| 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_COMPLEX_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_COMPLEX_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 |
/* |
| 34 |
* v0, v1 = dr |
| 35 |
* v2, v3 = di |
| 36 |
* v16, v17 = ar |
| 37 |
* v18, v19 = ai |
| 38 |
* v20, v21 = br |
| 39 |
* v22, v23 = bi |
| 40 |
*/ |
| 41 |
void complex_mul2(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t count) |
| 42 |
{ |
| 43 |
ARCH_AARCH64_ASM |
| 44 |
( |
| 45 |
// x8 blocks |
| 46 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 47 |
__ASM_EMIT("b.lo 2f") |
| 48 |
__ASM_EMIT("1:") |
| 49 |
__ASM_EMIT("ldp q16, q17, [%[src_re]]") // v16-v19 = ar |
| 50 |
__ASM_EMIT("ldp q18, q19, [%[src_im]]") // v18-v19 = ai |
| 51 |
__ASM_EMIT("ldp q20, q21, [%[dst_re]]") // v20-v21 = br |
| 52 |
__ASM_EMIT("ldp q22, q23, [%[dst_im]]") // v22-v23 = bi |
| 53 |
__ASM_EMIT("fmul v0.4s, v16.4s, v20.4s") // v0 = ar*br |
| 54 |
__ASM_EMIT("fmul v1.4s, v17.4s, v21.4s") |
| 55 |
__ASM_EMIT("add %[src_re], %[src_re], #0x20") |
| 56 |
__ASM_EMIT("fmul v2.4s, v18.4s, v20.4s") // v2 = ai*br |
| 57 |
__ASM_EMIT("fmul v3.4s, v19.4s, v21.4s") |
| 58 |
__ASM_EMIT("add %[src_im], %[src_im], #0x20") |
| 59 |
__ASM_EMIT("fmls v0.4s, v18.4s, v22.4s") // v0 = ar*br - ai*bi |
| 60 |
__ASM_EMIT("fmls v1.4s, v19.4s, v23.4s") |
| 61 |
__ASM_EMIT("fmla v2.4s, v16.4s, v22.4s") // v2 = ai*br + ar*bi |
| 62 |
__ASM_EMIT("fmla v3.4s, v17.4s, v23.4s") |
| 63 |
__ASM_EMIT("stp q0, q1, [%[dst_re]]") |
| 64 |
__ASM_EMIT("stp q2, q3, [%[dst_im]]") |
| 65 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 66 |
__ASM_EMIT("add %[dst_re], %[dst_re], #0x20") |
| 67 |
__ASM_EMIT("add %[dst_im], %[dst_im], #0x20") |
| 68 |
__ASM_EMIT("b.hs 1b") |
| 69 |
__ASM_EMIT("2:") |
| 70 |
// x4 blocks |
| 71 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 72 |
__ASM_EMIT("b.lo 4f") |
| 73 |
__ASM_EMIT("ldr q16, [%[src_re]]") // v16 = ar |
| 74 |
__ASM_EMIT("ldr q18, [%[src_im]]") // v18 = ai |
| 75 |
__ASM_EMIT("ldr q20, [%[dst_re]]") // v20 = br |
| 76 |
__ASM_EMIT("ldr q22, [%[dst_im]]") // v22 = bi |
| 77 |
__ASM_EMIT("fmul v0.4s, v16.4s, v20.4s") // v0 = ar*br |
| 78 |
__ASM_EMIT("add %[src_re], %[src_re], #0x10") |
| 79 |
__ASM_EMIT("fmul v2.4s, v18.4s, v20.4s") // v2 = ai*br |
| 80 |
__ASM_EMIT("add %[src_im], %[src_im], #0x10") |
| 81 |
__ASM_EMIT("fmls v0.4s, v18.4s, v22.4s") // v0 = ar*br - ai*bi |
| 82 |
__ASM_EMIT("fmla v2.4s, v16.4s, v22.4s") // v2 = ai*br + ar*bi |
| 83 |
__ASM_EMIT("str q0, [%[dst_re]]") |
| 84 |
__ASM_EMIT("str q2, [%[dst_im]]") |
| 85 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 86 |
__ASM_EMIT("add %[dst_re], %[dst_re], #0x10") |
| 87 |
__ASM_EMIT("add %[dst_im], %[dst_im], #0x10") |
| 88 |
__ASM_EMIT("4:") |
| 89 |
// x1 blocks |
| 90 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 91 |
__ASM_EMIT("b.lt 6f") |
| 92 |
__ASM_EMIT("5:") |
| 93 |
__ASM_EMIT("ld1r {v16.4s}, [%[src_re]]") // v16 = ar |
| 94 |
__ASM_EMIT("ld1r {v18.4s}, [%[src_im]]") // v18 = ai |
| 95 |
__ASM_EMIT("ld1r {v20.4s}, [%[dst_re]]") // v20 = br |
| 96 |
__ASM_EMIT("ld1r {v22.4s}, [%[dst_im]]") // v22 = bi |
| 97 |
__ASM_EMIT("fmul v0.4s, v16.4s, v20.4s") // v0 = ar*br |
| 98 |
__ASM_EMIT("add %[src_re], %[src_re], #0x04") |
| 99 |
__ASM_EMIT("fmul v2.4s, v18.4s, v20.4s") // v2 = ai*br |
| 100 |
__ASM_EMIT("add %[src_im], %[src_im], #0x04") |
| 101 |
__ASM_EMIT("fmls v0.4s, v18.4s, v22.4s") // v0 = ar*br - ai*bi |
| 102 |
__ASM_EMIT("fmla v2.4s, v16.4s, v22.4s") // v2 = ai*br + ar*bi |
| 103 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst_re]]") |
| 104 |
__ASM_EMIT("st1 {v2.s}[0], [%[dst_im]]") |
| 105 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 106 |
__ASM_EMIT("add %[dst_re], %[dst_re], #0x04") |
| 107 |
__ASM_EMIT("add %[dst_im], %[dst_im], #0x04") |
| 108 |
__ASM_EMIT("b.ge 5b") |
| 109 |
__ASM_EMIT("6:") |
| 110 |
/* end */ |
| 111 |
: [dst_re] "+r" (dst_re), [dst_im] "+r" (dst_im), |
| 112 |
[src_re] "+r" (src_re), [src_im] "+r" (src_im), |
| 113 |
[count] "+r" (count) |
| 114 |
: |
| 115 |
: "cc", "memory", |
| 116 |
"v0", "v1", |
| 117 |
"v16", "v17", "v18", "v19", |
| 118 |
"v20", "v21", "v22", "v23" |
| 119 |
); |
| 120 |
} |
| 121 |
|
| 122 |
void complex_mul3(float *dst_re, float *dst_im, const float *src1_re, const float *src1_im, const float *src2_re, const float *src2_im, size_t count) |
| 123 |
{ |
| 124 |
ARCH_AARCH64_ASM |
| 125 |
( |
| 126 |
// x16 blocks |
| 127 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 128 |
__ASM_EMIT("b.lo 2f") |
| 129 |
__ASM_EMIT("1:") |
| 130 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src1_re]]") // v16-v19 = sr |
| 131 |
__ASM_EMIT("ld1 {v20.4s-v23.4s}, [%[src1_im]]") // v20-v23 = si |
| 132 |
__ASM_EMIT("ld1 {v24.4s-v27.4s}, [%[src2_re]]") // v24-v27 = dr |
| 133 |
__ASM_EMIT("ld1 {v28.4s-v31.4s}, [%[src2_im]]") // v28-v31 = di |
| 134 |
__ASM_EMIT("add %[src1_re], %[src1_re], #0x40") |
| 135 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 136 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 137 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 138 |
__ASM_EMIT("fmul v3.4s, v19.4s, v27.4s") |
| 139 |
__ASM_EMIT("add %[src1_im], %[src1_im], #0x40") |
| 140 |
__ASM_EMIT("fmul v24.4s, v24.4s, v20.4s") // v24 = si*dr |
| 141 |
__ASM_EMIT("fmul v25.4s, v25.4s, v21.4s") |
| 142 |
__ASM_EMIT("fmul v26.4s, v26.4s, v22.4s") |
| 143 |
__ASM_EMIT("fmul v27.4s, v27.4s, v23.4s") |
| 144 |
__ASM_EMIT("add %[src2_re], %[src2_re], #0x40") |
| 145 |
__ASM_EMIT("fmls v0.4s, v20.4s, v28.4s") // v0 = sr*dr - si*di |
| 146 |
__ASM_EMIT("fmls v1.4s, v21.4s, v29.4s") |
| 147 |
__ASM_EMIT("fmls v2.4s, v22.4s, v30.4s") |
| 148 |
__ASM_EMIT("fmls v3.4s, v23.4s, v31.4s") |
| 149 |
__ASM_EMIT("add %[src2_im], %[src2_im], #0x40") |
| 150 |
