Home | History | Annotate | Download | only in tools
      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 static inline uint32_t
     79 field(uint32_t value, int start, int end)
     80 {
     81    uint32_t mask;
     82 
     83    mask = ~0U >> (31 - end + start);
     84 
     85    return (value >> start) & mask;
     86 }
     87 
     88 struct brw_instruction;
     89 
     90 static inline int
     91 valid_offset(uint32_t offset)
     92 {
     93    return offset < gtt_end;
     94 }
     95 
     96 static void
     97 print_dword_val(struct gen_field_iterator *iter, uint64_t offset,
     98                 int *dword_num)
     99 {
    100    struct gen_field *f;
    101 
    102    f = iter->group->fields[iter->i - 1];
    103    const int dword = f->start / 32;
    104 
    105    if (*dword_num != dword) {
    106       printf("0x%08"PRIx64":  0x%08x : Dword %d\n",
    107              offset + 4 * dword,  iter->p[dword], dword);
    108       *dword_num = dword;
    109    }
    110 }
    111 
    112 static char *
    113 print_iterator_values(struct gen_field_iterator *iter, int *idx)
    114 {
    115     char *token = NULL;
    116     if (strstr(iter->value, "struct") == NULL) {
    117        if (strlen(iter->description) > 0) {
    118           printf("    %s: %s (%s)\n",
    119                  iter->name, iter->value, iter->description);
    120        } else {
    121           printf("    %s: %s\n", iter->name, iter->value);
    122        }
    123     } else {
    124         token = strtok(iter->value, " ");
    125         if (token != NULL) {
    126             token = strtok(NULL, " ");
    127             *idx = atoi(strtok(NULL, ">"));
    128         } else {
    129             token = NULL;
    130         }
    131         printf("    %s:<struct %s>\n", iter->name, token);
    132     }
    133     return token;
    134 }
    135 
    136 static void
    137 decode_structure(struct gen_spec *spec, struct gen_group *strct,
    138                  const uint32_t *p)
    139 {
    140    struct gen_field_iterator iter;
    141    char *token = NULL;
    142    int idx = 0, dword_num = 0;
    143    uint64_t offset = 0;
    144 
    145    if (option_print_offsets)
    146       offset = (void *) p - gtt;
    147    else
    148       offset = 0;
    149 
    150    gen_field_iterator_init(&iter, strct, p,
    151                            option_color == COLOR_ALWAYS);
    152    while (gen_field_iterator_next(&iter)) {
    153       idx = 0;
    154       print_dword_val(&iter, offset, &dword_num);
    155       token = print_iterator_values(&iter, &idx);
    156       if (token != NULL) {
    157          struct gen_group *struct_val = gen_spec_find_struct(spec, token);
    158          decode_structure(spec, struct_val, &p[idx]);
    159          token = NULL;
    160       }
    161    }
    162 }
    163 
    164 static void
    165 handle_struct_decode(struct gen_spec *spec, char *struct_name, uint32_t *p)
    166 {
    167    if (struct_name == NULL)
    168       return;
    169    struct gen_group *struct_val = gen_spec_find_struct(spec, struct_name);
    170    decode_structure(spec, struct_val, p);
    171 }
    172 
    173 static void
    174 dump_binding_table(struct gen_spec *spec, uint32_t offset)
    175 {
    176    uint32_t *pointers, i;
    177    uint64_t start;
    178    struct gen_group *surface_state;
    179 
    180    surface_state = gen_spec_find_struct(spec, "RENDER_SURFACE_STATE");
    181    if (surface_state == NULL) {
    182       printf("did not find RENDER_SURFACE_STATE info\n");
    183       return;
    184    }
    185 
    186    start = surface_state_base + offset;
    187    pointers = gtt + start;
    188    for (i = 0; i < 16; i++) {
    189       if (pointers[i] == 0)
    190          continue;
    191       start = pointers[i] + surface_state_base;
    192       if (!valid_offset(start)) {
    193          printf("pointer %u: %08x <not valid>\n",
    194                 i, pointers[i]);
    195          continue;
    196       } else {
    197          printf("pointer %u: %08x\n", i, pointers[i]);
    198       }
    199 
    200       decode_structure(spec, surface_state, gtt + start);
    201    }
    202 }
    203 
    204 static void
    205 handle_3dstate_index_buffer(struct gen_spec *spec, uint32_t *p)
    206 {
    207    void *start;
    208    uint32_t length, i, type, size;
    209 
    210    start = gtt + p[2];
    211    type = (p[1] >> 8) & 3;
    212    size = 1 << type;
    213    length = p[4] / size;
    214    if (length > 10)
    215       length = 10;
    216 
    217    printf("\t");
    218 
    219    for (i = 0; i < length; i++) {
    220       switch (type) {
    221       case 0:
    222          printf("%3d ", ((uint8_t *)start)[i]);
    223          break;
    224       case 1:
    225          printf("%3d ", ((uint16_t *)start)[i]);
    226          break;
    227       case 2:
    228          printf("%3d ", ((uint32_t *)start)[i]);
    229          break;
    230       }
    231    }
    232    if (length < p[4] / size)
    233       printf("...\n");
    234    else
    235       printf("\n");
    236 }
    237 
    238 static inline uint64_t
    239 get_address(struct gen_spec *spec, uint32_t *p)
    240 {
    241    /* Addresses are always guaranteed to be page-aligned and sometimes
    242     * hardware packets have extra stuff stuffed in the bottom 12 bits.
    243     */
    244    uint64_t addr = p[0] & ~0xfffu;
    245 
    246    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
    247       /* On Broadwell and above, we have 48-bit addresses which consume two
    248        * dwords.  Some packets require that these get stored in a "canonical
    249        * form" which means that bit 47 is sign-extended through the upper
    250        * bits. In order to correctly handle those aub dumps, we need to mask
    251        * off the top 16 bits.
