src/mesa/drivers/dri/r300/compiler/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 <stdio.h>
  31
  32 #include "../r300_reg.h"
  33
  34 static struct rc_src_register shadow_ambient(struct radeon_compiler * c, int tmu)
  35 {
  36         struct rc_src_register reg = { 0, };
  37
  38         reg.File = RC_FILE_CONSTANT;
  39         reg.Index = rc_constants_add_state(&c->Program.Constants, RC_STATE_SHADOW_AMBIENT, tmu);
  40         reg.Swizzle = RC_SWIZZLE_WWWW;
  41         return reg;
  42 }
  43
  44 /**
  45  * Transform TEX, TXP, TXB, and KIL instructions in the following way:
  46  *  - implement texture compare (shadow extensions)
  47  *  - extract non-native source / destination operands
  48  */
  49 int r500_transform_TEX(
  50         struct radeon_compiler * c,
  51         struct rc_instruction * inst,
  52         void* data)
  53 {
  54         struct r300_fragment_program_compiler *compiler =
  55                 (struct r300_fragment_program_compiler*)data;
  56
  57         if (inst->U.I.Opcode != RC_OPCODE_TEX &&
  58             inst->U.I.Opcode != RC_OPCODE_TXB &&
  59             inst->U.I.Opcode != RC_OPCODE_TXP &&
  60             inst->U.I.Opcode != RC_OPCODE_KIL)
  61                 return 0;
  62
  63         /* ARB_shadow & EXT_shadow_funcs */
  64         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
  65             c->Program.ShadowSamplers & (1 << inst->U.I.TexSrcUnit)) {
  66                 rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;
  67
  68                 if (comparefunc == RC_COMPARE_FUNC_NEVER || comparefunc == RC_COMPARE_FUNC_ALWAYS) {
  69                         inst->U.I.Opcode = RC_OPCODE_MOV;
  70
  71                         if (comparefunc == RC_COMPARE_FUNC_ALWAYS) {
  72                                 inst->U.I.SrcReg[0].File = RC_FILE_NONE;
  73                                 inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
  74                         } else {
  75                                 inst->U.I.SrcReg[0] = shadow_ambient(c, inst->U.I.TexSrcUnit);
  76                         }
  77
  78                         return 1;
  79                 } else {
  80                         rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;
  81                         unsigned int depthmode = compiler->state.unit[inst->U.I.TexSrcUnit].depth_texture_mode;
  82                         struct rc_instruction * inst_rcp = rc_insert_new_instruction(c, inst);
  83                         struct rc_instruction * inst_mad = rc_insert_new_instruction(c, inst_rcp);
  84                         struct rc_instruction * inst_cmp = rc_insert_new_instruction(c, inst_mad);
  85                         int pass, fail;
  86
  87                         inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
  88                         inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
  89                         inst_rcp->U.I.DstReg.Index = rc_find_free_temporary(c);
  90                         inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
  91                         inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
  92                         inst_rcp->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_WWWW;
  93
  94                         inst_cmp->U.I.DstReg = inst->U.I.DstReg;
  95                         inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
  96                         inst->U.I.DstReg.Index = rc_find_free_temporary(c);
  97                         inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
  98
  99                         inst_mad->U.I.Opcode = RC_OPCODE_MAD;
 100                         inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
 101                         inst_mad->U.I.DstReg.Index = rc_find_free_temporary(c);
 102                         inst_mad->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 103                         inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_ZZZZ;
