src/mesa/state_tracker/st_mesa_to_tgsi.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007-2008 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 /*
  29  * \author
  30  * Michal Krol,
  31  * Keith Whitwell
  32  */
  33
  34 #include "pipe/p_compiler.h"
  35 #include "pipe/p_shader_tokens.h"
  36 #include "pipe/p_state.h"
  37 #include "pipe/p_context.h"
  38 #include "tgsi/tgsi_ureg.h"
  39 #include "st_mesa_to_tgsi.h"
  40 #include "st_context.h"
  41 #include "shader/prog_instruction.h"
  42 #include "shader/prog_parameter.h"
  43 #include "util/u_debug.h"
  44 #include "util/u_math.h"
  45 #include "util/u_memory.h"
  46
  47 struct label {
  48    unsigned branch_target;
  49    unsigned token;
  50 };
  51
  52
  53 /**
  54  * Intermediate state used during shader translation.
  55  */
  56 struct st_translate {
  57    struct ureg_program *ureg;
  58
  59    struct ureg_dst temps[MAX_PROGRAM_TEMPS];
  60    struct ureg_src *constants;
  61    struct ureg_dst outputs[PIPE_MAX_SHADER_OUTPUTS];
  62    struct ureg_src inputs[PIPE_MAX_SHADER_INPUTS];
  63    struct ureg_dst address[1];
  64    struct ureg_src samplers[PIPE_MAX_SAMPLERS];
  65    struct ureg_dst psizregreal;
  66    struct ureg_src pointSizeConst;
  67    GLint psizoutindex;
  68    GLboolean prevInstWrotePsiz;
  69
  70    const GLuint *inputMapping;
  71    const GLuint *outputMapping;
  72
  73    /* For every instruction that contains a label (eg CALL), keep
  74     * details so that we can go back afterwards and emit the correct
  75     * tgsi instruction number for each label.
  76     */
  77    struct label *labels;
  78    unsigned labels_size;
  79    unsigned labels_count;
  80
  81    /* Keep a record of the tgsi instruction number that each mesa
  82     * instruction starts at, will be used to fix up labels after
  83     * translation.
  84     */
  85    unsigned *insn;
  86    unsigned insn_size;
  87    unsigned insn_count;
  88
  89    unsigned procType;  /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
  90
  91    boolean error;
  92 };
  93
  94
  95 static unsigned *get_label( struct st_translate *t,
  96                             unsigned branch_target )
  97 {
  98    unsigned i;
  99
 100    if (t->labels_count + 1 >= t->labels_size) {
 101       unsigned old_size = t->labels_size;
 102       t->labels_size = 1 << (util_logbase2(t->labels_size) + 1);
 103       t->labels = REALLOC( t->labels,
 104                            old_size * sizeof t->labels[0],
 105                            t->labels_size * sizeof t->labels[0] );
 106       if (t->labels == NULL) {
 107          static unsigned dummy;
 108          t->error = TRUE;
 109          return &dummy;
 110       }
 111    }
 112
 113    i = t->labels_count++;
 114    t->labels[i].branch_target = branch_target;
 115    return &t->labels[i].token;
 116 }
 117
 118
 119 static void set_insn_start( struct st_translate *t,
 120                             unsigned start )
 121 {
 122    if (t->insn_count + 1 >= t->insn_size) {
 123       unsigned old_size = t->insn_size;
 124       t->insn_size = 1 << (util_logbase2(t->insn_size) + 1);
 125       t->insn = REALLOC( t->insn,
 126                          old_size * sizeof t->insn[0],
 127                          t->insn_size * sizeof t->insn[0] );
 128       if (t->insn == NULL) {
 129          t->error = TRUE;
 130          return;
 131       }
 132    }
 133
 134    t->insn[t->insn_count++] = start;
 135 }
 136
 137
 138 /*
 139  * Map mesa register file to TGSI register file.
