]>
Commit | Line | Data |
---|---|---|
2f9923e6 KS |
1 | use nihav_core::frame::*; |
2 | use nihav_codec_support::codecs::MV; | |
3 | use nihav_codec_support::codecs::blockdsp::*; | |
4 | ||
5 | #[cfg(not(debug_assertions))] | |
6 | mod release; | |
7 | #[cfg(not(debug_assertions))] | |
8 | use release::*; | |
9 | #[cfg(debug_assertions)] | |
10 | mod debug; | |
11 | #[cfg(debug_assertions)] | |
12 | use debug::*; | |
13 | ||
76431444 KS |
14 | type MCFunc = fn (dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, h: usize); |
15 | ||
2f9923e6 KS |
16 | fn clip_u8(val: i16) -> u8 { val.max(0).min(255) as u8 } |
17 | ||
3c506c7a KS |
18 | trait RegisterSIMD { |
19 | fn register_simd(&mut self); | |
20 | } | |
21 | ||
2f9923e6 KS |
22 | pub struct H264MC { |
23 | avg_buf: NAVideoBufferRef<u8>, | |
76431444 KS |
24 | pub put_block_weighted: [fn (dst: &mut [u8], stride: usize, src: &[u8], h: usize, wparams: [i8; 3]); 4], |
25 | pub put_block_weighted2: [fn (dst: &mut [u8], stride: usize, src0: &[u8], src1: &[u8], h: usize, wparams: [i8; 5]); 4], | |
2f9923e6 KS |
26 | } |
27 | ||
28 | impl H264MC { | |
29 | pub fn new(avg_buf: NAVideoBufferRef<u8>) -> Self { | |
3c506c7a | 30 | let mut obj = Self { |
76431444 KS |
31 | avg_buf, |
32 | put_block_weighted: [put_blk_w_2, put_blk_w_4, put_blk_w_8, put_blk_w_16], | |
33 | put_block_weighted2: [put_blk_w2_2, put_blk_w2_4, put_blk_w2_8, put_blk_w2_16], | |
3c506c7a KS |
34 | }; |
35 | obj.register_simd(); | |
36 | obj | |
2f9923e6 KS |
37 | } |
38 | pub fn do_mc(&mut self, frm: &mut NASimpleVideoFrame<u8>, refpic: NAVideoBufferRef<u8>, xpos: usize, ypos: usize, w: usize, h: usize, mv: MV) { | |
39 | let mut ebuf = [0u8; 22 * 22]; | |
40 | let mvx = mv.x >> 2; | |
41 | let mvy = mv.y >> 2; | |
42 | let mode = ((mv.x & 3) + (mv.y & 3) * 4) as usize; | |
43 | let pre = if mode != 0 { 2isize } else { 0 }; | |
44 | let post = if mode != 0 { 3isize } else { 0 }; | |
45 | let (yw, yh) = refpic.get_dimensions(0); | |
46 | let src = refpic.get_data(); | |
47 | let systride = refpic.get_stride(0); | |
48 | let src_x = (xpos as isize) + (mvx as isize); | |
49 | let src_y = (ypos as isize) + (mvy as isize); | |
50 | let (ysrc, ystride) = if (src_x - pre < 0) || (src_x + (w as isize) + post > (yw as isize)) || (src_y - pre < 0) || (src_y + (h as isize) + post > (yh as isize)) { | |
51 | let add = (pre + post) as usize; | |
52 | edge_emu(&refpic, src_x - pre, src_y - pre, w + add, h + add, &mut ebuf, 22, 0, 0); | |
53 | (ebuf.as_slice(), 22) | |
54 | } else { | |
55 | (&src[refpic.get_offset(0) + ((src_x - pre) as usize) + ((src_y - pre) as usize) * systride..], systride) | |
56 | }; | |
76431444 KS |
57 | let wmode = match w { |
58 | 4 => 0, | |
59 | 8 => 1, | |
60 | _ => 2, | |
61 | }; | |
62 | (H264_LUMA_INTERP[wmode][mode])(&mut frm.data[frm.offset[0] + xpos + ypos * frm.stride[0]..], frm.stride[0], ysrc, ystride, h); | |
2f9923e6 KS |
63 | |
64 | let (cw, ch) = refpic.get_dimensions(1); | |
65 | let mvx = mv.x >> 3; | |
66 | let mvy = mv.y >> 3; | |
