/* autogenerated from gstvideoscaleorc.orc */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <glib.h>

#ifndef _ORC_INTEGER_TYPEDEFS_
#define _ORC_INTEGER_TYPEDEFS_
#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
#include <stdint.h>
typedef int8_t orc_int8;
typedef int16_t orc_int16;
typedef int32_t orc_int32;
typedef int64_t orc_int64;
typedef uint8_t orc_uint8;
typedef uint16_t orc_uint16;
typedef uint32_t orc_uint32;
typedef uint64_t orc_uint64;
#define ORC_UINT64_C(x) UINT64_C(x)
#elif defined(_MSC_VER)
typedef signed __int8 orc_int8;
typedef signed __int16 orc_int16;
typedef signed __int32 orc_int32;
typedef signed __int64 orc_int64;
typedef unsigned __int8 orc_uint8;
typedef unsigned __int16 orc_uint16;
typedef unsigned __int32 orc_uint32;
typedef unsigned __int64 orc_uint64;
#define ORC_UINT64_C(x) (x##Ui64)
#define inline __inline
#else
#include <limits.h>
typedef signed char orc_int8;
typedef short orc_int16;
typedef int orc_int32;
typedef unsigned char orc_uint8;
typedef unsigned short orc_uint16;
typedef unsigned int orc_uint32;
#if INT_MAX == LONG_MAX
typedef long long orc_int64;
typedef unsigned long long orc_uint64;
#define ORC_UINT64_C(x) (x##ULL)
#else
typedef long orc_int64;
typedef unsigned long orc_uint64;
#define ORC_UINT64_C(x) (x##UL)
#endif
#endif
typedef union
{
  orc_int16 i;
  orc_int8 x2[2];
} orc_union16;
typedef union
{
  orc_int32 i;
  float f;
  orc_int16 x2[2];
  orc_int8 x4[4];
} orc_union32;
typedef union
{
  orc_int64 i;
  double f;
  orc_int32 x2[2];
  float x2f[2];
  orc_int16 x4[4];
} orc_union64;
#endif
#ifndef ORC_RESTRICT
#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
#define ORC_RESTRICT restrict
#elif defined(__GNUC__) && __GNUC__ >= 4
#define ORC_RESTRICT __restrict__
#else
#define ORC_RESTRICT
#endif
#endif

#ifndef DISABLE_ORC
#include <orc/orc.h>
#endif
void video_scale_orc_merge_linear_u8 (orc_uint8 * ORC_RESTRICT d1,
    const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
    int p1, int n);
void video_scale_orc_merge_linear_u16 (orc_uint16 * ORC_RESTRICT d1,
    const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
    int p1, int p2, int n);
void video_scale_orc_splat_u16 (orc_uint16 * ORC_RESTRICT d1, int p1, int n);
void video_scale_orc_splat_u32 (orc_uint32 * ORC_RESTRICT d1, int p1, int n);
void video_scale_orc_splat_u64 (orc_uint64 * ORC_RESTRICT d1, orc_int64 p1,
    int n);
void video_scale_orc_downsample_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n);
void video_scale_orc_downsample_u16 (guint16 * ORC_RESTRICT d1,
    const guint16 * ORC_RESTRICT s1, int n);
void video_scale_orc_downsample_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n);
void video_scale_orc_downsample_yuyv (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n);
void video_scale_orc_resample_nearest_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n);
void video_scale_orc_resample_bilinear_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n);
void video_scale_orc_resample_nearest_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n);
void video_scale_orc_resample_bilinear_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n);
void video_scale_orc_resample_merge_bilinear_u32 (guint8 * ORC_RESTRICT d1,
    guint8 * ORC_RESTRICT d2, const guint8 * ORC_RESTRICT s1,
    const guint8 * ORC_RESTRICT s2, int p1, int p2, int p3, int n);
void video_scale_orc_merge_bicubic_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
    int p2, int p3, int p4, int n);


/* begin Orc C target preamble */
#define ORC_CLAMP(x,a,b) ((x)<(a) ? (a) : ((x)>(b) ? (b) : (x)))
#define ORC_ABS(a) ((a)<0 ? -(a) : (a))
#define ORC_MIN(a,b) ((a)<(b) ? (a) : (b))
#define ORC_MAX(a,b) ((a)>(b) ? (a) : (b))
#define ORC_SB_MAX 127
#define ORC_SB_MIN (-1-ORC_SB_MAX)
#define ORC_UB_MAX 255
#define ORC_UB_MIN 0
#define ORC_SW_MAX 32767
#define ORC_SW_MIN (-1-ORC_SW_MAX)
#define ORC_UW_MAX 65535
#define ORC_UW_MIN 0
#define ORC_SL_MAX 2147483647
#define ORC_SL_MIN (-1-ORC_SL_MAX)
#define ORC_UL_MAX 4294967295U
#define ORC_UL_MIN 0
#define ORC_CLAMP_SB(x) ORC_CLAMP(x,ORC_SB_MIN,ORC_SB_MAX)
#define ORC_CLAMP_UB(x) ORC_CLAMP(x,ORC_UB_MIN,ORC_UB_MAX)
#define ORC_CLAMP_SW(x) ORC_CLAMP(x,ORC_SW_MIN,ORC_SW_MAX)
#define ORC_CLAMP_UW(x) ORC_CLAMP(x,ORC_UW_MIN,ORC_UW_MAX)
#define ORC_CLAMP_SL(x) ORC_CLAMP(x,ORC_SL_MIN,ORC_SL_MAX)
#define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX)
#define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))
#define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))
#define ORC_SWAP_Q(x) ((((x)&ORC_UINT64_C(0xff))<<56) | (((x)&ORC_UINT64_C(0xff00))<<40) | (((x)&ORC_UINT64_C(0xff0000))<<24) | (((x)&ORC_UINT64_C(0xff000000))<<8) | (((x)&ORC_UINT64_C(0xff00000000))>>8) | (((x)&ORC_UINT64_C(0xff0000000000))>>24) | (((x)&ORC_UINT64_C(0xff000000000000))>>40) | (((x)&ORC_UINT64_C(0xff00000000000000))>>56))
#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))
#define ORC_DENORMAL(x) ((x) & ((((x)&0x7f800000) == 0) ? 0xff800000 : 0xffffffff))
#define ORC_ISNAN(x) ((((x)&0x7f800000) == 0x7f800000) && (((x)&0x007fffff) != 0))
#define ORC_DENORMAL_DOUBLE(x) ((x) & ((((x)&ORC_UINT64_C(0x7ff0000000000000)) == 0) ? ORC_UINT64_C(0xfff0000000000000) : ORC_UINT64_C(0xffffffffffffffff)))
#define ORC_ISNAN_DOUBLE(x) ((((x)&ORC_UINT64_C(0x7ff0000000000000)) == ORC_UINT64_C(0x7ff0000000000000)) && (((x)&ORC_UINT64_C(0x000fffffffffffff)) != 0))
#ifndef ORC_RESTRICT
#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
#define ORC_RESTRICT restrict
#elif defined(__GNUC__) && __GNUC__ >= 4
#define ORC_RESTRICT __restrict__
#else
#define ORC_RESTRICT
#endif
#endif
/* end Orc C target preamble */