 140  */
 141 static struct ureg_dst
 142 dst_register( struct st_translate *t,
 143               gl_register_file file,
 144               GLuint index )
 145 {
 146    switch( file ) {
 147    case PROGRAM_UNDEFINED:
 148       return ureg_dst_undef();
 149
 150    case PROGRAM_TEMPORARY:
 151       if (ureg_dst_is_undef(t->temps[index]))
 152          t->temps[index] = ureg_DECL_temporary( t->ureg );
 153
 154       return t->temps[index];
 155
 156    case PROGRAM_OUTPUT:
 157       if (index == VERT_RESULT_PSIZ)
 158          t->prevInstWrotePsiz = GL_TRUE;
 159       return t->outputs[t->outputMapping[index]];
 160
 161    case PROGRAM_ADDRESS:
 162       return t->address[index];
 163
 164    default:
 165       debug_assert( 0 );
 166       return ureg_dst_undef();
 167    }
 168 }
 169
 170
 171 static struct ureg_src
 172 src_register( struct st_translate *t,
 173               gl_register_file file,
 174               GLint index )
 175 {
 176    switch( file ) {
 177    case PROGRAM_UNDEFINED:
 178       return ureg_src_undef();
 179
 180    case PROGRAM_TEMPORARY:
 181       ASSERT(index >= 0);
 182       if (ureg_dst_is_undef(t->temps[index]))
 183          t->temps[index] = ureg_DECL_temporary( t->ureg );
 184       assert(index < Elements(t->temps));
 185       return ureg_src(t->temps[index]);
 186
 187    case PROGRAM_NAMED_PARAM:
 188    case PROGRAM_ENV_PARAM:
 189    case PROGRAM_LOCAL_PARAM:
 190    case PROGRAM_UNIFORM:
 191       ASSERT(index >= 0);
 192       return t->constants[index];
 193    case PROGRAM_STATE_VAR:
 194    case PROGRAM_CONSTANT:       /* ie, immediate */
 195       if (index < 0)
 196          return ureg_DECL_constant( t->ureg, 0 );
 197       else
 198          return t->constants[index];
 199
 200    case PROGRAM_INPUT:
 201       assert(t->inputMapping[index] < Elements(t->inputs));
 202       return t->inputs[t->inputMapping[index]];
 203
 204    case PROGRAM_OUTPUT:
 205       assert(t->outputMapping[index] < Elements(t->outputs));
 206       return ureg_src(t->outputs[t->outputMapping[index]]); /* not needed? */
 207
 208    case PROGRAM_ADDRESS:
 209       return ureg_src(t->address[index]);
 210
 211    default:
 212       debug_assert( 0 );
 213       return ureg_src_undef();
 214    }
 215 }
 216
 217
 218 /**
 219  * Map mesa texture target to TGSI texture target.
 220  */
 221 static unsigned
 222 translate_texture_target( GLuint textarget,
 223                           GLboolean shadow )
 224 {
 225    if (shadow) {
 226       switch( textarget ) {
 227       case TEXTURE_1D_INDEX:   return TGSI_TEXTURE_SHADOW1D;
 228       case TEXTURE_2D_INDEX:   return TGSI_TEXTURE_SHADOW2D;
 229       case TEXTURE_RECT_INDEX: return TGSI_TEXTURE_SHADOWRECT;
 230       default: break;
 231       }
 232    }
 233
 234    switch( textarget ) {
 235    case TEXTURE_1D_INDEX:   return TGSI_TEXTURE_1D;
 236    case TEXTURE_2D_INDEX:   return TGSI_TEXTURE_2D;
 237    case TEXTURE_3D_INDEX:   return TGSI_TEXTURE_3D;
 238    case TEXTURE_CUBE_INDEX: return TGSI_TEXTURE_CUBE;
 239    case TEXTURE_RECT_INDEX: return TGSI_TEXTURE_RECT;
 240    default:
 241       debug_assert( 0 );
 242       return TGSI_TEXTURE_1D;
 243    }
 244 }
 245
 246
 247 static struct ureg_dst
 248 translate_dst( struct st_translate *t,
 249                const struct prog_dst_register *DstReg,
 250                boolean saturate )
 251 {
 252    struct ureg_dst dst = dst_register( t,
 253                                        DstReg->File,
 254                                        DstReg->Index );
 255
 256    dst = ureg_writemask( dst,
 257                          DstReg->WriteMask );
 258
 259    if (saturate)
 260       dst = ureg_saturate( dst );
 261
 262    if (DstReg->RelAddr)
 263       dst = ureg_dst_indirect( dst, ureg_src(t->address[0]) );
 264
 265    return dst;
 266 }
 267
 268
 269 static struct ureg_src
 270 translate_src( struct st_translate *t,
 271                const struct prog_src_register *SrcReg )
 272 {
 273    struct ureg_src src = src_register( t, SrcReg->File, SrcReg->Index );
 274
 275    src = ureg_swizzle( src,
 276                        GET_SWZ( SrcReg->Swizzle, 0 ) & 0x3,
 277                        GET_SWZ( SrcReg->Swizzle, 1 ) & 0x3,
 278                        GET_SWZ( SrcReg->Swizzle, 2 ) & 0x3,
 279                        GET_SWZ( SrcReg->Swizzle, 3 ) & 0x3);
 280
 281    if (SrcReg->Negate == NEGATE_XYZW)
 282       src = ureg_negate(src);
 283
 284    if (SrcReg->Abs)
 285       src = ureg_abs(src);
 286
 287    if (SrcReg->RelAddr) {
 288       src = ureg_src_indirect( src, ureg_src(t->address[0]));
 289       /* If SrcReg->Index was negative, it was set to zero in
 290        * src_register().  Reassign it now.
