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