src/mesa/drivers/dri/r300/compiler/r500_fragprog.c

   1 /*
   2  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  *
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining
   7  * a copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sublicense, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial
  16  * portions of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  */
  27
  28 #include "r500_fragprog.h"
  29
  30 #include "../r300_reg.h"
  31
  32 static struct prog_src_register shadow_ambient(struct radeon_compiler * c, int tmu)
  33 {
  34         struct prog_src_register reg = { 0, };
  35
  36         reg.File = PROGRAM_STATE_VAR;
  37         reg.Index = rc_constants_add_state(&c->Program.Constants, RC_STATE_SHADOW_AMBIENT, tmu);
  38         reg.Swizzle = SWIZZLE_WWWW;
  39         return reg;
  40 }
  41
  42 /**
  43  * Transform TEX, TXP, TXB, and KIL instructions in the following way:
  44  *  - implement texture compare (shadow extensions)
  45  *  - extract non-native source / destination operands
  46  */
  47 GLboolean r500_transform_TEX(
  48         struct radeon_compiler * c,
  49         struct rc_instruction * inst,
  50         void* data)
  51 {
  52         struct r300_fragment_program_compiler *compiler =
  53                 (struct r300_fragment_program_compiler*)data;
  54
  55         if (inst->I.Opcode != OPCODE_TEX &&
  56             inst->I.Opcode != OPCODE_TXB &&
  57             inst->I.Opcode != OPCODE_TXP &&
  58             inst->I.Opcode != OPCODE_KIL)
  59                 return GL_FALSE;
  60
  61         /* ARB_shadow & EXT_shadow_funcs */
  62         if (inst->I.Opcode != OPCODE_KIL &&
  63             c->Program.ShadowSamplers & (1 << inst->I.TexSrcUnit)) {
  64                 GLuint comparefunc = GL_NEVER + compiler->state.unit[inst->I.TexSrcUnit].texture_compare_func;
  65
  66                 if (comparefunc == GL_NEVER || comparefunc == GL_ALWAYS) {
  67                         inst->I.Opcode = OPCODE_MOV;
  68
  69                         if (comparefunc == GL_ALWAYS) {
  70                                 inst->I.SrcReg[0].File = PROGRAM_BUILTIN;
  71                                 inst->I.SrcReg[0].Swizzle = SWIZZLE_1111;
  72                         } else {
  73                                 inst->I.SrcReg[0] = shadow_ambient(c, inst->I.TexSrcUnit);
  74                         }
  75
  76                         return GL_TRUE;
  77                 } else {
  78                         GLuint comparefunc = GL_NEVER + compiler->state.unit[inst->I.TexSrcUnit].texture_compare_func;
  79                         GLuint depthmode = compiler->state.unit[inst->I.TexSrcUnit].depth_texture_mode;
  80                         struct rc_instruction * inst_rcp = rc_insert_new_instruction(c, inst);
  81                         struct rc_instruction * inst_mad = rc_insert_new_instruction(c, inst_rcp);
  82                         struct rc_instruction * inst_cmp = rc_insert_new_instruction(c, inst_mad);
  83                         int pass, fail;
  84
  85                         inst_rcp->I.Opcode = OPCODE_RCP;
  86                         inst_rcp->I.DstReg.File = PROGRAM_TEMPORARY;
  87                         inst_rcp->I.DstReg.Index = rc_find_free_temporary(c);
  88                         inst_rcp->I.DstReg.WriteMask = WRITEMASK_W;
  89                         inst_rcp->I.SrcReg[0] = inst->I.SrcReg[0];
  90                         inst_rcp->I.SrcReg[0].Swizzle = SWIZZLE_WWWW;
  91
  92                         inst_cmp->I.DstReg = inst->I.DstReg;
  93                         inst->I.DstReg.File = PROGRAM_TEMPORARY;
  94                         inst->I.DstReg.Index = rc_find_free_temporary(c);
  95                         inst->I.DstReg.WriteMask = WRITEMASK_XYZW;
