1 /*
2 * Copyright © 2014 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28 #include "brw_nir.h"
29 #include "compiler/nir/nir_builder.h"
30
31 /*
32 * Implements a small peephole optimization that looks for a multiply that
33 * is only ever used in an add and replaces both with an fma.
34 */
35
36 static inline bool
are_all_uses_fadd(nir_ssa_def * def)37 are_all_uses_fadd(nir_ssa_def *def)
38 {
39 if (!list_is_empty(&def->if_uses))
40 return false;
41
42 nir_foreach_use(use_src, def) {
43 nir_instr *use_instr = use_src->parent_instr;
44
45 if (use_instr->type != nir_instr_type_alu)
46 return false;
47
48 nir_alu_instr *use_alu = nir_instr_as_alu(use_instr);
49 switch (use_alu->op) {
50 case nir_op_fadd:
51 break; /* This one's ok */
52
53 case nir_op_mov:
54 case nir_op_fneg:
55 case nir_op_fabs:
56 assert(use_alu->dest.dest.is_ssa);
57 if (!are_all_uses_fadd(&use_alu->dest.dest.ssa))
58 return false;
59 break;
60
61 default:
62 return false;
63 }
64 }
65
66 return true;
67 }
68
69 static nir_alu_instr *
get_mul_for_src(nir_alu_src * src,unsigned num_components,uint8_t swizzle[4],bool * negate,bool * abs)70 get_mul_for_src(nir_alu_src *src, unsigned num_components,
71 uint8_t swizzle[4], bool *negate, bool *abs)
72 {
73 uint8_t swizzle_tmp[4];
74 assert(src->src.is_ssa && !src->abs && !src->negate);
75
76 nir_instr *instr = src->src.ssa->parent_instr;
77 if (instr->type != nir_instr_type_alu)
78 return NULL;
79
80 nir_alu_instr *alu = nir_instr_as_alu(instr);
81
82 /* We want to bail if any of the other ALU operations involved is labeled
83 * exact. One reason for this is that, while the value that is changing is
84 * actually the result of the add and not the multiply, the intention of
85 * the user when they specify an exact multiply is that they want *that*
86 * value and what they don't care about is the add. Another reason is that
87 * SPIR-V explicitly requires this behaviour.
88 */
89 if (alu->exact)
90 return NULL;
91
92 switch (alu->op) {
93 case nir_op_mov:
94 alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components,
95 swizzle, negate, abs);
96 break;
97
98 case nir_op_fneg:
99 alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components,
100 swizzle, negate, abs);
101 *negate = !*negate;
102 break;
103
104 case nir_op_fabs:
105 alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components,
106 swizzle, negate, abs);
107 *negate = false;
108 *abs = true;
109 break;
110
111 case nir_op_fmul:
112 /* Only absorb a fmul into a ffma if the fmul is only used in fadd
113 * operations. This prevents us from being too aggressive with our
114 * fusing which can actually lead to more instructions.
115 */
116 if (!are_all_uses_fadd(&alu->dest.dest.ssa))
117 return NULL;
118 break;
119
120 default:
121 return NULL;
122 }
123
124 if (!alu)
125 return NULL;
126
127 /* Copy swizzle data before overwriting it to avoid setting a wrong swizzle.
128 *
129 * Example:
130 * Former swizzle[] = xyzw
131 * src->swizzle[] = zyxx
132 *
133 * Expected output swizzle = zyxx
134 * If we reuse swizzle in the loop, then output swizzle would be zyzz.
135 */
136 memcpy(swizzle_tmp, swizzle, 4*sizeof(uint8_t));
137 for (int i = 0; i < num_components; i++)
138 swizzle[i] = swizzle_tmp[src->swizzle[i]];
139
140 return alu;
141 }
142
143 /**
144 * Given a list of (at least two) nir_alu_src's, tells if any of them is a
145 * constant value and is used only once.
