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
;
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
);
391 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, create_immed(b
, fui(0.0)), 0);
392 dst
[0]->cat2
.condition
= IR3_COND_NE
;
393 dst
[0] = ir3_n2b(b
, dst
[0]);
396 dst
[0] = ir3_COV(b
, ir3_b2n(b
, src
[0]), TYPE_U32
, TYPE_F16
);
399 dst
[0] = ir3_COV(b
, ir3_b2n(b
, src
[0]), TYPE_U32
, TYPE_F32
);
404 dst
[0] = ir3_b2n(b
, src
[0]);
407 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, create_immed(b
, 0), 0);
408 dst
[0]->cat2
.condition
= IR3_COND_NE
;
409 dst
[0] = ir3_n2b(b
, dst
[0]);
413 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FNEG
);
416 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FABS
);
419 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[1], 0);
422 dst
[0] = ir3_MIN_F(b
, src
[0], 0, src
[1], 0);
425 /* if there is just a single use of the src, and it supports
426 * (sat) bit, we can just fold the (sat) flag back to the
427 * src instruction and create a mov. This is easier for cp
430 * TODO probably opc_cat==4 is ok too
432 if (alu
->src
[0].src
.is_ssa
&&
433 (list_length(&alu
->src
[0].src
.ssa
->uses
) == 1) &&
434 ((opc_cat(src
[0]->opc
) == 2) || (opc_cat(src
[0]->opc
) == 3))) {
435 src
[0]->flags
|= IR3_INSTR_SAT
;
436 dst
[0] = ir3_MOV(b
, src
[0], dst_type
);
438 /* otherwise generate a max.f that saturates.. blob does
439 * similar (generating a cat2 mov using max.f)
441 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[0], 0);
442 dst
[0]->flags
|= IR3_INSTR_SAT
;
446 dst
[0] = ir3_MUL_F(b
, src
[0], 0, src
[1], 0);
449 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], 0);
452 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], IR3_REG_FNEG
);
455 dst
[0] = ir3_MAD_F32(b
, src
[0], 0, src
[1], 0, src
[2], 0);
458 case nir_op_fddx_coarse
:
459 dst
[0] = ir3_DSX(b
, src
[0], 0);
460 dst
[0]->cat5
.type
= TYPE_F32
;
463 case nir_op_fddy_coarse
:
464 dst
[0] = ir3_DSY(b
, src
[0], 0);
465 dst
[0]->cat5
.type
= TYPE_F32
;
469 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
470 dst
[0]->cat2
.condition
= IR3_COND_LT
;
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_GE
;
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_EQ
;
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_NE
;
486 dst
[0] = ir3_n2b(b
, dst
[0]);
489 dst
[0] = ir3_CEIL_F(b
, src
[0], 0);
492 dst
[0] = ir3_FLOOR_F(b
, src
[0], 0);
495 dst
[0] = ir3_TRUNC_F(b
, src
[0], 0);
497 case nir_op_fround_even
:
498 dst
[0] = ir3_RNDNE_F(b
, src
[0], 0);
501 dst
[0] = ir3_SIGN_F(b
, src
[0], 0);
505 dst
[0] = ir3_SIN(b
, src
[0], 0);
508 dst
[0] = ir3_COS(b
, src
[0], 0);
511 dst
[0] = ir3_RSQ(b
, src
[0], 0);
514 dst
[0] = ir3_RCP(b
, src
[0], 0);
517 dst
[0] = ir3_LOG2(b
, src
[0], 0);
520 dst
[0] = ir3_EXP2(b
, src
[0], 0);
523 dst
[0] = ir3_SQRT(b
, src
[0], 0);
527 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SABS
);
530 dst
[0] = ir3_ADD_U(b
, src
[0], 0, src
[1], 0);
533 dst
[0] = ir3_AND_B(b
, src
[0], 0, src
[1], 0);
536 dst
[0] = ir3_MAX_S(b
, src
[0], 0, src
[1], 0);
539 dst
[0] = ir3_MAX_U(b
, src
[0], 0, src
[1], 0);
542 dst
[0] = ir3_MIN_S(b
, src
[0], 0, src
[1], 0);
545 dst
[0] = ir3_MIN_U(b
, src
[0], 0, src
[1], 0);
547 case nir_op_umul_low
:
548 dst
[0] = ir3_MULL_U(b
, src
[0], 0, src
[1], 0);
550 case nir_op_imadsh_mix16
:
551 dst
[0] = ir3_MADSH_M16(b
, src
[0], 0, src
[1], 0, src
[2], 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
));
703 /* NOTE: since relative addressing is used, make sure constlen is
704 * at least big enough to cover all the UBO addresses, since the
705 * assembler won't know what the max address reg is.
707 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
708 const_state
->offsets
.ubo
+ (ctx
->s
->info
.num_ubos
* ptrsz
));
711 /* note: on 32bit gpu's base_hi is ignored and DCE'd */
714 if (nir_src_is_const(intr
->src
[1])) {
715 off
+= nir_src_as_uint(intr
->src
[1]);
717 /* For load_ubo_indirect, second src is indirect offset: */
718 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0];
720 /* and add offset to addr: */
721 addr
= ir3_ADD_S(b
, addr
, 0, src1
, 0);
724 /* if offset is to large to encode in the ldg, split it out: */
725 if ((off
+ (intr
->num_components
* 4)) > 1024) {
726 /* split out the minimal amount to improve the odds that
727 * cp can fit the immediate in the add.s instruction:
729 unsigned off2
= off
+ (intr
->num_components
* 4) - 1024;
730 addr
= ir3_ADD_S(b
, addr
, 0, create_immed(b
, off2
), 0);
735 struct ir3_instruction
*carry
;
737 /* handle 32b rollover, ie:
738 * if (addr < base_lo)
741 carry
= ir3_CMPS_U(b
, addr
, 0, base_lo
, 0);
742 carry
->cat2
.condition
= IR3_COND_LT
;
743 base_hi
= ir3_ADD_S(b
, base_hi
, 0, carry
, 0);
745 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){ addr
, base_hi
}, 2);
748 for (int i
= 0; i
< intr
->num_components
; i
++) {
749 struct ir3_instruction
*load
=
750 ir3_LDG(b
, addr
, 0, create_immed(b
, 1), 0);
751 load
->cat6
.type
= TYPE_U32
;
752 load
->cat6
.src_offset
= off
+ i
* 4; /* byte offset */
757 /* src[] = { block_index } */
759 emit_intrinsic_ssbo_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
760 struct ir3_instruction
**dst
)
762 /* SSBO size stored as a const starting at ssbo_sizes: */
763 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
764 unsigned blk_idx
= nir_src_as_uint(intr
->src
[0]);
765 unsigned idx
= regid(const_state
->offsets
.ssbo_sizes
, 0) +
766 const_state
->ssbo_size
.off
[blk_idx
];
768 debug_assert(const_state
->ssbo_size
.mask
& (1 << blk_idx
));
770 dst
[0] = create_uniform(ctx
->block
, idx
);
773 /* src[] = { offset }. const_index[] = { base } */
775 emit_intrinsic_load_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
776 struct ir3_instruction
**dst
)
778 struct ir3_block
*b
= ctx
->block
;
779 struct ir3_instruction
*ldl
, *offset
;
782 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
783 base
= nir_intrinsic_base(intr
);
785 ldl
= ir3_LDL(b
, offset
, 0, create_immed(b
, intr
->num_components
), 0);
786 ldl
->cat6
.src_offset
= base
;
787 ldl
->cat6
.type
= utype_dst(intr
->dest
);
788 ldl
->regs
[0]->wrmask
= MASK(intr
->num_components
);
790 ldl
->barrier_class
= IR3_BARRIER_SHARED_R
;
791 ldl
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
793 ir3_split_dest(b
, dst
, ldl
, 0, intr
->num_components
);
796 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
798 emit_intrinsic_store_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
800 struct ir3_block
*b
= ctx
->block
;
801 struct ir3_instruction
*stl
, *offset
;
802 struct ir3_instruction
* const *value
;
803 unsigned base
, wrmask
;
805 value
= ir3_get_src(ctx
, &intr
->src
[0]);
806 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
808 base
= nir_intrinsic_base(intr
);
809 wrmask
= nir_intrinsic_write_mask(intr
);
811 /* Combine groups of consecutive enabled channels in one write
812 * message. We use ffs to find the first enabled channel and then ffs on
813 * the bit-inverse, down-shifted writemask to determine the length of
814 * the block of enabled bits.
816 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
819 unsigned first_component
= ffs(wrmask
) - 1;
820 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
822 stl
= ir3_STL(b
, offset
, 0,
823 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
824 create_immed(b
, length
), 0);
825 stl
->cat6
.dst_offset
= first_component
+ base
;
826 stl
->cat6
.type
= utype_src(intr
->src
[0]);
827 stl
->barrier_class
= IR3_BARRIER_SHARED_W
;
828 stl
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
830 array_insert(b
, b
->keeps
, stl
);
832 /* Clear the bits in the writemask that we just wrote, then try
833 * again to see if more channels are left.
835 wrmask
&= (15 << (first_component
+ length
));
840 * CS shared variable atomic intrinsics
842 * All of the shared variable atomic memory operations read a value from
843 * memory, compute a new value using one of the operations below, write the
844 * new value to memory, and return the original value read.
846 * All operations take 2 sources except CompSwap that takes 3. These
849 * 0: The offset into the shared variable storage region that the atomic
850 * operation will operate on.
851 * 1: The data parameter to the atomic function (i.e. the value to add
852 * in shared_atomic_add, etc).
853 * 2: For CompSwap only: the second data parameter.
855 static struct ir3_instruction
*
856 emit_intrinsic_atomic_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
858 struct ir3_block
*b
= ctx
->block
;
859 struct ir3_instruction
*atomic
, *src0
, *src1
;
860 type_t type
= TYPE_U32
;
862 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0]; /* offset */
863 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0]; /* value */
865 switch (intr
->intrinsic
) {
866 case nir_intrinsic_shared_atomic_add
:
867 atomic
= ir3_ATOMIC_ADD(b
, src0
, 0, src1
, 0);
869 case nir_intrinsic_shared_atomic_imin
:
870 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
873 case nir_intrinsic_shared_atomic_umin
:
874 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
876 case nir_intrinsic_shared_atomic_imax
:
877 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
880 case nir_intrinsic_shared_atomic_umax
:
881 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
883 case nir_intrinsic_shared_atomic_and
:
884 atomic
= ir3_ATOMIC_AND(b
, src0
, 0, src1
, 0);
886 case nir_intrinsic_shared_atomic_or
:
887 atomic
= ir3_ATOMIC_OR(b
, src0
, 0, src1
, 0);
889 case nir_intrinsic_shared_atomic_xor
:
890 atomic
= ir3_ATOMIC_XOR(b
, src0
, 0, src1
, 0);
892 case nir_intrinsic_shared_atomic_exchange
:
893 atomic
= ir3_ATOMIC_XCHG(b
, src0
, 0, src1
, 0);
895 case nir_intrinsic_shared_atomic_comp_swap
:
896 /* for cmpxchg, src1 is [ui]vec2(data, compare): */
897 src1
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
898 ir3_get_src(ctx
, &intr
->src
[2])[0],
901 atomic
= ir3_ATOMIC_CMPXCHG(b
, src0
, 0, src1
, 0);
907 atomic
->cat6
.iim_val
= 1;
909 atomic
->cat6
.type
= type
;
910 atomic
->barrier_class
= IR3_BARRIER_SHARED_W
;
911 atomic
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
913 /* even if nothing consume the result, we can't DCE the instruction: */
914 array_insert(b
, b
->keeps
, atomic
);
919 /* TODO handle actual indirect/dynamic case.. which is going to be weird
920 * to handle with the image_mapping table..
