1 /*
2 * Copyright © 2018 Intel Corporation
3 * Copyright © 2019 Vasily Khoruzhick <anarsoul@gmail.com>
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25 #include "nir.h"
26 #include "nir_builder.h"
27
28 static bool
assert_ssa_def_is_not_int(nir_ssa_def * def,void * arg)29 assert_ssa_def_is_not_int(nir_ssa_def *def, void *arg)
30 {
31 ASSERTED BITSET_WORD *int_types = arg;
32 assert(!BITSET_TEST(int_types, def->index));
33 return true;
34 }
35
36 static bool
lower_alu_instr(nir_builder * b,nir_alu_instr * alu)37 lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
38 {
39 const nir_op_info *info = &nir_op_infos[alu->op];
40
41 bool is_bool_only = alu->dest.dest.ssa.bit_size == 1;
42 for (unsigned i = 0; i < info->num_inputs; i++) {
43 if (alu->src[i].src.ssa->bit_size != 1)
44 is_bool_only = false;
45 }
46
47 if (is_bool_only) {
48 /* avoid lowering integers ops are used for booleans (ieq,ine,etc) */
49 return false;
50 }
51
52 b->cursor = nir_before_instr(&alu->instr);
53
54 /* Replacement SSA value */
55 nir_ssa_def *rep = NULL;
56 switch (alu->op) {
57 case nir_op_mov:
58 case nir_op_vec2:
59 case nir_op_vec3:
60 case nir_op_vec4:
61 case nir_op_bcsel:
62 /* These we expect to have integers but the opcode doesn't change */
63 break;
64
65 case nir_op_b2i32: alu->op = nir_op_b2f32; break;
66 case nir_op_i2f32: alu->op = nir_op_mov; break;
67 case nir_op_u2f32: alu->op = nir_op_mov; break;
68
69 case nir_op_f2i32: {
70 alu->op = nir_op_ftrunc;
71
72 /* If the source was already integer, then we did't need to truncate and
73 * can switch it to a mov that can be copy-propagated away.
74 */
75 nir_alu_instr *src_alu = nir_src_as_alu_instr(alu->src[0].src);
76 if (src_alu) {
77 switch (src_alu->op) {
78 case nir_op_fround_even:
79 case nir_op_fceil:
80 case nir_op_ftrunc:
81 case nir_op_ffloor:
82 alu->op = nir_op_mov;
83 break;
84 default:
85 break;
86 }
87 }
88 break;
89 }
90
91 case nir_op_f2u32: alu->op = nir_op_ffloor; break;
92 case nir_op_i2b1: alu->op = nir_op_f2b1; break;
93
94 case nir_op_ilt: alu->op = nir_op_flt; break;
95 case nir_op_ige: alu->op = nir_op_fge; break;
96 case nir_op_ieq: alu->op = nir_op_feq; break;
97 case nir_op_ine: alu->op = nir_op_fneu; break;
98 case nir_op_ult: alu->op = nir_op_flt; break;
99 case nir_op_uge: alu->op = nir_op_fge; break;
100
101 case nir_op_iadd: alu->op = nir_op_fadd; break;
102 case nir_op_isub: alu->op = nir_op_fsub; break;
103 case nir_op_imul: alu->op = nir_op_fmul; break;
104
105 case nir_op_idiv: {
106 nir_ssa_def *x = nir_ssa_for_alu_src(b, alu, 0);
107 nir_ssa_def *y = nir_ssa_for_alu_src(b, alu, 1);
108
109 /* Hand-lower fdiv, since lower_int_to_float is after nir_opt_algebraic. */
110 if (b->shader->options->lower_fdiv) {
111 rep = nir_ftrunc(b, nir_fmul(b, x, nir_frcp(b, y)));
112 } else {
113 rep = nir_ftrunc(b, nir_fdiv(b, x, y));
114 }
115 break;
116 }
117
118 case nir_op_iabs: alu->op = nir_op_fabs; break;
119 case nir_op_ineg: alu->op = nir_op_fneg; break;
120 case nir_op_imax: alu->op = nir_op_fmax; break;