146 */
147 static bool
any_alu_src_is_a_constant(nir_alu_src srcs[])148 any_alu_src_is_a_constant(nir_alu_src srcs[])
149 {
150 for (unsigned i = 0; i < 2; i++) {
151 if (srcs[i].src.ssa->parent_instr->type == nir_instr_type_load_const) {
152 nir_load_const_instr *load_const =
153 nir_instr_as_load_const (srcs[i].src.ssa->parent_instr);
154
155 if (list_is_singular(&load_const->def.uses) &&
156 list_is_empty(&load_const->def.if_uses)) {
157 return true;
158 }
159 }
160 }
161
162 return false;
163 }
164
165 static bool
brw_nir_opt_peephole_ffma_instr(nir_builder * b,nir_instr * instr,UNUSED void * cb_data)166 brw_nir_opt_peephole_ffma_instr(nir_builder *b,
167 nir_instr *instr,
168 UNUSED void *cb_data)
169 {
170 if (instr->type != nir_instr_type_alu)
171 return false;
172
173 nir_alu_instr *add = nir_instr_as_alu(instr);
174 if (add->op != nir_op_fadd)
175 return false;
176
177 assert(add->dest.dest.is_ssa);
178 if (add->exact)
179 return false;
180
181 assert(add->src[0].src.is_ssa && add->src[1].src.is_ssa);
182
183 /* This, is the case a + a. We would rather handle this with an
184 * algebraic reduction than fuse it. Also, we want to only fuse
185 * things where the multiply is used only once and, in this case,
186 * it would be used twice by the same instruction.
187 */
188 if (add->src[0].src.ssa == add->src[1].src.ssa)
189 return false;
190
191 nir_alu_instr *mul;
192 uint8_t add_mul_src, swizzle[4];
193 bool negate, abs;
194 for (add_mul_src = 0; add_mul_src < 2; add_mul_src++) {
195 for (unsigned i = 0; i < 4; i++)
196 swizzle[i] = i;
197
198 negate = false;
199 abs = false;
200
201 mul = get_mul_for_src(&add->src[add_mul_src],
202 add->dest.dest.ssa.num_components,
203 swizzle, &negate, &abs);
204
205 if (mul != NULL)
206 break;
207 }
208
209 if (mul == NULL)
210 return false;
211
212 unsigned bit_size = add->dest.dest.ssa.bit_size;
213
214 nir_ssa_def *mul_src[2];
215 mul_src[0] = mul->src[0].src.ssa;
216 mul_src[1] = mul->src[1].src.ssa;
217
218 /* If any of the operands of the fmul and any of the fadd is a constant,
219 * we bypass because it will be more efficient as the constants will be
220 * propagated as operands, potentially saving two load_const instructions.
221 */
222 if (any_alu_src_is_a_constant(mul->src) &&
223 any_alu_src_is_a_constant(add->src)) {
224 return false;
225 }
226
227 b->cursor = nir_before_instr(&add->instr);
228
229 if (abs) {
230 for (unsigned i = 0; i < 2; i++)
231 mul_src[i] = nir_fabs(b, mul_src[i]);
232 }
233
234 if (negate)
235 mul_src[0] = nir_fneg(b, mul_src[0]);
236
237 nir_alu_instr *ffma = nir_alu_instr_create(b->shader, nir_op_ffma);
238 ffma->dest.saturate = add->dest.saturate;
239 ffma->dest.write_mask = add->dest.write_mask;
240
241 for (unsigned i = 0; i < 2; i++) {
242 ffma->src[i].src = nir_src_for_ssa(mul_src[i]);
243 for (unsigned j = 0; j < add->dest.dest.ssa.num_components; j++)
244 ffma->src[i].swizzle[j] = mul->src[i].swizzle[swizzle[j]];
245 }
246 nir_alu_src_copy(&ffma->src[2], &add->src[1 - add_mul_src]);
247
248 assert(add->dest.dest.is_ssa);
249
250 nir_ssa_dest_init(&ffma->instr, &ffma->dest.dest,
251 add->dest.dest.ssa.num_components,
252 bit_size, NULL);
253 nir_ssa_def_rewrite_uses(&add->dest.dest.ssa, &ffma->dest.dest.ssa);
254
255 nir_builder_instr_insert(b, &ffma->instr);
256 assert(list_is_empty(&add->dest.dest.ssa.uses));
257 nir_instr_remove(&add->instr);
258
259 return true;
260 }
261
262 bool
brw_nir_opt_peephole_ffma(nir_shader * shader)263 brw_nir_opt_peephole_ffma(nir_shader *shader)
264 {
265 return nir_shader_instructions_pass(shader, brw_nir_opt_peephole_ffma_instr,
266 nir_metadata_block_index |
267 nir_metadata_dominance,
268 NULL);
269 }
270