__ASM_EMIT("fmla v24.4s, v16.4s, v28.4s") // v24 = si*dr + sr*di |
| 151 |
__ASM_EMIT("fmla v25.4s, v17.4s, v29.4s") |
| 152 |
__ASM_EMIT("fmla v26.4s, v18.4s, v30.4s") |
| 153 |
__ASM_EMIT("fmla v27.4s, v19.4s, v31.4s") |
| 154 |
__ASM_EMIT("st1 {v0.4s-v3.4s}, [%[dst_re]]") |
| 155 |
__ASM_EMIT("st1 {v24.4s-v27.4s}, [%[dst_im]]") |
| 156 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 157 |
__ASM_EMIT("add %[dst_re], %[dst_re], #0x40") |
| 158 |
__ASM_EMIT("add %[dst_im], %[dst_im], #0x40") |
| 159 |
__ASM_EMIT("b.hs 1b") |
| 160 |
__ASM_EMIT("2:") |
| 161 |
// x8 blocks |
| 162 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 163 |
__ASM_EMIT("b.lt 4f") |
| 164 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[src1_re]]") // v16-v17 = sr |
| 165 |
__ASM_EMIT("ld1 {v20.4s, v21.4s}, [%[src1_im]]") // v20-v21 = si |
| 166 |
__ASM_EMIT("ld1 {v24.4s, v25.4s}, [%[src2_re]]") // v24-v25 = dr |
| 167 |
__ASM_EMIT("ld1 {v28.4s, v29.4s}, [%[src2_im]]") // v28-v29 = di |
| 168 |
__ASM_EMIT("add %[src1_re], %[src1_re], #0x20") |
| 169 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 170 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 171 |
__ASM_EMIT("add %[src1_im], %[src1_im], #0x20") |
| 172 |
__ASM_EMIT("fmul v24.4s, v24.4s, v20.4s") // v24 = si*dr |
| 173 |
__ASM_EMIT("fmul v25.4s, v25.4s, v21.4s") |
| 174 |
__ASM_EMIT("add %[src2_re], %[src2_re], #0x20") |
| 175 |
__ASM_EMIT("fmls v0.4s, v20.4s, v28.4s") // v0 = sr*dr - si*di |
| 176 |
__ASM_EMIT("fmls v1.4s, v21.4s, v29.4s") |
| 177 |
__ASM_EMIT("add %[src2_im], %[src2_im], #0x20") |
| 178 |
__ASM_EMIT("fmla v24.4s, v16.4s, v28.4s") // v24 = si*dr + sr*di |
| 179 |
__ASM_EMIT("fmla v25.4s, v17.4s, v29.4s") |
| 180 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 181 |
__ASM_EMIT("st1 {v0.4s, v1.4s}, [%[dst_re]]") |
| 182 |
__ASM_EMIT("st1 {v24.4s, v25.4s}, [%[dst_im]]") |
| 183 |
__ASM_EMIT("add %[dst_re], %[dst_re], #0x20") |
| 184 |
__ASM_EMIT("add %[dst_im], %[dst_im], #0x20") |
| 185 |
__ASM_EMIT("4:") |
| 186 |
// x4 blocks |
| 187 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 188 |
__ASM_EMIT("b.lt 6f") |
| 189 |
__ASM_EMIT("ld1 {v16.4s}, [%[src1_re]]") // v16-v17 = sr |
| 190 |
__ASM_EMIT("ld1 {v20.4s}, [%[src1_im]]") // v20-v21 = si |
| 191 |
__ASM_EMIT("ld1 {v24.4s}, [%[src2_re]]") // v24-v25 = dr |
| 192 |
__ASM_EMIT("ld1 {v28.4s}, [%[src2_im]]") // v28-v29 = di |
| 193 |
__ASM_EMIT("add %[src1_re], %[src1_re], #0x10") |
| 194 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 195 |
__ASM_EMIT("add %[src1_im], %[src1_im], #0x10") |
| 196 |
__ASM_EMIT("fmul v24.4s, v24.4s, v20.4s") // v24 = si*dr |
| 197 |
__ASM_EMIT("add %[src2_re], %[src2_re], #0x10") |
| 198 |
__ASM_EMIT("fmls v0.4s, v20.4s, v28.4s") // v0 = sr*dr - si*di |
| 199 |
__ASM_EMIT("add %[src2_im], %[src2_im], #0x10") |
| 200 |
__ASM_EMIT("fmla v24.4s, v16.4s, v28.4s") // v24 = si*dr + sr*di |
| 201 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 202 |
__ASM_EMIT("st1 {v0.4s}, [%[dst_re]]") |
| 203 |
__ASM_EMIT("st1 {v24.4s}, [%[dst_im]]") |
| 204 |
__ASM_EMIT("add %[dst_re], %[dst_re], #0x10") |
| 205 |
__ASM_EMIT("add %[dst_im], %[dst_im], #0x10") |
| 206 |
__ASM_EMIT("6:") |
| 207 |
// x1 blocks |
| 208 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 209 |
__ASM_EMIT("b.lt 8f") |
| 210 |
__ASM_EMIT("7:") |
| 211 |
__ASM_EMIT("ld1r {v16.4s}, [%[src1_re]]") // v16 = sr |
| 212 |
__ASM_EMIT("ld1r {v20.4s}, [%[src1_im]]") // v20 = si |
| 213 |
__ASM_EMIT("ld1r {v24.4s}, [%[src2_re]]") // v24 = dr |
| 214 |
__ASM_EMIT("ld1r {v28.4s}, [%[src2_im]]") // v28 = di |
| 215 |
__ASM_EMIT("add %[src1_re], %[src1_re], #0x04") |
| 216 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 217 |
__ASM_EMIT("add %[src1_im], %[src1_im], #0x04") |
| 218 |
__ASM_EMIT("fmul v24.4s, v24.4s, v20.4s") // v24 = si*dr |
| 219 |
__ASM_EMIT("add %[src2_re], %[src2_re], #0x04") |
| 220 |
__ASM_EMIT("fmls v0.4s, v20.4s, v28.4s") // v0 = sr*dr - si*di |
| 221 |
__ASM_EMIT("add %[src2_im], %[src2_im], #0x04") |
| 222 |
__ASM_EMIT("fmla v24.4s, v16.4s, v28.4s") // v24 = si*dr + sr*di |
| 223 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst_re]]") |
| 224 |
__ASM_EMIT("st1 {v24.s}[0], [%[dst_im]]") |
| 225 |
__ASM_EMIT("add %[dst_re], %[dst_re], #0x04") |
| 226 |
__ASM_EMIT("add %[dst_im], %[dst_im], #0x04") |
| 227 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 228 |
__ASM_EMIT("b.ge 7b") |
| 229 |
__ASM_EMIT("8:") |
| 230 |
|
| 231 |
: [dst_re] "+r" (dst_re), [dst_im] "+r" (dst_im), |
| 232 |
[src1_re] "+r" (src1_re), [src1_im] "+r" (src1_im), |
| 233 |
[src2_re] "+r" (src2_re), [src2_im] "+r" (src2_im), |
| 234 |
[count] "+r" (count) |
| 235 |
: |
| 236 |
: "cc", "memory", |
| 237 |
"v0", "v1", "v2", "v3", |
| 238 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 239 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 240 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 241 |
); |
| 242 |
} |
| 243 |
|
| 244 |
static const float complex_div_const[] __lsp_aligned16 = |
| 245 |
{ |
| 246 |
LSP_DSP_VEC4(1.0f) |
| 247 |
}; |
| 248 |
|
| 249 |
void complex_div2(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t count) |
| 250 |
{ |
| 251 |
ARCH_AARCH64_ASM |
| 252 |
( |
| 253 |
// x16 blocks |
| 254 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 255 |
__ASM_EMIT("ldr q8, [%[kdiv]]") |
| 256 |
__ASM_EMIT("b.lo 2f") |
| 257 |
__ASM_EMIT("1:") |
| 258 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src_re]], #0x40") // v16-v19 = sr |
| 259 |
__ASM_EMIT("ld1 {v20.4s-v23.4s}, [%[src_im]], #0x40") // v20-v23 = si |
| 260 |
__ASM_EMIT("ld1 {v24.4s-v27.4s}, [%[dst_re]]") // v24-v27 = dr |
| 261 |
__ASM_EMIT("ld1 {v28.4s-v31.4s}, [%[dst_im]]") // v28-v31 = di |
| 262 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 263 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 264 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 265 |
