2 * Copyright © 2014-2015 Broadcom
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 #include "nir_builder.h"
27 /** @file nir_lower_alu_to_scalar.c
29 * Replaces nir_alu_instr operations with more than one channel used in the
30 * arguments with individual per-channel operations.
34 nir_alu_ssa_dest_init(nir_alu_instr
*alu
, unsigned num_components
,
37 nir_ssa_dest_init(&alu
->instr
, &alu
->dest
.dest
, num_components
,
39 alu
->dest
.write_mask
= (1 << num_components
) - 1;
43 lower_reduction(nir_alu_instr
*alu
, nir_op chan_op
, nir_op merge_op
,
46 unsigned num_components
= nir_op_infos
[alu
->op
].input_sizes
[0];
48 nir_ssa_def
*last
= NULL
;
49 for (unsigned i
= 0; i
< num_components
; i
++) {
50 nir_alu_instr
*chan
= nir_alu_instr_create(builder
->shader
, chan_op
);
51 nir_alu_ssa_dest_init(chan
, 1, alu
->dest
.dest
.ssa
.bit_size
);
52 nir_alu_src_copy(&chan
->src
[0], &alu
->src
[0], chan
);
53 chan
->src
[0].swizzle
[0] = chan
->src
[0].swizzle
[i
];
54 if (nir_op_infos
[chan_op
].num_inputs
> 1) {
55 assert(nir_op_infos
[chan_op
].num_inputs
== 2);
56 nir_alu_src_copy(&chan
->src
[1], &alu
->src
[1], chan
);
57 chan
->src
[1].swizzle
[0] = chan
->src
[1].swizzle
[i
];
59 chan
->exact
= alu
->exact
;
61 nir_builder_instr_insert(builder
, &chan
->instr
);
64 last
= &chan
->dest
.dest
.ssa
;
66 last
= nir_build_alu(builder
, merge_op
,
67 last
, &chan
->dest
.dest
.ssa
, NULL
, NULL
);
71 assert(alu
->dest
.write_mask
== 1);
72 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(last
));
73 nir_instr_remove(&alu
->instr
);
77 lower_alu_instr_scalar(nir_alu_instr
*alu
, nir_builder
*b
, BITSET_WORD
*lower_set
)
79 unsigned num_src
= nir_op_infos
[alu
->op
].num_inputs
;
82 assert(alu
->dest
.dest
.is_ssa
);
83 assert(alu
->dest
.write_mask
!= 0);
85 b
->cursor
= nir_before_instr(&alu
->instr
);
86 b
->exact
= alu
->exact
;
88 if (lower_set
&& !BITSET_TEST(lower_set
, alu
->op
))
91 #define LOWER_REDUCTION(name, chan, merge) \
95 lower_reduction(alu, chan, merge, b); \
102 case nir_op_cube_face_coord
:
103 case nir_op_cube_face_index
:
104 /* We don't need to scalarize these ops, they're the ones generated to
105 * group up outputs into a value that can be SSAed.
109 case nir_op_pack_half_2x16
:
110 if (!b
->shader
->options
->lower_pack_half_2x16
)
113 nir_ssa_def
*src_vec2
= nir_ssa_for_alu_src(b
, alu
, 0);
116 nir_pack_half_2x16_split(b
, nir_channel(b
, src_vec2
, 0),
117 nir_channel(b
, src_vec2
, 1));
119 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(val
));
120 nir_instr_remove(&alu
->instr
);
123 case nir_op_unpack_unorm_4x8
:
124 case nir_op_unpack_snorm_4x8
:
125 case nir_op_unpack_unorm_2x16
:
126 case nir_op_unpack_snorm_2x16
:
127 /* There is no scalar version of these ops, unless we were to break it
128 * down to bitshifts and math (which is definitely not intended).
