| 1 |
#![allow(deprecated)] |
| 2 |
use crate::dirtyalpha::blurred_dirty_alpha; |
| 3 |
use crate::error::Error; |
| 4 |
#[cfg(not(feature = "threading"))] |
| 5 |
use crate::rayoff as rayon; |
| 6 |
use imgref::{Img, ImgVec}; |
| 7 |
use rav1e::prelude::*; |
| 8 |
use rgb::{RGB8, RGBA8}; |
| 9 |
|
| 10 |
/// For [`Encoder::with_internal_color_model`] |
| 11 |
#[derive(Debug, Copy, Clone, Eq, PartialEq)] |
| 12 |
pub enum ColorModel { |
| 13 |
/// Standard color model for photographic content. Usually the best choice. |
| 14 |
/// This library always uses full-resolution color (4:4:4). |
| 15 |
/// This library will automatically choose between BT.601 or BT.709. |
| 16 |
YCbCr, |
| 17 |
/// RGB channels are encoded without color space transformation. |
| 18 |
/// Usually results in larger file sizes, and is less compatible than `YCbCr`. |
| 19 |
/// Use only if the content really makes use of RGB, e.g. anaglyph images or RGB subpixel anti-aliasing. |
| 20 |
RGB, |
| 21 |
} |
| 22 |
|
| 23 |
/// Handling of color channels in transparent images. For [`Encoder::with_alpha_color_mode`] |
| 24 |
#[derive(Debug, Copy, Clone, Eq, PartialEq)] |
| 25 |
pub enum AlphaColorMode { |
| 26 |
/// Use unassociated alpha channel and leave color channels unchanged, even if there's redundant color data in transparent areas. |
| 27 |
UnassociatedDirty, |
| 28 |
/// Use unassociated alpha channel, but set color channels of transparent areas to a solid color to eliminate invisible data and improve compression. |
| 29 |
UnassociatedClean, |
| 30 |
/// Store color channels of transparent images in premultiplied form. |
| 31 |
/// This requires support for premultiplied alpha in AVIF decoders. |
| 32 |
/// |
| 33 |
/// It may reduce file sizes due to clearing of fully-transparent pixels, but |
| 34 |
/// may also increase file sizes due to creation of new edges in the color channels. |
| 35 |
/// |
| 36 |
/// Note that this is only internal detail for the AVIF file. |
| 37 |
/// It does not change meaning of `RGBA` in this library — it's always unassociated. |
| 38 |
Premultiplied, |
| 39 |
} |
| 40 |
|
| 41 |
#[derive(Default, Debug, Copy, Clone, Eq, PartialEq)] |
| 42 |
pub enum BitDepth { |
| 43 |
Eight, |
| 44 |
Ten, |
| 45 |
/// Same as `Ten` |
| 46 |
#[default] |
| 47 |
Auto, |
| 48 |
} |
| 49 |
|
| 50 |
/// The newly-created image file + extra info FYI |
| 51 |
#[non_exhaustive] |
| 52 |
#[derive(Clone)] |
| 53 |
pub struct EncodedImage { |
| 54 |
/// AVIF (HEIF+AV1) encoded image data |
| 55 |
pub avif_file: Vec<u8>, |
| 56 |
/// FYI: number of bytes of AV1 payload used for the color |
| 57 |
pub color_byte_size: usize, |
| 58 |
/// FYI: number of bytes of AV1 payload used for the alpha channel |
| 59 |
pub alpha_byte_size: usize, |
| 60 |
} |
| 61 |
|
| 62 |
/// Encoder config builder |
| 63 |
#[derive(Debug, Clone)] |
| 64 |
pub struct Encoder { |
| 65 |
/// 0-255 scale |
| 66 |
quantizer: u8, |
| 67 |
/// 0-255 scale |
| 68 |
alpha_quantizer: u8, |
| 69 |
/// rav1e preset 1 (slow) 10 (fast but crappy) |
| 70 |
speed: u8, |
| 71 |
/// True if RGBA input has already been premultiplied. It inserts appropriate metadata. |
| 72 |
premultiplied_alpha: bool, |
| 73 |
/// Which pixel format to use in AVIF file. RGB tends to give larger files. |
| 74 |
color_model: ColorModel, |
| 75 |
/// How many threads should be used (0 = match core count), None - use global rayon thread pool |
| 76 |
threads: Option<usize>, |
| 77 |
/// [`AlphaColorMode`] |
| 78 |
alpha_color_mode: AlphaColorMode, |
| 79 |
/// 8 or 10 |
| 80 |
output_depth: BitDepth, |
| 81 |
} |
| 82 |
|
| 83 |
/// Builder methods |
| 84 |
impl Encoder { |
| 85 |
/// Start here |
| 86 |
#[must_use] |
| 87 |
pub fn new() -> Self { |
| 88 |
Self { |
| 89 |
quantizer: quality_to_quantizer(80.), |
| 90 |
alpha_quantizer: quality_to_quantizer(80.), |
| 91 |
speed: 5, |
| 92 |
output_depth: BitDepth::default(), |
| 93 |
premultiplied_alpha: false, |
| 94 |
color_model: ColorModel::YCbCr, |
