src/panfrost/bifrost/bi_pack.c

   1 /*
   2  * Copyright (C) 2020 Collabora, Ltd.
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  */
  23
  24 #include "compiler.h"
  25
  26 #define RETURN_PACKED(str) { \
  27         uint64_t temp = 0; \
  28         memcpy(&temp, &str, sizeof(str)); \
  29         return temp; \
  30 }
  31
  32 /* This file contains the final passes of the compiler. Running after
  33  * scheduling and RA, the IR is now finalized, so we need to emit it to actual
  34  * bits on the wire (as well as fixup branches) */
  35
  36 static uint64_t
  37 bi_pack_header(bi_clause *clause, bi_clause *next, bool is_fragment)
  38 {
  39         struct bifrost_header header = {
  40                 .back_to_back = clause->back_to_back,
  41                 .no_end_of_shader = (next != NULL),
  42                 .elide_writes = is_fragment,
  43                 .branch_cond = clause->branch_conditional,
  44                 .datareg_writebarrier = clause->data_register_write_barrier,
  45                 .datareg = clause->data_register,
  46                 .scoreboard_deps = clause->dependencies,
  47                 .scoreboard_index = clause->scoreboard_id,
  48                 .clause_type = clause->clause_type,
  49                 .next_clause_type = next ? next->clause_type : 0,
  50         };
  51
  52         uint64_t u = 0;
  53         memcpy(&u, &header, sizeof(header));
  54         return u;
  55 }
  56
  57 /* Represents the assignment of ports for a given bundle */
  58
  59 struct bi_registers {
  60         /* Register to assign to each port */
  61         unsigned port[4];
  62
  63         /* Read ports can be disabled */
  64         bool enabled[2];
  65
  66         /* Should we write FMA? what about ADD? If only a single port is
  67          * enabled it is in port 2, else ADD/FMA is 2/3 respectively */
  68         bool write_fma, write_add;
  69
  70         /* Should we read with port 3? */
  71         bool read_port3;
  72
  73         /* Packed uniform/constant */
  74         uint8_t uniform_constant;
  75
  76         /* Whether writes are actually for the last instruction */
  77         bool first_instruction;
  78 };
  79
  80 /* The uniform/constant slot allows loading a contiguous 64-bit immediate or
  81  * pushed uniform per bundle. Figure out which one we need in the bundle (the
  82  * scheduler needs to ensure we only have one type per bundle), validate
  83  * everything, and rewrite away the register/uniform indices to use 3-bit
  84  * sources directly. */
  85
  86 static unsigned
  87 bi_lookup_constant(bi_clause *clause, uint64_t cons)
  88 {
  89         for (unsigned i = 0; i < clause->constant_count; ++i) {
  90                 /* Only check top 60-bits since that's what's actually embedded
  91                  * in the clause, the bottom 4-bits are bundle-inline */
  92
  93                 if ((cons >> 4) == (clause->constants[i] >> 4))
  94                         return i;
  95         }
  96
  97         unreachable("Invalid constant accessed");
  98 }
  99
 100 static unsigned
 101 bi_constant_field(unsigned idx)
 102 {
 103         assert(idx <= 5);
 104
 105         const unsigned values[] = {
 106                 4, 5, 6, 7, 2, 3
 107         };
 108
 109         return values[idx] << 4;
 110 }
 111
 112 static bool
 113 bi_assign_uniform_constant_single(
 114                 struct bi_registers *regs,
 115                 bi_clause *clause,
 116                 bi_instruction *ins, bool assigned, bool fast_zero)
 117 {
 118         if (!ins)
 119                 return assigned;
 120
 121         bi_foreach_src(ins, s) {
 122                 if (s == 0 && (ins->type == BI_LOAD_VAR_ADDRESS || ins->type == BI_LOAD_ATTR)) continue;
 123
