| 1 |
//! Utilities for doing the YUV -> RGB conversion |
| 2 |
//! The images are encoded in the Y'CbCr format as detailed here: <https://en.wikipedia.org/wiki/YCbCr> |
| 3 |
//! so need to be converted to RGB to be displayed |
| 4 |
//! To do the YUV -> RGB conversion we need to first decide how to map the yuv values to the pixels |
| 5 |
//! The y buffer is the same size as the pixel buffer so that maps 1-1 but the |
| 6 |
//! u and v buffers are half the size of the pixel buffer so we need to scale it up |
| 7 |
//! The simple way to upscale is just to take each u/v value and associate it with the 4 |
| 8 |
//! pixels around it e.g. for a 4x4 image: |
| 9 |
//! |
| 10 |
//! |||||| |
| 11 |
//! |yyyy| |
| 12 |
//! |yyyy| |
| 13 |
//! |yyyy| |
| 14 |
//! |yyyy| |
| 15 |
//! |||||| |
| 16 |
//! |
| 17 |
//! ||||||| |
| 18 |
//! |uu|vv| |
| 19 |
//! |uu|vv| |
| 20 |
//! ||||||| |
| 21 |
//! |
| 22 |
//! Then each of the 2x2 pixels would match the u/v from the same quadrant |
| 23 |
//! |
| 24 |
//! However fancy upsampling is the default for libwebp which does a little more work to make the values smoother |
| 25 |
//! It interpolates u and v so that for e.g. the pixel 1 down and 1 from the left the u value |
| 26 |
//! would be (9*u0 + 3*u1 + 3*u2 + u3 + 8) / 16 and similar for the other pixels |
| 27 |
//! The edges are mirrored, so for the pixel 1 down and 0 from the left it uses (9*u0 + 3*u2 + 3*u0 + u2 + 8) / 16 |
| 28 |
|
| 29 |
/// `_mm_mulhi_epu16` emulation |
| 30 |
fn mulhi(v: u8, coeff: u16) -> i32 { |
| 31 |
((u32::from(v) * u32::from(coeff)) >> 8) as i32 |
| 32 |
} |
| 33 |
|
| 34 |
/// This function has been rewritten to encourage auto-vectorization. |
| 35 |
/// |
| 36 |
/// Based on [src/dsp/yuv.h](https://github.com/webmproject/libwebp/blob/8534f53960befac04c9631e6e50d21dcb42dfeaf/src/dsp/yuv.h#L79) |
| 37 |
/// from the libwebp source. |
| 38 |
/// ```text |
| 39 |
/// const YUV_FIX2: i32 = 6; |
| 40 |
/// const YUV_MASK2: i32 = (256 << YUV_FIX2) - 1; |
| 41 |
/// fn clip(v: i32) -> u8 { |
| 42 |
/// if (v & !YUV_MASK2) == 0 { |
| 43 |
/// (v >> YUV_FIX2) as u8 |
| 44 |
/// } else if v < 0 { |
| 45 |
/// 0 |
| 46 |
/// } else { |
| 47 |
/// 255 |
| 48 |
/// } |
| 49 |
/// } |
| 50 |
/// ``` |
| 51 |
// Clippy suggests the clamp method, but it seems to optimize worse as of rustc 1.82.0 nightly. |
| 52 |
#[allow(clippy::manual_clamp)] |
| 53 |
fn clip(v: i32) -> u8 { |
| 54 |
const YUV_FIX2: i32 = 6; |
| 55 |
(v >> YUV_FIX2).max(0).min(255) as u8 |
| 56 |
} |
| 57 |
|
| 58 |
#[inline(always)] |
| 59 |
fn yuv_to_r(y: u8, v: u8) -> u8 { |
| 60 |
clip(mulhi(y, 19077) + mulhi(v, 26149) - 14234) |
| 61 |
} |
| 62 |
|
| 63 |
#[inline(always)] |
| 64 |
fn yuv_to_g(y: u8, u: u8, v: u8) -> u8 { |
| 65 |
clip(mulhi(y, 19077) - mulhi(u, 6419) - mulhi(v, 13320) + 8708) |
