Introduce LogL(uv) encodings

src/decoder/image.rs

@@ -409,11 +409,29 @@ impl Image {
             }
         };
 
+        let mut sample_format = sample_format;
+        let mut bits_per_sample = bits_per_sample[0];
+
+        // These compression methods only work for specific sample types. They pack each pixel into
+        // a pixel format. And when we decompress them we run an inverse transform inside the
+        // decompression step because that makes sense. (Even for 'raw' data you will not be able
+        // to handle the samples well otherwise). All we are doing is setting up the internal
+        // buffer here. There's an additional transform before the color readout as yo will receive
+        // the color as 32-bit floats (the encoding relies on very specific quantization so
+        // everything else would be very lossy).
+        if let CompressionMethod::SgiLog = compression_method {
+            sample_format = SampleFormat::IEEEFP;
+            bits_per_sample = 32;
+        } else if let CompressionMethod::SgiLog24 = compression_method {
+            sample_format = SampleFormat::IEEEFP;
+            bits_per_sample = 32;
+        }
+
         Ok(Image {
             ifd: Some(ifd),
             width,
             height,
-            bits_per_sample: bits_per_sample[0],
+            bits_per_sample,
             samples,
             extra_samples,
             photometric_samples,
@@ -523,6 +541,46 @@ impl Image {
                     vec![self.bits_per_sample; self.samples as usize],
                 ),
             )),
+            PhotometricInterpretation::CIELogL => match self.samples {
+                1 if matches!(self.bits_per_sample, 32)
+                    && matches!(self.sample_format, SampleFormat::IEEEFP) =>
+                {
+                    // We do this transform internally while reading.
+                    Ok(ColorType::Gray(32))
+                }
+                _ => Err(TiffError::UnsupportedError(
+                    TiffUnsupportedError::InterpretationWithBits(
+                        self.photometric_interpretation,
+                        vec![self.bits_per_sample; self.samples as usize],
+                    ),
+                )),
+            },
+            PhotometricInterpretation::CIELogLuv => match self.samples {
+                // Only fully supported for the compression format `SGILog`/`SGILog24` which must
+                // be used with 16-bit per sample.
+                //
+                // It's actually an interesting point what we should output here. Note that in
+                // libtiff there is a RAW mode (a single unsigned sample per pixel) as well as a
+                // 32-bit floating point mode for XYZ data. The `Lu'v'` variant is a heavily
+                // customized quantization so we should not return it directly. The format itself
+                // describes being based off `xyY` so that may make sense but there is a pole in
+                // the relation to `XYZ` (at y=0) which is absent in the chosen quantization since
+                // we assume a bias of `.5` so `v >= 1/820` and thus `y >= 0.49/820` in practice.
+                // Also we lose a bit of accuracy with just 16-bit precision.
+                //
+                // In `libtiff` data is always transformed through XYZ. So let's do that.
+                3 if matches!(self.bits_per_sample, 32)
+                    && matches!(self.sample_format, SampleFormat::IEEEFP) =>
+                {
+                    Ok(ColorType::RGB(32))
+                }
+                _ => Err(TiffError::UnsupportedError(
+                    TiffUnsupportedError::InterpretationWithBits(
+                        self.photometric_interpretation,
+                        vec![self.bits_per_sample; self.samples as usize],
+                    ),
+                )),
+            },
             // Unsupported due to extra unfiltering and conversion steps. We need to find the
             // Decode tag (SRATIONAL; 2 * SamplesPerPixel) and apply the following conversion:
             //
@@ -554,11 +612,35 @@ impl Image {
         compressed_length: u64,
         // FIXME: these should be `expect` attributes or we choose another way of passing them.
