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       case TGSI_TEXTURE_CUBE_ARRAY:
 141          readmask = TGSI_WRITEMASK_XYZW;
 142          break;
 143       case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
 144          readmask = TGSI_WRITEMASK_XYZW;
 145          break;
 146       default:
 147          assert(0);
 148          return;
 149       }
 150       /* XXX
 151        * For cube map arrays, this will not analyze lod or shadow argument.
 152        * For shadow cube, this will not analyze lod bias argument.
 153        * "Indirect" really has no meaning for such textures anyway though.
 154        */
 155
 156       if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 157          /* We don't track explicit derivatives, although we could */
 158          indirect = TRUE;
 159          tex_info->sampler_unit = inst->Src[3].Register.Index;
 160          tex_info->texture_unit = inst->Src[3].Register.Index;
 161       }  else {
 162          if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED ||
 163              modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
 164              modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
 165             readmask |= TGSI_WRITEMASK_W;
 166          }
 167          tex_info->sampler_unit = inst->Src[1].Register.Index;
 168          tex_info->texture_unit = inst->Src[1].Register.Index;
 169       }
 170
 171       for (chan = 0; chan < 4; ++chan) {
 172          struct lp_tgsi_channel_info *chan_info = &tex_info->coord[chan];
 173          if (readmask & (1 << chan)) {
 174             analyse_src(ctx, chan_info, &inst->Src[0].Register, chan);
 175             if (chan_info->file != TGSI_FILE_INPUT) {
 176                indirect = TRUE;
 177             }
 178          } else {
 179             memset(chan_info, 0, sizeof *chan_info);
 180          }
 181       }
 182
 183       if (indirect) {
 184          info->indirect_textures = TRUE;
 185       }
 186
 187       ++info->num_texs;
 188    } else {
 189       info->indirect_textures = TRUE;
 190    }
 191 }
 192
 193
 194 /**
 195  * Analyse properties of sample instructions, in particular used
 196  * to figure out if a texture is considered indirect.
 197  * Not actually used by much except the tgsi dumping code.
 198  */
 199 static void
 200 analyse_sample(struct analysis_context *ctx,
 201                const struct tgsi_full_instruction *inst,
 202                enum lp_build_tex_modifier modifier,
 203                boolean shadow)
 204 {
 205    struct lp_tgsi_info *info = ctx->info;
 206    unsigned chan;
 207
 208    if (info->num_texs < Elements(info->tex)) {
 209       struct lp_tgsi_texture_info *tex_info = &info->tex[info->num_texs];
 210       boolean indirect = FALSE;
 211       boolean shadow = FALSE;
 212       unsigned readmask;
 213
 214       /*
 215        * We don't really get much information here, in particular not
 216        * the target info, hence no useful writemask neither. Maybe should just
 217        * forget the whole function.
 218        */
 219       readmask = TGSI_WRITEMASK_XYZW;
 220
 221       tex_info->texture_unit = inst->Src[1].Register.Index;
 222       tex_info->sampler_unit = inst->Src[2].Register.Index;
 223
 224       if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV ||
 225           modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD ||
 226           modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS || shadow) {
 227          /* We don't track insts with additional regs, although we could */
 228          indirect = TRUE;
 229       }
 230
 231       for (chan = 0; chan < 4; ++chan) {
 232          struct lp_tgsi_channel_info *chan_info = &tex_info->coord[chan];
 233          if (readmask & (1 << chan)) {
 234             analyse_src(ctx, chan_info, &inst->Src[0].Register, chan);
 235             if (chan_info->file != TGSI_FILE_INPUT) {
 236                indirect = TRUE;
 237             }
 238          } else {
 239             memset(chan_info, 0, sizeof *chan_info);
 240          }
 241       }
 242
 243       if (indirect) {
 244          info->indirect_textures = TRUE;
 245       }
 246
 247       ++info->num_texs;
 248    } else {
 249       info->indirect_textures = TRUE;
 250    }
 251 }
 252
 253
 254 /**
 255  * Process an instruction, and update the register values accordingly.
