 625
 626 static struct ureg get_zero( struct texenv_fragment_program *p )
 627 {
 628    if (is_undef(p->zero))
 629       p->zero = register_scalar_const(p, 0.0);
 630    return p->zero;
 631 }
 632
 633
 634 static void program_error( struct texenv_fragment_program *p, const char *msg )
 635 {
 636    _mesa_problem(NULL, msg);
 637    p->error = 1;
 638 }
 639
 640 static struct ureg get_source( struct texenv_fragment_program *p,
 641                                GLuint src, GLuint unit )
 642 {
 643    switch (src) {
 644    case SRC_TEXTURE:
 645       assert(!is_undef(p->src_texture[unit]));
 646       return p->src_texture[unit];
 647
 648    case SRC_TEXTURE0:
 649    case SRC_TEXTURE1:
 650    case SRC_TEXTURE2:
 651    case SRC_TEXTURE3:
 652    case SRC_TEXTURE4:
 653    case SRC_TEXTURE5:
 654    case SRC_TEXTURE6:
 655    case SRC_TEXTURE7:
 656       assert(!is_undef(p->src_texture[src - SRC_TEXTURE0]));
 657       return p->src_texture[src - SRC_TEXTURE0];
 658
 659    case SRC_CONSTANT:
 660       return register_param2(p, STATE_TEXENV_COLOR, unit);
 661
 662    case SRC_PRIMARY_COLOR:
 663       return register_input(p, FRAG_ATTRIB_COL0);
 664
 665    case SRC_PREVIOUS:
 666    default:
 667       if (is_undef(p->src_previous))
 668          return register_input(p, FRAG_ATTRIB_COL0);
 669       else
 670          return p->src_previous;
 671    }
 672 }
 673
 674 static struct ureg emit_combine_source( struct texenv_fragment_program *p,
 675                                         GLuint mask,
 676                                         GLuint unit,
 677                                         GLuint source,
 678                                         GLuint operand )
 679 {
 680    struct ureg arg, src, one;
 681
 682    src = get_source(p, source, unit);
 683
 684    switch (operand) {
 685    case OPR_ONE_MINUS_SRC_COLOR:
 686       /* Get unused tmp,
 687        * Emit tmp = 1.0 - arg.xyzw
 688        */
 689       arg = get_temp( p );
 690       one = get_one( p );
 691       return emit_arith( p, OPCODE_SUB, arg, mask, 0, one, src, undef);
 692
 693    case OPR_SRC_ALPHA:
 694       if (mask == WRITEMASK_W)
 695          return src;
 696       else
 697          return swizzle1( src, SWIZZLE_W );
 698    case OPR_ONE_MINUS_SRC_ALPHA:
 699       /* Get unused tmp,
 700        * Emit tmp = 1.0 - arg.wwww
 701        */
 702       arg = get_temp(p);
 703       one = get_one(p);
 704       return emit_arith(p, OPCODE_SUB, arg, mask, 0,
 705                         one, swizzle1(src, SWIZZLE_W), undef);
 706    case OPR_ZERO:
 707       return get_zero(p);
 708    case OPR_ONE:
 709       return get_one(p);
 710    case OPR_SRC_COLOR:
 711    default:
 712       return src;
 713    }
 714 }
 715
 716 static GLboolean args_match( struct state_key *key, GLuint unit )
 717 {
 718    GLuint i, nr = key->unit[unit].NumArgsRGB;
 719
 720    for (i = 0 ; i < nr ; i++) {
 721       if (key->unit[unit].OptA[i].Source != key->unit[unit].OptRGB[i].Source)
 722          return GL_FALSE;
 723
 724       switch(key->unit[unit].OptA[i].Operand) {
 725       case OPR_SRC_ALPHA:
 726          switch(key->unit[unit].OptRGB[i].Operand) {
 727          case OPR_SRC_COLOR:
 728          case OPR_SRC_ALPHA:
 729             break;
 730          default:
 731             return GL_FALSE;
 732          }
 733          break;
 734       case OPR_ONE_MINUS_SRC_ALPHA:
 735          switch(key->unit[unit].OptRGB[i].Operand) {
 736          case OPR_ONE_MINUS_SRC_COLOR:
 737          case OPR_ONE_MINUS_SRC_ALPHA:
 738             break;
 739          default:
 740             return GL_FALSE;
 741          }
 742          break;
 743       default:
 744          return GL_FALSE;       /* impossible */
 745       }
 746    }
 747
