x86/me_cmp: port mmxext and sse2 sad functions to yasm

Also add a missing c->pix_abs[0][0] initialization, and sse2 versions of
sad16_x2, sad16_y2 and sad16_xy2 (%15 to %20 faster than mmxext).
Since the _xy2 versions are not bitexact, they are accordingly marked as
approximate.

Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

2 files changed, 379 insertions, 154 deletions

diff --git a/libavcodec/x86/me_cmp.asm b/libavcodec/x86/me_cmp.asm
index b0741f398b..27176f47ec 100644
--- a/libavcodec/x86/me_cmp.asm
+++ b/libavcodec/x86/me_cmp.asm
@@ -23,6 +23,10 @@
 
 %include "libavutil/x86/x86util.asm"
 
+SECTION_RODATA
+
+cextern pb_1
+
 SECTION .text
 
 %macro DIFF_PIXELS_1 4
@@ -465,3 +469,329 @@ cglobal hf_noise%1, 3,3,0, pix1, lsize, h
 INIT_MMX mmx
 HF_NOISE 8
 HF_NOISE 16
+
+;---------------------------------------------------------------------------------------
+;int ff_sad_<opt>(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, int stride, int h);
+;---------------------------------------------------------------------------------------
+INIT_MMX mmxext
+cglobal sad8, 4, 4, 0, v, pix1, pix2, stride
+    movu      m2, [pix2q]
+    movu      m1, [pix2q+strideq]
+    psadbw    m2, [pix1q]
+    psadbw    m1, [pix1q+strideq]
+    paddw     m2, m1
+
+%rep 3
+    lea    pix1q, [pix1q+strideq*2]
+    lea    pix2q, [pix2q+strideq*2]
+    movu      m0, [pix2q]
+    movu      m1, [pix2q+strideq]
+    psadbw    m0, [pix1q]
+    psadbw    m1, [pix1q+strideq]
+    paddw     m2, m0
+    paddw     m2, m1
+%endrep
+    movd     eax, m2
+    RET
+
+%macro SAD16 0
+cglobal sad16, 5, 5, 3, v, pix1, pix2, stride, h
+    pxor      m2, m2
+
+align 16
+.loop
+    movu      m0, [pix2q]
+    movu      m1, [pix2q+strideq]
+    psadbw    m0, [pix1q]
+    psadbw    m1, [pix1q+strideq]
+    paddw     m2, m0
+    paddw     m2, m1
+%if mmsize == 8
+    movu      m0, [pix2q+8]
+    movu      m1, [pix2q+strideq+8]
+    psadbw    m0, [pix1q+8]
+    psadbw    m1, [pix1q+strideq+8]
+    paddw     m2, m0
+    paddw     m2, m1
+%endif
+    lea    pix1q, [pix1q+strideq*2]
+    lea    pix2q, [pix2q+strideq*2]
+    sub       hd, 2
+    jg .loop
+%if mmsize == 16
+    movhlps   m0, m2
+    paddw     m2, m0
+%endif
+    movd     eax, m2
+    RET
+%endmacro
+
+INIT_MMX mmxext
+SAD16
+INIT_XMM sse2
+SAD16
+
+;------------------------------------------------------------------------------------------
+;int ff_sad_x2_<opt>(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, int stride, int h);
+;------------------------------------------------------------------------------------------
+INIT_MMX mmxext
+cglobal sad8_x2, 4, 4, 0, v, pix1, pix2, stride
+    movu      m0, [pix2q]
+    movu      m2, [pix2q+strideq]
+    pavgb     m0, [pix2q+1]
+    pavgb     m2, [pix2q+strideq+1]
+    psadbw    m0, [pix1q]
+    psadbw    m2, [pix1q+strideq]
+    paddw     m0, m2
+
+%rep 3
+    lea    pix1q, [pix1q+2*strideq]
+    lea    pix2q, [pix2q+2*strideq]
+    movu      m1, [pix2q]
+    movu      m2, [pix2q+strideq]
+    pavgb     m1, [pix2q+1]
+    pavgb     m2, [pix2q+strideq+1]
+    psadbw    m1, [pix1q]
+    psadbw    m2, [pix1q+strideq]
+    paddw     m0, m1
+    paddw     m0, m2
+%endrep
+    movd     eax, m0
+    RET
+
+%macro SAD16_X2 0
+cglobal sad16_x2, 5, 5, 5, v, pix1, pix2, stride, h
+    pxor      m0, m0
+
+align 16
+.loop:
+    movu      m1, [pix2q]
+    movu      m2, [pix2q+strideq]
+%if mmsize == 16
+    movu      m3, [pix2q+1]
+    movu      m4, [pix2q+strideq+1]
+    pavgb     m1, m3
+    pavgb     m2, m4
+%else
+    pavgb     m1, [pix2q+1]
+    pavgb     m2, [pix2q+strideq+1]
