src/mesa/drivers/dri/r300/compiler/radeon_program_tex.c

   1 /*
   2  * Copyright (C) 2010 Corbin Simpson
   3  * Copyright (C) 2010 Marek Olšák <maraeo@gmail.com>
   4  *
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining
   8  * a copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sublicense, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial
  17  * portions of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  20  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  22  * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  23  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  24  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  25  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  */
  28
  29 #include "radeon_program_tex.h"
  30
  31 #include "radeon_compiler_util.h"
  32
  33 /* Series of transformations to be done on textures. */
  34
  35 static struct rc_src_register shadow_fail_value(struct r300_fragment_program_compiler *compiler,
  36                                                 int tmu)
  37 {
  38         struct rc_src_register reg = { 0, };
  39
  40         if (compiler->enable_shadow_ambient) {
  41                 reg.File = RC_FILE_CONSTANT;
  42                 reg.Index = rc_constants_add_state(&compiler->Base.Program.Constants,
  43                                                    RC_STATE_SHADOW_AMBIENT, tmu);
  44                 reg.Swizzle = RC_SWIZZLE_WWWW;
  45         } else {
  46                 reg.File = RC_FILE_NONE;
  47                 reg.Swizzle = RC_SWIZZLE_0000;
  48         }
  49
  50         reg.Swizzle = combine_swizzles(reg.Swizzle,
  51                                 compiler->state.unit[tmu].texture_swizzle);
  52         return reg;
  53 }
  54
  55 static struct rc_src_register shadow_pass_value(struct r300_fragment_program_compiler *compiler,
  56                                                 int tmu)
  57 {
  58         struct rc_src_register reg = { 0, };
  59
  60         reg.File = RC_FILE_NONE;
  61         reg.Swizzle = combine_swizzles(RC_SWIZZLE_1111,
  62                                 compiler->state.unit[tmu].texture_swizzle);
  63         return reg;
  64 }
  65
  66 static void scale_texcoords(struct r300_fragment_program_compiler *compiler,
  67                             struct rc_instruction *inst,
  68                             unsigned state_constant)
  69 {
  70         struct rc_instruction *inst_mov;
  71
  72         unsigned temp = rc_find_free_temporary(&compiler->Base);
  73
  74         inst_mov = rc_insert_new_instruction(&compiler->Base, inst->Prev);
  75
  76         inst_mov->U.I.Opcode = RC_OPCODE_MUL;
  77         inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
  78         inst_mov->U.I.DstReg.Index = temp;
  79         inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
  80         inst_mov->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
  81         inst_mov->U.I.SrcReg[1].Index =
  82                         rc_constants_add_state(&compiler->Base.Program.Constants,
  83                                                state_constant, inst->U.I.TexSrcUnit);
  84
  85         reset_srcreg(&inst->U.I.SrcReg[0]);
  86         inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
  87         inst->U.I.SrcReg[0].Index = temp;
  88 }
  89
  90 static void projective_divide(struct r300_fragment_program_compiler *compiler,
  91                               struct rc_instruction *inst)
  92 {
  93         struct rc_instruction *inst_mul, *inst_rcp;
  94
  95         unsigned temp = rc_find_free_temporary(&compiler->Base);
  96
  97         inst_rcp = rc_insert_new_instruction(&compiler->Base, inst->Prev);
  98         inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
  99         inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
 100         inst_rcp->U.I.DstReg.Index = temp;
 101         inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
 102         inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 103         /* Because the input can be arbitrarily swizzled,
 104          * read the component mapped to W. */
 105         inst_rcp->U.I.SrcReg[0].Swizzle =
 106                 RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 3));
 107
 108         inst_mul = rc_insert_new_instruction(&compiler->Base, inst->Prev);
 109         inst_mul->U.I.Opcode = RC_OPCODE_MUL;
 110         inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 111         inst_mul->U.I.DstReg.Index = temp;
 112         inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 113         inst_mul->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 114         inst_mul->U.I.SrcReg[1].Index = temp;
 115         inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
 116
 117         reset_srcreg(&inst->U.I.SrcReg[0]);
 118         inst->U.I.Opcode = RC_OPCODE_TEX;
