annotate vp8dsp.c @ 12229:cf675a4d5195 libavcodec

Eliminate a LUT in escape decoding in VP8 decode_block_coeffs
author darkshikari
date Thu, 22 Jul 2010 22:08:09 +0000
parents 80b142c2e9f7
children e08d65897115
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
1 /**
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
2 * VP8 compatible video decoder
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
3 *
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
4 * Copyright (C) 2010 David Conrad
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
5 * Copyright (C) 2010 Ronald S. Bultje
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
6 *
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
7 * This file is part of FFmpeg.
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
8 *
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
9 * FFmpeg is free software; you can redistribute it and/or
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
10 * modify it under the terms of the GNU Lesser General Public
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
11 * License as published by the Free Software Foundation; either
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
12 * version 2.1 of the License, or (at your option) any later version.
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
13 *
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
14 * FFmpeg is distributed in the hope that it will be useful,
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
17 * Lesser General Public License for more details.
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
18 *
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
19 * You should have received a copy of the GNU Lesser General Public
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
20 * License along with FFmpeg; if not, write to the Free Software
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
22 */
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
23
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
24 #include "dsputil.h"
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
25 #include "vp8dsp.h"
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
26
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
27 // TODO: Maybe add dequant
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
28 static void vp8_luma_dc_wht_c(DCTELEM block[4][4][16], DCTELEM dc[16])
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
29 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
30 int i, t0, t1, t2, t3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
31
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
32 for (i = 0; i < 4; i++) {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
33 t0 = dc[0*4+i] + dc[3*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
34 t1 = dc[1*4+i] + dc[2*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
35 t2 = dc[1*4+i] - dc[2*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
36 t3 = dc[0*4+i] - dc[3*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
37
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
38 dc[0*4+i] = t0 + t1;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
39 dc[1*4+i] = t3 + t2;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
40 dc[2*4+i] = t0 - t1;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
41 dc[3*4+i] = t3 - t2;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
42 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
43
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
44 for (i = 0; i < 4; i++) {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
45 t0 = dc[i*4+0] + dc[i*4+3] + 3; // rounding
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
46 t1 = dc[i*4+1] + dc[i*4+2];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
47 t2 = dc[i*4+1] - dc[i*4+2];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
48 t3 = dc[i*4+0] - dc[i*4+3] + 3; // rounding
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
49
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
50 *block[i][0] = (t0 + t1) >> 3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
51 *block[i][1] = (t3 + t2) >> 3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
52 *block[i][2] = (t0 - t1) >> 3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
53 *block[i][3] = (t3 - t2) >> 3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
54 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
55 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
56
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
57
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
58 #define MUL_20091(a) ((((a)*20091) >> 16) + (a))
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
59 #define MUL_35468(a) (((a)*35468) >> 16)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
60
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
61 static void vp8_idct_add_c(uint8_t *dst, DCTELEM block[16], int stride)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
62 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
63 int i, t0, t1, t2, t3;
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
64 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
65 DCTELEM tmp[16];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
66
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
67 for (i = 0; i < 4; i++) {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
68 t0 = block[0*4+i] + block[2*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
69 t1 = block[0*4+i] - block[2*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
70 t2 = MUL_35468(block[1*4+i]) - MUL_20091(block[3*4+i]);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
71 t3 = MUL_20091(block[1*4+i]) + MUL_35468(block[3*4+i]);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
72
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
73 tmp[i*4+0] = t0 + t3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
74 tmp[i*4+1] = t1 + t2;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
75 tmp[i*4+2] = t1 - t2;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
76 tmp[i*4+3] = t0 - t3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
77 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
78
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
79 for (i = 0; i < 4; i++) {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
80 t0 = tmp[0*4+i] + tmp[2*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
81 t1 = tmp[0*4+i] - tmp[2*4+i];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
82 t2 = MUL_35468(tmp[1*4+i]) - MUL_20091(tmp[3*4+i]);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
83 t3 = MUL_20091(tmp[1*4+i]) + MUL_35468(tmp[3*4+i]);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
84
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
85 dst[0] = cm[dst[0] + ((t0 + t3 + 4) >> 3)];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
86 dst[1] = cm[dst[1] + ((t1 + t2 + 4) >> 3)];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
87 dst[2] = cm[dst[2] + ((t1 - t2 + 4) >> 3)];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
88 dst[3] = cm[dst[3] + ((t0 - t3 + 4) >> 3)];
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
89 dst += stride;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
90 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
91 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
92
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
93 static void vp8_idct_dc_add_c(uint8_t *dst, DCTELEM block[16], int stride)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
94 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
95 int i, dc = (block[0] + 4) >> 3;
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
96 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP + dc;
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
97
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
98 for (i = 0; i < 4; i++) {
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
99 dst[0] = cm[dst[0]];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
100 dst[1] = cm[dst[1]];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
101 dst[2] = cm[dst[2]];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
102 dst[3] = cm[dst[3]];
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
103 dst += stride;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
104 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
105 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
106
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
107
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
108 // because I like only having two parameters to pass functions...
