spe.h

linux下的gcc编译器

static inline float
__ev_get_sfix32_fs_internal (__ev64_opaque__ a, uint32_t pos)
{
  __ev64_fs__ v;

  v = __builtin_spe_evfscfsf ((__v2sf) a);
  return __ev_get_fs_internal ((__ev64_opaque__) v, pos);
}

static inline float
__ev_get_ufix32_fs_internal (__ev64_opaque__ a, uint32_t pos)
{
  __ev64_fs__ v;

  v = __builtin_spe_evfscfuf ((__v2sf) a);
  return __ev_get_fs_internal ((__ev64_opaque__) v, pos);
}

static inline uint16_t
__ev_get_u16_internal (__ev64_opaque__ a, uint32_t pos)
{
  union
  {
    __ev64_opaque__ v;
    uint16_t i[4];
  } u;

  u.v = a;
  return u.i[pos];
}

static inline int16_t
__ev_get_s16_internal (__ev64_opaque__ a, uint32_t pos)
{
  union
  {
    __ev64_opaque__ v;
    int16_t i[4];
  } u;

  u.v = a;
  return u.i[pos];
}

/* __ev_set_* functions.  */

#define __ev_set_u32(a, b, c) __ev_set_u32_internal ((__ev64_opaque__) a, b, c)
#define __ev_set_s32(a, b, c) __ev_set_s32_internal ((__ev64_opaque__) a, b, c)
#define __ev_set_fs(a, b, c) __ev_set_fs_internal ((__ev64_opaque__) a, b, c)
#define __ev_set_u16(a, b, c) __ev_set_u16_internal ((__ev64_opaque__) a, b, c)
#define __ev_set_s16(a, b, c) __ev_set_s16_internal ((__ev64_opaque__) a, b, c)

#define __ev_set_ufix32_u32 __ev_set_u32
#define __ev_set_sfix32_s32 __ev_set_s32

#define __ev_set_sfix32_fs(a, b, c)  __ev_set_sfix32_fs_internal ((__ev64_opaque__) (a), b, c)
#define __ev_set_ufix32_fs(a, b, c)  __ev_set_ufix32_fs_internal ((__ev64_opaque__) (a), b, c)

#define __ev_set_upper_u32(a, b) __ev_set_u32 (a, b, 0)
#define __ev_set_lower_u32(a, b) __ev_set_u32 (a, b, 1)
#define __ev_set_upper_s32(a, b) __ev_set_s32 (a, b, 0)
#define __ev_set_lower_s32(a, b) __ev_set_s32 (a, b, 1)
#define __ev_set_upper_fs(a, b) __ev_set_fs (a, b, 0)
#define __ev_set_lower_fs(a, b) __ev_set_fs (a, b, 1)
#define __ev_set_upper_ufix32_u32 __ev_set_upper_u32
#define __ev_set_lower_ufix32_u32 __ev_set_lower_u32
#define __ev_set_upper_sfix32_s32 __ev_set_upper_s32
#define __ev_set_lower_sfix32_s32 __ev_set_lower_s32
#define __ev_set_upper_sfix32_fs(a, b)  __ev_set_sfix32_fs (a, b, 0)
#define __ev_set_lower_sfix32_fs(a, b)  __ev_set_sfix32_fs (a, b, 1)
#define __ev_set_upper_ufix32_fs(a, b)  __ev_set_ufix32_fs (a, b, 0)
#define __ev_set_lower_ufix32_fs(a, b)  __ev_set_ufix32_fs (a, b, 1)

#define __ev_set_acc_vec64(a) __builtin_spe_evmra ((__ev64_opaque__)(a))

static inline __ev64_opaque__
__ev_set_acc_u64 (uint64_t a)
{
  __ev64_opaque__ ev32;
  ev32 = __ev_create_u64 (a);
  __ev_mra (ev32);
  return ev32;
}

static inline __ev64_opaque__
__ev_set_acc_s64 (int64_t a)
{
  __ev64_opaque__ ev32;
  ev32 = __ev_create_s64 (a);
  __ev_mra (ev32);
  return ev32;
}

static inline __ev64_opaque__
__ev_set_u32_internal (__ev64_opaque__ a, uint32_t b, uint32_t pos)
{
  union
  {
    __ev64_opaque__ v;
    uint32_t i[2];
  } u;

  u.v = a;
  u.i[pos] = b;
  return u.v;
}

static inline __ev64_opaque__
__ev_set_s32_internal (__ev64_opaque__ a, int32_t b, uint32_t pos)
{
  union
  {
    __ev64_opaque__ v;
    int32_t i[2];
  } u;

