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         unsigned uniform_constant;
  75
  76         /* Whether writes are actually for the last instruction */
  77         bool first_instruction;
  78 };
  79
  80 /* Assigns a port for reading, before anything is written */
  81
  82 static void
  83 bi_assign_port_read(struct bi_registers *regs, unsigned src)
  84 {
  85         /* We only assign for registers */
  86         if (!(src & BIR_INDEX_REGISTER))
  87                 return;
  88
  89         unsigned reg = src & ~BIR_INDEX_REGISTER;
  90
  91         /* Check if we already assigned the port */
  92         for (unsigned i = 0; i <= 1; ++i) {
  93                 if (regs->port[i] == reg && regs->enabled[i])
  94                         return;
  95         }
  96
  97         if (regs->port[3] == reg && regs->read_port3)
  98                 return;
  99
 100         /* Assign it now */
 101
 102         for (unsigned i = 0; i <= 1; ++i) {
 103                 if (!regs->enabled[i]) {
 104                         regs->port[i] = reg;
 105                         regs->enabled[i] = true;
 106                         return;
 107                 }
 108         }
 109
 110         if (!regs->read_port3) {
 111                 regs->port[3] = reg;
 112                 regs->read_port3 = true;
 113         }
 114 }
 115
 116 static struct bi_registers
 117 bi_assign_ports(bi_bundle now, bi_bundle prev)
 118 {
 119         struct bi_registers regs = { 0 };
 120
 121         /* We assign ports for the main register mechanism. Special ops
 122          * use the data registers, which has its own mechanism entirely
 123          * and thus gets skipped over here. */
 124
 125         unsigned read_dreg = now.add &&
 126                 bi_class_props[now.add->type] & BI_DATA_REG_SRC;
 127
 128         unsigned write_dreg = prev.add &&
 129                 bi_class_props[prev.add->type] & BI_DATA_REG_DEST;
 130
 131         /* First, assign reads */
 132
 133         if (now.fma)
 134                 bi_foreach_src(now.fma, src)
 135                         bi_assign_port_read(&regs, now.fma->src[src]);
 136
 137         if (now.add) {
 138                 bi_foreach_src(now.add, src) {
 139                         if (!(src == 0 && read_dreg))
 140                                 bi_assign_port_read(&regs, now.add->src[src]);
 141                 }
 142         }
 143
 144         /* Next, assign writes */
 145
 146         if (prev.fma && prev.fma->dest & BIR_INDEX_REGISTER) {
 147                 regs.port[2] = prev.fma->dest & ~BIR_INDEX_REGISTER;
 148                 regs.write_fma = true;
 149         }
 150
 151         if (prev.add && prev.add->dest & BIR_INDEX_REGISTER && !write_dreg) {
 152                 unsigned r = prev.add->dest & ~BIR_INDEX_REGISTER;
 153
 154                 if (regs.write_fma) {
 155                         /* Scheduler constraint: cannot read 3 and write 2 */
 156                         assert(!regs.read_port3);
 157                         regs.port[3] = r;
 158                 } else {
 159                         regs.port[2] = r;
 160                 }
 161
 162                 regs.write_add = true;
 163         }
 164
 165         /* Finally, ensure port 1 > port 0 for the 63-x trick to function */
 166
 167         if (regs.enabled[0] && regs.enabled[1] && regs.port[1] < regs.port[0]) {
 168                 unsigned temp = regs.port[0];
 169                 regs.port[0] = regs.port[1];
 170                 regs.port[1] = temp;
 171         }
 172
 173         return regs;
 174 }
 175
 176 /* Determines the register control field, ignoring the first? flag */
 177
 178 static enum bifrost_reg_control
 179 bi_pack_register_ctrl_lo(struct bi_registers r)
 180 {
 181         if (r.write_fma) {
 182                 if (r.write_add) {
 183                         assert(!r.read_port3);
 184                         return BIFROST_WRITE_ADD_P2_FMA_P3;
 185                 } else {
 186                         if (r.read_port3)
 187                                 return BIFROST_WRITE_FMA_P2_READ_P3;
 188                         else
 189                                 return BIFROST_WRITE_FMA_P2;
 190                 }
 191         } else if (r.write_add) {
 192                 if (r.read_port3)
 193                         return BIFROST_WRITE_ADD_P2_READ_P3;
 194                 else
 195                         return BIFROST_WRITE_ADD_P2;
 196         } else if (r.read_port3)
 197                 return BIFROST_READ_P3;
