lavc: introduce VideoDSPContext

Move some functions from dsputil. The idea is that videodsp contains
functions that are useful for a large and varied set of video decoders.
Currently, it contains emulated_edge_mc() and prefetch().

Signed-off-by: Luca Barbato <lu_zero@gentoo.org>

37 files changed, 1189 insertions, 927 deletions

diff --git a/configure b/configure
index 08f1d82af5..f099118b66 100755
--- a/configure
+++ b/configure
@@ -1328,6 +1328,7 @@ CONFIG_EXTRA="
     rangecoder
     rtpdec
     sinewin
+    videodsp
     vp3dsp
 "
 
@@ -1544,6 +1545,7 @@ mpeg4_decoder_select="h263_decoder mpeg4video_parser"
 mpeg4_encoder_select="h263_encoder"
 mpeg4_vaapi_hwaccel_select="vaapi mpeg4_decoder"
 mpeg4_vdpau_decoder_select="vdpau mpeg4_decoder"
+mpegvideo_select="videodsp"
 msmpeg4v1_decoder_select="h263_decoder"
 msmpeg4v1_encoder_select="h263_encoder"
 msmpeg4v2_decoder_select="h263_decoder"
@@ -1589,12 +1591,12 @@ vc1_vdpau_decoder_select="vdpau vc1_decoder"
 vc1image_decoder_select="vc1_decoder"
 vorbis_decoder_select="mdct"
 vorbis_encoder_select="mdct"
-vp3_decoder_select="vp3dsp"
-vp5_decoder_select="vp3dsp"
-vp6_decoder_select="huffman vp3dsp"
+vp3_decoder_select="vp3dsp videodsp"
+vp5_decoder_select="vp3dsp videodsp"
+vp6_decoder_select="huffman vp3dsp videodsp"
 vp6a_decoder_select="vp6_decoder"
 vp6f_decoder_select="vp6_decoder"
-vp8_decoder_select="h264pred h264qpel"
+vp8_decoder_select="h264pred videodsp"
 wmapro_decoder_select="mdct sinewin"
 wmav1_decoder_select="mdct sinewin"
 wmav1_encoder_select="mdct sinewin"
diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index b5354306b7..de638dde20 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -67,6 +67,7 @@ OBJS-$(CONFIG_RDFT)                    += rdft.o $(RDFT-OBJS-yes)
 OBJS-$(CONFIG_SINEWIN)                 += sinewin.o
 OBJS-$(CONFIG_VAAPI)                   += vaapi.o
 OBJS-$(CONFIG_VDPAU)                   += vdpau.o
+OBJS-$(CONFIG_VIDEODSP)                += videodsp.o
 OBJS-$(CONFIG_VP3DSP)                  += vp3dsp.o
 
 # decoders/encoders/hardware accelerators
diff --git a/libavcodec/arm/Makefile b/libavcodec/arm/Makefile
index b315d8da0e..ac486f47dc 100644
--- a/libavcodec/arm/Makefile
+++ b/libavcodec/arm/Makefile
@@ -30,6 +30,8 @@ OBJS-$(CONFIG_RV30_DECODER)            += arm/rv34dsp_init_arm.o
 OBJS-$(CONFIG_RV40_DECODER)            += arm/rv34dsp_init_arm.o        \
                                           arm/rv40dsp_init_arm.o        \
 
+OBJS-$(CONFIG_VIDEODSP)                += arm/videodsp_init_arm.o       \
+
 OBJS                                   += arm/dsputil_init_arm.o        \
                                           arm/dsputil_arm.o             \
                                           arm/fft_init_arm.o            \
@@ -41,6 +43,9 @@ OBJS                                   += arm/dsputil_init_arm.o        \
 ARMV5TE-OBJS-$(CONFIG_MPEGVIDEO)       += arm/mpegvideo_armv5te.o       \
                                           arm/mpegvideo_armv5te_s.o     \
 
+ARMV5TE-OBJS-$(CONFIG_VIDEODSP)        += arm/videodsp_init_armv5te.o   \
+                                          arm/videodsp_armv5te.o        \
+
 ARMV5TE-OBJS                           += arm/dsputil_init_armv5te.o    \
                                           arm/simple_idct_armv5te.o     \
 
diff --git a/libavcodec/arm/dsputil_arm.S b/libavcodec/arm/dsputil_arm.S
index ffb6eb4b50..8504032bf4 100644
--- a/libavcodec/arm/dsputil_arm.S
+++ b/libavcodec/arm/dsputil_arm.S
@@ -22,15 +22,7 @@
 #include "config.h"
 #include "libavutil/arm/asm.S"
 
-#if HAVE_ARMV5TE_EXTERNAL
-function ff_prefetch_arm, export=1
-        subs            r2,  r2,  #1
-        pld             [r0]
-        add             r0,  r0,  r1
-        bne             ff_prefetch_arm
-        bx              lr
-endfunc
-#else
+#if !HAVE_ARMV5TE_EXTERNAL
 #define pld @
 #endif
 
diff --git a/libavcodec/arm/dsputil_init_armv5te.c b/libavcodec/arm/dsputil_init_armv5te.c
index f37ffc3e9c..17017aa17d 100644
--- a/libavcodec/arm/dsputil_init_armv5te.c
+++ b/libavcodec/arm/dsputil_init_armv5te.c
@@ -25,8 +25,6 @@ void ff_simple_idct_armv5te(DCTELEM *data);
 void ff_simple_idct_put_armv5te(uint8_t *dest, int line_size, DCTELEM *data);
 void ff_simple_idct_add_armv5te(uint8_t *dest, int line_size, DCTELEM *data);
 
-void ff_prefetch_arm(void *mem, int stride, int h);
-
 av_cold void ff_dsputil_init_armv5te(DSPContext *c, AVCodecContext *avctx)
 {
     if (avctx->bits_per_raw_sample <= 8 &&
@@ -37,6 +35,4 @@ av_cold void ff_dsputil_init_armv5te(DSPContext *c, AVCodecContext *avctx)
         c->idct                  = ff_simple_idct_armv5te;
         c->idct_permutation_type = FF_NO_IDCT_PERM;
     }
-
-    c->prefetch = ff_prefetch_arm;
 }
diff --git a/libavcodec/arm/videodsp_armv5te.S b/libavcodec/arm/videodsp_armv5te.S
new file mode 100644
index 0000000000..96e92d5a73
--- /dev/null
+++ b/libavcodec/arm/videodsp_armv5te.S
@@ -0,0 +1,33 @@
+@
+@ ARMv5te optimized DSP utils
+@ Copyright (c) 2004 AGAWA Koji <i (AT) atty (DOT) jp>
+@
+@ This file is part of Libav.
+@
+@ Libav is free software; you can redistribute it and/or
+@ modify it under the terms of the GNU Lesser General Public
+@ License as published by the Free Software Foundation; either
+@ version 2.1 of the License, or (at your option) any later version.
+@
+@ Libav 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
+@ Lesser General Public License for more details.
+@
+@ You should have received a copy of the GNU Lesser General Public
+@ License along with Libav; if not, write to the Free Software
+@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+@
+
+#include "config.h"
+#include "libavutil/arm/asm.S"
+
+#if HAVE_ARMV5TE_EXTERNAL
+function ff_prefetch_arm, export=1
+        subs            r2,  r2,  #1
+        pld             [r0]
+        add             r0,  r0,  r1
+        bne             ff_prefetch_arm
+        bx              lr
+endfunc
+#endif
diff --git a/libavcodec/arm/videodsp_init_arm.c b/libavcodec/arm/videodsp_init_arm.c
new file mode 100644
index 0000000000..6c97b2ce5e
--- /dev/null
+++ b/libavcodec/arm/videodsp_init_arm.c
@@ -0,0 +1,29 @@
+/*
+ * Copyright (C) 2012 Ronald S. Bultje
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/arm/cpu.h"
+#include "libavcodec/videodsp.h"
+#include "videodsp_arm.h"
+
+void ff_videodsp_init_arm(VideoDSPContext *ctx, int bpc)
+{
+    int cpu_flags = av_get_cpu_flags();
+    if (have_armv5te(cpu_flags)) ff_videodsp_init_armv5te(ctx, bpc);
+}
diff --git a/libavcodec/cavs.c b/libavcodec/cavs.c
index a68a62801b..cec1d23683 100644
--- a/libavcodec/cavs.c
+++ b/libavcodec/cavs.c
@@ -392,7 +392,7 @@ static inline void mc_dir_part(AVSContext *h,Picture *pic,
           || full_my < 0-extra_height
           || full_mx + 16/*FIXME*/ > pic_width + extra_width
           || full_my + 16/*FIXME*/ > pic_height + extra_height){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->l_stride, h->l_stride,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->l_stride, h->l_stride,
                             16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
         src_y= s->edge_emu_buffer + 2 + 2*h->l_stride;
         emu=1;
@@ -401,14 +401,14 @@ static inline void mc_dir_part(AVSContext *h,Picture *pic,
     qpix_op[luma_xy](dest_y, src_y, h->l_stride); //FIXME try variable height perhaps?
 
     if(emu){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->c_stride,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->c_stride,
                             9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
         src_cb= s->edge_emu_buffer;
     }
     chroma_op(dest_cb, src_cb, h->c_stride, chroma_height, mx&7, my&7);
 
     if(emu){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->c_stride,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->c_stride,
                             9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
         src_cr= s->edge_emu_buffer;
     }
diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index d4471dc24c..4696bc7980 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -2615,8 +2615,6 @@ static void ff_jref_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
     add_pixels_clamped_c(block, dest, line_size);
 }
 
-static void just_return(void *mem av_unused, int stride av_unused, int h av_unused) { return; }
-
 /* init static data */
 av_cold void ff_dsputil_static_init(void)
 {
@@ -2867,8 +2865,6 @@ av_cold void ff_dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->shrink[2]= ff_shrink44;
     c->shrink[3]= ff_shrink88;
 
-    c->prefetch= just_return;
-
     memset(c->put_2tap_qpel_pixels_tab, 0, sizeof(c->put_2tap_qpel_pixels_tab));
     memset(c->avg_2tap_qpel_pixels_tab, 0, sizeof(c->avg_2tap_qpel_pixels_tab));
 
@@ -2905,7 +2901,6 @@ av_cold void ff_dsputil_init(DSPContext* c, AVCodecContext *avctx)
 #define BIT_DEPTH_FUNCS(depth, dct)\
     c->get_pixels                    = FUNCC(get_pixels   ## dct   , depth);\
     c->draw_edges                    = FUNCC(draw_edges            , depth);\
-    c->emulated_edge_mc              = FUNC (ff_emulated_edge_mc   , depth);\
     c->clear_block                   = FUNCC(clear_block  ## dct   , depth);\
     c->clear_blocks                  = FUNCC(clear_blocks ## dct   , depth);\
     c->add_pixels8                   = FUNCC(add_pixels8  ## dct   , depth);\
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 364270b0b3..57c8bebb63 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -188,15 +188,6 @@ void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
 void ff_init_scantable_permutation(uint8_t *idct_permutation,
                                    int idct_permutation_type);
 
-#define EMULATED_EDGE(depth) \
-void ff_emulated_edge_mc_ ## depth (uint8_t *buf, const uint8_t *src, int linesize,\
-                         int block_w, int block_h,\
-                         int src_x, int src_y, int w, int h);
-
-EMULATED_EDGE(8)
-EMULATED_EDGE(9)
-EMULATED_EDGE(10)
-
 /**
  * DSPContext.
  */
@@ -216,21 +207,6 @@ typedef struct DSPContext {
     void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
     int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
     /**
-     * Motion estimation with emulated edge values.
-     * @param buf pointer to destination buffer (unaligned)
-     * @param src pointer to pixel source (unaligned)
-     * @param linesize width (in pixels) for src/buf
-     * @param block_w number of pixels (per row) to copy to buf
-     * @param block_h nummber of pixel rows to copy to buf
-     * @param src_x offset of src to start of row - this may be negative
-     * @param src_y offset of src to top of image - this may be negative
-     * @param w width of src in pixels
-     * @param h height of src in pixels
-     */
-    void (*emulated_edge_mc)(uint8_t *buf, const uint8_t *src, int linesize,
-                             int block_w, int block_h,
-                             int src_x, int src_y, int w, int h);
-    /**
      * translational global motion compensation.
      */
     void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
@@ -465,8 +441,6 @@ typedef struct DSPContext {
 #define EDGE_TOP    1
 #define EDGE_BOTTOM 2
 
-    void (*prefetch)(void *mem, int stride, int h);
-
     void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 
     /**
diff --git a/libavcodec/dsputil_template.c b/libavcodec/dsputil_template.c
index 72ed6bfd2a..13f7628b2e 100644
--- a/libavcodec/dsputil_template.c
+++ b/libavcodec/dsputil_template.c
@@ -113,85 +113,6 @@ static void FUNCC(draw_edges)(uint8_t *_buf, int _wrap, int width, int height, i
             memcpy(last_line + (i + 1) * wrap, last_line, (width + w + w) * sizeof(pixel)); // bottom
 }
 
