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
->inout
.block
= ctx
->in_block
;
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
;
298 if (alu
->dest
.dest
.is_ssa
) {
299 dst_sz
= alu
->dest
.dest
.ssa
.num_components
;
300 wrmask
= (1 << dst_sz
) - 1;
302 dst_sz
= alu
->dest
.dest
.reg
.reg
->num_components
;
303 wrmask
= alu
->dest
.write_mask
;
306 dst
= ir3_get_dst(ctx
, &alu
->dest
.dest
, dst_sz
);
308 /* Vectors are special in that they have non-scalarized writemasks,
309 * and just take the first swizzle channel for each argument in
310 * order into each writemask channel.
312 if ((alu
->op
== nir_op_vec2
) ||
313 (alu
->op
== nir_op_vec3
) ||
314 (alu
->op
== nir_op_vec4
)) {
316 for (int i
= 0; i
< info
->num_inputs
; i
++) {
317 nir_alu_src
*asrc
= &alu
->src
[i
];
319 compile_assert(ctx
, !asrc
->abs
);
320 compile_assert(ctx
, !asrc
->negate
);
322 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[0]];
324 src
[i
] = create_immed(ctx
->block
, 0);
325 dst
[i
] = ir3_MOV(b
, src
[i
], TYPE_U32
);
328 ir3_put_dst(ctx
, &alu
->dest
.dest
);
332 /* We also get mov's with more than one component for mov's so
333 * handle those specially:
335 if (alu
->op
== nir_op_mov
) {
336 type_t type
= TYPE_U32
;
337 nir_alu_src
*asrc
= &alu
->src
[0];
338 struct ir3_instruction
*const *src0
= ir3_get_src(ctx
, &asrc
->src
);
340 for (unsigned i
= 0; i
< dst_sz
; i
++) {
341 if (wrmask
& (1 << i
)) {
342 dst
[i
] = ir3_MOV(b
, src0
[asrc
->swizzle
[i
]], type
);
348 ir3_put_dst(ctx
, &alu
->dest
.dest
);
352 /* General case: We can just grab the one used channel per src. */
353 for (int i
= 0; i
< info
->num_inputs
; i
++) {
354 unsigned chan
= ffs(alu
->dest
.write_mask
) - 1;
355 nir_alu_src
*asrc
= &alu
->src
[i
];
357 compile_assert(ctx
, !asrc
->abs
);
358 compile_assert(ctx
, !asrc
->negate
);
360 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[chan
]];
361 bs
[i
] = nir_src_bit_size(asrc
->src
);
363 compile_assert(ctx
, src
[i
]);
368 case nir_op_f2f16_rtne
:
369 case nir_op_f2f16_rtz
:
387 dst
[0] = create_cov(ctx
, src
[0], bs
[0], alu
->op
);
390 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, create_immed(b
, fui(0.0)), 0);
391 dst
[0]->cat2
.condition
= IR3_COND_NE
;
392 dst
[0] = ir3_n2b(b
, dst
[0]);
396 dst
[0] = ir3_COV(b
, ir3_b2n(b
, src
[0]), TYPE_U32
, TYPE_F32
);
401 dst
[0] = ir3_b2n(b
, src
[0]);
404 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, create_immed(b
, 0), 0);
405 dst
[0]->cat2
.condition
= IR3_COND_NE
;
406 dst
[0] = ir3_n2b(b
, dst
[0]);
410 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FNEG
);
413 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FABS
);
416 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[1], 0);
419 dst
[0] = ir3_MIN_F(b
, src
[0], 0, src
[1], 0);
422 /* if there is just a single use of the src, and it supports
423 * (sat) bit, we can just fold the (sat) flag back to the
424 * src instruction and create a mov. This is easier for cp
427 * TODO probably opc_cat==4 is ok too
429 if (alu
->src
[0].src
.is_ssa
&&
430 (list_length(&alu
->src
[0].src
.ssa
->uses
) == 1) &&
431 ((opc_cat(src
[0]->opc
) == 2) || (opc_cat(src
[0]->opc
) == 3))) {
432 src
[0]->flags
|= IR3_INSTR_SAT
;
433 dst
[0] = ir3_MOV(b
, src
[0], TYPE_U32
);
435 /* otherwise generate a max.f that saturates.. blob does
436 * similar (generating a cat2 mov using max.f)
438 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[0], 0);
439 dst
[0]->flags
|= IR3_INSTR_SAT
;
443 dst
[0] = ir3_MUL_F(b
, src
[0], 0, src
[1], 0);
446 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], 0);
449 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], IR3_REG_FNEG
);
452 dst
[0] = ir3_MAD_F32(b
, src
[0], 0, src
[1], 0, src
[2], 0);
455 dst
[0] = ir3_DSX(b
, src
[0], 0);
456 dst
[0]->cat5
.type
= TYPE_F32
;
459 dst
[0] = ir3_DSY(b
, src
[0], 0);
460 dst
[0]->cat5
.type
= TYPE_F32
;
464 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
465 dst
[0]->cat2
.condition
= IR3_COND_LT
;
466 dst
[0] = ir3_n2b(b
, dst
[0]);
469 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
470 dst
[0]->cat2
.condition
= IR3_COND_GE
;
471 dst
[0] = ir3_n2b(b
, dst
[0]);
474 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
475 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
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_NE
;
481 dst
[0] = ir3_n2b(b
, dst
[0]);
484 dst
[0] = ir3_CEIL_F(b
, src
[0], 0);
487 dst
[0] = ir3_FLOOR_F(b
, src
[0], 0);
490 dst
[0] = ir3_TRUNC_F(b
, src
[0], 0);
492 case nir_op_fround_even
:
493 dst
[0] = ir3_RNDNE_F(b
, src
[0], 0);
496 dst
[0] = ir3_SIGN_F(b
, src
[0], 0);
500 dst
[0] = ir3_SIN(b
, src
[0], 0);
503 dst
[0] = ir3_COS(b
, src
[0], 0);
506 dst
[0] = ir3_RSQ(b
, src
[0], 0);
509 dst
[0] = ir3_RCP(b
, src
[0], 0);
512 dst
[0] = ir3_LOG2(b
, src
[0], 0);
515 dst
[0] = ir3_EXP2(b
, src
[0], 0);
518 dst
[0] = ir3_SQRT(b
, src
[0], 0);
522 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SABS
);
525 dst
[0] = ir3_ADD_U(b
, src
[0], 0, src
[1], 0);
528 dst
[0] = ir3_AND_B(b
, src
[0], 0, src
[1], 0);
531 dst
[0] = ir3_MAX_S(b
, src
[0], 0, src
[1], 0);
534 dst
[0] = ir3_MAX_U(b
, src
[0], 0, src
[1], 0);
537 dst
[0] = ir3_MIN_S(b
, src
[0], 0, src
[1], 0);
540 dst
[0] = ir3_MIN_U(b
, src
[0], 0, src
[1], 0);
544 * dst = (al * bl) + (ah * bl << 16) + (al * bh << 16)
545 * mull.u tmp0, a, b ; mul low, i.e. al * bl
546 * madsh.m16 tmp1, a, b, tmp0 ; mul-add shift high mix, i.e. ah * bl << 16
547 * madsh.m16 dst, b, a, tmp1 ; i.e. al * bh << 16
549 dst
[0] = ir3_MADSH_M16(b
, src
[1], 0, src
[0], 0,
550 ir3_MADSH_M16(b
, src
[0], 0, src
[1], 0,
551 ir3_MULL_U(b
, src
[0], 0, src
[1], 0), 0), 0);
554 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SNEG
);
557 dst
[0] = ir3_NOT_B(b
, src
[0], 0);
560 dst
[0] = ir3_OR_B(b
, src
[0], 0, src
[1], 0);
563 dst
[0] = ir3_SHL_B(b
, src
[0], 0, src
[1], 0);
566 dst
[0] = ir3_ASHR_B(b
, src
[0], 0, src
[1], 0);
569 dst
[0] = ir3_SUB_U(b
, src
[0], 0, src
[1], 0);
572 dst
[0] = ir3_XOR_B(b
, src
[0], 0, src
[1], 0);
575 dst
[0] = ir3_SHR_B(b
, src
[0], 0, src
[1], 0);
578 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
579 dst
[0]->cat2
.condition
= IR3_COND_LT
;
580 dst
[0] = ir3_n2b(b
, dst
[0]);
583 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
584 dst
[0]->cat2
.condition
= IR3_COND_GE
;
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_EQ
;
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_NE
;
595 dst
[0] = ir3_n2b(b
, dst
[0]);
598 dst
[0] = ir3_CMPS_U(b
, src
[0], 0, src
[1], 0);
599 dst
[0]->cat2
.condition
= IR3_COND_LT
;
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_GE
;
605 dst
[0] = ir3_n2b(b
, dst
[0]);
608 case nir_op_b32csel
: {
609 struct ir3_instruction
*cond
= ir3_b2n(b
, src
[0]);
610 compile_assert(ctx
, bs
[1] == bs
[2]);
611 /* the boolean condition is 32b even if src[1] and src[2] are
612 * half-precision, but sel.b16 wants all three src's to be the
616 cond
= ir3_COV(b
, cond
, TYPE_U32
, TYPE_U16
);
617 dst
[0] = ir3_SEL_B32(b
, src
[1], 0, cond
, 0, src
[2], 0);
620 case nir_op_bit_count
: {
621 // TODO, we need to do this 16b at a time on a5xx+a6xx.. need to
622 // double check on earlier gen's. Once half-precision support is
623 // in place, this should probably move to a NIR lowering pass:
624 struct ir3_instruction
*hi
, *lo
;
626 hi
= ir3_COV(b
, ir3_SHR_B(b
, src
[0], 0, create_immed(b
, 16), 0),
628 lo
= ir3_COV(b
, src
[0], TYPE_U32
, TYPE_U16
);
630 hi
= ir3_CBITS_B(b
, hi
, 0);
631 lo
= ir3_CBITS_B(b
, lo
, 0);
633 // TODO maybe the builders should default to making dst half-precision
634 // if the src's were half precision, to make this less awkward.. otoh
635 // we should probably just do this lowering in NIR.