         #[cfg_attr(not(feature = "jpeg"), allow(unused_variables))] jpeg_tables: Option<&[u8]>,
-        #[cfg_attr(not(feature = "fax"), allow(unused_variables))] dimensions: (u32, u32),
+        dimensions: (u32, u32),
-        #[cfg_attr(not(feature = "webp"), allow(unused_variables))] samples: u16,
+        samples: u16,
     ) -> TiffResult<Box<dyn Read + 'r>> {
         Ok(match compression_method {
             CompressionMethod::None => Box::new(reader),
+            CompressionMethod::SgiLog => match samples {
+                1 => Box::new(super::logluv::LogLuv16::new(
+                    reader,
+                    compressed_length,
+                    dimensions.0,
+                )?),
+                3 => Box::new(super::logluv::LogLuv32::new(
+                    reader,
+                    compressed_length,
+                    dimensions.0,
+                )?),
+                _ => {
+                    return Err(TiffError::UnsupportedError(
+                        TiffUnsupportedError::UnsupportedCompressionMethod(
+                            CompressionMethod::SgiLog,
+                        ),
+                    ))
+                }
+            },
+            CompressionMethod::SgiLog24 => Box::new(super::logluv::LogLuv24::new(
+                reader,
+                compressed_length,
+                dimensions.0,
+            )?),
             #[cfg(feature = "lzw")]
             CompressionMethod::LZW => Box::new(super::stream::LZWReader::new(
                 reader,

src/decoder/logluv.rs

@@ -0,0 +1,541 @@
+use std::io::{Read, Take};
+
+use crate::TiffError;
+
+pub struct LogLuv24<R> {
+    reader: Take<R>,
+    /// The required buffer size. We only decompressed all samples at once.
+    chunk_sz: usize,
+}
+
+impl<R> LogLuv24<R> {
+    pub fn new(reader: R, n: u64, npixels: u32) -> Result<Self, TiffError>
+    where
+        R: Read,
+    {
+        let chunk_sz = usize::try_from(
+            u64::from(npixels)
+                .checked_mul(4 * 3)
+                .ok_or(TiffError::LimitsExceeded)?,
+        )?;
+
+        let reader = reader.take(n);
+        Ok(LogLuv24 { reader, chunk_sz })
+    }
+
+    fn expand_buf(buf: &mut [u8]) {
+        let data_len = buf.len() / 4;
+
+        for i in (0..data_len / 3).rev() {
+            let ch = buf.as_chunks_mut::<3>().0[i];
+            let [a, b, c] = ch.map(u32::from);
+
+            let raw = a << 16 | b << 8 | c;
+            let l = (raw >> 14) & 0x3ff;
+            let uv = decode_quantized_uv(raw & 0x3fff);
+
+            let by = dequantize_24_luma(l);
+            let (u, v) = dequantize_24(uv);
+
+            let s = 1. / (6. * u - 16. * v + 12.);
+            let x = 9. * u * s;
+            let y = 4. * v * s;
+
+            let bx = x / y * by;
+            let bz = (1. - x - y) / y * by;
+
+            let [r, g, b] = xyz_to_rgb([bx, by, bz]);
+
+            let into = &mut buf.as_chunks_mut::<12>().0[i];
+            into[0..4].copy_from_slice(&f32::to_ne_bytes(r));
+            into[4..8].copy_from_slice(&f32::to_ne_bytes(g));
+            into[8..12].copy_from_slice(&f32::to_ne_bytes(b));
+        }
+    }
+}
+
+impl<R: Read> Read for LogLuv24<R> {
+    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
+        self.read_exact(buf)?;
+        Ok(buf.len())
+    }
+
+    // FIXME: this is the interface we have for compressed chunks but it is not optimal for what we
+    // want. Note `fill_buf` is not super helpful since it does not guarantee any size of the
+    // filled buffer but we need 24-bits at a time. Two code paths is even more awkward.
