| 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_FFT_PSCRAMBLE_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_FFT_PSCRAMBLE_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 packed_scramble_self_direct(float *dst, size_t rank) |
| 34 |
{ |
| 35 |
IF_ARCH_AARCH64( |
| 36 |
size_t count = 1 << rank; |
| 37 |
size_t i, j, rrank = 64 - rank; |
| 38 |
float *dp, *src = dst; |
| 39 |
); |
| 40 |
|
| 41 |
// Self algorithm |
| 42 |
ARCH_AARCH64_ASM( |
| 43 |
// Do bit-reverse shuffle |
| 44 |
__ASM_EMIT("mov %[i], #1") // i = 1 |
| 45 |
__ASM_EMIT("1:") |
| 46 |
__ASM_EMIT("rbit %[j], %[i]") // j = reverse_bits(i) |
| 47 |
__ASM_EMIT("add %[src], %[src], #8") |
| 48 |
__ASM_EMIT("lsr %[j], %[j], %[rrank]") // j = reverse_bits(i) >> rank |
| 49 |
__ASM_EMIT("cmp %[i], %[j]") // i <=> j |
| 50 |
__ASM_EMIT("b.hs 2f") // if (i >= j) continue |
| 51 |
__ASM_EMIT("add %[dp], %[dst], %[j], LSL #3") // d_re = &dst[j] |
| 52 |
__ASM_EMIT("ldr d0, [%[src]]") // d0 = *src |
| 53 |
__ASM_EMIT("ldr d1, [%[dp]]") // d1 = *dp |
| 54 |
__ASM_EMIT("str d0, [%[dp]]") // *dp = d0 |
| 55 |
__ASM_EMIT("str d1, [%[src]]") // *src = d1 |
| 56 |
__ASM_EMIT("2:") |
| 57 |
__ASM_EMIT("add %[i], %[i], #1") // i++ |
| 58 |
__ASM_EMIT("cmp %[i], %[count]") // i <=> count |
| 59 |
__ASM_EMIT("b.lo 1b") |
| 60 |
// Perform x8 butterflies |
| 61 |
__ASM_EMIT("3:") |
| 62 |
__ASM_EMIT("ld4 {v0.4s-v3.4s}, [%[dst]]") |
| 63 |
// v0 = r0 r2 r4 r6 |
| 64 |
// v1 = i0 i2 i4 i6 |
| 65 |
// v2 = r1 r3 r5 r7 |
| 66 |
// v3 = i1 i3 i5 i7 |
| 67 |
__ASM_EMIT("fadd v6.4s, v0.4s, v2.4s") // v6 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' |
| 68 |
__ASM_EMIT("fadd v7.4s, v1.4s, v3.4s") // v7 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' |
| 69 |
__ASM_EMIT("fsub v0.4s, v0.4s, v2.4s") // v0 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' |
| 70 |
__ASM_EMIT("fsub v1.4s, v1.4s, v3.4s") // v1 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' |
| 71 |
// v6 = r0' r2' r4' r6' |
| 72 |
// v0 = r1' r3' r5' r7' |
| 73 |
// v7 = i0' i2' i4' i6' |
| 74 |
// v1 = i1' i3' i5' i7' |
| 75 |
__ASM_EMIT("uzp1 v4.4s, v6.4s, v0.4s") // v4 = r0' r4' r1' r5' |
| 76 |
__ASM_EMIT("uzp2 v0.4s, v6.4s, v0.4s") // v0 = r2' r6' r3' r7' |
| 77 |
__ASM_EMIT("uzp1 v5.4s, v7.4s, v1.4s") // v5 = i0' i4' i1' i5' |
| 78 |
__ASM_EMIT("uzp2 v1.4s, v7.4s, v1.4s") // v1 = i2' i6' i3' i7' |
| 79 |
__ASM_EMIT("mov v2.16b, v0.16b") // v2 = r2' r6' r3' r7' |
| 80 |
__ASM_EMIT("ins v0.d[1], v1.d[1]") // v0 = r2' r6' i3' i7' |
| 81 |
__ASM_EMIT("ins v1.d[1], v2.d[1]") // v1 = i2' i6' r3' r7' |
| 82 |
__ASM_EMIT("fadd v2.4s, v4.4s, v0.4s") // v2 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r1" r5" |
| 83 |
__ASM_EMIT("fsub v3.4s, v4.4s, v0.4s") // v3 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r3" r7" |
| 84 |
