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 = next ? next->dependencies : 0,
  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, bool *hi, bool b64)
  88 {
  89         uint64_t want = (cons >> 4);
  90
  91         for (unsigned i = 0; i < clause->constant_count; ++i) {
  92                 /* Only check top 60-bits since that's what's actually embedded
  93                  * in the clause, the bottom 4-bits are bundle-inline */
  94
  95                 unsigned candidates[2] = {
  96                         clause->constants[i] >> 4,
  97                         clause->constants[i] >> 36
  98                 };
  99
 100                 if (!b64)
 101                         candidates[0] &= 0xFFFFFFFF;
 102
 103                 if (candidates[0] == want)
 104                         return i;
 105
 106                 if (candidates[1] == want && !b64) {
 107                         *hi = true;
 108                         return i;
 109                 }
 110         }
 111
 112         unreachable("Invalid constant accessed");
 113 }
 114
 115 static unsigned
 116 bi_constant_field(unsigned idx)
 117 {
 118         assert(idx <= 5);
 119
 120         const unsigned values[] = {
 121                 4, 5, 6, 7, 2, 3
 122         };
 123
 124         return values[idx] << 4;
 125 }
 126
 127 static bool
 128 bi_assign_uniform_constant_single(
 129                 struct bi_registers *regs,
 130                 bi_clause *clause,
 131                 bi_instruction *ins, bool assigned, bool fast_zero)
 132 {
 133         if (!ins)
 134                 return assigned;
 135
 136         bi_foreach_src(ins, s) {
 137                 if (s == 0 && (ins->type == BI_LOAD_VAR_ADDRESS || ins->type == BI_LOAD_ATTR)) continue;
 138
 139                 if (ins->src[s] & BIR_INDEX_CONSTANT) {
 140                         bool hi = false;
 141                         bool b64 = nir_alu_type_get_type_size(ins->src_types[s]) > 32;
 142                         uint64_t cons = bi_get_immediate(ins, ins->src[s]);
 143                         unsigned idx = bi_lookup_constant(clause, cons, &hi, b64);
 144                         unsigned f = bi_constant_field(idx) | (cons & 0xF);
 145
 146                         if (assigned && regs->uniform_constant != f)
 147                                 unreachable("Mismatched uniform/const field: imm");
 148
 149                         regs->uniform_constant = f;
 150                         ins->src[s] = BIR_INDEX_PASS | (hi ? BIFROST_SRC_CONST_HI : BIFROST_SRC_CONST_LO);
 151                         assigned = true;
 152                 } else if (ins->src[s] & BIR_INDEX_ZERO && (ins->type == BI_LOAD_UNIFORM || ins->type == BI_LOAD_VAR)) {
 153                         /* XXX: HACK UNTIL WE HAVE HI MATCHING DUE TO OVERFLOW XXX */
 154                         ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_HI;
 155                 } else if (ins->src[s] & BIR_INDEX_ZERO && !fast_zero) {
 156                         /* FMAs have a fast zero port, ADD needs to use the
 157                          * uniform/const port's special 0 mode handled here */
 158                         unsigned f = 0;
 159
 160                         if (assigned && regs->uniform_constant != f)
 161                                 unreachable("Mismatched uniform/const field: 0");
 162
 163                         regs->uniform_constant = f;
 164                         ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_LO;
 165                         assigned = true;
 166                 } else if (s & BIR_INDEX_UNIFORM) {
 167                         unreachable("Push uniforms not implemented yet");
 168                 }
 169         }
 170
 171         return assigned;
 172 }
 173
 174 static void
 175 bi_assign_uniform_constant(
 176                 bi_clause *clause,
 177                 struct bi_registers *regs,
 178                 bi_bundle bundle)
 179 {
 180         bool assigned =
 181                 bi_assign_uniform_constant_single(regs, clause, bundle.fma, false, true);
 182
 183         bi_assign_uniform_constant_single(regs, clause, bundle.add, assigned, false);
 184 }
 185
 186 /* Assigns a port for reading, before anything is written */
 187
 188 static void
 189 bi_assign_port_read(struct bi_registers *regs, unsigned src)
 190 {
 191         /* We only assign for registers */
 192         if (!(src & BIR_INDEX_REGISTER))
 193                 return;
 194
 195         unsigned reg = src & ~BIR_INDEX_REGISTER;
 196
 197         /* Check if we already assigned the port */
 198         for (unsigned i = 0; i <= 1; ++i) {
 199                 if (regs->port[i] == reg && regs->enabled[i])
 200                         return;
 201         }
 202
 203         if (regs->port[3] == reg && regs->read_port3)
 204                 return;
 205
 206         /* Assign it now */
 207
 208         for (unsigned i = 0; i <= 1; ++i) {
