Created the pblend packet primitive and implemented it using SSE and AVX instructions.

This commit is contained in:
Benoit Steiner
2014-06-06 20:18:44 -07:00
parent 79085e08e9
commit 29aebf96e6
5 changed files with 90 additions and 4 deletions

View File

@@ -59,6 +59,7 @@ template<> struct packet_traits<float> : default_packet_traits
HasLog = 0,
HasExp = 0,
HasSqrt = 0
HasBlend = 1,
};
};
template<> struct packet_traits<double> : default_packet_traits
@@ -73,6 +74,7 @@ template<> struct packet_traits<double> : default_packet_traits
HasDiv = 1,
HasExp = 0
HasBlend = 1,
};
};
@@ -557,6 +559,19 @@ ptranspose(PacketBlock<Packet4d,4>& kernel) {
kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49);
}
template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) {
const __m256 zero = _mm256_setzero_ps();
const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
__m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ);
return _mm256_blendv_ps(thenPacket, elsePacket, false_mask);
}
template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) {
const __m256d zero = _mm256_setzero_pd();
const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
__m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ);
return _mm256_blendv_pd(thenPacket, elsePacket, false_mask);
}
} // end namespace internal
} // end namespace Eigen