annotate libmpcodecs/vf_spp.c @ 11993:4e2d99dbef78

spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
author michael
date Tue, 24 Feb 2004 11:23:27 +0000
parents cf76671b3d77
children a7751694c177
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
1 /*
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
2 Copyright (C) 2003 Michael Niedermayer <michaelni@gmx.at>
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
3
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
4 This program is free software; you can redistribute it and/or modify
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
5 it under the terms of the GNU General Public License as published by
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
6 the Free Software Foundation; either version 2 of the License, or
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
7 (at your option) any later version.
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
8
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
9 This program is distributed in the hope that it will be useful,
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
12 GNU General Public License for more details.
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
13
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
14 You should have received a copy of the GNU General Public License
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
15 along with this program; if not, write to the Free Software
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
17 */
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
18
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
19 /*
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
20 * This implementation is based on an algorithm described in
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
21 * "Aria Nosratinia Embedded Post-Processing for
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
22 * Enhancement of Compressed Images (1999)"
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
23 * (http://citeseer.nj.nec.com/nosratinia99embedded.html)
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
24 */
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
25
11334
michael
parents: 11308
diff changeset
26
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
27 #include <stdio.h>
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
28 #include <stdlib.h>
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
29 #include <string.h>
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
30 #include <inttypes.h>
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
31 #include <math.h>
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
32
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
33 #include "../config.h"
11335
michael
parents: 11334
diff changeset
34
michael
parents: 11334
diff changeset
35 #ifdef USE_LIBAVCODEC
michael
parents: 11334
diff changeset
36
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
37 #include "../mp_msg.h"
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
38 #include "../cpudetect.h"
11568
cf76671b3d77 Fix compilation if lavc is compiled as a shared lib. Patch by Panagiotis Issaris <takis@lumumba.luc.ac.be>
alex
parents: 11477
diff changeset
39
cf76671b3d77 Fix compilation if lavc is compiled as a shared lib. Patch by Panagiotis Issaris <takis@lumumba.luc.ac.be>
alex
parents: 11477
diff changeset
40 #ifdef USE_LIBAVCODEC_SO
cf76671b3d77 Fix compilation if lavc is compiled as a shared lib. Patch by Panagiotis Issaris <takis@lumumba.luc.ac.be>
alex
parents: 11477
diff changeset
41 #include <ffmpeg/avcodec.h>
cf76671b3d77 Fix compilation if lavc is compiled as a shared lib. Patch by Panagiotis Issaris <takis@lumumba.luc.ac.be>
alex
parents: 11477
diff changeset
42 #include <ffmpeg/dsputil.h>
cf76671b3d77 Fix compilation if lavc is compiled as a shared lib. Patch by Panagiotis Issaris <takis@lumumba.luc.ac.be>
alex
parents: 11477
diff changeset
43 #else
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
44 #include "../libavcodec/avcodec.h"
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
45 #include "../libavcodec/dsputil.h"
11568
cf76671b3d77 Fix compilation if lavc is compiled as a shared lib. Patch by Panagiotis Issaris <takis@lumumba.luc.ac.be>
alex
parents: 11477
diff changeset
46 #endif
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
47
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
48 #ifdef HAVE_MALLOC_H
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
49 #include <malloc.h>
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
50 #endif
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
51
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
52 #include "img_format.h"
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
53 #include "mp_image.h"
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
54 #include "vf.h"
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
55 #include "../libvo/fastmemcpy.h"
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
56
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
57 #define XMIN(a,b) ((a) < (b) ? (a) : (b))
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
58
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
59 //===========================================================================//
11477
9785bff83777 memcpy pix instead of black screen if no DR and codec provides no qscale table
michael
parents: 11335
diff changeset
60 static const uint8_t __attribute__((aligned(8))) dither[8][8]={
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
61 { 0, 48, 12, 60, 3, 51, 15, 63, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
