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 "common/gen_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 decode_group(struct gen_group *strct, const uint32_t *p, int starting_dword)
 100 {
 101    uint64_t offset = option_print_offsets ? (void *) p - gtt : 0;
 102
 103    gen_print_group(outfile, strct, offset, p, option_color == COLOR_ALWAYS);
 104 }
 105
 106 static void
 107 dump_binding_table(struct gen_spec *spec, uint32_t offset)
 108 {
 109    uint32_t *pointers, i;
 110    uint64_t start;
 111    struct gen_group *surface_state;
 112
 113    surface_state = gen_spec_find_struct(spec, "RENDER_SURFACE_STATE");
 114    if (surface_state == NULL) {
 115       fprintf(outfile, "did not find RENDER_SURFACE_STATE info\n");
 116       return;
 117    }
 118
 119    start = surface_state_base + offset;
 120    pointers = gtt + start;
 121    for (i = 0; i < 16; i++) {
 122       if (pointers[i] == 0)
 123          continue;
 124       start = pointers[i] + surface_state_base;
 125       if (!valid_offset(start)) {
 126          fprintf(outfile, "pointer %u: %08x <not valid>\n",
 127                  i, pointers[i]);
 128          continue;
 129       } else {
 130          fprintf(outfile, "pointer %u: %08x\n", i, pointers[i]);
 131       }
 132
 133       decode_group(surface_state, gtt + start, 0);
 134    }
 135 }
 136
 137 static void
 138 handle_3dstate_index_buffer(struct gen_spec *spec, uint32_t *p)
 139 {
 140    void *start;
 141    uint32_t length, i, type, size;
 142
 143    start = gtt + p[2];
 144    type = (p[1] >> 8) & 3;
 145    size = 1 << type;
 146    length = p[4] / size;
 147    if (length > 10)
 148       length = 10;
 149
 150    fprintf(outfile, "\t");
 151
 152    for (i = 0; i < length; i++) {
 153       switch (type) {
 154       case 0:
 155          fprintf(outfile, "%3d ", ((uint8_t *)start)[i]);
 156          break;
 157       case 1:
 158          fprintf(outfile, "%3d ", ((uint16_t *)start)[i]);
 159          break;
 160       case 2:
 161          fprintf(outfile, "%3d ", ((uint32_t *)start)[i]);
 162          break;
 163       }
 164    }
 165    if (length < p[4] / size)
 166       fprintf(outfile, "...\n");
 167    else
 168       fprintf(outfile, "\n");
 169 }
 170
 171 static inline uint64_t
 172 get_address(struct gen_spec *spec, uint32_t *p)
 173 {
 174    /* Addresses are always guaranteed to be page-aligned and sometimes
 175     * hardware packets have extra stuff stuffed in the bottom 12 bits.
 176     */
 177    uint64_t addr = p[0] & ~0xfffu;
 178
 179    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 180       /* On Broadwell and above, we have 48-bit addresses which consume two
 181        * dwords.  Some packets require that these get stored in a "canonical
 182        * form" which means that bit 47 is sign-extended through the upper
 183        * bits. In order to correctly handle those aub dumps, we need to mask
 184        * off the top 16 bits.
 185        */
 186       addr |= ((uint64_t)p[1] & 0xffff) << 32;
 187    }
 188
 189    return addr;
 190 }
 191
 192 static inline uint64_t
 193 get_offset(uint32_t *p, uint32_t start, uint32_t end)
 194 {
 195    assert(start <= end);
 196    assert(end < 64);
 197
 198    uint64_t mask = (~0ull >> (64 - (end - start + 1))) << start;
 199
 200    uint64_t offset = p[0];
 201    if (end >= 32)
 202       offset |= (uint64_t) p[1] << 32;
 203
 204    return offset & mask;
 205 }
 206
 207 static void
 208 handle_state_base_address(struct gen_spec *spec, uint32_t *p)
 209 {
 210    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 211       if (p[1] & 1)
 212          general_state_base = get_address(spec, &p[1]);
 213       if (p[4] & 1)
 214          surface_state_base = get_address(spec, &p[4]);
 215       if (p[6] & 1)
 216          dynamic_state_base = get_address(spec, &p[6]);
 217       if (p[10] & 1)
 218          instruction_base = get_address(spec, &p[10]);
 219       if (p[15] & 1)
 220          instruction_bound = p[15] & 0xfff;
 221    } else {
 222       if (p[2] & 1)
 223          surface_state_base = get_address(spec, &p[2]);
 224       if (p[3] & 1)
 225          dynamic_state_base = get_address(spec, &p[3]);
 226       if (p[5] & 1)
 227          instruction_base = get_address(spec, &p[5]);
 228       if (p[9] & 1)
 229          instruction_bound = get_address(spec, &p[9]);
 230    }
 231 }
 232
 233 static void
 234 dump_samplers(struct gen_spec *spec, uint32_t offset)
 235 {
 236    uint32_t i;
 237    uint64_t start;
 238    struct gen_group *sampler_state;
 239
 240    sampler_state = gen_spec_find_struct(spec, "SAMPLER_STATE");
 241
 242    start = dynamic_state_base + offset;
 243    for (i = 0; i < 4; i++) {
 244       fprintf(outfile, "sampler state %d\n", i);
 245       decode_group(sampler_state, gtt + start + i * 16, 0);
 246    }
 247 }
 248
 249 static void
 250 handle_media_interface_descriptor_load(struct gen_spec *spec, uint32_t *p)
 251 {
 252    int i, length = p[2] / 32;
