src/intel/common/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                           unsigned (*get_state_size)(void *, uint32_t),
  37                           void *user_data)
  38 {
  39    memset(ctx, 0, sizeof(*ctx));
  40
  41    ctx->get_bo = get_bo;
  42    ctx->get_state_size = get_state_size;
  43    ctx->user_data = user_data;
  44    ctx->fp = fp;
  45    ctx->flags = flags;
  46
  47    if (xml_path == NULL)
  48       ctx->spec = gen_spec_load(devinfo);
  49    else
  50       ctx->spec = gen_spec_load_from_path(devinfo, xml_path);
  51    ctx->disasm = gen_disasm_create(devinfo);
  52 }
  53
  54 void
  55 gen_batch_decode_ctx_finish(struct gen_batch_decode_ctx *ctx)
  56 {
  57    gen_spec_destroy(ctx->spec);
  58    gen_disasm_destroy(ctx->disasm);
  59 }
  60
  61 #define CSI "\e["
  62 #define RED_COLOR    CSI "31m"
  63 #define BLUE_HEADER  CSI "0;44m"
  64 #define GREEN_HEADER CSI "1;42m"
  65 #define NORMAL       CSI "0m"
  66
  67 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
  68
  69 static void
  70 ctx_print_group(struct gen_batch_decode_ctx *ctx,
  71                 struct gen_group *group,
  72                 uint64_t address, const void *map)
  73 {
  74    gen_print_group(ctx->fp, group, address, map, 0,
  75                    (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) != 0);
  76 }
  77
  78 static struct gen_batch_decode_bo
  79 ctx_get_bo(struct gen_batch_decode_ctx *ctx, uint64_t addr)
  80 {
  81    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0)) {
  82       /* On Broadwell and above, we have 48-bit addresses which consume two
  83        * dwords.  Some packets require that these get stored in a "canonical
  84        * form" which means that bit 47 is sign-extended through the upper
  85        * bits. In order to correctly handle those aub dumps, we need to mask
  86        * off the top 16 bits.
  87        */
  88       addr &= (~0ull >> 16);
  89    }
  90
  91    struct gen_batch_decode_bo bo = ctx->get_bo(ctx->user_data, addr);
  92
  93    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0))
  94       bo.addr &= (~0ull >> 16);
  95
  96    /* We may actually have an offset into the bo */
  97    if (bo.map != NULL) {
  98       assert(bo.addr <= addr);
  99       uint64_t offset = addr - bo.addr;
 100       bo.map += offset;
 101       bo.addr += offset;
 102       bo.size -= offset;
 103    }
 104
 105    return bo;
 106 }
 107
 108 static int
 109 update_count(struct gen_batch_decode_ctx *ctx,
 110              uint32_t offset_from_dsba,
 111              unsigned element_dwords,
 112              unsigned guess)
 113 {
 114    unsigned size = 0;
 115
 116    if (ctx->get_state_size)
 117       size = ctx->get_state_size(ctx->user_data, offset_from_dsba);
 118
 119    if (size > 0)
 120       return size / (sizeof(uint32_t) * element_dwords);
 121
 122    /* In the absence of any information, just guess arbitrarily. */
 123    return guess;
 124 }
 125
 126 static void
 127 ctx_disassemble_program(struct gen_batch_decode_ctx *ctx,
 128                         uint32_t ksp, const char *type)
 129 {
 130    if (!ctx->instruction_base.map)
 131       return;
 132
 133    printf("\nReferenced %s:\n", type);
 134    gen_disasm_disassemble(ctx->disasm,
 135                           (void *)ctx->instruction_base.map, ksp,
 136                           ctx->fp);
 137 }
 138
 139 /* Heuristic to determine whether a uint32_t is probably actually a float
 140  * (http://stackoverflow.com/a/2953466)
 141  */
 142
 143 static bool
 144 probably_float(uint32_t bits)
 145 {
 146    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 147    uint32_t mant = bits & 0x007fffff;
 148
 149    /* +- 0.0 */
 150    if (exp == -127 && mant == 0)
