2 * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
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
24 * Rob Clark <robclark@freedesktop.org>
29 #include "util/u_string.h"
30 #include "util/u_memory.h"
31 #include "util/u_math.h"
33 #include "ir3_compiler.h"
34 #include "ir3_shader.h"
37 #include "instr-a3xx.h"
39 #include "ir3_context.h"
42 static struct ir3_instruction
*
43 create_indirect_load(struct ir3_context
*ctx
, unsigned arrsz
, int n
,
44 struct ir3_instruction
*address
, struct ir3_instruction
*collect
)
46 struct ir3_block
*block
= ctx
->block
;
47 struct ir3_instruction
*mov
;
48 struct ir3_register
*src
;
50 mov
= ir3_instr_create(block
, OPC_MOV
);
51 mov
->cat1
.src_type
= TYPE_U32
;
52 mov
->cat1
.dst_type
= TYPE_U32
;
53 ir3_reg_create(mov
, 0, 0);
54 src
= ir3_reg_create(mov
, 0, IR3_REG_SSA
| IR3_REG_RELATIV
);
57 src
->array
.offset
= n
;
59 ir3_instr_set_address(mov
, address
);
64 static struct ir3_instruction
*
65 create_input_compmask(struct ir3_context
*ctx
, unsigned n
, unsigned compmask
)
67 struct ir3_instruction
*in
;
69 in
= ir3_instr_create(ctx
->in_block
, OPC_META_INPUT
);
70 in
->inout
.block
= ctx
->in_block
;
71 ir3_reg_create(in
, n
, 0);
73 in
->regs
[0]->wrmask
= compmask
;
78 static struct ir3_instruction
*
79 create_input(struct ir3_context
*ctx
, unsigned n
)
81 return create_input_compmask(ctx
, n
, 0x1);
84 static struct ir3_instruction
*
85 create_frag_input(struct ir3_context
*ctx
, bool use_ldlv
)
87 struct ir3_block
*block
= ctx
->block
;
88 struct ir3_instruction
*instr
;
89 /* actual inloc is assigned and fixed up later: */
90 struct ir3_instruction
*inloc
= create_immed(block
, 0);
93 instr
= ir3_LDLV(block
, inloc
, 0, create_immed(block
, 1), 0);
94 instr
->cat6
.type
= TYPE_U32
;
95 instr
->cat6
.iim_val
= 1;
97 instr
= ir3_BARY_F(block
, inloc
, 0, ctx
->frag_vcoord
, 0);
98 instr
->regs
[2]->wrmask
= 0x3;
104 static struct ir3_instruction
*
105 create_driver_param(struct ir3_context
*ctx
, enum ir3_driver_param dp
)
107 /* first four vec4 sysval's reserved for UBOs: */
108 /* NOTE: dp is in scalar, but there can be >4 dp components: */
109 unsigned n
= ctx
->so
->constbase
.driver_param
;
110 unsigned r
= regid(n
+ dp
/ 4, dp
% 4);
111 return create_uniform(ctx
->block
, r
);
115 * Adreno uses uint rather than having dedicated bool type,
116 * which (potentially) requires some conversion, in particular
117 * when using output of an bool instr to int input, or visa
121 * -------+---------+-------+-
125 * To convert from an adreno bool (uint) to nir, use:
127 * absneg.s dst, (neg)src
129 * To convert back in the other direction:
131 * absneg.s dst, (abs)arc
133 * The CP step can clean up the absneg.s that cancel each other
134 * out, and with a slight bit of extra cleverness (to recognize
135 * the instructions which produce either a 0 or 1) can eliminate
136 * the absneg.s's completely when an instruction that wants
137 * 0/1 consumes the result. For example, when a nir 'bcsel'
138 * consumes the result of 'feq'. So we should be able to get by
139 * without a boolean resolve step, and without incuring any
140 * extra penalty in instruction count.
143 /* NIR bool -> native (adreno): */
144 static struct ir3_instruction
*
145 ir3_b2n(struct ir3_block
*block
, struct ir3_instruction
*instr
)
147 return ir3_ABSNEG_S(block
, instr
, IR3_REG_SABS
);
150 /* native (adreno) -> NIR bool: */
151 static struct ir3_instruction
*
152 ir3_n2b(struct ir3_block
*block
, struct ir3_instruction
*instr
)
154 return ir3_ABSNEG_S(block
, instr
, IR3_REG_SNEG
);
158 * alu/sfu instructions:
161 static struct ir3_instruction
*
162 create_cov(struct ir3_context
*ctx
, struct ir3_instruction
*src
,
163 unsigned src_bitsize
, nir_op op
)
165 type_t src_type
, dst_type
;
169 case nir_op_f2f16_rtne
:
170 case nir_op_f2f16_rtz
:
178 switch (src_bitsize
) {
186 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
195 switch (src_bitsize
) {
206 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
215 switch (src_bitsize
) {
226 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
231 ir3_context_error(ctx
, "invalid conversion op: %u", op
);
241 case nir_op_f2f16_rtne
:
242 case nir_op_f2f16_rtz
:
244 /* TODO how to handle rounding mode? */
281 ir3_context_error(ctx
, "invalid conversion op: %u", op
);
284 return ir3_COV(ctx
->block
, src
, src_type
, dst_type
);
288 emit_alu(struct ir3_context
*ctx
, nir_alu_instr
*alu
)
290 const nir_op_info
*info
= &nir_op_infos
[alu
->op
];
291 struct ir3_instruction
**dst
, *src
[info
->num_inputs
];
292 unsigned bs
[info
->num_inputs
]; /* bit size */
293 struct ir3_block
*b
= ctx
->block
;
294 unsigned dst_sz
, wrmask
;
296 if (alu
->dest
.dest
.is_ssa
) {
297 dst_sz
= alu
->dest
.dest
.ssa
.num_components
;
298 wrmask
= (1 << dst_sz
) - 1;
300 dst_sz
= alu
->dest
.dest
.reg
.reg
->num_components
;
301 wrmask
= alu
->dest
.write_mask
;
304 dst
= ir3_get_dst(ctx
, &alu
->dest
.dest
, dst_sz
);
306 /* Vectors are special in that they have non-scalarized writemasks,
307 * and just take the first swizzle channel for each argument in
308 * order into each writemask channel.
310 if ((alu
->op
== nir_op_vec2
) ||
311 (alu
->op
== nir_op_vec3
) ||
312 (alu
->op
== nir_op_vec4
)) {
314 for (int i
= 0; i
< info
->num_inputs
; i
++) {
315 nir_alu_src
*asrc
= &alu
->src
[i
];
317 compile_assert(ctx
, !asrc
->abs
);
318 compile_assert(ctx
, !asrc
->negate
);
320 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[0]];
322 src
[i
] = create_immed(ctx
->block
, 0);
323 dst
[i
] = ir3_MOV(b
, src
[i
], TYPE_U32
);
326 put_dst(ctx
, &alu
->dest
.dest
);
330 /* We also get mov's with more than one component for mov's so
331 * handle those specially:
333 if ((alu
->op
== nir_op_imov
) || (alu
->op
== nir_op_fmov
)) {
334 type_t type
= (alu
->op
== nir_op_imov
) ? TYPE_U32
: TYPE_F32
;
335 nir_alu_src
*asrc
= &alu
->src
[0];
336 struct ir3_instruction
*const *src0
= ir3_get_src(ctx
, &asrc
->src
);
338 for (unsigned i
= 0; i
< dst_sz
; i
++) {
339 if (wrmask
& (1 << i
)) {
340 dst
[i
] = ir3_MOV(b
, src0
[asrc
->swizzle
[i
]], type
);
346 put_dst(ctx
, &alu
->dest
.dest
);
350 /* General case: We can just grab the one used channel per src. */
351 for (int i
= 0; i
< info
->num_inputs
; i
++) {
352 unsigned chan
= ffs(alu
->dest
.write_mask
) - 1;
353 nir_alu_src
*asrc
= &alu
->src
[i
];
355 compile_assert(ctx
, !asrc
->abs
);
356 compile_assert(ctx
, !asrc
->negate
);
358 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[chan
]];
359 bs
[i
] = nir_src_bit_size(asrc
->src
);
361 compile_assert(ctx
, src
[i
]);
366 case nir_op_f2f16_rtne
:
367 case nir_op_f2f16_rtz
:
385 dst
[0] = create_cov(ctx
, src
[0], bs
[0], alu
->op
);
388 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, create_immed(b
, fui(0.0)), 0);
389 dst
[0]->cat2
.condition
= IR3_COND_NE
;
390 dst
[0] = ir3_n2b(b
, dst
[0]);
394 dst
[0] = ir3_COV(b
, ir3_b2n(b
, src
[0]), TYPE_U32
, TYPE_F32
);
399 dst
[0] = ir3_b2n(b
, src
[0]);
402 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, create_immed(b
, 0), 0);
403 dst
[0]->cat2
.condition
= IR3_COND_NE
;
404 dst
[0] = ir3_n2b(b
, dst
[0]);
408 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FNEG
);
411 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FABS
);
414 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[1], 0);
417 dst
[0] = ir3_MIN_F(b
, src
[0], 0, src
[1], 0);
420 /* if there is just a single use of the src, and it supports
421 * (sat) bit, we can just fold the (sat) flag back to the
422 * src instruction and create a mov. This is easier for cp
425 * TODO probably opc_cat==4 is ok too
427 if (alu
->src
[0].src
.is_ssa
&&
428 (list_length(&alu
->src
[0].src
.ssa
->uses
) == 1) &&
429 ((opc_cat(src
[0]->opc
) == 2) || (opc_cat(src
[0]->opc
) == 3))) {
430 src
[0]->flags
|= IR3_INSTR_SAT
;
431 dst
[0] = ir3_MOV(b
, src
[0], TYPE_U32
);
433 /* otherwise generate a max.f that saturates.. blob does
434 * similar (generating a cat2 mov using max.f)
436 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[0], 0);
437 dst
[0]->flags
|= IR3_INSTR_SAT
;
441 dst
[0] = ir3_MUL_F(b
, src
[0], 0, src
[1], 0);
444 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], 0);
447 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], IR3_REG_FNEG
);
450 dst
[0] = ir3_MAD_F32(b
, src
[0], 0, src
[1], 0, src
[2], 0);
453 dst
[0] = ir3_DSX(b
, src
[0], 0);
454 dst
[0]->cat5
.type
= TYPE_F32
;
457 dst
[0] = ir3_DSY(b
, src
[0], 0);
458 dst
[0]->cat5
.type
= TYPE_F32
;
462 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
463 dst
[0]->cat2
.condition
= IR3_COND_LT
;
464 dst
[0] = ir3_n2b(b
, dst
[0]);
467 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
468 dst
[0]->cat2
.condition
= IR3_COND_GE
;
469 dst
[0] = ir3_n2b(b
, dst
[0]);
472 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
473 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
474 dst
[0] = ir3_n2b(b
, dst
[0]);
477 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
478 dst
[0]->cat2
.condition
= IR3_COND_NE
;
479 dst
[0] = ir3_n2b(b
, dst
[0]);
482 dst
[0] = ir3_CEIL_F(b
, src
[0], 0);
485 dst
[0] = ir3_FLOOR_F(b
, src
[0], 0);
488 dst
[0] = ir3_TRUNC_F(b
, src
[0], 0);
490 case nir_op_fround_even
:
491 dst
[0] = ir3_RNDNE_F(b
, src
[0], 0);
494 dst
[0] = ir3_SIGN_F(b
, src
[0], 0);
498 dst
[0] = ir3_SIN(b
, src
[0], 0);
501 dst
[0] = ir3_COS(b
, src
[0], 0);
504 dst
[0] = ir3_RSQ(b
, src
[0], 0);
507 dst
[0] = ir3_RCP(b
, src
[0], 0);
510 dst
[0] = ir3_LOG2(b
, src
[0], 0);
513 dst
[0] = ir3_EXP2(b
, src
[0], 0);
516 dst
[0] = ir3_SQRT(b
, src
[0], 0);
