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
  28 #include <unistd.h>
  29 #include <fcntl.h>
  30 #include <string.h>
  31 #include <signal.h>
  32 #include <errno.h>
  33 #include <error.h>
  34 #include <sys/types.h>
  35 #include <sys/stat.h>
  36 #include <sys/wait.h>
  37 #include <sys/mman.h>
  38
  39 #include "decoder.h"
  40 #include "intel_aub.h"
  41 #include "gen_disasm.h"
  42
  43 /* Below is the only command missing from intel_aub.h in libdrm
  44  * So, reuse intel_aub.h from libdrm and #define the
  45  * AUB_MI_BATCH_BUFFER_END as below
  46  */
  47 #define AUB_MI_BATCH_BUFFER_END (0x0500 << 16)
  48
  49 #define CSI "\e["
  50 #define HEADER CSI "37;44m"
  51 #define NORMAL CSI "0m"
  52 #define CLEAR_TO_EOL CSI "0K"
  53
  54 /* options */
  55
  56 static bool option_full_decode = true;
  57 static bool option_print_offsets = true;
  58 static enum { COLOR_AUTO, COLOR_ALWAYS, COLOR_NEVER } option_color;
  59
  60 /* state */
  61
  62 struct gen_disasm *disasm;
  63
  64 uint64_t gtt_size, gtt_end;
  65 void *gtt;
  66 uint64_t general_state_base;
  67 uint64_t surface_state_base;
  68 uint64_t dynamic_state_base;
  69 uint64_t instruction_base;
  70 uint64_t instruction_bound;
  71
  72 static inline uint32_t
  73 field(uint32_t value, int start, int end)
  74 {
  75    uint32_t mask;
  76
  77    mask = ~0U >> (31 - end + start);
  78
  79    return (value >> start) & mask;
  80 }
  81
  82 struct brw_instruction;
  83
  84 static inline int
  85 valid_offset(uint32_t offset)
  86 {
  87    return offset < gtt_end;
  88 }
  89
  90 static void
  91 print_dword_val(struct gen_field_iterator *iter, uint64_t offset,
  92                 int *dword_num)
  93 {
  94    struct gen_field *f;
  95    union {
  96       uint32_t dw;
  97       float f;
  98    } v;
  99
 100    f = iter->group->fields[iter->i - 1];
 101    v.dw = iter->p[f->start / 32];
 102
 103    if (*dword_num != (f->start / 32)) {
 104       printf("0x%08lx:  0x%08x : Dword %d\n",
 105              offset + 4 * (f->start / 32), v.dw, f->start / 32);
 106       *dword_num = (f->start / 32);
 107    }
 108 }
 109
 110 static char *
 111 print_iterator_values(struct gen_field_iterator *iter, int *idx)
 112 {
 113     char *token = NULL;
 114     if (strstr(iter->value, "struct") == NULL) {
 115         printf("    %s: %s\n", iter->name, iter->value);
 116     } else {
 117         token = strtok(iter->value, " ");
 118         if (token != NULL) {
 119             token = strtok(NULL, " ");
 120             *idx = atoi(strtok(NULL, ">"));
 121         } else {
 122             token = NULL;
 123         }
 124         printf("    %s:<struct %s>\n", iter->name, token);
 125     }
 126     return token;
 127 }
 128
 129 static void
 130 decode_structure(struct gen_spec *spec, struct gen_group *strct,
 131                  const uint32_t *p)
 132 {
 133    struct gen_field_iterator iter;
 134    char *token = NULL;
 135    int idx = 0, dword_num = 0;
 136    uint64_t offset = 0;
 137
 138    if (option_print_offsets)
 139       offset = (void *) p - gtt;
 140    else
 141       offset = 0;
 142
 143    gen_field_iterator_init(&iter, strct, p);
 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: %08lx <not valid>\n", start);
 305          continue;
 306       } else {
 307          printf("kernel: %08lx\n", start);
 308       }
 309