636 hi
->regs
[0]->flags
|= IR3_REG_HALF
;
637 lo
->regs
[0]->flags
|= IR3_REG_HALF
;
639 dst
[0] = ir3_ADD_S(b
, hi
, 0, lo
, 0);
640 dst
[0]->regs
[0]->flags
|= IR3_REG_HALF
;
641 dst
[0] = ir3_COV(b
, dst
[0], TYPE_U16
, TYPE_U32
);
644 case nir_op_ifind_msb
: {
645 struct ir3_instruction
*cmp
;
646 dst
[0] = ir3_CLZ_S(b
, src
[0], 0);
647 cmp
= ir3_CMPS_S(b
, dst
[0], 0, create_immed(b
, 0), 0);
648 cmp
->cat2
.condition
= IR3_COND_GE
;
649 dst
[0] = ir3_SEL_B32(b
,
650 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
654 case nir_op_ufind_msb
:
655 dst
[0] = ir3_CLZ_B(b
, src
[0], 0);
656 dst
[0] = ir3_SEL_B32(b
,
657 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
658 src
[0], 0, dst
[0], 0);
660 case nir_op_find_lsb
:
661 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
662 dst
[0] = ir3_CLZ_B(b
, dst
[0], 0);
664 case nir_op_bitfield_reverse
:
665 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
669 ir3_context_error(ctx
, "Unhandled ALU op: %s\n",
670 nir_op_infos
[alu
->op
].name
);
674 ir3_put_dst(ctx
, &alu
->dest
.dest
);
677 /* handles direct/indirect UBO reads: */
679 emit_intrinsic_load_ubo(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
680 struct ir3_instruction
**dst
)
682 struct ir3_block
*b
= ctx
->block
;
683 struct ir3_instruction
*base_lo
, *base_hi
, *addr
, *src0
, *src1
;
684 /* UBO addresses are the first driver params, but subtract 2 here to
685 * account for nir_lower_uniforms_to_ubo rebasing the UBOs such that UBO 0
686 * is the uniforms: */
687 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
688 unsigned ubo
= regid(const_state
->offsets
.ubo
, 0) - 2;
689 const unsigned ptrsz
= ir3_pointer_size(ctx
->compiler
);
693 /* First src is ubo index, which could either be an immed or not: */
694 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0];
695 if (is_same_type_mov(src0
) &&
696 (src0
->regs
[1]->flags
& IR3_REG_IMMED
)) {
697 base_lo
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
));
698 base_hi
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
) + 1);
700 base_lo
= create_uniform_indirect(b
, ubo
, ir3_get_addr(ctx
, src0
, ptrsz
));
701 base_hi
= create_uniform_indirect(b
, ubo
+ 1, ir3_get_addr(ctx
, src0
, ptrsz
));
704 /* note: on 32bit gpu's base_hi is ignored and DCE'd */
707 if (nir_src_is_const(intr
->src
[1])) {
708 off
+= nir_src_as_uint(intr
->src
[1]);
710 /* For load_ubo_indirect, second src is indirect offset: */
711 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0];
713 /* and add offset to addr: */
714 addr
= ir3_ADD_S(b
, addr
, 0, src1
, 0);
717 /* if offset is to large to encode in the ldg, split it out: */
718 if ((off
+ (intr
->num_components
* 4)) > 1024) {
719 /* split out the minimal amount to improve the odds that
720 * cp can fit the immediate in the add.s instruction:
722 unsigned off2
= off
+ (intr
->num_components
* 4) - 1024;
723 addr
= ir3_ADD_S(b
, addr
, 0, create_immed(b
, off2
), 0);
728 struct ir3_instruction
*carry
;
730 /* handle 32b rollover, ie:
731 * if (addr < base_lo)
734 carry
= ir3_CMPS_U(b
, addr
, 0, base_lo
, 0);
735 carry
->cat2
.condition
= IR3_COND_LT
;
736 base_hi
= ir3_ADD_S(b
, base_hi
, 0, carry
, 0);
738 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){ addr
, base_hi
}, 2);
741 for (int i
= 0; i
< intr
->num_components
; i
++) {
742 struct ir3_instruction
*load
=
743 ir3_LDG(b
, addr
, 0, create_immed(b
, 1), 0);
744 load
->cat6
.type
= TYPE_U32
;
745 load
->cat6
.src_offset
= off
+ i
* 4; /* byte offset */
750 /* src[] = { block_index } */
752 emit_intrinsic_ssbo_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
753 struct ir3_instruction
**dst
)
755 /* SSBO size stored as a const starting at ssbo_sizes: */
756 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
757 unsigned blk_idx
= nir_src_as_uint(intr
->src
[0]);
758 unsigned idx
= regid(const_state
->offsets
.ssbo_sizes
, 0) +
759 const_state
->ssbo_size
.off
[blk_idx
];
761 debug_assert(const_state
->ssbo_size
.mask
& (1 << blk_idx
));
763 dst
[0] = create_uniform(ctx
->block
, idx
);
766 /* src[] = { offset }. const_index[] = { base } */
768 emit_intrinsic_load_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
769 struct ir3_instruction
**dst
)
771 struct ir3_block
*b
= ctx
->block
;
772 struct ir3_instruction
*ldl
, *offset
;
775 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
776 base
= nir_intrinsic_base(intr
);
778 ldl
= ir3_LDL(b
, offset
, 0, create_immed(b
, intr
->num_components
), 0);
779 ldl
->cat6
.src_offset
= base
;
780 ldl
->cat6
.type
= utype_dst(intr
->dest
);
781 ldl
->regs
[0]->wrmask
= MASK(intr
->num_components
);
783 ldl
->barrier_class
= IR3_BARRIER_SHARED_R
;
784 ldl
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
786 ir3_split_dest(b
, dst
, ldl
, 0, intr
->num_components
);
789 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
791 emit_intrinsic_store_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
793 struct ir3_block
*b
= ctx
->block
;
794 struct ir3_instruction
*stl
, *offset
;
795 struct ir3_instruction
* const *value
;
796 unsigned base
, wrmask
;
798 value
= ir3_get_src(ctx
, &intr
->src
[0]);
799 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
801 base
= nir_intrinsic_base(intr
);
802 wrmask
= nir_intrinsic_write_mask(intr
);
804 /* Combine groups of consecutive enabled channels in one write
805 * message. We use ffs to find the first enabled channel and then ffs on
806 * the bit-inverse, down-shifted writemask to determine the length of
807 * the block of enabled bits.
809 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
812 unsigned first_component
= ffs(wrmask
) - 1;
813 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
815 stl
= ir3_STL(b
, offset
, 0,
816 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
817 create_immed(b
, length
), 0);
818 stl
->cat6
.dst_offset
= first_component
+ base
;
819 stl
->cat6
.type
= utype_src(intr
->src
[0]);
820 stl
->barrier_class
= IR3_BARRIER_SHARED_W
;
821 stl
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
823 array_insert(b
, b
->keeps
, stl
);
825 /* Clear the bits in the writemask that we just wrote, then try
826 * again to see if more channels are left.
828 wrmask
&= (15 << (first_component
+ length
));
833 * CS shared variable atomic intrinsics
835 * All of the shared variable atomic memory operations read a value from
836 * memory, compute a new value using one of the operations below, write the
837 * new value to memory, and return the original value read.
839 * All operations take 2 sources except CompSwap that takes 3. These
842 * 0: The offset into the shared variable storage region that the atomic
843 * operation will operate on.
844 * 1: The data parameter to the atomic function (i.e. the value to add
845 * in shared_atomic_add, etc).
846 * 2: For CompSwap only: the second data parameter.
848 static struct ir3_instruction
*
849 emit_intrinsic_atomic_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
851 struct ir3_block
*b
= ctx
->block
;
852 struct ir3_instruction
*atomic
, *src0
, *src1
;
853 type_t type
= TYPE_U32
;
855 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0]; /* offset */
856 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0]; /* value */
858 switch (intr
->intrinsic
) {
859 case nir_intrinsic_shared_atomic_add
:
860 atomic
= ir3_ATOMIC_ADD(b
, src0
, 0, src1
, 0);
862 case nir_intrinsic_shared_atomic_imin
:
863 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
866 case nir_intrinsic_shared_atomic_umin
:
867 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
869 case nir_intrinsic_shared_atomic_imax
:
870 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
873 case nir_intrinsic_shared_atomic_umax
:
874 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
876 case nir_intrinsic_shared_atomic_and
:
877 atomic
= ir3_ATOMIC_AND(b
, src0
, 0, src1
, 0);
879 case nir_intrinsic_shared_atomic_or
:
880 atomic
= ir3_ATOMIC_OR(b
, src0
, 0, src1
, 0);
882 case nir_intrinsic_shared_atomic_xor
:
883 atomic
= ir3_ATOMIC_XOR(b
, src0
, 0, src1
, 0);
885 case nir_intrinsic_shared_atomic_exchange
:
886 atomic
= ir3_ATOMIC_XCHG(b
, src0
, 0, src1
, 0);
888 case nir_intrinsic_shared_atomic_comp_swap
:
889 /* for cmpxchg, src1 is [ui]vec2(data, compare): */
890 src1
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
891 ir3_get_src(ctx
, &intr
->src
[2])[0],
894 atomic
= ir3_ATOMIC_CMPXCHG(b
, src0
, 0, src1
, 0);
900 atomic
->cat6
.iim_val
= 1;
902 atomic
->cat6
.type
= type
;
903 atomic
->barrier_class
= IR3_BARRIER_SHARED_W
;
904 atomic
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
906 /* even if nothing consume the result, we can't DCE the instruction: */
907 array_insert(b
, b
->keeps
, atomic
);
912 /* TODO handle actual indirect/dynamic case.. which is going to be weird
913 * to handle with the image_mapping table..
915 static struct ir3_instruction
*
916 get_image_samp_tex_src(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
918 unsigned slot
= ir3_get_image_slot(nir_src_as_deref(intr
->src
[0]));
919 unsigned tex_idx
= ir3_image_to_tex(&ctx
->so
->image_mapping
, slot
);
920 struct ir3_instruction
*texture
, *sampler
;
922 texture
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
923 sampler
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
925 return ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
931 /* src[] = { deref, coord, sample_index }. const_index[] = {} */
933 emit_intrinsic_load_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
934 struct ir3_instruction
**dst
)
936 struct ir3_block
*b
= ctx
->block
;
937 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
938 struct ir3_instruction
*samp_tex
= get_image_samp_tex_src(ctx
, intr
);
939 struct ir3_instruction
*sam
;
940 struct ir3_instruction
* const *src0
= ir3_get_src(ctx
, &intr
->src
[1]);
941 struct ir3_instruction
*coords
[4];
942 unsigned flags
, ncoords
= ir3_get_image_coords(var
, &flags
);
943 type_t type
= ir3_get_image_type(var
);
945 /* hmm, this seems a bit odd, but it is what blob does and (at least
946 * a5xx) just faults on bogus addresses otherwise:
948 if (flags
& IR3_INSTR_3D
) {
949 flags
&= ~IR3_INSTR_3D
;
950 flags
|= IR3_INSTR_A
;
953 for (unsigned i
= 0; i
< ncoords
; i
++)
957 coords
[ncoords
++] = create_immed(b
, 0);
959 sam
= ir3_SAM(b
, OPC_ISAM
, type
, 0b1111, flags
,
960 samp_tex
, ir3_create_collect(ctx
, coords
, ncoords
), NULL
);
962 sam
->barrier_class
= IR3_BARRIER_IMAGE_R
;
963 sam
->barrier_conflict
= IR3_BARRIER_IMAGE_W
;
965 ir3_split_dest(b
, dst
, sam
, 0, 4);
969 emit_intrinsic_image_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
970 struct ir3_instruction
**dst
)
972 struct ir3_block
*b
= ctx
->block
;
973 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
974 struct ir3_instruction
*samp_tex
= get_image_samp_tex_src(ctx
, intr
);
975 struct ir3_instruction
*sam
, *lod
;
976 unsigned flags
, ncoords
= ir3_get_image_coords(var
, &flags
);
978 lod
= create_immed(b
, 0);
979 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
980 samp_tex
, lod
, NULL
);
982 /* Array size actually ends up in .w rather than .z. This doesn't
983 * matter for miplevel 0, but for higher mips the value in z is
984 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
985 * returned, which means that we have to add 1 to it for arrays for
988 * Note use a temporary dst and then copy, since the size of the dst
989 * array that is passed in is based on nir's understanding of the
990 * result size, not the hardware's
992 struct ir3_instruction
*tmp
[4];
994 ir3_split_dest(b
, tmp
, sam
, 0, 4);
996 /* get_size instruction returns size in bytes instead of texels
997 * for imageBuffer, so we need to divide it by the pixel size
998 * of the image format.
1000 * TODO: This is at least true on a5xx. Check other gens.
1002 enum glsl_sampler_dim dim
=
1003 glsl_get_sampler_dim(glsl_without_array(var
->type
));
1004 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
1005 /* Since all the possible values the divisor can take are
1006 * power-of-two (4, 8, or 16), the division is implemented
1008 * During shader setup, the log2 of the image format's
1009 * bytes-per-pixel should have been emitted in 2nd slot of
1010 * image_dims. See ir3_shader::emit_image_dims().