 748    return GL_TRUE;
 749 }
 750
 751 static struct ureg emit_combine( struct texenv_fragment_program *p,
 752                                  struct ureg dest,
 753                                  GLuint mask,
 754                                  GLboolean saturate,
 755                                  GLuint unit,
 756                                  GLuint nr,
 757                                  GLuint mode,
 758                                  const struct mode_opt *opt)
 759 {
 760    struct ureg src[3];
 761    struct ureg tmp, half;
 762    GLuint i;
 763
 764    tmp = undef; /* silence warning (bug 5318) */
 765
 766    for (i = 0; i < nr; i++)
 767       src[i] = emit_combine_source( p, mask, unit, opt[i].Source, opt[i].Operand );
 768
 769    switch (mode) {
 770    case MODE_REPLACE:
 771       if (mask == WRITEMASK_XYZW && !saturate)
 772          return src[0];
 773       else
 774          return emit_arith( p, OPCODE_MOV, dest, mask, saturate, src[0], undef, undef );
 775    case MODE_MODULATE:
 776       return emit_arith( p, OPCODE_MUL, dest, mask, saturate,
 777                          src[0], src[1], undef );
 778    case MODE_ADD:
 779       return emit_arith( p, OPCODE_ADD, dest, mask, saturate,
 780                          src[0], src[1], undef );
 781    case MODE_ADD_SIGNED:
 782       /* tmp = arg0 + arg1
 783        * result = tmp - .5
 784        */
 785       half = get_half(p);
 786       tmp = get_temp( p );
 787       emit_arith( p, OPCODE_ADD, tmp, mask, 0, src[0], src[1], undef );
 788       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp, half, undef );
 789       return dest;
 790    case MODE_INTERPOLATE:
 791       /* Arg0 * (Arg2) + Arg1 * (1-Arg2) -- note arguments are reordered:
 792        */
 793       return emit_arith( p, OPCODE_LRP, dest, mask, saturate, src[2], src[0], src[1] );
 794
 795    case MODE_SUBTRACT:
 796       return emit_arith( p, OPCODE_SUB, dest, mask, saturate, src[0], src[1], undef );
 797
 798    case MODE_DOT3_RGBA:
 799    case MODE_DOT3_RGBA_EXT:
 800    case MODE_DOT3_RGB_EXT:
 801    case MODE_DOT3_RGB: {
 802       struct ureg tmp0 = get_temp( p );
 803       struct ureg tmp1 = get_temp( p );
 804       struct ureg neg1 = register_scalar_const(p, -1);
 805       struct ureg two  = register_scalar_const(p, 2);
 806
 807       /* tmp0 = 2*src0 - 1
 808        * tmp1 = 2*src1 - 1
 809        *
 810        * dst = tmp0 dot3 tmp1
 811        */
 812       emit_arith( p, OPCODE_MAD, tmp0, WRITEMASK_XYZW, 0,
 813                   two, src[0], neg1);
 814
 815       if (_mesa_memcmp(&src[0], &src[1], sizeof(struct ureg)) == 0)
 816          tmp1 = tmp0;
 817       else
 818          emit_arith( p, OPCODE_MAD, tmp1, WRITEMASK_XYZW, 0,
 819                      two, src[1], neg1);
 820       emit_arith( p, OPCODE_DP3, dest, mask, saturate, tmp0, tmp1, undef);
 821       return dest;
 822    }
 823    case MODE_MODULATE_ADD_ATI:
 824       /* Arg0 * Arg2 + Arg1 */
 825       return emit_arith( p, OPCODE_MAD, dest, mask, saturate,
 826                          src[0], src[2], src[1] );
 827    case MODE_MODULATE_SIGNED_ADD_ATI: {
 828       /* Arg0 * Arg2 + Arg1 - 0.5 */
 829       struct ureg tmp0 = get_temp(p);
 830       half = get_half(p);
 831       emit_arith( p, OPCODE_MAD, tmp0, mask, 0, src[0], src[2], src[1] );
 832       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp0, half, undef );
 833       return dest;
 834    }
 835    case MODE_MODULATE_SUBTRACT_ATI:
 836       /* Arg0 * Arg2 - Arg1 */
 837       emit_arith( p, OPCODE_MAD, dest, mask, 0, src[0], src[2], negate(src[1]) );
 838       return dest;
 839    default:
 840       return src[0];
 841    }
 842 }
 843
 844
 845 /**
 846  * Generate instructions for one texture unit's env/combiner mode.
