2 * Copyright (C) 2020 Collabora Ltd.
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:
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
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
23 * Authors (Collabora):
24 * Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
27 #ifndef __BIFROST_COMPILER_H
28 #define __BIFROST_COMPILER_H
31 #include "compiler/nir/nir.h"
32 #include "panfrost/util/pan_ir.h"
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.
45 * Please keep this list alphabetized. Please use a dictionary if you
46 * don't know how to do that.
73 BI_SPECIAL
, /* _FAST, _TABLE on supported GPUs */
80 /* Properties of a class... */
81 extern unsigned bi_class_props
[BI_NUM_CLASSES
];
83 /* abs/neg/outmod valid for a float op */
84 #define BI_MODS (1 << 0)
86 /* Generic enough that little class-specific information is required. In other
87 * words, it acts as a "normal" ALU op, even if the encoding ends up being
88 * irregular enough to warrant a separate class */
89 #define BI_GENERIC (1 << 1)
91 /* Accepts a bifrost_roundmode */
92 #define BI_ROUNDMODE (1 << 2)
94 /* Can be scheduled to FMA */
95 #define BI_SCHED_FMA (1 << 3)
97 /* Can be scheduled to ADD */
98 #define BI_SCHED_ADD (1 << 4)
100 /* Most ALU ops can do either, actually */
101 #define BI_SCHED_ALL (BI_SCHED_FMA | BI_SCHED_ADD)
103 /* Along with setting BI_SCHED_ADD, eats up the entire cycle, so FMA must be
104 * nopped out. Used for _FAST operations. */
105 #define BI_SCHED_SLOW (1 << 5)
107 /* Swizzling allowed for the 8/16-bit source */
108 #define BI_SWIZZLABLE (1 << 6)
110 /* For scheduling purposes this is a high latency instruction and must be at
111 * the end of a clause. Implies ADD */
112 #define BI_SCHED_HI_LATENCY (1 << 7)
114 /* Intrinsic is vectorized and should read 4 components in the first source
115 * regardless of writemask */
116 #define BI_VECTOR (1 << 8)
118 /* Use a data register for src0/dest respectively, bypassing the usual
119 * register accessor. Mutually exclusive. */
120 #define BI_DATA_REG_SRC (1 << 9)
121 #define BI_DATA_REG_DEST (1 << 10)
123 /* It can't get any worse than csel4... can it? */
124 #define BIR_SRC_COUNT 4
127 struct bi_load_vary
{
128 enum bifrost_interp_mode interp_mode
;
133 /* BI_BRANCH encoding the details of the branch itself as well as a pointer to
134 * the target. We forward declare bi_block since this is mildly circular (not
135 * strictly, but this order of the file makes more sense I think)
137 * We define our own enum of conditions since the conditions in the hardware
138 * packed in crazy ways that would make manipulation unweildly (meaning changes
139 * based on port swapping, etc), so we defer dealing with that until emit time.
