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    ctx->max_vbo_decoded_lines = -1; /* No limit! */
  47
  48    if (xml_path == NULL)
  49       ctx->spec = gen_spec_load(devinfo);
  50    else
  51       ctx->spec = gen_spec_load_from_path(devinfo, xml_path);
  52    ctx->disasm = gen_disasm_create(devinfo);
  53 }
  54
  55 void
  56 gen_batch_decode_ctx_finish(struct gen_batch_decode_ctx *ctx)
  57 {
  58    gen_spec_destroy(ctx->spec);
  59    gen_disasm_destroy(ctx->disasm);
  60 }
  61
  62 #define CSI "\e["
  63 #define RED_COLOR    CSI "31m"
  64 #define BLUE_HEADER  CSI "0;44m"
  65 #define GREEN_HEADER CSI "1;42m"
  66 #define NORMAL       CSI "0m"
  67
  68 #define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
  69
  70 static void
  71 ctx_print_group(struct gen_batch_decode_ctx *ctx,
  72                 struct gen_group *group,
  73                 uint64_t address, const void *map)
  74 {
  75    gen_print_group(ctx->fp, group, address, map, 0,
  76                    (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) != 0);
  77 }
  78
  79 static struct gen_batch_decode_bo
  80 ctx_get_bo(struct gen_batch_decode_ctx *ctx, uint64_t addr)
  81 {
  82    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0)) {
  83       /* On Broadwell and above, we have 48-bit addresses which consume two
  84        * dwords.  Some packets require that these get stored in a "canonical
  85        * form" which means that bit 47 is sign-extended through the upper
  86        * bits. In order to correctly handle those aub dumps, we need to mask
  87        * off the top 16 bits.
  88        */
  89       addr &= (~0ull >> 16);
  90    }
  91
  92    struct gen_batch_decode_bo bo = ctx->get_bo(ctx->user_data, addr);
  93
  94    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0))
  95       bo.addr &= (~0ull >> 16);
  96
  97    /* We may actually have an offset into the bo */
  98    if (bo.map != NULL) {
  99       assert(bo.addr <= addr);
 100       uint64_t offset = addr - bo.addr;
 101       bo.map += offset;
 102       bo.addr += offset;
 103       bo.size -= offset;
 104    }
 105
 106    return bo;
 107 }
 108
 109 static int
 110 update_count(struct gen_batch_decode_ctx *ctx,
 111              uint32_t offset_from_dsba,
 112              unsigned element_dwords,
 113              unsigned guess)
 114 {
 115    unsigned size = 0;
 116
 117    if (ctx->get_state_size)
 118       size = ctx->get_state_size(ctx->user_data, offset_from_dsba);
 119
 120    if (size > 0)
 121       return size / (sizeof(uint32_t) * element_dwords);
 122
 123    /* In the absence of any information, just guess arbitrarily. */
 124    return guess;
 125 }
 126
 127 static void
 128 ctx_disassemble_program(struct gen_batch_decode_ctx *ctx,
 129                         uint32_t ksp, const char *type)
 130 {
 131    uint64_t addr = ctx->instruction_base + ksp;
 132    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, addr);
 133    if (!bo.map)
 134       return;
 135
 136    fprintf(ctx->fp, "\nReferenced %s:\n", type);
 137    gen_disasm_disassemble(ctx->disasm, bo.map, 0, ctx->fp);
 138 }
 139
 140 /* Heuristic to determine whether a uint32_t is probably actually a float
 141  * (http://stackoverflow.com/a/2953466)
 142  */
 143
 144 static bool
 145 probably_float(uint32_t bits)
 146 {
 147    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 148    uint32_t mant = bits & 0x007fffff;
 149
 150    /* +- 0.0 */
 151    if (exp == -127 && mant == 0)
 152       return true;
 153
 154    /* +- 1 billionth to 1 billion */
 155    if (-30 <= exp && exp <= 30)
 156       return true;
 157
 158    /* some value with only a few binary digits */
 159    if ((mant & 0x0000ffff) == 0)
 160       return true;
 161
 162    return false;
 163 }
 164
 165 static void