 124                 if (ins->src[s] & BIR_INDEX_CONSTANT) {
 125                         /* TODO: lo/hi matching? */
 126                         uint64_t cons = ins->constant.u64;
 127                         unsigned idx = bi_lookup_constant(clause, cons);
 128                         unsigned f = bi_constant_field(idx) | (cons & 0xF);
 129
 130                         if (assigned && regs->uniform_constant != f)
 131                                 unreachable("Mismatched uniform/const field: imm");
 132
 133                         regs->uniform_constant = f;
 134                         ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_LO;
 135                         assigned = true;
 136                 } else if (ins->src[s] & BIR_INDEX_ZERO && (ins->type == BI_LOAD_UNIFORM || ins->type == BI_LOAD_VAR)) {
 137                         /* XXX: HACK UNTIL WE HAVE HI MATCHING DUE TO OVERFLOW XXX */
 138                         ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_HI;
 139                 } else if (ins->src[s] & BIR_INDEX_ZERO && !fast_zero) {
 140                         /* FMAs have a fast zero port, ADD needs to use the
 141                          * uniform/const port's special 0 mode handled here */
 142                         unsigned f = 0;
 143
 144                         if (assigned && regs->uniform_constant != f)
 145                                 unreachable("Mismatched uniform/const field: 0");
 146
 147                         regs->uniform_constant = f;
 148                         ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_LO;
 149                         assigned = true;
 150                 } else if (s & BIR_INDEX_UNIFORM) {
 151                         unreachable("Push uniforms not implemented yet");
 152                 }
 153         }
 154
 155         return assigned;
 156 }
 157
 158 static void
 159 bi_assign_uniform_constant(
 160                 bi_clause *clause,
 161                 struct bi_registers *regs,
 162                 bi_bundle bundle)
 163 {
 164         bool assigned =
 165                 bi_assign_uniform_constant_single(regs, clause, bundle.fma, false, true);
 166
 167         bi_assign_uniform_constant_single(regs, clause, bundle.add, assigned, false);
 168 }
 169
 170 /* Assigns a port for reading, before anything is written */
 171
 172 static void
 173 bi_assign_port_read(struct bi_registers *regs, unsigned src)
 174 {
 175         /* We only assign for registers */
 176         if (!(src & BIR_INDEX_REGISTER))
 177                 return;
 178
 179         unsigned reg = src & ~BIR_INDEX_REGISTER;
 180
 181         /* Check if we already assigned the port */
 182         for (unsigned i = 0; i <= 1; ++i) {
 183                 if (regs->port[i] == reg && regs->enabled[i])
 184                         return;
 185         }
 186
 187         if (regs->port[3] == reg && regs->read_port3)
 188                 return;
 189
 190         /* Assign it now */
 191
 192         for (unsigned i = 0; i <= 1; ++i) {
 193                 if (!regs->enabled[i]) {
 194                         regs->port[i] = reg;
 195                         regs->enabled[i] = true;
 196                         return;
 197                 }
 198         }
 199
 200         if (!regs->read_port3) {
 201                 regs->port[3] = reg;
 202                 regs->read_port3 = true;
 203         }
 204 }
 205
 206 static struct bi_registers
 207 bi_assign_ports(bi_bundle now, bi_bundle prev)
 208 {
 209         struct bi_registers regs = { 0 };
 210
 211         /* We assign ports for the main register mechanism. Special ops
 212          * use the data registers, which has its own mechanism entirely
 213          * and thus gets skipped over here. */
 214
 215         unsigned read_dreg = now.add &&
