mirror of
https://github.com/BLAKE3-team/BLAKE3
synced 2024-03-28 10:19:52 +01:00
Write _mm_blend_epi16 emulation without multiplication
Use _mm_and_si128 and _mm_cmpeq_epi16 rather than expensive multiplication _mm_mullo_epi16 with _mm_srai_epi16 that compiler may not be able to optimize.
This commit is contained in:
parent
90e2a924a4
commit
c33a8462d1
|
@ -79,10 +79,11 @@ INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
|
|||
}
|
||||
|
||||
INLINE __m128i blend_epi16(__m128i a, __m128i b, const int imm8) {
|
||||
__m128i mask = _mm_set1_epi16(imm8);
|
||||
mask = _mm_mullo_epi16(mask, set4(0x40008000, 0x10002000, 0x04000800, 0x01000200));
|
||||
mask = _mm_srai_epi16(mask, 15);
|
||||
return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
|
||||
const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
|
||||
__m128i mask = _mm_set1_epi16(imm8);
|
||||
mask = _mm_and_si128(mask, bits);
|
||||
mask = _mm_cmpeq_epi16(mask, bits);
|
||||
return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
|
||||
}
|
||||
|
||||
INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
|
||||
|
|
|
@ -139,9 +139,10 @@ unsafe fn undiagonalize(row0: &mut __m128i, row2: &mut __m128i, row3: &mut __m12
|
|||
|
||||
#[inline(always)]
|
||||
unsafe fn blend_epi16(a: __m128i, b: __m128i, imm8: i32) -> __m128i {
|
||||
let bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
|
||||
let mut mask = _mm_set1_epi16(imm8 as i16);
|
||||
mask = _mm_mullo_epi16(mask, set4(0x40008000, 0x10002000, 0x04000800, 0x01000200));
|
||||
mask = _mm_srai_epi16(mask, 15);
|
||||
mask = _mm_and_si128(mask, bits);
|
||||
mask = _mm_cmpeq_epi16(mask, bits);
|
||||
_mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a))
|
||||
}
|
||||
|
||||
|
|
Loading…
Reference in New Issue