 310       insns = (struct brw_instruction *) (gtt + start);
 311       gen_disasm_disassemble(disasm, insns, 0, 8192, 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 %08lx, start %08lx\n",
 406              instruction_base, start);
 407
 408       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 409       gen_disasm_disassemble(disasm, insns, 0, 8192, 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 %08lx, start %08lx\n",
 430              instruction_base, start);
 431
 432       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 433       gen_disasm_disassemble(disasm, insns, 0, 8192, 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, 8192, 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, 8192, 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, 8192, 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 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
 625
 626 #define STATE_BASE_ADDRESS                  0x61010000
 627
 628 #define MEDIA_INTERFACE_DESCRIPTOR_LOAD     0x70020000
 629
 630 #define _3DSTATE_INDEX_BUFFER               0x780a0000
 631 #define _3DSTATE_VERTEX_BUFFERS             0x78080000
 632
 633 #define _3DSTATE_VS                         0x78100000
 634 #define _3DSTATE_GS                         0x78110000
 635 #define _3DSTATE_HS                         0x781b0000
 636 #define _3DSTATE_DS                         0x781d0000
 637
 638 #define _3DSTATE_CONSTANT_VS                0x78150000
 639 #define _3DSTATE_CONSTANT_GS                0x78160000
 640 #define _3DSTATE_CONSTANT_PS                0x78170000
 641 #define _3DSTATE_CONSTANT_HS                0x78190000
 642 #define _3DSTATE_CONSTANT_DS                0x781A0000
 643
 644 #define _3DSTATE_PS                         0x78200000
 645
 646 #define _3DSTATE_BINDING_TABLE_POINTERS_VS  0x78260000
 647 #define _3DSTATE_BINDING_TABLE_POINTERS_HS  0x78270000
 648 #define _3DSTATE_BINDING_TABLE_POINTERS_DS  0x78280000
 649 #define _3DSTATE_BINDING_TABLE_POINTERS_GS  0x78290000
 650 #define _3DSTATE_BINDING_TABLE_POINTERS_PS  0x782a0000
 651
 652 #define _3DSTATE_SAMPLER_STATE_POINTERS_VS  0x782b0000
 653 #define _3DSTATE_SAMPLER_STATE_POINTERS_GS  0x782e0000
 654 #define _3DSTATE_SAMPLER_STATE_POINTERS_PS  0x782f0000
 655
 656 #define _3DSTATE_VIEWPORT_STATE_POINTERS_CC 0x78230000
 657 #define _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP 0x78210000
 658 #define _3DSTATE_BLEND_STATE_POINTERS       0x78240000
 659 #define _3DSTATE_CC_STATE_POINTERS          0x780e0000
 660 #define _3DSTATE_SCISSOR_STATE_POINTERS     0x780f0000
 661
 662 struct custom_handler {
 663    uint32_t opcode;
 664    void (*handle)(struct gen_spec *spec, uint32_t *p);
 665 } custom_handlers[] = {
 666    { STATE_BASE_ADDRESS, handle_state_base_address },
 667    { MEDIA_INTERFACE_DESCRIPTOR_LOAD, handle_media_interface_descriptor_load },
 668    { _3DSTATE_VERTEX_BUFFERS, handle_3dstate_vertex_buffers },
 669    { _3DSTATE_INDEX_BUFFER, handle_3dstate_index_buffer },
 670    { _3DSTATE_VS, handle_3dstate_vs },
 671    { _3DSTATE_GS, handle_3dstate_vs },
 672    { _3DSTATE_DS, handle_3dstate_vs },
 673    { _3DSTATE_HS, handle_3dstate_hs },
 674    { _3DSTATE_CONSTANT_VS, handle_3dstate_constant },
 675    { _3DSTATE_CONSTANT_GS, handle_3dstate_constant },
 676    { _3DSTATE_CONSTANT_PS, handle_3dstate_constant },
 677    { _3DSTATE_CONSTANT_HS, handle_3dstate_constant },