-/**
- * Copy a rectangular area of samples to a temporary buffer and replicate the border samples.
- * @param buf destination buffer
- * @param src source buffer
- * @param linesize number of bytes between 2 vertically adjacent samples in both the source and destination buffers
- * @param block_w width of block
- * @param block_h height of block
- * @param src_x x coordinate of the top left sample of the block in the source buffer
- * @param src_y y coordinate of the top left sample of the block in the source buffer
- * @param w width of the source buffer
- * @param h height of the source buffer
- */
-void FUNC(ff_emulated_edge_mc)(uint8_t *buf, const uint8_t *src, int linesize, int block_w, int block_h,
-                                    int src_x, int src_y, int w, int h){
-    int x, y;
-    int start_y, start_x, end_y, end_x;
-
-    if(src_y>= h){
-        src+= (h-1-src_y)*linesize;
-        src_y=h-1;
-    }else if(src_y<=-block_h){
-        src+= (1-block_h-src_y)*linesize;
-        src_y=1-block_h;
-    }
-    if(src_x>= w){
-        src+= (w-1-src_x)*sizeof(pixel);
-        src_x=w-1;
-    }else if(src_x<=-block_w){
-        src+= (1-block_w-src_x)*sizeof(pixel);
-        src_x=1-block_w;
-    }
-
-    start_y= FFMAX(0, -src_y);
-    start_x= FFMAX(0, -src_x);
-    end_y= FFMIN(block_h, h-src_y);
-    end_x= FFMIN(block_w, w-src_x);
-    assert(start_y < end_y && block_h);
-    assert(start_x < end_x && block_w);
-
-    w    = end_x - start_x;
-    src += start_y*linesize + start_x*sizeof(pixel);
-    buf += start_x*sizeof(pixel);
-
-    //top
-    for(y=0; y<start_y; y++){
-        memcpy(buf, src, w*sizeof(pixel));
-        buf += linesize;
-    }
-
-    // copy existing part
-    for(; y<end_y; y++){
-        memcpy(buf, src, w*sizeof(pixel));
-        src += linesize;
-        buf += linesize;
-    }
-
-    //bottom
-    src -= linesize;
-    for(; y<block_h; y++){
-        memcpy(buf, src, w*sizeof(pixel));
-        buf += linesize;
-    }
-
-    buf -= block_h * linesize + start_x*sizeof(pixel);
-    while (block_h--){
-        pixel *bufp = (pixel*)buf;
-       //left
-        for(x=0; x<start_x; x++){
-            bufp[x] = bufp[start_x];
-        }
-
-       //right
-        for(x=end_x; x<block_w; x++){
-            bufp[x] = bufp[end_x - 1];
-        }
-        buf += linesize;
-    }
-}
-
 #define DCTELEM_FUNCS(dctcoef, suffix)                                  \
 static void FUNCC(get_pixels ## suffix)(DCTELEM *restrict _block,       \
                                         const uint8_t *_pixels,         \
diff --git a/libavcodec/h264.c b/libavcodec/h264.c
index ad3f15c3f3..1a0b341641 100644
--- a/libavcodec/h264.c
+++ b/libavcodec/h264.c
@@ -486,11 +486,11 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
         full_my                <          0 - extra_height ||
         full_mx + 16 /*FIXME*/ > pic_width  + extra_width  ||
         full_my + 16 /*FIXME*/ > pic_height + extra_height) {
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer,
-                                src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
-                                h->mb_linesize,
-                                16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
-                                full_my - 2, pic_width, pic_height);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+                                 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
+                                 h->mb_linesize,
+                                 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
+                                 full_my - 2, pic_width, pic_height);
         src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
         emu   = 1;
     }
@@ -505,12 +505,12 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
     if (chroma_idc == 3 /* yuv444 */) {
         src_cb = pic->f.data[1] + offset;
         if (emu) {
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer,
-                                    src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
-                                    h->mb_linesize,
-                                    16 + 5, 16 + 5 /*FIXME*/,
-                                    full_mx - 2, full_my - 2,
-                                    pic_width, pic_height);
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+                                     src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
+                                     h->mb_linesize,
+                                     16 + 5, 16 + 5 /*FIXME*/,
+                                     full_mx - 2, full_my - 2,
+                                     pic_width, pic_height);
             src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
         }
         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
@@ -519,12 +519,12 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
 
         src_cr = pic->f.data[2] + offset;
         if (emu) {
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer,
-                                    src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
-                                    h->mb_linesize,
-                                    16 + 5, 16 + 5 /*FIXME*/,
-                                    full_mx - 2, full_my - 2,
-                                    pic_width, pic_height);
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+                                     src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
+                                     h->mb_linesize,
+                                     16 + 5, 16 + 5 /*FIXME*/,
+                                     full_mx - 2, full_my - 2,
+                                     pic_width, pic_height);
             src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
         }
         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
@@ -546,9 +546,9 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
              (my >> ysh) * h->mb_uvlinesize;
 
     if (emu) {
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
-                                9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
-                                pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
+                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
+                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
         src_cb = s->edge_emu_buffer;
     }
     chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
@@ -556,9 +556,9 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
 
     if (emu) {
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
-                                9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
-                                pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
+                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
+                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
         src_cr = s->edge_emu_buffer;
     }
     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
@@ -735,15 +735,15 @@ static av_always_inline void prefetch_motion(H264Context *h, int list,
         int off       = (mx << pixel_shift) +
                         (my + (s->mb_x & 3) * 4) * h->mb_linesize +
                         (64 << pixel_shift);
-        s->dsp.prefetch(src[0] + off, s->linesize, 4);
+        s->vdsp.prefetch(src[0] + off, s->linesize, 4);
         if (chroma_idc == 3 /* yuv444 */) {
-            s->dsp.prefetch(src[1] + off, s->linesize, 4);
-            s->dsp.prefetch(src[2] + off, s->linesize, 4);
+            s->vdsp.prefetch(src[1] + off, s->linesize, 4);
+            s->vdsp.prefetch(src[2] + off, s->linesize, 4);
         } else {
             off = ((mx >> 1) << pixel_shift) +
                   ((my >> 1) + (s->mb_x & 7)) * s->uvlinesize +
                   (64 << pixel_shift);
-            s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
+            s->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
         }
     }
 }
@@ -973,6 +973,7 @@ static av_cold void common_init(H264Context *h)
 
     /* needed so that IDCT permutation is known early */
     ff_dsputil_init(&s->dsp, s->avctx);
+    ff_videodsp_init(&s->vdsp, 8);
 
     memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
     memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
@@ -2439,6 +2440,7 @@ static int h264_set_parameter_from_sps(H264Context *h)
                               h->sps.chroma_format_idc);
             s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
             ff_dsputil_init(&s->dsp, s->avctx);
+            ff_videodsp_init(&s->vdsp, h->sps.bit_depth_luma);
         } else {
             av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n",
                    h->sps.bit_depth_luma);
diff --git a/libavcodec/h264_mb_template.c b/libavcodec/h264_mb_template.c
index 1e5b7f1f4e..e27fd1330d 100644
--- a/libavcodec/h264_mb_template.c
+++ b/libavcodec/h264_mb_template.c
@@ -60,8 +60,8 @@ static av_noinline void FUNC(hl_decode_mb)(H264Context *h)
     dest_cb = s->current_picture.f.data[1] +  (mb_x << PIXEL_SHIFT) * 8 + mb_y * s->uvlinesize * block_h;
     dest_cr = s->current_picture.f.data[2] +  (mb_x << PIXEL_SHIFT) * 8 + mb_y * s->uvlinesize * block_h;
 
-    s->dsp.prefetch(dest_y  + (s->mb_x & 3) * 4 * s->linesize   + (64 << PIXEL_SHIFT), s->linesize,       4);
-    s->dsp.prefetch(dest_cb + (s->mb_x & 7)     * s->uvlinesize + (64 << PIXEL_SHIFT), dest_cr - dest_cb, 2);
+    s->vdsp.prefetch(dest_y  + (s->mb_x & 3) * 4 * s->linesize   + (64 << PIXEL_SHIFT), s->linesize,       4);
+    s->vdsp.prefetch(dest_cb + (s->mb_x & 7)     * s->uvlinesize + (64 << PIXEL_SHIFT), dest_cr - dest_cb, 2);
 
     h->list_counts[mb_xy] = h->list_count;
 
@@ -292,8 +292,8 @@ static av_noinline void FUNC(hl_decode_mb_444)(H264Context *h)
     for (p = 0; p < plane_count; p++) {
         dest[p] = s->current_picture.f.data[p] +
                   ((mb_x << PIXEL_SHIFT) + mb_y * s->linesize) * 16;
-        s->dsp.prefetch(dest[p] + (s->mb_x & 3) * 4 * s->linesize + (64 << PIXEL_SHIFT),
-                        s->linesize, 4);
+        s->vdsp.prefetch(dest[p] + (s->mb_x & 3) * 4 * s->linesize + (64 << PIXEL_SHIFT),
+                         s->linesize, 4);
     }
 
     h->list_counts[mb_xy] = h->list_count;
diff --git a/libavcodec/mpegvideo.c b/libavcodec/mpegvideo.c
index edd52532b2..0fab9af326 100644
--- a/libavcodec/mpegvideo.c
+++ b/libavcodec/mpegvideo.c
@@ -175,6 +175,7 @@ const uint8_t *avpriv_mpv_find_start_code(const uint8_t *restrict p,
 av_cold int ff_dct_common_init(MpegEncContext *s)
 {
     ff_dsputil_init(&s->dsp, s->avctx);
+    ff_videodsp_init(&s->vdsp, 8);
 
     s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_c;
     s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_c;
diff --git a/libavcodec/mpegvideo.h b/libavcodec/mpegvideo.h
index 65997ce90b..359bebc74b 100644
--- a/libavcodec/mpegvideo.h
+++ b/libavcodec/mpegvideo.h
@@ -36,6 +36,7 @@
 #include "parser.h"
 #include "mpeg12data.h"
 #include "rl.h"
+#include "videodsp.h"
 
 #include "libavutil/opt.h"
 
@@ -358,6 +359,7 @@ typedef struct MpegEncContext {
     int h263_long_vectors;      ///< use horrible h263v1 long vector mode
 
     DSPContext dsp;             ///< pointers for accelerated dsp functions
+    VideoDSPContext vdsp;
     int f_code;                 ///< forward MV resolution
     int b_code;                 ///< backward MV resolution for B Frames (mpeg4)
     int16_t (*p_mv_table_base)[2];
diff --git a/libavcodec/mpegvideo_enc.c b/libavcodec/mpegvideo_enc.c
index 816f44caf2..f783dc7892 100644
--- a/libavcodec/mpegvideo_enc.c
+++ b/libavcodec/mpegvideo_enc.c
@@ -1782,16 +1782,16 @@ static av_always_inline void encode_mb_internal(MpegEncContext *s,
 
     if (mb_x * 16 + 16 > s->width || mb_y * 16 + 16 > s->height) {
         uint8_t *ebuf = s->edge_emu_buffer + 32;
-        s->dsp.emulated_edge_mc(ebuf, ptr_y, wrap_y, 16, 16, mb_x * 16,
-                                mb_y * 16, s->width, s->height);
+        s->vdsp.emulated_edge_mc(ebuf, ptr_y, wrap_y, 16, 16, mb_x * 16,
+                                 mb_y * 16, s->width, s->height);
         ptr_y = ebuf;
-        s->dsp.emulated_edge_mc(ebuf + 18 * wrap_y, ptr_cb, wrap_c, 8,
-                                mb_block_height, mb_x * 8, mb_y * 8,
-                                s->width >> 1, s->height >> 1);
+        s->vdsp.emulated_edge_mc(ebuf + 18 * wrap_y, ptr_cb, wrap_c, 8,
+                                 mb_block_height, mb_x * 8, mb_y * 8,
+                                 s->width >> 1, s->height >> 1);
         ptr_cb = ebuf + 18 * wrap_y;
-        s->dsp.emulated_edge_mc(ebuf + 18 * wrap_y + 8, ptr_cr, wrap_c, 8,
-                                mb_block_height, mb_x * 8, mb_y * 8,
-                                s->width >> 1, s->height >> 1);
+        s->vdsp.emulated_edge_mc(ebuf + 18 * wrap_y + 8, ptr_cr, wrap_c, 8,
+                                 mb_block_height, mb_x * 8, mb_y * 8,
+                                 s->width >> 1, s->height >> 1);
         ptr_cr = ebuf + 18 * wrap_y + 8;
     }
 
diff --git a/libavcodec/mpegvideo_motion.c b/libavcodec/mpegvideo_motion.c
index 4ea31ad252..536377a46f 100644
--- a/libavcodec/mpegvideo_motion.c
+++ b/libavcodec/mpegvideo_motion.c
@@ -59,7 +59,7 @@ static void gmc1_motion(MpegEncContext *s,
     if(s->flags&CODEC_FLAG_EMU_EDGE){
         if(   (unsigned)src_x >= FFMAX(s->h_edge_pos - 17, 0)
            || (unsigned)src_y >= FFMAX(s->v_edge_pos - 17, 0)){
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr, linesize, 17, 17, src_x, src_y, s->h_edge_pos, s->v_edge_pos);
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, linesize, 17, 17, src_x, src_y, s->h_edge_pos, s->v_edge_pos);
             ptr= s->edge_emu_buffer;
         }
     }
@@ -98,7 +98,7 @@ static void gmc1_motion(MpegEncContext *s,
     if(s->flags&CODEC_FLAG_EMU_EDGE){
         if(   (unsigned)src_x >= FFMAX((s->h_edge_pos>>1) - 9, 0)
            || (unsigned)src_y >= FFMAX((s->v_edge_pos>>1) - 9, 0)){
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr, uvlinesize, 9, 9, src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1);
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, uvlinesize, 9, 9, src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1);
             ptr= s->edge_emu_buffer;
             emu=1;
         }
@@ -107,7 +107,7 @@ static void gmc1_motion(MpegEncContext *s,
 
     ptr = ref_picture[2] + offset;
     if(emu){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr, uvlinesize, 9, 9, src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, uvlinesize, 9, 9, src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1);
         ptr= s->edge_emu_buffer;
     }
     s->dsp.gmc1(dest_cr, ptr, uvlinesize, 8, motion_x&15, motion_y&15, 128 - s->no_rounding);
@@ -195,7 +195,7 @@ static inline int hpel_motion(MpegEncContext *s,
     if(s->unrestricted_mv && (s->flags&CODEC_FLAG_EMU_EDGE)){
         if(   (unsigned)src_x > FFMAX(s->h_edge_pos - (motion_x&1) - 8, 0)
            || (unsigned)src_y > FFMAX(s->v_edge_pos - (motion_y&1) - 8, 0)){
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer, src, s->linesize, 9, 9,
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src, s->linesize, 9, 9,
                              src_x, src_y, s->h_edge_pos, s->v_edge_pos);
             src= s->edge_emu_buffer;
             emu=1;
@@ -285,19 +285,19 @@ if(s->quarter_sample)
                         "MPEG motion vector out of boundary (%d %d)\n", src_x, src_y);
                 return;
             }
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr_y, s->linesize,
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr_y, s->linesize,
                                 17, 17+field_based,
                                 src_x, src_y<<field_based,
                                 s->h_edge_pos, s->v_edge_pos);
             ptr_y = s->edge_emu_buffer;
             if(!CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
                 uint8_t *uvbuf= s->edge_emu_buffer+18*s->linesize;
-                s->dsp.emulated_edge_mc(uvbuf ,
+                s->vdsp.emulated_edge_mc(uvbuf ,
                                     ptr_cb, s->uvlinesize,
                                     9, 9+field_based,
                                     uvsrc_x, uvsrc_y<<field_based,
                                     s->h_edge_pos>>1, s->v_edge_pos>>1);
-                s->dsp.emulated_edge_mc(uvbuf+16,
+                s->vdsp.emulated_edge_mc(uvbuf+16,
                                     ptr_cr, s->uvlinesize,
                                     9, 9+field_based,
                                     uvsrc_x, uvsrc_y<<field_based,
@@ -498,17 +498,17 @@ static inline void qpel_motion(MpegEncContext *s,
 
