ssse3 h264 motion compensation.

25% faster tham mmx on core2, 35% if you discount fullpel, 4% overall decoding.

Originally committed as revision 11871 to svn://svn.ffmpeg.org/ffmpeg/trunk

3 files changed, 512 insertions, 60 deletions

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index bfa829628f..48da6b1910 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -49,10 +49,10 @@ DECLARE_ALIGNED_16(const uint64_t, ff_pdw_80000000[2]) =
 
 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_3  ) = 0x0003000300030003ULL;
 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_4  ) = 0x0004000400040004ULL;
-DECLARE_ALIGNED_8 (const uint64_t, ff_pw_5  ) = 0x0005000500050005ULL;
+DECLARE_ALIGNED_16(const xmm_t,    ff_pw_5  ) = {0x0005000500050005ULL, 0x0005000500050005ULL};
 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_8  ) = 0x0008000800080008ULL;
 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_15 ) = 0x000F000F000F000FULL;
-DECLARE_ALIGNED_8 (const uint64_t, ff_pw_16 ) = 0x0010001000100010ULL;
+DECLARE_ALIGNED_16(const xmm_t,    ff_pw_16 ) = {0x0010001000100010ULL, 0x0010001000100010ULL};
 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_20 ) = 0x0014001400140014ULL;
 DECLARE_ALIGNED_16(const xmm_t,    ff_pw_32 ) = {0x0020002000200020ULL, 0x0020002000200020ULL};
 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_42 ) = 0x002A002A002A002AULL;
@@ -478,6 +478,54 @@ static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_siz
         );
 }
 
+static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    __asm __volatile(
+         "1:                            \n\t"
+         "movdqu (%1), %%xmm0           \n\t"
+         "movdqu (%1,%3), %%xmm1        \n\t"
+         "movdqu (%1,%3,2), %%xmm2      \n\t"
+         "movdqu (%1,%4), %%xmm3        \n\t"
+         "movdqa %%xmm0, (%2)           \n\t"
+         "movdqa %%xmm1, (%2,%3)        \n\t"
+         "movdqa %%xmm2, (%2,%3,2)      \n\t"
+         "movdqa %%xmm3, (%2,%4)        \n\t"
+         "subl $4, %0                   \n\t"
+         "lea (%1,%3,4), %1             \n\t"
+         "lea (%2,%3,4), %2             \n\t"
+         "jnz 1b                        \n\t"
+         : "+g"(h), "+r" (pixels),  "+r" (block)
+         : "r"((long)line_size), "r"(3L*line_size)
+         : "memory"
+        );
+}
+
+static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    __asm __volatile(
+         "1:                            \n\t"
+         "movdqu (%1), %%xmm0           \n\t"
+         "movdqu (%1,%3), %%xmm1        \n\t"
+         "movdqu (%1,%3,2), %%xmm2      \n\t"
+         "movdqu (%1,%4), %%xmm3        \n\t"
+         "pavgb  (%2), %%xmm0           \n\t"
+         "pavgb  (%2,%3), %%xmm1        \n\t"
+         "pavgb  (%2,%3,2), %%xmm2      \n\t"
+         "pavgb  (%2,%4), %%xmm3        \n\t"
+         "movdqa %%xmm0, (%2)           \n\t"
+         "movdqa %%xmm1, (%2,%3)        \n\t"
+         "movdqa %%xmm2, (%2,%3,2)      \n\t"
+         "movdqa %%xmm3, (%2,%4)        \n\t"
+         "subl $4, %0                   \n\t"
+         "lea (%1,%3,4), %1             \n\t"
+         "lea (%2,%3,4), %2             \n\t"
+         "jnz 1b                        \n\t"
+         : "+g"(h), "+r" (pixels),  "+r" (block)
+         : "r"((long)line_size), "r"(3L*line_size)
+         : "memory"
+        );
+}
+
 static void clear_blocks_mmx(DCTELEM *blocks)
 {
     __asm __volatile(
@@ -3475,6 +3523,23 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow;
         }
 
