av1_grain/create.rs
1// Copyright (c) 2022-2022, The rav1e contributors. All rights reserved
2//
3// This source code is subject to the terms of the BSD 2 Clause License and
4// the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
5// was not distributed with this source code in the LICENSE file, you can
6// obtain it at www.aomedia.org/license/software. If the Alliance for Open
7// Media Patent License 1.0 was not distributed with this source code in the
8// PATENTS file, you can obtain it at www.aomedia.org/license/patent.
9
10// The original work for this formula was implmented in aomenc, and this is
11// an adaptation of that work:
12// https://aomedia.googlesource.com/aom/+/refs/heads/main/examples/photon_noise_table.c
13
14// This implementation creates a film grain table, for use in stills and videos,
15// representing the noise that one would get by shooting with a digital camera
16// at a given light level. Much of the noise in digital images is photon shot
17// noise, which is due to the characteristics of photon arrival and grows in
18// standard deviation as the square root of the expected number of photons
19// captured.
20// https://www.photonstophotos.net/Emil%20Martinec/noise.html#shotnoise
21//
22// The proxy used by this implementation for the amount of light captured
23// is the ISO value such that the focal plane exposure at the time of capture
24// would have been mapped by a 35mm camera to the output lightness observed
25// in the image. That is, if one were to shoot on a 35mm camera (36×24mm sensor)
26// at the nominal exposure for that ISO setting, the resulting image should
27// contain noise of the same order of magnitude as generated by this
28// implementation.
29//
30// The (mostly) square-root relationship between light intensity and noise
31// amplitude holds in linear light, but AV1 streams are most often encoded
32// non-linearly, and the film grain is applied to those non-linear values.
33// Therefore, this implementation must account for the non-linearity, and this
34// is controlled by the transfer function parameter, which specifies the tone
35// response curve that will be used when encoding the actual image. The default
36// for this implementation is BT.1886, which is approximately similar to an
37// encoding gamma of 1/2.8 (i.e. a decoding gamma of 2.8) though not quite
38// identical.
39//
40// As alluded to above, the implementation assumes that the image is taken from
41// the entirety of a 36×24mm (“35mm format”) sensor. If that assumption does not
42// hold, then a “35mm-equivalent ISO value” that can be passed to the
43// implementation can be obtained by multiplying the true ISO value by the ratio
44// of 36×24mm to the area that was actually used. For formats that approximately
45// share the same aspect ratio, this is often expressed as the square of the
46// “equivalence ratio” which is the ratio of their diagonals. For example, APS-C
47// (often ~24×16mm) is said to have an equivalence ratio of 1.5 relative to the
48// 35mm format, and therefore ISO 1000 on APS-C and ISO 1000×1.5² = 2250 on 35mm
49// produce an image of the same lightness from the same amount of light spread
50// onto their respective surface areas (resulting in different focal plane
51// exposures), and those images will thus have similar amounts of noise if the
52// cameras are of similar technology. https://doi.org/10.1117/1.OE.57.11.110801
53//
54// The implementation needs to know the resolution of the images to which its
55// grain tables will be applied so that it can know how the light on the sensor
56// was shared between its pixels. As a general rule, while a higher pixel count
57// will lead to more noise per pixel, when the final image is viewed at the same
58// physical size, that noise will tend to “average out” to the same amount over
59// a given area, since there will be more pixels in it which, in aggregate, will
60// have received essentially as much light. Put differently, the amount of noise
61// depends on the scale at which it is measured, and the decision for this
62// implementation was to make that scale relative to the image instead of its
63// constituent samples. For more on this, see:
64//
65// https://www.photonstophotos.net/Emil%20Martinec/noise-p3.html#pixelsize
66// https://www.dpreview.com/articles/5365920428/the-effect-of-pixel-and-sensor-sizes-on-noise/2
67// https://www.dpreview.com/videos/7940373140/dpreview-tv-why-lower-resolution-sensors-are-not-better-in-low-light
68
69use std::{
70 fs::File,
71 io::{BufWriter, Write},
72 path::Path,
73};
74
75use arrayvec::ArrayVec;
76
77use crate::{DEFAULT_GRAIN_SEED, GrainTableSegment, NUM_Y_POINTS, ScalingPoints};
78
79const PQ_M1: f32 = 2610. / 16384.;
80const PQ_M2: f32 = 128. * 2523. / 4096.;
81const PQ_C1: f32 = 3424. / 4096.;
82const PQ_C2: f32 = 32. * 2413. / 4096.;
83const PQ_C3: f32 = 32. * 2392. / 4096.;
84
85const BT1886_WHITEPOINT: f32 = 203.;
86const BT1886_BLACKPOINT: f32 = 0.1;
87const BT1886_GAMMA: f32 = 2.4;
88
89// BT.1886 formula from https://en.wikipedia.org/wiki/ITU-R_BT.1886.
