src/mesa/drivers/dri/r600/r700_assembler.h

   1 /*
   2  * Copyright (C) 2008-2009  Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included
  12  * in all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  15  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  18  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  19  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  */
  21
  22 /*
  23  * Authors:
  24  *   Richard Li <RichardZ.Li@amd.com>, <richardradeon@gmail.com>
  25  */
  26
  27 #ifndef _R700_ASSEMBLER_H_
  28 #define _R700_ASSEMBLER_H_
  29
  30 #include "main/mtypes.h"
  31 #include "program/prog_instruction.h"
  32
  33 #include "r700_chip.h"
  34 #include "r700_shaderinst.h"
  35 #include "r700_shader.h"
  36
  37 typedef enum LOADABLE_SCRIPT_SIGNITURE
  38 {
  39     GLSL_NOISE1 = 0x10000001,
  40     GLSL_NOISE2 = 0x10000002,
  41     GLSL_NOISE3 = 0x10000003,
  42     GLSL_NOISE4 = 0x10000004
  43 }LOADABLE_SCRIPT_SIGNITURE;
  44
  45 typedef struct COMPILED_SUB
  46 {
  47     struct  prog_instruction *Instructions;
  48     GLuint  NumInstructions;
  49     GLuint  NumTemporaries;
  50     GLuint  NumParameters;
  51     GLuint  MinRegIndex;
  52     GLfloat (*ParameterValues)[4];
  53     GLbyte  outputSwizzleX;
  54     GLbyte  outputSwizzleY;
  55     GLbyte  outputSwizzleZ;
  56     GLbyte  outputSwizzleW;
  57     GLshort srcRegIndex[3];
  58     GLushort dstRegIndex;
  59 }COMPILED_SUB;
  60
  61 typedef struct PRESUB_DESCtag
  62 {
  63     LOADABLE_SCRIPT_SIGNITURE sptSigniture;
  64     GLint  subIL_Shift;
  65     struct prog_src_register InReg[3];
  66     struct prog_dst_register OutReg;
  67
  68     GLushort maxStartReg;
  69     GLushort number_used_registers;
  70
  71     GLuint   unConstantsStart;
  72
  73     COMPILED_SUB * pCompiledSub;
  74 } PRESUB_DESC;
  75
  76 typedef enum SHADER_PIPE_TYPE
  77 {
  78     SPT_VP = 0,
  79     SPT_FP = 1
  80 } SHADER_PIPE_TYPE;
  81
  82 typedef enum ConstantCycles
  83 {
  84     NUMBER_OF_CYCLES     = 3,
  85     NUMBER_OF_COMPONENTS = 4
  86 } ConstantCycles;
  87
  88 typedef enum  HARDWARE_LIMIT_VALUES
  89 {
  90    TEMPORARY_REGISTER_OFFSET = SQ_ALU_SRC_GPR_BASE,
  91    MAX_TEMPORARY_REGISTERS   = SQ_ALU_SRC_GPR_SIZE,
  92    MAX_CONSTANT_REGISTERS    = SQ_ALU_SRC_CFILE_SIZE,
  93    CFILE_REGISTER_OFFSET     = SQ_ALU_SRC_CFILE_BASE,
  94    NUMBER_OF_INPUT_COLORS    = 2,
  95    NUMBER_OF_OUTPUT_COLORS   = 8,
  96    NUMBER_OF_TEXTURE_UNITS   = 16,
  97    MEGA_FETCH_BYTES          = 32
  98 } HARDWARE_LIMIT_VALUES;
  99
 100 typedef enum AddressMode
 101 {
 102     ADDR_ABSOLUTE          = 0,
 103     ADDR_RELATIVE_A0       = 1,
 104     ADDR_RELATIVE_FLI_0    = 2,
 105     NUMBER_OF_ADDR_MOD     = 3
 106 } AddressMode;
 107
 108 typedef enum SrcRegisterType
 109 {
 110     SRC_REG_TEMPORARY      = 0,
 111     SRC_REG_GPR            = 1,
 112     SRC_REG_CONSTANT       = 2,
 113     SRC_REG_ALT_TEMPORARY  = 3,
 114     SRC_REC_LITERAL        = 4,
 115     NUMBER_OF_SRC_REG_TYPE = 5
 116 } SrcRegisterType;
 117
 118 typedef enum DstRegisterType
 119 {
 120     DST_REG_TEMPORARY      = 0,
 121     DST_REG_A0             = 1,
 122     DST_REG_OUT            = 2,
 123     DST_REG_OUT_X_REPL     = 3,
 124     DST_REG_ALT_TEMPORARY  = 4,
 125     DST_REG_INPUT          = 5,
 126     NUMBER_OF_DST_REG_TYPE = 6
 127 } DstRegisterType;
 128
 129 typedef unsigned int BITS;
 130
 131 typedef struct PVSDSTtag
 132 {
 133 #ifdef MESA_BIG_ENDIAN
 134         BITS addrmode1:1; //32
 135         BITS addrmode0:1; //31   //29
 136
 137         BITS dualop:1;    // 30  //26
 138
 139         BITS op3:1;       // 29  Represents *_OP3_* ALU opcode