__ASM_EMIT("fmul v3.4s, v19.4s, v27.4s") |
| 266 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 267 |
__ASM_EMIT("fmul v5.4s, v17.4s, v29.4s") |
| 268 |
__ASM_EMIT("fmul v6.4s, v18.4s, v30.4s") |
| 269 |
__ASM_EMIT("fmul v7.4s, v19.4s, v31.4s") |
| 270 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 271 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 272 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 273 |
__ASM_EMIT("fmul v19.4s, v19.4s, v19.4s") |
| 274 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 275 |
__ASM_EMIT("fmla v1.4s, v21.4s, v29.4s") |
| 276 |
__ASM_EMIT("fmla v2.4s, v22.4s, v30.4s") |
| 277 |
__ASM_EMIT("fmla v3.4s, v23.4s, v31.4s") |
| 278 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 279 |
__ASM_EMIT("fmla v17.4s, v21.4s, v21.4s") |
| 280 |
__ASM_EMIT("fmla v18.4s, v22.4s, v22.4s") |
| 281 |
__ASM_EMIT("fmla v19.4s, v23.4s, v23.4s") |
| 282 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 283 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v17.4s") |
| 284 |
__ASM_EMIT("fdiv v14.4s, v8.4s, v18.4s") |
| 285 |
__ASM_EMIT("fdiv v15.4s, v8.4s, v19.4s") |
| 286 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 287 |
__ASM_EMIT("fmls v5.4s, v21.4s, v25.4s") |
| 288 |
__ASM_EMIT("fmls v6.4s, v22.4s, v26.4s") |
| 289 |
__ASM_EMIT("fmls v7.4s, v23.4s, v27.4s") |
| 290 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 291 |
__ASM_EMIT("fmul v1.4s, v1.4s, v13.4s") |
| 292 |
__ASM_EMIT("fmul v2.4s, v2.4s, v14.4s") |
| 293 |
__ASM_EMIT("fmul v3.4s, v3.4s, v15.4s") |
| 294 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 295 |
__ASM_EMIT("fmul v5.4s, v5.4s, v13.4s") |
| 296 |
__ASM_EMIT("fmul v6.4s, v6.4s, v14.4s") |
| 297 |
__ASM_EMIT("fmul v7.4s, v7.4s, v15.4s") |
| 298 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 299 |
__ASM_EMIT("st1 {v0.4s-v3.4s}, [%[dst_re]], #0x40") |
| 300 |
__ASM_EMIT("st1 {v4.4s-v7.4s}, [%[dst_im]], #0x40") |
| 301 |
__ASM_EMIT("b.hs 1b") |
| 302 |
__ASM_EMIT("2:") |
| 303 |
// x8 blocks |
| 304 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 305 |
__ASM_EMIT("b.lt 4f") |
| 306 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[src_re]], #0x20") // v16-v17 = sr |
| 307 |
__ASM_EMIT("ld1 {v20.4s, v21.4s}, [%[src_im]], #0x20") // v20-v21 = si |
| 308 |
__ASM_EMIT("ld1 {v24.4s, v25.4s}, [%[dst_re]]") // v24-v25 = dr |
| 309 |
__ASM_EMIT("ld1 {v28.4s, v29.4s}, [%[dst_im]]") // v28-v29 = di |
| 310 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 311 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 312 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 313 |
__ASM_EMIT("fmul v5.4s, v17.4s, v29.4s") |
| 314 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 315 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 316 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 317 |
__ASM_EMIT("fmla v1.4s, v21.4s, v29.4s") |
| 318 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 319 |
__ASM_EMIT("fmla v17.4s, v21.4s, v21.4s") |
| 320 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 321 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v17.4s") |
| 322 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 323 |
__ASM_EMIT("fmls v5.4s, v21.4s, v25.4s") |
| 324 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 325 |
__ASM_EMIT("fmul v1.4s, v1.4s, v13.4s") |
| 326 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 327 |
__ASM_EMIT("fmul v5.4s, v5.4s, v13.4s") |
| 328 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 329 |
__ASM_EMIT("st1 {v0.4s, v1.4s}, [%[dst_re]], #0x20") |
| 330 |
__ASM_EMIT("st1 {v4.4s, v5.4s}, [%[dst_im]], #0x20") |
| 331 |
__ASM_EMIT("4:") |
| 332 |
// x4 blocks |
| 333 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 334 |
__ASM_EMIT("b.lt 6f") |
| 335 |
__ASM_EMIT("ld1 {v16.4s}, [%[src_re]], #0x10") // v16-v17 = sr |
| 336 |
__ASM_EMIT("ld1 {v20.4s}, [%[src_im]], #0x10") // v20-v21 = si |
| 337 |
__ASM_EMIT("ld1 {v24.4s}, [%[dst_re]]") // v24-v25 = dr |
| 338 |
__ASM_EMIT("ld1 {v28.4s}, [%[dst_im]]") // v28-v29 = di |
| 339 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 340 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 341 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 342 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 343 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 344 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 345 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 346 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 347 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 348 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 349 |
__ASM_EMIT("st1 {v0.4s}, [%[dst_re]], #0x10") |
| 350 |
__ASM_EMIT("st1 {v4.4s}, [%[dst_im]], #0x10") |
| 351 |
__ASM_EMIT("6:") |
| 352 |
// x1 blocks |
| 353 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 354 |
__ASM_EMIT("b.lt 8f") |
| 355 |
__ASM_EMIT("7:") |
| 356 |
__ASM_EMIT("ld1r {v16.4s}, [%[src_re]], #0x04") // v16 = sr |
| 357 |
__ASM_EMIT("ld1r {v20.4s}, [%[src_im]], #0x04") // v20 = si |
| 358 |
__ASM_EMIT("ld1r {v24.4s}, [%[dst_re]]") // v24 = dr |
| 359 |
__ASM_EMIT("ld1r {v28.4s}, [%[dst_im]]") // v28 = di |
| 360 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 361 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 362 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 363 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 364 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 365 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 366 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 367 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 368 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 369 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst_re]], #0x04") |
