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

   1 /*
   2  * Copyright (C) 2010 Corbin Simpson
   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 "radeon_program_tex.h"
  29
  30 /* Series of transformations to be done on textures. */
  31
  32 static struct rc_src_register shadow_ambient(struct r300_fragment_program_compiler *compiler,
  33                                                                                          int tmu)
  34 {
  35         struct rc_src_register reg = { 0, };
  36
  37         if (compiler->enable_shadow_ambient) {
  38                 reg.File = RC_FILE_CONSTANT;
  39                 reg.Index = rc_constants_add_state(&compiler->Base.Program.Constants,
  40                                                                                    RC_STATE_SHADOW_AMBIENT, tmu);
  41                 reg.Swizzle = RC_SWIZZLE_WWWW;
  42         } else {
  43                 reg.File = RC_FILE_NONE;
  44                 reg.Swizzle = RC_SWIZZLE_0000;
  45         }
  46         return reg;
  47 }
  48
  49 static void lower_texture_rect(struct r300_fragment_program_compiler *compiler,
  50                                                            struct rc_instruction *inst)
  51 {
  52         struct rc_instruction *inst_rect;
  53         unsigned temp = rc_find_free_temporary(&compiler->Base);
  54
  55         if (inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
  56                 compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords) {
  57                 inst_rect = rc_insert_new_instruction(&compiler->Base, inst->Prev);
  58
  59                 inst_rect->U.I.Opcode = RC_OPCODE_MUL;
  60                 inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
  61                 inst_rect->U.I.DstReg.Index = temp;
  62                 inst_rect->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
  63                 inst_rect->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
  64                 inst_rect->U.I.SrcReg[1].Index =
  65                                 rc_constants_add_state(&compiler->Base.Program.Constants,
  66                                                                            RC_STATE_R300_TEXRECT_FACTOR, inst->U.I.TexSrcUnit);
  67
  68                 reset_srcreg(&inst->U.I.SrcReg[0]);
  69                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
  70                 inst->U.I.SrcReg[0].Index = temp;
  71
  72                 inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
  73         }
  74 }
  75
  76 /**
  77  * Transform TEX, TXP, TXB, and KIL instructions in the following ways:
  78  *  - implement texture compare (shadow extensions)
  79  *  - extract non-native source / destination operands
  80  *  - premultiply texture coordinates for RECT
  81  *  - extract operand swizzles
  82  *  - introduce a temporary register when write masks are needed
  83  */
  84 int radeonTransformTEX(
  85         struct radeon_compiler * c,
  86         struct rc_instruction * inst,
  87         void* data)
  88 {
  89         struct r300_fragment_program_compiler *compiler =
  90                 (struct r300_fragment_program_compiler*)data;
  91
  92         if (inst->U.I.Opcode != RC_OPCODE_TEX &&
  93                 inst->U.I.Opcode != RC_OPCODE_TXB &&
  94                 inst->U.I.Opcode != RC_OPCODE_TXP &&
  95                 inst->U.I.Opcode != RC_OPCODE_KIL)
  96                 return 0;
  97
  98         /* ARB_shadow & EXT_shadow_funcs */
  99         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 100                 c->Program.ShadowSamplers & (1 << inst->U.I.TexSrcUnit)) {
 101                 rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;
 102
 103                 if (comparefunc == RC_COMPARE_FUNC_NEVER || comparefunc == RC_COMPARE_FUNC_ALWAYS) {
 104                         inst->U.I.Opcode = RC_OPCODE_MOV;
 105
 106                         if (comparefunc == RC_COMPARE_FUNC_ALWAYS) {
 107                                 inst->U.I.SrcReg[0].File = RC_FILE_NONE;
 108                                 inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
 109                         } else {
 110                                 inst->U.I.SrcReg[0] = shadow_ambient(compiler, inst->U.I.TexSrcUnit);
 111                         }
 112
 113                         return 1;
 114                 } else {
 115                         rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;
 116                         unsigned int depthmode = compiler->state.unit[inst->U.I.TexSrcUnit].depth_texture_mode;
 117                         struct rc_instruction * inst_rcp = rc_insert_new_instruction(c, inst);
 118                         struct rc_instruction * inst_mad = rc_insert_new_instruction(c, inst_rcp);
 119                         struct rc_instruction * inst_cmp = rc_insert_new_instruction(c, inst_mad);
 120                         int pass, fail;
 121
 122                         inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
 123                         inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
 124                         inst_rcp->U.I.DstReg.Index = rc_find_free_temporary(c);
 125                         inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
 126                         inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 127                         inst_rcp->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_WWWW;
 128
 129                         inst_cmp->U.I.DstReg = inst->U.I.DstReg;
 130                         inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 131                         inst->U.I.DstReg.Index = rc_find_free_temporary(c);
 132                         inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 133
 134                         inst_mad->U.I.Opcode = RC_OPCODE_MAD;
 135                         inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
 136                         inst_mad->U.I.DstReg.Index = rc_find_free_temporary(c);
 137                         inst_mad->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 138                         inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_ZZZZ;
 139                         inst_mad->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 140                         inst_mad->U.I.SrcReg[1].Index = inst_rcp->U.I.DstReg.Index;
 141                         inst_mad->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
 142                         inst_mad->U.I.SrcReg[2].File = RC_FILE_TEMPORARY;
 143                         inst_mad->U.I.SrcReg[2].Index = inst->U.I.DstReg.Index;
