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)
  38 {
  39         struct bifrost_header header = {
  40                 /* stub */
  41                 .no_end_of_shader = (next != NULL),
  42         };
  43
  44         uint64_t u = 0;
  45         memcpy(&u, &header, sizeof(header));
  46         return u;
  47 }
  48
  49 /* Represents the assignment of ports for a given bundle */
  50
  51 struct bi_registers {
  52         /* Register to assign to each port */
  53         unsigned port[4];
  54
  55         /* Read ports can be disabled */
  56         bool enabled[2];
  57
  58         /* Should we write FMA? what about ADD? If only a single port is
  59          * enabled it is in port 2, else ADD/FMA is 2/3 respectively */
  60         bool write_fma, write_add;
  61
  62         /* Should we read with port 3? */
  63         bool read_port3;
  64
  65         /* Packed uniform/constant */
  66         unsigned uniform_constant;
  67
  68         /* Whether writes are actually for the last instruction */
  69         bool first_instruction;
  70 };
  71
  72 /* Assigns a port for reading, before anything is written */
  73
  74 static void
  75 bi_assign_port_read(struct bi_registers *regs, unsigned src)
  76 {
  77         /* We only assign for registers */
  78         if (!(src & BIR_INDEX_REGISTER))
  79                 return;
  80
  81         unsigned reg = src & ~BIR_INDEX_REGISTER;
  82
  83         /* Check if we already assigned the port */
  84         for (unsigned i = 0; i <= 1; ++i) {
  85                 if (regs->port[i] == reg && regs->enabled[i])
  86                         return;
  87         }
  88
  89         if (regs->port[3] == reg && regs->read_port3)
  90                 return;
  91
  92         /* Assign it now */
  93
  94         for (unsigned i = 0; i <= 1; ++i) {
  95                 if (!regs->enabled[i]) {
  96                         regs->port[i] = reg;
  97                         regs->enabled[i] = true;
  98                         return;
  99                 }
 100         }
 101
 102         if (!regs->read_port3) {
 103                 regs->port[3] = reg;
 104                 regs->read_port3 = true;
 105         }
 106 }
 107
 108 static struct bi_registers
 109 bi_assign_ports(bi_bundle now, bi_bundle prev)
 110 {
 111         struct bi_registers regs = { 0 };
 112
 113         /* First, assign reads */
 114
 115         if (now.fma)
 116                 bi_foreach_src(now.fma, src)
 117                         bi_assign_port_read(&regs, now.fma->src[src]);
 118
 119         if (now.add)
 120                 bi_foreach_src(now.add, src)
 121                         bi_assign_port_read(&regs, now.add->src[src]);
 122
 123         /* Next, assign writes */
 124
 125         if (prev.fma && prev.fma->dest & BIR_INDEX_REGISTER) {
 126                 regs.port[2] = prev.fma->dest & ~BIR_INDEX_REGISTER;
 127                 regs.write_fma = true;
 128         }
 129
 130         if (prev.add && prev.add->dest & BIR_INDEX_REGISTER) {
 131                 unsigned r = prev.add->dest & ~BIR_INDEX_REGISTER;
 132
 133                 if (regs.write_fma) {
 134                         /* Scheduler constraint: cannot read 3 and write 2 */
 135                         assert(!regs.read_port3);
 136                         regs.port[3] = r;
 137                 } else {
 138                         regs.port[2] = r;
 139                 }
 140
 141                 regs.write_add = true;
 142         }
 143
 144         /* Finally, ensure port 1 > port 0 for the 63-x trick to function */
 145
 146         if (regs.enabled[0] && regs.enabled[1] && regs.port[1] < regs.port[0]) {
 147                 unsigned temp = regs.port[0];
 148                 regs.port[0] = regs.port[1];
 149                 regs.port[1] = temp;
 150         }
 151
 152         return regs;
 153 }
 154
 155 /* Determines the register control field, ignoring the first? flag */
 156
 157 static enum bifrost_reg_control
