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 decompression routine, 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 | #[cfg_attr (target_endian = "big" , allow(unused, unreachable_code))]
|
43 | pub fn compress(channels: &ChannelList, remaining_bytes: ByteVec, area: IntegerBounds) -> Result<ByteVec> {
|
44 | #[cfg (target_endian = "big" )] {
|
45 | return Err(Error::unsupported(
|
46 | "PXR24 compression method not supported yet on big endian processor architecture"
|
47 | ))
|
48 | }
|
49 |
|
50 | if remaining_bytes.is_empty() { return Ok(Vec::new()); }
|
51 |
|
52 | // see https://github.com/AcademySoftwareFoundation/openexr/blob/3bd93f85bcb74c77255f28cdbb913fdbfbb39dfe/OpenEXR/IlmImf/ImfTiledOutputFile.cpp#L750-L842
|
53 | let remaining_bytes = super::convert_current_to_little_endian(remaining_bytes, channels, area);
|
54 | let mut remaining_bytes = remaining_bytes.as_slice(); // TODO less allocation
|
55 |
|
56 | let bytes_per_pixel: usize = channels.list.iter()
|
57 | .map(|channel| match channel.sample_type {
|
58 | SampleType::F16 => 2, SampleType::F32 => 3, SampleType::U32 => 4,
|
59 | })
|
60 | .sum();
|
61 |
|
62 | let mut raw = vec![0_u8; bytes_per_pixel * area.size.area()];
|
63 |
|
64 | {
|
65 | let mut write = raw.as_mut_slice();
|
66 |
|
67 | // TODO this loop should be an iterator in the `IntegerBounds` class, as it is used in all compressio methods
|
68 | for y in area.position.1..area.end().1 {
|
69 | for channel in &channels.list {
|
70 | if mod_p(y, usize_to_i32(channel.sampling.1)) != 0 { continue; }
|
71 |
|
72 | // this apparently can't be a closure in Rust 1.43 due to borrowing ambiguity
|
73 | let sample_count_x = channel.subsampled_resolution(area.size).0;
|
74 | macro_rules! split_off_write_slice { () => {{
|
75 | let (slice, rest) = write.split_at_mut(sample_count_x);
|
76 | write = rest;
|
77 | slice
|
78 | }}; }
|
79 |
|
80 | let mut previous_pixel: u32 = 0;
|
81 |
|
82 | match channel.sample_type {
|
83 | SampleType::F16 => {
|
84 | let out_byte_tuples = split_off_write_slice!().iter_mut()
|
85 | .zip(split_off_write_slice!());
|
86 |
|
87 | for (out_byte_0, out_byte_1) in out_byte_tuples {
|
88 | let pixel = u16::read_from_native_endian(&mut remaining_bytes).unwrap() as u32;
|
89 | let [byte_1, byte_0] = (pixel.wrapping_sub(previous_pixel) as u16).to_ne_bytes();
|
90 |
|
91 | *out_byte_0 = byte_0;
|
92 | *out_byte_1 = byte_1;
|
93 | previous_pixel = pixel;
|
94 | }
|
95 | },
|
96 |
|
97 | SampleType::U32 => {
|
98 | let out_byte_quadruplets = split_off_write_slice!().iter_mut()
|
99 | .zip(split_off_write_slice!())
|
100 | .zip(split_off_write_slice!())
|
101 | .zip(split_off_write_slice!());
|
102 |
|
103 | for (((out_byte_0, out_byte_1), out_byte_2), out_byte_3) in out_byte_quadruplets {
|
104 | let pixel = u32::read_from_native_endian(&mut remaining_bytes).unwrap();
|
105 | let [byte_3, byte_2, byte_1, byte_0] = pixel.wrapping_sub(previous_pixel).to_ne_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 | for ((out_byte_0, out_byte_1), out_byte_2) in out_byte_triplets {
|
121 | let pixel = f32_to_f24(f32::read_from_native_endian(&mut remaining_bytes).unwrap());
|
122 | let [byte_2, byte_1, byte_0, _] = pixel.wrapping_sub(previous_pixel).to_ne_bytes();
|
123 | previous_pixel = pixel;
|
124 |
|
125 | *out_byte_0 = byte_0;
|
126 | *out_byte_1 = byte_1;
|
127 | *out_byte_2 = byte_2;
|
128 | }
|
129 | },
|
130 | }
|
131 | }
|
132 | }
|
133 |
|
134 | debug_assert_eq!(write.len(), 0, "bytes left after compression" );
|
135 | }
|
136 |
|
137 | Ok(miniz_oxide::deflate::compress_to_vec_zlib(raw.as_slice(), 4))
|
138 | }
|
139 |
|
140 | #[cfg_attr (target_endian = "big" , allow(unused, unreachable_code))]
|
141 | pub fn decompress(channels: &ChannelList, bytes: ByteVec, area: IntegerBounds, expected_byte_size: usize, pedantic: bool) -> Result<ByteVec> {
|
142 | #[cfg (target_endian = "big" )] {
|
143 | return Err(Error::unsupported(
|
144 | "PXR24 decompression method not supported yet on big endian processor architecture"
|
145 | ))
|
146 | }
|
147 |
|
148 | let options = zune_inflate::DeflateOptions::default().set_limit(expected_byte_size).set_size_hint(expected_byte_size);
|
149 | let mut decoder = zune_inflate::DeflateDecoder::new_with_options(&bytes, options);
|
150 | let raw = decoder.decode_zlib()
|
151 | .map_err(|_| Error::invalid("zlib-compressed data malformed" ))?; // TODO share code with zip?
