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