src/mesa/drivers/dri/r300/r500_fragprog.c

   1 /*
   2  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  *
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining
   7  * a copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sublicense, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial
  16  * portions of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  */
  27
  28 #include "r500_fragprog.h"
  29
  30 #include "radeon_nqssadce.h"
  31 #include "radeon_program_alu.h"
  32
  33
  34 static void reset_srcreg(struct prog_src_register* reg)
  35 {
  36         _mesa_bzero(reg, sizeof(*reg));
  37         reg->Swizzle = SWIZZLE_NOOP;
  38 }
  39
  40 static struct prog_src_register shadow_ambient(struct gl_program *program, int tmu)
  41 {
  42         gl_state_index fail_value_tokens[STATE_LENGTH] = {
  43                 STATE_INTERNAL, STATE_SHADOW_AMBIENT, 0, 0, 0
  44         };
  45         struct prog_src_register reg = { 0, };
  46
  47         fail_value_tokens[2] = tmu;
  48         reg.File = PROGRAM_STATE_VAR;
  49         reg.Index = _mesa_add_state_reference(program->Parameters, fail_value_tokens);
  50         reg.Swizzle = SWIZZLE_WWWW;
  51         return reg;
  52 }
  53
  54 /**
  55  * Transform TEX, TXP, TXB, and KIL instructions in the following way:
  56  *  - premultiply texture coordinates for RECT
  57  *  - extract operand swizzles
  58  *  - introduce a temporary register when write masks are needed
  59  *
  60  */
  61 static GLboolean transform_TEX(
  62         struct radeon_transform_context *t,
  63         struct prog_instruction* orig_inst, void* data)
  64 {
  65         struct r500_fragment_program_compiler *compiler =
  66                 (struct r500_fragment_program_compiler*)data;
  67         struct prog_instruction inst = *orig_inst;
  68         struct prog_instruction* tgt;
  69         GLboolean destredirect = GL_FALSE;
  70
  71         if (inst.Opcode != OPCODE_TEX &&
  72             inst.Opcode != OPCODE_TXB &&
  73             inst.Opcode != OPCODE_TXP &&
  74             inst.Opcode != OPCODE_KIL)
  75                 return GL_FALSE;
  76
  77         /* ARB_shadow & EXT_shadow_funcs */
  78         if (inst.Opcode != OPCODE_KIL &&
  79             t->Program->ShadowSamplers & (1 << inst.TexSrcUnit)) {
  80                 GLuint comparefunc = GL_NEVER + compiler->fp->state.unit[inst.TexSrcUnit].texture_compare_func;
  81
  82                 if (comparefunc == GL_NEVER || comparefunc == GL_ALWAYS) {
  83                         tgt = radeonAppendInstructions(t->Program, 1);
  84
  85                         tgt->Opcode = OPCODE_MOV;
  86                         tgt->DstReg = inst.DstReg;
  87                         if (comparefunc == GL_ALWAYS) {
  88                                 tgt->SrcReg[0].File = PROGRAM_BUILTIN;
  89                                 tgt->SrcReg[0].Swizzle = SWIZZLE_1111;
  90                         } else {
  91                                 tgt->SrcReg[0] = shadow_ambient(t->Program, inst.TexSrcUnit);
  92                         }
  93                         return GL_TRUE;
  94                 }
  95
  96                 inst.DstReg.File = PROGRAM_TEMPORARY;
  97                 inst.DstReg.Index = radeonFindFreeTemporary(t);
  98                 inst.DstReg.WriteMask = WRITEMASK_XYZW;
  99         } else if (inst.Opcode != OPCODE_KIL && inst.DstReg.File != PROGRAM_TEMPORARY) {
 100                 int tempreg = radeonFindFreeTemporary(t);
 101
 102                 inst.DstReg.File = PROGRAM_TEMPORARY;
 103                 inst.DstReg.Index = tempreg;
 104                 inst.DstReg.WriteMask = WRITEMASK_XYZW;
 105                 destredirect = GL_TRUE;
