| 1 |
// Copyright (c) 2022-2023, The rav1e contributors. All rights reserved |
| 2 |
// |
| 3 |
// This source code is subject to the terms of the BSD 2 Clause License and |
| 4 |
// the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
| 5 |
// was not distributed with this source code in the LICENSE file, you can |
| 6 |
// obtain it at www.aomedia.org/license/software. If the Alliance for Open |
| 7 |
// Media Patent License 1.0 was not distributed with this source code in the |
| 8 |
// PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
| 9 |
|
| 10 |
use crate::activity::apply_ssim_boost; |
| 11 |
use crate::cpu_features::CpuFeatureLevel; |
| 12 |
use crate::dist::*; |
| 13 |
use crate::tiling::PlaneRegion; |
| 14 |
use crate::util::Pixel; |
| 15 |
use crate::util::PixelType; |
| 16 |
use std::arch::x86_64::*; |
| 17 |
|
| 18 |
type CdefDistKernelFn = unsafe extern fn( |
| 19 |
src: *const u8, |
| 20 |
src_stride: isize, |
| 21 |
dst: *const u8, |
| 22 |
dst_stride: isize, |
| 23 |
ret_ptr: *mut u32, |
| 24 |
); |
| 25 |
|
| 26 |
type CdefDistKernelHBDFn = unsafe fn( |
| 27 |
src: *const u16, |
| 28 |
src_stride: isize, |
| 29 |
dst: *const u16, |
| 30 |
dst_stride: isize, |
| 31 |
) -> (u32, u32, u32); |
| 32 |
|
| 33 |
extern { |
| 34 |
fn rav1e_cdef_dist_kernel_4x4_sse2( |
| 35 |
src: *const u8, src_stride: isize, dst: *const u8, dst_stride: isize, |
| 36 |
ret_ptr: *mut u32, |
| 37 |
); |
| 38 |
fn rav1e_cdef_dist_kernel_4x8_sse2( |
| 39 |
src: *const u8, src_stride: isize, dst: *const u8, dst_stride: isize, |
| 40 |
ret_ptr: *mut u32, |
| 41 |
); |
| 42 |
fn rav1e_cdef_dist_kernel_8x4_sse2( |
| 43 |
src: *const u8, src_stride: isize, dst: *const u8, dst_stride: isize, |
| 44 |
ret_ptr: *mut u32, |
| 45 |
); |
| 46 |
fn rav1e_cdef_dist_kernel_8x8_sse2( |
| 47 |
src: *const u8, src_stride: isize, dst: *const u8, dst_stride: isize, |
| 48 |
ret_ptr: *mut u32, |
| 49 |
); |
| 50 |
} |
| 51 |
|
| 52 |
/// # Panics |
| 53 |
/// |
| 54 |
/// - If in `check_asm` mode, panics on mismatch between native and ASM results. |
| 55 |
#[allow(clippy::let_and_return)] |
| 56 |
pub fn cdef_dist_kernel<T: Pixel>( |
| 57 |
src: &PlaneRegion<'_, T>, dst: &PlaneRegion<'_, T>, w: usize, h: usize, |
| 58 |
bit_depth: usize, cpu: CpuFeatureLevel, |
| 59 |
) -> u32 { |
| 60 |
debug_assert!(src.plane_cfg.xdec == 0); |
| 61 |
debug_assert!(src.plane_cfg.ydec == 0); |
| 62 |
debug_assert!(dst.plane_cfg.xdec == 0); |
| 63 |
debug_assert!(dst.plane_cfg.ydec == 0); |
| 64 |
|
| 65 |
// Limit kernel to 8x8 |
| 66 |
debug_assert!(w <= 8); |
| 67 |
debug_assert!(h <= 8); |
| 68 |
|
| 69 |
let call_rust = |
| 70 |
|| -> u32 { rust::cdef_dist_kernel(dst, src, w, h, bit_depth, cpu) }; |
| 71 |
#[cfg(feature = "check_asm")] |
| 72 |
let ref_dist = call_rust(); |
| 73 |
|
| 74 |
let (svar, dvar, sse) = match T::type_enum() { |
| 75 |
PixelType::U8 => { |
| 76 |
if let Some(func) = |
| 77 |
CDEF_DIST_KERNEL_FNS[cpu.as_index()][kernel_fn_index(w, h)] |
