vector_avx2.h File Reference

Include dependency graph for vector_avx2.h:

Go to the source code of this file.

Macros
#define	foreach_avx2_vec256i   _(i,8,32,epi8) _(i,16,16,epi16) _(i,32,8,epi32) _(i,64,4,epi64)
#define	foreach_avx2_vec256u   _(u,8,32,epi8) _(u,16,16,epi16) _(u,32,8,epi32) _(u,64,4,epi64)
#define	foreach_avx2_vec256f   _(f,32,8,ps) _(f,64,4,pd)
#define	u64x4_permute(v, m0, m1, m2, m3)
#define	u8x32_align_right(a, b, imm)   (u8x32) _mm256_alignr_epi8 ((__m256i) a, (__m256i) b, imm)
#define	u32x8_blend(a, b, m)   (u32x8) _mm256_blend_epi32 ((__m256i) a, (__m256i) b, m)
#define	u16x16_blend(v1, v2, mask)   (u16x16) _mm256_blend_epi16 ((__m256i) (v1), (__m256i) (v2), mask)
#define	u32x8_permute_lanes(a, b, m)   (u32x8) _mm256_permute2x128_si256 ((__m256i) a, (__m256i) b, m)
#define	u64x4_permute_lanes(a, b, m)   (u64x4) _mm256_permute2x128_si256 ((__m256i) a, (__m256i) b, m)

Functions
foreach_avx2_vec256i static foreach_avx2_vec256u u32x8	u32x8_permute (u32x8 v, u32x8 idx)
_mm256_packus_epi16 _mm256_packus_epi32 static_always_inline u32	u8x32_msb_mask (u8x32 v)
static_always_inline u32	i8x32_msb_mask (i8x32 v)
epu16_epi64 epu8_epi16 epu8_epi64 epi16_epi64 epi8_epi16 epi8_epi64 static_always_inline u64x4	u64x4_byte_swap (u64x4 v)
static_always_inline u32x8	u32x8_byte_swap (u32x8 v)
static_always_inline u16x16	u16x16_byte_swap (u16x16 v)
static_always_inline u8x32	u8x32_shuffle (u8x32 v, u8x32 m)
static_always_inline u32	u32x8_sum_elts (u32x8 sum8)
static_always_inline u32x8	u32x8_hadd (u32x8 v1, u32x8 v2)
static_always_inline u16x16	u16x16_mask_last (u16x16 v, u8 n_last)
static_always_inline f32x8	f32x8_from_u32x8 (u32x8 v)
static_always_inline u32x8	u32x8_from_f32x8 (f32x8 v)
static_always_inline u64x4	u64x4_gather (void *p0, void *p1, void *p2, void *p3)
static_always_inline u32x8	u32x8_gather (void *p0, void *p1, void *p2, void *p3, void *p4, void *p5, void *p6, void *p7)
static_always_inline void	u64x4_scatter (u64x4 r, void *p0, void *p1, void *p2, void *p3)
static_always_inline void	u32x8_scatter (u32x8 r, void *p0, void *p1, void *p2, void *p3, void *p4, void *p5, void *p6, void *p7)
static_always_inline void	u64x4_scatter_one (u64x4 r, int index, void *p)
static_always_inline void	u32x8_scatter_one (u32x8 r, int index, void *p)
static_always_inline u8x32	u8x32_is_greater (u8x32 v1, u8x32 v2)
static_always_inline u8x32	u8x32_blend (u8x32 v1, u8x32 v2, u8x32 mask)
static_always_inline u32x8	u32x8_min (u32x8 a, u32x8 b)
static_always_inline u32	u32x8_min_scalar (u32x8 v)
static_always_inline void	u32x8_transpose (u32x8 a[8])
static_always_inline void	u64x4_transpose (u64x4 a[8])

Variables
	u8x32
_mm256_packus_epi16	u16x16
	u64x4
epu16_epi64 epu8_epi16 epu8_epi64	i64x4
epu16_epi64 epu8_epi16 epu8_epi64 epi16_epi64	i16x16