| 95 |
threads: None, |
| 96 |
alpha_color_mode: AlphaColorMode::UnassociatedClean, |
| 97 |
} |
| 98 |
} |
| 99 |
|
| 100 |
/// Quality `1..=100`. Panics if out of range. |
| 101 |
#[inline(always)] |
| 102 |
#[track_caller] |
| 103 |
#[must_use] |
| 104 |
pub fn with_quality(mut self, quality: f32) -> Self { |
| 105 |
assert!(quality >= 1. && quality <= 100.); |
| 106 |
self.quantizer = quality_to_quantizer(quality); |
| 107 |
self |
| 108 |
} |
| 109 |
|
| 110 |
#[doc(hidden)] |
| 111 |
#[deprecated(note = "Renamed to with_bit_depth")] |
| 112 |
pub fn with_depth(self, depth: Option<u8>) -> Self { |
| 113 |
self.with_bit_depth(depth.map(|d| if d >= 10 { BitDepth::Ten } else { BitDepth::Eight }).unwrap_or(BitDepth::Auto)) |
| 114 |
} |
| 115 |
|
| 116 |
/// Internal precision to use in the encoded AV1 data, for both color and alpha. 10-bit depth works best, even for 8-bit inputs/outputs. |
| 117 |
/// |
| 118 |
/// Use 8-bit depth only as a workaround for decoders that need it. |
| 119 |
/// |
| 120 |
/// This setting does not affect pixel inputs for this library. |
| 121 |
#[inline(always)] |
| 122 |
#[must_use] |
| 123 |
pub fn with_bit_depth(mut self, depth: BitDepth) -> Self { |
| 124 |
self.output_depth = depth; |
| 125 |
self |
| 126 |
} |
| 127 |
|
| 128 |
/// Quality for the alpha channel only. `1..=100`. Panics if out of range. |
| 129 |
#[inline(always)] |
| 130 |
#[track_caller] |
| 131 |
#[must_use] |
| 132 |
pub fn with_alpha_quality(mut self, quality: f32) -> Self { |
| 133 |
assert!(quality >= 1. && quality <= 100.); |
| 134 |
self.alpha_quantizer = quality_to_quantizer(quality); |
| 135 |
self |
| 136 |
} |
| 137 |
|
| 138 |
/// * 1 = very very slow, but max compression. |
| 139 |
/// * 10 = quick, but larger file sizes and lower quality. |
| 140 |
/// |
| 141 |
/// Panics if outside `1..=10`. |
| 142 |
#[inline(always)] |
| 143 |
#[track_caller] |
| 144 |
#[must_use] |
| 145 |
pub fn with_speed(mut self, speed: u8) -> Self { |
| 146 |
assert!(speed >= 1 && speed <= 10); |
| 147 |
self.speed = speed; |
| 148 |
self |
| 149 |
} |
| 150 |
|
| 151 |
/// Changes how color channels are stored in the image. The default is YCbCr. |
| 152 |
/// |
| 153 |
/// Note that this is only internal detail for the AVIF file, and doesn't |
| 154 |
/// change color model of inputs to encode functions. |
| 155 |
#[inline(always)] |
| 156 |
#[must_use] |
| 157 |
pub fn with_internal_color_model(mut self, color_model: ColorModel) -> Self { |
| 158 |
self.color_model = color_model; |
| 159 |
self |
| 160 |
} |
| 161 |
|
| 162 |
#[doc(hidden)] |
| 163 |
#[deprecated = "Renamed to `with_internal_color_model()`"] |
| 164 |
pub fn with_internal_color_space(self, color_model: ColorModel) -> Self { |
| 165 |
self.with_internal_color_model(color_model) |
| 166 |
} |
| 167 |
|
| 168 |
/// Configures `rayon` thread pool size. |
| 169 |
/// The default `None` is to use all threads in the default `rayon` thread pool. |
| 170 |
#[inline(always)] |
| 171 |
#[track_caller] |
| 172 |
#[must_use] |
| 173 |
pub fn with_num_threads(mut self, num_threads: Option<usize>) -> Self { |
| 174 |
assert!(num_threads.map_or(true, |n| n > 0)); |
| 175 |
self.threads = num_threads; |
| 176 |
self |
| 177 |
} |
| 178 |
|
| 179 |
/// Configure handling of color channels in transparent images |
| 180 |
/// |
| 181 |
/// Note that this doesn't affect input format for this library, |
| 182 |
/// which for RGBA is always uncorrelated alpha. |
| 183 |
#[inline(always)] |
| 184 |
#[must_use] |
| 185 |
pub fn with_alpha_color_mode(mut self, mode: AlphaColorMode) -> Self { |
| 186 |
self.alpha_color_mode = mode; |
| 187 |
self.premultiplied_alpha = mode == AlphaColorMode::Premultiplied; |
| 188 |
self |
| 189 |
} |
| 190 |
} |
| 191 |
|
| 192 |
/// Once done with config, call one of the `encode_*` functions |
| 193 |
impl Encoder { |
| 194 |
/// Make a new AVIF image from RGBA pixels (non-premultiplied, alpha last) |
| 195 |
/// |
| 196 |
