| 1 |
|
| 2 |
//! Lossy compression for F32 data, but lossless compression for U32 and F16 data. |
| 3 |
// see https://github.com/AcademySoftwareFoundation/openexr/blob/master/OpenEXR/IlmImf/ImfPxr24Compressor.cpp |
| 4 |
|
| 5 |
// This compressor is based on source code that was contributed to |
| 6 |
// OpenEXR by Pixar Animation Studios. The compression method was |
| 7 |
// developed by Loren Carpenter. |
| 8 |
|
| 9 |
|
| 10 |
// The compressor preprocesses the pixel data to reduce entropy, and then calls zlib. |
| 11 |
// Compression of HALF and UINT channels is lossless, but compressing |
| 12 |
// FLOAT channels is lossy: 32-bit floating-point numbers are converted |
| 13 |
// to 24 bits by rounding the significand to 15 bits. |
| 14 |
// |
| 15 |
// When the compressor is invoked, the caller has already arranged |
| 16 |
// the pixel data so that the values for each channel appear in a |
| 17 |
// contiguous block of memory. The compressor converts the pixel |
| 18 |
// values to unsigned integers: For UINT, this is a no-op. HALF |
| 19 |
// values are simply re-interpreted as 16-bit integers. FLOAT |
| 20 |
// values are converted to 24 bits, and the resulting bit patterns |
| 21 |
// are interpreted as integers. The compressor then replaces each |
| 22 |
// value with the difference between the value and its left neighbor. |
| 23 |
// This turns flat fields in the image into zeroes, and ramps into |
| 24 |
// strings of similar values. Next, each difference is split into |
| 25 |
// 2, 3 or 4 bytes, and the bytes are transposed so that all the |
| 26 |
// most significant bytes end up in a contiguous block, followed |
| 27 |
// by the second most significant bytes, and so on. The resulting |
| 28 |
// string of bytes is compressed with zlib. |
| 29 |
|
| 30 |
use super::*; |
| 31 |
|
| 32 |
use crate::error::Result; |
| 33 |
use lebe::io::ReadPrimitive; |
| 34 |
|
| 35 |
|
| 36 |
// scanline decompreroussion tine, see https://github.com/openexr/openexr/blob/master/OpenEXR/IlmImf/ImfScanLineInputFile.cpp |
| 37 |
// 1. Uncompress the data, if necessary (If the line is uncompressed, it's in XDR format, regardless of the compressor's output format.) |
| 38 |
// 3. Convert one scan line's worth of pixel data back from the machine-independent representation |
| 39 |
// 4. Fill the frame buffer with pixel data, respective to sampling and whatnot |
| 40 |
|
| 41 |
|
| 42 |
pub fn compress(channels: &ChannelList, bytes_ne: ByteVec, area: IntegerBounds) -> Result<ByteVec> { |
| 43 |
if bytes_ne.is_empty() { return Ok(Vec::new()); } |
| 44 |
|
| 45 |
let mut remaining_bytes_ne = bytes_ne.as_slice(); // TODO less allocation |
| 46 |
|
| 47 |
let bytes_per_pixel: usize = channels.list.iter() |
| 48 |
.map(|channel| match channel.sample_type { |
| 49 |
SampleType::F16 => 2, |
| 50 |
SampleType::F32 => 3, |
| 51 |
SampleType::U32 => 4, |
| 52 |
}) |
| 53 |
.sum(); |
| 54 |
|
| 55 |
let mut encoded_be = vec![0_u8; bytes_per_pixel * area.size.area()]; |
| 56 |
|
| 57 |
{ |
| 58 |
let mut write = encoded_be.as_mut_slice(); |
| 59 |
|
| 60 |
// TODO this loop should be an iterator in the `IntegerBounds` class, as it is used in all compression methods |
| 61 |
for y in area.position.1..area.end().1 { |
| 62 |
for channel in &channels.list { |
| 63 |
if mod_p(y, usize_to_i32(channel.sampling.1, "sampling factor")?) != 0 { continue; } |
| 64 |
|
| 65 |
let sample_count_x = channel.subsampled_resolution(area.size).0; |
| 66 |
|
| 67 |
// this apparently can't be a closure in Rust 1.43 due to borrowing ambiguity |
| 68 |
macro_rules! split_off_write_slice { () => {{ |
| 69 |
let (slice, rest) = write.split_at_mut(sample_count_x); |
