Revamp alpha optimization

benches/filters.rs

@@ -13,7 +13,7 @@ fn filters_16_bits_filter_0(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::None);
+        png.raw.filter_image(RowFilter::None, false);
     });
 }
 
@@ -23,7 +23,7 @@ fn filters_8_bits_filter_0(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::None);
+        png.raw.filter_image(RowFilter::None, false);
     });
 }
 
@@ -35,7 +35,7 @@ fn filters_4_bits_filter_0(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::None);
+        png.raw.filter_image(RowFilter::None, false);
     });
 }
 
@@ -47,7 +47,7 @@ fn filters_2_bits_filter_0(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::None);
+        png.raw.filter_image(RowFilter::None, false);
     });
 }
 
@@ -59,7 +59,7 @@ fn filters_1_bits_filter_0(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::None);
+        png.raw.filter_image(RowFilter::None, false);
     });
 }
 
@@ -69,7 +69,7 @@ fn filters_16_bits_filter_1(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Sub);
+        png.raw.filter_image(RowFilter::Sub, false);
     });
 }
 
@@ -79,7 +79,7 @@ fn filters_8_bits_filter_1(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Sub);
+        png.raw.filter_image(RowFilter::Sub, false);
     });
 }
 
@@ -91,7 +91,7 @@ fn filters_4_bits_filter_1(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Sub);
+        png.raw.filter_image(RowFilter::Sub, false);
     });
 }
 
@@ -103,7 +103,7 @@ fn filters_2_bits_filter_1(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Sub);
+        png.raw.filter_image(RowFilter::Sub, false);
     });
 }
 
@@ -115,7 +115,7 @@ fn filters_1_bits_filter_1(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Sub);
+        png.raw.filter_image(RowFilter::Sub, false);
     });
 }
 
@@ -125,7 +125,7 @@ fn filters_16_bits_filter_2(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Up);
+        png.raw.filter_image(RowFilter::Up, false);
     });
 }
 
@@ -135,7 +135,7 @@ fn filters_8_bits_filter_2(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Up);
+        png.raw.filter_image(RowFilter::Up, false);
     });
 }
 
@@ -147,7 +147,7 @@ fn filters_4_bits_filter_2(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Up);
+        png.raw.filter_image(RowFilter::Up, false);
     });
 }
 
@@ -159,7 +159,7 @@ fn filters_2_bits_filter_2(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Up);
+        png.raw.filter_image(RowFilter::Up, false);
     });
 }
 
@@ -171,7 +171,7 @@ fn filters_1_bits_filter_2(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Up);
+        png.raw.filter_image(RowFilter::Up, false);
     });
 }
 
@@ -181,7 +181,7 @@ fn filters_16_bits_filter_3(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Average);
+        png.raw.filter_image(RowFilter::Average, false);
     });
 }
 
@@ -191,7 +191,7 @@ fn filters_8_bits_filter_3(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Average);
+        png.raw.filter_image(RowFilter::Average, false);
     });
 }
 
@@ -203,7 +203,7 @@ fn filters_4_bits_filter_3(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Average);
+        png.raw.filter_image(RowFilter::Average, false);
     });
 }
 
@@ -215,7 +215,7 @@ fn filters_2_bits_filter_3(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Average);
+        png.raw.filter_image(RowFilter::Average, false);
     });
 }
 
@@ -227,7 +227,7 @@ fn filters_1_bits_filter_3(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Average);
+        png.raw.filter_image(RowFilter::Average, false);
     });
 }
 
@@ -237,7 +237,7 @@ fn filters_16_bits_filter_4(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Paeth);
+        png.raw.filter_image(RowFilter::Paeth, false);
     });
 }
 
@@ -247,7 +247,7 @@ fn filters_8_bits_filter_4(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Paeth);
+        png.raw.filter_image(RowFilter::Paeth, false);
     });
 }
 
@@ -259,7 +259,7 @@ fn filters_4_bits_filter_4(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Paeth);
+        png.raw.filter_image(RowFilter::Paeth, false);
     });
 }
 