 291        */
 292       src.Index = SrcReg->Index;
 293    }
 294
 295    return src;
 296 }
 297
 298
 299 static struct ureg_src swizzle_4v( struct ureg_src src,
 300                                    const unsigned *swz )
 301 {
 302    return ureg_swizzle( src, swz[0], swz[1], swz[2], swz[3] );
 303 }
 304
 305
 306 /**
 307  * Translate a SWZ instruction into a MOV, MUL or MAD instruction.  EG:
 308  *
 309  *   SWZ dst, src.x-y10
 310  *
 311  * becomes:
 312  *
 313  *   MAD dst {1,-1,0,0}, src.xyxx, {0,0,1,0}
 314  */
 315 static void emit_swz( struct st_translate *t,
 316                       struct ureg_dst dst,
 317                       const struct prog_src_register *SrcReg )
 318 {
 319    struct ureg_program *ureg = t->ureg;
 320    struct ureg_src src = src_register( t, SrcReg->File, SrcReg->Index );
 321
 322    unsigned negate_mask =  SrcReg->Negate;
 323
 324    unsigned one_mask = ((GET_SWZ(SrcReg->Swizzle, 0) == SWIZZLE_ONE) << 0 |
 325                         (GET_SWZ(SrcReg->Swizzle, 1) == SWIZZLE_ONE) << 1 |
 326                         (GET_SWZ(SrcReg->Swizzle, 2) == SWIZZLE_ONE) << 2 |
 327                         (GET_SWZ(SrcReg->Swizzle, 3) == SWIZZLE_ONE) << 3);
 328
 329    unsigned zero_mask = ((GET_SWZ(SrcReg->Swizzle, 0) == SWIZZLE_ZERO) << 0 |
 330                          (GET_SWZ(SrcReg->Swizzle, 1) == SWIZZLE_ZERO) << 1 |
 331                          (GET_SWZ(SrcReg->Swizzle, 2) == SWIZZLE_ZERO) << 2 |
 332                          (GET_SWZ(SrcReg->Swizzle, 3) == SWIZZLE_ZERO) << 3);
 333
 334    unsigned negative_one_mask = one_mask & negate_mask;
 335    unsigned positive_one_mask = one_mask & ~negate_mask;
 336
 337    struct ureg_src imm;
 338    unsigned i;
 339    unsigned mul_swizzle[4] = {0,0,0,0};
 340    unsigned add_swizzle[4] = {0,0,0,0};
 341    unsigned src_swizzle[4] = {0,0,0,0};
 342    boolean need_add = FALSE;
 343    boolean need_mul = FALSE;
 344
 345    if (dst.WriteMask == 0)
 346       return;
 347
 348    /* Is this just a MOV?
 349     */
 350    if (zero_mask == 0 &&
 351        one_mask == 0 &&
 352        (negate_mask == 0 || negate_mask == TGSI_WRITEMASK_XYZW))
 353    {
 354       ureg_MOV( ureg, dst, translate_src( t, SrcReg ));
 355       return;
 356    }
 357
 358 #define IMM_ZERO    0
 359 #define IMM_ONE     1
 360 #define IMM_NEG_ONE 2
 361
 362    imm = ureg_imm3f( ureg, 0, 1, -1 );
 363
 364    for (i = 0; i < 4; i++) {
 365       unsigned bit = 1 << i;
 366
 367       if (dst.WriteMask & bit) {
 368          if (positive_one_mask & bit) {
 369             mul_swizzle[i] = IMM_ZERO;
 370             add_swizzle[i] = IMM_ONE;
 371             need_add = TRUE;
 372          }
 373          else if (negative_one_mask & bit) {
 374             mul_swizzle[i] = IMM_ZERO;
 375             add_swizzle[i] = IMM_NEG_ONE;
 376             need_add = TRUE;
 377          }
 378          else if (zero_mask & bit) {
 379             mul_swizzle[i] = IMM_ZERO;
 380             add_swizzle[i] = IMM_ZERO;
 381             need_add = TRUE;
 382          }
 383          else {
 384             add_swizzle[i] = IMM_ZERO;
 385             src_swizzle[i] = GET_SWZ(SrcReg->Swizzle, i);
 386             need_mul = TRUE;
 387             if (negate_mask & bit) {
 388                mul_swizzle[i] = IMM_NEG_ONE;
 389             }
 390             else {
 391                mul_swizzle[i] = IMM_ONE;
 392             }
 393          }
 394       }
 395    }
 396
 397    if (need_mul && need_add) {
 398       ureg_MAD( ureg,
 399                 dst,
 400                 swizzle_4v( src, src_swizzle ),
 401                 swizzle_4v( imm, mul_swizzle ),
 402                 swizzle_4v( imm, add_swizzle ) );
 403    }
 404    else if (need_mul) {
 405       ureg_MUL( ureg,
 406                 dst,
 407                 swizzle_4v( src, src_swizzle ),
 408                 swizzle_4v( imm, mul_swizzle ) );
 409    }
 410    else if (need_add) {
 411       ureg_MOV( ureg,
 412                 dst,
 413                 swizzle_4v( imm, add_swizzle ) );
 414    }
 415    else {
 416       debug_assert(0);
 417    }
 418
 419 #undef IMM_ZERO
 420 #undef IMM_ONE
 421 #undef IMM_NEG_ONE
 422 }
 423
 424
 425 /**
 426  * Negate the value of DDY to match GL semantics where (0,0) is the
 427  * lower-left corner of the window.
 428  * Note that the GL_ARB_fragment_coord_conventions extension will
 429  * effect this someday.