Macro Definition Documentation

foreach_avx2_vec256f

#define foreach_avx2_vec256f   _(f,32,8,ps) _(f,64,4,pd)

Definition at line 27 of file vector_avx2.h.

foreach_avx2_vec256i

#define foreach_avx2_vec256i   _(i,8,32,epi8) _(i,16,16,epi16) _(i,32,8,epi32) _(i,64,4,epi64)

Definition at line 23 of file vector_avx2.h.

foreach_avx2_vec256u

#define foreach_avx2_vec256u   _(u,8,32,epi8) _(u,16,16,epi16) _(u,32,8,epi32) _(u,64,4,epi64)

Definition at line 25 of file vector_avx2.h.

u16x16_blend

#define u16x16_blend	(		v1,
			v2,
			mask
	)		(u16x16) _mm256_blend_epi16 ((__m256i) (v1), (__m256i) (v2), mask)

Definition at line 252 of file vector_avx2.h.

u32x8_blend

#define u32x8_blend	(		a,
			b,
			m
	)		(u32x8) _mm256_blend_epi32 ((__m256i) a, (__m256i) b, m)

Definition at line 249 of file vector_avx2.h.

u32x8_permute_lanes

#define u32x8_permute_lanes	(		a,
			b,
			m
	)		(u32x8) _mm256_permute2x128_si256 ((__m256i) a, (__m256i) b, m)

Definition at line 324 of file vector_avx2.h.

u64x4_permute

#define u64x4_permute	(		v,
			m0,
			m1,
			m2,
			m3
	)

Value:

  (u64x4) _mm256_permute4x64_epi64 (                                          \
    (__m256i) v, ((m0) | (m1) << 2 | (m2) << 4 | (m3) << 6))

Definition at line 78 of file vector_avx2.h.

u64x4_permute_lanes

#define u64x4_permute_lanes	(		a,
			b,
			m
	)		(u64x4) _mm256_permute2x128_si256 ((__m256i) a, (__m256i) b, m)

Definition at line 326 of file vector_avx2.h.

u8x32_align_right

#define u8x32_align_right	(		a,
			b,
			imm
	)		(u8x32) _mm256_alignr_epi8 ((__m256i) a, (__m256i) b, imm)

Definition at line 192 of file vector_avx2.h.

Function Documentation

f32x8_from_u32x8()

static_always_inline f32x8 f32x8_from_u32x8

(

u32x8

v

)

Definition at line 238 of file vector_avx2.h.

Here is the caller graph for this function:

i8x32_msb_mask()

static_always_inline u32 i8x32_msb_mask

(

i8x32

v

)

Definition at line 129 of file vector_avx2.h.

Here is the caller graph for this function:

u16x16_byte_swap()

static_always_inline u16x16 u16x16_byte_swap

(

u16x16

v

)

Definition at line 177 of file vector_avx2.h.

Here is the caller graph for this function:

u16x16_mask_last()

static_always_inline u16x16 u16x16_mask_last	(	u16x16	v,
		u8	n_last
	)

Definition at line 210 of file vector_avx2.h.

Here is the caller graph for this function:

u32x8_byte_swap()

static_always_inline u32x8 u32x8_byte_swap

(

u32x8

v

)

Definition at line 167 of file vector_avx2.h.

Here is the caller graph for this function:

u32x8_from_f32x8()

static_always_inline u32x8 u32x8_from_f32x8

(

f32x8

v

)

Definition at line 244 of file vector_avx2.h.

Here is the caller graph for this function:

u32x8_gather()

static_always_inline u32x8 u32x8_gather	(	void *	p0,
		void *	p1,
		void *	p2,
		void *	p3,
		void *	p4,
		void *	p5,
		void *	p6,
		void *	p7
	)

Definition at line 265 of file vector_avx2.h.

u32x8_hadd()

static_always_inline u32x8 u32x8_hadd	(	u32x8	v1,
		u32x8	v2
	)

Definition at line 204 of file vector_avx2.h.

Here is the caller graph for this function:

u32x8_min()

static_always_inline u32x8 u32x8_min	(	u32x8	a,
		u32x8	b
	)

Definition at line 330 of file vector_avx2.h.