+%endif
+    psadbw    m1, [pix1q]
+    psadbw    m2, [pix1q+strideq]
+    paddw     m0, m1
+    paddw     m0, m2
+%if mmsize == 8
+    movu      m1, [pix2q+8]
+    movu      m2, [pix2q+strideq+8]
+    pavgb     m1, [pix2q+9]
+    pavgb     m2, [pix2q+strideq+9]
+    psadbw    m1, [pix1q+8]
+    psadbw    m2, [pix1q+strideq+8]
+    paddw     m0, m1
+    paddw     m0, m2
+%endif
+    lea    pix1q, [pix1q+2*strideq]
+    lea    pix2q, [pix2q+2*strideq]
+    sub       hd, 2
+    jg .loop
+%if mmsize == 16
+    movhlps   m1, m0
+    paddw     m0, m1
+%endif
+    movd     eax, m0
+    RET
+%endmacro
+
+INIT_MMX mmxext
+SAD16_X2
+INIT_XMM sse2
+SAD16_X2
+
+;------------------------------------------------------------------------------------------
+;int ff_sad_y2_<opt>(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, int stride, int h);
+;------------------------------------------------------------------------------------------
+INIT_MMX mmxext
+cglobal sad8_y2, 4, 4, 0, v, pix1, pix2, stride
+    movu      m1, [pix2q]
+    movu      m0, [pix2q+strideq]
+    movu      m3, [pix2q+2*strideq]
+    pavgb     m1, m0
+    pavgb     m0, m3
+    add    pix2q, strideq
+    psadbw    m1, [pix1q]
+    psadbw    m0, [pix1q+strideq]
+    paddw     m0, m1
+    mova      m1, m3
+
+%rep 3
+    lea    pix1q, [pix1q+2*strideq]
+    lea    pix2q, [pix2q+2*strideq]
+    movu      m2, [pix2q]
+    movu      m3, [pix2q+strideq]
+    pavgb     m1, m2
+    pavgb     m2, m3
+    psadbw    m1, [pix1q]
+    psadbw    m2, [pix1q+strideq]
+    paddw     m0, m1
+    paddw     m0, m2
+    mova      m1, m3
+%endrep
+    movd     eax, m0
+    RET
+
+%macro SAD16_Y2 0
+cglobal sad16_y2, 5, 5, 4, v, pix1, pix2, stride, h
+    movu      m1, [pix2q]
+%if mmsize == 8
+    movu      m4, [pix2q+8]
+%endif
+    pxor      m0, m0
+    add    pix2q, strideq
+
+align 16
+.loop:
+    movu      m2, [pix2q]
+    movu      m3, [pix2q+strideq]
+    pavgb     m1, m2
+    pavgb     m2, m3
+    psadbw    m1, [pix1q]
+    psadbw    m2, [pix1q+strideq]
+    paddw     m0, m1
+    paddw     m0, m2
+    mova      m1, m3
+%if mmsize == 8
+    movu      m5, [pix2q+8]
+    movu      m6, [pix2q+strideq+8]
+    pavgb     m4, m5
+    pavgb     m5, m6
+    psadbw    m4, [pix1q+8]
+    psadbw    m5, [pix1q+strideq+8]
+    paddw     m0, m4
+    paddw     m0, m5
+    mova      m4, m6
+%endif
+    lea    pix1q, [pix1q+2*strideq]
+    lea    pix2q, [pix2q+2*strideq]
+    sub       hd, 2
+    jg .loop
+%if mmsize == 16
+    movhlps   m1, m0
+    paddw     m0, m1
+%endif
+    movd     eax, m0
+    RET
+%endmacro
+
+INIT_MMX mmxext
+SAD16_Y2
+INIT_XMM sse2
+SAD16_Y2
+
+;-------------------------------------------------------------------------------------------
+;int ff_sad_approx_xy2_<opt>(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, int stride, int h);
+;-------------------------------------------------------------------------------------------
+INIT_MMX mmxext
+cglobal sad8_approx_xy2, 4, 4, 0, v, pix1, pix2, stride
+    pxor      m0, m0
+    mova      m4, [pb_1]
+    movu      m1, [pix2q]
+    movu      m0, [pix2q+strideq]
+    movu      m3, [pix2q+2*strideq]
+    pavgb     m1, [pix2q+1]
+    pavgb     m0, [pix2q+strideq+1]
+    pavgb     m3, [pix2q+2*strideq+1]
+    psubusb   m0, m4
+    pavgb     m1, m0
+    pavgb     m0, m3
+    add    pix2q, strideq
+    psadbw    m1, [pix1q]
+    psadbw    m0, [pix1q+strideq]
+    paddw     m0, m1
+    mova      m1, m3
+
+%rep 3
+    lea    pix1q, [pix1q+2*strideq]
+    lea    pix2q, [pix2q+2*strideq]
+    movu      m2, [pix2q]
