| 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::LutBarycentricReduction; |
| 30 |
use crate::conversions::interpolator::BarycentricWeight; |
| 31 |
use crate::conversions::lut_transforms::Lut4x3Factory; |
| 32 |
use crate::conversions::neon::interpolator::*; |
| 33 |
use crate::conversions::neon::interpolator_q0_15::NeonAlignedI16x4; |
| 34 |
use crate::conversions::neon::lut4_to_3_q0_15::TransformLut4To3NeonQ0_15; |
| 35 |
use crate::conversions::neon::rgb_xyz::NeonAlignedF32; |
| 36 |
use crate::transform::PointeeSizeExpressible; |
| 37 |
use crate::{ |
| 38 |
BarycentricWeightScale, CmsError, DataColorSpace, InterpolationMethod, Layout, |
| 39 |
TransformExecutor, TransformOptions, |
| 40 |
}; |
| 41 |
use num_traits::AsPrimitive; |
| 42 |
use std::arch::aarch64::*; |
| 43 |
use std::marker::PhantomData; |
| 44 |
|
| 45 |
struct TransformLut4To3Neon< |
| 46 |
T, |
| 47 |
U, |
| 48 |
const LAYOUT: u8, |
| 49 |
const GRID_SIZE: usize, |
| 50 |
const BIT_DEPTH: usize, |
| 51 |
const BINS: usize, |
| 52 |
const BARYCENTRIC_BINS: usize, |
| 53 |
> { |
| 54 |
lut: Vec<NeonAlignedF32>, |
| 55 |
_phantom: PhantomData<T>, |
| 56 |
_phantom1: PhantomData<U>, |
| 57 |
interpolation_method: InterpolationMethod, |
| 58 |
weights: Box<[BarycentricWeight<f32>; BINS]>, |
| 59 |
color_space: DataColorSpace, |
| 60 |
is_linear: bool, |
| 61 |
} |
| 62 |
|
| 63 |
impl< |
| 64 |
T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible, |
| 65 |
U: AsPrimitive<usize>, |
| 66 |
const LAYOUT: u8, |
| 67 |
const GRID_SIZE: usize, |
| 68 |
const BIT_DEPTH: usize, |
| 69 |
const BINS: usize, |
| 70 |
const BARYCENTRIC_BINS: usize, |
| 71 |
> TransformLut4To3Neon<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS> |
| 72 |
where |
| 73 |
f32: AsPrimitive<T>, |
| 74 |
u32: AsPrimitive<T>, |
| 75 |
(): LutBarycentricReduction<T, U>, |
| 76 |
{ |
| 77 |
#[allow(unused_unsafe)] |
| 78 |
#[inline(never)] |
| 79 |
fn transform_chunk( |
| 80 |
&self, |
| 81 |
src: &[T], |
| 82 |
dst: &mut [T], |
| 83 |
interpolator: Box<dyn NeonMdInterpolationDouble + Send + Sync>, |
| 84 |
) { |
| 85 |
let cn = Layout::from(LAYOUT); |
| 86 |
let channels = cn.channels(); |
| 87 |
let grid_size = GRID_SIZE as i32; |
| 88 |
let grid_size3 = grid_size * grid_size * grid_size; |
| 89 |
|
| 90 |
let value_scale = unsafe { vdupq_n_f32(((1 << BIT_DEPTH) - 1) as f32) }; |
| 91 |
let max_value = ((1 << BIT_DEPTH) - 1u32).as_(); |
| 92 |
|
| 93 |
for (src, dst) in src.chunks_exact(4).zip(dst.chunks_exact_mut(channels)) { |
| 94 |
let c = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>( |
| 95 |
src[0], |
| 96 |
); |
| 97 |
let m = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>( |
| 98 |
src[1], |
| 99 |
); |
| 100 |
let y = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>( |
| 101 |
src[2], |
| 102 |
); |
| 103 |
let k = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>( |
| 104 |
src[3], |
| 105 |
); |
| 106 |
|
| 107 |
let k_weights = self.weights[k.as_()]; |
| 108 |
|
| 109 |
let w: i32 = k_weights.x; |
| 110 |
let w_n: i32 = k_weights.x_n; |
| 111 |
let t: f32 = k_weights.w; |
| 112 |
|
| 113 |
let table1 = &self.lut[(w * grid_size3) as usize..]; |
| 114 |
let table2 = &self.lut[(w_n * grid_size3) as usize..]; |
| 115 |
|
| 116 |
let (a0, b0) = interpolator.inter3_neon( |
| 117 |
table1, |
| 118 |
table2, |
| 119 |
c.as_(), |
| 120 |
m.as_(), |
| 121 |
y.as_(), |
| 122 |
self.weights.as_slice(), |
