| 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 |
*/ |
| 29 |
use crate::conversions::avx::rgb_xyz::AvxAlignedU16; |
| 30 |
use crate::conversions::rgbxyz::TransformMatrixShaperOptimizedV; |
| 31 |
use crate::transform::PointeeSizeExpressible; |
| 32 |
use crate::{CmsError, Layout, TransformExecutor}; |
| 33 |
use num_traits::AsPrimitive; |
| 34 |
use std::arch::x86_64::*; |
| 35 |
|
| 36 |
pub(crate) struct TransformShaperRgbOptAvx< |
| 37 |
T: Clone + Copy + 'static + PointeeSizeExpressible + Default, |
| 38 |
const SRC_LAYOUT: u8, |
| 39 |
const DST_LAYOUT: u8, |
| 40 |
> { |
| 41 |
pub(crate) profile: TransformMatrixShaperOptimizedV<T>, |
| 42 |
pub(crate) bit_depth: usize, |
| 43 |
pub(crate) gamma_lut: usize, |
| 44 |
} |
| 45 |
|
| 46 |
impl< |
| 47 |
T: Clone + Copy + 'static + PointeeSizeExpressible + Default, |
| 48 |
const SRC_LAYOUT: u8, |
| 49 |
const DST_LAYOUT: u8, |
| 50 |
> TransformShaperRgbOptAvx<T, SRC_LAYOUT, DST_LAYOUT> |
| 51 |
where |
| 52 |
u32: AsPrimitive<T>, |
| 53 |
{ |
| 54 |
#[target_feature(enable = "avx2", enable = "fma")] |
| 55 |
unsafe fn transform_impl<const FMA: bool>( |
| 56 |
&self, |
| 57 |
src: &[T], |
| 58 |
dst: &mut [T], |
| 59 |
) -> Result<(), CmsError> { |
| 60 |
let src_cn = Layout::from(SRC_LAYOUT); |
| 61 |
let dst_cn = Layout::from(DST_LAYOUT); |
| 62 |
let src_channels = src_cn.channels(); |
| 63 |
let dst_channels = dst_cn.channels(); |
| 64 |
|
| 65 |
let mut temporary0 = AvxAlignedU16([0; 16]); |
| 66 |
|
| 67 |
if src.len() / src_channels != dst.len() / dst_channels { |
| 68 |
return Err(CmsError::LaneSizeMismatch); |
| 69 |
} |
| 70 |
if src.len() % src_channels != 0 { |
| 71 |
return Err(CmsError::LaneMultipleOfChannels); |
| 72 |
} |
| 73 |
if dst.len() % dst_channels != 0 { |
| 74 |
return Err(CmsError::LaneMultipleOfChannels); |
| 75 |
} |
| 76 |
|
| 77 |
let t = self.profile.adaptation_matrix.transpose(); |
| 78 |
|
| 79 |
let scale = (self.gamma_lut - 1) as f32; |
| 80 |
let max_colors: T = ((1 << self.bit_depth) - 1).as_(); |
| 81 |
|
| 82 |
// safety precondition for linearization table |
| 83 |
if T::FINITE { |
| 84 |
let cap = (1 << self.bit_depth) - 1; |
| 85 |
assert!(self.profile.linear.len() >= cap); |
| 86 |
} else { |
| 87 |
assert!(self.profile.linear.len() >= T::NOT_FINITE_LINEAR_TABLE_SIZE); |
| 88 |
} |
| 89 |
|
| 90 |
let lut_lin = &self.profile.linear; |
| 91 |
|
| 92 |
unsafe { |
| 93 |
let m0 = _mm256_setr_ps( |
| 94 |
t.v[0][0], t.v[0][1], t.v[0][2], 0., t.v[0][0], t.v[0][1], t.v[0][2], 0., |
| 95 |
); |
| 96 |
let m1 = _mm256_setr_ps( |
| 97 |
t.v[1][0], t.v[1][1], t.v[1][2], 0., t.v[1][0], t.v[1][1], t.v[1][2], 0., |
| 98 |
); |
| 99 |
let m2 = _mm256_setr_ps( |
| 100 |
t.v[2][0], t.v[2][1], t.v[2][2], 0., t.v[2][0], t.v[2][1], t.v[2][2], 0., |
