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_image.h"
35 #include "ir3_shader.h"
38 #include "instr-a3xx.h"
40 #include "ir3_context.h"
43 static struct ir3_instruction
*
44 create_indirect_load(struct ir3_context
*ctx
, unsigned arrsz
, int n
,
45 struct ir3_instruction
*address
, struct ir3_instruction
*collect
)
47 struct ir3_block
*block
= ctx
->block
;
48 struct ir3_instruction
*mov
;
49 struct ir3_register
*src
;
51 mov
= ir3_instr_create(block
, OPC_MOV
);
52 mov
->cat1
.src_type
= TYPE_U32
;
53 mov
->cat1
.dst_type
= TYPE_U32
;
54 ir3_reg_create(mov
, 0, 0);
55 src
= ir3_reg_create(mov
, 0, IR3_REG_SSA
| IR3_REG_RELATIV
);
58 src
->array
.offset
= n
;
60 ir3_instr_set_address(mov
, address
);
65 static struct ir3_instruction
*
66 create_input_compmask(struct ir3_context
*ctx
, unsigned n
, unsigned compmask
)
68 struct ir3_instruction
*in
;
70 in
= ir3_instr_create(ctx
->in_block
, OPC_META_INPUT
);
71 in
->input
.sysval
= ~0;
72 ir3_reg_create(in
, n
, 0);
74 in
->regs
[0]->wrmask
= compmask
;
79 static struct ir3_instruction
*
80 create_input(struct ir3_context
*ctx
, unsigned n
)
82 return create_input_compmask(ctx
, n
, 0x1);
85 static struct ir3_instruction
*
86 create_frag_input(struct ir3_context
*ctx
, bool use_ldlv
, unsigned n
)
88 struct ir3_block
*block
= ctx
->block
;
89 struct ir3_instruction
*instr
;
90 /* packed inloc is fixed up later: */
91 struct ir3_instruction
*inloc
= create_immed(block
, n
);
94 instr
= ir3_LDLV(block
, inloc
, 0, create_immed(block
, 1), 0);
95 instr
->cat6
.type
= TYPE_U32
;
96 instr
->cat6
.iim_val
= 1;
98 instr
= ir3_BARY_F(block
, inloc
, 0, ctx
->ij_pixel
, 0);
99 instr
->regs
[2]->wrmask
= 0x3;
105 static struct ir3_instruction
*
106 create_driver_param(struct ir3_context
*ctx
, enum ir3_driver_param dp
)
108 /* first four vec4 sysval's reserved for UBOs: */
109 /* NOTE: dp is in scalar, but there can be >4 dp components: */
110 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
111 unsigned n
= const_state
->offsets
.driver_param
;
112 unsigned r
= regid(n
+ dp
/ 4, dp
% 4);
113 return create_uniform(ctx
->block
, r
);
117 * Adreno uses uint rather than having dedicated bool type,
118 * which (potentially) requires some conversion, in particular
119 * when using output of an bool instr to int input, or visa
123 * -------+---------+-------+-
127 * To convert from an adreno bool (uint) to nir, use:
129 * absneg.s dst, (neg)src
131 * To convert back in the other direction:
133 * absneg.s dst, (abs)arc
135 * The CP step can clean up the absneg.s that cancel each other
136 * out, and with a slight bit of extra cleverness (to recognize
137 * the instructions which produce either a 0 or 1) can eliminate
138 * the absneg.s's completely when an instruction that wants
139 * 0/1 consumes the result. For example, when a nir 'bcsel'
140 * consumes the result of 'feq'. So we should be able to get by
141 * without a boolean resolve step, and without incuring any
142 * extra penalty in instruction count.
145 /* NIR bool -> native (adreno): */
146 static struct ir3_instruction
*
147 ir3_b2n(struct ir3_block
*block
, struct ir3_instruction
*instr
)
149 return ir3_ABSNEG_S(block
, instr
, IR3_REG_SABS
);
152 /* native (adreno) -> NIR bool: */
153 static struct ir3_instruction
*
154 ir3_n2b(struct ir3_block
*block
, struct ir3_instruction
*instr
)
156 return ir3_ABSNEG_S(block
, instr
, IR3_REG_SNEG
);
160 * alu/sfu instructions:
163 static struct ir3_instruction
*
164 create_cov(struct ir3_context
*ctx
, struct ir3_instruction
*src
,
165 unsigned src_bitsize
, nir_op op
)
167 type_t src_type
, dst_type
;
171 case nir_op_f2f16_rtne
:
172 case nir_op_f2f16_rtz
:
180 switch (src_bitsize
) {
188 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
197 switch (src_bitsize
) {
208 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
217 switch (src_bitsize
) {
228 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
233 ir3_context_error(ctx
, "invalid conversion op: %u", op
);
243 case nir_op_f2f16_rtne
:
244 case nir_op_f2f16_rtz
:
246 /* TODO how to handle rounding mode? */
283 ir3_context_error(ctx
, "invalid conversion op: %u", op
);
286 return ir3_COV(ctx
->block
, src
, src_type
, dst_type
);
290 emit_alu(struct ir3_context
*ctx
, nir_alu_instr
*alu
)
292 const nir_op_info
*info
= &nir_op_infos
[alu
->op
];
293 struct ir3_instruction
**dst
, *src
[info
->num_inputs
];
294 unsigned bs
[info
->num_inputs
]; /* bit size */
295 struct ir3_block
*b
= ctx
->block
;
296 unsigned dst_sz
, wrmask
;
297 type_t dst_type
= nir_dest_bit_size(alu
->dest
.dest
) < 32 ?
300 if (alu
->dest
.dest
.is_ssa
) {
301 dst_sz
= alu
->dest
.dest
.ssa
.num_components
;
302 wrmask
= (1 << dst_sz
) - 1;
304 dst_sz
= alu
->dest
.dest
.reg
.reg
->num_components
;
305 wrmask
= alu
->dest
.write_mask
;
308 dst
= ir3_get_dst(ctx
, &alu
->dest
.dest
, dst_sz
);
310 /* Vectors are special in that they have non-scalarized writemasks,
311 * and just take the first swizzle channel for each argument in
312 * order into each writemask channel.
314 if ((alu
->op
== nir_op_vec2
) ||
315 (alu
->op
== nir_op_vec3
) ||
316 (alu
->op
== nir_op_vec4
)) {
318 for (int i
= 0; i
< info
->num_inputs
; i
++) {
319 nir_alu_src
*asrc
= &alu
->src
[i
];
321 compile_assert(ctx
, !asrc
->abs
);
322 compile_assert(ctx
, !asrc
->negate
);
324 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[0]];
326 src
[i
] = create_immed_typed(ctx
->block
, 0, dst_type
);
327 dst
[i
] = ir3_MOV(b
, src
[i
], dst_type
);
330 ir3_put_dst(ctx
, &alu
->dest
.dest
);
334 /* We also get mov's with more than one component for mov's so
335 * handle those specially:
337 if (alu
->op
== nir_op_mov
) {
338 nir_alu_src
*asrc
= &alu
->src
[0];
339 struct ir3_instruction
*const *src0
= ir3_get_src(ctx
, &asrc
->src
);
341 for (unsigned i
= 0; i
< dst_sz
; i
++) {
342 if (wrmask
& (1 << i
)) {
343 dst
[i
] = ir3_MOV(b
, src0
[asrc
->swizzle
[i
]], dst_type
);
349 ir3_put_dst(ctx
, &alu
->dest
.dest
);
353 /* General case: We can just grab the one used channel per src. */
354 for (int i
= 0; i
< info
->num_inputs
; i
++) {
355 unsigned chan
= ffs(alu
->dest
.write_mask
) - 1;
356 nir_alu_src
*asrc
= &alu
->src
[i
];
358 compile_assert(ctx
, !asrc
->abs
);
359 compile_assert(ctx
, !asrc
->negate
);
361 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[chan
]];
362 bs
[i
] = nir_src_bit_size(asrc
->src
);
364 compile_assert(ctx
, src
[i
]);
369 case nir_op_f2f16_rtne
:
370 case nir_op_f2f16_rtz
:
388 dst
[0] = create_cov(ctx
, src
[0], bs
[0], alu
->op
);
390 case nir_op_fquantize2f16
:
391 dst
[0] = create_cov(ctx
,
392 create_cov(ctx
, src
[0], 32, nir_op_f2f16
),
396 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, create_immed(b
, fui(0.0)), 0);
397 dst
[0]->cat2
.condition
= IR3_COND_NE
;
398 dst
[0] = ir3_n2b(b
, dst
[0]);
401 dst
[0] = ir3_COV(b
, ir3_b2n(b
, src
[0]), TYPE_U32
, TYPE_F16
);
404 dst
[0] = ir3_COV(b
, ir3_b2n(b
, src
[0]), TYPE_U32
, TYPE_F32
);
409 dst
[0] = ir3_b2n(b
, src
[0]);
412 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, create_immed(b
, 0), 0);
413 dst
[0]->cat2
.condition
= IR3_COND_NE
;
414 dst
[0] = ir3_n2b(b
, dst
[0]);
418 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FNEG
);
421 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FABS
);
424 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[1], 0);
427 dst
[0] = ir3_MIN_F(b
, src
[0], 0, src
[1], 0);
430 /* if there is just a single use of the src, and it supports
431 * (sat) bit, we can just fold the (sat) flag back to the
432 * src instruction and create a mov. This is easier for cp
435 * TODO probably opc_cat==4 is ok too
437 if (alu
->src
[0].src
.is_ssa
&&
438 (list_length(&alu
->src
[0].src
.ssa
->uses
) == 1) &&
439 ((opc_cat(src
[0]->opc
) == 2) || (opc_cat(src
[0]->opc
) == 3))) {
440 src
[0]->flags
|= IR3_INSTR_SAT
;
441 dst
[0] = ir3_MOV(b
, src
[0], dst_type
);
443 /* otherwise generate a max.f that saturates.. blob does
444 * similar (generating a cat2 mov using max.f)
446 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[0], 0);
447 dst
[0]->flags
|= IR3_INSTR_SAT
;
451 dst
[0] = ir3_MUL_F(b
, src
[0], 0, src
[1], 0);
454 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], 0);
457 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], IR3_REG_FNEG
);
460 dst
[0] = ir3_MAD_F32(b
, src
[0], 0, src
[1], 0, src
[2], 0);
463 case nir_op_fddx_coarse
:
464 dst
[0] = ir3_DSX(b
, src
[0], 0);
465 dst
[0]->cat5
.type
= TYPE_F32
;
468 case nir_op_fddy_coarse
:
469 dst
[0] = ir3_DSY(b
, src
[0], 0);
470 dst
[0]->cat5
.type
= TYPE_F32
;
474 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
475 dst
[0]->cat2
.condition
= IR3_COND_LT
;
476 dst
[0] = ir3_n2b(b
, dst
[0]);
479 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
480 dst
[0]->cat2
.condition
= IR3_COND_GE
;
481 dst
[0] = ir3_n2b(b
, dst
[0]);
484 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
485 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
486 dst
[0] = ir3_n2b(b
, dst
[0]);
489 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
490 dst
[0]->cat2
.condition
= IR3_COND_NE
;
491 dst
[0] = ir3_n2b(b
, dst
[0]);
494 dst
[0] = ir3_CEIL_F(b
, src
[0], 0);
497 dst
[0] = ir3_FLOOR_F(b
, src
[0], 0);
500 dst
[0] = ir3_TRUNC_F(b
, src
[0], 0);
502 case nir_op_fround_even
:
503 dst
[0] = ir3_RNDNE_F(b
, src
[0], 0);
506 dst
[0] = ir3_SIGN_F(b
, src
[0], 0);
510 dst
[0] = ir3_SIN(b
, src
[0], 0);
513 dst
[0] = ir3_COS(b
, src
[0], 0);
516 dst
[0] = ir3_RSQ(b
, src
[0], 0);
519 dst
[0] = ir3_RCP(b
, src
[0], 0);
522 dst
[0] = ir3_LOG2(b
, src
[0], 0);
525 dst
[0] = ir3_EXP2(b
, src
[0], 0);
528 dst
[0] = ir3_SQRT(b
, src
[0], 0);
532 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SABS
);
535 dst
[0] = ir3_ADD_U(b
, src
[0], 0, src
[1], 0);
538 dst
[0] = ir3_AND_B(b
, src
[0], 0, src
[1], 0);
541 dst
[0] = ir3_MAX_S(b
, src
[0], 0, src
[1], 0);
544 dst
[0] = ir3_MAX_U(b
, src
[0], 0, src
[1], 0);
547 dst
[0] = ir3_MIN_S(b
, src
[0], 0, src
[1], 0);
550 dst
[0] = ir3_MIN_U(b
, src
[0], 0, src
[1], 0);
552 case nir_op_umul_low
:
553 dst
[0] = ir3_MULL_U(b
, src
[0], 0, src
[1], 0);
555 case nir_op_imadsh_mix16
:
556 dst
[0] = ir3_MADSH_M16(b
, src
[0], 0, src
[1], 0, src
[2], 0);
559 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SNEG
);
562 dst
[0] = ir3_NOT_B(b
, src
[0], 0);
565 dst
[0] = ir3_OR_B(b
, src
[0], 0, src
[1], 0);
568 dst
[0] = ir3_SHL_B(b
, src
[0], 0, src
[1], 0);
571 dst
[0] = ir3_ASHR_B(b
, src
[0], 0, src
[1], 0);
574 dst
[0] = ir3_SUB_U(b
, src
[0], 0, src
[1], 0);
577 dst
[0] = ir3_XOR_B(b
, src
[0], 0, src
[1], 0);
580 dst
[0] = ir3_SHR_B(b
, src
[0], 0, src
[1], 0);
583 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
584 dst
[0]->cat2
.condition
= IR3_COND_LT
;
585 dst
[0] = ir3_n2b(b
, dst
[0]);
588 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
589 dst
[0]->cat2
.condition
= IR3_COND_GE
;
590 dst
[0] = ir3_n2b(b
, dst
[0]);
593 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
594 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
595 dst
[0] = ir3_n2b(b
, dst
[0]);
598 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
599 dst
[0]->cat2
.condition
= IR3_COND_NE
;
600 dst
[0] = ir3_n2b(b
, dst
[0]);
603 dst
[0] = ir3_CMPS_U(b
, src
[0], 0, src
[1], 0);
604 dst
[0]->cat2
.condition
= IR3_COND_LT
;
605 dst
[0] = ir3_n2b(b
, dst
[0]);
608 dst
[0] = ir3_CMPS_U(b
, src
[0], 0, src
[1], 0);
609 dst
[0]->cat2
.condition
= IR3_COND_GE
;
610 dst
[0] = ir3_n2b(b
, dst
[0]);
613 case nir_op_b32csel
: {
614 struct ir3_instruction
*cond
= ir3_b2n(b
, src
[0]);
615 compile_assert(ctx
, bs
[1] == bs
[2]);
616 /* the boolean condition is 32b even if src[1] and src[2] are
617 * half-precision, but sel.b16 wants all three src's to be the
621 cond
= ir3_COV(b
, cond
, TYPE_U32
, TYPE_U16
);
622 dst
[0] = ir3_SEL_B32(b
, src
[1], 0, cond
, 0, src
[2], 0);
625 case nir_op_bit_count
: {
626 // TODO, we need to do this 16b at a time on a5xx+a6xx.. need to
627 // double check on earlier gen's. Once half-precision support is
628 // in place, this should probably move to a NIR lowering pass:
629 struct ir3_instruction
*hi
, *lo
;
631 hi
= ir3_COV(b
, ir3_SHR_B(b
, src
[0], 0, create_immed(b
, 16), 0),
633 lo
= ir3_COV(b
, src
[0], TYPE_U32
, TYPE_U16
);
635 hi
= ir3_CBITS_B(b
, hi
, 0);
636 lo
= ir3_CBITS_B(b
, lo
, 0);
638 // TODO maybe the builders should default to making dst half-precision
639 // if the src's were half precision, to make this less awkward.. otoh
640 // we should probably just do this lowering in NIR.
