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 <sys/types.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/mman.h>
  37
  38 #include "decoder.h"
  39 #include "intel_aub.h"
  40 #include "gen_disasm.h"
  41
  42 /* Below is the only command missing from intel_aub.h in libdrm
  43  * So, reuse intel_aub.h from libdrm and #define the
  44  * AUB_MI_BATCH_BUFFER_END as below
  45  */
  46 #define AUB_MI_BATCH_BUFFER_END (0x0500 << 16)
  47
  48 #define CSI "\e["
  49 #define HEADER CSI "37;44m"
  50 #define NORMAL CSI "0m"
  51 #define CLEAR_TO_EOL CSI "0K"
  52
  53 /* options */
  54
  55 static bool option_full_decode = true;
  56 static bool option_print_offsets = true;
  57 static enum { COLOR_AUTO, COLOR_ALWAYS, COLOR_NEVER } option_color;
  58
  59 /* state */
  60
  61 struct gen_disasm *disasm;
  62
  63 uint64_t gtt_size, gtt_end;
  64 void *gtt;
  65 uint64_t general_state_base;
  66 uint64_t surface_state_base;
  67 uint64_t dynamic_state_base;
  68 uint64_t instruction_base;
  69 uint64_t instruction_bound;
  70
  71 static inline uint32_t
  72 field(uint32_t value, int start, int end)
  73 {
  74    uint32_t mask;
  75
  76    mask = ~0U >> (31 - end + start);
  77
  78    return (value >> start) & mask;
  79 }
  80
  81 struct brw_instruction;
  82
  83 static inline int
  84 valid_offset(uint32_t offset)
  85 {
  86    return offset < gtt_end;
  87 }
  88
  89 static void
  90 print_dword_val(struct gen_field_iterator *iter, uint64_t offset,
  91                 int *dword_num)
  92 {
  93    struct gen_field *f;
  94
  95    f = iter->group->fields[iter->i - 1];
  96    const int dword = f->start / 32;
  97
  98    if (*dword_num != dword) {
  99       printf("0x%08lx:  0x%08x : Dword %d\n",
 100              offset + 4 * dword,  iter->p[dword], dword);
 101       *dword_num = dword;
 102    }
 103 }
 104
 105 static char *
 106 print_iterator_values(struct gen_field_iterator *iter, int *idx)
 107 {
 108     char *token = NULL;
 109     if (strstr(iter->value, "struct") == NULL) {
 110         printf("    %s: %s\n", iter->name, iter->value);
 111     } else {
 112         token = strtok(iter->value, " ");
 113         if (token != NULL) {
 114             token = strtok(NULL, " ");
 115             *idx = atoi(strtok(NULL, ">"));
 116         } else {
 117             token = NULL;
 118         }
 119         printf("    %s:<struct %s>\n", iter->name, token);
 120     }
 121     return token;
 122 }
 123
 124 static void
 125 decode_structure(struct gen_spec *spec, struct gen_group *strct,
 126                  const uint32_t *p)
 127 {
 128    struct gen_field_iterator iter;
 129    char *token = NULL;
 130    int idx = 0, dword_num = 0;
 131    uint64_t offset = 0;
 132
 133    if (option_print_offsets)
 134       offset = (void *) p - gtt;
 135    else
 136       offset = 0;
 137
 138    gen_field_iterator_init(&iter, strct, p);
 139    while (gen_field_iterator_next(&iter)) {
 140       idx = 0;
 141       print_dword_val(&iter, offset, &dword_num);
 142       token = print_iterator_values(&iter, &idx);
 143       if (token != NULL) {
 144          struct gen_group *struct_val = gen_spec_find_struct(spec, token);
 145          decode_structure(spec, struct_val, &p[idx]);
 146          token = NULL;
 147       }
 148    }
 149 }
 150
 151 static void
 152 handle_struct_decode(struct gen_spec *spec, char *struct_name, uint32_t *p)
 153 {
 154    if (struct_name == NULL)
 155       return;
 156    struct gen_group *struct_val = gen_spec_find_struct(spec, struct_name);
 157    decode_structure(spec, struct_val, p);
 158 }
 159
 160 static void
 161 dump_binding_table(struct gen_spec *spec, uint32_t offset)
 162 {
 163    uint32_t *pointers, i;
 164    uint64_t start;
 165    struct gen_group *surface_state;
 166
 167    surface_state = gen_spec_find_struct(spec, "RENDER_SURFACE_STATE");
 168    if (surface_state == NULL) {
 169       printf("did not find RENDER_SURFACE_STATE info\n");
 170       return;
 171    }
 172
 173    start = surface_state_base + offset;
 174    pointers = gtt + start;
 175    for (i = 0; i < 16; i++) {
 176       if (pointers[i] == 0)
 177          continue;
 178       start = pointers[i] + surface_state_base;
 179       if (!valid_offset(start)) {
 180          printf("pointer %u: %08x <not valid>\n",
 181                 i, pointers[i]);
 182          continue;
 183       } else {
 184          printf("pointer %u: %08x\n", i, pointers[i]);
