| 1 |
/* |
| 2 |
* Copyright (C) 2023 Linux Studio Plugins Project <https://lsp-plug.in/> |
| 3 |
* (C) 2023 Vladimir Sadovnikov <[email protected]> |
| 4 |
* |
| 5 |
* This file is part of lsp-dsp-lib |
| 6 |
* Created on: 20 окт. 2023 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_DYNAMICS_GATE_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_DYNAMICS_GATE_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 |
#include <private/dsp/arch/aarch64/asimd/pmath/exp.h> |
| 30 |
#include <private/dsp/arch/aarch64/asimd/pmath/log.h> |
| 31 |
|
| 32 |
namespace lsp |
| 33 |
{ |
| 34 |
namespace asimd |
| 35 |
{ |
| 36 |
#define PROCESS_KNEE_SINGLE_X8 \ |
| 37 |
/* in: v0 = lx0, v1 = lx1 */ \ |
| 38 |
__ASM_EMIT("add %[off], %[gate], #0x10") \ |
| 39 |
__ASM_EMIT("ld4r {v8.4s, v9.4s, v10.4s, v11.4s}, [%[off]]" ) /* v8=herm[0], v9=herm[1], v10=herm[2], v11=herm[3] */ \ |
| 40 |
__ASM_EMIT("fmul v2.4s, v0.4s, v8.4s") /* v2 = herm[0]*lx0 */ \ |
| 41 |
__ASM_EMIT("fmul v3.4s, v1.4s, v8.4s") \ |
| 42 |
__ASM_EMIT("fadd v2.4s, v2.4s, v9.4s") /* v2 = herm[0]*lx0+herm[1] */ \ |
| 43 |
__ASM_EMIT("fadd v3.4s, v3.4s, v9.4s") \ |
| 44 |
__ASM_EMIT("fmul v2.4s, v2.4s, v0.4s") /* v2 = (herm[0]*lx0+herm[1])*lx0 */ \ |
| 45 |
__ASM_EMIT("fmul v3.4s, v3.4s, v1.4s") \ |
| 46 |
__ASM_EMIT("fadd v2.4s, v2.4s, v10.4s") /* v2 = (herm[0]*lx0+herm[1])*lx0+herm[2] */ \ |
| 47 |
__ASM_EMIT("fadd v3.4s, v3.4s, v10.4s") \ |
| 48 |
__ASM_EMIT("fmul v2.4s, v2.4s, v0.4s") /* v2 = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0 */ \ |
| 49 |
__ASM_EMIT("fmul v3.4s, v3.4s, v1.4s") \ |
| 50 |
__ASM_EMIT("fadd v0.4s, v2.4s, v11.4s") /* v0 = TV = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0+herm[3] */ \ |
| 51 |
__ASM_EMIT("fadd v1.4s, v3.4s, v11.4s") \ |
| 52 |
EXP_CORE_X8 /* v0 = KV = expf(TV) */ \ |
| 53 |
__ASM_EMIT("ldp q6, q7, [%[mem], #0x00]") /* v6 = x0 */ \ |
| 54 |
__ASM_EMIT("ld4r {v2.4s, v3.4s, v4.4s, v5.4s}, [%[gate]]" ) /* v2=start, v3=end, v4=gain_start, v5=gain_end */ \ |
| 55 |
__ASM_EMIT("fcmgt v8.4s, v6.4s, v2.4s") /* v8 = [x0 > start] */ \ |
| 56 |
__ASM_EMIT("fcmgt v9.4s, v7.4s, v2.4s") \ |
| 57 |
__ASM_EMIT("fcmge v6.4s, v6.4s, v3.4s") /* v6 = [x0 >= end] */ \ |
| 58 |
__ASM_EMIT("fcmge v7.4s, v7.4s, v3.4s") \ |
| 59 |
__ASM_EMIT("bif v0.16b, v4.16b, v8.16b") /* v0 = [x0 <= start] ? gain_start : KV */ \ |
| 60 |
__ASM_EMIT("bif v1.16b, v4.16b, v9.16b") \ |
| 61 |
__ASM_EMIT("bit v0.16b, v5.16b, v6.16b") /* v0 = [x0 >= end] ? gain_end : [x0 <= start] ? gain_start : KV */ \ |
| 62 |