 119         inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 120         inst->U.I.SrcReg[0].Index = temp;
 121 }
 122
 123 /**
 124  * Transform TEX, TXP, TXB, and KIL instructions in the following ways:
 125  *  - implement texture compare (shadow extensions)
 126  *  - extract non-native source / destination operands
 127  *  - premultiply texture coordinates for RECT
 128  *  - extract operand swizzles
 129  *  - introduce a temporary register when write masks are needed
 130  */
 131 int radeonTransformTEX(
 132         struct radeon_compiler * c,
 133         struct rc_instruction * inst,
 134         void* data)
 135 {
 136         struct r300_fragment_program_compiler *compiler =
 137                 (struct r300_fragment_program_compiler*)data;
 138         rc_wrap_mode wrapmode = compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode;
 139         int is_rect = inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
 140                       compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords;
 141
 142         if (inst->U.I.Opcode != RC_OPCODE_TEX &&
 143                 inst->U.I.Opcode != RC_OPCODE_TXB &&
 144                 inst->U.I.Opcode != RC_OPCODE_TXP &&
 145                 inst->U.I.Opcode != RC_OPCODE_TXD &&
 146                 inst->U.I.Opcode != RC_OPCODE_TXL &&
 147                 inst->U.I.Opcode != RC_OPCODE_KIL)
 148                 return 0;
 149
 150         /* ARB_shadow & EXT_shadow_funcs */
 151         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 152                 ((c->Program.ShadowSamplers & (1 << inst->U.I.TexSrcUnit)) ||
 153                  (compiler->state.unit[inst->U.I.TexSrcUnit].compare_mode_enabled))) {
 154                 rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;
 155
 156                 if (comparefunc == RC_COMPARE_FUNC_NEVER || comparefunc == RC_COMPARE_FUNC_ALWAYS) {
 157                         inst->U.I.Opcode = RC_OPCODE_MOV;
 158
 159                         if (comparefunc == RC_COMPARE_FUNC_ALWAYS) {
 160                                 inst->U.I.SrcReg[0] = shadow_pass_value(compiler, inst->U.I.TexSrcUnit);
 161                         } else {
 162                                 inst->U.I.SrcReg[0] = shadow_fail_value(compiler, inst->U.I.TexSrcUnit);
 163                         }
 164
 165                         return 1;
 166                 } else {
 167                         struct rc_instruction * inst_rcp = NULL;
 168                         struct rc_instruction *inst_mul, *inst_add, *inst_cmp;
 169                         unsigned tmp_texsample;
 170                         unsigned tmp_sum;
 171                         int pass, fail;
 172
 173                         /* Save the output register. */
 174                         struct rc_dst_register output_reg = inst->U.I.DstReg;
 175                         unsigned saturate_mode = inst->U.I.SaturateMode;
 176
 177                         /* Redirect TEX to a new temp. */
 178                         tmp_texsample = rc_find_free_temporary(c);
 179                         inst->U.I.SaturateMode = 0;
 180                         inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 181                         inst->U.I.DstReg.Index = tmp_texsample;
 182                         inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 183
 184                         tmp_sum = rc_find_free_temporary(c);
 185
 186                         if (inst->U.I.Opcode == RC_OPCODE_TXP) {
 187                                 /* Compute 1/W. */
 188                                 inst_rcp = rc_insert_new_instruction(c, inst);
 189                                 inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
 190                                 inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
 191                                 inst_rcp->U.I.DstReg.Index = tmp_sum;
 192                                 inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
 193                                 inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 194                                 inst_rcp->U.I.SrcReg[0].Swizzle =
 195                                         RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 3));
 196                         }
 197
 198                         /* Divide Z by W (if it's TXP) and saturate. */
 199                         inst_mul = rc_insert_new_instruction(c, inst_rcp ? inst_rcp : inst);
 200                         inst_mul->U.I.Opcode = inst->U.I.Opcode == RC_OPCODE_TXP ? RC_OPCODE_MUL : RC_OPCODE_MOV;
 201                         inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 202                         inst_mul->U.I.DstReg.Index = tmp_sum;
 203                         inst_mul->U.I.DstReg.WriteMask = RC_MASK_W;
 204                         inst_mul->U.I.SaturateMode = RC_SATURATE_ZERO_ONE;
