#include "blake3_impl.h"
#include <immintrin.h>

Macros
#define	DEGREE 4

#define	_mm_shuffle_ps2(a, b, c)

Functions
INLINE __m128i	loadu (const uint8_t src[16])

INLINE void	storeu (__m128i src, uint8_t dest[16])

INLINE __m128i	addv (__m128i a, __m128i b)

INLINE __m128i	xorv (__m128i a, __m128i b)

INLINE __m128i	set1 (uint32_t x)

INLINE __m128i	set4 (uint32_t a, uint32_t b, uint32_t c, uint32_t d)

INLINE __m128i	rot16 (__m128i x)

INLINE __m128i	rot12 (__m128i x)

INLINE __m128i	rot8 (__m128i x)

INLINE __m128i	rot7 (__m128i x)

INLINE void	g1 (__m128i row0, __m128i row1, __m128i row2, __m128i row3, __m128i m)

INLINE void	g2 (__m128i row0, __m128i row1, __m128i row2, __m128i row3, __m128i m)

INLINE void	diagonalize (__m128i row0, __m128i row2, __m128i *row3)

INLINE void	undiagonalize (__m128i row0, __m128i row2, __m128i *row3)

INLINE void	compress_pre (__m128i rows[4], const uint32_t cv[8], const uint8_t block[BLAKE3_BLOCK_LEN], uint8_t block_len, uint64_t counter, uint8_t flags)

void	blake3_compress_in_place_sse41 (uint32_t cv[8], const uint8_t block[BLAKE3_BLOCK_LEN], uint8_t block_len, uint64_t counter, uint8_t flags)

void	blake3_compress_xof_sse41 (const uint32_t cv[8], const uint8_t block[BLAKE3_BLOCK_LEN], uint8_t block_len, uint64_t counter, uint8_t flags, uint8_t out[64])

INLINE void	round_fn (__m128i v[16], __m128i m[16], size_t r)

INLINE void	transpose_vecs (__m128i vecs[DEGREE])

INLINE void	transpose_msg_vecs (const uint8_t const inputs, size_t block_offset, __m128i out[16])

INLINE void	load_counters (uint64_t counter, bool increment_counter, __m128i out_lo, __m128i out_hi)

static void	blake3_hash4_sse41 (const uint8_t const inputs, size_t blocks, const uint32_t key[8], uint64_t counter, bool increment_counter, uint8_t flags, uint8_t flags_start, uint8_t flags_end, uint8_t *out)

INLINE void	hash_one_sse41 (const uint8_t *input, size_t blocks, const uint32_t key[8], uint64_t counter, uint8_t flags, uint8_t flags_start, uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN])

void	blake3_hash_many_sse41 (const uint8_t const inputs, size_t num_inputs, size_t blocks, const uint32_t key[8], uint64_t counter, bool increment_counter, uint8_t flags, uint8_t flags_start, uint8_t flags_end, uint8_t *out)

Macro Definition Documentation

◆ _mm_shuffle_ps2

#define _mm_shuffle_ps2	(	a,
		b,
		c
	)

Value:

(_mm_castps_si128( \

_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))

Definition at line 7 of file blake3_sse41.c.

◆ DEGREE

#define DEGREE 4

Definition at line 5 of file blake3_sse41.c.

Function Documentation

◆ addv()

INLINE __m128i addv	(	__m128i	a,
		__m128i	b
	)

Definition at line 19 of file blake3_sse41.c.

Referenced by g1(), g2(), and round_fn().

◆ blake3_compress_in_place_sse41()

void blake3_compress_in_place_sse41	(	uint32_t	cv[8],
		const uint8_t	block[BLAKE3_BLOCK_LEN],
		uint8_t	block_len,
		uint64_t	counter,
		uint8_t	flags
	)

Definition at line 254 of file blake3_sse41.c.

References block, compress_pre(), storeu(), and xorv().