 158 bi_pack_register_ctrl_lo(struct bi_registers r)
 159 {
 160         if (r.write_fma) {
 161                 if (r.write_add) {
 162                         assert(!r.read_port3);
 163                         return BIFROST_WRITE_ADD_P2_FMA_P3;
 164                 } else {
 165                         if (r.read_port3)
 166                                 return BIFROST_WRITE_FMA_P2_READ_P3;
 167                         else
 168                                 return BIFROST_WRITE_FMA_P2;
 169                 }
 170         } else if (r.write_add) {
 171                 if (r.read_port3)
 172                         return BIFROST_WRITE_ADD_P2_READ_P3;
 173                 else
 174                         return BIFROST_WRITE_ADD_P2;
 175         } else if (r.read_port3)
 176                 return BIFROST_READ_P3;
 177         else
 178                 return BIFROST_REG_NONE;
 179 }
 180
 181 /* Ditto but account for the first? flag this time */
 182
 183 static enum bifrost_reg_control
 184 bi_pack_register_ctrl(struct bi_registers r)
 185 {
 186         enum bifrost_reg_control ctrl = bi_pack_register_ctrl_lo(r);
 187
 188         if (r.first_instruction) {
 189                 if (ctrl == BIFROST_REG_NONE)
 190                         ctrl = BIFROST_FIRST_NONE;
 191                 else
 192                         ctrl |= BIFROST_FIRST_NONE;
 193         }
 194
 195         return ctrl;
 196 }
 197
 198 static uint64_t
 199 bi_pack_registers(struct bi_registers regs)
 200 {
 201         enum bifrost_reg_control ctrl = bi_pack_register_ctrl(regs);
 202         struct bifrost_regs s;
 203         uint64_t packed = 0;
 204
 205         if (regs.enabled[1]) {
 206                 /* Gotta save that bit!~ Required by the 63-x trick */
 207                 assert(regs.port[1] > regs.port[0]);
 208                 assert(regs.enabled[0]);
 209
 210                 /* Do the 63-x trick, see docs/disasm */
 211                 if (regs.port[0] > 31) {
 212                         regs.port[0] = 63 - regs.port[0];
 213                         regs.port[1] = 63 - regs.port[1];
 214                 }
 215
 216                 assert(regs.port[0] <= 31);
 217                 assert(regs.port[1] <= 63);
 218
 219                 s.ctrl = ctrl;
 220                 s.reg1 = regs.port[1];
 221                 s.reg0 = regs.port[0];
 222         } else {
 223                 /* Port 1 disabled, so set to zero and use port 1 for ctrl */
 224                 s.reg1 = ctrl << 2;
 225
 226                 if (regs.enabled[0]) {
 227                         /* Bit 0 upper bit of port 0 */
 228                         s.reg1 |= (regs.port[0] >> 5);
 229
 230                         /* Rest of port 0 in usual spot */
 231                         s.reg0 = (regs.port[0] & 0b11111);
 232                 } else {
 233                         /* Bit 1 set if port 0 also disabled */
 234                         s.reg1 |= (1 << 1);
 235                 }
 236         }
 237
 238         s.reg3 = regs.port[3];
 239         s.reg2 = regs.port[2];
 240         s.uniform_const = regs.uniform_constant;
 241
 242         memcpy(&packed, &s, sizeof(s));
 243         return packed;
 244 }
 245
 246 static enum bifrost_packed_src
 247 bi_get_src_reg_port(struct bi_registers *regs, unsigned src)
 248 {
 249         unsigned reg = src & ~BIR_INDEX_REGISTER;
 250
 251         if (regs->port[0] == reg && regs->enabled[0])
 252                 return BIFROST_SRC_PORT0;
 253         else if (regs->port[1] == reg && regs->enabled[1])
 254                 return BIFROST_SRC_PORT1;
 255         else if (regs->port[3] == reg && regs->read_port3)
 256                 return BIFROST_SRC_PORT3;
 257         else
 258                 unreachable("Tried to access register with no port");
 259 }
 260
 261 static enum bifrost_packed_src
 262 bi_get_src_const(struct bi_registers *regs, unsigned constant)
 263 {
 264         if (regs->uniform_constant & (1 << 7))
 265                 unreachable("Tried to get constant but loading uniforms");
