| 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_HMATH_HSUM_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_HMATH_HSUM_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 |
float h_sum(const float *src, size_t count) |
| 34 |
{ |
| 35 |
IF_ARCH_AARCH64(float res); |
| 36 |
ARCH_AARCH64_ASM |
| 37 |
( |
| 38 |
__ASM_EMIT("eor v0.16b, v0.16b, v0.16b") |
| 39 |
__ASM_EMIT("eor v1.16b, v1.16b, v1.16b") |
| 40 |
__ASM_EMIT("subs %[count], %[count], #24") |
| 41 |
__ASM_EMIT("b.lt 2f") |
| 42 |
/* 24x block */ |
| 43 |
__ASM_EMIT("1:") |
| 44 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 45 |
__ASM_EMIT("ldp q4, q5, [%[src], #0x20]") |
| 46 |
__ASM_EMIT("ldp q6, q7, [%[src], #0x40]") |
| 47 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 48 |
__ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") |
| 49 |
__ASM_EMIT("fadd v0.4s, v0.4s, v4.4s") |
| 50 |
__ASM_EMIT("fadd v1.4s, v1.4s, v5.4s") |
| 51 |
__ASM_EMIT("fadd v0.4s, v0.4s, v6.4s") |
| 52 |
__ASM_EMIT("fadd v1.4s, v1.4s, v7.4s") |
| 53 |
__ASM_EMIT("subs %[count], %[count], #24") |
| 54 |
__ASM_EMIT("add %[src], %[src], #0x60") |
| 55 |
__ASM_EMIT("b.hs 1b") |
| 56 |
/* 16x block */ |
| 57 |
__ASM_EMIT("2:") |
| 58 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 59 |
__ASM_EMIT("b.lt 4f") |
| 60 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 61 |
__ASM_EMIT("ldp q4, q5, [%[src], #0x20]") |
| 62 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 63 |
__ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") |
| 64 |
__ASM_EMIT("fadd v0.4s, v0.4s, v4.4s") |
| 65 |
__ASM_EMIT("fadd v1.4s, v1.4s, v5.4s") |
| 66 |
__ASM_EMIT("sub %[count], %[count], #16") |
| 67 |
__ASM_EMIT("add %[src], %[src], #0x40") |
| 68 |
/* 8x block */ |
| 69 |
__ASM_EMIT("4:") |
| 70 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 71 |
__ASM_EMIT("b.lt 6f") |
| 72 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 73 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 74 |
__ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") |
| 75 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 76 |
__ASM_EMIT("add %[src], %[src], #0x20") |
| 77 |
/* 4x block */ |
| 78 |
__ASM_EMIT("6:") |
| 79 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 80 |
__ASM_EMIT("fadd v0.4s, v0.4s, v1.4s") |
| 81 |
__ASM_EMIT("b.lt 8f") |
| 82 |
__ASM_EMIT("ldr q2, [%[src], #0x00]") |
| 83 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 84 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 85 |
__ASM_EMIT("add %[src], %[src], #0x10") |
| 86 |
/* 1x block */ |
| 87 |
__ASM_EMIT("8:") |
| 88 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 89 |
__ASM_EMIT("b.lt 10f") |
| 90 |
__ASM_EMIT("eor v2.16b, v2.16b, v2.16b") |
| 91 |
__ASM_EMIT("9:") |
| 92 |
