src/mesa/drivers/dri/r600/evergreen_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
  36 #include "program/prog_parameter.h"
  37 #include "program/prog_statevars.h"
  38 #include "program/program.h"
  39
  40 #include "r600_context.h"
  41 #include "r600_cmdbuf.h"
  42 #include "r600_emit.h"
  43
  44 #include "evergreen_vertprog.h"
  45 #include "evergreen_fragprog.h"
  46
  47 void evergreen_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_SIZE, 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_SIZE */
  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     /* invert wpos.y
  78      * wpos_temp.xyzw = wpos.x-yzw + winsize.0y00 */
  79     newInst[0].Opcode = OPCODE_ADD;
  80     newInst[0].DstReg.File = PROGRAM_TEMPORARY;
  81     newInst[0].DstReg.Index = wpos_temp;
  82     newInst[0].DstReg.WriteMask = WRITEMASK_XYZW;
  83
  84     newInst[0].SrcReg[0].File = PROGRAM_INPUT;
  85     newInst[0].SrcReg[0].Index = FRAG_ATTRIB_WPOS;
  86     newInst[0].SrcReg[0].Swizzle = SWIZZLE_XYZW;
  87     newInst[0].SrcReg[0].Negate = NEGATE_Y;
  88
  89     newInst[0].SrcReg[1].File = PROGRAM_STATE_VAR;
  90     newInst[0].SrcReg[1].Index = win_size;
  91     newInst[0].SrcReg[1].Swizzle = MAKE_SWIZZLE4(SWIZZLE_ZERO, SWIZZLE_Y, SWIZZLE_ZERO, SWIZZLE_ZERO);
  92
  93 }
  94
  95 //TODO : Validate FP input with VP output.
  96 void evergreen_Map_Fragment_Program(r700_AssemblerBase         *pAsm,
  97                                                   struct gl_fragment_program *mesa_fp,
  98                           struct gl_context *ctx)
  99 {
 100         unsigned int unBit;
 101     unsigned int i;
 102
 103     /* match fp inputs with vp exports. */
 104     struct evergreen_vertex_program_cont *vpc =
 105                        (struct evergreen_vertex_program_cont *)ctx->VertexProgram._Current;
 106     GLbitfield OutputsWritten = vpc->mesa_program.Base.OutputsWritten;
 107
 108         pAsm->number_used_registers = 0;
 109
 110 //Input mapping : mesa_fp->Base.InputsRead set the flag, set in
 111         //The flags parsed in parse_attrib_binding. FRAG_ATTRIB_COLx, FRAG_ATTRIB_TEXx, ...
 112         //MUST match order in Map_Vertex_Output
 113         unBit = 1 << FRAG_ATTRIB_WPOS;
 114         if(mesa_fp->Base.InputsRead & unBit)
 115         {
 116                 pAsm->uiFP_AttributeMap[FRAG_ATTRIB_WPOS] = pAsm->number_used_registers++;
 117         }
 118
 119     unBit = 1 << VERT_RESULT_COL0;
 120         if(OutputsWritten & unBit)
 121         {
 122                 pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL0] = pAsm->number_used_registers++;
 123         }
 124
 125         unBit = 1 << VERT_RESULT_COL1;
 126         if(OutputsWritten & unBit)
 127         {
 128                 pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL1] = pAsm->number_used_registers++;
 129         }
 130
 131     unBit = 1 << VERT_RESULT_FOGC;
 132     if(OutputsWritten & unBit)
 133     {
 134         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FOGC] = pAsm->number_used_registers++;
 135     }
 136
 137         for(i=0; i<8; i++)
 138         {
 139                 unBit = 1 << (VERT_RESULT_TEX0 + i);
 140                 if(OutputsWritten & unBit)
 141                 {
 142                         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_TEX0 + i] = pAsm->number_used_registers++;