+    movu      m3, [pix2q+strideq]
+    pavgb     m2, [pix2q+1]
+    pavgb     m3, [pix2q+strideq+1]
+    psubusb   m2, m4
+    pavgb     m1, m2
+    pavgb     m2, m3
+    psadbw    m1, [pix1q]
+    psadbw    m2, [pix1q+strideq]
+    paddw     m0, m1
+    paddw     m0, m2
+    mova      m1, m3
+%endrep
+    movd     eax, m0
+    RET
+
+%macro SAD16_APPROX_XY2 0
+cglobal sad16_approx_xy2, 5, 5, 7, v, pix1, pix2, stride, h
+    pxor      m0, m0
+    mova      m4, [pb_1]
+
+    movu      m1, [pix2q]
+%if mmsize == 16
+    movu      m2, [pix2q+1]
+    pavgb     m1, m2
+%else
+    movu      m5, [pix2q+8]
+    pavgb     m1, [pix2q+1]
+    pavgb     m5, [pix2q+8+1]
+%endif
+    add    pix2q, strideq
+
+align 16
+.loop:
+    movu      m2, [pix2q]
+    movu      m3, [pix2q+strideq]
+%if mmsize == 16
+    movu      m5, [pix2q+1]
+    movu      m6, [pix2q+strideq+1]
+    pavgb     m2, m5
+    pavgb     m3, m6
+%else
+    pavgb     m2, [pix2q+1]
+    pavgb     m3, [pix2q+strideq+1]
+%endif
+    psubusb   m2, m4
+    pavgb     m1, m2
+    pavgb     m2, m3
+    psadbw    m1, [pix1q]
+    psadbw    m2, [pix1q+strideq]
+    paddw     m0, m1
+    paddw     m0, m2
+    mova      m1, m3
+%if mmsize == 8
+    movu      m6, [pix2q+8]
+    movu      m7, [pix2q+strideq+8]
+    pavgb     m6, [pix2q+8+1]
+    pavgb     m7, [pix2q+strideq+8+1]
+    psubusb   m6, m4
+    pavgb     m5, m6
+    pavgb     m6, m7
+    psadbw    m5, [pix1q+8]
+    psadbw    m6, [pix1q+strideq+8]
+    paddw     m0, m5
+    paddw     m0, m6
+    mova      m5, m7
+%endif
+    lea    pix1q, [pix1q+2*strideq]
+    lea    pix2q, [pix2q+2*strideq]
+    sub       hd, 2
+    jg .loop
+%if mmsize == 16
+    movhlps   m1, m0
+    paddw     m0, m1
+%endif
+    movd     eax, m0
+    RET
+%endmacro
+
+INIT_MMX mmxext
+SAD16_APPROX_XY2
+INIT_XMM sse2
+SAD16_APPROX_XY2
diff --git a/libavcodec/x86/me_cmp_init.c b/libavcodec/x86/me_cmp_init.c
index 21db2214c6..cb47d631f9 100644
--- a/libavcodec/x86/me_cmp_init.c
+++ b/libavcodec/x86/me_cmp_init.c
@@ -41,6 +41,30 @@ int ff_sse16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
                   int line_size, int h);
 int ff_hf_noise8_mmx(uint8_t *pix1, int lsize, int h);
 int ff_hf_noise16_mmx(uint8_t *pix1, int lsize, int h);
+int ff_sad8_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                   int stride, int h);
+int ff_sad16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                    int stride, int h);
+int ff_sad16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                  int stride, int h);
+int ff_sad8_x2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                      int stride, int h);
+int ff_sad16_x2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                       int stride, int h);
+int ff_sad16_x2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                     int stride, int h);
+int ff_sad8_y2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                      int stride, int h);
+int ff_sad16_y2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                       int stride, int h);
+int ff_sad16_y2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                     int stride, int h);
+int ff_sad8_approx_xy2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                              int stride, int h);
+int ff_sad16_approx_xy2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                               int stride, int h);
+int ff_sad16_approx_xy2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                             int stride, int h);
 
