moxcms/conversions/avx/
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::avx::interpolator::*;
31use crate::conversions::avx::interpolator_q0_15::AvxAlignedI16;
32use crate::conversions::avx::lut4_to_3_q0_15::TransformLut4To3AvxQ0_15;
33use crate::conversions::interpolator::BarycentricWeight;
34use crate::conversions::lut_transforms::Lut4x3Factory;
35use crate::transform::PointeeSizeExpressible;
36use crate::{
37    BarycentricWeightScale, CmsError, DataColorSpace, InterpolationMethod, Layout,
38    TransformExecutor, TransformOptions,
39};
40use num_traits::AsPrimitive;
41use std::arch::x86_64::*;
42use std::marker::PhantomData;
43
44struct TransformLut4To3Avx<
45    T,
46    U,
47    const LAYOUT: u8,
48    const GRID_SIZE: usize,
49    const BIT_DEPTH: usize,
50    const BINS: usize,
51    const BARYCENTRIC_BINS: usize,
52> {
53    lut: Vec<SseAlignedF32>,
54    _phantom: PhantomData<T>,
55    _phantom1: PhantomData<U>,
56    interpolation_method: InterpolationMethod,
57    weights: Box<[BarycentricWeight<f32>; BINS]>,
58    color_space: DataColorSpace,
59    is_linear: bool,
60}
61
62impl<
63    T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
64    U: AsPrimitive<usize>,
65    const LAYOUT: u8,
66    const GRID_SIZE: usize,
67    const BIT_DEPTH: usize,
68    const BINS: usize,
69    const BARYCENTRIC_BINS: usize,
70> TransformLut4To3Avx<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
71where
72    f32: AsPrimitive<T>,
73    u32: AsPrimitive<T>,
74    (): LutBarycentricReduction<T, U>,
75{
76    #[allow(unused_unsafe)]
77    #[target_feature(enable = "avx2", enable = "fma")]
78    unsafe fn transform_chunk(
79        &self,
80        src: &[T],
81        dst: &mut [T],
82        interpolator: Box<dyn AvxMdInterpolationDouble + Send + Sync>,
83    ) {
84        let cn = Layout::from(LAYOUT);
85        let channels = cn.channels();
86        let grid_size = GRID_SIZE as i32;
87        let grid_size3 = grid_size * grid_size * grid_size;
88
89        let value_scale = unsafe { _mm_set1_ps(((1 << BIT_DEPTH) - 1) as f32) };
90        let max_value = ((1 << BIT_DEPTH) - 1u32).as_();
91
92        for (src, dst) in src.chunks_exact(4).zip(dst.chunks_exact_mut(channels)) {
93            let c = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
94                src[0],
95            );
96            let m = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
97                src[1],
98            );
99            let y = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
100                src[2],
101            );
102            let k = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
103                src[3],
104            );
105
106            let k_weights = self.weights[k.as_()];
107
108            let w: i32 = k_weights.x;
109            let w_n: i32 = k_weights.x_n;
110            let t: f32 = k_weights.w;
111
112            let table1 = &self.lut[(w * grid_size3) as usize..];
113            let table2 = &self.lut[(w_n * grid_size3) as usize..];
114
115            let v = interpolator.inter3_sse(
116                table1,
117                table2,
118                c.as_(),
119                m.as_(),
120                y.as_(),
121                self.weights.as_slice(),
122            );
123            let (a0, b0) = (v.0.v, v.1.v);
124
125            if T::FINITE {
126                unsafe {
127                    let t0 = _mm_set1_ps(t);
128                    let hp = _mm_fnmadd_ps(a0, t0, a0);
129                    let mut v = _mm_fmadd_ps(b0, t0, hp);
130                    v = _mm_max_ps(v, _mm_setzero_ps());
131                    v = _mm_mul_ps(v, value_scale);
132                    v = _mm_min_ps(v, value_scale);
133                    let jvz = _mm_cvtps_epi32(v);
134
135                    let x = _mm_extract_epi32::<0>(jvz);
136                    let y = _mm_extract_epi32::<1>(jvz);
137                    let z = _mm_extract_epi32::<2>(jvz);
138
139                    dst[cn.r_i()] = (x as u32).as_();
140                    dst[cn.g_i()] = (y as u32).as_();
141                    dst[cn.b_i()] = (z as u32).as_();
142                }