62 { 32, 16, 44, 28, 35, 19, 47, 31, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
63 { 8, 56, 4, 52, 11, 59, 7, 55, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
64 { 40, 24, 36, 20, 43, 27, 39, 23, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
65 { 2, 50, 14, 62, 1, 49, 13, 61, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
66 { 34, 18, 46, 30, 33, 17, 45, 29, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
67 { 10, 58, 6, 54, 9, 57, 5, 53, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
68 { 42, 26, 38, 22, 41, 25, 37, 21, },
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
69 };
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
70
11477
9785bff83777 memcpy pix instead of black screen if no DR and codec provides no qscale table
michael
parents: 11335
diff changeset
71 static const uint8_t offset[127][2]= {
11298
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
72 {0,0},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
73 {0,0}, {4,4},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
74 {0,0}, {2,2}, {6,4}, {4,6},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
75 {0,0}, {5,1}, {2,2}, {7,3}, {4,4}, {1,5}, {6,6}, {3,7},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
76
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
77 {0,0}, {4,0}, {1,1}, {5,1}, {3,2}, {7,2}, {2,3}, {6,3},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
78 {0,4}, {4,4}, {1,5}, {5,5}, {3,6}, {7,6}, {2,7}, {6,7},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
79
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
80 {0,0}, {0,2}, {0,4}, {0,6}, {1,1}, {1,3}, {1,5}, {1,7},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
81 {2,0}, {2,2}, {2,4}, {2,6}, {3,1}, {3,3}, {3,5}, {3,7},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
82 {4,0}, {4,2}, {4,4}, {4,6}, {5,1}, {5,3}, {5,5}, {5,7},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
83 {6,0}, {6,2}, {6,4}, {6,6}, {7,1}, {7,3}, {7,5}, {7,7},
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
84
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
85 {0,0}, {4,4}, {0,4}, {4,0}, {2,2}, {6,6}, {2,6}, {6,2},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
86 {0,2}, {4,6}, {0,6}, {4,2}, {2,0}, {6,4}, {2,4}, {6,0},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
87 {1,1}, {5,5}, {1,5}, {5,1}, {3,3}, {7,7}, {3,7}, {7,3},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
88 {1,3}, {5,7}, {1,7}, {5,3}, {3,1}, {7,5}, {3,5}, {7,1},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
89 {0,1}, {4,5}, {0,5}, {4,1}, {2,3}, {6,7}, {2,7}, {6,3},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
90 {0,3}, {4,7}, {0,7}, {4,3}, {2,1}, {6,5}, {2,5}, {6,1},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
91 {1,0}, {5,4}, {1,4}, {5,0}, {3,2}, {7,6}, {3,6}, {7,2},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
92 {1,2}, {5,6}, {1,6}, {5,2}, {3,0}, {7,4}, {3,4}, {7,0},
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
93 };
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
94
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
95 struct vf_priv_s {
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
96 int log2_count;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
97 int qp;
11993
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
98 int mode;
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
99 int mpeg2;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
100 unsigned int outfmt;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
101 int temp_stride;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
102 uint8_t *src;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
103 int16_t *temp;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
104 AVCodecContext *avctx;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
105 DSPContext dsp;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
106 };
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
107
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
108 #define SHIFT 22
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
109
11993
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
110 static void hardthresh_c(DCTELEM dst[64], DCTELEM src[64], int qp, uint8_t *permutation){
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
111 int i;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
112 int bias= 0; //FIXME
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
113 unsigned int threshold1, threshold2;
11296
86916e46d445 different / faster / simpler "quantization"
michael
parents: 11295
diff changeset
114
11301
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
115 threshold1= qp*((1<<4) - bias) - 1;
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
116 threshold2= (threshold1<<1);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
117
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
118 memset(dst, 0, 64*sizeof(DCTELEM));
11296
86916e46d445 different / faster / simpler "quantization"
michael
parents: 11295
diff changeset
119 dst[0]= (src[0] + 4)>>3;
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
120
11296
86916e46d445 different / faster / simpler "quantization"
michael
parents: 11295
diff changeset
121 for(i=1; i<64; i++){
86916e46d445 different / faster / simpler "quantization"
michael
parents: 11295
diff changeset
122 int level= src[i];
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
123 if(((unsigned)(level+threshold1))>threshold2){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
124 const int j= permutation[i];
11296
86916e46d445 different / faster / simpler "quantization"
michael
parents: 11295
diff changeset
125 dst[j]= (level + 4)>>3;
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
