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