GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-21 20:17:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

Added an optimized version of rsqrt for SSE and AVX that is used when EIGEN_FAST_MATH is defined.

Browse Source
This commit is contained in:
Benoit Steiner 2015-03-02 09:38:47 -08:00
parent ae73859a0a
commit 4fd7f47692
2 changed files with 65 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
Eigen/src/Core/arch/AVX/MathFunctions.h
View File

@ -309,13 +309,44 @@ template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4d psqrt<Packet4d>(const Packet4d& x) { Packet4d psqrt<Packet4d>(const Packet4d& x) {
return _mm256_sqrt_pd(x); return _mm256_sqrt_pd(x);
} }
#if EIGEN_FAST_MATH
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4f prsqrt<Packet4f>(const Packet4f& _x) {
_EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000);
_EIGEN_DECLARE_CONST_Packet4f_FROM_INT(nan, 0x7fc00000);
_EIGEN_DECLARE_CONST_Packet4f(one_point_five, 1.5f);
_EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f);
_EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000);
Packet4f neg_half = pmul(_x, p4f_minus_half);
// select only the inverse sqrt of positive normal inputs (denormals are
// flushed to zero and cause infs as well).
Packet4f le_zero_mask = _mm_cmple_ps(_x, p4f_flt_min);
Packet4f x = _mm_andnot_ps(le_zero_mask, _mm_rsqrt_ps(_x));
// Fill in NaNs and Infs for the negative/zero entries.
Packet4f neg_mask = _mm_cmplt_ps(_x, _mm_setzero_ps());
Packet4f zero_mask = _mm_andnot_ps(neg_mask, le_zero_mask);
Packet4f infs_and_nans = _mm_or_ps(_mm_and_ps(neg_mask, p4f_nan),
_mm_and_ps(zero_mask, p4f_inf));
// Do a single step of Newton's iteration.
x = pmul(x, pmadd(neg_half, pmul(x, x), p4f_one_point_five));
// Insert NaNs and Infs in all the right places.
return _mm_or_ps(x, infs_and_nans);
}
#else
template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet8f prsqrt<Packet8f>(const Packet8f& x) { Packet8f prsqrt<Packet8f>(const Packet8f& x) {
_EIGEN_DECLARE_CONST_Packet8f(one, 1.0f); _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
return _mm256_div_ps(p8f_one, _mm256_sqrt_ps(x)); return _mm256_div_ps(p8f_one, _mm256_sqrt_ps(x));
} }
#endif #endif
template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4d prsqrt<Packet4d>(const Packet4d& x) { Packet4d prsqrt<Packet4d>(const Packet4d& x) {
_EIGEN_DECLARE_CONST_Packet4d(one, 1.0); _EIGEN_DECLARE_CONST_Packet4d(one, 1.0);

34
Eigen/src/Core/arch/SSE/MathFunctions.h
View File

@ -470,12 +470,46 @@ Packet4f psqrt<Packet4f>(const Packet4f& x) { return _mm_sqrt_ps(x); }
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); } Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); }
#if EIGEN_FAST_MATH
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4f prsqrt<Packet4f>(const Packet4f& _x) {
_EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000);
_EIGEN_DECLARE_CONST_Packet4f_FROM_INT(nan, 0x7fc00000);
_EIGEN_DECLARE_CONST_Packet4f(one_point_five, 1.5f);
_EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f);
_EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000);
Packet4f neg_half = pmul(_x, p4f_minus_half);
// select only the inverse sqrt of positive normal inputs (denormals are
// flushed to zero and cause infs as well).
Packet4f le_zero_mask = _mm_cmple_ps(_x, p4f_flt_min);
Packet4f x = _mm_andnot_ps(le_zero_mask, _mm_rsqrt_ps(_x));
// Fill in NaNs and Infs for the negative/zero entries.
Packet4f neg_mask = _mm_cmplt_ps(_x, _mm_setzero_ps());
Packet4f zero_mask = _mm_andnot_ps(neg_mask, le_zero_mask);
Packet4f infs_and_nans = _mm_or_ps(_mm_and_ps(neg_mask, p4f_nan),
_mm_and_ps(zero_mask, p4f_inf));
// Do a single step of Newton's iteration.
x = pmul(x, pmadd(neg_half, pmul(x, x), p4f_one_point_five));
// Insert NaNs and Infs in all the right places.
return _mm_or_ps(x, infs_and_nans);
}
#else
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4f prsqrt<Packet4f>(const Packet4f& x) { Packet4f prsqrt<Packet4f>(const Packet4f& x) {
// Unfortunately we can't use the much faster mm_rqsrt_ps since it only provides an approximation. // Unfortunately we can't use the much faster mm_rqsrt_ps since it only provides an approximation.
return _mm_div_ps(pset1<Packet4f>(1.0f), _mm_sqrt_ps(x)); return _mm_div_ps(pset1<Packet4f>(1.0f), _mm_sqrt_ps(x));
} }
#endif
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet2d prsqrt<Packet2d>(const Packet2d& x) { Packet2d prsqrt<Packet2d>(const Packet2d& x) {
// Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation. // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.