src/intel/tools/aubinator.c

   1 /*
   2  * Copyright © 2016 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <stdint.h>
  27 #include <getopt.h>
  28
  29 #include <unistd.h>
  30 #include <fcntl.h>
  31 #include <string.h>
  32 #include <signal.h>
  33 #include <errno.h>
  34 #include <inttypes.h>
  35 #include <sys/types.h>
  36 #include <sys/stat.h>
  37 #include <sys/wait.h>
  38 #include <sys/mman.h>
  39
  40 #include "util/macros.h"
  41
  42 #include "decoder.h"
  43 #include "intel_aub.h"
  44 #include "gen_disasm.h"
  45
  46 /* Below is the only command missing from intel_aub.h in libdrm
  47  * So, reuse intel_aub.h from libdrm and #define the
  48  * AUB_MI_BATCH_BUFFER_END as below
  49  */
  50 #define AUB_MI_BATCH_BUFFER_END (0x0500 << 16)
  51
  52 #define CSI "\e["
  53 #define BLUE_HEADER  CSI "0;44m"
  54 #define GREEN_HEADER CSI "1;42m"
  55 #define NORMAL       CSI "0m"
  56
  57 /* options */
  58
  59 static bool option_full_decode = true;
  60 static bool option_print_offsets = true;
  61 static enum { COLOR_AUTO, COLOR_ALWAYS, COLOR_NEVER } option_color;
  62
  63 /* state */
  64
  65 uint16_t pci_id = 0;
  66 char *input_file = NULL, *xml_path = NULL;
  67 struct gen_spec *spec;
  68 struct gen_disasm *disasm;
  69
  70 uint64_t gtt_size, gtt_end;
  71 void *gtt;
  72 uint64_t general_state_base;
  73 uint64_t surface_state_base;
  74 uint64_t dynamic_state_base;
  75 uint64_t instruction_base;
  76 uint64_t instruction_bound;
  77
  78 static inline uint32_t
  79 field(uint32_t value, int start, int end)
  80 {
  81    uint32_t mask;
  82
  83    mask = ~0U >> (31 - end + start);
  84
  85    return (value >> start) & mask;
  86 }
  87
  88 struct brw_instruction;
  89
  90 static inline int
  91 valid_offset(uint32_t offset)
  92 {
  93    return offset < gtt_end;
  94 }
  95
  96 static void
  97 print_dword_val(struct gen_field_iterator *iter, uint64_t offset,
  98                 int *dword_num)
  99 {
 100    struct gen_field *f;
 101
 102    f = iter->group->fields[iter->i - 1];
 103    const int dword = f->start / 32;
 104
 105    if (*dword_num != dword) {
 106       printf("0x%08"PRIx64":  0x%08x : Dword %d\n",
 107              offset + 4 * dword,  iter->p[dword], dword);
 108       *dword_num = dword;
 109    }
 110 }
 111
 112 static char *
 113 print_iterator_values(struct gen_field_iterator *iter, int *idx)
 114 {
 115     char *token = NULL;
 116     if (strstr(iter->value, "struct") == NULL) {
 117        if (strlen(iter->description) > 0) {
 118           printf("    %s: %s (%s)\n",
 119                  iter->name, iter->value, iter->description);
 120        } else {
 121           printf("    %s: %s\n", iter->name, iter->value);
 122        }
 123     } else {
 124         token = strtok(iter->value, " ");
 125         if (token != NULL) {
 126             token = strtok(NULL, " ");
 127             *idx = atoi(strtok(NULL, ">"));
 128         } else {
 129             token = NULL;
 130         }
 131         printf("    %s:<struct %s>\n", iter->name, token);
 132     }
 133     return token;
 134 }
 135
 136 static void
 137 decode_structure(struct gen_spec *spec, struct gen_group *strct,
 138                  const uint32_t *p)
 139 {
 140    struct gen_field_iterator iter;
 141    char *token = NULL;
 142    int idx = 0, dword_num = 0;
 143    uint64_t offset = 0;
 144
 145    if (option_print_offsets)
 146       offset = (void *) p - gtt;
 147    else
 148       offset = 0;
 149
 150    gen_field_iterator_init(&iter, strct, p,
 151                            option_color == COLOR_ALWAYS);
 152    while (gen_field_iterator_next(&iter)) {
 153       idx = 0;
 154       print_dword_val(&iter, offset, &dword_num);
 155       token = print_iterator_values(&iter, &idx);
 156       if (token != NULL) {
 157          struct gen_group *struct_val = gen_spec_find_struct(spec, token);
 158          decode_structure(spec, struct_val, &p[idx]);
 159          token = NULL;
 160       }
 161    }
 162 }
 163
 164 static void
 165 handle_struct_decode(struct gen_spec *spec, char *struct_name, uint32_t *p)
 166 {
 167    if (struct_name == NULL)
 168       return;
 169    struct gen_group *struct_val = gen_spec_find_struct(spec, struct_name);
 170    decode_structure(spec, struct_val, p);
 171 }
 172
 173 static void
 174 dump_binding_table(struct gen_spec *spec, uint32_t offset)
 175 {
 176    uint32_t *pointers, i;
 177    uint64_t start;
 178    struct gen_group *surface_state;
 179
 180    surface_state = gen_spec_find_struct(spec, "RENDER_SURFACE_STATE");
 181    if (surface_state == NULL) {
 182       printf("did not find RENDER_SURFACE_STATE info\n");
 183       return;
 184    }
 185
 186    start = surface_state_base + offset;
 187    pointers = gtt + start;
 188    for (i = 0; i < 16; i++) {
 189       if (pointers[i] == 0)
 190          continue;
 191       start = pointers[i] + surface_state_base;
 192       if (!valid_offset(start)) {
 193          printf("pointer %u: %08x <not valid>\n",
 194                 i, pointers[i]);
 195          continue;
 196       } else {
 197          printf("pointer %u: %08x\n", i, pointers[i]);
 198       }
 199
 200       decode_structure(spec, surface_state, gtt + start);
 201    }
 202 }
 203
 204 static void
 205 handle_3dstate_index_buffer(struct gen_spec *spec, uint32_t *p)
 206 {
 207    void *start;
 208    uint32_t length, i, type, size;
 209
 210    start = gtt + p[2];
 211    type = (p[1] >> 8) & 3;
 212    size = 1 << type;
 213    length = p[4] / size;
 214    if (length > 10)
 215       length = 10;
 216
 217    printf("\t");
 218
 219    for (i = 0; i < length; i++) {
 220       switch (type) {
 221       case 0:
 222          printf("%3d ", ((uint8_t *)start)[i]);
 223          break;
 224       case 1:
 225          printf("%3d ", ((uint16_t *)start)[i]);
 226          break;
 227       case 2:
 228          printf("%3d ", ((uint32_t *)start)[i]);
 229          break;
 230       }
 231    }
 232    if (length < p[4] / size)
 233       printf("...\n");
 234    else
 235       printf("\n");
 236 }
 237
 238 static inline uint64_t
 239 get_address(struct gen_spec *spec, uint32_t *p)
 240 {
 241    /* Addresses are always guaranteed to be page-aligned and sometimes
 242     * hardware packets have extra stuff stuffed in the bottom 12 bits.