__ASM_EMIT("ld1 {v2.s}[0], [%[src]]") |
| 93 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 94 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 95 |
__ASM_EMIT("add %[src], %[src], #0x04") |
| 96 |
__ASM_EMIT("b.ge 9b") |
| 97 |
/* end of sum */ |
| 98 |
__ASM_EMIT("10:") |
| 99 |
__ASM_EMIT("ext v1.16b, v0.16b, v0.16b, #8") /* v0 = a0 a1 a2 a3, v1 = a2 a3 a0 a1 */ |
| 100 |
__ASM_EMIT("fadd v0.4s, v0.4s, v1.4s") /* v0 = a0+a2 a1+a3 a0+a2 a1+a3 */ |
| 101 |
__ASM_EMIT("ext v1.16b, v0.16b, v0.16b, #4") /* v1 = a1+a3 a0+a2 a1+a3 a0+a2 */ |
| 102 |
__ASM_EMIT("fadd %[res].4s, v0.4s, v1.4s") /* v0 = a0+a1+a2+a3 */ |
| 103 |
: [res] "=w" (res), |
| 104 |
[src] "+r" (src), [count] "+r" (count) |
| 105 |
: |
| 106 |
: "cc", "memory", |
| 107 |
"v1", "v2", "v3", "v4", "v5", "v6", "v7" |
| 108 |
); |
| 109 |
|
| 110 |
return res; |
| 111 |
} |
| 112 |
|
| 113 |
float h_sqr_sum(const float *src, size_t count) |
| 114 |
{ |
| 115 |
IF_ARCH_AARCH64(float res); |
| 116 |
ARCH_AARCH64_ASM |
| 117 |
( |
| 118 |
__ASM_EMIT("eor v0.16b, v0.16b, v0.16b") |
| 119 |
__ASM_EMIT("eor v1.16b, v1.16b, v1.16b") |
| 120 |
__ASM_EMIT("subs %[count], %[count], #24") |
| 121 |
__ASM_EMIT("b.lt 2f") |
| 122 |
/* 24x block */ |
| 123 |
__ASM_EMIT("1:") |
| 124 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 125 |
__ASM_EMIT("ldp q4, q5, [%[src], #0x20]") |
| 126 |
__ASM_EMIT("ldp q6, q7, [%[src], #0x40]") |
| 127 |
__ASM_EMIT("fmla v0.4s, v2.4s, v2.4s") |
| 128 |
__ASM_EMIT("fmla v1.4s, v3.4s, v3.4s") |
| 129 |
__ASM_EMIT("fmla v0.4s, v4.4s, v4.4s") |
| 130 |
__ASM_EMIT("fmla v1.4s, v5.4s, v5.4s") |
| 131 |
__ASM_EMIT("fmla v0.4s, v6.4s, v6.4s") |
| 132 |
__ASM_EMIT("fmla v1.4s, v7.4s, v7.4s") |
| 133 |
__ASM_EMIT("subs %[count], %[count], #24") |
| 134 |
__ASM_EMIT("add %[src], %[src], #0x60") |
| 135 |
__ASM_EMIT("b.hs 1b") |
| 136 |
/* 16x block */ |
| 137 |
__ASM_EMIT("2:") |
| 138 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 139 |
__ASM_EMIT("b.lt 4f") |
| 140 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 141 |
__ASM_EMIT("ldp q4, q5, [%[src], #0x20]") |
| 142 |
__ASM_EMIT("fmla v0.4s, v2.4s, v2.4s") |
| 143 |
__ASM_EMIT("fmla v1.4s, v3.4s, v3.4s") |
| 144 |
__ASM_EMIT("fmla v0.4s, v4.4s, v4.4s") |
| 145 |
__ASM_EMIT("fmla v1.4s, v5.4s, v5.4s") |
| 146 |
__ASM_EMIT("sub %[count], %[count], #16") |
| 147 |
__ASM_EMIT("add %[src], %[src], #0x40") |
| 148 |
/* 8x block */ |
| 149 |
__ASM_EMIT("4:") |
| 150 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 151 |
__ASM_EMIT("b.lt 6f") |
| 152 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 153 |
__ASM_EMIT("fmla v0.4s, v2.4s, v2.4s") |
| 154 |
__ASM_EMIT("fmla v1.4s, v3.4s, v3.4s") |
| 155 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 156 |
__ASM_EMIT("add %[src], %[src], #0x20") |
| 157 |
/* 4x block */ |
| 158 |