__ASM_EMIT("bit v1.16b, v5.16b, v7.16b") \ |
| 63 |
/* out: v0 = g0, v1 = g1 */ |
| 64 |
|
| 65 |
#define PROCESS_KNEE_SINGLE_X4 \ |
| 66 |
/* in: v0 = lx0 */ \ |
| 67 |
__ASM_EMIT("add %[off], %[gate], #0x10") \ |
| 68 |
__ASM_EMIT("ld4r {v8.4s, v9.4s, v10.4s, v11.4s}, [%[off]]" ) /* v8=herm[0], v9=herm[1], v10=herm[2], v11=herm[3] */ \ |
| 69 |
__ASM_EMIT("fmul v2.4s, v0.4s, v8.4s") /* v2 = herm[0]*lx0 */ \ |
| 70 |
__ASM_EMIT("fadd v2.4s, v2.4s, v9.4s") /* v2 = herm[0]*lx0+herm[1] */ \ |
| 71 |
__ASM_EMIT("fmul v2.4s, v2.4s, v0.4s") /* v2 = (herm[0]*lx0+herm[1])*lx0 */ \ |
| 72 |
__ASM_EMIT("fadd v2.4s, v2.4s, v10.4s") /* v2 = (herm[0]*lx0+herm[1])*lx0+herm[2] */ \ |
| 73 |
__ASM_EMIT("fmul v2.4s, v2.4s, v0.4s") /* v2 = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0 */ \ |
| 74 |
__ASM_EMIT("fadd v0.4s, v2.4s, v11.4s") /* v0 = TV = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0+herm[3] */ \ |
| 75 |
EXP_CORE_X4 /* v0 = KV = expf(TV) */ \ |
| 76 |
__ASM_EMIT("ldp q6, q7, [%[mem], #0x00]") /* v6 = x0 */ \ |
| 77 |
__ASM_EMIT("ld4r {v2.4s, v3.4s, v4.4s, v5.4s}, [%[gate]]" ) /* v2=start, v3=end, v4=gain_start, v5=gain_end */ \ |
| 78 |
__ASM_EMIT("fcmgt v8.4s, v6.4s, v2.4s") /* v8 = [x0 > start] */ \ |
| 79 |
__ASM_EMIT("fcmge v6.4s, v6.4s, v3.4s") /* v6 = [x0 >= end] */ \ |
| 80 |
__ASM_EMIT("bif v0.16b, v4.16b, v8.16b") /* v0 = [x0 <= start] ? gain_start : KV */ \ |
| 81 |
__ASM_EMIT("bit v0.16b, v5.16b, v6.16b") /* v0 = [x0 >= end] ? gain_end : [x0 <= start] ? gain_start : KV */ \ |
| 82 |
/* out: v0 = g0 */ |
| 83 |
|
| 84 |
#define PROCESS_GATE_FULL_X8 \ |
| 85 |
/* in: v0 = x0, v1 = x1 */ \ |
| 86 |
__ASM_EMIT("ld2r {v12.4s, v13.4s}, [%[gate]]") /* v12 = begin, v13 = end */ \ |
| 87 |
__ASM_EMIT("fabs v0.4s, v0.4s") /* v0 = fabsf(x0) */ \ |
| 88 |
__ASM_EMIT("fabs v1.4s, v1.4s") \ |
| 89 |
__ASM_EMIT("fcmgt v2.4s, v0.4s, v12.4s") /* v2 = [fabs(x0) > begin] */ \ |
| 90 |
__ASM_EMIT("fcmgt v3.4s, v1.4s, v12.4s") \ |
| 91 |
__ASM_EMIT("fcmgt v4.4s, v13.4s, v0.4s") /* v4 = [fabs(x0) < end] */ \ |
| 92 |
__ASM_EMIT("fcmgt v5.4s, v13.4s, v1.4s") \ |
| 93 |
__ASM_EMIT("and v6.16b, v2.16b, v4.16b") /* v6 = [fabs(x0) > begin] && [fabs(x0) < end] */ \ |
| 94 |
__ASM_EMIT("and v7.16b, v3.16b, v5.16b") \ |
| 95 |
__ASM_EMIT("orr v6.16b, v6.16b, v7.16b") \ |
| 96 |
__ASM_EMIT("ext v7.16b, v6.16b, v6.16b, #8") \ |
| 97 |
__ASM_EMIT("orr v6.16b, v6.16b, v7.16b") \ |
| 98 |
__ASM_EMIT("ext v7.16b, v6.16b, v6.16b, #4") \ |
| 99 |
