Add FP16 support (#257)

src/decoder/image.rs

@@ -1,7 +1,7 @@
 use super::ifd::{Directory, Value};
 use super::stream::{ByteOrder, DeflateReader, LZWReader, PackBitsReader};
 use super::tag_reader::TagReader;
-use super::{predict_f32, predict_f64, Limits};
+use super::{predict_f16, predict_f32, predict_f64, Limits};
 use super::{stream::SmartReader, ChunkType};
 use crate::tags::{
     CompressionMethod, PhotometricInterpretation, PlanarConfiguration, Predictor, SampleFormat, Tag,
@@ -592,7 +592,10 @@ impl Image {
 
         // Validate that the predictor is supported for the sample type.
         match (self.predictor, self.sample_format) {
-            (Predictor::Horizontal, SampleFormat::Int | SampleFormat::Uint) => {}
+            (
+                Predictor::Horizontal,
+                SampleFormat::Int | SampleFormat::Uint | SampleFormat::IEEEFP,
+            ) => {}
             (Predictor::Horizontal, _) => {
                 return Err(TiffError::UnsupportedError(
                     TiffUnsupportedError::HorizontalPredictor(color_type),
@@ -672,6 +675,7 @@ impl Image {
 
                 let row = &mut row[..data_row_bytes];
                 match color_type.bit_depth() {
+                    16 => predict_f16(&mut encoded, row, samples),
                     32 => predict_f32(&mut encoded, row, samples),
                     64 => predict_f64(&mut encoded, row, samples),
                     _ => unreachable!(),

src/decoder/mod.rs

@@ -8,6 +8,7 @@ use crate::tags::{
 use crate::{
     bytecast, ColorType, TiffError, TiffFormatError, TiffResult, TiffUnsupportedError, UsageError,
 };
+use half::f16;
 
 use self::ifd::Directory;
 use self::image::Image;
@@ -29,6 +30,8 @@ pub enum DecodingResult {
     U32(Vec<u32>),
     /// A vector of 64 bit unsigned ints
     U64(Vec<u64>),
+    /// A vector of 16 bit IEEE floats (held in u16)
+    F16(Vec<f16>),
     /// A vector of 32 bit IEEE floats
     F32(Vec<f32>),
     /// A vector of 64 bit IEEE floats
@@ -92,6 +95,14 @@ impl DecodingResult {
         }
     }
 
+    fn new_f16(size: usize, limits: &Limits) -> TiffResult<DecodingResult> {
+        if size > limits.decoding_buffer_size / std::mem::size_of::<u16>() {
+            Err(TiffError::LimitsExceeded)
+        } else {
+            Ok(DecodingResult::F16(vec![f16::ZERO; size]))
+        }
+    }
+
     fn new_i8(size: usize, limits: &Limits) -> TiffResult<DecodingResult> {
         if size > limits.decoding_buffer_size / std::mem::size_of::<i8>() {
             Err(TiffError::LimitsExceeded)
@@ -130,6 +141,7 @@ impl DecodingResult {
             DecodingResult::U16(ref mut buf) => DecodingBuffer::U16(&mut buf[start..]),
             DecodingResult::U32(ref mut buf) => DecodingBuffer::U32(&mut buf[start..]),
             DecodingResult::U64(ref mut buf) => DecodingBuffer::U64(&mut buf[start..]),
+            DecodingResult::F16(ref mut buf) => DecodingBuffer::F16(&mut buf[start..]),
             DecodingResult::F32(ref mut buf) => DecodingBuffer::F32(&mut buf[start..]),
             DecodingResult::F64(ref mut buf) => DecodingBuffer::F64(&mut buf[start..]),
             DecodingResult::I8(ref mut buf) => DecodingBuffer::I8(&mut buf[start..]),
@@ -150,6 +162,8 @@ pub enum DecodingBuffer<'a> {
     U32(&'a mut [u32]),
     /// A slice of 64 bit unsigned ints
     U64(&'a mut [u64]),
+    /// A slice of 16 bit IEEE floats
+    F16(&'a mut [f16]),
     /// A slice of 32 bit IEEE floats
     F32(&'a mut [f32]),
     /// A slice of 64 bit IEEE floats
@@ -175,6 +189,7 @@ impl<'a> DecodingBuffer<'a> {
             DecodingBuffer::I32(buf) => bytecast::i32_as_ne_mut_bytes(buf),
             DecodingBuffer::U64(buf) => bytecast::u64_as_ne_mut_bytes(buf),
             DecodingBuffer::I64(buf) => bytecast::i64_as_ne_mut_bytes(buf),
+            DecodingBuffer::F16(buf) => bytecast::f16_as_ne_mut_bytes(buf),
             DecodingBuffer::F32(buf) => bytecast::f32_as_ne_mut_bytes(buf),
             DecodingBuffer::F64(buf) => bytecast::f64_as_ne_mut_bytes(buf),
         }
@@ -303,6 +318,19 @@ fn predict_f32(input: &mut [u8], output: &mut [u8], samples: usize) {
     }
 }
 
+fn predict_f16(input: &mut [u8], output: &mut [u8], samples: usize) {
+    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],
+        ])));
+    }
+}
+
 fn predict_f64(input: &mut [u8], output: &mut [u8], samples: usize) {
     for i in samples..input.len() {
         input[i] = input[i].wrapping_add(input[i - samples]);
@@ -340,6 +368,7 @@ fn fix_endianness_and_predict(
         Predictor::FloatingPoint => {
             let mut buffer_copy = buf.to_vec();
             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),
                 _ => unreachable!("Caller should have validated arguments. Please file a bug."),
@@ -1004,6 +1033,7 @@ impl<R: Read + Seek> Decoder<R> {
                 )),
             },
             SampleFormat::IEEEFP => match max_sample_bits {
+                16 => DecodingResult::new_f16(buffer_size, &self.limits),
                 32 => DecodingResult::new_f32(buffer_size, &self.limits),
                 64 => DecodingResult::new_f64(buffer_size, &self.limits),
                 n => Err(TiffError::UnsupportedError(