132 case nir_op_unpack_half_2x16
: {
133 if (!b
->shader
->options
->lower_unpack_half_2x16
)
136 nir_ssa_def
*packed
= nir_ssa_for_alu_src(b
, alu
, 0);
138 nir_ssa_def
*comps
[2];
139 comps
[0] = nir_unpack_half_2x16_split_x(b
, packed
);
140 comps
[1] = nir_unpack_half_2x16_split_y(b
, packed
);
141 nir_ssa_def
*vec
= nir_vec(b
, comps
, 2);
143 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(vec
));
144 nir_instr_remove(&alu
->instr
);
148 case nir_op_pack_uvec2_to_uint
: {
149 assert(b
->shader
->options
->lower_pack_snorm_2x16
||
150 b
->shader
->options
->lower_pack_unorm_2x16
);
152 nir_ssa_def
*word
= nir_extract_u16(b
, nir_ssa_for_alu_src(b
, alu
, 0),
155 nir_ior(b
, nir_ishl(b
, nir_channel(b
, word
, 1), nir_imm_int(b
, 16)),
156 nir_channel(b
, word
, 0));
158 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(val
));
159 nir_instr_remove(&alu
->instr
);
163 case nir_op_pack_uvec4_to_uint
: {
164 assert(b
->shader
->options
->lower_pack_snorm_4x8
||
165 b
->shader
->options
->lower_pack_unorm_4x8
);
167 nir_ssa_def
*byte
= nir_extract_u8(b
, nir_ssa_for_alu_src(b
, alu
, 0),
170 nir_ior(b
, nir_ior(b
, nir_ishl(b
, nir_channel(b
, byte
, 3), nir_imm_int(b
, 24)),
171 nir_ishl(b
, nir_channel(b
, byte
, 2), nir_imm_int(b
, 16))),
172 nir_ior(b
, nir_ishl(b
, nir_channel(b
, byte
, 1), nir_imm_int(b
, 8)),
173 nir_channel(b
, byte
, 0)));
175 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(val
));
176 nir_instr_remove(&alu
->instr
);
181 nir_ssa_def
*src0_vec
= nir_ssa_for_alu_src(b
, alu
, 0);
182 nir_ssa_def
*src1_vec
= nir_ssa_for_alu_src(b
, alu
, 1);
185 for (unsigned i
= 0; i
< 3; i
++) {
186 sum
[i
] = nir_fmul(b
, nir_channel(b
, src0_vec
, i
),
187 nir_channel(b
, src1_vec
, i
));
189 sum
[3] = nir_channel(b
, src1_vec
, 3);
191 nir_ssa_def
*val
= nir_fadd(b
, nir_fadd(b
, sum
[0], sum
[1]),
192 nir_fadd(b
, sum
[2], sum
[3]));
194 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(val
));
195 nir_instr_remove(&alu
->instr
);
199 case nir_op_unpack_64_2x32
:
200 case nir_op_unpack_32_2x16
:
203 LOWER_REDUCTION(nir_op_fdot
, nir_op_fmul
, nir_op_fadd
);
204 LOWER_REDUCTION(nir_op_ball_fequal
, nir_op_feq
, nir_op_iand
);
205 LOWER_REDUCTION(nir_op_ball_iequal
, nir_op_ieq
, nir_op_iand
);
206 LOWER_REDUCTION(nir_op_bany_fnequal
, nir_op_fne
, nir_op_ior
);
207 LOWER_REDUCTION(nir_op_bany_inequal
, nir_op_ine
, nir_op_ior
);
208 LOWER_REDUCTION(nir_op_b32all_fequal
, nir_op_feq32
, nir_op_iand
);
209 LOWER_REDUCTION(nir_op_b32all_iequal
, nir_op_ieq32
, nir_op_iand
);
210 LOWER_REDUCTION(nir_op_b32any_fnequal
, nir_op_fne32
, nir_op_ior
);
211 LOWER_REDUCTION(nir_op_b32any_inequal
, nir_op_ine32
, nir_op_ior
);
212 LOWER_REDUCTION(nir_op_fall_equal
, nir_op_seq
, nir_op_fmin
);
213 LOWER_REDUCTION(nir_op_fany_nequal
, nir_op_sne
, nir_op_fmax
);
219 if (alu
->dest
.dest
.ssa
.num_components
== 1)
222 unsigned num_components
= alu
->dest
.dest
.ssa
.num_components
;
223 nir_ssa_def
*comps
[NIR_MAX_VEC_COMPONENTS
] = { NULL
};
225 for (chan
= 0; chan
< NIR_MAX_VEC_COMPONENTS
; chan
++) {
226 if (!(alu
->dest
.write_mask
& (1 << chan
)))
229 nir_alu_instr
*lower
= nir_alu_instr_create(b
->shader
, alu
->op
);
230 for (i
= 0; i
< num_src
; i
++) {
231 /* We only handle same-size-as-dest (input_sizes[] == 0) or scalar
232 * args (input_sizes[] == 1).
234 assert(nir_op_infos
[alu
->op
].input_sizes
[i
] < 2);
235 unsigned src_chan
= (nir_op_infos
[alu
->op
].input_sizes
[i
] == 1 ?
238 nir_alu_src_copy(&lower
->src
[i
], &alu
->src
[i
], lower
);
239 for (int j
= 0; j
< NIR_MAX_VEC_COMPONENTS
; j
++)
240 lower
->src
[i
].swizzle
[j
] = alu
->src
[i
].swizzle
[src_chan
];
243 nir_alu_ssa_dest_init(lower
, 1, alu
->dest
.dest
.ssa
.bit_size
);
244 lower
->dest
.saturate
= alu
->dest
.saturate
;
245 comps
[chan
] = &lower
->dest
.dest
.ssa
;
246 lower
->exact
= alu
->exact
;
248 nir_builder_instr_insert(b
, &lower
->instr
);
251 nir_ssa_def
*vec
= nir_vec(b
, comps
, num_components
);
253 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(vec
));
255 nir_instr_remove(&alu
->instr
);
260 nir_lower_alu_to_scalar_impl(nir_function_impl
*impl
, BITSET_WORD
*lower_set
)
263 nir_builder_init(&builder
, impl
);
264 bool progress
= false;
266 nir_foreach_block(block
, impl
) {
267 nir_foreach_instr_safe(instr
, block
) {
268 if (instr
->type
== nir_instr_type_alu
) {
269 progress
= lower_alu_instr_scalar(nir_instr_as_alu(instr
),
271 lower_set
) || progress
;
276 nir_metadata_preserve(impl
, nir_metadata_block_index
|
277 nir_metadata_dominance
);
283 nir_lower_alu_to_scalar(nir_shader
*shader
, BITSET_WORD
*lower_set
)
285 bool progress
= false;
287 nir_foreach_function(function
, shader
) {
289 progress
= nir_lower_alu_to_scalar_impl(function
->impl
,
290 lower_set
) || progress
;