Simplified the code that dispatches vectorized reductions on GPU

3 files changed, 31 insertions, 19 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index e6ff70460..a8e48fced 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -131,7 +131,7 @@ template <typename T, typename Device>
 struct reducer_traits<SumReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = packet_traits<typename PacketType<T, Device>::type>::HasAdd
+    PacketAccess = PacketType<T, Device>::HasAdd
   };
 };
 
@@ -183,7 +183,7 @@ template <typename T, typename Device>
 struct reducer_traits<MeanReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = packet_traits<typename PacketType<T, Device>::type>::HasAdd
+    PacketAccess = PacketType<T, Device>::HasAdd
   };
 };
 
@@ -225,7 +225,7 @@ template <typename T, typename Device>
 struct reducer_traits<MaxReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = packet_traits<typename PacketType<T, Device>::type>::HasMax
+    PacketAccess = PacketType<T, Device>::HasMax
   };
 };
 
@@ -267,7 +267,7 @@ template <typename T, typename Device>
 struct reducer_traits<MinReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = packet_traits<typename PacketType<T, Device>::type>::HasMin
+    PacketAccess = PacketType<T, Device>::HasMin
   };
 };
 
@@ -310,7 +310,7 @@ template <typename T, typename Device>
 struct reducer_traits<ProdReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::MulCost,
-    PacketAccess = packet_traits<typename PacketType<T, Device>::type>::HasMul
+    PacketAccess = PacketType<T, Device>::HasMul
   };
 };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
index 82a905a65..0f3778e6e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
@@ -47,27 +47,39 @@ template <> struct max_n_1<0> {
 
 // Default packet types
 template <typename Scalar, typename Device>
-struct PacketType {
+struct PacketType : internal::packet_traits<Scalar> {
   typedef typename internal::packet_traits<Scalar>::type type;
-  enum { size = internal::unpacket_traits<type>::size };
 };
 
 // For CUDA packet types when using a GpuDevice
 #if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
 template <>
-  struct PacketType<half, GpuDevice> {
+struct PacketType<half, GpuDevice> {
   typedef half2 type;
   static const int size = 2;
- };
-template <>
-struct PacketType<float, GpuDevice> {
-  typedef float4 type;
-  static const int size = 4;
-};
-template <>
-struct PacketType<double, GpuDevice> {
-  typedef double2 type;
-  static const int size = 2;
+  enum {
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasNegate = 1,
+    HasAbs    = 1,
+    HasArg    = 0,
+    HasAbs2   = 0,
+    HasMin    = 1,
+    HasMax    = 1,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasBlend  = 0,
+
+    HasDiv    = 1,
+    HasSqrt   = 1,
+    HasRsqrt  = 1,
+    HasExp    = 1,
+    HasLog    = 1,
+    HasLog1p  = 0,
+    HasLog10  = 0,
+    HasPow    = 1,
+  };
 };
 #endif
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index 1b4fdd03f..d9bbcd858 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -328,7 +328,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
   // Unfortunately nvidia doesn't support well exotic types such as complex,
   // so reduce the scope of the optimized version of the code to the simple case
   // of floats and half floats.
-  #ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_CUDA_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
        (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));