 209                 if (!regs->enabled[i]) {
 210                         regs->port[i] = reg;
 211                         regs->enabled[i] = true;
 212                         return;
 213                 }
 214         }
 215
 216         if (!regs->read_port3) {
 217                 regs->port[3] = reg;
 218                 regs->read_port3 = true;
 219         }
 220 }
 221
 222 static struct bi_registers
 223 bi_assign_ports(bi_bundle now, bi_bundle prev)
 224 {
 225         struct bi_registers regs = { 0 };
 226
 227         /* We assign ports for the main register mechanism. Special ops
 228          * use the data registers, which has its own mechanism entirely
 229          * and thus gets skipped over here. */
 230
 231         unsigned read_dreg = now.add &&
 232                 bi_class_props[now.add->type] & BI_DATA_REG_SRC;
 233
 234         unsigned write_dreg = prev.add &&
 235                 bi_class_props[prev.add->type] & BI_DATA_REG_DEST;
 236
 237         /* First, assign reads */
 238
 239         if (now.fma)
 240                 bi_foreach_src(now.fma, src)
 241                         bi_assign_port_read(&regs, now.fma->src[src]);
 242
 243         if (now.add) {
 244                 bi_foreach_src(now.add, src) {
 245                         if (!(src == 0 && read_dreg))
 246                                 bi_assign_port_read(&regs, now.add->src[src]);
 247                 }
 248         }
 249
 250         /* Next, assign writes */
 251
 252         if (prev.fma && prev.fma->dest & BIR_INDEX_REGISTER) {
 253                 regs.port[2] = prev.fma->dest & ~BIR_INDEX_REGISTER;
 254                 regs.write_fma = true;
 255         }
 256
 257         if (prev.add && prev.add->dest & BIR_INDEX_REGISTER && !write_dreg) {
 258                 unsigned r = prev.add->dest & ~BIR_INDEX_REGISTER;
 259
 260                 if (regs.write_fma) {
 261                         /* Scheduler constraint: cannot read 3 and write 2 */
 262                         assert(!regs.read_port3);
 263                         regs.port[3] = r;
 264                 } else {
 265                         regs.port[2] = r;
 266                 }
 267
 268                 regs.write_add = true;
 269         }
 270
 271         /* Finally, ensure port 1 > port 0 for the 63-x trick to function */
 272
 273         if (regs.enabled[0] && regs.enabled[1] && regs.port[1] < regs.port[0]) {
 274                 unsigned temp = regs.port[0];
 275                 regs.port[0] = regs.port[1];
 276                 regs.port[1] = temp;
 277         }
 278
 279         return regs;
 280 }
 281
 282 /* Determines the register control field, ignoring the first? flag */
 283
 284 static enum bifrost_reg_control
 285 bi_pack_register_ctrl_lo(struct bi_registers r)
 286 {
 287         if (r.write_fma) {
 288                 if (r.write_add) {
 289                         assert(!r.read_port3);
 290                         return BIFROST_WRITE_ADD_P2_FMA_P3;
 291                 } else {
 292                         if (r.read_port3)
 293                                 return BIFROST_WRITE_FMA_P2_READ_P3;
 294                         else
 295                                 return BIFROST_WRITE_FMA_P2;
 296                 }
 297         } else if (r.write_add) {
 298                 if (r.read_port3)
 299                         return BIFROST_WRITE_ADD_P2_READ_P3;
 300                 else
 301                         return BIFROST_WRITE_ADD_P2;
 302         } else if (r.read_port3)
 303                 return BIFROST_READ_P3;
 304         else
 305                 return BIFROST_REG_NONE;
 306 }
 307
 308 /* Ditto but account for the first? flag this time */
 309
 310 static enum bifrost_reg_control
 311 bi_pack_register_ctrl(struct bi_registers r)
 312 {
 313         enum bifrost_reg_control ctrl = bi_pack_register_ctrl_lo(r);
 314
 315         if (r.first_instruction) {
 316                 if (ctrl == BIFROST_REG_NONE)
 317                         ctrl = BIFROST_FIRST_NONE;
 318                 else
 319                         ctrl |= BIFROST_FIRST_NONE;
 320         }
 321
 322         return ctrl;
 323 }
 324
 325 static uint64_t
 326 bi_pack_registers(struct bi_registers regs)
 327 {
 328         enum bifrost_reg_control ctrl = bi_pack_register_ctrl(regs);
 329         struct bifrost_regs s;
 330         uint64_t packed = 0;
 331
 332         if (regs.enabled[1]) {
 333                 /* Gotta save that bit!~ Required by the 63-x trick */
 334                 assert(regs.port[1] > regs.port[0]);
 335                 assert(regs.enabled[0]);
 336
 337                 /* Do the 63-x trick, see docs/disasm */
 338                 if (regs.port[0] > 31) {
 339                         regs.port[0] = 63 - regs.port[0];
 340                         regs.port[1] = 63 - regs.port[1];
 341                 }
 342
 343                 assert(regs.port[0] <= 31);
 344                 assert(regs.port[1] <= 63);
 345
