/* * 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 */ #![cfg(target_arch = "aarch64")] //! AVX optimised IDCT. //! //! Okay not thaat optimised. //! //! //! # The implementation //! The implementation is neatly broken down into two operations. //! //! 1. Test for zeroes //! > There is a shortcut method for idct where when all AC values are zero, we can get the answer really quickly. //! by scaling the 1/8th of the DCT coefficient of the block to the whole block and level shifting. //! //! 2. If above fails, we proceed to carry out IDCT as a two pass one dimensional algorithm. //! IT does two whole scans where it carries out IDCT on all items //! After each successive scan, data is transposed in register(thank you x86 SIMD powers). and the second //! pass is carried out. //! //! The code is not super optimized, it produces bit identical results with scalar code hence it's //! `mm256_add_epi16` //! and it also has the advantage of making this implementation easy to maintain. #![cfg(feature = "neon")] use core::arch::aarch64::*; use crate::unsafe_utils::{transpose, YmmRegister}; const SCALE_BITS: i32 = 512 + 65536 + (128 << 17); #[inline] #[target_feature(enable = "neon")] unsafe fn pack_16(a: int32x4x2_t) -> int16x8_t { vcombine_s16(vqmovn_s32(a.0), vqmovn_s32(a.1)) } #[inline] #[target_feature(enable = "neon")] unsafe fn condense_bottom_16(a: int32x4x2_t, b: int32x4x2_t) -> int16x8x2_t { int16x8x2_t(pack_16(a), pack_16(b)) } #[target_feature(enable = "neon")] #[allow( clippy::too_many_lines, clippy::cast_possible_truncation, clippy::similar_names, clippy::op_ref, unused_assignments, clippy::zero_prefixed_literal )] pub unsafe fn idct_neon( in_vector: &mut [i32; 64], out_vector: &mut [i16], stride: usize ) { let mut pos = 0; // load into registers // // We sign extend i16's to i32's and calculate them with extended precision and // later reduce them to i16's when we are done carrying out IDCT let mut row0 = YmmRegister::load(in_vector[00..].as_ptr().cast()); let mut row1 = YmmRegister::load(in_vector[08..].as_ptr().cast()); let mut row2 = YmmRegister::load(in_vector[16..].as_ptr().cast()); let mut row3 = YmmRegister::load(in_vector[24..].as_ptr().cast()); let mut row4 = YmmRegister::load(in_vector[32..].as_ptr().cast()); let mut row5 = YmmRegister::load(in_vector[40..].as_ptr().cast()); let mut row6 = YmmRegister::load(in_vector[48..].as_ptr().cast()); let mut row7 = YmmRegister::load(in_vector[56..].as_ptr().cast()); // Forward DCT and quantization may cause all the AC terms to be zero, for such // cases we can try to accelerate it // Basically the poop is that whenever the array has 63 zeroes, its idct is // (arr[0]>>3)or (arr[0]/8) propagated to all the elements. // We first test to see if the array contains zero elements and if it does, we go the // short way. // // This reduces IDCT overhead from about 39% to 18 %, almost half // Do another load for the first row, we don't want to check DC value, because // we only care about AC terms // TODO this should be a shift/shuffle, not a likely unaligned load let row8 = YmmRegister::load(in_vector[1..].as_ptr().cast()); let or_tree = (((row1 | row8) | (row2 | row3)) | ((row4 | row5) | (row6 | row7))); if or_tree.all_zero() { // AC terms all zero, idct of the block is ( coeff[0] * qt[0] )/8 + 128 (bias) // (and clamped to 255) // Round by adding 0.5 * (1 << 3) and offset by adding (128 << 3) before scaling let coeff = ((in_vector[0] + 4 + 1024) >> 3).clamp(0, 255) as i16; let idct_value = vdupq_n_s16(coeff); macro_rules! store { ($pos:tt,$value:tt) => { // store vst1q_s16( out_vector .get_mut($pos..$pos + 8) .unwrap() .as_mut_ptr() .cast(), $value ); $pos += stride; }; } store!