Add dedicated implementations of predux_any for AVX512, NEON, and Altivec/VSE

This commit is contained in:
Gael Guennebaud 2019-01-09 16:40:42 +01:00
parent 3f14e0d19e
commit 47810cf5b7
3 changed files with 19 additions and 0 deletions

View File

@ -969,6 +969,13 @@ EIGEN_STRONG_INLINE double predux_max<Packet8d>(const Packet8d& a) {
return pfirst(_mm256_max_pd(res, _mm256_shuffle_pd(res, res, 1)));
}
template<> EIGEN_STRONG_INLINE bool predux_any(const Packet16f& x)
{
Packet16i xi = _mm512_castps_si512(x);
__mmask16 tmp = _mm512_test_epi32_mask(xi,xi);
return !_mm512_kortestz(tmp,tmp);
}
template <int Offset>
struct palign_impl<Offset, Packet16f> {
static EIGEN_STRONG_INLINE void run(Packet16f& first,

View File

@ -720,6 +720,11 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
return pfirst(res);
}
template<> EIGEN_STRONG_INLINE bool predux_any(const Packet4f& x)
{
return vec_any_ne(x, pzero(x));
}
template<int Offset>
struct palign_impl<Offset,Packet4f>
{

View File

@ -551,6 +551,13 @@ template<> EIGEN_STRONG_INLINE int32_t predux_max<Packet4i>(const Packet4i& a)
return vget_lane_s32(max, 0);
}
template<> EIGEN_STRONG_INLINE bool predux_any(const Packet4f& x)
{
uint32x2_t tmp = vorr_u32(vget_low_u32( vreinterpretq_u32_f32(x)),
vget_high_u32(vreinterpretq_u32_f32(x)));
return vget_lane_u32(vpmax_u32(tmp,tmp),0);
}
// this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors,
// see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074
#define PALIGN_NEON(Offset,Type,Command) \