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 RED_COLOR    CSI "31m"
  61 #define BLUE_HEADER  CSI "0;44m"
  62 #define GREEN_HEADER CSI "1;42m"
  63 #define NORMAL       CSI "0m"
  64
  65 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
  66
  67 static void
  68 ctx_print_group(struct gen_batch_decode_ctx *ctx,
  69                 struct gen_group *group,
  70                 uint64_t address, const void *map)
  71 {
  72    gen_print_group(ctx->fp, group, address, map, 0,
  73                    (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) != 0);
  74 }
  75
  76 static struct gen_batch_decode_bo
  77 ctx_get_bo(struct gen_batch_decode_ctx *ctx, uint64_t addr)
  78 {
  79    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0)) {
  80       /* On Broadwell and above, we have 48-bit addresses which consume two
  81        * dwords.  Some packets require that these get stored in a "canonical
  82        * form" which means that bit 47 is sign-extended through the upper
  83        * bits. In order to correctly handle those aub dumps, we need to mask
  84        * off the top 16 bits.
  85        */
  86       addr &= (~0ull >> 16);
  87    }
  88
  89    struct gen_batch_decode_bo bo = ctx->get_bo(ctx->user_data, addr);
  90
  91    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0))
  92       bo.addr &= (~0ull >> 16);
  93
  94    /* We may actually have an offset into the bo */
  95    if (bo.map != NULL) {
  96       assert(bo.addr <= addr);
  97       uint64_t offset = addr - bo.addr;
  98       bo.map += offset;
  99       bo.addr += offset;
 100       bo.size -= offset;
 101    }
 102
 103    return bo;
 104 }
 105
 106 static void
 107 ctx_disassemble_program(struct gen_batch_decode_ctx *ctx,
 108                         uint32_t ksp, const char *type)
 109 {
 110    if (!ctx->instruction_base.map)
 111       return;
 112
 113    printf("\nReferenced %s:\n", type);
 114    gen_disasm_disassemble(ctx->disasm,
 115                           (void *)ctx->instruction_base.map, ksp,
 116                           ctx->fp);
 117 }
 118
 119 /* Heuristic to determine whether a uint32_t is probably actually a float
 120  * (http://stackoverflow.com/a/2953466)
 121  */
 122
 123 static bool
 124 probably_float(uint32_t bits)
 125 {
 126    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 127    uint32_t mant = bits & 0x007fffff;
 128
 129    /* +- 0.0 */
 130    if (exp == -127 && mant == 0)
 131       return true;
 132
 133    /* +- 1 billionth to 1 billion */
 134    if (-30 <= exp && exp <= 30)
 135       return true;
 136
 137    /* some value with only a few binary digits */
 138    if ((mant & 0x0000ffff) == 0)
 139       return true;
 140
 141    return false;
 142 }
 143
 144 static void
 145 ctx_print_buffer(struct gen_batch_decode_ctx *ctx,
 146                  struct gen_batch_decode_bo bo,
 147                  uint32_t read_length,
 148                  uint32_t pitch)
 149 {
 150    const uint32_t *dw_end = bo.map + MIN2(bo.size, read_length);
 151
 152    unsigned line_count = 0;
 153    for (const uint32_t *dw = bo.map; dw < dw_end; dw++) {
 154       if (line_count * 4 == pitch || line_count == 8) {
 155          fprintf(ctx->fp, "\n");
 156          line_count = 0;
 157       }
 158       fprintf(ctx->fp, line_count == 0 ? "  " : " ");
 159
 160       if ((ctx->flags & GEN_BATCH_DECODE_FLOATS) && probably_float(*dw))
 161          fprintf(ctx->fp, "  %8.2f", *(float *) dw);
 162       else
 163          fprintf(ctx->fp, "  0x%08x", *dw);
 164
 165       line_count++;
 166    }
 167    fprintf(ctx->fp, "\n");