 143                 }
 144         }
 145
 146 /* order has been taken care of */
 147 #if 1
 148     for(i=VERT_RESULT_VAR0; i<VERT_RESULT_MAX; i++)
 149         {
 150         unBit = 1 << i;
 151         if(OutputsWritten & unBit)
 152                 {
 153             pAsm->uiFP_AttributeMap[i-VERT_RESULT_VAR0+FRAG_ATTRIB_VAR0] = pAsm->number_used_registers++;
 154         }
 155     }
 156 #else
 157     if( (mesa_fp->Base.InputsRead >> FRAG_ATTRIB_VAR0) > 0 )
 158     {
 159             struct evergreen_vertex_program_cont *vpc =
 160                        (struct evergreen_vertex_program_cont *)ctx->VertexProgram._Current;
 161         struct gl_program_parameter_list * VsVarying = vpc->mesa_program.Base.Varying;
 162         struct gl_program_parameter_list * PsVarying = mesa_fp->Base.Varying;
 163         struct gl_program_parameter      * pVsParam;
 164         struct gl_program_parameter      * pPsParam;
 165         GLuint j, k;
 166         GLuint unMaxVarying = 0;
 167
 168         for(i=0; i<VsVarying->NumParameters; i++)
 169         {
 170             pAsm->uiFP_AttributeMap[i + FRAG_ATTRIB_VAR0] = 0;
 171         }
 172
 173         for(i=FRAG_ATTRIB_VAR0; i<FRAG_ATTRIB_MAX; i++)
 174             {
 175             unBit = 1 << i;
 176             if(mesa_fp->Base.InputsRead & unBit)
 177                     {
 178                 j = i - FRAG_ATTRIB_VAR0;
 179                 pPsParam = PsVarying->Parameters + j;
 180
 181                 for(k=0; k<VsVarying->NumParameters; k++)
 182                 {
 183                     pVsParam = VsVarying->Parameters + k;
 184
 185                                 if( strcmp(pPsParam->Name, pVsParam->Name) == 0)
 186                     {
 187                         pAsm->uiFP_AttributeMap[i] = pAsm->number_used_registers + k;
 188                         if(k > unMaxVarying)
 189                         {
 190                             unMaxVarying = k;
 191                         }
 192                         break;
 193                     }
 194                 }
 195                     }
 196         }
 197
 198         pAsm->number_used_registers += unMaxVarying + 1;
 199     }
 200 #endif
 201     unBit = 1 << FRAG_ATTRIB_FACE;
 202     if(mesa_fp->Base.InputsRead & unBit)
 203     {
 204         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FACE] = pAsm->number_used_registers++;
 205     }
 206
 207     unBit = 1 << FRAG_ATTRIB_PNTC;
 208     if(mesa_fp->Base.InputsRead & unBit)
 209     {
 210         pAsm->uiFP_AttributeMap[FRAG_ATTRIB_PNTC] = pAsm->number_used_registers++;
 211     }
 212
 213     pAsm->uIIns = pAsm->number_used_registers;
 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         unBit = 1 << FRAG_RESULT_COLOR;
 232         if(mesa_fp->Base.OutputsWritten & unBit)
 233         {
 234                 pAsm->uiFP_OutputMap[FRAG_RESULT_COLOR] = pAsm->number_used_registers++;
 235                 pAsm->number_of_exports++;
 236                 pAsm->number_of_colorandz_exports++;
 237         }
 238         unBit = 1 << FRAG_RESULT_DEPTH;
 239         if(mesa_fp->Base.OutputsWritten & unBit)
 240         {
 241         pAsm->depth_export_register_number = pAsm->number_used_registers;
