src/gallium/auxiliary/tgsi/tgsi_scan.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2008 VMware, Inc.
   4  * All Rights Reserved.
   5  * Copyright 2008 VMware, Inc.  All rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * 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, sub license, 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 portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * TGSI program scan utility.
  31  * Used to determine which registers and instructions are used by a shader.
  32  *
  33  * Authors:  Brian Paul
  34  */
  35
  36
  37 #include "util/u_debug.h"
  38 #include "util/u_math.h"
  39 #include "util/u_memory.h"
  40 #include "util/u_prim.h"
  41 #include "tgsi/tgsi_parse.h"
  42 #include "tgsi/tgsi_util.h"
  43 #include "tgsi/tgsi_scan.h"
  44
  45
  46
  47
  48 /**
  49  * Scan the given TGSI shader to collect information such as number of
  50  * registers used, special instructions used, etc.
  51  * \return info  the result of the scan
  52  */
  53 void
  54 tgsi_scan_shader(const struct tgsi_token *tokens,
  55                  struct tgsi_shader_info *info)
  56 {
  57    uint procType, i;
  58    struct tgsi_parse_context parse;
  59    unsigned current_depth = 0;
  60
  61    memset(info, 0, sizeof(*info));
  62    for (i = 0; i < TGSI_FILE_COUNT; i++)
  63       info->file_max[i] = -1;
  64    for (i = 0; i < Elements(info->const_file_max); i++)
  65       info->const_file_max[i] = -1;
  66    info->properties[TGSI_PROPERTY_GS_INVOCATIONS] = 1;
  67
  68    /**
  69     ** Setup to begin parsing input shader
  70     **/
  71    if (tgsi_parse_init( &parse, tokens ) != TGSI_PARSE_OK) {
  72       debug_printf("tgsi_parse_init() failed in tgsi_scan_shader()!\n");
  73       return;
  74    }
  75    procType = parse.FullHeader.Processor.Processor;
  76    assert(procType == TGSI_PROCESSOR_FRAGMENT ||
  77           procType == TGSI_PROCESSOR_VERTEX ||
  78           procType == TGSI_PROCESSOR_GEOMETRY ||
  79           procType == TGSI_PROCESSOR_TESS_CTRL ||
  80           procType == TGSI_PROCESSOR_TESS_EVAL ||
  81           procType == TGSI_PROCESSOR_COMPUTE);
  82    info->processor = procType;
  83
  84
  85    /**
  86     ** Loop over incoming program tokens/instructions
  87     */
  88    while( !tgsi_parse_end_of_tokens( &parse ) ) {
  89
  90       info->num_tokens++;
  91
  92       tgsi_parse_token( &parse );
  93
  94       switch( parse.FullToken.Token.Type ) {
  95       case TGSI_TOKEN_TYPE_INSTRUCTION:
  96          {
  97             const struct tgsi_full_instruction *fullinst
  98                = &parse.FullToken.FullInstruction;
  99             uint i;
 100
 101             assert(fullinst->Instruction.Opcode < TGSI_OPCODE_LAST);
 102             info->opcode_count[fullinst->Instruction.Opcode]++;
 103
 104             switch (fullinst->Instruction.Opcode) {
 105             case TGSI_OPCODE_IF:
 106             case TGSI_OPCODE_UIF:
 107             case TGSI_OPCODE_BGNLOOP:
 108                current_depth++;
 109                info->max_depth = MAX2(info->max_depth, current_depth);
 110                break;
 111             case TGSI_OPCODE_ENDIF:
 112             case TGSI_OPCODE_ENDLOOP:
 113                current_depth--;
 114                break;
 115             default:
 116                break;
 117             }
 118
 119             if (fullinst->Instruction.Opcode == TGSI_OPCODE_INTERP_CENTROID ||
 120                 fullinst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
 121                 fullinst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
 122                const struct tgsi_full_src_register *src0 = &fullinst->Src[0];
 123                unsigned input;
 124
 125                if (src0->Register.Indirect && src0->Indirect.ArrayID)
 126                   input = info->input_array_first[src0->Indirect.ArrayID];
 127                else
 128                   input = src0->Register.Index;
 129
 130                /* For the INTERP opcodes, the interpolation is always
 131                 * PERSPECTIVE unless LINEAR is specified.
