| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 3/2025. All rights reserved. |
| 3 |
* // |
| 4 |
* // Redistribution and use in source and binary forms, with or without modification, |
| 5 |
* // are permitted provided that the following conditions are met: |
| 6 |
* // |
| 7 |
* // 1. Redistributions of source code must retain the above copyright notice, this |
| 8 |
* // list of conditions and the following disclaimer. |
| 9 |
* // |
| 10 |
* // 2. Redistributions in binary form must reproduce the above copyright notice, |
| 11 |
* // this list of conditions and the following disclaimer in the documentation |
| 12 |
* // and/or other materials provided with the distribution. |
| 13 |
* // |
| 14 |
* // 3. Neither the name of the copyright holder nor the names of its |
| 15 |
* // contributors may be used to endorse or promote products derived from |
| 16 |
* // this software without specific prior written permission. |
| 17 |
* // |
| 18 |
* // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 19 |
* // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 20 |
* // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE |
| 21 |
* // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE |
| 22 |
* // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 23 |
* // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR |
| 24 |
* // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER |
| 25 |
* // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, |
| 26 |
* // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 27 |
* // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 28 |
*/ |
| 29 |
use crate::conversions::avx::rgb_xyz::AvxAlignedU16; |
| 30 |
use crate::conversions::rgbxyz_fixed::TransformMatrixShaperFp; |
| 31 |
use crate::transform::PointeeSizeExpressible; |
| 32 |
use crate::{CmsError, Layout, TransformExecutor}; |
| 33 |
use num_traits::AsPrimitive; |
| 34 |
use std::arch::x86_64::*; |
| 35 |
|
| 36 |
pub(crate) struct TransformShaperRgbQ2_13Avx< |
| 37 |
T: Copy, |
| 38 |
const SRC_LAYOUT: u8, |
| 39 |
const DST_LAYOUT: u8, |
| 40 |
const PRECISION: i32, |
| 41 |
> { |
| 42 |
pub(crate) profile: TransformMatrixShaperFp<i32, T>, |
| 43 |
pub(crate) bit_depth: usize, |
| 44 |
pub(crate) gamma_lut: usize, |
| 45 |
} |
| 46 |
|
| 47 |
#[inline(always)] |
| 48 |
pub(crate) unsafe fn _xmm_broadcast_epi32(f: &i32) -> __m128i { |
| 49 |
let float_ref: &f32 = unsafe { &*(f as *const i32 as *const f32) }; |
| 50 |
unsafe { _mm_castps_si128(_mm_broadcast_ss(float_ref)) } |
| 51 |
} |
| 52 |
|
| 53 |
impl< |
| 54 |
T: Copy + PointeeSizeExpressible + 'static, |
| 55 |
const SRC_LAYOUT: u8, |
| 56 |
const DST_LAYOUT: u8, |
| 57 |
const PRECISION: i32, |
| 58 |
> TransformShaperRgbQ2_13Avx<T, SRC_LAYOUT, DST_LAYOUT, PRECISION> |
| 59 |
where |
| 60 |
u32: AsPrimitive<T>, |
| 61 |
{ |
| 62 |
#[target_feature(enable = "avx2")] |
| 63 |
unsafe fn transform_avx2(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> { |
| 64 |
let src_cn = Layout::from(SRC_LAYOUT); |
| 65 |
let dst_cn = Layout::from(DST_LAYOUT); |
| 66 |
let src_channels = src_cn.channels(); |
| 67 |
let dst_channels = dst_cn.channels(); |
| 68 |
|
| 69 |
let mut temporary0 = AvxAlignedU16([0; 16]); |
| 70 |
|
| 71 |
