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