MMX for H.264 deblocking filter

Originally committed as revision 4158 to svn://svn.ffmpeg.org/ffmpeg/trunk

4 files changed, 333 insertions, 149 deletions

diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index 926832ff16..29e1b0d3c2 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -2640,6 +2640,92 @@ static void h261_loop_filter_c(uint8_t *src, int stride){
     }
 }
 
+static inline void h264_loop_filter_luma_c(uint8_t *pix, int xstride, int ystride, int alpha, int beta, int *tc0)
+{
+    int i, d;
+    for( i = 0; i < 4; i++ ) {
+        if( tc0[i] < 0 ) {
+            pix += 4*ystride;
+            continue;
+        }
+        for( d = 0; d < 4; d++ ) {
+            const int p0 = pix[-1*xstride];
+            const int p1 = pix[-2*xstride];
+            const int p2 = pix[-3*xstride];
+            const int q0 = pix[0];
+            const int q1 = pix[1*xstride];
+            const int q2 = pix[2*xstride];
+    
+            if( ABS( p0 - q0 ) < alpha &&
+                ABS( p1 - p0 ) < beta &&
+                ABS( q1 - q0 ) < beta ) {
+    
+                int tc = tc0[i];
+                int i_delta;
+    
+                if( ABS( p2 - p0 ) < beta ) {
+                    pix[-2*xstride] = p1 + clip( ( p2 + ( ( p0 + q0 + 1 ) >> 1 ) - ( p1 << 1 ) ) >> 1, -tc0[i], tc0[i] );
+                    tc++;
+                }
+                if( ABS( q2 - q0 ) < beta ) {
+                    pix[xstride] = q1 + clip( ( q2 + ( ( p0 + q0 + 1 ) >> 1 ) - ( q1 << 1 ) ) >> 1, -tc0[i], tc0[i] );
+                    tc++;
+                }
+    
+                i_delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
+                pix[-xstride] = clip_uint8( p0 + i_delta );    /* p0' */
+                pix[0]        = clip_uint8( q0 - i_delta );    /* q0' */
+            }
+            pix += ystride;
+        }
+    }
+}
+static void h264_v_loop_filter_luma_c(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    h264_loop_filter_luma_c(pix, stride, 1, alpha, beta, tc0);
+}
+static void h264_h_loop_filter_luma_c(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    h264_loop_filter_luma_c(pix, 1, stride, alpha, beta, tc0);
+}
+
+static inline void h264_loop_filter_chroma_c(uint8_t *pix, int xstride, int ystride, int alpha, int beta, int *tc0)
+{
+    int i, d;
+    for( i = 0; i < 4; i++ ) {
+        const int tc = tc0[i];
+        if( tc <= 0 ) {
+            pix += 2*ystride;
+            continue;
+        }
+        for( d = 0; d < 2; d++ ) {
+            const int p0 = pix[-1*xstride];
+            const int p1 = pix[-2*xstride];
+            const int q0 = pix[0];
+            const int q1 = pix[1*xstride];
+
+            if( ABS( p0 - q0 ) < alpha &&
+                ABS( p1 - p0 ) < beta &&
+                ABS( q1 - q0 ) < beta ) {
+
+                int delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
+
+                pix[-xstride] = clip_uint8( p0 + delta );    /* p0' */
+                pix[0]        = clip_uint8( q0 - delta );    /* q0' */
+            }
+            pix += ystride;
+        }
+    }
+}
+static void h264_v_loop_filter_chroma_c(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    h264_loop_filter_chroma_c(pix, stride, 1, alpha, beta, tc0);
+}
+static void h264_h_loop_filter_chroma_c(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    h264_loop_filter_chroma_c(pix, 1, stride, alpha, beta, tc0);
+}
+
 static inline int pix_abs16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 {
     int s, i;
@@ -3739,6 +3825,11 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->diff_bytes= diff_bytes_c;
     c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_c;
     c->bswap_buf= bswap_buf;
+
+    c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_c;
+    c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_c;
+    c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_c;
+    c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_c;
     