 168 }
 169
 170 static void
 171 handle_state_base_address(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 172 {
 173    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 174
 175    struct gen_field_iterator iter;
 176    gen_field_iterator_init(&iter, inst, p, 0, false);
 177
 178    while (gen_field_iterator_next(&iter)) {
 179       if (strcmp(iter.name, "Surface State Base Address") == 0) {
 180          ctx->surface_base = ctx_get_bo(ctx, iter.raw_value);
 181       } else if (strcmp(iter.name, "Dynamic State Base Address") == 0) {
 182          ctx->dynamic_base = ctx_get_bo(ctx, iter.raw_value);
 183       } else if (strcmp(iter.name, "Instruction Base Address") == 0) {
 184          ctx->instruction_base = ctx_get_bo(ctx, iter.raw_value);
 185       }
 186    }
 187 }
 188
 189 static void
 190 dump_binding_table(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 191 {
 192    struct gen_group *strct =
 193       gen_spec_find_struct(ctx->spec, "RENDER_SURFACE_STATE");
 194    if (strct == NULL) {
 195       fprintf(ctx->fp, "did not find RENDER_SURFACE_STATE info\n");
 196       return;
 197    }
 198
 199    /* If we don't know the actual count, guess. */
 200    if (count < 0)
 201       count = 8;
 202
 203    if (ctx->surface_base.map == NULL) {
 204       fprintf(ctx->fp, "  binding table unavailable\n");
 205       return;
 206    }
 207
 208    if (offset % 32 != 0 || offset >= UINT16_MAX ||
 209        offset >= ctx->surface_base.size) {
 210       fprintf(ctx->fp, "  invalid binding table pointer\n");
 211       return;
 212    }
 213
 214    const uint32_t *pointers = ctx->surface_base.map + offset;
 215    for (int i = 0; i < count; i++) {
 216       if (pointers[i] == 0)
 217          continue;
 218
 219       if (pointers[i] % 32 != 0 ||
 220           (pointers[i] + strct->dw_length * 4) >= ctx->surface_base.size) {
 221          fprintf(ctx->fp, "pointer %u: %08x <not valid>\n", i, pointers[i]);
 222          continue;
 223       }
 224
 225       fprintf(ctx->fp, "pointer %u: %08x\n", i, pointers[i]);
 226       ctx_print_group(ctx, strct, ctx->surface_base.addr + pointers[i],
 227                       ctx->surface_base.map + pointers[i]);
 228    }
 229 }
 230
 231 static void
 232 dump_samplers(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 233 {
 234    struct gen_group *strct = gen_spec_find_struct(ctx->spec, "SAMPLER_STATE");
 235
 236    /* If we don't know the actual count, guess. */
 237    if (count < 0)
 238       count = 4;
 239
 240    if (ctx->dynamic_base.map == NULL) {
 241       fprintf(ctx->fp, "  samplers unavailable\n");
 242       return;
 243    }
 244
 245    if (offset % 32 != 0 || offset >= ctx->dynamic_base.size) {
 246       fprintf(ctx->fp, "  invalid sampler state pointer\n");
 247       return;
 248    }
 249
 250    uint64_t state_addr = ctx->dynamic_base.addr + offset;
 251    const void *state_map = ctx->dynamic_base.map + offset;
 252    for (int i = 0; i < count; i++) {
 253       fprintf(ctx->fp, "sampler state %d\n", i);
 254       ctx_print_group(ctx, strct, state_addr, state_map);
 255       state_addr += 16;
 256       state_map += 16;
 257    }
 258 }
 259
 260 static void
 261 handle_media_interface_descriptor_load(struct gen_batch_decode_ctx *ctx,
 262                                        const uint32_t *p)
 263 {
 264    if (ctx->dynamic_base.map == NULL)
 265       return;
 266
 267    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 268    struct gen_group *desc =
 269       gen_spec_find_struct(ctx->spec, "INTERFACE_DESCRIPTOR_DATA");
 270