__ASM_EMIT("fadd v0.4s, v5.4s, v1.4s") // v0 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i3" i7" |
| 85 |
__ASM_EMIT("fsub v1.4s, v5.4s, v1.4s") // v1 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i1" i5" |
| 86 |
// v0 = i0" i4" i3" i7" |
| 87 |
// v1 = i2" i6" i1" i5" |
| 88 |
// v2 = r0" r4" r1" r5" |
| 89 |
// v3 = r2" r6" r3" r7" |
| 90 |
__ASM_EMIT("mov v4.16b, v0.16b") // v4 = i0" i4" i3" i7" |
| 91 |
__ASM_EMIT("ins v0.d[1], v1.d[1]") // v0 = i0" i4" i1" i5" |
| 92 |
__ASM_EMIT("ins v1.d[1], v4.d[1]") // v1 = i2" i6" i3" i7" |
| 93 |
__ASM_EMIT("uzp1 v4.4s, v2.4s, v3.4s") // v4 = r0" r1" r2" r3" |
| 94 |
__ASM_EMIT("uzp2 v5.4s, v2.4s, v3.4s") // v5 = r4" r5" r6" r7" |
| 95 |
__ASM_EMIT("uzp1 v6.4s, v0.4s, v1.4s") // v6 = i0" i1" i2" i3" |
| 96 |
__ASM_EMIT("uzp2 v7.4s, v0.4s, v1.4s") // v7 = i4" i5" i6" i7" |
| 97 |
// Store |
| 98 |
__ASM_EMIT("stp q4, q6, [%[dst], #0x00]") |
| 99 |
__ASM_EMIT("stp q5, q7, [%[dst], #0x20]") |
| 100 |
__ASM_EMIT("subs %[count], %[count], #8") // i <=> count |
| 101 |
__ASM_EMIT("add %[dst], %[dst], #0x40") |
| 102 |
__ASM_EMIT("b.ne 3b") |
| 103 |
|
| 104 |
: [src] "+r" (src), [dst] "+r" (dst), [dp] "=&r" (dp), |
| 105 |
[rrank] "+r" (rrank), [i] "=&r" (i), [j] "=&r" (j), |
| 106 |
[count] "+r" (count) |
| 107 |
: |
| 108 |
: "cc", "memory", |
| 109 |
"v0", "v1", "v2", "v3", |
| 110 |
"v4", "v5", "v6", "v7" |
| 111 |
); |
| 112 |
} |
| 113 |
|
| 114 |
void packed_scramble_copy_direct(float *dst, const float *src, size_t rank) |
| 115 |
{ |
| 116 |
IF_ARCH_AARCH64( |
| 117 |
size_t i, j; |
| 118 |
float *sr1, *sr2; |
| 119 |
size_t regs = 1 << (rank - 3), rrank = 64 - (rank - 3); |
| 120 |
); |
| 121 |
|
| 122 |
ARCH_AARCH64_ASM( |
| 123 |
__ASM_EMIT("eor %[i], %[i], %[i]") // i = 0 |
| 124 |
__ASM_EMIT("1:") |
| 125 |
__ASM_EMIT("rbit %[j], %[i]") // j = reverse_bits(i) |
| 126 |
__ASM_EMIT("lsr %[j], %[j], %[rrank]") // j = reverse_bits(i) >> rank |
| 127 |
|
| 128 |
__ASM_EMIT("add %[sr1], %[src], %[j], LSL #3") // sr1 = &src[j] |
| 129 |
__ASM_EMIT("add %[sr2], %[sr1], %[regs], LSL #3") // sr2 = &src[j + regs] |
| 130 |
__ASM_EMIT("ld2 {v0.s, v1.s}[0], [%[sr1]]") // v0 = r0 ? ? ?, v1 = i0 ? ? ? |
| 131 |
__ASM_EMIT("ld2 {v0.s, v1.s}[2], [%[sr2]]") // v0 = r0 ? r4 ?, v1 = i0 ? i4 ? |
| 132 |
__ASM_EMIT("add %[sr1], %[sr1], %[regs], LSL #4") // sr1 = &src[j + regs*2] |
| 133 |
__ASM_EMIT("add %[sr2], %[sr2], %[regs], LSL #4") // sr2 = &src[j + regs*3] |
| 134 |
__ASM_EMIT("ld2 {v0.s, v1.s}[1], [%[sr1]]") // v0 = r0 r2 r4 ?, v1 = i0 i2 i4 ? |
| 135 |
__ASM_EMIT("ld2 {v0.s, v1.s}[3], [%[sr2]]") // v0 = r0 r2 r4 r6, v1 = i0 i2 i4 i6 |
| 136 |
|
| 137 |
__ASM_EMIT("add %[sr1], %[sr1], %[regs], LSL #4") // sr1 = &src[j + regs*4] |
| 138 |
__ASM_EMIT("add %[sr2], %[sr2], %[regs], LSL #4") // sr2 = &src[j + regs*5] |
| 139 |
__ASM_EMIT("ld2 {v2.s, v3.s}[0], [%[sr1]]") // v2 = r1 ? ? ?, v3 = i1 ? ? ? |
| 140 |