 185       }
 186
 187       decode_structure(spec, surface_state, gtt + start);
 188    }
 189 }
 190
 191 static void
 192 handle_3dstate_index_buffer(struct gen_spec *spec, uint32_t *p)
 193 {
 194    void *start;
 195    uint32_t length, i, type, size;
 196
 197    start = gtt + p[2];
 198    type = (p[1] >> 8) & 3;
 199    size = 1 << type;
 200    length = p[4] / size;
 201    if (length > 10)
 202       length = 10;
 203
 204    printf("\t");
 205
 206    for (i = 0; i < length; i++) {
 207       switch (type) {
 208       case 0:
 209          printf("%3d ", ((uint8_t *)start)[i]);
 210          break;
 211       case 1:
 212          printf("%3d ", ((uint16_t *)start)[i]);
 213          break;
 214       case 2:
 215          printf("%3d ", ((uint32_t *)start)[i]);
 216          break;
 217       }
 218    }
 219    if (length < p[4] / size)
 220       printf("...\n");
 221    else
 222       printf("\n");
 223 }
 224
 225 static inline uint64_t
 226 get_qword(uint32_t *p)
 227 {
 228    return ((uint64_t) p[1] << 32) | p[0];
 229 }
 230
 231 static void
 232 handle_state_base_address(struct gen_spec *spec, uint32_t *p)
 233 {
 234    uint64_t mask = ~((1 << 12) - 1);
 235
 236    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 237       if (p[1] & 1)
 238          general_state_base = get_qword(&p[1]) & mask;
 239       if (p[4] & 1)
 240          surface_state_base = get_qword(&p[4]) & mask;
 241       if (p[6] & 1)
 242          dynamic_state_base = get_qword(&p[6]) & mask;
 243       if (p[10] & 1)
 244          instruction_base = get_qword(&p[10]) & mask;
 245       if (p[15] & 1)
 246          instruction_bound = p[15] & mask;
 247    } else {
 248       if (p[2] & 1)
 249          surface_state_base = p[2] & mask;
 250       if (p[3] & 1)
 251          dynamic_state_base = p[3] & mask;
 252       if (p[5] & 1)
 253          instruction_base = p[5] & mask;
 254       if (p[9] & 1)
 255          instruction_bound = p[9] & mask;
 256    }
 257 }
 258
 259 static void
 260 dump_samplers(struct gen_spec *spec, uint32_t offset)
 261 {
 262    uint32_t i;
 263    uint64_t start;
 264    struct gen_group *sampler_state;
 265
 266    sampler_state = gen_spec_find_struct(spec, "SAMPLER_STATE");
 267
 268    start = dynamic_state_base + offset;
 269    for (i = 0; i < 4; i++) {
 270       printf("sampler state %d\n", i);
 271       decode_structure(spec, sampler_state, gtt + start + i * 16);
 272    }
 273 }
 274
 275 static void
 276 handle_media_interface_descriptor_load(struct gen_spec *spec, uint32_t *p)
 277 {
 278    int i, length = p[2] / 32;
 279    struct gen_group *descriptor_structure;
 280    uint32_t *descriptors;
 281    uint64_t start;
 282    struct brw_instruction *insns;
 283
 284    descriptor_structure =
 285       gen_spec_find_struct(spec, "INTERFACE_DESCRIPTOR_DATA");
 286    if (descriptor_structure == NULL) {
 287       printf("did not find INTERFACE_DESCRIPTOR_DATA info\n");
 288       return;
 289    }
 290
 291    start = dynamic_state_base + p[3];
 292    descriptors = gtt + start;
 293    for (i = 0; i < length; i++, descriptors += 8) {
 294       printf("descriptor %u: %08x\n", i, *descriptors);
 295       decode_structure(spec, descriptor_structure, descriptors);
 296
 297       start = instruction_base + descriptors[0];
 298       if (!valid_offset(start)) {
 299          printf("kernel: %08lx <not valid>\n", start);
 300          continue;
 301       } else {
 302          printf("kernel: %08lx\n", start);
 303       }
 304
 305       insns = (struct brw_instruction *) (gtt + start);
 306       gen_disasm_disassemble(disasm, insns, 0, stdout);
 307
 308       dump_samplers(spec, descriptors[3] & ~0x1f);
 309       dump_binding_table(spec, descriptors[4] & ~0x1f);
 310    }
 311 }
 312
 313 /* Heuristic to determine whether a uint32_t is probably actually a float
 314  * (http://stackoverflow.com/a/2953466)
 315  */
 316
 317 static bool
 318 probably_float(uint32_t bits)
 319 {
 320    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 321    uint32_t mant = bits & 0x007fffff;
 322
 323    /* +- 0.0 */
 324    if (exp == -127 && mant == 0)
 325       return true;
 326
 327    /* +- 1 billionth to 1 billion */
 328    if (-30 <= exp && exp <= 30)
 329       return true;
 330
 331    /* some value with only a few binary digits */
 332    if ((mant & 0x0000ffff) == 0)
 333       return true;
 334