| 70 |
write = rest; |
| 71 |
slice |
| 72 |
}}; } |
| 73 |
|
| 74 |
match channel.sample_type { |
| 75 |
SampleType::F16 => { |
| 76 |
let out_byte_tuples = split_off_write_slice!().iter_mut() |
| 77 |
.zip(split_off_write_slice!()); |
| 78 |
|
| 79 |
let mut previous_pixel: u32 = 0; |
| 80 |
for (out_byte_0, out_byte_1) in out_byte_tuples { |
| 81 |
let pixel = u16::read_from_native_endian(&mut remaining_bytes_ne) |
| 82 |
.expect("failed to read from in-memory bytes") as u32; |
| 83 |
|
| 84 |
let [byte_0, byte_1] = (pixel.wrapping_sub(previous_pixel) as u16) |
| 85 |
.to_be_bytes(); |
| 86 |
|
| 87 |
*out_byte_0 = byte_0; |
| 88 |
*out_byte_1 = byte_1; |
| 89 |
previous_pixel = pixel; |
| 90 |
} |
| 91 |
}, |
| 92 |
|
| 93 |
SampleType::U32 => { |
| 94 |
let out_byte_quadruplets = split_off_write_slice!().iter_mut() |
| 95 |
.zip(split_off_write_slice!()) |
| 96 |
.zip(split_off_write_slice!()) |
| 97 |
.zip(split_off_write_slice!()); |
| 98 |
|
| 99 |
let mut previous_pixel: u32 = 0; |
| 100 |
for (((out_byte_0, out_byte_1), out_byte_2), out_byte_3) in out_byte_quadruplets { |
| 101 |
let pixel = u32::read_from_native_endian(&mut remaining_bytes_ne) |
| 102 |
.expect("failed to read from in-memory bytes"); |
| 103 |
|
| 104 |
let [byte_0, byte_1, byte_2, byte_3] = pixel |
| 105 |
.wrapping_sub(previous_pixel).to_be_bytes(); |
| 106 |
|
| 107 |
*out_byte_0 = byte_0; |
| 108 |
*out_byte_1 = byte_1; |
| 109 |
*out_byte_2 = byte_2; |
| 110 |
*out_byte_3 = byte_3; |
| 111 |
previous_pixel = pixel; |
| 112 |
} |
| 113 |
}, |
| 114 |
|
| 115 |
SampleType::F32 => { |
| 116 |
let out_byte_triplets = split_off_write_slice!().iter_mut() |
| 117 |
.zip(split_off_write_slice!()) |
| 118 |
.zip(split_off_write_slice!()); |
| 119 |
|
| 120 |
let mut previous_pixel: u32 = 0; |
| 121 |
for ((out_byte_0, out_byte_1), out_byte_2) in out_byte_triplets { |
| 122 |
let pixel = f32_to_f24( |
| 123 |
f32::read_from_native_endian(&mut remaining_bytes_ne) |
| 124 |
.expect("failed to read from in-memory bytes") |
| 125 |
); |
| 126 |
|
| 127 |
let [_, byte_0, byte_1, byte_2] = pixel |
| 128 |
.wrapping_sub(previous_pixel).to_be_bytes(); |
| 129 |
|
| 130 |
*out_byte_0 = byte_0; |
| 131 |
*out_byte_1 = byte_1; |
| 132 |
*out_byte_2 = byte_2; |
| 133 |
previous_pixel = pixel; |
| 134 |
} |
| 135 |
}, |
| 136 |
} |
| 137 |
} |
| 138 |
} |
| 139 |
|
| 140 |
debug_assert_eq!(write.len(), 0, "bytes left after compression"); |
| 141 |
} |
| 142 |
|
| 143 |
Ok(miniz_oxide::deflate::compress_to_vec_zlib(encoded_be.as_slice(), 4)) |
| 144 |
} |
| 145 |
|
| 146 |
pub fn decompress(channels: &ChannelList, bytes_le: ByteVec, area: IntegerBounds, expected_byte_size: usize, pedantic: bool) -> Result<ByteVec> { |
| 147 |
let options = zune_inflate::DeflateOptions::default().set_limit(expected_byte_size).set_size_hint(expected_byte_size); |
| 148 |
let mut decompressor = zune_inflate::DeflateDecoder::new_with_options(&bytes_le, options); |
| 149 |
|
| 150 |
let encoded_be = decompressor.decode_zlib() |
| 151 |
.map_err(|_| Error::invalid("zlib-compressed data malformed"))?; // TODO share code with zip? |
| 152 |
|
| 153 |
let mut encoded_be = encoded_be.as_slice(); |
| 154 |
let mut out = Vec::with_capacity(expected_byte_size.min(2048*4)); |
| 155 |
|
| 156 |
for y in area.position.1 .. area.end().1 { |
| 157 |
for channel in &channels.list { |
| 158 |
if mod_p(y, usize_to_i32(channel.sampling.1, "sampling")?) != 0 { continue; } |
| 159 |
|
| 160 |
let sample_count_x = channel.subsampled_resolution(area.size).0; |
| 161 |
let mut read_sample_line = ||{ |
| 162 |
if sample_count_x > encoded_be.len() { return Err(Error::invalid("not enough data")) } |
| 163 |
