src/intel/tools/gen_batch_decoder.c

   1 /*
   2  * Copyright © 2017 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 "common/gen_decoder.h"
  25 #include "gen_disasm.h"
  26
  27 #include <string.h>
  28
  29 void
  30 gen_batch_decode_ctx_init(struct gen_batch_decode_ctx *ctx,
  31                           const struct gen_device_info *devinfo,
  32                           FILE *fp, enum gen_batch_decode_flags flags,
  33                           const char *xml_path,
  34                           struct gen_batch_decode_bo (*get_bo)(void *,
  35                                                                uint64_t),
  36                           void *user_data)
  37 {
  38    memset(ctx, 0, sizeof(*ctx));
  39
  40    ctx->get_bo = get_bo;
  41    ctx->user_data = user_data;
  42    ctx->fp = fp;
  43    ctx->flags = flags;
  44
  45    if (xml_path == NULL)
  46       ctx->spec = gen_spec_load(devinfo);
  47    else
  48       ctx->spec = gen_spec_load_from_path(devinfo, xml_path);
  49    ctx->disasm = gen_disasm_create(devinfo);
  50 }
  51
  52 void
  53 gen_batch_decode_ctx_finish(struct gen_batch_decode_ctx *ctx)
  54 {
  55    gen_spec_destroy(ctx->spec);
  56    gen_disasm_destroy(ctx->disasm);
  57 }
  58
  59 #define CSI "\e["
  60 #define BLUE_HEADER  CSI "0;44m"
  61 #define GREEN_HEADER CSI "1;42m"
  62 #define NORMAL       CSI "0m"
  63
  64 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
  65
  66 static void
  67 ctx_print_group(struct gen_batch_decode_ctx *ctx,
  68                 struct gen_group *group,
  69                 uint64_t address, const void *map)
  70 {
  71    gen_print_group(ctx->fp, group, address, map, 0,
  72                    (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) != 0);
  73 }
  74
  75 static struct gen_batch_decode_bo
  76 ctx_get_bo(struct gen_batch_decode_ctx *ctx, uint64_t addr)
  77 {
  78    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0)) {
  79       /* On Broadwell and above, we have 48-bit addresses which consume two
  80        * dwords.  Some packets require that these get stored in a "canonical
  81        * form" which means that bit 47 is sign-extended through the upper
  82        * bits. In order to correctly handle those aub dumps, we need to mask
  83        * off the top 16 bits.
  84        */
  85       addr &= (~0ull >> 16);
  86    }
  87
  88    struct gen_batch_decode_bo bo = ctx->get_bo(ctx->user_data, addr);
  89
  90    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0))
  91       bo.addr &= (~0ull >> 16);
  92
  93    /* We may actually have an offset into the bo */
  94    if (bo.map != NULL) {
  95       assert(bo.addr <= addr);
  96       uint64_t offset = addr - bo.addr;
  97       bo.map += offset;
  98       bo.addr += offset;
  99       bo.size -= offset;
 100    }
 101
 102    return bo;
 103 }
 104
 105 static void
 106 ctx_disassemble_program(struct gen_batch_decode_ctx *ctx,
 107                         uint32_t ksp, const char *type)
 108 {
 109    if (!ctx->instruction_base.map)
 110       return;
 111
 112    printf("\nReferenced %s:\n", type);
 113    gen_disasm_disassemble(ctx->disasm,
 114                           (void *)ctx->instruction_base.map, ksp,
 115                           ctx->fp);
 116 }
 117
 118 /* Heuristic to determine whether a uint32_t is probably actually a float
 119  * (http://stackoverflow.com/a/2953466)
 120  */
 121
 122 static bool
 123 probably_float(uint32_t bits)
 124 {
 125    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 126    uint32_t mant = bits & 0x007fffff;
