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 #include "main/macros.h" /* Needed for ROUND_DOWN_TO */
  28
  29 #include <string.h>
  30
  31 void
  32 gen_batch_decode_ctx_init(struct gen_batch_decode_ctx *ctx,
  33                           const struct gen_device_info *devinfo,
  34                           FILE *fp, enum gen_batch_decode_flags flags,
  35                           const char *xml_path,
  36                           struct gen_batch_decode_bo (*get_bo)(void *,
  37                                                                bool,
  38                                                                uint64_t),
  39                           unsigned (*get_state_size)(void *, uint32_t),
  40                           void *user_data)
  41 {
  42    memset(ctx, 0, sizeof(*ctx));
  43
  44    ctx->get_bo = get_bo;
  45    ctx->get_state_size = get_state_size;
  46    ctx->user_data = user_data;
  47    ctx->fp = fp;
  48    ctx->flags = flags;
  49    ctx->max_vbo_decoded_lines = -1; /* No limit! */
  50    ctx->engine = I915_ENGINE_CLASS_RENDER;
  51
  52    if (xml_path == NULL)
  53       ctx->spec = gen_spec_load(devinfo);
  54    else
  55       ctx->spec = gen_spec_load_from_path(devinfo, xml_path);
  56    ctx->disasm = gen_disasm_create(devinfo);
  57 }
  58
  59 void
  60 gen_batch_decode_ctx_finish(struct gen_batch_decode_ctx *ctx)
  61 {
  62    gen_spec_destroy(ctx->spec);
  63    gen_disasm_destroy(ctx->disasm);
  64 }
  65
  66 #define CSI "\e["
  67 #define RED_COLOR    CSI "31m"
  68 #define BLUE_HEADER  CSI "0;44m"
  69 #define GREEN_HEADER CSI "1;42m"
  70 #define NORMAL       CSI "0m"
  71
  72 static void
  73 ctx_print_group(struct gen_batch_decode_ctx *ctx,
  74                 struct gen_group *group,
  75                 uint64_t address, const void *map)
  76 {
  77    gen_print_group(ctx->fp, group, address, map, 0,
  78                    (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) != 0);
  79 }
  80
  81 static struct gen_batch_decode_bo
  82 ctx_get_bo(struct gen_batch_decode_ctx *ctx, bool ppgtt, uint64_t addr)
  83 {
  84    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0)) {
  85       /* On Broadwell and above, we have 48-bit addresses which consume two
  86        * dwords.  Some packets require that these get stored in a "canonical
  87        * form" which means that bit 47 is sign-extended through the upper
  88        * bits. In order to correctly handle those aub dumps, we need to mask
  89        * off the top 16 bits.
  90        */
  91       addr &= (~0ull >> 16);
  92    }
  93
  94    struct gen_batch_decode_bo bo = ctx->get_bo(ctx->user_data, ppgtt, addr);
  95
  96    if (gen_spec_get_gen(ctx->spec) >= gen_make_gen(8,0))
  97       bo.addr &= (~0ull >> 16);
  98
  99    /* We may actually have an offset into the bo */
 100    if (bo.map != NULL) {
 101       assert(bo.addr <= addr);
 102       uint64_t offset = addr - bo.addr;
 103       bo.map += offset;
 104       bo.addr += offset;
 105       bo.size -= offset;
 106    }
 107
 108    return bo;
 109 }
 110
 111 static int
 112 update_count(struct gen_batch_decode_ctx *ctx,
 113              uint32_t offset_from_dsba,
 114              unsigned element_dwords,
 115              unsigned guess)
 116 {
 117    unsigned size = 0;
 118
 119    if (ctx->get_state_size)
 120       size = ctx->get_state_size(ctx->user_data, offset_from_dsba);
 121
 122    if (size > 0)
 123       return size / (sizeof(uint32_t) * element_dwords);
 124
 125    /* In the absence of any information, just guess arbitrarily. */
 126    return guess;
 127 }
 128
 129 static void
 130 ctx_disassemble_program(struct gen_batch_decode_ctx *ctx,
 131                         uint32_t ksp, const char *type)
 132 {
 133    uint64_t addr = ctx->instruction_base + ksp;
 134    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, true, addr);
 135    if (!bo.map)
 136       return;
 137
 138    fprintf(ctx->fp, "\nReferenced %s:\n", type);
 139    gen_disasm_disassemble(ctx->disasm, bo.map, 0, ctx->fp);
 140 }
 141
