| 1 |
|
| 2 |
|
| 3 |
//! The PIZ compression method is a wavelet compression, |
| 4 |
//! based on the PIZ image format, customized for OpenEXR. |
| 5 |
// inspired by https://github.com/AcademySoftwareFoundation/openexr/blob/master/OpenEXR/IlmImf/ImfPizCompressor.cpp |
| 6 |
|
| 7 |
mod huffman; |
| 8 |
mod wavelet; |
| 9 |
|
| 10 |
use crate::prelude::*; |
| 11 |
use crate::io::Data; |
| 12 |
use crate::meta::attribute::*; |
| 13 |
use crate::compression::{ByteVec, Bytes, mod_p}; |
| 14 |
use crate::error::{usize_to_i32, usize_to_u16}; |
| 15 |
use std::convert::TryFrom; |
| 16 |
|
| 17 |
|
| 18 |
const U16_RANGE: usize = (1_i32 << 16_i32) as usize; |
| 19 |
const BITMAP_SIZE: usize = (U16_RANGE as i32 >> 3_i32) as usize; |
| 20 |
|
| 21 |
#[derive(Debug)] |
| 22 |
struct ChannelData { |
| 23 |
tmp_start_index: usize, |
| 24 |
tmp_end_index: usize, |
| 25 |
|
| 26 |
resolution: Vec2<usize>, |
| 27 |
y_sampling: usize, |
| 28 |
samples_per_pixel: usize, |
| 29 |
} |
| 30 |
|
| 31 |
|
| 32 |
pub fn decompress( |
| 33 |
channels: &ChannelList, |
| 34 |
compressed_le: ByteVec, |
| 35 |
rectangle: IntegerBounds, |
| 36 |
expected_byte_size: usize, // TODO remove expected byte size as it can be computed with `rectangle.size.area() * channels.bytes_per_pixel` |
| 37 |
pedantic: bool |
| 38 |
) -> Result<ByteVec> |
| 39 |
{ |
| 40 |
let expected_u16_count = expected_byte_size / 2; |
| 41 |
debug_assert_eq!(expected_byte_size, rectangle.size.area() * channels.bytes_per_pixel); |
| 42 |
debug_assert!(!channels.list.is_empty()); |
| 43 |
|
| 44 |
if compressed_le.is_empty() { |
| 45 |
return Ok(Vec::new()); |
| 46 |
} |
| 47 |
|
| 48 |
debug_assert_ne!(expected_u16_count, 0); |
| 49 |
|
| 50 |
let mut bitmap = vec![0_u8; BITMAP_SIZE]; // FIXME use bit_vec! |
| 51 |
|
| 52 |
let mut remaining_input_le = compressed_le.as_slice(); |
| 53 |
let min_non_zero = u16::read_le(&mut remaining_input_le)? as usize; |
| 54 |
let max_non_zero = u16::read_le(&mut remaining_input_le)? as usize; |
| 55 |
|
| 56 |
if max_non_zero >= BITMAP_SIZE || min_non_zero >= BITMAP_SIZE { |
| 57 |
return Err(Error::invalid("compression data")); |
| 58 |
} |
| 59 |
|
| 60 |
if min_non_zero <= max_non_zero { |
| 61 |
u8::read_slice_ne(&mut remaining_input_le, &mut bitmap[min_non_zero ..= max_non_zero])?; |
| 62 |
} |
| 63 |
|
| 64 |
let (lookup_table, max_value) = reverse_lookup_table_from_bitmap(&bitmap); |
| 65 |
|
| 66 |
{ |
| 67 |
let length = i32::read_le(&mut remaining_input_le)?; |
| 68 |
if pedantic && length as i64 != remaining_input_le.len() as i64 { |
| 69 |
// TODO length might be smaller than remaining?? |
| 70 |
return Err(Error::invalid("compression data")); |
| 71 |
} |
| 72 |
} |
| 73 |
|
| 74 |
let mut tmp_u16_buffer = huffman::decompress(remaining_input_le, expected_u16_count)?; |