| 370 |
__ASM_EMIT("st1 {v4.s}[0], [%[dst_im]], #0x04") |
| 371 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 372 |
__ASM_EMIT("b.ge 7b") |
| 373 |
__ASM_EMIT("8:") |
| 374 |
|
| 375 |
: [dst_re] "+r" (dst_re), [dst_im] "+r" (dst_im), |
| 376 |
[src_re] "+r" (src_re), [src_im] "+r" (src_im), |
| 377 |
[count] "+r" (count) |
| 378 |
: [kdiv] "r" (&complex_div_const[0]) |
| 379 |
: "cc", "memory", |
| 380 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 381 |
"v8", /* "v9", "v10", "v11", */ "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 382 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 383 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 384 |
); |
| 385 |
} |
| 386 |
|
| 387 |
void complex_rdiv2(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t count) |
| 388 |
{ |
| 389 |
ARCH_AARCH64_ASM |
| 390 |
( |
| 391 |
// x16 blocks |
| 392 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 393 |
__ASM_EMIT("ldr q8, [%[kdiv]]") |
| 394 |
__ASM_EMIT("b.lo 2f") |
| 395 |
__ASM_EMIT("1:") |
| 396 |
__ASM_EMIT("ld1 {v24.4s-v27.4s}, [%[src_re]], #0x40") // v24-v27 = sr |
| 397 |
__ASM_EMIT("ld1 {v28.4s-v31.4s}, [%[src_im]], #0x40") // v28-v31 = si |
| 398 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[dst_re]]") // v16-v19 = dr |
| 399 |
__ASM_EMIT("ld1 {v20.4s-v23.4s}, [%[dst_im]]") // v20-v23 = di |
| 400 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 401 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 402 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 403 |
__ASM_EMIT("fmul v3.4s, v19.4s, v27.4s") |
| 404 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 405 |
__ASM_EMIT("fmul v5.4s, v17.4s, v29.4s") |
| 406 |
__ASM_EMIT("fmul v6.4s, v18.4s, v30.4s") |
| 407 |
__ASM_EMIT("fmul v7.4s, v19.4s, v31.4s") |
| 408 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 409 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 410 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 411 |
__ASM_EMIT("fmul v19.4s, v19.4s, v19.4s") |
| 412 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 413 |
__ASM_EMIT("fmla v1.4s, v21.4s, v29.4s") |
| 414 |
__ASM_EMIT("fmla v2.4s, v22.4s, v30.4s") |
| 415 |
__ASM_EMIT("fmla v3.4s, v23.4s, v31.4s") |
| 416 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 417 |
__ASM_EMIT("fmla v17.4s, v21.4s, v21.4s") |
| 418 |
__ASM_EMIT("fmla v18.4s, v22.4s, v22.4s") |
| 419 |
__ASM_EMIT("fmla v19.4s, v23.4s, v23.4s") |
| 420 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 421 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v17.4s") |
| 422 |
__ASM_EMIT("fdiv v14.4s, v8.4s, v18.4s") |
| 423 |
__ASM_EMIT("fdiv v15.4s, v8.4s, v19.4s") |
| 424 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 425 |
__ASM_EMIT("fmls v5.4s, v21.4s, v25.4s") |
| 426 |
__ASM_EMIT("fmls v6.4s, v22.4s, v26.4s") |
| 427 |
__ASM_EMIT("fmls v7.4s, v23.4s, v27.4s") |
| 428 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 429 |
__ASM_EMIT("fmul v1.4s, v1.4s, v13.4s") |
| 430 |
__ASM_EMIT("fmul v2.4s, v2.4s, v14.4s") |
| 431 |
__ASM_EMIT("fmul v3.4s, v3.4s, v15.4s") |
| 432 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 433 |
__ASM_EMIT("fmul v5.4s, v5.4s, v13.4s") |
| 434 |
__ASM_EMIT("fmul v6.4s, v6.4s, v14.4s") |
| 435 |
__ASM_EMIT("fmul v7.4s, v7.4s, v15.4s") |
| 436 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 437 |
__ASM_EMIT("st1 {v0.4s-v3.4s}, [%[dst_re]], #0x40") |
| 438 |
__ASM_EMIT("st1 {v4.4s-v7.4s}, [%[dst_im]], #0x40") |
| 439 |
__ASM_EMIT("b.hs 1b") |
| 440 |
__ASM_EMIT("2:") |
| 441 |
// x8 blocks |
| 442 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 443 |
__ASM_EMIT("b.lt 4f") |
| 444 |
__ASM_EMIT("ld1 {v24.4s, v25.4s}, [%[src_re]], #0x20") // v24-v25 = sr |
| 445 |
__ASM_EMIT("ld1 {v28.4s, v29.4s}, [%[src_im]], #0x20") // v28-v29 = si |
| 446 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[dst_re]]") // v16-v17 = dr |
| 447 |
__ASM_EMIT("ld1 {v20.4s, v21.4s}, [%[dst_im]]") // v20-v21 = di |
| 448 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 449 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 450 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 451 |
__ASM_EMIT("fmul v5.4s, v17.4s, v29.4s") |
| 452 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 453 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 454 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 455 |
__ASM_EMIT("fmla v1.4s, v21.4s, v29.4s") |
| 456 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 457 |
__ASM_EMIT("fmla v17.4s, v21.4s, v21.4s") |
| 458 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 459 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v17.4s") |
| 460 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 461 |
__ASM_EMIT("fmls v5.4s, v21.4s, v25.4s") |
| 462 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 463 |
__ASM_EMIT("fmul v1.4s, v1.4s, v13.4s") |
| 464 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 465 |
__ASM_EMIT("fmul v5.4s, v5.4s, v13.4s") |
| 466 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 467 |
__ASM_EMIT("st1 {v0.4s, v1.4s}, [%[dst_re]], #0x20") |
| 468 |
__ASM_EMIT("st1 {v4.4s, v5.4s}, [%[dst_im]], #0x20") |
| 469 |
__ASM_EMIT("4:") |
| 470 |
// x4 blocks |
| 471 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 472 |
__ASM_EMIT("b.lt 6f") |
| 473 |
__ASM_EMIT("ld1 {v24.4s}, [%[src_re]], #0x10") // v24-v25 = sr |
| 474 |
__ASM_EMIT("ld1 {v28.4s}, [%[src_im]], #0x10") // v28-v29 = si |
| 475 |
__ASM_EMIT("ld1 {v16.4s}, [%[dst_re]]") // v16-v17 = dr |
| 476 |
