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_KIL)
 146                 return 0;
 147
 148         /* ARB_shadow & EXT_shadow_funcs */
 149         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 150                 ((c->Program.ShadowSamplers & (1 << inst->U.I.TexSrcUnit)) ||
 151                  (compiler->state.unit[inst->U.I.TexSrcUnit].compare_mode_enabled))) {
 152                 rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;
 153
 154                 if (comparefunc == RC_COMPARE_FUNC_NEVER || comparefunc == RC_COMPARE_FUNC_ALWAYS) {
 155                         inst->U.I.Opcode = RC_OPCODE_MOV;
 156
 157                         if (comparefunc == RC_COMPARE_FUNC_ALWAYS) {
 158                                 inst->U.I.SrcReg[0] = shadow_pass_value(compiler, inst->U.I.TexSrcUnit);
 159                         } else {
 160                                 inst->U.I.SrcReg[0] = shadow_fail_value(compiler, inst->U.I.TexSrcUnit);
 161                         }
 162
 163                         return 1;
 164                 } else {
 165                         struct rc_instruction * inst_rcp = NULL;
 166                         struct rc_instruction *inst_mul, *inst_add, *inst_cmp;
 167                         unsigned tmp_texsample;
 168                         unsigned tmp_sum;
 169                         int pass, fail;
 170
 171                         /* Save the output register. */
 172                         struct rc_dst_register output_reg = inst->U.I.DstReg;
 173                         unsigned saturate_mode = inst->U.I.SaturateMode;
 174
 175                         /* Redirect TEX to a new temp. */
 176                         tmp_texsample = rc_find_free_temporary(c);
 177                         inst->U.I.SaturateMode = 0;
 178                         inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 179                         inst->U.I.DstReg.Index = tmp_texsample;
 180                         inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 181
 182                         tmp_sum = rc_find_free_temporary(c);
 183
 184                         if (inst->U.I.Opcode == RC_OPCODE_TXP) {
 185                                 /* Compute 1/W. */
 186                                 inst_rcp = rc_insert_new_instruction(c, inst);
 187                                 inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
 188                                 inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
 189                                 inst_rcp->U.I.DstReg.Index = tmp_sum;
 190                                 inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
 191                                 inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 192                                 inst_rcp->U.I.SrcReg[0].Swizzle =
 193                                         RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 3));
 194                         }
 195
 196                         /* Divide Z by W (if it's TXP) and saturate. */
 197                         inst_mul = rc_insert_new_instruction(c, inst_rcp ? inst_rcp : inst);
 198                         inst_mul->U.I.Opcode = inst->U.I.Opcode == RC_OPCODE_TXP ? RC_OPCODE_MUL : RC_OPCODE_MOV;
 199                         inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 200                         inst_mul->U.I.DstReg.Index = tmp_sum;
 201                         inst_mul->U.I.DstReg.WriteMask = RC_MASK_W;
 202                         inst_mul->U.I.SaturateMode = RC_SATURATE_ZERO_ONE;
 203                         inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 204                         inst_mul->U.I.SrcReg[0].Swizzle =
 205                                 RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 2));
 206                         if (inst->U.I.Opcode == RC_OPCODE_TXP) {
 207                                 inst_mul->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 208                                 inst_mul->U.I.SrcReg[1].Index = tmp_sum;
 209                                 inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
 210                         }
 211
 212                         /* Add the depth texture value. */
 213                         inst_add = rc_insert_new_instruction(c, inst_mul);
 214                         inst_add->U.I.Opcode = RC_OPCODE_ADD;
 215                         inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
 216                         inst_add->U.I.DstReg.Index = tmp_sum;
 217                         inst_add->U.I.DstReg.WriteMask = RC_MASK_W;
 218                         inst_add->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 219                         inst_add->U.I.SrcReg[0].Index = tmp_sum;