 430  */
 431 static void emit_ddy( struct st_translate *t,
 432                       struct ureg_dst dst,
 433                       const struct prog_src_register *SrcReg )
 434 {
 435    struct ureg_program *ureg = t->ureg;
 436    struct ureg_src src = translate_src( t, SrcReg );
 437    src = ureg_negate( src );
 438    ureg_DDY( ureg, dst, src );
 439 }
 440
 441
 442
 443 static unsigned
 444 translate_opcode( unsigned op )
 445 {
 446    switch( op ) {
 447    case OPCODE_ARL:
 448       return TGSI_OPCODE_ARL;
 449    case OPCODE_ABS:
 450       return TGSI_OPCODE_ABS;
 451    case OPCODE_ADD:
 452       return TGSI_OPCODE_ADD;
 453    case OPCODE_BGNLOOP:
 454       return TGSI_OPCODE_BGNLOOP;
 455    case OPCODE_BGNSUB:
 456       return TGSI_OPCODE_BGNSUB;
 457    case OPCODE_BRA:
 458       return TGSI_OPCODE_BRA;
 459    case OPCODE_BRK:
 460       return TGSI_OPCODE_BRK;
 461    case OPCODE_CAL:
 462       return TGSI_OPCODE_CAL;
 463    case OPCODE_CMP:
 464       return TGSI_OPCODE_CMP;
 465    case OPCODE_CONT:
 466       return TGSI_OPCODE_CONT;
 467    case OPCODE_COS:
 468       return TGSI_OPCODE_COS;
 469    case OPCODE_DDX:
 470       return TGSI_OPCODE_DDX;
 471    case OPCODE_DDY:
 472       return TGSI_OPCODE_DDY;
 473    case OPCODE_DP2:
 474       return TGSI_OPCODE_DP2;
 475    case OPCODE_DP2A:
 476       return TGSI_OPCODE_DP2A;
 477    case OPCODE_DP3:
 478       return TGSI_OPCODE_DP3;
 479    case OPCODE_DP4:
 480       return TGSI_OPCODE_DP4;
 481    case OPCODE_DPH:
 482       return TGSI_OPCODE_DPH;
 483    case OPCODE_DST:
 484       return TGSI_OPCODE_DST;
 485    case OPCODE_ELSE:
 486       return TGSI_OPCODE_ELSE;
 487    case OPCODE_ENDIF:
 488       return TGSI_OPCODE_ENDIF;
 489    case OPCODE_ENDLOOP:
 490       return TGSI_OPCODE_ENDLOOP;
 491    case OPCODE_ENDSUB:
 492       return TGSI_OPCODE_ENDSUB;
 493    case OPCODE_EX2:
 494       return TGSI_OPCODE_EX2;
 495    case OPCODE_EXP:
 496       return TGSI_OPCODE_EXP;
 497    case OPCODE_FLR:
 498       return TGSI_OPCODE_FLR;
 499    case OPCODE_FRC:
 500       return TGSI_OPCODE_FRC;
 501    case OPCODE_IF:
 502       return TGSI_OPCODE_IF;
 503    case OPCODE_TRUNC:
 504       return TGSI_OPCODE_TRUNC;
 505    case OPCODE_KIL:
 506       return TGSI_OPCODE_KIL;
 507    case OPCODE_KIL_NV:
 508       return TGSI_OPCODE_KILP;
 509    case OPCODE_LG2:
 510       return TGSI_OPCODE_LG2;
 511    case OPCODE_LOG:
 512       return TGSI_OPCODE_LOG;
 513    case OPCODE_LIT:
 514       return TGSI_OPCODE_LIT;
 515    case OPCODE_LRP:
 516       return TGSI_OPCODE_LRP;
 517    case OPCODE_MAD:
 518       return TGSI_OPCODE_MAD;
 519    case OPCODE_MAX:
 520       return TGSI_OPCODE_MAX;
 521    case OPCODE_MIN:
 522       return TGSI_OPCODE_MIN;
 523    case OPCODE_MOV:
 524       return TGSI_OPCODE_MOV;
 525    case OPCODE_MUL:
 526       return TGSI_OPCODE_MUL;
 527    case OPCODE_NOP:
 528       return TGSI_OPCODE_NOP;
 529    case OPCODE_NRM3:
 530       return TGSI_OPCODE_NRM;
 531    case OPCODE_NRM4:
 532       return TGSI_OPCODE_NRM4;
 533    case OPCODE_POW:
 534       return TGSI_OPCODE_POW;
 535    case OPCODE_RCP:
 536       return TGSI_OPCODE_RCP;
 537    case OPCODE_RET:
 538       return TGSI_OPCODE_RET;
 539    case OPCODE_RSQ:
 540       return TGSI_OPCODE_RSQ;
 541    case OPCODE_SCS:
 542       return TGSI_OPCODE_SCS;
 543    case OPCODE_SEQ:
 544       return TGSI_OPCODE_SEQ;
 545    case OPCODE_SGE:
 546       return TGSI_OPCODE_SGE;
 547    case OPCODE_SGT:
 548       return TGSI_OPCODE_SGT;
 549    case OPCODE_SIN:
 550       return TGSI_OPCODE_SIN;
 551    case OPCODE_SLE:
 552       return TGSI_OPCODE_SLE;
 553    case OPCODE_SLT:
 554       return TGSI_OPCODE_SLT;
 555    case OPCODE_SNE:
 556       return TGSI_OPCODE_SNE;
 557    case OPCODE_SSG:
 558       return TGSI_OPCODE_SSG;
 559    case OPCODE_SUB:
 560       return TGSI_OPCODE_SUB;
 561    case OPCODE_TEX:
 562       return TGSI_OPCODE_TEX;
 563    case OPCODE_TXB:
 564       return TGSI_OPCODE_TXB;
 565    case OPCODE_TXD:
 566       return TGSI_OPCODE_TXD;
 567    case OPCODE_TXL:
 568       return TGSI_OPCODE_TXL;