__ASM_EMIT("ld2 {v2.s, v3.s}[2], [%[sr2]]") // v2 = r1 ? r5 ?, v3 = i1 ? i5 ? |
| 141 |
__ASM_EMIT("add %[sr1], %[sr1], %[regs], LSL #4") // sr1 = &src[j + regs*6] |
| 142 |
__ASM_EMIT("add %[sr2], %[sr2], %[regs], LSL #4") // sr2 = &src[j + regs*7] |
| 143 |
__ASM_EMIT("ld2 {v2.s, v3.s}[1], [%[sr1]]") // v2 = r1 r3 r5 ?, v3 = i1 i3 i5 ? |
| 144 |
__ASM_EMIT("ld2 {v2.s, v3.s}[3], [%[sr2]]") // v2 = r1 r3 r5 r7, v3 = i1 i3 i5 i7 |
| 145 |
// v0 = r0 r2 r4 r6 |
| 146 |
// v1 = i0 i2 i4 i6 |
| 147 |
// v2 = r1 r3 r5 r7 |
| 148 |
// v3 = i1 i3 i5 i7 |
| 149 |
__ASM_EMIT("fadd v6.4s, v0.4s, v2.4s") // v6 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' |
| 150 |
__ASM_EMIT("fadd v7.4s, v1.4s, v3.4s") // v7 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' |
| 151 |
__ASM_EMIT("fsub v0.4s, v0.4s, v2.4s") // v0 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' |
| 152 |
__ASM_EMIT("fsub v1.4s, v1.4s, v3.4s") // v1 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' |
| 153 |
// v6 = r0' r2' r4' r6' |
| 154 |
// v0 = r1' r3' r5' r7' |
| 155 |
// v7 = i0' i2' i4' i6' |
| 156 |
// v1 = i1' i3' i5' i7' |
| 157 |
__ASM_EMIT("uzp1 v4.4s, v6.4s, v0.4s") // v4 = r0' r4' r1' r5' |
| 158 |
__ASM_EMIT("uzp2 v0.4s, v6.4s, v0.4s") // v0 = r2' r6' r3' r7' |
| 159 |
__ASM_EMIT("uzp1 v5.4s, v7.4s, v1.4s") // v5 = i0' i4' i1' i5' |
| 160 |
__ASM_EMIT("uzp2 v1.4s, v7.4s, v1.4s") // v1 = i2' i6' i3' i7' |
| 161 |
__ASM_EMIT("mov v2.16b, v0.16b") // v2 = r2' r6' r3' r7' |
| 162 |
__ASM_EMIT("ins v0.d[1], v1.d[1]") // v0 = r2' r6' i3' i7' |
| 163 |
__ASM_EMIT("ins v1.d[1], v2.d[1]") // v1 = i2' i6' r3' r7' |
| 164 |
__ASM_EMIT("fadd v2.4s, v4.4s, v0.4s") // v2 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r1" r5" |
| 165 |
__ASM_EMIT("fsub v3.4s, v4.4s, v0.4s") // v3 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r3" r7" |
| 166 |
__ASM_EMIT("fadd v0.4s, v5.4s, v1.4s") // v0 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i3" i7" |
| 167 |
__ASM_EMIT("fsub v1.4s, v5.4s, v1.4s") // v1 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i1" i5" |
| 168 |
// v0 = i0" i4" i3" i7" |
| 169 |
// v1 = i2" i6" i1" i5" |
| 170 |
// v2 = r0" r4" r1" r5" |
| 171 |
// v3 = r2" r6" r3" r7" |
| 172 |
__ASM_EMIT("mov v4.16b, v0.16b") // v4 = i0" i4" i3" i7" |
| 173 |
__ASM_EMIT("ins v0.d[1], v1.d[1]") // v0 = i0" i4" i1" i5" |
| 174 |
__ASM_EMIT("ins v1.d[1], v4.d[1]") // v1 = i2" i6" i3" i7" |
| 175 |
__ASM_EMIT("uzp1 v4.4s, v2.4s, v3.4s") // v4 = r0" r1" r2" r3" |
| 176 |
__ASM_EMIT("uzp2 v5.4s, v2.4s, v3.4s") // v5 = r4" r5" r6" r7" |
| 177 |
__ASM_EMIT("uzp1 v6.4s, v0.4s, v1.4s") // v6 = i0" i1" i2" i3" |
| 178 |
__ASM_EMIT("uzp2 v7.4s, v0.4s, v1.4s") // v7 = i4" i5" i6" i7" |
| 179 |
// Store |
| 180 |
__ASM_EMIT("add %[i], %[i], #1") |
| 181 |
__ASM_EMIT("stp q4, q6, [%[dst], #0x00]") |
| 182 |
__ASM_EMIT("stp q5, q7, [%[dst], #0x20]") |
| 183 |