 242                 pAsm->uiFP_OutputMap[FRAG_RESULT_DEPTH] = pAsm->number_used_registers++;
 243                 pAsm->number_of_exports++;
 244                 pAsm->number_of_colorandz_exports++;
 245                 pAsm->pR700Shader->depthIsExported = 1;
 246         }
 247
 248     pAsm->flag_reg_index = pAsm->number_used_registers++;
 249
 250     pAsm->uFirstHelpReg = pAsm->number_used_registers;
 251 }
 252
 253 GLboolean evergreen_Find_Instruction_Dependencies_fp(struct evergreen_fragment_program *fp,
 254                                                                 struct gl_fragment_program   *mesa_fp)
 255 {
 256     GLuint i, j;
 257     GLint * puiTEMPwrites;
 258     GLint * puiTEMPreads;
 259     struct prog_instruction * pILInst;
 260     InstDeps         *pInstDeps;
 261     struct prog_instruction * texcoord_DepInst;
 262     GLint              nDepInstID;
 263
 264     puiTEMPwrites = (GLint*) MALLOC(sizeof(GLuint)*mesa_fp->Base.NumTemporaries);
 265     puiTEMPreads = (GLint*) MALLOC(sizeof(GLuint)*mesa_fp->Base.NumTemporaries);
 266
 267     for(i=0; i<mesa_fp->Base.NumTemporaries; i++)
 268     {
 269         puiTEMPwrites[i] = -1;
 270         puiTEMPreads[i] = -1;
 271     }
 272
 273     pInstDeps = (InstDeps*)MALLOC(sizeof(InstDeps)*mesa_fp->Base.NumInstructions);
 274
 275     for(i=0; i<mesa_fp->Base.NumInstructions; i++)
 276     {
 277         pInstDeps[i].nDstDep = -1;
 278         pILInst = &(mesa_fp->Base.Instructions[i]);
 279
 280         //Dst
 281         if(pILInst->DstReg.File == PROGRAM_TEMPORARY)
 282         {
 283             //Set lastwrite for the temp
 284             puiTEMPwrites[pILInst->DstReg.Index] = i;
 285         }
 286
 287         //Src
 288         for(j=0; j<3; j++)
 289         {
 290             if(pILInst->SrcReg[j].File == PROGRAM_TEMPORARY)
 291             {
 292                 //Set dep.
 293                 pInstDeps[i].nSrcDeps[j] = puiTEMPwrites[pILInst->SrcReg[j].Index];
 294                 //Set first read
 295                 if(puiTEMPreads[pILInst->SrcReg[j].Index] < 0 )
 296                 {
 297                     puiTEMPreads[pILInst->SrcReg[j].Index] = i;
 298                 }
 299             }
 300             else
 301             {
 302                 pInstDeps[i].nSrcDeps[j] = -1;
 303             }
 304         }
 305     }
 306
 307     fp->r700AsmCode.pInstDeps = pInstDeps;
 308
 309     //Find dep for tex inst
 310     for(i=0; i<mesa_fp->Base.NumInstructions; i++)
 311     {
 312         pILInst = &(mesa_fp->Base.Instructions[i]);
 313
 314         if(GL_TRUE == IsTex(pILInst->Opcode))
 315         {   //src0 is the tex coord register, src1 is texunit, src2 is textype
 316             nDepInstID = pInstDeps[i].nSrcDeps[0];
 317             if(nDepInstID >= 0)
 318             {
 319                 texcoord_DepInst = &(mesa_fp->Base.Instructions[nDepInstID]);
 320                 if(GL_TRUE == IsAlu(texcoord_DepInst->Opcode) )
 321                 {
 322                     pInstDeps[nDepInstID].nDstDep = i;
 323                     pInstDeps[i].nDstDep = i;
 324                 }
 325                 else if(GL_TRUE == IsTex(texcoord_DepInst->Opcode) )
 326                 {
 327                     pInstDeps[i].nDstDep = i;
 328                 }
 329                 else
 330                 {   //... other deps?