 569    case OPCODE_TXP:
 570       return TGSI_OPCODE_TXP;
 571    case OPCODE_XPD:
 572       return TGSI_OPCODE_XPD;
 573    case OPCODE_END:
 574       return TGSI_OPCODE_END;
 575    default:
 576       debug_assert( 0 );
 577       return TGSI_OPCODE_NOP;
 578    }
 579 }
 580
 581
 582 static void
 583 compile_instruction(
 584    struct st_translate *t,
 585    const struct prog_instruction *inst )
 586 {
 587    struct ureg_program *ureg = t->ureg;
 588    GLuint i;
 589    struct ureg_dst dst[1];
 590    struct ureg_src src[4];
 591    unsigned num_dst;
 592    unsigned num_src;
 593
 594    num_dst = _mesa_num_inst_dst_regs( inst->Opcode );
 595    num_src = _mesa_num_inst_src_regs( inst->Opcode );
 596
 597    if (num_dst)
 598       dst[0] = translate_dst( t,
 599                               &inst->DstReg,
 600                               inst->SaturateMode );
 601
 602    for (i = 0; i < num_src; i++)
 603       src[i] = translate_src( t, &inst->SrcReg[i] );
 604
 605    switch( inst->Opcode ) {
 606    case OPCODE_SWZ:
 607       emit_swz( t, dst[0], &inst->SrcReg[0] );
 608       return;
 609
 610    case OPCODE_BGNLOOP:
 611    case OPCODE_CAL:
 612    case OPCODE_ELSE:
 613    case OPCODE_ENDLOOP:
 614    case OPCODE_IF:
 615       debug_assert(num_dst == 0);
 616       ureg_label_insn( ureg,
 617                        translate_opcode( inst->Opcode ),
 618                        src, num_src,
 619                        get_label( t, inst->BranchTarget ));
 620       return;
 621
 622    case OPCODE_TEX:
 623    case OPCODE_TXB:
 624    case OPCODE_TXD:
 625    case OPCODE_TXL:
 626    case OPCODE_TXP:
 627       src[num_src++] = t->samplers[inst->TexSrcUnit];
 628       ureg_tex_insn( ureg,
 629                      translate_opcode( inst->Opcode ),
 630                      dst, num_dst,
 631                      translate_texture_target( inst->TexSrcTarget,
 632                                                inst->TexShadow ),
 633                      src, num_src );
 634       return;
 635
 636    case OPCODE_SCS:
 637       dst[0] = ureg_writemask(dst[0], TGSI_WRITEMASK_XY );
 638       ureg_insn( ureg,
 639                  translate_opcode( inst->Opcode ),
 640                  dst, num_dst,
 641                  src, num_src );
 642       break;
 643
 644    case OPCODE_XPD:
 645       dst[0] = ureg_writemask(dst[0], TGSI_WRITEMASK_XYZ );
 646       ureg_insn( ureg,
 647                  translate_opcode( inst->Opcode ),
 648                  dst, num_dst,
 649                  src, num_src );
 650       break;
 651
 652    case OPCODE_NOISE1:
 653    case OPCODE_NOISE2:
 654    case OPCODE_NOISE3:
 655    case OPCODE_NOISE4:
 656       /* At some point, a motivated person could add a better
 657        * implementation of noise.  Currently not even the nvidia
 658        * binary drivers do anything more than this.  In any case, the
 659        * place to do this is in the GL state tracker, not the poor
 660        * driver.
 661        */
 662       ureg_MOV( ureg, dst[0], ureg_imm1f(ureg, 0.5) );
 663       break;
 664
 665    case OPCODE_DDY:
 666       emit_ddy( t, dst[0], &inst->SrcReg[0] );
 667       break;
 668
 669    default:
 670       ureg_insn( ureg,
 671                  translate_opcode( inst->Opcode ),
 672                  dst, num_dst,
 673                  src, num_src );
 674       break;
 675    }
 676 }
 677
 678 /**
 679  * Emit the TGSI instructions to adjust the WPOS pixel center convention
 680  */
 681 static void
 682 emit_adjusted_wpos( struct st_translate *t,
 683                     const struct gl_program *program, GLfloat value)
 684 {
 685    struct ureg_program *ureg = t->ureg;
 686    struct ureg_dst wpos_temp = ureg_DECL_temporary(ureg);
 687    struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]];
 688
 689    ureg_ADD(ureg, ureg_writemask(wpos_temp, TGSI_WRITEMASK_X | TGSI_WRITEMASK_Y),
 690                    wpos_input, ureg_imm1f(ureg, value));
 691
 692    t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]] = ureg_src(wpos_temp);
 693 }
 694
 695 /**
 696  * Emit the TGSI instructions for inverting the WPOS y coordinate.
