Merge pull request #379 from lilith/main

Optimize horizontal and float predictor decode (4-43% faster)

.gitignore

@@ -2,3 +2,4 @@
 /target/
 **/*.rs.bk
 Cargo.lock
+tests/benches/corpus/

Cargo.toml

@@ -54,6 +54,10 @@ harness = false
 name = "encode"
 harness = false
 
+[[bench]]
+name = "decode"
+harness = false
+
 
 [[test]]
 name = "decode_libtiffpic"

benches/decode.rs

@@ -0,0 +1,461 @@
+extern crate criterion;
+extern crate tiff;
+
+use criterion::{black_box, Criterion, Throughput};
+use std::io::Cursor;
+use tiff::decoder::Decoder;
+use tiff::encoder::colortype::{Gray16, Gray8, RGB32Float, RGB16, RGB8, RGBA8};
+use tiff::encoder::{Compression, Predictor, TiffEncoder};
+
+const WIDTH: u32 = 2048;
+const HEIGHT: u32 = 2048;
+
+// === Corpus paths (real photo-based TIFFs) ===
+// Set TIFF_BENCH_DIR / TIFF_GEOTIFF_DIR env vars to a directory containing the corpus files.
+// If the files are not found, the corresponding benchmarks are silently skipped.
+// The corpus can be generated with vips/imagemagick from any large photo; see the benchmark
+// names for the expected filenames and compression settings.
+const CORPUS_DIR: &str = "tests/benches/corpus";
+
+fn bench_dir() -> String {
+    std::env::var("TIFF_BENCH_DIR").unwrap_or_else(|_| CORPUS_DIR.to_string())
+}
+
+fn geotiff_dir() -> String {
+    std::env::var("TIFF_GEOTIFF_DIR").unwrap_or_else(|_| CORPUS_DIR.to_string())
+}
+
+/// Load a file from the corpus directory, returning None if not found.
+fn load_corpus(dir: &str, name: &str) -> Option<Vec<u8>> {
+    let path = format!("{}/{}", dir, name);
+    std::fs::read(&path).ok()
+}
+
+fn decode_tiff_bytes(data: &[u8]) {
+    let cursor = Cursor::new(data);
+    let mut decoder = Decoder::open(black_box(cursor)).unwrap();
+    decoder.next_image().unwrap();
+    decoder.read_image().unwrap();
+}
+
+// === Synthetic image generators ===
+
+fn generate_rgb8(w: u32, h: u32) -> Vec<u8> {
+    (0..w * h * 3)
+        .map(|i| ((i * 7 + i / w * 3) & 0xFF) as u8)
+        .collect()
+}
+
+fn generate_rgba8(w: u32, h: u32) -> Vec<u8> {
+    (0..w * h * 4)
+        .map(|i| ((i * 11 + i / w * 5) & 0xFF) as u8)
+        .collect()
+}
+
+fn generate_gray8(w: u32, h: u32) -> Vec<u8> {
+    (0..w * h)
+        .map(|i| ((i * 7 + i / w * 3) & 0xFF) as u8)
+        .collect()
+}
+
+fn generate_rgb16(w: u32, h: u32) -> Vec<u16> {
+    (0..w * h * 3)
+        .map(|i| ((i * 257 + i / w * 131) & 0xFFFF) as u16)
+        .collect()
+}
+
+fn generate_gray16(w: u32, h: u32) -> Vec<u16> {
+    (0..w * h)
+        .map(|i| ((i * 257 + i / w * 131) & 0xFFFF) as u16)
+        .collect()
+}
+
+fn generate_rgb_f32(w: u32, h: u32) -> Vec<f32> {
+    let mut data = Vec::with_capacity((w * h * 3) as usize);
+    for y in 0..h {
+        for x in 0..w {
+            let fx = x as f32 / w as f32;
+            let fy = y as f32 / h as f32;
+            data.push(fx * 0.5 + fy * 0.3);
+            data.push(fy * 0.7 + fx * 0.2);
+            data.push((fx + fy) * 0.25);
+        }
+    }
+    data
+}
+
+// === Synthetic encoders (strip-only, encoder doesn't support tiles) ===
+
+fn encode<C: tiff::encoder::colortype::ColorType>(
+    data: &[C::Inner],
+    w: u32,
+    h: u32,
+    compression: Compression,
+    predictor: Predictor,
+) -> Vec<u8>
+where
+    [C::Inner]: tiff::encoder::TiffValue,
+{
+    let mut buf = Cursor::new(Vec::new());
+    TiffEncoder::new(&mut buf)
+        .unwrap()
+        .with_predictor(predictor)
+        .with_compression(compression)
+        .write_image::<C>(w, h, data)
+        .unwrap();
+    buf.into_inner()
+}
+
+fn main() {
+    let mut c = Criterion::default().configure_from_args();
+    let bd = bench_dir();
+    let gd = geotiff_dir();
+
+    // ========================================================================
+    // PART 1: Real photo corpus (5616x3744, ~21MP)
+    // ========================================================================
+    // Raw pixel sizes for throughput calculation:
+    //   RGB8:  5616*3744*3 = 63,078,912
+    //   RGB16: 5616*3744*6 = 126,157,824
+    //   F32x3: 5616*3744*12 = 252,315,648
+    //   Gray8: 5616*3744*1 = 21,026,304
+    //   GrayF32: 5616*3744*4 = 84,105,216
+    //   Palette8: 5616*3744*1 = 21,026,304 (output)
+
+    struct CorpusBench {
+        file: &'static str,
+        name: &'static str,
+        raw_bytes: u64,
+    }
+
+    let photo_benches = [
+        // Strip-based
+        CorpusBench {
+            file: "photo-rgb8-none.tif",
+            name: "strip-rgb8-none",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-rgb8-packbits.tif",
+            name: "strip-rgb8-packbits",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-rgb8-lzw.tif",