 216                 bi_class_props[now.add->type] & BI_DATA_REG_SRC;
 217
 218         unsigned write_dreg = prev.add &&
 219                 bi_class_props[prev.add->type] & BI_DATA_REG_DEST;
 220
 221         /* First, assign reads */
 222
 223         if (now.fma)
 224                 bi_foreach_src(now.fma, src)
 225                         bi_assign_port_read(&regs, now.fma->src[src]);
 226
 227         if (now.add) {
 228                 bi_foreach_src(now.add, src) {
 229                         if (!(src == 0 && read_dreg))
 230                                 bi_assign_port_read(&regs, now.add->src[src]);
 231                 }
 232         }
 233
 234         /* Next, assign writes */
 235
 236         if (prev.fma && prev.fma->dest & BIR_INDEX_REGISTER) {
 237                 regs.port[2] = prev.fma->dest & ~BIR_INDEX_REGISTER;
 238                 regs.write_fma = true;
 239         }
 240
 241         if (prev.add && prev.add->dest & BIR_INDEX_REGISTER && !write_dreg) {
 242                 unsigned r = prev.add->dest & ~BIR_INDEX_REGISTER;
 243
 244                 if (regs.write_fma) {
 245                         /* Scheduler constraint: cannot read 3 and write 2 */
 246                         assert(!regs.read_port3);
 247                         regs.port[3] = r;
 248                 } else {
 249                         regs.port[2] = r;
 250                 }
 251
 252                 regs.write_add = true;
 253         }
 254
 255         /* Finally, ensure port 1 > port 0 for the 63-x trick to function */
 256
 257         if (regs.enabled[0] && regs.enabled[1] && regs.port[1] < regs.port[0]) {
 258                 unsigned temp = regs.port[0];
 259                 regs.port[0] = regs.port[1];
 260                 regs.port[1] = temp;
 261         }
 262
 263         return regs;
 264 }
 265
 266 /* Determines the register control field, ignoring the first? flag */
 267
 268 static enum bifrost_reg_control
 269 bi_pack_register_ctrl_lo(struct bi_registers r)
 270 {
 271         if (r.write_fma) {
 272                 if (r.write_add) {
 273                         assert(!r.read_port3);
 274                         return BIFROST_WRITE_ADD_P2_FMA_P3;
 275                 } else {
 276                         if (r.read_port3)
 277                                 return BIFROST_WRITE_FMA_P2_READ_P3;
 278                         else
 279                                 return BIFROST_WRITE_FMA_P2;
 280                 }
 281         } else if (r.write_add) {
 282                 if (r.read_port3)
 283                         return BIFROST_WRITE_ADD_P2_READ_P3;
 284                 else
 285                         return BIFROST_WRITE_ADD_P2;
 286         } else if (r.read_port3)
 287                 return BIFROST_READ_P3;
 288         else
 289                 return BIFROST_REG_NONE;
 290 }
 291
 292 /* Ditto but account for the first? flag this time */
 293
 294 static enum bifrost_reg_control
 295 bi_pack_register_ctrl(struct bi_registers r)
 296 {
 297         enum bifrost_reg_control ctrl = bi_pack_register_ctrl_lo(r);
 298
 299         if (r.first_instruction) {
 300                 if (ctrl == BIFROST_REG_NONE)
 301                         ctrl = BIFROST_FIRST_NONE;
 302                 else
 303                         ctrl |= BIFROST_FIRST_NONE;
 304         }
 305
 306         return ctrl;
 307 }
 308
 309 static uint64_t
 310 bi_pack_registers(struct bi_registers regs)
 311 {
 312         enum bifrost_reg_control ctrl = bi_pack_register_ctrl(regs);
 313         struct bifrost_regs s;
 314         uint64_t packed = 0;
 315
 316         if (regs.enabled[1]) {
 317                 /* Gotta save that bit!~ Required by the 63-x trick */
 318                 assert(regs.port[1] > regs.port[0]);