 697  */
 698 static void
 699 emit_inverted_wpos( struct st_translate *t,
 700                     const struct gl_program *program )
 701 {
 702    struct ureg_program *ureg = t->ureg;
 703
 704    /* Fragment program uses fragment position input.
 705     * Need to replace instances of INPUT[WPOS] with temp T
 706     * where T = INPUT[WPOS] by y is inverted.
 707     */
 708    static const gl_state_index winSizeState[STATE_LENGTH]
 709       = { STATE_INTERNAL, STATE_FB_SIZE, 0, 0, 0 };
 710
 711    /* XXX: note we are modifying the incoming shader here!  Need to
 712     * do this before emitting the constant decls below, or this
 713     * will be missed:
 714     */
 715    unsigned winHeightConst = _mesa_add_state_reference(program->Parameters,
 716                                                        winSizeState);
 717
 718    struct ureg_src winsize = ureg_DECL_constant( ureg, winHeightConst );
 719    struct ureg_dst wpos_temp;
 720    struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]];
 721
 722    /* MOV wpos_temp, input[wpos]
 723     */
 724    if (wpos_input.File == TGSI_FILE_TEMPORARY)
 725       wpos_temp = ureg_dst(wpos_input);
 726    else {
 727       wpos_temp = ureg_DECL_temporary( ureg );
 728       ureg_MOV( ureg, wpos_temp, wpos_input );
 729    }
 730
 731    /* SUB wpos_temp.y, winsize_const, wpos_input
 732     */
 733    ureg_SUB( ureg,
 734              ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y ),
 735              winsize,
 736              wpos_input);
 737
 738    /* Use wpos_temp as position input from here on:
 739     */
 740    t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]] = ureg_src(wpos_temp);
 741 }
 742
 743
 744 /**
 745  * Emit fragment position/ooordinate code.
 746  */
 747 static void
 748 emit_wpos(struct st_context *st,
 749           struct st_translate *t,
 750           const struct gl_program *program,
 751           struct ureg_program *ureg)
 752 {
 753    const struct gl_fragment_program *fp =
 754       (const struct gl_fragment_program *) program;
 755    struct pipe_screen *pscreen = st->pipe->screen;
 756    boolean invert = FALSE;
 757
 758    if (fp->OriginUpperLeft) {
 759       if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) {
 760       }
 761       else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) {
 762          ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
 763          invert = TRUE;
 764       }
 765       else
 766          assert(0);
 767    }
 768    else {
 769       if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT))
 770          ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
 771       else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT))
 772          invert = TRUE;
 773       else
 774          assert(0);
 775    }
 776
 777    if (fp->PixelCenterInteger) {
 778       if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER))
 779          ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
 780       else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER))
 781          emit_adjusted_wpos(t, program, invert ? 0.5f : -0.5f);
 782       else
 783          assert(0);
 784    }
 785    else {
 786       if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) {
 787       }
 788       else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) {
 789          ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
 790          emit_adjusted_wpos(t, program, invert ? -0.5f : 0.5f);
 791       }
 792       else
 793          assert(0);
 794    }
 795
 796    /* we invert after adjustment so that we avoid the MOV to temporary,
 797     * and reuse the adjustment ADD instead */
 798    if (invert)
 799       emit_inverted_wpos(t, program);
 800 }
 801
 802
 803 /**
 804  * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
 805  * TGSI uses +1 for front, -1 for back.
 806  * This function converts the TGSI value to the GL value.  Simply clamping/
 807  * saturating the value to [0,1] does the job.
 808  */
 809 static void
 810 emit_face_var( struct st_translate *t,
 811                const struct gl_program *program )
 812 {
 813    struct ureg_program *ureg = t->ureg;
 814    struct ureg_dst face_temp = ureg_DECL_temporary( ureg );
 815    struct ureg_src face_input = t->inputs[t->inputMapping[FRAG_ATTRIB_FACE]];
 816
 817    /* MOV_SAT face_temp, input[face]
 818     */
 819    face_temp = ureg_saturate( face_temp );
 820    ureg_MOV( ureg, face_temp, face_input );
 821
 822    /* Use face_temp as face input from here on:
 823     */
 824    t->inputs[t->inputMapping[FRAG_ATTRIB_FACE]] = ureg_src(face_temp);
 825 }
 826
 827
 828 static void
 829 emit_edgeflags( struct st_translate *t,
 830                  const struct gl_program *program )
 831 {
 832    struct ureg_program *ureg = t->ureg;
 833    struct ureg_dst edge_dst = t->outputs[t->outputMapping[VERT_RESULT_EDGE]];
 834    struct ureg_src edge_src = t->inputs[t->inputMapping[VERT_ATTRIB_EDGEFLAG]];
 835
 836    ureg_MOV( ureg, edge_dst, edge_src );
 837 }
 838
 839
 840 /**
 841  * Translate Mesa program to TGSI format.