+/* FIXME works in most codecs, but crashes svq1 due to unaligned chroma
+        if((mm_flags & MM_SSE2) && !(mm_flags & MM_3DNOW)){
+            // these functions are slower than mmx on AMD, but faster on Intel
+            c->put_pixels_tab[0][0] = put_pixels16_sse2;
+            c->avg_pixels_tab[0][0] = avg_pixels16_sse2;
+        }
+*/
+
+#ifdef HAVE_SSSE3
+        if(mm_flags & MM_SSSE3){
+            SET_QPEL_FUNCS(put_h264_qpel, 0, 16, ssse3);
+            SET_QPEL_FUNCS(put_h264_qpel, 1, 8, ssse3);
+            SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, ssse3);
+            SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, ssse3);
+        }
+#endif
+
 #ifdef CONFIG_ENCODERS
         if(mm_flags & MM_SSE2){
             c->sum_abs_dctelem= sum_abs_dctelem_sse2;
diff --git a/libavcodec/i386/dsputil_mmx.h b/libavcodec/i386/dsputil_mmx.h
index c0846ca160..7747487b3b 100644
--- a/libavcodec/i386/dsputil_mmx.h
+++ b/libavcodec/i386/dsputil_mmx.h
@@ -33,10 +33,10 @@ extern const uint64_t ff_pdw_80000000[2];
 
 extern const uint64_t ff_pw_3;
 extern const uint64_t ff_pw_4;
-extern const uint64_t ff_pw_5;
+extern const xmm_t    ff_pw_5;
 extern const uint64_t ff_pw_8;
 extern const uint64_t ff_pw_15;
-extern const uint64_t ff_pw_16;
+extern const xmm_t    ff_pw_16;
 extern const uint64_t ff_pw_20;
 extern const xmm_t    ff_pw_32;
 extern const uint64_t ff_pw_42;
diff --git a/libavcodec/i386/h264dsp_mmx.c b/libavcodec/i386/h264dsp_mmx.c
index 657c96bb32..e71ee7f8b1 100644
--- a/libavcodec/i386/h264dsp_mmx.c
+++ b/libavcodec/i386/h264dsp_mmx.c
@@ -721,38 +721,44 @@ static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40]
 /***********************************/
 /* motion compensation */
 
-#define QPEL_H264V(A,B,C,D,E,F,OP)\
-        "movd (%0), "#F"            \n\t"\
-        "movq "#C", %%mm6           \n\t"\
-        "paddw "#D", %%mm6          \n\t"\
-        "psllw $2, %%mm6            \n\t"\
-        "psubw "#B", %%mm6          \n\t"\
-        "psubw "#E", %%mm6          \n\t"\
-        "pmullw %4, %%mm6           \n\t"\
-        "add %2, %0                 \n\t"\
-        "punpcklbw %%mm7, "#F"      \n\t"\
+#define QPEL_H264V_MM(A,B,C,D,E,F,OP,T,Z,d,q)\
+        "mov"#q" "#C", "#T"         \n\t"\
+        "mov"#d" (%0), "#F"         \n\t"\
+        "paddw "#D", "#T"           \n\t"\
+        "psllw $2, "#T"             \n\t"\
+        "psubw "#B", "#T"           \n\t"\
+        "psubw "#E", "#T"           \n\t"\
+        "punpcklbw "#Z", "#F"       \n\t"\
+        "pmullw %4, "#T"            \n\t"\
         "paddw %5, "#A"             \n\t"\
+        "add %2, %0                 \n\t"\
         "paddw "#F", "#A"           \n\t"\
-        "paddw "#A", %%mm6          \n\t"\
-        "psraw $5, %%mm6            \n\t"\
-        "packuswb %%mm6, %%mm6      \n\t"\
-        OP(%%mm6, (%1), A, d)\
+        "paddw "#A", "#T"           \n\t"\
+        "psraw $5, "#T"             \n\t"\
+        "packuswb "#T", "#T"        \n\t"\
+        OP(T, (%1), A, d)\
         "add %3, %1                 \n\t"
 
