src/mesa/drivers/dri/r600/r700_fragprog.c

   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  *   CooperYuan <cooper.yuan@amd.com>, <cooperyuan@gmail.com>
  26  */
  27
  28 #include <stdio.h>
  29 #include <stdarg.h>
  30 #include <stdlib.h>
  31 #include <string.h>
  32 #include <math.h>
  33
  34 #include "main/imports.h"
  35 #include "program/prog_parameter.h"
  36 #include "program/prog_statevars.h"
  37 #include "program/program.h"
  38
  39 #include "r600_context.h"
  40 #include "r600_cmdbuf.h"
  41 #include "r600_emit.h"
  42
  43 #include "r700_fragprog.h"
  44
  45 #include "r700_debug.h"
  46
  47 void insert_wpos_code(struct gl_context *ctx, struct gl_fragment_program *fprog)
  48 {
  49     static const gl_state_index winstate[STATE_LENGTH]
  50          = { STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM, 0, 0, 0};
  51     struct prog_instruction *newInst, *inst;
  52     GLint  win_size;  /* state reference */
  53     GLuint wpos_temp; /* temp register */
  54     int i, j;
  55
  56     /* PARAM win_size = STATE_FB_WPOS_Y_TRANSFORM */
  57     win_size = _mesa_add_state_reference(fprog->Base.Parameters, winstate);
  58
  59     wpos_temp = fprog->Base.NumTemporaries++;
  60
  61     /* scan program where WPOS is used and replace with wpos_temp */
  62     inst = fprog->Base.Instructions;
  63     for (i = 0; i < fprog->Base.NumInstructions; i++) {
  64         for (j=0; j < 3; j++) {
  65             if(inst->SrcReg[j].File == PROGRAM_INPUT &&
  66                inst->SrcReg[j].Index == FRAG_ATTRIB_WPOS) {
  67                 inst->SrcReg[j].File = PROGRAM_TEMPORARY;
  68                 inst->SrcReg[j].Index = wpos_temp;
  69             }
  70         }
  71         inst++;
  72     }
  73
  74     _mesa_insert_instructions(&(fprog->Base), 0, 1);
  75
  76     newInst = fprog->Base.Instructions;
  77     /* possibly invert wpos.y depending on STATE_FB_WPOS_Y_TRANSFORM var */
  78     newInst[0].Opcode = OPCODE_MAD;
  79     newInst[0].DstReg.File = PROGRAM_TEMPORARY;
  80     newInst[0].DstReg.Index = wpos_temp;
  81     newInst[0].DstReg.WriteMask = WRITEMASK_XYZW;
  82
  83     newInst[0].SrcReg[0].File = PROGRAM_INPUT;
  84     newInst[0].SrcReg[0].Index = FRAG_ATTRIB_WPOS;
  85     newInst[0].SrcReg[0].Swizzle = SWIZZLE_XYZW;
  86
  87     newInst[0].SrcReg[1].File = PROGRAM_STATE_VAR;
  88     newInst[0].SrcReg[1].Index = win_size;
  89     newInst[0].SrcReg[1].Swizzle = MAKE_SWIZZLE4(SWIZZLE_ONE, SWIZZLE_X, SWIZZLE_ONE, SWIZZLE_ONE);
  90
  91     newInst[0].SrcReg[2].File = PROGRAM_STATE_VAR;
  92     newInst[0].SrcReg[2].Index = win_size;
  93     newInst[0].SrcReg[2].Swizzle = MAKE_SWIZZLE4(SWIZZLE_ZERO, SWIZZLE_Y, SWIZZLE_ZERO, SWIZZLE_ZERO);
  94
  95 }
  96
  97 //TODO : Validate FP input with VP output.
  98 void Map_Fragment_Program(r700_AssemblerBase         *pAsm,
  99                                                   struct gl_fragment_program *mesa_fp,
 100                           struct gl_context *ctx)
 101 {
 102         unsigned int unBit;
 103     unsigned int i;
 104
 105     /* match fp inputs with vp exports. */
 106     struct r700_vertex_program_cont *vpc =
 107                        (struct r700_vertex_program_cont *)ctx->VertexProgram._Current;
 108     GLbitfield OutputsWritten = vpc->mesa_program.Base.OutputsWritten;
 109
 110         pAsm->number_used_registers = 0;
 111
 112 //Input mapping : mesa_fp->Base.InputsRead set the flag, set in
 113         //The flags parsed in parse_attrib_binding. FRAG_ATTRIB_COLx, FRAG_ATTRIB_TEXx, ...