  u.v = a;
  u.i[pos] = b;
  return u.v;
}

static inline __ev64_opaque__
__ev_set_fs_internal (__ev64_opaque__ a, float b, uint32_t pos)
{
  union
  {
    __ev64_opaque__ v;
    float f[2];
  } u;

  u.v = a;
  u.f[pos] = b;
  return u.v;
}

static inline __ev64_opaque__
__ev_set_sfix32_fs_internal (__ev64_opaque__ a, float b, uint32_t pos)
{
  __ev64_opaque__ v;
  float other;

  /* Get other half.  */
  other = __ev_get_fs_internal (a, pos ^ 1);

  /* Make an sfix32 with 'b'.  */
  v = __ev_create_sfix32_fs (b, b);

  /* Set other half to what it used to be.  */
  return __ev_set_fs_internal (v, other, pos ^ 1);
}

static inline __ev64_opaque__
__ev_set_ufix32_fs_internal (__ev64_opaque__ a, float b, uint32_t pos)
{
  __ev64_opaque__ v;
  float other;

  /* Get other half.  */
  other = __ev_get_fs_internal (a, pos ^ 1);

  /* Make an ufix32 with 'b'.  */
  v = __ev_create_ufix32_fs (b, b);

  /* Set other half to what it used to be.  */
  return __ev_set_fs_internal (v, other, pos ^ 1);
}

static inline __ev64_opaque__
__ev_set_u16_internal (__ev64_opaque__ a, uint16_t b, uint32_t pos)
{
  union
  {
    __ev64_opaque__ v;
    uint16_t i[4];
  } u;

  u.v = a;
  u.i[pos] = b;
  return u.v;
}

static inline __ev64_opaque__
__ev_set_s16_internal (__ev64_opaque__ a, int16_t b, uint32_t pos)
{
  union
  {
    __ev64_opaque__ v;
    int16_t i[4];
  } u;

  u.v = a;
  u.i[pos] = b;
  return u.v;
}

/* Predicates.  */

#define __pred_all	0
#define __pred_any	1
#define __pred_upper	2
#define __pred_lower	3

#define __ev_any_gts(a, b)		__builtin_spe_evcmpgts (__pred_any, (__v2si) (a), (__v2si) (b))
#define __ev_all_gts(a, b)		__builtin_spe_evcmpgts (__pred_all, (__v2si) (a), (__v2si) (b))
#define __ev_upper_gts(a, b)		__builtin_spe_evcmpgts (__pred_upper, (__v2si) (a), (__v2si) (b))
#define __ev_lower_gts(a, b)		__builtin_spe_evcmpgts (__pred_lower, (__v2si) (a), (__v2si) (b))
#define __ev_select_gts(a, b, c, d) 	((__v2si) __builtin_spe_evsel_gts ((__v2si) (a), (__v2si) (b), (__v2si) (c), (__v2si) (d)))

#define __ev_any_gtu(a, b)		__builtin_spe_evcmpgtu (__pred_any, (__v2si) (a), (__v2si) (b))
#define __ev_all_gtu(a, b)		__builtin_spe_evcmpgtu (__pred_all, (__v2si) (a), (__v2si) (b))
#define __ev_upper_gtu(a, b)		__builtin_spe_evcmpgtu (__pred_upper, (__v2si) (a), (__v2si) (b))
#define __ev_lower_gtu(a, b)		__builtin_spe_evcmpgtu (__pred_lower, (__v2si) (a), (__v2si) (b))
#define __ev_select_gtu(a, b, c, d) 	((__v2si) __builtin_spe_evsel_gtu ((__v2si) (a), (__v2si) (b), (__v2si) (c), (__v2si) (d)))

#define __ev_any_lts(a, b)		__builtin_spe_evcmplts (__pred_any, (__v2si) (a), (__v2si) (b))
#define __ev_all_lts(a, b)		__builtin_spe_evcmplts (__pred_all, (__v2si) (a), (__v2si) (b))
#define __ev_upper_lts(a, b)		__builtin_spe_evcmplts (__pred_upper, (__v2si) (a), (__v2si) (b))
#define __ev_lower_lts(a, b)		__builtin_spe_evcmplts (__pred_lower, (__v2si) (a), (__v2si) (b))
#define __ev_select_lts(a, b, c, d) 	((__v2si) __builtin_spe_evsel_lts ((__v2si) (a), (__v2si) (b), (__v2si) (c), (__v2si) (d)))