 132                 */
 133                switch (info->input_interpolate[input]) {
 134                case TGSI_INTERPOLATE_COLOR:
 135                case TGSI_INTERPOLATE_CONSTANT:
 136                case TGSI_INTERPOLATE_PERSPECTIVE:
 137                   switch (fullinst->Instruction.Opcode) {
 138                   case TGSI_OPCODE_INTERP_CENTROID:
 139                      info->uses_persp_opcode_interp_centroid = true;
 140                      break;
 141                   case TGSI_OPCODE_INTERP_OFFSET:
 142                      info->uses_persp_opcode_interp_offset = true;
 143                      break;
 144                   case TGSI_OPCODE_INTERP_SAMPLE:
 145                      info->uses_persp_opcode_interp_sample = true;
 146                      break;
 147                   }
 148                   break;
 149
 150                case TGSI_INTERPOLATE_LINEAR:
 151                   switch (fullinst->Instruction.Opcode) {
 152                   case TGSI_OPCODE_INTERP_CENTROID:
 153                      info->uses_linear_opcode_interp_centroid = true;
 154                      break;
 155                   case TGSI_OPCODE_INTERP_OFFSET:
 156                      info->uses_linear_opcode_interp_offset = true;
 157                      break;
 158                   case TGSI_OPCODE_INTERP_SAMPLE:
 159                      info->uses_linear_opcode_interp_sample = true;
 160                      break;
 161                   }
 162                   break;
 163                }
 164             }
 165
 166             if (fullinst->Instruction.Opcode >= TGSI_OPCODE_F2D &&
 167                 fullinst->Instruction.Opcode <= TGSI_OPCODE_DSSG)
 168                info->uses_doubles = true;
 169
 170             for (i = 0; i < fullinst->Instruction.NumSrcRegs; i++) {
 171                const struct tgsi_full_src_register *src =
 172                   &fullinst->Src[i];
 173                int ind = src->Register.Index;
 174
 175                /* Mark which inputs are effectively used */
 176                if (src->Register.File == TGSI_FILE_INPUT) {
 177                   unsigned usage_mask;
 178                   usage_mask = tgsi_util_get_inst_usage_mask(fullinst, i);
 179                   if (src->Register.Indirect) {
 180                      for (ind = 0; ind < info->num_inputs; ++ind) {
 181                         info->input_usage_mask[ind] |= usage_mask;
 182                      }
 183                   } else {
 184                      assert(ind >= 0);
 185                      assert(ind < PIPE_MAX_SHADER_INPUTS);
 186                      info->input_usage_mask[ind] |= usage_mask;
 187                   }
 188
 189                   if (procType == TGSI_PROCESSOR_FRAGMENT &&
 190                       !src->Register.Indirect) {
 191                      unsigned name =
 192                         info->input_semantic_name[src->Register.Index];
 193                      unsigned index =
 194                         info->input_semantic_index[src->Register.Index];
 195
 196                      if (name == TGSI_SEMANTIC_POSITION &&
 197                          (src->Register.SwizzleX == TGSI_SWIZZLE_Z ||
 198                           src->Register.SwizzleY == TGSI_SWIZZLE_Z ||
 199                           src->Register.SwizzleZ == TGSI_SWIZZLE_Z ||
 200                           src->Register.SwizzleW == TGSI_SWIZZLE_Z))
 201                         info->reads_z = TRUE;
 202
 203                      if (name == TGSI_SEMANTIC_COLOR) {
 204                         unsigned mask =
 205                               (1 << src->Register.SwizzleX) |
 206                               (1 << src->Register.SwizzleY) |
 207                               (1 << src->Register.SwizzleZ) |
 208                               (1 << src->Register.SwizzleW);
 209
 210                         info->colors_read |= mask << (index * 4);
 211                      }
 212                   }
 213                }
 214
 215                /* check for indirect register reads */
 216                if (src->Register.Indirect) {
