| 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_FASTCONV_APPLY_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_FASTCONV_APPLY_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 fastconv_apply(float *dst, float *tmp, const float *c1, const float *c2, size_t rank) |
| 34 |
{ |
| 35 |
size_t items = 1 << rank; // number of complex numbers per half of block (16 for rank=4) |
| 36 |
size_t n; |
| 37 |
|
| 38 |
IF_ARCH_AARCH64( |
| 39 |
float *a, *b, *ptr; |
| 40 |
size_t k, p; |
| 41 |
); |
| 42 |
|
| 43 |
ARCH_AARCH64_ASM( |
| 44 |
__ASM_EMIT("subs %[n], %[items], #8") |
| 45 |
__ASM_EMIT("ldp q24, q25, [%[MASK]]") |
| 46 |
__ASM_EMIT("mov %[ptr], %[tmp]") |
| 47 |
__ASM_EMIT("b.lo 2f") |
| 48 |
// Load data and perform complex multiplication |
| 49 |
__ASM_EMIT("1:") |
| 50 |
__ASM_EMIT("ldp q16, q17, [%[c1], #0x00]") // v16 = ar0, v17 = ai0 |
| 51 |
__ASM_EMIT("ldp q18, q19, [%[c1], #0x20]") // v18 = ar1, v19 = ai1 |
| 52 |
__ASM_EMIT("ldp q20, q21, [%[c2], #0x00]") // v20 = br0, v21 = bi0 |
| 53 |
__ASM_EMIT("ldp q22, q23, [%[c2], #0x20]") // v22 = br1, v23 = bi1 |
| 54 |
__ASM_EMIT("fmul v0.4s, v16.4s, v20.4s") // v0 = ar0 * br0 |
| 55 |
__ASM_EMIT("fmul v2.4s, v18.4s, v22.4s") // v2 = ar1 * br1 |
| 56 |
__ASM_EMIT("fmul v1.4s, v16.4s, v21.4s") // v1 = ar0 * bi0 |
| 57 |
__ASM_EMIT("fmul v3.4s, v18.4s, v23.4s") // v3 = ar1 * bi1 |
| 58 |
__ASM_EMIT("fmls v0.4s, v17.4s, v21.4s") // v0 = ar0*br0 - ai0*bi0 |
| 59 |
__ASM_EMIT("fmls v2.4s, v19.4s, v23.4s") // v2 = ar1*br1 - ai1*bi1 |
| 60 |
__ASM_EMIT("fmla v1.4s, v17.4s, v20.4s") // v1 = ar0*bi0 + ai0*br0 |
| 61 |
__ASM_EMIT("fmla v3.4s, v19.4s, v22.4s") // v3 = ar1*bi1 + ai1*br1 |
| 62 |
// v0 = r0 r1 r2 r3 |
| 63 |
// v1 = i0 i1 i2 i3 |
| 64 |
// v2 = r4 r5 r6 r7 |
| 65 |
// v3 = i4 i5 i6 i7 |
| 66 |
__ASM_EMIT("trn1 v4.4s, v0.4s, v1.4s") // v4 = r0 i0 r2 i2 |
| 67 |
__ASM_EMIT("trn1 v5.4s, v2.4s, v3.4s") // v5 = r4 i4 r6 i6 |
| 68 |
__ASM_EMIT("trn2 v6.4s, v0.4s, v1.4s") // v6 = r1 i1 r3 i3 |
| 69 |
__ASM_EMIT("trn2 v7.4s, v2.4s, v3.4s") // v7 = r5 i5 r7 i7 |
| 70 |
__ASM_EMIT("trn1 v0.2d, v4.2d, v5.2d") // v0 = r0 i0 r4 i4 |
| 71 |
__ASM_EMIT("trn1 v1.2d, v6.2d, v7.2d") // v1 = r1 i1 r5 i5 |
| 72 |
__ASM_EMIT("trn2 v2.2d, v4.2d, v5.2d") // v2 = r2 i2 r6 i6 |
| 73 |
__ASM_EMIT("trn2 v3.2d, v6.2d, v7.2d") // v3 = r3 i3 r7 i7 |
| 74 |
// v0 = r0 i0 r4 i4 |
| 75 |
// v1 = r1 i1 r5 i5 |
| 76 |
// v2 = r2 i2 r6 i6 |
| 77 |
// v3 = r3 i3 r7 i7 |
| 78 |
__ASM_EMIT("fsub v6.4s, v0.4s, v1.4s") // v6 = r0-r1 i0-i1 r4-r5 i4-i5 = r1' i1' r5' i5' |
| 79 |
__ASM_EMIT("fsub v7.4s, v2.4s, v3.4s") // v7 = r2-r3 i2-i3 r6-r7 i6-i7 = r3' i3' r7' i7' |
| 80 |
__ASM_EMIT("fadd v4.4s, v0.4s, v1.4s") // v4 = r0+r1 i0+i1 r4+r5 i4+i5 = r0' i0' r4' i4' |
| 81 |
