src/intel/tools/aubinator.c

   1 /*
   2  * Copyright © 2016 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <stdint.h>
  27
  28 #include <unistd.h>
  29 #include <fcntl.h>
  30 #include <string.h>
  31 #include <signal.h>
  32 #include <errno.h>
  33 #include <error.h>
  34 #include <sys/types.h>
  35 #include <sys/stat.h>
  36 #include <sys/wait.h>
  37 #include <sys/mman.h>
  38
  39 #include "decoder.h"
  40 #include "intel_aub.h"
  41 #include "gen_disasm.h"
  42
  43 /* Below is the only command missing from intel_aub.h in libdrm
  44  * So, reuse intel_aub.h from libdrm and #define the
  45  * AUB_MI_BATCH_BUFFER_END as below
  46  */
  47 #define AUB_MI_BATCH_BUFFER_END (0x0500 << 16)
  48
  49 #define CSI "\e["
  50 #define HEADER CSI "37;44m"
  51 #define NORMAL CSI "0m"
  52 #define CLEAR_TO_EOL CSI "0K"
  53
  54 /* options */
  55
  56 static bool option_full_decode = true;
  57 static bool option_print_offsets = true;
  58 static enum { COLOR_AUTO, COLOR_ALWAYS, COLOR_NEVER } option_color;
  59
  60 /* state */
  61
  62 struct gen_disasm *disasm;
  63
  64 uint64_t gtt_size, gtt_end;
  65 void *gtt;
  66 uint64_t general_state_base;
  67 uint64_t surface_state_base;
  68 uint64_t dynamic_state_base;
  69 uint64_t instruction_base;
  70 uint64_t instruction_bound;
  71
  72 static inline uint32_t
  73 field(uint32_t value, int start, int end)
  74 {
  75    uint32_t mask;
  76
  77    mask = ~0U >> (31 - end + start);
  78
  79    return (value >> start) & mask;
  80 }
  81
  82 struct brw_instruction;
  83
  84 static inline int
  85 valid_offset(uint32_t offset)
  86 {
  87    return offset < gtt_end;
  88 }
  89
  90 static void
  91 print_dword_val(struct gen_field_iterator *iter, uint64_t offset, int *dword_num)
  92 {
  93    struct gen_field *f;
  94    union {
  95       uint32_t dw;
  96       float f;
  97    } v;
  98
  99    f = iter->group->fields[iter->i - 1];
 100    v.dw = iter->p[f->start / 32];
 101
 102    if (*dword_num != (f->start / 32)) {
 103       printf("0x%08lx:  0x%08x : Dword %d\n",(offset + 4 * (f->start / 32)), v.dw, f->start / 32);
 104       *dword_num = (f->start / 32);
 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, 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: %08lx <not valid>\n", start);
 302          continue;
 303       } else {
 304          printf("kernel: %08lx\n", start);
 305       }
 306
 307       insns = (struct brw_instruction *) (gtt + start);
 308       gen_disasm_disassemble(disasm, insns, 0, 8192, 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 %08lx, start %08lx\n",
 403              instruction_base, start);
 404
 405       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 406       gen_disasm_disassemble(disasm, insns, 0, 8192, 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 %08lx, start %08lx\n",
 427              instruction_base, start);
 428
 429       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 430       gen_disasm_disassemble(disasm, insns, 0, 8192, 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, 8192, 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, 8192, 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, 8192, 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, uint32_t *p)
 571 {
 572    uint64_t start;
 573    struct gen_group *sf_clip_viewport;
 574
 575    sf_clip_viewport = gen_spec_find_struct(spec, "SF_CLIP_VIEWPORT");
 576
 577    start = dynamic_state_base + (p[1] & ~0x3fu);
