| 1 |
/* |
| 2 |
* Copyright (c) 2023. |
| 3 |
* |
| 4 |
* This software is free software; |
| 5 |
* |
| 6 |
* You can redistribute it or modify it under terms of the MIT, Apache License or Zlib license |
| 7 |
*/ |
| 8 |
|
| 9 |
//! Routines for IDCT |
| 10 |
//! |
| 11 |
//! Essentially we provide 2 routines for IDCT, a scalar implementation and a not super optimized |
| 12 |
//! AVX2 one, i'll talk about them here. |
| 13 |
//! |
| 14 |
//! There are 2 reasons why we have the avx one |
| 15 |
//! 1. No one compiles with -C target-features=avx2 hence binaries won't probably take advantage(even |
| 16 |
//! if it exists). |
| 17 |
//! 2. AVX employs zero short circuit in a way the scalar code cannot employ it. |
| 18 |
//! - AVX does this by checking for MCU's whose 63 AC coefficients are zero and if true, it writes |
| 19 |
//! values directly, if false, it goes the long way of calculating. |
| 20 |
//! - Although this can be trivially implemented in the scalar version, it generates code |
| 21 |
//! I'm not happy width(scalar version that basically loops and that is too many branches for me) |
| 22 |
//! The avx one does a better job of using bitwise or's with (`_mm256_or_si256`) which is magnitudes of faster |
| 23 |
//! than anything I could come up with |
| 24 |
//! |
| 25 |
//! The AVX code also has some cool transpose_u16 instructions which look so complicated to be cool |
| 26 |
//! (spoiler alert, i barely understand how it works, that's why I credited the owner). |
| 27 |
//! |
| 28 |
#![allow( |
| 29 |
clippy::excessive_precision, |
| 30 |
clippy::unreadable_literal, |
| 31 |
clippy::module_name_repetitions, |
| 32 |
unused_parens, |
| 33 |
clippy::wildcard_imports |
| 34 |
)] |
| 35 |
|
| 36 |
use zune_core::log::debug; |
| 37 |
use zune_core::options::DecoderOptions; |
| 38 |
|
| 39 |
use crate::decoder::IDCTPtr; |
| 40 |
use crate::idct::scalar::{idct_int, idct_int_1x1}; |
| 41 |
|
| 42 |
#[cfg(feature = "x86")] |
| 43 |
pub mod avx2; |
| 44 |
#[cfg(feature = "neon")] |
| 45 |
pub mod neon; |
| 46 |
|
| 47 |
pub mod scalar; |
| 48 |
|
| 49 |
/// Choose an appropriate IDCT function |
| 50 |
#[allow(unused_variables)] |
| 51 |
pub fn choose_idct_func(options: &DecoderOptions) -> IDCTPtr { |
| 52 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 53 |
#[cfg(feature = "x86")] |
| 54 |
{ |
| 55 |
if options.use_avx2() { |
| 56 |
debug!("Using vector integer IDCT"); |
| 57 |
return |a: &mut [i32; 64], b: &mut [i16], c: usize| { |
| 58 |
// SAFETY: `options.use_avx2()` only returns true if avx2 is supported. |
| 59 |
unsafe { avx2::idct_avx2(a,b,c) } |
| 60 |
}; |
| 61 |
} |
| 62 |
} |
| 63 |
#[cfg(target_arch = "aarch64")] |
| 64 |
#[cfg(feature = "neon")] |
| 65 |
{ |
| 66 |
if options.use_neon() { |
| 67 |
debug!("Using vector integer IDCT"); |
| 68 |
return |a: &mut [i32; 64], b: &mut [i16], c: usize| { |
| 69 |
// SAFETY: `options.use_neon()` only returns true if neon is supported. |
| 70 |
unsafe { neon::idct_neon(a,b,c) } |
| 71 |
}; |
| 72 |
} |
| 73 |
} |
| 74 |
debug!("Using scalar integer IDCT"); |
| 75 |
// use generic one |
| 76 |
return idct_int; |
| 77 |
} |
| 78 |
|
| 79 |
/// Choose a function to implement 4x4 IDCT. |
| 80 |
/// |
| 81 |
/// These functions get the same input but have an extra contract: Only the first 4x4 block of |
| 82 |
/// coefficients are non-zero. All other entries are zeroed. |
| 83 |
/// |
| 84 |
/// **The callee must uphold that contract on return** |
| 85 |
pub fn choose_idct_4x4_func(_options: &DecoderOptions) -> IDCTPtr { |
| 86 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 87 |
#[cfg(feature = "x86")] |
| 88 |
{ |
| 89 |
if _options.use_avx2() { |
| 90 |
debug!("Using vector integer IDCT"); |
| 91 |
return |a: &mut [i32; 64], b: &mut [i16], c: usize| { |
| 92 |
// SAFETY: `options.use_avx2()` only returns true if avx2 is supported. |
