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 /* UBO addresses are the first driver params, but subtract 2 here to
764 * account for nir_lower_uniforms_to_ubo rebasing the UBOs such that UBO 0
765 * is the uniforms: */
766 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
767 unsigned ubo
= regid(const_state
->offsets
.ubo
, 0) - 2;
768 const unsigned ptrsz
= ir3_pointer_size(ctx
->compiler
);
772 /* First src is ubo index, which could either be an immed or not: */
773 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0];
774 if (is_same_type_mov(src0
) &&
775 (src0
->regs
[1]->flags
& IR3_REG_IMMED
)) {
776 base_lo
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
));
777 base_hi
= create_uniform(b
, ubo
+ (src0
->regs
[1]->iim_val
* ptrsz
) + 1);
779 base_lo
= create_uniform_indirect(b
, ubo
, ir3_get_addr0(ctx
, src0
, ptrsz
));
780 base_hi
= create_uniform_indirect(b
, ubo
+ 1, ir3_get_addr0(ctx
, src0
, ptrsz
));
782 /* NOTE: since relative addressing is used, make sure constlen is
783 * at least big enough to cover all the UBO addresses, since the
784 * assembler won't know what the max address reg is.
786 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
787 const_state
->offsets
.ubo
+ (ctx
->s
->info
.num_ubos
* ptrsz
));
790 /* note: on 32bit gpu's base_hi is ignored and DCE'd */
793 if (nir_src_is_const(intr
->src
[1])) {
794 off
+= nir_src_as_uint(intr
->src
[1]);
796 /* For load_ubo_indirect, second src is indirect offset: */
797 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0];
799 /* and add offset to addr: */
800 addr
= ir3_ADD_S(b
, addr
, 0, src1
, 0);
803 /* if offset is to large to encode in the ldg, split it out: */
804 if ((off
+ (intr
->num_components
* 4)) > 1024) {
805 /* split out the minimal amount to improve the odds that
806 * cp can fit the immediate in the add.s instruction:
808 unsigned off2
= off
+ (intr
->num_components
* 4) - 1024;
809 addr
= ir3_ADD_S(b
, addr
, 0, create_immed(b
, off2
), 0);
814 struct ir3_instruction
*carry
;
816 /* handle 32b rollover, ie:
817 * if (addr < base_lo)
820 carry
= ir3_CMPS_U(b
, addr
, 0, base_lo
, 0);
821 carry
->cat2
.condition
= IR3_COND_LT
;
822 base_hi
= ir3_ADD_S(b
, base_hi
, 0, carry
, 0);
824 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){ addr
, base_hi
}, 2);
827 for (int i
= 0; i
< intr
->num_components
; i
++) {
828 struct ir3_instruction
*load
=
829 ir3_LDG(b
, addr
, 0, create_immed(b
, 1), 0, /* num components */
830 create_immed(b
, off
+ i
* 4), 0);
831 load
->cat6
.type
= TYPE_U32
;
836 /* src[] = { block_index } */
838 emit_intrinsic_ssbo_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
839 struct ir3_instruction
**dst
)
841 /* SSBO size stored as a const starting at ssbo_sizes: */
842 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
843 unsigned blk_idx
= nir_src_as_uint(intr
->src
[0]);
844 unsigned idx
= regid(const_state
->offsets
.ssbo_sizes
, 0) +
845 const_state
->ssbo_size
.off
[blk_idx
];
847 debug_assert(const_state
->ssbo_size
.mask
& (1 << blk_idx
));
849 dst
[0] = create_uniform(ctx
->block
, idx
);
852 /* src[] = { offset }. const_index[] = { base } */
854 emit_intrinsic_load_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
855 struct ir3_instruction
**dst
)
857 struct ir3_block
*b
= ctx
->block
;
858 struct ir3_instruction
*ldl
, *offset
;
861 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
862 base
= nir_intrinsic_base(intr
);
864 ldl
= ir3_LDL(b
, offset
, 0,
865 create_immed(b
, intr
->num_components
), 0,
866 create_immed(b
, base
), 0);
868 ldl
->cat6
.type
= utype_dst(intr
->dest
);
869 ldl
->regs
[0]->wrmask
= MASK(intr
->num_components
);
871 ldl
->barrier_class
= IR3_BARRIER_SHARED_R
;
872 ldl
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
874 ir3_split_dest(b
, dst
, ldl
, 0, intr
->num_components
);
877 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
879 emit_intrinsic_store_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
881 struct ir3_block
*b
= ctx
->block
;
882 struct ir3_instruction
*stl
, *offset
;
883 struct ir3_instruction
* const *value
;
884 unsigned base
, wrmask
;
886 value
= ir3_get_src(ctx
, &intr
->src
[0]);
887 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
889 base
= nir_intrinsic_base(intr
);
890 wrmask
= nir_intrinsic_write_mask(intr
);
892 /* Combine groups of consecutive enabled channels in one write
893 * message. We use ffs to find the first enabled channel and then ffs on
894 * the bit-inverse, down-shifted writemask to determine the length of
895 * the block of enabled bits.
897 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
900 unsigned first_component
= ffs(wrmask
) - 1;
901 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
903 stl
= ir3_STL(b
, offset
, 0,
904 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
905 create_immed(b
, length
), 0);
906 stl
->cat6
.dst_offset
= first_component
+ base
;
907 stl
->cat6
.type
= utype_src(intr
->src
[0]);
908 stl
->barrier_class
= IR3_BARRIER_SHARED_W
;
909 stl
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
911 array_insert(b
, b
->keeps
, stl
);
913 /* Clear the bits in the writemask that we just wrote, then try
914 * again to see if more channels are left.
916 wrmask
&= (15 << (first_component
+ length
));
920 /* src[] = { offset }. const_index[] = { base } */
922 emit_intrinsic_load_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
923 struct ir3_instruction
**dst
)
925 struct ir3_block
*b
= ctx
->block
;
926 struct ir3_instruction
*load
, *offset
;
929 offset
= ir3_get_src(ctx
, &intr
->src
[0])[0];
930 base
= nir_intrinsic_base(intr
);
932 load
= ir3_LDLW(b
, offset
, 0,
933 create_immed(b
, intr
->num_components
), 0,
934 create_immed(b
, base
), 0);
936 load
->cat6
.type
= utype_dst(intr
->dest
);
937 load
->regs
[0]->wrmask
= MASK(intr
->num_components
);
939 load
->barrier_class
= IR3_BARRIER_SHARED_R
;
940 load
->barrier_conflict
= IR3_BARRIER_SHARED_W
;
942 ir3_split_dest(b
, dst
, load
, 0, intr
->num_components
);
945 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
947 emit_intrinsic_store_shared_ir3(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
949 struct ir3_block
*b
= ctx
->block
;
950 struct ir3_instruction
*store
, *offset
;
951 struct ir3_instruction
* const *value
;
952 unsigned base
, wrmask
;
954 value
= ir3_get_src(ctx
, &intr
->src
[0]);
955 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
957 base
= nir_intrinsic_base(intr
);
958 wrmask
= nir_intrinsic_write_mask(intr
);
960 /* Combine groups of consecutive enabled channels in one write
961 * message. We use ffs to find the first enabled channel and then ffs on
962 * the bit-inverse, down-shifted writemask to determine the length of
963 * the block of enabled bits.
965 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
968 unsigned first_component
= ffs(wrmask
) - 1;
969 unsigned length
= ffs(~(wrmask
>> first_component
)) - 1;
971 store
= ir3_STLW(b
, offset
, 0,
972 ir3_create_collect(ctx
, &value
[first_component
], length
), 0,
973 create_immed(b
, length
), 0);
975 store
->cat6
.dst_offset
= first_component
+ base
;
976 store
->cat6
.type
= utype_src(intr
->src
[0]);
977 store
->barrier_class
= IR3_BARRIER_SHARED_W
;
978 store
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
980 array_insert(b
, b
->keeps
, store
);
982 /* Clear the bits in the writemask that we just wrote, then try
983 * again to see if more channels are left.
985 wrmask
&= (15 << (first_component
+ length
));
990 * CS shared variable atomic intrinsics
992 * All of the shared variable atomic memory operations read a value from
993 * memory, compute a new value using one of the operations below, write the
994 * new value to memory, and return the original value read.
996 * All operations take 2 sources except CompSwap that takes 3. These
999 * 0: The offset into the shared variable storage region that the atomic
1000 * operation will operate on.
1001 * 1: The data parameter to the atomic function (i.e. the value to add
1002 * in shared_atomic_add, etc).
1003 * 2: For CompSwap only: the second data parameter.
1005 static struct ir3_instruction
*
1006 emit_intrinsic_atomic_shared(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1008 struct ir3_block
*b
= ctx
->block
;
1009 struct ir3_instruction
*atomic
, *src0
, *src1
;
1010 type_t type
= TYPE_U32
;
1012 src0
= ir3_get_src(ctx
, &intr
->src
[0])[0]; /* offset */
1013 src1
= ir3_get_src(ctx
, &intr
->src
[1])[0]; /* value */
1015 switch (intr
->intrinsic
) {
1016 case nir_intrinsic_shared_atomic_add
:
1017 atomic
= ir3_ATOMIC_ADD(b
, src0
, 0, src1
, 0);
1019 case nir_intrinsic_shared_atomic_imin
:
1020 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
1023 case nir_intrinsic_shared_atomic_umin
:
1024 atomic
= ir3_ATOMIC_MIN(b
, src0
, 0, src1
, 0);
1026 case nir_intrinsic_shared_atomic_imax
:
1027 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
1030 case nir_intrinsic_shared_atomic_umax
:
1031 atomic
= ir3_ATOMIC_MAX(b
, src0
, 0, src1
, 0);
1033 case nir_intrinsic_shared_atomic_and
:
1034 atomic
= ir3_ATOMIC_AND(b
, src0
, 0, src1
, 0);
1036 case nir_intrinsic_shared_atomic_or
:
1037 atomic
= ir3_ATOMIC_OR(b
, src0
, 0, src1
, 0);
1039 case nir_intrinsic_shared_atomic_xor
:
1040 atomic
= ir3_ATOMIC_XOR(b
, src0
, 0, src1
, 0);
1042 case nir_intrinsic_shared_atomic_exchange
:
1043 atomic
= ir3_ATOMIC_XCHG(b
, src0
, 0, src1
, 0);
1045 case nir_intrinsic_shared_atomic_comp_swap
:
1046 /* for cmpxchg, src1 is [ui]vec2(data, compare): */
1047 src1
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1048 ir3_get_src(ctx
, &intr
->src
[2])[0],
1051 atomic
= ir3_ATOMIC_CMPXCHG(b
, src0
, 0, src1
, 0);
1057 atomic
->cat6
.iim_val
= 1;
1059 atomic
->cat6
.type
= type
;
1060 atomic
->barrier_class
= IR3_BARRIER_SHARED_W
;
1061 atomic
->barrier_conflict
= IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
;
1063 /* even if nothing consume the result, we can't DCE the instruction: */
1064 array_insert(b
, b
->keeps
, atomic
);
1069 struct tex_src_info
{
1071 unsigned tex_base
, samp_base
, tex_idx
, samp_idx
;
1072 /* For normal tex instructions */
1073 unsigned base
, combined_idx
, a1_val
, flags
;
1074 struct ir3_instruction
*samp_tex
;
1077 /* TODO handle actual indirect/dynamic case.. which is going to be weird
1078 * to handle with the image_mapping table..
1080 static struct tex_src_info
1081 get_image_samp_tex_src(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1083 struct ir3_block
*b
= ctx
->block
;
1084 struct tex_src_info info
= { 0 };
1085 nir_intrinsic_instr
*bindless_tex
= ir3_bindless_resource(intr
->src
[0]);
1086 ctx
->so
->bindless_tex
= true;
1090 info
.flags
|= IR3_INSTR_B
;
1092 /* Gather information required to determine which encoding to
1093 * choose as well as for prefetch.
1095 info
.tex_base
= nir_intrinsic_desc_set(bindless_tex
);
1096 bool tex_const
= nir_src_is_const(bindless_tex
->src
[0]);
1098 info
.tex_idx
= nir_src_as_uint(bindless_tex
->src
[0]);
1101 /* Choose encoding. */
1102 if (tex_const
&& info
.tex_idx
< 256) {
1103 if (info
.tex_idx
< 16) {
1104 /* Everything fits within the instruction */
1105 info
.base
= info
.tex_base
;
1106 info
.combined_idx
= info
.samp_idx
| (info
.tex_idx
<< 4);
1108 info
.base
= info
.tex_base
;
1109 info
.a1_val
= info
.tex_idx
<< 3;
1110 info
.combined_idx
= 0;
1111 info
.flags
|= IR3_INSTR_A1EN
;
1113 info
.samp_tex
= NULL
;
1115 info
.flags
|= IR3_INSTR_S2EN
;
1116 info
.base
= info
.tex_base
;
1118 /* Note: the indirect source is now a vec2 instead of hvec2 */
1119 struct ir3_instruction
*texture
, *sampler
;
1121 texture
= ir3_get_src(ctx
, &intr
->src
[0])[0];
1122 sampler
= create_immed(b
, 0);
1123 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1129 info
.flags
|= IR3_INSTR_S2EN
;
1130 unsigned slot
= nir_src_as_uint(intr
->src
[0]);
1131 unsigned tex_idx
= ir3_image_to_tex(&ctx
->so
->image_mapping
, slot
);
1132 struct ir3_instruction
*texture
, *sampler
;
1134 texture
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1135 sampler
= create_immed_typed(ctx
->block
, tex_idx
, TYPE_U16
);
1137 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1146 static struct ir3_instruction
*
1147 emit_sam(struct ir3_context
*ctx
, opc_t opc
, struct tex_src_info info
,
1148 type_t type
, unsigned wrmask
, struct ir3_instruction
*src0
,
1149 struct ir3_instruction
*src1
)
1151 struct ir3_instruction
*sam
, *addr
;
1152 if (info
.flags
& IR3_INSTR_A1EN
) {
1153 addr
= ir3_get_addr1(ctx
, info
.a1_val
);
1155 sam
= ir3_SAM(ctx
->block
, opc
, type
, 0b1111, info
.flags
,
1156 info
.samp_tex
, src0
, src1
);
1157 if (info
.flags
& IR3_INSTR_A1EN
) {
1158 ir3_instr_set_address(sam
, addr
);
1160 if (info
.flags
& IR3_INSTR_B
) {
1161 sam
->cat5
.tex_base
= info
.base
;
1162 sam
->cat5
.samp
= info
.combined_idx
;
1167 /* src[] = { deref, coord, sample_index }. const_index[] = {} */
1169 emit_intrinsic_load_image(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1170 struct ir3_instruction
**dst
)
1172 struct ir3_block
*b
= ctx
->block
;
1173 struct tex_src_info info
= get_image_samp_tex_src(ctx
, intr
);
1174 struct ir3_instruction
*sam
;
1175 struct ir3_instruction
* const *src0
= ir3_get_src(ctx
, &intr
->src
[1]);
1176 struct ir3_instruction
*coords
[4];
1177 unsigned flags
, ncoords
= ir3_get_image_coords(intr
, &flags
);
1178 type_t type
= ir3_get_type_for_image_intrinsic(intr
);
1180 /* hmm, this seems a bit odd, but it is what blob does and (at least
1181 * a5xx) just faults on bogus addresses otherwise:
1183 if (flags
& IR3_INSTR_3D
) {
1184 flags
&= ~IR3_INSTR_3D
;
1185 flags
|= IR3_INSTR_A
;
1187 info
.flags
|= flags
;
1189 for (unsigned i
= 0; i
< ncoords
; i
++)
1190 coords
[i
] = src0
[i
];
1193 coords
[ncoords
++] = create_immed(b
, 0);
1195 sam
= emit_sam(ctx
, OPC_ISAM
, info
, type
, 0b1111,
1196 ir3_create_collect(ctx
, coords
, ncoords
), NULL
);
1198 sam
->barrier_class
= IR3_BARRIER_IMAGE_R
;
1199 sam
->barrier_conflict
= IR3_BARRIER_IMAGE_W
;
1201 ir3_split_dest(b
, dst
, sam
, 0, 4);
1205 emit_intrinsic_image_size(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
,
1206 struct ir3_instruction
**dst
)
1208 struct ir3_block
*b
= ctx
->block
;
1209 struct tex_src_info info
= get_image_samp_tex_src(ctx
, intr
);
1210 struct ir3_instruction
*sam
, *lod
;
1211 unsigned flags
, ncoords
= ir3_get_image_coords(intr
, &flags
);
1212 type_t dst_type
= nir_dest_bit_size(intr
->dest
) == 16 ?
