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