src/gallium/auxiliary/gallivm/lp_bld_tgsi_info.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2010 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * 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, sub license, 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 SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  18  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  19  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  20  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * The above copyright notice and this permission notice (including the
  23  * next paragraph) shall be included in all copies or substantial portions
  24  * of the Software.
  25  *
  26  **************************************************************************/
  27
  28
  29 #include "util/u_memory.h"
  30 #include "util/u_math.h"
  31 #include "tgsi/tgsi_parse.h"
  32 #include "tgsi/tgsi_util.h"
  33 #include "tgsi/tgsi_dump.h"
  34 #include "tgsi/tgsi_strings.h"
  35 #include "lp_bld_debug.h"
  36 #include "lp_bld_tgsi.h"
  37
  38
  39 /**
  40  * Analysis context.
  41  *
  42  * This is where we keep store the value of each channel of the IMM/TEMP/OUT
  43  * register values, as we walk the shader.
  44  */
  45 struct analysis_context
  46 {
  47    struct lp_tgsi_info *info;
  48
  49    unsigned num_imms;
  50    float imm[LP_MAX_TGSI_IMMEDIATES][4];
  51
  52    struct lp_tgsi_channel_info temp[32][4];
  53 };
  54
  55
  56 /**
  57  * Describe the specified channel of the src register.
  58  */
  59 static void
  60 analyse_src(struct analysis_context *ctx,
  61             struct lp_tgsi_channel_info *chan_info,
  62             const struct tgsi_src_register *src,
  63             unsigned chan)
  64 {
  65    chan_info->file = TGSI_FILE_NULL;
  66    if (!src->Indirect && !src->Absolute && !src->Negate) {
  67       unsigned swizzle = tgsi_util_get_src_register_swizzle(src, chan);
  68       if (src->File == TGSI_FILE_TEMPORARY) {
  69          if (src->Index < Elements(ctx->temp)) {
  70             *chan_info = ctx->temp[src->Index][swizzle];
  71          }
  72       } else {
  73          chan_info->file = src->File;
  74          if (src->File == TGSI_FILE_IMMEDIATE) {
  75             assert(src->Index < Elements(ctx->imm));
  76             if (src->Index < Elements(ctx->imm)) {
  77                chan_info->u.value = ctx->imm[src->Index][swizzle];
  78             }
  79          } else {
  80             chan_info->u.index = src->Index;
  81             chan_info->swizzle = swizzle;
  82          }
  83       }
  84    }
  85 }
  86
  87
  88 /**
  89  * Whether this register channel refers to a specific immediate value.
  90  */
  91 static boolean
  92 is_immediate(const struct lp_tgsi_channel_info *chan_info, float value)
  93 {
  94    return chan_info->file == TGSI_FILE_IMMEDIATE &&
  95           chan_info->u.value == value;
  96 }
  97
  98
  99 /**
 100  * Analyse properties of tex instructions, in particular used
 101  * to figure out if a texture is considered indirect.
 102  * Not actually used by much except the tgsi dumping code.