| 78 |
{ |
| 79 |
let mut ret_buf = [0u32; 3]; |
| 80 |
// SAFETY: Calls Assembly code. |
| 81 |
unsafe { |
| 82 |
func( |
| 83 |
src.data_ptr() as *const _, |
| 84 |
T::to_asm_stride(src.plane_cfg.stride), |
| 85 |
dst.data_ptr() as *const _, |
| 86 |
T::to_asm_stride(dst.plane_cfg.stride), |
| 87 |
ret_buf.as_mut_ptr(), |
| 88 |
) |
| 89 |
} |
| 90 |
|
| 91 |
(ret_buf[0], ret_buf[1], ret_buf[2]) |
| 92 |
} else { |
| 93 |
return call_rust(); |
| 94 |
} |
| 95 |
} |
| 96 |
PixelType::U16 => { |
| 97 |
if let Some(func) = |
| 98 |
CDEF_DIST_KERNEL_HBD_FNS[cpu.as_index()][kernel_fn_index(w, h)] |
| 99 |
{ |
| 100 |
// SAFETY: Calls Assembly code. |
| 101 |
unsafe { |
| 102 |
func( |
| 103 |
src.data_ptr() as *const _, |
| 104 |
T::to_asm_stride(src.plane_cfg.stride), |
| 105 |
dst.data_ptr() as *const _, |
| 106 |
T::to_asm_stride(dst.plane_cfg.stride), |
| 107 |
) |
| 108 |
} |
| 109 |
} else { |
| 110 |
return call_rust(); |
| 111 |
} |
| 112 |
} |
| 113 |
}; |
| 114 |
|
| 115 |
let dist = apply_ssim_boost(sse, svar, dvar, bit_depth); |
| 116 |
#[cfg(feature = "check_asm")] |
| 117 |
assert_eq!( |
| 118 |
dist, ref_dist, |
| 119 |
"CDEF Distortion {}x{}: Assembly doesn't match reference code.", |
| 120 |
w, h |
| 121 |
); |
| 122 |
|
| 123 |
dist |
| 124 |
} |
| 125 |
|
| 126 |
/// Store functions in a 8x8 grid. Most will be empty. |
| 127 |
const CDEF_DIST_KERNEL_FNS_LENGTH: usize = 8 * 8; |
| 128 |
|
| 129 |
const fn kernel_fn_index(w: usize, h: usize) -> usize { |
| 130 |
((w - 1) << 3) | (h - 1) |
| 131 |
} |
| 132 |
|
| 133 |
static CDEF_DIST_KERNEL_FNS_SSE2: [Option<CdefDistKernelFn>; |
| 134 |
CDEF_DIST_KERNEL_FNS_LENGTH] = { |
| 135 |
let mut out: [Option<CdefDistKernelFn>; CDEF_DIST_KERNEL_FNS_LENGTH] = |
| 136 |
[None; CDEF_DIST_KERNEL_FNS_LENGTH]; |
| 137 |
|
| 138 |
out[kernel_fn_index(4, 4)] = Some(rav1e_cdef_dist_kernel_4x4_sse2); |
| 139 |
out[kernel_fn_index(4, 8)] = Some(rav1e_cdef_dist_kernel_4x8_sse2); |
| 140 |
out[kernel_fn_index(8, 4)] = Some(rav1e_cdef_dist_kernel_8x4_sse2); |
| 141 |
out[kernel_fn_index(8, 8)] = Some(rav1e_cdef_dist_kernel_8x8_sse2); |
| 142 |
|
| 143 |
out |
| 144 |
}; |
| 145 |
|
| 146 |
cpu_function_lookup_table!( |
| 147 |
CDEF_DIST_KERNEL_FNS: |
| 148 |
[[Option<CdefDistKernelFn>; CDEF_DIST_KERNEL_FNS_LENGTH]], |
| 149 |
default: [None; CDEF_DIST_KERNEL_FNS_LENGTH], |
| 150 |
[SSE2] |
| 151 |
); |
| 152 |
|
| 153 |
#[target_feature(enable = "avx2")] |
| 154 |
#[inline] |
| 155 |
unsafe fn mm256_sum_i32(ymm: __m256i) -> i32 { |
| 156 |
// We split the vector in half and then add the two halves and sum. |
| 157 |
let m1 = _mm256_extracti128_si256(ymm, 1); |
| 158 |
let m2 = _mm256_castsi256_si128(ymm); |
| 159 |
let m2 = _mm_add_epi32(m2, m1); |
| 160 |
let m1 = _mm_shuffle_epi32(m2, 0b11_10_11_10); |
| 161 |
let m2 = _mm_add_epi32(m2, m1); |
| 162 |
let m1 = _mm_shuffle_epi32(m2, 0b01_01_01_01); |
| 163 |
let m2 = _mm_add_epi32(m2, m1); |