◆ blake3_compress_xof_sse41()

void blake3_compress_xof_sse41	(	const uint32_t	cv[8],
		const uint8_t	block[BLAKE3_BLOCK_LEN],
		uint8_t	block_len,
		uint64_t	counter,
		uint8_t	flags,
		uint8_t	out[64]
	)

Definition at line 264 of file blake3_sse41.c.

References block, compress_pre(), loadu(), storeu(), and xorv().

◆ blake3_hash4_sse41()

static void blake3_hash4_sse41	(	const uint8_t const	inputs,
		size_t	blocks,
		const uint32_t	key[8],
		uint64_t	counter,
		bool	increment_counter,
		uint8_t	flags,
		uint8_t	flags_start,
		uint8_t	flags_end,
		uint8_t *	out
	)

static

Definition at line 454 of file blake3_sse41.c.

References BLAKE3_BLOCK_LEN, block, blocks, IV, load_counters(), round_fn(), set1(), storeu(), transpose_msg_vecs(), transpose_vecs(), and xorv().

Referenced by blake3_hash_many_sse41().

◆ blake3_hash_many_sse41()

void blake3_hash_many_sse41	(	const uint8_t const	inputs,
		size_t	num_inputs,
		size_t	blocks,
		const uint32_t	key[8],
		uint64_t	counter,
		bool	increment_counter,
		uint8_t	flags,
		uint8_t	flags_start,
		uint8_t	flags_end,
		uint8_t *	out
	)

Definition at line 535 of file blake3_sse41.c.

References blake3_hash4_sse41(), BLAKE3_OUT_LEN, blocks, DEGREE, and hash_one_sse41().

◆ compress_pre()

INLINE void compress_pre	(	__m128i	rows[4],
		const uint32_t	cv[8],
		const uint8_t	block[BLAKE3_BLOCK_LEN],
		uint8_t	block_len,
		uint64_t	counter,
		uint8_t	flags
	)

Definition at line 83 of file blake3_sse41.c.

References _mm_shuffle_ps2, block, counter_high(), counter_low(), diagonalize(), g1(), g2(), IV, loadu(), set4(), and undiagonalize().

Referenced by blake3_compress_in_place_sse41(), and blake3_compress_xof_sse41().

◆ diagonalize()

INLINE void diagonalize	(	__m128i *	row0,
		__m128i *	row2,
		__m128i *	row3
	)

Definition at line 71 of file blake3_sse41.c.

Referenced by compress_pre().

◆ g1()

INLINE void g1	(	__m128i *	row0,
		__m128i *	row1,
		__m128i *	row2,
		__m128i *	row3,
		__m128i	m
	)

Definition at line 48 of file blake3_sse41.c.

References addv(), rot12(), rot16(), and xorv().

Referenced by compress_pre().

◆ g2()

INLINE void g2	(	__m128i *	row0,
		__m128i *	row1,
		__m128i *	row2,
		__m128i *	row3,
		__m128i	m
	)

Definition at line 58 of file blake3_sse41.c.

References addv(), rot7(), rot8(), and xorv().

Referenced by compress_pre().

◆ hash_one_sse41()

INLINE void hash_one_sse41	(	const uint8_t *	input,
		size_t	blocks,
		const uint32_t	key[8],
		uint64_t	counter,
		uint8_t	flags,
		uint8_t	flags_start,
		uint8_t	flags_end,
		uint8_t	out[BLAKE3_OUT_LEN]
	)

Definition at line 515 of file blake3_sse41.c.

References BLAKE3_BLOCK_LEN, blake3_compress_in_place_sse41, BLAKE3_KEY_LEN, BLAKE3_OUT_LEN, and blocks.

Referenced by blake3_hash_many_sse41().

◆ load_counters()

INLINE void load_counters	(	uint64_t	counter,
		bool	increment_counter,
		__m128i *	out_lo,
		__m128i *	out_hi
	)

Definition at line 440 of file blake3_sse41.c.

