1#![cfg(feature = "sse_luts")]
30use crate::conversions::LutBarycentricReduction;
31use crate::conversions::interpolator::BarycentricWeight;
32use crate::conversions::sse::interpolator_q0_15::*;
33use crate::transform::PointeeSizeExpressible;
34use crate::{CmsError, DataColorSpace, InterpolationMethod, Layout, TransformExecutor};
35use num_traits::AsPrimitive;
36#[cfg(target_arch = "x86")]
37use std::arch::x86::*;
38#[cfg(target_arch = "x86_64")]
39use std::arch::x86_64::*;
40use std::marker::PhantomData;
41
42pub(crate) struct TransformLut4To3SseQ0_15<
43 T,
44 U,
45 const LAYOUT: u8,
46 const GRID_SIZE: usize,
47 const BIT_DEPTH: usize,
48 const BINS: usize,
49 const BARYCENTRIC_BINS: usize,
50> {
51 pub(crate) lut: Vec<SseAlignedI16x4>,
52 pub(crate) _phantom: PhantomData<T>,
53 pub(crate) _phantom1: PhantomData<U>,
54 pub(crate) interpolation_method: InterpolationMethod,
55 pub(crate) weights: Box<[BarycentricWeight<i16>; BINS]>,
56 pub(crate) color_space: DataColorSpace,
57 pub(crate) is_linear: bool,
58}
59
60impl<
61 T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
62 U: AsPrimitive<usize>,
63 const LAYOUT: u8,
64 const GRID_SIZE: usize,
65 const BIT_DEPTH: usize,
66 const BINS: usize,
67 const BARYCENTRIC_BINS: usize,
68> TransformLut4To3SseQ0_15<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
69where
70 f32: AsPrimitive<T>,
71 u32: AsPrimitive<T>,
72 (): LutBarycentricReduction<T, U>,
73{
74 #[allow(unused_unsafe)]
75 #[target_feature(enable = "sse4.1")]
76 unsafe fn transform_chunk(
77 &self,
78 src: &[T],
79 dst: &mut [T],
80 interpolator: Box<dyn SseMdInterpolationQ0_15 + Send + Sync>,
81 ) {
82 unsafe {
83 let cn = Layout::from(LAYOUT);
84 let channels = cn.channels();
85 let grid_size = GRID_SIZE as i32;
86 let grid_size3 = grid_size * grid_size * grid_size;
87
88 let f_value_scale = _mm_set1_ps(1. / ((1 << 14i32) - 1) as f32);
89 let max_value = ((1u32 << BIT_DEPTH) - 1).as_();
90 let v_max_scale = if T::FINITE {
91 _mm_set1_epi16(((1i32 << BIT_DEPTH) - 1) as i16)
92 } else {
93 _mm_set1_epi16(((1i32 << 14i32) - 1) as i16)
94 };
95
96 for (src, dst) in src.chunks_exact(4).zip(dst.chunks_exact_mut(channels)) {
97 let c = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
98 src[0],
99 );
100 let m = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
101 src[1],
102 );
103 let y = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
104 src[2],
105 );
106 let k = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
107 src[3],
108 );
109
110 let k_weights = self.weights[k.as_()];
111
112 let w: i32 = k_weights.x;
113 let w_n: i32 = k_weights.x_n;
114 const Q: i16 = ((1i32 << 15) - 1) as i16;
115 let t: i16 = k_weights.w;
116 let t_n: i16 = Q - t;
117
118 let table1 = &self.lut[(w * grid_size3) as usize..];
119 let table2 = &self.lut[(w_n * grid_size3) as usize..];
120
121 let a0 = interpolator
122 .inter3_sse(table1, c.as_(), m.as_(), y.as_(), self.weights.as_slice())
123 .v;
124 let b0 = interpolator
125 .inter3_sse(table2, c.as_(), m.as_(), y.as_(), self.weights.as_slice())
126 .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
135 let x = _mm_extract_epi16::<0>(o);
136 let y = _mm_extract_epi16::<1>(o);
137 let z = _mm_extract_epi16::<2>(o);
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 } else {
143 let mut r = _mm_cvtepi32_ps(_mm_cvtepi16_epi32(v));
144 r = _mm_mul_ps(r, f_value_scale);
145 dst[cn.r_i()] = f32::from_bits(_mm_extract_ps::<0>(r) as u32).as_();
146 dst[cn.g_i()] = f32::from_bits(_mm_extract_ps::<1>(r) as u32).as_();
147 dst[cn.b_i()] = f32::from_bits(_mm_extract_ps::<2>(r) as u32).as_();
148 }
149 if channels == 4 {
150 dst[cn.a_i()] = max_value;
151 }
152 }
153 }
154 }
155}
156
157impl<
158 T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
159 U: AsPrimitive<usize>,
160 const LAYOUT: u8,
161 const GRID_SIZE: usize,
162 const BIT_DEPTH: usize,
163 const BINS: usize,
164 const BARYCENTRIC_BINS: usize,
165> TransformExecutor<T>
166 for TransformLut4To3SseQ0_15<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
167where
168 f32: AsPrimitive<T>,
169 u32: AsPrimitive<T>,
170 (): LutBarycentricReduction<T, U>,
171{
172 fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
173 let cn = Layout::from(LAYOUT);
174 let channels = cn.channels();
175 if src.len() % 4 != 0 {
176 return Err(CmsError::LaneMultipleOfChannels);
177 }
178 if dst.len() % channels != 0 {
179 return Err(CmsError::LaneMultipleOfChannels);
180 }
181 let src_chunks = src.len() / 4;
182 let dst_chunks = dst.len() / channels;
183 if src_chunks != dst_chunks {
184 return Err(CmsError::LaneSizeMismatch);
185 }
186
187 unsafe {
188 if self.color_space == DataColorSpace::Lab
189 || (self.is_linear && self.color_space == DataColorSpace::Rgb)
190 || self.color_space == DataColorSpace::Xyz
191 {
192 self.transform_chunk(src, dst, Box::new(TrilinearSseQ0_15::<GRID_SIZE> {}));
193 } else {
194 match self.interpolation_method {
195 #[cfg(feature = "options")]
196 InterpolationMethod::Tetrahedral => {
197 self.transform_chunk(
198 src,
199 dst,
200 Box::new(TetrahedralSseQ0_15::<GRID_SIZE> {}),
201 );
202 }
203 #[cfg(feature = "options")]
204 InterpolationMethod::Pyramid => {
205 self.transform_chunk(src, dst, Box::new(PyramidalSseQ0_15::<GRID_SIZE> {}));
206 }
207 #[cfg(feature = "options")]
208 InterpolationMethod::Prism => {
209 self.transform_chunk(src, dst, Box::new(PrismaticSseQ0_15::<GRID_SIZE> {}));
210 }
211 InterpolationMethod::Linear => {
212 self.transform_chunk(src, dst, Box::new(TrilinearSseQ0_15::<GRID_SIZE> {}));
213 }
214 }
215 }
216 }
217
218 Ok(())
219 }
220}