1213 TYPE_U16
: TYPE_U32
;
1215 info
.flags
|= flags
;
1216 lod
= create_immed(b
, 0);
1217 sam
= emit_sam(ctx
, OPC_GETSIZE
, info
, dst_type
, 0b1111, lod
, NULL
);
1219 /* Array size actually ends up in .w rather than .z. This doesn't
1220 * matter for miplevel 0, but for higher mips the value in z is
1221 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
1222 * returned, which means that we have to add 1 to it for arrays for
1225 * Note use a temporary dst and then copy, since the size of the dst
1226 * array that is passed in is based on nir's understanding of the
1227 * result size, not the hardware's
1229 struct ir3_instruction
*tmp
[4];
1231 ir3_split_dest(b
, tmp
, sam
, 0, 4);
1233 /* get_size instruction returns size in bytes instead of texels
1234 * for imageBuffer, so we need to divide it by the pixel size
1235 * of the image format.
1237 * TODO: This is at least true on a5xx. Check other gens.
1239 if (nir_intrinsic_image_dim(intr
) == GLSL_SAMPLER_DIM_BUF
) {
1240 /* Since all the possible values the divisor can take are
1241 * power-of-two (4, 8, or 16), the division is implemented
1243 * During shader setup, the log2 of the image format's
1244 * bytes-per-pixel should have been emitted in 2nd slot of
1245 * image_dims. See ir3_shader::emit_image_dims().
1247 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
1248 unsigned cb
= regid(const_state
->offsets
.image_dims
, 0) +
1249 const_state
->image_dims
.off
[nir_src_as_uint(intr
->src
[0])];
1250 struct ir3_instruction
*aux
= create_uniform(b
, cb
+ 1);
1252 tmp
[0] = ir3_SHR_B(b
, tmp
[0], 0, aux
, 0);
1255 for (unsigned i
= 0; i
< ncoords
; i
++)
1258 if (flags
& IR3_INSTR_A
) {
1259 if (ctx
->compiler
->levels_add_one
) {
1260 dst
[ncoords
-1] = ir3_ADD_U(b
, tmp
[3], 0, create_immed(b
, 1), 0);
1262 dst
[ncoords
-1] = ir3_MOV(b
, tmp
[3], TYPE_U32
);
1268 emit_intrinsic_barrier(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1270 struct ir3_block
*b
= ctx
->block
;
1271 struct ir3_instruction
*barrier
;
1273 switch (intr
->intrinsic
) {
1274 case nir_intrinsic_control_barrier
:
1275 barrier
= ir3_BAR(b
);
1276 barrier
->cat7
.g
= true;
1277 barrier
->cat7
.l
= true;
1278 barrier
->flags
= IR3_INSTR_SS
| IR3_INSTR_SY
;
1279 barrier
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1281 case nir_intrinsic_memory_barrier
:
1282 barrier
= ir3_FENCE(b
);
1283 barrier
->cat7
.g
= true;
1284 barrier
->cat7
.r
= true;
1285 barrier
->cat7
.w
= true;
1286 barrier
->cat7
.l
= true;
1287 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
|
1288 IR3_BARRIER_BUFFER_W
;
1289 barrier
->barrier_conflict
=
1290 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1291 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1293 case nir_intrinsic_memory_barrier_buffer
:
1294 barrier
= ir3_FENCE(b
);
1295 barrier
->cat7
.g
= true;
1296 barrier
->cat7
.r
= true;
1297 barrier
->cat7
.w
= true;
1298 barrier
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1299 barrier
->barrier_conflict
= IR3_BARRIER_BUFFER_R
|
1300 IR3_BARRIER_BUFFER_W
;
1302 case nir_intrinsic_memory_barrier_image
:
1303 // TODO double check if this should have .g set
1304 barrier
= ir3_FENCE(b
);
1305 barrier
->cat7
.g
= true;
1306 barrier
->cat7
.r
= true;
1307 barrier
->cat7
.w
= true;
1308 barrier
->barrier_class
= IR3_BARRIER_IMAGE_W
;
1309 barrier
->barrier_conflict
= IR3_BARRIER_IMAGE_R
|
1310 IR3_BARRIER_IMAGE_W
;
1312 case nir_intrinsic_memory_barrier_shared
:
1313 barrier
= ir3_FENCE(b
);
1314 barrier
->cat7
.g
= true;
1315 barrier
->cat7
.l
= true;
1316 barrier
->cat7
.r
= true;
1317 barrier
->cat7
.w
= true;
1318 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
;
1319 barrier
->barrier_conflict
= IR3_BARRIER_SHARED_R
|
1320 IR3_BARRIER_SHARED_W
;
1322 case nir_intrinsic_group_memory_barrier
:
1323 barrier
= ir3_FENCE(b
);
1324 barrier
->cat7
.g
= true;
1325 barrier
->cat7
.l
= true;
1326 barrier
->cat7
.r
= true;
1327 barrier
->cat7
.w
= true;
1328 barrier
->barrier_class
= IR3_BARRIER_SHARED_W
|
1329 IR3_BARRIER_IMAGE_W
|
1330 IR3_BARRIER_BUFFER_W
;
1331 barrier
->barrier_conflict
=
1332 IR3_BARRIER_SHARED_R
| IR3_BARRIER_SHARED_W
|
1333 IR3_BARRIER_IMAGE_R
| IR3_BARRIER_IMAGE_W
|
1334 IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1340 /* make sure barrier doesn't get DCE'd */
1341 array_insert(b
, b
->keeps
, barrier
);
1344 static void add_sysval_input_compmask(struct ir3_context
*ctx
,
1345 gl_system_value slot
, unsigned compmask
,
1346 struct ir3_instruction
*instr
)
1348 struct ir3_shader_variant
*so
= ctx
->so
;
1349 unsigned n
= so
->inputs_count
++;
1351 assert(instr
->opc
== OPC_META_INPUT
);
1352 instr
->input
.inidx
= n
;
1353 instr
->input
.sysval
= slot
;
1355 so
->inputs
[n
].sysval
= true;
1356 so
->inputs
[n
].slot
= slot
;
1357 so
->inputs
[n
].compmask
= compmask
;
1358 so
->inputs
[n
].interpolate
= INTERP_MODE_FLAT
;
1362 static struct ir3_instruction
*
1363 create_sysval_input(struct ir3_context
*ctx
, gl_system_value slot
,
1367 struct ir3_instruction
*sysval
= create_input(ctx
, compmask
);
1368 add_sysval_input_compmask(ctx
, slot
, compmask
, sysval
);
1372 static struct ir3_instruction
*
1373 get_barycentric_centroid(struct ir3_context
*ctx
)
1375 if (!ctx
->ij_centroid
) {
1376 struct ir3_instruction
*xy
[2];
1377 struct ir3_instruction
*ij
;
1379 ij
= create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID
, 0x3);
1380 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1382 ctx
->ij_centroid
= ir3_create_collect(ctx
, xy
, 2);
1385 return ctx
->ij_centroid
;
1388 static struct ir3_instruction
*
1389 get_barycentric_sample(struct ir3_context
*ctx
)
1391 if (!ctx
->ij_sample
) {
1392 struct ir3_instruction
*xy
[2];
1393 struct ir3_instruction
*ij
;
1395 ij
= create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE
, 0x3);
1396 ir3_split_dest(ctx
->block
, xy
, ij
, 0, 2);
1398 ctx
->ij_sample
= ir3_create_collect(ctx
, xy
, 2);
1401 return ctx
->ij_sample
;
1404 static struct ir3_instruction
*
1405 get_barycentric_pixel(struct ir3_context
*ctx
)
1407 /* TODO when tgsi_to_nir supports "new-style" FS inputs switch
1408 * this to create ij_pixel only on demand:
1410 return ctx
->ij_pixel
;
1413 static struct ir3_instruction
*
1414 get_frag_coord(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1416 if (!ctx
->frag_coord
) {
1417 struct ir3_block
*b
= ctx
->in_block
;
1418 struct ir3_instruction
*xyzw
[4];
1419 struct ir3_instruction
*hw_frag_coord
;
1421 hw_frag_coord
= create_sysval_input(ctx
, SYSTEM_VALUE_FRAG_COORD
, 0xf);
1422 ir3_split_dest(b
, xyzw
, hw_frag_coord
, 0, 4);
1424 /* for frag_coord.xy, we get unsigned values.. we need
1425 * to subtract (integer) 8 and divide by 16 (right-
1426 * shift by 4) then convert to float:
1430 * mov.u32f32 dst, tmp
1433 for (int i
= 0; i
< 2; i
++) {
1434 xyzw
[i
] = ir3_COV(b
, xyzw
[i
], TYPE_U32
, TYPE_F32
);
1435 xyzw
[i
] = ir3_MUL_F(b
, xyzw
[i
], 0, create_immed(b
, fui(1.0 / 16.0)), 0);
1438 ctx
->frag_coord
= ir3_create_collect(ctx
, xyzw
, 4);
1441 ctx
->so
->fragcoord_compmask
|=
1442 nir_ssa_def_components_read(&intr
->dest
.ssa
);
1444 return ctx
->frag_coord
;
1448 emit_intrinsic(struct ir3_context
*ctx
, nir_intrinsic_instr
*intr
)
1450 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1451 struct ir3_instruction
**dst
;
1452 struct ir3_instruction
* const *src
;
1453 struct ir3_block
*b
= ctx
->block
;
1456 if (info
->has_dest
) {
1457 unsigned n
= nir_intrinsic_dest_components(intr
);
1458 dst
= ir3_get_dst(ctx
, &intr
->dest
, n
);
1463 const unsigned primitive_param
= ctx
->so
->shader
->const_state
.offsets
.primitive_param
* 4;
1464 const unsigned primitive_map
= ctx
->so
->shader
->const_state
.offsets
.primitive_map
* 4;
1466 switch (intr
->intrinsic
) {
1467 case nir_intrinsic_load_uniform
:
1468 idx
= nir_intrinsic_base(intr
);
1469 if (nir_src_is_const(intr
->src
[0])) {
1470 idx
+= nir_src_as_uint(intr
->src
[0]);
1471 for (int i
= 0; i
< intr
->num_components
; i
++) {
1472 dst
[i
] = create_uniform_typed(b
, idx
+ i
,
1473 nir_dest_bit_size(intr
->dest
) == 16 ? TYPE_F16
: TYPE_F32
);
1476 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1477 for (int i
= 0; i
< intr
->num_components
; i
++) {
1478 dst
[i
] = create_uniform_indirect(b
, idx
+ i
,
1479 ir3_get_addr0(ctx
, src
[0], 1));
1481 /* NOTE: if relative addressing is used, we set
1482 * constlen in the compiler (to worst-case value)
1483 * since we don't know in the assembler what the max
1484 * addr reg value can be:
1486 ctx
->so
->constlen
= MAX2(ctx
->so
->constlen
,
1487 ctx
->so
->shader
->ubo_state
.size
/ 16);
1491 case nir_intrinsic_load_vs_primitive_stride_ir3
:
1492 dst
[0] = create_uniform(b
, primitive_param
+ 0);
1494 case nir_intrinsic_load_vs_vertex_stride_ir3
:
1495 dst
[0] = create_uniform(b
, primitive_param
+ 1);
1497 case nir_intrinsic_load_hs_patch_stride_ir3
:
1498 dst
[0] = create_uniform(b
, primitive_param
+ 2);
1500 case nir_intrinsic_load_patch_vertices_in
:
1501 dst
[0] = create_uniform(b
, primitive_param
+ 3);
1503 case nir_intrinsic_load_tess_param_base_ir3
:
1504 dst
[0] = create_uniform(b
, primitive_param
+ 4);
1505 dst
[1] = create_uniform(b
, primitive_param
+ 5);
1507 case nir_intrinsic_load_tess_factor_base_ir3
:
1508 dst
[0] = create_uniform(b
, primitive_param
+ 6);
1509 dst
[1] = create_uniform(b
, primitive_param
+ 7);
1512 case nir_intrinsic_load_primitive_location_ir3
:
1513 idx
= nir_intrinsic_driver_location(intr
);
1514 dst
[0] = create_uniform(b
, primitive_map
+ idx
);
1517 case nir_intrinsic_load_gs_header_ir3
:
1518 dst
[0] = ctx
->gs_header
;
1520 case nir_intrinsic_load_tcs_header_ir3
:
1521 dst
[0] = ctx
->tcs_header
;
1524 case nir_intrinsic_load_primitive_id
:
1525 dst
[0] = ctx
->primitive_id
;
1528 case nir_intrinsic_load_tess_coord
:
1529 if (!ctx
->tess_coord
) {
1531 create_sysval_input(ctx
, SYSTEM_VALUE_TESS_COORD
, 0x3);
1533 ir3_split_dest(b
, dst
, ctx
->tess_coord
, 0, 2);
1535 /* Unused, but ir3_put_dst() below wants to free something */
1536 dst
[2] = create_immed(b
, 0);
1539 case nir_intrinsic_end_patch_ir3
:
1540 assert(ctx
->so
->type
== MESA_SHADER_TESS_CTRL
);
1541 struct ir3_instruction
*end
= ir3_PREDE(b
);
1542 array_insert(b
, b
->keeps
, end
);
1544 end
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1545 end
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
1548 case nir_intrinsic_store_global_ir3
: {
1549 struct ir3_instruction
*value
, *addr
, *offset
;
1551 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1552 ir3_get_src(ctx
, &intr
->src
[1])[0],
1553 ir3_get_src(ctx
, &intr
->src
[1])[1]
1556 offset
= ir3_get_src(ctx
, &intr
->src
[2])[0];
1558 value
= ir3_create_collect(ctx
, ir3_get_src(ctx
, &intr
->src
[0]),
1559 intr
->num_components
);
1561 struct ir3_instruction
*stg
=
1562 ir3_STG_G(ctx
->block
, addr
, 0, value
, 0,
1563 create_immed(ctx
->block
, intr
->num_components
), 0, offset
, 0);
1564 stg
->cat6
.type
= TYPE_U32
;
1565 stg
->cat6
.iim_val
= 1;
1567 array_insert(b
, b
->keeps
, stg
);
1569 stg
->barrier_class
= IR3_BARRIER_BUFFER_W
;
1570 stg
->barrier_conflict
= IR3_BARRIER_BUFFER_R
| IR3_BARRIER_BUFFER_W
;
1574 case nir_intrinsic_load_global_ir3
: {
1575 struct ir3_instruction
*addr
, *offset
;
1577 addr
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
1578 ir3_get_src(ctx
, &intr
->src
[0])[0],
1579 ir3_get_src(ctx
, &intr
->src
[0])[1]
1582 offset
= ir3_get_src(ctx
, &intr
->src
[1])[0];
1584 struct ir3_instruction
*load
=
1585 ir3_LDG(b
, addr
, 0, create_immed(ctx
->block
, intr
->num_components
),
1587 load
->cat6
.type
= TYPE_U32
;
1588 load
->regs
[0]->wrmask
= MASK(intr
->num_components
);
1590 load
->barrier_class
= IR3_BARRIER_BUFFER_R
;
1591 load
->barrier_conflict
= IR3_BARRIER_BUFFER_W
;
1593 ir3_split_dest(b
, dst
, load
, 0, intr
->num_components
);
1597 case nir_intrinsic_load_ubo
:
1598 emit_intrinsic_load_ubo(ctx
, intr
, dst
);
1600 case nir_intrinsic_load_ubo_ir3
:
1601 emit_intrinsic_load_ubo_ldc(ctx
, intr
, dst
);
1603 case nir_intrinsic_load_frag_coord
:
1604 ir3_split_dest(b
, dst
, get_frag_coord(ctx
, intr
), 0, 4);
1606 case nir_intrinsic_load_sample_pos_from_id
: {
1607 /* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
1608 * but that doesn't seem necessary.