 271    struct gen_field_iterator iter;
 272    gen_field_iterator_init(&iter, inst, p, 0, false);
 273    uint32_t descriptor_offset = 0;
 274    int descriptor_count = 0;
 275    while (gen_field_iterator_next(&iter)) {
 276       if (strcmp(iter.name, "Interface Descriptor Data Start Address") == 0) {
 277          descriptor_offset = strtol(iter.value, NULL, 16);
 278       } else if (strcmp(iter.name, "Interface Descriptor Total Length") == 0) {
 279          descriptor_count =
 280             strtol(iter.value, NULL, 16) / (desc->dw_length * 4);
 281       }
 282    }
 283
 284    uint64_t desc_addr = ctx->dynamic_base.addr + descriptor_offset;
 285    const uint32_t *desc_map = ctx->dynamic_base.map + descriptor_offset;
 286    for (int i = 0; i < descriptor_count; i++) {
 287       fprintf(ctx->fp, "descriptor %d: %08x\n", i, descriptor_offset);
 288
 289       ctx_print_group(ctx, desc, desc_addr, desc_map);
 290
 291       gen_field_iterator_init(&iter, desc, desc_map, 0, false);
 292       uint64_t ksp;
 293       uint32_t sampler_offset, sampler_count;
 294       uint32_t binding_table_offset, binding_entry_count;
 295       while (gen_field_iterator_next(&iter)) {
 296          if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 297             ksp = strtoll(iter.value, NULL, 16);
 298          } else if (strcmp(iter.name, "Sampler State Pointer") == 0) {
 299             sampler_offset = strtol(iter.value, NULL, 16);
 300          } else if (strcmp(iter.name, "Sampler Count") == 0) {
 301             sampler_count = strtol(iter.value, NULL, 10);
 302          } else if (strcmp(iter.name, "Binding Table Pointer") == 0) {
 303             binding_table_offset = strtol(iter.value, NULL, 16);
 304          } else if (strcmp(iter.name, "Binding Table Entry Count") == 0) {
 305             binding_entry_count = strtol(iter.value, NULL, 10);
 306          }
 307       }
 308
 309       ctx_disassemble_program(ctx, ksp, "compute shader");
 310       printf("\n");
 311
 312       dump_samplers(ctx, sampler_offset, sampler_count);
 313       dump_binding_table(ctx, binding_table_offset, binding_entry_count);
 314
 315       desc_map += desc->dw_length;
 316       desc_addr += desc->dw_length * 4;
 317    }
 318 }
 319
 320 static void
 321 handle_3dstate_vertex_buffers(struct gen_batch_decode_ctx *ctx,
 322                               const uint32_t *p)
 323 {
 324    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 325    struct gen_group *vbs = gen_spec_find_struct(ctx->spec, "VERTEX_BUFFER_STATE");
 326
 327    struct gen_batch_decode_bo vb = {};
 328    uint32_t vb_size = 0;
 329    int index = -1;
 330    int pitch = -1;
 331    bool ready = false;
 332
 333    struct gen_field_iterator iter;
 334    gen_field_iterator_init(&iter, inst, p, 0, false);
 335    while (gen_field_iterator_next(&iter)) {
 336       if (iter.struct_desc != vbs)
 337          continue;
 338
 339       struct gen_field_iterator vbs_iter;
 340       gen_field_iterator_init(&vbs_iter, vbs, &iter.p[iter.start_bit / 32], 0, false);
 341       while (gen_field_iterator_next(&vbs_iter)) {
 342          if (strcmp(vbs_iter.name, "Vertex Buffer Index") == 0) {
 343             index = vbs_iter.raw_value;
 344          } else if (strcmp(vbs_iter.name, "Buffer Pitch") == 0) {
 345             pitch = vbs_iter.raw_value;
 346          } else if (strcmp(vbs_iter.name, "Buffer Starting Address") == 0) {
 347             vb = ctx_get_bo(ctx, vbs_iter.raw_value);
 348          } else if (strcmp(vbs_iter.name, "Buffer Size") == 0) {
 349             vb_size = vbs_iter.raw_value;