__ASM_EMIT("orr v6.16b, v6.16b, v7.16b") \ |
| 100 |
__ASM_EMIT("mov %w[off], v6.s[0]") \ |
| 101 |
__ASM_EMIT("cmp %w[off], #0") \ |
| 102 |
__ASM_EMIT("b.ne 100f") \ |
| 103 |
__ASM_EMIT("add %[off], %[gate], #0x08") \ |
| 104 |
__ASM_EMIT("ld1r {v0.4s}, [%[off]]") /* v0 = gain_begin */ \ |
| 105 |
__ASM_EMIT("add %[off], %[gate], #0x0c") \ |
| 106 |
__ASM_EMIT("ld1r {v4.4s}, [%[off]]") /* v4 = gain_end */ \ |
| 107 |
__ASM_EMIT("mov v1.16b, v0.16b") \ |
| 108 |
__ASM_EMIT("bit v0.16b, v4.16b, v2.16b") /* v0 = [fabs(x0) > start] ? gain_end : gain_start */ \ |
| 109 |
__ASM_EMIT("bit v1.16b, v4.16b, v3.16b") \ |
| 110 |
__ASM_EMIT("b 200f") \ |
| 111 |
__ASM_EMIT("100:") \ |
| 112 |
__ASM_EMIT("stp q0, q1, [%[mem], #0x00]") /* mem[0x00] = fabfs(x0) */ \ |
| 113 |
LOGE_CORE_X8 /* v0= lx0 = logf(fabsf(x0)) */ \ |
| 114 |
PROCESS_KNEE_SINGLE_X8 /* apply knee 0 */ \ |
| 115 |
__ASM_EMIT("200:") \ |
| 116 |
/* out: v0 = G0, v1= G1 */ |
| 117 |
|
| 118 |
#define PROCESS_GATE_FULL_X4 \ |
| 119 |
/* in: v0 = x0 */ \ |
| 120 |
__ASM_EMIT("fabs v0.4s, v0.4s") /* v0 = fabsf(x0) */ \ |
| 121 |
__ASM_EMIT("str q0, [%[mem], #0x00]") /* mem[0x00] = fabfs(x0) */ \ |
| 122 |
LOGE_CORE_X4 /* v0= lx0 = logf(fabsf(x0)) */ \ |
| 123 |
PROCESS_KNEE_SINGLE_X4 /* apply knee 0 */ \ |
| 124 |
/* out: v0 = G0 */ |
| 125 |
|
| 126 |
void gate_x1_gain(float *dst, const float *src, const dsp::gate_knee_t *c, size_t count) |
| 127 |
{ |
| 128 |
IF_ARCH_AARCH64( |
| 129 |
float mem[8] __lsp_aligned16; |
| 130 |
size_t off; |
| 131 |
); |
| 132 |
|
| 133 |
ARCH_AARCH64_ASM( |
| 134 |
// x8 blocks |
| 135 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 136 |
__ASM_EMIT("b.lo 2f") |
| 137 |
__ASM_EMIT("1:") |
| 138 |
__ASM_EMIT("ldp q0, q1, [%[src], #0x00]") |
| 139 |
PROCESS_GATE_FULL_X8 |
| 140 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 141 |
__ASM_EMIT("stp q0, q1, [%[dst], #0x00]") |
| 142 |
__ASM_EMIT("add %[src], %[src], 0x20") |
| 143 |
__ASM_EMIT("add %[dst], %[dst], 0x20") |
| 144 |
__ASM_EMIT("b.hs 1b") |
| 145 |
__ASM_EMIT("2:") |
| 146 |
// x4 block |
| 147 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 148 |
__ASM_EMIT("b.lt 4f") |
| 149 |
__ASM_EMIT("ldr q0, [%[src], #0x00]") |
| 150 |
PROCESS_GATE_FULL_X4 |
| 151 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 152 |
__ASM_EMIT("str q0, [%[dst], #0x00]") |
| 153 |
__ASM_EMIT("add %[src], %[src], 0x10") |
| 154 |
__ASM_EMIT("add %[dst], %[dst], 0x10") |
| 155 |