 106         }
 107
 108         if (inst.SrcReg[0].File != PROGRAM_TEMPORARY && inst.SrcReg[0].File != PROGRAM_INPUT) {
 109                 int tmpreg = radeonFindFreeTemporary(t);
 110                 tgt = radeonAppendInstructions(t->Program, 1);
 111                 tgt->Opcode = OPCODE_MOV;
 112                 tgt->DstReg.File = PROGRAM_TEMPORARY;
 113                 tgt->DstReg.Index = tmpreg;
 114                 tgt->SrcReg[0] = inst.SrcReg[0];
 115
 116                 reset_srcreg(&inst.SrcReg[0]);
 117                 inst.SrcReg[0].File = PROGRAM_TEMPORARY;
 118                 inst.SrcReg[0].Index = tmpreg;
 119         }
 120
 121         tgt = radeonAppendInstructions(t->Program, 1);
 122         _mesa_copy_instructions(tgt, &inst, 1);
 123
 124         if (inst.Opcode != OPCODE_KIL &&
 125             t->Program->ShadowSamplers & (1 << inst.TexSrcUnit)) {
 126                 GLuint comparefunc = GL_NEVER + compiler->fp->state.unit[inst.TexSrcUnit].texture_compare_func;
 127                 GLuint depthmode = compiler->fp->state.unit[inst.TexSrcUnit].depth_texture_mode;
 128                 int rcptemp = radeonFindFreeTemporary(t);
 129                 int pass, fail;
 130
 131                 tgt = radeonAppendInstructions(t->Program, 3);
 132
 133                 tgt[0].Opcode = OPCODE_RCP;
 134                 tgt[0].DstReg.File = PROGRAM_TEMPORARY;
 135                 tgt[0].DstReg.Index = rcptemp;
 136                 tgt[0].DstReg.WriteMask = WRITEMASK_W;
 137                 tgt[0].SrcReg[0] = inst.SrcReg[0];
 138                 tgt[0].SrcReg[0].Swizzle = SWIZZLE_WWWW;
 139
 140                 tgt[1].Opcode = OPCODE_MAD;
 141                 tgt[1].DstReg = inst.DstReg;
 142                 tgt[1].DstReg.WriteMask = orig_inst->DstReg.WriteMask;
 143                 tgt[1].SrcReg[0] = inst.SrcReg[0];
 144                 tgt[1].SrcReg[0].Swizzle = SWIZZLE_ZZZZ;
 145                 tgt[1].SrcReg[1].File = PROGRAM_TEMPORARY;
 146                 tgt[1].SrcReg[1].Index = rcptemp;
 147                 tgt[1].SrcReg[1].Swizzle = SWIZZLE_WWWW;
 148                 tgt[1].SrcReg[2].File = PROGRAM_TEMPORARY;
 149                 tgt[1].SrcReg[2].Index = inst.DstReg.Index;
 150                 if (depthmode == 0) /* GL_LUMINANCE */
 151                         tgt[1].SrcReg[2].Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_Z);
 152                 else if (depthmode == 2) /* GL_ALPHA */
 153                         tgt[1].SrcReg[2].Swizzle = SWIZZLE_WWWW;
 154
 155                 /* Recall that SrcReg[0] is tex, SrcReg[2] is r and:
 156                  *   r  < tex  <=>      -tex+r < 0
 157                  *   r >= tex  <=> not (-tex+r < 0 */
 158                 if (comparefunc == GL_LESS || comparefunc == GL_GEQUAL)
 159                         tgt[1].SrcReg[2].Negate = tgt[0].SrcReg[2].Negate ^ NEGATE_XYZW;
 160                 else
 161                         tgt[1].SrcReg[0].Negate = tgt[0].SrcReg[0].Negate ^ NEGATE_XYZW;
 162
 163                 tgt[2].Opcode = OPCODE_CMP;
 164                 tgt[2].DstReg = orig_inst->DstReg;
 165                 tgt[2].SrcReg[0].File = PROGRAM_TEMPORARY;
 166                 tgt[2].SrcReg[0].Index = tgt[1].DstReg.Index;
 167
 168                 if (comparefunc == GL_LESS || comparefunc == GL_GREATER) {
 169                         pass = 1;
 170                         fail = 2;
 171                 } else {
 172                         pass = 2;
 173                         fail = 1;
 174                 }
 175
 176                 tgt[2].SrcReg[pass].File = PROGRAM_BUILTIN;
 177                 tgt[2].SrcReg[pass].Swizzle = SWIZZLE_1111;
 178                 tgt[2].SrcReg[fail] = shadow_ambient(t->Program, inst.TexSrcUnit);
