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
;
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
);
889 /* Combine groups of consecutive enabled channels in one write
890 * message. We use ffs to find the first enabled channel and then ffs on
891 * the bit-inverse, down-shifted writemask to determine the length of
892 * the block of enabled bits.
894 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
897 unsigned first_component
= ffs(wrmask
) - 1;
898 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
900 stl
= ir3_STL(b
, offset
, 0,
901 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
902 create_immed(b
, length
), 0);
903 stl
->cat6
.dst_offset
= first_component
+ base
;
904 stl
->cat6
.type
= utype_src(intr
->src
[0]);
905 stl
->barrier_class
= IR3_BARRIER_SHARED_W
;
906 stl
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
908 array_insert(b
, b
->keeps
, stl
);
910 /* Clear the bits in the writemask that we just wrote, then try
911 * again to see if more channels are left.
913 wrmask
&= (15 << (first_component
+ length
));
917 /* src[] = { offset }. const_index[] = { base } */
919 emit_intrinsic_load_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
920 struct ir3_instruction
**dst
)
922 struct ir3_block
*b
= ctx
->block
;
923 struct ir3_instruction
*load
, *offset
;
926 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
927 base
= nir_intrinsic_base(intr
);
929 load
= ir3_LDLW(b
, offset
, 0,
930 create_immed(b
, intr
->num_components
), 0,
931 create_immed(b
, base
), 0);
933 load
->cat6
.type
= utype_dst(intr
->dest
);
934 load
->regs
[0]->wrmask
= MASK(intr
->num_components
);
936 load
->barrier_class
= IR3_BARRIER_SHARED_R
;
937 load
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
939 ir3_split_dest(b
, dst
, load
, 0, intr
->num_components
);
942 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
944 emit_intrinsic_store_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
946 struct ir3_block
*b
= ctx
->block
;
947 struct ir3_instruction
*store
, *offset
;
948 struct ir3_instruction
* const *value
;
949 unsigned base
, wrmask
;
951 value
= ir3_get_src(ctx
, &intr
->src
[0]);
952 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
954 base
= nir_intrinsic_base(intr
);
955 wrmask
= nir_intrinsic_write_mask(intr
);
957 /* Combine groups of consecutive enabled channels in one write
958 * message. We use ffs to find the first enabled channel and then ffs on
959 * the bit-inverse, down-shifted writemask to determine the length of
960 * the block of enabled bits.
962 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
965 unsigned first_component
= ffs(wrmask
) - 1;
966 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
968 store
= ir3_STLW(b
, offset
, 0,
969 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
970 create_immed(b
, length
), 0);
972 store
->cat6
.dst_offset
= first_component
+ base
;
973 store
->cat6
.type
= utype_src(intr
->src
[0]);
974 store
->barrier_class
= IR3_BARRIER_SHARED_W
;
975 store
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
977 array_insert(b
, b
->keeps
, store
);
979 /* Clear the bits in the writemask that we just wrote, then try
980 * again to see if more channels are left.
982 wrmask
&= (15 << (first_component
+ length
));
987 * CS shared variable atomic intrinsics
989 * All of the shared variable atomic memory operations read a value from
990 * memory, compute a new value using one of the operations below, write the
991 * new value to memory, and return the original value read.
993 * All operations take 2 sources except CompSwap that takes 3. These
996 * 0: The offset into the shared variable storage region that the atomic
997 * operation will operate on.
998 * 1: The data parameter to the atomic function (i.e. the value to add
999 * in shared_atomic_add, etc).
1000 * 2: For CompSwap only: the second data parameter.
1002 static struct ir3_instruction
*
1003 emit_intrinsic_atomic_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1005 struct ir3_block
*b
= ctx
->block
;
1006 struct ir3_instruction
*atomic
, *src0
, *src1
;
1007 type_t type
= TYPE_U32
;
1009 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0]; /* offset */
1010 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0]; /* value */
1012 switch (intr
->intrinsic
) {
1013 case nir_intrinsic_shared_atomic_add
:
1014 atomic
= ir3_ATOMIC_ADD(b
, src0
, 0, src1
, 0);
1016 case nir_intrinsic_shared_atomic_imin
:
1017 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
1020 case nir_intrinsic_shared_atomic_umin
:
1021 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
1023 case nir_intrinsic_shared_atomic_imax
:
1024 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
1027 case nir_intrinsic_shared_atomic_umax
:
1028 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
1030 case nir_intrinsic_shared_atomic_and
:
1031 atomic
= ir3_ATOMIC_AND(b
, src0
, 0, src1
, 0);
1033 case nir_intrinsic_shared_atomic_or
:
1034 atomic
= ir3_ATOMIC_OR(b
, src0
, 0, src1
, 0);
1036 case nir_intrinsic_shared_atomic_xor
:
1037 atomic
= ir3_ATOMIC_XOR(b
, src0
, 0, src1
, 0);
1039 case nir_intrinsic_shared_atomic_exchange
:
1040 atomic
= ir3_ATOMIC_XCHG(b
, src0
, 0, src1
, 0);
1042 case nir_intrinsic_shared_atomic_comp_swap
:
1043 /* for cmpxchg, src1 is [ui]vec2(data, compare): */
1044 src1
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1045 ir3_get_src(ctx
, &intr
->src
[2])[0],
1048 atomic
= ir3_ATOMIC_CMPXCHG(b
, src0
, 0, src1
, 0);
1054 atomic
->cat6
.iim_val
= 1;
1056 atomic
->cat6
.type
= type
;
1057 atomic
->barrier_class
= IR3_BARRIER_SHARED_W
;
1058 atomic
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
1060 /* even if nothing consume the result, we can't DCE the instruction: */
1061 array_insert(b
, b
->keeps
, atomic
);
1066 struct tex_src_info
{
1068 unsigned tex_base
, samp_base
, tex_idx
, samp_idx
;
1069 /* For normal tex instructions */
1070 unsigned base
, combined_idx
, a1_val
, flags
;
1071 struct ir3_instruction
*samp_tex
;
1074 /* TODO handle actual indirect/dynamic case.. which is going to be weird
1075 * to handle with the image_mapping table..
1077 static struct tex_src_info
1078 get_image_samp_tex_src(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1080 struct ir3_block
*b
= ctx
->block
;
1081 struct tex_src_info info
= { 0 };
1082 nir_intrinsic_instr
*bindless_tex
= ir3_bindless_resource(intr
->src
[0]);
1083 ctx
->so
->bindless_tex
= true;
1087 info
.flags
|= IR3_INSTR_B
;
1089 /* Gather information required to determine which encoding to
1090 * choose as well as for prefetch.
1092 info
.tex_base
= nir_intrinsic_desc_set(bindless_tex
);
1093 bool tex_const
= nir_src_is_const(bindless_tex
->src
[0]);
1095 info
.tex_idx
= nir_src_as_uint(bindless_tex
->src
[0]);
1098 /* Choose encoding. */
1099 if (tex_const
&& info
.tex_idx
< 256) {
1100 if (info
.tex_idx
< 16) {
1101 /* Everything fits within the instruction */
1102 info
.base
= info
.tex_base
;
1103 info
.combined_idx
= info
.samp_idx
| (info
.tex_idx
<< 4);
1105 info
.base
= info
.tex_base
;
1106 info
.a1_val
= info
.tex_idx
<< 3;
1107 info
.combined_idx
= 0;
1108 info
.flags
|= IR3_INSTR_A1EN
;
1110 info
.samp_tex
= NULL
;
1112 info
.flags
|= IR3_INSTR_S2EN
;
1113 info
.base
= info
.tex_base
;
1115 /* Note: the indirect source is now a vec2 instead of hvec2 */
1116 struct ir3_instruction
*texture
, *sampler
;
1118 texture
= ir3_get_src(ctx
, &intr
->src
[0])[0];
1119 sampler
= create_immed(b
, 0);
1120 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1126 info
.flags
|= IR3_INSTR_S2EN
;
1127 unsigned slot
= nir_src_as_uint(intr
->src
[0]);
1128 unsigned tex_idx
= ir3_image_to_tex(&ctx
->so
->image_mapping
, slot
);
1129 struct ir3_instruction
*texture
, *sampler
;
1131 texture
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1132 sampler
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1134 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1143 static struct ir3_instruction
*
1144 emit_sam(struct ir3_context
*ctx
, opc_t opc
, struct tex_src_info info
,
1145 type_t type
, unsigned wrmask
, struct ir3_instruction
*src0
,
1146 struct ir3_instruction
*src1
)
1148 struct ir3_instruction
*sam
, *addr
;
1149 if (info
.flags
& IR3_INSTR_A1EN
) {
1150 addr
= ir3_get_addr1(ctx
, info
.a1_val
);
1152 sam
= ir3_SAM(ctx
->block
, opc
, type
, 0b1111, info
.flags
,
1153 info
.samp_tex
, src0
, src1
);
1154 if (info
.flags
& IR3_INSTR_A1EN
) {
1155 ir3_instr_set_address(sam
, addr
);
1157 if (info
.flags
& IR3_INSTR_B
) {
1158 sam
->cat5
.tex_base
= info
.base
;
1159 sam
->cat5
.samp
= info
.combined_idx
;
1164 /* src[] = { deref, coord, sample_index }. const_index[] = {} */
1166 emit_intrinsic_load_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1167 struct ir3_instruction
**dst
)
1169 struct ir3_block
*b
= ctx
->block
;
1170 struct tex_src_info info
= get_image_samp_tex_src(ctx
, intr
);
1171 struct ir3_instruction
*sam
;
1172 struct ir3_instruction
* const *src0
= ir3_get_src(ctx
, &intr
->src
[1]);
1173 struct ir3_instruction
*coords
[4];
1174 unsigned flags
, ncoords
= ir3_get_image_coords(intr
, &flags
);
1175 type_t type
= ir3_get_type_for_image_intrinsic(intr
);
1177 /* hmm, this seems a bit odd, but it is what blob does and (at least
1178 * a5xx) just faults on bogus addresses otherwise:
1180 if (flags
& IR3_INSTR_3D
) {
1181 flags
&= ~IR3_INSTR_3D
;
1182 flags
|= IR3_INSTR_A
;
1184 info
.flags
|= flags
;
1186 for (unsigned i
= 0; i
< ncoords
; i
++)
1187 coords
[i
] = src0
[i
];
1190 coords
[ncoords
++] = create_immed(b
, 0);
1192 sam
= emit_sam(ctx
, OPC_ISAM
, info
, type
, 0b1111,
1193 ir3_create_collect(ctx
, coords
, ncoords
), NULL
);
1195 sam
->barrier_class
= IR3_BARRIER_IMAGE_R
;
1196 sam
->barrier_conflict
= IR3_BARRIER_IMAGE_W
;
1198 ir3_split_dest(b
, dst
, sam
, 0, 4);
1202 emit_intrinsic_image_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1203 struct ir3_instruction
**dst
)
1205 struct ir3_block
*b
= ctx
->block
;
1206 struct tex_src_info info
= get_image_samp_tex_src(ctx
, intr
);
1207 struct ir3_instruction
*sam
, *lod
;
1208 unsigned flags
, ncoords
= ir3_get_image_coords(intr
, &flags
);
1209 type_t dst_type
= nir_dest_bit_size(intr
->dest
) == 16 ?
1210 TYPE_U16
: TYPE_U32
;
1212 info
.flags
|= flags
;
1213 lod
= create_immed(b
, 0);
1214 sam
= emit_sam(ctx
, OPC_GETSIZE
, info
, dst_type
, 0b1111, lod
, NULL
);
1216 /* Array size actually ends up in .w rather than .z. This doesn't
1217 * matter for miplevel 0, but for higher mips the value in z is
1218 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
1219 * returned, which means that we have to add 1 to it for arrays for
1222 * Note use a temporary dst and then copy, since the size of the dst
1223 * array that is passed in is based on nir's understanding of the
1224 * result size, not the hardware's
1226 struct ir3_instruction
*tmp
[4];
1228 ir3_split_dest(b
, tmp
, sam
, 0, 4);
1230 /* get_size instruction returns size in bytes instead of texels
1231 * for imageBuffer, so we need to divide it by the pixel size
1232 * of the image format.
1234 * TODO: This is at least true on a5xx. Check other gens.
1236 if (nir_intrinsic_image_dim(intr
) == GLSL_SAMPLER_DIM_BUF
) {
1237 /* Since all the possible values the divisor can take are
1238 * power-of-two (4, 8, or 16), the division is implemented
1240 * During shader setup, the log2 of the image format's
1241 * bytes-per-pixel should have been emitted in 2nd slot of
1242 * image_dims. See ir3_shader::emit_image_dims().
