1 /*
2 * Copyright (C) 2021 Collabora, Ltd.
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 #include "compiler.h"
25 #include "compiler/nir/nir_builder.h"
26
27 /* Divergent attribute access is undefined behaviour. To avoid divergence,
28 * lower to an if-chain like:
29 *
30 * value = 0;
31 * if (lane == 0)
32 * value = ld()
33 * else if (lane == 1)
34 * value = ld()
35 * ...
36 * else if (lane == MAX_LANE)
37 * value = ld()
38 */
39
40 static bool
bi_lower_divergent_indirects_impl(nir_builder * b,nir_instr * instr,void * data)41 bi_lower_divergent_indirects_impl(nir_builder *b, nir_instr *instr, void *data)
42 {
43 if (instr->type != nir_instr_type_intrinsic)
44 return false;
45
46 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
47 gl_shader_stage stage = b->shader->info.stage;
48 nir_src *offset;
49
50 /* Not all indirect access needs this workaround */
51 switch (intr->intrinsic) {
52 case nir_intrinsic_load_input:
53 case nir_intrinsic_load_interpolated_input:
54 /* Attributes and varyings */
55 offset = nir_get_io_offset_src(intr);
56 break;
57
58 case nir_intrinsic_store_output:
59 /* Varyings only */
60 if (stage == MESA_SHADER_FRAGMENT)
61 return false;
62
63 offset = nir_get_io_offset_src(intr);
64 break;
65
66 case nir_intrinsic_image_atomic_add:
67 case nir_intrinsic_image_atomic_imin:
68 case nir_intrinsic_image_atomic_umin:
69 case nir_intrinsic_image_atomic_imax:
70 case nir_intrinsic_image_atomic_umax:
71 case nir_intrinsic_image_atomic_and:
72 case nir_intrinsic_image_atomic_or:
73 case nir_intrinsic_image_atomic_xor:
74 case nir_intrinsic_image_load:
75 case nir_intrinsic_image_store:
76 /* Any image access */
77 offset = &intr->src[0];
78 break;
79 default:
80 return false;
81 }
82
83 if (!nir_src_is_divergent(*offset))
84 return false;
85
86 /* This indirect does need it */
87
88 b->cursor = nir_before_instr(instr);
89 nir_ssa_def *lane = nir_load_subgroup_invocation(b);
90 unsigned *lanes = data;
91
92 /* Write zero in a funny way to bypass lower_load_const_to_scalar */
93 bool has_dest = nir_intrinsic_infos[intr->intrinsic].has_dest;
94 unsigned size = has_dest ? nir_dest_bit_size(intr->dest) : 32;
95 nir_ssa_def *zero = has_dest ? nir_imm_zero(b, 1, size) : NULL;
96 nir_ssa_def *zeroes[4] = { zero, zero, zero, zero };
97 nir_ssa_def *res = has_dest ?
98 nir_vec(b, zeroes, nir_dest_num_components(intr->dest)) : NULL;
99
100 for (unsigned i = 0; i < (*lanes); ++i) {
101 nir_push_if(b, nir_ieq_imm(b, lane, i));
102
103 nir_instr *c = nir_instr_clone(b->shader, instr);
104 nir_intrinsic_instr *c_intr = nir_instr_as_intrinsic(c);
105 nir_builder_instr_insert(b, c);
106 nir_pop_if(b, NULL);
107
108 if (has_dest) {
109 assert(c_intr->dest.is_ssa);
110 nir_ssa_def *c_ssa = &c_intr->dest.ssa;
111 res = nir_if_phi(b, c_ssa, res);
112 }
113 }
114
115 if (has_dest)
116 nir_ssa_def_rewrite_uses(&intr->dest.ssa, res);
117
118 nir_instr_remove(instr);
119 return true;
120 }
121
122 bool
bi_lower_divergent_indirects(nir_shader * shader,unsigned lanes)123 bi_lower_divergent_indirects(nir_shader *shader, unsigned lanes)
124 {
125 return nir_shader_instructions_pass(shader,
126 bi_lower_divergent_indirects_impl,
127 nir_metadata_none, &lanes);
128 }
129