Commit acbac567 authored by Michael Niedermayer's avatar Michael Niedermayer

yadif filter, based on stefanos port of my yadif from mplayer.

Compared to stefanos, 2 frame output works with ffplay.

Originally committed as revision 25196 to svn://svn.ffmpeg.org/ffmpeg/trunk
parent b299c4e4
......@@ -1402,6 +1402,7 @@ udp_protocol_deps="network"
# filters
ocv_smooth_filter_deps="libopencv"
yadif_filter_deps="gpl"
# libraries
avdevice_deps="avcodec avformat"
......
......@@ -448,6 +448,36 @@ Flip the input video vertically.
./ffmpeg -i in.avi -vf "vflip" out.avi
@end example
@section yadif
yadif is "yet another deinterlacing filter".
It accepts the syntax:
@example
yadif=[@var{mode}[:@var{parity}]]
@end example
@table @option
@item mode
Specify the interlacing mode to adopt, accepts one of the following values.
0: Output 1 frame for each frame.
1: Output 1 frame for each field.
2: Like 0 but skips spatial interlacing check.
3: Like 1 but skips spatial interlacing check.
Default value is 0.
@item parity
0 if is bottom field first, 1 if the interlaced video is top field
first, -1 to enable automatic detection.
@end table
@c man end VIDEO FILTERS
@chapter Video Sources
......
......@@ -36,6 +36,9 @@ OBJS-$(CONFIG_SCALE_FILTER) += vf_scale.o
OBJS-$(CONFIG_SLICIFY_FILTER) += vf_slicify.o
OBJS-$(CONFIG_UNSHARP_FILTER) += vf_unsharp.o
OBJS-$(CONFIG_VFLIP_FILTER) += vf_vflip.o
OBJS-$(CONFIG_YADIF_FILTER) += vf_yadif.o
OBJS-$(ARCH_X86) += x86/yadif.o
OBJS-$(CONFIG_BUFFER_FILTER) += vsrc_buffer.o
OBJS-$(CONFIG_COLOR_FILTER) += vf_pad.o
......
......@@ -56,6 +56,7 @@ void avfilter_register_all(void)
REGISTER_FILTER (SLICIFY, slicify, vf);
REGISTER_FILTER (UNSHARP, unsharp, vf);
REGISTER_FILTER (VFLIP, vflip, vf);
REGISTER_FILTER (YADIF, yadif, vf);
REGISTER_FILTER (BUFFER, buffer, vsrc);
REGISTER_FILTER (COLOR, color, vsrc);
......
/*
* Copyright (C) 2006-2010 Michael Niedermayer <michaelni@gmx.at>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License along
* with FFmpeg; if not, write to the Free Software Foundation, Inc.,
* 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
*/
#include "libavutil/cpu.h"
#include "libavutil/x86_cpu.h"
#include "libavutil/common.h"
#include "avfilter.h"
#include "yadif.h"
#undef NDEBUG
#include <assert.h>
typedef struct {
/**
* 0: send 1 frame for each frame
* 1: send 1 frame for each field
* 2: like 0 but skips spatial interlacing check
* 3: like 1 but skips spatial interlacing check
*/
int mode;
/**
* 0: bottom field first
* 1: top field first
* -1: auto-detection
*/
int parity;
int frame_pending;
AVFilterBufferRef *cur;
AVFilterBufferRef *next;