1244 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
1245 unsigned cb
= regid(const_state
->offsets
.image_dims
, 0) +
1246 const_state
->image_dims
.off
[nir_src_as_uint(intr
->src
[0])];
1247 struct ir3_instruction
*aux
= create_uniform(b
, cb
+ 1);
1249 tmp
[0] = ir3_SHR_B(b
, tmp
[0], 0, aux
, 0);
1252 for (unsigned i
= 0; i
< ncoords
; i
++)
1255 if (flags
& IR3_INSTR_A
) {
1256 if (ctx
->compiler
->levels_add_one
) {
1257 dst
[ncoords
-1] = ir3_ADD_U(b
, tmp
[3], 0, create_immed(b
, 1), 0);
1259 dst
[ncoords
-1] = ir3_MOV(b
, tmp
[3], TYPE_U32
);
1265 emit_intrinsic_barrier(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1267 struct ir3_block
*b
= ctx
->block
;
1268 struct ir3_instruction
*barrier
;
1270 switch (intr
->intrinsic
) {
1271 case nir_intrinsic_control_barrier
:
1272 barrier
= ir3_BAR(b
);
1273 barrier
->cat7
.g
= true;
1274 barrier
->cat7
.l
= true;
1275 barrier
->flags
= IR3_INSTR_SS
| IR3_INSTR_SY
;
1276 barrier
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1278 case nir_intrinsic_memory_barrier
:
1279 barrier
= ir3_FENCE(b
);
1280 barrier
->cat7
.g
= true;
1281 barrier
->cat7
.r
= true;
1282 barrier
->cat7
.w
= true;
1283 barrier
->cat7
.l
= true;
1284 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
|
1285 IR3_BARRIER_BUFFER_W
;
1286 barrier
->barrier_conflict
=
1287 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1288 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1290 case nir_intrinsic_memory_barrier_buffer
:
1291 barrier
= ir3_FENCE(b
);
1292 barrier
->cat7
.g
= true;
1293 barrier
->cat7
.r
= true;
1294 barrier
->cat7
.w
= true;
1295 barrier
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1296 barrier
->barrier_conflict
= IR3_BARRIER_BUFFER_R
|
1297 IR3_BARRIER_BUFFER_W
;
1299 case nir_intrinsic_memory_barrier_image
:
1300 // TODO double check if this should have .g set
1301 barrier
= ir3_FENCE(b
);
1302 barrier
->cat7
.g
= true;
1303 barrier
->cat7
.r
= true;
1304 barrier
->cat7
.w
= true;
1305 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1306 barrier
->barrier_conflict
= IR3_BARRIER_IMAGE_R
|
1307 IR3_BARRIER_IMAGE_W
;
1309 case nir_intrinsic_memory_barrier_shared
:
1310 barrier
= ir3_FENCE(b
);
1311 barrier
->cat7
.g
= true;
1312 barrier
->cat7
.l
= true;
1313 barrier
->cat7
.r
= true;
1314 barrier
->cat7
.w
= true;
1315 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
;
1316 barrier
->barrier_conflict
= IR3_BARRIER_SHARED_R
|
1317 IR3_BARRIER_SHARED_W
;
1319 case nir_intrinsic_group_memory_barrier
:
1320 barrier
= ir3_FENCE(b
);
1321 barrier
->cat7
.g
= true;
1322 barrier
->cat7
.l
= true;
1323 barrier
->cat7
.r
= true;
1324 barrier
->cat7
.w
= true;
1325 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
|
1326 IR3_BARRIER_IMAGE_W
|
1327 IR3_BARRIER_BUFFER_W
;
1328 barrier
->barrier_conflict
=
1329 IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
|
1330 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1331 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1337 /* make sure barrier doesn't get DCE'd */
1338 array_insert(b
, b
->keeps
, barrier
);
1341 static void add_sysval_input_compmask(struct ir3_context
*ctx
,
1342 gl_system_value slot
, unsigned compmask
,
1343 struct ir3_instruction
*instr
)
1345 struct ir3_shader_variant
*so
= ctx
->so
;
1346 unsigned n
= so
->inputs_count
++;
1348 assert(instr
->opc
== OPC_META_INPUT
);
1349 instr
->input
.inidx
= n
;
1350 instr
->input
.sysval
= slot
;
1352 so
->inputs
[n
].sysval
= true;
1353 so
->inputs
[n
].slot
= slot
;
1354 so
->inputs
[n
].compmask
= compmask
;
1355 so
->inputs
[n
].interpolate
= INTERP_MODE_FLAT
;
1359 static struct ir3_instruction
*
1360 create_sysval_input(struct ir3_context
*ctx
, gl_system_value slot
,
1364 struct ir3_instruction
*sysval
= create_input(ctx
, compmask
);
1365 add_sysval_input_compmask(ctx
, slot
, compmask
, sysval
);
1369 static struct ir3_instruction
*
1370 get_barycentric_centroid(struct ir3_context
*ctx
)
1372 if (!ctx
->ij_centroid
) {
1373 struct ir3_instruction
*xy
[2];
1374 struct ir3_instruction
*ij
;
1376 ij
= create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID
, 0x3);
1377 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1379 ctx
->ij_centroid
= ir3_create_collect(ctx
, xy
, 2);
1382 return ctx
->ij_centroid
;
1385 static struct ir3_instruction
*
1386 get_barycentric_sample(struct ir3_context
*ctx
)
1388 if (!ctx
->ij_sample
) {
1389 struct ir3_instruction
*xy
[2];
1390 struct ir3_instruction
*ij
;
1392 ij
= create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE
, 0x3);
1393 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1395 ctx
->ij_sample
= ir3_create_collect(ctx
, xy
, 2);
1398 return ctx
->ij_sample
;
1401 static struct ir3_instruction
*
1402 get_barycentric_pixel(struct ir3_context
*ctx
)
1404 /* TODO when tgsi_to_nir supports "new-style" FS inputs switch
1405 * this to create ij_pixel only on demand:
1407 return ctx
->ij_pixel
;
1410 static struct ir3_instruction
*
1411 get_frag_coord(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1413 if (!ctx
->frag_coord
) {
1414 struct ir3_block
*b
= ctx
->in_block
;
1415 struct ir3_instruction
*xyzw
[4];
1416 struct ir3_instruction
*hw_frag_coord
;
1418 hw_frag_coord
= create_sysval_input(ctx
, SYSTEM_VALUE_FRAG_COORD
, 0xf);
1419 ir3_split_dest(b
, xyzw
, hw_frag_coord
, 0, 4);
1421 /* for frag_coord.xy, we get unsigned values.. we need
1422 * to subtract (integer) 8 and divide by 16 (right-
1423 * shift by 4) then convert to float:
1427 * mov.u32f32 dst, tmp
1430 for (int i
= 0; i
< 2; i
++) {
1431 xyzw
[i
] = ir3_COV(b
, xyzw
[i
], TYPE_U32
, TYPE_F32
);
1432 xyzw
[i
] = ir3_MUL_F(b
, xyzw
[i
], 0, create_immed(b
, fui(1.0 / 16.0)), 0);
1435 ctx
->frag_coord
= ir3_create_collect(ctx
, xyzw
, 4);
1438 ctx
->so
->fragcoord_compmask
|=
1439 nir_ssa_def_components_read(&intr
->dest
.ssa
);
1441 return ctx
->frag_coord
;
1445 emit_intrinsic(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1447 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1448 struct ir3_instruction
**dst
;
1449 struct ir3_instruction
* const *src
;
1450 struct ir3_block
*b
= ctx
->block
;
1453 if (info
->has_dest
) {
1454 unsigned n
= nir_intrinsic_dest_components(intr
);
1455 dst
= ir3_get_dst(ctx
, &intr
->dest
, n
);
1460 const unsigned primitive_param
= ctx
->so
->shader
->const_state
.offsets
.primitive_param
* 4;
1461 const unsigned primitive_map
= ctx
->so
->shader
->const_state
.offsets
.primitive_map
* 4;
1463 switch (intr
->intrinsic
) {
1464 case nir_intrinsic_load_uniform
:
1465 idx
= nir_intrinsic_base(intr
);
1466 if (nir_src_is_const(intr
->src
[0])) {
1467 idx
+= nir_src_as_uint(intr
->src
[0]);
1468 for (int i
= 0; i
< intr
->num_components
; i
++) {
1469 dst
[i
] = create_uniform_typed(b
, idx
+ i
,
1470 nir_dest_bit_size(intr
->dest
) == 16 ? TYPE_F16
: TYPE_F32
);
1473 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1474 for (int i
= 0; i
< intr
->num_components
; i
++) {
1475 dst
[i
] = create_uniform_indirect(b
, idx
+ i
,
1476 ir3_get_addr0(ctx
, src
[0], 1));
1478 /* NOTE: if relative addressing is used, we set
1479 * constlen in the compiler (to worst-case value)
1480 * since we don't know in the assembler what the max
1481 * addr reg value can be:
1483 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
1484 ctx
->so
->shader
->ubo_state
.size
/ 16);
1488 case nir_intrinsic_load_vs_primitive_stride_ir3
:
1489 dst
[0] = create_uniform(b
, primitive_param
+ 0);
1491 case nir_intrinsic_load_vs_vertex_stride_ir3
:
1492 dst
[0] = create_uniform(b
, primitive_param
+ 1);
1494 case nir_intrinsic_load_hs_patch_stride_ir3
:
1495 dst
[0] = create_uniform(b
, primitive_param
+ 2);
1497 case nir_intrinsic_load_patch_vertices_in
:
1498 dst
[0] = create_uniform(b
, primitive_param
+ 3);
1500 case nir_intrinsic_load_tess_param_base_ir3
:
1501 dst
[0] = create_uniform(b
, primitive_param
+ 4);
1502 dst
[1] = create_uniform(b
, primitive_param
+ 5);
1504 case nir_intrinsic_load_tess_factor_base_ir3
:
1505 dst
[0] = create_uniform(b
, primitive_param
+ 6);
1506 dst
[1] = create_uniform(b
, primitive_param
+ 7);
1509 case nir_intrinsic_load_primitive_location_ir3
:
1510 idx
= nir_intrinsic_driver_location(intr
);
1511 dst
[0] = create_uniform(b
, primitive_map
+ idx
);
1514 case nir_intrinsic_load_gs_header_ir3
:
1515 dst
[0] = ctx
->gs_header
;
1517 case nir_intrinsic_load_tcs_header_ir3
:
1518 dst
[0] = ctx
->tcs_header
;
1521 case nir_intrinsic_load_primitive_id
:
1522 dst
[0] = ctx
->primitive_id
;
1525 case nir_intrinsic_load_tess_coord
:
1526 if (!ctx
->tess_coord
) {
1528 create_sysval_input(ctx
, SYSTEM_VALUE_TESS_COORD
, 0x3);
1530 ir3_split_dest(b
, dst
, ctx
->tess_coord
, 0, 2);
1532 /* Unused, but ir3_put_dst() below wants to free something */
1533 dst
[2] = create_immed(b
, 0);
1536 case nir_intrinsic_end_patch_ir3
:
1537 assert(ctx
->so
->type
== MESA_SHADER_TESS_CTRL
);
1538 struct ir3_instruction
*end
= ir3_PREDE(b
);
1539 array_insert(b
, b
->keeps
, end
);
1541 end
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1542 end
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
1545 case nir_intrinsic_store_global_ir3
: {
1546 struct ir3_instruction
*value
, *addr
, *offset
;
1548 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1549 ir3_get_src(ctx
, &intr
->src
[1])[0],
1550 ir3_get_src(ctx
, &intr
->src
[1])[1]
1553 offset
= ir3_get_src(ctx
, &intr
->src
[2])[0];
1555 value
= ir3_create_collect(ctx
, ir3_get_src(ctx
, &intr
->src
[0]),
1556 intr
->num_components
);
1558 struct ir3_instruction
*stg
=
1559 ir3_STG_G(ctx
->block
, addr
, 0, value
, 0,
1560 create_immed(ctx
->block
, intr
->num_components
), 0, offset
, 0);
1561 stg
->cat6
.type
= TYPE_U32
;
1562 stg
->cat6
.iim_val
= 1;
1564 array_insert(b
, b
->keeps
, stg
);
1566 stg
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1567 stg
->barrier_conflict
= IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1571 case nir_intrinsic_load_global_ir3
: {
1572 struct ir3_instruction
*addr
, *offset
;
1574 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1575 ir3_get_src(ctx
, &intr
->src
[0])[0],
1576 ir3_get_src(ctx
, &intr
->src
[0])[1]
1579 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
1581 struct ir3_instruction
*load
=
1582 ir3_LDG(b
, addr
, 0, create_immed(ctx
->block
, intr
->num_components
),
1584 load
->cat6
.type
= TYPE_U32
;
1585 load
->regs
[0]->wrmask
= MASK(intr
->num_components
);
1587 load
->barrier_class
= IR3_BARRIER_BUFFER_R
;
1588 load
->barrier_conflict
= IR3_BARRIER_BUFFER_W
;
1590 ir3_split_dest(b
, dst
, load
, 0, intr
->num_components
);
1594 case nir_intrinsic_load_ubo
:
1595 emit_intrinsic_load_ubo(ctx
, intr
, dst
);
1597 case nir_intrinsic_load_ubo_ir3
:
1598 emit_intrinsic_load_ubo_ldc(ctx
, intr
, dst
);
1600 case nir_intrinsic_load_frag_coord
:
1601 ir3_split_dest(b
, dst
, get_frag_coord(ctx
, intr
), 0, 4);
1603 case nir_intrinsic_load_sample_pos_from_id
: {
1604 /* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
1605 * but that doesn't seem necessary.