 114         //MUST match order in Map_Vertex_Output
 115         unBit = 1 << FRAG_ATTRIB_WPOS;
 116         if(mesa_fp->Base.InputsRead & unBit)
 117         {
 118                 pAsm->uiFP_AttributeMap[FRAG_ATTRIB_WPOS] = pAsm->number_used_registers++;
 119         }
 120
 121     unBit = 1 << VERT_RESULT_COL0;
 122         if(OutputsWritten & unBit)
 123         {
 124                 pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL0] = pAsm->number_used_registers++;
 125         }
 126
 127         unBit = 1 << VERT_RESULT_COL1;
 128         if(OutputsWritten & unBit)
 129         {
 130                 pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL1] = pAsm->number_used_registers++;
 131         }
 132
 133     unBit = 1 << VERT_RESULT_FOGC;
 134     if(OutputsWritten & unBit)
 135     {
 136         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FOGC] = pAsm->number_used_registers++;
 137     }
 138
 139         for(i=0; i<8; i++)
 140         {
 141                 unBit = 1 << (VERT_RESULT_TEX0 + i);
 142                 if(OutputsWritten & unBit)
 143                 {
 144                         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_TEX0 + i] = pAsm->number_used_registers++;
 145                 }
 146         }
 147
 148 /* order has been taken care of */
 149 #if 1
 150     for(i=VERT_RESULT_VAR0; i<VERT_RESULT_MAX; i++)
 151         {
 152         unBit = 1 << i;
 153         if(OutputsWritten & unBit)
 154                 {
 155             pAsm->uiFP_AttributeMap[i-VERT_RESULT_VAR0+FRAG_ATTRIB_VAR0] = pAsm->number_used_registers++;
 156         }
 157     }
 158 #else
 159     if( (mesa_fp->Base.InputsRead >> FRAG_ATTRIB_VAR0) > 0 )
 160     {
 161             struct r700_vertex_program_cont *vpc =
 162                        (struct r700_vertex_program_cont *)ctx->VertexProgram._Current;
 163         struct gl_program_parameter_list * VsVarying = vpc->mesa_program.Base.Varying;
 164         struct gl_program_parameter_list * PsVarying = mesa_fp->Base.Varying;
 165         struct gl_program_parameter      * pVsParam;
 166         struct gl_program_parameter      * pPsParam;
 167         GLuint j, k;
 168         GLuint unMaxVarying = 0;
 169
 170         for(i=0; i<VsVarying->NumParameters; i++)
 171         {
 172             pAsm->uiFP_AttributeMap[i + FRAG_ATTRIB_VAR0] = 0;
 173         }
 174
 175         for(i=FRAG_ATTRIB_VAR0; i<FRAG_ATTRIB_MAX; i++)
 176             {
 177             unBit = 1 << i;
 178             if(mesa_fp->Base.InputsRead & unBit)
 179                     {
 180                 j = i - FRAG_ATTRIB_VAR0;
 181                 pPsParam = PsVarying->Parameters + j;
 182
 183                 for(k=0; k<VsVarying->NumParameters; k++)
 184                 {
 185                     pVsParam = VsVarying->Parameters + k;
 186
 187                                 if( strcmp(pPsParam->Name, pVsParam->Name) == 0)
 188                     {
 189                         pAsm->uiFP_AttributeMap[i] = pAsm->number_used_registers + k;
 190                         if(k > unMaxVarying)
 191                         {
 192                             unMaxVarying = k;
 193                         }
 194                         break;
 195                     }
 196                 }
 197                     }
 198         }
 199
 200         pAsm->number_used_registers += unMaxVarying + 1;
 201     }
 202 #endif
 203     unBit = 1 << FRAG_ATTRIB_FACE;
 204     if(mesa_fp->Base.InputsRead & unBit)
 205     {
 206         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FACE] = pAsm->number_used_registers++;
 207     }
 208
 209     unBit = 1 << FRAG_ATTRIB_PNTC;
 210     if(mesa_fp->Base.InputsRead & unBit)
 211     {
 212         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_PNTC] = pAsm->number_used_registers++;
 213     }
 214
 215 /* Map temporary registers (GPRs) */
 216     pAsm->starting_temp_register_number = pAsm->number_used_registers;
 217
 218     if(mesa_fp->Base.NumNativeTemporaries >= mesa_fp->Base.NumTemporaries)
 219     {
