diff --git a/Cargo.toml b/Cargo.toml
index c1a7f0f..d1531f7 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -50,6 +50,10 @@ zstd = ["dep:zstd"]
 name = "lzw"
 harness = false
 
+[[bench]]
+name = "encode"
+harness = false
+
 [[test]]
 name = "decode_libtiffpic"
 path = "tests/decode_libtiffpic.rs"
diff --git a/benches/encode.rs b/benches/encode.rs
new file mode 100644
index 0000000..23b2cf9
--- /dev/null
+++ b/benches/encode.rs
@@ -0,0 +1,44 @@
+extern crate criterion;
+extern crate tiff;
+
+use criterion::{black_box, Criterion, Throughput};
+use std::io::Cursor;
+use tiff::encoder::colortype::RGB32Float;
+use tiff::encoder::{Compression, DeflateLevel, Predictor, TiffEncoder};
+
+const WIDTH: u32 = 500;
+const HEIGHT: u32 = 500;
+
+fn encode_image_f32_float_predictor_deflate(data: &[f32]) {
+    let mut buf = Cursor::new(Vec::with_capacity(WIDTH as usize * HEIGHT as usize * 4));
+    TiffEncoder::new(&mut buf)
+        .unwrap()
+        .with_predictor(Predictor::FloatingPoint)
+        .with_compression(Compression::Deflate(DeflateLevel::Fast))
+        .write_image::<RGB32Float>(WIDTH, HEIGHT, black_box(data))
+        .unwrap();
+}
+
+fn main() {
+    // Generate synthetic f32 image data with a gradient pattern
+    let data: Vec<f32> = (0..WIDTH as usize * HEIGHT as usize * 3)
+        .map(|i| {
+            let pixel = i / 3;
+            let x = (pixel % WIDTH as usize) as f32 / WIDTH as f32;
+            let y = (pixel / WIDTH as usize) as f32 / HEIGHT as f32;
+            x * 0.5 + y * 0.5
+        })
+        .collect();
+
+    let mut c = Criterion::default().configure_from_args();
+    let mut group = c.benchmark_group("tiff-encode");
+
+    group
+        .sample_size(20)
+        .throughput(Throughput::Bytes(
+            (WIDTH as u64) * (HEIGHT as u64) * 3 * std::mem::size_of::<f32>() as u64,
+        ))
+        .bench_function("f32-float-predictor-deflate-fast", |b| {
+            b.iter(|| encode_image_f32_float_predictor_deflate(&data))
+        });
+}
diff --git a/src/encoder/colortype.rs b/src/encoder/colortype.rs
index ef4b9a8..189d80c 100644
--- a/src/encoder/colortype.rs
+++ b/src/encoder/colortype.rs
@@ -18,15 +18,15 @@ macro_rules! fp_predict {
             let byte_count = num_values * BYTE_SIZE;
             let start_len = result.len();
 
-            // Reserve space for the output
-            result.reserve(byte_count);
+            // Allocate zeroed space so we can write to arbitrary positions
+            result.resize(start_len + byte_count, 0);
+            let output = &mut result[start_len..];
 
-            // Step 1 & 2: Convert to big-endian and shuffle bytes
-            // All first bytes, then all second bytes, etc.
-            for byte_idx in 0..BYTE_SIZE {
-                for &value in row {
-                    let bytes = value.to_be_bytes();
-                    result.push(bytes[byte_idx]);
+            // Step 1 & 2: Convert to big-endian and scatter bytes
+            for (i, &value) in row.iter().enumerate() {
+                let bytes = value.to_be_bytes();
+                for b in 0..BYTE_SIZE {
+                    output[b * num_values + i] = bytes[b];
                 }
             }