#define __ev_any_ltu(a, b)		__builtin_spe_evcmpltu (__pred_any, (__v2si) (a), (__v2si) (b))
#define __ev_all_ltu(a, b)		__builtin_spe_evcmpltu (__pred_all, (__v2si) (a), (__v2si) (b))
#define __ev_upper_ltu(a, b)		__builtin_spe_evcmpltu (__pred_upper, (__v2si) (a), (__v2si) (b))
#define __ev_lower_ltu(a, b)		__builtin_spe_evcmpltu (__pred_lower, (__v2si) (a), (__v2si) (b))
#define __ev_select_ltu(a, b, c, d) 	((__v2si) __builtin_spe_evsel_ltu ((__v2si) (a), (__v2si) (b), (__v2si) (c), (__v2si) (d)))
#define __ev_any_eq(a, b)		__builtin_spe_evcmpeq (__pred_any, (__v2si) (a), (__v2si) (b))
#define __ev_all_eq(a, b)		__builtin_spe_evcmpeq (__pred_all, (__v2si) (a), (__v2si) (b))
#define __ev_upper_eq(a, b)		__builtin_spe_evcmpeq (__pred_upper, (__v2si) (a), (__v2si) (b))
#define __ev_lower_eq(a, b)		__builtin_spe_evcmpeq (__pred_lower, (__v2si) (a), (__v2si) (b))
#define __ev_select_eq(a, b, c, d) 	((__v2si) __builtin_spe_evsel_eq ((__v2si) (a), (__v2si) (b), (__v2si) (c), (__v2si) (d)))

#define __ev_any_fs_gt(a, b)		__builtin_spe_evfscmpgt (__pred_any, (__v2sf) (a), (__v2sf) (b))
#define __ev_all_fs_gt(a, b)		__builtin_spe_evfscmpgt (__pred_all, (__v2sf) (a), (__v2sf) (b))
#define __ev_upper_fs_gt(a, b)		__builtin_spe_evfscmpgt (__pred_upper, (__v2sf) (a), (__v2sf) (b))
#define __ev_lower_fs_gt(a, b)		__builtin_spe_evfscmpgt (__pred_lower, (__v2sf) (a), (__v2sf) (b))
#define __ev_select_fs_gt(a, b, c, d)	((__v2si) __builtin_spe_evsel_fsgt ((__v2sf) (a), (__v2sf) (b), (__v2sf) (c), (__v2sf) (d)))

#define __ev_any_fs_lt(a, b)		__builtin_spe_evfscmplt (__pred_any, (__v2sf) (a), (__v2sf) (b))
#define __ev_all_fs_lt(a, b)		__builtin_spe_evfscmplt (__pred_all, (__v2sf) (a), (__v2sf) (b))
#define __ev_upper_fs_lt(a, b)		__builtin_spe_evfscmplt (__pred_upper, (__v2sf) (a), (__v2sf) (b))
#define __ev_lower_fs_lt(a, b)		__builtin_spe_evfscmplt (__pred_lower, (__v2sf) (a), (__v2sf) (b))
#define __ev_select_fs_lt(a, b, c, d)	((__v2si) __builtin_spe_evsel_fslt ((__v2sf) (a), (__v2sf) (b), (__v2sf) (c), (__v2sf) (d)))

#define __ev_any_fs_eq(a, b)		__builtin_spe_evfscmpeq (__pred_any, (__v2sf) (a), (__v2sf) (b))
#define __ev_all_fs_eq(a, b)		__builtin_spe_evfscmpeq (__pred_all, (__v2sf) (a), (__v2sf) (b))
#define __ev_upper_fs_eq(a, b)		__builtin_spe_evfscmpeq (__pred_upper, (__v2sf) (a), (__v2sf) (b))
#define __ev_lower_fs_eq(a, b)		__builtin_spe_evfscmpeq (__pred_lower, (__v2sf) (a), (__v2sf) (b))
#define __ev_select_fs_eq(a, b, c, d)	((__v2si) __builtin_spe_evsel_fseq ((__v2sf) (a), (__v2sf) (b), (__v2sf) (c), (__v2sf) (d)))

