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
;
55 src
= __ssa_src(mov
, collect
, IR3_REG_RELATIV
);
57 src
->array
.offset
= n
;
59 ir3_instr_set_address(mov
, address
);
64 static struct ir3_instruction
*
65 create_input(struct ir3_context
*ctx
, unsigned compmask
)
67 struct ir3_instruction
*in
;
69 in
= ir3_instr_create(ctx
->in_block
, OPC_META_INPUT
);
70 in
->input
.sysval
= ~0;
71 __ssa_dst(in
)->wrmask
= compmask
;
73 array_insert(ctx
->ir
, ctx
->ir
->inputs
, in
);
78 static struct ir3_instruction
*
79 create_frag_input(struct ir3_context
*ctx
, bool use_ldlv
, unsigned n
)
81 struct ir3_block
*block
= ctx
->block
;
82 struct ir3_instruction
*instr
;
83 /* packed inloc is fixed up later: */
84 struct ir3_instruction
*inloc
= create_immed(block
, n
);
87 instr
= ir3_LDLV(block
, inloc
, 0, create_immed(block
, 1), 0);
88 instr
->cat6
.type
= TYPE_U32
;
89 instr
->cat6
.iim_val
= 1;
91 instr
= ir3_BARY_F(block
, inloc
, 0, ctx
->ij_pixel
, 0);
92 instr
->regs
[2]->wrmask
= 0x3;
98 static struct ir3_instruction
*
99 create_driver_param(struct ir3_context
*ctx
, enum ir3_driver_param dp
)
101 /* first four vec4 sysval's reserved for UBOs: */
102 /* NOTE: dp is in scalar, but there can be >4 dp components: */
103 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
104 unsigned n
= const_state
->offsets
.driver_param
;
105 unsigned r
= regid(n
+ dp
/ 4, dp
% 4);
106 return create_uniform(ctx
->block
, r
);
110 * Adreno's comparisons produce a 1 for true and 0 for false, in either 16 or
111 * 32-bit registers. We use NIR's 1-bit integers to represent bools, and
112 * trust that we will only see and/or/xor on those 1-bit values, so we can
113 * safely store NIR i1s in a 32-bit reg while always containing either a 1 or
118 * alu/sfu instructions:
121 static struct ir3_instruction
*
122 create_cov(struct ir3_context
*ctx
, struct ir3_instruction
*src
,
123 unsigned src_bitsize
, nir_op op
)
125 type_t src_type
, dst_type
;
129 case nir_op_f2f16_rtne
:
130 case nir_op_f2f16_rtz
:
138 switch (src_bitsize
) {
146 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
155 switch (src_bitsize
) {
166 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
175 switch (src_bitsize
) {
186 ir3_context_error(ctx
, "invalid src bit size: %u", src_bitsize
);
199 ir3_context_error(ctx
, "invalid conversion op: %u", op
);
210 case nir_op_f2f16_rtne
:
211 case nir_op_f2f16_rtz
:
253 ir3_context_error(ctx
, "invalid conversion op: %u", op
);
256 if (src_type
== dst_type
)
259 struct ir3_instruction
*cov
=
260 ir3_COV(ctx
->block
, src
, src_type
, dst_type
);
262 if (op
== nir_op_f2f16_rtne
)
263 cov
->regs
[0]->flags
|= IR3_REG_EVEN
;
269 emit_alu(struct ir3_context
*ctx
, nir_alu_instr
*alu
)
271 const nir_op_info
*info
= &nir_op_infos
[alu
->op
];
272 struct ir3_instruction
**dst
, *src
[info
->num_inputs
];
273 unsigned bs
[info
->num_inputs
]; /* bit size */
274 struct ir3_block
*b
= ctx
->block
;
275 unsigned dst_sz
, wrmask
;
276 type_t dst_type
= nir_dest_bit_size(alu
->dest
.dest
) == 16 ?
279 if (alu
->dest
.dest
.is_ssa
) {
280 dst_sz
= alu
->dest
.dest
.ssa
.num_components
;
281 wrmask
= (1 << dst_sz
) - 1;
283 dst_sz
= alu
->dest
.dest
.reg
.reg
->num_components
;
284 wrmask
= alu
->dest
.write_mask
;
287 dst
= ir3_get_dst(ctx
, &alu
->dest
.dest
, dst_sz
);
289 /* Vectors are special in that they have non-scalarized writemasks,
290 * and just take the first swizzle channel for each argument in
291 * order into each writemask channel.
293 if ((alu
->op
== nir_op_vec2
) ||
294 (alu
->op
== nir_op_vec3
) ||
295 (alu
->op
== nir_op_vec4
)) {
297 for (int i
= 0; i
< info
->num_inputs
; i
++) {
298 nir_alu_src
*asrc
= &alu
->src
[i
];
300 compile_assert(ctx
, !asrc
->abs
);
301 compile_assert(ctx
, !asrc
->negate
);
303 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[0]];
305 src
[i
] = create_immed_typed(ctx
->block
, 0, dst_type
);
306 dst
[i
] = ir3_MOV(b
, src
[i
], dst_type
);
309 ir3_put_dst(ctx
, &alu
->dest
.dest
);
313 /* We also get mov's with more than one component for mov's so
314 * handle those specially:
316 if (alu
->op
== nir_op_mov
) {
317 nir_alu_src
*asrc
= &alu
->src
[0];
318 struct ir3_instruction
*const *src0
= ir3_get_src(ctx
, &asrc
->src
);
320 for (unsigned i
= 0; i
< dst_sz
; i
++) {
321 if (wrmask
& (1 << i
)) {
322 dst
[i
] = ir3_MOV(b
, src0
[asrc
->swizzle
[i
]], dst_type
);
328 ir3_put_dst(ctx
, &alu
->dest
.dest
);
332 /* General case: We can just grab the one used channel per src. */
333 for (int i
= 0; i
< info
->num_inputs
; i
++) {
334 unsigned chan
= ffs(alu
->dest
.write_mask
) - 1;
335 nir_alu_src
*asrc
= &alu
->src
[i
];
337 compile_assert(ctx
, !asrc
->abs
);
338 compile_assert(ctx
, !asrc
->negate
);
340 src
[i
] = ir3_get_src(ctx
, &asrc
->src
)[asrc
->swizzle
[chan
]];
341 bs
[i
] = nir_src_bit_size(asrc
->src
);
343 compile_assert(ctx
, src
[i
]);
348 case nir_op_f2f16_rtne
:
349 case nir_op_f2f16_rtz
:
372 dst
[0] = create_cov(ctx
, src
[0], bs
[0], alu
->op
);
375 case nir_op_fquantize2f16
:
376 dst
[0] = create_cov(ctx
,
377 create_cov(ctx
, src
[0], 32, nir_op_f2f16
),
381 dst
[0] = ir3_CMPS_F(b
,
383 create_immed_typed(b
, 0, bs
[0] == 16 ? TYPE_F16
: TYPE_F32
), 0);
384 dst
[0]->cat2
.condition
= IR3_COND_NE
;
388 /* i2b1 will appear when translating from nir_load_ubo or
389 * nir_intrinsic_load_ssbo, where any non-zero value is true.
391 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, create_immed(b
, 0), 0);
392 dst
[0]->cat2
.condition
= IR3_COND_NE
;
396 /* b2b1 will appear when translating from
398 * - nir_intrinsic_load_shared of a 32-bit 0/~0 value.
399 * - nir_intrinsic_load_constant of a 32-bit 0/~0 value
401 * A negate can turn those into a 1 or 0 for us.
403 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SNEG
);
407 /* b2b32 will appear when converting our 1-bit bools to a store_shared
410 * A negate can turn those into a ~0 for us.
412 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SNEG
);
416 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FNEG
);
419 dst
[0] = ir3_ABSNEG_F(b
, src
[0], IR3_REG_FABS
);
422 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[1], 0);
425 dst
[0] = ir3_MIN_F(b
, src
[0], 0, src
[1], 0);
428 /* if there is just a single use of the src, and it supports
429 * (sat) bit, we can just fold the (sat) flag back to the
430 * src instruction and create a mov. This is easier for cp
433 * TODO probably opc_cat==4 is ok too
435 if (alu
->src
[0].src
.is_ssa
&&
436 (list_length(&alu
->src
[0].src
.ssa
->uses
) == 1) &&
437 ((opc_cat(src
[0]->opc
) == 2) || (opc_cat(src
[0]->opc
) == 3))) {
438 src
[0]->flags
|= IR3_INSTR_SAT
;
439 dst
[0] = ir3_MOV(b
, src
[0], dst_type
);
441 /* otherwise generate a max.f that saturates.. blob does
442 * similar (generating a cat2 mov using max.f)
444 dst
[0] = ir3_MAX_F(b
, src
[0], 0, src
[0], 0);
445 dst
[0]->flags
|= IR3_INSTR_SAT
;
449 dst
[0] = ir3_MUL_F(b
, src
[0], 0, src
[1], 0);
452 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], 0);
455 dst
[0] = ir3_ADD_F(b
, src
[0], 0, src
[1], IR3_REG_FNEG
);
458 dst
[0] = ir3_MAD_F32(b
, src
[0], 0, src
[1], 0, src
[2], 0);
461 case nir_op_fddx_coarse
:
462 dst
[0] = ir3_DSX(b
, src
[0], 0);
463 dst
[0]->cat5
.type
= TYPE_F32
;
465 case nir_op_fddx_fine
:
466 dst
[0] = ir3_DSXPP_1(b
, src
[0], 0);
467 dst
[0]->cat5
.type
= TYPE_F32
;
470 case nir_op_fddy_coarse
:
471 dst
[0] = ir3_DSY(b
, src
[0], 0);
472 dst
[0]->cat5
.type
= TYPE_F32
;
475 case nir_op_fddy_fine
:
476 dst
[0] = ir3_DSYPP_1(b
, src
[0], 0);
477 dst
[0]->cat5
.type
= TYPE_F32
;
480 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
481 dst
[0]->cat2
.condition
= IR3_COND_LT
;
484 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
485 dst
[0]->cat2
.condition
= IR3_COND_GE
;
488 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
489 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
492 dst
[0] = ir3_CMPS_F(b
, src
[0], 0, src
[1], 0);
493 dst
[0]->cat2
.condition
= IR3_COND_NE
;
496 dst
[0] = ir3_CEIL_F(b
, src
[0], 0);
499 dst
[0] = ir3_FLOOR_F(b
, src
[0], 0);
502 dst
[0] = ir3_TRUNC_F(b
, src
[0], 0);
504 case nir_op_fround_even
:
505 dst
[0] = ir3_RNDNE_F(b
, src
[0], 0);
508 dst
[0] = ir3_SIGN_F(b
, src
[0], 0);
512 dst
[0] = ir3_SIN(b
, src
[0], 0);
515 dst
[0] = ir3_COS(b
, src
[0], 0);
518 dst
[0] = ir3_RSQ(b
, src
[0], 0);
521 dst
[0] = ir3_RCP(b
, src
[0], 0);
524 dst
[0] = ir3_LOG2(b
, src
[0], 0);
527 dst
[0] = ir3_EXP2(b
, src
[0], 0);
530 dst
[0] = ir3_SQRT(b
, src
[0], 0);
534 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SABS
);
537 dst
[0] = ir3_ADD_U(b
, src
[0], 0, src
[1], 0);
540 dst
[0] = ir3_AND_B(b
, src
[0], 0, src
[1], 0);
543 dst
[0] = ir3_MAX_S(b
, src
[0], 0, src
[1], 0);
546 dst
[0] = ir3_MAX_U(b
, src
[0], 0, src
[1], 0);
549 dst
[0] = ir3_MIN_S(b
, src
[0], 0, src
[1], 0);
552 dst
[0] = ir3_MIN_U(b
, src
[0], 0, src
[1], 0);
554 case nir_op_umul_low
:
555 dst
[0] = ir3_MULL_U(b
, src
[0], 0, src
[1], 0);
557 case nir_op_imadsh_mix16
:
558 dst
[0] = ir3_MADSH_M16(b
, src
[0], 0, src
[1], 0, src
[2], 0);
560 case nir_op_imad24_ir3
:
561 dst
[0] = ir3_MAD_S24(b
, src
[0], 0, src
[1], 0, src
[2], 0);
564 dst
[0] = ir3_MUL_S24(b
, src
[0], 0, src
[1], 0);
567 dst
[0] = ir3_ABSNEG_S(b
, src
[0], IR3_REG_SNEG
);
571 dst
[0] = ir3_SUB_U(b
, create_immed(ctx
->block
, 1), 0, src
[0], 0);
573 dst
[0] = ir3_NOT_B(b
, src
[0], 0);
577 dst
[0] = ir3_OR_B(b
, src
[0], 0, src
[1], 0);
580 dst
[0] = ir3_SHL_B(b
, src
[0], 0, src
[1], 0);
583 dst
[0] = ir3_ASHR_B(b
, src
[0], 0, src
[1], 0);
586 dst
[0] = ir3_SUB_U(b
, src
[0], 0, src
[1], 0);
589 dst
[0] = ir3_XOR_B(b
, src
[0], 0, src
[1], 0);
592 dst
[0] = ir3_SHR_B(b
, src
[0], 0, src
[1], 0);
595 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
596 dst
[0]->cat2
.condition
= IR3_COND_LT
;
599 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
600 dst
[0]->cat2
.condition
= IR3_COND_GE
;
603 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
604 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
607 dst
[0] = ir3_CMPS_S(b
, src
[0], 0, src
[1], 0);
608 dst
[0]->cat2
.condition
= IR3_COND_NE
;
611 dst
[0] = ir3_CMPS_U(b
, src
[0], 0, src
[1], 0);
612 dst
[0]->cat2
.condition
= IR3_COND_LT
;
615 dst
[0] = ir3_CMPS_U(b
, src
[0], 0, src
[1], 0);
616 dst
[0]->cat2
.condition
= IR3_COND_GE
;
620 struct ir3_instruction
*cond
= src
[0];
622 /* If src[0] is a negation (likely as a result of an ir3_b2n(cond)),
623 * we can ignore that and use original cond, since the nonzero-ness of
624 * cond stays the same.
626 if (cond
->opc
== OPC_ABSNEG_S
&&
628 (cond
->regs
[1]->flags
& (IR3_REG_SNEG
| IR3_REG_SABS
)) == IR3_REG_SNEG
) {
629 cond
= cond
->regs
[1]->instr
;
632 compile_assert(ctx
, bs
[1] == bs
[2]);
633 /* The condition's size has to match the other two arguments' size, so
634 * convert down if necessary.
637 struct hash_entry
*prev_entry
=
638 _mesa_hash_table_search(ctx
->sel_cond_conversions
, src
[0]);
640 cond
= prev_entry
->data
;
642 cond
= ir3_COV(b
, cond
, TYPE_U32
, TYPE_U16
);
643 _mesa_hash_table_insert(ctx
->sel_cond_conversions
, src
[0], cond
);
648 dst
[0] = ir3_SEL_B32(b
, src
[1], 0, cond
, 0, src
[2], 0);
650 dst
[0] = ir3_SEL_B16(b
, src
[1], 0, cond
, 0, src
[2], 0);
653 case nir_op_bit_count
: {
654 // TODO, we need to do this 16b at a time on a5xx+a6xx.. need to
655 // double check on earlier gen's. Once half-precision support is
656 // in place, this should probably move to a NIR lowering pass:
657 struct ir3_instruction
*hi
, *lo
;
659 hi
= ir3_COV(b
, ir3_SHR_B(b
, src
[0], 0, create_immed(b
, 16), 0),
661 lo
= ir3_COV(b
, src
[0], TYPE_U32
, TYPE_U16
);
663 hi
= ir3_CBITS_B(b
, hi
, 0);
664 lo
= ir3_CBITS_B(b
, lo
, 0);
666 // TODO maybe the builders should default to making dst half-precision
667 // if the src's were half precision, to make this less awkward.. otoh
668 // we should probably just do this lowering in NIR.