 220             pAsm->number_used_registers += mesa_fp->Base.NumNativeTemporaries;
 221     }
 222     else
 223     {
 224         pAsm->number_used_registers += mesa_fp->Base.NumTemporaries;
 225     }
 226
 227 /* Output mapping */
 228         pAsm->number_of_exports = 0;
 229         pAsm->number_of_colorandz_exports = 0; /* don't include stencil and mask out. */
 230         pAsm->starting_export_register_number = pAsm->number_used_registers;
 231
 232     for (i = 0; i < FRAG_RESULT_MAX; ++i)
 233     {
 234         unBit = 1 << i;
 235         if (mesa_fp->Base.OutputsWritten & unBit)
 236         {
 237             if (i == FRAG_RESULT_DEPTH)
 238             {
 239                 pAsm->depth_export_register_number = pAsm->number_used_registers;
 240                 pAsm->pR700Shader->depthIsExported = 1;
 241             }
 242
 243             pAsm->uiFP_OutputMap[i] = pAsm->number_used_registers++;
 244             ++pAsm->number_of_exports;
 245             ++pAsm->number_of_colorandz_exports;
 246         }
 247     }
 248
 249     pAsm->flag_reg_index = pAsm->number_used_registers++;
 250
 251     pAsm->uFirstHelpReg = pAsm->number_used_registers;
 252 }
 253
 254 GLboolean Find_Instruction_Dependencies_fp(struct r700_fragment_program *fp,
 255                                                                 struct gl_fragment_program   *mesa_fp)
 256 {
 257     GLuint i, j;
 258     GLint * puiTEMPwrites;
 259     GLint * puiTEMPreads;
 260     struct prog_instruction * pILInst;
 261     InstDeps         *pInstDeps;
 262     struct prog_instruction * texcoord_DepInst;
 263     GLint              nDepInstID;
 264
 265     puiTEMPwrites = (GLint*) MALLOC(sizeof(GLuint)*mesa_fp->Base.NumTemporaries);
 266     puiTEMPreads = (GLint*) MALLOC(sizeof(GLuint)*mesa_fp->Base.NumTemporaries);
 267
 268     for(i=0; i<mesa_fp->Base.NumTemporaries; i++)
 269     {
 270         puiTEMPwrites[i] = -1;
 271         puiTEMPreads[i] = -1;
 272     }
 273
 274     pInstDeps = (InstDeps*)MALLOC(sizeof(InstDeps)*mesa_fp->Base.NumInstructions);
 275
 276     for(i=0; i<mesa_fp->Base.NumInstructions; i++)
 277     {
 278         pInstDeps[i].nDstDep = -1;
 279         pILInst = &(mesa_fp->Base.Instructions[i]);
 280
 281         //Dst
 282         if(pILInst->DstReg.File == PROGRAM_TEMPORARY)
 283         {
 284             //Set lastwrite for the temp
 285             puiTEMPwrites[pILInst->DstReg.Index] = i;
 286         }
 287
 288         //Src
 289         for(j=0; j<3; j++)
 290         {
 291             if(pILInst->SrcReg[j].File == PROGRAM_TEMPORARY)
 292             {
 293                 //Set dep.
 294                 pInstDeps[i].nSrcDeps[j] = puiTEMPwrites[pILInst->SrcReg[j].Index];
 295                 //Set first read
 296                 if(puiTEMPreads[pILInst->SrcReg[j].Index] < 0 )
 297                 {
 298                     puiTEMPreads[pILInst->SrcReg[j].Index] = i;
 299                 }
 300             }
 301             else
 302             {
 303                 pInstDeps[i].nSrcDeps[j] = -1;
 304             }
 305         }
 306     }
 307
 308     fp->r700AsmCode.pInstDeps = pInstDeps;
 309
 310     //Find dep for tex inst
 311     for(i=0; i<mesa_fp->Base.NumInstructions; i++)
 312     {
 313         pILInst = &(mesa_fp->Base.Instructions[i]);
 314
 315         if(GL_TRUE == IsTex(pILInst->Opcode))
 316         {   //src0 is the tex coord register, src1 is texunit, src2 is textype
 317             nDepInstID = pInstDeps[i].nSrcDeps[0];
 318             if(nDepInstID >= 0)
 319             {
 320                 texcoord_DepInst = &(mesa_fp->Base.Instructions[nDepInstID]);
 321                 if(GL_TRUE == IsAlu(texcoord_DepInst->Opcode) )
 322                 {
 323                     pInstDeps[nDepInstID].nDstDep = i;
 324                     pInstDeps[i].nDstDep = i;
 325                 }
 326                 else if(GL_TRUE == IsTex(texcoord_DepInst->Opcode) )
 327                 {
 328                     pInstDeps[i].nDstDep = i;
 329                 }
 330                 else
 331                 {   //... other deps?