 127
 128    /* +- 0.0 */
 129    if (exp == -127 && mant == 0)
 130       return true;
 131
 132    /* +- 1 billionth to 1 billion */
 133    if (-30 <= exp && exp <= 30)
 134       return true;
 135
 136    /* some value with only a few binary digits */
 137    if ((mant & 0x0000ffff) == 0)
 138       return true;
 139
 140    return false;
 141 }
 142
 143 static void
 144 ctx_print_buffer(struct gen_batch_decode_ctx *ctx,
 145                  struct gen_batch_decode_bo bo,
 146                  uint32_t read_length,
 147                  uint32_t pitch)
 148 {
 149    const uint32_t *dw_end = bo.map + MIN2(bo.size, read_length);
 150
 151    unsigned line_count = 0;
 152    for (const uint32_t *dw = bo.map; dw < dw_end; dw++) {
 153       if (line_count * 4 == pitch || line_count == 8) {
 154          fprintf(ctx->fp, "\n");
 155          line_count = 0;
 156       }
 157       fprintf(ctx->fp, line_count == 0 ? "  " : " ");
 158
 159       if ((ctx->flags & GEN_BATCH_DECODE_FLOATS) && probably_float(*dw))
 160          fprintf(ctx->fp, "  %8.2f", *(float *) dw);
 161       else
 162          fprintf(ctx->fp, "  0x%08x", *dw);
 163
 164       line_count++;
 165    }
 166    fprintf(ctx->fp, "\n");
 167 }
 168
 169 static void
 170 handle_state_base_address(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 171 {
 172    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 173
 174    struct gen_field_iterator iter;
 175    gen_field_iterator_init(&iter, inst, p, 0, false);
 176
 177    do {
 178       if (strcmp(iter.name, "Surface State Base Address") == 0) {
 179          ctx->surface_base = ctx_get_bo(ctx, iter.raw_value);
 180       } else if (strcmp(iter.name, "Dynamic State Base Address") == 0) {
 181          ctx->dynamic_base = ctx_get_bo(ctx, iter.raw_value);
 182       } else if (strcmp(iter.name, "Instruction Base Address") == 0) {
 183          ctx->instruction_base = ctx_get_bo(ctx, iter.raw_value);
 184       }
 185    } while (gen_field_iterator_next(&iter));
 186 }
 187
 188 static void
 189 dump_binding_table(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 190 {
 191    struct gen_group *strct =
 192       gen_spec_find_struct(ctx->spec, "RENDER_SURFACE_STATE");
 193    if (strct == NULL) {
 194       fprintf(ctx->fp, "did not find RENDER_SURFACE_STATE info\n");
 195       return;
 196    }
 197
 198    /* If we don't know the actual count, guess. */
 199    if (count < 0)
 200       count = 8;
 201
 202    if (ctx->surface_base.map == NULL) {
 203       fprintf(ctx->fp, "  binding table unavailable\n");
 204       return;
 205    }
 206
 207    if (offset % 32 != 0 || offset >= UINT16_MAX ||
 208        offset >= ctx->surface_base.size) {
 209       fprintf(ctx->fp, "  invalid binding table pointer\n");
 210       return;
 211    }
 212
 213    const uint32_t *pointers = ctx->surface_base.map + offset;
 214    for (int i = 0; i < count; i++) {
 215       if (pointers[i] == 0)
 216          continue;
 217
 218       if (pointers[i] % 32 != 0 ||
 219           (pointers[i] + strct->dw_length * 4) >= ctx->surface_base.size) {
 220          fprintf(ctx->fp, "pointer %u: %08x <not valid>\n", i, pointers[i]);
 221          continue;
 222       }
 223
 224       fprintf(ctx->fp, "pointer %u: %08x\n", i, pointers[i]);
 225       ctx_print_group(ctx, strct, ctx->surface_base.addr + pointers[i],
 226                       ctx->surface_base.map + pointers[i]);
 227    }
 228 }
 229
 230 static void
 231 dump_samplers(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 232 {