1607 struct ir3_instruction
*offset
=
1608 ir3_RGETPOS(b
, ir3_get_src(ctx
, &intr
->src
[0])[0], 0);
1609 offset
->regs
[0]->wrmask
= 0x3;
1610 offset
->cat5
.type
= TYPE_F32
;
1612 ir3_split_dest(b
, dst
, offset
, 0, 2);
1616 case nir_intrinsic_load_size_ir3
:
1617 if (!ctx
->ij_size
) {
1619 create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE
, 0x1);
1621 dst
[0] = ctx
->ij_size
;
1623 case nir_intrinsic_load_barycentric_centroid
:
1624 ir3_split_dest(b
, dst
, get_barycentric_centroid(ctx
), 0, 2);
1626 case nir_intrinsic_load_barycentric_sample
:
1627 if (ctx
->so
->key
.msaa
) {
1628 ir3_split_dest(b
, dst
, get_barycentric_sample(ctx
), 0, 2);
1630 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1633 case nir_intrinsic_load_barycentric_pixel
:
1634 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1636 case nir_intrinsic_load_interpolated_input
:
1637 idx
= nir_intrinsic_base(intr
);
1638 comp
= nir_intrinsic_component(intr
);
1639 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1640 if (nir_src_is_const(intr
->src
[1])) {
1641 struct ir3_instruction
*coord
= ir3_create_collect(ctx
, src
, 2);
1642 idx
+= nir_src_as_uint(intr
->src
[1]);
1643 for (int i
= 0; i
< intr
->num_components
; i
++) {
1644 unsigned inloc
= idx
* 4 + i
+ comp
;
1645 if (ctx
->so
->inputs
[idx
].bary
&&
1646 !ctx
->so
->inputs
[idx
].use_ldlv
) {
1647 dst
[i
] = ir3_BARY_F(b
, create_immed(b
, inloc
), 0, coord
, 0);
1649 /* for non-varyings use the pre-setup input, since
1650 * that is easier than mapping things back to a
1651 * nir_variable to figure out what it is.
1653 dst
[i
] = ctx
->inputs
[inloc
];
1654 compile_assert(ctx
, dst
[i
]);
1658 ir3_context_error(ctx
, "unhandled");
1661 case nir_intrinsic_load_input
:
1662 idx
= nir_intrinsic_base(intr
);
1663 comp
= nir_intrinsic_component(intr
);
1664 if (nir_src_is_const(intr
->src
[0])) {
1665 idx
+= nir_src_as_uint(intr
->src
[0]);
1666 for (int i
= 0; i
< intr
->num_components
; i
++) {
1667 unsigned n
= idx
* 4 + i
+ comp
;
1668 dst
[i
] = ctx
->inputs
[n
];
1669 compile_assert(ctx
, ctx
->inputs
[n
]);
1672 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1673 struct ir3_instruction
*collect
=
1674 ir3_create_collect(ctx
, ctx
->ir
->inputs
, ctx
->ninputs
);
1675 struct ir3_instruction
*addr
= ir3_get_addr0(ctx
, src
[0], 4);
1676 for (int i
= 0; i
< intr
->num_components
; i
++) {
1677 unsigned n
= idx
* 4 + i
+ comp
;
1678 dst
[i
] = create_indirect_load(ctx
, ctx
->ninputs
,
1683 /* All SSBO intrinsics should have been lowered by 'lower_io_offsets'
1684 * pass and replaced by an ir3-specifc version that adds the
1685 * dword-offset in the last source.
1687 case nir_intrinsic_load_ssbo_ir3
:
1688 ctx
->funcs
->emit_intrinsic_load_ssbo(ctx
, intr
, dst
);
1690 case nir_intrinsic_store_ssbo_ir3
:
1691 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1692 !ctx
->s
->info
.fs
.early_fragment_tests
)
1693 ctx
->so
->no_earlyz
= true;
1694 ctx
->funcs
->emit_intrinsic_store_ssbo(ctx
, intr
);
1696 case nir_intrinsic_get_buffer_size
:
1697 emit_intrinsic_ssbo_size(ctx
, intr
, dst
);
1699 case nir_intrinsic_ssbo_atomic_add_ir3
:
1700 case nir_intrinsic_ssbo_atomic_imin_ir3
:
1701 case nir_intrinsic_ssbo_atomic_umin_ir3
:
1702 case nir_intrinsic_ssbo_atomic_imax_ir3
:
1703 case nir_intrinsic_ssbo_atomic_umax_ir3
:
1704 case nir_intrinsic_ssbo_atomic_and_ir3
:
1705 case nir_intrinsic_ssbo_atomic_or_ir3
:
1706 case nir_intrinsic_ssbo_atomic_xor_ir3
:
1707 case nir_intrinsic_ssbo_atomic_exchange_ir3
:
1708 case nir_intrinsic_ssbo_atomic_comp_swap_ir3
:
1709 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1710 !ctx
->s
->info
.fs
.early_fragment_tests
)
1711 ctx
->so
->no_earlyz
= true;
1712 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_ssbo(ctx
, intr
);
1714 case nir_intrinsic_load_shared
:
1715 emit_intrinsic_load_shared(ctx
, intr
, dst
);
1717 case nir_intrinsic_store_shared
:
1718 emit_intrinsic_store_shared(ctx
, intr
);
1720 case nir_intrinsic_shared_atomic_add
:
1721 case nir_intrinsic_shared_atomic_imin
:
1722 case nir_intrinsic_shared_atomic_umin
:
1723 case nir_intrinsic_shared_atomic_imax
:
1724 case nir_intrinsic_shared_atomic_umax
:
1725 case nir_intrinsic_shared_atomic_and
:
1726 case nir_intrinsic_shared_atomic_or
:
1727 case nir_intrinsic_shared_atomic_xor
:
1728 case nir_intrinsic_shared_atomic_exchange
:
1729 case nir_intrinsic_shared_atomic_comp_swap
:
1730 dst
[0] = emit_intrinsic_atomic_shared(ctx
, intr
);
1732 case nir_intrinsic_image_load
:
1733 emit_intrinsic_load_image(ctx
, intr
, dst
);
1735 case nir_intrinsic_bindless_image_load
:
1736 /* Bindless uses the IBO state, which doesn't have swizzle filled out,
1737 * so using isam doesn't work.
1739 * TODO: can we use isam if we fill out more fields?
1741 ctx
->funcs
->emit_intrinsic_load_image(ctx
, intr
, dst
);
1743 case nir_intrinsic_image_store
:
1744 case nir_intrinsic_bindless_image_store
:
1745 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1746 !ctx
->s
->info
.fs
.early_fragment_tests
)
1747 ctx
->so
->no_earlyz
= true;
1748 ctx
->funcs
->emit_intrinsic_store_image(ctx
, intr
);
1750 case nir_intrinsic_image_size
:
1751 case nir_intrinsic_bindless_image_size
:
1752 emit_intrinsic_image_size(ctx
, intr
, dst
);
1754 case nir_intrinsic_image_atomic_add
:
1755 case nir_intrinsic_bindless_image_atomic_add
:
1756 case nir_intrinsic_image_atomic_imin
:
1757 case nir_intrinsic_bindless_image_atomic_imin
:
1758 case nir_intrinsic_image_atomic_umin
:
1759 case nir_intrinsic_bindless_image_atomic_umin
:
1760 case nir_intrinsic_image_atomic_imax
:
1761 case nir_intrinsic_bindless_image_atomic_imax
:
1762 case nir_intrinsic_image_atomic_umax
:
1763 case nir_intrinsic_bindless_image_atomic_umax
:
1764 case nir_intrinsic_image_atomic_and
:
1765 case nir_intrinsic_bindless_image_atomic_and
:
1766 case nir_intrinsic_image_atomic_or
:
1767 case nir_intrinsic_bindless_image_atomic_or
:
1768 case nir_intrinsic_image_atomic_xor
:
1769 case nir_intrinsic_bindless_image_atomic_xor
:
1770 case nir_intrinsic_image_atomic_exchange
:
1771 case nir_intrinsic_bindless_image_atomic_exchange
:
1772 case nir_intrinsic_image_atomic_comp_swap
:
1773 case nir_intrinsic_bindless_image_atomic_comp_swap
:
1774 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1775 !ctx
->s
->info
.fs
.early_fragment_tests
)
1776 ctx
->so
->no_earlyz
= true;
1777 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_image(ctx
, intr
);
1779 case nir_intrinsic_control_barrier
:
1780 case nir_intrinsic_memory_barrier
:
1781 case nir_intrinsic_group_memory_barrier
:
1782 case nir_intrinsic_memory_barrier_buffer
:
1783 case nir_intrinsic_memory_barrier_image
:
1784 case nir_intrinsic_memory_barrier_shared
:
1785 emit_intrinsic_barrier(ctx
, intr
);
1786 /* note that blk ptr no longer valid, make that obvious: */
1789 case nir_intrinsic_store_output
:
1790 idx
= nir_intrinsic_base(intr
);
1791 comp
= nir_intrinsic_component(intr
);
1792 compile_assert(ctx
, nir_src_is_const(intr
->src
[1]));
1793 idx
+= nir_src_as_uint(intr
->src
[1]);
1795 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1796 for (int i
= 0; i
< intr
->num_components
; i
++) {
1797 unsigned n
= idx
* 4 + i
+ comp
;
1798 ctx
->outputs
[n
] = src
[i
];
1801 case nir_intrinsic_load_base_vertex
:
1802 case nir_intrinsic_load_first_vertex
:
1803 if (!ctx
->basevertex
) {
1804 ctx
->basevertex
= create_driver_param(ctx
, IR3_DP_VTXID_BASE
);
1806 dst
[0] = ctx
->basevertex
;
1808 case nir_intrinsic_load_base_instance
:
1809 if (!ctx
->base_instance
) {
1810 ctx
->base_instance
= create_driver_param(ctx
, IR3_DP_INSTID_BASE
);
1812 dst
[0] = ctx
->base_instance
;
1814 case nir_intrinsic_load_vertex_id_zero_base
:
1815 case nir_intrinsic_load_vertex_id
:
1816 if (!ctx
->vertex_id
) {
1817 gl_system_value sv
= (intr
->intrinsic
== nir_intrinsic_load_vertex_id
) ?
1818 SYSTEM_VALUE_VERTEX_ID
: SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1819 ctx
->vertex_id
= create_sysval_input(ctx
, sv
, 0x1);
1821 dst
[0] = ctx
->vertex_id
;
1823 case nir_intrinsic_load_instance_id
:
1824 if (!ctx
->instance_id
) {
1825 ctx
->instance_id
= create_sysval_input(ctx
, SYSTEM_VALUE_INSTANCE_ID
, 0x1);
1827 dst
[0] = ctx
->instance_id
;
1829 case nir_intrinsic_load_sample_id
:
1830 ctx
->so
->per_samp
= true;
1832 case nir_intrinsic_load_sample_id_no_per_sample
:
1833 if (!ctx
->samp_id
) {
1834 ctx
->samp_id
= create_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_ID
, 0x1);
1835 ctx
->samp_id
->regs
[0]->flags
|= IR3_REG_HALF
;
1837 dst
[0] = ir3_COV(b
, ctx
->samp_id
, TYPE_U16
, TYPE_U32
);
1839 case nir_intrinsic_load_sample_mask_in
:
1840 if (!ctx
->samp_mask_in
) {
1841 ctx
->samp_mask_in
= create_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_MASK_IN
, 0x1);
1843 dst
[0] = ctx
->samp_mask_in
;
1845 case nir_intrinsic_load_user_clip_plane
:
1846 idx
= nir_intrinsic_ucp_id(intr
);
1847 for (int i
= 0; i
< intr
->num_components
; i
++) {
1848 unsigned n
= idx
* 4 + i
;
1849 dst
[i
] = create_driver_param(ctx
, IR3_DP_UCP0_X
+ n
);
1852 case nir_intrinsic_load_front_face
:
1853 if (!ctx
->frag_face
) {
1854 ctx
->so
->frag_face
= true;
1855 ctx
->frag_face
= create_sysval_input(ctx
, SYSTEM_VALUE_FRONT_FACE
, 0x1);
1856 ctx
->frag_face
->regs
[0]->flags
|= IR3_REG_HALF
;
1858 /* for fragface, we get -1 for back and 0 for front. However this is
1859 * the inverse of what nir expects (where ~0 is true).