520 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SABS
);
523 dst
[0] = ir3_ADD_U(b
, src
[0], 0, src
[1], 0);
526 dst
[0] = ir3_AND_B(b
, src
[0], 0, src
[1], 0);
529 dst
[0] = ir3_MAX_S(b
, src
[0], 0, src
[1], 0);
532 dst
[0] = ir3_MAX_U(b
, src
[0], 0, src
[1], 0);
535 dst
[0] = ir3_MIN_S(b
, src
[0], 0, src
[1], 0);
538 dst
[0] = ir3_MIN_U(b
, src
[0], 0, src
[1], 0);
542 * dst = (al * bl) + (ah * bl << 16) + (al * bh << 16)
543 * mull.u tmp0, a, b ; mul low, i.e. al * bl
544 * madsh.m16 tmp1, a, b, tmp0 ; mul-add shift high mix, i.e. ah * bl << 16
545 * madsh.m16 dst, b, a, tmp1 ; i.e. al * bh << 16
547 dst
[0] = ir3_MADSH_M16(b
, src
[1], 0, src
[0], 0,
548 ir3_MADSH_M16(b
, src
[0], 0, src
[1], 0,
549 ir3_MULL_U(b
, src
[0], 0, src
[1], 0), 0), 0);
552 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SNEG
);
555 dst
[0] = ir3_NOT_B(b
, src
[0], 0);
558 dst
[0] = ir3_OR_B(b
, src
[0], 0, src
[1], 0);
561 dst
[0] = ir3_SHL_B(b
, src
[0], 0, src
[1], 0);
564 dst
[0] = ir3_ASHR_B(b
, src
[0], 0, src
[1], 0);
567 /* maybe this would be sane to lower in nir.. */
568 struct ir3_instruction
*neg
, *pos
;
570 neg
= ir3_CMPS_S(b
, src
[0], 0, create_immed(b
, 0), 0);
571 neg
->cat2
.condition
= IR3_COND_LT
;
573 pos
= ir3_CMPS_S(b
, src
[0], 0, create_immed(b
, 0), 0);
574 pos
->cat2
.condition
= IR3_COND_GT
;
576 dst
[0] = ir3_SUB_U(b
, pos
, 0, neg
, 0);
581 dst
[0] = ir3_SUB_U(b
, src
[0], 0, src
[1], 0);
584 dst
[0] = ir3_XOR_B(b
, src
[0], 0, src
[1], 0);
587 dst
[0] = ir3_SHR_B(b
, src
[0], 0, src
[1], 0);
590 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
591 dst
[0]->cat2
.condition
= IR3_COND_LT
;
592 dst
[0] = ir3_n2b(b
, dst
[0]);
595 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
596 dst
[0]->cat2
.condition
= IR3_COND_GE
;
597 dst
[0] = ir3_n2b(b
, dst
[0]);
600 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
601 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
602 dst
[0] = ir3_n2b(b
, dst
[0]);
605 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
606 dst
[0]->cat2
.condition
= IR3_COND_NE
;
607 dst
[0] = ir3_n2b(b
, dst
[0]);
610 dst
[0] = ir3_CMPS_U(b
, src
[0], 0, src
[1], 0);
611 dst
[0]->cat2
.condition
= IR3_COND_LT
;
612 dst
[0] = ir3_n2b(b
, dst
[0]);
615 dst
[0] = ir3_CMPS_U(b
, src
[0], 0, src
[1], 0);
616 dst
[0]->cat2
.condition
= IR3_COND_GE
;
617 dst
[0] = ir3_n2b(b
, dst
[0]);
620 case nir_op_b32csel
: {
621 struct ir3_instruction
*cond
= ir3_b2n(b
, src
[0]);
622 compile_assert(ctx
, bs
[1] == bs
[2]);
623 /* the boolean condition is 32b even if src[1] and src[2] are
624 * half-precision, but sel.b16 wants all three src's to be the
628 cond
= ir3_COV(b
, cond
, TYPE_U32
, TYPE_U16
);
629 dst
[0] = ir3_SEL_B32(b
, src
[1], 0, cond
, 0, src
[2], 0);
632 case nir_op_bit_count
:
633 dst
[0] = ir3_CBITS_B(b
, src
[0], 0);
635 case nir_op_ifind_msb
: {
636 struct ir3_instruction
*cmp
;
637 dst
[0] = ir3_CLZ_S(b
, src
[0], 0);
638 cmp
= ir3_CMPS_S(b
, dst
[0], 0, create_immed(b
, 0), 0);
639 cmp
->cat2
.condition
= IR3_COND_GE
;
640 dst
[0] = ir3_SEL_B32(b
,
641 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
645 case nir_op_ufind_msb
:
646 dst
[0] = ir3_CLZ_B(b
, src
[0], 0);
647 dst
[0] = ir3_SEL_B32(b
,
648 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
649 src
[0], 0, dst
[0], 0);
651 case nir_op_find_lsb
:
652 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
653 dst
[0] = ir3_CLZ_B(b
, dst
[0], 0);
655 case nir_op_bitfield_reverse
:
656 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
660 ir3_context_error(ctx
, "Unhandled ALU op: %s\n",
661 nir_op_infos
[alu
->op
].name
);
665 put_dst(ctx
, &alu
->dest
.dest
);
668 /* handles direct/indirect UBO reads: */
670 emit_intrinsic_load_ubo(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
671 struct ir3_instruction
**dst
)
673 struct ir3_block
*b
= ctx
->block
;
674 struct ir3_instruction
*base_lo
, *base_hi
, *addr
, *src0
, *src1
;
675 nir_const_value
*const_offset
;
676 /* UBO addresses are the first driver params: */
677 unsigned ubo
= regid(ctx
->so
->constbase
.ubo
, 0);
678 const unsigned ptrsz
= ir3_pointer_size(ctx
);
682 /* First src is ubo index, which could either be an immed or not: */
683 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0];
684 if (is_same_type_mov(src0
) &&
685 (src0
->regs
[1]->flags
& IR3_REG_IMMED
)) {
686 base_lo
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
));
687 base_hi
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
) + 1);
689 base_lo
= create_uniform_indirect(b
, ubo
, ir3_get_addr(ctx
, src0
, 4));
690 base_hi
= create_uniform_indirect(b
, ubo
+ 1, ir3_get_addr(ctx
, src0
, 4));
693 /* note: on 32bit gpu's base_hi is ignored and DCE'd */
696 const_offset
= nir_src_as_const_value(intr
->src
[1]);
698 off
+= const_offset
->u32
[0];
700 /* For load_ubo_indirect, second src is indirect offset: */
701 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0];
703 /* and add offset to addr: */
704 addr
= ir3_ADD_S(b
, addr
, 0, src1
, 0);
707 /* if offset is to large to encode in the ldg, split it out: */
708 if ((off
+ (intr
->num_components
* 4)) > 1024) {
709 /* split out the minimal amount to improve the odds that
710 * cp can fit the immediate in the add.s instruction:
712 unsigned off2
= off
+ (intr
->num_components
* 4) - 1024;
713 addr
= ir3_ADD_S(b
, addr
, 0, create_immed(b
, off2
), 0);
718 struct ir3_instruction
*carry
;
720 /* handle 32b rollover, ie:
721 * if (addr < base_lo)
724 carry
= ir3_CMPS_U(b
, addr
, 0, base_lo
, 0);
725 carry
->cat2
.condition
= IR3_COND_LT
;
726 base_hi
= ir3_ADD_S(b
, base_hi
, 0, carry
, 0);
728 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){ addr
, base_hi
}, 2);
731 for (int i
= 0; i
< intr
->num_components
; i
++) {
732 struct ir3_instruction
*load
=
733 ir3_LDG(b
, addr
, 0, create_immed(b
, 1), 0);
734 load
->cat6
.type
= TYPE_U32
;
735 load
->cat6
.src_offset
= off
+ i
* 4; /* byte offset */
740 /* src[] = { buffer_index, offset }. No const_index */
742 emit_intrinsic_load_ssbo(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
743 struct ir3_instruction
**dst
)
745 struct ir3_block
*b
= ctx
->block
;
746 struct ir3_instruction
*ldgb
, *src0
, *src1
, *offset
;
747 nir_const_value
*const_offset
;
749 /* can this be non-const buffer_index? how do we handle that? */
750 const_offset
= nir_src_as_const_value(intr
->src
[0]);
751 compile_assert(ctx
, const_offset
);
753 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
755 /* src0 is uvec2(offset*4, 0), src1 is offset.. nir already *= 4: */
756 src0
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
760 src1
= ir3_SHR_B(b
, offset
, 0, create_immed(b
, 2), 0);
762 ldgb
= ir3_LDGB(b
, create_immed(b
, const_offset
->u32
[0]), 0,
764 ldgb
->regs
[0]->wrmask
= MASK(intr
->num_components
);
765 ldgb
->cat6
.iim_val
= intr
->num_components
;
767 ldgb
->cat6
.type
= TYPE_U32
;
768 ldgb
->barrier_class
= IR3_BARRIER_BUFFER_R
;
769 ldgb
->barrier_conflict
= IR3_BARRIER_BUFFER_W
;
771 ir3_split_dest(b
, dst
, ldgb
, 0, intr
->num_components
);
774 /* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
776 emit_intrinsic_store_ssbo(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
778 struct ir3_block
*b
= ctx
->block
;
779 struct ir3_instruction
*stgb
, *src0
, *src1
, *src2
, *offset
;
780 nir_const_value
*const_offset
;
781 /* TODO handle wrmask properly, see _store_shared().. but I think
782 * it is more a PITA than that, since blob ends up loading the
783 * masked components and writing them back out.
785 unsigned wrmask
= intr
->const_index
[0];
786 unsigned ncomp
= ffs(~wrmask
) - 1;
788 /* can this be non-const buffer_index? how do we handle that? */
789 const_offset
= nir_src_as_const_value(intr
->src
[1]);
790 compile_assert(ctx
, const_offset
);
792 offset
= ir3_get_src(ctx
, &intr
->src
[2])[0];
794 /* src0 is value, src1 is offset, src2 is uvec2(offset*4, 0)..
797 src0
= ir3_create_collect(ctx
, ir3_get_src(ctx
, &intr
->src
[0]), ncomp
);
798 src1
= ir3_SHR_B(b
, offset
, 0, create_immed(b
, 2), 0);
799 src2
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
804 stgb
= ir3_STGB(b
, create_immed(b
, const_offset
->u32
[0]), 0,
805 src0
, 0, src1
, 0, src2
, 0);
806 stgb
->cat6
.iim_val
= ncomp
;
808 stgb
->cat6
.type
= TYPE_U32
;
809 stgb
->barrier_class
= IR3_BARRIER_BUFFER_W
;
810 stgb
->barrier_conflict
= IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
812 array_insert(b
, b
->keeps
, stgb
);
815 /* src[] = { block_index } */
817 emit_intrinsic_ssbo_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
818 struct ir3_instruction
**dst
)
820 /* SSBO size stored as a const starting at ssbo_sizes: */
821 unsigned blk_idx
= nir_src_as_const_value(intr
->src
[0])->u32
[0];
822 unsigned idx
= regid(ctx
->so
->constbase
.ssbo_sizes
, 0) +
823 ctx
->so
->const_layout
.ssbo_size
.off
[blk_idx
];
825 debug_assert(ctx
->so
->const_layout
.ssbo_size
.mask
& (1 << blk_idx
));
827 dst
[0] = create_uniform(ctx
->block
, idx
);
831 * SSBO atomic intrinsics
833 * All of the SSBO atomic memory operations read a value from memory,
834 * compute a new value using one of the operations below, write the new
835 * value to memory, and return the original value read.
837 * All operations take 3 sources except CompSwap that takes 4. These
840 * 0: The SSBO buffer index.
841 * 1: The offset into the SSBO buffer of the variable that the atomic
842 * operation will operate on.
843 * 2: The data parameter to the atomic function (i.e. the value to add
844 * in ssbo_atomic_add, etc).
845 * 3: For CompSwap only: the second data parameter.