 253    struct gen_group *descriptor_structure;
 254    uint32_t *descriptors;
 255    uint64_t start;
 256    struct brw_instruction *insns;
 257
 258    descriptor_structure =
 259       gen_spec_find_struct(spec, "INTERFACE_DESCRIPTOR_DATA");
 260    if (descriptor_structure == NULL) {
 261       fprintf(outfile, "did not find INTERFACE_DESCRIPTOR_DATA info\n");
 262       return;
 263    }
 264
 265    start = dynamic_state_base + p[3];
 266    descriptors = gtt + start;
 267    for (i = 0; i < length; i++, descriptors += 8) {
 268       fprintf(outfile, "descriptor %u: %08x\n", i, *descriptors);
 269       decode_group(descriptor_structure, descriptors, 0);
 270
 271       start = instruction_base + descriptors[0];
 272       if (!valid_offset(start)) {
 273          fprintf(outfile, "kernel: %08"PRIx64" <not valid>\n", start);
 274          continue;
 275       } else {
 276          fprintf(outfile, "kernel: %08"PRIx64"\n", start);
 277       }
 278
 279       insns = (struct brw_instruction *) (gtt + start);
 280       gen_disasm_disassemble(disasm, insns, 0, stdout);
 281
 282       dump_samplers(spec, descriptors[3] & ~0x1f);
 283       dump_binding_table(spec, descriptors[4] & ~0x1f);
 284    }
 285 }
 286
 287 /* Heuristic to determine whether a uint32_t is probably actually a float
 288  * (http://stackoverflow.com/a/2953466)
 289  */
 290
 291 static bool
 292 probably_float(uint32_t bits)
 293 {
 294    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 295    uint32_t mant = bits & 0x007fffff;
 296
 297    /* +- 0.0 */
 298    if (exp == -127 && mant == 0)
 299       return true;
 300
 301    /* +- 1 billionth to 1 billion */
 302    if (-30 <= exp && exp <= 30)
 303       return true;
 304
 305    /* some value with only a few binary digits */
 306    if ((mant & 0x0000ffff) == 0)
 307       return true;
 308
 309    return false;
 310 }
 311
 312 static void
 313 handle_3dstate_vertex_buffers(struct gen_spec *spec, uint32_t *p)
 314 {
 315    uint32_t *end, *s, *dw, *dwend;
 316    uint64_t offset;
 317    int n, i, count, stride;
 318
 319    end = (p[0] & 0xff) + p + 2;
 320    for (s = &p[1], n = 0; s < end; s += 4, n++) {
 321       if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 322          offset = *(uint64_t *) &s[1];
 323          dwend = gtt + offset + s[3];
 324       } else {
 325          offset = s[1];
 326          dwend = gtt + s[2] + 1;
 327       }
 328
 329       stride = field(s[0], 0, 11);
 330       count = 0;
 331       fprintf(outfile, "vertex buffer %d, size %d\n", n, s[3]);
 332       for (dw = gtt + offset, i = 0; dw < dwend && i < 256; dw++) {
 333          if (count == 0 && count % (8 * 4) == 0)
 334             fprintf(outfile, "  ");
 335
 336          if (probably_float(*dw))
 337             fprintf(outfile, "  %8.2f", *(float *) dw);
 338          else
 339             fprintf(outfile, "  0x%08x", *dw);
 340
 341          i++;
 342          count += 4;
 343
 344          if (count == stride) {
 345             fprintf(outfile, "\n");
 346             count = 0;
 347          } else if (count % (8 * 4) == 0) {
 348             fprintf(outfile, "\n");
 349          } else {
 350             fprintf(outfile, " ");
 351          }
 352       }
 353       if (count > 0 && count % (8 * 4) != 0)
 354          fprintf(outfile, "\n");
 355    }
 356 }
 357
 358 static void
 359 handle_3dstate_vs(struct gen_spec *spec, uint32_t *p)
 360 {
 361    uint64_t start;
 362    struct brw_instruction *insns;
 363    int vs_enable;
 364
 365    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 366       start = get_offset(&p[1], 6, 63);
 367       vs_enable = p[7] & 1;
 368    } else {
 369       start = get_offset(&p[1], 6, 31);
 370       vs_enable = p[5] & 1;
 371    }
 372
 373    if (vs_enable) {
 374       fprintf(outfile, "instruction_base %08"PRIx64", start %08"PRIx64"\n",
 375               instruction_base, start);
 376
 377       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 378       gen_disasm_disassemble(disasm, insns, 0, stdout);
 379    }
 380 }
 381
 382 static void
 383 handle_3dstate_hs(struct gen_spec *spec, uint32_t *p)
 384 {
 385    uint64_t start;
 386    struct brw_instruction *insns;
 387    int hs_enable;
 388
 389    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 390       start = get_offset(&p[3], 6, 63);
 391    } else {
 392       start = get_offset(&p[3], 6, 31);
 393    }
 394
 395    hs_enable = p[2] & 0x80000000;
 396
 397    if (hs_enable) {
 398       fprintf(outfile, "instruction_base %08"PRIx64", start %08"PRIx64"\n",
 399               instruction_base, start);
 400
 401       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 402       gen_disasm_disassemble(disasm, insns, 0, stdout);
 403    }
 404 }
 405
 406 static void
 407 handle_3dstate_constant(struct gen_spec *spec, uint32_t *p)
 408 {
 409    int i, j, length;
 410    uint32_t *dw;
 411    float *f;
 412
 413    for (i = 0; i < 4; i++) {
