src/intel/tools/aubinator.c

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