| 93 |
unsafe { avx2::idct_avx2_4x4(a,b,c) } |
| 94 |
}; |
| 95 |
} |
| 96 |
} |
| 97 |
|
| 98 |
scalar::idct4x4 |
| 99 |
} |
| 100 |
|
| 101 |
pub fn choose_idct_1x1_func(_: &DecoderOptions) -> IDCTPtr { |
| 102 |
// These are simple stores, no alternative implementation for now |
| 103 |
idct_int_1x1 |
| 104 |
} |
| 105 |
|
| 106 |
#[cfg(test)] |
| 107 |
#[allow(unreachable_code)] |
| 108 |
#[allow(dead_code)] |
| 109 |
mod tests { |
| 110 |
use super::*; |
| 111 |
|
| 112 |
#[test] |
| 113 |
fn idct_test0() { |
| 114 |
let stride = 8; |
| 115 |
let mut coeff = [10; 64]; |
| 116 |
let mut coeff2 = [10; 64]; |
| 117 |
let mut output_scalar = [0; 64]; |
| 118 |
let mut output_vector = [0; 64]; |
| 119 |
let idct_func = choose_idct_func(&DecoderOptions::new_fast()); |
| 120 |
idct_func(&mut coeff, &mut output_vector, stride); |
| 121 |
idct_int(&mut coeff2, &mut output_scalar, stride); |
| 122 |
assert_eq!(output_scalar, output_vector, "IDCT and scalar do not match"); |
| 123 |
} |
| 124 |
|
| 125 |
#[test] |
| 126 |
fn do_idct_test1() { |
| 127 |
let stride = 8; |
| 128 |
let mut coeff = [14; 64]; |
| 129 |
let mut coeff2 = [14; 64]; |
| 130 |
let mut output_scalar = [0; 64]; |
| 131 |
let mut output_vector = [0; 64]; |
| 132 |
let idct_func = choose_idct_func(&DecoderOptions::new_fast()); |
| 133 |
idct_func(&mut coeff, &mut output_vector, stride); |
| 134 |
idct_int(&mut coeff2, &mut output_scalar, stride); |
| 135 |
assert_eq!(output_scalar, output_vector, "IDCT and scalar do not match"); |
| 136 |
} |
| 137 |
|
| 138 |
#[test] |
| 139 |
fn do_idct_test2() { |
| 140 |
let stride = 8; |
| 141 |
let mut coeff = [0; 64]; |
| 142 |
coeff[0] = 255; |
| 143 |
coeff[63] = -256; |
| 144 |
let mut coeff2 = coeff; |
| 145 |
let mut output_scalar = [0; 64]; |
| 146 |
let mut output_vector = [0; 64]; |
| 147 |
let idct_func = choose_idct_func(&DecoderOptions::new_fast()); |
| 148 |
idct_func(&mut coeff, &mut output_vector, stride); |
| 149 |
idct_int(&mut coeff2, &mut output_scalar, stride); |
| 150 |
assert_eq!(output_scalar, output_vector, "IDCT and scalar do not match"); |
| 151 |
} |
| 152 |
|
| 153 |
#[test] |
| 154 |
fn do_idct_zeros() { |
| 155 |
let stride = 8; |
| 156 |
let mut coeff = [0; 64]; |
| 157 |
let mut coeff2 = [0; 64]; |
| 158 |
let mut output_scalar = [0; 64]; |
| 159 |
let mut output_vector = [0; 64]; |
| 160 |
let idct_func = choose_idct_func(&DecoderOptions::new_fast()); |
| 161 |
idct_func(&mut coeff, &mut output_vector, stride); |
| 162 |
idct_int(&mut coeff2, &mut output_scalar, stride); |
| 163 |
assert_eq!(output_scalar, output_vector, "IDCT and scalar do not match"); |
| 164 |
} |
| 165 |
|
| 166 |
#[test] |
| 167 |
fn idct_4x4() { |
| 168 |
#[rustfmt::skip] |
| 169 |
const A: [i32; 32] = [ |
| 170 |
-254, -7, 0, 0, 0, 0, 0, 0, |
| 171 |
7, 0, -30, 32, 0, 0, 0, 0, |
| 172 |
7, 0, -30, 32, 0, 0, 0, 0, |
| 173 |
7, 0, -30, 32, 0, 0, 0, 0, |
| 174 |
]; |
| 175 |
|
| 176 |
let v: Vec<IDCTPtr> = vec![ |
| 177 |
choose_idct_func(&DecoderOptions::new_safe()), |
| 178 |
choose_idct_4x4_func(&DecoderOptions::new_safe()), |
| 179 |
choose_idct_func(&DecoderOptions::new_fast()), |
| 180 |
choose_idct_4x4_func(&DecoderOptions::new_fast()), |
| 181 |
]; |
| 182 |
let dct_names = vec![ |
| 183 |
"safe idct", |
| 184 |
"safe idct 4x4", |
| 185 |
"fast idct", |
| 186 |
"fast idct 4x4", |
| 187 |
]; |
| 188 |
|
| 189 |
let mut color = vec![]; |
| 190 |
|
| 191 |
for idct in v { |
| 192 |
let mut a = [0i32; 64]; |
| 193 |
a[..32].copy_from_slice(&A); |
| 194 |
let mut b = [0i16; 64]; |
| 195 |
|
| 196 |
idct(&mut a, &mut b, 8); |
| 197 |
|
| 198 |
color.push(b); |
| 199 |
} |
| 200 |
|
| 201 |
for (wnd, name) in color.windows(2).zip(&dct_names) { |
| 202 |
let [a, b] = wnd else { unreachable!() }; |
| 203 |
assert_eq!(a, b, "{name}"); |
| 204 |
} |
| 205 |
} |
| 206 |
} |