zune-jpeg 0.5.15

/*
 * Copyright (c) 2023.
 *
 * This software is free software;
 *
 * You can redistribute it or modify it under terms of the MIT, Apache License or Zlib license
 */

//! AVX color conversion routines
//!
//! Okay these codes are cool
//!
//! Herein lies super optimized codes to do color conversions.
//!
//!
//! 1. The YCbCr to RGB use integer approximations and not the floating point equivalent.
//! That means we may be +- 2 of pixels generated by libjpeg-turbo jpeg decoding
//! (also libjpeg uses routines like `Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G`)
//!
//! Firstly, we use integers (fun fact:there is no part of this code base where were dealing with
//! floating points.., fun fact: the first fun fact wasn't even fun.)
//!
//! Secondly ,we have cool clamping code, especially for rgba , where we don't need clamping and we
//! spend our time cursing that Intel decided permute instructions to work like 2 128 bit vectors(the compiler opitmizes
//! it out to something cool).
//!
//! There isn't a lot here (not as fun as bitstream ) but I hope you find what you're looking for.
//!
//! O and ~~subscribe to my youtube channel~~

#![cfg(any(target_arch = "x86", target_arch = "x86_64"))]
#![cfg(feature = "x86")]
#![allow(
    clippy::wildcard_imports,
    clippy::cast_possible_truncation,
    clippy::too_many_arguments,
    clippy::inline_always,
    clippy::doc_markdown,
    dead_code
)]

#[cfg(target_arch = "x86")]
use core::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use core::arch::x86_64::*;

use crate::color_convert::scalar::{CB_CF, CR_CF, C_G_CB_COEF_2, C_G_CR_COEF_1, YUV_RND, Y_CF};

pub union YmmRegister {
    // both are 32 when using std::mem::size_of
    mm256: __m256i,
    // for avx color conversion
    array: [i16; 16]
}

const R_AVX_COEF: i32 = i32::from_ne_bytes([CR_CF.to_ne_bytes()[0], CR_CF.to_ne_bytes()[1], 0, 0]);
const B_AVX_COEF: i32 = i32::from_ne_bytes([0, 0, CB_CF.to_ne_bytes()[0], CB_CF.to_ne_bytes()[1]]);
const G_COEF_AVX_COEF: i32 = i32::from_ne_bytes([
    C_G_CR_COEF_1.to_ne_bytes()[0],
    C_G_CR_COEF_1.to_ne_bytes()[1],
    C_G_CB_COEF_2.to_ne_bytes()[0],
    C_G_CB_COEF_2.to_ne_bytes()[1]
]);

//--------------------------------------------------------------------------------------------------
// AVX conversion routines
//--------------------------------------------------------------------------------------------------

///
/// Convert YCBCR to RGB using AVX instructions
///
///  # Note
///**IT IS THE RESPONSIBILITY OF THE CALLER TO CALL THIS IN CPUS SUPPORTING
/// AVX2 OTHERWISE THIS IS UB**
///
/// *Peace*
///
/// This library itself will ensure that it's never called in CPU's not
/// supporting AVX2
///
/// # Arguments
/// - `y`,`cb`,`cr`: A reference of 8 i32's
/// - `out`: The output  array where we store our converted items
/// - `offset`: The position from 0 where we write these RGB values
#[inline(always)]
pub fn ycbcr_to_rgb_avx2(
    y: &[i16; 16], cb: &[i16; 16], cr: &[i16; 16], out: &mut [u8], offset: &mut usize
) {
    // call this in another function to tell RUST to vectorize this
    // storing
    unsafe {
        ycbcr_to_rgb_avx2_1(y, cb, cr, out, offset);
    }
}

#[inline]
#[target_feature(enable = "avx2")]
unsafe fn ycbcr_to_rgb_avx2_1(
    y: &[i16; 16], cb: &[i16; 16], cr: &[i16; 16], out: &mut [u8], offset: &mut usize
) {
    let (mut r, mut g, mut b) = ycbcr_to_rgb_baseline_no_clamp(y, cb, cr);

    r = _mm256_packus_epi16(r, _mm256_setzero_si256());
    g = _mm256_packus_epi16(g, _mm256_setzero_si256());
    b = _mm256_packus_epi16(b, _mm256_setzero_si256());

    r = _mm256_permute4x64_epi64::<{ shuffle(3, 1, 2, 0) }>(r);
    g = _mm256_permute4x64_epi64::<{ shuffle(3, 1, 2, 0) }>(g);
    b = _mm256_permute4x64_epi64::<{ shuffle(3, 1, 2, 0) }>(b);

