| 1 | /* Copyright (C) 2010-2020 The RetroArch team |
| 2 | * |
| 3 | * --------------------------------------------------------------------------------------- |
| 4 | * The following license statement only applies to this file (sinc_resampler.c). |
| 5 | * --------------------------------------------------------------------------------------- |
| 6 | * |
| 7 | * Permission is hereby granted, free of charge, |
| 8 | * to any person obtaining a copy of this software and associated documentation files (the "Software"), |
| 9 | * to deal in the Software without restriction, including without limitation the rights to |
| 10 | * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, |
| 11 | * and to permit persons to whom the Software is furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. |
| 14 | * |
| 15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, |
| 16 | * INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. |
| 18 | * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, |
| 19 | * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 20 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. |
| 21 | */ |
| 22 | |
| 23 | /* Bog-standard windowed SINC implementation. */ |
| 24 | |
| 25 | #include <stdint.h> |
| 26 | #include <stdlib.h> |
| 27 | #include <math.h> |
| 28 | #include <string.h> |
| 29 | |
| 30 | #include <retro_environment.h> |
| 31 | #include <retro_inline.h> |
| 32 | #include <filters.h> |
| 33 | #include <memalign.h> |
| 34 | |
| 35 | #include <audio/audio_resampler.h> |
| 36 | #include <filters.h> |
| 37 | |
| 38 | #ifdef __SSE__ |
| 39 | #include <xmmintrin.h> |
| 40 | #endif |
| 41 | |
| 42 | #if defined(__AVX__) |
| 43 | #include <immintrin.h> |
| 44 | #endif |
| 45 | |
| 46 | /* Rough SNR values for upsampling: |
| 47 | * LOWEST: 40 dB |
| 48 | * LOWER: 55 dB |
| 49 | * NORMAL: 70 dB |
| 50 | * HIGHER: 110 dB |
| 51 | * HIGHEST: 140 dB |
| 52 | */ |
| 53 | |
| 54 | /* TODO, make all this more configurable. */ |
| 55 | |
| 56 | enum sinc_window |
| 57 | { |
| 58 | SINC_WINDOW_NONE = 0, |
| 59 | SINC_WINDOW_KAISER, |
| 60 | SINC_WINDOW_LANCZOS |
| 61 | }; |
| 62 | |
| 63 | /* For the little amount of taps we're using, |
| 64 | * SSE1 is faster than AVX for some reason. |
| 65 | * AVX code is kept here though as by increasing number |
| 66 | * of sinc taps, the AVX code is clearly faster than SSE1. |
| 67 | */ |
| 68 | |
| 69 | typedef struct rarch_sinc_resampler |
| 70 | { |
| 71 | /* A buffer for phase_table, buffer_l and buffer_r |
| 72 | * are created in a single calloc(). |
| 73 | * Ensure that we get as good cache locality as we can hope for. */ |
| 74 | float *main_buffer; |
| 75 | float *phase_table; |
| 76 | float *buffer_l; |
| 77 | float *buffer_r; |
| 78 | unsigned phase_bits; |
| 79 | unsigned subphase_bits; |
| 80 | unsigned subphase_mask; |
| 81 | unsigned taps; |
| 82 | unsigned ptr; |
| 83 | uint32_t time; |
| 84 | float subphase_mod; |
| 85 | float kaiser_beta; |
| 86 | } rarch_sinc_resampler_t; |
| 87 | |
| 88 | #if (defined(__ARM_NEON__) || defined(HAVE_NEON)) |
| 89 | |
| 90 | #ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS |
| 91 | void process_sinc_neon_asm(float *out, const float *left, |
| 92 | const float *right, const float *coeff, unsigned taps); |
| 93 | #else |
| 94 | #include <arm_neon.h> |
| 95 | |
| 96 | /* Assumes that taps >= 8, and that taps is a multiple of 8. |
| 97 | * Not bothering to reimplement this one for the external .S |
| 98 | */ |
| 99 | static void resampler_sinc_process_neon_kaiser(void *re_, struct resampler_data *data) |
| 100 | { |
| 101 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 102 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 103 | uint32_t ratio = phases / data->ratio; |
| 104 | const float *input = data->data_in; |
| 105 | float *output = data->data_out; |
| 106 | size_t frames = data->input_frames; |
| 107 | size_t out_frames = 0; |
| 108 | unsigned taps = resamp->taps; |
| 109 | while (frames) |
| 110 | { |
| 111 | while (frames && resamp->time >= phases) |
| 112 | { |
| 113 | /* Push in reverse to make filter more obvious. */ |
| 114 | if (!resamp->ptr) |
| 115 | resamp->ptr = taps; |
| 116 | resamp->ptr--; |
| 117 | |
| 118 | resamp->buffer_l[resamp->ptr + taps] = |
| 119 | resamp->buffer_l[resamp->ptr] = *input++; |
| 120 | |
| 121 | resamp->buffer_r[resamp->ptr + taps] = |
| 122 | resamp->buffer_r[resamp->ptr] = *input++; |