 243     */
 244    uint64_t addr = p[0] & ~0xfffu;
 245
 246    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 247       /* On Broadwell and above, we have 48-bit addresses which consume two
 248        * dwords.  Some packets require that these get stored in a "canonical
 249        * form" which means that bit 47 is sign-extended through the upper
 250        * bits. In order to correctly handle those aub dumps, we need to mask
 251        * off the top 16 bits.
 252        */
 253       addr |= ((uint64_t)p[1] & 0xffff) << 32;
 254    }
 255
 256    return addr;
 257 }
 258
 259 static inline uint64_t
 260 get_offset(uint32_t *p, uint32_t start, uint32_t end)
 261 {
 262    assert(start <= end);
 263    assert(end < 64);
 264
 265    uint64_t mask = (~0ull >> (64 - (end - start + 1))) << start;
 266
 267    uint64_t offset = p[0];
 268    if (end >= 32)
 269       offset |= (uint64_t) p[1] << 32;
 270
 271    return offset & mask;
 272 }
 273
 274 static void
 275 handle_state_base_address(struct gen_spec *spec, uint32_t *p)
 276 {
 277    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 278       if (p[1] & 1)
 279          general_state_base = get_address(spec, &p[1]);
 280       if (p[4] & 1)
 281          surface_state_base = get_address(spec, &p[4]);
 282       if (p[6] & 1)
 283          dynamic_state_base = get_address(spec, &p[6]);
 284       if (p[10] & 1)
 285          instruction_base = get_address(spec, &p[10]);
 286       if (p[15] & 1)
 287          instruction_bound = p[15] & 0xfff;
 288    } else {
 289       if (p[2] & 1)
 290          surface_state_base = get_address(spec, &p[2]);
 291       if (p[3] & 1)
 292          dynamic_state_base = get_address(spec, &p[3]);
 293       if (p[5] & 1)
 294          instruction_base = get_address(spec, &p[5]);
 295       if (p[9] & 1)
 296          instruction_bound = get_address(spec, &p[9]);
 297    }
 298 }
 299
 300 static void
 301 dump_samplers(struct gen_spec *spec, uint32_t offset)
 302 {
 303    uint32_t i;
 304    uint64_t start;
 305    struct gen_group *sampler_state;
 306
 307    sampler_state = gen_spec_find_struct(spec, "SAMPLER_STATE");
 308
 309    start = dynamic_state_base + offset;
 310    for (i = 0; i < 4; i++) {
 311       printf("sampler state %d\n", i);
 312       decode_structure(spec, sampler_state, gtt + start + i * 16);
 313    }
 314 }
 315
 316 static void
 317 handle_media_interface_descriptor_load(struct gen_spec *spec, uint32_t *p)
 318 {
 319    int i, length = p[2] / 32;
 320    struct gen_group *descriptor_structure;
 321    uint32_t *descriptors;
 322    uint64_t start;
 323    struct brw_instruction *insns;
 324
 325    descriptor_structure =
 326       gen_spec_find_struct(spec, "INTERFACE_DESCRIPTOR_DATA");
 327    if (descriptor_structure == NULL) {
 328       printf("did not find INTERFACE_DESCRIPTOR_DATA info\n");
 329       return;
 330    }
 331
 332    start = dynamic_state_base + p[3];
 333    descriptors = gtt + start;
 334    for (i = 0; i < length; i++, descriptors += 8) {
 335       printf("descriptor %u: %08x\n", i, *descriptors);
 336       decode_structure(spec, descriptor_structure, descriptors);
 337
 338       start = instruction_base + descriptors[0];
 339       if (!valid_offset(start)) {
 340          printf("kernel: %08"PRIx64" <not valid>\n", start);
 341          continue;
 342       } else {
 343          printf("kernel: %08"PRIx64"\n", start);
 344       }
 345
 346       insns = (struct brw_instruction *) (gtt + start);
 347       gen_disasm_disassemble(disasm, insns, 0, stdout);
 348
 349       dump_samplers(spec, descriptors[3] & ~0x1f);
 350       dump_binding_table(spec, descriptors[4] & ~0x1f);
 351    }
 352 }
 353
 354 /* Heuristic to determine whether a uint32_t is probably actually a float
 355  * (http://stackoverflow.com/a/2953466)
 356  */
 357
 358 static bool
 359 probably_float(uint32_t bits)
 360 {
 361    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 362    uint32_t mant = bits & 0x007fffff;
 363
 364    /* +- 0.0 */
 365    if (exp == -127 && mant == 0)
 366       return true;
 367
 368    /* +- 1 billionth to 1 billion */
 369    if (-30 <= exp && exp <= 30)
 370       return true;
 371
 372    /* some value with only a few binary digits */
 373    if ((mant & 0x0000ffff) == 0)
 374       return true;
 375
 376    return false;
 377 }
 378
 379 static void
 380 handle_3dstate_vertex_buffers(struct gen_spec *spec, uint32_t *p)
 381 {
 382    uint32_t *end, *s, *dw, *dwend;
 383    uint64_t offset;
 384    int n, i, count, stride;
 385
 386    end = (p[0] & 0xff) + p + 2;
 387    for (s = &p[1], n = 0; s < end; s += 4, n++) {
 388       if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 389          offset = *(uint64_t *) &s[1];
 390          dwend = gtt + offset + s[3];
 391       } else {
 392          offset = s[1];
 393          dwend = gtt + s[2] + 1;
 394       }
 395
 396       stride = field(s[0], 0, 11);
 397       count = 0;