 217                   info->indirect_files |= (1 << src->Register.File);
 218                   info->indirect_files_read |= (1 << src->Register.File);
 219                }
 220
 221                /* MSAA samplers */
 222                if (src->Register.File == TGSI_FILE_SAMPLER) {
 223                   assert(fullinst->Instruction.Texture);
 224                   assert(src->Register.Index < Elements(info->is_msaa_sampler));
 225
 226                   if (fullinst->Instruction.Texture &&
 227                       (fullinst->Texture.Texture == TGSI_TEXTURE_2D_MSAA ||
 228                        fullinst->Texture.Texture == TGSI_TEXTURE_2D_ARRAY_MSAA)) {
 229                      info->is_msaa_sampler[src->Register.Index] = TRUE;
 230                   }
 231                }
 232             }
 233
 234             /* check for indirect register writes */
 235             for (i = 0; i < fullinst->Instruction.NumDstRegs; i++) {
 236                const struct tgsi_full_dst_register *dst = &fullinst->Dst[i];
 237                if (dst->Register.Indirect) {
 238                   info->indirect_files |= (1 << dst->Register.File);
 239                   info->indirect_files_written |= (1 << dst->Register.File);
 240                }
 241             }
 242
 243             info->num_instructions++;
 244          }
 245          break;
 246
 247       case TGSI_TOKEN_TYPE_DECLARATION:
 248          {
 249             const struct tgsi_full_declaration *fulldecl
 250                = &parse.FullToken.FullDeclaration;
 251             const uint file = fulldecl->Declaration.File;
 252             uint reg;
 253
 254             if (fulldecl->Declaration.Array) {
 255                unsigned array_id = fulldecl->Array.ArrayID;
 256
 257                switch (file) {
 258                case TGSI_FILE_INPUT:
 259                   assert(array_id < ARRAY_SIZE(info->input_array_first));
 260                   info->input_array_first[array_id] = fulldecl->Range.First;
 261                   info->input_array_last[array_id] = fulldecl->Range.Last;
 262                   break;
 263                case TGSI_FILE_OUTPUT:
 264                   assert(array_id < ARRAY_SIZE(info->output_array_first));
 265                   info->output_array_first[array_id] = fulldecl->Range.First;
 266                   info->output_array_last[array_id] = fulldecl->Range.Last;
 267                   break;
 268                }
 269                info->array_max[file] = MAX2(info->array_max[file], array_id);
 270             }
 271
 272             for (reg = fulldecl->Range.First;
 273                  reg <= fulldecl->Range.Last;
 274                  reg++) {
 275                unsigned semName = fulldecl->Semantic.Name;
 276                unsigned semIndex =
 277                   fulldecl->Semantic.Index + (reg - fulldecl->Range.First);
 278
 279                /* only first 32 regs will appear in this bitfield */
 280                info->file_mask[file] |= (1 << reg);
 281                info->file_count[file]++;
 282                info->file_max[file] = MAX2(info->file_max[file], (int)reg);
 283
 284                if (file == TGSI_FILE_CONSTANT) {
 285                   int buffer = 0;
 286
 287                   if (fulldecl->Declaration.Dimension)
 288                      buffer = fulldecl->Dim.Index2D;
 289
 290                   info->const_file_max[buffer] =
 291                         MAX2(info->const_file_max[buffer], (int)reg);
 292                }
 293                else if (file == TGSI_FILE_INPUT) {
 294                   info->input_semantic_name[reg] = (ubyte) semName;
 295                   info->input_semantic_index[reg] = (ubyte) semIndex;
 296                   info->input_interpolate[reg] = (ubyte)fulldecl->Interp.Interpolate;
 297                   info->input_interpolate_loc[reg] = (ubyte)fulldecl->Interp.Location;
 298                   info->input_cylindrical_wrap[reg] = (ubyte)fulldecl->Interp.CylindricalWrap;
 299                   info->num_inputs++;
 300
 301                   /* Only interpolated varyings. Don't include POSITION.