| 123 | |
| 124 | resamp->time -= phases; |
| 125 | frames--; |
| 126 | } |
| 127 | |
| 128 | { |
| 129 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 130 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 131 | while (resamp->time < phases) |
| 132 | { |
| 133 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 134 | const float *phase_table = resamp->phase_table + phase * taps * 2; |
| 135 | const float *delta_table = phase_table + taps; |
| 136 | float32x4_t delta = vdupq_n_f32((resamp->time & resamp->subphase_mask) * resamp->subphase_mod); |
| 137 | int i; |
| 138 | float32x4_t p1 = {0, 0, 0, 0}, p2 = {0, 0, 0, 0}; |
| 139 | float32x2_t p3, p4; |
| 140 | |
| 141 | for (i = 0; i < (int)taps; i += 8) |
| 142 | { |
| 143 | float32x4x2_t coeff8 = vld2q_f32(&phase_table[i]); |
| 144 | float32x4x2_t delta8 = vld2q_f32(&delta_table[i]); |
| 145 | float32x4x2_t left8 = vld2q_f32(&buffer_l[i]); |
| 146 | float32x4x2_t right8 = vld2q_f32(&buffer_r[i]); |
| 147 | |
| 148 | coeff8.val[0] = vmlaq_f32(coeff8.val[0], delta8.val[0], delta); |
| 149 | coeff8.val[1] = vmlaq_f32(coeff8.val[1], delta8.val[1], delta); |
| 150 | |
| 151 | p1 = vmlaq_f32(p1, left8.val[0], coeff8.val[0]); |
| 152 | p2 = vmlaq_f32(p2, right8.val[0], coeff8.val[0]); |
| 153 | p1 = vmlaq_f32(p1, left8.val[1], coeff8.val[1]); |
| 154 | p2 = vmlaq_f32(p2, right8.val[1], coeff8.val[1]); |
| 155 | } |
| 156 | |
| 157 | p3 = vadd_f32(vget_low_f32(p1), vget_high_f32(p1)); |
| 158 | p4 = vadd_f32(vget_low_f32(p2), vget_high_f32(p2)); |
| 159 | vst1_f32(output, vpadd_f32(p3, p4)); |
| 160 | output += 2; |
| 161 | out_frames++; |
| 162 | resamp->time += ratio; |
| 163 | } |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | data->output_frames = out_frames; |
| 168 | } |
| 169 | #endif |
| 170 | |
| 171 | /* Assumes that taps >= 8, and that taps is a multiple of 8. */ |
| 172 | static void resampler_sinc_process_neon(void *re_, struct resampler_data *data) |
| 173 | { |
| 174 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 175 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 176 | uint32_t ratio = phases / data->ratio; |
| 177 | const float *input = data->data_in; |
| 178 | float *output = data->data_out; |
| 179 | size_t frames = data->input_frames; |
| 180 | size_t out_frames = 0; |
| 181 | unsigned taps = resamp->taps; |
| 182 | |
| 183 | while (frames) |
| 184 | { |
| 185 | while (frames && resamp->time >= phases) |
| 186 | { |
| 187 | /* Push in reverse to make filter more obvious. */ |
| 188 | if (!resamp->ptr) |
| 189 | resamp->ptr = taps; |
| 190 | resamp->ptr--; |
| 191 | |
| 192 | resamp->buffer_l[resamp->ptr + taps] = |
| 193 | resamp->buffer_l[resamp->ptr] = *input++; |
| 194 | |
| 195 | resamp->buffer_r[resamp->ptr + taps] = |
| 196 | resamp->buffer_r[resamp->ptr] = *input++; |
| 197 | |
| 198 | resamp->time -= phases; |
| 199 | frames--; |
| 200 | } |
| 201 | |
| 202 | { |
| 203 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 204 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 205 | while (resamp->time < phases) |
| 206 | { |
| 207 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 208 | const float *phase_table = resamp->phase_table + phase * taps; |
| 209 | #ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS |
| 210 | process_sinc_neon_asm(output, buffer_l, buffer_r, phase_table, taps); |
| 211 | #else |
| 212 | int i; |
| 213 | float32x4_t p1 = {0, 0, 0, 0}, p2 = {0, 0, 0, 0}; |
| 214 | float32x2_t p3, p4; |
| 215 | |
| 216 | for (i = 0; i < (int)taps; i += 8) |
| 217 | { |
| 218 | float32x4x2_t coeff8 = vld2q_f32(&phase_table[i]); |
| 219 | float32x4x2_t left8 = vld2q_f32(&buffer_l[i]); |
| 220 | float32x4x2_t right8 = vld2q_f32(&buffer_r[i]); |
| 221 | |
| 222 | p1 = vmlaq_f32(p1, left8.val[0], coeff8.val[0]); |
| 223 | p2 = vmlaq_f32(p2, right8.val[0], coeff8.val[0]); |
| 224 | p1 = vmlaq_f32(p1, left8.val[1], coeff8.val[1]); |
| 225 | p2 = vmlaq_f32(p2, right8.val[1], coeff8.val[1]); |
| 226 | } |
| 227 | |
| 228 | p3 = vadd_f32(vget_low_f32(p1), vget_high_f32(p1)); |
| 229 | p4 = vadd_f32(vget_low_f32(p2), vget_high_f32(p2)); |
| 230 | vst1_f32(output, vpadd_f32(p3, p4)); |
| 231 | #endif |
| 232 | output += 2; |
| 233 | out_frames++; |
| 234 | resamp->time += ratio; |
| 235 | } |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | data->output_frames = out_frames; |
| 240 | } |
| 241 | #endif |