669 hi
->regs
[0]->flags
|= IR3_REG_HALF
;
670 lo
->regs
[0]->flags
|= IR3_REG_HALF
;
672 dst
[0] = ir3_ADD_S(b
, hi
, 0, lo
, 0);
673 dst
[0]->regs
[0]->flags
|= IR3_REG_HALF
;
674 dst
[0] = ir3_COV(b
, dst
[0], TYPE_U16
, TYPE_U32
);
677 case nir_op_ifind_msb
: {
678 struct ir3_instruction
*cmp
;
679 dst
[0] = ir3_CLZ_S(b
, src
[0], 0);
680 cmp
= ir3_CMPS_S(b
, dst
[0], 0, create_immed(b
, 0), 0);
681 cmp
->cat2
.condition
= IR3_COND_GE
;
682 dst
[0] = ir3_SEL_B32(b
,
683 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
687 case nir_op_ufind_msb
:
688 dst
[0] = ir3_CLZ_B(b
, src
[0], 0);
689 dst
[0] = ir3_SEL_B32(b
,
690 ir3_SUB_U(b
, create_immed(b
, 31), 0, dst
[0], 0), 0,
691 src
[0], 0, dst
[0], 0);
693 case nir_op_find_lsb
:
694 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
695 dst
[0] = ir3_CLZ_B(b
, dst
[0], 0);
697 case nir_op_bitfield_reverse
:
698 dst
[0] = ir3_BFREV_B(b
, src
[0], 0);
702 ir3_context_error(ctx
, "Unhandled ALU op: %s\n",
703 nir_op_infos
[alu
->op
].name
);
707 if (nir_alu_type_get_base_type(info
->output_type
) == nir_type_bool
) {
708 assert(nir_dest_bit_size(alu
->dest
.dest
) == 1 ||
709 alu
->op
== nir_op_b2b32
);
712 /* 1-bit values stored in 32-bit registers are only valid for certain
723 compile_assert(ctx
, nir_dest_bit_size(alu
->dest
.dest
) != 1);
727 ir3_put_dst(ctx
, &alu
->dest
.dest
);
731 emit_intrinsic_load_ubo_ldc(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
732 struct ir3_instruction
**dst
)
734 struct ir3_block
*b
= ctx
->block
;
736 unsigned ncomp
= intr
->num_components
;
737 struct ir3_instruction
*offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
738 struct ir3_instruction
*idx
= ir3_get_src(ctx
, &intr
->src
[0])[0];
739 struct ir3_instruction
*ldc
= ir3_LDC(b
, idx
, 0, offset
, 0);
740 ldc
->regs
[0]->wrmask
= MASK(ncomp
);
741 ldc
->cat6
.iim_val
= ncomp
;
742 ldc
->cat6
.d
= nir_intrinsic_base(intr
);
743 ldc
->cat6
.type
= TYPE_U32
;
745 nir_intrinsic_instr
*bindless
= ir3_bindless_resource(intr
->src
[0]);
747 ldc
->flags
|= IR3_INSTR_B
;
748 ldc
->cat6
.base
= nir_intrinsic_desc_set(bindless
);
749 ctx
->so
->bindless_ubo
= true;
752 ir3_split_dest(b
, dst
, ldc
, 0, ncomp
);
756 /* handles direct/indirect UBO reads: */
758 emit_intrinsic_load_ubo(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
759 struct ir3_instruction
**dst
)
761 struct ir3_block
*b
= ctx
->block
;
762 struct ir3_instruction
*base_lo
, *base_hi
, *addr
, *src0
, *src1
;
763 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
764 unsigned ubo
= regid(const_state
->offsets
.ubo
, 0);
765 const unsigned ptrsz
= ir3_pointer_size(ctx
->compiler
);
769 /* First src is ubo index, which could either be an immed or not: */
770 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0];
771 if (is_same_type_mov(src0
) &&
772 (src0
->regs
[1]->flags
& IR3_REG_IMMED
)) {
773 base_lo
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
));
774 base_hi
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
) + 1);
776 base_lo
= create_uniform_indirect(b
, ubo
, ir3_get_addr0(ctx
, src0
, ptrsz
));
777 base_hi
= create_uniform_indirect(b
, ubo
+ 1, ir3_get_addr0(ctx
, src0
, ptrsz
));
779 /* NOTE: since relative addressing is used, make sure constlen is
780 * at least big enough to cover all the UBO addresses, since the
781 * assembler won't know what the max address reg is.
783 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
784 const_state
->offsets
.ubo
+ (ctx
->s
->info
.num_ubos
* ptrsz
));
787 /* note: on 32bit gpu's base_hi is ignored and DCE'd */
790 if (nir_src_is_const(intr
->src
[1])) {
791 off
+= nir_src_as_uint(intr
->src
[1]);
793 /* For load_ubo_indirect, second src is indirect offset: */
794 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0];
796 /* and add offset to addr: */
797 addr
= ir3_ADD_S(b
, addr
, 0, src1
, 0);
800 /* if offset is to large to encode in the ldg, split it out: */
801 if ((off
+ (intr
->num_components
* 4)) > 1024) {
802 /* split out the minimal amount to improve the odds that
803 * cp can fit the immediate in the add.s instruction:
805 unsigned off2
= off
+ (intr
->num_components
* 4) - 1024;
806 addr
= ir3_ADD_S(b
, addr
, 0, create_immed(b
, off2
), 0);
811 struct ir3_instruction
*carry
;
813 /* handle 32b rollover, ie:
814 * if (addr < base_lo)
817 carry
= ir3_CMPS_U(b
, addr
, 0, base_lo
, 0);
818 carry
->cat2
.condition
= IR3_COND_LT
;
819 base_hi
= ir3_ADD_S(b
, base_hi
, 0, carry
, 0);
821 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){ addr
, base_hi
}, 2);
824 for (int i
= 0; i
< intr
->num_components
; i
++) {
825 struct ir3_instruction
*load
=
826 ir3_LDG(b
, addr
, 0, create_immed(b
, 1), 0, /* num components */
827 create_immed(b
, off
+ i
* 4), 0);
828 load
->cat6
.type
= TYPE_U32
;
833 /* src[] = { block_index } */
835 emit_intrinsic_ssbo_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
836 struct ir3_instruction
**dst
)
838 /* SSBO size stored as a const starting at ssbo_sizes: */
839 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
840 unsigned blk_idx
= nir_src_as_uint(intr
->src
[0]);
841 unsigned idx
= regid(const_state
->offsets
.ssbo_sizes
, 0) +
842 const_state
->ssbo_size
.off
[blk_idx
];
844 debug_assert(const_state
->ssbo_size
.mask
& (1 << blk_idx
));
846 dst
[0] = create_uniform(ctx
->block
, idx
);
849 /* src[] = { offset }. const_index[] = { base } */
851 emit_intrinsic_load_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
852 struct ir3_instruction
**dst
)
854 struct ir3_block
*b
= ctx
->block
;
855 struct ir3_instruction
*ldl
, *offset
;
858 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
859 base
= nir_intrinsic_base(intr
);
861 ldl
= ir3_LDL(b
, offset
, 0,
862 create_immed(b
, intr
->num_components
), 0,
863 create_immed(b
, base
), 0);
865 ldl
->cat6
.type
= utype_dst(intr
->dest
);
866 ldl
->regs
[0]->wrmask
= MASK(intr
->num_components
);
868 ldl
->barrier_class
= IR3_BARRIER_SHARED_R
;
869 ldl
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
871 ir3_split_dest(b
, dst
, ldl
, 0, intr
->num_components
);
874 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
876 emit_intrinsic_store_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
878 struct ir3_block
*b
= ctx
->block
;
879 struct ir3_instruction
*stl
, *offset
;
880 struct ir3_instruction
* const *value
;
881 unsigned base
, wrmask
, ncomp
;
883 value
= ir3_get_src(ctx
, &intr
->src
[0]);
884 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
886 base
= nir_intrinsic_base(intr
);
887 wrmask
= nir_intrinsic_write_mask(intr
);
888 ncomp
= ffs(~wrmask
) - 1;
890 assert(wrmask
== BITFIELD_MASK(intr
->num_components
));
892 stl
= ir3_STL(b
, offset
, 0,
893 ir3_create_collect(ctx
, value
, ncomp
), 0,
894 create_immed(b
, ncomp
), 0);
895 stl
->cat6
.dst_offset
= base
;
896 stl
->cat6
.type
= utype_src(intr
->src
[0]);
897 stl
->barrier_class
= IR3_BARRIER_SHARED_W
;
898 stl
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
900 array_insert(b
, b
->keeps
, stl
);
903 /* src[] = { offset }. const_index[] = { base } */
905 emit_intrinsic_load_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
906 struct ir3_instruction
**dst
)
908 struct ir3_block
*b
= ctx
->block
;
909 struct ir3_instruction
*load
, *offset
;
912 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
913 base
= nir_intrinsic_base(intr
);
915 load
= ir3_LDLW(b
, offset
, 0,
916 create_immed(b
, intr
->num_components
), 0,
917 create_immed(b
, base
), 0);
919 load
->cat6
.type
= utype_dst(intr
->dest
);
920 load
->regs
[0]->wrmask
= MASK(intr
->num_components
);
922 load
->barrier_class
= IR3_BARRIER_SHARED_R
;
923 load
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
925 ir3_split_dest(b
, dst
, load
, 0, intr
->num_components
);
928 /* src[] = { value, offset }. const_index[] = { base } */
930 emit_intrinsic_store_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
932 struct ir3_block
*b
= ctx
->block
;
933 struct ir3_instruction
*store
, *offset
;
934 struct ir3_instruction
* const *value
;
936 value
= ir3_get_src(ctx
, &intr
->src
[0]);
937 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
939 store
= ir3_STLW(b
, offset
, 0,
940 ir3_create_collect(ctx
, value
, intr
->num_components
), 0,
941 create_immed(b
, intr
->num_components
), 0);
943 store
->cat6
.dst_offset
= nir_intrinsic_base(intr
);
944 store
->cat6
.type
= utype_src(intr
->src
[0]);
945 store
->barrier_class
= IR3_BARRIER_SHARED_W
;
946 store
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
948 array_insert(b
, b
->keeps
, store
);
952 * CS shared variable atomic intrinsics
954 * All of the shared variable atomic memory operations read a value from
955 * memory, compute a new value using one of the operations below, write the
956 * new value to memory, and return the original value read.
958 * All operations take 2 sources except CompSwap that takes 3. These
961 * 0: The offset into the shared variable storage region that the atomic
962 * operation will operate on.
963 * 1: The data parameter to the atomic function (i.e. the value to add
964 * in shared_atomic_add, etc).
965 * 2: For CompSwap only: the second data parameter.
967 static struct ir3_instruction
*
968 emit_intrinsic_atomic_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
970 struct ir3_block
*b
= ctx
->block
;
971 struct ir3_instruction
*atomic
, *src0
, *src1
;
972 type_t type
= TYPE_U32
;
974 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0]; /* offset */
975 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0]; /* value */
977 switch (intr
->intrinsic
) {
978 case nir_intrinsic_shared_atomic_add
:
979 atomic
= ir3_ATOMIC_ADD(b
, src0
, 0, src1
, 0);
981 case nir_intrinsic_shared_atomic_imin
:
982 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
985 case nir_intrinsic_shared_atomic_umin
:
986 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
988 case nir_intrinsic_shared_atomic_imax
:
989 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
992 case nir_intrinsic_shared_atomic_umax
:
993 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
995 case nir_intrinsic_shared_atomic_and
:
996 atomic
= ir3_ATOMIC_AND(b
, src0
, 0, src1
, 0);
998 case nir_intrinsic_shared_atomic_or
:
999 atomic
= ir3_ATOMIC_OR(b
, src0
, 0, src1
, 0);
1001 case nir_intrinsic_shared_atomic_xor
:
1002 atomic
= ir3_ATOMIC_XOR(b
, src0
, 0, src1
, 0);
1004 case nir_intrinsic_shared_atomic_exchange
:
1005 atomic
= ir3_ATOMIC_XCHG(b
, src0
, 0, src1
, 0);
1007 case nir_intrinsic_shared_atomic_comp_swap
:
1008 /* for cmpxchg, src1 is [ui]vec2(data, compare): */
1009 src1
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1010 ir3_get_src(ctx
, &intr
->src
[2])[0],
1013 atomic
= ir3_ATOMIC_CMPXCHG(b
, src0
, 0, src1
, 0);
1019 atomic
->cat6
.iim_val
= 1;
1021 atomic
->cat6
.type
= type
;
1022 atomic
->barrier_class
= IR3_BARRIER_SHARED_W
;
1023 atomic
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
1025 /* even if nothing consume the result, we can't DCE the instruction: */
1026 array_insert(b
, b
->keeps
, atomic
);
1031 struct tex_src_info
{
1033 unsigned tex_base
, samp_base
, tex_idx
, samp_idx
;
1034 /* For normal tex instructions */
1035 unsigned base
, combined_idx
, a1_val
, flags
;
1036 struct ir3_instruction
*samp_tex
;
1039 /* TODO handle actual indirect/dynamic case.. which is going to be weird
1040 * to handle with the image_mapping table..
1042 static struct tex_src_info
1043 get_image_samp_tex_src(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1045 struct ir3_block
*b
= ctx
->block
;
1046 struct tex_src_info info
= { 0 };
1047 nir_intrinsic_instr
*bindless_tex
= ir3_bindless_resource(intr
->src
[0]);
1048 ctx
->so
->bindless_tex
= true;
1052 info
.flags
|= IR3_INSTR_B
;
1054 /* Gather information required to determine which encoding to
1055 * choose as well as for prefetch.
1057 info
.tex_base
= nir_intrinsic_desc_set(bindless_tex
);
1058 bool tex_const
= nir_src_is_const(bindless_tex
->src
[0]);
1060 info
.tex_idx
= nir_src_as_uint(bindless_tex
->src
[0]);
1063 /* Choose encoding. */
1064 if (tex_const
&& info
.tex_idx
< 256) {
1065 if (info
.tex_idx
< 16) {
1066 /* Everything fits within the instruction */
1067 info
.base
= info
.tex_base
;
1068 info
.combined_idx
= info
.samp_idx
| (info
.tex_idx
<< 4);
1070 info
.base
= info
.tex_base
;
1071 info
.a1_val
= info
.tex_idx
<< 3;
1072 info
.combined_idx
= 0;
1073 info
.flags
|= IR3_INSTR_A1EN
;
1075 info
.samp_tex
= NULL
;
1077 info
.flags
|= IR3_INSTR_S2EN
;
1078 info
.base
= info
.tex_base
;
1080 /* Note: the indirect source is now a vec2 instead of hvec2 */
1081 struct ir3_instruction
*texture
, *sampler
;
1083 texture
= ir3_get_src(ctx
, &intr
->src
[0])[0];
1084 sampler
= create_immed(b
, 0);
1085 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1091 info
.flags
|= IR3_INSTR_S2EN
;
1092 unsigned slot
= nir_src_as_uint(intr
->src
[0]);
1093 unsigned tex_idx
= ir3_image_to_tex(&ctx
->so
->image_mapping
, slot
);
1094 struct ir3_instruction
*texture
, *sampler
;
1096 texture
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1097 sampler
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1099 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1108 static struct ir3_instruction
*
1109 emit_sam(struct ir3_context
*ctx
, opc_t opc
, struct tex_src_info info
,
1110 type_t type
, unsigned wrmask
, struct ir3_instruction
*src0
,
1111 struct ir3_instruction
*src1
)
1113 struct ir3_instruction
*sam
, *addr
;
1114 if (info
.flags
& IR3_INSTR_A1EN
) {
1115 addr
= ir3_get_addr1(ctx
, info
.a1_val
);
1117 sam
= ir3_SAM(ctx
->block
, opc
, type
, 0b1111, info
.flags
,
1118 info
.samp_tex
, src0
, src1
);
1119 if (info
.flags
& IR3_INSTR_A1EN
) {
1120 ir3_instr_set_address(sam
, addr
);
1122 if (info
.flags
& IR3_INSTR_B
) {
1123 sam
->cat5
.tex_base
= info
.base
;
1124 sam
->cat5
.samp
= info
.combined_idx
;
1129 /* src[] = { deref, coord, sample_index }. const_index[] = {} */
1131 emit_intrinsic_load_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1132 struct ir3_instruction
**dst
)
1134 struct ir3_block
*b
= ctx
->block
;
1135 struct tex_src_info info
= get_image_samp_tex_src(ctx
, intr
);
1136 struct ir3_instruction
*sam
;
1137 struct ir3_instruction
* const *src0
= ir3_get_src(ctx
, &intr
->src
[1]);
1138 struct ir3_instruction
*coords
[4];
1139 unsigned flags
, ncoords
= ir3_get_image_coords(intr
, &flags
);
1140 type_t type
= ir3_get_type_for_image_intrinsic(intr
);
1142 /* hmm, this seems a bit odd, but it is what blob does and (at least
1143 * a5xx) just faults on bogus addresses otherwise:
1145 if (flags
& IR3_INSTR_3D
) {
1146 flags
&= ~IR3_INSTR_3D
;
1147 flags
|= IR3_INSTR_A
;
1149 info
.flags
|= flags
;
1151 for (unsigned i
= 0; i
< ncoords
; i
++)
1152 coords
[i
] = src0
[i
];
1155 coords
[ncoords
++] = create_immed(b
, 0);
1157 sam
= emit_sam(ctx
, OPC_ISAM
, info
, type
, 0b1111,
1158 ir3_create_collect(ctx
, coords
, ncoords
), NULL
);
1160 sam
->barrier_class
= IR3_BARRIER_IMAGE_R
;
1161 sam
->barrier_conflict
= IR3_BARRIER_IMAGE_W
;
1163 ir3_split_dest(b
, dst
, sam
, 0, 4);
1167 emit_intrinsic_image_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1168 struct ir3_instruction
**dst
)
1170 struct ir3_block
*b
= ctx
->block
;
1171 struct tex_src_info info
= get_image_samp_tex_src(ctx
, intr
);
1172 struct ir3_instruction
*sam
, *lod
;
1173 unsigned flags
, ncoords
= ir3_get_image_coords(intr
, &flags
);
1174 type_t dst_type
= nir_dest_bit_size(intr
->dest
) == 16 ?