 233    struct gen_group *strct = gen_spec_find_struct(ctx->spec, "SAMPLER_STATE");
 234
 235    /* If we don't know the actual count, guess. */
 236    if (count < 0)
 237       count = 4;
 238
 239    if (ctx->dynamic_base.map == NULL) {
 240       fprintf(ctx->fp, "  samplers unavailable\n");
 241       return;
 242    }
 243
 244    if (offset % 32 != 0 || offset >= ctx->dynamic_base.size) {
 245       fprintf(ctx->fp, "  invalid sampler state pointer\n");
 246       return;
 247    }
 248
 249    uint64_t state_addr = ctx->dynamic_base.addr + offset;
 250    const void *state_map = ctx->dynamic_base.map + offset;
 251    for (int i = 0; i < count; i++) {
 252       fprintf(ctx->fp, "sampler state %d\n", i);
 253       ctx_print_group(ctx, strct, state_addr, state_map);
 254       state_addr += 16;
 255       state_map += 16;
 256    }
 257 }
 258
 259 static void
 260 handle_media_interface_descriptor_load(struct gen_batch_decode_ctx *ctx,
 261                                        const uint32_t *p)
 262 {
 263    if (ctx->dynamic_base.map == NULL)
 264       return;
 265
 266    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 267    struct gen_group *desc =
 268       gen_spec_find_struct(ctx->spec, "INTERFACE_DESCRIPTOR_DATA");
 269
 270    struct gen_field_iterator iter;
 271    gen_field_iterator_init(&iter, inst, p, 0, false);
 272    uint32_t descriptor_offset = 0;
 273    int descriptor_count = 0;
 274    do {
 275       if (strcmp(iter.name, "Interface Descriptor Data Start Address") == 0) {
 276          descriptor_offset = strtol(iter.value, NULL, 16);
 277       } else if (strcmp(iter.name, "Interface Descriptor Total Length") == 0) {
 278          descriptor_count =
 279             strtol(iter.value, NULL, 16) / (desc->dw_length * 4);
 280       }
 281    } while (gen_field_iterator_next(&iter));
 282
 283    uint64_t desc_addr = ctx->dynamic_base.addr + descriptor_offset;
 284    const uint32_t *desc_map = ctx->dynamic_base.map + descriptor_offset;
 285    for (int i = 0; i < descriptor_count; i++) {
 286       fprintf(ctx->fp, "descriptor %d: %08x\n", i, descriptor_offset);
 287
 288       ctx_print_group(ctx, inst, desc_addr, desc_map);
 289
 290       gen_field_iterator_init(&iter, desc, desc_map, 0, false);
 291       uint64_t ksp;
 292       uint32_t sampler_offset, sampler_count;
 293       uint32_t binding_table_offset, binding_entry_count;
 294       do {
 295          if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 296             ksp = strtoll(iter.value, NULL, 16);
 297          } else if (strcmp(iter.name, "Sampler State Pointer") == 0) {
 298             sampler_offset = strtol(iter.value, NULL, 16);
 299          } else if (strcmp(iter.name, "Sampler Count") == 0) {
 300             sampler_count = strtol(iter.value, NULL, 10);
 301          } else if (strcmp(iter.name, "Binding Table Pointer") == 0) {
 302             binding_table_offset = strtol(iter.value, NULL, 16);
 303          } else if (strcmp(iter.name, "Binding Table Entry Count") == 0) {
 304             binding_entry_count = strtol(iter.value, NULL, 10);
 305          }
 306       } while (gen_field_iterator_next(&iter));
 307
 308       ctx_disassemble_program(ctx, ksp, "compute shader");
 309       printf("\n");
 310
 311       dump_samplers(ctx, sampler_offset, sampler_count);
 312       dump_binding_table(ctx, binding_table_offset, binding_entry_count);
 313
 314       desc_map += desc->dw_length;
 315       desc_addr += desc->dw_length * 4;
 316    }
 317 }
 318
 319 static void
 320 handle_3dstate_vertex_buffers(struct gen_batch_decode_ctx *ctx,