-#define QPEL_H264HV(A,B,C,D,E,F,OF)\
-        "movd (%0), "#F"            \n\t"\
-        "movq "#C", %%mm6           \n\t"\
-        "paddw "#D", %%mm6          \n\t"\
-        "psllw $2, %%mm6            \n\t"\
-        "psubw "#B", %%mm6          \n\t"\
-        "psubw "#E", %%mm6          \n\t"\
-        "pmullw %3, %%mm6           \n\t"\
-        "add %2, %0                 \n\t"\
-        "punpcklbw %%mm7, "#F"      \n\t"\
+#define QPEL_H264HV_MM(A,B,C,D,E,F,OF,T,Z,d,q)\
+        "mov"#q" "#C", "#T"         \n\t"\
+        "mov"#d" (%0), "#F"         \n\t"\
+        "paddw "#D", "#T"           \n\t"\
+        "psllw $2, "#T"             \n\t"\
         "paddw %4, "#A"             \n\t"\
+        "psubw "#B", "#T"           \n\t"\
+        "psubw "#E", "#T"           \n\t"\
+        "punpcklbw "#Z", "#F"       \n\t"\
+        "pmullw %3, "#T"            \n\t"\
         "paddw "#F", "#A"           \n\t"\
-        "paddw "#A", %%mm6          \n\t"\
-        "movq %%mm6, "#OF"(%1)      \n\t"
+        "add %2, %0                 \n\t"\
+        "paddw "#A", "#T"           \n\t"\
+        "mov"#q" "#T", "#OF"(%1)    \n\t"
+
+#define QPEL_H264V(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%mm6,%%mm7,d,q)
+#define QPEL_H264HV(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%mm6,%%mm7,d,q)
+#define QPEL_H264V_XMM(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%xmm6,%%xmm7,q,dqa)
+#define QPEL_H264HV_XMM(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%xmm6,%%xmm7,q,dqa)
+
 
 #define QPEL_H264(OPNAME, OP, MMX)\
 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
@@ -1270,21 +1276,29 @@ static av_noinline void OPNAME ## pixels4_l2_shift5_ ## MMX(uint8_t *dst, int16_
 }\
 static av_noinline void OPNAME ## pixels8_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
 {\
-    while(h--){\
+    do{\
     asm volatile(\
         "movq      (%1), %%mm0          \n\t"\
         "movq     8(%1), %%mm1          \n\t"\
+        "movq    48(%1), %%mm2          \n\t"\
+        "movq  8+48(%1), %%mm3          \n\t"\
         "psraw      $5,  %%mm0          \n\t"\
         "psraw      $5,  %%mm1          \n\t"\
+        "psraw      $5,  %%mm2          \n\t"\
+        "psraw      $5,  %%mm3          \n\t"\
         "packuswb %%mm1, %%mm0          \n\t"\
+        "packuswb %%mm3, %%mm2          \n\t"\
         PAVGB"     (%0), %%mm0          \n\t"\
+        PAVGB"  (%0,%3), %%mm2          \n\t"\
         OP(%%mm0, (%2), %%mm5, q)\
-        ::"a"(src8), "c"(src16), "d"(dst)\
+        OP(%%mm2, (%2,%4), %%mm5, q)\
+        ::"a"(src8), "c"(src16), "d"(dst),\
+          "r"((long)src8Stride), "r"((long)dstStride)\
         :"memory");\
-        src8 += src8Stride;\
-        src16 += 24;\
-        dst += dstStride;\
-    }\
+        src8 += 2L*src8Stride;\
+        src16 += 48;\
+        dst += 2L*dstStride;\
+    }while(h-=2);\
 }\
 static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
 {\
@@ -1292,8 +1306,364 @@ static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, u
     OPNAME ## pixels8_l2_shift5_ ## MMX(dst+8, src16+8, src8+8, dstStride, src8Stride, h);\
 }\
 
