From: Kostya Shishkov Date: Fri, 8 Nov 2019 18:18:59 +0000 (+0100) Subject: duck/avc: improve decoder a bit X-Git-Url: https://git.nihav.org/?a=commitdiff_plain;h=80aaa7ef4cf5ce03978f64d07c344e4b18718b1b;p=nihav.git duck/avc: improve decoder a bit --- diff --git a/nihav-duck/src/codecs/on2avc.rs b/nihav-duck/src/codecs/on2avc.rs index 35b9c76..46880fa 100644 --- a/nihav-duck/src/codecs/on2avc.rs +++ b/nihav-duck/src/codecs/on2avc.rs @@ -144,6 +144,7 @@ struct AVCDecoder { delay: [[f32; COEFFS]; 2], tmp: [f32; COEFFS * 2], ew_buf: [f32; 1152], + use_generic: bool, qtab: [f32; QTAB_SIZE], scale_tab: [f32; 128], @@ -194,6 +195,7 @@ impl AVCDecoder { delay: [[0.0; COEFFS]; 2], tmp: [0.0; COEFFS * 2], ew_buf: [0.0; 1152], + use_generic: true, qtab, scale_tab, @@ -335,11 +337,14 @@ impl AVCDecoder { match self.cur_win { 0 | 1 => { - self.imdct_long.imdct(coeffs, &mut self.tmp); - overlap(dst, &self.tmp, delay, self.win_long); - delay.copy_from_slice(&self.tmp[COEFFS..]); + self.imdct_long.imdct_half(coeffs, &mut self.tmp); + overlap_half(dst, &self.tmp, delay, self.win_long); + (&mut delay[0..COEFFS/2]).copy_from_slice(&self.tmp[COEFFS/2..COEFFS]); + for i in COEFFS/2..COEFFS { + delay[i] = delay[COEFFS - 1 - i]; + } }, - 2 => { + 2 | 7 => { self.imdct_long.imdct(coeffs, &mut self.tmp); for i in 0..SHORT_WIN_POINT0 { dst[i] = delay[i]; @@ -348,7 +353,7 @@ impl AVCDecoder { &self.tmp[SHORT_WIN_POINT0..SHORT_WIN_POINT1], &delay[SHORT_WIN_POINT0..SHORT_WIN_POINT1], AVC_WIN_SHORT); for i in SHORT_WIN_POINT1..COEFFS { - dst[i] = self.tmp[i]; + dst[i] = self.tmp[i - SHORT_WIN_POINT1 + 128]; } delay.copy_from_slice(&self.tmp[COEFFS..]); }, @@ -389,30 +394,34 @@ impl AVCDecoder { for i in 0..SHORT_WIN_POINT0 { dst[i] = delay[i]; } - synth1024(&mut self.dsp, coeffs, &mut self.ew_buf, &mut self.tmp, self.is_40khz); - let tmp = &mut self.ew_buf[0..COEFFS]; - tmp.reverse(); - overlap(&mut dst[SHORT_WIN_POINT0..SHORT_WIN_POINT1], - &self.ew_buf[SHORT_WIN_POINT0..SHORT_WIN_POINT1], - &delay[SHORT_WIN_POINT0..SHORT_WIN_POINT1], AVC_WIN_SHORT); + if !self.use_generic { + synth1024(&mut self.dsp, coeffs, &mut self.ew_buf, &mut self.tmp, self.is_40khz); + } else { + synth_generic(coeffs, &mut self.ew_buf, &mut self.tmp, self.is_40khz, 1024); + } + overlap_half(&mut dst[SHORT_WIN_POINT0..SHORT_WIN_POINT1], + &self.ew_buf[0..64], + &delay[SHORT_WIN_POINT0..SHORT_WIN_POINT1], AVC_WIN_SHORT); for i in SHORT_WIN_POINT1..COEFFS { - dst[i] = self.ew_buf[i]; + dst[i] = self.ew_buf[i - SHORT_WIN_POINT1 + 64]; } - *delay = [0.0; COEFFS]; - for i in 0..SHORT_WIN_POINT1 { - delay[i] = self.ew_buf[i]; + for i in 0..COEFFS/2 { + delay[i] = self.ew_buf[COEFFS/2 + i]; + } + for i in COEFFS/2..COEFFS { + delay[i] = delay[COEFFS - 1 - i]; } }, 5 => { for i in 0..SHORT_WIN_POINT0 { dst[i] = delay[i]; } - synth512(&mut self.dsp, coeffs, &mut self.ew_buf, &mut self.tmp, self.is_40khz); + if !self.use_generic { + synth512(&mut self.dsp, coeffs, &mut self.ew_buf, &mut self.tmp, self.is_40khz); + } else { + synth_generic(coeffs, &mut self.ew_buf, &mut self.tmp, self.is_40khz, 512); + } self.imdct_mid.imdct(&coeffs[512..], &mut self.ew_buf[512..]); - let tmp = &mut self.ew_buf[512..1024]; - tmp.reverse(); - let tmp = &mut self.ew_buf[0..1024]; - tmp.reverse(); overlap(&mut dst[SHORT_WIN_POINT0..SHORT_WIN_POINT1], &self.ew_buf[SHORT_WIN_POINT0..SHORT_WIN_POINT1], &delay[SHORT_WIN_POINT0..SHORT_WIN_POINT1], AVC_WIN_SHORT); @@ -429,11 +438,11 @@ impl AVCDecoder { dst[i] = delay[i]; } self.imdct_mid.imdct(coeffs, &mut self.ew_buf); - let tmp = &mut self.ew_buf[0..512]; - tmp.reverse(); - synth512(&mut self.dsp, &coeffs[512..], &mut self.ew_buf[512..], &mut self.tmp, self.is_40khz); - let tmp = &mut self.ew_buf[0..1024]; - tmp.reverse(); + if !self.use_generic { + synth512(&mut self.dsp, &coeffs[512..], &mut self.ew_buf[512..], &mut self.tmp, self.is_40khz); + } else { + synth_generic(&coeffs[512..], &mut self.ew_buf[512..], &mut self.tmp, self.is_40khz, 512); + } overlap(&mut dst[SHORT_WIN_POINT0..SHORT_WIN_POINT1], &self.ew_buf[SHORT_WIN_POINT0..SHORT_WIN_POINT1], &delay[SHORT_WIN_POINT0..SHORT_WIN_POINT1], AVC_WIN_SHORT); @@ -445,22 +454,6 @@ impl AVCDecoder { delay[i] = self.ew_buf[i]; } }, - 7 => { - for i in 0..SHORT_WIN_POINT0 { - dst[i] = delay[i]; - } - self.imdct_long.imdct(coeffs, &mut self.tmp); - overlap(&mut dst[SHORT_WIN_POINT0..SHORT_WIN_POINT1], - &self.tmp[SHORT_WIN_POINT0..SHORT_WIN_POINT1], - &delay[SHORT_WIN_POINT0..SHORT_WIN_POINT1], AVC_WIN_SHORT); - for i in SHORT_WIN_POINT1..COEFFS { - dst[i] = self.tmp[i]; - } - *delay = [0.0; COEFFS]; - for i in 0..SHORT_WIN_POINT1 { - delay[i] = self.ew_buf[i]; - } - }, _ => unreachable!(), }; } @@ -472,6 +465,19 @@ fn overlap(dst: &mut [f32], src: &[f32], delay: &[f32], win: &[f32]) { } } +fn overlap_half(dst: &mut [f32], src: &[f32], delay: &[f32], win: &[f32]) { + let n = win.len(); + for i in 0..n/2 { + let ii = n - 1 - i; + let c = src [n/2 - 1 - i]; + let d = delay[i]; + let w0 = win[i]; + let w1 = win[ii]; + dst[i] = d * w1 - c * w0; + dst[ii] = d * w0 + c * w1; + } +} + #[derive(Clone,Copy)] struct SynthParams { p0: usize, @@ -594,12 +600,92 @@ fn merge_bands(src: &[FFTComplex], dst: &mut [FFTComplex], size: usize) { } } +const SPARAMS10: [SynthParams; 2] = [ + SynthParams { p0: 1, p1: 3, idx: 0 }, + SynthParams { p0: 3, p1: 1, idx: 1 } ]; +const SPARAMS20: [SynthParams; 2] = [ + SynthParams { p0: 5, p1: 4, idx: 0 }, + SynthParams { p0: 4, p1: 5, idx: 1 } ]; +const SPARAMS40: [SynthParams; 2] = [ + SynthParams { p0: 11, p1: 8, idx: 0 }, + SynthParams { p0: 8, p1: 11, idx: 1 } ]; const SPARAMS84: [SynthParams; 4] = [ SynthParams { p0: 16, p1: 4, idx: 0 }, SynthParams { p0: 16, p1: 4, idx: 1 }, SynthParams { p0: 13, p1: 7, idx: 2 }, SynthParams { p0: 15, p1: 5, idx: 3 } ]; +const MERGE_ORDER_44K: &[u8] = &[ + 4, 4, 2, 2, 0, 0, 2, 0, 0, 2, 2, 0, 0, 2, 0, 0, 2, 0, 0, + 2, 0, 0, 4, 0, 0, 0, 0, 2, 0, 0, 0 +]; +const MERGE_ORDER_40K: &[u8] = &[ + 4, 4, 2, 2, 0, 0, 2, 0, 0, 2, 2, 0, 0, 2, 0, 0, 2, 0, 0, + 2, 0, 0, 