 142 /* Heuristic to determine whether a uint32_t is probably actually a float
 143  * (http://stackoverflow.com/a/2953466)
 144  */
 145
 146 static bool
 147 probably_float(uint32_t bits)
 148 {
 149    int exp = ((bits & 0x7f800000U) >> 23) - 127;
 150    uint32_t mant = bits & 0x007fffff;
 151
 152    /* +- 0.0 */
 153    if (exp == -127 && mant == 0)
 154       return true;
 155
 156    /* +- 1 billionth to 1 billion */
 157    if (-30 <= exp && exp <= 30)
 158       return true;
 159
 160    /* some value with only a few binary digits */
 161    if ((mant & 0x0000ffff) == 0)
 162       return true;
 163
 164    return false;
 165 }
 166
 167 static void
 168 ctx_print_buffer(struct gen_batch_decode_ctx *ctx,
 169                  struct gen_batch_decode_bo bo,
 170                  uint32_t read_length,
 171                  uint32_t pitch,
 172                  int max_lines)
 173 {
 174    const uint32_t *dw_end =
 175          bo.map + ROUND_DOWN_TO(MIN2(bo.size, read_length), 4);
 176
 177    int column_count = 0, line_count = -1;
 178    for (const uint32_t *dw = bo.map; dw < dw_end; dw++) {
 179       if (column_count * 4 == pitch || column_count == 8) {
 180          fprintf(ctx->fp, "\n");
 181          column_count = 0;
 182          line_count++;
 183
 184          if (max_lines >= 0 && line_count >= max_lines)
 185             break;
 186       }
 187       fprintf(ctx->fp, column_count == 0 ? "  " : " ");
 188
 189       if ((ctx->flags & GEN_BATCH_DECODE_FLOATS) && probably_float(*dw))
 190          fprintf(ctx->fp, "  %8.2f", *(float *) dw);
 191       else
 192          fprintf(ctx->fp, "  0x%08x", *dw);
 193
 194       column_count++;
 195    }
 196    fprintf(ctx->fp, "\n");
 197 }
 198
 199 static struct gen_group *
 200 gen_ctx_find_instruction(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 201 {
 202    return gen_spec_find_instruction(ctx->spec, ctx->engine, p);
 203 }
 204
 205 static void
 206 handle_state_base_address(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 207 {
 208    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 209
 210    struct gen_field_iterator iter;
 211    gen_field_iterator_init(&iter, inst, p, 0, false);
 212
 213    uint64_t surface_base = 0, dynamic_base = 0, instruction_base = 0;
 214    bool surface_modify = 0, dynamic_modify = 0, instruction_modify = 0;
 215
 216    while (gen_field_iterator_next(&iter)) {
 217       if (strcmp(iter.name, "Surface State Base Address") == 0) {
 218          surface_base = iter.raw_value;
 219       } else if (strcmp(iter.name, "Dynamic State Base Address") == 0) {
 220          dynamic_base = iter.raw_value;
 221       } else if (strcmp(iter.name, "Instruction Base Address") == 0) {
 222          instruction_base = iter.raw_value;
 223       } else if (strcmp(iter.name, "Surface State Base Address Modify Enable") == 0) {
 224          surface_modify = iter.raw_value;
 225       } else if (strcmp(iter.name, "Dynamic State Base Address Modify Enable") == 0) {
 226          dynamic_modify = iter.raw_value;
 227       } else if (strcmp(iter.name, "Instruction Base Address Modify Enable") == 0) {
 228          instruction_modify = iter.raw_value;
 229       }
 230    }
 231
 232    if (dynamic_modify)
 233       ctx->dynamic_base = dynamic_base;
 234
 235    if (surface_modify)
 236       ctx->surface_base = surface_base;
 237
 238    if (instruction_modify)
 239       ctx->instruction_base = instruction_base;
 240 }
 241
 242 static void
 243 dump_binding_table(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 244 {
 245    struct gen_group *strct =
 246       gen_spec_find_struct(ctx->spec, "RENDER_SURFACE_STATE");
 247    if (strct == NULL) {
 248       fprintf(ctx->fp, "did not find RENDER_SURFACE_STATE info\n");
 249       return;
 250    }
 251
 252    if (count < 0)
 253       count = update_count(ctx, offset, 1, 8);
 254
 255    if (offset % 32 != 0 || offset >= UINT16_MAX) {
 256       fprintf(ctx->fp, "  invalid binding table pointer\n");
 257       return;
 258    }
 259
 260    struct gen_batch_decode_bo bind_bo =
