From 18ce00da888883413f937f96371ddef11c2ff075 Mon Sep 17 00:00:00 2001 From: vangel76 Date: Sun, 25 Oct 2026 12:02:23 +0100 Subject: [PATCH 1/2] av1enc: specialise the inverse transform per size, size the forward tmp buffer The inverse transform computed its bit-reversal permutations and loop bounds at run time and went through one generic path for every size. inverse_dct, the ADST permutations and inverse_2d are now const generic over the size and inlined, so the permutations fold at compile time. forward_2d zeroed a 4 KiB tmp buffer on every call, even for 4x4. It now dispatches to a fixed-size version for 4, 8 and 16 (tmp is exactly n*n floats); 32 keeps the run-time version, which measured faster than a const 32 variant. The float operation order is unchanged. Output is bit-identical. Co-Authored-By: Claude Sonnet 5.5 --- crates/av1enc/src/transform.rs | 95 ++++++++++++++++++++++++++++++---- 1 file changed, 86 insertions(+), 9 deletions(-) diff --git a/crates/av1enc/src/transform.rs b/crates/av1enc/src/transform.rs index 90042e26a..a7eb99d34 100644 --- a/crates/av1enc/src/transform.rs +++ b/crates/av1enc/src/transform.rs @@ -32,6 +32,7 @@ fn sin128(angle: i32) -> i64 { cos128(angle - 64) } +#[inline(always)] fn brev(num_bits: u32, x: usize) -> usize { let mut t = 0; for i in 0..num_bits { @@ -63,8 +64,11 @@ fn hd(t: &mut [i32], a: usize, b: usize, flip: bool, r: u32) { t[b] = (x - y).clamp(lo, hi); } -/// Inverse DCT of length 2^n, n = 2..=5 (7.13.2.3), in place. -fn inverse_dct(t: &mut [i32], n: u32, r: u32) { +/// Inverse DCT of length 2^N, N = 2..=5 (7.13.2.3), in place. `N` is a const generic so that the +/// bit-reversal permutations and loop bounds fold at compile time. +#[inline(always)] +fn inverse_dct_n(t: &mut [i32], r: u32) { + let n = N; let len = 1usize << n; let mut copy = [0i32; 32]; copy[..len].copy_from_slice(&t[..len]); @@ -180,6 +184,7 @@ const SINPI_2_9: i64 = 2482; const SINPI_3_9: i64 = 3344; const SINPI_4_9: i64 = 3803; +#[inline(always)] fn inverse_adst4(t: &mut [i32]) { let (t0, t1, t2, t3) = (t[0] as i64, t[1] as i64, t[2] as i64, t[3] as i64); let mut s = [SINPI_1_9 * t0, SINPI_2_9 * t0, SINPI_3_9 * t1, SINPI_4_9 * t2, SINPI_1_9 * t2, SINPI_2_9 * t3, SINPI_4_9 * t3]; @@ -197,7 +202,9 @@ fn inverse_adst4(t: &mut [i32]) { } } -fn adst_in_permute(t: &mut [i32], n: u32) { +#[inline(always)] +fn adst_in_permute(t: &mut [i32]) { + let n = N; let n0 = 1usize << n; let mut copy = [0i32; 16]; copy[..n0].copy_from_slice(&t[..n0]); @@ -207,7 +214,9 @@ fn adst_in_permute(t: &mut [i32], n: u32) { } } -fn adst_out_permute(t: &mut [i32], n: u32) { +#[inline(always)] +fn adst_out_permute(t: &mut [i32]) { + let n = N; let n0 = 1usize << n; let mut copy = [0i32; 16]; copy[..n0].copy_from_slice(&t[..n0]); @@ -221,8 +230,9 @@ fn adst_out_permute(t: &mut [i32], n: u32) { } } +#[inline(always)] fn inverse_adst8(t: &mut [i32], r: u32) { - adst_in_permute(t, 3); + adst_in_permute::<3>(t); for i in 0..4 { bf(t, 2 * i, 2 * i + 1, 60 - 16 * i as i32, true); } @@ -240,11 +250,12 @@ fn inverse_adst8(t: &mut [i32], r: u32) { for i in 0..2 { bf(t, 2 + 4 * i, 3 + 4 * i, 32, true); } - adst_out_permute(t, 3); + adst_out_permute::<3>(t); } +#[inline(always)] fn inverse_adst16(t: &mut [i32], r: u32) { - adst_in_permute(t, 4); + adst_in_permute::<4>(t); for i in 0..8 { bf(t, 2 * i, 2 * i + 1, 62 - 8 * i as i32, true); } @@ -273,7 +284,7 @@ fn inverse_adst16(t: &mut [i32], r: u32) { for i in 0..4 { bf(t, 2 + 4 * i, 3 + 4 * i, 32, true); } - adst_out_permute(t, 4); + adst_out_permute::<4>(t); } /// 1-D transform kind. @@ -298,6 +309,17 @@ pub(crate) fn kinds(tx_type: usize) -> (Kind, Kind) { } } +#[inline(always)] +fn inverse_dct(t: &mut [i32], n: u32, r: u32) { + match n { + 2 => inverse_dct_n::<2>(t, r), + 3 => inverse_dct_n::<3>(t, r), + 4 => inverse_dct_n::<4>(t, r), + _ => inverse_dct_n::<5>(t, r), + } +} + +#[inline(always)] fn inverse_1d(t: &mut [i32], kind: Kind, n: u32, r: u32) { match kind { Kind::Dct => inverse_dct(t, n, r), @@ -312,7 +334,17 @@ fn inverse_1d(t: &mut [i32], kind: Kind, n: u32, r: u32) { /// 2-D inverse transform (7.13.3) of a square `n x n` block (n = 4..