src/intel/tools/aubinator.c

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