 414       length = (p[1 + i / 2] >> (i & 1) * 16) & 0xffff;
 415       f = (float *) (gtt + p[3 + i * 2] + dynamic_state_base);
 416       dw = (uint32_t *) f;
 417       for (j = 0; j < length * 8; j++) {
 418          if (probably_float(dw[j]))
 419             fprintf(outfile, "  %04.3f", f[j]);
 420          else
 421             fprintf(outfile, "  0x%08x", dw[j]);
 422
 423          if ((j & 7) == 7)
 424             fprintf(outfile, "\n");
 425       }
 426    }
 427 }
 428
 429 static void
 430 handle_3dstate_ps(struct gen_spec *spec, uint32_t *p)
 431 {
 432    uint32_t mask = ~((1 << 6) - 1);
 433    uint64_t start;
 434    struct brw_instruction *insns;
 435    static const char unused[] = "unused";
 436    static const char *pixel_type[3] = {"8 pixel", "16 pixel", "32 pixel"};
 437    const char *k0, *k1, *k2;
 438    uint32_t k_mask, k1_offset, k2_offset;
 439
 440    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 441       k_mask = p[6] & 7;
 442       k1_offset = 8;
 443       k2_offset = 10;
 444    } else {
 445       k_mask = p[4] & 7;
 446       k1_offset = 6;
 447       k2_offset = 7;
 448    }
 449
 450 #define DISPATCH_8 1
 451 #define DISPATCH_16 2
 452 #define DISPATCH_32 4
 453
 454    switch (k_mask) {
 455    case DISPATCH_8:
 456       k0 = pixel_type[0];
 457       k1 = unused;
 458       k2 = unused;
 459       break;
 460    case DISPATCH_16:
 461       k0 = pixel_type[1];
 462       k1 = unused;
 463       k2 = unused;
 464       break;
 465    case DISPATCH_8 | DISPATCH_16:
 466       k0 = pixel_type[0];
 467       k1 = unused;
 468       k2 = pixel_type[1];
 469       break;
 470    case DISPATCH_32:
 471       k0 = pixel_type[2];
 472       k1 = unused;
 473       k2 = unused;
 474       break;
 475    case DISPATCH_16 | DISPATCH_32:
 476       k0 = unused;
 477       k1 = pixel_type[2];
 478       k2 = pixel_type[1];
 479       break;
 480    case DISPATCH_8 | DISPATCH_16 | DISPATCH_32:
 481       k0 = pixel_type[0];
 482       k1 = pixel_type[2];
 483       k2 = pixel_type[1];
 484       break;
 485    default:
 486       k0 = unused;
 487       k1 = unused;
 488       k2 = unused;
 489       break;
 490    }
 491
 492    start = instruction_base + (p[1] & mask);
 493    fprintf(outfile, "  Kernel[0] %s\n", k0);
 494    if (k0 != unused) {
 495       insns = (struct brw_instruction *) (gtt + start);
 496       gen_disasm_disassemble(disasm, insns, 0, stdout);
 497    }
 498
 499    start = instruction_base + (p[k1_offset] & mask);
 500    fprintf(outfile, "  Kernel[1] %s\n", k1);
 501    if (k1 != unused) {
 502       insns = (struct brw_instruction *) (gtt + start);
 503       gen_disasm_disassemble(disasm, insns, 0, stdout);
 504    }
 505
 506    start = instruction_base + (p[k2_offset] & mask);
 507    fprintf(outfile, "  Kernel[2] %s\n", k2);
 508    if (k2 != unused) {
 509       insns = (struct brw_instruction *) (gtt + start);
 510       gen_disasm_disassemble(disasm, insns, 0, stdout);
 511    }
 512 }
 513
 514 static void
 515 handle_3dstate_binding_table_pointers(struct gen_spec *spec, uint32_t *p)
 516 {
 517    dump_binding_table(spec, p[1]);
 518 }
 519
 520 static void
 521 handle_3dstate_sampler_state_pointers(struct gen_spec *spec, uint32_t *p)
 522 {
 523    dump_samplers(spec, p[1]);
 524 }
 525
 526 static void
 527 handle_3dstate_sampler_state_pointers_gen6(struct gen_spec *spec, uint32_t *p)
 528 {
 529    dump_samplers(spec, p[1]);
 530    dump_samplers(spec, p[2]);
 531    dump_samplers(spec, p[3]);
 532 }
 533
 534 static void
 535 handle_3dstate_viewport_state_pointers_cc(struct gen_spec *spec, uint32_t *p)
 536 {
 537    uint64_t start;
 538    struct gen_group *cc_viewport;
 539
 540    cc_viewport = gen_spec_find_struct(spec, "CC_VIEWPORT");
 541
 542    start = dynamic_state_base + (p[1] & ~0x1fu);
 543    for (uint32_t i = 0; i < 4; i++) {
 544       fprintf(outfile, "viewport %d\n", i);
 545       decode_group(cc_viewport, gtt + start + i * 8, 0);
 546    }
 547 }
 548
 549 static void
 550 handle_3dstate_viewport_state_pointers_sf_clip(struct gen_spec *spec,
 551                                                uint32_t *p)
 552 {
 553    uint64_t start;
 554    struct gen_group *sf_clip_viewport;
 555
 556    sf_clip_viewport = gen_spec_find_struct(spec, "SF_CLIP_VIEWPORT");
 557
 558    start = dynamic_state_base + (p[1] & ~0x3fu);
 559    for (uint32_t i = 0; i < 4; i++) {
 560       fprintf(outfile, "viewport %d\n", i);
 561       decode_group(sf_clip_viewport, gtt + start + i * 64, 0);
 562    }
 563 }
 564
 565 static void
 566 handle_3dstate_blend_state_pointers(struct gen_spec *spec, uint32_t *p)
 567 {
 568    uint64_t start;
 569    struct gen_group *blend_state;
 570
 571    blend_state = gen_spec_find_struct(spec, "BLEND_STATE");
 572
 573    start = dynamic_state_base + (p[1] & ~0x3fu);
 574    decode_group(blend_state, gtt + start, 0);
 575 }
 576
 577 static void
 578 handle_3dstate_cc_state_pointers(struct gen_spec *spec, uint32_t *p)
