| 1 |
/* |
| 2 |
* Copyright (C) 2024 Linux Studio Plugins Project <https://lsp-plug.in/> |
| 3 |
* (C) 2024 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_PCOMPLEX_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_PCOMPLEX_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 pcomplex_mul2(float *dst, const float *src, size_t count) |
| 34 |
{ |
| 35 |
IF_ARCH_AARCH64(float *ptr1, *ptr2); |
| 36 |
ARCH_AARCH64_ASM |
| 37 |
( |
| 38 |
// x8 blocks |
| 39 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 40 |
__ASM_EMIT("add %[src2], %[src1], #0x20") |
| 41 |
__ASM_EMIT("add %[dst2], %[dst1], #0x20") |
| 42 |
__ASM_EMIT("b.lo 2f") |
| 43 |
__ASM_EMIT("1:") |
| 44 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src1]]") // v16 = sr1, v17 = si1 |
| 45 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[dst1]]") // v24 = dr1, v25 = di1 |
| 46 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src2]]") // v18 = sr2, v19 = si2 |
| 47 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[dst2]]") // v26 = dr2, v27 = di2 |
| 48 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 49 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 50 |
__ASM_EMIT("add %[src1], %[src1], #0x40") |
| 51 |
__ASM_EMIT("fmul v1.4s, v24.4s, v17.4s") // v1 = si*dr |
| 52 |
__ASM_EMIT("fmul v3.4s, v26.4s, v19.4s") |
| 53 |
__ASM_EMIT("add %[src2], %[src2], #0x40") |
| 54 |
__ASM_EMIT("fmls v0.4s, v17.4s, v25.4s") // v0 = sr*dr - si*di |
| 55 |
__ASM_EMIT("fmls v2.4s, v19.4s, v27.4s") |
| 56 |
__ASM_EMIT("fmla v1.4s, v16.4s, v25.4s") // v1 = si*dr + sr*di |
| 57 |
__ASM_EMIT("fmla v3.4s, v18.4s, v27.4s") |
| 58 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst1]]") |
| 59 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst2]]") |
| 60 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 61 |
__ASM_EMIT("add %[dst1], %[dst1], #0x40") |
| 62 |
__ASM_EMIT("add %[dst2], %[dst2], #0x40") |
| 63 |
__ASM_EMIT("b.hs 1b") |
| 64 |
__ASM_EMIT("2:") |
| 65 |
// x4 blocks |
| 66 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 67 |
__ASM_EMIT("b.lt 4f") |
| 68 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src1]]") // v16 = sr1, v17 = si1 |
| 69 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[dst1]]") // v24 = dr1, v25 = di1 |
| 70 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 71 |
__ASM_EMIT("fmul v1.4s, v24.4s, v17.4s") // v1 = si*dr |
| 72 |
__ASM_EMIT("add %[src1], %[src1], #0x20") |
| 73 |
__ASM_EMIT("fmls v0.4s, v17.4s, v25.4s") // v0 = sr*dr - si*di |
| 74 |
__ASM_EMIT("fmla v1.4s, v16.4s, v25.4s") // v1 = si*dr + sr*di |
| 75 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 76 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst1]]") |
| 77 |
__ASM_EMIT("add %[dst1], %[dst1], #0x20") |
| 78 |
__ASM_EMIT("4:") |
| 79 |
// x1 block |
| 80 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 81 |
__ASM_EMIT("b.lt 6f") |
| 82 |
__ASM_EMIT("5:") |
| 83 |
__ASM_EMIT("ld2 {v16.s, v17.s}[0], [%[src1]]") // v16 = sr1, v17 = si1 |
| 84 |
__ASM_EMIT("ld2 {v24.s, v25.s}[0], [%[dst1]]") // v24 = dr1, v25 = di1 |
| 85 |
__ASM_EMIT("fmul v0.2s, v16.2s, v24.2s") // v0 = sr*dr |
| 86 |
__ASM_EMIT("fmul v1.2s, v24.2s, v17.2s") // v1 = si*dr |
| 87 |
__ASM_EMIT("add %[src1], %[src1], #0x08") |
| 88 |
__ASM_EMIT("fmls v0.2s, v17.2s, v25.2s") // v0 = sr*dr - si*di |
| 89 |
__ASM_EMIT("fmla v1.2s, v16.2s, v25.2s") // v1 = si*dr + sr*di |
| 90 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst1]]") |
| 91 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 92 |
__ASM_EMIT("add %[dst1], %[dst1], #0x08") |
| 93 |
__ASM_EMIT("b.ge 5b") |
| 94 |
__ASM_EMIT("6:") |
| 95 |
|
| 96 |
: [dst1] "+r" (dst), [src1] "+r" (src), |
| 97 |
[dst2] "=&r" (ptr1), [src2] "=&r" (ptr2), |
| 98 |
[count] "+r" (count) |
| 99 |
: |
| 100 |
: "cc", "memory", |
| 101 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 102 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 103 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 104 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 105 |
); |
| 106 |
} |
| 107 |
|
| 108 |
void pcomplex_mul3(float *dst, const float *src1, const float *src2, size_t count) |
| 109 |
{ |
| 110 |
ARCH_AARCH64_ASM |
| 111 |
( |
| 112 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 113 |
__ASM_EMIT("b.lo 2f") |
| 114 |
|
| 115 |
// x16 blocks |
| 116 |
__ASM_EMIT("1:") |
| 117 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src1]], #0x20") // v16 = sr1, v17 = si1 |
| 118 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[src2]], #0x20") // v24 = dr1, v25 = di1 |
| 119 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src1]], #0x20") // v18 = sr2, v19 = si2 |
| 120 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[src2]], #0x20") // v26 = dr2, v27 = di2 |
| 121 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[src1]], #0x20") // v20 = sr3, v21 = si3 |
| 122 |
__ASM_EMIT("ld2 {v28.4s, v29.4s}, [%[src2]], #0x20") // v28 = dr3, v29 = di3 |
| 123 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[src1]], #0x20") // v22 = sr4, v23 = si4 |
| 124 |
__ASM_EMIT("ld2 {v30.4s, v31.4s}, [%[src2]], #0x20") // v30 = dr4, v31 = di4 |
| 125 |
|
| 126 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 127 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 128 |
__ASM_EMIT("fmul v4.4s, v20.4s, v28.4s") |
| 129 |
__ASM_EMIT("fmul v6.4s, v22.4s, v30.4s") |
| 130 |
__ASM_EMIT("fmul v1.4s, v24.4s, v17.4s") // v1 = si*dr |
| 131 |
__ASM_EMIT("fmul v3.4s, v26.4s, v19.4s") |
| 132 |
__ASM_EMIT("fmul v5.4s, v28.4s, v21.4s") |
| 133 |
__ASM_EMIT("fmul v7.4s, v30.4s, v23.4s") |
| 134 |
__ASM_EMIT("fmls v0.4s, v17.4s, v25.4s") // v0 = sr*dr - si*di |
| 135 |
__ASM_EMIT("fmls v2.4s, v19.4s, v27.4s") |
| 136 |
__ASM_EMIT("fmls v4.4s, v21.4s, v29.4s") |
| 137 |
__ASM_EMIT("fmls v6.4s, v23.4s, v31.4s") |
| 138 |
__ASM_EMIT("fmla v1.4s, v16.4s, v25.4s") // v1 = si*dr + sr*di |
| 139 |
__ASM_EMIT("fmla v3.4s, v18.4s, v27.4s") |
| 140 |
__ASM_EMIT("fmla v5.4s, v20.4s, v29.4s") |
| 141 |
__ASM_EMIT("fmla v7.4s, v22.4s, v31.4s") |
| 142 |
|
| 143 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 144 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 145 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 146 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 147 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 148 |
__ASM_EMIT("b.hs 1b") |
| 149 |
|
| 150 |
// x8 blocks |
| 151 |
__ASM_EMIT("2:") |
| 152 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 153 |
__ASM_EMIT("b.lt 4f") |
| 154 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src1]], #0x20") // v16 = sr1, v17 = si1 |
| 155 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[src2]], #0x20") // v24 = dr1, v25 = di1 |
| 156 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src1]], #0x20") // v18 = sr2, v19 = si2 |
| 157 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[src2]], #0x20") // v26 = dr2, v27 = di2 |
| 158 |
|
| 159 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 160 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 161 |
__ASM_EMIT("fmul v1.4s, v24.4s, v17.4s") // v1 = si*dr |
| 162 |
__ASM_EMIT("fmul v3.4s, v26.4s, v19.4s") |
| 163 |
__ASM_EMIT("fmls v0.4s, v17.4s, v25.4s") // v0 = sr*dr - si*di |
| 164 |
__ASM_EMIT("fmls v2.4s, v19.4s, v27.4s") |
| 165 |
__ASM_EMIT("fmla v1.4s, v16.4s, v25.4s") // v1 = si*dr + sr*di |
| 166 |
__ASM_EMIT("fmla v3.4s, v18.4s, v27.4s") |
| 167 |
|
| 168 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 169 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 170 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 171 |
|
| 172 |
// x4 blocks |
| 173 |
__ASM_EMIT("4:") |
| 174 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 175 |
__ASM_EMIT("b.lt 6f") |
| 176 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src1]], #0x20") // v16 = sr1, v17 = si1 |
| 177 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[src2]], #0x20") // v24 = dr1, v25 = di1 |
| 178 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 179 |
__ASM_EMIT("fmul v1.4s, v24.4s, v17.4s") // v1 = si*dr |
| 180 |
__ASM_EMIT("fmls v0.4s, v17.4s, v25.4s") // v0 = sr*dr - si*di |
| 181 |
__ASM_EMIT("fmla v1.4s, v16.4s, v25.4s") // v1 = si*dr + sr*di |
| 182 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 183 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 184 |
|
| 185 |
// x1 blocks |
| 186 |
__ASM_EMIT("6:") |
| 187 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 188 |
__ASM_EMIT("b.lt 8f") |
| 189 |
__ASM_EMIT("7:") |
| 190 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[src1]], #0x08") // v16 = sr1, v17 = si1 |
| 191 |
__ASM_EMIT("ld2r {v24.4s, v25.4s}, [%[src2]], #0x08") // v24 = dr1, v25 = di1 |
| 192 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 193 |
__ASM_EMIT("fmul v1.4s, v24.4s, v17.4s") // v1 = si*dr |
| 194 |
__ASM_EMIT("fmls v0.4s, v17.4s, v25.4s") // v0 = sr*dr - si*di |
| 195 |
__ASM_EMIT("fmla v1.4s, v16.4s, v25.4s") // v1 = si*dr + sr*di |
| 196 |
__ASM_EMIT("str s0, [%[dst]], #0x04") |
| 197 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 198 |
__ASM_EMIT("str s1, [%[dst]], #0x04") |
| 199 |
__ASM_EMIT("b.ge 7b") |
| 200 |
|
| 201 |
__ASM_EMIT("8:") |
| 202 |
|
| 203 |
: [dst] "+r" (dst), [src1] "+r" (src1), [src2] "+r" (src2), |
| 204 |
[count] "+r" (count) |
| 205 |
: [kdiv] "r" (&complex_div_const[0]) |
| 206 |
: "cc", "memory", |
| 207 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 208 |
"v8", /* "v9", "v10", "v11", */ "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 209 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 210 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 211 |
); |
| 212 |
} |
| 213 |
|
| 214 |
static const float pcomplex_div_const[] __lsp_aligned16 = |
| 215 |
{ |
| 216 |
LSP_DSP_VEC4(1.0f) |
| 217 |
}; |
| 218 |
|
| 219 |
void pcomplex_div2(float *dst, const float *src, size_t count) |
| 220 |
{ |
| 221 |
ARCH_AARCH64_ASM |
| 222 |
( |
| 223 |
__ASM_EMIT("ldr q8, [%[kdiv]]") |
| 224 |
// x16 blocks |
| 225 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 226 |
__ASM_EMIT("b.lo 2f") |
| 227 |
__ASM_EMIT("1:") |
| 228 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 229 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[dst]], #0x20") // v24 = dr1, v25 = di1 |
| 230 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") // v18 = sr2, v19 = si2 |
| 231 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[dst]], #0x20") // v26 = dr2, v27 = di2 |
| 232 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[src]], #0x20") // v20 = sr3, v21 = si3 |
| 233 |
__ASM_EMIT("ld2 {v28.4s, v29.4s}, [%[dst]], #0x20") // v28 = dr3, v29 = di3 |
| 234 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[src]], #0x20") // v22 = sr4, v23 = si4 |
| 235 |
__ASM_EMIT("ld2 {v30.4s, v31.4s}, [%[dst]], #0x20") // v30 = dr4, v31 = di4 |
| 236 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 237 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 238 |
__ASM_EMIT("fmul v4.4s, v20.4s, v28.4s") |
| 239 |