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
109 #define LOAD_PIXELS\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
110 int av_unused p3 = p[-4*stride];\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
111 int av_unused p2 = p[-3*stride];\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
112 int av_unused p1 = p[-2*stride];\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
113 int av_unused p0 = p[-1*stride];\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
114 int av_unused q0 = p[ 0*stride];\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
115 int av_unused q1 = p[ 1*stride];\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
116 int av_unused q2 = p[ 2*stride];\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
117 int av_unused q3 = p[ 3*stride];
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
118
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
119 #define clip_int8(n) (cm[n+0x80]-0x80)
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
120
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
121 static av_always_inline void filter_common(uint8_t *p, int stride, int is4tap)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
122 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
123 LOAD_PIXELS
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
124 int a, f1, f2;
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
125 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
126
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
127 a = 3*(q0 - p0);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
128
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
129 if (is4tap)
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
130 a += clip_int8(p1 - q1);
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
131
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
132 a = clip_int8(a);
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
133
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
134 // We deviate from the spec here with c(a+3) >> 3
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
135 // since that's what libvpx does.
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
136 f1 = FFMIN(a+4, 127) >> 3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
137 f2 = FFMIN(a+3, 127) >> 3;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
138
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
139 // Despite what the spec says, we do need to clamp here to
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
140 // be bitexact with libvpx.
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
141 p[-1*stride] = cm[p0 + f2];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
142 p[ 0*stride] = cm[q0 - f1];
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
143
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
144 // only used for _inner on blocks without high edge variance
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
145 if (!is4tap) {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
146 a = (f1+1)>>1;
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
147 p[-2*stride] = cm[p1 + a];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
148 p[ 1*stride] = cm[q1 - a];
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
149 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
150 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
151
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
152 static av_always_inline int simple_limit(uint8_t *p, int stride, int flim)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
153 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
154 LOAD_PIXELS
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
155 return 2*FFABS(p0-q0) + (FFABS(p1-q1) >> 1) <= flim;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
156 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
157
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
158 /**
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
159 * E - limit at the macroblock edge
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
160 * I - limit for interior difference
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
161 */
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
162 static av_always_inline int normal_limit(uint8_t *p, int stride, int E, int I)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
163 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
164 LOAD_PIXELS
12081
812e23197d64 VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents: 12011
diff changeset
165 return simple_limit(p, stride, E)
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
166 && FFABS(p3-p2) <= I && FFABS(p2-p1) <= I && FFABS(p1-p0) <= I
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
167 && FFABS(q3-q2) <= I && FFABS(q2-q1) <= I && FFABS(q1-q0) <= I;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
168 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
169
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
170 // high edge variance
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
171 static av_always_inline int hev(uint8_t *p, int stride, int thresh)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
172 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
173 LOAD_PIXELS
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
174 return FFABS(p1-p0) > thresh || FFABS(q1-q0) > thresh;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
175 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
176
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
177 static av_always_inline void filter_mbedge(uint8_t *p, int stride)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
178 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
179 int a0, a1, a2, w;
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
180 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
181
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
182 LOAD_PIXELS
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
183
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
184 w = clip_int8(p1-q1);
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
185 w = clip_int8(w + 3*(q0-p0));
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
186
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
