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