 256  */
 257 static void
 258 analyse_instruction(struct analysis_context *ctx,
 259                     struct tgsi_full_instruction *inst)
 260 {
 261    struct lp_tgsi_info *info = ctx->info;
 262    struct lp_tgsi_channel_info (*regs)[4];
 263    unsigned max_regs;
 264    unsigned i;
 265    unsigned index;
 266    unsigned chan;
 267
 268    for (i = 0; i < inst->Instruction.NumDstRegs; ++i) {
 269       const struct tgsi_dst_register *dst = &inst->Dst[i].Register;
 270
 271       /*
 272        * Get the lp_tgsi_channel_info array corresponding to the destination
 273        * register file.
 274        */
 275
 276       if (dst->File == TGSI_FILE_TEMPORARY) {
 277          regs = ctx->temp;
 278          max_regs = Elements(ctx->temp);
 279       } else if (dst->File == TGSI_FILE_OUTPUT) {
 280          regs = info->output;
 281          max_regs = Elements(info->output);
 282       } else if (dst->File == TGSI_FILE_ADDRESS ||
 283                  dst->File == TGSI_FILE_PREDICATE) {
 284          continue;
 285       } else {
 286          assert(0);
 287          continue;
 288       }
 289
 290       /*
 291        * Detect direct TEX instructions
 292        */
 293
 294       switch (inst->Instruction.Opcode) {
 295       case TGSI_OPCODE_TEX:
 296          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_NONE);
 297          break;
 298       case TGSI_OPCODE_TXD:
 299          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 300          break;
 301       case TGSI_OPCODE_TXB:
 302          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 303          break;
 304       case TGSI_OPCODE_TXL:
 305          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 306          break;
 307       case TGSI_OPCODE_TXP:
 308          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 309          break;
 310       case TGSI_OPCODE_TEX2:
 311          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_NONE);
 312          break;
 313       case TGSI_OPCODE_TXB2:
 314          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 315          break;
 316       case TGSI_OPCODE_TXL2:
 317          analyse_tex(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 318          break;
 319       case TGSI_OPCODE_SAMPLE:
 320          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_NONE, FALSE);
 321          break;
 322       case TGSI_OPCODE_SAMPLE_C:
 323          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_NONE, TRUE);
 324          break;
 325       case TGSI_OPCODE_SAMPLE_C_LZ:
 326          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_LOD_ZERO, TRUE);
 327          break;
 328       case TGSI_OPCODE_SAMPLE_D:
 329          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV, FALSE);
 330          break;
 331       case TGSI_OPCODE_SAMPLE_B:
 332          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS, FALSE);
 333          break;
 334       case TGSI_OPCODE_SAMPLE_L:
 335          analyse_sample(ctx, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD, FALSE);
 336          break;
 337       default:
 338          break;
 339       }
 340
 341       /*
 342        * Keep track of assignments and writes
 343        */
 344
 345       if (dst->Indirect) {
 346          /*
 347           * It could be any register index so clear all register indices.
 348           */
 349
 350          for (chan = 0; chan < 4; ++chan) {
 351             if (dst->WriteMask & (1 << chan)) {
 352                for (index = 0; index < max_regs; ++index) {
 353                   regs[index][chan].file = TGSI_FILE_NULL;
 354                }
 355             }
 356          }
 357       } else if (dst->Index < max_regs) {
 358          /*
 359           * Update this destination register value.
 360           */
 361
 362          struct lp_tgsi_channel_info res[4];
 363
 364          memset(res, 0, sizeof res);
 365
 366          if (!inst->Instruction.Predicate &&
 367              !inst->Instruction.Saturate) {
 368             for (chan = 0; chan < 4; ++chan) {
 369                if (dst->WriteMask & (1 << chan)) {
 370                   if (inst->Instruction.Opcode == TGSI_OPCODE_MOV) {
 371                      analyse_src(ctx, &res[chan],
 372                                  &inst->Src[0].Register, chan);
 373                   } else if (inst->Instruction.Opcode == TGSI_OPCODE_MUL) {
 374                      /*
 375                       * Propagate values across 1.0 and 0.0 multiplications.