1610 struct ir3_instruction
*offset
=
1611 ir3_RGETPOS(b
, ir3_get_src(ctx
, &intr
->src
[0])[0], 0);
1612 offset
->regs
[0]->wrmask
= 0x3;
1613 offset
->cat5
.type
= TYPE_F32
;
1615 ir3_split_dest(b
, dst
, offset
, 0, 2);
1619 case nir_intrinsic_load_size_ir3
:
1620 if (!ctx
->ij_size
) {
1622 create_sysval_input(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE
, 0x1);
1624 dst
[0] = ctx
->ij_size
;
1626 case nir_intrinsic_load_barycentric_centroid
:
1627 ir3_split_dest(b
, dst
, get_barycentric_centroid(ctx
), 0, 2);
1629 case nir_intrinsic_load_barycentric_sample
:
1630 if (ctx
->so
->key
.msaa
) {
1631 ir3_split_dest(b
, dst
, get_barycentric_sample(ctx
), 0, 2);
1633 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1636 case nir_intrinsic_load_barycentric_pixel
:
1637 ir3_split_dest(b
, dst
, get_barycentric_pixel(ctx
), 0, 2);
1639 case nir_intrinsic_load_interpolated_input
:
1640 idx
= nir_intrinsic_base(intr
);
1641 comp
= nir_intrinsic_component(intr
);
1642 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1643 if (nir_src_is_const(intr
->src
[1])) {
1644 struct ir3_instruction
*coord
= ir3_create_collect(ctx
, src
, 2);
1645 idx
+= nir_src_as_uint(intr
->src
[1]);
1646 for (int i
= 0; i
< intr
->num_components
; i
++) {
1647 unsigned inloc
= idx
* 4 + i
+ comp
;
1648 if (ctx
->so
->inputs
[idx
].bary
&&
1649 !ctx
->so
->inputs
[idx
].use_ldlv
) {
1650 dst
[i
] = ir3_BARY_F(b
, create_immed(b
, inloc
), 0, coord
, 0);
1652 /* for non-varyings use the pre-setup input, since
1653 * that is easier than mapping things back to a
1654 * nir_variable to figure out what it is.
1656 dst
[i
] = ctx
->inputs
[inloc
];
1657 compile_assert(ctx
, dst
[i
]);
1661 ir3_context_error(ctx
, "unhandled");
1664 case nir_intrinsic_load_input
:
1665 idx
= nir_intrinsic_base(intr
);
1666 comp
= nir_intrinsic_component(intr
);
1667 if (nir_src_is_const(intr
->src
[0])) {
1668 idx
+= nir_src_as_uint(intr
->src
[0]);
1669 for (int i
= 0; i
< intr
->num_components
; i
++) {
1670 unsigned n
= idx
* 4 + i
+ comp
;
1671 dst
[i
] = ctx
->inputs
[n
];
1672 compile_assert(ctx
, ctx
->inputs
[n
]);
1675 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1676 struct ir3_instruction
*collect
=
1677 ir3_create_collect(ctx
, ctx
->ir
->inputs
, ctx
->ninputs
);
1678 struct ir3_instruction
*addr
= ir3_get_addr0(ctx
, src
[0], 4);
1679 for (int i
= 0; i
< intr
->num_components
; i
++) {
1680 unsigned n
= idx
* 4 + i
+ comp
;
1681 dst
[i
] = create_indirect_load(ctx
, ctx
->ninputs
,
1686 /* All SSBO intrinsics should have been lowered by 'lower_io_offsets'
1687 * pass and replaced by an ir3-specifc version that adds the
1688 * dword-offset in the last source.
1690 case nir_intrinsic_load_ssbo_ir3
:
1691 ctx
->funcs
->emit_intrinsic_load_ssbo(ctx
, intr
, dst
);
1693 case nir_intrinsic_store_ssbo_ir3
:
1694 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1695 !ctx
->s
->info
.fs
.early_fragment_tests
)
1696 ctx
->so
->no_earlyz
= true;
1697 ctx
->funcs
->emit_intrinsic_store_ssbo(ctx
, intr
);
1699 case nir_intrinsic_get_buffer_size
:
1700 emit_intrinsic_ssbo_size(ctx
, intr
, dst
);
1702 case nir_intrinsic_ssbo_atomic_add_ir3
:
1703 case nir_intrinsic_ssbo_atomic_imin_ir3
:
1704 case nir_intrinsic_ssbo_atomic_umin_ir3
:
1705 case nir_intrinsic_ssbo_atomic_imax_ir3
:
1706 case nir_intrinsic_ssbo_atomic_umax_ir3
:
1707 case nir_intrinsic_ssbo_atomic_and_ir3
:
1708 case nir_intrinsic_ssbo_atomic_or_ir3
:
1709 case nir_intrinsic_ssbo_atomic_xor_ir3
:
1710 case nir_intrinsic_ssbo_atomic_exchange_ir3
:
1711 case nir_intrinsic_ssbo_atomic_comp_swap_ir3
:
1712 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1713 !ctx
->s
->info
.fs
.early_fragment_tests
)
1714 ctx
->so
->no_earlyz
= true;
1715 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_ssbo(ctx
, intr
);
1717 case nir_intrinsic_load_shared
:
1718 emit_intrinsic_load_shared(ctx
, intr
, dst
);
1720 case nir_intrinsic_store_shared
:
1721 emit_intrinsic_store_shared(ctx
, intr
);
1723 case nir_intrinsic_shared_atomic_add
:
1724 case nir_intrinsic_shared_atomic_imin
:
1725 case nir_intrinsic_shared_atomic_umin
:
1726 case nir_intrinsic_shared_atomic_imax
:
1727 case nir_intrinsic_shared_atomic_umax
:
1728 case nir_intrinsic_shared_atomic_and
:
1729 case nir_intrinsic_shared_atomic_or
:
1730 case nir_intrinsic_shared_atomic_xor
:
1731 case nir_intrinsic_shared_atomic_exchange
:
1732 case nir_intrinsic_shared_atomic_comp_swap
:
1733 dst
[0] = emit_intrinsic_atomic_shared(ctx
, intr
);
1735 case nir_intrinsic_image_load
:
1736 emit_intrinsic_load_image(ctx
, intr
, dst
);
1738 case nir_intrinsic_bindless_image_load
:
1739 /* Bindless uses the IBO state, which doesn't have swizzle filled out,
1740 * so using isam doesn't work.
1742 * TODO: can we use isam if we fill out more fields?
1744 ctx
->funcs
->emit_intrinsic_load_image(ctx
, intr
, dst
);
1746 case nir_intrinsic_image_store
:
1747 case nir_intrinsic_bindless_image_store
:
1748 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1749 !ctx
->s
->info
.fs
.early_fragment_tests
)
1750 ctx
->so
->no_earlyz
= true;
1751 ctx
->funcs
->emit_intrinsic_store_image(ctx
, intr
);
1753 case nir_intrinsic_image_size
:
1754 case nir_intrinsic_bindless_image_size
:
1755 emit_intrinsic_image_size(ctx
, intr
, dst
);
1757 case nir_intrinsic_image_atomic_add
:
1758 case nir_intrinsic_bindless_image_atomic_add
:
1759 case nir_intrinsic_image_atomic_imin
:
1760 case nir_intrinsic_bindless_image_atomic_imin
:
1761 case nir_intrinsic_image_atomic_umin
:
1762 case nir_intrinsic_bindless_image_atomic_umin
:
1763 case nir_intrinsic_image_atomic_imax
:
1764 case nir_intrinsic_bindless_image_atomic_imax
:
1765 case nir_intrinsic_image_atomic_umax
:
1766 case nir_intrinsic_bindless_image_atomic_umax
:
1767 case nir_intrinsic_image_atomic_and
:
1768 case nir_intrinsic_bindless_image_atomic_and
:
1769 case nir_intrinsic_image_atomic_or
:
1770 case nir_intrinsic_bindless_image_atomic_or
:
1771 case nir_intrinsic_image_atomic_xor
:
1772 case nir_intrinsic_bindless_image_atomic_xor
:
1773 case nir_intrinsic_image_atomic_exchange
:
1774 case nir_intrinsic_bindless_image_atomic_exchange
:
1775 case nir_intrinsic_image_atomic_comp_swap
:
1776 case nir_intrinsic_bindless_image_atomic_comp_swap
:
1777 if ((ctx
->so
->type
== MESA_SHADER_FRAGMENT
) &&
1778 !ctx
->s
->info
.fs
.early_fragment_tests
)
1779 ctx
->so
->no_earlyz
= true;
1780 dst
[0] = ctx
->funcs
->emit_intrinsic_atomic_image(ctx
, intr
);
1782 case nir_intrinsic_control_barrier
:
1783 case nir_intrinsic_memory_barrier
:
1784 case nir_intrinsic_group_memory_barrier
:
1785 case nir_intrinsic_memory_barrier_buffer
:
1786 case nir_intrinsic_memory_barrier_image
:
1787 case nir_intrinsic_memory_barrier_shared
:
1788 emit_intrinsic_barrier(ctx
, intr
);
1789 /* note that blk ptr no longer valid, make that obvious: */
1792 case nir_intrinsic_store_output
:
1793 idx
= nir_intrinsic_base(intr
);
1794 comp
= nir_intrinsic_component(intr
);
1795 compile_assert(ctx
, nir_src_is_const(intr
->src
[1]));
1796 idx
+= nir_src_as_uint(intr
->src
[1]);
1798 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1799 for (int i
= 0; i
< intr
->num_components
; i
++) {
1800 unsigned n
= idx
* 4 + i
+ comp
;
1801 ctx
->outputs
[n
] = src
[i
];
1804 case nir_intrinsic_load_base_vertex
:
1805 case nir_intrinsic_load_first_vertex
:
1806 if (!ctx
->basevertex
) {
1807 ctx
->basevertex
= create_driver_param(ctx
, IR3_DP_VTXID_BASE
);
1809 dst
[0] = ctx
->basevertex
;
1811 case nir_intrinsic_load_base_instance
:
1812 if (!ctx
->base_instance
) {
1813 ctx
->base_instance
= create_driver_param(ctx
, IR3_DP_INSTID_BASE
);
1815 dst
[0] = ctx
->base_instance
;
1817 case nir_intrinsic_load_vertex_id_zero_base
:
1818 case nir_intrinsic_load_vertex_id
:
1819 if (!ctx
->vertex_id
) {
1820 gl_system_value sv
= (intr
->intrinsic
== nir_intrinsic_load_vertex_id
) ?