 335    return false;
 336 }
 337
 338 static void
 339 handle_3dstate_vertex_buffers(struct gen_spec *spec, uint32_t *p)
 340 {
 341    uint32_t *end, *s, *dw, *dwend;
 342    uint64_t offset;
 343    int n, i, count, stride;
 344
 345    end = (p[0] & 0xff) + p + 2;
 346    for (s = &p[1], n = 0; s < end; s += 4, n++) {
 347       if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 348          offset = *(uint64_t *) &s[1];
 349          dwend = gtt + offset + s[3];
 350       } else {
 351          offset = s[1];
 352          dwend = gtt + s[2] + 1;
 353       }
 354
 355       stride = field(s[0], 0, 11);
 356       count = 0;
 357       printf("vertex buffer %d, size %d\n", n, s[3]);
 358       for (dw = gtt + offset, i = 0; dw < dwend && i < 256; dw++) {
 359          if (count == 0 && count % (8 * 4) == 0)
 360             printf("  ");
 361
 362          if (probably_float(*dw))
 363             printf("  %8.2f", *(float *) dw);
 364          else
 365             printf("  0x%08x", *dw);
 366
 367          i++;
 368          count += 4;
 369
 370          if (count == stride) {
 371             printf("\n");
 372             count = 0;
 373          } else if (count % (8 * 4) == 0) {
 374             printf("\n");
 375          } else {
 376             printf(" ");
 377          }
 378       }
 379       if (count > 0 && count % (8 * 4) != 0)
 380          printf("\n");
 381    }
 382 }
 383
 384 static void
 385 handle_3dstate_vs(struct gen_spec *spec, uint32_t *p)
 386 {
 387    uint64_t start;
 388    struct brw_instruction *insns;
 389    int vs_enable;
 390
 391    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 392       start = get_qword(&p[1]);
 393       vs_enable = p[7] & 1;
 394    } else {
 395       start = p[1];
 396       vs_enable = p[5] & 1;
 397    }
 398
 399    if (vs_enable) {
 400       printf("instruction_base %08lx, start %08lx\n",
 401              instruction_base, start);
 402
 403       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 404       gen_disasm_disassemble(disasm, insns, 0, stdout);
 405    }
 406 }
 407
 408 static void
 409 handle_3dstate_hs(struct gen_spec *spec, uint32_t *p)
 410 {
 411    uint64_t start;
 412    struct brw_instruction *insns;
 413    int hs_enable;
 414
 415    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 416       start = get_qword(&p[4]);
 417    } else {
 418       start = p[4];
 419    }
 420
 421    hs_enable = p[2] & 0x80000000;
 422
 423    if (hs_enable) {
 424       printf("instruction_base %08lx, start %08lx\n",
 425              instruction_base, start);
 426
 427       insns = (struct brw_instruction *) (gtt + instruction_base + start);
 428       gen_disasm_disassemble(disasm, insns, 0, stdout);
 429    }
 430 }
 431
 432 static void
 433 handle_3dstate_constant(struct gen_spec *spec, uint32_t *p)
 434 {
 435    int i, j, length;
 436    uint32_t *dw;
 437    float *f;
 438
 439    for (i = 0; i < 4; i++) {
 440       length = (p[1 + i / 2] >> (i & 1) * 16) & 0xffff;
 441       f = (float *) (gtt + p[3 + i * 2] + dynamic_state_base);
 442       dw = (uint32_t *) f;
 443       for (j = 0; j < length * 8; j++) {
 444          if (probably_float(dw[j]))
 445             printf("  %04.3f", f[j]);
 446          else
 447             printf("  0x%08x", dw[j]);
 448
 449          if ((j & 7) == 7)
 450             printf("\n");
 451       }
 452    }
 453 }
 454
 455 static void
 456 handle_3dstate_ps(struct gen_spec *spec, uint32_t *p)
 457 {
 458    uint32_t mask = ~((1 << 6) - 1);
 459    uint64_t start;
 460    struct brw_instruction *insns;
 461    static const char unused[] = "unused";
 462    static const char *pixel_type[3] = {"8 pixel", "16 pixel", "32 pixel"};
 463    const char *k0, *k1, *k2;
 464    uint32_t k_mask, k1_offset, k2_offset;
 465
 466    if (gen_spec_get_gen(spec) >= gen_make_gen(8, 0)) {
 467       k_mask = p[6] & 7;
 468       k1_offset = 8;
 469       k2_offset = 10;
 470    } else {
 471       k_mask = p[4] & 7;
 472       k1_offset = 6;
 473       k2_offset = 7;
 474    }
 475
 476 #define DISPATCH_8 1
 477 #define DISPATCH_16 2
 478 #define DISPATCH_32 4
 479
 480    switch (k_mask) {
 481    case DISPATCH_8:
 482       k0 = pixel_type[0];
 483       k1 = unused;
 484       k2 = unused;
 485       break;
 486    case DISPATCH_16:
 487       k0 = pixel_type[1];
 488       k1 = unused;
 489       k2 = unused;
 490       break;