| 242 | |
| 243 | #if defined(__AVX__) |
| 244 | static void resampler_sinc_process_avx_kaiser(void *re_, struct resampler_data *data) |
| 245 | { |
| 246 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 247 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 248 | |
| 249 | uint32_t ratio = phases / data->ratio; |
| 250 | const float *input = data->data_in; |
| 251 | float *output = data->data_out; |
| 252 | size_t frames = data->input_frames; |
| 253 | size_t out_frames = 0; |
| 254 | unsigned taps = resamp->taps; |
| 255 | |
| 256 | { |
| 257 | while (frames) |
| 258 | { |
| 259 | while (frames && resamp->time >= phases) |
| 260 | { |
| 261 | /* Push in reverse to make filter more obvious. */ |
| 262 | if (!resamp->ptr) |
| 263 | resamp->ptr = taps; |
| 264 | resamp->ptr--; |
| 265 | |
| 266 | resamp->buffer_l[resamp->ptr + taps] = |
| 267 | resamp->buffer_l[resamp->ptr] = *input++; |
| 268 | |
| 269 | resamp->buffer_r[resamp->ptr + taps] = |
| 270 | resamp->buffer_r[resamp->ptr] = *input++; |
| 271 | |
| 272 | resamp->time -= phases; |
| 273 | frames--; |
| 274 | } |
| 275 | |
| 276 | { |
| 277 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 278 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 279 | while (resamp->time < phases) |
| 280 | { |
| 281 | int i; |
| 282 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 283 | |
| 284 | float *phase_table = resamp->phase_table + phase * taps * 2; |
| 285 | float *delta_table = phase_table + taps; |
| 286 | __m256 delta = _mm256_set1_ps((float) |
| 287 | (resamp->time & resamp->subphase_mask) * resamp->subphase_mod); |
| 288 | |
| 289 | __m256 sum_l = _mm256_setzero_ps(); |
| 290 | __m256 sum_r = _mm256_setzero_ps(); |
| 291 | |
| 292 | for (i = 0; i < (int)taps; i += 8) |
| 293 | { |
| 294 | __m256 buf_l = _mm256_loadu_ps(buffer_l + i); |
| 295 | __m256 buf_r = _mm256_loadu_ps(buffer_r + i); |
| 296 | __m256 deltas = _mm256_load_ps(delta_table + i); |
| 297 | __m256 sinc = _mm256_add_ps(_mm256_load_ps((const float*)phase_table + i), |
| 298 | _mm256_mul_ps(deltas, delta)); |
| 299 | |
| 300 | sum_l = _mm256_add_ps(sum_l, _mm256_mul_ps(buf_l, sinc)); |
| 301 | sum_r = _mm256_add_ps(sum_r, _mm256_mul_ps(buf_r, sinc)); |
| 302 | } |
| 303 | |
| 304 | /* hadd on AVX is weird, and acts on low-lanes |
| 305 | * and high-lanes separately. */ |
| 306 | __m256 res_l = _mm256_hadd_ps(sum_l, sum_l); |
| 307 | __m256 res_r = _mm256_hadd_ps(sum_r, sum_r); |
| 308 | res_l = _mm256_hadd_ps(res_l, res_l); |
| 309 | res_r = _mm256_hadd_ps(res_r, res_r); |
| 310 | res_l = _mm256_add_ps(_mm256_permute2f128_ps(res_l, res_l, 1), res_l); |
| 311 | res_r = _mm256_add_ps(_mm256_permute2f128_ps(res_r, res_r, 1), res_r); |
| 312 | |
| 313 | /* This is optimized to mov %xmmN, [mem]. |
| 314 | * There doesn't seem to be any _mm256_store_ss intrinsic. */ |
| 315 | _mm_store_ss(output + 0, _mm256_extractf128_ps(res_l, 0)); |
| 316 | _mm_store_ss(output + 1, _mm256_extractf128_ps(res_r, 0)); |
| 317 | |
| 318 | output += 2; |
| 319 | out_frames++; |
| 320 | resamp->time += ratio; |
| 321 | } |
| 322 | } |
| 323 | } |
| 324 | } |
| 325 | |
| 326 | data->output_frames = out_frames; |
| 327 | } |
| 328 | |
| 329 | static void resampler_sinc_process_avx(void *re_, struct resampler_data *data) |
| 330 | { |
| 331 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 332 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 333 | |
| 334 | uint32_t ratio = phases / data->ratio; |
| 335 | const float *input = data->data_in; |
| 336 | float *output = data->data_out; |
| 337 | size_t frames = data->input_frames; |
| 338 | size_t out_frames = 0; |
| 339 | unsigned taps = resamp->taps; |
| 340 | |
| 341 | { |
| 342 | while (frames) |
| 343 | { |
| 344 | while (frames && resamp->time >= phases) |
| 345 | { |
| 346 | /* Push in reverse to make filter more obvious. */ |
| 347 | if (!resamp->ptr) |
| 348 | resamp->ptr = taps; |
| 349 | resamp->ptr--; |
| 350 | |
| 351 | resamp->buffer_l[resamp->ptr + taps] = |
| 352 | resamp->buffer_l[resamp->ptr] = *input++; |
| 353 | |
| 354 | resamp->buffer_r[resamp->ptr + taps] = |
| 355 | resamp->buffer_r[resamp->ptr] = *input++; |
| 356 | |
| 357 | resamp->time -= phases; |
| 358 | frames--; |
| 359 | } |
| 360 | |
| 361 | { |
| 362 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 363 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 364 | while (resamp->time < phases) |
