| 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_ARM_NEON_D32_DYNAMICS_GATE_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_ARM_NEON_D32_DYNAMICS_GATE_H_ |
| 24 |
|
| 25 |
#include <private/dsp/arch/arm/neon-d32/pmath/exp.h> |
| 26 |
#include <private/dsp/arch/arm/neon-d32/pmath/log.h> |
| 27 |
|
| 28 |
namespace lsp |
| 29 |
{ |
| 30 |
namespace neon_d32 |
| 31 |
{ |
| 32 |
#define PROCESS_KNEE_SINGLE_X8 \ |
| 33 |
/* in: q0 = lx0, q1 = lx1 */ \ |
| 34 |
__ASM_EMIT("add %[off], %[gate], #0x10") \ |
| 35 |
__ASM_EMIT("vld4.f32 {d16[], d18[], d20[], d22[]}, [%[off]]") /* q8 = herm[0] q9=herm[1] q10=herm[2], q11=herm[3] */ \ |
| 36 |
__ASM_EMIT("vld4.f32 {d17[], d19[], d21[], d23[]}, [%[off]]") /* q8 = herm[0] q9=herm[1] q10=herm[2], q11=herm[3] */ \ |
| 37 |
__ASM_EMIT("vmul.f32 q2, q0, q8") /* q2 = herm[0]*lx0 */ \ |
| 38 |
__ASM_EMIT("vmul.f32 q3, q1, q8") \ |
| 39 |
__ASM_EMIT("vadd.f32 q2, q2, q9") /* q2 = herm[0]*lx0+herm[1] */ \ |
| 40 |
__ASM_EMIT("vadd.f32 q3, q3, q9") \ |
| 41 |
__ASM_EMIT("vmul.f32 q2, q2, q0") /* q2 = (herm[0]*lx0+herm[1])*lx0 */ \ |
| 42 |
__ASM_EMIT("vmul.f32 q3, q3, q1") \ |
| 43 |
__ASM_EMIT("vadd.f32 q2, q2, q10") /* q2 = (herm[0]*lx0+herm[1])*lx0+herm[2] */ \ |
| 44 |
__ASM_EMIT("vadd.f32 q3, q3, q10") \ |
| 45 |
__ASM_EMIT("vldm %[LOG2E], {q15}") \ |
| 46 |
__ASM_EMIT("vmul.f32 q2, q2, q0") /* q2 = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0 */ \ |
| 47 |
__ASM_EMIT("vmul.f32 q3, q3, q1") \ |
| 48 |
__ASM_EMIT("vadd.f32 q0, q2, q11") /* q0 = TV = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0+herm[3] */ \ |
| 49 |
__ASM_EMIT("vadd.f32 q1, q3, q11") \ |
| 50 |
EXP_CORE_X8 /* q0 = KV = expf(TV) */ \ |
| 51 |
__ASM_EMIT("vldm %[mem], {q6-q7}") /* q6 = x0 */ \ |
| 52 |
__ASM_EMIT("vld4.f32 {d4[], d6[], d8[], d10[]}, [%[gate]]") /* q2 = start, q3 = end, q4 = gain_start, q5 = gain_end */ \ |
| 53 |
__ASM_EMIT("vld4.f32 {d5[], d7[], d9[], d11[]}, [%[gate]]") /* q2 = start, q3 = end, q4 = gain_start, q5 = gain_end */ \ |
| 54 |
__ASM_EMIT("vcgt.f32 q8, q6, q2") /* q8 = [x0 > start] */ \ |
| 55 |
__ASM_EMIT("vcgt.f32 q9, q7, q2") \ |
| 56 |
__ASM_EMIT("vcge.f32 q6, q6, q3") /* q3 = [x0 >= end] */ \ |
| 57 |
__ASM_EMIT("vcge.f32 q7, q7, q3") \ |
| 58 |
__ASM_EMIT("vbif q0, q4, q8") /* q0 = [x0 <= start] ? gain_start : KV */ \ |
| 59 |