| 66 |
} |
| 67 |
|
| 68 |
#[inline(always)] |
| 69 |
fn yuv_to_b(y: u8, u: u8) -> u8 { |
| 70 |
clip(mulhi(y, 19077) + mulhi(u, 33050) - 17685) |
| 71 |
} |
| 72 |
|
| 73 |
/// Fills an rgb buffer with the image from the yuv buffers |
| 74 |
/// Size of the buffer is assumed to be correct |
| 75 |
/// BPP is short for bytes per pixel, allows both rgb and rgba to be decoded |
| 76 |
pub(crate) fn fill_rgb_buffer_fancy<const BPP: usize>( |
| 77 |
buffer: &mut [u8], |
| 78 |
y_buffer: &[u8], |
| 79 |
u_buffer: &[u8], |
| 80 |
v_buffer: &[u8], |
| 81 |
width: usize, |
| 82 |
height: usize, |
| 83 |
buffer_width: usize, |
| 84 |
) { |
| 85 |
// buffer width is always even so don't need to do div_ceil |
| 86 |
let chroma_buffer_width = buffer_width / 2; |
| 87 |
let chroma_width = width.div_ceil(2); |
| 88 |
|
| 89 |
// fill top row first since it only uses the top u/v row |
| 90 |
let top_row_y = &y_buffer[..width]; |
| 91 |
let top_row_u = &u_buffer[..chroma_width]; |
| 92 |
let top_row_v = &v_buffer[..chroma_width]; |
| 93 |
let top_row_buffer = &mut buffer[..width * BPP]; |
| 94 |
fill_row_fancy_with_1_uv_row::<BPP>(top_row_buffer, top_row_y, top_row_u, top_row_v); |
| 95 |
|
| 96 |
let mut main_row_chunks = buffer[width * BPP..].chunks_exact_mut(width * BPP * 2); |
| 97 |
// the y buffer iterator limits the end of the row iterator so we need this end index |
| 98 |
let end_y_index = height * buffer_width; |
| 99 |
let mut main_y_chunks = y_buffer[buffer_width..end_y_index].chunks_exact(buffer_width * 2); |
| 100 |
let mut main_u_windows = u_buffer |
| 101 |
.windows(chroma_buffer_width * 2) |
| 102 |
.step_by(chroma_buffer_width); |
| 103 |
let mut main_v_windows = v_buffer |
| 104 |
.windows(chroma_buffer_width * 2) |
| 105 |
.step_by(chroma_buffer_width); |
| 106 |
|
| 107 |
for (((row_buffer, y_rows), u_rows), v_rows) in (&mut main_row_chunks) |
| 108 |
.zip(&mut main_y_chunks) |
| 109 |
.zip(&mut main_u_windows) |
| 110 |
.zip(&mut main_v_windows) |
| 111 |
{ |
| 112 |
let (u_row_1, u_row_2) = u_rows.split_at(chroma_buffer_width); |
| 113 |
let (v_row_1, v_row_2) = v_rows.split_at(chroma_buffer_width); |
| 114 |
let (row_buf_1, row_buf_2) = row_buffer.split_at_mut(width * BPP); |
| 115 |
let (y_row_1, y_row_2) = y_rows.split_at(buffer_width); |
| 116 |
fill_row_fancy_with_2_uv_rows::<BPP>( |
| 117 |
row_buf_1, |
| 118 |
&y_row_1[..width], |
| 119 |
&u_row_1[..chroma_width], |
| 120 |
&u_row_2[..chroma_width], |
| 121 |
&v_row_1[..chroma_width], |
| 122 |
&v_row_2[..chroma_width], |
| 123 |
); |
| 124 |
fill_row_fancy_with_2_uv_rows::<BPP>( |
| 125 |
row_buf_2, |
| 126 |
&y_row_2[..width], |
| 127 |
&u_row_2[..chroma_width], |
| 128 |
&u_row_1[..chroma_width], |
| 129 |
&v_row_2[..chroma_width], |
| 130 |
&v_row_1[..chroma_width], |
| 131 |
); |
| 132 |
} |
| 133 |