 331                 }
 332             }
 333             // make sure that we dont overwrite src used earlier
 334             nDepInstID = puiTEMPreads[pILInst->DstReg.Index];
 335             if(nDepInstID < i)
 336             {
 337                 pInstDeps[i].nDstDep = puiTEMPreads[pILInst->DstReg.Index];
 338                 texcoord_DepInst = &(mesa_fp->Base.Instructions[nDepInstID]);
 339                 if(GL_TRUE == IsAlu(texcoord_DepInst->Opcode) )
 340                 {
 341                     pInstDeps[nDepInstID].nDstDep = i;
 342                 }
 343
 344             }
 345
 346         }
 347         }
 348
 349     FREE(puiTEMPwrites);
 350     FREE(puiTEMPreads);
 351
 352     return GL_TRUE;
 353 }
 354
 355 GLboolean evergreenTranslateFragmentShader(struct evergreen_fragment_program *fp,
 356                                                              struct gl_fragment_program   *mesa_fp,
 357                                  struct gl_context *ctx)
 358 {
 359         GLuint    number_of_colors_exported;
 360         GLboolean z_enabled = GL_FALSE;
 361         GLuint    unBit, shadow_unit;
 362         int i;
 363         struct prog_instruction *inst;
 364         gl_state_index shadow_ambient[STATE_LENGTH]
 365             = { STATE_INTERNAL, STATE_SHADOW_AMBIENT, 0, 0, 0};
 366
 367     //Init_Program
 368         Init_r700_AssemblerBase( SPT_FP, &(fp->r700AsmCode), &(fp->r700Shader) );
 369
 370     fp->constbo0 = NULL;
 371     fp->r700AsmCode.bUseMemConstant = GL_TRUE;
 372     fp->r700AsmCode.unAsic = 8;
 373
 374     if(mesa_fp->Base.InputsRead & FRAG_BIT_WPOS)
 375     {
 376         evergreen_insert_wpos_code(ctx, mesa_fp);
 377     }
 378
 379     /* add/map  consts for ARB_shadow_ambient */
 380     if(mesa_fp->Base.ShadowSamplers)
 381     {
 382         inst = mesa_fp->Base.Instructions;
 383         for (i = 0; i < mesa_fp->Base.NumInstructions; i++)
 384         {
 385             if(inst->TexShadow == 1)
 386             {
 387                 shadow_unit = inst->TexSrcUnit;
 388                 shadow_ambient[2] = shadow_unit;
 389                 fp->r700AsmCode.shadow_regs[shadow_unit] =
 390                     _mesa_add_state_reference(mesa_fp->Base.Parameters, shadow_ambient);
 391             }
 392             inst++;
 393         }
 394     }
 395
 396     evergreen_Map_Fragment_Program(&(fp->r700AsmCode), mesa_fp, ctx);
 397
 398     if( GL_FALSE == evergreen_Find_Instruction_Dependencies_fp(fp, mesa_fp) )
 399         {
 400                 return GL_FALSE;
 401     }
 402
 403     InitShaderProgram(&(fp->r700AsmCode));
 404
 405     for(i=0; i < MAX_SAMPLERS; i++)
 406     {
 407          fp->r700AsmCode.SamplerUnits[i] = fp->mesa_program.Base.SamplerUnits[i];
 408     }
 409
 410     fp->r700AsmCode.unCurNumILInsts = mesa_fp->Base.NumInstructions;
 411
 412         if( GL_FALSE == AssembleInstr(0,
 413                                   0,
 414                                   mesa_fp->Base.NumInstructions,
 415                                   &(mesa_fp->Base.Instructions[0]),
 416                                   &(fp->r700AsmCode)) )
 417         {
 418                 return GL_FALSE;
 419         }
 420
 421     if(GL_FALSE == Process_Fragment_Exports(&(fp->r700AsmCode), mesa_fp->Base.OutputsWritten) )
 422     {
 423         return GL_FALSE;
 424     }
 425
 426     if( GL_FALSE == RelocProgram(&(fp->r700AsmCode), &(mesa_fp->Base)) )
 427     {
 428         return GL_FALSE;
 429     }
 430
 431     fp->r700Shader.nRegs = (fp->r700AsmCode.number_used_registers == 0) ? 0
 432                          : (fp->r700AsmCode.number_used_registers - 1);
 433
 434         fp->r700Shader.nParamExports = fp->r700AsmCode.number_of_exports;
 435
 436         number_of_colors_exported = fp->r700AsmCode.number_of_colorandz_exports;
 437
 438         unBit = 1 << FRAG_RESULT_DEPTH;
 439         if(mesa_fp->Base.OutputsWritten & unBit)
 440         {