143            } else {
144                unsafe {
145                    let t0 = _mm_set1_ps(t);
146                    let hp = _mm_fnmadd_ps(a0, t0, a0);
147                    let v = _mm_fmadd_ps(b0, t0, hp);
148                    dst[cn.r_i()] = f32::from_bits(_mm_extract_ps::<0>(v) as u32).as_();
149                    dst[cn.g_i()] = f32::from_bits(_mm_extract_ps::<1>(v) as u32).as_();
150                    dst[cn.b_i()] = f32::from_bits(_mm_extract_ps::<2>(v) as u32).as_();
151                }
152            }
153            if channels == 4 {
154                dst[cn.a_i()] = max_value;
155            }
156        }
157    }
158}
159
160impl<
161    T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
162    U: AsPrimitive<usize>,
163    const LAYOUT: u8,
164    const GRID_SIZE: usize,
165    const BIT_DEPTH: usize,
166    const BINS: usize,
167    const BARYCENTRIC_BINS: usize,
168> TransformExecutor<T>
169    for TransformLut4To3Avx<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
170where
171    f32: AsPrimitive<T>,
172    u32: AsPrimitive<T>,
173    (): LutBarycentricReduction<T, U>,
174{
175    fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
176        let cn = Layout::from(LAYOUT);
177        let channels = cn.channels();
178        if src.len() % 4 != 0 {
179            return Err(CmsError::LaneMultipleOfChannels);
180        }
181        if dst.len() % channels != 0 {
182            return Err(CmsError::LaneMultipleOfChannels);
183        }
184        let src_chunks = src.len() / 4;
185        let dst_chunks = dst.len() / channels;
186        if src_chunks != dst_chunks {
187            return Err(CmsError::LaneSizeMismatch);
188        }
189
190        unsafe {
191            if self.color_space == DataColorSpace::Lab
192                || (self.is_linear && self.color_space == DataColorSpace::Rgb)
193                || self.color_space == DataColorSpace::Xyz
194            {
195                self.transform_chunk(src, dst, Box::new(TrilinearAvxFmaDouble::<GRID_SIZE> {}));
196            } else {
197                match self.interpolation_method {
198                    #[cfg(feature = "options")]
199                    InterpolationMethod::Tetrahedral => {
200                        self.transform_chunk(
201                            src,
202                            dst,
203                            Box::new(TetrahedralAvxFmaDouble::<GRID_SIZE> {}),
204                        );
205                    }
206                    #[cfg(feature = "options")]
207                    InterpolationMethod::Pyramid => {
208                        self.transform_chunk(
209                            src,
210                            dst,
211                            Box::new(PyramidAvxFmaDouble::<GRID_SIZE> {}),
212                        );
213                    }
214                    #[cfg(feature = "options")]
215                    InterpolationMethod::Prism => {
216                        self.transform_chunk(
217                            src,
218                            dst,
219                            Box::new(PrismaticAvxFmaDouble::<GRID_SIZE> {}),
220                        );
221                    }
222                    InterpolationMethod::Linear => {
223                        self.transform_chunk(
224                            src,
225                            dst,
226                            Box::new(TrilinearAvxFmaDouble::<GRID_SIZE> {}),
227                        );
228                    }
229                }
230            }
231        }
232
233        Ok(())
234    }
235}
236
237pub(crate) struct AvxLut4x3Factory {}
238
239impl Lut4x3Factory for AvxLut4x3Factory {
240    fn make_transform_4x3<
241        T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible + 'static + Send + Sync,
242        const LAYOUT: u8,
243        const GRID_SIZE: usize,
244        const BIT_DEPTH: usize,
245    >(
246        lut: Vec<f32>,
247        options: TransformOptions,
248        color_space: DataColorSpace,
249        is_linear: bool,
250    ) -> Box<dyn TransformExecutor<T> + Send + Sync>
251    where
252        f32: AsPrimitive<T>,