+#ifdef ARCH_X86_64
+#define QPEL_H264_HL2_XMM(OPNAME, OP, MMX)\
+static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
+    int h=16;\
+    asm volatile(\
+        "pxor %%xmm15, %%xmm15      \n\t"\
+        "movdqa %6, %%xmm14         \n\t"\
+        "movdqa %7, %%xmm13         \n\t"\
+        "1:                         \n\t"\
+        "lddqu    3(%0), %%xmm1     \n\t"\
+        "lddqu   -5(%0), %%xmm7     \n\t"\
+        "movdqa  %%xmm1, %%xmm0     \n\t"\
+        "punpckhbw %%xmm15, %%xmm1  \n\t"\
+        "punpcklbw %%xmm15, %%xmm0  \n\t"\
+        "punpcklbw %%xmm15, %%xmm7  \n\t"\
+        "movdqa  %%xmm1, %%xmm2     \n\t"\
+        "movdqa  %%xmm0, %%xmm6     \n\t"\
+        "movdqa  %%xmm1, %%xmm3     \n\t"\
+        "movdqa  %%xmm0, %%xmm8     \n\t"\
+        "movdqa  %%xmm1, %%xmm4     \n\t"\
+        "movdqa  %%xmm0, %%xmm9     \n\t"\
+        "movdqa  %%xmm1, %%xmm5     \n\t"\
+        "movdqa  %%xmm0, %%xmm10    \n\t"\
+        "palignr $6, %%xmm0, %%xmm5 \n\t"\
+        "palignr $6, %%xmm7, %%xmm10\n\t"\
+        "palignr $8, %%xmm0, %%xmm4 \n\t"\
+        "palignr $8, %%xmm7, %%xmm9 \n\t"\
+        "palignr $10,%%xmm0, %%xmm3 \n\t"\
+        "palignr $10,%%xmm7, %%xmm8 \n\t"\
+        "paddw   %%xmm1, %%xmm5     \n\t"\
+        "paddw   %%xmm0, %%xmm10    \n\t"\
+        "palignr $12,%%xmm0, %%xmm2 \n\t"\
+        "palignr $12,%%xmm7, %%xmm6 \n\t"\
+        "palignr $14,%%xmm0, %%xmm1 \n\t"\
+        "palignr $14,%%xmm7, %%xmm0 \n\t"\
+        "paddw   %%xmm3, %%xmm2     \n\t"\
+        "paddw   %%xmm8, %%xmm6     \n\t"\
+        "paddw   %%xmm4, %%xmm1     \n\t"\
+        "paddw   %%xmm9, %%xmm0     \n\t"\
+        "psllw   $2,     %%xmm2     \n\t"\
+        "psllw   $2,     %%xmm6     \n\t"\
+        "psubw   %%xmm1, %%xmm2     \n\t"\
+        "psubw   %%xmm0, %%xmm6     \n\t"\
+        "paddw   %%xmm13,%%xmm5     \n\t"\
+        "paddw   %%xmm13,%%xmm10    \n\t"\
+        "pmullw  %%xmm14,%%xmm2     \n\t"\
+        "pmullw  %%xmm14,%%xmm6     \n\t"\
+        "lddqu   (%2),   %%xmm3     \n\t"\
+        "paddw   %%xmm5, %%xmm2     \n\t"\
+        "paddw   %%xmm10,%%xmm6     \n\t"\
+        "psraw   $5,     %%xmm2     \n\t"\
+        "psraw   $5,     %%xmm6     \n\t"\
+        "packuswb %%xmm2,%%xmm6     \n\t"\
+        "pavgb   %%xmm3, %%xmm6     \n\t"\
+        OP(%%xmm6, (%1), %%xmm4, dqa)\
+        "add %5, %0                 \n\t"\
+        "add %5, %1                 \n\t"\
+        "add %4, %2                 \n\t"\
+        "dec %3                     \n\t"\
+        "jg 1b                      \n\t"\
+        : "+a"(src), "+c"(dst), "+d"(src2), "+g"(h)\
+        : "D"((long)src2Stride), "S"((long)dstStride),\
+          "m"(ff_pw_5), "m"(ff_pw_16)\
+        : "memory"\
+    );\
+}
+#else // ARCH_X86_64
+#define QPEL_H264_HL2_XMM(OPNAME, OP, MMX)\
+static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
+    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst  , src  , src2  , dstStride, src2Stride);\
+    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
+    src += 8*dstStride;\
+    dst += 8*dstStride;\
+    src2 += 8*src2Stride;\
+    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst  , src  , src2  , dstStride, src2Stride);\
+    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
+}
+#endif // ARCH_X86_64
+
+#define QPEL_H264_XMM(OPNAME, OP, MMX)\
+static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
+    int h=8;\
+    asm volatile(\
+        "pxor %%xmm7, %%xmm7        \n\t"\
+        "movdqa %0, %%xmm6          \n\t"\
+        :: "m"(ff_pw_5)\
+    );\
+    do{\
+    asm volatile(\
+        "lddqu   -5(%0), %%xmm1     \n\t"\
+        "movdqa  %%xmm1, %%xmm0     \n\t"\
+        "punpckhbw %%xmm7, %%xmm1   \n\t"\
+        "punpcklbw %%xmm7, %%xmm0   \n\t"\
+        "movdqa  %%xmm1, %%xmm2     \n\t"\
+        "movdqa  %%xmm1, %%xmm3     \n\t"\
+        "movdqa  %%xmm1, %%xmm4     \n\t"\
+        "movdqa  %%xmm1, %%xmm5     \n\t"\
+        "palignr $6, %%xmm0, %%xmm5 \n\t"\
+        "palignr $8, %%xmm0, %%xmm4 \n\t"\
+        "palignr $10,%%xmm0, %%xmm3 \n\t"\
+        "paddw   %%xmm1, %%xmm5     \n\t"\
+        "palignr $12,%%xmm0, %%xmm2 \n\t"\
+        "palignr $14,%%xmm0, %%xmm1 \n\t"\
+        "paddw   %%xmm3, %%xmm2     \n\t"\
+        "paddw   %%xmm4, %%xmm1     \n\t"\
+        "psllw   $2,     %%xmm2     \n\t"\
+        "movq    (%2),   %%xmm3     \n\t"\
+        "psubw   %%xmm1, %%xmm2     \n\t"\
+        "paddw   %5,     %%xmm5     \n\t"\
+        "pmullw  %%xmm6, %%xmm2     \n\t"\
+        "paddw   %%xmm5, %%xmm2     \n\t"\
+        "psraw   $5,     %%xmm2     \n\t"\
+        "packuswb %%xmm2, %%xmm2    \n\t"\
+        "pavgb   %%xmm3, %%xmm2     \n\t"\
+        OP(%%xmm2, (%1), %%xmm4, q)\
+        "add %4, %0                 \n\t"\
+        "add %4, %1                 \n\t"\
+        "add %3, %2                 \n\t"\
+        : "+a"(src), "+c"(dst), "+d"(src2)\
+        : "D"((long)src2Stride), "S"((long)dstStride),\
+          "m"(ff_pw_16)\
+        : "memory"\
+    );\
+    }while(--h);\
+}\
+QPEL_H264_HL2_XMM(OPNAME, OP, MMX)\
+\
+static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
+    src -= 2*srcStride;\
+    \
+    asm volatile(\