AVFilterBufferRef *prev;
AVFilterBufferRef *out;
void (*filter_line)(uint8_t *dst,
uint8_t *prev, uint8_t *cur, uint8_t *next,
int w, int refs, int parity, int mode);
} YADIFContext;
static void filter_line_c(uint8_t *dst,
uint8_t *prev, uint8_t *cur, uint8_t *next,
int w, int refs, int parity, int mode)
{
int x;
uint8_t *prev2 = parity ? prev : cur ;
uint8_t *next2 = parity ? cur : next;
for (x = 0; x < w; x++) {
int c = cur[-refs];
int d = (prev2[0] + next2[0])>>1;
int e = cur[+refs];
int temporal_diff0 = FFABS(prev2[0] - next2[0]);
int temporal_diff1 =(FFABS(prev[-refs] - c) + FFABS(prev[+refs] - e) )>>1;
int temporal_diff2 =(FFABS(next[-refs] - c) + FFABS(next[+refs] - e) )>>1;
int diff = FFMAX3(temporal_diff0>>1, temporal_diff1, temporal_diff2);
int spatial_pred = (c+e)>>1;
int spatial_score = FFABS(cur[-refs-1] - cur[+refs-1]) + FFABS(c-e)
+ FFABS(cur[-refs+1] - cur[+refs+1]) - 1;
#define CHECK(j)\
{ int score = FFABS(cur[-refs-1+j] - cur[+refs-1-j])\
+ FFABS(cur[-refs +j] - cur[+refs -j])\
+ FFABS(cur[-refs+1+j] - cur[+refs+1-j]);\
if (score < spatial_score) {\
spatial_score= score;\
spatial_pred= (cur[-refs +j] + cur[+refs -j])>>1;\
CHECK(-1) CHECK(-2) }} }}
CHECK( 1) CHECK( 2) }} }}
if (mode < 2) {
int b = (prev2[-2*refs] + next2[-2*refs])>>1;
int f = (prev2[+2*refs] + next2[+2*refs])>>1;
#if 0
int a = cur[-3*refs];
int g = cur[+3*refs];
int max = FFMAX3(d-e, d-c, FFMIN3(FFMAX(b-c,f-e),FFMAX(b-c,b-a),FFMAX(f-g,f-e)) );
int min = FFMIN3(d-e, d-c, FFMAX3(FFMIN(b-c,f-e),FFMIN(b-c,b-a),FFMIN(f-g,f-e)) );
#else
int max = FFMAX3(d-e, d-c, FFMIN(b-c, f-e));
int min = FFMIN3(d-e, d-c, FFMAX(b-c, f-e));
#endif
diff = FFMAX3(diff, min, -max);
}
if (spatial_pred > d + diff)
spatial_pred = d + diff;
else if (spatial_pred < d - diff)
spatial_pred = d - diff;
dst[0] = spatial_pred;
dst++;
cur++;
prev++;
next++;
prev2++;
next2++;
}
}
static void filter(AVFilterContext *ctx, AVFilterBufferRef *dstpic,
int parity, int tff)
{
YADIFContext *yadif = ctx->priv;
int y, i;
for (i = 0; i < 3; i++) {
int is_chroma = !!i;
int w = dstpic->video->w >> is_chroma;
int h = dstpic->video->h >> is_chroma;
int refs = yadif->cur->linesize[i];
for (y = 0; y < h; y++) {
if ((y ^ parity) & 1) {
uint8_t *prev = &yadif->prev->data[i][y*refs];
uint8_t *cur = &yadif->cur ->data[i][y*refs];
uint8_t *next = &yadif->next->data[i][y*refs];
uint8_t *dst = &dstpic->data[i][y*dstpic->linesize[i]];
yadif->filter_line(dst, prev, cur, next, w, refs, parity ^ tff, yadif->mode);
} else {
memcpy(&dstpic->data[i][y*dstpic->linesize[i]],
&yadif->cur->data[i][y*refs], w);
}
}
}
#if HAVE_MMX
__asm__ volatile("emms \n\t" : : : "memory");
#endif