90//
91// TODO: the inverses, alpha, and beta should all be constants
92// once floats in const fns are stabilized and `powf` is const.
93// Until then, `inline(always)` gets us close enough.
94
95fn bt1886_inv_whitepoint() -> f32 {
96 BT1886_WHITEPOINT.powf(1.0 / BT1886_GAMMA)
97}
98
99fn bt1886_inv_blackpoint() -> f32 {
100 BT1886_BLACKPOINT.powf(1.0 / BT1886_GAMMA)
101}
102
103/// The variable for user gain:
104/// `α = (Lw^(1/λ) - Lb^(1/λ)) ^ λ`
105fn bt1886_alpha() -> f32 {
106 (bt1886_inv_whitepoint() - bt1886_inv_blackpoint()).powf(BT1886_GAMMA)
107}
108
109/// The variable for user black level lift:
110/// `β = Lb^(1/λ) / (Lw^(1/λ) - Lb^(1/λ))`
111fn bt1886_beta() -> f32 {
112 bt1886_inv_blackpoint() / (bt1886_inv_whitepoint() - bt1886_inv_blackpoint())
113}
114
115/// Settings and video data defining how to generate the film grain params.
116#[derive(Debug, Clone, Copy)]
117pub struct NoiseGenArgs {
118 pub iso_setting: u32,
119 pub width: u32,
120 pub height: u32,
121 pub transfer_function: TransferFunction,
122 /// Whether the input is full range or limited range
123 pub full_range: bool,
124 pub chroma_grain: bool,
125 pub random_seed: Option<u16>,
126}
127
128/// Generates a set of photon noise parameters for a segment of video
129/// given a set of `args`.
130#[must_use]
131#[inline]
132pub fn generate_photon_noise_params(
133 start_time: u64,
134 end_time: u64,
135 args: NoiseGenArgs,
136) -> GrainTableSegment {
137 GrainTableSegment {
138 start_time,
139 end_time,
140 scaling_points_y: generate_luma_noise_points(args),
141 scaling_points_cb: ArrayVec::new(),
142 scaling_points_cr: ArrayVec::new(),
143 scaling_shift: 8,
144 ar_coeff_lag: 0,
145 ar_coeffs_y: ArrayVec::new(),
146 ar_coeffs_cb: ArrayVec::try_from([0].as_slice())
147 .expect("Cannot fail creation from const array"),
148 ar_coeffs_cr: ArrayVec::try_from([0].as_slice())
149 .expect("Cannot fail creation from const array"),
150 ar_coeff_shift: 6,
151 cb_mult: 0,
152 cb_luma_mult: 0,
153 cb_offset: 0,
154 cr_mult: 0,
155 cr_luma_mult: 0,
156 cr_offset: 0,
157 overlap_flag: true,
158 chroma_scaling_from_luma: args.chroma_grain,
159 grain_scale_shift: 0,
160 random_seed: args.random_seed.unwrap_or(DEFAULT_GRAIN_SEED),
161 }
162}
163
164/// Generates a set of film grain parameters for a segment of video
165/// given a set of `args`.