 678    { _3DSTATE_CONSTANT_DS, handle_3dstate_constant },
 679    { _3DSTATE_PS, handle_3dstate_ps },
 680
 681    { _3DSTATE_BINDING_TABLE_POINTERS_VS, handle_3dstate_binding_table_pointers },
 682    { _3DSTATE_BINDING_TABLE_POINTERS_HS, handle_3dstate_binding_table_pointers },
 683    { _3DSTATE_BINDING_TABLE_POINTERS_DS, handle_3dstate_binding_table_pointers },
 684    { _3DSTATE_BINDING_TABLE_POINTERS_GS, handle_3dstate_binding_table_pointers },
 685    { _3DSTATE_BINDING_TABLE_POINTERS_PS, handle_3dstate_binding_table_pointers },
 686
 687    { _3DSTATE_SAMPLER_STATE_POINTERS_VS, handle_3dstate_sampler_state_pointers },
 688    { _3DSTATE_SAMPLER_STATE_POINTERS_GS, handle_3dstate_sampler_state_pointers },
 689    { _3DSTATE_SAMPLER_STATE_POINTERS_PS, handle_3dstate_sampler_state_pointers },
 690
 691    { _3DSTATE_VIEWPORT_STATE_POINTERS_CC, handle_3dstate_viewport_state_pointers_cc },
 692    { _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP, handle_3dstate_viewport_state_pointers_sf_clip },
 693    { _3DSTATE_BLEND_STATE_POINTERS, handle_3dstate_blend_state_pointers },
 694    { _3DSTATE_CC_STATE_POINTERS, handle_3dstate_cc_state_pointers },
 695    { _3DSTATE_SCISSOR_STATE_POINTERS, handle_3dstate_scissor_state_pointers }
 696 };
 697
 698 static void
 699 parse_commands(struct gen_spec *spec, uint32_t *cmds, int size, int engine)
 700 {
 701    uint32_t *p, *end = cmds + size / 4;
 702    unsigned int length, i;
 703    struct gen_group *inst;
 704
 705    for (p = cmds; p < end; p += length) {
 706       inst = gen_spec_find_instruction(spec, p);
 707       if (inst == NULL) {
 708          printf("unknown instruction %08x\n", p[0]);
 709          length = (p[0] & 0xff) + 2;
 710          continue;
 711       }
 712       length = gen_group_get_length(inst, p);
 713
 714       const char *color, *reset_color = CLEAR_TO_EOL NORMAL;
 715       uint64_t offset;
 716
 717       if (option_full_decode)
 718          color = HEADER;
 719       else
 720          color = NORMAL;
 721
 722       if (option_color == COLOR_NEVER) {
 723          color = "";
 724          reset_color = "";
 725       }
 726
 727       if (option_print_offsets)
 728          offset = (void *) p - gtt;
 729       else
 730          offset = 0;
 731
 732       printf("%s0x%08lx:  0x%08x:  %s%s\n",
 733              color, offset, p[0],
 734              gen_group_get_name(inst), reset_color);
 735
 736       if (option_full_decode) {
 737          struct gen_field_iterator iter;
 738          char *token = NULL;
 739          int idx = 0, dword_num = 0;
 740          gen_field_iterator_init(&iter, inst, p);
 741          while (gen_field_iterator_next(&iter)) {
 742             idx = 0;
 743             print_dword_val(&iter, offset, &dword_num);
 744             if (dword_num > 0)
 745                 token = print_iterator_values(&iter, &idx);
 746             if (token != NULL) {
 747                 printf("0x%08lx:  0x%08x : Dword %d\n",
 748                        offset + 4 * idx, p[idx], idx);
 749                 handle_struct_decode(spec,token, &p[idx]);
 750                 token = NULL;
 751             }
 752          }
 753
 754          for (i = 0; i < ARRAY_LENGTH(custom_handlers); i++) {
 755             if (gen_group_get_opcode(inst) ==
 756                 custom_handlers[i].opcode)
 757                custom_handlers[i].handle(spec, p);
 758          }