 346                 s.ctrl = ctrl;
 347                 s.reg1 = regs.port[1];
 348                 s.reg0 = regs.port[0];
 349         } else {
 350                 /* Port 1 disabled, so set to zero and use port 1 for ctrl */
 351                 s.reg1 = ctrl << 2;
 352
 353                 if (regs.enabled[0]) {
 354                         /* Bit 0 upper bit of port 0 */
 355                         s.reg1 |= (regs.port[0] >> 5);
 356
 357                         /* Rest of port 0 in usual spot */
 358                         s.reg0 = (regs.port[0] & 0b11111);
 359                 } else {
 360                         /* Bit 1 set if port 0 also disabled */
 361                         s.reg1 |= (1 << 1);
 362                 }
 363         }
 364
 365         s.reg3 = regs.port[3];
 366         s.reg2 = regs.port[2];
 367         s.uniform_const = regs.uniform_constant;
 368
 369         memcpy(&packed, &s, sizeof(s));
 370         return packed;
 371 }
 372
 373 static void
 374 bi_set_data_register(bi_clause *clause, unsigned idx)
 375 {
 376         assert(idx & BIR_INDEX_REGISTER);
 377         unsigned reg = idx & ~BIR_INDEX_REGISTER;
 378         assert(reg <= 63);
 379         clause->data_register = reg;
 380 }
 381
 382 static void
 383 bi_read_data_register(bi_clause *clause, bi_instruction *ins)
 384 {
 385         bi_set_data_register(clause, ins->src[0]);
 386 }
 387
 388 static void
 389 bi_write_data_register(bi_clause *clause, bi_instruction *ins)
 390 {
 391         bi_set_data_register(clause, ins->dest);
 392 }
 393
 394 static enum bifrost_packed_src
 395 bi_get_src_reg_port(struct bi_registers *regs, unsigned src)
 396 {
 397         unsigned reg = src & ~BIR_INDEX_REGISTER;
 398
 399         if (regs->port[0] == reg && regs->enabled[0])
 400                 return BIFROST_SRC_PORT0;
 401         else if (regs->port[1] == reg && regs->enabled[1])
 402                 return BIFROST_SRC_PORT1;
 403         else if (regs->port[3] == reg && regs->read_port3)
 404                 return BIFROST_SRC_PORT3;
 405         else
 406                 unreachable("Tried to access register with no port");
 407 }
 408
 409 static enum bifrost_packed_src
 410 bi_get_src(bi_instruction *ins, struct bi_registers *regs, unsigned s, bool is_fma)
 411 {
 412         unsigned src = ins->src[s];
 413
 414         if (src & BIR_INDEX_REGISTER)
 415                 return bi_get_src_reg_port(regs, src);
 416         else if (src & BIR_INDEX_ZERO && is_fma)
 417                 return BIFROST_SRC_STAGE;
 418         else if (src & BIR_INDEX_PASS)
 419                 return src & ~BIR_INDEX_PASS;
 420         else
 421                 unreachable("Unknown src");
 422 }
 423
 424 /* Constructs a packed 2-bit swizzle for a 16-bit vec2 source. Source must be
 425  * 16-bit and written components must correspond to valid swizzles (component x
 426  * or y). */
 427
 428 static unsigned
 429 bi_swiz16(bi_instruction *ins, unsigned src)
 430 {
 431         assert(nir_alu_type_get_type_size(ins->src_types[src]) == 16);
 432         unsigned swizzle = 0;
 433
 434         for (unsigned c = 0; c < 2; ++c) {
 435                 if (!bi_writes_component(ins, src)) continue;
 436
 437                 unsigned k = ins->swizzle[src][c];
 438                 assert(k < 1);
 439                 swizzle |= (k << c);
 440         }
 441
 442         return swizzle;
 443 }
 444
 445 static unsigned
 446 bi_pack_fma_fma(bi_instruction *ins, struct bi_registers *regs)
 447 {
 448         /* (-a)(-b) = ab, so we only need one negate bit */
 449         bool negate_mul = ins->src_neg[0] ^ ins->src_neg[1];
 450
 451         if (ins->dest_type == nir_type_float32) {
 452                 struct bifrost_fma_fma pack = {
 453                         .src0 = bi_get_src(ins, regs, 0, true),
 454                         .src1 = bi_get_src(ins, regs, 1, true),
 455                         .src2 = bi_get_src(ins, regs, 2, true),
 456                         .src0_abs = ins->src_abs[0],
 457                         .src1_abs = ins->src_abs[1],
 458                         .src2_abs = ins->src_abs[2],
 459                         .src0_neg = negate_mul,
 460                         .src2_neg = ins->src_neg[2],
 461                         .outmod = ins->outmod,
 462                         .roundmode = ins->roundmode,
 463                         .op = BIFROST_FMA_OP_FMA
 464                 };
 465
 466                 RETURN_PACKED(pack);
 467         } else if (ins->dest_type == nir_type_float16) {
 468                 struct bifrost_fma_fma16 pack = {
 469                         .src0 = bi_get_src(ins, regs, 0, true),
 470                         .src1 = bi_get_src(ins, regs, 1, true),
 471                         .src2 = bi_get_src(ins, regs, 2, true),
 472                         .swizzle_0 = bi_swiz16(ins, 0),
 473                         .swizzle_1 = bi_swiz16(ins, 1),
 474                         .swizzle_2 = bi_swiz16(ins, 2),