187 a0 = (27*w + 63) >> 7;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
188 a1 = (18*w + 63) >> 7;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
189 a2 = ( 9*w + 63) >> 7;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
190
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
191 p[-3*stride] = cm[p2 + a2];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
192 p[-2*stride] = cm[p1 + a1];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
193 p[-1*stride] = cm[p0 + a0];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
194 p[ 0*stride] = cm[q0 - a0];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
195 p[ 1*stride] = cm[q1 - a1];
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
196 p[ 2*stride] = cm[q2 - a2];
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
197 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
198
12194
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
199 #define LOOP_FILTER(dir, size, stridea, strideb, maybe_inline) \
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
200 static maybe_inline void vp8_ ## dir ## _loop_filter ## size ## _c(uint8_t *dst, int stride,\
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
201 int flim_E, int flim_I, int hev_thresh)\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
202 {\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
203 int i;\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
204 \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
205 for (i = 0; i < size; i++)\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
206 if (normal_limit(dst+i*stridea, strideb, flim_E, flim_I)) {\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
207 if (hev(dst+i*stridea, strideb, hev_thresh))\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
208 filter_common(dst+i*stridea, strideb, 1);\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
209 else\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
210 filter_mbedge(dst+i*stridea, strideb);\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
211 }\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
212 }\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
213 \
12194
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
214 static maybe_inline void vp8_ ## dir ## _loop_filter ## size ## _inner_c(uint8_t *dst, int stride,\
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
215 int flim_E, int flim_I, int hev_thresh)\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
216 {\
12008
98fd80705850 Faster C VP8 normal inner loop filter
darkshikari
parents: 12007
diff changeset
217 int i;\
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
218 \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
219 for (i = 0; i < size; i++)\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
220 if (normal_limit(dst+i*stridea, strideb, flim_E, flim_I)) {\
12008
98fd80705850 Faster C VP8 normal inner loop filter
darkshikari
parents: 12007
diff changeset
221 int hv = hev(dst+i*stridea, strideb, hev_thresh);\
98fd80705850 Faster C VP8 normal inner loop filter
darkshikari
parents: 12007
diff changeset
222 if (hv) \
98fd80705850 Faster C VP8 normal inner loop filter
darkshikari
parents: 12007
diff changeset
223 filter_common(dst+i*stridea, strideb, 1);\
98fd80705850 Faster C VP8 normal inner loop filter
darkshikari
parents: 12007
diff changeset
224 else \
98fd80705850 Faster C VP8 normal inner loop filter
darkshikari
parents: 12007
diff changeset
225 filter_common(dst+i*stridea, strideb, 0);\
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
226 }\
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
227 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
228
12194
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
229 LOOP_FILTER(v, 16, 1, stride,)
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
230 LOOP_FILTER(h, 16, stride, 1,)
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
231
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
232 #define UV_LOOP_FILTER(dir, stridea, strideb) \
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
233 LOOP_FILTER(dir, 8, stridea, strideb, av_always_inline) \
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
234 static void vp8_ ## dir ## _loop_filter8uv_c(uint8_t *dstU, uint8_t *dstV, int stride,\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
235 int fE, int fI, int hev_thresh)\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
236 {\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
237 vp8_ ## dir ## _loop_filter8_c(dstU, stride, fE, fI, hev_thresh);\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
238 vp8_ ## dir ## _loop_filter8_c(dstV, stride, fE, fI, hev_thresh);\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
239 }\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
240 static void vp8_ ## dir ## _loop_filter8uv_inner_c(uint8_t *dstU, uint8_t *dstV, int stride,\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
241 int fE, int fI, int hev_thresh)\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
242 {\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
243 vp8_ ## dir ## _loop_filter8_inner_c(dstU, stride, fE, fI, hev_thresh);\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
244 vp8_ ## dir ## _loop_filter8_inner_c(dstV, stride, fE, fI, hev_thresh);\
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
245 }
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
246
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
247 UV_LOOP_FILTER(v, 1, stride)
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
248 UV_LOOP_FILTER(h, stride, 1)
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
249
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
250 static void vp8_v_loop_filter_simple_c(uint8_t *dst, int stride, int flim)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
251 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
252 int i;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
253
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
254 for (i = 0; i < 16; i++)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
255 if (simple_limit(dst+i, stride, flim))
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
256 filter_common(dst+i, stride, 1);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