|
| 134 |
let final_row_buffer = main_row_chunks.into_remainder(); |
| 135 |
|
| 136 |
// if the image has even height there will be one final row with only one u/v row matching it |
| 137 |
if !final_row_buffer.is_empty() { |
| 138 |
let final_y_row = main_y_chunks.remainder(); |
| 139 |
|
| 140 |
let chroma_height = height.div_ceil(2); |
| 141 |
let start_chroma_index = (chroma_height - 1) * chroma_buffer_width; |
| 142 |
|
| 143 |
let final_u_row = &u_buffer[start_chroma_index..]; |
| 144 |
let final_v_row = &v_buffer[start_chroma_index..]; |
| 145 |
fill_row_fancy_with_1_uv_row::<BPP>( |
| 146 |
final_row_buffer, |
| 147 |
&final_y_row[..width], |
| 148 |
&final_u_row[..chroma_width], |
| 149 |
&final_v_row[..chroma_width], |
| 150 |
); |
| 151 |
} |
| 152 |
} |
| 153 |
|
| 154 |
/// Fills a row with the fancy interpolation as detailed |
| 155 |
fn fill_row_fancy_with_2_uv_rows<const BPP: usize>( |
| 156 |
row_buffer: &mut [u8], |
| 157 |
y_row: &[u8], |
| 158 |
u_row_1: &[u8], |
| 159 |
u_row_2: &[u8], |
| 160 |
v_row_1: &[u8], |
| 161 |
v_row_2: &[u8], |
| 162 |
) { |
| 163 |
// need to do left pixel separately since it will only have one u/v value |
| 164 |
{ |
| 165 |
let rgb1 = &mut row_buffer[0..3]; |
| 166 |
let y_value = y_row[0]; |
| 167 |
// first pixel uses the first u/v as the main one |
| 168 |
let u_value = get_fancy_chroma_value(u_row_1[0], u_row_1[0], u_row_2[0], u_row_2[0]); |
| 169 |
let v_value = get_fancy_chroma_value(v_row_1[0], v_row_1[0], v_row_2[0], v_row_2[0]); |
| 170 |
set_pixel(rgb1, y_value, u_value, v_value); |
| 171 |
} |
| 172 |
|
| 173 |
let rest_row_buffer = &mut row_buffer[BPP..]; |
| 174 |
let rest_y_row = &y_row[1..]; |
| 175 |
|
| 176 |
// we do two pixels at a time since they share the same u/v values |
| 177 |
let mut main_row_chunks = rest_row_buffer.chunks_exact_mut(BPP * 2); |
| 178 |
let mut main_y_chunks = rest_y_row.chunks_exact(2); |
| 179 |
|
| 180 |
for (((((rgb, y_val), u_val_1), u_val_2), v_val_1), v_val_2) in (&mut main_row_chunks) |
| 181 |
.zip(&mut main_y_chunks) |
| 182 |
.zip(u_row_1.windows(2)) |
| 183 |
.zip(u_row_2.windows(2)) |
| 184 |
.zip(v_row_1.windows(2)) |
| 185 |
.zip(v_row_2.windows(2)) |
| 186 |
{ |
| 187 |
{ |
| 188 |
let rgb1 = &mut rgb[0..3]; |
| 189 |
let y_value = y_val[0]; |
| 190 |
// first pixel uses the first u/v as the main one |
| 191 |
let u_value = get_fancy_chroma_value(u_val_1[0], u_val_1[1], u_val_2[0], u_val_2[1]); |
| 192 |
let v_value = get_fancy_chroma_value(v_val_1[0], v_val_1[1], v_val_2[0], v_val_2[1]); |
| 193 |
set_pixel(rgb1, y_value, u_value, v_value); |
| 194 |
} |
| 195 |
{ |
| 196 |
let rgb2 = &mut rgb[BPP..]; |
| 197 |
let y_value = y_val[1]; |
| 198 |
let u_value = get_fancy_chroma_value(u_val_1[1], u_val_1[0], u_val_2[1], u_val_2[0]); |
| 199 |