 103  */
 104 static void
 105 analyse_tex(struct analysis_context *ctx,
 106             const struct tgsi_full_instruction *inst,
 107             enum lp_build_tex_modifier modifier)
 108 {
 109    struct lp_tgsi_info *info = ctx->info;
 110    unsigned chan;
 111
 112    if (info->num_texs < Elements(info->tex)) {
 113       struct lp_tgsi_texture_info *tex_info = &info->tex[info->num_texs];
 114       boolean indirect = FALSE;
 115       unsigned readmask = 0;
 116
 117       tex_info->target = inst->Texture.Texture;
 118       switch (inst->Texture.Texture) {
 119       case TGSI_TEXTURE_1D:
 120          readmask = TGSI_WRITEMASK_X;
 121          break;
 122       case TGSI_TEXTURE_1D_ARRAY:
 123       case TGSI_TEXTURE_2D:
 124       case TGSI_TEXTURE_RECT:
 125          readmask = TGSI_WRITEMASK_XY;
 126          break;
 127       case TGSI_TEXTURE_SHADOW1D:
 128       case TGSI_TEXTURE_SHADOW1D_ARRAY:
 129       case TGSI_TEXTURE_SHADOW2D:
 130       case TGSI_TEXTURE_SHADOWRECT:
 131       case TGSI_TEXTURE_2D_ARRAY:
 132       case TGSI_TEXTURE_3D:
 133       case TGSI_TEXTURE_CUBE:
 134          readmask = TGSI_WRITEMASK_XYZ;
 135          break;
 136       case TGSI_TEXTURE_SHADOW2D_ARRAY:
 137       case TGSI_TEXTURE_SHADOWCUBE:
 138          readmask = TGSI_WRITEMASK_XYZW;
 139          break;
 140       default:
 141          assert(0);
 142          return;
 143       }
 144
 145       if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 146          /* We don't track explicit derivatives, although we could */
 147          indirect = TRUE;
 148          tex_info->sampler_unit = inst->Src[3].Register.Index;
 149          tex_info->texture_unit = inst->Src[3].Register.Index;
 150       }  else {
 151          if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED ||
 152              modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
 153              modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
 154             readmask |= TGSI_WRITEMASK_W;
 155          }
 156          tex_info->sampler_unit = inst->Src[1].Register.Index;
 157          tex_info->texture_unit = inst->Src[1].Register.Index;
 158       }
 159
 160       for (chan = 0; chan < 4; ++chan) {
 161          struct lp_tgsi_channel_info *chan_info = &tex_info->coord[chan];
 162          if (readmask & (1 << chan)) {
 163             analyse_src(ctx, chan_info, &inst->Src[0].Register, chan);
 164             if (chan_info->file != TGSI_FILE_INPUT) {
 165                indirect = TRUE;
 166             }
 167          } else {
 168             memset(chan_info, 0, sizeof *chan_info);
 169          }
 170       }
 171
 172       if (indirect) {
 173          info->indirect_textures = TRUE;
 174       }
 175
 176       ++info->num_texs;
 177    } else {
 178       info->indirect_textures = TRUE;
 179    }
 180 }
 181
 182
 183 /**
 184  * Analyse properties of sample instructions, in particular used
 185  * to figure out if a texture is considered indirect.
 186  * Not actually used by much except the tgsi dumping code.
 187  */
 188 static void
 189 analyse_sample(struct analysis_context *ctx,
 190                const struct tgsi_full_instruction *inst,
 191                enum lp_build_tex_modifier modifier,
 192                boolean shadow)
 193 {
 194    struct lp_tgsi_info *info = ctx->info;
 195    unsigned chan;
 196
 197    if (info->num_texs < Elements(info->tex)) {
 198       struct lp_tgsi_texture_info *tex_info = &info->tex[info->num_texs];
 199       boolean indirect = FALSE;
 200       boolean shadow = FALSE;
 201       unsigned readmask;
 202
 203       /*
 204        * We don't really get much information here, in particular not
 205        * the target info, hence no useful writemask neither. Maybe should just
 206        * forget the whole function.
 207        */
 208       readmask = TGSI_WRITEMASK_XYZW;
 209
 210       tex_info->texture_unit = inst->Src[1].Register.Index;
 211       tex_info->sampler_unit = inst->Src[2].Register.Index;
 212
 213       if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV ||
 214           modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD ||
 215           modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS || shadow) {
 216          /* We don't track insts with additional regs, although we could */
 217          indirect = TRUE;
 218       }
 219
 220       for (chan = 0; chan < 4; ++chan) {
 221          struct lp_tgsi_channel_info *chan_info = &tex_info->coord[chan];
 222          if (readmask & (1 << chan)) {
 223             analyse_src(ctx, chan_info, &inst->Src[0].Register, chan);
 224             if (chan_info->file != TGSI_FILE_INPUT) {
 225                indirect = TRUE;
 226             }
 227          } else {
 228             memset(chan_info, 0, sizeof *chan_info);
 229          }
 230       }
 231
 232       if (indirect) {
 233          info->indirect_textures = TRUE;
 234       }
 235
 236       ++info->num_texs;
 237    } else {
 238       info->indirect_textures = TRUE;
 239    }
 240 }
 241
 242
 243 /**
 244  * Process an instruction, and update the register values accordingly.