 198         else
 199                 return BIFROST_REG_NONE;
 200 }
 201
 202 /* Ditto but account for the first? flag this time */
 203
 204 static enum bifrost_reg_control
 205 bi_pack_register_ctrl(struct bi_registers r)
 206 {
 207         enum bifrost_reg_control ctrl = bi_pack_register_ctrl_lo(r);
 208
 209         if (r.first_instruction) {
 210                 if (ctrl == BIFROST_REG_NONE)
 211                         ctrl = BIFROST_FIRST_NONE;
 212                 else
 213                         ctrl |= BIFROST_FIRST_NONE;
 214         }
 215
 216         return ctrl;
 217 }
 218
 219 static uint64_t
 220 bi_pack_registers(struct bi_registers regs)
 221 {
 222         enum bifrost_reg_control ctrl = bi_pack_register_ctrl(regs);
 223         struct bifrost_regs s;
 224         uint64_t packed = 0;
 225
 226         if (regs.enabled[1]) {
 227                 /* Gotta save that bit!~ Required by the 63-x trick */
 228                 assert(regs.port[1] > regs.port[0]);
 229                 assert(regs.enabled[0]);
 230
 231                 /* Do the 63-x trick, see docs/disasm */
 232                 if (regs.port[0] > 31) {
 233                         regs.port[0] = 63 - regs.port[0];
 234                         regs.port[1] = 63 - regs.port[1];
 235                 }
 236
 237                 assert(regs.port[0] <= 31);
 238                 assert(regs.port[1] <= 63);
 239
 240                 s.ctrl = ctrl;
 241                 s.reg1 = regs.port[1];
 242                 s.reg0 = regs.port[0];
 243         } else {
 244                 /* Port 1 disabled, so set to zero and use port 1 for ctrl */
 245                 s.reg1 = ctrl << 2;
 246
 247                 if (regs.enabled[0]) {
 248                         /* Bit 0 upper bit of port 0 */
 249                         s.reg1 |= (regs.port[0] >> 5);
 250
 251                         /* Rest of port 0 in usual spot */
 252                         s.reg0 = (regs.port[0] & 0b11111);
 253                 } else {
 254                         /* Bit 1 set if port 0 also disabled */
 255                         s.reg1 |= (1 << 1);
 256                 }
 257         }
 258
 259         s.reg3 = regs.port[3];
 260         s.reg2 = regs.port[2];
 261         s.uniform_const = regs.uniform_constant;
 262
 263         memcpy(&packed, &s, sizeof(s));
 264         return packed;
 265 }
 266
 267 static enum bifrost_packed_src
 268 bi_get_src_reg_port(struct bi_registers *regs, unsigned src)
 269 {
 270         unsigned reg = src & ~BIR_INDEX_REGISTER;
 271
 272         if (regs->port[0] == reg && regs->enabled[0])
 273                 return BIFROST_SRC_PORT0;
 274         else if (regs->port[1] == reg && regs->enabled[1])
 275                 return BIFROST_SRC_PORT1;
 276         else if (regs->port[3] == reg && regs->read_port3)
 277                 return BIFROST_SRC_PORT3;
 278         else
 279                 unreachable("Tried to access register with no port");
 280 }
 281
 282 static enum bifrost_packed_src
 283 bi_get_src_const(struct bi_registers *regs, unsigned constant)
 284 {
 285         if (regs->uniform_constant & (1 << 7))
 286                 unreachable("Tried to get constant but loading uniforms");
 287
 288         unsigned loc = (regs->uniform_constant >> 4) & 0x7;
 289
 290         if (loc != 0)
 291                 unreachable("TODO: constants in clauses");
 292
 293         unsigned lo = regs->uniform_constant & 0xF;
 294
 295         if (lo == 0) {
 296                 if (constant != 0)
 297                         unreachable("Tried to load !0 in 0 slot");
 298
 299                 return BIFROST_SRC_CONST_LO;
 300         } else {
 301                 unreachable("Special slot is not a fixed immediate");
 302         }
 303 }
 304
 305 static enum bifrost_packed_src
 306 bi_get_src(bi_instruction *ins, struct bi_registers *regs, unsigned s, bool is_fma)
 307 {
 308         unsigned src = ins->src[s];
 309
 310         if (src & BIR_INDEX_REGISTER)
 311                 return bi_get_src_reg_port(regs, src);
 312         else if (src & BIR_INDEX_ZERO && is_fma)
 313                 return BIFROST_SRC_STAGE;
 314         else if (src & BIR_INDEX_ZERO)
 315                 return bi_get_src_const(regs, 0);
 316         else if (src & BIR_INDEX_PASS)
 317                 return src & ~BIR_INDEX_PASS;
 318         else
 319                 unreachable("Unknown src");
 320 }
 321
 322 static unsigned
 323 bi_pack_fma_fma(bi_instruction *ins, struct bi_registers *regs)
 324 {