  96
  97                         inst_mad->I.Opcode = OPCODE_MAD;
  98                         inst_mad->I.DstReg.File = PROGRAM_TEMPORARY;
  99                         inst_mad->I.DstReg.Index = rc_find_free_temporary(c);
 100                         inst_mad->I.SrcReg[0] = inst->I.SrcReg[0];
 101                         inst_mad->I.SrcReg[0].Swizzle = SWIZZLE_ZZZZ;
 102                         inst_mad->I.SrcReg[1].File = PROGRAM_TEMPORARY;
 103                         inst_mad->I.SrcReg[1].Index = inst_rcp->I.DstReg.Index;
 104                         inst_mad->I.SrcReg[1].Swizzle = SWIZZLE_WWWW;
 105                         inst_mad->I.SrcReg[2].File = PROGRAM_TEMPORARY;
 106                         inst_mad->I.SrcReg[2].Index = inst->I.DstReg.Index;
 107                         if (depthmode == 0) /* GL_LUMINANCE */
 108                                 inst_mad->I.SrcReg[2].Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_Z);
 109                         else if (depthmode == 2) /* GL_ALPHA */
 110                                 inst_mad->I.SrcReg[2].Swizzle = SWIZZLE_WWWW;
 111
 112                         /* Recall that SrcReg[0] is tex, SrcReg[2] is r and:
 113                          *   r  < tex  <=>      -tex+r < 0
 114                          *   r >= tex  <=> not (-tex+r < 0 */
 115                         if (comparefunc == GL_LESS || comparefunc == GL_GEQUAL)
 116                                 inst_mad->I.SrcReg[2].Negate = inst_mad->I.SrcReg[2].Negate ^ NEGATE_XYZW;
 117                         else
 118                                 inst_mad->I.SrcReg[0].Negate = inst_mad->I.SrcReg[0].Negate ^ NEGATE_XYZW;
 119
 120                         inst_cmp->I.Opcode = OPCODE_CMP;
 121                         /* DstReg has been filled out above */
 122                         inst_cmp->I.SrcReg[0].File = PROGRAM_TEMPORARY;
 123                         inst_cmp->I.SrcReg[0].Index = inst_mad->I.DstReg.Index;
 124
 125                         if (comparefunc == GL_LESS || comparefunc == GL_GREATER) {
 126                                 pass = 1;
 127                                 fail = 2;
 128                         } else {
 129                                 pass = 2;
 130                                 fail = 1;
 131                         }
 132
 133                         inst_cmp->I.SrcReg[pass].File = PROGRAM_BUILTIN;
 134                         inst_cmp->I.SrcReg[pass].Swizzle = SWIZZLE_1111;
 135                         inst_cmp->I.SrcReg[fail] = shadow_ambient(c, inst->I.TexSrcUnit);
 136                 }
 137         }
 138
 139         /* Cannot write texture to output registers */
 140         if (inst->I.Opcode != OPCODE_KIL && inst->I.DstReg.File != PROGRAM_TEMPORARY) {
 141                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst);
 142
 143                 inst_mov->I.Opcode = OPCODE_MOV;
 144                 inst_mov->I.DstReg = inst->I.DstReg;
 145                 inst_mov->I.SrcReg[0].File = PROGRAM_TEMPORARY;
 146                 inst_mov->I.SrcReg[0].Index = rc_find_free_temporary(c);
 147
 148                 inst->I.DstReg.File = PROGRAM_TEMPORARY;
 149                 inst->I.DstReg.Index = inst_mov->I.SrcReg[0].Index;
 150                 inst->I.DstReg.WriteMask = WRITEMASK_XYZW;
 151         }
 152
 153         /* Cannot read texture coordinate from constants file */
 154         if (inst->I.SrcReg[0].File != PROGRAM_TEMPORARY && inst->I.SrcReg[0].File != PROGRAM_INPUT) {
 155                 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
 156
 157                 inst_mov->I.Opcode = OPCODE_MOV;
 158                 inst_mov->I.DstReg.File = PROGRAM_TEMPORARY;
 159                 inst_mov->I.DstReg.Index = rc_find_free_temporary(c);