__ASM_EMIT("fmul v6.4s, v22.4s, v30.4s") |
| 240 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 241 |
__ASM_EMIT("fmul v3.4s, v18.4s, v27.4s") |
| 242 |
__ASM_EMIT("fmul v5.4s, v20.4s, v29.4s") |
| 243 |
__ASM_EMIT("fmul v7.4s, v22.4s, v31.4s") |
| 244 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 245 |
__ASM_EMIT("fmla v2.4s, v19.4s, v27.4s") |
| 246 |
__ASM_EMIT("fmla v4.4s, v21.4s, v29.4s") |
| 247 |
__ASM_EMIT("fmla v6.4s, v23.4s, v31.4s") |
| 248 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 249 |
__ASM_EMIT("fmls v3.4s, v26.4s, v19.4s") |
| 250 |
__ASM_EMIT("fmls v5.4s, v28.4s, v21.4s") |
| 251 |
__ASM_EMIT("fmls v7.4s, v30.4s, v23.4s") |
| 252 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 253 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 254 |
__ASM_EMIT("fmul v20.4s, v20.4s, v20.4s") |
| 255 |
__ASM_EMIT("fmul v22.4s, v22.4s, v22.4s") |
| 256 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 257 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 258 |
__ASM_EMIT("fmla v20.4s, v21.4s, v21.4s") |
| 259 |
__ASM_EMIT("fmla v22.4s, v23.4s, v23.4s") |
| 260 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 261 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v18.4s") |
| 262 |
__ASM_EMIT("fdiv v14.4s, v8.4s, v20.4s") |
| 263 |
__ASM_EMIT("fdiv v15.4s, v8.4s, v22.4s") |
| 264 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 265 |
__ASM_EMIT("fmul v2.4s, v2.4s, v13.4s") |
| 266 |
__ASM_EMIT("fmul v4.4s, v4.4s, v14.4s") |
| 267 |
__ASM_EMIT("fmul v6.4s, v6.4s, v15.4s") |
| 268 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 269 |
__ASM_EMIT("fmul v3.4s, v3.4s, v13.4s") |
| 270 |
__ASM_EMIT("fmul v5.4s, v5.4s, v14.4s") |
| 271 |
__ASM_EMIT("fmul v7.4s, v7.4s, v15.4s") |
| 272 |
__ASM_EMIT("sub %[dst], %[dst], #0x80") // dst -= 0x80 |
| 273 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 274 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 275 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 276 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 277 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 278 |
__ASM_EMIT("b.hs 1b") |
| 279 |
__ASM_EMIT("2:") |
| 280 |
// x8 blocks |
| 281 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 282 |
__ASM_EMIT("b.lt 4f") |
| 283 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 284 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[dst]], #0x20") // v24 = dr1, v25 = di1 |
| 285 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") // v18 = sr2, v19 = si2 |
| 286 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[dst]], #0x20") // v26 = dr2, v27 = di2 |
| 287 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 288 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 289 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 290 |
__ASM_EMIT("fmul v3.4s, v18.4s, v27.4s") |
| 291 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 292 |
__ASM_EMIT("fmla v2.4s, v19.4s, v27.4s") |
| 293 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 294 |
__ASM_EMIT("fmls v3.4s, v26.4s, v19.4s") |
| 295 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 296 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 297 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 298 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 299 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 300 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v18.4s") |
| 301 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 302 |
__ASM_EMIT("fmul v2.4s, v2.4s, v13.4s") |
| 303 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 304 |
__ASM_EMIT("fmul v3.4s, v3.4s, v13.4s") |
| 305 |
__ASM_EMIT("sub %[dst], %[dst], #0x40") // dst -= 0x40 |
| 306 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 307 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 308 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 309 |
__ASM_EMIT("4:") |
| 310 |
// x4 blocks |
| 311 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 312 |
__ASM_EMIT("b.lt 6f") |
| 313 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 314 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[dst]]") // v24 = dr1, v25 = di1 |
| 315 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 316 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 317 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 318 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 319 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 320 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 321 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 322 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 323 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 324 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 325 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 326 |
__ASM_EMIT("6:") |
| 327 |
// x1 blocks |
| 328 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 329 |
__ASM_EMIT("b.lt 8f") |
| 330 |
__ASM_EMIT("7:") |
| 331 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[src]], #0x08") // v16 = sr1, v17 = si1 |
| 332 |
__ASM_EMIT("ld2r {v24.4s, v25.4s}, [%[dst]]") // v24 = dr1, v25 = di1 |
| 333 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 334 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 335 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 336 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 337 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 338 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 339 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 340 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 341 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 342 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 343 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 344 |
__ASM_EMIT("b.ge 7b") |
| 345 |
__ASM_EMIT("8:") |
| 346 |
|
| 347 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 348 |
[count] "+r" (count) |
| 349 |
: [kdiv] "r" (&pcomplex_div_const[0]) |
| 350 |
: "cc", "memory", |
| 351 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 352 |
"v8", /* "v9", "v10", "v11", */ "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 353 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 354 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 355 |
); |
| 356 |
} |
| 357 |
|
| 358 |
void pcomplex_rdiv2(float *dst, const float *src, size_t count) |
| 359 |
{ |
| 360 |
ARCH_AARCH64_ASM |
| 361 |
( |
| 362 |
__ASM_EMIT("ldr q8, [%[kdiv]]") |
| 363 |
// x16 blocks |
| 364 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 365 |
__ASM_EMIT("b.lo 2f") |
| 366 |
__ASM_EMIT("1:") |
| 367 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[src]], #0x20") // v24 = dr1, v25 = di1 |
| 368 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]], #0x20") // v16 = sr1, v17 = si1 |
| 369 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[src]], #0x20") // v26 = dr2, v27 = di2 |
| 370 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[dst]], #0x20") // v18 = sr2, v19 = si2 |
| 371 |
__ASM_EMIT("ld2 {v28.4s, v29.4s}, [%[src]], #0x20") // v28 = dr3, v29 = di3 |
| 372 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[dst]], #0x20") // v20 = sr3, v21 = si3 |
| 373 |
__ASM_EMIT("ld2 {v30.4s, v31.4s}, [%[src]], #0x20") // v30 = dr4, v31 = di4 |
| 374 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[dst]], #0x20") // v22 = sr4, v23 = si4 |
| 375 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 376 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 377 |
__ASM_EMIT("fmul v4.4s, v20.4s, v28.4s") |
| 378 |
__ASM_EMIT("fmul v6.4s, v22.4s, v30.4s") |
| 379 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 380 |
__ASM_EMIT("fmul v3.4s, v18.4s, v27.4s") |
| 381 |
__ASM_EMIT("fmul v5.4s, v20.4s, v29.4s") |
| 382 |
__ASM_EMIT("fmul v7.4s, v22.4s, v31.4s") |
| 383 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 384 |
__ASM_EMIT("fmla v2.4s, v19.4s, v27.4s") |
| 385 |
__ASM_EMIT("fmla v4.4s, v21.4s, v29.4s") |
| 386 |
__ASM_EMIT("fmla v6.4s, v23.4s, v31.4s") |
| 387 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 388 |
__ASM_EMIT("fmls v3.4s, v26.4s, v19.4s") |
| 389 |
__ASM_EMIT("fmls v5.4s, v28.4s, v21.4s") |
| 390 |
__ASM_EMIT("fmls v7.4s, v30.4s, v23.4s") |
| 391 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 392 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 393 |
__ASM_EMIT("fmul v20.4s, v20.4s, v20.4s") |
| 394 |
__ASM_EMIT("fmul v22.4s, v22.4s, v22.4s") |
| 395 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 396 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 397 |
__ASM_EMIT("fmla v20.4s, v21.4s, v21.4s") |
| 398 |
__ASM_EMIT("fmla v22.4s, v23.4s, v23.4s") |
| 399 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 400 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v18.4s") |
| 401 |
__ASM_EMIT("fdiv v14.4s, v8.4s, v20.4s") |
| 402 |
__ASM_EMIT("fdiv v15.4s, v8.4s, v22.4s") |
| 403 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 404 |
__ASM_EMIT("fmul v2.4s, v2.4s, v13.4s") |
| 405 |
__ASM_EMIT("fmul v4.4s, v4.4s, v14.4s") |
| 406 |
__ASM_EMIT("fmul v6.4s, v6.4s, v15.4s") |
| 407 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 408 |
__ASM_EMIT("fmul v3.4s, v3.4s, v13.4s") |
| 409 |
__ASM_EMIT("fmul v5.4s, v5.4s, v14.4s") |
| 410 |
__ASM_EMIT("fmul v7.4s, v7.4s, v15.4s") |
| 411 |
__ASM_EMIT("sub %[dst], %[dst], #0x80") // dst -= 0x80 |
| 412 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 413 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 414 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 415 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 416 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 417 |
__ASM_EMIT("b.hs 1b") |
| 418 |
__ASM_EMIT("2:") |
| 419 |
// x8 blocks |
| 420 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 421 |
__ASM_EMIT("b.lt 4f") |
| 422 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[src]], #0x20") // v24 = dr1, v25 = di1 |
| 423 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]], #0x20") // v16 = sr1, v17 = si1 |
| 424 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[src]], #0x20") // v26 = dr2, v27 = di2 |
| 425 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[dst]], #0x20") // v18 = sr2, v19 = si2 |
| 426 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 427 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 428 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 429 |
__ASM_EMIT("fmul v3.4s, v18.4s, v27.4s") |
| 430 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 431 |
__ASM_EMIT("fmla v2.4s, v19.4s, v27.4s") |
| 432 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 433 |
__ASM_EMIT("fmls v3.4s, v26.4s, v19.4s") |
| 434 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 435 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 436 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 437 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 438 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 439 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v18.4s") |