 205                         inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 206                         inst_mul->U.I.SrcReg[0].Swizzle =
 207                                 RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 2));
 208                         if (inst->U.I.Opcode == RC_OPCODE_TXP) {
 209                                 inst_mul->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 210                                 inst_mul->U.I.SrcReg[1].Index = tmp_sum;
 211                                 inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
 212                         }
 213
 214                         /* Add the depth texture value. */
 215                         inst_add = rc_insert_new_instruction(c, inst_mul);
 216                         inst_add->U.I.Opcode = RC_OPCODE_ADD;
 217                         inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
 218                         inst_add->U.I.DstReg.Index = tmp_sum;
 219                         inst_add->U.I.DstReg.WriteMask = RC_MASK_W;
 220                         inst_add->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 221                         inst_add->U.I.SrcReg[0].Index = tmp_sum;
 222                         inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_WWWW;
 223                         inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 224                         inst_add->U.I.SrcReg[1].Index = tmp_texsample;
 225                         inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XXXX;
 226
 227                         /* Note that SrcReg[0] is r, SrcReg[1] is tex and:
 228                          *   LESS:    r  < tex  <=>      -tex+r < 0
 229                          *   GEQUAL:  r >= tex  <=> not (-tex+r < 0)
 230                          *   GREATER: r  > tex  <=>       tex-r < 0
 231                          *   LEQUAL:  r <= tex  <=> not ( tex-r < 0)
 232                          *   EQUAL:   GEQUAL
 233                          *   NOTEQUAL:LESS
 234                          */
 235
 236                         /* This negates either r or tex: */
 237                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GEQUAL ||
 238                             comparefunc == RC_COMPARE_FUNC_EQUAL || comparefunc == RC_COMPARE_FUNC_NOTEQUAL)
 239                                 inst_add->U.I.SrcReg[1].Negate = inst_add->U.I.SrcReg[1].Negate ^ RC_MASK_XYZW;
 240                         else
 241                                 inst_add->U.I.SrcReg[0].Negate = inst_add->U.I.SrcReg[0].Negate ^ RC_MASK_XYZW;
 242
 243                         /* This negates the whole expresion: */
 244                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GREATER ||
 245                             comparefunc == RC_COMPARE_FUNC_NOTEQUAL) {
 246                                 pass = 1;
 247                                 fail = 2;
 248                         } else {
 249                                 pass = 2;
 250                                 fail = 1;
 251                         }
 252
 253                         inst_cmp = rc_insert_new_instruction(c, inst_add);
 254                         inst_cmp->U.I.Opcode = RC_OPCODE_CMP;
 255                         inst_cmp->U.I.SaturateMode = saturate_mode;
 256                         inst_cmp->U.I.DstReg = output_reg;
 257                         inst_cmp->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 258                         inst_cmp->U.I.SrcReg[0].Index = tmp_sum;
 259                         inst_cmp->U.I.SrcReg[0].Swizzle =
 260                                         combine_swizzles(RC_SWIZZLE_WWWW,
 261                                                          compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle);
 262                         inst_cmp->U.I.SrcReg[pass] = shadow_pass_value(compiler, inst->U.I.TexSrcUnit);
 263                         inst_cmp->U.I.SrcReg[fail] = shadow_fail_value(compiler, inst->U.I.TexSrcUnit);
 264
 265                         assert(tmp_texsample != tmp_sum);
 266                 }
 267         }
 268
 269         /* R300 cannot sample from rectangles and the wrap mode fallback needs
 270          * normalized coordinates anyway. */
 271         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 272             is_rect && (!c->is_r500 || wrapmode != RC_WRAP_NONE)) {
 273                 scale_texcoords(compiler, inst, RC_STATE_R300_TEXRECT_FACTOR);
 274                 inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
 275         }
 276
 277         /* Divide by W if needed. */
 278         if (inst->U.I.Opcode == RC_OPCODE_TXP &&
 279             (wrapmode == RC_WRAP_REPEAT || wrapmode == RC_WRAP_MIRRORED_REPEAT ||
 280              compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch)) {
 281                 projective_divide(compiler, inst);
 282         }
 283
 284         /* Texture wrap modes don't work on NPOT textures.