@@ -271,7 +271,7 @@ fn filters_2_bits_filter_4(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Paeth);
+        png.raw.filter_image(RowFilter::Paeth, false);
     });
 }
 
@@ -283,7 +283,7 @@ fn filters_1_bits_filter_4(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Paeth);
+        png.raw.filter_image(RowFilter::Paeth, false);
     });
 }
 
@@ -293,7 +293,7 @@ fn filters_16_bits_filter_5(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::MinSum);
+        png.raw.filter_image(RowFilter::MinSum, false);
     });
 }
 
@@ -303,7 +303,7 @@ fn filters_8_bits_filter_5(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::MinSum);
+        png.raw.filter_image(RowFilter::MinSum, false);
     });
 }
 
@@ -315,7 +315,7 @@ fn filters_4_bits_filter_5(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::MinSum);
+        png.raw.filter_image(RowFilter::MinSum, false);
     });
 }
 
@@ -327,7 +327,7 @@ fn filters_2_bits_filter_5(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::MinSum);
+        png.raw.filter_image(RowFilter::MinSum, false);
     });
 }
 
@@ -339,6 +339,6 @@ fn filters_1_bits_filter_5(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::MinSum);
+        png.raw.filter_image(RowFilter::MinSum, false);
     });
 }

benches/strategies.rs

@@ -13,7 +13,7 @@ fn filters_minsum(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::MinSum);
+        png.raw.filter_image(RowFilter::MinSum, false);
     });
 }
 
@@ -23,7 +23,7 @@ fn filters_entropy(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Entropy);
+        png.raw.filter_image(RowFilter::Entropy, false);
     });
 }
 
@@ -33,7 +33,7 @@ fn filters_bigrams(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Bigrams);
+        png.raw.filter_image(RowFilter::Bigrams, false);
     });
 }
 
@@ -43,7 +43,7 @@ fn filters_bigent(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::BigEnt);
+        png.raw.filter_image(RowFilter::BigEnt, false);
     });
 }
 
@@ -53,6 +53,6 @@ fn filters_brute(b: &mut Bencher) {
     let png = PngData::new(&input, false).unwrap();
 
     b.iter(|| {
-        png.raw.filter_image(RowFilter::Brute);
+        png.raw.filter_image(RowFilter::Brute, false);
     });
 }

src/evaluate.rs

@@ -44,6 +44,7 @@ pub(crate) struct Evaluator {
     deadline: Arc<Deadline>,
     filters: IndexSet<RowFilter>,
     compression: u8,
+    optimize_alpha: bool,
     nth: AtomicUsize,
     best_candidate_size: Arc<AtomicMin>,
     /// images are sent to the caller thread for evaluation
@@ -55,13 +56,19 @@ pub(crate) struct Evaluator {
 }
 
 impl Evaluator {
-    pub fn new(deadline: Arc<Deadline>, filters: IndexSet<RowFilter>, compression: u8) -> Self {
+    pub fn new(
+        deadline: Arc<Deadline>,
+        filters: IndexSet<RowFilter>,
+        compression: u8,
+        optimize_alpha: bool,
+    ) -> Self {
         #[cfg(feature = "parallel")]
         let eval_channel = unbounded();
         Self {
             deadline,
             filters,
             compression,
+            optimize_alpha,
             best_candidate_size: Arc::new(AtomicMin::new(None)),
             nth: AtomicUsize::new(0),
             #[cfg(feature = "parallel")]
@@ -106,6 +113,7 @@ impl Evaluator {
         let deadline = self.deadline.clone();
         let filters = self.filters.clone();
         let compression = self.compression;
+        let optimize_alpha = self.optimize_alpha;
         let best_candidate_size = self.best_candidate_size.clone();
         // sends it off asynchronously for compression,
         // but results will be collected via the message queue
@@ -122,7 +130,7 @@ impl Evaluator {
                 if deadline.passed() {
                     return;
                 }
-                let filtered = image.filter_image(filter);
+                let filtered = image.filter_image(filter, optimize_alpha);
                 if let Ok(idat_data) =
                     deflate::deflate(&filtered, compression, &best_candidate_size)
                 {

src/filters.rs

@@ -56,9 +56,21 @@ impl RowFilter {
     pub const STANDARD: [Self; 5] = [Self::None, Self::Sub, Self::Up, Self::Average, Self::Paeth];
     pub const SINGLE_LINE: [Self; 2] = [Self::None, Self::Sub];
 