| 440 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 441 |
__ASM_EMIT("fmul v2.4s, v2.4s, v13.4s") |
| 442 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 443 |
__ASM_EMIT("fmul v3.4s, v3.4s, v13.4s") |
| 444 |
__ASM_EMIT("sub %[dst], %[dst], #0x40") // dst -= 0x40 |
| 445 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 446 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 447 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 448 |
__ASM_EMIT("4:") |
| 449 |
// x4 blocks |
| 450 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 451 |
__ASM_EMIT("b.lt 6f") |
| 452 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[src]], #0x20") // v24 = dr1, v25 = di1 |
| 453 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]]") // v16 = sr1, v17 = si1 |
| 454 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 455 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 456 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 457 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 458 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 459 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 460 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 461 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 462 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 463 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 464 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 465 |
__ASM_EMIT("6:") |
| 466 |
// x1 blocks |
| 467 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 468 |
__ASM_EMIT("b.lt 8f") |
| 469 |
__ASM_EMIT("7:") |
| 470 |
__ASM_EMIT("ld2r {v24.4s, v25.4s}, [%[src]], #0x08") // v24 = dr1, v25 = di1 |
| 471 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[dst]]") // v16 = sr1, v17 = si1 |
| 472 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 473 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 474 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 475 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 476 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 477 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 478 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 479 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 480 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 481 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 482 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 483 |
__ASM_EMIT("b.ge 7b") |
| 484 |
__ASM_EMIT("8:") |
| 485 |
|
| 486 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 487 |
[count] "+r" (count) |
| 488 |
: [kdiv] "r" (&pcomplex_div_const[0]) |
| 489 |
: "cc", "memory", |
| 490 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 491 |
"v8", /* "v9", "v10", "v11", */ "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 492 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 493 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 494 |
); |
| 495 |
} |
| 496 |
|
| 497 |
void pcomplex_div3(float *dst, const float *t, const float *b, size_t count) |
| 498 |
{ |
| 499 |
ARCH_AARCH64_ASM |
| 500 |
( |
| 501 |
__ASM_EMIT("ldr q8, [%[kdiv]]") |
| 502 |
// x16 blocks |
| 503 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 504 |
__ASM_EMIT("b.lo 2f") |
| 505 |
__ASM_EMIT("1:") |
| 506 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[b]], #0x20") // v16 = sr1, v17 = si1 |
| 507 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[t]], #0x20") // v24 = dr1, v25 = di1 |
| 508 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[b]], #0x20") // v18 = sr2, v19 = si2 |
| 509 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[t]], #0x20") // v26 = dr2, v27 = di2 |
| 510 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[b]], #0x20") // v20 = sr3, v21 = si3 |
| 511 |
__ASM_EMIT("ld2 {v28.4s, v29.4s}, [%[t]], #0x20") // v28 = dr3, v29 = di3 |
| 512 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[b]], #0x20") // v22 = sr4, v23 = si4 |
| 513 |
__ASM_EMIT("ld2 {v30.4s, v31.4s}, [%[t]], #0x20") // v30 = dr4, v31 = di4 |
| 514 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 515 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 516 |
__ASM_EMIT("fmul v4.4s, v20.4s, v28.4s") |
| 517 |
__ASM_EMIT("fmul v6.4s, v22.4s, v30.4s") |
| 518 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 519 |
__ASM_EMIT("fmul v3.4s, v18.4s, v27.4s") |
| 520 |
__ASM_EMIT("fmul v5.4s, v20.4s, v29.4s") |
| 521 |
__ASM_EMIT("fmul v7.4s, v22.4s, v31.4s") |
| 522 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 523 |
__ASM_EMIT("fmla v2.4s, v19.4s, v27.4s") |
| 524 |
__ASM_EMIT("fmla v4.4s, v21.4s, v29.4s") |
| 525 |
__ASM_EMIT("fmla v6.4s, v23.4s, v31.4s") |
| 526 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 527 |
__ASM_EMIT("fmls v3.4s, v26.4s, v19.4s") |
| 528 |
__ASM_EMIT("fmls v5.4s, v28.4s, v21.4s") |
| 529 |
__ASM_EMIT("fmls v7.4s, v30.4s, v23.4s") |
| 530 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 531 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 532 |
__ASM_EMIT("fmul v20.4s, v20.4s, v20.4s") |
| 533 |
__ASM_EMIT("fmul v22.4s, v22.4s, v22.4s") |
| 534 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 535 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 536 |
__ASM_EMIT("fmla v20.4s, v21.4s, v21.4s") |
| 537 |
__ASM_EMIT("fmla v22.4s, v23.4s, v23.4s") |
| 538 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 539 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v18.4s") |
| 540 |
__ASM_EMIT("fdiv v14.4s, v8.4s, v20.4s") |
| 541 |
__ASM_EMIT("fdiv v15.4s, v8.4s, v22.4s") |
| 542 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 543 |
__ASM_EMIT("fmul v2.4s, v2.4s, v13.4s") |
| 544 |
__ASM_EMIT("fmul v4.4s, v4.4s, v14.4s") |
| 545 |
__ASM_EMIT("fmul v6.4s, v6.4s, v15.4s") |
| 546 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 547 |
__ASM_EMIT("fmul v3.4s, v3.4s, v13.4s") |
| 548 |
__ASM_EMIT("fmul v5.4s, v5.4s, v14.4s") |
| 549 |
__ASM_EMIT("fmul v7.4s, v7.4s, v15.4s") |
| 550 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 551 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 552 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 553 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 554 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 555 |
__ASM_EMIT("b.hs 1b") |
| 556 |
__ASM_EMIT("2:") |
| 557 |
// x8 blocks |
| 558 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 559 |
__ASM_EMIT("b.lt 4f") |
| 560 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[b]], #0x20") // v16 = sr1, v17 = si1 |
| 561 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[t]], #0x20") // v24 = dr1, v25 = di1 |
| 562 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[b]], #0x20") // v18 = sr2, v19 = si2 |
| 563 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[t]], #0x20") // v26 = dr2, v27 = di2 |
| 564 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 565 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 566 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 567 |
__ASM_EMIT("fmul v3.4s, v18.4s, v27.4s") |
| 568 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 569 |
__ASM_EMIT("fmla v2.4s, v19.4s, v27.4s") |
| 570 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 571 |
__ASM_EMIT("fmls v3.4s, v26.4s, v19.4s") |
| 572 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 573 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 574 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 575 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 576 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 577 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v18.4s") |
| 578 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 579 |
__ASM_EMIT("fmul v2.4s, v2.4s, v13.4s") |
| 580 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 581 |
__ASM_EMIT("fmul v3.4s, v3.4s, v13.4s") |
| 582 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 583 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 584 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 585 |
__ASM_EMIT("4:") |
| 586 |
// x4 blocks |
| 587 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 588 |
__ASM_EMIT("b.lt 6f") |
| 589 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[b]], #0x20") // v16 = sr1, v17 = si1 |
| 590 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[t]], #0x20") // v24 = dr1, v25 = di1 |
| 591 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 592 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 593 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 594 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 595 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 596 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 597 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 598 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 599 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 600 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 601 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 602 |
__ASM_EMIT("6:") |
| 603 |
// x1 blocks |
| 604 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 605 |
__ASM_EMIT("b.lt 8f") |
| 606 |
__ASM_EMIT("7:") |
| 607 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[b]], #0x08") // v16 = sr1, v17 = si1 |
| 608 |
__ASM_EMIT("ld2r {v24.4s, v25.4s}, [%[t]], #0x08") // v24 = dr1, v25 = di1 |
| 609 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 610 |
__ASM_EMIT("fmul v1.4s, v16.4s, v25.4s") // v1 = sr*di |
| 611 |
__ASM_EMIT("fmla v0.4s, v17.4s, v25.4s") // v0 = sr*dr + si*di |
| 612 |
__ASM_EMIT("fmls v1.4s, v24.4s, v17.4s") // v1 = sr*di-si*dr |
| 613 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 614 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*sr+si*si |
| 615 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 616 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = (sr*dr + si*di) / (sr*sr+si*si) |
| 617 |
__ASM_EMIT("fmul v1.4s, v1.4s, v12.4s") // v1 = (sr*di - si*dr) / (sr*sr+si*si) |
| 618 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 619 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 620 |
__ASM_EMIT("b.ge 7b") |
| 621 |
__ASM_EMIT("8:") |
| 622 |
|
| 623 |
: [dst] "+r" (dst), [t] "+r" (t), [b] "+r" (b), |
| 624 |
[count] "+r" (count) |
| 625 |
: [kdiv] "r" (&pcomplex_div_const[0]) |
| 626 |
: "cc", "memory", |
| 627 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 628 |
"v8", /* "v9", "v10", "v11", */ "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 629 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 630 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 631 |
); |
| 632 |
} |
| 633 |
|
| 634 |
void pcomplex_mod(float *dst, const float *src, size_t count) |
| 635 |
{ |
| 636 |
ARCH_AARCH64_ASM |
| 637 |
( |
| 638 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 639 |
__ASM_EMIT("b.lo 2f") |
| 640 |
|
| 641 |
// x16 blocks |
| 642 |
__ASM_EMIT("1:") |
| 643 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 644 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") // v18 = sr2, v19 = si2 |