 578    for (uint32_t i = 0; i < 4; i++) {
 579       printf("viewport %d\n", i);
 580       decode_structure(spec, sf_clip_viewport, gtt + start + i * 64);
 581    }
 582 }
 583
 584 static void
 585 handle_3dstate_blend_state_pointers(struct gen_spec *spec, uint32_t *p)
 586 {
 587    uint64_t start;
 588    struct gen_group *blend_state;
 589
 590    blend_state = gen_spec_find_struct(spec, "BLEND_STATE");
 591
 592    start = dynamic_state_base + (p[1] & ~0x3fu);
 593    decode_structure(spec, blend_state, gtt + start);
 594 }
 595
 596 static void
 597 handle_3dstate_cc_state_pointers(struct gen_spec *spec, uint32_t *p)
 598 {
 599    uint64_t start;
 600    struct gen_group *cc_state;
 601
 602    cc_state = gen_spec_find_struct(spec, "COLOR_CALC_STATE");
 603
 604    start = dynamic_state_base + (p[1] & ~0x3fu);
 605    decode_structure(spec, cc_state, gtt + start);
 606 }
 607
 608 static void
 609 handle_3dstate_scissor_state_pointers(struct gen_spec *spec, uint32_t *p)
 610 {
 611    uint64_t start;
 612    struct gen_group *scissor_rect;
 613
 614    scissor_rect = gen_spec_find_struct(spec, "SCISSOR_RECT");
 615
 616    start = dynamic_state_base + (p[1] & ~0x1fu);
 617    decode_structure(spec, scissor_rect, gtt + start);
 618 }
 619
 620 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
 621
 622 #define STATE_BASE_ADDRESS                  0x61010000
 623
 624 #define MEDIA_INTERFACE_DESCRIPTOR_LOAD     0x70020000
 625
 626 #define _3DSTATE_INDEX_BUFFER               0x780a0000
 627 #define _3DSTATE_VERTEX_BUFFERS             0x78080000
 628
 629 #define _3DSTATE_VS                         0x78100000
 630 #define _3DSTATE_GS                         0x78110000
 631 #define _3DSTATE_HS                         0x781b0000
 632 #define _3DSTATE_DS                         0x781d0000
 633
 634 #define _3DSTATE_CONSTANT_VS                0x78150000
 635 #define _3DSTATE_CONSTANT_GS                0x78160000
 636 #define _3DSTATE_CONSTANT_PS                0x78170000
 637 #define _3DSTATE_CONSTANT_HS                0x78190000
 638 #define _3DSTATE_CONSTANT_DS                0x781A0000
 639
 640 #define _3DSTATE_PS                         0x78200000
 641
 642 #define _3DSTATE_BINDING_TABLE_POINTERS_VS  0x78260000
 643 #define _3DSTATE_BINDING_TABLE_POINTERS_HS  0x78270000
 644 #define _3DSTATE_BINDING_TABLE_POINTERS_DS  0x78280000
 645 #define _3DSTATE_BINDING_TABLE_POINTERS_GS  0x78290000
 646 #define _3DSTATE_BINDING_TABLE_POINTERS_PS  0x782a0000
 647
 648 #define _3DSTATE_SAMPLER_STATE_POINTERS_VS  0x782b0000
 649 #define _3DSTATE_SAMPLER_STATE_POINTERS_GS  0x782e0000
 650 #define _3DSTATE_SAMPLER_STATE_POINTERS_PS  0x782f0000
 651
 652 #define _3DSTATE_VIEWPORT_STATE_POINTERS_CC 0x78230000
 653 #define _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP 0x78210000
 654 #define _3DSTATE_BLEND_STATE_POINTERS       0x78240000
 655 #define _3DSTATE_CC_STATE_POINTERS          0x780e0000
 656 #define _3DSTATE_SCISSOR_STATE_POINTERS     0x780f0000
 657
 658 struct custom_handler {
 659    uint32_t opcode;
 660    void (*handle)(struct gen_spec *spec, uint32_t *p);
 661 } custom_handlers[] = {
 662    { STATE_BASE_ADDRESS, handle_state_base_address },
 663    { MEDIA_INTERFACE_DESCRIPTOR_LOAD, handle_media_interface_descriptor_load },
 664    { _3DSTATE_VERTEX_BUFFERS, handle_3dstate_vertex_buffers },
 665    { _3DSTATE_INDEX_BUFFER, handle_3dstate_index_buffer },
