moxcms/conversions/sse/
lut4_to_3.rs

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 */
29use crate::conversions::LutBarycentricReduction;
30use crate::conversions::interpolator::BarycentricWeight;
31use crate::conversions::lut_transforms::Lut4x3Factory;
32use crate::conversions::sse::interpolator::*;
33use crate::conversions::sse::interpolator_q0_15::SseAlignedI16x4;
34use crate::conversions::sse::lut4_to_3_q0_15::TransformLut4To3SseQ0_15;
35use crate::transform::PointeeSizeExpressible;
36use crate::{
37    BarycentricWeightScale, CmsError, DataColorSpace, InterpolationMethod, Layout,
38    TransformExecutor, TransformOptions,
39};
40use num_traits::AsPrimitive;
41#[cfg(target_arch = "x86")]
42use std::arch::x86::*;
43#[cfg(target_arch = "x86_64")]
44use std::arch::x86_64::*;
45use std::marker::PhantomData;
46
47struct TransformLut4To3Sse<
48    T,
49    U,
50    const LAYOUT: u8,
51    const GRID_SIZE: usize,
52    const BIT_DEPTH: usize,
53    const BINS: usize,
54    const BARYCENTRIC_BINS: usize,
55> {
56    lut: Vec<SseAlignedF32>,
57    _phantom: PhantomData<T>,
58    _phantom1: PhantomData<U>,
59    interpolation_method: InterpolationMethod,
60    weights: Box<[BarycentricWeight<f32>; BINS]>,
61    color_space: DataColorSpace,
62    is_linear: bool,
63}
64
65impl<
66    T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
67    U: AsPrimitive<usize>,
68    const LAYOUT: u8,
69    const GRID_SIZE: usize,
70    const BIT_DEPTH: usize,
71    const BINS: usize,
72    const BARYCENTRIC_BINS: usize,
73> TransformLut4To3Sse<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
74where
75    f32: AsPrimitive<T>,
76    u32: AsPrimitive<T>,
77    (): LutBarycentricReduction<T, U>,
78{
79    #[allow(unused_unsafe)]
80    #[target_feature(enable = "sse4.1")]
81    unsafe fn transform_chunk(
82        &self,
83        src: &[T],
84        dst: &mut [T],
85        interpolator: Box<dyn SseMdInterpolation + Send + Sync>,
86    ) {
87        let cn = Layout::from(LAYOUT);
88        let channels = cn.channels();
89        let grid_size = GRID_SIZE as i32;
90        let grid_size3 = grid_size * grid_size * grid_size;
91
92        let value_scale = unsafe { _mm_set1_ps(((1 << BIT_DEPTH) - 1) as f32) };
93        let max_value = ((1 << BIT_DEPTH) - 1u32).as_();
94
95        for (src, dst) in src.chunks_exact(4).zip(dst.chunks_exact_mut(channels)) {
96            let c = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
97                src[0],
98            );
99            let m = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
100                src[1],
101            );
102            let y = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
103                src[2],
104            );
105            let k = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
106                src[3],
107            );
108
109            let k_weights = self.weights[k.as_()];
110
111            let w: i32 = k_weights.x;
112            let w_n: i32 = k_weights.x_n;
113            let t: f32 = k_weights.w;
114
115            let table1 = &self.lut[(w * grid_size3) as usize..];
116            let table2 = &self.lut[(w_n * grid_size3) as usize..];
117
118            let a0 = interpolator
119                .inter3_sse(table1, c.as_(), m.as_(), y.as_(), self.weights.as_slice())
120                .v;
121            let b0 = interpolator
122                .inter3_sse(table2, c.as_(), m.as_(), y.as_(), self.weights.as_slice())
123                .v;
124
125            if T::FINITE {
126                unsafe {
127                    let t0 = _mm_set1_ps(t);
128                    let ones = _mm_set1_ps(1f32);
129                    let hp = _mm_mul_ps(a0, _mm_sub_ps(ones, t0));
130                    let mut v = _mm_add_ps(_mm_mul_ps(b0, t0), hp);
131                    v = _mm_max_ps(v, _mm_setzero_ps());
132                    v = _mm_mul_ps(v, value_scale);
133                    v = _mm_min_ps(v, value_scale);
134                    let jvz = _mm_cvtps_epi32(v);