| 645 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[src]], #0x20") // v20 = sr3, v21 = si3 |
| 646 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[src]], #0x20") // v22 = sr4, v23 = si4 |
| 647 |
|
| 648 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 649 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 650 |
__ASM_EMIT("fmul v20.4s, v20.4s, v20.4s") |
| 651 |
__ASM_EMIT("fmul v22.4s, v22.4s, v22.4s") |
| 652 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*dr + si*di |
| 653 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 654 |
__ASM_EMIT("fmla v20.4s, v21.4s, v21.4s") |
| 655 |
__ASM_EMIT("fmla v22.4s, v23.4s, v23.4s") |
| 656 |
__ASM_EMIT("fsqrt v0.4s, v16.4s") |
| 657 |
__ASM_EMIT("fsqrt v1.4s, v18.4s") |
| 658 |
__ASM_EMIT("fsqrt v2.4s, v20.4s") |
| 659 |
__ASM_EMIT("fsqrt v3.4s, v22.4s") |
| 660 |
|
| 661 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 662 |
__ASM_EMIT("st1 {v0.4s-v3.4s}, [%[dst]], #0x40") |
| 663 |
__ASM_EMIT("b.hs 1b") |
| 664 |
|
| 665 |
// x8 blocks |
| 666 |
__ASM_EMIT("2:") |
| 667 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 668 |
__ASM_EMIT("b.lt 4f") |
| 669 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 670 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") // v18 = sr2, v19 = si2 |
| 671 |
|
| 672 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 673 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 674 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*dr + si*di |
| 675 |
__ASM_EMIT("fmla v18.4s, v19.4s, v19.4s") |
| 676 |
__ASM_EMIT("fsqrt v0.4s, v16.4s") |
| 677 |
__ASM_EMIT("fsqrt v1.4s, v18.4s") |
| 678 |
|
| 679 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 680 |
__ASM_EMIT("st1 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 681 |
|
| 682 |
// x4 blocks |
| 683 |
__ASM_EMIT("4:") |
| 684 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 685 |
__ASM_EMIT("b.lt 6f") |
| 686 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 687 |
|
| 688 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 689 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*dr + si*di |
| 690 |
__ASM_EMIT("fsqrt v0.4s, v16.4s") |
| 691 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 692 |
__ASM_EMIT("st1 {v0.4s}, [%[dst]], #0x10") |
| 693 |
|
| 694 |
// x1 blocks |
| 695 |
__ASM_EMIT("6:") |
| 696 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 697 |
__ASM_EMIT("b.lt 8f") |
| 698 |
__ASM_EMIT("7:") |
| 699 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[src]], #0x08") // v16 = sr1, v17 = si1 |
| 700 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 701 |
__ASM_EMIT("fmla v16.4s, v17.4s, v17.4s") // v16 = sr*dr + si*di |
| 702 |
__ASM_EMIT("fsqrt v0.4s, v16.4s") |
| 703 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 704 |
__ASM_EMIT("str s0, [%[dst]], #0x04") |
| 705 |
__ASM_EMIT("b.ge 7b") |
| 706 |
|
| 707 |
__ASM_EMIT("8:") |
| 708 |
|
| 709 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 710 |
[count] "+r" (count) |
| 711 |
: |
| 712 |
: "cc", "memory", |
| 713 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 714 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 715 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 716 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 717 |
); |
| 718 |
} |
| 719 |
|
| 720 |
void pcomplex_rcp1(float *dst, size_t count) |
| 721 |
{ |
| 722 |
ARCH_AARCH64_ASM |
| 723 |
( |
| 724 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 725 |
__ASM_EMIT("b.lo 2f") |
| 726 |
|
| 727 |
// x16 blocks |
| 728 |
__ASM_EMIT("1:") |
| 729 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]], #0x20") // v16 = sr1, v17 = si1 |
| 730 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[dst]], #0x20") // v18 = sr2, v19 = si2 |
| 731 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[dst]], #0x20") // v20 = sr3, v21 = si3 |
| 732 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[dst]], #0x20") // v22 = sr4, v23 = si4 |
| 733 |
|
| 734 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 735 |
__ASM_EMIT("fmul v25.4s, v18.4s, v18.4s") |
| 736 |
__ASM_EMIT("fmul v26.4s, v20.4s, v20.4s") |
| 737 |
__ASM_EMIT("fmul v27.4s, v22.4s, v22.4s") |
| 738 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 739 |
__ASM_EMIT("fmla v25.4s, v19.4s, v19.4s") |
| 740 |
__ASM_EMIT("fmla v26.4s, v21.4s, v21.4s") |
| 741 |
__ASM_EMIT("fmla v27.4s, v23.4s, v23.4s") |
| 742 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 743 |
__ASM_EMIT("fneg v19.4s, v19.4s") |
| 744 |
__ASM_EMIT("fneg v21.4s, v21.4s") |
| 745 |
__ASM_EMIT("fneg v23.4s, v23.4s") |
| 746 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 747 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v25.4s") |
| 748 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v26.4s") |
| 749 |
__ASM_EMIT("fdiv v22.4s, v22.4s, v27.4s") |
| 750 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 751 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v25.4s") |
| 752 |
__ASM_EMIT("fdiv v21.4s, v21.4s, v26.4s") |
| 753 |
__ASM_EMIT("fdiv v23.4s, v23.4s, v27.4s") |
| 754 |
|
| 755 |
__ASM_EMIT("sub %[dst], %[dst], 0x80") |
| 756 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 757 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 758 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 759 |
__ASM_EMIT("st2 {v20.4s, v21.4s}, [%[dst]], #0x20") |
| 760 |
__ASM_EMIT("st2 {v22.4s, v23.4s}, [%[dst]], #0x20") |
| 761 |
__ASM_EMIT("b.hs 1b") |
| 762 |
|
| 763 |
// x8 blocks |
| 764 |
__ASM_EMIT("2:") |
| 765 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 766 |
__ASM_EMIT("b.lt 4f") |
| 767 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]], #0x20") // v16 = sr1, v17 = si1 |
| 768 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[dst]], #0x20") // v18 = sr2, v19 = si2 |
| 769 |
|
| 770 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 771 |
__ASM_EMIT("fmul v25.4s, v18.4s, v18.4s") |
| 772 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 773 |
__ASM_EMIT("fmla v25.4s, v19.4s, v19.4s") |
| 774 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 775 |
__ASM_EMIT("fneg v19.4s, v19.4s") |
| 776 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 777 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v25.4s") |
| 778 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 779 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v25.4s") |
| 780 |
|
| 781 |
__ASM_EMIT("sub %[dst], %[dst], 0x40") |
| 782 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 783 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 784 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 785 |
|
| 786 |
// x4 blocks |
| 787 |
__ASM_EMIT("4:") |
| 788 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 789 |
__ASM_EMIT("b.lt 6f") |
| 790 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]]") // v16 = sr1, v17 = si1 |
| 791 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 792 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 793 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 794 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 795 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 796 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 797 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 798 |
|
| 799 |
// x1 blocks |
| 800 |
__ASM_EMIT("6:") |
| 801 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 802 |
__ASM_EMIT("b.lt 8f") |
| 803 |
__ASM_EMIT("7:") |
| 804 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[dst]]") // v16 = sr1, v17 = si1 |
| 805 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 806 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 807 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 808 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 809 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 810 |
__ASM_EMIT("str s16, [%[dst]], #0x04") |
| 811 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 812 |
__ASM_EMIT("str s17, [%[dst]], #0x04") |
| 813 |
__ASM_EMIT("b.ge 7b") |
| 814 |
|
| 815 |
__ASM_EMIT("8:") |
| 816 |
|
| 817 |
: [dst] "+r" (dst), |
| 818 |
[count] "+r" (count) |
| 819 |
: |
| 820 |
: "cc", "memory", |
| 821 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 822 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 823 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 824 |
); |
| 825 |
} |
| 826 |
|
| 827 |
void pcomplex_rcp2(float *dst, const float *src, size_t count) |
| 828 |
{ |
| 829 |
ARCH_AARCH64_ASM |
| 830 |
( |
| 831 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 832 |
__ASM_EMIT("b.lo 2f") |
| 833 |
|
| 834 |
// x16 blocks |
| 835 |
__ASM_EMIT("1:") |
| 836 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 837 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") // v18 = sr2, v19 = si2 |
| 838 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[src]], #0x20") // v20 = sr3, v21 = si3 |
| 839 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[src]], #0x20") // v22 = sr4, v23 = si4 |
| 840 |
|
| 841 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 842 |
__ASM_EMIT("fmul v25.4s, v18.4s, v18.4s") |
| 843 |
__ASM_EMIT("fmul v26.4s, v20.4s, v20.4s") |
| 844 |
__ASM_EMIT("fmul v27.4s, v22.4s, v22.4s") |
| 845 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 846 |
__ASM_EMIT("fmla v25.4s, v19.4s, v19.4s") |
| 847 |
__ASM_EMIT("fmla v26.4s, v21.4s, v21.4s") |
| 848 |
__ASM_EMIT("fmla v27.4s, v23.4s, v23.4s") |
| 849 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 850 |
__ASM_EMIT("fneg v19.4s, v19.4s") |
| 851 |
__ASM_EMIT("fneg v21.4s, v21.4s") |
| 852 |
__ASM_EMIT("fneg v23.4s, v23.4s") |
| 853 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 854 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v25.4s") |
| 855 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v26.4s") |
| 856 |
__ASM_EMIT("fdiv v22.4s, v22.4s, v27.4s") |
| 857 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 858 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v25.4s") |
| 859 |
__ASM_EMIT("fdiv v21.4s, v21.4s, v26.4s") |
| 860 |
__ASM_EMIT("fdiv v23.4s, v23.4s, v27.4s") |
| 861 |
|
| 862 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 863 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 864 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 865 |
__ASM_EMIT("st2 {v20.4s, v21.4s}, [%[dst]], #0x20") |
| 866 |
__ASM_EMIT("st2 {v22.4s, v23.4s}, [%[dst]], #0x20") |
| 867 |
__ASM_EMIT("b.hs 1b") |
| 868 |
|
| 869 |
// x8 blocks |
| 870 |
__ASM_EMIT("2:") |
| 871 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 872 |
__ASM_EMIT("b.lt 4f") |
| 873 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 874 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") // v18 = sr2, v19 = si2 |
| 875 |
|
| 876 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 877 |
__ASM_EMIT("fmul v25.4s, v18.4s, v18.4s") |
| 878 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 879 |
