#ifndef DSPUTIL_H
#define DSPUTIL_H

#include "common.h"

/* dct code */
typedef short DCTELEM;

void jpeg_fdct_ifast (DCTELEM *data);

void j_rev_dct (DCTELEM *data);

void fdct_mmx(DCTELEM *block);

void (*av_fdct)(DCTELEM *block);

/* encoding scans */
extern UINT8 ff_alternate_horizontal_scan[64];
extern UINT8 ff_alternate_vertical_scan[64];
extern UINT8 zigzag_direct[64];

/* pixel operations */
#define MAX_NEG_CROP 384

/* temporary */
extern UINT32 squareTbl[512];
extern UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];

void dsputil_init(void);

/* pixel ops : interface with DCT */

extern void (*ff_idct)(DCTELEM *block);
extern void (*get_pixels)(DCTELEM *block, const UINT8 *pixels, int line_size);
extern void (*put_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);
extern void (*add_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);

void get_pixels_c(DCTELEM *block, const UINT8 *pixels, int line_size);
void put_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size);
void add_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size);

/* add and put pixel (decoding) */
typedef void (*op_pixels_func)(UINT8 *block, const UINT8 *pixels, int line_size, int h);

extern op_pixels_func put_pixels_tab[4];
extern op_pixels_func avg_pixels_tab[4];
extern op_pixels_func put_no_rnd_pixels_tab[4];
extern op_pixels_func avg_no_rnd_pixels_tab[4];

/* sub pixel (encoding) */
extern void (*sub_pixels_tab[4])(DCTELEM *block, const UINT8 *pixels, int line_size, int h);

#define sub_pixels_2(block, pixels, line_size, dxy) \
   sub_pixels_tab[dxy](block, pixels, line_size, 8)

/* motion estimation */

typedef int (*op_pixels_abs_func)(UINT8 *blk1, UINT8 *blk2, int line_size, int h);

extern op_pixels_abs_func pix_abs16x16;
extern op_pixels_abs_func pix_abs16x16_x2;
extern op_pixels_abs_func pix_abs16x16_y2;
extern op_pixels_abs_func pix_abs16x16_xy2;

int pix_abs16x16_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_x2_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_y2_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_xy2_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);

static inline int block_permute_op(int j)
{
    return (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
}

void block_permute(INT16 *block);

#if defined(ARCH_X86)

#include "cpu_accel.h"

extern int mm_flags;

#define emms_c() \
{\
    if (mm_flags & MM_MMX)\
        emms();\
}

#define __align8 __attribute__ ((aligned (8)))

void dsputil_init_mmx(void);

#elif defined(ARCH_ARMV4L)

/* This is to use 4 bytes read to the IDCT pointers for some 'zero'
   line ptimizations */
#define __align8 __attribute__ ((aligned (4)))

void dsputil_init_armv4l(void);   

#endif



#if defined(HAVE_MLIB)

/* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
#define __align8 __attribute__ ((aligned (8)))

void dsputil_init_mlib(void);

#endif	/* HAVE_MLIB */


/*
 * provide empty defaults, if the target specific accelerated dsputils did
 * not define these:
 */
#ifndef	__align8
#define	__align8
#endif

#ifndef	emms_c
#define	emms_c()
#endif

#endif