257 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
258
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
259 static void vp8_h_loop_filter_simple_c(uint8_t *dst, int stride, int flim)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
260 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
261 int i;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
262
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
263 for (i = 0; i < 16; i++)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
264 if (simple_limit(dst+i*stride, 1, flim))
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
265 filter_common(dst+i*stride, 1, 1);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
266 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
267
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
268 static const uint8_t subpel_filters[7][6] = {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
269 { 0, 6, 123, 12, 1, 0 },
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
270 { 2, 11, 108, 36, 8, 1 },
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
271 { 0, 9, 93, 50, 6, 0 },
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
272 { 3, 16, 77, 77, 16, 3 },
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
273 { 0, 6, 50, 93, 9, 0 },
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
274 { 1, 8, 36, 108, 11, 2 },
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
275 { 0, 1, 12, 123, 6, 0 },
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
276 };
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
277
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
278 #define PUT_PIXELS(WIDTH) \
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
279 static void put_vp8_pixels ## WIDTH ##_c(uint8_t *dst, int dststride, uint8_t *src, int srcstride, int h, int x, int y) { \
11956
496d1300204d Really fix r23782
darkshikari
parents: 11954
diff changeset
280 int i; \
496d1300204d Really fix r23782
darkshikari
parents: 11954
diff changeset
281 for (i = 0; i < h; i++, dst+= dststride, src+= srcstride) { \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
282 memcpy(dst, src, WIDTH); \
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
283 } \
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
284 }
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
285
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
286 PUT_PIXELS(16)
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
287 PUT_PIXELS(8)
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
288 PUT_PIXELS(4)
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
289
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
290 #define FILTER_6TAP(src, F, stride) \
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
291 cm[(F[2]*src[x+0*stride] - F[1]*src[x-1*stride] + F[0]*src[x-2*stride] + \
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
292 F[3]*src[x+1*stride] - F[4]*src[x+2*stride] + F[5]*src[x+3*stride] + 64) >> 7]
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
293
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
294 #define FILTER_4TAP(src, F, stride) \
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
295 cm[(F[2]*src[x+0*stride] - F[1]*src[x-1*stride] + \
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
296 F[3]*src[x+1*stride] - F[4]*src[x+2*stride] + 64) >> 7]
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
297
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
298 #define VP8_EPEL_H(SIZE, FILTER, FILTERNAME) \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
299 static void put_vp8_epel ## SIZE ## _ ## FILTERNAME ## _c(uint8_t *dst, int dststride, uint8_t *src, int srcstride, int h, int mx, int my) \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
300 { \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
301 const uint8_t *filter = subpel_filters[mx-1]; \
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
302 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
303 int x, y; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
304 \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
305 for (y = 0; y < h; y++) { \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
306 for (x = 0; x < SIZE; x++) \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
307 dst[x] = FILTER(src, filter, 1); \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
308 dst += dststride; \
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
309 src += srcstride; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
310 } \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
311 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
312 #define VP8_EPEL_V(SIZE, FILTER, FILTERNAME) \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
313 static void put_vp8_epel ## SIZE ## _ ## FILTERNAME ## _c(uint8_t *dst, int dststride, uint8_t *src, int srcstride, int h, int mx, int my) \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
314 { \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
315 const uint8_t *filter = subpel_filters[my-1]; \
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
316 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
317 int x, y; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
318 \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
319 for (y = 0; y < h; y++) { \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
320 for (x = 0; x < SIZE; x++) \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
321 dst[x] = FILTER(src, filter, srcstride); \
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
322 dst += dststride; \
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
323 src += srcstride; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
324 } \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
325 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
326 #define VP8_EPEL_HV(SIZE, FILTERX, FILTERY, FILTERNAME) \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
327 static void put_vp8_epel ## SIZE ## _ ## FILTERNAME ## _c(uint8_t *dst, int dststride, uint8_t *src, int srcstride, int h, int mx, int my) \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
328 { \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
329 const uint8_t *filter = subpel_filters[mx-1]; \
12007
ec7be1d7d5b4 Use crop table in C implementations of VP8 DSP functions.