__ASM_EMIT("fmla v25.4s, v19.4s, v19.4s") |
| 880 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 881 |
__ASM_EMIT("fneg v19.4s, v19.4s") |
| 882 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 883 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v25.4s") |
| 884 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 885 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v25.4s") |
| 886 |
|
| 887 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 888 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 889 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 890 |
|
| 891 |
// x4 blocks |
| 892 |
__ASM_EMIT("4:") |
| 893 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 894 |
__ASM_EMIT("b.lt 6f") |
| 895 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") // v16 = sr1, v17 = si1 |
| 896 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 897 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 898 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 899 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 900 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 901 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 902 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 903 |
|
| 904 |
// x1 blocks |
| 905 |
__ASM_EMIT("6:") |
| 906 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 907 |
__ASM_EMIT("b.lt 8f") |
| 908 |
__ASM_EMIT("7:") |
| 909 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[src]], #0x08") // v16 = sr1, v17 = si1 |
| 910 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 911 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 912 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 913 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 914 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 915 |
__ASM_EMIT("str s16, [%[dst]], #0x04") |
| 916 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 917 |
__ASM_EMIT("str s17, [%[dst]], #0x04") |
| 918 |
__ASM_EMIT("b.ge 7b") |
| 919 |
|
| 920 |
__ASM_EMIT("8:") |
| 921 |
|
| 922 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 923 |
[count] "+r" (count) |
| 924 |
: |
| 925 |
: "cc", "memory", |
| 926 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 927 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 928 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 929 |
); |
| 930 |
} |
| 931 |
|
| 932 |
void pcomplex_r2c(float *dst, const float *src, size_t count) |
| 933 |
{ |
| 934 |
ARCH_AARCH64_ASM |
| 935 |
( |
| 936 |
__ASM_EMIT("subs %[count], %[count], #32") |
| 937 |
__ASM_EMIT("b.lo 2f") |
| 938 |
|
| 939 |
// x32 blocks |
| 940 |
__ASM_EMIT("eor v25.16b, v25.16b, v25.16b") |
| 941 |
__ASM_EMIT("eor v27.16b, v27.16b, v27.16b") |
| 942 |
__ASM_EMIT("eor v29.16b, v29.16b, v29.16b") |
| 943 |
__ASM_EMIT("eor v31.16b, v31.16b, v31.16b") |
| 944 |
|
| 945 |
__ASM_EMIT("1:") |
| 946 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src]], #0x40") |
| 947 |
__ASM_EMIT("ld1 {v20.4s-v23.4s}, [%[src]], #0x40") |
| 948 |
|
| 949 |
__ASM_EMIT("mov v30.16b, v23.16b") |
| 950 |
__ASM_EMIT("mov v28.16b, v22.16b") |
| 951 |
__ASM_EMIT("mov v26.16b, v21.16b") |
| 952 |
__ASM_EMIT("mov v24.16b, v20.16b") |
| 953 |
__ASM_EMIT("mov v22.16b, v19.16b") |
| 954 |
__ASM_EMIT("mov v20.16b, v18.16b") |
| 955 |
__ASM_EMIT("mov v18.16b, v17.16b") |
| 956 |
__ASM_EMIT("eor v17.16b, v17.16b, v17.16b") |
| 957 |
__ASM_EMIT("eor v19.16b, v19.16b, v19.16b") |
| 958 |
__ASM_EMIT("eor v21.16b, v21.16b, v21.16b") |
| 959 |
__ASM_EMIT("eor v23.16b, v23.16b, v23.16b") |
| 960 |
|
| 961 |
__ASM_EMIT("subs %[count], %[count], #32") |
| 962 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 963 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 964 |
__ASM_EMIT("st2 {v20.4s, v21.4s}, [%[dst]], #0x20") |
| 965 |
__ASM_EMIT("st2 {v22.4s, v23.4s}, [%[dst]], #0x20") |
| 966 |
__ASM_EMIT("st2 {v24.4s, v25.4s}, [%[dst]], #0x20") |
| 967 |
__ASM_EMIT("st2 {v26.4s, v27.4s}, [%[dst]], #0x20") |
| 968 |
__ASM_EMIT("st2 {v28.4s, v29.4s}, [%[dst]], #0x20") |
| 969 |
__ASM_EMIT("st2 {v30.4s, v31.4s}, [%[dst]], #0x20") |
| 970 |
__ASM_EMIT("b.hs 1b") |
| 971 |
|
| 972 |
// x16 blocks |
| 973 |
__ASM_EMIT("2:") |
| 974 |
__ASM_EMIT("adds %[count], %[count], #16") |
| 975 |
__ASM_EMIT("b.lt 4f") |
| 976 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src]], #0x40") |
| 977 |
|
| 978 |
__ASM_EMIT("mov v22.16b, v19.16b") |
| 979 |
__ASM_EMIT("mov v20.16b, v18.16b") |
| 980 |
__ASM_EMIT("mov v18.16b, v17.16b") |
| 981 |
__ASM_EMIT("eor v17.16b, v17.16b, v17.16b") |
| 982 |
__ASM_EMIT("eor v19.16b, v19.16b, v19.16b") |
| 983 |
__ASM_EMIT("eor v21.16b, v21.16b, v21.16b") |
| 984 |
__ASM_EMIT("eor v23.16b, v23.16b, v23.16b") |
| 985 |
|
| 986 |
__ASM_EMIT("sub %[count], %[count], #16") |
| 987 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 988 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 989 |
__ASM_EMIT("st2 {v20.4s, v21.4s}, [%[dst]], #0x20") |
| 990 |
__ASM_EMIT("st2 {v22.4s, v23.4s}, [%[dst]], #0x20") |
| 991 |
|
| 992 |
// x8 block |
| 993 |
__ASM_EMIT("4:") |
| 994 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 995 |
__ASM_EMIT("b.lt 6f") |
| 996 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[src]], #0x20") |
| 997 |
|
| 998 |
__ASM_EMIT("mov v18.16b, v17.16b") |
| 999 |
__ASM_EMIT("eor v17.16b, v17.16b, v17.16b") |
| 1000 |
__ASM_EMIT("eor v19.16b, v19.16b, v19.16b") |
| 1001 |
|
| 1002 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1003 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 1004 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 1005 |
|
| 1006 |
// x4 blocks |
| 1007 |
__ASM_EMIT("6:") |
| 1008 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1009 |
__ASM_EMIT("b.lt 8f") |
| 1010 |
__ASM_EMIT("ld1 {v16.4s}, [%[src]], #0x10") |
| 1011 |
__ASM_EMIT("mov v18.16b, v17.16b") |
| 1012 |
__ASM_EMIT("eor v17.16b, v17.16b, v17.16b") |
| 1013 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1014 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 1015 |
|
| 1016 |
// x1 blocks |
| 1017 |
__ASM_EMIT("8:") |
| 1018 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1019 |
__ASM_EMIT("b.lt 10f") |
| 1020 |
__ASM_EMIT("9:") |
| 1021 |
__ASM_EMIT("ldr s0, [%[src]], #0x04") |
| 1022 |
__ASM_EMIT("str s0, [%[dst]], #0x04") |
| 1023 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1024 |
__ASM_EMIT("str s31, [%[dst]], #0x04") |
| 1025 |
__ASM_EMIT("b.ge 9b") |
| 1026 |
|
| 1027 |
__ASM_EMIT("10:") |
| 1028 |
|
| 1029 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 1030 |
[count] "+r" (count) |
| 1031 |
: |
| 1032 |
: "cc", "memory", |
| 1033 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 1034 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 1035 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 1036 |
); |
| 1037 |
} |
| 1038 |
|
| 1039 |
void pcomplex_c2r(float *dst, const float *src, size_t count) |
| 1040 |
{ |
| 1041 |
ARCH_AARCH64_ASM |
| 1042 |
( |
| 1043 |
__ASM_EMIT("subs %[count], %[count], #32") |
| 1044 |
__ASM_EMIT("b.lo 2f") |
| 1045 |
|
| 1046 |
// x32 blocks |
| 1047 |
__ASM_EMIT("1:") |
| 1048 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") |
| 1049 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") |
| 1050 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[src]], #0x20") |
| 1051 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[src]], #0x20") |
| 1052 |
__ASM_EMIT("ld2 {v24.4s, v25.4s}, [%[src]], #0x20") |
| 1053 |
__ASM_EMIT("ld2 {v26.4s, v27.4s}, [%[src]], #0x20") |
| 1054 |
__ASM_EMIT("ld2 {v28.4s, v29.4s}, [%[src]], #0x20") |
| 1055 |
__ASM_EMIT("ld2 {v30.4s, v31.4s}, [%[src]], #0x20") |
| 1056 |
|
| 1057 |
__ASM_EMIT("mov v17.16b, v18.16b") |
| 1058 |
__ASM_EMIT("mov v18.16b, v20.16b") |
| 1059 |
__ASM_EMIT("mov v19.16b, v22.16b") |
| 1060 |
__ASM_EMIT("mov v20.16b, v24.16b") |
| 1061 |
__ASM_EMIT("mov v21.16b, v26.16b") |
| 1062 |
__ASM_EMIT("mov v22.16b, v28.16b") |
| 1063 |
__ASM_EMIT("mov v23.16b, v30.16b") |
| 1064 |
|
| 1065 |
__ASM_EMIT("subs %[count], %[count], #32") |
| 1066 |
__ASM_EMIT("st1 {v16.4s-v19.4s}, [%[dst]], #0x40") |
| 1067 |
__ASM_EMIT("st1 {v20.4s-v23.4s}, [%[dst]], #0x40") |
| 1068 |
__ASM_EMIT("b.hs 1b") |
| 1069 |
|
| 1070 |
// x16 blocks |
| 1071 |
__ASM_EMIT("2:") |
| 1072 |
__ASM_EMIT("adds %[count], %[count], #16") |
| 1073 |
__ASM_EMIT("b.lt 4f") |
| 1074 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") |
| 1075 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") |
| 1076 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[src]], #0x20") |
| 1077 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[src]], #0x20") |
| 1078 |
|
| 1079 |
__ASM_EMIT("mov v17.16b, v18.16b") |
| 1080 |
__ASM_EMIT("mov v18.16b, v20.16b") |
| 1081 |
__ASM_EMIT("mov v19.16b, v22.16b") |
| 1082 |
|
| 1083 |
__ASM_EMIT("sub %[count], %[count], #16") |
| 1084 |
__ASM_EMIT("st1 {v16.4s-v19.4s}, [%[dst]], #0x40") |
| 1085 |
|
| 1086 |
// x8 block |
| 1087 |
__ASM_EMIT("4:") |
| 1088 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 1089 |
__ASM_EMIT("b.lt 6f") |
| 1090 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") |
| 1091 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[src]], #0x20") |
| 1092 |
__ASM_EMIT("mov v17.16b, v18.16b") |
| 1093 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1094 |
__ASM_EMIT("st1 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 1095 |
|
| 1096 |
// x4 blocks |
| 1097 |
__ASM_EMIT("6:") |
| 1098 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1099 |
__ASM_EMIT("b.lt 8f") |
| 1100 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[src]], #0x20") |
| 1101 |
__ASM_EMIT("mov v17.16b, v18.16b") |
| 1102 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1103 |
__ASM_EMIT("st1 {v16.4s}, [%[dst]], #0x10") |
| 1104 |
|
| 1105 |
// x1 blocks |
| 1106 |
__ASM_EMIT("8:") |
| 1107 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1108 |
__ASM_EMIT("b.lt 10f") |
| 1109 |
__ASM_EMIT("9:") |
| 1110 |
__ASM_EMIT("ld2 {v16.s, v17.s}[0], [%[src]], #0x08") |
| 1111 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1112 |
__ASM_EMIT("str s16, [%[dst]], #0x04") |
| 1113 |
__ASM_EMIT("b.ge 9b") |
| 1114 |
|
| 1115 |
__ASM_EMIT("10:") |
| 1116 |
|
| 1117 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 1118 |
[count] "+r" (count) |
| 1119 |
: |
| 1120 |
: "cc", "memory", |
| 1121 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 1122 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 1123 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 1124 |
); |
| 1125 |
} |
| 1126 |
|
| 1127 |
void pcomplex_r2c_add2(float *dst, const float *src, size_t count) |
| 1128 |
{ |
| 1129 |
ARCH_AARCH64_ASM |
| 1130 |
( |
| 1131 |
// x16 blocks |
| 1132 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1133 |
__ASM_EMIT("b.lo 2f") |
| 1134 |
__ASM_EMIT("1:") |
| 1135 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1136 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1137 |