if src.len() / src_channels != dst.len() / dst_channels { |
| 72 |
return Err(CmsError::LaneSizeMismatch); |
| 73 |
} |
| 74 |
if src.len() % src_channels != 0 { |
| 75 |
return Err(CmsError::LaneMultipleOfChannels); |
| 76 |
} |
| 77 |
if dst.len() % dst_channels != 0 { |
| 78 |
return Err(CmsError::LaneMultipleOfChannels); |
| 79 |
} |
| 80 |
|
| 81 |
let t = self.profile.adaptation_matrix.transpose(); |
| 82 |
|
| 83 |
let max_colors = ((1 << self.bit_depth) - 1).as_(); |
| 84 |
|
| 85 |
// safety precondition for linearization table |
| 86 |
if T::FINITE { |
| 87 |
let cap = (1 << self.bit_depth) - 1; |
| 88 |
assert!(self.profile.r_linear.len() >= cap); |
| 89 |
assert!(self.profile.g_linear.len() >= cap); |
| 90 |
assert!(self.profile.b_linear.len() >= cap); |
| 91 |
} else { |
| 92 |
assert!(self.profile.r_linear.len() >= T::NOT_FINITE_LINEAR_TABLE_SIZE); |
| 93 |
assert!(self.profile.g_linear.len() >= T::NOT_FINITE_LINEAR_TABLE_SIZE); |
| 94 |
assert!(self.profile.b_linear.len() >= T::NOT_FINITE_LINEAR_TABLE_SIZE); |
| 95 |
} |
| 96 |
|
| 97 |
let r_lin = &self.profile.r_linear; |
| 98 |
let g_lin = &self.profile.g_linear; |
| 99 |
let b_lin = &self.profile.b_linear; |
| 100 |
|
| 101 |
unsafe { |
| 102 |
let m0 = _mm256_setr_epi16( |
| 103 |
t.v[0][0], t.v[1][0], t.v[0][1], t.v[1][1], t.v[0][2], t.v[1][2], 0, 0, t.v[0][0], |
| 104 |
t.v[1][0], t.v[0][1], t.v[1][1], t.v[0][2], t.v[1][2], 0, 0, |
| 105 |
); |
| 106 |
let m2 = _mm256_setr_epi16( |
| 107 |
t.v[2][0], 1, t.v[2][1], 1, t.v[2][2], 1, 0, 0, t.v[2][0], 1, t.v[2][1], 1, |
| 108 |
t.v[2][2], 1, 0, 0, |
| 109 |
); |
| 110 |
|
| 111 |
let rnd_val = ((1i32 << (PRECISION - 1)) as i16).to_ne_bytes(); |
| 112 |
let rnd = _mm256_set1_epi32(i32::from_ne_bytes([0, 0, rnd_val[0], rnd_val[1]])); |
| 113 |
|
| 114 |
let zeros = _mm256_setzero_si256(); |
| 115 |
|
| 116 |
let v_max_value = _mm256_set1_epi32(self.gamma_lut as i32 - 1); |
| 117 |
|
| 118 |
let mut src = src; |
| 119 |
let mut dst = dst; |
| 120 |
|
| 121 |
let mut src_iter = src.chunks_exact(src_channels * 2); |
| 122 |
let dst_iter = dst.chunks_exact_mut(dst_channels * 2); |
| 123 |
|
| 124 |
let (mut r0, mut g0, mut b0, mut a0); |
| 125 |
let (mut r1, mut g1, mut b1, mut a1); |
| 126 |
|
| 127 |
if let Some(src) = src_iter.next() { |
| 128 |
r0 = _xmm_broadcast_epi32(r_lin.get_unchecked(src[src_cn.r_i()]._as_usize())); |
| 129 |
g0 = _xmm_broadcast_epi32(g_lin.get_unchecked(src[src_cn.g_i()]._as_usize())); |
| 130 |
b0 = _xmm_broadcast_epi32(b_lin.get_unchecked(src[src_cn.b_i()]._as_usize())); |
| 131 |
r1 = _xmm_broadcast_epi32( |
| 132 |
r_lin.get_unchecked(src[src_cn.r_i() + src_channels]._as_usize()), |
| 133 |
); |
| 134 |
g1 = _xmm_broadcast_epi32( |
| 135 |
g_lin.get_unchecked(src[src_cn.g_i() + src_channels]._as_usize()), |
| 136 |
); |
| 137 |
b1 = _xmm_broadcast_epi32( |
| 138 |
b_lin.get_unchecked(src[src_cn.b_i() + src_channels]._as_usize()), |
| 139 |
); |
| 140 |
a0 = if src_channels == 4 { |
| 141 |
src[src_cn.a_i()] |
| 142 |
} else { |
| 143 |
max_colors |
| 144 |
}; |
| 145 |
a1 = if src_channels == 4 { |
| 146 |
src[src_cn.a_i() + src_channels] |
| 147 |
} else { |
| 148 |
max_colors |
| 149 |
}; |
| 150 |
} else { |
| 151 |
r0 = _mm_setzero_si128(); |
| 152 |
g0 = _mm_setzero_si128(); |
| 153 |