 398       printf("vertex buffer %d, size %d\n", n, s[3]);
 399       for (dw = gtt + offset, i = 0; dw < dwend && i < 256; dw++) {
 400          if (count == 0 && count % (8 * 4) == 0)
 401             printf("  ");
 402
 403          if (probably_float(*dw))
 404             printf("  %8.2f", *(float *) dw);
 405          else
 406             printf("  0x%08x", *dw);
 407
 408          i++;
 409          count += 4;
 410
 411          if (count == stride) {
 412             printf("\n");
 413             count = 0;
 414          } else if (count % (8 * 4) == 0) {
 415             printf("\n");
 416          } else {
 417             printf(" ");
 418          }
 419       }
 420       if (count > 0 && count % (8 * 4) != 0)
 421          printf("\n");
 422    }
 423 }
 424
 425 static void
 426 handle_3dstate_vs(struct gen_spec *spec, uint32_t *p)
 427 {
 428    uint64_t start;
 429    struct brw_instruction *insns;
 430    int vs_enable;
 431
 432    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 433       start = get_offset(&p[1], 6, 63);
 434       vs_enable = p[7] & 1;
 435    } else {
 436       start = get_offset(&p[1], 6, 31);
 437       vs_enable = p[5] & 1;
 438    }
 439
 440    if (vs_enable) {
 441       printf("instruction_base %08"PRIx64", start %08"PRIx64"\n",
 442              instruction_base, start);
 443
 444       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 445       gen_disasm_disassemble(disasm, insns, 0, stdout);
 446    }
 447 }
 448
 449 static void
 450 handle_3dstate_hs(struct gen_spec *spec, uint32_t *p)
 451 {
 452    uint64_t start;
 453    struct brw_instruction *insns;
 454    int hs_enable;
 455
 456    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 457       start = get_offset(&p[3], 6, 63);
 458    } else {
 459       start = get_offset(&p[3], 6, 31);
 460    }
 461
 462    hs_enable = p[2] & 0x80000000;
 463
 464    if (hs_enable) {
 465       printf("instruction_base %08"PRIx64", start %08"PRIx64"\n",
 466              instruction_base, start);
 467
 468       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 469       gen_disasm_disassemble(disasm, insns, 0, stdout);
 470    }
 471 }
 472
 473 static void
 474 handle_3dstate_constant(struct gen_spec *spec, uint32_t *p)
 475 {
 476    int i, j, length;
 477    uint32_t *dw;
 478    float *f;
 479
 480    for (i = 0; i < 4; i++) {
 481       length = (p[1 + i / 2] >> (i & 1) * 16) & 0xffff;
 482       f = (float *) (gtt + p[3 + i * 2] + dynamic_state_base);
 483       dw = (uint32_t *) f;
 484       for (j = 0; j < length * 8; j++) {
 485          if (probably_float(dw[j]))
 486             printf("  %04.3f", f[j]);
 487          else
 488             printf("  0x%08x", dw[j]);
 489
 490          if ((j & 7) == 7)
 491             printf("\n");
 492       }
 493    }
 494 }
 495
 496 static void
 497 handle_3dstate_ps(struct gen_spec *spec, uint32_t *p)
 498 {
 499    uint32_t mask = ~((1 << 6) - 1);
 500    uint64_t start;
 501    struct brw_instruction *insns;
 502    static const char unused[] = "unused";
 503    static const char *pixel_type[3] = {"8 pixel", "16 pixel", "32 pixel"};
 504    const char *k0, *k1, *k2;
 505    uint32_t k_mask, k1_offset, k2_offset;
 506
 507    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 508       k_mask = p[6] & 7;
 509       k1_offset = 8;
 510       k2_offset = 10;
 511    } else {
 512       k_mask = p[4] & 7;
 513       k1_offset = 6;
 514       k2_offset = 7;
 515    }
 516
 517 #define DISPATCH_8 1
 518 #define DISPATCH_16 2
 519 #define DISPATCH_32 4
 520
 521    switch (k_mask) {
 522    case DISPATCH_8:
 523       k0 = pixel_type[0];
 524       k1 = unused;
 525       k2 = unused;
 526       break;
 527    case DISPATCH_16:
 528       k0 = pixel_type[1];
 529       k1 = unused;
 530       k2 = unused;
 531       break;
 532    case DISPATCH_8 | DISPATCH_16:
 533       k0 = pixel_type[0];
 534       k1 = unused;
 535       k2 = pixel_type[1];
 536       break;
 537    case DISPATCH_32:
 538       k0 = pixel_type[2];
 539       k1 = unused;
 540       k2 = unused;
 541       break;
 542    case DISPATCH_16 | DISPATCH_32:
 543       k0 = unused;
 544       k1 = pixel_type[2];
 545       k2 = pixel_type[1];
 546       break;
 547    case DISPATCH_8 | DISPATCH_16 | DISPATCH_32:
 548       k0 = pixel_type[0];
 549       k1 = pixel_type[2];
 550       k2 = pixel_type[1];
 551       break;
 552    default:
 553       k0 = unused;
 554       k1 = unused;
 555       k2 = unused;
 556       break;
 557    }
 558
 559    start = instruction_base + (p[1] & mask);
 560    printf("  Kernel[0] %s\n", k0);
 561    if (k0 != unused) {
 562       insns = (struct brw_instruction *) (gtt + start);
 563       gen_disasm_disassemble(disasm, insns, 0, stdout);
 564    }
 565
 566    start = instruction_base + (p[k1_offset] & mask);
 567    printf("  Kernel[1] %s\n", k1);
 568    if (k1 != unused) {
 569       insns = (struct brw_instruction *) (gtt + start);