__ASM_EMIT("ld2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1138 |
__ASM_EMIT("ld2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1139 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src]], #0x40") |
| 1140 |
__ASM_EMIT("fadd v0.4s, v0.4s, v16.4s") |
| 1141 |
__ASM_EMIT("fadd v2.4s, v2.4s, v17.4s") |
| 1142 |
__ASM_EMIT("fadd v4.4s, v4.4s, v18.4s") |
| 1143 |
__ASM_EMIT("fadd v6.4s, v6.4s, v19.4s") |
| 1144 |
__ASM_EMIT("sub %[dst], %[dst], #0x80") |
| 1145 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1146 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1147 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1148 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1149 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1150 |
__ASM_EMIT("b.hs 1b") |
| 1151 |
// x8 block |
| 1152 |
__ASM_EMIT("2:") |
| 1153 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 1154 |
__ASM_EMIT("b.lt 4f") |
| 1155 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1156 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1157 |
__ASM_EMIT("ld1 {v16.4s-v17.4s}, [%[src]], #0x20") |
| 1158 |
__ASM_EMIT("fadd v0.4s, v0.4s, v16.4s") |
| 1159 |
__ASM_EMIT("fadd v2.4s, v2.4s, v17.4s") |
| 1160 |
__ASM_EMIT("sub %[dst], %[dst], #0x40") |
| 1161 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1162 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1163 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1164 |
// x4 block |
| 1165 |
__ASM_EMIT("4:") |
| 1166 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1167 |
__ASM_EMIT("b.lt 6f") |
| 1168 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]]") |
| 1169 |
__ASM_EMIT("ld1 {v16.4s}, [%[src]], #0x10") |
| 1170 |
__ASM_EMIT("fadd v0.4s, v0.4s, v16.4s") |
| 1171 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1172 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1173 |
// x1 blocks |
| 1174 |
__ASM_EMIT("6:") |
| 1175 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1176 |
__ASM_EMIT("b.lt 8f") |
| 1177 |
__ASM_EMIT("7:") |
| 1178 |
__ASM_EMIT("ld2r {v0.4s, v1.4s}, [%[dst]]") |
| 1179 |
__ASM_EMIT("ld1r {v16.4s}, [%[src]], #0x04") |
| 1180 |
__ASM_EMIT("fadd v0.4s, v0.4s, v16.4s") |
| 1181 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1182 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 1183 |
__ASM_EMIT("b.ge 7b") |
| 1184 |
|
| 1185 |
__ASM_EMIT("8:") |
| 1186 |
|
| 1187 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 1188 |
[count] "+r" (count) |
| 1189 |
: |
| 1190 |
: "cc", "memory", |
| 1191 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 1192 |
"v16", "v17", "v18", "v19" |
| 1193 |
); |
| 1194 |
} |
| 1195 |
|
| 1196 |
void pcomplex_r2c_sub2(float *dst, const float *src, size_t count) |
| 1197 |
{ |
| 1198 |
ARCH_AARCH64_ASM |
| 1199 |
( |
| 1200 |
// x16 blocks |
| 1201 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1202 |
__ASM_EMIT("b.lo 2f") |
| 1203 |
__ASM_EMIT("1:") |
| 1204 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1205 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1206 |
__ASM_EMIT("ld2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1207 |
__ASM_EMIT("ld2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1208 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src]], #0x40") |
| 1209 |
__ASM_EMIT("fsub v0.4s, v0.4s, v16.4s") |
| 1210 |
__ASM_EMIT("fsub v2.4s, v2.4s, v17.4s") |
| 1211 |
__ASM_EMIT("fsub v4.4s, v4.4s, v18.4s") |
| 1212 |
__ASM_EMIT("fsub v6.4s, v6.4s, v19.4s") |
| 1213 |
__ASM_EMIT("sub %[dst], %[dst], #0x80") |
| 1214 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1215 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1216 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1217 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1218 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1219 |
__ASM_EMIT("b.hs 1b") |
| 1220 |
// x8 block |
| 1221 |
__ASM_EMIT("2:") |
| 1222 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 1223 |
__ASM_EMIT("b.lt 4f") |
| 1224 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1225 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1226 |
__ASM_EMIT("ld1 {v16.4s-v17.4s}, [%[src]], #0x20") |
| 1227 |
__ASM_EMIT("fsub v0.4s, v0.4s, v16.4s") |
| 1228 |
__ASM_EMIT("fsub v2.4s, v2.4s, v17.4s") |
| 1229 |
__ASM_EMIT("sub %[dst], %[dst], #0x40") |
| 1230 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1231 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1232 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1233 |
// x4 block |
| 1234 |
__ASM_EMIT("4:") |
| 1235 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1236 |
__ASM_EMIT("b.lt 6f") |
| 1237 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]]") |
| 1238 |
__ASM_EMIT("ld1 {v16.4s}, [%[src]], #0x10") |
| 1239 |
__ASM_EMIT("fsub v0.4s, v0.4s, v16.4s") |
| 1240 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1241 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1242 |
// x1 blocks |
| 1243 |
__ASM_EMIT("6:") |
| 1244 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1245 |
__ASM_EMIT("b.lt 8f") |
| 1246 |
__ASM_EMIT("7:") |
| 1247 |
__ASM_EMIT("ld2r {v0.4s, v1.4s}, [%[dst]]") |
| 1248 |
__ASM_EMIT("ld1r {v16.4s}, [%[src]], #0x04") |
| 1249 |
__ASM_EMIT("fsub v0.4s, v0.4s, v16.4s") |
| 1250 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1251 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 1252 |
__ASM_EMIT("b.ge 7b") |
| 1253 |
|
| 1254 |
__ASM_EMIT("8:") |
| 1255 |
|
| 1256 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 1257 |
[count] "+r" (count) |
| 1258 |
: |
| 1259 |
: "cc", "memory", |
| 1260 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 1261 |
"v16", "v17", "v18", "v19" |
| 1262 |
); |
| 1263 |
} |
| 1264 |
|
| 1265 |
void pcomplex_r2c_rsub2(float *dst, const float *src, size_t count) |
| 1266 |
{ |
| 1267 |
ARCH_AARCH64_ASM |
| 1268 |
( |
| 1269 |
// x16 blocks |
| 1270 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1271 |
__ASM_EMIT("b.lo 2f") |
| 1272 |
__ASM_EMIT("1:") |
| 1273 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1274 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1275 |
__ASM_EMIT("ld2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1276 |
__ASM_EMIT("ld2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1277 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src]], #0x40") |
| 1278 |
__ASM_EMIT("fsub v0.4s, v16.4s, v0.4s") |
| 1279 |
__ASM_EMIT("fneg v1.4s, v1.4s") |
| 1280 |
__ASM_EMIT("fsub v2.4s, v17.4s, v2.4s") |
| 1281 |
__ASM_EMIT("fneg v3.4s, v3.4s") |
| 1282 |
__ASM_EMIT("fsub v4.4s, v18.4s, v4.4s") |
| 1283 |
__ASM_EMIT("fneg v5.4s, v5.4s") |
| 1284 |
__ASM_EMIT("fsub v6.4s, v19.4s, v6.4s") |
| 1285 |
__ASM_EMIT("fneg v7.4s, v7.4s") |
| 1286 |
__ASM_EMIT("sub %[dst], %[dst], #0x80") |
| 1287 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1288 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1289 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1290 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1291 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1292 |
__ASM_EMIT("b.hs 1b") |
| 1293 |
// x8 block |
| 1294 |
__ASM_EMIT("2:") |
| 1295 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 1296 |
__ASM_EMIT("b.lt 4f") |
| 1297 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1298 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1299 |
__ASM_EMIT("ld1 {v16.4s-v17.4s}, [%[src]], #0x20") |
| 1300 |
__ASM_EMIT("fsub v0.4s, v16.4s, v0.4s") |
| 1301 |
__ASM_EMIT("fneg v1.4s, v1.4s") |
| 1302 |
__ASM_EMIT("fsub v2.4s, v17.4s, v2.4s") |
| 1303 |
__ASM_EMIT("fneg v3.4s, v3.4s") |
| 1304 |
__ASM_EMIT("sub %[dst], %[dst], #0x40") |
| 1305 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1306 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1307 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1308 |
// x4 block |
| 1309 |
__ASM_EMIT("4:") |
| 1310 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1311 |
__ASM_EMIT("b.lt 6f") |
| 1312 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]]") |
| 1313 |
__ASM_EMIT("ld1 {v16.4s}, [%[src]], #0x10") |
| 1314 |
__ASM_EMIT("fsub v0.4s, v16.4s, v0.4s") |
| 1315 |
__ASM_EMIT("fneg v1.4s, v1.4s") |
| 1316 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1317 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1318 |
// x1 blocks |
| 1319 |
__ASM_EMIT("6:") |
| 1320 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1321 |
__ASM_EMIT("b.lt 8f") |
| 1322 |
__ASM_EMIT("7:") |
| 1323 |
__ASM_EMIT("ld2r {v0.4s, v1.4s}, [%[dst]]") |
| 1324 |
__ASM_EMIT("ld1r {v16.4s}, [%[src]], #0x04") |
| 1325 |
__ASM_EMIT("fsub v0.4s, v16.4s, v0.4s") |
| 1326 |
__ASM_EMIT("fneg v1.4s, v1.4s") |
| 1327 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1328 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 1329 |
__ASM_EMIT("b.ge 7b") |
| 1330 |
|
| 1331 |
__ASM_EMIT("8:") |
| 1332 |
|
| 1333 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 1334 |
[count] "+r" (count) |
| 1335 |
: |
| 1336 |
: "cc", "memory", |
| 1337 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 1338 |
"v16", "v17", "v18", "v19" |
| 1339 |
); |
| 1340 |
} |
| 1341 |
|
| 1342 |
void pcomplex_r2c_mul2(float *dst, const float *src, size_t count) |
| 1343 |
{ |
| 1344 |
ARCH_AARCH64_ASM |
| 1345 |
( |
| 1346 |
// x16 blocks |
| 1347 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1348 |
__ASM_EMIT("b.lo 2f") |
| 1349 |
__ASM_EMIT("1:") |
| 1350 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1351 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1352 |
__ASM_EMIT("ld2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1353 |
__ASM_EMIT("ld2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1354 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src]], #0x40") |
| 1355 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1356 |
__ASM_EMIT("fmul v2.4s, v2.4s, v17.4s") |
| 1357 |
__ASM_EMIT("fmul v4.4s, v4.4s, v18.4s") |
| 1358 |
__ASM_EMIT("fmul v6.4s, v6.4s, v19.4s") |
| 1359 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1360 |
__ASM_EMIT("fmul v3.4s, v3.4s, v17.4s") |
| 1361 |
__ASM_EMIT("fmul v5.4s, v5.4s, v18.4s") |
| 1362 |
__ASM_EMIT("fmul v7.4s, v7.4s, v19.4s") |
| 1363 |
__ASM_EMIT("sub %[dst], %[dst], #0x80") |
| 1364 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1365 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1366 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1367 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1368 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1369 |
__ASM_EMIT("b.hs 1b") |
| 1370 |
// x8 block |
| 1371 |
__ASM_EMIT("2:") |
| 1372 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 1373 |
__ASM_EMIT("b.lt 4f") |
| 1374 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1375 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1376 |
__ASM_EMIT("ld1 {v16.4s-v17.4s}, [%[src]], #0x20") |
| 1377 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1378 |
__ASM_EMIT("fmul v2.4s, v2.4s, v17.4s") |
| 1379 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1380 |
__ASM_EMIT("fmul v3.4s, v3.4s, v17.4s") |
| 1381 |