darkshikari
parents: 11985
diff changeset
330 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
331 int x, y; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
332 uint8_t tmp_array[(2*SIZE+5)*SIZE]; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
333 uint8_t *tmp = tmp_array; \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
334 src -= 2*srcstride; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
335 \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
336 for (y = 0; y < h+5; y++) { \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
337 for (x = 0; x < SIZE; x++) \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
338 tmp[x] = FILTERX(src, filter, 1); \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
339 tmp += SIZE; \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
340 src += srcstride; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
341 } \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
342 \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
343 tmp = tmp_array + 2*SIZE; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
344 filter = subpel_filters[my-1]; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
345 \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
346 for (y = 0; y < h; y++) { \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
347 for (x = 0; x < SIZE; x++) \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
348 dst[x] = FILTERY(tmp, filter, SIZE); \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
349 dst += dststride; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
350 tmp += SIZE; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
351 } \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
352 }
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
353
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
354 VP8_EPEL_H(16, FILTER_4TAP, h4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
355 VP8_EPEL_H(8, FILTER_4TAP, h4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
356 VP8_EPEL_H(4, FILTER_4TAP, h4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
357 VP8_EPEL_H(16, FILTER_6TAP, h6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
358 VP8_EPEL_H(8, FILTER_6TAP, h6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
359 VP8_EPEL_H(4, FILTER_6TAP, h6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
360 VP8_EPEL_V(16, FILTER_4TAP, v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
361 VP8_EPEL_V(8, FILTER_4TAP, v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
362 VP8_EPEL_V(4, FILTER_4TAP, v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
363 VP8_EPEL_V(16, FILTER_6TAP, v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
364 VP8_EPEL_V(8, FILTER_6TAP, v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
365 VP8_EPEL_V(4, FILTER_6TAP, v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
366 VP8_EPEL_HV(16, FILTER_4TAP, FILTER_4TAP, h4v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
367 VP8_EPEL_HV(8, FILTER_4TAP, FILTER_4TAP, h4v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
368 VP8_EPEL_HV(4, FILTER_4TAP, FILTER_4TAP, h4v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
369 VP8_EPEL_HV(16, FILTER_4TAP, FILTER_6TAP, h4v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
370 VP8_EPEL_HV(8, FILTER_4TAP, FILTER_6TAP, h4v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
371 VP8_EPEL_HV(4, FILTER_4TAP, FILTER_6TAP, h4v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
372 VP8_EPEL_HV(16, FILTER_6TAP, FILTER_4TAP, h6v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
373 VP8_EPEL_HV(8, FILTER_6TAP, FILTER_4TAP, h6v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
374 VP8_EPEL_HV(4, FILTER_6TAP, FILTER_4TAP, h6v4)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
375 VP8_EPEL_HV(16, FILTER_6TAP, FILTER_6TAP, h6v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
376 VP8_EPEL_HV(8, FILTER_6TAP, FILTER_6TAP, h6v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
377 VP8_EPEL_HV(4, FILTER_6TAP, FILTER_6TAP, h6v6)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
378
11974
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
379 #define VP8_BILINEAR(SIZE) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
380 static void put_vp8_bilinear ## SIZE ## _h_c(uint8_t *dst, int stride, uint8_t *src, int s2, int h, int mx, int my) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
381 { \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
382 int a = 8-mx, b = mx; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
383 int x, y; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
384 \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
385 for (y = 0; y < h; y++) { \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