 261       ctx_get_bo(ctx, true, ctx->surface_base + offset);
 262
 263    if (bind_bo.map == NULL) {
 264       fprintf(ctx->fp, "  binding table unavailable\n");
 265       return;
 266    }
 267
 268    const uint32_t *pointers = bind_bo.map;
 269    for (int i = 0; i < count; i++) {
 270       if (pointers[i] == 0)
 271          continue;
 272
 273       uint64_t addr = ctx->surface_base + pointers[i];
 274       struct gen_batch_decode_bo bo = ctx_get_bo(ctx, true, addr);
 275       uint32_t size = strct->dw_length * 4;
 276
 277       if (pointers[i] % 32 != 0 ||
 278           addr < bo.addr || addr + size >= bo.addr + bo.size) {
 279          fprintf(ctx->fp, "pointer %u: 0x%08x <not valid>\n", i, pointers[i]);
 280          continue;
 281       }
 282
 283       fprintf(ctx->fp, "pointer %u: 0x%08x\n", i, pointers[i]);
 284       ctx_print_group(ctx, strct, addr, bo.map + (addr - bo.addr));
 285    }
 286 }
 287
 288 static void
 289 dump_samplers(struct gen_batch_decode_ctx *ctx, uint32_t offset, int count)
 290 {
 291    struct gen_group *strct = gen_spec_find_struct(ctx->spec, "SAMPLER_STATE");
 292
 293    if (count < 0)
 294       count = update_count(ctx, offset, strct->dw_length, 4);
 295
 296    uint64_t state_addr = ctx->dynamic_base + offset;
 297    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, true, state_addr);
 298    const void *state_map = bo.map;
 299
 300    if (state_map == NULL) {
 301       fprintf(ctx->fp, "  samplers unavailable\n");
 302       return;
 303    }
 304
 305    if (offset % 32 != 0 || state_addr - bo.addr >= bo.size) {
 306       fprintf(ctx->fp, "  invalid sampler state pointer\n");
 307       return;
 308    }
 309
 310    for (int i = 0; i < count; i++) {
 311       fprintf(ctx->fp, "sampler state %d\n", i);
 312       ctx_print_group(ctx, strct, state_addr, state_map);
 313       state_addr += 16;
 314       state_map += 16;
 315    }
 316 }
 317
 318 static void
 319 handle_media_interface_descriptor_load(struct gen_batch_decode_ctx *ctx,
 320                                        const uint32_t *p)
 321 {
 322    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 323    struct gen_group *desc =
 324       gen_spec_find_struct(ctx->spec, "INTERFACE_DESCRIPTOR_DATA");
 325
 326    struct gen_field_iterator iter;
 327    gen_field_iterator_init(&iter, inst, p, 0, false);
 328    uint32_t descriptor_offset = 0;
 329    int descriptor_count = 0;
 330    while (gen_field_iterator_next(&iter)) {
 331       if (strcmp(iter.name, "Interface Descriptor Data Start Address") == 0) {
 332          descriptor_offset = strtol(iter.value, NULL, 16);
 333       } else if (strcmp(iter.name, "Interface Descriptor Total Length") == 0) {
 334          descriptor_count =
 335             strtol(iter.value, NULL, 16) / (desc->dw_length * 4);
 336       }
 337    }
 338
 339    uint64_t desc_addr = ctx->dynamic_base + descriptor_offset;
 340    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, true, desc_addr);
 341    const void *desc_map = bo.map;
 342
 343    if (desc_map == NULL) {
 344       fprintf(ctx->fp, "  interface descriptors unavailable\n");
 345       return;
 346    }
 347
 348    for (int i = 0; i < descriptor_count; i++) {
 349       fprintf(ctx->fp, "descriptor %d: %08x\n", i, descriptor_offset);
 350
 351       ctx_print_group(ctx, desc, desc_addr, desc_map);
 352
 353       gen_field_iterator_init(&iter, desc, desc_map, 0, false);
 354       uint64_t ksp = 0;
 355       uint32_t sampler_offset = 0, sampler_count = 0;
 356       uint32_t binding_table_offset = 0, binding_entry_count = 0;
 357       while (gen_field_iterator_next(&iter)) {
 358          if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 359             ksp = strtoll(iter.value, NULL, 16);
 360          } else if (strcmp(iter.name, "Sampler State Pointer") == 0) {
 361             sampler_offset = strtol(iter.value, NULL, 16);
 362          } else if (strcmp(iter.name, "Sampler Count") == 0) {
 363             sampler_count = strtol(iter.value, NULL, 10);
