| 1 |
#[cfg(asm_neon)] |
| 2 |
mod simd_neon; |
| 3 |
#[cfg(asm_x86_64)] |
| 4 |
mod simd_x86; |
| 5 |
#[cfg(test)] |
| 6 |
mod tests; |
| 7 |
|
| 8 |
#[cfg(not(any(asm_x86_64, asm_neon)))] |
| 9 |
use rust::*; |
| 10 |
#[cfg(asm_neon)] |
| 11 |
use simd_neon::*; |
| 12 |
#[cfg(asm_x86_64)] |
| 13 |
use simd_x86::*; |
| 14 |
use v_frame::pixel::Pixel; |
| 15 |
|
| 16 |
use super::{block::BlockSize, plane::PlaneRegion}; |
| 17 |
use crate::cpu::CpuFeatureLevel; |
| 18 |
|
| 19 |
mod rust { |
| 20 |
use v_frame::{ |
| 21 |
math::msb, |
| 22 |
pixel::{CastFromPrimitive, Pixel}, |
| 23 |
}; |
| 24 |
|
| 25 |
use crate::{ |
| 26 |
cpu::CpuFeatureLevel, |
| 27 |
data::{ |
| 28 |
hadamard::{hadamard4x4, hadamard8x8}, |
| 29 |
plane::{Area, PlaneRegion, Rect}, |
| 30 |
sad::get_sad, |
| 31 |
}, |
| 32 |
}; |
| 33 |
|
| 34 |
/// Sum of absolute transformed differences over a block. |
| 35 |
/// w and h can be at most 128, the size of the largest block. |
| 36 |
/// Use the sum of 4x4 and 8x8 hadamard transforms for the transform, but |
| 37 |
/// revert to sad on edges when these transforms do not fit into w and h. |
| 38 |
/// 4x4 transforms instead of 8x8 transforms when width or height < 8. |
| 39 |
#[cfg_attr(all(asm_x86_64, target_feature = "avx2"), cold)] |
| 40 |
#[cfg_attr(asm_neon, cold)] |
| 41 |
pub(super) fn get_satd_internal<T: Pixel>( |
| 42 |
plane_org: &PlaneRegion<'_, T>, |
| 43 |
plane_ref: &PlaneRegion<'_, T>, |
| 44 |
w: usize, |
| 45 |
h: usize, |
| 46 |
bit_depth: usize, |
| 47 |
cpu: CpuFeatureLevel, |
| 48 |
) -> u32 { |
| 49 |
assert!(w <= 128 && h <= 128); |
| 50 |
assert!(plane_org.rect().width >= w && plane_org.rect().height >= h); |
| 51 |
assert!(plane_ref.rect().width >= w && plane_ref.rect().height >= h); |
| 52 |
|
| 53 |
// Size of hadamard transform should be 4x4 or 8x8 |
| 54 |
// 4x* and *x4 use 4x4 and all other use 8x8 |
| 55 |
let size: usize = w.min(h).min(8); |
| 56 |
let tx2d = if size == 4 { hadamard4x4 } else { hadamard8x8 }; |
| 57 |
|
| 58 |
let mut sum: u64 = 0; |
| 59 |
|
| 60 |
// Loop over chunks the size of the chosen transform |
| 61 |
for chunk_y in (0..h).step_by(size) { |
| 62 |
let chunk_h = (h - chunk_y).min(size); |
| 63 |
for chunk_x in (0..w).step_by(size) { |
| 64 |
let chunk_w = (w - chunk_x).min(size); |
| 65 |
let chunk_area = Area::Rect(Rect { |
| 66 |
x: chunk_x as isize, |
| 67 |
y: chunk_y as isize, |
| 68 |
width: chunk_w, |
| 69 |
height: chunk_h, |
| 70 |
}); |
| 71 |
let chunk_org = plane_org.subregion(chunk_area); |
| 72 |
let chunk_ref = plane_ref.subregion(chunk_area); |
| 73 |
|
| 74 |
// Revert to sad on edge blocks (frame edges) |
| 75 |
if chunk_w != size || chunk_h != size { |
| 76 |
sum += get_sad(&chunk_org, &chunk_ref, chunk_w, chunk_h, bit_depth, cpu) as u64; |
| 77 |
continue; |
| 78 |
} |
| 79 |
|
| 80 |
let buf: &mut [i32] = &mut [0; 8 * 8][..size * size]; |
| 81 |
|
| 82 |
// Move the difference of the transforms to a buffer |
| 83 |
for (row_diff, (row_org, row_ref)) in buf |
| 84 |
.chunks_mut(size) |
| 85 |
.zip(chunk_org.rows_iter().zip(chunk_ref.rows_iter())) |
| 86 |
{ |
| 87 |
for (diff, (a, b)) in |
| 88 |
row_diff.iter_mut().zip(row_org.iter().zip(row_ref.iter())) |
| 89 |
{ |
| 90 |
*diff = i32::cast_from(*a) - i32::cast_from(*b); |
| 91 |
} |
| 92 |
} |
| 93 |
|
| 94 |
// Perform the hadamard transform on the differences |
| 95 |
// SAFETY: A sufficient number elements exist for the size of the transform. |
| 96 |
unsafe { |
| 97 |
tx2d(buf); |
| 98 |
} |
| 99 |
|
| 100 |
// Sum the absolute values of the transformed differences |
| 101 |
sum += buf.iter().map(|a| a.unsigned_abs() as u64).sum::<u64>(); |
| 102 |
} |
| 103 |
} |
| 104 |
|
| 105 |
// Normalize the results |
| 106 |
let ln = msb(size as i32) as u64; |
| 107 |
((sum + (1 << ln >> 1)) >> ln) as u32 |
| 108 |
} |
| 109 |
} |
| 110 |
|
| 111 |
// BlockSize::BLOCK_SIZES.next_power_of_two() |
| 112 |
const DIST_FNS_LENGTH: usize = 32; |
| 113 |
|
| 114 |
const fn to_index(bsize: BlockSize) -> usize { |
| 115 |
bsize as usize & (DIST_FNS_LENGTH - 1) |
| 116 |
} |
| 117 |
|
| 118 |
pub(crate) fn get_satd<T: Pixel>( |
| 119 |
src: &PlaneRegion<'_, T>, |
| 120 |
dst: &PlaneRegion<'_, T>, |
| 121 |
w: usize, |
| 122 |
h: usize, |
| 123 |
bit_depth: usize, |
| 124 |
cpu: CpuFeatureLevel, |
| 125 |
) -> u32 { |
| 126 |
get_satd_internal(src, dst, w, h, bit_depth, cpu) |
| 127 |
} |