__ASM_EMIT("fadd v5.4s, v2.4s, v3.4s") // v5 = r2+r3 i2+i3 r6+r7 i6+i7 = r2' i2' r6' i6' |
| 82 |
__ASM_EMIT("rev64 v7.4s, v7.4s") // v7 = i3' r3' i7' r7' |
| 83 |
// v4 = r0' i0' r4' i4' |
| 84 |
// v5 = r2' i2' r6' i6' |
| 85 |
// v6 = r1' i1' r5' i5' |
| 86 |
// v7 = i3' r3' i7' r7' |
| 87 |
__ASM_EMIT("fadd v0.4s, v4.4s, v5.4s") // v0 = r0'+r2' i0'+i2' r4'+r6' i4'+i6' = r0" i0" r4" i4" |
| 88 |
__ASM_EMIT("fsub v1.4s, v6.4s, v7.4s") // v2 = r1'-i3' i1'-r3' r5'-i7' i5'-r7' = r1" i3" r5" i7" |
| 89 |
__ASM_EMIT("fsub v2.4s, v4.4s, v5.4s") // v1 = r0'-r2' i0'-i2' r4'-r6' i4'-i6' = r2" i2" r6" i6" |
| 90 |
__ASM_EMIT("fadd v3.4s, v6.4s, v7.4s") // v3 = r1'+i3' i1'+r3' r5'+i7' i5'+r7' = r3" i1" r7" i5" |
| 91 |
__ASM_EMIT("mov v6.16b, v1.16b") // v6 = r1" i3" r5" i7" |
| 92 |
__ASM_EMIT("mov v7.16b, v3.16b") // v7 = r3" i1" r7" i5" |
| 93 |
__ASM_EMIT("bit v1.16b, v7.16b, v24.16b") // v1 = r1" i1" r5" i5" |
| 94 |
__ASM_EMIT("bit v3.16b, v6.16b, v25.16b") // v3 = r3" i3" r7" i7" |
| 95 |
__ASM_EMIT("st4 {v0.4s-v3.4s}, [%[ptr]]") |
| 96 |
__ASM_EMIT("subs %[n], %[n], #8") // n -= 8 |
| 97 |
__ASM_EMIT("add %[c1], %[c1], #0x40") |
| 98 |
__ASM_EMIT("add %[c2], %[c2], #0x40") |
| 99 |
__ASM_EMIT("add %[ptr], %[ptr], #0x40") |
| 100 |
__ASM_EMIT("b.hs 1b") |
| 101 |
__ASM_EMIT("2:") |
| 102 |
|
| 103 |
: [c1] "+r" (c1), [c2] "+r" (c2), [ptr] "=&r" (ptr), |
| 104 |
[n] "=&r" (n) |
| 105 |
: [tmp] "r" (tmp), [items] "r" (items), |
| 106 |
[MASK] "r" (&fastconv_swp_mask[0]) |
| 107 |
: "cc", "memory", |
| 108 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 109 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 110 |
"v24", "v25" |
| 111 |
); |
| 112 |
|
| 113 |
n = 8; |
| 114 |
const float *xfft_a = XFFT_A; |
| 115 |
|
| 116 |
// Perform butterflies without need of vector rotation |
| 117 |
if (n < items) |
| 118 |
{ |
| 119 |
ARCH_AARCH64_ASM( |
| 120 |
__ASM_EMIT("mov %[a], %[tmp]") |
| 121 |
__ASM_EMIT("subs %[k], %[items], #16") |
| 122 |
__ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr1, v17 = wr2 |
| 123 |
__ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi1, v19 = wi2 |
| 124 |
__ASM_EMIT("b.lo 2f") |
| 125 |
// 8x butterflies |
| 126 |
// Calculate complex c = w * b |
| 127 |
__ASM_EMIT("1:") |
| 128 |
// Calc cr and ci |
| 129 |
__ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1 |
| 130 |
__ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2 |
| 131 |
__ASM_EMIT("ldp q4, q5, [%[a], #0x40]") // v4 = ar3, v5 = ai3 |
| 132 |
__ASM_EMIT("ldp q6, q7, [%[a], #0x60]") // v6 = ar4, v7 = ai4 |
| 133 |
__ASM_EMIT("fmul v20.4s, v16.4s, v2.4s") // v20 = wr1 * br1 |
| 134 |
__ASM_EMIT("fmul v21.4s, v17.4s, v6.4s") // v21 = wr2 * br2 |
| 135 |
__ASM_EMIT("fmul v22.4s, v16.4s, v3.4s") // v22 = wr1 * bi1 |
| 136 |