 441                 z_enabled = GL_TRUE;
 442                 number_of_colors_exported--;
 443         }
 444
 445         /* illegal to set this to 0 */
 446         if(number_of_colors_exported || z_enabled)
 447         {
 448             fp->r700Shader.exportMode = number_of_colors_exported << 1 | z_enabled;
 449         }
 450         else
 451         {
 452             fp->r700Shader.exportMode = (1 << 1);
 453         }
 454
 455     fp->translated = GL_TRUE;
 456
 457         return GL_TRUE;
 458 }
 459
 460 void evergreenSelectFragmentShader(struct gl_context *ctx)
 461 {
 462     context_t *context = EVERGREEN_CONTEXT(ctx);
 463     struct evergreen_fragment_program *fp = (struct evergreen_fragment_program *)
 464             (ctx->FragmentProgram._Current);
 465     if (context->radeon.radeonScreen->chip_family < CHIP_FAMILY_RV770)
 466     {
 467             fp->r700AsmCode.bR6xx = 1;
 468     }
 469
 470     if (GL_FALSE == fp->translated)
 471             evergreenTranslateFragmentShader(fp, &(fp->mesa_program), ctx);
 472 }
 473
 474 void * evergreenGetActiveFpShaderBo(struct gl_context * ctx)
 475 {
 476     struct evergreen_fragment_program *fp = (struct evergreen_fragment_program *)
 477                                            (ctx->FragmentProgram._Current);
 478
 479     return fp->shaderbo;
 480 }
 481
 482 void * evergreenGetActiveFpShaderConstBo(struct gl_context * ctx)
 483 {
 484     struct evergreen_fragment_program *fp = (struct evergreen_fragment_program *)
 485                                            (ctx->FragmentProgram._Current);
 486
 487     return fp->constbo0;
 488 }
 489
 490 GLboolean evergreenSetupFragmentProgram(struct gl_context * ctx)
 491 {
 492     context_t *context = EVERGREEN_CONTEXT(ctx);
 493     EVERGREEN_CHIP_CONTEXT *evergreen = GET_EVERGREEN_CHIP(context);
 494     struct evergreen_fragment_program *fp = (struct evergreen_fragment_program *)
 495                                            (ctx->FragmentProgram._Current);
 496     r700_AssemblerBase         *pAsm = &(fp->r700AsmCode);
 497     struct gl_fragment_program *mesa_fp = &(fp->mesa_program);
 498     unsigned int ui, i;
 499     unsigned int unNumOfReg;
 500     unsigned int unBit;
 501     GLuint exportCount;
 502     GLboolean point_sprite = GL_FALSE;
 503
 504     if(GL_FALSE == fp->loaded)
 505     {
 506             if(fp->r700Shader.bNeedsAssembly == GL_TRUE)
 507             {
 508                     Assemble( &(fp->r700Shader) );
 509             }
 510
 511         r600EmitShader(ctx,
 512                        &(fp->shaderbo),
 513                        (GLvoid *)(fp->r700Shader.pProgram),
 514                        fp->r700Shader.uShaderBinaryDWORDSize,
 515                        "FS");
 516
 517         fp->loaded = GL_TRUE;
 518     }
 519
 520     /* TODO : enable this after MemUse fixed *=
 521     (context->chipobj.MemUse)(context, fp->shadercode.buf->id);
 522     */
 523
 524     EVERGREEN_STATECHANGE(context, sq);
 525
 526     evergreen->SQ_PGM_RESOURCES_PS.u32All = 0;
 527     SETbit(evergreen->SQ_PGM_RESOURCES_PS.u32All, PGM_RESOURCES__PRIME_CACHE_ON_DRAW_bit);
 528
 529     evergreen->ps.SQ_ALU_CONST_CACHE_PS_0.u32All = 0;
 530     evergreen->ps.SQ_PGM_START_PS.u32All = 0;
 531
 532     EVERGREEN_STATECHANGE(context, spi);
 533
 534     unNumOfReg = fp->r700Shader.nRegs + 1;
 535
 536     ui = (evergreen->SPI_PS_IN_CONTROL_0.u32All & NUM_INTERP_mask) / (1 << NUM_INTERP_shift);
 537
 538     /* PS uses fragment.position */
 539     if (mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_WPOS))
 540     {
 541         ui += 1;
 542         SETfield(evergreen->SPI_PS_IN_CONTROL_0.u32All, ui, NUM_INTERP_shift, NUM_INTERP_mask);
 543         SETfield(evergreen->SPI_PS_IN_CONTROL_0.u32All, CENTERS_ONLY, BARYC_SAMPLE_CNTL_shift, BARYC_SAMPLE_CNTL_mask);
 544         SETbit(evergreen->SPI_PS_IN_CONTROL_0.u32All, POSITION_ENA_bit);
 545         SETbit(evergreen->SPI_INPUT_Z.u32All, PROVIDE_Z_TO_SPI_bit);
