src/panfrost/bifrost/compiler.h

   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  * Authors (Collabora):
  24  *      Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  25  */
  26
  27 #ifndef __BIFROST_COMPILER_H
  28 #define __BIFROST_COMPILER_H
  29
  30 #include "bifrost.h"
  31 #include "compiler/nir/nir.h"
  32 #include "panfrost/util/pan_ir.h"
  33
  34 /* Bifrost opcodes are tricky -- the same op may exist on both FMA and
  35  * ADD with two completely different opcodes, and opcodes can be varying
  36  * length in some cases. Then we have different opcodes for int vs float
  37  * and then sometimes even for different typesizes. Further, virtually
  38  * every op has a number of flags which depend on the op. In constrast
  39  * to Midgard where you have a strict ALU/LDST/TEX division and within
  40  * ALU you have strict int/float and that's it... here it's a *lot* more
  41  * involved. As such, we use something much higher level for our IR,
  42  * encoding "classes" of operations, letting the opcode details get
  43  * sorted out at emit time.
  44  *
  45  * Please keep this list alphabetized. Please use a dictionary if you
  46  * don't know how to do that.
  47  */
  48
  49 enum bi_class {
  50         BI_ADD,
  51         BI_ATEST,
  52         BI_BRANCH,
  53         BI_CMP,
  54         BI_BLEND,
  55         BI_BITWISE,
  56         BI_CONVERT,
  57         BI_CSEL,
  58         BI_DISCARD,
  59         BI_FMA,
  60         BI_FREXP,
  61         BI_ISUB,
  62         BI_LOAD,
  63         BI_LOAD_UNIFORM,
  64         BI_LOAD_ATTR,
  65         BI_LOAD_VAR,
  66         BI_LOAD_VAR_ADDRESS,
  67         BI_MINMAX,
  68         BI_MOV,
  69         BI_SHIFT,
  70         BI_STORE,
  71         BI_STORE_VAR,
  72         BI_SPECIAL, /* _FAST, _TABLE on supported GPUs */
  73         BI_SWIZZLE,
  74         BI_TEX,
  75         BI_ROUND,
  76         BI_NUM_CLASSES
  77 };
  78
  79 /* Properties of a class... */
  80 extern unsigned bi_class_props[BI_NUM_CLASSES];
  81
  82 /* abs/neg/outmod valid for a float op */
  83 #define BI_MODS (1 << 0)
  84
  85 /* Generic enough that little class-specific information is required. In other
  86  * words, it acts as a "normal" ALU op, even if the encoding ends up being
  87  * irregular enough to warrant a separate class */
  88 #define BI_GENERIC (1 << 1)
  89
  90 /* Accepts a bifrost_roundmode */
  91 #define BI_ROUNDMODE (1 << 2)
  92
  93 /* Can be scheduled to FMA */
  94 #define BI_SCHED_FMA (1 << 3)
  95
  96 /* Can be scheduled to ADD */
  97 #define BI_SCHED_ADD (1 << 4)
  98
  99 /* Most ALU ops can do either, actually */
 100 #define BI_SCHED_ALL (BI_SCHED_FMA | BI_SCHED_ADD)
 101
 102 /* Along with setting BI_SCHED_ADD, eats up the entire cycle, so FMA must be
 103  * nopped out. Used for _FAST operations. */
 104 #define BI_SCHED_SLOW (1 << 5)
 105
 106 /* Swizzling allowed for the 8/16-bit source */
 107 #define BI_SWIZZLABLE (1 << 6)
 108
 109 /* For scheduling purposes this is a high latency instruction and must be at
 110  * the end of a clause. Implies ADD */
 111 #define BI_SCHED_HI_LATENCY (1 << 7)
 112
 113 /* Intrinsic is vectorized and should read 4 components regardless of writemask */
 114 #define BI_VECTOR (1 << 8)
 115
 116 /* Use a data register for src0/dest respectively, bypassing the usual
 117  * register accessor. Mutually exclusive. */
 118 #define BI_DATA_REG_SRC (1 << 9)
 119 #define BI_DATA_REG_DEST (1 << 10)
 120
 121 /* It can't get any worse than csel4... can it? */
 122 #define BIR_SRC_COUNT 4
 123
 124 /* BI_LD_VARY */
 125 struct bi_load_vary {
 126         enum bifrost_interp_mode interp_mode;
 127         bool reuse;
 128         bool flat;
 129 };
 130
 131 /* BI_BRANCH encoding the details of the branch itself as well as a pointer to
 132  * the target. We forward declare bi_block since this is mildly circular (not
 133  * strictly, but this order of the file makes more sense I think)
 134  *
 135  * We define our own enum of conditions since the conditions in the hardware
 136  * packed in crazy ways that would make manipulation unweildly (meaning changes
 137  * based on port swapping, etc), so we defer dealing with that until emit time.