641 hi
->regs
[0]->flags
|= IR3_REG_HALF
;
642 lo
->regs
[0]->flags
|= IR3_REG_HALF
;
644 dst
[0] = ir3_ADD_S(b
, hi
, 0, lo
, 0);
645 dst
[0]->regs
[0]->flags
|= IR3_REG_HALF
;
646 dst
[0] = ir3_COV(b
, dst
[0], TYPE_U16
, TYPE_U32
);
649 case nir_op_ifind_msb
: {
650 struct ir3_instruction
*cmp
;
651 dst
[0] = ir3_CLZ_S(b
, src
[0], 0);
652 cmp
= ir3_CMPS_S(b
, dst
[0], 0, create_immed(b
, 0), 0);
653 cmp
->cat2
.condition
= IR3_COND_GE
;
654 dst
[0] = ir3_SEL_B32(b
,
655 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
659 case nir_op_ufind_msb
:
660 dst
[0] = ir3_CLZ_B(b
, src
[0], 0);
661 dst
[0] = ir3_SEL_B32(b
,
662 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
663 src
[0], 0, dst
[0], 0);
665 case nir_op_find_lsb
:
666 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
667 dst
[0] = ir3_CLZ_B(b
, dst
[0], 0);
669 case nir_op_bitfield_reverse
:
670 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
674 ir3_context_error(ctx
, "Unhandled ALU op: %s\n",
675 nir_op_infos
[alu
->op
].name
);
679 ir3_put_dst(ctx
, &alu
->dest
.dest
);
682 /* handles direct/indirect UBO reads: */
684 emit_intrinsic_load_ubo(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
685 struct ir3_instruction
**dst
)
687 struct ir3_block
*b
= ctx
->block
;
688 struct ir3_instruction
*base_lo
, *base_hi
, *addr
, *src0
, *src1
;
689 /* UBO addresses are the first driver params, but subtract 2 here to
690 * account for nir_lower_uniforms_to_ubo rebasing the UBOs such that UBO 0
691 * is the uniforms: */
692 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
693 unsigned ubo
= regid(const_state
->offsets
.ubo
, 0) - 2;
694 const unsigned ptrsz
= ir3_pointer_size(ctx
->compiler
);
698 /* First src is ubo index, which could either be an immed or not: */
699 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0];
700 if (is_same_type_mov(src0
) &&
701 (src0
->regs
[1]->flags
& IR3_REG_IMMED
)) {
702 base_lo
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
));
703 base_hi
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
) + 1);
705 base_lo
= create_uniform_indirect(b
, ubo
, ir3_get_addr(ctx
, src0
, ptrsz
));
706 base_hi
= create_uniform_indirect(b
, ubo
+ 1, ir3_get_addr(ctx
, src0
, ptrsz
));
708 /* NOTE: since relative addressing is used, make sure constlen is
709 * at least big enough to cover all the UBO addresses, since the
710 * assembler won't know what the max address reg is.
712 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
713 const_state
->offsets
.ubo
+ (ctx
->s
->info
.num_ubos
* ptrsz
));
716 /* note: on 32bit gpu's base_hi is ignored and DCE'd */
719 if (nir_src_is_const(intr
->src
[1])) {
720 off
+= nir_src_as_uint(intr
->src
[1]);
722 /* For load_ubo_indirect, second src is indirect offset: */
723 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0];
725 /* and add offset to addr: */
726 addr
= ir3_ADD_S(b
, addr
, 0, src1
, 0);
729 /* if offset is to large to encode in the ldg, split it out: */
730 if ((off
+ (intr
->num_components
* 4)) > 1024) {
731 /* split out the minimal amount to improve the odds that
732 * cp can fit the immediate in the add.s instruction:
734 unsigned off2
= off
+ (intr
->num_components
* 4) - 1024;
735 addr
= ir3_ADD_S(b
, addr
, 0, create_immed(b
, off2
), 0);
740 struct ir3_instruction
*carry
;
742 /* handle 32b rollover, ie:
743 * if (addr < base_lo)
746 carry
= ir3_CMPS_U(b
, addr
, 0, base_lo
, 0);
747 carry
->cat2
.condition
= IR3_COND_LT
;
748 base_hi
= ir3_ADD_S(b
, base_hi
, 0, carry
, 0);
750 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){ addr
, base_hi
}, 2);
753 for (int i
= 0; i
< intr
->num_components
; i
++) {
754 struct ir3_instruction
*load
=
755 ir3_LDG(b
, addr
, 0, create_immed(b
, 1), 0, /* num components */
756 create_immed(b
, off
+ i
* 4), 0);
757 load
->cat6
.type
= TYPE_U32
;
762 /* src[] = { block_index } */
764 emit_intrinsic_ssbo_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
765 struct ir3_instruction
**dst
)
767 /* SSBO size stored as a const starting at ssbo_sizes: */
768 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
769 unsigned blk_idx
= nir_src_as_uint(intr
->src
[0]);
770 unsigned idx
= regid(const_state
->offsets
.ssbo_sizes
, 0) +
771 const_state
->ssbo_size
.off
[blk_idx
];
773 debug_assert(const_state
->ssbo_size
.mask
& (1 << blk_idx
));
775 dst
[0] = create_uniform(ctx
->block
, idx
);
778 /* src[] = { offset }. const_index[] = { base } */
780 emit_intrinsic_load_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
781 struct ir3_instruction
**dst
)
783 struct ir3_block
*b
= ctx
->block
;
784 struct ir3_instruction
*ldl
, *offset
;
787 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
788 base
= nir_intrinsic_base(intr
);
790 ldl
= ir3_LDL(b
, offset
, 0,
791 create_immed(b
, intr
->num_components
), 0,
792 create_immed(b
, base
), 0);
794 ldl
->cat6
.type
= utype_dst(intr
->dest
);
795 ldl
->regs
[0]->wrmask
= MASK(intr
->num_components
);
797 ldl
->barrier_class
= IR3_BARRIER_SHARED_R
;
798 ldl
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
800 ir3_split_dest(b
, dst
, ldl
, 0, intr
->num_components
);
803 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
805 emit_intrinsic_store_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
807 struct ir3_block
*b
= ctx
->block
;
808 struct ir3_instruction
*stl
, *offset
;
809 struct ir3_instruction
* const *value
;
810 unsigned base
, wrmask
;
812 value
= ir3_get_src(ctx
, &intr
->src
[0]);
813 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
815 base
= nir_intrinsic_base(intr
);
816 wrmask
= nir_intrinsic_write_mask(intr
);
818 /* Combine groups of consecutive enabled channels in one write
819 * message. We use ffs to find the first enabled channel and then ffs on
820 * the bit-inverse, down-shifted writemask to determine the length of
821 * the block of enabled bits.
823 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
826 unsigned first_component
= ffs(wrmask
) - 1;
827 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
829 stl
= ir3_STL(b
, offset
, 0,
830 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
831 create_immed(b
, length
), 0);
832 stl
->cat6
.dst_offset
= first_component
+ base
;
833 stl
->cat6
.type
= utype_src(intr
->src
[0]);
834 stl
->barrier_class
= IR3_BARRIER_SHARED_W
;
835 stl
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
837 array_insert(b
, b
->keeps
, stl
);
839 /* Clear the bits in the writemask that we just wrote, then try
840 * again to see if more channels are left.
842 wrmask
&= (15 << (first_component
+ length
));
846 /* src[] = { offset }. const_index[] = { base } */
848 emit_intrinsic_load_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
849 struct ir3_instruction
**dst
)
851 struct ir3_block
*b
= ctx
->block
;
852 struct ir3_instruction
*load
, *offset
;
855 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
856 base
= nir_intrinsic_base(intr
);
858 load
= ir3_LDLW(b
, offset
, 0,
859 create_immed(b
, intr
->num_components
), 0,
860 create_immed(b
, base
), 0);
862 load
->cat6
.type
= utype_dst(intr
->dest
);
863 load
->regs
[0]->wrmask
= MASK(intr
->num_components
);
865 load
->barrier_class
= IR3_BARRIER_SHARED_R
;
866 load
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
868 ir3_split_dest(b
, dst
, load
, 0, intr
->num_components
);
871 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
873 emit_intrinsic_store_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
875 struct ir3_block
*b
= ctx
->block
;
876 struct ir3_instruction
*store
, *offset
;
877 struct ir3_instruction
* const *value
;
878 unsigned base
, wrmask
;
880 value
= ir3_get_src(ctx
, &intr
->src
[0]);
881 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
883 base
= nir_intrinsic_base(intr
);
884 wrmask
= nir_intrinsic_write_mask(intr
);
886 /* Combine groups of consecutive enabled channels in one write
887 * message. We use ffs to find the first enabled channel and then ffs on
888 * the bit-inverse, down-shifted writemask to determine the length of
889 * the block of enabled bits.
891 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
894 unsigned first_component
= ffs(wrmask
) - 1;
895 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
897 store
= ir3_STLW(b
, offset
, 0,
898 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
899 create_immed(b
, length
), 0);
901 store
->cat6
.dst_offset
= first_component
+ base
;
902 store
->cat6
.type
= utype_src(intr
->src
[0]);
903 store
->barrier_class
= IR3_BARRIER_SHARED_W
;
904 store
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
906 array_insert(b
, b
->keeps
, store
);
908 /* Clear the bits in the writemask that we just wrote, then try
909 * again to see if more channels are left.
911 wrmask
&= (15 << (first_component
+ length
));
916 * CS shared variable atomic intrinsics
918 * All of the shared variable atomic memory operations read a value from
919 * memory, compute a new value using one of the operations below, write the
920 * new value to memory, and return the original value read.
922 * All operations take 2 sources except CompSwap that takes 3. These
925 * 0: The offset into the shared variable storage region that the atomic
926 * operation will operate on.
927 * 1: The data parameter to the atomic function (i.e. the value to add
928 * in shared_atomic_add, etc).
929 * 2: For CompSwap only: the second data parameter.
931 static struct ir3_instruction
*
932 emit_intrinsic_atomic_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
934 struct ir3_block
*b
= ctx
->block
;
935 struct ir3_instruction
*atomic
, *src0
, *src1
;
936 type_t type
= TYPE_U32
;
938 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0]; /* offset */
939 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0]; /* value */
941 switch (intr
->intrinsic
) {
942 case nir_intrinsic_shared_atomic_add
:
943 atomic
= ir3_ATOMIC_ADD(b
, src0
, 0, src1
, 0);
945 case nir_intrinsic_shared_atomic_imin
:
946 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
949 case nir_intrinsic_shared_atomic_umin
:
950 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
952 case nir_intrinsic_shared_atomic_imax
:
953 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
956 case nir_intrinsic_shared_atomic_umax
:
957 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
959 case nir_intrinsic_shared_atomic_and
:
960 atomic
= ir3_ATOMIC_AND(b
, src0
, 0, src1
, 0);
962 case nir_intrinsic_shared_atomic_or
:
963 atomic
= ir3_ATOMIC_OR(b
, src0
, 0, src1
, 0);
965 case nir_intrinsic_shared_atomic_xor
:
966 atomic
= ir3_ATOMIC_XOR(b
, src0
, 0, src1
, 0);
968 case nir_intrinsic_shared_atomic_exchange
:
969 atomic
= ir3_ATOMIC_XCHG(b
, src0
, 0, src1
, 0);
971 case nir_intrinsic_shared_atomic_comp_swap
:
972 /* for cmpxchg, src1 is [ui]vec2(data, compare): */
973 src1
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
974 ir3_get_src(ctx
, &intr
->src
[2])[0],
977 atomic
= ir3_ATOMIC_CMPXCHG(b
, src0
, 0, src1
, 0);
983 atomic
->cat6
.iim_val
= 1;
985 atomic
->cat6
.type
= type
;
986 atomic
->barrier_class
= IR3_BARRIER_SHARED_W
;
987 atomic
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
989 /* even if nothing consume the result, we can't DCE the instruction: */
990 array_insert(b
, b
->keeps
, atomic
);
995 /* TODO handle actual indirect/dynamic case.. which is going to be weird
996 * to handle with the image_mapping table..