922 static struct ir3_instruction
*
923 get_image_samp_tex_src(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
925 unsigned slot
= ir3_get_image_slot(nir_src_as_deref(intr
->src
[0]));
926 unsigned tex_idx
= ir3_image_to_tex(&ctx
->so
->image_mapping
, slot
);
927 struct ir3_instruction
*texture
, *sampler
;
929 texture
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
930 sampler
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
932 return ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
938 /* src[] = { deref, coord, sample_index }. const_index[] = {} */
940 emit_intrinsic_load_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
941 struct ir3_instruction
**dst
)
943 struct ir3_block
*b
= ctx
->block
;
944 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
945 struct ir3_instruction
*samp_tex
= get_image_samp_tex_src(ctx
, intr
);
946 struct ir3_instruction
*sam
;
947 struct ir3_instruction
* const *src0
= ir3_get_src(ctx
, &intr
->src
[1]);
948 struct ir3_instruction
*coords
[4];
949 unsigned flags
, ncoords
= ir3_get_image_coords(var
, &flags
);
950 type_t type
= ir3_get_image_type(var
);
952 /* hmm, this seems a bit odd, but it is what blob does and (at least
953 * a5xx) just faults on bogus addresses otherwise:
955 if (flags
& IR3_INSTR_3D
) {
956 flags
&= ~IR3_INSTR_3D
;
957 flags
|= IR3_INSTR_A
;
960 for (unsigned i
= 0; i
< ncoords
; i
++)
964 coords
[ncoords
++] = create_immed(b
, 0);
966 sam
= ir3_SAM(b
, OPC_ISAM
, type
, 0b1111, flags
,
967 samp_tex
, ir3_create_collect(ctx
, coords
, ncoords
), NULL
);
969 sam
->barrier_class
= IR3_BARRIER_IMAGE_R
;
970 sam
->barrier_conflict
= IR3_BARRIER_IMAGE_W
;
972 ir3_split_dest(b
, dst
, sam
, 0, 4);
976 emit_intrinsic_image_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
977 struct ir3_instruction
**dst
)
979 struct ir3_block
*b
= ctx
->block
;
980 const nir_variable
*var
= nir_intrinsic_get_var(intr
, 0);
981 struct ir3_instruction
*samp_tex
= get_image_samp_tex_src(ctx
, intr
);
982 struct ir3_instruction
*sam
, *lod
;
983 unsigned flags
, ncoords
= ir3_get_image_coords(var
, &flags
);
985 lod
= create_immed(b
, 0);
986 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
987 samp_tex
, lod
, NULL
);
989 /* Array size actually ends up in .w rather than .z. This doesn't
990 * matter for miplevel 0, but for higher mips the value in z is
991 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
992 * returned, which means that we have to add 1 to it for arrays for
995 * Note use a temporary dst and then copy, since the size of the dst
996 * array that is passed in is based on nir's understanding of the
997 * result size, not the hardware's
999 struct ir3_instruction
*tmp
[4];
1001 ir3_split_dest(b
, tmp
, sam
, 0, 4);
1003 /* get_size instruction returns size in bytes instead of texels
1004 * for imageBuffer, so we need to divide it by the pixel size
1005 * of the image format.
1007 * TODO: This is at least true on a5xx. Check other gens.
1009 enum glsl_sampler_dim dim
=
1010 glsl_get_sampler_dim(glsl_without_array(var
->type
));
1011 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
1012 /* Since all the possible values the divisor can take are
1013 * power-of-two (4, 8, or 16), the division is implemented
1015 * During shader setup, the log2 of the image format's
1016 * bytes-per-pixel should have been emitted in 2nd slot of
1017 * image_dims. See ir3_shader::emit_image_dims().
1019 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
1020 unsigned cb
= regid(const_state
->offsets
.image_dims
, 0) +
1021 const_state
->image_dims
.off
[var
->data
.driver_location
];
1022 struct ir3_instruction
*aux
= create_uniform(b
, cb
+ 1);
1024 tmp
[0] = ir3_SHR_B(b
, tmp
[0], 0, aux
, 0);
1027 for (unsigned i
= 0; i
< ncoords
; i
++)
1030 if (flags
& IR3_INSTR_A
) {
1031 if (ctx
->compiler
->levels_add_one
) {
1032 dst
[ncoords
-1] = ir3_ADD_U(b
, tmp
[3], 0, create_immed(b
, 1), 0);
1034 dst
[ncoords
-1] = ir3_MOV(b
, tmp
[3], TYPE_U32
);
1040 emit_intrinsic_barrier(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1042 struct ir3_block
*b
= ctx
->block
;
1043 struct ir3_instruction
*barrier
;
1045 switch (intr
->intrinsic
) {
1046 case nir_intrinsic_barrier
:
1047 barrier
= ir3_BAR(b
);
1048 barrier
->cat7
.g
= true;
1049 barrier
->cat7
.l
= true;
1050 barrier
->flags
= IR3_INSTR_SS
| IR3_INSTR_SY
;
1051 barrier
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1053 case nir_intrinsic_memory_barrier
:
1054 barrier
= ir3_FENCE(b
);
1055 barrier
->cat7
.g
= true;
1056 barrier
->cat7
.r
= true;
1057 barrier
->cat7
.w
= true;
1058 barrier
->cat7
.l
= true;
1059 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
|
1060 IR3_BARRIER_BUFFER_W
;
1061 barrier
->barrier_conflict
=
1062 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1063 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1065 case nir_intrinsic_memory_barrier_atomic_counter
:
1066 case nir_intrinsic_memory_barrier_buffer
:
1067 barrier
= ir3_FENCE(b
);
1068 barrier
->cat7
.g
= true;
1069 barrier
->cat7
.r
= true;
1070 barrier
->cat7
.w
= true;
1071 barrier
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1072 barrier
->barrier_conflict
= IR3_BARRIER_BUFFER_R
|
1073 IR3_BARRIER_BUFFER_W
;
1075 case nir_intrinsic_memory_barrier_image
:
1076 // TODO double check if this should have .g set
1077 barrier
= ir3_FENCE(b
);
1078 barrier
->cat7
.g
= true;
1079 barrier
->cat7
.r
= true;
1080 barrier
->cat7
.w
= true;
1081 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1082 barrier
->barrier_conflict
= IR3_BARRIER_IMAGE_R
|
1083 IR3_BARRIER_IMAGE_W
;
1085 case nir_intrinsic_memory_barrier_shared
:
1086 barrier
= ir3_FENCE(b
);
1087 barrier
->cat7
.g
= true;
1088 barrier
->cat7
.l
= true;
1089 barrier
->cat7
.r
= true;
1090 barrier
->cat7
.w
= true;
1091 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
;
1092 barrier
->barrier_conflict
= IR3_BARRIER_SHARED_R
|
1093 IR3_BARRIER_SHARED_W
;
1095 case nir_intrinsic_group_memory_barrier
:
1096 barrier
= ir3_FENCE(b
);
1097 barrier
->cat7
.g
= true;
1098 barrier
->cat7
.l
= true;
1099 barrier
->cat7
.r
= true;
1100 barrier
->cat7
.w
= true;
1101 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
|
1102 IR3_BARRIER_IMAGE_W
|
1103 IR3_BARRIER_BUFFER_W
;
1104 barrier
->barrier_conflict
=
1105 IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
|
1106 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1107 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1113 /* make sure barrier doesn't get DCE'd */
1114 array_insert(b
, b
->keeps
, barrier
);
1117 static void add_sysval_input_compmask(struct ir3_context
*ctx
,
1118 gl_system_value slot
, unsigned compmask
,
1119 struct ir3_instruction
*instr
)
1121 struct ir3_shader_variant
*so
= ctx
->so
;
1122 unsigned r
= regid(so
->inputs_count
, 0);
1123 unsigned n
= so
->inputs_count
++;
1125 so
->inputs
[n
].sysval
= true;
1126 so
->inputs
[n
].slot
= slot
;
1127 so
->inputs
[n
].compmask
= compmask
;
1128 so
->inputs
[n
].regid
= r
;
1129 so
->inputs
[n
].interpolate
= INTERP_MODE_FLAT
;
1132 ctx
->ir
->ninputs
= MAX2(ctx
->ir
->ninputs
, r
+ 1);
1133 ctx
->ir
->inputs
[r
] = instr
;
1136 static void add_sysval_input(struct ir3_context
*ctx
, gl_system_value slot
,
1137 struct ir3_instruction
*instr
)
1139 add_sysval_input_compmask(ctx
, slot
, 0x1, instr
);
1142 static struct ir3_instruction
*
1143 get_barycentric_centroid(struct ir3_context
*ctx
)
1145 if (!ctx
->ij_centroid
) {
1146 struct ir3_instruction
*xy
[2];
1147 struct ir3_instruction
*ij
;
1149 ij
= create_input_compmask(ctx
, 0, 0x3);
1150 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1152 ctx
->ij_centroid
= ir3_create_collect(ctx
, xy
, 2);
1154 add_sysval_input_compmask(ctx
,
1155 SYSTEM_VALUE_BARYCENTRIC_CENTROID
,
1159 return ctx
->ij_centroid
;
1162 static struct ir3_instruction
*
1163 get_barycentric_sample(struct ir3_context
*ctx
)
1165 if (!ctx
->ij_sample
) {
1166 struct ir3_instruction
*xy
[2];
1167 struct ir3_instruction
*ij
;
1169 ij
= create_input_compmask(ctx
, 0, 0x3);
1170 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1172 ctx
->ij_sample
= ir3_create_collect(ctx
, xy
, 2);
1174 add_sysval_input_compmask(ctx
,
1175 SYSTEM_VALUE_BARYCENTRIC_SAMPLE
,
1179 return ctx
->ij_sample
;
1182 static struct ir3_instruction
*
1183 get_barycentric_pixel(struct ir3_context
*ctx
)
1185 /* TODO when tgsi_to_nir supports "new-style" FS inputs switch
1186 * this to create ij_pixel only on demand:
1188 return ctx
->ij_pixel
;
1191 static struct ir3_instruction
*
1192 get_frag_coord(struct ir3_context
*ctx
)
1194 if (!ctx
->frag_coord
) {
1195 struct ir3_block
*b
= ctx
->block
;
1196 struct ir3_instruction
*xyzw
[4];
1197 struct ir3_instruction
*hw_frag_coord
;
1199 hw_frag_coord
= create_input_compmask(ctx
, 0, 0xf);
1200 ir3_split_dest(ctx
->block
, xyzw
, hw_frag_coord
, 0, 4);
1202 /* for frag_coord.xy, we get unsigned values.. we need
1203 * to subtract (integer) 8 and divide by 16 (right-
1204 * shift by 4) then convert to float:
1208 * mov.u32f32 dst, tmp
1211 for (int i
= 0; i
