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 #include "util/macros.h"
  27
  28 #include <string.h>
  29
  30 void
  31 gen_batch_decode_ctx_init(struct gen_batch_decode_ctx *ctx,
  32                           const struct gen_device_info *devinfo,
  33                           FILE *fp, enum gen_batch_decode_flags flags,
  34                           const char *xml_path,
  35                           struct gen_batch_decode_bo (*get_bo)(void *,
  36                                                                uint64_t),
  37                           unsigned (*get_state_size)(void *, uint32_t),
  38                           void *user_data)
  39 {
  40    memset(ctx, 0, sizeof(*ctx));
  41
  42    ctx->get_bo = get_bo;
  43    ctx->get_state_size = get_state_size;
  44    ctx->user_data = user_data;
  45    ctx->fp = fp;
  46    ctx->flags = flags;
  47    ctx->max_vbo_decoded_lines = -1; /* No limit! */
  48
  49    if (xml_path == NULL)
  50       ctx->spec = gen_spec_load(devinfo);
  51    else
  52       ctx->spec = gen_spec_load_from_path(devinfo, xml_path);
  53    ctx->disasm = gen_disasm_create(devinfo);
  54 }
  55
  56 void
  57 gen_batch_decode_ctx_finish(struct gen_batch_decode_ctx *ctx)
  58 {
  59    gen_spec_destroy(ctx->spec);
  60    gen_disasm_destroy(ctx->disasm);
  61 }
  62
  63 #define CSI "\e["
  64 #define RED_COLOR    CSI "31m"
  65 #define BLUE_HEADER  CSI "0;44m"
  66 #define GREEN_HEADER CSI "1;42m"
  67 #define NORMAL       CSI "0m"
  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    uint64_t addr = ctx->instruction_base + ksp;
 131    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, addr);
 132    if (!bo.map)
 133       return;
 134
 135    fprintf(ctx->fp, "\nReferenced %s:\n", type);
 136    gen_disasm_disassemble(ctx->disasm, bo.map, 0, 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                  int max_lines)
 170 {
 171    const uint32_t *dw_end = bo.map + MIN2(bo.size, read_length);
 172
 173    int column_count = 0, line_count = -1;
 174    for (const uint32_t *dw = bo.map; dw < dw_end; dw++) {
 175       if (column_count * 4 == pitch || column_count == 8) {
 176          fprintf(ctx->fp, "\n");
 177          column_count = 0;
 178          line_count++;
 179
 180          if (max_lines >= 0 && line_count >= max_lines)
 181             break;
 182       }
 183       fprintf(ctx->fp, column_count == 0 ? "  " : " ");
 184
 185       if ((ctx->flags & GEN_BATCH_DECODE_FLOATS) && probably_float(*dw))
 186          fprintf(ctx->fp, "  %8.2f", *(float *) dw);
 187       else
 188          fprintf(ctx->fp, "  0x%08x", *dw);
 189
 190       column_count++;
 191    }
 192    fprintf(ctx->fp, "\n");
 193 }
 194
 195 static void
 196 handle_state_base_address(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 197 {
 198    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 199
 200    struct gen_field_iterator iter;
 201    gen_field_iterator_init(&iter, inst, p, 0, false);
 202
 203    while (gen_field_iterator_next(&iter)) {
 204       if (strcmp(iter.name, "Surface State Base Address") == 0) {
 205          ctx->surface_base = iter.raw_value;
 206       } else if (strcmp(iter.name, "Dynamic State Base Address") == 0) {
 207          ctx->dynamic_base = iter.raw_value;
 208       } else if (strcmp(iter.name, "Instruction Base Address") == 0) {
 209          ctx->instruction_base = iter.raw_value;
 210       }
 211    }
 212 }
 213
 214 static void
 215 dump_binding_table(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 216 {
 217    struct gen_group *strct =