 245  */
 246 static void
 247 analyse_instruction(struct analysis_context *ctx,
 248                     struct tgsi_full_instruction *inst)
 249 {
 250    struct lp_tgsi_info *info = ctx->info;
 251    struct lp_tgsi_channel_info (*regs)[4];
 252    unsigned max_regs;
 253    unsigned i;
 254    unsigned index;
 255    unsigned chan;
 256
 257    for (i = 0; i < inst->Instruction.NumDstRegs; ++i) {
 258       const struct tgsi_dst_register *dst = &inst->Dst[i].Register;
 259
 260       /*
 261        * Get the lp_tgsi_channel_info array corresponding to the destination
 262        * register file.
 263        */
 264
 265       if (dst->File == TGSI_FILE_TEMPORARY) {
 266          regs = ctx->temp;
 267          max_regs = Elements(ctx->temp);
 268       } else if (dst->File == TGSI_FILE_OUTPUT) {
 269          regs = info->output;
 270          max_regs = Elements(info->output);
 271       } else if (dst->File == TGSI_FILE_ADDRESS ||
 272                  dst->File == TGSI_FILE_PREDICATE) {
 273          continue;
 274       } else {
 275          assert(0);
 276          continue;
 277       }
 278
 279       /*
 280        * Detect direct TEX instructions
 281        */
 282
 283       switch (inst->Instruction.Opcode) {
 284       case TGSI_OPCODE_TEX:
 285          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_NONE);
 286          break;
 287       case TGSI_OPCODE_TXD:
 288          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 289          break;
 290       case TGSI_OPCODE_TXB:
 291          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 292          break;
 293       case TGSI_OPCODE_TXL:
 294          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 295          break;
 296       case TGSI_OPCODE_TXP:
 297          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 298          break;
 299       case TGSI_OPCODE_SAMPLE:
 300          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_NONE, FALSE);
 301          break;
 302       case TGSI_OPCODE_SAMPLE_C:
 303          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_NONE, TRUE);
 304          break;
 305       case TGSI_OPCODE_SAMPLE_C_LZ:
 306          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_LOD_ZERO, TRUE);
 307          break;
 308       case TGSI_OPCODE_SAMPLE_D:
 309          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV, FALSE);
 310          break;
 311       case TGSI_OPCODE_SAMPLE_B:
 312          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS, FALSE);
 313          break;
 314       case TGSI_OPCODE_SAMPLE_L:
 315          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD, FALSE);
 316          break;
 317       default:
 318          break;
 319       }
 320
 321       /*
 322        * Keep track of assignments and writes
 323        */
 324
 325       if (dst->Indirect) {
 326          /*
 327           * It could be any register index so clear all register indices.
 328           */
 329
 330          for (chan = 0; chan < 4; ++chan) {
 331             if (dst->WriteMask & (1 << chan)) {
 332                for (index = 0; index < max_regs; ++index) {
 333                   regs[index][chan].file = TGSI_FILE_NULL;
 334                }
 335             }
 336          }
 337       } else if (dst->Index < max_regs) {
 338          /*
 339           * Update this destination register value.
 340           */
 341
 342          struct lp_tgsi_channel_info res[4];
 343
 344          memset(res, 0, sizeof res);
 345
 346          if (!inst->Instruction.Predicate &&
 347              !inst->Instruction.Saturate) {
 348             for (chan = 0; chan < 4; ++chan) {
 349                if (dst->WriteMask & (1 << chan)) {
 350                   if (inst->Instruction.Opcode == TGSI_OPCODE_MOV) {
 351                      analyse_src(ctx, &res[chan],
 352                                  &inst->Src[0].Register, chan);
 353                   } else if (inst->Instruction.Opcode == TGSI_OPCODE_MUL) {
 354                      /*
 355                       * Propagate values across 1.0 and 0.0 multiplications.