 546     }
 547     else
 548     {
 549         CLEARbit(evergreen->SPI_PS_IN_CONTROL_0.u32All, POSITION_ENA_bit);
 550         CLEARbit(evergreen->SPI_INPUT_Z.u32All, PROVIDE_Z_TO_SPI_bit);
 551     }
 552
 553     if (mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_FACE))
 554     {
 555         ui += 1;
 556         SETfield(evergreen->SPI_PS_IN_CONTROL_0.u32All, ui, NUM_INTERP_shift, NUM_INTERP_mask);
 557         SETbit(evergreen->SPI_PS_IN_CONTROL_1.u32All, FRONT_FACE_ENA_bit);
 558         SETbit(evergreen->SPI_PS_IN_CONTROL_1.u32All, FRONT_FACE_ALL_BITS_bit);
 559         SETfield(evergreen->SPI_PS_IN_CONTROL_1.u32All, pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FACE], FRONT_FACE_ADDR_shift, FRONT_FACE_ADDR_mask);
 560     }
 561     else
 562     {
 563         CLEARbit(evergreen->SPI_PS_IN_CONTROL_1.u32All, FRONT_FACE_ENA_bit);
 564     }
 565
 566     /* see if we need any point_sprite replacements */
 567     for (i = VERT_RESULT_TEX0; i<= VERT_RESULT_TEX7; i++)
 568     {
 569         if(ctx->Point.CoordReplace[i - VERT_RESULT_TEX0] == GL_TRUE)
 570             point_sprite = GL_TRUE;
 571     }
 572
 573     if ((mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_PNTC)) || point_sprite)
 574     {
 575         /* for FRAG_ATTRIB_PNTC we need to increase num_interp */
 576         if(mesa_fp->Base.InputsRead & (1 << FRAG_ATTRIB_PNTC))
 577         {
 578             ui++;
 579             SETfield(evergreen->SPI_PS_IN_CONTROL_0.u32All, ui, NUM_INTERP_shift, NUM_INTERP_mask);
 580         }
 581         SETbit(evergreen->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_ENA_bit);
 582         SETfield(evergreen->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_S, PNT_SPRITE_OVRD_X_shift, PNT_SPRITE_OVRD_X_mask);
 583         SETfield(evergreen->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_T, PNT_SPRITE_OVRD_Y_shift, PNT_SPRITE_OVRD_Y_mask);
 584         SETfield(evergreen->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_0, PNT_SPRITE_OVRD_Z_shift, PNT_SPRITE_OVRD_Z_mask);
 585         SETfield(evergreen->SPI_INTERP_CONTROL_0.u32All, SPI_PNT_SPRITE_SEL_1, PNT_SPRITE_OVRD_W_shift, PNT_SPRITE_OVRD_W_mask);
 586         if(ctx->Point.SpriteOrigin == GL_LOWER_LEFT)
 587             SETbit(evergreen->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_TOP_1_bit);
 588         else
 589             CLEARbit(evergreen->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_TOP_1_bit);
 590     }
 591     else
 592     {
 593         CLEARbit(evergreen->SPI_INTERP_CONTROL_0.u32All, PNT_SPRITE_ENA_bit);
 594     }
 595
 596
 597     ui = (unNumOfReg < ui) ? ui : unNumOfReg;
 598
 599     SETfield(evergreen->SQ_PGM_RESOURCES_PS.u32All, ui, NUM_GPRS_shift, NUM_GPRS_mask);
 600
 601     CLEARbit(evergreen->SQ_PGM_RESOURCES_PS.u32All, UNCACHED_FIRST_INST_bit);
 602
 603     if(fp->r700Shader.uStackSize) /* we don't use branch for now, it should be zero. */
 604         {
 605         SETfield(evergreen->SQ_PGM_RESOURCES_PS.u32All, fp->r700Shader.uStackSize,
 606                  STACK_SIZE_shift, STACK_SIZE_mask);
 607     }
 608
 609     SETfield(evergreen->SQ_PGM_EXPORTS_PS.u32All, fp->r700Shader.exportMode,
 610              EXPORT_MODE_shift, EXPORT_MODE_mask);
 611
 612     // emit ps input map
 613     struct evergreen_vertex_program_cont *vpc =
 614                        (struct evergreen_vertex_program_cont *)ctx->VertexProgram._Current;
 615     GLbitfield OutputsWritten = vpc->mesa_program.Base.OutputsWritten;
 616
 617     for(ui = 0; ui < EVERGREEN_MAX_SHADER_EXPORTS; ui++)
 618         evergreen->SPI_PS_INPUT_CNTL[ui].u32All = 0;
 619
 620     unBit = 1 << FRAG_ATTRIB_WPOS;
 621     if(mesa_fp->Base.InputsRead & unBit)
 622     {
 623             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_WPOS];
 624             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 625             SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 626                      SEMANTIC_shift, SEMANTIC_mask);