1012 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
1013 unsigned cb
= regid(const_state
->offsets
.image_dims
, 0) +
1014 const_state
->image_dims
.off
[var
->data
.driver_location
];
1015 struct ir3_instruction
*aux
= create_uniform(b
, cb
+ 1);
1017 tmp
[0] = ir3_SHR_B(b
, tmp
[0], 0, aux
, 0);
1020 for (unsigned i
= 0; i
< ncoords
; i
++)
1023 if (flags
& IR3_INSTR_A
) {
1024 if (ctx
->compiler
->levels_add_one
) {
1025 dst
[ncoords
-1] = ir3_ADD_U(b
, tmp
[3], 0, create_immed(b
, 1), 0);
1027 dst
[ncoords
-1] = ir3_MOV(b
, tmp
[3], TYPE_U32
);
1033 emit_intrinsic_barrier(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1035 struct ir3_block
*b
= ctx
->block
;
1036 struct ir3_instruction
*barrier
;
1038 switch (intr
->intrinsic
) {
1039 case nir_intrinsic_barrier
:
1040 barrier
= ir3_BAR(b
);
1041 barrier
->cat7
.g
= true;
1042 barrier
->cat7
.l
= true;
1043 barrier
->flags
= IR3_INSTR_SS
| IR3_INSTR_SY
;
1044 barrier
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1046 case nir_intrinsic_memory_barrier
:
1047 barrier
= ir3_FENCE(b
);
1048 barrier
->cat7
.g
= true;
1049 barrier
->cat7
.r
= true;
1050 barrier
->cat7
.w
= true;
1051 barrier
->cat7
.l
= true;
1052 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
|
1053 IR3_BARRIER_BUFFER_W
;
1054 barrier
->barrier_conflict
=
1055 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1056 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1058 case nir_intrinsic_memory_barrier_atomic_counter
:
1059 case nir_intrinsic_memory_barrier_buffer
:
1060 barrier
= ir3_FENCE(b
);
1061 barrier
->cat7
.g
= true;
1062 barrier
->cat7
.r
= true;
1063 barrier
->cat7
.w
= true;
1064 barrier
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1065 barrier
->barrier_conflict
= IR3_BARRIER_BUFFER_R
|
1066 IR3_BARRIER_BUFFER_W
;
1068 case nir_intrinsic_memory_barrier_image
:
1069 // TODO double check if this should have .g set
1070 barrier
= ir3_FENCE(b
);
1071 barrier
->cat7
.g
= true;
1072 barrier
->cat7
.r
= true;
1073 barrier
->cat7
.w
= true;
1074 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1075 barrier
->barrier_conflict
= IR3_BARRIER_IMAGE_R
|
1076 IR3_BARRIER_IMAGE_W
;
1078 case nir_intrinsic_memory_barrier_shared
:
1079 barrier
= ir3_FENCE(b
);
1080 barrier
->cat7
.g
= true;
1081 barrier
->cat7
.l
= true;
1082 barrier
->cat7
.r
= true;
1083 barrier
->cat7
.w
= true;
1084 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
;
1085 barrier
->barrier_conflict
= IR3_BARRIER_SHARED_R
|
1086 IR3_BARRIER_SHARED_W
;
1088 case nir_intrinsic_group_memory_barrier
:
1089 barrier
= ir3_FENCE(b
);
1090 barrier
->cat7
.g
= true;
1091 barrier
->cat7
.l
= true;
1092 barrier
->cat7
.r
= true;
1093 barrier
->cat7
.w
= true;
1094 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
|
1095 IR3_BARRIER_IMAGE_W
|
1096 IR3_BARRIER_BUFFER_W
;
1097 barrier
->barrier_conflict
=
1098 IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
|
1099 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1100 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1106 /* make sure barrier doesn't get DCE'd */
1107 array_insert(b
, b
->keeps
, barrier
);
1110 static void add_sysval_input_compmask(struct ir3_context
*ctx
,
1111 gl_system_value slot
, unsigned compmask
,
1112 struct ir3_instruction
*instr
)
1114 struct ir3_shader_variant
*so
= ctx
->so
;
1115 unsigned r
= regid(so
->inputs_count
, 0);
1116 unsigned n
= so
->inputs_count
++;
1118 so
->inputs
[n
].sysval
= true;
1119 so
->inputs
[n
].slot
= slot
;
1120 so
->inputs
[n
].compmask
= compmask
;
1121 so
->inputs
[n
].regid
= r
;
1122 so
->inputs
[n
].interpolate
= INTERP_MODE_FLAT
;
1125 ctx
->ir
->ninputs
= MAX2(ctx
->ir
->ninputs
, r
+ 1);
1126 ctx
->ir
->inputs
[r
] = instr
;
1129 static void add_sysval_input(struct ir3_context
*ctx
, gl_system_value slot
,
1130 struct ir3_instruction
*instr
)
1132 add_sysval_input_compmask(ctx
, slot
, 0x1, instr
);
1135 static struct ir3_instruction
*
1136 get_barycentric_centroid(struct ir3_context
*ctx
)
1138 if (!ctx
->ij_centroid
) {
1139 struct ir3_instruction
*xy
[2];
1140 struct ir3_instruction
*ij
;
1142 ij
= create_input_compmask(ctx
, 0, 0x3);
1143 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1145 ctx
->ij_centroid
= ir3_create_collect(ctx
, xy
, 2);
1147 add_sysval_input_compmask(ctx
,
1148 SYSTEM_VALUE_BARYCENTRIC_CENTROID
,
1152 return ctx
->ij_centroid
;
1155 static struct ir3_instruction
*
1156 get_barycentric_sample(struct ir3_context
*ctx
)
1158 if (!ctx
->ij_sample
) {
1159 struct ir3_instruction
*xy
[2];
1160 struct ir3_instruction
*ij
;
1162 ij
= create_input_compmask(ctx
, 0, 0x3);
1163 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1165 ctx
->ij_sample
= ir3_create_collect(ctx
, xy
, 2);
1167 add_sysval_input_compmask(ctx
,
1168 SYSTEM_VALUE_BARYCENTRIC_SAMPLE
,
1172 return ctx
->ij_sample
;
1175 static struct ir3_instruction
*
1176 get_barycentric_pixel(struct ir3_context
*ctx
)
1178 /* TODO when tgsi_to_nir supports "new-style" FS inputs switch
1179 * this to create ij_pixel only on demand:
1181 return ctx
->ij_pixel
;
1184 static struct ir3_instruction
*
1185 get_frag_coord(struct ir3_context
*ctx
)
1187 if (!ctx
->frag_coord
) {
1188 struct ir3_block
*b
= ctx
->block
;
1189 struct ir3_instruction
*xyzw
[4];
1190 struct ir3_instruction
*hw_frag_coord
;
1192 hw_frag_coord
= create_input_compmask(ctx
, 0, 0xf);
1193 ir3_split_dest(ctx
->block
, xyzw
, hw_frag_coord
, 0, 4);
1195 /* for frag_coord.xy, we get unsigned values.. we need
1196 * to subtract (integer) 8 and divide by 16 (right-
1197 * shift by 4) then convert to float:
1201 * mov.u32f32 dst, tmp
1204 for (int i
= 0; i
< 2; i
++) {
1205 xyzw
[i
] = ir3_SUB_S(b
, xyzw
[i
], 0,
1206 create_immed(b
, 8), 0);
1207 xyzw
[i
] = ir3_SHR_B(b
, xyzw
[i
], 0,
1208 create_immed(b
, 4), 0);
1209 xyzw
[i
] = ir3_COV(b
, xyzw
[i
], TYPE_U32
, TYPE_F32
);
1212 ctx
->frag_coord
= ir3_create_collect(ctx
, xyzw
, 4);
1214 add_sysval_input_compmask(ctx
,
1215 SYSTEM_VALUE_FRAG_COORD
,
1216 0xf, hw_frag_coord
);
1218 ctx
->so
->frag_coord
= true;
1221 return ctx
->frag_coord
;
1225 emit_intrinsic(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1227 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1228 struct ir3_instruction
**dst
;
1229 struct ir3_instruction
* const *src
;
1230 struct ir3_block
*b
= ctx
->block
;
1233 if (info
->has_dest
) {
1234 unsigned n
= nir_intrinsic_dest_components(intr
);
1235 dst
= ir3_get_dst(ctx
, &intr
->dest
, n
);
1240 switch (intr
->intrinsic
) {
1241 case nir_intrinsic_load_uniform
:
1242 idx
= nir_intrinsic_base(intr
);
1243 if (nir_src_is_const(intr
->src
[0])) {
1244 idx
+= nir_src_as_uint(intr
->src
[0]);
1245 for (int i
= 0; i
< intr
->num_components
; i
++) {
1246 dst
[i
] = create_uniform(b
, idx
+ i
);
1249 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1250 for (int i
= 0; i
< intr
->num_components
; i
++) {
1251 dst
[i
] = create_uniform_indirect(b
, idx
+ i
,
1252 ir3_get_addr(ctx
, src
[0], 1));
1254 /* NOTE: if relative addressing is used, we set
1255 * constlen in the compiler (to worst-case value)
1256 * since we don't know in the assembler what the max
1257 * addr reg value can be:
1259 ctx
->so
->constlen
= ctx
->s
->num_uniforms
;
1262 case nir_intrinsic_load_ubo
:
1263 emit_intrinsic_load_ubo(ctx
, intr
, dst
);
1265 case nir_intrinsic_load_frag_coord
:
1266 ir3_split_dest(b
, dst
, get_frag_coord(ctx
), 0, 4);
1268 case nir_intrinsic_load_sample_pos_from_id
: {
1269 /* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
1270 * but that doesn't seem necessary.
1272 struct ir3_instruction
*offset
=
1273 ir3_RGETPOS(b
, ir3_get_src(ctx
, &intr
->src
[0])[0], 0);
1274 offset
->regs
[0]->wrmask
= 0x3;
1275 offset
->cat5
.type
= TYPE_F32
;
1277 ir3_split_dest(b
, dst
, offset
, 0, 2);
1281 case nir_intrinsic_load_size_ir3
:
1282 if (!ctx
->ij_size
) {
1283 ctx
->ij_size
= create_input(ctx
, 0);
1285 add_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_SIZE
,
1288 dst
[0] = ctx
->ij_size
;
1290 case nir_intrinsic_load_barycentric_centroid
:
1291 ir3_split_dest(b
, dst
, get_barycentric_centroid(ctx
), 0, 2);
1293 case nir_intrinsic_load_barycentric_sample
:
1294 if (ctx
->so
->key
.msaa
) {
1295 ir3_split_dest(b
, dst
, get_barycentric_sample(ctx
), 0, 2);
1297 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1300 case nir_intrinsic_load_barycentric_pixel
:
1301 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1303 case nir_intrinsic_load_interpolated_input
:
1304 idx
= nir_intrinsic_base(intr
);
1305 comp
= nir_intrinsic_component(intr
);
1306 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1307 if (nir_src_is_const(intr
->src
[1])) {
1308 struct ir3_instruction
*coord
= ir3_create_collect(ctx
, src
, 2);
1309 idx
+= nir_src_as_uint(intr
->src
[1]);
1310 for (int i
= 0; i
< intr
->num_components
; i
++) {
1311 unsigned inloc
= idx
* 4 + i
+ comp
;
1312 if (ctx
->so
->inputs
[idx
].bary
&&
1313 !ctx
->so
->inputs
[idx
].use_ldlv
) {
1314 dst
[i
] = ir3_BARY_F(b
, create_immed(b
, inloc
), 0, coord
, 0);
1316 /* for non-varyings use the pre-setup input, since
1317 * that is easier than mapping things back to a
1318 * nir_variable to figure out what it is.