166///
167/// # Panics
168/// - This is not yet implemented, so it will always panic
169#[must_use]
170#[inline]
171#[cfg(feature = "unstable")]
172pub fn generate_film_grain_params(
173 start_time: u64,
174 end_time: u64,
175 args: NoiseGenArgs,
176) -> GrainTableSegment {
177 todo!("SCIENCE");
178 // GrainTableSegment {
179 // start_time,
180 // end_time,
181 // scaling_points_y: generate_luma_noise_points(args),
182 // scaling_points_cb: ArrayVec::new(),
183 // scaling_points_cr: ArrayVec::new(),
184 // scaling_shift: 8,
185 // ar_coeff_lag: 0,
186 // ar_coeffs_y: ArrayVec::new(),
187 // ar_coeffs_cb: ArrayVec::try_from([0].as_slice())
188 // .expect("Cannot fail creation from const array"),
189 // ar_coeffs_cr: ArrayVec::try_from([0].as_slice())
190 // .expect("Cannot fail creation from const array"),
191 // ar_coeff_shift: 6,
192 // cb_mult: 0,
193 // cb_luma_mult: 0,
194 // cb_offset: 0,
195 // cr_mult: 0,
196 // cr_luma_mult: 0,
197 // cr_offset: 0,
198 // overlap_flag: true,
199 // chroma_scaling_from_luma: args.chroma_grain,
200 // grain_scale_shift: 0,
201 // random_seed: args.random_seed.unwrap_or(DEFAULT_GRAIN_SEED),
202 // }
203}
204
205/// Write a set of generated film grain params to a table file,
206/// using the standard film grain table format supported by
207/// aomenc, rav1e, and svt-av1.
208///
209/// # Errors
210///
211/// - If the output file cannot be written to
212#[inline]
213pub fn write_grain_table<P: AsRef<Path>>(
214 filename: P,
215 params: &[GrainTableSegment],
216) -> anyhow::Result<()> {
217 let mut file = BufWriter::new(File::create(filename)?);
218 writeln!(&mut file, "filmgrn1")?;
219 for segment in params {
220 write_film_grain_segment(segment, &mut file)?;
221 }
222 file.flush()?;
223
224 Ok(())
225}
226
227fn write_film_grain_segment(
228 params: &GrainTableSegment,
229 output: &mut BufWriter<File>,
230) -> anyhow::Result<()> {
231 writeln!(
232 output,
233 "E {} {} 1 {} 1",
234 params.start_time, params.end_time, params.random_seed,
235 )?;
236 writeln!(
237 output,
238 "\tp {} {} {} {} {} {} {} {} {} {} {} {}",
239 params.ar_coeff_lag,
240 params.ar_coeff_shift,
241 params.grain_scale_shift,
242 params.scaling_shift,
243 u8::from(params.chroma_scaling_from_luma),
244 u8::from(params.overlap_flag),
245 params.cb_mult,
246 params.cb_luma_mult,
247 params.cb_offset,
248 params.cr_mult,
249 params.cr_luma_mult,
250 params.cr_offset
251 )?;
252
253 write!(output, "\tsY {} ", params.scaling_points_y.len())?;
254 for point in ¶ms.scaling_points_y {
255 write!(output, " {} {}", point[0], point[1])?;
256 }
257 writeln!(output)?;
258
259 write!(output, "\tsCb {}", params.scaling_points_cb.len())?;
260 for point in ¶ms.scaling_points_cb {
261 write!(output, " {} {}", point[0], point[1])?;
262 }
263 writeln!(output)?;
264
265 write!(output, "\tsCr {}", params.scaling_points_cr.len())?;
266 for point in ¶ms.scaling_points_cr {
267 write!(output, " {} {}", point[0], point[1])?;
268 }
269 writeln!(output)?;
270
271 write!(output, "\tcY")?;
272 for coeff in ¶ms.ar_coeffs_y {
273 write!(output, " {}", *coeff)?;
274 }
275 writeln!(output)?;
276
277 write!(output, "\tcCb")?;
278 for coeff in ¶ms.ar_coeffs_cb {
279 write!(output, " {}", *coeff)?;
280 }
281 writeln!(output)?;
282
283 write!(output, "\tcCr")?;
284 for coeff in ¶ms.ar_coeffs_cr {
285 write!(output, " {}", *coeff)?;
286 }
287 writeln!(output)?;
288
289 Ok(())