     c->h263_h_loop_filter= h263_h_loop_filter_c;
     c->h263_v_loop_filter= h263_v_loop_filter_c;
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 7598b3a251..57547431ed 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -274,6 +274,11 @@ typedef struct DSPContext {
      */
     void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
     void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
+
+    void (*h264_v_loop_filter_luma)(uint8_t *pix, int stride, int alpha, int beta, int *tc0);
+    void (*h264_h_loop_filter_luma)(uint8_t *pix, int stride, int alpha, int beta, int *tc0);
+    void (*h264_v_loop_filter_chroma)(uint8_t *pix, int stride, int alpha, int beta, int *tc0);
+    void (*h264_h_loop_filter_chroma)(uint8_t *pix, int stride, int alpha, int beta, int *tc0);
     
     void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
     void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
diff --git a/libavcodec/h264.c b/libavcodec/h264.c
index a06ba645b3..42f9b06cc2 100644
--- a/libavcodec/h264.c
+++ b/libavcodec/h264.c
@@ -5624,48 +5624,15 @@ static void filter_mb_edgev( H264Context *h, uint8_t *pix, int stride, int bS[4]
     const int alpha = alpha_table[index_a];
     const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];
 
-    for( i = 0; i < 4; i++ ) {
-        if( bS[i] == 0 ) {
-            pix += 4 * stride;
-            continue;
-        }
-
-        if( bS[i] < 4 ) {
-            const int tc0 = tc0_table[index_a][bS[i] - 1];
-            /* 4px edge length */
-            for( d = 0; d < 4; d++ ) {
-                const int p0 = pix[-1];
-                const int p1 = pix[-2];
-                const int p2 = pix[-3];
-                const int q0 = pix[0];
-                const int q1 = pix[1];
-                const int q2 = pix[2];
-
-                if( ABS( p0 - q0 ) < alpha &&
-                    ABS( p1 - p0 ) < beta &&
-                    ABS( q1 - q0 ) < beta ) {
-                    int tc = tc0;
-                    int i_delta;
-
-                    if( ABS( p2 - p0 ) < beta ) {
-                        pix[-2] = p1 + clip( ( p2 + ( ( p0 + q0 + 1 ) >> 1 ) - ( p1 << 1 ) ) >> 1, -tc0, tc0 );
-                        tc++;
-                    }
-                    if( ABS( q2 - q0 ) < beta ) {
-                        pix[1] = q1 + clip( ( q2 + ( ( p0 + q0 + 1 ) >> 1 ) - ( q1 << 1 ) ) >> 1, -tc0, tc0 );
-                        tc++;
-                    }
-
-                    i_delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
-                    pix[-1] = clip_uint8( p0 + i_delta );    /* p0' */
-                    pix[0]  = clip_uint8( q0 - i_delta );    /* q0' */
-                    tprintf("filter_mb_edgev i:%d d:%d, qp:%d, indexA:%d, alpha:%d, beta:%d, tc:%d\n# bS:%d -> [%02x, %02x, %02x, %02x, %02x, %02x] =>[%02x, %02x, %02x, %02x]\n", i, d, qp, index_a, alpha, beta, tc, bS[i], pix[-3], p1, p0, q0, q1, pix[2], pix[-2], pix[-1], pix[0], pix[1]);
-                }
-                pix += stride;
-            }
-        }else{
-            /* 4px edge length */
-            for( d = 0; d < 4; d++ ) {
+    if( bS[0] < 4 ) {
+        int tc[4];
+        for(i=0; i<4; i++)
+            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] : -1;
+        h->s.dsp.h264_h_loop_filter_luma(pix, stride, alpha, beta, tc);
+    } else {
+        /* 16px edge length, because bS=4 is triggered by being at
+         * the edge of an intra MB, so all 4 bS are the same */
+            for( d = 0; d < 16; d++ ) {
                 const int p0 = pix[-1];
                 const int p1 = pix[-2];
                 const int p2 = pix[-3];
@@ -5710,7 +5677,6 @@ static void filter_mb_edgev( H264Context *h, uint8_t *pix, int stride, int bS[4]
                 }
                 pix += stride;
             }
-        }
     }
 }
 static void filter_mb_edgecv( H264Context *h, uint8_t *pix, int stride, int bS[4], int qp ) {
@@ -5719,35 +5685,14 @@ static void filter_mb_edgecv( H264Context *h, uint8_t *pix, int stride, int bS[4
     const int alpha = alpha_table[index_a];
     const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];
 