}
static AVFilterBufferRef *get_video_buffer(AVFilterLink *link, int perms, int w, int h)
{
AVFilterBufferRef *picref;
int width = FFALIGN(w, 32);
int height= FFALIGN(h+6, 32);
int i;
picref = avfilter_default_get_video_buffer(link, perms, width, height);
picref->video->w = w;
picref->video->h = h;
for (i = 0; i < 3; i++)
picref->data[i] += 3 * picref->linesize[i];
return picref;
}
static void return_frame(AVFilterContext *ctx, int is_second){
YADIFContext *yadif = ctx->priv;
AVFilterLink *link= ctx->outputs[0];
int tff = yadif->parity == -1 ? yadif->cur->video->top_field_first : yadif->parity;
if(is_second)
yadif->out = avfilter_get_video_buffer(link, AV_PERM_WRITE | AV_PERM_PRESERVE |
AV_PERM_REUSE, link->w, link->h);
filter(ctx, yadif->out, tff ^ !is_second, tff);
if(is_second)
avfilter_start_frame(ctx->outputs[0], yadif->out);
avfilter_draw_slice(ctx->outputs[0], 0, link->h, 1);
avfilter_end_frame(ctx->outputs[0]);
yadif->frame_pending= (yadif->mode&1) && !is_second;
}
static void start_frame(AVFilterLink *link, AVFilterBufferRef *picref)
{
AVFilterContext *ctx = link->dst;
YADIFContext *yadif = ctx->priv;
if(yadif->frame_pending)
return_frame(ctx, 1);
if (yadif->prev)
avfilter_unref_buffer(yadif->prev);
yadif->prev = yadif->cur;
yadif->cur = yadif->next;
yadif->next = picref;
if (!yadif->prev)
yadif->prev = avfilter_get_video_buffer(link, AV_PERM_WRITE | AV_PERM_PRESERVE |
AV_PERM_REUSE, link->w, link->h);
if(!yadif->cur)
return;
yadif->out = avfilter_get_video_buffer(ctx->outputs[0], AV_PERM_WRITE | AV_PERM_PRESERVE |
AV_PERM_REUSE, link->w, link->h);
yadif->out->pts = yadif->cur->pts;
avfilter_start_frame(ctx->outputs[0], yadif->out);
}
static void end_frame(AVFilterLink *link)
{
AVFilterContext *ctx = link->dst;
YADIFContext *yadif = ctx->priv;
if(!yadif->out)
return;
return_frame(ctx, 0);
}
static int request_frame(AVFilterLink *link)
{
AVFilterContext *ctx = link->src;
YADIFContext *yadif = ctx->priv;
if(yadif->frame_pending){
return_frame(ctx, 1);
return 0;
}
do{
int ret;
if ((ret = avfilter_request_frame(link->src->inputs[0])))
return ret;
}while(!yadif->cur);
return 0;
}
static int poll_frame(AVFilterLink *link)
{
YADIFContext *yadif = link->src->priv;
int ret, val;
if(yadif->frame_pending)
return 1;
val= avfilter_poll_frame(link->src->inputs[0]);
if(val==1 && !yadif->next){ //FIXME change API to not requre this red tape
if ((ret = avfilter_request_frame(link->src->inputs[0])) < 0)
return ret;
val = avfilter_poll_frame(link->src->inputs[0]);
}
assert(yadif->next);
return val * ((yadif->mode&1)+1);
}
static av_cold void uninit(AVFilterContext *ctx)
{