__ASM_EMIT("cmp %[i], %[regs]") |
| 184 |
__ASM_EMIT("add %[dst], %[dst], #0x40") |
| 185 |
__ASM_EMIT("b.lo 1b") |
| 186 |
|
| 187 |
: [dst] "+r" (dst), [sr1] "=&r" (sr1), [sr2] "=&r" (sr2), |
| 188 |
[rrank] "+r" (rrank), [i] "=&r" (i), [j] "=&r" (j) |
| 189 |
: [src] "r" (src), [regs] "r" (regs) |
| 190 |
: "cc", "memory", |
| 191 |
"v0", "v1", "v2", "v3", |
| 192 |
"v4", "v5", "v6", "v7" |
| 193 |
); |
| 194 |
} |
| 195 |
|
| 196 |
void packed_scramble_self_reverse(float *dst, size_t rank) |
| 197 |
{ |
| 198 |
IF_ARCH_AARCH64( |
| 199 |
size_t count = 1 << rank; |
| 200 |
size_t i, j, rrank = 64 - rank; |
| 201 |
float *dp, *src = dst; |
| 202 |
); |
| 203 |
|
| 204 |
// Self algorithm |
| 205 |
ARCH_AARCH64_ASM( |
| 206 |
// Do bit-reverse shuffle |
| 207 |
__ASM_EMIT("mov %[i], #1") // i = 1 |
| 208 |
__ASM_EMIT("1:") |
| 209 |
__ASM_EMIT("rbit %[j], %[i]") // j = reverse_bits(i) |
| 210 |
__ASM_EMIT("add %[src], %[src], #8") |
| 211 |
__ASM_EMIT("lsr %[j], %[j], %[rrank]") // j = reverse_bits(i) >> rank |
| 212 |
__ASM_EMIT("cmp %[i], %[j]") // i <=> j |
| 213 |
__ASM_EMIT("b.hs 2f") // if (i >= j) continue |
| 214 |
__ASM_EMIT("add %[dp], %[dst], %[j], LSL #3") // d_re = &dst[j] |
| 215 |
__ASM_EMIT("ldr d0, [%[src]]") // d0 = *src |
| 216 |
__ASM_EMIT("ldr d1, [%[dp]]") // d1 = *dp |
| 217 |
__ASM_EMIT("str d0, [%[dp]]") // *dp = d0 |
| 218 |
__ASM_EMIT("str d1, [%[src]]") // *src = d1 |
| 219 |
__ASM_EMIT("2:") |
| 220 |
__ASM_EMIT("add %[i], %[i], #1") // i++ |
| 221 |
__ASM_EMIT("cmp %[i], %[count]") // i <=> count |
| 222 |
__ASM_EMIT("b.lo 1b") |
| 223 |
// Perform x8 butterflies |
| 224 |
__ASM_EMIT("3:") |
| 225 |
__ASM_EMIT("ld4 {v0.4s-v3.4s}, [%[dst]]") |
| 226 |
// v0 = r0 r2 r4 r6 |
| 227 |
// v1 = i0 i2 i4 i6 |
| 228 |
// v2 = r1 r3 r5 r7 |
| 229 |
// v3 = i1 i3 i5 i7 |
| 230 |
__ASM_EMIT("fadd v6.4s, v0.4s, v2.4s") // v6 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' |
| 231 |
__ASM_EMIT("fadd v7.4s, v1.4s, v3.4s") // v7 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' |
| 232 |
__ASM_EMIT("fsub v0.4s, v0.4s, v2.4s") // v0 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' |
| 233 |
__ASM_EMIT("fsub v1.4s, v1.4s, v3.4s") // v1 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' |
| 234 |
// v6 = r0' r2' r4' r6' |
| 235 |
// v0 = r1' r3' r5' r7' |
| 236 |
// v7 = i0' i2' i4' i6' |
| 237 |
// v1 = i1' i3' i5' i7' |
| 238 |
__ASM_EMIT("uzp1 v4.4s, v6.4s, v0.4s") // v4 = r0' r4' r1' r5' |
| 239 |
__ASM_EMIT("uzp2 v0.4s, v6.4s, v0.4s") // v0 = r2' r6' r3' r7' |
| 240 |
__ASM_EMIT("uzp1 v5.4s, v7.4s, v1.4s") // v5 = i0' i4' i1' i5' |
| 241 |
__ASM_EMIT("uzp2 v1.4s, v7.4s, v1.4s") // v1 = i2' i6' i3' i7' |
| 242 |
__ASM_EMIT("mov v2.16b, v0.16b") // v2 = r2' r6' r3' r7' |
| 243 |
__ASM_EMIT("ins v0.d[1], v1.d[1]") // v0 = r2' r6' i3' i7' |
| 244 |
__ASM_EMIT("ins v1.d[1], v2.d[1]") // v1 = i2' i6' r3' r7' |
| 245 |
