| 1 |
/* |
| 2 |
* Copyright (C) 2025 Linux Studio Plugins Project <https://lsp-plug.in/> |
| 3 |
* (C) 2025 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_X86_AVX_IMPL |
| 23 |
#error "This header should not be included directly" |
| 24 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX_IMPL */ |
| 25 |
|
| 26 |
#include <private/dsp/arch/x86/avx/fft/const.h> |
| 27 |
#include <private/dsp/arch/x86/avx/fft/butterfly.h> |
| 28 |
#include <private/dsp/arch/x86/avx/fft/normalize.h> |
| 29 |
|
| 30 |
#define FFT_SCRAMBLE_SELF_DIRECT_NAME scramble_self_direct8 |
| 31 |
#define FFT_SCRAMBLE_SELF_REVERSE_NAME scramble_self_reverse8 |
| 32 |
#define FFT_SCRAMBLE_COPY_DIRECT_NAME scramble_copy_direct8 |
| 33 |
#define FFT_SCRAMBLE_COPY_REVERSE_NAME scramble_copy_reverse8 |
| 34 |
#define FFT_TYPE uint8_t |
| 35 |
#define FFT_FMA(a, b) a |
| 36 |
#include <private/dsp/arch/x86/avx/fft/scramble.h> |
| 37 |
|
| 38 |
#define FFT_SCRAMBLE_SELF_DIRECT_NAME scramble_self_direct16 |
| 39 |
#define FFT_SCRAMBLE_SELF_REVERSE_NAME scramble_self_reverse16 |
| 40 |
#define FFT_SCRAMBLE_COPY_DIRECT_NAME scramble_copy_direct16 |
| 41 |
#define FFT_SCRAMBLE_COPY_REVERSE_NAME scramble_copy_reverse16 |
| 42 |
#define FFT_TYPE uint16_t |
| 43 |
#define FFT_FMA(a, b) a |
| 44 |
#include <private/dsp/arch/x86/avx/fft/scramble.h> |
| 45 |
|
| 46 |
#define FFT_SCRAMBLE_SELF_DIRECT_NAME scramble_self_direct8_fma3 |
| 47 |
#define FFT_SCRAMBLE_SELF_REVERSE_NAME scramble_self_reverse8_fma3 |
| 48 |
#define FFT_SCRAMBLE_COPY_DIRECT_NAME scramble_copy_direct8_fma3 |
| 49 |
#define FFT_SCRAMBLE_COPY_REVERSE_NAME scramble_copy_reverse8_fma3 |
| 50 |
#define FFT_TYPE uint8_t |
| 51 |
#define FFT_FMA(a, b) b |
| 52 |
#include <private/dsp/arch/x86/avx/fft/scramble.h> |
| 53 |
|
| 54 |
#define FFT_SCRAMBLE_SELF_DIRECT_NAME scramble_self_direct16_fma3 |
| 55 |
#define FFT_SCRAMBLE_SELF_REVERSE_NAME scramble_self_reverse16_fma3 |
| 56 |
#define FFT_SCRAMBLE_COPY_DIRECT_NAME scramble_copy_direct16_fma3 |
| 57 |
#define FFT_SCRAMBLE_COPY_REVERSE_NAME scramble_copy_reverse16_fma3 |
| 58 |
#define FFT_TYPE uint16_t |
| 59 |
#define FFT_FMA(a, b) b |
| 60 |
#include <private/dsp/arch/x86/avx/fft/scramble.h> |
| 61 |
|
| 62 |
namespace lsp |
| 63 |
{ |
| 64 |
namespace avx |
| 65 |
{ |
| 66 |
static inline void small_direct_fft(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t rank) |
| 67 |
{ |
| 68 |
if (rank == 2) |
| 69 |
{ |
| 70 |
float s0_re = src_re[0] + src_re[2]; |
| 71 |
float s1_re = src_re[0] - src_re[2]; |
| 72 |
float s2_re = src_re[1] + src_re[3]; |
| 73 |
float s3_re = src_re[1] - src_re[3]; |
| 74 |
|
| 75 |
float s0_im = src_im[0] + src_im[2]; |
| 76 |
float s1_im = src_im[0] - src_im[2]; |
| 77 |
float s2_im = src_im[1] + src_im[3]; |
| 78 |
float s3_im = src_im[1] - src_im[3]; |
| 79 |
|
| 80 |
dst_re[0] = s0_re + s2_re; |
| 81 |
dst_re[1] = s1_re + s3_im; |
| 82 |
dst_re[2] = s0_re - s2_re; |
| 83 |
dst_re[3] = s1_re - s3_im; |
| 84 |
|
| 85 |
dst_im[0] = s0_im + s2_im; |
| 86 |
dst_im[1] = s1_im - s3_re; |
| 87 |
dst_im[2] = s0_im - s2_im; |