| 164 |
_mm_cvtsi128_si32(m2) |
| 165 |
} |
| 166 |
|
| 167 |
#[target_feature(enable = "avx2")] |
| 168 |
#[inline] |
| 169 |
unsafe fn rav1e_cdef_dist_kernel_8x8_hbd_avx2( |
| 170 |
src: *const u16, src_stride: isize, dst: *const u16, dst_stride: isize, |
| 171 |
) -> (u32, u32, u32) { |
| 172 |
let src = src as *const u8; |
| 173 |
let dst = dst as *const u8; |
| 174 |
|
| 175 |
unsafe fn sum16(src: *const u8, src_stride: isize) -> u32 { |
| 176 |
let h = 8; |
| 177 |
let res = (0..h) |
| 178 |
.map(|row| _mm_load_si128(src.offset(row * src_stride) as *const _)) |
| 179 |
.reduce(|a, b| _mm_add_epi16(a, b)) |
| 180 |
.unwrap(); |
| 181 |
|
| 182 |
let m32 = _mm256_cvtepi16_epi32(res); |
| 183 |
mm256_sum_i32(m32) as u32 |
| 184 |
} |
| 185 |
unsafe fn mpadd32( |
| 186 |
src: *const u8, src_stride: isize, dst: *const u8, dst_stride: isize, |
| 187 |
) -> u32 { |
| 188 |
let h = 8; |
| 189 |
let res = (0..h / 2) |
| 190 |
.map(|row| { |
| 191 |
let s = _mm256_loadu2_m128i( |
| 192 |
src.offset(2 * row * src_stride) as *const _, |
| 193 |
src.offset((2 * row + 1) * src_stride) as *const _, |
| 194 |
); |
| 195 |
|
| 196 |
let d = _mm256_loadu2_m128i( |
| 197 |
dst.offset(2 * row * dst_stride) as *const _, |
| 198 |
dst.offset((2 * row + 1) * dst_stride) as *const _, |
| 199 |
); |
| 200 |
|
| 201 |
_mm256_madd_epi16(s, d) |
| 202 |
}) |
| 203 |
.reduce(|a, b| _mm256_add_epi32(a, b)) |
| 204 |
.unwrap(); |
| 205 |
mm256_sum_i32(res) as u32 |
| 206 |
} |
| 207 |
|
| 208 |
let sum_s = sum16(src, src_stride); |
| 209 |
let sum_d = sum16(dst, dst_stride); |
| 210 |
let sum_s2 = mpadd32(src, src_stride, src, src_stride); |
| 211 |
let sum_d2 = mpadd32(dst, dst_stride, dst, dst_stride); |
| 212 |
let sum_sd = mpadd32(src, src_stride, dst, dst_stride); |
| 213 |
|
| 214 |
// To get the distortion, compute sum of squared error and apply a weight |
| 215 |
// based on the variance of the two planes. |
| 216 |
let sse = sum_d2 + sum_s2 - 2 * sum_sd; |
| 217 |
|
| 218 |
// Convert to 64-bits to avoid overflow when squaring |
| 219 |
let sum_s = sum_s as u64; |
| 220 |
let sum_d = sum_d as u64; |
| 221 |
|
| 222 |
let svar = (sum_s2 as u64 - (sum_s * sum_s + 32) / 64) as u32; |
| 223 |
let dvar = (sum_d2 as u64 - (sum_d * sum_d + 32) / 64) as u32; |
| 224 |
|
| 225 |
(svar, dvar, sse) |
| 226 |
} |
| 227 |
|
| 228 |
static CDEF_DIST_KERNEL_HBD_FNS_AVX2: [Option<CdefDistKernelHBDFn>; |
| 229 |
CDEF_DIST_KERNEL_FNS_LENGTH] = { |
| 230 |
let mut out: [Option<CdefDistKernelHBDFn>; CDEF_DIST_KERNEL_FNS_LENGTH] = |
| 231 |
[None; CDEF_DIST_KERNEL_FNS_LENGTH]; |
| 232 |
|
| 233 |
out[kernel_fn_index(8, 8)] = Some(rav1e_cdef_dist_kernel_8x8_hbd_avx2); |
| 234 |
|
| 235 |
out |
| 236 |
}; |
| 237 |
|
| 238 |
cpu_function_lookup_table!( |
| 239 |
CDEF_DIST_KERNEL_HBD_FNS: |
| 240 |
[[Option<CdefDistKernelHBDFn>; CDEF_DIST_KERNEL_FNS_LENGTH]], |
| 241 |
default: [None; CDEF_DIST_KERNEL_FNS_LENGTH], |
| 242 |
[AVX2] |
| 243 |
); |