let v_value = get_fancy_chroma_value(v_val_1[1], v_val_1[0], v_val_2[1], v_val_2[0]); |
| 200 |
set_pixel(rgb2, y_value, u_value, v_value); |
| 201 |
} |
| 202 |
} |
| 203 |
|
| 204 |
let final_pixel = main_row_chunks.into_remainder(); |
| 205 |
let final_y = main_y_chunks.remainder(); |
| 206 |
|
| 207 |
if let (rgb, [y_value]) = (final_pixel, final_y) { |
| 208 |
let final_u_1 = *u_row_1.last().unwrap(); |
| 209 |
let final_u_2 = *u_row_2.last().unwrap(); |
| 210 |
|
| 211 |
let final_v_1 = *v_row_1.last().unwrap(); |
| 212 |
let final_v_2 = *v_row_2.last().unwrap(); |
| 213 |
|
| 214 |
let rgb1 = &mut rgb[0..3]; |
| 215 |
// first pixel uses the first u/v as the main one |
| 216 |
let u_value = get_fancy_chroma_value(final_u_1, final_u_1, final_u_2, final_u_2); |
| 217 |
let v_value = get_fancy_chroma_value(final_v_1, final_v_1, final_v_2, final_v_2); |
| 218 |
set_pixel(rgb1, *y_value, u_value, v_value); |
| 219 |
} |
| 220 |
} |
| 221 |
|
| 222 |
fn fill_row_fancy_with_1_uv_row<const BPP: usize>( |
| 223 |
row_buffer: &mut [u8], |
| 224 |
y_row: &[u8], |
| 225 |
u_row: &[u8], |
| 226 |
v_row: &[u8], |
| 227 |
) { |
| 228 |
// doing left pixel first |
| 229 |
{ |
| 230 |
let rgb1 = &mut row_buffer[0..3]; |
| 231 |
let y_value = y_row[0]; |
| 232 |
|
| 233 |
let u_value = u_row[0]; |
| 234 |
let v_value = v_row[0]; |
| 235 |
set_pixel(rgb1, y_value, u_value, v_value); |
| 236 |
} |
| 237 |
|
| 238 |
// two pixels at a time since they share the same u/v value |
| 239 |
let mut main_row_chunks = row_buffer[BPP..].chunks_exact_mut(BPP * 2); |
| 240 |
let mut main_y_row_chunks = y_row[1..].chunks_exact(2); |
| 241 |
|
| 242 |
for (((rgb, y_val), u_val), v_val) in (&mut main_row_chunks) |
| 243 |
.zip(&mut main_y_row_chunks) |
| 244 |
.zip(u_row.windows(2)) |
| 245 |
.zip(v_row.windows(2)) |
| 246 |
{ |
| 247 |
{ |
| 248 |
let rgb1 = &mut rgb[0..3]; |
| 249 |
let y_value = y_val[0]; |
| 250 |
// first pixel uses the first u/v as the main one |
| 251 |
let u_value = get_fancy_chroma_value(u_val[0], u_val[1], u_val[0], u_val[1]); |
| 252 |
let v_value = get_fancy_chroma_value(v_val[0], v_val[1], v_val[0], v_val[1]); |
| 253 |
set_pixel(rgb1, y_value, u_value, v_value); |
| 254 |
} |
| 255 |
{ |
| 256 |
let rgb2 = &mut rgb[BPP..]; |
| 257 |
let y_value = y_val[1]; |
| 258 |
let u_value = get_fancy_chroma_value(u_val[1], u_val[0], u_val[1], u_val[0]); |
| 259 |
let v_value = get_fancy_chroma_value(v_val[1], v_val[0], v_val[1], v_val[0]); |
| 260 |
set_pixel(rgb2, y_value, u_value, v_value); |
| 261 |
} |
| 262 |
} |
| 263 |
|
| 264 |
let final_pixel = main_row_chunks.into_remainder(); |
| 265 |
let final_y = main_y_row_chunks.remainder(); |
| 266 |
|
| 267 |
if let (rgb, [final_y]) = (final_pixel, final_y) { |
| 268 |
let final_u = *u_row.last().unwrap(); |
| 269 |
let final_v = *v_row.last().unwrap(); |