 140
 141         BITS writew:1;     //28
 142         BITS writez:1;
 143         BITS writey:1;
 144         BITS writex:1;
 145
 146         BITS reg:10;       //24   //20
 147         BITS rtype:3;
 148
 149         BITS pred_inv  :1; //11   //8
 150         BITS predicated:1; //10   //8
 151         BITS math:1;
 152         BITS opcode:8;     //(:6)  //@@@ really should be 10 bits for OP2
 153 #else
 154         BITS opcode:8;     //(:6)  //@@@ really should be 10 bits for OP2
 155         BITS math:1;
 156         BITS predicated:1; //10   //8
 157         BITS pred_inv  :1; //11   //8
 158
 159         BITS rtype:3;
 160         BITS reg:10;       //24   //20
 161
 162         BITS writex:1;
 163         BITS writey:1;
 164         BITS writez:1;
 165         BITS writew:1;     //28
 166
 167         BITS op3:1;       // 29  Represents *_OP3_* ALU opcode
 168
 169         BITS dualop:1;    // 30  //26
 170
 171         BITS addrmode0:1; //31   //29
 172         BITS addrmode1:1; //32
 173 #endif
 174 } PVSDST;
 175
 176 typedef struct PVSINSTtag
 177 {
 178 #ifdef MESA_BIG_ENDIAN
 179     BITS index_mode   :3;
 180     BITS SaturateMode :2;
 181     BITS literal_slots      :2;
 182 #else
 183     BITS literal_slots      :2;
 184     BITS SaturateMode :2;
 185     BITS index_mode   :3;
 186 #endif
 187 } PVSINST;
 188
 189 typedef struct PVSSRCtag
 190 {
 191 #ifdef MESA_BIG_ENDIAN
 192         BITS addrmode1:1; //32
 193         //BITS addrsel:2;
 194         BITS negw:1;      //31
 195         BITS negz:1;
 196         BITS negy:1;
 197         BITS negx:1;
 198         BITS abs:1;
 199
 200         BITS swizzlew:3;  //26
 201         BITS swizzlez:3;
 202         BITS swizzley:3;
 203         BITS swizzlex:3;
 204         BITS reg:10;      //14     (8)
 205         BITS addrmode0:1;
 206         BITS rtype:3;
 207 #else
 208         BITS rtype:3;
 209         BITS addrmode0:1;
 210         BITS reg:10;      //14     (8)
 211         BITS swizzlex:3;
 212         BITS swizzley:3;
 213         BITS swizzlez:3;
 214         BITS swizzlew:3;  //26
 215
 216         BITS abs:1;
 217         BITS negx:1;
 218         BITS negy:1;
 219         BITS negz:1;
 220         BITS negw:1;      //31
 221         //BITS addrsel:2;
 222         BITS addrmode1:1; //32
 223 #endif
 224 } PVSSRC;
 225
 226 typedef struct PVSMATHtag
 227 {
 228 #ifdef MESA_BIG_ENDIAN
 229         BITS spare2:3;
 230         BITS dstcomp:2; // select dest component
 231         BITS negy:1;
 232         BITS negx:1;
 233         BITS opcode:4;
 234         BITS dstoff:2; // 2 bits of dest offset into alt ram
 235         BITS swizzley:3;
 236         BITS swizzlex:3;
 237         BITS reg:8;
 238         BITS spare:1;
 239         BITS rtype:4;
 240 #else
 241         BITS rtype:4;
 242         BITS spare:1;
 243         BITS reg:8;
 244         BITS swizzlex:3;
 245         BITS swizzley:3;
 246         BITS dstoff:2; // 2 bits of dest offset into alt ram
 247         BITS opcode:4;
 248         BITS negx:1;
 249         BITS negy:1;
 250         BITS dstcomp:2; // select dest component
 251         BITS spare2:3;
 252 #endif
 253 } PVSMATH;
 254
 255 typedef union PVSDWORDtag
 256 {
 257         BITS    bits;
 258         PVSDST  dst;
 259     PVSINST dst2;
 260         PVSSRC  src;
 261         PVSMATH math;
 262         float   f;
 263 } PVSDWORD;
 264
 265 typedef struct VAP_OUT_VTX_FMT_0tag
 266 {
 267 #ifdef MESA_BIG_ENDIAN
 268         BITS resvd1:12;        // 20
 269
 270         BITS viewport_index:1; // 19
 271         BITS kill_flag:1;
 272         BITS rta_index:1;      //     shares same channel as kill_flag
 273         BITS edge_flag:1;