 160                 inst_mov->I.SrcReg[0] = inst->I.SrcReg[0];
 161
 162                 reset_srcreg(&inst->I.SrcReg[0]);
 163                 inst->I.SrcReg[0].File = PROGRAM_TEMPORARY;
 164                 inst->I.SrcReg[0].Index = inst_mov->I.DstReg.Index;
 165         }
 166
 167         return GL_TRUE;
 168 }
 169
 170 GLboolean r500FPIsNativeSwizzle(GLuint opcode, struct prog_src_register reg)
 171 {
 172         GLuint relevant;
 173         int i;
 174
 175         if (opcode == OPCODE_TEX ||
 176             opcode == OPCODE_TXB ||
 177             opcode == OPCODE_TXP ||
 178             opcode == OPCODE_KIL) {
 179                 if (reg.Abs)
 180                         return GL_FALSE;
 181
 182                 if (opcode == OPCODE_KIL && (reg.Swizzle != SWIZZLE_NOOP || reg.Negate != NEGATE_NONE))
 183                         return GL_FALSE;
 184
 185                 if (reg.Negate)
 186                         reg.Negate ^= NEGATE_XYZW;
 187
 188                 for(i = 0; i < 4; ++i) {
 189                         GLuint swz = GET_SWZ(reg.Swizzle, i);
 190                         if (swz == SWIZZLE_NIL) {
 191                                 reg.Negate &= ~(1 << i);
 192                                 continue;
 193                         }
 194                         if (swz >= 4)
 195                                 return GL_FALSE;
 196                 }
 197
 198                 if (reg.Negate)
 199                         return GL_FALSE;
 200
 201                 return GL_TRUE;
 202         } else if (opcode == OPCODE_DDX || opcode == OPCODE_DDY) {
 203                 /* DDX/MDH and DDY/MDV explicitly ignore incoming swizzles;
 204                  * if it doesn't fit perfectly into a .xyzw case... */
 205                 if (reg.Swizzle == SWIZZLE_NOOP && !reg.Abs && !reg.Negate)
 206                         return GL_TRUE;
 207
 208                 return GL_FALSE;
 209         } else {
 210                 /* ALU instructions support almost everything */
 211                 if (reg.Abs)
 212                         return GL_TRUE;
 213
 214                 relevant = 0;
 215                 for(i = 0; i < 3; ++i) {
 216                         GLuint swz = GET_SWZ(reg.Swizzle, i);
 217                         if (swz != SWIZZLE_NIL && swz != SWIZZLE_ZERO)
 218                                 relevant |= 1 << i;
 219                 }
 220                 if ((reg.Negate & relevant) && ((reg.Negate & relevant) != relevant))
 221                         return GL_FALSE;
 222
 223                 return GL_TRUE;
 224         }
 225 }
 226
 227 /**
 228  * Implement a MOV with a potentially non-native swizzle.
 229  *
 230  * The only thing we *cannot* do in an ALU instruction is per-component
 231  * negation. Therefore, we split the MOV into two instructions when necessary.
 232  */
 233 void r500FPBuildSwizzle(struct nqssadce_state *s, struct prog_dst_register dst, struct prog_src_register src)
 234 {
 235         GLuint negatebase[2] = { 0, 0 };
 236         int i;
 237
 238         for(i = 0; i < 4; ++i) {
 239                 GLuint swz = GET_SWZ(src.Swizzle, i);
 240                 if (swz == SWIZZLE_NIL)
 241                         continue;
 242                 negatebase[GET_BIT(src.Negate, i)] |= 1 << i;
 243         }
 244
 245         for(i = 0; i <= 1; ++i) {
 246                 if (!negatebase[i])
 247                         continue;
 248
 249                 struct rc_instruction *inst = rc_insert_new_instruction(s->Compiler, s->IP->Prev);
 250                 inst->I.Opcode = OPCODE_MOV;
 251                 inst->I.DstReg = dst;
 252                 inst->I.DstReg.WriteMask = negatebase[i];
 253                 inst->I.SrcReg[0] = src;