1861 dst
[0] = ir3_CMPS_S(b
,
1863 create_immed_typed(b
, 0, TYPE_U16
), 0);
1864 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
1866 case nir_intrinsic_load_local_invocation_id
:
1867 if (!ctx
->local_invocation_id
) {
1868 ctx
->local_invocation_id
=
1869 create_sysval_input(ctx
, SYSTEM_VALUE_LOCAL_INVOCATION_ID
, 0x7);
1871 ir3_split_dest(b
, dst
, ctx
->local_invocation_id
, 0, 3);
1873 case nir_intrinsic_load_work_group_id
:
1874 if (!ctx
->work_group_id
) {
1875 ctx
->work_group_id
=
1876 create_sysval_input(ctx
, SYSTEM_VALUE_WORK_GROUP_ID
, 0x7);
1877 ctx
->work_group_id
->regs
[0]->flags
|= IR3_REG_HIGH
;
1879 ir3_split_dest(b
, dst
, ctx
->work_group_id
, 0, 3);
1881 case nir_intrinsic_load_num_work_groups
:
1882 for (int i
= 0; i
< intr
->num_components
; i
++) {
1883 dst
[i
] = create_driver_param(ctx
, IR3_DP_NUM_WORK_GROUPS_X
+ i
);
1886 case nir_intrinsic_load_local_group_size
:
1887 for (int i
= 0; i
< intr
->num_components
; i
++) {
1888 dst
[i
] = create_driver_param(ctx
, IR3_DP_LOCAL_GROUP_SIZE_X
+ i
);
1891 case nir_intrinsic_discard_if
:
1892 case nir_intrinsic_discard
: {
1893 struct ir3_instruction
*cond
, *kill
;
1895 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
1896 /* conditional discard: */
1897 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1900 /* unconditional discard: */
1901 cond
= create_immed(b
, 1);
1904 /* NOTE: only cmps.*.* can write p0.x: */
1905 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1906 cond
->cat2
.condition
= IR3_COND_NE
;
1908 /* condition always goes in predicate register: */
1909 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1910 cond
->regs
[0]->flags
&= ~IR3_REG_SSA
;
1912 kill
= ir3_KILL(b
, cond
, 0);
1913 kill
->regs
[1]->num
= regid(REG_P0
, 0);
1914 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1916 array_insert(b
, b
->keeps
, kill
);
1917 ctx
->so
->no_earlyz
= true;
1922 case nir_intrinsic_cond_end_ir3
: {
1923 struct ir3_instruction
*cond
, *kill
;
1925 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1928 /* NOTE: only cmps.*.* can write p0.x: */
1929 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1930 cond
->cat2
.condition
= IR3_COND_NE
;
1932 /* condition always goes in predicate register: */
1933 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1935 kill
= ir3_PREDT(b
, cond
, 0);
1937 kill
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1938 kill
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
1940 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1941 array_insert(b
, b
->keeps
, kill
);
1945 case nir_intrinsic_load_shared_ir3
:
1946 emit_intrinsic_load_shared_ir3(ctx
, intr
, dst
);
1948 case nir_intrinsic_store_shared_ir3
:
1949 emit_intrinsic_store_shared_ir3(ctx
, intr
);
1951 case nir_intrinsic_bindless_resource_ir3
:
1952 dst
[0] = ir3_get_src(ctx
, &intr
->src
[0])[0];
1955 ir3_context_error(ctx
, "Unhandled intrinsic type: %s\n",
1956 nir_intrinsic_infos
[intr
->intrinsic
].name
);
1961 ir3_put_dst(ctx
, &intr
->dest
);
1965 emit_load_const(struct ir3_context
*ctx
, nir_load_const_instr
*instr
)
1967 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &instr
->def
,
1968 instr
->def
.num_components
);
1970 if (instr
->def
.bit_size
== 16) {
1971 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1972 dst
[i
] = create_immed_typed(ctx
->block
,
1973 instr
->value
[i
].u16
,
1976 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1977 dst
[i
] = create_immed_typed(ctx
->block
,
1978 instr
->value
[i
].u32
,
1985 emit_undef(struct ir3_context
*ctx
, nir_ssa_undef_instr
*undef
)
1987 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &undef
->def
,
1988 undef
->def
.num_components
);
1989 type_t type
= (undef
->def
.bit_size
== 16) ? TYPE_U16
: TYPE_U32
;
1991 /* backend doesn't want undefined instructions, so just plug
1994 for (int i
= 0; i
< undef
->def
.num_components
; i
++)
1995 dst
[i
] = create_immed_typed(ctx
->block
, fui(0.0), type
);
1999 * texture fetch/sample instructions:
2003 get_tex_dest_type(nir_tex_instr
*tex
)
2007 switch (nir_alu_type_get_base_type(tex
->dest_type
)) {
2008 case nir_type_invalid
:
2009 case nir_type_float
:
2010 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_F16
: TYPE_F32
;
2013 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_S16
: TYPE_S32
;
2017 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_U16
: TYPE_U32
;
2020 unreachable("bad dest_type");
2027 tex_info(nir_tex_instr
*tex
, unsigned *flagsp
, unsigned *coordsp
)
2029 unsigned coords
= glsl_get_sampler_dim_coordinate_components(tex
->sampler_dim
);
2032 /* note: would use tex->coord_components.. except txs.. also,
2033 * since array index goes after shadow ref, we don't want to
2037 flags
|= IR3_INSTR_3D
;
2039 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
2040 flags
|= IR3_INSTR_S
;
2042 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
)
2043 flags
|= IR3_INSTR_A
;
2049 /* Gets the sampler/texture idx as a hvec2. Which could either be dynamic
2050 * or immediate (in which case it will get lowered later to a non .s2en
2051 * version of the tex instruction which encode tex/samp as immediates:
2053 static struct tex_src_info
2054 get_tex_samp_tex_src(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2056 struct ir3_block
*b
= ctx
->block
;
2057 struct tex_src_info info
= { 0 };
2058 int texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_handle
);
2059 int sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_handle
);
2060 struct ir3_instruction
*texture
, *sampler
;
2062 if (texture_idx
>= 0 || sampler_idx
>= 0) {
2064 info
.flags
|= IR3_INSTR_B
;
2066 /* Gather information required to determine which encoding to
2067 * choose as well as for prefetch.
2069 nir_intrinsic_instr
*bindless_tex
= NULL
;
2071 if (texture_idx
>= 0) {
2072 ctx
->so
->bindless_tex
= true;
2073 bindless_tex
= ir3_bindless_resource(tex
->src
[texture_idx
].src
);
2074 assert(bindless_tex
);
2075 info
.tex_base
= nir_intrinsic_desc_set(bindless_tex
);
2076 tex_const
= nir_src_is_const(bindless_tex
->src
[0]);
2078 info
.tex_idx
= nir_src_as_uint(bindless_tex
->src
[0]);
2080 /* To simplify some of the logic below, assume the index is
2081 * constant 0 when it's not enabled.
2086 nir_intrinsic_instr
*bindless_samp
= NULL
;
2088 if (sampler_idx
>= 0) {
2089 ctx
->so
->bindless_samp
= true;
2090 bindless_samp
= ir3_bindless_resource(tex
->src
[sampler_idx
].src
);
2091 assert(bindless_samp
);
2092 info
.samp_base
= nir_intrinsic_desc_set(bindless_samp
);
2093 samp_const
= nir_src_is_const(bindless_samp
->src
[0]);
2095 info
.samp_idx
= nir_src_as_uint(bindless_samp
->src
[0]);
2101 /* Choose encoding. */
2102 if (tex_const
&& samp_const
&& info
.tex_idx
< 256 && info
.samp_idx
< 256) {
2103 if (info
.tex_idx
< 16 && info
.samp_idx
< 16 &&
2104 (!bindless_tex
|| !bindless_samp
|| info
.tex_base
== info
.samp_base
)) {
2105 /* Everything fits within the instruction */
2106 info
.base
= info
.tex_base
;
2107 info
.combined_idx
= info
.samp_idx
| (info
.tex_idx
<< 4);
2109 info
.base
= info
.tex_base
;
2110 info
.a1_val
= info
.tex_idx
<< 3 | info
.samp_base
;
2111 info
.combined_idx
= info
.samp_idx
;
2112 info
.flags
|= IR3_INSTR_A1EN
;
2114 info
.samp_tex
= NULL
;
2116 info
.flags
|= IR3_INSTR_S2EN
;
2117 /* In the indirect case, we only use a1.x to store the sampler
2118 * base if it differs from the texture base.
2120 if (!bindless_tex
|| !bindless_samp
|| info
.tex_base
== info
.samp_base
) {
2121 info
.base
= info
.tex_base
;
2123 info
.base
= info
.tex_base
;
2124 info
.a1_val
= info
.samp_base
;
2125 info
.flags
|= IR3_INSTR_A1EN
;
2128 /* Note: the indirect source is now a vec2 instead of hvec2, and
2129 * for some reason the texture and sampler are swapped.
2131 struct ir3_instruction
*texture
, *sampler
;
2134 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
2136 texture
= create_immed(b
, 0);
2139 if (bindless_samp
) {
2140 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
2142 sampler
= create_immed(b
, 0);
2144 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2150 info
.flags
|= IR3_INSTR_S2EN
;
2151 texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
);
2152 sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
);
2153 if (texture_idx
>= 0) {
2154 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
2155 texture
= ir3_COV(ctx
->block
, texture
, TYPE_U32
, TYPE_U16
);
2157 /* TODO what to do for dynamic case? I guess we only need the
2158 * max index for astc srgb workaround so maybe not a problem
2159 * to worry about if we don't enable indirect samplers for
2162 ctx
->max_texture_index
= MAX2(ctx
->max_texture_index
, tex
->texture_index
);
2163 texture
= create_immed_typed(ctx
->block
, tex
->texture_index
, TYPE_U16
);
2164 info
.tex_idx
= tex
->texture_index
;
2167 if (sampler_idx
>= 0) {
2168 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
2169 sampler
= ir3_COV(ctx
->block
, sampler
, TYPE_U32
, TYPE_U16
);
2171 sampler
= create_immed_typed(ctx
->block
, tex
->sampler_index
, TYPE_U16
);
2172 info
.samp_idx
= tex
->texture_index
;
2175 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2185 emit_tex(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2187 struct ir3_block
*b
= ctx
->block
;
2188 struct ir3_instruction
**dst
, *sam
, *src0
[12], *src1
[4];
2189 struct ir3_instruction
* const *coord
, * const *off
, * const *ddx
, * const *ddy
;
2190 struct ir3_instruction
*lod
, *compare
, *proj
, *sample_index
;
2191 struct tex_src_info info
= { 0 };
2192 bool has_bias
= false, has_lod
= false, has_proj
= false, has_off
= false;
2193 unsigned i
, coords
, flags
, ncomp
;
2194 unsigned nsrc0
= 0, nsrc1
= 0;
2198 ncomp
= nir_dest_num_components(tex
->dest
);
2200 coord
= off
= ddx
= ddy
= NULL
;
2201 lod
= proj
= compare
= sample_index
= NULL
;
2203 dst
= ir3_get_dst(ctx
, &tex
->dest
, ncomp
);
2205 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
2206 switch (tex
->src
[i
].src_type
) {
2207 case nir_tex_src_coord
:
2208 coord
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2210 case nir_tex_src_bias
:
2211 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2214 case nir_tex_src_lod
:
2215 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2218 case nir_tex_src_comparator
: /* shadow comparator */
2219 compare
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2221 case nir_tex_src_projector
:
2222 proj
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2225 case nir_tex_src_offset
:
2226 off
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2229 case nir_tex_src_ddx
:
2230 ddx
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2232 case nir_tex_src_ddy
:
2233 ddy
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2235 case nir_tex_src_ms_index
:
2236 sample_index
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2238 case nir_tex_src_texture_offset
:
2239 case nir_tex_src_sampler_offset
:
2240 case nir_tex_src_texture_handle
:
2241 case nir_tex_src_sampler_handle
:
2242 /* handled in get_tex_samp_src() */
2245 ir3_context_error(ctx
, "Unhandled NIR tex src type: %d\n",
2246 tex
->src
[i
].src_type
);
2252 case nir_texop_tex_prefetch
:
2253 compile_assert(ctx
, !has_bias
);
2254 compile_assert(ctx
, !has_lod
);
2255 compile_assert(ctx
, !compare
);
2256 compile_assert(ctx
, !has_proj
);
2257 compile_assert(ctx
, !has_off
);
2258 compile_assert(ctx
, !ddx
);
2259 compile_assert(ctx
, !ddy
);
2260 compile_assert(ctx
, !sample_index
);
2261 compile_assert(ctx
, nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
) < 0);
2262 compile_assert(ctx
, nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
) < 0);
2264 if (ctx
->so
->num_sampler_prefetch
< IR3_MAX_SAMPLER_PREFETCH
) {
2265 opc
= OPC_META_TEX_PREFETCH
;
2266 ctx
->so
->num_sampler_prefetch
++;
2270 case nir_texop_tex
: opc
= has_lod
? OPC_SAML
: OPC_SAM
; break;
2271 case nir_texop_txb
: opc
= OPC_SAMB
; break;
2272 case nir_texop_txl
: opc
= OPC_SAML
; break;
2273 case nir_texop_txd
: opc
= OPC_SAMGQ
; break;
2274 case nir_texop_txf
: opc
= OPC_ISAML
; break;
2275 case nir_texop_lod
: opc
= OPC_GETLOD
; break;
2277 /* NOTE: a4xx might need to emulate gather w/ txf (this is
2278 * what blob does, seems gather is broken?), and a3xx did
2279 * not support it (but probably could also emulate).