1320 dst
[i
] = ctx
->ir
->inputs
[inloc
];
1324 ir3_context_error(ctx
, "unhandled");
1327 case nir_intrinsic_load_input
:
1328 idx
= nir_intrinsic_base(intr
);
1329 comp
= nir_intrinsic_component(intr
);
1330 if (nir_src_is_const(intr
->src
[0])) {
1331 idx
+= nir_src_as_uint(intr
->src
[0]);
1332 for (int i
= 0; i
< intr
->num_components
; i
++) {
1333 unsigned n
= idx
* 4 + i
+ comp
;
1334 dst
[i
] = ctx
->ir
->inputs
[n
];
1335 compile_assert(ctx
, ctx
->ir
->inputs
[n
]);
1338 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1339 struct ir3_instruction
*collect
=
1340 ir3_create_collect(ctx
, ctx
->ir
->inputs
, ctx
->ir
->ninputs
);
1341 struct ir3_instruction
*addr
= ir3_get_addr(ctx
, src
[0], 4);
1342 for (int i
= 0; i
< intr
->num_components
; i
++) {
1343 unsigned n
= idx
* 4 + i
+ comp
;
1344 dst
[i
] = create_indirect_load(ctx
, ctx
->ir
->ninputs
,
1349 /* All SSBO intrinsics should have been lowered by 'lower_io_offsets'
1350 * pass and replaced by an ir3-specifc version that adds the
1351 * dword-offset in the last source.
1353 case nir_intrinsic_load_ssbo_ir3
:
1354 ctx
->funcs
->emit_intrinsic_load_ssbo(ctx
, intr
, dst
);
1356 case nir_intrinsic_store_ssbo_ir3
:
1357 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1358 !ctx
->s
->info
.fs
.early_fragment_tests
)
1359 ctx
->so
->no_earlyz
= true;
1360 ctx
->funcs
->emit_intrinsic_store_ssbo(ctx
, intr
);
1362 case nir_intrinsic_get_buffer_size
:
1363 emit_intrinsic_ssbo_size(ctx
, intr
, dst
);
1365 case nir_intrinsic_ssbo_atomic_add_ir3
:
1366 case nir_intrinsic_ssbo_atomic_imin_ir3
:
1367 case nir_intrinsic_ssbo_atomic_umin_ir3
:
1368 case nir_intrinsic_ssbo_atomic_imax_ir3
:
1369 case nir_intrinsic_ssbo_atomic_umax_ir3
:
1370 case nir_intrinsic_ssbo_atomic_and_ir3
:
1371 case nir_intrinsic_ssbo_atomic_or_ir3
:
1372 case nir_intrinsic_ssbo_atomic_xor_ir3
:
1373 case nir_intrinsic_ssbo_atomic_exchange_ir3
:
1374 case nir_intrinsic_ssbo_atomic_comp_swap_ir3
:
1375 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1376 !ctx
->s
->info
.fs
.early_fragment_tests
)
1377 ctx
->so
->no_earlyz
= true;
1378 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_ssbo(ctx
, intr
);
1380 case nir_intrinsic_load_shared
:
1381 emit_intrinsic_load_shared(ctx
, intr
, dst
);
1383 case nir_intrinsic_store_shared
:
1384 emit_intrinsic_store_shared(ctx
, intr
);
1386 case nir_intrinsic_shared_atomic_add
:
1387 case nir_intrinsic_shared_atomic_imin
:
1388 case nir_intrinsic_shared_atomic_umin
:
1389 case nir_intrinsic_shared_atomic_imax
:
1390 case nir_intrinsic_shared_atomic_umax
:
1391 case nir_intrinsic_shared_atomic_and
:
1392 case nir_intrinsic_shared_atomic_or
:
1393 case nir_intrinsic_shared_atomic_xor
:
1394 case nir_intrinsic_shared_atomic_exchange
:
1395 case nir_intrinsic_shared_atomic_comp_swap
:
1396 dst
[0] = emit_intrinsic_atomic_shared(ctx
, intr
);
1398 case nir_intrinsic_image_deref_load
:
1399 emit_intrinsic_load_image(ctx
, intr
, dst
);
1401 case nir_intrinsic_image_deref_store
:
1402 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1403 !ctx
->s
->info
.fs
.early_fragment_tests
)
1404 ctx
->so
->no_earlyz
= true;
1405 ctx
->funcs
->emit_intrinsic_store_image(ctx
, intr
);
1407 case nir_intrinsic_image_deref_size
:
1408 emit_intrinsic_image_size(ctx
, intr
, dst
);
1410 case nir_intrinsic_image_deref_atomic_add
:
1411 case nir_intrinsic_image_deref_atomic_min
:
1412 case nir_intrinsic_image_deref_atomic_max
:
1413 case nir_intrinsic_image_deref_atomic_and
:
1414 case nir_intrinsic_image_deref_atomic_or
:
1415 case nir_intrinsic_image_deref_atomic_xor
:
1416 case nir_intrinsic_image_deref_atomic_exchange
:
1417 case nir_intrinsic_image_deref_atomic_comp_swap
:
1418 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1419 !ctx
->s
->info
.fs
.early_fragment_tests
)
1420 ctx
->so
->no_earlyz
= true;
1421 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_image(ctx
, intr
);
1423 case nir_intrinsic_barrier
:
1424 case nir_intrinsic_memory_barrier
:
1425 case nir_intrinsic_group_memory_barrier
:
1426 case nir_intrinsic_memory_barrier_atomic_counter
:
1427 case nir_intrinsic_memory_barrier_buffer
:
1428 case nir_intrinsic_memory_barrier_image
:
1429 case nir_intrinsic_memory_barrier_shared
:
1430 emit_intrinsic_barrier(ctx
, intr
);
1431 /* note that blk ptr no longer valid, make that obvious: */
1434 case nir_intrinsic_store_output
:
1435 idx
= nir_intrinsic_base(intr
);
1436 comp
= nir_intrinsic_component(intr
);
1437 compile_assert(ctx
, nir_src_is_const(intr
->src
[1]));
1438 idx
+= nir_src_as_uint(intr
->src
[1]);
1440 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1441 for (int i
= 0; i
< intr
->num_components
; i
++) {
1442 unsigned n
= idx
* 4 + i
+ comp
;
1443 ctx
->ir
->outputs
[n
] = src
[i
];
1446 case nir_intrinsic_load_base_vertex
:
1447 case nir_intrinsic_load_first_vertex
:
1448 if (!ctx
->basevertex
) {
1449 ctx
->basevertex
= create_driver_param(ctx
, IR3_DP_VTXID_BASE
);
1450 add_sysval_input(ctx
, SYSTEM_VALUE_FIRST_VERTEX
, ctx
->basevertex
);
1452 dst
[0] = ctx
->basevertex
;
1454 case nir_intrinsic_load_vertex_id_zero_base
:
1455 case nir_intrinsic_load_vertex_id
:
1456 if (!ctx
->vertex_id
) {
1457 gl_system_value sv
= (intr
->intrinsic
== nir_intrinsic_load_vertex_id
) ?
1458 SYSTEM_VALUE_VERTEX_ID
: SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1459 ctx
->vertex_id
= create_input(ctx
, 0);
1460 add_sysval_input(ctx
, sv
, ctx
->vertex_id
);
1462 dst
[0] = ctx
->vertex_id
;
1464 case nir_intrinsic_load_instance_id
:
1465 if (!ctx
->instance_id
) {
1466 ctx
->instance_id
= create_input(ctx
, 0);
1467 add_sysval_input(ctx
, SYSTEM_VALUE_INSTANCE_ID
,
1470 dst
[0] = ctx
->instance_id
;
1472 case nir_intrinsic_load_sample_id
:
1473 ctx
->so
->per_samp
= true;
1475 case nir_intrinsic_load_sample_id_no_per_sample
:
1476 if (!ctx
->samp_id
) {
1477 ctx
->samp_id
= create_input(ctx
, 0);
1478 ctx
->samp_id
->regs
[0]->flags
|= IR3_REG_HALF
;
1479 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_ID
,
1482 dst
[0] = ir3_COV(b
, ctx
->samp_id
, TYPE_U16
, TYPE_U32
);
1484 case nir_intrinsic_load_sample_mask_in
:
1485 if (!ctx
->samp_mask_in
) {
1486 ctx
->samp_mask_in
= create_input(ctx
, 0);
1487 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_MASK_IN
,
1490 dst
[0] = ctx
->samp_mask_in
;
1492 case nir_intrinsic_load_user_clip_plane
:
1493 idx
= nir_intrinsic_ucp_id(intr
);
1494 for (int i
= 0; i
< intr
->num_components
; i
++) {
1495 unsigned n
= idx
* 4 + i
;
1496 dst
[i
] = create_driver_param(ctx
, IR3_DP_UCP0_X
+ n
);
1499 case nir_intrinsic_load_front_face
:
1500 if (!ctx
->frag_face
) {
1501 ctx
->so
->frag_face
= true;
1502 ctx
->frag_face
= create_input(ctx
, 0);
1503 add_sysval_input(ctx
, SYSTEM_VALUE_FRONT_FACE
, ctx
->frag_face
);
1504 ctx
->frag_face
->regs
[0]->flags
|= IR3_REG_HALF
;
1506 /* for fragface, we get -1 for back and 0 for front. However this is
1507 * the inverse of what nir expects (where ~0 is true).
