1 /*
2 * Copyright © 2018 Intel Corporation
3 * Copyright © 2018 Broadcom
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 "v3d_compiler.h"
26 #include "compiler/nir/nir_builder.h"
27 #include "compiler/nir/nir_format_convert.h"
28
29 /** @file v3d_nir_lower_scratch.c
30 *
31 * Swizzles around the addresses of
32 * nir_intrinsic_load_scratch/nir_intrinsic_store_scratch so that a QPU stores
33 * a cacheline at a time per dword of scratch access, scalarizing and removing
34 * writemasks in the process.
35 */
36
37 static nir_ssa_def *
v3d_nir_scratch_offset(nir_builder * b,nir_intrinsic_instr * instr)38 v3d_nir_scratch_offset(nir_builder *b, nir_intrinsic_instr *instr)
39 {
40 bool is_store = instr->intrinsic == nir_intrinsic_store_scratch;
41 nir_ssa_def *offset = nir_ssa_for_src(b, instr->src[is_store ? 1 : 0], 1);
42
43 assert(nir_intrinsic_align_mul(instr) >= 4);
44 assert(nir_intrinsic_align_offset(instr) == 0);
45
46 /* The spill_offset register will already have the subgroup ID (EIDX)
47 * shifted and ORed in at bit 2, so all we need to do is to move the
48 * dword index up above V3D_CHANNELS.
49 */
50 return nir_imul_imm(b, offset, V3D_CHANNELS);
51 }
52
53 static void
v3d_nir_lower_load_scratch(nir_builder * b,nir_intrinsic_instr * instr)54 v3d_nir_lower_load_scratch(nir_builder *b, nir_intrinsic_instr *instr)
55 {
56 b->cursor = nir_before_instr(&instr->instr);
57
58 nir_ssa_def *offset = v3d_nir_scratch_offset(b,instr);
59
60 nir_ssa_def *chans[NIR_MAX_VEC_COMPONENTS];
61 for (int i = 0; i < instr->num_components; i++) {
62 nir_ssa_def *chan_offset =
63 nir_iadd_imm(b, offset, V3D_CHANNELS * i * 4);
64
65 nir_intrinsic_instr *chan_instr =
66 nir_intrinsic_instr_create(b->shader, instr->intrinsic);
67 chan_instr->num_components = 1;
68 nir_ssa_dest_init(&chan_instr->instr, &chan_instr->dest, 1,
69 instr->dest.ssa.bit_size, NULL);
70
71 chan_instr->src[0] = nir_src_for_ssa(chan_offset);
72
73 nir_intrinsic_set_align(chan_instr, 4, 0);
74
75 nir_builder_instr_insert(b, &chan_instr->instr);
76
77 chans[i] = &chan_instr->dest.ssa;
78 }
79
80 nir_ssa_def *result = nir_vec(b, chans, instr->num_components);
81 nir_ssa_def_rewrite_uses(&instr->dest.ssa, result);
82 nir_instr_remove(&instr->instr);
83 }
84
85 static void
v3d_nir_lower_store_scratch(nir_builder * b,nir_intrinsic_instr * instr)86 v3d_nir_lower_store_scratch(nir_builder *b, nir_intrinsic_instr *instr)
87 {
88 b->cursor = nir_before_instr(&instr->instr);
89
90 nir_ssa_def *offset = v3d_nir_scratch_offset(b, instr);
91 nir_ssa_def *value = nir_ssa_for_src(b, instr->src[0],
92 instr->num_components);
93
94 for (int i = 0; i < instr->num_components; i++) {
95 if (!(nir_intrinsic_write_mask(instr) & (1 << i)))
96 continue;
97
98 nir_ssa_def *chan_offset =
99 nir_iadd_imm(b, offset, V3D_CHANNELS * i * 4);
100
101 nir_intrinsic_instr *chan_instr =
102 nir_intrinsic_instr_create(b->shader, instr->intrinsic);
103 chan_instr->num_components = 1;
104
105 chan_instr->src[0] = nir_src_for_ssa(nir_channel(b,
106 value,
107 i));
108 chan_instr->src[1] = nir_src_for_ssa(chan_offset);
109 nir_intrinsic_set_write_mask(chan_instr, 0x1);
110 nir_intrinsic_set_align(chan_instr, 4, 0);
111
112 nir_builder_instr_insert(b, &chan_instr->instr);
113 }
114
115 nir_instr_remove(&instr->instr);
116 }
117
118 static bool
v3d_nir_lower_scratch_cb(nir_builder * b,nir_instr * instr,void * _state)119 v3d_nir_lower_scratch_cb(nir_builder *b,
120 nir_instr *instr,
121 void *_state)
122 {
123 if (instr->type != nir_instr_type_intrinsic)
124 return false;
125
126 nir_intrinsic_instr *intr =
127 nir_instr_as_intrinsic(instr);
128
129 switch (intr->intrinsic) {
130 case nir_intrinsic_load_scratch:
131 v3d_nir_lower_load_scratch(b, intr);
132 return true;
133 case nir_intrinsic_store_scratch:
134 v3d_nir_lower_store_scratch(b, intr);
135 return true;
136 default:
137 return false;
138 }
139
140 return false;
141 }
142
143 bool
v3d_nir_lower_scratch(nir_shader * s)144 v3d_nir_lower_scratch(nir_shader *s)
145 {
146 return nir_shader_instructions_pass(s, v3d_nir_lower_scratch_cb,
147 nir_metadata_block_index |
148 nir_metadata_dominance, NULL);
149 }
150