| 75 |
|
| 76 |
let mut channel_data: SmallVec<[ChannelData; 6]> = { |
| 77 |
let mut tmp_read_index = 0; |
| 78 |
|
| 79 |
let channel_data = channels.list.iter().map(|channel| { |
| 80 |
let channel_data = ChannelData { |
| 81 |
tmp_start_index: tmp_read_index, |
| 82 |
tmp_end_index: tmp_read_index, |
| 83 |
y_sampling: channel.sampling.y(), |
| 84 |
resolution: channel.subsampled_resolution(rectangle.size), |
| 85 |
samples_per_pixel: channel.sample_type.bytes_per_sample() / SampleType::F16.bytes_per_sample() |
| 86 |
}; |
| 87 |
|
| 88 |
tmp_read_index += channel_data.resolution.area() * channel_data.samples_per_pixel; |
| 89 |
channel_data |
| 90 |
}).collect(); |
| 91 |
|
| 92 |
debug_assert_eq!(tmp_read_index, expected_u16_count); |
| 93 |
channel_data |
| 94 |
}; |
| 95 |
|
| 96 |
for channel in &channel_data { |
| 97 |
let u16_count = channel.resolution.area() * channel.samples_per_pixel; |
| 98 |
let u16s = &mut tmp_u16_buffer[channel.tmp_start_index .. channel.tmp_start_index + u16_count]; |
| 99 |
|
| 100 |
for offset in 0..channel.samples_per_pixel { // if channel is 32 bit, compress interleaved as two 16 bit values |
| 101 |
wavelet::decode( |
| 102 |
&mut u16s[offset..], |
| 103 |
channel.resolution, |
| 104 |
Vec2(channel.samples_per_pixel, channel.resolution.x() * channel.samples_per_pixel), |
| 105 |
max_value |
| 106 |
)?; |
| 107 |
} |
| 108 |
} |
| 109 |
|
| 110 |
// Expand the pixel data to their original range |
| 111 |
apply_lookup_table(&mut tmp_u16_buffer, &lookup_table); |
| 112 |
|
| 113 |
// let out_buffer_size = (max_scan_line_size * scan_line_count) + 65536 + 8192; // TODO not use expected byte size? |
| 114 |
let mut out = Vec::with_capacity(expected_byte_size); |
| 115 |
|
| 116 |
for y in rectangle.position.y() .. rectangle.end().y() { |
| 117 |
for channel in &mut channel_data { |
| 118 |
if mod_p(y, usize_to_i32(channel.y_sampling, "sampling factor")?) != 0 { |
| 119 |
continue; |
| 120 |
} |
| 121 |
|
| 122 |
let u16s_per_line = channel.resolution.x() * channel.samples_per_pixel; |
| 123 |
let next_tmp_end_index = channel.tmp_end_index + u16s_per_line; |
| 124 |
let values = &tmp_u16_buffer[channel.tmp_end_index .. next_tmp_end_index]; |
| 125 |
channel.tmp_end_index = next_tmp_end_index; |
| 126 |
|
| 127 |
// TODO do not convert endianness for f16-only images |
| 128 |
// see https://github.com/AcademySoftwareFoundation/openexr/blob/3bd93f85bcb74c77255f28cdbb913fdbfbb39dfe/OpenEXR/IlmImf/ImfTiledOutputFile.cpp#L750-L842 |
| 129 |
// We can support uncompressed data in the machine's native format |
| 130 |
// if all image channels are of type HALF, and if the Xdr and the |
| 131 |
// native representations of a half have the same size. |
| 132 |
u16::write_slice_le(&mut out, values).expect("write to in-memory failed"); |