__ASM_EMIT("6:") |
| 159 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 160 |
__ASM_EMIT("fadd v0.4s, v0.4s, v1.4s") |
| 161 |
__ASM_EMIT("b.lt 8f") |
| 162 |
__ASM_EMIT("ldr q2, [%[src], #0x00]") |
| 163 |
__ASM_EMIT("fmla v0.4s, v2.4s, v2.4s") |
| 164 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 165 |
__ASM_EMIT("add %[src], %[src], #0x10") |
| 166 |
/* 1x block */ |
| 167 |
__ASM_EMIT("8:") |
| 168 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 169 |
__ASM_EMIT("b.lt 10f") |
| 170 |
__ASM_EMIT("eor v2.16b, v2.16b, v2.16b") |
| 171 |
__ASM_EMIT("9:") |
| 172 |
__ASM_EMIT("ld1 {v2.s}[0], [%[src]]") |
| 173 |
__ASM_EMIT("fmla v0.4s, v2.4s, v2.4s") |
| 174 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 175 |
__ASM_EMIT("add %[src], %[src], #0x04") |
| 176 |
__ASM_EMIT("b.ge 9b") |
| 177 |
/* end of sum */ |
| 178 |
__ASM_EMIT("10:") |
| 179 |
__ASM_EMIT("ext v1.16b, v0.16b, v0.16b, #8") /* v0 = a0 a1 a2 a3, v1 = a2 a3 a0 a1 */ |
| 180 |
__ASM_EMIT("fadd v0.4s, v0.4s, v1.4s") /* v0 = a0+a2 a1+a3 a0+a2 a1+a3 */ |
| 181 |
__ASM_EMIT("ext v1.16b, v0.16b, v0.16b, #4") /* v1 = a1+a3 a0+a2 a1+a3 a0+a2 */ |
| 182 |
__ASM_EMIT("fadd %[res].4s, v0.4s, v1.4s") /* v0 = a0+a1+a2+a3 */ |
| 183 |
: [res] "=w" (res), |
| 184 |
[src] "+r" (src), [count] "+r" (count) |
| 185 |
: |
| 186 |
: "cc", "memory", |
| 187 |
"v1", "v2", "v3", "v4", "v5", "v6", "v7" |
| 188 |
); |
| 189 |
|
| 190 |
return res; |
| 191 |
} |
| 192 |
|
| 193 |
float h_abs_sum(const float *src, size_t count) |
| 194 |
{ |
| 195 |
IF_ARCH_AARCH64(float res); |
| 196 |
ARCH_AARCH64_ASM |
| 197 |
( |
| 198 |
__ASM_EMIT("eor v0.16b, v0.16b, v0.16b") |
| 199 |
__ASM_EMIT("eor v1.16b, v1.16b, v1.16b") |
| 200 |
__ASM_EMIT("subs %[count], %[count], #24") |
| 201 |
__ASM_EMIT("b.lt 2f") |
| 202 |
/* 24x block */ |
| 203 |
__ASM_EMIT("1:") |
| 204 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 205 |
__ASM_EMIT("ldp q4, q5, [%[src], #0x20]") |
| 206 |
__ASM_EMIT("ldp q6, q7, [%[src], #0x40]") |
| 207 |
__ASM_EMIT("fabs v2.4s, v2.4s") |
| 208 |
__ASM_EMIT("fabs v3.4s, v3.4s") |
| 209 |
__ASM_EMIT("fabs v4.4s, v4.4s") |
| 210 |
__ASM_EMIT("fabs v5.4s, v5.4s") |
| 211 |
__ASM_EMIT("fabs v6.4s, v6.4s") |
| 212 |
__ASM_EMIT("fabs v7.4s, v7.4s") |
| 213 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 214 |
__ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") |
| 215 |
__ASM_EMIT("fadd v0.4s, v0.4s, v4.4s") |
| 216 |
__ASM_EMIT("fadd v1.4s, v1.4s, v5.4s") |
| 217 |
__ASM_EMIT("fadd v0.4s, v0.4s, v6.4s") |
| 218 |
__ASM_EMIT("fadd v1.4s, v1.4s, v7.4s") |
| 219 |
__ASM_EMIT("subs %[count], %[count], #24") |
| 220 |
__ASM_EMIT("add %[src], %[src], #0x60") |
| 221 |
__ASM_EMIT("b.hs 1b") |
| 222 |
/* 16x block */ |
| 223 |
__ASM_EMIT("2:") |
| 224 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 225 |