 274         BITS point_size:1;     // 15
 275
 276         BITS depth:1;          // 14
 277
 278         BITS normal:1;
 279
 280         BITS color7:1;
 281         BITS color6:1;
 282         BITS color5:1;
 283         BITS color4:1;
 284         BITS color3:1;
 285         BITS color2:1;
 286         BITS color1:1;
 287         BITS color0:1;
 288
 289         BITS pos_param:1; // 4
 290         BITS clip_dist1:1;
 291         BITS clip_dist0:1;
 292         BITS misc:1;
 293         BITS pos:1;      // 0
 294 #else
 295         BITS pos:1;      // 0
 296         BITS misc:1;
 297         BITS clip_dist0:1;
 298         BITS clip_dist1:1;
 299         BITS pos_param:1; // 4
 300
 301         BITS color0:1;    // 5
 302         BITS color1:1;
 303         BITS color2:1;
 304         BITS color3:1;
 305         BITS color4:1;
 306         BITS color5:1;
 307         BITS color6:1;
 308         BITS color7:1;
 309
 310         BITS normal:1;
 311
 312         BITS depth:1;          // 14
 313
 314         BITS point_size:1;     // 15
 315         BITS edge_flag:1;
 316         BITS rta_index:1;      //     shares same channel as kill_flag
 317         BITS kill_flag:1;
 318         BITS viewport_index:1; // 19
 319
 320         BITS resvd1:12;        // 20
 321 #endif
 322 } VAP_OUT_VTX_FMT_0;
 323
 324 typedef struct VAP_OUT_VTX_FMT_1tag
 325 {
 326 #ifdef MESA_BIG_ENDIAN
 327         BITS resvd:8;
 328
 329         BITS tex7comp:3;
 330         BITS tex6comp:3;
 331         BITS tex5comp:3;
 332         BITS tex4comp:3;
 333         BITS tex3comp:3;
 334         BITS tex2comp:3;
 335         BITS tex1comp:3;
 336         BITS tex0comp:3;
 337 #else
 338         BITS tex0comp:3;
 339         BITS tex1comp:3;
 340         BITS tex2comp:3;
 341         BITS tex3comp:3;
 342         BITS tex4comp:3;
 343         BITS tex5comp:3;
 344         BITS tex6comp:3;
 345         BITS tex7comp:3;
 346
 347         BITS resvd:8;
 348 #endif
 349 } VAP_OUT_VTX_FMT_1;
 350
 351 typedef struct VAP_OUT_VTX_FMT_2tag
 352 {
 353 #ifdef MESA_BIG_ENDIAN
 354         BITS resvd:8;
 355
 356         BITS tex15comp:3;
 357         BITS tex14comp:3;
 358         BITS tex13comp:3;
 359         BITS tex12comp:3;
 360         BITS tex11comp:3;
 361         BITS tex10comp:3;
 362         BITS tex9comp:3;
 363         BITS tex8comp:3;
 364 #else
 365         BITS tex8comp :3;
 366         BITS tex9comp :3;
 367         BITS tex10comp:3;
 368         BITS tex11comp:3;
 369         BITS tex12comp:3;
 370         BITS tex13comp:3;
 371         BITS tex14comp:3;
 372         BITS tex15comp:3;
 373
 374         BITS resvd:8;
 375 #endif
 376 } VAP_OUT_VTX_FMT_2;
 377
 378 typedef struct OUT_FRAGMENT_FMT_0tag
 379 {
 380 #ifdef MESA_BIG_ENDIAN
 381         BITS resvd1:20;
 382
 383         BITS mask:1;
 384         BITS coverage_to_mask:1;
 385         BITS stencil_ref:1;
 386         BITS depth:1;
 387
 388         BITS color7:1;
 389         BITS color6:1;
 390         BITS color5:1;
 391         BITS color4:1;
 392         BITS color3:1;
 393         BITS color2:1;
 394         BITS color1:1;
 395         BITS color0:1;
 396 #else
 397         BITS color0:1;
 398         BITS color1:1;
 399         BITS color2:1;
 400         BITS color3:1;
 401         BITS color4:1;
 402         BITS color5:1;
 403         BITS color6:1;
 404         BITS color7:1;
 405
 406         BITS depth:1;
 407         BITS stencil_ref:1;
 408         BITS coverage_to_mask:1;
 409         BITS mask:1;
 410
 411         BITS resvd1:20;
 412 #endif
 413 } OUT_FRAGMENT_FMT_0;
 414
 415 typedef enum  CF_CLAUSE_TYPE
 416 {
 417    CF_EXPORT_CLAUSE,
 418    CF_ALU_CLAUSE,
 419    CF_TEX_CLAUSE,