< 2; i
++) {
1212 xyzw
[i
] = ir3_SUB_S(b
, xyzw
[i
], 0,
1213 create_immed(b
, 8), 0);
1214 xyzw
[i
] = ir3_SHR_B(b
, xyzw
[i
], 0,
1215 create_immed(b
, 4), 0);
1216 xyzw
[i
] = ir3_COV(b
, xyzw
[i
], TYPE_U32
, TYPE_F32
);
1219 ctx
->frag_coord
= ir3_create_collect(ctx
, xyzw
, 4);
1221 add_sysval_input_compmask(ctx
,
1222 SYSTEM_VALUE_FRAG_COORD
,
1223 0xf, hw_frag_coord
);
1225 ctx
->so
->frag_coord
= true;
1228 return ctx
->frag_coord
;
1232 emit_intrinsic(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1234 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1235 struct ir3_instruction
**dst
;
1236 struct ir3_instruction
* const *src
;
1237 struct ir3_block
*b
= ctx
->block
;
1240 if (info
->has_dest
) {
1241 unsigned n
= nir_intrinsic_dest_components(intr
);
1242 dst
= ir3_get_dst(ctx
, &intr
->dest
, n
);
1247 switch (intr
->intrinsic
) {
1248 case nir_intrinsic_load_uniform
:
1249 idx
= nir_intrinsic_base(intr
);
1250 if (nir_src_is_const(intr
->src
[0])) {
1251 idx
+= nir_src_as_uint(intr
->src
[0]);
1252 for (int i
= 0; i
< intr
->num_components
; i
++) {
1253 dst
[i
] = create_uniform_typed(b
, idx
+ i
,
1254 nir_dest_bit_size(intr
->dest
) < 32 ? TYPE_F16
: TYPE_F32
);
1257 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1258 for (int i
= 0; i
< intr
->num_components
; i
++) {
1259 dst
[i
] = create_uniform_indirect(b
, idx
+ i
,
1260 ir3_get_addr(ctx
, src
[0], 1));
1262 /* NOTE: if relative addressing is used, we set
1263 * constlen in the compiler (to worst-case value)
1264 * since we don't know in the assembler what the max
1265 * addr reg value can be:
1267 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
1268 ctx
->so
->shader
->ubo_state
.size
/ 16);
1271 case nir_intrinsic_load_ubo
:
1272 emit_intrinsic_load_ubo(ctx
, intr
, dst
);
1274 case nir_intrinsic_load_frag_coord
:
1275 ir3_split_dest(b
, dst
, get_frag_coord(ctx
), 0, 4);
1277 case nir_intrinsic_load_sample_pos_from_id
: {
1278 /* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
1279 * but that doesn't seem necessary.
1281 struct ir3_instruction
*offset
=
1282 ir3_RGETPOS(b
, ir3_get_src(ctx
, &intr
->src
[0])[0], 0);
1283 offset
->regs
[0]->wrmask
= 0x3;
1284 offset
->cat5
.type
= TYPE_F32
;
1286 ir3_split_dest(b
, dst
, offset
, 0, 2);
1290 case nir_intrinsic_load_size_ir3
:
1291 if (!ctx
->ij_size
) {
1292 ctx
->ij_size
= create_input(ctx
, 0);
1294 add_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_SIZE
,
1297 dst
[0] = ctx
->ij_size
;
1299 case nir_intrinsic_load_barycentric_centroid
:
1300 ir3_split_dest(b
, dst
, get_barycentric_centroid(ctx
), 0, 2);
1302 case nir_intrinsic_load_barycentric_sample
:
1303 if (ctx
->so
->key
.msaa
) {
1304 ir3_split_dest(b
, dst
, get_barycentric_sample(ctx
), 0, 2);
1306 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1309 case nir_intrinsic_load_barycentric_pixel
:
1310 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1312 case nir_intrinsic_load_interpolated_input
:
1313 idx
= nir_intrinsic_base(intr
);
1314 comp
= nir_intrinsic_component(intr
);
1315 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1316 if (nir_src_is_const(intr
->src
[1])) {
1317 struct ir3_instruction
*coord
= ir3_create_collect(ctx
, src
, 2);
1318 idx
+= nir_src_as_uint(intr
->src
[1]);
1319 for (int i
= 0; i
< intr
->num_components
; i
++) {
1320 unsigned inloc
= idx
* 4 + i
+ comp
;
1321 if (ctx
->so
->inputs
[idx
].bary
&&
1322 !ctx
->so
->inputs
[idx
].use_ldlv
) {
1323 dst
[i
] = ir3_BARY_F(b
, create_immed(b
, inloc
), 0, coord
, 0);
1325 /* for non-varyings use the pre-setup input, since
1326 * that is easier than mapping things back to a
1327 * nir_variable to figure out what it is.
1329 dst
[i
] = ctx
->ir
->inputs
[inloc
];
1333 ir3_context_error(ctx
, "unhandled");
1336 case nir_intrinsic_load_input
:
1337 idx
= nir_intrinsic_base(intr
);
1338 comp
= nir_intrinsic_component(intr
);
1339 if (nir_src_is_const(intr
->src
[0])) {
1340 idx
+= nir_src_as_uint(intr
->src
[0]);
1341 for (int i
= 0; i
< intr
->num_components
; i
++) {
1342 unsigned n
= idx
* 4 + i
+ comp
;
1343 dst
[i
] = ctx
->ir
->inputs
[n
];
1344 compile_assert(ctx
, ctx
->ir
->inputs
[n
]);
1347 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1348 struct ir3_instruction
*collect
=
1349 ir3_create_collect(ctx
, ctx
->ir
->inputs
, ctx
->ir
->ninputs
);
1350 struct ir3_instruction
*addr
= ir3_get_addr(ctx
, src
[0], 4);
1351 for (int i
= 0; i
< intr
->num_components
; i
++) {
1352 unsigned n
= idx
* 4 + i
+ comp
;
1353 dst
[i
] = create_indirect_load(ctx
, ctx
->ir
->ninputs
,
1358 /* All SSBO intrinsics should have been lowered by 'lower_io_offsets'
1359 * pass and replaced by an ir3-specifc version that adds the
1360 * dword-offset in the last source.
1362 case nir_intrinsic_load_ssbo_ir3
:
1363 ctx
->funcs
->emit_intrinsic_load_ssbo(ctx
, intr
, dst
);
1365 case nir_intrinsic_store_ssbo_ir3
:
1366 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1367 !ctx
->s
->info
.fs
.early_fragment_tests
)
1368 ctx
->so
->no_earlyz
= true;
1369 ctx
->funcs
->emit_intrinsic_store_ssbo(ctx
, intr
);
1371 case nir_intrinsic_get_buffer_size
:
1372 emit_intrinsic_ssbo_size(ctx
, intr
, dst
);
1374 case nir_intrinsic_ssbo_atomic_add_ir3
:
1375 case nir_intrinsic_ssbo_atomic_imin_ir3
:
1376 case nir_intrinsic_ssbo_atomic_umin_ir3
:
1377 case nir_intrinsic_ssbo_atomic_imax_ir3
:
1378 case nir_intrinsic_ssbo_atomic_umax_ir3
:
1379 case nir_intrinsic_ssbo_atomic_and_ir3
:
1380 case nir_intrinsic_ssbo_atomic_or_ir3
:
1381 case nir_intrinsic_ssbo_atomic_xor_ir3
:
1382 case nir_intrinsic_ssbo_atomic_exchange_ir3
:
1383 case nir_intrinsic_ssbo_atomic_comp_swap_ir3
:
1384 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1385 !ctx
->s
->info
.fs
.early_fragment_tests
)
1386 ctx
->so
->no_earlyz
= true;
1387 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_ssbo(ctx
, intr
);
1389 case nir_intrinsic_load_shared
:
1390 emit_intrinsic_load_shared(ctx
, intr
, dst
);
1392 case nir_intrinsic_store_shared
:
1393 emit_intrinsic_store_shared(ctx
, intr
);
1395 case nir_intrinsic_shared_atomic_add
:
1396 case nir_intrinsic_shared_atomic_imin
:
1397 case nir_intrinsic_shared_atomic_umin
:
1398 case nir_intrinsic_shared_atomic_imax
:
1399 case nir_intrinsic_shared_atomic_umax
:
1400 case nir_intrinsic_shared_atomic_and
:
1401 case nir_intrinsic_shared_atomic_or
:
1402 case nir_intrinsic_shared_atomic_xor
:
1403 case nir_intrinsic_shared_atomic_exchange
:
1404 case nir_intrinsic_shared_atomic_comp_swap
:
1405 dst
[0] = emit_intrinsic_atomic_shared(ctx
, intr
);
1407 case nir_intrinsic_image_deref_load
:
1408 emit_intrinsic_load_image(ctx
, intr
, dst
);
1410 case nir_intrinsic_image_deref_store
:
1411 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1412 !ctx
->s
->info
.fs
.early_fragment_tests
)
1413 ctx
->so
->no_earlyz
= true;
1414 ctx
->funcs
->emit_intrinsic_store_image(ctx
, intr
);
1416 case nir_intrinsic_image_deref_size
:
1417 emit_intrinsic_image_size(ctx
, intr
, dst
);
1419 case nir_intrinsic_image_deref_atomic_add
:
1420 case nir_intrinsic_image_deref_atomic_imin
:
1421 case nir_intrinsic_image_deref_atomic_umin
:
1422 case nir_intrinsic_image_deref_atomic_imax
:
1423 case nir_intrinsic_image_deref_atomic_umax
:
1424 case nir_intrinsic_image_deref_atomic_and
:
1425 case nir_intrinsic_image_deref_atomic_or
:
1426 case nir_intrinsic_image_deref_atomic_xor
:
1427 case nir_intrinsic_image_deref_atomic_exchange
:
1428 case nir_intrinsic_image_deref_atomic_comp_swap
:
1429 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1430 !ctx
->s
->info
.fs
.early_fragment_tests
)
1431 ctx
->so
->no_earlyz
= true;
1432 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_image(ctx
, intr
);
1434 case nir_intrinsic_barrier
:
1435 case nir_intrinsic_memory_barrier
:
1436 case nir_intrinsic_group_memory_barrier
:
1437 case nir_intrinsic_memory_barrier_atomic_counter
:
1438 case nir_intrinsic_memory_barrier_buffer
:
1439 case nir_intrinsic_memory_barrier_image
:
1440 case nir_intrinsic_memory_barrier_shared
:
1441 emit_intrinsic_barrier(ctx
, intr
);
1442 /* note that blk ptr no longer valid, make that obvious: */
1445 case nir_intrinsic_store_output
:
1446 idx
= nir_intrinsic_base(intr
);
1447 comp
= nir_intrinsic_component(intr
);
1448 compile_assert(ctx
, nir_src_is_const(intr
->src
[1]));
1449 idx
+= nir_src_as_uint(intr
->src
[1]);
1451 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1452 for (int i
= 0; i
< intr
->num_components
; i
++) {
1453 unsigned n
= idx
* 4 + i
+ comp
;
1454 ctx
->ir
->outputs
[n
] = src
[i
];
1457 case nir_intrinsic_load_base_vertex
:
1458 case nir_intrinsic_load_first_vertex
:
1459 if (!ctx
->basevertex
) {
1460 ctx
->basevertex
= create_driver_param(ctx
, IR3_DP_VTXID_BASE
);
1461 add_sysval_input(ctx
, SYSTEM_VALUE_FIRST_VERTEX
, ctx
->basevertex
);
1463 dst
[0] = ctx
->basevertex
;
1465 case nir_intrinsic_load_vertex_id_zero_base
:
1466 case nir_intrinsic_load_vertex_id
:
1467 if (!ctx
->vertex_id
) {
1468 gl_system_value sv
= (intr
->intrinsic
== nir_intrinsic_load_vertex_id
) ?