1821 SYSTEM_VALUE_VERTEX_ID
: SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1822 ctx
->vertex_id
= create_sysval_input(ctx
, sv
, 0x1);
1824 dst
[0] = ctx
->vertex_id
;
1826 case nir_intrinsic_load_instance_id
:
1827 if (!ctx
->instance_id
) {
1828 ctx
->instance_id
= create_sysval_input(ctx
, SYSTEM_VALUE_INSTANCE_ID
, 0x1);
1830 dst
[0] = ctx
->instance_id
;
1832 case nir_intrinsic_load_sample_id
:
1833 ctx
->so
->per_samp
= true;
1835 case nir_intrinsic_load_sample_id_no_per_sample
:
1836 if (!ctx
->samp_id
) {
1837 ctx
->samp_id
= create_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_ID
, 0x1);
1838 ctx
->samp_id
->regs
[0]->flags
|= IR3_REG_HALF
;
1840 dst
[0] = ir3_COV(b
, ctx
->samp_id
, TYPE_U16
, TYPE_U32
);
1842 case nir_intrinsic_load_sample_mask_in
:
1843 if (!ctx
->samp_mask_in
) {
1844 ctx
->samp_mask_in
= create_sysval_input(ctx
, SYSTEM_VALUE_SAMPLE_MASK_IN
, 0x1);
1846 dst
[0] = ctx
->samp_mask_in
;
1848 case nir_intrinsic_load_user_clip_plane
:
1849 idx
= nir_intrinsic_ucp_id(intr
);
1850 for (int i
= 0; i
< intr
->num_components
; i
++) {
1851 unsigned n
= idx
* 4 + i
;
1852 dst
[i
] = create_driver_param(ctx
, IR3_DP_UCP0_X
+ n
);
1855 case nir_intrinsic_load_front_face
:
1856 if (!ctx
->frag_face
) {
1857 ctx
->so
->frag_face
= true;
1858 ctx
->frag_face
= create_sysval_input(ctx
, SYSTEM_VALUE_FRONT_FACE
, 0x1);
1859 ctx
->frag_face
->regs
[0]->flags
|= IR3_REG_HALF
;
1861 /* for fragface, we get -1 for back and 0 for front. However this is
1862 * the inverse of what nir expects (where ~0 is true).
1864 dst
[0] = ir3_CMPS_S(b
,
1866 create_immed_typed(b
, 0, TYPE_U16
), 0);
1867 dst
[0]->cat2
.condition
= IR3_COND_EQ
;
1869 case nir_intrinsic_load_local_invocation_id
:
1870 if (!ctx
->local_invocation_id
) {
1871 ctx
->local_invocation_id
=
1872 create_sysval_input(ctx
, SYSTEM_VALUE_LOCAL_INVOCATION_ID
, 0x7);
1874 ir3_split_dest(b
, dst
, ctx
->local_invocation_id
, 0, 3);
1876 case nir_intrinsic_load_work_group_id
:
1877 if (!ctx
->work_group_id
) {
1878 ctx
->work_group_id
=
1879 create_sysval_input(ctx
, SYSTEM_VALUE_WORK_GROUP_ID
, 0x7);
1880 ctx
->work_group_id
->regs
[0]->flags
|= IR3_REG_HIGH
;
1882 ir3_split_dest(b
, dst
, ctx
->work_group_id
, 0, 3);
1884 case nir_intrinsic_load_num_work_groups
:
1885 for (int i
= 0; i
< intr
->num_components
; i
++) {
1886 dst
[i
] = create_driver_param(ctx
, IR3_DP_NUM_WORK_GROUPS_X
+ i
);
1889 case nir_intrinsic_load_local_group_size
:
1890 for (int i
= 0; i
< intr
->num_components
; i
++) {
1891 dst
[i
] = create_driver_param(ctx
, IR3_DP_LOCAL_GROUP_SIZE_X
+ i
);
1894 case nir_intrinsic_discard_if
:
1895 case nir_intrinsic_discard
: {
1896 struct ir3_instruction
*cond
, *kill
;
1898 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
1899 /* conditional discard: */
1900 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1903 /* unconditional discard: */
1904 cond
= create_immed(b
, 1);
1907 /* NOTE: only cmps.*.* can write p0.x: */
1908 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1909 cond
->cat2
.condition
= IR3_COND_NE
;
1911 /* condition always goes in predicate register: */
1912 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1913 cond
->regs
[0]->flags
&= ~IR3_REG_SSA
;
1915 kill
= ir3_KILL(b
, cond
, 0);
1916 kill
->regs
[1]->num
= regid(REG_P0
, 0);
1917 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1919 array_insert(b
, b
->keeps
, kill
);
1920 ctx
->so
->no_earlyz
= true;
1925 case nir_intrinsic_cond_end_ir3
: {
1926 struct ir3_instruction
*cond
, *kill
;
1928 src
= ir3_get_src(ctx
, &intr
->src
[0]);
1931 /* NOTE: only cmps.*.* can write p0.x: */
1932 cond
= ir3_CMPS_S(b
, cond
, 0, create_immed(b
, 0), 0);
1933 cond
->cat2
.condition
= IR3_COND_NE
;
1935 /* condition always goes in predicate register: */
1936 cond
->regs
[0]->num
= regid(REG_P0
, 0);
1938 kill
= ir3_PREDT(b
, cond
, 0);
1940 kill
->barrier_class
= IR3_BARRIER_EVERYTHING
;
1941 kill
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
1943 array_insert(ctx
->ir
, ctx
->ir
->predicates
, kill
);
1944 array_insert(b
, b
->keeps
, kill
);
1948 case nir_intrinsic_load_shared_ir3
:
1949 emit_intrinsic_load_shared_ir3(ctx
, intr
, dst
);
1951 case nir_intrinsic_store_shared_ir3
:
1952 emit_intrinsic_store_shared_ir3(ctx
, intr
);
1954 case nir_intrinsic_bindless_resource_ir3
:
1955 dst
[0] = ir3_get_src(ctx
, &intr
->src
[0])[0];
1958 ir3_context_error(ctx
, "Unhandled intrinsic type: %s\n",
1959 nir_intrinsic_infos
[intr
->intrinsic
].name
);
1964 ir3_put_dst(ctx
, &intr
->dest
);
1968 emit_load_const(struct ir3_context
*ctx
, nir_load_const_instr
*instr
)
1970 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &instr
->def
,
1971 instr
->def
.num_components
);
1973 if (instr
->def
.bit_size
== 16) {
1974 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1975 dst
[i
] = create_immed_typed(ctx
->block
,
1976 instr
->value
[i
].u16
,
1979 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1980 dst
[i
] = create_immed_typed(ctx
->block
,
1981 instr
->value
[i
].u32
,
1988 emit_undef(struct ir3_context
*ctx
, nir_ssa_undef_instr
*undef
)
1990 struct ir3_instruction
**dst
= ir3_get_dst_ssa(ctx
, &undef
->def
,
1991 undef
->def
.num_components
);
1992 type_t type
= (undef
->def
.bit_size
== 16) ? TYPE_U16
: TYPE_U32
;
1994 /* backend doesn't want undefined instructions, so just plug
1997 for (int i
= 0; i
< undef
->def
.num_components
; i
++)
1998 dst
[i
] = create_immed_typed(ctx
->block
, fui(0.0), type
);
2002 * texture fetch/sample instructions:
2006 get_tex_dest_type(nir_tex_instr
*tex
)
2010 switch (nir_alu_type_get_base_type(tex
->dest_type
)) {
2011 case nir_type_invalid
:
2012 case nir_type_float
:
2013 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_F16
: TYPE_F32
;
2016 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_S16
: TYPE_S32
;
2020 type
= nir_dest_bit_size(tex
->dest
) == 16 ? TYPE_U16
: TYPE_U32
;
2023 unreachable("bad dest_type");
2030 tex_info(nir_tex_instr
*tex
, unsigned *flagsp
, unsigned *coordsp
)
2032 unsigned coords
= glsl_get_sampler_dim_coordinate_components(tex
->sampler_dim
);
2035 /* note: would use tex->coord_components.. except txs.. also,
2036 * since array index goes after shadow ref, we don't want to
2040 flags
|= IR3_INSTR_3D
;
2042 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
2043 flags
|= IR3_INSTR_S
;
2045 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
)
2046 flags
|= IR3_INSTR_A
;
2052 /* Gets the sampler/texture idx as a hvec2. Which could either be dynamic
2053 * or immediate (in which case it will get lowered later to a non .s2en
2054 * version of the tex instruction which encode tex/samp as immediates:
2056 static struct tex_src_info
2057 get_tex_samp_tex_src(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2059 struct ir3_block
*b
= ctx
->block
;
2060 struct tex_src_info info
= { 0 };
2061 int texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_handle
);
2062 int sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_handle
);
2063 struct ir3_instruction
*texture
, *sampler
;
2065 if (texture_idx
>= 0 || sampler_idx
>= 0) {
2067 info
.flags
|= IR3_INSTR_B
;
2069 /* Gather information required to determine which encoding to
2070 * choose as well as for prefetch.
2072 nir_intrinsic_instr
*bindless_tex
= NULL
;
2074 if (texture_idx
>= 0) {
2075 ctx
->so
->bindless_tex
= true;
2076 bindless_tex
= ir3_bindless_resource(tex
->src
[texture_idx
].src
);
2077 assert(bindless_tex
);
2078 info
.tex_base
= nir_intrinsic_desc_set(bindless_tex
);
2079 tex_const
= nir_src_is_const(bindless_tex
->src
[0]);
2081 info
.tex_idx
= nir_src_as_uint(bindless_tex
->src
[0]);
2083 /* To simplify some of the logic below, assume the index is
2084 * constant 0 when it's not enabled.
2089 nir_intrinsic_instr
*bindless_samp
= NULL
;
2091 if (sampler_idx
>= 0) {
2092 ctx
->so
->bindless_samp
= true;
2093 bindless_samp
= ir3_bindless_resource(tex
->src
[sampler_idx
].src
);
2094 assert(bindless_samp
);
2095 info
.samp_base
= nir_intrinsic_desc_set(bindless_samp
);
2096 samp_const
= nir_src_is_const(bindless_samp
->src
[0]);
2098 info
.samp_idx
= nir_src_as_uint(bindless_samp
->src
[0]);
2104 /* Choose encoding. */
2105 if (tex_const
&& samp_const
&& info
.tex_idx
< 256 && info
.samp_idx
< 256) {
2106 if (info
.tex_idx
< 16 && info
.samp_idx
< 16 &&
2107 (!bindless_tex
|| !bindless_samp
|| info
.tex_base
== info
.samp_base
)) {
2108 /* Everything fits within the instruction */
2109 info
.base
= info
.tex_base
;
2110 info
.combined_idx
= info
.samp_idx
| (info
.tex_idx
<< 4);
2112 info
.base
= info
.tex_base
;
2113 info
.a1_val
= info
.tex_idx
<< 3 | info
.samp_base
;
2114 info
.combined_idx
= info
.samp_idx
;
2115 info
.flags
|= IR3_INSTR_A1EN
;
2117 info
.samp_tex
= NULL
;
2119 info
.flags
|= IR3_INSTR_S2EN
;
2120 /* In the indirect case, we only use a1.x to store the sampler
2121 * base if it differs from the texture base.
2123 if (!bindless_tex
|| !bindless_samp
|| info
.tex_base
== info
.samp_base
) {
2124 info
.base
= info
.tex_base
;
2126 info
.base
= info
.tex_base
;
2127 info
.a1_val
= info
.samp_base
;
2128 info
.flags
|= IR3_INSTR_A1EN
;
2131 /* Note: the indirect source is now a vec2 instead of hvec2, and
2132 * for some reason the texture and sampler are swapped.