/* video_scale_orc_merge_linear_u8 */
#ifdef DISABLE_ORC
void
video_scale_orc_merge_linear_u8 (orc_uint8 * ORC_RESTRICT d1,
    const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
    int p1, int n)
{
  int i;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  const orc_int8 *ORC_RESTRICT ptr5;
  orc_int8 var36;
  orc_int8 var37;
  orc_union16 var38;
  orc_union16 var39;
  orc_int8 var40;
  orc_int8 var41;
  orc_union16 var42;
  orc_union16 var43;
  orc_union16 var44;
  orc_union16 var45;
  orc_union16 var46;
  orc_int8 var47;

  ptr0 = (orc_int8 *) d1;
  ptr4 = (orc_int8 *) s1;
  ptr5 = (orc_int8 *) s2;

  /* 6: loadpw */
  var38.i = p1;
  /* 8: loadpw */
  var39.i = (int) 0x00000080;   /* 128 or 6.32404e-322f */

  for (i = 0; i < n; i++) {
    /* 0: loadb */
    var41 = ptr4[i];
    /* 1: loadb */
    var36 = ptr4[i];
    /* 2: convubw */
    var42.i = (orc_uint8) var36;
    /* 3: loadb */
    var37 = ptr5[i];
    /* 4: convubw */
    var43.i = (orc_uint8) var37;
    /* 5: subw */
    var44.i = var43.i - var42.i;
    /* 7: mullw */
    var45.i = (var44.i * var38.i) & 0xffff;
    /* 9: addw */
    var46.i = var45.i + var39.i;
    /* 10: convhwb */
    var47 = ((orc_uint16) var46.i) >> 8;
    /* 11: addb */
    var40 = var47 + var41;
    /* 12: storeb */
    ptr0[i] = var40;
  }

}

#else
static void
_backup_video_scale_orc_merge_linear_u8 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  const orc_int8 *ORC_RESTRICT ptr5;
  orc_int8 var36;
  orc_int8 var37;
  orc_union16 var38;
  orc_union16 var39;
  orc_int8 var40;
  orc_int8 var41;
  orc_union16 var42;
  orc_union16 var43;
  orc_union16 var44;
  orc_union16 var45;
  orc_union16 var46;
  orc_int8 var47;

  ptr0 = (orc_int8 *) ex->arrays[0];
  ptr4 = (orc_int8 *) ex->arrays[4];
  ptr5 = (orc_int8 *) ex->arrays[5];

  /* 6: loadpw */
  var38.i = ex->params[24];
  /* 8: loadpw */
  var39.i = (int) 0x00000080;   /* 128 or 6.32404e-322f */

  for (i = 0; i < n; i++) {
    /* 0: loadb */
    var41 = ptr4[i];
    /* 1: loadb */
    var36 = ptr4[i];
    /* 2: convubw */
    var42.i = (orc_uint8) var36;
    /* 3: loadb */
    var37 = ptr5[i];
    /* 4: convubw */
    var43.i = (orc_uint8) var37;
    /* 5: subw */
    var44.i = var43.i - var42.i;
    /* 7: mullw */
    var45.i = (var44.i * var38.i) & 0xffff;
    /* 9: addw */
    var46.i = var45.i + var39.i;
    /* 10: convhwb */
    var47 = ((orc_uint16) var46.i) >> 8;
    /* 11: addb */
    var40 = var47 + var41;
    /* 12: storeb */
    ptr0[i] = var40;
  }

}

void
video_scale_orc_merge_linear_u8 (orc_uint8 * ORC_RESTRICT d1,
    const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
    int p1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_merge_linear_u8");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_merge_linear_u8);
      orc_program_add_destination (p, 1, "d1");
      orc_program_add_source (p, 1, "s1");
      orc_program_add_source (p, 1, "s2");
      orc_program_add_constant (p, 4, 0x00000080, "c1");
      orc_program_add_parameter (p, 1, "p1");
      orc_program_add_temporary (p, 2, "t1");
      orc_program_add_temporary (p, 2, "t2");
      orc_program_add_temporary (p, 1, "t3");
      orc_program_add_temporary (p, 1, "t4");

      orc_program_append_2 (p, "loadb", 0, ORC_VAR_T3, ORC_VAR_S1, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convhwb", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "addb", 0, ORC_VAR_D1, ORC_VAR_T4, ORC_VAR_T3,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->arrays[ORC_VAR_S2] = (void *) s2;
  ex->params[ORC_VAR_P1] = p1;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_merge_linear_u16 */
#ifdef DISABLE_ORC
void
video_scale_orc_merge_linear_u16 (orc_uint16 * ORC_RESTRICT d1,
    const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
    int p1, int p2, int n)
{
  int i;
  orc_union16 *ORC_RESTRICT ptr0;
  const orc_union16 *ORC_RESTRICT ptr4;
  const orc_union16 *ORC_RESTRICT ptr5;
  orc_union16 var34;
  orc_union16 var35;
  orc_union16 var36;
  orc_union16 var37;
  orc_union16 var38;
  orc_union32 var39;
  orc_union32 var40;
  orc_union32 var41;
  orc_union32 var42;

  ptr0 = (orc_union16 *) d1;
  ptr4 = (orc_union16 *) s1;
  ptr5 = (orc_union16 *) s2;

  /* 1: loadpw */
  var35.i = p1;
  /* 4: loadpw */
  var37.i = p2;

  for (i = 0; i < n; i++) {
    /* 0: loadw */
    var34 = ptr4[i];
    /* 2: muluwl */
    var39.i = ((orc_uint16) var34.i) * ((orc_uint16) var35.i);
    /* 3: loadw */
    var36 = ptr5[i];
    /* 5: muluwl */
    var40.i = ((orc_uint16) var36.i) * ((orc_uint16) var37.i);
    /* 6: addl */
    var41.i = var39.i + var40.i;
    /* 7: shrul */
    var42.i = ((orc_uint32) var41.i) >> 16;
    /* 8: convlw */
    var38.i = var42.i;
    /* 9: storew */
    ptr0[i] = var38;
  }

}

#else
static void
_backup_video_scale_orc_merge_linear_u16 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union16 *ORC_RESTRICT ptr0;
  const orc_union16 *ORC_RESTRICT ptr4;
  const orc_union16 *ORC_RESTRICT ptr5;
  orc_union16 var34;
  orc_union16 var35;
  orc_union16 var36;
  orc_union16 var37;
  orc_union16 var38;
  orc_union32 var39;
  orc_union32 var40;
  orc_union32 var41;
  orc_union32 var42;

  ptr0 = (orc_union16 *) ex->arrays[0];
  ptr4 = (orc_union16 *) ex->arrays[4];
  ptr5 = (orc_union16 *) ex->arrays[5];

  /* 1: loadpw */
  var35.i = ex->params[24];
  /* 4: loadpw */
  var37.i = ex->params[25];

  for (i = 0; i < n; i++) {
    /* 0: loadw */
    var34 = ptr4[i];
    /* 2: muluwl */
    var39.i = ((orc_uint16) var34.i) * ((orc_uint16) var35.i);
    /* 3: loadw */
    var36 = ptr5[i];
    /* 5: muluwl */
    var40.i = ((orc_uint16) var36.i) * ((orc_uint16) var37.i);
    /* 6: addl */
    var41.i = var39.i + var40.i;
    /* 7: shrul */
    var42.i = ((orc_uint32) var41.i) >> 16;
    /* 8: convlw */
    var38.i = var42.i;
    /* 9: storew */
    ptr0[i] = var38;
  }