+    fn read_exact(&mut self, buf: &mut [u8]) -> std::io::Result<()> {
+        if !buf.len().is_multiple_of(self.chunk_sz) {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidInput,
+                format!(
+                    "buffer size {0} must match the chunk size for compression method SGILog24 {1}",
+                    buf.len(),
+                    self.chunk_sz,
+                ),
+            ));
+        }
+
+        let data_len = buf.len() / 4;
+        self.reader.read_exact(&mut buf[..data_len])?;
+        Self::expand_buf(buf);
+
+        Ok(())
+    }
+}
+
+fn packbits_decode<R: Read, const N: usize>(
+    reader: &mut R,
+    buf: &mut [[u8; N]],
+) -> std::io::Result<()> {
+    for offset in (0..N).rev() {
+        let mut i = 0;
+        let mut code = [0u8; 1];
+
+        loop {
+            let Some(pixels) = buf.get_mut(i..) else {
+                break;
+            };
+
+            if pixels.is_empty() {
+                break;
+            }
+
+            let pixels = pixels.iter_mut();
+            reader.read_exact(&mut code)?;
+            let [repeat] = code;
+
+            if repeat >= 128 {
+                reader.read_exact(&mut code)?;
+                let [value] = code;
+                let rc = (repeat - 128) + 2;
+
+                for (_, px) in (0..rc).zip(pixels) {
+                    px[offset] = value;
+                }
+
+                i += usize::from(rc);
+            } else {
+                for (_, px) in (0..repeat).zip(pixels) {
+                    reader.read_exact(&mut code)?;
+                    let [value] = code;
+                    px[offset] = value;
+                }
+
+                i += usize::from(repeat);
+            }
+        }
+    }
+
+    Ok(())
+}
+
+pub struct LogLuv32<R> {
+    reader: Take<R>,
+    /// The required buffer size. We only decompressed all samples at once.
+    chunk_sz: usize,
+}
+
+impl<R: Read> LogLuv32<R> {
+    pub fn new(reader: R, n: u64, npixels: u32) -> Result<Self, TiffError> {
+        let chunk_sz = usize::try_from(
+            u64::from(npixels)
+                .checked_mul(4 * 3)
+                .ok_or(TiffError::LimitsExceeded)?,
+        )?;
+
+        let reader = reader.take(n);
+        Ok(LogLuv32 { reader, chunk_sz })
+    }
+
+    fn decode_rle_row(&mut self, row: &mut [u8]) -> std::io::Result<()>
+    where
+        R: Read,
+    {
+        // 32-bit per pixel format.
+        let n_pixels = self.chunk_sz / 12;
+        // buffer for reading (using only 32-bit per pixel)
+        let raw = &mut row.as_chunks_mut::<4>().0[..n_pixels];
+        packbits_decode::<_, 4>(&mut self.reader, raw)?;
+
+        Ok(())
+    }
+
+    fn expand_row(row: &mut [u8]) {
+        let uvscale = 1.0 / 410.0;
+        let n_pixels = row.len() / 12;
+
+        // Work backwards, the 32-bit compressed raw data is at the front of the row buffer and
+        // this avoids overwriting data before we read it.
+        for i in (0..n_pixels).rev() {
+            let raw = row.as_chunks_mut::<4>().0[i];
+
+            let l = u16::from(raw[0]) << 8 | u16::from(raw[1]);
+            let by = log16_to_y(l);
+
+            let u = uvscale * (f32::from(raw[2]) + 0.5);
+            let v = uvscale * (f32::from(raw[3]) + 0.5);
+
+            let s = 1. / (6. * u - 16. * v + 12.);
+            let x = 9. * u * s;
+            let y = 4. * v * s;
+
+            let bx = x / y * by;
+            let bz = (1. - x - y) / y * by;
+
+            let [r, g, b] = xyz_to_rgb([bx, by, bz]);
+            let into = &mut row.as_chunks_mut::<12>().0[i];
+            into[0..4].copy_from_slice(&f32::to_ne_bytes(r));
+            into[4..8].copy_from_slice(&f32::to_ne_bytes(g));
+            into[8..12].copy_from_slice(&f32::to_ne_bytes(b));
+        }
+    }
+}
+
+impl<R: Read> Read for LogLuv32<R> {
+    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
+        self.read_exact(buf)?;
+        Ok(buf.len())
+    }
+
+    // FIXME: See `LogLuv24::read_exact`
+    fn read_exact(&mut self, buf: &mut [u8]) -> std::io::Result<()> {
+        for row in buf.chunks_exact_mut(self.chunk_sz) {
+            self.decode_rle_row(row)?;
+            Self::expand_row(row);
+        }
+
+        Ok(())
+    }
+}
+
+/// Pack-Bits like encoding for a single 16-bit log-L sample.