135
136                    let x = _mm_extract_epi32::<0>(jvz);
137                    let y = _mm_extract_epi32::<1>(jvz);
138                    let z = _mm_extract_epi32::<2>(jvz);
139
140                    dst[cn.r_i()] = (x as u32).as_();
141                    dst[cn.g_i()] = (y as u32).as_();
142                    dst[cn.b_i()] = (z as u32).as_();
143                }
144            } else {
145                unsafe {
146                    let t0 = _mm_set1_ps(t);
147                    let ones = _mm_set1_ps(1f32);
148                    let hp = _mm_mul_ps(a0, _mm_sub_ps(ones, t0));
149                    let v = _mm_add_ps(_mm_mul_ps(b0, t0), hp);
150
151                    dst[cn.r_i()] = f32::from_bits(_mm_extract_ps::<0>(v) as u32).as_();
152                    dst[cn.g_i()] = f32::from_bits(_mm_extract_ps::<1>(v) as u32).as_();
153                    dst[cn.b_i()] = f32::from_bits(_mm_extract_ps::<2>(v) as u32).as_();
154                }
155            }
156            if channels == 4 {
157                dst[cn.a_i()] = max_value;
158            }
159        }
160    }
161}
162
163impl<
164    T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
165    U: AsPrimitive<usize>,
166    const LAYOUT: u8,
167    const GRID_SIZE: usize,
168    const BIT_DEPTH: usize,
169    const BINS: usize,
170    const BARYCENTRIC_BINS: usize,
171> TransformExecutor<T>
172    for TransformLut4To3Sse<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
173where
174    f32: AsPrimitive<T>,
175    u32: AsPrimitive<T>,
176    (): LutBarycentricReduction<T, U>,
177{
178    fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
179        let cn = Layout::from(LAYOUT);
180        let channels = cn.channels();
181        if src.len() % 4 != 0 {
182            return Err(CmsError::LaneMultipleOfChannels);
183        }
184        if dst.len() % channels != 0 {
185            return Err(CmsError::LaneMultipleOfChannels);
186        }
187        let src_chunks = src.len() / 4;
188        let dst_chunks = dst.len() / channels;
189        if src_chunks != dst_chunks {
190            return Err(CmsError::LaneSizeMismatch);
191        }
192
193        unsafe {
194            if self.color_space == DataColorSpace::Lab
195                || (self.is_linear && self.color_space == DataColorSpace::Rgb)
196                || self.color_space == DataColorSpace::Xyz
197            {
198                self.transform_chunk(src, dst, Box::new(TrilinearSse::<GRID_SIZE> {}));
199            } else {
200                match self.interpolation_method {
201                    #[cfg(feature = "options")]
202                    InterpolationMethod::Tetrahedral => {
203                        self.transform_chunk(src, dst, Box::new(TetrahedralSse::<GRID_SIZE> {}));
204                    }
205                    #[cfg(feature = "options")]
206                    InterpolationMethod::Pyramid => {
207                        self.transform_chunk(src, dst, Box::new(PyramidalSse::<GRID_SIZE> {}));
208                    }
209                    #[cfg(feature = "options")]
210                    InterpolationMethod::Prism => {
211                        self.transform_chunk(src, dst, Box::new(PrismaticSse::<GRID_SIZE> {}));
212                    }
213                    InterpolationMethod::Linear => {
214                        self.transform_chunk(src, dst, Box::new(TrilinearSse::<GRID_SIZE> {}));
215                    }
216                }
217            }
218        }
219
220        Ok(())
221    }
222}
223
224pub(crate) struct SseLut4x3Factory {}
225
226impl Lut4x3Factory for SseLut4x3Factory {
227    fn make_transform_4x3<
228        T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible + 'static + Send + Sync,
229        const LAYOUT: u8,
230        const GRID_SIZE: usize,
231        const BIT_DEPTH: usize,
232    >(
233        lut: Vec<f32>,
234        options: TransformOptions,
235        color_space: DataColorSpace,
236        is_linear: bool,
237    ) -> Box<dyn TransformExecutor<T> + Sync + Send>