-    pub fn filter_line(self, bpp: usize, data: &[u8], last_line: &[u8], buf: &mut Vec<u8>) {
+    pub fn filter_line(
+        self,
+        bpp: usize,
+        data: &mut [u8],
+        prev_line: &[u8],
+        buf: &mut Vec<u8>,
+        alpha_bytes: usize,
+    ) {
         assert!(data.len() >= bpp);
-        assert!(last_line.is_empty() || data.len() == last_line.len());
+        assert_eq!(data.len(), prev_line.len());
+
+        if alpha_bytes != 0 {
+            self.optimize_alpha(bpp, data, prev_line, bpp - alpha_bytes);
+        }
+
         buf.clear();
         buf.reserve(data.len() + 1);
         buf.push(self as u8);
@@ -76,68 +88,101 @@ impl RowFilter {
                 );
             }
             Self::Up => {
-                if last_line.is_empty() {
-                    buf.extend_from_slice(data);
-                } else {
-                    assert_eq!(data.len(), last_line.len());
                 buf.extend(
                     data.iter()
-                            .zip(last_line.iter())
+                        .zip(prev_line.iter())
                         .map(|(cur, last)| cur.wrapping_sub(*last)),
                 );
-                };
             }
             Self::Average => {
                 for (i, byte) in data.iter().enumerate() {
-                    if last_line.is_empty() {
-                        buf.push(match i.checked_sub(bpp) {
-                            Some(x) => byte.wrapping_sub(data[x] >> 1),
-                            None => *byte,
-                        });
-                    } else {
                     buf.push(match i.checked_sub(bpp) {
                         Some(x) => byte.wrapping_sub(
-                                ((u16::from(data[x]) + u16::from(last_line[i])) >> 1) as u8,
+                            ((u16::from(data[x]) + u16::from(prev_line[i])) >> 1) as u8,
                         ),
-                            None => byte.wrapping_sub(last_line[i] >> 1),
+                        None => byte.wrapping_sub(prev_line[i] >> 1),
                     });
-                    };
                 }
             }
             Self::Paeth => {
                 for (i, byte) in data.iter().enumerate() {
-                    if last_line.is_empty() {
                     buf.push(match i.checked_sub(bpp) {
-                            Some(x) => byte.wrapping_sub(data[x]),
+                        Some(x) => {
-                            None => *byte,
+                            byte.wrapping_sub(paeth_predictor(data[x], prev_line[i], prev_line[x]))
+                        }
+                        None => byte.wrapping_sub(prev_line[i]),
                     });
-                    } else {
-                        buf.push(match i.checked_sub(bpp) {
-                            Some(x) => byte.wrapping_sub(paeth_predictor(
-                                data[x],
-                                last_line[i],
-                                last_line[x],
-                            )),
-                            None => byte.wrapping_sub(last_line[i]),
-                        });
-                    };
                 }
             }
             _ => unreachable!(),
         }
     }
 