2281 switch (tex
->component
) {
2282 case 0: opc
= OPC_GATHER4R
; break;
2283 case 1: opc
= OPC_GATHER4G
; break;
2284 case 2: opc
= OPC_GATHER4B
; break;
2285 case 3: opc
= OPC_GATHER4A
; break;
2288 case nir_texop_txf_ms_fb
:
2289 case nir_texop_txf_ms
: opc
= OPC_ISAMM
; break;
2291 ir3_context_error(ctx
, "Unhandled NIR tex type: %d\n", tex
->op
);
2295 tex_info(tex
, &flags
, &coords
);
2298 * lay out the first argument in the proper order:
2299 * - actual coordinates first
2300 * - shadow reference
2303 * - starting at offset 4, dpdx.xy, dpdy.xy
2305 * bias/lod go into the second arg
2308 /* insert tex coords: */
2309 for (i
= 0; i
< coords
; i
++)
2314 /* scale up integer coords for TXF based on the LOD */
2315 if (ctx
->compiler
->unminify_coords
&& (opc
== OPC_ISAML
)) {
2317 for (i
= 0; i
< coords
; i
++)
2318 src0
[i
] = ir3_SHL_B(b
, src0
[i
], 0, lod
, 0);
2322 /* hw doesn't do 1d, so we treat it as 2d with
2323 * height of 1, and patch up the y coord.
2326 src0
[nsrc0
++] = create_immed(b
, 0);
2328 src0
[nsrc0
++] = create_immed(b
, fui(0.5));
2332 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
2333 src0
[nsrc0
++] = compare
;
2335 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
) {
2336 struct ir3_instruction
*idx
= coord
[coords
];
2338 /* the array coord for cube arrays needs 0.5 added to it */
2339 if (ctx
->compiler
->array_index_add_half
&& !is_isam(opc
))
2340 idx
= ir3_ADD_F(b
, idx
, 0, create_immed(b
, fui(0.5)), 0);
2342 src0
[nsrc0
++] = idx
;
2346 src0
[nsrc0
++] = proj
;
2347 flags
|= IR3_INSTR_P
;
2350 /* pad to 4, then ddx/ddy: */
2351 if (tex
->op
== nir_texop_txd
) {
2353 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2354 for (i
= 0; i
< coords
; i
++)
2355 src0
[nsrc0
++] = ddx
[i
];
2357 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2358 for (i
= 0; i
< coords
; i
++)
2359 src0
[nsrc0
++] = ddy
[i
];
2361 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2364 /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
2365 * with scaled x coord according to requested sample:
2367 if (opc
== OPC_ISAMM
) {
2368 if (ctx
->compiler
->txf_ms_with_isaml
) {
2369 /* the samples are laid out in x dimension as
2371 * x_ms = (x << ms) + sample_index;
2373 struct ir3_instruction
*ms
;
2374 ms
= create_immed(b
, (ctx
->samples
>> (2 * tex
->texture_index
)) & 3);
2376 src0
[0] = ir3_SHL_B(b
, src0
[0], 0, ms
, 0);
2377 src0
[0] = ir3_ADD_U(b
, src0
[0], 0, sample_index
, 0);
2381 src0
[nsrc0
++] = sample_index
;
2386 * second argument (if applicable):
2391 if (has_off
| has_lod
| has_bias
) {
2393 unsigned off_coords
= coords
;
2394 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2396 for (i
= 0; i
< off_coords
; i
++)
2397 src1
[nsrc1
++] = off
[i
];
2399 src1
[nsrc1
++] = create_immed(b
, fui(0.0));
2400 flags
|= IR3_INSTR_O
;
2403 if (has_lod
| has_bias
)
2404 src1
[nsrc1
++] = lod
;
2407 type
= get_tex_dest_type(tex
);
2409 if (opc
== OPC_GETLOD
)
2413 if (tex
->op
== nir_texop_txf_ms_fb
) {
2414 /* only expect a single txf_ms_fb per shader: */
2415 compile_assert(ctx
, !ctx
->so
->fb_read
);
2416 compile_assert(ctx
, ctx
->so
->type
== MESA_SHADER_FRAGMENT
);
2418 ctx
->so
->fb_read
= true;
2419 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2420 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2421 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2423 info
.flags
= IR3_INSTR_S2EN
;
2425 ctx
->so
->num_samp
++;
2427 info
= get_tex_samp_tex_src(ctx
, tex
);
2430 struct ir3_instruction
*col0
= ir3_create_collect(ctx
, src0
, nsrc0
);
2431 struct ir3_instruction
*col1
= ir3_create_collect(ctx
, src1
, nsrc1
);
2433 if (opc
== OPC_META_TEX_PREFETCH
) {
2434 int idx
= nir_tex_instr_src_index(tex
, nir_tex_src_coord
);
2436 compile_assert(ctx
, tex
->src
[idx
].src
.is_ssa
);
2438 sam
= ir3_META_TEX_PREFETCH(b
);
2439 __ssa_dst(sam
)->wrmask
= MASK(ncomp
); /* dst */
2440 __ssa_src(sam
, get_barycentric_pixel(ctx
), 0);
2441 sam
->prefetch
.input_offset
=
2442 ir3_nir_coord_offset(tex
->src
[idx
].src
.ssa
);
2443 /* make sure not to add irrelevant flags like S2EN */
2444 sam
->flags
= flags
| (info
.flags
& IR3_INSTR_B
);
2445 sam
->prefetch
.tex
= info
.tex_idx
;
2446 sam
->prefetch
.samp
= info
.samp_idx
;
2447 sam
->prefetch
.tex_base
= info
.tex_base
;
2448 sam
->prefetch
.samp_base
= info
.samp_base
;
2450 info
.flags
|= flags
;
2451 sam
= emit_sam(ctx
, opc
, info
, type
, MASK(ncomp
), col0
, col1
);
2454 if ((ctx
->astc_srgb
& (1 << tex
->texture_index
)) && !nir_tex_instr_is_query(tex
)) {
2455 assert(opc
!= OPC_META_TEX_PREFETCH
);
2457 /* only need first 3 components: */
2458 sam
->regs
[0]->wrmask
= 0x7;
2459 ir3_split_dest(b
, dst
, sam
, 0, 3);
2461 /* we need to sample the alpha separately with a non-ASTC
2464 sam
= ir3_SAM(b
, opc
, type
, 0b1000, flags
| info
.flags
,
2465 info
.samp_tex
, col0
, col1
);
2467 array_insert(ctx
->ir
, ctx
->ir
->astc_srgb
, sam
);
2469 /* fixup .w component: */
2470 ir3_split_dest(b
, &dst
[3], sam
, 3, 1);
2472 /* normal (non-workaround) case: */
2473 ir3_split_dest(b
, dst
, sam
, 0, ncomp
);
2476 /* GETLOD returns results in 4.8 fixed point */
2477 if (opc
== OPC_GETLOD
) {
2478 struct ir3_instruction
*factor
= create_immed(b
, fui(1.0 / 256));
2480 compile_assert(ctx
, tex
->dest_type
== nir_type_float
);
2481 for (i
= 0; i
< 2; i
++) {
2482 dst
[i
] = ir3_MUL_F(b
, ir3_COV(b
, dst
[i
], TYPE_S32
, TYPE_F32
), 0,
2487 ir3_put_dst(ctx
, &tex
->dest
);
2491 emit_tex_info(struct ir3_context
*ctx
, nir_tex_instr
*tex
, unsigned idx
)
2493 struct ir3_block
*b
= ctx
->block
;
2494 struct ir3_instruction
**dst
, *sam
;
2495 type_t dst_type
= get_tex_dest_type(tex
);
2496 struct tex_src_info info
= get_tex_samp_tex_src(ctx
, tex
);
2498 dst
= ir3_get_dst(ctx
, &tex
->dest
, 1);
2500 sam
= emit_sam(ctx
, OPC_GETINFO
, info
, dst_type
, 1 << idx
, NULL
, NULL
);
2502 /* even though there is only one component, since it ends
2503 * up in .y/.z/.w rather than .x, we need a split_dest()
2505 ir3_split_dest(b
, dst
, sam
, idx
, 1);
2507 /* The # of levels comes from getinfo.z. We need to add 1 to it, since
2508 * the value in TEX_CONST_0 is zero-based.
2510 if (ctx
->compiler
->levels_add_one
)
2511 dst
[0] = ir3_ADD_U(b
, dst
[0], 0, create_immed(b
, 1), 0);
2513 ir3_put_dst(ctx
, &tex
->dest
);
2517 emit_tex_txs(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2519 struct ir3_block
*b
= ctx
->block
;
2520 struct ir3_instruction
**dst
, *sam
;
2521 struct ir3_instruction
*lod
;
2522 unsigned flags
, coords
;
2523 type_t dst_type
= get_tex_dest_type(tex
);
2524 struct tex_src_info info
= get_tex_samp_tex_src(ctx
, tex
);
2526 tex_info(tex
, &flags
, &coords
);
2527 info
.flags
|= flags
;
2529 /* Actually we want the number of dimensions, not coordinates. This
2530 * distinction only matters for cubes.
2532 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2535 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
2537 int lod_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_lod
);
2538 compile_assert(ctx
, lod_idx
>= 0);
2540 lod
= ir3_get_src(ctx
, &tex
->src
[lod_idx
].src
)[0];
2542 sam
= emit_sam(ctx
, OPC_GETSIZE
, info
, dst_type
, 0b1111, lod
, NULL
);
2543 ir3_split_dest(b
, dst
, sam
, 0, 4);
2545 /* Array size actually ends up in .w rather than .z. This doesn't
2546 * matter for miplevel 0, but for higher mips the value in z is
2547 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2548 * returned, which means that we have to add 1 to it for arrays.
2550 if (tex
->is_array
) {
2551 if (ctx
->compiler
->levels_add_one
) {
2552 dst
[coords
] = ir3_ADD_U(b
, dst
[3], 0, create_immed(b
, 1), 0);
2554 dst
[coords
] = ir3_MOV(b
, dst
[3], TYPE_U32
);
2558 ir3_put_dst(ctx
, &tex
->dest
);
2562 emit_jump(struct ir3_context
*ctx
, nir_jump_instr
*jump
)
2564 switch (jump
->type
) {
2565 case nir_jump_break
:
2566 case nir_jump_continue
:
2567 case nir_jump_return
:
2568 /* I *think* we can simply just ignore this, and use the
2569 * successor block link to figure out where we need to
2570 * jump to for break/continue
2574 ir3_context_error(ctx
, "Unhandled NIR jump type: %d\n", jump
->type
);
2580 emit_instr(struct ir3_context
*ctx
, nir_instr
*instr
)
2582 switch (instr
->type
) {
2583 case nir_instr_type_alu
:
2584 emit_alu(ctx
, nir_instr_as_alu(instr
));
2586 case nir_instr_type_deref
:
2587 /* ignored, handled as part of the intrinsic they are src to */
2589 case nir_instr_type_intrinsic
:
2590 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2592 case nir_instr_type_load_const
:
2593 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2595 case nir_instr_type_ssa_undef
:
2596 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
2598 case nir_instr_type_tex
: {
2599 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
2600 /* couple tex instructions get special-cased:
2604 emit_tex_txs(ctx
, tex
);
2606 case nir_texop_query_levels
:
2607 emit_tex_info(ctx
, tex
, 2);
2609 case nir_texop_texture_samples
:
2610 emit_tex_info(ctx
, tex
, 3);
2618 case nir_instr_type_jump
:
2619 emit_jump(ctx
, nir_instr_as_jump(instr
));
2621 case nir_instr_type_phi
:
2622 /* we have converted phi webs to regs in NIR by now */
2623 ir3_context_error(ctx
, "Unexpected NIR instruction type: %d\n", instr
->type
);
2625 case nir_instr_type_call
:
2626 case nir_instr_type_parallel_copy
:
2627 ir3_context_error(ctx
, "Unhandled NIR instruction type: %d\n", instr
->type
);
2632 static struct ir3_block
*
2633 get_block(struct ir3_context
*ctx
, const nir_block
*nblock
)
2635 struct ir3_block
*block
;
2636 struct hash_entry
*hentry
;
2638 hentry
= _mesa_hash_table_search(ctx
->block_ht
, nblock
);
2640 return hentry
->data
;
2642 block
= ir3_block_create(ctx
->ir
);
2643 block
->nblock
= nblock
;
2644 _mesa_hash_table_insert(ctx
->block_ht
, nblock
, block
);
2646 set_foreach(nblock
->predecessors
, sentry
) {
2647 _mesa_set_add(block
->predecessors
, get_block(ctx
, sentry
->key
));
2654 emit_block(struct ir3_context
*ctx
, nir_block
*nblock
)
2656 struct ir3_block
*block
= get_block(ctx
, nblock
);
2658 for (int i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
2659 if (nblock
->successors
[i
]) {
2660 block
->successors
[i
] =
2661 get_block(ctx
, nblock
->successors
[i
]);
2666 list_addtail(&block
->node
, &ctx
->ir
->block_list
);
2668 /* re-emit addr register in each block if needed: */
2669 for (int i
= 0; i
< ARRAY_SIZE(ctx
->addr0_ht
); i
++) {
2670 _mesa_hash_table_destroy(ctx
->addr0_ht
[i
], NULL
);
2671 ctx
->addr0_ht
[i
] = NULL
;
2674 _mesa_hash_table_u64_destroy(ctx
->addr1_ht
, NULL
);
2675 ctx
->addr1_ht
= NULL
;
2677 nir_foreach_instr (instr
, nblock
) {
2678 ctx
->cur_instr
= instr
;
2679 emit_instr(ctx
, instr
);
2680 ctx
->cur_instr
= NULL
;
2685 _mesa_hash_table_clear(ctx
->sel_cond_conversions
, NULL
);
2688 static void emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
);
2691 emit_if(struct ir3_context
*ctx
, nir_if
*nif
)
2693 struct ir3_instruction
*condition
= ir3_get_src(ctx
, &nif
->condition
)[0];
2695 ctx
->block
->condition
= ir3_get_predicate(ctx
, condition
);
2697 emit_cf_list(ctx
, &nif
->then_list
);
2698 emit_cf_list(ctx
, &nif
->else_list
);
2702 emit_loop(struct ir3_context
*ctx
, nir_loop
*nloop
)
2704 emit_cf_list(ctx
, &nloop
->body
);
2709 stack_push(struct ir3_context
*ctx
)
2712 ctx
->max_stack
= MAX2(ctx
->max_stack
, ctx
->stack
);
2716 stack_pop(struct ir3_context
*ctx
)
2718 compile_assert(ctx
, ctx
->stack
> 0);
2723 emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
)
2725 foreach_list_typed (nir_cf_node
, node
, node
, list
) {
2726 switch (node
->type
) {
2727 case nir_cf_node_block
:
2728 emit_block(ctx
, nir_cf_node_as_block(node
));
2730 case nir_cf_node_if
:
2732 emit_if(ctx
, nir_cf_node_as_if(node
));
2735 case nir_cf_node_loop
:
2737 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2740 case nir_cf_node_function
:
2741 ir3_context_error(ctx
, "TODO\n");
2747 /* emit stream-out code. At this point, the current block is the original
2748 * (nir) end block, and nir ensures that all flow control paths terminate
2749 * into the end block. We re-purpose the original end block to generate
2750 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2751 * block holding stream-out write instructions, followed by the new end
2755 * p0.x = (vtxcnt < maxvtxcnt)
2756 * // succs: blockStreamOut, blockNewEnd
2759 * // preds: blockOrigEnd
2760 * ... stream-out instructions ...