998 static struct ir3_instruction
*
999 get_image_samp_tex_src(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1001 unsigned slot
= ir3_get_image_slot(nir_src_as_deref(intr
->src
[0]));
1002 unsigned tex_idx
= ir3_image_to_tex(&ctx
->so
->image_mapping
, slot
);
1003 struct ir3_instruction
*texture
, *sampler
;
1005 texture
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1006 sampler
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1008 return ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1014 /* src[] = { deref, coord, sample_index }. const_index[] = {} */
1016 emit_intrinsic_load_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1017 struct ir3_instruction
**dst
)
1019 struct ir3_block
*b
= ctx
->block
;
1020 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
1021 struct ir3_instruction
*samp_tex
= get_image_samp_tex_src(ctx
, intr
);
1022 struct ir3_instruction
*sam
;
1023 struct ir3_instruction
* const *src0
= ir3_get_src(ctx
, &intr
->src
[1]);
1024 struct ir3_instruction
*coords
[4];
1025 unsigned flags
, ncoords
= ir3_get_image_coords(var
, &flags
);
1026 type_t type
= ir3_get_image_type(var
);
1028 /* hmm, this seems a bit odd, but it is what blob does and (at least
1029 * a5xx) just faults on bogus addresses otherwise:
1031 if (flags
& IR3_INSTR_3D
) {
1032 flags
&= ~IR3_INSTR_3D
;
1033 flags
|= IR3_INSTR_A
;
1036 for (unsigned i
= 0; i
< ncoords
; i
++)
1037 coords
[i
] = src0
[i
];
1040 coords
[ncoords
++] = create_immed(b
, 0);
1042 sam
= ir3_SAM(b
, OPC_ISAM
, type
, 0b1111, flags
,
1043 samp_tex
, ir3_create_collect(ctx
, coords
, ncoords
), NULL
);
1045 sam
->barrier_class
= IR3_BARRIER_IMAGE_R
;
1046 sam
->barrier_conflict
= IR3_BARRIER_IMAGE_W
;
1048 ir3_split_dest(b
, dst
, sam
, 0, 4);
1052 emit_intrinsic_image_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1053 struct ir3_instruction
**dst
)
1055 struct ir3_block
*b
= ctx
->block
;
1056 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
1057 struct ir3_instruction
*samp_tex
= get_image_samp_tex_src(ctx
, intr
);
1058 struct ir3_instruction
*sam
, *lod
;
1059 unsigned flags
, ncoords
= ir3_get_image_coords(var
, &flags
);
1061 lod
= create_immed(b
, 0);
1062 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
1063 samp_tex
, lod
, NULL
);
1065 /* Array size actually ends up in .w rather than .z. This doesn't
1066 * matter for miplevel 0, but for higher mips the value in z is
1067 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
1068 * returned, which means that we have to add 1 to it for arrays for
1071 * Note use a temporary dst and then copy, since the size of the dst
1072 * array that is passed in is based on nir's understanding of the
1073 * result size, not the hardware's
1075 struct ir3_instruction
*tmp
[4];
1077 ir3_split_dest(b
, tmp
, sam
, 0, 4);
1079 /* get_size instruction returns size in bytes instead of texels
1080 * for imageBuffer, so we need to divide it by the pixel size
1081 * of the image format.
1083 * TODO: This is at least true on a5xx. Check other gens.
1085 enum glsl_sampler_dim dim
=
1086 glsl_get_sampler_dim(glsl_without_array(var
->type
));
1087 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
1088 /* Since all the possible values the divisor can take are
1089 * power-of-two (4, 8, or 16), the division is implemented
1091 * During shader setup, the log2 of the image format's
1092 * bytes-per-pixel should have been emitted in 2nd slot of
1093 * image_dims. See ir3_shader::emit_image_dims().
1095 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
1096 unsigned cb
= regid(const_state
->offsets
.image_dims
, 0) +
1097 const_state
->image_dims
.off
[var
->data
.driver_location
];
1098 struct ir3_instruction
*aux
= create_uniform(b
, cb
+ 1);
1100 tmp
[0] = ir3_SHR_B(b
, tmp
[0], 0, aux
, 0);
1103 for (unsigned i
= 0; i
< ncoords
; i
++)
1106 if (flags
& IR3_INSTR_A
) {
1107 if (ctx
->compiler
->levels_add_one
) {
1108 dst
[ncoords
-1] = ir3_ADD_U(b
, tmp
[3], 0, create_immed(b
, 1), 0);
1110 dst
[ncoords
-1] = ir3_MOV(b
, tmp
[3], TYPE_U32
);
1116 emit_intrinsic_barrier(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1118 struct ir3_block
*b
= ctx
->block
;
1119 struct ir3_instruction
*barrier
;
1121 switch (intr
->intrinsic
) {
1122 case nir_intrinsic_barrier
:
1123 barrier
= ir3_BAR(b
);
1124 barrier
->cat7
.g
= true;
1125 barrier
->cat7
.l
= true;
1126 barrier
->flags
= IR3_INSTR_SS
| IR3_INSTR_SY
;
1127 barrier
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1129 case nir_intrinsic_memory_barrier
:
1130 barrier
= ir3_FENCE(b
);
1131 barrier
->cat7
.g
= true;
1132 barrier
->cat7
.r
= true;
1133 barrier
->cat7
.w
= true;
1134 barrier
->cat7
.l
= true;
1135 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
|
1136 IR3_BARRIER_BUFFER_W
;
1137 barrier
->barrier_conflict
=
1138 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1139 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1141 case nir_intrinsic_memory_barrier_atomic_counter
:
1142 case nir_intrinsic_memory_barrier_buffer
:
1143 barrier
= ir3_FENCE(b
);
1144 barrier
->cat7
.g
= true;
1145 barrier
->cat7
.r
= true;
1146 barrier
->cat7
.w
= true;
1147 barrier
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1148 barrier
->barrier_conflict
= IR3_BARRIER_BUFFER_R
|
1149 IR3_BARRIER_BUFFER_W
;
1151 case nir_intrinsic_memory_barrier_image
:
1152 // TODO double check if this should have .g set
1153 barrier
= ir3_FENCE(b
);
1154 barrier
->cat7
.g
= true;
1155 barrier
->cat7
.r
= true;
1156 barrier
->cat7
.w
= true;
1157 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1158 barrier
->barrier_conflict
= IR3_BARRIER_IMAGE_R
|
1159 IR3_BARRIER_IMAGE_W
;
1161 case nir_intrinsic_memory_barrier_shared
:
1162 barrier
= ir3_FENCE(b
);
1163 barrier
->cat7
.g
= true;
1164 barrier
->cat7
.l
= true;
1165 barrier
->cat7
.r
= true;
1166 barrier
->cat7
.w
= true;
1167 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
;
1168 barrier
->barrier_conflict
= IR3_BARRIER_SHARED_R
|
1169 IR3_BARRIER_SHARED_W
;
1171 case nir_intrinsic_group_memory_barrier
:
1172 barrier
= ir3_FENCE(b
);
1173 barrier
->cat7
.g
= true;
1174 barrier
->cat7
.l
= true;
1175 barrier
->cat7
.r
= true;
1176 barrier
->cat7
.w
= true;
1177 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
|
1178 IR3_BARRIER_IMAGE_W
|
1179 IR3_BARRIER_BUFFER_W
;
1180 barrier
->barrier_conflict
=
1181 IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
|
1182 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1183 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1189 /* make sure barrier doesn't get DCE'd */
1190 array_insert(b
, b
->keeps
, barrier
);
1193 static void add_sysval_input_compmask(struct ir3_context
*ctx
,
1194 gl_system_value slot
, unsigned compmask
,
1195 struct ir3_instruction
*instr
)
1197 struct ir3_shader_variant
*so
= ctx
->so
;
1198 unsigned r
= regid(so
->inputs_count
, 0);
1199 unsigned n
= so
->inputs_count
++;
1201 assert(instr
->opc
== OPC_META_INPUT
);
1202 instr
->input
.sysval
= slot
;
1204 so
->inputs
[n
].sysval
= true;
1205 so
->inputs
[n
].slot
= slot
;
1206 so
->inputs
[n
].compmask
= compmask
;
1207 so
->inputs
[n
].regid
= r
;
1208 so
->inputs
[n
].interpolate
= INTERP_MODE_FLAT
;
1211 ctx
->ir
->ninputs
= MAX2(ctx
->ir
->ninputs
, r
+ 1);
1212 ctx
->ir
->inputs
[r
] = instr
;
1215 static void add_sysval_input(struct ir3_context
*ctx
, gl_system_value slot
,
1216 struct ir3_instruction
*instr
)
1218 add_sysval_input_compmask(ctx
, slot
, 0x1, instr
);
1221 static struct ir3_instruction
*
1222 get_barycentric_centroid(struct ir3_context
*ctx
)
1224 if (!ctx
->ij_centroid
) {
1225 struct ir3_instruction
*xy
[2];
1226 struct ir3_instruction
*ij
;
1228 ij
= create_input_compmask(ctx
, 0, 0x3);
1229 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1231 ctx
->ij_centroid
= ir3_create_collect(ctx
, xy
, 2);
1233 add_sysval_input_compmask(ctx
,
1234 SYSTEM_VALUE_BARYCENTRIC_CENTROID
,
1238 return ctx
->ij_centroid
;
1241 static struct ir3_instruction
*
1242 get_barycentric_sample(struct ir3_context
*ctx
)
1244 if (!ctx
->ij_sample
) {
1245 struct ir3_instruction
*xy
[2];
1246 struct ir3_instruction
*ij
;
1248 ij
= create_input_compmask(ctx
, 0, 0x3);
1249 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1251 ctx
->ij_sample
= ir3_create_collect(ctx
, xy
, 2);
1253 add_sysval_input_compmask(ctx
,
1254 SYSTEM_VALUE_BARYCENTRIC_SAMPLE
,
1258 return ctx
->ij_sample
;
1261 static struct ir3_instruction
*
1262 get_barycentric_pixel(struct ir3_context
*ctx
)
1264 /* TODO when tgsi_to_nir supports "new-style" FS inputs switch
1265 * this to create ij_pixel only on demand:
1267 return ctx
->ij_pixel
;
1270 static struct ir3_instruction
*
1271 get_frag_coord(struct ir3_context
*ctx
)
1273 if (!ctx
->frag_coord
) {
1274 struct ir3_block
*b
= ctx
->block
;
1275 struct ir3_instruction
*xyzw
[4];
1276 struct ir3_instruction
*hw_frag_coord
;
1278 hw_frag_coord
= create_input_compmask(ctx
, 0, 0xf);
1279 ir3_split_dest(ctx
->block
, xyzw
, hw_frag_coord
, 0, 4);
1281 /* for frag_coord.xy, we get unsigned values.. we need
1282 * to subtract (integer) 8 and divide by 16 (right-
1283 * shift by 4) then convert to float:
1287 * mov.u32f32 dst, tmp
1290 for (int i
= 0; i
< 2; i
++) {
1291 xyzw
[i
] = ir3_SUB_S(b
, xyzw
[i
], 0,
1292 create_immed(b
, 8), 0);
1293 xyzw
[i
] = ir3_SHR_B(b
, xyzw
[i
], 0,
1294 create_immed(b
, 4), 0);
1295 xyzw
[i
] = ir3_COV(b
, xyzw
[i
], TYPE_U32
, TYPE_F32
);
1298 ctx
->frag_coord
= ir3_create_collect(ctx
, xyzw
, 4);
1300 add_sysval_input_compmask(ctx
,
1301 SYSTEM_VALUE_FRAG_COORD
,
1302 0xf, hw_frag_coord
);
1304 ctx
->so
->frag_coord
= true;
1307 return ctx
->frag_coord
;
1311 emit_intrinsic(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1313 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1314 struct ir3_instruction
**dst
;
1315 struct ir3_instruction
* const *src
;
1316 struct ir3_block
*b
= ctx
->block
;
1319 if (info
->has_dest
) {
1320 unsigned n
= nir_intrinsic_dest_components(intr
);
1321 dst
= ir3_get_dst(ctx
, &intr
->dest
, n
);
1326 const unsigned primitive_param
= ctx
->so
->shader
->const_state
.offsets
.primitive_param
* 4;
1327 const unsigned primitive_map
= ctx
->so
->shader
->const_state
.offsets
.primitive_map
* 4;
1329 switch (intr
->intrinsic
) {
1330 case nir_intrinsic_load_uniform
:
1331 idx
= nir_intrinsic_base(intr
);
1332 if (nir_src_is_const(intr
->src
[0])) {
1333 idx
+= nir_src_as_uint(intr
->src
[0]);
1334 for (int i
= 0; i
< intr
->num_components
; i
++) {
1335 dst
[i
] = create_uniform_typed(b
, idx
+ i
,
1336 nir_dest_bit_size(intr
->dest
) < 32 ? TYPE_F16
: TYPE_F32
);
1339 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1340 for (int i
= 0; i
< intr
->num_components
; i
++) {
1341 dst
[i
] = create_uniform_indirect(b
, idx
+ i
,
1342 ir3_get_addr(ctx
, src
[0], 1));
1344 /* NOTE: if relative addressing is used, we set
1345 * constlen in the compiler (to worst-case value)
1346 * since we don't know in the assembler what the max
1347 * addr reg value can be:
1349 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
1350 ctx
->so
->shader
->ubo_state
.size
/ 16);
1354 case nir_intrinsic_load_vs_primitive_stride_ir3
:
1355 dst
[0] = create_uniform(b
, primitive_param
+ 0);
1357 case nir_intrinsic_load_vs_vertex_stride_ir3
:
1358 dst
[0] = create_uniform(b
, primitive_param
+ 1);
1360 case nir_intrinsic_load_primitive_location_ir3
:
1361 idx
= nir_intrinsic_driver_location(intr
);
1362 dst
[0] = create_uniform(b
, primitive_map
+ idx
);
1365 case nir_intrinsic_load_gs_header_ir3
:
1366 dst
[0] = ctx
->gs_header
;
1369 case nir_intrinsic_load_primitive_id
:
1370 dst
[0] = ctx
->primitive_id
;
1373 case nir_intrinsic_load_ubo
:
1374 emit_intrinsic_load_ubo(ctx
, intr
, dst
);
1376 case nir_intrinsic_load_frag_coord
:
1377 ir3_split_dest(b
, dst
, get_frag_coord(ctx
), 0, 4);
1379 case nir_intrinsic_load_sample_pos_from_id
: {
1380 /* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
1381 * but that doesn't seem necessary.