| 365 | { |
| 366 | int i; |
| 367 | __m256 delta; |
| 368 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 369 | float *phase_table = resamp->phase_table + phase * taps; |
| 370 | |
| 371 | __m256 sum_l = _mm256_setzero_ps(); |
| 372 | __m256 sum_r = _mm256_setzero_ps(); |
| 373 | |
| 374 | for (i = 0; i < (int)taps; i += 8) |
| 375 | { |
| 376 | __m256 buf_l = _mm256_loadu_ps(buffer_l + i); |
| 377 | __m256 buf_r = _mm256_loadu_ps(buffer_r + i); |
| 378 | __m256 sinc = _mm256_load_ps((const float*)phase_table + i); |
| 379 | |
| 380 | sum_l = _mm256_add_ps(sum_l, _mm256_mul_ps(buf_l, sinc)); |
| 381 | sum_r = _mm256_add_ps(sum_r, _mm256_mul_ps(buf_r, sinc)); |
| 382 | } |
| 383 | |
| 384 | /* hadd on AVX is weird, and acts on low-lanes |
| 385 | * and high-lanes separately. */ |
| 386 | __m256 res_l = _mm256_hadd_ps(sum_l, sum_l); |
| 387 | __m256 res_r = _mm256_hadd_ps(sum_r, sum_r); |
| 388 | res_l = _mm256_hadd_ps(res_l, res_l); |
| 389 | res_r = _mm256_hadd_ps(res_r, res_r); |
| 390 | res_l = _mm256_add_ps(_mm256_permute2f128_ps(res_l, res_l, 1), res_l); |
| 391 | res_r = _mm256_add_ps(_mm256_permute2f128_ps(res_r, res_r, 1), res_r); |
| 392 | |
| 393 | /* This is optimized to mov %xmmN, [mem]. |
| 394 | * There doesn't seem to be any _mm256_store_ss intrinsic. */ |
| 395 | _mm_store_ss(output + 0, _mm256_extractf128_ps(res_l, 0)); |
| 396 | _mm_store_ss(output + 1, _mm256_extractf128_ps(res_r, 0)); |
| 397 | |
| 398 | output += 2; |
| 399 | out_frames++; |
| 400 | resamp->time += ratio; |
| 401 | } |
| 402 | } |
| 403 | } |
| 404 | } |
| 405 | |
| 406 | data->output_frames = out_frames; |
| 407 | } |
| 408 | #endif |
| 409 | |
| 410 | #if defined(__SSE__) |
| 411 | static void resampler_sinc_process_sse_kaiser(void *re_, struct resampler_data *data) |
| 412 | { |
| 413 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 414 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 415 | |
| 416 | uint32_t ratio = phases / data->ratio; |
| 417 | const float *input = data->data_in; |
| 418 | float *output = data->data_out; |
| 419 | size_t frames = data->input_frames; |
| 420 | size_t out_frames = 0; |
| 421 | unsigned taps = resamp->taps; |
| 422 | |
| 423 | { |
| 424 | while (frames) |
| 425 | { |
| 426 | while (frames && resamp->time >= phases) |
| 427 | { |
| 428 | /* Push in reverse to make filter more obvious. */ |
| 429 | if (!resamp->ptr) |
| 430 | resamp->ptr = taps; |
| 431 | resamp->ptr--; |
| 432 | |
| 433 | resamp->buffer_l[resamp->ptr + taps] = |
| 434 | resamp->buffer_l[resamp->ptr] = *input++; |
| 435 | |
| 436 | resamp->buffer_r[resamp->ptr + taps] = |
| 437 | resamp->buffer_r[resamp->ptr] = *input++; |
| 438 | |
| 439 | resamp->time -= phases; |
| 440 | frames--; |
| 441 | } |
| 442 | |
| 443 | { |
| 444 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 445 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 446 | while (resamp->time < phases) |
| 447 | { |
| 448 | int i; |
| 449 | __m128 sum; |
| 450 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 451 | float *phase_table = resamp->phase_table + phase * taps * 2; |
| 452 | float *delta_table = phase_table + taps; |
| 453 | __m128 delta = _mm_set1_ps((float) |
| 454 | (resamp->time & resamp->subphase_mask) * resamp->subphase_mod); |
| 455 | |
| 456 | __m128 sum_l = _mm_setzero_ps(); |
| 457 | __m128 sum_r = _mm_setzero_ps(); |
| 458 | |
| 459 | for (i = 0; i < (int)taps; i += 4) |
| 460 | { |
| 461 | __m128 buf_l = _mm_loadu_ps(buffer_l + i); |
| 462 | __m128 buf_r = _mm_loadu_ps(buffer_r + i); |
| 463 | __m128 deltas = _mm_load_ps(delta_table + i); |
| 464 | __m128 _sinc = _mm_add_ps(_mm_load_ps((const float*)phase_table + i), |
| 465 | _mm_mul_ps(deltas, delta)); |
| 466 | sum_l = _mm_add_ps(sum_l, _mm_mul_ps(buf_l, _sinc)); |
| 467 | sum_r = _mm_add_ps(sum_r, _mm_mul_ps(buf_r, _sinc)); |
| 468 | } |
| 469 | |
| 470 | /* Them annoying shuffles. |
| 471 | * sum_l = { l3, l2, l1, l0 } |
| 472 | * sum_r = { r3, r2, r1, r0 } |
| 473 | */ |
| 474 | |
| 475 | sum = _mm_add_ps(_mm_shuffle_ps(sum_l, sum_r, |
| 476 | _MM_SHUFFLE(1, 0, 1, 0)), |
| 477 | _mm_shuffle_ps(sum_l, sum_r, _MM_SHUFFLE(3, 2, 3, 2))); |
| 478 | |
| 479 | /* sum = { r1, r0, l1, l0 } + { r3, r2, l3, l2 } |
| 480 | * sum = { R1, R0, L1, L0 } |
| 481 | */ |
| 482 | |
| 483 | sum = _mm_add_ps(_mm_shuffle_ps(sum, sum, _MM_SHUFFLE(3, 3, 1, 1)), sum); |
| 484 | |