2761 * // succs: blockNewEnd
2764 * // preds: blockOrigEnd, blockStreamOut
2768 emit_stream_out(struct ir3_context
*ctx
)
2770 struct ir3
*ir
= ctx
->ir
;
2771 struct ir3_stream_output_info
*strmout
=
2772 &ctx
->so
->shader
->stream_output
;
2773 struct ir3_block
*orig_end_block
, *stream_out_block
, *new_end_block
;
2774 struct ir3_instruction
*vtxcnt
, *maxvtxcnt
, *cond
;
2775 struct ir3_instruction
*bases
[IR3_MAX_SO_BUFFERS
];
2777 /* create vtxcnt input in input block at top of shader,
2778 * so that it is seen as live over the entire duration
2781 vtxcnt
= create_sysval_input(ctx
, SYSTEM_VALUE_VERTEX_CNT
, 0x1);
2782 maxvtxcnt
= create_driver_param(ctx
, IR3_DP_VTXCNT_MAX
);
2784 /* at this point, we are at the original 'end' block,
2785 * re-purpose this block to stream-out condition, then
2786 * append stream-out block and new-end block
2788 orig_end_block
= ctx
->block
;
2790 // maybe w/ store_global intrinsic, we could do this
2791 // stuff in nir->nir pass
2793 stream_out_block
= ir3_block_create(ir
);
2794 list_addtail(&stream_out_block
->node
, &ir
->block_list
);
2796 new_end_block
= ir3_block_create(ir
);
2797 list_addtail(&new_end_block
->node
, &ir
->block_list
);
2799 orig_end_block
->successors
[0] = stream_out_block
;
2800 orig_end_block
->successors
[1] = new_end_block
;
2802 stream_out_block
->successors
[0] = new_end_block
;
2803 _mesa_set_add(stream_out_block
->predecessors
, orig_end_block
);
2805 _mesa_set_add(new_end_block
->predecessors
, orig_end_block
);
2806 _mesa_set_add(new_end_block
->predecessors
, stream_out_block
);
2808 /* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2809 cond
= ir3_CMPS_S(ctx
->block
, vtxcnt
, 0, maxvtxcnt
, 0);
2810 cond
->regs
[0]->num
= regid(REG_P0
, 0);
2811 cond
->regs
[0]->flags
&= ~IR3_REG_SSA
;
2812 cond
->cat2
.condition
= IR3_COND_LT
;
2814 /* condition goes on previous block to the conditional,
2815 * since it is used to pick which of the two successor
2818 orig_end_block
->condition
= cond
;
2820 /* switch to stream_out_block to generate the stream-out
2823 ctx
->block
= stream_out_block
;
2825 /* Calculate base addresses based on vtxcnt. Instructions
2826 * generated for bases not used in following loop will be
2827 * stripped out in the backend.
2829 for (unsigned i
= 0; i
< IR3_MAX_SO_BUFFERS
; i
++) {
2830 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
2831 unsigned stride
= strmout
->stride
[i
];
2832 struct ir3_instruction
*base
, *off
;
2834 base
= create_uniform(ctx
->block
, regid(const_state
->offsets
.tfbo
, i
));
2836 /* 24-bit should be enough: */
2837 off
= ir3_MUL_U24(ctx
->block
, vtxcnt
, 0,
2838 create_immed(ctx
->block
, stride
* 4), 0);
2840 bases
[i
] = ir3_ADD_S(ctx
->block
, off
, 0, base
, 0);
2843 /* Generate the per-output store instructions: */
2844 for (unsigned i
= 0; i
< strmout
->num_outputs
; i
++) {
2845 for (unsigned j
= 0; j
< strmout
->output
[i
].num_components
; j
++) {
2846 unsigned c
= j
+ strmout
->output
[i
].start_component
;
2847 struct ir3_instruction
*base
, *out
, *stg
;
2849 base
= bases
[strmout
->output
[i
].output_buffer
];
2850 out
= ctx
->outputs
[regid(strmout
->output
[i
].register_index
, c
)];
2852 stg
= ir3_STG(ctx
->block
, base
, 0, out
, 0,
2853 create_immed(ctx
->block
, 1), 0);
2854 stg
->cat6
.type
= TYPE_U32
;
2855 stg
->cat6
.dst_offset
= (strmout
->output
[i
].dst_offset
+ j
) * 4;
2857 array_insert(ctx
->block
, ctx
->block
->keeps
, stg
);
2861 /* and finally switch to the new_end_block: */
2862 ctx
->block
= new_end_block
;
2866 emit_function(struct ir3_context
*ctx
, nir_function_impl
*impl
)
2868 nir_metadata_require(impl
, nir_metadata_block_index
);
2870 compile_assert(ctx
, ctx
->stack
== 0);
2872 emit_cf_list(ctx
, &impl
->body
);
2873 emit_block(ctx
, impl
->end_block
);
2875 compile_assert(ctx
, ctx
->stack
== 0);
2877 /* at this point, we should have a single empty block,
2878 * into which we emit the 'end' instruction.
2880 compile_assert(ctx
, list_is_empty(&ctx
->block
->instr_list
));
2882 /* If stream-out (aka transform-feedback) enabled, emit the
2883 * stream-out instructions, followed by a new empty block (into
2884 * which the 'end' instruction lands).
2886 * NOTE: it is done in this order, rather than inserting before
2887 * we emit end_block, because NIR guarantees that all blocks
2888 * flow into end_block, and that end_block has no successors.
2889 * So by re-purposing end_block as the first block of stream-
2890 * out, we guarantee that all exit paths flow into the stream-
2893 if ((ctx
->compiler
->gpu_id
< 500) &&
2894 (ctx
->so
->shader
->stream_output
.num_outputs
> 0) &&
2895 !ctx
->so
->binning_pass
) {
2896 debug_assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
2897 emit_stream_out(ctx
);
2900 /* Vertex shaders in a tessellation or geometry pipeline treat END as a
2901 * NOP and has an epilogue that writes the VS outputs to local storage, to
2902 * be read by the HS. Then it resets execution mask (chmask) and chains
2903 * to the next shader (chsh).
2905 if ((ctx
->so
->type
== MESA_SHADER_VERTEX
&&
2906 (ctx
->so
->key
.has_gs
|| ctx
->so
->key
.tessellation
)) ||
2907 (ctx
->so
->type
== MESA_SHADER_TESS_EVAL
&& ctx
->so
->key
.has_gs
)) {
2908 struct ir3_instruction
*chmask
=
2909 ir3_CHMASK(ctx
->block
);
2910 chmask
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2911 chmask
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2913 struct ir3_instruction
*chsh
=
2914 ir3_CHSH(ctx
->block
);
2915 chsh
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2916 chsh
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2918 ir3_END(ctx
->block
);
2923 setup_input(struct ir3_context
*ctx
, nir_variable
*in
)
2925 struct ir3_shader_variant
*so
= ctx
->so
;
2926 unsigned ncomp
= glsl_get_components(in
->type
);
2927 unsigned n
= in
->data
.driver_location
;
2928 unsigned frac
= in
->data
.location_frac
;
2929 unsigned slot
= in
->data
.location
;
2931 /* Inputs are loaded using ldlw or ldg for these stages. */
2932 if (ctx
->so
->type
== MESA_SHADER_TESS_CTRL
||
2933 ctx
->so
->type
== MESA_SHADER_TESS_EVAL
||
2934 ctx
->so
->type
== MESA_SHADER_GEOMETRY
)
2937 /* skip unread inputs, we could end up with (for example), unsplit
2938 * matrix/etc inputs in the case they are not read, so just silently
2944 so
->inputs
[n
].slot
= slot
;
2945 so
->inputs
[n
].compmask
|= (1 << (ncomp
+ frac
)) - 1;
2946 so
->inputs_count
= MAX2(so
->inputs_count
, n
+ 1);
2947 so
->inputs
[n
].interpolate
= in
->data
.interpolation
;
2949 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2951 /* if any varyings have 'sample' qualifer, that triggers us
2952 * to run in per-sample mode:
2954 so
->per_samp
|= in
->data
.sample
;
2956 for (int i
= 0; i
< ncomp
; i
++) {
2957 struct ir3_instruction
*instr
= NULL
;
2958 unsigned idx
= (n
* 4) + i
+ frac
;
2960 if (slot
== VARYING_SLOT_POS
) {
2961 ir3_context_error(ctx
, "fragcoord should be a sysval!\n");
2963 /* detect the special case for front/back colors where
2964 * we need to do flat vs smooth shading depending on
2967 if (in
->data
.interpolation
== INTERP_MODE_NONE
) {
2969 case VARYING_SLOT_COL0
:
2970 case VARYING_SLOT_COL1
:
2971 case VARYING_SLOT_BFC0
:
2972 case VARYING_SLOT_BFC1
:
2973 so
->inputs
[n
].rasterflat
= true;
2980 if (ctx
->compiler
->flat_bypass
) {
2981 if ((so
->inputs
[n
].interpolate
== INTERP_MODE_FLAT
) ||
2982 (so
->inputs
[n
].rasterflat
&& ctx
->so
->key
.rasterflat
))
2983 so
->inputs
[n
].use_ldlv
= true;
2986 so
->inputs
[n
].bary
= true;
2988 instr
= create_frag_input(ctx
, so
->inputs
[n
].use_ldlv
, idx
);
2991 compile_assert(ctx
, idx
< ctx
->ninputs
);
2993 ctx
->inputs
[idx
] = instr
;
2995 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2996 struct ir3_instruction
*input
= NULL
, *in
;
2997 struct ir3_instruction
*components
[4];
2998 unsigned mask
= (1 << (ncomp
+ frac
)) - 1;
3000 foreach_input (in
, ctx
->ir
) {
3001 if (in
->input
.inidx
== n
) {
3008 input
= create_input(ctx
, mask
);
3009 input
->input
.inidx
= n
;
3011 input
->regs
[0]->wrmask
|= mask
;
3014 ir3_split_dest(ctx
->block
, components
, input
, frac
, ncomp
);
3016 for (int i
= 0; i
< ncomp
; i
++) {
3017 unsigned idx
= (n
* 4) + i
+ frac
;
3018 compile_assert(ctx
, idx
< ctx
->ninputs
);
3019 ctx
->inputs
[idx
] = components
[i
];
3022 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
3025 if (so
->inputs
[n
].bary
|| (ctx
->so
->type
== MESA_SHADER_VERTEX
)) {
3026 so
->total_in
+= ncomp
;
3030 /* Initially we assign non-packed inloc's for varyings, as we don't really
3031 * know up-front which components will be unused. After all the compilation
3032 * stages we scan the shader to see which components are actually used, and
3033 * re-pack the inlocs to eliminate unneeded varyings.