1383 struct ir3_instruction
*offset
=
1384 ir3_RGETPOS(b
, ir3_get_src(ctx
, &intr
->src
[0])[0], 0);
1385 offset
->regs
[0]->wrmask
= 0x3;
1386 offset
->cat5
.type
= TYPE_F32
;
1388 ir3_split_dest(b
, dst
, offset
, 0, 2);
1392 case nir_intrinsic_load_size_ir3
:
1393 if (!ctx
->ij_size
) {
1394 ctx
->ij_size
= create_input(ctx
, 0);
1396 add_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_SIZE
,
1399 dst
[0] = ctx
->ij_size
;
1401 case nir_intrinsic_load_barycentric_centroid
:
1402 ir3_split_dest(b
, dst
, get_barycentric_centroid(ctx
), 0, 2);
1404 case nir_intrinsic_load_barycentric_sample
:
1405 if (ctx
->so
->key
.msaa
) {
1406 ir3_split_dest(b
, dst
, get_barycentric_sample(ctx
), 0, 2);
1408 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1411 case nir_intrinsic_load_barycentric_pixel
:
1412 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1414 case nir_intrinsic_load_interpolated_input
:
1415 idx
= nir_intrinsic_base(intr
);
1416 comp
= nir_intrinsic_component(intr
);
1417 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1418 if (nir_src_is_const(intr
->src
[1])) {
1419 struct ir3_instruction
*coord
= ir3_create_collect(ctx
, src
, 2);
1420 idx
+= nir_src_as_uint(intr
->src
[1]);
1421 for (int i
= 0; i
< intr
->num_components
; i
++) {
1422 unsigned inloc
= idx
* 4 + i
+ comp
;
1423 if (ctx
->so
->inputs
[idx
].bary
&&
1424 !ctx
->so
->inputs
[idx
].use_ldlv
) {
1425 dst
[i
] = ir3_BARY_F(b
, create_immed(b
, inloc
), 0, coord
, 0);
1427 /* for non-varyings use the pre-setup input, since
1428 * that is easier than mapping things back to a
1429 * nir_variable to figure out what it is.
1431 dst
[i
] = ctx
->ir
->inputs
[inloc
];
1435 ir3_context_error(ctx
, "unhandled");
1438 case nir_intrinsic_load_input
:
1439 idx
= nir_intrinsic_base(intr
);
1440 comp
= nir_intrinsic_component(intr
);
1441 if (nir_src_is_const(intr
->src
[0])) {
1442 idx
+= nir_src_as_uint(intr
->src
[0]);
1443 for (int i
= 0; i
< intr
->num_components
; i
++) {
1444 unsigned n
= idx
* 4 + i
+ comp
;
1445 dst
[i
] = ctx
->ir
->inputs
[n
];
1446 compile_assert(ctx
, ctx
->ir
->inputs
[n
]);
1449 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1450 struct ir3_instruction
*collect
=
1451 ir3_create_collect(ctx
, ctx
->ir
->inputs
, ctx
->ir
->ninputs
);
1452 struct ir3_instruction
*addr
= ir3_get_addr(ctx
, src
[0], 4);
1453 for (int i
= 0; i
< intr
->num_components
; i
++) {
1454 unsigned n
= idx
* 4 + i
+ comp
;
1455 dst
[i
] = create_indirect_load(ctx
, ctx
->ir
->ninputs
,
1460 /* All SSBO intrinsics should have been lowered by 'lower_io_offsets'
1461 * pass and replaced by an ir3-specifc version that adds the
1462 * dword-offset in the last source.
1464 case nir_intrinsic_load_ssbo_ir3
:
1465 ctx
->funcs
->emit_intrinsic_load_ssbo(ctx
, intr
, dst
);
1467 case nir_intrinsic_store_ssbo_ir3
:
1468 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1469 !ctx
->s
->info
.fs
.early_fragment_tests
)
1470 ctx
->so
->no_earlyz
= true;
1471 ctx
->funcs
->emit_intrinsic_store_ssbo(ctx
, intr
);
1473 case nir_intrinsic_get_buffer_size
:
1474 emit_intrinsic_ssbo_size(ctx
, intr
, dst
);
1476 case nir_intrinsic_ssbo_atomic_add_ir3
:
1477 case nir_intrinsic_ssbo_atomic_imin_ir3
:
1478 case nir_intrinsic_ssbo_atomic_umin_ir3
:
1479 case nir_intrinsic_ssbo_atomic_imax_ir3
:
1480 case nir_intrinsic_ssbo_atomic_umax_ir3
:
1481 case nir_intrinsic_ssbo_atomic_and_ir3
:
1482 case nir_intrinsic_ssbo_atomic_or_ir3
:
1483 case nir_intrinsic_ssbo_atomic_xor_ir3
:
1484 case nir_intrinsic_ssbo_atomic_exchange_ir3
:
1485 case nir_intrinsic_ssbo_atomic_comp_swap_ir3
:
1486 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1487 !ctx
->s
->info
.fs
.early_fragment_tests
)
1488 ctx
->so
->no_earlyz
= true;
1489 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_ssbo(ctx
, intr
);
1491 case nir_intrinsic_load_shared
:
1492 emit_intrinsic_load_shared(ctx
, intr
, dst
);
1494 case nir_intrinsic_store_shared
:
1495 emit_intrinsic_store_shared(ctx
, intr
);
1497 case nir_intrinsic_shared_atomic_add
:
1498 case nir_intrinsic_shared_atomic_imin
:
1499 case nir_intrinsic_shared_atomic_umin
:
1500 case nir_intrinsic_shared_atomic_imax
:
1501 case nir_intrinsic_shared_atomic_umax
:
1502 case nir_intrinsic_shared_atomic_and
:
1503 case nir_intrinsic_shared_atomic_or
:
1504 case nir_intrinsic_shared_atomic_xor
:
1505 case nir_intrinsic_shared_atomic_exchange
:
1506 case nir_intrinsic_shared_atomic_comp_swap
:
1507 dst
[0] = emit_intrinsic_atomic_shared(ctx
, intr
);
1509 case nir_intrinsic_image_deref_load
:
1510 emit_intrinsic_load_image(ctx
, intr
, dst
);
1512 case nir_intrinsic_image_deref_store
:
1513 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1514 !ctx
->s
->info
.fs
.early_fragment_tests
)
1515 ctx
->so
->no_earlyz
= true;
1516 ctx
->funcs
->emit_intrinsic_store_image(ctx
, intr
);
1518 case nir_intrinsic_image_deref_size
:
1519 emit_intrinsic_image_size(ctx
, intr
, dst
);
1521 case nir_intrinsic_image_deref_atomic_add
:
1522 case nir_intrinsic_image_deref_atomic_imin
:
1523 case nir_intrinsic_image_deref_atomic_umin
:
1524 case nir_intrinsic_image_deref_atomic_imax
:
1525 case nir_intrinsic_image_deref_atomic_umax
:
1526 case nir_intrinsic_image_deref_atomic_and
:
1527 case nir_intrinsic_image_deref_atomic_or
:
1528 case nir_intrinsic_image_deref_atomic_xor
:
1529 case nir_intrinsic_image_deref_atomic_exchange
:
1530 case nir_intrinsic_image_deref_atomic_comp_swap
:
1531 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1532 !ctx
->s
->info
.fs
.early_fragment_tests
)
1533 ctx
->so
->no_earlyz
= true;
1534 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_image(ctx
, intr
);
1536 case nir_intrinsic_barrier
:
1537 case nir_intrinsic_memory_barrier
:
1538 case nir_intrinsic_group_memory_barrier
:
1539 case nir_intrinsic_memory_barrier_atomic_counter
:
1540 case nir_intrinsic_memory_barrier_buffer
:
1541 case nir_intrinsic_memory_barrier_image
:
1542 case nir_intrinsic_memory_barrier_shared
:
1543 emit_intrinsic_barrier(ctx
, intr
);
1544 /* note that blk ptr no longer valid, make that obvious: */
1547 case nir_intrinsic_store_output
:
1548 idx
= nir_intrinsic_base(intr
);
1549 comp
= nir_intrinsic_component(intr
);
1550 compile_assert(ctx
, nir_src_is_const(intr
->src
[1]));
1551 idx
+= nir_src_as_uint(intr
->src
[1]);
1553 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1554 for (int i
= 0; i
< intr
->num_components
; i
++) {
1555 unsigned n
= idx
* 4 + i
+ comp
;
1556 ctx
->ir
->outputs
[n
] = src
[i
];
1559 case nir_intrinsic_load_base_vertex
:
1560 case nir_intrinsic_load_first_vertex
:
1561 if (!ctx
->basevertex
) {
1562 ctx
->basevertex
= create_driver_param(ctx
, IR3_DP_VTXID_BASE
);
1563 add_sysval_input(ctx
, SYSTEM_VALUE_FIRST_VERTEX
, ctx
->basevertex
);
1565 dst
[0] = ctx
->basevertex
;
1567 case nir_intrinsic_load_vertex_id_zero_base
:
1568 case nir_intrinsic_load_vertex_id
:
1569 if (!ctx
->vertex_id
) {
1570 gl_system_value sv
= (intr
->intrinsic
== nir_intrinsic_load_vertex_id
) ?
1571 SYSTEM_VALUE_VERTEX_ID
: SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1572 ctx
->vertex_id
= create_input(ctx
, 0);
1573 add_sysval_input(ctx
, sv
, ctx
->vertex_id
);
1575 dst
[0] = ctx
->vertex_id
;
1577 case nir_intrinsic_load_instance_id
:
1578 if (!ctx
->instance_id
) {
1579 ctx
->instance_id
= create_input(ctx
, 0);
1580 add_sysval_input(ctx
, SYSTEM_VALUE_INSTANCE_ID
,
1583 dst
[0] = ctx
->instance_id
;
1585 case nir_intrinsic_load_sample_id
:
1586 ctx
->so
->per_samp
= true;
1588 case nir_intrinsic_load_sample_id_no_per_sample
:
1589 if (!ctx
->samp_id
) {
1590 ctx
->samp_id
= create_input(ctx
, 0);
1591 ctx
->samp_id
->regs
[0]->flags
|= IR3_REG_HALF
;
1592 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_ID
,
1595 dst
[0] = ir3_COV(b
, ctx
->samp_id
, TYPE_U16
, TYPE_U32
);
1597 case nir_intrinsic_load_sample_mask_in
:
1598 if (!ctx
->samp_mask_in
) {
1599 ctx
->samp_mask_in
= create_input(ctx
, 0);
1600 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_MASK_IN
,
1603 dst
[0] = ctx
->samp_mask_in
;
1605 case nir_intrinsic_load_user_clip_plane
:
1606 idx
= nir_intrinsic_ucp_id(intr
);
1607 for (int i
= 0; i
< intr
->num_components
; i
++) {
1608 unsigned n
= idx
* 4 + i
;
1609 dst
[i
] = create_driver_param(ctx
, IR3_DP_UCP0_X
+ n
);
1612 case nir_intrinsic_load_front_face
:
1613 if (!ctx
->frag_face
) {
1614 ctx
->so
->frag_face
= true;
1615 ctx
->frag_face
= create_input(ctx
, 0);
1616 add_sysval_input(ctx
, SYSTEM_VALUE_FRONT_FACE
, ctx
->frag_face
);
1617 ctx
->frag_face
->regs
[0]->flags
|= IR3_REG_HALF
;
1619 /* for fragface, we get -1 for back and 0 for front. However this is
1620 * the inverse of what nir expects (where ~0 is true).