4, 2, 0, 0, 2, 0, 0, 0, 0, 4, 0, 0, 0, 0, 0 +]; + +fn synth_filter(src: &[f32], dst: &mut [f32], len: usize, step: usize, bands: usize, idx: usize) +{ + let off = (len - step) / bands + 1; + let params = match step { + 10 => &SPARAMS10[idx], + 20 => &SPARAMS20[idx], + 40 => &SPARAMS40[idx], + _ => &SPARAMS84[idx], + }; + + if step == 10 { + synth_step0_10(src, dst, step, off, params); + synth_step1(src, dst, len, bands, step, off, params.p0, &AVC_TAB10[idx]); + } else if step == 20 { + synth_step0_20(src, dst, step, off, params); + synth_step1(src, dst, len, bands, step, off, params.p0, &AVC_TAB20[idx]); + } else if step == 40 { + synth_step0_40(src, dst, step, off, params); + synth_step1(src, dst, len, bands, step, off, params.p0, &AVC_TAB40[idx]); + } else { + synth_step0_84(src, dst, step, off, params); + synth_step1(src, dst, len, bands, step, off, params.p0, &AVC_TAB84[idx]); + } +} + +fn synth_recursive(dst: &mut [f32], tmp: &mut [f32], size: usize, order: &[u8], order_idx: &mut usize, dir: bool) { + let bands = order[*order_idx] as usize; + *order_idx += 1; + if bands == 0 { return; } + let sub_size = size / bands; + let mut sub_dir = false; + for (dst, tmp) in dst.chunks_mut(sub_size).take(bands).zip(tmp.chunks_mut(sub_size)) { + synth_recursive(dst, tmp, sub_size, order, order_idx, sub_dir); + sub_dir = !sub_dir; + } + for el in tmp.iter_mut().take(size) { *el = 0.0; } + let step = if bands == 2 { + if sub_size <= 20 { + 10 + } else if sub_size <= 40 { + 20 + } else { + 40 + } + } else { + 84 + }; + for (i, src) in dst.chunks_mut(sub_size).take(bands).enumerate() { + let idx = if !dir { i } else { bands - 1 - i }; + synth_filter(src, tmp, size, step, bands, idx); + } + (&mut dst[..size]).copy_from_slice(&tmp[..size]); +} + +fn synth_generic(src: &[f32], dst: &mut [f32], tmpbuf: &mut [f32; COEFFS * 2], is_40khz: bool, size: usize) { + let order = if is_40khz { MERGE_ORDER_40K } else { MERGE_ORDER_44K }; + (&mut dst[..size]).copy_from_slice(&src[..size]); + let mut order_idx = 0; + synth_recursive(dst, tmpbuf, size, order, &mut order_idx, false); + for i in 0..COEFFS { dst[i] *= 0.125; } +} + fn synth1024(dsp: &mut SynthDSP, src: &[f32], dst: &mut [f32], tmpbuf: &mut [f32; COEFFS * 2], is_40khz: bool) { for el in tmpbuf.iter_mut() { *el = 0.0; diff --git a/nihav-duck/src/codecs/on2avcdata.rs b/nihav-duck/src/codecs/on2avcdata.rs index c4fb1e0..b190acd 100644 --- a/nihav-duck/src/codecs/on2avcdata.rs +++ b/nihav-duck/src/codecs/on2avcdata.rs @@ -7140,6 +7140,15 @@ pub const AVC_TAB10_1: &[f64; 22] = &[ -0.80689017, 0.332667 ]; +pub const AVC_TAB10: [[f64; 10]; 2] = [ + [ + 0.0033357253, -0.012580752, -0.0062414902, 0.077571494, -0.03224487, + -0.24229489, 0.13842815, 0.72430853, 0.60382927, 0.1601024 + ], [ + -0.1601024, 0.60382927, -0.72430853, 0.13842815, 0.24229489, + -0.03224487, -0.077571494, -0.0062414902, 0.012580752, 0.0033357253, + ] +]; pub const AVC_TAB20: [[f64; 20]; 2] = [ [ -0.0014416961, -0.00056214974, 0.020948911, 0.025363335, -0.10427845,