238    where
239        f32: AsPrimitive<T>,
240        u32: AsPrimitive<T>,
241        (): LutBarycentricReduction<T, u8>,
242        (): LutBarycentricReduction<T, u16>,
243    {
244        if options.prefer_fixed_point && BIT_DEPTH < 16 {
245            let q: f32 = if T::FINITE {
246                ((1i32 << BIT_DEPTH as i32) - 1) as f32
247            } else {
248                ((1i32 << 14i32) - 1) as f32
249            };
250            let lut = lut
251                .chunks_exact(3)
252                .map(|x| {
253                    SseAlignedI16x4([
254                        (x[0] * q).round() as i16,
255                        (x[1] * q).round() as i16,
256                        (x[2] * q).round() as i16,
257                        0,
258                    ])
259                })
260                .collect::<Vec<_>>();
261            return match options.barycentric_weight_scale {
262                BarycentricWeightScale::Low => Box::new(TransformLut4To3SseQ0_15::<
263                    T,
264                    u8,
265                    LAYOUT,
266                    GRID_SIZE,
267                    BIT_DEPTH,
268                    256,
269                    256,
270                > {
271                    lut,
272                    interpolation_method: options.interpolation_method,
273                    weights: BarycentricWeight::<i16>::create_ranged_256::<GRID_SIZE>(),
274                    _phantom: PhantomData,
275                    _phantom1: PhantomData,
276                    color_space,
277                    is_linear,
278                }),
279                #[cfg(feature = "options")]
280                BarycentricWeightScale::High => Box::new(TransformLut4To3SseQ0_15::<
281                    T,
282                    u16,
283                    LAYOUT,
284                    GRID_SIZE,
285                    BIT_DEPTH,
286                    65536,
287                    65536,
288                > {
289                    lut,
290                    interpolation_method: options.interpolation_method,
291                    weights: BarycentricWeight::<i16>::create_binned::<GRID_SIZE, 65536>(),
292                    _phantom: PhantomData,
293                    _phantom1: PhantomData,
294                    color_space,
295                    is_linear,
296                }),
297            };
298        }
299        let lut = lut
300            .chunks_exact(3)
301            .map(|x| SseAlignedF32([x[0], x[1], x[2], 0f32]))
302            .collect::<Vec<_>>();
303        match options.barycentric_weight_scale {
304            BarycentricWeightScale::Low => {
305                Box::new(
306                    TransformLut4To3Sse::<T, u8, LAYOUT, GRID_SIZE, BIT_DEPTH, 256, 256> {
307                        lut,
308                        _phantom: PhantomData,
309                        _phantom1: PhantomData,
310                        interpolation_method: options.interpolation_method,
311                        weights: BarycentricWeight::<f32>::create_ranged_256::<GRID_SIZE>(),
312                        color_space,
313                        is_linear,
314                    },
315                )
316            }
317            #[cfg(feature = "options")]
318            BarycentricWeightScale::High => {
319                Box::new(
320                    TransformLut4To3Sse::<T, u16, LAYOUT, GRID_SIZE, BIT_DEPTH, 65536, 65536> {
321                        lut,
322                        _phantom: PhantomData,
323                        _phantom1: PhantomData,
324                        interpolation_method: options.interpolation_method,
325                        weights: BarycentricWeight::<f32>::create_binned::<GRID_SIZE, 65536>(),
326                        color_space,
327                        is_linear,
328                    },
329                )
330            }
331        }
332    }
333}
moxcms/conversions/sse/lut4_to_3.rs

moxcms/conversions/sse/
lut4_to_3.rs