 220                         inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_WWWW;
 221                         inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 222                         inst_add->U.I.SrcReg[1].Index = tmp_texsample;
 223                         inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XXXX;
 224
 225                         /* Note that SrcReg[0] is r, SrcReg[1] is tex and:
 226                          *   LESS:    r  < tex  <=>      -tex+r < 0
 227                          *   GEQUAL:  r >= tex  <=> not (-tex+r < 0)
 228                          *   GREATER: r  > tex  <=>       tex-r < 0
 229                          *   LEQUAL:  r <= tex  <=> not ( tex-r < 0)
 230                          *   EQUAL:   GEQUAL
 231                          *   NOTEQUAL:LESS
 232                          */
 233
 234                         /* This negates either r or tex: */
 235                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GEQUAL ||
 236                             comparefunc == RC_COMPARE_FUNC_EQUAL || comparefunc == RC_COMPARE_FUNC_NOTEQUAL)
 237                                 inst_add->U.I.SrcReg[1].Negate = inst_add->U.I.SrcReg[1].Negate ^ RC_MASK_XYZW;
 238                         else
 239                                 inst_add->U.I.SrcReg[0].Negate = inst_add->U.I.SrcReg[0].Negate ^ RC_MASK_XYZW;
 240
 241                         /* This negates the whole expresion: */
 242                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GREATER ||
 243                             comparefunc == RC_COMPARE_FUNC_NOTEQUAL) {
 244                                 pass = 1;
 245                                 fail = 2;
 246                         } else {
 247                                 pass = 2;
 248                                 fail = 1;
 249                         }
 250
 251                         inst_cmp = rc_insert_new_instruction(c, inst_add);
 252                         inst_cmp->U.I.Opcode = RC_OPCODE_CMP;
 253                         inst_cmp->U.I.SaturateMode = saturate_mode;
 254                         inst_cmp->U.I.DstReg = output_reg;
 255                         inst_cmp->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 256                         inst_cmp->U.I.SrcReg[0].Index = tmp_sum;
 257                         inst_cmp->U.I.SrcReg[0].Swizzle =
 258                                         combine_swizzles(RC_SWIZZLE_WWWW,
 259                                                          compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle);
 260                         inst_cmp->U.I.SrcReg[pass] = shadow_pass_value(compiler, inst->U.I.TexSrcUnit);
 261                         inst_cmp->U.I.SrcReg[fail] = shadow_fail_value(compiler, inst->U.I.TexSrcUnit);
 262
 263                         assert(tmp_texsample != tmp_sum);
 264                 }
 265         }
 266
 267         /* R300 cannot sample from rectangles and the wrap mode fallback needs
 268          * normalized coordinates anyway. */
 269         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 270             is_rect && (!c->is_r500 || wrapmode != RC_WRAP_NONE)) {
 271                 scale_texcoords(compiler, inst, RC_STATE_R300_TEXRECT_FACTOR);
 272                 inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
 273         }
 274
 275         /* Divide by W if needed. */
 276         if (inst->U.I.Opcode == RC_OPCODE_TXP &&
 277             (wrapmode == RC_WRAP_REPEAT || wrapmode == RC_WRAP_MIRRORED_REPEAT ||
 278              compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch)) {
 279                 projective_divide(compiler, inst);
 280         }
 281
 282         /* Texture wrap modes don't work on NPOT textures.
 283          *
 284          * Non-wrapped/clamped texcoords with NPOT are free in HW. Repeat and
 285          * mirroring are not. If we need to repeat, we do:
 286          *
 287          * MUL temp, texcoord, <scaling factor constant>
 288          * FRC temp, temp ; Discard integer portion of coords
 289          *
 290          * This gives us coords in [0, 1].
 291          *
 292          * Mirroring is trickier. We're going to start out like repeat:
 293          *
 294          * MUL temp, texcoord, <scaling factor constant> ; De-mirror across axes
 295          * MUL temp, temp, 0.5 ; Pattern repeats in [0, 2]
 296          *                            ; so scale to [0, 1]
 297          * FRC temp, temp ; Make the pattern repeat
 298          * MAD temp, temp, 2, -1 ; Move the pattern to [-1, 1]
 299          * ADD temp, 1, -abs(temp) ; Now comes a neat trick: use abs to mirror the pattern.
 300          *                              ; The pattern is backwards, so reverse it (1-x).
 301          *
 302          * This gives us coords in [0, 1].