 842  * \param program  the program to translate
 843  * \param numInputs  number of input registers used
 844  * \param inputMapping  maps Mesa fragment program inputs to TGSI generic
 845  *                      input indexes
 846  * \param inputSemanticName  the TGSI_SEMANTIC flag for each input
 847  * \param inputSemanticIndex  the semantic index (ex: which texcoord) for
 848  *                            each input
 849  * \param interpMode  the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
 850  * \param numOutputs  number of output registers used
 851  * \param outputMapping  maps Mesa fragment program outputs to TGSI
 852  *                       generic outputs
 853  * \param outputSemanticName  the TGSI_SEMANTIC flag for each output
 854  * \param outputSemanticIndex  the semantic index (ex: which texcoord) for
 855  *                             each output
 856  *
 857  * \return  PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
 858  */
 859 enum pipe_error
 860 st_translate_mesa_program(
 861    GLcontext *ctx,
 862    uint procType,
 863    struct ureg_program *ureg,
 864    const struct gl_program *program,
 865    GLuint numInputs,
 866    const GLuint inputMapping[],
 867    const ubyte inputSemanticName[],
 868    const ubyte inputSemanticIndex[],
 869    const GLuint interpMode[],
 870    GLuint numOutputs,
 871    const GLuint outputMapping[],
 872    const ubyte outputSemanticName[],
 873    const ubyte outputSemanticIndex[],
 874    boolean passthrough_edgeflags )
 875 {
 876    struct st_translate translate, *t;
 877    unsigned i;
 878    enum pipe_error ret = PIPE_OK;
 879
 880    t = &translate;
 881    memset(t, 0, sizeof *t);
 882
 883    t->procType = procType;
 884    t->inputMapping = inputMapping;
 885    t->outputMapping = outputMapping;
 886    t->ureg = ureg;
 887    t->psizoutindex = -1;
 888    t->prevInstWrotePsiz = GL_FALSE;
 889
 890    /*_mesa_print_program(program);*/
 891
 892    /*
 893     * Declare input attributes.
 894     */
 895    if (procType == TGSI_PROCESSOR_FRAGMENT) {
 896       for (i = 0; i < numInputs; i++) {
 897          if (program->InputFlags[0] & PROG_PARAM_BIT_CYL_WRAP) {
 898             t->inputs[i] = ureg_DECL_fs_input_cyl(ureg,
 899                                                   inputSemanticName[i],
 900                                                   inputSemanticIndex[i],
 901                                                   interpMode[i],
 902                                                   TGSI_CYLINDRICAL_WRAP_X);
 903          }
 904          else {
 905             t->inputs[i] = ureg_DECL_fs_input(ureg,
 906                                               inputSemanticName[i],
 907                                               inputSemanticIndex[i],
 908                                               interpMode[i]);
 909          }
 910       }
 911
 912       if (program->InputsRead & FRAG_BIT_WPOS) {
 913          /* Must do this after setting up t->inputs, and before
 914           * emitting constant references, below:
 915           */
 916          emit_wpos(st_context(ctx), t, program, ureg);
 917       }
 918
 919       if (program->InputsRead & FRAG_BIT_FACE) {
 920          emit_face_var( t, program );
 921       }
 922
 923       /*
 924        * Declare output attributes.
 925        */
 926       for (i = 0; i < numOutputs; i++) {
 927          switch (outputSemanticName[i]) {
 928          case TGSI_SEMANTIC_POSITION:
 929             t->outputs[i] = ureg_DECL_output( ureg,
 930                                               TGSI_SEMANTIC_POSITION, /* Z / Depth */
 931                                               outputSemanticIndex[i] );
 932
 933             t->outputs[i] = ureg_writemask( t->outputs[i],
 934                                             TGSI_WRITEMASK_Z );
 935             break;
 936          case TGSI_SEMANTIC_COLOR:
 937             t->outputs[i] = ureg_DECL_output( ureg,
 938                                               TGSI_SEMANTIC_COLOR,
 939                                               outputSemanticIndex[i] );
 940             break;
 941          default:
 942             debug_assert(0);
 943             return 0;
 944          }
 945       }
 946    }
 947    else {
 948       for (i = 0; i < numInputs; i++) {
 949          t->inputs[i] = ureg_DECL_vs_input(ureg, i);
 950       }
 951
 952       for (i = 0; i < numOutputs; i++) {
 953          t->outputs[i] = ureg_DECL_output( ureg,
 954                                            outputSemanticName[i],
 955                                            outputSemanticIndex[i] );
 956          if ((outputSemanticName[i] == TGSI_SEMANTIC_PSIZE) && program->Id) {
 957             static const gl_state_index pointSizeClampState[STATE_LENGTH]
 958                = { STATE_INTERNAL, STATE_POINT_SIZE_IMPL_CLAMP, 0, 0, 0 };
 959                /* XXX: note we are modifying the incoming shader here!  Need to
 960                * do this before emitting the constant decls below, or this
 961                * will be missed:
 962                */
 963             unsigned pointSizeClampConst =
 964                _mesa_add_state_reference(program->Parameters,
 965                                          pointSizeClampState);
 966             struct ureg_dst psizregtemp = ureg_DECL_temporary( ureg );
 967             t->pointSizeConst = ureg_DECL_constant( ureg, pointSizeClampConst );
 968             t->psizregreal = t->outputs[i];
 969             t->psizoutindex = i;
 970             t->outputs[i] = psizregtemp;
 971          }
 972       }
 973       if (passthrough_edgeflags)
 974          emit_edgeflags( t, program );
 975    }
 976
 977    /* Declare address register.