 579 {
 580    uint64_t start;
 581    struct gen_group *cc_state;
 582
 583    cc_state = gen_spec_find_struct(spec, "COLOR_CALC_STATE");
 584
 585    start = dynamic_state_base + (p[1] & ~0x3fu);
 586    decode_group(cc_state, gtt + start, 0);
 587 }
 588
 589 static void
 590 handle_3dstate_scissor_state_pointers(struct gen_spec *spec, uint32_t *p)
 591 {
 592    uint64_t start;
 593    struct gen_group *scissor_rect;
 594
 595    scissor_rect = gen_spec_find_struct(spec, "SCISSOR_RECT");
 596
 597    start = dynamic_state_base + (p[1] & ~0x1fu);
 598    decode_group(scissor_rect, gtt + start, 0);
 599 }
 600
 601 static void
 602 handle_load_register_imm(struct gen_spec *spec, uint32_t *p)
 603 {
 604    struct gen_group *reg = gen_spec_find_register(spec, p[1]);
 605
 606    if (reg != NULL) {
 607       fprintf(outfile, "register %s (0x%x): 0x%x\n",
 608               reg->name, reg->register_offset, p[2]);
 609       decode_group(reg, &p[2], 0);
 610    }
 611 }
 612
 613 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
 614
 615 #define STATE_BASE_ADDRESS                  0x61010000
 616
 617 #define MEDIA_INTERFACE_DESCRIPTOR_LOAD     0x70020000
 618
 619 #define _3DSTATE_INDEX_BUFFER               0x780a0000
 620 #define _3DSTATE_VERTEX_BUFFERS             0x78080000
 621
 622 #define _3DSTATE_VS                         0x78100000
 623 #define _3DSTATE_GS                         0x78110000
 624 #define _3DSTATE_HS                         0x781b0000
 625 #define _3DSTATE_DS                         0x781d0000
 626
 627 #define _3DSTATE_CONSTANT_VS                0x78150000
 628 #define _3DSTATE_CONSTANT_GS                0x78160000
 629 #define _3DSTATE_CONSTANT_PS                0x78170000
 630 #define _3DSTATE_CONSTANT_HS                0x78190000
 631 #define _3DSTATE_CONSTANT_DS                0x781A0000
 632
 633 #define _3DSTATE_PS                         0x78200000
 634
 635 #define _3DSTATE_BINDING_TABLE_POINTERS_VS  0x78260000
 636 #define _3DSTATE_BINDING_TABLE_POINTERS_HS  0x78270000
 637 #define _3DSTATE_BINDING_TABLE_POINTERS_DS  0x78280000
 638 #define _3DSTATE_BINDING_TABLE_POINTERS_GS  0x78290000
 639 #define _3DSTATE_BINDING_TABLE_POINTERS_PS  0x782a0000
 640
 641 #define _3DSTATE_SAMPLER_STATE_POINTERS_VS  0x782b0000
 642 #define _3DSTATE_SAMPLER_STATE_POINTERS_GS  0x782e0000
 643 #define _3DSTATE_SAMPLER_STATE_POINTERS_PS  0x782f0000
 644
 645 #define _3DSTATE_SAMPLER_STATE_POINTERS     0x78020000
 646
 647 #define _3DSTATE_VIEWPORT_STATE_POINTERS_CC 0x78230000
 648 #define _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP 0x78210000
 649 #define _3DSTATE_BLEND_STATE_POINTERS       0x78240000
 650 #define _3DSTATE_CC_STATE_POINTERS          0x780e0000
 651 #define _3DSTATE_SCISSOR_STATE_POINTERS     0x780f0000
 652
 653 #define _MI_LOAD_REGISTER_IMM               0x11000000
 654
 655 struct custom_handler {
 656    uint32_t opcode;
 657    void (*handle)(struct gen_spec *spec, uint32_t *p);
 658 } custom_handlers[] = {
 659    { STATE_BASE_ADDRESS, handle_state_base_address },
 660    { MEDIA_INTERFACE_DESCRIPTOR_LOAD, handle_media_interface_descriptor_load },
 661    { _3DSTATE_VERTEX_BUFFERS, handle_3dstate_vertex_buffers },
 662    { _3DSTATE_INDEX_BUFFER, handle_3dstate_index_buffer },
 663    { _3DSTATE_VS, handle_3dstate_vs },
 664    { _3DSTATE_GS, handle_3dstate_vs },
 665    { _3DSTATE_DS, handle_3dstate_vs },
 666    { _3DSTATE_HS, handle_3dstate_hs },
 667    { _3DSTATE_CONSTANT_VS, handle_3dstate_constant },
 668    { _3DSTATE_CONSTANT_GS, handle_3dstate_constant },
 669    { _3DSTATE_CONSTANT_PS, handle_3dstate_constant },
 670    { _3DSTATE_CONSTANT_HS, handle_3dstate_constant },
 671    { _3DSTATE_CONSTANT_DS, handle_3dstate_constant },
 672    { _3DSTATE_PS, handle_3dstate_ps },
 673
 674    { _3DSTATE_BINDING_TABLE_POINTERS_VS, handle_3dstate_binding_table_pointers },
 675    { _3DSTATE_BINDING_TABLE_POINTERS_HS, handle_3dstate_binding_table_pointers },
 676    { _3DSTATE_BINDING_TABLE_POINTERS_DS, handle_3dstate_binding_table_pointers },
 677    { _3DSTATE_BINDING_TABLE_POINTERS_GS, handle_3dstate_binding_table_pointers },
 678    { _3DSTATE_BINDING_TABLE_POINTERS_PS, handle_3dstate_binding_table_pointers },
 679
 680    { _3DSTATE_SAMPLER_STATE_POINTERS_VS, handle_3dstate_sampler_state_pointers },
 681    { _3DSTATE_SAMPLER_STATE_POINTERS_GS, handle_3dstate_sampler_state_pointers },
 682    { _3DSTATE_SAMPLER_STATE_POINTERS_PS, handle_3dstate_sampler_state_pointers },
 683    { _3DSTATE_SAMPLER_STATE_POINTERS, handle_3dstate_sampler_state_pointers_gen6 },
 684
 685    { _3DSTATE_VIEWPORT_STATE_POINTERS_CC, handle_3dstate_viewport_state_pointers_cc },
 686    { _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP, handle_3dstate_viewport_state_pointers_sf_clip },
 687    { _3DSTATE_BLEND_STATE_POINTERS, handle_3dstate_blend_state_pointers },
 688    { _3DSTATE_CC_STATE_POINTERS, handle_3dstate_cc_state_pointers },
 689    { _3DSTATE_SCISSOR_STATE_POINTERS, handle_3dstate_scissor_state_pointers },