 666    { _3DSTATE_VS, handle_3dstate_vs },
 667    { _3DSTATE_GS, handle_3dstate_vs },
 668    { _3DSTATE_DS, handle_3dstate_vs },
 669    { _3DSTATE_HS, handle_3dstate_hs },
 670    { _3DSTATE_CONSTANT_VS, handle_3dstate_constant },
 671    { _3DSTATE_CONSTANT_GS, handle_3dstate_constant },
 672    { _3DSTATE_CONSTANT_PS, handle_3dstate_constant },
 673    { _3DSTATE_CONSTANT_HS, handle_3dstate_constant },
 674    { _3DSTATE_CONSTANT_DS, handle_3dstate_constant },
 675    { _3DSTATE_PS, handle_3dstate_ps },
 676
 677    { _3DSTATE_BINDING_TABLE_POINTERS_VS, handle_3dstate_binding_table_pointers },
 678    { _3DSTATE_BINDING_TABLE_POINTERS_HS, handle_3dstate_binding_table_pointers },
 679    { _3DSTATE_BINDING_TABLE_POINTERS_DS, handle_3dstate_binding_table_pointers },
 680    { _3DSTATE_BINDING_TABLE_POINTERS_GS, handle_3dstate_binding_table_pointers },
 681    { _3DSTATE_BINDING_TABLE_POINTERS_PS, handle_3dstate_binding_table_pointers },
 682
 683    { _3DSTATE_SAMPLER_STATE_POINTERS_VS, handle_3dstate_sampler_state_pointers },
 684    { _3DSTATE_SAMPLER_STATE_POINTERS_GS, handle_3dstate_sampler_state_pointers },
 685    { _3DSTATE_SAMPLER_STATE_POINTERS_PS, handle_3dstate_sampler_state_pointers },
 686
 687    { _3DSTATE_VIEWPORT_STATE_POINTERS_CC, handle_3dstate_viewport_state_pointers_cc },
 688    { _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP, handle_3dstate_viewport_state_pointers_sf_clip },
 689    { _3DSTATE_BLEND_STATE_POINTERS, handle_3dstate_blend_state_pointers },
 690    { _3DSTATE_CC_STATE_POINTERS, handle_3dstate_cc_state_pointers },
 691    { _3DSTATE_SCISSOR_STATE_POINTERS, handle_3dstate_scissor_state_pointers }
 692 };
 693
 694 static void
 695 parse_commands(struct gen_spec *spec, uint32_t *cmds, int size, int engine)
 696 {
 697    uint32_t *p, *end = cmds + size / 4;
 698    unsigned int length, i;
 699    struct gen_group *inst;
 700
 701    for (p = cmds; p < end; p += length) {
 702       inst = gen_spec_find_instruction(spec, p);
 703       if (inst == NULL) {
 704          printf("unknown instruction %08x\n", p[0]);
 705          length = (p[0] & 0xff) + 2;
 706          continue;
 707       }
 708       length = gen_group_get_length(inst, p);
 709
 710       const char *color, *reset_color = CLEAR_TO_EOL NORMAL;
 711       uint64_t offset;
 712
 713       if (option_full_decode)
 714          color = HEADER;
 715       else
 716          color = NORMAL;
 717
 718       if (option_color == COLOR_NEVER) {
 719          color = "";
 720          reset_color = "";
 721       }
 722
 723       if (option_print_offsets)
 724          offset = (void *) p - gtt;
 725       else
 726          offset = 0;
 727
 728       printf("%s0x%08lx:  0x%08x:  %s%s\n",
 729              color, offset, p[0],
 730              gen_group_get_name(inst), reset_color);
 731
 732       if (option_full_decode) {
 733          struct gen_field_iterator iter;
 734          char* token = NULL;
 735          int idx = 0, dword_num = 0;
 736          gen_field_iterator_init(&iter, inst, p);
 737          while (gen_field_iterator_next(&iter)) {
 738             idx = 0;
 739             print_dword_val(&iter, offset, &dword_num);
 740             if (dword_num > 0)
 741                 token = print_iterator_values(&iter, &idx);
 742             if (token != NULL) {
 743                 printf("0x%08lx:  0x%08x : Dword %d\n",(offset+4*idx), p[idx], idx);
 744                 handle_struct_decode(spec,token, &p[idx]);