 364          } else if (strcmp(iter.name, "Binding Table Pointer") == 0) {
 365             binding_table_offset = strtol(iter.value, NULL, 16);
 366          } else if (strcmp(iter.name, "Binding Table Entry Count") == 0) {
 367             binding_entry_count = strtol(iter.value, NULL, 10);
 368          }
 369       }
 370
 371       ctx_disassemble_program(ctx, ksp, "compute shader");
 372       printf("\n");
 373
 374       dump_samplers(ctx, sampler_offset, sampler_count);
 375       dump_binding_table(ctx, binding_table_offset, binding_entry_count);
 376
 377       desc_map += desc->dw_length;
 378       desc_addr += desc->dw_length * 4;
 379    }
 380 }
 381
 382 static void
 383 handle_3dstate_vertex_buffers(struct gen_batch_decode_ctx *ctx,
 384                               const uint32_t *p)
 385 {
 386    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 387    struct gen_group *vbs = gen_spec_find_struct(ctx->spec, "VERTEX_BUFFER_STATE");
 388
 389    struct gen_batch_decode_bo vb = {};
 390    uint32_t vb_size = 0;
 391    int index = -1;
 392    int pitch = -1;
 393    bool ready = false;
 394
 395    struct gen_field_iterator iter;
 396    gen_field_iterator_init(&iter, inst, p, 0, false);
 397    while (gen_field_iterator_next(&iter)) {
 398       if (iter.struct_desc != vbs)
 399          continue;
 400
 401       struct gen_field_iterator vbs_iter;
 402       gen_field_iterator_init(&vbs_iter, vbs, &iter.p[iter.start_bit / 32], 0, false);
 403       while (gen_field_iterator_next(&vbs_iter)) {
 404          if (strcmp(vbs_iter.name, "Vertex Buffer Index") == 0) {
 405             index = vbs_iter.raw_value;
 406          } else if (strcmp(vbs_iter.name, "Buffer Pitch") == 0) {
 407             pitch = vbs_iter.raw_value;
 408          } else if (strcmp(vbs_iter.name, "Buffer Starting Address") == 0) {
 409             vb = ctx_get_bo(ctx, true, vbs_iter.raw_value);
 410          } else if (strcmp(vbs_iter.name, "Buffer Size") == 0) {
 411             vb_size = vbs_iter.raw_value;
 412             ready = true;
 413          } else if (strcmp(vbs_iter.name, "End Address") == 0) {
 414             if (vb.map && vbs_iter.raw_value >= vb.addr)
 415                vb_size = (vbs_iter.raw_value + 1) - vb.addr;
 416             else
 417                vb_size = 0;
 418             ready = true;
 419          }
 420
 421          if (!ready)
 422             continue;
 423
 424          fprintf(ctx->fp, "vertex buffer %d, size %d\n", index, vb_size);
 425
 426          if (vb.map == NULL) {
 427             fprintf(ctx->fp, "  buffer contents unavailable\n");
 428             continue;
 429          }
 430
 431          if (vb.map == 0 || vb_size == 0)
 432             continue;
 433
 434          ctx_print_buffer(ctx, vb, vb_size, pitch, ctx->max_vbo_decoded_lines);
 435
 436          vb.map = NULL;
 437          vb_size = 0;
 438          index = -1;
 439          pitch = -1;
 440          ready = false;
 441       }
 442    }
 443 }
 444
 445 static void
 446 handle_3dstate_index_buffer(struct gen_batch_decode_ctx *ctx,
 447                             const uint32_t *p)
 448 {
 449    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 450
 451    struct gen_batch_decode_bo ib = {};
 452    uint32_t ib_size = 0;
 453    uint32_t format = 0;
 454
 455    struct gen_field_iterator iter;
 456    gen_field_iterator_init(&iter, inst, p, 0, false);
 457    while (gen_field_iterator_next(&iter)) {
 458       if (strcmp(iter.name, "Index Format") == 0) {
 459          format = iter.raw_value;
 460       } else if (strcmp(iter.name, "Buffer Starting Address") == 0) {
 461          ib = ctx_get_bo(ctx, true, iter.raw_value);
 462       } else if (strcmp(iter.name, "Buffer Size") == 0) {
 463          ib_size = iter.raw_value;
 464       }
 465    }
 466
 467    if (ib.map == NULL) {
 468       fprintf(ctx->fp, "  buffer contents unavailable\n");
 469       return;
 470    }
 471
 472    const void *m = ib.map;
 473    const void *ib_end = ib.map + MIN2(ib.size, ib_size);
 474    for (int i = 0; m < ib_end && i < 10; i++) {
 475       switch (format) {