| 485 | /* sum = {R1, R1, L1, L1 } + { R1, R0, L1, L0 } |
| 486 | * sum = { X, R, X, L } |
| 487 | */ |
| 488 | |
| 489 | /* Store L */ |
| 490 | _mm_store_ss(output + 0, sum); |
| 491 | |
| 492 | /* movehl { X, R, X, L } == { X, R, X, R } */ |
| 493 | _mm_store_ss(output + 1, _mm_movehl_ps(sum, sum)); |
| 494 | |
| 495 | output += 2; |
| 496 | out_frames++; |
| 497 | resamp->time += ratio; |
| 498 | } |
| 499 | } |
| 500 | } |
| 501 | } |
| 502 | |
| 503 | data->output_frames = out_frames; |
| 504 | } |
| 505 | |
| 506 | static void resampler_sinc_process_sse(void *re_, struct resampler_data *data) |
| 507 | { |
| 508 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 509 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 510 | |
| 511 | uint32_t ratio = phases / data->ratio; |
| 512 | const float *input = data->data_in; |
| 513 | float *output = data->data_out; |
| 514 | size_t frames = data->input_frames; |
| 515 | size_t out_frames = 0; |
| 516 | unsigned taps = resamp->taps; |
| 517 | |
| 518 | { |
| 519 | while (frames) |
| 520 | { |
| 521 | while (frames && resamp->time >= phases) |
| 522 | { |
| 523 | /* Push in reverse to make filter more obvious. */ |
| 524 | if (!resamp->ptr) |
| 525 | resamp->ptr = taps; |
| 526 | resamp->ptr--; |
| 527 | |
| 528 | resamp->buffer_l[resamp->ptr + taps] = |
| 529 | resamp->buffer_l[resamp->ptr] = *input++; |
| 530 | |
| 531 | resamp->buffer_r[resamp->ptr + taps] = |
| 532 | resamp->buffer_r[resamp->ptr] = *input++; |
| 533 | |
| 534 | resamp->time -= phases; |
| 535 | frames--; |
| 536 | } |
| 537 | |
| 538 | { |
| 539 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 540 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 541 | while (resamp->time < phases) |
| 542 | { |
| 543 | int i; |
| 544 | __m128 sum; |
| 545 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 546 | float *phase_table = resamp->phase_table + phase * taps; |
| 547 | |
| 548 | __m128 sum_l = _mm_setzero_ps(); |
| 549 | __m128 sum_r = _mm_setzero_ps(); |
| 550 | |
| 551 | for (i = 0; i < (int)taps; i += 4) |
| 552 | { |
| 553 | __m128 buf_l = _mm_loadu_ps(buffer_l + i); |
| 554 | __m128 buf_r = _mm_loadu_ps(buffer_r + i); |
| 555 | __m128 _sinc = _mm_load_ps((const float*)phase_table + i); |
| 556 | sum_l = _mm_add_ps(sum_l, _mm_mul_ps(buf_l, _sinc)); |
| 557 | sum_r = _mm_add_ps(sum_r, _mm_mul_ps(buf_r, _sinc)); |
| 558 | } |
| 559 | |
| 560 | /* Them annoying shuffles. |
| 561 | * sum_l = { l3, l2, l1, l0 } |
| 562 | * sum_r = { r3, r2, r1, r0 } |
| 563 | */ |
| 564 | |
| 565 | sum = _mm_add_ps(_mm_shuffle_ps(sum_l, sum_r, |
| 566 | _MM_SHUFFLE(1, 0, 1, 0)), |
| 567 | _mm_shuffle_ps(sum_l, sum_r, _MM_SHUFFLE(3, 2, 3, 2))); |
| 568 | |
| 569 | /* sum = { r1, r0, l1, l0 } + { r3, r2, l3, l2 } |
| 570 | * sum = { R1, R0, L1, L0 } |
| 571 | */ |
| 572 | |
| 573 | sum = _mm_add_ps(_mm_shuffle_ps(sum, sum, _MM_SHUFFLE(3, 3, 1, 1)), sum); |
| 574 | |
| 575 | /* sum = {R1, R1, L1, L1 } + { R1, R0, L1, L0 } |
| 576 | * sum = { X, R, X, L } |
| 577 | */ |
| 578 | |
| 579 | /* Store L */ |
| 580 | _mm_store_ss(output + 0, sum); |
| 581 | |
| 582 | /* movehl { X, R, X, L } == { X, R, X, R } */ |
| 583 | _mm_store_ss(output + 1, _mm_movehl_ps(sum, sum)); |
| 584 | |
| 585 | output += 2; |
| 586 | out_frames++; |
| 587 | resamp->time += ratio; |
| 588 | } |
| 589 | } |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | data->output_frames = out_frames; |
| 594 | } |
| 595 | #endif |
| 596 | |
| 597 | static void resampler_sinc_process_c_kaiser(void *re_, struct resampler_data *data) |
| 598 | { |
| 599 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 600 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 601 | |
| 602 | uint32_t ratio = phases / data->ratio; |
| 603 | const float *input = data->data_in; |
| 604 | float *output = data->data_out; |
| 605 | size_t frames = data->input_frames; |
| 606 | size_t out_frames = 0; |
| 607 | unsigned taps = resamp->taps; |
| 608 | |
| 609 | { |
| 610 | while (frames) |
| 611 | { |
| 612 | while (frames && resamp->time >= phases) |
| 613 | { |
| 614 | /* Push in reverse to make filter more obvious. */ |
| 615 | if (!resamp->ptr) |
| 616 | resamp->ptr = taps; |
| 617 | resamp->ptr--; |
| 618 | |
| 619 | resamp->buffer_l[resamp->ptr + taps] = |
| 620 | resamp->buffer_l[resamp->ptr] = *input++; |
| 621 | |
| 622 | resamp->buffer_r[resamp->ptr + taps] = |
| 623 | resamp->buffer_r[resamp->ptr] = *input++; |