 420    CF_VTX_CLAUSE,
 421    CF_OTHER_CLAUSE,
 422    CF_EMPTY_CLAUSE,
 423    NUMBER_CF_CLAUSE_TYPES
 424 } CF_CLAUSE_TYPE;
 425
 426 enum
 427 {
 428     MAX_BOOL_CONSTANTS   = 32,
 429     MAX_INT_CONSTANTS    = 32,
 430     MAX_FLOAT_CONSTANTS  = 256,
 431
 432     FC_NONE = 0,
 433     FC_IF = 1,
 434     FC_LOOP = 2,
 435     FC_REP = 3,
 436     FC_PUSH_VPM = 4,
 437     FC_PUSH_WQM = 5,
 438
 439     COND_NONE = 0,
 440     COND_BOOL = 1,
 441     COND_PRED = 2,
 442     COND_ALU = 3,
 443
 444     SAFEDIST_TEX = 6, ///< safe distance for using result of texture lookup in alu or another tex lookup
 445     SAFEDIST_ALU = 6 ///< the same for alu->fc
 446 };
 447
 448 typedef struct FC_LEVEL
 449 {
 450     R700ControlFlowGenericClause *  first;
 451     R700ControlFlowGenericClause ** mid;
 452     unsigned int unNumMid;
 453     unsigned int midLen;
 454     unsigned int type;
 455     unsigned int cond;
 456     unsigned int inv;
 457     int id; ///< id of bool or int variable
 458 } FC_LEVEL;
 459
 460 typedef struct VTX_FETCH_METHOD
 461 {
 462     GLboolean bEnableMini;
 463     GLuint mega_fetch_remainder;
 464 } VTX_FETCH_METHOD;
 465
 466 typedef struct SUB_OFFSET
 467 {
 468     GLint  subIL_Offset;
 469     GLuint unCFoffset;
 470     GLuint unStackDepthMax;
 471     PRESUB_DESC *   pPresubDesc;
 472     TypedShaderList lstCFInstructions_local;
 473 } SUB_OFFSET;
 474
 475 typedef struct CALLER_POINTER
 476 {
 477     GLint  subIL_Offset;
 478     GLint  subDescIndex;
 479     R700ControlFlowGenericClause* cf_ptr;
 480
 481     R700ControlFlowGenericClause* prelude_cf_ptr;
 482     R700ControlFlowGenericClause* finale_cf_ptr;
 483 } CALLER_POINTER;
 484
 485 #define SQ_MAX_CALL_DEPTH 0x00000020
 486
 487 typedef struct CALL_LEVEL
 488 {
 489     unsigned int      FCSP_BeforeEntry;
 490     GLint             subDescIndex;
 491     GLushort          current;
 492     GLushort          max;
 493     TypedShaderList * plstCFInstructions_local;
 494 } CALL_LEVEL;
 495
 496 #define HAS_CURRENT_LOOPRET 0x1L
 497 #define HAS_LOOPRET         0x2L
 498 #define LOOPRET_FLAGS       HAS_LOOPRET | HAS_CURRENT_LOOPRET
 499
 500 typedef struct r700_AssemblerBase
 501 {
 502         R700ControlFlowSXClause*      cf_last_export_ptr;
 503         R700ControlFlowSXClause*      cf_current_export_clause_ptr;
 504         R700ControlFlowALUClause*     cf_current_alu_clause_ptr;
 505         R700ControlFlowGenericClause* cf_current_tex_clause_ptr;
 506         R700ControlFlowGenericClause* cf_current_vtx_clause_ptr;
 507         R700ControlFlowGenericClause* cf_current_cf_clause_ptr;
 508
 509     //Result shader
 510     R700_Shader * pR700Shader;
 511
 512         // No clause has been created yet
 513         CF_CLAUSE_TYPE cf_current_clause_type;
 514
 515     BITS alu_x_opcode;
 516
 517         GLuint number_of_exports;
 518         GLuint number_of_colorandz_exports;
 519         GLuint number_of_export_opcodes;
 520
 521         PVSDWORD D;
 522     PVSDWORD D2;
 523         PVSDWORD S[3];
 524         PVSDWORD C[4];
 525
 526         unsigned int uLastPosUpdate;
 527         unsigned int last_cond_register;
 528
 529         OUT_FRAGMENT_FMT_0     fp_stOutFmt0;
 530
 531         unsigned int uIIns;
 532         unsigned int uOIns;
 533         unsigned int number_used_registers;
 534         unsigned int uUsedConsts;
 535
 536     unsigned int flag_reg_index;
 537
 538         // Fragment programs
 539         unsigned int uiFP_AttributeMap[FRAG_ATTRIB_MAX];
 540         unsigned int uiFP_OutputMap[FRAG_RESULT_MAX];
 541         unsigned int uBoolConsts;
 542         unsigned int uIntConsts;
 543         unsigned int uInsts;