 151       return true;
 152
 153    /* +- 1 billionth to 1 billion */
 154    if (-30 <= exp && exp <= 30)
 155       return true;
 156
 157    /* some value with only a few binary digits */
 158    if ((mant & 0x0000ffff) == 0)
 159       return true;
 160
 161    return false;
 162 }
 163
 164 static void
 165 ctx_print_buffer(struct gen_batch_decode_ctx *ctx,
 166                  struct gen_batch_decode_bo bo,
 167                  uint32_t read_length,
 168                  uint32_t pitch)
 169 {
 170    const uint32_t *dw_end = bo.map + MIN2(bo.size, read_length);
 171
 172    unsigned line_count = 0;
 173    for (const uint32_t *dw = bo.map; dw < dw_end; dw++) {
 174       if (line_count * 4 == pitch || line_count == 8) {
 175          fprintf(ctx->fp, "\n");
 176          line_count = 0;
 177       }
 178       fprintf(ctx->fp, line_count == 0 ? "  " : " ");
 179
 180       if ((ctx->flags & GEN_BATCH_DECODE_FLOATS) && probably_float(*dw))
 181          fprintf(ctx->fp, "  %8.2f", *(float *) dw);
 182       else
 183          fprintf(ctx->fp, "  0x%08x", *dw);
 184
 185       line_count++;
 186    }
 187    fprintf(ctx->fp, "\n");
 188 }
 189
 190 static void
 191 handle_state_base_address(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 192 {
 193    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 194
 195    struct gen_field_iterator iter;
 196    gen_field_iterator_init(&iter, inst, p, 0, false);
 197
 198    while (gen_field_iterator_next(&iter)) {
 199       if (strcmp(iter.name, "Surface State Base Address") == 0) {
 200          ctx->surface_base = ctx_get_bo(ctx, iter.raw_value);
 201       } else if (strcmp(iter.name, "Dynamic State Base Address") == 0) {
 202          ctx->dynamic_base = ctx_get_bo(ctx, iter.raw_value);
 203       } else if (strcmp(iter.name, "Instruction Base Address") == 0) {
 204          ctx->instruction_base = ctx_get_bo(ctx, iter.raw_value);
 205       }
 206    }
 207 }
 208
 209 static void
 210 dump_binding_table(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 211 {
 212    struct gen_group *strct =
 213       gen_spec_find_struct(ctx->spec, "RENDER_SURFACE_STATE");
 214    if (strct == NULL) {
 215       fprintf(ctx->fp, "did not find RENDER_SURFACE_STATE info\n");
 216       return;
 217    }
 218
 219    if (count < 0)
 220       count = update_count(ctx, offset, 1, 8);
 221
 222    if (ctx->surface_base.map == NULL) {
 223       fprintf(ctx->fp, "  binding table unavailable\n");
 224       return;
 225    }
 226
 227    if (offset % 32 != 0 || offset >= UINT16_MAX ||
 228        offset >= ctx->surface_base.size) {
 229       fprintf(ctx->fp, "  invalid binding table pointer\n");
 230       return;
 231    }
 232
 233    const uint32_t *pointers = ctx->surface_base.map + offset;
 234    for (int i = 0; i < count; i++) {
 235       if (pointers[i] == 0)
 236          continue;
 237
 238       if (pointers[i] % 32 != 0 ||
 239           (pointers[i] + strct->dw_length * 4) >= ctx->surface_base.size) {
 240          fprintf(ctx->fp, "pointer %u: %08x <not valid>\n", i, pointers[i]);
 241          continue;
 242       }
 243
 244       fprintf(ctx->fp, "pointer %u: %08x\n", i, pointers[i]);
 245       ctx_print_group(ctx, strct, ctx->surface_base.addr + pointers[i],
 246                       ctx->surface_base.map + pointers[i]);
 247    }
 248 }
 249
 250 static void
 251 dump_samplers(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 252 {
 253    struct gen_group *strct = gen_spec_find_struct(ctx->spec, "SAMPLER_STATE");
 254
 255    if (count < 0)
 256       count = update_count(ctx, offset, strct->dw_length, 4);
 257
 258    if (ctx->dynamic_base.map == NULL) {
 259       fprintf(ctx->fp, "  samplers unavailable\n");
 260       return;
 261    }
 262