 285          *
 286          * Non-wrapped/clamped texcoords with NPOT are free in HW. Repeat and
 287          * mirroring are not. If we need to repeat, we do:
 288          *
 289          * MUL temp, texcoord, <scaling factor constant>
 290          * FRC temp, temp ; Discard integer portion of coords
 291          *
 292          * This gives us coords in [0, 1].
 293          *
 294          * Mirroring is trickier. We're going to start out like repeat:
 295          *
 296          * MUL temp, texcoord, <scaling factor constant> ; De-mirror across axes
 297          * MUL temp, temp, 0.5 ; Pattern repeats in [0, 2]
 298          *                            ; so scale to [0, 1]
 299          * FRC temp, temp ; Make the pattern repeat
 300          * MAD temp, temp, 2, -1 ; Move the pattern to [-1, 1]
 301          * ADD temp, 1, -abs(temp) ; Now comes a neat trick: use abs to mirror the pattern.
 302          *                              ; The pattern is backwards, so reverse it (1-x).
 303          *
 304          * This gives us coords in [0, 1].
 305          *
 306          * ~ C & M. ;)
 307          */
 308         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 309             wrapmode != RC_WRAP_NONE) {
 310                 struct rc_instruction *inst_mov;
 311                 unsigned temp = rc_find_free_temporary(c);
 312
 313                 if (wrapmode == RC_WRAP_REPEAT) {
 314                         /* Both instructions will be paired up. */
 315                         struct rc_instruction *inst_frc = rc_insert_new_instruction(c, inst->Prev);
 316
 317                         inst_frc->U.I.Opcode = RC_OPCODE_FRC;
 318                         inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
 319                         inst_frc->U.I.DstReg.Index = temp;
 320                         inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 321                         inst_frc->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 322                 } else if (wrapmode == RC_WRAP_MIRRORED_REPEAT) {
 323                         /*
 324                          * Function:
 325                          *   f(v) = 1 - abs(frac(v * 0.5) * 2 - 1)
 326                          *
 327                          * Code:
 328                          *   MUL temp, src0, 0.5
 329                          *   FRC temp, temp
 330                          *   MAD temp, temp, 2, -1
 331                          *   ADD temp, 1, -abs(temp)
 332                          */
 333
 334                         struct rc_instruction *inst_mul, *inst_frc, *inst_mad, *inst_add;
 335                         unsigned two, two_swizzle;
 336
 337                         inst_mul = rc_insert_new_instruction(c, inst->Prev);
 338
 339                         inst_mul->U.I.Opcode = RC_OPCODE_MUL;
 340                         inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 341                         inst_mul->U.I.DstReg.Index = temp;
 342                         inst_mul->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 343                         inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 344                         inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_HHHH;
 345
 346                         inst_frc = rc_insert_new_instruction(c, inst->Prev);
 347
 348                         inst_frc->U.I.Opcode = RC_OPCODE_FRC;
 349                         inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
 350                         inst_frc->U.I.DstReg.Index = temp;
 351                         inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 352                         inst_frc->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 353                         inst_frc->U.I.SrcReg[0].Index = temp;
 354                         inst_frc->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
 355
 356                         two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2, &two_swizzle);
 357                         inst_mad = rc_insert_new_instruction(c, inst->Prev);
 358
 359                         inst_mad->U.I.Opcode = RC_OPCODE_MAD;
 360                         inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
 361                         inst_mad->U.I.DstReg.Index = temp;
 362                         inst_mad->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 363                         inst_mad->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 364                         inst_mad->U.I.SrcReg[0].Index = temp;
 365                         inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