1175 TYPE_U16
: TYPE_U32
;
1177 info
.flags
|= flags
;
1178 lod
= create_immed(b
, 0);
1179 sam
= emit_sam(ctx
, OPC_GETSIZE
, info
, dst_type
, 0b1111, lod
, NULL
);
1181 /* Array size actually ends up in .w rather than .z. This doesn't
1182 * matter for miplevel 0, but for higher mips the value in z is
1183 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
1184 * returned, which means that we have to add 1 to it for arrays for
1187 * Note use a temporary dst and then copy, since the size of the dst
1188 * array that is passed in is based on nir's understanding of the
1189 * result size, not the hardware's
1191 struct ir3_instruction
*tmp
[4];
1193 ir3_split_dest(b
, tmp
, sam
, 0, 4);
1195 /* get_size instruction returns size in bytes instead of texels
1196 * for imageBuffer, so we need to divide it by the pixel size
1197 * of the image format.
1199 * TODO: This is at least true on a5xx. Check other gens.
1201 if (nir_intrinsic_image_dim(intr
) == GLSL_SAMPLER_DIM_BUF
) {
1202 /* Since all the possible values the divisor can take are
1203 * power-of-two (4, 8, or 16), the division is implemented
1205 * During shader setup, the log2 of the image format's
1206 * bytes-per-pixel should have been emitted in 2nd slot of
1207 * image_dims. See ir3_shader::emit_image_dims().
1209 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
1210 unsigned cb
= regid(const_state
->offsets
.image_dims
, 0) +
1211 const_state
->image_dims
.off
[nir_src_as_uint(intr
->src
[0])];
1212 struct ir3_instruction
*aux
= create_uniform(b
, cb
+ 1);
1214 tmp
[0] = ir3_SHR_B(b
, tmp
[0], 0, aux
, 0);
1217 for (unsigned i
= 0; i
< ncoords
; i
++)
1220 if (flags
& IR3_INSTR_A
) {
1221 if (ctx
->compiler
->levels_add_one
) {
1222 dst
[ncoords
-1] = ir3_ADD_U(b
, tmp
[3], 0, create_immed(b
, 1), 0);
1224 dst
[ncoords
-1] = ir3_MOV(b
, tmp
[3], TYPE_U32
);
1230 emit_intrinsic_barrier(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1232 struct ir3_block
*b
= ctx
->block
;
1233 struct ir3_instruction
*barrier
;
1235 switch (intr
->intrinsic
) {
1236 case nir_intrinsic_control_barrier
:
1237 barrier
= ir3_BAR(b
);
1238 barrier
->cat7
.g
= true;
1239 barrier
->cat7
.l
= true;
1240 barrier
->flags
= IR3_INSTR_SS
| IR3_INSTR_SY
;
1241 barrier
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1243 case nir_intrinsic_memory_barrier
:
1244 barrier
= ir3_FENCE(b
);
1245 barrier
->cat7
.g
= true;
1246 barrier
->cat7
.r
= true;
1247 barrier
->cat7
.w
= true;
1248 barrier
->cat7
.l
= true;
1249 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
|
1250 IR3_BARRIER_BUFFER_W
;
1251 barrier
->barrier_conflict
=
1252 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1253 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1255 case nir_intrinsic_memory_barrier_buffer
:
1256 barrier
= ir3_FENCE(b
);
1257 barrier
->cat7
.g
= true;
1258 barrier
->cat7
.r
= true;
1259 barrier
->cat7
.w
= true;
1260 barrier
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1261 barrier
->barrier_conflict
= IR3_BARRIER_BUFFER_R
|
1262 IR3_BARRIER_BUFFER_W
;
1264 case nir_intrinsic_memory_barrier_image
:
1265 // TODO double check if this should have .g set
1266 barrier
= ir3_FENCE(b
);
1267 barrier
->cat7
.g
= true;
1268 barrier
->cat7
.r
= true;
1269 barrier
->cat7
.w
= true;
1270 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1271 barrier
->barrier_conflict
= IR3_BARRIER_IMAGE_R
|
1272 IR3_BARRIER_IMAGE_W
;
1274 case nir_intrinsic_memory_barrier_shared
:
1275 barrier
= ir3_FENCE(b
);
1276 barrier
->cat7
.g
= true;
1277 barrier
->cat7
.l
= true;
1278 barrier
->cat7
.r
= true;
1279 barrier
->cat7
.w
= true;
1280 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
;
1281 barrier
->barrier_conflict
= IR3_BARRIER_SHARED_R
|
1282 IR3_BARRIER_SHARED_W
;
1284 case nir_intrinsic_group_memory_barrier
:
1285 barrier
= ir3_FENCE(b
);
1286 barrier
->cat7
.g
= true;
1287 barrier
->cat7
.l
= true;
1288 barrier
->cat7
.r
= true;
1289 barrier
->cat7
.w
= true;
1290 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
|
1291 IR3_BARRIER_IMAGE_W
|
1292 IR3_BARRIER_BUFFER_W
;
1293 barrier
->barrier_conflict
=
1294 IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
|
1295 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1296 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1302 /* make sure barrier doesn't get DCE'd */
1303 array_insert(b
, b
->keeps
, barrier
);
1306 static void add_sysval_input_compmask(struct ir3_context
*ctx
,
1307 gl_system_value slot
, unsigned compmask
,
1308 struct ir3_instruction
*instr
)
1310 struct ir3_shader_variant
*so
= ctx
->so
;
1311 unsigned n
= so
->inputs_count
++;
1313 assert(instr
->opc
== OPC_META_INPUT
);
1314 instr
->input
.inidx
= n
;
1315 instr
->input
.sysval
= slot
;
1317 so
->inputs
[n
].sysval
= true;
1318 so
->inputs
[n
].slot
= slot
;
1319 so
->inputs
[n
].compmask
= compmask
;
1320 so
->inputs
[n
].interpolate
= INTERP_MODE_FLAT
;
1324 static struct ir3_instruction
*
1325 create_sysval_input(struct ir3_context
*ctx
, gl_system_value slot
,
1329 struct ir3_instruction
*sysval
= create_input(ctx
, compmask
);
1330 add_sysval_input_compmask(ctx
, slot
, compmask
, sysval
);
1334 static struct ir3_instruction
*
1335 get_barycentric_centroid(struct ir3_context
*ctx
)
1337 if (!ctx
->ij_centroid
) {
1338 struct ir3_instruction
*xy
[2];
1339 struct ir3_instruction
*ij
;
1341 ij
= create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID
, 0x3);
1342 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1344 ctx
->ij_centroid
= ir3_create_collect(ctx
, xy
, 2);
1347 return ctx
->ij_centroid
;
1350 static struct ir3_instruction
*
1351 get_barycentric_sample(struct ir3_context
*ctx
)
1353 if (!ctx
->ij_sample
) {
1354 struct ir3_instruction
*xy
[2];
1355 struct ir3_instruction
*ij
;
1357 ij
= create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE
, 0x3);
1358 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1360 ctx
->ij_sample
= ir3_create_collect(ctx
, xy
, 2);
1363 return ctx
->ij_sample
;
1366 static struct ir3_instruction
*
1367 get_barycentric_pixel(struct ir3_context
*ctx
)
1369 /* TODO when tgsi_to_nir supports "new-style" FS inputs switch
1370 * this to create ij_pixel only on demand:
1372 return ctx
->ij_pixel
;
1375 static struct ir3_instruction
*
1376 get_frag_coord(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1378 if (!ctx
->frag_coord
) {
1379 struct ir3_block
*b
= ctx
->in_block
;
1380 struct ir3_instruction
*xyzw
[4];
1381 struct ir3_instruction
*hw_frag_coord
;
1383 hw_frag_coord
= create_sysval_input(ctx
, SYSTEM_VALUE_FRAG_COORD
, 0xf);
1384 ir3_split_dest(b
, xyzw
, hw_frag_coord
, 0, 4);
1386 /* for frag_coord.xy, we get unsigned values.. we need
1387 * to subtract (integer) 8 and divide by 16 (right-
1388 * shift by 4) then convert to float:
1392 * mov.u32f32 dst, tmp
1395 for (int i
= 0; i
< 2; i
++) {
1396 xyzw
[i
] = ir3_COV(b
, xyzw
[i
], TYPE_U32
, TYPE_F32
);
1397 xyzw
[i
] = ir3_MUL_F(b
, xyzw
[i
], 0, create_immed(b
, fui(1.0 / 16.0)), 0);
1400 ctx
->frag_coord
= ir3_create_collect(ctx
, xyzw
, 4);
1403 ctx
->so
->fragcoord_compmask
|=
1404 nir_ssa_def_components_read(&intr
->dest
.ssa
);
1406 return ctx
->frag_coord
;
1410 emit_intrinsic(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1412 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1413 struct ir3_instruction
**dst
;
1414 struct ir3_instruction
* const *src
;
1415 struct ir3_block
*b
= ctx
->block
;
1418 if (info
->has_dest
) {
1419 unsigned n
= nir_intrinsic_dest_components(intr
);
1420 dst
= ir3_get_dst(ctx
, &intr
->dest
, n
);
1425 const unsigned primitive_param
= ctx
->so
->shader
->const_state
.offsets
.primitive_param
* 4;
1426 const unsigned primitive_map
= ctx
->so
->shader
->const_state
.offsets
.primitive_map
* 4;
1428 switch (intr
->intrinsic
) {
1429 case nir_intrinsic_load_uniform
:
1430 idx
= nir_intrinsic_base(intr
);
1431 if (nir_src_is_const(intr
->src
[0])) {
1432 idx
+= nir_src_as_uint(intr
->src
[0]);
1433 for (int i
= 0; i
< intr
->num_components
; i
++) {
1434 dst
[i
] = create_uniform_typed(b
, idx
+ i
,
1435 nir_dest_bit_size(intr
->dest
) == 16 ? TYPE_F16
: TYPE_F32
);
1438 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1439 for (int i
= 0; i
< intr
->num_components
; i
++) {
1440 dst
[i
] = create_uniform_indirect(b
, idx
+ i
,
1441 ir3_get_addr0(ctx
, src
[0], 1));
1443 /* NOTE: if relative addressing is used, we set
1444 * constlen in the compiler (to worst-case value)
1445 * since we don't know in the assembler what the max
1446 * addr reg value can be:
1448 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
1449 ctx
->so
->shader
->ubo_state
.size
/ 16);
1453 case nir_intrinsic_load_vs_primitive_stride_ir3
:
1454 dst
[0] = create_uniform(b
, primitive_param
+ 0);
1456 case nir_intrinsic_load_vs_vertex_stride_ir3
:
1457 dst
[0] = create_uniform(b
, primitive_param
+ 1);
1459 case nir_intrinsic_load_hs_patch_stride_ir3
:
1460 dst
[0] = create_uniform(b
, primitive_param
+ 2);
1462 case nir_intrinsic_load_patch_vertices_in
:
1463 dst
[0] = create_uniform(b
, primitive_param
+ 3);
1465 case nir_intrinsic_load_tess_param_base_ir3
:
1466 dst
[0] = create_uniform(b
, primitive_param
+ 4);
1467 dst
[1] = create_uniform(b
, primitive_param
+ 5);
1469 case nir_intrinsic_load_tess_factor_base_ir3
:
1470 dst
[0] = create_uniform(b
, primitive_param
+ 6);
1471 dst
[1] = create_uniform(b
, primitive_param
+ 7);
1474 case nir_intrinsic_load_primitive_location_ir3
:
1475 idx
= nir_intrinsic_driver_location(intr
);
1476 dst
[0] = create_uniform(b
, primitive_map
+ idx
);
1479 case nir_intrinsic_load_gs_header_ir3
:
1480 dst
[0] = ctx
->gs_header
;
1482 case nir_intrinsic_load_tcs_header_ir3
:
1483 dst
[0] = ctx
->tcs_header
;
1486 case nir_intrinsic_load_primitive_id
:
1487 dst
[0] = ctx
->primitive_id
;
1490 case nir_intrinsic_load_tess_coord
:
1491 if (!ctx
->tess_coord
) {
1493 create_sysval_input(ctx
, SYSTEM_VALUE_TESS_COORD
, 0x3);
1495 ir3_split_dest(b
, dst
, ctx
->tess_coord
, 0, 2);
1497 /* Unused, but ir3_put_dst() below wants to free something */
1498 dst
[2] = create_immed(b
, 0);
1501 case nir_intrinsic_end_patch_ir3
:
1502 assert(ctx
->so
->type
== MESA_SHADER_TESS_CTRL
);
1503 struct ir3_instruction
*end
= ir3_PREDE(b
);
1504 array_insert(b
, b
->keeps
, end
);
1506 end
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1507 end
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
1510 case nir_intrinsic_store_global_ir3
: {
1511 struct ir3_instruction
*value
, *addr
, *offset
;
1513 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1514 ir3_get_src(ctx
, &intr
->src
[1])[0],
1515 ir3_get_src(ctx
, &intr
->src
[1])[1]
1518 offset
= ir3_get_src(ctx
, &intr
->src
[2])[0];
1520 value
= ir3_create_collect(ctx
, ir3_get_src(ctx
, &intr
->src
[0]),
1521 intr
->num_components
);
1523 struct ir3_instruction
*stg
=
1524 ir3_STG_G(ctx
->block
, addr
, 0, value
, 0,
1525 create_immed(ctx
->block
, intr
->num_components
), 0, offset
, 0);
1526 stg
->cat6
.type
= TYPE_U32
;
1527 stg
->cat6
.iim_val
= 1;
1529 array_insert(b
, b
->keeps
, stg
);
1531 stg
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1532 stg
->barrier_conflict
= IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1536 case nir_intrinsic_load_global_ir3
: {
1537 struct ir3_instruction
*addr
, *offset
;
1539 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1540 ir3_get_src(ctx
, &intr
->src
[0])[0],
1541 ir3_get_src(ctx
, &intr
->src
[0])[1]
1544 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
1546 struct ir3_instruction
*load
=
1547 ir3_LDG(b
, addr
, 0, create_immed(ctx
->block
, intr
->num_components
),
1549 load
->cat6
.type
= TYPE_U32
;
1550 load
->regs
[0]->wrmask
= MASK(intr
->num_components
);
1552 load
->barrier_class
= IR3_BARRIER_BUFFER_R
;
1553 load
->barrier_conflict
= IR3_BARRIER_BUFFER_W
;
1555 ir3_split_dest(b
, dst
, load
, 0, intr
->num_components
);
1559 case nir_intrinsic_load_ubo
:
1560 emit_intrinsic_load_ubo(ctx
, intr
, dst
);
1562 case nir_intrinsic_load_ubo_ir3
:
1563 emit_intrinsic_load_ubo_ldc(ctx
, intr
, dst
);
1565 case nir_intrinsic_load_frag_coord
:
1566 ir3_split_dest(b
, dst
, get_frag_coord(ctx
, intr
), 0, 4);
1568 case nir_intrinsic_load_sample_pos_from_id
: {
1569 /* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
1570 * but that doesn't seem necessary.