 476       case 0:
 477          fprintf(ctx->fp, "%3d ", *(uint8_t *)m);
 478          m += 1;
 479          break;
 480       case 1:
 481          fprintf(ctx->fp, "%3d ", *(uint16_t *)m);
 482          m += 2;
 483          break;
 484       case 2:
 485          fprintf(ctx->fp, "%3d ", *(uint32_t *)m);
 486          m += 4;
 487          break;
 488       }
 489    }
 490
 491    if (m < ib_end)
 492       fprintf(ctx->fp, "...");
 493    fprintf(ctx->fp, "\n");
 494 }
 495
 496 static void
 497 decode_single_ksp(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 498 {
 499    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 500
 501    uint64_t ksp = 0;
 502    bool is_simd8 = false; /* vertex shaders on Gen8+ only */
 503    bool is_enabled = true;
 504
 505    struct gen_field_iterator iter;
 506    gen_field_iterator_init(&iter, inst, p, 0, false);
 507    while (gen_field_iterator_next(&iter)) {
 508       if (strcmp(iter.name, "Kernel Start Pointer") == 0) {
 509          ksp = iter.raw_value;
 510       } else if (strcmp(iter.name, "SIMD8 Dispatch Enable") == 0) {
 511          is_simd8 = iter.raw_value;
 512       } else if (strcmp(iter.name, "Dispatch Mode") == 0) {
 513          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 514       } else if (strcmp(iter.name, "Dispatch Enable") == 0) {
 515          is_simd8 = strcmp(iter.value, "SIMD8") == 0;
 516       } else if (strcmp(iter.name, "Enable") == 0) {
 517          is_enabled = iter.raw_value;
 518       }
 519    }
 520
 521    const char *type =
 522       strcmp(inst->name,   "VS_STATE") == 0 ? "vertex shader" :
 523       strcmp(inst->name,   "GS_STATE") == 0 ? "geometry shader" :
 524       strcmp(inst->name,   "SF_STATE") == 0 ? "strips and fans shader" :
 525       strcmp(inst->name, "CLIP_STATE") == 0 ? "clip shader" :
 526       strcmp(inst->name, "3DSTATE_DS") == 0 ? "tessellation evaluation shader" :
 527       strcmp(inst->name, "3DSTATE_HS") == 0 ? "tessellation control shader" :
 528       strcmp(inst->name, "3DSTATE_VS") == 0 ? (is_simd8 ? "SIMD8 vertex shader" : "vec4 vertex shader") :
 529       strcmp(inst->name, "3DSTATE_GS") == 0 ? (is_simd8 ? "SIMD8 geometry shader" : "vec4 geometry shader") :
 530       NULL;
 531
 532    if (is_enabled) {
 533       ctx_disassemble_program(ctx, ksp, type);
 534       printf("\n");
 535    }
 536 }
 537
 538 static void
 539 decode_ps_kernels(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 540 {
 541    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 542
 543    uint64_t ksp[3] = {0, 0, 0};
 544    bool enabled[3] = {false, false, false};
 545
 546    struct gen_field_iterator iter;
 547    gen_field_iterator_init(&iter, inst, p, 0, false);
 548    while (gen_field_iterator_next(&iter)) {
 549       if (strncmp(iter.name, "Kernel Start Pointer ",
 550                   strlen("Kernel Start Pointer ")) == 0) {
 551          int idx = iter.name[strlen("Kernel Start Pointer ")] - '0';
 552          ksp[idx] = strtol(iter.value, NULL, 16);
 553       } else if (strcmp(iter.name, "8 Pixel Dispatch Enable") == 0) {
 554          enabled[0] = strcmp(iter.value, "true") == 0;
 555       } else if (strcmp(iter.name, "16 Pixel Dispatch Enable") == 0) {
 556          enabled[1] = strcmp(iter.value, "true") == 0;
 557       } else if (strcmp(iter.name, "32 Pixel Dispatch Enable") == 0) {
 558          enabled[2] = strcmp(iter.value, "true") == 0;
 559       }
 560    }
 561
 562    /* Reorder KSPs to be [8, 16, 32] instead of the hardware order. */
 563    if (enabled[0] + enabled[1] + enabled[2] == 1) {
 564       if (enabled[1]) {
 565          ksp[1] = ksp[0];
 566          ksp[0] = 0;
 567       } else if (enabled[2]) {
 568          ksp[2] = ksp[0];
 569          ksp[0] = 0;
 570       }
 571    } else {
 572       uint64_t tmp = ksp[1];
 573       ksp[1] = ksp[2];
 574       ksp[2] = tmp;
 575    }
 576
 577    if (enabled[0])
 578       ctx_disassemble_program(ctx, ksp[0], "SIMD8 fragment shader");
 579    if (enabled[1])
 580       ctx_disassemble_program(ctx, ksp[1], "SIMD16 fragment shader");
