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
26 #define RETURN_PACKED(str) { \
28 memcpy(&temp, &str, sizeof(str)); \
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) */
37 bi_pack_header(bi_clause
*clause
, bi_clause
*next
, bool is_fragment
)
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,
53 memcpy(&u
, &header
, sizeof(header
));
57 /* Represents the assignment of ports for a given bundle */
60 /* Register to assign to each port */
63 /* Read ports can be disabled */
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
;
70 /* Should we read with port 3? */
73 /* Packed uniform/constant */
74 uint8_t uniform_constant
;
76 /* Whether writes are actually for the last instruction */
77 bool first_instruction
;
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. */
87 bi_lookup_constant(bi_clause
*clause
, uint64_t cons
)
89 for (unsigned i
= 0; i
< clause
->constant_count
; ++i
) {
90 /* Only check top 60-bits since that's what's actually embedded
91 * in the clause, the bottom 4-bits are bundle-inline */
93 if ((cons
>> 4) == (clause
->constants
[i
] >> 4))
97 unreachable("Invalid constant accessed");
101 bi_constant_field(unsigned idx
)
105 const unsigned values
[] = {
109 return values
[idx
] << 4;
113 bi_assign_uniform_constant_single(
114 struct bi_registers
*regs
,
116 bi_instruction
*ins
, bool assigned
, bool fast_zero
)
121 bi_foreach_src(ins
, s
) {
122 if (ins
->src
[s
] & BIR_INDEX_CONSTANT
) {
123 /* TODO: lo/hi matching? */
124 uint64_t cons
= ins
->constant
.u64
;
125 unsigned idx
= bi_lookup_constant(clause
, cons
);
126 unsigned f
= bi_constant_field(idx
) | (cons
& 0xF);
128 if (assigned
&& regs
->uniform_constant
!= f
)
129 unreachable("Mismatched uniform/const field: imm");
131 regs
->uniform_constant
= f
;
132 ins
->src
[s
] = BIR_INDEX_PASS
| BIFROST_SRC_CONST_LO
;
134 } else if (ins
->src
[s
] & BIR_INDEX_ZERO
&& (ins
->type
== BI_LOAD_UNIFORM
|| ins
->type
== BI_LOAD_VAR
)) {
135 /* XXX: HACK UNTIL WE HAVE HI MATCHING DUE TO OVERFLOW XXX */
136 ins
->src
[s
] = BIR_INDEX_PASS
| BIFROST_SRC_CONST_HI
;
137 } else if (ins
->src
[s
] & BIR_INDEX_ZERO
&& !fast_zero
) {
138 /* FMAs have a fast zero port, ADD needs to use the
139 * uniform/const port's special 0 mode handled here */
142 if (assigned
&& regs
->uniform_constant
!= f
)
143 unreachable("Mismatched uniform/const field: 0");
145 regs
->uniform_constant
= f
;
146 ins
->src
[s
] = BIR_INDEX_PASS
| BIFROST_SRC_CONST_LO
;
148 } else if (s
& BIR_INDEX_UNIFORM
) {
149 unreachable("Push uniforms not implemented yet");
157 bi_assign_uniform_constant(
159 struct bi_registers
*regs
,
163 bi_assign_uniform_constant_single(regs
, clause
, bundle
.fma
, false, true);
165 bi_assign_uniform_constant_single(regs
, clause
, bundle
.add
, assigned
, false);
168 /* Assigns a port for reading, before anything is written */
171 bi_assign_port_read(struct bi_registers
*regs
, unsigned src
)
173 /* We only assign for registers */
174 if (!(src
& BIR_INDEX_REGISTER
))
177 unsigned reg
= src
& ~BIR_INDEX_REGISTER
;
179 /* Check if we already assigned the port */
180 for (unsigned i
= 0; i
<= 1; ++i
) {
181 if (regs
->port
[i
] == reg
&& regs
->enabled
[i
])
185 if (regs
->port
[3] == reg
&& regs
->read_port3
)