1572 struct ir3_instruction
*offset
=
1573 ir3_RGETPOS(b
, ir3_get_src(ctx
, &intr
->src
[0])[0], 0);
1574 offset
->regs
[0]->wrmask
= 0x3;
1575 offset
->cat5
.type
= TYPE_F32
;
1577 ir3_split_dest(b
, dst
, offset
, 0, 2);
1581 case nir_intrinsic_load_size_ir3
:
1582 if (!ctx
->ij_size
) {
1584 create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE
, 0x1);
1586 dst
[0] = ctx
->ij_size
;
1588 case nir_intrinsic_load_barycentric_centroid
:
1589 ir3_split_dest(b
, dst
, get_barycentric_centroid(ctx
), 0, 2);
1591 case nir_intrinsic_load_barycentric_sample
:
1592 if (ctx
->so
->key
.msaa
) {
1593 ir3_split_dest(b
, dst
, get_barycentric_sample(ctx
), 0, 2);
1595 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1598 case nir_intrinsic_load_barycentric_pixel
:
1599 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1601 case nir_intrinsic_load_interpolated_input
:
1602 idx
= nir_intrinsic_base(intr
);
1603 comp
= nir_intrinsic_component(intr
);
1604 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1605 if (nir_src_is_const(intr
->src
[1])) {
1606 struct ir3_instruction
*coord
= ir3_create_collect(ctx
, src
, 2);
1607 idx
+= nir_src_as_uint(intr
->src
[1]);
1608 for (int i
= 0; i
< intr
->num_components
; i
++) {
1609 unsigned inloc
= idx
* 4 + i
+ comp
;
1610 if (ctx
->so
->inputs
[idx
].bary
&&
1611 !ctx
->so
->inputs
[idx
].use_ldlv
) {
1612 dst
[i
] = ir3_BARY_F(b
, create_immed(b
, inloc
), 0, coord
, 0);
1614 /* for non-varyings use the pre-setup input, since
1615 * that is easier than mapping things back to a
1616 * nir_variable to figure out what it is.
1618 dst
[i
] = ctx
->inputs
[inloc
];
1619 compile_assert(ctx
, dst
[i
]);
1623 ir3_context_error(ctx
, "unhandled");
1626 case nir_intrinsic_load_input
:
1627 idx
= nir_intrinsic_base(intr
);
1628 comp
= nir_intrinsic_component(intr
);
1629 if (nir_src_is_const(intr
->src
[0])) {
1630 idx
+= nir_src_as_uint(intr
->src
[0]);
1631 for (int i
= 0; i
< intr
->num_components
; i
++) {
1632 unsigned n
= idx
* 4 + i
+ comp
;
1633 dst
[i
] = ctx
->inputs
[n
];
1634 compile_assert(ctx
, ctx
->inputs
[n
]);
1637 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1638 struct ir3_instruction
*collect
=
1639 ir3_create_collect(ctx
, ctx
->ir
->inputs
, ctx
->ninputs
);
1640 struct ir3_instruction
*addr
= ir3_get_addr0(ctx
, src
[0], 4);
1641 for (int i
= 0; i
< intr
->num_components
; i
++) {
1642 unsigned n
= idx
* 4 + i
+ comp
;
1643 dst
[i
] = create_indirect_load(ctx
, ctx
->ninputs
,
1648 /* All SSBO intrinsics should have been lowered by 'lower_io_offsets'
1649 * pass and replaced by an ir3-specifc version that adds the
1650 * dword-offset in the last source.
1652 case nir_intrinsic_load_ssbo_ir3
:
1653 ctx
->funcs
->emit_intrinsic_load_ssbo(ctx
, intr
, dst
);
1655 case nir_intrinsic_store_ssbo_ir3
:
1656 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1657 !ctx
->s
->info
.fs
.early_fragment_tests
)
1658 ctx
->so
->no_earlyz
= true;
1659 ctx
->funcs
->emit_intrinsic_store_ssbo(ctx
, intr
);
1661 case nir_intrinsic_get_buffer_size
:
1662 emit_intrinsic_ssbo_size(ctx
, intr
, dst
);
1664 case nir_intrinsic_ssbo_atomic_add_ir3
:
1665 case nir_intrinsic_ssbo_atomic_imin_ir3
:
1666 case nir_intrinsic_ssbo_atomic_umin_ir3
:
1667 case nir_intrinsic_ssbo_atomic_imax_ir3
:
1668 case nir_intrinsic_ssbo_atomic_umax_ir3
:
1669 case nir_intrinsic_ssbo_atomic_and_ir3
:
1670 case nir_intrinsic_ssbo_atomic_or_ir3
:
1671 case nir_intrinsic_ssbo_atomic_xor_ir3
:
1672 case nir_intrinsic_ssbo_atomic_exchange_ir3
:
1673 case nir_intrinsic_ssbo_atomic_comp_swap_ir3
:
1674 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1675 !ctx
->s
->info
.fs
.early_fragment_tests
)
1676 ctx
->so
->no_earlyz
= true;
1677 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_ssbo(ctx
, intr
);
1679 case nir_intrinsic_load_shared
:
1680 emit_intrinsic_load_shared(ctx
, intr
, dst
);
1682 case nir_intrinsic_store_shared
:
1683 emit_intrinsic_store_shared(ctx
, intr
);
1685 case nir_intrinsic_shared_atomic_add
:
1686 case nir_intrinsic_shared_atomic_imin
:
1687 case nir_intrinsic_shared_atomic_umin
:
1688 case nir_intrinsic_shared_atomic_imax
:
1689 case nir_intrinsic_shared_atomic_umax
:
1690 case nir_intrinsic_shared_atomic_and
:
1691 case nir_intrinsic_shared_atomic_or
:
1692 case nir_intrinsic_shared_atomic_xor
:
1693 case nir_intrinsic_shared_atomic_exchange
:
1694 case nir_intrinsic_shared_atomic_comp_swap
:
1695 dst
[0] = emit_intrinsic_atomic_shared(ctx
, intr
);
1697 case nir_intrinsic_image_load
:
1698 emit_intrinsic_load_image(ctx
, intr
, dst
);
1700 case nir_intrinsic_bindless_image_load
:
1701 /* Bindless uses the IBO state, which doesn't have swizzle filled out,
1702 * so using isam doesn't work.
1704 * TODO: can we use isam if we fill out more fields?
1706 ctx
->funcs
->emit_intrinsic_load_image(ctx
, intr
, dst
);
1708 case nir_intrinsic_image_store
:
1709 case nir_intrinsic_bindless_image_store
:
1710 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1711 !ctx
->s
->info
.fs
.early_fragment_tests
)
1712 ctx
->so
->no_earlyz
= true;
1713 ctx
->funcs
->emit_intrinsic_store_image(ctx
, intr
);
1715 case nir_intrinsic_image_size
:
1716 case nir_intrinsic_bindless_image_size
:
1717 emit_intrinsic_image_size(ctx
, intr
, dst
);
1719 case nir_intrinsic_image_atomic_add
:
1720 case nir_intrinsic_bindless_image_atomic_add
:
1721 case nir_intrinsic_image_atomic_imin
:
1722 case nir_intrinsic_bindless_image_atomic_imin
:
1723 case nir_intrinsic_image_atomic_umin
:
1724 case nir_intrinsic_bindless_image_atomic_umin
:
1725 case nir_intrinsic_image_atomic_imax
:
1726 case nir_intrinsic_bindless_image_atomic_imax
:
1727 case nir_intrinsic_image_atomic_umax
:
1728 case nir_intrinsic_bindless_image_atomic_umax
:
1729 case nir_intrinsic_image_atomic_and
:
1730 case nir_intrinsic_bindless_image_atomic_and
:
1731 case nir_intrinsic_image_atomic_or
:
1732 case nir_intrinsic_bindless_image_atomic_or
:
1733 case nir_intrinsic_image_atomic_xor
:
1734 case nir_intrinsic_bindless_image_atomic_xor
:
1735 case nir_intrinsic_image_atomic_exchange
:
1736 case nir_intrinsic_bindless_image_atomic_exchange
:
1737 case nir_intrinsic_image_atomic_comp_swap
:
1738 case nir_intrinsic_bindless_image_atomic_comp_swap
:
1739 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1740 !ctx
->s
->info
.fs
.early_fragment_tests
)
1741 ctx
->so
->no_earlyz
= true;
1742 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_image(ctx
, intr
);
1744 case nir_intrinsic_control_barrier
:
1745 case nir_intrinsic_memory_barrier
:
1746 case nir_intrinsic_group_memory_barrier
:
1747 case nir_intrinsic_memory_barrier_buffer
:
1748 case nir_intrinsic_memory_barrier_image
:
1749 case nir_intrinsic_memory_barrier_shared
:
1750 emit_intrinsic_barrier(ctx
, intr
);
1751 /* note that blk ptr no longer valid, make that obvious: */
1754 case nir_intrinsic_store_output
:
1755 idx
= nir_intrinsic_base(intr
);
1756 comp
= nir_intrinsic_component(intr
);
1757 compile_assert(ctx
, nir_src_is_const(intr
->src
[1]));
1758 idx
+= nir_src_as_uint(intr
->src
[1]);
1760 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1761 for (int i
= 0; i
< intr
->num_components
; i
++) {
1762 unsigned n
= idx
* 4 + i
+ comp
;
1763 ctx
->outputs
[n
] = src
[i
];
1766 case nir_intrinsic_load_base_vertex
:
1767 case nir_intrinsic_load_first_vertex
:
1768 if (!ctx
->basevertex
) {
1769 ctx
->basevertex
= create_driver_param(ctx
, IR3_DP_VTXID_BASE
);
1771 dst
[0] = ctx
->basevertex
;
1773 case nir_intrinsic_load_base_instance
:
1774 if (!ctx
->base_instance
) {
1775 ctx
->base_instance
= create_driver_param(ctx
, IR3_DP_INSTID_BASE
);
1777 dst
[0] = ctx
->base_instance
;
1779 case nir_intrinsic_load_vertex_id_zero_base
:
1780 case nir_intrinsic_load_vertex_id
:
1781 if (!ctx
->vertex_id
) {
1782 gl_system_value sv
= (intr
->intrinsic
== nir_intrinsic_load_vertex_id
) ?
1783 SYSTEM_VALUE_VERTEX_ID
: SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1784 ctx
->vertex_id
= create_sysval_input(ctx
, sv
, 0x1);
1786 dst
[0] = ctx
->vertex_id
;
1788 case nir_intrinsic_load_instance_id
:
1789 if (!ctx
->instance_id
) {
1790 ctx
->instance_id
= create_sysval_input(ctx
, SYSTEM_VALUE_INSTANCE_ID
, 0x1);
1792 dst
[0] = ctx
->instance_id
;
1794 case nir_intrinsic_load_sample_id
:
1795 ctx
->so
->per_samp
= true;
1797 case nir_intrinsic_load_sample_id_no_per_sample
:
1798 if (!ctx
->samp_id
) {
1799 ctx
->samp_id
= create_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_ID
, 0x1);
1800 ctx
->samp_id
->regs
[0]->flags
|= IR3_REG_HALF
;
1802 dst
[0] = ir3_COV(b
, ctx
->samp_id
, TYPE_U16
, TYPE_U32
);
1804 case nir_intrinsic_load_sample_mask_in
:
1805 if (!ctx
->samp_mask_in
) {
1806 ctx
->samp_mask_in
= create_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_MASK_IN
, 0x1);
1808 dst
[0] = ctx
->samp_mask_in
;
1810 case nir_intrinsic_load_user_clip_plane
:
1811 idx
= nir_intrinsic_ucp_id(intr
);
1812 for (int i
= 0; i
< intr
->num_components
; i
++) {
1813 unsigned n
= idx
* 4 + i
;
1814 dst
[i
] = create_driver_param(ctx
, IR3_DP_UCP0_X
+ n
);
1817 case nir_intrinsic_load_front_face
:
1818 if (!ctx
->frag_face
) {
1819 ctx
->so
->frag_face
= true;
1820 ctx
->frag_face
= create_sysval_input(ctx
, SYSTEM_VALUE_FRONT_FACE
, 0x1);
1821 ctx
->frag_face
->regs
[0]->flags
|= IR3_REG_HALF
;
1823 /* for fragface, we get -1 for back and 0 for front. However this is
1824 * the inverse of what nir expects (where ~0 is true).
