yadif filter, based on stefanos port of my yadif from mplayer.
authorMichael Niedermayer <michaelni@gmx.at>
Sat, 25 Sep 2010 16:43:42 +0000 (16:43 +0000)
committerMichael Niedermayer <michaelni@gmx.at>
Sat, 25 Sep 2010 16:43:42 +0000 (16:43 +0000)
Compared to stefanos, 2 frame output works with ffplay.

Originally committed as revision 25196 to svn://svn.ffmpeg.org/ffmpeg/trunk

configure
doc/filters.texi
libavfilter/Makefile
libavfilter/allfilters.c
libavfilter/vf_yadif.c [new file with mode: 0644]
libavfilter/x86/yadif.c [new file with mode: 0644]
libavfilter/yadif.h [new file with mode: 0644]

index 781b09d..f4e4feb 100755 (executable)
--- a/configure
+++ b/configure
@@ -1402,6 +1402,7 @@ udp_protocol_deps="network"
 
 # filters
 ocv_smooth_filter_deps="libopencv"
+yadif_filter_deps="gpl"
 
 # libraries
 avdevice_deps="avcodec avformat"
index 54cdf58..d000276 100644 (file)
@@ -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
index 39c4b3b..c3f4a49 100644 (file)
@@ -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
index ac54bd0..ec7d933 100644 (file)
@@ -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);
diff --git a/libavfilter/vf_yadif.c b/libavfilter/vf_yadif.c
new file mode 100644 (file)
index 0000000..b2b2fe5
--- /dev/null
@@ -0,0 +1,332 @@
+/*
+ * 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}},
+};
diff --git a/libavfilter/x86/yadif.c b/libavfilter/x86/yadif.c
new file mode 100644 (file)
index 0000000..e92da8c
--- /dev/null
@@ -0,0 +1,240 @@
+/*
+ * 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
+#undef next2
+    } else {
+#define prev2 "cur"
+#define next2 "next"
+        FILTER
+#undef prev2
+#undef next2
+    }
+}
+#undef LOAD4
+#undef PABS
+#undef CHECK
+#undef CHECK1
+#undef CHECK2
+#undef FILTER
+
+#endif /* HAVE_MMX */
diff --git a/libavfilter/yadif.h b/libavfilter/yadif.h
new file mode 100644 (file)
index 0000000..4686aa2
--- /dev/null
@@ -0,0 +1,28 @@
+/*
+ * 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.
+ */
+
+#ifndef AVFILTER_YADIF_H
+#define AVFILTER_YADIF_H
+
+#include "avfilter.h"
+
+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);
+
+#endif /* AVFILTER_YADIF_H */