1469 SYSTEM_VALUE_VERTEX_ID
: SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1470 ctx
->vertex_id
= create_input(ctx
, 0);
1471 add_sysval_input(ctx
, sv
, ctx
->vertex_id
);
1473 dst
[0] = ctx
->vertex_id
;
1475 case nir_intrinsic_load_instance_id
:
1476 if (!ctx
->instance_id
) {
1477 ctx
->instance_id
= create_input(ctx
, 0);
1478 add_sysval_input(ctx
, SYSTEM_VALUE_INSTANCE_ID
,
1481 dst
[0] = ctx
->instance_id
;
1483 case nir_intrinsic_load_sample_id
:
1484 ctx
->so
->per_samp
= true;
1486 case nir_intrinsic_load_sample_id_no_per_sample
:
1487 if (!ctx
->samp_id
) {
1488 ctx
->samp_id
= create_input(ctx
, 0);
1489 ctx
->samp_id
->regs
[0]->flags
|= IR3_REG_HALF
;
1490 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_ID
,
1493 dst
[0] = ir3_COV(b
, ctx
->samp_id
, TYPE_U16
, TYPE_U32
);
1495 case nir_intrinsic_load_sample_mask_in
:
1496 if (!ctx
->samp_mask_in
) {
1497 ctx
->samp_mask_in
= create_input(ctx
, 0);
1498 add_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_MASK_IN
,
1501 dst
[0] = ctx
->samp_mask_in
;
1503 case nir_intrinsic_load_user_clip_plane
:
1504 idx
= nir_intrinsic_ucp_id(intr
);
1505 for (int i
= 0; i
< intr
->num_components
; i
++) {
1506 unsigned n
= idx
* 4 + i
;
1507 dst
[i
] = create_driver_param(ctx
, IR3_DP_UCP0_X
+ n
);
1510 case nir_intrinsic_load_front_face
:
1511 if (!ctx
->frag_face
) {
1512 ctx
->so
->frag_face
= true;
1513 ctx
->frag_face
= create_input(ctx
, 0);
1514 add_sysval_input(ctx
, SYSTEM_VALUE_FRONT_FACE
, ctx
->frag_face
);
1515 ctx
->frag_face
->regs
[0]->flags
|= IR3_REG_HALF
;
1517 /* for fragface, we get -1 for back and 0 for front. However this is
1518 * the inverse of what nir expects (where ~0 is true).
1520 dst
[0] = ir3_COV(b
, ctx
->frag_face
, TYPE_S16
, TYPE_S32
);
1521 dst
[0] = ir3_NOT_B(b
, dst
[0], 0);
1523 case nir_intrinsic_load_local_invocation_id
:
1524 if (!ctx
->local_invocation_id
) {
1525 ctx
->local_invocation_id
= create_input_compmask(ctx
, 0, 0x7);
1526 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_LOCAL_INVOCATION_ID
,
1527 0x7, ctx
->local_invocation_id
);
1529 ir3_split_dest(b
, dst
, ctx
->local_invocation_id
, 0, 3);
1531 case nir_intrinsic_load_work_group_id
:
1532 if (!ctx
->work_group_id
) {
1533 ctx
->work_group_id
= create_input_compmask(ctx
, 0, 0x7);
1534 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_WORK_GROUP_ID
,
1535 0x7, ctx
->work_group_id
);
1536 ctx
->work_group_id
->regs
[0]->flags
|= IR3_REG_HIGH
;
1538 ir3_split_dest(b
, dst
, ctx
->work_group_id
, 0, 3);
1540 case nir_intrinsic_load_num_work_groups
:
1541 for (int i
= 0; i
< intr
->num_components
; i
++) {
1542 dst
[i
] = create_driver_param(ctx
, IR3_DP_NUM_WORK_GROUPS_X
+ i
);
1545 case nir_intrinsic_load_local_group_size
:
1546 for (int i
= 0; i
< intr
->num_components
; i
++) {
1547 dst
[i
] = create_driver_param(ctx
, IR3_DP_LOCAL_GROUP_SIZE_X
+ i
);
1550 case nir_intrinsic_discard_if
:
1551 case nir_intrinsic_discard
: {
1552 struct ir3_instruction
*cond
, *kill
;
1554 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
1555 /* conditional discard: */
1556 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1557 cond
= ir3_b2n(b
, src
[0]);
1559 /* unconditional discard: */
1560 cond
= create_immed(b
, 1);
1563 /* NOTE: only cmps.*.* can write p0.x: */
1564 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1565 cond
->cat2
.condition
= IR3_COND_NE
;
1567 /* condition always goes in predicate register: */
1568 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1570 kill
= ir3_KILL(b
, cond
, 0);
1571 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1573 array_insert(b
, b
->keeps
, kill
);
1574 ctx
->so
->no_earlyz
= true;
1579 ir3_context_error(ctx
, "Unhandled intrinsic type: %s\n",
1580 nir_intrinsic_infos
[intr
->intrinsic
].name
);
1585 ir3_put_dst(ctx
, &intr
->dest
);
1589 emit_load_const(struct ir3_context
*ctx
, nir_load_const_instr
*instr
)
1591 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &instr
->def
,
1592 instr
->def
.num_components
);
1594 if (instr
->def
.bit_size
< 32) {
1595 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1596 dst
[i
] = create_immed_typed(ctx
->block
,
1597 instr
->value
[i
].u16
,
1600 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1601 dst
[i
] = create_immed_typed(ctx
->block
,
1602 instr
->value
[i
].u32
,
1609 emit_undef(struct ir3_context
*ctx
, nir_ssa_undef_instr
*undef
)
1611 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &undef
->def
,
1612 undef
->def
.num_components
);
1613 type_t type
= (undef
->def
.bit_size
< 32) ? TYPE_U16
: TYPE_U32
;
1615 /* backend doesn't want undefined instructions, so just plug
1618 for (int i
= 0; i
< undef
->def
.num_components
; i
++)
1619 dst
[i
] = create_immed_typed(ctx
->block
, fui(0.0), type
);
1623 * texture fetch/sample instructions:
1627 tex_info(nir_tex_instr
*tex
, unsigned *flagsp
, unsigned *coordsp
)
1629 unsigned coords
, flags
= 0;
1631 /* note: would use tex->coord_components.. except txs.. also,
1632 * since array index goes after shadow ref, we don't want to
1635 switch (tex
->sampler_dim
) {
1636 case GLSL_SAMPLER_DIM_1D
:
1637 case GLSL_SAMPLER_DIM_BUF
:
1640 case GLSL_SAMPLER_DIM_2D
:
1641 case GLSL_SAMPLER_DIM_RECT
:
1642 case GLSL_SAMPLER_DIM_EXTERNAL
:
1643 case GLSL_SAMPLER_DIM_MS
:
1646 case GLSL_SAMPLER_DIM_3D
:
1647 case GLSL_SAMPLER_DIM_CUBE
:
1649 flags
|= IR3_INSTR_3D
;
1652 unreachable("bad sampler_dim");
1655 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
1656 flags
|= IR3_INSTR_S
;
1658 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
)
1659 flags
|= IR3_INSTR_A
;
1665 /* Gets the sampler/texture idx as a hvec2. Which could either be dynamic
1666 * or immediate (in which case it will get lowered later to a non .s2en
1667 * version of the tex instruction which encode tex/samp as immediates:
1669 static struct ir3_instruction
*
1670 get_tex_samp_tex_src(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1672 int texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
);
1673 int sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
);
1674 struct ir3_instruction
*texture
, *sampler
;
1676 if (texture_idx
>= 0) {
1677 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
1678 texture
= ir3_COV(ctx
->block
, texture
, TYPE_U32
, TYPE_U16
);
1680 /* TODO what to do for dynamic case? I guess we only need the
1681 * max index for astc srgb workaround so maybe not a problem
1682 * to worry about if we don't enable indirect samplers for
1685 ctx
->max_texture_index
= MAX2(ctx
->max_texture_index
, tex
->texture_index
);
1686 texture
= create_immed_typed(ctx
->block
, tex
->texture_index
, TYPE_U16
);
1689 if (sampler_idx
>= 0) {
1690 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
1691 sampler
= ir3_COV(ctx
->block
, sampler
, TYPE_U32
, TYPE_U16
);
1693 sampler
= create_immed_typed(ctx
->block
, tex
->sampler_index
, TYPE_U16
);
1696 return ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1703 emit_tex(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1705 struct ir3_block
*b
= ctx
->block
;
1706 struct ir3_instruction
**dst
, *sam
, *src0
[12], *src1
[4];
1707 struct ir3_instruction
* const *coord
, * const *off
, * const *ddx
, * const *ddy
;
1708 struct ir3_instruction
*lod
, *compare
, *proj
, *sample_index
;
1709 bool has_bias
= false, has_lod
= false, has_proj
= false, has_off
= false;
1710 unsigned i
, coords
, flags
, ncomp
;
1711 unsigned nsrc0
= 0, nsrc1
= 0;
1715 ncomp
= nir_dest_num_components(tex
->dest
);
1717 coord
= off
= ddx
= ddy
= NULL
;
1718 lod
= proj
= compare
= sample_index
= NULL
;
1720 dst
= ir3_get_dst(ctx
, &tex
->dest
, ncomp
);
1722 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1723 switch (tex
->src
[i
].src_type
) {
1724 case nir_tex_src_coord
:
1725 coord
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1727 case nir_tex_src_bias
:
1728 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1731 case nir_tex_src_lod
:
1732 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1735 case nir_tex_src_comparator
: /* shadow comparator */
1736 compare
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1738 case nir_tex_src_projector
:
1739 proj
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1742 case nir_tex_src_offset
:
1743 off
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1746 case nir_tex_src_ddx
:
1747 ddx
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1749 case nir_tex_src_ddy
:
1750 ddy
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
1752 case nir_tex_src_ms_index
:
1753 sample_index
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
1755 case nir_tex_src_texture_offset
:
1756 case nir_tex_src_sampler_offset
:
1757 /* handled in get_tex_samp_src() */
1760 ir3_context_error(ctx
, "Unhandled NIR tex src type: %d\n",
1761 tex
->src
[i
].src_type
);
1767 case nir_texop_tex
: opc
= has_lod
? OPC_SAML
: OPC_SAM
; break;
1768 case nir_texop_txb
: opc
= OPC_SAMB
; break;
1769 case nir_texop_txl
: opc
= OPC_SAML
; break;
1770 case nir_texop_txd
: opc
= OPC_SAMGQ
; break;
1771 case nir_texop_txf
: opc
= OPC_ISAML
; break;
1772 case nir_texop_lod
: opc
= OPC_GETLOD
; break;
1774 /* NOTE: a4xx might need to emulate gather w/ txf (this is
1775 * what blob does, seems gather is broken?), and a3xx did
1776 * not support it (but probably could also emulate).