__ASM_EMIT("fmul v23.4s, v17.4s, v7.4s") // v23 = wr2 * bi2 |
| 137 |
__ASM_EMIT("fmls v20.4s, v18.4s, v3.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1 |
| 138 |
__ASM_EMIT("fmls v21.4s, v19.4s, v7.4s") // v21 = wr2 * br2 - wi2 * bi2 = cr2 |
| 139 |
__ASM_EMIT("fmla v22.4s, v18.4s, v2.4s") // v22 = wr1 * bi1 + wi1 * br1 = ci1 |
| 140 |
__ASM_EMIT("fmla v23.4s, v19.4s, v6.4s") // v23 = wr1 * bi1 + wi2 * br2 = ci2 |
| 141 |
// Apply butterfly |
| 142 |
__ASM_EMIT("fsub v2.4s, v0.4s, v20.4s") // v2 = ar1 - cr1 |
| 143 |
__ASM_EMIT("fsub v3.4s, v1.4s, v22.4s") // v3 = ai1 - ci1 |
| 144 |
__ASM_EMIT("fsub v6.4s, v4.4s, v21.4s") // v6 = ar2 - cr2 |
| 145 |
__ASM_EMIT("fsub v7.4s, v5.4s, v23.4s") // v7 = ai2 - ci2 |
| 146 |
__ASM_EMIT("fadd v0.4s, v0.4s, v20.4s") // v0 = ar1 + cr1 |
| 147 |
__ASM_EMIT("fadd v1.4s, v1.4s, v22.4s") // v1 = ai1 + ci1 |
| 148 |
__ASM_EMIT("fadd v4.4s, v4.4s, v21.4s") // v4 = ar2 + cr2 |
| 149 |
__ASM_EMIT("fadd v5.4s, v5.4s, v23.4s") // v5 = ai2 + ci2 |
| 150 |
__ASM_EMIT("stp q0, q1, [%[a], #0x00]") |
| 151 |
__ASM_EMIT("stp q2, q3, [%[a], #0x20]") |
| 152 |
__ASM_EMIT("stp q4, q5, [%[a], #0x40]") |
| 153 |
__ASM_EMIT("stp q6, q7, [%[a], #0x60]") |
| 154 |
__ASM_EMIT("subs %[k], %[k], #16") |
| 155 |
__ASM_EMIT("add %[a], %[a], #0x80") |
| 156 |
__ASM_EMIT("b.ge 1b") |
| 157 |
|
| 158 |
__ASM_EMIT("2:") |
| 159 |
__ASM_EMIT("adds %[k], %[k], #8") |
| 160 |
__ASM_EMIT("b.lo 4f") |
| 161 |
// 4x butterflies |
| 162 |
__ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1 |
| 163 |
__ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2 |
| 164 |
__ASM_EMIT("fmul v20.4s, v16.4s, v2.4s") // v20 = wr1 * br1 |
| 165 |
__ASM_EMIT("fmul v22.4s, v16.4s, v3.4s") // v22 = wr1 * bi1 |
| 166 |
__ASM_EMIT("fmls v20.4s, v18.4s, v3.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1 |
| 167 |
__ASM_EMIT("fmla v22.4s, v18.4s, v2.4s") // v22 = wr1 * bi1 + wi1 * br1 = ci1 |
| 168 |
// Apply butterfly |
| 169 |
__ASM_EMIT("fsub v2.4s, v0.4s, v20.4s") // v2 = ar1 - cr1 |
| 170 |
__ASM_EMIT("fsub v3.4s, v1.4s, v22.4s") // v3 = ai1 - ci1 |
| 171 |
__ASM_EMIT("fadd v0.4s, v0.4s, v20.4s") // v0 = ar1 + cr1 |
| 172 |
__ASM_EMIT("fadd v1.4s, v1.4s, v22.4s") // v1 = ai1 + ci1 |
| 173 |
__ASM_EMIT("stp q0, q1, [%[a], #0x00]") |
| 174 |
__ASM_EMIT("stp q2, q3, [%[a], #0x20]") |
| 175 |
__ASM_EMIT("4:") |
| 176 |
__ASM_EMIT("add %[XFFT_A], %[XFFT_A], #0x40") |
| 177 |
|
| 178 |
: [a] "=&r" (a), [k] "=&r" (k), |
| 179 |
[XFFT_A] "+r" (xfft_a) |
| 180 |
: [tmp] "r" (tmp), [items] "r" (items) |
| 181 |
: "cc", "memory", |
| 182 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 183 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" |
| 184 |
); |
| 185 |
|
| 186 |
n <<= 1; |
| 187 |