| 101 |
); |
| 102 |
|
| 103 |
let zeros = _mm_setzero_ps(); |
| 104 |
|
| 105 |
let v_scale = _mm256_set1_ps(scale); |
| 106 |
|
| 107 |
let mut src = src; |
| 108 |
let mut dst = dst; |
| 109 |
|
| 110 |
let mut src_iter = src.chunks_exact(src_channels * 2); |
| 111 |
let dst_iter = dst.chunks_exact_mut(dst_channels * 2); |
| 112 |
|
| 113 |
let (mut r0, mut g0, mut b0, mut a0); |
| 114 |
let (mut r1, mut g1, mut b1, mut a1); |
| 115 |
|
| 116 |
if let Some(src) = src_iter.next() { |
| 117 |
r0 = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.r_i()]._as_usize())); |
| 118 |
g0 = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.g_i()]._as_usize())); |
| 119 |
b0 = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.b_i()]._as_usize())); |
| 120 |
r1 = _mm_broadcast_ss( |
| 121 |
lut_lin.get_unchecked(src[src_cn.r_i() + src_channels]._as_usize()), |
| 122 |
); |
| 123 |
g1 = _mm_broadcast_ss( |
| 124 |
lut_lin.get_unchecked(src[src_cn.g_i() + src_channels]._as_usize()), |
| 125 |
); |
| 126 |
b1 = _mm_broadcast_ss( |
| 127 |
lut_lin.get_unchecked(src[src_cn.b_i() + src_channels]._as_usize()), |
| 128 |
); |
| 129 |
a0 = if src_channels == 4 { |
| 130 |
src[src_cn.a_i()] |
| 131 |
} else { |
| 132 |
max_colors |
| 133 |
}; |
| 134 |
a1 = if src_channels == 4 { |
| 135 |
src[src_cn.a_i() + src_channels] |
| 136 |
} else { |
| 137 |
max_colors |
| 138 |
}; |
| 139 |
} else { |
| 140 |
r0 = _mm_setzero_ps(); |
| 141 |
g0 = _mm_setzero_ps(); |
| 142 |
b0 = _mm_setzero_ps(); |
| 143 |
a0 = max_colors; |
| 144 |
r1 = _mm_setzero_ps(); |
| 145 |
g1 = _mm_setzero_ps(); |
| 146 |
b1 = _mm_setzero_ps(); |
| 147 |
a1 = max_colors; |
| 148 |
} |
| 149 |
|
| 150 |
for (src, dst) in src_iter.zip(dst_iter) { |
| 151 |
let r = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(r0), r1); |
| 152 |
let g = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(g0), g1); |
| 153 |
let b = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(b0), b1); |
| 154 |
|
| 155 |
let mut v = if FMA { |
| 156 |
let v0 = _mm256_mul_ps(r, m0); |
| 157 |
let v1 = _mm256_fmadd_ps(g, m1, v0); |
| 158 |
_mm256_fmadd_ps(b, m2, v1) |
| 159 |
} else { |
| 160 |
let v0 = _mm256_mul_ps(r, m0); |
| 161 |
let v1 = _mm256_mul_ps(g, m1); |
| 162 |
let v2 = _mm256_mul_ps(b, m2); |
| 163 |
|
| 164 |
_mm256_add_ps(_mm256_add_ps(v0, v1), v2) |
| 165 |
}; |
| 166 |
|
| 167 |
v = _mm256_max_ps(v, _mm256_setzero_ps()); |
| 168 |
v = _mm256_mul_ps(v, v_scale); |
| 169 |
v = _mm256_min_ps(v, v_scale); |
| 170 |
|
| 171 |
let zx = _mm256_cvtps_epi32(v); |
| 172 |
_mm256_store_si256(temporary0.0.as_mut_ptr() as *mut _, zx); |
| 173 |
|
| 174 |
r0 = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.r_i()]._as_usize())); |
| 175 |
g0 = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.g_i()]._as_usize())); |