__ASM_EMIT("ld1 {v20.4s}, [%[dst_im]]") // v20-v21 = di |
| 477 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 478 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 479 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 480 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 481 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 482 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 483 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 484 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 485 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 486 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 487 |
__ASM_EMIT("st1 {v0.4s}, [%[dst_re]], #0x10") |
| 488 |
__ASM_EMIT("st1 {v4.4s}, [%[dst_im]], #0x10") |
| 489 |
__ASM_EMIT("6:") |
| 490 |
// x1 blocks |
| 491 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 492 |
__ASM_EMIT("b.lt 8f") |
| 493 |
__ASM_EMIT("7:") |
| 494 |
__ASM_EMIT("ld1r {v24.4s}, [%[src_re]], #0x04") // v24 = sr |
| 495 |
__ASM_EMIT("ld1r {v28.4s}, [%[src_im]], #0x04") // v28 = si |
| 496 |
__ASM_EMIT("ld1r {v16.4s}, [%[dst_re]]") // v16 = dr |
| 497 |
__ASM_EMIT("ld1r {v20.4s}, [%[dst_im]]") // v20 = di |
| 498 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 499 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 500 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 501 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 502 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 503 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 504 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 505 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 506 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 507 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst_re]], #0x04") |
| 508 |
__ASM_EMIT("st1 {v4.s}[0], [%[dst_im]], #0x04") |
| 509 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 510 |
__ASM_EMIT("b.ge 7b") |
| 511 |
__ASM_EMIT("8:") |
| 512 |
|
| 513 |
: [dst_re] "+r" (dst_re), [dst_im] "+r" (dst_im), |
| 514 |
[src_re] "+r" (src_re), [src_im] "+r" (src_im), |
| 515 |
[count] "+r" (count) |
| 516 |
: [kdiv] "r" (&complex_div_const[0]) |
| 517 |
: "cc", "memory", |
| 518 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 519 |
"v8", /* "v9", "v10", "v11", */ "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 520 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 521 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 522 |
); |
| 523 |
} |
| 524 |
|
| 525 |
void complex_div3(float *dst_re, float *dst_im, const float *t_re, const float *t_im, const float *b_re, const float *b_im, size_t count) |
| 526 |
{ |
| 527 |
ARCH_AARCH64_ASM |
| 528 |
( |
| 529 |
// x16 blocks |
| 530 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 531 |
__ASM_EMIT("ldr q8, [%[kdiv]]") |
| 532 |
__ASM_EMIT("b.lo 2f") |
| 533 |
__ASM_EMIT("1:") |
| 534 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[b_re]], #0x40") // v16-v19 = sr |
| 535 |
__ASM_EMIT("ld1 {v20.4s-v23.4s}, [%[b_im]], #0x40") // v20-v23 = si |
| 536 |
__ASM_EMIT("ld1 {v24.4s-v27.4s}, [%[t_re]], #0x40") // v24-v27 = dr |
| 537 |
__ASM_EMIT("ld1 {v28.4s-v31.4s}, [%[t_im]], #0x40") // v28-v31 = di |
| 538 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 539 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 540 |
__ASM_EMIT("fmul v2.4s, v18.4s, v26.4s") |
| 541 |
__ASM_EMIT("fmul v3.4s, v19.4s, v27.4s") |
| 542 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 543 |
__ASM_EMIT("fmul v5.4s, v17.4s, v29.4s") |
| 544 |
__ASM_EMIT("fmul v6.4s, v18.4s, v30.4s") |
| 545 |
__ASM_EMIT("fmul v7.4s, v19.4s, v31.4s") |
| 546 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 547 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 548 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 549 |
__ASM_EMIT("fmul v19.4s, v19.4s, v19.4s") |
| 550 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 551 |
__ASM_EMIT("fmla v1.4s, v21.4s, v29.4s") |
| 552 |
__ASM_EMIT("fmla v2.4s, v22.4s, v30.4s") |
| 553 |
__ASM_EMIT("fmla v3.4s, v23.4s, v31.4s") |
| 554 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 555 |
__ASM_EMIT("fmla v17.4s, v21.4s, v21.4s") |
| 556 |
__ASM_EMIT("fmla v18.4s, v22.4s, v22.4s") |
| 557 |
__ASM_EMIT("fmla v19.4s, v23.4s, v23.4s") |
| 558 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 559 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v17.4s") |
| 560 |
__ASM_EMIT("fdiv v14.4s, v8.4s, v18.4s") |
| 561 |
__ASM_EMIT("fdiv v15.4s, v8.4s, v19.4s") |
| 562 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 563 |
__ASM_EMIT("fmls v5.4s, v21.4s, v25.4s") |
| 564 |
__ASM_EMIT("fmls v6.4s, v22.4s, v26.4s") |
| 565 |
__ASM_EMIT("fmls v7.4s, v23.4s, v27.4s") |
| 566 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 567 |
__ASM_EMIT("fmul v1.4s, v1.4s, v13.4s") |
| 568 |
__ASM_EMIT("fmul v2.4s, v2.4s, v14.4s") |
| 569 |
__ASM_EMIT("fmul v3.4s, v3.4s, v15.4s") |
| 570 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 571 |
__ASM_EMIT("fmul v5.4s, v5.4s, v13.4s") |
| 572 |
__ASM_EMIT("fmul v6.4s, v6.4s, v14.4s") |
| 573 |
__ASM_EMIT("fmul v7.4s, v7.4s, v15.4s") |
| 574 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 575 |
__ASM_EMIT("st1 {v0.4s-v3.4s}, [%[dst_re]], #0x40") |
| 576 |
__ASM_EMIT("st1 {v4.4s-v7.4s}, [%[dst_im]], #0x40") |
| 577 |
__ASM_EMIT("b.hs 1b") |
| 578 |
__ASM_EMIT("2:") |
| 579 |
// x8 blocks |
| 580 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 581 |
__ASM_EMIT("b.lt 4f") |
| 582 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[b_re]], #0x20") // v16-v17 = sr |
| 583 |
__ASM_EMIT("ld1 {v20.4s, v21.4s}, [%[b_im]], #0x20") // v20-v21 = si |
| 584 |
__ASM_EMIT("ld1 {v24.4s, v25.4s}, [%[t_re]], #0x20") // v24-v25 = dr |
| 585 |
__ASM_EMIT("ld1 {v28.4s, v29.4s}, [%[t_im]], #0x20") // v28-v29 = di |
| 586 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 587 |
__ASM_EMIT("fmul v1.4s, v17.4s, v25.4s") |
| 588 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 589 |