/// Make the `Img` for the `buffer` like this: |
| 197 |
/// |
| 198 |
/// ```rust,ignore |
| 199 |
/// Img::new(&pixels_rgba[..], width, height) |
| 200 |
/// ``` |
| 201 |
/// |
| 202 |
/// If you have pixels as `u8` slice, then use the `rgb` crate, and do: |
| 203 |
/// |
| 204 |
/// ```rust,ignore |
| 205 |
/// use rgb::ComponentSlice; |
| 206 |
/// let pixels_rgba = pixels_u8.as_rgba(); |
| 207 |
/// ``` |
| 208 |
/// |
| 209 |
/// If all pixels are opaque, the alpha channel will be left out automatically. |
| 210 |
/// |
| 211 |
/// This function takes 8-bit inputs, but will generate an AVIF file using 10-bit depth. |
| 212 |
/// |
| 213 |
/// returns AVIF file with info about sizes about AV1 payload. |
| 214 |
pub fn encode_rgba(&self, in_buffer: Img<&[rgb::RGBA<u8>]>) -> Result<EncodedImage, Error> { |
| 215 |
let new_alpha = self.convert_alpha_8bit(in_buffer); |
| 216 |
let buffer = new_alpha.as_ref().map(|b| b.as_ref()).unwrap_or(in_buffer); |
| 217 |
let use_alpha = buffer.pixels().any(|px| px.a != 255); |
| 218 |
if !use_alpha { |
| 219 |
return self.encode_rgb_internal_from_8bit(buffer.width(), buffer.height(), buffer.pixels().map(|px| px.rgb())); |
| 220 |
} |
| 221 |
|
| 222 |
let width = buffer.width(); |
| 223 |
let height = buffer.height(); |
| 224 |
let matrix_coefficients = match self.color_model { |
| 225 |
ColorModel::YCbCr => MatrixCoefficients::BT601, |
| 226 |
ColorModel::RGB => MatrixCoefficients::Identity, |
| 227 |
}; |
| 228 |
match self.output_depth { |
| 229 |
BitDepth::Eight => { |
| 230 |
let planes = buffer.pixels().map(|px| { |
| 231 |
let (y, u, v) = match self.color_model { |
| 232 |
ColorModel::YCbCr => rgb_to_8_bit_ycbcr(px.rgb(), BT601), |
| 233 |
ColorModel::RGB => rgb_to_8_bit_gbr(px.rgb()), |
| 234 |
}; |
| 235 |
[y, u, v] |
| 236 |
}); |
| 237 |
let alpha = buffer.pixels().map(|px| px.a); |
| 238 |
self.encode_raw_planes_8_bit(width, height, planes, Some(alpha), PixelRange::Full, matrix_coefficients) |
| 239 |
}, |
| 240 |
BitDepth::Ten | BitDepth::Auto => { |
| 241 |
let planes = buffer.pixels().map(|px| { |
| 242 |
let (y, u, v) = match self.color_model { |
| 243 |
ColorModel::YCbCr => rgb_to_10_bit_ycbcr(px.rgb(), BT601), |
| 244 |
ColorModel::RGB => rgb_to_10_bit_gbr(px.rgb()), |
| 245 |
}; |
| 246 |
[y, u, v] |
| 247 |
}); |
| 248 |
let alpha = buffer.pixels().map(|px| to_ten(px.a)); |
| 249 |
self.encode_raw_planes_10_bit(width, height, planes, Some(alpha), PixelRange::Full, matrix_coefficients) |
| 250 |
}, |
| 251 |
} |
| 252 |
} |
| 253 |
|
| 254 |
fn convert_alpha_8bit(&self, in_buffer: Img<&[RGBA8]>) -> Option<ImgVec<RGBA8>> { |
| 255 |
match self.alpha_color_mode { |
| 256 |
AlphaColorMode::UnassociatedDirty => None, |
| 257 |
AlphaColorMode::UnassociatedClean => blurred_dirty_alpha(in_buffer), |
| 258 |
AlphaColorMode::Premultiplied => { |
| 259 |
let prem = in_buffer.pixels() |
| 260 |
.map(|px| { |
| 261 |
if px.a == 0 || px.a == 255 { |
| 262 |
RGBA8::default() |
| 263 |
} else { |
| 264 |
RGBA8::new( |
| 265 |
(u16::from(px.r) * 255 / u16::from(px.a)) as u8, |
| 266 |
(u16::from(px.g) * 255 / u16::from(px.a)) as u8, |
| 267 |
(u16::from(px.b) * 255 / u16::from(px.a)) as u8, |
| 268 |
px.a, |
| 269 |
) |
| 270 |
} |
| 271 |
}) |
| 272 |
.collect(); |
| 273 |
Some(ImgVec::new(prem, in_buffer.width(), in_buffer.height())) |
| 274 |
}, |
| 275 |
} |
| 276 |
} |
| 277 |
|
| 278 |
/// Make a new AVIF image from RGB pixels |
| 279 |
/// |
| 280 |
/// Make the `Img` for the `buffer` like this: |
| 281 |
/// |
| 282 |
/// ```rust,ignore |
| 283 |
/// Img::new(&pixels_rgb[..], width, height) |
| 284 |
/// ``` |
| 285 |
/// |
| 286 |
/// If you have pixels as `u8` slice, then first do: |
| 287 |
/// |
| 288 |
/// ```rust,ignore |
| 289 |
/// use rgb::ComponentSlice; |
| 290 |
/// let pixels_rgb = pixels_u8.as_rgb(); |
| 291 |
/// ``` |
| 292 |
/// |
| 293 |
/// returns AVIF file, size of color metadata |
| 294 |
#[inline] |
| 295 |
pub fn encode_rgb(&self, buffer: Img<&[RGB8]>) -> Result<EncodedImage, Error> { |