 #define hadamard_func(cpu)                                              \
     int ff_hadamard8_diff_ ## cpu(MpegEncContext *s, uint8_t *src1,     \
@@ -345,8 +369,6 @@ DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {
     0x0002000200020002ULL,
 };
 
-DECLARE_ASM_CONST(8, uint64_t, bone) = 0x0101010101010101LL;
-
 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 {
     x86_reg len = -(x86_reg)stride * h;
@@ -382,130 +404,6 @@ static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
         : "r" (blk1 - len), "r" (blk2 - len), "r" ((x86_reg) stride));
 }
 
-static inline void sad8_1_mmxext(uint8_t *blk1, uint8_t *blk2,
-                                 int stride, int h)
-{
-    __asm__ volatile (
-        ".p2align 4                     \n\t"
-        "1:                             \n\t"
-        "movq (%1), %%mm0               \n\t"
-        "movq (%1, %3), %%mm1           \n\t"
-        "psadbw (%2), %%mm0             \n\t"
-        "psadbw (%2, %3), %%mm1         \n\t"
-        "paddw %%mm0, %%mm6             \n\t"
-        "paddw %%mm1, %%mm6             \n\t"
-        "lea (%1,%3,2), %1              \n\t"
-        "lea (%2,%3,2), %2              \n\t"
-        "sub $2, %0                     \n\t"
-        " jg 1b                         \n\t"
-        : "+r" (h), "+r" (blk1), "+r" (blk2)
-        : "r" ((x86_reg) stride));
-}
-
-static int sad16_sse2(MpegEncContext *v, uint8_t *blk2, uint8_t *blk1,
-                      int stride, int h)
-{
-    int ret;
-    __asm__ volatile (
-        "pxor %%xmm2, %%xmm2            \n\t"
-        ".p2align 4                     \n\t"
-        "1:                             \n\t"
-        "movdqu (%1), %%xmm0            \n\t"
-        "movdqu (%1, %4), %%xmm1        \n\t"
-        "psadbw (%2), %%xmm0            \n\t"
-        "psadbw (%2, %4), %%xmm1        \n\t"
-        "paddw %%xmm0, %%xmm2           \n\t"
-        "paddw %%xmm1, %%xmm2           \n\t"
-        "lea (%1,%4,2), %1              \n\t"
-        "lea (%2,%4,2), %2              \n\t"
-        "sub $2, %0                     \n\t"
-        " jg 1b                         \n\t"
-        "movhlps %%xmm2, %%xmm0         \n\t"
-        "paddw   %%xmm0, %%xmm2         \n\t"
-        "movd    %%xmm2, %3             \n\t"
-        : "+r" (h), "+r" (blk1), "+r" (blk2), "=r" (ret)
-        : "r" ((x86_reg) stride));
-    return ret;
-}
-
-static inline void sad8_x2a_mmxext(uint8_t *blk1, uint8_t *blk2,
-                                   int stride, int h)
-{
-    __asm__ volatile (
-        ".p2align 4                     \n\t"
-        "1:                             \n\t"
-        "movq (%1), %%mm0               \n\t"
-        "movq (%1, %3), %%mm1           \n\t"
-        "pavgb 1(%1), %%mm0             \n\t"
-        "pavgb 1(%1, %3), %%mm1         \n\t"
-        "psadbw (%2), %%mm0             \n\t"
-        "psadbw (%2, %3), %%mm1         \n\t"