140 * Likewise, we expose NIR types instead of the crazy branch types, although
141 * the restrictions do eventually apply of course. */
156 /* Types are specified in src_types and must be compatible (either both
157 * int, or both float, 16/32, and same size or 32/16 if float. Types
158 * ignored if BI_COND_ALWAYS is set for an unconditional branch. */
161 struct bi_block
*target
;
164 /* Opcodes within a class */
177 BI_ROUND_MODE
, /* use round mode */
178 BI_ROUND_ROUND
/* i.e.: fround() */
193 struct list_head link
; /* Must be first */
196 /* Indices, see bir_ssa_index etc. Note zero is special cased
197 * to "no argument" */
199 unsigned src
[BIR_SRC_COUNT
];
201 /* If one of the sources has BIR_INDEX_CONSTANT */
209 /* Floating-point modifiers, type/class permitting. If not
210 * allowed for the type/class, these are ignored. */
211 enum bifrost_outmod outmod
;
212 bool src_abs
[BIR_SRC_COUNT
];
213 bool src_neg
[BIR_SRC_COUNT
];
215 /* Round mode (requires BI_ROUNDMODE) */
216 enum bifrost_roundmode roundmode
;
218 /* Writemask (bit for each affected byte). This is quite restricted --
219 * ALU ops can only write to a single channel (exception: <32 in which
220 * you can write to 32/N contiguous aligned channels). Load/store can
221 * only write to all channels at once, in a sense. But it's still
222 * better to use this generic form than have synthetic ops flying
223 * about, since we're not essentially vector for RA purposes. */
226 /* Destination type. Usually the type of the instruction
227 * itself, but if sources and destination have different
228 * types, the type of the destination wins (so f2i would be
229 * int). Zero if there is no destination. Bitsize included */
230 nir_alu_type dest_type
;
232 /* Source types if required by the class */
233 nir_alu_type src_types
[BIR_SRC_COUNT
];
235 /* If the source type is 8-bit or 16-bit such that SIMD is possible,
236 * and the class has BI_SWIZZLABLE, this is a swizzle in the usual
237 * sense. On non-SIMD instructions, it can be used for component
238 * selection, so we don't have to special case extraction. */
239 uint8_t swizzle
[BIR_SRC_COUNT
][NIR_MAX_VEC_COMPONENTS
];
241 /* A class-specific op from which the actual opcode can be derived
242 * (along with the above information) */
245 enum bi_minmax_op minmax
;
246 enum bi_bitwise_op bitwise
;
247 enum bi_round_op round
;
248 enum bi_special_op special
;
249 enum bi_cond compare
;
252 /* Union for class-specific information */
254 enum bifrost_minmax_mode minmax
;
255 struct bi_load_vary load_vary
;
256 struct bi_branch branch
;
258 /* For CSEL, the comparison op. BI_COND_ALWAYS doesn't make
259 * sense here but you can always just use a move for that */
260 enum bi_cond csel_cond
;
262 /* For BLEND -- the location 0-7 */
263 unsigned blend_location
;
265 /* For STORE, STORE_VAR -- channel count */
266 unsigned store_channels
;
270 /* Scheduling takes place in two steps. Step 1 groups instructions within a
271 * block into distinct clauses (bi_clause). Step 2 schedules instructions
272 * within a clause into FMA/ADD pairs (bi_bundle).
274 * A bi_bundle contains two paired instruction pointers. If a slot is unfilled,
275 * leave it NULL; the emitter will fill in a nop.
284 struct list_head link
;
286 /* A clause can have 8 instructions in bundled FMA/ADD sense, so there
287 * can be 8 bundles. But each bundle can have both an FMA and an ADD,
288 * so a clause can have up to 16 bi_instructions. Whether bundles or
289 * instructions are used depends on where in scheduling we are. */
291 unsigned instruction_count
;
292 unsigned bundle_count
;
295 bi_instruction
*instructions
[16];
296 bi_bundle bundles
[8];
299 /* For scoreboarding -- the clause ID (this is not globally unique!)
300 * and its dependencies in terms of other clauses, computed during