| 88 |
dst_im[3] = s1_im + s3_re; |
| 89 |
} |
| 90 |
else if (rank == 1) |
| 91 |
{ |
| 92 |
// s0' = s0 + s1 |
| 93 |
// s1' = s0 - s1 |
| 94 |
float s1_re = src_re[1]; |
| 95 |
float s1_im = src_im[1]; |
| 96 |
dst_re[1] = src_re[0] - s1_re; |
| 97 |
dst_im[1] = src_im[0] - s1_im; |
| 98 |
dst_re[0] = src_re[0] + s1_re; |
| 99 |
dst_im[0] = src_im[0] + s1_im; |
| 100 |
} |
| 101 |
else |
| 102 |
{ |
| 103 |
dst_re[0] = src_re[0]; |
| 104 |
dst_im[0] = src_im[0]; |
| 105 |
} |
| 106 |
} |
| 107 |
|
| 108 |
static inline void small_reverse_fft(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t rank) |
| 109 |
{ |
| 110 |
if (rank == 2) |
| 111 |
{ |
| 112 |
float s0_re = src_re[0] + src_re[2]; |
| 113 |
float s1_re = src_re[0] - src_re[2]; |
| 114 |
float s2_re = src_re[1] + src_re[3]; |
| 115 |
float s3_re = src_re[1] - src_re[3]; |
| 116 |
|
| 117 |
float s0_im = src_im[0] + src_im[2]; |
| 118 |
float s1_im = src_im[0] - src_im[2]; |
| 119 |
float s2_im = src_im[1] + src_im[3]; |
| 120 |
float s3_im = src_im[1] - src_im[3]; |
| 121 |
|
| 122 |
dst_re[0] = (s0_re + s2_re) * 0.25f; |
| 123 |
dst_re[1] = (s1_re - s3_im) * 0.25f; |
| 124 |
dst_re[2] = (s0_re - s2_re) * 0.25f; |
| 125 |
dst_re[3] = (s1_re + s3_im) * 0.25f; |
| 126 |
|
| 127 |
dst_im[0] = (s0_im + s2_im) * 0.25f; |
| 128 |
dst_im[1] = (s1_im + s3_re) * 0.25f; |
| 129 |
dst_im[2] = (s0_im - s2_im) * 0.25f; |
| 130 |
dst_im[3] = (s1_im - s3_re) * 0.25f; |
| 131 |
} |
| 132 |
else if (rank == 1) |
| 133 |
{ |
| 134 |
// s0' = s0 + s1 |
| 135 |
// s1' = s0 - s1 |
| 136 |
float s1_re = src_re[1]; |
| 137 |
float s1_im = src_im[1]; |
| 138 |
dst_re[1] = src_re[0] - s1_re; |
| 139 |
dst_im[1] = src_im[0] - s1_im; |
| 140 |
dst_re[0] = src_re[0] + s1_re; |
| 141 |
dst_im[0] = src_im[0] + s1_im; |
| 142 |
} |
| 143 |
else |
| 144 |
{ |
| 145 |
dst_re[0] = src_re[0]; |
| 146 |
dst_im[0] = src_im[0]; |
| 147 |
} |
| 148 |
} |
| 149 |
|
| 150 |
void direct_fft(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t rank) |
| 151 |
{ |
| 152 |
// Check bounds |
| 153 |
if (rank <= 2) |
| 154 |
{ |
| 155 |
small_direct_fft(dst_re, dst_im, src_re, src_im, rank); |
| 156 |
return; |
| 157 |
} |
| 158 |
|
| 159 |
if ((dst_re == src_re) || (dst_im == src_im) || (rank < 4)) |
| 160 |
{ |
| 161 |
dsp::move(dst_re, src_re, 1 << rank); |
| 162 |
dsp::move(dst_im, src_im, 1 << rank); |
| 163 |
if (rank <= 8) |
| 164 |
scramble_self_direct8(dst_re, dst_im, rank); |
| 165 |
else |
| 166 |
scramble_self_direct16(dst_re, dst_im, rank); |
| 167 |
} |
| 168 |
else |
| 169 |
{ |
| 170 |
if (rank <= 12) |
| 171 |
scramble_copy_direct8(dst_re, dst_im, src_re, src_im, rank-4); |
| 172 |
else |
| 173 |
scramble_copy_direct16(dst_re, dst_im, src_re, src_im, rank-4); |
| 174 |
} |
| 175 |
|
| 176 |
for (size_t i=3; i < rank; ++i) |
| 177 |
butterfly_direct8p(dst_re, dst_im, i, 1 << (rank - i - 1)); |
| 178 |
} |
| 179 |
|
| 180 |
void direct_fft_fma3(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t rank) |
| 181 |
{ |
| 182 |
// Check bounds |
| 183 |
if (rank <= 2) |