     if(   (unsigned)src_x > FFMAX(s->h_edge_pos - (motion_x&3) - 16, 0)
        || (unsigned)src_y > FFMAX(   v_edge_pos - (motion_y&3) - h , 0)){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr_y, s->linesize,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr_y, s->linesize,
                             17, 17+field_based, src_x, src_y<<field_based,
                             s->h_edge_pos, s->v_edge_pos);
         ptr_y= s->edge_emu_buffer;
         if(!CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
             uint8_t *uvbuf= s->edge_emu_buffer + 18*s->linesize;
-            s->dsp.emulated_edge_mc(uvbuf, ptr_cb, s->uvlinesize,
+            s->vdsp.emulated_edge_mc(uvbuf, ptr_cb, s->uvlinesize,
                                 9, 9 + field_based,
                                 uvsrc_x, uvsrc_y<<field_based,
                                 s->h_edge_pos>>1, s->v_edge_pos>>1);
-            s->dsp.emulated_edge_mc(uvbuf + 16, ptr_cr, s->uvlinesize,
+            s->vdsp.emulated_edge_mc(uvbuf + 16, ptr_cr, s->uvlinesize,
                                 9, 9 + field_based,
                                 uvsrc_x, uvsrc_y<<field_based,
                                 s->h_edge_pos>>1, s->v_edge_pos>>1);
@@ -577,7 +577,7 @@ static void chroma_4mv_motion(MpegEncContext *s,
     if(s->flags&CODEC_FLAG_EMU_EDGE){
         if(   (unsigned)src_x > FFMAX((s->h_edge_pos>>1) - (dxy &1) - 8, 0)
            || (unsigned)src_y > FFMAX((s->v_edge_pos>>1) - (dxy>>1) - 8, 0)){
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize,
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize,
                                 9, 9, src_x, src_y,
                                 s->h_edge_pos>>1, s->v_edge_pos>>1);
             ptr= s->edge_emu_buffer;
@@ -588,7 +588,7 @@ static void chroma_4mv_motion(MpegEncContext *s,
 
     ptr = ref_picture[2] + offset;
     if(emu){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize,
                             9, 9, src_x, src_y,
                             s->h_edge_pos>>1, s->v_edge_pos>>1);
         ptr= s->edge_emu_buffer;
@@ -603,9 +603,9 @@ static inline void prefetch_motion(MpegEncContext *s, uint8_t **pix, int dir){
     const int mx= (s->mv[dir][0][0]>>shift) + 16*s->mb_x + 8;
     const int my= (s->mv[dir][0][1]>>shift) + 16*s->mb_y;
     int off= mx + (my + (s->mb_x&3)*4)*s->linesize + 64;
-    s->dsp.prefetch(pix[0]+off, s->linesize, 4);
+    s->vdsp.prefetch(pix[0]+off, s->linesize, 4);
     off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
-    s->dsp.prefetch(pix[1]+off, pix[2]-pix[1], 2);
+    s->vdsp.prefetch(pix[1]+off, pix[2]-pix[1], 2);
 }
 
 /**
@@ -757,7 +757,7 @@ static av_always_inline void MPV_motion_internal(MpegEncContext *s,
                 if(s->flags&CODEC_FLAG_EMU_EDGE){
                     if(   (unsigned)src_x > FFMAX(s->h_edge_pos - (motion_x&3) - 8, 0)
                        || (unsigned)src_y > FFMAX(s->v_edge_pos - (motion_y&3) - 8, 0)){
-                        s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr,
+                        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr,
                                             s->linesize, 9, 9,
                                             src_x, src_y,
                                             s->h_edge_pos, s->v_edge_pos);
diff --git a/libavcodec/ppc/Makefile b/libavcodec/ppc/Makefile
index f7548e375a..c8c3025bf0 100644
--- a/libavcodec/ppc/Makefile
+++ b/libavcodec/ppc/Makefile
@@ -1,4 +1,5 @@
 OBJS                                   += ppc/dsputil_ppc.o             \
+                                          ppc/videodsp_ppc.o            \
 
 OBJS-$(CONFIG_VP3DSP)                  += ppc/vp3dsp_altivec.o
 
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index 6b894e02aa..a3069402da 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -137,21 +137,11 @@ static long check_dcbzl_effect(void)
 }
 #endif
 
-static void prefetch_ppc(void *mem, int stride, int h)
-{
-    register const uint8_t *p = mem;
-    do {
-        __asm__ volatile ("dcbt 0,%0" : : "r" (p));
-        p+= stride;
-    } while(--h);
-}
-
 void ff_dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx)
 {
     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
 
     // Common optimizations whether AltiVec is available or not
-    c->prefetch = prefetch_ppc;
     if (!high_bit_depth) {
     switch (check_dcbzl_effect()) {
         case 32:
diff --git a/libavcodec/ppc/videodsp_ppc.c b/libavcodec/ppc/videodsp_ppc.c
new file mode 100644
index 0000000000..1865b198b0
--- /dev/null
+++ b/libavcodec/ppc/videodsp_ppc.c
@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) 2003-2004 Romain Dolbeau
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavcodec/videodsp.h"
+
+static void prefetch_ppc(uint8_t *mem, ptrdiff_t stride, int h)
+{
+    register const uint8_t *p = mem;
+    do {
+        __asm__ volatile ("dcbt 0,%0" : : "r" (p));
+        p += stride;
+    } while(--h);
+}
+
+void ff_videodsp_init_ppc(VideoDSPContext *ctx, int bpc)
+{
+    ctx->prefetch = prefetch_ppc;
+}
diff --git a/libavcodec/rv34.c b/libavcodec/rv34.c
index ffc8649f12..3106bfb52c 100644
--- a/libavcodec/rv34.c
+++ b/libavcodec/rv34.c
@@ -725,12 +725,12 @@ static inline void rv34_mc(RV34DecContext *r, const int block_type,
         uint8_t *uvbuf = s->edge_emu_buffer + 22 * s->linesize;
 
         srcY -= 2 + 2*s->linesize;
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize, (width<<3)+6, (height<<3)+6,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize, (width<<3)+6, (height<<3)+6,
                             src_x - 2, src_y - 2, s->h_edge_pos, s->v_edge_pos);
         srcY = s->edge_emu_buffer + 2 + 2*s->linesize;
-        s->dsp.emulated_edge_mc(uvbuf     , srcU, s->uvlinesize, (width<<2)+1, (height<<2)+1,
+        s->vdsp.emulated_edge_mc(uvbuf     , srcU, s->uvlinesize, (width<<2)+1, (height<<2)+1,
                             uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, s->v_edge_pos >> 1);
-        s->dsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, (width<<2)+1, (height<<2)+1,
+        s->vdsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, (width<<2)+1, (height<<2)+1,
                             uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, s->v_edge_pos >> 1);
         srcU = uvbuf;
         srcV = uvbuf + 16;
diff --git a/libavcodec/svq3.c b/libavcodec/svq3.c
index 0bfc73ec8d..266e22f985 100644
--- a/libavcodec/svq3.c
+++ b/libavcodec/svq3.c
@@ -293,9 +293,9 @@ static inline void svq3_mc_dir_part(MpegEncContext *s,
     src  = pic->f.data[0] + mx + my * s->linesize;
 
     if (emu) {
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, src, s->linesize,
-                                width + 1, height + 1,
-                                mx, my, s->h_edge_pos, s->v_edge_pos);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src, s->linesize,
+                                 width + 1, height + 1,
+                                 mx, my, s->h_edge_pos, s->v_edge_pos);
         src = s->edge_emu_buffer;
     }
     if (thirdpel)
@@ -319,10 +319,10 @@ static inline void svq3_mc_dir_part(MpegEncContext *s,
             src  = pic->f.data[i] + mx + my * s->uvlinesize;
 
             if (emu) {
-                s->dsp.emulated_edge_mc(s->edge_emu_buffer, src, s->uvlinesize,
-                                        width + 1, height + 1,
-                                        mx, my, (s->h_edge_pos >> 1),
-                                        s->v_edge_pos >> 1);
+                s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src, s->uvlinesize,
+                                         width + 1, height + 1,
+                                         mx, my, (s->h_edge_pos >> 1),
+                                         s->v_edge_pos >> 1);
                 src = s->edge_emu_buffer;
             }
             if (thirdpel)
diff --git a/libavcodec/vc1dec.c b/libavcodec/vc1dec.c
index 3f10c85628..d5d6884a73 100644
--- a/libavcodec/vc1dec.c
+++ b/libavcodec/vc1dec.c
@@ -434,15 +434,15 @@ static void vc1_mc_1mv(VC1Context *v, int dir)
         uint8_t *uvbuf = s->edge_emu_buffer + 19 * s->linesize;
 
         srcY -= s->mspel * (1 + s->linesize);
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize,
-                                17 + s->mspel * 2, 17 + s->mspel * 2,
-                                src_x - s->mspel, src_y - s->mspel,
-                                s->h_edge_pos, v_edge_pos);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize,
+                                 17 + s->mspel * 2, 17 + s->mspel * 2,
+                                 src_x - s->mspel, src_y - s->mspel,
+                                 s->h_edge_pos, v_edge_pos);
         srcY = s->edge_emu_buffer;
-        s->dsp.emulated_edge_mc(uvbuf     , srcU, s->uvlinesize, 8 + 1, 8 + 1,
-                                uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
-        s->dsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, 8 + 1, 8 + 1,
-                                uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
+        s->vdsp.emulated_edge_mc(uvbuf     , srcU, s->uvlinesize, 8 + 1, 8 + 1,
+                                 uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
+        s->vdsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, 8 + 1, 8 + 1,
+                                 uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
         srcU = uvbuf;
         srcV = uvbuf + 16;
         /* if we deal with range reduction we need to scale source blocks */
@@ -667,10 +667,10 @@ static void vc1_mc_4mv_luma(VC1Context *v, int n, int dir)
         || (unsigned)(src_y - (s->mspel << fieldmv)) > v_edge_pos - (my & 3) - ((8 + s->mspel * 2) << fieldmv)) {
         srcY -= s->mspel * (1 + (s->linesize << fieldmv));
         /* check emulate edge stride and offset */
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize,
-                                9 + s->mspel * 2, (9 + s->mspel * 2) << fieldmv,
-                                src_x - s->mspel, src_y - (s->mspel << fieldmv),
-                                s->h_edge_pos, v_edge_pos);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize,
+                                 9 + s->mspel * 2, (9 + s->mspel * 2) << fieldmv,
+                                 src_x - s->mspel, src_y - (s->mspel << fieldmv),
+                                 s->h_edge_pos, v_edge_pos);
         srcY = s->edge_emu_buffer;
         /* if we deal with range reduction we need to scale source blocks */
         if (v->rangeredfrm) {
@@ -868,12 +868,12 @@ static void vc1_mc_4mv_chroma(VC1Context *v, int dir)
         || s->h_edge_pos < 18 || v_edge_pos < 18
         || (unsigned)uvsrc_x > (s->h_edge_pos >> 1) - 9
         || (unsigned)uvsrc_y > (v_edge_pos    >> 1) - 9) {
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer     , srcU, s->uvlinesize,
-                                8 + 1, 8 + 1, uvsrc_x, uvsrc_y,
-                                s->h_edge_pos >> 1, v_edge_pos >> 1);
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer + 16, srcV, s->uvlinesize,
-                                8 + 1, 8 + 1, uvsrc_x, uvsrc_y,
-                                s->h_edge_pos >> 1, v_edge_pos >> 1);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer     , srcU, s->uvlinesize,
+                                 8 + 1, 8 + 1, uvsrc_x, uvsrc_y,
+                                 s->h_edge_pos >> 1, v_edge_pos >> 1);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer + 16, srcV, s->uvlinesize,
+                                 8 + 1, 8 + 1, uvsrc_x, uvsrc_y,
+                                 s->h_edge_pos >> 1, v_edge_pos >> 1);
         srcU = s->edge_emu_buffer;
         srcV = s->edge_emu_buffer + 16;
 
@@ -973,12 +973,12 @@ static void vc1_mc_4mv_chroma4(VC1Context *v)
             || s->h_edge_pos < 10 || v_edge_pos < (5 << fieldmv)
             || (unsigned)uvsrc_x > (s->h_edge_pos >> 1) - 5
             || (unsigned)uvsrc_y > v_edge_pos - (5 << fieldmv)) {
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer, srcU, s->uvlinesize,
-                                    5, (5 << fieldmv), uvsrc_x, uvsrc_y,
-                                    s->h_edge_pos >> 1, v_edge_pos);
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer + 16, srcV, s->uvlinesize,
-                                    5, (5 << fieldmv), uvsrc_x, uvsrc_y,
-                                    s->h_edge_pos >> 1, v_edge_pos);
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcU, s->uvlinesize,
+                                     5, (5 << fieldmv), uvsrc_x, uvsrc_y,
+                                     s->h_edge_pos >> 1, v_edge_pos);
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer + 16, srcV, s->uvlinesize,
+                                     5, (5 << fieldmv), uvsrc_x, uvsrc_y,
+                                     s->h_edge_pos >> 1, v_edge_pos);
             srcU = s->edge_emu_buffer;
             srcV = s->edge_emu_buffer + 16;
 