-    for( i = 0; i < 4; i++ ) {
-        if( bS[i] == 0 ) {
-            pix += 2 * stride;
-            continue;
-        }
-
-        if( bS[i] < 4 ) {
-            const int tc = tc0_table[index_a][bS[i] - 1] + 1;
-            /* 2px edge length (because we use same bS than the one for luma) */
-            for( d = 0; d < 2; d++ ){
-                const int p0 = pix[-1];
-                const int p1 = pix[-2];
-                const int q0 = pix[0];
-                const int q1 = pix[1];
-
-                if( ABS( p0 - q0 ) < alpha &&
-                    ABS( p1 - p0 ) < beta &&
-                    ABS( q1 - q0 ) < beta ) {
-                    const int i_delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
-
-                    pix[-1] = clip_uint8( p0 + i_delta );    /* p0' */
-                    pix[0]  = clip_uint8( q0 - i_delta );    /* q0' */
-                    tprintf("filter_mb_edgecv i:%d d:%d, qp:%d, indexA:%d, alpha:%d, beta:%d, tc:%d\n# bS:%d -> [%02x, %02x, %02x, %02x, %02x, %02x] =>[%02x, %02x, %02x, %02x]\n", i, d, qp, index_a, alpha, beta, tc, bS[i], pix[-3], p1, p0, q0, q1, pix[2], p1, pix[-1], pix[0], q1);
-                }
-                pix += stride;
-            }
-        }else{
-            /* 2px edge length (because we use same bS than the one for luma) */
-            for( d = 0; d < 2; d++ ){
+    if( bS[0] < 4 ) {
+        int tc[4];
+        for(i=0; i<4; i++)
+            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] + 1 : 0;
+        h->s.dsp.h264_h_loop_filter_chroma(pix, stride, alpha, beta, tc);
+    } else {
+        /* 8px edge length, see filter_mb_edgev */
+            for( d = 0; d < 8; d++ ){
                 const int p0 = pix[-1];
                 const int p1 = pix[-2];
                 const int q0 = pix[0];
@@ -5763,7 +5708,6 @@ static void filter_mb_edgecv( H264Context *h, uint8_t *pix, int stride, int bS[4
                 }
                 pix += stride;
             }
-        }
     }
 }
 
@@ -5928,49 +5872,14 @@ static void filter_mb_edgeh( H264Context *h, uint8_t *pix, int stride, int bS[4]
     const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];
     const int pix_next  = stride;
 
-    for( i = 0; i < 4; i++ ) {
-        if( bS[i] == 0 ) {
-            pix += 4;
-            continue;
-        }
-
-        if( bS[i] < 4 ) {
-            const int tc0 = tc0_table[index_a][bS[i] - 1];
-            /* 4px edge length */
-            for( d = 0; d < 4; d++ ) {
-                const int p0 = pix[-1*pix_next];
-                const int p1 = pix[-2*pix_next];
-                const int p2 = pix[-3*pix_next];
-                const int q0 = pix[0];
-                const int q1 = pix[1*pix_next];
-                const int q2 = pix[2*pix_next];
-
-                if( ABS( p0 - q0 ) < alpha &&
-                    ABS( p1 - p0 ) < beta &&
-                    ABS( q1 - q0 ) < beta ) {
-
-                    int tc = tc0;
-                    int i_delta;
-
-                    if( ABS( p2 - p0 ) < beta ) {
-                        pix[-2*pix_next] = p1 + clip( ( p2 + ( ( p0 + q0 + 1 ) >> 1 ) - ( p1 << 1 ) ) >> 1, -tc0, tc0 );
-                        tc++;
-                    }
-                    if( ABS( q2 - q0 ) < beta ) {
-                        pix[pix_next] = q1 + clip( ( q2 + ( ( p0 + q0 + 1 ) >> 1 ) - ( q1 << 1 ) ) >> 1, -tc0, tc0 );
-                        tc++;
-                    }
-
-                    i_delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
-                    pix[-pix_next] = clip_uint8( p0 + i_delta );    /* p0' */
-                    pix[0]         = clip_uint8( q0 - i_delta );    /* q0' */
-                    tprintf("filter_mb_edgeh i:%d d:%d, qp:%d, indexA:%d, alpha:%d, beta:%d, tc:%d\n# bS:%d -> [%02x, %02x, %02x, %02x, %02x, %02x] =>[%02x, %02x, %02x, %02x]\n", i, d, qp, index_a, alpha, beta, tc, bS[i], p2, p1, p0, q0, q1, q2, pix[-2*pix_next], pix[-pix_next], pix[0], pix[pix_next]);
-                }
-                pix++;
-            }
-        }else{
-            /* 4px edge length */
-            for( d = 0; d < 4; d++ ) {
+    if( bS[0] < 4 ) {
+        int tc[4];
+        for(i=0; i<4; i++)
+            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] : -1;
+        h->s.dsp.h264_v_loop_filter_luma(pix, stride, alpha, beta, tc);
+    } else {
+        /* 16px edge length, see filter_mb_edgev */
+            for( d = 0; d < 16; d++ ) {
                 const int p0 = pix[-1*pix_next];
                 const int p1 = pix[-2*pix_next];
                 const int p2 = pix[-3*pix_next];
@@ -6013,7 +5922,6 @@ static void filter_mb_edgeh( H264Context *h, uint8_t *pix, int stride, int bS[4]
                 }
                 pix++;
             }
-        }
     }
 }
 