1778 switch (tex
->component
) {
1779 case 0: opc
= OPC_GATHER4R
; break;
1780 case 1: opc
= OPC_GATHER4G
; break;
1781 case 2: opc
= OPC_GATHER4B
; break;
1782 case 3: opc
= OPC_GATHER4A
; break;
1785 case nir_texop_txf_ms_fb
:
1786 case nir_texop_txf_ms
: opc
= OPC_ISAMM
; break;
1788 ir3_context_error(ctx
, "Unhandled NIR tex type: %d\n", tex
->op
);
1792 tex_info(tex
, &flags
, &coords
);
1795 * lay out the first argument in the proper order:
1796 * - actual coordinates first
1797 * - shadow reference
1800 * - starting at offset 4, dpdx.xy, dpdy.xy
1802 * bias/lod go into the second arg
1805 /* insert tex coords: */
1806 for (i
= 0; i
< coords
; i
++)
1811 /* scale up integer coords for TXF based on the LOD */
1812 if (ctx
->compiler
->unminify_coords
&& (opc
== OPC_ISAML
)) {
1814 for (i
= 0; i
< coords
; i
++)
1815 src0
[i
] = ir3_SHL_B(b
, src0
[i
], 0, lod
, 0);
1819 /* hw doesn't do 1d, so we treat it as 2d with
1820 * height of 1, and patch up the y coord.
1823 src0
[nsrc0
++] = create_immed(b
, 0);
1825 src0
[nsrc0
++] = create_immed(b
, fui(0.5));
1829 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
1830 src0
[nsrc0
++] = compare
;
1832 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
) {
1833 struct ir3_instruction
*idx
= coord
[coords
];
1835 /* the array coord for cube arrays needs 0.5 added to it */
1836 if (ctx
->compiler
->array_index_add_half
&& !is_isam(opc
))
1837 idx
= ir3_ADD_F(b
, idx
, 0, create_immed(b
, fui(0.5)), 0);
1839 src0
[nsrc0
++] = idx
;
1843 src0
[nsrc0
++] = proj
;
1844 flags
|= IR3_INSTR_P
;
1847 /* pad to 4, then ddx/ddy: */
1848 if (tex
->op
== nir_texop_txd
) {
1850 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1851 for (i
= 0; i
< coords
; i
++)
1852 src0
[nsrc0
++] = ddx
[i
];
1854 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1855 for (i
= 0; i
< coords
; i
++)
1856 src0
[nsrc0
++] = ddy
[i
];
1858 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
1861 /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
1862 * with scaled x coord according to requested sample:
1864 if (opc
== OPC_ISAMM
) {
1865 if (ctx
->compiler
->txf_ms_with_isaml
) {
1866 /* the samples are laid out in x dimension as
1868 * x_ms = (x << ms) + sample_index;
1870 struct ir3_instruction
*ms
;
1871 ms
= create_immed(b
, (ctx
->samples
>> (2 * tex
->texture_index
)) & 3);
1873 src0
[0] = ir3_SHL_B(b
, src0
[0], 0, ms
, 0);
1874 src0
[0] = ir3_ADD_U(b
, src0
[0], 0, sample_index
, 0);
1878 src0
[nsrc0
++] = sample_index
;
1883 * second argument (if applicable):
1888 if (has_off
| has_lod
| has_bias
) {
1890 unsigned off_coords
= coords
;
1891 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
1893 for (i
= 0; i
< off_coords
; i
++)
1894 src1
[nsrc1
++] = off
[i
];
1896 src1
[nsrc1
++] = create_immed(b
, fui(0.0));
1897 flags
|= IR3_INSTR_O
;
1900 if (has_lod
| has_bias
)
1901 src1
[nsrc1
++] = lod
;
1904 switch (tex
->dest_type
) {
1905 case nir_type_invalid
:
1906 case nir_type_float
:
1917 unreachable("bad dest_type");
1920 if (opc
== OPC_GETLOD
)
1923 struct ir3_instruction
*samp_tex
;
1925 if (tex
->op
== nir_texop_txf_ms_fb
) {
1926 /* only expect a single txf_ms_fb per shader: */
1927 compile_assert(ctx
, !ctx
->so
->fb_read
);
1928 compile_assert(ctx
, ctx
->so
->type
== MESA_SHADER_FRAGMENT
);
1930 ctx
->so
->fb_read
= true;
1931 samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1932 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
1933 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
1936 ctx
->so
->num_samp
++;
1938 samp_tex
= get_tex_samp_tex_src(ctx
, tex
);
1941 struct ir3_instruction
*col0
= ir3_create_collect(ctx
, src0
, nsrc0
);
1942 struct ir3_instruction
*col1
= ir3_create_collect(ctx
, src1
, nsrc1
);
1944 sam
= ir3_SAM(b
, opc
, type
, MASK(ncomp
), flags
,
1945 samp_tex
, col0
, col1
);
1947 if ((ctx
->astc_srgb
& (1 << tex
->texture_index
)) && !nir_tex_instr_is_query(tex
)) {
1948 /* only need first 3 components: */
1949 sam
->regs
[0]->wrmask
= 0x7;
1950 ir3_split_dest(b
, dst
, sam
, 0, 3);
1952 /* we need to sample the alpha separately with a non-ASTC
1955 sam
= ir3_SAM(b
, opc
, type
, 0b1000, flags
,
1956 samp_tex
, col0
, col1
);
1958 array_insert(ctx
->ir
, ctx
->ir
->astc_srgb
, sam
);
1960 /* fixup .w component: */
1961 ir3_split_dest(b
, &dst
[3], sam
, 3, 1);
1963 /* normal (non-workaround) case: */
1964 ir3_split_dest(b
, dst
, sam
, 0, ncomp
);
1967 /* GETLOD returns results in 4.8 fixed point */
1968 if (opc
== OPC_GETLOD
) {
1969 struct ir3_instruction
*factor
= create_immed(b
, fui(1.0 / 256));
1971 compile_assert(ctx
, tex
->dest_type
== nir_type_float
);
1972 for (i
= 0; i
< 2; i
++) {
1973 dst
[i
] = ir3_MUL_F(b
, ir3_COV(b
, dst
[i
], TYPE_U32
, TYPE_F32
), 0,
1978 ir3_put_dst(ctx
, &tex
->dest
);
1982 emit_tex_query_levels(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
1984 struct ir3_block
*b
= ctx
->block
;
1985 struct ir3_instruction
**dst
, *sam
;
1987 dst
= ir3_get_dst(ctx
, &tex
->dest
, 1);
1989 sam
= ir3_SAM(b
, OPC_GETINFO
, TYPE_U32
, 0b0100, 0,
1990 get_tex_samp_tex_src(ctx
, tex
), NULL
, NULL
);
1992 /* even though there is only one component, since it ends
1993 * up in .z rather than .x, we need a split_dest()
1995 ir3_split_dest(b
, dst
, sam
, 0, 3);
1997 /* The # of levels comes from getinfo.z. We need to add 1 to it, since
1998 * the value in TEX_CONST_0 is zero-based.
2000 if (ctx
->compiler
->levels_add_one
)
2001 dst
[0] = ir3_ADD_U(b
, dst
[0], 0, create_immed(b
, 1), 0);
2003 ir3_put_dst(ctx
, &tex
->dest
);
2007 emit_tex_txs(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2009 struct ir3_block
*b
= ctx
->block
;
2010 struct ir3_instruction
**dst
, *sam
;
2011 struct ir3_instruction
*lod
;
2012 unsigned flags
, coords
;
2014 tex_info(tex
, &flags
, &coords
);
2016 /* Actually we want the number of dimensions, not coordinates. This
2017 * distinction only matters for cubes.
2019 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2022 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
2024 compile_assert(ctx
, tex
->num_srcs
== 1);
2025 compile_assert(ctx
, tex
->src
[0].src_type
== nir_tex_src_lod
);
2027 lod
= ir3_get_src(ctx
, &tex
->src
[0].src
)[0];
2029 sam
= ir3_SAM(b
, OPC_GETSIZE
, TYPE_U32
, 0b1111, flags
,
2030 get_tex_samp_tex_src(ctx
, tex
), lod
, NULL
);
2032 ir3_split_dest(b
, dst
, sam
, 0, 4);
2034 /* Array size actually ends up in .w rather than .z. This doesn't
2035 * matter for miplevel 0, but for higher mips the value in z is
2036 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2037 * returned, which means that we have to add 1 to it for arrays.