 570       gen_disasm_disassemble(disasm, insns, 0, stdout);
 571    }
 572
 573    start = instruction_base + (p[k2_offset] & mask);
 574    printf("  Kernel[2] %s\n", k2);
 575    if (k2 != unused) {
 576       insns = (struct brw_instruction *) (gtt + start);
 577       gen_disasm_disassemble(disasm, insns, 0, stdout);
 578    }
 579 }
 580
 581 static void
 582 handle_3dstate_binding_table_pointers(struct gen_spec *spec, uint32_t *p)
 583 {
 584    dump_binding_table(spec, p[1]);
 585 }
 586
 587 static void
 588 handle_3dstate_sampler_state_pointers(struct gen_spec *spec, uint32_t *p)
 589 {
 590    dump_samplers(spec, p[1]);
 591 }
 592
 593 static void
 594 handle_3dstate_viewport_state_pointers_cc(struct gen_spec *spec, uint32_t *p)
 595 {
 596    uint64_t start;
 597    struct gen_group *cc_viewport;
 598
 599    cc_viewport = gen_spec_find_struct(spec, "CC_VIEWPORT");
 600
 601    start = dynamic_state_base + (p[1] & ~0x1fu);
 602    for (uint32_t i = 0; i < 4; i++) {
 603       printf("viewport %d\n", i);
 604       decode_structure(spec, cc_viewport, gtt + start + i * 8);
 605    }
 606 }
 607
 608 static void
 609 handle_3dstate_viewport_state_pointers_sf_clip(struct gen_spec *spec,
 610                                                uint32_t *p)
 611 {
 612    uint64_t start;
 613    struct gen_group *sf_clip_viewport;
 614
 615    sf_clip_viewport = gen_spec_find_struct(spec, "SF_CLIP_VIEWPORT");
 616
 617    start = dynamic_state_base + (p[1] & ~0x3fu);
 618    for (uint32_t i = 0; i < 4; i++) {
 619       printf("viewport %d\n", i);
 620       decode_structure(spec, sf_clip_viewport, gtt + start + i * 64);
 621    }
 622 }
 623
 624 static void
 625 handle_3dstate_blend_state_pointers(struct gen_spec *spec, uint32_t *p)
 626 {
 627    uint64_t start;
 628    struct gen_group *blend_state;
 629
 630    blend_state = gen_spec_find_struct(spec, "BLEND_STATE");
 631
 632    start = dynamic_state_base + (p[1] & ~0x3fu);
 633    decode_structure(spec, blend_state, gtt + start);
 634 }
 635
 636 static void
 637 handle_3dstate_cc_state_pointers(struct gen_spec *spec, uint32_t *p)
 638 {
 639    uint64_t start;
 640    struct gen_group *cc_state;
 641
 642    cc_state = gen_spec_find_struct(spec, "COLOR_CALC_STATE");
 643
 644    start = dynamic_state_base + (p[1] & ~0x3fu);
 645    decode_structure(spec, cc_state, gtt + start);
 646 }
 647
 648 static void
 649 handle_3dstate_scissor_state_pointers(struct gen_spec *spec, uint32_t *p)
 650 {
 651    uint64_t start;
 652    struct gen_group *scissor_rect;
 653
 654    scissor_rect = gen_spec_find_struct(spec, "SCISSOR_RECT");
 655
 656    start = dynamic_state_base + (p[1] & ~0x1fu);
 657    decode_structure(spec, scissor_rect, gtt + start);
 658 }
 659
 660 static void
 661 handle_load_register_imm(struct gen_spec *spec, uint32_t *p)
 662 {
 663    struct gen_group *reg = gen_spec_find_register(spec, p[1]);
 664
 665    if (reg != NULL) {
 666       printf("register %s (0x%x): 0x%x\n",
 667              reg->name, reg->register_offset, p[2]);
 668       decode_structure(spec, reg, &p[2]);
 669    }
 670 }
 671
 672 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
 673
 674 #define STATE_BASE_ADDRESS                  0x61010000
 675
 676 #define MEDIA_INTERFACE_DESCRIPTOR_LOAD     0x70020000
 677
 678 #define _3DSTATE_INDEX_BUFFER               0x780a0000
 679 #define _3DSTATE_VERTEX_BUFFERS             0x78080000
 680
 681 #define _3DSTATE_VS                         0x78100000
 682 #define _3DSTATE_GS                         0x78110000
 683 #define _3DSTATE_HS                         0x781b0000
 684 #define _3DSTATE_DS                         0x781d0000
 685
 686 #define _3DSTATE_CONSTANT_VS                0x78150000
 687 #define _3DSTATE_CONSTANT_GS                0x78160000
 688 #define _3DSTATE_CONSTANT_PS                0x78170000
 689 #define _3DSTATE_CONSTANT_HS                0x78190000
 690 #define _3DSTATE_CONSTANT_DS                0x781A0000
 691
 692 #define _3DSTATE_PS                         0x78200000
 693
 694 #define _3DSTATE_BINDING_TABLE_POINTERS_VS  0x78260000
 695 #define _3DSTATE_BINDING_TABLE_POINTERS_HS  0x78270000
 696 #define _3DSTATE_BINDING_TABLE_POINTERS_DS  0x78280000
 697 #define _3DSTATE_BINDING_TABLE_POINTERS_GS  0x78290000
 698 #define _3DSTATE_BINDING_TABLE_POINTERS_PS  0x782a0000
 699
 700 #define _3DSTATE_SAMPLER_STATE_POINTERS_VS  0x782b0000
 701 #define _3DSTATE_SAMPLER_STATE_POINTERS_GS  0x782e0000
 702 #define _3DSTATE_SAMPLER_STATE_POINTERS_PS  0x782f0000
 703
 704 #define _3DSTATE_VIEWPORT_STATE_POINTERS_CC 0x78230000
 705 #define _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP 0x78210000
 706 #define _3DSTATE_BLEND_STATE_POINTERS       0x78240000
 707 #define _3DSTATE_CC_STATE_POINTERS          0x780e0000
 708 #define _3DSTATE_SCISSOR_STATE_POINTERS     0x780f0000
 709
 710 #define _MI_LOAD_REGISTER_IMM               0x11000000
 711
 712 struct custom_handler {
 713    uint32_t opcode;