 218       gen_spec_find_struct(ctx->spec, "RENDER_SURFACE_STATE");
 219    if (strct == NULL) {
 220       fprintf(ctx->fp, "did not find RENDER_SURFACE_STATE info\n");
 221       return;
 222    }
 223
 224    if (count < 0)
 225       count = update_count(ctx, offset, 1, 8);
 226
 227    if (offset % 32 != 0 || offset >= UINT16_MAX) {
 228       fprintf(ctx->fp, "  invalid binding table pointer\n");
 229       return;
 230    }
 231
 232    struct gen_batch_decode_bo bind_bo =
 233       ctx_get_bo(ctx, ctx->surface_base + offset);
 234
 235    if (bind_bo.map == NULL) {
 236       fprintf(ctx->fp, "  binding table unavailable\n");
 237       return;
 238    }
 239
 240    const uint32_t *pointers = bind_bo.map;
 241    for (int i = 0; i < count; i++) {
 242       if (pointers[i] == 0)
 243          continue;
 244
 245       uint64_t addr = ctx->surface_base + pointers[i];
 246       struct gen_batch_decode_bo bo = ctx_get_bo(ctx, addr);
 247       uint32_t size = strct->dw_length * 4;
 248
 249       if (pointers[i] % 32 != 0 ||
 250           addr < bo.addr || addr + size >= bo.addr + bo.size) {
 251          fprintf(ctx->fp, "pointer %u: %08x <not valid>\n", i, pointers[i]);
 252          continue;
 253       }
 254
 255       fprintf(ctx->fp, "pointer %u: %08x\n", i, pointers[i]);
 256       ctx_print_group(ctx, strct, addr, bo.map + (addr - bo.addr));
 257    }
 258 }
 259
 260 static void
 261 dump_samplers(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 262 {
 263    struct gen_group *strct = gen_spec_find_struct(ctx->spec, "SAMPLER_STATE");
 264
 265    if (count < 0)
 266       count = update_count(ctx, offset, strct->dw_length, 4);
 267
 268    uint64_t state_addr = ctx->dynamic_base + offset;
 269    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, state_addr);
 270    const void *state_map = bo.map;
 271
 272    if (state_map == NULL) {
 273       fprintf(ctx->fp, "  samplers unavailable\n");
 274       return;
 275    }
 276
 277    if (offset % 32 != 0 || state_addr - bo.addr >= bo.size) {
 278       fprintf(ctx->fp, "  invalid sampler state pointer\n");
 279       return;
 280    }
 281
 282    for (int i = 0; i < count; i++) {
 283       fprintf(ctx->fp, "sampler state %d\n", i);
 284       ctx_print_group(ctx, strct, state_addr, state_map);
 285       state_addr += 16;
 286       state_map += 16;
 287    }
 288 }
 289
 290 static void
 291 handle_media_interface_descriptor_load(struct gen_batch_decode_ctx *ctx,
 292                                        const uint32_t *p)
 293 {
 294    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 295    struct gen_group *desc =
 296       gen_spec_find_struct(ctx->spec, "INTERFACE_DESCRIPTOR_DATA");
 297
 298    struct gen_field_iterator iter;
 299    gen_field_iterator_init(&iter, inst, p, 0, false);
 300    uint32_t descriptor_offset = 0;
 301    int descriptor_count = 0;
 302    while (gen_field_iterator_next(&iter)) {
 303       if (strcmp(iter.name, "Interface Descriptor Data Start Address") == 0) {
 304          descriptor_offset = strtol(iter.value, NULL, 16);
 305       } else if (strcmp(iter.name, "Interface Descriptor Total Length") == 0) {
 306          descriptor_count =
 307             strtol(iter.value, NULL, 16) / (desc->dw_length * 4);
 308       }
 309    }
 310
 311    uint64_t desc_addr = ctx->dynamic_base + descriptor_offset;
 312    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, desc_addr);
 313    const void *desc_map = bo.map;
 314
 315    if (desc_map == NULL) {
 316       fprintf(ctx->fp, "  interface descriptors unavailable\n");
 317       return;
 318    }
 319
 320    for (int i = 0; i < descriptor_count; i++) {