2039 if (tex
->is_array
) {
2040 if (ctx
->compiler
->levels_add_one
) {
2041 dst
[coords
] = ir3_ADD_U(b
, dst
[3], 0, create_immed(b
, 1), 0);
2043 dst
[coords
] = ir3_MOV(b
, dst
[3], TYPE_U32
);
2047 ir3_put_dst(ctx
, &tex
->dest
);
2051 emit_jump(struct ir3_context
*ctx
, nir_jump_instr
*jump
)
2053 switch (jump
->type
) {
2054 case nir_jump_break
:
2055 case nir_jump_continue
:
2056 case nir_jump_return
:
2057 /* I *think* we can simply just ignore this, and use the
2058 * successor block link to figure out where we need to
2059 * jump to for break/continue
2063 ir3_context_error(ctx
, "Unhandled NIR jump type: %d\n", jump
->type
);
2069 emit_instr(struct ir3_context
*ctx
, nir_instr
*instr
)
2071 switch (instr
->type
) {
2072 case nir_instr_type_alu
:
2073 emit_alu(ctx
, nir_instr_as_alu(instr
));
2075 case nir_instr_type_deref
:
2076 /* ignored, handled as part of the intrinsic they are src to */
2078 case nir_instr_type_intrinsic
:
2079 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2081 case nir_instr_type_load_const
:
2082 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2084 case nir_instr_type_ssa_undef
:
2085 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
2087 case nir_instr_type_tex
: {
2088 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
2089 /* couple tex instructions get special-cased:
2093 emit_tex_txs(ctx
, tex
);
2095 case nir_texop_query_levels
:
2096 emit_tex_query_levels(ctx
, tex
);
2104 case nir_instr_type_jump
:
2105 emit_jump(ctx
, nir_instr_as_jump(instr
));
2107 case nir_instr_type_phi
:
2108 /* we have converted phi webs to regs in NIR by now */
2109 ir3_context_error(ctx
, "Unexpected NIR instruction type: %d\n", instr
->type
);
2111 case nir_instr_type_call
:
2112 case nir_instr_type_parallel_copy
:
2113 ir3_context_error(ctx
, "Unhandled NIR instruction type: %d\n", instr
->type
);
2118 static struct ir3_block
*
2119 get_block(struct ir3_context
*ctx
, const nir_block
*nblock
)
2121 struct ir3_block
*block
;
2122 struct hash_entry
*hentry
;
2124 hentry
= _mesa_hash_table_search(ctx
->block_ht
, nblock
);
2126 return hentry
->data
;
2128 block
= ir3_block_create(ctx
->ir
);
2129 block
->nblock
= nblock
;
2130 _mesa_hash_table_insert(ctx
->block_ht
, nblock
, block
);
2132 block
->predecessors
= _mesa_pointer_set_create(block
);
2133 set_foreach(nblock
->predecessors
, sentry
) {
2134 _mesa_set_add(block
->predecessors
, get_block(ctx
, sentry
->key
));
2141 emit_block(struct ir3_context
*ctx
, nir_block
*nblock
)
2143 struct ir3_block
*block
= get_block(ctx
, nblock
);
2145 for (int i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
2146 if (nblock
->successors
[i
]) {
2147 block
->successors
[i
] =
2148 get_block(ctx
, nblock
->successors
[i
]);
2153 list_addtail(&block
->node
, &ctx
->ir
->block_list
);
2155 /* re-emit addr register in each block if needed: */
2156 for (int i
= 0; i
< ARRAY_SIZE(ctx
->addr_ht
); i
++) {
2157 _mesa_hash_table_destroy(ctx
->addr_ht
[i
], NULL
);
2158 ctx
->addr_ht
[i
] = NULL
;
2161 nir_foreach_instr(instr
, nblock
) {
2162 ctx
->cur_instr
= instr
;
2163 emit_instr(ctx
, instr
);
2164 ctx
->cur_instr
= NULL
;
2170 static void emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
);
2173 emit_if(struct ir3_context
*ctx
, nir_if
*nif
)
2175 struct ir3_instruction
*condition
= ir3_get_src(ctx
, &nif
->condition
)[0];
2177 ctx
->block
->condition
=
2178 ir3_get_predicate(ctx
, ir3_b2n(condition
->block
, condition
));
2180 emit_cf_list(ctx
, &nif
->then_list
);
2181 emit_cf_list(ctx
, &nif
->else_list
);
2185 emit_loop(struct ir3_context
*ctx
, nir_loop
*nloop
)
2187 emit_cf_list(ctx
, &nloop
->body
);
2192 stack_push(struct ir3_context
*ctx
)
2195 ctx
->max_stack
= MAX2(ctx
->max_stack
, ctx
->stack
);
2199 stack_pop(struct ir3_context
*ctx
)
2201 compile_assert(ctx
, ctx
->stack
> 0);
2206 emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
)
2208 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2209 switch (node
->type
) {
2210 case nir_cf_node_block
:
2211 emit_block(ctx
, nir_cf_node_as_block(node
));
2213 case nir_cf_node_if
:
2215 emit_if(ctx
, nir_cf_node_as_if(node
));
2218 case nir_cf_node_loop
:
2220 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2223 case nir_cf_node_function
:
2224 ir3_context_error(ctx
, "TODO\n");
2230 /* emit stream-out code. At this point, the current block is the original
2231 * (nir) end block, and nir ensures that all flow control paths terminate
2232 * into the end block. We re-purpose the original end block to generate
2233 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2234 * block holding stream-out write instructions, followed by the new end
2238 * p0.x = (vtxcnt < maxvtxcnt)
2239 * // succs: blockStreamOut, blockNewEnd
2242 * ... stream-out instructions ...
2243 * // succs: blockNewEnd
2249 emit_stream_out(struct ir3_context
*ctx
)
2251 struct ir3
*ir
= ctx
->ir
;
2252 struct ir3_stream_output_info
*strmout
=
2253 &ctx
->so
->shader
->stream_output
;
2254 struct ir3_block
*orig_end_block
, *stream_out_block
, *new_end_block
;
2255 struct ir3_instruction
*vtxcnt
, *maxvtxcnt
, *cond
;
2256 struct ir3_instruction
*bases
[IR3_MAX_SO_BUFFERS
];
2258 /* create vtxcnt input in input block at top of shader,
2259 * so that it is seen as live over the entire duration
2262 vtxcnt
= create_input(ctx
, 0);
2263 add_sysval_input(ctx
, SYSTEM_VALUE_VERTEX_CNT
, vtxcnt
);
2265 maxvtxcnt
= create_driver_param(ctx
, IR3_DP_VTXCNT_MAX
);
2267 /* at this point, we are at the original 'end' block,
2268 * re-purpose this block to stream-out condition, then
2269 * append stream-out block and new-end block
2271 orig_end_block
= ctx
->block
;
2273 // TODO these blocks need to update predecessors..
2274 // maybe w/ store_global intrinsic, we could do this
2275 // stuff in nir->nir pass
2277 stream_out_block
= ir3_block_create(ir
);
2278 list_addtail(&stream_out_block
->node
, &ir
->block_list
);
2280 new_end_block
= ir3_block_create(ir
);
2281 list_addtail(&new_end_block
->node
, &ir
->block_list
);
2283 orig_end_block
->successors
[0] = stream_out_block
;
2284 orig_end_block
->successors
[1] = new_end_block
;
2285 stream_out_block
->successors
[0] = new_end_block
;
2287 /* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2288 cond
= ir3_CMPS_S(ctx
->block
, vtxcnt
, 0, maxvtxcnt
, 0);
2289 cond
->regs
[0]->num
= regid(REG_P0
, 0);
2290 cond
->cat2
.condition
= IR3_COND_LT
;
2292 /* condition goes on previous block to the conditional,
2293 * since it is used to pick which of the two successor
2296 orig_end_block
->condition
= cond
;
2298 /* switch to stream_out_block to generate the stream-out
2301 ctx
->block
= stream_out_block
;
2303 /* Calculate base addresses based on vtxcnt. Instructions
2304 * generated for bases not used in following loop will be
2305 * stripped out in the backend.
2307 for (unsigned i
= 0; i
< IR3_MAX_SO_BUFFERS
; i
++) {
2308 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
2309 unsigned stride
= strmout
->stride
[i
];
2310 struct ir3_instruction
*base
, *off
;
2312 base
= create_uniform(ctx
->block
, regid(const_state
->offsets
.tfbo
, i
));
2314 /* 24-bit should be enough: */
2315 off
= ir3_MUL_U(ctx
->block
, vtxcnt
, 0,
2316 create_immed(ctx
->block
, stride
* 4), 0);
2318 bases
[i
] = ir3_ADD_S(ctx
->block
, off
, 0, base
, 0);
2321 /* Generate the per-output store instructions: */
2322 for (unsigned i
= 0; i
< strmout
->num_outputs
; i
++) {
2323 for (unsigned j
= 0; j
< strmout
->output
[i
].num_components
; j
++) {
2324 unsigned c
= j
+ strmout
->output
[i
].start_component
;
2325 struct ir3_instruction
*base
, *out
, *stg
;
2327 base
= bases
[strmout
->output
[i
].output_buffer
];
2328 out
= ctx
->ir
->outputs
[regid(strmout
->output
[i
].register_index
, c
)];
2330 stg
= ir3_STG(ctx
->block
, base
, 0, out
, 0,
2331 create_immed(ctx
->block
, 1), 0);
2332 stg
->cat6
.type
= TYPE_U32
;
2333 stg
->cat6
.dst_offset
= (strmout
->output
[i
].dst_offset
+ j
) * 4;
2335 array_insert(ctx
->block
, ctx
->block
->keeps
, stg
);
2339 /* and finally switch to the new_end_block: */
2340 ctx
->block
= new_end_block
;
2344 emit_function(struct ir3_context
*ctx
, nir_function_impl
*impl
)
2346 nir_metadata_require(impl
, nir_metadata_block_index
);
2348 compile_assert(ctx
, ctx
->stack
== 0);
2350 emit_cf_list(ctx
, &impl
->body
);
2351 emit_block(ctx
, impl
->end_block
);
2353 compile_assert(ctx
, ctx
->stack
== 0);
2355 /* at this point, we should have a single empty block,
2356 * into which we emit the 'end' instruction.
2358 compile_assert(ctx
, list_empty(&ctx
->block
->instr_list
));
2360 /* If stream-out (aka transform-feedback) enabled, emit the
2361 * stream-out instructions, followed by a new empty block (into
2362 * which the 'end' instruction lands).
2364 * NOTE: it is done in this order, rather than inserting before
2365 * we emit end_block, because NIR guarantees that all blocks
2366 * flow into end_block, and that end_block has no successors.
