motion_comp_mmx.c

从FFMPEG转换而来的H264解码程序,VC下编译..

} while (0)

#define pavg_m2r(src,dest)		\
do {					\
    if (cpu == CPU_MMXEXT)		\
	pavgb_m2r (src, dest);		\
/*    else				\
	pavgusb_m2r (src, dest);*/	\
} while (0)


/* CPU_MMXEXT code */


static inline void MC_put1_8 (int height, uint8_t * dest, const uint8_t * ref,
			      const int stride)
{
    do {
        __m64 mm0;
	movq_m2r (*ref, mm0);
	movq_r2m (mm0, *dest);
	ref += stride;
	dest += stride;
    } while (--height);
}

static inline void MC_put1_16 (int height, uint8_t * dest, const uint8_t * ref,
			       const int stride)
{
    do {
        __m64 mm0,mm1;
	movq_m2r (*ref, mm0);
	movq_m2r (*(ref+8), mm1);
	ref += stride;
	movq_r2m (mm0, *dest);
	movq_r2m (mm1, *(dest+8));
	dest += stride;
    } while (--height);
}

static inline void MC_avg1_8 (int height, uint8_t * dest, const uint8_t * ref,
			      const int stride, const int cpu)
{
    do {
        __m64 mm0;
	movq_m2r (*ref, mm0);
	pavg_m2r (*dest, mm0);
	ref += stride;
	movq_r2m (mm0, *dest);
	dest += stride;
    } while (--height);
}

static inline void MC_avg1_16 (int height, uint8_t * dest, const uint8_t * ref,
			       const int stride, const int cpu)
{
    do {
        __m64 mm0,mm1;
	movq_m2r (*ref, mm0);
	movq_m2r (*(ref+8), mm1);
	pavg_m2r (*dest, mm0);
	pavg_m2r (*(dest+8), mm1);
	movq_r2m (mm0, *dest);
	ref += stride;
	movq_r2m (mm1, *(dest+8));
	dest += stride;
    } while (--height);
}

static inline void MC_put2_8 (int height, uint8_t * dest, const uint8_t * ref,
			      const int stride, const int offset,
			      const int cpu)
{
    do {
        __m64 mm0;
	movq_m2r (*ref, mm0);
	pavg_m2r (*(ref+offset), mm0);
	ref += stride;
	movq_r2m (mm0, *dest);
	dest += stride;
    } while (--height);
}

static inline void MC_put2_16 (int height, uint8_t * dest, const uint8_t * ref,
			       const int stride, const int offset,
			       const int cpu)
{
    do {
        __m64 mm0,mm1;
	movq_m2r (*ref, mm0);
	movq_m2r (*(ref+8), mm1);
	pavg_m2r (*(ref+offset), mm0);
	pavg_m2r (*(ref+offset+8), mm1);
	movq_r2m (mm0, *dest);
	ref += stride;
	movq_r2m (mm1, *(dest+8));
	dest += stride;
    } while (--height);
}

static inline void MC_avg2_8 (int height, uint8_t * dest, const uint8_t * ref,
			      const int stride, const int offset,
			      const int cpu)
{
    do {
        __m64 mm0;
	movq_m2r (*ref, mm0);
	pavg_m2r (*(ref+offset), mm0);
	pavg_m2r (*dest, mm0);
	ref += stride;
	movq_r2m (mm0, *dest);
	dest += stride;
    } while (--height);
}

static inline void MC_avg2_16 (int height, uint8_t * dest, const uint8_t * ref,
			       const int stride, const int offset,
			       const int cpu)
{
    do {
        __m64 mm0,mm1;
	movq_m2r (*ref, mm0);
	movq_m2r (*(ref+8), mm1);
	pavg_m2r (*(ref+offset), mm0);
	pavg_m2r (*(ref+offset+8), mm1);
	pavg_m2r (*dest, mm0);
	pavg_m2r (*(dest+8), mm1);
	ref += stride;
	movq_r2m (mm0, *dest);
	movq_r2m (mm1, *(dest+8));
	dest += stride;
    } while (--height);
}