+            name: "strip-rgb8-lzw",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-rgb8-lzw-hpred.tif",
+            name: "strip-rgb8-lzw-hpred",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-rgb8-deflate-hpred.tif",
+            name: "strip-rgb8-deflate-hpred",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-rgb16-deflate-hpred.tif",
+            name: "strip-rgb16-deflate-hpred",
+            raw_bytes: 126_157_824,
+        },
+        CorpusBench {
+            file: "photo-gray8-lzw.tif",
+            name: "strip-gray8-lzw",
+            raw_bytes: 21_026_304,
+        },
+        CorpusBench {
+            file: "photo-gray8-deflate-hpred.tif",
+            name: "strip-gray8-deflate-hpred",
+            raw_bytes: 21_026_304,
+        },
+        CorpusBench {
+            file: "photo-rgb-f32-deflate-fpred.tif",
+            name: "strip-rgb-f32-deflate-fpred",
+            raw_bytes: 252_315_648,
+        },
+        CorpusBench {
+            file: "photo-gray-f32-deflate-fpred.tif",
+            name: "strip-gray-f32-deflate-fpred",
+            raw_bytes: 84_105_216,
+        },
+        CorpusBench {
+            file: "photo-palette8-lzw.tif",
+            name: "strip-palette8-lzw",
+            raw_bytes: 21_026_304,
+        },
+        CorpusBench {
+            file: "scan-bilevel-fax4.tif",
+            name: "strip-bilevel-fax4",
+            raw_bytes: 5616 * 3744 / 8,
+        },
+        // Tiled
+        CorpusBench {
+            file: "photo-rgb8-lzw-hpred-tiled256.tif",
+            name: "tiled256-rgb8-lzw-hpred",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-rgb8-deflate-hpred-tiled256.tif",
+            name: "tiled256-rgb8-deflate-hpred",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-rgb8-jpeg-tiled256.tif",
+            name: "tiled256-rgb8-jpeg",
+            raw_bytes: 63_078_912,
+        },
+        CorpusBench {
+            file: "photo-gray-f32-deflate-fpred-tiled512.tif",
+            name: "tiled512-gray-f32-deflate-fpred",
+            raw_bytes: 84_105_216,
+        },
+    ];
+
+    {
+        let mut group = c.benchmark_group("photo-21mp");
+        group.sample_size(10);
+
+        for bench in &photo_benches {
+            if let Some(data) = load_corpus(&bd, bench.file) {
+                group.throughput(Throughput::Bytes(bench.raw_bytes));
+                let data_ref = &data;
+                group.bench_function(bench.name, |b| b.iter(|| decode_tiff_bytes(data_ref)));
+            }
+        }
+
+        group.finish();
+    }
+
+    // ========================================================================
+    // PART 2: GeoTIFF (real tiled DEM data)
+    // ========================================================================
+
+    let geotiff_benches = [
+        // Copernicus DEM: 3600x3600, tiled 1024x1024, Gray f32, deflate + fpredict
+        CorpusBench {
+            file: "Copernicus_DSM_COG_10_N40_00_W075_00_DEM.tif",
+            name: "copernicus-dem-f32-deflate-fpred-1024tile",
+            raw_bytes: 3600 * 3600 * 4,
+        },
+        // USGS DEM: 3612x3612, tiled 512x512, Gray f32, LZW + fpredict
+        CorpusBench {
+            file: "USGS_1_n40w075.tif",
+            name: "usgs-dem-f32-lzw-fpred-512tile",
+            raw_bytes: 3612 * 3612 * 4,
+        },
+    ];
+
+    {
+        let mut group = c.benchmark_group("geotiff");
+        group.sample_size(10);
+
+        for bench in &geotiff_benches {
+            if let Some(data) = load_corpus(&gd, bench.file) {
+                group.throughput(Throughput::Bytes(bench.raw_bytes));
+                let data_ref = &data;
+                group.bench_function(bench.name, |b| b.iter(|| decode_tiff_bytes(data_ref)));
+            }
+        }
+
+        group.finish();
+    }
+
+    // ========================================================================
+    // PART 3: Synthetic (2048x2048, controlled comparisons)
+    // ========================================================================
+
+    let rgb8 = generate_rgb8(WIDTH, HEIGHT);
+    let rgba8 = generate_rgba8(WIDTH, HEIGHT);
+    let gray8 = generate_gray8(WIDTH, HEIGHT);
+    let rgb16 = generate_rgb16(WIDTH, HEIGHT);
+    let gray16 = generate_gray16(WIDTH, HEIGHT);
+    let rgb_f32 = generate_rgb_f32(WIDTH, HEIGHT);
+
+    let raw_rgb8 = (WIDTH * HEIGHT * 3) as u64;
+    let raw_rgba8 = (WIDTH * HEIGHT * 4) as u64;
+    let raw_gray8 = (WIDTH * HEIGHT) as u64;
+    let raw_rgb16 = (WIDTH * HEIGHT * 6) as u64;
+    let raw_gray16 = (WIDTH * HEIGHT * 2) as u64;
+    let raw_f32x3 = (WIDTH * HEIGHT * 12) as u64;
+
+    // --- Uncompressed baseline ---
+    {
+        let mut group = c.benchmark_group("synth-uncompressed");