2367 * So by re-purposing end_block as the first block of stream-
2368 * out, we guarantee that all exit paths flow into the stream-
2371 if ((ctx
->compiler
->gpu_id
< 500) &&
2372 (ctx
->so
->shader
->stream_output
.num_outputs
> 0) &&
2373 !ctx
->so
->binning_pass
) {
2374 debug_assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
2375 emit_stream_out(ctx
);
2378 ir3_END(ctx
->block
);
2382 setup_input(struct ir3_context
*ctx
, nir_variable
*in
)
2384 struct ir3_shader_variant
*so
= ctx
->so
;
2385 unsigned ncomp
= glsl_get_components(in
->type
);
2386 unsigned n
= in
->data
.driver_location
;
2387 unsigned frac
= in
->data
.location_frac
;
2388 unsigned slot
= in
->data
.location
;
2390 /* skip unread inputs, we could end up with (for example), unsplit
2391 * matrix/etc inputs in the case they are not read, so just silently
2397 so
->inputs
[n
].slot
= slot
;
2398 so
->inputs
[n
].compmask
= (1 << (ncomp
+ frac
)) - 1;
2399 so
->inputs_count
= MAX2(so
->inputs_count
, n
+ 1);
2400 so
->inputs
[n
].interpolate
= in
->data
.interpolation
;
2402 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2404 /* if any varyings have 'sample' qualifer, that triggers us
2405 * to run in per-sample mode:
2407 so
->per_samp
|= in
->data
.sample
;
2409 for (int i
= 0; i
< ncomp
; i
++) {
2410 struct ir3_instruction
*instr
= NULL
;
2411 unsigned idx
= (n
* 4) + i
+ frac
;
2413 if (slot
== VARYING_SLOT_POS
) {
2414 ir3_context_error(ctx
, "fragcoord should be a sysval!\n");
2415 } else if (slot
== VARYING_SLOT_PNTC
) {
2416 /* see for example st_nir_fixup_varying_slots().. this is
2417 * maybe a bit mesa/st specific. But we need things to line
2418 * up for this in fdN_program:
2419 * unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
2420 * if (emit->sprite_coord_enable & texmask) {
2424 so
->inputs
[n
].slot
= VARYING_SLOT_VAR8
;
2425 so
->inputs
[n
].bary
= true;
2426 instr
= create_frag_input(ctx
, false, idx
);
2428 /* detect the special case for front/back colors where
2429 * we need to do flat vs smooth shading depending on
2432 if (in
->data
.interpolation
== INTERP_MODE_NONE
) {
2434 case VARYING_SLOT_COL0
:
2435 case VARYING_SLOT_COL1
:
2436 case VARYING_SLOT_BFC0
:
2437 case VARYING_SLOT_BFC1
:
2438 so
->inputs
[n
].rasterflat
= true;
2445 if (ctx
->compiler
->flat_bypass
) {
2446 if ((so
->inputs
[n
].interpolate
== INTERP_MODE_FLAT
) ||
2447 (so
->inputs
[n
].rasterflat
&& ctx
->so
->key
.rasterflat
))
2448 so
->inputs
[n
].use_ldlv
= true;
2451 so
->inputs
[n
].bary
= true;
2453 instr
= create_frag_input(ctx
, so
->inputs
[n
].use_ldlv
, idx
);
2456 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2458 ctx
->ir
->inputs
[idx
] = instr
;
2460 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2461 for (int i
= 0; i
< ncomp
; i
++) {
2462 unsigned idx
= (n
* 4) + i
+ frac
;
2463 compile_assert(ctx
, idx
< ctx
->ir
->ninputs
);
2464 ctx
->ir
->inputs
[idx
] = create_input(ctx
, idx
);
2467 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2470 if (so
->inputs
[n
].bary
|| (ctx
->so
->type
== MESA_SHADER_VERTEX
)) {
2471 so
->total_in
+= ncomp
;
2475 /* Initially we assign non-packed inloc's for varyings, as we don't really
2476 * know up-front which components will be unused. After all the compilation
2477 * stages we scan the shader to see which components are actually used, and
2478 * re-pack the inlocs to eliminate unneeded varyings.
2481 pack_inlocs(struct ir3_context
*ctx
)
2483 struct ir3_shader_variant
*so
= ctx
->so
;
2484 uint8_t used_components
[so
->inputs_count
];
2486 memset(used_components
, 0, sizeof(used_components
));
2489 * First Step: scan shader to find which bary.f/ldlv remain:
2492 list_for_each_entry (struct ir3_block
, block
, &ctx
->ir
->block_list
, node
) {
2493 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
2494 if (is_input(instr
)) {
2495 unsigned inloc
= instr
->regs
[1]->iim_val
;
2496 unsigned i
= inloc
/ 4;
2497 unsigned j
= inloc
% 4;
2499 compile_assert(ctx
, instr
->regs
[1]->flags
& IR3_REG_IMMED
);
2500 compile_assert(ctx
, i
< so
->inputs_count
);
2502 used_components
[i
] |= 1 << j
;
2508 * Second Step: reassign varying inloc/slots:
2511 unsigned actual_in
= 0;
2514 for (unsigned i
= 0; i
< so
->inputs_count
; i
++) {
2515 unsigned compmask
= 0, maxcomp
= 0;
2517 so
->inputs
[i
].inloc
= inloc
;
2518 so
->inputs
[i
].bary
= false;
2520 for (unsigned j
= 0; j
< 4; j
++) {
2521 if (!(used_components
[i
] & (1 << j
)))
2524 compmask
|= (1 << j
);
2528 /* at this point, since used_components[i] mask is only
2529 * considering varyings (ie. not sysvals) we know this
2532 so
->inputs
[i
].bary
= true;
2535 if (so
->inputs
[i
].bary
) {
2537 so
->inputs
[i
].compmask
= (1 << maxcomp
) - 1;
2543 * Third Step: reassign packed inloc's:
2546 list_for_each_entry (struct ir3_block
, block
, &ctx
->ir
->block_list
, node
) {
2547 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
2548 if (is_input(instr
)) {
2549 unsigned inloc
= instr
->regs
[1]->iim_val
;
2550 unsigned i
= inloc
/ 4;
2551 unsigned j
= inloc
% 4;
2553 instr
->regs
[1]->iim_val
= so
->inputs
[i
].inloc
+ j
;
2560 setup_output(struct ir3_context
*ctx
, nir_variable
*out
)
2562 struct ir3_shader_variant
*so
= ctx
->so
;
2563 unsigned ncomp
= glsl_get_components(out
->type
);
2564 unsigned n
= out
->data
.driver_location
;
2565 unsigned frac
= out
->data
.location_frac
;
2566 unsigned slot
= out
->data
.location
;
2569 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2571 case FRAG_RESULT_DEPTH
:
2572 comp
= 2; /* tgsi will write to .z component */
2573 so
->writes_pos
= true;
2575 case FRAG_RESULT_COLOR
:
2578 case FRAG_RESULT_SAMPLE_MASK
:
2579 so
->writes_smask
= true;
2582 if (slot
>= FRAG_RESULT_DATA0
)
2584 ir3_context_error(ctx
, "unknown FS output name: %s\n",
2585 gl_frag_result_name(slot
));
2587 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2589 case VARYING_SLOT_POS
:
2590 so
->writes_pos
= true;
2592 case VARYING_SLOT_PSIZ
:
2593 so
->writes_psize
= true;
2595 case VARYING_SLOT_COL0
:
2596 case VARYING_SLOT_COL1
:
2597 case VARYING_SLOT_BFC0
:
2598 case VARYING_SLOT_BFC1
:
2599 case VARYING_SLOT_FOGC
:
2600 case VARYING_SLOT_CLIP_DIST0
:
2601 case VARYING_SLOT_CLIP_DIST1
:
2602 case VARYING_SLOT_CLIP_VERTEX
:
2605 if (slot
>= VARYING_SLOT_VAR0
)
2607 if ((VARYING_SLOT_TEX0
<= slot
) && (slot
<= VARYING_SLOT_TEX7
))
2609 ir3_context_error(ctx
, "unknown VS output name: %s\n",
2610 gl_varying_slot_name(slot
));
2613 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2616 compile_assert(ctx
, n
< ARRAY_SIZE(so
->outputs
));
2618 so
->outputs
[n
].slot
= slot
;
2619 so
->outputs
[n
].regid
= regid(n
, comp
);
2620 so
->outputs_count
= MAX2(so
->outputs_count
, n
+ 1);
2622 for (int i
= 0; i
< ncomp
; i
++) {
2623 unsigned idx
= (n
* 4) + i
+ frac
;
2624 compile_assert(ctx
, idx
< ctx
->ir
->noutputs
);
2625 ctx
->ir
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
2628 /* if varying packing doesn't happen, we could end up in a situation
2629 * with "holes" in the output, and since the per-generation code that
2630 * sets up varying linkage registers doesn't expect to have more than
2631 * one varying per vec4 slot, pad the holes.
2633 * Note that this should probably generate a performance warning of
2636 for (int i
= 0; i
< frac
; i
++) {
2637 unsigned idx
= (n
* 4) + i
;
2638 if (!ctx
->ir
->outputs
[idx
]) {
2639 ctx
->ir
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
2645 max_drvloc(struct exec_list
*vars
)
2648 nir_foreach_variable(var
, vars
) {
2649 drvloc
= MAX2(drvloc
, (int)var
->data
.driver_location
);
2654 static const unsigned max_sysvals
[] = {
2655 [MESA_SHADER_FRAGMENT
] = 24, // TODO
2656 [MESA_SHADER_VERTEX
] = 16,
2657 [MESA_SHADER_COMPUTE
] = 16, // TODO how many do we actually need?
2658 [MESA_SHADER_KERNEL
] = 16, // TODO how many do we actually need?
2662 emit_instructions(struct ir3_context
*ctx
)
2664 unsigned ninputs
, noutputs
;
2665 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->s
);
2667 ninputs
= (max_drvloc(&ctx
->s
->inputs
) + 1) * 4;
2668 noutputs
= (max_drvloc(&ctx
->s
->outputs
) + 1) * 4;
2670 /* we need to leave room for sysvals:
2672 ninputs
+= max_sysvals
[ctx
->so
->type
];
2674 ctx
->ir
= ir3_create(ctx
->compiler
, ctx
->so
->type
, ninputs
, noutputs
);
2676 /* Create inputs in first block: */
2677 ctx
->block
= get_block(ctx
, nir_start_block(fxn
));
2678 ctx
->in_block
= ctx
->block
;
2679 list_addtail(&ctx
->block
->node
, &ctx
->ir
->block_list
);
2681 ninputs
-= max_sysvals
[ctx
->so
->type
];
2683 /* for fragment shader, the vcoord input register is used as the
2684 * base for bary.f varying fetch instrs:
2686 * TODO defer creating ctx->ij_pixel and corresponding sysvals
2687 * until emit_intrinsic when we know they are actually needed.
2688 * For now, we defer creating ctx->ij_centroid, etc, since we
2689 * only need ij_pixel for "old style" varying inputs (ie.
2692 struct ir3_instruction
*vcoord
= NULL
;
2693 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2694 struct ir3_instruction
*xy
[2];
2696 vcoord
= create_input_compmask(ctx
, 0, 0x3);
2697 ir3_split_dest(ctx
->block
, xy
, vcoord
, 0, 2);
2699 ctx
->ij_pixel
= ir3_create_collect(ctx
, xy
, 2);
2703 nir_foreach_variable(var
, &ctx
->s
->inputs
) {
2704 setup_input(ctx
, var
);
2707 /* Defer add_sysval_input() stuff until after setup_inputs(),
2708 * because sysvals need to be appended after varyings:
2711 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_BARYCENTRIC_PIXEL
,
2715 /* Setup outputs: */
2716 nir_foreach_variable(var
, &ctx
->s
->outputs
) {
2717 setup_output(ctx
, var
);
2720 /* Find # of samplers: */
2721 nir_foreach_variable(var
, &ctx
->s
->uniforms
) {
2722 ctx
->so
->num_samp
+= glsl_type_get_sampler_count(var
->type
);
2723 /* just assume that we'll be reading from images.. if it
2724 * is write-only we don't have to count it, but not sure
2725 * if there is a good way to know?