301 * scheduling and used when emitting code. Dependencies expressed as a
302 * bitfield matching the hardware, except shifted by a clause (the
303 * shift back to the ISA's off-by-one encoding is worked out when
304 * emitting clauses) */
305 unsigned scoreboard_id
;
306 uint8_t dependencies
;
308 /* Back-to-back corresponds directly to the back-to-back bit. Branch
309 * conditional corresponds to the branch conditional bit except that in
310 * the emitted code it's always set if back-to-bit is, whereas we use
311 * the actual value (without back-to-back so to speak) internally */
313 bool branch_conditional
;
315 /* Assigned data register */
316 unsigned data_register
;
318 /* Corresponds to the usual bit but shifted by a clause */
319 bool data_register_write_barrier
;
321 /* Constants read by this clause. ISA limit. */
322 uint64_t constants
[8];
323 unsigned constant_count
;
325 /* What type of high latency instruction is here, basically */
326 unsigned clause_type
;
329 typedef struct bi_block
{
330 pan_block base
; /* must be first */
332 /* If true, uses clauses; if false, uses instructions */
334 struct list_head clauses
; /* list of bi_clause */
339 gl_shader_stage stage
;
340 struct list_head blocks
; /* list of bi_block */
341 struct panfrost_sysvals sysvals
;
344 /* During NIR->BIR */
345 nir_function_impl
*impl
;
346 bi_block
*current_block
;
347 unsigned block_name_count
;
348 bi_block
*after_block
;
349 bi_block
*break_block
;
350 bi_block
*continue_block
;
353 /* For creating temporaries */
356 /* Analysis results */
359 /* Stats for shader-db */
360 unsigned instruction_count
;
364 static inline bi_instruction
*
365 bi_emit(bi_context
*ctx
, bi_instruction ins
)
367 bi_instruction
*u
= rzalloc(ctx
, bi_instruction
);
368 memcpy(u
, &ins
, sizeof(ins
));
369 list_addtail(&u
->link
, &ctx
->current_block
->base
.instructions
);
373 static inline bi_instruction
*
374 bi_emit_before(bi_context
*ctx
, bi_instruction
*tag
, bi_instruction ins
)
376 bi_instruction
*u
= rzalloc(ctx
, bi_instruction
);
377 memcpy(u
, &ins
, sizeof(ins
));
378 list_addtail(&u
->link
, &tag
->link
);
383 bi_remove_instruction(bi_instruction
*ins
)
385 list_del(&ins
->link
);
388 /* So we can distinguish between SSA/reg/sentinel quickly */
389 #define BIR_NO_ARG (0)
390 #define BIR_IS_REG (1)
392 /* If high bits are set, instead of SSA/registers, we have specials indexed by
393 * the low bits if necessary.
395 * Fixed register: do not allocate register, do not collect $200.
396 * Uniform: access a uniform register given by low bits.
397 * Constant: access the specified constant (specifies a bit offset / shift)
398 * Zero: special cased to avoid wasting a constant
399 * Passthrough: a bifrost_packed_src to passthrough T/T0/T1
402 #define BIR_INDEX_REGISTER (1 << 31)
403 #define BIR_INDEX_UNIFORM (1 << 30)
404 #define BIR_INDEX_CONSTANT (1 << 29)
405 #define BIR_INDEX_ZERO (1 << 28)
406 #define BIR_INDEX_PASS (1 << 27)
408 /* Keep me synced please so we can check src & BIR_SPECIAL */
410 #define BIR_SPECIAL ((BIR_INDEX_REGISTER | BIR_INDEX_UNIFORM) | \
411 (BIR_INDEX_CONSTANT | BIR_INDEX_ZERO | BIR_INDEX_PASS))
413 static inline unsigned
414 bi_max_temp(bi_context
*ctx
)
416 unsigned alloc
= MAX2(ctx
->impl
->reg_alloc
, ctx
->impl
->ssa_alloc
);
417 return ((alloc
+ 2 + ctx
->temp_alloc
) << 1);
420 static inline unsigned
421 bi_make_temp(bi_context
*ctx
)
423 return (ctx
->impl
->ssa_alloc
+ 1 + ctx
->temp_alloc
++) << 1;
426 static inline unsigned
427 bi_make_temp_reg(bi_context
*ctx
)
429 return ((ctx
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | BIR_IS_REG
;
432 static inline unsigned
433 bir_ssa_index(nir_ssa_def
*ssa
)
435 /* Off-by-one ensures BIR_NO_ARG is skipped */
436 return ((ssa
->index
+ 1) << 1) | 0;
439 static inline unsigned
440 bir_src_index(nir_src
*src
)
443 return bir_ssa_index(src
->ssa
);