67 | let dx = (mv.x & 7) as u16; | |
68 | let dy = (mv.y & 7) as u16; | |
69 | let src_x = ((xpos >> 1) as isize) + (mvx as isize); | |
70 | let src_y = ((ypos >> 1) as isize) + (mvy as isize); | |
71 | let suoff = refpic.get_offset(1); | |
72 | let svoff = refpic.get_offset(2); | |
73 | let sustride = refpic.get_stride(1); | |
74 | let svstride = refpic.get_stride(2); | |
75 | let cbw = w / 2; | |
76 | let cbh = h / 2; | |
77 | let (csrc, cstride) = if (src_x < 0) || (src_x + (cbw as isize) + 1 > (cw as isize)) || (src_y < 0) || (src_y + (cbh as isize) + 1 > (ch as isize)) { | |
78 | edge_emu(&refpic, src_x, src_y, cbw+1, cbh+1, &mut ebuf, 18, 1, 4); | |
79 | edge_emu(&refpic, src_x, src_y, cbw+1, cbh+1, &mut ebuf[9..], 18, 2, 4); | |
80 | ([&ebuf, &ebuf[9..]], [18, 18]) | |
81 | } else { | |
82 | ([&src[suoff + (src_x as usize) + (src_y as usize) * sustride..], | |
83 | &src[svoff + (src_x as usize) + (src_y as usize) * svstride..]], | |
84 | [sustride, svstride]) | |
85 | }; | |
86 | for chroma in 1..3 { | |
87 | let off = frm.offset[chroma] + xpos / 2 + (ypos / 2) * frm.stride[chroma]; | |
88 | chroma_interp(&mut frm.data[off..], frm.stride[chroma], csrc[chroma - 1], cstride[chroma - 1], dx, dy, cbw, cbh); | |
89 | } | |
90 | } | |
91 | ||
92 | pub fn mc_blocks(&mut self, ydst: &mut [u8], udst: &mut [u8], vdst: &mut [u8], refpic: NAVideoBufferRef<u8>, xpos: usize, ypos: usize, w: usize, h: usize, mv: MV) { | |
93 | let mode = ((mv.x & 3) + (mv.y & 3) * 4) as usize; | |
94 | ||
95 | let pre = if mode != 0 { 2 } else { 0 }; | |
96 | let post = if mode != 0 { 3 } else { 0 }; | |
97 | let (width, height) = refpic.get_dimensions(0); | |
98 | let sx = (xpos as isize) + ((mv.x >> 2) as isize); | |
99 | let sy = (ypos as isize) + ((mv.y >> 2) as isize); | |
100 | ||
101 | const EBUF_STRIDE: usize = 32; | |
102 | let mut ebuf = [0u8; EBUF_STRIDE * (16 + 2 + 3)]; | |
103 | ||
76431444 KS |
104 | let wmode = match w { |
105 | 4 => 0, | |
106 | 8 => 1, | |
107 | _ => 2, | |
108 | }; | |
2f9923e6 KS |
109 | if (sx - pre < 0) || (sx + (w as isize) + post > (width as isize)) || |
110 | (sy - pre < 0) || (sy + (h as isize) + post > (height as isize)) { | |
111 | let edge = (pre + post) as usize; | |
112 | edge_emu(&refpic, sx - pre, sy - pre, w + edge, h + edge, | |
113 | &mut ebuf, EBUF_STRIDE, 0, 0); | |
76431444 | 114 | (H264_LUMA_INTERP[wmode][mode])(ydst, 16, &ebuf, EBUF_STRIDE, h); |
2f9923e6 KS |
115 | } else { |
116 | let sstride = refpic.get_stride(0); | |
117 | let soff = refpic.get_offset(0); | |
118 | let sdta = refpic.get_data(); | |
119 | let sbuf: &[u8] = sdta.as_slice(); | |
120 | let saddr = soff + ((sx - pre) as usize) + ((sy - pre) as usize) * sstride; | |
76431444 | 121 | (H264_LUMA_INTERP[wmode][mode])(ydst, 16, &sbuf[saddr..], sstride, h); |
2f9923e6 KS |
122 | } |
123 | ||
124 | let (cw, ch) = refpic.get_dimensions(1); | |
125 | let mvx = mv.x >> 3; | |
126 | let mvy = mv.y >> 3; | |
127 | let dx = (mv.x & 7) as u16; | |
128 | let dy = (mv.y & 7) as u16; | |