 179         } else if (destredirect) {
 180                 tgt = radeonAppendInstructions(t->Program, 1);
 181
 182                 tgt->Opcode = OPCODE_MOV;
 183                 tgt->DstReg = orig_inst->DstReg;
 184                 tgt->SrcReg[0].File = PROGRAM_TEMPORARY;
 185                 tgt->SrcReg[0].Index = inst.DstReg.Index;
 186         }
 187
 188         return GL_TRUE;
 189 }
 190
 191
 192 static void update_params(r300ContextPtr r300, struct r500_fragment_program *fp)
 193 {
 194         struct gl_fragment_program *mp = &fp->mesa_program;
 195
 196         /* Ask Mesa nicely to fill in ParameterValues for us */
 197         if (mp->Base.Parameters)
 198                 _mesa_load_state_parameters(r300->radeon.glCtx, mp->Base.Parameters);
 199 }
 200
 201
 202 /**
 203  * Transform the program to support fragment.position.
 204  *
 205  * Introduce a small fragment at the start of the program that will be
 206  * the only code that directly reads the FRAG_ATTRIB_WPOS input.
 207  * All other code pieces that reference that input will be rewritten
 208  * to read from a newly allocated temporary.
 209  *
 210  * \todo if/when r5xx supports the radeon_program architecture, this is a
 211  * likely candidate for code sharing.
 212  */
 213 static void insert_WPOS_trailer(struct r500_fragment_program_compiler *compiler)
 214 {
 215         GLuint InputsRead = compiler->fp->mesa_program.Base.InputsRead;
 216
 217         if (!(InputsRead & FRAG_BIT_WPOS))
 218                 return;
 219
 220         static gl_state_index tokens[STATE_LENGTH] = {
 221                 STATE_INTERNAL, STATE_R300_WINDOW_DIMENSION, 0, 0, 0
 222         };
 223         struct prog_instruction *fpi;
 224         GLuint window_index;
 225         int i = 0;
 226         GLuint tempregi = _mesa_find_free_register(compiler->program, PROGRAM_TEMPORARY);
 227
 228         _mesa_insert_instructions(compiler->program, 0, 3);
 229         fpi = compiler->program->Instructions;
 230
 231         /* perspective divide */
 232         fpi[i].Opcode = OPCODE_RCP;
 233
 234         fpi[i].DstReg.File = PROGRAM_TEMPORARY;
 235         fpi[i].DstReg.Index = tempregi;
 236         fpi[i].DstReg.WriteMask = WRITEMASK_W;
 237         fpi[i].DstReg.CondMask = COND_TR;
 238
 239         fpi[i].SrcReg[0].File = PROGRAM_INPUT;
 240         fpi[i].SrcReg[0].Index = FRAG_ATTRIB_WPOS;
 241         fpi[i].SrcReg[0].Swizzle = SWIZZLE_WWWW;
 242         i++;
 243
 244         fpi[i].Opcode = OPCODE_MUL;
 245
 246         fpi[i].DstReg.File = PROGRAM_TEMPORARY;
 247         fpi[i].DstReg.Index = tempregi;
 248         fpi[i].DstReg.WriteMask = WRITEMASK_XYZ;
 249         fpi[i].DstReg.CondMask = COND_TR;
 250
 251         fpi[i].SrcReg[0].File = PROGRAM_INPUT;
 252         fpi[i].SrcReg[0].Index = FRAG_ATTRIB_WPOS;
 253         fpi[i].SrcReg[0].Swizzle = SWIZZLE_XYZW;
 254
 255         fpi[i].SrcReg[1].File = PROGRAM_TEMPORARY;
 256         fpi[i].SrcReg[1].Index = tempregi;
 257         fpi[i].SrcReg[1].Swizzle = SWIZZLE_WWWW;
 258         i++;
 259
 260         /* viewport transformation */
 261         window_index = _mesa_add_state_reference(compiler->program->Parameters, tokens);
 262
 263         fpi[i].Opcode = OPCODE_MAD;
 264
 265         fpi[i].DstReg.File = PROGRAM_TEMPORARY;
 266         fpi[i].DstReg.Index = tempregi;
 267         fpi[i].DstReg.WriteMask = WRITEMASK_XYZ;
 268         fpi[i].DstReg.CondMask = COND_TR;
 269
 270         fpi[i].SrcReg[0].File = PROGRAM_TEMPORARY;
 271         fpi[i].SrcReg[0].Index = tempregi;
 272         fpi[i].SrcReg[0].Swizzle =
 273             MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_ZERO);
 274