445 assert(!src
->reg
.indirect
);
446 return (src
->reg
.reg
->index
<< 1) | BIR_IS_REG
;
450 static inline unsigned
451 bir_dest_index(nir_dest
*dst
)
454 return bir_ssa_index(&dst
->ssa
);
456 assert(!dst
->reg
.indirect
);
457 return (dst
->reg
.reg
->index
<< 1) | BIR_IS_REG
;
461 /* Iterators for Bifrost IR */
463 #define bi_foreach_block(ctx, v) \
464 list_for_each_entry(pan_block, v, &ctx->blocks, link)
466 #define bi_foreach_block_from(ctx, from, v) \
467 list_for_each_entry_from(pan_block, v, from, &ctx->blocks, link)
469 #define bi_foreach_instr_in_block(block, v) \
470 list_for_each_entry(bi_instruction, v, &(block)->base.instructions, link)
472 #define bi_foreach_instr_in_block_rev(block, v) \
473 list_for_each_entry_rev(bi_instruction, v, &(block)->base.instructions, link)
475 #define bi_foreach_instr_in_block_safe(block, v) \
476 list_for_each_entry_safe(bi_instruction, v, &(block)->base.instructions, link)
478 #define bi_foreach_instr_in_block_safe_rev(block, v) \
479 list_for_each_entry_safe_rev(bi_instruction, v, &(block)->base.instructions, link)
481 #define bi_foreach_instr_in_block_from(block, v, from) \
482 list_for_each_entry_from(bi_instruction, v, from, &(block)->base.instructions, link)
484 #define bi_foreach_instr_in_block_from_rev(block, v, from) \
485 list_for_each_entry_from_rev(bi_instruction, v, from, &(block)->base.instructions, link)
487 #define bi_foreach_clause_in_block(block, v) \
488 list_for_each_entry(bi_clause, v, &(block)->clauses, link)
490 #define bi_foreach_instr_global(ctx, v) \
491 bi_foreach_block(ctx, v_block) \
492 bi_foreach_instr_in_block((bi_block *) v_block, v)
494 #define bi_foreach_instr_global_safe(ctx, v) \
495 bi_foreach_block(ctx, v_block) \
496 bi_foreach_instr_in_block_safe((bi_block *) v_block, v)
498 /* Based on set_foreach, expanded with automatic type casts */
500 #define bi_foreach_predecessor(blk, v) \
501 struct set_entry *_entry_##v; \
503 for (_entry_##v = _mesa_set_next_entry(blk->base.predecessors, NULL), \
504 v = (bi_block *) (_entry_##v ? _entry_##v->key : NULL); \
505 _entry_##v != NULL; \
506 _entry_##v = _mesa_set_next_entry(blk->base.predecessors, _entry_##v), \
507 v = (bi_block *) (_entry_##v ? _entry_##v->key : NULL))
509 #define bi_foreach_src(ins, v) \
510 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
512 static inline bi_instruction
*
513 bi_prev_op(bi_instruction
*ins
)
515 return list_last_entry(&(ins
->link
), bi_instruction
, link
);
518 static inline bi_instruction
*
519 bi_next_op(bi_instruction
*ins
)
521 return list_first_entry(&(ins
->link
), bi_instruction
, link
);
524 static inline pan_block
*
525 pan_next_block(pan_block
*block
)
527 return list_first_entry(&(block
->link
), pan_block
, link
);
530 /* BIR manipulation */
532 bool bi_has_outmod(bi_instruction
*ins
);
533 bool bi_has_source_mods(bi_instruction
*ins
);
534 bool bi_is_src_swizzled(bi_instruction
*ins
, unsigned s
);
535 bool bi_has_arg(bi_instruction
*ins
, unsigned arg
);
536 uint16_t bi_from_bytemask(uint16_t bytemask
, unsigned bytes
);
537 unsigned bi_get_component_count(bi_instruction
*ins
, unsigned s
);
538 unsigned bi_load32_components(bi_instruction
*ins
);
539 uint16_t bi_bytemask_of_read_components(bi_instruction
*ins
, unsigned node
);
540 uint64_t bi_get_immediate(bi_instruction
*ins
, unsigned index
);
544 bool bi_opt_dead_code_eliminate(bi_context
*ctx
, bi_block
*block
);
545 void bi_schedule(bi_context
*ctx
);
546 void bi_register_allocate(bi_context
*ctx
);
550 void bi_compute_liveness(bi_context
*ctx
);
551 void bi_liveness_ins_update(uint16_t *live
, bi_instruction
*ins
, unsigned max
);
552 void bi_invalidate_liveness(bi_context
*ctx
);
553 bool bi_is_live_after(bi_context
*ctx
, bi_block
*block
, bi_instruction
*start
, int src
);
557 void bi_pack(bi_context
*ctx
, struct util_dynarray
*emission
);