847 static struct ir3_instruction
*
848 emit_intrinsic_atomic_ssbo(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
850 struct ir3_block
*b
= ctx
->block
;
851 struct ir3_instruction
*atomic
, *ssbo
, *src0
, *src1
, *src2
, *offset
;
852 nir_const_value
*const_offset
;
853 type_t type
= TYPE_U32
;
855 /* can this be non-const buffer_index? how do we handle that? */
856 const_offset
= nir_src_as_const_value(intr
->src
[0]);
857 compile_assert(ctx
, const_offset
);
858 ssbo
= create_immed(b
, const_offset
->u32
[0]);
860 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
862 /* src0 is data (or uvec2(data, compare))
864 * src2 is uvec2(offset*4, 0) (appears to be 64b byte offset)
866 * Note that nir already multiplies the offset by four
868 src0
= ir3_get_src(ctx
, &intr
->src
[2])[0];
869 src1
= ir3_SHR_B(b
, offset
, 0, create_immed(b
, 2), 0);
870 src2
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
875 switch (intr
->intrinsic
) {
876 case nir_intrinsic_ssbo_atomic_add
:
877 atomic
= ir3_ATOMIC_ADD_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
879 case nir_intrinsic_ssbo_atomic_imin
:
880 atomic
= ir3_ATOMIC_MIN_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
883 case nir_intrinsic_ssbo_atomic_umin
:
884 atomic
= ir3_ATOMIC_MIN_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
886 case nir_intrinsic_ssbo_atomic_imax
:
887 atomic
= ir3_ATOMIC_MAX_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
890 case nir_intrinsic_ssbo_atomic_umax
:
891 atomic
= ir3_ATOMIC_MAX_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
893 case nir_intrinsic_ssbo_atomic_and
:
894 atomic
= ir3_ATOMIC_AND_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
896 case nir_intrinsic_ssbo_atomic_or
:
897 atomic
= ir3_ATOMIC_OR_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
899 case nir_intrinsic_ssbo_atomic_xor
:
900 atomic
= ir3_ATOMIC_XOR_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
902 case nir_intrinsic_ssbo_atomic_exchange
:
903 atomic
= ir3_ATOMIC_XCHG_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
905 case nir_intrinsic_ssbo_atomic_comp_swap
:
906 /* for cmpxchg, src0 is [ui]vec2(data, compare): */
907 src0
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
908 ir3_get_src(ctx
, &intr
->src
[3])[0],
911 atomic
= ir3_ATOMIC_CMPXCHG_G(b
, ssbo
, 0, src0
, 0, src1
, 0, src2
, 0);
917 atomic
->cat6
.iim_val
= 1;
919 atomic
->cat6
.type
= type
;
920 atomic
->barrier_class
= IR3_BARRIER_BUFFER_W
;
921 atomic
->barrier_conflict
= IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
923 /* even if nothing consume the result, we can't DCE the instruction: */
924 array_insert(b
, b
->keeps
, atomic
);
929 /* src[] = { offset }. const_index[] = { base } */
931 emit_intrinsic_load_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
932 struct ir3_instruction
**dst
)
934 struct ir3_block
*b
= ctx
->block
;
935 struct ir3_instruction
*ldl
, *offset
;
938 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
939 base
= nir_intrinsic_base(intr
);
941 ldl
= ir3_LDL(b
, offset
, 0, create_immed(b
, intr
->num_components
), 0);
942 ldl
->cat6
.src_offset
= base
;
943 ldl
->cat6
.type
= utype_dst(intr
->dest
);
944 ldl
->regs
[0]->wrmask
= MASK(intr
->num_components
);
946 ldl
->barrier_class
= IR3_BARRIER_SHARED_R
;
947 ldl
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
949 ir3_split_dest(b
, dst
, ldl
, 0, intr
->num_components
);
952 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
954 emit_intrinsic_store_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
956 struct ir3_block
*b
= ctx
->block
;
957 struct ir3_instruction
*stl
, *offset
;
958 struct ir3_instruction
* const *value
;
959 unsigned base
, wrmask
;
961 value
= ir3_get_src(ctx
, &intr
->src
[0]);
962 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
964 base
= nir_intrinsic_base(intr
);
965 wrmask
= nir_intrinsic_write_mask(intr
);
967 /* Combine groups of consecutive enabled channels in one write
968 * message. We use ffs to find the first enabled channel and then ffs on
969 * the bit-inverse, down-shifted writemask to determine the length of
970 * the block of enabled bits.
972 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
975 unsigned first_component
= ffs(wrmask
) - 1;
976 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
978 stl
= ir3_STL(b
, offset
, 0,
979 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
980 create_immed(b
, length
), 0);
981 stl
->cat6
.dst_offset
= first_component
+ base
;
982 stl
->cat6
.type
= utype_src(intr
->src
[0]);
983 stl
->barrier_class
= IR3_BARRIER_SHARED_W
;
984 stl
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
986 array_insert(b
, b
->keeps
, stl
);
988 /* Clear the bits in the writemask that we just wrote, then try
989 * again to see if more channels are left.
991 wrmask
&= (15 << (first_component
+ length
));
996 * CS shared variable atomic intrinsics
998 * All of the shared variable atomic memory operations read a value from
999 * memory, compute a new value using one of the operations below, write the
1000 * new value to memory, and return the original value read.
1002 * All operations take 2 sources except CompSwap that takes 3. These
1003 * sources represent:
1005 * 0: The offset into the shared variable storage region that the atomic
1006 * operation will operate on.
1007 * 1: The data parameter to the atomic function (i.e. the value to add
1008 * in shared_atomic_add, etc).
1009 * 2: For CompSwap only: the second data parameter.
1011 static struct ir3_instruction
*
1012 emit_intrinsic_atomic_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1014 struct ir3_block
*b
= ctx
->block
;
1015 struct ir3_instruction
*atomic
, *src0
, *src1
;
1016 type_t type
= TYPE_U32
;
1018 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0]; /* offset */
1019 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0]; /* value */
1021 switch (intr
->intrinsic
) {
1022 case nir_intrinsic_shared_atomic_add
:
1023 atomic
= ir3_ATOMIC_ADD(b
, src0
, 0, src1
, 0);
1025 case nir_intrinsic_shared_atomic_imin
:
1026 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
1029 case nir_intrinsic_shared_atomic_umin
:
1030 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
1032 case nir_intrinsic_shared_atomic_imax
:
1033 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
1036 case nir_intrinsic_shared_atomic_umax
:
1037 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
1039 case nir_intrinsic_shared_atomic_and
:
1040 atomic
= ir3_ATOMIC_AND(b
, src0
, 0, src1
, 0);
1042 case nir_intrinsic_shared_atomic_or
:
1043 atomic
= ir3_ATOMIC_OR(b
, src0
, 0, src1
, 0);
1045 case nir_intrinsic_shared_atomic_xor
:
1046 atomic
= ir3_ATOMIC_XOR(b
, src0
, 0, src1
, 0);
1048 case nir_intrinsic_shared_atomic_exchange
:
1049 atomic
= ir3_ATOMIC_XCHG(b
, src0
, 0, src1
, 0);
1051 case nir_intrinsic_shared_atomic_comp_swap
:
1052 /* for cmpxchg, src1 is [ui]vec2(data, compare): */
1053 src1
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1054 ir3_get_src(ctx
, &intr
->src
[2])[0],
1057 atomic
= ir3_ATOMIC_CMPXCHG(b
, src0
, 0, src1
, 0);
1063 atomic
->cat6
.iim_val
= 1;
1065 atomic
->cat6
.type
= type
;
1066 atomic
->barrier_class
= IR3_BARRIER_SHARED_W
;
1067 atomic
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
1069 /* even if nothing consume the result, we can't DCE the instruction: */
1070 array_insert(b
, b
->keeps
, atomic
);
1075 /* Images get mapped into SSBO/image state (for store/atomic) and texture
1076 * state block (for load). To simplify things, invert the image id and
1077 * map it from end of state block, ie. image 0 becomes num-1, image 1
1078 * becomes num-2, etc. This potentially avoids needing to re-emit texture
1079 * state when switching shaders.
1081 * TODO is max # of samplers and SSBOs the same. This shouldn't be hard-
1082 * coded. Also, since all the gl shader stages (ie. everything but CS)
1083 * share the same SSBO/image state block, this might require some more
1084 * logic if we supported images in anything other than FS..
1087 get_image_slot(struct ir3_context
*ctx
, nir_deref_instr
*deref
)
1089 unsigned int loc
= 0;
1090 unsigned inner_size
= 1;
1092 while (deref
->deref_type
!= nir_deref_type_var
) {
1093 assert(deref
->deref_type
== nir_deref_type_array
);
1094 nir_const_value
*const_index
= nir_src_as_const_value(deref
->arr
.index
);
1095 assert(const_index
);
1097 /* Go to the next instruction */
1098 deref
= nir_deref_instr_parent(deref
);
1100 assert(glsl_type_is_array(deref
->type
));
1101 const unsigned array_len
= glsl_get_length(deref
->type
);
1102 loc
+= MIN2(const_index
->u32
[0], array_len
- 1) * inner_size
;
1104 /* Update the inner size */
1105 inner_size
*= array_len
;
1108 loc
+= deref
->var
->data
.driver_location
;
1110 /* TODO figure out real limit per generation, and don't hardcode: */
1111 const unsigned max_samplers
= 16;
1112 return max_samplers
- loc
- 1;
1115 /* see tex_info() for equiv logic for texture instructions.. it would be
1116 * nice if this could be better unified..
1119 get_image_coords(const nir_variable
*var
, unsigned *flagsp
)
1121 const struct glsl_type
*type
= glsl_without_array(var
->type
);
1122 unsigned coords
, flags
= 0;
1124 switch (glsl_get_sampler_dim(type
)) {
1125 case GLSL_SAMPLER_DIM_1D
:
1126 case GLSL_SAMPLER_DIM_BUF
:
1129 case GLSL_SAMPLER_DIM_2D
:
1130 case GLSL_SAMPLER_DIM_RECT
:
1131 case GLSL_SAMPLER_DIM_EXTERNAL
:
1132 case GLSL_SAMPLER_DIM_MS
:
1135 case GLSL_SAMPLER_DIM_3D
:
1136 case GLSL_SAMPLER_DIM_CUBE
:
1137 flags
|= IR3_INSTR_3D
;
1141 unreachable("bad sampler dim");
1145 if (glsl_sampler_type_is_array(type
)) {
1146 /* note: unlike tex_info(), adjust # of coords to include array idx: */
1148 flags
|= IR3_INSTR_A
;
1158 get_image_type(const nir_variable
*var
)
1160 switch (glsl_get_sampler_result_type(glsl_without_array(var
->type
))) {
1161 case GLSL_TYPE_UINT
:
1165 case GLSL_TYPE_FLOAT
:
1168 unreachable("bad sampler type.");
1173 static struct ir3_instruction
*
1174 get_image_offset(struct ir3_context
*ctx
, const nir_variable
*var
,
1175 struct ir3_instruction
* const *coords
, bool byteoff
)
1177 struct ir3_block
*b
= ctx
->block
;
1178 struct ir3_instruction
*offset
;
1179 unsigned ncoords
= get_image_coords(var
, NULL
);
1181 /* to calculate the byte offset (yes, uggg) we need (up to) three
1182 * const values to know the bytes per pixel, and y and z stride:
1184 unsigned cb
= regid(ctx
->so
->constbase
.image_dims
, 0) +
1185 ctx
->so
->const_layout
.image_dims
.off
[var
->data
.driver_location
];
1187 debug_assert(ctx
->so
->const_layout
.image_dims
.mask
&
1188 (1 << var
->data
.driver_location
));
1190 /* offset = coords.x * bytes_per_pixel: */
1191 offset
= ir3_MUL_S(b
, coords
[0], 0, create_uniform(b
, cb
+ 0), 0);
1193 /* offset += coords.y * y_pitch: */
1194 offset
= ir3_MAD_S24(b
, create_uniform(b
, cb
+ 1), 0,
1195 coords
[1], 0, offset
, 0);
1198 /* offset += coords.z * z_pitch: */
1199 offset
= ir3_MAD_S24(b
, create_uniform(b
, cb
+ 2), 0,
1200 coords
[2], 0, offset
, 0);
1204 /* Some cases, like atomics, seem to use dword offset instead
1205 * of byte offsets.. blob just puts an extra shr.b in there
1208 offset
= ir3_SHR_B(b
, offset
, 0, create_immed(b
, 2), 0);
1211 return ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1217 /* src[] = { deref, coord, sample_index }. const_index[] = {} */
1219 emit_intrinsic_load_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1220 struct ir3_instruction
**dst
)
1222 struct ir3_block
*b
= ctx
->block
;
1223 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
1224 struct ir3_instruction
*sam
;
1225 struct ir3_instruction
* const *src0
= ir3_get_src(ctx
, &intr
->src
[1]);
1226 struct ir3_instruction
*coords
[4];
1227 unsigned flags
, ncoords
= get_image_coords(var
, &flags
);
1228 unsigned tex_idx
= get_image_slot(ctx
, nir_src_as_deref(intr
->src
[0]));
1229 type_t type
= get_image_type(var
);
1231 /* hmm, this seems a bit odd, but it is what blob does and (at least
1232 * a5xx) just faults on bogus addresses otherwise:
1234 if (flags
& IR3_INSTR_3D
) {
1235 flags
&= ~IR3_INSTR_3D
;
1236 flags
|= IR3_INSTR_A
;
1239 for (unsigned i
= 0; i
< ncoords
; i
++)
1240 coords
[i
] = src0
[i
];
1243 coords
[ncoords
++] = create_immed(b
, 0);
1245 sam
= ir3_SAM(b
, OPC_ISAM
, type
, 0b1111, flags
,
1246 tex_idx
, tex_idx
, ir3_create_collect(ctx
, coords
, ncoords
), NULL
);
1248 sam
->barrier_class
= IR3_BARRIER_IMAGE_R
;
1249 sam
->barrier_conflict
= IR3_BARRIER_IMAGE_W
;
1251 ir3_split_dest(b
, dst
, sam
, 0, 4);
1254 /* src[] = { deref, coord, sample_index, value }. const_index[] = {} */
1256 emit_intrinsic_store_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1258 struct ir3_block
*b
= ctx
->block
;
1259 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
1260 struct ir3_instruction
*stib
, *offset
;
1261 struct ir3_instruction
* const *value
= ir3_get_src(ctx
, &intr
->src
[3]);
1262 struct ir3_instruction
* const *coords
= ir3_get_src(ctx
, &intr
->src
[1]);
1263 unsigned ncoords
= get_image_coords(var
, NULL
);
1264 unsigned tex_idx
= get_image_slot(ctx
, nir_src_as_deref(intr
->src
[0]));
1268 * src2 is 64b byte offset
1271 offset
= get_image_offset(ctx
, var
, coords
, true);
1273 /* NOTE: stib seems to take byte offset, but stgb.typed can be used
1274 * too and takes a dword offset.. not quite sure yet why blob uses
1275 * one over the other in various cases.