126 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
127 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
128 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
129
11993
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
130 static void softthresh_c(DCTELEM dst[64], DCTELEM src[64], int qp, uint8_t *permutation){
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
131 int i;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
132 int bias= 0; //FIXME
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
133 unsigned int threshold1, threshold2;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
134
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
135 threshold1= qp*((1<<4) - bias) - 1;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
136 threshold2= (threshold1<<1);
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
137
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
138 memset(dst, 0, 64*sizeof(DCTELEM));
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
139 dst[0]= (src[0] + 4)>>3;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
140
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
141 for(i=1; i<64; i++){
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
142 int level= src[i];
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
143 if(((unsigned)(level+threshold1))>threshold2){
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
144 const int j= permutation[i];
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
145 if(level>0)
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
146 dst[j]= (level - threshold1 + 4)>>3;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
147 else
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
148 dst[j]= (level + threshold1 + 4)>>3;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
149 }
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
150 }
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
151 }
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
152
11301
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
153 #ifdef HAVE_MMX
11993
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
154 static void hardthresh_mmx(DCTELEM dst[64], DCTELEM src[64], int qp, uint8_t *permutation){
11301
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
155 int bias= 0; //FIXME
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
156 unsigned int threshold1;
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
157
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
158 threshold1= qp*((1<<4) - bias) - 1;
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
159
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
160 asm volatile(
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
161 #define REQUANT_CORE(dst0, dst1, dst2, dst3, src0, src1, src2, src3) \
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
162 "movq " #src0 ", %%mm0 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
163 "movq " #src1 ", %%mm1 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
164 "movq " #src2 ", %%mm2 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
165 "movq " #src3 ", %%mm3 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
166 "psubw %%mm4, %%mm0 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
167 "psubw %%mm4, %%mm1 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
168 "psubw %%mm4, %%mm2 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
169 "psubw %%mm4, %%mm3 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
170 "paddusw %%mm5, %%mm0 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
171 "paddusw %%mm5, %%mm1 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
172 "paddusw %%mm5, %%mm2 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
173 "paddusw %%mm5, %%mm3 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
174 "paddw %%mm6, %%mm0 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
175 "paddw %%mm6, %%mm1 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
176 "paddw %%mm6, %%mm2 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
177 "paddw %%mm6, %%mm3 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
178 "psubusw %%mm6, %%mm0 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
179 "psubusw %%mm6, %%mm1 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
180 "psubusw %%mm6, %%mm2 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
181 "psubusw %%mm6, %%mm3 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
182 "psraw $3, %%mm0 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
183 "psraw $3, %%mm1 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
184 "psraw $3, %%mm2 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
185 "psraw $3, %%mm3 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
186 \
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
187 "movq %%mm0, %%mm7 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
188 "punpcklwd %%mm2, %%mm0 \n\t" /*A*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
189 "punpckhwd %%mm2, %%mm7 \n\t" /*C*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
190 "movq %%mm1, %%mm2 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
191 "punpcklwd %%mm3, %%mm1 \n\t" /*B*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
192 "punpckhwd %%mm3, %%mm2 \n\t" /*D*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
193 "movq %%mm0, %%mm3 \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
194 "punpcklwd %%mm1, %%mm0 \n\t" /*A*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
195 "punpckhwd %%mm7, %%mm3 \n\t" /*C*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
196 "punpcklwd %%mm1, %%mm7 \n\t" /*B*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