@@ -1888,15 +1888,15 @@ static void vc1_interp_mc(VC1Context *v)
         uint8_t *uvbuf = s->edge_emu_buffer + 19 * s->linesize;
 
         srcY -= s->mspel * (1 + s->linesize);
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize,
-                                17 + s->mspel * 2, 17 + s->mspel * 2,
-                                src_x - s->mspel, src_y - s->mspel,
-                                s->h_edge_pos, v_edge_pos);
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize,
+                                 17 + s->mspel * 2, 17 + s->mspel * 2,
+                                 src_x - s->mspel, src_y - s->mspel,
+                                 s->h_edge_pos, v_edge_pos);
         srcY = s->edge_emu_buffer;
-        s->dsp.emulated_edge_mc(uvbuf     , srcU, s->uvlinesize, 8 + 1, 8 + 1,
-                                uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
-        s->dsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, 8 + 1, 8 + 1,
-                                uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
+        s->vdsp.emulated_edge_mc(uvbuf     , srcU, s->uvlinesize, 8 + 1, 8 + 1,
+                                 uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
+        s->vdsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, 8 + 1, 8 + 1,
+                                 uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1);
         srcU = uvbuf;
         srcV = uvbuf + 16;
         /* if we deal with range reduction we need to scale source blocks */
diff --git a/libavcodec/videodsp.c b/libavcodec/videodsp.c
new file mode 100644
index 0000000000..d14b0a13f1
--- /dev/null
+++ b/libavcodec/videodsp.c
@@ -0,0 +1,51 @@
+/*
+ * Copyright (C) 2012 Ronald S. Bultje
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/common.h"
+#include "videodsp.h"
+
+#define BIT_DEPTH 8
+#include "videodsp_template.c"
+#undef BIT_DEPTH
+
+#define BIT_DEPTH 16
+#include "videodsp_template.c"
+#undef BIT_DEPTH
+
+static void just_return(uint8_t *buf, ptrdiff_t stride, int h)
+{
+}
+
+void ff_videodsp_init(VideoDSPContext *ctx, int bpc)
+{
+    ctx->prefetch = just_return;
+    if (bpc <= 8) {
+        ctx->emulated_edge_mc = ff_emulated_edge_mc_8;
+    } else {
+        ctx->emulated_edge_mc = ff_emulated_edge_mc_16;
+    }
+
+    if (ARCH_ARM)
+        ff_videodsp_init_arm(ctx, bpc);
+    if (ARCH_PPC)
+        ff_videodsp_init_ppc(ctx, bpc);
+    if (ARCH_X86)
+        ff_videodsp_init_x86(ctx, bpc);
+}
diff --git a/libavcodec/videodsp.h b/libavcodec/videodsp.h
new file mode 100644
index 0000000000..7d785939f3
--- /dev/null
+++ b/libavcodec/videodsp.h
@@ -0,0 +1,71 @@
+/*
+ * Copyright (C) 2012 Ronald S. Bultje
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * Core video DSP helper functions
+ */
+
+#ifndef AVCODEC_VIDEODSP_H
+#define AVCODEC_VIDEODSP_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+typedef struct VideoDSPContext {
+    /**
+     * Copy a rectangular area of samples to a temporary buffer and replicate
+     * the border samples.
+     *
+     * @param buf destination buffer
+     * @param src source buffer
+     * @param linesize number of bytes between 2 vertically adjacent samples
+     *                 in both the source and destination buffers
+     * @param block_w width of block
+     * @param block_h height of block
+     * @param src_x x coordinate of the top left sample of the block in the
+     *                source buffer
+     * @param src_y y coordinate of the top left sample of the block in the
+     *                source buffer
+     * @param w width of the source buffer
+     * @param h height of the source buffer
+     */
+    void (*emulated_edge_mc)(uint8_t *buf, const uint8_t *src,
+                             ptrdiff_t linesize, int block_w, int block_h,
+                             int src_x, int src_y, int w, int h);
+
+    /**
+     * Prefetch memory into cache (if supported by hardware).
+     *
+     * @buf pointer to buffer to prefetch memory from
+     * @stride distance between two lines of buf (in bytes)
+     * @h number of lines to prefetch
+     */
+    void (*prefetch)(uint8_t *buf, ptrdiff_t stride, int h);
+} VideoDSPContext;
+
+void ff_videodsp_init(VideoDSPContext *ctx, int bpc);
+
+/* for internal use only (i.e. called by ff_videodsp_init() */
+void ff_videodsp_init_arm(VideoDSPContext *ctx, int bpc);
+void ff_videodsp_init_ppc(VideoDSPContext *ctx, int bpc);
+void ff_videodsp_init_x86(VideoDSPContext *ctx, int bpc);
+
+#endif /* AVCODEC_VIDEODSP_H */
diff --git a/libavcodec/videodsp_template.c b/libavcodec/videodsp_template.c
new file mode 100644
index 0000000000..5d5a32ff60
--- /dev/null
+++ b/libavcodec/videodsp_template.c
@@ -0,0 +1,93 @@
+/*
+ * Copyright (c) 2002-2004 Michael Niedermayer
+ * Copyright (C) 2012 Ronald S. Bultje
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "bit_depth_template.c"
+
+static void FUNC(ff_emulated_edge_mc)(uint8_t *buf, const uint8_t *src,
+                                      ptrdiff_t linesize,
+                                      int block_w, int block_h,
+                                      int src_x, int src_y, int w, int h)
+{
+    int x, y;
+    int start_y, start_x, end_y, end_x;
+
+    if (src_y >= h) {
+        src  += (h - 1 - src_y) * linesize;
+        src_y = h - 1;
+    } else if (src_y <= -block_h) {
+        src  += (1 - block_h - src_y) * linesize;
+        src_y = 1 - block_h;
+    }
+    if (src_x >= w) {
+        src  += (w - 1 - src_x) * sizeof(pixel);
+        src_x = w - 1;
+    } else if (src_x <= -block_w) {
+        src  += (1 - block_w - src_x) * sizeof(pixel);
+        src_x = 1 - block_w;
+    }
+
+    start_y = FFMAX(0, -src_y);
+    start_x = FFMAX(0, -src_x);
+    end_y = FFMIN(block_h, h-src_y);
+    end_x = FFMIN(block_w, w-src_x);
+    assert(start_y < end_y && block_h);
+    assert(start_x < end_x && block_w);
+
+    w    = end_x - start_x;
+    src += start_y * linesize + start_x * sizeof(pixel);
+    buf += start_x * sizeof(pixel);
+
+    // top
+    for (y = 0; y < start_y; y++) {
+        memcpy(buf, src, w * sizeof(pixel));
+        buf += linesize;
+    }
+
+    // copy existing part
+    for (; y < end_y; y++) {
+        memcpy(buf, src, w * sizeof(pixel));
+        src += linesize;
+        buf += linesize;
+    }
+
+    // bottom
+    src -= linesize;
+    for (; y < block_h; y++) {
+        memcpy(buf, src, w * sizeof(pixel));
+        buf += linesize;
+    }
+
+    buf -= block_h * linesize + start_x * sizeof(pixel);
+    while (block_h--) {
+        pixel *bufp = (pixel *) buf;
+
+        // left
+        for(x = 0; x < start_x; x++) {
+            bufp[x] = bufp[start_x];
+        }
+
+        // right
+        for (x = end_x; x < block_w; x++) {
+            bufp[x] = bufp[end_x - 1];
+        }
+        buf += linesize;
+    }
+}
diff --git a/libavcodec/vp3.c b/libavcodec/vp3.c
index de23f285e5..6e85b90964 100644
--- a/libavcodec/vp3.c
+++ b/libavcodec/vp3.c
@@ -38,7 +38,7 @@
 #include "internal.h"
 #include "dsputil.h"
 #include "get_bits.h"
-
+#include "videodsp.h"
 #include "vp3data.h"
 #include "vp3dsp.h"
 #include "xiph.h"
@@ -136,6 +136,7 @@ typedef struct Vp3DecodeContext {
     AVFrame current_frame;
     int keyframe;
     DSPContext dsp;
+    VideoDSPContext vdsp;
     VP3DSPContext vp3dsp;
     int flipped_image;
     int last_slice_end;
@@ -1543,7 +1544,7 @@ static void render_slice(Vp3DecodeContext *s, int slice)
                             uint8_t *temp= s->edge_emu_buffer;
                             if(stride<0) temp -= 8*stride;
 
-                            s->dsp.emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height);
+                            s->vdsp.emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height);
                             motion_source= temp;
                         }
                     }
@@ -1677,6 +1678,7 @@ static av_cold int vp3_decode_init(AVCodecContext *avctx)
         avctx->pix_fmt = AV_PIX_FMT_YUV420P;
     avctx->chroma_sample_location = AVCHROMA_LOC_CENTER;
     ff_dsputil_init(&s->dsp, avctx);
+    ff_videodsp_init(&s->vdsp, 8);
     ff_vp3dsp_init(&s->vp3dsp, avctx->flags);
 
     ff_init_scantable_permutation(s->dsp.idct_permutation, s->vp3dsp.idct_perm);
diff --git a/libavcodec/vp56.c b/libavcodec/vp56.c
index 5bd0a1abef..c3c9cb2914 100644
--- a/libavcodec/vp56.c
+++ b/libavcodec/vp56.c
@@ -340,7 +340,7 @@ static void vp56_mc(VP56Context *s, int b, int plane, uint8_t *src,
 
     if (x<0 || x+12>=s->plane_width[plane] ||
         y<0 || y+12>=s->plane_height[plane]) {
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
                             src + s->block_offset[b] + (dy-2)*stride + (dx-2),
                             stride, 12, 12, x, y,
                             s->plane_width[plane],
@@ -674,6 +674,7 @@ av_cold void ff_vp56_init(AVCodecContext *avctx, int flip, int has_alpha)
     avctx->pix_fmt = has_alpha ? AV_PIX_FMT_YUVA420P : AV_PIX_FMT_YUV420P;
 
     ff_dsputil_init(&s->dsp, avctx);
+    ff_videodsp_init(&s->vdsp, 8);
     ff_vp3dsp_init(&s->vp3dsp, avctx->flags);
     ff_vp56dsp_init(&s->vp56dsp, avctx->codec->id);
     ff_init_scantable_permutation(s->dsp.idct_permutation, s->vp3dsp.idct_perm);
diff --git a/libavcodec/vp56.h b/libavcodec/vp56.h
index 9be2eabdf9..431d1a9465 100644
--- a/libavcodec/vp56.h
+++ b/libavcodec/vp56.h
@@ -30,6 +30,7 @@
 #include "dsputil.h"
 #include "get_bits.h"
 #include "bytestream.h"
+#include "videodsp.h"
 #include "vp3dsp.h"
 #include "vp56dsp.h"
 
@@ -94,6 +95,7 @@ typedef struct VP56Model {
 struct vp56_context {
     AVCodecContext *avctx;
     DSPContext dsp;
+    VideoDSPContext vdsp;
     VP3DSPContext vp3dsp;
     VP56DSPContext vp56dsp;
     ScanTable scantable;
diff --git a/libavcodec/vp8.c b/libavcodec/vp8.c
index 8044e8db33..3b8f7d2a24 100644
--- a/libavcodec/vp8.c
+++ b/libavcodec/vp8.c
@@ -1198,9 +1198,9 @@ void vp8_mc_luma(VP8Context *s, VP8ThreadData *td, uint8_t *dst,
         src += y_off * linesize + x_off;
         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||
             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {
-            s->dsp.emulated_edge_mc(td->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,
-                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
-                                    x_off - mx_idx, y_off - my_idx, width, height);
+            s->vdsp.emulated_edge_mc(td->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,
+                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
+                                     x_off - mx_idx, y_off - my_idx, width, height);
             src = td->edge_emu_buffer + mx_idx + linesize * my_idx;
         }
         mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
@@ -1248,15 +1248,15 @@ void vp8_mc_chroma(VP8Context *s, VP8ThreadData *td, uint8_t *dst1, uint8_t *dst
         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 3, 0);
         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||
             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {
-            s->dsp.emulated_edge_mc(td->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,
-                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
-                                    x_off - mx_idx, y_off - my_idx, width, height);
+            s->vdsp.emulated_edge_mc(td->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,
+                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
+                                     x_off - mx_idx, y_off - my_idx, width, height);
             src1 = td->edge_emu_buffer + mx_idx + linesize * my_idx;
             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);
 
-            s->dsp.emulated_edge_mc(td->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,
-                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
-                                    x_off - mx_idx, y_off - my_idx, width, height);
+            s->vdsp.emulated_edge_mc(td->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,
+                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
+                                     x_off - mx_idx, y_off - my_idx, width, height);
             src2 = td->edge_emu_buffer + mx_idx + linesize * my_idx;
             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);
         } else {
@@ -1315,9 +1315,9 @@ static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, i
         /* For threading, a ff_thread_await_progress here might be useful, but
          * it actually slows down the decoder. Since a bad prefetch doesn't
          * generate bad decoder output, we don't run it here. */
-        s->dsp.prefetch(src[0]+off, s->linesize, 4);
+        s->vdsp.prefetch(src[0]+off, s->linesize, 4);
         off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
-        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
+        s->vdsp.prefetch(src[1]+off, src[2]-src[1], 2);
     }
 }
 
@@ -1716,8 +1716,8 @@ static void vp8_decode_mb_row_no_filter(AVCodecContext *avctx, void *tdata,
             }
         }
 
-        s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
-        s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
+        s->vdsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
+        s->vdsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
 
         if (!s->mb_layout)
             decode_mb_mode(s, mb, mb_x, mb_y, curframe->ref_index[0] + mb_xy,
@@ -2020,7 +2020,7 @@ static av_cold int vp8_decode_init(AVCodecContext *avctx)
     s->avctx = avctx;
     avctx->pix_fmt = AV_PIX_FMT_YUV420P;
 