 138  * Likewise, we expose NIR types instead of the crazy branch types, although
 139  * the restrictions do eventually apply of course. */
 140
 141 struct bi_block;
 142
 143 enum bi_cond {
 144         BI_COND_ALWAYS,
 145         BI_COND_LT,
 146         BI_COND_LE,
 147         BI_COND_GE,
 148         BI_COND_GT,
 149         BI_COND_EQ,
 150         BI_COND_NE,
 151 };
 152
 153 struct bi_branch {
 154         /* Types are specified in src_types and must be compatible (either both
 155          * int, or both float, 16/32, and same size or 32/16 if float. Types
 156          * ignored if BI_COND_ALWAYS is set for an unconditional branch. */
 157
 158         enum bi_cond cond;
 159         struct bi_block *target;
 160 };
 161
 162 /* Opcodes within a class */
 163 enum bi_minmax_op {
 164         BI_MINMAX_MIN,
 165         BI_MINMAX_MAX
 166 };
 167
 168 enum bi_bitwise_op {
 169         BI_BITWISE_AND,
 170         BI_BITWISE_OR,
 171         BI_BITWISE_XOR
 172 };
 173
 174 enum bi_round_op {
 175         BI_ROUND_MODE, /* use round mode */
 176         BI_ROUND_ROUND /* i.e.: fround() */
 177 };
 178
 179 enum bi_special_op {
 180         BI_SPECIAL_FRCP,
 181         BI_SPECIAL_FRSQ,
 182         BI_SPECIAL_FATAN,
 183         BI_SPECIAL_FSIN,
 184         BI_SPECIAL_FCOS,
 185         BI_SPECIAL_FEXP,
 186         BI_SPECIAL_FLOG2,
 187         BI_SPECIAL_FLOGE
 188 };
 189
 190 typedef struct {
 191         struct list_head link; /* Must be first */
 192         enum bi_class type;
 193
 194         /* Indices, see bir_ssa_index etc. Note zero is special cased
 195          * to "no argument" */
 196         unsigned dest;
 197         unsigned src[BIR_SRC_COUNT];
 198
 199         /* If one of the sources has BIR_INDEX_CONSTANT */
 200         union {
 201                 uint64_t u64;
 202                 uint32_t u32;
 203                 uint16_t u16[2];
 204                 uint8_t u8[4];
 205         } constant;
 206
 207         /* Floating-point modifiers, type/class permitting. If not
 208          * allowed for the type/class, these are ignored. */
 209         enum bifrost_outmod outmod;
 210         bool src_abs[BIR_SRC_COUNT];
 211         bool src_neg[BIR_SRC_COUNT];
 212
 213         /* Round mode (requires BI_ROUNDMODE) */
 214         enum bifrost_roundmode roundmode;
 215
 216         /* Writemask (bit for each affected byte). This is quite restricted --
 217          * ALU ops can only write to a single channel (exception: <32 in which
 218          * you can write to 32/N contiguous aligned channels). Load/store can
 219          * only write to all channels at once, in a sense. But it's still
 220          * better to use this generic form than have synthetic ops flying
 221          * about, since we're not essentially vector for RA purposes. */
 222         uint16_t writemask;
 223
 224         /* Destination type. Usually the type of the instruction
 225          * itself, but if sources and destination have different
 226          * types, the type of the destination wins (so f2i would be
 227          * int). Zero if there is no destination. Bitsize included */
 228         nir_alu_type dest_type;
 229
 230         /* Source types if required by the class */
 231         nir_alu_type src_types[BIR_SRC_COUNT];
 232
 233         /* If the source type is 8-bit or 16-bit such that SIMD is possible,
 234          * and the class has BI_SWIZZLABLE, this is a swizzle in the usual
 235          * sense. On non-SIMD instructions, it can be used for component
 236          * selection, so we don't have to special case extraction. */
 237         uint8_t swizzle[BIR_SRC_COUNT][NIR_MAX_VEC_COMPONENTS];
 238
 239         /* A class-specific op from which the actual opcode can be derived
 240          * (along with the above information) */
 241
 242         union {
 243                 enum bi_minmax_op minmax;
 244                 enum bi_bitwise_op bitwise;
 245                 enum bi_round_op round;
 246                 enum bi_special_op special;
 247                 enum bi_cond compare;
 248         } op;
 249
 250         /* Union for class-specific information */
 251         union {
 252                 enum bifrost_minmax_mode minmax;
 253                 struct bi_load_vary load_vary;
 254                 struct bi_branch branch;
 255
 256                 /* For CSEL, the comparison op. BI_COND_ALWAYS doesn't make
 257                  * sense here but you can always just use a move for that */