 332                 }
 333             }
 334             // make sure that we dont overwrite src used earlier
 335             nDepInstID = puiTEMPreads[pILInst->DstReg.Index];
 336             if(nDepInstID < i)
 337             {
 338                 pInstDeps[i].nDstDep = puiTEMPreads[pILInst->DstReg.Index];
 339                 texcoord_DepInst = &(mesa_fp->Base.Instructions[nDepInstID]);
 340                 if(GL_TRUE == IsAlu(texcoord_DepInst->Opcode) )
 341                 {
 342                     pInstDeps[nDepInstID].nDstDep = i;
 343                 }
 344
 345             }
 346
 347         }
 348         }
 349
 350     FREE(puiTEMPwrites);
 351     FREE(puiTEMPreads);
 352
 353     return GL_TRUE;
 354 }
 355
 356 GLboolean r700TranslateFragmentShader(struct r700_fragment_program *fp,
 357                                                              struct gl_fragment_program   *mesa_fp,
 358                                  struct gl_context *ctx)
 359 {
 360     context_t *context = R700_CONTEXT(ctx);
 361     R700_CHIP_CONTEXT *r700 = (R700_CHIP_CONTEXT*)(&context->hw);
 362
 363         GLuint    number_of_colors_exported;
 364         GLboolean z_enabled = GL_FALSE;
 365         GLuint    unBit, shadow_unit;
 366         int i;
 367         struct prog_instruction *inst;
 368         gl_state_index shadow_ambient[STATE_LENGTH]
 369             = { STATE_INTERNAL, STATE_SHADOW_AMBIENT, 0, 0, 0};
 370
 371     //Init_Program
 372         Init_r700_AssemblerBase( SPT_FP, &(fp->r700AsmCode), &(fp->r700Shader) );
 373
 374     if(GL_TRUE == r700->bShaderUseMemConstant)
 375     {
 376         fp->r700AsmCode.bUseMemConstant = GL_TRUE;
 377     }
 378     else
 379     {
 380         fp->r700AsmCode.bUseMemConstant = GL_FALSE;
 381     }
 382
 383     fp->r700AsmCode.unAsic = 7;
 384
 385     if(mesa_fp->Base.InputsRead & FRAG_BIT_WPOS)
 386     {
 387         insert_wpos_code(ctx, mesa_fp);
 388     }
 389
 390     /* add/map  consts for ARB_shadow_ambient */
 391     if(mesa_fp->Base.ShadowSamplers)
 392     {
 393         inst = mesa_fp->Base.Instructions;
 394         for (i = 0; i < mesa_fp->Base.NumInstructions; i++)
 395         {
 396             if(inst->TexShadow == 1)
 397             {
 398                 shadow_unit = inst->TexSrcUnit;
 399                 shadow_ambient[2] = shadow_unit;
 400                 fp->r700AsmCode.shadow_regs[shadow_unit] =
 401                     _mesa_add_state_reference(mesa_fp->Base.Parameters, shadow_ambient);
 402             }
 403             inst++;
 404         }
 405     }
 406
 407     Map_Fragment_Program(&(fp->r700AsmCode), mesa_fp, ctx);
 408
 409     if( GL_FALSE == Find_Instruction_Dependencies_fp(fp, mesa_fp) )
 410         {
 411                 return GL_FALSE;
 412     }
 413
 414     InitShaderProgram(&(fp->r700AsmCode));
 415
 416     for(i=0; i < MAX_SAMPLERS; i++)
 417     {
 418          fp->r700AsmCode.SamplerUnits[i] = fp->mesa_program.Base.SamplerUnits[i];
 419     }
 420
 421     fp->r700AsmCode.unCurNumILInsts = mesa_fp->Base.NumInstructions;
 422
 423         if( GL_FALSE == AssembleInstr(0,
 424                                   0,
 425                                   mesa_fp->Base.NumInstructions,
 426                                   &(mesa_fp->Base.Instructions[0]),
 427                                   &(fp->r700AsmCode)) )
 428         {
 429                 return GL_FALSE;
 430         }
 431
 432     if(GL_FALSE == Process_Fragment_Exports(&(fp->r700AsmCode), mesa_fp->Base.OutputsWritten) )
 433     {
 434         return GL_FALSE;
 435     }
 436
 437     if( GL_FALSE == RelocProgram(&(fp->r700AsmCode), &(mesa_fp->Base)) )
 438     {
 439         return GL_FALSE;
 440     }
 441
 442     fp->r700Shader.nRegs = (fp->r700AsmCode.number_used_registers == 0) ? 0
 443                          : (fp->r700AsmCode.number_used_registers - 1);
 444
 445         fp->r700Shader.nParamExports = fp->r700AsmCode.number_of_exports;
 446
 447         number_of_colors_exported = fp->r700AsmCode.number_of_colorandz_exports;