253        u32: AsPrimitive<T>,
254        (): LutBarycentricReduction<T, u8>,
255        (): LutBarycentricReduction<T, u16>,
256    {
257        if options.prefer_fixed_point && BIT_DEPTH < 16 {
258            let q: f32 = if T::FINITE {
259                ((1i32 << BIT_DEPTH as i32) - 1) as f32
260            } else {
261                ((1i32 << 14i32) - 1) as f32
262            };
263            let lut = lut
264                .chunks_exact(3)
265                .map(|x| {
266                    AvxAlignedI16([
267                        (x[0] * q).round() as i16,
268                        (x[1] * q).round() as i16,
269                        (x[2] * q).round() as i16,
270                        0,
271                    ])
272                })
273                .collect::<Vec<_>>();
274            return match options.barycentric_weight_scale {
275                BarycentricWeightScale::Low => Box::new(TransformLut4To3AvxQ0_15::<
276                    T,
277                    u8,
278                    LAYOUT,
279                    GRID_SIZE,
280                    BIT_DEPTH,
281                    256,
282                    256,
283                > {
284                    lut,
285                    interpolation_method: options.interpolation_method,
286                    weights: BarycentricWeight::<i16>::create_ranged_256::<GRID_SIZE>(),
287                    _phantom: PhantomData,
288                    _phantom1: PhantomData,
289                    color_space,
290                    is_linear,
291                }),
292                #[cfg(feature = "options")]
293                BarycentricWeightScale::High => Box::new(TransformLut4To3AvxQ0_15::<
294                    T,
295                    u16,
296                    LAYOUT,
297                    GRID_SIZE,
298                    BIT_DEPTH,
299                    65536,
300                    65536,
301                > {
302                    lut,
303                    interpolation_method: options.interpolation_method,
304                    weights: BarycentricWeight::<i16>::create_binned::<GRID_SIZE, 65536>(),
305                    _phantom: PhantomData,
306                    _phantom1: PhantomData,
307                    color_space,
308                    is_linear,
309                }),
310            };
311        }
312        assert!(
313            std::arch::is_x86_feature_detected!("fma"),
314            "Internal configuration error, this feature might not be called without `fma` feature"
315        );
316        let lut = lut
317            .chunks_exact(3)
318            .map(|x| SseAlignedF32([x[0], x[1], x[2], 0f32]))
319            .collect::<Vec<_>>();
320        match options.barycentric_weight_scale {
321            BarycentricWeightScale::Low => {
322                Box::new(
323                    TransformLut4To3Avx::<T, u8, LAYOUT, GRID_SIZE, BIT_DEPTH, 256, 256> {
324                        lut,
325                        interpolation_method: options.interpolation_method,
326                        weights: BarycentricWeight::<f32>::create_ranged_256::<GRID_SIZE>(),
327                        _phantom: PhantomData,
328                        _phantom1: PhantomData,
329                        color_space,
330                        is_linear,
331                    },
332                )
333            }
334            #[cfg(feature = "options")]
335            BarycentricWeightScale::High => {
336                Box::new(
337                    TransformLut4To3Avx::<T, u16, LAYOUT, GRID_SIZE, BIT_DEPTH, 65536, 65536> {
338                        lut,
339                        interpolation_method: options.interpolation_method,
340                        weights: BarycentricWeight::<f32>::create_binned::<GRID_SIZE, 65536>(),
341                        _phantom: PhantomData,
342                        _phantom1: PhantomData,
343                        color_space,
344                        is_linear,
345                    },
346                )
347            }
348        }
349    }
350}
moxcms/conversions/avx/lut4_to_3.rs

moxcms/conversions/avx/
lut4_to_3.rs