197 "punpckhwd %%mm2, %%mm1 \n\t" /*D*/\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
198 \
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
199 "movq %%mm0, " #dst0 " \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
200 "movq %%mm7, " #dst1 " \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
201 "movq %%mm3, " #dst2 " \n\t"\
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
202 "movq %%mm1, " #dst3 " \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
203
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
204 "movd %2, %%mm4 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
205 "movd %3, %%mm5 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
206 "movd %4, %%mm6 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
207 "packssdw %%mm4, %%mm4 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
208 "packssdw %%mm5, %%mm5 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
209 "packssdw %%mm6, %%mm6 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
210 "packssdw %%mm4, %%mm4 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
211 "packssdw %%mm5, %%mm5 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
212 "packssdw %%mm6, %%mm6 \n\t"
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
213 REQUANT_CORE( (%1), 8(%1), 16(%1), 24(%1), (%0), 8(%0), 64(%0), 72(%0))
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
214 REQUANT_CORE(32(%1), 40(%1), 48(%1), 56(%1),16(%0),24(%0), 48(%0), 56(%0))
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
215 REQUANT_CORE(64(%1), 72(%1), 80(%1), 88(%1),32(%0),40(%0), 96(%0),104(%0))
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
216 REQUANT_CORE(96(%1),104(%1),112(%1),120(%1),80(%0),88(%0),112(%0),120(%0))
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
217 : : "r" (src), "r" (dst), "g" (threshold1+1), "g" (threshold1+5), "g" (threshold1-4) //FIXME maybe more accurate then needed?
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
218 );
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
219 dst[0]= (src[0] + 4)>>3;
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
220 }
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
221 #endif
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
222
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
223 static inline void add_block(int16_t *dst, int stride, DCTELEM block[64]){
11280
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
224 int y;
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
225
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
226 for(y=0; y<8; y++){
11280
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
227 *(uint32_t*)&dst[0 + y*stride]+= *(uint32_t*)&block[0 + y*8];
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
228 *(uint32_t*)&dst[2 + y*stride]+= *(uint32_t*)&block[2 + y*8];
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
229 *(uint32_t*)&dst[4 + y*stride]+= *(uint32_t*)&block[4 + y*8];
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
230 *(uint32_t*)&dst[6 + y*stride]+= *(uint32_t*)&block[6 + y*8];
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
231 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
232 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
233
11305
michael
parents: 11304
diff changeset
234 static void store_slice_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale){
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
235 int y, x;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
236
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
237 #define STORE(pos) \
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
238 temp= ((src[x + y*src_stride + pos]<<log2_scale) + d[pos])>>6;\
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
239 if(temp & 0x100) temp= ~(temp>>31);\
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
240 dst[x + y*dst_stride + pos]= temp;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
241
11305
michael
parents: 11304
diff changeset
242 for(y=0; y<height; y++){
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
243 uint8_t *d= dither[y];
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
244 for(x=0; x<width; x+=8){
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
245 int temp;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
246 STORE(0);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
247 STORE(1);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
248 STORE(2);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
249 STORE(3);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
250 STORE(4);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
251 STORE(5);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
252 STORE(6);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
253 STORE(7);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
254 }
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
255 }
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
256 }
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
257
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
258 #ifdef HAVE_MMX
11305
michael
parents: 11304
diff changeset
259 static void store_slice_mmx(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale){
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
260 int y;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
261
11305
michael
parents: 11304
diff changeset
262 for(y=0; y<height; y++){
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
263 uint8_t *dst1= dst;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
264 int16_t *src1= src;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
265 asm volatile(
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