let (samples, rest) = encoded_be.split_at(sample_count_x); |
| 164 |
encoded_be = rest; |
| 165 |
Ok(samples) |
| 166 |
}; |
| 167 |
|
| 168 |
|
| 169 |
match channel.sample_type { |
| 170 |
SampleType::F16 => { |
| 171 |
let sample_byte_pairs = read_sample_line()?.iter() |
| 172 |
.zip(read_sample_line()?); |
| 173 |
|
| 174 |
let mut pixel_accumulation: u32 = 0; |
| 175 |
for (&in_byte_0, &in_byte_1) in sample_byte_pairs { |
| 176 |
let difference = u16::from_be_bytes([in_byte_0, in_byte_1]) as u32; |
| 177 |
pixel_accumulation = pixel_accumulation.overflowing_add(difference).0; |
| 178 |
out.extend_from_slice(&(pixel_accumulation as u16).to_ne_bytes()); |
| 179 |
} |
| 180 |
}, |
| 181 |
|
| 182 |
SampleType::U32 => { |
| 183 |
let sample_byte_quads = read_sample_line()?.iter() |
| 184 |
.zip(read_sample_line()?) |
| 185 |
.zip(read_sample_line()?) |
| 186 |
.zip(read_sample_line()?); |
| 187 |
|
| 188 |
let mut pixel_accumulation: u32 = 0; |
| 189 |
for (((&in_byte_0, &in_byte_1), &in_byte_2), &in_byte_3) in sample_byte_quads { |
| 190 |
let difference = u32::from_be_bytes([in_byte_0, in_byte_1, in_byte_2, in_byte_3]); |
| 191 |
pixel_accumulation = pixel_accumulation.overflowing_add(difference).0; |
| 192 |
out.extend_from_slice(&pixel_accumulation.to_ne_bytes()); |
| 193 |
} |
| 194 |
}, |
| 195 |
|
| 196 |
SampleType::F32 => { |
| 197 |
let sample_byte_triplets = read_sample_line()?.iter() |
| 198 |
.zip(read_sample_line()?).zip(read_sample_line()?); |
| 199 |
|
| 200 |
let mut pixel_accumulation: u32 = 0; |
| 201 |
for ((&in_byte_0, &in_byte_1), &in_byte_2) in sample_byte_triplets { |
| 202 |
let difference = u32::from_be_bytes([in_byte_0, in_byte_1, in_byte_2, 0]); |
| 203 |
pixel_accumulation = pixel_accumulation.overflowing_add(difference).0; |
| 204 |
out.extend_from_slice(&pixel_accumulation.to_ne_bytes()); |
| 205 |
} |
| 206 |
} |
| 207 |
} |
| 208 |
} |
| 209 |
} |
| 210 |
|
| 211 |
if pedantic && !encoded_be.is_empty() { |
| 212 |
return Err(Error::invalid("too much data")); |
| 213 |
} |
| 214 |
|
| 215 |
Ok(out) |
| 216 |
} |
| 217 |
|
| 218 |
|
| 219 |
|
| 220 |
|
| 221 |
/// Conversion from 32-bit to 24-bit floating-point numbers. |
| 222 |
/// Reverse conversion is just a simple 8-bit left shift. |
| 223 |
pub fn f32_to_f24(float: f32) -> u32 { |
| 224 |
let bits = float.to_bits(); |
| 225 |
|
| 226 |
let sign = bits & 0x80000000; |
| 227 |
let exponent = bits & 0x7f800000; |
| 228 |
let mantissa = bits & 0x007fffff; |
| 229 |
|
| 230 |
let result = if exponent == 0x7f800000 { |
| 231 |
if mantissa != 0 { |
| 232 |
// F is a NAN; we preserve the sign bit and |
| 233 |
// the 15 leftmost bits of the significand, |
| 234 |
// with one exception: If the 15 leftmost |
| 235 |
// bits are all zero, the NAN would turn |
| 236 |
// into an infinity, so we have to set at |
| 237 |
// least one bit in the significand. |
| 238 |
|
| 239 |
let mantissa = mantissa >> 8; |
| 240 |
(exponent >> 8) | mantissa | if mantissa == 0 { 1 } else { 0 } |
| 241 |
} |
| 242 |
else { // F is an infinity. |
| 243 |
exponent >> 8 |
| 244 |
} |
| 245 |
} |
| 246 |
else { // F is finite, round the significand to 15 bits. |
| 247 |
let result = ((exponent | mantissa) + (mantissa & 0x00000080)) >> 8; |
| 248 |
|
| 249 |
if result >= 0x7f8000 { |
| 250 |
// F was close to FLT_MAX, and the significand was |
| 251 |
// rounded up, resulting in an exponent overflow. |
| 252 |
// Avoid the overflow by truncating the significand |
| 253 |
// instead of rounding it. |
| 254 |
|
| 255 |
(exponent | mantissa) >> 8 |
| 256 |
} |
| 257 |
else { |
| 258 |
result |
| 259 |
} |
| 260 |
}; |
| 261 |
|
| 262 |
return (sign >> 8) | result; |
| 263 |
} |