|
152 |
|
153 | let mut read = raw.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)) != 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 > read.len() { return Err(Error::invalid("not enough data" )) }
|
163 | let (samples, rest) = read.split_at(sample_count_x);
|
164 | read = rest;
|
165 | Ok(samples)
|
166 | };
|
167 |
|
168 | let mut pixel_accumulation: u32 = 0;
|
169 |
|
170 | match channel.sample_type {
|
171 | SampleType::F16 => {
|
172 | let sample_byte_pairs = read_sample_line()?.iter()
|
173 | .zip(read_sample_line()?);
|
174 |
|
175 | for (&in_byte_0, &in_byte_1) in sample_byte_pairs {
|
176 | let difference = u16::from_ne_bytes([in_byte_1, in_byte_0]) 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 | for (((&in_byte_0, &in_byte_1), &in_byte_2), &in_byte_3) in sample_byte_quads {
|
189 | let difference = u32::from_ne_bytes([in_byte_3, in_byte_2, in_byte_1, in_byte_0]);
|
190 | pixel_accumulation = pixel_accumulation.overflowing_add(difference).0;
|
191 | out.extend_from_slice(&pixel_accumulation.to_ne_bytes());
|
192 | }
|
193 | },
|
194 |
|
195 | SampleType::F32 => {
|
196 | let sample_byte_triplets = read_sample_line()?.iter()
|
197 | .zip(read_sample_line()?).zip(read_sample_line()?);
|
198 |
|
199 | for ((&in_byte_0, &in_byte_1), &in_byte_2) in sample_byte_triplets {
|
200 | let difference = u32::from_ne_bytes([0, in_byte_2, in_byte_1, in_byte_0]);
|
201 | pixel_accumulation = pixel_accumulation.overflowing_add(difference).0;
|
202 | out.extend_from_slice(&pixel_accumulation.to_ne_bytes());
|
203 | }
|
204 | }
|
205 | }
|
206 | }
|
207 | }
|
208 |
|
209 | if pedantic && !read.is_empty() {
|
210 | return Err(Error::invalid("too much data" ));
|
211 | }
|
212 |
|
213 | Ok(super::convert_little_endian_to_current(out, channels, area))
|
214 | }
|
215 |
|
216 |
|
217 |
|
218 |
|
219 | /// Conversion from 32-bit to 24-bit floating-point numbers.
|
220 | /// Reverse conversion is just a simple 8-bit left shift.
|
221 | pub fn f32_to_f24(float: f32) -> u32 {
|
222 | let bits = float.to_bits();
|
223 |
|
224 | let sign = bits & 0x80000000;
|
225 | let exponent = bits & 0x7f800000;
|
226 | let mantissa = bits & 0x007fffff;
|
227 |
|
228 | let result = if exponent == 0x7f800000 {
|
229 | if mantissa != 0 {
|
230 | // F is a NAN; we preserve the sign bit and
|
231 | // the 15 leftmost bits of the significand,
|
232 | // with one exception: If the 15 leftmost
|
233 | // bits are all zero, the NAN would turn
|
234 | // into an infinity, so we have to set at
|
235 | // least one bit in the significand.
|
236 |
|
237 | let mantissa = mantissa >> 8;
|
238 | (exponent >> 8) | mantissa | if mantissa == 0 { 1 } else { 0 }
|
239 | }
|
240 | else { // F is an infinity.
|
241 | exponent >> 8
|
242 | }
|
243 | }
|
244 | else { // F is finite, round the significand to 15 bits.
|
245 | let result = ((exponent | mantissa) + (mantissa & 0x00000080)) >> 8;
|
246 |
|
247 | if result >= 0x7f8000 {
|
248 | // F was close to FLT_MAX, and the significand was
|
249 | // rounded up, resulting in an exponent overflow.
|
250 | // Avoid the overflow by truncating the significand
|
251 | // instead of rounding it.
|
252 |
|
253 | (exponent | mantissa) >> 8
|
254 | }
|
255 | else {
|
256 | result
|
257 | }
|
258 | };
|
259 |
|
260 | return (sign >> 8) | result;
|
261 | } |