+    // Optimize fully transparent pixels of a scanline such that they will be zeroed when filtered
+    fn optimize_alpha(self, bpp: usize, data: &mut [u8], prev_line: &[u8], color_bytes: usize) {
+        if self == Self::None {
+            // Assume transparent pixels already set to 0
+            return;
+        }
+
+        let mut pixels: Vec<_> = data.chunks_mut(bpp).collect();
+        let prev_pixels: Vec<_> = prev_line.chunks(bpp).collect();
+        for i in 0..pixels.len() {
+            if pixels[i].iter().skip(color_bytes).all(|b| *b == 0) {
+                match self {
+                    Self::Sub => {
+                        for j in 0..color_bytes {
+                            pixels[i][j] = match i {
+                                0 => 0,
+                                _ => pixels[i - 1][j],
+                            };
+                        }
+                    }
+                    Self::Up => {
+                        pixels[i][0..color_bytes].copy_from_slice(&prev_pixels[i][0..color_bytes]);
+                    }
+                    Self::Average => {
+                        for j in 0..color_bytes {
+                            pixels[i][j] = match i {
+                                0 => prev_pixels[i][j] >> 1,
+                                _ => {
+                                    ((u16::from(pixels[i - 1][j]) + u16::from(prev_pixels[i][j]))
+                                        >> 1) as u8
+                                }
+                            };
+                        }
+                    }
+                    Self::Paeth => {
+                        if i == 0 {
+                            pixels[i][0..color_bytes]
+                                .copy_from_slice(&prev_pixels[i][0..color_bytes]);
+                        } else {
+                            for j in 0..color_bytes {
+                                pixels[i][j] = paeth_predictor(
+                                    pixels[i - 1][j],
+                                    prev_pixels[i][j],
+                                    prev_pixels[i - 1][j],
+                                );
+                            }
+                        }
+                    }
+                    _ => unreachable!(),
+                }
+            }
+        }
+    }
+
     pub fn unfilter_line(
         self,
         bpp: usize,
         data: &[u8],
-        last_line: &[u8],
+        prev_line: &[u8],
         buf: &mut Vec<u8>,
     ) -> Result<(), PngError> {
         buf.clear();
         buf.reserve(data.len());
         assert!(data.len() >= bpp);
-        assert_eq!(data.len(), last_line.len());
+        assert_eq!(data.len(), prev_line.len());
         match self {
             Self::None => {
                 buf.extend_from_slice(data);
@@ -154,12 +199,12 @@ impl RowFilter {
             Self::Up => {
                 buf.extend(
                     data.iter()
-                        .zip(last_line)
+                        .zip(prev_line)
                         .map(|(&cur, &last)| cur.wrapping_add(last)),
                 );
             }
             Self::Average => {
-                for (i, (&cur, &last)) in data.iter().zip(last_line).enumerate() {
+                for (i, (&cur, &last)) in data.iter().zip(prev_line).enumerate() {
                     let prev_byte = i.checked_sub(bpp).and_then(|x| buf.get(x).copied());
                     buf.push(match prev_byte {
                         Some(b) => cur.wrapping_add(((u16::from(b) + u16::from(last)) >> 1) as u8),
@@ -168,11 +213,11 @@ impl RowFilter {
                 }
             }
             Self::Paeth => {
-                for (i, (&cur, &up)) in data.iter().zip(last_line).enumerate() {
+                for (i, (&cur, &up)) in data.iter().zip(prev_line).enumerate() {
                     buf.push(
                         match i
                             .checked_sub(bpp)
-                            .map(|x| (buf.get(x).copied(), last_line.get(x).copied()))
+                            .map(|x| (buf.get(x).copied(), prev_line.get(x).copied()))
                         {
                             Some((Some(left), Some(left_up))) => {
                                 cur.wrapping_add(paeth_predictor(left, up, left_up))

src/lib.rs

@@ -491,13 +491,25 @@ fn optimize_png(
     perform_strip(png, opts);
     let stripped_png = png.clone();
 