266 "movq (%3), %%mm3 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
267 "movq (%3), %%mm4 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
268 "movd %4, %%mm2 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
269 "pxor %%mm0, %%mm0 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
270 "punpcklbw %%mm0, %%mm3 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
271 "punpckhbw %%mm0, %%mm4 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
272 "psraw %%mm2, %%mm3 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
273 "psraw %%mm2, %%mm4 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
274 "movd %5, %%mm2 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
275 "1: \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
276 "movq (%0), %%mm0 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
277 "movq 8(%0), %%mm1 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
278 "paddw %%mm3, %%mm0 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
279 "paddw %%mm4, %%mm1 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
280 "psraw %%mm2, %%mm0 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
281 "psraw %%mm2, %%mm1 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
282 "packuswb %%mm1, %%mm0 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
283 "movq %%mm0, (%1) \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
284 "addl $16, %0 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
285 "addl $8, %1 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
286 "cmpl %2, %1 \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
287 " jb 1b \n\t"
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
288 : "+r" (src1), "+r"(dst1)
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
289 : "r"(dst + width), "r"(dither[y]), "g"(log2_scale), "g"(6-log2_scale)
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
290 );
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
291 src += src_stride;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
292 dst += dst_stride;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
293 }
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
294 // if(width != mmxw)
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
295 // store_slice_c(dst + mmxw, src + mmxw, dst_stride, src_stride, width - mmxw, log2_scale);
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
296 }
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
297 #endif
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
298
11305
michael
parents: 11304
diff changeset
299 static void (*store_slice)(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale)= store_slice_c;
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
300
11993
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
301 static void (*requantize)(DCTELEM dst[64], DCTELEM src[64], int qp, uint8_t *permutation)= hardthresh_c;
11301
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
302
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
303 static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src, int dst_stride, int src_stride, int width, int height, uint8_t *qp_store, int qp_stride, int is_luma){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
304 int x, y, i;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
305 const int count= 1<<p->log2_count;
11307
michael
parents: 11305
diff changeset
306 const int stride= is_luma ? p->temp_stride : ((width+16+15)&(~15));
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
307 uint64_t block_align[32];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
308 DCTELEM *block = (DCTELEM *)block_align;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
309 DCTELEM *block2= (DCTELEM *)(block_align+16);
11301
a8dfeb7180ff optimize
michael
parents: 11299
diff changeset
310
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
311 for(y=0; y<height; y++){
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
312 int index= 8 + 8*stride + y*stride;
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
313 memcpy(p->src + index, src + y*src_stride, width);
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
314 for(x=0; x<8; x++){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
315 p->src[index - x - 1]= p->src[index + x ];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
316 p->src[index + width + x ]= p->src[index + width - x - 1];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
317 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
318 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
319 for(y=0; y<8; y++){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
320 memcpy(p->src + ( 7-y)*stride, p->src + ( y+8)*stride, stride);
11281
michael
parents: 11280
diff changeset
321 memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride);
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
322 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
323 //FIXME (try edge emu)
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
324
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
325 for(y=0; y<height+8; y+=8){
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
326 memset(p->temp + (8+y)*stride, 0, 8*stride*sizeof(int16_t));
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
327 for(x=0; x<width+8; x+=8){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
328 const int qps= 3 + is_luma;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
329 int qp;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
330
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
331 if(p->qp)
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
332 qp= p->qp;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