| 296 |
self.encode_rgb_internal_from_8bit(buffer.width(), buffer.height(), buffer.pixels()) |
| 297 |
} |
| 298 |
|
| 299 |
fn encode_rgb_internal_from_8bit(&self, width: usize, height: usize, pixels: impl Iterator<Item = RGB8> + Send + Sync) -> Result<EncodedImage, Error> { |
| 300 |
let matrix_coefficients = match self.color_model { |
| 301 |
ColorModel::YCbCr => MatrixCoefficients::BT601, |
| 302 |
ColorModel::RGB => MatrixCoefficients::Identity, |
| 303 |
}; |
| 304 |
|
| 305 |
match self.output_depth { |
| 306 |
BitDepth::Eight => { |
| 307 |
let planes = pixels.map(|px| { |
| 308 |
let (y, u, v) = match self.color_model { |
| 309 |
ColorModel::YCbCr => rgb_to_8_bit_ycbcr(px, BT601), |
| 310 |
ColorModel::RGB => rgb_to_8_bit_gbr(px), |
| 311 |
}; |
| 312 |
[y, u, v] |
| 313 |
}); |
| 314 |
self.encode_raw_planes_8_bit(width, height, planes, None::<[_; 0]>, PixelRange::Full, matrix_coefficients) |
| 315 |
}, |
| 316 |
BitDepth::Ten | BitDepth::Auto => { |
| 317 |
let planes = pixels.map(|px| { |
| 318 |
let (y, u, v) = match self.color_model { |
| 319 |
ColorModel::YCbCr => rgb_to_10_bit_ycbcr(px, BT601), |
| 320 |
ColorModel::RGB => rgb_to_10_bit_gbr(px), |
| 321 |
}; |
| 322 |
[y, u, v] |
| 323 |
}); |
| 324 |
self.encode_raw_planes_10_bit(width, height, planes, None::<[_; 0]>, PixelRange::Full, matrix_coefficients) |
| 325 |
}, |
| 326 |
} |
| 327 |
} |
| 328 |
|
| 329 |
/// Encodes AVIF from 3 planar channels that are in the color space described by `matrix_coefficients`, |
| 330 |
/// with sRGB transfer characteristics and color primaries. |
| 331 |
/// |
| 332 |
/// Alpha always uses full range. Chroma subsampling is not supported, and it's a bad idea for AVIF anyway. |
| 333 |
/// If there's no alpha, use `None::<[_; 0]>`. |
| 334 |
/// |
| 335 |
/// `color_pixel_range` should be `PixelRange::Full` to avoid worsening already small 8-bit dynamic range. |
| 336 |
/// Support for limited range may be removed in the future. |
| 337 |
/// |
| 338 |
/// If `AlphaColorMode::Premultiplied` has been set, the alpha pixels must be premultiplied. |
| 339 |
/// `AlphaColorMode::UnassociatedClean` has no effect in this function, and is equivalent to `AlphaColorMode::UnassociatedDirty`. |
| 340 |
/// |
| 341 |
/// returns AVIF file, size of color metadata, size of alpha metadata overhead |
| 342 |
#[inline] |
| 343 |
pub fn encode_raw_planes_8_bit( |
| 344 |
&self, width: usize, height: usize, planes: impl IntoIterator<Item = [u8; 3]> + Send, alpha: Option<impl IntoIterator<Item = u8> + Send>, |
| 345 |
color_pixel_range: PixelRange, matrix_coefficients: MatrixCoefficients, |
| 346 |
) -> Result<EncodedImage, Error> { |
| 347 |
self.encode_raw_planes_internal(width, height, planes, alpha, color_pixel_range, matrix_coefficients, 8) |
| 348 |
} |
| 349 |
|
| 350 |
/// Encodes AVIF from 3 planar channels that are in the color space described by `matrix_coefficients`, |
| 351 |
/// with sRGB transfer characteristics and color primaries. |
| 352 |
/// |
| 353 |
/// The pixels are 10-bit (values `0.=1023`). |
| 354 |
/// |
| 355 |
/// Alpha always uses full range. Chroma subsampling is not supported, and it's a bad idea for AVIF anyway. |
| 356 |
/// If there's no alpha, use `None::<[_; 0]>`. |
| 357 |
/// |
| 358 |
/// `color_pixel_range` should be `PixelRange::Full`. Support for limited range may be removed in the future. |
| 359 |
/// |
| 360 |
/// If `AlphaColorMode::Premultiplied` has been set, the alpha pixels must be premultiplied. |
| 361 |
/// `AlphaColorMode::UnassociatedClean` has no effect in this function, and is equivalent to `AlphaColorMode::UnassociatedDirty`. |
| 362 |
/// |
| 363 |
/// returns AVIF file, size of color metadata, size of alpha metadata overhead |
| 364 |
#[inline] |
| 365 |
pub fn encode_raw_planes_10_bit( |
| 366 |
&self, width: usize, height: usize, planes: impl IntoIterator<Item = [u16; 3]> + Send, alpha: Option<impl IntoIterator<Item = u16> + Send>, |
| 367 |