 321       fprintf(ctx->fp, "descriptor %d: %08x\n", i, descriptor_offset);
 322
 323       ctx_print_group(ctx, desc, desc_addr, desc_map);
 324
 325       gen_field_iterator_init(&iter, desc, desc_map, 0, false);
 326       uint64_t ksp = 0;
 327       uint32_t sampler_offset = 0, sampler_count = 0;
 328       uint32_t binding_table_offset = 0, binding_entry_count = 0;
 329       while (gen_field_iterator_next(&iter)) {
 330          if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 331             ksp = strtoll(iter.value, NULL, 16);
 332          } else if (strcmp(iter.name, "Sampler State Pointer") == 0) {
 333             sampler_offset = strtol(iter.value, NULL, 16);
 334          } else if (strcmp(iter.name, "Sampler Count") == 0) {
 335             sampler_count = strtol(iter.value, NULL, 10);
 336          } else if (strcmp(iter.name, "Binding Table Pointer") == 0) {
 337             binding_table_offset = strtol(iter.value, NULL, 16);
 338          } else if (strcmp(iter.name, "Binding Table Entry Count") == 0) {
 339             binding_entry_count = strtol(iter.value, NULL, 10);
 340          }
 341       }
 342
 343       ctx_disassemble_program(ctx, ksp, "compute shader");
 344       printf("\n");
 345
 346       dump_samplers(ctx, sampler_offset, sampler_count);
 347       dump_binding_table(ctx, binding_table_offset, binding_entry_count);
 348
 349       desc_map += desc->dw_length;
 350       desc_addr += desc->dw_length * 4;
 351    }
 352 }
 353
 354 static void
 355 handle_3dstate_vertex_buffers(struct gen_batch_decode_ctx *ctx,
 356                               const uint32_t *p)
 357 {
 358    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 359    struct gen_group *vbs = gen_spec_find_struct(ctx->spec, "VERTEX_BUFFER_STATE");
 360
 361    struct gen_batch_decode_bo vb = {};
 362    uint32_t vb_size = 0;
 363    int index = -1;
 364    int pitch = -1;
 365    bool ready = false;
 366
 367    struct gen_field_iterator iter;
 368    gen_field_iterator_init(&iter, inst, p, 0, false);
 369    while (gen_field_iterator_next(&iter)) {
 370       if (iter.struct_desc != vbs)
 371          continue;
 372
 373       struct gen_field_iterator vbs_iter;
 374       gen_field_iterator_init(&vbs_iter, vbs, &iter.p[iter.start_bit / 32], 0, false);
 375       while (gen_field_iterator_next(&vbs_iter)) {
 376          if (strcmp(vbs_iter.name, "Vertex Buffer Index") == 0) {
 377             index = vbs_iter.raw_value;
 378          } else if (strcmp(vbs_iter.name, "Buffer Pitch") == 0) {
 379             pitch = vbs_iter.raw_value;
 380          } else if (strcmp(vbs_iter.name, "Buffer Starting Address") == 0) {
 381             vb = ctx_get_bo(ctx, vbs_iter.raw_value);
 382          } else if (strcmp(vbs_iter.name, "Buffer Size") == 0) {
 383             vb_size = vbs_iter.raw_value;
 384             ready = true;
 385          } else if (strcmp(vbs_iter.name, "End Address") == 0) {
 386             if (vb.map && vbs_iter.raw_value >= vb.addr)
 387                vb_size = vbs_iter.raw_value - vb.addr;
 388             else
 389                vb_size = 0;
 390             ready = true;
 391          }
 392
 393          if (!ready)
 394             continue;
 395
 396          fprintf(ctx->fp, "vertex buffer %d, size %d\n", index, vb_size);
 397
 398          if (vb.map == NULL) {
 399             fprintf(ctx->fp, "  buffer contents unavailable\n");
 400             continue;
 401          }
 402
 403          if (vb.map == 0 || vb_size == 0)
 404             continue;
 405
 406          ctx_print_buffer(ctx, vb, vb_size, pitch, ctx->max_vbo_decoded_lines);
