Fix banding from linear RGB color space conversion

codecs/resize/src/lib.rs

@@ -5,12 +5,12 @@ extern crate wasm_bindgen;
 mod utils;
 
 use cfg_if::cfg_if;
-use resize::Pixel::RGBA;
+use resize::Pixel;
 use resize::Type;
 use wasm_bindgen::prelude::*;
 
 mod srgb;
-use srgb::Clamp;
+use srgb::{linear_to_srgb, srgb_to_linear, Clamp};
 
 cfg_if! {
     // When the `wee_alloc` feature is enabled, use `wee_alloc` as the global
@@ -27,14 +27,17 @@ include!("./lut.inc");
 // If `with_space_conversion` is true, this function returns 2 functions that
 // convert from sRGB to linear RGB and vice versa. If `with_space_conversion` is
 // false, the 2 functions returned do nothing.
-fn converter_funcs(with_space_conversion: bool) -> ((fn(u8) -> f32), (fn(f32) -> u8)) {
+fn srgb_converter_funcs(with_space_conversion: bool) -> (fn(u8) -> f32, fn(f32) -> u8) {
     if with_space_conversion {
         (
-            |v| SRGB_TO_LINEAR_LUT[v as usize] * 255.0,
-            |v| (LINEAR_TO_SRGB_LUT[v as usize] * 255.0) as u8,
+            |v| SRGB_TO_LINEAR_LUT[v as usize],
+            |v| (linear_to_srgb(v) * 255.0).clamp(0.0, 255.0) as u8,
         )
     } else {
-        (|v| v as f32, |v| v as u8)
+        (
+            |v| (v as f32) / 255.0,
+            |v| (v * 255.0).clamp(0.0, 255.0) as u8,
+        )
     }
 }
 
@@ -44,21 +47,24 @@ fn converter_funcs(with_space_conversion: bool) -> ((fn(u8) -> f32), (fn(f32) ->
 // false, the functions just return the channel value.
 fn alpha_multiplier_funcs(
     with_alpha_premultiplication: bool,
-) -> ((fn(f32, u8) -> u8), (fn(u8, u8) -> f32)) {
+) -> (fn(f32, f32) -> f32, fn(f32, f32) -> f32) {
     if with_alpha_premultiplication {
-        (
-            |v, a| (v * (a as f32) / 255.0) as u8,
-            |v, a| ((v as f32) * 255.0 / (a as f32)).clamp(0.0, 255.0),
-        )
+        (|v, a| v * a, |v, a| v / a)
     } else {
-        (|v, _a| v as u8, |v, _a| v as f32)
+        (|v, _a| v, |v, _a| v)
     }
 }
 
+fn vec_with_len<T: Clone>(len: usize, item: T) -> Vec<T> {
+    let mut v: Vec<T> = Vec::with_capacity(len);
+    v.resize(len, item);
+    return v;
+}
+
 #[wasm_bindgen]
 #[no_mangle]
 pub fn resize(
-    mut input_image: Vec<u8>,
+    input_image: Vec<u8>,
     input_width: usize,
     input_height: usize,
     output_width: usize,
@@ -77,44 +83,64 @@ pub fn resize(
     let num_input_pixels = input_width * input_height;
     let num_output_pixels = output_width * output_height;
 
-    let (to_linear, to_color_space) = converter_funcs(color_space_conversion);
-    let (premultiplier, demultiplier) = alpha_multiplier_funcs(premultiply);
+    let mut output_image: Vec<u8> = vec_with_len(num_output_pixels * 4, 0);
 
     // If both options are false, there is no preprocessing on the pixel valus
     // and we can skip the loop.
-    if premultiply || color_space_conversion {
-        for i in 0..num_input_pixels {
-            for j in 0..3 {
-                input_image[4 * i + j] =
-                    premultiplier(to_linear(input_image[4 * i + j]), input_image[4 * i + 3]);
-            }
+    if !premultiply && !color_space_conversion {
+        let mut resizer = resize::new(
+            input_width,
+            input_height,
+            output_width,
+            output_height,
+            Pixel::RGBA,
+            typ,
+        );
+        resizer.resize(input_image.as_slice(), output_image.as_mut_slice());
+        return output_image;
+    }
+
+    // Otherwise, we convert to f32 images so we don’t introduce
+    // banding through the conversions.
+    let (to_linear, to_srgb) = srgb_converter_funcs(color_space_conversion);
+    let (premultiplier, demultiplier) = alpha_multiplier_funcs(premultiply);
+
+    let mut preprocessed_input_image: Vec<f32> = Vec::with_capacity(input_image.len());
+    preprocessed_input_image.resize(input_image.len(), 0.0f32);
+    for i in 0..num_input_pixels {
+        for j in 0..3 {
+            preprocessed_input_image[4 * i + j] = premultiplier(
+                to_linear(input_image[4 * i + j]),
+                (input_image[4 * i + 3] as f32) / 255.0,
+            );
+            preprocessed_input_image[4 * i + 3] = (input_image[4 * i + 3] as f32) / 255.0;
         }
     }
 
+    let mut unprocessed_output_image: Vec<f32> = vec_with_len(num_output_pixels * 4, 0.0f32);
+
     let mut resizer = resize::new(
         input_width,
         input_height,
         output_width,
         output_height,
-        RGBA,
+        Pixel::RGBAF32,
         typ,
     );
-    let mut output_image = Vec::<u8>::with_capacity(num_output_pixels * 4);
-    output_image.resize(num_output_pixels * 4, 0);
-    resizer.resize(input_image.as_slice(), output_image.as_mut_slice());
+    resizer.resize(
+        preprocessed_input_image.as_slice(),
+        unprocessed_output_image.as_mut_slice(),
+    );
 
-    if premultiply || color_space_conversion {
-        for i in 0..num_output_pixels {
-            for j in 0..3 {
-                // We don’t need to worry about division by zero, as division by zero
-                // is well-defined on floats to return ±Inf. ±Inf is converted to 0
-                // when casting to integers.
-                output_image[4 * i + j] = to_color_space(demultiplier(
-                    output_image[4 * i + j],
-                    output_image[4 * i + 3],
-                ));
-            }
+    for i in 0..num_output_pixels {
+        for j in 0..3 {
+            output_image[4 * i + j] = to_srgb(demultiplier(
+                unprocessed_output_image[4 * i + j],
+                unprocessed_output_image[4 * i + 3],
+            ));
         }
+        output_image[4 * i + 3] =
+            (unprocessed_output_image[4 * i + 3] * 255.0).clamp(0.0, 255.0) as u8;
     }
 
     return output_image;