 319                 assert(regs.enabled[0]);
 320
 321                 /* Do the 63-x trick, see docs/disasm */
 322                 if (regs.port[0] > 31) {
 323                         regs.port[0] = 63 - regs.port[0];
 324                         regs.port[1] = 63 - regs.port[1];
 325                 }
 326
 327                 assert(regs.port[0] <= 31);
 328                 assert(regs.port[1] <= 63);
 329
 330                 s.ctrl = ctrl;
 331                 s.reg1 = regs.port[1];
 332                 s.reg0 = regs.port[0];
 333         } else {
 334                 /* Port 1 disabled, so set to zero and use port 1 for ctrl */
 335                 s.reg1 = ctrl << 2;
 336
 337                 if (regs.enabled[0]) {
 338                         /* Bit 0 upper bit of port 0 */
 339                         s.reg1 |= (regs.port[0] >> 5);
 340
 341                         /* Rest of port 0 in usual spot */
 342                         s.reg0 = (regs.port[0] & 0b11111);
 343                 } else {
 344                         /* Bit 1 set if port 0 also disabled */
 345                         s.reg1 |= (1 << 1);
 346                 }
 347         }
 348
 349         s.reg3 = regs.port[3];
 350         s.reg2 = regs.port[2];
 351         s.uniform_const = regs.uniform_constant;
 352
 353         memcpy(&packed, &s, sizeof(s));
 354         return packed;
 355 }
 356
 357 static void
 358 bi_set_data_register(bi_clause *clause, unsigned idx)
 359 {
 360         assert(idx & BIR_INDEX_REGISTER);
 361         unsigned reg = idx & ~BIR_INDEX_REGISTER;
 362         assert(reg <= 63);
 363         clause->data_register = reg;
 364 }
 365
 366 static void
 367 bi_read_data_register(bi_clause *clause, bi_instruction *ins)
 368 {
 369         bi_set_data_register(clause, ins->src[0]);
 370 }
 371
 372 static void
 373 bi_write_data_register(bi_clause *clause, bi_instruction *ins)
 374 {
 375         bi_set_data_register(clause, ins->dest);
 376 }
 377
 378 static enum bifrost_packed_src
 379 bi_get_src_reg_port(struct bi_registers *regs, unsigned src)
 380 {
 381         unsigned reg = src & ~BIR_INDEX_REGISTER;
 382
 383         if (regs->port[0] == reg && regs->enabled[0])
 384                 return BIFROST_SRC_PORT0;
 385         else if (regs->port[1] == reg && regs->enabled[1])
 386                 return BIFROST_SRC_PORT1;
 387         else if (regs->port[3] == reg && regs->read_port3)
 388                 return BIFROST_SRC_PORT3;
 389         else
 390                 unreachable("Tried to access register with no port");
 391 }
 392
 393 static enum bifrost_packed_src
 394 bi_get_src(bi_instruction *ins, struct bi_registers *regs, unsigned s, bool is_fma)
 395 {
 396         unsigned src = ins->src[s];
 397
 398         if (src & BIR_INDEX_REGISTER)
 399                 return bi_get_src_reg_port(regs, src);
 400         else if (src & BIR_INDEX_ZERO && is_fma)
 401                 return BIFROST_SRC_STAGE;
 402         else if (src & BIR_INDEX_PASS)
 403                 return src & ~BIR_INDEX_PASS;
 404         else
 405                 unreachable("Unknown src");
 406 }
 407
 408 static unsigned
 409 bi_pack_fma_fma(bi_instruction *ins, struct bi_registers *regs)
 410 {
 411         /* (-a)(-b) = ab, so we only need one negate bit */
 412         bool negate_mul = ins->src_neg[0] ^ ins->src_neg[1];
 413
 414         struct bifrost_fma_fma pack = {
 415                 .src0 = bi_get_src(ins, regs, 0, true),
 416                 .src1 = bi_get_src(ins, regs, 1, true),
 417                 .src2 = bi_get_src(ins, regs, 2, true),
 418                 .src0_abs = ins->src_abs[0],
 419                 .src1_abs = ins->src_abs[1],
 420                 .src2_abs = ins->src_abs[2],
 421                 .src0_neg = negate_mul,