 104                         inst_mad->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 105                         inst_mad->U.I.SrcReg[1].Index = inst_rcp->U.I.DstReg.Index;
 106                         inst_mad->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
 107                         inst_mad->U.I.SrcReg[2].File = RC_FILE_TEMPORARY;
 108                         inst_mad->U.I.SrcReg[2].Index = inst->U.I.DstReg.Index;
 109                         if (depthmode == 0) /* GL_LUMINANCE */
 110                                 inst_mad->U.I.SrcReg[2].Swizzle = RC_MAKE_SWIZZLE(RC_SWIZZLE_X, RC_SWIZZLE_Y, RC_SWIZZLE_Z, RC_SWIZZLE_Z);
 111                         else if (depthmode == 2) /* GL_ALPHA */
 112                                 inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_WWWW;
 113
 114                         /* Recall that SrcReg[0] is tex, SrcReg[2] is r and:
 115                          *   r  < tex  <=>      -tex+r < 0
 116                          *   r >= tex  <=> not (-tex+r < 0 */
 117                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GEQUAL)
 118                                 inst_mad->U.I.SrcReg[2].Negate = inst_mad->U.I.SrcReg[2].Negate ^ RC_MASK_XYZW;
 119                         else
 120                                 inst_mad->U.I.SrcReg[0].Negate = inst_mad->U.I.SrcReg[0].Negate ^ RC_MASK_XYZW;
 121
 122                         inst_cmp->U.I.Opcode = RC_OPCODE_CMP;
 123                         /* DstReg has been filled out above */
 124                         inst_cmp->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 125                         inst_cmp->U.I.SrcReg[0].Index = inst_mad->U.I.DstReg.Index;
 126
 127                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GREATER) {
 128                                 pass = 1;
 129                                 fail = 2;
 130                         } else {
 131                                 pass = 2;
 132                                 fail = 1;
 133                         }
 134
 135                         inst_cmp->U.I.SrcReg[pass].File = RC_FILE_NONE;
 136                         inst_cmp->U.I.SrcReg[pass].Swizzle = RC_SWIZZLE_1111;
 137                         inst_cmp->U.I.SrcReg[fail] = shadow_ambient(c, inst->U.I.TexSrcUnit);
 138                 }
 139         }
 140
 141         /* Texture wrap modes don't work on NPOT textures or texrects.
 142          *
 143          * The game plan is simple. We have two flags, fake_npot and
 144          * non_normalized_coords, as well as a tex target. The RECT tex target
 145          * will make the emitted code use non-scaled texcoords.
 146          *
 147          * Non-wrapped/clamped texcoords with NPOT are free in HW. Repeat and
 148          * mirroring are not. If we need to repeat, we do:
 149          *
 150          * MUL temp, texcoord, <scaling factor constant>
 151          * FRC temp, temp ; Discard integer portion of coords
 152          *
 153          * This gives us coords in [0, 1].
 154          *
 155          * Mirroring is trickier. We're going to start out like repeat:
 156          *
 157          * MUL temp0, texcoord, <scaling factor constant> ; De-mirror across axes
 158          * MUL temp0, abs(temp0), 0.5 ; Pattern repeats in [0, 2]
 159          *                            ; so scale to [0, 1]
 160          * FRC temp0, temp0 ; Make the pattern repeat
 161          * SGE temp1, temp0, 0.5 ; Select components that need to be "reflected"
 162          *                       ; across the mirror
 163          * MAD temp0, neg(0.5), temp1, temp0 ; Add -0.5 to the
 164          *                                   ; selected components
 165          * ADD temp0, temp0, temp0 ; Poor man's 2x to undo earlier MUL
 166          *
 167          * This gives us coords in [0, 1].
 168          *
 169          * ~ C.