+pub struct LogLuv16<R> {
+    reader: Take<R>,
+    /// The required buffer size. We only decompressed all samples at once.
+    chunk_sz: usize,
+}
+
+impl<R: Read> LogLuv16<R> {
+    pub fn new(reader: R, n: u64, npixels: u32) -> Result<Self, TiffError> {
+        let chunk_sz = usize::try_from(
+            u64::from(npixels)
+                .checked_mul(4)
+                .ok_or(TiffError::LimitsExceeded)?,
+        )?;
+
+        let reader = reader.take(n);
+        Ok(LogLuv16 { reader, chunk_sz })
+    }
+
+    fn decode_rle_row(&mut self, row: &mut [u8]) -> std::io::Result<()>
+    where
+        R: Read,
+    {
+        // 32-bit per pixel format.
+        let n_pixels = self.chunk_sz / 4;
+        // buffer for reading (16-bit per pixel)
+        let raw = &mut row.as_chunks_mut::<2>().0[..n_pixels];
+        packbits_decode::<_, 2>(&mut self.reader, raw)?;
+
+        Ok(())
+    }
+
+    fn expand_row(row: &mut [u8]) {
+        let n_pixels = row.len() / 4;
+
+        // Work backwards, the 32-bit compressed raw data is at the front of the row buffer and
+        // this avoids overwriting data before we read it.
+        for i in (0..n_pixels).rev() {
+            let raw = row.as_chunks_mut::<2>().0[i];
+
+            let l = u16::from(raw[0]) << 8 | u16::from(raw[1]);
+            let by = log16_to_y(l);
+
+            let into = &mut row.as_chunks_mut::<4>().0[i];
+            into[0..4].copy_from_slice(&f32::to_ne_bytes(by));
+        }
+    }
+}
+
+impl<R: Read> Read for LogLuv16<R> {
+    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
+        self.read_exact(buf)?;
+        Ok(buf.len())
+    }
+
+    // FIXME: See `LogLuv24::read_exact`
+    fn read_exact(&mut self, buf: &mut [u8]) -> std::io::Result<()> {
+        for row in buf.chunks_exact_mut(self.chunk_sz) {
+            self.decode_rle_row(row)?;
+            Self::expand_row(row);
+        }
+
+        Ok(())
+    }
+}
+
+// Table information from: `libtiff`, 1997 by Greg Ward Larson, SGI
+#[doc(alias = "UV_SQSIZ")]
+const UV_QUANTIZATION_WIDTH: f32 = 0.003500;
+
+// The `v` value corresonding to the bottom most row (quantized v' = 0).
+const UV_VSTART: f32 = 0.016940;
+
+/// Number of distinct 'u' rows, each at a different v level. Each is quantized individually with
+/// the first quantization level starting at its `ustart` and `UV_QUANTIZATION_WIDTH` wide.
+const UV_NVS: usize = 163;
+
+/// Contrary to the original C code we calculate the dependent values (cumulative sum).