1826 dst
[0] = ir3_CMPS_S(b
,
1828 create_immed_typed(b
, 0, TYPE_U16
), 0);
1829 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
1831 case nir_intrinsic_load_local_invocation_id
:
1832 if (!ctx
->local_invocation_id
) {
1833 ctx
->local_invocation_id
=
1834 create_sysval_input(ctx
, SYSTEM_VALUE_LOCAL_INVOCATION_ID
, 0x7);
1836 ir3_split_dest(b
, dst
, ctx
->local_invocation_id
, 0, 3);
1838 case nir_intrinsic_load_work_group_id
:
1839 if (!ctx
->work_group_id
) {
1840 ctx
->work_group_id
=
1841 create_sysval_input(ctx
, SYSTEM_VALUE_WORK_GROUP_ID
, 0x7);
1842 ctx
->work_group_id
->regs
[0]->flags
|= IR3_REG_HIGH
;
1844 ir3_split_dest(b
, dst
, ctx
->work_group_id
, 0, 3);
1846 case nir_intrinsic_load_num_work_groups
:
1847 for (int i
= 0; i
< intr
->num_components
; i
++) {
1848 dst
[i
] = create_driver_param(ctx
, IR3_DP_NUM_WORK_GROUPS_X
+ i
);
1851 case nir_intrinsic_load_local_group_size
:
1852 for (int i
= 0; i
< intr
->num_components
; i
++) {
1853 dst
[i
] = create_driver_param(ctx
, IR3_DP_LOCAL_GROUP_SIZE_X
+ i
);
1856 case nir_intrinsic_discard_if
:
1857 case nir_intrinsic_discard
: {
1858 struct ir3_instruction
*cond
, *kill
;
1860 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
1861 /* conditional discard: */
1862 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1865 /* unconditional discard: */
1866 cond
= create_immed(b
, 1);
1869 /* NOTE: only cmps.*.* can write p0.x: */
1870 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1871 cond
->cat2
.condition
= IR3_COND_NE
;
1873 /* condition always goes in predicate register: */
1874 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1875 cond
->regs
[0]->flags
&= ~IR3_REG_SSA
;
1877 kill
= ir3_KILL(b
, cond
, 0);
1878 kill
->regs
[1]->num
= regid(REG_P0
, 0);
1879 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1881 array_insert(b
, b
->keeps
, kill
);
1882 ctx
->so
->no_earlyz
= true;
1887 case nir_intrinsic_cond_end_ir3
: {
1888 struct ir3_instruction
*cond
, *kill
;
1890 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1893 /* NOTE: only cmps.*.* can write p0.x: */
1894 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1895 cond
->cat2
.condition
= IR3_COND_NE
;
1897 /* condition always goes in predicate register: */
1898 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1900 kill
= ir3_PREDT(b
, cond
, 0);
1902 kill
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1903 kill
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
1905 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1906 array_insert(b
, b
->keeps
, kill
);
1910 case nir_intrinsic_load_shared_ir3
:
1911 emit_intrinsic_load_shared_ir3(ctx
, intr
, dst
);
1913 case nir_intrinsic_store_shared_ir3
:
1914 emit_intrinsic_store_shared_ir3(ctx
, intr
);
1916 case nir_intrinsic_bindless_resource_ir3
:
1917 dst
[0] = ir3_get_src(ctx
, &intr
->src
[0])[0];
1920 ir3_context_error(ctx
, "Unhandled intrinsic type: %s\n",
1921 nir_intrinsic_infos
[intr
->intrinsic
].name
);
1926 ir3_put_dst(ctx
, &intr
->dest
);
1930 emit_load_const(struct ir3_context
*ctx
, nir_load_const_instr
*instr
)
1932 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &instr
->def
,
1933 instr
->def
.num_components
);
1935 if (instr
->def
.bit_size
== 16) {
1936 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1937 dst
[i
] = create_immed_typed(ctx
->block
,
1938 instr
->value
[i
].u16
,
1941 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1942 dst
[i
] = create_immed_typed(ctx
->block
,
1943 instr
->value
[i
].u32
,
1950 emit_undef(struct ir3_context
*ctx
, nir_ssa_undef_instr
*undef
)
1952 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &undef
->def
,
1953 undef
->def
.num_components
);
1954 type_t type
= (undef
->def
.bit_size
== 16) ? TYPE_U16
: TYPE_U32
;
1956 /* backend doesn't want undefined instructions, so just plug
1959 for (int i
= 0; i
< undef
->def
.num_components
; i
++)
1960 dst
[i
] = create_immed_typed(ctx
->block
, fui(0.0), type
);
1964 * texture fetch/sample instructions:
1968 get_tex_dest_type(nir_tex_instr
*tex
)
1972 switch (nir_alu_type_get_base_type(tex
->dest_type
)) {
1973 case nir_type_invalid
:
1974 case nir_type_float
:
1975 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_F16
: TYPE_F32
;
1978 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_S16
: TYPE_S32
;
1982 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_U16
: TYPE_U32
;
1985 unreachable("bad dest_type");
1992 tex_info(nir_tex_instr
*tex
, unsigned *flagsp
, unsigned *coordsp
)
1994 unsigned coords
= glsl_get_sampler_dim_coordinate_components(tex
->sampler_dim
);
1997 /* note: would use tex->coord_components.. except txs.. also,
1998 * since array index goes after shadow ref, we don't want to
2002 flags
|= IR3_INSTR_3D
;
2004 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
2005 flags
|= IR3_INSTR_S
;
2007 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
)
2008 flags
|= IR3_INSTR_A
;
2014 /* Gets the sampler/texture idx as a hvec2. Which could either be dynamic
2015 * or immediate (in which case it will get lowered later to a non .s2en
2016 * version of the tex instruction which encode tex/samp as immediates:
2018 static struct tex_src_info
2019 get_tex_samp_tex_src(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2021 struct ir3_block
*b
= ctx
->block
;
2022 struct tex_src_info info
= { 0 };
2023 int texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_handle
);
2024 int sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_handle
);
2025 struct ir3_instruction
*texture
, *sampler
;
2027 if (texture_idx
>= 0 || sampler_idx
>= 0) {
2029 info
.flags
|= IR3_INSTR_B
;
2031 /* Gather information required to determine which encoding to
2032 * choose as well as for prefetch.
2034 nir_intrinsic_instr
*bindless_tex
= NULL
;
2036 if (texture_idx
>= 0) {
2037 ctx
->so
->bindless_tex
= true;
2038 bindless_tex
= ir3_bindless_resource(tex
->src
[texture_idx
].src
);
2039 assert(bindless_tex
);
2040 info
.tex_base
= nir_intrinsic_desc_set(bindless_tex
);
2041 tex_const
= nir_src_is_const(bindless_tex
->src
[0]);
2043 info
.tex_idx
= nir_src_as_uint(bindless_tex
->src
[0]);
2045 /* To simplify some of the logic below, assume the index is
2046 * constant 0 when it's not enabled.
2051 nir_intrinsic_instr
*bindless_samp
= NULL
;
2053 if (sampler_idx
>= 0) {
2054 ctx
->so
->bindless_samp
= true;
2055 bindless_samp
= ir3_bindless_resource(tex
->src
[sampler_idx
].src
);
2056 assert(bindless_samp
);
2057 info
.samp_base
= nir_intrinsic_desc_set(bindless_samp
);
2058 samp_const
= nir_src_is_const(bindless_samp
->src
[0]);
2060 info
.samp_idx
= nir_src_as_uint(bindless_samp
->src
[0]);
2066 /* Choose encoding. */
2067 if (tex_const
&& samp_const
&& info
.tex_idx
< 256 && info
.samp_idx
< 256) {
2068 if (info
.tex_idx
< 16 && info
.samp_idx
< 16 &&
2069 (!bindless_tex
|| !bindless_samp
|| info
.tex_base
== info
.samp_base
)) {
2070 /* Everything fits within the instruction */
2071 info
.base
= info
.tex_base
;
2072 info
.combined_idx
= info
.samp_idx
| (info
.tex_idx
<< 4);
2074 info
.base
= info
.tex_base
;
2075 info
.a1_val
= info
.tex_idx
<< 3 | info
.samp_base
;
2076 info
.combined_idx
= info
.samp_idx
;
2077 info
.flags
|= IR3_INSTR_A1EN
;
2079 info
.samp_tex
= NULL
;
2081 info
.flags
|= IR3_INSTR_S2EN
;
2082 /* In the indirect case, we only use a1.x to store the sampler
2083 * base if it differs from the texture base.
2085 if (!bindless_tex
|| !bindless_samp
|| info
.tex_base
== info
.samp_base
) {
2086 info
.base
= info
.tex_base
;
2088 info
.base
= info
.tex_base
;
2089 info
.a1_val
= info
.samp_base
;
2090 info
.flags
|= IR3_INSTR_A1EN
;
2093 /* Note: the indirect source is now a vec2 instead of hvec2, and
2094 * for some reason the texture and sampler are swapped.
2096 struct ir3_instruction
*texture
, *sampler
;
2099 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
2101 texture
= create_immed(b
, 0);
2104 if (bindless_samp
) {
2105 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
2107 sampler
= create_immed(b
, 0);
2109 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2115 info
.flags
|= IR3_INSTR_S2EN
;
2116 texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
);
2117 sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
);
2118 if (texture_idx
>= 0) {
2119 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
2120 texture
= ir3_COV(ctx
->block
, texture
, TYPE_U32
, TYPE_U16
);
2122 /* TODO what to do for dynamic case? I guess we only need the
2123 * max index for astc srgb workaround so maybe not a problem
2124 * to worry about if we don't enable indirect samplers for
2127 ctx
->max_texture_index
= MAX2(ctx
->max_texture_index
, tex
->texture_index
);
2128 texture
= create_immed_typed(ctx
->block
, tex
->texture_index
, TYPE_U16
);
2129 info
.tex_idx
= tex
->texture_index
;
2132 if (sampler_idx
>= 0) {
2133 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
2134 sampler
= ir3_COV(ctx
->block
, sampler
, TYPE_U32
, TYPE_U16
);
2136 sampler
= create_immed_typed(ctx
->block
, tex
->sampler_index
, TYPE_U16
);
2137 info
.samp_idx
= tex
->texture_index
;
2140 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2150 emit_tex(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2152 struct ir3_block
*b
= ctx
->block
;
2153 struct ir3_instruction
**dst
, *sam
, *src0
[12], *src1
[4];
2154 struct ir3_instruction
* const *coord
, * const *off
, * const *ddx
, * const *ddy
;
2155 struct ir3_instruction
*lod
, *compare
, *proj
, *sample_index
;
2156 struct tex_src_info info
= { 0 };
2157 bool has_bias
= false, has_lod
= false, has_proj
= false, has_off
= false;
2158 unsigned i
, coords
, flags
, ncomp
;
2159 unsigned nsrc0
= 0, nsrc1
= 0;
2163 ncomp
= nir_dest_num_components(tex
->dest
);
2165 coord
= off
= ddx
= ddy
= NULL
;
2166 lod
= proj
= compare
= sample_index
= NULL
;
2168 dst
= ir3_get_dst(ctx
, &tex
->dest
, ncomp
);
2170 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
2171 switch (tex
->src
[i
].src_type
) {
2172 case nir_tex_src_coord
:
2173 coord
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2175 case nir_tex_src_bias
:
2176 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2179 case nir_tex_src_lod
:
2180 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2183 case nir_tex_src_comparator
: /* shadow comparator */
2184 compare
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2186 case nir_tex_src_projector
:
2187 proj
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2190 case nir_tex_src_offset
:
2191 off
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2194 case nir_tex_src_ddx
:
2195 ddx
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2197 case nir_tex_src_ddy
:
2198 ddy
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2200 case nir_tex_src_ms_index
:
2201 sample_index
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2203 case nir_tex_src_texture_offset
:
2204 case nir_tex_src_sampler_offset
:
2205 case nir_tex_src_texture_handle
:
2206 case nir_tex_src_sampler_handle
:
2207 /* handled in get_tex_samp_src() */
2210 ir3_context_error(ctx
, "Unhandled NIR tex src type: %d\n",
2211 tex
->src
[i
].src_type
);
2217 case nir_texop_tex_prefetch
:
2218 compile_assert(ctx
, !has_bias
);
2219 compile_assert(ctx
, !has_lod
);
2220 compile_assert(ctx
, !compare
);
2221 compile_assert(ctx
, !has_proj
);
2222 compile_assert(ctx
, !has_off
);
2223 compile_assert(ctx
, !ddx
);
2224 compile_assert(ctx
, !ddy
);
2225 compile_assert(ctx
, !sample_index
);
2226 compile_assert(ctx
, nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
) < 0);
2227 compile_assert(ctx
, nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
) < 0);
2229 if (ctx
->so
->num_sampler_prefetch
< ctx
->prefetch_limit
) {
2230 opc
= OPC_META_TEX_PREFETCH
;
2231 ctx
->so
->num_sampler_prefetch
++;
2235 case nir_texop_tex
: opc
= has_lod
? OPC_SAML
: OPC_SAM
; break;
2236 case nir_texop_txb
: opc
= OPC_SAMB
; break;
2237 case nir_texop_txl
: opc
= OPC_SAML
; break;
2238 case nir_texop_txd
: opc
= OPC_SAMGQ
; break;
2239 case nir_texop_txf
: opc
= OPC_ISAML
; break;
2240 case nir_texop_lod
: opc
= OPC_GETLOD
; break;
2242 /* NOTE: a4xx might need to emulate gather w/ txf (this is
2243 * what blob does, seems gather is broken?), and a3xx did
2244 * not support it (but probably could also emulate).
2246 switch (tex
->component
) {
2247 case 0: opc
= OPC_GATHER4R
; break;
2248 case 1: opc
= OPC_GATHER4G
; break;
2249 case 2: opc
= OPC_GATHER4B
; break;
2250 case 3: opc
= OPC_GATHER4A
; break;
2253 case nir_texop_txf_ms_fb
:
2254 case nir_texop_txf_ms
: opc
= OPC_ISAMM
; break;
2256 ir3_context_error(ctx
, "Unhandled NIR tex type: %d\n", tex
->op
);
2260 tex_info(tex
, &flags
, &coords
);
2263 * lay out the first argument in the proper order:
2264 * - actual coordinates first
2265 * - shadow reference
2268 * - starting at offset 4, dpdx.xy, dpdy.xy
2270 * bias/lod go into the second arg
2273 /* insert tex coords: */
2274 for (i
= 0; i
< coords
; i
++)
2279 /* scale up integer coords for TXF based on the LOD */
2280 if (ctx
->compiler
->unminify_coords
&& (opc
== OPC_ISAML
)) {
2282 for (i
= 0; i
< coords
; i
++)
2283 src0
[i
] = ir3_SHL_B(b
, src0
[i
], 0, lod
, 0);
2287 /* hw doesn't do 1d, so we treat it as 2d with
2288 * height of 1, and patch up the y coord.