 978     */
 979    if (program->NumAddressRegs > 0) {
 980       debug_assert( program->NumAddressRegs == 1 );
 981       t->address[0] = ureg_DECL_address( ureg );
 982    }
 983
 984    /* Emit constants and immediates.  Mesa uses a single index space
 985     * for these, so we put all the translated regs in t->constants.
 986     */
 987    if (program->Parameters) {
 988       t->constants = CALLOC( program->Parameters->NumParameters,
 989                              sizeof t->constants[0] );
 990       if (t->constants == NULL) {
 991          ret = PIPE_ERROR_OUT_OF_MEMORY;
 992          goto out;
 993       }
 994
 995       for (i = 0; i < program->Parameters->NumParameters; i++) {
 996          switch (program->Parameters->Parameters[i].Type) {
 997          case PROGRAM_ENV_PARAM:
 998          case PROGRAM_LOCAL_PARAM:
 999          case PROGRAM_STATE_VAR:
1000          case PROGRAM_NAMED_PARAM:
1001          case PROGRAM_UNIFORM:
1002             t->constants[i] = ureg_DECL_constant( ureg, i );
1003             break;
1004
1005             /* Emit immediates only when there is no address register
1006              * in use.  FIXME: Be smarter and recognize param arrays:
1007              * indirect addressing is only valid within the referenced
1008              * array.
1009              */
1010          case PROGRAM_CONSTANT:
1011             if (program->NumAddressRegs > 0)
1012                t->constants[i] = ureg_DECL_constant( ureg, i );
1013             else
1014                t->constants[i] =
1015                   ureg_DECL_immediate( ureg,
1016                                        program->Parameters->ParameterValues[i],
1017                                        4 );
1018             break;
1019          default:
1020             break;
1021          }
1022       }
1023    }
1024
1025    /* texture samplers */
1026    for (i = 0; i < ctx->Const.MaxTextureImageUnits; i++) {
1027       if (program->SamplersUsed & (1 << i)) {
1028          t->samplers[i] = ureg_DECL_sampler( ureg, i );
1029       }
1030    }
1031
1032    /* Emit each instruction in turn:
1033     */
1034    for (i = 0; i < program->NumInstructions; i++) {
1035       set_insn_start( t, ureg_get_instruction_number( ureg ));
1036       compile_instruction( t, &program->Instructions[i] );
1037
1038       /* note can't do that easily at the end of prog due to
1039          possible early return */
1040       if (t->prevInstWrotePsiz && program->Id) {
1041          set_insn_start( t, ureg_get_instruction_number( ureg ));
1042          ureg_MAX( t->ureg,
1043                    ureg_writemask(t->outputs[t->psizoutindex], WRITEMASK_X),
1044                    ureg_src(t->outputs[t->psizoutindex]),
1045                    ureg_swizzle(t->pointSizeConst, 1,1,1,1));
1046          ureg_MIN( t->ureg, ureg_writemask(t->psizregreal, WRITEMASK_X),
1047                    ureg_src(t->outputs[t->psizoutindex]),
1048                    ureg_swizzle(t->pointSizeConst, 2,2,2,2));
1049       }
1050       t->prevInstWrotePsiz = GL_FALSE;
1051    }
1052
1053    /* Fix up all emitted labels:
1054     */
1055    for (i = 0; i < t->labels_count; i++) {
1056       ureg_fixup_label( ureg,
1057                         t->labels[i].token,
1058                         t->insn[t->labels[i].branch_target] );
1059    }
1060
1061 out:
1062    FREE(t->insn);
1063    FREE(t->labels);
1064    FREE(t->constants);
1065
1066    if (t->error) {
1067       debug_printf("%s: translate error flag set\n", __FUNCTION__);
1068    }
1069
1070    return ret;
1071 }
1072
1073
1074 /**
1075  * Tokens cannot be free with free otherwise the builtin gallium
1076  * malloc debugging will get confused.
1077  */
1078 void
1079 st_free_tokens(const struct tgsi_token *tokens)
1080 {
1081    FREE((void *)tokens);
1082 }