1#![cfg(feature = "avx_luts")]
30use crate::conversions::LutBarycentricReduction;
31use crate::conversions::avx::interpolator_q0_15::*;
32use crate::conversions::interpolator::BarycentricWeight;
33use crate::transform::PointeeSizeExpressible;
34use crate::{CmsError, DataColorSpace, InterpolationMethod, Layout, TransformExecutor};
35use num_traits::AsPrimitive;
36use std::arch::x86_64::*;
37use std::marker::PhantomData;
38
39pub(crate) struct TransformLut4To3AvxQ0_15<
40 T,
41 U,
42 const LAYOUT: u8,
43 const GRID_SIZE: usize,
44 const BIT_DEPTH: usize,
45 const BINS: usize,
46 const BARYCENTRIC_BINS: usize,
47> {
48 pub(crate) lut: Vec<AvxAlignedI16>,
49 pub(crate) _phantom: PhantomData<T>,
50 pub(crate) _phantom1: PhantomData<U>,
51 pub(crate) interpolation_method: InterpolationMethod,
52 pub(crate) weights: Box<[BarycentricWeight<i16>; BINS]>,
53 pub(crate) color_space: DataColorSpace,
54 pub(crate) is_linear: bool,
55}
56
57impl<
58 T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
59 U: AsPrimitive<usize>,
60 const LAYOUT: u8,
61 const GRID_SIZE: usize,
62 const BIT_DEPTH: usize,
63 const BINS: usize,
64 const BARYCENTRIC_BINS: usize,
65> TransformLut4To3AvxQ0_15<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
66where
67 f32: AsPrimitive<T>,
68 u32: AsPrimitive<T>,
69 (): LutBarycentricReduction<T, U>,
70{
71 #[allow(unused_unsafe)]
72 #[target_feature(enable = "avx2")]
73 unsafe fn transform_chunk(
74 &self,
75 src: &[T],
76 dst: &mut [T],
77 interpolator: Box<dyn AvxMdInterpolationQ0_15Double + Send + Sync>,
78 ) {
79 unsafe {
80 let cn = Layout::from(LAYOUT);
81 let channels = cn.channels();
82 let grid_size = GRID_SIZE as i32;
83 let grid_size3 = grid_size * grid_size * grid_size;
84
85 let f_value_scale = _mm_set1_ps(1. / ((1 << 14i32) - 1) as f32);
86 let max_value = ((1u32 << BIT_DEPTH) - 1).as_();
87 let v_max_scale = if T::FINITE {
88 _mm_set1_epi16(((1i32 << BIT_DEPTH) - 1) as i16)
89 } else {
90 _mm_set1_epi16(((1i32 << 14i32) - 1) as i16)
91 };
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 const Q: i16 = ((1i32 << 15) - 1) as i16;
112 let t: i16 = k_weights.w;
113 let t_n: i16 = Q - t;
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 v = interpolator.inter3_sse(
119 table1,
120 table2,
121 c.as_(),
122 m.as_(),
123 y.as_(),
124 self.weights.as_slice(),
125 );
126 let (a0, b0) = (v.0.v, v.1.v);
127
128 let hp = _mm_mulhrs_epi16(_mm_set1_epi16(t_n), a0);
129 let v = _mm_add_epi16(hp, _mm_mulhrs_epi16(b0, _mm_set1_epi16(t)));
130
131 if T::FINITE {
132 let mut o = _mm_max_epi16(v, _mm_setzero_si128());
133 o = _mm_min_epi16(o, v_max_scale);
134 let x = _mm_extract_epi16::<0>(o);
135 let y = _mm_extract_epi16::<1>(o);
136 let z = _mm_extract_epi16::<2>(o);
137
138 dst[cn.r_i()] = (x as u32).as_();
139 dst[cn.g_i()] = (y as u32).as_();
140 dst[cn.b_i()] = (z as u32).as_();
141 } else {
142 let mut r = _mm_cvtepi32_ps(_mm_cvtepi16_epi32(v));
143 r = _mm_mul_ps(r, f_value_scale);
144 dst[cn.r_i()] = f32::from_bits(_mm_extract_ps::<0>(r) as u32).as_();
145 dst[cn.g_i()] = f32::from_bits(_mm_extract_ps::<1>(r) as u32).as_();
146 dst[cn.b_i()] = f32::from_bits(_mm_extract_ps::<2>(r) as u32).as_();
147 }
148 if channels == 4 {
149 dst[cn.a_i()] = max_value;
150 }
151 }
152 }
153 }
154}
155
156impl<
157 T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
158 U: AsPrimitive<usize>,
159 const LAYOUT: u8,
160 const GRID_SIZE: usize,
161 const BIT_DEPTH: usize,
162 const BINS: usize,
163 const BARYCENTRIC_BINS: usize,
164> TransformExecutor<T>
165 for TransformLut4To3AvxQ0_15<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
166where
167 f32: AsPrimitive<T>,
168 u32: AsPrimitive<T>,
169 (): LutBarycentricReduction<T, U>,
170{
171 fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
172 let cn = Layout::from(LAYOUT);
173 let channels = cn.channels();
174 if src.len() % 4 != 0 {
175 return Err(CmsError::LaneMultipleOfChannels);
176 }
177 if dst.len() % channels != 0 {
178 return Err(CmsError::LaneMultipleOfChannels);
179 }
180 let src_chunks = src.len() / 4;
181 let dst_chunks = dst.len() / channels;
182 if src_chunks != dst_chunks {
183 return Err(CmsError::LaneSizeMismatch);
184 }
185
186 unsafe {
187 if self.color_space == DataColorSpace::Lab
188 || (self.is_linear && self.color_space == DataColorSpace::Rgb)
189 || self.color_space == DataColorSpace::Xyz
190 {
191 self.transform_chunk(src, dst, Box::new(TrilinearAvxQ0_15Double::<GRID_SIZE> {}));
192 } else {
193 match self.interpolation_method {
194 #[cfg(feature = "options")]
195 InterpolationMethod::Tetrahedral => {
196 self.transform_chunk(
197 src,
198 dst,
199 Box::new(TetrahedralAvxQ0_15Double::<GRID_SIZE> {}),
200 );
201 }
202 #[cfg(feature = "options")]
203 InterpolationMethod::Pyramid => {
204 self.transform_chunk(
205 src,
206 dst,
207 Box::new(PyramidAvxFmaQ0_15Double::<GRID_SIZE> {}),
208 );
209 }
210 #[cfg(feature = "options")]
211 InterpolationMethod::Prism => {
212 self.transform_chunk(
213 src,
214 dst,
215 Box::new(PrismaticAvxQ0_15Double::<GRID_SIZE> {}),
216 );
217 }
218 InterpolationMethod::Linear => {
219 self.transform_chunk(
220 src,
221 dst,
222 Box::new(TrilinearAvxQ0_15Double::<GRID_SIZE> {}),
223 );
224 }
225 }
226 }
227 }
228
229 Ok(())
230 }
231}