+    // If alpha optimization is enabled, first perform a black alpha reduction
+    // This can allow reductions from alpha to indexed which may not have been possible otherwise
+    if !opts.alphas.is_empty() {
+        if let Some(reduced) = filtered_alpha_channel(&png.raw, AlphaOptim::Black) {
+            png.raw = Arc::new(reduced);
+        }
+    }
+
     // Must use normal (lazy) compression, as faster ones (greedy) are not representative
-    // Alpha reductions can benefit from higher compression but otherwise it's not beneficial
     let eval_compression = 5;
     // None and Bigrams work well together, especially for alpha reductions
     let eval_filters = indexset! {RowFilter::None, RowFilter::Bigrams};
     // This will collect all versions of images and pick one that compresses best
-    let eval = Evaluator::new(deadline.clone(), eval_filters.clone(), eval_compression);
+    let eval = Evaluator::new(
+        deadline.clone(),
+        eval_filters.clone(),
+        eval_compression,
+        false,
+    );
     perform_reductions(png.raw.clone(), opts, &deadline, &eval);
     let (reduction_occurred, mut eval_filter) = if let Some(result) = eval.get_best_candidate() {
         *png = result.image;
@@ -519,7 +531,8 @@ fn optimize_png(
 
             if !filters.is_empty() {
                 debug!("Evaluating: {} filters", filters.len());
-                let eval = Evaluator::new(deadline, filters, eval_compression);
+                let eval =
+                    Evaluator::new(deadline, filters, eval_compression, !opts.alphas.is_empty());
                 if eval_filter.is_some() {
                     eval.set_best_size(png.idat_data.len());
                 }
@@ -585,7 +598,7 @@ fn optimize_png(
                 if deadline.passed() {
                     return None;
                 }
-                let filtered = &png.raw.filter_image(trial.filter);
+                let filtered = &png.raw.filter_image(trial.filter, !opts.alphas.is_empty());
                 perform_trial(filtered, opts, trial, &best_size)
             });
             best.reduce_with(|i, j| {
@@ -606,7 +619,7 @@ fn optimize_png(
                 opts.filter,
                 png.idat_data.len()
             );
-        } else if eval_filter.is_some() {
+        } else {
             *png = stripped_png;
         }
     } else if png.idat_data.len() >= idat_original_size {
@@ -726,13 +739,6 @@ fn perform_reductions(
     }
 
     if opts.color_type_reduction {
-        // Perform a black alpha reduction before color type reductions
-        // This can allow reductions from alpha to indexed which may not have been possible otherwise
-        if !opts.alphas.is_empty() {
-            if let Some(reduced) = filtered_alpha_channel(&png, AlphaOptim::Black) {
-                png = Arc::new(reduced);
-            }
-        }
         if let Some(reduced) = reduce_color_type(&png, opts.grayscale_reduction) {
             png = Arc::new(reduced);
             eval.try_image(png.clone());
@@ -744,10 +750,6 @@ fn perform_reductions(
         }
     }
 
-    if try_alpha_reductions(png, &opts.alphas, eval) {
-        reduction_occurred = true;
-    }
-
     if let Some(baseline) = baseline {
         if reduction_occurred {
             eval.set_baseline(baseline);

src/png/mod.rs

@@ -314,31 +314,43 @@ impl PngImage {
     }
 
     /// Apply the specified filter type to all rows in the image
-    pub fn filter_image(&self, filter: RowFilter) -> Vec<u8> {
+    pub fn filter_image(&self, filter: RowFilter, optimize_alpha: bool) -> Vec<u8> {
         let mut filtered = Vec::with_capacity(self.data.len());
         let bpp = ((self.ihdr.bit_depth.as_u8() * self.channels_per_pixel() + 7) / 8) as usize;
-        let mut last_line: &[u8] = &[];
+        // If alpha optimization is enabled, determine how many bytes of alpha there are per pixel
-        let mut last_pass: Option<u8> = None;
+        let alpha_bytes = match self.ihdr.color_type {
+            ColorType::RGBA | ColorType::GrayscaleAlpha if optimize_alpha => {
+                (self.ihdr.bit_depth.as_u8() / 8) as usize
+            }
+            _ => 0,
+        };
+
+        let mut prev_line = Vec::new();
+        let mut prev_pass: Option<u8> = None;
         let mut f_buf = Vec::new();
         for line in self.scan_lines() {
-            if last_pass != line.pass {
+            if prev_pass != line.pass || line.data.len() != prev_line.len() {
-                last_line = &[];
+                prev_line = vec![0; line.data.len()];
             }
+            // Alpha optimisation may alter the line data, so we need a mutable copy of it
+            let mut line_data = line.data.to_vec();
 