Referenced by blake3_hash4_sse41().

◆ loadu()

INLINE __m128i loadu ( const uint8_t src[16] )

Definition at line 11 of file blake3_sse41.c.

Referenced by blake3_compress_xof_sse41(), compress_pre(), and transpose_msg_vecs().

◆ rot12()

INLINE __m128i rot12 ( __m128i x )

Definition at line 35 of file blake3_sse41.c.

References xorv().

Referenced by g1(), and round_fn().

◆ rot16()

INLINE __m128i rot16 ( __m128i x )

Definition at line 30 of file blake3_sse41.c.

Referenced by g1(), and round_fn().

◆ rot7()

INLINE __m128i rot7 ( __m128i x )

Definition at line 44 of file blake3_sse41.c.

References xorv().

Referenced by g2(), and round_fn().

◆ rot8()

INLINE __m128i rot8 ( __m128i x )

Definition at line 39 of file blake3_sse41.c.

Referenced by g2(), and round_fn().

◆ round_fn()

INLINE void round_fn	(	__m128i	v[16],
		__m128i	m[16],
		size_t	r
	)

Definition at line 276 of file blake3_sse41.c.

References addv(), MSG_SCHEDULE, rot12(), rot16(), rot7(), rot8(), and xorv().

Referenced by blake3_hash4_sse41().

◆ set1()

INLINE __m128i set1 ( uint32_t x )

Definition at line 24 of file blake3_sse41.c.

Referenced by blake3_hash4_sse41().

◆ set4()

INLINE __m128i set4	(	uint32_t	a,
		uint32_t	b,
		uint32_t	c,
		uint32_t	d
	)

Definition at line 26 of file blake3_sse41.c.

Referenced by compress_pre().

◆ storeu()

INLINE void storeu	(	__m128i	src,
		uint8_t	dest[16]
	)

Definition at line 15 of file blake3_sse41.c.

Referenced by blake3_compress_in_place_sse41(), blake3_compress_xof_sse41(), and blake3_hash4_sse41().

◆ transpose_msg_vecs()

INLINE void transpose_msg_vecs	(	const uint8_t const	inputs,
		size_t	block_offset,
		__m128i	out[16]
	)

Definition at line 413 of file blake3_sse41.c.

References loadu(), and transpose_vecs().

Referenced by blake3_hash4_sse41().

◆ transpose_vecs()

INLINE void transpose_vecs ( __m128i vecs[DEGREE] )

Definition at line 392 of file blake3_sse41.c.

Referenced by blake3_hash4_sse41(), and transpose_msg_vecs().

◆ undiagonalize()

INLINE void undiagonalize	(	__m128i *	row0,
		__m128i *	row2,
		__m128i *	row3
	)

Definition at line 77 of file blake3_sse41.c.

Referenced by compress_pre().

◆ xorv()

INLINE __m128i xorv	(	__m128i	a,
		__m128i	b
	)

Definition at line 22 of file blake3_sse41.c.

Referenced by blake3_compress_in_place_sse41(), blake3_compress_xof_sse41(), blake3_hash4_sse41(), g1(), g2(), rot12(), rot7(), and round_fn().

Macros

Functions

Macro Definition Documentation

◆ _mm_shuffle_ps2

◆ DEGREE

Function Documentation

◆ addv()

◆ blake3_compress_in_place_sse41()

◆ blake3_compress_xof_sse41()

◆ blake3_hash4_sse41()

◆ blake3_hash_many_sse41()

◆ compress_pre()

◆ diagonalize()

◆ g1()

◆ g2()

◆ hash_one_sse41()

◆ load_counters()

◆ loadu()

◆ rot12()

◆ rot16()

◆ rot7()

◆ rot8()

◆ round_fn()

◆ set1()

◆ set4()

◆ storeu()

◆ transpose_msg_vecs()

◆ transpose_vecs()

◆ undiagonalize()

◆ xorv()