-    ff_dsputil_init(&s->dsp, avctx);
+    ff_videodsp_init(&s->vdsp, 8);
     ff_h264_pred_init(&s->hpc, AV_CODEC_ID_VP8, 8, 1);
     ff_vp8dsp_init(&s->vp8dsp);
 
diff --git a/libavcodec/vp8.h b/libavcodec/vp8.h
index 073a0335e0..4a85541f37 100644
--- a/libavcodec/vp8.h
+++ b/libavcodec/vp8.h
@@ -247,7 +247,7 @@ typedef struct VP8Context {
      */
     int num_coeff_partitions;
     VP56RangeCoder coeff_partition[8];
-    DSPContext dsp;
+    VideoDSPContext vdsp;
     VP8DSPContext vp8dsp;
     H264PredContext hpc;
     vp8_mc_func put_pixels_tab[3][3][3];
diff --git a/libavcodec/wmv2.c b/libavcodec/wmv2.c
index f113790e7d..a3dcbb37ae 100644
--- a/libavcodec/wmv2.c
+++ b/libavcodec/wmv2.c
@@ -102,7 +102,7 @@ void ff_mspel_motion(MpegEncContext *s,
     if(s->flags&CODEC_FLAG_EMU_EDGE){
         if(src_x<1 || src_y<1 || src_x + 17  >= s->h_edge_pos
                               || src_y + h+1 >= v_edge_pos){
-            s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr - 1 - s->linesize, s->linesize, 19, 19,
+            s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr - 1 - s->linesize, s->linesize, 19, 19,
                              src_x-1, src_y-1, s->h_edge_pos, s->v_edge_pos);
             ptr= s->edge_emu_buffer + 1 + s->linesize;
             emu=1;
@@ -143,7 +143,7 @@ void ff_mspel_motion(MpegEncContext *s,
     offset = (src_y * uvlinesize) + src_x;
     ptr = ref_picture[1] + offset;
     if(emu){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, 9, 9,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, 9, 9,
                          src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1);
         ptr= s->edge_emu_buffer;
     }
@@ -151,7 +151,7 @@ void ff_mspel_motion(MpegEncContext *s,
 
     ptr = ref_picture[2] + offset;
     if(emu){
-        s->dsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, 9, 9,
+        s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, 9, 9,
                          src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1);
         ptr= s->edge_emu_buffer;
     }
diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index 1e70d535f5..4fc7ac0514 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -19,6 +19,7 @@ OBJS-$(CONFIG_RV40_DECODER)            += x86/rv34dsp_init.o            \
                                           x86/rv40dsp_init.o
 OBJS-$(CONFIG_TRUEHD_DECODER)          += x86/mlpdsp.o
 OBJS-$(CONFIG_VC1_DECODER)             += x86/vc1dsp_init.o
+OBJS-$(CONFIG_VIDEODSP)                += x86/videodsp_init.o
 OBJS-$(CONFIG_VP3DSP)                  += x86/vp3dsp_init.o
 OBJS-$(CONFIG_VP5_DECODER)             += x86/vp56dsp_init.o
 OBJS-$(CONFIG_VP6_DECODER)             += x86/vp56dsp_init.o
@@ -60,6 +61,7 @@ YASM-OBJS-$(CONFIG_RV30_DECODER)       += x86/rv34dsp.o
 YASM-OBJS-$(CONFIG_RV40_DECODER)       += x86/rv34dsp.o                 \
                                           x86/rv40dsp.o
 YASM-OBJS-$(CONFIG_VC1_DECODER)        += x86/vc1dsp.o
+YASM-OBJS-$(CONFIG_VIDEODSP)           += x86/videodsp.o
 YASM-OBJS-$(CONFIG_VP3DSP)             += x86/vp3dsp.o
 YASM-OBJS-$(CONFIG_VP6_DECODER)        += x86/vp56dsp.o
 YASM-OBJS-$(CONFIG_VP8_DECODER)        += x86/vp8dsp.o
diff --git a/libavcodec/x86/dsputil.asm b/libavcodec/x86/dsputil.asm
index dd0c795235..b9990bb1ba 100644
--- a/libavcodec/x86/dsputil.asm
+++ b/libavcodec/x86/dsputil.asm
@@ -489,577 +489,6 @@ cglobal scalarproduct_float, 3,3,2, v1, v2, offset
 %endif
     RET
 
-; extern void ff_emu_edge_core(uint8_t *buf, const uint8_t *src, x86_reg linesize,
-;                              x86_reg start_y, x86_reg end_y, x86_reg block_h,
-;                              x86_reg start_x, x86_reg end_x, x86_reg block_w);
-;
-; The actual function itself is below. It basically wraps a very simple
-; w = end_x - start_x
-; if (w) {
-;   if (w > 22) {
-;     jump to the slow loop functions
-;   } else {
-;     jump to the fast loop functions
-;   }
-; }
-;
-; ... and then the same for left/right extend also. See below for loop
-; function implementations. Fast are fixed-width, slow is variable-width
-
-%macro EMU_EDGE_FUNC 0
-%if ARCH_X86_64
-%define w_reg r7
-cglobal emu_edge_core, 6, 9, 1
-    mov         r8, r5          ; save block_h
-%else
-%define w_reg r6
-cglobal emu_edge_core, 2, 7, 0
-    mov         r4, r4m         ; end_y
-    mov         r5, r5m         ; block_h
-%endif
-
-    ; start with vertical extend (top/bottom) and body pixel copy
-    mov      w_reg, r7m
-    sub      w_reg, r6m         ; w = start_x - end_x
-    sub         r5, r4
-%if ARCH_X86_64
-    sub         r4, r3
-%else
-    sub         r4, dword r3m
-%endif
-    cmp      w_reg, 22
-    jg .slow_v_extend_loop
-%if ARCH_X86_32
-    mov         r2, r2m         ; linesize
-%endif
-    sal      w_reg, 7           ; w * 128
-%ifdef PIC
-    lea        rax, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)]
-    add      w_reg, rax
-%else
-    lea      w_reg, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)+w_reg]
-%endif
-    call     w_reg              ; fast top extend, body copy and bottom extend
-.v_extend_end:
-
-    ; horizontal extend (left/right)
-    mov      w_reg, r6m         ; start_x
-    sub         r0, w_reg
-%if ARCH_X86_64
-    mov         r3, r0          ; backup of buf+block_h*linesize
-    mov         r5, r8
-%else
-    mov        r0m, r0          ; backup of buf+block_h*linesize
-    mov         r5, r5m
-%endif
-    test     w_reg, w_reg
-    jz .right_extend
-    cmp      w_reg, 22
-    jg .slow_left_extend_loop
-    mov         r1, w_reg
-    dec      w_reg
-    ; FIXME we can do a if size == 1 here if that makes any speed difference, test me
-    sar      w_reg, 1
-    sal      w_reg, 6
-    ; r0=buf+block_h*linesize,r7(64)/r6(32)=start_x offset for funcs
-    ; r6(rax)/r3(ebx)=val,r2=linesize,r1=start_x,r5=block_h
-%ifdef PIC
-    lea        rax, [.emuedge_extend_left_2]
-    add      w_reg, rax
-%else
-    lea      w_reg, [.emuedge_extend_left_2+w_reg]
-%endif
-    call     w_reg
-
-    ; now r3(64)/r0(32)=buf,r2=linesize,r8/r5=block_h,r6/r3=val, r7/r6=end_x, r1=block_w
-.right_extend:
-%if ARCH_X86_32
-    mov         r0, r0m
-    mov         r5, r5m
-%endif
-    mov      w_reg, r7m         ; end_x
-    mov         r1, r8m         ; block_w
-    mov         r4, r1
-    sub         r1, w_reg
-    jz .h_extend_end            ; if (end_x == block_w) goto h_extend_end
-    cmp         r1, 22
-    jg .slow_right_extend_loop
-    dec         r1
-    ; FIXME we can do a if size == 1 here if that makes any speed difference, test me
-    sar         r1, 1
-    sal         r1, 6
-%ifdef PIC
-    lea        rax, [.emuedge_extend_right_2]
-    add         r1, rax
-%else
-    lea         r1, [.emuedge_extend_right_2+r1]
-%endif
-    call        r1
-.h_extend_end:
-    RET
-
-%if ARCH_X86_64
-%define vall  al
-%define valh  ah
-%define valw  ax
-%define valw2 r7w
-%define valw3 r3w
-%if WIN64
-%define valw4 r7w
-%else ; unix64
-%define valw4 r3w
-%endif
-%define vald eax
-%else
-%define vall  bl
-%define valh  bh
-%define valw  bx
-%define valw2 r6w
-%define valw3 valw2
-%define valw4 valw3
-%define vald ebx
-%define stack_offset 0x14
-%endif
-
-%endmacro
-
-; macro to read/write a horizontal number of pixels (%2) to/from registers
-; on x86-64, - fills xmm0-15 for consecutive sets of 16 pixels
-;            - if (%2 & 15 == 8) fills the last 8 bytes into rax
-;            - else if (%2 & 8)  fills 8 bytes into mm0
-;            - if (%2 & 7 == 4)  fills the last 4 bytes into rax
-;            - else if (%2 & 4)  fills 4 bytes into mm0-1
-;            - if (%2 & 3 == 3)  fills 2 bytes into r7/r3, and 1 into eax
-;              (note that we're using r3 for body/bottom because it's a shorter
-;               opcode, and then the loop fits in 128 bytes)
-;            - else              fills remaining bytes into rax
-; on x86-32, - fills mm0-7 for consecutive sets of 8 pixels
-;            - if (%2 & 7 == 4)  fills 4 bytes into ebx
-;            - else if (%2 & 4)  fills 4 bytes into mm0-7
-;            - if (%2 & 3 == 3)  fills 2 bytes into r6, and 1 into ebx
-;            - else              fills remaining bytes into ebx
-; writing data out is in the same way
-%macro READ_NUM_BYTES 2
-%assign %%src_off 0 ; offset in source buffer
-%assign %%smidx   0 ; mmx register idx
-%assign %%sxidx   0 ; xmm register idx
-
-%if cpuflag(sse)
-%rep %2/16
-    movups xmm %+ %%sxidx, [r1+%%src_off]
-%assign %%src_off %%src_off+16
-%assign %%sxidx   %%sxidx+1
-%endrep ; %2/16
-%endif
-
-%if ARCH_X86_64
-%if (%2-%%src_off) == 8
-    mov           rax, [r1+%%src_off]
-%assign %%src_off %%src_off+8
-%endif ; (%2-%%src_off) == 8
-%endif ; x86-64
-
-%rep (%2-%%src_off)/8
-    movq    mm %+ %%smidx, [r1+%%src_off]
-%assign %%src_off %%src_off+8
-%assign %%smidx   %%smidx+1
-%endrep ; (%2-%%dst_off)/8
-
-%if (%2-%%src_off) == 4
-    mov          vald, [r1+%%src_off]
-%elif (%2-%%src_off) & 4
-    movd    mm %+ %%smidx, [r1+%%src_off]
-%assign %%src_off %%src_off+4
-%endif ; (%2-%%src_off) ==/& 4
-
-%if (%2-%%src_off) == 1
-    mov          vall, [r1+%%src_off]
-%elif (%2-%%src_off) == 2
-    mov          valw, [r1+%%src_off]
-%elif (%2-%%src_off) == 3
-%ifidn %1, top
-    mov         valw2, [r1+%%src_off]
-%elifidn %1, body
-    mov         valw3, [r1+%%src_off]
-%elifidn %1, bottom
-    mov         valw4, [r1+%%src_off]
-%endif ; %1 ==/!= top
-    mov          vall, [r1+%%src_off+2]
-%endif ; (%2-%%src_off) == 1/2/3
-%endmacro ; READ_NUM_BYTES
-
-%macro WRITE_NUM_BYTES 2
-%assign %%dst_off 0 ; offset in destination buffer
-%assign %%dmidx   0 ; mmx register idx
-%assign %%dxidx   0 ; xmm register idx
-
-%if cpuflag(sse)
-%rep %2/16
-    movups [r0+%%dst_off], xmm %+ %%dxidx
-%assign %%dst_off %%dst_off+16
-%assign %%dxidx   %%dxidx+1
-%endrep ; %2/16
-%endif
-
-%if ARCH_X86_64
-%if (%2-%%dst_off) == 8
-    mov    [r0+%%dst_off], rax
-%assign %%dst_off %%dst_off+8
-%endif ; (%2-%%dst_off) == 8
-%endif ; x86-64
-
-%rep (%2-%%dst_off)/8
-    movq   [r0+%%dst_off], mm %+ %%dmidx
-%assign %%dst_off %%dst_off+8
-%assign %%dmidx   %%dmidx+1
-%endrep ; (%2-%%dst_off)/8
-
-%if (%2-%%dst_off) == 4
-    mov    [r0+%%dst_off], vald
-%elif (%2-%%dst_off) & 4
-    movd   [r0+%%dst_off], mm %+ %%dmidx
-%assign %%dst_off %%dst_off+4
-%endif ; (%2-%%dst_off) ==/& 4
-
-%if (%2-%%dst_off) == 1
-    mov    [r0+%%dst_off], vall
-%elif (%2-%%dst_off) == 2
-    mov    [r0+%%dst_off], valw
-%elif (%2-%%dst_off) == 3
-%ifidn %1, top
-    mov    [r0+%%dst_off], valw2
-%elifidn %1, body
-    mov    [r0+%%dst_off], valw3
-%elifidn %1, bottom
-    mov    [r0+%%dst_off], valw4
-%endif ; %1 ==/!= top
-    mov  [r0+%%dst_off+2], vall
-%endif ; (%2-%%dst_off) == 1/2/3
-%endmacro ; WRITE_NUM_BYTES
-
-; vertical top/bottom extend and body copy fast loops
-; these are function pointers to set-width line copy functions, i.e.
-; they read a fixed number of pixels into set registers, and write
-; those out into the destination buffer
-; r0=buf,r1=src,r2=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h
-; r6(eax/64)/r3(ebx/32)=val_reg
-%macro VERTICAL_EXTEND 0
-%assign %%n 1
-%rep 22
-ALIGN 128
-.emuedge_v_extend_ %+ %%n:
-    ; extend pixels above body
-%if ARCH_X86_64
-    test           r3 , r3                   ; if (!start_y)
-    jz .emuedge_copy_body_ %+ %%n %+ _loop   ;   goto body
-%else ; ARCH_X86_32
-    cmp      dword r3m, 0
-    je .emuedge_copy_body_ %+ %%n %+ _loop
-%endif ; ARCH_X86_64/32
-    READ_NUM_BYTES  top,    %%n              ; read bytes
-.emuedge_extend_top_ %+ %%n %+ _loop:        ; do {
-    WRITE_NUM_BYTES top,    %%n              ;   write bytes
-    add            r0 , r2                   ;   dst += linesize
-%if ARCH_X86_64
-    dec            r3d
-%else ; ARCH_X86_32
-    dec      dword r3m
-%endif ; ARCH_X86_64/32
-    jnz .emuedge_extend_top_ %+ %%n %+ _loop ; } while (--start_y)
-
-    ; copy body pixels
-.emuedge_copy_body_ %+ %%n %+ _loop:         ; do {
-    READ_NUM_BYTES  body,   %%n              ;   read bytes
-    WRITE_NUM_BYTES body,   %%n              ;   write bytes
-    add            r0 , r2                   ;   dst += linesize
-    add            r1 , r2                   ;   src += linesize
-    dec            r4d
-    jnz .emuedge_copy_body_ %+ %%n %+ _loop  ; } while (--end_y)
-
-    ; copy bottom pixels
-    test           r5 , r5                   ; if (!block_h)
-    jz .emuedge_v_extend_end_ %+ %%n         ;   goto end
-    sub            r1 , r2                   ; src -= linesize
-    READ_NUM_BYTES  bottom, %%n              ; read bytes
-.emuedge_extend_bottom_ %+ %%n %+ _loop:     ; do {
-    WRITE_NUM_BYTES bottom, %%n              ;   write bytes
-    add            r0 , r2                   ;   dst += linesize
-    dec            r5d
-    jnz .emuedge_extend_bottom_ %+ %%n %+ _loop ; } while (--block_h)
-
-.emuedge_v_extend_end_ %+ %%n:
-%if ARCH_X86_64
-    ret
-%else ; ARCH_X86_32
-    rep ret
-%endif ; ARCH_X86_64/32
-%assign %%n %%n+1
-%endrep
-%endmacro VERTICAL_EXTEND
-
-; left/right (horizontal) fast extend functions
-; these are essentially identical to the vertical extend ones above,
-; just left/right separated because number of pixels to extend is
-; obviously not the same on both sides.
-; for reading, pixels are placed in eax (x86-64) or ebx (x86-64) in the
-; lowest two bytes of the register (so val*0x0101), and are splatted
-; into each byte of mm0 as well if n_pixels >= 8
-
-%macro READ_V_PIXEL 2
-    mov        vall, %2
-    mov        valh, vall
-%if %1 >= 8
-    movd        mm0, vald
-%if cpuflag(mmxext)
-    pshufw      mm0, mm0, 0
-%else ; mmx
-    punpcklwd   mm0, mm0
-    punpckldq   mm0, mm0
-%endif ; sse
-%endif ; %1 >= 8
-%endmacro
-
-%macro WRITE_V_PIXEL 2
-%assign %%dst_off 0
-%rep %1/8
-    movq [%2+%%dst_off], mm0
-%assign %%dst_off %%dst_off+8
-%endrep
-%if %1 & 4
-%if %1 >= 8
-    movd [%2+%%dst_off], mm0
-%else ; %1 < 8
-    mov  [%2+%%dst_off]  , valw
-    mov  [%2+%%dst_off+2], valw
-%endif ; %1 >=/< 8
-%assign %%dst_off %%dst_off+4
-%endif ; %1 & 4
-%if %1&2
-    mov  [%2+%%dst_off], valw
-%endif ; %1 & 2
-%endmacro
-
-; r0=buf+block_h*linesize, r1=start_x, r2=linesize, r5=block_h, r6/r3=val
-%macro LEFT_EXTEND 0
-%assign %%n 2
-%rep 11
-ALIGN 64
-.emuedge_extend_left_ %+ %%n:          ; do {
-    sub         r0, r2                 ;   dst -= linesize
-    READ_V_PIXEL  %%n, [r0+r1]         ;   read pixels
-    WRITE_V_PIXEL %%n, r0              ;   write pixels
-    dec         r5
-    jnz .emuedge_extend_left_ %+ %%n   ; } while (--block_h)
-%if ARCH_X86_64
-    ret
-%else ; ARCH_X86_32
-    rep ret
-%endif ; ARCH_X86_64/32
-%assign %%n %%n+2
-%endrep
-%endmacro ; LEFT_EXTEND
-
-; r3/r0=buf+block_h*linesize, r2=linesize, r8/r5=block_h, r0/r6=end_x, r6/r3=val
-%macro RIGHT_EXTEND 0
-%assign %%n 2
-%rep 11
-ALIGN 64
-.emuedge_extend_right_ %+ %%n:          ; do {
-%if ARCH_X86_64
-    sub        r3, r2                   ;   dst -= linesize
-    READ_V_PIXEL  %%n, [r3+w_reg-1]     ;   read pixels
-    WRITE_V_PIXEL %%n, r3+r4-%%n        ;   write pixels
-    dec       r8
-%else ; ARCH_X86_32
-    sub        r0, r2                   ;   dst -= linesize
-    READ_V_PIXEL  %%n, [r0+w_reg-1]     ;   read pixels
-    WRITE_V_PIXEL %%n, r0+r4-%%n        ;   write pixels
-    dec     r5
-%endif ; ARCH_X86_64/32
-    jnz .emuedge_extend_right_ %+ %%n   ; } while (--block_h)
-%if ARCH_X86_64
-    ret
-%else ; ARCH_X86_32
-    rep ret
-%endif ; ARCH_X86_64/32
-%assign %%n %%n+2
-%endrep
-
-%if ARCH_X86_32
-%define stack_offset 0x10
-%endif
-%endmacro ; RIGHT_EXTEND
-
-; below follow the "slow" copy/extend functions, these act on a non-fixed
-; width specified in a register, and run a loop to copy the full amount
-; of bytes. They are optimized for copying of large amounts of pixels per
-; line, so they unconditionally splat data into mm registers to copy 8
-; bytes per loop iteration. It could be considered to use xmm for x86-64
-; also, but I haven't optimized this as much (i.e. FIXME)
-%macro V_COPY_NPX 4-5
-%if %0 == 4
-    test     w_reg, %4
-    jz .%1_skip_%4_px
-%else ; %0 == 5
-.%1_%4_px_loop:
-%endif
-    %3          %2, [r1+cnt_reg]
-    %3 [r0+cnt_reg], %2
-    add    cnt_reg, %4
-%if %0 == 5
-    sub      w_reg, %4
-    test     w_reg, %5
-    jnz .%1_%4_px_loop
-%endif
-.%1_skip_%4_px:
-%endmacro
-
-%macro V_COPY_ROW 2
-%ifidn %1, bottom
-    sub         r1, linesize
-%endif
-.%1_copy_loop:
-    xor    cnt_reg, cnt_reg
-%if notcpuflag(sse)
-%define linesize r2m
-    V_COPY_NPX %1,  mm0, movq,    8, 0xFFFFFFF8
-%else ; sse
-    V_COPY_NPX %1, xmm0, movups, 16, 0xFFFFFFF0
-%if ARCH_X86_64
-%define linesize r2
-    V_COPY_NPX %1, rax , mov,     8
-%else ; ARCH_X86_32
-%define linesize r2m
-    V_COPY_NPX %1,  mm0, movq,    8
-%endif ; ARCH_X86_64/32
-%endif ; sse
-    V_COPY_NPX %1, vald, mov,     4
-    V_COPY_NPX %1, valw, mov,     2
-    V_COPY_NPX %1, vall, mov,     1
-    mov      w_reg, cnt_reg
-%ifidn %1, body
-    add         r1, linesize
-%endif
-    add         r0, linesize
-    dec         %2
-    jnz .%1_copy_loop
-%endmacro
-
-%macro SLOW_V_EXTEND 0
-.slow_v_extend_loop:
-; r0=buf,r1=src,r2(64)/r2m(32)=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h
-; r8(64)/r3(later-64)/r2(32)=cnt_reg,r6(64)/r3(32)=val_reg,r7(64)/r6(32)=w=end_x-start_x
-%if ARCH_X86_64
-    push        r8              ; save old value of block_h
-    test        r3, r3
-%define cnt_reg r8
-    jz .do_body_copy            ; if (!start_y) goto do_body_copy
-    V_COPY_ROW top, r3
-%else
-    cmp  dword r3m, 0
-%define cnt_reg r2
-    je .do_body_copy            ; if (!start_y) goto do_body_copy
-    V_COPY_ROW top, dword r3m
-%endif
-
-.do_body_copy:
-    V_COPY_ROW body, r4
-
-%if ARCH_X86_64
-    pop         r8              ; restore old value of block_h
-%define cnt_reg r3
-%endif
-    test        r5, r5
-%if ARCH_X86_64
-    jz .v_extend_end
-%else
-    jz .skip_bottom_extend
-%endif
-    V_COPY_ROW bottom, r5
-%if ARCH_X86_32
-.skip_bottom_extend:
-    mov         r2, r2m
-%endif
-    jmp .v_extend_end
-%endmacro
-
-%macro SLOW_LEFT_EXTEND 0
-.slow_left_extend_loop:
-; r0=buf+block_h*linesize,r2=linesize,r6(64)/r3(32)=val,r5=block_h,r4=cntr,r7/r6=start_x
-    mov         r4, 8
-    sub         r0, linesize
-    READ_V_PIXEL 8, [r0+w_reg]
-.left_extend_8px_loop:
-    movq [r0+r4-8], mm0
-    add         r4, 8
-    cmp         r4, w_reg
-    jle .left_extend_8px_loop
-    sub         r4, 8
-    cmp         r4, w_reg
-    jge .left_extend_loop_end
-.left_extend_2px_loop:
-    mov    [r0+r4], valw
-    add         r4, 2
-    cmp         r4, w_reg
-    jl .left_extend_2px_loop
-.left_extend_loop_end:
-    dec         r5
-    jnz .slow_left_extend_loop
-%if ARCH_X86_32
-    mov         r2, r2m
-%endif
-    jmp .right_extend
-%endmacro
-
-%macro SLOW_RIGHT_EXTEND 0
-.slow_right_extend_loop:
-; r3(64)/r0(32)=buf+block_h*linesize,r2=linesize,r4=block_w,r8(64)/r5(32)=block_h,
-; r7(64)/r6(32)=end_x,r6/r3=val,r1=cntr
-%if ARCH_X86_64
-%define buf_reg r3
-%define bh_reg r8
-%else
-%define buf_reg r0
-%define bh_reg r5
-%endif
-    lea         r1, [r4-8]
-    sub    buf_reg, linesize
-    READ_V_PIXEL 8, [buf_reg+w_reg-1]
-.right_extend_8px_loop:
-    movq [buf_reg+r1], mm0
-    sub         r1, 8
-    cmp         r1, w_reg
-    jge .right_extend_8px_loop
-    add         r1, 8
-    cmp         r1, w_reg
-    je .right_extend_loop_end
-.right_extend_2px_loop:
-    sub         r1, 2
-    mov [buf_reg+r1], valw
-    cmp         r1, w_reg
-    jg .right_extend_2px_loop
-.right_extend_loop_end:
-    dec         bh_reg
-    jnz .slow_right_extend_loop
-    jmp .h_extend_end
-%endmacro
-
-%macro emu_edge 1
-INIT_XMM %1
-EMU_EDGE_FUNC
-VERTICAL_EXTEND
-LEFT_EXTEND
-RIGHT_EXTEND
-SLOW_V_EXTEND
-SLOW_LEFT_EXTEND
-SLOW_RIGHT_EXTEND
-%endmacro
-
-emu_edge sse
-%if ARCH_X86_32
-emu_edge mmx
-%endif
-
 ;-----------------------------------------------------------------------------
 ; void ff_vector_clip_int32(int32_t *dst, const int32_t *src, int32_t min,
 ;                           int32_t max, unsigned int len)
diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c
index 58b6389753..b16f7e4de3 100644
--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -1635,78 +1635,6 @@ void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
   avg_pixels16_xy2_mmx(dst, src, stride, 16);
 }
 