2134 struct ir3_instruction
*texture
, *sampler
;
2137 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
2139 texture
= create_immed(b
, 0);
2142 if (bindless_samp
) {
2143 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
2145 sampler
= create_immed(b
, 0);
2147 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2153 info
.flags
|= IR3_INSTR_S2EN
;
2154 texture_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
);
2155 sampler_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
);
2156 if (texture_idx
>= 0) {
2157 texture
= ir3_get_src(ctx
, &tex
->src
[texture_idx
].src
)[0];
2158 texture
= ir3_COV(ctx
->block
, texture
, TYPE_U32
, TYPE_U16
);
2160 /* TODO what to do for dynamic case? I guess we only need the
2161 * max index for astc srgb workaround so maybe not a problem
2162 * to worry about if we don't enable indirect samplers for
2165 ctx
->max_texture_index
= MAX2(ctx
->max_texture_index
, tex
->texture_index
);
2166 texture
= create_immed_typed(ctx
->block
, tex
->texture_index
, TYPE_U16
);
2167 info
.tex_idx
= tex
->texture_index
;
2170 if (sampler_idx
>= 0) {
2171 sampler
= ir3_get_src(ctx
, &tex
->src
[sampler_idx
].src
)[0];
2172 sampler
= ir3_COV(ctx
->block
, sampler
, TYPE_U32
, TYPE_U16
);
2174 sampler
= create_immed_typed(ctx
->block
, tex
->sampler_index
, TYPE_U16
);
2175 info
.samp_idx
= tex
->texture_index
;
2178 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2188 emit_tex(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2190 struct ir3_block
*b
= ctx
->block
;
2191 struct ir3_instruction
**dst
, *sam
, *src0
[12], *src1
[4];
2192 struct ir3_instruction
* const *coord
, * const *off
, * const *ddx
, * const *ddy
;
2193 struct ir3_instruction
*lod
, *compare
, *proj
, *sample_index
;
2194 struct tex_src_info info
= { 0 };
2195 bool has_bias
= false, has_lod
= false, has_proj
= false, has_off
= false;
2196 unsigned i
, coords
, flags
, ncomp
;
2197 unsigned nsrc0
= 0, nsrc1
= 0;
2201 ncomp
= nir_dest_num_components(tex
->dest
);
2203 coord
= off
= ddx
= ddy
= NULL
;
2204 lod
= proj
= compare
= sample_index
= NULL
;
2206 dst
= ir3_get_dst(ctx
, &tex
->dest
, ncomp
);
2208 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
2209 switch (tex
->src
[i
].src_type
) {
2210 case nir_tex_src_coord
:
2211 coord
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2213 case nir_tex_src_bias
:
2214 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2217 case nir_tex_src_lod
:
2218 lod
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2221 case nir_tex_src_comparator
: /* shadow comparator */
2222 compare
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2224 case nir_tex_src_projector
:
2225 proj
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2228 case nir_tex_src_offset
:
2229 off
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2232 case nir_tex_src_ddx
:
2233 ddx
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2235 case nir_tex_src_ddy
:
2236 ddy
= ir3_get_src(ctx
, &tex
->src
[i
].src
);
2238 case nir_tex_src_ms_index
:
2239 sample_index
= ir3_get_src(ctx
, &tex
->src
[i
].src
)[0];
2241 case nir_tex_src_texture_offset
:
2242 case nir_tex_src_sampler_offset
:
2243 case nir_tex_src_texture_handle
:
2244 case nir_tex_src_sampler_handle
:
2245 /* handled in get_tex_samp_src() */
2248 ir3_context_error(ctx
, "Unhandled NIR tex src type: %d\n",
2249 tex
->src
[i
].src_type
);
2255 case nir_texop_tex_prefetch
:
2256 compile_assert(ctx
, !has_bias
);
2257 compile_assert(ctx
, !has_lod
);
2258 compile_assert(ctx
, !compare
);
2259 compile_assert(ctx
, !has_proj
);
2260 compile_assert(ctx
, !has_off
);
2261 compile_assert(ctx
, !ddx
);
2262 compile_assert(ctx
, !ddy
);
2263 compile_assert(ctx
, !sample_index
);
2264 compile_assert(ctx
, nir_tex_instr_src_index(tex
, nir_tex_src_texture_offset
) < 0);
2265 compile_assert(ctx
, nir_tex_instr_src_index(tex
, nir_tex_src_sampler_offset
) < 0);
2267 if (ctx
->so
->num_sampler_prefetch
< IR3_MAX_SAMPLER_PREFETCH
) {
2268 opc
= OPC_META_TEX_PREFETCH
;
2269 ctx
->so
->num_sampler_prefetch
++;
2273 case nir_texop_tex
: opc
= has_lod
? OPC_SAML
: OPC_SAM
; break;
2274 case nir_texop_txb
: opc
= OPC_SAMB
; break;
2275 case nir_texop_txl
: opc
= OPC_SAML
; break;
2276 case nir_texop_txd
: opc
= OPC_SAMGQ
; break;
2277 case nir_texop_txf
: opc
= OPC_ISAML
; break;
2278 case nir_texop_lod
: opc
= OPC_GETLOD
; break;
2280 /* NOTE: a4xx might need to emulate gather w/ txf (this is
2281 * what blob does, seems gather is broken?), and a3xx did
2282 * not support it (but probably could also emulate).
2284 switch (tex
->component
) {
2285 case 0: opc
= OPC_GATHER4R
; break;
2286 case 1: opc
= OPC_GATHER4G
; break;
2287 case 2: opc
= OPC_GATHER4B
; break;
2288 case 3: opc
= OPC_GATHER4A
; break;
2291 case nir_texop_txf_ms_fb
:
2292 case nir_texop_txf_ms
: opc
= OPC_ISAMM
; break;
2294 ir3_context_error(ctx
, "Unhandled NIR tex type: %d\n", tex
->op
);
2298 tex_info(tex
, &flags
, &coords
);
2301 * lay out the first argument in the proper order:
2302 * - actual coordinates first
2303 * - shadow reference
2306 * - starting at offset 4, dpdx.xy, dpdy.xy
2308 * bias/lod go into the second arg
2311 /* insert tex coords: */
2312 for (i
= 0; i
< coords
; i
++)
2317 /* scale up integer coords for TXF based on the LOD */
2318 if (ctx
->compiler
->unminify_coords
&& (opc
== OPC_ISAML
)) {
2320 for (i
= 0; i
< coords
; i
++)
2321 src0
[i
] = ir3_SHL_B(b
, src0
[i
], 0, lod
, 0);
2325 /* hw doesn't do 1d, so we treat it as 2d with
2326 * height of 1, and patch up the y coord.
2329 src0
[nsrc0
++] = create_immed(b
, 0);
2331 src0
[nsrc0
++] = create_immed(b
, fui(0.5));
2335 if (tex
->is_shadow
&& tex
->op
!= nir_texop_lod
)
2336 src0
[nsrc0
++] = compare
;
2338 if (tex
->is_array
&& tex
->op
!= nir_texop_lod
) {
2339 struct ir3_instruction
*idx
= coord
[coords
];
2341 /* the array coord for cube arrays needs 0.5 added to it */
2342 if (ctx
->compiler
->array_index_add_half
&& !is_isam(opc
))
2343 idx
= ir3_ADD_F(b
, idx
, 0, create_immed(b
, fui(0.5)), 0);
2345 src0
[nsrc0
++] = idx
;
2349 src0
[nsrc0
++] = proj
;
2350 flags
|= IR3_INSTR_P
;
2353 /* pad to 4, then ddx/ddy: */
2354 if (tex
->op
== nir_texop_txd
) {
2356 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2357 for (i
= 0; i
< coords
; i
++)
2358 src0
[nsrc0
++] = ddx
[i
];
2360 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2361 for (i
= 0; i
< coords
; i
++)
2362 src0
[nsrc0
++] = ddy
[i
];
2364 src0
[nsrc0
++] = create_immed(b
, fui(0.0));
2367 /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
2368 * with scaled x coord according to requested sample:
2370 if (opc
== OPC_ISAMM
) {
2371 if (ctx
->compiler
->txf_ms_with_isaml
) {
2372 /* the samples are laid out in x dimension as
2374 * x_ms = (x << ms) + sample_index;
2376 struct ir3_instruction
*ms
;
2377 ms
= create_immed(b
, (ctx
->samples
>> (2 * tex
->texture_index
)) & 3);
2379 src0
[0] = ir3_SHL_B(b
, src0
[0], 0, ms
, 0);
2380 src0
[0] = ir3_ADD_U(b
, src0
[0], 0, sample_index
, 0);
2384 src0
[nsrc0
++] = sample_index
;
2389 * second argument (if applicable):
2394 if (has_off
| has_lod
| has_bias
) {
2396 unsigned off_coords
= coords
;
2397 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2399 for (i
= 0; i
< off_coords
; i
++)
2400 src1
[nsrc1
++] = off
[i
];
2402 src1
[nsrc1
++] = create_immed(b
, fui(0.0));
2403 flags
|= IR3_INSTR_O
;
2406 if (has_lod
| has_bias
)
2407 src1
[nsrc1
++] = lod
;
2410 type
= get_tex_dest_type(tex
);
2412 if (opc
== OPC_GETLOD
)
2416 if (tex
->op
== nir_texop_txf_ms_fb
) {
2417 /* only expect a single txf_ms_fb per shader: */
2418 compile_assert(ctx
, !ctx
->so
->fb_read
);
2419 compile_assert(ctx
, ctx
->so
->type
== MESA_SHADER_FRAGMENT
);
2421 ctx
->so
->fb_read
= true;
2422 info
.samp_tex
= ir3_create_collect(ctx
, (struct ir3_instruction
*[]){
2423 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2424 create_immed_typed(ctx
->block
, ctx
->so
->num_samp
, TYPE_U16
),
2426 info
.flags
= IR3_INSTR_S2EN
;
2428 ctx
->so
->num_samp
++;
2430 info
= get_tex_samp_tex_src(ctx
, tex
);
2433 struct ir3_instruction
*col0
= ir3_create_collect(ctx
, src0
, nsrc0
);
2434 struct ir3_instruction
*col1
= ir3_create_collect(ctx
, src1
, nsrc1
);
2436 if (opc
== OPC_META_TEX_PREFETCH
) {
2437 int idx
= nir_tex_instr_src_index(tex
, nir_tex_src_coord
);
2439 compile_assert(ctx
, tex
->src
[idx
].src
.is_ssa
);
2441 sam
= ir3_META_TEX_PREFETCH(b
);
2442 __ssa_dst(sam
)->wrmask
= MASK(ncomp
); /* dst */
2443 __ssa_src(sam
, get_barycentric_pixel(ctx
), 0);
2444 sam
->prefetch
.input_offset
=
2445 ir3_nir_coord_offset(tex
->src
[idx
].src
.ssa
);
2446 /* make sure not to add irrelevant flags like S2EN */
2447 sam
->flags
= flags
| (info
.flags
& IR3_INSTR_B
);
2448 sam
->prefetch
.tex
= info
.tex_idx
;
2449 sam
->prefetch
.samp
= info
.samp_idx
;
2450 sam
->prefetch
.tex_base
= info
.tex_base
;
2451 sam
->prefetch
.samp_base
= info
.samp_base
;
2453 info
.flags
|= flags
;
2454 sam
= emit_sam(ctx
, opc
, info
, type
, MASK(ncomp
), col0
, col1
);
2457 if ((ctx
->astc_srgb
& (1 << tex
->texture_index
)) && !nir_tex_instr_is_query(tex
)) {
2458 assert(opc
!= OPC_META_TEX_PREFETCH
);
2460 /* only need first 3 components: */
2461 sam
->regs
[0]->wrmask
= 0x7;
2462 ir3_split_dest(b
, dst
, sam
, 0, 3);
2464 /* we need to sample the alpha separately with a non-ASTC
2467 sam
= ir3_SAM(b
, opc
, type
, 0b1000, flags
| info
.flags
,
2468 info
.samp_tex
, col0
, col1
);
2470 array_insert(ctx
->ir
, ctx
->ir
->astc_srgb
, sam
);
2472 /* fixup .w component: */
2473 ir3_split_dest(b
, &dst
[3], sam
, 3, 1);
2475 /* normal (non-workaround) case: */
2476 ir3_split_dest(b
, dst
, sam
, 0, ncomp
);
2479 /* GETLOD returns results in 4.8 fixed point */
2480 if (opc
== OPC_GETLOD
) {
2481 struct ir3_instruction
*factor
= create_immed(b
, fui(1.0 / 256));
2483 compile_assert(ctx
, tex
->dest_type
== nir_type_float
);
2484 for (i
= 0; i
< 2; i
++) {
2485 dst
[i
] = ir3_MUL_F(b
, ir3_COV(b
, dst
[i
], TYPE_S32
, TYPE_F32
), 0,
2490 ir3_put_dst(ctx
, &tex
->dest
);
2494 emit_tex_info(struct ir3_context
*ctx
, nir_tex_instr
*tex
, unsigned idx
)
2496 struct ir3_block
*b
= ctx
->block
;
2497 struct ir3_instruction
**dst
, *sam
;
2498 type_t dst_type
= get_tex_dest_type(tex
);
2499 struct tex_src_info info
= get_tex_samp_tex_src(ctx
, tex
);
2501 dst
= ir3_get_dst(ctx
, &tex
->dest
, 1);
2503 sam
= emit_sam(ctx
, OPC_GETINFO
, info
, dst_type
, 1 << idx
, NULL
, NULL
);
2505 /* even though there is only one component, since it ends
2506 * up in .y/.z/.w rather than .x, we need a split_dest()
2508 ir3_split_dest(b
, dst
, sam
, idx
, 1);
2510 /* The # of levels comes from getinfo.z. We need to add 1 to it, since
2511 * the value in TEX_CONST_0 is zero-based.
2513 if (ctx
->compiler
->levels_add_one
)
2514 dst
[0] = ir3_ADD_U(b
, dst
[0], 0, create_immed(b
, 1), 0);
2516 ir3_put_dst(ctx
, &tex
->dest
);
2520 emit_tex_txs(struct ir3_context
*ctx
, nir_tex_instr
*tex
)
2522 struct ir3_block
*b
= ctx
->block
;
2523 struct ir3_instruction
**dst
, *sam
;
2524 struct ir3_instruction
*lod
;
2525 unsigned flags
, coords
;
2526 type_t dst_type
= get_tex_dest_type(tex
);
2527 struct tex_src_info info
= get_tex_samp_tex_src(ctx
, tex
);
2529 tex_info(tex
, &flags
, &coords
);
2530 info
.flags
|= flags
;
2532 /* Actually we want the number of dimensions, not coordinates. This
2533 * distinction only matters for cubes.
2535 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2538 dst
= ir3_get_dst(ctx
, &tex
->dest
, 4);
2540 int lod_idx
= nir_tex_instr_src_index(tex
, nir_tex_src_lod
);
2541 compile_assert(ctx
, lod_idx
>= 0);
2543 lod
= ir3_get_src(ctx
, &tex
->src
[lod_idx
].src
)[0];
2545 sam
= emit_sam(ctx
, OPC_GETSIZE
, info
, dst_type
, 0b1111, lod
, NULL
);
2546 ir3_split_dest(b
, dst
, sam
, 0, 4);
2548 /* Array size actually ends up in .w rather than .z. This doesn't
2549 * matter for miplevel 0, but for higher mips the value in z is
2550 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2551 * returned, which means that we have to add 1 to it for arrays.