1509 dst
[0] = ir3_COV(b
, ctx
->frag_face
, TYPE_S16
, TYPE_S32
);
1510 dst
[0] = ir3_NOT_B(b
, dst
[0], 0);
1512 case nir_intrinsic_load_local_invocation_id
:
1513 if (!ctx
->local_invocation_id
) {
1514 ctx
->local_invocation_id
= create_input_compmask(ctx
, 0, 0x7);
1515 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_LOCAL_INVOCATION_ID
,
1516 0x7, ctx
->local_invocation_id
);
1518 ir3_split_dest(b
, dst
, ctx
->local_invocation_id
, 0, 3);
1520 case nir_intrinsic_load_work_group_id
:
1521 if (!ctx
->work_group_id
) {
1522 ctx
->work_group_id
= create_input_compmask(ctx
, 0, 0x7);
1523 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_WORK_GROUP_ID
,
1524 0x7, ctx
->work_group_id
);
1525 ctx
->work_group_id
->regs
[0]->flags
|= IR3_REG_HIGH
;
1527 ir3_split_dest(b
, dst
, ctx
->work_group_id
, 0, 3);
1529 case nir_intrinsic_load_num_work_groups
:
1530 for (int i
= 0; i
< intr
->num_components
; i
++) {
1531 dst
[i
] = create_driver_param(ctx
, IR3_DP_NUM_WORK_GROUPS_X
+ i
);
1534 case nir_intrinsic_load_local_group_size
:
1535 for (int i
= 0; i
< intr
->num_components
; i
++) {
1536 dst
[i
] = create_driver_param(ctx
, IR3_DP_LOCAL_GROUP_SIZE_X
+ i
);
1539 case nir_intrinsic_discard_if
:
1540 case nir_intrinsic_discard
: {
1541 struct ir3_instruction
*cond
, *kill
;
1543 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
1544 /* conditional discard: */
1545 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1546 cond
= ir3_b2n(b
, src
[0]);
1548 /* unconditional discard: */
1549 cond
= create_immed(b
, 1);
1552 /* NOTE: only cmps.*.* can write p0.x: */
1553 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1554 cond
->cat2
.condition
= IR3_COND_NE
;
1556 /* condition always goes in predicate register: */
1557 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1559 kill
= ir3_KILL(b
, cond
, 0);
1560 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1562 array_insert(b
, b
->keeps
, kill
);
1563 ctx
->so
->no_earlyz
= true;
1568 ir3_context_error(ctx
, "Unhandled intrinsic type: %s\n",
1569 nir_intrinsic_infos
[intr
->intrinsic
].name
);
1574 ir3_put_dst(ctx
, &intr
->dest
);
1578 emit_load_const(struct ir3_context
*ctx
, nir_load_const_instr
*instr
)
1580 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &instr
->def
,
1581 instr
->def
.num_components
);
1582 type_t type
= (instr
->def
.bit_size
< 32) ? TYPE_U16
: TYPE_U32
;
1584 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1585 dst
[i
] = create_immed_typed(ctx
->block
, instr
->value
[i
].u32
, type
);
1589 emit_undef(struct ir3_context
*ctx
, nir_ssa_undef_instr
*undef
)
1591 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &undef
->def
,
1592 undef
->def
.num_components
);
1593 type_t type
= (undef
->def
.bit_size
< 32) ? TYPE_U16
: TYPE_U32
;
1595 /* backend doesn't want undefined instructions, so just plug
1598 for (int i
= 0; i
< undef
->def
.num_components
; i
++)
1599 dst
[i
] = create_immed_typed(ctx
->block
, fui(0.0), type
);
1603 * texture fetch/sample instructions:
1607 tex_info(nir_tex_instr
*tex
, unsigned *flagsp
, unsigned *coordsp
)
1609 unsigned coords
, flags
= 0;
1611 /* note: would use tex->coord_components.. except txs.. also,
1612 * since array index goes after shadow ref, we don't want to
1615 switch (tex
->sampler_dim
) {
1616 case GLSL_SAMPLER_DIM_1D
:
1617 case GLSL_SAMPLER_DIM_BUF
:
1620 case GLSL_SAMPLER_DIM_2D
:
1621 case GLSL_SAMPLER_DIM_RECT
:
1622 case GLSL_SAMPLER_DIM_EXTERNAL
:
1623 case GLSL_SAMPLER_DIM_MS
:
1626 case GLSL_SAMPLER_DIM_3D
:
1627 case GLSL_SAMPLER_DIM_CUBE
:
1629 flags
|= IR3_INSTR_3D
;
1632 unreachable("bad sampler_dim");
1635 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
1636 flags
|= IR3_INSTR_S
;
1638 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
)
1639 flags
|= IR3_INSTR_A
;
1645 /* Gets the sampler/texture idx as a hvec2. Which could either be dynamic
1646 * or immediate (in which case it will get lowered later to a non .s2en
1647 * version of the tex instruction which encode tex/samp as immediates:
1649 static struct ir3_instruction
*
1650 get_tex_samp_tex_src(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1652 int texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
);
1653 int sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
);
1654 struct ir3_instruction
*texture
, *sampler
;
1656 if (texture_idx
>= 0) {
1657 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
1658 texture
= ir3_COV(ctx
->block
, texture
, TYPE_U32
, TYPE_U16
);
1660 /* TODO what to do for dynamic case? I guess we only need the
1661 * max index for astc srgb workaround so maybe not a problem
1662 * to worry about if we don't enable indirect samplers for
1665 ctx
->max_texture_index
= MAX2(ctx
->max_texture_index
, tex
->texture_index
);
1666 texture
= create_immed_typed(ctx
->block
, tex
->texture_index
, TYPE_U16
);
1669 if (sampler_idx
>= 0) {
1670 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
1671 sampler
= ir3_COV(ctx
->block
, sampler
, TYPE_U32
, TYPE_U16
);
1673 sampler
= create_immed_typed(ctx
->block
, tex
->sampler_index
, TYPE_U16
);
1676 return ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1683 emit_tex(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1685 struct ir3_block
*b
= ctx
->block
;
1686 struct ir3_instruction
**dst
, *sam
, *src0
[12], *src1
[4];
1687 struct ir3_instruction
* const *coord
, * const *off
, * const *ddx
, * const *ddy
;
1688 struct ir3_instruction
*lod
, *compare
, *proj
, *sample_index
;
1689 bool has_bias
= false, has_lod
= false, has_proj
= false, has_off
= false;
1690 unsigned i
, coords
, flags
, ncomp
;
1691 unsigned nsrc0
= 0, nsrc1
= 0;
1695 ncomp
= nir_dest_num_components(tex
->dest
);
1697 coord
= off
= ddx
= ddy
= NULL
;
1698 lod
= proj
= compare
= sample_index
= NULL
;
1700 dst
= ir3_get_dst(ctx
, &tex
->dest
, ncomp
);
1702 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1703 switch (tex
->src
[i
].src_type
) {
1704 case nir_tex_src_coord
:
1705 coord
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1707 case nir_tex_src_bias
:
1708 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1711 case nir_tex_src_lod
:
1712 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1715 case nir_tex_src_comparator
: /* shadow comparator */
1716 compare
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1718 case nir_tex_src_projector
:
1719 proj
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1722 case nir_tex_src_offset
:
1723 off
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1726 case nir_tex_src_ddx
:
1727 ddx
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1729 case nir_tex_src_ddy
:
1730 ddy
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1732 case nir_tex_src_ms_index
:
1733 sample_index
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1735 case nir_tex_src_texture_offset
:
1736 case nir_tex_src_sampler_offset
:
1737 /* handled in get_tex_samp_src() */
1740 ir3_context_error(ctx
, "Unhandled NIR tex src type: %d\n",
1741 tex
->src
[i
].src_type
);
1747 case nir_texop_tex
: opc
= has_lod
? OPC_SAML
: OPC_SAM
; break;
1748 case nir_texop_txb
: opc
= OPC_SAMB
; break;
1749 case nir_texop_txl
: opc
= OPC_SAML
; break;
1750 case nir_texop_txd
: opc
= OPC_SAMGQ
; break;
1751 case nir_texop_txf
: opc
= OPC_ISAML
; break;
1752 case nir_texop_lod
: opc
= OPC_GETLOD
; break;
1754 /* NOTE: a4xx might need to emulate gather w/ txf (this is
1755 * what blob does, seems gather is broken?), and a3xx did
1756 * not support it (but probably could also emulate).
1758 switch (tex
->component
) {
1759 case 0: opc
= OPC_GATHER4R
; break;
1760 case 1: opc
= OPC_GATHER4G
; break;
1761 case 2: opc
= OPC_GATHER4B
; break;
1762 case 3: opc
= OPC_GATHER4A
; break;
1765 case nir_texop_txf_ms_fb
:
1766 case nir_texop_txf_ms
: opc
= OPC_ISAMM
; break;
1768 ir3_context_error(ctx
, "Unhandled NIR tex type: %d\n", tex
->op
);
1772 tex_info(tex
, &flags
, &coords
);
1775 * lay out the first argument in the proper order:
1776 * - actual coordinates first
1777 * - shadow reference
1780 * - starting at offset 4, dpdx.xy, dpdy.xy
1782 * bias/lod go into the second arg
1785 /* insert tex coords: */
1786 for (i
= 0; i
< coords
; i
++)
1791 /* scale up integer coords for TXF based on the LOD */
1792 if (ctx
->compiler
->unminify_coords
&& (opc
== OPC_ISAML
)) {
1794 for (i
= 0; i
< coords
; i
++)
1795 src0
[i
] = ir3_SHL_B(b
, src0
[i
], 0, lod
, 0);
1799 /* hw doesn't do 1d, so we treat it as 2d with
1800 * height of 1, and patch up the y coord.
1803 src0
[nsrc0
++] = create_immed(b
, 0);
1805 src0
[nsrc0
++] = create_immed(b
, fui(0.5));
1809 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
1810 src0
[nsrc0
++] = compare
;
1812 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
) {
1813 struct ir3_instruction
*idx
= coord
[coords
];
1815 /* the array coord for cube arrays needs 0.5 added to it */
1816 if (ctx
->compiler
->array_index_add_half
&& !is_isam(opc
))
1817 idx
= ir3_ADD_F(b
, idx
, 0, create_immed(b
, fui(0.5)), 0);
1819 src0
[nsrc0
++] = idx
;
1823 src0
[nsrc0
++] = proj
;
1824 flags
|= IR3_INSTR_P
;
1827 /* pad to 4, then ddx/ddy: */
1828 if (tex
->op
== nir_texop_txd
) {
1830 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1831 for (i
= 0; i
< coords
; i
++)
1832 src0
[nsrc0
++] = ddx
[i
];
1834 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1835 for (i
= 0; i
< coords
; i
++)
1836 src0
[nsrc0
++] = ddy
[i
];
1838 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1841 /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
1842 * with scaled x coord according to requested sample:
1844 if (opc
== OPC_ISAMM
) {
1845 if (ctx
->compiler
->txf_ms_with_isaml
) {
1846 /* the samples are laid out in x dimension as
1848 * x_ms = (x << ms) + sample_index;
1850 struct ir3_instruction
*ms
;
1851 ms
= create_immed(b
, (ctx
->samples
>> (2 * tex
->texture_index
)) & 3);
1853 src0
[0] = ir3_SHL_B(b
, src0
[0], 0, ms
, 0);
1854 src0
[0] = ir3_ADD_U(b
, src0
[0], 0, sample_index
, 0);
1858 src0
[nsrc0
++] = sample_index
;
1863 * second argument (if applicable):
1868 if (has_off
| has_lod
| has_bias
) {
1870 unsigned off_coords
= coords
;
1871 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
1873 for (i
= 0; i
< off_coords
; i
++)
1874 src1
[nsrc1
++] = off
[i
];
1876 src1
[nsrc1
++] = create_immed(b
, fui(0.0));
1877 flags
|= IR3_INSTR_O
;
1880 if (has_lod
| has_bias
)
1881 src1
[nsrc1
++] = lod
;
1884 switch (tex
->dest_type
) {
1885 case nir_type_invalid
:
1886 case nir_type_float
:
1897 unreachable("bad dest_type");
1900 if (opc
== OPC_GETLOD
)
1903 struct ir3_instruction
*samp_tex
;
1905 if (tex
->op
== nir_texop_txf_ms_fb
) {
1906 /* only expect a single txf_ms_fb per shader: */
1907 compile_assert(ctx
, !ctx
->so
->fb_read
);
1908 compile_assert(ctx
, ctx
->so
->type
== MESA_SHADER_FRAGMENT
);
1910 ctx
->so
->fb_read
= true;
1911 samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1912 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
1913 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
1916 ctx
->so
->num_samp
++;
1918 samp_tex
= get_tex_samp_tex_src(ctx
, tex
);
1921 struct ir3_instruction
*col0
= ir3_create_collect(ctx
, src0
, nsrc0
);
1922 struct ir3_instruction
*col1
= ir3_create_collect(ctx
, src1
, nsrc1
);
1924 sam
= ir3_SAM(b
, opc
, type
, MASK(ncomp
), flags
,
1925 samp_tex
, col0
, col1
);
1927 if ((ctx
->astc_srgb
& (1 << tex
->texture_index
)) && !nir_tex_instr_is_query(tex
)) {
1928 /* only need first 3 components: */
1929 sam
->regs
[0]->wrmask
= 0x7;
1930 ir3_split_dest(b
, dst
, sam
, 0, 3);
1932 /* we need to sample the alpha separately with a non-ASTC
1935 sam
= ir3_SAM(b
, opc
, type
, 0b1000, flags
,
1936 samp_tex
, col0
, col1
);
1938 array_insert(ctx
->ir
, ctx
->ir
->astc_srgb
, sam
);
1940 /* fixup .w component: */
1941 ir3_split_dest(b
, &dst
[3], sam
, 3, 1);
1943 /* normal (non-workaround) case: */
1944 ir3_split_dest(b
, dst
, sam
, 0, ncomp
);
1947 /* GETLOD returns results in 4.8 fixed point */
1948 if (opc
== OPC_GETLOD
) {
1949 struct ir3_instruction
*factor
= create_immed(b
, fui(1.0 / 256));
1951 compile_assert(ctx
, tex
->dest_type
== nir_type_float
);
1952 for (i
= 0; i
< 2; i
++) {
1953 dst
[i
] = ir3_MUL_F(b
, ir3_COV(b
, dst
[i
], TYPE_U32
, TYPE_F32
), 0,
1958 ir3_put_dst(ctx
, &tex
->dest
);
1962 emit_tex_query_levels(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1964 struct ir3_block
*b
= ctx
->block
;
1965 struct ir3_instruction
**dst
, *sam
;
1967 dst
= ir3_get_dst(ctx
, &tex
->dest
, 1);
1969 sam
= ir3_SAM(b
, OPC_GETINFO
, TYPE_U32
, 0b0100, 0,
1970 get_tex_samp_tex_src(ctx
, tex
), NULL
, NULL
);
1972 /* even though there is only one component, since it ends
1973 * up in .z rather than .x, we need a split_dest()
1975 ir3_split_dest(b
, dst
, sam
, 0, 3);
1977 /* The # of levels comes from getinfo.z. We need to add 1 to it, since
1978 * the value in TEX_CONST_0 is zero-based.