}

void
video_scale_orc_merge_linear_u16 (orc_uint16 * ORC_RESTRICT d1,
    const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
    int p1, int p2, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_merge_linear_u16");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_merge_linear_u16);
      orc_program_add_destination (p, 2, "d1");
      orc_program_add_source (p, 2, "s1");
      orc_program_add_source (p, 2, "s2");
      orc_program_add_constant (p, 4, 0x00000010, "c1");
      orc_program_add_parameter (p, 2, "p1");
      orc_program_add_parameter (p, 2, "p2");
      orc_program_add_temporary (p, 4, "t1");
      orc_program_add_temporary (p, 4, "t2");

      orc_program_append_2 (p, "muluwl", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_P1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "muluwl", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_P2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "addl", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "shrul", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convlw", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->arrays[ORC_VAR_S2] = (void *) s2;
  ex->params[ORC_VAR_P1] = p1;
  ex->params[ORC_VAR_P2] = p2;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_splat_u16 */
#ifdef DISABLE_ORC
void
video_scale_orc_splat_u16 (orc_uint16 * ORC_RESTRICT d1, int p1, int n)
{
  int i;
  orc_union16 *ORC_RESTRICT ptr0;
  orc_union16 var32;
  orc_union16 var33;

  ptr0 = (orc_union16 *) d1;

  /* 0: loadpw */
  var32.i = p1;

  for (i = 0; i < n; i++) {
    /* 1: copyw */
    var33.i = var32.i;
    /* 2: storew */
    ptr0[i] = var33;
  }

}

#else
static void
_backup_video_scale_orc_splat_u16 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union16 *ORC_RESTRICT ptr0;
  orc_union16 var32;
  orc_union16 var33;

  ptr0 = (orc_union16 *) ex->arrays[0];

  /* 0: loadpw */
  var32.i = ex->params[24];

  for (i = 0; i < n; i++) {
    /* 1: copyw */
    var33.i = var32.i;
    /* 2: storew */
    ptr0[i] = var33;
  }

}

void
video_scale_orc_splat_u16 (orc_uint16 * ORC_RESTRICT d1, int p1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_splat_u16");
      orc_program_set_backup_function (p, _backup_video_scale_orc_splat_u16);
      orc_program_add_destination (p, 2, "d1");
      orc_program_add_parameter (p, 2, "p1");

      orc_program_append_2 (p, "copyw", 0, ORC_VAR_D1, ORC_VAR_P1, ORC_VAR_D1,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->params[ORC_VAR_P1] = p1;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_splat_u32 */
#ifdef DISABLE_ORC
void
video_scale_orc_splat_u32 (orc_uint32 * ORC_RESTRICT d1, int p1, int n)
{
  int i;
  orc_union32 *ORC_RESTRICT ptr0;
  orc_union32 var32;
  orc_union32 var33;

  ptr0 = (orc_union32 *) d1;

  /* 0: loadpl */
  var32.i = p1;

  for (i = 0; i < n; i++) {
    /* 1: copyl */
    var33.i = var32.i;
    /* 2: storel */
    ptr0[i] = var33;
  }

}

#else
static void
_backup_video_scale_orc_splat_u32 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union32 *ORC_RESTRICT ptr0;
  orc_union32 var32;
  orc_union32 var33;

  ptr0 = (orc_union32 *) ex->arrays[0];

  /* 0: loadpl */
  var32.i = ex->params[24];

  for (i = 0; i < n; i++) {
    /* 1: copyl */
    var33.i = var32.i;
    /* 2: storel */
    ptr0[i] = var33;
  }

}

void
video_scale_orc_splat_u32 (orc_uint32 * ORC_RESTRICT d1, int p1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_splat_u32");
      orc_program_set_backup_function (p, _backup_video_scale_orc_splat_u32);
      orc_program_add_destination (p, 4, "d1");
      orc_program_add_parameter (p, 4, "p1");

      orc_program_append_2 (p, "copyl", 0, ORC_VAR_D1, ORC_VAR_P1, ORC_VAR_D1,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->params[ORC_VAR_P1] = p1;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_splat_u64 */
#ifdef DISABLE_ORC
void
video_scale_orc_splat_u64 (orc_uint64 * ORC_RESTRICT d1, orc_int64 p1, int n)
{
  int i;
  orc_union64 *ORC_RESTRICT ptr0;
  orc_union64 var32;
  orc_union64 var33;

  ptr0 = (orc_union64 *) d1;

  /* 0: loadpq */
  var32.i = p1;

  for (i = 0; i < n; i++) {
    /* 1: copyq */
    var33.i = var32.i;
    /* 2: storeq */
    ptr0[i] = var33;
  }

}

#else
static void
_backup_video_scale_orc_splat_u64 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union64 *ORC_RESTRICT ptr0;
  orc_union64 var32;
  orc_union64 var33;

  ptr0 = (orc_union64 *) ex->arrays[0];

  /* 0: loadpq */
  var32.i =
      (ex->params[24] & 0xffffffff) | ((orc_uint64) (ex->params[24 +
              (ORC_VAR_T1 - ORC_VAR_P1)]) << 32);

  for (i = 0; i < n; i++) {
    /* 1: copyq */
    var33.i = var32.i;
    /* 2: storeq */
    ptr0[i] = var33;
  }

}

void
video_scale_orc_splat_u64 (orc_uint64 * ORC_RESTRICT d1, orc_int64 p1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_splat_u64");
      orc_program_set_backup_function (p, _backup_video_scale_orc_splat_u64);
      orc_program_add_destination (p, 8, "d1");
      orc_program_add_parameter_int64 (p, 8, "p1");

      orc_program_append_2 (p, "copyq", 0, ORC_VAR_D1, ORC_VAR_P1, ORC_VAR_D1,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  {
    orc_union64 tmp;
    tmp.i = p1;
    ex->params[ORC_VAR_P1] = tmp.x2[0];
    ex->params[ORC_VAR_T1] = tmp.x2[1];
  }

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_downsample_u8 */
#ifdef DISABLE_ORC
void
video_scale_orc_downsample_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n)
{
  int i;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_union16 *ORC_RESTRICT ptr4;
  orc_union16 var34;
  orc_int8 var35;
  orc_int8 var36;
  orc_int8 var37;

  ptr0 = (orc_int8 *) d1;
  ptr4 = (orc_union16 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: loadw */
    var34 = ptr4[i];
    /* 1: splitwb */
    {
      orc_union16 _src;
      _src.i = var34.i;
      var36 = _src.x2[1];
      var37 = _src.x2[0];
    }
    /* 2: avgub */
    var35 = ((orc_uint8) var36 + (orc_uint8) var37 + 1) >> 1;
    /* 3: storeb */
    ptr0[i] = var35;
  }

}

#else
static void
_backup_video_scale_orc_downsample_u8 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_union16 *ORC_RESTRICT ptr4;
  orc_union16 var34;
  orc_int8 var35;
  orc_int8 var36;
  orc_int8 var37;

  ptr0 = (orc_int8 *) ex->arrays[0];
  ptr4 = (orc_union16 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: loadw */
    var34 = ptr4[i];
    /* 1: splitwb */
    {
      orc_union16 _src;
      _src.i = var34.i;
      var36 = _src.x2[1];
      var37 = _src.x2[0];
    }
    /* 2: avgub */
    var35 = ((orc_uint8) var36 + (orc_uint8) var37 + 1) >> 1;
    /* 3: storeb */
    ptr0[i] = var35;
  }