 321                               const uint32_t *p)
 322 {
 323    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 324
 325    struct gen_batch_decode_bo vb = {};
 326    uint32_t vb_size = 0;
 327    int index = -1;
 328    int pitch = -1;
 329    bool ready = false;
 330
 331    struct gen_field_iterator iter;
 332    gen_field_iterator_init(&iter, inst, p, 0, false);
 333    do {
 334       if (strcmp(iter.name, "Vertex Buffer Index") == 0) {
 335          index = iter.raw_value;
 336       } else if (strcmp(iter.name, "Buffer Pitch") == 0) {
 337          pitch = iter.raw_value;
 338       } else if (strcmp(iter.name, "Buffer Starting Address") == 0) {
 339          vb = ctx_get_bo(ctx, iter.raw_value);
 340       } else if (strcmp(iter.name, "Buffer Size") == 0) {
 341          vb_size = iter.raw_value;
 342          ready = true;
 343       } else if (strcmp(iter.name, "End Address") == 0) {
 344          if (vb.map && iter.raw_value >= vb.addr)
 345             vb_size = iter.raw_value - vb.addr;
 346          else
 347             vb_size = 0;
 348          ready = true;
 349       }
 350
 351       if (!ready)
 352          continue;
 353
 354       fprintf(ctx->fp, "vertex buffer %d, size %d\n", index, vb_size);
 355
 356       if (vb.map == NULL) {
 357          fprintf(ctx->fp, "  buffer contents unavailable\n");
 358          continue;
 359       }
 360
 361       if (vb.map == 0 || vb_size == 0)
 362          continue;
 363
 364       ctx_print_buffer(ctx, vb, vb_size, pitch);
 365
 366       vb.map = NULL;
 367       vb_size = 0;
 368       index = -1;
 369       pitch = -1;
 370       ready = false;
 371    } while (gen_field_iterator_next(&iter));
 372 }
 373
 374 static void
 375 handle_3dstate_index_buffer(struct gen_batch_decode_ctx *ctx,
 376                             const uint32_t *p)
 377 {
 378    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 379
 380    struct gen_batch_decode_bo ib = {};
 381    uint32_t ib_size = 0;
 382    uint32_t format = 0;
 383
 384    struct gen_field_iterator iter;
 385    gen_field_iterator_init(&iter, inst, p, 0, false);
 386    do {
 387       if (strcmp(iter.name, "Index Format") == 0) {
 388          format = iter.raw_value;
 389       } else if (strcmp(iter.name, "Buffer Starting Address") == 0) {
 390          ib = ctx_get_bo(ctx, iter.raw_value);
 391       } else if (strcmp(iter.name, "Buffer Size") == 0) {
 392          ib_size = iter.raw_value;
 393       }
 394    } while (gen_field_iterator_next(&iter));
 395
 396    if (ib.map == NULL) {
 397       fprintf(ctx->fp, "  buffer contents unavailable\n");
 398       return;
 399    }
 400
 401    const void *m = ib.map;
 402    const void *ib_end = ib.map + MIN2(ib.size, ib_size);
 403    for (int i = 0; m < ib_end && i < 10; i++) {
 404       switch (format) {
 405       case 0:
 406          fprintf(ctx->fp, "%3d ", *(uint8_t *)m);
 407          m += 1;
 408          break;
 409       case 1:
 410          fprintf(ctx->fp, "%3d ", *(uint16_t *)m);
 411          m += 2;
 412          break;
 413       case 2:
 414          fprintf(ctx->fp, "%3d ", *(uint32_t *)m);
 415          m += 4;
 416          break;
 417       }
 418    }
 419
 420    if (m < ib_end)
 421       fprintf(ctx->fp, "...");
 422    fprintf(ctx->fp, "\n");
 423 }
 424
 425 static void
 426 decode_single_ksp(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 427 {
 428    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 429
 430    uint64_t ksp = 0;
 431    bool is_simd8 = false; /* vertex shaders on Gen8+ only */
 432    bool is_enabled = true;