 356                       */
 357
 358                      struct lp_tgsi_channel_info src0;
 359                      struct lp_tgsi_channel_info src1;
 360
 361                      analyse_src(ctx, &src0, &inst->Src[0].Register, chan);
 362                      analyse_src(ctx, &src1, &inst->Src[1].Register, chan);
 363
 364                      if (is_immediate(&src0, 0.0f)) {
 365                         res[chan] = src0;
 366                      } else if (is_immediate(&src1, 0.0f)) {
 367                         res[chan] = src1;
 368                      } else if (is_immediate(&src0, 1.0f)) {
 369                         res[chan] = src1;
 370                      } else if (is_immediate(&src1, 1.0f)) {
 371                         res[chan] = src0;
 372                      }
 373                   }
 374                }
 375             }
 376          }
 377
 378          for (chan = 0; chan < 4; ++chan) {
 379             if (dst->WriteMask & (1 << chan)) {
 380                regs[dst->Index][chan] = res[chan];
 381             }
 382          }
 383       }
 384    }
 385
 386    /*
 387     * Clear all temporaries information in presence of a control flow opcode.
 388     */
 389
 390    switch (inst->Instruction.Opcode) {
 391    case TGSI_OPCODE_IF:
 392    case TGSI_OPCODE_UIF:
 393    case TGSI_OPCODE_ELSE:
 394    case TGSI_OPCODE_ENDIF:
 395    case TGSI_OPCODE_BGNLOOP:
 396    case TGSI_OPCODE_BRK:
 397    case TGSI_OPCODE_BREAKC:
 398    case TGSI_OPCODE_CONT:
 399    case TGSI_OPCODE_ENDLOOP:
 400    case TGSI_OPCODE_CALLNZ:
 401    case TGSI_OPCODE_CAL:
 402    case TGSI_OPCODE_BGNSUB:
 403    case TGSI_OPCODE_ENDSUB:
 404    case TGSI_OPCODE_SWITCH:
 405    case TGSI_OPCODE_CASE:
 406    case TGSI_OPCODE_DEFAULT:
 407    case TGSI_OPCODE_ENDSWITCH:
 408    case TGSI_OPCODE_RET:
 409    case TGSI_OPCODE_END:
 410       /* XXX: Are there more cases? */
 411       memset(&ctx->temp, 0, sizeof ctx->temp);
 412       memset(&info->output, 0, sizeof info->output);
 413    default:
 414       break;
 415    }
 416 }
 417
 418
 419 static INLINE void
 420 dump_info(const struct tgsi_token *tokens,
 421           struct lp_tgsi_info *info)
 422 {
 423    unsigned index;
 424    unsigned chan;
 425
 426    tgsi_dump(tokens, 0);
 427
 428    for (index = 0; index < info->num_texs; ++index) {
 429       const struct lp_tgsi_texture_info *tex_info = &info->tex[index];
 430       debug_printf("TEX[%u] =", index);
 431       for (chan = 0; chan < 4; ++chan) {
 432          const struct lp_tgsi_channel_info *chan_info =
 433                &tex_info->coord[chan];
 434          if (chan_info->file != TGSI_FILE_NULL) {
 435             debug_printf(" %s[%u].%c",
 436                          tgsi_file_name(chan_info->file),
 437                          chan_info->u.index,
 438                          "xyzw01"[chan_info->swizzle]);
 439          } else {
 440             debug_printf(" _");
 441          }
 442       }
 443       debug_printf(", RES[%u], SAMP[%u], %s\n",
 444                    tex_info->texture_unit,
 445                    tex_info->sampler_unit,
 446                    tgsi_texture_names[tex_info->target]);
 447    }
 448
 449    for (index = 0; index < PIPE_MAX_SHADER_OUTPUTS; ++index) {
 450       for (chan = 0; chan < 4; ++chan) {
 451          const struct lp_tgsi_channel_info *chan_info =
 452                &info->output[index][chan];
 453          if (chan_info->file != TGSI_FILE_NULL) {
 454             debug_printf("OUT[%u].%c = ", index, "xyzw"[chan]);
 455             if (chan_info->file == TGSI_FILE_IMMEDIATE) {
 456                debug_printf("%f", chan_info->u.value);
 457             } else {
 458                const char *file_name;