1278 stib
= ir3_STIB(b
, create_immed(b
, tex_idx
), 0,
1279 ir3_create_collect(ctx
, value
, 4), 0,
1280 ir3_create_collect(ctx
, coords
, ncoords
), 0,
1282 stib
->cat6
.iim_val
= 4;
1283 stib
->cat6
.d
= ncoords
;
1284 stib
->cat6
.type
= get_image_type(var
);
1285 stib
->cat6
.typed
= true;
1286 stib
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1287 stib
->barrier_conflict
= IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
;
1289 array_insert(b
, b
->keeps
, stib
);
1293 emit_intrinsic_image_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1294 struct ir3_instruction
**dst
)
1296 struct ir3_block
*b
= ctx
->block
;
1297 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
1298 unsigned tex_idx
= get_image_slot(ctx
, nir_src_as_deref(intr
->src
[0]));
1299 struct ir3_instruction
*sam
, *lod
;
1300 unsigned flags
, ncoords
= get_image_coords(var
, &flags
);
1302 lod
= create_immed(b
, 0);
1303 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
1304 tex_idx
, tex_idx
, lod
, NULL
);
1306 /* Array size actually ends up in .w rather than .z. This doesn't
1307 * matter for miplevel 0, but for higher mips the value in z is
1308 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
1309 * returned, which means that we have to add 1 to it for arrays for
1312 * Note use a temporary dst and then copy, since the size of the dst
1313 * array that is passed in is based on nir's understanding of the
1314 * result size, not the hardware's
1316 struct ir3_instruction
*tmp
[4];
1318 ir3_split_dest(b
, tmp
, sam
, 0, 4);
1320 /* get_size instruction returns size in bytes instead of texels
1321 * for imageBuffer, so we need to divide it by the pixel size
1322 * of the image format.
1324 * TODO: This is at least true on a5xx. Check other gens.
1326 enum glsl_sampler_dim dim
=
1327 glsl_get_sampler_dim(glsl_without_array(var
->type
));
1328 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
1329 /* Since all the possible values the divisor can take are
1330 * power-of-two (4, 8, or 16), the division is implemented
1332 * During shader setup, the log2 of the image format's
1333 * bytes-per-pixel should have been emitted in 2nd slot of
1334 * image_dims. See ir3_shader::emit_image_dims().
1336 unsigned cb
= regid(ctx
->so
->constbase
.image_dims
, 0) +
1337 ctx
->so
->const_layout
.image_dims
.off
[var
->data
.driver_location
];
1338 struct ir3_instruction
*aux
= create_uniform(b
, cb
+ 1);
1340 tmp
[0] = ir3_SHR_B(b
, tmp
[0], 0, aux
, 0);
1343 for (unsigned i
= 0; i
< ncoords
; i
++)
1346 if (flags
& IR3_INSTR_A
) {
1347 if (ctx
->compiler
->levels_add_one
) {
1348 dst
[ncoords
-1] = ir3_ADD_U(b
, tmp
[3], 0, create_immed(b
, 1), 0);
1350 dst
[ncoords
-1] = ir3_MOV(b
, tmp
[3], TYPE_U32
);
1355 /* src[] = { deref, coord, sample_index, value, compare }. const_index[] = {} */
1356 static struct ir3_instruction
*
1357 emit_intrinsic_atomic_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1359 struct ir3_block
*b
= ctx
->block
;
1360 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
1361 struct ir3_instruction
*atomic
, *image
, *src0
, *src1
, *src2
;
1362 struct ir3_instruction
* const *coords
= ir3_get_src(ctx
, &intr
->src
[1]);
1363 unsigned ncoords
= get_image_coords(var
, NULL
);
1365 image
= create_immed(b
, get_image_slot(ctx
, nir_src_as_deref(intr
->src
[0])));
1367 /* src0 is value (or uvec2(value, compare))
1369 * src2 is 64b byte offset
1371 src0
= ir3_get_src(ctx
, &intr
->src
[3])[0];
1372 src1
= ir3_create_collect(ctx
, coords
, ncoords
);
1373 src2
= get_image_offset(ctx
, var
, coords
, false);
1375 switch (intr
->intrinsic
) {
1376 case nir_intrinsic_image_deref_atomic_add
:
1377 atomic
= ir3_ATOMIC_ADD_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1379 case nir_intrinsic_image_deref_atomic_min
:
1380 atomic
= ir3_ATOMIC_MIN_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1382 case nir_intrinsic_image_deref_atomic_max
:
1383 atomic
= ir3_ATOMIC_MAX_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1385 case nir_intrinsic_image_deref_atomic_and
:
1386 atomic
= ir3_ATOMIC_AND_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1388 case nir_intrinsic_image_deref_atomic_or
:
1389 atomic
= ir3_ATOMIC_OR_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1391 case nir_intrinsic_image_deref_atomic_xor
:
1392 atomic
= ir3_ATOMIC_XOR_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1394 case nir_intrinsic_image_deref_atomic_exchange
:
1395 atomic
= ir3_ATOMIC_XCHG_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1397 case nir_intrinsic_image_deref_atomic_comp_swap
:
1398 /* for cmpxchg, src0 is [ui]vec2(data, compare): */
1399 src0
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1400 ir3_get_src(ctx
, &intr
->src
[4])[0],
1403 atomic
= ir3_ATOMIC_CMPXCHG_G(b
, image
, 0, src0
, 0, src1
, 0, src2
, 0);
1409 atomic
->cat6
.iim_val
= 1;
1410 atomic
->cat6
.d
= ncoords
;
1411 atomic
->cat6
.type
= get_image_type(var
);
1412 atomic
->cat6
.typed
= true;
1413 atomic
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1414 atomic
->barrier_conflict
= IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
;
1416 /* even if nothing consume the result, we can't DCE the instruction: */
1417 array_insert(b
, b
->keeps
, atomic
);
1423 emit_intrinsic_barrier(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1425 struct ir3_block
*b
= ctx
->block
;
1426 struct ir3_instruction
*barrier
;
1428 switch (intr
->intrinsic
) {
1429 case nir_intrinsic_barrier
:
1430 barrier
= ir3_BAR(b
);
1431 barrier
->cat7
.g
= true;
1432 barrier
->cat7
.l
= true;
1433 barrier
->flags
= IR3_INSTR_SS
| IR3_INSTR_SY
;
1434 barrier
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1436 case nir_intrinsic_memory_barrier
:
1437 barrier
= ir3_FENCE(b
);
1438 barrier
->cat7
.g
= true;
1439 barrier
->cat7
.r
= true;
1440 barrier
->cat7
.w
= true;
1441 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
|
1442 IR3_BARRIER_BUFFER_W
;
1443 barrier
->barrier_conflict
=
1444 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1445 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1447 case nir_intrinsic_memory_barrier_atomic_counter
:
1448 case nir_intrinsic_memory_barrier_buffer
:
1449 barrier
= ir3_FENCE(b
);
1450 barrier
->cat7
.g
= true;
1451 barrier
->cat7
.r
= true;
1452 barrier
->cat7
.w
= true;
1453 barrier
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1454 barrier
->barrier_conflict
= IR3_BARRIER_BUFFER_R
|
1455 IR3_BARRIER_BUFFER_W
;
1457 case nir_intrinsic_memory_barrier_image
:
1458 // TODO double check if this should have .g set
1459 barrier
= ir3_FENCE(b
);
1460 barrier
->cat7
.g
= true;
1461 barrier
->cat7
.r
= true;
1462 barrier
->cat7
.w
= true;
1463 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1464 barrier
->barrier_conflict
= IR3_BARRIER_IMAGE_R
|
1465 IR3_BARRIER_IMAGE_W
;
1467 case nir_intrinsic_memory_barrier_shared
:
1468 barrier
= ir3_FENCE(b
);
1469 barrier
->cat7
.g
= true;
1470 barrier
->cat7
.l
= true;
1471 barrier
->cat7
.r
= true;
1472 barrier
->cat7
.w
= true;
1473 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
;
1474 barrier
->barrier_conflict
= IR3_BARRIER_SHARED_R
|
1475 IR3_BARRIER_SHARED_W
;
1477 case nir_intrinsic_group_memory_barrier
:
1478 barrier
= ir3_FENCE(b
);
1479 barrier
->cat7
.g
= true;
1480 barrier
->cat7
.l
= true;
1481 barrier
->cat7
.r
= true;
1482 barrier
->cat7
.w
= true;
1483 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
|
1484 IR3_BARRIER_IMAGE_W
|
1485 IR3_BARRIER_BUFFER_W
;
1486 barrier
->barrier_conflict
=
1487 IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
|
1488 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1489 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1495 /* make sure barrier doesn't get DCE'd */
1496 array_insert(b
, b
->keeps
, barrier
);
1499 static void add_sysval_input_compmask(struct ir3_context
*ctx
,
1500 gl_system_value slot
, unsigned compmask
,
1501 struct ir3_instruction
*instr
)
1503 struct ir3_shader_variant
*so
= ctx
->so
;
1504 unsigned r
= regid(so
->inputs_count
, 0);
1505 unsigned n
= so
->inputs_count
++;
1507 so
->inputs
[n
].sysval
= true;
1508 so
->inputs
[n
].slot
= slot
;
1509 so
->inputs
[n
].compmask
= compmask
;
1510 so
->inputs
[n
].regid
= r
;
1511 so
->inputs
[n
].interpolate
= INTERP_MODE_FLAT
;
1514 ctx
->ir
->ninputs
= MAX2(ctx
->ir
->ninputs
, r
+ 1);
1515 ctx
->ir
->inputs
[r
] = instr
;
1518 static void add_sysval_input(struct ir3_context
*ctx
, gl_system_value slot
,
1519 struct ir3_instruction
*instr
)
1521 add_sysval_input_compmask(ctx
, slot
, 0x1, instr
);
1525 emit_intrinsic(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1527 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1528 struct ir3_instruction
**dst
;
1529 struct ir3_instruction
* const *src
;
1530 struct ir3_block
*b
= ctx
->block
;
1531 nir_const_value
*const_offset
;
1534 if (info
->has_dest
) {
1535 unsigned n
= nir_intrinsic_dest_components(intr
);
1536 dst
= ir3_get_dst(ctx
, &intr
->dest
, n
);
1541 switch (intr
->intrinsic
) {
1542 case nir_intrinsic_load_uniform
:
1543 idx
= nir_intrinsic_base(intr
);
1544 const_offset
= nir_src_as_const_value(intr
->src
[0]);
1546 idx
+= const_offset
->u32
[0];
1547 for (int i
= 0; i
< intr
->num_components
; i
++) {
1548 unsigned n
= idx
* 4 + i
;
1549 dst
[i
] = create_uniform(b
, n
);
1552 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1553 for (int i
= 0; i
< intr
->num_components
; i
++) {
1554 int n
= idx
* 4 + i
;
1555 dst
[i
] = create_uniform_indirect(b
, n
,
1556 ir3_get_addr(ctx
, src
[0], 4));
1558 /* NOTE: if relative addressing is used, we set
1559 * constlen in the compiler (to worst-case value)
1560 * since we don't know in the assembler what the max
1561 * addr reg value can be:
1563 ctx
->so
->constlen
= ctx
->s
->num_uniforms
;
1566 case nir_intrinsic_load_ubo
:
1567 emit_intrinsic_load_ubo(ctx
, intr
, dst
);
1569 case nir_intrinsic_load_input
:
1570 idx
= nir_intrinsic_base(intr
);
1571 comp
= nir_intrinsic_component(intr
);
1572 const_offset
= nir_src_as_const_value(intr
->src
[0]);
1574 idx
+= const_offset
->u32
[0];
1575 for (int i
= 0; i
< intr
->num_components
; i
++) {
1576 unsigned n
= idx
* 4 + i
+ comp
;
1577 dst
[i
] = ctx
->ir
->inputs
[n
];
1580 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1581 struct ir3_instruction
*collect
=
1582 ir3_create_collect(ctx
, ctx
->ir
->inputs
, ctx
->ir
->ninputs
);
1583 struct ir3_instruction
*addr
= ir3_get_addr(ctx
, src
[0], 4);
1584 for (int i
= 0; i
< intr
->num_components
; i
++) {
1585 unsigned n
= idx
* 4 + i
+ comp
;
1586 dst
[i
] = create_indirect_load(ctx
, ctx
->ir
->ninputs
,
1591 case nir_intrinsic_load_ssbo
:
1592 emit_intrinsic_load_ssbo(ctx
, intr
, dst
);
1594 case nir_intrinsic_store_ssbo
:
1595 emit_intrinsic_store_ssbo(ctx
, intr
);
1597 case nir_intrinsic_get_buffer_size
:
1598 emit_intrinsic_ssbo_size(ctx
, intr
, dst
);
1600 case nir_intrinsic_ssbo_atomic_add
:
