| 1 |
/* |
| 2 |
* Copyright (c) 2023. |
| 3 |
* |
| 4 |
* This software is free software; |
| 5 |
* |
| 6 |
* You can redistribute it or modify it under terms of the MIT, Apache License or Zlib license |
| 7 |
*/ |
| 8 |
|
| 9 |
//! Up-sampling routines |
| 10 |
//! |
| 11 |
//! The main upsampling method is a bi-linear interpolation or a "triangle |
| 12 |
//! filter " or libjpeg turbo `fancy_upsampling` which is a good compromise |
| 13 |
//! between speed and visual quality |
| 14 |
//! |
| 15 |
//! # The filter |
| 16 |
//! Each output pixel is made from `(3*A+B)/4` where A is the original |
| 17 |
//! pixel closer to the output and B is the one further. |
| 18 |
//! |
| 19 |
//! ```text |
| 20 |
//!+---+---+ |
| 21 |
//! | A | B | |
| 22 |
//! +---+---+ |
| 23 |
//! +-+-+-+-+ |
| 24 |
//! | |P| | | |
| 25 |
//! +-+-+-+-+ |
| 26 |
//! ``` |
| 27 |
//! |
| 28 |
//! # Horizontal Bi-linear filter |
| 29 |
//! ```text |
| 30 |
//! |---+-----------+---+ |
| 31 |
//! | | | | |
| 32 |
//! | A | |p1 | p2| | B | |
| 33 |
//! | | | | |
| 34 |
//! |---+-----------+---+ |
| 35 |
//! |
| 36 |
//! ``` |
| 37 |
//! For a horizontal bi-linear it's trivial to implement, |
| 38 |
//! |
| 39 |
//! `A` becomes the input closest to the output. |
| 40 |
//! |
| 41 |
//! `B` varies depending on output. |
| 42 |
//! - For odd positions, input is the `next` pixel after A |
| 43 |
//! - For even positions, input is the `previous` value before A. |
| 44 |
//! |
| 45 |
//! We iterate in a classic 1-D sliding window with a window of 3. |
| 46 |
//! For our sliding window approach, `A` is the 1st and `B` is either the 0th term or 2nd term |
| 47 |
//! depending on position we are writing.(see scalar code). |
| 48 |
//! |
| 49 |
//! For vector code see module sse for explanation. |
| 50 |
//! |
| 51 |
//! # Vertical bi-linear. |
| 52 |
//! Vertical up-sampling is a bit trickier. |
| 53 |
//! |
| 54 |
//! ```text |
| 55 |
//! +----+----+ |
| 56 |
//! | A1 | A2 | |
| 57 |
//! +----+----+ |
| 58 |
//! +----+----+ |
| 59 |
//! | p1 | p2 | |
| 60 |
//! +----+-+--+ |
| 61 |
//! +----+-+--+ |
| 62 |
//! | p3 | p4 | |
| 63 |
//! +----+-+--+ |
| 64 |
//! +----+----+ |
| 65 |
//! | B1 | B2 | |
| 66 |
//! +----+----+ |
| 67 |
//! ``` |
| 68 |
//! |
| 69 |
//! For `p1` |
| 70 |
//! - `A1` is given a weight of `3` and `B1` is given a weight of 1. |
| 71 |
//! |
| 72 |
//! For `p3` |
| 73 |
//! - `B1` is given a weight of `3` and `A1` is given a weight of 1 |
| 74 |
//! |
| 75 |
//! # Horizontal vertical downsampling/chroma quartering. |
| 76 |
//! |
| 77 |
//! Carry out a vertical filter in the first pass, then a horizontal filter in the second pass. |
| 78 |
#![allow(unreachable_code)] |
| 79 |
use zune_core::options::DecoderOptions; |
| 80 |
|
| 81 |
use crate::components::UpSampler; |
| 82 |
|
| 83 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 84 |
#[cfg(feature = "x86")] |
| 85 |
mod avx2; |
| 86 |
#[cfg(target_arch = "aarch64")] |
| 87 |
#[cfg(feature = "neon")] |
| 88 |
mod neon; |
| 89 |
#[cfg(feature = "portable_simd")] |
| 90 |
mod portable_simd; |
| 91 |
mod scalar; |
| 92 |
|
| 93 |