Here is the caller graph for this function:

u32x8_min_scalar()

static_always_inline u32 u32x8_min_scalar

(

u32x8

v

)

Definition at line 336 of file vector_avx2.h.

Here is the call graph for this function:

Here is the caller graph for this function:

u32x8_permute()

foreach_avx2_vec256i static foreach_avx2_vec256u u32x8 u32x8_permute ( u32x8  v, u32x8  idx  )

inlinestatic

Definition at line 73 of file vector_avx2.h.

Here is the caller graph for this function:

u32x8_scatter()

static_always_inline void u32x8_scatter	(	u32x8	r,
		void *	p0,
		void *	p1,
		void *	p2,
		void *	p3,
		void *	p4,
		void *	p5,
		void *	p6,
		void *	p7
	)

Definition at line 286 of file vector_avx2.h.

u32x8_scatter_one()

static_always_inline void u32x8_scatter_one	(	u32x8	r,
		int	index,
		void *	p
	)

Definition at line 306 of file vector_avx2.h.

Here is the caller graph for this function:

u32x8_sum_elts()

static_always_inline u32 u32x8_sum_elts

(

u32x8

sum8

)

Definition at line 196 of file vector_avx2.h.

u32x8_transpose()

static_always_inline void u32x8_transpose

(

u32x8

a[8]

)

Definition at line 343 of file vector_avx2.h.

u64x4_byte_swap()

epu16_epi64 epu8_epi16 epu8_epi64 epi16_epi64 epi8_epi16 epi8_epi64 static_always_inline u64x4 u64x4_byte_swap

(

u64x4

v

)

Definition at line 157 of file vector_avx2.h.

u64x4_gather()

static_always_inline u64x4 u64x4_gather	(	void *	p0,
		void *	p1,
		void *	p2,
		void *	p3
	)

Definition at line 256 of file vector_avx2.h.

Here is the caller graph for this function:

u64x4_scatter()

static_always_inline void u64x4_scatter	(	u64x4	r,
		void *	p0,
		void *	p1,
		void *	p2,
		void *	p3
	)

Definition at line 277 of file vector_avx2.h.

Here is the caller graph for this function:

u64x4_scatter_one()

static_always_inline void u64x4_scatter_one	(	u64x4	r,
		int	index,
		void *	p
	)

Definition at line 300 of file vector_avx2.h.

u64x4_transpose()

static_always_inline void u64x4_transpose

(

u64x4

a[8]

)

Definition at line 378 of file vector_avx2.h.

u8x32_blend()

static_always_inline u8x32 u8x32_blend	(	u8x32	v1,
		u8x32	v2,
		u8x32	mask
	)

Definition at line 318 of file vector_avx2.h.

Here is the caller graph for this function:

u8x32_is_greater()

static_always_inline u8x32 u8x32_is_greater	(	u8x32	v1,
		u8x32	v2
	)

Definition at line 312 of file vector_avx2.h.

Here is the caller graph for this function:

u8x32_msb_mask()

_mm256_packus_epi16 _mm256_packus_epi32 static_always_inline u32 u8x32_msb_mask

(

u8x32

v

)

Definition at line 123 of file vector_avx2.h.

Here is the caller graph for this function:

u8x32_shuffle()

static_always_inline u8x32 u8x32_shuffle	(	u8x32	v,
		u8x32	m
	)

Definition at line 187 of file vector_avx2.h.

Variable Documentation

i16x16

epu16_epi64 epu8_epi16 epu8_epi64 epi16_epi64 i16x16

Definition at line 150 of file vector_avx2.h.

i64x4

epu16_epi64 epu8_epi16 epu8_epi64 epi16_epi64 epi8_epi16 i64x4

Definition at line 148 of file vector_avx2.h.

u16x16

epu16_epi64 u16x16

Definition at line 118 of file vector_avx2.h.

u64x4

epu16_epi64 epu8_epi16 u64x4

Definition at line 142 of file vector_avx2.h.

u8x32

u8x32

Definition at line 116 of file vector_avx2.h.