2 * Copyright © 2014 Intel Corporation
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
24 * Jason Ekstrand (jason@jlekstrand.net)
31 * Implements a small peephole optimization that looks for a multiply that
32 * is only ever used in an add and replaces both with an fma.
35 struct peephole_ffma_state
{
37 nir_function_impl
*impl
;
42 are_all_uses_fadd(nir_ssa_def
*def
)
44 if (!list_empty(&def
->if_uses
))
47 nir_foreach_use(def
, use_src
) {
48 nir_instr
*use_instr
= use_src
->parent_instr
;
50 if (use_instr
->type
!= nir_instr_type_alu
)
53 nir_alu_instr
*use_alu
= nir_instr_as_alu(use_instr
);
54 switch (use_alu
->op
) {
56 break; /* This one's ok */
62 assert(use_alu
->dest
.dest
.is_ssa
);
63 if (!are_all_uses_fadd(&use_alu
->dest
.dest
.ssa
))
75 static nir_alu_instr
*
76 get_mul_for_src(nir_alu_src
*src
, int num_components
,
77 uint8_t swizzle
[4], bool *negate
, bool *abs
)
79 assert(src
->src
.is_ssa
&& !src
->abs
&& !src
->negate
);
81 nir_instr
*instr
= src
->src
.ssa
->parent_instr
;
82 if (instr
->type
!= nir_instr_type_alu
)
85 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
89 alu
= get_mul_for_src(&alu
->src
[0], num_components
, swizzle
, negate
, abs
);
93 alu
= get_mul_for_src(&alu
->src
[0], num_components
, swizzle
, negate
, abs
);
98 alu
= get_mul_for_src(&alu
->src
[0], num_components
, swizzle
, negate
, abs
);
104 /* Only absorb a fmul into a ffma if the fmul is is only used in fadd
105 * operations. This prevents us from being too aggressive with our
106 * fusing which can actually lead to more instructions.
108 if (!are_all_uses_fadd(&alu
->dest
.dest
.ssa
))
119 for (unsigned i
= 0; i
< num_components
; i
++)
120 swizzle
[i
] = swizzle
[src
->swizzle
[i
]];
126 nir_opt_peephole_ffma_block(nir_block
*block
, void *void_state
)
128 struct peephole_ffma_state
*state
= void_state
;
130 nir_foreach_instr_safe(block
, instr
) {
131 if (instr
->type
!= nir_instr_type_alu
)
134 nir_alu_instr
*add
= nir_instr_as_alu(instr
);
135 if (add
->op
!= nir_op_fadd
)
138 /* TODO: Maybe bail if this expression is considered "precise"? */
140 assert(add
->src
[0].src
.is_ssa
&& add
->src
[1].src
.is_ssa
);
142 /* This, is the case a + a. We would rather handle this with an
143 * algebraic reduction than fuse it. Also, we want to only fuse
144 * things where the multiply is used only once and, in this case,
145 * it would be used twice by the same instruction.
147 if (add
->src
[0].src
.ssa
== add
->src
[1].src
.ssa
)
151 uint8_t add_mul_src
, swizzle
[4];
153 for (add_mul_src
= 0; add_mul_src
< 2; add_mul_src
++) {
154 for (unsigned i
= 0; i
< 4; i
++)
160 mul
= get_mul_for_src(&add
->src
[add_mul_src
],
161 add
->dest
.dest
.ssa
.num_components
,
162 swizzle
, &negate
, &abs
);
171 nir_ssa_def
*mul_src
[2];
172 mul_src
[0] = mul
->src
[0].src
.ssa
;
173 mul_src
[1] = mul
->src
[1].src
.ssa
;
176 for (unsigned i
= 0; i
< 2; i
++) {
177 nir_alu_instr
*abs
= nir_alu_instr_create(state
->mem_ctx
,
179 abs
->src
[0].src
= nir_src_for_ssa(mul_src
[i
]);
180 nir_ssa_dest_init(&abs
->instr
, &abs
->dest
.dest
,
181 mul_src
[i
]->num_components
, NULL
);
182 abs
->dest
.write_mask
= (1 << mul_src
[i
]->num_components
) - 1;
183 nir_instr_insert_before(&add
->instr
, &abs
->instr
);
184 mul_src
[i
] = &abs
->dest
.dest
.ssa
;
189 nir_alu_instr
*neg
= nir_alu_instr_create(state
->mem_ctx
,
191 neg
->src
[0].src
= nir_src_for_ssa(mul_src
[0]);
192 nir_ssa_dest_init(&neg
->instr
, &neg
->dest
.dest
,
193 mul_src
[0]->num_components
, NULL
);
194 neg
->dest
.write_mask
= (1 << mul_src
[0]->num_components
) - 1;
195 nir_instr_insert_before(&add
->instr
, &neg
->instr
);
196 mul_src
[0] = &neg
->dest
.dest
.ssa
;
199 nir_alu_instr
*ffma
= nir_alu_instr_create(state
->mem_ctx
, nir_op_ffma
);
200 ffma
->dest
.saturate
= add
->dest
.saturate
;
201 ffma
->dest
.write_mask
= add
->dest
.write_mask
;
203 for (unsigned i
= 0; i
< 2; i
++) {
204 ffma
->src
[i
].src
= nir_src_for_ssa(mul_src
[i
]);
205 for (unsigned j
= 0; j
< add
->dest
.dest
.ssa
.num_components
; j
++)
206 ffma
->src
[i
].swizzle
[j
] = mul
->src
[i
].swizzle
[swizzle
[j
]];
208 nir_alu_src_copy(&ffma
->src
[2], &add
->src
[1 - add_mul_src
],
211 assert(add
->dest
.dest
.is_ssa
);
213 nir_ssa_dest_init(&ffma
->instr
, &ffma
->dest
.dest
,
214 add
->dest
.dest
.ssa
.num_components
,
215 add
->dest
.dest
.ssa
.name
);
216 nir_ssa_def_rewrite_uses(&add
->dest
.dest
.ssa
,
217 nir_src_for_ssa(&ffma
->dest
.dest
.ssa
),
220 nir_instr_insert_before(&add
->instr
, &ffma
->instr
);
221 assert(list_empty(&add
->dest
.dest
.ssa
.uses
));
222 nir_instr_remove(&add
->instr
);
224 state
->progress
= true;
231 nir_opt_peephole_ffma_impl(nir_function_impl
*impl
)
233 struct peephole_ffma_state state
;
235 state
.mem_ctx
= ralloc_parent(impl
);
237 state
.progress
= false;
239 nir_foreach_block(impl
, nir_opt_peephole_ffma_block
, &state
);
242 nir_metadata_preserve(impl
, nir_metadata_block_index
|
243 nir_metadata_dominance
);
245 return state
.progress
;
249 nir_opt_peephole_ffma(nir_shader
*shader
)
251 bool progress
= false;
253 nir_foreach_overload(shader
, overload
) {
255 progress
|= nir_opt_peephole_ffma_impl(overload
->impl
);