b0 = _mm_setzero_si128(); |
| 154 |
a0 = max_colors; |
| 155 |
r1 = _mm_setzero_si128(); |
| 156 |
g1 = _mm_setzero_si128(); |
| 157 |
b1 = _mm_setzero_si128(); |
| 158 |
a1 = max_colors; |
| 159 |
} |
| 160 |
|
| 161 |
for (src, dst) in src_iter.zip(dst_iter) { |
| 162 |
let zr0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(r0), r1); |
| 163 |
let mut zg0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(g0), g1); |
| 164 |
let zb0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(b0), b1); |
| 165 |
zg0 = _mm256_slli_epi32::<16>(zg0); |
| 166 |
|
| 167 |
let zrg0 = _mm256_or_si256(zr0, zg0); |
| 168 |
let zbz0 = _mm256_or_si256(zb0, rnd); |
| 169 |
|
| 170 |
let va0 = _mm256_madd_epi16(zrg0, m0); |
| 171 |
let va1 = _mm256_madd_epi16(zbz0, m2); |
| 172 |
|
| 173 |
let mut v0 = _mm256_add_epi32(va0, va1); |
| 174 |
|
| 175 |
v0 = _mm256_srai_epi32::<PRECISION>(v0); |
| 176 |
v0 = _mm256_max_epi32(v0, zeros); |
| 177 |
v0 = _mm256_min_epi32(v0, v_max_value); |
| 178 |
|
| 179 |
_mm256_store_si256(temporary0.0.as_mut_ptr() as *mut _, v0); |
| 180 |
|
| 181 |
r0 = _xmm_broadcast_epi32(r_lin.get_unchecked(src[src_cn.r_i()]._as_usize())); |
| 182 |
g0 = _xmm_broadcast_epi32(g_lin.get_unchecked(src[src_cn.g_i()]._as_usize())); |
| 183 |
b0 = _xmm_broadcast_epi32(b_lin.get_unchecked(src[src_cn.b_i()]._as_usize())); |
| 184 |
r1 = _xmm_broadcast_epi32( |
| 185 |
r_lin.get_unchecked(src[src_cn.r_i() + src_channels]._as_usize()), |
| 186 |
); |
| 187 |
g1 = _xmm_broadcast_epi32( |
| 188 |
g_lin.get_unchecked(src[src_cn.g_i() + src_channels]._as_usize()), |
| 189 |
); |
| 190 |
b1 = _xmm_broadcast_epi32( |
| 191 |
b_lin.get_unchecked(src[src_cn.b_i() + src_channels]._as_usize()), |
| 192 |
); |
| 193 |
|
| 194 |
dst[dst_cn.r_i()] = self.profile.r_gamma[temporary0.0[0] as usize]; |
| 195 |
dst[dst_cn.g_i()] = self.profile.g_gamma[temporary0.0[2] as usize]; |
| 196 |
dst[dst_cn.b_i()] = self.profile.b_gamma[temporary0.0[4] as usize]; |
| 197 |
if dst_channels == 4 { |
| 198 |
dst[dst_cn.a_i()] = a0; |
| 199 |
} |
| 200 |
|
| 201 |
dst[dst_cn.r_i() + dst_channels] = self.profile.r_gamma[temporary0.0[8] as usize]; |
| 202 |
dst[dst_cn.g_i() + dst_channels] = self.profile.g_gamma[temporary0.0[10] as usize]; |
| 203 |
dst[dst_cn.b_i() + dst_channels] = self.profile.b_gamma[temporary0.0[12] as usize]; |
| 204 |
if dst_channels == 4 { |
| 205 |
dst[dst_cn.a_i() + dst_channels] = a1; |
| 206 |
} |
| 207 |
|
| 208 |
a0 = if src_channels == 4 { |
| 209 |
src[src_cn.a_i()] |
| 210 |
} else { |
| 211 |
max_colors |
| 212 |
}; |
| 213 |
a1 = if src_channels == 4 { |
| 214 |
src[src_cn.a_i() + src_channels] |
| 215 |
} else { |
| 216 |
max_colors |
| 217 |
}; |
| 218 |
} |
| 219 |
|
| 220 |
if let Some(dst) = dst.chunks_exact_mut(dst_channels * 2).last() { |
| 221 |
let zr0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(r0), r1); |
| 222 |
let mut zg0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(g0), g1); |
| 223 |
let zb0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(b0), b1); |
| 224 |
zg0 = _mm256_slli_epi32::<16>(zg0); |
| 225 |
|
| 226 |
let zrg0 = _mm256_or_si256(zr0, zg0); |
| 227 |
let zbz0 = _mm256_or_si256(zb0, rnd); |
| 228 |
|
| 229 |
let va0 = _mm256_madd_epi16(zrg0, m0); |
| 230 |