__ASM_EMIT("vbif q1, q4, q9") \ |
| 60 |
__ASM_EMIT("vbit q0, q5, q6") /* q0 = [x0 >= end] ? gain_end : [x0 <= start] ? gain_start : KV */ \ |
| 61 |
__ASM_EMIT("vbit q1, q5, q7") \ |
| 62 |
/* out: q0 = g0, q1 = g1 */ |
| 63 |
|
| 64 |
#define PROCESS_KNEE_SINGLE_X4 \ |
| 65 |
/* in: q0 = lx0 */ \ |
| 66 |
__ASM_EMIT("add %[off], %[gate], #0x10") \ |
| 67 |
__ASM_EMIT("vld4.f32 {d16[], d18[], d20[], d22[]}, [%[off]]") /* q8 = herm[0] q9=herm[1] q10=herm[2], q11=herm[3] */ \ |
| 68 |
__ASM_EMIT("vld4.f32 {d17[], d19[], d21[], d23[]}, [%[off]]") /* q8 = herm[0] q9=herm[1] q10=herm[2], q11=herm[3] */ \ |
| 69 |
__ASM_EMIT("vmul.f32 q2, q0, q8") /* q2 = herm[0]*lx0 */ \ |
| 70 |
__ASM_EMIT("vadd.f32 q2, q2, q9") /* q2 = herm[0]*lx0+herm[1] */ \ |
| 71 |
__ASM_EMIT("vmul.f32 q2, q2, q0") /* q2 = (herm[0]*lx0+herm[1])*lx0 */ \ |
| 72 |
__ASM_EMIT("vadd.f32 q2, q2, q10") /* q2 = (herm[0]*lx0+herm[1])*lx0+herm[2] */ \ |
| 73 |
__ASM_EMIT("vldm %[LOG2E], {q15}") \ |
| 74 |
__ASM_EMIT("vmul.f32 q2, q2, q0") /* q2 = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0 */ \ |
| 75 |
__ASM_EMIT("vadd.f32 q0, q2, q11") /* q0 = TV = ((herm[0]*lx0+herm[1])*lx0+herm[2])*lx0+herm[3] */ \ |
| 76 |
EXP_CORE_X4 /* q0 = KV = expf(TV) */ \ |
| 77 |
__ASM_EMIT("vldm %[mem], {q6}") /* q6 = x0 */ \ |
| 78 |
__ASM_EMIT("vld4.f32 {d4[], d6[], d8[], d10[]}, [%[gate]]") /* q2 = start, q3 = end, q4 = gain_start, q5 = gain_end */ \ |
| 79 |
__ASM_EMIT("vld4.f32 {d5[], d7[], d9[], d11[]}, [%[gate]]") /* q2 = start, q3 = end, q4 = gain_start, q5 = gain_end */ \ |
| 80 |
__ASM_EMIT("vcgt.f32 q8, q6, q2") /* q8 = [x0 > start] */ \ |
| 81 |
__ASM_EMIT("vcge.f32 q6, q6, q3") /* q3 = [x0 >= end] */ \ |
| 82 |
__ASM_EMIT("vbif q0, q4, q8") /* q0 = [x0 <= start] ? gain_start : KV */ \ |
| 83 |
__ASM_EMIT("vbit q0, q5, q6") /* q0 = [x0 >= end] ? gain_end : [x0 <= start] ? gain_start : KV */ \ |
| 84 |
/* out: q0 = g0 */ |
| 85 |
|
| 86 |
#define PROCESS_GATE_FULL_X8 \ |
| 87 |
/* in: q0 = x0, q1 = x1 */ \ |
| 88 |
__ASM_EMIT("vld2.f32 {d26[], d28[]}, [%[gate]]") /* q13 = start, q14 = end */ \ |
| 89 |
__ASM_EMIT("vld2.f32 {d27[], d29[]}, [%[gate]]") /* q13 = start, q14 = end */ \ |
| 90 |
__ASM_EMIT("vabs.f32 q0, q0") /* q0 = fabsf(x0) */ \ |
| 91 |
__ASM_EMIT("vabs.f32 q1, q1") \ |
| 92 |