    252        */
    253       addr |= ((uint64_t)p[1] & 0xffff) << 32;
    254    }
    255 
    256    return addr;
    257 }
    258 
    259 static inline uint64_t
    260 get_offset(uint32_t *p, uint32_t start, uint32_t end)
    261 {
    262    assert(start <= end);
    263    assert(end < 64);
    264 
    265    uint64_t mask = (~0ull >> (64 - (end - start + 1))) << start;
    266 
    267    uint64_t offset = p[0];
    268    if (end >= 32)
    269       offset |= (uint64_t) p[1] << 32;
    270 
    271    return offset & mask;
    272 }
    273 
    274 static void
    275 handle_state_base_address(struct gen_spec *spec, uint32_t *p)
    276 {
    277    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
    278       if (p[1] & 1)
    279          general_state_base = get_address(spec, &p[1]);
    280       if (p[4] & 1)
    281          surface_state_base = get_address(spec, &p[4]);
    282       if (p[6] & 1)
    283          dynamic_state_base = get_address(spec, &p[6]);
    284       if (p[10] & 1)
    285          instruction_base = get_address(spec, &p[10]);
    286       if (p[15] & 1)
    287          instruction_bound = p[15] & 0xfff;
    288    } else {
    289       if (p[2] & 1)
    290          surface_state_base = get_address(spec, &p[2]);
    291       if (p[3] & 1)
    292          dynamic_state_base = get_address(spec, &p[3]);
    293       if (p[5] & 1)
    294          instruction_base = get_address(spec, &p[5]);
    295       if (p[9] & 1)
    296          instruction_bound = get_address(spec, &p[9]);
    297    }
    298 }
    299 
    300 static void
    301 dump_samplers(struct gen_spec *spec, uint32_t offset)
    302 {
    303    uint32_t i;
    304    uint64_t start;
    305    struct gen_group *sampler_state;
    306 
    307    sampler_state = gen_spec_find_struct(spec, "SAMPLER_STATE");
    308 
    309    start = dynamic_state_base + offset;
    310    for (i = 0; i < 4; i++) {
    311       printf("sampler state %d\n", i);
    312       decode_structure(spec, sampler_state, gtt + start + i * 16);
    313    }
    314 }
    315 
    316 static void
    317 handle_media_interface_descriptor_load(struct gen_spec *spec, uint32_t *p)
    318 {
    319    int i, length = p[2] / 32;
    320    struct gen_group *descriptor_structure;
    321    uint32_t *descriptors;
    322    uint64_t start;
    323    struct brw_instruction *insns;
    324 
    325    descriptor_structure =
    326       gen_spec_find_struct(spec, "INTERFACE_DESCRIPTOR_DATA");
    327    if (descriptor_structure == NULL) {
    328       printf("did not find INTERFACE_DESCRIPTOR_DATA info\n");
    329       return;
    330    }
    331 
    332    start = dynamic_state_base + p[3];
    333    descriptors = gtt + start;
    334    for (i = 0; i < length; i++, descriptors += 8) {
    335       printf("descriptor %u: %08x\n", i, *descriptors);
    336       decode_structure(spec, descriptor_structure, descriptors);
    337 
    338       start = instruction_base + descriptors[0];
    339       if (!valid_offset(start)) {
    340          printf("kernel: %08"PRIx64" <not valid>\n", start);
    341          continue;
    342       } else {
    343          printf("kernel: %08"PRIx64"\n", start);
    344       }
    345 
    346       insns = (struct brw_instruction *) (gtt + start);
    347       gen_disasm_disassemble(disasm, insns, 0, stdout);
    348 
    349       dump_samplers(spec, descriptors[3] & ~0x1f);
    350       dump_binding_table(spec, descriptors[4] & ~0x1f);
    351    }
    352 }
    353 
    354 /* Heuristic to determine whether a uint32_t is probably actually a float
    355  * (http://stackoverflow.com/a/2953466)
    356  */
    357 
    358 static bool
    359 probably_float(uint32_t bits)
    360 {
    361    int exp = ((bits & 0x7f800000U) >> 23) - 127;
    362    uint32_t mant = bits & 0x007fffff;
    363 
    364    /* +- 0.0 */
    365    if (exp == -127 && mant == 0)
    366       return true;
    367 
    368    /* +- 1 billionth to 1 billion */
    369    if (-30 <= exp && exp <= 30)
    370       return true;
    371 
    372    /* some value with only a few binary digits */
    373    if ((mant & 0x0000ffff) == 0)
    374       return true;
    375 
    376    return false;
    377 }
    378 
    379 static void
    380 handle_3dstate_vertex_buffers(struct gen_spec *spec, uint32_t *p)
    381 {
    382    uint32_t *end, *s, *dw, *dwend;
    383    uint64_t offset;
    384    int n, i, count, stride;
    385 
    386    end = (p[0] & 0xff) + p + 2;
    387    for (s = &p[1], n = 0; s < end; s += 4, n++) {
    388       if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
    389          offset = *(uint64_t *) &s[1];
    390          dwend = gtt + offset + s[3];
    391       } else {
    392          offset = s[1];
    393          dwend = gtt + s[2] + 1;
    394       }
    395 
    396       stride = field(s[0], 0, 11);
    397       count = 0;
    398       printf("vertex buffer %d, size %d\n", n, s[3]);
    399       for (dw = gtt + offset, i = 0; dw < dwend && i < 256; dw++) {
    400          if (count == 0 && count % (8 * 4) == 0)
    401             printf("  ");
    402 
    403          if (probably_float(*dw))
    404             printf("  %8.2f", *(float *) dw);
    405          else
    406             printf("  0x%08x", *dw);
    407 
    408          i++;
    409          count += 4;
    410 
    411          if (count == stride) {
    412             printf("\n");
    413             count = 0;
    414          } else if (count % (8 * 4) == 0) {
    415             printf("\n");
    416          } else {
    417             printf(" ");
    418          }
    419       }
    420       if (count > 0 && count % (8 * 4) != 0)
    421          printf("\n");
    422    }
    423 }
    424 
    425 static void
    426 handle_3dstate_vs(struct gen_spec *spec, uint32_t *p)
    427 {
    428    uint64_t start;
    429    struct brw_instruction *insns;
    430    int vs_enable;
    431 