| 624 | |
| 625 | resamp->time -= phases; |
| 626 | frames--; |
| 627 | } |
| 628 | |
| 629 | { |
| 630 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 631 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 632 | while (resamp->time < phases) |
| 633 | { |
| 634 | int i; |
| 635 | float sum_l = 0.0f; |
| 636 | float sum_r = 0.0f; |
| 637 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 638 | float *phase_table = resamp->phase_table + phase * taps * 2; |
| 639 | float *delta_table = phase_table + taps; |
| 640 | float delta = (float) |
| 641 | (resamp->time & resamp->subphase_mask) * resamp->subphase_mod; |
| 642 | |
| 643 | for (i = 0; i < (int)taps; i++) |
| 644 | { |
| 645 | float sinc_val = phase_table[i] + delta_table[i] * delta; |
| 646 | |
| 647 | sum_l += buffer_l[i] * sinc_val; |
| 648 | sum_r += buffer_r[i] * sinc_val; |
| 649 | } |
| 650 | |
| 651 | output[0] = sum_l; |
| 652 | output[1] = sum_r; |
| 653 | |
| 654 | output += 2; |
| 655 | out_frames++; |
| 656 | resamp->time += ratio; |
| 657 | } |
| 658 | } |
| 659 | |
| 660 | } |
| 661 | } |
| 662 | |
| 663 | data->output_frames = out_frames; |
| 664 | } |
| 665 | |
| 666 | static void resampler_sinc_process_c(void *re_, struct resampler_data *data) |
| 667 | { |
| 668 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)re_; |
| 669 | unsigned phases = 1 << (resamp->phase_bits + resamp->subphase_bits); |
| 670 | |
| 671 | uint32_t ratio = phases / data->ratio; |
| 672 | const float *input = data->data_in; |
| 673 | float *output = data->data_out; |
| 674 | size_t frames = data->input_frames; |
| 675 | size_t out_frames = 0; |
| 676 | unsigned taps = resamp->taps; |
| 677 | |
| 678 | { |
| 679 | while (frames) |
| 680 | { |
| 681 | while (frames && resamp->time >= phases) |
| 682 | { |
| 683 | /* Push in reverse to make filter more obvious. */ |
| 684 | if (!resamp->ptr) |
| 685 | resamp->ptr = taps; |
| 686 | resamp->ptr--; |
| 687 | |
| 688 | resamp->buffer_l[resamp->ptr + taps] = |
| 689 | resamp->buffer_l[resamp->ptr] = *input++; |
| 690 | |
| 691 | resamp->buffer_r[resamp->ptr + taps] = |
| 692 | resamp->buffer_r[resamp->ptr] = *input++; |
| 693 | |
| 694 | resamp->time -= phases; |
| 695 | frames--; |
| 696 | } |
| 697 | |
| 698 | { |
| 699 | const float *buffer_l = resamp->buffer_l + resamp->ptr; |
| 700 | const float *buffer_r = resamp->buffer_r + resamp->ptr; |
| 701 | while (resamp->time < phases) |
| 702 | { |
| 703 | int i; |
| 704 | float sum_l = 0.0f; |
| 705 | float sum_r = 0.0f; |
| 706 | unsigned phase = resamp->time >> resamp->subphase_bits; |
| 707 | float *phase_table = resamp->phase_table + phase * taps; |
| 708 | |
| 709 | for (i = 0; i < (int)taps; i++) |
| 710 | { |
| 711 | float sinc_val = phase_table[i]; |
| 712 | |
| 713 | sum_l += buffer_l[i] * sinc_val; |
| 714 | sum_r += buffer_r[i] * sinc_val; |
| 715 | } |
| 716 | |
| 717 | output[0] = sum_l; |
| 718 | output[1] = sum_r; |
| 719 | |
| 720 | output += 2; |
| 721 | out_frames++; |
| 722 | resamp->time += ratio; |
| 723 | } |
| 724 | } |
| 725 | |
| 726 | } |
| 727 | } |
| 728 | |
| 729 | data->output_frames = out_frames; |
| 730 | } |
| 731 | |
| 732 | static void resampler_sinc_free(void *data) |
| 733 | { |
| 734 | rarch_sinc_resampler_t *resamp = (rarch_sinc_resampler_t*)data; |
| 735 | if (resamp) |
| 736 | memalign_free(resamp->main_buffer); |
| 737 | free(resamp); |
| 738 | } |
| 739 | |
| 740 | static void sinc_init_table_kaiser(rarch_sinc_resampler_t *resamp, |
| 741 | double cutoff, |
| 742 | float *phase_table, int phases, int taps, bool calculate_delta) |
| 743 | { |
| 744 | int i, j; |
| 745 | /* Kaiser window function - need to normalize w(0) to 1.0f */ |
| 746 | float kaiser_beta = resamp->kaiser_beta; |
| 747 | double window_mod = besseli0(kaiser_beta); |
| 748 | int stride = calculate_delta ? 2 : 1; |
| 749 | double sidelobes = taps / 2.0; |
| 750 | |
| 751 | for (i = 0; i < phases; i++) |
| 752 | { |
| 753 | for (j = 0; j < taps; j++) |
| 754 | { |
| 755 | float val; |
| 756 | double sinc_phase; |
| 757 | int n = j * phases + i; |
| 758 | double window_phase = (double)n / (phases * taps); /* [0, 1). */ |
| 759 | window_phase = 2.0 * window_phase - 1.0; /* [-1, 1) */ |
| 760 | sinc_phase = sidelobes * window_phase; |
| 761 | val = cutoff * sinc(M_PI * sinc_phase * cutoff) * |
| 762 | besseli0(kaiser_beta * sqrtf(1 - window_phase * window_phase)) |