 350             ready = true;
 351          } else if (strcmp(vbs_iter.name, "End Address") == 0) {
 352             if (vb.map && vbs_iter.raw_value >= vb.addr)
 353                vb_size = vbs_iter.raw_value - vb.addr;
 354             else
 355                vb_size = 0;
 356             ready = true;
 357          }
 358
 359          if (!ready)
 360             continue;
 361
 362          fprintf(ctx->fp, "vertex buffer %d, size %d\n", index, vb_size);
 363
 364          if (vb.map == NULL) {
 365             fprintf(ctx->fp, "  buffer contents unavailable\n");
 366             continue;
 367          }
 368
 369          if (vb.map == 0 || vb_size == 0)
 370             continue;
 371
 372          ctx_print_buffer(ctx, vb, vb_size, pitch);
 373
 374          vb.map = NULL;
 375          vb_size = 0;
 376          index = -1;
 377          pitch = -1;
 378          ready = false;
 379       }
 380    }
 381 }
 382
 383 static void
 384 handle_3dstate_index_buffer(struct gen_batch_decode_ctx *ctx,
 385                             const uint32_t *p)
 386 {
 387    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 388
 389    struct gen_batch_decode_bo ib = {};
 390    uint32_t ib_size = 0;
 391    uint32_t format = 0;
 392
 393    struct gen_field_iterator iter;
 394    gen_field_iterator_init(&iter, inst, p, 0, false);
 395    while (gen_field_iterator_next(&iter)) {
 396       if (strcmp(iter.name, "Index Format") == 0) {
 397          format = iter.raw_value;
 398       } else if (strcmp(iter.name, "Buffer Starting Address") == 0) {
 399          ib = ctx_get_bo(ctx, iter.raw_value);
 400       } else if (strcmp(iter.name, "Buffer Size") == 0) {
 401          ib_size = iter.raw_value;
 402       }
 403    }
 404
 405    if (ib.map == NULL) {
 406       fprintf(ctx->fp, "  buffer contents unavailable\n");
 407       return;
 408    }
 409
 410    const void *m = ib.map;
 411    const void *ib_end = ib.map + MIN2(ib.size, ib_size);
 412    for (int i = 0; m < ib_end && i < 10; i++) {
 413       switch (format) {
 414       case 0:
 415          fprintf(ctx->fp, "%3d ", *(uint8_t *)m);
 416          m += 1;
 417          break;
 418       case 1:
 419          fprintf(ctx->fp, "%3d ", *(uint16_t *)m);
 420          m += 2;
 421          break;
 422       case 2:
 423          fprintf(ctx->fp, "%3d ", *(uint32_t *)m);
 424          m += 4;
 425          break;
 426       }
 427    }
 428
 429    if (m < ib_end)
 430       fprintf(ctx->fp, "...");
 431    fprintf(ctx->fp, "\n");
 432 }
 433
 434 static void
 435 decode_single_ksp(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 436 {
 437    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 438
 439    uint64_t ksp = 0;
 440    bool is_simd8 = false; /* vertex shaders on Gen8+ only */
 441    bool is_enabled = true;
 442
 443    struct gen_field_iterator iter;
 444    gen_field_iterator_init(&iter, inst, p, 0, false);
 445    while (gen_field_iterator_next(&iter)) {
 446       if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 447          ksp = iter.raw_value;
 448       } else if (strcmp(iter.name, "SIMD8 Dispatch Enable") == 0) {
 449          is_simd8 = iter.raw_value;
 450       } else if (strcmp(iter.name, "Dispatch Mode") == 0) {
 451          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 452       } else if (strcmp(iter.name, "Dispatch Enable") == 0) {
 453          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 454       } else if (strcmp(iter.name, "Enable") == 0) {
 455          is_enabled = iter.raw_value;
 456       }
 457    }
 458
 459    const char *type =