1601 case nir_intrinsic_ssbo_atomic_imin
:
1602 case nir_intrinsic_ssbo_atomic_umin
:
1603 case nir_intrinsic_ssbo_atomic_imax
:
1604 case nir_intrinsic_ssbo_atomic_umax
:
1605 case nir_intrinsic_ssbo_atomic_and
:
1606 case nir_intrinsic_ssbo_atomic_or
:
1607 case nir_intrinsic_ssbo_atomic_xor
:
1608 case nir_intrinsic_ssbo_atomic_exchange
:
1609 case nir_intrinsic_ssbo_atomic_comp_swap
:
1610 dst
[0] = emit_intrinsic_atomic_ssbo(ctx
, intr
);
1612 case nir_intrinsic_load_shared
:
1613 emit_intrinsic_load_shared(ctx
, intr
, dst
);
1615 case nir_intrinsic_store_shared
:
1616 emit_intrinsic_store_shared(ctx
, intr
);
1618 case nir_intrinsic_shared_atomic_add
:
1619 case nir_intrinsic_shared_atomic_imin
:
1620 case nir_intrinsic_shared_atomic_umin
:
1621 case nir_intrinsic_shared_atomic_imax
:
1622 case nir_intrinsic_shared_atomic_umax
:
1623 case nir_intrinsic_shared_atomic_and
:
1624 case nir_intrinsic_shared_atomic_or
:
1625 case nir_intrinsic_shared_atomic_xor
:
1626 case nir_intrinsic_shared_atomic_exchange
:
1627 case nir_intrinsic_shared_atomic_comp_swap
:
1628 dst
[0] = emit_intrinsic_atomic_shared(ctx
, intr
);
1630 case nir_intrinsic_image_deref_load
:
1631 emit_intrinsic_load_image(ctx
, intr
, dst
);
1633 case nir_intrinsic_image_deref_store
:
1634 emit_intrinsic_store_image(ctx
, intr
);
1636 case nir_intrinsic_image_deref_size
:
1637 emit_intrinsic_image_size(ctx
, intr
, dst
);
1639 case nir_intrinsic_image_deref_atomic_add
:
1640 case nir_intrinsic_image_deref_atomic_min
:
1641 case nir_intrinsic_image_deref_atomic_max
:
1642 case nir_intrinsic_image_deref_atomic_and
:
1643 case nir_intrinsic_image_deref_atomic_or
:
1644 case nir_intrinsic_image_deref_atomic_xor
:
1645 case nir_intrinsic_image_deref_atomic_exchange
:
1646 case nir_intrinsic_image_deref_atomic_comp_swap
:
1647 dst
[0] = emit_intrinsic_atomic_image(ctx
, intr
);
1649 case nir_intrinsic_barrier
:
1650 case nir_intrinsic_memory_barrier
:
1651 case nir_intrinsic_group_memory_barrier
:
1652 case nir_intrinsic_memory_barrier_atomic_counter
:
1653 case nir_intrinsic_memory_barrier_buffer
:
1654 case nir_intrinsic_memory_barrier_image
:
1655 case nir_intrinsic_memory_barrier_shared
:
1656 emit_intrinsic_barrier(ctx
, intr
);
1657 /* note that blk ptr no longer valid, make that obvious: */
1660 case nir_intrinsic_store_output
:
1661 idx
= nir_intrinsic_base(intr
);
1662 comp
= nir_intrinsic_component(intr
);
1663 const_offset
= nir_src_as_const_value(intr
->src
[1]);
1664 compile_assert(ctx
, const_offset
!= NULL
);
1665 idx
+= const_offset
->u32
[0];
1667 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1668 for (int i
= 0; i
< intr
->num_components
; i
++) {
1669 unsigned n
= idx
* 4 + i
+ comp
;
1670 ctx
->ir
->outputs
[n
] = src
[i
];
1673 case nir_intrinsic_load_base_vertex
:
1674 case nir_intrinsic_load_first_vertex
:
1675 if (!ctx
->basevertex
) {
1676 ctx
->basevertex
= create_driver_param(ctx
, IR3_DP_VTXID_BASE
);
1677 add_sysval_input(ctx
, SYSTEM_VALUE_FIRST_VERTEX
, ctx
->basevertex
);
1679 dst
[0] = ctx
->basevertex
;
1681 case nir_intrinsic_load_vertex_id_zero_base
:
1682 case nir_intrinsic_load_vertex_id
:
1683 if (!ctx
->vertex_id
) {
1684 gl_system_value sv
= (intr
->intrinsic
== nir_intrinsic_load_vertex_id
) ?
1685 SYSTEM_VALUE_VERTEX_ID
: SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1686 ctx
->vertex_id
= create_input(ctx
, 0);
1687 add_sysval_input(ctx
, sv
, ctx
->vertex_id
);
1689 dst
[0] = ctx
->vertex_id
;
1691 case nir_intrinsic_load_instance_id
:
1692 if (!ctx
->instance_id
) {
1693 ctx
->instance_id
= create_input(ctx
, 0);
1694 add_sysval_input(ctx
, SYSTEM_VALUE_INSTANCE_ID
,
1697 dst
[0] = ctx
->instance_id
;
1699 case nir_intrinsic_load_sample_id
:
1700 case nir_intrinsic_load_sample_id_no_per_sample
:
1701 if (!ctx
->samp_id
) {
1702 ctx
->samp_id
= create_input(ctx
, 0);
1703 ctx
->samp_id
->regs
[0]->flags
|= IR3_REG_HALF
;
1704 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_ID
,
1707 dst
[0] = ir3_COV(b
, ctx
->samp_id
, TYPE_U16
, TYPE_U32
);
1709 case nir_intrinsic_load_sample_mask_in
:
1710 if (!ctx
->samp_mask_in
) {
1711 ctx
->samp_mask_in
= create_input(ctx
, 0);
1712 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_MASK_IN
,
1715 dst
[0] = ctx
->samp_mask_in
;
1717 case nir_intrinsic_load_user_clip_plane
:
1718 idx
= nir_intrinsic_ucp_id(intr
);
1719 for (int i
= 0; i
< intr
->num_components
; i
++) {
1720 unsigned n
= idx
* 4 + i
;
1721 dst
[i
] = create_driver_param(ctx
, IR3_DP_UCP0_X
+ n
);
1724 case nir_intrinsic_load_front_face
:
1725 if (!ctx
->frag_face
) {
1726 ctx
->so
->frag_face
= true;
1727 ctx
->frag_face
= create_input(ctx
, 0);
1728 add_sysval_input(ctx
, SYSTEM_VALUE_FRONT_FACE
, ctx
->frag_face
);
1729 ctx
->frag_face
->regs
[0]->flags
|= IR3_REG_HALF
;
1731 /* for fragface, we get -1 for back and 0 for front. However this is
1732 * the inverse of what nir expects (where ~0 is true).
1734 dst
[0] = ir3_COV(b
, ctx
->frag_face
, TYPE_S16
, TYPE_S32
);
1735 dst
[0] = ir3_NOT_B(b
, dst
[0], 0);
1737 case nir_intrinsic_load_local_invocation_id
:
1738 if (!ctx
->local_invocation_id
) {
1739 ctx
->local_invocation_id
= create_input_compmask(ctx
, 0, 0x7);
1740 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_LOCAL_INVOCATION_ID
,
1741 0x7, ctx
->local_invocation_id
);
1743 ir3_split_dest(b
, dst
, ctx
->local_invocation_id
, 0, 3);
1745 case nir_intrinsic_load_work_group_id
:
1746 if (!ctx
->work_group_id
) {
1747 ctx
->work_group_id
= create_input_compmask(ctx
, 0, 0x7);
1748 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_WORK_GROUP_ID
,
1749 0x7, ctx
->work_group_id
);
1750 ctx
->work_group_id
->regs
[0]->flags
|= IR3_REG_HIGH
;
1752 ir3_split_dest(b
, dst
, ctx
->work_group_id
, 0, 3);
1754 case nir_intrinsic_load_num_work_groups
:
1755 for (int i
= 0; i
< intr
->num_components
; i
++) {
1756 dst
[i
] = create_driver_param(ctx
, IR3_DP_NUM_WORK_GROUPS_X
+ i
);
1759 case nir_intrinsic_load_local_group_size
:
1760 for (int i
= 0; i
< intr
->num_components
; i
++) {
1761 dst
[i
] = create_driver_param(ctx
, IR3_DP_LOCAL_GROUP_SIZE_X
+ i
);
1764 case nir_intrinsic_discard_if
:
1765 case nir_intrinsic_discard
: {
1766 struct ir3_instruction
*cond
, *kill
;
1768 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
1769 /* conditional discard: */
1770 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1771 cond
= ir3_b2n(b
, src
[0]);
1773 /* unconditional discard: */
1774 cond
= create_immed(b
, 1);
1777 /* NOTE: only cmps.*.* can write p0.x: */
1778 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1779 cond
->cat2
.condition
= IR3_COND_NE
;
1781 /* condition always goes in predicate register: */
1782 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1784 kill
= ir3_KILL(b
, cond
, 0);
1785 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1787 array_insert(b
, b
->keeps
, kill
);
1788 ctx
->so
->has_kill
= true;
1793 ir3_context_error(ctx
, "Unhandled intrinsic type: %s\n",
1794 nir_intrinsic_infos
[intr
->intrinsic
].name
);
1799 put_dst(ctx
, &intr
->dest
);
1803 emit_load_const(struct ir3_context
*ctx
, nir_load_const_instr
*instr
)
1805 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &instr
->def
,
1806 instr
->def
.num_components
);
1807 type_t type
= (instr
->def
.bit_size
< 32) ? TYPE_U16
: TYPE_U32
;
1809 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1810 dst
[i
] = create_immed_typed(ctx
->block
, instr
->value
.u32
[i
], type
);
1814 emit_undef(struct ir3_context
*ctx
, nir_ssa_undef_instr
*undef
)
1816 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &undef
->def
,
1817 undef
->def
.num_components
);
1818 type_t type
= (undef
->def
.bit_size
< 32) ? TYPE_U16
: TYPE_U32
;
1820 /* backend doesn't want undefined instructions, so just plug
1823 for (int i
= 0; i
< undef
->def
.num_components
; i
++)
1824 dst
[i
] = create_immed_typed(ctx
->block
, fui(0.0), type
);
1828 * texture fetch/sample instructions:
1832 tex_info(nir_tex_instr
*tex
, unsigned *flagsp
, unsigned *coordsp
)
1834 unsigned coords
, flags
= 0;
1836 /* note: would use tex->coord_components.. except txs.. also,
1837 * since array index goes after shadow ref, we don't want to
1840 switch (tex
->sampler_dim
) {
1841 case GLSL_SAMPLER_DIM_1D
:
1842 case GLSL_SAMPLER_DIM_BUF
:
1845 case GLSL_SAMPLER_DIM_2D
:
1846 case GLSL_SAMPLER_DIM_RECT
:
1847 case GLSL_SAMPLER_DIM_EXTERNAL
:
1848 case GLSL_SAMPLER_DIM_MS
:
1851 case GLSL_SAMPLER_DIM_3D
:
1852 case GLSL_SAMPLER_DIM_CUBE
:
1854 flags
|= IR3_INSTR_3D
;
1857 unreachable("bad sampler_dim");
1860 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
1861 flags
|= IR3_INSTR_S
;
1863 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
)
1864 flags
|= IR3_INSTR_A
;
1871 emit_tex(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1873 struct ir3_block
*b
= ctx
->block
;
1874 struct ir3_instruction
**dst
, *sam
, *src0
[12], *src1
[4];
1875 struct ir3_instruction
* const *coord
, * const *off
, * const *ddx
, * const *ddy
;
1876 struct ir3_instruction
*lod
, *compare
, *proj
, *sample_index
;
1877 bool has_bias
= false, has_lod
= false, has_proj
= false, has_off
= false;
1878 unsigned i
, coords
, flags
;
1879 unsigned nsrc0
= 0, nsrc1
= 0;
1883 coord
= off
= ddx
= ddy
= NULL
;
1884 lod
= proj
= compare
= sample_index
= NULL
;
1886 /* TODO: might just be one component for gathers? */
1887 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
1889 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1890 switch (tex
->src
[i
].src_type
) {
1891 case nir_tex_src_coord
:
1892 coord
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1894 case nir_tex_src_bias
:
1895 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1898 case nir_tex_src_lod
:
1899 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1902 case nir_tex_src_comparator
: /* shadow comparator */
1903 compare
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1905 case nir_tex_src_projector
:
1906 proj
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1909 case nir_tex_src_offset
:
1910 off
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1913 case nir_tex_src_ddx
:
1914 ddx
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1916 case nir_tex_src_ddy
:
1917 ddy
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1919 case nir_tex_src_ms_index
:
1920 sample_index
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1923 ir3_context_error(ctx
, "Unhandled NIR tex src type: %d\n",
1924 tex
->src
[i
].src_type
);
1930 case nir_texop_tex
: opc
= has_lod
? OPC_SAML
: OPC_SAM
; break;
1931 case nir_texop_txb
: opc
= OPC_SAMB
; break;
1932 case nir_texop_txl
: opc
= OPC_SAML
; break;
1933 case nir_texop_txd
: opc
= OPC_SAMGQ
; break;
1934 case nir_texop_txf
: opc
= OPC_ISAML
; break;
1935 case nir_texop_lod
: opc
= OPC_GETLOD
; break;
1937 /* NOTE: a4xx might need to emulate gather w/ txf (this is
1938 * what blob does, seems gather is broken?), and a3xx did
1939 * not support it (but probably could also emulate).