 366                         inst_mad->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
 367                         inst_mad->U.I.SrcReg[1].Index = two;
 368                         inst_mad->U.I.SrcReg[1].Swizzle = two_swizzle;
 369                         inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_1111;
 370                         inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZ;
 371
 372                         inst_add = rc_insert_new_instruction(c, inst->Prev);
 373
 374                         inst_add->U.I.Opcode = RC_OPCODE_ADD;
 375                         inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
 376                         inst_add->U.I.DstReg.Index = temp;
 377                         inst_add->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 378                         inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
 379                         inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 380                         inst_add->U.I.SrcReg[1].Index = temp;
 381                         inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XYZ0;
 382                         inst_add->U.I.SrcReg[1].Abs = 1;
 383                         inst_add->U.I.SrcReg[1].Negate = RC_MASK_XYZ;
 384                 } else if (wrapmode == RC_WRAP_MIRRORED_CLAMP) {
 385                         /*
 386                          * Mirrored clamp modes are bloody simple, we just use abs
 387                          * to mirror [0, 1] into [-1, 0]. This works for
 388                          * all modes i.e. CLAMP, CLAMP_TO_EDGE, and CLAMP_TO_BORDER.
 389                          */
 390                         struct rc_instruction *inst_mov;
 391
 392                         inst_mov = rc_insert_new_instruction(c, inst->Prev);
 393
 394                         inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 395                         inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 396                         inst_mov->U.I.DstReg.Index = temp;
 397                         inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 398                         inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 399                         inst_mov->U.I.SrcReg[0].Abs = 1;
 400                 }
 401
 402                 /* Preserve W for TXP/TXB. */
 403                 inst_mov = rc_insert_new_instruction(c, inst->Prev);
 404
 405                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 406                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 407                 inst_mov->U.I.DstReg.Index = temp;
 408                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
 409                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 410
 411                 reset_srcreg(&inst->U.I.SrcReg[0]);
 412                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 413                 inst->U.I.SrcReg[0].Index = temp;
 414         }
 415
 416         /* NPOT -> POT conversion for 3D textures. */
 417         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 418             compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch) {
 419                 struct rc_instruction *inst_mov;
 420                 unsigned temp = rc_find_free_temporary(c);
 421
 422                 /* Saturate XYZ. */
 423                 inst_mov = rc_insert_new_instruction(c, inst->Prev);
 424                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 425                 inst_mov->U.I.SaturateMode = RC_SATURATE_ZERO_ONE;
 426                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 427                 inst_mov->U.I.DstReg.Index = temp;
 428                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 429                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 430
 431                 /* Copy W. */
 432                 inst_mov = rc_insert_new_instruction(c, inst->Prev);
 433                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 434                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 435                 inst_mov->U.I.DstReg.Index = temp;
 436                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
 437                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 438
 439                 reset_srcreg(&inst->U.I.SrcReg[0]);
 440                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 441                 inst->U.I.SrcReg[0].Index = temp;
 442
 443                 scale_texcoords(compiler, inst, RC_STATE_R300_TEXSCALE_FACTOR);
 444         }
 445
 446         /* Convert SNORM-encoded ATI1N sampled as UNORM to SNORM.