 491    case DISPATCH_8 | DISPATCH_16:
 492       k0 = pixel_type[0];
 493       k1 = unused;
 494       k2 = pixel_type[1];
 495       break;
 496    case DISPATCH_32:
 497       k0 = pixel_type[2];
 498       k1 = unused;
 499       k2 = unused;
 500       break;
 501    case DISPATCH_16 | DISPATCH_32:
 502       k0 = unused;
 503       k1 = pixel_type[2];
 504       k2 = pixel_type[1];
 505       break;
 506    case DISPATCH_8 | DISPATCH_16 | DISPATCH_32:
 507       k0 = pixel_type[0];
 508       k1 = pixel_type[2];
 509       k2 = pixel_type[1];
 510       break;
 511    default:
 512       k0 = unused;
 513       k1 = unused;
 514       k2 = unused;
 515       break;
 516    }
 517
 518    start = instruction_base + (p[1] & mask);
 519    printf("  Kernel[0] %s\n", k0);
 520    if (k0 != unused) {
 521       insns = (struct brw_instruction *) (gtt + start);
 522       gen_disasm_disassemble(disasm, insns, 0, stdout);
 523    }
 524
 525    start = instruction_base + (p[k1_offset] & mask);
 526    printf("  Kernel[1] %s\n", k1);
 527    if (k1 != unused) {
 528       insns = (struct brw_instruction *) (gtt + start);
 529       gen_disasm_disassemble(disasm, insns, 0, stdout);
 530    }
 531
 532    start = instruction_base + (p[k2_offset] & mask);
 533    printf("  Kernel[2] %s\n", k2);
 534    if (k2 != unused) {
 535       insns = (struct brw_instruction *) (gtt + start);
 536       gen_disasm_disassemble(disasm, insns, 0, stdout);
 537    }
 538 }
 539
 540 static void
 541 handle_3dstate_binding_table_pointers(struct gen_spec *spec, uint32_t *p)
 542 {
 543    dump_binding_table(spec, p[1]);
 544 }
 545
 546 static void
 547 handle_3dstate_sampler_state_pointers(struct gen_spec *spec, uint32_t *p)
 548 {
 549    dump_samplers(spec, p[1]);
 550 }
 551
 552 static void
 553 handle_3dstate_viewport_state_pointers_cc(struct gen_spec *spec, uint32_t *p)
 554 {
 555    uint64_t start;
 556    struct gen_group *cc_viewport;
 557
 558    cc_viewport = gen_spec_find_struct(spec, "CC_VIEWPORT");
 559
 560    start = dynamic_state_base + (p[1] & ~0x1fu);
 561    for (uint32_t i = 0; i < 4; i++) {
 562       printf("viewport %d\n", i);
 563       decode_structure(spec, cc_viewport, gtt + start + i * 8);
 564    }
 565 }
 566
 567 static void
 568 handle_3dstate_viewport_state_pointers_sf_clip(struct gen_spec *spec,
 569                                                uint32_t *p)
 570 {
 571    uint64_t start;
 572    struct gen_group *sf_clip_viewport;
 573
 574    sf_clip_viewport = gen_spec_find_struct(spec, "SF_CLIP_VIEWPORT");
 575
 576    start = dynamic_state_base + (p[1] & ~0x3fu);
 577    for (uint32_t i = 0; i < 4; i++) {
 578       printf("viewport %d\n", i);
 579       decode_structure(spec, sf_clip_viewport, gtt + start + i * 64);
 580    }
 581 }
 582
 583 static void
 584 handle_3dstate_blend_state_pointers(struct gen_spec *spec, uint32_t *p)
 585 {
 586    uint64_t start;
 587    struct gen_group *blend_state;
 588
 589    blend_state = gen_spec_find_struct(spec, "BLEND_STATE");
 590
 591    start = dynamic_state_base + (p[1] & ~0x3fu);
 592    decode_structure(spec, blend_state, gtt + start);
 593 }
 594
 595 static void
 596 handle_3dstate_cc_state_pointers(struct gen_spec *spec, uint32_t *p)
 597 {
 598    uint64_t start;
 599    struct gen_group *cc_state;
 600
 601    cc_state = gen_spec_find_struct(spec, "COLOR_CALC_STATE");
 602
 603    start = dynamic_state_base + (p[1] & ~0x3fu);
 604    decode_structure(spec, cc_state, gtt + start);
 605 }
 606
 607 static void
 608 handle_3dstate_scissor_state_pointers(struct gen_spec *spec, uint32_t *p)
 609 {
 610    uint64_t start;
 611    struct gen_group *scissor_rect;
 612
 613    scissor_rect = gen_spec_find_struct(spec, "SCISSOR_RECT");
 614
 615    start = dynamic_state_base + (p[1] & ~0x1fu);
 616    decode_structure(spec, scissor_rect, gtt + start);
 617 }
 618
 619 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
 620
 621 #define STATE_BASE_ADDRESS                  0x61010000
 622
 623 #define MEDIA_INTERFACE_DESCRIPTOR_LOAD     0x70020000
 624
 625 #define _3DSTATE_INDEX_BUFFER               0x780a0000
 626 #define _3DSTATE_VERTEX_BUFFERS             0x78080000
 627
 628 #define _3DSTATE_VS                         0x78100000
 629 #define _3DSTATE_GS                         0x78110000
 630 #define _3DSTATE_HS                         0x781b0000
 631 #define _3DSTATE_DS                         0x781d0000
 632