 407
 408          vb.map = NULL;
 409          vb_size = 0;
 410          index = -1;
 411          pitch = -1;
 412          ready = false;
 413       }
 414    }
 415 }
 416
 417 static void
 418 handle_3dstate_index_buffer(struct gen_batch_decode_ctx *ctx,
 419                             const uint32_t *p)
 420 {
 421    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 422
 423    struct gen_batch_decode_bo ib = {};
 424    uint32_t ib_size = 0;
 425    uint32_t format = 0;
 426
 427    struct gen_field_iterator iter;
 428    gen_field_iterator_init(&iter, inst, p, 0, false);
 429    while (gen_field_iterator_next(&iter)) {
 430       if (strcmp(iter.name, "Index Format") == 0) {
 431          format = iter.raw_value;
 432       } else if (strcmp(iter.name, "Buffer Starting Address") == 0) {
 433          ib = ctx_get_bo(ctx, iter.raw_value);
 434       } else if (strcmp(iter.name, "Buffer Size") == 0) {
 435          ib_size = iter.raw_value;
 436       }
 437    }
 438
 439    if (ib.map == NULL) {
 440       fprintf(ctx->fp, "  buffer contents unavailable\n");
 441       return;
 442    }
 443
 444    const void *m = ib.map;
 445    const void *ib_end = ib.map + MIN2(ib.size, ib_size);
 446    for (int i = 0; m < ib_end && i < 10; i++) {
 447       switch (format) {
 448       case 0:
 449          fprintf(ctx->fp, "%3d ", *(uint8_t *)m);
 450          m += 1;
 451          break;
 452       case 1:
 453          fprintf(ctx->fp, "%3d ", *(uint16_t *)m);
 454          m += 2;
 455          break;
 456       case 2:
 457          fprintf(ctx->fp, "%3d ", *(uint32_t *)m);
 458          m += 4;
 459          break;
 460       }
 461    }
 462
 463    if (m < ib_end)
 464       fprintf(ctx->fp, "...");
 465    fprintf(ctx->fp, "\n");
 466 }
 467
 468 static void
 469 decode_single_ksp(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 470 {
 471    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 472
 473    uint64_t ksp = 0;
 474    bool is_simd8 = false; /* vertex shaders on Gen8+ only */
 475    bool is_enabled = true;
 476
 477    struct gen_field_iterator iter;
 478    gen_field_iterator_init(&iter, inst, p, 0, false);
 479    while (gen_field_iterator_next(&iter)) {
 480       if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 481          ksp = iter.raw_value;
 482       } else if (strcmp(iter.name, "SIMD8 Dispatch Enable") == 0) {
 483          is_simd8 = iter.raw_value;
 484       } else if (strcmp(iter.name, "Dispatch Mode") == 0) {
 485          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 486       } else if (strcmp(iter.name, "Dispatch Enable") == 0) {
 487          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 488       } else if (strcmp(iter.name, "Enable") == 0) {
 489          is_enabled = iter.raw_value;
 490       }
 491    }
 492
 493    const char *type =
 494       strcmp(inst->name,   "VS_STATE") == 0 ? "vertex shader" :
 495       strcmp(inst->name,   "GS_STATE") == 0 ? "geometry shader" :
 496       strcmp(inst->name,   "SF_STATE") == 0 ? "strips and fans shader" :
 497       strcmp(inst->name, "CLIP_STATE") == 0 ? "clip shader" :
 498       strcmp(inst->name, "3DSTATE_DS") == 0 ? "tessellation evaluation shader" :
 499       strcmp(inst->name, "3DSTATE_HS") == 0 ? "tessellation control shader" :
 500       strcmp(inst->name, "3DSTATE_VS") == 0 ? (is_simd8 ? "SIMD8 vertex shader" : "vec4 vertex shader") :
 501       strcmp(inst->name, "3DSTATE_GS") == 0 ? (is_simd8 ? "SIMD8 geometry shader" : "vec4 geometry shader") :
 502       NULL;
 503
 504    if (is_enabled) {
