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