 166 ctx_print_buffer(struct gen_batch_decode_ctx *ctx,
 167                  struct gen_batch_decode_bo bo,
 168                  uint32_t read_length,
 169                  uint32_t pitch,
 170                  int max_lines)
 171 {
 172    const uint32_t *dw_end = bo.map + MIN2(bo.size, read_length);
 173
 174    int column_count = 0, line_count = -1;
 175    for (const uint32_t *dw = bo.map; dw < dw_end; dw++) {
 176       if (column_count * 4 == pitch || column_count == 8) {
 177          fprintf(ctx->fp, "\n");
 178          column_count = 0;
 179          line_count++;
 180
 181          if (max_lines >= 0 && line_count >= max_lines)
 182             break;
 183       }
 184       fprintf(ctx->fp, column_count == 0 ? "  " : " ");
 185
 186       if ((ctx->flags & GEN_BATCH_DECODE_FLOATS) && probably_float(*dw))
 187          fprintf(ctx->fp, "  %8.2f", *(float *) dw);
 188       else
 189          fprintf(ctx->fp, "  0x%08x", *dw);
 190
 191       column_count++;
 192    }
 193    fprintf(ctx->fp, "\n");
 194 }
 195
 196 static void
 197 handle_state_base_address(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 198 {
 199    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 200
 201    struct gen_field_iterator iter;
 202    gen_field_iterator_init(&iter, inst, p, 0, false);
 203
 204    while (gen_field_iterator_next(&iter)) {
 205       if (strcmp(iter.name, "Surface State Base Address") == 0) {
 206          ctx->surface_base = iter.raw_value;
 207       } else if (strcmp(iter.name, "Dynamic State Base Address") == 0) {
 208          ctx->dynamic_base = iter.raw_value;
 209       } else if (strcmp(iter.name, "Instruction Base Address") == 0) {
 210          ctx->instruction_base = iter.raw_value;
 211       }
 212    }
 213 }
 214
 215 static void
 216 dump_binding_table(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 217 {
 218    struct gen_group *strct =
 219       gen_spec_find_struct(ctx->spec, "RENDER_SURFACE_STATE");
 220    if (strct == NULL) {
 221       fprintf(ctx->fp, "did not find RENDER_SURFACE_STATE info\n");
 222       return;
 223    }
 224
 225    if (count < 0)
 226       count = update_count(ctx, offset, 1, 8);
 227
 228    if (offset % 32 != 0 || offset >= UINT16_MAX) {
 229       fprintf(ctx->fp, "  invalid binding table pointer\n");
 230       return;
 231    }
 232
 233    struct gen_batch_decode_bo bind_bo =
 234       ctx_get_bo(ctx, ctx->surface_base + offset);
 235
 236    if (bind_bo.map == NULL) {
 237       fprintf(ctx->fp, "  binding table unavailable\n");
 238       return;
 239    }
 240
 241    const uint32_t *pointers = bind_bo.map;
 242    for (int i = 0; i < count; i++) {
 243       if (pointers[i] == 0)
 244          continue;
 245
 246       uint64_t addr = ctx->surface_base + pointers[i];
 247       struct gen_batch_decode_bo bo = ctx_get_bo(ctx, addr);
 248       uint32_t size = strct->dw_length * 4;
 249
 250       if (pointers[i] % 32 != 0 ||
 251           addr < bo.addr || addr + size >= bo.addr + bo.size) {
 252          fprintf(ctx->fp, "pointer %u: %08x <not valid>\n", i, pointers[i]);
 253          continue;
 254       }
 255
 256       fprintf(ctx->fp, "pointer %u: %08x\n", i, pointers[i]);
 257       ctx_print_group(ctx, strct, addr, bo.map + (addr - bo.addr));
 258    }
 259 }
 260
 261 static void
 262 dump_samplers(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 263 {
 264    struct gen_group *strct = gen_spec_find_struct(ctx->spec, "SAMPLER_STATE");
 265
 266    if (count < 0)
 267       count = update_count(ctx, offset, strct->dw_length, 4);
 268
 269    uint64_t state_addr = ctx->dynamic_base + offset;
 270    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, state_addr);
 271    const void *state_map = bo.map;
 272
 273    if (state_map == NULL) {
 274       fprintf(ctx->fp, "  samplers unavailable\n");