 690    { _MI_LOAD_REGISTER_IMM, handle_load_register_imm }
 691 };
 692
 693 static void
 694 parse_commands(struct gen_spec *spec, uint32_t *cmds, int size, int engine)
 695 {
 696    uint32_t *p, *end = cmds + size / 4;
 697    int length, i;
 698    struct gen_group *inst;
 699
 700    for (p = cmds; p < end; p += length) {
 701       inst = gen_spec_find_instruction(spec, p);
 702       length = gen_group_get_length(inst, p);
 703       assert(inst == NULL || length > 0);
 704       length = MAX2(1, length);
 705       if (inst == NULL) {
 706          fprintf(outfile, "unknown instruction %08x\n", p[0]);
 707          continue;
 708       }
 709
 710       const char *color, *reset_color = NORMAL;
 711       uint64_t offset;
 712
 713       if (option_full_decode) {
 714          if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START ||
 715              (p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END)
 716             color = GREEN_HEADER;
 717          else
 718             color = BLUE_HEADER;
 719       } else
 720          color = NORMAL;
 721
 722       if (option_color == COLOR_NEVER) {
 723          color = "";
 724          reset_color = "";
 725       }
 726
 727       if (option_print_offsets)
 728          offset = (void *) p - gtt;
 729       else
 730          offset = 0;
 731
 732       fprintf(outfile, "%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 733               color, offset, p[0],
 734               gen_group_get_name(inst), reset_color);
 735
 736       if (option_full_decode) {
 737          decode_group(inst, p, 0);
 738
 739          for (i = 0; i < ARRAY_LENGTH(custom_handlers); i++) {
 740             if (gen_group_get_opcode(inst) == custom_handlers[i].opcode) {
 741                custom_handlers[i].handle(spec, p);
 742                break;
 743             }
 744          }
 745       }
 746
 747       if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START) {
 748          uint64_t start = get_address(spec, &p[1]);
 749
 750          if (p[0] & (1 << 22)) {
 751             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 752              * like a subroutine call.  Commands that come afterwards get
 753              * processed once the 2nd level batch buffer returns with
 754              * MI_BATCH_BUFFER_END.
 755              */
 756             parse_commands(spec, gtt + start, gtt_end - start, engine);
 757          } else {
 758             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 759              * like a goto.  Nothing after it will ever get processed.  In
 760              * order to prevent the recursion from growing, we just reset the
 761              * loop and continue;
 762              */
 763             p = gtt + start;
 764             /* We don't know where secondaries end so use the GTT end */
 765             end = gtt + gtt_end;
 766             length = 0;
 767             continue;
 768          }
 769       } else if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END) {
 770          break;
 771       }
 772    }
 773 }
 774
 775 #define GEN_ENGINE_RENDER 1
 776 #define GEN_ENGINE_BLITTER 2
 777
 778 static void
 779 handle_trace_block(uint32_t *p)
 780 {
 781    int operation = p[1] & AUB_TRACE_OPERATION_MASK;
 782    int type = p[1] & AUB_TRACE_TYPE_MASK;
 783    int address_space = p[1] & AUB_TRACE_ADDRESS_SPACE_MASK;
 784    uint64_t offset = p[3];
 785    uint32_t size = p[4];
 786    int header_length = p[0] & 0xffff;
 787    uint32_t *data = p + header_length + 2;
 788    int engine = GEN_ENGINE_RENDER;
 789
 790    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0))
 791       offset += (uint64_t) p[5] << 32;
 792
 793    switch (operation) {
 794    case AUB_TRACE_OP_DATA_WRITE:
 795       if (address_space != AUB_TRACE_MEMTYPE_GTT)
 796          break;
 797       if (gtt_size < offset + size) {
 798          fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
 799          exit(EXIT_FAILURE);
 800       }
 801       memcpy((char *) gtt + offset, data, size);
 802       if (gtt_end < offset + size)
 803          gtt_end = offset + size;
 804       break;
 805    case AUB_TRACE_OP_COMMAND_WRITE:
 806       switch (type) {
 807       case AUB_TRACE_TYPE_RING_PRB0:
 808          engine = GEN_ENGINE_RENDER;
 809          break;
 810       case AUB_TRACE_TYPE_RING_PRB2:
 811          engine = GEN_ENGINE_BLITTER;
 812          break;
 813       default:
 814          fprintf(outfile, "command write to unknown ring %d\n", type);
 815          break;
 816       }
 817
 818       parse_commands(spec, data, size, engine);
 819       gtt_end = 0;
 820       break;
 821    }
 822 }
 823
 824 static void
 825 handle_trace_header(uint32_t *p)
 826 {
 827    /* The intel_aubdump tool from IGT is kind enough to put a PCI-ID= tag in
 828     * the AUB header comment.  If the user hasn't specified a hardware
 829     * generation, try to use the one from the AUB file.