 254                 inst->I.SrcReg[0].Negate = (i == 0) ? NEGATE_NONE : NEGATE_XYZW;
 255         }
 256 }
 257
 258
 259 static char *toswiz(int swiz_val) {
 260   switch(swiz_val) {
 261   case 0: return "R";
 262   case 1: return "G";
 263   case 2: return "B";
 264   case 3: return "A";
 265   case 4: return "0";
 266   case 5: return "1/2";
 267   case 6: return "1";
 268   case 7: return "U";
 269   }
 270   return NULL;
 271 }
 272
 273 static char *toop(int op_val)
 274 {
 275   char *str = NULL;
 276   switch (op_val) {
 277   case 0: str = "MAD"; break;
 278   case 1: str = "DP3"; break;
 279   case 2: str = "DP4"; break;
 280   case 3: str = "D2A"; break;
 281   case 4: str = "MIN"; break;
 282   case 5: str = "MAX"; break;
 283   case 6: str = "Reserved"; break;
 284   case 7: str = "CND"; break;
 285   case 8: str = "CMP"; break;
 286   case 9: str = "FRC"; break;
 287   case 10: str = "SOP"; break;
 288   case 11: str = "MDH"; break;
 289   case 12: str = "MDV"; break;
 290   }
 291   return str;
 292 }
 293
 294 static char *to_alpha_op(int op_val)
 295 {
 296   char *str = NULL;
 297   switch (op_val) {
 298   case 0: str = "MAD"; break;
 299   case 1: str = "DP"; break;
 300   case 2: str = "MIN"; break;
 301   case 3: str = "MAX"; break;
 302   case 4: str = "Reserved"; break;
 303   case 5: str = "CND"; break;
 304   case 6: str = "CMP"; break;
 305   case 7: str = "FRC"; break;
 306   case 8: str = "EX2"; break;
 307   case 9: str = "LN2"; break;
 308   case 10: str = "RCP"; break;
 309   case 11: str = "RSQ"; break;
 310   case 12: str = "SIN"; break;
 311   case 13: str = "COS"; break;
 312   case 14: str = "MDH"; break;
 313   case 15: str = "MDV"; break;
 314   }
 315   return str;
 316 }
 317
 318 static char *to_mask(int val)
 319 {
 320   char *str = NULL;
 321   switch(val) {
 322   case 0: str = "NONE"; break;
 323   case 1: str = "R"; break;
 324   case 2: str = "G"; break;
 325   case 3: str = "RG"; break;
 326   case 4: str = "B"; break;
 327   case 5: str = "RB"; break;
 328   case 6: str = "GB"; break;
 329   case 7: str = "RGB"; break;
 330   case 8: str = "A"; break;
 331   case 9: str = "AR"; break;
 332   case 10: str = "AG"; break;
 333   case 11: str = "ARG"; break;
 334   case 12: str = "AB"; break;
 335   case 13: str = "ARB"; break;
 336   case 14: str = "AGB"; break;
 337   case 15: str = "ARGB"; break;
 338   }
 339   return str;
 340 }
 341
 342 static char *to_texop(int val)
 343 {
 344   switch(val) {
 345   case 0: return "NOP";
 346   case 1: return "LD";
 347   case 2: return "TEXKILL";
 348   case 3: return "PROJ";
 349   case 4: return "LODBIAS";
 350   case 5: return "LOD";
 351   case 6: return "DXDY";
 352   }
 353   return NULL;
 354 }
 355
 356 void r500FragmentProgramDump(struct rX00_fragment_program_code *c)
 357 {
 358   struct r500_fragment_program_code *code = &c->code.r500;
 359   fprintf(stderr, "R500 Fragment Program:\n--------\n");
 360
 361   int n;
 362   uint32_t inst;
 363   uint32_t inst0;
 364   char *str = NULL;
 365
 366   for (n = 0; n < code->inst_end+1; n++) {
 367     inst0 = inst = code->inst[n].inst0;
 368     fprintf(stderr,"%d\t0:CMN_INST   0x%08x:", n, inst);
 369     switch(inst & 0x3) {
 370     case R500_INST_TYPE_ALU: str = "ALU"; break;
 371     case R500_INST_TYPE_OUT: str = "OUT"; break;
 372     case R500_INST_TYPE_FC: str = "FC"; break;
 373     case R500_INST_TYPE_TEX: str = "TEX"; break;
 374     };
 375     fprintf(stderr,"%s %s %s %s %s ", str,
 376             inst & R500_INST_TEX_SEM_WAIT ? "TEX_WAIT" : "",
 377             inst & R500_INST_LAST ? "LAST" : "",
 378             inst & R500_INST_NOP ? "NOP" : "",
 379             inst & R500_INST_ALU_WAIT ? "ALU WAIT" : "");