} |
| 188 |
|
| 189 |
const float *xfft_dw = &XFFT_DW[8]; |
| 190 |
|
| 191 |
// Perform loops except last one |
| 192 |
while (n < items) |
| 193 |
{ |
| 194 |
ARCH_AARCH64_ASM( |
| 195 |
__ASM_EMIT("mov %[k], %[items]") |
| 196 |
__ASM_EMIT("mov %[a], %[tmp]") |
| 197 |
__ASM_EMIT("1:") |
| 198 |
// Initialize sub-loop |
| 199 |
__ASM_EMIT("mov %[p], %[n]") |
| 200 |
__ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr1, v17 = wr2 |
| 201 |
__ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi1, v19 = wi2 |
| 202 |
__ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, v25 = di |
| 203 |
__ASM_EMIT("add %[b], %[a], %[n], LSL #2") // b = &a[n*4] |
| 204 |
// 8x butterflies |
| 205 |
__ASM_EMIT("3:") |
| 206 |
// Calc cr and ci |
| 207 |
__ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1 |
| 208 |
__ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2 |
| 209 |
__ASM_EMIT("ldp q4, q5, [%[b], #0x00]") // v4 = br1, v5 = bi1 |
| 210 |
__ASM_EMIT("ldp q6, q7, [%[b], #0x20]") // v6 = br2, v7 = bi2 |
| 211 |
__ASM_EMIT("fmul v20.4s, v16.4s, v4.4s") // v20 = wr1 * br1 |
| 212 |
__ASM_EMIT("fmul v21.4s, v17.4s, v6.4s") // v21 = wr2 * br2 |
| 213 |
__ASM_EMIT("fmul v22.4s, v16.4s, v5.4s") // v22 = wr1 * bi1 |
| 214 |
__ASM_EMIT("fmul v23.4s, v17.4s, v7.4s") // v23 = wr2 * bi2 |
| 215 |
__ASM_EMIT("fmls v20.4s, v18.4s, v5.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1 |
| 216 |
__ASM_EMIT("fmls v21.4s, v19.4s, v7.4s") // v21 = wr2 * br2 - wi2 * bi2 = cr2 |
| 217 |
__ASM_EMIT("fmla v22.4s, v18.4s, v4.4s") // v22 = wr1 * bi1 + wi1 * br1 = ci1 |
| 218 |
__ASM_EMIT("fmla v23.4s, v19.4s, v6.4s") // v23 = wr1 * bi1 + wi2 * br2 = ci2 |
| 219 |
// Apply butterfly |
| 220 |
__ASM_EMIT("fsub v4.4s, v0.4s, v20.4s") // v4 = ar1 - cr1 |
| 221 |
__ASM_EMIT("fsub v5.4s, v1.4s, v22.4s") // v5 = ai1 - ci1 |
| 222 |
__ASM_EMIT("fsub v6.4s, v2.4s, v21.4s") // v6 = ar2 - cr2 |
| 223 |
__ASM_EMIT("fsub v7.4s, v3.4s, v23.4s") // v7 = ai2 - ci2 |
| 224 |
__ASM_EMIT("fadd v0.4s, v0.4s, v20.4s") // v0 = ar1 + cr1 |
| 225 |
__ASM_EMIT("fadd v1.4s, v1.4s, v22.4s") // v1 = ai1 + ci1 |
| 226 |
__ASM_EMIT("fadd v2.4s, v2.4s, v21.4s") // v2 = ar2 + cr2 |
| 227 |
__ASM_EMIT("fadd v3.4s, v3.4s, v23.4s") // v3 = ai2 + ci2 |
| 228 |
__ASM_EMIT("stp q0, q1, [%[a], #0x00]") |
| 229 |
__ASM_EMIT("stp q2, q3, [%[a], #0x20]") |
| 230 |
__ASM_EMIT("stp q4, q5, [%[b], #0x00]") |
| 231 |
__ASM_EMIT("stp q6, q7, [%[b], #0x20]") |
| 232 |
__ASM_EMIT("subs %[p], %[p], #16") |
| 233 |
__ASM_EMIT("add %[a], %[a], #0x40") |
| 234 |
__ASM_EMIT("add %[b], %[b], #0x40") |
| 235 |
__ASM_EMIT("b.eq 4f") |
| 236 |
// Prepare next loop: rotate angle |
| 237 |
__ASM_EMIT("fmul v20.4s, v16.4s, v25.4s") // v20 = wr1 * di |
| 238 |