 275       return;
 276    }
 277
 278    if (offset % 32 != 0 || state_addr - bo.addr >= bo.size) {
 279       fprintf(ctx->fp, "  invalid sampler state pointer\n");
 280       return;
 281    }
 282
 283    for (int i = 0; i < count; i++) {
 284       fprintf(ctx->fp, "sampler state %d\n", i);
 285       ctx_print_group(ctx, strct, state_addr, state_map);
 286       state_addr += 16;
 287       state_map += 16;
 288    }
 289 }
 290
 291 static void
 292 handle_media_interface_descriptor_load(struct gen_batch_decode_ctx *ctx,
 293                                        const uint32_t *p)
 294 {
 295    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 296    struct gen_group *desc =
 297       gen_spec_find_struct(ctx->spec, "INTERFACE_DESCRIPTOR_DATA");
 298
 299    struct gen_field_iterator iter;
 300    gen_field_iterator_init(&iter, inst, p, 0, false);
 301    uint32_t descriptor_offset = 0;
 302    int descriptor_count = 0;
 303    while (gen_field_iterator_next(&iter)) {
 304       if (strcmp(iter.name, "Interface Descriptor Data Start Address") == 0) {
 305          descriptor_offset = strtol(iter.value, NULL, 16);
 306       } else if (strcmp(iter.name, "Interface Descriptor Total Length") == 0) {
 307          descriptor_count =
 308             strtol(iter.value, NULL, 16) / (desc->dw_length * 4);
 309       }
 310    }
 311
 312    uint64_t desc_addr = ctx->dynamic_base + descriptor_offset;
 313    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, desc_addr);
 314    const void *desc_map = bo.map;
 315
 316    if (desc_map == NULL) {
 317       fprintf(ctx->fp, "  interface descriptors unavailable\n");
 318       return;
 319    }
 320
 321    for (int i = 0; i < descriptor_count; i++) {
 322       fprintf(ctx->fp, "descriptor %d: %08x\n", i, descriptor_offset);
 323
 324       ctx_print_group(ctx, desc, desc_addr, desc_map);
 325
 326       gen_field_iterator_init(&iter, desc, desc_map, 0, false);
 327       uint64_t ksp = 0;
 328       uint32_t sampler_offset = 0, sampler_count = 0;
 329       uint32_t binding_table_offset = 0, binding_entry_count = 0;
 330       while (gen_field_iterator_next(&iter)) {
 331          if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 332             ksp = strtoll(iter.value, NULL, 16);
 333          } else if (strcmp(iter.name, "Sampler State Pointer") == 0) {
 334             sampler_offset = strtol(iter.value, NULL, 16);
 335          } else if (strcmp(iter.name, "Sampler Count") == 0) {
 336             sampler_count = strtol(iter.value, NULL, 10);
 337          } else if (strcmp(iter.name, "Binding Table Pointer") == 0) {
 338             binding_table_offset = strtol(iter.value, NULL, 16);
 339          } else if (strcmp(iter.name, "Binding Table Entry Count") == 0) {
 340             binding_entry_count = strtol(iter.value, NULL, 10);
 341          }
 342       }
 343
 344       ctx_disassemble_program(ctx, ksp, "compute shader");
 345       printf("\n");
 346
 347       dump_samplers(ctx, sampler_offset, sampler_count);
 348       dump_binding_table(ctx, binding_table_offset, binding_entry_count);
 349
 350       desc_map += desc->dw_length;
 351       desc_addr += desc->dw_length * 4;
 352    }
 353 }
 354
 355 static void
 356 handle_3dstate_vertex_buffers(struct gen_batch_decode_ctx *ctx,
 357                               const uint32_t *p)
 358 {
 359    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 360    struct gen_group *vbs = gen_spec_find_struct(ctx->spec, "VERTEX_BUFFER_STATE");
 361
 362    struct gen_batch_decode_bo vb = {};
 363    uint32_t vb_size = 0;
 364    int index = -1;
 365    int pitch = -1;
 366    bool ready = false;
 367
 368    struct gen_field_iterator iter;
 369    gen_field_iterator_init(&iter, inst, p, 0, false);
 370    while (gen_field_iterator_next(&iter)) {