 448
 449         unBit = 1 << FRAG_RESULT_DEPTH;
 450         if(mesa_fp->Base.OutputsWritten & unBit)
 451         {
 452                 z_enabled = GL_TRUE;
 453                 number_of_colors_exported--;
 454         }
 455
 456         /* illegal to set this to 0 */
 457         if(number_of_colors_exported || z_enabled)
 458         {
 459             fp->r700Shader.exportMode = number_of_colors_exported << 1 | z_enabled;
 460         }
 461         else
 462         {
 463             fp->r700Shader.exportMode = (1 << 1);
 464         }
 465
 466     fp->translated = GL_TRUE;
 467
 468         return GL_TRUE;
 469 }
 470
 471 void r700SelectFragmentShader(struct gl_context *ctx)
 472 {
 473     context_t *context = R700_CONTEXT(ctx);
 474     struct r700_fragment_program *fp = (struct r700_fragment_program *)
 475             (ctx->FragmentProgram._Current);
 476     if (context->radeon.radeonScreen->chip_family < CHIP_FAMILY_RV770)
 477     {
 478             fp->r700AsmCode.bR6xx = 1;
 479     }
 480
 481     if (GL_FALSE == fp->translated)
 482             r700TranslateFragmentShader(fp, &(fp->mesa_program), ctx);
 483 }
 484
 485 void * r700GetActiveFpShaderBo(struct gl_context * ctx)
 486 {
 487     struct r700_fragment_program *fp = (struct r700_fragment_program *)
 488                                            (ctx->FragmentProgram._Current);
 489
 490     return fp->shaderbo;
 491 }
 492
 493 void * r700GetActiveFpShaderConstBo(struct gl_context * ctx)
 494 {
 495     struct r700_fragment_program *fp = (struct r700_fragment_program *)
 496                                            (ctx->FragmentProgram._Current);
 497
 498     return fp->constbo0;
 499 }
 500
 501 GLboolean r700SetupFragmentProgram(struct gl_context * ctx)
 502 {
 503     context_t *context = R700_CONTEXT(ctx);
 504     R700_CHIP_CONTEXT *r700 = (R700_CHIP_CONTEXT*)(&context->hw);
 505     struct r700_fragment_program *fp = (struct r700_fragment_program *)
 506                                            (ctx->FragmentProgram._Current);
 507     r700_AssemblerBase         *pAsm = &(fp->r700AsmCode);
 508     struct gl_fragment_program *mesa_fp = &(fp->mesa_program);
 509     struct gl_program_parameter_list *paramList;
 510     unsigned int unNumParamData;
 511     unsigned int ui, i;
 512     unsigned int unNumOfReg;
 513     unsigned int unBit;
 514     unsigned int num_sq_ps_gprs;
 515     GLuint exportCount;
 516     GLboolean point_sprite = GL_FALSE;
 517
 518     if(GL_FALSE == fp->loaded)
 519     {
 520             if(fp->r700Shader.bNeedsAssembly == GL_TRUE)
 521             {
 522                     Assemble( &(fp->r700Shader) );
 523             }
 524
 525         /* Load fp to gpu */
 526         r600EmitShader(ctx,
 527                        &(fp->shaderbo),
 528                        (GLvoid *)(fp->r700Shader.pProgram),
 529                        fp->r700Shader.uShaderBinaryDWORDSize,
 530                        "FS");
 531
 532         fp->loaded = GL_TRUE;
 533     }
 534
 535     DumpHwBinary(DUMP_PIXEL_SHADER, (GLvoid *)(fp->r700Shader.pProgram),
 536                  fp->r700Shader.uShaderBinaryDWORDSize);
 537
 538     /* TODO : enable this after MemUse fixed *=
 539     (context->chipobj.MemUse)(context, fp->shadercode.buf->id);
 540     */
 541
 542     R600_STATECHANGE(context, ps);
 543
 544     r700->ps.SQ_PGM_RESOURCES_PS.u32All = 0;
 545     SETbit(r700->ps.SQ_PGM_RESOURCES_PS.u32All, PGM_RESOURCES__PRIME_CACHE_ON_DRAW_bit);
 546
 547     r700->ps.SQ_PGM_START_PS.u32All = 0; /* set from buffer obj */
 548
 549     R600_STATECHANGE(context, spi);
 550
 551     unNumOfReg = fp->r700Shader.nRegs + 1;
 552
 553     ui = (r700->SPI_PS_IN_CONTROL_0.u32All & NUM_INTERP_mask) / (1 << NUM_INTERP_shift);
 554
 555     /* PS uses fragment.position */
 556     if (mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_WPOS))
 557     {
 558         ui += 1;
 559         SETfield(r700->SPI_PS_IN_CONTROL_0.u32All, ui, NUM_INTERP_shift, NUM_INTERP_mask);