+struct UvRow {
+    ustart: f32,
+    num_us: u16,
+}
+
+#[rustfmt::skip]
+const UV_ROWS: [UvRow; UV_NVS] = [
+    UvRow { ustart: 0.247663, num_us: 4, },
+    UvRow { ustart: 0.243779, num_us: 6, },
+    UvRow { ustart: 0.241684, num_us: 7, },
+    UvRow { ustart: 0.237874, num_us: 9, },
+    UvRow { ustart: 0.235906, num_us: 10, },
+    UvRow { ustart: 0.232153, num_us: 12, },
+    UvRow { ustart: 0.228352, num_us: 14, },
+    UvRow { ustart: 0.226259, num_us: 15, },
+    UvRow { ustart: 0.222371, num_us: 17, },
+    UvRow { ustart: 0.220410, num_us: 18, },
+    UvRow { ustart: 0.214710, num_us: 21, },
+    UvRow { ustart: 0.212714, num_us: 22, },
+    UvRow { ustart: 0.210721, num_us: 23, },
+    UvRow { ustart: 0.204976, num_us: 26, },
+    UvRow { ustart: 0.202986, num_us: 27, },
+    UvRow { ustart: 0.199245, num_us: 29, },
+    UvRow { ustart: 0.195525, num_us: 31, },
+    UvRow { ustart: 0.193560, num_us: 32, },
+    UvRow { ustart: 0.189878, num_us: 34, },
+    UvRow { ustart: 0.186216, num_us: 36, },
+    UvRow { ustart: 0.186216, num_us: 36, },
+    UvRow { ustart: 0.182592, num_us: 38, },
+    UvRow { ustart: 0.179003, num_us: 40, },
+    UvRow { ustart: 0.175466, num_us: 42, },
+    UvRow { ustart: 0.172001, num_us: 44, },
+    UvRow { ustart: 0.172001, num_us: 44, },
+    UvRow { ustart: 0.168612, num_us: 46, },
+    UvRow { ustart: 0.168612, num_us: 46, },
+    UvRow { ustart: 0.163575, num_us: 49, },
+    UvRow { ustart: 0.158642, num_us: 52, },
+    UvRow { ustart: 0.158642, num_us: 52, },
+    UvRow { ustart: 0.158642, num_us: 52, },
+    UvRow { ustart: 0.153815, num_us: 55, },
+    UvRow { ustart: 0.153815, num_us: 55, },
+    UvRow { ustart: 0.149097, num_us: 58, },
+    UvRow { ustart: 0.149097, num_us: 58, },
+    UvRow { ustart: 0.142746, num_us: 62, },
+    UvRow { ustart: 0.142746, num_us: 62, },
+    UvRow { ustart: 0.142746, num_us: 62, },
+    UvRow { ustart: 0.138270, num_us: 65, },
+    UvRow { ustart: 0.138270, num_us: 65, },
+    UvRow { ustart: 0.138270, num_us: 65, },
+    UvRow { ustart: 0.132166, num_us: 69, },
+    UvRow { ustart: 0.132166, num_us: 69, },
+    UvRow { ustart: 0.126204, num_us: 73, },
+    UvRow { ustart: 0.126204, num_us: 73, },
+    UvRow { ustart: 0.126204, num_us: 73, },
+    UvRow { ustart: 0.120381, num_us: 77, },
+    UvRow { ustart: 0.120381, num_us: 77, },
+    UvRow { ustart: 0.120381, num_us: 77, },
+    UvRow { ustart: 0.120381, num_us: 77, },
+    UvRow { ustart: 0.112962, num_us: 82, },
+    UvRow { ustart: 0.112962, num_us: 82, },
+    UvRow { ustart: 0.112962, num_us: 82, },
+    UvRow { ustart: 0.107450, num_us: 86, },
+    UvRow { ustart: 0.107450, num_us: 86, },
+    UvRow { ustart: 0.107450, num_us: 86, },
+    UvRow { ustart: 0.107450, num_us: 86, },
+    UvRow { ustart: 0.100343, num_us: 91, },
+    UvRow { ustart: 0.100343, num_us: 91, },
+    UvRow { ustart: 0.100343, num_us: 91, },
+    UvRow { ustart: 0.095126, num_us: 95, },
+    UvRow { ustart: 0.095126, num_us: 95, },
+    UvRow { ustart: 0.095126, num_us: 95, },
+    UvRow { ustart: 0.095126, num_us: 95, },
+    UvRow { ustart: 0.088276, num_us: 100, },
+    UvRow { ustart: 0.088276, num_us: 100, },
+    UvRow { ustart: 0.088276, num_us: 100, },
