mirror of
https://github.com/NGSolve/netgen.git
synced 2025-04-12 16:17:29 +05:00
less special HSum
This commit is contained in:
parent
3e52c44aeb
commit
32f1b6177e
@ -36,6 +36,7 @@
|
|||||||
|
|
||||||
namespace ngcore
|
namespace ngcore
|
||||||
{
|
{
|
||||||
|
/*
|
||||||
#ifndef __CUDA_ARCH__
|
#ifndef __CUDA_ARCH__
|
||||||
#ifdef NETGEN_ARCH_AMD64
|
#ifdef NETGEN_ARCH_AMD64
|
||||||
NETGEN_INLINE auto HSum (SIMD<double,2> v1, SIMD<double,2> v2, SIMD<double,2> v3, SIMD<double,2> v4)
|
NETGEN_INLINE auto HSum (SIMD<double,2> v1, SIMD<double,2> v2, SIMD<double,2> v3, SIMD<double,2> v4)
|
||||||
@ -51,7 +52,8 @@ namespace ngcore
|
|||||||
}
|
}
|
||||||
#endif
|
#endif
|
||||||
#endif // __CUDA_ARCH__
|
#endif // __CUDA_ARCH__
|
||||||
|
*/
|
||||||
|
|
||||||
NETGEN_INLINE void SIMDTranspose (SIMD<double,4> a1, SIMD<double,4> a2, SIMD <double,4> a3, SIMD<double,4> a4,
|
NETGEN_INLINE void SIMDTranspose (SIMD<double,4> a1, SIMD<double,4> a2, SIMD <double,4> a3, SIMD<double,4> a4,
|
||||||
SIMD<double,4> & b1, SIMD<double,4> & b2, SIMD<double,4> & b3, SIMD<double,4> & b4)
|
SIMD<double,4> & b1, SIMD<double,4> & b2, SIMD<double,4> & b3, SIMD<double,4> & b4)
|
||||||
{
|
{
|
||||||
@ -82,7 +84,8 @@ namespace ngcore
|
|||||||
template<int N>
|
template<int N>
|
||||||
NETGEN_INLINE auto HSum (SIMD<double,N> s1, SIMD<double,N> s2, SIMD<double,N> s3, SIMD<double,N> s4 )
|
NETGEN_INLINE auto HSum (SIMD<double,N> s1, SIMD<double,N> s2, SIMD<double,N> s3, SIMD<double,N> s4 )
|
||||||
{
|
{
|
||||||
return SIMD<double,4>(HSum(s1), HSum(s2), HSum(s3), HSum(s4));
|
// return SIMD<double,4>(HSum(s1), HSum(s2), HSum(s3), HSum(s4));
|
||||||
|
return SIMD<double,4>(HSum(s1, s2), HSum(s3,s4));
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
Loading…
x
Reference in New Issue
Block a user