commit abdbe8562e889a0ca0877d607cfd5c4cbf937e3a
author Benoit Steiner <benoit.steiner.goog@gmail.com> Tue Mar 24 10:45:46 2015 -0700
committer Benoit Steiner <benoit.steiner.goog@gmail.com> Tue Mar 24 10:45:46 2015 -0700
tree d7dcce3c42887a81d2d2fcf0a66d1ade96e53ca8
parent 4472f3e57884202b52c551da0b294c2883e2ccbf

Fixed the CUDA packet primitives

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

diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/CUDA/PacketMath.h
index 19749c8..ceed1d1 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMath.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -197,21 +197,21 @@
 }
 #endif
 
-template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, int stride) {
+template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, Index stride) {
   return make_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]);
 }
 
-template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, int stride) {
+template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, Index stride) {
   return make_double2(from[0*stride], from[1*stride]);
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, int stride) {
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, Index stride) {
   to[stride*0] = from.x;
   to[stride*1] = from.y;
   to[stride*2] = from.z;
   to[stride*3] = from.w;
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, int stride) {
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, Index stride) {
   to[stride*0] = from.x;
   to[stride*1] = from.y;
 }
@@ -245,14 +245,14 @@
 }
 
 template<> EIGEN_DEVICE_FUNC inline float4  pabs<float4>(const float4& a) {
-  return make_float4(fabs(a.x), fabs(a.y), fabs(a.z), fabs(a.w));
+  return make_float4(fabsf(a.x), fabsf(a.y), fabsf(a.z), fabsf(a.w));
 }
 template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) {
-  return make_double2(abs(a.x), abs(a.y));
+  return make_double2(fabs(a.x), fabs(a.y));
 }
 
 
-template<> EIGEN_DEVICE_FUNC inline void
+EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<float4,4>& kernel) {
   double tmp = kernel.packet[0].y;
   kernel.packet[0].y = kernel.packet[1].x;
@@ -279,7 +279,7 @@
   kernel.packet[3].z = tmp;
 }
 
-template<> EIGEN_DEVICE_FUNC inline void
+EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<double2,2>& kernel) {
   double tmp = kernel.packet[0].y;
   kernel.packet[0].y = kernel.packet[1].x;