333 else{
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
334 qp= qp_store[ (XMIN(x, width-1)>>qps) + (XMIN(y, height-1)>>qps) * qp_stride];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
335 if(p->mpeg2) qp>>=1;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
336 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
337 for(i=0; i<count; i++){
11298
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
338 const int x1= x + offset[i+count-1][0];
14dbc88adaf0 change offsets (quality 2-4 looks better now)
michael
parents: 11296
diff changeset
339 const int y1= y + offset[i+count-1][1];
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
340 const int index= x1 + y1*stride;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
341 p->dsp.get_pixels(block, p->src + index, stride);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
342 p->dsp.fdct(block);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
343 requantize(block2, block, qp, p->dsp.idct_permutation);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
344 p->dsp.idct(block2);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
345 add_block(p->temp + index, stride, block2);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
346 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
347 }
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
348 if(y)
11305
michael
parents: 11304
diff changeset
349 store_slice(dst + (y-8)*dst_stride, p->temp + 8 + y*stride, dst_stride, stride, width, XMIN(8, height+8-y), 6-p->log2_count);
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
350 }
11295
michael
parents: 11281
diff changeset
351 #if 0
michael
parents: 11281
diff changeset
352 for(y=0; y<height; y++){
michael
parents: 11281
diff changeset
353 for(x=0; x<width; x++){
michael
parents: 11281
diff changeset
354 if((((x>>6) ^ (y>>6)) & 1) == 0)
michael
parents: 11281
diff changeset
355 dst[x + y*dst_stride]= p->src[8 + 8*stride + x + y*stride];
michael
parents: 11281
diff changeset
356 if((x&63) == 0 || (y&63)==0)
michael
parents: 11281
diff changeset
357 dst[x + y*dst_stride] += 128;
michael
parents: 11281
diff changeset
358 }
michael
parents: 11281
diff changeset
359 }
michael
parents: 11281
diff changeset
360 #endif
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
361 //FIXME reorder for better caching
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
362 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
363
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
364 static int config(struct vf_instance_s* vf,
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
365 int width, int height, int d_width, int d_height,
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
366 unsigned int flags, unsigned int outfmt){
11305
michael
parents: 11304
diff changeset
367 int h= (height+16+15)&(~15);
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
368
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
369 vf->priv->temp_stride= (width+16+15)&(~15);
11305
michael
parents: 11304
diff changeset
370 vf->priv->temp= malloc(vf->priv->temp_stride*h*sizeof(int16_t));
michael
parents: 11304
diff changeset
371 vf->priv->src = malloc(vf->priv->temp_stride*h*sizeof(uint8_t));
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
372
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
373 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
374 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
375
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
376 static void get_image(struct vf_instance_s* vf, mp_image_t *mpi){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
377 if(mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
378 if(mpi->imgfmt!=vf->priv->outfmt) return; // colorspace differ
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
379 // ok, we can do pp in-place (or pp disabled):
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
380 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
381 mpi->type, mpi->flags, mpi->w, mpi->h);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
382 mpi->planes[0]=vf->dmpi->planes[0];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
383 mpi->stride[0]=vf->dmpi->stride[0];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
384 mpi->width=vf->dmpi->width;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
385 if(mpi->flags&MP_IMGFLAG_PLANAR){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
386 mpi->planes[1]=vf->dmpi->planes[1];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
387 mpi->planes[2]=vf->dmpi->planes[2];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
388 mpi->stride[1]=vf->dmpi->stride[1];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
389 mpi->stride[2]=vf->dmpi->stride[2];
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
390 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
391 mpi->flags|=MP_IMGFLAG_DIRECT;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
392 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
393
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
394 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
395 mp_image_t *dmpi;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
396
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
397 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
398 // no DR, so get a new image! hope we'll get DR buffer:
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
399 vf->dmpi=vf_get_image(vf->next,vf->priv->outfmt,
11307
michael
parents: 11305
diff changeset
400 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
401 mpi->w,mpi->h);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
402 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