__ASM_EMIT("vcgt.f32 q2, q0, q13") /* q2 = [fabs(x0) > start] */ \ |
| 93 |
__ASM_EMIT("vcgt.f32 q3, q1, q13") \ |
| 94 |
__ASM_EMIT("vcgt.f32 q4, q14, q0") /* q4 = [fabs(x0) < end] */ \ |
| 95 |
__ASM_EMIT("vcgt.f32 q5, q14, q1") \ |
| 96 |
__ASM_EMIT("vand q6, q2, q4") /* q6 = [fabs(x0) > start] && [fabs(x0) < end] */ \ |
| 97 |
__ASM_EMIT("vand q7, q3, q5") \ |
| 98 |
__ASM_EMIT("vorr q6, q6, q7") \ |
| 99 |
__ASM_EMIT("vext.32 q7, q6, q6, #2") \ |
| 100 |
__ASM_EMIT("vorr q6, q6, q7") \ |
| 101 |
__ASM_EMIT("vext.32 q7, q6, q6, #1") \ |
| 102 |
__ASM_EMIT("vorr q6, q6, q7") \ |
| 103 |
__ASM_EMIT("vmov.32 %[off], d12[0]") \ |
| 104 |
__ASM_EMIT("cmp %[off], #0") \ |
| 105 |
__ASM_EMIT("bne 100f") \ |
| 106 |
__ASM_EMIT("add %[off], %[gate], #0x08") \ |
| 107 |
__ASM_EMIT("vld1.f32 {d0[], d1[]}, [%[off]]") /* q0 = gain_start */ \ |
| 108 |
__ASM_EMIT("add %[off], %[gate], #0x0c") \ |
| 109 |
__ASM_EMIT("vld1.f32 {d8[], d9[]}, [%[off]]") /* q4 = gain_end */ \ |
| 110 |
__ASM_EMIT("vmov q1, q0") \ |
| 111 |
__ASM_EMIT("vbit q0, q4, q2") /* q0 = [fabs(x0) > start] ? gain_end : gain_start */ \ |
| 112 |
__ASM_EMIT("vbit q1, q4, q3") \ |
| 113 |
__ASM_EMIT("b 200f") \ |
| 114 |
__ASM_EMIT("100:") \ |
| 115 |
__ASM_EMIT("vldm %[LOGC], {q14-q15}") \ |
| 116 |
__ASM_EMIT("vstm %[mem], {q0-q1}") /* mem[0x00] = fabfs(x0) */ \ |
| 117 |
LOGE_CORE_X8 /* q0 = lx0 = logf(fabsf(x0)) */ \ |
| 118 |
PROCESS_KNEE_SINGLE_X8 /* apply knee 0 */ \ |
| 119 |
__ASM_EMIT("200:") \ |
| 120 |
/* out: q0 = G0, q1= G1 */ |
| 121 |
|
| 122 |
#define PROCESS_GATE_FULL_X4 \ |
| 123 |
/* in: q0 = x0, q1 = x1 */ \ |
| 124 |
__ASM_EMIT("vldm %[LOGC], {q14-q15}") \ |
| 125 |
__ASM_EMIT("vabs.f32 q0, q0") /* q0 = fabsf(x0) */ \ |
| 126 |
__ASM_EMIT("vstm %[mem], {q0}") /* mem[0x00] = fabfs(x0) */ \ |
| 127 |
LOGE_CORE_X4 /* q0 = lx0 = logf(fabsf(x0)) */ \ |
| 128 |
PROCESS_KNEE_SINGLE_X4 /* apply knee 0 */ \ |
| 129 |
/* out: q0 = G0 */ |
| 130 |
|
| 131 |
void gate_x1_gain(float *dst, const float *src, const dsp::gate_knee_t *c, size_t count) |
| 132 |
{ |
| 133 |
IF_ARCH_ARM( |
| 134 |
float mem[8] __lsp_aligned16; |
| 135 |
size_t off; |
| 136 |
); |
| 137 |
|
| 138 |
ARCH_ARM_ASM( |
| 139 |
// x8 blocks |
| 140 |
__ASM_EMIT("subs %[count], #8") |
| 141 |
__ASM_EMIT("blo 2f") |
| 142 |