static inline void MC_put4_8 (int height, uint8_t * dest, const uint8_t * ref,
			      const int stride, const int cpu)
{
    __m64 mm0,mm1,mm2,mm3,mm5,mm6,mm7;
    __m64 mask_one = _mm_set1_pi8(1);

    movq_m2r (*ref, mm0);
    movq_m2r (*(ref+1), mm1);
    movq_r2r (mm0, mm7);
    pxor_r2r (mm1, mm7);
    pavg_r2r (mm1, mm0);
    ref += stride;

    do {
	movq_m2r (*ref, mm2);
	movq_r2r (mm0, mm5);

	movq_m2r (*(ref+1), mm3);
	movq_r2r (mm2, mm6);

	pxor_r2r (mm3, mm6);
	pavg_r2r (mm3, mm2);

	por_r2r (mm6, mm7);
	pxor_r2r (mm2, mm5);

	pand_r2r (mm5, mm7);
	pavg_r2r (mm2, mm0);

	pand_m2r (mask_one, mm7);

	psubusb_r2r (mm7, mm0);

	ref += stride;
	movq_r2m (mm0, *dest);
	dest += stride;

	movq_r2r (mm6, mm7);	/* unroll ! */
	movq_r2r (mm2, mm0);	/* unroll ! */
    } while (--height);
}

static inline void MC_put4_16 (int height, uint8_t * dest, const uint8_t * ref,
			       const int stride, const int cpu)
{
    __m64 mask_one = _mm_set1_pi8(1);
    do {
            __m64 mm0,mm1,mm2,mm3,mm6,mm7;

	movq_m2r (*ref, mm0);
	movq_m2r (*(ref+stride+1), mm1);
	movq_r2r (mm0, mm7);
	movq_m2r (*(ref+1), mm2);
	pxor_r2r (mm1, mm7);
	movq_m2r (*(ref+stride), mm3);
	movq_r2r (mm2, mm6);
	pxor_r2r (mm3, mm6);
	pavg_r2r (mm1, mm0);
	pavg_r2r (mm3, mm2);
	por_r2r (mm6, mm7);
	movq_r2r (mm0, mm6);
	pxor_r2r (mm2, mm6);
	pand_r2r (mm6, mm7);
	pand_m2r (mask_one, mm7);
	pavg_r2r (mm2, mm0);
	psubusb_r2r (mm7, mm0);
	movq_r2m (mm0, *dest);

	movq_m2r (*(ref+8), mm0);
	movq_m2r (*(ref+stride+9), mm1);
	movq_r2r (mm0, mm7);
	movq_m2r (*(ref+9), mm2);
	pxor_r2r (mm1, mm7);
	movq_m2r (*(ref+stride+8), mm3);
	movq_r2r (mm2, mm6);
	pxor_r2r (mm3, mm6);
	pavg_r2r (mm1, mm0);
	pavg_r2r (mm3, mm2);
	por_r2r (mm6, mm7);
	movq_r2r (mm0, mm6);
	pxor_r2r (mm2, mm6);
	pand_r2r (mm6, mm7);
	pand_m2r (mask_one, mm7);
	pavg_r2r (mm2, mm0);
	psubusb_r2r (mm7, mm0);
	ref += stride;
	movq_r2m (mm0, *(dest+8));
	dest += stride;
    } while (--height);
}

static inline void MC_avg4_8 (int height, uint8_t * dest, const uint8_t * ref,
			      const int stride, const int cpu)
{
    __m64 mask_one = _mm_set1_pi8(1);
    do {
            __m64 mm0,mm1,mm2,mm3,mm6,mm7;

