Fixes from testing on an avx512f-capable machine

This commit is contained in:
2024-06-29 22:41:39 -07:00
parent fd3ea2c2a8
commit daa8e02d4f
+5 -16
View File
@@ -1786,9 +1786,9 @@ compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
const auto rvVec = _mm_set1_epi32(r);
const auto zero = _mm_setzero_si128();
for (int i = 0; i < 4; ++i) {
compared |=
_mm_movemask_ps(_mm_cmpgt_epi32(_mm_sub_epi32(w[i], rvVec), zero))
<< (i * 4);
compared |= _mm_movemask_ps(
__m128(_mm_cmpgt_epi32(_mm_sub_epi32(w[i], rvVec), zero)))
<< (i * 4);
}
return compared;
}
@@ -1801,7 +1801,7 @@ compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
uint32_t r;
memcpy(&r, &rv, sizeof(r));
return _mm512_cmpgt_epi32_mask(_mm512_sub_epi32(w, _mm512_set1_epi32(r)),
_mm512_setzero());
_mm512_setzero_epi32());
}
#endif
#endif
@@ -1926,18 +1926,7 @@ inline __attribute((always_inline)) bool scan16(const InternalVersionT *vs,
conflict >>= begin << 2;
return !conflict;
#elif INTERNAL_VERSION_32_BIT && defined(HAS_AVX)
__m128i w4[4];
memcpy(w4, vs, sizeof(w4));
uint32_t rv;
memcpy(&rv, &readVersion, sizeof(rv));
const auto rvVec = _mm_set1_epi32(rv);
const auto zero = _mm_setzero_si128();
uint32_t conflict = 0;
for (int i = 0; i < 4; ++i) {
conflict |=
_mm_movemask_ps(_mm_cmpgt_epi32(_mm_sub_epi32(w4[i], rvVec), zero))
<< (i * 4);
}
uint32_t conflict = compare16_32bit(vs, readVersion);
conflict &= (1 << end) - 1;
conflict >>= begin;
return !conflict;