}

void
video_scale_orc_downsample_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_downsample_u8");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_downsample_u8);
      orc_program_add_destination (p, 1, "d1");
      orc_program_add_source (p, 2, "s1");
      orc_program_add_temporary (p, 1, "t1");
      orc_program_add_temporary (p, 1, "t2");

      orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_downsample_u16 */
#ifdef DISABLE_ORC
void
video_scale_orc_downsample_u16 (guint16 * ORC_RESTRICT d1,
    const guint16 * ORC_RESTRICT s1, int n)
{
  int i;
  orc_union16 *ORC_RESTRICT ptr0;
  const orc_union32 *ORC_RESTRICT ptr4;
  orc_union32 var34;
  orc_union16 var35;
  orc_union16 var36;
  orc_union16 var37;

  ptr0 = (orc_union16 *) d1;
  ptr4 = (orc_union32 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: loadl */
    var34 = ptr4[i];
    /* 1: splitlw */
    {
      orc_union32 _src;
      _src.i = var34.i;
      var36.i = _src.x2[1];
      var37.i = _src.x2[0];
    }
    /* 2: avguw */
    var35.i = ((orc_uint16) var36.i + (orc_uint16) var37.i + 1) >> 1;
    /* 3: storew */
    ptr0[i] = var35;
  }

}

#else
static void
_backup_video_scale_orc_downsample_u16 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union16 *ORC_RESTRICT ptr0;
  const orc_union32 *ORC_RESTRICT ptr4;
  orc_union32 var34;
  orc_union16 var35;
  orc_union16 var36;
  orc_union16 var37;

  ptr0 = (orc_union16 *) ex->arrays[0];
  ptr4 = (orc_union32 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: loadl */
    var34 = ptr4[i];
    /* 1: splitlw */
    {
      orc_union32 _src;
      _src.i = var34.i;
      var36.i = _src.x2[1];
      var37.i = _src.x2[0];
    }
    /* 2: avguw */
    var35.i = ((orc_uint16) var36.i + (orc_uint16) var37.i + 1) >> 1;
    /* 3: storew */
    ptr0[i] = var35;
  }

}

void
video_scale_orc_downsample_u16 (guint16 * ORC_RESTRICT d1,
    const guint16 * ORC_RESTRICT s1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_downsample_u16");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_downsample_u16);
      orc_program_add_destination (p, 2, "d1");
      orc_program_add_source (p, 4, "s1");
      orc_program_add_temporary (p, 2, "t1");
      orc_program_add_temporary (p, 2, "t2");

      orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "avguw", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_downsample_u32 */
#ifdef DISABLE_ORC
void
video_scale_orc_downsample_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n)
{
  int i;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union64 *ORC_RESTRICT ptr4;
  orc_union64 var34;
  orc_union32 var35;
  orc_union32 var36;
  orc_union32 var37;

  ptr0 = (orc_union32 *) d1;
  ptr4 = (orc_union64 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: loadq */
    var34 = ptr4[i];
    /* 1: splitql */
    {
      orc_union64 _src;
      _src.i = var34.i;
      var36.i = _src.x2[1];
      var37.i = _src.x2[0];
    }
    /* 2: avgub */
    var35.x4[0] = ((orc_uint8) var36.x4[0] + (orc_uint8) var37.x4[0] + 1) >> 1;
    var35.x4[1] = ((orc_uint8) var36.x4[1] + (orc_uint8) var37.x4[1] + 1) >> 1;
    var35.x4[2] = ((orc_uint8) var36.x4[2] + (orc_uint8) var37.x4[2] + 1) >> 1;
    var35.x4[3] = ((orc_uint8) var36.x4[3] + (orc_uint8) var37.x4[3] + 1) >> 1;
    /* 3: storel */
    ptr0[i] = var35;
  }

}

#else
static void
_backup_video_scale_orc_downsample_u32 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union64 *ORC_RESTRICT ptr4;
  orc_union64 var34;
  orc_union32 var35;
  orc_union32 var36;
  orc_union32 var37;

  ptr0 = (orc_union32 *) ex->arrays[0];
  ptr4 = (orc_union64 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: loadq */
    var34 = ptr4[i];
    /* 1: splitql */
    {
      orc_union64 _src;
      _src.i = var34.i;
      var36.i = _src.x2[1];
      var37.i = _src.x2[0];
    }
    /* 2: avgub */
    var35.x4[0] = ((orc_uint8) var36.x4[0] + (orc_uint8) var37.x4[0] + 1) >> 1;
    var35.x4[1] = ((orc_uint8) var36.x4[1] + (orc_uint8) var37.x4[1] + 1) >> 1;
    var35.x4[2] = ((orc_uint8) var36.x4[2] + (orc_uint8) var37.x4[2] + 1) >> 1;
    var35.x4[3] = ((orc_uint8) var36.x4[3] + (orc_uint8) var37.x4[3] + 1) >> 1;
    /* 3: storel */
    ptr0[i] = var35;
  }

}

void
video_scale_orc_downsample_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_downsample_u32");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_downsample_u32);
      orc_program_add_destination (p, 4, "d1");
      orc_program_add_source (p, 8, "s1");
      orc_program_add_temporary (p, 4, "t1");
      orc_program_add_temporary (p, 4, "t2");

      orc_program_append_2 (p, "splitql", 0, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "avgub", 2, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_downsample_yuyv */
#ifdef DISABLE_ORC
void
video_scale_orc_downsample_yuyv (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n)
{
  int i;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union64 *ORC_RESTRICT ptr4;
  orc_union64 var38;
  orc_union32 var39;
  orc_union32 var40;
  orc_union32 var41;
  orc_union16 var42;
  orc_union16 var43;
  orc_union16 var44;
  orc_union16 var45;
  orc_union16 var46;
  orc_union16 var47;

  ptr0 = (orc_union32 *) d1;
  ptr4 = (orc_union64 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: loadq */
    var38 = ptr4[i];
    /* 1: splitwb */
    {
      orc_union16 _src;
      _src.i = var38.x4[0];
      var40.x4[0] = _src.x2[1];
      var41.x4[0] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var38.x4[1];
      var40.x4[1] = _src.x2[1];
      var41.x4[1] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var38.x4[2];
      var40.x4[2] = _src.x2[1];
      var41.x4[2] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var38.x4[3];
      var40.x4[3] = _src.x2[1];
      var41.x4[3] = _src.x2[0];
    }
    /* 2: splitwb */
    {
      orc_union16 _src;
      _src.i = var40.x2[0];
      var42.x2[0] = _src.x2[1];
      var43.x2[0] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var40.x2[1];
      var42.x2[1] = _src.x2[1];
      var43.x2[1] = _src.x2[0];
    }
    /* 3: avgub */
    var44.x2[0] = ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1;
    var44.x2[1] = ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1;
    /* 4: splitlw */
    {
      orc_union32 _src;
      _src.i = var41.i;
      var45.i = _src.x2[1];
      var46.i = _src.x2[0];
    }
    /* 5: avgub */
    var47.x2[0] = ((orc_uint8) var45.x2[0] + (orc_uint8) var46.x2[0] + 1) >> 1;
    var47.x2[1] = ((orc_uint8) var45.x2[1] + (orc_uint8) var46.x2[1] + 1) >> 1;
    /* 6: mergebw */
    {
      orc_union16 _dest;
      _dest.x2[0] = var44.x2[0];
      _dest.x2[1] = var47.x2[0];
      var39.x2[0] = _dest.i;
    }
    {
      orc_union16 _dest;
      _dest.x2[0] = var44.x2[1];
      _dest.x2[1] = var47.x2[1];
      var39.x2[1] = _dest.i;
    }
    /* 7: storel */
    ptr0[i] = var39;
  }