 544         unsigned int uConsts;
 545
 546         // Vertex programs
 547         unsigned char ucVP_AttributeMap[VERT_ATTRIB_MAX];
 548         unsigned char ucVP_OutputMap[VERT_RESULT_MAX];
 549
 550         //-----------------------------------------------------------------------------------
 551         // flow control members
 552         //-----------------------------------------------------------------------------------
 553         unsigned int FCSP;
 554         FC_LEVEL fc_stack[32];
 555
 556         //-----------------------------------------------------------------------------------
 557         // ArgSubst used in Assemble_Source() function
 558         //-----------------------------------------------------------------------------------
 559         int aArgSubst[4];
 560
 561     GLint hw_gpr[ NUMBER_OF_CYCLES ][ NUMBER_OF_COMPONENTS ];
 562     GLint hw_cfile_addr[ NUMBER_OF_COMPONENTS ];
 563     GLint hw_cfile_chan[ NUMBER_OF_COMPONENTS ];
 564
 565     GLuint uOutputs;
 566
 567     GLint color_export_register_number[NUMBER_OF_OUTPUT_COLORS];
 568         GLint depth_export_register_number;
 569
 570         GLint stencil_export_register_number;
 571         GLint coverage_to_mask_export_register_number;
 572         GLint mask_export_register_number;
 573
 574         GLuint starting_export_register_number;
 575         GLuint starting_vfetch_register_number;
 576         GLuint starting_temp_register_number;
 577         GLuint uHelpReg;
 578         GLuint uFirstHelpReg;
 579
 580         GLboolean input_position_is_used;
 581         GLboolean input_normal_is_used;
 582
 583         GLboolean input_color_is_used[NUMBER_OF_INPUT_COLORS];
 584
 585         GLboolean input_texture_unit_is_used[NUMBER_OF_TEXTURE_UNITS];
 586
 587     R700VertexGenericFetch* vfetch_instruction_ptr_array[VERT_ATTRIB_MAX];
 588
 589         GLuint number_of_inputs;
 590
 591     InstDeps *pInstDeps;
 592
 593     SHADER_PIPE_TYPE currentShaderType;
 594     struct prog_instruction * pILInst;
 595     GLuint             uiCurInst;
 596     GLubyte SamplerUnits[MAX_SAMPLERS];
 597     GLboolean   bR6xx;
 598
 599     /* TODO : merge bR6xx */
 600     GLuint  unAsic;
 601
 602     /* helper to decide which type of instruction to assemble */
 603     GLboolean is_tex;
 604     /* we inserted helper intructions and need barrier on next TEX ins */
 605     GLboolean need_tex_barrier;
 606
 607     SUB_OFFSET     * subs;
 608     GLuint           unSubArraySize;
 609     GLuint           unSubArrayPointer;
 610     CALLER_POINTER * callers;
 611     GLuint           unCallerArraySize;
 612     GLuint           unCallerArrayPointer;
 613     unsigned int     CALLSP;
 614     CALL_LEVEL       CALLSTACK[SQ_MAX_CALL_DEPTH];
 615
 616     GLuint unCFflags;
 617
 618     PRESUB_DESC * presubs;
 619     GLuint        unPresubArraySize;
 620     GLuint        unNumPresub;
 621     GLuint        unCurNumILInsts;
 622
 623     GLuint    unVetTexBits;
 624
 625     GLuint    shadow_regs[R700_MAX_TEXTURE_UNITS];
 626
 627     GLboolean bUseMemConstant;
 628     GLuint    kcacheUsed;
 629
 630 } r700_AssemblerBase;
 631
 632 //Internal use
 633 BITS addrmode_PVSDST(PVSDST * pPVSDST);
 634 void setaddrmode_PVSDST(PVSDST * pPVSDST, BITS addrmode);
 635 void nomask_PVSDST(PVSDST * pPVSDST);
 636 BITS addrmode_PVSSRC(PVSSRC* pPVSSRC);
 637 void setaddrmode_PVSSRC(PVSSRC* pPVSSRC, BITS addrmode);
 638 void setswizzle_PVSSRC(PVSSRC* pPVSSRC, BITS swz);
 639 void noswizzle_PVSSRC(PVSSRC* pPVSSRC);
 640 void swizzleagain_PVSSRC(PVSSRC * pPVSSRC, BITS x, BITS y, BITS z, BITS w);
 641 void neg_PVSSRC(PVSSRC* pPVSSRC);
 642 void noneg_PVSSRC(PVSSRC* pPVSSRC);
