1 /* adler32_avx2_p.h -- adler32 avx2 utility functions
2 * Copyright (C) 2022 Adam Stylinski
3 * For conditions of distribution and use, see copyright notice in zlib.h
4 */
5
6 #ifndef ADLER32_AVX2_P_H_
7 #define ADLER32_AVX2_P_H_
8
9 #if defined(X86_AVX2_ADLER32) || defined(X86_AVX512VNNI_ADLER32)
10
11 /* 32 bit horizontal sum, adapted from Agner Fog's vector library. */
hsum256(__m256i x)12 static inline uint32_t hsum256(__m256i x) {
13 __m128i sum1 = _mm_add_epi32(_mm256_extracti128_si256(x, 1),
14 _mm256_castsi256_si128(x));
15 __m128i sum2 = _mm_add_epi32(sum1, _mm_unpackhi_epi64(sum1, sum1));
16 __m128i sum3 = _mm_add_epi32(sum2, _mm_shuffle_epi32(sum2, 1));
17 return (uint32_t)_mm_cvtsi128_si32(sum3);
18 }
19
partial_hsum256(__m256i x)20 static inline uint32_t partial_hsum256(__m256i x) {
21 /* We need a permutation vector to extract every other integer. The
22 * rest are going to be zeros */
23 const __m256i perm_vec = _mm256_setr_epi32(0, 2, 4, 6, 1, 1, 1, 1);
24 __m256i non_zero = _mm256_permutevar8x32_epi32(x, perm_vec);
25 __m128i non_zero_sse = _mm256_castsi256_si128(non_zero);
26 __m128i sum2 = _mm_add_epi32(non_zero_sse,_mm_unpackhi_epi64(non_zero_sse, non_zero_sse));
27 __m128i sum3 = _mm_add_epi32(sum2, _mm_shuffle_epi32(sum2, 1));
28 return (uint32_t)_mm_cvtsi128_si32(sum3);
29 }
30 #endif
31
32 #endif
33