+        "pxor %%xmm7, %%xmm7        \n\t"\
+        "movq (%0), %%xmm0          \n\t"\
+        "add %2, %0                 \n\t"\
+        "movq (%0), %%xmm1          \n\t"\
+        "add %2, %0                 \n\t"\
+        "movq (%0), %%xmm2          \n\t"\
+        "add %2, %0                 \n\t"\
+        "movq (%0), %%xmm3          \n\t"\
+        "add %2, %0                 \n\t"\
+        "movq (%0), %%xmm4          \n\t"\
+        "add %2, %0                 \n\t"\
+        "punpcklbw %%xmm7, %%xmm0   \n\t"\
+        "punpcklbw %%xmm7, %%xmm1   \n\t"\
+        "punpcklbw %%xmm7, %%xmm2   \n\t"\
+        "punpcklbw %%xmm7, %%xmm3   \n\t"\
+        "punpcklbw %%xmm7, %%xmm4   \n\t"\
+        QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
+        QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
+        QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
+        QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
+        QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
+        QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
+        QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
+        QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
+         \
+        : "+a"(src), "+c"(dst)\
+        : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
+        : "memory"\
+    );\
+    if(h==16){\
+        asm volatile(\
+            QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
+            QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
+            QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
+            QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
+            QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
+            QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
+            QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
+            QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
+            \
+            : "+a"(src), "+c"(dst)\
+            : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
+            : "memory"\
+        );\
+    }\
+}\
+static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\
+    int h = size;\
+    int w = (size+8)>>3;\
+    src -= 2*srcStride+2;\
+    while(w--){\
+        asm volatile(\
+            "pxor %%xmm7, %%xmm7        \n\t"\
+            "movq (%0), %%xmm0          \n\t"\
+            "add %2, %0                 \n\t"\
+            "movq (%0), %%xmm1          \n\t"\
+            "add %2, %0                 \n\t"\
+            "movq (%0), %%xmm2          \n\t"\
+            "add %2, %0                 \n\t"\
+            "movq (%0), %%xmm3          \n\t"\
+            "add %2, %0                 \n\t"\
+            "movq (%0), %%xmm4          \n\t"\
+            "add %2, %0                 \n\t"\
+            "punpcklbw %%xmm7, %%xmm0   \n\t"\
+            "punpcklbw %%xmm7, %%xmm1   \n\t"\
+            "punpcklbw %%xmm7, %%xmm2   \n\t"\
+            "punpcklbw %%xmm7, %%xmm3   \n\t"\
+            "punpcklbw %%xmm7, %%xmm4   \n\t"\
+            QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 0*48)\
+            QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 1*48)\
+            QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 2*48)\
+            QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 3*48)\
+            QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 4*48)\
+            QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 5*48)\
+            QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 6*48)\
+            QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 7*48)\
+            : "+a"(src)\
+            : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
+            : "memory"\
+        );\
+        if(size==16){\
+            asm volatile(\
+                QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1,  8*48)\
+                QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2,  9*48)\
+                QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 10*48)\
+                QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 11*48)\
+                QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 12*48)\
+                QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 13*48)\
+                QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 14*48)\
+                QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 15*48)\
+                : "+a"(src)\
+                : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
+                : "memory"\
+            );\
+        }\
+        tmp += 8;\
+        src += 8 - (size+5)*srcStride;\
+    }\
+    tmp -= size+8;\
+    h = size;\
+    if(size==16){\
+        asm volatile(\
+            "1:                         \n\t"\
+            "movdqa 32(%0), %%xmm4      \n\t"\
+            "movdqa 16(%0), %%xmm5      \n\t"\
+            "movdqa   (%0), %%xmm7      \n\t"\
+            "movdqa %%xmm4, %%xmm3      \n\t"\
+            "movdqa %%xmm4, %%xmm2      \n\t"\
+            "movdqa %%xmm4, %%xmm1      \n\t"\
+            "movdqa %%xmm4, %%xmm0      \n\t"\
+            "palignr $10, %%xmm5, %%xmm0 \n\t"\