190 for (unsigned i
= 0; i
<= 1; ++i
) {
191 if (!regs
->enabled
[i
]) {
193 regs
->enabled
[i
] = true;
198 if (!regs
->read_port3
) {
200 regs
->read_port3
= true;
204 static struct bi_registers
205 bi_assign_ports(bi_bundle now
, bi_bundle prev
)
207 struct bi_registers regs
= { 0 };
209 /* We assign ports for the main register mechanism. Special ops
210 * use the data registers, which has its own mechanism entirely
211 * and thus gets skipped over here. */
213 unsigned read_dreg
= now
.add
&&
214 bi_class_props
[now
.add
->type
] & BI_DATA_REG_SRC
;
216 unsigned write_dreg
= prev
.add
&&
217 bi_class_props
[prev
.add
->type
] & BI_DATA_REG_DEST
;
219 /* First, assign reads */
222 bi_foreach_src(now
.fma
, src
)
223 bi_assign_port_read(®s
, now
.fma
->src
[src
]);
226 bi_foreach_src(now
.add
, src
) {
227 if (!(src
== 0 && read_dreg
))
228 bi_assign_port_read(®s
, now
.add
->src
[src
]);
232 /* Next, assign writes */
234 if (prev
.fma
&& prev
.fma
->dest
& BIR_INDEX_REGISTER
) {
235 regs
.port
[2] = prev
.fma
->dest
& ~BIR_INDEX_REGISTER
;
236 regs
.write_fma
= true;
239 if (prev
.add
&& prev
.add
->dest
& BIR_INDEX_REGISTER
&& !write_dreg
) {
240 unsigned r
= prev
.add
->dest
& ~BIR_INDEX_REGISTER
;
242 if (regs
.write_fma
) {
243 /* Scheduler constraint: cannot read 3 and write 2 */
244 assert(!regs
.read_port3
);
250 regs
.write_add
= true;
253 /* Finally, ensure port 1 > port 0 for the 63-x trick to function */
255 if (regs
.enabled
[0] && regs
.enabled
[1] && regs
.port
[1] < regs
.port
[0]) {
256 unsigned temp
= regs
.port
[0];
257 regs
.port
[0] = regs
.port
[1];
264 /* Determines the register control field, ignoring the first? flag */
266 static enum bifrost_reg_control
267 bi_pack_register_ctrl_lo(struct bi_registers r
)
271 assert(!r
.read_port3
);
272 return BIFROST_WRITE_ADD_P2_FMA_P3
;
275 return BIFROST_WRITE_FMA_P2_READ_P3
;
277 return BIFROST_WRITE_FMA_P2
;
279 } else if (r
.write_add
) {
281 return BIFROST_WRITE_ADD_P2_READ_P3
;
283 return BIFROST_WRITE_ADD_P2
;
284 } else if (r
.read_port3
)
285 return BIFROST_READ_P3
;
287 return BIFROST_REG_NONE
;
290 /* Ditto but account for the first? flag this time */
292 static enum bifrost_reg_control
293 bi_pack_register_ctrl(struct bi_registers r
)
295 enum bifrost_reg_control ctrl
= bi_pack_register_ctrl_lo(r
);
297 if (r
.first_instruction
) {
298 if (ctrl
== BIFROST_REG_NONE
)
299 ctrl
= BIFROST_FIRST_NONE
;
301 ctrl
|= BIFROST_FIRST_NONE
;
308 bi_pack_registers(struct bi_registers regs
)
310 enum bifrost_reg_control ctrl
= bi_pack_register_ctrl(regs
);
311 struct bifrost_regs s
;
314 if (regs
.enabled
[1]) {
315 /* Gotta save that bit!~ Required by the 63-x trick */
316 assert(regs
.port
[1] > regs
.port
[0]);
317 assert(regs
.enabled
[0]);
319 /* Do the 63-x trick, see docs/disasm */
320 if (regs
.port
[0] > 31) {
321 regs
.port
[0] = 63 - regs
.port
[0];
322 regs
.port
[1] = 63 - regs
.port
[1];
325 assert(regs
.port
[0] <= 31);
326 assert(regs
.port
[1] <= 63);
329 s
.reg1
= regs
.port
[1];
330 s
.reg0
= regs
.port
[0];
332 /* Port 1 disabled, so set to zero and use port 1 for ctrl */
335 if (regs
.enabled
[0]) {
336 /* Bit 0 upper bit of port 0 */
337 s
.reg1
|= (regs
.port
[0] >> 5);
339 /* Rest of port 0 in usual spot */
340 s
.reg0