 627             CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 628     }
 629
 630     unBit = 1 << VERT_RESULT_COL0;
 631     if(OutputsWritten & unBit)
 632     {
 633             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL0];
 634             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 635             SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 636                      SEMANTIC_shift, SEMANTIC_mask);
 637             if (evergreen->SPI_INTERP_CONTROL_0.u32All & FLAT_SHADE_ENA_bit)
 638                     SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 639             else
 640                     CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 641     }
 642
 643     unBit = 1 << VERT_RESULT_COL1;
 644     if(OutputsWritten & unBit)
 645     {
 646             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_COL1];
 647             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 648             SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 649                      SEMANTIC_shift, SEMANTIC_mask);
 650             if (evergreen->SPI_INTERP_CONTROL_0.u32All & FLAT_SHADE_ENA_bit)
 651                     SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 652             else
 653                     CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 654     }
 655
 656     unBit = 1 << VERT_RESULT_FOGC;
 657     if(OutputsWritten & unBit)
 658     {
 659             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FOGC];
 660             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 661             SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 662                      SEMANTIC_shift, SEMANTIC_mask);
 663             CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 664     }
 665
 666     for(i=0; i<8; i++)
 667     {
 668             unBit = 1 << (VERT_RESULT_TEX0 + i);
 669             if(OutputsWritten & unBit)
 670             {
 671                     ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_TEX0 + i];
 672                     SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 673                     SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 674                              SEMANTIC_shift, SEMANTIC_mask);
 675                     CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 676                     /* ARB_point_sprite */
 677                     if(ctx->Point.CoordReplace[i] == GL_TRUE)
 678                     {
 679                              SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, PT_SPRITE_TEX_bit);
 680                     }
 681             }
 682     }
 683
 684     unBit = 1 << FRAG_ATTRIB_FACE;
 685     if(mesa_fp->Base.InputsRead & unBit)
 686     {
 687             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_FACE];
 688             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 689             SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 690                      SEMANTIC_shift, SEMANTIC_mask);
 691             CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 692     }
 693     unBit = 1 << FRAG_ATTRIB_PNTC;
 694     if(mesa_fp->Base.InputsRead & unBit)
 695     {
 696             ui = pAsm->uiFP_AttributeMap[FRAG_ATTRIB_PNTC];
 697             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 698             SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 699                      SEMANTIC_shift, SEMANTIC_mask);
 700             CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 701             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, PT_SPRITE_TEX_bit);
 702     }
 703
 704
 705
 706
 707     for(i=VERT_RESULT_VAR0; i<VERT_RESULT_MAX; i++)
 708         {
 709         unBit = 1 << i;
 710         if(OutputsWritten & unBit)
 711                 {
 712             ui = pAsm->uiFP_AttributeMap[i-VERT_RESULT_VAR0+FRAG_ATTRIB_VAR0];