386 for (x = 0; x < SIZE; x++) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
387 dst[x] = (a*src[x] + b*src[x+1] + 4) >> 3; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
388 dst += stride; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
389 src += stride; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
390 } \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
391 } \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
392 static void put_vp8_bilinear ## SIZE ## _v_c(uint8_t *dst, int stride, uint8_t *src, int s2, int h, int mx, int my) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
393 { \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
394 int c = 8-my, d = my; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
395 int x, y; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
396 \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
397 for (y = 0; y < h; y++) { \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
398 for (x = 0; x < SIZE; x++) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
399 dst[x] = (c*src[x] + d*src[x+stride] + 4) >> 3; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
400 dst += stride; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
401 src += stride; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
402 } \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
403 } \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
404 \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
405 static void put_vp8_bilinear ## SIZE ## _hv_c(uint8_t *dst, int stride, uint8_t *src, int s2, int h, int mx, int my) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
406 { \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
407 int a = 8-mx, b = mx; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
408 int c = 8-my, d = my; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
409 int x, y; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
410 uint8_t tmp_array[(2*SIZE+1)*SIZE]; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
411 uint8_t *tmp = tmp_array; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
412 \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
413 for (y = 0; y < h+1; y++) { \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
414 for (x = 0; x < SIZE; x++) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
415 tmp[x] = (a*src[x] + b*src[x+1] + 4) >> 3; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
416 tmp += SIZE; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
417 src += stride; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
418 } \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
419 \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
420 tmp = tmp_array; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
421 \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
422 for (y = 0; y < h; y++) { \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
423 for (x = 0; x < SIZE; x++) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
424 dst[x] = (c*tmp[x] + d*tmp[x+SIZE] + 4) >> 3; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
425 dst += stride; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
426 tmp += SIZE; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
427 } \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
428 }
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
429
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
430 VP8_BILINEAR(16)
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
431 VP8_BILINEAR(8)
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
432 VP8_BILINEAR(4)
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
433
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
434 #define VP8_MC_FUNC(IDX, SIZE) \
11950
56aba5a9761c Make VP8 DSP functions take two strides
darkshikari
parents: 11921
diff changeset
435 dsp->put_vp8_epel_pixels_tab[IDX][0][0] = put_vp8_pixels ## SIZE ## _c; \
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
436 dsp->put_vp8_epel_pixels_tab[IDX][0][1] = put_vp8_epel ## SIZE ## _h4_c; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
437 dsp->put_vp8_epel_pixels_tab[IDX][0][2] = put_vp8_epel ## SIZE ## _h6_c; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
438 dsp->put_vp8_epel_pixels_tab[IDX][1][0] = put_vp8_epel ## SIZE ## _v4_c; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
439 dsp->put_vp8_epel_pixels_tab[IDX][1][1] = put_vp8_epel ## SIZE ## _h4v4_c; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
440 dsp->put_vp8_epel_pixels_tab[IDX][1][2] = put_vp8_epel ## SIZE ## _h6v4_c; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
441 dsp->put_vp8_epel_pixels_tab[IDX][2][0] = put_vp8_epel ## SIZE ## _v6_c; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
