moxcms/conversions/avx/
rgb_xyz_q2_13_opt.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::avx::rgb_xyz::AvxAlignedU16;
30use crate::conversions::avx::rgb_xyz_q2_13::_xmm_broadcast_epi32;
31use crate::conversions::rgbxyz_fixed::TransformMatrixShaperFpOptVec;
32use crate::transform::PointeeSizeExpressible;
33use crate::{CmsError, Layout, TransformExecutor};
34use num_traits::AsPrimitive;
35use std::arch::x86_64::*;
36
37pub(crate) struct TransformShaperRgbQ2_13OptAvx<
38    T: Copy,
39    const SRC_LAYOUT: u8,
40    const DST_LAYOUT: u8,
41    const PRECISION: i32,
42> {
43    pub(crate) profile: TransformMatrixShaperFpOptVec<i32, i16, T>,
44    pub(crate) bit_depth: usize,
45    pub(crate) gamma_lut: usize,
46}
47
48impl<
49    T: Copy + PointeeSizeExpressible + 'static,
50    const SRC_LAYOUT: u8,
51    const DST_LAYOUT: u8,
52    const PRECISION: i32,
53> TransformShaperRgbQ2_13OptAvx<T, SRC_LAYOUT, DST_LAYOUT, PRECISION>
54where
55    u32: AsPrimitive<T>,
56{
57    #[target_feature(enable = "avx2")]
58    unsafe fn transform_avx2(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
59        let src_cn = Layout::from(SRC_LAYOUT);
60        let dst_cn = Layout::from(DST_LAYOUT);
61        let src_channels = src_cn.channels();
62        let dst_channels = dst_cn.channels();
63
64        let mut temporary0 = AvxAlignedU16([0; 16]);
65
66        if src.len() / src_channels != dst.len() / dst_channels {
67            return Err(CmsError::LaneSizeMismatch);
68        }
69        if src.len() % src_channels != 0 {
70            return Err(CmsError::LaneMultipleOfChannels);
71        }
72        if dst.len() % dst_channels != 0 {
73            return Err(CmsError::LaneMultipleOfChannels);
74        }
75
76        let t = self.profile.adaptation_matrix.transpose();
77
78        let max_colors = ((1 << self.bit_depth) - 1).as_();
79
80        // safety precondition for linearization table
81        if T::FINITE {
82            let cap = (1 << self.bit_depth) - 1;
83            assert!(self.profile.linear.len() >= cap);
84        } else {
85            assert!(self.profile.linear.len() >= T::NOT_FINITE_LINEAR_TABLE_SIZE);
86        }
87
88        let lut_lin = &self.profile.linear;
89
90        unsafe {
91            let m0 = _mm256_setr_epi16(
92                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],
93                t.v[1][0], t.v[0][1], t.v[1][1], t.v[0][2], t.v[1][2], 0, 0,
94            );
95            let m2 = _mm256_setr_epi16(
96                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,
97                t.v[2][2], 1, 0, 0,
98            );
99
100            let rnd_val = ((1i32 << (PRECISION - 1)) as i16).to_ne_bytes();
101            let rnd = _mm256_set1_epi32(i32::from_ne_bytes([0, 0, rnd_val[0], rnd_val[1]]));
102
103            let zeros = _mm256_setzero_si256();
104
105            let v_max_value = _mm256_set1_epi32(self.gamma_lut as i32 - 1);
106
107            let (mut r0, mut g0, mut b0, mut a0);
108            let (mut r1, mut g1, mut b1, mut a1);
109
110            let mut src_iter = src.chunks_exact(src_channels * 2);
111
112            if let Some(src0) = src_iter.next() {
113                r0 = _xmm_broadcast_epi32(lut_lin.get_unchecked(src0[src_cn.r_i()]._as_usize()));
114                g0 = _xmm_broadcast_epi32(lut_lin.get_unchecked(src0[src_cn.g_i()]._as_usize()));
115                b0 = _xmm_broadcast_epi32(lut_lin.get_unchecked(src0[src_cn.b_i()]._as_usize()));
116
117                r1 = _xmm_broadcast_epi32(
118                    lut_lin.get_unchecked(src0[src_cn.r_i() + src_channels]._as_usize()),
119                );
120                g1 = _xmm_broadcast_epi32(
121                    lut_lin.get_unchecked(src0[src_cn.g_i() + src_channels]._as_usize()),
122                );
123                b1 = _xmm_broadcast_epi32(
124                    lut_lin.get_unchecked(src0[src_cn.b_i() + src_channels]._as_usize()),
125                );
126