+        group.sample_size(10);
+
+        let t = encode::<RGB8>(
+            &rgb8,
+            WIDTH,
+            HEIGHT,
+            Compression::Uncompressed,
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb8))
+            .bench_function("rgb8", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGBA8>(
+            &rgba8,
+            WIDTH,
+            HEIGHT,
+            Compression::Uncompressed,
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgba8))
+            .bench_function("rgba8", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<Gray8>(
+            &gray8,
+            WIDTH,
+            HEIGHT,
+            Compression::Uncompressed,
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_gray8))
+            .bench_function("gray8", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB16>(
+            &rgb16,
+            WIDTH,
+            HEIGHT,
+            Compression::Uncompressed,
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb16))
+            .bench_function("rgb16", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<Gray16>(
+            &gray16,
+            WIDTH,
+            HEIGHT,
+            Compression::Uncompressed,
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_gray16))
+            .bench_function("gray16", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        group.finish();
+    }
+
+    // --- Predictor overhead isolation (deflate-1, with vs without predictor) ---
+    {
+        let mut group = c.benchmark_group("synth-predictor");
+        group.sample_size(10);
+
+        let t = encode::<RGB8>(
+            &rgb8,
+            WIDTH,
+            HEIGHT,
+            Compression::Deflate(1),
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb8))
+            .bench_function("rgb8-deflate-none", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB8>(
+            &rgb8,
+            WIDTH,
+            HEIGHT,
+            Compression::Deflate(1),
+            Predictor::Horizontal,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb8))
+            .bench_function("rgb8-deflate-hpred", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB16>(
+            &rgb16,
+            WIDTH,
+            HEIGHT,
+            Compression::Deflate(1),
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb16))
+            .bench_function("rgb16-deflate-none", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB16>(
+            &rgb16,
+            WIDTH,
+            HEIGHT,
+            Compression::Deflate(1),
+            Predictor::Horizontal,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb16))
+            .bench_function("rgb16-deflate-hpred", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB32Float>(
+            &rgb_f32,
+            WIDTH,
+            HEIGHT,
+            Compression::Deflate(1),
+            Predictor::None,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_f32x3))
+            .bench_function("rgb-f32-deflate-none", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB32Float>(
+            &rgb_f32,
+            WIDTH,
+            HEIGHT,
+            Compression::Deflate(1),
+            Predictor::FloatingPoint,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_f32x3))
+            .bench_function("rgb-f32-deflate-fpred", |b| {
+                b.iter(|| decode_tiff_bytes(&t))
+            });
+
+        group.finish();
+    }
+
+    // --- LZW vs Deflate comparison ---
+    {
+        let mut group = c.benchmark_group("synth-compression");
+        group.sample_size(10);
+
+        let t = encode::<RGB8>(
+            &rgb8,
+            WIDTH,
+            HEIGHT,
+            Compression::Deflate(1),
+            Predictor::Horizontal,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb8))
+            .bench_function("rgb8-deflate-hpred", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB8>(
+            &rgb8,
+            WIDTH,
+            HEIGHT,
+            Compression::Lzw,
+            Predictor::Horizontal,
+        );
+        group
+            .throughput(Throughput::Bytes(raw_rgb8))
+            .bench_function("rgb8-lzw-hpred", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB8>(&rgb8, WIDTH, HEIGHT, Compression::Lzw, Predictor::None);
+        group
+            .throughput(Throughput::Bytes(raw_rgb8))
+            .bench_function("rgb8-lzw-none", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        let t = encode::<RGB8>(&rgb8, WIDTH, HEIGHT, Compression::Packbits, Predictor::None);
+        group
+            .throughput(Throughput::Bytes(raw_rgb8))
+            .bench_function("rgb8-packbits", |b| b.iter(|| decode_tiff_bytes(&t)));
+
+        group.finish();
+    }
+}

