| 1 |
use std::f64; |
| 2 |
|
| 3 |
use rustfft::Length; |
| 4 |
|
| 5 |
use crate::common::mdct_error_inplace; |
| 6 |
use crate::RequiredScratch; |
| 7 |
use crate::{mdct::Mdct, DctNum}; |
| 8 |
|
| 9 |
/// Naive O(n^2 ) MDCT implementation |
| 10 |
/// |
| 11 |
/// This implementation is primarily used to test other MDCT algorithms. |
| 12 |
/// |
| 13 |
/// ~~~ |
| 14 |
/// // Computes a naive MDCT of output size 124, using the MP3 window function |
| 15 |
/// use rustdct::mdct::{Mdct, MdctNaive, window_fn}; |
| 16 |
/// use rustdct::RequiredScratch; |
| 17 |
/// |
| 18 |
/// let len = 124; |
| 19 |
/// |
| 20 |
/// let dct = MdctNaive::new(len, window_fn::mp3); |
| 21 |
/// |
| 22 |
/// let input = vec![0f32; len * 2]; |
| 23 |
/// let (input_a, input_b) = input.split_at(len); |
| 24 |
/// let mut output = vec![0f32; len]; |
| 25 |
/// let mut scratch = vec![0f32; dct.get_scratch_len()]; |
| 26 |
/// |
| 27 |
/// dct.process_mdct_with_scratch(input_a, input_b, &mut output, &mut scratch); |
| 28 |
/// ~~~ |
| 29 |
pub struct MdctNaive<T> { |
| 30 |
twiddles: Box<[T]>, |
| 31 |
window: Box<[T]>, |
| 32 |
} |
| 33 |
|
| 34 |
impl<T: DctNum> MdctNaive<T> { |
| 35 |
/// Creates a new MDCT context that will process inputs of length `output_len * 2` and produce |
| 36 |
/// outputs of length `output_len` |
| 37 |
/// |
| 38 |
/// `output_len` must be even. |
| 39 |
/// |
| 40 |
/// `window_fn` is a function that takes a `size` and returns a `Vec` containing `size` window values. |
| 41 |
/// See the [`window_fn`](mdct/window_fn/index.html) module for provided window functions. |
| 42 |
pub fn new<F>(output_len: usize, window_fn: F) -> Self |
| 43 |
where |
| 44 |
F: FnOnce(usize) -> Vec<T>, |
| 45 |
{ |
| 46 |
assert!( |
| 47 |
output_len % 2 == 0, |
| 48 |
"The MDCT len must be even. Got {}", |
| 49 |
output_len |
| 50 |
); |
| 51 |
|
| 52 |
let constant_factor = 0.5f64 * f64::consts::PI / (output_len as f64); |
| 53 |
let twiddles: Vec<T> = (0..output_len * 4) |
| 54 |
.map(|i| (constant_factor * (i as f64 + 0.5_f64)).cos()) |
| 55 |
.map(|c| T::from_f64(c).unwrap()) |
| 56 |
.collect(); |
| 57 |
|
| 58 |
let window = window_fn(output_len * 2); |
| 59 |
assert_eq!( |
| 60 |
window.len(), |
| 61 |
output_len * 2, |
| 62 |
"Window function returned incorrect number of values" |
| 63 |
); |
| 64 |
|
| 65 |
Self { |
| 66 |
twiddles: twiddles.into_boxed_slice(), |
| 67 |
window: window.into_boxed_slice(), |
| 68 |
} |
| 69 |
} |
| 70 |
} |
| 71 |
|
| 72 |
impl<T: DctNum> Mdct<T> for MdctNaive<T> { |
| 73 |
fn process_mdct_with_scratch( |
| 74 |
&self, |
| 75 |
input_a: &[T], |
| 76 |
input_b: &[T], |
| 77 |
output: &mut [T], |
| 78 |
scratch: &mut [T], |
| 79 |
) { |
| 80 |
validate_buffers_mdct!( |
| 81 |
input_a, |
| 82 |
input_b, |
| 83 |
output, |
| 84 |
scratch, |
| 85 |
self.len(), |
| 86 |
self.get_scratch_len() |
| 87 |
); |
| 88 |
|
| 89 |
let output_len = output.len(); |
| 90 |
let half_output = output.len() / 2; |
| 91 |
|
| 92 |
for k in 0..output_len { |
| 93 |
let output_cell = output.get_mut(k).unwrap(); |
| 94 |
*output_cell = T::zero(); |
| 95 |
|
| 96 |
let mut twiddle_index = (half_output + k * (output_len + 1)) % self.twiddles.len(); |