 275         fpi[i].SrcReg[1].File = PROGRAM_STATE_VAR;
 276         fpi[i].SrcReg[1].Index = window_index;
 277         fpi[i].SrcReg[1].Swizzle =
 278             MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_ZERO);
 279
 280         fpi[i].SrcReg[2].File = PROGRAM_STATE_VAR;
 281         fpi[i].SrcReg[2].Index = window_index;
 282         fpi[i].SrcReg[2].Swizzle =
 283             MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_ZERO);
 284         i++;
 285
 286         for (; i < compiler->program->NumInstructions; ++i) {
 287                 int reg;
 288                 for (reg = 0; reg < 3; reg++) {
 289                         if (fpi[i].SrcReg[reg].File == PROGRAM_INPUT &&
 290                             fpi[i].SrcReg[reg].Index == FRAG_ATTRIB_WPOS) {
 291                                 fpi[i].SrcReg[reg].File = PROGRAM_TEMPORARY;
 292                                 fpi[i].SrcReg[reg].Index = tempregi;
 293                         }
 294                 }
 295         }
 296 }
 297
 298
 299 static void nqssadce_init(struct nqssadce_state* s)
 300 {
 301         s->Outputs[FRAG_RESULT_COLOR].Sourced = WRITEMASK_XYZW;
 302         s->Outputs[FRAG_RESULT_DEPTH].Sourced = WRITEMASK_W;
 303 }
 304
 305 static GLboolean is_native_swizzle(GLuint opcode, struct prog_src_register reg)
 306 {
 307         GLuint relevant;
 308         int i;
 309
 310         if (opcode == OPCODE_TEX ||
 311             opcode == OPCODE_TXB ||
 312             opcode == OPCODE_TXP ||
 313             opcode == OPCODE_KIL) {
 314                 if (reg.Abs)
 315                         return GL_FALSE;
 316
 317                 if (reg.Negate)
 318                         reg.Negate ^= NEGATE_XYZW;
 319
 320                 if (opcode == OPCODE_KIL) {
 321                         if (reg.Swizzle != SWIZZLE_NOOP)
 322                                 return GL_FALSE;
 323                 } else {
 324                         for(i = 0; i < 4; ++i) {
 325                                 GLuint swz = GET_SWZ(reg.Swizzle, i);
 326                                 if (swz == SWIZZLE_NIL) {
 327                                         reg.Negate &= ~(1 << i);
 328                                         continue;
 329                                 }
 330                                 if (swz >= 4)
 331                                         return GL_FALSE;
 332                         }
 333                 }
 334
 335                 if (reg.Negate)
 336                         return GL_FALSE;
 337
 338                 return GL_TRUE;
 339         } else if (opcode == OPCODE_DDX || opcode == OPCODE_DDY) {
 340                 /* DDX/MDH and DDY/MDV explicitly ignore incoming swizzles;
 341                  * if it doesn't fit perfectly into a .xyzw case... */
 342                 if (reg.Swizzle == SWIZZLE_NOOP && !reg.Abs && !reg.Negate)
 343                         return GL_TRUE;
 344
 345                 return GL_FALSE;
 346         } else {
 347                 /* ALU instructions support almost everything */
 348                 if (reg.Abs)
 349                         return GL_TRUE;
 350
 351                 relevant = 0;
 352                 for(i = 0; i < 3; ++i) {
 353                         GLuint swz = GET_SWZ(reg.Swizzle, i);
 354                         if (swz != SWIZZLE_NIL && swz != SWIZZLE_ZERO)
 355                                 relevant |= 1 << i;
 356                 }
 357                 if ((reg.Negate & relevant) && ((reg.Negate & relevant) != relevant))
 358                         return GL_FALSE;
 359
 360                 return GL_TRUE;
 361         }
 362 }
 363
 364 /**
 365  * Implement a MOV with a potentially non-native swizzle.