| 270 |
|
| 271 |
set_pixel(rgb, *final_y, final_u, final_v); |
| 272 |
} |
| 273 |
} |
| 274 |
|
| 275 |
#[inline] |
| 276 |
fn get_fancy_chroma_value(main: u8, secondary1: u8, secondary2: u8, tertiary: u8) -> u8 { |
| 277 |
let val0 = u16::from(main); |
| 278 |
let val1 = u16::from(secondary1); |
| 279 |
let val2 = u16::from(secondary2); |
| 280 |
let val3 = u16::from(tertiary); |
| 281 |
((9 * val0 + 3 * val1 + 3 * val2 + val3 + 8) / 16) as u8 |
| 282 |
} |
| 283 |
|
| 284 |
#[inline] |
| 285 |
fn set_pixel(rgb: &mut [u8], y: u8, u: u8, v: u8) { |
| 286 |
rgb[0] = yuv_to_r(y, v); |
| 287 |
rgb[1] = yuv_to_g(y, u, v); |
| 288 |
rgb[2] = yuv_to_b(y, u); |
| 289 |
} |
| 290 |
|
| 291 |
/// Simple conversion, not currently used but could add a config to allow for using the simple |
| 292 |
#[allow(unused)] |
| 293 |
pub(crate) fn fill_rgb_buffer_simple<const BPP: usize>( |
| 294 |
buffer: &mut [u8], |
| 295 |
y_buffer: &[u8], |
| 296 |
u_buffer: &[u8], |
| 297 |
v_buffer: &[u8], |
| 298 |
width: usize, |
| 299 |
chroma_width: usize, |
| 300 |
buffer_width: usize, |
| 301 |
) { |
| 302 |
let u_row_twice_iter = u_buffer |
| 303 |
.chunks_exact(buffer_width / 2) |
| 304 |
.flat_map(|n| std::iter::repeat(n).take(2)); |
| 305 |
let v_row_twice_iter = v_buffer |
| 306 |
.chunks_exact(buffer_width / 2) |
| 307 |
.flat_map(|n| std::iter::repeat(n).take(2)); |
| 308 |
|
| 309 |
for (((row, y_row), u_row), v_row) in buffer |
| 310 |
.chunks_exact_mut(width * BPP) |
| 311 |
.zip(y_buffer.chunks_exact(buffer_width)) |
| 312 |
.zip(u_row_twice_iter) |
| 313 |
.zip(v_row_twice_iter) |
| 314 |
{ |
| 315 |
fill_rgba_row_simple::<BPP>( |
| 316 |
&y_row[..width], |
| 317 |
&u_row[..chroma_width], |
| 318 |
&v_row[..chroma_width], |
| 319 |
row, |
| 320 |
); |
| 321 |
} |
| 322 |
} |
| 323 |
|
| 324 |
fn fill_rgba_row_simple<const BPP: usize>( |
| 325 |
y_vec: &[u8], |
| 326 |
u_vec: &[u8], |
| 327 |
v_vec: &[u8], |
| 328 |
rgba: &mut [u8], |
| 329 |
) { |
| 330 |
// Fill 2 pixels per iteration: these pixels share `u` and `v` components |
| 331 |
let mut rgb_chunks = rgba.chunks_exact_mut(BPP * 2); |
| 332 |
let mut y_chunks = y_vec.chunks_exact(2); |
| 333 |
let mut u_iter = u_vec.iter(); |
| 334 |
let mut v_iter = v_vec.iter(); |
| 335 |
|
| 336 |
for (((rgb, y), &u), &v) in (&mut rgb_chunks) |
| 337 |
.zip(&mut y_chunks) |
| 338 |
.zip(&mut u_iter) |
| 339 |
.zip(&mut v_iter) |
| 340 |
{ |
| 341 |
let coeffs = [ |
| 342 |
mulhi(v, 26149), |
| 343 |
mulhi(u, 6419), |
| 344 |
mulhi(v, 13320), |
| 345 |
mulhi(u, 33050), |
| 346 |
]; |
| 347 |
|
| 348 |
let get_r = |y: u8| clip(mulhi(y, 19077) + coeffs[0] - 14234); |
| 349 |
let get_g = |y: u8| clip(mulhi(y, 19077) - coeffs[1] - coeffs[2] + 8708); |
| 350 |
let get_b = |y: u8| clip(mulhi(y, 19077) + coeffs[3] - 17685); |
| 351 |