-#endif /* HAVE_INLINE_ASM */
-
-#if HAVE_YASM
-typedef void emu_edge_core_func(uint8_t *buf, const uint8_t *src,
-                                x86_reg linesize, x86_reg start_y,
-                                x86_reg end_y, x86_reg block_h,
-                                x86_reg start_x, x86_reg end_x,
-                                x86_reg block_w);
-extern emu_edge_core_func ff_emu_edge_core_mmx;
-extern emu_edge_core_func ff_emu_edge_core_sse;
-
-static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src,
-                                              int linesize,
-                                              int block_w, int block_h,
-                                              int src_x, int src_y,
-                                              int w, int h,
-                                              emu_edge_core_func *core_fn)
-{
-    int start_y, start_x, end_y, end_x, src_y_add = 0;
-
-    if (src_y >= h) {
-        src_y_add = h - 1 - src_y;
-        src_y     = h - 1;
-    } else if (src_y <= -block_h) {
-        src_y_add = 1 - block_h - src_y;
-        src_y     = 1 - block_h;
-    }
-    if (src_x >= w) {
-        src   += w - 1 - src_x;
-        src_x  = w - 1;
-    } else if (src_x <= -block_w) {
-        src   += 1 - block_w - src_x;
-        src_x  = 1 - block_w;
-    }
-
-    start_y = FFMAX(0, -src_y);
-    start_x = FFMAX(0, -src_x);
-    end_y   = FFMIN(block_h, h-src_y);
-    end_x   = FFMIN(block_w, w-src_x);
-    assert(start_x < end_x && block_w > 0);
-    assert(start_y < end_y && block_h > 0);
-
-    // fill in the to-be-copied part plus all above/below
-    src += (src_y_add + start_y) * linesize + start_x;
-    buf += start_x;
-    core_fn(buf, src, linesize, start_y, end_y,
-            block_h, start_x, end_x, block_w);
-}
-
-#if ARCH_X86_32
-static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src,
-                                             int linesize,
-                                             int block_w, int block_h,
-                                             int src_x, int src_y, int w, int h)
-{
-    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
-                     w, h, &ff_emu_edge_core_mmx);
-}
-#endif
-
-static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
-                                             int linesize,
-                                             int block_w, int block_h,
-                                             int src_x, int src_y, int w, int h)
-{
-    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
-                     w, h, &ff_emu_edge_core_sse);
-}
-#endif /* HAVE_YASM */
-
-#if HAVE_INLINE_ASM
-
 static void gmc_mmx(uint8_t *dst, uint8_t *src,
                     int stride, int h, int ox, int oy,
                     int dxx, int dxy, int dyx, int dyy,
@@ -1822,21 +1750,6 @@ static void gmc_mmx(uint8_t *dst, uint8_t *src,
         src += 4 - h * stride;
     }
 }
-
-#define PREFETCH(name, op)                      \
-static void name(void *mem, int stride, int h)  \
-{                                               \
-    const uint8_t *p = mem;                     \
-    do {                                        \
-        __asm__ volatile (#op" %0" :: "m"(*p)); \
-        p += stride;                            \
-    } while (--h);                              \
-}
-
-PREFETCH(prefetch_mmxext, prefetcht0)
-PREFETCH(prefetch_3dnow, prefetch)
-#undef PREFETCH
-
 #endif /* HAVE_INLINE_ASM */
 