 170          */
 171         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 172                 (inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
 173                         compiler->state.unit[inst->U.I.TexSrcUnit].fake_npot ||
 174                         compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords)) {
 175                 rc_wrap_mode wrapmode = compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode;
 176                 struct rc_instruction *inst_rect = NULL;
 177                 unsigned temp = rc_find_free_temporary(c);
 178
 179                 if (compiler->state.unit[inst->U.I.TexSrcUnit].fake_npot &&
 180                         wrapmode != RC_WRAP_NONE) {
 181
 182                         if ((inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
 183                                 compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords)) {
 184                                 inst_rect = rc_insert_new_instruction(c, inst->Prev);
 185
 186                                 inst_rect->U.I.Opcode = RC_OPCODE_MUL;
 187                                 inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
 188                                 inst_rect->U.I.DstReg.Index = temp;
 189                                 inst_rect->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 190                                 inst_rect->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
 191                                 inst_rect->U.I.SrcReg[1].Index =
 192                                         rc_constants_add_state(&c->Program.Constants,
 193                                                 RC_STATE_R300_TEXRECT_FACTOR, inst->U.I.TexSrcUnit);
 194
 195                                 reset_srcreg(&inst->U.I.SrcReg[0]);
 196                                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 197                                 inst->U.I.SrcReg[0].Index = temp;
 198
 199                                 inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
 200                         }
 201
 202                         if (wrapmode == RC_WRAP_REPEAT) {
 203                                 /* Both instructions will be paired up. */
 204                                 struct rc_instruction *inst_frc = rc_insert_new_instruction(c, inst->Prev);
 205                                 struct rc_instruction *inst_mov = rc_insert_new_instruction(c, inst_frc);
 206
 207                                 inst_frc->U.I.Opcode = RC_OPCODE_FRC;
 208                                 inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
 209                                 inst_frc->U.I.DstReg.Index = temp;
 210                                 inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 211                                 inst_frc->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 212
 213                                 /* Preserve W for TXP. */
 214                                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 215                                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 216                                 inst_mov->U.I.DstReg.Index = temp;
 217                                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
 218                                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 219
 220                                 reset_srcreg(&inst->U.I.SrcReg[0]);
 221                                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 222                                 inst->U.I.SrcReg[0].Index = temp;
 223                         } else if (wrapmode == RC_WRAP_MIRROR) {
 224                                 unsigned temp1;
 225                                 /*
 226                                  * MUL temp0, abs(temp0), 0.5
 227                                  * FRC temp0, temp0
 228                                  * SGE temp1, temp0, 0.5
 229                                  * MAD temp0, neg(0.5), temp1, temp0
 230                                  * ADD temp0, temp0, temp0
 231                                  */
 232                                 inst_rect = rc_insert_new_instruction(c, inst->Prev);
 233
 234                                 inst_rect->U.I.Opcode = RC_OPCODE_MUL;