 433
 434    struct gen_field_iterator iter;
 435    gen_field_iterator_init(&iter, inst, p, 0, false);
 436    do {
 437       if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 438          ksp = iter.raw_value;
 439       } else if (strcmp(iter.name, "SIMD8 Dispatch Enable") == 0) {
 440          is_simd8 = iter.raw_value;
 441       } else if (strcmp(iter.name, "Dispatch Mode") == 0) {
 442          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 443       } else if (strcmp(iter.name, "Dispatch Enable") == 0) {
 444          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 445       } else if (strcmp(iter.name, "Enable") == 0) {
 446          is_enabled = iter.raw_value;
 447       }
 448    } while (gen_field_iterator_next(&iter));
 449
 450    const char *type =
 451       strcmp(inst->name,   "VS_STATE") == 0 ? "vertex shader" :
 452       strcmp(inst->name,   "GS_STATE") == 0 ? "geometry shader" :
 453       strcmp(inst->name,   "SF_STATE") == 0 ? "strips and fans shader" :
 454       strcmp(inst->name, "CLIP_STATE") == 0 ? "clip shader" :
 455       strcmp(inst->name, "3DSTATE_DS") == 0 ? "tessellation control shader" :
 456       strcmp(inst->name, "3DSTATE_HS") == 0 ? "tessellation evaluation shader" :
 457       strcmp(inst->name, "3DSTATE_VS") == 0 ? (is_simd8 ? "SIMD8 vertex shader" : "vec4 vertex shader") :
 458       strcmp(inst->name, "3DSTATE_GS") == 0 ? (is_simd8 ? "SIMD8 geometry shader" : "vec4 geometry shader") :
 459       NULL;
 460
 461    if (is_enabled) {
 462       ctx_disassemble_program(ctx, ksp, type);
 463       printf("\n");
 464    }
 465 }
 466
 467 static void
 468 decode_ps_kernels(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 469 {
 470    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 471
 472    uint64_t ksp[3] = {0, 0, 0};
 473    bool enabled[3] = {false, false, false};
 474
 475    struct gen_field_iterator iter;
 476    gen_field_iterator_init(&iter, inst, p, 0, false);
 477    do {
 478       if (strncmp(iter.name, "Kernel Start Pointer ",
 479                   strlen("Kernel Start Pointer ")) == 0) {
 480          int idx = iter.name[strlen("Kernel Start Pointer ")] - '0';
 481          ksp[idx] = strtol(iter.value, NULL, 16);
 482       } else if (strcmp(iter.name, "8 Pixel Dispatch Enable") == 0) {
 483          enabled[0] = strcmp(iter.value, "true") == 0;
 484       } else if (strcmp(iter.name, "16 Pixel Dispatch Enable") == 0) {
 485          enabled[1] = strcmp(iter.value, "true") == 0;
 486       } else if (strcmp(iter.name, "32 Pixel Dispatch Enable") == 0) {
 487          enabled[2] = strcmp(iter.value, "true") == 0;
 488       }
 489    } while (gen_field_iterator_next(&iter));
 490
 491    /* Reorder KSPs to be [8, 16, 32] instead of the hardware order. */
 492    if (enabled[0] + enabled[1] + enabled[2] == 1) {
 493       if (enabled[1]) {
 494          ksp[1] = ksp[0];
 495          ksp[0] = 0;
 496       } else if (enabled[2]) {
 497          ksp[2] = ksp[0];
 498          ksp[0] = 0;
 499       }
 500    } else {
 501       uint64_t tmp = ksp[1];
 502       ksp[1] = ksp[2];
 503       ksp[2] = tmp;
 504    }
 505
 506    if (enabled[0])
 507       ctx_disassemble_program(ctx, ksp[0], "SIMD8 fragment shader");
 508    if (enabled[1])
 509       ctx_disassemble_program(ctx, ksp[1], "SIMD16 fragment shader");
 510    if (enabled[2])
 511       ctx_disassemble_program(ctx, ksp[2], "SIMD32 fragment shader");
 512    fprintf(ctx->fp, "\n");
 513 }
 514
 515 static void
 516 decode_3dstate_constant(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 517 {
