annotate libmpcodecs/vf_gradfun.c @ 34522:d40c46fad187

palctrl exists no longer and has been unused a long time, so no need to free it.
author reimar
date Sat, 28 Jan 2012 10:32:49 +0000
parents 8b6727d2d479
children b4ce15212bfc
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
1 /*
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
2 * Copyright (C) 2009 Loren Merritt <lorenm@u.washignton.edu>
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
3 *
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
4 * This file is part of MPlayer.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
5 *
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
6 * MPlayer is free software; you can redistribute it and/or modify
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
7 * it under the terms of the GNU General Public License as published by
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
8 * the Free Software Foundation; either version 2 of the License, or
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
9 * (at your option) any later version.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
10 *
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
11 * MPlayer is distributed in the hope that it will be useful,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
14 * GNU General Public License for more details.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
15 *
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
16 * You should have received a copy of the GNU General Public License along
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
17 * with MPlayer; if not, write to the Free Software Foundation, Inc.,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
19 */
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
20
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
21 /*
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
22 * Debanding algorithm (from gradfun2db by prunedtree):
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
23 * Boxblur.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
24 * Foreach pixel, if it's within threshold of the blurred value, make it closer.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
25 * So now we have a smoothed and higher bitdepth version of all the shallow
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
26 * gradients, while leaving detailed areas untouched.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
27 * Dither it back to 8bit.
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
28 */
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
29
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
30 #include <stdio.h>
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
31 #include <stdlib.h>
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
32 #include <string.h>
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
33 #include <inttypes.h>
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
34
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
35 #include "config.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
36 #include "cpudetect.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
37 #include "img_format.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
38 #include "mp_image.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
39 #include "vf.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
40 #include "libvo/fastmemcpy.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
41 #include "libavutil/avutil.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
42 #include "libavutil/x86_cpu.h"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
43
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
44 struct vf_priv_s {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
45 int thresh;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
46 int radius;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
47 uint16_t *buf;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
48 void (*filter_line)(uint8_t *dst, uint8_t *src, uint16_t *dc,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
49 int width, int thresh, const uint16_t *dithers);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
50 void (*blur_line)(uint16_t *dc, uint16_t *buf, uint16_t *buf1,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
51 uint8_t *src, int sstride, int width);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
52 };
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
53
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
54 static const uint16_t __attribute__((aligned(16))) pw_7f[8] = {127,127,127,127,127,127,127,127};
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
55 static const uint16_t __attribute__((aligned(16))) pw_ff[8] = {255,255,255,255,255,255,255,255};
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
56 static const uint16_t __attribute__((aligned(16))) dither[8][8] = {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
57 { 0, 96, 24,120, 6,102, 30,126 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
58 { 64, 32, 88, 56, 70, 38, 94, 62 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
59 { 16,112, 8,104, 22,118, 14,110 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
60 { 80, 48, 72, 40, 86, 54, 78, 46 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
61 { 4,100, 28,124, 2, 98, 26,122 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
62 { 68, 36, 92, 60, 66, 34, 90, 58 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
63 { 20,116, 12,108, 18,114, 10,106 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
64 { 84, 52, 76, 44, 82, 50, 74, 42 },
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
65 };
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
66
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
67 static void filter_line_c(uint8_t *dst, uint8_t *src, uint16_t *dc,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
68 int width, int thresh, const uint16_t *dithers)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
69 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
70 int x;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
71 for (x=0; x<width; x++, dc+=x&1) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
72 int pix = src[x]<<7;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
73 int delta = dc[0] - pix;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
74 int m = abs(delta) * thresh >> 16;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
75 m = FFMAX(0, 127-m);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
76 m = m*m*delta >> 14;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
77 pix += m + dithers[x&7];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
78 dst[x] = av_clip_uint8(pix>>7);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
79 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
80 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
81
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
82 static void blur_line_c(uint16_t *dc, uint16_t *buf, uint16_t *buf1,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
83 uint8_t *src, int sstride, int width)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
84 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
85 int x, v, old;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
86 for (x=0; x<width; x++) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
87 v = buf1[x] + src[2*x] + src[2*x+1] + src[2*x+sstride] + src[2*x+1+sstride];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
88 old = buf[x];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
89 buf[x] = v;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
90 dc[x] = v - old;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
91 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
92 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
93
30976
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
94 #if HAVE_MMX2
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
95 static void filter_line_mmx2(uint8_t *dst, uint8_t *src, uint16_t *dc,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
96 int width, int thresh, const uint16_t *dithers)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
97 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
98 intptr_t x;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
99 if (width&3) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
100 x = width&~3;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
101 filter_line_c(dst+x, src+x, dc+x/2, width-x, thresh, dithers);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
102 width = x;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
103 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
104 x = -width;
30920
570e001f2e53 Convert asm keyword to __asm__.