| 763 | / window_mod; |
| 764 | phase_table[i * stride * taps + j] = val; |
| 765 | } |
| 766 | } |
| 767 | |
| 768 | if (calculate_delta) |
| 769 | { |
| 770 | int phase; |
| 771 | int p; |
| 772 | |
| 773 | for (p = 0; p < phases - 1; p++) |
| 774 | { |
| 775 | for (j = 0; j < taps; j++) |
| 776 | { |
| 777 | float delta = phase_table[(p + 1) * stride * taps + j] - |
| 778 | phase_table[p * stride * taps + j]; |
| 779 | phase_table[(p * stride + 1) * taps + j] = delta; |
| 780 | } |
| 781 | } |
| 782 | |
| 783 | phase = phases - 1; |
| 784 | for (j = 0; j < taps; j++) |
| 785 | { |
| 786 | float val, delta; |
| 787 | double sinc_phase; |
| 788 | int n = j * phases + (phase + 1); |
| 789 | double window_phase = (double)n / (phases * taps); /* (0, 1]. */ |
| 790 | window_phase = 2.0 * window_phase - 1.0; /* (-1, 1] */ |
| 791 | sinc_phase = sidelobes * window_phase; |
| 792 | |
| 793 | val = cutoff * sinc(M_PI * sinc_phase * cutoff) * |
| 794 | besseli0(resamp->kaiser_beta * sqrtf(1 - window_phase * |
| 795 | window_phase)) / window_mod; |
| 796 | delta = (val - phase_table[phase * stride * taps + j]); |
| 797 | phase_table[(phase * stride + 1) * taps + j] = delta; |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | static void sinc_init_table_lanczos( |
| 803 | rarch_sinc_resampler_t *resamp, double cutoff, |
| 804 | float *phase_table, int phases, int taps, bool calculate_delta) |
| 805 | { |
| 806 | int i, j; |
| 807 | /* Lanczos window function - need to normalize w(0) to 1.0f */ |
| 808 | double window_mod = 1.0; |
| 809 | int stride = calculate_delta ? 2 : 1; |
| 810 | double sidelobes = taps / 2.0; |
| 811 | |
| 812 | for (i = 0; i < phases; i++) |
| 813 | { |
| 814 | for (j = 0; j < taps; j++) |
| 815 | { |
| 816 | double sinc_phase; |
| 817 | float val; |
| 818 | int n = j * phases + i; |
| 819 | double window_phase = (double)n / (phases * taps); /* [0, 1). */ |
| 820 | window_phase = 2.0 * window_phase - 1.0; /* [-1, 1) */ |
| 821 | sinc_phase = sidelobes * window_phase; |
| 822 | val = cutoff * sinc(M_PI * sinc_phase * cutoff) * |
| 823 | sinc(M_PI * window_phase) / window_mod; |
| 824 | phase_table[i * stride * taps + j] = val; |
| 825 | } |
| 826 | } |
| 827 | |
| 828 | if (calculate_delta) |
| 829 | { |
| 830 | int phase; |
| 831 | int p; |
| 832 | |
| 833 | for (p = 0; p < phases - 1; p++) |
| 834 | { |
| 835 | for (j = 0; j < taps; j++) |
| 836 | { |
| 837 | float delta = phase_table[(p + 1) * stride * taps + j] - |
| 838 | phase_table[p * stride * taps + j]; |
| 839 | phase_table[(p * stride + 1) * taps + j] = delta; |
| 840 | } |
| 841 | } |
| 842 | |
| 843 | phase = phases - 1; |
| 844 | for (j = 0; j < taps; j++) |
| 845 | { |
| 846 | float val, delta; |
| 847 | double sinc_phase; |
| 848 | int n = j * phases + (phase + 1); |
| 849 | double window_phase = (double)n / (phases * taps); /* (0, 1]. */ |
| 850 | window_phase = 2.0 * window_phase - 1.0; /* (-1, 1] */ |
| 851 | sinc_phase = sidelobes * window_phase; |
| 852 | |
| 853 | val = cutoff * sinc(M_PI * sinc_phase * cutoff) * |
| 854 | sinc(M_PI * window_phase) / window_mod; |
| 855 | delta = (val - phase_table[phase * stride * taps + j]); |
| 856 | phase_table[(phase * stride + 1) * taps + j] = delta; |
| 857 | } |
| 858 | } |
| 859 | } |
| 860 | |
| 861 | static void *resampler_sinc_new(const struct resampler_config *config, |
| 862 | double bandwidth_mod, enum resampler_quality quality, |
| 863 | resampler_simd_mask_t mask) |
| 864 | { |
| 865 | double cutoff = 0.0; |
| 866 | size_t phase_elems = 0; |
| 867 | size_t elems = 0; |
| 868 | unsigned enable_avx = 0; |
| 869 | unsigned sidelobes = 0; |
| 870 | enum sinc_window window_type = SINC_WINDOW_NONE; |
| 871 | rarch_sinc_resampler_t *re = (rarch_sinc_resampler_t*) |
| 872 | calloc(1, sizeof(*re)); |
| 873 | |
| 874 | if (!re) |
| 875 | return NULL; |
| 876 | |
| 877 | switch (quality) |
| 878 | { |
| 879 | case RESAMPLER_QUALITY_LOWEST: |
| 880 | cutoff = 0.98; |
| 881 | sidelobes = 2; |
| 882 | re->phase_bits = 12; |
| 883 | re->subphase_bits = 10; |
| 884 | window_type = SINC_WINDOW_LANCZOS; |
| 885 | break; |
| 886 | case RESAMPLER_QUALITY_LOWER: |
| 887 | cutoff = 0.98; |
| 888 | sidelobes = 4; |
| 889 | re->phase_bits = 12; |
| 890 | re->subphase_bits = 10; |
| 891 | window_type = SINC_WINDOW_LANCZOS; |
| 892 | break; |
| 893 | case RESAMPLER_QUALITY_HIGHER: |
| 894 | cutoff = 0.90; |