 847  */
 848 static struct ureg
 849 emit_texenv(struct texenv_fragment_program *p, GLuint unit)
 850 {
 851    struct state_key *key = p->state;
 852    GLboolean saturate = (unit < p->last_tex_stage);
 853    GLuint rgb_shift, alpha_shift;
 854    struct ureg out, shift;
 855    struct ureg dest;
 856
 857    if (!key->unit[unit].enabled) {
 858       return get_source(p, SRC_PREVIOUS, 0);
 859    }
 860
 861    switch (key->unit[unit].ModeRGB) {
 862    case MODE_DOT3_RGB_EXT:
 863       alpha_shift = key->unit[unit].ScaleShiftA;
 864       rgb_shift = 0;
 865       break;
 866    case MODE_DOT3_RGBA_EXT:
 867       alpha_shift = 0;
 868       rgb_shift = 0;
 869       break;
 870    default:
 871       rgb_shift = key->unit[unit].ScaleShiftRGB;
 872       alpha_shift = key->unit[unit].ScaleShiftA;
 873       break;
 874    }
 875
 876    /* If this is the very last calculation, emit direct to output reg:
 877     */
 878    if (key->separate_specular ||
 879        unit != p->last_tex_stage ||
 880        alpha_shift ||
 881        rgb_shift)
 882       dest = get_temp( p );
 883    else
 884       dest = make_ureg(PROGRAM_OUTPUT, FRAG_RESULT_COLR);
 885
 886    /* Emit the RGB and A combine ops
 887     */
 888    if (key->unit[unit].ModeRGB == key->unit[unit].ModeA &&
 889        args_match(key, unit)) {
 890       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 891                           unit,
 892                           key->unit[unit].NumArgsRGB,
 893                           key->unit[unit].ModeRGB,
 894                           key->unit[unit].OptRGB);
 895    }
 896    else if (key->unit[unit].ModeRGB == MODE_DOT3_RGBA_EXT ||
 897             key->unit[unit].ModeRGB == MODE_DOT3_RGBA) {
 898
 899       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
 900                           unit,
 901                           key->unit[unit].NumArgsRGB,
 902                           key->unit[unit].ModeRGB,
 903                           key->unit[unit].OptRGB);
 904    }
 905    else {
 906       /* Need to do something to stop from re-emitting identical
 907        * argument calculations here:
 908        */
 909       out = emit_combine( p, dest, WRITEMASK_XYZ, saturate,
 910                           unit,
 911                           key->unit[unit].NumArgsRGB,
 912                           key->unit[unit].ModeRGB,
 913                           key->unit[unit].OptRGB);
 914       out = emit_combine( p, dest, WRITEMASK_W, saturate,
 915                           unit,
 916                           key->unit[unit].NumArgsA,
 917                           key->unit[unit].ModeA,
 918                           key->unit[unit].OptA);
 919    }
 920
 921    /* Deal with the final shift:
 922     */
 923    if (alpha_shift || rgb_shift) {
 924       if (rgb_shift == alpha_shift) {
 925          shift = register_scalar_const(p, (GLfloat)(1<<rgb_shift));
 926       }
 927       else {
 928          shift = register_const4f(p,
 929                                   (GLfloat)(1<<rgb_shift),
 930                                   (GLfloat)(1<<rgb_shift),
 931                                   (GLfloat)(1<<rgb_shift),
 932                                   (GLfloat)(1<<alpha_shift));
 933       }
 934       return emit_arith( p, OPCODE_MUL, dest, WRITEMASK_XYZW,
 935                          saturate, out, shift, undef );
 936    }
 937    else
 938       return out;
 939 }
 940
 941
 942 /**
 943  * Generate instruction for getting a texture source term.