color_pixel_range: PixelRange, matrix_coefficients: MatrixCoefficients, |
| 368 |
) -> Result<EncodedImage, Error> { |
| 369 |
self.encode_raw_planes_internal(width, height, planes, alpha, color_pixel_range, matrix_coefficients, 10) |
| 370 |
} |
| 371 |
|
| 372 |
#[inline(never)] |
| 373 |
fn encode_raw_planes_internal<P: rav1e::Pixel + Default>( |
| 374 |
&self, width: usize, height: usize, planes: impl IntoIterator<Item = [P; 3]> + Send, alpha: Option<impl IntoIterator<Item = P> + Send>, |
| 375 |
color_pixel_range: PixelRange, matrix_coefficients: MatrixCoefficients, input_pixels_bit_depth: u8, |
| 376 |
) -> Result<EncodedImage, Error> { |
| 377 |
let color_description = Some(ColorDescription { |
| 378 |
transfer_characteristics: TransferCharacteristics::SRGB, |
| 379 |
color_primaries: ColorPrimaries::BT709, // sRGB-compatible |
| 380 |
matrix_coefficients, |
| 381 |
}); |
| 382 |
|
| 383 |
let threads = self.threads.map(|threads| { |
| 384 |
if threads > 0 { threads } else { rayon::current_num_threads() } |
| 385 |
}); |
| 386 |
|
| 387 |
let encode_color = move || { |
| 388 |
encode_to_av1::<P>( |
| 389 |
&Av1EncodeConfig { |
| 390 |
width, |
| 391 |
height, |
| 392 |
bit_depth: input_pixels_bit_depth.into(), |
| 393 |
quantizer: self.quantizer.into(), |
| 394 |
speed: SpeedTweaks::from_my_preset(self.speed, self.quantizer), |
| 395 |
threads, |
| 396 |
pixel_range: color_pixel_range, |
| 397 |
chroma_sampling: ChromaSampling::Cs444, |
| 398 |
color_description, |
| 399 |
}, |
| 400 |
move |frame| init_frame_3(width, height, planes, frame), |
| 401 |
) |
| 402 |
}; |
| 403 |
let encode_alpha = move || { |
| 404 |
alpha.map(|alpha| { |
| 405 |
encode_to_av1::<P>( |
| 406 |
&Av1EncodeConfig { |
| 407 |
width, |
| 408 |
height, |
| 409 |
bit_depth: input_pixels_bit_depth.into(), |
| 410 |
quantizer: self.alpha_quantizer.into(), |
| 411 |
speed: SpeedTweaks::from_my_preset(self.speed, self.alpha_quantizer), |
| 412 |
threads, |
| 413 |
pixel_range: PixelRange::Full, |
| 414 |
chroma_sampling: ChromaSampling::Cs400, |
| 415 |
color_description: None, |
| 416 |
}, |
| 417 |
|frame| init_frame_1(width, height, alpha, frame), |
| 418 |
) |
| 419 |
}) |
| 420 |
}; |
| 421 |
#[cfg(all(target_arch = "wasm32", not(target_feature = "atomics")))] |
| 422 |
let (color, alpha) = (encode_color(), encode_alpha()); |
| 423 |
#[cfg(not(all(target_arch = "wasm32", not(target_feature = "atomics"))))] |
| 424 |
let (color, alpha) = rayon::join(encode_color, encode_alpha); |
| 425 |
let (color, alpha) = (color?, alpha.transpose()?); |
| 426 |
|
| 427 |
let avif_file = avif_serialize::Aviffy::new() |
| 428 |
.matrix_coefficients(match matrix_coefficients { |
| 429 |
MatrixCoefficients::Identity => avif_serialize::constants::MatrixCoefficients::Rgb, |
| 430 |
MatrixCoefficients::BT709 => avif_serialize::constants::MatrixCoefficients::Bt709, |
| 431 |
MatrixCoefficients::Unspecified => avif_serialize::constants::MatrixCoefficients::Unspecified, |
| 432 |
MatrixCoefficients::BT601 => avif_serialize::constants::MatrixCoefficients::Bt601, |
| 433 |
MatrixCoefficients::YCgCo => avif_serialize::constants::MatrixCoefficients::Ycgco, |
| 434 |
MatrixCoefficients::BT2020NCL => avif_serialize::constants::MatrixCoefficients::Bt2020Ncl, |
| 435 |
MatrixCoefficients::BT2020CL => avif_serialize::constants::MatrixCoefficients::Bt2020Cl, |
| 436 |
_ => return Err(Error::Unsupported("matrix coefficients")), |
| 437 |
}) |
| 438 |
.premultiplied_alpha(self.premultiplied_alpha) |
| 439 |
.to_vec(&color, alpha.as_deref(), width as u32, height as u32, input_pixels_bit_depth); |
| 440 |
let color_byte_size = color.len(); |
| 441 |
let alpha_byte_size = alpha.as_ref().map_or(0, |a| a.len()); |
| 442 |
|
| 443 |
Ok(EncodedImage { |
| 444 |
avif_file, color_byte_size, alpha_byte_size, |
| 445 |
}) |
| 446 |
} |
| 447 |
} |
| 448 |
|
| 449 |
#[inline(always)] |
| 450 |
fn to_ten(x: u8) -> u16 { |
| 451 |
(u16::from(x) << 2) | (u16::from(x) >> 6) |
| 452 |
} |
| 453 |