+            "palignr  $8, %%xmm5, %%xmm1 \n\t"\
+            "palignr  $6, %%xmm5, %%xmm2 \n\t"\
+            "palignr  $4, %%xmm5, %%xmm3 \n\t"\
+            "palignr  $2, %%xmm5, %%xmm4 \n\t"\
+            "paddw  %%xmm5, %%xmm0      \n\t"\
+            "paddw  %%xmm4, %%xmm1      \n\t"\
+            "paddw  %%xmm3, %%xmm2      \n\t"\
+            "movdqa %%xmm5, %%xmm6      \n\t"\
+            "movdqa %%xmm5, %%xmm4      \n\t"\
+            "movdqa %%xmm5, %%xmm3      \n\t"\
+            "palignr  $8, %%xmm7, %%xmm4 \n\t"\
+            "palignr  $2, %%xmm7, %%xmm6 \n\t"\
+            "palignr $10, %%xmm7, %%xmm3 \n\t"\
+            "paddw  %%xmm6, %%xmm4      \n\t"\
+            "movdqa %%xmm5, %%xmm6      \n\t"\
+            "palignr  $6, %%xmm7, %%xmm5 \n\t"\
+            "palignr  $4, %%xmm7, %%xmm6 \n\t"\
+            "paddw  %%xmm7, %%xmm3      \n\t"\
+            "paddw  %%xmm6, %%xmm5      \n\t"\
+            \
+            "psubw  %%xmm1, %%xmm0      \n\t"\
+            "psubw  %%xmm4, %%xmm3      \n\t"\
+            "psraw      $2, %%xmm0      \n\t"\
+            "psraw      $2, %%xmm3      \n\t"\
+            "psubw  %%xmm1, %%xmm0      \n\t"\
+            "psubw  %%xmm4, %%xmm3      \n\t"\
+            "paddw  %%xmm2, %%xmm0      \n\t"\
+            "paddw  %%xmm5, %%xmm3      \n\t"\
+            "psraw      $2, %%xmm0      \n\t"\
+            "psraw      $2, %%xmm3      \n\t"\
+            "paddw  %%xmm2, %%xmm0      \n\t"\
+            "paddw  %%xmm5, %%xmm3      \n\t"\
+            "psraw      $6, %%xmm0      \n\t"\
+            "psraw      $6, %%xmm3      \n\t"\
+            "packuswb %%xmm0, %%xmm3    \n\t"\
+            OP(%%xmm3, (%1), %%xmm7, dqa)\
+            "add $48, %0                \n\t"\
+            "add %3, %1                 \n\t"\
+            "decl %2                    \n\t"\
+            " jnz 1b                    \n\t"\
+            : "+a"(tmp), "+c"(dst), "+g"(h)\
+            : "S"((long)dstStride)\
+            : "memory"\
+        );\
+    }else{\
+        asm volatile(\
+            "1:                         \n\t"\
+            "movdqa 16(%0), %%xmm1      \n\t"\
+            "movdqa   (%0), %%xmm0      \n\t"\
+            "movdqa %%xmm1, %%xmm2      \n\t"\
+            "movdqa %%xmm1, %%xmm3      \n\t"\
+            "movdqa %%xmm1, %%xmm4      \n\t"\
+            "movdqa %%xmm1, %%xmm5      \n\t"\
+            "palignr $10, %%xmm0, %%xmm5 \n\t"\
+            "palignr  $8, %%xmm0, %%xmm4 \n\t"\
+            "palignr  $6, %%xmm0, %%xmm3 \n\t"\
+            "palignr  $4, %%xmm0, %%xmm2 \n\t"\
+            "palignr  $2, %%xmm0, %%xmm1 \n\t"\
+            "paddw  %%xmm5, %%xmm0      \n\t"\
+            "paddw  %%xmm4, %%xmm1      \n\t"\
+            "paddw  %%xmm3, %%xmm2      \n\t"\
+            "psubw  %%xmm1, %%xmm0      \n\t"\
+            "psraw      $2, %%xmm0      \n\t"\
+            "psubw  %%xmm1, %%xmm0      \n\t"\
+            "paddw  %%xmm2, %%xmm0      \n\t"\
+            "psraw      $2, %%xmm0      \n\t"\
+            "paddw  %%xmm2, %%xmm0      \n\t"\
+            "psraw      $6, %%xmm0      \n\t"\
+            "packuswb %%xmm0, %%xmm0    \n\t"\
+            OP(%%xmm0, (%1), %%xmm7, q)\
+            "add $48, %0                \n\t"\
+            "add %3, %1                 \n\t"\
+            "decl %2                    \n\t"\
+            " jnz 1b                    \n\t"\
+            : "+a"(tmp), "+c"(dst), "+g"(h)\
+            : "S"((long)dstStride)\
+            : "memory"\
+        );\
+    }\
+}\
+\
+static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
+    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst  , src  , dstStride, srcStride, 8);\
+}\
+static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
+    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst  , src  , dstStride, srcStride, 16);\
+    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
+}\
+static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
+    OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst  , tmp  , src  , dstStride, tmpStride, srcStride, 8);\
+}\
+static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
+    OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst  , tmp  , src  , dstStride, tmpStride, srcStride, 16);\
+}\
+
+#define put_pixels8_ssse3 put_pixels8_mmx2
+#define put_pixels16_ssse3 put_pixels16_sse2
+#define avg_pixels8_ssse3 avg_pixels8_mmx2
+#define avg_pixels16_ssse3 avg_pixels16_sse2
+
+#define put_pixels8_l2_ssse3 put_pixels8_l2_mmx2
+#define put_pixels16_l2_ssse3 put_pixels16_l2_mmx2
+#define avg_pixels8_l2_ssse3 avg_pixels8_l2_mmx2
+#define avg_pixels16_l2_ssse3 avg_pixels16_l2_mmx2
+
+#define put_pixels8_l2_shift5_ssse3 put_pixels8_l2_shift5_mmx2
+#define put_pixels16_l2_shift5_ssse3 put_pixels16_l2_shift5_mmx2
+#define avg_pixels8_l2_shift5_ssse3 avg_pixels8_l2_shift5_mmx2
+#define avg_pixels16_l2_shift5_ssse3 avg_pixels16_l2_shift5_mmx2
+
+#define put_h264_qpel8_h_lowpass_ssse3 put_h264_qpel8_h_lowpass_mmx2
+#define put_h264_qpel16_h_lowpass_ssse3 put_h264_qpel16_h_lowpass_mmx2
+#define avg_h264_qpel8_h_lowpass_ssse3 avg_h264_qpel8_h_lowpass_mmx2
+#define avg_h264_qpel16_h_lowpass_ssse3 avg_h264_qpel16_h_lowpass_mmx2
 