}

#else
static void
_backup_video_scale_orc_downsample_yuyv (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union64 *ORC_RESTRICT ptr4;
  orc_union64 var38;
  orc_union32 var39;
  orc_union32 var40;
  orc_union32 var41;
  orc_union16 var42;
  orc_union16 var43;
  orc_union16 var44;
  orc_union16 var45;
  orc_union16 var46;
  orc_union16 var47;

  ptr0 = (orc_union32 *) ex->arrays[0];
  ptr4 = (orc_union64 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: loadq */
    var38 = ptr4[i];
    /* 1: splitwb */
    {
      orc_union16 _src;
      _src.i = var38.x4[0];
      var40.x4[0] = _src.x2[1];
      var41.x4[0] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var38.x4[1];
      var40.x4[1] = _src.x2[1];
      var41.x4[1] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var38.x4[2];
      var40.x4[2] = _src.x2[1];
      var41.x4[2] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var38.x4[3];
      var40.x4[3] = _src.x2[1];
      var41.x4[3] = _src.x2[0];
    }
    /* 2: splitwb */
    {
      orc_union16 _src;
      _src.i = var40.x2[0];
      var42.x2[0] = _src.x2[1];
      var43.x2[0] = _src.x2[0];
    }
    {
      orc_union16 _src;
      _src.i = var40.x2[1];
      var42.x2[1] = _src.x2[1];
      var43.x2[1] = _src.x2[0];
    }
    /* 3: avgub */
    var44.x2[0] = ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1;
    var44.x2[1] = ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1;
    /* 4: splitlw */
    {
      orc_union32 _src;
      _src.i = var41.i;
      var45.i = _src.x2[1];
      var46.i = _src.x2[0];
    }
    /* 5: avgub */
    var47.x2[0] = ((orc_uint8) var45.x2[0] + (orc_uint8) var46.x2[0] + 1) >> 1;
    var47.x2[1] = ((orc_uint8) var45.x2[1] + (orc_uint8) var46.x2[1] + 1) >> 1;
    /* 6: mergebw */
    {
      orc_union16 _dest;
      _dest.x2[0] = var44.x2[0];
      _dest.x2[1] = var47.x2[0];
      var39.x2[0] = _dest.i;
    }
    {
      orc_union16 _dest;
      _dest.x2[0] = var44.x2[1];
      _dest.x2[1] = var47.x2[1];
      var39.x2[1] = _dest.i;
    }
    /* 7: storel */
    ptr0[i] = var39;
  }

}

void
video_scale_orc_downsample_yuyv (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_downsample_yuyv");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_downsample_yuyv);
      orc_program_add_destination (p, 4, "d1");
      orc_program_add_source (p, 8, "s1");
      orc_program_add_temporary (p, 4, "t1");
      orc_program_add_temporary (p, 4, "t2");
      orc_program_add_temporary (p, 2, "t3");
      orc_program_add_temporary (p, 2, "t4");
      orc_program_add_temporary (p, 2, "t5");
      orc_program_add_temporary (p, 2, "t6");

      orc_program_append_2 (p, "splitwb", 2, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "avgub", 1, ORC_VAR_T5, ORC_VAR_T3, ORC_VAR_T4,
          ORC_VAR_D1);
      orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "avgub", 1, ORC_VAR_T6, ORC_VAR_T3, ORC_VAR_T4,
          ORC_VAR_D1);
      orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T5, ORC_VAR_T6,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_resample_nearest_u8 */
#ifdef DISABLE_ORC
void
video_scale_orc_resample_nearest_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  int i;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  orc_int8 var32;

  ptr0 = (orc_int8 *) d1;
  ptr4 = (orc_int8 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: ldresnearb */
    var32 = ptr4[(p1 + i * p2) >> 16];
    /* 1: storeb */
    ptr0[i] = var32;
  }

}

#else
static void
_backup_video_scale_orc_resample_nearest_u8 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  orc_int8 var32;

  ptr0 = (orc_int8 *) ex->arrays[0];
  ptr4 = (orc_int8 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: ldresnearb */
    var32 = ptr4[(ex->params[24] + i * ex->params[25]) >> 16];
    /* 1: storeb */
    ptr0[i] = var32;
  }

}

void
video_scale_orc_resample_nearest_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_resample_nearest_u8");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_resample_nearest_u8);
      orc_program_add_destination (p, 1, "d1");
      orc_program_add_source (p, 1, "s1");
      orc_program_add_parameter (p, 4, "p1");
      orc_program_add_parameter (p, 4, "p2");

      orc_program_append_2 (p, "ldresnearb", 0, ORC_VAR_D1, ORC_VAR_S1,
          ORC_VAR_P1, ORC_VAR_P2);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->params[ORC_VAR_P1] = p1;
  ex->params[ORC_VAR_P2] = p2;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_resample_bilinear_u8 */
#ifdef DISABLE_ORC
void
video_scale_orc_resample_bilinear_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  int i;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  orc_int8 var32;

  ptr0 = (orc_int8 *) d1;
  ptr4 = (orc_int8 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: ldreslinb */
    {
      int tmp = p1 + i * p2;
      var32 =
          ((orc_uint8) ptr4[tmp >> 16] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) ptr4[(tmp >> 16) + 1] * ((tmp >> 8) & 0xff)) >> 8;
    }
    /* 1: storeb */
    ptr0[i] = var32;
  }

}

#else
static void
_backup_video_scale_orc_resample_bilinear_u8 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  orc_int8 var32;

  ptr0 = (orc_int8 *) ex->arrays[0];
  ptr4 = (orc_int8 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: ldreslinb */
    {
      int tmp = ex->params[24] + i * ex->params[25];
      var32 =
          ((orc_uint8) ptr4[tmp >> 16] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) ptr4[(tmp >> 16) + 1] * ((tmp >> 8) & 0xff)) >> 8;
    }
    /* 1: storeb */
    ptr0[i] = var32;
  }

}

void
video_scale_orc_resample_bilinear_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_resample_bilinear_u8");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_resample_bilinear_u8);
      orc_program_add_destination (p, 1, "d1");
      orc_program_add_source (p, 1, "s1");
      orc_program_add_parameter (p, 4, "p1");
      orc_program_add_parameter (p, 4, "p2");

      orc_program_append_2 (p, "ldreslinb", 0, ORC_VAR_D1, ORC_VAR_S1,
          ORC_VAR_P1, ORC_VAR_P2);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->params[ORC_VAR_P1] = p1;
  ex->params[ORC_VAR_P2] = p2;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_resample_nearest_u32 */
#ifdef DISABLE_ORC
void
video_scale_orc_resample_nearest_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  int i;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union32 *ORC_RESTRICT ptr4;
  orc_union32 var32;

  ptr0 = (orc_union32 *) d1;
  ptr4 = (orc_union32 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: ldresnearl */
    var32 = ptr4[(p1 + i * p2) >> 16];
    /* 1: storel */
    ptr0[i] = var32;
  }

}

#else
static void
_backup_video_scale_orc_resample_nearest_u32 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union32 *ORC_RESTRICT ptr4;
  orc_union32 var32;

  ptr0 = (orc_union32 *) ex->arrays[0];
  ptr4 = (orc_union32 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: ldresnearl */
    var32 = ptr4[(ex->params[24] + i * ex->params[25]) >> 16];
    /* 1: storel */
    ptr0[i] = var32;
  }