 505       ctx_disassemble_program(ctx, ksp, type);
 506       printf("\n");
 507    }
 508 }
 509
 510 static void
 511 decode_ps_kernels(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 512 {
 513    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 514
 515    uint64_t ksp[3] = {0, 0, 0};
 516    bool enabled[3] = {false, false, false};
 517
 518    struct gen_field_iterator iter;
 519    gen_field_iterator_init(&iter, inst, p, 0, false);
 520    while (gen_field_iterator_next(&iter)) {
 521       if (strncmp(iter.name, "Kernel Start Pointer ",
 522                   strlen("Kernel Start Pointer ")) == 0) {
 523          int idx = iter.name[strlen("Kernel Start Pointer ")] - '0';
 524          ksp[idx] = strtol(iter.value, NULL, 16);
 525       } else if (strcmp(iter.name, "8 Pixel Dispatch Enable") == 0) {
 526          enabled[0] = strcmp(iter.value, "true") == 0;
 527       } else if (strcmp(iter.name, "16 Pixel Dispatch Enable") == 0) {
 528          enabled[1] = strcmp(iter.value, "true") == 0;
 529       } else if (strcmp(iter.name, "32 Pixel Dispatch Enable") == 0) {
 530          enabled[2] = strcmp(iter.value, "true") == 0;
 531       }
 532    }
 533
 534    /* Reorder KSPs to be [8, 16, 32] instead of the hardware order. */
 535    if (enabled[0] + enabled[1] + enabled[2] == 1) {
 536       if (enabled[1]) {
 537          ksp[1] = ksp[0];
 538          ksp[0] = 0;
 539       } else if (enabled[2]) {
 540          ksp[2] = ksp[0];
 541          ksp[0] = 0;
 542       }
 543    } else {
 544       uint64_t tmp = ksp[1];
 545       ksp[1] = ksp[2];
 546       ksp[2] = tmp;
 547    }
 548
 549    if (enabled[0])
 550       ctx_disassemble_program(ctx, ksp[0], "SIMD8 fragment shader");
 551    if (enabled[1])
 552       ctx_disassemble_program(ctx, ksp[1], "SIMD16 fragment shader");
 553    if (enabled[2])
 554       ctx_disassemble_program(ctx, ksp[2], "SIMD32 fragment shader");
 555    fprintf(ctx->fp, "\n");
 556 }
 557
 558 static void
 559 decode_3dstate_constant(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 560 {
 561    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 562    struct gen_group *body =
 563       gen_spec_find_struct(ctx->spec, "3DSTATE_CONSTANT_BODY");
 564
 565    uint32_t read_length[4] = {0};
 566    uint64_t read_addr[4];
 567
 568    struct gen_field_iterator outer;
 569    gen_field_iterator_init(&outer, inst, p, 0, false);
 570    while (gen_field_iterator_next(&outer)) {
 571       if (outer.struct_desc != body)
 572          continue;
 573
 574       struct gen_field_iterator iter;
 575       gen_field_iterator_init(&iter, body, &outer.p[outer.start_bit / 32],
 576                               0, false);
 577
 578       while (gen_field_iterator_next(&iter)) {
 579          int idx;
 580          if (sscanf(iter.name, "Read Length[%d]", &idx) == 1) {
 581             read_length[idx] = iter.raw_value;
 582          } else if (sscanf(iter.name, "Buffer[%d]", &idx) == 1) {
 583             read_addr[idx] = iter.raw_value;
 584          }
 585       }
 586
 587       for (int i = 0; i < 4; i++) {
 588          if (read_length[i] == 0)
 589             continue;
 590
 591          struct gen_batch_decode_bo buffer = ctx_get_bo(ctx, read_addr[i]);
 592          if (!buffer.map) {
 593             fprintf(ctx->fp, "constant buffer %d unavailable\n", i);
 594             continue;
 595          }
 596
 597          unsigned size = read_length[i] * 32;
 598          fprintf(ctx->fp, "constant buffer %d, size %u\n", i, size);
 599
 600          ctx_print_buffer(ctx, buffer, size, 0, -1);
 601       }
 602    }
 603 }
 604
