vector_sse42.h File Reference

Include dependency graph for vector_sse42.h:

Go to the source code of this file.

Macros
#define	foreach_sse42_vec128i   _(i,8,16,epi8) _(i,16,8,epi16) _(i,32,4,epi32) _(i,64,2,epi64x)
#define	foreach_sse42_vec128u   _(u,8,16,epi8) _(u,16,8,epi16) _(u,32,4,epi32) _(u,64,2,epi64x)
#define	foreach_sse42_vec128f   _(f,32,4,ps) _(f,64,2,pd)
#define	CLIB_VEC128_SPLAT_DEFINED
#define	CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE
#define	u8x16_word_shift_left(a, n)   (u8x16) _mm_slli_si128((__m128i) a, n)
#define	u8x16_word_shift_right(a, n)   (u8x16) _mm_srli_si128((__m128i) a, n)
#define	i8x16_word_shift_left(a, n)   ((i8x16) u8x16_word_shift_left((u8x16) (a), (n)))
#define	i8x16_word_shift_right(a, n)   ((i8x16) u8x16_word_shift_right((u8x16) (a), (n)))
#define	u16x8_word_shift_left(a, n)   ((u16x8) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u16)))
#define	i16x8_word_shift_left(a, n)   ((u16x8) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u16)))
#define	u16x8_word_shift_right(a, n)   ((u16x8) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u16)))
#define	i16x8_word_shift_right(a, n)   ((i16x8) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u16)))
#define	u32x4_word_shift_left(a, n)   ((u32x4) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u32)))
#define	i32x4_word_shift_left(a, n)   ((u32x4) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u32)))
#define	u32x4_word_shift_right(a, n)   ((u32x4) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u32)))
#define	i32x4_word_shift_right(a, n)   ((i32x4) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u32)))
#define	u64x2_word_shift_left(a, n)   ((u64x2) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u64)))
#define	i64x2_word_shift_left(a, n)   ((u64x2) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u64)))
#define	u64x2_word_shift_right(a, n)   ((u64x2) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u64)))
#define	i64x2_word_shift_right(a, n)   ((i64x2) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u64)))
#define	u8x16_align_right(a, b, imm)   (u8x16) _mm_alignr_epi8 ((__m128i) a, (__m128i) b, imm)
#define	CLIB_HAVE_VEC128_MSB_MASK

Functions
static u8x16	u8x16_interleave_hi (u8x16 a, u8x16 b)
static u8x16	u8x16_interleave_lo (u8x16 a, u8x16 b)
static u16x8	u16x8_interleave_hi (u16x8 a, u16x8 b)
static u16x8	u16x8_interleave_lo (u16x8 a, u16x8 b)
static u32x4	u32x4_interleave_hi (u32x4 a, u32x4 b)
static u32x4	u32x4_interleave_lo (u32x4 a, u32x4 b)
static u64x2	u64x2_interleave_hi (u64x2 a, u64x2 b)
static u64x2	u64x2_interleave_lo (u64x2 a, u64x2 b)
adds_epu static subs_epu i16x8	i16x8_mul_lo (i16x8 x, i16x8 y)
static u16x8	u16x8_mul_lo (u16x8 x, u16x8 y)
static i16x8	i16x8_mul_hi (i16x8 x, i16x8 y)
static u16x8	u16x8_mul_hi (u16x8 x, u16x8 y)
static u32	u8x16_max_scalar (u8x16 x)
static u8	u8x16_min_scalar (u8x16 x)
static i16	i16x8_max_scalar (i16x8 x)
static i16	i16x8_min_scalar (i16x8 x)
static_always_inline u32	u32x4_min_scalar (u32x4 v)
static_always_inline u32	u32x4_max_scalar (u32x4 v)
static_always_inline u32	i32x4_min_scalar (i32x4 v)
static_always_inline u32	i32x4_max_scalar (i32x4 v)
static_always_inline u16	u8x16_msb_mask (u8x16 v)
static_always_inline u16	i8x16_msb_mask (i8x16 v)
static_always_inline u32x4	u32x4_byte_swap (u32x4 v)
static_always_inline u16x8	u16x8_byte_swap (u16x8 v)
static_always_inline u8x16	u8x16_reflect (u8x16 v)
static_always_inline u32x4	u32x4_hadd (u32x4 v1, u32x4 v2)
static_always_inline u32 __clib_unused	u32x4_sum_elts (u32x4 sum4)
static_always_inline u8x16	u8x16_shuffle (u8x16 v, u8x16 m)
static_always_inline u32x4	u32x4_shuffle (u32x4 v, const int a, const int b, const int c, const int d)
epu8_epi32 epu16_epi32 epu32_epi64 epi8_epi32 epi16_epi32 epi32_epi64 static_always_inline u64x2	u64x2_gather (void *p0, void *p1)
static_always_inline u32x4	u32x4_gather (void *p0, void *p1, void *p2, void *p3)
static_always_inline void	u64x2_scatter (u64x2 r, void *p0, void *p1)
static_always_inline void	u32x4_scatter (u32x4 r, void *p0, void *p1, void *p2, void *p3)
static_always_inline void	u64x2_scatter_one (u64x2 r, int index, void *p)
static_always_inline void	u32x4_scatter_one (u32x4 r, int index, void *p)
static_always_inline u8x16	u8x16_is_greater (u8x16 v1, u8x16 v2)
static_always_inline u8x16	u8x16_blend (u8x16 v1, u8x16 v2, u8x16 mask)
static_always_inline u8x16	u8x16_xor3 (u8x16 a, u8x16 b, u8x16 c)

