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