-#define H264_MC(OPNAME, SIZE, MMX) \
+#define H264_MC(OPNAME, SIZE, MMX, ALIGN) \
 static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
     OPNAME ## pixels ## SIZE ## _ ## MMX(dst, src, stride, SIZE);\
 }\
@@ -1311,7 +1681,7 @@ static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## MMX(uint8_t *dst, uint8_t *
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
     uint8_t * const half= (uint8_t*)temp;\
     put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\
     OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src, half, stride, stride, SIZE);\
@@ -1322,48 +1692,48 @@ static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
     uint8_t * const half= (uint8_t*)temp;\
     put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\
     OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src+stride, half, stride, stride, SIZE);\
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
     uint8_t * const halfV= (uint8_t*)temp;\
     put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\
     OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
     uint8_t * const halfV= (uint8_t*)temp;\
     put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\
     OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
     uint8_t * const halfV= (uint8_t*)temp;\
     put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\
     OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
     uint8_t * const halfV= (uint8_t*)temp;\
     put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\
     OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*(SIZE<8?12:24)/4];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4]);\
     int16_t * const tmp= (int16_t*)temp;\
     OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(dst, tmp, src, stride, SIZE, stride);\
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
     uint8_t * const halfHV= (uint8_t*)temp;\
     int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\
     assert(((int)temp & 7) == 0);\
@@ -1372,7 +1742,7 @@ static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
     uint8_t * const halfHV= (uint8_t*)temp;\
     int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\
     assert(((int)temp & 7) == 0);\
@@ -1381,7 +1751,7 @@ static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
     int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\
     uint8_t * const halfHV= ((uint8_t*)temp);\
     assert(((int)temp & 7) == 0);\
@@ -1390,7 +1760,7 @@ static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *
 }\
 \
 static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
-    uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\
+    DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
     int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\
     uint8_t * const halfHV= ((uint8_t*)temp);\
     assert(((int)temp & 7) == 0);\
@@ -1415,20 +1785,37 @@ QPEL_H264(avg_, AVG_3DNOW_OP, 3dnow)
 #define PAVGB "pavgb"
 QPEL_H264(put_,       PUT_OP, mmx2)
 QPEL_H264(avg_,  AVG_MMX2_OP, mmx2)
+#ifdef HAVE_SSSE3
+QPEL_H264_XMM(put_,       PUT_OP, ssse3)
+QPEL_H264_XMM(avg_,  AVG_MMX2_OP, ssse3)
+#endif
 #undef PAVGB
 