__ASM_EMIT("1:") |
| 143 |
__ASM_EMIT("vld1.32 {q0-q1}, [%[src]]!") |
| 144 |
PROCESS_GATE_FULL_X8 |
| 145 |
__ASM_EMIT("subs %[count], #8") |
| 146 |
__ASM_EMIT("vst1.32 {q0-q1}, [%[dst]]!") |
| 147 |
__ASM_EMIT("bhs 1b") |
| 148 |
__ASM_EMIT("2:") |
| 149 |
// x4 block |
| 150 |
__ASM_EMIT("adds %[count], #4") |
| 151 |
__ASM_EMIT("blt 4f") |
| 152 |
__ASM_EMIT("vld1.32 {q0}, [%[src]]!") |
| 153 |
PROCESS_GATE_FULL_X4 |
| 154 |
__ASM_EMIT("sub %[count], #4") |
| 155 |
__ASM_EMIT("vst1.32 {q0}, [%[dst]]!") |
| 156 |
__ASM_EMIT("4:") |
| 157 |
// Tail: 1x-3x block |
| 158 |
__ASM_EMIT("adds %[count], #4") |
| 159 |
__ASM_EMIT("bls 12f") |
| 160 |
__ASM_EMIT("tst %[count], #1") |
| 161 |
__ASM_EMIT("beq 6f") |
| 162 |
__ASM_EMIT("vld1.32 {d0[0]}, [%[src]]!") |
| 163 |
__ASM_EMIT("6:") |
| 164 |
__ASM_EMIT("tst %[count], #2") |
| 165 |
__ASM_EMIT("beq 8f") |
| 166 |
__ASM_EMIT("vld1.32 {d1}, [%[src]]") |
| 167 |
__ASM_EMIT("8:") |
| 168 |
PROCESS_GATE_FULL_X4 |
| 169 |
__ASM_EMIT("tst %[count], #1") |
| 170 |
__ASM_EMIT("beq 10f") |
| 171 |
__ASM_EMIT("vst1.32 {d0[0]}, [%[dst]]!") |
| 172 |
__ASM_EMIT("10:") |
| 173 |
__ASM_EMIT("tst %[count], #2") |
| 174 |
__ASM_EMIT("beq 12f") |
| 175 |
__ASM_EMIT("vst1.32 {d1}, [%[dst]]") |
| 176 |
__ASM_EMIT("12:") |
| 177 |
|
| 178 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 179 |
[count] "+r" (count), |
| 180 |
[off] "=&r" (off) |
| 181 |
: [gate] "r" (c), |
| 182 |
[L2C] "r" (&LOG2_CONST[0]), |
| 183 |
[LOGC] "r" (&LOGE_C[0]), |
| 184 |
[mem] "r" (&mem[0]), |
| 185 |
[E2C] "r" (&EXP2_CONST[0]), |
| 186 |
[LOG2E] "r" (&EXP_LOG2E[0]) |
| 187 |
: "cc", "memory", |
| 188 |
"q0", "q1", "q2", "q3", |
| 189 |
"q4", "q5", "q6", "q7", |
| 190 |
"q8", "q9", "q10", "q11", |
| 191 |
"q12", "q13", "q14", "q15" |
| 192 |
); |
| 193 |
} |
| 194 |
|
| 195 |
void gate_x1_curve(float *dst, const float *src, const dsp::gate_knee_t *c, size_t count) |
| 196 |
{ |
| 197 |
IF_ARCH_ARM( |
| 198 |
float mem[8] __lsp_aligned16; |
| 199 |
size_t off; |
| 200 |
); |
| 201 |
|
| 202 |
ARCH_ARM_ASM( |
| 203 |
// x8 blocks |
| 204 |
__ASM_EMIT("subs %[count], #8") |
| 205 |
__ASM_EMIT("blo 2f") |
| 206 |
__ASM_EMIT("1:") |
| 207 |
__ASM_EMIT("vld1.32 {q0-q1}, [%[src]]") |
| 208 |
PROCESS_GATE_FULL_X8 |
| 209 |
__ASM_EMIT("vld1.32 {q2-q3}, [%[src]]!") |
| 210 |