 258                 enum bi_cond csel_cond;
 259
 260                 /* For BLEND -- the location 0-7 */
 261                 unsigned blend_location;
 262         };
 263 } bi_instruction;
 264
 265 /* Scheduling takes place in two steps. Step 1 groups instructions within a
 266  * block into distinct clauses (bi_clause). Step 2 schedules instructions
 267  * within a clause into FMA/ADD pairs (bi_bundle).
 268  *
 269  * A bi_bundle contains two paired instruction pointers. If a slot is unfilled,
 270  * leave it NULL; the emitter will fill in a nop.
 271  */
 272
 273 typedef struct {
 274         bi_instruction *fma;
 275         bi_instruction *add;
 276 } bi_bundle;
 277
 278 typedef struct {
 279         struct list_head link;
 280
 281         /* A clause can have 8 instructions in bundled FMA/ADD sense, so there
 282          * can be 8 bundles. But each bundle can have both an FMA and an ADD,
 283          * so a clause can have up to 16 bi_instructions. Whether bundles or
 284          * instructions are used depends on where in scheduling we are. */
 285
 286         unsigned instruction_count;
 287         unsigned bundle_count;
 288
 289         union {
 290                 bi_instruction *instructions[16];
 291                 bi_bundle bundles[8];
 292         };
 293
 294         /* For scoreboarding -- the clause ID (this is not globally unique!)
 295          * and its dependencies in terms of other clauses, computed during
 296          * scheduling and used when emitting code. Dependencies expressed as a
 297          * bitfield matching the hardware, except shifted by a clause (the
 298          * shift back to the ISA's off-by-one encoding is worked out when
 299          * emitting clauses) */
 300         unsigned scoreboard_id;
 301         uint8_t dependencies;
 302
 303         /* Back-to-back corresponds directly to the back-to-back bit. Branch
 304          * conditional corresponds to the branch conditional bit except that in
 305          * the emitted code it's always set if back-to-bit is, whereas we use
 306          * the actual value (without back-to-back so to speak) internally */
 307         bool back_to_back;
 308         bool branch_conditional;
 309
 310         /* Assigned data register */
 311         unsigned data_register;
 312
 313         /* Corresponds to the usual bit but shifted by a clause */
 314         bool data_register_write_barrier;
 315
 316         /* Constants read by this clause. ISA limit. */
 317         uint64_t constants[8];
 318         unsigned constant_count;
 319
 320         /* What type of high latency instruction is here, basically */
 321         unsigned clause_type;
 322 } bi_clause;
 323
 324 typedef struct bi_block {
 325         pan_block base; /* must be first */
 326
 327         /* If true, uses clauses; if false, uses instructions */
 328         bool scheduled;
 329         struct list_head clauses; /* list of bi_clause */
 330 } bi_block;
 331
 332 typedef struct {
 333        nir_shader *nir;
 334        gl_shader_stage stage;
 335        struct list_head blocks; /* list of bi_block */
 336        struct panfrost_sysvals sysvals;
 337        uint32_t quirks;
 338
 339        /* During NIR->BIR */
 340        nir_function_impl *impl;
 341        bi_block *current_block;
 342        unsigned block_name_count;
 343        bi_block *after_block;
 344        bi_block *break_block;
 345        bi_block *continue_block;
 346        bool emitted_atest;
 347
 348        /* For creating temporaries */
 349        unsigned temp_alloc;
 350
 351        /* Analysis results */
 352        bool has_liveness;
 353
 354        /* Stats for shader-db */
 355        unsigned instruction_count;
 356        unsigned loop_count;
 357 } bi_context;
 358
 359 static inline bi_instruction *
 360 bi_emit(bi_context *ctx, bi_instruction ins)
 361 {
 362         bi_instruction *u = rzalloc(ctx, bi_instruction);
 363         memcpy(u, &ins, sizeof(ins));
 364         list_addtail(&u->link, &ctx->current_block->base.instructions);
 365         return u;
 366 }
 367
 368 static inline void
 369 bi_remove_instruction(bi_instruction *ins)
 370 {
 371         list_del(&ins->link);
 372 }
 373
 374 /* So we can distinguish between SSA/reg/sentinel quickly */
 375 #define BIR_NO_ARG (0)
 376 #define BIR_IS_REG (1)
 377
 378 /* If high bits are set, instead of SSA/registers, we have specials indexed by
 379  * the low bits if necessary.