 366  *
 367  * The only thing we *cannot* do in an ALU instruction is per-component
 368  * negation. Therefore, we split the MOV into two instructions when necessary.
 369  */
 370 static void nqssadce_build_swizzle(struct nqssadce_state *s,
 371         struct prog_dst_register dst, struct prog_src_register src)
 372 {
 373         struct prog_instruction *inst;
 374         GLuint negatebase[2] = { 0, 0 };
 375         int i;
 376
 377         for(i = 0; i < 4; ++i) {
 378                 GLuint swz = GET_SWZ(src.Swizzle, i);
 379                 if (swz == SWIZZLE_NIL)
 380                         continue;
 381                 negatebase[GET_BIT(src.Negate, i)] |= 1 << i;
 382         }
 383
 384         _mesa_insert_instructions(s->Program, s->IP, (negatebase[0] ? 1 : 0) + (negatebase[1] ? 1 : 0));
 385         inst = s->Program->Instructions + s->IP;
 386
 387         for(i = 0; i <= 1; ++i) {
 388                 if (!negatebase[i])
 389                         continue;
 390
 391                 inst->Opcode = OPCODE_MOV;
 392                 inst->DstReg = dst;
 393                 inst->DstReg.WriteMask = negatebase[i];
 394                 inst->SrcReg[0] = src;
 395                 inst++;
 396                 s->IP++;
 397         }
 398 }
 399
 400 static GLuint build_dtm(GLuint depthmode)
 401 {
 402         switch(depthmode) {
 403         default:
 404         case GL_LUMINANCE: return 0;
 405         case GL_INTENSITY: return 1;
 406         case GL_ALPHA: return 2;
 407         }
 408 }
 409
 410 static GLuint build_func(GLuint comparefunc)
 411 {
 412         return comparefunc - GL_NEVER;
 413 }
 414
 415
 416 /**
 417  * Collect all external state that is relevant for compiling the given
 418  * fragment program.
 419  */
 420 static void build_state(
 421         r300ContextPtr r300,
 422         struct r500_fragment_program *fp,
 423         struct r500_fragment_program_external_state *state)
 424 {
 425         int unit;
 426
 427         _mesa_bzero(state, sizeof(*state));
 428
 429         for(unit = 0; unit < 16; ++unit) {
 430                 if (fp->mesa_program.Base.ShadowSamplers & (1 << unit)) {
 431                         struct gl_texture_object* tex = r300->radeon.glCtx->Texture.Unit[unit]._Current;
 432
 433                         state->unit[unit].depth_texture_mode = build_dtm(tex->DepthMode);
 434                         state->unit[unit].texture_compare_func = build_func(tex->CompareFunc);
 435                 }
 436         }
 437 }
 438
 439 static void dump_program(struct r500_fragment_program_code *code);
 440
 441 void r500TranslateFragmentShader(r300ContextPtr r300,
 442                                  struct r500_fragment_program *fp)
 443 {
 444         struct r500_fragment_program_external_state state;
 445
 446         build_state(r300, fp, &state);
 447         if (_mesa_memcmp(&fp->state, &state, sizeof(state))) {
 448                 /* TODO: cache compiled programs */
 449                 fp->translated = GL_FALSE;
 450                 _mesa_memcpy(&fp->state, &state, sizeof(state));
 451         }
 452
 453         if (!fp->translated) {
 454                 struct r500_fragment_program_compiler compiler;
 455
 456                 compiler.r300 = r300;
 457                 compiler.fp = fp;
 458                 compiler.code = &fp->code;
 459                 compiler.program = _mesa_clone_program(r300->radeon.glCtx, &fp->mesa_program.Base);