 #include "h264_qpel.c"
@@ -2239,11 +2152,6 @@ static void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
 #endif /* HAVE_INLINE_ASM */
 
 #if HAVE_YASM
-#if ARCH_X86_32
-    if (!high_bit_depth)
-        c->emulated_edge_mc = emulated_edge_mc_mmx;
-#endif
-
     if (!high_bit_depth && CONFIG_H264CHROMA) {
         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_mmx;
         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_mmx;
@@ -2261,8 +2169,6 @@ static void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
     const int high_bit_depth = bit_depth > 8;
 
 #if HAVE_INLINE_ASM
-    c->prefetch = prefetch_mmxext;
-
     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
     SET_QPEL_FUNCS(avg_2tap_qpel,   0, 16, mmxext, );
@@ -2371,8 +2277,6 @@ static void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
 
 #if HAVE_INLINE_ASM
-    c->prefetch = prefetch_3dnow;
-
     if (!high_bit_depth) {
         c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
         c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
@@ -2452,9 +2356,6 @@ static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
 
     c->scalarproduct_float          = ff_scalarproduct_float_sse;
     c->butterflies_float_interleave = ff_butterflies_float_interleave_sse;
-
-    if (!high_bit_depth)
-        c->emulated_edge_mc = emulated_edge_mc_sse;
 #endif /* HAVE_YASM */
 }
 