 518    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 519
 520    uint32_t read_length[4];
 521    struct gen_batch_decode_bo buffer[4];
 522    memset(buffer, 0, sizeof(buffer));
 523
 524    int rlidx = 0, bidx = 0;
 525
 526    struct gen_field_iterator iter;
 527    gen_field_iterator_init(&iter, inst, p, 0, false);
 528    do {
 529       if (strcmp(iter.name, "Read Length") == 0) {
 530          read_length[rlidx++] = iter.raw_value;
 531       } else if (strcmp(iter.name, "Buffer") == 0) {
 532          buffer[bidx++] = ctx_get_bo(ctx, iter.raw_value);
 533       }
 534    } while (gen_field_iterator_next(&iter));
 535
 536    for (int i = 0; i < 4; i++) {
 537       if (read_length[i] == 0 || buffer[i].map == NULL)
 538          continue;
 539
 540       unsigned size = read_length[i] * 32;
 541       fprintf(ctx->fp, "constant buffer %d, size %u\n", i, size);
 542
 543       ctx_print_buffer(ctx, buffer[i], size, 0);
 544    }
 545 }
 546
 547 static void
 548 decode_3dstate_binding_table_pointers(struct gen_batch_decode_ctx *ctx,
 549                                       const uint32_t *p)
 550 {
 551    dump_binding_table(ctx, p[1], -1);
 552 }
 553
 554 static void
 555 decode_3dstate_sampler_state_pointers(struct gen_batch_decode_ctx *ctx,
 556                                       const uint32_t *p)
 557 {
 558    dump_samplers(ctx, p[1], -1);
 559 }
 560
 561 static void
 562 decode_3dstate_sampler_state_pointers_gen6(struct gen_batch_decode_ctx *ctx,
 563                                            const uint32_t *p)
 564 {
 565    dump_samplers(ctx, p[1], -1);
 566    dump_samplers(ctx, p[2], -1);
 567    dump_samplers(ctx, p[3], -1);
 568 }
 569
 570 static bool
 571 str_ends_with(const char *str, const char *end)
 572 {
 573    int offset = strlen(str) - strlen(end);
 574    if (offset < 0)
 575       return false;
 576
 577    return strcmp(str + offset, end) == 0;
 578 }
 579
 580 static void
 581 decode_dynamic_state_pointers(struct gen_batch_decode_ctx *ctx,
 582                               const char *struct_type, const uint32_t *p,
 583                               int count)
 584 {
 585    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 586    struct gen_group *state = gen_spec_find_struct(ctx->spec, struct_type);
 587
 588    uint32_t state_offset;
 589
 590    struct gen_field_iterator iter;
 591    gen_field_iterator_init(&iter, inst, p, 0, false);
 592    do {
 593       if (str_ends_with(iter.name, "Pointer")) {
 594          state_offset = iter.raw_value;
 595          break;
 596       }
 597    } while (gen_field_iterator_next(&iter));
 598
 599    uint32_t state_addr = ctx->dynamic_base.addr + state_offset;
 600    const uint32_t *state_map = ctx->dynamic_base.map + state_offset;
 601    for (int i = 0; i < count; i++) {
 602       fprintf(ctx->fp, "%s %d\n", struct_type, i);
 603       ctx_print_group(ctx, state, state_offset, state_map);
 604
 605       state_addr += state->dw_length * 4;
 606       state_map += state->dw_length;
 607    }
 608 }
 609
 610 static void
 611 decode_3dstate_viewport_state_pointers_cc(struct gen_batch_decode_ctx *ctx,
 612                                           const uint32_t *p)
 613 {
 614    decode_dynamic_state_pointers(ctx, "CC_VIEWPORT", p, 4);
 615 }
 616
 617 static void
 618 decode_3dstate_viewport_state_pointers_sf_clip(struct gen_batch_decode_ctx *ctx,
 619                                                const uint32_t *p)
 620 {
 621    decode_dynamic_state_pointers(ctx, "SF_CLIP_VIEWPORT", p, 4);
 622 }
 623
 624 static void