__ASM_EMIT("fadd v2.4s, v4.4s, v0.4s") // v2 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" |
| 246 |
__ASM_EMIT("fsub v3.4s, v4.4s, v0.4s") // v3 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" |
| 247 |
__ASM_EMIT("fadd v0.4s, v5.4s, v1.4s") // v0 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" |
| 248 |
__ASM_EMIT("fsub v1.4s, v5.4s, v1.4s") // v1 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" |
| 249 |
// v0 = i0" i4" i1" i5" |
| 250 |
// v1 = i2" i6" i3" i7" |
| 251 |
// v2 = r0" r4" r3" r7" |
| 252 |
// v3 = r2" r6" r1" r5" |
| 253 |
__ASM_EMIT("mov v4.16b, v2.16b") // v4 = r0" r4" r3" r7" |
| 254 |
__ASM_EMIT("ins v2.d[1], v3.d[1]") // v2 = r0" r4" r1" r5" |
| 255 |
__ASM_EMIT("ins v3.d[1], v4.d[1]") // v3 = r2" r6" r3" r7" |
| 256 |
__ASM_EMIT("uzp1 v4.4s, v2.4s, v3.4s") // v4 = r0" r1" r2" r3" |
| 257 |
__ASM_EMIT("uzp2 v5.4s, v2.4s, v3.4s") // v5 = r4" r5" r6" r7" |
| 258 |
__ASM_EMIT("uzp1 v6.4s, v0.4s, v1.4s") // v6 = i0" i1" i2" i3" |
| 259 |
__ASM_EMIT("uzp2 v7.4s, v0.4s, v1.4s") // v7 = i4" i5" i6" i7" |
| 260 |
// Store |
| 261 |
__ASM_EMIT("stp q4, q6, [%[dst], #0x00]") |
| 262 |
__ASM_EMIT("stp q5, q7, [%[dst], #0x20]") |
| 263 |
__ASM_EMIT("subs %[count], %[count], #8") // i <=> count |
| 264 |
__ASM_EMIT("add %[dst], %[dst], #0x40") |
| 265 |
__ASM_EMIT("b.ne 3b") |
| 266 |
|
| 267 |
: [src] "+r" (src), [dst] "+r" (dst), [dp] "=&r" (dp), |
| 268 |
[rrank] "+r" (rrank), [i] "=&r" (i), [j] "=&r" (j), |
| 269 |
[count] "+r" (count) |
| 270 |
: |
| 271 |
: "cc", "memory", |
| 272 |
"v0", "v1", "v2", "v3", |
| 273 |
"v4", "v5", "v6", "v7" |
| 274 |
); |
| 275 |
} |
| 276 |
|
| 277 |
void packed_scramble_copy_reverse(float *dst, const float *src, size_t rank) |
| 278 |
{ |
| 279 |
IF_ARCH_AARCH64( |
| 280 |
size_t i, j; |
| 281 |
float *sr1, *sr2; |
| 282 |
size_t regs = 1 << (rank - 3), rrank = 64 - (rank - 3); |
| 283 |
); |
| 284 |
|
| 285 |
ARCH_AARCH64_ASM( |
| 286 |
__ASM_EMIT("eor %[i], %[i], %[i]") // i = 0 |
| 287 |
__ASM_EMIT("1:") |
| 288 |
__ASM_EMIT("rbit %[j], %[i]") // j = reverse_bits(i) |
| 289 |
__ASM_EMIT("lsr %[j], %[j], %[rrank]") // j = reverse_bits(i) >> rank |
| 290 |
|
| 291 |
__ASM_EMIT("add %[sr1], %[src], %[j], LSL #3") // sr1 = &src[j] |
| 292 |
__ASM_EMIT("add %[sr2], %[sr1], %[regs], LSL #3") // sr2 = &src[j + regs] |
| 293 |
__ASM_EMIT("ld2 {v0.s, v1.s}[0], [%[sr1]]") // v0 = r0 ? ? ?, v1 = i0 ? ? ? |
| 294 |
__ASM_EMIT("ld2 {v0.s, v1.s}[2], [%[sr2]]") // v0 = r0 ? r4 ?, v1 = i0 ? i4 ? |
| 295 |
__ASM_EMIT("add %[sr1], %[sr1], %[regs], LSL #4") // sr1 = &src[j + regs*2] |
| 296 |
__ASM_EMIT("add %[sr2], %[sr2], %[regs], LSL #4") // sr2 = &src[j + regs*3] |
| 297 |
__ASM_EMIT("ld2 {v0.s, v1.s}[1], [%[sr1]]") // v0 = r0 r2 r4 ?, v1 = i0 i2 i4 ? |
| 298 |
__ASM_EMIT("ld2 {v0.s, v1.s}[3], [%[sr2]]") // v0 = r0 r2 r4 r6, v1 = i0 i2 i4 i6 |
| 299 |
|
| 300 |