1941 switch (tex
->component
) {
1942 case 0: opc
= OPC_GATHER4R
; break;
1943 case 1: opc
= OPC_GATHER4G
; break;
1944 case 2: opc
= OPC_GATHER4B
; break;
1945 case 3: opc
= OPC_GATHER4A
; break;
1948 case nir_texop_txf_ms
: opc
= OPC_ISAMM
; break;
1950 case nir_texop_query_levels
:
1951 case nir_texop_texture_samples
:
1952 case nir_texop_samples_identical
:
1953 case nir_texop_txf_ms_mcs
:
1954 ir3_context_error(ctx
, "Unhandled NIR tex type: %d\n", tex
->op
);
1958 tex_info(tex
, &flags
, &coords
);
1961 * lay out the first argument in the proper order:
1962 * - actual coordinates first
1963 * - shadow reference
1966 * - starting at offset 4, dpdx.xy, dpdy.xy
1968 * bias/lod go into the second arg
1971 /* insert tex coords: */
1972 for (i
= 0; i
< coords
; i
++)
1977 /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
1978 * with scaled x coord according to requested sample:
1980 if (tex
->op
== nir_texop_txf_ms
) {
1981 if (ctx
->compiler
->txf_ms_with_isaml
) {
1982 /* the samples are laid out in x dimension as
1984 * x_ms = (x << ms) + sample_index;
1986 struct ir3_instruction
*ms
;
1987 ms
= create_immed(b
, (ctx
->samples
>> (2 * tex
->texture_index
)) & 3);
1989 src0
[0] = ir3_SHL_B(b
, src0
[0], 0, ms
, 0);
1990 src0
[0] = ir3_ADD_U(b
, src0
[0], 0, sample_index
, 0);
1994 src0
[nsrc0
++] = sample_index
;
1998 /* scale up integer coords for TXF based on the LOD */
1999 if (ctx
->compiler
->unminify_coords
&& (opc
== OPC_ISAML
)) {
2001 for (i
= 0; i
< coords
; i
++)
2002 src0
[i
] = ir3_SHL_B(b
, src0
[i
], 0, lod
, 0);
2006 /* hw doesn't do 1d, so we treat it as 2d with
2007 * height of 1, and patch up the y coord.
2008 * TODO: y coord should be (int)0 in some cases..
2010 src0
[nsrc0
++] = create_immed(b
, fui(0.5));
2013 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
2014 src0
[nsrc0
++] = compare
;
2016 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
) {
2017 struct ir3_instruction
*idx
= coord
[coords
];
2019 /* the array coord for cube arrays needs 0.5 added to it */
2020 if (ctx
->compiler
->array_index_add_half
&& (opc
!= OPC_ISAML
))
2021 idx
= ir3_ADD_F(b
, idx
, 0, create_immed(b
, fui(0.5)), 0);
2023 src0
[nsrc0
++] = idx
;
2027 src0
[nsrc0
++] = proj
;
2028 flags
|= IR3_INSTR_P
;
2031 /* pad to 4, then ddx/ddy: */
2032 if (tex
->op
== nir_texop_txd
) {
2034 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2035 for (i
= 0; i
< coords
; i
++)
2036 src0
[nsrc0
++] = ddx
[i
];
2038 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2039 for (i
= 0; i
< coords
; i
++)
2040 src0
[nsrc0
++] = ddy
[i
];
2042 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2046 * second argument (if applicable):
2051 if (has_off
| has_lod
| has_bias
) {
2053 unsigned off_coords
= coords
;
2054 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2056 for (i
= 0; i
< off_coords
; i
++)
2057 src1
[nsrc1
++] = off
[i
];
2059 src1
[nsrc1
++] = create_immed(b
, fui(0.0));
2060 flags
|= IR3_INSTR_O
;
2063 if (has_lod
| has_bias
)
2064 src1
[nsrc1
++] = lod
;
2067 switch (tex
->dest_type
) {
2068 case nir_type_invalid
:
2069 case nir_type_float
:
2080 unreachable("bad dest_type");
2083 if (opc
== OPC_GETLOD
)
2086 unsigned tex_idx
= tex
->texture_index
;
2088 ctx
->max_texture_index
= MAX2(ctx
->max_texture_index
, tex_idx
);
2090 struct ir3_instruction
*col0
= ir3_create_collect(ctx
, src0
, nsrc0
);
2091 struct ir3_instruction
*col1
= ir3_create_collect(ctx
, src1
, nsrc1
);
2093 sam
= ir3_SAM(b
, opc
, type
, 0b1111, flags
,
2094 tex_idx
, tex_idx
, col0
, col1
);
2096 if ((ctx
->astc_srgb
& (1 << tex_idx
)) && !nir_tex_instr_is_query(tex
)) {
2097 /* only need first 3 components: */
2098 sam
->regs
[0]->wrmask
= 0x7;
2099 ir3_split_dest(b
, dst
, sam
, 0, 3);
2101 /* we need to sample the alpha separately with a non-ASTC
2104 sam
= ir3_SAM(b
, opc
, type
, 0b1000, flags
,
2105 tex_idx
, tex_idx
, col0
, col1
);
2107 array_insert(ctx
->ir
, ctx
->ir
->astc_srgb
, sam
);
2109 /* fixup .w component: */
2110 ir3_split_dest(b
, &dst
[3], sam
, 3, 1);
2112 /* normal (non-workaround) case: */
2113 ir3_split_dest(b
, dst
, sam
, 0, 4);
2116 /* GETLOD returns results in 4.8 fixed point */
2117 if (opc
== OPC_GETLOD
) {
2118 struct ir3_instruction
*factor
= create_immed(b
, fui(1.0 / 256));
2120 compile_assert(ctx
, tex
->dest_type
== nir_type_float
);
2121 for (i
= 0; i
< 2; i
++) {
2122 dst
[i
] = ir3_MUL_F(b
, ir3_COV(b
, dst
[i
], TYPE_U32
, TYPE_F32
), 0,
2127 put_dst(ctx
, &tex
->dest
);
2131 emit_tex_query_levels(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2133 struct ir3_block
*b
= ctx
->block
;
2134 struct ir3_instruction
**dst
, *sam
;
2136 dst
= ir3_get_dst(ctx
, &tex
->dest
, 1);
2138 sam
= ir3_SAM(b
, OPC_GETINFO
, TYPE_U32
, 0b0100, 0,
2139 tex
->texture_index
, tex
->texture_index
, NULL
, NULL
);
2141 /* even though there is only one component, since it ends
2142 * up in .z rather than .x, we need a split_dest()
2144 ir3_split_dest(b
, dst
, sam
, 0, 3);
2146 /* The # of levels comes from getinfo.z. We need to add 1 to it, since
2147 * the value in TEX_CONST_0 is zero-based.
2149 if (ctx
->compiler
->levels_add_one
)
2150 dst
[0] = ir3_ADD_U(b
, dst
[0], 0, create_immed(b
, 1), 0);
2152 put_dst(ctx
, &tex
->dest
);
2156 emit_tex_txs(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2158 struct ir3_block
*b
= ctx
->block
;
2159 struct ir3_instruction
**dst
, *sam
;
2160 struct ir3_instruction
*lod
;
2161 unsigned flags
, coords
;
2163 tex_info(tex
, &flags
, &coords
);
2165 /* Actually we want the number of dimensions, not coordinates. This
2166 * distinction only matters for cubes.
2168 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2171 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
2173 compile_assert(ctx
, tex
->num_srcs
== 1);
2174 compile_assert(ctx
, tex
->src
[0].src_type
== nir_tex_src_lod
);
2176 lod
= ir3_get_src(ctx
, &tex
->src
[0].src
)[0];
2178 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
2179 tex
->texture_index
, tex
->texture_index
, lod
, NULL
);
2181 ir3_split_dest(b
, dst
, sam
, 0, 4);
2183 /* Array size actually ends up in .w rather than .z. This doesn't
2184 * matter for miplevel 0, but for higher mips the value in z is
2185 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2186 * returned, which means that we have to add 1 to it for arrays.
2188 if (tex
->is_array
) {
2189 if (ctx
->compiler
->levels_add_one
) {
2190 dst
[coords
] = ir3_ADD_U(b
, dst
[3], 0, create_immed(b
, 1), 0);
2192 dst
[coords
] = ir3_MOV(b
, dst
[3], TYPE_U32
);
2196 put_dst(ctx
, &tex
->dest
);
2200 emit_jump(struct ir3_context
*ctx
, nir_jump_instr
*jump
)
2202 switch (jump
->type
) {
2203 case nir_jump_break
:
2204 case nir_jump_continue
:
2205 case nir_jump_return
:
2206 /* I *think* we can simply just ignore this, and use the
2207 * successor block link to figure out where we need to
2208 * jump to for break/continue
2212 ir3_context_error(ctx
, "Unhandled NIR jump type: %d\n", jump
->type
);
2218 emit_instr(struct ir3_context
*ctx
, nir_instr
*instr
)
2220 switch (instr
->type
) {
2221 case nir_instr_type_alu
:
2222 emit_alu(ctx
, nir_instr_as_alu(instr
));
2224 case nir_instr_type_deref
:
2225 /* ignored, handled as part of the intrinsic they are src to */
2227 case nir_instr_type_intrinsic
:
2228 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2230 case nir_instr_type_load_const
:
2231 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2233 case nir_instr_type_ssa_undef
:
2234 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
2236 case nir_instr_type_tex
: {
2237 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
2238 /* couple tex instructions get special-cased:
2242 emit_tex_txs(ctx
, tex
);
2244 case nir_texop_query_levels
:
2245 emit_tex_query_levels(ctx
, tex
);
2253 case nir_instr_type_jump
:
2254 emit_jump(ctx
, nir_instr_as_jump(instr
));
2256 case nir_instr_type_phi
:
2257 /* we have converted phi webs to regs in NIR by now */
2258 ir3_context_error(ctx
, "Unexpected NIR instruction type: %d\n", instr
->type
);
2260 case nir_instr_type_call
:
2261 case nir_instr_type_parallel_copy
:
2262 ir3_context_error(ctx
, "Unhandled NIR instruction type: %d\n", instr
->type
);
2267 static struct ir3_block
*
2268 get_block(struct ir3_context
*ctx
, const nir_block
*nblock
)
2270 struct ir3_block
*block
;
2271 struct hash_entry
*hentry
;
2274 hentry
= _mesa_hash_table_search(ctx
->block_ht
, nblock
);
2276 return hentry
->data
;
2278 block
= ir3_block_create(ctx
->ir
);
2279 block
->nblock
= nblock
;
2280 _mesa_hash_table_insert(ctx
->block_ht
, nblock
, block
);
2282 block
->predecessors_count
= nblock
->predecessors
->entries
;
2283 block
->predecessors
= ralloc_array_size(block
,
2284 sizeof(block
->predecessors
[0]), block
->predecessors_count
);
2286 set_foreach(nblock
->predecessors
, sentry
) {
2287 block
->predecessors
[i
++] = get_block(ctx
, sentry
->key
);
2294 emit_block(struct ir3_context
*ctx
, nir_block
*nblock
)
2296 struct ir3_block
*block
= get_block(ctx
, nblock
);
2298 for (int i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
2299 if (nblock
->successors
[i
]) {
2300 block
->successors
[i
] =
2301 get_block(ctx
, nblock
->successors
[i
]);
2306 list_addtail(&block
->node
, &ctx
->ir
->block_list
);
2308 /* re-emit addr register in each block if needed: */
2309 for (int i
= 0; i
< ARRAY_SIZE(ctx
->addr_ht
); i
++) {
2310 _mesa_hash_table_destroy(ctx
->addr_ht
[i
], NULL
);
2311 ctx
->addr_ht
[i
] = NULL
;
2314 nir_foreach_instr(instr
, nblock
) {
2315 ctx
->cur_instr
= instr
;
2316 emit_instr(ctx
, instr
);
2317 ctx
->cur_instr
= NULL
;
2323 static void emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
);
2326 emit_if(struct ir3_context
*ctx
, nir_if
*nif
)
2328 struct ir3_instruction
*condition
= ir3_get_src(ctx
, &nif
->condition
)[0];
2330 ctx
->block
->condition
=
2331 ir3_get_predicate(ctx
, ir3_b2n(condition
->block
, condition
));
2333 emit_cf_list(ctx
, &nif
->then_list
);
2334 emit_cf_list(ctx
, &nif
->else_list
);
2338 emit_loop(struct ir3_context
*ctx
, nir_loop
*nloop
)
2340 emit_cf_list(ctx
, &nloop
->body
);
2344 stack_push(struct ir3_context
*ctx
)
2347 ctx
->max_stack
= MAX2(ctx
->max_stack
, ctx
->stack
);
2351 stack_pop(struct ir3_context
*ctx
)
2353 compile_assert(ctx
, ctx
->stack
> 0);
2358 emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
)
2360 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2361 switch (node
->type
) {
2362 case nir_cf_node_block
:
2363 emit_block(ctx
, nir_cf_node_as_block(node
));
2365 case nir_cf_node_if
:
2367 emit_if(ctx
, nir_cf_node_as_if(node
));
2370 case nir_cf_node_loop
:
2372 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2375 case nir_cf_node_function
:
2376 ir3_context_error(ctx
, "TODO\n");
2382 /* emit stream-out code. At this point, the current block is the original
2383 * (nir) end block, and nir ensures that all flow control paths terminate
2384 * into the end block. We re-purpose the original end block to generate
2385 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2386 * block holding stream-out write instructions, followed by the new end
2390 * p0.x = (vtxcnt < maxvtxcnt)
2391 * // succs: blockStreamOut, blockNewEnd
2394 * ... stream-out instructions ...