__ASM_EMIT("fmul v5.4s, v17.4s, v29.4s") |
| 590 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 591 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 592 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 593 |
__ASM_EMIT("fmla v1.4s, v21.4s, v29.4s") |
| 594 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 595 |
__ASM_EMIT("fmla v17.4s, v21.4s, v21.4s") |
| 596 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 597 |
__ASM_EMIT("fdiv v13.4s, v8.4s, v17.4s") |
| 598 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 599 |
__ASM_EMIT("fmls v5.4s, v21.4s, v25.4s") |
| 600 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 601 |
__ASM_EMIT("fmul v1.4s, v1.4s, v13.4s") |
| 602 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 603 |
__ASM_EMIT("fmul v5.4s, v5.4s, v13.4s") |
| 604 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 605 |
__ASM_EMIT("st1 {v0.4s, v1.4s}, [%[dst_re]], #0x20") |
| 606 |
__ASM_EMIT("st1 {v4.4s, v5.4s}, [%[dst_im]], #0x20") |
| 607 |
__ASM_EMIT("4:") |
| 608 |
// x4 blocks |
| 609 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 610 |
__ASM_EMIT("b.lt 6f") |
| 611 |
__ASM_EMIT("ld1 {v16.4s}, [%[b_re]], #0x10") // v16-v17 = sr |
| 612 |
__ASM_EMIT("ld1 {v20.4s}, [%[b_im]], #0x10") // v20-v21 = si |
| 613 |
__ASM_EMIT("ld1 {v24.4s}, [%[t_re]], #0x10") // v24-v25 = dr |
| 614 |
__ASM_EMIT("ld1 {v28.4s}, [%[t_im]], #0x10") // v28-v29 = di |
| 615 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 616 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 617 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 618 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 619 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 620 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 621 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 622 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 623 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 624 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 625 |
__ASM_EMIT("st1 {v0.4s}, [%[dst_re]], #0x10") |
| 626 |
__ASM_EMIT("st1 {v4.4s}, [%[dst_im]], #0x10") |
| 627 |
__ASM_EMIT("6:") |
| 628 |
// x1 blocks |
| 629 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 630 |
__ASM_EMIT("b.lt 8f") |
| 631 |
__ASM_EMIT("7:") |
| 632 |
__ASM_EMIT("ld1r {v16.4s}, [%[b_re]], #0x04") // v16 = sr |
| 633 |
__ASM_EMIT("ld1r {v20.4s}, [%[b_im]], #0x04") // v20 = si |
| 634 |
__ASM_EMIT("ld1r {v24.4s}, [%[t_re]], #0x04") // v24 = dr |
| 635 |
__ASM_EMIT("ld1r {v28.4s}, [%[t_im]], #0x04") // v28 = di |
| 636 |
__ASM_EMIT("fmul v0.4s, v16.4s, v24.4s") // v0 = sr*dr |
| 637 |
__ASM_EMIT("fmul v4.4s, v16.4s, v28.4s") // v4 = sr*di |
| 638 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = sr*sr |
| 639 |
__ASM_EMIT("fmla v0.4s, v20.4s, v28.4s") // v0 = sr*dr + si*di |
| 640 |
__ASM_EMIT("fmla v16.4s, v20.4s, v20.4s") // v16 = r = sr*sr + si*si |
| 641 |
__ASM_EMIT("fdiv v12.4s, v8.4s, v16.4s") // v12 = R = 1/(sr*sr + si*si) |
| 642 |
__ASM_EMIT("fmls v4.4s, v20.4s, v24.4s") // v4 = i = sr*di - si*dr |
| 643 |
__ASM_EMIT("fmul v0.4s, v0.4s, v12.4s") // v0 = r/R |
| 644 |
__ASM_EMIT("fmul v4.4s, v4.4s, v12.4s") // v4 = i/R |
| 645 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst_re]], #0x04") |
| 646 |
__ASM_EMIT("st1 {v4.s}[0], [%[dst_im]], #0x04") |
| 647 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 648 |
__ASM_EMIT("b.ge 7b") |
| 649 |
__ASM_EMIT("8:") |
| 650 |
|
| 651 |
: [dst_re] "+r" (dst_re), [dst_im] "+r" (dst_im), |
| 652 |
[t_re] "+r" (t_re), [t_im] "+r" (t_im), |
| 653 |
[b_re] "+r" (b_re), [b_im] "+r" (b_im), |
| 654 |
[count] "+r" (count) |
| 655 |
: [kdiv] "r" (&complex_div_const[0]) |
| 656 |
: "cc", "memory", |
| 657 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 658 |
"v8", /* "v9", "v10", "v11", */ "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 659 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 660 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 661 |
); |
| 662 |
} |
| 663 |
|
| 664 |
void complex_mod(float *dst, const float *src_re, const float *src_im, size_t count) |
| 665 |
{ |
| 666 |
ARCH_AARCH64_ASM |
| 667 |
( |
| 668 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 669 |
__ASM_EMIT("b.lo 2f") |
| 670 |
|
| 671 |
// x16 blocks |
| 672 |
__ASM_EMIT("1:") |
| 673 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src_re]], #0x40") // v16-v19 = r |
| 674 |
__ASM_EMIT("ld1 {v24.4s-v27.4s}, [%[src_im]], #0x40") // v24-v27 = i |
| 675 |
|
| 676 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = r*r |
| 677 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 678 |
__ASM_EMIT("fmul v18.4s, v18.4s, v18.4s") |
| 679 |
__ASM_EMIT("fmul v19.4s, v19.4s, v19.4s") |
| 680 |
__ASM_EMIT("fmla v16.4s, v24.4s, v24.4s") // v16 = R = r*r + i*i |
| 681 |
__ASM_EMIT("fmla v17.4s, v25.4s, v25.4s") |
| 682 |
__ASM_EMIT("fmla v18.4s, v26.4s, v26.4s") |
| 683 |
__ASM_EMIT("fmla v19.4s, v27.4s, v27.4s") |
| 684 |
__ASM_EMIT("fsqrt v16.4s, v16.4s") // v16 = sqrt(R) |
| 685 |
__ASM_EMIT("fsqrt v17.4s, v17.4s") |
| 686 |
__ASM_EMIT("fsqrt v18.4s, v18.4s") |
| 687 |
__ASM_EMIT("fsqrt v19.4s, v19.4s") |
| 688 |
|
| 689 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 690 |
__ASM_EMIT("st1 {v16.4s-v19.4s}, [%[dst]], #0x40") |
| 691 |
__ASM_EMIT("b.hs 1b") |
| 692 |
|
| 693 |
// x8 blocks |
| 694 |
__ASM_EMIT("2:") |
| 695 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 696 |
__ASM_EMIT("b.lt 4f") |
| 697 |
|
| 698 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[src_re]], #0x20") // v16-v17 = r |
| 699 |
__ASM_EMIT("ld1 {v24.4s, v25.4s}, [%[src_im]], #0x20") // v24-v25 = i |
| 700 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = r*r |
| 701 |
__ASM_EMIT("fmul v17.4s, v17.4s, v17.4s") |
| 702 |
__ASM_EMIT("fmla v16.4s, v24.4s, v24.4s") // v16 = R = r*r + i*i |
| 703 |
__ASM_EMIT("fmla v17.4s, v25.4s, v25.4s") |
| 704 |