#define __ev_any_fs_tst_gt(a, b)	__builtin_spe_evfststgt (__pred_any, (__v2sf) (a), (__v2sf) (b))
#define __ev_all_fs_tst_gt(a, b)	__builtin_spe_evfststgt (__pred_all, (__v2sf) (a), (__v2sf) (b))
#define __ev_upper_fs_tst_gt(a, b)	__builtin_spe_evfststgt (__pred_upper, (__v2sf) (a), (__v2sf) (b))
#define __ev_lower_fs_tst_gt(a, b)	__builtin_spe_evfststgt (__pred_lower, (__v2sf) (a), (__v2sf) (b))
#define __ev_select_fs_tst_gt(a, b, c, d)	((__v2si) __builtin_spe_evsel_fststgt ((__v2sf) (a), (__v2sf) (b), (__v2sf) (c), (__v2sf) (d)))

#define __ev_any_fs_tst_lt(a, b)	__builtin_spe_evfststlt (__pred_any, (__v2sf) (a), (__v2sf) (b))
#define __ev_all_fs_tst_lt(a, b)	__builtin_spe_evfststlt (__pred_all, (__v2sf) (a), (__v2sf) (b))
#define __ev_upper_fs_tst_lt(a, b)	__builtin_spe_evfststlt (__pred_upper, (__v2sf) (a), (__v2sf) (b))
#define __ev_lower_fs_tst_lt(a, b)	__builtin_spe_evfststlt (__pred_lower, (__v2sf) (a), (__v2sf) (b))
#define __ev_select_fs_tst_lt(a, b, c, d)	((__v2si) __builtin_spe_evsel_fststlt ((__v2sf) (a), (__v2sf) (b), (__v2sf) (c), (__v2sf) (d)))

#define __ev_any_fs_tst_eq(a, b)	__builtin_spe_evfststeq (__pred_any, (__v2sf) (a), (__v2sf) (b))
#define __ev_all_fs_tst_eq(a, b)	__builtin_spe_evfststeq (__pred_all, (__v2sf) (a), (__v2sf) (b))
#define __ev_upper_fs_tst_eq(a, b)	__builtin_spe_evfststeq (__pred_upper, (__v2sf) (a), (__v2sf) (b))
#define __ev_lower_fs_tst_eq(a, b)	__builtin_spe_evfststeq (__pred_lower, (__v2sf) (a), (__v2sf) (b))
#define __ev_select_fs_tst_eq(a, b, c, d)	((__v2si) __builtin_spe_evsel_fststeq ((__v2sf) (a), (__v2sf) (b), (__v2sf) (c), (__v2sf) (d)))

/* SPEFSCR accesor functions.  */

#define __SPEFSCR_SOVH		0x80000000
#define __SPEFSCR_OVH		0x40000000
#define __SPEFSCR_FGH		0x20000000
#define __SPEFSCR_FXH		0x10000000
#define __SPEFSCR_FINVH		0x08000000
#define __SPEFSCR_FDBZH		0x04000000
#define __SPEFSCR_FUNFH		0x02000000
#define __SPEFSCR_FOVFH		0x01000000
/* 2 unused bits.  */
#define __SPEFSCR_FINXS		0x00200000
#define __SPEFSCR_FINVS		0x00100000
#define __SPEFSCR_FDBZS		0x00080000
#define __SPEFSCR_FUNFS		0x00040000
#define __SPEFSCR_FOVFS		0x00020000
#define __SPEFSCR_MODE		0x00010000
#define __SPEFSCR_SOV		0x00008000
#define __SPEFSCR_OV		0x00004000
#define __SPEFSCR_FG		0x00002000
#define __SPEFSCR_FX		0x00001000
#define __SPEFSCR_FINV		0x00000800
#define __SPEFSCR_FDBZ		0x00000400
#define __SPEFSCR_FUNF		0x00000200
#define __SPEFSCR_FOVF		0x00000100
/* 1 unused bit.  */
#define __SPEFSCR_FINXE		0x00000040
#define __SPEFSCR_FINVE		0x00000020
#define __SPEFSCR_FDBZE		0x00000010
#define __SPEFSCR_FUNFE		0x00000008
#define __SPEFSCR_FOVFE		0x00000004
#define __SPEFSCR_FRMC		0x00000003