 302                    * Don't include integer varyings, because they are not
 303                    * interpolated.
 304                    */
 305                   if (semName == TGSI_SEMANTIC_GENERIC ||
 306                       semName == TGSI_SEMANTIC_TEXCOORD ||
 307                       semName == TGSI_SEMANTIC_COLOR ||
 308                       semName == TGSI_SEMANTIC_BCOLOR ||
 309                       semName == TGSI_SEMANTIC_FOG ||
 310                       semName == TGSI_SEMANTIC_CLIPDIST ||
 311                       semName == TGSI_SEMANTIC_CULLDIST) {
 312                      switch (fulldecl->Interp.Interpolate) {
 313                      case TGSI_INTERPOLATE_COLOR:
 314                      case TGSI_INTERPOLATE_PERSPECTIVE:
 315                         switch (fulldecl->Interp.Location) {
 316                         case TGSI_INTERPOLATE_LOC_CENTER:
 317                            info->uses_persp_center = true;
 318                            break;
 319                         case TGSI_INTERPOLATE_LOC_CENTROID:
 320                            info->uses_persp_centroid = true;
 321                            break;
 322                         case TGSI_INTERPOLATE_LOC_SAMPLE:
 323                            info->uses_persp_sample = true;
 324                            break;
 325                         }
 326                         break;
 327                      case TGSI_INTERPOLATE_LINEAR:
 328                         switch (fulldecl->Interp.Location) {
 329                         case TGSI_INTERPOLATE_LOC_CENTER:
 330                            info->uses_linear_center = true;
 331                            break;
 332                         case TGSI_INTERPOLATE_LOC_CENTROID:
 333                            info->uses_linear_centroid = true;
 334                            break;
 335                         case TGSI_INTERPOLATE_LOC_SAMPLE:
 336                            info->uses_linear_sample = true;
 337                            break;
 338                         }
 339                         break;
 340                      /* TGSI_INTERPOLATE_CONSTANT doesn't do any interpolation. */
 341                      }
 342                   }
 343
 344                   if (semName == TGSI_SEMANTIC_PRIMID)
 345                      info->uses_primid = TRUE;
 346                   else if (procType == TGSI_PROCESSOR_FRAGMENT) {
 347                      if (semName == TGSI_SEMANTIC_POSITION)
 348                         info->reads_position = TRUE;
 349                      else if (semName == TGSI_SEMANTIC_FACE)
 350                         info->uses_frontface = TRUE;
 351                   }
 352                }
 353                else if (file == TGSI_FILE_SYSTEM_VALUE) {
 354                   unsigned index = fulldecl->Range.First;
 355
 356                   info->system_value_semantic_name[index] = semName;
 357                   info->num_system_values = MAX2(info->num_system_values,
 358                                                  index + 1);
 359
 360                   if (semName == TGSI_SEMANTIC_INSTANCEID) {
 361                      info->uses_instanceid = TRUE;
 362                   }
 363                   else if (semName == TGSI_SEMANTIC_VERTEXID) {
 364                      info->uses_vertexid = TRUE;
 365                   }
 366                   else if (semName == TGSI_SEMANTIC_VERTEXID_NOBASE) {
 367                      info->uses_vertexid_nobase = TRUE;
 368                   }
 369                   else if (semName == TGSI_SEMANTIC_BASEVERTEX) {
 370                      info->uses_basevertex = TRUE;
 371                   }
 372                   else if (semName == TGSI_SEMANTIC_PRIMID) {
 373                      info->uses_primid = TRUE;
 374                   } else if (semName == TGSI_SEMANTIC_INVOCATIONID) {
 375                      info->uses_invocationid = TRUE;
 376                   }