__ASM_EMIT("fsqrt v16.4s, v16.4s") // v16 = sqrt(R) |
| 705 |
__ASM_EMIT("fsqrt v17.4s, v17.4s") |
| 706 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 707 |
__ASM_EMIT("st1 {v16.4s, v17.4s}, [%[dst]], #0x20") |
| 708 |
|
| 709 |
// x4 blocks |
| 710 |
__ASM_EMIT("4:") |
| 711 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 712 |
__ASM_EMIT("b.lt 6f") |
| 713 |
|
| 714 |
__ASM_EMIT("ld1 {v16.4s}, [%[src_re]], #0x10") // v16 = r |
| 715 |
__ASM_EMIT("ld1 {v24.4s}, [%[src_im]], #0x10") // v24 = i |
| 716 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = r*r |
| 717 |
__ASM_EMIT("fmla v16.4s, v24.4s, v24.4s") // v16 = R = r*r + i*i |
| 718 |
__ASM_EMIT("fsqrt v16.4s, v16.4s") // v16 = sqrt(R) |
| 719 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 720 |
__ASM_EMIT("st1 {v16.4s}, [%[dst]], #0x10") |
| 721 |
|
| 722 |
// x1 blocks |
| 723 |
__ASM_EMIT("6:") |
| 724 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 725 |
__ASM_EMIT("b.lt 8f") |
| 726 |
__ASM_EMIT("7:") |
| 727 |
|
| 728 |
__ASM_EMIT("ld1r {v16.4s}, [%[src_re]], #0x04") // v16 = r |
| 729 |
__ASM_EMIT("ld1r {v24.4s}, [%[src_im]], #0x04") // v24 = i |
| 730 |
__ASM_EMIT("fmul v16.4s, v16.4s, v16.4s") // v16 = r*r |
| 731 |
__ASM_EMIT("fmla v16.4s, v24.4s, v24.4s") // v16 = R = r*r + i*i |
| 732 |
__ASM_EMIT("fsqrt v16.4s, v16.4s") // v16 = sqrt(R) |
| 733 |
__ASM_EMIT("str s16, [%[dst]], #0x04") |
| 734 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 735 |
__ASM_EMIT("b.ge 7b") |
| 736 |
|
| 737 |
__ASM_EMIT("8:") |
| 738 |
|
| 739 |
: [dst] "+r" (dst), [src_re] "+r" (src_re), [src_im] "+r" (src_im), |
| 740 |
[count] "+r" (count) |
| 741 |
: |
| 742 |
: "cc", "memory", |
| 743 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 744 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 745 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 746 |
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" |
| 747 |
); |
| 748 |
} |
| 749 |
|
| 750 |
void complex_rcp1(float *dst_re, float *dst_im, size_t count) |
| 751 |
{ |
| 752 |
ARCH_AARCH64_ASM |
| 753 |
( |
| 754 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 755 |
__ASM_EMIT("b.lo 2f") |
| 756 |
|
| 757 |
// x16 blocks |
| 758 |
__ASM_EMIT("1:") |
| 759 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[dst_re]]") // v16-v19 = dr |
| 760 |
__ASM_EMIT("ld1 {v20.4s-v23.4s}, [%[dst_im]]") // v20-v23 = di |
| 761 |
|
| 762 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 763 |
__ASM_EMIT("fmul v1.4s, v17.4s, v17.4s") |
| 764 |
__ASM_EMIT("fmul v2.4s, v18.4s, v18.4s") |
| 765 |
__ASM_EMIT("fmul v3.4s, v19.4s, v19.4s") |
| 766 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 767 |
__ASM_EMIT("fmla v1.4s, v21.4s, v21.4s") |
| 768 |
__ASM_EMIT("fmla v2.4s, v22.4s, v22.4s") |
| 769 |
__ASM_EMIT("fmla v3.4s, v23.4s, v23.4s") |
| 770 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 771 |
__ASM_EMIT("fneg v21.4s, v21.4s") |
| 772 |
__ASM_EMIT("fneg v22.4s, v22.4s") |
| 773 |
__ASM_EMIT("fneg v23.4s, v23.4s") |
| 774 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 775 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v1.4s") |
| 776 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v2.4s") |
| 777 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v3.4s") |
| 778 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 779 |
__ASM_EMIT("fdiv v21.4s, v21.4s, v1.4s") |
| 780 |
__ASM_EMIT("fdiv v22.4s, v22.4s, v2.4s") |
| 781 |
__ASM_EMIT("fdiv v23.4s, v23.4s, v3.4s") |
| 782 |
|
| 783 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 784 |
__ASM_EMIT("st1 {v16.4s-v19.4s}, [%[dst_re]], #0x40") |
| 785 |
__ASM_EMIT("st1 {v20.4s-v23.4s}, [%[dst_im]], #0x40") |
| 786 |
__ASM_EMIT("b.hs 1b") |
| 787 |
|
| 788 |
// x8 blocks |
| 789 |
__ASM_EMIT("2:") |
| 790 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 791 |
__ASM_EMIT("b.lt 4f") |
| 792 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[dst_re]]") // v16-v17 = dr |
| 793 |
__ASM_EMIT("ld1 {v20.4s, v21.4s}, [%[dst_im]]") // v20-v21 = di |
| 794 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 795 |
__ASM_EMIT("fmul v1.4s, v17.4s, v17.4s") |
| 796 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 797 |
__ASM_EMIT("fmla v1.4s, v21.4s, v21.4s") |
| 798 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 799 |
__ASM_EMIT("fneg v21.4s, v21.4s") |
| 800 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 801 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v1.4s") |
| 802 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 803 |
__ASM_EMIT("fdiv v21.4s, v21.4s, v1.4s") |
| 804 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 805 |
__ASM_EMIT("st1 {v16.4s, v17.4s}, [%[dst_re]], #0x20") |
| 806 |
__ASM_EMIT("st1 {v20.4s, v21.4s}, [%[dst_im]], #0x20") |
| 807 |
|
| 808 |
// x4 blocks |
| 809 |
__ASM_EMIT("4:") |
| 810 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 811 |
__ASM_EMIT("b.lt 6f") |
| 812 |
__ASM_EMIT("ld1 {v16.4s}, [%[dst_re]]") // v16-v17 = dr |
| 813 |
__ASM_EMIT("ld1 {v20.4s}, [%[dst_im]]") // v20-v21 = di |
| 814 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 815 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 816 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 817 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 818 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 819 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 820 |
__ASM_EMIT("st1 {v16.4s}, [%[dst_re]], #0x10") |
| 821 |
__ASM_EMIT("st1 {v20.4s}, [%[dst_im]], #0x10") |
| 822 |
|
| 823 |
// x1 blocks |
| 824 |
__ASM_EMIT("6:") |
| 825 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 826 |
__ASM_EMIT("b.lt 8f") |
| 827 |
__ASM_EMIT("7:") |
| 828 |
__ASM_EMIT("ld1r {v16.4s}, [%[dst_re]]") // v16 = dr |
| 829 |