 625 decode_3dstate_blend_state_pointers(struct gen_batch_decode_ctx *ctx,
 626                                     const uint32_t *p)
 627 {
 628    decode_dynamic_state_pointers(ctx, "BLEND_STATE", p, 1);
 629 }
 630
 631 static void
 632 decode_3dstate_cc_state_pointers(struct gen_batch_decode_ctx *ctx,
 633                                  const uint32_t *p)
 634 {
 635    decode_dynamic_state_pointers(ctx, "COLOR_CALC_STATE", p, 1);
 636 }
 637
 638 static void
 639 decode_3dstate_scissor_state_pointers(struct gen_batch_decode_ctx *ctx,
 640                                       const uint32_t *p)
 641 {
 642    decode_dynamic_state_pointers(ctx, "SCISSOR_RECT", p, 1);
 643 }
 644
 645 static void
 646 decode_load_register_imm(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 647 {
 648    struct gen_group *reg = gen_spec_find_register(ctx->spec, p[1]);
 649
 650    if (reg != NULL) {
 651       fprintf(ctx->fp, "register %s (0x%x): 0x%x\n",
 652               reg->name, reg->register_offset, p[2]);
 653       ctx_print_group(ctx, reg, reg->register_offset, &p[2]);
 654    }
 655 }
 656
 657 struct custom_decoder {
 658    const char *cmd_name;
 659    void (*decode)(struct gen_batch_decode_ctx *ctx, const uint32_t *p);
 660 } custom_decoders[] = {
 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", decode_single_ksp },
 666    { "3DSTATE_GS", decode_single_ksp },
 667    { "3DSTATE_DS", decode_single_ksp },
 668    { "3DSTATE_HS", decode_single_ksp },
 669    { "3DSTATE_PS", decode_ps_kernels },
 670    { "3DSTATE_CONSTANT_VS", decode_3dstate_constant },
 671    { "3DSTATE_CONSTANT_GS", decode_3dstate_constant },
 672    { "3DSTATE_CONSTANT_PS", decode_3dstate_constant },
 673    { "3DSTATE_CONSTANT_HS", decode_3dstate_constant },
 674    { "3DSTATE_CONSTANT_DS", decode_3dstate_constant },
 675
 676    { "3DSTATE_BINDING_TABLE_POINTERS_VS", decode_3dstate_binding_table_pointers },
 677    { "3DSTATE_BINDING_TABLE_POINTERS_HS", decode_3dstate_binding_table_pointers },
 678    { "3DSTATE_BINDING_TABLE_POINTERS_DS", decode_3dstate_binding_table_pointers },
 679    { "3DSTATE_BINDING_TABLE_POINTERS_GS", decode_3dstate_binding_table_pointers },
 680    { "3DSTATE_BINDING_TABLE_POINTERS_PS", decode_3dstate_binding_table_pointers },
 681
 682    { "3DSTATE_SAMPLER_STATE_POINTERS_VS", decode_3dstate_sampler_state_pointers },
 683    { "3DSTATE_SAMPLER_STATE_POINTERS_HS", decode_3dstate_sampler_state_pointers },
 684    { "3DSTATE_SAMPLER_STATE_POINTERS_DS", decode_3dstate_sampler_state_pointers },
 685    { "3DSTATE_SAMPLER_STATE_POINTERS_GS", decode_3dstate_sampler_state_pointers },
 686    { "3DSTATE_SAMPLER_STATE_POINTERS_PS", decode_3dstate_sampler_state_pointers },
 687    { "3DSTATE_SAMPLER_STATE_POINTERS", decode_3dstate_sampler_state_pointers_gen6 },
 688
 689    { "3DSTATE_VIEWPORT_STATE_POINTERS_CC", decode_3dstate_viewport_state_pointers_cc },
 690    { "3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP", decode_3dstate_viewport_state_pointers_sf_clip },
 691    { "3DSTATE_BLEND_STATE_POINTERS", decode_3dstate_blend_state_pointers },
 692    { "3DSTATE_CC_STATE_POINTERS", decode_3dstate_cc_state_pointers },
 693    { "3DSTATE_SCISSOR_STATE_POINTERS", decode_3dstate_scissor_state_pointers },
 694    { "MI_LOAD_REGISTER_IMM", decode_load_register_imm }
 695 };
 696
 697 static inline uint64_t
 698 get_address(struct gen_spec *spec, const uint32_t *p)
 699 {
 700    /* Addresses are always guaranteed to be page-aligned and sometimes
 701     * hardware packets have extra stuff stuffed in the bottom 12 bits.