             if filter <= RowFilter::Paeth {
                 // Standard filters
-                let filter = if last_pass == line.pass || filter <= RowFilter::Sub {
+                let filter = if prev_pass == line.pass || filter <= RowFilter::Sub {
                     filter
                 } else {
                     RowFilter::None
                 };
-                filter.filter_line(bpp, line.data, last_line, &mut f_buf);
+                filter.filter_line(bpp, &mut line_data, &prev_line, &mut f_buf, alpha_bytes);
                 filtered.extend_from_slice(&f_buf);
+                prev_line = line_data;
             } else {
                 // Heuristic filter selection strategies
                 let mut best_line = Vec::new();
+                let mut best_line_raw = Vec::new();
                 // Avoid vertical filtering on first line of each interlacing pass
-                let try_filters = if last_pass == line.pass {
+                let try_filters = if prev_pass == line.pass {
                     RowFilter::STANDARD.iter()
                 } else {
                     RowFilter::SINGLE_LINE.iter()
@@ -348,8 +360,8 @@ impl PngImage {
                         // MSAD algorithm mentioned in libpng reference docs
                         // http://www.libpng.org/pub/png/book/chapter09.html
                         let mut best_size = usize::MAX;
-                        for try_filter in try_filters {
+                        for f in try_filters {
-                            try_filter.filter_line(bpp, line.data, last_line, &mut f_buf);
+                            f.filter_line(bpp, &mut line_data, &prev_line, &mut f_buf, alpha_bytes);
                             let size = f_buf.iter().fold(0, |acc, &x| {
                                 let signed = x as i8;
                                 acc + signed.unsigned_abs() as usize
@@ -357,6 +369,7 @@ impl PngImage {
                             if size < best_size {
                                 best_size = size;
                                 std::mem::swap(&mut best_line, &mut f_buf);
+                                best_line_raw = line_data.clone();
                             }
                         }
                     }
@@ -364,8 +377,8 @@ impl PngImage {
                         // Shannon entropy algorithm, from LodePNG
                         // https://github.com/lvandeve/lodepng
                         let mut best_size = i32::MIN;
-                        for try_filter in try_filters {
+                        for f in try_filters {
-                            try_filter.filter_line(bpp, line.data, last_line, &mut f_buf);
+                            f.filter_line(bpp, &mut line_data, &prev_line, &mut f_buf, alpha_bytes);
                             let mut counts = vec![0; 0x100];
                             for &i in f_buf.iter() {
                                 counts[i as usize] += 1;
@@ -379,6 +392,7 @@ impl PngImage {
                             if size > best_size {
                                 best_size = size;
                                 std::mem::swap(&mut best_line, &mut f_buf);
+                                best_line_raw = line_data.clone();
                             }
                         }
                     }
@@ -386,8 +400,8 @@ impl PngImage {
                         // Count distinct bigrams, from pngwolf
                         // https://bjoern.hoehrmann.de/pngwolf/
                         let mut best_size = usize::MAX;
-                        for try_filter in try_filters {
+                        for f in try_filters {
-                            try_filter.filter_line(bpp, line.data, last_line, &mut f_buf);
+                            f.filter_line(bpp, &mut line_data, &prev_line, &mut f_buf, alpha_bytes);
                             let mut set = bitarr![0; 0x10000];
                             for pair in f_buf.windows(2) {
                                 let bigram = (pair[0] as usize) << 8 | pair[1] as usize;
@@ -397,6 +411,7 @@ impl PngImage {
                             if size < best_size {
                                 best_size = size;
                                 std::mem::swap(&mut best_line, &mut f_buf);
+                                best_line_raw = line_data.clone();
                             }
                         }
                     }
@@ -405,8 +420,8 @@ impl PngImage {
                         let mut best_size = i32::MIN;
                         // FxHasher is the fastest rust hasher currently available for this purpose
                         let mut counts = FxHashMap::<u16, u32>::default();
-                        for try_filter in try_filters {
+                        for f in try_filters {
-                            try_filter.filter_line(bpp, line.data, last_line, &mut f_buf);
+                            f.filter_line(bpp, &mut line_data, &prev_line, &mut f_buf, alpha_bytes);
                             counts.clear();
                             for pair in f_buf.windows(2) {
                                 let bigram = (pair[0] as u16) << 8 | pair[1] as u16;
@@ -416,6 +431,7 @@ impl PngImage {
                             if size > best_size {
                                 best_size = size;
                                 std::mem::swap(&mut best_line, &mut f_buf);
+                                best_line_raw = line_data.clone();
                             }
                         }
                     }
@@ -431,8 +447,8 @@ impl PngImage {
                         let capacity = compressor.zlib_compress_bound(limit);
                         let mut dest = vec![0; capacity];
 