examples/profile_decode.rs

@@ -0,0 +1,130 @@
+//! Profiling helper: encodes and decodes several TIFF configurations.
+//! Run under callgrind: `valgrind --tool=callgrind target/release/examples/profile_decode`
+use std::io::Cursor;
+use tiff::decoder::Decoder;
+use tiff::encoder::colortype::{Gray8, RGB32Float, RGB16, RGB8};
+use tiff::encoder::{Compression, Predictor, TiffEncoder};
+
+const WIDTH: u32 = 2048;
+const HEIGHT: u32 = 2048;
+const ITERS: usize = 3;
+
+fn generate_rgb8(w: u32, h: u32) -> Vec<u8> {
+    (0..w * h * 3).map(|i| (i % 256) as u8).collect()
+}
+
+fn generate_rgb16(w: u32, h: u32) -> Vec<u16> {
+    (0..w * h * 3).map(|i| (i % 65536) as u16).collect()
+}
+
+fn generate_rgb_f32(w: u32, h: u32) -> Vec<f32> {
+    (0..w * h * 3)
+        .map(|i| (i as f32) / (w * h * 3) as f32)
+        .collect()
+}
+
+fn generate_gray8(w: u32, h: u32) -> Vec<u8> {
+    (0..w * h).map(|i| (i % 256) as u8).collect()
+}
+
+fn encode<C: tiff::encoder::colortype::ColorType>(
+    data: &[C::Inner],
+    w: u32,
+    h: u32,
+    compression: Compression,
+    predictor: Predictor,
+) -> Vec<u8>
+where
+    [C::Inner]: tiff::encoder::TiffValue,
+{
+    let mut buf = Cursor::new(Vec::new());
+    TiffEncoder::new(&mut buf)
+        .unwrap()
+        .with_predictor(predictor)
+        .with_compression(compression)
+        .write_image::<C>(w, h, data)
+        .unwrap();
+    buf.into_inner()
+}
+
+fn decode(data: &[u8]) {
+    let cursor = Cursor::new(data);
+    let mut decoder = Decoder::open(cursor).unwrap();
+    decoder.next_image().unwrap();
+    let _ = decoder.read_image().unwrap();
+}
+
+fn main() {
+    let rgb8 = generate_rgb8(WIDTH, HEIGHT);
+    let rgb16 = generate_rgb16(WIDTH, HEIGHT);
+    let rgb_f32 = generate_rgb_f32(WIDTH, HEIGHT);
+    let gray8 = generate_gray8(WIDTH, HEIGHT);
+
+    let cases: Vec<(&str, Vec<u8>)> = vec![
+        (
+            "deflate-rgb8-hpredict",
+            encode::<RGB8>(
+                &rgb8,
+                WIDTH,
+                HEIGHT,
+                Compression::Deflate(1),
+                Predictor::Horizontal,
+            ),
+        ),
+        (
+            "deflate-gray8-hpredict",
+            encode::<Gray8>(
+                &gray8,
+                WIDTH,
+                HEIGHT,
+                Compression::Deflate(1),
+                Predictor::Horizontal,
+            ),
+        ),
+        (
+            "deflate-rgb16-hpredict",
+            encode::<RGB16>(
+                &rgb16,
+                WIDTH,
+                HEIGHT,
+                Compression::Deflate(1),
+                Predictor::Horizontal,
+            ),
+        ),
+        (
+            "deflate-rgb-f32-fpredict",
+            encode::<RGB32Float>(
+                &rgb_f32,
+                WIDTH,
+                HEIGHT,
+                Compression::Deflate(1),
+                Predictor::FloatingPoint,
+            ),
+        ),
+        (
+            "lzw-rgb8-no-pred",
+            encode::<RGB8>(&rgb8, WIDTH, HEIGHT, Compression::Lzw, Predictor::None),
+        ),
+        (
+            "lzw-rgb8-hpredict",
+            encode::<RGB8>(
+                &rgb8,
+                WIDTH,
+                HEIGHT,
+                Compression::Lzw,
+                Predictor::Horizontal,
+            ),
+        ),
+    ];
+
+    for (name, tiff_data) in &cases {
+        eprintln!(
+            "=== Profiling: {} ({} bytes encoded) ===",
+            name,
+            tiff_data.len()
+        );
+        for _ in 0..ITERS {
+            decode(tiff_data);
+        }
+    }
+}