Variables
	u8x16
_mm_packus_epi16	u16x8
	add_saturate
adds_epu	sub_saturate
	left
sll	right
	u32x4
epu8_epi32 epu16_epi32	u64x2
epu8_epi32 epu16_epi32 epu32_epi64	i32x4
epu8_epi32 epu16_epi32 epu32_epi64 epi8_epi32 epi16_epi32	i64x2

Macro Definition Documentation

CLIB_HAVE_VEC128_MSB_MASK

#define CLIB_HAVE_VEC128_MSB_MASK

Definition at line 369 of file vector_sse42.h.

CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE

#define CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE

Definition at line 98 of file vector_sse42.h.

CLIB_VEC128_SPLAT_DEFINED

#define CLIB_VEC128_SPLAT_DEFINED

Definition at line 97 of file vector_sse42.h.

foreach_sse42_vec128f

#define foreach_sse42_vec128f   _(f,32,4,ps) _(f,64,2,pd)

Definition at line 49 of file vector_sse42.h.

foreach_sse42_vec128i

#define foreach_sse42_vec128i   _(i,8,16,epi8) _(i,16,8,epi16) _(i,32,4,epi32) _(i,64,2,epi64x)

Definition at line 45 of file vector_sse42.h.

foreach_sse42_vec128u

#define foreach_sse42_vec128u   _(u,8,16,epi8) _(u,16,8,epi16) _(u,32,4,epi32) _(u,64,2,epi64x)

Definition at line 47 of file vector_sse42.h.

i16x8_word_shift_left

#define i16x8_word_shift_left	(		a,
			n
	)		((u16x8) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u16)))

i16x8_word_shift_right

#define i16x8_word_shift_right	(		a,
			n
	)		((i16x8) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u16)))

i32x4_word_shift_left

#define i32x4_word_shift_left	(		a,
			n
	)		((u32x4) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u32)))

i32x4_word_shift_right

#define i32x4_word_shift_right	(		a,
			n
	)		((i32x4) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u32)))

i64x2_word_shift_left

#define i64x2_word_shift_left	(		a,
			n
	)		((u64x2) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u64)))

i64x2_word_shift_right

#define i64x2_word_shift_right	(		a,
			n
	)		((i64x2) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u64)))

i8x16_word_shift_left

#define i8x16_word_shift_left	(		a,
			n
	)		((i8x16) u8x16_word_shift_left((u8x16) (a), (n)))

i8x16_word_shift_right

#define i8x16_word_shift_right	(		a,
			n
	)		((i8x16) u8x16_word_shift_right((u8x16) (a), (n)))

u16x8_word_shift_left

#define u16x8_word_shift_left	(		a,
			n
	)		((u16x8) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u16)))

u16x8_word_shift_right

#define u16x8_word_shift_right	(		a,
			n
	)		((u16x8) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u16)))

u32x4_word_shift_left

#define u32x4_word_shift_left	(		a,
			n
	)		((u32x4) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u32)))

u32x4_word_shift_right

#define u32x4_word_shift_right	(		a,
			n
	)		((u32x4) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u32)))

u64x2_word_shift_left

#define u64x2_word_shift_left	(		a,
			n
	)		((u64x2) u8x16_word_shift_left((u8x16) (a), (n) * sizeof (u64)))

u64x2_word_shift_right

#define u64x2_word_shift_right	(		a,
			n
	)		((u64x2) u8x16_word_shift_right((u8x16) (a), (n) * sizeof (u64)))

u8x16_align_right

#define u8x16_align_right	(		a,
			b,
			imm
	)		(u8x16) _mm_alignr_epi8 ((__m128i) a, (__m128i) b, imm)