 235                                 inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
 236                                 inst_rect->U.I.DstReg.Index = temp;
 237                                 inst_rect->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 238                                 inst_rect->U.I.SrcReg[1].Swizzle = RC_MAKE_SWIZZLE_SMEAR(RC_SWIZZLE_HALF);
 239
 240                                 inst_rect = rc_insert_new_instruction(c, inst->Prev);
 241
 242                                 inst_rect->U.I.Opcode = RC_OPCODE_FRC;
 243                                 inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
 244                                 inst_rect->U.I.DstReg.Index = temp;
 245                                 inst_rect->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 246                                 inst_rect->U.I.SrcReg[0].Index = temp;
 247
 248                                 temp1 = rc_find_free_temporary(c);
 249                                 inst_rect = rc_insert_new_instruction(c, inst->Prev);
 250
 251                                 inst_rect->U.I.Opcode = RC_OPCODE_SGE;
 252                                 inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
 253                                 inst_rect->U.I.DstReg.Index = temp1;
 254                                 inst_rect->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 255                                 inst_rect->U.I.SrcReg[0].Index = temp;
 256                                 inst_rect->U.I.SrcReg[1].Swizzle = RC_MAKE_SWIZZLE_SMEAR(RC_SWIZZLE_HALF);
 257
 258                                 inst_rect = rc_insert_new_instruction(c, inst->Prev);
 259
 260                                 inst_rect->U.I.Opcode = RC_OPCODE_MAD;
 261                                 inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
 262                                 inst_rect->U.I.DstReg.Index = temp;
 263                                 inst_rect->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 264                                 inst_rect->U.I.SrcReg[0].Index = temp1;
 265                                 inst_rect->U.I.SrcReg[1].Swizzle = RC_MAKE_SWIZZLE_SMEAR(RC_SWIZZLE_HALF);
 266                                 inst_rect->U.I.SrcReg[1].Negate = 1;
 267                                 inst_rect->U.I.SrcReg[2].File = RC_FILE_TEMPORARY;
 268                                 inst_rect->U.I.SrcReg[2].Index = temp;
 269
 270                                 inst_rect = rc_insert_new_instruction(c, inst->Prev);
 271
 272                                 inst_rect->U.I.Opcode = RC_OPCODE_ADD;
 273                                 inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
 274                                 inst_rect->U.I.DstReg.Index = temp;
 275                                 inst_rect->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 276                                 inst_rect->U.I.SrcReg[0].Index = temp;
 277                                 inst_rect->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 278                                 inst_rect->U.I.SrcReg[1].Index = temp;
 279
 280                                 reset_srcreg(&inst->U.I.SrcReg[0]);
 281                                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 282                                 inst->U.I.SrcReg[0].Index = temp;
 283                         }
 284                 }
 285         }
 286
 287         /* Cannot write texture to output registers */
 288         if (inst->U.I.Opcode != RC_OPCODE_KIL && inst->U.I.DstReg.File != RC_FILE_TEMPORARY) {
 289                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst);
 290
 291                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 292                 inst_mov->U.I.DstReg = inst->U.I.DstReg;
 293                 inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 294                 inst_mov->U.I.SrcReg[0].Index = rc_find_free_temporary(c);
 295
 296                 inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 297                 inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
 298                 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 299         }
 300