src/decoder/image.rs

@@ -1009,6 +1009,7 @@ impl Image {
         buf: &mut [u8],
         layout: &ReadoutLayout,
         chunk_index: u32,
+        scratch: &mut Vec<u8>,
     ) -> TiffResult<()> {
         let ValueReader {
             reader,
@@ -1191,6 +1192,7 @@ impl Image {
                     samples,
                     ByteOrder::native(),
                     predictor,
+                    scratch,
                 );
 
                 if photometric_interpretation == PhotometricInterpretation::WhiteIsZero {
@@ -1209,6 +1211,7 @@ impl Image {
                     samples,
                     decompressed_byte_order,
                     predictor,
+                    scratch,
                 );
             }
 
@@ -1255,6 +1258,7 @@ impl Image {
                     samples,
                     decompressed_byte_order,
                     predictor,
+                    scratch,
                 );
 
                 if photometric_interpretation == PhotometricInterpretation::WhiteIsZero {
@@ -1296,6 +1300,7 @@ impl Image {
                     samples,
                     decompressed_byte_order,
                     predictor,
+                    scratch,
                 );
 
                 if photometric_interpretation == PhotometricInterpretation::WhiteIsZero {

src/decoder/mod.rs

@@ -254,6 +254,9 @@ pub struct Decoder<R> {
     /// Map from the ifd into the `ifd_offsets` ordered list.
     seen_ifds: cycles::IfdCycles,
     image: Option<Image>,
+    /// Reusable scratch buffer for the floating-point predictor. Stored here so that the
+    /// allocation persists across strips/tiles and images, avoiding per-row heap allocation.
+    scratch: Vec<u8>,
 }
 
 /// All the information needed to read and interpret byte slices from the underlying file, i.e. to
@@ -273,13 +276,37 @@ pub struct IfdDecoder<'lt> {
     inner: tag_reader::TagReader<'lt, dyn tag_reader::EntryDecoder + 'lt>,
 }
 
-fn rev_hpredict_nsamp(buf: &mut [u8], bit_depth: u8, samples: u16) {
-    fn one_byte_predict<const N: usize>(buf: &mut [u8]) {
-        for i in N..buf.len() {
-            buf[i] = buf[i].wrapping_add(buf[i - N]);
+/// Byte-level prefix sum with compile-time stride. Each element is the wrapping sum of itself and
+/// the element N positions earlier. For N >= 2, the N interleaved channels are independent streams.
+///
+/// For N=3 (RGB) and N=4 (RGBA), we use an explicit `[u8; N]` accumulator so LLVM sees each
+/// channel as an independent scalar chain and can unroll effectively. Without this, LLVM emits a
+/// scalar byte-at-a-time loop for N=3 because 3 doesn't map cleanly to SIMD widths.
+fn one_byte_predict<const N: usize>(buf: &mut [u8]) {
+    if N == 0 || buf.len() <= N {
+        return;
+    }
+
+    // Copy the first N bytes as the initial accumulator state.
+    let mut acc = [0u8; N];
+    acc.copy_from_slice(&buf[..N]);
+
+    // Process N bytes per iteration. Each channel's prefix sum is independent.
+    let mut chunks = buf[N..].chunks_exact_mut(N);
+    for chunk in &mut chunks {
+        for c in 0..N {
+            acc[c] = chunk[c].wrapping_add(acc[c]);
+            chunk[c] = acc[c];
         }
     }
 
+    // Handle trailing bytes (when buf.len() - N is not a multiple of N).
+    for (c, byte) in chunks.into_remainder().iter_mut().enumerate() {
+        *byte = byte.wrapping_add(acc[c]);
+    }
+}
+
+fn rev_hpredict_nsamp(buf: &mut [u8], bit_depth: u8, samples: u16) {
     fn two_bytes_predict<const N: usize>(buf: &mut [u8]) {
         for i in (2 * N..buf.len()).step_by(2) {
             let v = u16::from_ne_bytes(buf[i..][..2].try_into().unwrap());
@@ -364,65 +391,150 @@ fn rev_hpredict_nsamp(buf: &mut [u8], bit_depth: u8, samples: u16) {
     }
 }
 