129 | let src_x = ((xpos >> 1) as isize) + (mvx as isize); | |
130 | let src_y = ((ypos >> 1) as isize) + (mvy as isize); | |
131 | let suoff = refpic.get_offset(1); | |
132 | let svoff = refpic.get_offset(2); | |
133 | let sustride = refpic.get_stride(1); | |
134 | let svstride = refpic.get_stride(2); | |
135 | let src = refpic.get_data(); | |
136 | let cbw = w / 2; | |
137 | let cbh = h / 2; | |
138 | let (csrc, cstride) = if (src_x < 0) || (src_x + (cbw as isize) + 1 > (cw as isize)) || (src_y < 0) || (src_y + (cbh as isize) + 1 > (ch as isize)) { | |
139 | edge_emu(&refpic, src_x, src_y, cbw+1, cbh+1, &mut ebuf, 18, 1, 4); | |
140 | edge_emu(&refpic, src_x, src_y, cbw+1, cbh+1, &mut ebuf[9..], 18, 2, 4); | |
141 | ([&ebuf, &ebuf[9..]], [18, 18]) | |
142 | } else { | |
143 | ([&src[suoff + (src_x as usize) + (src_y as usize) * sustride..], | |
144 | &src[svoff + (src_x as usize) + (src_y as usize) * svstride..]], | |
145 | [sustride, svstride]) | |
146 | }; | |
147 | chroma_interp(udst, 16, csrc[0], cstride[0], dx, dy, cbw, cbh); | |
148 | chroma_interp(vdst, 16, csrc[1], cstride[1], dx, dy, cbw, cbh); | |
149 | } | |
150 | ||
151 | pub fn avg(&mut self, dst: &mut [u8], dstride: usize, bw: usize, bh: usize, comp: usize) { | |
152 | let afrm = NASimpleVideoFrame::from_video_buf(&mut self.avg_buf).unwrap(); | |
153 | let src = &afrm.data[afrm.offset[comp]..]; | |
154 | let sstride = afrm.stride[comp]; | |
155 | for (dline, sline) in dst.chunks_mut(dstride).zip(src.chunks(sstride)).take(bh) { | |
156 | for (dst, src) in dline.iter_mut().zip(sline.iter()).take(bw) { | |
157 | *dst = ((u16::from(*dst) + u16::from(*src) + 1) >> 1) as u8; | |
158 | } | |
159 | } | |
160 | } | |
161 | ||
162 | pub fn do_mc_avg(&mut self, frm: &mut NASimpleVideoFrame<u8>, refpic: NAVideoBufferRef<u8>, xpos: usize, ypos: usize, w: usize, h: usize, mv: MV) { | |
163 | let mut abuf = self.avg_buf.clone(); | |
164 | let mut afrm = NASimpleVideoFrame::from_video_buf(&mut abuf).unwrap(); | |
165 | let amv = MV { x: mv.x + (xpos as i16) * 4, y: mv.y + (ypos as i16) * 4 }; | |
166 | self.do_mc(&mut afrm, refpic, 0, 0, w, h, amv); | |
167 | for comp in 0..3 { | |
168 | let shift = if comp == 0 { 0 } else { 1 }; | |
169 | self.avg(&mut frm.data[frm.offset[comp] + (xpos >> shift) + (ypos >> shift) * frm.stride[comp]..], frm.stride[comp], w >> shift, h >> shift, comp); | |
170 | } | |
171 | } | |
172 | ||
2f9923e6 KS |
173 | pub fn gray_block(&mut self, frm: &mut NASimpleVideoFrame<u8>, x: usize, y: usize, w: usize, h: usize) { |
174 | let yoff = frm.offset[0] + x + y * frm.stride[0]; | |
175 | let coff = [frm.offset[1] + x / 2 + y / 2 * frm.stride[1], | |
176 | frm.offset[2] + x / 2 + y / 2 * frm.stride[2]]; | |
177 | for row in frm.data[yoff..].chunks_mut(frm.stride[0]).take(h) { | |
178 | for el in row[..w].iter_mut() { | |
179 | *el = 128; | |
180 | } | |
181 | } | |
182 | for chroma in 0..2 { | |
183 | for row in frm.data[coff[chroma]..].chunks_mut(frm.stride[chroma + 1]).take(h / 2) { | |