Definition at line 322 of file vector_sse42.h.

u8x16_word_shift_left

#define u8x16_word_shift_left	(		a,
			n
	)		(u8x16) _mm_slli_si128((__m128i) a, n)

u8x16_word_shift_right

#define u8x16_word_shift_right	(		a,
			n
	)		(u8x16) _mm_srli_si128((__m128i) a, n)

Function Documentation

i16x8_max_scalar()

static i16 i16x8_max_scalar ( i16x8  x )

inlinestatic

Definition at line 305 of file vector_sse42.h.

i16x8_min_scalar()

static i16 i16x8_min_scalar ( i16x8  x )

inlinestatic

Definition at line 314 of file vector_sse42.h.

i16x8_mul_hi()

static i16x8 i16x8_mul_hi ( i16x8  x, i16x8  y  )

inlinestatic

Definition at line 191 of file vector_sse42.h.

i16x8_mul_lo()

adds_epu static subs_epu i16x8 i16x8_mul_lo ( i16x8  x, i16x8  y  )

inlinestatic

Definition at line 179 of file vector_sse42.h.

i32x4_max_scalar()

static_always_inline u32 i32x4_max_scalar

(

i32x4

v

)

Definition at line 350 of file vector_sse42.h.

i32x4_min_scalar()

static_always_inline u32 i32x4_min_scalar

(

i32x4

v

)

Definition at line 342 of file vector_sse42.h.

i8x16_msb_mask()

static_always_inline u16 i8x16_msb_mask

(

i8x16

v

)

Definition at line 364 of file vector_sse42.h.

Here is the caller graph for this function:

u16x8_byte_swap()

static_always_inline u16x8 u16x8_byte_swap

(

u16x8

v

)

Definition at line 383 of file vector_sse42.h.

u16x8_interleave_hi()

static u16x8 u16x8_interleave_hi ( u16x8  a, u16x8  b  )

inlinestatic

Definition at line 114 of file vector_sse42.h.

u16x8_interleave_lo()

static u16x8 u16x8_interleave_lo ( u16x8  a, u16x8  b  )

inlinestatic

Definition at line 120 of file vector_sse42.h.

u16x8_mul_hi()

static u16x8 u16x8_mul_hi ( u16x8  x, u16x8  y  )

inlinestatic

Definition at line 197 of file vector_sse42.h.

u16x8_mul_lo()

static u16x8 u16x8_mul_lo ( u16x8  x, u16x8  y  )

inlinestatic

Definition at line 185 of file vector_sse42.h.

u32x4_byte_swap()

static_always_inline u32x4 u32x4_byte_swap

(

u32x4

v

)

Definition at line 374 of file vector_sse42.h.

u32x4_gather()

static_always_inline u32x4 u32x4_gather	(	void *	p0,
		void *	p1,
		void *	p2,
		void *	p3
	)

Definition at line 463 of file vector_sse42.h.

u32x4_hadd()

static_always_inline u32x4 u32x4_hadd	(	u32x4	v1,
		u32x4	v2
	)

Definition at line 401 of file vector_sse42.h.

u32x4_interleave_hi()

static u32x4 u32x4_interleave_hi ( u32x4  a, u32x4  b  )

inlinestatic

Definition at line 126 of file vector_sse42.h.

u32x4_interleave_lo()

static u32x4 u32x4_interleave_lo ( u32x4  a, u32x4  b  )

inlinestatic

Definition at line 132 of file vector_sse42.h.

u32x4_max_scalar()

static_always_inline u32 u32x4_max_scalar

(

u32x4

v

)

Definition at line 334 of file vector_sse42.h.

u32x4_min_scalar()

static_always_inline u32 u32x4_min_scalar

(

u32x4

v

)

Definition at line 326 of file vector_sse42.h.

u32x4_scatter()

static_always_inline void u32x4_scatter	(	u32x4	r,
		void *	p0,
		void *	p1,
		void *	p2,
		void *	p3
	)

Definition at line 478 of file vector_sse42.h.

u32x4_scatter_one()

static_always_inline void u32x4_scatter_one	(	u32x4	r,
		int	index,
		void *	p
	)

Definition at line 493 of file vector_sse42.h.

u32x4_shuffle()

static_always_inline u32x4 u32x4_shuffle	(	u32x4	v,
		const int	a,
		const int	b,
		const int	c,
		const int	d
	)

Definition at line 421 of file vector_sse42.h.