 377                }
 378                else if (file == TGSI_FILE_OUTPUT) {
 379                   info->output_semantic_name[reg] = (ubyte) semName;
 380                   info->output_semantic_index[reg] = (ubyte) semIndex;
 381                   info->num_outputs++;
 382
 383                   if (semName == TGSI_SEMANTIC_COLOR)
 384                      info->colors_written |= 1 << semIndex;
 385
 386                   if (procType == TGSI_PROCESSOR_VERTEX ||
 387                       procType == TGSI_PROCESSOR_GEOMETRY ||
 388                       procType == TGSI_PROCESSOR_TESS_CTRL ||
 389                       procType == TGSI_PROCESSOR_TESS_EVAL) {
 390                      if (semName == TGSI_SEMANTIC_VIEWPORT_INDEX) {
 391                         info->writes_viewport_index = TRUE;
 392                      }
 393                      else if (semName == TGSI_SEMANTIC_LAYER) {
 394                         info->writes_layer = TRUE;
 395                      }
 396                      else if (semName == TGSI_SEMANTIC_PSIZE) {
 397                         info->writes_psize = TRUE;
 398                      }
 399                      else if (semName == TGSI_SEMANTIC_CLIPVERTEX) {
 400                         info->writes_clipvertex = TRUE;
 401                      }
 402                   }
 403
 404                   if (procType == TGSI_PROCESSOR_FRAGMENT) {
 405                      if (semName == TGSI_SEMANTIC_POSITION) {
 406                         info->writes_z = TRUE;
 407                      }
 408                      else if (semName == TGSI_SEMANTIC_STENCIL) {
 409                         info->writes_stencil = TRUE;
 410                      } else if (semName == TGSI_SEMANTIC_SAMPLEMASK) {
 411                         info->writes_samplemask = TRUE;
 412                      }
 413                   }
 414
 415                   if (procType == TGSI_PROCESSOR_VERTEX) {
 416                      if (semName == TGSI_SEMANTIC_EDGEFLAG) {
 417                         info->writes_edgeflag = TRUE;
 418                      }
 419                   }
 420                } else if (file == TGSI_FILE_SAMPLER) {
 421                   info->samplers_declared |= 1 << reg;
 422                }
 423             }
 424          }
 425          break;
 426
 427       case TGSI_TOKEN_TYPE_IMMEDIATE:
 428          {
 429             uint reg = info->immediate_count++;
 430             uint file = TGSI_FILE_IMMEDIATE;
 431
 432             info->file_mask[file] |= (1 << reg);
 433             info->file_count[file]++;
 434             info->file_max[file] = MAX2(info->file_max[file], (int)reg);
 435          }
 436          break;
 437
 438       case TGSI_TOKEN_TYPE_PROPERTY:
 439          {
 440             const struct tgsi_full_property *fullprop
 441                = &parse.FullToken.FullProperty;
 442             unsigned name = fullprop->Property.PropertyName;
 443             unsigned value = fullprop->u[0].Data;
 444
 445             assert(name < Elements(info->properties));
 446             info->properties[name] = value;
 447
 448             switch (name) {
 449             case TGSI_PROPERTY_NUM_CLIPDIST_ENABLED:
 450                info->num_written_clipdistance = value;
 451                info->clipdist_writemask |= (1 << value) - 1;
 452                break;
 453             case TGSI_PROPERTY_NUM_CULLDIST_ENABLED:
 454                info->num_written_culldistance = value;
 455                info->culldist_writemask |= (1 << value) - 1;
 456                break;
 457             }
 458          }
 459          break;
 460
 461       default:
 462          assert( 0 );
 463       }
 464    }
 465
 466    info->uses_kill = (info->opcode_count[TGSI_OPCODE_KILL_IF] ||
 467                       info->opcode_count[TGSI_OPCODE_KILL]);
 468
 469    /* The dimensions of the IN decleration in geometry shader have
 470     * to be deduced from the type of the input primitive.