 263    if (offset % 32 != 0 || offset >= ctx->dynamic_base.size) {
 264       fprintf(ctx->fp, "  invalid sampler state pointer\n");
 265       return;
 266    }
 267
 268    uint64_t state_addr = ctx->dynamic_base.addr + offset;
 269    const void *state_map = ctx->dynamic_base.map + offset;
 270    for (int i = 0; i < count; i++) {
 271       fprintf(ctx->fp, "sampler state %d\n", i);
 272       ctx_print_group(ctx, strct, state_addr, state_map);
 273       state_addr += 16;
 274       state_map += 16;
 275    }
 276 }
 277
 278 static void
 279 handle_media_interface_descriptor_load(struct gen_batch_decode_ctx *ctx,
 280                                        const uint32_t *p)
 281 {
 282    if (ctx->dynamic_base.map == NULL)
 283       return;
 284
 285    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 286    struct gen_group *desc =
 287       gen_spec_find_struct(ctx->spec, "INTERFACE_DESCRIPTOR_DATA");
 288
 289    struct gen_field_iterator iter;
 290    gen_field_iterator_init(&iter, inst, p, 0, false);
 291    uint32_t descriptor_offset = 0;
 292    int descriptor_count = 0;
 293    while (gen_field_iterator_next(&iter)) {
 294       if (strcmp(iter.name, "Interface Descriptor Data Start Address") == 0) {
 295          descriptor_offset = strtol(iter.value, NULL, 16);
 296       } else if (strcmp(iter.name, "Interface Descriptor Total Length") == 0) {
 297          descriptor_count =
 298             strtol(iter.value, NULL, 16) / (desc->dw_length * 4);
 299       }
 300    }
 301
 302    uint64_t desc_addr = ctx->dynamic_base.addr + descriptor_offset;
 303    const uint32_t *desc_map = ctx->dynamic_base.map + descriptor_offset;
 304    for (int i = 0; i < descriptor_count; i++) {
 305       fprintf(ctx->fp, "descriptor %d: %08x\n", i, descriptor_offset);
 306
 307       ctx_print_group(ctx, desc, desc_addr, desc_map);
 308
 309       gen_field_iterator_init(&iter, desc, desc_map, 0, false);
 310       uint64_t ksp;
 311       uint32_t sampler_offset, sampler_count;
 312       uint32_t binding_table_offset, binding_entry_count;
 313       while (gen_field_iterator_next(&iter)) {
 314          if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 315             ksp = strtoll(iter.value, NULL, 16);
 316          } else if (strcmp(iter.name, "Sampler State Pointer") == 0) {
 317             sampler_offset = strtol(iter.value, NULL, 16);
 318          } else if (strcmp(iter.name, "Sampler Count") == 0) {
 319             sampler_count = strtol(iter.value, NULL, 10);
 320          } else if (strcmp(iter.name, "Binding Table Pointer") == 0) {
 321             binding_table_offset = strtol(iter.value, NULL, 16);
 322          } else if (strcmp(iter.name, "Binding Table Entry Count") == 0) {
 323             binding_entry_count = strtol(iter.value, NULL, 10);
 324          }
 325       }
 326
 327       ctx_disassemble_program(ctx, ksp, "compute shader");
 328       printf("\n");
 329
 330       dump_samplers(ctx, sampler_offset, sampler_count);
 331       dump_binding_table(ctx, binding_table_offset, binding_entry_count);
 332
 333       desc_map += desc->dw_length;
 334       desc_addr += desc->dw_length * 4;
 335    }
 336 }
 337
 338 static void
 339 handle_3dstate_vertex_buffers(struct gen_batch_decode_ctx *ctx,
 340                               const uint32_t *p)
 341 {
 342    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 343    struct gen_group *vbs = gen_spec_find_struct(ctx->spec, "VERTEX_BUFFER_STATE");
 344
 345    struct gen_batch_decode_bo vb = {};
 346    uint32_t vb_size = 0;
 347    int index = -1;
 348    int pitch = -1;
 349    bool ready = false;
 350
 351    struct gen_field_iterator iter;
 352    gen_field_iterator_init(&iter, inst, p, 0, false);
 353    while (gen_field_iterator_next(&iter)) {
