1use crate::conversions::LutBarycentricReduction;
30use crate::conversions::avx::interpolator_q0_15::*;
31use crate::conversions::interpolator::BarycentricWeight;
32use crate::transform::PointeeSizeExpressible;
33use crate::{CmsError, DataColorSpace, InterpolationMethod, Layout, TransformExecutor};
34use num_traits::AsPrimitive;
35use std::arch::x86_64::*;
36use std::marker::PhantomData;
37
38pub(crate) struct TransformLut4To3AvxQ0_15<
39 T,
40 U,
41 const LAYOUT: u8,
42 const GRID_SIZE: usize,
43 const BIT_DEPTH: usize,
44 const BINS: usize,
45 const BARYCENTRIC_BINS: usize,
46> {
47 pub(crate) lut: Vec<AvxAlignedI16>,
48 pub(crate) _phantom: PhantomData<T>,
49 pub(crate) _phantom1: PhantomData<U>,
50 pub(crate) interpolation_method: InterpolationMethod,
51 pub(crate) weights: Box<[BarycentricWeight<i16>; BINS]>,
52 pub(crate) color_space: DataColorSpace,
53 pub(crate) is_linear: bool,
54}
55
56impl<
57 T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
58 U: AsPrimitive<usize>,
59 const LAYOUT: u8,
60 const GRID_SIZE: usize,
61 const BIT_DEPTH: usize,
62 const BINS: usize,
63 const BARYCENTRIC_BINS: usize,
64> TransformLut4To3AvxQ0_15<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
65where
66 f32: AsPrimitive<T>,
67 u32: AsPrimitive<T>,
68 (): LutBarycentricReduction<T, U>,
69{
70 #[allow(unused_unsafe)]
71 #[target_feature(enable = "avx2")]
72 unsafe fn transform_chunk<'b, Interpolator: AvxMdInterpolationQ0_15Double<'b, GRID_SIZE>>(
73 &'b self,
74 src: &[T],
75 dst: &mut [T],
76 ) {
77 unsafe {
78 let cn = Layout::from(LAYOUT);
79 let channels = cn.channels();
80 let grid_size = GRID_SIZE as i32;
81 let grid_size3 = grid_size * grid_size * grid_size;
82
83 let f_value_scale = _mm_set1_ps(1. / ((1 << 14i32) - 1) as f32);
84 let max_value = ((1u32 << BIT_DEPTH) - 1).as_();
85 let v_max_scale = if T::FINITE {
86 _mm_set1_epi16(((1i32 << BIT_DEPTH) - 1) as i16)
87 } else {
88 _mm_set1_epi16(((1i32 << 14i32) - 1) as i16)
89 };
90
91 for (src, dst) in src.chunks_exact(4).zip(dst.chunks_exact_mut(channels)) {
92 let c = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
93 src[0],
94 );
95 let m = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
96 src[1],
97 );
98 let y = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
99 src[2],
100 );
101 let k = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
102 src[3],
103 );
104
105 let k_weights = self.weights[k.as_()];
106
107 let w: i32 = k_weights.x;
108 let w_n: i32 = k_weights.x_n;
109 const Q: i16 = ((1i32 << 15) - 1) as i16;
110 let t: i16 = k_weights.w;
111 let t_n: i16 = Q - t;
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 interpolator = Interpolator::new(table1, table2);
117 let v = interpolator.inter3_sse(c, m, y, &self.weights);
118 let (a0, b0) = (v.0.v, v.1.v);
119
120 let hp = _mm_mulhrs_epi16(_mm_set1_epi16(t_n), a0);
121 let v = _mm_add_epi16(hp, _mm_mulhrs_epi16(b0, _mm_set1_epi16(t)));
122
123 if T::FINITE {
124 let mut o = _mm_max_epi16(v, _mm_setzero_si128());
125 o = _mm_min_epi16(o, v_max_scale);
126 let x = _mm_extract_epi16::<0>(o);
127 let y = _mm_extract_epi16::<1>(o);
128 let z = _mm_extract_epi16::<2>(o);
129
130 dst[cn.r_i()] = (x as u32).as_();
131 dst[cn.g_i()] = (y as u32).as_();
132 dst[cn.b_i()] = (z as u32).as_();
133 } else {
134 let mut r = _mm_cvtepi32_ps(_mm_cvtepi16_epi32(v));
135 r = _mm_mul_ps(r, f_value_scale);
136 dst[cn.r_i()] = f32::from_bits(_mm_extract_ps::<0>(r) as u32).as_();
137 dst[cn.g_i()] = f32::from_bits(_mm_extract_ps::<1>(r) as u32).as_();
138 dst[cn.b_i()] = f32::from_bits(_mm_extract_ps::<2>(r) as u32).as_();
139 }
140 if channels == 4 {
141 dst[cn.a_i()] = max_value;
142 }
143 }
144 }
145 }
146}
147
148impl<
149 T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
150 U: AsPrimitive<usize>,
151 const LAYOUT: u8,
152 const GRID_SIZE: usize,
153 const BIT_DEPTH: usize,
154 const BINS: usize,
155 const BARYCENTRIC_BINS: usize,
156> TransformExecutor<T>
157 for TransformLut4To3AvxQ0_15<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
158where
159 f32: AsPrimitive<T>,
160 u32: AsPrimitive<T>,
161 (): LutBarycentricReduction<T, U>,
162{
163 fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
164 let cn = Layout::from(LAYOUT);
165 let channels = cn.channels();
166 if src.len() % 4 != 0 {
167 return Err(CmsError::LaneMultipleOfChannels);
168 }
169 if dst.len() % channels != 0 {
170 return Err(CmsError::LaneMultipleOfChannels);
171 }
172 let src_chunks = src.len() / 4;
173 let dst_chunks = dst.len() / channels;
174 if src_chunks != dst_chunks {
175 return Err(CmsError::LaneSizeMismatch);
176 }
177
178 unsafe {
179 if self.color_space == DataColorSpace::Lab
180 || (self.is_linear && self.color_space == DataColorSpace::Rgb)
181 || self.color_space == DataColorSpace::Xyz
182 {
183 self.transform_chunk::<TrilinearAvxQ0_15Double<GRID_SIZE>>(src, dst);
184 } else {
185 match self.interpolation_method {
186 #[cfg(feature = "options")]
187 InterpolationMethod::Tetrahedral => {
188 self.transform_chunk::<TetrahedralAvxQ0_15Double<GRID_SIZE>>(src, dst);
189 }
190 #[cfg(feature = "options")]
191 InterpolationMethod::Pyramid => {
192 self.transform_chunk::<PyramidAvxFmaQ0_15Double<GRID_SIZE>>(src, dst);
193 }
194 #[cfg(feature = "options")]
195 InterpolationMethod::Prism => {
196 self.transform_chunk::<PrismaticAvxQ0_15Double<GRID_SIZE>>(src, dst);
197 }
198 InterpolationMethod::Linear => {
199 self.transform_chunk::<TrilinearAvxQ0_15Double<GRID_SIZE>>(src, dst);
200 }
201 }
202 }
203 }
204
205 Ok(())
206 }
207}