 713             SETbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, SEL_CENTROID_bit);
 714             SETfield(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, ui,
 715                              SEMANTIC_shift, SEMANTIC_mask);
 716             CLEARbit(evergreen->SPI_PS_INPUT_CNTL[ui].u32All, FLAT_SHADE_bit);
 717         }
 718     }
 719
 720     exportCount = (evergreen->SQ_PGM_EXPORTS_PS.u32All & EXPORT_MODE_mask) / (1 << EXPORT_MODE_shift);
 721
 722     return GL_TRUE;
 723 }
 724
 725 GLboolean evergreenSetupFPconstants(struct gl_context * ctx)
 726 {
 727     context_t *context = EVERGREEN_CONTEXT(ctx);
 728     EVERGREEN_CHIP_CONTEXT *evergreen = GET_EVERGREEN_CHIP(context);
 729     struct evergreen_fragment_program *fp = (struct evergreen_fragment_program *)
 730                                            (ctx->FragmentProgram._Current);
 731     r700_AssemblerBase *pAsm = &(fp->r700AsmCode);
 732
 733     struct gl_program_parameter_list *paramList;
 734     unsigned int unNumParamData;
 735     unsigned int ui;
 736     int alloc_size;
 737
 738     /* sent out shader constants. */
 739     paramList = fp->mesa_program.Base.Parameters;
 740
 741     if(NULL != paramList)
 742     {
 743             _mesa_load_state_parameters(ctx, paramList);
 744
 745             if (paramList->NumParameters > EVERGREEN_MAX_DX9_CONSTS)
 746                     return GL_FALSE;
 747
 748             EVERGREEN_STATECHANGE(context, sq);
 749
 750             evergreen->ps.num_consts = paramList->NumParameters;
 751
 752             unNumParamData = paramList->NumParameters;
 753
 754             for(ui=0; ui<unNumParamData; ui++) {
 755                         evergreen->ps.consts[ui][0].f32All = paramList->ParameterValues[ui][0];
 756                         evergreen->ps.consts[ui][1].f32All = paramList->ParameterValues[ui][1];
 757                         evergreen->ps.consts[ui][2].f32All = paramList->ParameterValues[ui][2];
 758                         evergreen->ps.consts[ui][3].f32All = paramList->ParameterValues[ui][3];
 759             }
 760
 761             /* alloc multiple of 16 constants */
 762             alloc_size = ((unNumParamData * 4 * 4) + 255) & ~255;
 763
 764         /* Load fp constants to gpu */
 765         if(unNumParamData > 0)
 766         {
 767             radeonAllocDmaRegion(&context->radeon,
 768                                 &context->fp_Constbo,
 769                                 &context->fp_bo_offset,
 770                                 alloc_size,
 771                                 256);
 772             r600EmitShaderConsts(ctx,
 773                                  context->fp_Constbo,
 774                                  context->fp_bo_offset,
 775                                  (GLvoid *)&(evergreen->ps.consts[0][0]),
 776                                  unNumParamData * 4 * 4);
 777         }
 778     } else
 779             evergreen->ps.num_consts = 0;
 780
 781     COMPILED_SUB * pCompiledSub;
 782     GLuint uj;
 783     GLuint unConstOffset = evergreen->ps.num_consts;
 784     for(ui=0; ui<pAsm->unNumPresub; ui++)
 785     {
 786         pCompiledSub = pAsm->presubs[ui].pCompiledSub;
 787
 788         evergreen->ps.num_consts += pCompiledSub->NumParameters;
 789
 790         for(uj=0; uj<pCompiledSub->NumParameters; uj++)
 791         {
 792             evergreen->ps.consts[uj + unConstOffset][0].f32All = pCompiledSub->ParameterValues[uj][0];
 793                     evergreen->ps.consts[uj + unConstOffset][1].f32All = pCompiledSub->ParameterValues[uj][1];
 794                     evergreen->ps.consts[uj + unConstOffset][2].f32All = pCompiledSub->ParameterValues[uj][2];
 795                     evergreen->ps.consts[uj + unConstOffset][3].f32All = pCompiledSub->ParameterValues[uj][3];
 796         }
 797         unConstOffset += pCompiledSub->NumParameters;
 798     }
 799
 800     return GL_TRUE;
 801 }
 802