}

void
video_scale_orc_resample_nearest_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_resample_nearest_u32");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_resample_nearest_u32);
      orc_program_add_destination (p, 4, "d1");
      orc_program_add_source (p, 4, "s1");
      orc_program_add_parameter (p, 4, "p1");
      orc_program_add_parameter (p, 4, "p2");

      orc_program_append_2 (p, "ldresnearl", 0, ORC_VAR_D1, ORC_VAR_S1,
          ORC_VAR_P1, ORC_VAR_P2);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->params[ORC_VAR_P1] = p1;
  ex->params[ORC_VAR_P2] = p2;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_resample_bilinear_u32 */
#ifdef DISABLE_ORC
void
video_scale_orc_resample_bilinear_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  int i;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union32 *ORC_RESTRICT ptr4;
  orc_union32 var32;

  ptr0 = (orc_union32 *) d1;
  ptr4 = (orc_union32 *) s1;


  for (i = 0; i < n; i++) {
    /* 0: ldreslinl */
    {
      int tmp = p1 + i * p2;
      orc_union32 a = ptr4[tmp >> 16];
      orc_union32 b = ptr4[(tmp >> 16) + 1];
      var32.x4[0] =
          ((orc_uint8) a.x4[0] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[0] * ((tmp >> 8) & 0xff)) >> 8;
      var32.x4[1] =
          ((orc_uint8) a.x4[1] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[1] * ((tmp >> 8) & 0xff)) >> 8;
      var32.x4[2] =
          ((orc_uint8) a.x4[2] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[2] * ((tmp >> 8) & 0xff)) >> 8;
      var32.x4[3] =
          ((orc_uint8) a.x4[3] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[3] * ((tmp >> 8) & 0xff)) >> 8;
    }
    /* 1: storel */
    ptr0[i] = var32;
  }

}

#else
static void
_backup_video_scale_orc_resample_bilinear_u32 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union32 *ORC_RESTRICT ptr0;
  const orc_union32 *ORC_RESTRICT ptr4;
  orc_union32 var32;

  ptr0 = (orc_union32 *) ex->arrays[0];
  ptr4 = (orc_union32 *) ex->arrays[4];


  for (i = 0; i < n; i++) {
    /* 0: ldreslinl */
    {
      int tmp = ex->params[24] + i * ex->params[25];
      orc_union32 a = ptr4[tmp >> 16];
      orc_union32 b = ptr4[(tmp >> 16) + 1];
      var32.x4[0] =
          ((orc_uint8) a.x4[0] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[0] * ((tmp >> 8) & 0xff)) >> 8;
      var32.x4[1] =
          ((orc_uint8) a.x4[1] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[1] * ((tmp >> 8) & 0xff)) >> 8;
      var32.x4[2] =
          ((orc_uint8) a.x4[2] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[2] * ((tmp >> 8) & 0xff)) >> 8;
      var32.x4[3] =
          ((orc_uint8) a.x4[3] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[3] * ((tmp >> 8) & 0xff)) >> 8;
    }
    /* 1: storel */
    ptr0[i] = var32;
  }

}

void
video_scale_orc_resample_bilinear_u32 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, int p1, int p2, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_resample_bilinear_u32");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_resample_bilinear_u32);
      orc_program_add_destination (p, 4, "d1");
      orc_program_add_source (p, 4, "s1");
      orc_program_add_parameter (p, 4, "p1");
      orc_program_add_parameter (p, 4, "p2");

      orc_program_append_2 (p, "ldreslinl", 0, ORC_VAR_D1, ORC_VAR_S1,
          ORC_VAR_P1, ORC_VAR_P2);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->params[ORC_VAR_P1] = p1;
  ex->params[ORC_VAR_P2] = p2;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_resample_merge_bilinear_u32 */
#ifdef DISABLE_ORC
void
video_scale_orc_resample_merge_bilinear_u32 (guint8 * ORC_RESTRICT d1,
    guint8 * ORC_RESTRICT d2, const guint8 * ORC_RESTRICT s1,
    const guint8 * ORC_RESTRICT s2, int p1, int p2, int p3, int n)
{
  int i;
  orc_union32 *ORC_RESTRICT ptr0;
  orc_union32 *ORC_RESTRICT ptr1;
  const orc_union32 *ORC_RESTRICT ptr4;
  const orc_union32 *ORC_RESTRICT ptr5;
  orc_union64 var37;
  orc_union32 var38;
  orc_union32 var39;
  orc_union32 var40;
  orc_union64 var41;
  orc_union64 var42;
  orc_union64 var43;
  orc_union64 var44;
  orc_union32 var45;

  ptr0 = (orc_union32 *) d1;
  ptr1 = (orc_union32 *) d2;
  ptr4 = (orc_union32 *) s1;
  ptr5 = (orc_union32 *) s2;

  /* 6: loadpw */
  var37.x4[0] = p1;
  var37.x4[1] = p1;
  var37.x4[2] = p1;
  var37.x4[3] = p1;

  for (i = 0; i < n; i++) {
    /* 0: ldreslinl */
    {
      int tmp = p2 + i * p3;
      orc_union32 a = ptr5[tmp >> 16];
      orc_union32 b = ptr5[(tmp >> 16) + 1];
      var39.x4[0] =
          ((orc_uint8) a.x4[0] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[0] * ((tmp >> 8) & 0xff)) >> 8;
      var39.x4[1] =
          ((orc_uint8) a.x4[1] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[1] * ((tmp >> 8) & 0xff)) >> 8;
      var39.x4[2] =
          ((orc_uint8) a.x4[2] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[2] * ((tmp >> 8) & 0xff)) >> 8;
      var39.x4[3] =
          ((orc_uint8) a.x4[3] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[3] * ((tmp >> 8) & 0xff)) >> 8;
    }
    /* 1: storel */
    ptr1[i] = var39;
    /* 2: loadl */
    var40 = ptr4[i];
    /* 3: convubw */
    var41.x4[0] = (orc_uint8) var40.x4[0];
    var41.x4[1] = (orc_uint8) var40.x4[1];
    var41.x4[2] = (orc_uint8) var40.x4[2];
    var41.x4[3] = (orc_uint8) var40.x4[3];
    /* 4: convubw */
    var42.x4[0] = (orc_uint8) var39.x4[0];
    var42.x4[1] = (orc_uint8) var39.x4[1];
    var42.x4[2] = (orc_uint8) var39.x4[2];
    var42.x4[3] = (orc_uint8) var39.x4[3];
    /* 5: subw */
    var43.x4[0] = var42.x4[0] - var41.x4[0];
    var43.x4[1] = var42.x4[1] - var41.x4[1];
    var43.x4[2] = var42.x4[2] - var41.x4[2];
    var43.x4[3] = var42.x4[3] - var41.x4[3];
    /* 7: mullw */
    var44.x4[0] = (var43.x4[0] * var37.x4[0]) & 0xffff;
    var44.x4[1] = (var43.x4[1] * var37.x4[1]) & 0xffff;
    var44.x4[2] = (var43.x4[2] * var37.x4[2]) & 0xffff;
    var44.x4[3] = (var43.x4[3] * var37.x4[3]) & 0xffff;
    /* 8: convhwb */
    var45.x4[0] = ((orc_uint16) var44.x4[0]) >> 8;
    var45.x4[1] = ((orc_uint16) var44.x4[1]) >> 8;
    var45.x4[2] = ((orc_uint16) var44.x4[2]) >> 8;
    var45.x4[3] = ((orc_uint16) var44.x4[3]) >> 8;
    /* 9: addb */
    var38.x4[0] = var45.x4[0] + var40.x4[0];
    var38.x4[1] = var45.x4[1] + var40.x4[1];
    var38.x4[2] = var45.x4[2] + var40.x4[2];
    var38.x4[3] = var45.x4[3] + var40.x4[3];
    /* 10: storel */
    ptr0[i] = var38;
  }