 605 static void
 606 decode_3dstate_binding_table_pointers(struct gen_batch_decode_ctx *ctx,
 607                                       const uint32_t *p)
 608 {
 609    dump_binding_table(ctx, p[1], -1);
 610 }
 611
 612 static void
 613 decode_3dstate_sampler_state_pointers(struct gen_batch_decode_ctx *ctx,
 614                                       const uint32_t *p)
 615 {
 616    dump_samplers(ctx, p[1], -1);
 617 }
 618
 619 static void
 620 decode_3dstate_sampler_state_pointers_gen6(struct gen_batch_decode_ctx *ctx,
 621                                            const uint32_t *p)
 622 {
 623    dump_samplers(ctx, p[1], -1);
 624    dump_samplers(ctx, p[2], -1);
 625    dump_samplers(ctx, p[3], -1);
 626 }
 627
 628 static bool
 629 str_ends_with(const char *str, const char *end)
 630 {
 631    int offset = strlen(str) - strlen(end);
 632    if (offset < 0)
 633       return false;
 634
 635    return strcmp(str + offset, end) == 0;
 636 }
 637
 638 static void
 639 decode_dynamic_state_pointers(struct gen_batch_decode_ctx *ctx,
 640                               const char *struct_type, const uint32_t *p,
 641                               int count)
 642 {
 643    struct gen_group *inst = gen_spec_find_instruction(ctx->spec, p);
 644
 645    uint32_t state_offset = 0;
 646
 647    struct gen_field_iterator iter;
 648    gen_field_iterator_init(&iter, inst, p, 0, false);
 649    while (gen_field_iterator_next(&iter)) {
 650       if (str_ends_with(iter.name, "Pointer")) {
 651          state_offset = iter.raw_value;
 652          break;
 653       }
 654    }
 655
 656    uint64_t state_addr = ctx->dynamic_base + state_offset;
 657    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, state_addr);
 658    const void *state_map = bo.map;
 659
 660    if (state_map == NULL) {
 661       fprintf(ctx->fp, "  dynamic %s state unavailable\n", struct_type);
 662       return;
 663    }
 664
 665    struct gen_group *state = gen_spec_find_struct(ctx->spec, struct_type);
 666    if (strcmp(struct_type, "BLEND_STATE") == 0) {
 667       /* Blend states are different from the others because they have a header
 668        * struct called BLEND_STATE which is followed by a variable number of
 669        * BLEND_STATE_ENTRY structs.
 670        */
 671       fprintf(ctx->fp, "%s\n", struct_type);
 672       ctx_print_group(ctx, state, state_addr, state_map);
 673
 674       state_addr += state->dw_length * 4;
 675       state_map += state->dw_length * 4;
 676
 677       struct_type = "BLEND_STATE_ENTRY";
 678       state = gen_spec_find_struct(ctx->spec, struct_type);
 679    }
 680
 681    for (int i = 0; i < count; i++) {
 682       fprintf(ctx->fp, "%s %d\n", struct_type, i);
 683       ctx_print_group(ctx, state, state_addr, state_map);
 684
 685       state_addr += state->dw_length * 4;
 686       state_map += state->dw_length * 4;
 687    }
 688 }
 689
 690 static void
 691 decode_3dstate_viewport_state_pointers_cc(struct gen_batch_decode_ctx *ctx,
 692                                           const uint32_t *p)
 693 {
 694    decode_dynamic_state_pointers(ctx, "CC_VIEWPORT", p, 4);
 695 }
 696
 697 static void
 698 decode_3dstate_viewport_state_pointers_sf_clip(struct gen_batch_decode_ctx *ctx,
 699                                                const uint32_t *p)
 700 {
 701    decode_dynamic_state_pointers(ctx, "SF_CLIP_VIEWPORT", p, 4);
 702 }
 703
 704 static void
 705 decode_3dstate_blend_state_pointers(struct gen_batch_decode_ctx *ctx,
 706                                     const uint32_t *p)
 707 {
 708    decode_dynamic_state_pointers(ctx, "BLEND_STATE", p, 1);