 460
 461                 if (RADEON_DEBUG & DEBUG_PIXEL) {
 462                         _mesa_printf("Compiler: Initial program:\n");
 463                         _mesa_print_program(compiler.program);
 464                 }
 465
 466                 insert_WPOS_trailer(&compiler);
 467
 468                 struct radeon_program_transformation transformations[] = {
 469                         { &transform_TEX, &compiler },
 470                         { &radeonTransformALU, 0 },
 471                         { &radeonTransformDeriv, 0 },
 472                         { &radeonTransformTrigScale, 0 }
 473                 };
 474                 radeonLocalTransform(r300->radeon.glCtx, compiler.program,
 475                         4, transformations);
 476
 477                 if (RADEON_DEBUG & DEBUG_PIXEL) {
 478                         _mesa_printf("Compiler: after native rewrite:\n");
 479                         _mesa_print_program(compiler.program);
 480                 }
 481
 482                 struct radeon_nqssadce_descr nqssadce = {
 483                         .Init = &nqssadce_init,
 484                         .IsNativeSwizzle = &is_native_swizzle,
 485                         .BuildSwizzle = &nqssadce_build_swizzle,
 486                         .RewriteDepthOut = GL_TRUE
 487                 };
 488                 radeonNqssaDce(r300->radeon.glCtx, compiler.program, &nqssadce);
 489
 490                 if (RADEON_DEBUG & DEBUG_PIXEL) {
 491                         _mesa_printf("Compiler: after NqSSA-DCE:\n");
 492                         _mesa_print_program(compiler.program);
 493                 }
 494
 495                 fp->translated = r500FragmentProgramEmit(&compiler);
 496
 497                 /* Subtle: Rescue any parameters that have been added during transformations */
 498                 _mesa_free_parameter_list(fp->mesa_program.Base.Parameters);
 499                 fp->mesa_program.Base.Parameters = compiler.program->Parameters;
 500                 compiler.program->Parameters = 0;
 501
 502                 _mesa_reference_program(r300->radeon.glCtx, &compiler.program, 0);
 503
 504                 r300UpdateStateParameters(r300->radeon.glCtx, _NEW_PROGRAM |
 505                                           _NEW_PROGRAM_CONSTANTS);
 506
 507                 if (RADEON_DEBUG & DEBUG_PIXEL) {
 508                         if (fp->translated) {
 509                                 _mesa_printf("Machine-readable code:\n");
 510                                 dump_program(&fp->code);
 511                         }
 512                 }
 513
 514         }
 515
 516         update_params(r300, fp);
 517
 518 }
 519
 520 static char *toswiz(int swiz_val) {
 521   switch(swiz_val) {
 522   case 0: return "R";
 523   case 1: return "G";
 524   case 2: return "B";
 525   case 3: return "A";
 526   case 4: return "0";
 527   case 5: return "1/2";
 528   case 6: return "1";
 529   case 7: return "U";
 530   }
 531   return NULL;
 532 }
 533
 534 static char *toop(int op_val)
 535 {
 536   char *str = NULL;
 537   switch (op_val) {
 538   case 0: str = "MAD"; break;
 539   case 1: str = "DP3"; break;
 540   case 2: str = "DP4"; break;
 541   case 3: str = "D2A"; break;
 542   case 4: str = "MIN"; break;
 543   case 5: str = "MAX"; break;
 544   case 6: str = "Reserved"; break;
 545   case 7: str = "CND"; break;
 546   case 8: str = "CMP"; break;
 547   case 9: str = "FRC"; break;