 633 #define _3DSTATE_CONSTANT_VS                0x78150000
 634 #define _3DSTATE_CONSTANT_GS                0x78160000
 635 #define _3DSTATE_CONSTANT_PS                0x78170000
 636 #define _3DSTATE_CONSTANT_HS                0x78190000
 637 #define _3DSTATE_CONSTANT_DS                0x781A0000
 638
 639 #define _3DSTATE_PS                         0x78200000
 640
 641 #define _3DSTATE_BINDING_TABLE_POINTERS_VS  0x78260000
 642 #define _3DSTATE_BINDING_TABLE_POINTERS_HS  0x78270000
 643 #define _3DSTATE_BINDING_TABLE_POINTERS_DS  0x78280000
 644 #define _3DSTATE_BINDING_TABLE_POINTERS_GS  0x78290000
 645 #define _3DSTATE_BINDING_TABLE_POINTERS_PS  0x782a0000
 646
 647 #define _3DSTATE_SAMPLER_STATE_POINTERS_VS  0x782b0000
 648 #define _3DSTATE_SAMPLER_STATE_POINTERS_GS  0x782e0000
 649 #define _3DSTATE_SAMPLER_STATE_POINTERS_PS  0x782f0000
 650
 651 #define _3DSTATE_VIEWPORT_STATE_POINTERS_CC 0x78230000
 652 #define _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP 0x78210000
 653 #define _3DSTATE_BLEND_STATE_POINTERS       0x78240000
 654 #define _3DSTATE_CC_STATE_POINTERS          0x780e0000
 655 #define _3DSTATE_SCISSOR_STATE_POINTERS     0x780f0000
 656
 657 struct custom_handler {
 658    uint32_t opcode;
 659    void (*handle)(struct gen_spec *spec, uint32_t *p);
 660 } custom_handlers[] = {
 661    { STATE_BASE_ADDRESS, handle_state_base_address },
 662    { MEDIA_INTERFACE_DESCRIPTOR_LOAD, handle_media_interface_descriptor_load },
 663    { _3DSTATE_VERTEX_BUFFERS, handle_3dstate_vertex_buffers },
 664    { _3DSTATE_INDEX_BUFFER, handle_3dstate_index_buffer },
 665    { _3DSTATE_VS, handle_3dstate_vs },
 666    { _3DSTATE_GS, handle_3dstate_vs },
 667    { _3DSTATE_DS, handle_3dstate_vs },
 668    { _3DSTATE_HS, handle_3dstate_hs },
 669    { _3DSTATE_CONSTANT_VS, handle_3dstate_constant },
 670    { _3DSTATE_CONSTANT_GS, handle_3dstate_constant },
 671    { _3DSTATE_CONSTANT_PS, handle_3dstate_constant },
 672    { _3DSTATE_CONSTANT_HS, handle_3dstate_constant },
 673    { _3DSTATE_CONSTANT_DS, handle_3dstate_constant },
 674    { _3DSTATE_PS, handle_3dstate_ps },
 675
 676    { _3DSTATE_BINDING_TABLE_POINTERS_VS, handle_3dstate_binding_table_pointers },
 677    { _3DSTATE_BINDING_TABLE_POINTERS_HS, handle_3dstate_binding_table_pointers },
 678    { _3DSTATE_BINDING_TABLE_POINTERS_DS, handle_3dstate_binding_table_pointers },
 679    { _3DSTATE_BINDING_TABLE_POINTERS_GS, handle_3dstate_binding_table_pointers },
 680    { _3DSTATE_BINDING_TABLE_POINTERS_PS, handle_3dstate_binding_table_pointers },
 681
 682    { _3DSTATE_SAMPLER_STATE_POINTERS_VS, handle_3dstate_sampler_state_pointers },
 683    { _3DSTATE_SAMPLER_STATE_POINTERS_GS, handle_3dstate_sampler_state_pointers },
 684    { _3DSTATE_SAMPLER_STATE_POINTERS_PS, handle_3dstate_sampler_state_pointers },
 685
 686    { _3DSTATE_VIEWPORT_STATE_POINTERS_CC, handle_3dstate_viewport_state_pointers_cc },
 687    { _3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP, handle_3dstate_viewport_state_pointers_sf_clip },
 688    { _3DSTATE_BLEND_STATE_POINTERS, handle_3dstate_blend_state_pointers },
 689    { _3DSTATE_CC_STATE_POINTERS, handle_3dstate_cc_state_pointers },
 690    { _3DSTATE_SCISSOR_STATE_POINTERS, handle_3dstate_scissor_state_pointers }
 691 };
 692
 693 static void
 694 parse_commands(struct gen_spec *spec, uint32_t *cmds, int size, int engine)
 695 {
 696    uint32_t *p, *end = cmds + size / 4;
 697    unsigned int length, i;
 698    struct gen_group *inst;
 699
 700    for (p = cmds; p < end; p += length) {
 701       inst = gen_spec_find_instruction(spec, p);
 702       if (inst == NULL) {
 703          printf("unknown instruction %08x\n", p[0]);
 704          length = (p[0] & 0xff) + 2;
 705          continue;
 706       }
 707       length = gen_group_get_length(inst, p);
 708
 709       const char *color, *reset_color = CLEAR_TO_EOL NORMAL;
 710       uint64_t offset;
 711
 712       if (option_full_decode)
 713          color = HEADER;
 714       else
 715          color = NORMAL;
 716
 717       if (option_color == COLOR_NEVER) {
 718          color = "";
 719          reset_color = "";
 720       }
 721
 722       if (option_print_offsets)
 723          offset = (void *) p - gtt;
 724       else
 725          offset = 0;
 726
 727       printf("%s0x%08lx:  0x%08x:  %s%s\n",
 728              color, offset, p[0],
 729              gen_group_get_name(inst), reset_color);
 730