= (regs
.port
[0] & 0b11111);
342 /* Bit 1 set if port 0 also disabled */
347 s
.reg3
= regs
.port
[3];
348 s
.reg2
= regs
.port
[2];
349 s
.uniform_const
= regs
.uniform_constant
;
351 memcpy(&packed
, &s
, sizeof(s
));
356 bi_write_data_register(bi_clause
*clause
, bi_instruction
*ins
)
358 assert(ins
->dest
& BIR_INDEX_REGISTER
);
359 unsigned reg
= ins
->dest
& ~BIR_INDEX_REGISTER
;
361 clause
->data_register
= reg
;
364 static enum bifrost_packed_src
365 bi_get_src_reg_port(struct bi_registers
*regs
, unsigned src
)
367 unsigned reg
= src
& ~BIR_INDEX_REGISTER
;
369 if (regs
->port
[0] == reg
&& regs
->enabled
[0])
370 return BIFROST_SRC_PORT0
;
371 else if (regs
->port
[1] == reg
&& regs
->enabled
[1])
372 return BIFROST_SRC_PORT1
;
373 else if (regs
->port
[3] == reg
&& regs
->read_port3
)
374 return BIFROST_SRC_PORT3
;
376 unreachable("Tried to access register with no port");
379 static enum bifrost_packed_src
380 bi_get_src(bi_instruction
*ins
, struct bi_registers
*regs
, unsigned s
, bool is_fma
)
382 unsigned src
= ins
->src
[s
];
384 if (src
& BIR_INDEX_REGISTER
)
385 return bi_get_src_reg_port(regs
, src
);
386 else if (src
& BIR_INDEX_ZERO
&& is_fma
)
387 return BIFROST_SRC_STAGE
;
388 else if (src
& BIR_INDEX_PASS
)
389 return src
& ~BIR_INDEX_PASS
;
391 unreachable("Unknown src");
395 bi_pack_fma_fma(bi_instruction
*ins
, struct bi_registers
*regs
)
397 /* (-a)(-b) = ab, so we only need one negate bit */
398 bool negate_mul
= ins
->src_neg
[0] ^ ins
->src_neg
[1];
400 struct bifrost_fma_fma pack
= {
401 .src0
= bi_get_src(ins
, regs
, 0, true),
402 .src1
= bi_get_src(ins
, regs
, 1, true),
403 .src2
= bi_get_src(ins
, regs
, 2, true),
404 .src0_abs
= ins
->src_abs
[0],
405 .src1_abs
= ins
->src_abs
[1],
406 .src2_abs
= ins
->src_abs
[2],
407 .src0_neg
= negate_mul
,
408 .src2_neg
= ins
->src_neg
[2],
409 .op
= BIFROST_FMA_OP_FMA
416 bi_pack_fma_add(bi_instruction
*ins
, struct bi_registers
*regs
)
418 /* TODO: fadd16 packing is a bit different */
419 assert(ins
->dest_type
== nir_type_float32
);
421 struct bifrost_fma_add pack
= {
422 .src0
= bi_get_src(ins
, regs
, 0, true),
423 .src1
= bi_get_src(ins
, regs
, 1, true),
424 .src0_abs
= ins
->src_abs
[0],
425 .src1_abs
= ins
->src_abs
[1],
426 .src0_neg
= ins
->src_neg
[0],
427 .src1_neg
= ins
->src_neg
[1],
429 .outmod
= ins
->outmod
,
430 .roundmode
= ins
->roundmode
,
431 .op
= BIFROST_FMA_OP_FADD32
438 bi_pack_fma_1src(bi_instruction
*ins
, struct bi_registers
*regs
, unsigned op
)
440 struct bifrost_fma_inst pack
= {
441 .src0
= bi_get_src(ins
, regs
, 0, true),
449 bi_pack_fma(bi_clause
*clause
, bi_bundle bundle
, struct bi_registers
*regs
)
452 return BIFROST_FMA_NOP
;
454 switch (bundle
.fma
->type
) {
456 return bi_pack_fma_add(bundle
.fma
, regs
);
461 return BIFROST_FMA_NOP
;
463 return bi_pack_fma_fma(bundle
.fma
, regs
);
467 return BIFROST_FMA_NOP
;
469 return bi_pack_fma_1src(bundle
.fma
, regs
, BIFROST_FMA_OP_MOV
);
474 return BIFROST_FMA_NOP
;
476 unreachable("Cannot encode class as FMA");
481 bi_pack_add_ld_vary(bi_clause
*clause
, bi_instruction
*ins
, struct bi_registers
*regs
)
483 unsigned size
= nir_alu_type_get_type_size(ins
->dest_type
);
484 assert(size
== 32 || size
== 16);
486 unsigned op
= (size
== 32) ?