 325         /* (-a)(-b) = ab, so we only need one negate bit */
 326         bool negate_mul = ins->src_neg[0] ^ ins->src_neg[1];
 327
 328         struct bifrost_fma_fma pack = {
 329                 .src0 = bi_get_src(ins, regs, 0, true),
 330                 .src1 = bi_get_src(ins, regs, 1, true),
 331                 .src2 = bi_get_src(ins, regs, 2, true),
 332                 .src0_abs = ins->src_abs[0],
 333                 .src1_abs = ins->src_abs[1],
 334                 .src2_abs = ins->src_abs[2],
 335                 .src0_neg = negate_mul,
 336                 .src2_neg = ins->src_neg[2],
 337                 .op = BIFROST_FMA_OP_FMA
 338         };
 339
 340         RETURN_PACKED(pack);
 341 }
 342
 343 static unsigned
 344 bi_pack_fma_add(bi_instruction *ins, struct bi_registers *regs)
 345 {
 346         /* TODO: fadd16 packing is a bit different */
 347         assert(ins->dest_type == nir_type_float32);
 348
 349         struct bifrost_fma_add pack = {
 350                 .src0 = bi_get_src(ins, regs, 0, true),
 351                 .src1 = bi_get_src(ins, regs, 1, true),
 352                 .src0_abs = ins->src_abs[0],
 353                 .src1_abs = ins->src_abs[1],
 354                 .src0_neg = ins->src_neg[0],
 355                 .src1_neg = ins->src_neg[1],
 356                 .unk = 0x0,
 357                 .outmod = ins->outmod,
 358                 .roundmode = ins->roundmode,
 359                 .op = BIFROST_FMA_OP_FADD32
 360         };
 361
 362         RETURN_PACKED(pack);
 363 }
 364
 365 static unsigned
 366 bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 367 {
 368         if (!bundle.fma)
 369                 return BIFROST_FMA_NOP;
 370
 371         switch (bundle.fma->type) {
 372         case BI_ADD:
 373                 return bi_pack_fma_add(bundle.fma, regs);
 374         case BI_CMP:
 375         case BI_BITWISE:
 376         case BI_CONVERT:
 377         case BI_CSEL:
 378                 return BIFROST_FMA_NOP;
 379         case BI_FMA:
 380                 return bi_pack_fma_fma(bundle.fma, regs);
 381         case BI_FREXP:
 382         case BI_ISUB:
 383         case BI_MINMAX:
 384         case BI_MOV:
 385         case BI_SHIFT:
 386         case BI_SWIZZLE:
 387         case BI_ROUND:
 388                 return BIFROST_FMA_NOP;
 389         default:
 390                 unreachable("Cannot encode class as FMA");
 391         }
 392 }
 393
 394 static unsigned
 395 bi_pack_add_ld_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 396 {
 397         unsigned size = nir_alu_type_get_type_size(ins->dest_type);
 398         assert(size == 32 || size == 16);
 399
 400         unsigned op = (size == 32) ?
 401                 BIFROST_ADD_OP_LD_VAR_32 :
 402                 BIFROST_ADD_OP_LD_VAR_16;
 403
 404         unsigned cmask = bi_from_bytemask(ins->writemask, size / 8);
 405         unsigned channels = util_bitcount(cmask);
 406         assert(cmask == ((1 << channels) - 1));
 407
 408         unsigned packed_addr = 0;
 409
 410         if (ins->src[0] & BIR_INDEX_CONSTANT) {
 411                 /* Direct uses address field directly */
 412                 packed_addr = ins->src[0] & ~BIR_INDEX_CONSTANT;
 413                 assert(packed_addr < 0b1000);
 414         } else {
 415                 /* Indirect gets an extra source */
 416                 packed_addr = bi_get_src(ins, regs, 0, false) | 0b11000;
 417         }
 418
 419         /* The destination is thrown in the data register */
 420         assert(ins->dest & BIR_INDEX_REGISTER);
 421         clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
 422
 423         assert(channels >= 1 && channels <= 4);
 424
 425         struct bifrost_ld_var pack = {
 426                 .src0 = bi_get_src(ins, regs, 1, false),
 427                 .addr = packed_addr,
 428                 .channels = MALI_POSITIVE(channels),
 429                 .interp_mode = ins->load_vary.interp_mode,
 430                 .reuse = ins->load_vary.reuse,
 431                 .flat = ins->load_vary.flat,
 432                 .op = op
 433         };
 434
 435         RETURN_PACKED(pack);
 436 }
 437
 438 static unsigned
 439 bi_pack_add_atest(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
 440 {
 441         /* TODO: fp16 */
 442         assert(ins->src_types[1] == nir_type_float32);
 443
 444         struct bifrost_add_atest pack = {
 445                 .src0 = bi_get_src(ins, regs, 0, false),
 446                 .src1 = bi_get_src(ins, regs, 1, false),