 643 void flipneg_PVSSRC(PVSSRC* pPVSSRC);
 644 void zerocomp_PVSSRC(PVSSRC* pPVSSRC, int c);
 645 void onecomp_PVSSRC(PVSSRC* pPVSSRC, int c);
 646 BITS is_misc_component_exported(VAP_OUT_VTX_FMT_0* pOutVTXFmt0);
 647 BITS is_depth_component_exported(OUT_FRAGMENT_FMT_0* pFPOutFmt) ;
 648 GLboolean is_reduction_opcode(PVSDWORD * dest);
 649 GLuint GetSurfaceFormat(GLenum eType, GLuint nChannels, GLuint * pClient_size);
 650
 651 unsigned int r700GetNumOperands(GLuint opcode, GLuint nIsOp3);
 652
 653 unsigned int EG_GetNumOperands(GLuint opcode, GLuint nIsOp3);
 654
 655 GLboolean IsTex(gl_inst_opcode Opcode);
 656 GLboolean IsAlu(gl_inst_opcode Opcode);
 657 int check_current_clause(r700_AssemblerBase* pAsm,
 658                                              CF_CLAUSE_TYPE      new_clause_type);
 659 GLboolean add_vfetch_instruction(r700_AssemblerBase*     pAsm,
 660                                                                  R700VertexInstruction*  vertex_instruction_ptr);
 661 GLboolean add_tex_instruction(r700_AssemblerBase*     pAsm,
 662                               R700TextureInstruction* tex_instruction_ptr);
 663 GLboolean assemble_vfetch_instruction(r700_AssemblerBase* pAsm,
 664                                                                 GLuint gl_client_id,
 665                                 GLuint destination_register,
 666                                                                 GLuint number_of_elements,
 667                                 GLenum dataElementType,
 668                                                                 VTX_FETCH_METHOD* pFetchMethod);
 669 GLboolean assemble_vfetch_instruction2(r700_AssemblerBase* pAsm,
 670                                        GLuint              destination_register,
 671                                        GLenum              type,
 672                                        GLint               size,
 673                                        GLubyte             element,
 674                                        GLuint              _signed,
 675                                        GLboolean           normalize,
 676                                        GLenum              format,
 677                                        VTX_FETCH_METHOD  * pFetchMethod);
 678
 679 GLboolean EG_assemble_vfetch_instruction(r700_AssemblerBase* pAsm,
 680                                        GLuint              destination_register,
 681                                        GLenum              type,
 682                                        GLint               size,
 683                                        GLubyte             element,
 684                                        GLuint              _signed,
 685                                        GLboolean           normalize,
 686                                        GLenum              format,
 687                                        VTX_FETCH_METHOD  * pFetchMethod);
 688 //-----------------------
 689
 690 GLboolean cleanup_vfetch_instructions(r700_AssemblerBase* pAsm);
 691 GLuint gethelpr(r700_AssemblerBase* pAsm);
 692 void resethelpr(r700_AssemblerBase* pAsm);
 693 void checkop_init(r700_AssemblerBase* pAsm);
 694 GLboolean mov_temp(r700_AssemblerBase* pAsm, int src);
 695 GLboolean checkop1(r700_AssemblerBase* pAsm);
 696 GLboolean checkop2(r700_AssemblerBase* pAsm);
 697 GLboolean checkop3(r700_AssemblerBase* pAsm);
 698 GLboolean assemble_src(r700_AssemblerBase *pAsm,
 699                        int src,
 700                        int fld);
 701 GLboolean assemble_dst(r700_AssemblerBase *pAsm);
 702 GLboolean tex_dst(r700_AssemblerBase *pAsm);