 745                 token = NULL;
 746             }
 747          }
 748
 749          for (i = 0; i < ARRAY_LENGTH(custom_handlers); i++) {
 750             if (gen_group_get_opcode(inst) ==
 751                 custom_handlers[i].opcode)
 752                custom_handlers[i].handle(spec, p);
 753          }
 754       }
 755
 756       if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_START) {
 757          uint64_t start;
 758          if (gen_spec_get_gen(spec) >= gen_make_gen(8,0))
 759             start = get_qword(&p[1]);
 760          else
 761             start = p[1];
 762
 763          parse_commands(spec, gtt + start, 1 << 20, engine);
 764       } else if ((p[0] & 0xffff0000) == AUB_MI_BATCH_BUFFER_END) {
 765          break;
 766       }
 767    }
 768 }
 769
 770 #define GEN_ENGINE_RENDER 1
 771 #define GEN_ENGINE_BLITTER 2
 772
 773 static void
 774 handle_trace_block(struct gen_spec *spec, uint32_t *p)
 775 {
 776    int operation = p[1] & AUB_TRACE_OPERATION_MASK;
 777    int type = p[1] & AUB_TRACE_TYPE_MASK;
 778    int address_space = p[1] & AUB_TRACE_ADDRESS_SPACE_MASK;
 779    uint64_t offset = p[3];
 780    uint32_t size = p[4];
 781    int header_length = p[0] & 0xffff;
 782    uint32_t *data = p + header_length + 2;
 783    int engine = GEN_ENGINE_RENDER;
 784
 785    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0))
 786       offset += (uint64_t) p[5] << 32;
 787
 788    switch (operation) {
 789    case AUB_TRACE_OP_DATA_WRITE:
 790       if (address_space != AUB_TRACE_MEMTYPE_GTT)
 791          break;
 792       if (gtt_size < offset + size)
 793          error(EXIT_FAILURE, errno, "overflow gtt space");
 794       memcpy((char *) gtt + offset, data, size);
 795       if (gtt_end < offset + size)
 796          gtt_end = offset + size;
 797       break;
 798    case AUB_TRACE_OP_COMMAND_WRITE:
 799       switch (type) {
 800       case AUB_TRACE_TYPE_RING_PRB0:
 801          engine = GEN_ENGINE_RENDER;
 802          break;
 803       case AUB_TRACE_TYPE_RING_PRB2:
 804          engine = GEN_ENGINE_BLITTER;
 805          break;
 806       default:
 807          printf("command write to unknown ring %d\n", type);
 808          break;
 809       }
 810
 811       parse_commands(spec, data, size, engine);
 812       gtt_end = 0;
 813       break;
 814    }
 815 }
 816
 817 struct aub_file {
 818    char *filename;
 819    int fd;
 820    struct stat sb;
 821    uint32_t *map, *end, *cursor;
 822 };
 823
 824 static struct aub_file *
 825 aub_file_open(const char *filename)
 826 {
 827    struct aub_file *file;
 828
 829    file = malloc(sizeof *file);
 830    file->filename = strdup(filename);
 831    file->fd = open(file->filename, O_RDONLY);
 832    if (file->fd == -1)
 833       error(EXIT_FAILURE, errno, "open %s failed", file->filename);
 834
 835    if (fstat(file->fd, &file->sb) == -1)
 836       error(EXIT_FAILURE, errno, "stat failed");
 837
 838    file->map = mmap(NULL, file->sb.st_size,
 839                     PROT_READ, MAP_SHARED, file->fd, 0);
 840    if (file->map == MAP_FAILED)
 841       error(EXIT_FAILURE, errno, "mmap failed");
 842
 843    file->cursor = file->map;
 844    file->end = file->map + file->sb.st_size / 4;
 845
 846    /* mmap a terabyte for our gtt space. */
 847    gtt_size = 1ul << 40;
 848    gtt = mmap(NULL, gtt_size, PROT_READ | PROT_WRITE,
 849               MAP_PRIVATE | MAP_ANONYMOUS |  MAP_NORESERVE, -1, 0);