 422                 .src2_neg = ins->src_neg[2],
 423                 .op = BIFROST_FMA_OP_FMA
 424         };
 425
 426         RETURN_PACKED(pack);
 427 }
 428
 429 static unsigned
 430 bi_pack_fma_add(bi_instruction *ins, struct bi_registers *regs)
 431 {
 432         /* TODO: fadd16 packing is a bit different */
 433         assert(ins->dest_type == nir_type_float32);
 434
 435         struct bifrost_fma_add pack = {
 436                 .src0 = bi_get_src(ins, regs, 0, true),
 437                 .src1 = bi_get_src(ins, regs, 1, true),
 438                 .src0_abs = ins->src_abs[0],
 439                 .src1_abs = ins->src_abs[1],
 440                 .src0_neg = ins->src_neg[0],
 441                 .src1_neg = ins->src_neg[1],
 442                 .unk = 0x0,
 443                 .outmod = ins->outmod,
 444                 .roundmode = ins->roundmode,
 445                 .op = BIFROST_FMA_OP_FADD32
 446         };
 447
 448         RETURN_PACKED(pack);
 449 }
 450
 451 static unsigned
 452 bi_pack_fma_1src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
 453 {
 454         struct bifrost_fma_inst pack = {
 455                 .src0 = bi_get_src(ins, regs, 0, true),
 456                 .op = op
 457         };
 458
 459         RETURN_PACKED(pack);
 460 }
 461
 462 static unsigned
 463 bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 464 {
 465         if (!bundle.fma)
 466                 return BIFROST_FMA_NOP;
 467
 468         switch (bundle.fma->type) {
 469         case BI_ADD:
 470                 return bi_pack_fma_add(bundle.fma, regs);
 471         case BI_CMP:
 472         case BI_BITWISE:
 473         case BI_CONVERT:
 474         case BI_CSEL:
 475                 return BIFROST_FMA_NOP;
 476         case BI_FMA:
 477                 return bi_pack_fma_fma(bundle.fma, regs);
 478         case BI_FREXP:
 479         case BI_ISUB:
 480         case BI_MINMAX:
 481                 return BIFROST_FMA_NOP;
 482         case BI_MOV:
 483                 return bi_pack_fma_1src(bundle.fma, regs, BIFROST_FMA_OP_MOV);
 484         case BI_FMOV:
 485         case BI_SHIFT:
 486         case BI_SWIZZLE:
 487         case BI_ROUND:
 488                 return BIFROST_FMA_NOP;
 489         default:
 490                 unreachable("Cannot encode class as FMA");
 491         }
 492 }
 493
 494 static unsigned
 495 bi_pack_add_ld_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 496 {
 497         unsigned size = nir_alu_type_get_type_size(ins->dest_type);
 498         assert(size == 32 || size == 16);
 499
 500         unsigned op = (size == 32) ?
 501                 BIFROST_ADD_OP_LD_VAR_32 :
 502                 BIFROST_ADD_OP_LD_VAR_16;
 503
 504         unsigned cmask = bi_from_bytemask(ins->writemask, size / 8);
 505         unsigned channels = util_bitcount(cmask);
 506         assert(cmask == ((1 << channels) - 1));
 507
 508         unsigned packed_addr = 0;
 509
 510         if (ins->src[0] & BIR_INDEX_CONSTANT) {
 511                 /* Direct uses address field directly */
 512                 packed_addr = ins->src[0] & ~BIR_INDEX_CONSTANT;
 513                 assert(packed_addr < 0b1000);
 514         } else {
 515                 /* Indirect gets an extra source */
 516                 packed_addr = bi_get_src(ins, regs, 0, false) | 0b11000;
 517         }
 518
 519         /* The destination is thrown in the data register */
 520         assert(ins->dest & BIR_INDEX_REGISTER);
 521         clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
 522
 523         assert(channels >= 1 && channels <= 4);
 524
 525         struct bifrost_ld_var pack = {
 526                 .src0 = bi_get_src(ins, regs, 1, false),
 527                 .addr = packed_addr,
 528                 .channels = MALI_POSITIVE(channels),