// choose the best possible implementation for this platform |
| 94 |
#[allow(unused_variables)] |
| 95 |
pub fn choose_horizontal_samp_function(options: &DecoderOptions) -> UpSampler { |
| 96 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 97 |
#[cfg(feature = "x86")] |
| 98 |
if options.use_avx2() { |
| 99 |
return |a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 100 |
// SAFETY: `options.use_avx2()` only returns true if avx2 is supported. |
| 101 |
unsafe { avx2::upsample_horizontal_avx2(a, b, c, d, e) } |
| 102 |
}; |
| 103 |
} |
| 104 |
#[cfg(target_arch = "aarch64")] |
| 105 |
#[cfg(feature = "neon")] |
| 106 |
if options.use_neon() { |
| 107 |
return |a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 108 |
// SAFETY: `options.use_neon()` only returns true if neon is supported. |
| 109 |
unsafe { neon::upsample_horizontal_neon(a, b, c, d, e) } |
| 110 |
}; |
| 111 |
} |
| 112 |
#[cfg(feature = "portable_simd")] |
| 113 |
return portable_simd::upsample_horizontal_simd; |
| 114 |
return scalar::upsample_horizontal; |
| 115 |
} |
| 116 |
|
| 117 |
#[allow(unused_variables)] |
| 118 |
pub fn choose_hv_samp_function(options: &DecoderOptions) -> UpSampler { |
| 119 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 120 |
#[cfg(feature = "x86")] |
| 121 |
if options.use_avx2() { |
| 122 |
return |a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 123 |
// SAFETY: `options.use_avx2()` only returns true if avx2 is supported. |
| 124 |
unsafe { avx2::upsample_hv_avx2(a, b, c, d, e) } |
| 125 |
}; |
| 126 |
} |
| 127 |
#[cfg(target_arch = "aarch64")] |
| 128 |
#[cfg(feature = "neon")] |
| 129 |
if options.use_neon() { |
| 130 |
return |a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 131 |
// SAFETY: `options.use_neon()` only returns true if neon is supported. |
| 132 |
unsafe { neon::upsample_hv_neon(a, b, c, d, e) } |
| 133 |
}; |
| 134 |
} |
| 135 |
#[cfg(feature = "portable_simd")] |
| 136 |
return portable_simd::upsample_hv_simd; |
| 137 |
return scalar::upsample_hv; |
| 138 |
} |
| 139 |
|
| 140 |
#[allow(unused_variables)] |
| 141 |
pub fn choose_v_samp_function(options: &DecoderOptions) -> UpSampler { |
| 142 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 143 |
#[cfg(feature = "x86")] |
| 144 |
if options.use_avx2() { |
| 145 |
return |a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 146 |
// SAFETY: `options.use_avx2()` only returns true if avx2 is supported. |
| 147 |
unsafe { avx2::upsample_vertical_avx2(a, b, c, d, e) } |
| 148 |
}; |
| 149 |
} |
| 150 |
#[cfg(target_arch = "aarch64")] |
| 151 |
#[cfg(feature = "neon")] |
| 152 |
if options.use_neon() { |
| 153 |
return |a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 154 |
// SAFETY: `options.use_neon()` only returns true if neon is supported. |
| 155 |
unsafe { neon::upsample_vertical_neon(a, b, c, d, e) } |
| 156 |
}; |
| 157 |
} |
| 158 |
#[cfg(feature = "portable_simd")] |
| 159 |
return portable_simd::upsample_vertical_simd; |
| 160 |
return scalar::upsample_vertical; |
| 161 |
} |
| 162 |
|
| 163 |
/// Upsample nothing |
| 164 |
|
| 165 |
pub fn upsample_no_op( |
| 166 |