__ASM_EMIT("add %[sr1], %[sr1], %[regs], LSL #4") // sr1 = &src[j + regs*4] |
| 301 |
__ASM_EMIT("add %[sr2], %[sr2], %[regs], LSL #4") // sr2 = &src[j + regs*5] |
| 302 |
__ASM_EMIT("ld2 {v2.s, v3.s}[0], [%[sr1]]") // v2 = r1 ? ? ?, v3 = i1 ? ? ? |
| 303 |
__ASM_EMIT("ld2 {v2.s, v3.s}[2], [%[sr2]]") // v2 = r1 ? r5 ?, v3 = i1 ? i5 ? |
| 304 |
__ASM_EMIT("add %[sr1], %[sr1], %[regs], LSL #4") // sr1 = &src[j + regs*6] |
| 305 |
__ASM_EMIT("add %[sr2], %[sr2], %[regs], LSL #4") // sr2 = &src[j + regs*7] |
| 306 |
__ASM_EMIT("ld2 {v2.s, v3.s}[1], [%[sr1]]") // v2 = r1 r3 r5 ?, v3 = i1 i3 i5 ? |
| 307 |
__ASM_EMIT("ld2 {v2.s, v3.s}[3], [%[sr2]]") // v2 = r1 r3 r5 r7, v3 = i1 i3 i5 i7 |
| 308 |
// v0 = r0 r2 r4 r6 |
| 309 |
// v1 = i0 i2 i4 i6 |
| 310 |
// v2 = r1 r3 r5 r7 |
| 311 |
// v3 = i1 i3 i5 i7 |
| 312 |
__ASM_EMIT("fadd v6.4s, v0.4s, v2.4s") // v6 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' |
| 313 |
__ASM_EMIT("fadd v7.4s, v1.4s, v3.4s") // v7 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' |
| 314 |
__ASM_EMIT("fsub v0.4s, v0.4s, v2.4s") // v0 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' |
| 315 |
__ASM_EMIT("fsub v1.4s, v1.4s, v3.4s") // v1 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' |
| 316 |
// v6 = r0' r2' r4' r6' |
| 317 |
// v0 = r1' r3' r5' r7' |
| 318 |
// v7 = i0' i2' i4' i6' |
| 319 |
// v1 = i1' i3' i5' i7' |
| 320 |
__ASM_EMIT("uzp1 v4.4s, v6.4s, v0.4s") // v4 = r0' r4' r1' r5' |
| 321 |
__ASM_EMIT("uzp2 v0.4s, v6.4s, v0.4s") // v0 = r2' r6' r3' r7' |
| 322 |
__ASM_EMIT("uzp1 v5.4s, v7.4s, v1.4s") // v5 = i0' i4' i1' i5' |
| 323 |
__ASM_EMIT("uzp2 v1.4s, v7.4s, v1.4s") // v1 = i2' i6' i3' i7' |
| 324 |
__ASM_EMIT("mov v2.16b, v0.16b") // v2 = r2' r6' r3' r7' |
| 325 |
__ASM_EMIT("ins v0.d[1], v1.d[1]") // v0 = r2' r6' i3' i7' |
| 326 |
__ASM_EMIT("ins v1.d[1], v2.d[1]") // v1 = i2' i6' r3' r7' |
| 327 |
__ASM_EMIT("fadd v2.4s, v4.4s, v0.4s") // v2 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" |
| 328 |
__ASM_EMIT("fsub v3.4s, v4.4s, v0.4s") // v3 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" |
| 329 |
__ASM_EMIT("fadd v0.4s, v5.4s, v1.4s") // v0 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" |
| 330 |
__ASM_EMIT("fsub v1.4s, v5.4s, v1.4s") // v1 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" |
| 331 |
// v0 = i0" i4" i1" i5" |
| 332 |
// v1 = i2" i6" i3" i7" |
| 333 |
// v2 = r0" r4" r3" r7" |
| 334 |
// v3 = r2" r6" r1" r5" |
| 335 |
__ASM_EMIT("mov v4.16b, v2.16b") // v4 = r0" r4" r3" r7" |
| 336 |
__ASM_EMIT("ins v2.d[1], v3.d[1]") // v2 = r0" r4" r1" r5" |
| 337 |
__ASM_EMIT("ins v3.d[1], v4.d[1]") // v3 = r2" r6" r3" r7" |
| 338 |
__ASM_EMIT("uzp1 v4.4s, v2.4s, v3.4s") // v4 = r0" r1" r2" r3" |
| 339 |
__ASM_EMIT("uzp2 v5.4s, v2.4s, v3.4s") // v5 = r4" r5" r6" r7" |
| 340 |
__ASM_EMIT("uzp1 v6.4s, v0.4s, v1.4s") // v6 = i0" i1" i2" i3" |
| 341 |