 850    if (gtt == MAP_FAILED)
 851       error(EXIT_FAILURE, errno, "failed to alloc gtt space");
 852
 853    return file;
 854 }
 855
 856 #define TYPE(dw)       (((dw) >> 29) & 7)
 857 #define OPCODE(dw)     (((dw) >> 23) & 0x3f)
 858 #define SUBOPCODE(dw)  (((dw) >> 16) & 0x7f)
 859
 860 #define MAKE_HEADER(type, opcode, subopcode) \
 861                    (((type) << 29) | ((opcode) << 23) | ((subopcode) << 16))
 862
 863 #define TYPE_AUB            0x7
 864
 865 /* Classic AUB opcodes */
 866 #define OPCODE_AUB          0x01
 867 #define SUBOPCODE_HEADER    0x05
 868 #define SUBOPCODE_BLOCK     0x41
 869 #define SUBOPCODE_BMP       0x1e
 870
 871 /* Newer version AUB opcode*/
 872 #define OPCODE_NEW_AUB      0x2e
 873 #define SUBOPCODE_VERSION   0x00
 874 #define SUBOPCODE_REG_WRITE 0x03
 875 #define SUBOPCODE_MEM_POLL  0x05
 876 #define SUBOPCODE_MEM_WRITE 0x06
 877
 878 #define MAKE_GEN(major, minor) ( ((major) << 8) | (minor) )
 879
 880 struct {
 881    const char *name;
 882    uint32_t gen;
 883 } device_map[] = {
 884    { "bwr", MAKE_GEN(4, 0) },
 885    { "cln", MAKE_GEN(4, 0) },
 886    { "blc", MAKE_GEN(4, 0) },
 887    { "ctg", MAKE_GEN(4, 0) },
 888    { "el", MAKE_GEN(4, 0) },
 889    { "il", MAKE_GEN(4, 0) },
 890    { "sbr", MAKE_GEN(6, 0) },
 891    { "ivb", MAKE_GEN(7, 0) },
 892    { "lrb2", MAKE_GEN(0, 0) },
 893    { "hsw", MAKE_GEN(7, 5) },
 894    { "vlv", MAKE_GEN(7, 0) },
 895    { "bdw", MAKE_GEN(8, 0) },
 896    { "skl", MAKE_GEN(9, 0) },
 897    { "chv", MAKE_GEN(8, 0) },
 898    { "bxt", MAKE_GEN(9, 0) }
 899 };
 900
 901 static void
 902 aub_file_decode_batch(struct aub_file *file, struct gen_spec *spec)
 903 {
 904    uint32_t *p, h, device, data_type;
 905    int header_length, payload_size, bias;
 906
 907    p = file->cursor;
 908    h = *p;
 909    header_length = h & 0xffff;
 910
 911    switch (OPCODE(h)) {
 912    case OPCODE_AUB:
 913       bias = 2;
 914       break;
 915    case OPCODE_NEW_AUB:
 916       bias = 1;
 917       break;
 918    default:
 919       printf("unknown opcode %d at %ld/%ld\n",
 920              OPCODE(h), file->cursor - file->map,
 921              file->end - file->map);
 922       file->cursor = file->end;
 923       return;
 924    }
 925
 926    payload_size = 0;
 927    switch (h & 0xffff0000) {
 928    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_HEADER):
 929       payload_size = p[12];
 930       break;
 931    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK):
 932       payload_size = p[4];
 933       handle_trace_block(spec, p);
 934       break;
 935    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BMP):
 936       break;
 937
 938    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_VERSION):
 939       printf("version block: dw1 %08x\n", p[1]);
 940       device = (p[1] >> 8) & 0xff;
 941       printf("  device %s\n", device_map[device].name);
 942       break;
 943    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_REG_WRITE):
 944       printf("register write block: (dwords %d)\n", h & 0xffff);
 945       printf("  reg 0x%x, data 0x%x\n", p[1], p[5]);
 946       break;
 947    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_WRITE):
 948       printf("memory write block (dwords %d):\n", h & 0xffff);
 949       printf("  address 0x%lx\n", *(uint64_t *) &p[1]);
 950       data_type = (p[3] >> 20) & 0xff;
 951       if (data_type != 0)
 952          printf("  data type 0x%x\n", data_type);