2553 if (tex
->is_array
) {
2554 if (ctx
->compiler
->levels_add_one
) {
2555 dst
[coords
] = ir3_ADD_U(b
, dst
[3], 0, create_immed(b
, 1), 0);
2557 dst
[coords
] = ir3_MOV(b
, dst
[3], TYPE_U32
);
2561 ir3_put_dst(ctx
, &tex
->dest
);
2565 emit_jump(struct ir3_context
*ctx
, nir_jump_instr
*jump
)
2567 switch (jump
->type
) {
2568 case nir_jump_break
:
2569 case nir_jump_continue
:
2570 case nir_jump_return
:
2571 /* I *think* we can simply just ignore this, and use the
2572 * successor block link to figure out where we need to
2573 * jump to for break/continue
2577 ir3_context_error(ctx
, "Unhandled NIR jump type: %d\n", jump
->type
);
2583 emit_instr(struct ir3_context
*ctx
, nir_instr
*instr
)
2585 switch (instr
->type
) {
2586 case nir_instr_type_alu
:
2587 emit_alu(ctx
, nir_instr_as_alu(instr
));
2589 case nir_instr_type_deref
:
2590 /* ignored, handled as part of the intrinsic they are src to */
2592 case nir_instr_type_intrinsic
:
2593 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2595 case nir_instr_type_load_const
:
2596 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2598 case nir_instr_type_ssa_undef
:
2599 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
2601 case nir_instr_type_tex
: {
2602 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
2603 /* couple tex instructions get special-cased:
2607 emit_tex_txs(ctx
, tex
);
2609 case nir_texop_query_levels
:
2610 emit_tex_info(ctx
, tex
, 2);
2612 case nir_texop_texture_samples
:
2613 emit_tex_info(ctx
, tex
, 3);
2621 case nir_instr_type_jump
:
2622 emit_jump(ctx
, nir_instr_as_jump(instr
));
2624 case nir_instr_type_phi
:
2625 /* we have converted phi webs to regs in NIR by now */
2626 ir3_context_error(ctx
, "Unexpected NIR instruction type: %d\n", instr
->type
);
2628 case nir_instr_type_call
:
2629 case nir_instr_type_parallel_copy
:
2630 ir3_context_error(ctx
, "Unhandled NIR instruction type: %d\n", instr
->type
);
2635 static struct ir3_block
*
2636 get_block(struct ir3_context
*ctx
, const nir_block
*nblock
)
2638 struct ir3_block
*block
;
2639 struct hash_entry
*hentry
;
2641 hentry
= _mesa_hash_table_search(ctx
->block_ht
, nblock
);
2643 return hentry
->data
;
2645 block
= ir3_block_create(ctx
->ir
);
2646 block
->nblock
= nblock
;
2647 _mesa_hash_table_insert(ctx
->block_ht
, nblock
, block
);
2649 set_foreach(nblock
->predecessors
, sentry
) {
2650 _mesa_set_add(block
->predecessors
, get_block(ctx
, sentry
->key
));
2657 emit_block(struct ir3_context
*ctx
, nir_block
*nblock
)
2659 struct ir3_block
*block
= get_block(ctx
, nblock
);
2661 for (int i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
2662 if (nblock
->successors
[i
]) {
2663 block
->successors
[i
] =
2664 get_block(ctx
, nblock
->successors
[i
]);
2669 list_addtail(&block
->node
, &ctx
->ir
->block_list
);
2671 /* re-emit addr register in each block if needed: */
2672 for (int i
= 0; i
< ARRAY_SIZE(ctx
->addr0_ht
); i
++) {
2673 _mesa_hash_table_destroy(ctx
->addr0_ht
[i
], NULL
);
2674 ctx
->addr0_ht
[i
] = NULL
;
2677 _mesa_hash_table_u64_destroy(ctx
->addr1_ht
, NULL
);
2678 ctx
->addr1_ht
= NULL
;
2680 nir_foreach_instr (instr
, nblock
) {
2681 ctx
->cur_instr
= instr
;
2682 emit_instr(ctx
, instr
);
2683 ctx
->cur_instr
= NULL
;
2688 _mesa_hash_table_clear(ctx
->sel_cond_conversions
, NULL
);
2691 static void emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
);
2694 emit_if(struct ir3_context
*ctx
, nir_if
*nif
)
2696 struct ir3_instruction
*condition
= ir3_get_src(ctx
, &nif
->condition
)[0];
2698 ctx
->block
->condition
= ir3_get_predicate(ctx
, condition
);
2700 emit_cf_list(ctx
, &nif
->then_list
);
2701 emit_cf_list(ctx
, &nif
->else_list
);
2705 emit_loop(struct ir3_context
*ctx
, nir_loop
*nloop
)
2707 emit_cf_list(ctx
, &nloop
->body
);
2712 stack_push(struct ir3_context
*ctx
)
2715 ctx
->max_stack
= MAX2(ctx
->max_stack
, ctx
->stack
);
2719 stack_pop(struct ir3_context
*ctx
)
2721 compile_assert(ctx
, ctx
->stack
> 0);
2726 emit_cf_list(struct ir3_context
*ctx
, struct exec_list
*list
)
2728 foreach_list_typed (nir_cf_node
, node
, node
, list
) {
2729 switch (node
->type
) {
2730 case nir_cf_node_block
:
2731 emit_block(ctx
, nir_cf_node_as_block(node
));
2733 case nir_cf_node_if
:
2735 emit_if(ctx
, nir_cf_node_as_if(node
));
2738 case nir_cf_node_loop
:
2740 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2743 case nir_cf_node_function
:
2744 ir3_context_error(ctx
, "TODO\n");
2750 /* emit stream-out code. At this point, the current block is the original
2751 * (nir) end block, and nir ensures that all flow control paths terminate
2752 * into the end block. We re-purpose the original end block to generate
2753 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2754 * block holding stream-out write instructions, followed by the new end
2758 * p0.x = (vtxcnt < maxvtxcnt)
2759 * // succs: blockStreamOut, blockNewEnd
2762 * // preds: blockOrigEnd
2763 * ... stream-out instructions ...
2764 * // succs: blockNewEnd
2767 * // preds: blockOrigEnd, blockStreamOut
2771 emit_stream_out(struct ir3_context
*ctx
)
2773 struct ir3
*ir
= ctx
->ir
;
2774 struct ir3_stream_output_info
*strmout
=
2775 &ctx
->so
->shader
->stream_output
;
2776 struct ir3_block
*orig_end_block
, *stream_out_block
, *new_end_block
;
2777 struct ir3_instruction
*vtxcnt
, *maxvtxcnt
, *cond
;
2778 struct ir3_instruction
*bases
[IR3_MAX_SO_BUFFERS
];
2780 /* create vtxcnt input in input block at top of shader,
2781 * so that it is seen as live over the entire duration
2784 vtxcnt
= create_sysval_input(ctx
, SYSTEM_VALUE_VERTEX_CNT
, 0x1);
2785 maxvtxcnt
= create_driver_param(ctx
, IR3_DP_VTXCNT_MAX
);
2787 /* at this point, we are at the original 'end' block,
2788 * re-purpose this block to stream-out condition, then
2789 * append stream-out block and new-end block
2791 orig_end_block
= ctx
->block
;
2793 // maybe w/ store_global intrinsic, we could do this
2794 // stuff in nir->nir pass
2796 stream_out_block
= ir3_block_create(ir
);
2797 list_addtail(&stream_out_block
->node
, &ir
->block_list
);
2799 new_end_block
= ir3_block_create(ir
);
2800 list_addtail(&new_end_block
->node
, &ir
->block_list
);
2802 orig_end_block
->successors
[0] = stream_out_block
;
2803 orig_end_block
->successors
[1] = new_end_block
;
2805 stream_out_block
->successors
[0] = new_end_block
;
2806 _mesa_set_add(stream_out_block
->predecessors
, orig_end_block
);
2808 _mesa_set_add(new_end_block
->predecessors
, orig_end_block
);
2809 _mesa_set_add(new_end_block
->predecessors
, stream_out_block
);
2811 /* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2812 cond
= ir3_CMPS_S(ctx
->block
, vtxcnt
, 0, maxvtxcnt
, 0);
2813 cond
->regs
[0]->num
= regid(REG_P0
, 0);
2814 cond
->regs
[0]->flags
&= ~IR3_REG_SSA
;
2815 cond
->cat2
.condition
= IR3_COND_LT
;
2817 /* condition goes on previous block to the conditional,
2818 * since it is used to pick which of the two successor
2821 orig_end_block
->condition
= cond
;
2823 /* switch to stream_out_block to generate the stream-out
2826 ctx
->block
= stream_out_block
;
2828 /* Calculate base addresses based on vtxcnt. Instructions
2829 * generated for bases not used in following loop will be
2830 * stripped out in the backend.
2832 for (unsigned i
= 0; i
< IR3_MAX_SO_BUFFERS
; i
++) {
2833 struct ir3_const_state
*const_state
= &ctx
->so
->shader
->const_state
;
2834 unsigned stride
= strmout
->stride
[i
];
2835 struct ir3_instruction
*base
, *off
;
2837 base
= create_uniform(ctx
->block
, regid(const_state
->offsets
.tfbo
, i
));
2839 /* 24-bit should be enough: */
2840 off
= ir3_MUL_U24(ctx
->block
, vtxcnt
, 0,
2841 create_immed(ctx
->block
, stride
* 4), 0);
2843 bases
[i
] = ir3_ADD_S(ctx
->block
, off
, 0, base
, 0);
2846 /* Generate the per-output store instructions: */
2847 for (unsigned i
= 0; i
< strmout
->num_outputs
; i
++) {
2848 for (unsigned j
= 0; j
< strmout
->output
[i
].num_components
; j
++) {
2849 unsigned c
= j
+ strmout
->output
[i
].start_component
;
2850 struct ir3_instruction
*base
, *out
, *stg
;
2852 base
= bases
[strmout
->output
[i
].output_buffer
];
2853 out
= ctx
->outputs
[regid(strmout
->output
[i
].register_index
, c
)];
2855 stg
= ir3_STG(ctx
->block
, base
, 0, out
, 0,
2856 create_immed(ctx
->block
, 1), 0);
2857 stg
->cat6
.type
= TYPE_U32
;
2858 stg
->cat6
.dst_offset
= (strmout
->output
[i
].dst_offset
+ j
) * 4;
2860 array_insert(ctx
->block
, ctx
->block
->keeps
, stg
);
2864 /* and finally switch to the new_end_block: */
2865 ctx
->block
= new_end_block
;
2869 emit_function(struct ir3_context
*ctx
, nir_function_impl
*impl
)
2871 nir_metadata_require(impl
, nir_metadata_block_index
);
2873 compile_assert(ctx
, ctx
->stack
== 0);
2875 emit_cf_list(ctx
, &impl
->body
);
2876 emit_block(ctx
, impl
->end_block
);
2878 compile_assert(ctx
, ctx
->stack
== 0);
2880 /* at this point, we should have a single empty block,
2881 * into which we emit the 'end' instruction.
2883 compile_assert(ctx
, list_is_empty(&ctx
->block
->instr_list
));
2885 /* If stream-out (aka transform-feedback) enabled, emit the
2886 * stream-out instructions, followed by a new empty block (into
2887 * which the 'end' instruction lands).
2889 * NOTE: it is done in this order, rather than inserting before
2890 * we emit end_block, because NIR guarantees that all blocks
2891 * flow into end_block, and that end_block has no successors.
2892 * So by re-purposing end_block as the first block of stream-
2893 * out, we guarantee that all exit paths flow into the stream-
2896 if ((ctx
->compiler
->gpu_id
< 500) &&
2897 (ctx
->so
->shader
->stream_output
.num_outputs
> 0) &&
2898 !ctx
->so
->binning_pass
) {
2899 debug_assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
2900 emit_stream_out(ctx
);
2903 /* Vertex shaders in a tessellation or geometry pipeline treat END as a
2904 * NOP and has an epilogue that writes the VS outputs to local storage, to
2905 * be read by the HS. Then it resets execution mask (chmask) and chains
2906 * to the next shader (chsh).
2908 if ((ctx
->so
->type
== MESA_SHADER_VERTEX
&&
2909 (ctx
->so
->key
.has_gs
|| ctx
->so
->key
.tessellation
)) ||
2910 (ctx
->so
->type
== MESA_SHADER_TESS_EVAL
&& ctx
->so
->key
.has_gs
)) {
2911 struct ir3_instruction
*chmask
=
2912 ir3_CHMASK(ctx
->block
);
2913 chmask
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2914 chmask
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2916 struct ir3_instruction
*chsh
=
2917 ir3_CHSH(ctx
->block
);
2918 chsh
->barrier_class
= IR3_BARRIER_EVERYTHING
;
2919 chsh
->barrier_conflict
= IR3_BARRIER_EVERYTHING
;
2921 ir3_END(ctx
->block
);
2926 setup_input(struct ir3_context
*ctx
, nir_variable
*in
)
2928 struct ir3_shader_variant
*so
= ctx
->so
;
2929 unsigned ncomp
= glsl_get_components(in
->type
);
2930 unsigned n
= in
->data
.driver_location
;
2931 unsigned frac
= in
->data
.location_frac
;
2932 unsigned slot
= in
->data
.location
;
2934 /* Inputs are loaded using ldlw or ldg for these stages. */
2935 if (ctx
->so
->type
== MESA_SHADER_TESS_CTRL
||
2936 ctx
->so
->type
== MESA_SHADER_TESS_EVAL
||
2937 ctx
->so
->type
== MESA_SHADER_GEOMETRY
)
2940 /* skip unread inputs, we could end up with (for example), unsplit
2941 * matrix/etc inputs in the case they are not read, so just silently
2947 so
->inputs
[n
].slot
= slot
;
2948 so
->inputs
[n
].compmask
|= (1 << (ncomp
+ frac
)) - 1;
2949 so
->inputs_count
= MAX2(so
->inputs_count
, n
+ 1);
2950 so
->inputs
[n
].interpolate
= in
->data
.interpolation
;
2952 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
2954 /* if any varyings have 'sample' qualifer, that triggers us
2955 * to run in per-sample mode:
2957 so
->per_samp
|= in
->data
.sample
;
2959 for (int i
= 0; i
< ncomp
; i
++) {
2960 struct ir3_instruction
*instr
= NULL
;
2961 unsigned idx
= (n
* 4) + i
+ frac
;
2963 if (slot
== VARYING_SLOT_POS
) {
2964 ir3_context_error(ctx
, "fragcoord should be a sysval!\n");
2966 /* detect the special case for front/back colors where
2967 * we need to do flat vs smooth shading depending on
2970 if (in
->data
.interpolation
== INTERP_MODE_NONE
) {
2972 case VARYING_SLOT_COL0
:
2973 case VARYING_SLOT_COL1
:
2974 case VARYING_SLOT_BFC0
:
2975 case VARYING_SLOT_BFC1
:
2976 so
->inputs
[n
].rasterflat
= true;
2983 if (ctx
->compiler
->flat_bypass
) {
2984 if ((so
->inputs
[n
].interpolate
== INTERP_MODE_FLAT
) ||
2985 (so
->inputs
[n
].rasterflat
&& ctx
->so
->key
.rasterflat
))
2986 so
->inputs
[n
].use_ldlv
= true;
2989 so
->inputs
[n
].bary
= true;
2991 instr
= create_frag_input(ctx
, so
->inputs
[n
].use_ldlv
, idx
);
2994 compile_assert(ctx
, idx
< ctx
->ninputs
);
2996 ctx
->inputs
[idx
] = instr
;
2998 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
2999 struct ir3_instruction
*input
= NULL
, *in
;
3000 struct ir3_instruction
*components
[4];
3001 unsigned mask
= (1 << (ncomp
+ frac
)) - 1;
3003 foreach_input (in
, ctx
->ir
) {
3004 if (in
->input
.inidx
== n
) {
3011 input
= create_input(ctx
, mask
);
3012 input
->input
.inidx
= n
;
3014 input
->regs
[0]->wrmask
|= mask
;
3017 ir3_split_dest(ctx
->block
, components
, input
, frac
, ncomp
);
3019 for (int i
= 0; i
< ncomp
; i
++) {
3020 unsigned idx
= (n
* 4) + i
+ frac
;
3021 compile_assert(ctx
, idx
< ctx
->ninputs
);
3022 ctx
->inputs
[idx
] = components
[i
];
3025 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
3028 if (so
->inputs
[n
].bary
|| (ctx
->so
->type
== MESA_SHADER_VERTEX
)) {
3029 so
->total_in
+= ncomp
;
3033 /* Initially we assign non-packed inloc's for varyings, as we don't really
3034 * know up-front which components will be unused. After all the compilation
3035 * stages we scan the shader to see which components are actually used, and
3036 * re-pack the inlocs to eliminate unneeded varyings.