#define __ev_get_spefscr_sovh() (__builtin_spe_mfspefscr () & __SPEFSCR_SOVH)
#define __ev_get_spefscr_ovh() (__builtin_spe_mfspefscr () & __SPEFSCR_OVH)
#define __ev_get_spefscr_fgh() (__builtin_spe_mfspefscr () & __SPEFSCR_FGH)
#define __ev_get_spefscr_fxh() (__builtin_spe_mfspefscr () & __SPEFSCR_FXH)
#define __ev_get_spefscr_finvh() (__builtin_spe_mfspefscr () & __SPEFSCR_FINVH)
#define __ev_get_spefscr_fdbzh() (__builtin_spe_mfspefscr () & __SPEFSCR_FDBZH)
#define __ev_get_spefscr_funfh() (__builtin_spe_mfspefscr () & __SPEFSCR_FUNFH)
#define __ev_get_spefscr_fovfh() (__builtin_spe_mfspefscr () & __SPEFSCR_FOVFH)
#define __ev_get_spefscr_finxs() (__builtin_spe_mfspefscr () & __SPEFSCR_FINXS)
#define __ev_get_spefscr_finvs() (__builtin_spe_mfspefscr () & __SPEFSCR_FINVS)
#define __ev_get_spefscr_fdbzs() (__builtin_spe_mfspefscr () & __SPEFSCR_FDBZS)
#define __ev_get_spefscr_funfs() (__builtin_spe_mfspefscr () & __SPEFSCR_FUNFS)
#define __ev_get_spefscr_fovfs() (__builtin_spe_mfspefscr () & __SPEFSCR_FOVFS)
#define __ev_get_spefscr_mode() (__builtin_spe_mfspefscr () & __SPEFSCR_MODE)
#define __ev_get_spefscr_sov() (__builtin_spe_mfspefscr () & __SPEFSCR_SOV)
#define __ev_get_spefscr_ov() (__builtin_spe_mfspefscr () & __SPEFSCR_OV)
#define __ev_get_spefscr_fg() (__builtin_spe_mfspefscr () & __SPEFSCR_FG)
#define __ev_get_spefscr_fx() (__builtin_spe_mfspefscr () & __SPEFSCR_FX)
#define __ev_get_spefscr_finv() (__builtin_spe_mfspefscr () & __SPEFSCR_FINV)
#define __ev_get_spefscr_fdbz() (__builtin_spe_mfspefscr () & __SPEFSCR_FDBZ)
#define __ev_get_spefscr_funf() (__builtin_spe_mfspefscr () & __SPEFSCR_FUNF)
#define __ev_get_spefscr_fovf() (__builtin_spe_mfspefscr () & __SPEFSCR_FOVF)
#define __ev_get_spefscr_finxe() (__builtin_spe_mfspefscr () & __SPEFSCR_FINXE)
#define __ev_get_spefscr_finve() (__builtin_spe_mfspefscr () & __SPEFSCR_FINVE)
#define __ev_get_spefscr_fdbze() (__builtin_spe_mfspefscr () & __SPEFSCR_FDBZE)
#define __ev_get_spefscr_funfe() (__builtin_spe_mfspefscr () & __SPEFSCR_FUNFE)
#define __ev_get_spefscr_fovfe() (__builtin_spe_mfspefscr () & __SPEFSCR_FOVFE)
#define __ev_get_spefscr_frmc() (__builtin_spe_mfspefscr () & __SPEFSCR_FRMC)

static inline void
__ev_clr_spefscr_field (int mask)
{
  int i;

  i = __builtin_spe_mfspefscr ();
  i &= ~mask;
  __builtin_spe_mtspefscr (i);
}

#define __ev_clr_spefscr_sovh() __ev_clr_spefscr_field (__SPEFSCR_SOVH)
#define __ev_clr_spefscr_sov() __ev_clr_spefscr_field (__SPEFSCR_SOV)
#define __ev_clr_spefscr_finxs() __ev_clr_spefscr_field (__SPEFSCR_FINXS)
#define __ev_clr_spefscr_finvs() __ev_clr_spefscr_field (__SPEFSCR_FINVS)
#define __ev_clr_spefscr_fdbzs() __ev_clr_spefscr_field (__SPEFSCR_FDBZS)
#define __ev_clr_spefscr_funfs() __ev_clr_spefscr_field (__SPEFSCR_FUNFS)
#define __ev_clr_spefscr_fovfs() __ev_clr_spefscr_field (__SPEFSCR_FOVFS)

/* Set rounding mode:
     rnd = 0 (nearest)
     rnd = 1 (zero)
     rnd = 2 (+inf)
     rnd = 3 (-inf).  */

static inline void
__ev_set_spefscr_frmc (int rnd)
{
  int i;

  i = __builtin_spe_mfspefscr ();
  i &= ~__SPEFSCR_FRMC;
  i |= rnd;
}

#endif /* _SPE_H */