 144                         if (depthmode == 0) /* GL_LUMINANCE */
 145                                 inst_mad->U.I.SrcReg[2].Swizzle = RC_MAKE_SWIZZLE(RC_SWIZZLE_X, RC_SWIZZLE_Y, RC_SWIZZLE_Z, RC_SWIZZLE_Z);
 146                         else if (depthmode == 2) /* GL_ALPHA */
 147                                 inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_WWWW;
 148
 149                         /* Recall that SrcReg[0] is tex, SrcReg[2] is r and:
 150                          *   r  < tex  <=>      -tex+r < 0
 151                          *   r >= tex  <=> not (-tex+r < 0 */
 152                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GEQUAL)
 153                                 inst_mad->U.I.SrcReg[2].Negate = inst_mad->U.I.SrcReg[2].Negate ^ RC_MASK_XYZW;
 154                         else
 155                                 inst_mad->U.I.SrcReg[0].Negate = inst_mad->U.I.SrcReg[0].Negate ^ RC_MASK_XYZW;
 156
 157                         inst_cmp->U.I.Opcode = RC_OPCODE_CMP;
 158                         /* DstReg has been filled out above */
 159                         inst_cmp->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 160                         inst_cmp->U.I.SrcReg[0].Index = inst_mad->U.I.DstReg.Index;
 161
 162                         if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GREATER) {
 163                                 pass = 1;
 164                                 fail = 2;
 165                         } else {
 166                                 pass = 2;
 167                                 fail = 1;
 168                         }
 169
 170                         inst_cmp->U.I.SrcReg[pass].File = RC_FILE_NONE;
 171                         inst_cmp->U.I.SrcReg[pass].Swizzle = RC_SWIZZLE_1111;
 172                         inst_cmp->U.I.SrcReg[fail] = shadow_ambient(compiler, inst->U.I.TexSrcUnit);
 173                 }
 174         }
 175
 176         /* Texture wrap modes don't work on NPOT textures or texrects.
 177          *
 178          * The game plan is simple. We have two flags, fake_npot and
 179          * non_normalized_coords, as well as a tex target. The RECT tex target
 180          * will make the emitted code use non-scaled texcoords.
 181          *
 182          * Non-wrapped/clamped texcoords with NPOT are free in HW. Repeat and
 183          * mirroring are not. If we need to repeat, we do:
 184          *
 185          * MUL temp, texcoord, <scaling factor constant>
 186          * FRC temp, temp ; Discard integer portion of coords
 187          *
 188          * This gives us coords in [0, 1].
 189          *
 190          * Mirroring is trickier. We're going to start out like repeat:
 191          *
 192          * MUL temp, texcoord, <scaling factor constant> ; De-mirror across axes
 193          * MUL temp, temp, 0.5 ; Pattern repeats in [0, 2]
 194          *                            ; so scale to [0, 1]
 195          * FRC temp, temp ; Make the pattern repeat
 196          * MAD temp, temp, 2, -1 ; Move the pattern to [-1, 1]
 197          * ADD temp, 1, -abs(temp) ; Now comes a neat trick: use abs to mirror the pattern.
 198          *                              ; The pattern is backwards, so reverse it (1-x).
 199          *
 200          * This gives us coords in [0, 1].
 201          *
 202          * ~ C & M. ;)
 203          */
 204         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 205                 (inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
 206                         compiler->state.unit[inst->U.I.TexSrcUnit].fake_npot ||
 207                         compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords)) {
 208                 rc_wrap_mode wrapmode = compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode;
 209
 210                 /* R300 cannot sample from rectangles. */
 211                 if (!compiler->is_r500) {
 212                         lower_texture_rect(compiler, inst);
 213                 }
 214
 215                 if (compiler->state.unit[inst->U.I.TexSrcUnit].fake_npot &&
 216                         wrapmode != RC_WRAP_NONE) {
 217                         struct rc_instruction *inst_mov;
 218                         unsigned temp = rc_find_free_temporary(c);
 219
 220                         /* For NPOT fallback, we need normalized coordinates anyway. */
 221                         if (compiler->is_r500) {
 222                                 lower_texture_rect(compiler, inst);
 223                         }
 224
 225                         if (wrapmode == RC_WRAP_REPEAT) {
 226                                 /* Both instructions will be paired up. */
 227                                 struct rc_instruction *inst_frc = rc_insert_new_instruction(c, inst->Prev);
 228
 229                                 inst_frc->U.I.Opcode = RC_OPCODE_FRC;
 230                                 inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
 231                                 inst_frc->U.I.DstReg.Index = temp;
 232                                 inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 233                                 inst_frc->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 234                         } else if (wrapmode == RC_WRAP_MIRRORED_REPEAT) {
 235                                 /*
 236                                  * Function:
 237                                  *   f(v) = 1 - abs(frac(v * 0.5) * 2 - 1)
 238                                  *
 239                                  * Code:
 240                                  *   MUL temp, src0, 0.5
 241                                  *   FRC temp, temp
 242                                  *   MAD temp, temp, 2, -1
 243                                  *   ADD temp, 1, -abs(temp)
 244                                  */
 245
 246                                 struct rc_instruction *inst_mul, *inst_frc, *inst_mad, *inst_add;
 247                                 unsigned two, two_swizzle;
 248
 249                                 inst_mul = rc_insert_new_instruction(c, inst->Prev);
 250
 251                                 inst_mul->U.I.Opcode = RC_OPCODE_MUL;
 252                                 inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
 253                                 inst_mul->U.I.DstReg.Index = temp;
 254                                 inst_mul->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 255                                 inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 256                                 inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_HHHH;
