commit 4d7f31710299fd869def962f2070c252ae1aaa67
author Rasmus Munk Larsen <rmlarsen@google.com> Thu Feb 21 13:32:13 2019 -0800
committer Rasmus Munk Larsen <rmlarsen@google.com> Thu Feb 21 13:32:13 2019 -0800
tree f068bcdfb8e2c29bb761d5aa19a392a53824b824
parent 2a39659d793fcde656593bbf01948bc0bd568181

Add a few missing packet ops: cmp_eq for NEON. pfloor for GPU.

Eigen/src/Core/arch/GPU/PacketMath.h

diff --git a/Eigen/src/Core/arch/GPU/PacketMath.h b/Eigen/src/Core/arch/GPU/PacketMath.h
index c1b097f..cd4615a 100644
--- a/Eigen/src/Core/arch/GPU/PacketMath.h
+++ b/Eigen/src/Core/arch/GPU/PacketMath.h
@@ -53,6 +53,7 @@
     HasBetaInc = 1,
 
     HasBlend = 0,
+    HasFloor = 1,
   };
 };
 
@@ -86,6 +87,7 @@
     HasBetaInc = 1,
 
     HasBlend = 0,
+    HasFloor = 1,
   };
 };
 
@@ -408,6 +410,13 @@
   return make_double2(fabs(a.x), fabs(a.y));
 }
 
+template<> EIGEN_DEVICE_FUNC inline float4  pfloor<float4>(const float4& a) {
+  return make_float4(floorf(a.x), floorf(a.y), floorf(a.z), floorf(a.w));
+}
+template<> EIGEN_DEVICE_FUNC inline double2 pfloor<double2>(const double2& a) {
+  return make_double2(floor(a.x), floor(a.y));
+}
+
 EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<float4,4>& kernel) {
   float tmp = kernel.packet[0].y;