 830     */
 831    uint32_t *end = p + (p[0] & 0xffff) + 2;
 832    int aub_pci_id = 0;
 833    if (end > &p[12] && p[12] > 0)
 834       sscanf((char *)&p[13], "PCI-ID=%i", &aub_pci_id);
 835
 836    if (pci_id == 0)
 837       pci_id = aub_pci_id;
 838
 839    struct gen_device_info devinfo;
 840    if (!gen_get_device_info(pci_id, &devinfo)) {
 841       fprintf(stderr, "can't find device information: pci_id=0x%x\n", pci_id);
 842       exit(EXIT_FAILURE);
 843    }
 844
 845    if (xml_path == NULL)
 846       spec = gen_spec_load(&devinfo);
 847    else
 848       spec = gen_spec_load_from_path(&devinfo, xml_path);
 849    disasm = gen_disasm_create(pci_id);
 850
 851    if (spec == NULL || disasm == NULL)
 852       exit(EXIT_FAILURE);
 853
 854    fprintf(outfile, "%sAubinator: Intel AUB file decoder.%-80s%s\n",
 855            GREEN_HEADER, "", NORMAL);
 856
 857    if (input_file)
 858       fprintf(outfile, "File name:        %s\n", input_file);
 859
 860    if (aub_pci_id)
 861       fprintf(outfile, "PCI ID:           0x%x\n", aub_pci_id);
 862
 863    char app_name[33];
 864    strncpy(app_name, (char *)&p[2], 32);
 865    app_name[32] = 0;
 866    fprintf(outfile, "Application name: %s\n", app_name);
 867
 868    fprintf(outfile, "Decoding as:      %s\n", gen_get_device_name(pci_id));
 869
 870    /* Throw in a new line before the first batch */
 871    fprintf(outfile, "\n");
 872 }
 873
 874 struct aub_file {
 875    FILE *stream;
 876
 877    uint32_t *map, *end, *cursor;
 878    uint32_t *mem_end;
 879 };
 880
 881 static struct aub_file *
 882 aub_file_open(const char *filename)
 883 {
 884    struct aub_file *file;
 885    struct stat sb;
 886    int fd;
 887
 888    file = calloc(1, sizeof *file);
 889    fd = open(filename, O_RDONLY);
 890    if (fd == -1) {
 891       fprintf(stderr, "open %s failed: %s\n", filename, strerror(errno));
 892       exit(EXIT_FAILURE);
 893    }
 894
 895    if (fstat(fd, &sb) == -1) {
 896       fprintf(stderr, "stat failed: %s\n", strerror(errno));
 897       exit(EXIT_FAILURE);
 898    }
 899
 900    file->map = mmap(NULL, sb.st_size,
 901                     PROT_READ, MAP_SHARED, fd, 0);
 902    if (file->map == MAP_FAILED) {
 903       fprintf(stderr, "mmap failed: %s\n", strerror(errno));
 904       exit(EXIT_FAILURE);
 905    }
 906
 907    close(fd);
 908
 909    file->cursor = file->map;
 910    file->end = file->map + sb.st_size / 4;
 911
 912    return file;
 913 }
 914
 915 static struct aub_file *
 916 aub_file_stdin(void)
 917 {
 918    struct aub_file *file;
 919
 920    file = calloc(1, sizeof *file);
 921    file->stream = stdin;
 922
 923    return file;
 924 }
 925
 926 #define TYPE(dw)       (((dw) >> 29) & 7)
 927 #define OPCODE(dw)     (((dw) >> 23) & 0x3f)
 928 #define SUBOPCODE(dw)  (((dw) >> 16) & 0x7f)
 929
 930 #define MAKE_HEADER(type, opcode, subopcode) \
 931                    (((type) << 29) | ((opcode) << 23) | ((subopcode) << 16))
 932
 933 #define TYPE_AUB            0x7
 934
 935 /* Classic AUB opcodes */
 936 #define OPCODE_AUB          0x01
 937 #define SUBOPCODE_HEADER    0x05
 938 #define SUBOPCODE_BLOCK     0x41
 939 #define SUBOPCODE_BMP       0x1e
 940
 941 /* Newer version AUB opcode */
 942 #define OPCODE_NEW_AUB      0x2e
 943 #define SUBOPCODE_VERSION   0x00
 944 #define SUBOPCODE_REG_WRITE 0x03
 945 #define SUBOPCODE_MEM_POLL  0x05
 946 #define SUBOPCODE_MEM_WRITE 0x06
 947
 948 #define MAKE_GEN(major, minor) ( ((major) << 8) | (minor) )
 949
 950 struct {
 951    const char *name;
 952    uint32_t gen;
 953 } device_map[] = {
 954    { "bwr", MAKE_GEN(4, 0) },
 955    { "cln", MAKE_GEN(4, 0) },
 956    { "blc", MAKE_GEN(4, 0) },
 957    { "ctg", MAKE_GEN(4, 0) },
 958    { "el", MAKE_GEN(4, 0) },
 959    { "il", MAKE_GEN(4, 0) },
 960    { "sbr", MAKE_GEN(6, 0) },
 961    { "ivb", MAKE_GEN(7, 0) },
 962    { "lrb2", MAKE_GEN(0, 0) },
 963    { "hsw", MAKE_GEN(7, 5) },
 964    { "vlv", MAKE_GEN(7, 0) },
 965    { "bdw", MAKE_GEN(8, 0) },
 966    { "skl", MAKE_GEN(9, 0) },
 967    { "chv", MAKE_GEN(8, 0) },
 968    { "bxt", MAKE_GEN(9, 0) },
 969    { "cnl", MAKE_GEN(10, 0) },
 970 };
 971
 972 enum {
 973    AUB_ITEM_DECODE_OK,
 974    AUB_ITEM_DECODE_FAILED,
 975    AUB_ITEM_DECODE_NEED_MORE_DATA,
 976 };
 977
 978 static int
 979 aub_file_decode_batch(struct aub_file *file)
 980 {
 981    uint32_t *p, h, device, data_type, *new_cursor;
 982    int header_length, bias;
 983
 984    if (file->end - file->cursor < 1)
 985       return AUB_ITEM_DECODE_NEED_MORE_DATA;
 986
 987    p = file->cursor;
 988    h = *p;
 989    header_length = h & 0xffff;
 990
 991    switch (OPCODE(h)) {
 992    case OPCODE_AUB:
 993       bias = 2;
 994       break;
 995    case OPCODE_NEW_AUB:
 996       bias = 1;
 997       break;
 998    default:
 999       fprintf(outfile, "unknown opcode %d at %td/%td\n",
1000               OPCODE(h), file->cursor - file->map,
1001               file->end - file->map);