 560         SETfield(r700->SPI_PS_IN_CONTROL_0.u32All, CENTERS_ONLY, BARYC_SAMPLE_CNTL_shift, BARYC_SAMPLE_CNTL_mask);
 561         SETbit(r700->SPI_PS_IN_CONTROL_0.u32All, POSITION_ENA_bit);
 562         SETbit(r700->SPI_INPUT_Z.u32All, PROVIDE_Z_TO_SPI_bit);
 563     }
 564     else
 565     {
 566         CLEARbit(r700->SPI_PS_IN_CONTROL_0.u32All, POSITION_ENA_bit);
 567         CLEARbit(r700->SPI_INPUT_Z.u32All, PROVIDE_Z_TO_SPI_bit);
 568     }
 569
 570     if (mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_FACE))
 571     {
 572         ui += 1;
 573         SETfield(r700->SPI_PS_IN_CONTROL_0.u32All, ui, NUM_INTERP_shift, NUM_INTERP_mask);
 574         SETbit(r700->SPI_PS_IN_CONTROL_1.u32All, FRONT_FACE_ENA_bit);
 575         SETbit(r700->SPI_PS_IN_CONTROL_1.u32All, FRONT_FACE_ALL_BITS_bit);
 576         SETfield(r700->SPI_PS_IN_CONTROL_1.u32All, pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FACE], FRONT_FACE_ADDR_shift, FRONT_FACE_ADDR_mask);
 577     }
 578     else
 579     {
 580         CLEARbit(r700->SPI_PS_IN_CONTROL_1.u32All, FRONT_FACE_ENA_bit);
 581     }
 582
 583     /* see if we need any point_sprite replacements, also increase num_interp
 584      * as there's no vp output for them */
 585     if (ctx->Point.PointSprite)
 586     {
 587         for (i = FRAG_ATTRIB_TEX0; i<= FRAG_ATTRIB_TEX7; i++)
 588         {
 589             if (ctx->Point.CoordReplace[i - FRAG_ATTRIB_TEX0] == GL_TRUE)
 590             {
 591                 ui++;
 592                 point_sprite = GL_TRUE;
 593             }
 594         }
 595     }
 596
 597     if( mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_PNTC))
 598         ui++;
 599
 600     if ((mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_PNTC)) || point_sprite)
 601     {
 602         SETfield(r700->SPI_PS_IN_CONTROL_0.u32All, ui, NUM_INTERP_shift, NUM_INTERP_mask);
 603         SETbit(r700->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_ENA_bit);
 604         SETfield(r700->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_S, PNT_SPRITE_OVRD_X_shift, PNT_SPRITE_OVRD_X_mask);
 605         SETfield(r700->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_T, PNT_SPRITE_OVRD_Y_shift, PNT_SPRITE_OVRD_Y_mask);
 606         SETfield(r700->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_0, PNT_SPRITE_OVRD_Z_shift, PNT_SPRITE_OVRD_Z_mask);
 607         SETfield(r700->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_1, PNT_SPRITE_OVRD_W_shift, PNT_SPRITE_OVRD_W_mask);
 608         /* Like e.g. viewport and winding, point sprite coordinates are
 609          * inverted when rendering to FBO. */
 610         if ((ctx->Point.SpriteOrigin == GL_LOWER_LEFT) == !ctx->DrawBuffer->Name)
 611             SETbit(r700->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_TOP_1_bit);
 612         else
 613             CLEARbit(r700->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_TOP_1_bit);
 614     }
 615     else
 616     {
 617         CLEARbit(r700->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_ENA_bit);
 618     }
 619
 620
 621     ui = (unNumOfReg < ui) ? ui : unNumOfReg;
 622
 623     SETfield(r700->ps.SQ_PGM_RESOURCES_PS.u32All, ui, NUM_GPRS_shift, NUM_GPRS_mask);
 624
 625     num_sq_ps_gprs = ((r700->sq_config.SQ_GPR_RESOURCE_MGMT_1.u32All & NUM_PS_GPRS_mask) >> NUM_PS_GPRS_shift);
 626
 627     if(ui > num_sq_ps_gprs)
 628     {
 629         /* care! thich changes sq - needs idle state */
 630         R600_STATECHANGE(context, sq);
 631         SETfield(r700->sq_config.SQ_GPR_RESOURCE_MGMT_1.u32All, ui, NUM_PS_GPRS_shift, NUM_PS_GPRS_mask);
 632     }
 633
 634     CLEARbit(r700->ps.SQ_PGM_RESOURCES_PS.u32All, UNCACHED_FIRST_INST_bit);
 635
 636     if(fp->r700Shader.uStackSize) /* we don't use branch for now, it should be zero. */
 637         {
 638         SETfield(r700->ps.SQ_PGM_RESOURCES_PS.u32All, fp->r700Shader.uStackSize,
 639                  STACK_SIZE_shift, STACK_SIZE_mask);
 640     }
 641
 642     SETfield(r700->ps.SQ_PGM_EXPORTS_PS.u32All, fp->r700Shader.exportMode,