@@ -6024,37 +5932,14 @@ static void filter_mb_edgech( H264Context *h, uint8_t *pix, int stride, int bS[4
     const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];
     const int pix_next  = stride;
 
-    for( i = 0; i < 4; i++ )
-    {
-        if( bS[i] == 0 ) {
-            pix += 2;
-            continue;
-        }
-
-        if( bS[i] < 4 ) {
-            int tc = tc0_table[index_a][bS[i] - 1] + 1;
-            /* 2px edge length (see deblocking_filter_edgecv) */
-            for( d = 0; d < 2; d++ ) {
-                const int p0 = pix[-1*pix_next];
-                const int p1 = pix[-2*pix_next];
-                const int q0 = pix[0];
-                const int q1 = pix[1*pix_next];
-
-                if( ABS( p0 - q0 ) < alpha &&
-                    ABS( p1 - p0 ) < beta &&
-                    ABS( q1 - q0 ) < beta ) {
-
-                    int i_delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
-
-                    pix[-pix_next] = clip_uint8( p0 + i_delta );    /* p0' */
-                    pix[0]         = clip_uint8( q0 - i_delta );    /* q0' */
-                    tprintf("filter_mb_edgech i:%d d:%d, qp:%d, indexA:%d, alpha:%d, beta:%d, tc:%d\n# bS:%d -> [%02x, %02x, %02x, %02x, %02x, %02x] =>[%02x, %02x, %02x, %02x]\n", i, d, qp, index_a, alpha, beta, tc, bS[i], pix[-3*pix_next], p1, p0, q0, q1, pix[2*pix_next], pix[-2*pix_next], pix[-pix_next], pix[0], pix[pix_next]);
-                }
-                pix++;
-            }
-        }else{
-            /* 2px edge length (see deblocking_filter_edgecv) */
-            for( d = 0; d < 2; d++ ) {
+    if( bS[0] < 4 ) {
+        int tc[4];
+        for(i=0; i<4; i++)
+            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] + 1 : 0;
+        h->s.dsp.h264_v_loop_filter_chroma(pix, stride, alpha, beta, tc);
+    } else {
+        /* 8px edge length, see filter_mb_edgev */
+            for( d = 0; d < 8; d++ ) {
                 const int p0 = pix[-1*pix_next];
                 const int p1 = pix[-2*pix_next];
                 const int q0 = pix[0];
@@ -6070,7 +5955,6 @@ static void filter_mb_edgech( H264Context *h, uint8_t *pix, int stride, int bS[4
                 }
                 pix++;
             }
-        }
     }
 }
 
diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 5501226736..1d1f2b2137 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -39,6 +39,7 @@ static const uint64_t mm_wtwo attribute_used __attribute__ ((aligned(8))) = 0x00
 