 257
 258                                 inst_frc = rc_insert_new_instruction(c, inst->Prev);
 259
 260                                 inst_frc->U.I.Opcode = RC_OPCODE_FRC;
 261                                 inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
 262                                 inst_frc->U.I.DstReg.Index = temp;
 263                                 inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 264                                 inst_frc->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 265                                 inst_frc->U.I.SrcReg[0].Index = temp;
 266                                 inst_frc->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
 267
 268                                 two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2, &two_swizzle);
 269                                 inst_mad = rc_insert_new_instruction(c, inst->Prev);
 270
 271                                 inst_mad->U.I.Opcode = RC_OPCODE_MAD;
 272                                 inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
 273                                 inst_mad->U.I.DstReg.Index = temp;
 274                                 inst_mad->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 275                                 inst_mad->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 276                                 inst_mad->U.I.SrcReg[0].Index = temp;
 277                                 inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
 278                                 inst_mad->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
 279                                 inst_mad->U.I.SrcReg[1].Index = two;
 280                                 inst_mad->U.I.SrcReg[1].Swizzle = two_swizzle;
 281                                 inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_1111;
 282                                 inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZ;
 283
 284                                 inst_add = rc_insert_new_instruction(c, inst->Prev);
 285
 286                                 inst_add->U.I.Opcode = RC_OPCODE_ADD;
 287                                 inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
 288                                 inst_add->U.I.DstReg.Index = temp;
 289                                 inst_add->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 290                                 inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
 291                                 inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 292                                 inst_add->U.I.SrcReg[1].Index = temp;
 293                                 inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XYZ0;
 294                                 inst_add->U.I.SrcReg[1].Abs = 1;
 295                                 inst_add->U.I.SrcReg[1].Negate = RC_MASK_XYZ;
 296                         } else if (wrapmode == RC_WRAP_MIRRORED_CLAMP) {
 297                                 /*
 298                                  * Mirrored clamp modes are bloody simple, we just use abs
 299                                  * to mirror [0, 1] into [-1, 0]. This works for
 300                                  * all modes i.e. CLAMP, CLAMP_TO_EDGE, and CLAMP_TO_BORDER.
 301                                  */
 302                                 struct rc_instruction *inst_mov;
 303
 304                                 inst_mov = rc_insert_new_instruction(c, inst->Prev);
 305
 306                                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 307                                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 308                                 inst_mov->U.I.DstReg.Index = temp;
 309                                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
 310                                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 311                                 inst_mov->U.I.SrcReg[0].Abs = 1;
 312                         }
 313
 314                         /* Preserve W for TXP/TXB. */
 315                         inst_mov = rc_insert_new_instruction(c, inst->Prev);
 316
 317                         inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 318                         inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 319                         inst_mov->U.I.DstReg.Index = temp;
 320                         inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
 321                         inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 322
 323                         reset_srcreg(&inst->U.I.SrcReg[0]);
 324                         inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 325                         inst->U.I.SrcReg[0].Index = temp;
 326                 }
 327         }
 328
 329         /* Cannot write texture to output registers (all chips) or with masks (non-r500) */
 330         if (inst->U.I.Opcode != RC_OPCODE_KIL &&
 331                 (inst->U.I.DstReg.File != RC_FILE_TEMPORARY ||
 332                  (!compiler->is_r500 && inst->U.I.DstReg.WriteMask != RC_MASK_XYZW))) {
 333                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst);
 334
 335                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 336                 inst_mov->U.I.DstReg = inst->U.I.DstReg;
 337                 inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 338                 inst_mov->U.I.SrcReg[0].Index = rc_find_free_temporary(c);
 339
 340                 inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
 341                 inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
 342                 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 343         }
 344
 345         /* Cannot read texture coordinate from constants file */
 346         if (inst->U.I.SrcReg[0].File != RC_FILE_TEMPORARY && inst->U.I.SrcReg[0].File != RC_FILE_INPUT) {
 347                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
 348
 349                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 350                 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 351                 inst_mov->U.I.DstReg.Index = rc_find_free_temporary(c);
 352                 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
 353
 354                 reset_srcreg(&inst->U.I.SrcReg[0]);
 355                 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 356                 inst->U.I.SrcReg[0].Index = inst_mov->U.I.DstReg.Index;
 357         }
 358
 359         return 1;
 360 }