 266
 267         unsigned loc = (regs->uniform_constant >> 4) & 0x7;
 268
 269         if (loc != 0)
 270                 unreachable("TODO: constants in clauses");
 271
 272         unsigned lo = regs->uniform_constant & 0xF;
 273
 274         if (lo == 0) {
 275                 if (constant != 0)
 276                         unreachable("Tried to load !0 in 0 slot");
 277
 278                 return BIFROST_SRC_CONST_LO;
 279         } else {
 280                 unreachable("Special slot is not a fixed immediate");
 281         }
 282 }
 283
 284 static enum bifrost_packed_src
 285 bi_get_src(bi_instruction *ins, struct bi_registers *regs, unsigned s, bool is_fma)
 286 {
 287         unsigned src = ins->src[s];
 288
 289         if (src & BIR_INDEX_REGISTER)
 290                 return bi_get_src_reg_port(regs, src);
 291         else if (src & BIR_INDEX_ZERO && is_fma)
 292                 return BIFROST_SRC_STAGE;
 293         else if (src & BIR_INDEX_ZERO)
 294                 return bi_get_src_const(regs, 0);
 295         else if (src & BIR_INDEX_PASS)
 296                 return src & ~BIR_INDEX_PASS;
 297         else
 298                 unreachable("Unknown src");
 299 }
 300
 301 static unsigned
 302 bi_pack_fma_fma(bi_instruction *ins, struct bi_registers *regs)
 303 {
 304         /* (-a)(-b) = ab, so we only need one negate bit */
 305         bool negate_mul = ins->src_neg[0] ^ ins->src_neg[1];
 306
 307         struct bifrost_fma_fma pack = {
 308                 .src0 = bi_get_src(ins, regs, 0, true),
 309                 .src1 = bi_get_src(ins, regs, 1, true),
 310                 .src2 = bi_get_src(ins, regs, 2, true),
 311                 .src0_abs = ins->src_abs[0],
 312                 .src1_abs = ins->src_abs[1],
 313                 .src2_abs = ins->src_abs[2],
 314                 .src0_neg = negate_mul,
 315                 .src2_neg = ins->src_neg[2],
 316                 .op = BIFROST_FMA_OP_FMA
 317         };
 318
 319         RETURN_PACKED(pack);
 320 }
 321
 322 static unsigned
 323 bi_pack_fma_add(bi_instruction *ins, struct bi_registers *regs)
 324 {
 325         /* TODO: fadd16 packing is a bit different */
 326         assert(ins->dest_type == nir_type_float32);
 327
 328         struct bifrost_fma_add pack = {
 329                 .src0 = bi_get_src(ins, regs, 0, true),
 330                 .src1 = bi_get_src(ins, regs, 1, true),
 331                 .src0_abs = ins->src_abs[0],
 332                 .src1_abs = ins->src_abs[1],
 333                 .src0_neg = ins->src_neg[0],
 334                 .src1_neg = ins->src_neg[1],
 335                 .unk = 0x0,
 336                 .outmod = ins->outmod,
 337                 .roundmode = ins->roundmode,
 338                 .op = BIFROST_FMA_OP_FADD32
 339         };
 340
 341         RETURN_PACKED(pack);
 342 }
 343
 344 static unsigned
 345 bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 346 {
 347         if (!bundle.fma)
 348                 return BIFROST_FMA_NOP;
 349
 350         switch (bundle.fma->type) {
 351         case BI_ADD:
 352                 return bi_pack_fma_add(bundle.fma, regs);
 353         case BI_CMP:
 354         case BI_BITWISE:
 355         case BI_CONVERT:
 356         case BI_CSEL:
 357                 return BIFROST_FMA_NOP;
 358         case BI_FMA:
 359                 return bi_pack_fma_fma(bundle.fma, regs);
 360         case BI_FREXP:
 361         case BI_ISUB:
 362         case BI_MINMAX:
 363         case BI_MOV:
 364         case BI_SHIFT:
 365         case BI_SWIZZLE:
 366         case BI_ROUND:
 367                 return BIFROST_FMA_NOP;
 368         default:
 369                 unreachable("Cannot encode class as FMA");
 370         }
 371 }
 372
 373 static unsigned
 374 bi_pack_add_ld_vary(bi_instruction *ins, struct bi_registers *regs)
 375 {
 376         unsigned size = nir_alu_type_get_type_size(ins->dest_type);
 377         assert(size == 32 || size == 16);
 378
 379         unsigned op = (size == 32) ?