__ASM_EMIT("subs %[count], #8") |
| 211 |
__ASM_EMIT("vmul.f32 q0, q0, q2") |
| 212 |
__ASM_EMIT("vmul.f32 q1, q1, q3") |
| 213 |
__ASM_EMIT("vst1.32 {q0-q1}, [%[dst]]!") |
| 214 |
__ASM_EMIT("bhs 1b") |
| 215 |
__ASM_EMIT("2:") |
| 216 |
// x4 block |
| 217 |
__ASM_EMIT("adds %[count], #4") |
| 218 |
__ASM_EMIT("blt 4f") |
| 219 |
__ASM_EMIT("vld1.32 {q0}, [%[src]]") |
| 220 |
PROCESS_GATE_FULL_X4 |
| 221 |
__ASM_EMIT("vld1.32 {q2}, [%[src]]!") |
| 222 |
__ASM_EMIT("sub %[count], #4") |
| 223 |
__ASM_EMIT("vmul.f32 q0, q0, q2") |
| 224 |
__ASM_EMIT("vst1.32 {q0}, [%[dst]]!") |
| 225 |
__ASM_EMIT("4:") |
| 226 |
// Tail: 1x-3x block |
| 227 |
__ASM_EMIT("adds %[count], #4") |
| 228 |
__ASM_EMIT("bls 12f") |
| 229 |
__ASM_EMIT("tst %[count], #1") |
| 230 |
__ASM_EMIT("beq 6f") |
| 231 |
__ASM_EMIT("vld1.32 {d0[0]}, [%[src]]!") |
| 232 |
__ASM_EMIT("6:") |
| 233 |
__ASM_EMIT("tst %[count], #2") |
| 234 |
__ASM_EMIT("beq 8f") |
| 235 |
__ASM_EMIT("vld1.32 {d1}, [%[src]]") |
| 236 |
__ASM_EMIT("8:") |
| 237 |
__ASM_EMIT("vmov q1, q0") |
| 238 |
PROCESS_GATE_FULL_X4 |
| 239 |
__ASM_EMIT("vmul.f32 q0, q0, q1") |
| 240 |
__ASM_EMIT("tst %[count], #1") |
| 241 |
__ASM_EMIT("beq 10f") |
| 242 |
__ASM_EMIT("vst1.32 {d0[0]}, [%[dst]]!") |
| 243 |
__ASM_EMIT("10:") |
| 244 |
__ASM_EMIT("tst %[count], #2") |
| 245 |
__ASM_EMIT("beq 12f") |
| 246 |
__ASM_EMIT("vst1.32 {d1}, [%[dst]]") |
| 247 |
__ASM_EMIT("12:") |
| 248 |
|
| 249 |
: [dst] "+r" (dst), [src] "+r" (src), |
| 250 |
[count] "+r" (count), |
| 251 |
[off] "=&r" (off) |
| 252 |
: [gate] "r" (c), |
| 253 |
[L2C] "r" (&LOG2_CONST[0]), |
| 254 |
[LOGC] "r" (&LOGE_C[0]), |
| 255 |
[mem] "r" (&mem[0]), |
| 256 |
[E2C] "r" (&EXP2_CONST[0]), |
| 257 |
[LOG2E] "r" (&EXP_LOG2E[0]) |
| 258 |
: "cc", "memory", |
| 259 |
"q0", "q1", "q2", "q3", |
| 260 |
"q4", "q5", "q6", "q7", |
| 261 |
"q8", "q9", "q10", "q11", |
| 262 |
"q12", "q13", "q14", "q15" |
| 263 |
); |
| 264 |
} |
| 265 |
|
| 266 |
#undef PROCESS_GATE_FULL_X4 |
| 267 |
#undef PROCESS_GATE_FULL_X8 |
| 268 |
#undef PROCESS_KNEE_SINGLE_X4 |
| 269 |
#undef PROCESS_KNEE_SINGLE_X8 |
| 270 |
|
| 271 |
} /* namespace neon_d32 */ |
| 272 |
} /* namespace lsp */ |
| 273 |
|
| 274 |
|
| 275 |
|
| 276 |
|
| 277 |
#endif /* PRIVATE_DSP_ARCH_ARM_NEON_D32_DYNAMICS_GATE_H_ */ |