 475                         .src0_neg = negate_mul,
 476                         .src2_neg = ins->src_neg[2],
 477                         .outmod = ins->outmod,
 478                         .roundmode = ins->roundmode,
 479                         .op = BIFROST_FMA_OP_FMA16
 480                 };
 481
 482                 RETURN_PACKED(pack);
 483         } else {
 484                 unreachable("Invalid fma dest type");
 485         }
 486 }
 487
 488 static unsigned
 489 bi_pack_fma_add_f32(bi_instruction *ins, struct bi_registers *regs)
 490 {
 491         struct bifrost_fma_add pack = {
 492                 .src0 = bi_get_src(ins, regs, 0, true),
 493                 .src1 = bi_get_src(ins, regs, 1, true),
 494                 .src0_abs = ins->src_abs[0],
 495                 .src1_abs = ins->src_abs[1],
 496                 .src0_neg = ins->src_neg[0],
 497                 .src1_neg = ins->src_neg[1],
 498                 .unk = 0x0,
 499                 .outmod = ins->outmod,
 500                 .roundmode = ins->roundmode,
 501                 .op = BIFROST_FMA_OP_FADD32
 502         };
 503
 504         RETURN_PACKED(pack);
 505 }
 506
 507 static unsigned
 508 bi_pack_fma_addmin_f16(bi_instruction *ins, struct bi_registers *regs)
 509 {
 510         unsigned op =
 511                 (ins->type == BI_ADD) ? BIFROST_FMA_OP_FADD16 :
 512                 (ins->op.minmax == BI_MINMAX_MIN) ? BIFROST_FMA_OP_FMIN16 :
 513                 BIFROST_FMA_OP_FMAX16;
 514
 515         /* Absolute values are packed in a quirky way. Let k = src1 < src0. Let
 516          * l be an auxiliary bit we encode. Then the hardware determines:
 517          *
 518          *      abs0 = l || k
 519          *      abs1 = l && k
 520          *
 521          * Since add/min/max are commutative, this saves a bit by using the
 522          * order of the operands as a bit (k). To pack this, first note:
 523          *
 524          *      (l && k) implies (l || k).
 525          *
 526          * That is, if the second argument is abs'd, then the first argument
 527          * also has abs. So there are three cases:
 528          *
 529          * Case 0: Neither src has absolute value. Then we have l = k = 0.
 530          *
 531          * Case 1: Exactly one src has absolute value. Assign that source to
 532          * src0 and the other source to src1. Compute k = src1 < src0 based on
 533          * that assignment. Then l = ~k.
 534          *
 535          * Case 2: Both sources have absolute value. Then we have l = k = 1.
 536          * Note to force k = 1 requires that (src1 < src0) OR (src0 < src1).
 537          * That is, this encoding is only valid if src1 and src0 are distinct.
 538          * This is a scheduling restriction (XXX); if an op of this type
 539          * requires both identical sources to have abs value, then we must
 540          * schedule to ADD (which does not use this ordering trick).
 541          */
 542
 543         unsigned abs_0 = ins->src_abs[0], abs_1 = ins->src_abs[1];
 544         unsigned src_0 = bi_get_src(ins, regs, 0, true);
 545         unsigned src_1 = bi_get_src(ins, regs, 0, true);
 546         bool l = false;
 547
 548         if (!abs_0 && !abs_1) {
 549                 /* Force k = 0 <===> NOT(src1 < src0) <==> src1 >= src0 */
 550                 if (src_0 < src_1) {
 551                         unsigned tmp = src_0;
 552                         src_0 = src_1;
 553                         src_1 = tmp;
 554                 }
 555         } else if (abs_0 && !abs_1) {
 556                 l = src_1 >= src_0;
 557         } else if (abs_1 && !abs_0) {
 558                 unsigned tmp = src_0;
 559                 src_0 = src_1;
 560                 src_0 = tmp;
 561
 562                 l = src_1 >= src_0;
 563         } else {
 564                 if (src_0 >= src_1) {
 565                         unsigned tmp = src_0;
 566                         src_0 = src_1;
 567                         src_1 = tmp;
 568                 }
 569
 570                 l = true;
 571         }
 572
 573         struct bifrost_fma_add_minmax16 pack = {
 574                 .src0 = src_0,
 575                 .src1 = src_1,
 576                 .src0_neg = ins->src_neg[0],
 577                 .src1_neg = ins->src_neg[1],
 578                 .abs1 = l,
 579                 .outmod = ins->outmod,
 580                 .mode = (ins->type == BI_ADD) ? ins->roundmode : ins->minmax,
 581                 .op = op
 582         };
 583
 584         RETURN_PACKED(pack);
 585 }
 586
 587 static unsigned
 588 bi_pack_fma_add(bi_instruction *ins, struct bi_registers *regs)
 589 {
 590         if (ins->dest_type == nir_type_float32)
 591                 return bi_pack_fma_add_f32(ins, regs);
 592         else if(ins->dest_type == nir_type_float16)
 593                 return bi_pack_fma_addmin_f16(ins, regs);
 594         else
 595                 unreachable("Unknown FMA/ADD type");