__ASM_EMIT("uzp2 v7.4s, v0.4s, v1.4s") // v7 = i4" i5" i6" i7" |
| 342 |
// Store |
| 343 |
__ASM_EMIT("add %[i], %[i], #1") |
| 344 |
__ASM_EMIT("stp q4, q6, [%[dst], #0x00]") |
| 345 |
__ASM_EMIT("stp q5, q7, [%[dst], #0x20]") |
| 346 |
__ASM_EMIT("cmp %[i], %[regs]") |
| 347 |
__ASM_EMIT("add %[dst], %[dst], #0x40") |
| 348 |
__ASM_EMIT("b.lo 1b") |
| 349 |
|
| 350 |
: [dst] "+r" (dst), [sr1] "=&r" (sr1), [sr2] "=&r" (sr2), |
| 351 |
[rrank] "+r" (rrank), [i] "=&r" (i), [j] "=&r" (j) |
| 352 |
: [src] "r" (src), [regs] "r" (regs) |
| 353 |
: "cc", "memory", |
| 354 |
"v0", "v1", "v2", "v3", |
| 355 |
"v4", "v5", "v6", "v7" |
| 356 |
); |
| 357 |
} |
| 358 |
|
| 359 |
void packed_unscramble_direct(float *dst, size_t rank) |
| 360 |
{ |
| 361 |
IF_ARCH_AARCH64( |
| 362 |
size_t count = 1 << rank; |
| 363 |
); |
| 364 |
|
| 365 |
ARCH_AARCH64_ASM( |
| 366 |
// 16x blocks |
| 367 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 368 |
__ASM_EMIT("b.lo 2f") |
| 369 |
__ASM_EMIT("1:") |
| 370 |
__ASM_EMIT("ldp q0, q16, [%[dst], #0x00]") |
| 371 |
__ASM_EMIT("ldp q2, q17, [%[dst], #0x20]") |
| 372 |
__ASM_EMIT("ldp q4, q18, [%[dst], #0x40]") |
| 373 |
__ASM_EMIT("ldp q6, q19, [%[dst], #0x60]") |
| 374 |
__ASM_EMIT("zip2 v1.4s, v0.4s, v16.4s") |
| 375 |
__ASM_EMIT("zip1 v0.4s, v0.4s, v16.4s") |
| 376 |
__ASM_EMIT("zip2 v3.4s, v2.4s, v17.4s") |
| 377 |
__ASM_EMIT("zip1 v2.4s, v2.4s, v17.4s") |
| 378 |
__ASM_EMIT("zip2 v5.4s, v4.4s, v18.4s") |
| 379 |
__ASM_EMIT("zip1 v4.4s, v4.4s, v18.4s") |
| 380 |
__ASM_EMIT("zip2 v7.4s, v6.4s, v19.4s") |
| 381 |
__ASM_EMIT("zip1 v6.4s, v6.4s, v19.4s") |
| 382 |
__ASM_EMIT("stp q0, q1, [%[dst], #0x00]") |
| 383 |
__ASM_EMIT("stp q2, q3, [%[dst], #0x20]") |
| 384 |
__ASM_EMIT("stp q4, q5, [%[dst], #0x40]") |
| 385 |
__ASM_EMIT("stp q6, q7, [%[dst], #0x60]") |
| 386 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 387 |
__ASM_EMIT("add %[dst], %[dst], #0x80") |
| 388 |
__ASM_EMIT("b.hs 1b") |
| 389 |
__ASM_EMIT("2:") |
| 390 |
// 8x block |
| 391 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 392 |
__ASM_EMIT("b.lt 4f") |
| 393 |
__ASM_EMIT("ldp q0, q16, [%[dst], #0x00]") |
| 394 |
__ASM_EMIT("ldp q2, q17, [%[dst], #0x20]") |
| 395 |
__ASM_EMIT("zip2 v1.4s, v0.4s, v16.4s") |
| 396 |
__ASM_EMIT("zip1 v0.4s, v0.4s, v16.4s") |
| 397 |
__ASM_EMIT("zip2 v3.4s, v2.4s, v17.4s") |
| 398 |
__ASM_EMIT("zip1 v2.4s, v2.4s, v17.4s") |
| 399 |
__ASM_EMIT("stp q0, q1, [%[dst], #0x00]") |
| 400 |
__ASM_EMIT("stp q2, q3, [%[dst], #0x20]") |
| 401 |
__ASM_EMIT("4:") |
| 402 |
: [dst] "+r" (dst), |
| 403 |
[count] "+r" (count) |
| 404 |
: |
| 405 |
: "cc", "memory", |
| 406 |
"v0", "v1", "v2", "v3", |
| 407 |
"v4", "v5", "v6", "v7", |
| 408 |
"v16", "v17", "v18", "v19" |
| 409 |
); |
| 410 |
} |
| 411 |
|
| 412 |
void packed_unscramble_reverse(float *dst, size_t rank) |
| 413 |
{ |
| 414 |
IF_ARCH_AARCH64( |