3039 pack_inlocs(struct ir3_context
*ctx
)
3041 struct ir3_shader_variant
*so
= ctx
->so
;
3042 uint8_t used_components
[so
->inputs_count
];
3044 memset(used_components
, 0, sizeof(used_components
));
3047 * First Step: scan shader to find which bary.f/ldlv remain:
3050 foreach_block (block
, &ctx
->ir
->block_list
) {
3051 foreach_instr (instr
, &block
->instr_list
) {
3052 if (is_input(instr
)) {
3053 unsigned inloc
= instr
->regs
[1]->iim_val
;
3054 unsigned i
= inloc
/ 4;
3055 unsigned j
= inloc
% 4;
3057 compile_assert(ctx
, instr
->regs
[1]->flags
& IR3_REG_IMMED
);
3058 compile_assert(ctx
, i
< so
->inputs_count
);
3060 used_components
[i
] |= 1 << j
;
3061 } else if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3062 for (int n
= 0; n
< 2; n
++) {
3063 unsigned inloc
= instr
->prefetch
.input_offset
+ n
;
3064 unsigned i
= inloc
/ 4;
3065 unsigned j
= inloc
% 4;
3067 compile_assert(ctx
, i
< so
->inputs_count
);
3069 used_components
[i
] |= 1 << j
;
3076 * Second Step: reassign varying inloc/slots:
3079 unsigned actual_in
= 0;
3082 for (unsigned i
= 0; i
< so
->inputs_count
; i
++) {
3083 unsigned compmask
= 0, maxcomp
= 0;
3085 so
->inputs
[i
].inloc
= inloc
;
3086 so
->inputs
[i
].bary
= false;
3088 for (unsigned j
= 0; j
< 4; j
++) {
3089 if (!(used_components
[i
] & (1 << j
)))
3092 compmask
|= (1 << j
);
3096 /* at this point, since used_components[i] mask is only
3097 * considering varyings (ie. not sysvals) we know this
3100 so
->inputs
[i
].bary
= true;
3103 if (so
->inputs
[i
].bary
) {
3105 so
->inputs
[i
].compmask
= (1 << maxcomp
) - 1;
3111 * Third Step: reassign packed inloc's:
3114 foreach_block (block
, &ctx
->ir
->block_list
) {
3115 foreach_instr (instr
, &block
->instr_list
) {
3116 if (is_input(instr
)) {
3117 unsigned inloc
= instr
->regs
[1]->iim_val
;
3118 unsigned i
= inloc
/ 4;
3119 unsigned j
= inloc
% 4;
3121 instr
->regs
[1]->iim_val
= so
->inputs
[i
].inloc
+ j
;
3122 } else if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3123 unsigned i
= instr
->prefetch
.input_offset
/ 4;
3124 unsigned j
= instr
->prefetch
.input_offset
% 4;
3125 instr
->prefetch
.input_offset
= so
->inputs
[i
].inloc
+ j
;
3132 setup_output(struct ir3_context
*ctx
, nir_variable
*out
)
3134 struct ir3_shader_variant
*so
= ctx
->so
;
3135 unsigned slots
= glsl_count_vec4_slots(out
->type
, false, false);
3136 unsigned ncomp
= glsl_get_components(glsl_without_array(out
->type
));
3137 unsigned n
= out
->data
.driver_location
;
3138 unsigned frac
= out
->data
.location_frac
;
3139 unsigned slot
= out
->data
.location
;
3141 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
3143 case FRAG_RESULT_DEPTH
:
3144 so
->writes_pos
= true;
3146 case FRAG_RESULT_COLOR
:
3149 case FRAG_RESULT_SAMPLE_MASK
:
3150 so
->writes_smask
= true;
3153 if (slot
>= FRAG_RESULT_DATA0
)
3155 ir3_context_error(ctx
, "unknown FS output name: %s\n",
3156 gl_frag_result_name(slot
));
3158 } else if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
3159 ctx
->so
->type
== MESA_SHADER_TESS_EVAL
||
3160 ctx
->so
->type
== MESA_SHADER_GEOMETRY
) {
3162 case VARYING_SLOT_POS
:
3163 so
->writes_pos
= true;
3165 case VARYING_SLOT_PSIZ
:
3166 so
->writes_psize
= true;
3168 case VARYING_SLOT_PRIMITIVE_ID
:
3169 case VARYING_SLOT_LAYER
:
3170 case VARYING_SLOT_GS_VERTEX_FLAGS_IR3
:
3171 debug_assert(ctx
->so
->type
== MESA_SHADER_GEOMETRY
);
3173 case VARYING_SLOT_COL0
:
3174 case VARYING_SLOT_COL1
:
3175 case VARYING_SLOT_BFC0
:
3176 case VARYING_SLOT_BFC1
:
3177 case VARYING_SLOT_FOGC
:
3178 case VARYING_SLOT_CLIP_DIST0
:
3179 case VARYING_SLOT_CLIP_DIST1
:
3180 case VARYING_SLOT_CLIP_VERTEX
:
3183 if (slot
>= VARYING_SLOT_VAR0
)
3185 if ((VARYING_SLOT_TEX0
<= slot
) && (slot
<= VARYING_SLOT_TEX7
))
3187 ir3_context_error(ctx
, "unknown %s shader output name: %s\n",
3188 _mesa_shader_stage_to_string(ctx
->so
->type
),
3189 gl_varying_slot_name(slot
));
3191 } else if (ctx
->so
->type
== MESA_SHADER_TESS_CTRL
) {
3192 /* output lowered to buffer writes. */
3195 ir3_context_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
3199 so
->outputs_count
= out
->data
.driver_location
+ slots
;
3200 compile_assert(ctx
, so
->outputs_count
< ARRAY_SIZE(so
->outputs
));
3202 for (int i
= 0; i
< slots
; i
++) {
3203 int slot_base
= n
+ i
;
3204 so
->outputs
[slot_base
].slot
= slot
+ i
;
3206 for (int i
= 0; i
< ncomp
; i
++) {
3207 unsigned idx
= (slot_base
* 4) + i
+ frac
;
3208 compile_assert(ctx
, idx
< ctx
->noutputs
);
3209 ctx
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
3212 /* if varying packing doesn't happen, we could end up in a situation
3213 * with "holes" in the output, and since the per-generation code that
3214 * sets up varying linkage registers doesn't expect to have more than
3215 * one varying per vec4 slot, pad the holes.
3217 * Note that this should probably generate a performance warning of
3220 for (int i
= 0; i
< frac
; i
++) {
3221 unsigned idx
= (slot_base
* 4) + i
;
3222 if (!ctx
->outputs
[idx
]) {
3223 ctx
->outputs
[idx
] = create_immed(ctx
->block
, fui(0.0));
3230 emit_instructions(struct ir3_context
*ctx
)
3232 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->s
);
3234 ctx
->ninputs
= ctx
->s
->num_inputs
* 4;
3235 ctx
->noutputs
= ctx
->s
->num_outputs
* 4;
3236 ctx
->inputs
= rzalloc_array(ctx
, struct ir3_instruction
*, ctx
->ninputs
);
3237 ctx
->outputs
= rzalloc_array(ctx
, struct ir3_instruction
*, ctx
->noutputs
);
3239 ctx
->ir
= ir3_create(ctx
->compiler
, ctx
->so
->type
);
3241 /* Create inputs in first block: */
3242 ctx
->block
= get_block(ctx
, nir_start_block(fxn
));
3243 ctx
->in_block
= ctx
->block
;
3245 /* for fragment shader, the vcoord input register is used as the
3246 * base for bary.f varying fetch instrs:
3248 * TODO defer creating ctx->ij_pixel and corresponding sysvals
3249 * until emit_intrinsic when we know they are actually needed.
3250 * For now, we defer creating ctx->ij_centroid, etc, since we
3251 * only need ij_pixel for "old style" varying inputs (ie.
3254 if (ctx
->so
->type
== MESA_SHADER_FRAGMENT
) {
3255 ctx
->ij_pixel
= create_input(ctx
, 0x3);
3259 nir_foreach_variable (var
, &ctx
->s
->inputs
) {
3260 setup_input(ctx
, var
);
3263 /* Defer add_sysval_input() stuff until after setup_inputs(),
3264 * because sysvals need to be appended after varyings:
3266 if (ctx
->ij_pixel
) {
3267 add_sysval_input_compmask(ctx
, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
,
3268 0x3, ctx
->ij_pixel
);
3272 /* Tesselation shaders always need primitive ID for indexing the
3273 * BO. Geometry shaders don't always need it but when they do it has be
3274 * delivered and unclobbered in the VS. To make things easy, we always
3275 * make room for it in VS/DS.
3277 bool has_tess
= ctx
->so
->key
.tessellation
!= IR3_TESS_NONE
;
3278 bool has_gs
= ctx
->so
->key
.has_gs
;
3279 switch (ctx
->so
->type
) {
3280 case MESA_SHADER_VERTEX
:
3282 ctx
->tcs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_TCS_HEADER_IR3
, 0x1);
3283 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3284 } else if (has_gs
) {
3285 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3286 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3289 case MESA_SHADER_TESS_CTRL
:
3290 ctx
->tcs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_TCS_HEADER_IR3
, 0x1);
3291 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3293 case MESA_SHADER_TESS_EVAL
:
3295 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3296 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3298 case MESA_SHADER_GEOMETRY
:
3299 ctx
->gs_header
= create_sysval_input(ctx
, SYSTEM_VALUE_GS_HEADER_IR3
, 0x1);
3300 ctx
->primitive_id
= create_sysval_input(ctx
, SYSTEM_VALUE_PRIMITIVE_ID
, 0x1);
3306 /* Setup outputs: */
3307 nir_foreach_variable (var
, &ctx
->s
->outputs
) {
3308 setup_output(ctx
, var
);
3311 /* Find # of samplers: */
3312 nir_foreach_variable (var
, &ctx
->s
->uniforms
) {
3313 ctx
->so
->num_samp
+= glsl_type_get_sampler_count(var
->type
);
3314 /* just assume that we'll be reading from images.. if it
3315 * is write-only we don't have to count it, but not sure
3316 * if there is a good way to know?