 static const uint64_t ff_pw_20 attribute_used __attribute__ ((aligned(8))) = 0x0014001400140014ULL;
 static const uint64_t ff_pw_3  attribute_used __attribute__ ((aligned(8))) = 0x0003000300030003ULL;
+static const uint64_t ff_pw_4  attribute_used __attribute__ ((aligned(8))) = 0x0004000400040004ULL;
 static const uint64_t ff_pw_5  attribute_used __attribute__ ((aligned(8))) = 0x0005000500050005ULL;
 static const uint64_t ff_pw_16 attribute_used __attribute__ ((aligned(8))) = 0x0010001000100010ULL;
 static const uint64_t ff_pw_32 attribute_used __attribute__ ((aligned(8))) = 0x0020002000200020ULL;
@@ -691,6 +692,204 @@ static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale){
     );
 }
 
+// dst = ABS( a - b )
+#define MMABS_DIFF_MMX2(a,b,dst,z)\
+    "movq    " #b ", " #dst " \n\t"\
+    "movq    " #a ", " #z   " \n\t"\
+    "psubusw " #b ", " #z   " \n\t"\
+    "psubusw " #a ", " #dst " \n\t"\
+    "pmaxsw  " #z ", " #dst " \n\t"
+
+// a = clip( a, -tc, tc )
+#define CLIP_MMX2(a,tc,z)\
+    "pxor    " #z  ", " #z "  \n\t"\
+    "psubw   " #tc ", " #z "  \n\t"\
+    "pmaxsw  " #z  ", " #a "  \n\t"\
+    "pminsw  " #tc ", " #a "  \n\t"
+
+// in: mm0=p1, mm1=p0, mm2=q0, mm3=q1
+// out: mm7 = do we filter this pixel?
+#define H264_DEBLOCK_THRESH(alpha,beta)\
+    "pxor      %%mm7, %%mm7     \n\t"\
+    "punpcklbw %%mm7, %%mm0     \n\t"\
+    "punpcklbw %%mm7, %%mm1     \n\t"\
+    "punpcklbw %%mm7, %%mm2     \n\t"\
+    "punpcklbw %%mm7, %%mm3     \n\t"\
+    MMABS_DIFF_MMX2(%%mm1, %%mm2, %%mm5, %%mm4)\
+    "movd " #alpha ", %%mm6     \n\t"\
+    "pshufw    $0, %%mm6, %%mm6 \n\t"\
+    "pcmpgtw   %%mm5, %%mm6     \n\t" /* ABS(p0-q0) < alpha */\
+    MMABS_DIFF_MMX2(%%mm0, %%mm1, %%mm5, %%mm4)\
+    MMABS_DIFF_MMX2(%%mm3, %%mm2, %%mm7, %%mm4)\
+    "pmaxsw    %%mm7, %%mm5     \n\t"\
+    "movd  " #beta ", %%mm7     \n\t"\
+    "pshufw    $0, %%mm7, %%mm7 \n\t"\
+    "movq      %%mm7, %%mm4     \n\t"\
+    "pcmpgtw   %%mm5, %%mm7     \n\t" /* ABS(p1-p0) < beta && ABS(q1-q0) < beta */\
+    "pand      %%mm6, %%mm7     \n\t"
+
+// in: mm0=p1, mm1=p0, mm2=q0, mm3=q1, mm6=tc
+// out: mm1=p0', mm2=q0'
+#define H264_DEBLOCK_P0_Q0(pw4)\
+    "movq   " #pw4 ", %%mm4     \n\t"\
+    "movq      %%mm2, %%mm5     \n\t"\
+    "paddw     %%mm4, %%mm0     \n\t"\
+    "psubw     %%mm1, %%mm5     \n\t"\
+    "psubw     %%mm3, %%mm0     \n\t"\
+    "psllw     $2,    %%mm5     \n\t"\
+    "paddw     %%mm0, %%mm5     \n\t"\
+    "psraw     $3,    %%mm5     \n\t" /* mm5 = (((q0 - p0) << 2) + (p1 - q1) + 4) >> 3 */\
+    CLIP_MMX2(%%mm5, %%mm6, %%mm4)    /* delta = clip( mm5, -tc, tc ) */\
+    "paddw     %%mm5, %%mm1     \n\t" /* p0 += delta */\
+    "psubw     %%mm5, %%mm2     \n\t" /* q0 -= delta */
+
+// in: mm1=p0, mm2=q0, mm6=tc0
+// out: mm5=delta
+#define H264_DEBLOCK_DELTA_PQ1(p1,p2,z)\
+    "movq      %%mm1, %%mm5     \n\t"\
+    "pavgb     %%mm2, %%mm5     \n\t"\
+    "paddw   " #p2 ", %%mm5     \n\t"\
