1use crate::conversions::LutBarycentricReduction;
30use crate::conversions::interpolator::BarycentricWeight;
31use crate::conversions::sse::interpolator_q0_15::*;
32use crate::transform::PointeeSizeExpressible;
33use crate::{CmsError, DataColorSpace, InterpolationMethod, Layout, TransformExecutor};
34use num_traits::AsPrimitive;
35#[cfg(target_arch = "x86")]
36use std::arch::x86::*;
37#[cfg(target_arch = "x86_64")]
38use std::arch::x86_64::*;
39use std::marker::PhantomData;
40
41pub(crate) struct TransformLut4To3SseQ0_15<
42 T,
43 U,
44 const LAYOUT: u8,
45 const GRID_SIZE: usize,
46 const BIT_DEPTH: usize,
47 const BINS: usize,
48 const BARYCENTRIC_BINS: usize,
49> {
50 pub(crate) lut: Vec<SseAlignedI16x4>,
51 pub(crate) _phantom: PhantomData<T>,
52 pub(crate) _phantom1: PhantomData<U>,
53 pub(crate) interpolation_method: InterpolationMethod,
54 pub(crate) weights: Box<[BarycentricWeight<i16>; BINS]>,
55 pub(crate) color_space: DataColorSpace,
56 pub(crate) is_linear: bool,
57}
58
59impl<
60 T: Copy + AsPrimitive<f32> + Default + PointeeSizeExpressible,
61 U: AsPrimitive<usize>,
62 const LAYOUT: u8,
63 const GRID_SIZE: usize,
64 const BIT_DEPTH: usize,
65 const BINS: usize,
66 const BARYCENTRIC_BINS: usize,
67> TransformLut4To3SseQ0_15<T, U, LAYOUT, GRID_SIZE, BIT_DEPTH, BINS, BARYCENTRIC_BINS>
68where
69 f32: AsPrimitive<T>,
70 u32: AsPrimitive<T>,
71 (): LutBarycentricReduction<T, U>,
72{
73 #[allow(unused_unsafe)]
74 #[target_feature(enable = "sse4.1")]
75 unsafe fn transform_chunk(
76 &self,
77 src: &[T],
78 dst: &mut [T],
79 interpolator: Box<dyn SseMdInterpolationQ0_15 + Send + Sync>,
80 ) {
81 unsafe {
82 let cn = Layout::from(LAYOUT);
83 let channels = cn.channels();
84 let grid_size = GRID_SIZE as i32;
85 let grid_size3 = grid_size * grid_size * grid_size;
86
87 let f_value_scale = _mm_set1_ps(1. / ((1 << 14i32) - 1) as f32);
88 let max_value = ((1u32 << BIT_DEPTH) - 1).as_();
89 let v_max_scale = if T::FINITE {
90 _mm_set1_epi16(((1i32 << BIT_DEPTH) - 1) as i16)
91 } else {
92 _mm_set1_epi16(((1i32 << 14i32) - 1) as i16)
93 };
94
95 for (src, dst) in src.chunks_exact(4).zip(dst.chunks_exact_mut(channels)) {
96 let c = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
97 src[0],
98 );
99 let m = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
100 src[1],
101 );
102 let y = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
103 src[2],
104 );
105 let k = <() as LutBarycentricReduction<T, U>>::reduce::<BIT_DEPTH, BARYCENTRIC_BINS>(
106 src[3],
107 );
108
109 let k_weights = self.weights[k.as_()];
110
111 let w: i32 = k_weights.x;
112 let w_n: i32 = k_weights.x_n;
113 const Q: i16 = ((1i32 << 15) - 1) as i16;
114 let t: i16 = k_weights.w;
115 let t_n: i16 = Q - t;
116
117 let table1 = &self.lut[(w * grid_size3) as usize..];
118 let table2 = &self.lut[(w_n * grid_size3) as usize..];
119
120 let a0 = interpolator
121 .inter3_sse(table1, c.as_(), m.as_(), y.as_(), self.weights.as_slice())
122 .v;
123 let b0 = interpolator
124 .inter3_sse(table2, c.as_(), m.as_(), y.as_(), self.weights.as_slice())
125 .v;
126
127 let hp = _mm_mulhrs_epi16(_mm_set1_epi16(t_n), a0);
128 let v = _mm_add_epi16(hp, _mm_mulhrs_epi16(b0, _mm_set1_epi16(t)));
129
130 if T::FINITE {
131 let mut o = _mm_max_epi16(v, _mm_setzero_si128());
132 o = _mm_min_epi16(o, v_max_scale);
133
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 TransformLut4To3SseQ0_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(TrilinearSseQ0_15::<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(TetrahedralSseQ0_15::<GRID_SIZE> {}),
200 );
201 }
202 #[cfg(feature = "options")]
203 InterpolationMethod::Pyramid => {
204 self.transform_chunk(src, dst, Box::new(PyramidalSseQ0_15::<GRID_SIZE> {}));
205 }
206 #[cfg(feature = "options")]
207 InterpolationMethod::Prism => {
208 self.transform_chunk(src, dst, Box::new(PrismaticSseQ0_15::<GRID_SIZE> {}));
209 }
210 InterpolationMethod::Linear => {
211 self.transform_chunk(src, dst, Box::new(TrilinearSseQ0_15::<GRID_SIZE> {}));
212 }
213 }
214 }
215 }
216
217 Ok(())
218 }
219}