 759       }
 760
 761       if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START) {
 762          uint64_t start;
 763          if (gen_spec_get_gen(spec) >= gen_make_gen(8,0))
 764             start = get_qword(&p[1]);
 765          else
 766             start = p[1];
 767
 768          parse_commands(spec, gtt + start, 1 << 20, engine);
 769       } else if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END) {
 770          break;
 771       }
 772    }
 773 }
 774
 775 #define GEN_ENGINE_RENDER 1
 776 #define GEN_ENGINE_BLITTER 2
 777
 778 static void
 779 handle_trace_block(struct gen_spec *spec, uint32_t *p)
 780 {
 781    int operation = p[1] & AUB_TRACE_OPERATION_MASK;
 782    int type = p[1] & AUB_TRACE_TYPE_MASK;
 783    int address_space = p[1] & AUB_TRACE_ADDRESS_SPACE_MASK;
 784    uint64_t offset = p[3];
 785    uint32_t size = p[4];
 786    int header_length = p[0] & 0xffff;
 787    uint32_t *data = p + header_length + 2;
 788    int engine = GEN_ENGINE_RENDER;
 789
 790    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0))
 791       offset += (uint64_t) p[5] << 32;
 792
 793    switch (operation) {
 794    case AUB_TRACE_OP_DATA_WRITE:
 795       if (address_space != AUB_TRACE_MEMTYPE_GTT)
 796          break;
 797       if (gtt_size < offset + size)
 798          error(EXIT_FAILURE, errno, "overflow gtt space");
 799       memcpy((char *) gtt + offset, data, size);
 800       if (gtt_end < offset + size)
 801          gtt_end = offset + size;
 802       break;
 803    case AUB_TRACE_OP_COMMAND_WRITE:
 804       switch (type) {
 805       case AUB_TRACE_TYPE_RING_PRB0:
 806          engine = GEN_ENGINE_RENDER;
 807          break;
 808       case AUB_TRACE_TYPE_RING_PRB2:
 809          engine = GEN_ENGINE_BLITTER;
 810          break;
 811       default:
 812          printf("command write to unknown ring %d\n", type);
 813          break;
 814       }
 815
 816       parse_commands(spec, data, size, engine);
 817       gtt_end = 0;
 818       break;
 819    }
 820 }
 821
 822 struct aub_file {
 823    char *filename;
 824    int fd;
 825    struct stat sb;
 826    uint32_t *map, *end, *cursor;
 827 };
 828
 829 static struct aub_file *
 830 aub_file_open(const char *filename)
 831 {
 832    struct aub_file *file;
 833
 834    file = malloc(sizeof *file);
 835    file->filename = strdup(filename);
 836    file->fd = open(file->filename, O_RDONLY);
 837    if (file->fd == -1)
 838       error(EXIT_FAILURE, errno, "open %s failed", file->filename);
 839
 840    if (fstat(file->fd, &file->sb) == -1)
 841       error(EXIT_FAILURE, errno, "stat failed");
 842
 843    file->map = mmap(NULL, file->sb.st_size,
 844                     PROT_READ, MAP_SHARED, file->fd, 0);
 845    if (file->map == MAP_FAILED)
 846       error(EXIT_FAILURE, errno, "mmap failed");
 847
 848    file->cursor = file->map;
 849    file->end = file->map + file->sb.st_size / 4;
 850
 851    /* mmap a terabyte for our gtt space. */
 852    gtt_size = 1ul << 40;
 853    gtt = mmap(NULL, gtt_size, PROT_READ | PROT_WRITE,
 854               MAP_PRIVATE | MAP_ANONYMOUS |  MAP_NORESERVE, -1, 0);
 855    if (gtt == MAP_FAILED)
 856       error(EXIT_FAILURE, errno, "failed to alloc gtt space");
 857
 858    return file;
 859 }
 860
 861 #define TYPE(dw)       (((dw) >> 29) & 7)
 862 #define OPCODE(dw)     (((dw) >> 23) & 0x3f)