    let sh_r = _mm256_setr_epi8(
        0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10, 5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14,
        9, 4, 15, 10, 5
    );
    let sh_g = _mm256_setr_epi8(
        5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10, 5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3,
        14, 9, 4, 15, 10
    );
    let sh_b = _mm256_setr_epi8(
        10, 5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10, 5, 0, 11, 6, 1, 12, 7, 2, 13, 8,
        3, 14, 9, 4, 15
    );

    let r0 = _mm256_shuffle_epi8(r, sh_r);
    let g0 = _mm256_shuffle_epi8(g, sh_g);
    let b0 = _mm256_shuffle_epi8(b, sh_b);

    let m0 = _mm256_setr_epi8(
        0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1,
        0, 0, -1, 0, 0
    );
    let m1 = _mm256_setr_epi8(
        0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0,
        -1, 0, 0, -1, 0
    );

    let p0 = _mm256_blendv_epi8(_mm256_blendv_epi8(r0, g0, m0), b0, m1);
    let p1 = _mm256_blendv_epi8(_mm256_blendv_epi8(g0, b0, m0), r0, m1);
    let p2 = _mm256_blendv_epi8(_mm256_blendv_epi8(b0, r0, m0), g0, m1);

    let rgb0 = _mm256_permute2x128_si256::<32>(p0, p1);
    let rgb1 = _mm256_permute2x128_si256::<48>(p2, p0);

    _mm256_storeu_si256(out.as_mut_ptr().cast(), rgb0);
    _mm_storeu_si128(out[32..].as_mut_ptr().cast(), _mm256_castsi256_si128(rgb1));

    *offset += 48;
}

// Enabled avx2 automatically enables avx.
#[inline]
#[target_feature(enable = "avx2")]
/// A baseline implementation of YCbCr to RGB conversion which does not carry
/// out clamping
///
/// This is used by the `ycbcr_to_rgba_avx` and `ycbcr_to_rgbx` conversion
/// routines
unsafe fn ycbcr_to_rgb_baseline_no_clamp(
    y: &[i16; 16], cb: &[i16; 16], cr: &[i16; 16]
) -> (__m256i, __m256i, __m256i) {
    // Load values into a register
    //
    let y_c = _mm256_loadu_si256(y.as_ptr().cast());
    let cb_c = _mm256_loadu_si256(cb.as_ptr().cast());
    let cr_c = _mm256_loadu_si256(cr.as_ptr().cast());

    // Here we want to use _mm256_madd_epi16 to perform 2 multiplications
    // and one addition per instruction.

    // At first, we have to pack i16 U and V that stores u8 into one u8 [U,V]
    // then zero extend, and keep in mind that lanes is already been permuted.

    let y_coeff = _mm256_set1_epi32(i32::from(Y_CF));
    let cr_coeff = _mm256_set1_epi32(R_AVX_COEF);
    let cb_coeff = _mm256_set1_epi32(B_AVX_COEF);
    let cg_coeff = _mm256_set1_epi32(G_COEF_AVX_COEF);
    let v_rnd = _mm256_set1_epi32(i32::from(YUV_RND));
    let uv_bias = _mm256_set1_epi16(128);

    // UV in memory because x86/x86_64 is always little endian
    let v_0 = _mm256_slli_epi16::<8>(cb_c);
    let u_v_8 = _mm256_or_si256(v_0, cr_c);

    let mut u_v_lo = _mm256_unpacklo_epi8(u_v_8, _mm256_setzero_si256());
    let mut u_v_hi = _mm256_unpackhi_epi8(u_v_8, _mm256_setzero_si256());

    let mut y_lo = _mm256_unpacklo_epi16(y_c, _mm256_setzero_si256());
    let mut y_hi = _mm256_unpackhi_epi16(y_c, _mm256_setzero_si256());

    u_v_lo = _mm256_sub_epi16(u_v_lo, uv_bias);
    u_v_hi = _mm256_sub_epi16(u_v_hi, uv_bias);

    y_lo = _mm256_madd_epi16(y_lo, y_coeff);
    y_hi = _mm256_madd_epi16(y_hi, y_coeff);

    let mut r_lo = _mm256_madd_epi16(u_v_lo, cr_coeff);
    let mut r_hi = _mm256_madd_epi16(u_v_hi, cr_coeff);

    let mut g_lo = _mm256_madd_epi16(u_v_lo, cg_coeff);
    let mut g_hi = _mm256_madd_epi16(u_v_hi, cg_coeff);