 709 }
 710
 711 static void
 712 decode_3dstate_cc_state_pointers(struct gen_batch_decode_ctx *ctx,
 713                                  const uint32_t *p)
 714 {
 715    decode_dynamic_state_pointers(ctx, "COLOR_CALC_STATE", p, 1);
 716 }
 717
 718 static void
 719 decode_3dstate_scissor_state_pointers(struct gen_batch_decode_ctx *ctx,
 720                                       const uint32_t *p)
 721 {
 722    decode_dynamic_state_pointers(ctx, "SCISSOR_RECT", p, 1);
 723 }
 724
 725 static void
 726 decode_load_register_imm(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 727 {
 728    struct gen_group *reg = gen_spec_find_register(ctx->spec, p[1]);
 729
 730    if (reg != NULL) {
 731       fprintf(ctx->fp, "register %s (0x%x): 0x%x\n",
 732               reg->name, reg->register_offset, p[2]);
 733       ctx_print_group(ctx, reg, reg->register_offset, &p[2]);
 734    }
 735 }
 736
 737 struct custom_decoder {
 738    const char *cmd_name;
 739    void (*decode)(struct gen_batch_decode_ctx *ctx, const uint32_t *p);
 740 } custom_decoders[] = {
 741    { "STATE_BASE_ADDRESS", handle_state_base_address },
 742    { "MEDIA_INTERFACE_DESCRIPTOR_LOAD", handle_media_interface_descriptor_load },
 743    { "3DSTATE_VERTEX_BUFFERS", handle_3dstate_vertex_buffers },
 744    { "3DSTATE_INDEX_BUFFER", handle_3dstate_index_buffer },
 745    { "3DSTATE_VS", decode_single_ksp },
 746    { "3DSTATE_GS", decode_single_ksp },
 747    { "3DSTATE_DS", decode_single_ksp },
 748    { "3DSTATE_HS", decode_single_ksp },
 749    { "3DSTATE_PS", decode_ps_kernels },
 750    { "3DSTATE_CONSTANT_VS", decode_3dstate_constant },
 751    { "3DSTATE_CONSTANT_GS", decode_3dstate_constant },
 752    { "3DSTATE_CONSTANT_PS", decode_3dstate_constant },
 753    { "3DSTATE_CONSTANT_HS", decode_3dstate_constant },
 754    { "3DSTATE_CONSTANT_DS", decode_3dstate_constant },
 755
 756    { "3DSTATE_BINDING_TABLE_POINTERS_VS", decode_3dstate_binding_table_pointers },
 757    { "3DSTATE_BINDING_TABLE_POINTERS_HS", decode_3dstate_binding_table_pointers },
 758    { "3DSTATE_BINDING_TABLE_POINTERS_DS", decode_3dstate_binding_table_pointers },
 759    { "3DSTATE_BINDING_TABLE_POINTERS_GS", decode_3dstate_binding_table_pointers },
 760    { "3DSTATE_BINDING_TABLE_POINTERS_PS", decode_3dstate_binding_table_pointers },
 761
 762    { "3DSTATE_SAMPLER_STATE_POINTERS_VS", decode_3dstate_sampler_state_pointers },
 763    { "3DSTATE_SAMPLER_STATE_POINTERS_HS", decode_3dstate_sampler_state_pointers },
 764    { "3DSTATE_SAMPLER_STATE_POINTERS_DS", decode_3dstate_sampler_state_pointers },
 765    { "3DSTATE_SAMPLER_STATE_POINTERS_GS", decode_3dstate_sampler_state_pointers },
 766    { "3DSTATE_SAMPLER_STATE_POINTERS_PS", decode_3dstate_sampler_state_pointers },
 767    { "3DSTATE_SAMPLER_STATE_POINTERS", decode_3dstate_sampler_state_pointers_gen6 },
 768
 769    { "3DSTATE_VIEWPORT_STATE_POINTERS_CC", decode_3dstate_viewport_state_pointers_cc },
 770    { "3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP", decode_3dstate_viewport_state_pointers_sf_clip },
 771    { "3DSTATE_BLEND_STATE_POINTERS", decode_3dstate_blend_state_pointers },
 772    { "3DSTATE_CC_STATE_POINTERS", decode_3dstate_cc_state_pointers },
 773    { "3DSTATE_SCISSOR_STATE_POINTERS", decode_3dstate_scissor_state_pointers },
 774    { "MI_LOAD_REGISTER_IMM", decode_load_register_imm }
 775 };
 776
 777 static inline uint64_t
 778 get_address(struct gen_spec *spec, const uint32_t *p)
 779 {
 780    /* Addresses are always guaranteed to be page-aligned and sometimes
 781     * hardware packets have extra stuff stuffed in the bottom 12 bits.