 460       strcmp(inst->name,   "VS_STATE") == 0 ? "vertex shader" :
 461       strcmp(inst->name,   "GS_STATE") == 0 ? "geometry shader" :
 462       strcmp(inst->name,   "SF_STATE") == 0 ? "strips and fans shader" :
 463       strcmp(inst->name, "CLIP_STATE") == 0 ? "clip shader" :
 464       strcmp(inst->name, "3DSTATE_DS") == 0 ? "tessellation evaluation shader" :
 465       strcmp(inst->name, "3DSTATE_HS") == 0 ? "tessellation control shader" :
 466       strcmp(inst->name, "3DSTATE_VS") == 0 ? (is_simd8 ? "SIMD8 vertex shader" : "vec4 vertex shader") :
 467       strcmp(inst->name, "3DSTATE_GS") == 0 ? (is_simd8 ? "SIMD8 geometry shader" : "vec4 geometry shader") :
 468       NULL;
 469
 470    if (is_enabled) {
 471       ctx_disassemble_program(ctx, ksp, type);
 472       printf("\n");
 473    }
 474 }
 475
 476 static void
 477 decode_ps_kernels(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 478 {
 479    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 480
 481    uint64_t ksp[3] = {0, 0, 0};
 482    bool enabled[3] = {false, false, false};
 483
 484    struct gen_field_iterator iter;
 485    gen_field_iterator_init(&iter, inst, p, 0, false);
 486    while (gen_field_iterator_next(&iter)) {
 487       if (strncmp(iter.name, "Kernel Start Pointer ",
 488                   strlen("Kernel Start Pointer ")) == 0) {
 489          int idx = iter.name[strlen("Kernel Start Pointer ")] - '0';
 490          ksp[idx] = strtol(iter.value, NULL, 16);
 491       } else if (strcmp(iter.name, "8 Pixel Dispatch Enable") == 0) {
 492          enabled[0] = strcmp(iter.value, "true") == 0;
 493       } else if (strcmp(iter.name, "16 Pixel Dispatch Enable") == 0) {
 494          enabled[1] = strcmp(iter.value, "true") == 0;
 495       } else if (strcmp(iter.name, "32 Pixel Dispatch Enable") == 0) {
 496          enabled[2] = strcmp(iter.value, "true") == 0;
 497       }
 498    }
 499
 500    /* Reorder KSPs to be [8, 16, 32] instead of the hardware order. */
 501    if (enabled[0] + enabled[1] + enabled[2] == 1) {
 502       if (enabled[1]) {
 503          ksp[1] = ksp[0];
 504          ksp[0] = 0;
 505       } else if (enabled[2]) {
 506          ksp[2] = ksp[0];
 507          ksp[0] = 0;
 508       }
 509    } else {
 510       uint64_t tmp = ksp[1];
 511       ksp[1] = ksp[2];
 512       ksp[2] = tmp;
 513    }
 514
 515    if (enabled[0])
 516       ctx_disassemble_program(ctx, ksp[0], "SIMD8 fragment shader");
 517    if (enabled[1])
 518       ctx_disassemble_program(ctx, ksp[1], "SIMD16 fragment shader");
 519    if (enabled[2])
 520       ctx_disassemble_program(ctx, ksp[2], "SIMD32 fragment shader");
 521    fprintf(ctx->fp, "\n");
 522 }
 523
 524 static void
 525 decode_3dstate_constant(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 526 {
 527    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 528
 529    uint32_t read_length[4];
 530    struct gen_batch_decode_bo buffer[4];
 531    memset(buffer, 0, sizeof(buffer));
 532
 533    int rlidx = 0, bidx = 0;
 534
 535    struct gen_field_iterator iter;
 536    gen_field_iterator_init(&iter, inst, p, 0, false);
 537    while (gen_field_iterator_next(&iter)) {
 538       if (strcmp(iter.name, "Read Length") == 0) {
 539          read_length[rlidx++] = iter.raw_value;
 540       } else if (strcmp(iter.name, "Buffer") == 0) {
 541          buffer[bidx++] = ctx_get_bo(ctx, iter.raw_value);
 542       }
 543    }
 544
 545    for (int i = 0; i < 4; i++) {
 546       if (read_length[i] == 0 || buffer[i].map == NULL)
 547          continue;
 548
 549       unsigned size = read_length[i] * 32;