YADIFContext *yadif = ctx->priv;
if (yadif->prev) avfilter_unref_buffer(yadif->prev);
if (yadif->cur ) avfilter_unref_buffer(yadif->cur );
if (yadif->next) avfilter_unref_buffer(yadif->next);
}
static int query_formats(AVFilterContext *ctx)
{
static const enum PixelFormat pix_fmts[] = {
PIX_FMT_YUV420P,
PIX_FMT_GRAY8,
PIX_FMT_NONE
};
avfilter_set_common_formats(ctx, avfilter_make_format_list(pix_fmts));
return 0;
}
static av_cold int init(AVFilterContext *ctx, const char *args, void *opaque)
{
YADIFContext *yadif = ctx->priv;
av_unused int cpu_flags = av_get_cpu_flags();
yadif->mode = 0;
yadif->parity = -1;
if (args) sscanf(args, "%d:%d", &yadif->mode, &yadif->parity);
yadif->filter_line = filter_line_c;
#if HAVE_MMX
if (cpu_flags & AV_CPU_FLAG_MMX)
yadif->filter_line = ff_yadif_filter_line_mmx;
#endif
av_log(ctx, AV_LOG_INFO, "mode:%d parity:%d\n", yadif->mode, yadif->parity);
return 0;
}
static void null_draw_slice(AVFilterLink *link, int y, int h, int slice_dir) { }
AVFilter avfilter_vf_yadif = {
.name = "yadif",
.description = NULL_IF_CONFIG_SMALL("Deinterlace the input image"),
.priv_size = sizeof(YADIFContext),
.init = init,
.uninit = uninit,
.query_formats = query_formats,
.inputs = (AVFilterPad[]) {{ .name = "default",
.type = AVMEDIA_TYPE_VIDEO,
.start_frame = start_frame,
.get_video_buffer = get_video_buffer,
.draw_slice = null_draw_slice,
.end_frame = end_frame, },
{ .name = NULL}},
.outputs = (AVFilterPad[]) {{ .name = "default",
.type = AVMEDIA_TYPE_VIDEO,
.poll_frame = poll_frame,
.request_frame = request_frame, },
{ .name = NULL}},
};
/*
* Copyright (C) 2006 Michael Niedermayer <michaelni@gmx.at>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License along
* with FFmpeg; if not, write to the Free Software Foundation, Inc.,
* 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
*/
#include "libavutil/cpu.h"
#include "libavutil/x86_cpu.h"
#include "libavfilter/yadif.h"
#if HAVE_MMX
#define LOAD4(mem,dst) \
"movd "mem", "#dst" \n\t"\
"punpcklbw %%mm7, "#dst" \n\t"
#define PABS(tmp,dst) \
"pxor "#tmp", "#tmp" \n\t"\
"psubw "#dst", "#tmp" \n\t"\
"pmaxsw "#tmp", "#dst" \n\t"
#define CHECK(pj,mj) \
"movq "#pj"(%[cur],%[mrefs]), %%mm2 \n\t" /* cur[x-refs-1+j] */\
"movq "#mj"(%[cur],%[prefs]), %%mm3 \n\t" /* cur[x+refs-1-j] */\
"movq %%mm2, %%mm4 \n\t"\
"movq %%mm2, %%mm5 \n\t"\
"pxor %%mm3, %%mm4 \n\t"\
"pavgb %%mm3, %%mm5 \n\t"\
"pand %[pb1], %%mm4 \n\t"\
"psubusb %%mm4, %%mm5 \n\t"\
"psrlq $8, %%mm5 \n\t"\
"punpcklbw %%mm7, %%mm5 \n\t" /* (cur[x-refs+j] + cur[x+refs-j])>>1 */\
"movq %%mm2, %%mm4 \n\t"\
"psubusb %%mm3, %%mm2 \n\t"\
"psubusb %%mm4, %%mm3 \n\t"\
"pmaxub %%mm3, %%mm2 \n\t"\