 953       printf("  address space 0x%x\n", (p[3] >> 28) & 0xf);
 954       break;
 955    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_POLL):
 956       printf("memory poll block (dwords %d):\n", h & 0xffff);
 957       break;
 958    default:
 959       printf("unknown block type=0x%x, opcode=0x%x, "
 960              "subopcode=0x%x (%08x)\n", TYPE(h), OPCODE(h), SUBOPCODE(h), h);
 961       break;
 962    }
 963    file->cursor = p + header_length + bias + payload_size / 4;
 964 }
 965
 966 static int
 967 aub_file_more_stuff(struct aub_file *file)
 968 {
 969    return file->cursor < file->end;
 970 }
 971
 972 static void
 973 setup_pager(void)
 974 {
 975    int fds[2];
 976    pid_t pid;
 977
 978    if (!isatty(1))
 979       return;
 980
 981    if (pipe(fds) == -1)
 982       return;
 983
 984    pid = fork();
 985    if (pid == -1)
 986       return;
 987
 988    if (pid == 0) {
 989       close(fds[1]);
 990       dup2(fds[0], 0);
 991       execlp("less", "less", "-rFi", NULL);
 992    }
 993
 994    close(fds[0]);
 995    dup2(fds[1], 1);
 996    close(fds[1]);
 997 }
 998
 999 static void
1000 print_help(FILE *file)
1001 {
1002    fprintf(file,
1003            "Usage: %s [OPTION]... FILE\n"
1004            "Decode aub file contents.\n\n"
1005            "A valid --gen option must be provided.\n\n"
1006            "      --help          display this help and exit\n"
1007            "      --gen=platform  decode for given platform (ivb, byt, hsw, bdw, chv, skl, kbl or bxt)\n"
1008            "      --headers       decode only command headers\n"
1009            "      --color[=WHEN]  colorize the output; WHEN can be 'auto' (default\n"
1010            "                        if omitted), 'always', or 'never'\n"
1011            "      --no-pager      don't launch pager\n"
1012            "      --no-offsets    don't print instruction offsets\n",
1013            basename(program_invocation_name));
1014 }
1015
1016 static bool
1017 is_prefix(const char *arg, const char *prefix, const char **value)
1018 {
1019    int l = strlen(prefix);
1020
1021    if (strncmp(arg, prefix, l) == 0 && (arg[l] == '\0' || arg[l] == '=')) {
1022       if (arg[l] == '=')
1023          *value = arg + l + 1;
1024       else
1025          *value = NULL;
1026
1027       return true;
1028    }
1029
1030    return false;
1031 }
1032
1033 int main(int argc, char *argv[])
1034 {
1035    struct gen_spec *spec;
1036    struct aub_file *file;
1037    int i, pci_id = 0;
1038    bool found_arg_gen = false, pager = true;
1039    int gen_major, gen_minor;
1040    const char *value;
1041    char gen_file[256], gen_val[24];
1042
1043    if (argc == 1) {
1044       print_help(stderr);
1045       exit(EXIT_FAILURE);
1046    }
1047
1048    for (i = 1; i < argc; ++i) {
1049       if (strcmp(argv[i], "--no-pager") == 0) {
1050          pager = false;
1051       } else if (strcmp(argv[i], "--no-offsets") == 0) {
1052          option_print_offsets = false;
1053       } else if (is_prefix(argv[i], "--gen", &value)) {
1054          if (value == NULL)
1055             error(EXIT_FAILURE, 0, "option '--gen' requires an argument\n");
1056          found_arg_gen = true;
1057          gen_major = 0;
1058          gen_minor = 0;
1059          snprintf(gen_val, sizeof(gen_val), "%s", value);
1060       } else if (strcmp(argv[i], "--headers") == 0) {
1061          option_full_decode = false;
1062       } else if (is_prefix(argv[i], "--color", &value)) {