1622 dst
[0] = ir3_COV(b
, ctx
->frag_face
, TYPE_S16
, TYPE_S32
);
1623 dst
[0] = ir3_NOT_B(b
, dst
[0], 0);
1625 case nir_intrinsic_load_local_invocation_id
:
1626 if (!ctx
->local_invocation_id
) {
1627 ctx
->local_invocation_id
= create_input_compmask(ctx
, 0, 0x7);
1628 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_LOCAL_INVOCATION_ID
,
1629 0x7, ctx
->local_invocation_id
);
1631 ir3_split_dest(b
, dst
, ctx
->local_invocation_id
, 0, 3);
1633 case nir_intrinsic_load_work_group_id
:
1634 if (!ctx
->work_group_id
) {
1635 ctx
->work_group_id
= create_input_compmask(ctx
, 0, 0x7);
1636 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_WORK_GROUP_ID
,
1637 0x7, ctx
->work_group_id
);
1638 ctx
->work_group_id
->regs
[0]->flags
|= IR3_REG_HIGH
;
1640 ir3_split_dest(b
, dst
, ctx
->work_group_id
, 0, 3);
1642 case nir_intrinsic_load_num_work_groups
:
1643 for (int i
= 0; i
< intr
->num_components
; i
++) {
1644 dst
[i
] = create_driver_param(ctx
, IR3_DP_NUM_WORK_GROUPS_X
+ i
);
1647 case nir_intrinsic_load_local_group_size
:
1648 for (int i
= 0; i
< intr
->num_components
; i
++) {
1649 dst
[i
] = create_driver_param(ctx
, IR3_DP_LOCAL_GROUP_SIZE_X
+ i
);
1652 case nir_intrinsic_discard_if
:
1653 case nir_intrinsic_discard
: {
1654 struct ir3_instruction
*cond
, *kill
;
1656 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
1657 /* conditional discard: */
1658 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1659 cond
= ir3_b2n(b
, src
[0]);
1661 /* unconditional discard: */
1662 cond
= create_immed(b
, 1);
1665 /* NOTE: only cmps.*.* can write p0.x: */
1666 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1667 cond
->cat2
.condition
= IR3_COND_NE
;
1669 /* condition always goes in predicate register: */
1670 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1672 kill
= ir3_KILL(b
, cond
, 0);
1673 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1675 array_insert(b
, b
->keeps
, kill
);
1676 ctx
->so
->no_earlyz
= true;
1680 case nir_intrinsic_load_shared_ir3
:
1681 emit_intrinsic_load_shared_ir3(ctx
, intr
, dst
);
1683 case nir_intrinsic_store_shared_ir3
:
1684 emit_intrinsic_store_shared_ir3(ctx
, intr
);
1687 ir3_context_error(ctx
, "Unhandled intrinsic type: %s\n",
1688 nir_intrinsic_infos
[intr
->intrinsic
].name
);
1693 ir3_put_dst(ctx
, &intr
->dest
);
1697 emit_load_const(struct ir3_context
*ctx
, nir_load_const_instr
*instr
)
1699 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &instr
->def
,
1700 instr
->def
.num_components
);
1702 if (instr
->def
.bit_size
< 32) {
1703 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1704 dst
[i
] = create_immed_typed(ctx
->block
,
1705 instr
->value
[i
].u16
,
1708 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1709 dst
[i
] = create_immed_typed(ctx
->block
,
1710 instr
->value
[i
].u32
,
1717 emit_undef(struct ir3_context
*ctx
, nir_ssa_undef_instr
*undef
)
1719 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &undef
->def
,
1720 undef
->def
.num_components
);
1721 type_t type
= (undef
->def
.bit_size
< 32) ? TYPE_U16
: TYPE_U32
;
1723 /* backend doesn't want undefined instructions, so just plug
1726 for (int i
= 0; i
< undef
->def
.num_components
; i
++)
1727 dst
[i
] = create_immed_typed(ctx
->block
, fui(0.0), type
);
1731 * texture fetch/sample instructions:
1735 tex_info(nir_tex_instr
*tex
, unsigned *flagsp
, unsigned *coordsp
)
1737 unsigned coords
, flags
= 0;
1739 /* note: would use tex->coord_components.. except txs.. also,
1740 * since array index goes after shadow ref, we don't want to
1743 switch (tex
->sampler_dim
) {
1744 case GLSL_SAMPLER_DIM_1D
:
1745 case GLSL_SAMPLER_DIM_BUF
:
1748 case GLSL_SAMPLER_DIM_2D
:
1749 case GLSL_SAMPLER_DIM_RECT
:
1750 case GLSL_SAMPLER_DIM_EXTERNAL
:
1751 case GLSL_SAMPLER_DIM_MS
:
1754 case GLSL_SAMPLER_DIM_3D
:
1755 case GLSL_SAMPLER_DIM_CUBE
:
1757 flags
|= IR3_INSTR_3D
;
1760 unreachable("bad sampler_dim");
1763 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
1764 flags
|= IR3_INSTR_S
;
1766 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
)
1767 flags
|= IR3_INSTR_A
;
1773 /* Gets the sampler/texture idx as a hvec2. Which could either be dynamic
1774 * or immediate (in which case it will get lowered later to a non .s2en
1775 * version of the tex instruction which encode tex/samp as immediates:
1777 static struct ir3_instruction
*
1778 get_tex_samp_tex_src(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1780 int texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
);
1781 int sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
);
1782 struct ir3_instruction
*texture
, *sampler
;
1784 if (texture_idx
>= 0) {
1785 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
1786 texture
= ir3_COV(ctx
->block
, texture
, TYPE_U32
, TYPE_U16
);
1788 /* TODO what to do for dynamic case? I guess we only need the
1789 * max index for astc srgb workaround so maybe not a problem
1790 * to worry about if we don't enable indirect samplers for
1793 ctx
->max_texture_index
= MAX2(ctx
->max_texture_index
, tex
->texture_index
);
1794 texture
= create_immed_typed(ctx
->block
, tex
->texture_index
, TYPE_U16
);
1797 if (sampler_idx
>= 0) {
1798 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
1799 sampler
= ir3_COV(ctx
->block
, sampler
, TYPE_U32
, TYPE_U16
);
1801 sampler
= create_immed_typed(ctx
->block
, tex
->sampler_index
, TYPE_U16
);
1804 return ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1811 emit_tex(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1813 struct ir3_block
*b
= ctx
->block
;
1814 struct ir3_instruction
**dst
, *sam
, *src0
[12], *src1
[4];
1815 struct ir3_instruction
* const *coord
, * const *off
, * const *ddx
, * const *ddy
;
1816 struct ir3_instruction
*lod
, *compare
, *proj
, *sample_index
;
1817 bool has_bias
= false, has_lod
= false, has_proj
= false, has_off
= false;
1818 unsigned i
, coords
, flags
, ncomp
;
1819 unsigned nsrc0
= 0, nsrc1
= 0;
1823 ncomp
= nir_dest_num_components(tex
->dest
);
1825 coord
= off
= ddx
= ddy
= NULL
;
1826 lod
= proj
= compare
= sample_index
= NULL
;
1828 dst
= ir3_get_dst(ctx
, &tex
->dest
, ncomp
);
1830 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1831 switch (tex
->src
[i
].src_type
) {
1832 case nir_tex_src_coord
:
1833 coord
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1835 case nir_tex_src_bias
:
1836 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1839 case nir_tex_src_lod
:
1840 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1843 case nir_tex_src_comparator
: /* shadow comparator */
1844 compare
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1846 case nir_tex_src_projector
:
1847 proj
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1850 case nir_tex_src_offset
:
1851 off
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1854 case nir_tex_src_ddx
:
1855 ddx
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1857 case nir_tex_src_ddy
:
1858 ddy
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1860 case nir_tex_src_ms_index
:
1861 sample_index
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1863 case nir_tex_src_texture_offset
:
1864 case nir_tex_src_sampler_offset
:
1865 /* handled in get_tex_samp_src() */
1868 ir3_context_error(ctx
, "Unhandled NIR tex src type: %d\n",
1869 tex
->src
[i
].src_type
);
1875 case nir_texop_tex
: opc
= has_lod
? OPC_SAML
: OPC_SAM
; break;
1876 case nir_texop_txb
: opc
= OPC_SAMB
; break;
1877 case nir_texop_txl
: opc
= OPC_SAML
; break;
1878 case nir_texop_txd
: opc
= OPC_SAMGQ
; break;
1879 case nir_texop_txf
: opc
= OPC_ISAML
; break;
1880 case nir_texop_lod
: opc
= OPC_GETLOD
; break;
1882 /* NOTE: a4xx might need to emulate gather w/ txf (this is
1883 * what blob does, seems gather is broken?), and a3xx did
1884 * not support it (but probably could also emulate).
1886 switch (tex
->component
) {
1887 case 0: opc
= OPC_GATHER4R
; break;
1888 case 1: opc
= OPC_GATHER4G
; break;
1889 case 2: opc
= OPC_GATHER4B
; break;
1890 case 3: opc
= OPC_GATHER4A
; break;
1893 case nir_texop_txf_ms_fb
:
1894 case nir_texop_txf_ms
: opc
= OPC_ISAMM
; break;
1896 ir3_context_error(ctx
, "Unhandled NIR tex type: %d\n", tex
->op
);
1900 tex_info(tex
, &flags
, &coords
);
1903 * lay out the first argument in the proper order:
1904 * - actual coordinates first
1905 * - shadow reference
1908 * - starting at offset 4, dpdx.xy, dpdy.xy
1910 * bias/lod go into the second arg
1913 /* insert tex coords: */
1914 for (i
= 0; i
< coords
; i
++)
1919 /* scale up integer coords for TXF based on the LOD */
1920 if (ctx
->compiler
->unminify_coords
&& (opc
== OPC_ISAML
)) {
1922 for (i
= 0; i
< coords
; i
++)
1923 src0
[i
] = ir3_SHL_B(b
, src0
[i
], 0, lod
, 0);
1927 /* hw doesn't do 1d, so we treat it as 2d with
1928 * height of 1, and patch up the y coord.
1931 src0
[nsrc0
++] = create_immed(b
, 0);
1933 src0
[nsrc0
++] = create_immed(b
, fui(0.5));
1937 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
1938 src0
[nsrc0
++] = compare
;
1940 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
) {
1941 struct ir3_instruction
*idx
= coord
[coords
];
1943 /* the array coord for cube arrays needs 0.5 added to it */
1944 if (ctx
->compiler
->array_index_add_half
&& !is_isam(opc
))
1945 idx
= ir3_ADD_F(b
, idx
, 0, create_immed(b
, fui(0.5)), 0);
1947 src0
[nsrc0
++] = idx
;
1951 src0
[nsrc0
++] = proj
;
1952 flags
|= IR3_INSTR_P
;
1955 /* pad to 4, then ddx/ddy: */
1956 if (tex
->op
== nir_texop_txd
) {
1958 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1959 for (i
= 0; i
< coords
; i
++)
1960 src0
[nsrc0
++] = ddx
[i
];
1962 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1963 for (i
= 0; i
< coords
; i
++)
1964 src0
[nsrc0
++] = ddy
[i
];
1966 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1969 /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
1970 * with scaled x coord according to requested sample:
1972 if (opc
== OPC_ISAMM
) {
1973 if (ctx
->compiler
->txf_ms_with_isaml
) {
1974 /* the samples are laid out in x dimension as
1976 * x_ms = (x << ms) + sample_index;
1978 struct ir3_instruction
*ms
;
1979 ms
= create_immed(b
, (ctx
->samples
>> (2 * tex
->texture_index
)) & 3);
1981 src0
[0] = ir3_SHL_B(b
, src0
[0], 0, ms
, 0);
1982 src0
[0] = ir3_ADD_U(b
, src0
[0], 0, sample_index
, 0);
1986 src0
[nsrc0
++] = sample_index
;
1991 * second argument (if applicable):
1996 if (has_off
| has_lod
| has_bias
) {
1998 unsigned off_coords
= coords
;
1999 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2001 for (i
= 0; i
< off_coords
; i
++)
2002 src1
[nsrc1
++] = off
[i
];
2004 src1
[nsrc1
++] = create_immed(b
, fui(0.0));
2005 flags
|= IR3_INSTR_O
;
2008 if (has_lod
| has_bias
)
2009 src1
[nsrc1
++] = lod
;
2012 switch (tex
->dest_type
) {
2013 case nir_type_invalid
:
2014 case nir_type_float
:
2025 unreachable("bad dest_type");
2028 if (opc
== OPC_GETLOD
)
2031 struct ir3_instruction
*samp_tex
;
2033 if (tex
->op
== nir_texop_txf_ms_fb
) {
2034 /* only expect a single txf_ms_fb per shader: */
2035 compile_assert(ctx
, !ctx
->so
->fb_read
);
2036 compile_assert(ctx
, ctx
->so
->type
== MESA_SHADER_FRAGMENT
);
2038 ctx
->so
->fb_read
= true;
2039 samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2040 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2041 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2044 ctx
->so
->num_samp
++;
2046 samp_tex
= get_tex_samp_tex_src(ctx
, tex
);
2049 struct ir3_instruction
*col0
= ir3_create_collect(ctx
, src0
, nsrc0
);
2050 struct ir3_instruction
*col1
= ir3_create_collect(ctx
, src1
, nsrc1
);
2052 sam
= ir3_SAM(b
, opc
, type
, MASK(ncomp
), flags
,
2053 samp_tex
, col0
, col1
);
2055 if ((ctx
->astc_srgb
& (1 << tex
->texture_index
)) && !nir_tex_instr_is_query(tex
)) {
2056 /* only need first 3 components: */
2057 sam
->regs
[0]->wrmask
= 0x7;
2058 ir3_split_dest(b
, dst
, sam
, 0, 3);
2060 /* we need to sample the alpha separately with a non-ASTC
2063 sam
= ir3_SAM(b
, opc
, type
, 0b1000, flags
,
2064 samp_tex
, col0
, col1
);
2066 array_insert(ctx
->ir
, ctx
->ir
->astc_srgb
, sam
);
2068 /* fixup .w component: */
2069 ir3_split_dest(b
, &dst
[3], sam
, 3, 1);
2071 /* normal (non-workaround) case: */
2072 ir3_split_dest(b
, dst
, sam
, 0, ncomp
);
2075 /* GETLOD returns results in 4.8 fixed point */
2076 if (opc
== OPC_GETLOD
) {
2077 struct ir3_instruction
*factor
= create_immed(b
, fui(1.0 / 256));
2079 compile_assert(ctx
, tex
->dest_type
== nir_type_float
);
2080 for (i
= 0; i
< 2; i
++) {
2081 dst
[i
] = ir3_MUL_F(b
, ir3_COV(b
, dst
[i
], TYPE_S32
, TYPE_F32
), 0,
2086 ir3_put_dst(ctx
, &tex
->dest
);
2090 emit_tex_info(struct ir3_context
*ctx
, nir_tex_instr
*tex
, unsigned idx
)
2092 struct ir3_block
*b
= ctx
->block
;
2093 struct ir3_instruction
**dst
, *sam
;
2095 dst
= ir3_get_dst(ctx
, &tex
->dest
, 1);
2097 sam
= ir3_SAM(b
, OPC_GETINFO
, TYPE_U32
, 1 << idx
, 0,
2098 get_tex_samp_tex_src(ctx
, tex
), NULL
, NULL
);
2100 /* even though there is only one component, since it ends
2101 * up in .y/.z/.w rather than .x, we need a split_dest()
2104 ir3_split_dest(b
, dst
, sam
, 0, idx
+ 1);
2106 /* The # of levels comes from getinfo.z. We need to add 1 to it, since
2107 * the value in TEX_CONST_0 is zero-based.