| 97 |
let twiddle_stride = k * 2 + 1; |
| 98 |
|
| 99 |
for i in 0..input_a.len() { |
| 100 |
let twiddle = self.twiddles[twiddle_index]; |
| 101 |
|
| 102 |
*output_cell = *output_cell + input_a[i] * self.window[i] * twiddle; |
| 103 |
|
| 104 |
twiddle_index += twiddle_stride; |
| 105 |
if twiddle_index >= self.twiddles.len() { |
| 106 |
twiddle_index -= self.twiddles.len(); |
| 107 |
} |
| 108 |
} |
| 109 |
|
| 110 |
for i in 0..input_b.len() { |
| 111 |
let twiddle = self.twiddles[twiddle_index]; |
| 112 |
|
| 113 |
*output_cell = *output_cell + input_b[i] * self.window[i + output_len] * twiddle; |
| 114 |
|
| 115 |
twiddle_index += twiddle_stride; |
| 116 |
if twiddle_index >= self.twiddles.len() { |
| 117 |
twiddle_index -= self.twiddles.len(); |
| 118 |
} |
| 119 |
} |
| 120 |
} |
| 121 |
} |
| 122 |
fn process_imdct_with_scratch( |
| 123 |
&self, |
| 124 |
input: &[T], |
| 125 |
output_a: &mut [T], |
| 126 |
output_b: &mut [T], |
| 127 |
scratch: &mut [T], |
| 128 |
) { |
| 129 |
validate_buffers_mdct!( |
| 130 |
input, |
| 131 |
output_a, |
| 132 |
output_b, |
| 133 |
scratch, |
| 134 |
self.len(), |
| 135 |
self.get_scratch_len() |
| 136 |
); |
| 137 |
|
| 138 |
let input_len = input.len(); |
| 139 |
let half_input = input_len / 2; |
| 140 |
|
| 141 |
for k in 0..input_len { |
| 142 |
let mut output_val = T::zero(); |
| 143 |
|
| 144 |
let mut twiddle_index = half_input + k; |
| 145 |
let twiddle_stride = input_len + k * 2 + 1; |
| 146 |
|
| 147 |
for i in 0..input.len() { |
| 148 |
let twiddle = self.twiddles[twiddle_index]; |
| 149 |
|
| 150 |
output_val = output_val + input[i] * twiddle; |
| 151 |
|
| 152 |
twiddle_index += twiddle_stride; |
| 153 |
if twiddle_index >= self.twiddles.len() { |
| 154 |
twiddle_index -= self.twiddles.len(); |
| 155 |
} |
| 156 |
} |
| 157 |
output_a[k] = output_a[k] + output_val * self.window[k]; |
| 158 |
} |
| 159 |
|
| 160 |
for k in 0..input_len { |
| 161 |
let mut output_val = T::zero(); |
| 162 |
|
| 163 |
let twiddle_k = if k < input_len / 2 { |
| 164 |
k |
| 165 |
} else { |
| 166 |
input_len - k - 1 |
| 167 |
}; |
| 168 |
|
| 169 |
let mut twiddle_index = |
| 170 |
self.twiddles.len() - input_len * 2 + half_input - twiddle_k - 1; |
| 171 |
let twiddle_stride = input_len - 1 - twiddle_k * 2; |
| 172 |
|
| 173 |
for i in 0..input.len() { |
| 174 |
let twiddle = self.twiddles[twiddle_index]; |
| 175 |
|
| 176 |
output_val = output_val + input[i] * twiddle; |
| 177 |
|
| 178 |
twiddle_index += twiddle_stride; |
| 179 |
if twiddle_index >= self.twiddles.len() { |
| 180 |
twiddle_index -= self.twiddles.len(); |
| 181 |
} |
| 182 |
} |
| 183 |
output_b[k] = output_b[k] + output_val * self.window[k + input_len]; |
| 184 |
} |
| 185 |
} |
| 186 |
} |
| 187 |
impl<T> Length for MdctNaive<T> { |
| 188 |
fn len(&self) -> usize { |
| 189 |
self.twiddles.len() / 4 |
| 190 |
} |
| 191 |
} |
| 192 |
impl<T> RequiredScratch for MdctNaive<T> { |
| 193 |
fn get_scratch_len(&self) -> usize { |
| 194 |
0 |
| 195 |
} |
| 196 |
} |
| 197 |
|
| 198 |
#[cfg(test)] |
| 199 |
mod unit_tests { |
| 200 |
use super::*; |
| 201 |
use crate::mdct::window_fn; |
| 202 |
use std::f32; |
| 203 |
|
| 204 |