 380     fprintf(stderr,"wmask: %s omask: %s\n", to_mask((inst >> 11) & 0xf),
 381             to_mask((inst >> 15) & 0xf));
 382
 383     switch(inst0 & 0x3) {
 384     case 0:
 385     case 1:
 386       fprintf(stderr,"\t1:RGB_ADDR   0x%08x:", code->inst[n].inst1);
 387       inst = code->inst[n].inst1;
 388
 389       fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
 390               inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
 391               (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
 392               (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
 393               (inst >> 30));
 394
 395       fprintf(stderr,"\t2:ALPHA_ADDR 0x%08x:", code->inst[n].inst2);
 396       inst = code->inst[n].inst2;
 397       fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
 398               inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
 399               (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
 400               (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
 401               (inst >> 30));
 402       fprintf(stderr,"\t3 RGB_INST:  0x%08x:", code->inst[n].inst3);
 403       inst = code->inst[n].inst3;
 404       fprintf(stderr,"rgb_A_src:%d %s/%s/%s %d rgb_B_src:%d %s/%s/%s %d\n",
 405               (inst) & 0x3, toswiz((inst >> 2) & 0x7), toswiz((inst >> 5) & 0x7), toswiz((inst >> 8) & 0x7),
 406               (inst >> 11) & 0x3,
 407               (inst >> 13) & 0x3, toswiz((inst >> 15) & 0x7), toswiz((inst >> 18) & 0x7), toswiz((inst >> 21) & 0x7),
 408               (inst >> 24) & 0x3);
 409
 410
 411       fprintf(stderr,"\t4 ALPHA_INST:0x%08x:", code->inst[n].inst4);
 412       inst = code->inst[n].inst4;
 413       fprintf(stderr,"%s dest:%d%s alp_A_src:%d %s %d alp_B_src:%d %s %d w:%d\n", to_alpha_op(inst & 0xf),
 414               (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
 415               (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), (inst >> 17) & 0x3,
 416               (inst >> 19) & 0x3, toswiz((inst >> 21) & 0x7), (inst >> 24) & 0x3,
 417               (inst >> 31) & 0x1);
 418
 419       fprintf(stderr,"\t5 RGBA_INST: 0x%08x:", code->inst[n].inst5);
 420       inst = code->inst[n].inst5;
 421       fprintf(stderr,"%s dest:%d%s rgb_C_src:%d %s/%s/%s %d alp_C_src:%d %s %d\n", toop(inst & 0xf),
 422               (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
 423               (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), toswiz((inst >> 17) & 0x7), toswiz((inst >> 20) & 0x7),
 424               (inst >> 23) & 0x3,
 425               (inst >> 25) & 0x3, toswiz((inst >> 27) & 0x7), (inst >> 30) & 0x3);
 426       break;
 427     case 2:
 428       break;
 429     case 3:
 430       inst = code->inst[n].inst1;
 431       fprintf(stderr,"\t1:TEX_INST:  0x%08x: id: %d op:%s, %s, %s %s\n", inst, (inst >> 16) & 0xf,
 432               to_texop((inst >> 22) & 0x7), (inst & (1<<25)) ? "ACQ" : "",
 433               (inst & (1<<26)) ? "IGNUNC" : "", (inst & (1<<27)) ? "UNSCALED" : "SCALED");
 434       inst = code->inst[n].inst2;
 435       fprintf(stderr,"\t2:TEX_ADDR:  0x%08x: src: %d%s %s/%s/%s/%s dst: %d%s %s/%s/%s/%s\n", inst,
 436               inst & 127, inst & (1<<7) ? "(rel)" : "",
 437               toswiz((inst >> 8) & 0x3), toswiz((inst >> 10) & 0x3),
 438               toswiz((inst >> 12) & 0x3), toswiz((inst >> 14) & 0x3),
 439               (inst >> 16) & 127, inst & (1<<23) ? "(rel)" : "",
 440               toswiz((inst >> 24) & 0x3), toswiz((inst >> 26) & 0x3),
 441               toswiz((inst >> 28) & 0x3), toswiz((inst >> 30) & 0x3));
 442
 443       fprintf(stderr,"\t3:TEX_DXDY:  0x%08x\n", code->inst[n].inst3);
 444       break;
 445     }
 446     fprintf(stderr,"\n");
 447   }
 448
 449 }