1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
|
/*
* Copyright (C) 2020 Linux Studio Plugins Project <https://lsp-plug.in/>
* (C) 2020 Vladimir Sadovnikov <sadko4u@gmail.com>
*
* This file is part of lsp-dsp-lib
* Created on: 31 мар. 2020 г.
*
* lsp-dsp-lib is free software: you can redistribute it and/or modify
* it under the terms of the GNU Lesser General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* any later version.
*
* lsp-dsp-lib is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public License
* along with lsp-dsp-lib. If not, see <https://www.gnu.org/licenses/>.
*/
#ifndef PRIVATE_DSP_ARCH_AARCH64_ASIMD_CONVOLUTION_H_
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_CONVOLUTION_H_
#ifndef PRIVATE_DSP_ARCH_AARCH64_ASIMD_IMPL
#error "This header should not be included directly"
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_IMPL */
namespace lsp
{
namespace asimd
{
void convolve(float *dst, const float *src, const float *conv, size_t length, size_t count)
{
IF_ARCH_AARCH64(
const float *c;
float *d;
size_t clen;
)
ARCH_AARCH64_ASM(
// 4x blocks
__ASM_EMIT("subs %[count], %[count], #4")
__ASM_EMIT("b.lo 200f")
__ASM_EMIT("eor v17.16b, v17.16b, v17.16b") // v17 = 0 0 0 0
__ASM_EMIT("100:")
__ASM_EMIT("mov %[d], %[dst]")
__ASM_EMIT("mov %[c], %[conv]")
__ASM_EMIT("ld4r {v4.4s, v5.4s, v6.4s, v7.4s}, [%[k]]")
__ASM_EMIT("eor v16.16b, v16.16b, v16.16b") // v16 = p0 p0 p0 p0 = 0 0 0 0 (history)
// v4 = k0 k0 k0 k0
// v5 = k1 k1 k1 k1
// v6 = k2 k2 k2 k2
// v7 = k3 k3 k3 k3
// 8x convolution
__ASM_EMIT("subs %[clen], %[length], #8")
__ASM_EMIT("b.lo 10f")
__ASM_EMIT("11:")
__ASM_EMIT("ldp q8, q9, [%[c]]") // v8 = c0 c1 c2 c3, v9 = c4 c5 c6 c7
__ASM_EMIT("ldp q0, q1, [%[d]]") // v0 = d0 d1 d2 d3, v1 = d4 d5 d6 d7
__ASM_EMIT("ext v10.16b, v16.16b, v8.16b, #12") // v10 = p3 c0 c1 c2
__ASM_EMIT("ext v11.16b, v8.16b, v9.16b, #12") // v11 = c3 c4 c5 c6
__ASM_EMIT("ext v12.16b, v16.16b, v8.16b, #8") // v12 = p2 p3 c0 c1
__ASM_EMIT("ext v13.16b, v8.16b, v9.16b, #8") // v13 = c2 c3 c4 c5
__ASM_EMIT("ext v14.16b, v16.16b, v8.16b, #4") // v14 = p1 p2 p3 c0
__ASM_EMIT("ext v15.16b, v8.16b, v9.16b, #4") // v15 = c1 c2 c3 c4
__ASM_EMIT("fmla v0.4s, v4.4s, v8.4s") // v0 = d0+k0*c0 d1+k0*c1 d2+k0*c2 d3+k0*c3
__ASM_EMIT("fmla v1.4s, v4.4s, v9.4s") // v1 = d4+k0*c4 d5+k0*c5 d6+k0*c6 d7+k0*c7
__ASM_EMIT("fmla v0.4s, v5.4s, v10.4s") // v0 = d0 + k0*c0 + k1*p3 ...
__ASM_EMIT("fmla v1.4s, v5.4s, v11.4s") // v1 = d1 + k0*c4 + k1*c3 ...
__ASM_EMIT("fmla v0.4s, v6.4s, v12.4s") // v0 = d0 + k0*c0 + k1*p3 + k2*p2 ...
__ASM_EMIT("fmla v1.4s, v6.4s, v13.4s") // v1 = d1 + k0*c4 + k1*c3 + k2*c2 ...