-                        for try_filter in try_filters {
+                        for f in try_filters {
-                            try_filter.filter_line(bpp, line.data, last_line, &mut f_buf);
+                            f.filter_line(bpp, &mut line_data, &prev_line, &mut f_buf, alpha_bytes);
                             filtered[line_start..].copy_from_slice(&f_buf);
                             let size = compressor
                                 .zlib_compress(&filtered[filtered.len() - limit..], &mut dest)
@@ -440,6 +456,7 @@ impl PngImage {
                             if size < best_size {
                                 best_size = size;
                                 std::mem::swap(&mut best_line, &mut f_buf);
+                                best_line_raw = line_data.clone();
                             }
                         }
                         filtered.resize(line_start, 0);
@@ -447,10 +464,10 @@ impl PngImage {
                     _ => unreachable!(),
                 }
                 filtered.extend_from_slice(&best_line);
+                prev_line = best_line_raw;
             }
 
-            last_line = line.data;
+            prev_pass = line.pass;
-            last_pass = line.pass;
         }
         filtered
     }

src/reduction/alpha.rs

@@ -1,32 +1,7 @@
 use crate::colors::AlphaOptim;
 use crate::colors::ColorType;
-use crate::evaluate::Evaluator;
 use crate::headers::IhdrData;
 use crate::png::PngImage;
-#[cfg(not(feature = "parallel"))]
-use crate::rayon::prelude::*;
-use indexmap::IndexSet;
-#[cfg(feature = "parallel")]
-use rayon::prelude::*;
-use std::sync::Arc;
-
-pub(crate) fn try_alpha_reductions(
-    png: Arc<PngImage>,
-    alphas: &IndexSet<AlphaOptim>,
-    eval: &Evaluator,
-) -> bool {
-    match png.ihdr.color_type {
-        ColorType::RGBA | ColorType::GrayscaleAlpha if !alphas.is_empty() => {
-            alphas
-                .par_iter()
-                .with_max_len(1)
-                .filter_map(|&alpha| filtered_alpha_channel(&png, alpha))
-                .for_each(|image| eval.try_image(Arc::new(image)));
-            true
-        }
-        _ => false,
-    }
-}
 
 pub fn filtered_alpha_channel(png: &PngImage, optim: AlphaOptim) -> Option<PngImage> {
     let (bpc, bpp) = match png.ihdr.color_type {

src/reduction/mod.rs

@@ -12,7 +12,7 @@ use crate::bit_depth::reduce_bit_depth_8_or_less;
 pub mod color;
 use crate::color::*;
 
-pub(crate) use crate::alpha::{filtered_alpha_channel, try_alpha_reductions};
+pub(crate) use crate::alpha::filtered_alpha_channel;
 pub(crate) use crate::bit_depth::reduce_bit_depth;
 
 /// Attempt to reduce the number of colors in the palette