	movq_m2r (*ref, mm0);
	movq_m2r (*(ref+stride+1), mm1);
	movq_r2r (mm0, mm7);
	movq_m2r (*(ref+1), mm2);
	pxor_r2r (mm1, mm7);
	movq_m2r (*(ref+stride), mm3);
	movq_r2r (mm2, mm6);
	pxor_r2r (mm3, mm6);
	pavg_r2r (mm1, mm0);
	pavg_r2r (mm3, mm2);
	por_r2r (mm6, mm7);
	movq_r2r (mm0, mm6);
	pxor_r2r (mm2, mm6);
	pand_r2r (mm6, mm7);
	pand_m2r (mask_one, mm7);
	pavg_r2r (mm2, mm0);
	psubusb_r2r (mm7, mm0);
	movq_m2r (*dest, mm1);
	pavg_r2r (mm1, mm0);
	ref += stride;
	movq_r2m (mm0, *dest);
	dest += stride;
    } while (--height);
}

static inline void MC_avg4_16 (int height, uint8_t * dest, const uint8_t * ref,
			       const int stride, const int cpu)
{
    __m64 mask_one = _mm_set1_pi8(1);
    do {
            __m64 mm0,mm1,mm2,mm3,mm6,mm7;

	movq_m2r (*ref, mm0);
	movq_m2r (*(ref+stride+1), mm1);
	movq_r2r (mm0, mm7);
	movq_m2r (*(ref+1), mm2);
	pxor_r2r (mm1, mm7);
	movq_m2r (*(ref+stride), mm3);
	movq_r2r (mm2, mm6);
	pxor_r2r (mm3, mm6);
	pavg_r2r (mm1, mm0);
	pavg_r2r (mm3, mm2);
	por_r2r (mm6, mm7);
	movq_r2r (mm0, mm6);
	pxor_r2r (mm2, mm6);
	pand_r2r (mm6, mm7);
	pand_m2r (mask_one, mm7);
	pavg_r2r (mm2, mm0);
	psubusb_r2r (mm7, mm0);
	movq_m2r (*dest, mm1);
	pavg_r2r (mm1, mm0);
	movq_r2m (mm0, *dest);

	movq_m2r (*(ref+8), mm0);
	movq_m2r (*(ref+stride+9), mm1);
	movq_r2r (mm0, mm7);
	movq_m2r (*(ref+9), mm2);
	pxor_r2r (mm1, mm7);
	movq_m2r (*(ref+stride+8), mm3);
	movq_r2r (mm2, mm6);
	pxor_r2r (mm3, mm6);
	pavg_r2r (mm1, mm0);
	pavg_r2r (mm3, mm2);
	por_r2r (mm6, mm7);
	movq_r2r (mm0, mm6);
	pxor_r2r (mm2, mm6);
	pand_r2r (mm6, mm7);
	pand_m2r (mask_one, mm7);
	pavg_r2r (mm2, mm0);
	psubusb_r2r (mm7, mm0);
	movq_m2r (*(dest+8), mm1);
	pavg_r2r (mm1, mm0);
	ref += stride;
	movq_r2m (mm0, *(dest+8));
	dest += stride;
    } while (--height);
}

static void MC_avg_o_16_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_avg1_16 (height, dest, ref, stride, CPU_MMXEXT);
}

static void MC_avg_o_8_mmxext (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_avg1_8 (height, dest, ref, stride, CPU_MMXEXT);
}

static void MC_put_o_16_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_put1_16 (height, dest, ref, stride);
}

static void MC_put_o_8_mmxext (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_put1_8 (height, dest, ref, stride);
}

static void MC_avg_x_16_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_avg2_16 (height, dest, ref, stride, 1, CPU_MMXEXT);
}

static void MC_avg_x_8_mmxext (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_avg2_8 (height, dest, ref, stride, 1, CPU_MMXEXT);
}

static void MC_put_x_16_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_put2_16 (height, dest, ref, stride, 1, CPU_MMXEXT);
}

static void MC_put_x_8_mmxext (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_put2_8 (height, dest, ref, stride, 1, CPU_MMXEXT);
}