 581    if (enabled[2])
 582       ctx_disassemble_program(ctx, ksp[2], "SIMD32 fragment shader");
 583    fprintf(ctx->fp, "\n");
 584 }
 585
 586 static void
 587 decode_3dstate_constant(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 588 {
 589    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 590    struct gen_group *body =
 591       gen_spec_find_struct(ctx->spec, "3DSTATE_CONSTANT_BODY");
 592
 593    uint32_t read_length[4] = {0};
 594    uint64_t read_addr[4];
 595
 596    struct gen_field_iterator outer;
 597    gen_field_iterator_init(&outer, inst, p, 0, false);
 598    while (gen_field_iterator_next(&outer)) {
 599       if (outer.struct_desc != body)
 600          continue;
 601
 602       struct gen_field_iterator iter;
 603       gen_field_iterator_init(&iter, body, &outer.p[outer.start_bit / 32],
 604                               0, false);
 605
 606       while (gen_field_iterator_next(&iter)) {
 607          int idx;
 608          if (sscanf(iter.name, "Read Length[%d]", &idx) == 1) {
 609             read_length[idx] = iter.raw_value;
 610          } else if (sscanf(iter.name, "Buffer[%d]", &idx) == 1) {
 611             read_addr[idx] = iter.raw_value;
 612          }
 613       }
 614
 615       for (int i = 0; i < 4; i++) {
 616          if (read_length[i] == 0)
 617             continue;
 618
 619          struct gen_batch_decode_bo buffer = ctx_get_bo(ctx, true, read_addr[i]);
 620          if (!buffer.map) {
 621             fprintf(ctx->fp, "constant buffer %d unavailable\n", i);
 622             continue;
 623          }
 624
 625          unsigned size = read_length[i] * 32;
 626          fprintf(ctx->fp, "constant buffer %d, size %u\n", i, size);
 627
 628          ctx_print_buffer(ctx, buffer, size, 0, -1);
 629       }
 630    }
 631 }
 632
 633 static void
 634 decode_3dstate_binding_table_pointers(struct gen_batch_decode_ctx *ctx,
 635                                       const uint32_t *p)
 636 {
 637    dump_binding_table(ctx, p[1], -1);
 638 }
 639
 640 static void
 641 decode_3dstate_sampler_state_pointers(struct gen_batch_decode_ctx *ctx,
 642                                       const uint32_t *p)
 643 {
 644    dump_samplers(ctx, p[1], -1);
 645 }
 646
 647 static void
 648 decode_3dstate_sampler_state_pointers_gen6(struct gen_batch_decode_ctx *ctx,
 649                                            const uint32_t *p)
 650 {
 651    dump_samplers(ctx, p[1], -1);
 652    dump_samplers(ctx, p[2], -1);
 653    dump_samplers(ctx, p[3], -1);
 654 }
 655
 656 static bool
 657 str_ends_with(const char *str, const char *end)
 658 {
 659    int offset = strlen(str) - strlen(end);
 660    if (offset < 0)
 661       return false;
 662
 663    return strcmp(str + offset, end) == 0;
 664 }
 665
 666 static void
 667 decode_dynamic_state_pointers(struct gen_batch_decode_ctx *ctx,
 668                               const char *struct_type, const uint32_t *p,
 669                               int count)
 670 {
 671    struct gen_group *inst = gen_ctx_find_instruction(ctx, p);
 672
 673    uint32_t state_offset = 0;
 674
 675    struct gen_field_iterator iter;
 676    gen_field_iterator_init(&iter, inst, p, 0, false);
 677    while (gen_field_iterator_next(&iter)) {
 678       if (str_ends_with(iter.name, "Pointer")) {
 679          state_offset = iter.raw_value;
 680          break;
 681       }
 682    }
 683
 684    uint64_t state_addr = ctx->dynamic_base + state_offset;
 685    struct gen_batch_decode_bo bo = ctx_get_bo(ctx, true, state_addr);
 686    const void *state_map = bo.map;
 687
 688    if (state_map == NULL) {
 689       fprintf(ctx->fp, "  dynamic %s state unavailable\n", struct_type);
 690       return;
 691    }
 692
 693    struct gen_group *state = gen_spec_find_struct(ctx->spec, struct_type);
 694    if (strcmp(struct_type, "BLEND_STATE") == 0) {
 695       /* Blend states are different from the others because they have a header
 696        * struct called BLEND_STATE which is followed by a variable number of
 697        * BLEND_STATE_ENTRY structs.