121 case nir_op_imin: alu->op = nir_op_fmin; break;
122 case nir_op_umax: alu->op = nir_op_fmax; break;
123 case nir_op_umin: alu->op = nir_op_fmin; break;
124
125 case nir_op_ball_iequal2: alu->op = nir_op_ball_fequal2; break;
126 case nir_op_ball_iequal3: alu->op = nir_op_ball_fequal3; break;
127 case nir_op_ball_iequal4: alu->op = nir_op_ball_fequal4; break;
128 case nir_op_bany_inequal2: alu->op = nir_op_bany_fnequal2; break;
129 case nir_op_bany_inequal3: alu->op = nir_op_bany_fnequal3; break;
130 case nir_op_bany_inequal4: alu->op = nir_op_bany_fnequal4; break;
131
132 default:
133 assert(nir_alu_type_get_base_type(info->output_type) != nir_type_int &&
134 nir_alu_type_get_base_type(info->output_type) != nir_type_uint);
135 for (unsigned i = 0; i < info->num_inputs; i++) {
136 assert(nir_alu_type_get_base_type(info->input_types[i]) != nir_type_int &&
137 nir_alu_type_get_base_type(info->input_types[i]) != nir_type_uint);
138 }
139 return false;
140 }
141
142 if (rep) {
143 /* We've emitted a replacement instruction */
144 nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, rep);
145 nir_instr_remove(&alu->instr);
146 }
147
148 return true;
149 }
150
151 static bool
nir_lower_int_to_float_impl(nir_function_impl * impl)152 nir_lower_int_to_float_impl(nir_function_impl *impl)
153 {
154 bool progress = false;
155 BITSET_WORD *float_types = NULL, *int_types = NULL;
156
157 nir_builder b;
158 nir_builder_init(&b, impl);
159
160 nir_index_ssa_defs(impl);
161 float_types = calloc(BITSET_WORDS(impl->ssa_alloc),
162 sizeof(BITSET_WORD));
163 int_types = calloc(BITSET_WORDS(impl->ssa_alloc),
164 sizeof(BITSET_WORD));
165 nir_gather_ssa_types(impl, float_types, int_types);
166
167 nir_foreach_block(block, impl) {
168 nir_foreach_instr_safe(instr, block) {
169 switch (instr->type) {
170 case nir_instr_type_alu:
171 progress |= lower_alu_instr(&b, nir_instr_as_alu(instr));
172 break;
173
174 case nir_instr_type_load_const: {
175 nir_load_const_instr *load = nir_instr_as_load_const(instr);
176 if (load->def.bit_size != 1 && BITSET_TEST(int_types, load->def.index)) {
177 for (unsigned i = 0; i < load->def.num_components; i++)
178 load->value[i].f32 = load->value[i].i32;
179 }
180 break;
181 }
182
183 case nir_instr_type_intrinsic:
184 case nir_instr_type_ssa_undef:
185 case nir_instr_type_phi:
186 case nir_instr_type_tex:
187 break;
188
189 default:
190 nir_foreach_ssa_def(instr, assert_ssa_def_is_not_int, (void *)int_types);
191 break;
192 }
193 }
194 }
195
196 if (progress) {
197 nir_metadata_preserve(impl, nir_metadata_block_index |
198 nir_metadata_dominance);
199 } else {
200 nir_metadata_preserve(impl, nir_metadata_all);
201 }
202
203 free(float_types);
204 free(int_types);
205
206 return progress;
207 }
208
209 bool
nir_lower_int_to_float(nir_shader * shader)210 nir_lower_int_to_float(nir_shader *shader)
211 {
212 bool progress = false;
213
214 nir_foreach_function(function, shader) {
215 if (function->impl && nir_lower_int_to_float_impl(function->impl))
216 progress = true;
217 }
218
219 return progress;
220 }
221