1980 if (ctx
->compiler
->levels_add_one
)
1981 dst
[0] = ir3_ADD_U(b
, dst
[0], 0, create_immed(b
, 1), 0);
1983 ir3_put_dst(ctx
, &tex
->dest
);
1987 emit_tex_txs(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1989 struct ir3_block
*b
= ctx
->block
;
1990 struct ir3_instruction
**dst
, *sam
;
1991 struct ir3_instruction
*lod
;
1992 unsigned flags
, coords
;
1994 tex_info(tex
, &flags
, &coords
);
1996 /* Actually we want the number of dimensions, not coordinates. This
1997 * distinction only matters for cubes.
1999 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2002 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
2004 compile_assert(ctx
, tex
->num_srcs
== 1);
2005 compile_assert(ctx
, tex
->src
[0].src_type
== nir_tex_src_lod
);
2007 lod
= ir3_get_src(ctx
, &tex
->src
[0].src
)[0];
2009 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
2010 get_tex_samp_tex_src(ctx
, tex
), lod
, NULL
);
2012 ir3_split_dest(b
, dst
, sam
, 0, 4);
2014 /* Array size actually ends up in .w rather than .z. This doesn't
2015 * matter for miplevel 0, but for higher mips the value in z is
2016 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2017 * returned, which means that we have to add 1 to it for arrays.
2019 if (tex
->is_array
) {
2020 if (ctx
->compiler
->levels_add_one
) {
2021 dst
[coords
] = ir3_ADD_U(b
, dst
[3], 0, create_immed(b
, 1), 0);
2023 dst
[coords
] = ir3_MOV(b
, dst
[3], TYPE_U32
);
2027 ir3_put_dst(ctx
, &tex
->dest
);
2031 emit_jump(struct ir3_context
*ctx
, nir_jump_instr
*jump
)
2033 switch (jump
->type
) {
2034 case nir_jump_break
:
2035 case nir_jump_continue
:
2036 case nir_jump_return
:
2037 /* I *think* we can simply just ignore this, and use the
2038 * successor block link to figure out where we need to
2039 * jump to for break/continue
2043 ir3_context_error(ctx
, "Unhandled NIR jump type: %d\n", jump
->type
);
2049 emit_instr(struct ir3_context
*ctx
, nir_instr
*instr
)
2051 switch (instr
->type
) {
2052 case nir_instr_type_alu
:
2053 emit_alu(ctx
, nir_instr_as_alu(instr
));
2055 case nir_instr_type_deref
:
2056 /* ignored, handled as part of the intrinsic they are src to */
2058 case nir_instr_type_intrinsic
:
2059 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2061 case nir_instr_type_load_const
:
2062 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2064 case nir_instr_type_ssa_undef
:
2065 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
2067 case nir_instr_type_tex
: {
2068 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
2069 /* couple tex instructions get special-cased:
2073 emit_tex_txs(ctx
, tex
);
2075 case nir_texop_query_levels
:
2076 emit_tex_query_levels(ctx
, tex
);
2084 case nir_instr_type_jump
:
2085 emit_jump(ctx
, nir_instr_as_jump(instr
));
2087 case nir_instr_type_phi
:
2088 /* we have converted phi webs to regs in NIR by now */
2089 ir3_context_error(ctx
, "Unexpected NIR instruction type: %d\n", instr
->type
);
2091 case nir_instr_type_call
:
2092 case nir_instr_type_parallel_copy
:
2093 ir3_context_error(ctx
, "Unhandled NIR instruction type: %d\n", instr
->type
);
2098 static struct ir3_block
*
2099 get_block(struct ir3_context
*ctx
, const nir_block
*nblock
)
2101 struct ir3_block
*block
;
2102 struct hash_entry
*hentry
;
2105 hentry
= _mesa_hash_table_search(ctx
->block_ht
, nblock
);
2107 return hentry
->data
;
2109 block
= ir3_block_create(ctx
->ir
);
2110 block
->nblock
= nblock
;
2111 _mesa_hash_table_insert(ctx
->block_ht
, nblock
, block
);
2113 block
->predecessors_count
= nblock
->predecessors
->entries
;
2114 block
->predecessors
= ralloc_array_size(block
,
2115 sizeof(block
->predecessors
[0]), block
->predecessors_count
);
2117 set_foreach(nblock
->predecessors
, sentry
) {
2118 block
->predecessors
[i
++] = get_block(ctx
, sentry
->key
);
2125 emit_block(struct ir3_context
*ctx
, nir_block
*nblock
)
2127 struct ir3_block
*block
= get_block(ctx
, nblock
);
2129 for (int i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
2130 if (nblock
->successors
[i
]) {
2131 block
->successors
[i
] =
2132 get_block(ctx
, nblock
->successors
[i
]);
2137 list_addtail(&block
->node
, &ctx
->ir
->block_list
);
2139 /* re-emit addr register in each block if needed: */
2140 for (int i
= 0; i
< ARRAY_SIZE(ctx
->addr_ht
); i
++) {
2141 _mesa_hash_table_destroy(ctx
->addr_ht
[i
], NULL
);
2142 ctx
->addr_ht
[i
] = NULL
;
2145 nir_foreach_instr(instr
, nblock
) {
2146 ctx
->cur_instr
= instr
;
2147 emit_instr(ctx
, instr
);
2148 ctx
->cur_instr
= NULL
;
2154 static void emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
);
2157 emit_if(struct ir3_context
*ctx
, nir_if
*nif
)
2159 struct ir3_instruction
*condition
= ir3_get_src(ctx
, &nif
->condition
)[0];
2161 ctx
->block
->condition
=
2162 ir3_get_predicate(ctx
, ir3_b2n(condition
->block
, condition
));
2164 emit_cf_list(ctx
, &nif
->then_list
);
2165 emit_cf_list(ctx
, &nif
->else_list
);
2169 emit_loop(struct ir3_context
*ctx
, nir_loop
*nloop
)
2171 emit_cf_list(ctx
, &nloop
->body
);
2176 stack_push(struct ir3_context
*ctx
)
2179 ctx
->max_stack
= MAX2(ctx
->max_stack
, ctx
->stack
);
2183 stack_pop(struct ir3_context
*ctx
)
2185 compile_assert(ctx
, ctx
->stack
> 0);
2190 emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
)
2192 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2193 switch (node
->type
) {
2194 case nir_cf_node_block
:
2195 emit_block(ctx
, nir_cf_node_as_block(node
));
2197 case nir_cf_node_if
:
2199 emit_if(ctx
, nir_cf_node_as_if(node
));
2202 case nir_cf_node_loop
:
2204 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2207 case nir_cf_node_function
:
2208 ir3_context_error(ctx
, "TODO\n");
2214 /* emit stream-out code. At this point, the current block is the original
2215 * (nir) end block, and nir ensures that all flow control paths terminate
2216 * into the end block. We re-purpose the original end block to generate
2217 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2218 * block holding stream-out write instructions, followed by the new end
2222 * p0.x = (vtxcnt < maxvtxcnt)
2223 * // succs: blockStreamOut, blockNewEnd
2226 * ... stream-out instructions ...
2227 * // succs: blockNewEnd
2233 emit_stream_out(struct ir3_context
*ctx
)
2235 struct ir3
*ir
= ctx
->ir
;
2236 struct ir3_stream_output_info
*strmout
=
2237 &ctx
->so
->shader
->stream_output
;
2238 struct ir3_block
*orig_end_block
, *stream_out_block
, *new_end_block
;
2239 struct ir3_instruction
*vtxcnt
, *maxvtxcnt
, *cond
;
2240 struct ir3_instruction
*bases
[IR3_MAX_SO_BUFFERS
];
2242 /* create vtxcnt input in input block at top of shader,
2243 * so that it is seen as live over the entire duration
2246 vtxcnt
= create_input(ctx
, 0);
2247 add_sysval_input(ctx
, SYSTEM_VALUE_VERTEX_CNT
, vtxcnt
);
2249 maxvtxcnt
= create_driver_param(ctx
, IR3_DP_VTXCNT_MAX
);
2251 /* at this point, we are at the original 'end' block,
2252 * re-purpose this block to stream-out condition, then
2253 * append stream-out block and new-end block
2255 orig_end_block
= ctx
->block
;
2257 // TODO these blocks need to update predecessors..
2258 // maybe w/ store_global intrinsic, we could do this
2259 // stuff in nir->nir pass
2261 stream_out_block
= ir3_block_create(ir
);
2262 list_addtail(&stream_out_block
->node
, &ir
->block_list
);
2264 new_end_block
= ir3_block_create(ir
);
2265 list_addtail(&new_end_block
->node
, &ir
->block_list
);
2267 orig_end_block
->successors
[0] = stream_out_block
;
2268 orig_end_block
->successors
[1] = new_end_block
;
2269 stream_out_block
->successors
[0] = new_end_block
;
2271 /* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2272 cond
= ir3_CMPS_S(ctx
->block
, vtxcnt
, 0, maxvtxcnt
, 0);
2273 cond
->regs
[0]->num
= regid(REG_P0
, 0);
2274 cond
->cat2
.condition
= IR3_COND_LT
;
2276 /* condition goes on previous block to the conditional,
2277 * since it is used to pick which of the two successor
2280 orig_end_block
->condition
= cond
;
2282 /* switch to stream_out_block to generate the stream-out
2285 ctx
->block
= stream_out_block
;
2287 /* Calculate base addresses based on vtxcnt. Instructions
2288 * generated for bases not used in following loop will be
2289 * stripped out in the backend.