 698        */
 699       fprintf(ctx->fp, "%s\n", struct_type);
 700       ctx_print_group(ctx, state, state_addr, state_map);
 701
 702       state_addr += state->dw_length * 4;
 703       state_map += state->dw_length * 4;
 704
 705       struct_type = "BLEND_STATE_ENTRY";
 706       state = gen_spec_find_struct(ctx->spec, struct_type);
 707    }
 708
 709    for (int i = 0; i < count; i++) {
 710       fprintf(ctx->fp, "%s %d\n", struct_type, i);
 711       ctx_print_group(ctx, state, state_addr, state_map);
 712
 713       state_addr += state->dw_length * 4;
 714       state_map += state->dw_length * 4;
 715    }
 716 }
 717
 718 static void
 719 decode_3dstate_viewport_state_pointers_cc(struct gen_batch_decode_ctx *ctx,
 720                                           const uint32_t *p)
 721 {
 722    decode_dynamic_state_pointers(ctx, "CC_VIEWPORT", p, 4);
 723 }
 724
 725 static void
 726 decode_3dstate_viewport_state_pointers_sf_clip(struct gen_batch_decode_ctx *ctx,
 727                                                const uint32_t *p)
 728 {
 729    decode_dynamic_state_pointers(ctx, "SF_CLIP_VIEWPORT", p, 4);
 730 }
 731
 732 static void
 733 decode_3dstate_blend_state_pointers(struct gen_batch_decode_ctx *ctx,
 734                                     const uint32_t *p)
 735 {
 736    decode_dynamic_state_pointers(ctx, "BLEND_STATE", p, 1);
 737 }
 738
 739 static void
 740 decode_3dstate_cc_state_pointers(struct gen_batch_decode_ctx *ctx,
 741                                  const uint32_t *p)
 742 {
 743    decode_dynamic_state_pointers(ctx, "COLOR_CALC_STATE", p, 1);
 744 }
 745
 746 static void
 747 decode_3dstate_scissor_state_pointers(struct gen_batch_decode_ctx *ctx,
 748                                       const uint32_t *p)
 749 {
 750    decode_dynamic_state_pointers(ctx, "SCISSOR_RECT", p, 1);
 751 }
 752
 753 static void
 754 decode_load_register_imm(struct gen_batch_decode_ctx *ctx, const uint32_t *p)
 755 {
 756    struct gen_group *reg = gen_spec_find_register(ctx->spec, p[1]);
 757
 758    if (reg != NULL) {
 759       fprintf(ctx->fp, "register %s (0x%x): 0x%x\n",
 760               reg->name, reg->register_offset, p[2]);
 761       ctx_print_group(ctx, reg, reg->register_offset, &p[2]);
 762    }
 763 }
 764
 765 struct custom_decoder {
 766    const char *cmd_name;
 767    void (*decode)(struct gen_batch_decode_ctx *ctx, const uint32_t *p);
 768 } custom_decoders[] = {
 769    { "STATE_BASE_ADDRESS", handle_state_base_address },
 770    { "MEDIA_INTERFACE_DESCRIPTOR_LOAD", handle_media_interface_descriptor_load },
 771    { "3DSTATE_VERTEX_BUFFERS", handle_3dstate_vertex_buffers },
 772    { "3DSTATE_INDEX_BUFFER", handle_3dstate_index_buffer },
 773    { "3DSTATE_VS", decode_single_ksp },
 774    { "3DSTATE_GS", decode_single_ksp },
 775    { "3DSTATE_DS", decode_single_ksp },
 776    { "3DSTATE_HS", decode_single_ksp },
 777    { "3DSTATE_PS", decode_ps_kernels },
 778    { "3DSTATE_CONSTANT_VS", decode_3dstate_constant },
 779    { "3DSTATE_CONSTANT_GS", decode_3dstate_constant },
 780    { "3DSTATE_CONSTANT_PS", decode_3dstate_constant },
 781    { "3DSTATE_CONSTANT_HS", decode_3dstate_constant },
 782    { "3DSTATE_CONSTANT_DS", decode_3dstate_constant },
 783
 784    { "3DSTATE_BINDING_TABLE_POINTERS_VS", decode_3dstate_binding_table_pointers },
 785    { "3DSTATE_BINDING_TABLE_POINTERS_HS", decode_3dstate_binding_table_pointers },
 786    { "3DSTATE_BINDING_TABLE_POINTERS_DS", decode_3dstate_binding_table_pointers },
 787    { "3DSTATE_BINDING_TABLE_POINTERS_GS", decode_3dstate_binding_table_pointers },
 788    { "3DSTATE_BINDING_TABLE_POINTERS_PS", decode_3dstate_binding_table_pointers },
 789
 790    { "3DSTATE_SAMPLER_STATE_POINTERS_VS", decode_3dstate_sampler_state_pointers },
 791    { "3DSTATE_SAMPLER_STATE_POINTERS_HS", decode_3dstate_sampler_state_pointers },
 792    { "3DSTATE_SAMPLER_STATE_POINTERS_DS", decode_3dstate_sampler_state_pointers },
 793    { "3DSTATE_SAMPLER_STATE_POINTERS_GS", decode_3dstate_sampler_state_pointers },
 794    { "3DSTATE_SAMPLER_STATE_POINTERS_PS", decode_3dstate_sampler_state_pointers },
 795    { "3DSTATE_SAMPLER_STATE_POINTERS", decode_3dstate_sampler_state_pointers_gen6 },
 796
 797    { "3DSTATE_VIEWPORT_STATE_POINTERS_CC", decode_3dstate_viewport_state_pointers_cc },
 798    { "3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP", decode_3dstate_viewport_state_pointers_sf_clip },
 799    { "3DSTATE_BLEND_STATE_POINTERS", decode_3dstate_blend_state_pointers },
 800    { "3DSTATE_CC_STATE_POINTERS", decode_3dstate_cc_state_pointers },