-H264_MC(put_, 4, 3dnow)
-H264_MC(put_, 8, 3dnow)
-H264_MC(put_, 16,3dnow)
-H264_MC(avg_, 4, 3dnow)
-H264_MC(avg_, 8, 3dnow)
-H264_MC(avg_, 16,3dnow)
-H264_MC(put_, 4, mmx2)
-H264_MC(put_, 8, mmx2)
-H264_MC(put_, 16,mmx2)
-H264_MC(avg_, 4, mmx2)
-H264_MC(avg_, 8, mmx2)
-H264_MC(avg_, 16,mmx2)
+H264_MC(put_, 4, 3dnow, 8)
+H264_MC(put_, 8, 3dnow, 8)
+H264_MC(put_, 16,3dnow, 8)
+H264_MC(avg_, 4, 3dnow, 8)
+H264_MC(avg_, 8, 3dnow, 8)
+H264_MC(avg_, 16,3dnow, 8)
+H264_MC(put_, 4, mmx2, 8)
+H264_MC(put_, 8, mmx2, 8)
+H264_MC(put_, 16,mmx2, 8)
+H264_MC(avg_, 4, mmx2, 8)
+H264_MC(avg_, 8, mmx2, 8)
+H264_MC(avg_, 16,mmx2, 8)
+#ifdef HAVE_SSSE3
+// FIXME some of these only require sse2
+H264_MC(put_, 8, ssse3, 16)
+H264_MC(put_, 16,ssse3, 16)
+H264_MC(avg_, 8, ssse3, 16)
+H264_MC(avg_, 16,ssse3, 16)
+#endif
+
+// no mc20 because it wasn't significantly faster than mmx2
+#define put_h264_qpel8_mc20_ssse3 put_h264_qpel8_mc20_mmx2
+#define put_h264_qpel16_mc20_ssse3 put_h264_qpel16_mc20_mmx2
+#define avg_h264_qpel8_mc20_ssse3 avg_h264_qpel8_mc20_mmx2
+#define avg_h264_qpel16_mc20_ssse3 avg_h264_qpel16_mc20_mmx2
 
 
 #define H264_CHROMA_OP(S,D)