 731       if (option_full_decode) {
 732          struct gen_field_iterator iter;
 733          char *token = NULL;
 734          int idx = 0, dword_num = 0;
 735          gen_field_iterator_init(&iter, inst, p);
 736          while (gen_field_iterator_next(&iter)) {
 737             idx = 0;
 738             print_dword_val(&iter, offset, &dword_num);
 739             if (dword_num > 0)
 740                 token = print_iterator_values(&iter, &idx);
 741             if (token != NULL) {
 742                 printf("0x%08lx:  0x%08x : Dword %d\n",
 743                        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          fprintf(stderr, "overflow gtt space: %s", strerror(errno));
 794          exit(EXIT_FAILURE);
 795       }
 796       memcpy((char *) gtt + offset, data, size);
 797       if (gtt_end < offset + size)
 798          gtt_end = offset + size;
 799       break;
 800    case AUB_TRACE_OP_COMMAND_WRITE:
 801       switch (type) {
 802       case AUB_TRACE_TYPE_RING_PRB0:
 803          engine = GEN_ENGINE_RENDER;
 804          break;
 805       case AUB_TRACE_TYPE_RING_PRB2:
 806          engine = GEN_ENGINE_BLITTER;
 807          break;
 808       default:
 809          printf("command write to unknown ring %d\n", type);
 810          break;
 811       }
 812
 813       parse_commands(spec, data, size, engine);
 814       gtt_end = 0;
 815       break;
 816    }
 817 }
 818
 819 struct aub_file {
 820    char *filename;
 821    int fd;
 822    struct stat sb;
 823    uint32_t *map, *end, *cursor;
 824 };
 825
 826 static struct aub_file *
 827 aub_file_open(const char *filename)
 828 {
 829    struct aub_file *file;
 830
 831    file = malloc(sizeof *file);
 832    file->filename = strdup(filename);
 833    file->fd = open(file->filename, O_RDONLY);
 834    if (file->fd == -1) {
 835       fprintf(stderr, "open %s failed: %s", file->filename, strerror(errno));
 836       exit(EXIT_FAILURE);
 837    }
 838
 839    if (fstat(file->fd, &file->sb) == -1) {
 840       fprintf(stderr, "stat failed: %s", strerror(errno));
 841       exit(EXIT_FAILURE);
 842    }
 843
 844    file->map = mmap(NULL, file->sb.st_size,
 845                     PROT_READ, MAP_SHARED, file->fd, 0);
 846    if (file->map == MAP_FAILED) {
 847       fprintf(stderr, "mmap failed: %s", strerror(errno));
 848       exit(EXIT_FAILURE);
 849    }
 850
 851    file->cursor = file->map;
 852    file->end = file->map + file->sb.st_size / 4;
 853
 854    /* mmap a terabyte for our gtt space. */
 855    gtt_size = 1ul << 40;
 856    gtt = mmap(NULL, gtt_size, PROT_READ | PROT_WRITE,
 857               MAP_PRIVATE | MAP_ANONYMOUS |  MAP_NORESERVE, -1, 0);
 858    if (gtt == MAP_FAILED) {
 859       fprintf(stderr, "failed to alloc gtt space: %s", strerror(errno));
 860       exit(1);
 861    }
 862
 863    return file;
 864 }
 865
 866 #define TYPE(dw)       (((dw) >> 29) & 7)
 867 #define OPCODE(dw)     (((dw) >> 23) & 0x3f)
 868 #define SUBOPCODE(dw)  (((dw) >> 16) & 0x7f)
 869
 870 #define MAKE_HEADER(type, opcode, subopcode) \
 871                    (((type) << 29) | ((opcode) << 23) | ((subopcode) << 16))
 872
 873 #define TYPE_AUB            0x7
 874
 875 /* Classic AUB opcodes */
 876 #define OPCODE_AUB          0x01
 877 #define SUBOPCODE_HEADER    0x05
 878 #define SUBOPCODE_BLOCK     0x41
 879 #define SUBOPCODE_BMP       0x1e
 880
 881 /* Newer version AUB opcode */
 882 #define OPCODE_NEW_AUB      0x2e
 883 #define SUBOPCODE_VERSION   0x00
 884 #define SUBOPCODE_REG_WRITE 0x03
 885 #define SUBOPCODE_MEM_POLL  0x05
 886 #define SUBOPCODE_MEM_WRITE 0x06
 887
 888 #define MAKE_GEN(major, minor) ( ((major) << 8) | (minor) )
 889
 890 struct {
 891    const char *name;
 892    uint32_t gen;
 893 } device_map[] = {
 894    { "bwr", MAKE_GEN(4, 0) },
 895    { "cln", MAKE_GEN(4, 0) },
 896    { "blc", MAKE_GEN(4, 0) },
 897    { "ctg", MAKE_GEN(4, 0) },
 898    { "el", MAKE_GEN(4, 0) },
 899    { "il", MAKE_GEN(4, 0) },
 900    { "sbr", MAKE_GEN(6, 0) },
 901    { "ivb", MAKE_GEN(7, 0) },
 902    { "lrb2", MAKE_GEN(0, 0) },
 903    { "hsw", MAKE_GEN(7, 5) },
 904    { "vlv", MAKE_GEN(7, 0) },
 905    { "bdw", MAKE_GEN(8, 0) },
 906    { "skl", MAKE_GEN(9, 0) },
 907    { "chv", MAKE_GEN(8, 0) },
 908    { "bxt", MAKE_GEN(9, 0) }
 909 };
 910
 911 static void
 912 aub_file_decode_batch(struct aub_file *file, struct gen_spec *spec)
 913 {
 914    uint32_t *p, h, device, data_type;