1063          if (value == NULL || strcmp(value, "always") == 0)
1064             option_color = COLOR_ALWAYS;
1065          else if (strcmp(value, "never") == 0)
1066             option_color = COLOR_NEVER;
1067          else if (strcmp(value, "auto") == 0)
1068             option_color = COLOR_AUTO;
1069          else
1070             error(EXIT_FAILURE, 0, "invalid value for --color: %s", value);
1071       } else if (strcmp(argv[i], "--help") == 0) {
1072          print_help(stdout);
1073          exit(EXIT_SUCCESS);
1074       } else {
1075          if (argv[i][0] == '-') {
1076             fprintf(stderr, "unknown option %s\n", argv[i]);
1077             exit(EXIT_FAILURE);
1078          }
1079          break;
1080       }
1081    }
1082
1083    if (!found_arg_gen) {
1084       fprintf(stderr, "argument --gen is required\n");
1085       exit(EXIT_FAILURE);
1086    }
1087
1088    if (strstr(gen_val,"ivb") != NULL) {
1089       /* Intel(R) Ivybridge Mobile GT2 */
1090       pci_id = 0x0166;
1091       gen_major = 7;
1092       gen_minor = 0;
1093    } else if (strstr(gen_val,"hsw") != NULL) {
1094       /* Intel(R) Haswell Mobile GT2  */
1095       pci_id = 0x0416;
1096       gen_major = 7;
1097       gen_minor = 5;
1098    } else if (strstr(gen_val,"byt") != NULL) {
1099       /* Intel(R) Bay Trail  */
1100       pci_id = 0x0155;
1101       gen_major = 7;
1102       gen_minor = 5;
1103    } else if (strstr(gen_val,"bdw") != NULL) {
1104       /* Intel(R) HD Graphics 5500 (Broadwell GT2) */
1105       pci_id = 0x1616;
1106       gen_major = 8;
1107       gen_minor = 0;
1108    }  else if (strstr(gen_val,"chv") != NULL) {
1109       /* Intel(R) HD Graphics (Cherryview) */
1110       pci_id = 0x22B3;
1111       gen_major = 8;
1112       gen_minor = 0;
1113    } else if (strstr(gen_val,"skl") != NULL) {
1114       /* Intel(R) HD Graphics 530 (Skylake GT2) */
1115       pci_id = 0x1912;
1116       gen_major = 9;
1117       gen_minor = 0;
1118    } else if (strstr(gen_val,"kbl") != NULL) {
1119       /* Intel(R) Kabylake GT2 */
1120       pci_id = 0x591D;
1121       gen_major = 9;
1122       gen_minor = 0;
1123    } else if (strstr(gen_val,"bxt") != NULL) {
1124       /* Intel(R) HD Graphics (Broxton) */
1125       pci_id = 0x0A84;
1126       gen_major = 9;
1127       gen_minor = 0;
1128    } else {
1129       error(EXIT_FAILURE, 0, "can't parse gen: %s, expected ivb, byt, hsw, bdw, chv, skl, kbl or bxt\n", gen_val);
1130    }
1131
1132    /* Do this before we redirect stdout to pager. */
1133    if (option_color == COLOR_AUTO)
1134       option_color = isatty(1) ? COLOR_ALWAYS : COLOR_NEVER;
1135
1136    if (isatty(1) && pager)
1137       setup_pager();
1138
1139    if (gen_minor > 0)
1140       snprintf(gen_file, sizeof(gen_file), "../genxml/gen%d%d.xml", gen_major, gen_minor);
1141    else
1142       snprintf(gen_file, sizeof(gen_file), "../genxml/gen%d.xml", gen_major);
1143
1144    spec = gen_spec_load(gen_file);
1145    disasm = gen_disasm_create(pci_id);
1146
1147    if (argv[i] == NULL) {
1148        print_help(stderr);
1149        exit(EXIT_FAILURE);
1150    } else {
1151        file = aub_file_open(argv[i]);
1152    }
1153
1154    while (aub_file_more_stuff(file))
1155       aub_file_decode_batch(file, spec);
1156
1157    fflush(stdout);
1158    /* close the stdout which is opened to write the output */
1159    close(1);
1160
1161    wait(NULL);
1162
1163    return EXIT_SUCCESS;
1164 }