 371       if (iter.struct_desc != vbs)
 372          continue;
 373
 374       struct gen_field_iterator vbs_iter;
 375       gen_field_iterator_init(&vbs_iter, vbs, &iter.p[iter.start_bit / 32], 0, false);
 376       while (gen_field_iterator_next(&vbs_iter)) {
 377          if (strcmp(vbs_iter.name, "Vertex Buffer Index") == 0) {
 378             index = vbs_iter.raw_value;
 379          } else if (strcmp(vbs_iter.name, "Buffer Pitch") == 0) {
 380             pitch = vbs_iter.raw_value;
 381          } else if (strcmp(vbs_iter.name, "Buffer Starting Address") == 0) {
 382             vb = ctx_get_bo(ctx, vbs_iter.raw_value);
 383          } else if (strcmp(vbs_iter.name, "Buffer Size") == 0) {
 384             vb_size = vbs_iter.raw_value;
 385             ready = true;
 386          } else if (strcmp(vbs_iter.name, "End Address") == 0) {
 387             if (vb.map && vbs_iter.raw_value >= vb.addr)
 388                vb_size = vbs_iter.raw_value - vb.addr;
 389             else
 390                vb_size = 0;
 391             ready = true;
 392          }
 393
 394          if (!ready)
 395             continue;
 396
 397          fprintf(ctx->fp, "vertex buffer %d, size %d\n", index, vb_size);
 398
 399          if (vb.map == NULL) {
 400             fprintf(ctx->fp, "  buffer contents unavailable\n");
 401             continue;
 402          }
 403
 404          if (vb.map == 0 || vb_size == 0)
 405             continue;
 406
 407          ctx_print_buffer(ctx, vb, vb_size, pitch, ctx->max_vbo_decoded_lines);
 408
 409          vb.map = NULL;
 410          vb_size = 0;
 411          index = -1;
 412          pitch = -1;
 413          ready = false;
 414       }
 415    }
 416 }
 417
 418 static void
 419 handle_3dstate_index_buffer(struct gen_batch_decode_ctx *ctx,
 420                             const uint32_t *p)
 421 {
 422    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 423
 424    struct gen_batch_decode_bo ib = {};
 425    uint32_t ib_size = 0;
 426    uint32_t format = 0;
 427
 428    struct gen_field_iterator iter;
 429    gen_field_iterator_init(&iter, inst, p, 0, false);
 430    while (gen_field_iterator_next(&iter)) {
 431       if (strcmp(iter.name, "Index Format") == 0) {
 432          format = iter.raw_value;
 433       } else if (strcmp(iter.name, "Buffer Starting Address") == 0) {
 434          ib = ctx_get_bo(ctx, iter.raw_value);
 435       } else if (strcmp(iter.name, "Buffer Size") == 0) {
 436          ib_size = iter.raw_value;
 437       }
 438    }
 439
 440    if (ib.map == NULL) {
 441       fprintf(ctx->fp, "  buffer contents unavailable\n");
 442       return;
 443    }
 444
 445    const void *m = ib.map;
 446    const void *ib_end = ib.map + MIN2(ib.size, ib_size);
 447    for (int i = 0; m < ib_end && i < 10; i++) {
 448       switch (format) {
 449       case 0:
 450          fprintf(ctx->fp, "%3d ", *(uint8_t *)m);
 451          m += 1;
 452          break;
 453       case 1:
 454          fprintf(ctx->fp, "%3d ", *(uint16_t *)m);
 455          m += 2;
 456          break;
 457       case 2:
 458          fprintf(ctx->fp, "%3d ", *(uint32_t *)m);
 459          m += 4;
 460          break;
 461       }
 462    }
 463
 464    if (m < ib_end)
 465       fprintf(ctx->fp, "...");
 466    fprintf(ctx->fp, "\n");
 467 }
 468
 469 static void
 470 decode_single_ksp(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 471 {
 472    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 473
 474    uint64_t ksp = 0;
 475    bool is_simd8 = false; /* vertex shaders on Gen8+ only */
 476    bool is_enabled = true;
 477
 478    struct gen_field_iterator iter;
 479    gen_field_iterator_init(&iter, inst, p, 0, false);
 480    while (gen_field_iterator_next(&iter)) {
 481       if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 482          ksp = iter.raw_value;