 529                 .interp_mode = ins->load_vary.interp_mode,
 530                 .reuse = ins->load_vary.reuse,
 531                 .flat = ins->load_vary.flat,
 532                 .op = op
 533         };
 534
 535         RETURN_PACKED(pack);
 536 }
 537
 538 static unsigned
 539 bi_pack_add_2src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
 540 {
 541         struct bifrost_add_2src pack = {
 542                 .src0 = bi_get_src(ins, regs, 0, true),
 543                 .src1 = bi_get_src(ins, regs, 1, true),
 544                 .op = op
 545         };
 546
 547         RETURN_PACKED(pack);
 548 }
 549
 550 static unsigned
 551 bi_pack_add_ld_ubo(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 552 {
 553         unsigned components = bi_load32_components(ins);
 554
 555         const unsigned ops[4] = {
 556                 BIFROST_ADD_OP_LD_UBO_1,
 557                 BIFROST_ADD_OP_LD_UBO_2,
 558                 BIFROST_ADD_OP_LD_UBO_3,
 559                 BIFROST_ADD_OP_LD_UBO_4
 560         };
 561
 562         bi_write_data_register(clause, ins);
 563         return bi_pack_add_2src(ins, regs, ops[components - 1]);
 564 }
 565
 566 static enum bifrost_ldst_type
 567 bi_pack_ldst_type(nir_alu_type T)
 568 {
 569         switch (T) {
 570         case nir_type_float16: return BIFROST_LDST_F16;
 571         case nir_type_float32: return BIFROST_LDST_F32;
 572         case nir_type_int32:   return BIFROST_LDST_I32;
 573         case nir_type_uint32:  return BIFROST_LDST_U32;
 574         default: unreachable("Invalid type loaded");
 575         }
 576 }
 577
 578 static unsigned
 579 bi_pack_add_ld_var_addr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 580 {
 581         /* Only direct loads supported */
 582         assert(ins->src[0] == BIR_INDEX_CONSTANT);
 583
 584         struct bifrost_ld_var_addr pack = {
 585                 .src0 = bi_get_src(ins, regs, 1, false),
 586                 .src1 = bi_get_src(ins, regs, 2, false),
 587                 .location = ins->constant.u64,
 588                 .type = bi_pack_ldst_type(ins->src_types[3]),
 589                 .op = BIFROST_ADD_OP_LD_VAR_ADDR
 590         };
 591
 592         bi_write_data_register(clause, ins);
 593         RETURN_PACKED(pack);
 594 }
 595
 596 static unsigned
 597 bi_pack_add_ld_attr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 598 {
 599         /* Only direct loads supported */
 600         assert(ins->src[0] == BIR_INDEX_CONSTANT);
 601
 602         struct bifrost_ld_attr pack = {
 603                 .src0 = bi_get_src(ins, regs, 1, false),
 604                 .src1 = bi_get_src(ins, regs, 2, false),
 605                 .location = ins->constant.u64,
 606                 .channels = MALI_POSITIVE(bi_load32_components(ins)),
 607                 .type = bi_pack_ldst_type(ins->dest_type),
 608                 .op = BIFROST_ADD_OP_LD_ATTR
 609         };
 610
 611         bi_write_data_register(clause, ins);
 612         RETURN_PACKED(pack);
 613 }
 614
 615 static unsigned
 616 bi_pack_add_st_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 617 {
 618         assert(ins->store_channels >= 1 && ins->store_channels <= 4);
 619
 620         struct bifrost_st_vary pack = {
 621                 .src0 = bi_get_src(ins, regs, 1, false),
 622                 .src1 = bi_get_src(ins, regs, 2, false),
 623                 .src2 = bi_get_src(ins, regs, 3, false),
 624                 .channels = MALI_POSITIVE(ins->store_channels),
 625                 .op = BIFROST_ADD_OP_ST_VAR
 626         };
 627
 628         bi_read_data_register(clause, ins);
 629         RETURN_PACKED(pack);
 630 }
 631
 632 static unsigned
 633 bi_pack_add_atest(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 634 {
 635         /* TODO: fp16 */