-        "paddw %%mm0, %%mm6             \n\t"
-        "paddw %%mm1, %%mm6             \n\t"
-        "lea (%1,%3,2), %1              \n\t"
-        "lea (%2,%3,2), %2              \n\t"
-        "sub $2, %0                     \n\t"
-        " jg 1b                         \n\t"
-        : "+r" (h), "+r" (blk1), "+r" (blk2)
-        : "r" ((x86_reg) stride));
-}
-
-static inline void sad8_y2a_mmxext(uint8_t *blk1, uint8_t *blk2,
-                                   int stride, int h)
-{
-    __asm__ volatile (
-        "movq (%1), %%mm0               \n\t"
-        "add %3, %1                     \n\t"
-        ".p2align 4                     \n\t"
-        "1:                             \n\t"
-        "movq (%1), %%mm1               \n\t"
-        "movq (%1, %3), %%mm2           \n\t"
-        "pavgb %%mm1, %%mm0             \n\t"
-        "pavgb %%mm2, %%mm1             \n\t"
-        "psadbw (%2), %%mm0             \n\t"
-        "psadbw (%2, %3), %%mm1         \n\t"
-        "paddw %%mm0, %%mm6             \n\t"
-        "paddw %%mm1, %%mm6             \n\t"
-        "movq %%mm2, %%mm0              \n\t"
-        "lea (%1,%3,2), %1              \n\t"
-        "lea (%2,%3,2), %2              \n\t"
-        "sub $2, %0                     \n\t"
-        " jg 1b                         \n\t"
-        : "+r" (h), "+r" (blk1), "+r" (blk2)
-        : "r" ((x86_reg) stride));
-}
-
-static inline void sad8_4_mmxext(uint8_t *blk1, uint8_t *blk2,
-                                 int stride, int h)
-{
-    __asm__ volatile (
-        "movq "MANGLE(bone)", %%mm5     \n\t"
-        "movq (%1), %%mm0               \n\t"
-        "pavgb 1(%1), %%mm0             \n\t"
-        "add %3, %1                     \n\t"
-        ".p2align 4                     \n\t"
-        "1:                             \n\t"
-        "movq (%1), %%mm1               \n\t"
-        "movq (%1,%3), %%mm2            \n\t"
-        "pavgb 1(%1), %%mm1             \n\t"
-        "pavgb 1(%1,%3), %%mm2          \n\t"
-        "psubusb %%mm5, %%mm1           \n\t"
-        "pavgb %%mm1, %%mm0             \n\t"
-        "pavgb %%mm2, %%mm1             \n\t"
-        "psadbw (%2), %%mm0             \n\t"
-        "psadbw (%2,%3), %%mm1          \n\t"
-        "paddw %%mm0, %%mm6             \n\t"
-        "paddw %%mm1, %%mm6             \n\t"
-        "movq %%mm2, %%mm0              \n\t"
-        "lea (%1,%3,2), %1              \n\t"
-        "lea (%2,%3,2), %2              \n\t"
-        "sub $2, %0                     \n\t"
-        " jg 1b                         \n\t"
-        : "+r" (h), "+r" (blk1), "+r" (blk2)
-        : "r" ((x86_reg) stride)
-          NAMED_CONSTRAINTS_ADD(bone));
-}
-
 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2,
                               int stride, int h)
 {
@@ -613,15 +511,6 @@ static inline int sum_mmx(void)
     return ret & 0xFFFF;
 }
 