 483       } else if (strcmp(iter.name, "SIMD8 Dispatch Enable") == 0) {
 484          is_simd8 = iter.raw_value;
 485       } else if (strcmp(iter.name, "Dispatch Mode") == 0) {
 486          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 487       } else if (strcmp(iter.name, "Dispatch Enable") == 0) {
 488          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 489       } else if (strcmp(iter.name, "Enable") == 0) {
 490          is_enabled = iter.raw_value;
 491       }
 492    }
 493
 494    const char *type =
 495       strcmp(inst->name,   "VS_STATE") == 0 ? "vertex shader" :
 496       strcmp(inst->name,   "GS_STATE") == 0 ? "geometry shader" :
 497       strcmp(inst->name,   "SF_STATE") == 0 ? "strips and fans shader" :
 498       strcmp(inst->name, "CLIP_STATE") == 0 ? "clip shader" :
 499       strcmp(inst->name, "3DSTATE_DS") == 0 ? "tessellation evaluation shader" :
 500       strcmp(inst->name, "3DSTATE_HS") == 0 ? "tessellation control shader" :
 501       strcmp(inst->name, "3DSTATE_VS") == 0 ? (is_simd8 ? "SIMD8 vertex shader" : "vec4 vertex shader") :
 502       strcmp(inst->name, "3DSTATE_GS") == 0 ? (is_simd8 ? "SIMD8 geometry shader" : "vec4 geometry shader") :
 503       NULL;
 504
 505    if (is_enabled) {
 506       ctx_disassemble_program(ctx, ksp, type);
 507       printf("\n");
 508    }
 509 }
 510
 511 static void
 512 decode_ps_kernels(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 513 {
 514    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 515
 516    uint64_t ksp[3] = {0, 0, 0};
 517    bool enabled[3] = {false, false, false};
 518
 519    struct gen_field_iterator iter;
 520    gen_field_iterator_init(&iter, inst, p, 0, false);
 521    while (gen_field_iterator_next(&iter)) {
 522       if (strncmp(iter.name, "Kernel Start Pointer ",
 523                   strlen("Kernel Start Pointer ")) == 0) {
 524          int idx = iter.name[strlen("Kernel Start Pointer ")] - '0';
 525          ksp[idx] = strtol(iter.value, NULL, 16);
 526       } else if (strcmp(iter.name, "8 Pixel Dispatch Enable") == 0) {
 527          enabled[0] = strcmp(iter.value, "true") == 0;
 528       } else if (strcmp(iter.name, "16 Pixel Dispatch Enable") == 0) {
 529          enabled[1] = strcmp(iter.value, "true") == 0;
 530       } else if (strcmp(iter.name, "32 Pixel Dispatch Enable") == 0) {
 531          enabled[2] = strcmp(iter.value, "true") == 0;
 532       }
 533    }
 534
 535    /* Reorder KSPs to be [8, 16, 32] instead of the hardware order. */
 536    if (enabled[0] + enabled[1] + enabled[2] == 1) {
 537       if (enabled[1]) {
 538          ksp[1] = ksp[0];
 539          ksp[0] = 0;
 540       } else if (enabled[2]) {
 541          ksp[2] = ksp[0];
 542          ksp[0] = 0;
 543       }
 544    } else {
 545       uint64_t tmp = ksp[1];
 546       ksp[1] = ksp[2];
 547       ksp[2] = tmp;
 548    }
 549
 550    if (enabled[0])
 551       ctx_disassemble_program(ctx, ksp[0], "SIMD8 fragment shader");
 552    if (enabled[1])
 553       ctx_disassemble_program(ctx, ksp[1], "SIMD16 fragment shader");
 554    if (enabled[2])
 555       ctx_disassemble_program(ctx, ksp[2], "SIMD32 fragment shader");
 556    fprintf(ctx->fp, "\n");
 557 }
 558
 559 static void
 560 decode_3dstate_constant(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 561 {
 562    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 563    struct gen_group *body =
 564       gen_spec_find_struct(ctx->spec, "3DSTATE_CONSTANT_BODY");
 565
 566    uint32_t read_length[4] = {0};
 567    uint64_t read_addr[4];
 568
 569    struct gen_field_iterator outer;
 570    gen_field_iterator_init(&outer, inst, p, 0, false);
 571    while (gen_field_iterator_next(&outer)) {
 572       if (outer.struct_desc != body)
 573          continue;
 574
 575       struct gen_field_iterator iter;