diego
parents: 30642
diff changeset
105 __asm__ volatile(
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
106 "movd %4, %%mm5 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
107 "pxor %%mm7, %%mm7 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
108 "pshufw $0, %%mm5, %%mm5 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
109 "movq %6, %%mm6 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
110 "movq %5, %%mm4 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
111 "1: \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
112 "movd (%2,%0), %%mm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
113 "movd (%3,%0), %%mm1 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
114 "punpcklbw %%mm7, %%mm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
115 "punpcklwd %%mm1, %%mm1 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
116 "psllw $7, %%mm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
117 "pxor %%mm2, %%mm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
118 "psubw %%mm0, %%mm1 \n" // delta = dc - pix
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
119 "psubw %%mm1, %%mm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
120 "pmaxsw %%mm1, %%mm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
121 "pmulhuw %%mm5, %%mm2 \n" // m = abs(delta) * thresh >> 16
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
122 "psubw %%mm6, %%mm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
123 "pminsw %%mm7, %%mm2 \n" // m = -max(0, 127-m)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
124 "pmullw %%mm2, %%mm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
125 "paddw %%mm4, %%mm0 \n" // pix += dither
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
126 "pmulhw %%mm2, %%mm1 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
127 "psllw $2, %%mm1 \n" // m = m*m*delta >> 14
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
128 "paddw %%mm1, %%mm0 \n" // pix += m
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
129 "psraw $7, %%mm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
130 "packuswb %%mm0, %%mm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
131 "movd %%mm0, (%1,%0) \n" // dst = clip(pix>>7)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
132 "add $4, %0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
133 "jl 1b \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
134 "emms \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
135 :"+r"(x)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
136 :"r"(dst+width), "r"(src+width), "r"(dc+width/2),
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
137 "rm"(thresh), "m"(*dithers), "m"(*pw_7f)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
138 :"memory"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
139 );
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
140 }
30976
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
141 #endif
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
142
30976
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
143 #if HAVE_SSSE3
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
144 static void filter_line_ssse3(uint8_t *dst, uint8_t *src, uint16_t *dc,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
145 int width, int thresh, const uint16_t *dithers)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
146 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
147 intptr_t x;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
148 if (width&7) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
149 // could be 10% faster if I somehow eliminated this
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
150 x = width&~7;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
151 filter_line_c(dst+x, src+x, dc+x/2, width-x, thresh, dithers);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
152 width = x;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
153 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
154 x = -width;
30920
570e001f2e53 Convert asm keyword to __asm__.