use crate::test_utils::{compare_float_vectors, random_signal}; |
| 205 |
|
| 206 |
/// Verify our naive implementation against some known values |
| 207 |
#[test] |
| 208 |
fn test_known_values_mdct() { |
| 209 |
let input_list = vec![ |
| 210 |
vec![0_f32, 0_f32, 0_f32, 0_f32], |
| 211 |
vec![1_f32, 1_f32, -5_f32, 5_f32], |
| 212 |
vec![7_f32, 3_f32, 8_f32, 4_f32, -1_f32, 3_f32, 0_f32, 4_f32], |
| 213 |
vec![ |
| 214 |
7_f32, 3_f32, 8_f32, 4_f32, -1_f32, 3_f32, 0_f32, 4_f32, 1f32, 1f32, 1f32, 1f32, |
| 215 |
], |
| 216 |
]; |
| 217 |
let expected_list = vec![ |
| 218 |
vec![0_f32, 0_f32], |
| 219 |
vec![0_f32, 0_f32], |
| 220 |
vec![-4.7455063, -2.073643, -2.2964284, 8.479767], |
| 221 |
vec![ |
| 222 |
-2.90775651, |
| 223 |
-12.30026278, |
| 224 |
6.92661442, |
| 225 |
2.79403335, |
| 226 |
3.56420194, |
| 227 |
-2.40007133, |
| 228 |
], |
| 229 |
]; |
| 230 |
|
| 231 |
for (input, expected) in input_list.iter().zip(expected_list.iter()) { |
| 232 |
let output = slow_mdct(&input, window_fn::one); |
| 233 |
|
| 234 |
assert!(compare_float_vectors(&expected, &output)); |
| 235 |
} |
| 236 |
} |
| 237 |
|
| 238 |
/// Verify our naive windowed implementation against some known values |
| 239 |
#[test] |
| 240 |
fn test_known_values_windowed_mdct() { |
| 241 |
let input_list = vec![ |
| 242 |
vec![0_f32, 0_f32, 0_f32, 0_f32], |
| 243 |
vec![1_f32, 1_f32, -5_f32, 5_f32], |
| 244 |
vec![7_f32, 3_f32, 8_f32, 4_f32, -1_f32, 3_f32, 0_f32, 4_f32], |
| 245 |
vec![ |
| 246 |
7_f32, 3_f32, 8_f32, 4_f32, -1_f32, 3_f32, 0_f32, 4_f32, 1f32, 1f32, 1f32, 1f32, |
| 247 |
], |
| 248 |
]; |
| 249 |
let expected_list = vec![ |
| 250 |
vec![0_f32, 0_f32], |
| 251 |
vec![2.29289322, 1.53553391], |
| 252 |
vec![-4.67324308, 3.1647844, -6.22625186, 2.1647844], |
| 253 |
vec![ |
| 254 |
-5.50153067, |
| 255 |
-3.46580575, |
| 256 |
3.79375195, |
| 257 |
-1.25072987, |
| 258 |
4.6738204, |
| 259 |
3.16506351, |
| 260 |
], |
| 261 |
]; |
| 262 |
|
| 263 |
for (input, expected) in input_list.iter().zip(expected_list.iter()) { |
| 264 |
let output = slow_mdct(&input, window_fn::mp3); |
| 265 |
|
| 266 |
assert!(compare_float_vectors(&expected, &output)); |
| 267 |
} |
| 268 |
} |
| 269 |
|
| 270 |
/// Verify that our fast implementation of the MDCT gives the same output as the slow version, for many different inputs |
| 271 |
#[test] |
| 272 |
fn test_matches_mdct() { |
| 273 |
for current_window_fn in &[window_fn::one, window_fn::mp3, window_fn::vorbis] { |
| 274 |
for i in 1..10 { |
| 275 |
let input_len = i * 4; |
| 276 |
let output_len = i * 2; |
| 277 |
|
| 278 |
let input = random_signal(input_len); |
| 279 |
let (input_a, input_b) = input.split_at(output_len); |
| 280 |
|
| 281 |
let slow_output = slow_mdct(&input, current_window_fn); |
| 282 |
|
| 283 |
let mut fast_output = vec![0f32; output_len]; |
| 284 |
|
| 285 |
let dct = MdctNaive::new(output_len, current_window_fn); |
| 286 |
dct.process_mdct_with_scratch(&input_a, &input_b, &mut fast_output, &mut []); |
| 287 |
|
| 288 |
println!("{}", output_len); |
| 289 |
println!("expected: {:?}", slow_output); |
| 290 |
println!("actual: {:?}", fast_output); |
| 291 |
|
| 292 |
assert!( |
| 293 |
compare_float_vectors(&slow_output, &fast_output), |