"movq %%mm2, %%mm3 \n\t"\
"movq %%mm2, %%mm4 \n\t" /* ABS(cur[x-refs-1+j] - cur[x+refs-1-j]) */\
"psrlq $8, %%mm3 \n\t" /* ABS(cur[x-refs +j] - cur[x+refs -j]) */\
"psrlq $16, %%mm4 \n\t" /* ABS(cur[x-refs+1+j] - cur[x+refs+1-j]) */\
"punpcklbw %%mm7, %%mm2 \n\t"\
"punpcklbw %%mm7, %%mm3 \n\t"\
"punpcklbw %%mm7, %%mm4 \n\t"\
"paddw %%mm3, %%mm2 \n\t"\
"paddw %%mm4, %%mm2 \n\t" /* score */
#define CHECK1 \
"movq %%mm0, %%mm3 \n\t"\
"pcmpgtw %%mm2, %%mm3 \n\t" /* if(score < spatial_score) */\
"pminsw %%mm2, %%mm0 \n\t" /* spatial_score= score; */\
"movq %%mm3, %%mm6 \n\t"\
"pand %%mm3, %%mm5 \n\t"\
"pandn %%mm1, %%mm3 \n\t"\
"por %%mm5, %%mm3 \n\t"\
"movq %%mm3, %%mm1 \n\t" /* spatial_pred= (cur[x-refs+j] + cur[x+refs-j])>>1; */
#define CHECK2 /* pretend not to have checked dir=2 if dir=1 was bad.\
hurts both quality and speed, but matches the C version. */\
"paddw %[pw1], %%mm6 \n\t"\
"psllw $14, %%mm6 \n\t"\
"paddsw %%mm6, %%mm2 \n\t"\
"movq %%mm0, %%mm3 \n\t"\
"pcmpgtw %%mm2, %%mm3 \n\t"\
"pminsw %%mm2, %%mm0 \n\t"\
"pand %%mm3, %%mm5 \n\t"\
"pandn %%mm1, %%mm3 \n\t"\
"por %%mm5, %%mm3 \n\t"\
"movq %%mm3, %%mm1 \n\t"
void ff_yadif_filter_line_mmx(uint8_t *dst,
uint8_t *prev, uint8_t *cur, uint8_t *next,
int w, int refs, int parity, int mode)
{
static const uint64_t pw_1 = 0x0001000100010001ULL;
static const uint64_t pb_1 = 0x0101010101010101ULL;
uint64_t tmp0, tmp1, tmp2, tmp3;
int x;
#define FILTER\
for(x=0; x<w; x+=4){\
__asm__ volatile(\
"pxor %%mm7, %%mm7 \n\t"\
LOAD4("(%[cur],%[mrefs])", %%mm0) /* c = cur[x-refs] */\
LOAD4("(%[cur],%[prefs])", %%mm1) /* e = cur[x+refs] */\
LOAD4("(%["prev2"])", %%mm2) /* prev2[x] */\
LOAD4("(%["next2"])", %%mm3) /* next2[x] */\
"movq %%mm3, %%mm4 \n\t"\
"paddw %%mm2, %%mm3 \n\t"\
"psraw $1, %%mm3 \n\t" /* d = (prev2[x] + next2[x])>>1 */\
"movq %%mm0, %[tmp0] \n\t" /* c */\
"movq %%mm3, %[tmp1] \n\t" /* d */\
"movq %%mm1, %[tmp2] \n\t" /* e */\
"psubw %%mm4, %%mm2 \n\t"\
PABS( %%mm4, %%mm2) /* temporal_diff0 */\
LOAD4("(%[prev],%[mrefs])", %%mm3) /* prev[x-refs] */\
LOAD4("(%[prev],%[prefs])", %%mm4) /* prev[x+refs] */\
"psubw %%mm0, %%mm3 \n\t"\
"psubw %%mm1, %%mm4 \n\t"\
PABS( %%mm5, %%mm3)\
PABS( %%mm5, %%mm4)\
"paddw %%mm4, %%mm3 \n\t" /* temporal_diff1 */\
"psrlw $1, %%mm2 \n\t"\
"psrlw $1, %%mm3 \n\t"\
"pmaxsw %%mm3, %%mm2 \n\t"\
LOAD4("(%[next],%[mrefs])", %%mm3) /* next[x-refs] */\
LOAD4("(%[next],%[prefs])", %%mm4) /* next[x+refs] */\
"psubw %%mm0, %%mm3 \n\t"\
"psubw %%mm1, %%mm4 \n\t"\
PABS( %%mm5, %%mm3)\
PABS( %%mm5, %%mm4)\
"paddw %%mm4, %%mm3 \n\t" /* temporal_diff2 */\
"psrlw $1, %%mm3 \n\t"\
"pmaxsw %%mm3, %%mm2 \n\t"\