2291 src0
[nsrc0
++] = create_immed(b
, 0);
2293 src0
[nsrc0
++] = create_immed(b
, fui(0.5));
2297 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
2298 src0
[nsrc0
++] = compare
;
2300 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
) {
2301 struct ir3_instruction
*idx
= coord
[coords
];
2303 /* the array coord for cube arrays needs 0.5 added to it */
2304 if (ctx
->compiler
->array_index_add_half
&& !is_isam(opc
))
2305 idx
= ir3_ADD_F(b
, idx
, 0, create_immed(b
, fui(0.5)), 0);
2307 src0
[nsrc0
++] = idx
;
2311 src0
[nsrc0
++] = proj
;
2312 flags
|= IR3_INSTR_P
;
2315 /* pad to 4, then ddx/ddy: */
2316 if (tex
->op
== nir_texop_txd
) {
2318 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2319 for (i
= 0; i
< coords
; i
++)
2320 src0
[nsrc0
++] = ddx
[i
];
2322 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2323 for (i
= 0; i
< coords
; i
++)
2324 src0
[nsrc0
++] = ddy
[i
];
2326 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2329 /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
2330 * with scaled x coord according to requested sample:
2332 if (opc
== OPC_ISAMM
) {
2333 if (ctx
->compiler
->txf_ms_with_isaml
) {
2334 /* the samples are laid out in x dimension as
2336 * x_ms = (x << ms) + sample_index;
2338 struct ir3_instruction
*ms
;
2339 ms
= create_immed(b
, (ctx
->samples
>> (2 * tex
->texture_index
)) & 3);
2341 src0
[0] = ir3_SHL_B(b
, src0
[0], 0, ms
, 0);
2342 src0
[0] = ir3_ADD_U(b
, src0
[0], 0, sample_index
, 0);
2346 src0
[nsrc0
++] = sample_index
;
2351 * second argument (if applicable):
2356 if (has_off
| has_lod
| has_bias
) {
2358 unsigned off_coords
= coords
;
2359 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2361 for (i
= 0; i
< off_coords
; i
++)
2362 src1
[nsrc1
++] = off
[i
];
2364 src1
[nsrc1
++] = create_immed(b
, fui(0.0));
2365 flags
|= IR3_INSTR_O
;
2368 if (has_lod
| has_bias
)
2369 src1
[nsrc1
++] = lod
;
2372 type
= get_tex_dest_type(tex
);
2374 if (opc
== OPC_GETLOD
)
2378 if (tex
->op
== nir_texop_txf_ms_fb
) {
2379 /* only expect a single txf_ms_fb per shader: */
2380 compile_assert(ctx
, !ctx
->so
->fb_read
);
2381 compile_assert(ctx
, ctx
->so
->type
== MESA_SHADER_FRAGMENT
);
2383 ctx
->so
->fb_read
= true;
2384 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2385 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2386 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2388 info
.flags
= IR3_INSTR_S2EN
;
2390 ctx
->so
->num_samp
++;
2392 info
= get_tex_samp_tex_src(ctx
, tex
);
2395 struct ir3_instruction
*col0
= ir3_create_collect(ctx
, src0
, nsrc0
);
2396 struct ir3_instruction
*col1
= ir3_create_collect(ctx
, src1
, nsrc1
);
2398 if (opc
== OPC_META_TEX_PREFETCH
) {
2399 int idx
= nir_tex_instr_src_index(tex
, nir_tex_src_coord
);
2401 compile_assert(ctx
, tex
->src
[idx
].src
.is_ssa
);
2403 sam
= ir3_META_TEX_PREFETCH(b
);
2404 __ssa_dst(sam
)->wrmask
= MASK(ncomp
); /* dst */
2405 __ssa_src(sam
, get_barycentric_pixel(ctx
), 0);
2406 sam
->prefetch
.input_offset
=
2407 ir3_nir_coord_offset(tex
->src
[idx
].src
.ssa
);
2408 /* make sure not to add irrelevant flags like S2EN */
2409 sam
->flags
= flags
| (info
.flags
& IR3_INSTR_B
);
2410 sam
->prefetch
.tex
= info
.tex_idx
;
2411 sam
->prefetch
.samp
= info
.samp_idx
;
2412 sam
->prefetch
.tex_base
= info
.tex_base
;
2413 sam
->prefetch
.samp_base
= info
.samp_base
;
2415 info
.flags
|= flags
;
2416 sam
= emit_sam(ctx
, opc
, info
, type
, MASK(ncomp
), col0
, col1
);
2419 if ((ctx
->astc_srgb
& (1 << tex
->texture_index
)) && !nir_tex_instr_is_query(tex
)) {
2420 assert(opc
!= OPC_META_TEX_PREFETCH
);
2422 /* only need first 3 components: */
2423 sam
->regs
[0]->wrmask
= 0x7;
2424 ir3_split_dest(b
, dst
, sam
, 0, 3);
2426 /* we need to sample the alpha separately with a non-ASTC
2429 sam
= ir3_SAM(b
, opc
, type
, 0b1000, flags
| info
.flags
,
2430 info
.samp_tex
, col0
, col1
);
2432 array_insert(ctx
->ir
, ctx
->ir
->astc_srgb
, sam
);
2434 /* fixup .w component: */
2435 ir3_split_dest(b
, &dst
[3], sam
, 3, 1);
2437 /* normal (non-workaround) case: */
2438 ir3_split_dest(b
, dst
, sam
, 0, ncomp
);
2441 /* GETLOD returns results in 4.8 fixed point */
2442 if (opc
== OPC_GETLOD
) {
2443 struct ir3_instruction
*factor
= create_immed(b
, fui(1.0 / 256));
2445 compile_assert(ctx
, tex
->dest_type
== nir_type_float
);
2446 for (i
= 0; i
< 2; i
++) {
2447 dst
[i
] = ir3_MUL_F(b
, ir3_COV(b
, dst
[i
], TYPE_S32
, TYPE_F32
), 0,
2452 ir3_put_dst(ctx
, &tex
->dest
);
2456 emit_tex_info(struct ir3_context
*ctx
, nir_tex_instr
*tex
, unsigned idx
)
2458 struct ir3_block
*b
= ctx
->block
;
2459 struct ir3_instruction
**dst
, *sam
;
2460 type_t dst_type
= get_tex_dest_type(tex
);
2461 struct tex_src_info info
= get_tex_samp_tex_src(ctx
, tex
);
2463 dst
= ir3_get_dst(ctx
, &tex
->dest
, 1);
2465 sam
= emit_sam(ctx
, OPC_GETINFO
, info
, dst_type
, 1 << idx
, NULL
, NULL
);
2467 /* even though there is only one component, since it ends
2468 * up in .y/.z/.w rather than .x, we need a split_dest()
2470 ir3_split_dest(b
, dst
, sam
, idx
, 1);
2472 /* The # of levels comes from getinfo.z. We need to add 1 to it, since
2473 * the value in TEX_CONST_0 is zero-based.
2475 if (ctx
->compiler
->levels_add_one
)
2476 dst
[0] = ir3_ADD_U(b
, dst
[0], 0, create_immed(b
, 1), 0);
2478 ir3_put_dst(ctx
, &tex
->dest
);
2482 emit_tex_txs(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2484 struct ir3_block
*b
= ctx
->block
;
2485 struct ir3_instruction
**dst
, *sam
;
2486 struct ir3_instruction
*lod
;
2487 unsigned flags
, coords
;
2488 type_t dst_type
= get_tex_dest_type(tex
);
2489 struct tex_src_info info
= get_tex_samp_tex_src(ctx
, tex
);
2491 tex_info(tex
, &flags
, &coords
);
2492 info
.flags
|= flags
;
2494 /* Actually we want the number of dimensions, not coordinates. This
2495 * distinction only matters for cubes.
2497 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2500 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
2502 int lod_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_lod
);
2503 compile_assert(ctx
, lod_idx
>= 0);
2505 lod
= ir3_get_src(ctx
, &tex
->src
[lod_idx
].src
)[0];
2507 sam
= emit_sam(ctx
, OPC_GETSIZE
, info
, dst_type
, 0b1111, lod
, NULL
);
2508 ir3_split_dest(b
, dst
, sam
, 0, 4);
2510 /* Array size actually ends up in .w rather than .z. This doesn't
2511 * matter for miplevel 0, but for higher mips the value in z is
2512 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2513 * returned, which means that we have to add 1 to it for arrays.
2515 if (tex
->is_array
) {
2516 if (ctx
->compiler
->levels_add_one
) {
2517 dst
[coords
] = ir3_ADD_U(b
, dst
[3], 0, create_immed(b
, 1), 0);
2519 dst
[coords
] = ir3_MOV(b
, dst
[3], TYPE_U32
);
2523 ir3_put_dst(ctx
, &tex
->dest
);
2527 emit_jump(struct ir3_context
*ctx
, nir_jump_instr
*jump
)
2529 switch (jump
->type
) {
2530 case nir_jump_break
:
2531 case nir_jump_continue
:
2532 case nir_jump_return
:
2533 /* I *think* we can simply just ignore this, and use the
2534 * successor block link to figure out where we need to
2535 * jump to for break/continue
2539 ir3_context_error(ctx
, "Unhandled NIR jump type: %d\n", jump
->type
);
2545 emit_instr(struct ir3_context
*ctx
, nir_instr
*instr
)
2547 switch (instr
->type
) {
2548 case nir_instr_type_alu
:
2549 emit_alu(ctx
, nir_instr_as_alu(instr
));
2551 case nir_instr_type_deref
:
2552 /* ignored, handled as part of the intrinsic they are src to */
2554 case nir_instr_type_intrinsic
:
2555 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2557 case nir_instr_type_load_const
:
2558 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2560 case nir_instr_type_ssa_undef
:
2561 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
2563 case nir_instr_type_tex
: {
2564 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
2565 /* couple tex instructions get special-cased:
2569 emit_tex_txs(ctx
, tex
);
2571 case nir_texop_query_levels
:
2572 emit_tex_info(ctx
, tex
, 2);
2574 case nir_texop_texture_samples
:
2575 emit_tex_info(ctx
, tex
, 3);
2583 case nir_instr_type_jump
:
2584 emit_jump(ctx
, nir_instr_as_jump(instr
));
2586 case nir_instr_type_phi
:
2587 /* we have converted phi webs to regs in NIR by now */
2588 ir3_context_error(ctx
, "Unexpected NIR instruction type: %d\n", instr
->type
);
2590 case nir_instr_type_call
:
2591 case nir_instr_type_parallel_copy
:
2592 ir3_context_error(ctx
, "Unhandled NIR instruction type: %d\n", instr
->type
);
2597 static struct ir3_block
*
2598 get_block(struct ir3_context
*ctx
, const nir_block
*nblock
)
2600 struct ir3_block
*block
;
2601 struct hash_entry
*hentry
;
2603 hentry
= _mesa_hash_table_search(ctx
->block_ht
, nblock
);
2605 return hentry
->data
;
2607 block
= ir3_block_create(ctx
->ir
);
2608 block
->nblock
= nblock
;
2609 _mesa_hash_table_insert(ctx
->block_ht
, nblock
, block
);
2611 set_foreach(nblock
->predecessors
, sentry
) {
2612 _mesa_set_add(block
->predecessors
, get_block(ctx
, sentry
->key
));
2619 emit_block(struct ir3_context
*ctx
, nir_block
*nblock
)
2621 struct ir3_block
*block
= get_block(ctx
, nblock
);
2623 for (int i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
2624 if (nblock
->successors
[i
]) {
2625 block
->successors
[i
] =
2626 get_block(ctx
, nblock
->successors
[i
]);
2631 list_addtail(&block
->node
, &ctx
->ir
->block_list
);
2633 /* re-emit addr register in each block if needed: */
2634 for (int i
= 0; i
< ARRAY_SIZE(ctx
->addr0_ht
); i
++) {
2635 _mesa_hash_table_destroy(ctx
->addr0_ht
[i
], NULL
);
2636 ctx
->addr0_ht
[i
] = NULL
;
2639 _mesa_hash_table_u64_destroy(ctx
->addr1_ht
, NULL
);
2640 ctx
->addr1_ht
= NULL
;
2642 nir_foreach_instr (instr
, nblock
) {
2643 ctx
->cur_instr
= instr
;
2644 emit_instr(ctx
, instr
);
2645 ctx
->cur_instr
= NULL
;
2650 _mesa_hash_table_clear(ctx
->sel_cond_conversions
, NULL
);
2653 static void emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
);
2656 emit_if(struct ir3_context
*ctx
, nir_if
*nif
)
2658 struct ir3_instruction
*condition
= ir3_get_src(ctx
, &nif
->condition
)[0];
2660 ctx
->block
->condition
= ir3_get_predicate(ctx
, condition
);
2662 emit_cf_list(ctx
, &nif
->then_list
);
2663 emit_cf_list(ctx
, &nif
->else_list
);
2667 emit_loop(struct ir3_context
*ctx
, nir_loop
*nloop
)
2669 emit_cf_list(ctx
, &nloop
->body
);
2674 stack_push(struct ir3_context
*ctx
)
2677 ctx
->max_stack
= MAX2(ctx
->max_stack
, ctx
->stack
);
2681 stack_pop(struct ir3_context
*ctx
)
2683 compile_assert(ctx
, ctx
->stack
> 0);
2688 emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
)
2690 foreach_list_typed (nir_cf_node
, node
, node
, list
) {
2691 switch (node
->type
) {
2692 case nir_cf_node_block
:
2693 emit_block(ctx
, nir_cf_node_as_block(node
));
2695 case nir_cf_node_if
:
2697 emit_if(ctx
, nir_cf_node_as_if(node
));
2700 case nir_cf_node_loop
:
2702 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2705 case nir_cf_node_function
:
2706 ir3_context_error(ctx
, "TODO\n");
2712 /* emit stream-out code. At this point, the current block is the original
2713 * (nir) end block, and nir ensures that all flow control paths terminate
2714 * into the end block. We re-purpose the original end block to generate
2715 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2716 * block holding stream-out write instructions, followed by the new end
2720 * p0.x = (vtxcnt < maxvtxcnt)
2721 * // succs: blockStreamOut, blockNewEnd
2724 * // preds: blockOrigEnd
2725 * ... stream-out instructions ...
2726 * // succs: blockNewEnd
2729 * // preds: blockOrigEnd, blockStreamOut
2733 emit_stream_out(struct ir3_context
*ctx
)
2735 struct ir3
*ir
= ctx
->ir
;
2736 struct ir3_stream_output_info
*strmout
=
2737 &ctx
->so
->shader
->stream_output
;
2738 struct ir3_block
*orig_end_block
, *stream_out_block
, *new_end_block
;
2739 struct ir3_instruction
*vtxcnt
, *maxvtxcnt
, *cond
;
2740 struct ir3_instruction
*bases
[IR3_MAX_SO_BUFFERS
];
2742 /* create vtxcnt input in input block at top of shader,
2743 * so that it is seen as live over the entire duration
2746 vtxcnt
= create_sysval_input(ctx
, SYSTEM_VALUE_VERTEX_CNT
, 0x1);
2747 maxvtxcnt
= create_driver_param(ctx
, IR3_DP_VTXCNT_MAX
);
2749 /* at this point, we are at the original 'end' block,
2750 * re-purpose this block to stream-out condition, then
2751 * append stream-out block and new-end block
2753 orig_end_block
= ctx
->block
;
2755 // maybe w/ store_global intrinsic, we could do this
2756 // stuff in nir->nir pass
2758 stream_out_block
= ir3_block_create(ir
);
2759 list_addtail(&stream_out_block
->node
, &ir
->block_list
);
2761 new_end_block
= ir3_block_create(ir
);
2762 list_addtail(&new_end_block
->node
, &ir
->block_list
);
2764 orig_end_block
->successors
[0] = stream_out_block
;
2765 orig_end_block
->successors
[1] = new_end_block
;
2767 stream_out_block
->successors
[0] = new_end_block
;
2768 _mesa_set_add(stream_out_block
->predecessors
, orig_end_block
);
2770 _mesa_set_add(new_end_block
->predecessors
, orig_end_block
);
2771 _mesa_set_add(new_end_block
->predecessors
, stream_out_block
);
2773 /* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2774 cond
= ir3_CMPS_S(ctx
->block
, vtxcnt
, 0, maxvtxcnt
, 0);
2775 cond
->regs
[0]->num
= regid(REG_P0
, 0);
2776 cond
->regs
[0]->flags
&= ~IR3_REG_SSA
;
2777 cond
->cat2
.condition
= IR3_COND_LT
;
2779 /* condition goes on previous block to the conditional,
2780 * since it is used to pick which of the two successor
2783 orig_end_block
->condition
= cond
;
2785 /* switch to stream_out_block to generate the stream-out
2788 ctx
->block
= stream_out_block
;
2790 /* Calculate base addresses based on vtxcnt. Instructions
2791 * generated for bases not used in following loop will be
2792 * stripped out in the backend.
2794 for (unsigned i
= 0; i
< IR3_MAX_SO_BUFFERS
; i
++) {
2795 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
2796 unsigned stride
= strmout
->stride
[i
];
2797 struct ir3_instruction
*base
, *off
;
2799 base
= create_uniform(ctx
->block
, regid(const_state
->offsets
.tfbo
, i
));
2801 /* 24-bit should be enough: */
2802 off
= ir3_MUL_U24(ctx
->block
, vtxcnt
, 0,
2803 create_immed(ctx
->block
, stride
* 4), 0);
2805 bases
[i
] = ir3_ADD_S(ctx
->block
, off
, 0, base
, 0);
2808 /* Generate the per-output store instructions: */
2809 for (unsigned i
= 0; i
< strmout
->num_outputs
; i
++) {
2810 for (unsigned j
= 0; j
< strmout
->output
[i
].num_components
; j
++) {
2811 unsigned c
= j
+ strmout
->output
[i
].start_component
;
2812 struct ir3_instruction
*base
, *out
, *stg
;
2814 base
= bases
[strmout
->output
[i
].output_buffer
];
2815 out
= ctx
->outputs
[regid(strmout
->output
[i
].register_index
, c
)];
2817 stg
= ir3_STG(ctx
->block
, base
, 0, out
, 0,
2818 create_immed(ctx
->block
, 1), 0);
2819 stg
->cat6
.type
= TYPE_U32
;
2820 stg
->cat6
.dst_offset
= (strmout
->output
[i
].dst_offset
+ j
) * 4;
2822 array_insert(ctx
->block
, ctx
->block
->keeps
, stg
);
2826 /* and finally switch to the new_end_block: */
2827 ctx
->block
= new_end_block
;
2831 emit_function(struct ir3_context
*ctx
, nir_function_impl
*impl
)
2833 nir_metadata_require(impl
, nir_metadata_block_index
);
2835 compile_assert(ctx
, ctx
->stack
== 0);
2837 emit_cf_list(ctx
, &impl
->body
);
2838 emit_block(ctx
, impl
->end_block
);
2840 compile_assert(ctx
, ctx
->stack
== 0);
2842 /* at this point, we should have a single empty block,
2843 * into which we emit the 'end' instruction.
2845 compile_assert(ctx
, list_is_empty(&ctx
->block
->instr_list
));
2847 /* If stream-out (aka transform-feedback) enabled, emit the
2848 * stream-out instructions, followed by a new empty block (into
2849 * which the 'end' instruction lands).