 354       if (iter.struct_desc != vbs)
 355          continue;
 356
 357       struct gen_field_iterator vbs_iter;
 358       gen_field_iterator_init(&vbs_iter, vbs, &iter.p[iter.start_bit / 32], 0, false);
 359       while (gen_field_iterator_next(&vbs_iter)) {
 360          if (strcmp(vbs_iter.name, "Vertex Buffer Index") == 0) {
 361             index = vbs_iter.raw_value;
 362          } else if (strcmp(vbs_iter.name, "Buffer Pitch") == 0) {
 363             pitch = vbs_iter.raw_value;
 364          } else if (strcmp(vbs_iter.name, "Buffer Starting Address") == 0) {
 365             vb = ctx_get_bo(ctx, vbs_iter.raw_value);
 366          } else if (strcmp(vbs_iter.name, "Buffer Size") == 0) {
 367             vb_size = vbs_iter.raw_value;
 368             ready = true;
 369          } else if (strcmp(vbs_iter.name, "End Address") == 0) {
 370             if (vb.map && vbs_iter.raw_value >= vb.addr)
 371                vb_size = vbs_iter.raw_value - vb.addr;
 372             else
 373                vb_size = 0;
 374             ready = true;
 375          }
 376
 377          if (!ready)
 378             continue;
 379
 380          fprintf(ctx->fp, "vertex buffer %d, size %d\n", index, vb_size);
 381
 382          if (vb.map == NULL) {
 383             fprintf(ctx->fp, "  buffer contents unavailable\n");
 384             continue;
 385          }
 386
 387          if (vb.map == 0 || vb_size == 0)
 388             continue;
 389
 390          ctx_print_buffer(ctx, vb, vb_size, pitch);
 391
 392          vb.map = NULL;
 393          vb_size = 0;
 394          index = -1;
 395          pitch = -1;
 396          ready = false;
 397       }
 398    }
 399 }
 400
 401 static void
 402 handle_3dstate_index_buffer(struct gen_batch_decode_ctx *ctx,
 403                             const uint32_t *p)
 404 {
 405    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 406
 407    struct gen_batch_decode_bo ib = {};
 408    uint32_t ib_size = 0;
 409    uint32_t format = 0;
 410
 411    struct gen_field_iterator iter;
 412    gen_field_iterator_init(&iter, inst, p, 0, false);
 413    while (gen_field_iterator_next(&iter)) {
 414       if (strcmp(iter.name, "Index Format") == 0) {
 415          format = iter.raw_value;
 416       } else if (strcmp(iter.name, "Buffer Starting Address") == 0) {
 417          ib = ctx_get_bo(ctx, iter.raw_value);
 418       } else if (strcmp(iter.name, "Buffer Size") == 0) {
 419          ib_size = iter.raw_value;
 420       }
 421    }
 422
 423    if (ib.map == NULL) {
 424       fprintf(ctx->fp, "  buffer contents unavailable\n");
 425       return;
 426    }
 427
 428    const void *m = ib.map;
 429    const void *ib_end = ib.map + MIN2(ib.size, ib_size);
 430    for (int i = 0; m < ib_end && i < 10; i++) {
 431       switch (format) {
 432       case 0:
 433          fprintf(ctx->fp, "%3d ", *(uint8_t *)m);
 434          m += 1;
 435          break;
 436       case 1:
 437          fprintf(ctx->fp, "%3d ", *(uint16_t *)m);
 438          m += 2;
 439          break;
 440       case 2:
 441          fprintf(ctx->fp, "%3d ", *(uint32_t *)m);
 442          m += 4;
 443          break;
 444       }
 445    }
 446
 447    if (m < ib_end)
 448       fprintf(ctx->fp, "...");
 449    fprintf(ctx->fp, "\n");
 450 }
 451
 452 static void
 453 decode_single_ksp(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 454 {
 455    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 456
 457    uint64_t ksp = 0;
 458    bool is_simd8 = false; /* vertex shaders on Gen8+ only */
 459    bool is_enabled = true;
 460
 461    struct gen_field_iterator iter;
 462    gen_field_iterator_init(&iter, inst, p, 0, false);
 463    while (gen_field_iterator_next(&iter)) {
 464       if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 465          ksp = iter.raw_value;