| 415 |
size_t count = 1 << rank; |
| 416 |
float k = 1.0f/(1 << rank); |
| 417 |
); |
| 418 |
|
| 419 |
ARCH_AARCH64_ASM( |
| 420 |
__ASM_EMIT("dup v31.4s, %[k].s[0]") // v31 = k |
| 421 |
// 16x blocks |
| 422 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 423 |
__ASM_EMIT("mov v30.16b, v31.16b") |
| 424 |
__ASM_EMIT("b.lo 2f") |
| 425 |
__ASM_EMIT("1:") |
| 426 |
__ASM_EMIT("ldp q0, q16, [%[dst], #0x00]") |
| 427 |
__ASM_EMIT("ldp q2, q17, [%[dst], #0x20]") |
| 428 |
__ASM_EMIT("ldp q4, q18, [%[dst], #0x40]") |
| 429 |
__ASM_EMIT("ldp q6, q19, [%[dst], #0x60]") |
| 430 |
__ASM_EMIT("zip2 v1.4s, v0.4s, v16.4s") |
| 431 |
__ASM_EMIT("zip1 v0.4s, v0.4s, v16.4s") |
| 432 |
__ASM_EMIT("fmul v1.4s, v1.4s, v30.4s") |
| 433 |
__ASM_EMIT("fmul v0.4s, v0.4s, v31.4s") |
| 434 |
__ASM_EMIT("zip2 v3.4s, v2.4s, v17.4s") |
| 435 |
__ASM_EMIT("zip1 v2.4s, v2.4s, v17.4s") |
| 436 |
__ASM_EMIT("fmul v3.4s, v3.4s, v30.4s") |
| 437 |
__ASM_EMIT("fmul v2.4s, v2.4s, v31.4s") |
| 438 |
__ASM_EMIT("zip2 v5.4s, v4.4s, v18.4s") |
| 439 |
__ASM_EMIT("zip1 v4.4s, v4.4s, v18.4s") |
| 440 |
__ASM_EMIT("fmul v5.4s, v5.4s, v30.4s") |
| 441 |
__ASM_EMIT("fmul v4.4s, v4.4s, v31.4s") |
| 442 |
__ASM_EMIT("zip2 v7.4s, v6.4s, v19.4s") |
| 443 |
__ASM_EMIT("zip1 v6.4s, v6.4s, v19.4s") |
| 444 |
__ASM_EMIT("fmul v7.4s, v7.4s, v30.4s") |
| 445 |
__ASM_EMIT("fmul v6.4s, v6.4s, v31.4s") |
| 446 |
__ASM_EMIT("stp q0, q1, [%[dst], #0x00]") |
| 447 |
__ASM_EMIT("stp q2, q3, [%[dst], #0x20]") |
| 448 |
__ASM_EMIT("stp q4, q5, [%[dst], #0x40]") |
| 449 |
__ASM_EMIT("stp q6, q7, [%[dst], #0x60]") |
| 450 |
__ASM_EMIT("subs %[count], %[count], #16") |
| 451 |
__ASM_EMIT("add %[dst], %[dst], #0x80") |
| 452 |
__ASM_EMIT("b.hs 1b") |
| 453 |
__ASM_EMIT("2:") |
| 454 |
// 8x block |
| 455 |
__ASM_EMIT("adds %[count], %[count], #8") |
| 456 |
__ASM_EMIT("b.lt 4f") |
| 457 |
__ASM_EMIT("ldp q0, q16, [%[dst], #0x00]") |
| 458 |
__ASM_EMIT("ldp q2, q17, [%[dst], #0x20]") |
| 459 |
__ASM_EMIT("zip2 v1.4s, v0.4s, v16.4s") |
| 460 |
__ASM_EMIT("zip1 v0.4s, v0.4s, v16.4s") |
| 461 |
__ASM_EMIT("fmul v1.4s, v1.4s, v30.4s") |
| 462 |
__ASM_EMIT("fmul v0.4s, v0.4s, v31.4s") |
| 463 |
__ASM_EMIT("zip2 v3.4s, v2.4s, v17.4s") |
| 464 |
__ASM_EMIT("zip1 v2.4s, v2.4s, v17.4s") |
| 465 |
__ASM_EMIT("fmul v3.4s, v3.4s, v30.4s") |
| 466 |
__ASM_EMIT("fmul v2.4s, v2.4s, v31.4s") |
| 467 |
__ASM_EMIT("stp q0, q1, [%[dst], #0x00]") |
| 468 |
__ASM_EMIT("stp q2, q3, [%[dst], #0x20]") |
| 469 |
__ASM_EMIT("4:") |
| 470 |
: [dst] "+r" (dst), |
| 471 |
[count] "+r" (count), |
| 472 |
[k] "+w" (k) |
| 473 |
: |
| 474 |
: "cc", "memory", |
| 475 |
"v0", "v1", "v2", "v3", |
| 476 |
"v4", "v5", "v6", "v7", |
| 477 |
"v16", "v17", "v18", "v19", |
| 478 |
"v30", "v31" |
| 479 |
); |
| 480 |
} |
| 481 |
} |
| 482 |
} |
| 483 |
|
| 484 |
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_FFT_PSCRAMBLE_H_ */ |