 915    int header_length, payload_size, bias;
 916
 917    p = file->cursor;
 918    h = *p;
 919    header_length = h & 0xffff;
 920
 921    switch (OPCODE(h)) {
 922    case OPCODE_AUB:
 923       bias = 2;
 924       break;
 925    case OPCODE_NEW_AUB:
 926       bias = 1;
 927       break;
 928    default:
 929       printf("unknown opcode %d at %ld/%ld\n",
 930              OPCODE(h), file->cursor - file->map,
 931              file->end - file->map);
 932       file->cursor = file->end;
 933       return;
 934    }
 935
 936    payload_size = 0;
 937    switch (h & 0xffff0000) {
 938    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_HEADER):
 939       payload_size = p[12];
 940       break;
 941    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BLOCK):
 942       payload_size = p[4];
 943       handle_trace_block(spec, p);
 944       break;
 945    case MAKE_HEADER(TYPE_AUB, OPCODE_AUB, SUBOPCODE_BMP):
 946       break;
 947
 948    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_VERSION):
 949       printf("version block: dw1 %08x\n", p[1]);
 950       device = (p[1] >> 8) & 0xff;
 951       printf("  device %s\n", device_map[device].name);
 952       break;
 953    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_REG_WRITE):
 954       printf("register write block: (dwords %d)\n", h & 0xffff);
 955       printf("  reg 0x%x, data 0x%x\n", p[1], p[5]);
 956       break;
 957    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_WRITE):
 958       printf("memory write block (dwords %d):\n", h & 0xffff);
 959       printf("  address 0x%lx\n", *(uint64_t *) &p[1]);
 960       data_type = (p[3] >> 20) & 0xff;
 961       if (data_type != 0)
 962          printf("  data type 0x%x\n", data_type);
 963       printf("  address space 0x%x\n", (p[3] >> 28) & 0xf);
 964       break;
 965    case MAKE_HEADER(TYPE_AUB, OPCODE_NEW_AUB, SUBOPCODE_MEM_POLL):
 966       printf("memory poll block (dwords %d):\n", h & 0xffff);
 967       break;
 968    default:
 969       printf("unknown block type=0x%x, opcode=0x%x, "
 970              "subopcode=0x%x (%08x)\n", TYPE(h), OPCODE(h), SUBOPCODE(h), h);
 971       break;
 972    }
 973    file->cursor = p + header_length + bias + payload_size / 4;
 974 }
 975
 976 static int
 977 aub_file_more_stuff(struct aub_file *file)
 978 {
 979    return file->cursor < file->end;
 980 }
 981
 982 static void
 983 setup_pager(void)
 984 {
 985    int fds[2];
 986    pid_t pid;
 987
 988    if (!isatty(1))
 989       return;
 990
 991    if (pipe(fds) == -1)
 992       return;
 993
 994    pid = fork();
 995    if (pid == -1)
 996       return;
 997
 998    if (pid == 0) {
 999       close(fds[1]);
1000       dup2(fds[0], 0);
1001       execlp("less", "less", "-rFi", NULL);
1002    }
1003
1004    close(fds[0]);
1005    dup2(fds[1], 1);
1006    close(fds[1]);
1007 }
1008
1009 static void
1010 print_help(const char *progname, FILE *file)
1011 {
1012    fprintf(file,
1013            "Usage: %s [OPTION]... FILE\n"
1014            "Decode aub file contents.\n\n"
1015            "A valid --gen option must be provided.\n\n"
1016            "      --help          display this help and exit\n"
1017            "      --gen=platform  decode for given platform (ivb, byt, hsw, bdw, chv, skl, kbl or bxt)\n"
1018            "      --headers       decode only command headers\n"
1019            "      --color[=WHEN]  colorize the output; WHEN can be 'auto' (default\n"
1020            "                        if omitted), 'always', or 'never'\n"
1021            "      --no-pager      don't launch pager\n"
1022            "      --no-offsets    don't print instruction offsets\n",
1023            progname);
1024 }
1025
1026 static bool
1027 is_prefix(const char *arg, const char *prefix, const char **value)
1028 {
1029    int l = strlen(prefix);
1030
1031    if (strncmp(arg, prefix, l) == 0 && (arg[l] == '\0' || arg[l] == '=')) {
1032       if (arg[l] == '=')
1033          *value = arg + l + 1;
1034       else
1035          *value = NULL;
1036
1037       return true;
1038    }
1039
1040    return false;
1041 }
1042
1043 int main(int argc, char *argv[])
1044 {
1045    struct gen_spec *spec;
1046    struct aub_file *file;
1047    int i, pci_id = 0;
1048    bool found_arg_gen = false, pager = true;
1049    int gen_major, gen_minor;
1050    const char *value;
1051    char gen_file[256], gen_val[24];
1052
1053    if (argc == 1) {
1054       print_help(argv[0], stderr);