2291 for (unsigned i
= 0; i
< IR3_MAX_SO_BUFFERS
; i
++) {
2292 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
2293 unsigned stride
= strmout
->stride
[i
];
2294 struct ir3_instruction
*base
, *off
;
2296 base
= create_uniform(ctx
->block
, regid(const_state
->offsets
.tfbo
, i
));
2298 /* 24-bit should be enough: */
2299 off
= ir3_MUL_U(ctx
->block
, vtxcnt
, 0,
2300 create_immed(ctx
->block
, stride
* 4), 0);
2302 bases
[i
] = ir3_ADD_S(ctx
->block
, off
, 0, base
, 0);
2305 /* Generate the per-output store instructions: */
2306 for (unsigned i
= 0; i
< strmout
->num_outputs
; i
++) {
2307 for (unsigned j
= 0; j
< strmout
->output
[i
].num_components
; j
++) {
2308 unsigned c
= j
+ strmout
->output
[i
].start_component
;
2309 struct ir3_instruction
*base
, *out
, *stg
;
2311 base
= bases
[strmout
->output
[i
].output_buffer
];
2312 out
= ctx
->ir
->outputs
[regid(strmout
->output
[i
].register_index
, c
)];
2314 stg
= ir3_STG(ctx
->block
, base
, 0, out
, 0,
2315 create_immed(ctx
->block
, 1), 0);
2316 stg
->cat6
.type
= TYPE_U32
;
2317 stg
->cat6
.dst_offset
= (strmout
->output
[i
].dst_offset
+ j
) * 4;
2319 array_insert(ctx
->block
, ctx
->block
->keeps
, stg
);
2323 /* and finally switch to the new_end_block: */
2324 ctx
->block
= new_end_block
;
2328 emit_function(struct ir3_context
*ctx
, nir_function_impl
*impl
)
2330 nir_metadata_require(impl
, nir_metadata_block_index
);
2332 compile_assert(ctx
, ctx
->stack
== 0);
2334 emit_cf_list(ctx
, &impl
->body
);
2335 emit_block(ctx
, impl
->end_block
);
2337 compile_assert(ctx
, ctx
->stack
== 0);
2339 /* at this point, we should have a single empty block,
2340 * into which we emit the 'end' instruction.
2342 compile_assert(ctx
, list_empty(&ctx
->block
->instr_list
));
2344 /* If stream-out (aka transform-feedback) enabled, emit the
2345 * stream-out instructions, followed by a new empty block (into
2346 * which the 'end' instruction lands).
2348 * NOTE: it is done in this order, rather than inserting before
2349 * we emit end_block, because NIR guarantees that all blocks
2350 * flow into end_block, and that end_block has no successors.
2351 * So by re-purposing end_block as the first block of stream-
2352 * out, we guarantee that all exit paths flow into the stream-
2355 if ((ctx
->compiler
->gpu_id
< 500) &&
2356 (ctx
->so
->shader
->stream_output
.num_outputs
> 0) &&
2357 !ctx
->so
->binning_pass
) {
2358 debug_assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
2359 emit_stream_out(ctx
);
2362 ir3_END(ctx
->block
);
2366 setup_input(struct ir3_context
*ctx
, nir_variable
*in
)
2368 struct ir3_shader_variant
*so
= ctx
->so
;
2369 unsigned ncomp
= glsl_get_components(in
->type
);
2370 unsigned n
= in
->data
.driver_location
;
2371 unsigned frac
= in
->data
.location_frac
;
2372 unsigned slot
= in
->data
.location
;
2374 /* skip unread inputs, we could end up with (for example), unsplit
2375 * matrix/etc inputs in the case they are not read, so just silently
2381 so
->inputs
[n
].slot
= slot
;
2382 so
->inputs
[n
].compmask
= (1 << (ncomp
+ frac
)) - 1;
2383 so
->inputs_count
= MAX2(so
->inputs_count
, n
+ 1);
2384 so
->inputs
[n
].interpolate
= in
->data
.interpolation
;
2385 so
->inputs
[n
].ncomp
= ncomp
;
2387 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2389 /* if any varyings have 'sample' qualifer, that triggers us
2390 * to run in per-sample mode:
2392 so
->per_samp
|= in
->data
.sample
;
2394 for (int i
= 0; i
< ncomp
; i
++) {
2395 struct ir3_instruction
*instr
= NULL
;
2396 unsigned idx
= (n
* 4) + i
+ frac
;
2398 if (slot
== VARYING_SLOT_POS
) {
2399 ir3_context_error(ctx
, "fragcoord should be a sysval!\n");
2400 } else if (slot
== VARYING_SLOT_PNTC
) {
2401 /* see for example st_nir_fixup_varying_slots().. this is
2402 * maybe a bit mesa/st specific. But we need things to line
2403 * up for this in fdN_program:
2404 * unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
2405 * if (emit->sprite_coord_enable & texmask) {
2409 so
->inputs
[n
].slot
= VARYING_SLOT_VAR8
;
2410 so
->inputs
[n
].bary
= true;
2411 instr
= create_frag_input(ctx
, false, idx
);
2413 /* detect the special case for front/back colors where
2414 * we need to do flat vs smooth shading depending on
2417 if (in
->data
.interpolation
== INTERP_MODE_NONE
) {
2419 case VARYING_SLOT_COL0
:
2420 case VARYING_SLOT_COL1
:
2421 case VARYING_SLOT_BFC0
:
2422 case VARYING_SLOT_BFC1
:
2423 so
->inputs
[n
].rasterflat
= true;
2430 if (ctx
->compiler
->flat_bypass
) {
2431 if ((so
->inputs
[n
].interpolate
== INTERP_MODE_FLAT
) ||
2432 (so
->inputs
[n
].rasterflat
&& ctx
->so
->key
.rasterflat
))
2433 so
->inputs
[n
].use_ldlv
= true;
2436 so
->inputs
[n
].bary
= true;
2438 instr
= create_frag_input(ctx
, so
->inputs
[n
].use_ldlv
, idx
);
2441 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2443 ctx
->ir
->inputs
[idx
] = instr
;
2445 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2446 for (int i
= 0; i
< ncomp
; i
++) {
2447 unsigned idx
= (n
* 4) + i
+ frac
;
2448 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2449 ctx
->ir
->inputs
[idx
] = create_input(ctx
, idx
);
2452 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2455 if (so
->inputs
[n
].bary
|| (ctx
->so
->type
== MESA_SHADER_VERTEX
)) {
2456 so
->total_in
+= ncomp
;
2460 /* Initially we assign non-packed inloc's for varyings, as we don't really
2461 * know up-front which components will be unused. After all the compilation
2462 * stages we scan the shader to see which components are actually used, and
2463 * re-pack the inlocs to eliminate unneeded varyings.
2466 pack_inlocs(struct ir3_context
*ctx
)
2468 struct ir3_shader_variant
*so
= ctx
->so
;
2469 uint8_t used_components
[so
->inputs_count
];
2471 memset(used_components
, 0, sizeof(used_components
));
2474 * First Step: scan shader to find which bary.f/ldlv remain:
2477 list_for_each_entry (struct ir3_block
, block
, &ctx
->ir
->block_list
, node
) {
2478 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
2479 if (is_input(instr
)) {
2480 unsigned inloc
= instr
->regs
[1]->iim_val
;
2481 unsigned i
= inloc
/ 4;
2482 unsigned j
= inloc
% 4;
2484 compile_assert(ctx
, instr
->regs
[1]->flags
& IR3_REG_IMMED
);
2485 compile_assert(ctx
, i
< so
->inputs_count
);
2487 used_components
[i
] |= 1 << j
;
2493 * Second Step: reassign varying inloc/slots:
2496 unsigned actual_in
= 0;
2499 for (unsigned i
= 0; i
< so
->inputs_count
; i
++) {
2500 unsigned compmask
= 0, maxcomp
= 0;
2502 so
->inputs
[i
].ncomp
= 0;
2503 so
->inputs
[i
].inloc
= inloc
;
2504 so
->inputs
[i
].bary
= false;
2506 for (unsigned j
= 0; j
< 4; j
++) {
2507 if (!(used_components
[i
] & (1 << j
)))
2510 compmask
|= (1 << j
);
2512 so
->inputs
[i
].ncomp
++;
2515 /* at this point, since used_components[i] mask is only
2516 * considering varyings (ie. not sysvals) we know this
2519 so
->inputs
[i
].bary
= true;
2522 if (so
->inputs
[i
].bary
) {
2524 so
->inputs
[i
].compmask
= (1 << maxcomp
) - 1;
2530 * Third Step: reassign packed inloc's:
2533 list_for_each_entry (struct ir3_block
, block
, &ctx
->ir
->block_list
, node
) {
2534 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
2535 if (is_input(instr
)) {
2536 unsigned inloc
= instr
->regs
[1]->iim_val
;
2537 unsigned i
= inloc
/ 4;
2538 unsigned j
= inloc
% 4;
2540 instr
->regs
[1]->iim_val
= so
->inputs
[i
].inloc
+ j
;
2547 setup_output(struct ir3_context
*ctx
, nir_variable
*out
)
2549 struct ir3_shader_variant
*so
= ctx
->so
;
2550 unsigned ncomp
= glsl_get_components(out
->type
);
2551 unsigned n
= out
->data
.driver_location
;
2552 unsigned frac
= out
->data
.location_frac
;
2553 unsigned slot
= out
->data
.location
;
2556 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2558 case FRAG_RESULT_DEPTH
:
2559 comp
= 2; /* tgsi will write to .z component */
2560 so
->writes_pos
= true;
2562 case FRAG_RESULT_COLOR
:
2565 case FRAG_RESULT_SAMPLE_MASK
:
2566 so
->writes_smask
= true;
2569 if (slot
>= FRAG_RESULT_DATA0
)
2571 ir3_context_error(ctx
, "unknown FS output name: %s\n",
2572 gl_frag_result_name(slot
));
2574 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2576 case VARYING_SLOT_POS
:
2577 so
->writes_pos
= true;
2579 case VARYING_SLOT_PSIZ
:
2580 so
->writes_psize
= true;
2582 case VARYING_SLOT_COL0
:
2583 case VARYING_SLOT_COL1
:
2584 case VARYING_SLOT_BFC0
:
2585 case VARYING_SLOT_BFC1
:
2586 case VARYING_SLOT_FOGC
:
2587 case VARYING_SLOT_CLIP_DIST0
:
2588 case VARYING_SLOT_CLIP_DIST1
:
2589 case VARYING_SLOT_CLIP_VERTEX
:
2592 if (slot
>= VARYING_SLOT_VAR0
)
2594 if ((VARYING_SLOT_TEX0
<= slot
) && (slot
<= VARYING_SLOT_TEX7
))
2596 ir3_context_error(ctx
, "unknown VS output name: %s\n",
2597 gl_varying_slot_name(slot
));
2600 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2603 compile_assert(ctx
, n
< ARRAY_SIZE(so
->outputs
));
2605 so
->outputs
[n
].slot
= slot
;
2606 so
->outputs
[n
].regid
= regid(n
, comp
);
2607 so
->outputs_count
= MAX2(so
->outputs_count
, n
+ 1);
2609 for (int i
= 0; i
< ncomp
; i
++) {
2610 unsigned idx
= (n
* 4) + i
+ frac
;
2611 compile_assert(ctx
, idx
< ctx
->ir
->noutputs
);
2612 ctx
->ir
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
2615 /* if varying packing doesn't happen, we could end up in a situation
2616 * with "holes" in the output, and since the per-generation code that
2617 * sets up varying linkage registers doesn't expect to have more than
2618 * one varying per vec4 slot, pad the holes.
2620 * Note that this should probably generate a performance warning of
2623 for (int i
= 0; i
< frac
; i
++) {
2624 unsigned idx
= (n
* 4) + i
;
2625 if (!ctx
->ir
->outputs
[idx
]) {
2626 ctx
->ir
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
2632 max_drvloc(struct exec_list
*vars
)
2635 nir_foreach_variable(var
, vars
) {
2636 drvloc
= MAX2(drvloc
, (int)var
->data
.driver_location
);
2641 static const unsigned max_sysvals
[] = {
2642 [MESA_SHADER_FRAGMENT
] = 24, // TODO
2643 [MESA_SHADER_VERTEX
] = 16,
2644 [MESA_SHADER_COMPUTE
] = 16, // TODO how many do we actually need?
2645 [MESA_SHADER_KERNEL
] = 16, // TODO how many do we actually need?