|
| 454 |
#[inline(always)] |
| 455 |
fn rgb_to_10_bit_gbr(px: rgb::RGB<u8>) -> (u16, u16, u16) { |
| 456 |
(to_ten(px.g), to_ten(px.b), to_ten(px.r)) |
| 457 |
} |
| 458 |
|
| 459 |
#[inline(always)] |
| 460 |
fn rgb_to_8_bit_gbr(px: rgb::RGB<u8>) -> (u8, u8, u8) { |
| 461 |
(px.g, px.b, px.r) |
| 462 |
} |
| 463 |
|
| 464 |
// const REC709: [f32; 3] = [0.2126, 0.7152, 0.0722]; |
| 465 |
const BT601: [f32; 3] = [0.2990, 0.5870, 0.1140]; |
| 466 |
|
| 467 |
#[inline(always)] |
| 468 |
fn rgb_to_ycbcr(px: rgb::RGB<u8>, depth: u8, matrix: [f32; 3]) -> (f32, f32, f32) { |
| 469 |
let max_value = ((1 << depth) - 1) as f32; |
| 470 |
let scale = max_value / 255.; |
| 471 |
let shift = (max_value * 0.5).round(); |
| 472 |
let y = scale * matrix[0] * f32::from(px.r) + scale * matrix[1] * f32::from(px.g) + scale * matrix[2] * f32::from(px.b); |
| 473 |
let cb = (f32::from(px.b) * scale - y).mul_add(0.5 / (1. - matrix[2]), shift); |
| 474 |
let cr = (f32::from(px.r) * scale - y).mul_add(0.5 / (1. - matrix[0]), shift); |
| 475 |
(y.round(), cb.round(), cr.round()) |
| 476 |
} |
| 477 |
|
| 478 |
#[inline(always)] |
| 479 |
fn rgb_to_10_bit_ycbcr(px: rgb::RGB<u8>, matrix: [f32; 3]) -> (u16, u16, u16) { |
| 480 |
let (y, u, v) = rgb_to_ycbcr(px, 10, matrix); |
| 481 |
(y as u16, u as u16, v as u16) |
| 482 |
} |
| 483 |
|
| 484 |
#[inline(always)] |
| 485 |
fn rgb_to_8_bit_ycbcr(px: rgb::RGB<u8>, matrix: [f32; 3]) -> (u8, u8, u8) { |
| 486 |
let (y, u, v) = rgb_to_ycbcr(px, 8, matrix); |
| 487 |
(y as u8, u as u8, v as u8) |
| 488 |
} |
| 489 |
|
| 490 |
fn quality_to_quantizer(quality: f32) -> u8 { |
| 491 |
let q = quality / 100.; |
| 492 |
let x = if q >= 0.85 { (1. - q) * 3. } else if q > 0.25 { 1. - 0.125 - q * 0.5 } else { 1. - q }; |
| 493 |
(x * 255.).round() as u8 |
| 494 |
} |
| 495 |
|
| 496 |
#[derive(Debug, Copy, Clone)] |
| 497 |
struct SpeedTweaks { |
| 498 |
pub speed_preset: u8, |
| 499 |
|
| 500 |
pub fast_deblock: Option<bool>, |
| 501 |
pub reduced_tx_set: Option<bool>, |
| 502 |
pub tx_domain_distortion: Option<bool>, |
| 503 |
pub tx_domain_rate: Option<bool>, |
| 504 |
pub encode_bottomup: Option<bool>, |
| 505 |
pub rdo_tx_decision: Option<bool>, |
| 506 |
pub cdef: Option<bool>, |
| 507 |
/// loop restoration filter |
| 508 |
pub lrf: Option<bool>, |
| 509 |
pub sgr_complexity_full: Option<bool>, |
| 510 |
pub use_satd_subpel: Option<bool>, |
| 511 |
pub inter_tx_split: Option<bool>, |
| 512 |
pub fine_directional_intra: Option<bool>, |
| 513 |
pub complex_prediction_modes: Option<bool>, |
| 514 |
pub partition_range: Option<(u8, u8)>, |
| 515 |
pub min_tile_size: u16, |
| 516 |
} |
| 517 |
|
| 518 |
impl SpeedTweaks { |
| 519 |
pub fn from_my_preset(speed: u8, quantizer: u8) -> Self { |
| 520 |
let low_quality = quantizer < quality_to_quantizer(55.); |
| 521 |
let high_quality = quantizer > quality_to_quantizer(80.); |
| 522 |
let max_block_size = if high_quality { 16 } else { 64 }; |
| 523 |
|
| 524 |
Self { |
| 525 |
speed_preset: speed, |
| 526 |
|
| 527 |
partition_range: Some(match speed { |
| 528 |
0 => (4, 64.min(max_block_size)), |
| 529 |
1 if low_quality => (4, 64.min(max_block_size)), |
| 530 |
2 if low_quality => (4, 32.min(max_block_size)), |
| 531 |
1..=4 => (4, 16), |
| 532 |
5..=8 => (8, 16), |
| 533 |
_ => (16, 16), |
| 534 |
}), |
| 535 |
|
| 536 |
complex_prediction_modes: Some(speed <= 1), // 2x-3x slower, 2% better |
| 537 |
sgr_complexity_full: Some(speed <= 2), // 15% slower, barely improves anything -/+1% |
| 538 |
|
| 539 |
encode_bottomup: Some(speed <= 2), // may be costly (+60%), may even backfire |
| 540 |
|
| 541 |
// big blocks disabled at 3 |
| 542 |
|
| 543 |
// these two are together? |
| 544 |
rdo_tx_decision: Some(speed <= 4 && !high_quality), // it tends to blur subtle textures |
| 545 |
reduced_tx_set: Some(speed == 4 || speed >= 9), // It interacts with tx_domain_distortion too? |
| 546 |
|
| 547 |
// 4px blocks disabled at 5 |
| 548 |
|
| 549 |