 714    void (*handle)(struct gen_spec *spec, uint32_t *p);
 715 } custom_handlers[] = {
 716    { STATE_BASE_ADDRESS, handle_state_base_address },
 717    { MEDIA_INTERFACE_DESCRIPTOR_LOAD, handle_media_interface_descriptor_load },
 718    { _3DSTATE_VERTEX_BUFFERS, handle_3dstate_vertex_buffers },
 719    { _3DSTATE_INDEX_BUFFER, handle_3dstate_index_buffer },
 720    { _3DSTATE_VS, handle_3dstate_vs },
 721    { _3DSTATE_GS, handle_3dstate_vs },
 722    { _3DSTATE_DS, handle_3dstate_vs },
 723    { _3DSTATE_HS, handle_3dstate_hs },
 724    { _3DSTATE_CONSTANT_VS, handle_3dstate_constant },
 725    { _3DSTATE_CONSTANT_GS, handle_3dstate_constant },
 726    { _3DSTATE_CONSTANT_PS, handle_3dstate_constant },
 727    { _3DSTATE_CONSTANT_HS, handle_3dstate_constant },
 728    { _3DSTATE_CONSTANT_DS, handle_3dstate_constant },
 729    { _3DSTATE_PS, handle_3dstate_ps },
 730
 731    { _3DSTATE_BINDING_TABLE_POINTERS_VS, handle_3dstate_binding_table_pointers },
 732    { _3DSTATE_BINDING_TABLE_POINTERS_HS, handle_3dstate_binding_table_pointers },
 733    { _3DSTATE_BINDING_TABLE_POINTERS_DS, handle_3dstate_binding_table_pointers },
 734    { _3DSTATE_BINDING_TABLE_POINTERS_GS, handle_3dstate_binding_table_pointers },
 735    { _3DSTATE_BINDING_TABLE_POINTERS_PS, handle_3dstate_binding_table_pointers },
 736
 737    { _3DSTATE_SAMPLER_STATE_POINTERS_VS, handle_3dstate_sampler_state_pointers },
 738    { _3DSTATE_SAMPLER_STATE_POINTERS_GS, handle_3dstate_sampler_state_pointers },
 739    { _3DSTATE_SAMPLER_STATE_POINTERS_PS, handle_3dstate_sampler_state_pointers },
 740
 741    { _3DSTATE_VIEWPORT_STATE_POINTERS_CC, handle_3dstate_viewport_state_pointers_cc },
 742    { _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP, handle_3dstate_viewport_state_pointers_sf_clip },
 743    { _3DSTATE_BLEND_STATE_POINTERS, handle_3dstate_blend_state_pointers },
 744    { _3DSTATE_CC_STATE_POINTERS, handle_3dstate_cc_state_pointers },
 745    { _3DSTATE_SCISSOR_STATE_POINTERS, handle_3dstate_scissor_state_pointers },
 746    { _MI_LOAD_REGISTER_IMM, handle_load_register_imm }
 747 };
 748
 749 static void
 750 parse_commands(struct gen_spec *spec, uint32_t *cmds, int size, int engine)
 751 {
 752    uint32_t *p, *end = cmds + size / 4;
 753    unsigned int length, i;
 754    struct gen_group *inst;
 755
 756    for (p = cmds; p < end; p += length) {
 757       inst = gen_spec_find_instruction(spec, p);
 758       if (inst == NULL) {
 759          printf("unknown instruction %08x\n", p[0]);
 760          length = (p[0] & 0xff) + 2;
 761          continue;
 762       }
 763       length = gen_group_get_length(inst, p);
 764
 765       const char *color, *reset_color = NORMAL;
 766       uint64_t offset;
 767
 768       if (option_full_decode) {
 769          if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START ||
 770              (p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END)
 771             color = GREEN_HEADER;
 772          else
 773             color = BLUE_HEADER;
 774       } else
 775          color = NORMAL;
 776
 777       if (option_color == COLOR_NEVER) {
 778          color = "";
 779          reset_color = "";
 780       }
 781
 782       if (option_print_offsets)
 783          offset = (void *) p - gtt;
 784       else
 785          offset = 0;
 786
 787       printf("%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 788              color, offset, p[0],
 789              gen_group_get_name(inst), reset_color);
 790
 791       if (option_full_decode) {
 792          struct gen_field_iterator iter;
 793          char *token = NULL;
 794          int idx = 0, dword_num = 0;
 795          gen_field_iterator_init(&iter, inst, p,
 796                                  option_color == COLOR_ALWAYS);
 797          while (gen_field_iterator_next(&iter)) {
 798             idx = 0;
 799             print_dword_val(&iter, offset, &dword_num);
 800             if (dword_num > 0)
 801                token = print_iterator_values(&iter, &idx);
 802             if (token != NULL) {
 803                printf("0x%08"PRIx64":  0x%08x : Dword %d\n",
 804                       offset + 4 * idx, p[idx], idx);
 805                handle_struct_decode(spec,token, &p[idx]);
 806                token = NULL;
 807             }
 808          }
 809
 810          for (i = 0; i < ARRAY_LENGTH(custom_handlers); i++) {
 811             if (gen_group_get_opcode(inst) == custom_handlers[i].opcode)
 812                custom_handlers[i].handle(spec, p);
 813          }
 814       }
 815
 816       if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START) {
 817          uint64_t start = get_address(spec, &p[1]);
 818
 819          if (p[0] & (1 << 22)) {
 820             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 821              * like a subroutine call.  Commands that come afterwards get
 822              * processed once the 2nd level batch buffer returns with
 823              * MI_BATCH_BUFFER_END.