 303          *
 304          * ~ C & M. ;)
 305          */
 306         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 307             wrapmode != RC_WRAP_NONE) {
 308                 struct rc_instruction *inst_mov;
 309                 unsigned temp = rc_find_free_temporary(c);
 310
 311                 if (wrapmode == RC_WRAP_REPEAT) {
 312                         /* Both instructions will be paired up. */
 313                         struct rc_instruction *inst_frc = rc_insert_new_instruction(c, inst->Prev);
 314
 315                         inst_frc->U.I.Opcode = RC_OPCODE_FRC;
 316                         inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
 317                         inst_frc->U.I.DstReg.Index = temp;
 318                         inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 319                         inst_frc->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 320                 } else if (wrapmode == RC_WRAP_MIRRORED_REPEAT) {
 321                         /*
 322                          * Function:
 323                          *   f(v) = 1 - abs(frac(v * 0.5) * 2 - 1)
 324                          *
 325                          * Code:
 326                          *   MUL temp, src0, 0.5
 327                          *   FRC temp, temp
 328                          *   MAD temp, temp, 2, -1
 329                          *   ADD temp, 1, -abs(temp)
 330                          */
 331
 332                         struct rc_instruction *inst_mul, *inst_frc, *inst_mad, *inst_add;
 333                         unsigned two, two_swizzle;
 334
 335                         inst_mul = rc_insert_new_instruction(c, inst->Prev);
 336
 337                         inst_mul->U.I.Opcode = RC_OPCODE_MUL;
 338                         inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 339                         inst_mul->U.I.DstReg.Index = temp;
 340                         inst_mul->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 341                         inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 342                         inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_HHHH;
 343
 344                         inst_frc = rc_insert_new_instruction(c, inst->Prev);
 345
 346                         inst_frc->U.I.Opcode = RC_OPCODE_FRC;
 347                         inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
 348                         inst_frc->U.I.DstReg.Index = temp;
 349                         inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 350                         inst_frc->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 351                         inst_frc->U.I.SrcReg[0].Index = temp;
 352                         inst_frc->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
 353
 354                         two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2, &two_swizzle);
 355                         inst_mad = rc_insert_new_instruction(c, inst->Prev);
 356
 357                         inst_mad->U.I.Opcode = RC_OPCODE_MAD;
 358                         inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
 359                         inst_mad->U.I.DstReg.Index = temp;
 360                         inst_mad->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 361                         inst_mad->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 362                         inst_mad->U.I.SrcReg[0].Index = temp;
 363                         inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
 364                         inst_mad->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
 365                         inst_mad->U.I.SrcReg[1].Index = two;
 366                         inst_mad->U.I.SrcReg[1].Swizzle = two_swizzle;
 367                         inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_1111;
 368                         inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZ;
 369
 370                         inst_add = rc_insert_new_instruction(c, inst->Prev);
 371
 372                         inst_add->U.I.Opcode = RC_OPCODE_ADD;
 373                         inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
 374                         inst_add->U.I.DstReg.Index = temp;
 375                         inst_add->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 376                         inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
 377                         inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 378                         inst_add->U.I.SrcReg[1].Index = temp;
 379                         inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XYZ0;
 380                         inst_add->U.I.SrcReg[1].Abs = 1;
 381                         inst_add->U.I.SrcReg[1].Negate = RC_MASK_XYZ;
 382                 } else if (wrapmode == RC_WRAP_MIRRORED_CLAMP) {
 383                         /*
 384                          * Mirrored clamp modes are bloody simple, we just use abs
 385                          * to mirror [0, 1] into [-1, 0]. This works for
 386                          * all modes i.e. CLAMP, CLAMP_TO_EDGE, and CLAMP_TO_BORDER.
 387                          */
 388                         struct rc_instruction *inst_mov;
 389
 390                         inst_mov = rc_insert_new_instruction(c, inst->Prev);
 391
 392                         inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 393                         inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 394                         inst_mov->U.I.DstReg.Index = temp;
 395                         inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 396                         inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 397                         inst_mov->U.I.SrcReg[0].Abs = 1;
 398                 }
 399
 400                 /* Preserve W for TXP/TXB. */
 401                 inst_mov = rc_insert_new_instruction(c, inst->Prev);
 402
 403                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 404                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 405                 inst_mov->U.I.DstReg.Index = temp;
 406                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
 407                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 408
 409                 reset_srcreg(&inst->U.I.SrcReg[0]);
 410                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 411                 inst->U.I.SrcReg[0].Index = temp;
 412         }
 413
 414         /* NPOT -> POT conversion for 3D textures. */
 415         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 416             compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch) {
 417                 struct rc_instruction *inst_mov;
 418                 unsigned temp = rc_find_free_temporary(c);
 419
 420                 /* Saturate XYZ. */
 421                 inst_mov = rc_insert_new_instruction(c, inst->Prev);
 422                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 423                 inst_mov->U.I.SaturateMode = RC_SATURATE_ZERO_ONE;
 424                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 425                 inst_mov->U.I.DstReg.Index = temp;
 426                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 427                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 428
 429                 /* Copy W. */
 430                 inst_mov = rc_insert_new_instruction(c, inst->Prev);
 431                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 432                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 433                 inst_mov->U.I.DstReg.Index = temp;
 434                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
 435                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 436
 437                 reset_srcreg(&inst->U.I.SrcReg[0]);
 438                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 439                 inst->U.I.SrcReg[0].Index = temp;
 440
 441                 scale_texcoords(compiler, inst, RC_STATE_R300_TEXSCALE_FACTOR);
 442         }
 443
 444         /* Convert SNORM-encoded ATI1N sampled as UNORM to SNORM.