_input: &[i16], |
| 167 |
_in_ref: &[i16], |
| 168 |
_in_near: &[i16], |
| 169 |
_scratch_space: &mut [i16], |
| 170 |
_output: &mut [i16], |
| 171 |
) { |
| 172 |
} |
| 173 |
|
| 174 |
pub fn generic_sampler() -> UpSampler { |
| 175 |
scalar::upsample_generic |
| 176 |
} |
| 177 |
|
| 178 |
#[cfg(test)] |
| 179 |
mod tests { |
| 180 |
use super::*; |
| 181 |
|
| 182 |
#[cfg(feature = "portable_simd")] |
| 183 |
mod portable_simd_impl { |
| 184 |
use super::*; |
| 185 |
|
| 186 |
#[test] |
| 187 |
fn portable_simd_vertical() { |
| 188 |
_test_vertical(portable_simd::upsample_vertical_simd) |
| 189 |
} |
| 190 |
|
| 191 |
#[test] |
| 192 |
fn portable_simd_horizontal() { |
| 193 |
_test_horizontal(portable_simd::upsample_horizontal_simd) |
| 194 |
} |
| 195 |
|
| 196 |
#[test] |
| 197 |
fn portable_simd_hv() { |
| 198 |
_test_hv(portable_simd::upsample_hv_simd) |
| 199 |
} |
| 200 |
} |
| 201 |
|
| 202 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 203 |
#[cfg(feature = "x86")] |
| 204 |
#[cfg(target_feature = "avx2")] |
| 205 |
mod avx2_impl { |
| 206 |
use super::*; |
| 207 |
|
| 208 |
#[test] |
| 209 |
fn avx2_vertical() { |
| 210 |
_test_vertical(|a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 211 |
// SAFETY: Test guarded behind `target_feature` |
| 212 |
unsafe { avx2::upsample_vertical_avx2(a, b, c, d, e) } |
| 213 |
}) |
| 214 |
} |
| 215 |
|
| 216 |
#[test] |
| 217 |
fn avx2_horizontal() { |
| 218 |
_test_horizontal(|a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 219 |
// SAFETY: Test guarded behind `target_feature` |
| 220 |
unsafe { avx2::upsample_horizontal_avx2(a, b, c, d, e) } |
| 221 |
}) |
| 222 |
} |
| 223 |
|
| 224 |
#[test] |
| 225 |
fn avx2_hv() { |
| 226 |
_test_hv(|a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 227 |
// SAFETY: Test guarded behind `target_feature` |
| 228 |
unsafe { avx2::upsample_hv_avx2(a, b, c, d, e) } |
| 229 |
}) |
| 230 |
} |
| 231 |
} |
| 232 |
|
| 233 |
#[cfg(target_arch = "aarch64")] |
| 234 |
#[cfg(feature = "neon")] |
| 235 |
#[cfg(target_feature = "neon")] |
| 236 |
mod neon_impl { |
| 237 |
use super::*; |
| 238 |
|
| 239 |
#[test] |
| 240 |
fn neon_vertical() { |
| 241 |
_test_vertical(|a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 242 |
// SAFETY: Test guarded behind `target_feature` |
| 243 |
unsafe { neon::upsample_vertical_neon(a, b, c, d, e) } |
| 244 |
}) |
| 245 |
} |
| 246 |
|
| 247 |
#[test] |
| 248 |
fn neon_horizontal() { |
| 249 |
_test_horizontal(|a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 250 |
// SAFETY: Test guarded behind `target_feature` |
| 251 |
unsafe { neon::upsample_horizontal_neon(a, b, c, d, e) } |
| 252 |
}) |
| 253 |
} |
| 254 |
|
| 255 |
#[test] |
| 256 |
fn neon_hv() { |
| 257 |
_test_hv(|a: &[i16], b: &[i16], c: &[i16], d: &mut [i16], e: &mut [i16]| { |
| 258 |
// SAFETY: Test guarded behind `target_feature` |
| 259 |
unsafe { neon::upsample_hv_neon(a, b, c, d, e) } |
| 260 |
}) |
| 261 |
} |
| 262 |
} |
| 263 |
|
| 264 |
fn _test_vertical(upsampler: UpSampler) { |
| 265 |
let width = 1024; |
| 266 |
let input: Vec<i16> = (0..width).map(|x| ((x + 10) % 256) as i16).collect(); |