 447          * Formula: dst = tex > 0.5 ? tex*2-2 : tex*2
 448          */
 449         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 450             compiler->state.unit[inst->U.I.TexSrcUnit].convert_unorm_to_snorm) {
 451                 unsigned two, two_swizzle;
 452                 struct rc_instruction *inst_mul, *inst_mad, *inst_cnd;
 453
 454                 two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2.35, &two_swizzle);
 455
 456                 inst_mul = rc_insert_new_instruction(c, inst);
 457                 inst_mul->U.I.Opcode = RC_OPCODE_MUL;
 458                 inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 459                 inst_mul->U.I.DstReg.Index = rc_find_free_temporary(c);
 460                 inst_mul->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 461                 inst_mul->U.I.SrcReg[0].Index = rc_find_free_temporary(c); /* redirected TEX output */
 462                 inst_mul->U.I.SrcReg[1].File = RC_FILE_CONSTANT; /* 2 */
 463                 inst_mul->U.I.SrcReg[1].Index = two;
 464                 inst_mul->U.I.SrcReg[1].Swizzle = two_swizzle;
 465
 466                 inst_mad = rc_insert_new_instruction(c, inst_mul);
 467                 inst_mad->U.I.Opcode = RC_OPCODE_MAD;
 468                 inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
 469                 inst_mad->U.I.DstReg.Index = rc_find_free_temporary(c);
 470                 inst_mad->U.I.SrcReg[0] = inst_mul->U.I.SrcReg[0]; /* redirected TEX output */
 471                 inst_mad->U.I.SrcReg[1] = inst_mul->U.I.SrcReg[1]; /* 2 */
 472                 inst_mad->U.I.SrcReg[2] = inst_mul->U.I.SrcReg[1]; /* 2 */
 473                 inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZW;
 474
 475                 inst_cnd = rc_insert_new_instruction(c, inst_mad);
 476                 inst_cnd->U.I.Opcode = RC_OPCODE_CND;
 477                 inst_cnd->U.I.SaturateMode = inst->U.I.SaturateMode;
 478                 inst_cnd->U.I.DstReg = inst->U.I.DstReg;
 479                 inst_cnd->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 480                 inst_cnd->U.I.SrcReg[0].Index = inst_mad->U.I.DstReg.Index;
 481                 inst_cnd->U.I.SrcReg[0].Swizzle = compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle;
 482                 inst_cnd->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 483                 inst_cnd->U.I.SrcReg[1].Index = inst_mul->U.I.DstReg.Index;
 484                 inst_cnd->U.I.SrcReg[1].Swizzle = compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle;
 485                 inst_cnd->U.I.SrcReg[2] = inst_mul->U.I.SrcReg[0]; /* redirected TEX output */
 486
 487                 inst->U.I.SaturateMode = 0;
 488                 inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 489                 inst->U.I.DstReg.Index = inst_mul->U.I.SrcReg[0].Index;
 490                 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 491         }
 492
 493         /* Cannot write texture to output registers or with saturate (all chips),
 494          * or with masks (non-r500). */
 495         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 496                 (inst->U.I.DstReg.File != RC_FILE_TEMPORARY ||
 497                  inst->U.I.SaturateMode ||
 498                  (!c->is_r500 && inst->U.I.DstReg.WriteMask != RC_MASK_XYZW))) {
 499                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst);
 500
 501                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 502                 inst_mov->U.I.SaturateMode = inst->U.I.SaturateMode;
 503                 inst_mov->U.I.DstReg = inst->U.I.DstReg;
 504                 inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 505                 inst_mov->U.I.SrcReg[0].Index = rc_find_free_temporary(c);
 506
 507                 inst->U.I.SaturateMode = 0;
 508                 inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 509                 inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
 510                 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 511         }
 512
 513         /* Cannot read texture coordinate from constants file */
 514         if (inst->U.I.SrcReg[0].File != RC_FILE_TEMPORARY && inst->U.I.SrcReg[0].File != RC_FILE_INPUT) {
 515                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
 516
 517                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 518                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 519                 inst_mov->U.I.DstReg.Index = rc_find_free_temporary(c);
 520                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 521
 522                 reset_srcreg(&inst->U.I.SrcReg[0]);
 523                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 524                 inst->U.I.SrcReg[0].Index = inst_mov->U.I.DstReg.Index;
 525         }
 526
 527         return 1;
 528 }