3318 ctx
->so
->num_samp
+= glsl_type_get_image_count(var
->type
);
3321 /* NOTE: need to do something more clever when we support >1 fxn */
3322 nir_foreach_register (reg
, &fxn
->registers
) {
3323 ir3_declare_array(ctx
, reg
);
3325 /* And emit the body: */
3327 emit_function(ctx
, fxn
);
3330 /* Fixup tex sampler state for astc/srgb workaround instructions. We
3331 * need to assign the tex state indexes for these after we know the
3335 fixup_astc_srgb(struct ir3_context
*ctx
)
3337 struct ir3_shader_variant
*so
= ctx
->so
;
3338 /* indexed by original tex idx, value is newly assigned alpha sampler
3339 * state tex idx. Zero is invalid since there is at least one sampler
3342 unsigned alt_tex_state
[16] = {0};
3343 unsigned tex_idx
= ctx
->max_texture_index
+ 1;
3346 so
->astc_srgb
.base
= tex_idx
;
3348 for (unsigned i
= 0; i
< ctx
->ir
->astc_srgb_count
; i
++) {
3349 struct ir3_instruction
*sam
= ctx
->ir
->astc_srgb
[i
];
3351 compile_assert(ctx
, sam
->cat5
.tex
< ARRAY_SIZE(alt_tex_state
));
3353 if (alt_tex_state
[sam
->cat5
.tex
] == 0) {
3354 /* assign new alternate/alpha tex state slot: */
3355 alt_tex_state
[sam
->cat5
.tex
] = tex_idx
++;
3356 so
->astc_srgb
.orig_idx
[idx
++] = sam
->cat5
.tex
;
3357 so
->astc_srgb
.count
++;
3360 sam
->cat5
.tex
= alt_tex_state
[sam
->cat5
.tex
];
3365 fixup_binning_pass(struct ir3_context
*ctx
)
3367 struct ir3_shader_variant
*so
= ctx
->so
;
3368 struct ir3
*ir
= ctx
->ir
;
3371 /* first pass, remove unused outputs from the IR level outputs: */
3372 for (i
= 0, j
= 0; i
< ir
->outputs_count
; i
++) {
3373 struct ir3_instruction
*out
= ir
->outputs
[i
];
3374 assert(out
->opc
== OPC_META_COLLECT
);
3375 unsigned outidx
= out
->collect
.outidx
;
3376 unsigned slot
= so
->outputs
[outidx
].slot
;
3378 /* throw away everything but first position/psize */
3379 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
3380 ir
->outputs
[j
] = ir
->outputs
[i
];
3384 ir
->outputs_count
= j
;
3386 /* second pass, cleanup the unused slots in ir3_shader_variant::outputs
3389 for (i
= 0, j
= 0; i
< so
->outputs_count
; i
++) {
3390 unsigned slot
= so
->outputs
[i
].slot
;
3392 /* throw away everything but first position/psize */
3393 if ((slot
== VARYING_SLOT_POS
) || (slot
== VARYING_SLOT_PSIZ
)) {
3394 so
->outputs
[j
] = so
->outputs
[i
];
3396 /* fixup outidx to point to new output table entry: */
3397 struct ir3_instruction
*out
;
3398 foreach_output (out
, ir
) {
3399 if (out
->collect
.outidx
== i
) {
3400 out
->collect
.outidx
= j
;
3408 so
->outputs_count
= j
;
3412 collect_tex_prefetches(struct ir3_context
*ctx
, struct ir3
*ir
)
3416 /* Collect sampling instructions eligible for pre-dispatch. */
3417 foreach_block (block
, &ir
->block_list
) {
3418 foreach_instr_safe (instr
, &block
->instr_list
) {
3419 if (instr
->opc
== OPC_META_TEX_PREFETCH
) {
3420 assert(idx
< ARRAY_SIZE(ctx
->so
->sampler_prefetch
));
3421 struct ir3_sampler_prefetch
*fetch
=
3422 &ctx
->so
->sampler_prefetch
[idx
];
3425 if (instr
->flags
& IR3_INSTR_B
) {
3426 fetch
->cmd
= IR3_SAMPLER_BINDLESS_PREFETCH_CMD
;
3427 /* In bindless mode, the index is actually the base */
3428 fetch
->tex_id
= instr
->prefetch
.tex_base
;
3429 fetch
->samp_id
= instr
->prefetch
.samp_base
;
3430 fetch
->tex_bindless_id
= instr
->prefetch
.tex
;
3431 fetch
->samp_bindless_id
= instr
->prefetch
.samp
;
3433 fetch
->cmd
= IR3_SAMPLER_PREFETCH_CMD
;
3434 fetch
->tex_id
= instr
->prefetch
.tex
;
3435 fetch
->samp_id
= instr
->prefetch
.samp
;
3437 fetch
->wrmask
= instr
->regs
[0]->wrmask
;
3438 fetch
->dst
= instr
->regs
[0]->num
;
3439 fetch
->src
= instr
->prefetch
.input_offset
;
3442 MAX2(ctx
->so
->total_in
, instr
->prefetch
.input_offset
+ 2);
3444 /* Disable half precision until supported. */
3445 fetch
->half_precision
= !!(instr
->regs
[0]->flags
& IR3_REG_HALF
);
3447 /* Remove the prefetch placeholder instruction: */
3448 list_delinit(&instr
->node
);
3455 ir3_compile_shader_nir(struct ir3_compiler
*compiler
,
3456 struct ir3_shader_variant
*so
)
3458 struct ir3_context
*ctx
;
3460 int ret
= 0, max_bary
;
3464 ctx
= ir3_context_init(compiler
, so
);
3466 DBG("INIT failed!");
3471 emit_instructions(ctx
);
3474 DBG("EMIT failed!");
3479 ir
= so
->ir
= ctx
->ir
;
3481 assert((ctx
->noutputs
% 4) == 0);
3483 /* Setup IR level outputs, which are "collects" that gather
3484 * the scalar components of outputs.
3486 for (unsigned i
= 0; i
< ctx
->noutputs
; i
+= 4) {
3488 /* figure out the # of components written:
3490 * TODO do we need to handle holes, ie. if .x and .z
3491 * components written, but .y component not written?
3493 for (unsigned j
= 0; j
< 4; j
++) {
3494 if (!ctx
->outputs
[i
+ j
])
3499 /* Note that in some stages, like TCS, store_output is
3500 * lowered to memory writes, so no components of the
3501 * are "written" from the PoV of traditional store-
3502 * output instructions:
3507 struct ir3_instruction
*out
=
3508 ir3_create_collect(ctx
, &ctx
->outputs
[i
], ncomp
);
3511 assert(outidx
< so
->outputs_count
);
3513 /* stash index into so->outputs[] so we can map the
3514 * output back to slot/etc later:
3516 out
->collect
.outidx
= outidx
;
3518 array_insert(ir
, ir
->outputs
, out
);
3521 /* Set up the gs header as an output for the vertex shader so it won't
3522 * clobber it for the tess ctrl shader.
3524 * TODO this could probably be done more cleanly in a nir pass.
3526 if (ctx
->so
->type
== MESA_SHADER_VERTEX
||
3527 (ctx
->so
->key
.has_gs
&& ctx
->so
->type
== MESA_SHADER_TESS_EVAL
)) {
3528 if (ctx
->primitive_id
) {
3529 unsigned n
= so
->outputs_count
++;
3530 so
->outputs
[n
].slot
= VARYING_SLOT_PRIMITIVE_ID
;
3532 struct ir3_instruction
*out
=
3533 ir3_create_collect(ctx
, &ctx
->primitive_id
, 1);
3534 out
->collect
.outidx
= n
;
3535 array_insert(ir
, ir
->outputs
, out
);
3538 if (ctx
->gs_header
) {
3539 unsigned n
= so
->outputs_count
++;
3540 so
->outputs
[n
].slot
= VARYING_SLOT_GS_HEADER_IR3
;
3541 struct ir3_instruction
*out
=
3542 ir3_create_collect(ctx
, &ctx
->gs_header
, 1);
3543 out
->collect
.outidx
= n
;
3544 array_insert(ir
, ir
->outputs
, out
);
3547 if (ctx
->tcs_header
) {
3548 unsigned n
= so
->outputs_count
++;
3549 so
->outputs
[n
].slot
= VARYING_SLOT_TCS_HEADER_IR3
;
3550 struct ir3_instruction
*out
=
3551 ir3_create_collect(ctx
, &ctx
->tcs_header
, 1);
3552 out
->collect
.outidx
= n
;
3553 array_insert(ir
, ir
->outputs
, out
);
3557 /* at this point, for binning pass, throw away unneeded outputs: */
3558 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
< 600))
3559 fixup_binning_pass(ctx
);
3561 ir3_debug_print(ir
, "BEFORE CF");
3565 ir3_debug_print(ir
, "BEFORE CP");
3569 /* at this point, for binning pass, throw away unneeded outputs:
3570 * Note that for a6xx and later, we do this after ir3_cp to ensure
3571 * that the uniform/constant layout for BS and VS matches, so that
3572 * we can re-use same VS_CONST state group.
3574 if (so
->binning_pass
&& (ctx
->compiler
->gpu_id
>= 600))
3575 fixup_binning_pass(ctx
);
3577 /* for a6xx+, binning and draw pass VS use same VBO state, so we
3578 * need to make sure not to remove any inputs that are used by
3579 * the nonbinning VS.
3581 if (ctx
->compiler
->gpu_id
>= 600 && so
->binning_pass
&&
3582 so
->type
== MESA_SHADER_VERTEX
) {
3583 for (int i
= 0; i
< ctx
->ninputs
; i
++) {
3584 struct ir3_instruction
*in
= ctx
->inputs
[i
];
3592 debug_assert(n
< so
->nonbinning
->inputs_count
);
3594 if (so
->nonbinning
->inputs
[n
].sysval
)
3597 /* be sure to keep inputs, even if only used in VS */
3598 if (so
->nonbinning
->inputs
[n
].compmask
& (1 << c
))
3599 array_insert(in
->block
, in
->block
->keeps
, in
);
3603 ir3_debug_print(ir
, "BEFORE GROUPING");
3605 ir3_sched_add_deps(ir
);
3607 /* Group left/right neighbors, inserting mov's where needed to
3612 ir3_debug_print(ir
, "AFTER GROUPING");
3616 ir3_debug_print(ir
, "AFTER DCE");
3618 /* do Sethi–Ullman numbering before scheduling: */
3621 ret
= ir3_sched(ir
);
3623 DBG("SCHED failed!");
3627 ir3_debug_print(ir
, "AFTER SCHED");
3629 /* Pre-assign VS inputs on a6xx+ binning pass shader, to align
3630 * with draw pass VS, so binning and draw pass can both use the
3633 * Note that VS inputs are expected to be full precision.
3635 bool pre_assign_inputs
= (ir
->compiler
->gpu_id
>= 600) &&
3636 (ir
->type
== MESA_SHADER_VERTEX
) &&
3639 if (pre_assign_inputs
) {
3640 for (unsigned i
= 0; i
< ctx
->ninputs
; i
++) {
3641 struct ir3_instruction
*instr
= ctx
->inputs
[i
];
3648 unsigned regid
= so
->nonbinning
->inputs
[n
].regid
+ c
;
3650 instr
->regs
[0]->num
= regid
;
3653 ret
= ir3_ra(so
, ctx
->inputs
, ctx
->ninputs
);
3654 } else if (ctx
->tcs_header
) {
3655 /* We need to have these values in the same registers between VS and TCS
3656 * since the VS chains to TCS and doesn't get the sysvals redelivered.
3659 ctx
->tcs_header
->regs
[0]->num
= regid(0, 0);
3660 ctx
->primitive_id
->regs
[0]->num
= regid(0, 1);
3661 struct ir3_instruction
*precolor
[] = { ctx
->tcs_header
, ctx
->primitive_id
};
3662 ret
= ir3_ra(so
, precolor
, ARRAY_SIZE(precolor
));
3663 } else if (ctx
->gs_header
) {
3664 /* We need to have these values in the same registers between producer
3665 * (VS or DS) and GS since the producer chains to GS and doesn't get
3666 * the sysvals redelivered.
3669 ctx
->gs_header
->regs
[0]->num
= regid(0, 0);
3670 ctx
->primitive_id
->regs
[0]->num
= regid(0, 1);
3671 struct ir3_instruction
*precolor
[] = { ctx
->gs_header
, ctx
->primitive_id
};
3672 ret
= ir3_ra(so
, precolor
, ARRAY_SIZE(precolor
));
3673 } else if (so
->num_sampler_prefetch
) {
3674 assert(so
->type
== MESA_SHADER_FRAGMENT
);
3675 struct ir3_instruction
*instr
, *precolor
[2];
3678 foreach_input (instr
, ir
) {
3679 if (instr
->input
.sysval
!= SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
)
3682 assert(idx
< ARRAY_SIZE(precolor
));
3684 precolor
[idx
] = instr
;
3685 instr
->regs
[0]->num
= idx
;
3689 ret
= ir3_ra(so
, precolor
, idx
);
3691 ret
= ir3_ra(so
, NULL
, 0);
3700 ir3_debug_print(ir
, "AFTER POSTSCHED");
3702 if (compiler
->gpu_id
>= 600) {
3703 if (ir3_a6xx_fixup_atomic_dests(ir
, so
)) {
3704 ir3_debug_print(ir
, "AFTER ATOMIC FIXUP");
3708 if (so
->type
== MESA_SHADER_FRAGMENT
)
3712 * Fixup inputs/outputs to point to the actual registers assigned:
3714 * 1) initialize to r63.x (invalid/unused)
3715 * 2) iterate IR level inputs/outputs and update the variants
3716 * inputs/outputs table based on the assigned registers for
3717 * the remaining inputs/outputs.
3720 for (unsigned i
= 0; i
< so
->inputs_count
; i
++)
3721 so
->inputs
[i
].regid
= INVALID_REG
;
3722 for (unsigned i
= 0; i
< so
->outputs_count
; i
++)
3723 so
->outputs
[i
].regid
= INVALID_REG
;
3725 struct ir3_instruction
*out
;
3726 foreach_output (out
, ir
) {
3727 assert(out
->opc
== OPC_META_COLLECT
);
3728 unsigned outidx
= out
->collect
.outidx
;
3730 so
->outputs
[outidx
].regid
= out
->regs
[0]->num
;
3731 so
->outputs
[outidx
].half
= !!(out
->regs
[0]->flags
& IR3_REG_HALF
);
3734 struct ir3_instruction
*in
;
3735 foreach_input (in
, ir
) {
3736 assert(in
->opc
== OPC_META_INPUT
);
3737 unsigned inidx
= in
->input
.inidx
;
3739 if (pre_assign_inputs
&& !so
->inputs
[inidx
].sysval
) {
3740 if (VALIDREG(so
->nonbinning
->inputs
[inidx
].regid
)) {
3741 compile_assert(ctx
, in
->regs
[0]->num
==
3742 so
->nonbinning
->inputs
[inidx
].regid
);
3743 compile_assert(ctx
, !!(in
->regs
[0]->flags
& IR3_REG_HALF
) ==
3744 so
->nonbinning
->inputs
[inidx
].half
);
3746 so
->inputs
[inidx
].regid
= so
->nonbinning
->inputs
[inidx
].regid
;
3747 so
->inputs
[inidx
].half
= so
->nonbinning
->inputs
[inidx
].half
;
3749 so
->inputs
[inidx
].regid
= in
->regs
[0]->num
;
3750 so
->inputs
[inidx
].half
= !!(in
->regs
[0]->flags
& IR3_REG_HALF
);
3755 fixup_astc_srgb(ctx
);
3757 /* We need to do legalize after (for frag shader's) the "bary.f"
3758 * offsets (inloc) have been assigned.
3760 ir3_legalize(ir
, so
, &max_bary
);
3762 ir3_debug_print(ir
, "AFTER LEGALIZE");
3764 /* Set (ss)(sy) on first TCS and GEOMETRY instructions, since we don't
3765 * know what we might have to wait on when coming in from VS chsh.
3767 if (so
->type
== MESA_SHADER_TESS_CTRL
||
3768 so
->type
== MESA_SHADER_GEOMETRY
) {
3769 foreach_block (block
, &ir
->block_list
) {
3770 foreach_instr (instr
, &block
->instr_list
) {
3771 instr
->flags
|= IR3_INSTR_SS
| IR3_INSTR_SY
;
3777 so
->branchstack
= ctx
->max_stack
;
3779 /* Note that actual_in counts inputs that are not bary.f'd for FS: */
3780 if (so
->type
== MESA_SHADER_FRAGMENT
)
3781 so
->total_in
= max_bary
+ 1;
3783 so
->max_sun
= ir
->max_sun
;
3785 /* Collect sampling instructions eligible for pre-dispatch. */
3786 collect_tex_prefetches(ctx
, ir
);
3788 if (so
->type
== MESA_SHADER_FRAGMENT
&&
3789 ctx
->s
->info
.fs
.needs_helper_invocations
)
3790 so
->need_pixlod
= true;
3795 ir3_destroy(so
->ir
);
3798 ir3_context_free(ctx
);