| 123 |
); |
| 124 |
let (a0, b0) = (a0.v, b0.v); |
| 125 |
|
| 126 |
if T::FINITE { |
| 127 |
unsafe { |
| 128 |
let t0 = vdupq_n_f32(t); |
| 129 |
let hp = vfmsq_f32(a0, a0, t0); |
| 130 |
let mut v = vfmaq_f32(hp, b0, t0); |
| 131 |
v = vmulq_f32(v, value_scale); |
| 132 |
v = vminq_f32(v, value_scale); |
| 133 |
|
| 134 |
let jvx = vcvtaq_u32_f32(v); |
| 135 |
|
| 136 |
dst[cn.r_i()] = vgetq_lane_u32::<0>(jvx).as_(); |
| 137 |
dst[cn.g_i()] = vgetq_lane_u32::<1>(jvx).as_(); |
| 138 |
dst[cn.b_i()] = vgetq_lane_u32::<2>(jvx).as_(); |
| 139 |
} |
| 140 |
} else { |
| 141 |
unsafe { |
| 142 |
let t0 = vdupq_n_f32(t); |
| 143 |
let hp = vfmsq_f32(a0, a0, t0); |
| 144 |
let v = vfmaq_f32(hp, b0, t0); |
| 145 |
|
| 146 |
dst[cn.r_i()] = vgetq_lane_f32::<0>(v).as_(); |
| 147 |
dst[cn.g_i()] = vgetq_lane_f32::<1>(v).as_(); |
| 148 |
dst[cn.b_i()] = vgetq_lane_f32::<2>(v).as_(); |
| 149 |
} |
| 150 |
} |
| 151 |
if channels == 4 { |
| 152 |
dst[cn.a_i()] = max_value; |
| 153 |
} |
| 154 |
} |
| 155 |
} |
| 156 |
} |
| 157 |
|
| 158 |
impl< |
| 159 |
T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible, |
| 160 |
U: AsPrimitive<usize>, |
| 161 |
const LAYOUT: u8, |
| 162 |
const GRID_SIZE: usize, |
| 163 |
const BIT_DEPTH: usize, |
| 164 |
const BINS: usize, |
| 165 |
const BARYCENTRIC_BINS: usize, |
| 166 |
> TransformExecutor<T> |
| 167 |
for TransformLut4To3Neon<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS> |
| 168 |
where |
| 169 |
f32: AsPrimitive<T>, |
| 170 |
u32: AsPrimitive<T>, |
| 171 |
(): LutBarycentricReduction<T, U>, |
| 172 |
{ |
| 173 |
fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> { |
| 174 |
let cn = Layout::from(LAYOUT); |
| 175 |
let channels = cn.channels(); |
| 176 |
if src.len() % 4 != 0 { |
| 177 |
return Err(CmsError::LaneMultipleOfChannels); |
| 178 |
} |
| 179 |
if dst.len() % channels != 0 { |
| 180 |
return Err(CmsError::LaneMultipleOfChannels); |
| 181 |
} |
| 182 |
let src_chunks = src.len() / 4; |
| 183 |
let dst_chunks = dst.len() / channels; |
| 184 |
if src_chunks != dst_chunks { |
| 185 |
return Err(CmsError::LaneSizeMismatch); |
| 186 |
} |
| 187 |
|
| 188 |
if self.color_space == DataColorSpace::Lab |
| 189 |
|| (self.is_linear && self.color_space == DataColorSpace::Rgb) |
| 190 |
|| self.color_space == DataColorSpace::Xyz |
| 191 |
{ |
| 192 |
self.transform_chunk(src, dst, Box::new(TrilinearNeonDouble::<GRID_SIZE> {})); |
| 193 |
} else { |
| 194 |
match self.interpolation_method { |
| 195 |
#[cfg(feature = "options")] |
| 196 |
InterpolationMethod::Tetrahedral => { |
| 197 |
self.transform_chunk(src, dst, Box::new(TetrahedralNeonDouble::<GRID_SIZE> {})); |
| 198 |
} |
| 199 |
#[cfg(feature = "options")] |
| 200 |
InterpolationMethod::Pyramid => { |
| 201 |
self.transform_chunk(src, dst, Box::new(PyramidalNeonDouble::<GRID_SIZE> {})); |
| 202 |
} |
| 203 |
#[cfg(feature = "options")] |
| 204 |
InterpolationMethod::Prism => { |
| 205 |
self.transform_chunk(src, dst, Box::new(PrismaticNeonDouble::<GRID_SIZE> {})); |
| 206 |
} |
| 207 |
InterpolationMethod::Linear => { |
| 208 |
self.transform_chunk(src, dst, Box::new(TrilinearNeonDouble::<GRID_SIZE> {})); |
| 209 |
} |
| 210 |
} |
| 211 |
} |
| 212 |
|
| 213 |
Ok(()) |
| 214 |
} |
| 215 |
} |
| 216 |
|
| 217 |
pub(crate) struct NeonLut4x3Factory {} |
| 218 |
|
| 219 |
impl Lut4x3Factory for NeonLut4x3Factory { |
| 220 |
fn make_transform_4x3< |
| 221 |