| 294 |
"i = {}", |
| 295 |
i |
| 296 |
); |
| 297 |
} |
| 298 |
} |
| 299 |
} |
| 300 |
|
| 301 |
fn slow_mdct<F>(input: &[f32], window_fn: F) -> Vec<f32> |
| 302 |
where |
| 303 |
F: Fn(usize) -> Vec<f32>, |
| 304 |
{ |
| 305 |
let mut output = vec![0f32; input.len() / 2]; |
| 306 |
|
| 307 |
let size_float = output.len() as f32; |
| 308 |
|
| 309 |
let window = window_fn(input.len()); |
| 310 |
let windowed_input: Vec<f32> = input.iter().zip(window).map(|(i, w)| i * w).collect(); |
| 311 |
|
| 312 |
for k in 0..output.len() { |
| 313 |
let mut current_value = 0_f32; |
| 314 |
|
| 315 |
let k_float = k as f32; |
| 316 |
|
| 317 |
for n in 0..input.len() { |
| 318 |
let n_float = n as f32; |
| 319 |
|
| 320 |
let twiddle = (f32::consts::PI |
| 321 |
* (n_float + 0.5_f32 + size_float * 0.5) |
| 322 |
* (k_float + 0.5_f32) |
| 323 |
/ size_float) |
| 324 |
.cos(); |
| 325 |
|
| 326 |
current_value += windowed_input[n] * twiddle; |
| 327 |
} |
| 328 |
output[k] = current_value; |
| 329 |
} |
| 330 |
output |
| 331 |
} |
| 332 |
|
| 333 |
/// Verify our naive implementation against some known values |
| 334 |
#[test] |
| 335 |
fn test_known_values_imdct() { |
| 336 |
let input_list = vec![ |
| 337 |
vec![0f32, 0f32], |
| 338 |
vec![1f32, 5f32], |
| 339 |
vec![7f32, 3f32, 8f32, 4f32], |
| 340 |
vec![7f32, 3f32, 8f32, 4f32, -1f32, 3f32], |
| 341 |
]; |
| 342 |
let expected_list = vec![ |
| 343 |
vec![0f32, 0f32, 0f32, 0f32], |
| 344 |
vec![-4.2367144, 4.2367153, -2.837299, -2.8372989], |
| 345 |
vec![ |
| 346 |
5.833236, 2.4275358, -2.4275393, -5.833232, 4.8335495, -14.584825, -14.584811, |
| 347 |
4.8335423, |
| 348 |
], |
| 349 |
vec![ |
| 350 |
2.4138875, |
| 351 |
8.921771, |
| 352 |
-2.4359043, |
| 353 |
2.4359055, |
| 354 |
-8.921769, |
| 355 |
-2.4138737, |
| 356 |
3.1458342, |
| 357 |
-0.63405657, |
| 358 |
-18.502512, |
| 359 |
-18.502502, |
| 360 |
-0.6340414, |
| 361 |
3.1458292, |
| 362 |
], |
| 363 |
]; |
| 364 |
|
| 365 |
for (input, expected) in input_list.iter().zip(expected_list.iter()) { |
| 366 |
let slow_output = slow_imdct(&input, window_fn::one); |
| 367 |
|
| 368 |
let mut fast_output = vec![0f32; input.len() * 2]; |
| 369 |
let (fast_output_a, fast_output_b) = fast_output.split_at_mut(input.len()); |
| 370 |
|
| 371 |
let dct = MdctNaive::new(input.len(), window_fn::one); |
| 372 |
dct.process_imdct_with_scratch(&input, fast_output_a, fast_output_b, &mut []); |
| 373 |
|
| 374 |
assert!(compare_float_vectors(&expected, &slow_output)); |
| 375 |
assert!(compare_float_vectors(&expected, &fast_output)); |
| 376 |
} |
| 377 |
} |
| 378 |
|
| 379 |
/// Verify our naive windowed implementation against some known values |
| 380 |
#[test] |
| 381 |
fn test_known_values_windowed_imdct() { |
| 382 |
let input_list = vec![ |
| 383 |
vec![0f32, 0f32], |
| 384 |
vec![1_f32, 5_f32], |
| 385 |
vec![7_f32, 3_f32, 8_f32, 4_f32], |
| 386 |
vec![7_f32, 3_f32, 8_f32, 4_f32, -1_f32, 3_f32], |
| 387 |
]; |
| 388 |
let expected_list = vec![ |
| 389 |
vec![0f32, 0f32, 0f32, 0f32], |
| 390 |
vec![ |
| 391 |
-1.6213203435596431, |
| 392 |
3.9142135623730936, |
| 393 |
-2.6213203435596433, |
| 394 |
-1.0857864376269069, |
| 395 |
], |
| 396 |
vec![ |
| 397 |