__ASM_EMIT("4:") |
| 156 |
// Tail: 1x-3x block |
| 157 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 158 |
__ASM_EMIT("b.ls 12f") |
| 159 |
__ASM_EMIT("tst %[count], #1") |
| 160 |
__ASM_EMIT("b.eq 6f") |
| 161 |
__ASM_EMIT("ld1 {v0.s}[0], [%[src]]") |
| 162 |
__ASM_EMIT("add %[src], %[src], #0x04") |
| 163 |
__ASM_EMIT("6:") |
| 164 |
__ASM_EMIT("tst %[count], #2") |
| 165 |
__ASM_EMIT("b.eq 8f") |
| 166 |
__ASM_EMIT("ld1 {v0.d}[1], [%[src]]") |
| 167 |
__ASM_EMIT("8:") |
| 168 |
PROCESS_GATE_FULL_X4 |
| 169 |
__ASM_EMIT("tst %[count], #1") |
| 170 |
__ASM_EMIT("b.eq 10f") |
| 171 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst]]") |
| 172 |
__ASM_EMIT("add %[dst], %[dst], #0x04") |
| 173 |
__ASM_EMIT("10:") |
| 174 |
__ASM_EMIT("tst %[count], #2") |
| 175 |
__ASM_EMIT("b.eq 12f") |
| 176 |
__ASM_EMIT("st1 {v0.d}[1], [%[dst]]") |
| 177 |
__ASM_EMIT("12:") |
| 178 |
|
| 179 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 180 |
[count] "+r" (count), |
| 181 |
[off] "=&r" (off) |
| 182 |
: [gate] "r" (c), |
| 183 |
[L2C] "r" (&LOG2_CONST[0]), |
| 184 |
[LOGC] "r" (&LOGE_C[0]), |
| 185 |
[mem] "r" (&mem[0]), |
| 186 |
[E2C] "r" (&EXP2_CONST[0]), |
| 187 |
[LOG2E] "r" (&EXP_LOG2E[0]) |
| 188 |
: "cc", "memory", |
| 189 |
"v0", "v1", "v2", "v3", |
| 190 |
"v4", "v5", "v6", "v7", |
| 191 |
"v8", "v9", "v10", "v11", |
| 192 |
"v12", "v13", "v14", "v15", |
| 193 |
"v16", "v17", "v18", "v19", |
| 194 |
"v20", "v21", "v22", "v23", |
| 195 |
"v24", "v25", "v26", "v27", |
| 196 |
"v28", "v29", "v30", "v31" |
| 197 |
); |
| 198 |
} |
| 199 |
|
| 200 |
void gate_x1_curve(float *dst, const float *src, const dsp::gate_knee_t *c, size_t count) |
| 201 |
{ |
| 202 |
IF_ARCH_AARCH64( |
| 203 |
float mem[8] __lsp_aligned16; |
| 204 |
size_t off; |
| 205 |
); |
| 206 |
|
| 207 |
ARCH_AARCH64_ASM( |
| 208 |
// x8 blocks |
| 209 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 210 |
__ASM_EMIT("b.lo 2f") |
| 211 |
__ASM_EMIT("1:") |
| 212 |
__ASM_EMIT("ldp q0, q1, [%[src], #0x00]") |
| 213 |
PROCESS_GATE_FULL_X8 |
| 214 |
__ASM_EMIT("ldp q2, q3, [%[src], #0x00]") |
| 215 |
__ASM_EMIT("subs %[count], %[count], #8") |
| 216 |
__ASM_EMIT("fmul v0.4s, v0.4s, v2.4s") |
| 217 |
__ASM_EMIT("fmul v1.4s, v1.4s, v3.4s") |
| 218 |
__ASM_EMIT("stp q0, q1, [%[dst], #0x00]") |
| 219 |
__ASM_EMIT("add %[src], %[src], 0x20") |
| 220 |
__ASM_EMIT("add %[dst], %[dst], 0x20") |
| 221 |
__ASM_EMIT("b.hs 1b") |
| 222 |
__ASM_EMIT("2:") |
| 223 |
// x4 block |
