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