 548   case 10: str = "SOP"; break;
 549   case 11: str = "MDH"; break;
 550   case 12: str = "MDV"; break;
 551   }
 552   return str;
 553 }
 554
 555 static char *to_alpha_op(int op_val)
 556 {
 557   char *str = NULL;
 558   switch (op_val) {
 559   case 0: str = "MAD"; break;
 560   case 1: str = "DP"; break;
 561   case 2: str = "MIN"; break;
 562   case 3: str = "MAX"; break;
 563   case 4: str = "Reserved"; break;
 564   case 5: str = "CND"; break;
 565   case 6: str = "CMP"; break;
 566   case 7: str = "FRC"; break;
 567   case 8: str = "EX2"; break;
 568   case 9: str = "LN2"; break;
 569   case 10: str = "RCP"; break;
 570   case 11: str = "RSQ"; break;
 571   case 12: str = "SIN"; break;
 572   case 13: str = "COS"; break;
 573   case 14: str = "MDH"; break;
 574   case 15: str = "MDV"; break;
 575   }
 576   return str;
 577 }
 578
 579 static char *to_mask(int val)
 580 {
 581   char *str = NULL;
 582   switch(val) {
 583   case 0: str = "NONE"; break;
 584   case 1: str = "R"; break;
 585   case 2: str = "G"; break;
 586   case 3: str = "RG"; break;
 587   case 4: str = "B"; break;
 588   case 5: str = "RB"; break;
 589   case 6: str = "GB"; break;
 590   case 7: str = "RGB"; break;
 591   case 8: str = "A"; break;
 592   case 9: str = "AR"; break;
 593   case 10: str = "AG"; break;
 594   case 11: str = "ARG"; break;
 595   case 12: str = "AB"; break;
 596   case 13: str = "ARB"; break;
 597   case 14: str = "AGB"; break;
 598   case 15: str = "ARGB"; break;
 599   }
 600   return str;
 601 }
 602
 603 static char *to_texop(int val)
 604 {
 605   switch(val) {
 606   case 0: return "NOP";
 607   case 1: return "LD";
 608   case 2: return "TEXKILL";
 609   case 3: return "PROJ";
 610   case 4: return "LODBIAS";
 611   case 5: return "LOD";
 612   case 6: return "DXDY";
 613   }
 614   return NULL;
 615 }
 616
 617 static void dump_program(struct r500_fragment_program_code *code)
 618 {
 619
 620   fprintf(stderr, "R500 Fragment Program:\n--------\n");
 621
 622   int n;
 623   uint32_t inst;
 624   uint32_t inst0;
 625   char *str = NULL;
 626
 627   if (code->const_nr) {
 628     fprintf(stderr, "--------\nConstants:\n");
 629     for (n = 0; n < code->const_nr; n++) {
 630       fprintf(stderr, "Constant %d: %i[%i]\n", n,
 631         code->constant[n].File, code->constant[n].Index);
 632     }
 633     fprintf(stderr, "--------\n");
 634   }
 635
 636   for (n = 0; n < code->inst_end+1; n++) {
 637     inst0 = inst = code->inst[n].inst0;
 638     fprintf(stderr,"%d\t0:CMN_INST   0x%08x:", n, inst);
 639     switch(inst & 0x3) {
 640     case R500_INST_TYPE_ALU: str = "ALU"; break;
 641     case R500_INST_TYPE_OUT: str = "OUT"; break;
 642     case R500_INST_TYPE_FC: str = "FC"; break;
 643     case R500_INST_TYPE_TEX: str = "TEX"; break;
 644     };
 645     fprintf(stderr,"%s %s %s %s %s ", str,
 646             inst & R500_INST_TEX_SEM_WAIT ? "TEX_WAIT" : "",
 647             inst & R500_INST_LAST ? "LAST" : "",
 648             inst & R500_INST_NOP ? "NOP" : "",
 649             inst & R500_INST_ALU_WAIT ? "ALU WAIT" : "");
 650     fprintf(stderr,"wmask: %s omask: %s\n", to_mask((inst >> 11) & 0xf),
 651             to_mask((inst >> 15) & 0xf));
 652
 653     switch(inst0 & 0x3) {
 654     case 0:
 655     case 1:
 656       fprintf(stderr,"\t1:RGB_ADDR   0x%08x:", code->inst[n].inst1);
 657       inst = code->inst[n].inst1;
 658