T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible + 'static + Send + Sync, |
| 222 |
const LAYOUT: u8, |
| 223 |
const GRID_SIZE: usize, |
| 224 |
const BIT_DEPTH: usize, |
| 225 |
>( |
| 226 |
lut: Vec<f32>, |
| 227 |
options: TransformOptions, |
| 228 |
color_space: DataColorSpace, |
| 229 |
is_linear: bool, |
| 230 |
) -> Box<dyn TransformExecutor<T> + Sync + Send> |
| 231 |
where |
| 232 |
f32: AsPrimitive<T>, |
| 233 |
u32: AsPrimitive<T>, |
| 234 |
(): LutBarycentricReduction<T, u8>, |
| 235 |
(): LutBarycentricReduction<T, u16>, |
| 236 |
{ |
| 237 |
if options.prefer_fixed_point |
| 238 |
&& BIT_DEPTH < 16 |
| 239 |
&& std::arch::is_aarch64_feature_detected!("rdm") |
| 240 |
{ |
| 241 |
let q: f32 = if T::FINITE { |
| 242 |
((1i32 << BIT_DEPTH as i32) - 1) as f32 |
| 243 |
} else { |
| 244 |
((1i32 << 14i32) - 1) as f32 |
| 245 |
}; |
| 246 |
let lut = lut |
| 247 |
.chunks_exact(3) |
| 248 |
.map(|x| { |
| 249 |
NeonAlignedI16x4([ |
| 250 |
(x[0] * q).round() as i16, |
| 251 |
(x[1] * q).round() as i16, |
| 252 |
(x[2] * q).round() as i16, |
| 253 |
0, |
| 254 |
]) |
| 255 |
}) |
| 256 |
.collect::<Vec<_>>(); |
| 257 |
return match options.barycentric_weight_scale { |
| 258 |
BarycentricWeightScale::Low => Box::new(TransformLut4To3NeonQ0_15::< |
| 259 |
T, |
| 260 |
u8, |
| 261 |
LAYOUT, |
| 262 |
GRID_SIZE, |
| 263 |
BIT_DEPTH, |
| 264 |
256, |
| 265 |
256, |
| 266 |
> { |
| 267 |
lut, |
| 268 |
_phantom: PhantomData, |
| 269 |
_phantom1: PhantomData, |
| 270 |
interpolation_method: options.interpolation_method, |
| 271 |
weights: BarycentricWeight::<i16>::create_ranged_256::<GRID_SIZE>(), |
| 272 |
color_space, |
| 273 |
is_linear, |
| 274 |
}), |
| 275 |
#[cfg(feature = "options")] |
| 276 |
BarycentricWeightScale::High => Box::new(TransformLut4To3NeonQ0_15::< |
| 277 |
T, |
| 278 |
u16, |
| 279 |
LAYOUT, |
| 280 |
GRID_SIZE, |
| 281 |
BIT_DEPTH, |
| 282 |
65536, |
| 283 |
65536, |
| 284 |
> { |
| 285 |
lut, |
| 286 |
_phantom: PhantomData, |
| 287 |
_phantom1: PhantomData, |
| 288 |
interpolation_method: options.interpolation_method, |
| 289 |
weights: BarycentricWeight::<i16>::create_binned::<GRID_SIZE, 65536>(), |
| 290 |
color_space, |
| 291 |
is_linear, |
| 292 |
}), |
| 293 |
}; |
| 294 |
} |
| 295 |
let lut = lut |
| 296 |
.chunks_exact(3) |
| 297 |
.map(|x| NeonAlignedF32([x[0], x[1], x[2], 0f32])) |
| 298 |
.collect::<Vec<_>>(); |
| 299 |
match options.barycentric_weight_scale { |
| 300 |
BarycentricWeightScale::Low => { |
| 301 |
Box::new( |
| 302 |
TransformLut4To3Neon::<T, u8, LAYOUT, GRID_SIZE, BIT_DEPTH, 256, 256> { |
| 303 |
lut, |
| 304 |
_phantom: PhantomData, |
| 305 |
_phantom1: PhantomData, |
| 306 |
interpolation_method: options.interpolation_method, |
| 307 |
weights: BarycentricWeight::<f32>::create_ranged_256::<GRID_SIZE>(), |
| 308 |
color_space, |
| 309 |
is_linear, |
| 310 |
}, |
| 311 |
) |
| 312 |
} |
| 313 |
#[cfg(feature = "options")] |
| 314 |
BarycentricWeightScale::High => { |
| 315 |
Box::new( |
| 316 |
TransformLut4To3Neon::<T, u16, LAYOUT, GRID_SIZE, BIT_DEPTH, 65536, 65536> { |
| 317 |
lut, |
| 318 |
_phantom: PhantomData, |
| 319 |
_phantom1: PhantomData, |
| 320 |
interpolation_method: options.interpolation_method, |
| 321 |
weights: BarycentricWeight::<f32>::create_binned::<GRID_SIZE, 65536>(), |
| 322 |
color_space, |
| 323 |
is_linear, |
| 324 |
}, |
| 325 |
) |
| 326 |
} |
| 327 |
} |
| 328 |
} |
| 329 |
} |