 576       gen_field_iterator_init(&iter, body, &outer.p[outer.start_bit / 32],
 577                               0, false);
 578
 579       while (gen_field_iterator_next(&iter)) {
 580          int idx;
 581          if (sscanf(iter.name, "Read Length[%d]", &idx) == 1) {
 582             read_length[idx] = iter.raw_value;
 583          } else if (sscanf(iter.name, "Buffer[%d]", &idx) == 1) {
 584             read_addr[idx] = iter.raw_value;
 585          }
 586       }
 587
 588       for (int i = 0; i < 4; i++) {
 589          if (read_length[i] == 0)
 590             continue;
 591
 592          struct gen_batch_decode_bo buffer = ctx_get_bo(ctx, read_addr[i]);
 593          if (!buffer.map) {
 594             fprintf(ctx->fp, "constant buffer %d unavailable\n", i);
 595             continue;
 596          }
 597
 598          unsigned size = read_length[i] * 32;
 599          fprintf(ctx->fp, "constant buffer %d, size %u\n", i, size);
 600
 601          ctx_print_buffer(ctx, buffer, size, 0, -1);
 602       }
 603    }
 604 }
 605
 606 static void
 607 decode_3dstate_binding_table_pointers(struct gen_batch_decode_ctx *ctx,
 608                                       const uint32_t *p)
 609 {
 610    dump_binding_table(ctx, p[1], -1);
 611 }
 612
 613 static void
 614 decode_3dstate_sampler_state_pointers(struct gen_batch_decode_ctx *ctx,
 615                                       const uint32_t *p)
 616 {
 617    dump_samplers(ctx, p[1], -1);
 618 }
 619
 620 static void
 621 decode_3dstate_sampler_state_pointers_gen6(struct gen_batch_decode_ctx *ctx,
 622                                            const uint32_t *p)
 623 {
 624    dump_samplers(ctx, p[1], -1);
 625    dump_samplers(ctx, p[2], -1);
 626    dump_samplers(ctx, p[3], -1);
 627 }
 628
 629 static bool
 630 str_ends_with(const char *str, const char *end)
 631 {
 632    int offset = strlen(str) - strlen(end);
 633    if (offset < 0)
 634       return false;
 635
 636    return strcmp(str + offset, end) == 0;
 637 }
 638
 639 static void
 640 decode_dynamic_state_pointers(struct gen_batch_decode_ctx *ctx,
 641                               const char *struct_type, const uint32_t *p,
 642                               int count)
 643 {
 644    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 645    struct gen_group *state = gen_spec_find_struct(ctx->spec, struct_type);
 646
 647    uint32_t state_offset = 0;
 648
 649    struct gen_field_iterator iter;
 650    gen_field_iterator_init(&iter, inst, p, 0, false);
 651    while (gen_field_iterator_next(&iter)) {
 652       if (str_ends_with(iter.name, "Pointer")) {
 653          state_offset = iter.raw_value;
 654          break;
 655       }
 656    }
 657
 658    uint64_t state_addr = ctx->dynamic_base + state_offset;
 659    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, state_addr);
 660    const void *state_map = bo.map;
 661
 662    if (state_map == NULL) {
 663       fprintf(ctx->fp, "  dynamic %s state unavailable\n", struct_type);
 664       return;
 665    }
 666
 667    for (int i = 0; i < count; i++) {
 668       fprintf(ctx->fp, "%s %d\n", struct_type, i);
 669       ctx_print_group(ctx, state, state_offset, state_map);
 670
 671       state_addr += state->dw_length * 4;
 672       state_map += state->dw_length;
 673    }
 674 }
 675
 676 static void
 677 decode_3dstate_viewport_state_pointers_cc(struct gen_batch_decode_ctx *ctx,
 678                                           const uint32_t *p)
 679 {
 680    decode_dynamic_state_pointers(ctx, "CC_VIEWPORT", p, 4);
 681 }
 682
 683 static void
 684 decode_3dstate_viewport_state_pointers_sf_clip(struct gen_batch_decode_ctx *ctx,
 685                                                const uint32_t *p)
 686 {
 687    decode_dynamic_state_pointers(ctx, "SF_CLIP_VIEWPORT", p, 4);
 688 }
 689