 550       fprintf(ctx->fp, "constant buffer %d, size %u\n", i, size);
 551
 552       ctx_print_buffer(ctx, buffer[i], size, 0);
 553    }
 554 }
 555
 556 static void
 557 decode_3dstate_binding_table_pointers(struct gen_batch_decode_ctx *ctx,
 558                                       const uint32_t *p)
 559 {
 560    dump_binding_table(ctx, p[1], -1);
 561 }
 562
 563 static void
 564 decode_3dstate_sampler_state_pointers(struct gen_batch_decode_ctx *ctx,
 565                                       const uint32_t *p)
 566 {
 567    dump_samplers(ctx, p[1], -1);
 568 }
 569
 570 static void
 571 decode_3dstate_sampler_state_pointers_gen6(struct gen_batch_decode_ctx *ctx,
 572                                            const uint32_t *p)
 573 {
 574    dump_samplers(ctx, p[1], -1);
 575    dump_samplers(ctx, p[2], -1);
 576    dump_samplers(ctx, p[3], -1);
 577 }
 578
 579 static bool
 580 str_ends_with(const char *str, const char *end)
 581 {
 582    int offset = strlen(str) - strlen(end);
 583    if (offset < 0)
 584       return false;
 585
 586    return strcmp(str + offset, end) == 0;
 587 }
 588
 589 static void
 590 decode_dynamic_state_pointers(struct gen_batch_decode_ctx *ctx,
 591                               const char *struct_type, const uint32_t *p,
 592                               int count)
 593 {
 594    if (ctx->dynamic_base.map == NULL) {
 595       fprintf(ctx->fp, "  dynamic %s state unavailable\n", struct_type);
 596       return;
 597    }
 598
 599    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 600    struct gen_group *state = gen_spec_find_struct(ctx->spec, struct_type);
 601
 602    uint32_t state_offset;
 603
 604    struct gen_field_iterator iter;
 605    gen_field_iterator_init(&iter, inst, p, 0, false);
 606    while (gen_field_iterator_next(&iter)) {
 607       if (str_ends_with(iter.name, "Pointer")) {
 608          state_offset = iter.raw_value;
 609          break;
 610       }
 611    }
 612
 613    uint32_t state_addr = ctx->dynamic_base.addr + state_offset;
 614    const uint32_t *state_map = ctx->dynamic_base.map + state_offset;
 615    for (int i = 0; i < count; i++) {
 616       fprintf(ctx->fp, "%s %d\n", struct_type, i);
 617       ctx_print_group(ctx, state, state_offset, state_map);
 618
 619       state_addr += state->dw_length * 4;
 620       state_map += state->dw_length;
 621    }
 622 }
 623
 624 static void
 625 decode_3dstate_viewport_state_pointers_cc(struct gen_batch_decode_ctx *ctx,
 626                                           const uint32_t *p)
 627 {
 628    decode_dynamic_state_pointers(ctx, "CC_VIEWPORT", p, 4);
 629 }
 630
 631 static void
 632 decode_3dstate_viewport_state_pointers_sf_clip(struct gen_batch_decode_ctx *ctx,
 633                                                const uint32_t *p)
 634 {
 635    decode_dynamic_state_pointers(ctx, "SF_CLIP_VIEWPORT", p, 4);
 636 }
 637
 638 static void
 639 decode_3dstate_blend_state_pointers(struct gen_batch_decode_ctx *ctx,
 640                                     const uint32_t *p)
 641 {
 642    decode_dynamic_state_pointers(ctx, "BLEND_STATE", p, 1);
 643 }
 644
 645 static void
 646 decode_3dstate_cc_state_pointers(struct gen_batch_decode_ctx *ctx,
 647                                  const uint32_t *p)
 648 {
 649    decode_dynamic_state_pointers(ctx, "COLOR_CALC_STATE", p, 1);
 650 }
 651
 652 static void
 653 decode_3dstate_scissor_state_pointers(struct gen_batch_decode_ctx *ctx,
 654                                       const uint32_t *p)
 655 {
 656    decode_dynamic_state_pointers(ctx, "SCISSOR_RECT", p, 1);
 657 }
 658