3036 pack_inlocs(struct ir3_context
*ctx
)
3038 struct ir3_shader_variant
*so
= ctx
->so
;
3039 uint8_t used_components
[so
->inputs_count
];
3041 memset(used_components
, 0, sizeof(used_components
));
3044 * First Step: scan shader to find which bary.f/ldlv remain:
3047 foreach_block (block
, &ctx
->ir
->block_list
) {
3048 foreach_instr (instr
, &block
->instr_list
) {
3049 if (is_input(instr
)) {
3050 unsigned inloc
= instr
->regs
[1]->iim_val
;
3051 unsigned i
= inloc
/ 4;
3052 unsigned j
= inloc
% 4;
3054 compile_assert(ctx
, instr
->regs
[1]->flags
& IR3_REG_IMMED
);
3055 compile_assert(ctx
, i
< so
->inputs_count
);
3057 used_components
[i
] |= 1 << j
;
3058 } else if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3059 for (int n
= 0; n
< 2; n
++) {
3060 unsigned inloc
= instr
->prefetch
.input_offset
+ n
;
3061 unsigned i
= inloc
/ 4;
3062 unsigned j
= inloc
% 4;
3064 compile_assert(ctx
, i
< so
->inputs_count
);
3066 used_components
[i
] |= 1 << j
;
3073 * Second Step: reassign varying inloc/slots:
3076 unsigned actual_in
= 0;
3079 for (unsigned i
= 0; i
< so
->inputs_count
; i
++) {
3080 unsigned compmask
= 0, maxcomp
= 0;
3082 so
->inputs
[i
].inloc
= inloc
;
3083 so
->inputs
[i
].bary
= false;
3085 for (unsigned j
= 0; j
< 4; j
++) {
3086 if (!(used_components
[i
] & (1 << j
)))
3089 compmask
|= (1 << j
);
3093 /* at this point, since used_components[i] mask is only
3094 * considering varyings (ie. not sysvals) we know this
3097 so
->inputs
[i
].bary
= true;
3100 if (so
->inputs
[i
].bary
) {
3102 so
->inputs
[i
].compmask
= (1 << maxcomp
) - 1;
3108 * Third Step: reassign packed inloc's:
3111 foreach_block (block
, &ctx
->ir
->block_list
) {
3112 foreach_instr (instr
, &block
->instr_list
) {
3113 if (is_input(instr
)) {
3114 unsigned inloc
= instr
->regs
[1]->iim_val
;
3115 unsigned i
= inloc
/ 4;
3116 unsigned j
= inloc
% 4;
3118 instr
->regs
[1]->iim_val
= so
->inputs
[i
].inloc
+ j
;
3119 } else if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3120 unsigned i
= instr
->prefetch
.input_offset
/ 4;
3121 unsigned j
= instr
->prefetch
.input_offset
% 4;
3122 instr
->prefetch
.input_offset
= so
->inputs
[i
].inloc
+ j
;
3129 setup_output(struct ir3_context
*ctx
, nir_variable
*out
)
3131 struct ir3_shader_variant
*so
= ctx
->so
;
3132 unsigned slots
= glsl_count_vec4_slots(out
->type
, false, false);
3133 unsigned ncomp
= glsl_get_components(glsl_without_array(out
->type
));
3134 unsigned n
= out
->data
.driver_location
;
3135 unsigned frac
= out
->data
.location_frac
;
3136 unsigned slot
= out
->data
.location
;
3138 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
3140 case FRAG_RESULT_DEPTH
:
3141 so
->writes_pos
= true;
3143 case FRAG_RESULT_COLOR
:
3146 case FRAG_RESULT_SAMPLE_MASK
:
3147 so
->writes_smask
= true;
3150 if (slot
>= FRAG_RESULT_DATA0
)
3152 ir3_context_error(ctx
, "unknown FS output name: %s\n",
3153 gl_frag_result_name(slot
));
3155 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
3156 ctx
->so
->type
== MESA_SHADER_TESS_EVAL
||
3157 ctx
->so
->type
== MESA_SHADER_GEOMETRY
) {
3159 case VARYING_SLOT_POS
:
3160 so
->writes_pos
= true;
3162 case VARYING_SLOT_PSIZ
:
3163 so
->writes_psize
= true;
3165 case VARYING_SLOT_PRIMITIVE_ID
:
3166 case VARYING_SLOT_LAYER
:
3167 case VARYING_SLOT_GS_VERTEX_FLAGS_IR3
:
3168 debug_assert(ctx
->so
->type
== MESA_SHADER_GEOMETRY
);
3170 case VARYING_SLOT_COL0
:
3171 case VARYING_SLOT_COL1
:
3172 case VARYING_SLOT_BFC0
:
3173 case VARYING_SLOT_BFC1
:
3174 case VARYING_SLOT_FOGC
:
3175 case VARYING_SLOT_CLIP_DIST0
:
3176 case VARYING_SLOT_CLIP_DIST1
:
3177 case VARYING_SLOT_CLIP_VERTEX
:
3180 if (slot
>= VARYING_SLOT_VAR0
)
3182 if ((VARYING_SLOT_TEX0
<= slot
) && (slot
<= VARYING_SLOT_TEX7
))
3184 ir3_context_error(ctx
, "unknown %s shader output name: %s\n",
3185 _mesa_shader_stage_to_string(ctx
->so
->type
),
3186 gl_varying_slot_name(slot
));
3188 } else if (ctx
->so
->type
== MESA_SHADER_TESS_CTRL
) {
3189 /* output lowered to buffer writes. */
3192 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
3196 so
->outputs_count
= out
->data
.driver_location
+ slots
;
3197 compile_assert(ctx
, so
->outputs_count
< ARRAY_SIZE(so
->outputs
));
3199 for (int i
= 0; i
< slots
; i
++) {
3200 int slot_base
= n
+ i
;
3201 so
->outputs
[slot_base
].slot
= slot
+ i
;
3203 for (int i
= 0; i
< ncomp
; i
++) {
3204 unsigned idx
= (slot_base
* 4) + i
+ frac
;
3205 compile_assert(ctx
, idx
< ctx
->noutputs
);
3206 ctx
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
3209 /* if varying packing doesn't happen, we could end up in a situation
3210 * with "holes" in the output, and since the per-generation code that
3211 * sets up varying linkage registers doesn't expect to have more than
3212 * one varying per vec4 slot, pad the holes.
3214 * Note that this should probably generate a performance warning of
3217 for (int i
= 0; i
< frac
; i
++) {
3218 unsigned idx
= (slot_base
* 4) + i
;
3219 if (!ctx
->outputs
[idx
]) {
3220 ctx
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
3227 emit_instructions(struct ir3_context
*ctx
)
3229 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->s
);
3231 ctx
->ninputs
= ctx
->s
->num_inputs
* 4;
3232 ctx
->noutputs
= ctx
->s
->num_outputs
* 4;
3233 ctx
->inputs
= rzalloc_array(ctx
, struct ir3_instruction
*, ctx
->ninputs
);
3234 ctx
->outputs
= rzalloc_array(ctx
, struct ir3_instruction
*, ctx
->noutputs
);
3236 ctx
->ir
= ir3_create(ctx
->compiler
, ctx
->so
->type
);
3238 /* Create inputs in first block: */
3239 ctx
->block
= get_block(ctx
, nir_start_block(fxn
));
3240 ctx
->in_block
= ctx
->block
;
3242 /* for fragment shader, the vcoord input register is used as the
3243 * base for bary.f varying fetch instrs:
3245 * TODO defer creating ctx->ij_pixel and corresponding sysvals
3246 * until emit_intrinsic when we know they are actually needed.
3247 * For now, we defer creating ctx->ij_centroid, etc, since we
3248 * only need ij_pixel for "old style" varying inputs (ie.
3251 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
3252 ctx
->ij_pixel
= create_input(ctx
, 0x3);
3256 nir_foreach_variable (var
, &ctx
->s
->inputs
) {
3257 setup_input(ctx
, var
);
3260 /* Defer add_sysval_input() stuff until after setup_inputs(),
3261 * because sysvals need to be appended after varyings:
3263 if (ctx
->ij_pixel
) {
3264 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
,
3265 0x3, ctx
->ij_pixel
);
3269 /* Tesselation shaders always need primitive ID for indexing the
3270 * BO. Geometry shaders don't always need it but when they do it has be
3271 * delivered and unclobbered in the VS. To make things easy, we always
3272 * make room for it in VS/DS.
3274 bool has_tess
= ctx
->so
->key
.tessellation
!= IR3_TESS_NONE
;
3275 bool has_gs
= ctx
->so
->key
.has_gs
;
3276 switch (ctx
->so
->type
) {
3277 case MESA_SHADER_VERTEX
:
3279 ctx
->tcs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_TCS_HEADER_IR3
, 0x1);
3280 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3281 } else if (has_gs
) {
3282 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3283 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3286 case MESA_SHADER_TESS_CTRL
:
3287 ctx
->tcs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_TCS_HEADER_IR3
, 0x1);
3288 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3290 case MESA_SHADER_TESS_EVAL
:
3292 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3293 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3295 case MESA_SHADER_GEOMETRY
:
3296 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3297 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3303 /* Setup outputs: */
3304 nir_foreach_variable (var
, &ctx
->s
->outputs
) {
3305 setup_output(ctx
, var
);
3308 /* Find # of samplers: */
3309 nir_foreach_variable (var
, &ctx
->s
->uniforms
) {
3310 ctx
->so
->num_samp
+= glsl_type_get_sampler_count(var
->type
);
3311 /* just assume that we'll be reading from images.. if it
3312 * is write-only we don't have to count it, but not sure
3313 * if there is a good way to know?
3315 ctx
->so
->num_samp
+= glsl_type_get_image_count(var
->type
);
3318 /* NOTE: need to do something more clever when we support >1 fxn */
3319 nir_foreach_register (reg
, &fxn
->registers
) {
3320 ir3_declare_array(ctx
, reg
);
3322 /* And emit the body: */
3324 emit_function(ctx
, fxn
);
3327 /* Fixup tex sampler state for astc/srgb workaround instructions. We
3328 * need to assign the tex state indexes for these after we know the
3332 fixup_astc_srgb(struct ir3_context
*ctx
)
3334 struct ir3_shader_variant
*so
= ctx
->so
;
3335 /* indexed by original tex idx, value is newly assigned alpha sampler
3336 * state tex idx. Zero is invalid since there is at least one sampler
3339 unsigned alt_tex_state
[16] = {0};
3340 unsigned tex_idx
= ctx
->max_texture_index
+ 1;
3343 so
->astc_srgb
.base
= tex_idx
;
3345 for (unsigned i
= 0; i
< ctx
->ir
->astc_srgb_count
; i
++) {
3346 struct ir3_instruction
*sam
= ctx
->ir
->astc_srgb
[i
];
3348 compile_assert(ctx
, sam
->cat5
.tex
< ARRAY_SIZE(alt_tex_state
));
3350 if (alt_tex_state
[sam
->cat5
.tex
] == 0) {
3351 /* assign new alternate/alpha tex state slot: */
3352 alt_tex_state
[sam
->cat5
.tex
] = tex_idx
++;
3353 so
->astc_srgb
.orig_idx
[idx
++] = sam
->cat5
.tex
;
3354 so
->astc_srgb
.count
++;
3357 sam
->cat5
.tex
= alt_tex_state
[sam
->cat5
.tex
];
3362 fixup_binning_pass(struct ir3_context
*ctx
)
3364 struct ir3_shader_variant
*so
= ctx
->so
;
3365 struct ir3
*ir
= ctx
->ir
;
3368 /* first pass, remove unused outputs from the IR level outputs: */
3369 for (i
= 0, j
= 0; i
< ir
->outputs_count
; i
++) {
3370 struct ir3_instruction
*out
= ir
->outputs
[i
];
3371 assert(out
->opc
== OPC_META_COLLECT
);
3372 unsigned outidx
= out
->collect
.outidx
;
3373 unsigned slot
= so
->outputs
[outidx
].slot
;
3375 /* throw away everything but first position/psize */
3376 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
3377 ir
->outputs
[j
] = ir
->outputs
[i
];
3381 ir
->outputs_count
= j
;
3383 /* second pass, cleanup the unused slots in ir3_shader_variant::outputs
3386 for (i
= 0, j
= 0; i
< so
->outputs_count
; i
++) {
3387 unsigned slot
= so
->outputs
[i
].slot
;
3389 /* throw away everything but first position/psize */
3390 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
3391 so
->outputs
[j
] = so
->outputs
[i
];
3393 /* fixup outidx to point to new output table entry: */
3394 struct ir3_instruction
*out
;
3395 foreach_output (out
, ir
) {
3396 if (out
->collect
.outidx
== i
) {
3397 out
->collect
.outidx
= j
;
3405 so
->outputs_count
= j
;
3409 collect_tex_prefetches(struct ir3_context
*ctx
, struct ir3
*ir
)
3413 /* Collect sampling instructions eligible for pre-dispatch. */
3414 foreach_block (block
, &ir
->block_list
) {
3415 foreach_instr_safe (instr
, &block
->instr_list
) {
3416 if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3417 assert(idx
< ARRAY_SIZE(ctx
->so
->sampler_prefetch
));
3418 struct ir3_sampler_prefetch
*fetch
=
3419 &ctx
->so
->sampler_prefetch
[idx
];
3422 if (instr
->flags
& IR3_INSTR_B
) {
3423 fetch
->cmd
= IR3_SAMPLER_BINDLESS_PREFETCH_CMD
;
3424 /* In bindless mode, the index is actually the base */
3425 fetch
->tex_id
= instr
->prefetch
.tex_base
;
3426 fetch
->samp_id
= instr
->prefetch
.samp_base
;
3427 fetch
->tex_bindless_id
= instr
->prefetch
.tex
;
3428 fetch
->samp_bindless_id
= instr
->prefetch
.samp
;
3430 fetch
->cmd
= IR3_SAMPLER_PREFETCH_CMD
;
3431 fetch
->tex_id
= instr
->prefetch
.tex
;
3432 fetch
->samp_id
= instr
->prefetch
.samp
;
3434 fetch
->wrmask
= instr
->regs
[0]->wrmask
;
3435 fetch
->dst
= instr
->regs
[0]->num
;
3436 fetch
->src
= instr
->prefetch
.input_offset
;
3439 MAX2(ctx
->so
->total_in
, instr
->prefetch
.input_offset
+ 2);
3441 /* Disable half precision until supported. */
3442 fetch
->half_precision
= !!(instr
->regs
[0]->flags
& IR3_REG_HALF
);
3444 /* Remove the prefetch placeholder instruction: */
3445 list_delinit(&instr
->node
);
3452 ir3_compile_shader_nir(struct ir3_compiler
*compiler
,
3453 struct ir3_shader_variant
*so
)
3455 struct ir3_context
*ctx
;
3457 int ret
= 0, max_bary
;
3461 ctx
= ir3_context_init(compiler
, so
);
3463 DBG("INIT failed!");
3468 emit_instructions(ctx
);
3471 DBG("EMIT failed!");
3476 ir
= so
->ir
= ctx
->ir
;
3478 assert((ctx
->noutputs
% 4) == 0);
3480 /* Setup IR level outputs, which are "collects" that gather
3481 * the scalar components of outputs.