290}
291
292#[allow(clippy::upper_case_acronyms)]
293#[derive(Debug, Clone, Copy, PartialEq, Eq)]
294pub enum TransferFunction {
295 /// For SDR content
296 BT1886,
297 /// For HDR content
298 SMPTE2084,
299}
300
301impl TransferFunction {
302 #[must_use]
303 #[inline]
304 pub fn to_linear(self, x: f32) -> f32 {
305 match self {
306 TransferFunction::BT1886 => {
307 // The screen luminance in cd/m^2:
308 // L = α * (x + β)^λ
309 let luma = bt1886_alpha() * (x + bt1886_beta()).powf(BT1886_GAMMA);
310
311 // Normalize to between 0.0 and 1.0
312 luma / BT1886_WHITEPOINT
313 }
314 TransferFunction::SMPTE2084 => {
315 let pq_pow_inv_m2 = x.powf(1. / PQ_M2);
316 (0_f32.max(pq_pow_inv_m2 - PQ_C1) / PQ_C3.mul_add(-pq_pow_inv_m2, PQ_C2))
317 .powf(1. / PQ_M1)
318 }
319 }
320 }
321
322 #[allow(clippy::wrong_self_convention)]
323 #[must_use]
324 #[inline]
325 pub fn from_linear(self, x: f32) -> f32 {
326 match self {
327 TransferFunction::BT1886 => {
328 // Scale to a raw cd/m^2 value
329 let luma = x * BT1886_WHITEPOINT;
330
331 // The inverse of the `to_linear` formula:
332 // `(L / α)^(1 / λ) - β = x`
333 (luma / bt1886_alpha()).powf(1.0 / BT1886_GAMMA) - bt1886_beta()
334 }
335 TransferFunction::SMPTE2084 => {
336 if x < f32::EPSILON {
337 return 0.0;
338 }
339 let linear_pow_m1 = x.powf(PQ_M1);
340 (PQ_C2.mul_add(linear_pow_m1, PQ_C1) / PQ_C3.mul_add(linear_pow_m1, 1.)).powf(PQ_M2)
341 }
342 }
343 }
344
345 #[inline]
346 #[must_use]
347 pub fn mid_tone(self) -> f32 {
348 self.to_linear(0.5)
349 }
350}
351
352fn generate_luma_noise_points(args: NoiseGenArgs) -> ScalingPoints {
353 // Assumes a daylight-like spectrum.
354 // https://www.strollswithmydog.com/effective-quantum-efficiency-of-sensor/#:~:text=11%2C260%20photons/um%5E2/lx-s
355 const PHOTONS_PER_SQ_MICRON_PER_LUX_SECOND: f32 = 11260.;
356
357 // Order of magnitude for cameras in the 2010-2020 decade, taking the CFA into
358 // account.
359 const EFFECTIVE_QUANTUM_EFFICIENCY: f32 = 0.2;
360
361 // Also reasonable values for current cameras. The read noise is typically
362 // higher than this at low ISO settings but it matters less there.
363 const PHOTO_RESPONSE_NON_UNIFORMITY: f32 = 0.005;
364 const INPUT_REFERRED_READ_NOISE: f32 = 1.5;
365
366 // Assumes a 35mm sensor (36mm × 24mm).
367 const SENSOR_AREA: f32 = 36_000. * 24_000.;
368
369 // Focal plane exposure for a mid-tone (typically a 18% reflectance card), in
370 // lx·s.
371 let mid_tone_exposure = 10. / args.iso_setting as f32;
372
373 let pixel_area_microns = SENSOR_AREA / (args.width * args.height) as f32;
374
375 let mid_tone_electrons_per_pixel = EFFECTIVE_QUANTUM_EFFICIENCY
376 * PHOTONS_PER_SQ_MICRON_PER_LUX_SECOND
377 * mid_tone_exposure
378 * pixel_area_microns;
379 let max_electrons_per_pixel = mid_tone_electrons_per_pixel / args.transfer_function.mid_tone();
380
381 let max_value = if args.full_range { 255 } else { 235 };
382 let min_value = if args.full_range { 0 } else { 16 };
383 let range = max_value - min_value;
384 const RAMP_OFFSET: usize = 3;
385 const MIN_EDGE: usize = 0;
386 const MAX_EDGE: usize = NUM_Y_POINTS - 1;
387
388 let mut scaling_points = ScalingPoints::default();
389 for i in 0..NUM_Y_POINTS {
390 // Applying photon noise "as is" results in unwanted brightening of darkest and darkening of brightest luma values;
391 // clamping scaling points to a maximum of 1 at those min and max values prevents that.