 690 static void
 691 decode_3dstate_blend_state_pointers(struct gen_batch_decode_ctx *ctx,
 692                                     const uint32_t *p)
 693 {
 694    decode_dynamic_state_pointers(ctx, "BLEND_STATE", p, 1);
 695 }
 696
 697 static void
 698 decode_3dstate_cc_state_pointers(struct gen_batch_decode_ctx *ctx,
 699                                  const uint32_t *p)
 700 {
 701    decode_dynamic_state_pointers(ctx, "COLOR_CALC_STATE", p, 1);
 702 }
 703
 704 static void
 705 decode_3dstate_scissor_state_pointers(struct gen_batch_decode_ctx *ctx,
 706                                       const uint32_t *p)
 707 {
 708    decode_dynamic_state_pointers(ctx, "SCISSOR_RECT", p, 1);
 709 }
 710
 711 static void
 712 decode_load_register_imm(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 713 {
 714    struct gen_group *reg = gen_spec_find_register(ctx->spec, p[1]);
 715
 716    if (reg != NULL) {
 717       fprintf(ctx->fp, "register %s (0x%x): 0x%x\n",
 718               reg->name, reg->register_offset, p[2]);
 719       ctx_print_group(ctx, reg, reg->register_offset, &p[2]);
 720    }
 721 }
 722
 723 struct custom_decoder {
 724    const char *cmd_name;
 725    void (*decode)(struct gen_batch_decode_ctx *ctx, const uint32_t *p);
 726 } custom_decoders[] = {
 727    { "STATE_BASE_ADDRESS", handle_state_base_address },
 728    { "MEDIA_INTERFACE_DESCRIPTOR_LOAD", handle_media_interface_descriptor_load },
 729    { "3DSTATE_VERTEX_BUFFERS", handle_3dstate_vertex_buffers },
 730    { "3DSTATE_INDEX_BUFFER", handle_3dstate_index_buffer },
 731    { "3DSTATE_VS", decode_single_ksp },
 732    { "3DSTATE_GS", decode_single_ksp },
 733    { "3DSTATE_DS", decode_single_ksp },
 734    { "3DSTATE_HS", decode_single_ksp },
 735    { "3DSTATE_PS", decode_ps_kernels },
 736    { "3DSTATE_CONSTANT_VS", decode_3dstate_constant },
 737    { "3DSTATE_CONSTANT_GS", decode_3dstate_constant },
 738    { "3DSTATE_CONSTANT_PS", decode_3dstate_constant },
 739    { "3DSTATE_CONSTANT_HS", decode_3dstate_constant },
 740    { "3DSTATE_CONSTANT_DS", decode_3dstate_constant },
 741
 742    { "3DSTATE_BINDING_TABLE_POINTERS_VS", decode_3dstate_binding_table_pointers },
 743    { "3DSTATE_BINDING_TABLE_POINTERS_HS", decode_3dstate_binding_table_pointers },
 744    { "3DSTATE_BINDING_TABLE_POINTERS_DS", decode_3dstate_binding_table_pointers },
 745    { "3DSTATE_BINDING_TABLE_POINTERS_GS", decode_3dstate_binding_table_pointers },
 746    { "3DSTATE_BINDING_TABLE_POINTERS_PS", decode_3dstate_binding_table_pointers },
 747
 748    { "3DSTATE_SAMPLER_STATE_POINTERS_VS", decode_3dstate_sampler_state_pointers },
 749    { "3DSTATE_SAMPLER_STATE_POINTERS_HS", decode_3dstate_sampler_state_pointers },
 750    { "3DSTATE_SAMPLER_STATE_POINTERS_DS", decode_3dstate_sampler_state_pointers },
 751    { "3DSTATE_SAMPLER_STATE_POINTERS_GS", decode_3dstate_sampler_state_pointers },
 752    { "3DSTATE_SAMPLER_STATE_POINTERS_PS", decode_3dstate_sampler_state_pointers },
 753    { "3DSTATE_SAMPLER_STATE_POINTERS", decode_3dstate_sampler_state_pointers_gen6 },
 754
 755    { "3DSTATE_VIEWPORT_STATE_POINTERS_CC", decode_3dstate_viewport_state_pointers_cc },
 756    { "3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP", decode_3dstate_viewport_state_pointers_sf_clip },
 757    { "3DSTATE_BLEND_STATE_POINTERS", decode_3dstate_blend_state_pointers },
 758    { "3DSTATE_CC_STATE_POINTERS", decode_3dstate_cc_state_pointers },
 759    { "3DSTATE_SCISSOR_STATE_POINTERS", decode_3dstate_scissor_state_pointers },
 760    { "MI_LOAD_REGISTER_IMM", decode_load_register_imm }
 761 };
 762
 763 static inline uint64_t
 764 get_address(struct gen_spec *spec, const uint32_t *p)
 765 {
 766    /* Addresses are always guaranteed to be page-aligned and sometimes
 767     * hardware packets have extra stuff stuffed in the bottom 12 bits.