+/// Undo the floating-point predictor: byte-level prefix sum, then de-interleave byte planes back
+/// into native-endian samples.
+///
+/// The TIFF float predictor splits each N-byte sample into N separate byte planes within the row,
+/// then delta-encodes each plane. Decoding reverses this: prefix-sum to undo the deltas, then
+/// transpose the N planes back into packed samples.
 fn predict_f32(input: &mut [u8], output: &mut [u8], samples: u16) {
-    let samples = usize::from(samples);
-
-    for i in samples..input.len() {
-        input[i] = input[i].wrapping_add(input[i - samples]);
-    }
-
-    for (i, chunk) in output.chunks_mut(4).enumerate() {
-        chunk.copy_from_slice(&u32::to_ne_bytes(u32::from_be_bytes([
-            input[i],
-            input[input.len() / 4 + i],
-            input[input.len() / 4 * 2 + i],
-            input[input.len() / 4 * 3 + i],
-        ])));
-    }
+    one_byte_predict_generic(input, usize::from(samples));
+    deinterleave_f32_planes(input, output);
 }
 
 fn predict_f16(input: &mut [u8], output: &mut [u8], samples: u16) {
-    let samples = usize::from(samples);
-
-    for i in samples..input.len() {
-        input[i] = input[i].wrapping_add(input[i - samples]);
-    }
-
-    for (i, chunk) in output.chunks_mut(2).enumerate() {
-        chunk.copy_from_slice(&u16::to_ne_bytes(u16::from_be_bytes([
-            input[i],
-            input[input.len() / 2 + i],
-        ])));
-    }
+    one_byte_predict_generic(input, usize::from(samples));
+    deinterleave_f16_planes(input, output);
 }
 
 fn predict_f64(input: &mut [u8], output: &mut [u8], samples: u16) {
-    let samples = usize::from(samples);
+    one_byte_predict_generic(input, usize::from(samples));
+    deinterleave_f64_planes(input, output);
+}
 
-    for i in samples..input.len() {
-        input[i] = input[i].wrapping_add(input[i - samples]);
+/// Byte-level prefix sum: `buf[i] += buf[i - stride]` for all `i >= stride`.
+///
+/// Same operation as `one_byte_predict<N>` but dispatched at runtime. The most common float
+/// channel counts (1, 3, 4) get dedicated loops so LLVM can see the independent channel streams
+/// and vectorize each one separately.
+fn one_byte_predict_generic(buf: &mut [u8], stride: usize) {
+    match stride {
+        1 => one_byte_predict::<1>(buf),
+        2 => one_byte_predict::<2>(buf),
+        3 => one_byte_predict::<3>(buf),
+        4 => one_byte_predict::<4>(buf),
+        _ => {
+            for i in stride..buf.len() {
+                buf[i] = buf[i].wrapping_add(buf[i - stride]);
+            }
         }
-
-    for (i, chunk) in output.chunks_mut(8).enumerate() {
-        chunk.copy_from_slice(&u64::to_ne_bytes(u64::from_be_bytes([
-            input[i],
-            input[input.len() / 8 + i],
-            input[input.len() / 8 * 2 + i],
-            input[input.len() / 8 * 3 + i],
-            input[input.len() / 8 * 4 + i],
-            input[input.len() / 8 * 5 + i],
-            input[input.len() / 8 * 6 + i],
-            input[input.len() / 8 * 7 + i],
-        ])));
     }
 }
 