fine_directional_intra: Some(speed <= 6), |
| 550 |
fast_deblock: Some(speed >= 7 && !high_quality), // mixed bag? |
| 551 |
|
| 552 |
// 8px blocks disabled at 8 |
| 553 |
lrf: Some(low_quality && speed <= 8), // hardly any help for hi-q images. recovers some q at low quality |
| 554 |
cdef: Some(low_quality && speed <= 9), // hardly any help for hi-q images. recovers some q at low quality |
| 555 |
|
| 556 |
inter_tx_split: Some(speed >= 9), // mixed bag even when it works, and it backfires if not used together with reduced_tx_set |
| 557 |
tx_domain_rate: Some(speed >= 10), // 20% faster, but also 10% larger files! |
| 558 |
|
| 559 |
tx_domain_distortion: None, // very mixed bag, sometimes helps speed sometimes it doesn't |
| 560 |
use_satd_subpel: Some(false), // doesn't make sense |
| 561 |
min_tile_size: match speed { |
| 562 |
0 => 4096, |
| 563 |
1 => 2048, |
| 564 |
2 => 1024, |
| 565 |
3 => 512, |
| 566 |
4 => 256, |
| 567 |
_ => 128, |
| 568 |
} * if high_quality { 2 } else { 1 }, |
| 569 |
} |
| 570 |
} |
| 571 |
|
| 572 |
pub(crate) fn speed_settings(&self) -> SpeedSettings { |
| 573 |
let mut speed_settings = SpeedSettings::from_preset(self.speed_preset); |
| 574 |
|
| 575 |
speed_settings.multiref = false; |
| 576 |
speed_settings.rdo_lookahead_frames = 1; |
| 577 |
speed_settings.scene_detection_mode = SceneDetectionSpeed::None; |
| 578 |
speed_settings.motion.include_near_mvs = false; |
| 579 |
|
| 580 |
if let Some(v) = self.fast_deblock { speed_settings.fast_deblock = v; } |
| 581 |
if let Some(v) = self.reduced_tx_set { speed_settings.transform.reduced_tx_set = v; } |
| 582 |
if let Some(v) = self.tx_domain_distortion { speed_settings.transform.tx_domain_distortion = v; } |
| 583 |
if let Some(v) = self.tx_domain_rate { speed_settings.transform.tx_domain_rate = v; } |
| 584 |
if let Some(v) = self.encode_bottomup { speed_settings.partition.encode_bottomup = v; } |
| 585 |
if let Some(v) = self.rdo_tx_decision { speed_settings.transform.rdo_tx_decision = v; } |
| 586 |
if let Some(v) = self.cdef { speed_settings.cdef = v; } |
| 587 |
if let Some(v) = self.lrf { speed_settings.lrf = v; } |
| 588 |
if let Some(v) = self.inter_tx_split { speed_settings.transform.enable_inter_tx_split = v; } |
| 589 |
if let Some(v) = self.sgr_complexity_full { speed_settings.sgr_complexity = if v { SGRComplexityLevel::Full } else { SGRComplexityLevel::Reduced } }; |
| 590 |
if let Some(v) = self.use_satd_subpel { speed_settings.motion.use_satd_subpel = v; } |
| 591 |
if let Some(v) = self.fine_directional_intra { speed_settings.prediction.fine_directional_intra = v; } |
| 592 |
if let Some(v) = self.complex_prediction_modes { speed_settings.prediction.prediction_modes = if v { PredictionModesSetting::ComplexAll } else { PredictionModesSetting::Simple} }; |
| 593 |
if let Some((min, max)) = self.partition_range { |
| 594 |
debug_assert!(min <= max); |
| 595 |
fn sz(s: u8) -> BlockSize { |
| 596 |
match s { |
| 597 |
4 => BlockSize::BLOCK_4X4, |
| 598 |
8 => BlockSize::BLOCK_8X8, |
| 599 |
16 => BlockSize::BLOCK_16X16, |
| 600 |
32 => BlockSize::BLOCK_32X32, |
| 601 |
64 => BlockSize::BLOCK_64X64, |
| 602 |
128 => BlockSize::BLOCK_128X128, |
| 603 |
_ => panic!("bad size {s}"), |
| 604 |
} |
| 605 |
} |
| 606 |
speed_settings.partition.partition_range = PartitionRange::new(sz(min), sz(max)); |
| 607 |
} |
| 608 |
|
| 609 |
speed_settings |
| 610 |
} |
| 611 |
} |
| 612 |
|
| 613 |
struct Av1EncodeConfig { |
| 614 |
pub width: usize, |
| 615 |
pub height: usize, |
| 616 |
pub bit_depth: usize, |
| 617 |
pub quantizer: usize, |
| 618 |
pub speed: SpeedTweaks, |
| 619 |
/// 0 means num_cpus |
| 620 |
pub threads: Option<usize>, |
| 621 |
pub pixel_range: PixelRange, |
| 622 |
pub chroma_sampling: ChromaSampling, |
| 623 |
pub color_description: Option<ColorDescription>, |
| 624 |
} |
| 625 |
|
| 626 |
fn rav1e_config(p: &Av1EncodeConfig) -> Config { |
| 627 |
// AV1 needs all the CPU power you can give it, |
| 628 |