 944  */
 945 static void load_texture( struct texenv_fragment_program *p, GLuint unit )
 946 {
 947    if (is_undef(p->src_texture[unit])) {
 948       GLuint dim = p->state->unit[unit].source_index;
 949       struct ureg texcoord = register_input(p, FRAG_ATTRIB_TEX0+unit);
 950       struct ureg tmp = get_tex_temp( p );
 951
 952       if (dim == TEXTURE_UNKNOWN_INDEX)
 953          program_error(p, "TexSrcBit");
 954
 955       /* TODO: Use D0_MASK_XY where possible.
 956        */
 957       if (p->state->unit[unit].enabled) {
 958          p->src_texture[unit] = emit_texld( p, OPCODE_TXP,
 959                                             tmp, WRITEMASK_XYZW,
 960                                             unit, dim, texcoord );
 961          if (p->state->unit[unit].shadow)
 962             p->program->Base.ShadowSamplers |= 1 << unit;
 963
 964          p->program->Base.SamplersUsed |= (1 << unit);
 965          /* This identity mapping should already be in place
 966           * (see _mesa_init_program_struct()) but let's be safe.
 967           */
 968          p->program->Base.SamplerUnits[unit] = unit;
 969       }
 970       else
 971          p->src_texture[unit] = get_zero(p);
 972    }
 973 }
 974
 975 static GLboolean load_texenv_source( struct texenv_fragment_program *p,
 976                                      GLuint src, GLuint unit )
 977 {
 978    switch (src) {
 979    case SRC_TEXTURE:
 980       load_texture(p, unit);
 981       break;
 982
 983    case SRC_TEXTURE0:
 984    case SRC_TEXTURE1:
 985    case SRC_TEXTURE2:
 986    case SRC_TEXTURE3:
 987    case SRC_TEXTURE4:
 988    case SRC_TEXTURE5:
 989    case SRC_TEXTURE6:
 990    case SRC_TEXTURE7:
 991       load_texture(p, src - SRC_TEXTURE0);
 992       break;
 993
 994    default:
 995       break;
 996    }
 997
 998    return GL_TRUE;
 999 }
1000
1001
1002 /**
1003  * Generate instructions for loading all texture source terms.
1004  */
1005 static GLboolean
1006 load_texunit_sources( struct texenv_fragment_program *p, int unit )
1007 {
1008    struct state_key *key = p->state;
1009    GLuint i;
1010
1011    for (i = 0; i < key->unit[unit].NumArgsRGB; i++) {
1012       load_texenv_source( p, key->unit[unit].OptRGB[i].Source, unit);
1013    }
1014
1015    for (i = 0; i < key->unit[unit].NumArgsA; i++) {
1016       load_texenv_source( p, key->unit[unit].OptA[i].Source, unit );
1017    }
1018
1019    return GL_TRUE;
1020 }
1021
1022
1023 /**
1024  * Generate a new fragment program which implements the context's
1025  * current texture env/combine mode.