+    UvRow { ustart: 0.088276, num_us: 100, },
+    UvRow { ustart: 0.081523, num_us: 105, },
+    UvRow { ustart: 0.081523, num_us: 105, },
+    UvRow { ustart: 0.081523, num_us: 105, },
+    UvRow { ustart: 0.081523, num_us: 105, },
+    UvRow { ustart: 0.074861, num_us: 110, },
+    UvRow { ustart: 0.074861, num_us: 110, },
+    UvRow { ustart: 0.074861, num_us: 110, },
+    UvRow { ustart: 0.074861, num_us: 110, },
+    UvRow { ustart: 0.068290, num_us: 115, },
+    UvRow { ustart: 0.068290, num_us: 115, },
+    UvRow { ustart: 0.068290, num_us: 115, },
+    UvRow { ustart: 0.068290, num_us: 115, },
+    UvRow { ustart: 0.063573, num_us: 119, },
+    UvRow { ustart: 0.063573, num_us: 119, },
+    UvRow { ustart: 0.063573, num_us: 119, },
+    UvRow { ustart: 0.063573, num_us: 119, },
+    UvRow { ustart: 0.057219, num_us: 124, },
+    UvRow { ustart: 0.057219, num_us: 124, },
+    UvRow { ustart: 0.057219, num_us: 124, },
+    UvRow { ustart: 0.057219, num_us: 124, },
+    UvRow { ustart: 0.050985, num_us: 129, },
+    UvRow { ustart: 0.050985, num_us: 129, },
+    UvRow { ustart: 0.050985, num_us: 129, },
+    UvRow { ustart: 0.050985, num_us: 129, },
+    UvRow { ustart: 0.050985, num_us: 129, },
+    UvRow { ustart: 0.044859, num_us: 134, },
+    UvRow { ustart: 0.044859, num_us: 134, },
+    UvRow { ustart: 0.044859, num_us: 134, },
+    UvRow { ustart: 0.044859, num_us: 134, },
+    UvRow { ustart: 0.040571, num_us: 138, },
+    UvRow { ustart: 0.040571, num_us: 138, },
+    UvRow { ustart: 0.040571, num_us: 138, },
+    UvRow { ustart: 0.040571, num_us: 138, },
+    UvRow { ustart: 0.036339, num_us: 142, },
+    UvRow { ustart: 0.036339, num_us: 142, },
+    UvRow { ustart: 0.036339, num_us: 142, },
+    UvRow { ustart: 0.036339, num_us: 142, },
+    UvRow { ustart: 0.032139, num_us: 146, },
+    UvRow { ustart: 0.032139, num_us: 146, },
+    UvRow { ustart: 0.032139, num_us: 146, },
+    UvRow { ustart: 0.032139, num_us: 146, },
+    UvRow { ustart: 0.027947, num_us: 150, },
+    UvRow { ustart: 0.027947, num_us: 150, },
+    UvRow { ustart: 0.027947, num_us: 150, },
+    UvRow { ustart: 0.023739, num_us: 154, },
+    UvRow { ustart: 0.023739, num_us: 154, },
+    UvRow { ustart: 0.023739, num_us: 154, },
+    UvRow { ustart: 0.023739, num_us: 154, },
+    UvRow { ustart: 0.019504, num_us: 158, },
+    UvRow { ustart: 0.019504, num_us: 158, },
+    UvRow { ustart: 0.019504, num_us: 158, },
+    UvRow { ustart: 0.016976, num_us: 161, },
+    UvRow { ustart: 0.016976, num_us: 161, },
+    UvRow { ustart: 0.016976, num_us: 161, },
+    UvRow { ustart: 0.016976, num_us: 161, },
+    UvRow { ustart: 0.012639, num_us: 165, },
+    UvRow { ustart: 0.012639, num_us: 165, },
+    UvRow { ustart: 0.012639, num_us: 165, },
+    UvRow { ustart: 0.009991, num_us: 168, },
+    UvRow { ustart: 0.009991, num_us: 168, },
+    UvRow { ustart: 0.009991, num_us: 168, },
+    UvRow { ustart: 0.009016, num_us: 170, },
+    UvRow { ustart: 0.009016, num_us: 170, },
+    UvRow { ustart: 0.009016, num_us: 170, },
+    UvRow { ustart: 0.006217, num_us: 173, },
+    UvRow { ustart: 0.006217, num_us: 173, },
+    UvRow { ustart: 0.005097, num_us: 175, },
+    UvRow { ustart: 0.005097, num_us: 175, },