 824              */
 825             parse_commands(spec, gtt + start, gtt_end - start, engine);
 826          } else {
 827             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 828              * like a goto.  Nothing after it will ever get processed.  In
 829              * order to prevent the recursion from growing, we just reset the
 830              * loop and continue;
 831              */
 832             p = gtt + start;
 833             /* We don't know where secondaries end so use the GTT end */
 834             end = gtt + gtt_end;
 835             length = 0;
 836             continue;
 837          }
 838       } else if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END) {
 839          break;
 840       }
 841    }
 842 }
 843
 844 #define GEN_ENGINE_RENDER 1
 845 #define GEN_ENGINE_BLITTER 2
 846
 847 static void
 848 handle_trace_block(uint32_t *p)
 849 {
 850    int operation = p[1] & AUB_TRACE_OPERATION_MASK;
 851    int type = p[1] & AUB_TRACE_TYPE_MASK;
 852    int address_space = p[1] & AUB_TRACE_ADDRESS_SPACE_MASK;
 853    uint64_t offset = p[3];
 854    uint32_t size = p[4];
 855    int header_length = p[0] & 0xffff;
 856    uint32_t *data = p + header_length + 2;
 857    int engine = GEN_ENGINE_RENDER;
 858
 859    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0))
 860       offset += (uint64_t) p[5] << 32;
 861
 862    switch (operation) {
 863    case AUB_TRACE_OP_DATA_WRITE:
 864       if (address_space != AUB_TRACE_MEMTYPE_GTT)
 865          break;
 866       if (gtt_size < offset + size) {
 867          fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
 868          exit(EXIT_FAILURE);
 869       }
 870       memcpy((char *) gtt + offset, data, size);
 871       if (gtt_end < offset + size)
 872          gtt_end = offset + size;
 873       break;
 874    case AUB_TRACE_OP_COMMAND_WRITE:
 875       switch (type) {
 876       case AUB_TRACE_TYPE_RING_PRB0:
 877          engine = GEN_ENGINE_RENDER;
 878          break;
 879       case AUB_TRACE_TYPE_RING_PRB2:
 880          engine = GEN_ENGINE_BLITTER;
 881          break;
 882       default:
 883          printf("command write to unknown ring %d\n", type);
 884          break;
 885       }
 886
 887       parse_commands(spec, data, size, engine);
 888       gtt_end = 0;
 889       break;
 890    }
 891 }
 892
 893 static void
 894 handle_trace_header(uint32_t *p)
 895 {
 896    /* The intel_aubdump tool from IGT is kind enough to put a PCI-ID= tag in
 897     * the AUB header comment.  If the user hasn't specified a hardware
 898     * generation, try to use the one from the AUB file.
 899     */
 900    uint32_t *end = p + (p[0] & 0xffff) + 2;
 901    int aub_pci_id = 0;
 902    if (end > &p[12] && p[12] > 0)
 903       sscanf((char *)&p[13], "PCI-ID=%i", &aub_pci_id);
 904
 905    if (pci_id == 0)
 906       pci_id = aub_pci_id;
 907
 908    struct gen_device_info devinfo;
 909    if (!gen_get_device_info(pci_id, &devinfo)) {
 910       fprintf(stderr, "can't find device information: pci_id=0x%x\n", pci_id);
 911       exit(EXIT_FAILURE);
 912    }
 913
 914    if (xml_path == NULL)
 915       spec = gen_spec_load(&devinfo);
 916    else
 917       spec = gen_spec_load_from_path(&devinfo, xml_path);
 918    disasm = gen_disasm_create(pci_id);
 919
 920    if (spec == NULL || disasm == NULL)
 921       exit(EXIT_FAILURE);
 922
 923    printf("%sAubinator: Intel AUB file decoder.%-80s%s\n",
 924           GREEN_HEADER, "", NORMAL);
 925
 926    if (input_file)
 927       printf("File name:        %s\n", input_file);
 928
 929    if (aub_pci_id)
 930       printf("PCI ID:           0x%x\n", aub_pci_id);
 931
 932    char app_name[33];
 933    strncpy(app_name, (char *)&p[2], 32);
 934    app_name[32] = 0;
 935    printf("Application name: %s\n", app_name);
 936
 937    printf("Decoding as:      %s\n", gen_get_device_name(pci_id));
 938
 939    /* Throw in a new line before the first batch */
 940    printf("\n");
 941 }
 942
 943 struct aub_file {
 944    FILE *stream;
 945
 946    uint32_t *map, *end, *cursor;
 947    uint32_t *mem_end;
 948 };
 949
 950 static struct aub_file *
 951 aub_file_open(const char *filename)
 952 {
 953    struct aub_file *file;
 954    struct stat sb;
 955    int fd;
 956
 957    file = calloc(1, sizeof *file);
 958    fd = open(filename, O_RDONLY);
 959    if (fd == -1) {
 960       fprintf(stderr, "open %s failed: %s\n", filename, strerror(errno));
 961       exit(EXIT_FAILURE);
 962    }
 963
 964    if (fstat(fd, &sb) == -1) {
 965       fprintf(stderr, "stat failed: %s\n", strerror(errno));
 966       exit(EXIT_FAILURE);
 967    }
 968
 969    file->map = mmap(NULL, sb.st_size,
 970                     PROT_READ, MAP_SHARED, fd, 0);
 971    if (file->map == MAP_FAILED) {
 972       fprintf(stderr, "mmap failed: %s\n", strerror(errno));
 973       exit(EXIT_FAILURE);
 974    }
 975
 976    file->cursor = file->map;
 977    file->end = file->map + sb.st_size / 4;
 978
 979    return file;
 980 }
 981
 982 static struct aub_file *
 983 aub_file_stdin(void)
 984 {
 985    struct aub_file *file;
 986
 987    file = calloc(1, sizeof *file);
 988    file->stream = stdin;
 989
 990    return file;
 991 }
 992
 993 #define TYPE(dw)       (((dw) >> 29) & 7)
 994 #define OPCODE(dw)     (((dw) >> 23) & 0x3f)
 995 #define SUBOPCODE(dw)  (((dw) >> 16) & 0x7f)
 996
 997 #define MAKE_HEADER(type, opcode, subopcode) \
 998                    (((type) << 29) | ((opcode) << 23) | ((subopcode) << 16))
 999
1000 #define TYPE_AUB            0x7
1001
1002 /* Classic AUB opcodes */
1003 #define OPCODE_AUB          0x01
1004 #define SUBOPCODE_HEADER    0x05
1005 #define SUBOPCODE_BLOCK     0x41