-static inline int sum_mmxext(void)
-{
-    int ret;
-    __asm__ volatile (
-        "movd %%mm6, %0                 \n\t"
-        : "=r" (ret));
-    return ret;
-}
-
 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 {
     sad8_2_mmx(blk1, blk1 + 1, blk2, stride, h);
@@ -750,7 +639,6 @@ static int sad16_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2,          \
 }                                                                       \
 
 PIX_SAD(mmx)
-PIX_SAD(mmxext)
 
 #endif /* HAVE_INLINE_ASM */
 
@@ -782,29 +670,11 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
     if (INLINE_MMXEXT(cpu_flags)) {
         c->vsad[4] = vsad_intra16_mmxext;
 
-        c->pix_abs[0][0] = sad16_mmxext;
-        c->pix_abs[1][0] = sad8_mmxext;
-
-        c->sad[0] = sad16_mmxext;
-        c->sad[1] = sad8_mmxext;
-
-        c->pix_abs[0][1] = sad16_x2_mmxext;
-        c->pix_abs[0][2] = sad16_y2_mmxext;
-        c->pix_abs[1][1] = sad8_x2_mmxext;
-        c->pix_abs[1][2] = sad8_y2_mmxext;
-
         if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
-            c->pix_abs[0][3] = sad16_xy2_mmxext;
-            c->pix_abs[1][3] = sad8_xy2_mmxext;
-
             c->vsad[0] = vsad16_mmxext;
         }
     }
 
-    if (INLINE_SSE2(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_SSE2SLOW) && avctx->codec_id != AV_CODEC_ID_SNOW) {
-        c->sad[0] = sad16_sse2;
-    }
-
 #endif /* HAVE_INLINE_ASM */
 
     if (EXTERNAL_MMX(cpu_flags)) {
@@ -823,6 +693,21 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmxext;
         c->hadamard8_diff[1] = ff_hadamard8_diff_mmxext;
         c->sum_abs_dctelem   = ff_sum_abs_dctelem_mmxext;
+
+        c->sad[0] = ff_sad16_mmxext;
+        c->sad[1] = ff_sad8_mmxext;
+
+        c->pix_abs[0][0] = ff_sad16_mmxext;
+        c->pix_abs[0][1] = ff_sad16_x2_mmxext;
+        c->pix_abs[0][2] = ff_sad16_y2_mmxext;
+        c->pix_abs[1][0] = ff_sad8_mmxext;
+        c->pix_abs[1][1] = ff_sad8_x2_mmxext;
+        c->pix_abs[1][2] = ff_sad8_y2_mmxext;
+
+        if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
+            c->pix_abs[0][3] = ff_sad16_approx_xy2_mmxext;
+            c->pix_abs[1][3] = ff_sad8_approx_xy2_mmxext;
+        }
     }
 
     if (EXTERNAL_SSE2(cpu_flags)) {
@@ -833,6 +718,16 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
         c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;
         c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;
 #endif
+        if (!(cpu_flags & AV_CPU_FLAG_SSE2SLOW) && avctx->codec_id != AV_CODEC_ID_SNOW) {
+            c->sad[0]        = ff_sad16_sse2;
+            c->pix_abs[0][0] = ff_sad16_sse2;
+            c->pix_abs[0][1] = ff_sad16_x2_sse2;
+            c->pix_abs[0][2] = ff_sad16_y2_sse2;
+
+            if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
+                c->pix_abs[0][3] = ff_sad16_approx_xy2_sse2;
+            }
+        }
     }
 
     if (EXTERNAL_SSSE3(cpu_flags)) {