__ASM_EMIT("b.lt 4f") |
| 226 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 227 |
__ASM_EMIT("ldp q4, q5, [%[src], #0x20]") |
| 228 |
__ASM_EMIT("fabs v2.4s, v2.4s") |
| 229 |
__ASM_EMIT("fabs v3.4s, v3.4s") |
| 230 |
__ASM_EMIT("fabs v4.4s, v4.4s") |
| 231 |
__ASM_EMIT("fabs v5.4s, v5.4s") |
| 232 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 233 |
__ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") |
| 234 |
__ASM_EMIT("fadd v0.4s, v0.4s, v4.4s") |
| 235 |
__ASM_EMIT("fadd v1.4s, v1.4s, v5.4s") |
| 236 |
__ASM_EMIT("sub %[count], %[count], #16") |
| 237 |
__ASM_EMIT("add %[src], %[src], #0x40") |
| 238 |
/* 8x block */ |
| 239 |
__ASM_EMIT("4:") |
| 240 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 241 |
__ASM_EMIT("b.lt 6f") |
| 242 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 243 |
__ASM_EMIT("fabs v2.4s, v2.4s") |
| 244 |
__ASM_EMIT("fabs v3.4s, v3.4s") |
| 245 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 246 |
__ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") |
| 247 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 248 |
__ASM_EMIT("add %[src], %[src], #0x20") |
| 249 |
/* 4x block */ |
| 250 |
__ASM_EMIT("6:") |
| 251 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 252 |
__ASM_EMIT("fadd v0.4s, v0.4s, v1.4s") |
| 253 |
__ASM_EMIT("b.lt 8f") |
| 254 |
__ASM_EMIT("ldr q2, [%[src], #0x00]") |
| 255 |
__ASM_EMIT("fabs v2.4s, v2.4s") |
| 256 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 257 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 258 |
__ASM_EMIT("add %[src], %[src], #0x10") |
| 259 |
/* 1x block */ |
| 260 |
__ASM_EMIT("8:") |
| 261 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 262 |
__ASM_EMIT("b.lt 10f") |
| 263 |
__ASM_EMIT("eor v2.16b, v2.16b, v2.16b") |
| 264 |
__ASM_EMIT("9:") |
| 265 |
__ASM_EMIT("ld1 {v2.s}[0], [%[src]]") |
| 266 |
__ASM_EMIT("fabs v2.4s, v2.4s") |
| 267 |
__ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") |
| 268 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 269 |
__ASM_EMIT("add %[src], %[src], #0x04") |
| 270 |
__ASM_EMIT("b.ge 9b") |
| 271 |
/* end of sum */ |
| 272 |
__ASM_EMIT("10:") |
| 273 |
__ASM_EMIT("ext v1.16b, v0.16b, v0.16b, #8") /* v0 = a0 a1 a2 a3, v1 = a2 a3 a0 a1 */ |
| 274 |
__ASM_EMIT("fadd v0.4s, v0.4s, v1.4s") /* v0 = a0+a2 a1+a3 a0+a2 a1+a3 */ |
| 275 |
__ASM_EMIT("ext v1.16b, v0.16b, v0.16b, #4") /* v1 = a1+a3 a0+a2 a1+a3 a0+a2 */ |
| 276 |
__ASM_EMIT("fadd %[res].4s, v0.4s, v1.4s") /* v0 = a0+a1+a2+a3 */ |
| 277 |
: [res] "=w" (res), |
| 278 |
[src] "+r" (src), [count] "+r" (count) |
| 279 |
: |
| 280 |
: "cc", "memory", |
| 281 |
"v1", "v2", "v3", "v4", "v5", "v6", "v7" |
| 282 |
); |
| 283 |
|
| 284 |
return res; |
| 285 |
} |
| 286 |
} |
| 287 |
} |
| 288 |
|
| 289 |
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_HMATH_HSUM_H_ */ |