 863 #define SUBOPCODE(dw)  (((dw) >> 16) & 0x7f)
 864
 865 #define MAKE_HEADER(type, opcode, subopcode) \
 866                    (((type) << 29) | ((opcode) << 23) | ((subopcode) << 16))
 867
 868 #define TYPE_AUB            0x7
 869
 870 /* Classic AUB opcodes */
 871 #define OPCODE_AUB          0x01
 872 #define SUBOPCODE_HEADER    0x05
 873 #define SUBOPCODE_BLOCK     0x41
 874 #define SUBOPCODE_BMP       0x1e
 875
 876 /* Newer version AUB opcode */
 877 #define OPCODE_NEW_AUB      0x2e
 878 #define SUBOPCODE_VERSION   0x00
 879 #define SUBOPCODE_REG_WRITE 0x03
 880 #define SUBOPCODE_MEM_POLL  0x05
 881 #define SUBOPCODE_MEM_WRITE 0x06
 882
 883 #define MAKE_GEN(major, minor) ( ((major) << 8) | (minor) )
 884
 885 struct {
 886    const char *name;
 887    uint32_t gen;
 888 } device_map[] = {
 889    { "bwr", MAKE_GEN(4, 0) },
 890    { "cln", MAKE_GEN(4, 0) },
 891    { "blc", MAKE_GEN(4, 0) },
 892    { "ctg", MAKE_GEN(4, 0) },
 893    { "el", MAKE_GEN(4, 0) },
 894    { "il", MAKE_GEN(4, 0) },
 895    { "sbr", MAKE_GEN(6, 0) },
 896    { "ivb", MAKE_GEN(7, 0) },
 897    { "lrb2", MAKE_GEN(0, 0) },
 898    { "hsw", MAKE_GEN(7, 5) },
 899    { "vlv", MAKE_GEN(7, 0) },
 900    { "bdw", MAKE_GEN(8, 0) },
 901    { "skl", MAKE_GEN(9, 0) },
 902    { "chv", MAKE_GEN(8, 0) },
 903    { "bxt", MAKE_GEN(9, 0) }
 904 };
 905
 906 static void
 907 aub_file_decode_batch(struct aub_file *file, struct gen_spec *spec)
 908 {
 909    uint32_t *p, h, device, data_type;
 910    int header_length, payload_size, bias;
 911
 912    p = file->cursor;
 913    h = *p;
 914    header_length = h & 0xffff;
 915
 916    switch (OPCODE(h)) {
 917    case OPCODE_AUB:
 918       bias = 2;
 919       break;
 920    case OPCODE_NEW_AUB:
 921       bias = 1;
 922       break;
 923    default:
 924       printf("unknown opcode %d at %ld/%ld\n",
 925              OPCODE(h), file->cursor - file->map,
 926              file->end - file->map);
 927       file->cursor = file->end;
 928       return;
 929    }
 930
 931    payload_size = 0;
 932    switch (h & 0xffff0000) {
 933    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_HEADER):
 934       payload_size = p[12];
 935       break;
 936    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK):
 937       payload_size = p[4];
 938       handle_trace_block(spec, p);
 939       break;
 940    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BMP):
 941       break;
 942
 943    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_VERSION):
 944       printf("version block: dw1 %08x\n", p[1]);
 945       device = (p[1] >> 8) & 0xff;
 946       printf("  device %s\n", device_map[device].name);
 947       break;
 948    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_REG_WRITE):
 949       printf("register write block: (dwords %d)\n", h & 0xffff);
 950       printf("  reg 0x%x, data 0x%x\n", p[1], p[5]);
 951       break;
 952    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_WRITE):
 953       printf("memory write block (dwords %d):\n", h & 0xffff);
 954       printf("  address 0x%lx\n", *(uint64_t *) &p[1]);
 955       data_type = (p[3] >> 20) & 0xff;
 956       if (data_type != 0)
 957          printf("  data type 0x%x\n", data_type);
 958       printf("  address space 0x%x\n", (p[3] >> 28) & 0xf);
 959       break;
 960    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_POLL):
 961       printf("memory poll block (dwords %d):\n", h & 0xffff);