    432    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
    433       start = get_offset(&p[1], 6, 63);
    434       vs_enable = p[7] & 1;
    435    } else {
    436       start = get_offset(&p[1], 6, 31);
    437       vs_enable = p[5] & 1;
    438    }
    439 
    440    if (vs_enable) {
    441       printf("instruction_base %08"PRIx64", start %08"PRIx64"\n",
    442              instruction_base, start);
    443 
    444       insns = (struct brw_instruction *) (gtt + instruction_base + start);
    445       gen_disasm_disassemble(disasm, insns, 0, stdout);
    446    }
    447 }
    448 
    449 static void
    450 handle_3dstate_hs(struct gen_spec *spec, uint32_t *p)
    451 {
    452    uint64_t start;
    453    struct brw_instruction *insns;
    454    int hs_enable;
    455 
    456    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
    457       start = get_offset(&p[3], 6, 63);
    458    } else {
    459       start = get_offset(&p[3], 6, 31);
    460    }
    461 
    462    hs_enable = p[2] & 0x80000000;
    463 
    464    if (hs_enable) {
    465       printf("instruction_base %08"PRIx64", start %08"PRIx64"\n",
    466              instruction_base, start);
    467 
    468       insns = (struct brw_instruction *) (gtt + instruction_base + start);
    469       gen_disasm_disassemble(disasm, insns, 0, stdout);
    470    }
    471 }
    472 
    473 static void
    474 handle_3dstate_constant(struct gen_spec *spec, uint32_t *p)
    475 {
    476    int i, j, length;
    477    uint32_t *dw;
    478    float *f;
    479 
    480    for (i = 0; i < 4; i++) {
    481       length = (p[1 + i / 2] >> (i & 1) * 16) & 0xffff;
    482       f = (float *) (gtt + p[3 + i * 2] + dynamic_state_base);
    483       dw = (uint32_t *) f;
    484       for (j = 0; j < length * 8; j++) {
    485          if (probably_float(dw[j]))
    486             printf("  %04.3f", f[j]);
    487          else
    488             printf("  0x%08x", dw[j]);
    489 
    490          if ((j & 7) == 7)
    491             printf("\n");
    492       }
    493    }
    494 }
    495 
    496 static void
    497 handle_3dstate_ps(struct gen_spec *spec, uint32_t *p)
    498 {
    499    uint32_t mask = ~((1 << 6) - 1);
    500    uint64_t start;
    501    struct brw_instruction *insns;
    502    static const char unused[] = "unused";
    503    static const char *pixel_type[3] = {"8 pixel", "16 pixel", "32 pixel"};
    504    const char *k0, *k1, *k2;
    505    uint32_t k_mask, k1_offset, k2_offset;
    506 
    507    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
    508       k_mask = p[6] & 7;
    509       k1_offset = 8;
    510       k2_offset = 10;
    511    } else {
    512       k_mask = p[4] & 7;
    513       k1_offset = 6;
    514       k2_offset = 7;
    515    }
    516 
    517 #define DISPATCH_8 1
    518 #define DISPATCH_16 2
    519 #define DISPATCH_32 4
    520 
    521    switch (k_mask) {
    522    case DISPATCH_8:
    523       k0 = pixel_type[0];
    524       k1 = unused;
    525       k2 = unused;
    526       break;
    527    case DISPATCH_16:
    528       k0 = pixel_type[1];
    529       k1 = unused;
    530       k2 = unused;
    531       break;
    532    case DISPATCH_8 | DISPATCH_16:
    533       k0 = pixel_type[0];
    534       k1 = unused;
    535       k2 = pixel_type[1];
    536       break;
    537    case DISPATCH_32:
    538       k0 = pixel_type[2];
    539       k1 = unused;
    540       k2 = unused;
    541       break;
    542    case DISPATCH_16 | DISPATCH_32:
    543       k0 = unused;
    544       k1 = pixel_type[2];
    545       k2 = pixel_type[1];
    546       break;
    547    case DISPATCH_8 | DISPATCH_16 | DISPATCH_32:
    548       k0 = pixel_type[0];
    549       k1 = pixel_type[2];
    550       k2 = pixel_type[1];
    551       break;
    552    default:
    553       k0 = unused;
    554       k1 = unused;
    555       k2 = unused;
    556       break;
    557    }
    558 
    559    start = instruction_base + (p[1] & mask);
    560    printf("  Kernel[0] %s\n", k0);
    561    if (k0 != unused) {
    562       insns = (struct brw_instruction *) (gtt + start);
    563       gen_disasm_disassemble(disasm, insns, 0, stdout);
    564    }
    565 
    566    start = instruction_base + (p[k1_offset] & mask);
    567    printf("  Kernel[1] %s\n", k1);
    568    if (k1 != unused) {
    569       insns = (struct brw_instruction *) (gtt + start);
    570       gen_disasm_disassemble(disasm, insns, 0, stdout);
    571    }
    572 
    573    start = instruction_base + (p[k2_offset] & mask);
    574    printf("  Kernel[2] %s\n", k2);
    575    if (k2 != unused) {
    576       insns = (struct brw_instruction *) (gtt + start);
    577       gen_disasm_disassemble(disasm, insns, 0, stdout);
    578    }
    579 }
    580 
    581 static void
    582 handle_3dstate_binding_table_pointers(struct gen_spec *spec, uint32_t *p)
    583 {
    584    dump_binding_table(spec, p[1]);
    585 }
    586 
    587 static void
    588 handle_3dstate_sampler_state_pointers(struct gen_spec *spec, uint32_t *p)
    589 {
    590    dump_samplers(spec, p[1]);
    591 }
    592 
    593 static void
    594 handle_3dstate_viewport_state_pointers_cc(struct gen_spec *spec, uint32_t *p)
    595 {
    596    uint64_t start;
    597    struct gen_group *cc_viewport;
    598 
    599    cc_viewport = gen_spec_find_struct(spec, "CC_VIEWPORT");
    600 
    601    start = dynamic_state_base + (p[1] & ~0x1fu);
    602    for (uint32_t i = 0; i < 4; i++) {
    603       printf("viewport %d\n", i);
    604       decode_structure(spec, cc_viewport, gtt + start + i * 8);
    605    }
    606 }
    607 
    608 static void
    609 handle_3dstate_viewport_state_pointers_sf_clip(struct gen_spec *spec,
    610                                                uint32_t *p)
    611 {
    612    uint64_t start;
    613    struct gen_group *sf_clip_viewport;
    614 
    615    sf_clip_viewport = gen_spec_find_struct(spec, "SF_CLIP_VIEWPORT");
