Fixed the optimized AVX implementation of the fast rsqrt function

1 files changed, 15 insertions, 15 deletions

diff --git a/Eigen/src/Core/arch/AVX/MathFunctions.h b/Eigen/src/Core/arch/AVX/MathFunctions.h
index 3031e7233..ef80a8b2d 100644
--- a/Eigen/src/Core/arch/AVX/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX/MathFunctions.h
@@ -312,31 +312,31 @@ Packet4d psqrt<Packet4d>(const Packet4d& 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);
+Packet8f prsqrt<Packet8f>(const Packet8f& _x) {
+ _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
 
-  Packet4f neg_half = pmul(_x, p4f_minus_half);
+  Packet8f neg_half = pmul(_x, p8f_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));
+  Packet8f le_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_LT_OQ);
+  Packet8f x = _mm256_andnot_ps(le_zero_mask, _mm256_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));
+  Packet8f neg_mask = _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_LT_OQ);
+  Packet8f zero_mask = _mm256_andnot_ps(neg_mask, le_zero_mask);
+  Packet8f infs_and_nans = _mm256_or_ps(_mm256_and_ps(neg_mask, p8f_nan),
+                                        _mm256_and_ps(zero_mask, p8f_inf));
 
   // Do a single step of Newton's iteration.
-  x = pmul(x, pmadd(neg_half, pmul(x, x), p4f_one_point_five));
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five));
 
   // Insert NaNs and Infs in all the right places.
-  return _mm_or_ps(x, infs_and_nans);
+  return _mm256_or_ps(x, infs_and_nans);
 }
 
 #else