 471     */
 472    if (procType == TGSI_PROCESSOR_GEOMETRY) {
 473       unsigned input_primitive =
 474             info->properties[TGSI_PROPERTY_GS_INPUT_PRIM];
 475       int num_verts = u_vertices_per_prim(input_primitive);
 476       int j;
 477       info->file_count[TGSI_FILE_INPUT] = num_verts;
 478       info->file_max[TGSI_FILE_INPUT] =
 479             MAX2(info->file_max[TGSI_FILE_INPUT], num_verts - 1);
 480       for (j = 0; j < num_verts; ++j) {
 481          info->file_mask[TGSI_FILE_INPUT] |= (1 << j);
 482       }
 483    }
 484
 485    tgsi_parse_free (&parse);
 486 }
 487
 488
 489
 490 /**
 491  * Check if the given shader is a "passthrough" shader consisting of only
 492  * MOV instructions of the form:  MOV OUT[n], IN[n]
 493  *
 494  */
 495 boolean
 496 tgsi_is_passthrough_shader(const struct tgsi_token *tokens)
 497 {
 498    struct tgsi_parse_context parse;
 499
 500    /**
 501     ** Setup to begin parsing input shader
 502     **/
 503    if (tgsi_parse_init(&parse, tokens) != TGSI_PARSE_OK) {
 504       debug_printf("tgsi_parse_init() failed in tgsi_is_passthrough_shader()!\n");
 505       return FALSE;
 506    }
 507
 508    /**
 509     ** Loop over incoming program tokens/instructions
 510     */
 511    while (!tgsi_parse_end_of_tokens(&parse)) {
 512
 513       tgsi_parse_token(&parse);
 514
 515       switch (parse.FullToken.Token.Type) {
 516       case TGSI_TOKEN_TYPE_INSTRUCTION:
 517          {
 518             struct tgsi_full_instruction *fullinst =
 519                &parse.FullToken.FullInstruction;
 520             const struct tgsi_full_src_register *src =
 521                &fullinst->Src[0];
 522             const struct tgsi_full_dst_register *dst =
 523                &fullinst->Dst[0];
 524
 525             /* Do a whole bunch of checks for a simple move */
 526             if (fullinst->Instruction.Opcode != TGSI_OPCODE_MOV ||
 527                 (src->Register.File != TGSI_FILE_INPUT &&
 528                  src->Register.File != TGSI_FILE_SYSTEM_VALUE) ||
 529                 dst->Register.File != TGSI_FILE_OUTPUT ||
 530                 src->Register.Index != dst->Register.Index ||
 531
 532                 src->Register.Negate ||
 533                 src->Register.Absolute ||
 534
 535                 src->Register.SwizzleX != TGSI_SWIZZLE_X ||
 536                 src->Register.SwizzleY != TGSI_SWIZZLE_Y ||
 537                 src->Register.SwizzleZ != TGSI_SWIZZLE_Z ||
 538                 src->Register.SwizzleW != TGSI_SWIZZLE_W ||
 539
 540                 dst->Register.WriteMask != TGSI_WRITEMASK_XYZW)
 541             {
 542                tgsi_parse_free(&parse);
 543                return FALSE;
 544             }
 545          }
 546          break;
 547
 548       case TGSI_TOKEN_TYPE_DECLARATION:
 549          /* fall-through */
 550       case TGSI_TOKEN_TYPE_IMMEDIATE:
 551          /* fall-through */
 552       case TGSI_TOKEN_TYPE_PROPERTY:
 553          /* fall-through */
 554       default:
 555          ; /* no-op */
 556       }
 557    }
 558
 559    tgsi_parse_free(&parse);
 560
 561    /* if we get here, it's a pass-through shader */
 562    return TRUE;
 563 }