 447                 .component = 1, /* Set for fp32 */
 448                 .op = BIFROST_ADD_OP_ATEST,
 449         };
 450
 451         /* Despite *also* writing with the usual mechanism... quirky and
 452          * perhaps unnecessary, but let's match the blob */
 453         clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
 454
 455         RETURN_PACKED(pack);
 456 }
 457
 458 static unsigned
 459 bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 460 {
 461         if (!bundle.add)
 462                 return BIFROST_ADD_NOP;
 463
 464         switch (bundle.add->type) {
 465         case BI_ADD:
 466                 return BIFROST_ADD_NOP;
 467         case BI_ATEST:
 468                 return bi_pack_add_atest(clause, bundle.add, regs);
 469         case BI_BRANCH:
 470         case BI_CMP:
 471         case BI_BLEND:
 472         case BI_BITWISE:
 473         case BI_CONVERT:
 474         case BI_DISCARD:
 475         case BI_FREXP:
 476         case BI_ISUB:
 477         case BI_LOAD:
 478         case BI_LOAD_UNIFORM:
 479         case BI_LOAD_ATTR:
 480                 return BIFROST_ADD_NOP;
 481         case BI_LOAD_VAR:
 482                 return bi_pack_add_ld_vary(clause, bundle.add, regs);
 483         case BI_LOAD_VAR_ADDRESS:
 484         case BI_MINMAX:
 485         case BI_MOV:
 486         case BI_SHIFT:
 487         case BI_STORE:
 488         case BI_STORE_VAR:
 489         case BI_SPECIAL:
 490         case BI_SWIZZLE:
 491         case BI_TEX:
 492         case BI_ROUND:
 493                 return BIFROST_ADD_NOP;
 494         default:
 495                 unreachable("Cannot encode class as ADD");
 496         }
 497 }
 498
 499 struct bi_packed_bundle {
 500         uint64_t lo;
 501         uint64_t hi;
 502 };
 503
 504 static struct bi_packed_bundle
 505 bi_pack_bundle(bi_clause *clause, bi_bundle bundle, bi_bundle prev, bool first_bundle)
 506 {
 507         struct bi_registers regs = bi_assign_ports(bundle, prev);
 508         regs.first_instruction = first_bundle;
 509
 510         uint64_t reg = bi_pack_registers(regs);
 511         uint64_t fma = bi_pack_fma(clause, bundle, &regs);
 512         uint64_t add = bi_pack_add(clause, bundle, &regs);
 513
 514         struct bi_packed_bundle packed = {
 515                 .lo = reg | (fma << 35) | ((add & 0b111111) << 58),
 516                 .hi = add >> 6
 517         };
 518
 519         return packed;
 520 }
 521
 522 static void
 523 bi_pack_clause(bi_context *ctx, bi_clause *clause, bi_clause *next,
 524                 struct util_dynarray *emission)
 525 {
 526         struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true);
 527         assert(clause->bundle_count == 1);
 528
 529         /* Used to decide if we elide writes */
 530         bool is_fragment = ctx->stage == MESA_SHADER_FRAGMENT;
 531
 532         struct bifrost_fmt1 quad_1 = {
 533                 .tag = BIFROST_FMT1_FINAL,
 534                 .header = bi_pack_header(clause, next, is_fragment),
 535                 .ins_1 = ins_1.lo,
 536                 .ins_2 = ins_1.hi & ((1 << 11) - 1),
 537                 .ins_0 = (ins_1.hi >> 11) & 0b111,
 538         };
 539
 540         util_dynarray_append(emission, struct bifrost_fmt1, quad_1);
 541 }
 542
 543 static bi_clause *
 544 bi_next_clause(bi_context *ctx, pan_block *block, bi_clause *clause)
 545 {
 546         /* Try the next clause in this block */
 547         if (clause->link.next != &((bi_block *) block)->clauses)
 548                 return list_first_entry(&(clause->link), bi_clause, link);
 549
 550         /* Try the next block, or the one after that if it's empty, etc .*/
 551         pan_block *next_block = pan_next_block(block);
 552
 553         bi_foreach_block_from(ctx, next_block, block) {
 554                 bi_block *blk = (bi_block *) block;
 555
 556                 if (!list_is_empty(&blk->clauses))
 557                         return list_first_entry(&(blk->clauses), bi_clause, link);
 558         }
 559
 560         return NULL;
 561 }
 562
 563 void
 564 bi_pack(bi_context *ctx, struct util_dynarray *emission)
 565 {
 566         util_dynarray_init(emission, NULL);
 567
 568         bi_foreach_block(ctx, _block) {
 569                 bi_block *block = (bi_block *) _block;
 570
 571                 bi_foreach_clause_in_block(block, clause) {
 572                         bi_clause *next = bi_next_clause(ctx, _block, clause);
 573                         bi_pack_clause(ctx, clause, next, emission);
 574                 }
 575         }
 576 }