2109 if (ctx
->compiler
->levels_add_one
)
2110 dst
[0] = ir3_ADD_U(b
, dst
[0], 0, create_immed(b
, 1), 0);
2112 ir3_put_dst(ctx
, &tex
->dest
);
2116 emit_tex_txs(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2118 struct ir3_block
*b
= ctx
->block
;
2119 struct ir3_instruction
**dst
, *sam
;
2120 struct ir3_instruction
*lod
;
2121 unsigned flags
, coords
;
2123 tex_info(tex
, &flags
, &coords
);
2125 /* Actually we want the number of dimensions, not coordinates. This
2126 * distinction only matters for cubes.
2128 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2131 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
2133 compile_assert(ctx
, tex
->num_srcs
== 1);
2134 compile_assert(ctx
, tex
->src
[0].src_type
== nir_tex_src_lod
);
2136 lod
= ir3_get_src(ctx
, &tex
->src
[0].src
)[0];
2138 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
2139 get_tex_samp_tex_src(ctx
, tex
), lod
, NULL
);
2141 ir3_split_dest(b
, dst
, sam
, 0, 4);
2143 /* Array size actually ends up in .w rather than .z. This doesn't
2144 * matter for miplevel 0, but for higher mips the value in z is
2145 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2146 * returned, which means that we have to add 1 to it for arrays.
2148 if (tex
->is_array
) {
2149 if (ctx
->compiler
->levels_add_one
) {
2150 dst
[coords
] = ir3_ADD_U(b
, dst
[3], 0, create_immed(b
, 1), 0);
2152 dst
[coords
] = ir3_MOV(b
, dst
[3], TYPE_U32
);
2156 ir3_put_dst(ctx
, &tex
->dest
);
2160 emit_jump(struct ir3_context
*ctx
, nir_jump_instr
*jump
)
2162 switch (jump
->type
) {
2163 case nir_jump_break
:
2164 case nir_jump_continue
:
2165 case nir_jump_return
:
2166 /* I *think* we can simply just ignore this, and use the
2167 * successor block link to figure out where we need to
2168 * jump to for break/continue
2172 ir3_context_error(ctx
, "Unhandled NIR jump type: %d\n", jump
->type
);
2178 emit_instr(struct ir3_context
*ctx
, nir_instr
*instr
)
2180 switch (instr
->type
) {
2181 case nir_instr_type_alu
:
2182 emit_alu(ctx
, nir_instr_as_alu(instr
));
2184 case nir_instr_type_deref
:
2185 /* ignored, handled as part of the intrinsic they are src to */
2187 case nir_instr_type_intrinsic
:
2188 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2190 case nir_instr_type_load_const
:
2191 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2193 case nir_instr_type_ssa_undef
:
2194 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
2196 case nir_instr_type_tex
: {
2197 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
2198 /* couple tex instructions get special-cased:
2202 emit_tex_txs(ctx
, tex
);
2204 case nir_texop_query_levels
:
2205 emit_tex_info(ctx
, tex
, 2);
2207 case nir_texop_texture_samples
:
2208 emit_tex_info(ctx
, tex
, 3);
2216 case nir_instr_type_jump
:
2217 emit_jump(ctx
, nir_instr_as_jump(instr
));
2219 case nir_instr_type_phi
:
2220 /* we have converted phi webs to regs in NIR by now */
2221 ir3_context_error(ctx
, "Unexpected NIR instruction type: %d\n", instr
->type
);
2223 case nir_instr_type_call
:
2224 case nir_instr_type_parallel_copy
:
2225 ir3_context_error(ctx
, "Unhandled NIR instruction type: %d\n", instr
->type
);
2230 static struct ir3_block
*
2231 get_block(struct ir3_context
*ctx
, const nir_block
*nblock
)
2233 struct ir3_block
*block
;
2234 struct hash_entry
*hentry
;
2236 hentry
= _mesa_hash_table_search(ctx
->block_ht
, nblock
);
2238 return hentry
->data
;
2240 block
= ir3_block_create(ctx
->ir
);
2241 block
->nblock
= nblock
;
2242 _mesa_hash_table_insert(ctx
->block_ht
, nblock
, block
);
2244 block
->predecessors
= _mesa_pointer_set_create(block
);
2245 set_foreach(nblock
->predecessors
, sentry
) {
2246 _mesa_set_add(block
->predecessors
, get_block(ctx
, sentry
->key
));
2253 emit_block(struct ir3_context
*ctx
, nir_block
*nblock
)
2255 struct ir3_block
*block
= get_block(ctx
, nblock
);
2257 for (int i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
2258 if (nblock
->successors
[i
]) {
2259 block
->successors
[i
] =
2260 get_block(ctx
, nblock
->successors
[i
]);
2265 list_addtail(&block
->node
, &ctx
->ir
->block_list
);
2267 /* re-emit addr register in each block if needed: */
2268 for (int i
= 0; i
< ARRAY_SIZE(ctx
->addr_ht
); i
++) {
2269 _mesa_hash_table_destroy(ctx
->addr_ht
[i
], NULL
);
2270 ctx
->addr_ht
[i
] = NULL
;
2273 nir_foreach_instr(instr
, nblock
) {
2274 ctx
->cur_instr
= instr
;
2275 emit_instr(ctx
, instr
);
2276 ctx
->cur_instr
= NULL
;
2282 static void emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
);
2285 emit_if(struct ir3_context
*ctx
, nir_if
*nif
)
2287 struct ir3_instruction
*condition
= ir3_get_src(ctx
, &nif
->condition
)[0];
2289 ctx
->block
->condition
=
2290 ir3_get_predicate(ctx
, ir3_b2n(condition
->block
, condition
));
2292 emit_cf_list(ctx
, &nif
->then_list
);
2293 emit_cf_list(ctx
, &nif
->else_list
);
2297 emit_loop(struct ir3_context
*ctx
, nir_loop
*nloop
)
2299 emit_cf_list(ctx
, &nloop
->body
);
2304 stack_push(struct ir3_context
*ctx
)
2307 ctx
->max_stack
= MAX2(ctx
->max_stack
, ctx
->stack
);
2311 stack_pop(struct ir3_context
*ctx
)
2313 compile_assert(ctx
, ctx
->stack
> 0);
2318 emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
)
2320 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2321 switch (node
->type
) {
2322 case nir_cf_node_block
:
2323 emit_block(ctx
, nir_cf_node_as_block(node
));
2325 case nir_cf_node_if
:
2327 emit_if(ctx
, nir_cf_node_as_if(node
));
2330 case nir_cf_node_loop
:
2332 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2335 case nir_cf_node_function
:
2336 ir3_context_error(ctx
, "TODO\n");
2342 /* emit stream-out code. At this point, the current block is the original
2343 * (nir) end block, and nir ensures that all flow control paths terminate
2344 * into the end block. We re-purpose the original end block to generate
2345 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2346 * block holding stream-out write instructions, followed by the new end
2350 * p0.x = (vtxcnt < maxvtxcnt)
2351 * // succs: blockStreamOut, blockNewEnd
2354 * ... stream-out instructions ...
2355 * // succs: blockNewEnd
2361 emit_stream_out(struct ir3_context
*ctx
)
2363 struct ir3
*ir
= ctx
->ir
;
2364 struct ir3_stream_output_info
*strmout
=
2365 &ctx
->so
->shader
->stream_output
;
2366 struct ir3_block
*orig_end_block
, *stream_out_block
, *new_end_block
;
2367 struct ir3_instruction
*vtxcnt
, *maxvtxcnt
, *cond
;
2368 struct ir3_instruction
*bases
[IR3_MAX_SO_BUFFERS
];
2370 /* create vtxcnt input in input block at top of shader,
2371 * so that it is seen as live over the entire duration
2374 vtxcnt
= create_input(ctx
, 0);
2375 add_sysval_input(ctx
, SYSTEM_VALUE_VERTEX_CNT
, vtxcnt
);
2377 maxvtxcnt
= create_driver_param(ctx
, IR3_DP_VTXCNT_MAX
);
2379 /* at this point, we are at the original 'end' block,
2380 * re-purpose this block to stream-out condition, then
2381 * append stream-out block and new-end block
2383 orig_end_block
= ctx
->block
;
2385 // TODO these blocks need to update predecessors..
2386 // maybe w/ store_global intrinsic, we could do this
2387 // stuff in nir->nir pass
2389 stream_out_block
= ir3_block_create(ir
);
2390 list_addtail(&stream_out_block
->node
, &ir
->block_list
);
2392 new_end_block
= ir3_block_create(ir
);
2393 list_addtail(&new_end_block
->node
, &ir
->block_list
);
2395 orig_end_block
->successors
[0] = stream_out_block
;
2396 orig_end_block
->successors
[1] = new_end_block
;
2397 stream_out_block
->successors
[0] = new_end_block
;
2399 /* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2400 cond
= ir3_CMPS_S(ctx
->block
, vtxcnt
, 0, maxvtxcnt
, 0);
2401 cond
->regs
[0]->num
= regid(REG_P0
, 0);
2402 cond
->cat2
.condition
= IR3_COND_LT
;
2404 /* condition goes on previous block to the conditional,
2405 * since it is used to pick which of the two successor
2408 orig_end_block
->condition
= cond
;
2410 /* switch to stream_out_block to generate the stream-out
2413 ctx
->block
= stream_out_block
;
2415 /* Calculate base addresses based on vtxcnt. Instructions
2416 * generated for bases not used in following loop will be
2417 * stripped out in the backend.
2419 for (unsigned i
= 0; i
< IR3_MAX_SO_BUFFERS
; i
++) {
2420 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
2421 unsigned stride
= strmout
->stride
[i
];
2422 struct ir3_instruction
*base
, *off
;
2424 base
= create_uniform(ctx
->block
, regid(const_state
->offsets
.tfbo
, i
));
2426 /* 24-bit should be enough: */
2427 off
= ir3_MUL_U(ctx
->block
, vtxcnt
, 0,
2428 create_immed(ctx
->block
, stride
* 4), 0);
2430 bases
[i
] = ir3_ADD_S(ctx
->block
, off
, 0, base
, 0);
2433 /* Generate the per-output store instructions: */
2434 for (unsigned i
= 0; i
< strmout
->num_outputs
; i
++) {
2435 for (unsigned j
= 0; j
< strmout
->output
[i
].num_components
; j
++) {
2436 unsigned c
= j
+ strmout
->output
[i
].start_component
;
2437 struct ir3_instruction
*base
, *out
, *stg
;
2439 base
= bases
[strmout
->output
[i
].output_buffer
];
2440 out
= ctx
->ir
->outputs
[regid(strmout
->output
[i
].register_index
, c
)];
2442 stg
= ir3_STG(ctx
->block
, base
, 0, out
, 0,
2443 create_immed(ctx
->block
, 1), 0);
2444 stg
->cat6
.type
= TYPE_U32
;
2445 stg
->cat6
.dst_offset
= (strmout
->output
[i
].dst_offset
+ j
) * 4;
2447 array_insert(ctx
->block
, ctx
->block
->keeps
, stg
);
2451 /* and finally switch to the new_end_block: */
2452 ctx
->block
= new_end_block
;
2456 emit_function(struct ir3_context
*ctx
, nir_function_impl
*impl
)
2458 nir_metadata_require(impl
, nir_metadata_block_index
);
2460 compile_assert(ctx
, ctx
->stack
== 0);
2462 emit_cf_list(ctx
, &impl
->body
);
2463 emit_block(ctx
, impl
->end_block
);
2465 compile_assert(ctx
, ctx
->stack
== 0);
2467 /* at this point, we should have a single empty block,
2468 * into which we emit the 'end' instruction.
2470 compile_assert(ctx
, list_empty(&ctx
->block
->instr_list
));
2472 /* If stream-out (aka transform-feedback) enabled, emit the
2473 * stream-out instructions, followed by a new empty block (into
2474 * which the 'end' instruction lands).
2476 * NOTE: it is done in this order, rather than inserting before
2477 * we emit end_block, because NIR guarantees that all blocks
2478 * flow into end_block, and that end_block has no successors.
2479 * So by re-purposing end_block as the first block of stream-
2480 * out, we guarantee that all exit paths flow into the stream-
2483 if ((ctx
->compiler
->gpu_id
< 500) &&
2484 (ctx
->so
->shader
->stream_output
.num_outputs
> 0) &&
2485 !ctx
->so
->binning_pass
) {
2486 debug_assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
2487 emit_stream_out(ctx
);
2490 /* Vertex shaders in a tessellation or geometry pipeline treat END as a
2491 * NOP and has an epilogue that writes the VS outputs to local storage, to
2492 * be read by the HS. Then it resets execution mask (chmask) and chains
2493 * to the next shader (chsh).