184 | for el in row[..w / 2].iter_mut() { | |
185 | *el = 128; | |
186 | } | |
187 | } | |
188 | } | |
189 | } | |
190 | } | |
76431444 KS |
191 | |
192 | fn put_block_weighted(dst: &mut [u8], stride: usize, src: &[u8], w: usize, h: usize, wparams: [i8; 3]) { | |
193 | let weight = i16::from(wparams[0]); | |
194 | let offset = i16::from(wparams[1]); | |
195 | let wshift = wparams[2] as u8; | |
196 | let bias = (1 << wshift) >> 1; | |
197 | ||
198 | for (drow, srow) in dst.chunks_mut(stride).zip(src.chunks(16)).take(h) { | |
199 | for (dst, &src) in drow[..w].iter_mut().zip(srow.iter()) { | |
200 | *dst = clip_u8(((i16::from(src) * weight + bias) >> wshift) + offset); | |
201 | } | |
202 | } | |
203 | } | |
204 | ||
205 | fn put_blk_w_2(dst: &mut [u8], stride: usize, src: &[u8], h: usize, wparams: [i8; 3]) { | |
206 | put_block_weighted(dst, stride, src, 2, h, wparams); | |
207 | } | |
208 | fn put_blk_w_4(dst: &mut [u8], stride: usize, src: &[u8], h: usize, wparams: [i8; 3]) { | |
209 | put_block_weighted(dst, stride, src, 4, h, wparams); | |
210 | } | |
211 | fn put_blk_w_8(dst: &mut [u8], stride: usize, src: &[u8], h: usize, wparams: [i8; 3]) { | |
212 | put_block_weighted(dst, stride, src, 8, h, wparams); | |
213 | } | |
214 | fn put_blk_w_16(dst: &mut [u8], stride: usize, src: &[u8], h: usize, wparams: [i8; 3]) { | |
215 | put_block_weighted(dst, stride, src, 16, h, wparams); | |
216 | } | |
217 | ||
218 | fn put_block_weighted2(dst: &mut [u8], stride: usize, src0: &[u8], src1: &[u8], w: usize, h: usize, wparams: [i8; 5]) { | |
219 | let weight0 = i16::from(wparams[0]); | |
220 | let offset0 = i16::from(wparams[1]); | |
221 | let weight1 = i16::from(wparams[2]); | |
222 | let offset1 = i16::from(wparams[3]); | |
223 | let wshift = (wparams[4] as u8) + 1; | |
224 | let offset = (offset0 + offset1 + 1) >> 1; | |
225 | let bias = (1 << wshift) >> 1; | |
226 | ||
227 | for (drow, (srow0, srow1)) in dst.chunks_mut(stride).zip(src0.chunks(16).zip(src1.chunks(16))).take(h) { | |
228 | for (dst, (&src0, &src1)) in drow[..w].iter_mut().zip(srow0.iter().zip(srow1.iter())) { | |
229 | *dst = clip_u8(((i16::from(src0) * weight0 + i16::from(src1) * weight1 + bias) >> wshift) + offset); | |
230 | } | |
231 | } | |
232 | } | |
233 | ||
234 | fn put_blk_w2_2(dst: &mut [u8], stride: usize, src0: &[u8], src1: &[u8], h: usize, wparams: [i8; 5]) { | |
235 | put_block_weighted2(dst, stride, src0, src1, 2, h, wparams); | |
236 | } | |
237 | fn put_blk_w2_4(dst: &mut [u8], stride: usize, src0: &[u8], src1: &[u8], h: usize, wparams: [i8; 5]) { | |
238 | put_block_weighted2(dst, stride, src0, src1, 4, h, wparams); | |
239 | } | |
240 | fn put_blk_w2_8(dst: &mut [u8], stride: usize, src0: &[u8], src1: &[u8], h: usize, wparams: [i8; 5]) { | |
241 | put_block_weighted2(dst, stride, src0, src1, 8, h, wparams); | |
242 | } | |
243 | fn put_blk_w2_16(dst: &mut [u8], stride: usize, src0: &[u8], src1: &[u8], h: usize, wparams: [i8; 5]) { | |
244 | put_block_weighted2(dst, stride, src0, src1, 16, h, wparams); | |
245 | } |