2727 ctx
->so
->num_samp
+= glsl_type_get_image_count(var
->type
);
2730 /* NOTE: need to do something more clever when we support >1 fxn */
2731 nir_foreach_register(reg
, &fxn
->registers
) {
2732 ir3_declare_array(ctx
, reg
);
2734 /* And emit the body: */
2736 emit_function(ctx
, fxn
);
2739 /* from NIR perspective, we actually have varying inputs. But the varying
2740 * inputs, from an IR standpoint, are just bary.f/ldlv instructions. The
2741 * only actual inputs are the sysvals.
2744 fixup_frag_inputs(struct ir3_context
*ctx
)
2746 struct ir3_shader_variant
*so
= ctx
->so
;
2747 struct ir3
*ir
= ctx
->ir
;
2750 /* sysvals should appear at the end of the inputs, drop everything else: */
2751 while ((i
< so
->inputs_count
) && !so
->inputs
[i
].sysval
)
2754 /* at IR level, inputs are always blocks of 4 scalars: */
2757 ir
->inputs
= &ir
->inputs
[i
];
2761 /* Fixup tex sampler state for astc/srgb workaround instructions. We
2762 * need to assign the tex state indexes for these after we know the
2766 fixup_astc_srgb(struct ir3_context
*ctx
)
2768 struct ir3_shader_variant
*so
= ctx
->so
;
2769 /* indexed by original tex idx, value is newly assigned alpha sampler
2770 * state tex idx. Zero is invalid since there is at least one sampler
2773 unsigned alt_tex_state
[16] = {0};
2774 unsigned tex_idx
= ctx
->max_texture_index
+ 1;
2777 so
->astc_srgb
.base
= tex_idx
;
2779 for (unsigned i
= 0; i
< ctx
->ir
->astc_srgb_count
; i
++) {
2780 struct ir3_instruction
*sam
= ctx
->ir
->astc_srgb
[i
];
2782 compile_assert(ctx
, sam
->cat5
.tex
< ARRAY_SIZE(alt_tex_state
));
2784 if (alt_tex_state
[sam
->cat5
.tex
] == 0) {
2785 /* assign new alternate/alpha tex state slot: */
2786 alt_tex_state
[sam
->cat5
.tex
] = tex_idx
++;
2787 so
->astc_srgb
.orig_idx
[idx
++] = sam
->cat5
.tex
;
2788 so
->astc_srgb
.count
++;
2791 sam
->cat5
.tex
= alt_tex_state
[sam
->cat5
.tex
];
2796 fixup_binning_pass(struct ir3_context
*ctx
)
2798 struct ir3_shader_variant
*so
= ctx
->so
;
2799 struct ir3
*ir
= ctx
->ir
;
2802 for (i
= 0, j
= 0; i
< so
->outputs_count
; i
++) {
2803 unsigned slot
= so
->outputs
[i
].slot
;
2805 /* throw away everything but first position/psize */
2806 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
2808 so
->outputs
[j
] = so
->outputs
[i
];
2809 ir
->outputs
[(j
*4)+0] = ir
->outputs
[(i
*4)+0];
2810 ir
->outputs
[(j
*4)+1] = ir
->outputs
[(i
*4)+1];
2811 ir
->outputs
[(j
*4)+2] = ir
->outputs
[(i
*4)+2];
2812 ir
->outputs
[(j
*4)+3] = ir
->outputs
[(i
*4)+3];
2817 so
->outputs_count
= j
;
2818 ir
->noutputs
= j
* 4;
2822 ir3_compile_shader_nir(struct ir3_compiler
*compiler
,
2823 struct ir3_shader_variant
*so
)
2825 struct ir3_context
*ctx
;
2827 struct ir3_instruction
**inputs
;
2829 int ret
= 0, max_bary
;
2833 ctx
= ir3_context_init(compiler
, so
);
2835 DBG("INIT failed!");
2840 emit_instructions(ctx
);
2843 DBG("EMIT failed!");
2848 ir
= so
->ir
= ctx
->ir
;
2850 /* keep track of the inputs from TGSI perspective.. */
2851 inputs
= ir
->inputs
;
2853 /* but fixup actual inputs for frag shader: */
2854 if (so
->type
== MESA_SHADER_FRAGMENT
)
2855 fixup_frag_inputs(ctx
);
2857 /* at this point, for binning pass, throw away unneeded outputs: */
2858 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
< 600))
2859 fixup_binning_pass(ctx
);
2861 /* if we want half-precision outputs, mark the output registers
2864 if (so
->key
.half_precision
) {
2865 for (i
= 0; i
< ir
->noutputs
; i
++) {
2866 struct ir3_instruction
*out
= ir
->outputs
[i
];
2871 /* if frag shader writes z, that needs to be full precision: */
2872 if (so
->outputs
[i
/4].slot
== FRAG_RESULT_DEPTH
)
2875 out
->regs
[0]->flags
|= IR3_REG_HALF
;
2876 /* output could be a fanout (ie. texture fetch output)
2877 * in which case we need to propagate the half-reg flag
2878 * up to the definer so that RA sees it:
2880 if (out
->opc
== OPC_META_FO
) {
2881 out
= out
->regs
[1]->instr
;
2882 out
->regs
[0]->flags
|= IR3_REG_HALF
;
2885 if (out
->opc
== OPC_MOV
) {
2886 out
->cat1
.dst_type
= half_type(out
->cat1
.dst_type
);
2891 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2892 printf("BEFORE CP:\n");
2898 /* at this point, for binning pass, throw away unneeded outputs:
2899 * Note that for a6xx and later, we do this after ir3_cp to ensure
2900 * that the uniform/constant layout for BS and VS matches, so that
2901 * we can re-use same VS_CONST state group.
2903 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
>= 600))
2904 fixup_binning_pass(ctx
);
2906 /* for a6xx+, binning and draw pass VS use same VBO state, so we
2907 * need to make sure not to remove any inputs that are used by
2908 * the nonbinning VS.
2910 if (ctx
->compiler
->gpu_id
>= 600 && so
->binning_pass
) {
2911 debug_assert(so
->type
== MESA_SHADER_VERTEX
);
2912 for (int i
= 0; i
< ir
->ninputs
; i
++) {
2913 struct ir3_instruction
*in
= ir
->inputs
[i
];
2921 debug_assert(n
< so
->nonbinning
->inputs_count
);
2923 if (so
->nonbinning
->inputs
[n
].sysval
)
2926 /* be sure to keep inputs, even if only used in VS */
2927 if (so
->nonbinning
->inputs
[n
].compmask
& (1 << c
))
2928 array_insert(in
->block
, in
->block
->keeps
, in
);
2932 /* Insert mov if there's same instruction for each output.
2933 * eg. dEQP-GLES31.functional.shaders.opaque_type_indexing.sampler.const_expression.vertex.sampler2dshadow
2935 for (int i
= ir
->noutputs
- 1; i
>= 0; i
--) {
2936 if (!ir
->outputs
[i
])
2938 for (unsigned j
= 0; j
< i
; j
++) {
2939 if (ir
->outputs
[i
] == ir
->outputs
[j
]) {
2941 ir3_MOV(ir
->outputs
[i
]->block
, ir
->outputs
[i
], TYPE_F32
);
2946 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2947 printf("BEFORE GROUPING:\n");
2951 ir3_sched_add_deps(ir
);
2953 /* Group left/right neighbors, inserting mov's where needed to
2958 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2959 printf("AFTER GROUPING:\n");
2965 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2966 printf("AFTER DEPTH:\n");
2970 /* do Sethi–Ullman numbering before scheduling: */
2973 ret
= ir3_sched(ir
);
2975 DBG("SCHED failed!");
2979 if (compiler
->gpu_id
>= 600) {
2980 ir3_a6xx_fixup_atomic_dests(ir
, so
);
2983 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2984 printf("AFTER SCHED:\n");
2994 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
2995 printf("AFTER RA:\n");
2999 if (so
->type
== MESA_SHADER_FRAGMENT
)
3002 /* fixup input/outputs: */
3003 for (i
= 0; i
< so
->outputs_count
; i
++) {
3004 /* sometimes we get outputs that don't write the .x coord, like:
3006 * decl_var shader_out INTERP_MODE_NONE float Color (VARYING_SLOT_VAR9.z, 1, 0)
3008 * Presumably the result of varying packing and then eliminating
3009 * some unneeded varyings? Just skip head to the first valid
3010 * component of the output.
3012 for (unsigned j
= 0; j
< 4; j
++) {
3013 struct ir3_instruction
*instr
= ir
->outputs
[(i
*4) + j
];
3015 so
->outputs
[i
].regid
= instr
->regs
[0]->num
;
3016 so
->outputs
[i
].half
= !!(instr
->regs
[0]->flags
& IR3_REG_HALF
);
3022 /* Note that some or all channels of an input may be unused: */
3023 for (i
= 0; i
< so
->inputs_count
; i
++) {
3024 unsigned j
, reg
= regid(63,0);
3026 for (j
= 0; j
< 4; j
++) {
3027 struct ir3_instruction
*in
= inputs
[(i
*4) + j
];
3032 if (in
->flags
& IR3_INSTR_UNUSED
)
3035 reg
= in
->regs
[0]->num
- j
;
3037 compile_assert(ctx
, in
->regs
[0]->flags
& IR3_REG_HALF
);
3039 half
= !!(in
->regs
[0]->flags
& IR3_REG_HALF
);
3042 so
->inputs
[i
].regid
= reg
;
3043 so
->inputs
[i
].half
= half
;
3047 fixup_astc_srgb(ctx
);
3049 /* We need to do legalize after (for frag shader's) the "bary.f"
3050 * offsets (inloc) have been assigned.
3052 ir3_legalize(ir
, &so
->has_ssbo
, &so
->need_pixlod
, &max_bary
);
3054 if (ir3_shader_debug
& IR3_DBG_OPTMSGS
) {
3055 printf("AFTER LEGALIZE:\n");
3059 so
->branchstack
= ctx
->max_stack
;
3061 /* Note that actual_in counts inputs that are not bary.f'd for FS: */
3062 if (so
->type
== MESA_SHADER_FRAGMENT
)
3063 so
->total_in
= max_bary
+ 1;
3065 so
->max_sun
= ir
->max_sun
;
3070 ir3_destroy(so
->ir
);
3073 ir3_context_free(ctx
);