1026  */
1027 static void
1028 create_new_program(GLcontext *ctx, struct state_key *key,
1029                    struct gl_fragment_program *program)
1030 {
1031    struct prog_instruction instBuffer[MAX_INSTRUCTIONS];
1032    struct texenv_fragment_program p;
1033    GLuint unit;
1034    struct ureg cf, out;
1035
1036    _mesa_memset(&p, 0, sizeof(p));
1037    p.ctx = ctx;
1038    p.state = key;
1039    p.program = program;
1040
1041    /* During code generation, use locally-allocated instruction buffer,
1042     * then alloc dynamic storage below.
1043     */
1044    p.program->Base.Instructions = instBuffer;
1045    p.program->Base.Target = GL_FRAGMENT_PROGRAM_ARB;
1046    p.program->Base.NumTexIndirections = 1;      /* correct? */
1047    p.program->Base.NumTexInstructions = 0;
1048    p.program->Base.NumAluInstructions = 0;
1049    p.program->Base.String = NULL;
1050    p.program->Base.NumInstructions =
1051    p.program->Base.NumTemporaries =
1052    p.program->Base.NumParameters =
1053    p.program->Base.NumAttributes = p.program->Base.NumAddressRegs = 0;
1054    p.program->Base.Parameters = _mesa_new_parameter_list();
1055
1056    p.program->Base.InputsRead = 0;
1057    p.program->Base.OutputsWritten = 1 << FRAG_RESULT_COLR;
1058
1059    for (unit = 0; unit < MAX_TEXTURE_UNITS; unit++)
1060       p.src_texture[unit] = undef;
1061
1062    p.src_previous = undef;
1063    p.half = undef;
1064    p.zero = undef;
1065    p.one = undef;
1066
1067    p.last_tex_stage = 0;
1068    release_temps(ctx, &p);
1069
1070    if (key->enabled_units) {
1071       /* First pass - to support texture_env_crossbar, first identify
1072        * all referenced texture sources and emit texld instructions
1073        * for each:
1074        */
1075       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits ; unit++)
1076          if (key->unit[unit].enabled) {
1077             load_texunit_sources( &p, unit );
1078             p.last_tex_stage = unit;
1079          }
1080
1081       /* Second pass - emit combine instructions to build final color:
1082        */
1083       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits; unit++)
1084          if (key->enabled_units & (1<<unit)) {
1085             p.src_previous = emit_texenv( &p, unit );
1086             release_temps(ctx, &p);     /* release all temps */
1087          }
1088    }
1089
1090    cf = get_source( &p, SRC_PREVIOUS, 0 );
1091    out = make_ureg( PROGRAM_OUTPUT, FRAG_RESULT_COLR );
1092
1093    if (key->separate_specular) {
1094       /* Emit specular add.
1095        */
1096       struct ureg s = register_input(&p, FRAG_ATTRIB_COL1);
1097       emit_arith( &p, OPCODE_ADD, out, WRITEMASK_XYZ, 0, cf, s, undef );
1098       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_W, 0, cf, undef, undef );
1099    }
1100    else if (_mesa_memcmp(&cf, &out, sizeof(cf)) != 0) {
1101       /* Will wind up in here if no texture enabled or a couple of
1102        * other scenarios (GL_REPLACE for instance).