127                a0 = if src_channels == 4 {
128                    src0[src_cn.a_i()]
129                } else {
130                    max_colors
131                };
132                a1 = if src_channels == 4 {
133                    src0[src_cn.a_i() + src_channels]
134                } else {
135                    max_colors
136                };
137            } else {
138                r0 = _mm_setzero_si128();
139                g0 = _mm_setzero_si128();
140                b0 = _mm_setzero_si128();
141                a0 = max_colors;
142                r1 = _mm_setzero_si128();
143                g1 = _mm_setzero_si128();
144                b1 = _mm_setzero_si128();
145                a1 = max_colors;
146            }
147
148            for (src, dst) in src_iter.zip(dst.chunks_exact_mut(dst_channels * 2)) {
149                let zr0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(r0), r1);
150                let mut zg0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(g0), g1);
151                let zb0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(b0), b1);
152                zg0 = _mm256_slli_epi32::<16>(zg0);
153
154                let zrg0 = _mm256_or_si256(zr0, zg0);
155                let zbz0 = _mm256_or_si256(zb0, rnd);
156
157                let va0 = _mm256_madd_epi16(zrg0, m0);
158                let va1 = _mm256_madd_epi16(zbz0, m2);
159
160                let mut v0 = _mm256_add_epi32(va0, va1);
161
162                v0 = _mm256_srai_epi32::<PRECISION>(v0);
163                v0 = _mm256_max_epi32(v0, zeros);
164                v0 = _mm256_min_epi32(v0, v_max_value);
165
166                _mm256_store_si256(temporary0.0.as_mut_ptr() as *mut _, v0);
167
168                r0 = _xmm_broadcast_epi32(lut_lin.get_unchecked(src[src_cn.r_i()]._as_usize()));
169                g0 = _xmm_broadcast_epi32(lut_lin.get_unchecked(src[src_cn.g_i()]._as_usize()));
170                b0 = _xmm_broadcast_epi32(lut_lin.get_unchecked(src[src_cn.b_i()]._as_usize()));
171
172                r1 = _xmm_broadcast_epi32(
173                    lut_lin.get_unchecked(src[src_cn.r_i() + src_channels]._as_usize()),
174                );
175                g1 = _xmm_broadcast_epi32(
176                    lut_lin.get_unchecked(src[src_cn.g_i() + src_channels]._as_usize()),
177                );
178                b1 = _xmm_broadcast_epi32(
179                    lut_lin.get_unchecked(src[src_cn.b_i() + src_channels]._as_usize()),
180                );
181
182                dst[dst_cn.r_i()] = self.profile.gamma[temporary0.0[0] as usize];
183                dst[dst_cn.g_i()] = self.profile.gamma[temporary0.0[2] as usize];
184                dst[dst_cn.b_i()] = self.profile.gamma[temporary0.0[4] as usize];
185                if dst_channels == 4 {
186                    dst[dst_cn.a_i()] = a0;
187                }
188
189                dst[dst_cn.r_i() + dst_channels] = self.profile.gamma[temporary0.0[8] as usize];
190                dst[dst_cn.g_i() + dst_channels] = self.profile.gamma[temporary0.0[10] as usize];
191                dst[dst_cn.b_i() + dst_channels] = self.profile.gamma[temporary0.0[12] as usize];
192                if dst_channels == 4 {
193                    dst[dst_cn.a_i() + dst_channels] = a1;
194                }
195
196                a0 = if src_channels == 4 {
197                    src[src_cn.a_i()]
198                } else {
199                    max_colors
200                };
201                a1 = if src_channels == 4 {
202                    src[src_cn.a_i() + src_channels]
203                } else {
204                    max_colors
205                };
206            }
207
208            if let Some(dst) = dst.chunks_exact_mut(dst_channels * 2).last() {
209                let zr0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(r0), r1);
210                let mut zg0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(g0), g1);