 376                       */
 377
 378                      struct lp_tgsi_channel_info src0;
 379                      struct lp_tgsi_channel_info src1;
 380
 381                      analyse_src(ctx, &src0, &inst->Src[0].Register, chan);
 382                      analyse_src(ctx, &src1, &inst->Src[1].Register, chan);
 383
 384                      if (is_immediate(&src0, 0.0f)) {
 385                         res[chan] = src0;
 386                      } else if (is_immediate(&src1, 0.0f)) {
 387                         res[chan] = src1;
 388                      } else if (is_immediate(&src0, 1.0f)) {
 389                         res[chan] = src1;
 390                      } else if (is_immediate(&src1, 1.0f)) {
 391                         res[chan] = src0;
 392                      }
 393                   }
 394                }
 395             }
 396          }
 397
 398          for (chan = 0; chan < 4; ++chan) {
 399             if (dst->WriteMask & (1 << chan)) {
 400                regs[dst->Index][chan] = res[chan];
 401             }
 402          }
 403       }
 404    }
 405
 406    /*
 407     * Clear all temporaries information in presence of a control flow opcode.
 408     */
 409
 410    switch (inst->Instruction.Opcode) {
 411    case TGSI_OPCODE_IF:
 412    case TGSI_OPCODE_UIF:
 413    case TGSI_OPCODE_ELSE:
 414    case TGSI_OPCODE_ENDIF:
 415    case TGSI_OPCODE_BGNLOOP:
 416    case TGSI_OPCODE_BRK:
 417    case TGSI_OPCODE_BREAKC:
 418    case TGSI_OPCODE_CONT:
 419    case TGSI_OPCODE_ENDLOOP:
 420    case TGSI_OPCODE_CALLNZ:
 421    case TGSI_OPCODE_CAL:
 422    case TGSI_OPCODE_BGNSUB:
 423    case TGSI_OPCODE_ENDSUB:
 424    case TGSI_OPCODE_SWITCH:
 425    case TGSI_OPCODE_CASE:
 426    case TGSI_OPCODE_DEFAULT:
 427    case TGSI_OPCODE_ENDSWITCH:
 428    case TGSI_OPCODE_RET:
 429    case TGSI_OPCODE_END:
 430       /* XXX: Are there more cases? */
 431       memset(&ctx->temp, 0, sizeof ctx->temp);
 432       memset(&info->output, 0, sizeof info->output);
 433    default:
 434       break;
 435    }
 436 }
 437
 438
 439 static INLINE void
 440 dump_info(const struct tgsi_token *tokens,
 441           struct lp_tgsi_info *info)
 442 {
 443    unsigned index;
 444    unsigned chan;
 445
 446    tgsi_dump(tokens, 0);
 447
 448    for (index = 0; index < info->num_texs; ++index) {
 449       const struct lp_tgsi_texture_info *tex_info = &info->tex[index];
 450       debug_printf("TEX[%u] =", index);
 451       for (chan = 0; chan < 4; ++chan) {
 452          const struct lp_tgsi_channel_info *chan_info =
 453                &tex_info->coord[chan];
 454          if (chan_info->file != TGSI_FILE_NULL) {
 455             debug_printf(" %s[%u].%c",
 456                          tgsi_file_name(chan_info->file),
 457                          chan_info->u.index,
 458                          "xyzw01"[chan_info->swizzle]);
 459          } else {
 460             debug_printf(" _");
 461          }
 462       }
 463       debug_printf(", RES[%u], SAMP[%u], %s\n",
 464                    tex_info->texture_unit,
 465                    tex_info->sampler_unit,
 466                    tgsi_texture_names[tex_info->target]);
 467    }
 468
 469    for (index = 0; index < PIPE_MAX_SHADER_OUTPUTS; ++index) {
 470       for (chan = 0; chan < 4; ++chan) {
 471          const struct lp_tgsi_channel_info *chan_info =
 472                &info->output[index][chan];
 473          if (chan_info->file != TGSI_FILE_NULL) {
 474             debug_printf("OUT[%u].%c = ", index, "xyzw"[chan]);
 475             if (chan_info->file == TGSI_FILE_IMMEDIATE) {
 476                debug_printf("%f", chan_info->u.value);
 477             } else {
 478                const char *file_name;