403
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
404 dmpi= vf->dmpi;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
405
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
406 vf->priv->mpeg2= mpi->qscale_type;
11307
michael
parents: 11305
diff changeset
407 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
11308
0c8d12a58a29 skip filter if codec doesnt provide the QP array and user didnt force a QP (fixes diegos segfault)
michael
parents: 11307
diff changeset
408 if(mpi->qscale || vf->priv->qp){
11307
michael
parents: 11305
diff changeset
409 filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0], mpi->w, mpi->h, mpi->qscale, mpi->qstride, 1);
michael
parents: 11305
diff changeset
410 filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, mpi->qscale, mpi->qstride, 0);
michael
parents: 11305
diff changeset
411 filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, mpi->qscale, mpi->qstride, 0);
11477
9785bff83777 memcpy pix instead of black screen if no DR and codec provides no qscale table
michael
parents: 11335
diff changeset
412 }else{
9785bff83777 memcpy pix instead of black screen if no DR and codec provides no qscale table
michael
parents: 11335
diff changeset
413 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]);
9785bff83777 memcpy pix instead of black screen if no DR and codec provides no qscale table
michael
parents: 11335
diff changeset
414 memcpy_pic(dmpi->planes[1], mpi->planes[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[1], mpi->stride[1]);
9785bff83777 memcpy pix instead of black screen if no DR and codec provides no qscale table
michael
parents: 11335
diff changeset
415 memcpy_pic(dmpi->planes[2], mpi->planes[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[2], mpi->stride[2]);
11308
0c8d12a58a29 skip filter if codec doesnt provide the QP array and user didnt force a QP (fixes diegos segfault)
michael
parents: 11307
diff changeset
416 }
11307
michael
parents: 11305
diff changeset
417 }
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
418 vf_clone_mpi_attributes(dmpi, mpi);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
419
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
420 #ifdef HAVE_MMX
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
421 if(gCpuCaps.hasMMX) asm volatile ("emms\n\t");
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
422 #endif
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
423 #ifdef HAVE_MMX2
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
424 if(gCpuCaps.hasMMX2) asm volatile ("sfence\n\t");
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
425 #endif
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
426
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
427 return vf_next_put_image(vf,dmpi);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
428 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
429
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
430 static void uninit(struct vf_instance_s* vf){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
431 if(!vf->priv) return;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
432
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
433 if(vf->priv->temp) free(vf->priv->temp);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
434 vf->priv->temp= NULL;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
435 if(vf->priv->src) free(vf->priv->src);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
436 vf->priv->src= NULL;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
437 if(vf->priv->avctx) free(vf->priv->avctx);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
438 vf->priv->avctx= NULL;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
439
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
440 free(vf->priv);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
441 vf->priv=NULL;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
442 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
443
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
444 //===========================================================================//
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
445 static int query_format(struct vf_instance_s* vf, unsigned int fmt){
11307
michael
parents: 11305
diff changeset
446 switch(fmt){
michael
parents: 11305
diff changeset
447 case IMGFMT_YVU9:
michael
parents: 11305
diff changeset
448 case IMGFMT_IF09:
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
449 case IMGFMT_YV12:
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
450 case IMGFMT_I420:
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
451 case IMGFMT_IYUV:
11307
michael
parents: 11305
diff changeset
452 case IMGFMT_CLPL:
michael
parents: 11305
diff changeset
453 case IMGFMT_Y800:
michael
parents: 11305
diff changeset
454 case IMGFMT_Y8:
michael
parents: 11305
diff changeset
455 case IMGFMT_NV12:
michael
parents: 11305
diff changeset
456 case IMGFMT_444P:
michael
parents: 11305
diff changeset
457 case IMGFMT_422P:
michael
parents: 11305
diff changeset
458 case IMGFMT_411P:
michael
parents: 11305
diff changeset
459 return vf_next_query_format(vf,fmt);
michael
parents: 11305
diff changeset
460 }
michael
parents: 11305
diff changeset
461 return 0;
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
462 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
463
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
464 static unsigned int fmt_list[]={
11307
michael
parents: 11305
diff changeset
465 IMGFMT_YVU9,
michael
parents: 11305
diff changeset
466 IMGFMT_IF09,
michael
parents: 11305
diff changeset
467 IMGFMT_YV12,
michael
parents: 11305
diff changeset
468 IMGFMT_I420,
michael