+    "psraw     $1, %%mm5        \n\t"\
+    "psubw   " #p1 ", %%mm5     \n\t" /* ( ( q2 + ((p0+q0+1)>>1) ) >> 1 ) - q1 */\
+    CLIP_MMX2(%%mm5, %%mm6, z)
+
+static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int tc0)
+{
+    uint64_t tmp0, tmp1;
+    asm volatile(
+        "movd  (%2,%4),   %%mm0        \n\t" //p1
+        "movd  (%2,%4,2), %%mm1        \n\t" //p0
+        "movd  (%3),      %%mm2        \n\t" //q0
+        "movd  (%3,%4),   %%mm3        \n\t" //q1
+        H264_DEBLOCK_THRESH(%6,%7)
+        "movq      %%mm7, %0           \n\t"
+
+// filter p1 if ABS(p2-p0) < beta
+        "movd      (%2),  %%mm3        \n\t"
+        "pxor      %%mm6, %%mm6        \n\t"
+        "punpcklbw %%mm6, %%mm3        \n\t" //p2
+        MMABS_DIFF_MMX2(%%mm1, %%mm3, %%mm5, %%mm6)
+        "pcmpgtw   %%mm5, %%mm4        \n\t"
+        "pand      %%mm7, %%mm4        \n\t" // mm4 = ( ABS( p2 - p0 ) < beta && filterp )
+        "movd      %5,    %%mm6        \n\t"
+        "pshufw    $0, %%mm6, %%mm6    \n\t" //tc
+
+        H264_DEBLOCK_DELTA_PQ1(%%mm0, %%mm3, %%mm7) // delta = clip( ( p2 + ((p0+q0+1)>>1) ) >> 1 ) - p1 )
+        "pand      %%mm4, %%mm5        \n\t"
+        "paddw     %%mm0, %%mm5        \n\t"
+        "packuswb  %%mm5, %%mm5        \n\t"
+        "movd      %%mm5, (%2,%4)      \n\t" // *p1 += delta
+        "psrlw     $15, %%mm4          \n\t"
+        "paddw     %%mm6, %%mm4        \n\t" // tc++
+        "movq      %%mm4, %1           \n\t"
+
+// filter q1 if ABS(q2-q0) < beta
+        "pxor      %%mm7, %%mm7        \n\t"
+        "movd  (%3,%4),   %%mm3        \n\t" //q1
+        "movd  (%3,%4,2), %%mm4        \n\t" //q2
+        "punpcklbw %%mm7, %%mm3        \n\t"
+        "punpcklbw %%mm7, %%mm4        \n\t"
+        MMABS_DIFF_MMX2(%%mm2, %%mm4, %%mm5, %%mm7)
+        "movd      %7,    %%mm7        \n\t"
+        "pshufw    $0, %%mm7, %%mm7    \n\t"
+        "pcmpgtw   %%mm5, %%mm7        \n\t"
+
+        H264_DEBLOCK_DELTA_PQ1(%%mm3, %%mm4, %%mm4) // delta = clip( ( q2 + ((p0+q0+1)>>1) ) >> 1 ) - q1 )
+        "movq      %0,    %%mm4        \n\t"
+        "pand      %%mm4, %%mm7        \n\t" // mm7 = ( ABS( q2 - q0 ) < beta && filterp )
+        "pand      %%mm7, %%mm5        \n\t"
+        "paddw     %%mm3, %%mm5        \n\t"
+        "packuswb  %%mm5, %%mm5        \n\t"
+        "movd      %%mm5, (%3,%4)      \n\t" // *q1 += delta
+        "movq      %1, %%mm6           \n\t"
+        "psrlw     $15, %%mm7          \n\t"
+        "paddw     %%mm7, %%mm6        \n\t" // tc++
+        "movq      %0,    %%mm4        \n\t"
+        "pand      %%mm4, %%mm6        \n\t"
+
+        H264_DEBLOCK_P0_Q0(%8)
+        "packuswb  %%mm1, %%mm1        \n\t"
+        "packuswb  %%mm2, %%mm2        \n\t"
+        "movd      %%mm1, (%2,%4,2)    \n\t"
+        "movd      %%mm2, (%3)         \n\t"
+
+        : "=m"(tmp0), "=m"(tmp1)
+        : "r"(pix-3*stride), "r"(pix), "r"((long)stride),