2851 * NOTE: it is done in this order, rather than inserting before
2852 * we emit end_block, because NIR guarantees that all blocks
2853 * flow into end_block, and that end_block has no successors.
2854 * So by re-purposing end_block as the first block of stream-
2855 * out, we guarantee that all exit paths flow into the stream-
2858 if ((ctx
->compiler
->gpu_id
< 500) &&
2859 (ctx
->so
->shader
->stream_output
.num_outputs
> 0) &&
2860 !ctx
->so
->binning_pass
) {
2861 debug_assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
2862 emit_stream_out(ctx
);
2865 /* Vertex shaders in a tessellation or geometry pipeline treat END as a
2866 * NOP and has an epilogue that writes the VS outputs to local storage, to
2867 * be read by the HS. Then it resets execution mask (chmask) and chains
2868 * to the next shader (chsh).
2870 if ((ctx
->so
->type
== MESA_SHADER_VERTEX
&&
2871 (ctx
->so
->key
.has_gs
|| ctx
->so
->key
.tessellation
)) ||
2872 (ctx
->so
->type
== MESA_SHADER_TESS_EVAL
&& ctx
->so
->key
.has_gs
)) {
2873 struct ir3_instruction
*chmask
=
2874 ir3_CHMASK(ctx
->block
);
2875 chmask
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2876 chmask
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2878 struct ir3_instruction
*chsh
=
2879 ir3_CHSH(ctx
->block
);
2880 chsh
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2881 chsh
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2883 ir3_END(ctx
->block
);
2888 setup_input(struct ir3_context
*ctx
, nir_variable
*in
)
2890 struct ir3_shader_variant
*so
= ctx
->so
;
2891 unsigned ncomp
= glsl_get_components(in
->type
);
2892 unsigned n
= in
->data
.driver_location
;
2893 unsigned frac
= in
->data
.location_frac
;
2894 unsigned slot
= in
->data
.location
;
2896 /* Inputs are loaded using ldlw or ldg for these stages. */
2897 if (ctx
->so
->type
== MESA_SHADER_TESS_CTRL
||
2898 ctx
->so
->type
== MESA_SHADER_TESS_EVAL
||
2899 ctx
->so
->type
== MESA_SHADER_GEOMETRY
)
2902 /* skip unread inputs, we could end up with (for example), unsplit
2903 * matrix/etc inputs in the case they are not read, so just silently
2909 so
->inputs
[n
].slot
= slot
;
2910 so
->inputs
[n
].compmask
|= (1 << (ncomp
+ frac
)) - 1;
2911 so
->inputs_count
= MAX2(so
->inputs_count
, n
+ 1);
2912 so
->inputs
[n
].interpolate
= in
->data
.interpolation
;
2914 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2916 /* if any varyings have 'sample' qualifer, that triggers us
2917 * to run in per-sample mode:
2919 so
->per_samp
|= in
->data
.sample
;
2921 for (int i
= 0; i
< ncomp
; i
++) {
2922 struct ir3_instruction
*instr
= NULL
;
2923 unsigned idx
= (n
* 4) + i
+ frac
;
2925 if (slot
== VARYING_SLOT_POS
) {
2926 ir3_context_error(ctx
, "fragcoord should be a sysval!\n");
2928 /* detect the special case for front/back colors where
2929 * we need to do flat vs smooth shading depending on
2932 if (in
->data
.interpolation
== INTERP_MODE_NONE
) {
2934 case VARYING_SLOT_COL0
:
2935 case VARYING_SLOT_COL1
:
2936 case VARYING_SLOT_BFC0
:
2937 case VARYING_SLOT_BFC1
:
2938 so
->inputs
[n
].rasterflat
= true;
2945 if (ctx
->compiler
->flat_bypass
) {
2946 if ((so
->inputs
[n
].interpolate
== INTERP_MODE_FLAT
) ||
2947 (so
->inputs
[n
].rasterflat
&& ctx
->so
->key
.rasterflat
))
2948 so
->inputs
[n
].use_ldlv
= true;
2951 so
->inputs
[n
].bary
= true;
2953 instr
= create_frag_input(ctx
, so
->inputs
[n
].use_ldlv
, idx
);
2956 compile_assert(ctx
, idx
< ctx
->ninputs
);
2958 ctx
->inputs
[idx
] = instr
;
2960 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2961 struct ir3_instruction
*input
= NULL
, *in
;
2962 struct ir3_instruction
*components
[4];
2963 unsigned mask
= (1 << (ncomp
+ frac
)) - 1;
2965 foreach_input (in
, ctx
->ir
) {
2966 if (in
->input
.inidx
== n
) {
2973 input
= create_input(ctx
, mask
);
2974 input
->input
.inidx
= n
;
2976 input
->regs
[0]->wrmask
|= mask
;
2979 ir3_split_dest(ctx
->block
, components
, input
, frac
, ncomp
);
2981 for (int i
= 0; i
< ncomp
; i
++) {
2982 unsigned idx
= (n
* 4) + i
+ frac
;
2983 compile_assert(ctx
, idx
< ctx
->ninputs
);
2984 ctx
->inputs
[idx
] = components
[i
];
2987 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
2990 if (so
->inputs
[n
].bary
|| (ctx
->so
->type
== MESA_SHADER_VERTEX
)) {
2991 so
->total_in
+= ncomp
;
2995 /* Initially we assign non-packed inloc's for varyings, as we don't really
2996 * know up-front which components will be unused. After all the compilation
2997 * stages we scan the shader to see which components are actually used, and
2998 * re-pack the inlocs to eliminate unneeded varyings.
3001 pack_inlocs(struct ir3_context
*ctx
)
3003 struct ir3_shader_variant
*so
= ctx
->so
;
3004 uint8_t used_components
[so
->inputs_count
];
3006 memset(used_components
, 0, sizeof(used_components
));
3009 * First Step: scan shader to find which bary.f/ldlv remain:
3012 foreach_block (block
, &ctx
->ir
->block_list
) {
3013 foreach_instr (instr
, &block
->instr_list
) {
3014 if (is_input(instr
)) {
3015 unsigned inloc
= instr
->regs
[1]->iim_val
;
3016 unsigned i
= inloc
/ 4;
3017 unsigned j
= inloc
% 4;
3019 compile_assert(ctx
, instr
->regs
[1]->flags
& IR3_REG_IMMED
);
3020 compile_assert(ctx
, i
< so
->inputs_count
);
3022 used_components
[i
] |= 1 << j
;
3023 } else if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3024 for (int n
= 0; n
< 2; n
++) {
3025 unsigned inloc
= instr
->prefetch
.input_offset
+ n
;
3026 unsigned i
= inloc
/ 4;
3027 unsigned j
= inloc
% 4;
3029 compile_assert(ctx
, i
< so
->inputs_count
);
3031 used_components
[i
] |= 1 << j
;
3038 * Second Step: reassign varying inloc/slots:
3041 unsigned actual_in
= 0;
3044 for (unsigned i
= 0; i
< so
->inputs_count
; i
++) {
3045 unsigned compmask
= 0, maxcomp
= 0;
3047 so
->inputs
[i
].inloc
= inloc
;
3048 so
->inputs
[i
].bary
= false;
3050 for (unsigned j
= 0; j
< 4; j
++) {
3051 if (!(used_components
[i
] & (1 << j
)))
3054 compmask
|= (1 << j
);
3058 /* at this point, since used_components[i] mask is only
3059 * considering varyings (ie. not sysvals) we know this
3062 so
->inputs
[i
].bary
= true;
3065 if (so
->inputs
[i
].bary
) {
3067 so
->inputs
[i
].compmask
= (1 << maxcomp
) - 1;
3073 * Third Step: reassign packed inloc's:
3076 foreach_block (block
, &ctx
->ir
->block_list
) {
3077 foreach_instr (instr
, &block
->instr_list
) {
3078 if (is_input(instr
)) {
3079 unsigned inloc
= instr
->regs
[1]->iim_val
;
3080 unsigned i
= inloc
/ 4;
3081 unsigned j
= inloc
% 4;
3083 instr
->regs
[1]->iim_val
= so
->inputs
[i
].inloc
+ j
;
3084 } else if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3085 unsigned i
= instr
->prefetch
.input_offset
/ 4;
3086 unsigned j
= instr
->prefetch
.input_offset
% 4;
3087 instr
->prefetch
.input_offset
= so
->inputs
[i
].inloc
+ j
;
3094 setup_output(struct ir3_context
*ctx
, nir_variable
*out
)
3096 struct ir3_shader_variant
*so
= ctx
->so
;
3097 unsigned slots
= glsl_count_vec4_slots(out
->type
, false, false);
3098 unsigned ncomp
= glsl_get_components(glsl_without_array(out
->type
));
3099 unsigned n
= out
->data
.driver_location
;
3100 unsigned frac
= out
->data
.location_frac
;
3101 unsigned slot
= out
->data
.location
;
3103 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
3105 case FRAG_RESULT_DEPTH
:
3106 so
->writes_pos
= true;
3108 case FRAG_RESULT_COLOR
:
3111 case FRAG_RESULT_SAMPLE_MASK
:
3112 so
->writes_smask
= true;
3115 slot
+= out
->data
.index
; /* For dual-src blend */
3116 if (slot
>= FRAG_RESULT_DATA0
)
3118 ir3_context_error(ctx
, "unknown FS output name: %s\n",
3119 gl_frag_result_name(slot
));
3121 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
3122 ctx
->so
->type
== MESA_SHADER_TESS_EVAL
||
3123 ctx
->so
->type
== MESA_SHADER_GEOMETRY
) {
3125 case VARYING_SLOT_POS
:
3126 so
->writes_pos
= true;
3128 case VARYING_SLOT_PSIZ
:
3129 so
->writes_psize
= true;
3131 case VARYING_SLOT_PRIMITIVE_ID
:
3132 case VARYING_SLOT_LAYER
:
3133 case VARYING_SLOT_GS_VERTEX_FLAGS_IR3
:
3134 debug_assert(ctx
->so
->type
== MESA_SHADER_GEOMETRY
);
3136 case VARYING_SLOT_COL0
:
3137 case VARYING_SLOT_COL1
:
3138 case VARYING_SLOT_BFC0
:
3139 case VARYING_SLOT_BFC1
:
3140 case VARYING_SLOT_FOGC
:
3141 case VARYING_SLOT_CLIP_DIST0
:
3142 case VARYING_SLOT_CLIP_DIST1
:
3143 case VARYING_SLOT_CLIP_VERTEX
:
3146 if (slot
>= VARYING_SLOT_VAR0
)
3148 if ((VARYING_SLOT_TEX0
<= slot
) && (slot
<= VARYING_SLOT_TEX7
))
3150 ir3_context_error(ctx
, "unknown %s shader output name: %s\n",
3151 _mesa_shader_stage_to_string(ctx
->so
->type
),
3152 gl_varying_slot_name(slot
));
3154 } else if (ctx
->so
->type
== MESA_SHADER_TESS_CTRL
) {
3155 /* output lowered to buffer writes. */
3158 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
3162 so
->outputs_count
= out
->data
.driver_location
+ slots
;
3163 compile_assert(ctx
, so
->outputs_count
< ARRAY_SIZE(so
->outputs
));
3165 for (int i
= 0; i
< slots
; i
++) {
3166 int slot_base
= n
+ i
;
3167 so
->outputs
[slot_base
].slot
= slot
+ i
;
3169 for (int i
= 0; i
< ncomp
; i
++) {
3170 unsigned idx
= (slot_base
* 4) + i
+ frac
;
3171 compile_assert(ctx
, idx
< ctx
->noutputs
);
3172 ctx
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
3175 /* if varying packing doesn't happen, we could end up in a situation
3176 * with "holes" in the output, and since the per-generation code that
3177 * sets up varying linkage registers doesn't expect to have more than
3178 * one varying per vec4 slot, pad the holes.
3180 * Note that this should probably generate a performance warning of
3183 for (int i
= 0; i
< frac
; i
++) {
3184 unsigned idx
= (slot_base
* 4) + i
;
3185 if (!ctx
->outputs
[idx
]) {
3186 ctx
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
3193 emit_instructions(struct ir3_context
*ctx
)
3195 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->s
);
3197 ctx
->ninputs
= ctx
->s
->num_inputs
* 4;
3198 ctx
->noutputs
= ctx
->s
->num_outputs
* 4;
3199 ctx
->inputs
= rzalloc_array(ctx
, struct ir3_instruction
*, ctx
->ninputs
);
3200 ctx
->outputs
= rzalloc_array(ctx
, struct ir3_instruction
*, ctx
->noutputs
);
3202 ctx
->ir
= ir3_create(ctx
->compiler
, ctx
->so
->type
);
3204 /* Create inputs in first block: */
3205 ctx
->block
= get_block(ctx
, nir_start_block(fxn
));
3206 ctx
->in_block
= ctx
->block
;
3208 /* for fragment shader, the vcoord input register is used as the
3209 * base for bary.f varying fetch instrs:
3211 * TODO defer creating ctx->ij_pixel and corresponding sysvals
3212 * until emit_intrinsic when we know they are actually needed.
3213 * For now, we defer creating ctx->ij_centroid, etc, since we
3214 * only need ij_pixel for "old style" varying inputs (ie.
3217 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
3218 ctx
->ij_pixel
= create_input(ctx
, 0x3);
3222 nir_foreach_variable (var
, &ctx
->s
->inputs
) {
3223 setup_input(ctx
, var
);
3226 /* Defer add_sysval_input() stuff until after setup_inputs(),
3227 * because sysvals need to be appended after varyings:
3229 if (ctx
->ij_pixel
) {
3230 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
,
3231 0x3, ctx
->ij_pixel
);
3235 /* Tesselation shaders always need primitive ID for indexing the
3236 * BO. Geometry shaders don't always need it but when they do it has be
3237 * delivered and unclobbered in the VS. To make things easy, we always
3238 * make room for it in VS/DS.
3240 bool has_tess
= ctx
->so
->key
.tessellation
!= IR3_TESS_NONE
;
3241 bool has_gs
= ctx
->so
->key
.has_gs
;
3242 switch (ctx
->so
->type
) {
3243 case MESA_SHADER_VERTEX
:
3245 ctx
->tcs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_TCS_HEADER_IR3
, 0x1);
3246 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3247 } else if (has_gs
) {
3248 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3249 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3252 case MESA_SHADER_TESS_CTRL
:
3253 ctx
->tcs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_TCS_HEADER_IR3
, 0x1);
3254 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3256 case MESA_SHADER_TESS_EVAL
:
3258 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3259 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3261 case MESA_SHADER_GEOMETRY
:
3262 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3263 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3269 /* Setup outputs: */
3270 nir_foreach_variable (var
, &ctx
->s
->outputs
) {
3271 setup_output(ctx
, var
);
3274 /* Find # of samplers: */
3275 nir_foreach_variable (var
, &ctx
->s
->uniforms
) {
3276 ctx
->so
->num_samp
+= glsl_type_get_sampler_count(var
->type
);
3277 /* just assume that we'll be reading from images.. if it
3278 * is write-only we don't have to count it, but not sure
3279 * if there is a good way to know?