| 895 | sidelobes = 32; |
| 896 | re->phase_bits = 10; |
| 897 | re->subphase_bits = 14; |
| 898 | window_type = SINC_WINDOW_KAISER; |
| 899 | re->kaiser_beta = 10.5; |
| 900 | enable_avx = 1; |
| 901 | break; |
| 902 | case RESAMPLER_QUALITY_HIGHEST: |
| 903 | cutoff = 0.962; |
| 904 | sidelobes = 128; |
| 905 | re->phase_bits = 10; |
| 906 | re->subphase_bits = 14; |
| 907 | window_type = SINC_WINDOW_KAISER; |
| 908 | re->kaiser_beta = 14.5; |
| 909 | enable_avx = 1; |
| 910 | break; |
| 911 | case RESAMPLER_QUALITY_NORMAL: |
| 912 | case RESAMPLER_QUALITY_DONTCARE: |
| 913 | cutoff = 0.825; |
| 914 | sidelobes = 8; |
| 915 | re->phase_bits = 8; |
| 916 | re->subphase_bits = 16; |
| 917 | window_type = SINC_WINDOW_KAISER; |
| 918 | re->kaiser_beta = 5.5; |
| 919 | break; |
| 920 | } |
| 921 | |
| 922 | re->subphase_mask = (1 << re->subphase_bits) - 1; |
| 923 | re->subphase_mod = 1.0f / (1 << re->subphase_bits); |
| 924 | re->taps = sidelobes * 2; |
| 925 | |
| 926 | /* Downsampling, must lower cutoff, and extend number of |
| 927 | * taps accordingly to keep same stopband attenuation. */ |
| 928 | if (bandwidth_mod < 1.0) |
| 929 | { |
| 930 | cutoff *= bandwidth_mod; |
| 931 | re->taps = (unsigned)ceil(re->taps / bandwidth_mod); |
| 932 | } |
| 933 | |
| 934 | /* Be SIMD-friendly. */ |
| 935 | #if defined(__AVX__) |
| 936 | if (enable_avx) |
| 937 | re->taps = (re->taps + 7) & ~7; |
| 938 | else |
| 939 | #endif |
| 940 | { |
| 941 | #if (defined(__ARM_NEON__) || defined(HAVE_NEON)) |
| 942 | re->taps = (re->taps + 7) & ~7; |
| 943 | #else |
| 944 | re->taps = (re->taps + 3) & ~3; |
| 945 | #endif |
| 946 | } |
| 947 | |
| 948 | phase_elems = ((1 << re->phase_bits) * re->taps); |
| 949 | if (window_type == SINC_WINDOW_KAISER) |
| 950 | phase_elems = phase_elems * 2; |
| 951 | elems = phase_elems + 4 * re->taps; |
| 952 | |
| 953 | re->main_buffer = (float*)memalign_alloc(128, sizeof(float) * elems); |
| 954 | if (!re->main_buffer) |
| 955 | goto error; |
| 956 | |
| 957 | memset(re->main_buffer, 0, sizeof(float) * elems); |
| 958 | |
| 959 | re->phase_table = re->main_buffer; |
| 960 | re->buffer_l = re->main_buffer + phase_elems; |
| 961 | re->buffer_r = re->buffer_l + 2 * re->taps; |
| 962 | |
| 963 | switch (window_type) |
| 964 | { |
| 965 | case SINC_WINDOW_LANCZOS: |
| 966 | sinc_init_table_lanczos(re, cutoff, re->phase_table, |
| 967 | 1 << re->phase_bits, re->taps, false); |
| 968 | break; |
| 969 | case SINC_WINDOW_KAISER: |
| 970 | sinc_init_table_kaiser(re, cutoff, re->phase_table, |
| 971 | 1 << re->phase_bits, re->taps, true); |
| 972 | break; |
| 973 | case SINC_WINDOW_NONE: |
| 974 | goto error; |
| 975 | } |
| 976 | |
| 977 | sinc_resampler.process = resampler_sinc_process_c; |
| 978 | if (window_type == SINC_WINDOW_KAISER) |
| 979 | sinc_resampler.process = resampler_sinc_process_c_kaiser; |
| 980 | |
| 981 | if (mask & RESAMPLER_SIMD_AVX && enable_avx) |
| 982 | { |
| 983 | #if defined(__AVX__) |
| 984 | sinc_resampler.process = resampler_sinc_process_avx; |
| 985 | if (window_type == SINC_WINDOW_KAISER) |
| 986 | sinc_resampler.process = resampler_sinc_process_avx_kaiser; |
| 987 | #endif |
| 988 | } |
| 989 | else if (mask & RESAMPLER_SIMD_SSE) |
| 990 | { |
| 991 | #if defined(__SSE__) |
| 992 | sinc_resampler.process = resampler_sinc_process_sse; |
| 993 | if (window_type == SINC_WINDOW_KAISER) |
| 994 | sinc_resampler.process = resampler_sinc_process_sse_kaiser; |
| 995 | #endif |
| 996 | } |
| 997 | else if (mask & RESAMPLER_SIMD_NEON) |
| 998 | { |
| 999 | #if (defined(__ARM_NEON__) || defined(HAVE_NEON)) |
| 1000 | #ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS |
| 1001 | if (window_type != SINC_WINDOW_KAISER) |
| 1002 | sinc_resampler.process = resampler_sinc_process_neon; |
| 1003 | #else |
| 1004 | sinc_resampler.process = resampler_sinc_process_neon; |
| 1005 | if (window_type == SINC_WINDOW_KAISER) |
| 1006 | sinc_resampler.process = resampler_sinc_process_neon_kaiser; |
| 1007 | #endif |
| 1008 | #endif |
| 1009 | } |
| 1010 | |
| 1011 | return re; |
| 1012 | |
| 1013 | error: |
| 1014 | resampler_sinc_free(re); |
| 1015 | return NULL; |
| 1016 | } |
| 1017 | |
| 1018 | retro_resampler_t sinc_resampler = { |
| 1019 | resampler_sinc_new, |
| 1020 | resampler_sinc_process_c, |
| 1021 | resampler_sinc_free, |
| 1022 | RESAMPLER_API_VERSION, |
| 1023 | "sinc", |
| 1024 | "sinc" |
| 1025 | }; |