| 184 |
{ |
| 185 |
small_direct_fft(dst_re, dst_im, src_re, src_im, rank); |
| 186 |
return; |
| 187 |
} |
| 188 |
|
| 189 |
if ((dst_re == src_re) || (dst_im == src_im) || (rank < 4)) |
| 190 |
{ |
| 191 |
dsp::move(dst_re, src_re, 1 << rank); |
| 192 |
dsp::move(dst_im, src_im, 1 << rank); |
| 193 |
if (rank <= 8) |
| 194 |
scramble_self_direct8_fma3(dst_re, dst_im, rank); |
| 195 |
else |
| 196 |
scramble_self_direct16_fma3(dst_re, dst_im, rank); |
| 197 |
} |
| 198 |
else |
| 199 |
{ |
| 200 |
if (rank <= 12) |
| 201 |
scramble_copy_direct8_fma3(dst_re, dst_im, src_re, src_im, rank-4); |
| 202 |
else |
| 203 |
scramble_copy_direct16_fma3(dst_re, dst_im, src_re, src_im, rank-4); |
| 204 |
} |
| 205 |
|
| 206 |
for (size_t i=3; i < rank; ++i) |
| 207 |
butterfly_direct8p_fma3(dst_re, dst_im, i, 1 << (rank - i - 1)); |
| 208 |
} |
| 209 |
|
| 210 |
void reverse_fft(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t rank) |
| 211 |
{ |
| 212 |
// Check bounds |
| 213 |
if (rank <= 2) |
| 214 |
{ |
| 215 |
small_reverse_fft(dst_re, dst_im, src_re, src_im, rank); |
| 216 |
return; |
| 217 |
} |
| 218 |
|
| 219 |
if ((dst_re == src_re) || (dst_im == src_im) || (rank < 4)) |
| 220 |
{ |
| 221 |
dsp::move(dst_re, src_re, 1 << rank); |
| 222 |
dsp::move(dst_im, src_im, 1 << rank); |
| 223 |
if (rank <= 8) |
| 224 |
scramble_self_reverse8(dst_re, dst_im, rank); |
| 225 |
else |
| 226 |
scramble_self_reverse16(dst_re, dst_im, rank); |
| 227 |
} |
| 228 |
else |
| 229 |
{ |
| 230 |
if (rank <= 12) |
| 231 |
scramble_copy_reverse8(dst_re, dst_im, src_re, src_im, rank-4); |
| 232 |
else |
| 233 |
scramble_copy_reverse16(dst_re, dst_im, src_re, src_im, rank-4); |
| 234 |
} |
| 235 |
|
| 236 |
for (size_t i=3; i < rank; ++i) |
| 237 |
butterfly_reverse8p(dst_re, dst_im, i, 1 << (rank - i - 1)); |
| 238 |
|
| 239 |
dsp::normalize_fft2(dst_re, dst_im, rank); |
| 240 |
} |
| 241 |
|
| 242 |
void reverse_fft_fma3(float *dst_re, float *dst_im, const float *src_re, const float *src_im, size_t rank) |
| 243 |
{ |
| 244 |
// Check bounds |
| 245 |
if (rank <= 2) |
| 246 |
{ |
| 247 |
small_reverse_fft(dst_re, dst_im, src_re, src_im, rank); |
| 248 |
return; |
| 249 |
} |
| 250 |
|
| 251 |
if ((dst_re == src_re) || (dst_im == src_im) || (rank < 4)) |
| 252 |
{ |
| 253 |
dsp::move(dst_re, src_re, 1 << rank); |
| 254 |
dsp::move(dst_im, src_im, 1 << rank); |
| 255 |
if (rank <= 8) |
| 256 |
scramble_self_reverse8_fma3(dst_re, dst_im, rank); |
| 257 |
else |
| 258 |
scramble_self_reverse16_fma3(dst_re, dst_im, rank); |
| 259 |
} |
| 260 |
else |
| 261 |
{ |
| 262 |
if (rank <= 12) |
| 263 |
scramble_copy_reverse8_fma3(dst_re, dst_im, src_re, src_im, rank-4); |
| 264 |
else |
| 265 |
scramble_copy_reverse16_fma3(dst_re, dst_im, src_re, src_im, rank-4); |
| 266 |
} |
| 267 |
|
| 268 |
for (size_t i=3; i < rank; ++i) |
| 269 |
butterfly_reverse8p_fma3(dst_re, dst_im, i, 1 << (rank - i - 1)); |
| 270 |
|
| 271 |
dsp::normalize_fft2(dst_re, dst_im, rank); |
| 272 |
} |
| 273 |
|
| 274 |
} /* namespace avx */ |
| 275 |
} /* namespace lsp */ |