 703 GLboolean tex_src(r700_AssemblerBase *pAsm);
 704 GLboolean assemble_tex_instruction(r700_AssemblerBase *pAsm, GLboolean normalized);
 705 void initialize(r700_AssemblerBase *pAsm);
 706 GLboolean assemble_alu_src(R700ALUInstruction*  alu_instruction_ptr,
 707                            int                  source_index,
 708                            PVSSRC*              pSource,
 709                            BITS                 scalar_channel_index,
 710                            r700_AssemblerBase  *pAsm);
 711
 712 GLboolean add_alu_instruction(r700_AssemblerBase* pAsm,
 713                               R700ALUInstruction* alu_instruction_ptr,
 714                               GLuint              contiguous_slots_needed);
 715
 716 GLboolean add_cf_instruction(r700_AssemblerBase* pAsm);
 717 void add_return_inst(r700_AssemblerBase *pAsm);
 718
 719 void get_src_properties(R700ALUInstruction*  alu_instruction_ptr,
 720                         int                  source_index,
 721                         BITS*                psrc_sel,
 722                         BITS*                psrc_rel,
 723                         BITS*                psrc_chan,
 724                         BITS*                psrc_neg);
 725 int is_cfile(BITS sel);
 726 int is_const(BITS sel);
 727 int is_gpr(BITS sel);
 728 GLboolean reserve_cfile(r700_AssemblerBase* pAsm,
 729                         GLuint sel,
 730                         GLuint chan);
 731 GLboolean reserve_gpr(r700_AssemblerBase* pAsm, GLuint sel, GLuint chan, GLuint cycle);
 732 GLboolean cycle_for_scalar_bank_swizzle(const int swiz, const int sel, GLuint* pCycle);
 733 GLboolean cycle_for_vector_bank_swizzle(const int swiz, const int sel, GLuint* pCycle);
 734 GLboolean check_scalar(r700_AssemblerBase* pAsm,
 735                        R700ALUInstruction* alu_instruction_ptr);
 736 GLboolean check_vector(r700_AssemblerBase* pAsm,
 737                        R700ALUInstruction* alu_instruction_ptr);
 738 GLboolean assemble_alu_instruction(r700_AssemblerBase *pAsm);
 739
 740 GLboolean pops(r700_AssemblerBase *pAsm, GLuint pops);
 741 GLboolean jumpToOffest(r700_AssemblerBase *pAsm, GLuint pops, GLint offset);
 742 GLboolean setRetInLoopFlag(r700_AssemblerBase *pAsm, GLuint flagValue);
 743 GLboolean testFlag(r700_AssemblerBase *pAsm);
 744 GLboolean breakLoopOnFlag(r700_AssemblerBase *pAsm, GLuint unFCSP);
 745 GLboolean returnOnFlag(r700_AssemblerBase *pAsm, GLuint unIF);
 746
 747 GLboolean assemble_math_function(r700_AssemblerBase* pAsm, BITS opcode);
 748 GLboolean assemble_ABS(r700_AssemblerBase *pAsm);
 749 GLboolean assemble_ADD(r700_AssemblerBase *pAsm);
 750 GLboolean assemble_ARL(r700_AssemblerBase *pAsm);
 751 GLboolean assemble_BAD(char *opcode_str);
 752 GLboolean assemble_CMP(r700_AssemblerBase *pAsm);
 753 GLboolean assemble_DOT(r700_AssemblerBase *pAsm);
 754 GLboolean assemble_DST(r700_AssemblerBase *pAsm);
 755 GLboolean assemble_EX2(r700_AssemblerBase *pAsm);
 756 GLboolean assemble_EXP(r700_AssemblerBase *pAsm);
 757 GLboolean assemble_FLR(r700_AssemblerBase *pAsm);
 758 GLboolean assemble_FLR_INT(r700_AssemblerBase *pAsm);
 759 GLboolean assemble_FRC(r700_AssemblerBase *pAsm);
 760 GLboolean assemble_KIL(r700_AssemblerBase *pAsm, GLuint opcode);
 761 GLboolean assemble_LG2(r700_AssemblerBase *pAsm);
 762 GLboolean assemble_LRP(r700_AssemblerBase *pAsm);
 763 GLboolean assemble_LOG(r700_AssemblerBase *pAsm);
 764 GLboolean assemble_MAD(r700_AssemblerBase *pAsm);
 765 GLboolean assemble_LIT(r700_AssemblerBase *pAsm);
 766 GLboolean assemble_MAX(r700_AssemblerBase *pAsm);
 767 GLboolean assemble_MIN(r700_AssemblerBase *pAsm);
 768 GLboolean assemble_MOV(r700_AssemblerBase *pAsm);