 380  *
 381  *  Fixed register: do not allocate register, do not collect $200.
 382  *  Uniform: access a uniform register given by low bits.
 383  *  Constant: access the specified constant
 384  *  Zero: special cased to avoid wasting a constant
 385  *  Passthrough: a bifrost_packed_src to passthrough T/T0/T1
 386  */
 387
 388 #define BIR_INDEX_REGISTER (1 << 31)
 389 #define BIR_INDEX_UNIFORM  (1 << 30)
 390 #define BIR_INDEX_CONSTANT (1 << 29)
 391 #define BIR_INDEX_ZERO     (1 << 28)
 392 #define BIR_INDEX_PASS     (1 << 27)
 393
 394 /* Keep me synced please so we can check src & BIR_SPECIAL */
 395
 396 #define BIR_SPECIAL        ((BIR_INDEX_REGISTER | BIR_INDEX_UNIFORM) | \
 397         (BIR_INDEX_CONSTANT | BIR_INDEX_ZERO | BIR_INDEX_PASS))
 398
 399 static inline unsigned
 400 bi_max_temp(bi_context *ctx)
 401 {
 402         unsigned alloc = MAX2(ctx->impl->reg_alloc, ctx->impl->ssa_alloc);
 403         return ((alloc + 2 + ctx->temp_alloc) << 1);
 404 }
 405
 406 static inline unsigned
 407 bi_make_temp(bi_context *ctx)
 408 {
 409         return (ctx->impl->ssa_alloc + 1 + ctx->temp_alloc++) << 1;
 410 }
 411
 412 static inline unsigned
 413 bi_make_temp_reg(bi_context *ctx)
 414 {
 415         return ((ctx->impl->reg_alloc + ctx->temp_alloc++) << 1) | BIR_IS_REG;
 416 }
 417
 418 static inline unsigned
 419 bir_ssa_index(nir_ssa_def *ssa)
 420 {
 421         /* Off-by-one ensures BIR_NO_ARG is skipped */
 422         return ((ssa->index + 1) << 1) | 0;
 423 }
 424
 425 static inline unsigned
 426 bir_src_index(nir_src *src)
 427 {
 428         if (src->is_ssa)
 429                 return bir_ssa_index(src->ssa);
 430         else {
 431                 assert(!src->reg.indirect);
 432                 return (src->reg.reg->index << 1) | BIR_IS_REG;
 433         }
 434 }
 435
 436 static inline unsigned
 437 bir_dest_index(nir_dest *dst)
 438 {
 439         if (dst->is_ssa)
 440                 return bir_ssa_index(&dst->ssa);
 441         else {
 442                 assert(!dst->reg.indirect);
 443                 return (dst->reg.reg->index << 1) | BIR_IS_REG;
 444         }
 445 }
 446
 447 /* Iterators for Bifrost IR */
 448
 449 #define bi_foreach_block(ctx, v) \
 450         list_for_each_entry(pan_block, v, &ctx->blocks, link)
 451
 452 #define bi_foreach_block_from(ctx, from, v) \
 453         list_for_each_entry_from(pan_block, v, from, &ctx->blocks, link)
 454
 455 #define bi_foreach_instr_in_block(block, v) \
 456         list_for_each_entry(bi_instruction, v, &(block)->base.instructions, link)
 457
 458 #define bi_foreach_instr_in_block_rev(block, v) \
 459         list_for_each_entry_rev(bi_instruction, v, &(block)->base.instructions, link)
 460
 461 #define bi_foreach_instr_in_block_safe(block, v) \
 462         list_for_each_entry_safe(bi_instruction, v, &(block)->base.instructions, link)
 463
 464 #define bi_foreach_instr_in_block_safe_rev(block, v) \
 465         list_for_each_entry_safe_rev(bi_instruction, v, &(block)->base.instructions, link)
 466