 643              EXPORT_MODE_shift, EXPORT_MODE_mask);
 644
 645     // emit ps input map
 646     struct r700_vertex_program_cont *vpc =
 647                        (struct r700_vertex_program_cont *)ctx->VertexProgram._Current;
 648     GLbitfield OutputsWritten = vpc->mesa_program.Base.OutputsWritten;
 649
 650     for(ui = 0; ui < R700_MAX_SHADER_EXPORTS; ui++)
 651         r700->SPI_PS_INPUT_CNTL[ui].u32All = 0;
 652
 653     unBit = 1 << FRAG_ATTRIB_WPOS;
 654     if(mesa_fp->Base.InputsRead & unBit)
 655     {
 656             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_WPOS];
 657             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 658             SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 659                      SEMANTIC_shift, SEMANTIC_mask);
 660             CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 661     }
 662
 663     unBit = 1 << VERT_RESULT_COL0;
 664     if(OutputsWritten & unBit)
 665     {
 666             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL0];
 667             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 668             SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 669                      SEMANTIC_shift, SEMANTIC_mask);
 670             if (r700->SPI_INTERP_CONTROL_0.u32All & FLAT_SHADE_ENA_bit)
 671                     SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 672             else
 673                     CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 674     }
 675
 676     unBit = 1 << VERT_RESULT_COL1;
 677     if(OutputsWritten & unBit)
 678     {
 679             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL1];
 680             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 681             SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 682                      SEMANTIC_shift, SEMANTIC_mask);
 683             if (r700->SPI_INTERP_CONTROL_0.u32All & FLAT_SHADE_ENA_bit)
 684                     SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 685             else
 686                     CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 687     }
 688
 689     unBit = 1 << VERT_RESULT_FOGC;
 690     if(OutputsWritten & unBit)
 691     {
 692             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FOGC];
 693             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 694             SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 695                      SEMANTIC_shift, SEMANTIC_mask);
 696             CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 697     }
 698
 699     for(i=0; i<8; i++)
 700     {
 701             GLboolean coord_replace = ctx->Point.PointSprite && ctx->Point.CoordReplace[i];
 702             unBit = 1 << (VERT_RESULT_TEX0 + i);
 703             if ((OutputsWritten & unBit) || coord_replace)
 704             {
 705                     ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_TEX0 + i];
 706                     SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 707                     SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 708                              SEMANTIC_shift, SEMANTIC_mask);
 709                     CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 710                     /* ARB_point_sprite */
 711                     if (coord_replace)
 712                     {
 713                              SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, PT_SPRITE_TEX_bit);
 714                     }
 715             }
 716     }
 717
 718     unBit = 1 << FRAG_ATTRIB_FACE;
 719     if(mesa_fp->Base.InputsRead & unBit)
 720     {
 721             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FACE];
 722             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 723             SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 724                      SEMANTIC_shift, SEMANTIC_mask);
 725             CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 726     }
 727     unBit = 1 << FRAG_ATTRIB_PNTC;