    616 
    617    start = dynamic_state_base + (p[1] & ~0x3fu);
    618    for (uint32_t i = 0; i < 4; i++) {
    619       printf("viewport %d\n", i);
    620       decode_structure(spec, sf_clip_viewport, gtt + start + i * 64);
    621    }
    622 }
    623 
    624 static void
    625 handle_3dstate_blend_state_pointers(struct gen_spec *spec, uint32_t *p)
    626 {
    627    uint64_t start;
    628    struct gen_group *blend_state;
    629 
    630    blend_state = gen_spec_find_struct(spec, "BLEND_STATE");
    631 
    632    start = dynamic_state_base + (p[1] & ~0x3fu);
    633    decode_structure(spec, blend_state, gtt + start);
    634 }
    635 
    636 static void
    637 handle_3dstate_cc_state_pointers(struct gen_spec *spec, uint32_t *p)
    638 {
    639    uint64_t start;
    640    struct gen_group *cc_state;
    641 
    642    cc_state = gen_spec_find_struct(spec, "COLOR_CALC_STATE");
    643 
    644    start = dynamic_state_base + (p[1] & ~0x3fu);
    645    decode_structure(spec, cc_state, gtt + start);
    646 }
    647 
    648 static void
    649 handle_3dstate_scissor_state_pointers(struct gen_spec *spec, uint32_t *p)
    650 {
    651    uint64_t start;
    652    struct gen_group *scissor_rect;
    653 
    654    scissor_rect = gen_spec_find_struct(spec, "SCISSOR_RECT");
    655 
    656    start = dynamic_state_base + (p[1] & ~0x1fu);
    657    decode_structure(spec, scissor_rect, gtt + start);
    658 }
    659 
    660 static void
    661 handle_load_register_imm(struct gen_spec *spec, uint32_t *p)
    662 {
    663    struct gen_group *reg = gen_spec_find_register(spec, p[1]);
    664 
    665    if (reg != NULL) {
    666       printf("register %s (0x%x): 0x%x\n",
    667              reg->name, reg->register_offset, p[2]);
    668       decode_structure(spec, reg, &p[2]);
    669    }
    670 }
    671 
    672 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
    673 
    674 #define STATE_BASE_ADDRESS                  0x61010000
    675 
    676 #define MEDIA_INTERFACE_DESCRIPTOR_LOAD     0x70020000
    677 
    678 #define _3DSTATE_INDEX_BUFFER               0x780a0000
    679 #define _3DSTATE_VERTEX_BUFFERS             0x78080000
    680 
    681 #define _3DSTATE_VS                         0x78100000
    682 #define _3DSTATE_GS                         0x78110000
    683 #define _3DSTATE_HS                         0x781b0000
    684 #define _3DSTATE_DS                         0x781d0000
    685 
    686 #define _3DSTATE_CONSTANT_VS                0x78150000
    687 #define _3DSTATE_CONSTANT_GS                0x78160000
    688 #define _3DSTATE_CONSTANT_PS                0x78170000
    689 #define _3DSTATE_CONSTANT_HS                0x78190000
    690 #define _3DSTATE_CONSTANT_DS                0x781A0000
    691 
    692 #define _3DSTATE_PS                         0x78200000
    693 
    694 #define _3DSTATE_BINDING_TABLE_POINTERS_VS  0x78260000
    695 #define _3DSTATE_BINDING_TABLE_POINTERS_HS  0x78270000
    696 #define _3DSTATE_BINDING_TABLE_POINTERS_DS  0x78280000
    697 #define _3DSTATE_BINDING_TABLE_POINTERS_GS  0x78290000
    698 #define _3DSTATE_BINDING_TABLE_POINTERS_PS  0x782a0000
    699 
    700 #define _3DSTATE_SAMPLER_STATE_POINTERS_VS  0x782b0000
    701 #define _3DSTATE_SAMPLER_STATE_POINTERS_GS  0x782e0000
    702 #define _3DSTATE_SAMPLER_STATE_POINTERS_PS  0x782f0000
    703 
    704 #define _3DSTATE_VIEWPORT_STATE_POINTERS_CC 0x78230000
    705 #define _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP 0x78210000
    706 #define _3DSTATE_BLEND_STATE_POINTERS       0x78240000
    707 #define _3DSTATE_CC_STATE_POINTERS          0x780e0000
    708 #define _3DSTATE_SCISSOR_STATE_POINTERS     0x780f0000
    709 
    710 #define _MI_LOAD_REGISTER_IMM               0x11000000
    711 
    712 struct custom_handler {
    713    uint32_t opcode;
    714    void (*handle)(struct gen_spec *spec, uint32_t *p);
    715 } custom_handlers[] = {
    716    { STATE_BASE_ADDRESS, handle_state_base_address },
    717    { MEDIA_INTERFACE_DESCRIPTOR_LOAD, handle_media_interface_descriptor_load },
    718    { _3DSTATE_VERTEX_BUFFERS, handle_3dstate_vertex_buffers },
    719    { _3DSTATE_INDEX_BUFFER, handle_3dstate_index_buffer },
    720    { _3DSTATE_VS, handle_3dstate_vs },
    721    { _3DSTATE_GS, handle_3dstate_vs },
    722    { _3DSTATE_DS, handle_3dstate_vs },
    723    { _3DSTATE_HS, handle_3dstate_hs },
    724    { _3DSTATE_CONSTANT_VS, handle_3dstate_constant },
    725    { _3DSTATE_CONSTANT_GS, handle_3dstate_constant },
    726    { _3DSTATE_CONSTANT_PS, handle_3dstate_constant },
    727    { _3DSTATE_CONSTANT_HS, handle_3dstate_constant },
    728    { _3DSTATE_CONSTANT_DS, handle_3dstate_constant },
    729    { _3DSTATE_PS, handle_3dstate_ps },
    730 
    731    { _3DSTATE_BINDING_TABLE_POINTERS_VS, handle_3dstate_binding_table_pointers },
    732    { _3DSTATE_BINDING_TABLE_POINTERS_HS, handle_3dstate_binding_table_pointers },
    733    { _3DSTATE_BINDING_TABLE_POINTERS_DS, handle_3dstate_binding_table_pointers },
    734    { _3DSTATE_BINDING_TABLE_POINTERS_GS, handle_3dstate_binding_table_pointers },
    735    { _3DSTATE_BINDING_TABLE_POINTERS_PS, handle_3dstate_binding_table_pointers },
    736 
    737    { _3DSTATE_SAMPLER_STATE_POINTERS_VS, handle_3dstate_sampler_state_pointers },
    738    { _3DSTATE_SAMPLER_STATE_POINTERS_GS, handle_3dstate_sampler_state_pointers },
    739    { _3DSTATE_SAMPLER_STATE_POINTERS_PS, handle_3dstate_sampler_state_pointers },
    740 
    741    { _3DSTATE_VIEWPORT_STATE_POINTERS_CC, handle_3dstate_viewport_state_pointers_cc },
    742    { _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP, handle_3dstate_viewport_state_pointers_sf_clip },
    743    { _3DSTATE_BLEND_STATE_POINTERS, handle_3dstate_blend_state_pointers },
    744    { _3DSTATE_CC_STATE_POINTERS, handle_3dstate_cc_state_pointers },
    745    { _3DSTATE_SCISSOR_STATE_POINTERS, handle_3dstate_scissor_state_pointers },
    746    { _MI_LOAD_REGISTER_IMM, handle_load_register_imm }