1002       return AUB_ITEM_DECODE_FAILED;
1003    }
1004
1005    new_cursor = p + header_length + bias;
1006    if ((h & 0xffff0000) == MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK)) {
1007       if (file->end - file->cursor < 4)
1008          return AUB_ITEM_DECODE_NEED_MORE_DATA;
1009       new_cursor += p[4] / 4;
1010    }
1011
1012    if (new_cursor > file->end)
1013       return AUB_ITEM_DECODE_NEED_MORE_DATA;
1014
1015    switch (h & 0xffff0000) {
1016    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_HEADER):
1017       handle_trace_header(p);
1018       break;
1019    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK):
1020       handle_trace_block(p);
1021       break;
1022    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BMP):
1023       break;
1024    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_VERSION):
1025       fprintf(outfile, "version block: dw1 %08x\n", p[1]);
1026       device = (p[1] >> 8) & 0xff;
1027       fprintf(outfile, "  device %s\n", device_map[device].name);
1028       break;
1029    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_REG_WRITE):
1030       fprintf(outfile, "register write block: (dwords %d)\n", h & 0xffff);
1031       fprintf(outfile, "  reg 0x%x, data 0x%x\n", p[1], p[5]);
1032       break;
1033    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_WRITE):
1034       fprintf(outfile, "memory write block (dwords %d):\n", h & 0xffff);
1035       fprintf(outfile, "  address 0x%"PRIx64"\n", *(uint64_t *) &p[1]);
1036       data_type = (p[3] >> 20) & 0xff;
1037       if (data_type != 0)
1038          fprintf(outfile, "  data type 0x%x\n", data_type);
1039       fprintf(outfile, "  address space 0x%x\n", (p[3] >> 28) & 0xf);
1040       break;
1041    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_POLL):
1042       fprintf(outfile, "memory poll block (dwords %d):\n", h & 0xffff);
1043       break;
1044    default:
1045       fprintf(outfile, "unknown block type=0x%x, opcode=0x%x, "
1046              "subopcode=0x%x (%08x)\n", TYPE(h), OPCODE(h), SUBOPCODE(h), h);
1047       break;
1048    }
1049    file->cursor = new_cursor;
1050
1051    return AUB_ITEM_DECODE_OK;
1052 }
1053
1054 static int
1055 aub_file_more_stuff(struct aub_file *file)
1056 {
1057    return file->cursor < file->end || (file->stream && !feof(file->stream));
1058 }
1059
1060 #define AUB_READ_BUFFER_SIZE (4096)
1061 #define MAX(a, b) ((a) < (b) ? (b) : (a))
1062
1063 static void
1064 aub_file_data_grow(struct aub_file *file)
1065 {
1066    size_t old_size = (file->mem_end - file->map) * 4;
1067    size_t new_size = MAX(old_size * 2, AUB_READ_BUFFER_SIZE);
1068    uint32_t *new_start = realloc(file->map, new_size);
1069
1070    file->cursor = new_start + (file->cursor - file->map);
1071    file->end = new_start + (file->end - file->map);
1072    file->map = new_start;
1073    file->mem_end = file->map + (new_size / 4);
1074 }
1075
1076 static bool
1077 aub_file_data_load(struct aub_file *file)
1078 {
1079    size_t r;
1080
1081    if (file->stream == NULL)
1082       return false;
1083
1084    /* First remove any consumed data */
1085    if (file->cursor > file->map) {
1086       memmove(file->map, file->cursor,
1087               (file->end - file->cursor) * 4);
1088       file->end -= file->cursor - file->map;
1089       file->cursor = file->map;
1090    }
1091
1092    /* Then load some new data in */
1093    if ((file->mem_end - file->end) < (AUB_READ_BUFFER_SIZE / 4))
1094       aub_file_data_grow(file);
1095
1096    r = fread(file->end, 1, (file->mem_end - file->end) * 4, file->stream);
1097    file->end += r / 4;
1098
1099    return r != 0;
1100 }
1101
1102 static void
1103 setup_pager(void)
1104 {
1105    int fds[2];
1106    pid_t pid;
1107
1108    if (!isatty(1))
1109       return;
1110
1111    if (pipe(fds) == -1)
1112       return;
1113
1114    pid = fork();
1115    if (pid == -1)
1116       return;
1117
1118    if (pid == 0) {
1119       close(fds[1]);
1120       dup2(fds[0], 0);
1121       execlp("less", "less", "-FRSi", NULL);
1122    }
1123
1124    close(fds[0]);
1125    dup2(fds[1], 1);
1126    close(fds[1]);
1127 }
1128
1129 static void
1130 print_help(const char *progname, FILE *file)
1131 {
1132    fprintf(file,
1133            "Usage: %s [OPTION]... [FILE]\n"
1134            "Decode aub file contents from either FILE or the standard input.\n\n"
1135            "A valid --gen option must be provided.\n\n"
1136            "      --help          display this help and exit\n"
1137            "      --gen=platform  decode for given platform (ivb, byt, hsw, bdw, chv, skl, kbl, bxt or cnl)\n"
1138            "      --headers       decode only command headers\n"
1139            "      --color[=WHEN]  colorize the output; WHEN can be 'auto' (default\n"
1140            "                        if omitted), 'always', or 'never'\n"
1141            "      --no-pager      don't launch pager\n"
1142            "      --no-offsets    don't print instruction offsets\n"