}

#else
static void
_backup_video_scale_orc_resample_merge_bilinear_u32 (OrcExecutor *
    ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_union32 *ORC_RESTRICT ptr0;
  orc_union32 *ORC_RESTRICT ptr1;
  const orc_union32 *ORC_RESTRICT ptr4;
  const orc_union32 *ORC_RESTRICT ptr5;
  orc_union64 var37;
  orc_union32 var38;
  orc_union32 var39;
  orc_union32 var40;
  orc_union64 var41;
  orc_union64 var42;
  orc_union64 var43;
  orc_union64 var44;
  orc_union32 var45;

  ptr0 = (orc_union32 *) ex->arrays[0];
  ptr1 = (orc_union32 *) ex->arrays[1];
  ptr4 = (orc_union32 *) ex->arrays[4];
  ptr5 = (orc_union32 *) ex->arrays[5];

  /* 6: loadpw */
  var37.x4[0] = ex->params[24];
  var37.x4[1] = ex->params[24];
  var37.x4[2] = ex->params[24];
  var37.x4[3] = ex->params[24];

  for (i = 0; i < n; i++) {
    /* 0: ldreslinl */
    {
      int tmp = ex->params[25] + i * ex->params[26];
      orc_union32 a = ptr5[tmp >> 16];
      orc_union32 b = ptr5[(tmp >> 16) + 1];
      var39.x4[0] =
          ((orc_uint8) a.x4[0] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[0] * ((tmp >> 8) & 0xff)) >> 8;
      var39.x4[1] =
          ((orc_uint8) a.x4[1] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[1] * ((tmp >> 8) & 0xff)) >> 8;
      var39.x4[2] =
          ((orc_uint8) a.x4[2] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[2] * ((tmp >> 8) & 0xff)) >> 8;
      var39.x4[3] =
          ((orc_uint8) a.x4[3] * (256 - ((tmp >> 8) & 0xff)) +
          (orc_uint8) b.x4[3] * ((tmp >> 8) & 0xff)) >> 8;
    }
    /* 1: storel */
    ptr1[i] = var39;
    /* 2: loadl */
    var40 = ptr4[i];
    /* 3: convubw */
    var41.x4[0] = (orc_uint8) var40.x4[0];
    var41.x4[1] = (orc_uint8) var40.x4[1];
    var41.x4[2] = (orc_uint8) var40.x4[2];
    var41.x4[3] = (orc_uint8) var40.x4[3];
    /* 4: convubw */
    var42.x4[0] = (orc_uint8) var39.x4[0];
    var42.x4[1] = (orc_uint8) var39.x4[1];
    var42.x4[2] = (orc_uint8) var39.x4[2];
    var42.x4[3] = (orc_uint8) var39.x4[3];
    /* 5: subw */
    var43.x4[0] = var42.x4[0] - var41.x4[0];
    var43.x4[1] = var42.x4[1] - var41.x4[1];
    var43.x4[2] = var42.x4[2] - var41.x4[2];
    var43.x4[3] = var42.x4[3] - var41.x4[3];
    /* 7: mullw */
    var44.x4[0] = (var43.x4[0] * var37.x4[0]) & 0xffff;
    var44.x4[1] = (var43.x4[1] * var37.x4[1]) & 0xffff;
    var44.x4[2] = (var43.x4[2] * var37.x4[2]) & 0xffff;
    var44.x4[3] = (var43.x4[3] * var37.x4[3]) & 0xffff;
    /* 8: convhwb */
    var45.x4[0] = ((orc_uint16) var44.x4[0]) >> 8;
    var45.x4[1] = ((orc_uint16) var44.x4[1]) >> 8;
    var45.x4[2] = ((orc_uint16) var44.x4[2]) >> 8;
    var45.x4[3] = ((orc_uint16) var44.x4[3]) >> 8;
    /* 9: addb */
    var38.x4[0] = var45.x4[0] + var40.x4[0];
    var38.x4[1] = var45.x4[1] + var40.x4[1];
    var38.x4[2] = var45.x4[2] + var40.x4[2];
    var38.x4[3] = var45.x4[3] + var40.x4[3];
    /* 10: storel */
    ptr0[i] = var38;
  }

}

void
video_scale_orc_resample_merge_bilinear_u32 (guint8 * ORC_RESTRICT d1,
    guint8 * ORC_RESTRICT d2, const guint8 * ORC_RESTRICT s1,
    const guint8 * ORC_RESTRICT s2, int p1, int p2, int p3, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_resample_merge_bilinear_u32");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_resample_merge_bilinear_u32);
      orc_program_add_destination (p, 4, "d1");
      orc_program_add_destination (p, 4, "d2");
      orc_program_add_source (p, 4, "s1");
      orc_program_add_source (p, 4, "s2");
      orc_program_add_parameter (p, 4, "p1");
      orc_program_add_parameter (p, 4, "p2");
      orc_program_add_parameter (p, 4, "p3");
      orc_program_add_temporary (p, 4, "t1");
      orc_program_add_temporary (p, 4, "t2");
      orc_program_add_temporary (p, 4, "t3");
      orc_program_add_temporary (p, 8, "t4");
      orc_program_add_temporary (p, 8, "t5");

      orc_program_append_2 (p, "ldreslinl", 0, ORC_VAR_T2, ORC_VAR_S2,
          ORC_VAR_P2, ORC_VAR_P3);
      orc_program_append_2 (p, "storel", 0, ORC_VAR_D2, ORC_VAR_T2, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "loadl", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convubw", 2, ORC_VAR_T4, ORC_VAR_T1, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convubw", 2, ORC_VAR_T5, ORC_VAR_T2, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "subw", 2, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_T4,
          ORC_VAR_D1);
      orc_program_append_2 (p, "mullw", 2, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_P1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convhwb", 2, ORC_VAR_T3, ORC_VAR_T5, ORC_VAR_D1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T1,
          ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_D2] = d2;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->arrays[ORC_VAR_S2] = (void *) s2;
  ex->params[ORC_VAR_P1] = p1;
  ex->params[ORC_VAR_P2] = p2;
  ex->params[ORC_VAR_P3] = p3;

  func = p->code_exec;
  func (ex);
}
#endif


/* video_scale_orc_merge_bicubic_u8 */
#ifdef DISABLE_ORC
void
video_scale_orc_merge_bicubic_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
    int p2, int p3, int p4, int n)
{
  int i;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  const orc_int8 *ORC_RESTRICT ptr5;
  const orc_int8 *ORC_RESTRICT ptr6;
  const orc_int8 *ORC_RESTRICT ptr7;
  orc_int8 var34;
  orc_int8 var35;
  orc_int8 var36;
  orc_int8 var37;
  orc_int8 var38;
  orc_int8 var39;
  orc_int8 var40;
  orc_int8 var41;
  orc_union16 var42;
  orc_int8 var43;
  orc_union16 var44;
  orc_union16 var45;
  orc_union16 var46;
  orc_union16 var47;
  orc_union16 var48;
  orc_union16 var49;
  orc_union16 var50;
  orc_union16 var51;
  orc_union16 var52;

  ptr0 = (orc_int8 *) d1;
  ptr4 = (orc_int8 *) s1;
  ptr5 = (orc_int8 *) s2;
  ptr6 = (orc_int8 *) s3;
  ptr7 = (orc_int8 *) s4;

