[patch] patch

chromium-95-eigen-avx.patch (text/plain), 1.99 KB, created by Stephan Hartmann (RETIRED) on 2021-10-03 16:59:49 UTC

From 958a33e67a8dc043c53a8c449bb2763f017ed7da Mon Sep 17 00:00:00 2001
From: Stephan Hartmann <[e-mail]>
Date: Sun, 3 Oct 2021 15:44:23 +0000
Subject: [PATCH] Eigen: fix pmin/pmax for Packet8i if AVX2 is not available

_mm256_min_epi32 is a AVX2 intrinsic. Emulate behavior using AVX
and SSE4 intrinics if AVX2 is not available.
---
 third_party/eigen3/src/Eigen/src/Core/arch/AVX/PacketMath.h | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/third_party/eigen3/src/Eigen/src/Core/arch/AVX/PacketMath.h b/third_party/eigen3/src/Eigen/src/Core/arch/AVX/PacketMath.h
index 926761c..8134878 100644
--- a/third_party/eigen3/src/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/third_party/eigen3/src/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -419,7 +419,13 @@ template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const
 #endif
 }
 template<> EIGEN_STRONG_INLINE Packet8i pmin<Packet8i>(const Packet8i& a, const Packet8i& b) {
+#ifdef EIGEN_VECTORIZE_AVX2
   return _mm256_min_epi32(a, b);
+#else
+  __m128i lo = _mm_min_epi32(_mm256_extractf128_si256(a, 0), _mm256_extractf128_si256(b, 0));
+  __m128i hi = _mm_min_epi32(_mm256_extractf128_si256(a, 1), _mm256_extractf128_si256(b, 1));
+  return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1);
+#endif
 }
 
 template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) {
@@ -451,7 +451,13 @@ template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const
 #endif
 }
 template<> EIGEN_STRONG_INLINE Packet8i pmax<Packet8i>(const Packet8i& a, const Packet8i& b) {
+#ifdef EIGEN_VECTORIZE_AVX2
   return _mm256_max_epi32(a, b);
+#else
+  __m128i lo = _mm_max_epi32(_mm256_extractf128_si256(a, 0), _mm256_extractf128_si256(b, 0));
+  __m128i hi = _mm_max_epi32(_mm256_extractf128_si256(a, 1), _mm256_extractf128_si256(b, 1));
+  return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1);
+#endif
 }
 
 // Add specializations for min/max with prescribed NaN progation.
-- 
2.32.0