| 176 |
b0 = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.b_i()]._as_usize())); |
| 177 |
r1 = _mm_broadcast_ss( |
| 178 |
lut_lin.get_unchecked(src[src_cn.r_i() + src_channels]._as_usize()), |
| 179 |
); |
| 180 |
g1 = _mm_broadcast_ss( |
| 181 |
lut_lin.get_unchecked(src[src_cn.g_i() + src_channels]._as_usize()), |
| 182 |
); |
| 183 |
b1 = _mm_broadcast_ss( |
| 184 |
lut_lin.get_unchecked(src[src_cn.b_i() + src_channels]._as_usize()), |
| 185 |
); |
| 186 |
|
| 187 |
dst[dst_cn.r_i()] = self.profile.gamma[temporary0.0[0] as usize]; |
| 188 |
dst[dst_cn.g_i()] = self.profile.gamma[temporary0.0[2] as usize]; |
| 189 |
dst[dst_cn.b_i()] = self.profile.gamma[temporary0.0[4] as usize]; |
| 190 |
if dst_channels == 4 { |
| 191 |
dst[dst_cn.a_i()] = a0; |
| 192 |
} |
| 193 |
|
| 194 |
dst[dst_cn.r_i() + dst_channels] = self.profile.gamma[temporary0.0[8] as usize]; |
| 195 |
dst[dst_cn.g_i() + dst_channels] = self.profile.gamma[temporary0.0[10] as usize]; |
| 196 |
dst[dst_cn.b_i() + dst_channels] = self.profile.gamma[temporary0.0[12] as usize]; |
| 197 |
if dst_channels == 4 { |
| 198 |
dst[dst_cn.a_i() + dst_channels] = a1; |
| 199 |
} |
| 200 |
|
| 201 |
a0 = if src_channels == 4 { |
| 202 |
src[src_cn.a_i()] |
| 203 |
} else { |
| 204 |
max_colors |
| 205 |
}; |
| 206 |
a1 = if src_channels == 4 { |
| 207 |
src[src_cn.a_i() + src_channels] |
| 208 |
} else { |
| 209 |
max_colors |
| 210 |
}; |
| 211 |
} |
| 212 |
|
| 213 |
if let Some(dst) = dst.chunks_exact_mut(dst_channels * 2).last() { |
| 214 |
let r = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(r0), r1); |
| 215 |
let g = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(g0), g1); |
| 216 |
let b = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(b0), b1); |
| 217 |
|
| 218 |
let mut v = if FMA { |
| 219 |
let v0 = _mm256_mul_ps(r, m0); |
| 220 |
let v1 = _mm256_fmadd_ps(g, m1, v0); |
| 221 |
_mm256_fmadd_ps(b, m2, v1) |
| 222 |
} else { |
| 223 |
let v0 = _mm256_mul_ps(r, m0); |
| 224 |
let v1 = _mm256_mul_ps(g, m1); |
| 225 |
let v2 = _mm256_mul_ps(b, m2); |
| 226 |
|
| 227 |
_mm256_add_ps(_mm256_add_ps(v0, v1), v2) |
| 228 |
}; |
| 229 |
|
| 230 |
v = _mm256_max_ps(v, _mm256_setzero_ps()); |
| 231 |
v = _mm256_mul_ps(v, v_scale); |
| 232 |
v = _mm256_min_ps(v, v_scale); |
| 233 |
|
| 234 |
let zx = _mm256_cvtps_epi32(v); |
| 235 |
_mm256_store_si256(temporary0.0.as_mut_ptr() as *mut _, zx); |
| 236 |
|
| 237 |
dst[dst_cn.r_i()] = self.profile.gamma[temporary0.0[0] as usize]; |
| 238 |
dst[dst_cn.g_i()] = self.profile.gamma[temporary0.0[2] as usize]; |
| 239 |
dst[dst_cn.b_i()] = self.profile.gamma[temporary0.0[4] as usize]; |
| 240 |
if dst_channels == 4 { |
| 241 |
dst[dst_cn.a_i()] = a0; |
| 242 |
} |
| 243 |
|
| 244 |