+    UvRow { ustart: 0.005097, num_us: 175, },
+    UvRow { ustart: 0.003909, num_us: 177, },
+    UvRow { ustart: 0.003909, num_us: 177, },
+    UvRow { ustart: 0.002340, num_us: 177, },
+    UvRow { ustart: 0.002389, num_us: 170, },
+    UvRow { ustart: 0.001068, num_us: 164, },
+    UvRow { ustart: 0.001653, num_us: 157, },
+    UvRow { ustart: 0.000717, num_us: 150, },
+    UvRow { ustart: 0.001614, num_us: 143, },
+    UvRow { ustart: 0.000270, num_us: 136, },
+    UvRow { ustart: 0.000484, num_us: 129, },
+    UvRow { ustart: 0.001103, num_us: 123, },
+    UvRow { ustart: 0.001242, num_us: 115, },
+    UvRow { ustart: 0.001188, num_us: 109, },
+    UvRow { ustart: 0.001011, num_us: 103, },
+    UvRow { ustart: 0.000709, num_us: 97, },
+    UvRow { ustart: 0.000301, num_us: 89, },
+    UvRow { ustart: 0.002416, num_us: 82, },
+    UvRow { ustart: 0.003251, num_us: 76, },
+    UvRow { ustart: 0.003246, num_us: 69, },
+    UvRow { ustart: 0.004141, num_us: 62, },
+    UvRow { ustart: 0.005963, num_us: 55, },
+    UvRow { ustart: 0.008839, num_us: 47, },
+    UvRow { ustart: 0.010490, num_us: 40, },
+    UvRow { ustart: 0.016994, num_us: 31, },
+    UvRow { ustart: 0.023659, num_us: 21, },
+];
+
+const UV_CUMULATIVE: [u16; UV_NVS] = {
+    let mut arr = [0u16; UV_NVS];
+    let mut sum = 0;
+    let mut i = 0;
+    while i < UV_NVS {
+        arr[i] = sum;
+        sum += UV_ROWS[i].num_us;
+        i += 1;
+    }
+    arr
+};
+
+const UV_NDIVS: u16 = UV_CUMULATIVE[UV_NVS - 1] + UV_ROWS[UV_NVS - 1].num_us;
+
+fn decode_quantized_uv(uv: u32) -> (u8, u8) {
+    if uv >= UV_NDIVS as u32 {
+        return (0, 0);
+    }
+
+    // Find the start of the row, or the index after that.
+    let v_index = match UV_CUMULATIVE.binary_search(&(uv as u16)) {
+        Ok(n) => n,
+        Err(n) => n - 1,
+    };
+
+    let u_index = uv as u16 - UV_CUMULATIVE[v_index];
+    (u_index as u8, v_index as u8)
+}
+
+fn dequantize_24_luma(l: u32) -> f32 {
+    if l == 0 {
+        0.0
+    } else {
+        let le = (l & 0x3ff) as f32;
+        let exponent = ((le / 64.0).floor() as i32) - 24;
+        let mantissa = le - (exponent as f32 + 24.0) * 64.0;
+        (mantissa + 64.0) / 64.0 * 2f32.powi(exponent)
+    }
+}
+
+fn dequantize_24((u, v): (u8, u8)) -> (f32, f32) {
+    let v_f = UV_VSTART + (v as f32) * UV_QUANTIZATION_WIDTH;
+    let u_row = &UV_ROWS[v as usize];
+    let u_f = u_row.ustart + (u as f32) * UV_QUANTIZATION_WIDTH;
+    (u_f, v_f)
+}
+
+fn log16_to_y(l: u16) -> f32 {
+    if l == 0 {
+        return 0.0;
+    }
+
+    let le = f32::from(l & 0x7fff);
+    //     Y = exp(M_LN2 / 256. * (Le + .5) - M_LN2 * 64.);
+    let y = (std::f32::consts::LN_2 / 256.0 * (le + 0.5) - std::f32::consts::LN_2 * 64.0).exp();
+
+    if l & 0x8000 != 0 {
+        -y
+    } else {
+        y
+    }
+}
+
+fn xyz_to_rgb([l, u, v]: [f32; 3]) -> [f32; 3] {
+    // XYZ to sRGB primaries at D65
+    let r = 3.2404542 * l + -1.5371385 * u + -0.4985314 * v;
+    let g = -0.969266 * l + 1.8760108 * u + 0.0415560 * v;
+    let b = 0.0556434 * l + -0.2040259 * u + 1.0572252 * v;
+    [r, g, b]
+}
+
+#[test]
+fn test_decode_quantized_uv() {
+    // Test some known values from the original C code.