|
| 352 |
let rgb1 = &mut rgb[0..3]; |
| 353 |
rgb1[0] = get_r(y[0]); |
| 354 |
rgb1[1] = get_g(y[0]); |
| 355 |
rgb1[2] = get_b(y[0]); |
| 356 |
|
| 357 |
let rgb2 = &mut rgb[BPP..]; |
| 358 |
rgb2[0] = get_r(y[1]); |
| 359 |
rgb2[1] = get_g(y[1]); |
| 360 |
rgb2[2] = get_b(y[1]); |
| 361 |
} |
| 362 |
|
| 363 |
let remainder = rgb_chunks.into_remainder(); |
| 364 |
if remainder.len() >= 3 { |
| 365 |
if let (Some(&y), Some(&u), Some(&v)) = ( |
| 366 |
y_chunks.remainder().iter().next(), |
| 367 |
u_iter.next(), |
| 368 |
v_iter.next(), |
| 369 |
) { |
| 370 |
let coeffs = [ |
| 371 |
mulhi(v, 26149), |
| 372 |
mulhi(u, 6419), |
| 373 |
mulhi(v, 13320), |
| 374 |
mulhi(u, 33050), |
| 375 |
]; |
| 376 |
|
| 377 |
remainder[0] = clip(mulhi(y, 19077) + coeffs[0] - 14234); |
| 378 |
remainder[1] = clip(mulhi(y, 19077) - coeffs[1] - coeffs[2] + 8708); |
| 379 |
remainder[2] = clip(mulhi(y, 19077) + coeffs[3] - 17685); |
| 380 |
} |
| 381 |
} |
| 382 |
} |
| 383 |
|
| 384 |
#[cfg(test)] |
| 385 |
mod tests { |
| 386 |
use super::*; |
| 387 |
|
| 388 |
#[test] |
| 389 |
fn test_fancy_grid() { |
| 390 |
#[rustfmt::skip] |
| 391 |
let y_buffer = [ |
| 392 |
77, 162, 202, 185, |
| 393 |
28, 13, 199, 182, |
| 394 |
135, 147, 164, 135, |
| 395 |
66, 27, 171, 130, |
| 396 |
]; |
| 397 |
|
| 398 |
#[rustfmt::skip] |
| 399 |
let u_buffer = [ |
| 400 |
34, 101, |
| 401 |
123, 163 |
| 402 |
]; |
| 403 |
|
| 404 |
#[rustfmt::skip] |
| 405 |
let v_buffer = [ |
| 406 |
97, 167, |
| 407 |
149, 23, |
| 408 |
]; |
| 409 |
|
| 410 |
let mut rgb_buffer = [0u8; 16 * 3]; |
| 411 |
fill_rgb_buffer_fancy::<3>(&mut rgb_buffer, &y_buffer, &u_buffer, &v_buffer, 4, 4, 4); |
| 412 |
|
| 413 |
#[rustfmt::skip] |
| 414 |
let upsampled_u_buffer = [ |
| 415 |
34, 51, 84, 101, |
| 416 |
56, 71, 101, 117, |
| 417 |
101, 112, 136, 148, |
| 418 |
123, 133, 153, 163, |
| 419 |
]; |
| 420 |
|
| 421 |
#[rustfmt::skip] |
| 422 |
let upsampled_v_buffer = [ |
| 423 |
97, 115, 150, 167, |
| 424 |
110, 115, 126, 131, |
| 425 |
136, 117, 78, 59, |
| 426 |
149, 118, 55, 23, |
| 427 |
]; |
| 428 |
|
| 429 |
let mut upsampled_rgb_buffer = [0u8; 16 * 3]; |
| 430 |
for (((rgb_val, y), u), v) in upsampled_rgb_buffer |
| 431 |
.chunks_exact_mut(3) |
| 432 |
.zip(y_buffer) |
| 433 |
.zip(upsampled_u_buffer) |
| 434 |
.zip(upsampled_v_buffer) |
| 435 |
{ |
| 436 |
rgb_val[0] = yuv_to_r(y, v); |
| 437 |
rgb_val[1] = yuv_to_g(y, u, v); |
| 438 |
rgb_val[2] = yuv_to_b(y, u); |
| 439 |
} |
| 440 |
|
| 441 |
assert_eq!(rgb_buffer, upsampled_rgb_buffer); |
| 442 |
} |
| 443 |
|
| 444 |
#[test] |
| 445 |
fn test_yuv_conversions() { |
| 446 |
let (y, u, v) = (203, 40, 42); |
| 447 |
|
| 448 |
assert_eq!(yuv_to_r(y, v), 80); |
| 449 |
assert_eq!(yuv_to_g(y, u, v), 255); |
| 450 |
assert_eq!(yuv_to_b(y, u), 40); |
| 451 |
} |
| 452 |
} |