 301         /* Cannot read texture coordinate from constants file */
 302         if (inst->U.I.SrcReg[0].File != RC_FILE_TEMPORARY && inst->U.I.SrcReg[0].File != RC_FILE_INPUT) {
 303                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
 304
 305                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 306                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 307                 inst_mov->U.I.DstReg.Index = rc_find_free_temporary(c);
 308                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 309
 310                 reset_srcreg(&inst->U.I.SrcReg[0]);
 311                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 312                 inst->U.I.SrcReg[0].Index = inst_mov->U.I.DstReg.Index;
 313         }
 314
 315         return 1;
 316 }
 317
 318 /**
 319  * Rewrite IF instructions to use the ALU result special register.
 320  */
 321 int r500_transform_IF(
 322         struct radeon_compiler * c,
 323         struct rc_instruction * inst,
 324         void* data)
 325 {
 326         if (inst->U.I.Opcode != RC_OPCODE_IF)
 327                 return 0;
 328
 329         struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
 330         inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 331         inst_mov->U.I.DstReg.WriteMask = 0;
 332         inst_mov->U.I.WriteALUResult = RC_ALURESULT_W;
 333         inst_mov->U.I.ALUResultCompare = RC_COMPARE_FUNC_NOTEQUAL;
 334         inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 335         inst_mov->U.I.SrcReg[0].Swizzle = combine_swizzles4(inst_mov->U.I.SrcReg[0].Swizzle,
 336                         RC_SWIZZLE_UNUSED, RC_SWIZZLE_UNUSED, RC_SWIZZLE_UNUSED, RC_SWIZZLE_X);
 337
 338         inst->U.I.SrcReg[0].File = RC_FILE_SPECIAL;
 339         inst->U.I.SrcReg[0].Index = RC_SPECIAL_ALU_RESULT;
 340         inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;
 341         inst->U.I.SrcReg[0].Negate = 0;
 342
 343         return 1;
 344 }
 345
 346 static int r500_swizzle_is_native(rc_opcode opcode, struct rc_src_register reg)
 347 {
 348         unsigned int relevant;
 349         int i;
 350
 351         if (opcode == RC_OPCODE_TEX ||
 352             opcode == RC_OPCODE_TXB ||
 353             opcode == RC_OPCODE_TXP ||
 354             opcode == RC_OPCODE_KIL) {
 355                 if (reg.Abs)
 356                         return 0;
 357
 358                 if (opcode == RC_OPCODE_KIL && (reg.Swizzle != RC_SWIZZLE_XYZW || reg.Negate != RC_MASK_NONE))
 359                         return 0;
 360
 361                 if (reg.Negate)
 362                         reg.Negate ^= RC_MASK_XYZW;
 363
 364                 for(i = 0; i < 4; ++i) {
 365                         unsigned int swz = GET_SWZ(reg.Swizzle, i);
 366                         if (swz == RC_SWIZZLE_UNUSED) {
 367                                 reg.Negate &= ~(1 << i);
 368                                 continue;
 369                         }
 370                         if (swz >= 4)
 371                                 return 0;
 372                 }
 373
 374                 if (reg.Negate)
 375                         return 0;
 376
 377                 return 1;
 378         } else if (opcode == RC_OPCODE_DDX || opcode == RC_OPCODE_DDY) {
 379                 /* DDX/MDH and DDY/MDV explicitly ignore incoming swizzles;
 380                  * if it doesn't fit perfectly into a .xyzw case... */
 381                 if (reg.Swizzle == RC_SWIZZLE_XYZW && !reg.Abs && !reg.Negate)
 382                         return 1;
 383
 384                 return 0;
 385         } else {
 386                 /* ALU instructions support almost everything */
 387                 if (reg.Abs)
 388                         return 1;
 389
 390                 relevant = 0;
 391                 for(i = 0; i < 3; ++i) {
 392                         unsigned int swz = GET_SWZ(reg.Swizzle, i);
 393                         if (swz != RC_SWIZZLE_UNUSED && swz != RC_SWIZZLE_ZERO)
 394                                 relevant |= 1 << i;
 395                 }
 396                 if ((reg.Negate & relevant) && ((reg.Negate & relevant) != relevant))
 397                         return 0;
 398
 399                 return 1;
 400         }
 401 }
