devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove unused packet op "preduxp".

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2020-04-23 18:17:14 +00:00
parent 0aebe19aca
commit e80ec24357
16 changed files with 1 additions and 1009 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
Eigen/src/Core/arch/AVX/Complex.h
View File

@@ -151,21 +151,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet4cf>(const Packe
Packet2cf(_mm256_extractf128_ps(a.v,1))));
}
template<> EIGEN_STRONG_INLINE Packet4cf preduxp<Packet4cf>(const Packet4cf* vecs)
{
Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0));
Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0));
t0 = _mm256_hadd_ps(t0,t1);
Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0));
Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0));
t2 = _mm256_hadd_ps(t2,t3);
t1 = _mm256_permute2f128_ps(t0,t2, 0 + (2<<4));
t3 = _mm256_permute2f128_ps(t0,t2, 1 + (3<<4));
return Packet4cf(_mm256_add_ps(t1,t3));
}
template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet4cf>(const Packet4cf& a)
{
return predux_mul(pmul(Packet2cf(_mm256_extractf128_ps(a.v, 0)),
@@ -348,14 +333,6 @@ template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet2cd>(const Pack
Packet1cd(_mm256_extractf128_pd(a.v,1))));
}
template<> EIGEN_STRONG_INLINE Packet2cd preduxp<Packet2cd>(const Packet2cd* vecs)
{
Packet4d t0 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 0 + (2<<4));
Packet4d t1 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 1 + (3<<4));
return Packet2cd(_mm256_add_pd(t0,t1));
}
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet2cd>(const Packet2cd& a)
{
return predux(pmul(Packet1cd(_mm256_extractf128_pd(a.v,0)),

57
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@@ -627,49 +627,6 @@ template<> EIGEN_STRONG_INLINE Packet4d pldexp<Packet4d>(const Packet4d& a, cons
return pmul(a,_mm256_castsi256_pd(e));
}
// preduxp should be ok
// FIXME: why is this ok? why isn't the simply implementation working as expected?
template<> EIGEN_STRONG_INLINE Packet8f preduxp<Packet8f>(const Packet8f* vecs)
{
__m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]);
__m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]);
__m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]);
__m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]);
__m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1);
__m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2);
__m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3);
__m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4);
__m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
__m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
__m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
__m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
__m256 sum1 = _mm256_add_ps(perm1, hsum5);
__m256 sum2 = _mm256_add_ps(perm2, hsum6);
__m256 sum3 = _mm256_add_ps(perm3, hsum7);
__m256 sum4 = _mm256_add_ps(perm4, hsum8);
__m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
__m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
__m256 final = _mm256_blend_ps(blend1, blend2, 0xf0);
return final;
}
template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs)
{
Packet4d tmp0, tmp1;
tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]);
tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]);
tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
return _mm256_blend_pd(tmp0, tmp1, 0xC);
}
template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a)
{
return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1))));
@@ -1105,20 +1062,6 @@ template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h&
return Eigen::half(reduced);
}
template<> EIGEN_STRONG_INLINE Packet8h preduxp<Packet8h>(const Packet8h* p) {
Packet8f pf[8];
pf[0] = half2float(p[0]);
pf[1] = half2float(p[1]);
pf[2] = half2float(p[2]);
pf[3] = half2float(p[3]);
pf[4] = half2float(p[4]);
pf[5] = half2float(p[5]);
pf[6] = half2float(p[6]);
pf[7] = half2float(p[7]);
Packet8f reduced = preduxp<Packet8f>(pf);
return float2half(reduced);
}
template<> EIGEN_STRONG_INLINE Packet8h preverse(const Packet8h& a)
{
__m128i m = _mm_setr_epi8(14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1);