 596 }
 597
 598 static unsigned
 599 bi_pack_fma_1src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
 600 {
 601         struct bifrost_fma_inst pack = {
 602                 .src0 = bi_get_src(ins, regs, 0, true),
 603                 .op = op
 604         };
 605
 606         RETURN_PACKED(pack);
 607 }
 608
 609 static enum bifrost_csel_cond
 610 bi_cond_to_csel(enum bi_cond cond, bool *flip, bool *invert, nir_alu_type T)
 611 {
 612         nir_alu_type B = nir_alu_type_get_base_type(T);
 613         unsigned idx = (B == nir_type_float) ? 0 :
 614                 ((B == nir_type_int) ? 1 : 2);
 615
 616         switch (cond){
 617         case BI_COND_LT:
 618                 *flip = true;
 619         case BI_COND_GT: {
 620                 const enum bifrost_csel_cond ops[] = {
 621                         BIFROST_FGT_F,
 622                         BIFROST_IGT_I,
 623                         BIFROST_UGT_I
 624                 };
 625
 626                 return ops[idx];
 627         }
 628         case BI_COND_LE:
 629                 *flip = true;
 630         case BI_COND_GE: {
 631                 const enum bifrost_csel_cond ops[] = {
 632                         BIFROST_FGE_F,
 633                         BIFROST_IGE_I,
 634                         BIFROST_UGE_I
 635                 };
 636
 637                 return ops[idx];
 638         }
 639         case BI_COND_NE:
 640                 *invert = true;
 641         case BI_COND_EQ: {
 642                 const enum bifrost_csel_cond ops[] = {
 643                         BIFROST_FEQ_F,
 644                         BIFROST_IEQ_F,
 645                         BIFROST_IEQ_F /* sign is irrelevant */
 646                 };
 647
 648                 return ops[idx];
 649         }
 650         default:
 651                 unreachable("Invalid op for csel");
 652         }
 653 }
 654
 655 static unsigned
 656 bi_pack_fma_csel(bi_instruction *ins, struct bi_registers *regs)
 657 {
 658         /* TODO: Use csel3 as well */
 659         bool flip = false, invert = false;
 660
 661         enum bifrost_csel_cond cond =
 662                 bi_cond_to_csel(ins->csel_cond, &flip, &invert, ins->src_types[0]);
 663
 664         unsigned size = nir_alu_type_get_type_size(ins->dest_type);
 665
 666         unsigned cmp_0 = (flip ? 3 : 0);
 667         unsigned cmp_1 = (flip ? 0 : 3);
 668         unsigned res_0 = (invert ? 2 : 1);
 669         unsigned res_1 = (invert ? 1 : 2);
 670
 671         struct bifrost_csel4 pack = {
 672                 .src0 = bi_get_src(ins, regs, cmp_0, true),
 673                 .src1 = bi_get_src(ins, regs, cmp_1, true),
 674                 .src2 = bi_get_src(ins, regs, res_0, true),
 675                 .src3 = bi_get_src(ins, regs, res_1, true),
 676                 .cond = cond,
 677                 .op = (size == 16) ? BIFROST_FMA_OP_CSEL4_V16 :
 678                         BIFROST_FMA_OP_CSEL4
 679         };
 680
 681         RETURN_PACKED(pack);
 682 }
 683
 684 /* We have a single convert opcode in the IR but a number of opcodes that could
 685  * come out. In particular we have native opcodes for:
 686  *
 687  * [ui]16 --> [fui]32           -- int16_to_32
 688  * f16     --> f32              -- float16_to_32
 689  * f32     --> f16              -- float32_to_16
 690  * f32     --> [ui]32           -- float32_to_int
 691  * [ui]32  --> f32              -- int_to_float32
 692  * [fui]16 --> [fui]16          -- f2i_i2f16
 693  */
 694
 695 static unsigned
 696 bi_pack_fma_convert(bi_instruction *ins, struct bi_registers *regs)
 697 {
 698         nir_alu_type from_base = nir_alu_type_get_base_type(ins->src_types[0]);
 699         unsigned from_size = nir_alu_type_get_type_size(ins->src_types[0]);
 700         bool from_unsigned = from_base == nir_type_uint;
 701
 702         nir_alu_type to_base = nir_alu_type_get_base_type(ins->dest_type);
 703         unsigned to_size = nir_alu_type_get_type_size(ins->dest_type);
 704         bool to_unsigned = to_base == nir_type_uint;
 705
 706         /* Sanity check */
 707         assert((from_base != to_base) || (from_size != to_size));
 708         assert((MAX2(from_size, to_size) / MIN2(from_size, to_size)) <= 2);
 709
 710         if (from_size == 16 && to_size == 16) {
 711                 /* f2i_i2f16 */
 712                 unreachable("i16 not yet implemented");
 713         } else if (from_size == 32 && to_size == 32) {
 714                 unsigned op = 0;
 715
 716                 if (from_base == nir_type_float) {
 717                         op = BIFROST_FMA_FLOAT32_TO_INT(to_unsigned);
 718                 } else {
 719                         op = BIFROST_FMA_INT_TO_FLOAT32(from_unsigned);
 720                 }
 721
 722                 return bi_pack_fma_1src(ins, regs, op);
 723         } else if (from_size == 16 && to_size == 32) {
 724                 bool from_y = ins->swizzle[0][0];
 725