2495 if (ctx
->so
->type
== MESA_SHADER_VERTEX
&& ctx
->so
->key
.has_gs
) {
2496 struct ir3_instruction
*chmask
=
2497 ir3_CHMASK(ctx
->block
);
2498 chmask
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2499 chmask
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2501 struct ir3_instruction
*chsh
=
2502 ir3_CHSH(ctx
->block
);
2503 chsh
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2504 chsh
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2506 ir3_END(ctx
->block
);
2511 setup_input(struct ir3_context
*ctx
, nir_variable
*in
)
2513 struct ir3_shader_variant
*so
= ctx
->so
;
2514 unsigned ncomp
= glsl_get_components(in
->type
);
2515 unsigned n
= in
->data
.driver_location
;
2516 unsigned frac
= in
->data
.location_frac
;
2517 unsigned slot
= in
->data
.location
;
2519 /* Inputs are loaded using ldlw or ldg for these stages. */
2520 if (ctx
->so
->type
== MESA_SHADER_TESS_CTRL
||
2521 ctx
->so
->type
== MESA_SHADER_TESS_EVAL
||
2522 ctx
->so
->type
== MESA_SHADER_GEOMETRY
)
2525 /* skip unread inputs, we could end up with (for example), unsplit
2526 * matrix/etc inputs in the case they are not read, so just silently
2532 so
->inputs
[n
].slot
= slot
;
2533 so
->inputs
[n
].compmask
= (1 << (ncomp
+ frac
)) - 1;
2534 so
->inputs_count
= MAX2(so
->inputs_count
, n
+ 1);
2535 so
->inputs
[n
].interpolate
= in
->data
.interpolation
;
2537 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2539 /* if any varyings have 'sample' qualifer, that triggers us
2540 * to run in per-sample mode:
2542 so
->per_samp
|= in
->data
.sample
;
2544 for (int i
= 0; i
< ncomp
; i
++) {
2545 struct ir3_instruction
*instr
= NULL
;
2546 unsigned idx
= (n
* 4) + i
+ frac
;
2548 if (slot
== VARYING_SLOT_POS
) {
2549 ir3_context_error(ctx
, "fragcoord should be a sysval!\n");
2550 } else if (slot
== VARYING_SLOT_PNTC
) {
2551 /* see for example st_nir_fixup_varying_slots().. this is
2552 * maybe a bit mesa/st specific. But we need things to line
2553 * up for this in fdN_program:
2554 * unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
2555 * if (emit->sprite_coord_enable & texmask) {
2559 so
->inputs
[n
].slot
= VARYING_SLOT_VAR8
;
2560 so
->inputs
[n
].bary
= true;
2561 instr
= create_frag_input(ctx
, false, idx
);
2563 /* detect the special case for front/back colors where
2564 * we need to do flat vs smooth shading depending on
2567 if (in
->data
.interpolation
== INTERP_MODE_NONE
) {
2569 case VARYING_SLOT_COL0
:
2570 case VARYING_SLOT_COL1
:
2571 case VARYING_SLOT_BFC0
:
2572 case VARYING_SLOT_BFC1
:
2573 so
->inputs
[n
].rasterflat
= true;
2580 if (ctx
->compiler
->flat_bypass
) {
2581 if ((so
->inputs
[n
].interpolate
== INTERP_MODE_FLAT
) ||
2582 (so
->inputs
[n
].rasterflat
&& ctx
->so
->key
.rasterflat
))
2583 so
->inputs
[n
].use_ldlv
= true;
2586 so
->inputs
[n
].bary
= true;
2588 instr
= create_frag_input(ctx
, so
->inputs
[n
].use_ldlv
, idx
);
2591 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2593 ctx
->ir
->inputs
[idx
] = instr
;
2595 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2596 for (int i
= 0; i
< ncomp
; i
++) {
2597 unsigned idx
= (n
* 4) + i
+ frac
;
2598 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2599 ctx
->ir
->inputs
[idx
] = create_input(ctx
, idx
);
2602 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2605 if (so
->inputs
[n
].bary
|| (ctx
->so
->type
== MESA_SHADER_VERTEX
)) {
2606 so
->total_in
+= ncomp
;
2610 /* Initially we assign non-packed inloc's for varyings, as we don't really
2611 * know up-front which components will be unused. After all the compilation
2612 * stages we scan the shader to see which components are actually used, and
2613 * re-pack the inlocs to eliminate unneeded varyings.
2616 pack_inlocs(struct ir3_context
*ctx
)
2618 struct ir3_shader_variant
*so
= ctx
->so
;
2619 uint8_t used_components
[so
->inputs_count
];
2621 memset(used_components
, 0, sizeof(used_components
));
2624 * First Step: scan shader to find which bary.f/ldlv remain:
2627 list_for_each_entry (struct ir3_block
, block
, &ctx
->ir
->block_list
, node
) {
2628 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
2629 if (is_input(instr
)) {
2630 unsigned inloc
= instr
->regs
[1]->iim_val
;
2631 unsigned i
= inloc
/ 4;
2632 unsigned j
= inloc
% 4;
2634 compile_assert(ctx
, instr
->regs
[1]->flags
& IR3_REG_IMMED
);
2635 compile_assert(ctx
, i
< so
->inputs_count
);
2637 used_components
[i
] |= 1 << j
;
2638 } else if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
2639 for (int n
= 0; n
< 2; n
++) {
2640 unsigned inloc
= instr
->prefetch
.input_offset
+ n
;
2641 unsigned i
= inloc
/ 4;
2642 unsigned j
= inloc
% 4;
2644 compile_assert(ctx
, i
< so
->inputs_count
);
2646 used_components
[i
] |= 1 << j
;
2653 * Second Step: reassign varying inloc/slots:
2656 unsigned actual_in
= 0;
2659 for (unsigned i
= 0; i
< so
->inputs_count
; i
++) {
2660 unsigned compmask
= 0, maxcomp
= 0;
2662 so
->inputs
[i
].inloc
= inloc
;
2663 so
->inputs
[i
].bary
= false;
2665 for (unsigned j
= 0; j
< 4; j
++) {
2666 if (!(used_components
[i
] & (1 << j
)))
2669 compmask
|= (1 << j
);
2673 /* at this point, since used_components[i] mask is only
2674 * considering varyings (ie. not sysvals) we know this
2677 so
->inputs
[i
].bary
= true;
2680 if (so
->inputs
[i
].bary
) {
2682 so
->inputs
[i
].compmask
= (1 << maxcomp
) - 1;
2688 * Third Step: reassign packed inloc's:
2691 list_for_each_entry (struct ir3_block
, block
, &ctx
->ir
->block_list
, node
) {
2692 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
2693 if (is_input(instr
)) {
2694 unsigned inloc
= instr
->regs
[1]->iim_val
;
2695 unsigned i
= inloc
/ 4;
2696 unsigned j
= inloc
% 4;
2698 instr
->regs
[1]->iim_val
= so
->inputs
[i
].inloc
+ j
;
2705 setup_output(struct ir3_context
*ctx
, nir_variable
*out
)
2707 struct ir3_shader_variant
*so
= ctx
->so
;
2708 unsigned ncomp
= glsl_get_components(out
->type
);
2709 unsigned n
= out
->data
.driver_location
;
2710 unsigned frac
= out
->data
.location_frac
;
2711 unsigned slot
= out
->data
.location
;
2714 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2716 case FRAG_RESULT_DEPTH
:
2717 comp
= 2; /* tgsi will write to .z component */
2718 so
->writes_pos
= true;
2720 case FRAG_RESULT_COLOR
:
2723 case FRAG_RESULT_SAMPLE_MASK
:
2724 so
->writes_smask
= true;
2727 if (slot
>= FRAG_RESULT_DATA0
)
2729 ir3_context_error(ctx
, "unknown FS output name: %s\n",
2730 gl_frag_result_name(slot
));
2732 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
2733 ctx
->so
->type
== MESA_SHADER_GEOMETRY
) {
2735 case VARYING_SLOT_POS
:
2736 so
->writes_pos
= true;
2738 case VARYING_SLOT_PSIZ
:
2739 so
->writes_psize
= true;
2741 case VARYING_SLOT_PRIMITIVE_ID
:
2742 case VARYING_SLOT_LAYER
:
2743 case VARYING_SLOT_GS_VERTEX_FLAGS_IR3
:
2744 debug_assert(ctx
->so
->type
== MESA_SHADER_GEOMETRY
);
2746 case VARYING_SLOT_COL0
:
2747 case VARYING_SLOT_COL1
:
2748 case VARYING_SLOT_BFC0
:
2749 case VARYING_SLOT_BFC1
:
2750 case VARYING_SLOT_FOGC
:
2751 case VARYING_SLOT_CLIP_DIST0
:
2752 case VARYING_SLOT_CLIP_DIST1
:
2753 case VARYING_SLOT_CLIP_VERTEX
:
2756 if (slot
>= VARYING_SLOT_VAR0
)
2758 if ((VARYING_SLOT_TEX0
<= slot
) && (slot
<= VARYING_SLOT_TEX7
))
2760 ir3_context_error(ctx
, "unknown %s shader output name: %s\n",
2761 _mesa_shader_stage_to_string(ctx
->so
->type
),
2762 gl_varying_slot_name(slot
));
2765 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2768 compile_assert(ctx
, n
< ARRAY_SIZE(so
->outputs
));
2770 so
->outputs
[n
].slot
= slot
;
2771 so
->outputs
[n
].regid
= regid(n
, comp
);
2772 so
->outputs_count
= MAX2(so
->outputs_count
, n
+ 1);
2774 for (int i
= 0; i
< ncomp
; i
++) {
2775 unsigned idx
= (n
* 4) + i
+ frac
;
2776 compile_assert(ctx
, idx
< ctx
->ir
->noutputs
);
2777 ctx
->ir
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
2780 /* if varying packing doesn't happen, we could end up in a situation
2781 * with "holes" in the output, and since the per-generation code that
2782 * sets up varying linkage registers doesn't expect to have more than
2783 * one varying per vec4 slot, pad the holes.
2785 * Note that this should probably generate a performance warning of
2788 for (int i
= 0; i
< frac
; i
++) {
2789 unsigned idx
= (n
* 4) + i
;
2790 if (!ctx
->ir
->outputs
[idx
]) {
2791 ctx
->ir
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
2797 max_drvloc(struct exec_list
*vars
)
2800 nir_foreach_variable(var
, vars
) {
2801 drvloc
= MAX2(drvloc
, (int)var
->data
.driver_location
);
2806 static const unsigned max_sysvals
[] = {
2807 [MESA_SHADER_VERTEX
] = 16,
2808 [MESA_SHADER_GEOMETRY
] = 16,
2809 [MESA_SHADER_FRAGMENT
] = 24, // TODO
2810 [MESA_SHADER_COMPUTE
] = 16, // TODO how many do we actually need?
2811 [MESA_SHADER_KERNEL
] = 16, // TODO how many do we actually need?
2815 emit_instructions(struct ir3_context
*ctx
)
2817 unsigned ninputs
, noutputs
;
2818 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->s
);
2820 ninputs
= (max_drvloc(&ctx
->s
->inputs
) + 1) * 4;
2821 noutputs
= (max_drvloc(&ctx
->s
->outputs
) + 1) * 4;
2823 /* we need to leave room for sysvals:
2825 ninputs
+= max_sysvals
[ctx
->so
->type
];
2826 if (ctx
->so
->type
== MESA_SHADER_VERTEX
)
2827 noutputs
+= 8; /* gs or tess header + primitive_id */
2829 ctx
->ir
= ir3_create(ctx
->compiler
, ctx
->so
->type
, ninputs
, noutputs
);
2831 /* Create inputs in first block: */
2832 ctx
->block
= get_block(ctx
, nir_start_block(fxn
));
2833 ctx
->in_block
= ctx
->block
;
2834 list_addtail(&ctx
->block
->node
, &ctx
->ir
->block_list
);
2836 ninputs
-= max_sysvals
[ctx
->so
->type
];
2838 if (ctx
->so
->key
.has_gs
) {
2839 if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
2840 ctx
->so
->type
== MESA_SHADER_GEOMETRY
) {
2841 ctx
->gs_header
= create_input(ctx
, 0);
2842 ctx
->primitive_id
= create_input(ctx
, 0);
2846 /* for fragment shader, the vcoord input register is used as the
2847 * base for bary.f varying fetch instrs:
2849 * TODO defer creating ctx->ij_pixel and corresponding sysvals
2850 * until emit_intrinsic when we know they are actually needed.
2851 * For now, we defer creating ctx->ij_centroid, etc, since we
2852 * only need ij_pixel for "old style" varying inputs (ie.