 636         assert(ins->src_types[1] == nir_type_float32);
 637
 638         struct bifrost_add_atest pack = {
 639                 .src0 = bi_get_src(ins, regs, 0, false),
 640                 .src1 = bi_get_src(ins, regs, 1, false),
 641                 .component = 1, /* Set for fp32 */
 642                 .op = BIFROST_ADD_OP_ATEST,
 643         };
 644
 645         /* Despite *also* writing with the usual mechanism... quirky and
 646          * perhaps unnecessary, but let's match the blob */
 647         clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
 648
 649         RETURN_PACKED(pack);
 650 }
 651
 652 static unsigned
 653 bi_pack_add_blend(bi_instruction *ins, struct bi_registers *regs)
 654 {
 655         struct bifrost_add_inst pack = {
 656                 .src0 = bi_get_src(ins, regs, 0, false),
 657                 .op = BIFROST_ADD_OP_BLEND
 658         };
 659
 660         /* TODO: Pack location in uniform_const */
 661         assert(ins->blend_location == 0);
 662
 663         RETURN_PACKED(pack);
 664 }
 665
 666 static unsigned
 667 bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 668 {
 669         if (!bundle.add)
 670                 return BIFROST_ADD_NOP;
 671
 672         switch (bundle.add->type) {
 673         case BI_ADD:
 674                 return BIFROST_ADD_NOP;
 675         case BI_ATEST:
 676                 return bi_pack_add_atest(clause, bundle.add, regs);
 677         case BI_BRANCH:
 678         case BI_CMP:
 679                 return BIFROST_ADD_NOP;
 680         case BI_BLEND:
 681                 return bi_pack_add_blend(bundle.add, regs);
 682         case BI_BITWISE:
 683         case BI_CONVERT:
 684         case BI_DISCARD:
 685         case BI_FREXP:
 686         case BI_ISUB:
 687         case BI_LOAD:
 688                 return BIFROST_ADD_NOP;
 689         case BI_LOAD_ATTR:
 690                 return bi_pack_add_ld_attr(clause, bundle.add, regs);
 691         case BI_LOAD_UNIFORM:
 692                 return bi_pack_add_ld_ubo(clause, bundle.add, regs);
 693         case BI_LOAD_VAR:
 694                 return bi_pack_add_ld_vary(clause, bundle.add, regs);
 695         case BI_LOAD_VAR_ADDRESS:
 696                 return bi_pack_add_ld_var_addr(clause, bundle.add, regs);
 697         case BI_MINMAX:
 698         case BI_MOV:
 699         case BI_FMOV:
 700         case BI_SHIFT:
 701         case BI_STORE:
 702                 return BIFROST_ADD_NOP;
 703         case BI_STORE_VAR:
 704                 return bi_pack_add_st_vary(clause, bundle.add, regs);
 705         case BI_SPECIAL:
 706         case BI_SWIZZLE:
 707         case BI_TEX:
 708         case BI_ROUND:
 709                 return BIFROST_ADD_NOP;
 710         default:
 711                 unreachable("Cannot encode class as ADD");
 712         }
 713 }
 714
 715 struct bi_packed_bundle {
 716         uint64_t lo;
 717         uint64_t hi;
 718 };
 719
 720 static struct bi_packed_bundle
 721 bi_pack_bundle(bi_clause *clause, bi_bundle bundle, bi_bundle prev, bool first_bundle)
 722 {
 723         struct bi_registers regs = bi_assign_ports(bundle, prev);
 724         bi_assign_uniform_constant(clause, &regs, bundle);
 725         regs.first_instruction = first_bundle;
 726
 727         uint64_t reg = bi_pack_registers(regs);
 728         uint64_t fma = bi_pack_fma(clause, bundle, &regs);
 729         uint64_t add = bi_pack_add(clause, bundle, &regs);
 730
 731         struct bi_packed_bundle packed = {
 732                 .lo = reg | (fma << 35) | ((add & 0b111111) << 58),
 733                 .hi = add >> 6
 734         };
 735
 736         return packed;
 737 }
 738
 739 /* Packs the next two constants as a dedicated constant quadword at the end of
 740  * the clause, returning the number packed. */
