+/* Same as the above, except we do (mv + 2) >> 2 on the input. */
+#define x264_predictor_roundclip x264_predictor_roundclip_mmx2
+static int ALWAYS_INLINE x264_predictor_roundclip_mmx2( int16_t (*dst)[2], int16_t (*mvc)[2], int i_mvc, int16_t mv_limit[2][2], uint32_t pmv )
+{
+ static const uint64_t pw_2 = 0x0002000200020002ULL;
+ static const uint32_t pd_32 = 0x20;
+ intptr_t tmp = (intptr_t)mv_limit, mvc_max = i_mvc, i = 0;
+
+ asm(
+ "movq (%2), %%mm5 \n"
+ "movq %6, %%mm7 \n"
+ "movd %7, %%mm3 \n"
+ "pshufw $0xEE, %%mm5, %%mm6 \n"
+ "dec %k3 \n"
+ "jz 2f \n"
+ "punpckldq %%mm3, %%mm3 \n"
+ "punpckldq %%mm5, %%mm5 \n"
+ "movd %8, %%mm4 \n"
+ "lea (%0,%3,4), %3 \n"
+ "1: \n"
+ "movq (%0), %%mm0 \n"
+ "add $8, %0 \n"
+ "paddw %%mm7, %%mm0 \n"
+ "psraw $2, %%mm0 \n"
+ "movq %%mm3, %%mm1 \n"
+ "pxor %%mm2, %%mm2 \n"
+ "pcmpeqd %%mm0, %%mm1 \n"
+ "pcmpeqd %%mm0, %%mm2 \n"
+ "por %%mm1, %%mm2 \n"
+ "pmovmskb %%mm2, %k2 \n"
+ "pmaxsw %%mm5, %%mm0 \n"
+ "pminsw %%mm6, %%mm0 \n"
+ "pand %%mm4, %%mm2 \n"
+ "psrlq %%mm2, %%mm0 \n"
+ "movq %%mm0, (%5,%4,4) \n"
+ "and $24, %k2 \n"
+ "add $2, %4 \n"
+ "add $8, %k2 \n"
+ "shr $4, %k2 \n"
+ "sub %2, %4 \n"
+ "cmp %3, %0 \n"
+ "jl 1b \n"
+ "jg 3f \n"
+
+ /* Do the last iteration */
+ "2: \n"
+ "movd (%0), %%mm0 \n"
+ "paddw %%mm7, %%mm0 \n"
+ "psraw $2, %%mm0 \n"
+ "pxor %%mm2, %%mm2 \n"
+ "pcmpeqd %%mm0, %%mm3 \n"
+ "pcmpeqd %%mm0, %%mm2 \n"
+ "por %%mm3, %%mm2 \n"
+ "pmovmskb %%mm2, %k2 \n"
+ "pmaxsw %%mm5, %%mm0 \n"
+ "pminsw %%mm6, %%mm0 \n"
+ "movd %%mm0, (%5,%4,4) \n"
+ "inc %4 \n"
+ "and $1, %k2 \n"
+ "sub %2, %4 \n"
+ "3: \n"
+ :"+r"(mvc), "=m"(M64( dst )), "+r"(tmp), "+r"(mvc_max), "+r"(i)
+ :"r"(dst), "m"(pw_2), "g"(pmv), "m"(pd_32), "m"(M64( mvc ))
+ );
+ return i;
+}