// except when it'd create inefficiently tiny tiles |
| 629 |
let tiles = { |
| 630 |
let threads = p.threads.unwrap_or_else(rayon::current_num_threads); |
| 631 |
threads.min((p.width * p.height) / (p.speed.min_tile_size as usize).pow(2)) |
| 632 |
}; |
| 633 |
let speed_settings = p.speed.speed_settings(); |
| 634 |
let cfg = Config::new() |
| 635 |
.with_encoder_config(EncoderConfig { |
| 636 |
width: p.width, |
| 637 |
height: p.height, |
| 638 |
time_base: Rational::new(1, 1), |
| 639 |
sample_aspect_ratio: Rational::new(1, 1), |
| 640 |
bit_depth: p.bit_depth, |
| 641 |
chroma_sampling: p.chroma_sampling, |
| 642 |
chroma_sample_position: ChromaSamplePosition::Unknown, |
| 643 |
pixel_range: p.pixel_range, |
| 644 |
color_description: p.color_description, |
| 645 |
mastering_display: None, |
| 646 |
content_light: None, |
| 647 |
enable_timing_info: false, |
| 648 |
still_picture: true, |
| 649 |
error_resilient: false, |
| 650 |
switch_frame_interval: 0, |
| 651 |
min_key_frame_interval: 0, |
| 652 |
max_key_frame_interval: 0, |
| 653 |
reservoir_frame_delay: None, |
| 654 |
low_latency: false, |
| 655 |
quantizer: p.quantizer, |
| 656 |
min_quantizer: p.quantizer as _, |
| 657 |
bitrate: 0, |
| 658 |
tune: Tune::Psychovisual, |
| 659 |
tile_cols: 0, |
| 660 |
tile_rows: 0, |
| 661 |
tiles, |
| 662 |
film_grain_params: None, |
| 663 |
level_idx: None, |
| 664 |
speed_settings, |
| 665 |
}); |
| 666 |
|
| 667 |
if let Some(threads) = p.threads { |
| 668 |
cfg.with_threads(threads) |
| 669 |
} else { |
| 670 |
cfg |
| 671 |
} |
| 672 |
} |
| 673 |
|
| 674 |
fn init_frame_3<P: rav1e::Pixel + Default>( |
| 675 |
width: usize, height: usize, planes: impl IntoIterator<Item = [P; 3]> + Send, frame: &mut Frame<P>, |
| 676 |
) -> Result<(), Error> { |
| 677 |
let mut f = frame.planes.iter_mut(); |
| 678 |
let mut planes = planes.into_iter(); |
| 679 |
|
| 680 |
// it doesn't seem to be necessary to fill padding area |
| 681 |
let mut y = f.next().unwrap().mut_slice(Default::default()); |
| 682 |
let mut u = f.next().unwrap().mut_slice(Default::default()); |
| 683 |
let mut v = f.next().unwrap().mut_slice(Default::default()); |
| 684 |
|
| 685 |
for ((y, u), v) in y.rows_iter_mut().zip(u.rows_iter_mut()).zip(v.rows_iter_mut()).take(height) { |
| 686 |
let y = &mut y[..width]; |
| 687 |
let u = &mut u[..width]; |
| 688 |
let v = &mut v[..width]; |
| 689 |
for ((y, u), v) in y.iter_mut().zip(u).zip(v) { |
| 690 |
let px = planes.next().ok_or(Error::TooFewPixels)?; |
| 691 |
*y = px[0]; |
| 692 |
*u = px[1]; |
| 693 |
*v = px[2]; |
| 694 |
} |
| 695 |
} |
| 696 |
Ok(()) |
| 697 |
} |
| 698 |
|
| 699 |
fn init_frame_1<P: rav1e::Pixel + Default>(width: usize, height: usize, planes: impl IntoIterator<Item = P> + Send, frame: &mut Frame<P>) -> Result<(), Error> { |
| 700 |
let mut y = frame.planes[0].mut_slice(Default::default()); |
| 701 |
let mut planes = planes.into_iter(); |
| 702 |
|
| 703 |
for y in y.rows_iter_mut().take(height) { |
| 704 |
let y = &mut y[..width]; |
| 705 |
for y in y.iter_mut() { |
| 706 |
*y = planes.next().ok_or(Error::TooFewPixels)?; |
| 707 |
} |
| 708 |
} |
| 709 |
Ok(()) |
| 710 |
} |
| 711 |
|
| 712 |
#[inline(never)] |
| 713 |
fn encode_to_av1<P: rav1e::Pixel>(p: &Av1EncodeConfig, init: impl FnOnce(&mut Frame<P>) -> Result<(), Error>) -> Result<Vec<u8>, Error> { |
| 714 |
let mut ctx: Context<P> = rav1e_config(p).new_context()?; |
| 715 |
let mut frame = ctx.new_frame(); |
| 716 |
|
| 717 |
init(&mut frame)?; |
| 718 |
ctx.send_frame(frame)?; |
| 719 |
ctx.flush(); |
| 720 |
|
| 721 |
let mut out = Vec::new(); |
| 722 |
loop { |
| 723 |
match ctx.receive_packet() { |
| 724 |
Ok(mut packet) => match packet.frame_type { |
| 725 |
FrameType::KEY => { |
| 726 |
out.append(&mut packet.data); |
| 727 |
}, |
| 728 |
_ => continue, |
| 729 |
}, |
| 730 |
Err(EncoderStatus::Encoded) | |
| 731 |
Err(EncoderStatus::LimitReached) => break, |
| 732 |
Err(err) => Err(err)?, |
| 733 |
} |
| 734 |
} |
| 735 |
Ok(out) |
| 736 |
} |