  /* 1: loadpb */
  var35 = p2;
  /* 4: loadpb */
  var37 = p3;
  /* 8: loadpb */
  var39 = p1;
  /* 12: loadpb */
  var41 = p4;
  /* 15: loadpw */
  var42.i = (int) 0x00000020;   /* 32 or 1.58101e-322f */

  for (i = 0; i < n; i++) {
    /* 0: loadb */
    var34 = ptr5[i];
    /* 2: mulubw */
    var44.i = ((orc_uint8) var34) * ((orc_uint8) var35);
    /* 3: loadb */
    var36 = ptr6[i];
    /* 5: mulubw */
    var45.i = ((orc_uint8) var36) * ((orc_uint8) var37);
    /* 6: addw */
    var46.i = var44.i + var45.i;
    /* 7: loadb */
    var38 = ptr4[i];
    /* 9: mulubw */
    var47.i = ((orc_uint8) var38) * ((orc_uint8) var39);
    /* 10: subw */
    var48.i = var46.i - var47.i;
    /* 11: loadb */
    var40 = ptr7[i];
    /* 13: mulubw */
    var49.i = ((orc_uint8) var40) * ((orc_uint8) var41);
    /* 14: subw */
    var50.i = var48.i - var49.i;
    /* 16: addw */
    var51.i = var50.i + var42.i;
    /* 17: shrsw */
    var52.i = var51.i >> 6;
    /* 18: convsuswb */
    var43 = ORC_CLAMP_UB (var52.i);
    /* 19: storeb */
    ptr0[i] = var43;
  }

}

#else
static void
_backup_video_scale_orc_merge_bicubic_u8 (OrcExecutor * ORC_RESTRICT ex)
{
  int i;
  int n = ex->n;
  orc_int8 *ORC_RESTRICT ptr0;
  const orc_int8 *ORC_RESTRICT ptr4;
  const orc_int8 *ORC_RESTRICT ptr5;
  const orc_int8 *ORC_RESTRICT ptr6;
  const orc_int8 *ORC_RESTRICT ptr7;
  orc_int8 var34;
  orc_int8 var35;
  orc_int8 var36;
  orc_int8 var37;
  orc_int8 var38;
  orc_int8 var39;
  orc_int8 var40;
  orc_int8 var41;
  orc_union16 var42;
  orc_int8 var43;
  orc_union16 var44;
  orc_union16 var45;
  orc_union16 var46;
  orc_union16 var47;
  orc_union16 var48;
  orc_union16 var49;
  orc_union16 var50;
  orc_union16 var51;
  orc_union16 var52;

  ptr0 = (orc_int8 *) ex->arrays[0];
  ptr4 = (orc_int8 *) ex->arrays[4];
  ptr5 = (orc_int8 *) ex->arrays[5];
  ptr6 = (orc_int8 *) ex->arrays[6];
  ptr7 = (orc_int8 *) ex->arrays[7];

  /* 1: loadpb */
  var35 = ex->params[25];
  /* 4: loadpb */
  var37 = ex->params[26];
  /* 8: loadpb */
  var39 = ex->params[24];
  /* 12: loadpb */
  var41 = ex->params[27];
  /* 15: loadpw */
  var42.i = (int) 0x00000020;   /* 32 or 1.58101e-322f */

  for (i = 0; i < n; i++) {
    /* 0: loadb */
    var34 = ptr5[i];
    /* 2: mulubw */
    var44.i = ((orc_uint8) var34) * ((orc_uint8) var35);
    /* 3: loadb */
    var36 = ptr6[i];
    /* 5: mulubw */
    var45.i = ((orc_uint8) var36) * ((orc_uint8) var37);
    /* 6: addw */
    var46.i = var44.i + var45.i;
    /* 7: loadb */
    var38 = ptr4[i];
    /* 9: mulubw */
    var47.i = ((orc_uint8) var38) * ((orc_uint8) var39);
    /* 10: subw */
    var48.i = var46.i - var47.i;
    /* 11: loadb */
    var40 = ptr7[i];
    /* 13: mulubw */
    var49.i = ((orc_uint8) var40) * ((orc_uint8) var41);
    /* 14: subw */
    var50.i = var48.i - var49.i;
    /* 16: addw */
    var51.i = var50.i + var42.i;
    /* 17: shrsw */
    var52.i = var51.i >> 6;
    /* 18: convsuswb */
    var43 = ORC_CLAMP_UB (var52.i);
    /* 19: storeb */
    ptr0[i] = var43;
  }

}

void
video_scale_orc_merge_bicubic_u8 (guint8 * ORC_RESTRICT d1,
    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
    int p2, int p3, int p4, int n)
{
  OrcExecutor _ex, *ex = &_ex;
  static volatile int p_inited = 0;
  static OrcProgram *p = 0;
  void (*func) (OrcExecutor *);

  if (!p_inited) {
    orc_once_mutex_lock ();
    if (!p_inited) {

      p = orc_program_new ();
      orc_program_set_name (p, "video_scale_orc_merge_bicubic_u8");
      orc_program_set_backup_function (p,
          _backup_video_scale_orc_merge_bicubic_u8);
      orc_program_add_destination (p, 1, "d1");
      orc_program_add_source (p, 1, "s1");
      orc_program_add_source (p, 1, "s2");
      orc_program_add_source (p, 1, "s3");
      orc_program_add_source (p, 1, "s4");
      orc_program_add_constant (p, 4, 0x00000020, "c1");
      orc_program_add_constant (p, 4, 0x00000006, "c2");
      orc_program_add_parameter (p, 4, "p1");
      orc_program_add_parameter (p, 4, "p2");
      orc_program_add_parameter (p, 4, "p3");
      orc_program_add_parameter (p, 4, "p4");
      orc_program_add_temporary (p, 2, "t1");
      orc_program_add_temporary (p, 2, "t2");

      orc_program_append_2 (p, "mulubw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_P2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "mulubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_P3,
          ORC_VAR_D1);
      orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "mulubw", 0, ORC_VAR_T2, ORC_VAR_S1, ORC_VAR_P1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "mulubw", 0, ORC_VAR_T2, ORC_VAR_S4, ORC_VAR_P4,
          ORC_VAR_D1);
      orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
          ORC_VAR_D1);
      orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C2,
          ORC_VAR_D1);
      orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
          ORC_VAR_D1, ORC_VAR_D1);

      orc_program_compile (p);
    }
    p_inited = TRUE;
    orc_once_mutex_unlock ();
  }
  ex->program = p;

  ex->n = n;
  ex->arrays[ORC_VAR_D1] = d1;
  ex->arrays[ORC_VAR_S1] = (void *) s1;
  ex->arrays[ORC_VAR_S2] = (void *) s2;
  ex->arrays[ORC_VAR_S3] = (void *) s3;
  ex->arrays[ORC_VAR_S4] = (void *) s4;
  ex->params[ORC_VAR_P1] = p1;
  ex->params[ORC_VAR_P2] = p2;
  ex->params[ORC_VAR_P3] = p3;
  ex->params[ORC_VAR_P4] = p4;

  func = p->code_exec;
  func (ex);
}
#endif