Vectorize byte-level horizontal differencing in fp_predict

Process the difference step in reverse 16-byte chunks, reading both
curr and prev into local arrays before writing. This makes all 16
subtractions independent, enabling autovectorization (confirmed with
SSE2: movdqu + psubb).

Co-Authored-By: Claude <noreply@anthropic.com>

src/encoder/colortype.rs

@@ -30,12 +30,59 @@ macro_rules! fp_predict {
                 }
             }
 
-            // Step 3: Apply byte-level horizontal difference across the entire buffer
-            // We iterate backwards so that when we compute output[i] - output[i-samples],
-            // output[i-samples] still has its original value.
+            // Step 3: Apply byte-level horizontal difference across the entire buffer.
             // The prediction spans across quarter boundaries, just like the decoder's cumulative sum.
-            let output = &mut result[start_len..];
-            for i in (samples..byte_count).rev() {
+            //
+            // We process elements in reverse in batches of 16 to enable autovectorization.
+            // Reverse order ensures that when we compute output[i] - output[i-samples],
+            // the value at output[i-samples] hasn't been overwritten yet (critical when
+            // samples < 16, which is the common case: 1, 3, 4, 5).
+            //
+            // 8 bits * 16 = 128 bits, the width of a typical SIMD register.
+            // Even if the target has no vector instructions, this still benefits from
+            // instruction-level parallelism, since unlike in decoding, all subtractions
+            // within a chunk are truly independent of each other.
+            let output = &mut result[start_len..byte_count + start_len];
+            let diff_len = byte_count - samples;
+            let remainder_len = diff_len % 16;
+            // Process full 16-byte chunks in reverse.
+            // We iterate from the end so that predecessor values (at lower indices)
+            // are still in their original state when we read them.
+            let mut chunk_end = byte_count;
+            while chunk_end >= samples + remainder_len + 16 {
+                let chunk_start = chunk_end - 16;
+                let prev_start = chunk_start - samples;
+
+                // Read all values in the chunk and their predecessors first.
+                // This makes all subtractions below independent, enabling vectorization.
+                let curr: [u8; 16] = output[chunk_start..chunk_end].try_into().unwrap();
+                let prev: [u8; 16] = output[prev_start..prev_start + 16].try_into().unwrap();
+
+                let chunk = &mut output[chunk_start..chunk_end];
+                chunk[0] = curr[0].wrapping_sub(prev[0]);
+                chunk[1] = curr[1].wrapping_sub(prev[1]);
+                chunk[2] = curr[2].wrapping_sub(prev[2]);
+                chunk[3] = curr[3].wrapping_sub(prev[3]);
+                chunk[4] = curr[4].wrapping_sub(prev[4]);
+                chunk[5] = curr[5].wrapping_sub(prev[5]);
+                chunk[6] = curr[6].wrapping_sub(prev[6]);
+                chunk[7] = curr[7].wrapping_sub(prev[7]);
+                chunk[8] = curr[8].wrapping_sub(prev[8]);
+                chunk[9] = curr[9].wrapping_sub(prev[9]);
+                chunk[10] = curr[10].wrapping_sub(prev[10]);
+                chunk[11] = curr[11].wrapping_sub(prev[11]);
+                chunk[12] = curr[12].wrapping_sub(prev[12]);
+                chunk[13] = curr[13].wrapping_sub(prev[13]);
+                chunk[14] = curr[14].wrapping_sub(prev[14]);
+                chunk[15] = curr[15].wrapping_sub(prev[15]);
+
+                chunk_end = chunk_start;
+            }
+            // Handle the leading remainder that doesn't fill a complete chunk of 16.
+            // This must be done last (after the reverse chunk loop) because these are
+            // the lowest indices, and the chunks above may read predecessor values
+            // from this range.
+            for i in (samples..samples + remainder_len).rev() {
                 output[i] = output[i].wrapping_sub(output[i - samples]);
             }
         }