 728     if(mesa_fp->Base.InputsRead & unBit)
 729     {
 730             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_PNTC];
 731             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 732             SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 733                      SEMANTIC_shift, SEMANTIC_mask);
 734             if (r700->SPI_INTERP_CONTROL_0.u32All & FLAT_SHADE_ENA_bit)
 735                     SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 736             else
 737                     CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 738             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, PT_SPRITE_TEX_bit);
 739     }
 740
 741
 742
 743
 744     for(i=VERT_RESULT_VAR0; i<VERT_RESULT_MAX; i++)
 745         {
 746         unBit = 1 << i;
 747         if(OutputsWritten & unBit)
 748                 {
 749             ui = pAsm->uiFP_AttributeMap[i-VERT_RESULT_VAR0+FRAG_ATTRIB_VAR0];
 750             SETbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 751             SETfield(r700->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 752                              SEMANTIC_shift, SEMANTIC_mask);
 753             CLEARbit(r700->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 754         }
 755     }
 756
 757     exportCount = (r700->ps.SQ_PGM_EXPORTS_PS.u32All & EXPORT_MODE_mask) / (1 << EXPORT_MODE_shift);
 758     if (r700->CB_SHADER_CONTROL.u32All != ((1 << exportCount) - 1))
 759     {
 760             R600_STATECHANGE(context, cb);
 761             r700->CB_SHADER_CONTROL.u32All = (1 << exportCount) - 1;
 762     }
 763
 764     /* sent out shader constants. */
 765     paramList = fp->mesa_program.Base.Parameters;
 766
 767     if(NULL != paramList)
 768     {
 769             _mesa_load_state_parameters(ctx, paramList);
 770
 771             if (paramList->NumParameters > R700_MAX_DX9_CONSTS)
 772                     return GL_FALSE;
 773
 774             R600_STATECHANGE(context, ps_consts);
 775
 776             r700->ps.num_consts = paramList->NumParameters;
 777
 778             unNumParamData = paramList->NumParameters;
 779
 780             for(ui=0; ui<unNumParamData; ui++) {
 781                         r700->ps.consts[ui][0].f32All = paramList->ParameterValues[ui][0];
 782                         r700->ps.consts[ui][1].f32All = paramList->ParameterValues[ui][1];
 783                         r700->ps.consts[ui][2].f32All = paramList->ParameterValues[ui][2];
 784                         r700->ps.consts[ui][3].f32All = paramList->ParameterValues[ui][3];
 785             }
 786
 787         /* Load fp constants to gpu */
 788         if( (GL_TRUE == r700->bShaderUseMemConstant) && (unNumParamData > 0) )
 789         {
 790             r600EmitShader(ctx,
 791                            &(fp->constbo0),
 792                            (GLvoid *)&(paramList->ParameterValues[0][0]),
 793                            unNumParamData * 4,
 794                            "FS Const");
 795         }
 796
 797     } else
 798             r700->ps.num_consts = 0;
 799
 800     COMPILED_SUB * pCompiledSub;
 801     GLuint uj;
 802     GLuint unConstOffset = r700->ps.num_consts;
 803     for(ui=0; ui<pAsm->unNumPresub; ui++)
 804     {
 805         pCompiledSub = pAsm->presubs[ui].pCompiledSub;
 806
 807         r700->ps.num_consts += pCompiledSub->NumParameters;
 808
 809         for(uj=0; uj<pCompiledSub->NumParameters; uj++)
 810         {
 811             r700->ps.consts[uj + unConstOffset][0].f32All = pCompiledSub->ParameterValues[uj][0];
 812                     r700->ps.consts[uj + unConstOffset][1].f32All = pCompiledSub->ParameterValues[uj][1];
 813                     r700->ps.consts[uj + unConstOffset][2].f32All = pCompiledSub->ParameterValues[uj][2];
 814                     r700->ps.consts[uj + unConstOffset][3].f32All = pCompiledSub->ParameterValues[uj][3];
 815         }
 816         unConstOffset += pCompiledSub->NumParameters;
 817     }
 818
 819     return GL_TRUE;
 820 }
 821