 479                switch (chan_info->file) {
 480                case TGSI_FILE_CONSTANT:
 481                   file_name = "CONST";
 482                   break;
 483                case TGSI_FILE_INPUT:
 484                   file_name = "IN";
 485                   break;
 486                default:
 487                   file_name = "???";
 488                   break;
 489                }
 490                debug_printf("%s[%u].%c",
 491                             file_name,
 492                             chan_info->u.index,
 493                             "xyzw01"[chan_info->swizzle]);
 494             }
 495             debug_printf("\n");
 496          }
 497       }
 498    }
 499 }
 500
 501
 502 /**
 503  * Detect any direct relationship between the output color
 504  */
 505 void
 506 lp_build_tgsi_info(const struct tgsi_token *tokens,
 507                    struct lp_tgsi_info *info)
 508 {
 509    struct tgsi_parse_context parse;
 510    struct analysis_context *ctx;
 511    unsigned index;
 512    unsigned chan;
 513
 514    memset(info, 0, sizeof *info);
 515
 516    tgsi_scan_shader(tokens, &info->base);
 517
 518    ctx = CALLOC(1, sizeof(struct analysis_context));
 519    ctx->info = info;
 520
 521    tgsi_parse_init(&parse, tokens);
 522
 523    while (!tgsi_parse_end_of_tokens(&parse)) {
 524       tgsi_parse_token(&parse);
 525
 526       switch (parse.FullToken.Token.Type) {
 527       case TGSI_TOKEN_TYPE_DECLARATION:
 528          break;
 529
 530       case TGSI_TOKEN_TYPE_INSTRUCTION:
 531          {
 532             struct tgsi_full_instruction *inst =
 533                   &parse.FullToken.FullInstruction;
 534
 535             if (inst->Instruction.Opcode == TGSI_OPCODE_END ||
 536                 inst->Instruction.Opcode == TGSI_OPCODE_BGNSUB) {
 537                /* We reached the end of main function body. */
 538                goto finished;
 539             }
 540
 541             analyse_instruction(ctx, inst);
 542          }
 543          break;
 544
 545       case TGSI_TOKEN_TYPE_IMMEDIATE:
 546          {
 547             const unsigned size =
 548                   parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
 549             assert(size <= 4);
 550             if (ctx->num_imms < Elements(ctx->imm)) {
 551                for (chan = 0; chan < size; ++chan) {
 552                   float value = parse.FullToken.FullImmediate.u[chan].Float;
 553                   ctx->imm[ctx->num_imms][chan] = value;
 554
 555                   if (value < 0.0f || value > 1.0f) {
 556                      info->unclamped_immediates = TRUE;
 557                   }
 558                }
 559                ++ctx->num_imms;
 560             }
 561          }
 562          break;
 563
 564       case TGSI_TOKEN_TYPE_PROPERTY:
 565          break;
 566
 567       default:
 568          assert(0);
 569       }
 570    }
 571 finished:
 572
 573    tgsi_parse_free(&parse);
 574    FREE(ctx);
 575
 576
 577    /*
 578     * Link the output color values.
 579     */
 580
 581    for (index = 0; index < PIPE_MAX_COLOR_BUFS; ++index) {
 582       const struct lp_tgsi_channel_info null_output[4];
 583       info->cbuf[index] = null_output;
 584    }
 585
 586    for (index = 0; index < info->base.num_outputs; ++index) {
 587       unsigned semantic_name = info->base.output_semantic_name[index];
 588       unsigned semantic_index = info->base.output_semantic_index[index];
 589       if (semantic_name == TGSI_SEMANTIC_COLOR &&
 590           semantic_index < PIPE_MAX_COLOR_BUFS) {
 591          info->cbuf[semantic_index] = info->output[index];
 592       }
 593    }
 594
 595    if (gallivm_debug & GALLIVM_DEBUG_TGSI) {
 596       dump_info(tokens, info);
 597    }
 598 }