392 let x = if i == MIN_EDGE {
393 0.0
394 } else if i == MAX_EDGE {
395 1.0
396 } else {
397 // (ramp_offset + (range - 2 * ramp_offset) * ((i - 1) / (film_grain->num_y_points - 3.0))) / range
398 ((range - 2 * RAMP_OFFSET) as f32).mul_add(
399 (i - 1) as f32 / (NUM_Y_POINTS - 3) as f32,
400 RAMP_OFFSET as f32,
401 ) / range as f32
402 };
403
404 let linear = args.transfer_function.to_linear(x);
405 let electrons_per_pixel = max_electrons_per_pixel * linear;
406
407 // Quadrature sum of the relevant sources of noise, in electrons rms. Photon
408 // shot noise is sqrt(electrons) so we can skip the square root and the
409 // squaring.
410 // https://en.wikipedia.org/wiki/Addition_in_quadrature
411 // https://doi.org/10.1117/3.725073
412 let noise_in_electrons = (PHOTO_RESPONSE_NON_UNIFORMITY
413 * PHOTO_RESPONSE_NON_UNIFORMITY
414 * electrons_per_pixel)
415 .mul_add(
416 electrons_per_pixel,
417 INPUT_REFERRED_READ_NOISE.mul_add(INPUT_REFERRED_READ_NOISE, electrons_per_pixel),
418 )
419 .sqrt();
420 let linear_noise = noise_in_electrons / max_electrons_per_pixel;
421 let linear_range_start = 0_f32.max(2.0f32.mul_add(-linear_noise, linear));
422 let linear_range_end = 1_f32.min(2_f32.mul_add(linear_noise, linear));
423 let tf_slope = (args.transfer_function.from_linear(linear_range_end)
424 - args.transfer_function.from_linear(linear_range_start))
425 / (linear_range_end - linear_range_start);
426 let encoded_noise = linear_noise * tf_slope;
427
428 // min_value as f32 + range as f32 * x
429 let x = (range as f32).mul_add(x, min_value as f32).round() as u8;
430 let mut encoded_noise =
431 (range as f32).min((range as f32 * 7.88 * encoded_noise).round()) as u8;
432 if i == MIN_EDGE || i == MAX_EDGE {
433 encoded_noise = if encoded_noise >= 1 { 1 } else { 0 };
434 }
435
436 scaling_points.push([x, encoded_noise]);
437 }
438
439 scaling_points
440}
441
442#[cfg(test)]
443mod tests {
444 use quickcheck::TestResult;
445 use quickcheck_macros::quickcheck;
446
447 use super::*;
448
449 #[quickcheck]
450 fn bt1886_to_linear_within_range(x: f32) -> TestResult {
451 if !(0.0..=1.0).contains(&x) || x.is_nan() {
452 return TestResult::discard();
453 }
454
455 let tx = TransferFunction::BT1886;
456 let res = tx.to_linear(x);
457 TestResult::from_bool((0.0..=1.0).contains(&res))
458 }
459
460 #[quickcheck]
461 fn bt1886_to_linear_reverts_correctly(x: f32) -> TestResult {
462 if !(0.0..=1.0).contains(&x) || x.is_nan() {
463 return TestResult::discard();
464 }
465
466 let tx = TransferFunction::BT1886;
467 let res = tx.to_linear(x);
468 let res = tx.from_linear(res);
469 TestResult::from_bool((x - res).abs() < f32::EPSILON)
470 }
471
472 #[quickcheck]
473 fn smpte2084_to_linear_within_range(x: f32) -> TestResult {
474 if !(0.0..=1.0).contains(&x) || x.is_nan() {
475 return TestResult::discard();
476 }
477
478 let tx = TransferFunction::SMPTE2084;
479 let res = tx.to_linear(x);
480 TestResult::from_bool((0.0..=1.0).contains(&res))
481 }
482
483 #[quickcheck]
484 fn smpte2084_to_linear_reverts_correctly(x: f32) -> TestResult {
485 if !(0.0..=1.0).contains(&x) || x.is_nan() {
486 return TestResult::discard();
487 }
488
489 let tx = TransferFunction::SMPTE2084;
490 let res = tx.to_linear(x);
491 let res = tx.from_linear(res);
492 TestResult::from_bool((x - res).abs() < f32::EPSILON)
493 }
494}