2649 emit_instructions(struct ir3_context
*ctx
)
2651 unsigned ninputs
, noutputs
;
2652 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->s
);
2654 ninputs
= (max_drvloc(&ctx
->s
->inputs
) + 1) * 4;
2655 noutputs
= (max_drvloc(&ctx
->s
->outputs
) + 1) * 4;
2657 /* we need to leave room for sysvals:
2659 ninputs
+= max_sysvals
[ctx
->so
->type
];
2661 ctx
->ir
= ir3_create(ctx
->compiler
, ctx
->so
->type
, ninputs
, noutputs
);
2663 /* Create inputs in first block: */
2664 ctx
->block
= get_block(ctx
, nir_start_block(fxn
));
2665 ctx
->in_block
= ctx
->block
;
2666 list_addtail(&ctx
->block
->node
, &ctx
->ir
->block_list
);
2668 ninputs
-= max_sysvals
[ctx
->so
->type
];
2670 /* for fragment shader, the vcoord input register is used as the
2671 * base for bary.f varying fetch instrs:
2673 * TODO defer creating ctx->ij_pixel and corresponding sysvals
2674 * until emit_intrinsic when we know they are actually needed.
2675 * For now, we defer creating ctx->ij_centroid, etc, since we
2676 * only need ij_pixel for "old style" varying inputs (ie.
2679 struct ir3_instruction
*vcoord
= NULL
;
2680 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2681 struct ir3_instruction
*xy
[2];
2683 vcoord
= create_input_compmask(ctx
, 0, 0x3);
2684 ir3_split_dest(ctx
->block
, xy
, vcoord
, 0, 2);
2686 ctx
->ij_pixel
= ir3_create_collect(ctx
, xy
, 2);
2690 nir_foreach_variable(var
, &ctx
->s
->inputs
) {
2691 setup_input(ctx
, var
);
2694 /* Defer add_sysval_input() stuff until after setup_inputs(),
2695 * because sysvals need to be appended after varyings:
2698 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_BARYCENTRIC_PIXEL
,
2702 /* Setup outputs: */
2703 nir_foreach_variable(var
, &ctx
->s
->outputs
) {
2704 setup_output(ctx
, var
);
2707 /* Find # of samplers: */
2708 nir_foreach_variable(var
, &ctx
->s
->uniforms
) {
2709 ctx
->so
->num_samp
+= glsl_type_get_sampler_count(var
->type
);
2710 /* just assume that we'll be reading from images.. if it
2711 * is write-only we don't have to count it, but not sure
2712 * if there is a good way to know?
2714 ctx
->so
->num_samp
+= glsl_type_get_image_count(var
->type
);
2717 /* NOTE: need to do something more clever when we support >1 fxn */
2718 nir_foreach_register(reg
, &fxn
->registers
) {
2719 ir3_declare_array(ctx
, reg
);
2721 /* And emit the body: */
2723 emit_function(ctx
, fxn
);
2726 /* from NIR perspective, we actually have varying inputs. But the varying
2727 * inputs, from an IR standpoint, are just bary.f/ldlv instructions. The
2728 * only actual inputs are the sysvals.
2731 fixup_frag_inputs(struct ir3_context
*ctx
)
2733 struct ir3_shader_variant
*so
= ctx
->so
;
2734 struct ir3
*ir
= ctx
->ir
;
2737 /* sysvals should appear at the end of the inputs, drop everything else: */
2738 while ((i
< so
->inputs_count
) && !so
->inputs
[i
].sysval
)
2741 /* at IR level, inputs are always blocks of 4 scalars: */
2744 ir
->inputs
= &ir
->inputs
[i
];
2748 /* Fixup tex sampler state for astc/srgb workaround instructions. We
2749 * need to assign the tex state indexes for these after we know the
2753 fixup_astc_srgb(struct ir3_context
*ctx
)
2755 struct ir3_shader_variant
*so
= ctx
->so
;
2756 /* indexed by original tex idx, value is newly assigned alpha sampler
2757 * state tex idx. Zero is invalid since there is at least one sampler
2760 unsigned alt_tex_state
[16] = {0};
2761 unsigned tex_idx
= ctx
->max_texture_index
+ 1;
2764 so
->astc_srgb
.base
= tex_idx
;
2766 for (unsigned i
= 0; i
< ctx
->ir
->astc_srgb_count
; i
++) {
2767 struct ir3_instruction
*sam
= ctx
->ir
->astc_srgb
[i
];
2769 compile_assert(ctx
, sam
->cat5
.tex
< ARRAY_SIZE(alt_tex_state
));
2771 if (alt_tex_state
[sam
->cat5
.tex
] == 0) {
2772 /* assign new alternate/alpha tex state slot: */
2773 alt_tex_state
[sam
->cat5
.tex
] = tex_idx
++;
2774 so
->astc_srgb
.orig_idx
[idx
++] = sam
->cat5
.tex
;
2775 so
->astc_srgb
.count
++;
2778 sam
->cat5
.tex
= alt_tex_state
[sam
->cat5
.tex
];
2783 fixup_binning_pass(struct ir3_context
*ctx
)
2785 struct ir3_shader_variant
*so
= ctx
->so
;
2786 struct ir3
*ir
= ctx
->ir
;
2789 for (i
= 0, j
= 0; i
< so
->outputs_count
; i
++) {
2790 unsigned slot
= so
->outputs
[i
].slot
;
2792 /* throw away everything but first position/psize */
2793 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
2795 so
->outputs
[j
] = so
->outputs
[i
];
2796 ir
->outputs
[(j
*4)+0] = ir
->outputs
[(i
*4)+0];
2797 ir
->outputs
[(j
*4)+1] = ir
->outputs
[(i
*4)+1];
2798 ir
->outputs
[(j
*4)+2] = ir
->outputs
[(i
*4)+2];
2799 ir
->outputs
[(j
*4)+3] = ir
->outputs
[(i
*4)+3];
2804 so
->outputs_count
= j
;
2805 ir
->noutputs
= j
* 4;
2809 ir3_compile_shader_nir(struct ir3_compiler
*compiler
,
2810 struct ir3_shader_variant
*so
)
2812 struct ir3_context
*ctx
;
2814 struct ir3_instruction
**inputs
;
2816 int ret
= 0, max_bary
;
2820 ctx
= ir3_context_init(compiler
, so
);
2822 DBG("INIT failed!");
2827 emit_instructions(ctx
);
2830 DBG("EMIT failed!");
2835 ir
= so
->ir
= ctx
->ir
;
2837 /* keep track of the inputs from TGSI perspective.. */
2838 inputs
= ir
->inputs
;
2840 /* but fixup actual inputs for frag shader: */
2841 if (so
->type
== MESA_SHADER_FRAGMENT
)
2842 fixup_frag_inputs(ctx
);
2844 /* at this point, for binning pass, throw away unneeded outputs: */
2845 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
< 600))
2846 fixup_binning_pass(ctx
);
2848 /* if we want half-precision outputs, mark the output registers
2851 if (so
->key
.half_precision
) {
2852 for (i
= 0; i
< ir
->noutputs
; i
++) {
2853 struct ir3_instruction
*out
= ir
->outputs
[i
];
2858 /* if frag shader writes z, that needs to be full precision: */
2859 if (so
->outputs
[i
/4].slot
== FRAG_RESULT_DEPTH
)
2862 out
->regs
[0]->flags
|= IR3_REG_HALF
;
2863 /* output could be a fanout (ie. texture fetch output)
2864 * in which case we need to propagate the half-reg flag
2865 * up to the definer so that RA sees it:
2867 if (out
->opc
== OPC_META_FO
) {
2868 out
= out
->regs
[1]->instr
;
2869 out
->regs
[0]->flags
|= IR3_REG_HALF
;
2872 if (out
->opc
== OPC_MOV
) {
2873 out
->cat1
.dst_type
= half_type(out
->cat1
.dst_type
);
2878 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2879 printf("BEFORE CP:\n");
2885 /* at this point, for binning pass, throw away unneeded outputs:
2886 * Note that for a6xx and later, we do this after ir3_cp to ensure
2887 * that the uniform/constant layout for BS and VS matches, so that
2888 * we can re-use same VS_CONST state group.
2890 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
>= 600))
2891 fixup_binning_pass(ctx
);
2893 /* Insert mov if there's same instruction for each output.
2894 * eg. dEQP-GLES31.functional.shaders.opaque_type_indexing.sampler.const_expression.vertex.sampler2dshadow
2896 for (int i
= ir
->noutputs
- 1; i
>= 0; i
--) {
2897 if (!ir
->outputs
[i
])
2899 for (unsigned j
= 0; j
< i
; j
++) {
2900 if (ir
->outputs
[i
] == ir
->outputs
[j
]) {
2902 ir3_MOV(ir
->outputs
[i
]->block
, ir
->outputs
[i
], TYPE_F32
);
2907 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2908 printf("BEFORE GROUPING:\n");
2912 ir3_sched_add_deps(ir
);
2914 /* Group left/right neighbors, inserting mov's where needed to
2919 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2920 printf("AFTER GROUPING:\n");
2926 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2927 printf("AFTER DEPTH:\n");
2931 /* do Sethi–Ullman numbering before scheduling: */
2934 ret
= ir3_sched(ir
);
2936 DBG("SCHED failed!");
2940 if (compiler
->gpu_id
>= 600) {
2941 ir3_a6xx_fixup_atomic_dests(ir
, so
);
2944 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2945 printf("AFTER SCHED:\n");
2949 ret
= ir3_ra(ir
, so
->type
, so
->frag_coord
, so
->frag_face
);
2955 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2956 printf("AFTER RA:\n");
2960 if (so
->type
== MESA_SHADER_FRAGMENT
)
2963 /* fixup input/outputs: */
2964 for (i
= 0; i
< so
->outputs_count
; i
++) {
2965 /* sometimes we get outputs that don't write the .x coord, like:
2967 * decl_var shader_out INTERP_MODE_NONE float Color (VARYING_SLOT_VAR9.z, 1, 0)
2969 * Presumably the result of varying packing and then eliminating
2970 * some unneeded varyings? Just skip head to the first valid
2971 * component of the output.
2973 for (unsigned j
= 0; j
< 4; j
++) {
2974 struct ir3_instruction
*instr
= ir
->outputs
[(i
*4) + j
];
2976 so
->outputs
[i
].regid
= instr
->regs
[0]->num
;
2977 so
->outputs
[i
].half
= !!(instr
->regs
[0]->flags
& IR3_REG_HALF
);
2983 /* Note that some or all channels of an input may be unused: */
2984 for (i
= 0; i
< so
->inputs_count
; i
++) {
2985 unsigned j
, reg
= regid(63,0);
2987 for (j
= 0; j
< 4; j
++) {
2988 struct ir3_instruction
*in
= inputs
[(i
*4) + j
];
2990 if (in
&& !(in
->flags
& IR3_INSTR_UNUSED
)) {
2991 reg
= in
->regs
[0]->num
- j
;
2993 compile_assert(ctx
, in
->regs
[0]->flags
& IR3_REG_HALF
);
2995 half
= !!(in
->regs
[0]->flags
& IR3_REG_HALF
);
2999 so
->inputs
[i
].regid
= reg
;
3000 so
->inputs
[i
].half
= half
;
3004 fixup_astc_srgb(ctx
);
3006 /* We need to do legalize after (for frag shader's) the "bary.f"
3007 * offsets (inloc) have been assigned.
3009 ir3_legalize(ir
, &so
->has_ssbo
, &so
->need_pixlod
, &max_bary
);
3011 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3012 printf("AFTER LEGALIZE:\n");
3016 so
->branchstack
= ctx
->max_stack
;
3018 /* Note that actual_in counts inputs that are not bary.f'd for FS: */
3019 if (so
->type
== MESA_SHADER_FRAGMENT
)
3020 so
->total_in
= max_bary
+ 1;
3022 so
->max_sun
= ir
->max_sun
;
3027 ir3_destroy(so
->ir
);
3030 ir3_context_free(ctx
);