2855 struct ir3_instruction
*vcoord
= NULL
;
2856 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2857 struct ir3_instruction
*xy
[2];
2859 vcoord
= create_input_compmask(ctx
, 0, 0x3);
2860 ir3_split_dest(ctx
->block
, xy
, vcoord
, 0, 2);
2862 ctx
->ij_pixel
= ir3_create_collect(ctx
, xy
, 2);
2866 nir_foreach_variable(var
, &ctx
->s
->inputs
) {
2867 setup_input(ctx
, var
);
2870 /* Defer add_sysval_input() stuff until after setup_inputs(),
2871 * because sysvals need to be appended after varyings:
2874 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_BARYCENTRIC_PIXEL
,
2878 if (ctx
->primitive_id
)
2879 add_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, ctx
->primitive_id
);
2881 add_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, ctx
->gs_header
);
2883 /* Setup outputs: */
2884 nir_foreach_variable(var
, &ctx
->s
->outputs
) {
2885 setup_output(ctx
, var
);
2888 /* Set up the gs header as an output for the vertex shader so it won't
2889 * clobber it for the tess ctrl shader. */
2890 if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2891 struct ir3_shader_variant
*so
= ctx
->so
;
2892 if (ctx
->primitive_id
) {
2893 unsigned n
= so
->outputs_count
++;
2894 so
->outputs
[n
].slot
= VARYING_SLOT_PRIMITIVE_ID
;
2895 so
->outputs
[n
].regid
= regid(n
, 0);
2896 ctx
->ir
->outputs
[n
* 4] = ctx
->primitive_id
;
2898 compile_assert(ctx
, n
* 4 < ctx
->ir
->noutputs
);
2901 if (ctx
->gs_header
) {
2902 unsigned n
= so
->outputs_count
++;
2903 so
->outputs
[n
].slot
= VARYING_SLOT_GS_HEADER_IR3
;
2904 so
->outputs
[n
].regid
= regid(n
, 0);
2905 ctx
->ir
->outputs
[n
* 4] = ctx
->gs_header
;
2907 compile_assert(ctx
, n
* 4 < ctx
->ir
->noutputs
);
2912 /* Find # of samplers: */
2913 nir_foreach_variable(var
, &ctx
->s
->uniforms
) {
2914 ctx
->so
->num_samp
+= glsl_type_get_sampler_count(var
->type
);
2915 /* just assume that we'll be reading from images.. if it
2916 * is write-only we don't have to count it, but not sure
2917 * if there is a good way to know?
2919 ctx
->so
->num_samp
+= glsl_type_get_image_count(var
->type
);
2922 /* NOTE: need to do something more clever when we support >1 fxn */
2923 nir_foreach_register(reg
, &fxn
->registers
) {
2924 ir3_declare_array(ctx
, reg
);
2926 /* And emit the body: */
2928 emit_function(ctx
, fxn
);
2931 /* from NIR perspective, we actually have varying inputs. But the varying
2932 * inputs, from an IR standpoint, are just bary.f/ldlv instructions. The
2933 * only actual inputs are the sysvals.
2936 fixup_frag_inputs(struct ir3_context
*ctx
)
2938 struct ir3_shader_variant
*so
= ctx
->so
;
2939 struct ir3
*ir
= ctx
->ir
;
2942 /* sysvals should appear at the end of the inputs, drop everything else: */
2943 while ((i
< so
->inputs_count
) && !so
->inputs
[i
].sysval
)
2946 /* at IR level, inputs are always blocks of 4 scalars: */
2949 ir
->inputs
= &ir
->inputs
[i
];
2953 /* Fixup tex sampler state for astc/srgb workaround instructions. We
2954 * need to assign the tex state indexes for these after we know the
2958 fixup_astc_srgb(struct ir3_context
*ctx
)
2960 struct ir3_shader_variant
*so
= ctx
->so
;
2961 /* indexed by original tex idx, value is newly assigned alpha sampler
2962 * state tex idx. Zero is invalid since there is at least one sampler
2965 unsigned alt_tex_state
[16] = {0};
2966 unsigned tex_idx
= ctx
->max_texture_index
+ 1;
2969 so
->astc_srgb
.base
= tex_idx
;
2971 for (unsigned i
= 0; i
< ctx
->ir
->astc_srgb_count
; i
++) {
2972 struct ir3_instruction
*sam
= ctx
->ir
->astc_srgb
[i
];
2974 compile_assert(ctx
, sam
->cat5
.tex
< ARRAY_SIZE(alt_tex_state
));
2976 if (alt_tex_state
[sam
->cat5
.tex
] == 0) {
2977 /* assign new alternate/alpha tex state slot: */
2978 alt_tex_state
[sam
->cat5
.tex
] = tex_idx
++;
2979 so
->astc_srgb
.orig_idx
[idx
++] = sam
->cat5
.tex
;
2980 so
->astc_srgb
.count
++;
2983 sam
->cat5
.tex
= alt_tex_state
[sam
->cat5
.tex
];
2988 fixup_binning_pass(struct ir3_context
*ctx
)
2990 struct ir3_shader_variant
*so
= ctx
->so
;
2991 struct ir3
*ir
= ctx
->ir
;
2994 for (i
= 0, j
= 0; i
< so
->outputs_count
; i
++) {
2995 unsigned slot
= so
->outputs
[i
].slot
;
2997 /* throw away everything but first position/psize */
2998 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
3000 so
->outputs
[j
] = so
->outputs
[i
];
3001 ir
->outputs
[(j
*4)+0] = ir
->outputs
[(i
*4)+0];
3002 ir
->outputs
[(j
*4)+1] = ir
->outputs
[(i
*4)+1];
3003 ir
->outputs
[(j
*4)+2] = ir
->outputs
[(i
*4)+2];
3004 ir
->outputs
[(j
*4)+3] = ir
->outputs
[(i
*4)+3];
3009 so
->outputs_count
= j
;
3010 ir
->noutputs
= j
* 4;
3014 ir3_compile_shader_nir(struct ir3_compiler
*compiler
,
3015 struct ir3_shader_variant
*so
)
3017 struct ir3_context
*ctx
;
3019 struct ir3_instruction
**inputs
;
3021 int ret
= 0, max_bary
;
3025 ctx
= ir3_context_init(compiler
, so
);
3027 DBG("INIT failed!");
3032 emit_instructions(ctx
);
3035 DBG("EMIT failed!");
3040 ir
= so
->ir
= ctx
->ir
;
3042 /* keep track of the inputs from TGSI perspective.. */
3043 inputs
= ir
->inputs
;
3045 /* but fixup actual inputs for frag shader: */
3046 if (so
->type
== MESA_SHADER_FRAGMENT
)
3047 fixup_frag_inputs(ctx
);
3049 /* at this point, for binning pass, throw away unneeded outputs: */
3050 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
< 600))
3051 fixup_binning_pass(ctx
);
3053 /* if we want half-precision outputs, mark the output registers
3056 if (so
->key
.half_precision
) {
3057 for (i
= 0; i
< ir
->noutputs
; i
++) {
3058 struct ir3_instruction
*out
= ir
->outputs
[i
];
3063 /* if frag shader writes z, that needs to be full precision: */
3064 if (so
->outputs
[i
/4].slot
== FRAG_RESULT_DEPTH
)
3067 out
->regs
[0]->flags
|= IR3_REG_HALF
;
3068 /* output could be a fanout (ie. texture fetch output)
3069 * in which case we need to propagate the half-reg flag
3070 * up to the definer so that RA sees it:
3072 if (out
->opc
== OPC_META_FO
) {
3073 out
= out
->regs
[1]->instr
;
3074 out
->regs
[0]->flags
|= IR3_REG_HALF
;
3077 if (out
->opc
== OPC_MOV
) {
3078 out
->cat1
.dst_type
= half_type(out
->cat1
.dst_type
);
3083 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3084 printf("BEFORE CP:\n");
3090 /* at this point, for binning pass, throw away unneeded outputs:
3091 * Note that for a6xx and later, we do this after ir3_cp to ensure
3092 * that the uniform/constant layout for BS and VS matches, so that
3093 * we can re-use same VS_CONST state group.
3095 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
>= 600))
3096 fixup_binning_pass(ctx
);
3098 /* for a6xx+, binning and draw pass VS use same VBO state, so we
3099 * need to make sure not to remove any inputs that are used by
3100 * the nonbinning VS.
3102 if (ctx
->compiler
->gpu_id
>= 600 && so
->binning_pass
) {
3103 debug_assert(so
->type
== MESA_SHADER_VERTEX
);
3104 for (int i
= 0; i
< ir
->ninputs
; i
++) {
3105 struct ir3_instruction
*in
= ir
->inputs
[i
];
3113 debug_assert(n
< so
->nonbinning
->inputs_count
);
3115 if (so
->nonbinning
->inputs
[n
].sysval
)
3118 /* be sure to keep inputs, even if only used in VS */
3119 if (so
->nonbinning
->inputs
[n
].compmask
& (1 << c
))
3120 array_insert(in
->block
, in
->block
->keeps
, in
);
3124 /* Insert mov if there's same instruction for each output.
3125 * eg. dEQP-GLES31.functional.shaders.opaque_type_indexing.sampler.const_expression.vertex.sampler2dshadow
3127 for (int i
= ir
->noutputs
- 1; i
>= 0; i
--) {
3128 if (!ir
->outputs
[i
])
3130 for (unsigned j
= 0; j
< i
; j
++) {
3131 if (ir
->outputs
[i
] == ir
->outputs
[j
]) {
3133 ir3_MOV(ir
->outputs
[i
]->block
, ir
->outputs
[i
], TYPE_F32
);
3138 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3139 printf("BEFORE GROUPING:\n");
3143 ir3_sched_add_deps(ir
);
3145 /* Group left/right neighbors, inserting mov's where needed to
3150 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3151 printf("AFTER GROUPING:\n");
3157 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3158 printf("AFTER DEPTH:\n");
3162 /* do Sethi–Ullman numbering before scheduling: */
3165 ret
= ir3_sched(ir
);
3167 DBG("SCHED failed!");
3171 if (compiler
->gpu_id
>= 600) {
3172 ir3_a6xx_fixup_atomic_dests(ir
, so
);
3175 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3176 printf("AFTER SCHED:\n");
3180 /* Pre-assign VS inputs on a6xx+ binning pass shader, to align
3181 * with draw pass VS, so binning and draw pass can both use the
3184 * Note that VS inputs are expected to be full precision.
3186 bool pre_assign_inputs
= (ir
->compiler
->gpu_id
>= 600) &&
3187 (ir
->type
== MESA_SHADER_VERTEX
) &&
3190 if (pre_assign_inputs
) {
3191 for (unsigned i
= 0; i
< ir
->ninputs
; i
++) {
3192 struct ir3_instruction
*instr
= ir
->inputs
[i
];
3199 unsigned regid
= so
->nonbinning
->inputs
[n
].regid
+ c
;
3201 instr
->regs
[0]->num
= regid
;
3204 ret
= ir3_ra(so
, ir
->inputs
, ir
->ninputs
);
3205 } else if (ctx
->gs_header
) {
3206 /* We need to have these values in the same registers between VS and GS
3207 * since the VS chains to GS and doesn't get the sysvals redelivered.
3210 ctx
->gs_header
->regs
[0]->num
= 0;
3211 ctx
->primitive_id
->regs
[0]->num
= 1;
3212 struct ir3_instruction
*precolor
[] = { ctx
->gs_header
, ctx
->primitive_id
};
3213 ret
= ir3_ra(so
, precolor
, ARRAY_SIZE(precolor
));
3215 ret
= ir3_ra(so
, NULL
, 0);
3223 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3224 printf("AFTER RA:\n");
3228 if (so
->type
== MESA_SHADER_FRAGMENT
)
3231 /* fixup input/outputs: */
3232 for (i
= 0; i
< so
->outputs_count
; i
++) {
3233 /* sometimes we get outputs that don't write the .x coord, like:
3235 * decl_var shader_out INTERP_MODE_NONE float Color (VARYING_SLOT_VAR9.z, 1, 0)
3237 * Presumably the result of varying packing and then eliminating
3238 * some unneeded varyings? Just skip head to the first valid
3239 * component of the output.
3241 for (unsigned j
= 0; j
< 4; j
++) {
3242 struct ir3_instruction
*instr
= ir
->outputs
[(i
*4) + j
];
3244 so
->outputs
[i
].regid
= instr
->regs
[0]->num
;
3245 so
->outputs
[i
].half
= !!(instr
->regs
[0]->flags
& IR3_REG_HALF
);
3251 /* Note that some or all channels of an input may be unused: */
3252 for (i
= 0; i
< so
->inputs_count
; i
++) {
3253 unsigned j
, reg
= regid(63,0);
3255 for (j
= 0; j
< 4; j
++) {
3256 struct ir3_instruction
*in
= inputs
[(i
*4) + j
];
3261 if (in
->flags
& IR3_INSTR_UNUSED
)
3264 reg
= in
->regs
[0]->num
- j
;
3266 compile_assert(ctx
, in
->regs
[0]->flags
& IR3_REG_HALF
);
3268 half
= !!(in
->regs
[0]->flags
& IR3_REG_HALF
);
3271 so
->inputs
[i
].regid
= reg
;
3272 so
->inputs
[i
].half
= half
;
3276 fixup_astc_srgb(ctx
);
3278 /* We need to do legalize after (for frag shader's) the "bary.f"
3279 * offsets (inloc) have been assigned.
3281 ir3_legalize(ir
, &so
->has_ssbo
, &so
->need_pixlod
, &max_bary
);
3283 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3284 printf("AFTER LEGALIZE:\n");
3288 /* Set (ss)(sy) on first TCS and GEOMETRY instructions, since we don't
3289 * know what we might have to wait on when coming in from VS chsh.
3291 if (so
->type
== MESA_SHADER_TESS_CTRL
||
3292 so
->type
== MESA_SHADER_GEOMETRY
) {
3293 list_for_each_entry (struct ir3_block
, block
, &ir
->block_list
, node
) {
3294 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
3295 instr
->flags
|= IR3_INSTR_SS
| IR3_INSTR_SY
;
3301 so
->branchstack
= ctx
->max_stack
;
3303 /* Note that actual_in counts inputs that are not bary.f'd for FS: */
3304 if (so
->type
== MESA_SHADER_FRAGMENT
)
3305 so
->total_in
= max_bary
+ 1;
3307 so
->max_sun
= ir
->max_sun
;
3312 ir3_destroy(so
->ir
);
3315 ir3_context_free(ctx
);