3483 for (unsigned i
= 0; i
< ctx
->noutputs
; i
+= 4) {
3485 /* figure out the # of components written:
3487 * TODO do we need to handle holes, ie. if .x and .z
3488 * components written, but .y component not written?
3490 for (unsigned j
= 0; j
< 4; j
++) {
3491 if (!ctx
->outputs
[i
+ j
])
3496 /* Note that in some stages, like TCS, store_output is
3497 * lowered to memory writes, so no components of the
3498 * are "written" from the PoV of traditional store-
3499 * output instructions:
3504 struct ir3_instruction
*out
=
3505 ir3_create_collect(ctx
, &ctx
->outputs
[i
], ncomp
);
3508 assert(outidx
< so
->outputs_count
);
3510 /* stash index into so->outputs[] so we can map the
3511 * output back to slot/etc later:
3513 out
->collect
.outidx
= outidx
;
3515 array_insert(ir
, ir
->outputs
, out
);
3518 /* Set up the gs header as an output for the vertex shader so it won't
3519 * clobber it for the tess ctrl shader.
3521 * TODO this could probably be done more cleanly in a nir pass.
3523 if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
3524 (ctx
->so
->key
.has_gs
&& ctx
->so
->type
== MESA_SHADER_TESS_EVAL
)) {
3525 if (ctx
->primitive_id
) {
3526 unsigned n
= so
->outputs_count
++;
3527 so
->outputs
[n
].slot
= VARYING_SLOT_PRIMITIVE_ID
;
3529 struct ir3_instruction
*out
=
3530 ir3_create_collect(ctx
, &ctx
->primitive_id
, 1);
3531 out
->collect
.outidx
= n
;
3532 array_insert(ir
, ir
->outputs
, out
);
3535 if (ctx
->gs_header
) {
3536 unsigned n
= so
->outputs_count
++;
3537 so
->outputs
[n
].slot
= VARYING_SLOT_GS_HEADER_IR3
;
3538 struct ir3_instruction
*out
=
3539 ir3_create_collect(ctx
, &ctx
->gs_header
, 1);
3540 out
->collect
.outidx
= n
;
3541 array_insert(ir
, ir
->outputs
, out
);
3544 if (ctx
->tcs_header
) {
3545 unsigned n
= so
->outputs_count
++;
3546 so
->outputs
[n
].slot
= VARYING_SLOT_TCS_HEADER_IR3
;
3547 struct ir3_instruction
*out
=
3548 ir3_create_collect(ctx
, &ctx
->tcs_header
, 1);
3549 out
->collect
.outidx
= n
;
3550 array_insert(ir
, ir
->outputs
, out
);
3554 /* at this point, for binning pass, throw away unneeded outputs: */
3555 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
< 600))
3556 fixup_binning_pass(ctx
);
3558 ir3_debug_print(ir
, "BEFORE CF");
3562 ir3_debug_print(ir
, "BEFORE CP");
3566 /* at this point, for binning pass, throw away unneeded outputs:
3567 * Note that for a6xx and later, we do this after ir3_cp to ensure
3568 * that the uniform/constant layout for BS and VS matches, so that
3569 * we can re-use same VS_CONST state group.
3571 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
>= 600))
3572 fixup_binning_pass(ctx
);
3574 /* for a6xx+, binning and draw pass VS use same VBO state, so we
3575 * need to make sure not to remove any inputs that are used by
3576 * the nonbinning VS.
3578 if (ctx
->compiler
->gpu_id
>= 600 && so
->binning_pass
&&
3579 so
->type
== MESA_SHADER_VERTEX
) {
3580 for (int i
= 0; i
< ctx
->ninputs
; i
++) {
3581 struct ir3_instruction
*in
= ctx
->inputs
[i
];
3589 debug_assert(n
< so
->nonbinning
->inputs_count
);
3591 if (so
->nonbinning
->inputs
[n
].sysval
)
3594 /* be sure to keep inputs, even if only used in VS */
3595 if (so
->nonbinning
->inputs
[n
].compmask
& (1 << c
))
3596 array_insert(in
->block
, in
->block
->keeps
, in
);
3600 ir3_debug_print(ir
, "BEFORE GROUPING");
3602 ir3_sched_add_deps(ir
);
3604 /* Group left/right neighbors, inserting mov's where needed to
3609 ir3_debug_print(ir
, "AFTER GROUPING");
3613 ir3_debug_print(ir
, "AFTER DCE");
3615 /* do Sethi–Ullman numbering before scheduling: */
3618 ret
= ir3_sched(ir
);
3620 DBG("SCHED failed!");
3624 ir3_debug_print(ir
, "AFTER SCHED");
3626 /* Pre-assign VS inputs on a6xx+ binning pass shader, to align
3627 * with draw pass VS, so binning and draw pass can both use the
3630 * Note that VS inputs are expected to be full precision.
3632 bool pre_assign_inputs
= (ir
->compiler
->gpu_id
>= 600) &&
3633 (ir
->type
== MESA_SHADER_VERTEX
) &&
3636 if (pre_assign_inputs
) {
3637 for (unsigned i
= 0; i
< ctx
->ninputs
; i
++) {
3638 struct ir3_instruction
*instr
= ctx
->inputs
[i
];
3645 unsigned regid
= so
->nonbinning
->inputs
[n
].regid
+ c
;
3647 instr
->regs
[0]->num
= regid
;
3650 ret
= ir3_ra(so
, ctx
->inputs
, ctx
->ninputs
);
3651 } else if (ctx
->tcs_header
) {
3652 /* We need to have these values in the same registers between VS and TCS
3653 * since the VS chains to TCS and doesn't get the sysvals redelivered.
3656 ctx
->tcs_header
->regs
[0]->num
= regid(0, 0);
3657 ctx
->primitive_id
->regs
[0]->num
= regid(0, 1);
3658 struct ir3_instruction
*precolor
[] = { ctx
->tcs_header
, ctx
->primitive_id
};
3659 ret
= ir3_ra(so
, precolor
, ARRAY_SIZE(precolor
));
3660 } else if (ctx
->gs_header
) {
3661 /* We need to have these values in the same registers between producer
3662 * (VS or DS) and GS since the producer chains to GS and doesn't get
3663 * the sysvals redelivered.
3666 ctx
->gs_header
->regs
[0]->num
= regid(0, 0);
3667 ctx
->primitive_id
->regs
[0]->num
= regid(0, 1);
3668 struct ir3_instruction
*precolor
[] = { ctx
->gs_header
, ctx
->primitive_id
};
3669 ret
= ir3_ra(so
, precolor
, ARRAY_SIZE(precolor
));
3670 } else if (so
->num_sampler_prefetch
) {
3671 assert(so
->type
== MESA_SHADER_FRAGMENT
);
3672 struct ir3_instruction
*instr
, *precolor
[2];
3675 foreach_input (instr
, ir
) {
3676 if (instr
->input
.sysval
!= SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
)
3679 assert(idx
< ARRAY_SIZE(precolor
));
3681 precolor
[idx
] = instr
;
3682 instr
->regs
[0]->num
= idx
;
3686 ret
= ir3_ra(so
, precolor
, idx
);
3688 ret
= ir3_ra(so
, NULL
, 0);
3697 ir3_debug_print(ir
, "AFTER POSTSCHED");
3699 if (compiler
->gpu_id
>= 600) {
3700 if (ir3_a6xx_fixup_atomic_dests(ir
, so
)) {
3701 ir3_debug_print(ir
, "AFTER ATOMIC FIXUP");
3705 if (so
->type
== MESA_SHADER_FRAGMENT
)
3709 * Fixup inputs/outputs to point to the actual registers assigned:
3711 * 1) initialize to r63.x (invalid/unused)
3712 * 2) iterate IR level inputs/outputs and update the variants
3713 * inputs/outputs table based on the assigned registers for
3714 * the remaining inputs/outputs.
3717 for (unsigned i
= 0; i
< so
->inputs_count
; i
++)
3718 so
->inputs
[i
].regid
= INVALID_REG
;
3719 for (unsigned i
= 0; i
< so
->outputs_count
; i
++)
3720 so
->outputs
[i
].regid
= INVALID_REG
;
3722 struct ir3_instruction
*out
;
3723 foreach_output (out
, ir
) {
3724 assert(out
->opc
== OPC_META_COLLECT
);
3725 unsigned outidx
= out
->collect
.outidx
;
3727 so
->outputs
[outidx
].regid
= out
->regs
[0]->num
;
3728 so
->outputs
[outidx
].half
= !!(out
->regs
[0]->flags
& IR3_REG_HALF
);
3731 struct ir3_instruction
*in
;
3732 foreach_input (in
, ir
) {
3733 assert(in
->opc
== OPC_META_INPUT
);
3734 unsigned inidx
= in
->input
.inidx
;
3736 if (pre_assign_inputs
&& !so
->inputs
[inidx
].sysval
) {
3737 if (VALIDREG(so
->nonbinning
->inputs
[inidx
].regid
)) {
3738 compile_assert(ctx
, in
->regs
[0]->num
==
3739 so
->nonbinning
->inputs
[inidx
].regid
);
3740 compile_assert(ctx
, !!(in
->regs
[0]->flags
& IR3_REG_HALF
) ==
3741 so
->nonbinning
->inputs
[inidx
].half
);
3743 so
->inputs
[inidx
].regid
= so
->nonbinning
->inputs
[inidx
].regid
;
3744 so
->inputs
[inidx
].half
= so
->nonbinning
->inputs
[inidx
].half
;
3746 so
->inputs
[inidx
].regid
= in
->regs
[0]->num
;
3747 so
->inputs
[inidx
].half
= !!(in
->regs
[0]->flags
& IR3_REG_HALF
);
3752 fixup_astc_srgb(ctx
);
3754 /* We need to do legalize after (for frag shader's) the "bary.f"
3755 * offsets (inloc) have been assigned.
3757 ir3_legalize(ir
, so
, &max_bary
);
3759 ir3_debug_print(ir
, "AFTER LEGALIZE");
3761 /* Set (ss)(sy) on first TCS and GEOMETRY instructions, since we don't
3762 * know what we might have to wait on when coming in from VS chsh.
3764 if (so
->type
== MESA_SHADER_TESS_CTRL
||
3765 so
->type
== MESA_SHADER_GEOMETRY
) {
3766 foreach_block (block
, &ir
->block_list
) {
3767 foreach_instr (instr
, &block
->instr_list
) {
3768 instr
->flags
|= IR3_INSTR_SS
| IR3_INSTR_SY
;
3774 so
->branchstack
= ctx
->max_stack
;
3776 /* Note that actual_in counts inputs that are not bary.f'd for FS: */
3777 if (so
->type
== MESA_SHADER_FRAGMENT
)
3778 so
->total_in
= max_bary
+ 1;
3780 so
->max_sun
= ir
->max_sun
;
3782 /* Collect sampling instructions eligible for pre-dispatch. */
3783 collect_tex_prefetches(ctx
, ir
);
3785 if (so
->type
== MESA_SHADER_FRAGMENT
&&
3786 ctx
->s
->info
.fs
.needs_helper_invocations
)
3787 so
->need_pixlod
= true;
3792 ir3_destroy(so
->ir
);
3795 ir3_context_free(ctx
);