+    assert_eq!(decode_quantized_uv(0), (0, 0));
+    assert_eq!(decode_quantized_uv(3), (3, 0));
+    assert_eq!(decode_quantized_uv(4), (0, 1));
+    assert_eq!(decode_quantized_uv(9), (5, 1));
+    assert_eq!(decode_quantized_uv(16288), (20, 162));
+    assert_eq!(decode_quantized_uv(16289), (0, 0)); // Out of range
+}

src/decoder/mod.rs

@@ -19,6 +19,7 @@ use self::stream::{ByteOrder, EndianReader};
 mod cycles;
 pub mod ifd;
 mod image;
+mod logluv;
 mod stream;
 mod tag_reader;
 

src/tags.rs

@@ -283,6 +283,14 @@ pub enum CompressionMethod(u16) unknown(
 
     // Self-assigned by libtiff
     WebP = 0xC351,
+
+    // Self-assigned by libtiff/SGI
+    // A packbit like scheme on specific 32-bit samples (16 sint luma, 8 sint u,v). Only really
+    // compatible with PhotometricInterpretation CIELogLuv.
+    SgiLog = 34676,
+    // A packbit like scheme on specific 24-bit samples (10 packed bits luma, 14-bit uv index).
+    // Only really compatible with PhotometricInterpretation CIELogLuv.
+    SgiLog24 = 34677,
 }
 }
 
@@ -295,9 +303,41 @@ pub enum PhotometricInterpretation(u16) {
     TransparencyMask = 4,
     CMYK = 5,
     YCbCr = 6,
+    /// Biased Lab with all coefficients stored as unsigned integers.
     CIELab = 8,
     IccLab = 9,
     ItuLab = 10,
+    /// Logarithmic L* color space,
+    ///
+    /// see: <https://web.archive.org/web/20250121120358/https://www.anyhere.com/gward/pixformat/tiffluv.html>
+    /// References: <https://doi.org/10.2352/CIC.1998.6.1.art00046>
+    ///
+    /// The `uv` coordinates are based on:
+    ///
+    /// ```text
+    /// (6*u - 16*v + 12) * (-2x + 12y + 3) = 36
+    ///
+    /// u' = 4x / ( -2x + 12y + 3 )
+    /// v' = 9y / ( -2x + 12y + 3 )
+    /// x = 9u' / (6u' - 16v' + 12)
+    /// y = 4v' / (6u' - 16v' + 12)
+    /// ```
+    ///
+    /// Since `(-2x + 12y + 3) <= 15` it holds `(6u' - 16v' + 12) >= 36/15` for all XYZ colors
+    /// which makes the back transform to `xyY` well-defined. Also `v' <= 9/15 = 0.6` and also
+    /// by practical means `u' <= 2/3`.
+    ///
+    /// The compression format that is supposed to be used with this constrains them as well. It
+    /// encodes each component as a 16-bit unsigned integer scaled and clamped to a quantization
+    /// using 256 out of 410 levels or an index into a table based on a quantization to 177 levels
+    /// each `0.0035` (~1/286) wide and 163 levels of `v` (same quantization width).
+    ///
+    /// > Larson, G.W., ``Overcoming Gamut and Dynamic Range Limitations in Digital Images,'' Proceedings of the Sixth Color Imaging Conference, November 1998.
+    CIELogL = 32844,
+    /// Logarithmic L*u'v' color space,
+    ///
+    /// See references under [`Self::CIELogL`].
+    CIELogLuv = 32845,
 }
 }