442 dsp->put_vp8_epel_pixels_tab[IDX][2][1] = put_vp8_epel ## SIZE ## _h4v6_c; \
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
443 dsp->put_vp8_epel_pixels_tab[IDX][2][2] = put_vp8_epel ## SIZE ## _h6v6_c
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
444
11974
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
445 #define VP8_BILINEAR_MC_FUNC(IDX, SIZE) \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
446 dsp->put_vp8_bilinear_pixels_tab[IDX][0][0] = put_vp8_pixels ## SIZE ## _c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
447 dsp->put_vp8_bilinear_pixels_tab[IDX][0][1] = put_vp8_bilinear ## SIZE ## _h_c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
448 dsp->put_vp8_bilinear_pixels_tab[IDX][0][2] = put_vp8_bilinear ## SIZE ## _h_c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
449 dsp->put_vp8_bilinear_pixels_tab[IDX][1][0] = put_vp8_bilinear ## SIZE ## _v_c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
450 dsp->put_vp8_bilinear_pixels_tab[IDX][1][1] = put_vp8_bilinear ## SIZE ## _hv_c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
451 dsp->put_vp8_bilinear_pixels_tab[IDX][1][2] = put_vp8_bilinear ## SIZE ## _hv_c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
452 dsp->put_vp8_bilinear_pixels_tab[IDX][2][0] = put_vp8_bilinear ## SIZE ## _v_c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
453 dsp->put_vp8_bilinear_pixels_tab[IDX][2][1] = put_vp8_bilinear ## SIZE ## _hv_c; \
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
454 dsp->put_vp8_bilinear_pixels_tab[IDX][2][2] = put_vp8_bilinear ## SIZE ## _hv_c
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
455
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
456 av_cold void ff_vp8dsp_init(VP8DSPContext *dsp)
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
457 {
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
458 dsp->vp8_luma_dc_wht = vp8_luma_dc_wht_c;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
459 dsp->vp8_idct_add = vp8_idct_add_c;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
460 dsp->vp8_idct_dc_add = vp8_idct_dc_add_c;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
461
12194
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
462 dsp->vp8_v_loop_filter16y = vp8_v_loop_filter16_c;
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
463 dsp->vp8_h_loop_filter16y = vp8_h_loop_filter16_c;
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
464 dsp->vp8_v_loop_filter8uv = vp8_v_loop_filter8uv_c;
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
465 dsp->vp8_h_loop_filter8uv = vp8_h_loop_filter8uv_c;
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
466
12194
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
467 dsp->vp8_v_loop_filter16y_inner = vp8_v_loop_filter16_inner_c;
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
468 dsp->vp8_h_loop_filter16y_inner = vp8_h_loop_filter16_inner_c;
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
469 dsp->vp8_v_loop_filter8uv_inner = vp8_v_loop_filter8uv_inner_c;
80b142c2e9f7 Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents: 12081
diff changeset
470 dsp->vp8_h_loop_filter8uv_inner = vp8_h_loop_filter8uv_inner_c;
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
471
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
472 dsp->vp8_v_loop_filter_simple = vp8_v_loop_filter_simple_c;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
473 dsp->vp8_h_loop_filter_simple = vp8_h_loop_filter_simple_c;
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
474
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
475 VP8_MC_FUNC(0, 16);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
476 VP8_MC_FUNC(1, 8);
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
477 VP8_MC_FUNC(2, 4);
11974
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
478
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
479 VP8_BILINEAR_MC_FUNC(0, 16);
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
480 VP8_BILINEAR_MC_FUNC(1, 8);
356b20a6566d VP8 bilinear filter
conrad
parents: 11956
diff changeset
481 VP8_BILINEAR_MC_FUNC(2, 4);
11975
c3afb5be0d9b First shot at VP8 optimizations:
rbultje
parents: 11974
diff changeset
482
11985
1cffcc7b1470 Fix linking if MMX is disabled.
stefano
parents: 11975
diff changeset
483 if (HAVE_MMX)
11975
c3afb5be0d9b First shot at VP8 optimizations:
rbultje
parents: 11974
diff changeset
484 ff_vp8dsp_init_x86(dsp);
12011
f96187e79438 Altivec VP8 MC functions
conrad
parents: 12008
diff changeset
485 if (HAVE_ALTIVEC)
f96187e79438 Altivec VP8 MC functions
conrad
parents: 12008
diff changeset
486 ff_vp8dsp_init_altivec(dsp);
11921
f2007d7c3f1d Native VP8 decoder.
rbultje
parents:
diff changeset
487 }