2395 * // succs: blockNewEnd
2401 emit_stream_out(struct ir3_context
*ctx
)
2403 struct ir3_shader_variant
*v
= ctx
->so
;
2404 struct ir3
*ir
= ctx
->ir
;
2405 struct ir3_stream_output_info
*strmout
=
2406 &ctx
->so
->shader
->stream_output
;
2407 struct ir3_block
*orig_end_block
, *stream_out_block
, *new_end_block
;
2408 struct ir3_instruction
*vtxcnt
, *maxvtxcnt
, *cond
;
2409 struct ir3_instruction
*bases
[IR3_MAX_SO_BUFFERS
];
2411 /* create vtxcnt input in input block at top of shader,
2412 * so that it is seen as live over the entire duration
2415 vtxcnt
= create_input(ctx
, 0);
2416 add_sysval_input(ctx
, SYSTEM_VALUE_VERTEX_CNT
, vtxcnt
);
2418 maxvtxcnt
= create_driver_param(ctx
, IR3_DP_VTXCNT_MAX
);
2420 /* at this point, we are at the original 'end' block,
2421 * re-purpose this block to stream-out condition, then
2422 * append stream-out block and new-end block
2424 orig_end_block
= ctx
->block
;
2426 // TODO these blocks need to update predecessors..
2427 // maybe w/ store_global intrinsic, we could do this
2428 // stuff in nir->nir pass
2430 stream_out_block
= ir3_block_create(ir
);
2431 list_addtail(&stream_out_block
->node
, &ir
->block_list
);
2433 new_end_block
= ir3_block_create(ir
);
2434 list_addtail(&new_end_block
->node
, &ir
->block_list
);
2436 orig_end_block
->successors
[0] = stream_out_block
;
2437 orig_end_block
->successors
[1] = new_end_block
;
2438 stream_out_block
->successors
[0] = new_end_block
;
2440 /* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2441 cond
= ir3_CMPS_S(ctx
->block
, vtxcnt
, 0, maxvtxcnt
, 0);
2442 cond
->regs
[0]->num
= regid(REG_P0
, 0);
2443 cond
->cat2
.condition
= IR3_COND_LT
;
2445 /* condition goes on previous block to the conditional,
2446 * since it is used to pick which of the two successor
2449 orig_end_block
->condition
= cond
;
2451 /* switch to stream_out_block to generate the stream-out
2454 ctx
->block
= stream_out_block
;
2456 /* Calculate base addresses based on vtxcnt. Instructions
2457 * generated for bases not used in following loop will be
2458 * stripped out in the backend.
2460 for (unsigned i
= 0; i
< IR3_MAX_SO_BUFFERS
; i
++) {
2461 unsigned stride
= strmout
->stride
[i
];
2462 struct ir3_instruction
*base
, *off
;
2464 base
= create_uniform(ctx
->block
, regid(v
->constbase
.tfbo
, i
));
2466 /* 24-bit should be enough: */
2467 off
= ir3_MUL_U(ctx
->block
, vtxcnt
, 0,
2468 create_immed(ctx
->block
, stride
* 4), 0);
2470 bases
[i
] = ir3_ADD_S(ctx
->block
, off
, 0, base
, 0);
2473 /* Generate the per-output store instructions: */
2474 for (unsigned i
= 0; i
< strmout
->num_outputs
; i
++) {
2475 for (unsigned j
= 0; j
< strmout
->output
[i
].num_components
; j
++) {
2476 unsigned c
= j
+ strmout
->output
[i
].start_component
;
2477 struct ir3_instruction
*base
, *out
, *stg
;
2479 base
= bases
[strmout
->output
[i
].output_buffer
];
2480 out
= ctx
->ir
->outputs
[regid(strmout
->output
[i
].register_index
, c
)];
2482 stg
= ir3_STG(ctx
->block
, base
, 0, out
, 0,
2483 create_immed(ctx
->block
, 1), 0);
2484 stg
->cat6
.type
= TYPE_U32
;
2485 stg
->cat6
.dst_offset
= (strmout
->output
[i
].dst_offset
+ j
) * 4;
2487 array_insert(ctx
->block
, ctx
->block
->keeps
, stg
);
2491 /* and finally switch to the new_end_block: */
2492 ctx
->block
= new_end_block
;
2496 emit_function(struct ir3_context
*ctx
, nir_function_impl
*impl
)
2498 nir_metadata_require(impl
, nir_metadata_block_index
);
2500 compile_assert(ctx
, ctx
->stack
== 0);
2502 emit_cf_list(ctx
, &impl
->body
);
2503 emit_block(ctx
, impl
->end_block
);
2505 compile_assert(ctx
, ctx
->stack
== 0);
2507 /* at this point, we should have a single empty block,
2508 * into which we emit the 'end' instruction.
2510 compile_assert(ctx
, list_empty(&ctx
->block
->instr_list
));
2512 /* If stream-out (aka transform-feedback) enabled, emit the
2513 * stream-out instructions, followed by a new empty block (into
2514 * which the 'end' instruction lands).
2516 * NOTE: it is done in this order, rather than inserting before
2517 * we emit end_block, because NIR guarantees that all blocks
2518 * flow into end_block, and that end_block has no successors.
2519 * So by re-purposing end_block as the first block of stream-
2520 * out, we guarantee that all exit paths flow into the stream-
2523 if ((ctx
->compiler
->gpu_id
< 500) &&
2524 (ctx
->so
->shader
->stream_output
.num_outputs
> 0) &&
2525 !ctx
->so
->binning_pass
) {
2526 debug_assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
2527 emit_stream_out(ctx
);
2530 ir3_END(ctx
->block
);
2533 static struct ir3_instruction
*
2534 create_frag_coord(struct ir3_context
*ctx
, unsigned comp
)
2536 struct ir3_block
*block
= ctx
->block
;
2537 struct ir3_instruction
*instr
;
2539 if (!ctx
->frag_coord
) {
2540 ctx
->frag_coord
= create_input_compmask(ctx
, 0, 0xf);
2541 /* defer add_sysval_input() until after all inputs created */
2544 ir3_split_dest(block
, &instr
, ctx
->frag_coord
, comp
, 1);
2549 /* for frag_coord, we get unsigned values.. we need
2550 * to subtract (integer) 8 and divide by 16 (right-
2551 * shift by 4) then convert to float:
2555 * mov.u32f32 dst, tmp
2558 instr
= ir3_SUB_S(block
, instr
, 0,
2559 create_immed(block
, 8), 0);
2560 instr
= ir3_SHR_B(block
, instr
, 0,
2561 create_immed(block
, 4), 0);
2562 instr
= ir3_COV(block
, instr
, TYPE_U32
, TYPE_F32
);
2568 /* seems that we can use these as-is: */
2574 setup_input(struct ir3_context
*ctx
, nir_variable
*in
)
2576 struct ir3_shader_variant
*so
= ctx
->so
;
2577 unsigned ncomp
= glsl_get_components(in
->type
);
2578 unsigned n
= in
->data
.driver_location
;
2579 unsigned slot
= in
->data
.location
;
2581 /* let's pretend things other than vec4 don't exist: */
2582 ncomp
= MAX2(ncomp
, 4);
2584 /* skip unread inputs, we could end up with (for example), unsplit
2585 * matrix/etc inputs in the case they are not read, so just silently
2591 compile_assert(ctx
, ncomp
== 4);
2593 so
->inputs
[n
].slot
= slot
;
2594 so
->inputs
[n
].compmask
= (1 << ncomp
) - 1;
2595 so
->inputs_count
= MAX2(so
->inputs_count
, n
+ 1);
2596 so
->inputs
[n
].interpolate
= in
->data
.interpolation
;
2598 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2599 for (int i
= 0; i
< ncomp
; i
++) {
2600 struct ir3_instruction
*instr
= NULL
;
2601 unsigned idx
= (n
* 4) + i
;
2603 if (slot
== VARYING_SLOT_POS
) {
2604 so
->inputs
[n
].bary
= false;
2605 so
->frag_coord
= true;
2606 instr
= create_frag_coord(ctx
, i
);
2607 } else if (slot
== VARYING_SLOT_PNTC
) {
2608 /* see for example st_nir_fixup_varying_slots().. this is
2609 * maybe a bit mesa/st specific. But we need things to line
2610 * up for this in fdN_program:
2611 * unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
2612 * if (emit->sprite_coord_enable & texmask) {
2616 so
->inputs
[n
].slot
= VARYING_SLOT_VAR8
;
2617 so
->inputs
[n
].bary
= true;
2618 instr
= create_frag_input(ctx
, false);
2620 bool use_ldlv
= false;
2622 /* detect the special case for front/back colors where
2623 * we need to do flat vs smooth shading depending on
2626 if (in
->data
.interpolation
== INTERP_MODE_NONE
) {
2628 case VARYING_SLOT_COL0
:
2629 case VARYING_SLOT_COL1
:
2630 case VARYING_SLOT_BFC0
:
2631 case VARYING_SLOT_BFC1
:
2632 so
->inputs
[n
].rasterflat
= true;
2639 if (ctx
->compiler
->flat_bypass
) {
2640 if ((so
->inputs
[n
].interpolate
== INTERP_MODE_FLAT
) ||
2641 (so
->inputs
[n
].rasterflat
&& ctx
->so
->key
.rasterflat
))
2645 so
->inputs
[n
].bary
= true;
2647 instr
= create_frag_input(ctx
, use_ldlv
);
2650 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2652 ctx
->ir
->inputs
[idx
] = instr
;
2654 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2655 for (int i
= 0; i
< ncomp
; i
++) {
2656 unsigned idx
= (n
* 4) + i
;
2657 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2658 ctx
->ir
->inputs
[idx
] = create_input(ctx
, idx
);
2661 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2664 if (so
->inputs
[n
].bary
|| (ctx
->so
->type
== MESA_SHADER_VERTEX
)) {
2665 so
->total_in
+= ncomp
;
2670 setup_output(struct ir3_context
*ctx
, nir_variable
*out
)
2672 struct ir3_shader_variant
*so
= ctx
->so
;
2673 unsigned ncomp
= glsl_get_components(out
->type
);
2674 unsigned n
= out
->data
.driver_location
;
2675 unsigned slot
= out
->data
.location
;
2678 /* let's pretend things other than vec4 don't exist: */
2679 ncomp
= MAX2(ncomp
, 4);
2680 compile_assert(ctx
, ncomp
== 4);
2682 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2684 case FRAG_RESULT_DEPTH
:
2685 comp
= 2; /* tgsi will write to .z component */
2686 so
->writes_pos
= true;
2688 case FRAG_RESULT_COLOR
:
2692 if (slot
>= FRAG_RESULT_DATA0
)
2694 ir3_context_error(ctx
, "unknown FS output name: %s\n",
2695 gl_frag_result_name(slot
));
2697 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2699 case VARYING_SLOT_POS
:
2700 so
->writes_pos
= true;
2702 case VARYING_SLOT_PSIZ
:
2703 so
->writes_psize
= true;
2705 case VARYING_SLOT_COL0
:
2706 case VARYING_SLOT_COL1
:
2707 case VARYING_SLOT_BFC0
:
2708 case VARYING_SLOT_BFC1
:
2709 case VARYING_SLOT_FOGC
:
2710 case VARYING_SLOT_CLIP_DIST0
:
2711 case VARYING_SLOT_CLIP_DIST1
:
2712 case VARYING_SLOT_CLIP_VERTEX
:
2715 if (slot
>= VARYING_SLOT_VAR0
)
2717 if ((VARYING_SLOT_TEX0
<= slot
) && (slot
<= VARYING_SLOT_TEX7
))
2719 ir3_context_error(ctx
, "unknown VS output name: %s\n",
2720 gl_varying_slot_name(slot
));
2723 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2726 compile_assert(ctx
, n
< ARRAY_SIZE(so
->outputs
));
2728 so
->outputs
[n
].slot
= slot
;
2729 so
->outputs
[n
].regid
= regid(n
, comp
);
2730 so
->outputs_count
= MAX2(so
->outputs_count
, n
+ 1);
2732 for (int i
= 0; i
< ncomp
; i
++) {
2733 unsigned idx
= (n
* 4) + i
;
2734 compile_assert(ctx
, idx
< ctx
->ir
->noutputs
);
2735 ctx
->ir
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
2740 max_drvloc(struct exec_list
*vars
)
2743 nir_foreach_variable(var
, vars
) {
2744 drvloc
= MAX2(drvloc
, (int)var
->data
.driver_location
);
2749 static const unsigned max_sysvals
[] = {
2750 [MESA_SHADER_FRAGMENT
] = 24, // TODO
2751 [MESA_SHADER_VERTEX
] = 16,
2752 [MESA_SHADER_COMPUTE
] = 16, // TODO how many do we actually need?