 466       } else if (strcmp(iter.name, "SIMD8 Dispatch Enable") == 0) {
 467          is_simd8 = iter.raw_value;
 468       } else if (strcmp(iter.name, "Dispatch Mode") == 0) {
 469          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 470       } else if (strcmp(iter.name, "Dispatch Enable") == 0) {
 471          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 472       } else if (strcmp(iter.name, "Enable") == 0) {
 473          is_enabled = iter.raw_value;
 474       }
 475    }
 476
 477    const char *type =
 478       strcmp(inst->name,   "VS_STATE") == 0 ? "vertex shader" :
 479       strcmp(inst->name,   "GS_STATE") == 0 ? "geometry shader" :
 480       strcmp(inst->name,   "SF_STATE") == 0 ? "strips and fans shader" :
 481       strcmp(inst->name, "CLIP_STATE") == 0 ? "clip shader" :
 482       strcmp(inst->name, "3DSTATE_DS") == 0 ? "tessellation evaluation shader" :
 483       strcmp(inst->name, "3DSTATE_HS") == 0 ? "tessellation control shader" :
 484       strcmp(inst->name, "3DSTATE_VS") == 0 ? (is_simd8 ? "SIMD8 vertex shader" : "vec4 vertex shader") :
 485       strcmp(inst->name, "3DSTATE_GS") == 0 ? (is_simd8 ? "SIMD8 geometry shader" : "vec4 geometry shader") :
 486       NULL;
 487
 488    if (is_enabled) {
 489       ctx_disassemble_program(ctx, ksp, type);
 490       printf("\n");
 491    }
 492 }
 493
 494 static void
 495 decode_ps_kernels(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 496 {
 497    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 498
 499    uint64_t ksp[3] = {0, 0, 0};
 500    bool enabled[3] = {false, false, false};
 501
 502    struct gen_field_iterator iter;
 503    gen_field_iterator_init(&iter, inst, p, 0, false);
 504    while (gen_field_iterator_next(&iter)) {
 505       if (strncmp(iter.name, "Kernel Start Pointer ",
 506                   strlen("Kernel Start Pointer ")) == 0) {
 507          int idx = iter.name[strlen("Kernel Start Pointer ")] - '0';
 508          ksp[idx] = strtol(iter.value, NULL, 16);
 509       } else if (strcmp(iter.name, "8 Pixel Dispatch Enable") == 0) {
 510          enabled[0] = strcmp(iter.value, "true") == 0;
 511       } else if (strcmp(iter.name, "16 Pixel Dispatch Enable") == 0) {
 512          enabled[1] = strcmp(iter.value, "true") == 0;
 513       } else if (strcmp(iter.name, "32 Pixel Dispatch Enable") == 0) {
 514          enabled[2] = strcmp(iter.value, "true") == 0;
 515       }
 516    }
 517
 518    /* Reorder KSPs to be [8, 16, 32] instead of the hardware order. */
 519    if (enabled[0] + enabled[1] + enabled[2] == 1) {
 520       if (enabled[1]) {
 521          ksp[1] = ksp[0];
 522          ksp[0] = 0;
 523       } else if (enabled[2]) {
 524          ksp[2] = ksp[0];
 525          ksp[0] = 0;
 526       }
 527    } else {
 528       uint64_t tmp = ksp[1];
 529       ksp[1] = ksp[2];
 530       ksp[2] = tmp;
 531    }
 532
 533    if (enabled[0])
 534       ctx_disassemble_program(ctx, ksp[0], "SIMD8 fragment shader");
 535    if (enabled[1])
 536       ctx_disassemble_program(ctx, ksp[1], "SIMD16 fragment shader");
 537    if (enabled[2])
 538       ctx_disassemble_program(ctx, ksp[2], "SIMD32 fragment shader");
 539    fprintf(ctx->fp, "\n");
 540 }
 541
 542 static void
 543 decode_3dstate_constant(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 544 {
 545    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 546    struct gen_group *body =
 547       gen_spec_find_struct(ctx->spec, "3DSTATE_CONSTANT_BODY");
 548
 549    uint32_t read_length[4];
 550    struct gen_batch_decode_bo buffer[4];
 551    memset(buffer, 0, sizeof(buffer));
 552
 553    struct gen_field_iterator outer;
 554    gen_field_iterator_init(&outer, inst, p, 0, false);
 555    while (gen_field_iterator_next(&outer)) {