1.1380080486867217, |
| 398 |
1.3486674811260955, |
| 399 |
-2.0184235241728627, |
| 400 |
-5.7211528055198331, |
| 401 |
4.7406716077536428, |
| 402 |
-12.126842074178105, |
| 403 |
-8.1028968193867765, |
| 404 |
0.94297821246780911, |
| 405 |
], |
| 406 |
vec![ |
| 407 |
0.3150751815802082, |
| 408 |
3.4142135623730949, |
| 409 |
-1.4828837895525038, |
| 410 |
1.9325317795197492, |
| 411 |
-8.2426406871192732, |
| 412 |
-2.3932336063055089, |
| 413 |
3.1189227588735786, |
| 414 |
-0.58578643762689731, |
| 415 |
-14.679036212259122, |
| 416 |
-11.263620643186901, |
| 417 |
-0.24264068711929426, |
| 418 |
0.41061397098787894, |
| 419 |
], |
| 420 |
]; |
| 421 |
|
| 422 |
for (input, expected) in input_list.iter().zip(expected_list.iter()) { |
| 423 |
let slow_output = slow_imdct(&input, window_fn::mp3); |
| 424 |
|
| 425 |
let mut fast_output = vec![0f32; input.len() * 2]; |
| 426 |
let (fast_output_a, fast_output_b) = fast_output.split_at_mut(input.len()); |
| 427 |
|
| 428 |
let dct = MdctNaive::new(input.len(), window_fn::mp3); |
| 429 |
dct.process_imdct_with_scratch(&input, fast_output_a, fast_output_b, &mut []); |
| 430 |
|
| 431 |
assert!(compare_float_vectors(&expected, &slow_output)); |
| 432 |
assert!(compare_float_vectors(&expected, &fast_output)); |
| 433 |
} |
| 434 |
} |
| 435 |
|
| 436 |
/// Verify that our fast implementation of the MDCT gives the same output as the slow version, for many different inputs |
| 437 |
#[test] |
| 438 |
fn test_matches_imdct() { |
| 439 |
for current_window_fn in &[window_fn::one, window_fn::mp3, window_fn::vorbis] { |
| 440 |
for i in 1..10 { |
| 441 |
let input_len = i * 2; |
| 442 |
let output_len = i * 4; |
| 443 |
|
| 444 |
let input = random_signal(input_len); |
| 445 |
let slow_output = slow_imdct(&input, current_window_fn); |
| 446 |
|
| 447 |
let mut fast_output = vec![0f32; output_len]; |
| 448 |
let (fast_output_a, fast_output_b) = fast_output.split_at_mut(input_len); |
| 449 |
|
| 450 |
let dct = MdctNaive::new(input_len, current_window_fn); |
| 451 |
dct.process_imdct_with_scratch(&input, fast_output_a, fast_output_b, &mut []); |
| 452 |
|
| 453 |
assert!( |
| 454 |
compare_float_vectors(&slow_output, &fast_output), |
| 455 |
"i = {}", |
| 456 |
i |
| 457 |
); |
| 458 |
} |
| 459 |
} |
| 460 |
} |
| 461 |
|
| 462 |
fn slow_imdct<F>(input: &[f32], window_fn: F) -> Vec<f32> |
| 463 |
where |
| 464 |
F: Fn(usize) -> Vec<f32>, |
| 465 |
{ |
| 466 |
let mut output = vec![0f32; input.len() * 2]; |
| 467 |
|
| 468 |
let size_float = input.len() as f32; |
| 469 |
|
| 470 |
for n in 0..output.len() { |
| 471 |
let mut current_value = 0_f32; |
| 472 |
|
| 473 |
let n_float = n as f32; |
| 474 |
|
| 475 |
for k in 0..input.len() { |
| 476 |
let k_float = k as f32; |
| 477 |
|
| 478 |
let twiddle = (f32::consts::PI |
| 479 |
* (n_float + 0.5_f32 + size_float * 0.5) |
| 480 |
* (k_float + 0.5_f32) |
| 481 |
/ size_float) |
| 482 |
.cos(); |
| 483 |
|
| 484 |
current_value += input[k] * twiddle; |
| 485 |
} |
| 486 |
output[n] = current_value; |
| 487 |
} |
| 488 |
|
| 489 |
let window = window_fn(output.len()); |
| 490 |
output.iter().zip(window).map(|(e, w)| e * w).collect() |
| 491 |
} |
| 492 |
} |