    747 };
    748 
    749 static void
    750 parse_commands(struct gen_spec *spec, uint32_t *cmds, int size, int engine)
    751 {
    752    uint32_t *p, *end = cmds + size / 4;
    753    unsigned int length, i;
    754    struct gen_group *inst;
    755 
    756    for (p = cmds; p < end; p += length) {
    757       inst = gen_spec_find_instruction(spec, p);
    758       if (inst == NULL) {
    759          printf("unknown instruction %08x\n", p[0]);
    760          length = (p[0] & 0xff) + 2;
    761          continue;
    762       }
    763       length = gen_group_get_length(inst, p);
    764 
    765       const char *color, *reset_color = NORMAL;
    766       uint64_t offset;
    767 
    768       if (option_full_decode) {
    769          if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START ||
    770              (p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END)
    771             color = GREEN_HEADER;
    772          else
    773             color = BLUE_HEADER;
    774       } else
    775          color = NORMAL;
    776 
    777       if (option_color == COLOR_NEVER) {
    778          color = "";
    779          reset_color = "";
    780       }
    781 
    782       if (option_print_offsets)
    783          offset = (void *) p - gtt;
    784       else
    785          offset = 0;
    786 
    787       printf("%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
    788              color, offset, p[0],
    789              gen_group_get_name(inst), reset_color);
    790 
    791       if (option_full_decode) {
    792          struct gen_field_iterator iter;
    793          char *token = NULL;
    794          int idx = 0, dword_num = 0;
    795          gen_field_iterator_init(&iter, inst, p,
    796                                  option_color == COLOR_ALWAYS);
    797          while (gen_field_iterator_next(&iter)) {
    798             idx = 0;
    799             print_dword_val(&iter, offset, &dword_num);
    800             if (dword_num > 0)
    801                 token = print_iterator_values(&iter, &idx);
    802             if (token != NULL) {
    803                 printf("0x%08"PRIx64":  0x%08x : Dword %d\n",
    804                        offset + 4 * idx, p[idx], idx);
    805                 handle_struct_decode(spec,token, &p[idx]);
    806                 token = NULL;
    807             }
    808          }
    809 
    810          for (i = 0; i < ARRAY_LENGTH(custom_handlers); i++) {
    811             if (gen_group_get_opcode(inst) ==
    812                 custom_handlers[i].opcode)
    813                custom_handlers[i].handle(spec, p);
    814          }
    815       }
    816 
    817       if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START) {
    818          uint64_t start = get_address(spec, &p[1]);
    819 
    820          if (p[0] & (1 << 22)) {
    821             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
    822              * like a subroutine call.  Commands that come afterwards get
    823              * processed once the 2nd level batch buffer returns with
    824              * MI_BATCH_BUFFER_END.
    825              */
    826             parse_commands(spec, gtt + start, gtt_end - start, engine);
    827          } else {
    828             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
    829              * like a goto.  Nothing after it will ever get processed.  In
    830              * order to prevent the recursion from growing, we just reset the
    831              * loop and continue;
    832              */
    833             p = gtt + start;
    834             /* We don't know where secondaries end so use the GTT end */
    835             end = gtt + gtt_end;
    836             length = 0;
    837             continue;
    838          }
    839       } else if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END) {
    840          break;
    841       }
    842    }
    843 }
    844 
    845 #define GEN_ENGINE_RENDER 1
    846 #define GEN_ENGINE_BLITTER 2
    847 
    848 static void
    849 handle_trace_block(uint32_t *p)
    850 {
    851    int operation = p[1] & AUB_TRACE_OPERATION_MASK;
    852    int type = p[1] & AUB_TRACE_TYPE_MASK;
    853    int address_space = p[1] & AUB_TRACE_ADDRESS_SPACE_MASK;
    854    uint64_t offset = p[3];
    855    uint32_t size = p[4];
    856    int header_length = p[0] & 0xffff;
    857    uint32_t *data = p + header_length + 2;
    858    int engine = GEN_ENGINE_RENDER;
    859 
    860    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0))
    861       offset += (uint64_t) p[5] << 32;
    862 
    863    switch (operation) {
    864    case AUB_TRACE_OP_DATA_WRITE:
    865       if (address_space != AUB_TRACE_MEMTYPE_GTT)
    866          break;
    867       if (gtt_size < offset + size) {
    868          fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
    869          exit(EXIT_FAILURE);
    870       }
    871       memcpy((char *) gtt + offset, data, size);
    872       if (gtt_end < offset + size)
    873          gtt_end = offset + size;
    874       break;
    875    case AUB_TRACE_OP_COMMAND_WRITE:
    876       switch (type) {
    877       case AUB_TRACE_TYPE_RING_PRB0:
    878          engine = GEN_ENGINE_RENDER;
    879          break;
    880       case AUB_TRACE_TYPE_RING_PRB2:
    881          engine = GEN_ENGINE_BLITTER;
    882          break;
    883       default:
    884          printf("command write to unknown ring %d\n", type);
    885          break;
    886       }
    887 
    888       parse_commands(spec, data, size, engine);
    889       gtt_end = 0;
    890       break;
    891    }
    892 }
    893 
    894 static void
    895 handle_trace_header(uint32_t *p)
    896 {
    897    /* The intel_aubdump tool from IGT is kind enough to put a PCI-ID= tag in
    898     * the AUB header comment.  If the user hasn't specified a hardware
    899     * generation, try to use the one from the AUB file.