+/// Transpose byte-planes back into packed f16/f32/f64 samples with big-endian to native conversion.
+///
+/// Input layout: `[plane0: n bytes][plane1: n bytes]...[planeP: n bytes]` where each plane holds
+/// one byte position of each sample. Output: packed native-endian samples.
+///
+/// Each function splits the input into equal-sized plane slices, then gathers one byte per plane
+/// per sample. Using `from_be_bytes` for endian conversion lets LLVM recognise the byte-swap idiom
+/// and emit efficient code (bswap or movbe).
+fn deinterleave_f32_planes(input: &[u8], output: &mut [u8]) {
+    let n = output.len() / 4;
+    if n == 0 || input.len() < n * 4 {
+        return;
+    }
+    let (p0, rest) = input.split_at(n);
+    let (p1, rest) = rest.split_at(n);
+    let (p2, p3) = rest.split_at(n);
+
+    // Batch 4 samples (16 output bytes) per iteration. Fixed-size array slicing eliminates
+    // bounds checks in the inner loop.
+    let (out_chunks, out_rest) = output.as_chunks_mut::<16>();
+    let (p0_chunks, p0_rest) = p0.as_chunks::<4>();
+    let (p1_chunks, p1_rest) = p1.as_chunks::<4>();
+    let (p2_chunks, p2_rest) = p2.as_chunks::<4>();
+    let (p3_chunks, p3_rest) = p3.as_chunks::<4>();
+
+    for ((((out, b0), b1), b2), b3) in out_chunks
+        .iter_mut()
+        .zip(p0_chunks)
+        .zip(p1_chunks)
+        .zip(p2_chunks)
+        .zip(p3_chunks)
+    {
+        let s0 = u32::from_be_bytes([b0[0], b1[0], b2[0], b3[0]]).to_ne_bytes();
+        let s1 = u32::from_be_bytes([b0[1], b1[1], b2[1], b3[1]]).to_ne_bytes();
+        let s2 = u32::from_be_bytes([b0[2], b1[2], b2[2], b3[2]]).to_ne_bytes();
+        let s3 = u32::from_be_bytes([b0[3], b1[3], b2[3], b3[3]]).to_ne_bytes();
+        *out = [
+            s0[0], s0[1], s0[2], s0[3], s1[0], s1[1], s1[2], s1[3], s2[0], s2[1], s2[2], s2[3],
+            s3[0], s3[1], s3[2], s3[3],
+        ];
+    }
+
+    // Remaining 0–3 samples.
+    for (j, out) in out_rest.chunks_exact_mut(4).enumerate() {
+        out.copy_from_slice(
+            &u32::from_be_bytes([p0_rest[j], p1_rest[j], p2_rest[j], p3_rest[j]]).to_ne_bytes(),
+        );
+    }
+}
+
+fn deinterleave_f16_planes(input: &[u8], output: &mut [u8]) {
+    let n = output.len() / 2;
+    if n == 0 || input.len() < n * 2 {
+        return;
+    }
+    let (p0, p1) = input.split_at(n);
+
+    for (out, (b0, b1)) in output.chunks_exact_mut(2).zip(p0.iter().zip(p1)) {
+        out.copy_from_slice(&u16::from_be_bytes([*b0, *b1]).to_ne_bytes());
+    }
+}
+
+fn deinterleave_f64_planes(input: &[u8], output: &mut [u8]) {
+    let n = output.len() / 8;
+    if n == 0 || input.len() < n * 8 {
+        return;
+    }
+    let (p0, rest) = input.split_at(n);
+    let (p1, rest) = rest.split_at(n);
+    let (p2, rest) = rest.split_at(n);
+    let (p3, rest) = rest.split_at(n);
+    let (p4, rest) = rest.split_at(n);
+    let (p5, rest) = rest.split_at(n);
+    let (p6, p7) = rest.split_at(n);
+
+    // Zip all 8 planes with the output to avoid indexed access and bounds checks.
+    let iter = output.chunks_exact_mut(8).zip(
+        p0.iter()
+            .zip(p1)
+            .zip(p2)
+            .zip(p3)
+            .zip(p4)
+            .zip(p5)
+            .zip(p6)
+            .zip(p7),
+    );
+    for (out, (((((((b0, b1), b2), b3), b4), b5), b6), b7)) in iter {
+        out.copy_from_slice(
+            &u64::from_be_bytes([*b0, *b1, *b2, *b3, *b4, *b5, *b6, *b7]).to_ne_bytes(),
+        );
+    }
+}
+
+/// Fix byte order and apply the inverse predictor for one row of decoded data.
+///
+/// `scratch` is a reusable buffer for the floating-point predictor (which needs a working copy of
+/// the input). Callers should pass the same buffer across rows to avoid per-row heap allocation.
 fn fix_endianness_and_predict(
     buf: &mut [u8],
     bit_depth: u8,
     samples: u16,
     byte_order: ByteOrder,
     predictor: Predictor,
+    scratch: &mut Vec<u8>,
 ) {
     match predictor {
         Predictor::None => {
@@ -433,11 +545,14 @@ fn fix_endianness_and_predict(
             rev_hpredict_nsamp(buf, bit_depth, samples);
         }
         Predictor::FloatingPoint => {
-            let mut buffer_copy = buf.to_vec();
+            // The float predictor reads from scattered byte planes, so it needs a separate input
+            // buffer.
+            scratch.resize(buf.len(), 0);
+            scratch.copy_from_slice(buf);
             match bit_depth {
-                16 => predict_f16(&mut buffer_copy, buf, samples),
-                32 => predict_f32(&mut buffer_copy, buf, samples),
-                64 => predict_f64(&mut buffer_copy, buf, samples),
+                16 => predict_f16(scratch, buf, samples),
+                32 => predict_f32(scratch, buf, samples),
+                64 => predict_f64(scratch, buf, samples),
                 _ => unreachable!("Caller should have validated arguments. Please file a bug."),
             }
         }
@@ -627,6 +742,7 @@ impl TiffHeader {
             },
             seen_ifds: cycles::IfdCycles::new(),
             image: None,
+            scratch: Vec::new(),
         }
     }
 }
@@ -712,13 +828,23 @@ impl<R: Read + Seek> Decoder<R> {
     // FIXME: this tuple would be a good candidate for a proxy, like IfdDecoder. That has been
     // requested in some form or another in a number of issues.
     #[expect(clippy::unnecessary_unwrap)] // Lifetimes would overlap due to return.
-    fn ensure_image_and_reader(&mut self) -> TiffResult<(&mut Image, &mut ValueReader<R>)> {
+    fn ensure_image_and_reader(
+        &mut self,
+    ) -> TiffResult<(&mut Image, &mut ValueReader<R>, &mut Vec<u8>)> {
         if self.image.is_some() {
-            return Ok((self.image.as_mut().unwrap(), &mut self.value_reader));
+            return Ok((
+                self.image.as_mut().unwrap(),
+                &mut self.value_reader,
+                &mut self.scratch,
+            ));
         }
 