1006 #define SUBOPCODE_BMP       0x1e
1007
1008 /* Newer version AUB opcode */
1009 #define OPCODE_NEW_AUB      0x2e
1010 #define SUBOPCODE_VERSION   0x00
1011 #define SUBOPCODE_REG_WRITE 0x03
1012 #define SUBOPCODE_MEM_POLL  0x05
1013 #define SUBOPCODE_MEM_WRITE 0x06
1014
1015 #define MAKE_GEN(major, minor) ( ((major) << 8) | (minor) )
1016
1017 struct {
1018    const char *name;
1019    uint32_t gen;
1020 } device_map[] = {
1021    { "bwr", MAKE_GEN(4, 0) },
1022    { "cln", MAKE_GEN(4, 0) },
1023    { "blc", MAKE_GEN(4, 0) },
1024    { "ctg", MAKE_GEN(4, 0) },
1025    { "el", MAKE_GEN(4, 0) },
1026    { "il", MAKE_GEN(4, 0) },
1027    { "sbr", MAKE_GEN(6, 0) },
1028    { "ivb", MAKE_GEN(7, 0) },
1029    { "lrb2", MAKE_GEN(0, 0) },
1030    { "hsw", MAKE_GEN(7, 5) },
1031    { "vlv", MAKE_GEN(7, 0) },
1032    { "bdw", MAKE_GEN(8, 0) },
1033    { "skl", MAKE_GEN(9, 0) },
1034    { "chv", MAKE_GEN(8, 0) },
1035    { "bxt", MAKE_GEN(9, 0) }
1036 };
1037
1038 enum {
1039    AUB_ITEM_DECODE_OK,
1040    AUB_ITEM_DECODE_FAILED,
1041    AUB_ITEM_DECODE_NEED_MORE_DATA,
1042 };
1043
1044 static int
1045 aub_file_decode_batch(struct aub_file *file)
1046 {
1047    uint32_t *p, h, device, data_type, *new_cursor;
1048    int header_length, bias;
1049
1050    if (file->end - file->cursor < 1)
1051       return AUB_ITEM_DECODE_NEED_MORE_DATA;
1052
1053    p = file->cursor;
1054    h = *p;
1055    header_length = h & 0xffff;
1056
1057    switch (OPCODE(h)) {
1058    case OPCODE_AUB:
1059       bias = 2;
1060       break;
1061    case OPCODE_NEW_AUB:
1062       bias = 1;
1063       break;
1064    default:
1065       printf("unknown opcode %d at %td/%td\n",
1066              OPCODE(h), file->cursor - file->map,
1067              file->end - file->map);
1068       return AUB_ITEM_DECODE_FAILED;
1069    }
1070
1071    new_cursor = p + header_length + bias;
1072    if ((h & 0xffff0000) == MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK)) {
1073       if (file->end - file->cursor < 4)
1074          return AUB_ITEM_DECODE_NEED_MORE_DATA;
1075       new_cursor += p[4] / 4;
1076    }
1077
1078    if (new_cursor > file->end)
1079       return AUB_ITEM_DECODE_NEED_MORE_DATA;
1080
1081    switch (h & 0xffff0000) {
1082    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_HEADER):
1083       handle_trace_header(p);
1084       break;
1085    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK):
1086       handle_trace_block(p);
1087       break;
1088    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BMP):
1089       break;
1090    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_VERSION):
1091       printf("version block: dw1 %08x\n", p[1]);
1092       device = (p[1] >> 8) & 0xff;
1093       printf("  device %s\n", device_map[device].name);
1094       break;
1095    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_REG_WRITE):
1096       printf("register write block: (dwords %d)\n", h & 0xffff);
1097       printf("  reg 0x%x, data 0x%x\n", p[1], p[5]);
1098       break;
1099    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_WRITE):
1100       printf("memory write block (dwords %d):\n", h & 0xffff);
1101       printf("  address 0x%"PRIx64"\n", *(uint64_t *) &p[1]);
1102       data_type = (p[3] >> 20) & 0xff;
1103       if (data_type != 0)
1104          printf("  data type 0x%x\n", data_type);
1105       printf("  address space 0x%x\n", (p[3] >> 28) & 0xf);
1106       break;
1107    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_POLL):
1108       printf("memory poll block (dwords %d):\n", h & 0xffff);
1109       break;
1110    default:
1111       printf("unknown block type=0x%x, opcode=0x%x, "
1112              "subopcode=0x%x (%08x)\n", TYPE(h), OPCODE(h), SUBOPCODE(h), h);
1113       break;
1114    }
1115    file->cursor = new_cursor;
1116
1117    return AUB_ITEM_DECODE_OK;
1118 }
1119
1120 static int
1121 aub_file_more_stuff(struct aub_file *file)
1122 {
1123    return file->cursor < file->end || (file->stream && !feof(file->stream));
1124 }
1125
1126 #define AUB_READ_BUFFER_SIZE (4096)
1127 #define MAX(a, b) ((a) < (b) ? (b) : (a))
1128
1129 static void
1130 aub_file_data_grow(struct aub_file *file)
1131 {
1132    size_t old_size = (file->mem_end - file->map) * 4;
1133    size_t new_size = MAX(old_size * 2, AUB_READ_BUFFER_SIZE);
1134    uint32_t *new_start = realloc(file->map, new_size);
1135
1136    file->cursor = new_start + (file->cursor - file->map);
1137    file->end = new_start + (file->end - file->map);
1138    file->map = new_start;
1139    file->mem_end = file->map + (new_size / 4);
1140 }
1141
1142 static bool
1143 aub_file_data_load(struct aub_file *file)
1144 {
1145    size_t r;
1146
1147    if (file->stream == NULL)
1148       return false;
1149
1150    /* First remove any consumed data */
1151    if (file->cursor > file->map) {
1152       memmove(file->map, file->cursor,
1153               (file->end - file->cursor) * 4);
1154       file->end -= file->cursor - file->map;
1155       file->cursor = file->map;
1156    }
1157
1158    /* Then load some new data in */
1159    if ((file->mem_end - file->end) < (AUB_READ_BUFFER_SIZE / 4))
1160       aub_file_data_grow(file);
1161
1162    r = fread(file->end, 1, (file->mem_end - file->end) * 4, file->stream);
1163    file->end += r / 4;
1164
1165    return r != 0;
1166 }
1167
1168 static void
1169 setup_pager(void)
1170 {
1171    int fds[2];
1172    pid_t pid;
1173
1174    if (!isatty(1))
1175       return;
1176
1177    if (pipe(fds) == -1)
1178       return;
1179
1180    pid = fork();
1181    if (pid == -1)
1182       return;
1183
1184    if (pid == 0) {
1185       close(fds[1]);
1186       dup2(fds[0], 0);