 726                 if (from_base == nir_type_float) {
 727                         return bi_pack_fma_1src(ins, regs,
 728                                         BIFROST_FMA_FLOAT16_TO_32(from_y));
 729                 } else {
 730                         unreachable("i16 not yet implemented");
 731                 }
 732         } else if (from_size == 32 && to_size == 16) {
 733                 if (from_base == nir_type_float) {
 734                         /* TODO: second vectorized source? */
 735                         struct bifrost_fma_2src pack = {
 736                                 .src0 = bi_get_src(ins, regs, 0, true),
 737                                 .src1 = BIFROST_SRC_STAGE, /* 0 */
 738                                 .op = BIFROST_FMA_FLOAT32_TO_16
 739                         };
 740
 741                         RETURN_PACKED(pack);
 742                 } else {
 743                         unreachable("i16 not yet implemented");
 744                 }
 745         }
 746
 747         unreachable("Unknown convert");
 748 }
 749
 750 static unsigned
 751 bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 752 {
 753         if (!bundle.fma)
 754                 return BIFROST_FMA_NOP;
 755
 756         switch (bundle.fma->type) {
 757         case BI_ADD:
 758                 return bi_pack_fma_add(bundle.fma, regs);
 759         case BI_CMP:
 760         case BI_BITWISE:
 761                 return BIFROST_FMA_NOP;
 762         case BI_CONVERT:
 763                 return bi_pack_fma_convert(bundle.fma, regs);
 764         case BI_CSEL:
 765                 return bi_pack_fma_csel(bundle.fma, regs);
 766         case BI_FMA:
 767                 return bi_pack_fma_fma(bundle.fma, regs);
 768         case BI_FREXP:
 769         case BI_ISUB:
 770         case BI_MINMAX:
 771                 return BIFROST_FMA_NOP;
 772         case BI_MOV:
 773                 return bi_pack_fma_1src(bundle.fma, regs, BIFROST_FMA_OP_MOV);
 774         case BI_FMOV:
 775         case BI_SHIFT:
 776         case BI_SWIZZLE:
 777         case BI_ROUND:
 778                 return BIFROST_FMA_NOP;
 779         default:
 780                 unreachable("Cannot encode class as FMA");
 781         }
 782 }
 783
 784 static unsigned
 785 bi_pack_add_ld_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 786 {
 787         unsigned size = nir_alu_type_get_type_size(ins->dest_type);
 788         assert(size == 32 || size == 16);
 789
 790         unsigned op = (size == 32) ?
 791                 BIFROST_ADD_OP_LD_VAR_32 :
 792                 BIFROST_ADD_OP_LD_VAR_16;
 793
 794         unsigned cmask = bi_from_bytemask(ins->writemask, size / 8);
 795         unsigned channels = util_bitcount(cmask);
 796         assert(cmask == ((1 << channels) - 1));
 797
 798         unsigned packed_addr = 0;
 799
 800         if (ins->src[0] & BIR_INDEX_CONSTANT) {
 801                 /* Direct uses address field directly */
 802                 packed_addr = bi_get_immediate(ins, ins->src[0]);
 803                 assert(packed_addr < 0b1000);
 804         } else {
 805                 /* Indirect gets an extra source */
 806                 packed_addr = bi_get_src(ins, regs, 0, false) | 0b11000;
 807         }
 808
 809         /* The destination is thrown in the data register */
 810         assert(ins->dest & BIR_INDEX_REGISTER);
 811         clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
 812
 813         assert(channels >= 1 && channels <= 4);
 814
 815         struct bifrost_ld_var pack = {
 816                 .src0 = bi_get_src(ins, regs, 1, false),
 817                 .addr = packed_addr,
 818                 .channels = MALI_POSITIVE(channels),
 819                 .interp_mode = ins->load_vary.interp_mode,
 820                 .reuse = ins->load_vary.reuse,
 821                 .flat = ins->load_vary.flat,
 822                 .op = op
 823         };
 824
 825         RETURN_PACKED(pack);
 826 }
 827
 828 static unsigned
 829 bi_pack_add_2src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
 830 {
 831         struct bifrost_add_2src pack = {
 832                 .src0 = bi_get_src(ins, regs, 0, true),
 833                 .src1 = bi_get_src(ins, regs, 1, true),
 834                 .op = op
 835         };
 836
 837         RETURN_PACKED(pack);
 838 }
 839
 840 static unsigned
 841 bi_pack_add_ld_ubo(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 842 {
 843         unsigned components = bi_load32_components(ins);
 844
 845         const unsigned ops[4] = {
 846                 BIFROST_ADD_OP_LD_UBO_1,
 847                 BIFROST_ADD_OP_LD_UBO_2,
 848                 BIFROST_ADD_OP_LD_UBO_3,
 849                 BIFROST_ADD_OP_LD_UBO_4
 850         };
 851
 852         bi_write_data_register(clause, ins);
 853         return bi_pack_add_2src(ins, regs, ops[components - 1]);
 854 }
 855
 856 static enum bifrost_ldst_type
 857 bi_pack_ldst_type(nir_alu_type T)
 858 {
 859         switch (T) {