__ASM_EMIT("fmla v0.4s, v7.4s, v14.4s") // v0 = d0 + k0*c0 + k1*p3 + k2*p2 + k3*p1 ...
__ASM_EMIT("fmla v1.4s, v7.4s, v15.4s") // v1 = d1 + k0*c4 + k1*c3 + k2*c2 + k3*c1 ...
__ASM_EMIT("mov v16.16b, v9.16b") // v16 = c4 c5 c6 c7
__ASM_EMIT("stp q0, q1, [%[d]]") // v0 = d0 d1 d2 d3, v1 = d4 d5 d6 d7
__ASM_EMIT("subs %[clen], %[clen], #8")
__ASM_EMIT("add %[c], %[c], #0x20") // c += 8
__ASM_EMIT("add %[d], %[d], #0x20") // d += 8
__ASM_EMIT("b.hs 11b")
__ASM_EMIT("10:")
// 4x convolution
__ASM_EMIT("adds %[clen], %[clen], #4")
__ASM_EMIT("b.lo 12f")
__ASM_EMIT("ldr q8, [%[c]]") // v8 = c0 c1 c2 c3
__ASM_EMIT("ldr q0, [%[d]]") // v0 = d0 d1 d2 d3
__ASM_EMIT("ext v10.16b, v16.16b, v8.16b, #12") // v10 = p3 c0 c1 c2
__ASM_EMIT("ext v12.16b, v16.16b, v8.16b, #8") // v12 = p2 p3 c0 c1
__ASM_EMIT("ext v14.16b, v16.16b, v8.16b, #4") // v14 = p1 p2 p3 c0
__ASM_EMIT("fmla v0.4s, v4.4s, v8.4s") // v0 = d0+k0*c0 d1+k0*c1 d2+k0*c2 d3+k0*c3
__ASM_EMIT("fmla v0.4s, v5.4s, v10.4s") // v0 = d0 + k0*c0 + k1*p3 ...
__ASM_EMIT("fmla v0.4s, v6.4s, v12.4s") // v0 = d0 + k0*c0 + k1*p3 + k2*p2 ...
__ASM_EMIT("fmla v0.4s, v7.4s, v14.4s") // v0 = d0 + k0*c0 + k1*p3 + k2*p2 + k3*p1 ...
__ASM_EMIT("mov v16.16b, v8.16b") // v16 = c0 c1 c2 c3
__ASM_EMIT("str q0, [%[d]]")
__ASM_EMIT("sub %[clen], %[clen], #4")
__ASM_EMIT("add %[c], %[c], #0x10") // c += 4
__ASM_EMIT("add %[d], %[d], #0x10") // d += 4
__ASM_EMIT("12:")
// Apply tail: v16 = p0 p1 p2 p3
__ASM_EMIT("ldr d0, [%[d], #0x00]") // v0 = d0 d1
__ASM_EMIT("ldr s1, [%[d], #0x08]") // v1 = d2
__ASM_EMIT("ext v10.16b, v16.16b, v17.16b, #12") // v10 = p3 0 0 0
__ASM_EMIT("mov v0.s[2], v1.s[0]") // v0 = d0 d1 d2
__ASM_EMIT("ext v12.16b, v16.16b, v17.16b, #8") // v12 = p2 p3 0 0
__ASM_EMIT("ext v14.16b, v16.16b, v17.16b, #4") // v14 = p1 p2 p3 0
__ASM_EMIT("fmla v0.4s, v5.4s, v10.4s") // v0 = d0 + k1*p3 ...
__ASM_EMIT("fmla v0.4s, v6.4s, v12.4s") // v0 = d0 + k1*p3 + k2*p2 ...
__ASM_EMIT("fmla v0.4s, v7.4s, v14.4s") // v0 = d0 + k1*p3 + k2*p2 + k3*p1 ...
__ASM_EMIT("str d0, [%[d], #0x00]")
__ASM_EMIT("mov v1.s[0], v0.s[2]")
__ASM_EMIT("str s1, [%[d], #0x08]")
// 1x convolution
__ASM_EMIT("adds %[clen], %[clen], #3")
__ASM_EMIT("b.lo 14f")
__ASM_EMIT("ldr q4, [%[k]]") // q4 = k0 k1 k2 k3
__ASM_EMIT("13:")
__ASM_EMIT("ld1r {v8.4s}, [%[c]]") // v8 = c0 c0 c0 c0
__ASM_EMIT("ldr q0, [%[d]]") // v0 = d0 d1 d2 d3
__ASM_EMIT("fmla v0.4s, v4.4s, v8.4s") // v0 = d0+k0*c0 d1+k1*c0 d2+k2*c0 d3+k3*c0
__ASM_EMIT("str q0, [%[d]]")
__ASM_EMIT("subs %[clen], %[clen], #1")
__ASM_EMIT("add %[c], %[c], #0x04")
__ASM_EMIT("add %[d], %[d], #0x04")
__ASM_EMIT("b.ge 13b")
__ASM_EMIT("14:")
__ASM_EMIT("subs %[count], %[count], #4")
__ASM_EMIT("add %[dst], %[dst], #0x10")
__ASM_EMIT("add %[k], %[k], #0x10")
__ASM_EMIT("b.ge 100b")
// 1x blocks