| 224 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 225 |
__ASM_EMIT("b.lt 4f") |
| 226 |
__ASM_EMIT("ldr q0, [%[src], #0x00]") |
| 227 |
PROCESS_GATE_FULL_X4 |
| 228 |
__ASM_EMIT("ldr q2, [%[src], #0x00]") |
| 229 |
__ASM_EMIT("sub %[count], %[count], #4") |
| 230 |
__ASM_EMIT("fmul v0.4s, v0.4s, v2.4s") |
| 231 |
__ASM_EMIT("str q0, [%[dst], #0x00]") |
| 232 |
__ASM_EMIT("add %[src], %[src], 0x10") |
| 233 |
__ASM_EMIT("add %[dst], %[dst], 0x10") |
| 234 |
__ASM_EMIT("4:") |
| 235 |
// Tail: 1x-3x block |
| 236 |
__ASM_EMIT("adds %[count], %[count], #4") |
| 237 |
__ASM_EMIT("b.ls 12f") |
| 238 |
__ASM_EMIT("tst %[count], #1") |
| 239 |
__ASM_EMIT("b.eq 6f") |
| 240 |
__ASM_EMIT("ld1 {v0.s}[0], [%[src]]") |
| 241 |
__ASM_EMIT("add %[src], %[src], #0x04") |
| 242 |
__ASM_EMIT("6:") |
| 243 |
__ASM_EMIT("tst %[count], #2") |
| 244 |
__ASM_EMIT("b.eq 8f") |
| 245 |
__ASM_EMIT("ld1 {v0.d}[1], [%[src]]") |
| 246 |
__ASM_EMIT("8:") |
| 247 |
__ASM_EMIT("mov v1.16b, v0.16b") |
| 248 |
PROCESS_GATE_FULL_X4 |
| 249 |
__ASM_EMIT("fmul v0.4s, v0.4s, v1.4s") |
| 250 |
__ASM_EMIT("tst %[count], #1") |
| 251 |
__ASM_EMIT("b.eq 10f") |
| 252 |
__ASM_EMIT("st1 {v0.s}[0], [%[dst]]") |
| 253 |
__ASM_EMIT("add %[dst], %[dst], #0x04") |
| 254 |
__ASM_EMIT("10:") |
| 255 |
__ASM_EMIT("tst %[count], #2") |
| 256 |
__ASM_EMIT("b.eq 12f") |
| 257 |
__ASM_EMIT("st1 {v0.d}[1], [%[dst]]") |
| 258 |
__ASM_EMIT("12:") |
| 259 |
|
| 260 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 261 |
[count] "+r" (count), |
| 262 |
[off] "=&r" (off) |
| 263 |
: [gate] "r" (c), |
| 264 |
[L2C] "r" (&LOG2_CONST[0]), |
| 265 |
[LOGC] "r" (&LOGE_C[0]), |
| 266 |
[mem] "r" (&mem[0]), |
| 267 |
[E2C] "r" (&EXP2_CONST[0]), |
| 268 |
[LOG2E] "r" (&EXP_LOG2E[0]) |
| 269 |
: "cc", "memory", |
| 270 |
"v0", "v1", "v2", "v3", |
| 271 |
"v4", "v5", "v6", "v7", |
| 272 |
"v8", "v9", "v10", "v11", |
| 273 |
"v12", "v13", "v14", "v15", |
| 274 |
"v16", "v17", "v18", "v19", |
| 275 |
"v20", "v21", "v22", "v23", |
| 276 |
"v24", "v25", "v26", "v27", |
| 277 |
"v28", "v29", "v30", "v31" |
| 278 |
); |
| 279 |
} |
| 280 |
|
| 281 |
#undef KNEE_LOAD |
| 282 |
#undef PROCESS_KNEE_SINGLE_X4 |
| 283 |
#undef PROCESS_KNEE_SINGLE_X8 |
| 284 |
#undef PROCESS_GATE_FULL_X8 |
| 285 |
#undef PROCESS_GATE_FULL_X4 |
| 286 |
|
| 287 |
} /* namespace asimd */ |
| 288 |
} /* namespace lsp */ |
| 289 |
|
| 290 |
|
| 291 |
|
| 292 |
|
| 293 |
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_DYNAMICS_GATE_H_ */ |