__ASM_EMIT("sub %[dst], %[dst], #0x40") |
| 1382 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1383 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1384 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1385 |
// x4 block |
| 1386 |
__ASM_EMIT("4:") |
| 1387 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1388 |
__ASM_EMIT("b.lt 6f") |
| 1389 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]]") |
| 1390 |
__ASM_EMIT("ld1 {v16.4s}, [%[src]], #0x10") |
| 1391 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1392 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1393 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1394 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1395 |
// x1 blocks |
| 1396 |
__ASM_EMIT("6:") |
| 1397 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1398 |
__ASM_EMIT("b.lt 8f") |
| 1399 |
__ASM_EMIT("7:") |
| 1400 |
__ASM_EMIT("ld2r {v0.4s, v1.4s}, [%[dst]]") |
| 1401 |
__ASM_EMIT("ld1r {v16.4s}, [%[src]], #0x04") |
| 1402 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1403 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1404 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1405 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 1406 |
__ASM_EMIT("b.ge 7b") |
| 1407 |
|
| 1408 |
__ASM_EMIT("8:") |
| 1409 |
|
| 1410 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 1411 |
[count] "+r" (count) |
| 1412 |
: |
| 1413 |
: "cc", "memory", |
| 1414 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 1415 |
"v16", "v17", "v18", "v19" |
| 1416 |
); |
| 1417 |
} |
| 1418 |
|
| 1419 |
void pcomplex_r2c_div2(float *dst, const float *src, size_t count) |
| 1420 |
{ |
| 1421 |
ARCH_AARCH64_ASM |
| 1422 |
( |
| 1423 |
// x16 blocks |
| 1424 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1425 |
__ASM_EMIT("b.lo 2f") |
| 1426 |
__ASM_EMIT("1:") |
| 1427 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src]], #0x40") |
| 1428 |
__ASM_EMIT("frecpe v0.4s, v16.4s") /* v0.4s = s2 */ |
| 1429 |
__ASM_EMIT("frecpe v1.4s, v17.4s") |
| 1430 |
__ASM_EMIT("frecpe v2.4s, v18.4s") |
| 1431 |
__ASM_EMIT("frecpe v3.4s, v19.4s") |
| 1432 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2) */ |
| 1433 |
__ASM_EMIT("frecps v5.4s, v1.4s, v17.4s") |
| 1434 |
__ASM_EMIT("frecps v6.4s, v2.4s, v18.4s") |
| 1435 |
__ASM_EMIT("frecps v7.4s, v3.4s, v19.4s") |
| 1436 |
__ASM_EMIT("fmul v0.4s, v4.4s, v0.4s") /* v0.4s = s2' = s2 * (2 - R*s2) */ |
| 1437 |
__ASM_EMIT("fmul v1.4s, v5.4s, v1.4s") |
| 1438 |
__ASM_EMIT("fmul v2.4s, v6.4s, v2.4s") |
| 1439 |
__ASM_EMIT("fmul v3.4s, v7.4s, v3.4s") |
| 1440 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2') */ |
| 1441 |
__ASM_EMIT("frecps v5.4s, v1.4s, v17.4s") |
| 1442 |
__ASM_EMIT("frecps v6.4s, v2.4s, v18.4s") |
| 1443 |
__ASM_EMIT("frecps v7.4s, v3.4s, v19.4s") |
| 1444 |
__ASM_EMIT("fmul v16.4s, v4.4s, v0.4s") /* v16s = s2" = s2' * (2 - R*s2) = 1/s2 */ |
| 1445 |
__ASM_EMIT("fmul v17.4s, v5.4s, v1.4s") |
| 1446 |
__ASM_EMIT("fmul v18.4s, v6.4s, v2.4s") |
| 1447 |
__ASM_EMIT("fmul v19.4s, v7.4s, v3.4s") |
| 1448 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1449 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1450 |
__ASM_EMIT("ld2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1451 |
__ASM_EMIT("ld2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1452 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1453 |
__ASM_EMIT("fmul v2.4s, v2.4s, v17.4s") |
| 1454 |
__ASM_EMIT("fmul v4.4s, v4.4s, v18.4s") |
| 1455 |
__ASM_EMIT("fmul v6.4s, v6.4s, v19.4s") |
| 1456 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1457 |
__ASM_EMIT("fmul v3.4s, v3.4s, v17.4s") |
| 1458 |
__ASM_EMIT("fmul v5.4s, v5.4s, v18.4s") |
| 1459 |
__ASM_EMIT("fmul v7.4s, v7.4s, v19.4s") |
| 1460 |
__ASM_EMIT("sub %[dst], %[dst], #0x80") |
| 1461 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1462 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1463 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1464 |
__ASM_EMIT("st2 {v4.4s, v5.4s}, [%[dst]], #0x20") |
| 1465 |
__ASM_EMIT("st2 {v6.4s, v7.4s}, [%[dst]], #0x20") |
| 1466 |
__ASM_EMIT("b.hs 1b") |
| 1467 |
// x8 block |
| 1468 |
__ASM_EMIT("2:") |
| 1469 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 1470 |
__ASM_EMIT("b.lt 4f") |
| 1471 |
__ASM_EMIT("ld1 {v16.4s-v17.4s}, [%[src]], #0x20") |
| 1472 |
__ASM_EMIT("frecpe v0.4s, v16.4s") /* v0.4s = s2 */ |
| 1473 |
__ASM_EMIT("frecpe v1.4s, v17.4s") |
| 1474 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2) */ |
| 1475 |
__ASM_EMIT("frecps v5.4s, v1.4s, v17.4s") |
| 1476 |
__ASM_EMIT("fmul v0.4s, v4.4s, v0.4s") /* v0.4s = s2' = s2 * (2 - R*s2) */ |
| 1477 |
__ASM_EMIT("fmul v1.4s, v5.4s, v1.4s") |
| 1478 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2') */ |
| 1479 |
__ASM_EMIT("frecps v5.4s, v1.4s, v17.4s") |
| 1480 |
__ASM_EMIT("fmul v16.4s, v4.4s, v0.4s") /* v16s = s2" = s2' * (2 - R*s2) = 1/s2 */ |
| 1481 |
__ASM_EMIT("fmul v17.4s, v5.4s, v1.4s") |
| 1482 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1483 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1484 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1485 |
__ASM_EMIT("fmul v2.4s, v2.4s, v17.4s") |
| 1486 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1487 |
__ASM_EMIT("fmul v3.4s, v3.4s, v17.4s") |
| 1488 |
__ASM_EMIT("sub %[dst], %[dst], #0x40") |
| 1489 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1490 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1491 |
__ASM_EMIT("st2 {v2.4s, v3.4s}, [%[dst]], #0x20") |
| 1492 |
// x4 block |
| 1493 |
__ASM_EMIT("4:") |
| 1494 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1495 |
__ASM_EMIT("b.lt 6f") |
| 1496 |
__ASM_EMIT("ld1 {v16.4s}, [%[src]], #0x10") |
| 1497 |
__ASM_EMIT("frecpe v0.4s, v16.4s") /* v0.4s = s2 */ |
| 1498 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2) */ |
| 1499 |
__ASM_EMIT("fmul v0.4s, v4.4s, v0.4s") /* v0.4s = s2' = s2 * (2 - R*s2) */ |
| 1500 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2') */ |
| 1501 |
__ASM_EMIT("fmul v16.4s, v4.4s, v0.4s") /* v16s = s2" = s2' * (2 - R*s2) = 1/s2 */ |
| 1502 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[dst]]") |
| 1503 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1504 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1505 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1506 |
__ASM_EMIT("st2 {v0.4s, v1.4s}, [%[dst]], #0x20") |
| 1507 |
// x1 blocks |
| 1508 |
__ASM_EMIT("6:") |
| 1509 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1510 |
__ASM_EMIT("b.lt 8f") |
| 1511 |
__ASM_EMIT("7:") |
| 1512 |
__ASM_EMIT("ld1r {v16.4s}, [%[src]], #0x04") |
| 1513 |
__ASM_EMIT("frecpe v0.4s, v16.4s") /* v0.4s = s2 */ |
| 1514 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2) */ |
| 1515 |
__ASM_EMIT("fmul v0.4s, v4.4s, v0.4s") /* v0.4s = s2' = s2 * (2 - R*s2) */ |
| 1516 |
__ASM_EMIT("frecps v4.4s, v0.4s, v16.4s") /* v4.4s = (2 - R*s2') */ |
| 1517 |
__ASM_EMIT("fmul v16.4s, v4.4s, v0.4s") /* v16s = s2" = s2' * (2 - R*s2) = 1/s2 */ |
| 1518 |
__ASM_EMIT("ld2r {v0.4s, v1.4s}, [%[dst]]") |
| 1519 |
__ASM_EMIT("fmul v0.4s, v0.4s, v16.4s") |
| 1520 |
__ASM_EMIT("fmul v1.4s, v1.4s, v16.4s") |
| 1521 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1522 |
__ASM_EMIT("st2 {v0.s, v1.s}[0], [%[dst]], #0x08") |
| 1523 |
__ASM_EMIT("b.ge 7b") |
| 1524 |
|
| 1525 |
__ASM_EMIT("8:") |
| 1526 |
|
| 1527 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 1528 |
[count] "+r" (count) |
| 1529 |
: |
| 1530 |
: "cc", "memory", |
| 1531 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 1532 |
"v16", "v17", "v18", "v19" |
| 1533 |
); |
| 1534 |
} |
| 1535 |
|
| 1536 |
void pcomplex_r2c_rdiv2(float *dst, const float *src, size_t count) |
| 1537 |
{ |
| 1538 |
ARCH_AARCH64_ASM |
| 1539 |
( |
| 1540 |
// x16 blocks |
| 1541 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1542 |
__ASM_EMIT("b.lo 2f") |
| 1543 |
__ASM_EMIT("1:") |
| 1544 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]], #0x20") // v16 = sr1, v17 = si1 |
| 1545 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[dst]], #0x20") // v18 = sr2, v19 = si2 |
| 1546 |
__ASM_EMIT("ld2 {v20.4s, v21.4s}, [%[dst]], #0x20") // v20 = sr3, v21 = si3 |
| 1547 |
__ASM_EMIT("ld2 {v22.4s, v23.4s}, [%[dst]], #0x20") // v22 = sr4, v23 = si4 |
| 1548 |
|
| 1549 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 1550 |
__ASM_EMIT("fmul v25.4s, v18.4s, v18.4s") |
| 1551 |
__ASM_EMIT("fmul v26.4s, v20.4s, v20.4s") |
| 1552 |
__ASM_EMIT("fmul v27.4s, v22.4s, v22.4s") |
| 1553 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 1554 |
__ASM_EMIT("fmla v25.4s, v19.4s, v19.4s") |
| 1555 |
__ASM_EMIT("fmla v26.4s, v21.4s, v21.4s") |
| 1556 |
__ASM_EMIT("fmla v27.4s, v23.4s, v23.4s") |
| 1557 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 1558 |
__ASM_EMIT("fneg v19.4s, v19.4s") |
| 1559 |
__ASM_EMIT("fneg v21.4s, v21.4s") |
| 1560 |
__ASM_EMIT("fneg v23.4s, v23.4s") |
| 1561 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 1562 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v25.4s") |
| 1563 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v26.4s") |
| 1564 |
__ASM_EMIT("fdiv v22.4s, v22.4s, v27.4s") |
| 1565 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 1566 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v25.4s") |
| 1567 |
__ASM_EMIT("fdiv v21.4s, v21.4s, v26.4s") |
| 1568 |
__ASM_EMIT("fdiv v23.4s, v23.4s, v27.4s") |
| 1569 |
|
| 1570 |
__ASM_EMIT("ld1 {v0.4s-v3.4s}, [%[src]], #0x40") |
| 1571 |
__ASM_EMIT("sub %[dst], %[dst], 0x80") |
| 1572 |
__ASM_EMIT("fmul v16.4s, v16.4s, v0.4s") |
| 1573 |
__ASM_EMIT("fmul v18.4s, v18.4s, v1.4s") |
| 1574 |
__ASM_EMIT("fmul v20.4s, v20.4s, v2.4s") |
| 1575 |
__ASM_EMIT("fmul v22.4s, v22.4s, v3.4s") |
| 1576 |
__ASM_EMIT("fmul v17.4s, v17.4s, v0.4s") |
| 1577 |
__ASM_EMIT("fmul v19.4s, v19.4s, v1.4s") |
| 1578 |
__ASM_EMIT("fmul v21.4s, v21.4s, v2.4s") |
| 1579 |
__ASM_EMIT("fmul v23.4s, v23.4s, v3.4s") |
| 1580 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 1581 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 1582 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 1583 |
__ASM_EMIT("st2 {v20.4s, v21.4s}, [%[dst]], #0x20") |
| 1584 |
__ASM_EMIT("st2 {v22.4s, v23.4s}, [%[dst]], #0x20") |
| 1585 |
__ASM_EMIT("b.hs 1b") |
| 1586 |
|
| 1587 |
// x8 block |
| 1588 |
__ASM_EMIT("2:") |
| 1589 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 1590 |
__ASM_EMIT("b.lt 4f") |
| 1591 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]], #0x20") // v16 = sr1, v17 = si1 |
| 1592 |
__ASM_EMIT("ld2 {v18.4s, v19.4s}, [%[dst]], #0x20") // v18 = sr2, v19 = si2 |
| 1593 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 1594 |
__ASM_EMIT("fmul v25.4s, v18.4s, v18.4s") |
| 1595 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 1596 |
__ASM_EMIT("fmla v25.4s, v19.4s, v19.4s") |
| 1597 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 1598 |
__ASM_EMIT("fneg v19.4s, v19.4s") |
| 1599 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 1600 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v25.4s") |