diego
parents: 30642
diff changeset
155 __asm__ volatile(
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
156 "movd %4, %%xmm5 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
157 "pxor %%xmm7, %%xmm7 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
158 "pshuflw $0,%%xmm5, %%xmm5 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
159 "movdqa %6, %%xmm6 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
160 "punpcklqdq %%xmm5, %%xmm5 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
161 "movdqa %5, %%xmm4 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
162 "1: \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
163 "movq (%2,%0), %%xmm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
164 "movq (%3,%0), %%xmm1 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
165 "punpcklbw %%xmm7, %%xmm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
166 "punpcklwd %%xmm1, %%xmm1 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
167 "psllw $7, %%xmm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
168 "psubw %%xmm0, %%xmm1 \n" // delta = dc - pix
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
169 "pabsw %%xmm1, %%xmm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
170 "pmulhuw %%xmm5, %%xmm2 \n" // m = abs(delta) * thresh >> 16
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
171 "psubw %%xmm6, %%xmm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
172 "pminsw %%xmm7, %%xmm2 \n" // m = -max(0, 127-m)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
173 "pmullw %%xmm2, %%xmm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
174 "psllw $1, %%xmm2 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
175 "paddw %%xmm4, %%xmm0 \n" // pix += dither
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
176 "pmulhrsw %%xmm2, %%xmm1 \n" // m = m*m*delta >> 14
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
177 "paddw %%xmm1, %%xmm0 \n" // pix += m
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
178 "psraw $7, %%xmm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
179 "packuswb %%xmm0, %%xmm0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
180 "movq %%xmm0, (%1,%0) \n" // dst = clip(pix>>7)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
181 "add $8, %0 \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
182 "jl 1b \n"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
183 :"+&r"(x)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
184 :"r"(dst+width), "r"(src+width), "r"(dc+width/2),
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
185 "rm"(thresh), "m"(*dithers), "m"(*pw_7f)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
186 :"memory"
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
187 );
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
188 }
30976
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
189 #endif // HAVE_SSSE3
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
190
30983
8b6727d2d479 cosmetics: Reorder some x86-related preprocessor conditionals.
diego
parents: 30976
diff changeset
191 #if HAVE_SSE2 && HAVE_6REGS
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
192 #define BLURV(load)\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
193 intptr_t x = -2*width;\
30920
570e001f2e53 Convert asm keyword to __asm__.
diego
parents: 30642
diff changeset
194 __asm__ volatile(\
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
195 "movdqa %6, %%xmm7 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
196 "1: \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
197 load" (%4,%0), %%xmm0 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
198 load" (%5,%0), %%xmm1 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
199 "movdqa %%xmm0, %%xmm2 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
200 "movdqa %%xmm1, %%xmm3 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
201 "psrlw $8, %%xmm0 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
202 "psrlw $8, %%xmm1 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
203 "pand %%xmm7, %%xmm2 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
204 "pand %%xmm7, %%xmm3 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
205 "paddw %%xmm1, %%xmm0 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
206 "paddw %%xmm3, %%xmm2 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
207 "paddw %%xmm2, %%xmm0 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
208 "paddw (%2,%0), %%xmm0 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
209 "movdqa (%1,%0), %%xmm1 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
210 "movdqa %%xmm0, (%1,%0) \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
211 "psubw %%xmm1, %%xmm0 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
212 "movdqa %%xmm0, (%3,%0) \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
213 "add $16, %0 \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
214 "jl 1b \n"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
215 :"+&r"(x)\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
216 :"r"(buf+width),\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
217 "r"(buf1+width),\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
218 "r"(dc+width),\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
219 "r"(src+width*2),\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
220 "r"(src+width*2+sstride),\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
221 "m"(*pw_ff)\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
222 :"memory"\
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
223 );
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
224
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
225 static void blur_line_sse2(uint16_t *dc, uint16_t *buf, uint16_t *buf1,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
226 uint8_t *src, int sstride, int width)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
227 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
228 if (((intptr_t)src|sstride)&15) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
229 BLURV("movdqu");
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
230 } else {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
231 BLURV("movdqa");
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
232 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
233 }
30976
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
234 #endif // HAVE_6REGS && HAVE_SSE2
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
235
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
236 static void filter(struct vf_priv_s *ctx, uint8_t *dst, uint8_t *src,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
237 int width, int height, int dstride, int sstride, int r)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
238 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
239 int bstride = ((width+15)&~15)/2;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
240 int y;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
