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