211                let zb0 = _mm256_inserti128_si256::<1>(_mm256_castsi128_si256(b0), b1);
212                zg0 = _mm256_slli_epi32::<16>(zg0);
213
214                let zrg0 = _mm256_or_si256(zr0, zg0);
215                let zbz0 = _mm256_or_si256(zb0, rnd);
216
217                let va0 = _mm256_madd_epi16(zrg0, m0);
218                let va1 = _mm256_madd_epi16(zbz0, m2);
219
220                let mut v0 = _mm256_add_epi32(va0, va1);
221
222                v0 = _mm256_srai_epi32::<PRECISION>(v0);
223                v0 = _mm256_max_epi32(v0, zeros);
224                v0 = _mm256_min_epi32(v0, v_max_value);
225
226                _mm256_store_si256(temporary0.0.as_mut_ptr() as *mut _, v0);
227
228                dst[dst_cn.r_i()] = self.profile.gamma[temporary0.0[0] as usize];
229                dst[dst_cn.g_i()] = self.profile.gamma[temporary0.0[2] as usize];
230                dst[dst_cn.b_i()] = self.profile.gamma[temporary0.0[4] as usize];
231                if dst_channels == 4 {
232                    dst[dst_cn.a_i()] = a0;
233                }
234
235                dst[dst_cn.r_i() + dst_channels] = self.profile.gamma[temporary0.0[8] as usize];
236                dst[dst_cn.g_i() + dst_channels] = self.profile.gamma[temporary0.0[10] as usize];
237                dst[dst_cn.b_i() + dst_channels] = self.profile.gamma[temporary0.0[12] as usize];
238                if dst_channels == 4 {
239                    dst[dst_cn.a_i() + dst_channels] = a1;
240                }
241            }
242
243            let src = src.chunks_exact(src_channels * 2).remainder();
244            let dst = dst.chunks_exact_mut(dst_channels * 2).into_remainder();
245
246            for (src, dst) in src
247                .chunks_exact(src_channels)
248                .zip(dst.chunks_exact_mut(dst_channels))
249            {
250                let r = _xmm_broadcast_epi32(lut_lin.get_unchecked(src[src_cn.r_i()]._as_usize()));
251                let mut g =
252                    _xmm_broadcast_epi32(lut_lin.get_unchecked(src[src_cn.g_i()]._as_usize()));
253                let b = _xmm_broadcast_epi32(lut_lin.get_unchecked(src[src_cn.b_i()]._as_usize()));
254
255                g = _mm_slli_epi32::<16>(g);
256
257                let a = if src_channels == 4 {
258                    src[src_cn.a_i()]
259                } else {
260                    max_colors
261                };
262
263                let zrg0 = _mm_or_si128(r, g);
264                let zbz0 = _mm_or_si128(b, _mm256_castsi256_si128(rnd));
265
266                let v0 = _mm_madd_epi16(zrg0, _mm256_castsi256_si128(m0));
267                let v1 = _mm_madd_epi16(zbz0, _mm256_castsi256_si128(m2));
268
269                let mut v = _mm_add_epi32(v0, v1);
270
271                v = _mm_srai_epi32::<PRECISION>(v);
272                v = _mm_max_epi32(v, _mm_setzero_si128());
273                v = _mm_min_epi32(v, _mm256_castsi256_si128(v_max_value));
274
275                _mm_store_si128(temporary0.0.as_mut_ptr() as *mut _, v);
276
277                dst[dst_cn.r_i()] = self.profile.gamma[temporary0.0[0] as usize];
278                dst[dst_cn.g_i()] = self.profile.gamma[temporary0.0[2] as usize];
279                dst[dst_cn.b_i()] = self.profile.gamma[temporary0.0[4] as usize];
280                if dst_channels == 4 {
281                    dst[dst_cn.a_i()] = a;
282                }
283            }
284        }
285
286        Ok(())
287    }
288}
289
290impl<
291    T: Copy + PointeeSizeExpressible + 'static + Default,
292    const SRC_LAYOUT: u8,
293    const DST_LAYOUT: u8,
294    const PRECISION: i32,
295> TransformExecutor<T> for TransformShaperRgbQ2_13OptAvx<T, SRC_LAYOUT, DST_LAYOUT, PRECISION>
296where
297    u32: AsPrimitive<T>,
298{
299    fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
300        unsafe { self.transform_avx2(src, dst) }
301    }
302}
moxcms/conversions/avx/rgb_xyz_q2_13_opt.rs

moxcms/conversions/avx/
rgb_xyz_q2_13_opt.rs