241 uint32_t dc_factor = (1<<21)/(r*r);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
242 uint16_t *dc = ctx->buf+16;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
243 uint16_t *buf = ctx->buf+bstride+32;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
244 int thresh = ctx->thresh;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
245
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
246 memset(dc, 0, (bstride+16)*sizeof(*buf));
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
247 for (y=0; y<r; y++)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
248 ctx->blur_line(dc, buf+y*bstride, buf+(y-1)*bstride, src+2*y*sstride, sstride, width/2);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
249 for (;;) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
250 if (y < height-r) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
251 int mod = ((y+r)/2)%r;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
252 uint16_t *buf0 = buf+mod*bstride;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
253 uint16_t *buf1 = buf+(mod?mod-1:r-1)*bstride;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
254 int x, v;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
255 ctx->blur_line(dc, buf0, buf1, src+(y+r)*sstride, sstride, width/2);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
256 for (x=v=0; x<r; x++)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
257 v += dc[x];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
258 for (; x<width/2; x++) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
259 v += dc[x] - dc[x-r];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
260 dc[x-r] = v * dc_factor >> 16;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
261 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
262 for (; x<(width+r+1)/2; x++)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
263 dc[x-r] = v * dc_factor >> 16;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
264 for (x=-r/2; x<0; x++)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
265 dc[x] = dc[0];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
266 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
267 if (y == r) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
268 for (y=0; y<r; y++)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
269 ctx->filter_line(dst+y*dstride, src+y*sstride, dc-r/2, width, thresh, dither[y&7]);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
270 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
271 ctx->filter_line(dst+y*dstride, src+y*sstride, dc-r/2, width, thresh, dither[y&7]);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
272 if (++y >= height) break;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
273 ctx->filter_line(dst+y*dstride, src+y*sstride, dc-r/2, width, thresh, dither[y&7]);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
274 if (++y >= height) break;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
275 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
276 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
277
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
278 static void get_image(struct vf_instance *vf, mp_image_t *mpi)
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
279 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
280 if (mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
281 // ok, we can do pp in-place:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
282 vf->dmpi = vf_get_image(vf->next, mpi->imgfmt,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
283 mpi->type, mpi->flags, mpi->width, mpi->height);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
284 mpi->planes[0] = vf->dmpi->planes[0];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
285 mpi->stride[0] = vf->dmpi->stride[0];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
286 mpi->width = vf->dmpi->width;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
287 if (mpi->flags&MP_IMGFLAG_PLANAR){
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
288 mpi->planes[1] = vf->dmpi->planes[1];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
289 mpi->planes[2] = vf->dmpi->planes[2];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
290 mpi->stride[1] = vf->dmpi->stride[1];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
291 mpi->stride[2] = vf->dmpi->stride[2];
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
292 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
293 mpi->flags |= MP_IMGFLAG_DIRECT;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
294 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
295
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
296 static int put_image(struct vf_instance *vf, mp_image_t *mpi, double pts)
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
297 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
298 mp_image_t *dmpi = vf->dmpi;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
299 int p;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
300
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
301 if (!(mpi->flags&MP_IMGFLAG_DIRECT)) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
302 // no DR, so get a new image. hope we'll get DR buffer:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
303 dmpi = vf_get_image(vf->next,mpi->imgfmt, MP_IMGTYPE_TEMP,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
304 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
305 mpi->w, mpi->h);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
306 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
307 vf_clone_mpi_attributes(dmpi, mpi);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
308
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
309 for (p=0; p<mpi->num_planes; p++) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
310 int w = mpi->w;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
311 int h = mpi->h;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
312 int r = vf->priv->radius;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
313 if (p) {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
314 w >>= mpi->chroma_x_shift;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
315 h >>= mpi->chroma_y_shift;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
316 r = ((r>>mpi->chroma_x_shift) + (r>>mpi->chroma_y_shift)) / 2;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
317 r = av_clip((r+1)&~1,4,32);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
318 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
319 if (FFMIN(w,h) > 2*r)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
320 filter(vf->priv, dmpi->planes[p], mpi->planes[p], w, h,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
321 dmpi->stride[p], mpi->stride[p], r);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
322 else if (dmpi->planes[p] != mpi->planes[p])
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