1187       execlp("less", "less", "-FRSi", NULL);
1188    }
1189
1190    close(fds[0]);
1191    dup2(fds[1], 1);
1192    close(fds[1]);
1193 }
1194
1195 static void
1196 print_help(const char *progname, FILE *file)
1197 {
1198    fprintf(file,
1199            "Usage: %s [OPTION]... [FILE]\n"
1200            "Decode aub file contents from either FILE or the standard input.\n\n"
1201            "A valid --gen option must be provided.\n\n"
1202            "      --help          display this help and exit\n"
1203            "      --gen=platform  decode for given platform (ivb, byt, hsw, bdw, chv, skl, kbl or bxt)\n"
1204            "      --headers       decode only command headers\n"
1205            "      --color[=WHEN]  colorize the output; WHEN can be 'auto' (default\n"
1206            "                        if omitted), 'always', or 'never'\n"
1207            "      --no-pager      don't launch pager\n"
1208            "      --no-offsets    don't print instruction offsets\n"
1209            "      --xml=DIR       load hardware xml description from directory DIR\n",
1210            progname);
1211 }
1212
1213 int main(int argc, char *argv[])
1214 {
1215    struct aub_file *file;
1216    int c, i;
1217    bool help = false, pager = true;
1218    const struct {
1219       const char *name;
1220       int pci_id;
1221    } gens[] = {
1222       { "ivb", 0x0166 }, /* Intel(R) Ivybridge Mobile GT2 */
1223       { "hsw", 0x0416 }, /* Intel(R) Haswell Mobile GT2 */
1224       { "byt", 0x0155 }, /* Intel(R) Bay Trail */
1225       { "bdw", 0x1616 }, /* Intel(R) HD Graphics 5500 (Broadwell GT2) */
1226       { "chv", 0x22B3 }, /* Intel(R) HD Graphics (Cherryview) */
1227       { "skl", 0x1912 }, /* Intel(R) HD Graphics 530 (Skylake GT2) */
1228       { "kbl", 0x591D }, /* Intel(R) Kabylake GT2 */
1229       { "bxt", 0x0A84 }  /* Intel(R) HD Graphics (Broxton) */
1230    };
1231    const struct option aubinator_opts[] = {
1232       { "help",       no_argument,       (int *) &help,                 true },
1233       { "no-pager",   no_argument,       (int *) &pager,                false },
1234       { "no-offsets", no_argument,       (int *) &option_print_offsets, false },
1235       { "gen",        required_argument, NULL,                          'g' },
1236       { "headers",    no_argument,       (int *) &option_full_decode,   false },
1237       { "color",      required_argument, NULL,                          'c' },
1238       { "xml",        required_argument, NULL,                          'x' },
1239       { NULL,         0,                 NULL,                          0 }
1240    };
1241
1242    i = 0;
1243    while ((c = getopt_long(argc, argv, "", aubinator_opts, &i)) != -1) {
1244       switch (c) {
1245       case 'g':
1246          for (i = 0; i < ARRAY_SIZE(gens); i++) {
1247             if (!strcmp(optarg, gens[i].name)) {
1248                pci_id = gens[i].pci_id;
1249                break;
1250             }
1251          }
1252          if (i == ARRAY_SIZE(gens)) {
1253             fprintf(stderr, "can't parse gen: '%s', expected ivb, byt, hsw, "
1254                                    "bdw, chv, skl, kbl or bxt\n", optarg);
1255             exit(EXIT_FAILURE);
1256          }
1257          break;
1258       case 'c':
1259          if (optarg == NULL || strcmp(optarg, "always") == 0)
1260             option_color = COLOR_ALWAYS;
1261          else if (strcmp(optarg, "never") == 0)
1262             option_color = COLOR_NEVER;
1263          else if (strcmp(optarg, "auto") == 0)
1264             option_color = COLOR_AUTO;
1265          else {
1266             fprintf(stderr, "invalid value for --color: %s", optarg);
1267             exit(EXIT_FAILURE);
1268          }
1269          break;
1270       case 'x':
1271          xml_path = strdup(optarg);
1272          break;
1273       default:
1274          break;
1275       }
1276    }
1277
1278    if (help || argc == 1) {
1279       print_help(argv[0], stderr);
1280       exit(0);
1281    }
1282
1283    if (optind < argc)
1284       input_file = argv[optind];
1285
1286    /* Do this before we redirect stdout to pager. */
1287    if (option_color == COLOR_AUTO)
1288       option_color = isatty(1) ? COLOR_ALWAYS : COLOR_NEVER;
1289
1290    if (isatty(1) && pager)
1291       setup_pager();
1292
1293    if (input_file == NULL)
1294       file = aub_file_stdin();
1295    else
1296       file = aub_file_open(input_file);
1297
1298    /* mmap a terabyte for our gtt space. */
1299    gtt_size = 1ull << 40;
1300    gtt = mmap(NULL, gtt_size, PROT_READ | PROT_WRITE,
1301               MAP_PRIVATE | MAP_ANONYMOUS |  MAP_NORESERVE, -1, 0);
1302    if (gtt == MAP_FAILED) {
1303       fprintf(stderr, "failed to alloc gtt space: %s\n", strerror(errno));
1304       exit(EXIT_FAILURE);
1305    }
1306
1307    while (aub_file_more_stuff(file)) {
1308       switch (aub_file_decode_batch(file)) {
1309       case AUB_ITEM_DECODE_OK:
1310          break;
1311       case AUB_ITEM_DECODE_NEED_MORE_DATA:
1312          if (!file->stream) {
1313             file->cursor = file->end;
1314             break;
1315          }
1316          if (aub_file_more_stuff(file) && !aub_file_data_load(file)) {
1317             fprintf(stderr, "failed to load data from stdin\n");
1318             exit(EXIT_FAILURE);
1319          }
1320          break;
1321       default:
1322          fprintf(stderr, "failed to parse aubdump data\n");
1323          exit(EXIT_FAILURE);
1324       }
1325    }
1326
1327
1328    fflush(stdout);
1329    /* close the stdout which is opened to write the output */
1330    close(1);
1331    free(xml_path);
1332
1333    wait(NULL);
1334
1335    return EXIT_SUCCESS;
1336 }