 962       break;
 963    default:
 964       printf("unknown block type=0x%x, opcode=0x%x, "
 965              "subopcode=0x%x (%08x)\n", TYPE(h), OPCODE(h), SUBOPCODE(h), h);
 966       break;
 967    }
 968    file->cursor = p + header_length + bias + payload_size / 4;
 969 }
 970
 971 static int
 972 aub_file_more_stuff(struct aub_file *file)
 973 {
 974    return file->cursor < file->end;
 975 }
 976
 977 static void
 978 setup_pager(void)
 979 {
 980    int fds[2];
 981    pid_t pid;
 982
 983    if (!isatty(1))
 984       return;
 985
 986    if (pipe(fds) == -1)
 987       return;
 988
 989    pid = fork();
 990    if (pid == -1)
 991       return;
 992
 993    if (pid == 0) {
 994       close(fds[1]);
 995       dup2(fds[0], 0);
 996       execlp("less", "less", "-rFi", NULL);
 997    }
 998
 999    close(fds[0]);
1000    dup2(fds[1], 1);
1001    close(fds[1]);
1002 }
1003
1004 static void
1005 print_help(FILE *file)
1006 {
1007    fprintf(file,
1008            "Usage: %s [OPTION]... FILE\n"
1009            "Decode aub file contents.\n\n"
1010            "A valid --gen option must be provided.\n\n"
1011            "      --help          display this help and exit\n"
1012            "      --gen=platform  decode for given platform (ivb, byt, hsw, bdw, chv, skl, kbl or bxt)\n"
1013            "      --headers       decode only command headers\n"
1014            "      --color[=WHEN]  colorize the output; WHEN can be 'auto' (default\n"
1015            "                        if omitted), 'always', or 'never'\n"
1016            "      --no-pager      don't launch pager\n"
1017            "      --no-offsets    don't print instruction offsets\n",
1018            basename(program_invocation_name));
1019 }
1020
1021 static bool
1022 is_prefix(const char *arg, const char *prefix, const char **value)
1023 {
1024    int l = strlen(prefix);
1025
1026    if (strncmp(arg, prefix, l) == 0 && (arg[l] == '\0' || arg[l] == '=')) {
1027       if (arg[l] == '=')
1028          *value = arg + l + 1;
1029       else
1030          *value = NULL;
1031
1032       return true;
1033    }
1034
1035    return false;
1036 }
1037
1038 int main(int argc, char *argv[])
1039 {
1040    struct gen_spec *spec;
1041    struct aub_file *file;
1042    int i, pci_id = 0;
1043    bool found_arg_gen = false, pager = true;
1044    int gen_major, gen_minor;
1045    const char *value;
1046    char gen_file[256], gen_val[24];
1047
1048    if (argc == 1) {
1049       print_help(stderr);
1050       exit(EXIT_FAILURE);
1051    }
1052
1053    for (i = 1; i < argc; ++i) {
1054       if (strcmp(argv[i], "--no-pager") == 0) {
1055          pager = false;
1056       } else if (strcmp(argv[i], "--no-offsets") == 0) {
1057          option_print_offsets = false;
1058       } else if (is_prefix(argv[i], "--gen", &value)) {
1059          if (value == NULL)
1060             error(EXIT_FAILURE, 0, "option '--gen' requires an argument\n");
1061          found_arg_gen = true;
1062          gen_major = 0;
1063          gen_minor = 0;
1064          snprintf(gen_val, sizeof(gen_val), "%s", value);
1065       } else if (strcmp(argv[i], "--headers") == 0) {
1066          option_full_decode = false;
1067       } else if (is_prefix(argv[i], "--color", &value)) {
1068          if (value == NULL || strcmp(value, "always") == 0)
1069             option_color = COLOR_ALWAYS;