dst[dst_cn.r_i() + dst_channels] = self.profile.gamma[temporary0.0[8] as usize]; |
| 245 |
dst[dst_cn.g_i() + dst_channels] = self.profile.gamma[temporary0.0[10] as usize]; |
| 246 |
dst[dst_cn.b_i() + dst_channels] = self.profile.gamma[temporary0.0[12] as usize]; |
| 247 |
if dst_channels == 4 { |
| 248 |
dst[dst_cn.a_i() + dst_channels] = a1; |
| 249 |
} |
| 250 |
} |
| 251 |
|
| 252 |
src = src.chunks_exact(src_channels * 2).remainder(); |
| 253 |
dst = dst.chunks_exact_mut(dst_channels * 2).into_remainder(); |
| 254 |
|
| 255 |
for (src, dst) in src |
| 256 |
.chunks_exact(src_channels) |
| 257 |
.zip(dst.chunks_exact_mut(dst_channels)) |
| 258 |
{ |
| 259 |
let r = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.r_i()]._as_usize())); |
| 260 |
let g = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.g_i()]._as_usize())); |
| 261 |
let b = _mm_broadcast_ss(lut_lin.get_unchecked(src[src_cn.b_i()]._as_usize())); |
| 262 |
let a = if src_channels == 4 { |
| 263 |
src[src_cn.a_i()] |
| 264 |
} else { |
| 265 |
max_colors |
| 266 |
}; |
| 267 |
|
| 268 |
let mut v = if FMA { |
| 269 |
let v0 = _mm_mul_ps(r, _mm256_castps256_ps128(m0)); |
| 270 |
let v1 = _mm_fmadd_ps(g, _mm256_castps256_ps128(m1), v0); |
| 271 |
_mm_fmadd_ps(b, _mm256_castps256_ps128(m2), v1) |
| 272 |
} else { |
| 273 |
let v0 = _mm_mul_ps(r, _mm256_castps256_ps128(m0)); |
| 274 |
let v1 = _mm_mul_ps(g, _mm256_castps256_ps128(m1)); |
| 275 |
let v2 = _mm_mul_ps(b, _mm256_castps256_ps128(m2)); |
| 276 |
|
| 277 |
_mm_add_ps(_mm_add_ps(v0, v1), v2) |
| 278 |
}; |
| 279 |
|
| 280 |
v = _mm_max_ps(v, zeros); |
| 281 |
v = _mm_mul_ps(v, _mm256_castps256_ps128(v_scale)); |
| 282 |
v = _mm_min_ps(v, _mm256_castps256_ps128(v_scale)); |
| 283 |
|
| 284 |
let zx = _mm_cvtps_epi32(v); |
| 285 |
_mm_store_si128(temporary0.0.as_mut_ptr() as *mut _, zx); |
| 286 |
|
| 287 |
dst[dst_cn.r_i()] = self.profile.gamma[temporary0.0[0] as usize]; |
| 288 |
dst[dst_cn.g_i()] = self.profile.gamma[temporary0.0[2] as usize]; |
| 289 |
dst[dst_cn.b_i()] = self.profile.gamma[temporary0.0[4] as usize]; |
| 290 |
if dst_channels == 4 { |
| 291 |
dst[dst_cn.a_i()] = a; |
| 292 |
} |
| 293 |
} |
| 294 |
} |
| 295 |
|
| 296 |
Ok(()) |
| 297 |
} |
| 298 |
} |
| 299 |
|
| 300 |
impl< |
| 301 |
T: Clone + Copy + 'static + PointeeSizeExpressible + Default, |
| 302 |
const SRC_LAYOUT: u8, |
| 303 |
const DST_LAYOUT: u8, |
| 304 |
> TransformExecutor<T> for TransformShaperRgbOptAvx<T, SRC_LAYOUT, DST_LAYOUT> |
| 305 |
where |
| 306 |
u32: AsPrimitive<T>, |
| 307 |
{ |
| 308 |
fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> { |
| 309 |
unsafe { |
| 310 |
assert!(std::arch::is_x86_feature_detected!("fma")); |
| 311 |
self.transform_impl::<true>(src, dst) |
| 312 |
} |
| 313 |
} |
| 314 |
} |