| 133 |
} |
| 134 |
} |
| 135 |
|
| 136 |
for (previous, current) in channel_data.iter().zip(channel_data.iter().skip(1)) { |
| 137 |
debug_assert_eq!(previous.tmp_end_index, current.tmp_start_index); |
| 138 |
} |
| 139 |
|
| 140 |
debug_assert_eq!(channel_data.last().unwrap().tmp_end_index, tmp_u16_buffer.len()); |
| 141 |
debug_assert_eq!(out.len(), expected_byte_size); |
| 142 |
|
| 143 |
// TODO optimize for when all channels are f16! |
| 144 |
// we should be able to omit endianness conversions in that case |
| 145 |
// see https://github.com/AcademySoftwareFoundation/openexr/blob/3bd93f85bcb74c77255f28cdbb913fdbfbb39dfe/OpenEXR/IlmImf/ImfTiledOutputFile.cpp#L750-L842 |
| 146 |
super::convert_little_endian_to_current(out, channels, rectangle) |
| 147 |
} |
| 148 |
|
| 149 |
|
| 150 |
|
| 151 |
pub fn compress( |
| 152 |
channels: &ChannelList, |
| 153 |
uncompressed_ne: ByteVec, |
| 154 |
rectangle: IntegerBounds |
| 155 |
) -> Result<ByteVec> |
| 156 |
{ |
| 157 |
if uncompressed_ne.is_empty() { |
| 158 |
return Ok(Vec::new()); |
| 159 |
} |
| 160 |
|
| 161 |
// TODO do not convert endianness for f16-only images twice |
| 162 |
// see https://github.com/AcademySoftwareFoundation/openexr/blob/3bd93f85bcb74c77255f28cdbb913fdbfbb39dfe/OpenEXR/IlmImf/ImfTiledOutputFile.cpp#L750-L842 |
| 163 |
let uncompressed_le = super::convert_current_to_little_endian(uncompressed_ne, channels, rectangle)?; |
| 164 |
let uncompressed_le = uncompressed_le.as_slice();// TODO no alloc |
| 165 |
|
| 166 |
let mut tmp = vec![0_u16; uncompressed_le.len() / 2 ]; |
| 167 |
let mut channel_data: SmallVec<[ChannelData; 6]> = { |
| 168 |
let mut tmp_end_index = 0; |
| 169 |
|
| 170 |
let vec = channels.list.iter().map(|channel| { |
| 171 |
let number_samples = channel.subsampled_resolution(rectangle.size); |
| 172 |
let byte_size = channel.sample_type.bytes_per_sample() / SampleType::F16.bytes_per_sample(); |
| 173 |
let byte_count = byte_size * number_samples.area(); |
| 174 |
|
| 175 |
let channel = ChannelData { |
| 176 |
tmp_end_index, |
| 177 |
tmp_start_index: tmp_end_index, |
| 178 |
y_sampling: channel.sampling.y(), |
| 179 |
resolution: number_samples, |
| 180 |
samples_per_pixel: byte_size, |
| 181 |
}; |
| 182 |
|
| 183 |
tmp_end_index += byte_count; |
| 184 |
channel |
| 185 |
}).collect(); |
| 186 |
|
| 187 |
debug_assert_eq!(tmp_end_index, tmp.len()); |
| 188 |
vec |
| 189 |
}; |
| 190 |
|
| 191 |
let mut remaining_uncompressed_bytes_le = uncompressed_le; |
| 192 |
for y in rectangle.position.y() .. rectangle.end().y() { |
| 193 |
for channel in &mut channel_data { |
| 194 |
if mod_p(y, usize_to_i32(channel.y_sampling, "sampling")?) != 0 { continue; } |
| 195 |
let u16s_per_line = channel.resolution.x() * channel.samples_per_pixel; |
| 196 |