    900     */
    901    uint32_t *end = p + (p[0] & 0xffff) + 2;
    902    int aub_pci_id = 0;
    903    if (end > &p[12] && p[12] > 0)
    904       sscanf((char *)&p[13], "PCI-ID=%i", &aub_pci_id);
    905 
    906    if (pci_id == 0)
    907       pci_id = aub_pci_id;
    908 
    909    struct gen_device_info devinfo;
    910    if (!gen_get_device_info(pci_id, &devinfo)) {
    911       fprintf(stderr, "can't find device information: pci_id=0x%x\n", pci_id);
    912       exit(EXIT_FAILURE);
    913    }
    914 
    915    if (xml_path == NULL)
    916       spec = gen_spec_load(&devinfo);
    917    else
    918       spec = gen_spec_load_from_path(&devinfo, xml_path);
    919    disasm = gen_disasm_create(pci_id);
    920 
    921    if (spec == NULL || disasm == NULL)
    922       exit(EXIT_FAILURE);
    923 
    924    printf("%sAubinator: Intel AUB file decoder.%-80s%s\n",
    925           GREEN_HEADER, "", NORMAL);
    926 
    927    if (input_file)
    928       printf("File name:        %s\n", input_file);
    929 
    930    if (aub_pci_id)
    931       printf("PCI ID:           0x%x\n", aub_pci_id);
    932 
    933    char app_name[33];
    934    strncpy(app_name, (char *)&p[2], 32);
    935    app_name[32] = 0;
    936    printf("Application name: %s\n", app_name);
    937 
    938    printf("Decoding as:      %s\n", gen_get_device_name(pci_id));
    939 
    940    /* Throw in a new line before the first batch */
    941    printf("\n");
    942 }
    943 
    944 struct aub_file {
    945    FILE *stream;
    946 
    947    uint32_t *map, *end, *cursor;
    948    uint32_t *mem_end;
    949 };
    950 
    951 static struct aub_file *
    952 aub_file_open(const char *filename)
    953 {
    954    struct aub_file *file;
    955    struct stat sb;
    956    int fd;
    957 
    958    file = calloc(1, sizeof *file);
    959    fd = open(filename, O_RDONLY);
    960    if (fd == -1) {
    961       fprintf(stderr, "open %s failed: %s\n", filename, strerror(errno));
    962       exit(EXIT_FAILURE);
    963    }
    964 
    965    if (fstat(fd, &sb) == -1) {
    966       fprintf(stderr, "stat failed: %s\n", strerror(errno));
    967       exit(EXIT_FAILURE);
    968    }
    969 
    970    file->map = mmap(NULL, sb.st_size,
    971                     PROT_READ, MAP_SHARED, fd, 0);
    972    if (file->map == MAP_FAILED) {
    973       fprintf(stderr, "mmap failed: %s\n", strerror(errno));
    974       exit(EXIT_FAILURE);
    975    }
    976 
    977    file->cursor = file->map;
    978    file->end = file->map + sb.st_size / 4;
    979 
    980    return file;
    981 }
    982 
    983 static struct aub_file *
    984 aub_file_stdin(void)
    985 {
    986    struct aub_file *file;
    987 
    988    file = calloc(1, sizeof *file);
    989    file->stream = stdin;
    990 
    991    return file;
    992 }
    993 
    994 #define TYPE(dw)       (((dw) >> 29) & 7)
    995 #define OPCODE(dw)     (((dw) >> 23) & 0x3f)
    996 #define SUBOPCODE(dw)  (((dw) >> 16) & 0x7f)
    997 
    998 #define MAKE_HEADER(type, opcode, subopcode) \
    999                    (((type) << 29) | ((opcode) << 23) | ((subopcode) << 16))
   1000 
   1001 #define TYPE_AUB            0x7
   1002 
   1003 /* Classic AUB opcodes */
   1004 #define OPCODE_AUB          0x01
   1005 #define SUBOPCODE_HEADER    0x05
   1006 #define SUBOPCODE_BLOCK     0x41
   1007 #define SUBOPCODE_BMP       0x1e
   1008 
   1009 /* Newer version AUB opcode */
   1010 #define OPCODE_NEW_AUB      0x2e
   1011 #define SUBOPCODE_VERSION   0x00
   1012 #define SUBOPCODE_REG_WRITE 0x03
   1013 #define SUBOPCODE_MEM_POLL  0x05
   1014 #define SUBOPCODE_MEM_WRITE 0x06
   1015 
   1016 #define MAKE_GEN(major, minor) ( ((major) << 8) | (minor) )
   1017 
   1018 struct {
   1019    const char *name;
   1020    uint32_t gen;
   1021 } device_map[] = {
   1022    { "bwr", MAKE_GEN(4, 0) },
   1023    { "cln", MAKE_GEN(4, 0) },
   1024    { "blc", MAKE_GEN(4, 0) },
   1025    { "ctg", MAKE_GEN(4, 0) },
   1026    { "el", MAKE_GEN(4, 0) },
   1027    { "il", MAKE_GEN(4, 0) },
   1028    { "sbr", MAKE_GEN(6, 0) },
   1029    { "ivb", MAKE_GEN(7, 0) },
   1030    { "lrb2", MAKE_GEN(0, 0) },
   1031    { "hsw", MAKE_GEN(7, 5) },
   1032    { "vlv", MAKE_GEN(7, 0) },
   1033    { "bdw", MAKE_GEN(8, 0) },
   1034    { "skl", MAKE_GEN(9, 0) },
   1035    { "chv", MAKE_GEN(8, 0) },
   1036    { "bxt", MAKE_GEN(9, 0) }
   1037 };
   1038 
   1039 enum {
   1040    AUB_ITEM_DECODE_OK,
   1041    AUB_ITEM_DECODE_FAILED,
   1042    AUB_ITEM_DECODE_NEED_MORE_DATA,
   1043 };
   1044 
   1045 static int
   1046 aub_file_decode_batch(struct aub_file *file)
   1047 {
   1048    uint32_t *p, h, device, data_type, *new_cursor;
   1049    int header_length, bias;
   1050 
   1051    if (file->end - file->cursor < 1)
   1052       return AUB_ITEM_DECODE_NEED_MORE_DATA;
   1053 
   1054    p = file->cursor;
   1055    h = *p;
   1056    header_length = h & 0xffff;
   1057 
   1058    switch (OPCODE(h)) {
   1059    case OPCODE_AUB:
   1060       bias = 2;
   1061       break;
   1062    case OPCODE_NEW_AUB:
   1063       bias = 1;
   1064       break;
   1065    default:
   1066       printf("unknown opcode %d at %td/%td\n",