__ASM_EMIT("fmul v21.4s, v17.4s, v25.4s") // v21 = wr2 * di |
| 239 |
__ASM_EMIT("fmul v22.4s, v18.4s, v25.4s") // v22 = wi1 * di |
| 240 |
__ASM_EMIT("fmul v23.4s, v19.4s, v25.4s") // v23 = wi2 * di |
| 241 |
__ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16 = wr1 * dr |
| 242 |
__ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17 = wr2 * dr |
| 243 |
__ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18 = wi1 * dr |
| 244 |
__ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19 = wi2 * dr |
| 245 |
__ASM_EMIT("fsub v16.4s, v16.4s, v22.4s") // v16 = wr1*dr - wi1*di |
| 246 |
__ASM_EMIT("fsub v17.4s, v17.4s, v23.4s") // v17 = wr2*dr - wi2*di |
| 247 |
__ASM_EMIT("fadd v18.4s, v18.4s, v20.4s") // v18 = wi1*dr + wr1*di |
| 248 |
__ASM_EMIT("fadd v19.4s, v19.4s, v21.4s") // v19 = wi2*dr + wr2*di |
| 249 |
__ASM_EMIT("b 3b") |
| 250 |
__ASM_EMIT("4:") |
| 251 |
__ASM_EMIT("subs %[k], %[k], %[n]") |
| 252 |
__ASM_EMIT("mov %[a], %[b]") |
| 253 |
__ASM_EMIT("b.ne 1b") |
| 254 |
|
| 255 |
: [a] "=&r" (a), [b] "=&r" (b), |
| 256 |
[p] "=&r" (p), [k] "=&r" (k) |
| 257 |
: [tmp] "r" (tmp), [items] "r" (items), [n] "r" (n), |
| 258 |
[XFFT_A] "r" (xfft_a), [XFFT_DW] "r" (xfft_dw) |
| 259 |
: "cc", "memory", |
| 260 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 261 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 262 |
"v24", "v25" |
| 263 |
); |
| 264 |
|
| 265 |
n <<= 1; |
| 266 |
xfft_a += 16; |
| 267 |
xfft_dw += 8; |
| 268 |
} |
| 269 |
|
| 270 |
// Last one large loop |
| 271 |
IF_ARCH_AARCH64(float fftk = 1.0f / items); |
| 272 |
|
| 273 |
ARCH_AARCH64_ASM( |
| 274 |
// Initialize sub-loop |
| 275 |
__ASM_EMIT("add %[b], %[a], %[n], LSL #2") // b = &a[n*4] |
| 276 |
__ASM_EMIT("ld1r {v23.4s}, [%[fftk]]") // v23 = 1 / items |
| 277 |
__ASM_EMIT("add %[db], %[da], %[n], LSL #1") // db = &da[n*2] |
| 278 |
__ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr1, v17 = wr2 |
| 279 |
__ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi1, v19 = wi2 |
| 280 |
__ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, v25 = di |
| 281 |
__ASM_EMIT("mov v22.16b, v23.16b") // v22 = 1 / items |
| 282 |
// 8x butterflies |
| 283 |
__ASM_EMIT("1:") |
| 284 |
__ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1 |
| 285 |
__ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2 |
| 286 |
__ASM_EMIT("ldp q4, q5, [%[b], #0x00]") // v4 = br1, v5 = bi1 |
| 287 |
__ASM_EMIT("ldp q6, q7, [%[b], #0x20]") // v6 = br2, v7 = bi2 |
| 288 |
__ASM_EMIT("fmul v20.4s, v16.4s, v4.4s") // v20 = wr1 * br1 |
| 289 |
__ASM_EMIT("fmul v21.4s, v17.4s, v6.4s") // v21 = wr2 * br2 |
| 290 |
__ASM_EMIT("fmls v20.4s, v18.4s, v5.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1 |
| 291 |