 741
 742 static unsigned
 743 bi_pack_constants(bi_context *ctx, bi_clause *clause,
 744                 unsigned index,
 745                 struct util_dynarray *emission)
 746 {
 747         /* After these two, are we done? Determines tag */
 748         bool done = clause->constant_count <= (index + 2);
 749         bool only = clause->constant_count <= (index + 1);
 750
 751         /* TODO: Pos */
 752         assert(index == 0 && clause->bundle_count == 1);
 753
 754         struct bifrost_fmt_constant quad = {
 755                 .pos = 0, /* TODO */
 756                 .tag = done ? BIFROST_FMTC_FINAL : BIFROST_FMTC_CONSTANTS,
 757                 .imm_1 = clause->constants[index + 0] >> 4,
 758                 .imm_2 = only ? 0 : clause->constants[index + 1] >> 4
 759         };
 760
 761         /* XXX: On G71, Connor observed that the difference of the top 4 bits
 762          * of the second constant with the first must be less than 8, otherwise
 763          * we have to swap them. I am not able to reproduce this on G52,
 764          * further investigation needed. Possibly an errata. XXX */
 765
 766         util_dynarray_append(emission, struct bifrost_fmt_constant, quad);
 767
 768         return 2;
 769 }
 770
 771 static void
 772 bi_pack_clause(bi_context *ctx, bi_clause *clause, bi_clause *next,
 773                 struct util_dynarray *emission)
 774 {
 775         struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true);
 776         assert(clause->bundle_count == 1);
 777
 778         /* Used to decide if we elide writes */
 779         bool is_fragment = ctx->stage == MESA_SHADER_FRAGMENT;
 780
 781         /* State for packing constants throughout */
 782         unsigned constant_index = 0;
 783
 784         struct bifrost_fmt1 quad_1 = {
 785                 .tag = clause->constant_count ? BIFROST_FMT1_CONSTANTS : BIFROST_FMT1_FINAL,
 786                 .header = bi_pack_header(clause, next, is_fragment),
 787                 .ins_1 = ins_1.lo,
 788                 .ins_2 = ins_1.hi & ((1 << 11) - 1),
 789                 .ins_0 = (ins_1.hi >> 11) & 0b111,
 790         };
 791
 792         util_dynarray_append(emission, struct bifrost_fmt1, quad_1);
 793
 794         /* Pack the remaining constants */
 795
 796         while (constant_index < clause->constant_count) {
 797                 constant_index += bi_pack_constants(ctx, clause,
 798                                 constant_index, emission);
 799         }
 800 }
 801
 802 static bi_clause *
 803 bi_next_clause(bi_context *ctx, pan_block *block, bi_clause *clause)
 804 {
 805         /* Try the next clause in this block */
 806         if (clause->link.next != &((bi_block *) block)->clauses)
 807                 return list_first_entry(&(clause->link), bi_clause, link);
 808
 809         /* Try the next block, or the one after that if it's empty, etc .*/
 810         pan_block *next_block = pan_next_block(block);
 811
 812         bi_foreach_block_from(ctx, next_block, block) {
 813                 bi_block *blk = (bi_block *) block;
 814
 815                 if (!list_is_empty(&blk->clauses))
 816                         return list_first_entry(&(blk->clauses), bi_clause, link);
 817         }
 818
 819         return NULL;
 820 }
 821
 822 void
 823 bi_pack(bi_context *ctx, struct util_dynarray *emission)
 824 {
 825         util_dynarray_init(emission, NULL);
 826
 827         bi_foreach_block(ctx, _block) {
 828                 bi_block *block = (bi_block *) _block;
 829
 830                 bi_foreach_clause_in_block(block, clause) {
 831                         bi_clause *next = bi_next_clause(ctx, _block, clause);
 832                         bi_pack_clause(ctx, clause, next, emission);
 833                 }
 834         }
 835 }