 702     */
 703    uint64_t addr = p[0] & ~0xfffu;
 704
 705    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 706       /* On Broadwell and above, we have 48-bit addresses which consume two
 707        * dwords.  Some packets require that these get stored in a "canonical
 708        * form" which means that bit 47 is sign-extended through the upper
 709        * bits. In order to correctly handle those aub dumps, we need to mask
 710        * off the top 16 bits.
 711        */
 712       addr |= ((uint64_t)p[1] & 0xffff) << 32;
 713    }
 714
 715    return addr;
 716 }
 717
 718 void
 719 gen_print_batch(struct gen_batch_decode_ctx *ctx,
 720                 const uint32_t *batch, uint32_t batch_size,
 721                 uint64_t batch_addr)
 722 {
 723    const uint32_t *p, *end = batch + batch_size;
 724    int length;
 725    struct gen_group *inst;
 726
 727    for (p = batch; p < end; p += length) {
 728       inst = gen_spec_find_instruction(ctx->spec, p);
 729       length = gen_group_get_length(inst, p);
 730       assert(inst == NULL || length > 0);
 731       length = MAX2(1, length);
 732       if (inst == NULL) {
 733          fprintf(ctx->fp, "unknown instruction %08x\n", p[0]);
 734          continue;
 735       }
 736
 737       const char *color, *reset_color;
 738       uint64_t offset;
 739
 740       const char *inst_name = gen_group_get_name(inst);
 741       if (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) {
 742          reset_color = NORMAL;
 743          if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 744             if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0 ||
 745                 strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0)
 746                color = GREEN_HEADER;
 747             else
 748                color = BLUE_HEADER;
 749          } else {
 750             color = NORMAL;
 751          }
 752       } else {
 753          color = "";
 754          reset_color = "";
 755       }
 756
 757       if (ctx->flags & GEN_BATCH_DECODE_OFFSETS)
 758          offset = batch_addr + ((char *)p - (char *)batch);
 759       else
 760          offset = 0;
 761
 762       fprintf(ctx->fp, "%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 763               color, offset, p[0], inst_name, reset_color);
 764
 765       if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 766          ctx_print_group(ctx, inst, offset, p);
 767
 768          for (int i = 0; i < ARRAY_LENGTH(custom_decoders); i++) {
 769             if (strcmp(inst_name, custom_decoders[i].cmd_name) == 0) {
 770                custom_decoders[i].decode(ctx, p);
 771                break;
 772             }
 773          }
 774       }
 775
 776       if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0) {
 777          struct gen_batch_decode_bo next_batch;
 778          bool second_level;
 779          struct gen_field_iterator iter;
 780          gen_field_iterator_init(&iter, inst, p, 0, false);
 781          do {
 782             if (strcmp(iter.name, "Batch Buffer Start Address") == 0) {
 783                next_batch = ctx_get_bo(ctx, iter.raw_value);
 784             } else if (strcmp(iter.name, "Second Level Batch Buffer") == 0) {
 785                second_level = iter.raw_value;
 786             }
 787          } while (gen_field_iterator_next(&iter));
 788
 789          if (next_batch.map == NULL) {
 790             fprintf(ctx->fp, "Secondary batch at 0x%08"PRIx64" unavailable",
 791                     next_batch.addr);
 792          }
 793
 794          if (second_level) {
 795             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 796              * like a subroutine call.  Commands that come afterwards get
 797              * processed once the 2nd level batch buffer returns with
 798              * MI_BATCH_BUFFER_END.
 799              */
 800             if (next_batch.map) {
 801                gen_print_batch(ctx, next_batch.map, next_batch.size,
 802                                next_batch.addr);
 803             }
 804          } else {
 805             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 806              * like a goto.  Nothing after it will ever get processed.  In
 807              * order to prevent the recursion from growing, we just reset the
 808              * loop and continue;
 809              */
 810             if (next_batch.map) {
 811                p = next_batch.map;
 812                end = next_batch.map + next_batch.size;
 813                length = 0;
 814                continue;
 815             } else {
 816                /* Nothing we can do */
 817                break;
 818             }
 819          }
 820       } else if (strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0) {
 821          break;
 822       }
 823    }
 824 }