 659 static void
 660 decode_load_register_imm(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 661 {
 662    struct gen_group *reg = gen_spec_find_register(ctx->spec, p[1]);
 663
 664    if (reg != NULL) {
 665       fprintf(ctx->fp, "register %s (0x%x): 0x%x\n",
 666               reg->name, reg->register_offset, p[2]);
 667       ctx_print_group(ctx, reg, reg->register_offset, &p[2]);
 668    }
 669 }
 670
 671 struct custom_decoder {
 672    const char *cmd_name;
 673    void (*decode)(struct gen_batch_decode_ctx *ctx, const uint32_t *p);
 674 } custom_decoders[] = {
 675    { "STATE_BASE_ADDRESS", handle_state_base_address },
 676    { "MEDIA_INTERFACE_DESCRIPTOR_LOAD", handle_media_interface_descriptor_load },
 677    { "3DSTATE_VERTEX_BUFFERS", handle_3dstate_vertex_buffers },
 678    { "3DSTATE_INDEX_BUFFER", handle_3dstate_index_buffer },
 679    { "3DSTATE_VS", decode_single_ksp },
 680    { "3DSTATE_GS", decode_single_ksp },
 681    { "3DSTATE_DS", decode_single_ksp },
 682    { "3DSTATE_HS", decode_single_ksp },
 683    { "3DSTATE_PS", decode_ps_kernels },
 684    { "3DSTATE_CONSTANT_VS", decode_3dstate_constant },
 685    { "3DSTATE_CONSTANT_GS", decode_3dstate_constant },
 686    { "3DSTATE_CONSTANT_PS", decode_3dstate_constant },
 687    { "3DSTATE_CONSTANT_HS", decode_3dstate_constant },
 688    { "3DSTATE_CONSTANT_DS", decode_3dstate_constant },
 689
 690    { "3DSTATE_BINDING_TABLE_POINTERS_VS", decode_3dstate_binding_table_pointers },
 691    { "3DSTATE_BINDING_TABLE_POINTERS_HS", decode_3dstate_binding_table_pointers },
 692    { "3DSTATE_BINDING_TABLE_POINTERS_DS", decode_3dstate_binding_table_pointers },
 693    { "3DSTATE_BINDING_TABLE_POINTERS_GS", decode_3dstate_binding_table_pointers },
 694    { "3DSTATE_BINDING_TABLE_POINTERS_PS", decode_3dstate_binding_table_pointers },
 695
 696    { "3DSTATE_SAMPLER_STATE_POINTERS_VS", decode_3dstate_sampler_state_pointers },
 697    { "3DSTATE_SAMPLER_STATE_POINTERS_HS", decode_3dstate_sampler_state_pointers },
 698    { "3DSTATE_SAMPLER_STATE_POINTERS_DS", decode_3dstate_sampler_state_pointers },
 699    { "3DSTATE_SAMPLER_STATE_POINTERS_GS", decode_3dstate_sampler_state_pointers },
 700    { "3DSTATE_SAMPLER_STATE_POINTERS_PS", decode_3dstate_sampler_state_pointers },
 701    { "3DSTATE_SAMPLER_STATE_POINTERS", decode_3dstate_sampler_state_pointers_gen6 },
 702
 703    { "3DSTATE_VIEWPORT_STATE_POINTERS_CC", decode_3dstate_viewport_state_pointers_cc },
 704    { "3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP", decode_3dstate_viewport_state_pointers_sf_clip },
 705    { "3DSTATE_BLEND_STATE_POINTERS", decode_3dstate_blend_state_pointers },
 706    { "3DSTATE_CC_STATE_POINTERS", decode_3dstate_cc_state_pointers },
 707    { "3DSTATE_SCISSOR_STATE_POINTERS", decode_3dstate_scissor_state_pointers },
 708    { "MI_LOAD_REGISTER_IMM", decode_load_register_imm }
 709 };
 710
 711 static inline uint64_t
 712 get_address(struct gen_spec *spec, const uint32_t *p)
 713 {
 714    /* Addresses are always guaranteed to be page-aligned and sometimes
 715     * hardware packets have extra stuff stuffed in the bottom 12 bits.
 716     */
 717    uint64_t addr = p[0] & ~0xfffu;
 718
 719    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 720       /* On Broadwell and above, we have 48-bit addresses which consume two
 721        * dwords.  Some packets require that these get stored in a "canonical
 722        * form" which means that bit 47 is sign-extended through the upper
 723        * bits. In order to correctly handle those aub dumps, we need to mask
 724        * off the top 16 bits.