         let image = Image::from_reader(&mut self.value_reader, &self.directory)?;
-        Ok((self.image.insert(image), &mut self.value_reader))
+        Ok((
+            self.image.insert(image),
+            &mut self.value_reader,
+            &mut self.scratch,
+        ))
     }
 
     fn ensure_image(&mut self) -> TiffResult<&mut Image> {
@@ -848,6 +974,7 @@ impl<R: Read + Seek> Decoder<R> {
             directory: self.directory.clone(),
             current_ifd_pointer: self.current_ifd_pointer,
             image: core::mem::take(&mut self.image),
+            scratch: core::mem::take(&mut self.scratch),
         }
     }
 
@@ -1323,10 +1450,10 @@ impl<R: Read + Seek> Decoder<R> {
         chunk_index: u32,
         layout: &image::ReadoutLayout,
     ) -> TiffResult<()> {
-        let (image, value_reader) = self.ensure_image_and_reader()?;
+        let (image, value_reader, scratch) = self.ensure_image_and_reader()?;
         let offset = image.chunk_file_range(chunk_index)?.0;
         value_reader.reader.goto_offset(offset)?;
-        image.expand_chunk(value_reader, buffer, layout, chunk_index)?;
+        image.expand_chunk(value_reader, buffer, layout, chunk_index, scratch)?;
         Ok(())
     }
 
@@ -1478,7 +1605,7 @@ impl<R: Read + Seek> Decoder<R> {
         slice: TiffCodingUnit,
         buffer: &mut [u8],
     ) -> TiffResult<()> {
-        let (image, value_reader) = self.ensure_image_and_reader()?;
+        let (image, value_reader, scratch) = self.ensure_image_and_reader()?;
 
         let (width, height) = image.chunk_data_dimensions(slice.0)?;
         let readout = image.readout_for_size(width, height)?;
@@ -1505,7 +1632,13 @@ impl<R: Read + Seek> Decoder<R> {
             let chunk = slice.0 + idx as u32 * readout.chunks_per_plane;
             let offset = image.chunk_offsets[chunk as usize];
             value_reader.reader.goto_offset(offset)?;
-            image.expand_chunk(value_reader, &mut buffer[plane_offset..], readout, chunk)?;
+            image.expand_chunk(
+                value_reader,
+                &mut buffer[plane_offset..],
+                readout,
+                chunk,
+                scratch,
+            )?;
         }
 
         Ok(())
@@ -1676,7 +1809,7 @@ impl<R: Read + Seek> Decoder<R> {
     /// Returns an error if the buffer fits less than one plane. In particular, for non-planar
     /// images returns an error if the buffer does not fit the required size.
     pub fn read_image_bytes(&mut self, buffer: &mut [u8]) -> TiffResult<()> {
-        let (image, value_reader) = self.ensure_image_and_reader()?;
+        let (image, value_reader, scratch) = self.ensure_image_and_reader()?;
         let readout = image.readout_for_image()?;
 
         let ref layout @ image::PlaneLayout {
@@ -1709,6 +1842,7 @@ impl<R: Read + Seek> Decoder<R> {
                     &mut buffer[plane_offset..][buffer_offset..],
                     readout,
                     chunk,
+                    scratch,
                 )?;
             }
         }