1143            "      --xml=DIR       load hardware xml description from directory DIR\n",
1144            progname);
1145 }
1146
1147 int main(int argc, char *argv[])
1148 {
1149    struct aub_file *file;
1150    int c, i;
1151    bool help = false, pager = true;
1152    const struct {
1153       const char *name;
1154       int pci_id;
1155    } gens[] = {
1156       { "ilk", 0x0046 }, /* Intel(R) Ironlake Mobile */
1157       { "snb", 0x0126 }, /* Intel(R) Sandybridge Mobile GT2 */
1158       { "ivb", 0x0166 }, /* Intel(R) Ivybridge Mobile GT2 */
1159       { "hsw", 0x0416 }, /* Intel(R) Haswell Mobile GT2 */
1160       { "byt", 0x0155 }, /* Intel(R) Bay Trail */
1161       { "bdw", 0x1616 }, /* Intel(R) HD Graphics 5500 (Broadwell GT2) */
1162       { "chv", 0x22B3 }, /* Intel(R) HD Graphics (Cherryview) */
1163       { "skl", 0x1912 }, /* Intel(R) HD Graphics 530 (Skylake GT2) */
1164       { "kbl", 0x591D }, /* Intel(R) Kabylake GT2 */
1165       { "bxt", 0x0A84 },  /* Intel(R) HD Graphics (Broxton) */
1166       { "cnl", 0x5A52 },  /* Intel(R) HD Graphics (Cannonlake) */
1167    };
1168    const struct option aubinator_opts[] = {
1169       { "help",       no_argument,       (int *) &help,                 true },
1170       { "no-pager",   no_argument,       (int *) &pager,                false },
1171       { "no-offsets", no_argument,       (int *) &option_print_offsets, false },
1172       { "gen",        required_argument, NULL,                          'g' },
1173       { "headers",    no_argument,       (int *) &option_full_decode,   false },
1174       { "color",      required_argument, NULL,                          'c' },
1175       { "xml",        required_argument, NULL,                          'x' },
1176       { NULL,         0,                 NULL,                          0 }
1177    };
1178
1179    outfile = stdout;
1180
1181    i = 0;
1182    while ((c = getopt_long(argc, argv, "", aubinator_opts, &i)) != -1) {
1183       switch (c) {
1184       case 'g':
1185          for (i = 0; i < ARRAY_SIZE(gens); i++) {
1186             if (!strcmp(optarg, gens[i].name)) {
1187                pci_id = gens[i].pci_id;
1188                break;
1189             }
1190          }
1191          if (i == ARRAY_SIZE(gens)) {
1192             fprintf(stderr, "can't parse gen: '%s', expected ivb, byt, hsw, "
1193                                    "bdw, chv, skl, kbl or bxt\n", optarg);
1194             exit(EXIT_FAILURE);
1195          }
1196          break;
1197       case 'c':
1198          if (optarg == NULL || strcmp(optarg, "always") == 0)
1199             option_color = COLOR_ALWAYS;
1200          else if (strcmp(optarg, "never") == 0)
1201             option_color = COLOR_NEVER;
1202          else if (strcmp(optarg, "auto") == 0)
1203             option_color = COLOR_AUTO;
1204          else {
1205             fprintf(stderr, "invalid value for --color: %s", optarg);
1206             exit(EXIT_FAILURE);
1207          }
1208          break;
1209       case 'x':
1210          xml_path = strdup(optarg);
1211          break;
1212       default:
1213          break;
1214       }
1215    }
1216
1217    if (help || argc == 1) {
1218       print_help(argv[0], stderr);
1219       exit(0);
1220    }
1221
1222    if (optind < argc)
1223       input_file = argv[optind];
1224
1225    /* Do this before we redirect stdout to pager. */
1226    if (option_color == COLOR_AUTO)
1227       option_color = isatty(1) ? COLOR_ALWAYS : COLOR_NEVER;
1228
1229    if (isatty(1) && pager)
1230       setup_pager();
1231
1232    if (input_file == NULL)
1233       file = aub_file_stdin();
1234    else
1235       file = aub_file_open(input_file);
1236
1237    /* mmap a terabyte for our gtt space. */
1238    gtt_size = 1ull << 40;
1239    gtt = mmap(NULL, gtt_size, PROT_READ | PROT_WRITE,
1240               MAP_PRIVATE | MAP_ANONYMOUS |  MAP_NORESERVE, -1, 0);
1241    if (gtt == MAP_FAILED) {
1242       fprintf(stderr, "failed to alloc gtt space: %s\n", strerror(errno));
1243       exit(EXIT_FAILURE);
1244    }
1245
1246    while (aub_file_more_stuff(file)) {
1247       switch (aub_file_decode_batch(file)) {
1248       case AUB_ITEM_DECODE_OK:
1249          break;
1250       case AUB_ITEM_DECODE_NEED_MORE_DATA:
1251          if (!file->stream) {
1252             file->cursor = file->end;
1253             break;
1254          }
1255          if (aub_file_more_stuff(file) && !aub_file_data_load(file)) {
1256             fprintf(stderr, "failed to load data from stdin\n");
1257             exit(EXIT_FAILURE);
1258          }
1259          break;
1260       default:
1261          fprintf(stderr, "failed to parse aubdump data\n");
1262          exit(EXIT_FAILURE);
1263       }
1264    }
1265
1266
1267    fflush(stdout);
1268    /* close the stdout which is opened to write the output */
1269    close(1);
1270    free(xml_path);
1271
1272    wait(NULL);
1273
1274    return EXIT_SUCCESS;
1275 }