2756 emit_instructions(struct ir3_context
*ctx
)
2758 unsigned ninputs
, noutputs
;
2759 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->s
);
2761 ninputs
= (max_drvloc(&ctx
->s
->inputs
) + 1) * 4;
2762 noutputs
= (max_drvloc(&ctx
->s
->outputs
) + 1) * 4;
2764 /* we need to leave room for sysvals:
2766 ninputs
+= max_sysvals
[ctx
->so
->type
];
2768 ctx
->ir
= ir3_create(ctx
->compiler
, ninputs
, noutputs
);
2770 /* Create inputs in first block: */
2771 ctx
->block
= get_block(ctx
, nir_start_block(fxn
));
2772 ctx
->in_block
= ctx
->block
;
2773 list_addtail(&ctx
->block
->node
, &ctx
->ir
->block_list
);
2775 ninputs
-= max_sysvals
[ctx
->so
->type
];
2777 /* for fragment shader, the vcoord input register is used as the
2778 * base for bary.f varying fetch instrs:
2780 struct ir3_instruction
*vcoord
= NULL
;
2781 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2782 struct ir3_instruction
*xy
[2];
2784 vcoord
= create_input_compmask(ctx
, 0, 0x3);
2785 ir3_split_dest(ctx
->block
, xy
, vcoord
, 0, 2);
2787 ctx
->frag_vcoord
= ir3_create_collect(ctx
, xy
, 2);
2791 nir_foreach_variable(var
, &ctx
->s
->inputs
) {
2792 setup_input(ctx
, var
);
2795 /* Defer add_sysval_input() stuff until after setup_inputs(),
2796 * because sysvals need to be appended after varyings:
2799 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_VARYING_COORD
,
2803 if (ctx
->frag_coord
) {
2804 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_FRAG_COORD
,
2805 0xf, ctx
->frag_coord
);
2808 /* Setup outputs: */
2809 nir_foreach_variable(var
, &ctx
->s
->outputs
) {
2810 setup_output(ctx
, var
);
2813 /* Setup registers (which should only be arrays): */
2814 nir_foreach_register(reg
, &ctx
->s
->registers
) {
2815 ir3_declare_array(ctx
, reg
);
2818 /* NOTE: need to do something more clever when we support >1 fxn */
2819 nir_foreach_register(reg
, &fxn
->registers
) {
2820 ir3_declare_array(ctx
, reg
);
2822 /* And emit the body: */
2824 emit_function(ctx
, fxn
);
2827 /* from NIR perspective, we actually have varying inputs. But the varying
2828 * inputs, from an IR standpoint, are just bary.f/ldlv instructions. The
2829 * only actual inputs are the sysvals.
2832 fixup_frag_inputs(struct ir3_context
*ctx
)
2834 struct ir3_shader_variant
*so
= ctx
->so
;
2835 struct ir3
*ir
= ctx
->ir
;
2838 /* sysvals should appear at the end of the inputs, drop everything else: */
2839 while ((i
< so
->inputs_count
) && !so
->inputs
[i
].sysval
)
2842 /* at IR level, inputs are always blocks of 4 scalars: */
2845 ir
->inputs
= &ir
->inputs
[i
];
2849 /* Fixup tex sampler state for astc/srgb workaround instructions. We
2850 * need to assign the tex state indexes for these after we know the
2854 fixup_astc_srgb(struct ir3_context
*ctx
)
2856 struct ir3_shader_variant
*so
= ctx
->so
;
2857 /* indexed by original tex idx, value is newly assigned alpha sampler
2858 * state tex idx. Zero is invalid since there is at least one sampler
2861 unsigned alt_tex_state
[16] = {0};
2862 unsigned tex_idx
= ctx
->max_texture_index
+ 1;
2865 so
->astc_srgb
.base
= tex_idx
;
2867 for (unsigned i
= 0; i
< ctx
->ir
->astc_srgb_count
; i
++) {
2868 struct ir3_instruction
*sam
= ctx
->ir
->astc_srgb
[i
];
2870 compile_assert(ctx
, sam
->cat5
.tex
< ARRAY_SIZE(alt_tex_state
));
2872 if (alt_tex_state
[sam
->cat5
.tex
] == 0) {
2873 /* assign new alternate/alpha tex state slot: */
2874 alt_tex_state
[sam
->cat5
.tex
] = tex_idx
++;
2875 so
->astc_srgb
.orig_idx
[idx
++] = sam
->cat5
.tex
;
2876 so
->astc_srgb
.count
++;
2879 sam
->cat5
.tex
= alt_tex_state
[sam
->cat5
.tex
];
2884 fixup_binning_pass(struct ir3_context
*ctx
)
2886 struct ir3_shader_variant
*so
= ctx
->so
;
2887 struct ir3
*ir
= ctx
->ir
;
2890 for (i
= 0, j
= 0; i
< so
->outputs_count
; i
++) {
2891 unsigned slot
= so
->outputs
[i
].slot
;
2893 /* throw away everything but first position/psize */
2894 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
2896 so
->outputs
[j
] = so
->outputs
[i
];
2897 ir
->outputs
[(j
*4)+0] = ir
->outputs
[(i
*4)+0];
2898 ir
->outputs
[(j
*4)+1] = ir
->outputs
[(i
*4)+1];
2899 ir
->outputs
[(j
*4)+2] = ir
->outputs
[(i
*4)+2];
2900 ir
->outputs
[(j
*4)+3] = ir
->outputs
[(i
*4)+3];
2905 so
->outputs_count
= j
;
2906 ir
->noutputs
= j
* 4;
2910 ir3_compile_shader_nir(struct ir3_compiler
*compiler
,
2911 struct ir3_shader_variant
*so
)
2913 struct ir3_context
*ctx
;
2915 struct ir3_instruction
**inputs
;
2916 unsigned i
, actual_in
, inloc
;
2917 int ret
= 0, max_bary
;
2921 ctx
= ir3_context_init(compiler
, so
);
2923 DBG("INIT failed!");
2928 emit_instructions(ctx
);
2931 DBG("EMIT failed!");
2936 ir
= so
->ir
= ctx
->ir
;
2938 /* keep track of the inputs from TGSI perspective.. */
2939 inputs
= ir
->inputs
;
2941 /* but fixup actual inputs for frag shader: */
2942 if (so
->type
== MESA_SHADER_FRAGMENT
)
2943 fixup_frag_inputs(ctx
);
2945 /* at this point, for binning pass, throw away unneeded outputs: */
2946 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
< 600))
2947 fixup_binning_pass(ctx
);
2949 /* if we want half-precision outputs, mark the output registers
2952 if (so
->key
.half_precision
) {
2953 for (i
= 0; i
< ir
->noutputs
; i
++) {
2954 struct ir3_instruction
*out
= ir
->outputs
[i
];
2959 /* if frag shader writes z, that needs to be full precision: */
2960 if (so
->outputs
[i
/4].slot
== FRAG_RESULT_DEPTH
)
2963 out
->regs
[0]->flags
|= IR3_REG_HALF
;
2964 /* output could be a fanout (ie. texture fetch output)
2965 * in which case we need to propagate the half-reg flag
2966 * up to the definer so that RA sees it:
2968 if (out
->opc
== OPC_META_FO
) {
2969 out
= out
->regs
[1]->instr
;
2970 out
->regs
[0]->flags
|= IR3_REG_HALF
;
2973 if (out
->opc
== OPC_MOV
) {
2974 out
->cat1
.dst_type
= half_type(out
->cat1
.dst_type
);
2979 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2980 printf("BEFORE CP:\n");
2986 /* at this point, for binning pass, throw away unneeded outputs:
2987 * Note that for a6xx and later, we do this after ir3_cp to ensure
2988 * that the uniform/constant layout for BS and VS matches, so that
2989 * we can re-use same VS_CONST state group.
2991 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
>= 600))
2992 fixup_binning_pass(ctx
);
2994 /* Insert mov if there's same instruction for each output.
2995 * eg. dEQP-GLES31.functional.shaders.opaque_type_indexing.sampler.const_expression.vertex.sampler2dshadow
2997 for (int i
= ir
->noutputs
- 1; i
>= 0; i
--) {
2998 if (!ir
->outputs
[i
])
3000 for (unsigned j
= 0; j
< i
; j
++) {
3001 if (ir
->outputs
[i
] == ir
->outputs
[j
]) {
3003 ir3_MOV(ir
->outputs
[i
]->block
, ir
->outputs
[i
], TYPE_F32
);
3008 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3009 printf("BEFORE GROUPING:\n");
3013 ir3_sched_add_deps(ir
);
3015 /* Group left/right neighbors, inserting mov's where needed to
3020 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3021 printf("AFTER GROUPING:\n");
3027 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3028 printf("AFTER DEPTH:\n");
3032 ret
= ir3_sched(ir
);
3034 DBG("SCHED failed!");
3038 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3039 printf("AFTER SCHED:\n");
3043 ret
= ir3_ra(ir
, so
->type
, so
->frag_coord
, so
->frag_face
);
3049 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3050 printf("AFTER RA:\n");
3054 /* fixup input/outputs: */
3055 for (i
= 0; i
< so
->outputs_count
; i
++) {
3056 so
->outputs
[i
].regid
= ir
->outputs
[i
*4]->regs
[0]->num
;
3059 /* Note that some or all channels of an input may be unused: */
3062 for (i
= 0; i
< so
->inputs_count
; i
++) {
3063 unsigned j
, reg
= regid(63,0), compmask
= 0, maxcomp
= 0;
3064 so
->inputs
[i
].ncomp
= 0;
3065 so
->inputs
[i
].inloc
= inloc
;
3066 for (j
= 0; j
< 4; j
++) {
3067 struct ir3_instruction
*in
= inputs
[(i
*4) + j
];
3068 if (in
&& !(in
->flags
& IR3_INSTR_UNUSED
)) {
3069 compmask
|= (1 << j
);
3070 reg
= in
->regs
[0]->num
- j
;
3072 so
->inputs
[i
].ncomp
++;
3073 if ((so
->type
== MESA_SHADER_FRAGMENT
) && so
->inputs
[i
].bary
) {
3075 assert(in
->regs
[1]->flags
& IR3_REG_IMMED
);
3076 in
->regs
[1]->iim_val
= inloc
+ j
;
3081 if ((so
->type
== MESA_SHADER_FRAGMENT
) && compmask
&& so
->inputs
[i
].bary
) {
3083 so
->inputs
[i
].compmask
= (1 << maxcomp
) - 1;
3085 } else if (!so
->inputs
[i
].sysval
) {
3086 so
->inputs
[i
].compmask
= compmask
;
3088 so
->inputs
[i
].regid
= reg
;
3092 fixup_astc_srgb(ctx
);
3094 /* We need to do legalize after (for frag shader's) the "bary.f"
3095 * offsets (inloc) have been assigned.
3097 ir3_legalize(ir
, &so
->num_samp
, &so
->has_ssbo
, &max_bary
);
3099 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3100 printf("AFTER LEGALIZE:\n");
3104 so
->branchstack
= ctx
->max_stack
;
3106 /* Note that actual_in counts inputs that are not bary.f'd for FS: */
3107 if (so
->type
== MESA_SHADER_VERTEX
)
3108 so
->total_in
= actual_in
;
3110 so
->total_in
= max_bary
+ 1;
3115 ir3_destroy(so
->ir
);
3118 ir3_context_free(ctx
);