 445          * Formula: dst = tex > 0.5 ? tex*2-2 : tex*2
 446          */
 447         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 448             compiler->state.unit[inst->U.I.TexSrcUnit].convert_unorm_to_snorm) {
 449                 unsigned two, two_swizzle;
 450                 struct rc_instruction *inst_mul, *inst_mad, *inst_cnd;
 451
 452                 two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2.35, &two_swizzle);
 453
 454                 inst_mul = rc_insert_new_instruction(c, inst);
 455                 inst_mul->U.I.Opcode = RC_OPCODE_MUL;
 456                 inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 457                 inst_mul->U.I.DstReg.Index = rc_find_free_temporary(c);
 458                 inst_mul->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 459                 inst_mul->U.I.SrcReg[0].Index = rc_find_free_temporary(c); /* redirected TEX output */
 460                 inst_mul->U.I.SrcReg[1].File = RC_FILE_CONSTANT; /* 2 */
 461                 inst_mul->U.I.SrcReg[1].Index = two;
 462                 inst_mul->U.I.SrcReg[1].Swizzle = two_swizzle;
 463
 464                 inst_mad = rc_insert_new_instruction(c, inst_mul);
 465                 inst_mad->U.I.Opcode = RC_OPCODE_MAD;
 466                 inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
 467                 inst_mad->U.I.DstReg.Index = rc_find_free_temporary(c);
 468                 inst_mad->U.I.SrcReg[0] = inst_mul->U.I.SrcReg[0]; /* redirected TEX output */
 469                 inst_mad->U.I.SrcReg[1] = inst_mul->U.I.SrcReg[1]; /* 2 */
 470                 inst_mad->U.I.SrcReg[2] = inst_mul->U.I.SrcReg[1]; /* 2 */
 471                 inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZW;
 472
 473                 inst_cnd = rc_insert_new_instruction(c, inst_mad);
 474                 inst_cnd->U.I.Opcode = RC_OPCODE_CND;
 475                 inst_cnd->U.I.SaturateMode = inst->U.I.SaturateMode;
 476                 inst_cnd->U.I.DstReg = inst->U.I.DstReg;
 477                 inst_cnd->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 478                 inst_cnd->U.I.SrcReg[0].Index = inst_mad->U.I.DstReg.Index;
 479                 inst_cnd->U.I.SrcReg[0].Swizzle = compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle;
 480                 inst_cnd->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 481                 inst_cnd->U.I.SrcReg[1].Index = inst_mul->U.I.DstReg.Index;
 482                 inst_cnd->U.I.SrcReg[1].Swizzle = compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle;
 483                 inst_cnd->U.I.SrcReg[2] = inst_mul->U.I.SrcReg[0]; /* redirected TEX output */
 484
 485                 inst->U.I.SaturateMode = 0;
 486                 inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 487                 inst->U.I.DstReg.Index = inst_mul->U.I.SrcReg[0].Index;
 488                 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 489         }
 490
 491         /* Cannot write texture to output registers or with saturate (all chips),
 492          * or with masks (non-r500). */
 493         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 494                 (inst->U.I.DstReg.File != RC_FILE_TEMPORARY ||
 495                  inst->U.I.SaturateMode ||
 496                  (!c->is_r500 && inst->U.I.DstReg.WriteMask != RC_MASK_XYZW))) {
 497                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst);
 498
 499                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 500                 inst_mov->U.I.SaturateMode = inst->U.I.SaturateMode;
 501                 inst_mov->U.I.DstReg = inst->U.I.DstReg;
 502                 inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 503                 inst_mov->U.I.SrcReg[0].Index = rc_find_free_temporary(c);
 504
 505                 inst->U.I.SaturateMode = 0;
 506                 inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 507                 inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
 508                 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 509         }
 510
 511         /* Cannot read texture coordinate from constants file */
 512         if (inst->U.I.SrcReg[0].File != RC_FILE_TEMPORARY && inst->U.I.SrcReg[0].File != RC_FILE_INPUT) {
 513                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
 514
 515                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 516                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 517                 inst_mov->U.I.DstReg.Index = rc_find_free_temporary(c);
 518                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 519
 520                 reset_srcreg(&inst->U.I.SrcReg[0]);
 521                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 522                 inst->U.I.SrcReg[0].Index = inst_mov->U.I.DstReg.Index;
 523         }
 524
 525         return 1;
 526 }