Here is the caller graph for this function:

u32x4_sum_elts()

static_always_inline u32 __clib_unused u32x4_sum_elts

(

u32x4

sum4

)

Definition at line 407 of file vector_sse42.h.

u64x2_gather()

epu8_epi32 epu16_epi32 epu32_epi64 epi8_epi32 epi16_epi32 epi32_epi64 static_always_inline u64x2 u64x2_gather	(	void *	p0,
		void *	p1
	)

Definition at line 456 of file vector_sse42.h.

u64x2_interleave_hi()

static u64x2 u64x2_interleave_hi ( u64x2  a, u64x2  b  )

inlinestatic

Definition at line 138 of file vector_sse42.h.

u64x2_interleave_lo()

static u64x2 u64x2_interleave_lo ( u64x2  a, u64x2  b  )

inlinestatic

Definition at line 144 of file vector_sse42.h.

u64x2_scatter()

static_always_inline void u64x2_scatter	(	u64x2	r,
		void *	p0,
		void *	p1
	)

Definition at line 471 of file vector_sse42.h.

u64x2_scatter_one()

static_always_inline void u64x2_scatter_one	(	u64x2	r,
		int	index,
		void *	p
	)

Definition at line 487 of file vector_sse42.h.

u8x16_blend()

static_always_inline u8x16 u8x16_blend	(	u8x16	v1,
		u8x16	v2,
		u8x16	mask
	)

Definition at line 505 of file vector_sse42.h.

Here is the caller graph for this function:

u8x16_interleave_hi()

static u8x16 u8x16_interleave_hi ( u8x16  a, u8x16  b  )

inlinestatic

Definition at line 102 of file vector_sse42.h.

u8x16_interleave_lo()

static u8x16 u8x16_interleave_lo ( u8x16  a, u8x16  b  )

inlinestatic

Definition at line 108 of file vector_sse42.h.

u8x16_is_greater()

static_always_inline u8x16 u8x16_is_greater	(	u8x16	v1,
		u8x16	v2
	)

Definition at line 499 of file vector_sse42.h.

Here is the caller graph for this function:

u8x16_max_scalar()

static u32 u8x16_max_scalar ( u8x16  x )

inlinestatic

Definition at line 285 of file vector_sse42.h.

u8x16_min_scalar()

static u8 u8x16_min_scalar ( u8x16  x )

inlinestatic

Definition at line 295 of file vector_sse42.h.

u8x16_msb_mask()

static_always_inline u16 u8x16_msb_mask

(

u8x16

v

)

Definition at line 358 of file vector_sse42.h.

u8x16_reflect()

static_always_inline u8x16 u8x16_reflect

(

u8x16

v

)

Definition at line 392 of file vector_sse42.h.

u8x16_shuffle()

static_always_inline u8x16 u8x16_shuffle	(	u8x16	v,
		u8x16	m
	)

Definition at line 415 of file vector_sse42.h.

u8x16_xor3()

static_always_inline u8x16 u8x16_xor3	(	u8x16	a,
		u8x16	b,
		u8x16	c
	)

Definition at line 511 of file vector_sse42.h.

Variable Documentation

add_saturate

add_saturate

Definition at line 175 of file vector_sse42.h.

i32x4

epu8_epi32 epu16_epi32 epu32_epi64 epi8_epi32 i32x4

Definition at line 447 of file vector_sse42.h.

i64x2

epu8_epi32 epu16_epi32 epu32_epi64 epi8_epi32 epi16_epi32 i64x2

Definition at line 451 of file vector_sse42.h.

left

sll srl srl sll sra left

Definition at line 212 of file vector_sse42.h.

right

sll srl srl sll sra right

Definition at line 214 of file vector_sse42.h.

sub_saturate

adds_epu sub_saturate

Definition at line 177 of file vector_sse42.h.

u16x8

_mm_packus_epi16 u16x8

Definition at line 159 of file vector_sse42.h.

u32x4

epu8_epi32 u32x4

Definition at line 440 of file vector_sse42.h.

u64x2

epu8_epi32 epu16_epi32 u64x2

Definition at line 444 of file vector_sse42.h.

u8x16

u8x16

Definition at line 157 of file vector_sse42.h.