GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-01 08:14:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

Workaround alignment warnings

Browse Source
This commit is contained in:
Gael Guennebaud 2014-03-30 22:43:47 +02:00
parent e497a27ddc
commit 1c0728043a
3 changed files with 20 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
Eigen/src/Core/arch/AVX/Complex.h
View File

@ -87,15 +87,10 @@ template<> EIGEN_STRONG_INLINE Packet4cf pset1<Packet4cf>(const std::complex<flo
template<> EIGEN_STRONG_INLINE Packet4cf ploaddup<Packet4cf>(const std::complex<float>* from) template<> EIGEN_STRONG_INLINE Packet4cf ploaddup<Packet4cf>(const std::complex<float>* from)
{ {
// This should be optimized. // FIXME The following might be optimized using _mm256_movedup_pd
__m128 complex1 = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)from); Packet2cf a = ploaddup<Packet2cf>(from);
complex1 = _mm_movelh_ps(complex1, complex1); Packet2cf b = ploaddup<Packet2cf>(from+1);
__m128 complex2 = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from+1)); return Packet4cf(_mm256_insertf128_ps(_mm256_castps128_ps256(a.v), b.v, 1));
complex2 = _mm_movelh_ps(complex2, complex2);
__m256 result = _mm256_setzero_ps();
result = _mm256_insertf128_ps(result, complex1, 0);
result = _mm256_insertf128_ps(result, complex2, 1);
return Packet4cf(result);
} }
template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); }
@ -127,10 +122,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet4cf
template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet4cf>(const Packet4cf& a) template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet4cf>(const Packet4cf& a)
{ {
__m128 low = _mm256_extractf128_ps(a.v, 0); return pfirst(Packet2cf(_mm256_castps256_ps128(a.v)));
std::complex<float> res;
_mm_storel_pi((__m64*)&res, low);
return res;
} }
template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) {

5
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@ -173,6 +173,7 @@ template<> EIGEN_STRONG_INLINE Packet8i ploadu<Packet8i>(const int* from) { EIGE
// Loads 4 floats from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, a3} // Loads 4 floats from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, a3}
template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from) template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from)
{ {
// FIXME we should only load the first 128bits
Packet8f tmp = ploadu<Packet8f>(from); Packet8f tmp = ploadu<Packet8f>(from);
Packet8f tmp1 = _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2)); Packet8f tmp1 = _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2));
Packet8f tmp2 = _mm256_permute_ps(tmp, _MM_SHUFFLE(1,1,0,0)); Packet8f tmp2 = _mm256_permute_ps(tmp, _MM_SHUFFLE(1,1,0,0));
@ -181,6 +182,7 @@ template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from)
// Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1} // Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1}
template<> EIGEN_STRONG_INLINE Packet4d ploaddup<Packet4d>(const double* from) template<> EIGEN_STRONG_INLINE Packet4d ploaddup<Packet4d>(const double* from)
{ {
// FIXME we should only load the first 128bits
Packet4d tmp = ploadu<Packet4d>(from); Packet4d tmp = ploadu<Packet4d>(from);
Packet4d tmp1 = _mm256_permute_pd(tmp,0); Packet4d tmp1 = _mm256_permute_pd(tmp,0);
Packet4d tmp2 = _mm256_permute_pd(tmp,3); Packet4d tmp2 = _mm256_permute_pd(tmp,3);
@ -195,7 +197,8 @@ template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet8f&
template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); } template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); }
template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); }
// TODO: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available // NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available
// NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4);
template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, int stride) template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, int stride)
{ {
return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride], return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride],

3
Eigen/src/Core/arch/SSE/Complex.h
View File

@ -122,9 +122,6 @@ template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packe
template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, int stride) template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, int stride)
{ {
/* for (int i = 0; i < 2; i+=2) {
to[stride*i] = std::complex<float>(from.v[i], from.v[i+1]);
}*/
to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)), to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1))); _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)), to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),