parents: 11305
diff changeset
469 IMGFMT_IYUV,
michael
parents: 11305
diff changeset
470 IMGFMT_CLPL,
michael
parents: 11305
diff changeset
471 IMGFMT_Y800,
michael
parents: 11305
diff changeset
472 IMGFMT_Y8,
michael
parents: 11305
diff changeset
473 IMGFMT_NV12,
michael
parents: 11305
diff changeset
474 IMGFMT_444P,
michael
parents: 11305
diff changeset
475 IMGFMT_422P,
michael
parents: 11305
diff changeset
476 IMGFMT_411P,
michael
parents: 11305
diff changeset
477 0
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
478 };
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
479
11307
michael
parents: 11305
diff changeset
480 static int control(struct vf_instance_s* vf, int request, void* data){
michael
parents: 11305
diff changeset
481 switch(request){
michael
parents: 11305
diff changeset
482 case VFCTRL_QUERY_MAX_PP_LEVEL:
michael
parents: 11305
diff changeset
483 return 6;
michael
parents: 11305
diff changeset
484 case VFCTRL_SET_PP_LEVEL:
michael
parents: 11305
diff changeset
485 vf->priv->log2_count= *((unsigned int*)data);
michael
parents: 11305
diff changeset
486 return CONTROL_TRUE;
michael
parents: 11305
diff changeset
487 }
michael
parents: 11305
diff changeset
488 return vf_next_control(vf,request,data);
michael
parents: 11305
diff changeset
489 }
michael
parents: 11305
diff changeset
490
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
491 static int open(vf_instance_t *vf, char* args){
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
492 vf->config=config;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
493 vf->put_image=put_image;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
494 vf->get_image=get_image;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
495 vf->query_format=query_format;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
496 vf->uninit=uninit;
11307
michael
parents: 11305
diff changeset
497 vf->control= control;
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
498 vf->priv=malloc(sizeof(struct vf_priv_s));
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
499 memset(vf->priv, 0, sizeof(struct vf_priv_s));
11280
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
500
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
501 avcodec_init();
b677102fa650 10l and minor optimization
michael
parents: 11277
diff changeset
502
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
503 vf->priv->avctx= avcodec_alloc_context();
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
504 dsputil_init(&vf->priv->dsp, vf->priv->avctx);
11993
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
505
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
506 vf->priv->log2_count= 3;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
507
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
508 if (args) sscanf(args, "%d:%d:%d", &vf->priv->log2_count, &vf->priv->qp, &vf->priv->mode);
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
509 switch(vf->priv->mode){
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
510 case 0: requantize= hardthresh_c; break;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
511 case 1: requantize= softthresh_c; break;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
512 }
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
513
11304
f89e212dd5b9 another 10l ...
michael
parents: 11301
diff changeset
514 #ifdef HAVE_MMX
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
515 if(gCpuCaps.hasMMX){
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
516 store_slice= store_slice_mmx;
11993
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
517 switch(vf->priv->mode){
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
518 case 0: requantize= hardthresh_mmx; break;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
519 //case 1: requantize= softthresh_mmx; break;
4e2d99dbef78 spp soft thresholding patch by (James Crowson <jbcrowso at ncsu dot edu>)
michael
parents: 11568
diff changeset
520 }
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
521 }
11304
f89e212dd5b9 another 10l ...
michael
parents: 11301
diff changeset
522 #endif
11299
66bdf2002aab optimizing
michael
parents: 11298
diff changeset
523
11277
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
524 // check csp:
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
525 vf->priv->outfmt=vf_match_csp(&vf->next,fmt_list,IMGFMT_YV12);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
526 if(!vf->priv->outfmt)
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
527 {
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
528 uninit(vf);
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
529 return 0; // no csp match :(
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
530 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
531
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
532 return 1;
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
533 }
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
534
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
535 vf_info_t vf_info_spp = {
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
536 "simple postprocess",
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
537 "spp",
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
538 "Michael Niedermayer",
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
539 "",
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
540 open,
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
541 NULL
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
542 };
fffd200d81a7 simple alternative postprocessing filter
michael
parents:
diff changeset
543
11335
michael
parents: 11334
diff changeset
544 #endif //USE_LIBAVCODEC