 467 #define bi_foreach_instr_in_block_from(block, v, from) \
 468         list_for_each_entry_from(bi_instruction, v, from, &(block)->base.instructions, link)
 469
 470 #define bi_foreach_instr_in_block_from_rev(block, v, from) \
 471         list_for_each_entry_from_rev(bi_instruction, v, from, &(block)->base.instructions, link)
 472
 473 #define bi_foreach_clause_in_block(block, v) \
 474         list_for_each_entry(bi_clause, v, &(block)->clauses, link)
 475
 476 #define bi_foreach_instr_global(ctx, v) \
 477         bi_foreach_block(ctx, v_block) \
 478                 bi_foreach_instr_in_block((bi_block *) v_block, v)
 479
 480 #define bi_foreach_instr_global_safe(ctx, v) \
 481         bi_foreach_block(ctx, v_block) \
 482                 bi_foreach_instr_in_block_safe((bi_block *) v_block, v)
 483
 484 /* Based on set_foreach, expanded with automatic type casts */
 485
 486 #define bi_foreach_predecessor(blk, v) \
 487         struct set_entry *_entry_##v; \
 488         bi_block *v; \
 489         for (_entry_##v = _mesa_set_next_entry(blk->base.predecessors, NULL), \
 490                 v = (bi_block *) (_entry_##v ? _entry_##v->key : NULL);  \
 491                 _entry_##v != NULL; \
 492                 _entry_##v = _mesa_set_next_entry(blk->base.predecessors, _entry_##v), \
 493                 v = (bi_block *) (_entry_##v ? _entry_##v->key : NULL))
 494
 495 #define bi_foreach_src(ins, v) \
 496         for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
 497
 498 static inline bi_instruction *
 499 bi_prev_op(bi_instruction *ins)
 500 {
 501         return list_last_entry(&(ins->link), bi_instruction, link);
 502 }
 503
 504 static inline bi_instruction *
 505 bi_next_op(bi_instruction *ins)
 506 {
 507         return list_first_entry(&(ins->link), bi_instruction, link);
 508 }
 509
 510 static inline pan_block *
 511 pan_next_block(pan_block *block)
 512 {
 513         return list_first_entry(&(block->link), pan_block, link);
 514 }
 515
 516 /* BIR manipulation */
 517
 518 bool bi_has_outmod(bi_instruction *ins);
 519 bool bi_has_source_mods(bi_instruction *ins);
 520 bool bi_is_src_swizzled(bi_instruction *ins, unsigned s);
 521 bool bi_has_arg(bi_instruction *ins, unsigned arg);
 522 uint16_t bi_from_bytemask(uint16_t bytemask, unsigned bytes);
 523 unsigned bi_get_component_count(bi_instruction *ins);
 524 uint16_t bi_bytemask_of_read_components(bi_instruction *ins, unsigned node);
 525
 526 /* BIR passes */
 527
 528 bool bi_opt_dead_code_eliminate(bi_context *ctx, bi_block *block);
 529 void bi_schedule(bi_context *ctx);
 530 void bi_register_allocate(bi_context *ctx);
 531
 532 /* Liveness */
 533
 534 void bi_compute_liveness(bi_context *ctx);
 535 void bi_liveness_ins_update(uint16_t *live, bi_instruction *ins, unsigned max);
 536 void bi_invalidate_liveness(bi_context *ctx);
 537 bool bi_is_live_after(bi_context *ctx, bi_block *block, bi_instruction *start, int src);
 538
 539 /* Code emit */
 540
 541 void bi_pack(bi_context *ctx, struct util_dynarray *emission);
 542
 543 #endif