3281 ctx
->so
->num_samp
+= glsl_type_get_image_count(var
->type
);
3284 /* NOTE: need to do something more clever when we support >1 fxn */
3285 nir_foreach_register (reg
, &fxn
->registers
) {
3286 ir3_declare_array(ctx
, reg
);
3288 /* And emit the body: */
3290 emit_function(ctx
, fxn
);
3293 /* Fixup tex sampler state for astc/srgb workaround instructions. We
3294 * need to assign the tex state indexes for these after we know the
3298 fixup_astc_srgb(struct ir3_context
*ctx
)
3300 struct ir3_shader_variant
*so
= ctx
->so
;
3301 /* indexed by original tex idx, value is newly assigned alpha sampler
3302 * state tex idx. Zero is invalid since there is at least one sampler
3305 unsigned alt_tex_state
[16] = {0};
3306 unsigned tex_idx
= ctx
->max_texture_index
+ 1;
3309 so
->astc_srgb
.base
= tex_idx
;
3311 for (unsigned i
= 0; i
< ctx
->ir
->astc_srgb_count
; i
++) {
3312 struct ir3_instruction
*sam
= ctx
->ir
->astc_srgb
[i
];
3314 compile_assert(ctx
, sam
->cat5
.tex
< ARRAY_SIZE(alt_tex_state
));
3316 if (alt_tex_state
[sam
->cat5
.tex
] == 0) {
3317 /* assign new alternate/alpha tex state slot: */
3318 alt_tex_state
[sam
->cat5
.tex
] = tex_idx
++;
3319 so
->astc_srgb
.orig_idx
[idx
++] = sam
->cat5
.tex
;
3320 so
->astc_srgb
.count
++;
3323 sam
->cat5
.tex
= alt_tex_state
[sam
->cat5
.tex
];
3328 fixup_binning_pass(struct ir3_context
*ctx
)
3330 struct ir3_shader_variant
*so
= ctx
->so
;
3331 struct ir3
*ir
= ctx
->ir
;
3334 /* first pass, remove unused outputs from the IR level outputs: */
3335 for (i
= 0, j
= 0; i
< ir
->outputs_count
; i
++) {
3336 struct ir3_instruction
*out
= ir
->outputs
[i
];
3337 assert(out
->opc
== OPC_META_COLLECT
);
3338 unsigned outidx
= out
->collect
.outidx
;
3339 unsigned slot
= so
->outputs
[outidx
].slot
;
3341 /* throw away everything but first position/psize */
3342 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
3343 ir
->outputs
[j
] = ir
->outputs
[i
];
3347 ir
->outputs_count
= j
;
3349 /* second pass, cleanup the unused slots in ir3_shader_variant::outputs
3352 for (i
= 0, j
= 0; i
< so
->outputs_count
; i
++) {
3353 unsigned slot
= so
->outputs
[i
].slot
;
3355 /* throw away everything but first position/psize */
3356 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
3357 so
->outputs
[j
] = so
->outputs
[i
];
3359 /* fixup outidx to point to new output table entry: */
3360 struct ir3_instruction
*out
;
3361 foreach_output (out
, ir
) {
3362 if (out
->collect
.outidx
== i
) {
3363 out
->collect
.outidx
= j
;
3371 so
->outputs_count
= j
;
3375 collect_tex_prefetches(struct ir3_context
*ctx
, struct ir3
*ir
)
3379 /* Collect sampling instructions eligible for pre-dispatch. */
3380 foreach_block (block
, &ir
->block_list
) {
3381 foreach_instr_safe (instr
, &block
->instr_list
) {
3382 if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3383 assert(idx
< ARRAY_SIZE(ctx
->so
->sampler_prefetch
));
3384 struct ir3_sampler_prefetch
*fetch
=
3385 &ctx
->so
->sampler_prefetch
[idx
];
3388 if (instr
->flags
& IR3_INSTR_B
) {
3389 fetch
->cmd
= IR3_SAMPLER_BINDLESS_PREFETCH_CMD
;
3390 /* In bindless mode, the index is actually the base */
3391 fetch
->tex_id
= instr
->prefetch
.tex_base
;
3392 fetch
->samp_id
= instr
->prefetch
.samp_base
;
3393 fetch
->tex_bindless_id
= instr
->prefetch
.tex
;
3394 fetch
->samp_bindless_id
= instr
->prefetch
.samp
;
3396 fetch
->cmd
= IR3_SAMPLER_PREFETCH_CMD
;
3397 fetch
->tex_id
= instr
->prefetch
.tex
;
3398 fetch
->samp_id
= instr
->prefetch
.samp
;
3400 fetch
->wrmask
= instr
->regs
[0]->wrmask
;
3401 fetch
->dst
= instr
->regs
[0]->num
;
3402 fetch
->src
= instr
->prefetch
.input_offset
;
3405 MAX2(ctx
->so
->total_in
, instr
->prefetch
.input_offset
+ 2);
3407 /* Disable half precision until supported. */
3408 fetch
->half_precision
= !!(instr
->regs
[0]->flags
& IR3_REG_HALF
);
3410 /* Remove the prefetch placeholder instruction: */
3411 list_delinit(&instr
->node
);
3418 ir3_compile_shader_nir(struct ir3_compiler
*compiler
,
3419 struct ir3_shader_variant
*so
)
3421 struct ir3_context
*ctx
;
3423 int ret
= 0, max_bary
;
3427 ctx
= ir3_context_init(compiler
, so
);
3429 DBG("INIT failed!");
3434 emit_instructions(ctx
);
3437 DBG("EMIT failed!");
3442 ir
= so
->ir
= ctx
->ir
;
3444 assert((ctx
->noutputs
% 4) == 0);
3446 /* Setup IR level outputs, which are "collects" that gather
3447 * the scalar components of outputs.
3449 for (unsigned i
= 0; i
< ctx
->noutputs
; i
+= 4) {
3451 /* figure out the # of components written:
3453 * TODO do we need to handle holes, ie. if .x and .z
3454 * components written, but .y component not written?
3456 for (unsigned j
= 0; j
< 4; j
++) {
3457 if (!ctx
->outputs
[i
+ j
])
3462 /* Note that in some stages, like TCS, store_output is
3463 * lowered to memory writes, so no components of the
3464 * are "written" from the PoV of traditional store-
3465 * output instructions:
3470 struct ir3_instruction
*out
=
3471 ir3_create_collect(ctx
, &ctx
->outputs
[i
], ncomp
);
3474 assert(outidx
< so
->outputs_count
);
3476 /* stash index into so->outputs[] so we can map the
3477 * output back to slot/etc later:
3479 out
->collect
.outidx
= outidx
;
3481 array_insert(ir
, ir
->outputs
, out
);
3484 /* Set up the gs header as an output for the vertex shader so it won't
3485 * clobber it for the tess ctrl shader.
3487 * TODO this could probably be done more cleanly in a nir pass.
3489 if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
3490 (ctx
->so
->key
.has_gs
&& ctx
->so
->type
== MESA_SHADER_TESS_EVAL
)) {
3491 if (ctx
->primitive_id
) {
3492 unsigned n
= so
->outputs_count
++;
3493 so
->outputs
[n
].slot
= VARYING_SLOT_PRIMITIVE_ID
;
3495 struct ir3_instruction
*out
=
3496 ir3_create_collect(ctx
, &ctx
->primitive_id
, 1);
3497 out
->collect
.outidx
= n
;
3498 array_insert(ir
, ir
->outputs
, out
);
3501 if (ctx
->gs_header
) {
3502 unsigned n
= so
->outputs_count
++;
3503 so
->outputs
[n
].slot
= VARYING_SLOT_GS_HEADER_IR3
;
3504 struct ir3_instruction
*out
=
3505 ir3_create_collect(ctx
, &ctx
->gs_header
, 1);
3506 out
->collect
.outidx
= n
;
3507 array_insert(ir
, ir
->outputs
, out
);
3510 if (ctx
->tcs_header
) {
3511 unsigned n
= so
->outputs_count
++;
3512 so
->outputs
[n
].slot
= VARYING_SLOT_TCS_HEADER_IR3
;
3513 struct ir3_instruction
*out
=
3514 ir3_create_collect(ctx
, &ctx
->tcs_header
, 1);
3515 out
->collect
.outidx
= n
;
3516 array_insert(ir
, ir
->outputs
, out
);
3520 /* at this point, for binning pass, throw away unneeded outputs: */
3521 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
< 600))
3522 fixup_binning_pass(ctx
);
3524 ir3_debug_print(ir
, "BEFORE CF");
3528 ir3_debug_print(ir
, "BEFORE CP");
3532 /* at this point, for binning pass, throw away unneeded outputs:
3533 * Note that for a6xx and later, we do this after ir3_cp to ensure
3534 * that the uniform/constant layout for BS and VS matches, so that
3535 * we can re-use same VS_CONST state group.
3537 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
>= 600))
3538 fixup_binning_pass(ctx
);
3540 /* for a6xx+, binning and draw pass VS use same VBO state, so we
3541 * need to make sure not to remove any inputs that are used by
3542 * the nonbinning VS.
3544 if (ctx
->compiler
->gpu_id
>= 600 && so
->binning_pass
&&
3545 so
->type
== MESA_SHADER_VERTEX
) {
3546 for (int i
= 0; i
< ctx
->ninputs
; i
++) {
3547 struct ir3_instruction
*in
= ctx
->inputs
[i
];
3555 debug_assert(n
< so
->nonbinning
->inputs_count
);
3557 if (so
->nonbinning
->inputs
[n
].sysval
)
3560 /* be sure to keep inputs, even if only used in VS */
3561 if (so
->nonbinning
->inputs
[n
].compmask
& (1 << c
))
3562 array_insert(in
->block
, in
->block
->keeps
, in
);
3566 ir3_debug_print(ir
, "BEFORE GROUPING");
3568 ir3_sched_add_deps(ir
);
3570 /* Group left/right neighbors, inserting mov's where needed to
3575 ir3_debug_print(ir
, "AFTER GROUPING");
3579 ir3_debug_print(ir
, "AFTER DCE");
3581 /* do Sethi–Ullman numbering before scheduling: */
3584 ret
= ir3_sched(ir
);
3586 DBG("SCHED failed!");
3590 ir3_debug_print(ir
, "AFTER SCHED");
3592 /* Pre-assign VS inputs on a6xx+ binning pass shader, to align
3593 * with draw pass VS, so binning and draw pass can both use the
3596 * Note that VS inputs are expected to be full precision.
3598 bool pre_assign_inputs
= (ir
->compiler
->gpu_id
>= 600) &&
3599 (ir
->type
== MESA_SHADER_VERTEX
) &&
3602 if (pre_assign_inputs
) {
3603 for (unsigned i
= 0; i
< ctx
->ninputs
; i
++) {
3604 struct ir3_instruction
*instr
= ctx
->inputs
[i
];
3611 unsigned regid
= so
->nonbinning
->inputs
[n
].regid
+ c
;
3613 instr
->regs
[0]->num
= regid
;
3616 ret
= ir3_ra(so
, ctx
->inputs
, ctx
->ninputs
);
3617 } else if (ctx
->tcs_header
) {
3618 /* We need to have these values in the same registers between VS and TCS
3619 * since the VS chains to TCS and doesn't get the sysvals redelivered.
3622 ctx
->tcs_header
->regs
[0]->num
= regid(0, 0);
3623 ctx
->primitive_id
->regs
[0]->num
= regid(0, 1);
3624 struct ir3_instruction
*precolor
[] = { ctx
->tcs_header
, ctx
->primitive_id
};
3625 ret
= ir3_ra(so
, precolor
, ARRAY_SIZE(precolor
));
3626 } else if (ctx
->gs_header
) {
3627 /* We need to have these values in the same registers between producer
3628 * (VS or DS) and GS since the producer chains to GS and doesn't get
3629 * the sysvals redelivered.
3632 ctx
->gs_header
->regs
[0]->num
= regid(0, 0);
3633 ctx
->primitive_id
->regs
[0]->num
= regid(0, 1);
3634 struct ir3_instruction
*precolor
[] = { ctx
->gs_header
, ctx
->primitive_id
};
3635 ret
= ir3_ra(so
, precolor
, ARRAY_SIZE(precolor
));
3636 } else if (so
->num_sampler_prefetch
) {
3637 assert(so
->type
== MESA_SHADER_FRAGMENT
);
3638 struct ir3_instruction
*instr
, *precolor
[2];
3641 foreach_input (instr
, ir
) {
3642 if (instr
->input
.sysval
!= SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
)
3645 assert(idx
< ARRAY_SIZE(precolor
));
3647 precolor
[idx
] = instr
;
3648 instr
->regs
[0]->num
= idx
;
3652 ret
= ir3_ra(so
, precolor
, idx
);
3654 ret
= ir3_ra(so
, NULL
, 0);
3663 ir3_debug_print(ir
, "AFTER POSTSCHED");
3665 if (compiler
->gpu_id
>= 600) {
3666 if (ir3_a6xx_fixup_atomic_dests(ir
, so
)) {
3667 ir3_debug_print(ir
, "AFTER ATOMIC FIXUP");
3671 if (so
->type
== MESA_SHADER_FRAGMENT
)
3675 * Fixup inputs/outputs to point to the actual registers assigned:
3677 * 1) initialize to r63.x (invalid/unused)
3678 * 2) iterate IR level inputs/outputs and update the variants
3679 * inputs/outputs table based on the assigned registers for
3680 * the remaining inputs/outputs.
3683 for (unsigned i
= 0; i
< so
->inputs_count
; i
++)
3684 so
->inputs
[i
].regid
= INVALID_REG
;
3685 for (unsigned i
= 0; i
< so
->outputs_count
; i
++)
3686 so
->outputs
[i
].regid
= INVALID_REG
;
3688 struct ir3_instruction
*out
;
3689 foreach_output (out
, ir
) {
3690 assert(out
->opc
== OPC_META_COLLECT
);
3691 unsigned outidx
= out
->collect
.outidx
;
3693 so
->outputs
[outidx
].regid
= out
->regs
[0]->num
;
3694 so
->outputs
[outidx
].half
= !!(out
->regs
[0]->flags
& IR3_REG_HALF
);
3697 struct ir3_instruction
*in
;
3698 foreach_input (in
, ir
) {
3699 assert(in
->opc
== OPC_META_INPUT
);
3700 unsigned inidx
= in
->input
.inidx
;
3702 if (pre_assign_inputs
&& !so
->inputs
[inidx
].sysval
) {
3703 if (VALIDREG(so
->nonbinning
->inputs
[inidx
].regid
)) {
3704 compile_assert(ctx
, in
->regs
[0]->num
==
3705 so
->nonbinning
->inputs
[inidx
].regid
);
3706 compile_assert(ctx
, !!(in
->regs
[0]->flags
& IR3_REG_HALF
) ==
3707 so
->nonbinning
->inputs
[inidx
].half
);
3709 so
->inputs
[inidx
].regid
= so
->nonbinning
->inputs
[inidx
].regid
;
3710 so
->inputs
[inidx
].half
= so
->nonbinning
->inputs
[inidx
].half
;
3712 so
->inputs
[inidx
].regid
= in
->regs
[0]->num
;
3713 so
->inputs
[inidx
].half
= !!(in
->regs
[0]->flags
& IR3_REG_HALF
);
3718 fixup_astc_srgb(ctx
);
3720 /* We need to do legalize after (for frag shader's) the "bary.f"
3721 * offsets (inloc) have been assigned.
3723 ir3_legalize(ir
, so
, &max_bary
);
3725 ir3_debug_print(ir
, "AFTER LEGALIZE");
3727 /* Set (ss)(sy) on first TCS and GEOMETRY instructions, since we don't
3728 * know what we might have to wait on when coming in from VS chsh.
3730 if (so
->type
== MESA_SHADER_TESS_CTRL
||
3731 so
->type
== MESA_SHADER_GEOMETRY
) {
3732 foreach_block (block
, &ir
->block_list
) {
3733 foreach_instr (instr
, &block
->instr_list
) {
3734 instr
->flags
|= IR3_INSTR_SS
| IR3_INSTR_SY
;
3740 so
->branchstack
= ctx
->max_stack
;
3742 /* Note that actual_in counts inputs that are not bary.f'd for FS: */
3743 if (so
->type
== MESA_SHADER_FRAGMENT
)
3744 so
->total_in
= max_bary
+ 1;
3746 so
->max_sun
= ir
->max_sun
;
3748 /* Collect sampling instructions eligible for pre-dispatch. */
3749 collect_tex_prefetches(ctx
, ir
);
3751 if (so
->type
== MESA_SHADER_FRAGMENT
&&
3752 ctx
->s
->info
.fs
.needs_helper_invocations
)
3753 so
->need_pixlod
= true;
3758 ir3_destroy(so
->ir
);
3761 ir3_context_free(ctx
);