 459                switch (chan_info->file) {
 460                case TGSI_FILE_CONSTANT:
 461                   file_name = "CONST";
 462                   break;
 463                case TGSI_FILE_INPUT:
 464                   file_name = "IN";
 465                   break;
 466                default:
 467                   file_name = "???";
 468                   break;
 469                }
 470                debug_printf("%s[%u].%c",
 471                             file_name,
 472                             chan_info->u.index,
 473                             "xyzw01"[chan_info->swizzle]);
 474             }
 475             debug_printf("\n");
 476          }
 477       }
 478    }
 479 }
 480
 481
 482 /**
 483  * Detect any direct relationship between the output color
 484  */
 485 void
 486 lp_build_tgsi_info(const struct tgsi_token *tokens,
 487                    struct lp_tgsi_info *info)
 488 {
 489    struct tgsi_parse_context parse;
 490    struct analysis_context *ctx;
 491    unsigned index;
 492    unsigned chan;
 493
 494    memset(info, 0, sizeof *info);
 495
 496    tgsi_scan_shader(tokens, &info->base);
 497
 498    ctx = CALLOC(1, sizeof(struct analysis_context));
 499    ctx->info = info;
 500
 501    tgsi_parse_init(&parse, tokens);
 502
 503    while (!tgsi_parse_end_of_tokens(&parse)) {
 504       tgsi_parse_token(&parse);
 505
 506       switch (parse.FullToken.Token.Type) {
 507       case TGSI_TOKEN_TYPE_DECLARATION:
 508          break;
 509
 510       case TGSI_TOKEN_TYPE_INSTRUCTION:
 511          {
 512             struct tgsi_full_instruction *inst =
 513                   &parse.FullToken.FullInstruction;
 514
 515             if (inst->Instruction.Opcode == TGSI_OPCODE_END ||
 516                 inst->Instruction.Opcode == TGSI_OPCODE_BGNSUB) {
 517                /* We reached the end of main function body. */
 518                goto finished;
 519             }
 520
 521             analyse_instruction(ctx, inst);
 522          }
 523          break;
 524
 525       case TGSI_TOKEN_TYPE_IMMEDIATE:
 526          {
 527             const unsigned size =
 528                   parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
 529             assert(size <= 4);
 530             if (ctx->num_imms < Elements(ctx->imm)) {
 531                for (chan = 0; chan < size; ++chan) {
 532                   float value = parse.FullToken.FullImmediate.u[chan].Float;
 533                   ctx->imm[ctx->num_imms][chan] = value;
 534
 535                   if (value < 0.0f || value > 1.0f) {
 536                      info->unclamped_immediates = TRUE;
 537                   }
 538                }
 539                ++ctx->num_imms;
 540             }
 541          }
 542          break;
 543
 544       case TGSI_TOKEN_TYPE_PROPERTY:
 545          break;
 546
 547       default:
 548          assert(0);
 549       }
 550    }
 551 finished:
 552
 553    tgsi_parse_free(&parse);
 554    FREE(ctx);
 555
 556
 557    /*
 558     * Link the output color values.
 559     */
 560
 561    for (index = 0; index < PIPE_MAX_COLOR_BUFS; ++index) {
 562       const struct lp_tgsi_channel_info null_output[4];
 563       info->cbuf[index] = null_output;
 564    }
 565
 566    for (index = 0; index < info->base.num_outputs; ++index) {
 567       unsigned semantic_name = info->base.output_semantic_name[index];
 568       unsigned semantic_index = info->base.output_semantic_index[index];
 569       if (semantic_name == TGSI_SEMANTIC_COLOR &&
 570           semantic_index < PIPE_MAX_COLOR_BUFS) {
 571          info->cbuf[semantic_index] = info->output[index];
 572       }
 573    }
 574
 575    if (gallivm_debug & GALLIVM_DEBUG_TGSI) {
 576       dump_info(tokens, info);
 577    }
 578 }