 556       if (outer.struct_desc != body)
 557          continue;
 558
 559       struct gen_field_iterator iter;
 560       gen_field_iterator_init(&iter, body, &outer.p[outer.start_bit / 32],
 561                               0, false);
 562
 563       while (gen_field_iterator_next(&iter)) {
 564          int idx;
 565          if (sscanf(iter.name, "Read Length[%d]", &idx) == 1) {
 566             read_length[idx] = iter.raw_value;
 567          } else if (sscanf(iter.name, "Buffer[%d]", &idx) == 1) {
 568             buffer[idx] = ctx_get_bo(ctx, iter.raw_value);
 569          }
 570       }
 571
 572       for (int i = 0; i < 4; i++) {
 573          if (read_length[i] == 0 || buffer[i].map == NULL)
 574             continue;
 575
 576          unsigned size = read_length[i] * 32;
 577          fprintf(ctx->fp, "constant buffer %d, size %u\n", i, size);
 578
 579          ctx_print_buffer(ctx, buffer[i], size, 0);
 580       }
 581    }
 582 }
 583
 584 static void
 585 decode_3dstate_binding_table_pointers(struct gen_batch_decode_ctx *ctx,
 586                                       const uint32_t *p)
 587 {
 588    dump_binding_table(ctx, p[1], -1);
 589 }
 590
 591 static void
 592 decode_3dstate_sampler_state_pointers(struct gen_batch_decode_ctx *ctx,
 593                                       const uint32_t *p)
 594 {
 595    dump_samplers(ctx, p[1], -1);
 596 }
 597
 598 static void
 599 decode_3dstate_sampler_state_pointers_gen6(struct gen_batch_decode_ctx *ctx,
 600                                            const uint32_t *p)
 601 {
 602    dump_samplers(ctx, p[1], -1);
 603    dump_samplers(ctx, p[2], -1);
 604    dump_samplers(ctx, p[3], -1);
 605 }
 606
 607 static bool
 608 str_ends_with(const char *str, const char *end)
 609 {
 610    int offset = strlen(str) - strlen(end);
 611    if (offset < 0)
 612       return false;
 613
 614    return strcmp(str + offset, end) == 0;
 615 }
 616
 617 static void
 618 decode_dynamic_state_pointers(struct gen_batch_decode_ctx *ctx,
 619                               const char *struct_type, const uint32_t *p,
 620                               int count)
 621 {
 622    if (ctx->dynamic_base.map == NULL) {
 623       fprintf(ctx->fp, "  dynamic %s state unavailable\n", struct_type);
 624       return;
 625    }
 626
 627    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 628    struct gen_group *state = gen_spec_find_struct(ctx->spec, struct_type);
 629
 630    uint32_t state_offset;
 631
 632    struct gen_field_iterator iter;
 633    gen_field_iterator_init(&iter, inst, p, 0, false);
 634    while (gen_field_iterator_next(&iter)) {
 635       if (str_ends_with(iter.name, "Pointer")) {
 636          state_offset = iter.raw_value;
 637          break;
 638       }
 639    }
 640
 641    uint32_t state_addr = ctx->dynamic_base.addr + state_offset;
 642    const uint32_t *state_map = ctx->dynamic_base.map + state_offset;
 643    for (int i = 0; i < count; i++) {
 644       fprintf(ctx->fp, "%s %d\n", struct_type, i);
 645       ctx_print_group(ctx, state, state_offset, state_map);
 646
 647       state_addr += state->dw_length * 4;
 648       state_map += state->dw_length;
 649    }
 650 }
 651
 652 static void
 653 decode_3dstate_viewport_state_pointers_cc(struct gen_batch_decode_ctx *ctx,
 654                                           const uint32_t *p)
 655 {
 656    decode_dynamic_state_pointers(ctx, "CC_VIEWPORT", p, 4);
 657 }
 658
 659 static void
 660 decode_3dstate_viewport_state_pointers_sf_clip(struct gen_batch_decode_ctx *ctx,
 661                                                const uint32_t *p)
 662 {
 663    decode_dynamic_state_pointers(ctx, "SF_CLIP_VIEWPORT", p, 4);
 664 }
 665
 666 static void
 667 decode_3dstate_blend_state_pointers(struct gen_batch_decode_ctx *ctx,