1055       exit(EXIT_FAILURE);
1056    }
1057
1058    for (i = 1; i < argc; ++i) {
1059       if (strcmp(argv[i], "--no-pager") == 0) {
1060          pager = false;
1061       } else if (strcmp(argv[i], "--no-offsets") == 0) {
1062          option_print_offsets = false;
1063       } else if (is_prefix(argv[i], "--gen", &value)) {
1064          if (value == NULL) {
1065             fprintf(stderr, "option '--gen' requires an argument\n");
1066             exit(EXIT_FAILURE);
1067          }
1068          found_arg_gen = true;
1069          gen_major = 0;
1070          gen_minor = 0;
1071          snprintf(gen_val, sizeof(gen_val), "%s", value);
1072       } else if (strcmp(argv[i], "--headers") == 0) {
1073          option_full_decode = false;
1074       } else if (is_prefix(argv[i], "--color", &value)) {
1075          if (value == NULL || strcmp(value, "always") == 0)
1076             option_color = COLOR_ALWAYS;
1077          else if (strcmp(value, "never") == 0)
1078             option_color = COLOR_NEVER;
1079          else if (strcmp(value, "auto") == 0)
1080             option_color = COLOR_AUTO;
1081          else {
1082             fprintf(stderr, "invalid value for --color: %s", value);
1083             exit(EXIT_FAILURE);
1084          }
1085       } else if (strcmp(argv[i], "--help") == 0) {
1086          print_help(argv[0], stdout);
1087          exit(EXIT_SUCCESS);
1088       } else {
1089          if (argv[i][0] == '-') {
1090             fprintf(stderr, "unknown option %s\n", argv[i]);
1091             exit(EXIT_FAILURE);
1092          }
1093          break;
1094       }
1095    }
1096
1097    if (!found_arg_gen) {
1098       fprintf(stderr, "argument --gen is required\n");
1099       exit(EXIT_FAILURE);
1100    }
1101
1102    if (strstr(gen_val, "ivb") != NULL) {
1103       /* Intel(R) Ivybridge Mobile GT2 */
1104       pci_id = 0x0166;
1105       gen_major = 7;
1106       gen_minor = 0;
1107    } else if (strstr(gen_val, "hsw") != NULL) {
1108       /* Intel(R) Haswell Mobile GT2 */
1109       pci_id = 0x0416;
1110       gen_major = 7;
1111       gen_minor = 5;
1112    } else if (strstr(gen_val, "byt") != NULL) {
1113       /* Intel(R) Bay Trail */
1114       pci_id = 0x0155;
1115       gen_major = 7;
1116       gen_minor = 5;
1117    } else if (strstr(gen_val, "bdw") != NULL) {
1118       /* Intel(R) HD Graphics 5500 (Broadwell GT2) */
1119       pci_id = 0x1616;
1120       gen_major = 8;
1121       gen_minor = 0;
1122    }  else if (strstr(gen_val, "chv") != NULL) {
1123       /* Intel(R) HD Graphics (Cherryview) */
1124       pci_id = 0x22B3;
1125       gen_major = 8;
1126       gen_minor = 0;
1127    } else if (strstr(gen_val, "skl") != NULL) {
1128       /* Intel(R) HD Graphics 530 (Skylake GT2) */
1129       pci_id = 0x1912;
1130       gen_major = 9;
1131       gen_minor = 0;
1132    } else if (strstr(gen_val, "kbl") != NULL) {
1133       /* Intel(R) Kabylake GT2 */
1134       pci_id = 0x591D;
1135       gen_major = 9;
1136       gen_minor = 0;
1137    } else if (strstr(gen_val, "bxt") != NULL) {
1138       /* Intel(R) HD Graphics (Broxton) */
1139       pci_id = 0x0A84;
1140       gen_major = 9;
1141       gen_minor = 0;
1142    } else {
1143       fprintf(stderr, "can't parse gen: %s, expected ivb, byt, hsw, "
1144                              "bdw, chv, skl, kbl or bxt\n", gen_val);
1145       exit(EXIT_FAILURE);
1146    }
1147
1148    /* Do this before we redirect stdout to pager. */
1149    if (option_color == COLOR_AUTO)
1150       option_color = isatty(1) ? COLOR_ALWAYS : COLOR_NEVER;
1151
1152    if (isatty(1) && pager)
1153       setup_pager();
1154
1155    if (gen_minor > 0) {
1156       snprintf(gen_file, sizeof(gen_file), "../genxml/gen%d%d.xml",
1157                gen_major, gen_minor);
1158    } else {
1159       snprintf(gen_file, sizeof(gen_file), "../genxml/gen%d.xml", gen_major);
1160    }
1161
1162    spec = gen_spec_load(gen_file);
1163    disasm = gen_disasm_create(pci_id);
1164
1165    if (argv[i] == NULL) {
1166        print_help(argv[0], stderr);
1167        exit(EXIT_FAILURE);
1168    } else {
1169        file = aub_file_open(argv[i]);
1170    }
1171
1172    while (aub_file_more_stuff(file))
1173       aub_file_decode_batch(file, spec);
1174
1175    fflush(stdout);
1176    /* close the stdout which is opened to write the output */
1177    close(1);
1178
1179    wait(NULL);
1180
1181    return EXIT_SUCCESS;
1182 }