| 1601 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 1602 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v25.4s") |
| 1603 |
__ASM_EMIT("ld1 {v0.4s-v1.4s}, [%[src]], #0x20") |
| 1604 |
__ASM_EMIT("sub %[dst], %[dst], 0x40") |
| 1605 |
__ASM_EMIT("fmul v16.4s, v16.4s, v0.4s") |
| 1606 |
__ASM_EMIT("fmul v18.4s, v18.4s, v1.4s") |
| 1607 |
__ASM_EMIT("fmul v17.4s, v17.4s, v0.4s") |
| 1608 |
__ASM_EMIT("fmul v19.4s, v19.4s, v1.4s") |
| 1609 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 1610 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 1611 |
__ASM_EMIT("st2 {v18.4s, v19.4s}, [%[dst]], #0x20") |
| 1612 |
// x4 block |
| 1613 |
__ASM_EMIT("4:") |
| 1614 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1615 |
__ASM_EMIT("b.lt 6f") |
| 1616 |
__ASM_EMIT("ld2 {v16.4s, v17.4s}, [%[dst]]") // v16 = sr1, v17 = si1 |
| 1617 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 1618 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 1619 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 1620 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 1621 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 1622 |
__ASM_EMIT("ld1 {v0.4s}, [%[src]], #0x10") |
| 1623 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1624 |
__ASM_EMIT("fmul v16.4s, v16.4s, v0.4s") |
| 1625 |
__ASM_EMIT("fmul v17.4s, v17.4s, v0.4s") |
| 1626 |
__ASM_EMIT("st2 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 1627 |
// x1 block |
| 1628 |
__ASM_EMIT("6:") |
| 1629 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1630 |
__ASM_EMIT("b.lt 8f") |
| 1631 |
__ASM_EMIT("7:") |
| 1632 |
__ASM_EMIT("ld2r {v16.4s, v17.4s}, [%[dst]]") // v16 = sr1, v17 = si1 |
| 1633 |
__ASM_EMIT("fmul v24.4s, v16.4s, v16.4s") // v24 = sr*sr |
| 1634 |
__ASM_EMIT("fmla v24.4s, v17.4s, v17.4s") // v24 = sr*dr + si*di |
| 1635 |
__ASM_EMIT("fneg v17.4s, v17.4s") // v17 = -si |
| 1636 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v24.4s") // v16 = sr/(sr*dr + si*di) |
| 1637 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v24.4s") // v17 = -si/(sr*dr + si*di) |
| 1638 |
__ASM_EMIT("ld1r {v0.4s}, [%[src]], #0x04") |
| 1639 |
__ASM_EMIT("fmul v16.4s, v16.4s, v0.4s") |
| 1640 |
__ASM_EMIT("fmul v17.4s, v17.4s, v0.4s") |
| 1641 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1642 |
__ASM_EMIT("st2 {v16.s, v17.s}[0], [%[dst]], #0x08") |
| 1643 |
__ASM_EMIT("b.ge 7b") |
| 1644 |
|
| 1645 |
__ASM_EMIT("8:") |
| 1646 |
|
| 1647 |
: [dst] "+r" (dst), [src] "+r" (src), [count] "+r" (count) |
| 1648 |
: |
| 1649 |
: "cc", "memory", |
| 1650 |
"v0", "v1", "v2", "v3", |
| 1651 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 1652 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 1653 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 1654 |
); |
| 1655 |
} |
| 1656 |
|
| 1657 |
IF_ARCH_AARCH64( |
| 1658 |
static float pcomplex_corr_const[] __lsp_aligned16 = |
| 1659 |
{ |
| 1660 |
LSP_DSP_VEC4(1e-36f), |
| 1661 |
LSP_DSP_VEC4(1e-36f) |
| 1662 |
}; |
| 1663 |
); |
| 1664 |
|
| 1665 |
void pcomplex_corr(float *dst_corr, const float *src1, const float *src2, size_t count) |
| 1666 |
{ |
| 1667 |
/* |
| 1668 |
* src1[i] = a + j*b, src2[i] = c + j*d |
| 1669 |
* den = (a*a + b*b)*(c*c + d*d) |
| 1670 |
* nom = a*c + b*d |
| 1671 |
* corr[i] = (den > threshold) ? nom / sqrt(den) : 0.0 |
| 1672 |
*/ |
| 1673 |
ARCH_AARCH64_ASM |
| 1674 |
( |
| 1675 |
// x8 blocks |
| 1676 |
__ASM_EMIT("ldp q14, q15, [%[CC]]") /* v14-v15 = threshold */ |
| 1677 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 1678 |
__ASM_EMIT("b.lo 2f") |
| 1679 |
__ASM_EMIT("1:") |
| 1680 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[src1]], #0x20") /* v0 = A0, v1 = B0 */ |
| 1681 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[src2]], #0x20") /* v2 = C0, v3 = D0 */ |
| 1682 |
__ASM_EMIT("ld2 {v4.4s, v5.4s}, [%[src1]], #0x20") /* v4 = A1, v5 = B1 */ |
| 1683 |
__ASM_EMIT("ld2 {v6.4s, v7.4s}, [%[src2]], #0x20") /* v6 = C1, v7 = D1 */ |
| 1684 |
__ASM_EMIT("fmul v8.4s, v0.4s, v2.4s") /* v8 = A0*C0 */ |
| 1685 |
__ASM_EMIT("fmul v9.4s, v4.4s, v6.4s") /* v9 = A1*C1 */ |
| 1686 |
__ASM_EMIT("fmul v10.4s, v0.4s, v0.4s") /* v10 = A0*A0 */ |
| 1687 |
__ASM_EMIT("fmul v11.4s, v4.4s, v4.4s") /* v11 = A1*A1 */ |
| 1688 |
__ASM_EMIT("fmul v12.4s, v2.4s, v2.4s") /* v12 = C0*C0 */ |
| 1689 |
__ASM_EMIT("fmul v13.4s, v6.4s, v6.4s") /* v13 = C1*C1 */ |
| 1690 |
__ASM_EMIT("fmla v8.4s, v1.4s, v3.4s") /* v8 = nom0 = A0*C0 + B0*D0 */ |
| 1691 |
__ASM_EMIT("fmla v9.4s, v5.4s, v7.4s") /* v9 = nom1 = A1*C1 + B1*D1 */ |
| 1692 |
__ASM_EMIT("fmla v10.4s, v1.4s, v1.4s") /* v10 = A0*A0 + B0*B0 */ |
| 1693 |
__ASM_EMIT("fmla v11.4s, v5.4s, v5.4s") /* v11 = A1*A1 + B1*B1 */ |
| 1694 |
__ASM_EMIT("fmla v12.4s, v3.4s, v3.4s") /* v12 = C0*C0 + D0*D0 */ |
| 1695 |
__ASM_EMIT("fmla v13.4s, v7.4s, v7.4s") /* v13 = C1*C1 + D1*D1 */ |
| 1696 |
__ASM_EMIT("fmul v0.4s, v10.4s, v12.4s") /* v0 = den0 = (A0*A0 + B0*B0)*(C0*C0 + D0*D0) */ |
| 1697 |
__ASM_EMIT("fmul v1.4s, v11.4s, v13.4s") /* v1 = den1 = (A1*A1 + B1*B1)*(C1*C1 + D1*D1) */ |
| 1698 |
__ASM_EMIT("frsqrte v2.4s, v0.4s") /* v2 = x0 */ |
| 1699 |
__ASM_EMIT("frsqrte v3.4s, v1.4s") |
| 1700 |
__ASM_EMIT("fmul v4.4s, v2.4s, v0.4s") /* v4 = R * x0 */ |
| 1701 |
__ASM_EMIT("fmul v5.4s, v3.4s, v1.4s") |
| 1702 |
__ASM_EMIT("frsqrts v6.4s, v4.4s, v2.4s") /* v6 = (3 - R * x0 * x0) / 2 */ |
| 1703 |
__ASM_EMIT("frsqrts v7.4s, v5.4s, v3.4s") |
| 1704 |
__ASM_EMIT("fmul v2.4s, v2.4s, v6.4s") /* v2 = x1 = x0 * (3 - R * x0 * x0) / 2 */ |
| 1705 |
__ASM_EMIT("fmul v3.4s, v3.4s, v7.4s") |
| 1706 |
__ASM_EMIT("fmul v4.4s, v2.4s, v0.4s") /* v4 = R * x1 */ |
| 1707 |
__ASM_EMIT("fmul v5.4s, v3.4s, v1.4s") |
| 1708 |
__ASM_EMIT("frsqrts v6.4s, v4.4s, v2.4s") /* v6 = (3 - R * x1 * x1) / 2 */ |
| 1709 |
__ASM_EMIT("frsqrts v7.4s, v5.4s, v3.4s") |
| 1710 |
__ASM_EMIT("fmul v2.4s, v2.4s, v6.4s") /* v2 = 1/den0 = x1 * (3 - R * x1 * x1) / 2 */ |
| 1711 |
__ASM_EMIT("fmul v3.4s, v3.4s, v7.4s") /* v3 = 1/den1 */ |
| 1712 |
__ASM_EMIT("fcmge v0.4s, v0.4s, v14.4s") /* v0 = (den0 >= threshold) */ |
| 1713 |
__ASM_EMIT("fcmge v1.4s, v1.4s, v15.4s") /* v1 = (den1 >= threshold) */ |
| 1714 |
__ASM_EMIT("fmul v2.4s, v2.4s, v8.4s") /* v2 = nom0/den0 */ |
| 1715 |
__ASM_EMIT("fmul v3.4s, v3.4s, v9.4s") /* v3 = nom1/den1 */ |
| 1716 |
__ASM_EMIT("and v0.16b, v0.16b, v2.16b") /* v0 = (den0 >= threshold) ? nom0/den0 : 0.0f */ |
| 1717 |
__ASM_EMIT("and v1.16b, v1.16b, v3.16b") /* v1 = (den1 >= threshold) ? nom1/den1 : 0.0f */ |
| 1718 |
__ASM_EMIT("stp q0, q1, [%[dst]]") |
| 1719 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 1720 |
__ASM_EMIT("add %[dst], %[dst], #0x20") |
| 1721 |
__ASM_EMIT("b.hs 1b") |
| 1722 |
// x4 block |
| 1723 |
__ASM_EMIT("2:") |
| 1724 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 1725 |
__ASM_EMIT("b.lt 4f") |
| 1726 |
__ASM_EMIT("ld2 {v0.4s, v1.4s}, [%[src1]], #0x20") /* v0 = A0, v1 = B0 */ |
| 1727 |
__ASM_EMIT("ld2 {v2.4s, v3.4s}, [%[src2]], #0x20") /* v2 = C0, v3 = D0 */ |
| 1728 |
__ASM_EMIT("fmul v8.4s, v0.4s, v2.4s") /* v8 = A0*C0 */ |
| 1729 |
__ASM_EMIT("fmul v10.4s, v0.4s, v0.4s") /* v10 = A0*A0 */ |
| 1730 |
__ASM_EMIT("fmul v12.4s, v2.4s, v2.4s") /* v12 = C0*C0 */ |
| 1731 |
__ASM_EMIT("fmla v8.4s, v1.4s, v3.4s") /* v8 = nom0 = A0*C0 + B0*D0 */ |
| 1732 |
__ASM_EMIT("fmla v10.4s, v1.4s, v1.4s") /* v10 = A0*A0 + B0*B0 */ |
| 1733 |
__ASM_EMIT("fmla v12.4s, v3.4s, v3.4s") /* v12 = C0*C0 + D0*D0 */ |
| 1734 |
__ASM_EMIT("fmul v0.4s, v10.4s, v12.4s") /* v0 = den0 = (A0*A0 + B0*B0)*(C0*C0 + D0*D0) */ |
| 1735 |
__ASM_EMIT("frsqrte v2.4s, v0.4s") /* v2 = x0 */ |
| 1736 |
__ASM_EMIT("fmul v4.4s, v2.4s, v0.4s") /* v4 = R * x0 */ |
| 1737 |
__ASM_EMIT("frsqrts v6.4s, v4.4s, v2.4s") /* v6 = (3 - R * x0 * x0) / 2 */ |
| 1738 |
__ASM_EMIT("fmul v2.4s, v2.4s, v6.4s") /* v2 = x1 = x0 * (3 - R * x0 * x0) / 2 */ |
| 1739 |
__ASM_EMIT("fmul v4.4s, v2.4s, v0.4s") /* v4 = R * x1 */ |
| 1740 |
__ASM_EMIT("frsqrts v6.4s, v4.4s, v2.4s") /* v6 = (3 - R * x1 * x1) / 2 */ |
| 1741 |
__ASM_EMIT("fmul v2.4s, v2.4s, v6.4s") /* v2 = 1/den0 = x1 * (3 - R * x1 * x1) / 2 */ |
| 1742 |
__ASM_EMIT("fcmge v0.4s, v0.4s, v14.4s") /* v0 = (den0 >= threshold) */ |
| 1743 |
__ASM_EMIT("fmul v2.4s, v2.4s, v8.4s") /* v2 = nom0/den0 */ |
| 1744 |
__ASM_EMIT("and v0.16b, v0.16b, v2.16b") /* v0 = (den0 >= threshold) ? nom0/den0 : 0.0f */ |
| 1745 |
__ASM_EMIT("str q0, [%[dst]]") |
| 1746 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 1747 |
__ASM_EMIT("add %[dst], %[dst], #0x10") |
| 1748 |
// x1 blocks |
| 1749 |
__ASM_EMIT("4:") |
| 1750 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 1751 |
__ASM_EMIT("b.lt 6f") |
| 1752 |
__ASM_EMIT("5:") |
| 1753 |
__ASM_EMIT("ld2r {v0.4s, v1.4s}, [%[src1]], #0x08") /* v0 = A0, v1 = B0 */ |
| 1754 |
__ASM_EMIT("ld2r {v2.4s, v3.4s}, [%[src2]], #0x08") /* v2 = C0, v3 = D0 */ |
| 1755 |
__ASM_EMIT("fmul v8.4s, v0.4s, v2.4s") /* v8 = A0*C0 */ |
| 1756 |
__ASM_EMIT("fmul v10.4s, v0.4s, v0.4s") /* v10 = A0*A0 */ |
| 1757 |
__ASM_EMIT("fmul v12.4s, v2.4s, v2.4s") /* v12 = C0*C0 */ |
| 1758 |
__ASM_EMIT("fmla v8.4s, v1.4s, v3.4s") /* v8 = nom0 = A0*C0 + B0*D0 */ |
| 1759 |
__ASM_EMIT("fmla v10.4s, v1.4s, v1.4s") /* v10 = A0*A0 + B0*B0 */ |
| 1760 |
__ASM_EMIT("fmla v12.4s, v3.4s, v3.4s") /* v12 = C0*C0 + D0*D0 */ |
| 1761 |
__ASM_EMIT("fmul v0.4s, v10.4s, v12.4s") /* v0 = den0 = (A0*A0 + B0*B0)*(C0*C0 + D0*D0) */ |
| 1762 |
__ASM_EMIT("frsqrte v2.4s, v0.4s") /* v2 = x0 */ |
| 1763 |
__ASM_EMIT("fmul v4.4s, v2.4s, v0.4s") /* v4 = R * x0 */ |
| 1764 |
__ASM_EMIT("frsqrts v6.4s, v4.4s, v2.4s") /* v6 = (3 - R * x0 * x0) / 2 */ |
| 1765 |
__ASM_EMIT("fmul v2.4s, v2.4s, v6.4s") /* v2 = x1 = x0 * (3 - R * x0 * x0) / 2 */ |
| 1766 |
__ASM_EMIT("fmul v4.4s, v2.4s, v0.4s") /* v4 = R * x1 */ |
| 1767 |
__ASM_EMIT("frsqrts v6.4s, v4.4s, v2.4s") /* v6 = (3 - R * x1 * x1) / 2 */ |
| 1768 |
__ASM_EMIT("fmul v2.4s, v2.4s, v6.4s") /* v2 = 1/den0 = x1 * (3 - R * x1 * x1) / 2 */ |
| 1769 |
__ASM_EMIT("fcmge v0.4s, v0.4s, v14.4s") /* v0 = (den0 >= threshold) */ |
| 1770 |
__ASM_EMIT("fmul v2.4s, v2.4s, v8.4s") /* v2 = nom0/den0 */ |
| 1771 |
__ASM_EMIT("and v0.16b, v0.16b, v2.16b") /* v0 = (den0 >= threshold) ? nom0/den0 : 0.0f */ |
| 1772 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 1773 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst]], #0x04") |
| 1774 |
__ASM_EMIT("b.ge 5b") |
| 1775 |
// End |
| 1776 |
__ASM_EMIT("6:") |
| 1777 |
|
| 1778 |
: [dst] "+r" (dst_corr), [src1] "+r" (src1), [src2] "+r" (src2), [count] "+r" (count) |
| 1779 |
: [CC] "r" (&pcomplex_corr_const[0]) |
| 1780 |
: "cc", "memory", |
| 1781 |
"v0", "v1", "v2", "v3", |
| 1782 |
"v4", "v5", "v6", "v7", |
| 1783 |
"v8", "v9", "v10", "v11", |
| 1784 |
"v12", "v13", "v14", "v15" |
| 1785 |
); |
| 1786 |
} |
| 1787 |
|
| 1788 |
} /* namespace asimd */ |
| 1789 |
} /* namespace lsp */ |
| 1790 |
|
| 1791 |
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_PCOMPLEX_H_ */ |