 380                 BIFROST_ADD_OP_LD_VAR_32 :
 381                 BIFROST_ADD_OP_LD_VAR_16;
 382
 383         unsigned cmask = bi_from_bytemask(ins->writemask, size / 8);
 384         unsigned channels = util_bitcount(cmask);
 385         assert(cmask == ((1 << channels) - 1));
 386
 387         unsigned packed_addr = 0;
 388
 389         if (ins->src[0] & BIR_INDEX_CONSTANT) {
 390                 /* Direct uses address field directly */
 391                 packed_addr = ins->src[0] & ~BIR_INDEX_CONSTANT;
 392                 assert(packed_addr < 0b1000);
 393         } else {
 394                 /* Indirect gets an extra source */
 395                 packed_addr = bi_get_src(ins, regs, 0, false) | 0b11000;
 396         }
 397
 398         assert(channels >= 1 && channels <= 4);
 399
 400         struct bifrost_ld_var pack = {
 401                 .src0 = bi_get_src(ins, regs, 1, false),
 402                 .addr = packed_addr,
 403                 .channels = MALI_POSITIVE(channels),
 404                 .interp_mode = ins->load_vary.interp_mode,
 405                 .reuse = ins->load_vary.reuse,
 406                 .flat = ins->load_vary.flat,
 407                 .op = op
 408         };
 409
 410         RETURN_PACKED(pack);
 411 }
 412
 413 static unsigned
 414 bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
 415 {
 416         if (!bundle.add)
 417                 return BIFROST_ADD_NOP;
 418
 419         switch (bundle.add->type) {
 420         case BI_ADD:
 421         case BI_ATEST:
 422         case BI_BRANCH:
 423         case BI_CMP:
 424         case BI_BLEND:
 425         case BI_BITWISE:
 426         case BI_CONVERT:
 427         case BI_DISCARD:
 428         case BI_FREXP:
 429         case BI_ISUB:
 430         case BI_LOAD:
 431         case BI_LOAD_UNIFORM:
 432         case BI_LOAD_ATTR:
 433                 return BIFROST_ADD_NOP;
 434         case BI_LOAD_VAR:
 435                 return bi_pack_add_ld_vary(bundle.add, regs);
 436         case BI_LOAD_VAR_ADDRESS:
 437         case BI_MINMAX:
 438         case BI_MOV:
 439         case BI_SHIFT:
 440         case BI_STORE:
 441         case BI_STORE_VAR:
 442         case BI_SPECIAL:
 443         case BI_SWIZZLE:
 444         case BI_TEX:
 445         case BI_ROUND:
 446                 return BIFROST_ADD_NOP;
 447         default:
 448                 unreachable("Cannot encode class as ADD");
 449         }
 450 }
 451
 452 struct bi_packed_bundle {
 453         uint64_t lo;
 454         uint64_t hi;
 455 };
 456
 457 static struct bi_packed_bundle
 458 bi_pack_bundle(bi_clause *clause, bi_bundle bundle, bi_bundle prev, bool first_bundle)
 459 {
 460         struct bi_registers regs = bi_assign_ports(bundle, prev);
 461         regs.first_instruction = first_bundle;
 462
 463         uint64_t reg = bi_pack_registers(regs);
 464         uint64_t fma = bi_pack_fma(clause, bundle, &regs);
 465         uint64_t add = bi_pack_add(clause, bundle, &regs);
 466
 467         struct bi_packed_bundle packed = {
 468                 .lo = reg | (fma << 35) | ((add & 0b111111) << 58),
 469                 .hi = add >> 6
 470         };
 471
 472         return packed;
 473 }
 474
 475 static void
 476 bi_pack_clause(bi_context *ctx, bi_clause *clause, bi_clause *next,
 477                 struct util_dynarray *emission)
 478 {
 479         struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true);
 480         assert(clause->bundle_count == 1);
 481
 482         struct bifrost_fmt1 quad_1 = {
 483                 .tag = BIFROST_FMT1_FINAL,
 484                 .header = bi_pack_header(clause, next),
 485                 .ins_1 = ins_1.lo,
 486                 .ins_2 = ins_1.hi & ((1 << 11) - 1),
 487                 .ins_0 = (ins_1.hi >> 11) & 0b111,
 488         };
 489
 490         util_dynarray_append(emission, struct bifrost_fmt1, quad_1);
 491 }
 492
 493 static bi_clause *
 494 bi_next_clause(bi_context *ctx, pan_block *block, bi_clause *clause)
 495 {
 496         /* Try the next clause in this block */
 497         if (clause->link.next != &((bi_block *) block)->clauses)
 498                 return list_first_entry(&(clause->link), bi_clause, link);
 499
 500         /* Try the next block, or the one after that if it's empty, etc .*/
 501         pan_block *next_block = pan_next_block(block);
 502
 503         bi_foreach_block_from(ctx, next_block, block) {
 504                 bi_block *blk = (bi_block *) block;
 505
 506                 if (!list_is_empty(&blk->clauses))
 507                         return list_first_entry(&(blk->clauses), bi_clause, link);
 508         }
 509
 510         return NULL;
 511 }
 512
 513 void
 514 bi_pack(bi_context *ctx, struct util_dynarray *emission)
 515 {
 516         util_dynarray_init(emission, NULL);
 517
 518         bi_foreach_block(ctx, _block) {
 519                 bi_block *block = (bi_block *) _block;
 520
 521                 bi_foreach_clause_in_block(block, clause) {
 522                         bi_clause *next = bi_next_clause(ctx, _block, clause);
 523                         bi_pack_clause(ctx, clause, next, emission);
 524                 }
 525         }
 526 }