1070          else if (strcmp(value, "never") == 0)
1071             option_color = COLOR_NEVER;
1072          else if (strcmp(value, "auto") == 0)
1073             option_color = COLOR_AUTO;
1074          else
1075             error(EXIT_FAILURE, 0, "invalid value for --color: %s", value);
1076       } else if (strcmp(argv[i], "--help") == 0) {
1077          print_help(stdout);
1078          exit(EXIT_SUCCESS);
1079       } else {
1080          if (argv[i][0] == '-') {
1081             fprintf(stderr, "unknown option %s\n", argv[i]);
1082             exit(EXIT_FAILURE);
1083          }
1084          break;
1085       }
1086    }
1087
1088    if (!found_arg_gen) {
1089       fprintf(stderr, "argument --gen is required\n");
1090       exit(EXIT_FAILURE);
1091    }
1092
1093    if (strstr(gen_val, "ivb") != NULL) {
1094       /* Intel(R) Ivybridge Mobile GT2 */
1095       pci_id = 0x0166;
1096       gen_major = 7;
1097       gen_minor = 0;
1098    } else if (strstr(gen_val, "hsw") != NULL) {
1099       /* Intel(R) Haswell Mobile GT2 */
1100       pci_id = 0x0416;
1101       gen_major = 7;
1102       gen_minor = 5;
1103    } else if (strstr(gen_val, "byt") != NULL) {
1104       /* Intel(R) Bay Trail */
1105       pci_id = 0x0155;
1106       gen_major = 7;
1107       gen_minor = 5;
1108    } else if (strstr(gen_val, "bdw") != NULL) {
1109       /* Intel(R) HD Graphics 5500 (Broadwell GT2) */
1110       pci_id = 0x1616;
1111       gen_major = 8;
1112       gen_minor = 0;
1113    }  else if (strstr(gen_val, "chv") != NULL) {
1114       /* Intel(R) HD Graphics (Cherryview) */
1115       pci_id = 0x22B3;
1116       gen_major = 8;
1117       gen_minor = 0;
1118    } else if (strstr(gen_val, "skl") != NULL) {
1119       /* Intel(R) HD Graphics 530 (Skylake GT2) */
1120       pci_id = 0x1912;
1121       gen_major = 9;
1122       gen_minor = 0;
1123    } else if (strstr(gen_val, "kbl") != NULL) {
1124       /* Intel(R) Kabylake GT2 */
1125       pci_id = 0x591D;
1126       gen_major = 9;
1127       gen_minor = 0;
1128    } else if (strstr(gen_val, "bxt") != NULL) {
1129       /* Intel(R) HD Graphics (Broxton) */
1130       pci_id = 0x0A84;
1131       gen_major = 9;
1132       gen_minor = 0;
1133    } else {
1134       error(EXIT_FAILURE, 0, "can't parse gen: %s, expected ivb, byt, hsw, "
1135                              "bdw, chv, skl, kbl or bxt\n", gen_val);
1136    }
1137
1138    /* Do this before we redirect stdout to pager. */
1139    if (option_color == COLOR_AUTO)
1140       option_color = isatty(1) ? COLOR_ALWAYS : COLOR_NEVER;
1141
1142    if (isatty(1) && pager)
1143       setup_pager();
1144
1145    if (gen_minor > 0) {
1146       snprintf(gen_file, sizeof(gen_file), "../genxml/gen%d%d.xml",
1147                gen_major, gen_minor);
1148    } else {
1149       snprintf(gen_file, sizeof(gen_file), "../genxml/gen%d.xml", gen_major);
1150    }
1151
1152    spec = gen_spec_load(gen_file);
1153    disasm = gen_disasm_create(pci_id);
1154
1155    if (argv[i] == NULL) {
1156        print_help(stderr);
1157        exit(EXIT_FAILURE);
1158    } else {
1159        file = aub_file_open(argv[i]);
1160    }
1161
1162    while (aub_file_more_stuff(file))
1163       aub_file_decode_batch(file, spec);
1164
1165    fflush(stdout);
1166    /* close the stdout which is opened to write the output */
1167    close(1);
1168
1169    wait(NULL);
1170
1171    return EXIT_SUCCESS;
1172 }