"movq %%mm2, %[tmp3] \n\t" /* diff */\
\
"paddw %%mm0, %%mm1 \n\t"\
"paddw %%mm0, %%mm0 \n\t"\
"psubw %%mm1, %%mm0 \n\t"\
"psrlw $1, %%mm1 \n\t" /* spatial_pred */\
PABS( %%mm2, %%mm0) /* ABS(c-e) */\
\
"movq -1(%[cur],%[mrefs]), %%mm2 \n\t" /* cur[x-refs-1] */\
"movq -1(%[cur],%[prefs]), %%mm3 \n\t" /* cur[x+refs-1] */\
"movq %%mm2, %%mm4 \n\t"\
"psubusb %%mm3, %%mm2 \n\t"\
"psubusb %%mm4, %%mm3 \n\t"\
"pmaxub %%mm3, %%mm2 \n\t"\
"pshufw $9,%%mm2, %%mm3 \n\t"\
"punpcklbw %%mm7, %%mm2 \n\t" /* ABS(cur[x-refs-1] - cur[x+refs-1]) */\
"punpcklbw %%mm7, %%mm3 \n\t" /* ABS(cur[x-refs+1] - cur[x+refs+1]) */\
"paddw %%mm2, %%mm0 \n\t"\
"paddw %%mm3, %%mm0 \n\t"\
"psubw %[pw1], %%mm0 \n\t" /* spatial_score */\
\
CHECK(-2,0)\
CHECK1\
CHECK(-3,1)\
CHECK2\
CHECK(0,-2)\
CHECK1\
CHECK(1,-3)\
CHECK2\
\
/* if(p->mode<2) ... */\
"movq %[tmp3], %%mm6 \n\t" /* diff */\
"cmp $2, %[mode] \n\t"\
"jge 1f \n\t"\
LOAD4("(%["prev2"],%[mrefs],2)", %%mm2) /* prev2[x-2*refs] */\
LOAD4("(%["next2"],%[mrefs],2)", %%mm4) /* next2[x-2*refs] */\
LOAD4("(%["prev2"],%[prefs],2)", %%mm3) /* prev2[x+2*refs] */\
LOAD4("(%["next2"],%[prefs],2)", %%mm5) /* next2[x+2*refs] */\
"paddw %%mm4, %%mm2 \n\t"\
"paddw %%mm5, %%mm3 \n\t"\
"psrlw $1, %%mm2 \n\t" /* b */\
"psrlw $1, %%mm3 \n\t" /* f */\
"movq %[tmp0], %%mm4 \n\t" /* c */\
"movq %[tmp1], %%mm5 \n\t" /* d */\
"movq %[tmp2], %%mm7 \n\t" /* e */\
"psubw %%mm4, %%mm2 \n\t" /* b-c */\
"psubw %%mm7, %%mm3 \n\t" /* f-e */\
"movq %%mm5, %%mm0 \n\t"\
"psubw %%mm4, %%mm5 \n\t" /* d-c */\
"psubw %%mm7, %%mm0 \n\t" /* d-e */\
"movq %%mm2, %%mm4 \n\t"\
"pminsw %%mm3, %%mm2 \n\t"\
"pmaxsw %%mm4, %%mm3 \n\t"\
"pmaxsw %%mm5, %%mm2 \n\t"\
"pminsw %%mm5, %%mm3 \n\t"\
"pmaxsw %%mm0, %%mm2 \n\t" /* max */\
"pminsw %%mm0, %%mm3 \n\t" /* min */\
"pxor %%mm4, %%mm4 \n\t"\
"pmaxsw %%mm3, %%mm6 \n\t"\
"psubw %%mm2, %%mm4 \n\t" /* -max */\
"pmaxsw %%mm4, %%mm6 \n\t" /* diff= MAX3(diff, min, -max); */\
"1: \n\t"\
\
"movq %[tmp1], %%mm2 \n\t" /* d */\
"movq %%mm2, %%mm3 \n\t"\
"psubw %%mm6, %%mm2 \n\t" /* d-diff */\
"paddw %%mm6, %%mm3 \n\t" /* d+diff */\
"pmaxsw %%mm2, %%mm1 \n\t"\
"pminsw %%mm3, %%mm1 \n\t" /* d = clip(spatial_pred, d-diff, d+diff); */\
"packuswb %%mm1, %%mm1 \n\t"\
\
:[tmp0]"=m"(tmp0),\
[tmp1]"=m"(tmp1),\
[tmp2]"=m"(tmp2),\
[tmp3]"=m"(tmp3)\
:[prev] "r"(prev),\
[cur] "r"(cur),\
[next] "r"(next),\
[prefs]"r"((x86_reg)refs),\
[mrefs]"r"((x86_reg)-refs),\
[pw1] "m"(pw_1),\
[pb1] "m"(pb_1),\
[mode] "g"(mode)\
);\
__asm__ volatile("movd %%mm1, %0" :"=m"(*dst));\
dst += 4;\
prev+= 4;\
cur += 4;\
next+= 4;\
}
if (parity) {
#define prev2 "prev"
#define next2 "cur"
FILTER
#undef prev2