 769 GLboolean assemble_MUL(r700_AssemblerBase *pAsm);
 770 GLboolean assemble_POW(r700_AssemblerBase *pAsm);
 771 GLboolean assemble_RCP(r700_AssemblerBase *pAsm);
 772 GLboolean assemble_RSQ(r700_AssemblerBase *pAsm);
 773 GLboolean assemble_SCS(r700_AssemblerBase *pAsm);
 774 GLboolean assemble_SGE(r700_AssemblerBase *pAsm);
 775 GLboolean assemble_CONT(r700_AssemblerBase *pAsm);
 776
 777 GLboolean assemble_LOGIC(r700_AssemblerBase *pAsm, BITS opcode);
 778 GLboolean assemble_LOGIC_PRED(r700_AssemblerBase *pAsm, BITS opcode);
 779 GLboolean assemble_TRIG(r700_AssemblerBase *pAsm, BITS opcode);
 780
 781 GLboolean assemble_SLT(r700_AssemblerBase *pAsm);
 782 GLboolean assemble_SSG(r700_AssemblerBase *pAsm);
 783 GLboolean assemble_STP(r700_AssemblerBase *pAsm);
 784 GLboolean assemble_TEX(r700_AssemblerBase *pAsm);
 785 GLboolean assemble_XPD(r700_AssemblerBase *pAsm);
 786 GLboolean assemble_EXPORT(r700_AssemblerBase *pAsm);
 787 GLboolean assemble_IF(r700_AssemblerBase *pAsm, GLboolean bHasElse);
 788 GLboolean assemble_ELSE(r700_AssemblerBase *pAsm);
 789 GLboolean assemble_ENDIF(r700_AssemblerBase *pAsm);
 790
 791 GLboolean assemble_BGNLOOP(r700_AssemblerBase *pAsm);
 792 GLboolean assemble_BRK(r700_AssemblerBase *pAsm);
 793 GLboolean assemble_COND(r700_AssemblerBase *pAsm);
 794 GLboolean assemble_ENDLOOP(r700_AssemblerBase *pAsm);
 795
 796 GLboolean assemble_BGNSUB(r700_AssemblerBase *pAsm, GLint nILindex, GLuint uiIL_Shift);
 797 GLboolean assemble_ENDSUB(r700_AssemblerBase *pAsm);
 798 GLboolean assemble_RET(r700_AssemblerBase *pAsm);
 799 GLboolean assemble_CAL(r700_AssemblerBase *pAsm,
 800                        GLint nILindex,
 801                        GLuint uiIL_Offest,
 802                        GLuint uiNumberInsts,
 803                        struct prog_instruction *pILInst,
 804                        PRESUB_DESC * pPresubDesc);
 805
 806 GLboolean Process_Export(r700_AssemblerBase* pAsm,
 807                          GLuint type,
 808                          GLuint export_starting_index,
 809                          GLuint export_count,
 810                          GLuint starting_register_number,
 811                          GLboolean is_depth_export);
 812 GLboolean Move_Depth_Exports_To_Correct_Channels(r700_AssemblerBase *pAsm,
 813                                                  BITS depth_channel_select);
 814
 815 GLboolean callPreSub(r700_AssemblerBase* pAsm,
 816                      LOADABLE_SCRIPT_SIGNITURE scriptSigniture,
 817                      /* struct prog_instruction ** pILInstParent, */
 818                      COMPILED_SUB * pCompiledSub,
 819                      GLshort uOutReg,
 820                      GLshort uNumValidSrc);
 821 GLboolean EG_add_ps_interp(r700_AssemblerBase* pAsm);
 822
 823 //Interface
 824 GLboolean AssembleInstr(GLuint uiFirstInst,
 825                         GLuint uiIL_Shift,
 826                         GLuint uiNumberInsts,
 827                         struct prog_instruction *pILInst,
 828                                                 r700_AssemblerBase *pR700AsmCode);
 829 GLboolean Process_Fragment_Exports(r700_AssemblerBase *pR700AsmCode, GLbitfield OutputsWritten);
 830 GLboolean Process_Vertex_Exports(r700_AssemblerBase *pR700AsmCode, GLbitfield OutputsWritten);
 831
 832 GLboolean RelocProgram(r700_AssemblerBase * pAsm, struct gl_program * pILProg);
 833 GLboolean InitShaderProgram(r700_AssemblerBase * pAsm);
 834
 835 int       Init_r700_AssemblerBase(SHADER_PIPE_TYPE spt, r700_AssemblerBase* pAsm, R700_Shader* pShader);
 836 GLboolean Clean_Up_Assembler(r700_AssemblerBase *pR700AsmCode);
 837
 838 #endif //_R700_ASSEMBLER_H_