 402
 403 /**
 404  * Split source register access.
 405  *
 406  * The only thing we *cannot* do in an ALU instruction is per-component
 407  * negation.
 408  */
 409 static void r500_swizzle_split(struct rc_src_register src, unsigned int usemask,
 410                 struct rc_swizzle_split * split)
 411 {
 412         unsigned int negatebase[2] = { 0, 0 };
 413         int i;
 414
 415         for(i = 0; i < 4; ++i) {
 416                 unsigned int swz = GET_SWZ(src.Swizzle, i);
 417                 if (swz == RC_SWIZZLE_UNUSED || !GET_BIT(usemask, i))
 418                         continue;
 419                 negatebase[GET_BIT(src.Negate, i)] |= 1 << i;
 420         }
 421
 422         split->NumPhases = 0;
 423
 424         for(i = 0; i <= 1; ++i) {
 425                 if (!negatebase[i])
 426                         continue;
 427
 428                 split->Phase[split->NumPhases++] = negatebase[i];
 429         }
 430 }
 431
 432 struct rc_swizzle_caps r500_swizzle_caps = {
 433         .IsNative = r500_swizzle_is_native,
 434         .Split = r500_swizzle_split
 435 };
 436
 437 static char *toswiz(int swiz_val) {
 438   switch(swiz_val) {
 439   case 0: return "R";
 440   case 1: return "G";
 441   case 2: return "B";
 442   case 3: return "A";
 443   case 4: return "0";
 444   case 5: return "H";
 445   case 6: return "1";
 446   case 7: return "U";
 447   }
 448   return NULL;
 449 }
 450
 451 static char *toop(int op_val)
 452 {
 453   char *str = NULL;
 454   switch (op_val) {
 455   case 0: str = "MAD"; break;
 456   case 1: str = "DP3"; break;
 457   case 2: str = "DP4"; break;
 458   case 3: str = "D2A"; break;
 459   case 4: str = "MIN"; break;
 460   case 5: str = "MAX"; break;
 461   case 6: str = "Reserved"; break;
 462   case 7: str = "CND"; break;
 463   case 8: str = "CMP"; break;
 464   case 9: str = "FRC"; break;
 465   case 10: str = "SOP"; break;
 466   case 11: str = "MDH"; break;
 467   case 12: str = "MDV"; break;
 468   }
 469   return str;
 470 }
 471
 472 static char *to_alpha_op(int op_val)
 473 {
 474   char *str = NULL;
 475   switch (op_val) {
 476   case 0: str = "MAD"; break;
 477   case 1: str = "DP"; break;
 478   case 2: str = "MIN"; break;
 479   case 3: str = "MAX"; break;
 480   case 4: str = "Reserved"; break;
 481   case 5: str = "CND"; break;
 482   case 6: str = "CMP"; break;
 483   case 7: str = "FRC"; break;
 484   case 8: str = "EX2"; break;
 485   case 9: str = "LN2"; break;
 486   case 10: str = "RCP"; break;
 487   case 11: str = "RSQ"; break;
 488   case 12: str = "SIN"; break;
 489   case 13: str = "COS"; break;
 490   case 14: str = "MDH"; break;
 491   case 15: str = "MDV"; break;
 492   }
 493   return str;
 494 }
 495
 496 static char *to_mask(int val)
 497 {
 498   char *str = NULL;
 499   switch(val) {
 500   case 0: str = "NONE"; break;
 501   case 1: str = "R"; break;
 502   case 2: str = "G"; break;
 503   case 3: str = "RG"; break;
 504   case 4: str = "B"; break;
 505   case 5: str = "RB"; break;
 506   case 6: str = "GB"; break;
 507   case 7: str = "RGB"; break;
 508   case 8: str = "A"; break;
 509   case 9: str = "AR"; break;
 510   case 10: str = "AG"; break;
 511   case 11: str = "ARG"; break;
 512   case 12: str = "AB"; break;
 513   case 13: str = "ARB"; break;
 514   case 14: str = "AGB"; break;
 515   case 15: str = "ARGB"; break;
 516   }
 517   return str;
 518 }
 519
 520 static char *to_texop(int val)
 521 {
 522   switch(val) {
 523   case 0: return "NOP";
 524   case 1: return "LD";
 525   case 2: return "TEXKILL";
 526   case 3: return "PROJ";
 527   case 4: return "LODBIAS";
 528   case 5: return "LOD";
 529   case 6: return "DXDY";
 530   }
 531   return NULL;
 532 }
 533
 534 void r500FragmentProgramDump(struct rX00_fragment_program_code *c)
 535 {
 536   struct r500_fragment_program_code *code = &c->code.r500;
 537   fprintf(stderr, "R500 Fragment Program:\n--------\n");
 538
 539   int n;
 540   uint32_t inst;
 541   uint32_t inst0;
 542   char *str = NULL;
 543
 544   for (n = 0; n < code->inst_end+1; n++) {
 545     inst0 = inst = code->inst[n].inst0;
 546     fprintf(stderr,"%d\t0:CMN_INST   0x%08x:", n, inst);
 547     switch(inst & 0x3) {
 548     case R500_INST_TYPE_ALU: str = "ALU"; break;
 549     case R500_INST_TYPE_OUT: str = "OUT"; break;