__ASM_EMIT("fmls v21.4s, v19.4s, v7.4s") // v21 = wr2 * br2 - wi2 * bi2 = cr2 |
| 292 |
// Apply butterfly, we do not need imaginary part |
| 293 |
__ASM_EMIT("fadd v4.4s, v0.4s, v20.4s") // v4 = ar1 + cr1 |
| 294 |
__ASM_EMIT("fadd v5.4s, v2.4s, v21.4s") // v5 = ar2 + cr2 |
| 295 |
__ASM_EMIT("fsub v6.4s, v0.4s, v20.4s") // v6 = ar1 - cr1 |
| 296 |
__ASM_EMIT("fsub v7.4s, v2.4s, v21.4s") // v7 = ar2 - cr2 |
| 297 |
__ASM_EMIT("ldp q0, q1, [%[da]]") // v0 = dar1, v1 = dar2 |
| 298 |
__ASM_EMIT("ldp q2, q3, [%[db]]") // v2 = dbr1, v3 = dbr2 |
| 299 |
__ASM_EMIT("fmla v0.4s, v4.4s, v22.4s") // v4 = dar1 + (ar1 + cr1) / items |
| 300 |
__ASM_EMIT("fmla v1.4s, v5.4s, v23.4s") // v5 = dar2 + (ar2 + cr2) / items |
| 301 |
__ASM_EMIT("fmla v2.4s, v6.4s, v22.4s") // v6 = dbr1 + (ar1 - cr1) / items |
| 302 |
__ASM_EMIT("fmla v3.4s, v7.4s, v23.4s") // v7 = dbr2 + (ar2 - cr2) / items |
| 303 |
__ASM_EMIT("stp q0, q1, [%[da]]") |
| 304 |
__ASM_EMIT("stp q2, q3, [%[db]]") |
| 305 |
__ASM_EMIT("subs %[items], %[items], #16") |
| 306 |
__ASM_EMIT("add %[a], %[a], #0x40") |
| 307 |
__ASM_EMIT("add %[b], %[b], #0x40") |
| 308 |
__ASM_EMIT("add %[da], %[da], #0x20") |
| 309 |
__ASM_EMIT("add %[db], %[db], #0x20") |
| 310 |
__ASM_EMIT("b.ls 2f") |
| 311 |
// Prepare next loop: rotate angle |
| 312 |
__ASM_EMIT("fmul v2.4s, v16.4s, v25.4s") // v2 = wr1 * di |
| 313 |
__ASM_EMIT("fmul v3.4s, v17.4s, v25.4s") // v3 = wr2 * di |
| 314 |
__ASM_EMIT("fmul v4.4s, v18.4s, v25.4s") // v4 = wi1 * di |
| 315 |
__ASM_EMIT("fmul v5.4s, v19.4s, v25.4s") // v5 = wi2 * di |
| 316 |
__ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16 = wr1 * dr |
| 317 |
__ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17 = wr2 * dr |
| 318 |
__ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18 = wi1 * dr |
| 319 |
__ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19 = wi2 * dr |
| 320 |
__ASM_EMIT("fsub v16.4s, v16.4s, v4.4s") // v16 = wr1*dr - wi1*di |
| 321 |
__ASM_EMIT("fsub v17.4s, v17.4s, v5.4s") // v17 = wr2*dr - wi2*di |
| 322 |
__ASM_EMIT("fadd v18.4s, v18.4s, v2.4s") // v18 = wi1*dr + wr1*di |
| 323 |
__ASM_EMIT("fadd v19.4s, v19.4s, v3.4s") // v19 = wi2*dr + wr2*di |
| 324 |
__ASM_EMIT("b 1b") |
| 325 |
__ASM_EMIT("2:") |
| 326 |
|
| 327 |
: [da] "+r" (dst), [db] "=&r" (a), |
| 328 |
[a] "+r" (tmp), [b] "=&r" (b), |
| 329 |
[items] "+r" (items) |
| 330 |
: [n] "r" (n), [fftk] "r" (&fftk), |
| 331 |
[XFFT_A] "r" (xfft_a), [XFFT_DW] "r" (xfft_dw) |
| 332 |
: "cc", "memory", |
| 333 |
"v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", |
| 334 |
"v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", |
| 335 |
"v24", "v25" |
| 336 |
); |
| 337 |
} |
| 338 |
} |
| 339 |
} |
| 340 |
|
| 341 |
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_FASTCONV_APPLY_H_ */ |