let next_tmp_end_index = channel.tmp_end_index + u16s_per_line; |
| 197 |
let target = &mut tmp[channel.tmp_end_index .. next_tmp_end_index]; |
| 198 |
channel.tmp_end_index = next_tmp_end_index; |
| 199 |
|
| 200 |
// TODO do not convert endianness for f16-only images |
| 201 |
// see https://github.com/AcademySoftwareFoundation/openexr/blob/3bd93f85bcb74c77255f28cdbb913fdbfbb39dfe/OpenEXR/IlmImf/ImfTiledOutputFile.cpp#L750-L842 |
| 202 |
// We can support uncompressed data in the machine's native format |
| 203 |
// if all image channels are of type HALF, and if the Xdr and the |
| 204 |
// native representations of a half have the same size. |
| 205 |
u16::read_slice_le(&mut remaining_uncompressed_bytes_le, target).expect("in-memory read failed"); |
| 206 |
} |
| 207 |
} |
| 208 |
|
| 209 |
|
| 210 |
let (min_non_zero, max_non_zero, bitmap) = bitmap_from_data(&tmp); |
| 211 |
let (max_value, table) = forward_lookup_table_from_bitmap(&bitmap); |
| 212 |
apply_lookup_table(&mut tmp, &table); |
| 213 |
|
| 214 |
let mut piz_compressed = Vec::with_capacity(uncompressed_le.len() / 2); |
| 215 |
u16::try_from(min_non_zero)?.write_le(&mut piz_compressed)?; |
| 216 |
u16::try_from(max_non_zero)?.write_le(&mut piz_compressed)?; |
| 217 |
|
| 218 |
if min_non_zero <= max_non_zero { |
| 219 |
piz_compressed.extend_from_slice(&bitmap[min_non_zero ..= max_non_zero]); |
| 220 |
} |
| 221 |
|
| 222 |
for channel in channel_data { |
| 223 |
for offset in 0 .. channel.samples_per_pixel { |
| 224 |
wavelet::encode( |
| 225 |
&mut tmp[channel.tmp_start_index + offset .. channel.tmp_end_index], |
| 226 |
channel.resolution, |
| 227 |
Vec2(channel.samples_per_pixel, channel.resolution.x() * channel.samples_per_pixel), |
| 228 |
max_value |
| 229 |
)?; |
| 230 |
} |
| 231 |
} |
| 232 |
|
| 233 |
let huffman_compressed: Vec<u8> = huffman::compress(&tmp)?; |
| 234 |
u8::write_i32_sized_slice_le(&mut piz_compressed, &huffman_compressed).expect("in-memory write failed"); |
| 235 |
|
| 236 |
Ok(piz_compressed) |
| 237 |
} |
| 238 |
|
| 239 |
|
| 240 |
pub fn bitmap_from_data(data: &[u16]) -> (usize, usize, Vec<u8>) { |
| 241 |
let mut bitmap = vec![0_u8; BITMAP_SIZE]; |
| 242 |
|
| 243 |
for value in data { |
| 244 |
bitmap[*value as usize >> 3] |= 1 << (*value as u8 & 7); |
| 245 |
} |
| 246 |
|
| 247 |
bitmap[0] = bitmap[0] & !1; // zero is not explicitly stored in the bitmap; we assume that the data always contain zeroes |
| 248 |
|
| 249 |
let min_index = bitmap.iter().position(|&value| value != 0); |
| 250 |
let max_index = min_index.map(|min| // only if min was found |
| 251 |
min + bitmap[min..].iter().rposition(|&value| value != 0).expect("[min] not found") |
| 252 |
); |
| 253 |
|
| 254 |
(min_index.unwrap_or(0), max_index.unwrap_or(0), bitmap) |
| 255 |
} |
| 256 |
|