   1067              OPCODE(h), file->cursor - file->map,
   1068              file->end - file->map);
   1069       return AUB_ITEM_DECODE_FAILED;
   1070    }
   1071 
   1072    new_cursor = p + header_length + bias;
   1073    if ((h & 0xffff0000) == MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK)) {
   1074       if (file->end - file->cursor < 4)
   1075          return AUB_ITEM_DECODE_NEED_MORE_DATA;
   1076       new_cursor += p[4] / 4;
   1077    }
   1078 
   1079    if (new_cursor > file->end)
   1080       return AUB_ITEM_DECODE_NEED_MORE_DATA;
   1081 
   1082    switch (h & 0xffff0000) {
   1083    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_HEADER):
   1084       handle_trace_header(p);
   1085       break;
   1086    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK):
   1087       handle_trace_block(p);
   1088       break;
   1089    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BMP):
   1090       break;
   1091    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_VERSION):
   1092       printf("version block: dw1 %08x\n", p[1]);
   1093       device = (p[1] >> 8) & 0xff;
   1094       printf("  device %s\n", device_map[device].name);
   1095       break;
   1096    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_REG_WRITE):
   1097       printf("register write block: (dwords %d)\n", h & 0xffff);
   1098       printf("  reg 0x%x, data 0x%x\n", p[1], p[5]);
   1099       break;
   1100    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_WRITE):
   1101       printf("memory write block (dwords %d):\n", h & 0xffff);
   1102       printf("  address 0x%"PRIx64"\n", *(uint64_t *) &p[1]);
   1103       data_type = (p[3] >> 20) & 0xff;
   1104       if (data_type != 0)
   1105          printf("  data type 0x%x\n", data_type);
   1106       printf("  address space 0x%x\n", (p[3] >> 28) & 0xf);
   1107       break;
   1108    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_POLL):
   1109       printf("memory poll block (dwords %d):\n", h & 0xffff);
   1110       break;
   1111    default:
   1112       printf("unknown block type=0x%x, opcode=0x%x, "
   1113              "subopcode=0x%x (%08x)\n", TYPE(h), OPCODE(h), SUBOPCODE(h), h);
   1114       break;
   1115    }
   1116    file->cursor = new_cursor;
   1117 
   1118    return AUB_ITEM_DECODE_OK;
   1119 }
   1120 
   1121 static int
   1122 aub_file_more_stuff(struct aub_file *file)
   1123 {
   1124    return file->cursor < file->end || (file->stream && !feof(file->stream));
   1125 }
   1126 
   1127 #define AUB_READ_BUFFER_SIZE (4096)
   1128 #define MAX(a, b) ((a) < (b) ? (b) : (a))
   1129 
   1130 static void
   1131 aub_file_data_grow(struct aub_file *file)
   1132 {
   1133    size_t old_size = (file->mem_end - file->map) * 4;
   1134    size_t new_size = MAX(old_size * 2, AUB_READ_BUFFER_SIZE);
   1135    uint32_t *new_start = realloc(file->map, new_size);
   1136 
   1137    file->cursor = new_start + (file->cursor - file->map);
   1138    file->end = new_start + (file->end - file->map);
   1139    file->map = new_start;
   1140    file->mem_end = file->map + (new_size / 4);
   1141 }
   1142 
   1143 static bool
   1144 aub_file_data_load(struct aub_file *file)
   1145 {
   1146    size_t r;
   1147 
   1148    if (file->stream == NULL)
   1149       return false;
   1150 
   1151    /* First remove any consumed data */
   1152    if (file->cursor > file->map) {
   1153       memmove(file->map, file->cursor,
   1154               (file->end - file->cursor) * 4);
   1155       file->end -= file->cursor - file->map;
   1156       file->cursor = file->map;
   1157    }
   1158 
   1159    /* Then load some new data in */
   1160    if ((file->mem_end - file->end) < (AUB_READ_BUFFER_SIZE / 4))
   1161       aub_file_data_grow(file);
   1162 
   1163    r = fread(file->end, 1, (file->mem_end - file->end) * 4, file->stream);
   1164    file->end += r / 4;
   1165 
   1166    return r != 0;
   1167 }
   1168 
   1169 static void
   1170 setup_pager(void)
   1171 {
   1172    int fds[2];
   1173    pid_t pid;
   1174 
   1175    if (!isatty(1))
   1176       return;
   1177 
   1178    if (pipe(fds) == -1)
   1179       return;
   1180 
   1181    pid = fork();
   1182    if (pid == -1)
   1183       return;
   1184 
   1185    if (pid == 0) {
   1186       close(fds[1]);
   1187       dup2(fds[0], 0);
   1188       execlp("less", "less", "-FRSi", NULL);
   1189    }
   1190 
   1191    close(fds[0]);
   1192    dup2(fds[1], 1);
   1193    close(fds[1]);
   1194 }
   1195 
   1196 static void
   1197 print_help(const char *progname, FILE *file)
   1198 {
   1199    fprintf(file,
   1200            "Usage: %s [OPTION]... [FILE]\n"
   1201            "Decode aub file contents from either FILE or the standard input.\n\n"
   1202            "A valid --gen option must be provided.\n\n"
   1203            "      --help          display this help and exit\n"
   1204            "      --gen=platform  decode for given platform (ivb, byt, hsw, bdw, chv, skl, kbl or bxt)\n"
   1205            "      --headers       decode only command headers\n"
   1206            "      --color[=WHEN]  colorize the output; WHEN can be 'auto' (default\n"
   1207            "                        if omitted), 'always', or 'never'\n"