(pos, idct_value); store!(pos, idct_value); store!(pos, idct_value); store!(pos, idct_value); store!(pos, idct_value); store!(pos, idct_value); store!(pos, idct_value); store!(pos, idct_value); return; } macro_rules! dct_pass { ($SCALE_BITS:tt,$scale:tt) => { // There are a lot of ways to do this // but to keep it simple(and beautiful), ill make a direct translation of the // scalar code to also make this code fully transparent(this version and the non // avx one should produce identical code.) // Compiler does a pretty good job of optimizing add + mul pairs // into multiply-acumulate pairs // even part let p1 = (row2 + row6) * 2217; let mut t2 = p1 + row6 * -7567; let mut t3 = p1 + row2 * 3135; let mut t0 = (row0 + row4).const_shl::<12>(); let mut t1 = (row0 - row4).const_shl::<12>(); let x0 = t0 + t3 + $SCALE_BITS; let x3 = t0 - t3 + $SCALE_BITS; let x1 = t1 + t2 + $SCALE_BITS; let x2 = t1 - t2 + $SCALE_BITS; let p3 = row7 + row3; let p4 = row5 + row1; let p1 = row7 + row1; let p2 = row5 + row3; let p5 = (p3 + p4) * 4816; t0 = row7 * 1223; t1 = row5 * 8410; t2 = row3 * 12586; t3 = row1 * 6149; let p1 = p5 + p1 * -3685; let p2 = p5 + (p2 * -10497); let p3 = p3 * -8034; let p4 = p4 * -1597; t3 += p1 + p4; t2 += p2 + p3; t1 += p2 + p4; t0 += p1 + p3; row0 = (x0 + t3).const_shra::<$scale>(); row1 = (x1 + t2).const_shra::<$scale>(); row2 = (x2 + t1).const_shra::<$scale>(); row3 = (x3 + t0).const_shra::<$scale>(); row4 = (x3 - t0).const_shra::<$scale>(); row5 = (x2 - t1).const_shra::<$scale>(); row6 = (x1 - t2).const_shra::<$scale>(); row7 = (x0 - t3).const_shra::<$scale>(); }; } // Process rows dct_pass!(512, 10); transpose( &mut row0, &mut row1, &mut row2, &mut row3, &mut row4, &mut row5, &mut row6, &mut row7 ); // process columns dct_pass!(SCALE_BITS, 17); transpose( &mut row0, &mut row1, &mut row2, &mut row3, &mut row4, &mut row5, &mut row6, &mut row7 ); // Pack i32 to i16's, // clamp them to be between 0-255 // Undo shuffling // Store back to array // This could potentially be reorganized to take advantage of the multi-register stores macro_rules! permute_store { ($x:tt,$y:tt,$index:tt,$out:tt) => { let a = condense_bottom_16($x, $y); // Clamp the values after packing, we can clamp more values at once let b = clamp256_neon(a); // store first vector vst1q_s16( ($out) .get_mut($index..$index + 8) .unwrap() .as_mut_ptr() .cast(), b.0 ); $index += stride; // second vector vst1q_s16( ($out) .get_mut($index..$index + 8) .unwrap() .as_mut_ptr() .cast(), b.1 ); $index += stride; }; } // Pack and write the values back to the array permute_store!((row0.mm256), (row1.mm256), pos, out_vector); permute_store!((row2.mm256), (row3.mm256), pos, out_vector); permute_store!((row4.mm256), (row5.mm256), pos, out_vector); permute_store!((row6.mm256), (row7.mm256), pos, out_vector); } #[inline] #[target_feature(enable = "neon")] unsafe fn clamp_neon(reg: int16x8_t) -> int16x8_t { let min_s = vdupq_n_s16(0); let max_s = vdupq_n_s16(255); let max_v = vmaxq_s16(reg, min_s); //max(a,0) let min_v = vminq_s16(max_v, max_s); //min(max(a,0),255) min_v } #[inline] #[target_feature(enable = "neon")] unsafe fn clamp256_neon(reg: int16x8x2_t) -> int16x8x2_t { int16x8x2_t(clamp_neon(reg.0), clamp_neon(reg.1)) } #[cfg(test)] mod test { use super::*; #[test] fn test_neon_clamp_256() { unsafe { let vals: [i16; 16] = [-1, -2, -3, 4, 256, 257, 258, 240, -1, 290, 2, 3, 4, 5, 6, 7]; let loaded = vld1q_s16_x2(vals.as_ptr().cast()); let shuffled = clamp256_neon(loaded); let mut result: [i16; 16] = [0; 16]; vst1q_s16_x2(result.as_mut_ptr().cast(), shuffled); assert_eq!( result, [0, 0, 0, 4, 255, 255, 255, 240, 0, 255, 2, 3, 4, 5, 6, 7] ) } } }