static void MC_avg_y_16_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_avg2_16 (height, dest, ref, stride, stride, CPU_MMXEXT);
}

static void MC_avg_y_8_mmxext (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_avg2_8 (height, dest, ref, stride, stride, CPU_MMXEXT);
}

static void MC_put_y_16_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_put2_16 (height, dest, ref, stride, stride, CPU_MMXEXT);
}

static void MC_put_y_8_mmxext (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_put2_8 (height, dest, ref, stride, stride, CPU_MMXEXT);
}

static void MC_avg_xy_16_mmxext (uint8_t * dest, const uint8_t * ref,
				 int stride, int height)
{
    MC_avg4_16 (height, dest, ref, stride, CPU_MMXEXT);
}

static void MC_avg_xy_8_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_avg4_8 (height, dest, ref, stride, CPU_MMXEXT);
}

static void MC_put_xy_16_mmxext (uint8_t * dest, const uint8_t * ref,
				 int stride, int height)
{
    MC_put4_16 (height, dest, ref, stride, CPU_MMXEXT);
}

static void MC_put_xy_8_mmxext (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_put4_8 (height, dest, ref, stride, CPU_MMXEXT);
}


MPEG2_MC_EXTERN (mmxext)


/*
static void MC_avg_o_16_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_avg1_16 (height, dest, ref, stride, CPU_3DNOW);
}

static void MC_avg_o_8_3dnow (uint8_t * dest, const uint8_t * ref,
			      int stride, int height)
{
    MC_avg1_8 (height, dest, ref, stride, CPU_3DNOW);
}

static void MC_put_o_16_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_put1_16 (height, dest, ref, stride);
}

static void MC_put_o_8_3dnow (uint8_t * dest, const uint8_t * ref,
			      int stride, int height)
{
    MC_put1_8 (height, dest, ref, stride);
}

static void MC_avg_x_16_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_avg2_16 (height, dest, ref, stride, 1, CPU_3DNOW);
}

static void MC_avg_x_8_3dnow (uint8_t * dest, const uint8_t * ref,
			      int stride, int height)
{
    MC_avg2_8 (height, dest, ref, stride, 1, CPU_3DNOW);
}

static void MC_put_x_16_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_put2_16 (height, dest, ref, stride, 1, CPU_3DNOW);
}

static void MC_put_x_8_3dnow (uint8_t * dest, const uint8_t * ref,
			      int stride, int height)
{
    MC_put2_8 (height, dest, ref, stride, 1, CPU_3DNOW);
}

static void MC_avg_y_16_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_avg2_16 (height, dest, ref, stride, stride, CPU_3DNOW);
}

static void MC_avg_y_8_3dnow (uint8_t * dest, const uint8_t * ref,
			      int stride, int height)
{
    MC_avg2_8 (height, dest, ref, stride, stride, CPU_3DNOW);
}

static void MC_put_y_16_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_put2_16 (height, dest, ref, stride, stride, CPU_3DNOW);
}

static void MC_put_y_8_3dnow (uint8_t * dest, const uint8_t * ref,
			      int stride, int height)
{
    MC_put2_8 (height, dest, ref, stride, stride, CPU_3DNOW);
}

static void MC_avg_xy_16_3dnow (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_avg4_16 (height, dest, ref, stride, CPU_3DNOW);
}

static void MC_avg_xy_8_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_avg4_8 (height, dest, ref, stride, CPU_3DNOW);
}

static void MC_put_xy_16_3dnow (uint8_t * dest, const uint8_t * ref,
				int stride, int height)
{
    MC_put4_16 (height, dest, ref, stride, CPU_3DNOW);
}

static void MC_put_xy_8_3dnow (uint8_t * dest, const uint8_t * ref,
			       int stride, int height)
{
    MC_put4_8 (height, dest, ref, stride, CPU_3DNOW);
}


MPEG2_MC_EXTERN (3dnow)
*/
#endif