 860         case nir_type_float16: return BIFROST_LDST_F16;
 861         case nir_type_float32: return BIFROST_LDST_F32;
 862         case nir_type_int32:   return BIFROST_LDST_I32;
 863         case nir_type_uint32:  return BIFROST_LDST_U32;
 864         default: unreachable("Invalid type loaded");
 865         }
 866 }
 867
 868 static unsigned
 869 bi_pack_add_ld_var_addr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 870 {
 871         struct bifrost_ld_var_addr pack = {
 872                 .src0 = bi_get_src(ins, regs, 1, false),
 873                 .src1 = bi_get_src(ins, regs, 2, false),
 874                 .location = bi_get_immediate(ins, ins->src[0]),
 875                 .type = bi_pack_ldst_type(ins->src_types[3]),
 876                 .op = BIFROST_ADD_OP_LD_VAR_ADDR
 877         };
 878
 879         bi_write_data_register(clause, ins);
 880         RETURN_PACKED(pack);
 881 }
 882
 883 static unsigned
 884 bi_pack_add_ld_attr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 885 {
 886         struct bifrost_ld_attr pack = {
 887                 .src0 = bi_get_src(ins, regs, 1, false),
 888                 .src1 = bi_get_src(ins, regs, 2, false),
 889                 .location = bi_get_immediate(ins, ins->src[0]),
 890                 .channels = MALI_POSITIVE(bi_load32_components(ins)),
 891                 .type = bi_pack_ldst_type(ins->dest_type),
 892                 .op = BIFROST_ADD_OP_LD_ATTR
 893         };
 894
 895         bi_write_data_register(clause, ins);
 896         RETURN_PACKED(pack);
 897 }
 898
 899 static unsigned
 900 bi_pack_add_st_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 901 {
 902         assert(ins->store_channels >= 1 && ins->store_channels <= 4);
 903
 904         struct bifrost_st_vary pack = {
 905                 .src0 = bi_get_src(ins, regs, 1, false),
 906                 .src1 = bi_get_src(ins, regs, 2, false),
 907                 .src2 = bi_get_src(ins, regs, 3, false),
 908                 .channels = MALI_POSITIVE(ins->store_channels),
 909                 .op = BIFROST_ADD_OP_ST_VAR
 910         };
 911
 912         bi_read_data_register(clause, ins);
 913         RETURN_PACKED(pack);
 914 }
 915
 916 static unsigned
 917 bi_pack_add_atest(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 918 {
 919         /* TODO: fp16 */
 920         assert(ins->src_types[1] == nir_type_float32);
 921
 922         struct bifrost_add_atest pack = {
 923                 .src0 = bi_get_src(ins, regs, 0, false),
 924                 .src1 = bi_get_src(ins, regs, 1, false),
 925                 .component = 1, /* Set for fp32 */
 926                 .op = BIFROST_ADD_OP_ATEST,
 927         };
 928
 929         /* Despite *also* writing with the usual mechanism... quirky and
 930          * perhaps unnecessary, but let's match the blob */
 931         clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
 932
 933         RETURN_PACKED(pack);
 934 }
 935
 936 static unsigned
 937 bi_pack_add_blend(bi_instruction *ins, struct bi_registers *regs)
 938 {
 939         struct bifrost_add_inst pack = {
 940                 .src0 = bi_get_src(ins, regs, 0, false),
 941                 .op = BIFROST_ADD_OP_BLEND
 942         };
 943
 944         /* TODO: Pack location in uniform_const */
 945         assert(ins->blend_location == 0);
 946
 947         RETURN_PACKED(pack);
 948 }
 949
 950 static unsigned
 951 bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 952 {
 953         if (!bundle.add)
 954                 return BIFROST_ADD_NOP;
 955
 956         switch (bundle.add->type) {
 957         case BI_ADD:
 958                 return BIFROST_ADD_NOP;
 959         case BI_ATEST:
 960                 return bi_pack_add_atest(clause, bundle.add, regs);
 961         case BI_BRANCH:
 962         case BI_CMP:
 963                 return BIFROST_ADD_NOP;
 964         case BI_BLEND:
 965                 return bi_pack_add_blend(bundle.add, regs);
 966         case BI_BITWISE:
 967         case BI_CONVERT:
 968         case BI_DISCARD:
 969         case BI_FREXP:
 970         case BI_ISUB:
 971         case BI_LOAD:
 972                 return BIFROST_ADD_NOP;
 973         case BI_LOAD_ATTR:
 974                 return bi_pack_add_ld_attr(clause, bundle.add, regs);
 975         case BI_LOAD_UNIFORM:
 976                 return bi_pack_add_ld_ubo(clause, bundle.add, regs);
 977         case BI_LOAD_VAR:
 978                 return bi_pack_add_ld_vary(clause, bundle.add, regs);
 979         case BI_LOAD_VAR_ADDRESS:
 980                 return bi_pack_add_ld_var_addr(clause, bundle.add, regs);
 981         case BI_MINMAX:
 982         case BI_MOV:
 983         case BI_FMOV:
 984         case BI_SHIFT:
 985         case BI_STORE:
 986                 return BIFROST_ADD_NOP;
 987         case BI_STORE_VAR:
 988                 return bi_pack_add_st_vary(clause, bundle.add, regs);
 989         case BI_SPECIAL:
 990         case BI_SWIZZLE:
 991         case BI_TEX:
 992         case BI_ROUND:
 993                 return BIFROST_ADD_NOP;
 994         default:
 995                 unreachable("Cannot encode class as ADD");
 996         }
 997 }
 998
 999 struct bi_packed_bundle {
1000         uint64_t lo;
1001         uint64_t hi;
1002 };
1003
1004 static struct bi_packed_bundle
1005 bi_pack_bundle(bi_clause *clause, bi_bundle bundle, bi_bundle prev, bool first_bundle)
1006 {
1007         struct bi_registers regs = bi_assign_ports(bundle, prev);
1008         bi_assign_uniform_constant(clause, &regs, bundle);
1009         regs.first_instruction = first_bundle;
1010
1011         uint64_t reg = bi_pack_registers(regs);
1012         uint64_t fma = bi_pack_fma(clause, bundle, &regs);
1013         uint64_t add = bi_pack_add(clause, bundle, &regs);
1014
1015         struct bi_packed_bundle packed = {
1016                 .lo = reg | (fma << 35) | ((add & 0b111111) << 58),
1017                 .hi = add >> 6
1018         };
1019
1020         return packed;
1021 }
1022
1023 /* Packs the next two constants as a dedicated constant quadword at the end of
1024  * the clause, returning the number packed. */
1025
1026 static unsigned
1027 bi_pack_constants(bi_context *ctx, bi_clause *clause,
1028                 unsigned index,
1029                 struct util_dynarray *emission)
1030 {
1031         /* After these two, are we done? Determines tag */
1032         bool done = clause->constant_count <= (index + 2);
1033         bool only = clause->constant_count <= (index + 1);
1034
1035         /* TODO: Pos */
1036         assert(index == 0 && clause->bundle_count == 1);
1037
1038         struct bifrost_fmt_constant quad = {
1039                 .pos = 0, /* TODO */
1040                 .tag = done ? BIFROST_FMTC_FINAL : BIFROST_FMTC_CONSTANTS,
1041                 .imm_1 = clause->constants[index + 0] >> 4,
1042                 .imm_2 = only ? 0 : clause->constants[index + 1] >> 4
1043         };
1044
1045         /* XXX: On G71, Connor observed that the difference of the top 4 bits
1046          * of the second constant with the first must be less than 8, otherwise
1047          * we have to swap them. I am not able to reproduce this on G52,
1048          * further investigation needed. Possibly an errata. XXX */
1049
1050         util_dynarray_append(emission, struct bifrost_fmt_constant, quad);
1051
1052         return 2;
1053 }
1054
1055 static void
1056 bi_pack_clause(bi_context *ctx, bi_clause *clause, bi_clause *next,
1057                 struct util_dynarray *emission)
1058 {
1059         struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true);
1060         assert(clause->bundle_count == 1);
1061
1062         /* Used to decide if we elide writes */
1063         bool is_fragment = ctx->stage == MESA_SHADER_FRAGMENT;
1064
1065         /* State for packing constants throughout */
1066         unsigned constant_index = 0;
1067
1068         struct bifrost_fmt1 quad_1 = {
1069                 .tag = clause->constant_count ? BIFROST_FMT1_CONSTANTS : BIFROST_FMT1_FINAL,
1070                 .header = bi_pack_header(clause, next, is_fragment),
1071                 .ins_1 = ins_1.lo,
1072                 .ins_2 = ins_1.hi & ((1 << 11) - 1),
1073                 .ins_0 = (ins_1.hi >> 11) & 0b111,
1074         };
1075
1076         util_dynarray_append(emission, struct bifrost_fmt1, quad_1);
1077
1078         /* Pack the remaining constants */
1079
1080         while (constant_index < clause->constant_count) {
1081                 constant_index += bi_pack_constants(ctx, clause,
1082                                 constant_index, emission);
1083         }
1084 }
1085
1086 static bi_clause *
1087 bi_next_clause(bi_context *ctx, pan_block *block, bi_clause *clause)
1088 {
1089         /* Try the next clause in this block */
1090         if (clause->link.next != &((bi_block *) block)->clauses)
1091                 return list_first_entry(&(clause->link), bi_clause, link);
1092
1093         /* Try the next block, or the one after that if it's empty, etc .*/
1094         pan_block *next_block = pan_next_block(block);
1095
1096         bi_foreach_block_from(ctx, next_block, block) {
1097                 bi_block *blk = (bi_block *) block;
1098
1099                 if (!list_is_empty(&blk->clauses))
1100                         return list_first_entry(&(blk->clauses), bi_clause, link);
1101         }
1102
1103         return NULL;
1104 }
1105
1106 void
1107 bi_pack(bi_context *ctx, struct util_dynarray *emission)
1108 {
1109         util_dynarray_init(emission, NULL);
1110
1111         bi_foreach_block(ctx, _block) {
1112                 bi_block *block = (bi_block *) _block;
1113
1114                 bi_foreach_clause_in_block(block, clause) {
1115                         bi_clause *next = bi_next_clause(ctx, _block, clause);
1116                         bi_pack_clause(ctx, clause, next, emission);
1117                 }
1118         }
1119 }