__ASM_EMIT("ld1r {v20.4s}, [%[dst_im]]") // v20 = di |
| 830 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 831 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 832 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 833 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 834 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 835 |
__ASM_EMIT("str s16, [%[dst_re]], #0x04") |
| 836 |
__ASM_EMIT("str s20, [%[dst_im]], #0x04") |
| 837 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 838 |
__ASM_EMIT("b.ge 7b") |
| 839 |
|
| 840 |
__ASM_EMIT("8:") |
| 841 |
|
| 842 |
: [dst_re] "+r" (dst_re), [dst_im] "+r" (dst_im), |
| 843 |
[count] "+r" (count) |
| 844 |
: |
| 845 |
: "cc", "memory", |
| 846 |
"v0", "v1", "v2", "v3", |
| 847 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 848 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" |
| 849 |
); |
| 850 |
} |
| 851 |
|
| 852 |
void complex_rcp2(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t count) |
| 853 |
{ |
| 854 |
ARCH_AARCH64_ASM |
| 855 |
( |
| 856 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 857 |
__ASM_EMIT("b.lo 2f") |
| 858 |
|
| 859 |
// x16 blocks |
| 860 |
__ASM_EMIT("1:") |
| 861 |
__ASM_EMIT("ld1 {v16.4s-v19.4s}, [%[src_re]], #0x40") // v16-v19 = dr |
| 862 |
__ASM_EMIT("ld1 {v20.4s-v23.4s}, [%[src_im]], #0x40") // v20-v23 = di |
| 863 |
|
| 864 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 865 |
__ASM_EMIT("fmul v1.4s, v17.4s, v17.4s") |
| 866 |
__ASM_EMIT("fmul v2.4s, v18.4s, v18.4s") |
| 867 |
__ASM_EMIT("fmul v3.4s, v19.4s, v19.4s") |
| 868 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 869 |
__ASM_EMIT("fmla v1.4s, v21.4s, v21.4s") |
| 870 |
__ASM_EMIT("fmla v2.4s, v22.4s, v22.4s") |
| 871 |
__ASM_EMIT("fmla v3.4s, v23.4s, v23.4s") |
| 872 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 873 |
__ASM_EMIT("fneg v21.4s, v21.4s") |
| 874 |
__ASM_EMIT("fneg v22.4s, v22.4s") |
| 875 |
__ASM_EMIT("fneg v23.4s, v23.4s") |
| 876 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 877 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v1.4s") |
| 878 |
__ASM_EMIT("fdiv v18.4s, v18.4s, v2.4s") |
| 879 |
__ASM_EMIT("fdiv v19.4s, v19.4s, v3.4s") |
| 880 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 881 |
__ASM_EMIT("fdiv v21.4s, v21.4s, v1.4s") |
| 882 |
__ASM_EMIT("fdiv v22.4s, v22.4s, v2.4s") |
| 883 |
__ASM_EMIT("fdiv v23.4s, v23.4s, v3.4s") |
| 884 |
|
| 885 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 886 |
__ASM_EMIT("st1 {v16.4s-v19.4s}, [%[dst_re]], #0x40") |
| 887 |
__ASM_EMIT("st1 {v20.4s-v23.4s}, [%[dst_im]], #0x40") |
| 888 |
__ASM_EMIT("b.hs 1b") |
| 889 |
|
| 890 |
// x8 blocks |
| 891 |
__ASM_EMIT("2:") |
| 892 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 893 |
__ASM_EMIT("b.lt 4f") |
| 894 |
__ASM_EMIT("ld1 {v16.4s, v17.4s}, [%[src_re]], #0x20") // v16-v17 = dr |
| 895 |
__ASM_EMIT("ld1 {v20.4s, v21.4s}, [%[src_im]], #0x20") // v20-v21 = di |
| 896 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 897 |
__ASM_EMIT("fmul v1.4s, v17.4s, v17.4s") |
| 898 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 899 |
__ASM_EMIT("fmla v1.4s, v21.4s, v21.4s") |
| 900 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 901 |
__ASM_EMIT("fneg v21.4s, v21.4s") |
| 902 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 903 |
__ASM_EMIT("fdiv v17.4s, v17.4s, v1.4s") |
| 904 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 905 |
__ASM_EMIT("fdiv v21.4s, v21.4s, v1.4s") |
| 906 |
__ASM_EMIT("sub %[count], %[count], #8") |
| 907 |
__ASM_EMIT("st1 {v16.4s, v17.4s}, [%[dst_re]], #0x20") |
| 908 |
__ASM_EMIT("st1 {v20.4s, v21.4s}, [%[dst_im]], #0x20") |
| 909 |
|
| 910 |
// x4 blocks |
| 911 |
__ASM_EMIT("4:") |
| 912 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 913 |
__ASM_EMIT("b.lt 6f") |
| 914 |
__ASM_EMIT("ld1 {v16.4s}, [%[src_re]], #0x10") // v16-v17 = dr |
| 915 |
__ASM_EMIT("ld1 {v20.4s}, [%[src_im]], #0x10") // v20-v21 = di |
| 916 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 917 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 918 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 919 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 920 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 921 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 922 |
__ASM_EMIT("st1 {v16.4s}, [%[dst_re]], #0x10") |
| 923 |
__ASM_EMIT("st1 {v20.4s}, [%[dst_im]], #0x10") |
| 924 |
|
| 925 |
// x1 blocks |
| 926 |
__ASM_EMIT("6:") |
| 927 |
__ASM_EMIT("adds %[count], %[count], #3") |
| 928 |
__ASM_EMIT("b.lt 8f") |
| 929 |
__ASM_EMIT("7:") |
| 930 |
__ASM_EMIT("ld1r {v16.4s}, [%[src_re]], #0x04") // v16 = dr |
| 931 |
__ASM_EMIT("ld1r {v20.4s}, [%[src_im]], #0x04") // v20 = di |
| 932 |
__ASM_EMIT("fmul v0.4s, v16.4s, v16.4s") // v0 = dr*dr |
| 933 |
__ASM_EMIT("fmla v0.4s, v20.4s, v20.4s") // v0 = dr*dr + di*di |
| 934 |
__ASM_EMIT("fneg v20.4s, v20.4s") // v20 = -di |
| 935 |
__ASM_EMIT("fdiv v16.4s, v16.4s, v0.4s") // v16 = dr/(dr*dr + di*di) |
| 936 |
__ASM_EMIT("fdiv v20.4s, v20.4s, v0.4s") // v20 = -di/(dr*dr + di*di) |
| 937 |
__ASM_EMIT("str s16, [%[dst_re]], #0x04") |
| 938 |
__ASM_EMIT("str s20, [%[dst_im]], #0x04") |
| 939 |
__ASM_EMIT("subs %[count], %[count], #1") |
| 940 |
__ASM_EMIT("b.ge 7b") |
| 941 |
|
| 942 |
__ASM_EMIT("8:") |
| 943 |
|
| 944 |
: [dst_re] "+r" (dst_re), [dst_im] "+r" (dst_im), |
| 945 |
[src_re] "+r" (src_re), [src_im] "+r" (src_im), |
| 946 |
[count] "+r" (count) |
| 947 |
: |
| 948 |
: "cc", "memory", |
| 949 |
"v0", "v1", "v2", "v3", |
| 950 |
// "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", // Avoid usage if possible |
| 951 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" |
| 952 |
); |
| 953 |
} |
| 954 |
} |
| 955 |
} |
| 956 |
|
| 957 |
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_COMPLEX_H_ */ |