487 BIFROST_ADD_OP_LD_VAR_32
:
488 BIFROST_ADD_OP_LD_VAR_16
;
490 unsigned cmask
= bi_from_bytemask(ins
->writemask
, size
/ 8);
491 unsigned channels
= util_bitcount(cmask
);
492 assert(cmask
== ((1 << channels
) - 1));
494 unsigned packed_addr
= 0;
496 if (ins
->src
[0] & BIR_INDEX_CONSTANT
) {
497 /* Direct uses address field directly */
498 packed_addr
= ins
->src
[0] & ~BIR_INDEX_CONSTANT
;
499 assert(packed_addr
< 0b1000);
501 /* Indirect gets an extra source */
502 packed_addr
= bi_get_src(ins
, regs
, 0, false) | 0b11000;
505 /* The destination is thrown in the data register */
506 assert(ins
->dest
& BIR_INDEX_REGISTER
);
507 clause
->data_register
= ins
->dest
& ~BIR_INDEX_REGISTER
;
509 assert(channels
>= 1 && channels
<= 4);
511 struct bifrost_ld_var pack
= {
512 .src0
= bi_get_src(ins
, regs
, 1, false),
514 .channels
= MALI_POSITIVE(channels
),
515 .interp_mode
= ins
->load_vary
.interp_mode
,
516 .reuse
= ins
->load_vary
.reuse
,
517 .flat
= ins
->load_vary
.flat
,
525 bi_pack_add_2src(bi_instruction
*ins
, struct bi_registers
*regs
, unsigned op
)
527 struct bifrost_add_2src pack
= {
528 .src0
= bi_get_src(ins
, regs
, 0, true),
529 .src1
= bi_get_src(ins
, regs
, 1, true),
537 bi_pack_add_ld_ubo(bi_clause
*clause
, bi_instruction
*ins
, struct bi_registers
*regs
)
539 unsigned components
= bi_load32_components(ins
);
541 const unsigned ops
[4] = {
542 BIFROST_ADD_OP_LD_UBO_1
,
543 BIFROST_ADD_OP_LD_UBO_2
,
544 BIFROST_ADD_OP_LD_UBO_3
,
545 BIFROST_ADD_OP_LD_UBO_4
548 bi_write_data_register(clause
, ins
);
549 return bi_pack_add_2src(ins
, regs
, ops
[components
- 1]);
553 bi_pack_add_atest(bi_clause
*clause
, bi_instruction
*ins
, struct bi_registers
*regs
)
556 assert(ins
->src_types
[1] == nir_type_float32
);
558 struct bifrost_add_atest pack
= {
559 .src0
= bi_get_src(ins
, regs
, 0, false),
560 .src1
= bi_get_src(ins
, regs
, 1, false),
561 .component
= 1, /* Set for fp32 */
562 .op
= BIFROST_ADD_OP_ATEST
,
565 /* Despite *also* writing with the usual mechanism... quirky and
566 * perhaps unnecessary, but let's match the blob */
567 clause
->data_register
= ins
->dest
& ~BIR_INDEX_REGISTER
;
573 bi_pack_add_blend(bi_instruction
*ins
, struct bi_registers
*regs
)
575 struct bifrost_add_inst pack
= {
576 .src0
= bi_get_src(ins
, regs
, 0, false),
577 .op
= BIFROST_ADD_OP_BLEND
580 /* TODO: Pack location in uniform_const */
581 assert(ins
->blend_location
== 0);
587 bi_pack_add(bi_clause
*clause
, bi_bundle bundle
, struct bi_registers
*regs
)
590 return BIFROST_ADD_NOP
;
592 switch (bundle
.add
->type
) {
594 return BIFROST_ADD_NOP
;
596 return bi_pack_add_atest(clause
, bundle
.add
, regs
);
599 return BIFROST_ADD_NOP
;
601 return bi_pack_add_blend(bundle
.add
, regs
);
608 return BIFROST_ADD_NOP
;
609 case BI_LOAD_UNIFORM
:
610 return bi_pack_add_ld_ubo(clause
, bundle
.add
, regs
);
612 return BIFROST_ADD_NOP
;
614 return bi_pack_add_ld_vary(clause
, bundle
.add
, regs
);
615 case BI_LOAD_VAR_ADDRESS
:
626 return BIFROST_ADD_NOP
;
628 unreachable("Cannot encode class as ADD");
632 struct bi_packed_bundle
{
637 static struct bi_packed_bundle
638 bi_pack_bundle(bi_clause
*clause
, bi_bundle bundle
, bi_bundle prev
, bool first_bundle
)
640 struct bi_registers regs
= bi_assign_ports(bundle