 668                                     const uint32_t *p)
 669 {
 670    decode_dynamic_state_pointers(ctx, "BLEND_STATE", p, 1);
 671 }
 672
 673 static void
 674 decode_3dstate_cc_state_pointers(struct gen_batch_decode_ctx *ctx,
 675                                  const uint32_t *p)
 676 {
 677    decode_dynamic_state_pointers(ctx, "COLOR_CALC_STATE", p, 1);
 678 }
 679
 680 static void
 681 decode_3dstate_scissor_state_pointers(struct gen_batch_decode_ctx *ctx,
 682                                       const uint32_t *p)
 683 {
 684    decode_dynamic_state_pointers(ctx, "SCISSOR_RECT", p, 1);
 685 }
 686
 687 static void
 688 decode_load_register_imm(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 689 {
 690    struct gen_group *reg = gen_spec_find_register(ctx->spec, p[1]);
 691
 692    if (reg != NULL) {
 693       fprintf(ctx->fp, "register %s (0x%x): 0x%x\n",
 694               reg->name, reg->register_offset, p[2]);
 695       ctx_print_group(ctx, reg, reg->register_offset, &p[2]);
 696    }
 697 }
 698
 699 struct custom_decoder {
 700    const char *cmd_name;
 701    void (*decode)(struct gen_batch_decode_ctx *ctx, const uint32_t *p);
 702 } custom_decoders[] = {
 703    { "STATE_BASE_ADDRESS", handle_state_base_address },
 704    { "MEDIA_INTERFACE_DESCRIPTOR_LOAD", handle_media_interface_descriptor_load },
 705    { "3DSTATE_VERTEX_BUFFERS", handle_3dstate_vertex_buffers },
 706    { "3DSTATE_INDEX_BUFFER", handle_3dstate_index_buffer },
 707    { "3DSTATE_VS", decode_single_ksp },
 708    { "3DSTATE_GS", decode_single_ksp },
 709    { "3DSTATE_DS", decode_single_ksp },
 710    { "3DSTATE_HS", decode_single_ksp },
 711    { "3DSTATE_PS", decode_ps_kernels },
 712    { "3DSTATE_CONSTANT_VS", decode_3dstate_constant },
 713    { "3DSTATE_CONSTANT_GS", decode_3dstate_constant },
 714    { "3DSTATE_CONSTANT_PS", decode_3dstate_constant },
 715    { "3DSTATE_CONSTANT_HS", decode_3dstate_constant },
 716    { "3DSTATE_CONSTANT_DS", decode_3dstate_constant },
 717
 718    { "3DSTATE_BINDING_TABLE_POINTERS_VS", decode_3dstate_binding_table_pointers },
 719    { "3DSTATE_BINDING_TABLE_POINTERS_HS", decode_3dstate_binding_table_pointers },
 720    { "3DSTATE_BINDING_TABLE_POINTERS_DS", decode_3dstate_binding_table_pointers },
 721    { "3DSTATE_BINDING_TABLE_POINTERS_GS", decode_3dstate_binding_table_pointers },
 722    { "3DSTATE_BINDING_TABLE_POINTERS_PS", decode_3dstate_binding_table_pointers },
 723
 724    { "3DSTATE_SAMPLER_STATE_POINTERS_VS", decode_3dstate_sampler_state_pointers },
 725    { "3DSTATE_SAMPLER_STATE_POINTERS_HS", decode_3dstate_sampler_state_pointers },
 726    { "3DSTATE_SAMPLER_STATE_POINTERS_DS", decode_3dstate_sampler_state_pointers },
 727    { "3DSTATE_SAMPLER_STATE_POINTERS_GS", decode_3dstate_sampler_state_pointers },
 728    { "3DSTATE_SAMPLER_STATE_POINTERS_PS", decode_3dstate_sampler_state_pointers },
 729    { "3DSTATE_SAMPLER_STATE_POINTERS", decode_3dstate_sampler_state_pointers_gen6 },
 730
 731    { "3DSTATE_VIEWPORT_STATE_POINTERS_CC", decode_3dstate_viewport_state_pointers_cc },
 732    { "3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP", decode_3dstate_viewport_state_pointers_sf_clip },
 733    { "3DSTATE_BLEND_STATE_POINTERS", decode_3dstate_blend_state_pointers },
 734    { "3DSTATE_CC_STATE_POINTERS", decode_3dstate_cc_state_pointers },
 735    { "3DSTATE_SCISSOR_STATE_POINTERS", decode_3dstate_scissor_state_pointers },
 736    { "MI_LOAD_REGISTER_IMM", decode_load_register_imm }
 737 };
 738
 739 static inline uint64_t
 740 get_address(struct gen_spec *spec, const uint32_t *p)
 741 {
 742    /* Addresses are always guaranteed to be page-aligned and sometimes
 743     * hardware packets have extra stuff stuffed in the bottom 12 bits.