1103        */
1104       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_XYZW, 0, cf, undef, undef );
1105    }
1106
1107    /* Finish up:
1108     */
1109    emit_arith( &p, OPCODE_END, undef, WRITEMASK_XYZW, 0, undef, undef, undef);
1110
1111    if (key->fog_enabled) {
1112       /* Pull fog mode from GLcontext, the value in the state key is
1113        * a reduced value and not what is expected in FogOption
1114        */
1115       p.program->FogOption = ctx->Fog.Mode;
1116       p.program->Base.InputsRead |= FRAG_BIT_FOGC; /* XXX new */
1117    } else
1118       p.program->FogOption = GL_NONE;
1119
1120    if (p.program->Base.NumTexIndirections > ctx->Const.FragmentProgram.MaxTexIndirections)
1121       program_error(&p, "Exceeded max nr indirect texture lookups");
1122
1123    if (p.program->Base.NumTexInstructions > ctx->Const.FragmentProgram.MaxTexInstructions)
1124       program_error(&p, "Exceeded max TEX instructions");
1125
1126    if (p.program->Base.NumAluInstructions > ctx->Const.FragmentProgram.MaxAluInstructions)
1127       program_error(&p, "Exceeded max ALU instructions");
1128
1129    ASSERT(p.program->Base.NumInstructions <= MAX_INSTRUCTIONS);
1130
1131    /* Allocate final instruction array */
1132    p.program->Base.Instructions
1133       = _mesa_alloc_instructions(p.program->Base.NumInstructions);
1134    if (!p.program->Base.Instructions) {
1135       _mesa_error(ctx, GL_OUT_OF_MEMORY,
1136                   "generating tex env program");
1137       return;
1138    }
1139    _mesa_copy_instructions(p.program->Base.Instructions, instBuffer,
1140                            p.program->Base.NumInstructions);
1141
1142    if (p.program->FogOption) {
1143       _mesa_append_fog_code(ctx, p.program);
1144       p.program->FogOption = GL_NONE;
1145    }
1146
1147
1148    /* Notify driver the fragment program has (actually) changed.
1149     */
1150    if (ctx->Driver.ProgramStringNotify) {
1151       ctx->Driver.ProgramStringNotify( ctx, GL_FRAGMENT_PROGRAM_ARB,
1152                                        &p.program->Base );
1153    }
1154
1155    if (DISASSEM) {
1156       _mesa_print_program(&p.program->Base);
1157       _mesa_printf("\n");
1158    }
1159 }
1160
1161
1162 /**
1163  * Return a fragment program which implements the current
1164  * fixed-function texture, fog and color-sum operations.
1165  */
1166 struct gl_fragment_program *
1167 _mesa_get_fixed_func_fragment_program(GLcontext *ctx)
1168 {
1169    struct gl_fragment_program *prog;
1170    struct state_key key;
1171
1172    make_state_key(ctx, &key);
1173
1174    prog = (struct gl_fragment_program *)
1175       _mesa_search_program_cache(ctx->FragmentProgram.Cache,
1176                                  &key, sizeof(key));
1177
1178    if (!prog) {
1179       prog = (struct gl_fragment_program *)
1180          ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
1181
1182       create_new_program(ctx, &key, prog);
1183
1184       _mesa_program_cache_insert(ctx, ctx->FragmentProgram.Cache,
1185                                  &key, sizeof(key), &prog->Base);
1186    }
1187
1188    return prog;
1189 }
1190
1191
1192
1193 /**
1194  * If _MaintainTexEnvProgram is set we'll generate a fragment program that
1195  * implements the current texture env/combine mode.
1196  * This function generates that program and puts it into effect.
1197  */
1198 void
1199 _mesa_UpdateTexEnvProgram( GLcontext *ctx )
1200 {
1201    const struct gl_fragment_program *prev = ctx->FragmentProgram._Current;
1202
1203    ASSERT(ctx->FragmentProgram._MaintainTexEnvProgram);
1204
1205    /* If a conventional fragment program/shader isn't in effect... */
1206    if (!ctx->FragmentProgram._Enabled &&
1207        (!ctx->Shader.CurrentProgram ||
1208         !ctx->Shader.CurrentProgram->FragmentProgram) )
1209    {
1210       struct gl_fragment_program *newProg;
1211
1212       newProg = _mesa_get_fixed_func_fragment_program(ctx);
1213
1214       _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._Current, newProg);
1215       _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._TexEnvProgram, newProg);
1216    }
1217
1218    /* Tell the driver about the change.  Could define a new target for
1219     * this?
1220     */
1221    if (ctx->FragmentProgram._Current != prev && ctx->Driver.BindProgram) {
1222       ctx->Driver.BindProgram(ctx, GL_FRAGMENT_PROGRAM_ARB,
1223                          (struct gl_program *) ctx->FragmentProgram._Current);
1224    }
1225 }