let va1 = _mm256_madd_epi16(zbz0, m2); |
| 231 |
|
| 232 |
let mut v0 = _mm256_add_epi32(va0, va1); |
| 233 |
|
| 234 |
v0 = _mm256_srai_epi32::<PRECISION>(v0); |
| 235 |
v0 = _mm256_max_epi32(v0, zeros); |
| 236 |
v0 = _mm256_min_epi32(v0, v_max_value); |
| 237 |
|
| 238 |
_mm256_store_si256(temporary0.0.as_mut_ptr() as *mut _, v0); |
| 239 |
|
| 240 |
dst[dst_cn.r_i()] = self.profile.r_gamma[temporary0.0[0] as usize]; |
| 241 |
dst[dst_cn.g_i()] = self.profile.g_gamma[temporary0.0[2] as usize]; |
| 242 |
dst[dst_cn.b_i()] = self.profile.b_gamma[temporary0.0[4] as usize]; |
| 243 |
if dst_channels == 4 { |
| 244 |
dst[dst_cn.a_i()] = a0; |
| 245 |
} |
| 246 |
|
| 247 |
dst[dst_cn.r_i() + dst_channels] = self.profile.r_gamma[temporary0.0[8] as usize]; |
| 248 |
dst[dst_cn.g_i() + dst_channels] = self.profile.g_gamma[temporary0.0[10] as usize]; |
| 249 |
dst[dst_cn.b_i() + dst_channels] = self.profile.b_gamma[temporary0.0[12] as usize]; |
| 250 |
if dst_channels == 4 { |
| 251 |
dst[dst_cn.a_i() + dst_channels] = a1; |
| 252 |
} |
| 253 |
} |
| 254 |
|
| 255 |
src = src.chunks_exact(src_channels * 2).remainder(); |
| 256 |
dst = dst.chunks_exact_mut(dst_channels * 2).into_remainder(); |
| 257 |
|
| 258 |
for (src, dst) in src |
| 259 |
.chunks_exact(src_channels) |
| 260 |
.zip(dst.chunks_exact_mut(dst_channels)) |
| 261 |
{ |
| 262 |
let r = _xmm_broadcast_epi32(r_lin.get_unchecked(src[src_cn.r_i()]._as_usize())); |
| 263 |
let mut g = |
| 264 |
_xmm_broadcast_epi32(g_lin.get_unchecked(src[src_cn.g_i()]._as_usize())); |
| 265 |
let b = _xmm_broadcast_epi32(b_lin.get_unchecked(src[src_cn.b_i()]._as_usize())); |
| 266 |
|
| 267 |
g = _mm_slli_epi32::<16>(g); |
| 268 |
|
| 269 |
let a = if src_channels == 4 { |
| 270 |
src[src_cn.a_i()] |
| 271 |
} else { |
| 272 |
max_colors |
| 273 |
}; |
| 274 |
|
| 275 |
let zrg0 = _mm_or_si128(r, g); |
| 276 |
let zbz0 = _mm_or_si128(b, _mm256_castsi256_si128(rnd)); |
| 277 |
|
| 278 |
let v0 = _mm_madd_epi16(zrg0, _mm256_castsi256_si128(m0)); |
| 279 |
let v1 = _mm_madd_epi16(zbz0, _mm256_castsi256_si128(m2)); |
| 280 |
|
| 281 |
let mut v = _mm_add_epi32(v0, v1); |
| 282 |
|
| 283 |
v = _mm_srai_epi32::<PRECISION>(v); |
| 284 |
v = _mm_max_epi32(v, _mm_setzero_si128()); |
| 285 |
v = _mm_min_epi32(v, _mm256_castsi256_si128(v_max_value)); |
| 286 |
|
| 287 |
_mm_store_si128(temporary0.0.as_mut_ptr() as *mut _, v); |
| 288 |
|
| 289 |
dst[dst_cn.r_i()] = self.profile.r_gamma[temporary0.0[0] as usize]; |
| 290 |
dst[dst_cn.g_i()] = self.profile.g_gamma[temporary0.0[2] as usize]; |
| 291 |
dst[dst_cn.b_i()] = self.profile.b_gamma[temporary0.0[4] as usize]; |
| 292 |
if dst_channels == 4 { |
| 293 |
dst[dst_cn.a_i()] = a; |
| 294 |
} |
| 295 |
} |
| 296 |
} |
| 297 |
|
| 298 |
Ok(()) |
| 299 |
} |
| 300 |
} |
| 301 |
|
| 302 |
impl< |
| 303 |
T: Copy + PointeeSizeExpressible + 'static + Default, |
| 304 |
const SRC_LAYOUT: u8, |
| 305 |
const DST_LAYOUT: u8, |
| 306 |
const PRECISION: i32, |
| 307 |
> TransformExecutor<T> for TransformShaperRgbQ2_13Avx<T, SRC_LAYOUT, DST_LAYOUT, PRECISION> |
| 308 |
where |
| 309 |
u32: AsPrimitive<T>, |
| 310 |
{ |
| 311 |
fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> { |
| 312 |
unsafe { self.transform_avx2(src, dst) } |
| 313 |
} |
| 314 |
} |