aboutsummaryrefslogtreecommitdiffhomepage
path: root/Eigen/src/Core/arch/AVX
diff options
context:
space:
mode:
authorGravatar Rasmus Munk Larsen <rmlarsen@google.com>2016-10-04 15:09:11 -0700
committerGravatar Rasmus Munk Larsen <rmlarsen@google.com>2016-10-04 15:09:11 -0700
commit7f67e6dfdb635ed1e744c0794bb1f3dd4306678f (patch)
tree5d804948d2f5ef13054e79b3f765fcc460fcf5b2 /Eigen/src/Core/arch/AVX
parent765615609d40f485c718d60b76b35a277c36d5ce (diff)
Update comment for fast sqrt.
Diffstat (limited to 'Eigen/src/Core/arch/AVX')
-rw-r--r--Eigen/src/Core/arch/AVX/MathFunctions.h8
1 files changed, 5 insertions, 3 deletions
diff --git a/Eigen/src/Core/arch/AVX/MathFunctions.h b/Eigen/src/Core/arch/AVX/MathFunctions.h
index da70b6636..25cd992ef 100644
--- a/Eigen/src/Core/arch/AVX/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX/MathFunctions.h
@@ -355,9 +355,11 @@ pexp<Packet4d>(const Packet4d& _x) {
// Functions for sqrt.
// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
-// exact solution. The main advantage of this approach is not just speed, but
-// also the fact that it can be inlined and pipelined with other computations,
-// further reducing its effective latency.
+// exact solution. It does not handle +inf, or denormalized numbers correctly.
+// The main advantage of this approach is not just speed, but also the fact that
+// it can be inlined and pipelined with other computations, further reducing its
+// effective latency. This is similar to Quake3's fast inverse square root.
+// For detail see here: http://www.beyond3d.com/content/articles/8/
#if EIGEN_FAST_MATH
template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f