__ASM_EMIT("200:")
__ASM_EMIT("adds %[count], %[count], #3")
__ASM_EMIT("b.lt 400f")
__ASM_EMIT("300:")
__ASM_EMIT("ld1r {v8.4s}, [%[k]]") // v8 = k0 k0 k0 k0
__ASM_EMIT("mov %[d], %[dst]")
__ASM_EMIT("mov v9.16b, v8.16b") // v9 = k0 k0 k0 k0
__ASM_EMIT("mov %[c], %[conv]")
// 16x convolution
__ASM_EMIT("subs %[clen], %[length], #16")
__ASM_EMIT("b.lo 20f")
__ASM_EMIT("21:")
__ASM_EMIT("ldp q0, q1, [%[d], #0x00]")
__ASM_EMIT("ldp q2, q3, [%[d], #0x20]")
__ASM_EMIT("ldp q4, q5, [%[c], #0x00]")
__ASM_EMIT("ldp q6, q7, [%[c], #0x20]")
__ASM_EMIT("fmla v0.4s, v4.4s, v8.4s")
__ASM_EMIT("fmla v1.4s, v5.4s, v9.4s")
__ASM_EMIT("fmla v2.4s, v6.4s, v8.4s")
__ASM_EMIT("fmla v3.4s, v7.4s, v9.4s")
__ASM_EMIT("stp q0, q1, [%[d], #0x00]")
__ASM_EMIT("stp q2, q3, [%[d], #0x20]")
__ASM_EMIT("subs %[clen], %[clen], #16")
__ASM_EMIT("add %[c], %[c], #0x40")
__ASM_EMIT("add %[d], %[d], #0x40")
__ASM_EMIT("b.hs 21b")
__ASM_EMIT("20:")
// 8x convolution
__ASM_EMIT("adds %[clen], %[clen], #8")
__ASM_EMIT("b.lo 22f")
__ASM_EMIT("ldp q0, q1, [%[d], #0x00]")
__ASM_EMIT("ldp q4, q5, [%[c], #0x00]")
__ASM_EMIT("fmla v0.4s, v4.4s, v8.4s")
__ASM_EMIT("fmla v1.4s, v5.4s, v9.4s")
__ASM_EMIT("stp q0, q1, [%[d], #0x00]")
__ASM_EMIT("sub %[clen], %[clen], #8")
__ASM_EMIT("add %[c], %[c], #0x20")
__ASM_EMIT("add %[d], %[d], #0x20")
__ASM_EMIT("22:")
// 4x convolution
__ASM_EMIT("adds %[clen], %[clen], #4")
__ASM_EMIT("b.lo 24f")
__ASM_EMIT("ldr q0, [%[d], #0x00]")
__ASM_EMIT("ldr q4, [%[c], #0x00]")
__ASM_EMIT("fmla v0.4s, v4.4s, v8.4s")
__ASM_EMIT("str q0, [%[d], #0x00]")
__ASM_EMIT("sub %[clen], %[clen], #4")
__ASM_EMIT("add %[c], %[c], #0x10")
__ASM_EMIT("add %[d], %[d], #0x10")
__ASM_EMIT("24:")
// 1x tail convolution
__ASM_EMIT("adds %[clen], %[clen], #3")
__ASM_EMIT("b.lo 26f")
__ASM_EMIT("25:")
__ASM_EMIT("ld1r {v0.4s}, [%[d]]")
__ASM_EMIT("ld1r {v4.4s}, [%[c]]")
__ASM_EMIT("fmla v0.4s, v4.4s, v8.4s")
__ASM_EMIT("st1 {v0.s}[0], [%[d]]")
__ASM_EMIT("subs %[clen], %[clen], #1")
__ASM_EMIT("add %[c], %[c], #0x04")
__ASM_EMIT("add %[d], %[d], #0x04")
__ASM_EMIT("b.ge 25b")
__ASM_EMIT("26:")
__ASM_EMIT("subs %[count], %[count], #1")
__ASM_EMIT("add %[dst], %[dst], #0x04")
__ASM_EMIT("add %[k], %[k], #0x04")
__ASM_EMIT("b.ge 300b")
__ASM_EMIT("400:")
: [dst] "+r" (dst),
[k] "+r" (src), [count] "+r" (count),
[c] "=&r" (c), [d] "=&r" (d),
[clen] "=&r" (clen)
: [conv] "r" (conv), [length] "r" (length)
: "cc", "memory",
"v0", "v1", "v2", "v3",
"v4", "v5", "v6", "v7",
"v8", "v9", "v10", "v11",
"v12", "v13", "v14", "v15",
"v16", "v17"
);
}
}
}
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_CONVOLUTION_H_ */
|