 659       fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
 660               inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
 661               (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
 662               (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
 663               (inst >> 30));
 664
 665       fprintf(stderr,"\t2:ALPHA_ADDR 0x%08x:", code->inst[n].inst2);
 666       inst = code->inst[n].inst2;
 667       fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
 668               inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
 669               (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
 670               (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
 671               (inst >> 30));
 672       fprintf(stderr,"\t3 RGB_INST:  0x%08x:", code->inst[n].inst3);
 673       inst = code->inst[n].inst3;
 674       fprintf(stderr,"rgb_A_src:%d %s/%s/%s %d rgb_B_src:%d %s/%s/%s %d\n",
 675               (inst) & 0x3, toswiz((inst >> 2) & 0x7), toswiz((inst >> 5) & 0x7), toswiz((inst >> 8) & 0x7),
 676               (inst >> 11) & 0x3,
 677               (inst >> 13) & 0x3, toswiz((inst >> 15) & 0x7), toswiz((inst >> 18) & 0x7), toswiz((inst >> 21) & 0x7),
 678               (inst >> 24) & 0x3);
 679
 680
 681       fprintf(stderr,"\t4 ALPHA_INST:0x%08x:", code->inst[n].inst4);
 682       inst = code->inst[n].inst4;
 683       fprintf(stderr,"%s dest:%d%s alp_A_src:%d %s %d alp_B_src:%d %s %d w:%d\n", to_alpha_op(inst & 0xf),
 684               (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
 685               (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), (inst >> 17) & 0x3,
 686               (inst >> 19) & 0x3, toswiz((inst >> 21) & 0x7), (inst >> 24) & 0x3,
 687               (inst >> 31) & 0x1);
 688
 689       fprintf(stderr,"\t5 RGBA_INST: 0x%08x:", code->inst[n].inst5);
 690       inst = code->inst[n].inst5;
 691       fprintf(stderr,"%s dest:%d%s rgb_C_src:%d %s/%s/%s %d alp_C_src:%d %s %d\n", toop(inst & 0xf),
 692               (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
 693               (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), toswiz((inst >> 17) & 0x7), toswiz((inst >> 20) & 0x7),
 694               (inst >> 23) & 0x3,
 695               (inst >> 25) & 0x3, toswiz((inst >> 27) & 0x7), (inst >> 30) & 0x3);
 696       break;
 697     case 2:
 698       break;
 699     case 3:
 700       inst = code->inst[n].inst1;
 701       fprintf(stderr,"\t1:TEX_INST:  0x%08x: id: %d op:%s, %s, %s %s\n", inst, (inst >> 16) & 0xf,
 702               to_texop((inst >> 22) & 0x7), (inst & (1<<25)) ? "ACQ" : "",
 703               (inst & (1<<26)) ? "IGNUNC" : "", (inst & (1<<27)) ? "UNSCALED" : "SCALED");
 704       inst = code->inst[n].inst2;
 705       fprintf(stderr,"\t2:TEX_ADDR:  0x%08x: src: %d%s %s/%s/%s/%s dst: %d%s %s/%s/%s/%s\n", inst,
 706               inst & 127, inst & (1<<7) ? "(rel)" : "",
 707               toswiz((inst >> 8) & 0x3), toswiz((inst >> 10) & 0x3),
 708               toswiz((inst >> 12) & 0x3), toswiz((inst >> 14) & 0x3),
 709               (inst >> 16) & 127, inst & (1<<23) ? "(rel)" : "",
 710               toswiz((inst >> 24) & 0x3), toswiz((inst >> 26) & 0x3),
 711               toswiz((inst >> 28) & 0x3), toswiz((inst >> 30) & 0x3));
 712
 713       fprintf(stderr,"\t3:TEX_DXDY:  0x%08x\n", code->inst[n].inst3);
 714       break;
 715     }
 716     fprintf(stderr,"\n");
 717   }
 718
 719 }