=32) of dequantised /// coefficients (row stride `n`) into `residual` (row stride `n`). pub(crate) fn inverse_2d(dequant: &[i32], residual: &mut [i32], tx_sz: usize, tx_type: usize, bit_depth: u32) { - let log2 = TX_WIDTH_LOG2[tx_sz] as u32; + match TX_WIDTH_LOG2[tx_sz] { + 2 => inverse_2d_n::<2>(dequant, residual, tx_sz, tx_type, bit_depth), + 3 => inverse_2d_n::<3>(dequant, residual, tx_sz, tx_type, bit_depth), + 4 => inverse_2d_n::<4>(dequant, residual, tx_sz, tx_type, bit_depth), + _ => inverse_2d_n::<5>(dequant, residual, tx_sz, tx_type, bit_depth), + } +} + +/// `inverse_2d` for a block of width `1 << LOG2`. +fn inverse_2d_n(dequant: &[i32], residual: &mut [i32], tx_sz: usize, tx_type: usize, bit_depth: u32) { + let log2 = LOG2; let n = 1usize << log2; let row_shift = TRANSFORM_ROW_SHIFT[tx_sz] as u32; let row_clamp = bit_depth + 8; @@ -419,6 +451,51 @@ fn basis(kind: Kind, n: usize) -> &'static [f32] { /// Forward 2-D transform of the `n x n` residual (row stride `n`) into orthonormal-domain /// coefficients scaled by 8 (so that `level = out / q` with the spec's dequantiser). pub(crate) fn forward_2d(residual: &[i32], out: &mut [f32], n: usize, tx_type: usize) { + match n { + 4 => forward_2d_n::<4, 16>(residual, out, tx_type), + 8 => forward_2d_n::<8, 64>(residual, out, tx_type), + 16 => forward_2d_n::<16, 256>(residual, out, tx_type), + _ => forward_2d_dyn(residual, out, n, tx_type), + } +} + +/// `forward_2d` for a fixed size, so `tmp` is exactly `NN = N * N` floats (no 4 KiB zeroing). +#[inline(always)] +fn forward_2d_n(residual: &[i32], out: &mut [f32], tx_type: usize) { + let n = N; + let (ck, rk) = kinds(tx_type); + let bc = &basis(ck, n)[..NN]; + let brt = &basis_t(rk, n)[..NN]; + let residual = &residual[..NN]; + let out = &mut out[..NN]; + let mut tmp = [0f32; NN]; + for i in 0..n { + let t = &mut tmp[i * n..i * n + n]; + for j in 0..n { + let r = residual[i * n + j]; + if r == 0 { + continue; + } + let r = r as f32; + for (o, &b) in t.iter_mut().zip(&brt[j * n..j * n + n]) { + *o += r * b; + } + } + } + out.iter_mut().for_each(|v| *v = 0.0); + for k in 0..n { + let o = &mut out[k * n..k * n + n]; + for i in 0..n { + let b = 8.0 * bc[k * n + i]; + for (v, &t) in o.iter_mut().zip(&tmp[i * n..i * n + n]) { + *v += b * t; + } + } + } +} + +/// 32x32 (and any other size): a const-generic 32 version measured slower. +fn forward_2d_dyn(residual: &[i32], out: &mut [f32], n: usize, tx_type: usize) { let (ck, rk) = kinds(tx_type); let bc = basis(ck, n); let brt = basis_t(rk, n); From c99b20cac3fc15a85f5782a56eb2307c56702e6d Mon Sep 17 00:00:00 2001 From: bflatastic Date: Thu, 29 Oct 2026 23:15:48 +0700 Subject: [PATCH 2/2] av1enc: inline the DCT stage helpers so const-generic sizes fold through them The strict-clippy split of inverse_dct (#2496) passes n at run time; inlining keeps inverse_dct_n's compile-time sizes. Co-Authored-By: Claude Opus 5.5 --- crates/av1enc/src/transform.rs | 2 ++ 1 file changed, 2 insertions(+) diff --git a/crates/av1enc/src/transform.rs b/crates/av1enc/src/transform.rs index fce766eba..9d81ead0b 100644 --- a/crates/av1enc/src/transform.rs +++ b/crates/av1enc/src/transform.rs @@ -80,6 +80,7 @@ fn inverse_dct_n(t: &mut [i32], r: u32) { } /// Rotate the high-frequency branches before the four-point core. +#[inline(always)] fn inverse_dct_initial_stages(t: &mut [i32], n: u32, r: u32) { if n >= 5 { for i in 0..8 { @@ -116,6 +117,7 @@ fn inverse_dct_initial_stages(t: &mut [i32], n: u32, r: u32) { } /// Evaluate the four-point core and merge progressively larger branches. +#[inline(always)] fn inverse_dct_final_stages(t: &mut [i32], n: u32, r: u32) { for i in 0..2 { bf(t, 2 * i, 2 * i + 1, 32 + 16 * i as i32, i == 0);