    // This ordering is preferred to reduce register file pressure.

    y_lo = _mm256_add_epi32(y_lo, v_rnd);
    y_hi = _mm256_add_epi32(y_hi, v_rnd);

    let mut b_lo = _mm256_madd_epi16(u_v_lo, cb_coeff);
    let mut b_hi = _mm256_madd_epi16(u_v_hi, cb_coeff);

    r_lo = _mm256_add_epi32(r_lo, y_lo);
    r_hi = _mm256_add_epi32(r_hi, y_hi);

    g_lo = _mm256_add_epi32(g_lo, y_lo);
    g_hi = _mm256_add_epi32(g_hi, y_hi);

    b_lo = _mm256_add_epi32(b_lo, y_lo);
    b_hi = _mm256_add_epi32(b_hi, y_hi);

    r_lo = _mm256_srai_epi32::<14>(r_lo);
    r_hi = _mm256_srai_epi32::<14>(r_hi);

    g_lo = _mm256_srai_epi32::<14>(g_lo);
    g_hi = _mm256_srai_epi32::<14>(g_hi);

    b_lo = _mm256_srai_epi32::<14>(b_lo);
    b_hi = _mm256_srai_epi32::<14>(b_hi);

    let r = _mm256_packus_epi32(r_lo, r_hi);
    let g = _mm256_packus_epi32(g_lo, g_hi);
    let b = _mm256_packus_epi32(b_lo, b_hi);

    return (r, g, b);
}

#[inline(always)]
pub fn ycbcr_to_rgba_avx2(
    y: &[i16; 16], cb: &[i16; 16], cr: &[i16; 16], out: &mut [u8], offset: &mut usize
) {
    unsafe {
        ycbcr_to_rgba_unsafe(y, cb, cr, out, offset);
    }
}

#[inline]
#[target_feature(enable = "avx2")]
#[rustfmt::skip]
unsafe fn ycbcr_to_rgba_unsafe(
    y: &[i16; 16], cb: &[i16; 16], cr: &[i16; 16],
    out: &mut [u8],
    offset: &mut usize,
)
{
    // check if we have enough space to write.
    let tmp:& mut [u8; 64] = out.get_mut(*offset..*offset + 64).expect("Slice to small cannot write").try_into().unwrap();

    let (r, g, b) = ycbcr_to_rgb_baseline_no_clamp(y, cb, cr);

    // set alpha channel to 255 for opaque

    // And no these comments were not from me pressing the keyboard

    // Pack the integers into u8's using unsigned saturation.
    let c = _mm256_packus_epi16(r, g); //aaaaa_bbbbb_aaaaa_bbbbbb
    let d = _mm256_packus_epi16(b, _mm256_set1_epi16(255)); // cccccc_dddddd_ccccccc_ddddd
    // transpose_u16 and interleave channels
    let e = _mm256_unpacklo_epi8(c, d); //ab_ab_ab_ab_ab_ab_ab_ab
    let f = _mm256_unpackhi_epi8(c, d); //cd_cd_cd_cd_cd_cd_cd_cd
    // final transpose_u16
    let g = _mm256_unpacklo_epi8(e, f); //abcd_abcd_abcd_abcd_abcd
    let h = _mm256_unpackhi_epi8(e, f);
    
    // undo packus shuffling...
    let i = _mm256_permute2x128_si256::<{ shuffle(3, 2, 1, 0) }>(g, h);
    
    let j = _mm256_permute2x128_si256::<{ shuffle(1, 2, 3, 0) }>(g, h);
    
    let k = _mm256_permute2x128_si256::<{ shuffle(3, 2, 0, 1) }>(g, h);
    
    let l = _mm256_permute2x128_si256::<{ shuffle(0, 3, 2, 1) }>(g, h);
    
    let m = _mm256_blend_epi32::<0b1111_0000>(i, j);
    
    let n = _mm256_blend_epi32::<0b1111_0000>(k, l);
    
    // Store
    // Use streaming instructions to prevent polluting the cache?
    _mm256_storeu_si256(tmp.as_mut_ptr().cast(), m);
    
    _mm256_storeu_si256(tmp[32..].as_mut_ptr().cast(), n);

    *offset += 64;
}

#[inline]
const fn shuffle(z: i32, y: i32, x: i32, w: i32) -> i32 {
    (z << 6) | (y << 4) | (x << 2) | w
}