323 memcpy_pic(dmpi->planes[p], mpi->planes[p], w, h,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
324 dmpi->stride[p], mpi->stride[p]);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
325 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
326
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
327 return vf_next_put_image(vf, dmpi, pts);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
328 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
329
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
330 static int query_format(struct vf_instance *vf, unsigned int fmt)
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
331 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
332 switch (fmt){
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
333 case IMGFMT_YVU9:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
334 case IMGFMT_IF09:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
335 case IMGFMT_YV12:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
336 case IMGFMT_I420:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
337 case IMGFMT_IYUV:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
338 case IMGFMT_CLPL:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
339 case IMGFMT_Y800:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
340 case IMGFMT_Y8:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
341 case IMGFMT_NV12:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
342 case IMGFMT_NV21:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
343 case IMGFMT_444P:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
344 case IMGFMT_422P:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
345 case IMGFMT_411P:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
346 case IMGFMT_HM12:
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
347 return vf_next_query_format(vf,fmt);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
348 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
349 return 0;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
350 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
351
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
352 static int config(struct vf_instance *vf,
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
353 int width, int height, int d_width, int d_height,
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
354 unsigned int flags, unsigned int outfmt)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
355 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
356 free(vf->priv->buf);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
357 vf->priv->buf = av_mallocz((((width+15)&~15)*(vf->priv->radius+1)/2+32)*sizeof(uint16_t));
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
358 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
359 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
360
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
361 static void uninit(struct vf_instance *vf)
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
362 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
363 if (!vf->priv) return;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
364 av_free(vf->priv->buf);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
365 free(vf->priv);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
366 vf->priv = NULL;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
367 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
368
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 29371
diff changeset
369 static int vf_open(vf_instance_t *vf, char *args)
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
370 {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
371 float thresh = 1.2;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
372 int radius = 16;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
373
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
374 vf->get_image=get_image;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
375 vf->put_image=put_image;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
376 vf->query_format=query_format;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
377 vf->config=config;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
378 vf->uninit=uninit;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
379 vf->priv=malloc(sizeof(struct vf_priv_s));
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
380 memset(vf->priv, 0, sizeof(struct vf_priv_s));
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
381
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
382 if (args) sscanf(args, "%f:%d", &thresh, &radius);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
383 vf->priv->thresh = (1<<15)/av_clipf(thresh,0.51,255);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
384 vf->priv->radius = av_clip((radius+1)&~1,4,32);
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
385
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
386 vf->priv->blur_line = blur_line_c;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
387 vf->priv->filter_line = filter_line_c;
30983
8b6727d2d479 cosmetics: Reorder some x86-related preprocessor conditionals.
diego
parents: 30976
diff changeset
388 #if HAVE_SSE2 && HAVE_6REGS
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
389 if (gCpuCaps.hasSSE2)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
390 vf->priv->blur_line = blur_line_sse2;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
391 #endif
30976
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
392 #if HAVE_MMX2
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
393 if (gCpuCaps.hasMMX2)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
394 vf->priv->filter_line = filter_line_mmx2;
30976
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
395 #endif
6955998c187e Change ifdefs to make more sense: HAVE_SSSE3 should only be around SSSE3-code,
reimar
parents: 30920
diff changeset
396 #if HAVE_SSSE3
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
397 if (gCpuCaps.hasSSSE3)
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
398 vf->priv->filter_line = filter_line_ssse3;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
399 #endif
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
400
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
401 return 1;
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
402 }
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
403
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
404 const vf_info_t vf_info_gradfun = {
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
405 "gradient deband",
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
406 "gradfun",
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
407 "Loren Merritt",
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
408 "",
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 29371
diff changeset
409 vf_open,
29371
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
410 NULL
3fa15eca924e new debanding filter
lorenm
parents:
diff changeset
411 };