| 257 |
pub fn forward_lookup_table_from_bitmap(bitmap: &[u8]) -> (u16, Vec<u16>) { |
| 258 |
debug_assert_eq!(bitmap.len(), BITMAP_SIZE); |
| 259 |
|
| 260 |
let mut table = vec![0_u16; U16_RANGE]; |
| 261 |
let mut count = 0_usize; |
| 262 |
|
| 263 |
for (index, entry) in table.iter_mut().enumerate() { |
| 264 |
if index == 0 || bitmap[index >> 3] as usize & (1 << (index & 7)) != 0 { |
| 265 |
*entry = usize_to_u16(count, "piz entry").unwrap(); |
| 266 |
count += 1; |
| 267 |
} |
| 268 |
} |
| 269 |
|
| 270 |
(usize_to_u16(count - 1, "piz count").unwrap(), table) |
| 271 |
} |
| 272 |
|
| 273 |
fn reverse_lookup_table_from_bitmap(bitmap: Bytes<'_>) -> (Vec<u16>, u16) { |
| 274 |
let mut table = Vec::with_capacity(U16_RANGE); |
| 275 |
|
| 276 |
for index in 0 .. U16_RANGE { // cannot use iter because filter removes capacity sizehint |
| 277 |
if index == 0 || ((bitmap[index >> 3] as usize & (1 << (index & 7))) != 0) { |
| 278 |
table.push(usize_to_u16(index, "table entry").unwrap()); |
| 279 |
} |
| 280 |
} |
| 281 |
|
| 282 |
debug_assert!(!table.is_empty()); |
| 283 |
let max_value = usize_to_u16(table.len() - 1, "table size").unwrap(); |
| 284 |
|
| 285 |
// fill remaining up to u16 range |
| 286 |
assert!(table.len() <= U16_RANGE); |
| 287 |
table.resize(U16_RANGE, 0); |
| 288 |
|
| 289 |
(table, max_value) |
| 290 |
} |
| 291 |
|
| 292 |
fn apply_lookup_table(data: &mut [u16], table: &[u16]) { |
| 293 |
for data in data { |
| 294 |
*data = table[*data as usize]; |
| 295 |
} |
| 296 |
} |
| 297 |
|
| 298 |
#[cfg(test)] |
| 299 |
mod test { |
| 300 |
use crate::prelude::*; |
| 301 |
use crate::compression::ByteVec; |
| 302 |
use crate::compression::piz; |
| 303 |
use crate::meta::attribute::*; |
| 304 |
|
| 305 |
fn test_roundtrip_noise_with(channels: ChannelList, rectangle: IntegerBounds){ |
| 306 |
let pixel_bytes: ByteVec = (0 .. 37).map(|_| rand::random()).collect::<Vec<u8>>().into_iter() |
| 307 |
.cycle().take(channels.bytes_per_pixel * rectangle.size.area()) |
| 308 |
.collect(); |
| 309 |
|
| 310 |
let compressed = piz::compress(&channels, pixel_bytes.clone(), rectangle).unwrap(); |
| 311 |
let decompressed = piz::decompress(&channels, compressed, rectangle, pixel_bytes.len(), true).unwrap(); |
| 312 |
|
| 313 |
assert_eq!(pixel_bytes, decompressed); |
| 314 |
} |
| 315 |
|
| 316 |
|
| 317 |
#[test] |
| 318 |
fn roundtrip_any_sample_type(){ |
| 319 |
for &sample_type in &[SampleType::F16, SampleType::F32, SampleType::U32] { |
| 320 |
let channel = ChannelDescription { |
| 321 |
sample_type, |
| 322 |
|
| 323 |
name: Default::default(), |
| 324 |
quantize_linearly: false, |
| 325 |
sampling: Vec2(1,1) |
| 326 |
}; |
| 327 |
|
| 328 |
let channels = ChannelList::new(smallvec![ channel.clone(), channel ]); |
| 329 |
|
| 330 |
let rectangle = IntegerBounds { |