 744     */
 745    uint64_t addr = p[0] & ~0xfffu;
 746
 747    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 748       /* On Broadwell and above, we have 48-bit addresses which consume two
 749        * dwords.  Some packets require that these get stored in a "canonical
 750        * form" which means that bit 47 is sign-extended through the upper
 751        * bits. In order to correctly handle those aub dumps, we need to mask
 752        * off the top 16 bits.
 753        */
 754       addr |= ((uint64_t)p[1] & 0xffff) << 32;
 755    }
 756
 757    return addr;
 758 }
 759
 760 void
 761 gen_print_batch(struct gen_batch_decode_ctx *ctx,
 762                 const uint32_t *batch, uint32_t batch_size,
 763                 uint64_t batch_addr)
 764 {
 765    const uint32_t *p, *end = batch + batch_size;
 766    int length;
 767    struct gen_group *inst;
 768
 769    for (p = batch; p < end; p += length) {
 770       inst = gen_spec_find_instruction(ctx->spec, p);
 771       length = gen_group_get_length(inst, p);
 772       assert(inst == NULL || length > 0);
 773       length = MAX2(1, length);
 774
 775       const char *reset_color = ctx->flags & GEN_BATCH_DECODE_IN_COLOR ? NORMAL : "";
 776
 777       uint64_t offset;
 778       if (ctx->flags & GEN_BATCH_DECODE_OFFSETS)
 779          offset = batch_addr + ((char *)p - (char *)batch);
 780       else
 781          offset = 0;
 782
 783       if (inst == NULL) {
 784          fprintf(ctx->fp, "%s0x%08"PRIx64": unknown instruction %08x%s\n",
 785                  (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) ? RED_COLOR : "",
 786                  offset, p[0], reset_color);
 787          continue;
 788       }
 789
 790       const char *color;
 791       const char *inst_name = gen_group_get_name(inst);
 792       if (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) {
 793          reset_color = NORMAL;
 794          if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 795             if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0 ||
 796                 strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0)
 797                color = GREEN_HEADER;
 798             else
 799                color = BLUE_HEADER;
 800          } else {
 801             color = NORMAL;
 802          }
 803       } else {
 804          color = "";
 805          reset_color = "";
 806       }
 807
 808       fprintf(ctx->fp, "%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 809               color, offset, p[0], inst_name, reset_color);
 810
 811       if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 812          ctx_print_group(ctx, inst, offset, p);
 813
 814          for (int i = 0; i < ARRAY_LENGTH(custom_decoders); i++) {
 815             if (strcmp(inst_name, custom_decoders[i].cmd_name) == 0) {
 816                custom_decoders[i].decode(ctx, p);
 817                break;
 818             }
 819          }
 820       }
 821
 822       if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0) {
 823          struct gen_batch_decode_bo next_batch;
 824          bool second_level;
 825          struct gen_field_iterator iter;
 826          gen_field_iterator_init(&iter, inst, p, 0, false);
 827          while (gen_field_iterator_next(&iter)) {
 828             if (strcmp(iter.name, "Batch Buffer Start Address") == 0) {
 829                next_batch = ctx_get_bo(ctx, iter.raw_value);
 830             } else if (strcmp(iter.name, "Second Level Batch Buffer") == 0) {
 831                second_level = iter.raw_value;
 832             }
 833          }
 834
 835          if (next_batch.map == NULL) {
 836             fprintf(ctx->fp, "Secondary batch at 0x%08"PRIx64" unavailable",
 837                     next_batch.addr);
 838          }
 839
 840          if (second_level) {
 841             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 842              * like a subroutine call.  Commands that come afterwards get
 843              * processed once the 2nd level batch buffer returns with
 844              * MI_BATCH_BUFFER_END.
 845              */
 846             if (next_batch.map) {
 847                gen_print_batch(ctx, next_batch.map, next_batch.size,
 848                                next_batch.addr);
 849             }
 850          } else {
 851             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 852              * like a goto.  Nothing after it will ever get processed.  In
 853              * order to prevent the recursion from growing, we just reset the
 854              * loop and continue;
 855              */
 856             if (next_batch.map) {
 857                p = next_batch.map;
 858                end = next_batch.map + next_batch.size;
 859                length = 0;
 860                continue;
 861             } else {
 862                /* Nothing we can do */
 863                break;
 864             }
 865          }
 866       } else if (strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0) {
 867          break;
 868       }
 869    }
 870 }