+          "r"(tc0), "r"(alpha), "r"(beta), "m"(ff_pw_4)
+    );
+}
+
+static void h264_v_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    int i;
+    for(i=0; i<4; i++, pix+=4) {
+        if(tc0[i] < 0)
+            continue;
+        h264_loop_filter_luma_mmx2(pix, stride, alpha, beta, tc0[i]);
+    }
+}
+
+static void h264_h_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    uint8_t trans[4*8];
+    int i;
+    for(i=0; i<4; i++, pix+=4*stride) {
+        if(tc0[i] < 0)
+            continue;
+        //FIXME: could cut some load/stores by merging transpose with filter
+        transpose4x4(trans, pix-4, 4, stride);
+        transpose4x4(trans+4*4, pix, 4, stride);
+        h264_loop_filter_luma_mmx2(trans+4*4, 4, alpha, beta, tc0[i]);
+        transpose4x4(pix-2, trans+2*4, stride, 4);
+    }
+}
+
+static inline void h264_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    asm volatile(
+        "movd    (%0),    %%mm0     \n\t"
+        "movd    (%0,%2), %%mm1     \n\t"
+        "movd    (%1),    %%mm2     \n\t"
+        "movd    (%1,%2), %%mm3     \n\t"
+        H264_DEBLOCK_THRESH(%4,%5)
+        "movd      %3,    %%mm6     \n\t"
+        "pshufw $0x50, %%mm6, %%mm6 \n\t" // mm6 = tc[1], tc[1], tc[0], tc[0]
+        "pand      %%mm7, %%mm6     \n\t"
+        H264_DEBLOCK_P0_Q0(%6)
+        "packuswb  %%mm1, %%mm1     \n\t"
+        "packuswb  %%mm2, %%mm2     \n\t"
+        "movd      %%mm1, (%0,%2)   \n\t"
+        "movd      %%mm2, (%1)      \n\t"
+        :: "r"(pix-2*stride), "r"(pix), "r"((long)stride),
+           "r"(tc0[1]<<16 | tc0[0]),
+           "r"(alpha), "r"(beta), "m"(ff_pw_4)
+    );
+}
+
+static void h264_v_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    int i;
+    for(i=0; i<2; i++) {
+        h264_loop_filter_chroma_mmx2(pix, stride, alpha, beta, tc0);
+        pix += 4;
+        tc0 += 2;
+    }
+}
+
+static void h264_h_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int *tc0)
+{
+    uint8_t trans[4*4];
+    int i;
+    for(i=0; i<2; i++) {
+        //FIXME: could cut some load/stores by merging transpose with filter
+        transpose4x4(trans, pix-2, 4, stride);
+        h264_loop_filter_chroma_mmx2(trans+2*4, 4, alpha, beta, tc0);
+        transpose4x4(pix-2, trans, stride, 4);
+        pix += 4*stride;
+        tc0 += 2;
+    }
+}
+
 #ifdef CONFIG_ENCODERS
 static int pix_norm1_mmx(uint8_t *pix, int line_size) {
     int tmp;
@@ -3184,6 +3383,11 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             dspfunc(avg_h264_qpel, 2, 4);
 #undef dspfunc
 
+            c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2;
+            c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2;
+            c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2;
+            c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_mmx2;
+
 #ifdef CONFIG_ENCODERS
             c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_mmx2;
 #endif //CONFIG_ENCODERS