 782     */
 783    uint64_t addr = p[0] & ~0xfffu;
 784
 785    if (gen_spec_get_gen(spec) >= gen_make_gen(8,0)) {
 786       /* On Broadwell and above, we have 48-bit addresses which consume two
 787        * dwords.  Some packets require that these get stored in a "canonical
 788        * form" which means that bit 47 is sign-extended through the upper
 789        * bits. In order to correctly handle those aub dumps, we need to mask
 790        * off the top 16 bits.
 791        */
 792       addr |= ((uint64_t)p[1] & 0xffff) << 32;
 793    }
 794
 795    return addr;
 796 }
 797
 798 void
 799 gen_print_batch(struct gen_batch_decode_ctx *ctx,
 800                 const uint32_t *batch, uint32_t batch_size,
 801                 uint64_t batch_addr)
 802 {
 803    const uint32_t *p, *end = batch + batch_size;
 804    int length;
 805    struct gen_group *inst;
 806
 807    for (p = batch; p < end; p += length) {
 808       inst = gen_spec_find_instruction(ctx->spec, p);
 809       length = gen_group_get_length(inst, p);
 810       assert(inst == NULL || length > 0);
 811       length = MAX2(1, length);
 812
 813       const char *reset_color = ctx->flags & GEN_BATCH_DECODE_IN_COLOR ? NORMAL : "";
 814
 815       uint64_t offset;
 816       if (ctx->flags & GEN_BATCH_DECODE_OFFSETS)
 817          offset = batch_addr + ((char *)p - (char *)batch);
 818       else
 819          offset = 0;
 820
 821       if (inst == NULL) {
 822          fprintf(ctx->fp, "%s0x%08"PRIx64": unknown instruction %08x%s\n",
 823                  (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) ? RED_COLOR : "",
 824                  offset, p[0], reset_color);
 825          continue;
 826       }
 827
 828       const char *color;
 829       const char *inst_name = gen_group_get_name(inst);
 830       if (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) {
 831          reset_color = NORMAL;
 832          if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 833             if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0 ||
 834                 strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0)
 835                color = GREEN_HEADER;
 836             else
 837                color = BLUE_HEADER;
 838          } else {
 839             color = NORMAL;
 840          }
 841       } else {
 842          color = "";
 843          reset_color = "";
 844       }
 845
 846       fprintf(ctx->fp, "%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 847               color, offset, p[0], inst_name, reset_color);
 848
 849       if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 850          ctx_print_group(ctx, inst, offset, p);
 851
 852          for (int i = 0; i < ARRAY_SIZE(custom_decoders); i++) {
 853             if (strcmp(inst_name, custom_decoders[i].cmd_name) == 0) {
 854                custom_decoders[i].decode(ctx, p);
 855                break;
 856             }
 857          }
 858       }
 859
 860       if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0) {
 861          struct gen_batch_decode_bo next_batch = {};
 862          bool second_level;
 863          struct gen_field_iterator iter;
 864          gen_field_iterator_init(&iter, inst, p, 0, false);
 865          while (gen_field_iterator_next(&iter)) {
 866             if (strcmp(iter.name, "Batch Buffer Start Address") == 0) {
 867                next_batch = ctx_get_bo(ctx, iter.raw_value);
 868             } else if (strcmp(iter.name, "Second Level Batch Buffer") == 0) {
 869                second_level = iter.raw_value;
 870             }
 871          }
 872
 873          if (next_batch.map == NULL) {
 874             fprintf(ctx->fp, "Secondary batch at 0x%08"PRIx64" unavailable\n",
 875                     next_batch.addr);
 876          } else {
 877             gen_print_batch(ctx, next_batch.map, next_batch.size,
 878                             next_batch.addr);
 879          }
 880          if (second_level) {
 881             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 882              * like a subroutine call.  Commands that come afterwards get
 883              * processed once the 2nd level batch buffer returns with
 884              * MI_BATCH_BUFFER_END.
 885              */
 886             continue;
 887          } else {
 888             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 889              * like a goto.  Nothing after it will ever get processed.  In
 890              * order to prevent the recursion from growing, we just reset the
 891              * loop and continue;
 892              */
 893             break;
 894          }
 895       } else if (strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0) {
 896          break;
 897       }
 898    }
 899 }