 725        */
 726       addr |= ((uint64_t)p[1] & 0xffff) << 32;
 727    }
 728
 729    return addr;
 730 }
 731
 732 void
 733 gen_print_batch(struct gen_batch_decode_ctx *ctx,
 734                 const uint32_t *batch, uint32_t batch_size,
 735                 uint64_t batch_addr)
 736 {
 737    const uint32_t *p, *end = batch + batch_size;
 738    int length;
 739    struct gen_group *inst;
 740
 741    for (p = batch; p < end; p += length) {
 742       inst = gen_spec_find_instruction(ctx->spec, p);
 743       length = gen_group_get_length(inst, p);
 744       assert(inst == NULL || length > 0);
 745       length = MAX2(1, length);
 746
 747       const char *reset_color = ctx->flags & GEN_BATCH_DECODE_IN_COLOR ? NORMAL : "";
 748
 749       uint64_t offset;
 750       if (ctx->flags & GEN_BATCH_DECODE_OFFSETS)
 751          offset = batch_addr + ((char *)p - (char *)batch);
 752       else
 753          offset = 0;
 754
 755       if (inst == NULL) {
 756          fprintf(ctx->fp, "%s0x%08"PRIx64": unknown instruction %08x%s\n",
 757                  (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) ? RED_COLOR : "",
 758                  offset, p[0], reset_color);
 759          continue;
 760       }
 761
 762       const char *color;
 763       const char *inst_name = gen_group_get_name(inst);
 764       if (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) {
 765          reset_color = NORMAL;
 766          if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 767             if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0 ||
 768                 strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0)
 769                color = GREEN_HEADER;
 770             else
 771                color = BLUE_HEADER;
 772          } else {
 773             color = NORMAL;
 774          }
 775       } else {
 776          color = "";
 777          reset_color = "";
 778       }
 779
 780       fprintf(ctx->fp, "%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 781               color, offset, p[0], inst_name, reset_color);
 782
 783       if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 784          ctx_print_group(ctx, inst, offset, p);
 785
 786          for (int i = 0; i < ARRAY_LENGTH(custom_decoders); i++) {
 787             if (strcmp(inst_name, custom_decoders[i].cmd_name) == 0) {
 788                custom_decoders[i].decode(ctx, p);
 789                break;
 790             }
 791          }
 792       }
 793
 794       if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0) {
 795          struct gen_batch_decode_bo next_batch;
 796          bool second_level;
 797          struct gen_field_iterator iter;
 798          gen_field_iterator_init(&iter, inst, p, 0, false);
 799          while (gen_field_iterator_next(&iter)) {
 800             if (strcmp(iter.name, "Batch Buffer Start Address") == 0) {
 801                next_batch = ctx_get_bo(ctx, iter.raw_value);
 802             } else if (strcmp(iter.name, "Second Level Batch Buffer") == 0) {
 803                second_level = iter.raw_value;
 804             }
 805          }
 806
 807          if (next_batch.map == NULL) {
 808             fprintf(ctx->fp, "Secondary batch at 0x%08"PRIx64" unavailable",
 809                     next_batch.addr);
 810          }
 811
 812          if (second_level) {
 813             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 814              * like a subroutine call.  Commands that come afterwards get
 815              * processed once the 2nd level batch buffer returns with
 816              * MI_BATCH_BUFFER_END.
 817              */
 818             if (next_batch.map) {
 819                gen_print_batch(ctx, next_batch.map, next_batch.size,
 820                                next_batch.addr);
 821             }
 822          } else {
 823             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 824              * like a goto.  Nothing after it will ever get processed.  In
 825              * order to prevent the recursion from growing, we just reset the
 826              * loop and continue;
 827              */
 828             if (next_batch.map) {
 829                p = next_batch.map;
 830                end = next_batch.map + next_batch.size;
 831                length = 0;
 832                continue;
 833             } else {
 834                /* Nothing we can do */
 835                break;
 836             }
 837          }
 838       } else if (strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0) {
 839          break;
 840       }
 841    }
 842 }