 768     */
 769    uint64_t addr = p[0] & ~0xfffu;
 770
 771    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 772       /* On Broadwell and above, we have 48-bit addresses which consume two
 773        * dwords.  Some packets require that these get stored in a "canonical
 774        * form" which means that bit 47 is sign-extended through the upper
 775        * bits. In order to correctly handle those aub dumps, we need to mask
 776        * off the top 16 bits.
 777        */
 778       addr |= ((uint64_t)p[1] & 0xffff) << 32;
 779    }
 780
 781    return addr;
 782 }
 783
 784 void
 785 gen_print_batch(struct gen_batch_decode_ctx *ctx,
 786                 const uint32_t *batch, uint32_t batch_size,
 787                 uint64_t batch_addr)
 788 {
 789    const uint32_t *p, *end = batch + batch_size;
 790    int length;
 791    struct gen_group *inst;
 792
 793    for (p = batch; p < end; p += length) {
 794       inst = gen_spec_find_instruction(ctx->spec, p);
 795       length = gen_group_get_length(inst, p);
 796       assert(inst == NULL || length > 0);
 797       length = MAX2(1, length);
 798
 799       const char *reset_color = ctx->flags & GEN_BATCH_DECODE_IN_COLOR ? NORMAL : "";
 800
 801       uint64_t offset;
 802       if (ctx->flags & GEN_BATCH_DECODE_OFFSETS)
 803          offset = batch_addr + ((char *)p - (char *)batch);
 804       else
 805          offset = 0;
 806
 807       if (inst == NULL) {
 808          fprintf(ctx->fp, "%s0x%08"PRIx64": unknown instruction %08x%s\n",
 809                  (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) ? RED_COLOR : "",
 810                  offset, p[0], reset_color);
 811          continue;
 812       }
 813
 814       const char *color;
 815       const char *inst_name = gen_group_get_name(inst);
 816       if (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) {
 817          reset_color = NORMAL;
 818          if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 819             if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0 ||
 820                 strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0)
 821                color = GREEN_HEADER;
 822             else
 823                color = BLUE_HEADER;
 824          } else {
 825             color = NORMAL;
 826          }
 827       } else {
 828          color = "";
 829          reset_color = "";
 830       }
 831
 832       fprintf(ctx->fp, "%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 833               color, offset, p[0], inst_name, reset_color);
 834
 835       if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 836          ctx_print_group(ctx, inst, offset, p);
 837
 838          for (int i = 0; i < ARRAY_LENGTH(custom_decoders); i++) {
 839             if (strcmp(inst_name, custom_decoders[i].cmd_name) == 0) {
 840                custom_decoders[i].decode(ctx, p);
 841                break;
 842             }
 843          }
 844       }
 845
 846       if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0) {
 847          struct gen_batch_decode_bo next_batch = {};
 848          bool second_level;
 849          struct gen_field_iterator iter;
 850          gen_field_iterator_init(&iter, inst, p, 0, false);
 851          while (gen_field_iterator_next(&iter)) {
 852             if (strcmp(iter.name, "Batch Buffer Start Address") == 0) {
 853                next_batch = ctx_get_bo(ctx, iter.raw_value);
 854             } else if (strcmp(iter.name, "Second Level Batch Buffer") == 0) {
 855                second_level = iter.raw_value;
 856             }
 857          }
 858
 859          if (next_batch.map == NULL) {
 860             fprintf(ctx->fp, "Secondary batch at 0x%08"PRIx64" unavailable\n",
 861                     next_batch.addr);
 862          } else {
 863             gen_print_batch(ctx, next_batch.map, next_batch.size,
 864                             next_batch.addr);
 865          }
 866          if (second_level) {
 867             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 868              * like a subroutine call.  Commands that come afterwards get
 869              * processed once the 2nd level batch buffer returns with
 870              * MI_BATCH_BUFFER_END.
 871              */
 872             continue;
 873          } else {
 874             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 875              * like a goto.  Nothing after it will ever get processed.  In
 876              * order to prevent the recursion from growing, we just reset the
 877              * loop and continue;
 878              */
 879             break;
 880          }
 881       } else if (strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0) {
 882          break;
 883       }
 884    }
 885 }