 550     case R500_INST_TYPE_FC: str = "FC"; break;
 551     case R500_INST_TYPE_TEX: str = "TEX"; break;
 552     };
 553     fprintf(stderr,"%s %s %s %s %s ", str,
 554             inst & R500_INST_TEX_SEM_WAIT ? "TEX_WAIT" : "",
 555             inst & R500_INST_LAST ? "LAST" : "",
 556             inst & R500_INST_NOP ? "NOP" : "",
 557             inst & R500_INST_ALU_WAIT ? "ALU WAIT" : "");
 558     fprintf(stderr,"wmask: %s omask: %s\n", to_mask((inst >> 11) & 0xf),
 559             to_mask((inst >> 15) & 0xf));
 560
 561     switch(inst0 & 0x3) {
 562     case 0:
 563     case 1:
 564       fprintf(stderr,"\t1:RGB_ADDR   0x%08x:", code->inst[n].inst1);
 565       inst = code->inst[n].inst1;
 566
 567       fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
 568               inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
 569               (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
 570               (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
 571               (inst >> 30));
 572
 573       fprintf(stderr,"\t2:ALPHA_ADDR 0x%08x:", code->inst[n].inst2);
 574       inst = code->inst[n].inst2;
 575       fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
 576               inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
 577               (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
 578               (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
 579               (inst >> 30));
 580       fprintf(stderr,"\t3 RGB_INST:  0x%08x:", code->inst[n].inst3);
 581       inst = code->inst[n].inst3;
 582       fprintf(stderr,"rgb_A_src:%d %s/%s/%s %d rgb_B_src:%d %s/%s/%s %d targ: %d\n",
 583               (inst) & 0x3, toswiz((inst >> 2) & 0x7), toswiz((inst >> 5) & 0x7), toswiz((inst >> 8) & 0x7),
 584               (inst >> 11) & 0x3,
 585               (inst >> 13) & 0x3, toswiz((inst >> 15) & 0x7), toswiz((inst >> 18) & 0x7), toswiz((inst >> 21) & 0x7),
 586               (inst >> 24) & 0x3, (inst >> 29) & 0x3);
 587
 588
 589       fprintf(stderr,"\t4 ALPHA_INST:0x%08x:", code->inst[n].inst4);
 590       inst = code->inst[n].inst4;
 591       fprintf(stderr,"%s dest:%d%s alp_A_src:%d %s %d alp_B_src:%d %s %d targ %d w:%d\n", to_alpha_op(inst & 0xf),
 592               (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
 593               (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), (inst >> 17) & 0x3,
 594               (inst >> 19) & 0x3, toswiz((inst >> 21) & 0x7), (inst >> 24) & 0x3,
 595               (inst >> 29) & 0x3,
 596               (inst >> 31) & 0x1);
 597
 598       fprintf(stderr,"\t5 RGBA_INST: 0x%08x:", code->inst[n].inst5);
 599       inst = code->inst[n].inst5;
 600       fprintf(stderr,"%s dest:%d%s rgb_C_src:%d %s/%s/%s %d alp_C_src:%d %s %d\n", toop(inst & 0xf),
 601               (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
 602               (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), toswiz((inst >> 17) & 0x7), toswiz((inst >> 20) & 0x7),
 603               (inst >> 23) & 0x3,
 604               (inst >> 25) & 0x3, toswiz((inst >> 27) & 0x7), (inst >> 30) & 0x3);
 605       break;
 606     case 2:
 607       break;
 608     case 3:
 609       inst = code->inst[n].inst1;
 610       fprintf(stderr,"\t1:TEX_INST:  0x%08x: id: %d op:%s, %s, %s %s\n", inst, (inst >> 16) & 0xf,
 611               to_texop((inst >> 22) & 0x7), (inst & (1<<25)) ? "ACQ" : "",
 612               (inst & (1<<26)) ? "IGNUNC" : "", (inst & (1<<27)) ? "UNSCALED" : "SCALED");
 613       inst = code->inst[n].inst2;
 614       fprintf(stderr,"\t2:TEX_ADDR:  0x%08x: src: %d%s %s/%s/%s/%s dst: %d%s %s/%s/%s/%s\n", inst,
 615               inst & 127, inst & (1<<7) ? "(rel)" : "",
 616               toswiz((inst >> 8) & 0x3), toswiz((inst >> 10) & 0x3),
 617               toswiz((inst >> 12) & 0x3), toswiz((inst >> 14) & 0x3),
 618               (inst >> 16) & 127, inst & (1<<23) ? "(rel)" : "",
 619               toswiz((inst >> 24) & 0x3), toswiz((inst >> 26) & 0x3),
 620               toswiz((inst >> 28) & 0x3), toswiz((inst >> 30) & 0x3));
 621
 622       fprintf(stderr,"\t3:TEX_DXDY:  0x%08x\n", code->inst[n].inst3);
 623       break;
 624     }
 625     fprintf(stderr,"\n");
 626   }
 627
 628 }