From 99473f255b939bf893d653eb9492037c53f7d65f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Antonio=20S=C3=A1nchez?= Date: Fri, 10 Mar 2023 22:36:57 +0000 Subject: [PATCH] Fix failing MSVC tests due to compiler bugs. (cherry picked from commit 394aabb0a3976d95a5c6f286d49e43bb49558cc2) --- Eigen/src/Core/arch/AVX/PacketMath.h | 16 ++++++++++++++-- Eigen/src/Core/arch/AVX512/PacketMath.h | 11 +++++++++-- 2 files changed, 23 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 2906217ba..5be14ba11 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -630,11 +630,23 @@ template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet4d& template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet8f& from, uint8_t umask) { +#ifdef EIGEN_VECTORIZE_AVX512 + __mmask16 mask = static_cast<__mmask16>(umask & 0x00FF); + EIGEN_DEBUG_UNALIGNED_STORE _mm512_mask_storeu_ps(to, mask, _mm512_castps256_ps512(from)); +#else Packet8i mask = _mm256_set1_epi8(static_cast(umask)); - const Packet8i bit_mask = _mm256_set_epi32(0xffffff7f, 0xffffffbf, 0xffffffdf, 0xffffffef, 0xfffffff7, 0xfffffffb, 0xfffffffd, 0xfffffffe); + const Packet8i bit_mask = _mm256_set_epi32(0x7f7f7f7f, 0xbfbfbfbf, 0xdfdfdfdf, 0xefefefef, 0xf7f7f7f7, 0xfbfbfbfb, 0xfdfdfdfd, 0xfefefefe); mask = por(mask, bit_mask); mask = pcmp_eq(mask, _mm256_set1_epi32(0xffffffff)); - EIGEN_DEBUG_UNALIGNED_STORE return _mm256_maskstore_ps(to, mask, from); +#if EIGEN_COMP_MSVC + // MSVC sometimes seems to use a bogus mask with maskstore. + const __m256i ifrom = _mm256_castps_si256(from); + EIGEN_DEBUG_UNALIGNED_STORE _mm_maskmoveu_si128(_mm256_extractf128_si256(ifrom, 0), _mm256_extractf128_si256(mask, 0), reinterpret_cast(to)); + EIGEN_DEBUG_UNALIGNED_STORE _mm_maskmoveu_si128(_mm256_extractf128_si256(ifrom, 1), _mm256_extractf128_si256(mask, 1), reinterpret_cast(to + 4)); +#else + EIGEN_DEBUG_UNALIGNED_STORE _mm256_maskstore_ps(to, mask, from); +#endif +#endif } // NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h index 3110b2df9..4ab100cec 100644 --- a/Eigen/src/Core/arch/AVX512/PacketMath.h +++ b/Eigen/src/Core/arch/AVX512/PacketMath.h @@ -296,12 +296,19 @@ EIGEN_STRONG_INLINE Packet16i psub(const Packet16i& a, template <> EIGEN_STRONG_INLINE Packet16f pnegate(const Packet16f& a) { - const __m512i mask = _mm512_set1_epi32(0x80000000); + // NOTE: MSVC seems to struggle with _mm512_set1_epi32, leading to random results. + // The intel docs give it a relatively high latency as well, so we're probably + // better off with using _mm512_set_epi32 directly anyways. + const __m512i mask = _mm512_set_epi32(0x80000000,0x80000000,0x80000000,0x80000000, + 0x80000000,0x80000000,0x80000000,0x80000000, + 0x80000000,0x80000000,0x80000000,0x80000000, + 0x80000000,0x80000000,0x80000000,0x80000000); return _mm512_castsi512_ps(_mm512_xor_epi32(_mm512_castps_si512(a), mask)); } template <> EIGEN_STRONG_INLINE Packet8d pnegate(const Packet8d& a) { - const __m512i mask = _mm512_set1_epi64(0x8000000000000000ULL); + const __m512i mask = _mm512_set_epi64(0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL, + 0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL); return _mm512_castsi512_pd(_mm512_xor_epi64(_mm512_castpd_si512(a), mask)); }