| 267 |
let in_near: Vec<i16> = (0..width).map(|x| ((x + 20) % 256) as i16).collect(); |
| 268 |
let in_far: Vec<i16> = (0..width).map(|x| ((x + 30) % 256) as i16).collect(); |
| 269 |
let mut scratch = vec![0i16; width]; |
| 270 |
|
| 271 |
let mut output_scalar = vec![0i16; width * 2]; |
| 272 |
let mut output_fast = vec![0i16; width * 2]; |
| 273 |
|
| 274 |
scalar::upsample_vertical(&input, &in_near, &in_far, &mut scratch, &mut output_scalar); |
| 275 |
upsampler(&input, &in_near, &in_far, &mut scratch, &mut output_fast); |
| 276 |
|
| 277 |
assert_eq!(output_scalar, output_fast); |
| 278 |
} |
| 279 |
|
| 280 |
fn _test_horizontal(upsampler: UpSampler) { |
| 281 |
_test_horizontal_even_width(upsampler); |
| 282 |
_test_horizontal_odd_width(upsampler); |
| 283 |
} |
| 284 |
|
| 285 |
fn _test_horizontal_even_width(upsampler: UpSampler) { |
| 286 |
let width = 1024; |
| 287 |
let input: Vec<i16> = (0..width).map(|x| ((x + 10) % 256) as i16).collect(); |
| 288 |
|
| 289 |
let mut scratch = vec![0i16; width]; |
| 290 |
|
| 291 |
let mut output_scalar = vec![0i16; width * 2]; |
| 292 |
let mut output_fast = vec![0i16; width * 2]; |
| 293 |
|
| 294 |
scalar::upsample_horizontal(&input, &[], &[], &mut scratch, &mut output_scalar); |
| 295 |
upsampler(&input, &[], &[], &mut scratch, &mut output_fast); |
| 296 |
|
| 297 |
assert_eq!(output_scalar, output_fast); |
| 298 |
} |
| 299 |
|
| 300 |
fn _test_horizontal_odd_width(upsampler: UpSampler) { |
| 301 |
let width = 33; |
| 302 |
let input: Vec<i16> = (0..width).map(|x| ((x + 10) % 256) as i16).collect(); |
| 303 |
let mut scratch = vec![0i16; width]; |
| 304 |
let mut output_scalar = vec![0i16; width * 2]; |
| 305 |
let mut output_fast = vec![0i16; width * 2]; |
| 306 |
|
| 307 |
scalar::upsample_horizontal(&input, &[], &[], &mut scratch, &mut output_scalar); |
| 308 |
upsampler(&input, &[], &[], &mut scratch, &mut output_fast); |
| 309 |
|
| 310 |
assert_eq!(output_scalar, output_fast); |
| 311 |
} |
| 312 |
|
| 313 |
fn _test_hv(upsampler: UpSampler) { |
| 314 |
let width = 512; |
| 315 |
let input: Vec<i16> = (0..width).map(|x| ((x + 10) % 256) as i16).collect(); |
| 316 |
let in_near: Vec<i16> = (0..width).map(|x| ((x + 20) % 256) as i16).collect(); |
| 317 |
let in_far: Vec<i16> = (0..width).map(|x| ((x + 30) % 256) as i16).collect(); |
| 318 |
|
| 319 |
// Output len is width * 4 for HV (vertical * 2, then horizontal * 2 for each row) |
| 320 |
// scratch is width * 2 |
| 321 |
let mut scratch_scalar = vec![0i16; width * 2]; |
| 322 |
let mut scratch_fast = vec![0i16; width * 2]; |
| 323 |
let mut output_scalar = vec![0i16; width * 4]; |
| 324 |
let mut output_fast = vec![0i16; width * 4]; |
| 325 |
|
| 326 |
scalar::upsample_hv( |
| 327 |
&input, |
| 328 |
&in_near, |
| 329 |
&in_far, |
| 330 |
&mut scratch_scalar, |
| 331 |
&mut output_scalar, |
| 332 |
); |
| 333 |
upsampler( |
| 334 |
&input, |
| 335 |
&in_near, |
| 336 |
&in_far, |
| 337 |
&mut scratch_fast, |
| 338 |
&mut output_fast, |
| 339 |
); |
| 340 |
|
| 341 |
assert_eq!(output_scalar, output_fast); |
| 342 |
} |
| 343 |
} |