   1208            "      --no-pager      don't launch pager\n"
   1209            "      --no-offsets    don't print instruction offsets\n"
   1210            "      --xml=DIR       load hardware xml description from directory DIR\n",
   1211            progname);
   1212 }
   1213 
   1214 int main(int argc, char *argv[])
   1215 {
   1216    struct aub_file *file;
   1217    int c, i;
   1218    bool help = false, pager = true;
   1219    const struct {
   1220       const char *name;
   1221       int pci_id;
   1222    } gens[] = {
   1223       { "ivb", 0x0166 }, /* Intel(R) Ivybridge Mobile GT2 */
   1224       { "hsw", 0x0416 }, /* Intel(R) Haswell Mobile GT2 */
   1225       { "byt", 0x0155 }, /* Intel(R) Bay Trail */
   1226       { "bdw", 0x1616 }, /* Intel(R) HD Graphics 5500 (Broadwell GT2) */
   1227       { "chv", 0x22B3 }, /* Intel(R) HD Graphics (Cherryview) */
   1228       { "skl", 0x1912 }, /* Intel(R) HD Graphics 530 (Skylake GT2) */
   1229       { "kbl", 0x591D }, /* Intel(R) Kabylake GT2 */
   1230       { "bxt", 0x0A84 }  /* Intel(R) HD Graphics (Broxton) */
   1231    };
   1232    const struct option aubinator_opts[] = {
   1233       { "help",       no_argument,       (int *) &help,                 true },
   1234       { "no-pager",   no_argument,       (int *) &pager,                false },
   1235       { "no-offsets", no_argument,       (int *) &option_print_offsets, false },
   1236       { "gen",        required_argument, NULL,                          'g' },
   1237       { "headers",    no_argument,       (int *) &option_full_decode,   false },
   1238       { "color",      required_argument, NULL,                          'c' },
   1239       { "xml",        required_argument, NULL,                          'x' },
   1240       { NULL,         0,                 NULL,                          0 }
   1241    };
   1242 
   1243    i = 0;
   1244    while ((c = getopt_long(argc, argv, "", aubinator_opts, &i)) != -1) {
   1245       switch (c) {
   1246       case 'g':
   1247          for (i = 0; i < ARRAY_SIZE(gens); i++) {
   1248             if (!strcmp(optarg, gens[i].name)) {
   1249                pci_id = gens[i].pci_id;
   1250                break;
   1251             }
   1252          }
   1253          if (i == ARRAY_SIZE(gens)) {
   1254             fprintf(stderr, "can't parse gen: '%s', expected ivb, byt, hsw, "
   1255                                    "bdw, chv, skl, kbl or bxt\n", optarg);
   1256             exit(EXIT_FAILURE);
   1257          }
   1258          break;
   1259       case 'c':
   1260          if (optarg == NULL || strcmp(optarg, "always") == 0)
   1261             option_color = COLOR_ALWAYS;
   1262          else if (strcmp(optarg, "never") == 0)
   1263             option_color = COLOR_NEVER;
   1264          else if (strcmp(optarg, "auto") == 0)
   1265             option_color = COLOR_AUTO;
   1266          else {
   1267             fprintf(stderr, "invalid value for --color: %s", optarg);
   1268             exit(EXIT_FAILURE);
   1269          }
   1270          break;
   1271       case 'x':
   1272          xml_path = strdup(optarg);
   1273          break;
   1274       default:
   1275          break;
   1276       }
   1277    }
   1278 
   1279    if (help || argc == 1) {
   1280       print_help(argv[0], stderr);
   1281       exit(0);
   1282    }
   1283 
   1284    if (optind < argc)
   1285       input_file = argv[optind];
   1286 
   1287    /* Do this before we redirect stdout to pager. */
   1288    if (option_color == COLOR_AUTO)
   1289       option_color = isatty(1) ? COLOR_ALWAYS : COLOR_NEVER;
   1290 
   1291    if (isatty(1) && pager)
   1292       setup_pager();
   1293 
   1294    if (input_file == NULL)
   1295       file = aub_file_stdin();
   1296    else
   1297       file = aub_file_open(input_file);
   1298 
   1299    /* mmap a terabyte for our gtt space. */
   1300    gtt_size = 1ull << 40;
   1301    gtt = mmap(NULL, gtt_size, PROT_READ | PROT_WRITE,
   1302               MAP_PRIVATE | MAP_ANONYMOUS |  MAP_NORESERVE, -1, 0);
   1303    if (gtt == MAP_FAILED) {
   1304       fprintf(stderr, "failed to alloc gtt space: %s\n", strerror(errno));
   1305       exit(EXIT_FAILURE);
   1306    }
   1307 
   1308    while (aub_file_more_stuff(file)) {
   1309       switch (aub_file_decode_batch(file)) {
   1310       case AUB_ITEM_DECODE_OK:
   1311          break;
   1312       case AUB_ITEM_DECODE_NEED_MORE_DATA:
   1313          if (!file->stream) {
   1314             file->cursor = file->end;
   1315             break;
   1316          }
   1317          if (aub_file_more_stuff(file) && !aub_file_data_load(file)) {
   1318             fprintf(stderr, "failed to load data from stdin\n");
   1319             exit(EXIT_FAILURE);
   1320          }
   1321          break;
   1322       default:
   1323          fprintf(stderr, "failed to parse aubdump data\n");
   1324          exit(EXIT_FAILURE);
   1325       }
   1326    }
   1327 
   1328 
   1329    fflush(stdout);
   1330    /* close the stdout which is opened to write the output */
   1331    close(1);
   1332    free(xml_path);
   1333 
   1334    wait(NULL);
   1335 
   1336    return EXIT_SUCCESS;
   1337 }
   1338