 801    { "3DSTATE_SCISSOR_STATE_POINTERS", decode_3dstate_scissor_state_pointers },
 802    { "MI_LOAD_REGISTER_IMM", decode_load_register_imm }
 803 };
 804
 805 void
 806 gen_print_batch(struct gen_batch_decode_ctx *ctx,
 807                 const uint32_t *batch, uint32_t batch_size,
 808                 uint64_t batch_addr)
 809 {
 810    const uint32_t *p, *end = batch + batch_size / sizeof(uint32_t);
 811    int length;
 812    struct gen_group *inst;
 813
 814    for (p = batch; p < end; p += length) {
 815       inst = gen_ctx_find_instruction(ctx, p);
 816       length = gen_group_get_length(inst, p);
 817       assert(inst == NULL || length > 0);
 818       length = MAX2(1, length);
 819
 820       const char *reset_color = ctx->flags & GEN_BATCH_DECODE_IN_COLOR ? NORMAL : "";
 821
 822       uint64_t offset;
 823       if (ctx->flags & GEN_BATCH_DECODE_OFFSETS)
 824          offset = batch_addr + ((char *)p - (char *)batch);
 825       else
 826          offset = 0;
 827
 828       if (inst == NULL) {
 829          fprintf(ctx->fp, "%s0x%08"PRIx64": unknown instruction %08x%s\n",
 830                  (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) ? RED_COLOR : "",
 831                  offset, p[0], reset_color);
 832          continue;
 833       }
 834
 835       const char *color;
 836       const char *inst_name = gen_group_get_name(inst);
 837       if (ctx->flags & GEN_BATCH_DECODE_IN_COLOR) {
 838          reset_color = NORMAL;
 839          if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 840             if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0 ||
 841                 strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0)
 842                color = GREEN_HEADER;
 843             else
 844                color = BLUE_HEADER;
 845          } else {
 846             color = NORMAL;
 847          }
 848       } else {
 849          color = "";
 850          reset_color = "";
 851       }
 852
 853       fprintf(ctx->fp, "%s0x%08"PRIx64":  0x%08x:  %-80s%s\n",
 854               color, offset, p[0], inst_name, reset_color);
 855
 856       if (ctx->flags & GEN_BATCH_DECODE_FULL) {
 857          ctx_print_group(ctx, inst, offset, p);
 858
 859          for (int i = 0; i < ARRAY_SIZE(custom_decoders); i++) {
 860             if (strcmp(inst_name, custom_decoders[i].cmd_name) == 0) {
 861                custom_decoders[i].decode(ctx, p);
 862                break;
 863             }
 864          }
 865       }
 866
 867       if (strcmp(inst_name, "MI_BATCH_BUFFER_START") == 0) {
 868          uint64_t next_batch_addr;
 869          bool ppgtt = false;
 870          bool second_level;
 871          struct gen_field_iterator iter;
 872          gen_field_iterator_init(&iter, inst, p, 0, false);
 873          while (gen_field_iterator_next(&iter)) {
 874             if (strcmp(iter.name, "Batch Buffer Start Address") == 0) {
 875                next_batch_addr = iter.raw_value;
 876             } else if (strcmp(iter.name, "Second Level Batch Buffer") == 0) {
 877                second_level = iter.raw_value;
 878             } else if (strcmp(iter.name, "Address Space Indicator") == 0) {
 879                ppgtt = iter.raw_value;
 880             }
 881          }
 882
 883          struct gen_batch_decode_bo next_batch = ctx_get_bo(ctx, ppgtt, next_batch_addr);
 884
 885          if (next_batch.map == NULL) {
 886             fprintf(ctx->fp, "Secondary batch at 0x%08"PRIx64" unavailable\n",
 887                     next_batch_addr);
 888          } else {
 889             gen_print_batch(ctx, next_batch.map, next_batch.size,
 890                             next_batch.addr);
 891          }
 892          if (second_level) {
 893             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" set acts
 894              * like a subroutine call.  Commands that come afterwards get
 895              * processed once the 2nd level batch buffer returns with
 896              * MI_BATCH_BUFFER_END.
 897              */
 898             continue;
 899          } else {
 900             /* MI_BATCH_BUFFER_START with "2nd Level Batch Buffer" unset acts
 901              * like a goto.  Nothing after it will ever get processed.  In
 902              * order to prevent the recursion from growing, we just reset the
 903              * loop and continue;
 904              */
 905             break;
 906          }
 907       } else if (strcmp(inst_name, "MI_BATCH_BUFFER_END") == 0) {
 908          break;
 909       }
 910    }
 911 }