Add support for VDPAU video out, including hardware decoding.
[mplayer/glamo.git] / libmpcodecs / vf_decimate.c
blobe1cea34ae395ed91ce1577cac1158e1b6759a264
1 #include <stdio.h>
2 #include <stdlib.h>
3 #include <string.h>
5 #include "config.h"
6 #include "mp_msg.h"
7 #include "cpudetect.h"
9 #include "img_format.h"
10 #include "mp_image.h"
11 #include "vf.h"
13 #include "libvo/fastmemcpy.h"
16 struct vf_priv_s {
17 int hi, lo;
18 float frac;
19 int max, last, cnt;
22 #if HAVE_MMX
23 static int diff_MMX(unsigned char *old, unsigned char *new, int os, int ns)
25 volatile short out[4];
26 __asm__ (
27 "movl $8, %%ecx \n\t"
28 "pxor %%mm4, %%mm4 \n\t"
29 "pxor %%mm7, %%mm7 \n\t"
31 ASMALIGN(4)
32 "1: \n\t"
34 "movq (%%"REG_S"), %%mm0 \n\t"
35 "movq (%%"REG_S"), %%mm2 \n\t"
36 "add %%"REG_a", %%"REG_S" \n\t"
37 "movq (%%"REG_D"), %%mm1 \n\t"
38 "add %%"REG_b", %%"REG_D" \n\t"
39 "psubusb %%mm1, %%mm2 \n\t"
40 "psubusb %%mm0, %%mm1 \n\t"
41 "movq %%mm2, %%mm0 \n\t"
42 "movq %%mm1, %%mm3 \n\t"
43 "punpcklbw %%mm7, %%mm0 \n\t"
44 "punpcklbw %%mm7, %%mm1 \n\t"
45 "punpckhbw %%mm7, %%mm2 \n\t"
46 "punpckhbw %%mm7, %%mm3 \n\t"
47 "paddw %%mm0, %%mm4 \n\t"
48 "paddw %%mm1, %%mm4 \n\t"
49 "paddw %%mm2, %%mm4 \n\t"
50 "paddw %%mm3, %%mm4 \n\t"
52 "decl %%ecx \n\t"
53 "jnz 1b \n\t"
54 "movq %%mm4, (%%"REG_d") \n\t"
55 "emms \n\t"
57 : "S" (old), "D" (new), "a" ((long)os), "b" ((long)ns), "d" (out)
58 : "memory"
60 return out[0]+out[1]+out[2]+out[3];
62 #endif
64 static int diff_C(unsigned char *old, unsigned char *new, int os, int ns)
66 int x, y, d=0;
67 for (y = 8; y; y--) {
68 for (x = 8; x; x--) {
69 d += abs(new[x] - old[x]);
71 new += ns;
72 old += os;
74 return d;
77 static int (*diff)(unsigned char *, unsigned char *, int, int);
79 static int diff_to_drop_plane(int hi, int lo, float frac, unsigned char *old, unsigned char *new, int w, int h, int os, int ns)
81 int x, y;
82 int d, c=0;
83 int t = (w/16)*(h/16)*frac;
84 for (y = 0; y < h-7; y += 4) {
85 for (x = 8; x < w-7; x += 4) {
86 d = diff(old+x+y*os, new+x+y*ns, os, ns);
87 if (d > hi) return 0;
88 if (d > lo) {
89 c++;
90 if (c > t) return 0;
94 return 1;
97 static int diff_to_drop(int hi, int lo, float frac, mp_image_t *old, mp_image_t *new)
99 if (new->flags & MP_IMGFLAG_PLANAR) {
100 return diff_to_drop_plane(hi,lo,frac, old->planes[0], new->planes[0],
101 new->w, new->h, old->stride[0], new->stride[0])
102 && diff_to_drop_plane(hi,lo,frac, old->planes[1], new->planes[1],
103 new->chroma_width, new->chroma_height,
104 old->stride[1], new->stride[1])
105 && diff_to_drop_plane(hi,lo,frac, old->planes[2], new->planes[2],
106 new->chroma_width, new->chroma_height,
107 old->stride[2], new->stride[2]);
109 return diff_to_drop_plane(hi,lo,frac, old->planes[0], new->planes[0],
110 new->w*(new->bpp/8), new->h, old->stride[0], new->stride[0]);
113 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts)
115 mp_image_t *dmpi;
117 dmpi = vf_get_image(vf->next, mpi->imgfmt,
118 MP_IMGTYPE_STATIC, MP_IMGFLAG_ACCEPT_STRIDE |
119 MP_IMGFLAG_PRESERVE | MP_IMGFLAG_READABLE,
120 mpi->width, mpi->height);
121 dmpi->qscale = mpi->qscale;
122 dmpi->qstride = mpi->qstride;
123 dmpi->qscale_type = mpi->qscale_type;
125 if (diff_to_drop(vf->priv->hi, vf->priv->lo, vf->priv->frac, dmpi, mpi)) {
126 if (vf->priv->max == 0)
127 return 0;
128 else if ((vf->priv->max > 0) && (vf->priv->cnt++ < vf->priv->max))
129 return 0;
130 else if ((vf->priv->max < 0) && (vf->priv->last+1 >= -vf->priv->max))
131 return vf->priv->last=0;
133 vf->priv->last++;
134 vf->priv->cnt=0;
136 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h,
137 dmpi->stride[0], mpi->stride[0]);
138 if (mpi->flags & MP_IMGFLAG_PLANAR) {
139 memcpy_pic(dmpi->planes[1], mpi->planes[1],
140 mpi->chroma_width, mpi->chroma_height,
141 dmpi->stride[1], mpi->stride[1]);
142 memcpy_pic(dmpi->planes[2], mpi->planes[2],
143 mpi->chroma_width, mpi->chroma_height,
144 dmpi->stride[2], mpi->stride[2]);
146 return vf_next_put_image(vf, dmpi, pts);
149 static void uninit(struct vf_instance_s* vf)
151 free(vf->priv);
154 static int open(vf_instance_t *vf, char* args)
156 struct vf_priv_s *p;
157 vf->put_image = put_image;
158 vf->uninit = uninit;
159 vf->default_reqs = VFCAP_ACCEPT_STRIDE;
160 vf->priv = p = calloc(1, sizeof(struct vf_priv_s));
161 p->max = 0;
162 p->hi = 64*12;
163 p->lo = 64*5;
164 p->frac = 0.33;
165 if (args) sscanf(args, "%d:%d:%d:%f", &p->max, &p->hi, &p->lo, &p->frac);
166 diff = diff_C;
167 #if HAVE_MMX
168 if(gCpuCaps.hasMMX) diff = diff_MMX;
169 #endif
170 return 1;
173 const vf_info_t vf_info_decimate = {
174 "near-duplicate frame remover",
175 "decimate",
176 "Rich Felker",
178 open,
179 NULL