diff options
Diffstat (limited to 'src/vppinfra/vector_avx2.h')
| -rw-r--r-- | src/vppinfra/vector_avx2.h | 25 |
1 files changed, 25 insertions, 0 deletions
diff --git a/src/vppinfra/vector_avx2.h b/src/vppinfra/vector_avx2.h index 8cc1d77d63c..0511ec7e4e2 100644 --- a/src/vppinfra/vector_avx2.h +++ b/src/vppinfra/vector_avx2.h @@ -162,6 +162,23 @@ u16x16_byte_swap (u16x16 v) return (u16x16) _mm256_shuffle_epi8 ((__m256i) v, (__m256i) swap); } +static_always_inline u8x32 +u8x32_shuffle (u8x32 v, u8x32 m) +{ + return (u8x32) _mm256_shuffle_epi8 ((__m256i) v, (__m256i) m); +} + +#define u8x32_align_right(a, b, imm) \ + (u8x32) _mm256_alignr_epi8 ((__m256i) a, (__m256i) b, imm) + +static_always_inline u32 +u32x8_sum_elts (u32x8 sum8) +{ + sum8 += (u32x8) u8x32_align_right (sum8, sum8, 8); + sum8 += (u32x8) u8x32_align_right (sum8, sum8, 4); + return sum8[0] + sum8[4]; +} + static_always_inline u32x8 u32x8_hadd (u32x8 v1, u32x8 v2) { @@ -196,6 +213,14 @@ u16x16_mask_last (u16x16 v, u8 n_last) return v & masks[16 - n_last]; } +#ifdef __AVX512F__ +static_always_inline u8x32 +u8x32_mask_load (u8x32 a, void *p, u32 mask) +{ + return (u8x32) _mm256_mask_loadu_epi8 ((__m256i) a, mask, p); +} +#endif + static_always_inline f32x8 f32x8_from_u32x8 (u32x8 v) { |