diff options
| author | Paul Duncan <pabs@pablotron.org> | 2024-05-26 18:51:01 -0400 |
|---|---|---|
| committer | Paul Duncan <pabs@pablotron.org> | 2024-05-26 18:51:01 -0400 |
| commit | 52456fa6875189a942430ae684eae47648da4d50 (patch) | |
| tree | 65c22312e1786b3effc9590c7fd65221d5b2da00 | |
| parent | d710dae94359c5a1c014531162b75981d384f28d (diff) | |
| download | sha3-52456fa6875189a942430ae684eae47648da4d50.tar.bz2 sha3-52456fa6875189a942430ae684eae47648da4d50.zip | |
sha3.c: avx2: s/loadu_epi64/loadu_si256
| -rw-r--r-- | sha3.c | 20 |
1 files changed, 10 insertions, 10 deletions
@@ -536,11 +536,11 @@ static const __m256i LM0 = { ~0, 0, 0, 0 }, // mask, first lane only // load state array to avx2 registers // FIXME: remove macro, not needed #define AVX2_LOAD(s) __m256i \ - r0_lo = _mm256_loadu_epi64(s + 0), /* row 0, cols 0-3 */ \ - r1_lo = _mm256_loadu_epi64(s + 5), /* row 1, cols 0-3 */ \ - r2_lo = _mm256_loadu_epi64(s + 10), /* row 2, cols 0-3 */ \ - r3_lo = _mm256_loadu_epi64(s + 15), /* row 3, cols 0-3 */ \ - r4_lo = _mm256_loadu_epi64(s + 20), /* row 4, cols 0-3 */ \ + r0_lo = _mm256_loadu_si256((__m256i*) (s + 0)), /* row 0, cols 0-3 */ \ + r1_lo = _mm256_loadu_si256((__m256i*) (s + 5)), /* row 1, cols 0-3 */ \ + r2_lo = _mm256_loadu_si256((__m256i*) (s + 10)), /* row 2, cols 0-3 */ \ + r3_lo = _mm256_loadu_si256((__m256i*) (s + 15)), /* row 3, cols 0-3 */ \ + r4_lo = _mm256_loadu_si256((__m256i*) (s + 20)), /* row 4, cols 0-3 */ \ r0_hi = { s[ 4] }, /* row 0, col 4 */ \ r1_hi = { s[ 9] }, /* row 1, col 4 */ \ r2_hi = { s[14] }, /* row 2, col 4 */ \ @@ -552,11 +552,11 @@ static const __m256i LM0 = { ~0, 0, 0, 0 }, // mask, first lane only union { long long int *i64; uint64_t *u64; } p = { .u64 = s }; \ \ /* store rows */ \ - _mm256_storeu_epi64(p.i64 + 0, r0_lo); /* row 0, cols 0-3 */ \ - _mm256_storeu_epi64(p.i64 + 5, r1_lo); /* row 1, cols 0-3 */ \ - _mm256_storeu_epi64(p.i64 + 10, r2_lo); /* row 2, cols 0-3 */ \ - _mm256_storeu_epi64(p.i64 + 15, r3_lo); /* row 3, cols 0-3 */ \ - _mm256_storeu_epi64(p.i64 + 20, r4_lo); /* row 4, cols 0-3 */ \ + _mm256_storeu_si256((__m256i*) (p.i64 + 0), r0_lo); /* row 0, cols 0-3 */ \ + _mm256_storeu_si256((__m256i*) (p.i64 + 5), r1_lo); /* row 1, cols 0-3 */ \ + _mm256_storeu_si256((__m256i*) (p.i64 + 10), r2_lo); /* row 2, cols 0-3 */ \ + _mm256_storeu_si256((__m256i*) (p.i64 + 15), r3_lo); /* row 3, cols 0-3 */ \ + _mm256_storeu_si256((__m256i*) (p.i64 + 20), r4_lo); /* row 4, cols 0-3 */ \ _mm256_maskstore_epi64(p.i64 + 4, LM0, r0_hi); /* row 0, col 4 */ \ _mm256_maskstore_epi64(p.i64 + 9, LM0, r1_hi); /* row 1, col 4 */ \ _mm256_maskstore_epi64(p.i64 + 14, LM0, r2_hi); /* row 2, col 4 */ \ |