| 331 |
position: Vec2(-30, 100), |
| 332 |
size: Vec2(1080, 720), |
| 333 |
}; |
| 334 |
|
| 335 |
test_roundtrip_noise_with(channels, rectangle); |
| 336 |
} |
| 337 |
} |
| 338 |
|
| 339 |
#[test] |
| 340 |
fn roundtrip_two_channels(){ |
| 341 |
let channel = ChannelDescription { |
| 342 |
sample_type: SampleType::F16, |
| 343 |
|
| 344 |
name: Default::default(), |
| 345 |
quantize_linearly: false, |
| 346 |
sampling: Vec2(1,1) |
| 347 |
}; |
| 348 |
|
| 349 |
let channel2 = ChannelDescription { |
| 350 |
sample_type: SampleType::F32, |
| 351 |
|
| 352 |
name: Default::default(), |
| 353 |
quantize_linearly: false, |
| 354 |
sampling: Vec2(1,1) |
| 355 |
}; |
| 356 |
|
| 357 |
let channels = ChannelList::new(smallvec![ channel, channel2 ]); |
| 358 |
|
| 359 |
let rectangle = IntegerBounds { |
| 360 |
position: Vec2(-3, 1), |
| 361 |
size: Vec2(223, 3132), |
| 362 |
}; |
| 363 |
|
| 364 |
test_roundtrip_noise_with(channels, rectangle); |
| 365 |
} |
| 366 |
|
| 367 |
|
| 368 |
|
| 369 |
#[test] |
| 370 |
fn roundtrip_seven_channels(){ |
| 371 |
let channels = ChannelList::new(smallvec![ |
| 372 |
ChannelDescription { |
| 373 |
sample_type: SampleType::F32, |
| 374 |
|
| 375 |
name: Default::default(), |
| 376 |
quantize_linearly: false, |
| 377 |
sampling: Vec2(1,1) |
| 378 |
}, |
| 379 |
|
| 380 |
ChannelDescription { |
| 381 |
sample_type: SampleType::F32, |
| 382 |
|
| 383 |
name: Default::default(), |
| 384 |
quantize_linearly: false, |
| 385 |
sampling: Vec2(1,1) |
| 386 |
}, |
| 387 |
|
| 388 |
ChannelDescription { |
| 389 |
sample_type: SampleType::F32, |
| 390 |
|
| 391 |
name: Default::default(), |
| 392 |
quantize_linearly: false, |
| 393 |
sampling: Vec2(1,1) |
| 394 |
}, |
| 395 |
|
| 396 |
ChannelDescription { |
| 397 |
sample_type: SampleType::F16, |
| 398 |
|
| 399 |
name: Default::default(), |
| 400 |
quantize_linearly: false, |
| 401 |
sampling: Vec2(1,1) |
| 402 |
}, |
| 403 |
|
| 404 |
ChannelDescription { |
| 405 |
sample_type: SampleType::F32, |
| 406 |
|
| 407 |
name: Default::default(), |
| 408 |
quantize_linearly: false, |
| 409 |
sampling: Vec2(1,1) |
| 410 |
}, |
| 411 |
|
| 412 |
ChannelDescription { |
| 413 |
sample_type: SampleType::F32, |
| 414 |
|
| 415 |
name: Default::default(), |
| 416 |
quantize_linearly: false, |
| 417 |
sampling: Vec2(1,1) |
| 418 |
}, |
| 419 |
|
| 420 |
ChannelDescription { |
| 421 |
sample_type: SampleType::U32, |
| 422 |
|
| 423 |
name: Default::default(), |
| 424 |
quantize_linearly: false, |
| 425 |
sampling: Vec2(1,1) |
| 426 |
}, |
| 427 |
]); |
| 428 |
|
| 429 |
let rectangle = IntegerBounds { |
| 430 |
position: Vec2(-3, 1), |
| 431 |
size: Vec2(1323, 132), |
| 432 |
}; |
| 433 |
|
| 434 |
test_roundtrip_noise_with(channels, rectangle); |
| 435 |
} |
| 436 |
|
| 437 |
} |