, prev
);
641 bi_assign_uniform_constant(clause
, ®s
, bundle
);
642 regs
.first_instruction
= first_bundle
;
644 uint64_t reg
= bi_pack_registers(regs
);
645 uint64_t fma
= bi_pack_fma(clause
, bundle
, ®s
);
646 uint64_t add
= bi_pack_add(clause
, bundle
, ®s
);
648 struct bi_packed_bundle packed
= {
649 .lo
= reg
| (fma
<< 35) | ((add
& 0b111111) << 58),
656 /* Packs the next two constants as a dedicated constant quadword at the end of
657 * the clause, returning the number packed. */
660 bi_pack_constants(bi_context
*ctx
, bi_clause
*clause
,
662 struct util_dynarray
*emission
)
664 /* After these two, are we done? Determines tag */
665 bool done
= clause
->constant_count
<= (index
+ 2);
666 bool only
= clause
->constant_count
<= (index
+ 1);
669 assert(index
== 0 && clause
->bundle_count
== 1);
671 struct bifrost_fmt_constant quad
= {
673 .tag
= done
? BIFROST_FMTC_FINAL
: BIFROST_FMTC_CONSTANTS
,
674 .imm_1
= clause
->constants
[index
+ 0] >> 4,
675 .imm_2
= only
? 0 : clause
->constants
[index
+ 1] >> 4
678 /* XXX: On G71, Connor observed that the difference of the top 4 bits
679 * of the second constant with the first must be less than 8, otherwise
680 * we have to swap them. I am not able to reproduce this on G52,
681 * further investigation needed. Possibly an errata. XXX */
683 util_dynarray_append(emission
, struct bifrost_fmt_constant
, quad
);
689 bi_pack_clause(bi_context
*ctx
, bi_clause
*clause
, bi_clause
*next
,
690 struct util_dynarray
*emission
)
692 struct bi_packed_bundle ins_1
= bi_pack_bundle(clause
, clause
->bundles
[0], clause
->bundles
[0], true);
693 assert(clause
->bundle_count
== 1);
695 /* Used to decide if we elide writes */
696 bool is_fragment
= ctx
->stage
== MESA_SHADER_FRAGMENT
;
698 /* State for packing constants throughout */
699 unsigned constant_index
= 0;
701 struct bifrost_fmt1 quad_1
= {
702 .tag
= clause
->constant_count
? BIFROST_FMT1_CONSTANTS
: BIFROST_FMT1_FINAL
,
703 .header
= bi_pack_header(clause
, next
, is_fragment
),
705 .ins_2
= ins_1
.hi
& ((1 << 11) - 1),
706 .ins_0
= (ins_1
.hi
>> 11) & 0b111,
709 util_dynarray_append(emission
, struct bifrost_fmt1
, quad_1
);
711 /* Pack the remaining constants */
713 while (constant_index
< clause
->constant_count
) {
714 constant_index
+= bi_pack_constants(ctx
, clause
,
715 constant_index
, emission
);
720 bi_next_clause(bi_context
*ctx
, pan_block
*block
, bi_clause
*clause
)
722 /* Try the next clause in this block */
723 if (clause
->link
.next
!= &((bi_block
*) block
)->clauses
)
724 return list_first_entry(&(clause
->link
), bi_clause
, link
);
726 /* Try the next block, or the one after that if it's empty, etc .*/
727 pan_block
*next_block
= pan_next_block(block
);
729 bi_foreach_block_from(ctx
, next_block
, block
) {
730 bi_block
*blk
= (bi_block
*) block
;
732 if (!list_is_empty(&blk
->clauses
))
733 return list_first_entry(&(blk
->clauses
), bi_clause
, link
);
740 bi_pack(bi_context
*ctx
, struct util_dynarray
*emission
)
742 util_dynarray_init(emission
, NULL
);
744 bi_foreach_block(ctx
, _block
) {
745 bi_block
*block
= (bi_block
*) _block
;
747 bi_foreach_clause_in_block(block
, clause
) {
748 bi_clause
*next
= bi_next_clause(ctx
, _block
, clause
);
749 bi_pack_clause(ctx
, clause
, next
, emission
);