diff --git a/libavcodec/x86/videodsp.asm b/libavcodec/x86/videodsp.asm
new file mode 100644
index 0000000000..19b910b923
--- /dev/null
+++ b/libavcodec/x86/videodsp.asm
@@ -0,0 +1,612 @@
+;******************************************************************************
+;* Core video DSP functions
+;* Copyright (c) 2012 Ronald S. Bultje <rsbultje@gmail.com>
+;*
+;* This file is part of Libav.
+;*
+;* Libav is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 of the License, or (at your option) any later version.
+;*
+;* Libav 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
+;* Lesser General Public License for more details.
+;*
+;* You should have received a copy of the GNU Lesser General Public
+;* License along with Libav; if not, write to the Free Software
+;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+;******************************************************************************
+
+%include "libavutil/x86/x86util.asm"
+
+SECTION .text
+
+; extern void ff_emu_edge_core(uint8_t *buf, const uint8_t *src, x86_reg linesize,
+;                              x86_reg start_y, x86_reg end_y, x86_reg block_h,
+;                              x86_reg start_x, x86_reg end_x, x86_reg block_w);
+;
+; The actual function itself is below. It basically wraps a very simple
+; w = end_x - start_x
+; if (w) {
+;   if (w > 22) {
+;     jump to the slow loop functions
+;   } else {
+;     jump to the fast loop functions
+;   }
+; }
+;
+; ... and then the same for left/right extend also. See below for loop
+; function implementations. Fast are fixed-width, slow is variable-width
+
+%macro EMU_EDGE_FUNC 0
+%if ARCH_X86_64
+%define w_reg r7
+cglobal emu_edge_core, 6, 9, 1
+    mov         r8, r5          ; save block_h
+%else
+%define w_reg r6
+cglobal emu_edge_core, 2, 7, 0
+    mov         r4, r4m         ; end_y
+    mov         r5, r5m         ; block_h
+%endif
+
+    ; start with vertical extend (top/bottom) and body pixel copy
+    mov      w_reg, r7m
+    sub      w_reg, r6m         ; w = start_x - end_x
+    sub         r5, r4
+%if ARCH_X86_64
+    sub         r4, r3
+%else
+    sub         r4, dword r3m
+%endif
+    cmp      w_reg, 22
+    jg .slow_v_extend_loop
+%if ARCH_X86_32
+    mov         r2, r2m         ; linesize
+%endif
+    sal      w_reg, 7           ; w * 128
+%ifdef PIC
+    lea        rax, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)]
+    add      w_reg, rax
+%else
+    lea      w_reg, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)+w_reg]
+%endif
+    call     w_reg              ; fast top extend, body copy and bottom extend
+.v_extend_end:
+
+    ; horizontal extend (left/right)
+    mov      w_reg, r6m         ; start_x
+    sub         r0, w_reg
+%if ARCH_X86_64
+    mov         r3, r0          ; backup of buf+block_h*linesize
+    mov         r5, r8
+%else
+    mov        r0m, r0          ; backup of buf+block_h*linesize
+    mov         r5, r5m
+%endif
+    test     w_reg, w_reg
+    jz .right_extend
+    cmp      w_reg, 22
+    jg .slow_left_extend_loop
+    mov         r1, w_reg
+    dec      w_reg
+    ; FIXME we can do a if size == 1 here if that makes any speed difference, test me
+    sar      w_reg, 1
+    sal      w_reg, 6
+    ; r0=buf+block_h*linesize,r7(64)/r6(32)=start_x offset for funcs
+    ; r6(rax)/r3(ebx)=val,r2=linesize,r1=start_x,r5=block_h
+%ifdef PIC
+    lea        rax, [.emuedge_extend_left_2]
+    add      w_reg, rax
+%else
+    lea      w_reg, [.emuedge_extend_left_2+w_reg]
+%endif
+    call     w_reg
+
+    ; now r3(64)/r0(32)=buf,r2=linesize,r8/r5=block_h,r6/r3=val, r7/r6=end_x, r1=block_w
+.right_extend:
+%if ARCH_X86_32
+    mov         r0, r0m
+    mov         r5, r5m
+%endif
+    mov      w_reg, r7m         ; end_x
+    mov         r1, r8m         ; block_w
+    mov         r4, r1
+    sub         r1, w_reg
+    jz .h_extend_end            ; if (end_x == block_w) goto h_extend_end
+    cmp         r1, 22
+    jg .slow_right_extend_loop
+    dec         r1
+    ; FIXME we can do a if size == 1 here if that makes any speed difference, test me
+    sar         r1, 1
+    sal         r1, 6
+%ifdef PIC
+    lea        rax, [.emuedge_extend_right_2]
+    add         r1, rax
+%else
+    lea         r1, [.emuedge_extend_right_2+r1]
+%endif
+    call        r1
+.h_extend_end:
+    RET
+
+%if ARCH_X86_64
+%define vall  al
+%define valh  ah
+%define valw  ax
+%define valw2 r7w
+%define valw3 r3w
+%if WIN64
+%define valw4 r7w
+%else ; unix64
+%define valw4 r3w
+%endif
+%define vald eax
+%else
+%define vall  bl
+%define valh  bh
+%define valw  bx
+%define valw2 r6w
+%define valw3 valw2
+%define valw4 valw3
+%define vald ebx
+%define stack_offset 0x14
+%endif
+
+%endmacro
+
+; macro to read/write a horizontal number of pixels (%2) to/from registers
+; on x86-64, - fills xmm0-15 for consecutive sets of 16 pixels
+;            - if (%2 & 15 == 8) fills the last 8 bytes into rax
+;            - else if (%2 & 8)  fills 8 bytes into mm0
+;            - if (%2 & 7 == 4)  fills the last 4 bytes into rax
+;            - else if (%2 & 4)  fills 4 bytes into mm0-1
+;            - if (%2 & 3 == 3)  fills 2 bytes into r7/r3, and 1 into eax
+;              (note that we're using r3 for body/bottom because it's a shorter
+;               opcode, and then the loop fits in 128 bytes)
+;            - else              fills remaining bytes into rax
+; on x86-32, - fills mm0-7 for consecutive sets of 8 pixels
+;            - if (%2 & 7 == 4)  fills 4 bytes into ebx
+;            - else if (%2 & 4)  fills 4 bytes into mm0-7
+;            - if (%2 & 3 == 3)  fills 2 bytes into r6, and 1 into ebx
+;            - else              fills remaining bytes into ebx
+; writing data out is in the same way
+%macro READ_NUM_BYTES 2
+%assign %%src_off 0 ; offset in source buffer
+%assign %%smidx   0 ; mmx register idx
+%assign %%sxidx   0 ; xmm register idx
+
+%if cpuflag(sse)
+%rep %2/16
+    movups xmm %+ %%sxidx, [r1+%%src_off]
+%assign %%src_off %%src_off+16
+%assign %%sxidx   %%sxidx+1
+%endrep ; %2/16
+%endif
+
+%if ARCH_X86_64
+%if (%2-%%src_off) == 8
+    mov           rax, [r1+%%src_off]
+%assign %%src_off %%src_off+8
+%endif ; (%2-%%src_off) == 8
+%endif ; x86-64
+
+%rep (%2-%%src_off)/8
+    movq    mm %+ %%smidx, [r1+%%src_off]
+%assign %%src_off %%src_off+8
+%assign %%smidx   %%smidx+1
+%endrep ; (%2-%%dst_off)/8
+
+%if (%2-%%src_off) == 4
+    mov          vald, [r1+%%src_off]
+%elif (%2-%%src_off) & 4
+    movd    mm %+ %%smidx, [r1+%%src_off]
+%assign %%src_off %%src_off+4
+%endif ; (%2-%%src_off) ==/& 4
+
+%if (%2-%%src_off) == 1
+    mov          vall, [r1+%%src_off]
+%elif (%2-%%src_off) == 2
+    mov          valw, [r1+%%src_off]
+%elif (%2-%%src_off) == 3
+%ifidn %1, top
+    mov         valw2, [r1+%%src_off]
+%elifidn %1, body
+    mov         valw3, [r1+%%src_off]
+%elifidn %1, bottom
+    mov         valw4, [r1+%%src_off]
+%endif ; %1 ==/!= top
+    mov          vall, [r1+%%src_off+2]
+%endif ; (%2-%%src_off) == 1/2/3
+%endmacro ; READ_NUM_BYTES
+
+%macro WRITE_NUM_BYTES 2
+%assign %%dst_off 0 ; offset in destination buffer
+%assign %%dmidx   0 ; mmx register idx
+%assign %%dxidx   0 ; xmm register idx
+
+%if cpuflag(sse)
+%rep %2/16
+    movups [r0+%%dst_off], xmm %+ %%dxidx
+%assign %%dst_off %%dst_off+16
+%assign %%dxidx   %%dxidx+1
+%endrep ; %2/16
+%endif
+
+%if ARCH_X86_64
+%if (%2-%%dst_off) == 8
+    mov    [r0+%%dst_off], rax
+%assign %%dst_off %%dst_off+8
+%endif ; (%2-%%dst_off) == 8
+%endif ; x86-64
+
+%rep (%2-%%dst_off)/8
+    movq   [r0+%%dst_off], mm %+ %%dmidx
+%assign %%dst_off %%dst_off+8
+%assign %%dmidx   %%dmidx+1
+%endrep ; (%2-%%dst_off)/8
+
+%if (%2-%%dst_off) == 4
+    mov    [r0+%%dst_off], vald
+%elif (%2-%%dst_off) & 4
+    movd   [r0+%%dst_off], mm %+ %%dmidx
+%assign %%dst_off %%dst_off+4
+%endif ; (%2-%%dst_off) ==/& 4
+
+%if (%2-%%dst_off) == 1
+    mov    [r0+%%dst_off], vall
+%elif (%2-%%dst_off) == 2
+    mov    [r0+%%dst_off], valw
+%elif (%2-%%dst_off) == 3
+%ifidn %1, top
+    mov    [r0+%%dst_off], valw2
+%elifidn %1, body
+    mov    [r0+%%dst_off], valw3
+%elifidn %1, bottom
+    mov    [r0+%%dst_off], valw4
+%endif ; %1 ==/!= top
+    mov  [r0+%%dst_off+2], vall
+%endif ; (%2-%%dst_off) == 1/2/3
+%endmacro ; WRITE_NUM_BYTES
+
+; vertical top/bottom extend and body copy fast loops
+; these are function pointers to set-width line copy functions, i.e.
+; they read a fixed number of pixels into set registers, and write
+; those out into the destination buffer
+; r0=buf,r1=src,r2=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h
+; r6(eax/64)/r3(ebx/32)=val_reg
+%macro VERTICAL_EXTEND 0
+%assign %%n 1
+%rep 22
+ALIGN 128
+.emuedge_v_extend_ %+ %%n:
+    ; extend pixels above body
+%if ARCH_X86_64
+    test           r3 , r3                   ; if (!start_y)
+    jz .emuedge_copy_body_ %+ %%n %+ _loop   ;   goto body
+%else ; ARCH_X86_32
+    cmp      dword r3m, 0
+    je .emuedge_copy_body_ %+ %%n %+ _loop
+%endif ; ARCH_X86_64/32
+    READ_NUM_BYTES  top,    %%n              ; read bytes
+.emuedge_extend_top_ %+ %%n %+ _loop:        ; do {
+    WRITE_NUM_BYTES top,    %%n              ;   write bytes
+    add            r0 , r2                   ;   dst += linesize
+%if ARCH_X86_64
+    dec            r3d
+%else ; ARCH_X86_32
+    dec      dword r3m
+%endif ; ARCH_X86_64/32
+    jnz .emuedge_extend_top_ %+ %%n %+ _loop ; } while (--start_y)
+
+    ; copy body pixels
+.emuedge_copy_body_ %+ %%n %+ _loop:         ; do {
+    READ_NUM_BYTES  body,   %%n              ;   read bytes
+    WRITE_NUM_BYTES body,   %%n              ;   write bytes
+    add            r0 , r2                   ;   dst += linesize
+    add            r1 , r2                   ;   src += linesize
+    dec            r4d
+    jnz .emuedge_copy_body_ %+ %%n %+ _loop  ; } while (--end_y)
+
+    ; copy bottom pixels
+    test           r5 , r5                   ; if (!block_h)
+    jz .emuedge_v_extend_end_ %+ %%n         ;   goto end
+    sub            r1 , r2                   ; src -= linesize
+    READ_NUM_BYTES  bottom, %%n              ; read bytes
+.emuedge_extend_bottom_ %+ %%n %+ _loop:     ; do {
+    WRITE_NUM_BYTES bottom, %%n              ;   write bytes
+    add            r0 , r2                   ;   dst += linesize
+    dec            r5d
+    jnz .emuedge_extend_bottom_ %+ %%n %+ _loop ; } while (--block_h)
+
+.emuedge_v_extend_end_ %+ %%n:
+%if ARCH_X86_64
+    ret
+%else ; ARCH_X86_32
+    rep ret
+%endif ; ARCH_X86_64/32
+%assign %%n %%n+1
+%endrep
+%endmacro VERTICAL_EXTEND
+
+; left/right (horizontal) fast extend functions
+; these are essentially identical to the vertical extend ones above,
+; just left/right separated because number of pixels to extend is
+; obviously not the same on both sides.
+; for reading, pixels are placed in eax (x86-64) or ebx (x86-64) in the
+; lowest two bytes of the register (so val*0x0101), and are splatted
+; into each byte of mm0 as well if n_pixels >= 8
+
+%macro READ_V_PIXEL 2
+    mov        vall, %2
+    mov        valh, vall
+%if %1 >= 8
+    movd        mm0, vald
+%if cpuflag(mmxext)
+    pshufw      mm0, mm0, 0
+%else ; mmx
+    punpcklwd   mm0, mm0
+    punpckldq   mm0, mm0
+%endif ; sse
+%endif ; %1 >= 8
+%endmacro
+
+%macro WRITE_V_PIXEL 2
+%assign %%dst_off 0
+%rep %1/8
+    movq [%2+%%dst_off], mm0
+%assign %%dst_off %%dst_off+8
+%endrep
+%if %1 & 4
+%if %1 >= 8
+    movd [%2+%%dst_off], mm0
+%else ; %1 < 8
+    mov  [%2+%%dst_off]  , valw
+    mov  [%2+%%dst_off+2], valw
+%endif ; %1 >=/< 8
+%assign %%dst_off %%dst_off+4
+%endif ; %1 & 4
+%if %1&2
+    mov  [%2+%%dst_off], valw
+%endif ; %1 & 2
+%endmacro
+
+; r0=buf+block_h*linesize, r1=start_x, r2=linesize, r5=block_h, r6/r3=val
+%macro LEFT_EXTEND 0
+%assign %%n 2
+%rep 11
+ALIGN 64
+.emuedge_extend_left_ %+ %%n:          ; do {
+    sub         r0, r2                 ;   dst -= linesize
+    READ_V_PIXEL  %%n, [r0+r1]         ;   read pixels
+    WRITE_V_PIXEL %%n, r0              ;   write pixels
+    dec         r5
+    jnz .emuedge_extend_left_ %+ %%n   ; } while (--block_h)
+%if ARCH_X86_64
+    ret
+%else ; ARCH_X86_32
+    rep ret
+%endif ; ARCH_X86_64/32
+%assign %%n %%n+2
+%endrep
+%endmacro ; LEFT_EXTEND
+
+; r3/r0=buf+block_h*linesize, r2=linesize, r8/r5=block_h, r0/r6=end_x, r6/r3=val
+%macro RIGHT_EXTEND 0
+%assign %%n 2
+%rep 11
+ALIGN 64
+.emuedge_extend_right_ %+ %%n:          ; do {
+%if ARCH_X86_64
+    sub        r3, r2                   ;   dst -= linesize
+    READ_V_PIXEL  %%n, [r3+w_reg-1]     ;   read pixels
+    WRITE_V_PIXEL %%n, r3+r4-%%n        ;   write pixels
+    dec       r8
+%else ; ARCH_X86_32
+    sub        r0, r2                   ;   dst -= linesize
+    READ_V_PIXEL  %%n, [r0+w_reg-1]     ;   read pixels
+    WRITE_V_PIXEL %%n, r0+r4-%%n        ;   write pixels
+    dec     r5
+%endif ; ARCH_X86_64/32
+    jnz .emuedge_extend_right_ %+ %%n   ; } while (--block_h)
+%if ARCH_X86_64
+    ret
+%else ; ARCH_X86_32
+    rep ret
+%endif ; ARCH_X86_64/32
+%assign %%n %%n+2
+%endrep
+
+%if ARCH_X86_32
+%define stack_offset 0x10
+%endif
+%endmacro ; RIGHT_EXTEND
+
+; below follow the "slow" copy/extend functions, these act on a non-fixed
+; width specified in a register, and run a loop to copy the full amount
+; of bytes. They are optimized for copying of large amounts of pixels per
+; line, so they unconditionally splat data into mm registers to copy 8
+; bytes per loop iteration. It could be considered to use xmm for x86-64
+; also, but I haven't optimized this as much (i.e. FIXME)
+%macro V_COPY_NPX 4-5
+%if %0 == 4
+    test     w_reg, %4
+    jz .%1_skip_%4_px
+%else ; %0 == 5
+.%1_%4_px_loop:
+%endif
+    %3          %2, [r1+cnt_reg]
+    %3 [r0+cnt_reg], %2
+    add    cnt_reg, %4
+%if %0 == 5
+    sub      w_reg, %4
+    test     w_reg, %5
+    jnz .%1_%4_px_loop
+%endif
+.%1_skip_%4_px:
+%endmacro
+
+%macro V_COPY_ROW 2
+%ifidn %1, bottom
+    sub         r1, linesize
+%endif
+.%1_copy_loop:
+    xor    cnt_reg, cnt_reg
+%if notcpuflag(sse)
+%define linesize r2m
+    V_COPY_NPX %1,  mm0, movq,    8, 0xFFFFFFF8
+%else ; sse
+    V_COPY_NPX %1, xmm0, movups, 16, 0xFFFFFFF0
+%if ARCH_X86_64
+%define linesize r2
+    V_COPY_NPX %1, rax , mov,     8
+%else ; ARCH_X86_32
+%define linesize r2m
+    V_COPY_NPX %1,  mm0, movq,    8
+%endif ; ARCH_X86_64/32
+%endif ; sse
+    V_COPY_NPX %1, vald, mov,     4
+    V_COPY_NPX %1, valw, mov,     2
+    V_COPY_NPX %1, vall, mov,     1
+    mov      w_reg, cnt_reg
+%ifidn %1, body
+    add         r1, linesize
+%endif
+    add         r0, linesize
+    dec         %2
+    jnz .%1_copy_loop
+%endmacro
+
+%macro SLOW_V_EXTEND 0
+.slow_v_extend_loop:
+; r0=buf,r1=src,r2(64)/r2m(32)=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h
+; r8(64)/r3(later-64)/r2(32)=cnt_reg,r6(64)/r3(32)=val_reg,r7(64)/r6(32)=w=end_x-start_x
+%if ARCH_X86_64
+    push        r8              ; save old value of block_h
+    test        r3, r3
+%define cnt_reg r8
+    jz .do_body_copy            ; if (!start_y) goto do_body_copy
+    V_COPY_ROW top, r3
+%else
+    cmp  dword r3m, 0
+%define cnt_reg r2
+    je .do_body_copy            ; if (!start_y) goto do_body_copy
+    V_COPY_ROW top, dword r3m
+%endif
+
+.do_body_copy:
+    V_COPY_ROW body, r4
+
+%if ARCH_X86_64
+    pop         r8              ; restore old value of block_h
+%define cnt_reg r3
+%endif
+    test        r5, r5
+%if ARCH_X86_64
+    jz .v_extend_end
+%else
+    jz .skip_bottom_extend
+%endif
+    V_COPY_ROW bottom, r5
+%if ARCH_X86_32
+.skip_bottom_extend:
+    mov         r2, r2m
+%endif
+    jmp .v_extend_end
+%endmacro
+
+%macro SLOW_LEFT_EXTEND 0
+.slow_left_extend_loop:
+; r0=buf+block_h*linesize,r2=linesize,r6(64)/r3(32)=val,r5=block_h,r4=cntr,r7/r6=start_x
+    mov         r4, 8
+    sub         r0, linesize
+    READ_V_PIXEL 8, [r0+w_reg]
+.left_extend_8px_loop:
+    movq [r0+r4-8], mm0
+    add         r4, 8
+    cmp         r4, w_reg
+    jle .left_extend_8px_loop
+    sub         r4, 8
+    cmp         r4, w_reg
+    jge .left_extend_loop_end
+.left_extend_2px_loop:
+    mov    [r0+r4], valw
+    add         r4, 2
+    cmp         r4, w_reg
+    jl .left_extend_2px_loop
+.left_extend_loop_end:
+    dec         r5
+    jnz .slow_left_extend_loop
+%if ARCH_X86_32
+    mov         r2, r2m
+%endif
+    jmp .right_extend
+%endmacro
+
+%macro SLOW_RIGHT_EXTEND 0
+.slow_right_extend_loop:
+; r3(64)/r0(32)=buf+block_h*linesize,r2=linesize,r4=block_w,r8(64)/r5(32)=block_h,
+; r7(64)/r6(32)=end_x,r6/r3=val,r1=cntr
+%if ARCH_X86_64
+%define buf_reg r3
+%define bh_reg r8
+%else
+%define buf_reg r0
+%define bh_reg r5
+%endif
+    lea         r1, [r4-8]
+    sub    buf_reg, linesize
+    READ_V_PIXEL 8, [buf_reg+w_reg-1]
+.right_extend_8px_loop:
+    movq [buf_reg+r1], mm0
+    sub         r1, 8
+    cmp         r1, w_reg
+    jge .right_extend_8px_loop
+    add         r1, 8
+    cmp         r1, w_reg
+    je .right_extend_loop_end
+.right_extend_2px_loop:
+    sub         r1, 2
+    mov [buf_reg+r1], valw
+    cmp         r1, w_reg
+    jg .right_extend_2px_loop
+.right_extend_loop_end:
+    dec         bh_reg
+    jnz .slow_right_extend_loop
+    jmp .h_extend_end
+%endmacro
+
+%macro emu_edge 1
+INIT_XMM %1
+EMU_EDGE_FUNC
+VERTICAL_EXTEND
+LEFT_EXTEND
+RIGHT_EXTEND
+SLOW_V_EXTEND
+SLOW_LEFT_EXTEND
+SLOW_RIGHT_EXTEND
+%endmacro
+
+emu_edge sse
+%if ARCH_X86_32
+emu_edge mmx
+%endif
+
+%macro PREFETCH_FN 1
+cglobal prefetch, 3, 3, 0, buf, stride, h
+.loop:
+    %1      [bufq]
+    add      bufq, strideq
+    dec        hd
+    jg .loop
+    REP_RET
+%endmacro
+
+INIT_MMX mmxext
+PREFETCH_FN prefetcht0
+%if ARCH_X86_32
+INIT_MMX 3dnow
+PREFETCH_FN prefetch
+%endif
diff --git a/libavcodec/x86/videodsp_init.c b/libavcodec/x86/videodsp_init.c
new file mode 100644
index 0000000000..07a0324c69
--- /dev/null
+++ b/libavcodec/x86/videodsp_init.c
@@ -0,0 +1,119 @@
+/*
+ * Copyright (C) 2012 Ronald S. Bultje
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "config.h"
+#include "libavutil/common.h"
+#include "libavutil/cpu.h"
+#include "libavutil/mem.h"
+#include "libavutil/x86/asm.h"
+#include "libavcodec/videodsp.h"
+
+#if HAVE_YASM
+typedef void emu_edge_core_func(uint8_t *buf, const uint8_t *src,
+                                x86_reg linesize, x86_reg start_y,
+                                x86_reg end_y, x86_reg block_h,
+                                x86_reg start_x, x86_reg end_x,
+                                x86_reg block_w);
+extern emu_edge_core_func ff_emu_edge_core_mmx;
+extern emu_edge_core_func ff_emu_edge_core_sse;
+
+static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src,
+                                              ptrdiff_t linesize,
+                                              int block_w, int block_h,
+                                              int src_x, int src_y,
+                                              int w, int h,
+                                              emu_edge_core_func *core_fn)
+{
+    int start_y, start_x, end_y, end_x, src_y_add = 0;
+
+    if (src_y >= h) {
+        src_y_add = h - 1 - src_y;
+        src_y     = h - 1;
+    } else if (src_y <= -block_h) {
+        src_y_add = 1 - block_h - src_y;
+        src_y     = 1 - block_h;
+    }
+    if (src_x >= w) {
+        src   += w - 1 - src_x;
+        src_x  = w - 1;
+    } else if (src_x <= -block_w) {
+        src   += 1 - block_w - src_x;
+        src_x  = 1 - block_w;
+    }
+
+    start_y = FFMAX(0, -src_y);
+    start_x = FFMAX(0, -src_x);
+    end_y   = FFMIN(block_h, h-src_y);
+    end_x   = FFMIN(block_w, w-src_x);
+    assert(start_x < end_x && block_w > 0);
+    assert(start_y < end_y && block_h > 0);
+
+    // fill in the to-be-copied part plus all above/below
+    src += (src_y_add + start_y) * linesize + start_x;
+    buf += start_x;
+    core_fn(buf, src, linesize, start_y, end_y,
+            block_h, start_x, end_x, block_w);
+}
+
+#if ARCH_X86_32
+static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src,
+                                             ptrdiff_t linesize,
+                                             int block_w, int block_h,
+                                             int src_x, int src_y, int w, int h)
+{
+    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
+                     w, h, &ff_emu_edge_core_mmx);
+}
+#endif
+
+static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
+                                             ptrdiff_t linesize,
+                                             int block_w, int block_h,
+                                             int src_x, int src_y, int w, int h)
+{
+    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
+                     w, h, &ff_emu_edge_core_sse);
+}
+#endif /* HAVE_YASM */
+
+void ff_prefetch_mmxext(uint8_t *buf, ptrdiff_t stride, int h);
+void ff_prefetch_3dnow(uint8_t *buf, ptrdiff_t stride, int h);
+
+void ff_videodsp_init_x86(VideoDSPContext *ctx, int bpc)
+{
+#if HAVE_YASM
+    int mm_flags = av_get_cpu_flags();
+
+#if ARCH_X86_32
+    if (bpc <= 8 && mm_flags & AV_CPU_FLAG_MMX) {
+        ctx->emulated_edge_mc = emulated_edge_mc_mmx;
+    }
+    if (mm_flags & AV_CPU_FLAG_3DNOW) {
+        ctx->prefetch = ff_prefetch_3dnow;
+    }
+#endif /* ARCH_X86_32 */
+    if (mm_flags & AV_CPU_FLAG_MMXEXT) {
+        ctx->prefetch = ff_prefetch_mmxext;
+    }
+    if (bpc <= 8 && mm_flags & AV_CPU_FLAG_SSE) {
+        ctx->emulated_edge_mc = emulated_edge_mc_sse;
+    }
+#endif /* HAVE_YASM */
+}